1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  22  * Copyright (c) 1994, 2010, Oracle and/or its affiliates. All rights reserved.
  23  * Copyright (c) 2012, 2017 by Delphix. All rights reserved.
  24  * Copyright 2015 Nexenta Systems, Inc.  All rights reserved.
  25  */
  26 
  27 /*
  28  * Kernel memory allocator, as described in the following two papers and a
  29  * statement about the consolidator:
  30  *
  31  * Jeff Bonwick,
  32  * The Slab Allocator: An Object-Caching Kernel Memory Allocator.
  33  * Proceedings of the Summer 1994 Usenix Conference.
  34  * Available as /shared/sac/PSARC/1994/028/materials/kmem.pdf.
  35  *
  36  * Jeff Bonwick and Jonathan Adams,
  37  * Magazines and vmem: Extending the Slab Allocator to Many CPUs and
  38  * Arbitrary Resources.
  39  * Proceedings of the 2001 Usenix Conference.
  40  * Available as /shared/sac/PSARC/2000/550/materials/vmem.pdf.
  41  *
  42  * kmem Slab Consolidator Big Theory Statement:
  43  *
  44  * 1. Motivation
  45  *
  46  * As stated in Bonwick94, slabs provide the following advantages over other
  47  * allocation structures in terms of memory fragmentation:
  48  *
  49  *  - Internal fragmentation (per-buffer wasted space) is minimal.
  50  *  - Severe external fragmentation (unused buffers on the free list) is
  51  *    unlikely.
  52  *
  53  * Segregating objects by size eliminates one source of external fragmentation,
  54  * and according to Bonwick:
  55  *
  56  *   The other reason that slabs reduce external fragmentation is that all
  57  *   objects in a slab are of the same type, so they have the same lifetime
  58  *   distribution. The resulting segregation of short-lived and long-lived
  59  *   objects at slab granularity reduces the likelihood of an entire page being
  60  *   held hostage due to a single long-lived allocation [Barrett93, Hanson90].
  61  *
  62  * While unlikely, severe external fragmentation remains possible. Clients that
  63  * allocate both short- and long-lived objects from the same cache cannot
  64  * anticipate the distribution of long-lived objects within the allocator's slab
  65  * implementation. Even a small percentage of long-lived objects distributed
  66  * randomly across many slabs can lead to a worst case scenario where the client
  67  * frees the majority of its objects and the system gets back almost none of the
  68  * slabs. Despite the client doing what it reasonably can to help the system
  69  * reclaim memory, the allocator cannot shake free enough slabs because of
  70  * lonely allocations stubbornly hanging on. Although the allocator is in a
  71  * position to diagnose the fragmentation, there is nothing that the allocator
  72  * by itself can do about it. It only takes a single allocated object to prevent
  73  * an entire slab from being reclaimed, and any object handed out by
  74  * kmem_cache_alloc() is by definition in the client's control. Conversely,
  75  * although the client is in a position to move a long-lived object, it has no
  76  * way of knowing if the object is causing fragmentation, and if so, where to
  77  * move it. A solution necessarily requires further cooperation between the
  78  * allocator and the client.
  79  *
  80  * 2. Move Callback
  81  *
  82  * The kmem slab consolidator therefore adds a move callback to the
  83  * allocator/client interface, improving worst-case external fragmentation in
  84  * kmem caches that supply a function to move objects from one memory location
  85  * to another. In a situation of low memory kmem attempts to consolidate all of
  86  * a cache's slabs at once; otherwise it works slowly to bring external
  87  * fragmentation within the 1/8 limit guaranteed for internal fragmentation,
  88  * thereby helping to avoid a low memory situation in the future.
  89  *
  90  * The callback has the following signature:
  91  *
  92  *   kmem_cbrc_t move(void *old, void *new, size_t size, void *user_arg)
  93  *
  94  * It supplies the kmem client with two addresses: the allocated object that
  95  * kmem wants to move and a buffer selected by kmem for the client to use as the
  96  * copy destination. The callback is kmem's way of saying "Please get off of
  97  * this buffer and use this one instead." kmem knows where it wants to move the
  98  * object in order to best reduce fragmentation. All the client needs to know
  99  * about the second argument (void *new) is that it is an allocated, constructed
 100  * object ready to take the contents of the old object. When the move function
 101  * is called, the system is likely to be low on memory, and the new object
 102  * spares the client from having to worry about allocating memory for the
 103  * requested move. The third argument supplies the size of the object, in case a
 104  * single move function handles multiple caches whose objects differ only in
 105  * size (such as zio_buf_512, zio_buf_1024, etc). Finally, the same optional
 106  * user argument passed to the constructor, destructor, and reclaim functions is
 107  * also passed to the move callback.
 108  *
 109  * 2.1 Setting the Move Callback
 110  *
 111  * The client sets the move callback after creating the cache and before
 112  * allocating from it:
 113  *
 114  *      object_cache = kmem_cache_create(...);
 115  *      kmem_cache_set_move(object_cache, object_move);
 116  *
 117  * 2.2 Move Callback Return Values
 118  *
 119  * Only the client knows about its own data and when is a good time to move it.
 120  * The client is cooperating with kmem to return unused memory to the system,
 121  * and kmem respectfully accepts this help at the client's convenience. When
 122  * asked to move an object, the client can respond with any of the following:
 123  *
 124  *   typedef enum kmem_cbrc {
 125  *           KMEM_CBRC_YES,
 126  *           KMEM_CBRC_NO,
 127  *           KMEM_CBRC_LATER,
 128  *           KMEM_CBRC_DONT_NEED,
 129  *           KMEM_CBRC_DONT_KNOW
 130  *   } kmem_cbrc_t;
 131  *
 132  * The client must not explicitly kmem_cache_free() either of the objects passed
 133  * to the callback, since kmem wants to free them directly to the slab layer
 134  * (bypassing the per-CPU magazine layer). The response tells kmem which of the
 135  * objects to free:
 136  *
 137  *       YES: (Did it) The client moved the object, so kmem frees the old one.
 138  *        NO: (Never) The client refused, so kmem frees the new object (the
 139  *            unused copy destination). kmem also marks the slab of the old
 140  *            object so as not to bother the client with further callbacks for
 141  *            that object as long as the slab remains on the partial slab list.
 142  *            (The system won't be getting the slab back as long as the
 143  *            immovable object holds it hostage, so there's no point in moving
 144  *            any of its objects.)
 145  *     LATER: The client is using the object and cannot move it now, so kmem
 146  *            frees the new object (the unused copy destination). kmem still
 147  *            attempts to move other objects off the slab, since it expects to
 148  *            succeed in clearing the slab in a later callback. The client
 149  *            should use LATER instead of NO if the object is likely to become
 150  *            movable very soon.
 151  * DONT_NEED: The client no longer needs the object, so kmem frees the old along
 152  *            with the new object (the unused copy destination). This response
 153  *            is the client's opportunity to be a model citizen and give back as
 154  *            much as it can.
 155  * DONT_KNOW: The client does not know about the object because
 156  *            a) the client has just allocated the object and not yet put it
 157  *               wherever it expects to find known objects
 158  *            b) the client has removed the object from wherever it expects to
 159  *               find known objects and is about to free it, or
 160  *            c) the client has freed the object.
 161  *            In all these cases (a, b, and c) kmem frees the new object (the
 162  *            unused copy destination).  In the first case, the object is in
 163  *            use and the correct action is that for LATER; in the latter two
 164  *            cases, we know that the object is either freed or about to be
 165  *            freed, in which case it is either already in a magazine or about
 166  *            to be in one.  In these cases, we know that the object will either
 167  *            be reallocated and reused, or it will end up in a full magazine
 168  *            that will be reaped (thereby liberating the slab).  Because it
 169  *            is prohibitively expensive to differentiate these cases, and
 170  *            because the defrag code is executed when we're low on memory
 171  *            (thereby biasing the system to reclaim full magazines) we treat
 172  *            all DONT_KNOW cases as LATER and rely on cache reaping to
 173  *            generally clean up full magazines.  While we take the same action
 174  *            for these cases, we maintain their semantic distinction:  if
 175  *            defragmentation is not occurring, it is useful to know if this
 176  *            is due to objects in use (LATER) or objects in an unknown state
 177  *            of transition (DONT_KNOW).
 178  *
 179  * 2.3 Object States
 180  *
 181  * Neither kmem nor the client can be assumed to know the object's whereabouts
 182  * at the time of the callback. An object belonging to a kmem cache may be in
 183  * any of the following states:
 184  *
 185  * 1. Uninitialized on the slab
 186  * 2. Allocated from the slab but not constructed (still uninitialized)
 187  * 3. Allocated from the slab, constructed, but not yet ready for business
 188  *    (not in a valid state for the move callback)
 189  * 4. In use (valid and known to the client)
 190  * 5. About to be freed (no longer in a valid state for the move callback)
 191  * 6. Freed to a magazine (still constructed)
 192  * 7. Allocated from a magazine, not yet ready for business (not in a valid
 193  *    state for the move callback), and about to return to state #4
 194  * 8. Deconstructed on a magazine that is about to be freed
 195  * 9. Freed to the slab
 196  *
 197  * Since the move callback may be called at any time while the object is in any
 198  * of the above states (except state #1), the client needs a safe way to
 199  * determine whether or not it knows about the object. Specifically, the client
 200  * needs to know whether or not the object is in state #4, the only state in
 201  * which a move is valid. If the object is in any other state, the client should
 202  * immediately return KMEM_CBRC_DONT_KNOW, since it is unsafe to access any of
 203  * the object's fields.
 204  *
 205  * Note that although an object may be in state #4 when kmem initiates the move
 206  * request, the object may no longer be in that state by the time kmem actually
 207  * calls the move function. Not only does the client free objects
 208  * asynchronously, kmem itself puts move requests on a queue where thay are
 209  * pending until kmem processes them from another context. Also, objects freed
 210  * to a magazine appear allocated from the point of view of the slab layer, so
 211  * kmem may even initiate requests for objects in a state other than state #4.
 212  *
 213  * 2.3.1 Magazine Layer
 214  *
 215  * An important insight revealed by the states listed above is that the magazine
 216  * layer is populated only by kmem_cache_free(). Magazines of constructed
 217  * objects are never populated directly from the slab layer (which contains raw,
 218  * unconstructed objects). Whenever an allocation request cannot be satisfied
 219  * from the magazine layer, the magazines are bypassed and the request is
 220  * satisfied from the slab layer (creating a new slab if necessary). kmem calls
 221  * the object constructor only when allocating from the slab layer, and only in
 222  * response to kmem_cache_alloc() or to prepare the destination buffer passed in
 223  * the move callback. kmem does not preconstruct objects in anticipation of
 224  * kmem_cache_alloc().
 225  *
 226  * 2.3.2 Object Constructor and Destructor
 227  *
 228  * If the client supplies a destructor, it must be valid to call the destructor
 229  * on a newly created object (immediately after the constructor).
 230  *
 231  * 2.4 Recognizing Known Objects
 232  *
 233  * There is a simple test to determine safely whether or not the client knows
 234  * about a given object in the move callback. It relies on the fact that kmem
 235  * guarantees that the object of the move callback has only been touched by the
 236  * client itself or else by kmem. kmem does this by ensuring that none of the
 237  * cache's slabs are freed to the virtual memory (VM) subsystem while a move
 238  * callback is pending. When the last object on a slab is freed, if there is a
 239  * pending move, kmem puts the slab on a per-cache dead list and defers freeing
 240  * slabs on that list until all pending callbacks are completed. That way,
 241  * clients can be certain that the object of a move callback is in one of the
 242  * states listed above, making it possible to distinguish known objects (in
 243  * state #4) using the two low order bits of any pointer member (with the
 244  * exception of 'char *' or 'short *' which may not be 4-byte aligned on some
 245  * platforms).
 246  *
 247  * The test works as long as the client always transitions objects from state #4
 248  * (known, in use) to state #5 (about to be freed, invalid) by setting the low
 249  * order bit of the client-designated pointer member. Since kmem only writes
 250  * invalid memory patterns, such as 0xbaddcafe to uninitialized memory and
 251  * 0xdeadbeef to freed memory, any scribbling on the object done by kmem is
 252  * guaranteed to set at least one of the two low order bits. Therefore, given an
 253  * object with a back pointer to a 'container_t *o_container', the client can
 254  * test
 255  *
 256  *      container_t *container = object->o_container;
 257  *      if ((uintptr_t)container & 0x3) {
 258  *              return (KMEM_CBRC_DONT_KNOW);
 259  *      }
 260  *
 261  * Typically, an object will have a pointer to some structure with a list or
 262  * hash where objects from the cache are kept while in use. Assuming that the
 263  * client has some way of knowing that the container structure is valid and will
 264  * not go away during the move, and assuming that the structure includes a lock
 265  * to protect whatever collection is used, then the client would continue as
 266  * follows:
 267  *
 268  *      // Ensure that the container structure does not go away.
 269  *      if (container_hold(container) == 0) {
 270  *              return (KMEM_CBRC_DONT_KNOW);
 271  *      }
 272  *      mutex_enter(&container->c_objects_lock);
 273  *      if (container != object->o_container) {
 274  *              mutex_exit(&container->c_objects_lock);
 275  *              container_rele(container);
 276  *              return (KMEM_CBRC_DONT_KNOW);
 277  *      }
 278  *
 279  * At this point the client knows that the object cannot be freed as long as
 280  * c_objects_lock is held. Note that after acquiring the lock, the client must
 281  * recheck the o_container pointer in case the object was removed just before
 282  * acquiring the lock.
 283  *
 284  * When the client is about to free an object, it must first remove that object
 285  * from the list, hash, or other structure where it is kept. At that time, to
 286  * mark the object so it can be distinguished from the remaining, known objects,
 287  * the client sets the designated low order bit:
 288  *
 289  *      mutex_enter(&container->c_objects_lock);
 290  *      object->o_container = (void *)((uintptr_t)object->o_container | 0x1);
 291  *      list_remove(&container->c_objects, object);
 292  *      mutex_exit(&container->c_objects_lock);
 293  *
 294  * In the common case, the object is freed to the magazine layer, where it may
 295  * be reused on a subsequent allocation without the overhead of calling the
 296  * constructor. While in the magazine it appears allocated from the point of
 297  * view of the slab layer, making it a candidate for the move callback. Most
 298  * objects unrecognized by the client in the move callback fall into this
 299  * category and are cheaply distinguished from known objects by the test
 300  * described earlier. Because searching magazines is prohibitively expensive
 301  * for kmem, clients that do not mark freed objects (and therefore return
 302  * KMEM_CBRC_DONT_KNOW for large numbers of objects) may find defragmentation
 303  * efficacy reduced.
 304  *
 305  * Invalidating the designated pointer member before freeing the object marks
 306  * the object to be avoided in the callback, and conversely, assigning a valid
 307  * value to the designated pointer member after allocating the object makes the
 308  * object fair game for the callback:
 309  *
 310  *      ... allocate object ...
 311  *      ... set any initial state not set by the constructor ...
 312  *
 313  *      mutex_enter(&container->c_objects_lock);
 314  *      list_insert_tail(&container->c_objects, object);
 315  *      membar_producer();
 316  *      object->o_container = container;
 317  *      mutex_exit(&container->c_objects_lock);
 318  *
 319  * Note that everything else must be valid before setting o_container makes the
 320  * object fair game for the move callback. The membar_producer() call ensures
 321  * that all the object's state is written to memory before setting the pointer
 322  * that transitions the object from state #3 or #7 (allocated, constructed, not
 323  * yet in use) to state #4 (in use, valid). That's important because the move
 324  * function has to check the validity of the pointer before it can safely
 325  * acquire the lock protecting the collection where it expects to find known
 326  * objects.
 327  *
 328  * This method of distinguishing known objects observes the usual symmetry:
 329  * invalidating the designated pointer is the first thing the client does before
 330  * freeing the object, and setting the designated pointer is the last thing the
 331  * client does after allocating the object. Of course, the client is not
 332  * required to use this method. Fundamentally, how the client recognizes known
 333  * objects is completely up to the client, but this method is recommended as an
 334  * efficient and safe way to take advantage of the guarantees made by kmem. If
 335  * the entire object is arbitrary data without any markable bits from a suitable
 336  * pointer member, then the client must find some other method, such as
 337  * searching a hash table of known objects.
 338  *
 339  * 2.5 Preventing Objects From Moving
 340  *
 341  * Besides a way to distinguish known objects, the other thing that the client
 342  * needs is a strategy to ensure that an object will not move while the client
 343  * is actively using it. The details of satisfying this requirement tend to be
 344  * highly cache-specific. It might seem that the same rules that let a client
 345  * remove an object safely should also decide when an object can be moved
 346  * safely. However, any object state that makes a removal attempt invalid is
 347  * likely to be long-lasting for objects that the client does not expect to
 348  * remove. kmem knows nothing about the object state and is equally likely (from
 349  * the client's point of view) to request a move for any object in the cache,
 350  * whether prepared for removal or not. Even a low percentage of objects stuck
 351  * in place by unremovability will defeat the consolidator if the stuck objects
 352  * are the same long-lived allocations likely to hold slabs hostage.
 353  * Fundamentally, the consolidator is not aimed at common cases. Severe external
 354  * fragmentation is a worst case scenario manifested as sparsely allocated
 355  * slabs, by definition a low percentage of the cache's objects. When deciding
 356  * what makes an object movable, keep in mind the goal of the consolidator: to
 357  * bring worst-case external fragmentation within the limits guaranteed for
 358  * internal fragmentation. Removability is a poor criterion if it is likely to
 359  * exclude more than an insignificant percentage of objects for long periods of
 360  * time.
 361  *
 362  * A tricky general solution exists, and it has the advantage of letting you
 363  * move any object at almost any moment, practically eliminating the likelihood
 364  * that an object can hold a slab hostage. However, if there is a cache-specific
 365  * way to ensure that an object is not actively in use in the vast majority of
 366  * cases, a simpler solution that leverages this cache-specific knowledge is
 367  * preferred.
 368  *
 369  * 2.5.1 Cache-Specific Solution
 370  *
 371  * As an example of a cache-specific solution, the ZFS znode cache takes
 372  * advantage of the fact that the vast majority of znodes are only being
 373  * referenced from the DNLC. (A typical case might be a few hundred in active
 374  * use and a hundred thousand in the DNLC.) In the move callback, after the ZFS
 375  * client has established that it recognizes the znode and can access its fields
 376  * safely (using the method described earlier), it then tests whether the znode
 377  * is referenced by anything other than the DNLC. If so, it assumes that the
 378  * znode may be in active use and is unsafe to move, so it drops its locks and
 379  * returns KMEM_CBRC_LATER. The advantage of this strategy is that everywhere
 380  * else znodes are used, no change is needed to protect against the possibility
 381  * of the znode moving. The disadvantage is that it remains possible for an
 382  * application to hold a znode slab hostage with an open file descriptor.
 383  * However, this case ought to be rare and the consolidator has a way to deal
 384  * with it: If the client responds KMEM_CBRC_LATER repeatedly for the same
 385  * object, kmem eventually stops believing it and treats the slab as if the
 386  * client had responded KMEM_CBRC_NO. Having marked the hostage slab, kmem can
 387  * then focus on getting it off of the partial slab list by allocating rather
 388  * than freeing all of its objects. (Either way of getting a slab off the
 389  * free list reduces fragmentation.)
 390  *
 391  * 2.5.2 General Solution
 392  *
 393  * The general solution, on the other hand, requires an explicit hold everywhere
 394  * the object is used to prevent it from moving. To keep the client locking
 395  * strategy as uncomplicated as possible, kmem guarantees the simplifying
 396  * assumption that move callbacks are sequential, even across multiple caches.
 397  * Internally, a global queue processed by a single thread supports all caches
 398  * implementing the callback function. No matter how many caches supply a move
 399  * function, the consolidator never moves more than one object at a time, so the
 400  * client does not have to worry about tricky lock ordering involving several
 401  * related objects from different kmem caches.
 402  *
 403  * The general solution implements the explicit hold as a read-write lock, which
 404  * allows multiple readers to access an object from the cache simultaneously
 405  * while a single writer is excluded from moving it. A single rwlock for the
 406  * entire cache would lock out all threads from using any of the cache's objects
 407  * even though only a single object is being moved, so to reduce contention,
 408  * the client can fan out the single rwlock into an array of rwlocks hashed by
 409  * the object address, making it probable that moving one object will not
 410  * prevent other threads from using a different object. The rwlock cannot be a
 411  * member of the object itself, because the possibility of the object moving
 412  * makes it unsafe to access any of the object's fields until the lock is
 413  * acquired.
 414  *
 415  * Assuming a small, fixed number of locks, it's possible that multiple objects
 416  * will hash to the same lock. A thread that needs to use multiple objects in
 417  * the same function may acquire the same lock multiple times. Since rwlocks are
 418  * reentrant for readers, and since there is never more than a single writer at
 419  * a time (assuming that the client acquires the lock as a writer only when
 420  * moving an object inside the callback), there would seem to be no problem.
 421  * However, a client locking multiple objects in the same function must handle
 422  * one case of potential deadlock: Assume that thread A needs to prevent both
 423  * object 1 and object 2 from moving, and thread B, the callback, meanwhile
 424  * tries to move object 3. It's possible, if objects 1, 2, and 3 all hash to the
 425  * same lock, that thread A will acquire the lock for object 1 as a reader
 426  * before thread B sets the lock's write-wanted bit, preventing thread A from
 427  * reacquiring the lock for object 2 as a reader. Unable to make forward
 428  * progress, thread A will never release the lock for object 1, resulting in
 429  * deadlock.
 430  *
 431  * There are two ways of avoiding the deadlock just described. The first is to
 432  * use rw_tryenter() rather than rw_enter() in the callback function when
 433  * attempting to acquire the lock as a writer. If tryenter discovers that the
 434  * same object (or another object hashed to the same lock) is already in use, it
 435  * aborts the callback and returns KMEM_CBRC_LATER. The second way is to use
 436  * rprwlock_t (declared in common/fs/zfs/sys/rprwlock.h) instead of rwlock_t,
 437  * since it allows a thread to acquire the lock as a reader in spite of a
 438  * waiting writer. This second approach insists on moving the object now, no
 439  * matter how many readers the move function must wait for in order to do so,
 440  * and could delay the completion of the callback indefinitely (blocking
 441  * callbacks to other clients). In practice, a less insistent callback using
 442  * rw_tryenter() returns KMEM_CBRC_LATER infrequently enough that there seems
 443  * little reason to use anything else.
 444  *
 445  * Avoiding deadlock is not the only problem that an implementation using an
 446  * explicit hold needs to solve. Locking the object in the first place (to
 447  * prevent it from moving) remains a problem, since the object could move
 448  * between the time you obtain a pointer to the object and the time you acquire
 449  * the rwlock hashed to that pointer value. Therefore the client needs to
 450  * recheck the value of the pointer after acquiring the lock, drop the lock if
 451  * the value has changed, and try again. This requires a level of indirection:
 452  * something that points to the object rather than the object itself, that the
 453  * client can access safely while attempting to acquire the lock. (The object
 454  * itself cannot be referenced safely because it can move at any time.)
 455  * The following lock-acquisition function takes whatever is safe to reference
 456  * (arg), follows its pointer to the object (using function f), and tries as
 457  * often as necessary to acquire the hashed lock and verify that the object
 458  * still has not moved:
 459  *
 460  *      object_t *
 461  *      object_hold(object_f f, void *arg)
 462  *      {
 463  *              object_t *op;
 464  *
 465  *              op = f(arg);
 466  *              if (op == NULL) {
 467  *                      return (NULL);
 468  *              }
 469  *
 470  *              rw_enter(OBJECT_RWLOCK(op), RW_READER);
 471  *              while (op != f(arg)) {
 472  *                      rw_exit(OBJECT_RWLOCK(op));
 473  *                      op = f(arg);
 474  *                      if (op == NULL) {
 475  *                              break;
 476  *                      }
 477  *                      rw_enter(OBJECT_RWLOCK(op), RW_READER);
 478  *              }
 479  *
 480  *              return (op);
 481  *      }
 482  *
 483  * The OBJECT_RWLOCK macro hashes the object address to obtain the rwlock. The
 484  * lock reacquisition loop, while necessary, almost never executes. The function
 485  * pointer f (used to obtain the object pointer from arg) has the following type
 486  * definition:
 487  *
 488  *      typedef object_t *(*object_f)(void *arg);
 489  *
 490  * An object_f implementation is likely to be as simple as accessing a structure
 491  * member:
 492  *
 493  *      object_t *
 494  *      s_object(void *arg)
 495  *      {
 496  *              something_t *sp = arg;
 497  *              return (sp->s_object);
 498  *      }
 499  *
 500  * The flexibility of a function pointer allows the path to the object to be
 501  * arbitrarily complex and also supports the notion that depending on where you
 502  * are using the object, you may need to get it from someplace different.
 503  *
 504  * The function that releases the explicit hold is simpler because it does not
 505  * have to worry about the object moving:
 506  *
 507  *      void
 508  *      object_rele(object_t *op)
 509  *      {
 510  *              rw_exit(OBJECT_RWLOCK(op));
 511  *      }
 512  *
 513  * The caller is spared these details so that obtaining and releasing an
 514  * explicit hold feels like a simple mutex_enter()/mutex_exit() pair. The caller
 515  * of object_hold() only needs to know that the returned object pointer is valid
 516  * if not NULL and that the object will not move until released.
 517  *
 518  * Although object_hold() prevents an object from moving, it does not prevent it
 519  * from being freed. The caller must take measures before calling object_hold()
 520  * (afterwards is too late) to ensure that the held object cannot be freed. The
 521  * caller must do so without accessing the unsafe object reference, so any lock
 522  * or reference count used to ensure the continued existence of the object must
 523  * live outside the object itself.
 524  *
 525  * Obtaining a new object is a special case where an explicit hold is impossible
 526  * for the caller. Any function that returns a newly allocated object (either as
 527  * a return value, or as an in-out paramter) must return it already held; after
 528  * the caller gets it is too late, since the object cannot be safely accessed
 529  * without the level of indirection described earlier. The following
 530  * object_alloc() example uses the same code shown earlier to transition a new
 531  * object into the state of being recognized (by the client) as a known object.
 532  * The function must acquire the hold (rw_enter) before that state transition
 533  * makes the object movable:
 534  *
 535  *      static object_t *
 536  *      object_alloc(container_t *container)
 537  *      {
 538  *              object_t *object = kmem_cache_alloc(object_cache, 0);
 539  *              ... set any initial state not set by the constructor ...
 540  *              rw_enter(OBJECT_RWLOCK(object), RW_READER);
 541  *              mutex_enter(&container->c_objects_lock);
 542  *              list_insert_tail(&container->c_objects, object);
 543  *              membar_producer();
 544  *              object->o_container = container;
 545  *              mutex_exit(&container->c_objects_lock);
 546  *              return (object);
 547  *      }
 548  *
 549  * Functions that implicitly acquire an object hold (any function that calls
 550  * object_alloc() to supply an object for the caller) need to be carefully noted
 551  * so that the matching object_rele() is not neglected. Otherwise, leaked holds
 552  * prevent all objects hashed to the affected rwlocks from ever being moved.
 553  *
 554  * The pointer to a held object can be hashed to the holding rwlock even after
 555  * the object has been freed. Although it is possible to release the hold
 556  * after freeing the object, you may decide to release the hold implicitly in
 557  * whatever function frees the object, so as to release the hold as soon as
 558  * possible, and for the sake of symmetry with the function that implicitly
 559  * acquires the hold when it allocates the object. Here, object_free() releases
 560  * the hold acquired by object_alloc(). Its implicit object_rele() forms a
 561  * matching pair with object_hold():
 562  *
 563  *      void
 564  *      object_free(object_t *object)
 565  *      {
 566  *              container_t *container;
 567  *
 568  *              ASSERT(object_held(object));
 569  *              container = object->o_container;
 570  *              mutex_enter(&container->c_objects_lock);
 571  *              object->o_container =
 572  *                  (void *)((uintptr_t)object->o_container | 0x1);
 573  *              list_remove(&container->c_objects, object);
 574  *              mutex_exit(&container->c_objects_lock);
 575  *              object_rele(object);
 576  *              kmem_cache_free(object_cache, object);
 577  *      }
 578  *
 579  * Note that object_free() cannot safely accept an object pointer as an argument
 580  * unless the object is already held. Any function that calls object_free()
 581  * needs to be carefully noted since it similarly forms a matching pair with
 582  * object_hold().
 583  *
 584  * To complete the picture, the following callback function implements the
 585  * general solution by moving objects only if they are currently unheld:
 586  *
 587  *      static kmem_cbrc_t
 588  *      object_move(void *buf, void *newbuf, size_t size, void *arg)
 589  *      {
 590  *              object_t *op = buf, *np = newbuf;
 591  *              container_t *container;
 592  *
 593  *              container = op->o_container;
 594  *              if ((uintptr_t)container & 0x3) {
 595  *                      return (KMEM_CBRC_DONT_KNOW);
 596  *              }
 597  *
 598  *              // Ensure that the container structure does not go away.
 599  *              if (container_hold(container) == 0) {
 600  *                      return (KMEM_CBRC_DONT_KNOW);
 601  *              }
 602  *
 603  *              mutex_enter(&container->c_objects_lock);
 604  *              if (container != op->o_container) {
 605  *                      mutex_exit(&container->c_objects_lock);
 606  *                      container_rele(container);
 607  *                      return (KMEM_CBRC_DONT_KNOW);
 608  *              }
 609  *
 610  *              if (rw_tryenter(OBJECT_RWLOCK(op), RW_WRITER) == 0) {
 611  *                      mutex_exit(&container->c_objects_lock);
 612  *                      container_rele(container);
 613  *                      return (KMEM_CBRC_LATER);
 614  *              }
 615  *
 616  *              object_move_impl(op, np); // critical section
 617  *              rw_exit(OBJECT_RWLOCK(op));
 618  *
 619  *              op->o_container = (void *)((uintptr_t)op->o_container | 0x1);
 620  *              list_link_replace(&op->o_link_node, &np->o_link_node);
 621  *              mutex_exit(&container->c_objects_lock);
 622  *              container_rele(container);
 623  *              return (KMEM_CBRC_YES);
 624  *      }
 625  *
 626  * Note that object_move() must invalidate the designated o_container pointer of
 627  * the old object in the same way that object_free() does, since kmem will free
 628  * the object in response to the KMEM_CBRC_YES return value.
 629  *
 630  * The lock order in object_move() differs from object_alloc(), which locks
 631  * OBJECT_RWLOCK first and &container->c_objects_lock second, but as long as the
 632  * callback uses rw_tryenter() (preventing the deadlock described earlier), it's
 633  * not a problem. Holding the lock on the object list in the example above
 634  * through the entire callback not only prevents the object from going away, it
 635  * also allows you to lock the list elsewhere and know that none of its elements
 636  * will move during iteration.
 637  *
 638  * Adding an explicit hold everywhere an object from the cache is used is tricky
 639  * and involves much more change to client code than a cache-specific solution
 640  * that leverages existing state to decide whether or not an object is
 641  * movable. However, this approach has the advantage that no object remains
 642  * immovable for any significant length of time, making it extremely unlikely
 643  * that long-lived allocations can continue holding slabs hostage; and it works
 644  * for any cache.
 645  *
 646  * 3. Consolidator Implementation
 647  *
 648  * Once the client supplies a move function that a) recognizes known objects and
 649  * b) avoids moving objects that are actively in use, the remaining work is up
 650  * to the consolidator to decide which objects to move and when to issue
 651  * callbacks.
 652  *
 653  * The consolidator relies on the fact that a cache's slabs are ordered by
 654  * usage. Each slab has a fixed number of objects. Depending on the slab's
 655  * "color" (the offset of the first object from the beginning of the slab;
 656  * offsets are staggered to mitigate false sharing of cache lines) it is either
 657  * the maximum number of objects per slab determined at cache creation time or
 658  * else the number closest to the maximum that fits within the space remaining
 659  * after the initial offset. A completely allocated slab may contribute some
 660  * internal fragmentation (per-slab overhead) but no external fragmentation, so
 661  * it is of no interest to the consolidator. At the other extreme, slabs whose
 662  * objects have all been freed to the slab are released to the virtual memory
 663  * (VM) subsystem (objects freed to magazines are still allocated as far as the
 664  * slab is concerned). External fragmentation exists when there are slabs
 665  * somewhere between these extremes. A partial slab has at least one but not all
 666  * of its objects allocated. The more partial slabs, and the fewer allocated
 667  * objects on each of them, the higher the fragmentation. Hence the
 668  * consolidator's overall strategy is to reduce the number of partial slabs by
 669  * moving allocated objects from the least allocated slabs to the most allocated
 670  * slabs.
 671  *
 672  * Partial slabs are kept in an AVL tree ordered by usage. Completely allocated
 673  * slabs are kept separately in an unordered list. Since the majority of slabs
 674  * tend to be completely allocated (a typical unfragmented cache may have
 675  * thousands of complete slabs and only a single partial slab), separating
 676  * complete slabs improves the efficiency of partial slab ordering, since the
 677  * complete slabs do not affect the depth or balance of the AVL tree. This
 678  * ordered sequence of partial slabs acts as a "free list" supplying objects for
 679  * allocation requests.
 680  *
 681  * Objects are always allocated from the first partial slab in the free list,
 682  * where the allocation is most likely to eliminate a partial slab (by
 683  * completely allocating it). Conversely, when a single object from a completely
 684  * allocated slab is freed to the slab, that slab is added to the front of the
 685  * free list. Since most free list activity involves highly allocated slabs
 686  * coming and going at the front of the list, slabs tend naturally toward the
 687  * ideal order: highly allocated at the front, sparsely allocated at the back.
 688  * Slabs with few allocated objects are likely to become completely free if they
 689  * keep a safe distance away from the front of the free list. Slab misorders
 690  * interfere with the natural tendency of slabs to become completely free or
 691  * completely allocated. For example, a slab with a single allocated object
 692  * needs only a single free to escape the cache; its natural desire is
 693  * frustrated when it finds itself at the front of the list where a second
 694  * allocation happens just before the free could have released it. Another slab
 695  * with all but one object allocated might have supplied the buffer instead, so
 696  * that both (as opposed to neither) of the slabs would have been taken off the
 697  * free list.
 698  *
 699  * Although slabs tend naturally toward the ideal order, misorders allowed by a
 700  * simple list implementation defeat the consolidator's strategy of merging
 701  * least- and most-allocated slabs. Without an AVL tree to guarantee order, kmem
 702  * needs another way to fix misorders to optimize its callback strategy. One
 703  * approach is to periodically scan a limited number of slabs, advancing a
 704  * marker to hold the current scan position, and to move extreme misorders to
 705  * the front or back of the free list and to the front or back of the current
 706  * scan range. By making consecutive scan ranges overlap by one slab, the least
 707  * allocated slab in the current range can be carried along from the end of one
 708  * scan to the start of the next.
 709  *
 710  * Maintaining partial slabs in an AVL tree relieves kmem of this additional
 711  * task, however. Since most of the cache's activity is in the magazine layer,
 712  * and allocations from the slab layer represent only a startup cost, the
 713  * overhead of maintaining a balanced tree is not a significant concern compared
 714  * to the opportunity of reducing complexity by eliminating the partial slab
 715  * scanner just described. The overhead of an AVL tree is minimized by
 716  * maintaining only partial slabs in the tree and keeping completely allocated
 717  * slabs separately in a list. To avoid increasing the size of the slab
 718  * structure the AVL linkage pointers are reused for the slab's list linkage,
 719  * since the slab will always be either partial or complete, never stored both
 720  * ways at the same time. To further minimize the overhead of the AVL tree the
 721  * compare function that orders partial slabs by usage divides the range of
 722  * allocated object counts into bins such that counts within the same bin are
 723  * considered equal. Binning partial slabs makes it less likely that allocating
 724  * or freeing a single object will change the slab's order, requiring a tree
 725  * reinsertion (an avl_remove() followed by an avl_add(), both potentially
 726  * requiring some rebalancing of the tree). Allocation counts closest to
 727  * completely free and completely allocated are left unbinned (finely sorted) to
 728  * better support the consolidator's strategy of merging slabs at either
 729  * extreme.
 730  *
 731  * 3.1 Assessing Fragmentation and Selecting Candidate Slabs
 732  *
 733  * The consolidator piggybacks on the kmem maintenance thread and is called on
 734  * the same interval as kmem_cache_update(), once per cache every fifteen
 735  * seconds. kmem maintains a running count of unallocated objects in the slab
 736  * layer (cache_bufslab). The consolidator checks whether that number exceeds
 737  * 12.5% (1/8) of the total objects in the cache (cache_buftotal), and whether
 738  * there is a significant number of slabs in the cache (arbitrarily a minimum
 739  * 101 total slabs). Unused objects that have fallen out of the magazine layer's
 740  * working set are included in the assessment, and magazines in the depot are
 741  * reaped if those objects would lift cache_bufslab above the fragmentation
 742  * threshold. Once the consolidator decides that a cache is fragmented, it looks
 743  * for a candidate slab to reclaim, starting at the end of the partial slab free
 744  * list and scanning backwards. At first the consolidator is choosy: only a slab
 745  * with fewer than 12.5% (1/8) of its objects allocated qualifies (or else a
 746  * single allocated object, regardless of percentage). If there is difficulty
 747  * finding a candidate slab, kmem raises the allocation threshold incrementally,
 748  * up to a maximum 87.5% (7/8), so that eventually the consolidator will reduce
 749  * external fragmentation (unused objects on the free list) below 12.5% (1/8),
 750  * even in the worst case of every slab in the cache being almost 7/8 allocated.
 751  * The threshold can also be lowered incrementally when candidate slabs are easy
 752  * to find, and the threshold is reset to the minimum 1/8 as soon as the cache
 753  * is no longer fragmented.
 754  *
 755  * 3.2 Generating Callbacks
 756  *
 757  * Once an eligible slab is chosen, a callback is generated for every allocated
 758  * object on the slab, in the hope that the client will move everything off the
 759  * slab and make it reclaimable. Objects selected as move destinations are
 760  * chosen from slabs at the front of the free list. Assuming slabs in the ideal
 761  * order (most allocated at the front, least allocated at the back) and a
 762  * cooperative client, the consolidator will succeed in removing slabs from both
 763  * ends of the free list, completely allocating on the one hand and completely
 764  * freeing on the other. Objects selected as move destinations are allocated in
 765  * the kmem maintenance thread where move requests are enqueued. A separate
 766  * callback thread removes pending callbacks from the queue and calls the
 767  * client. The separate thread ensures that client code (the move function) does
 768  * not interfere with internal kmem maintenance tasks. A map of pending
 769  * callbacks keyed by object address (the object to be moved) is checked to
 770  * ensure that duplicate callbacks are not generated for the same object.
 771  * Allocating the move destination (the object to move to) prevents subsequent
 772  * callbacks from selecting the same destination as an earlier pending callback.
 773  *
 774  * Move requests can also be generated by kmem_cache_reap() when the system is
 775  * desperate for memory and by kmem_cache_move_notify(), called by the client to
 776  * notify kmem that a move refused earlier with KMEM_CBRC_LATER is now possible.
 777  * The map of pending callbacks is protected by the same lock that protects the
 778  * slab layer.
 779  *
 780  * When the system is desperate for memory, kmem does not bother to determine
 781  * whether or not the cache exceeds the fragmentation threshold, but tries to
 782  * consolidate as many slabs as possible. Normally, the consolidator chews
 783  * slowly, one sparsely allocated slab at a time during each maintenance
 784  * interval that the cache is fragmented. When desperate, the consolidator
 785  * starts at the last partial slab and enqueues callbacks for every allocated
 786  * object on every partial slab, working backwards until it reaches the first
 787  * partial slab. The first partial slab, meanwhile, advances in pace with the
 788  * consolidator as allocations to supply move destinations for the enqueued
 789  * callbacks use up the highly allocated slabs at the front of the free list.
 790  * Ideally, the overgrown free list collapses like an accordion, starting at
 791  * both ends and ending at the center with a single partial slab.
 792  *
 793  * 3.3 Client Responses
 794  *
 795  * When the client returns KMEM_CBRC_NO in response to the move callback, kmem
 796  * marks the slab that supplied the stuck object non-reclaimable and moves it to
 797  * front of the free list. The slab remains marked as long as it remains on the
 798  * free list, and it appears more allocated to the partial slab compare function
 799  * than any unmarked slab, no matter how many of its objects are allocated.
 800  * Since even one immovable object ties up the entire slab, the goal is to
 801  * completely allocate any slab that cannot be completely freed. kmem does not
 802  * bother generating callbacks to move objects from a marked slab unless the
 803  * system is desperate.
 804  *
 805  * When the client responds KMEM_CBRC_LATER, kmem increments a count for the
 806  * slab. If the client responds LATER too many times, kmem disbelieves and
 807  * treats the response as a NO. The count is cleared when the slab is taken off
 808  * the partial slab list or when the client moves one of the slab's objects.
 809  *
 810  * 4. Observability
 811  *
 812  * A kmem cache's external fragmentation is best observed with 'mdb -k' using
 813  * the ::kmem_slabs dcmd. For a complete description of the command, enter
 814  * '::help kmem_slabs' at the mdb prompt.
 815  */
 816 
 817 #include <sys/kmem_impl.h>
 818 #include <sys/vmem_impl.h>
 819 #include <sys/param.h>
 820 #include <sys/sysmacros.h>
 821 #include <sys/vm.h>
 822 #include <sys/proc.h>
 823 #include <sys/tuneable.h>
 824 #include <sys/systm.h>
 825 #include <sys/cmn_err.h>
 826 #include <sys/debug.h>
 827 #include <sys/sdt.h>
 828 #include <sys/mutex.h>
 829 #include <sys/bitmap.h>
 830 #include <sys/atomic.h>
 831 #include <sys/kobj.h>
 832 #include <sys/disp.h>
 833 #include <vm/seg_kmem.h>
 834 #include <sys/log.h>
 835 #include <sys/callb.h>
 836 #include <sys/taskq.h>
 837 #include <sys/modctl.h>
 838 #include <sys/reboot.h>
 839 #include <sys/id32.h>
 840 #include <sys/zone.h>
 841 #include <sys/netstack.h>
 842 #ifdef  DEBUG
 843 #include <sys/random.h>
 844 #endif
 845 
 846 extern void streams_msg_init(void);
 847 extern int segkp_fromheap;
 848 extern void segkp_cache_free(void);
 849 extern int callout_init_done;
 850 
 851 struct kmem_cache_kstat {
 852         kstat_named_t   kmc_buf_size;
 853         kstat_named_t   kmc_align;
 854         kstat_named_t   kmc_chunk_size;
 855         kstat_named_t   kmc_slab_size;
 856         kstat_named_t   kmc_alloc;
 857         kstat_named_t   kmc_alloc_fail;
 858         kstat_named_t   kmc_free;
 859         kstat_named_t   kmc_depot_alloc;
 860         kstat_named_t   kmc_depot_free;
 861         kstat_named_t   kmc_depot_contention;
 862         kstat_named_t   kmc_slab_alloc;
 863         kstat_named_t   kmc_slab_free;
 864         kstat_named_t   kmc_buf_constructed;
 865         kstat_named_t   kmc_buf_avail;
 866         kstat_named_t   kmc_buf_inuse;
 867         kstat_named_t   kmc_buf_total;
 868         kstat_named_t   kmc_buf_max;
 869         kstat_named_t   kmc_slab_create;
 870         kstat_named_t   kmc_slab_destroy;
 871         kstat_named_t   kmc_vmem_source;
 872         kstat_named_t   kmc_hash_size;
 873         kstat_named_t   kmc_hash_lookup_depth;
 874         kstat_named_t   kmc_hash_rescale;
 875         kstat_named_t   kmc_full_magazines;
 876         kstat_named_t   kmc_empty_magazines;
 877         kstat_named_t   kmc_magazine_size;
 878         kstat_named_t   kmc_reap; /* number of kmem_cache_reap() calls */
 879         kstat_named_t   kmc_defrag; /* attempts to defrag all partial slabs */
 880         kstat_named_t   kmc_scan; /* attempts to defrag one partial slab */
 881         kstat_named_t   kmc_move_callbacks; /* sum of yes, no, later, dn, dk */
 882         kstat_named_t   kmc_move_yes;
 883         kstat_named_t   kmc_move_no;
 884         kstat_named_t   kmc_move_later;
 885         kstat_named_t   kmc_move_dont_need;
 886         kstat_named_t   kmc_move_dont_know; /* obj unrecognized by client ... */
 887         kstat_named_t   kmc_move_hunt_found; /* ... but found in mag layer */
 888         kstat_named_t   kmc_move_slabs_freed; /* slabs freed by consolidator */
 889         kstat_named_t   kmc_move_reclaimable; /* buffers, if consolidator ran */
 890 } kmem_cache_kstat = {
 891         { "buf_size",           KSTAT_DATA_UINT64 },
 892         { "align",              KSTAT_DATA_UINT64 },
 893         { "chunk_size",         KSTAT_DATA_UINT64 },
 894         { "slab_size",          KSTAT_DATA_UINT64 },
 895         { "alloc",              KSTAT_DATA_UINT64 },
 896         { "alloc_fail",         KSTAT_DATA_UINT64 },
 897         { "free",               KSTAT_DATA_UINT64 },
 898         { "depot_alloc",        KSTAT_DATA_UINT64 },
 899         { "depot_free",         KSTAT_DATA_UINT64 },
 900         { "depot_contention",   KSTAT_DATA_UINT64 },
 901         { "slab_alloc",         KSTAT_DATA_UINT64 },
 902         { "slab_free",          KSTAT_DATA_UINT64 },
 903         { "buf_constructed",    KSTAT_DATA_UINT64 },
 904         { "buf_avail",          KSTAT_DATA_UINT64 },
 905         { "buf_inuse",          KSTAT_DATA_UINT64 },
 906         { "buf_total",          KSTAT_DATA_UINT64 },
 907         { "buf_max",            KSTAT_DATA_UINT64 },
 908         { "slab_create",        KSTAT_DATA_UINT64 },
 909         { "slab_destroy",       KSTAT_DATA_UINT64 },
 910         { "vmem_source",        KSTAT_DATA_UINT64 },
 911         { "hash_size",          KSTAT_DATA_UINT64 },
 912         { "hash_lookup_depth",  KSTAT_DATA_UINT64 },
 913         { "hash_rescale",       KSTAT_DATA_UINT64 },
 914         { "full_magazines",     KSTAT_DATA_UINT64 },
 915         { "empty_magazines",    KSTAT_DATA_UINT64 },
 916         { "magazine_size",      KSTAT_DATA_UINT64 },
 917         { "reap",               KSTAT_DATA_UINT64 },
 918         { "defrag",             KSTAT_DATA_UINT64 },
 919         { "scan",               KSTAT_DATA_UINT64 },
 920         { "move_callbacks",     KSTAT_DATA_UINT64 },
 921         { "move_yes",           KSTAT_DATA_UINT64 },
 922         { "move_no",            KSTAT_DATA_UINT64 },
 923         { "move_later",         KSTAT_DATA_UINT64 },
 924         { "move_dont_need",     KSTAT_DATA_UINT64 },
 925         { "move_dont_know",     KSTAT_DATA_UINT64 },
 926         { "move_hunt_found",    KSTAT_DATA_UINT64 },
 927         { "move_slabs_freed",   KSTAT_DATA_UINT64 },
 928         { "move_reclaimable",   KSTAT_DATA_UINT64 },
 929 };
 930 
 931 static kmutex_t kmem_cache_kstat_lock;
 932 
 933 /*
 934  * The default set of caches to back kmem_alloc().
 935  * These sizes should be reevaluated periodically.
 936  *
 937  * We want allocations that are multiples of the coherency granularity
 938  * (64 bytes) to be satisfied from a cache which is a multiple of 64
 939  * bytes, so that it will be 64-byte aligned.  For all multiples of 64,
 940  * the next kmem_cache_size greater than or equal to it must be a
 941  * multiple of 64.
 942  *
 943  * We split the table into two sections:  size <= 4k and size > 4k.  This
 944  * saves a lot of space and cache footprint in our cache tables.
 945  */
 946 static const int kmem_alloc_sizes[] = {
 947         1 * 8,
 948         2 * 8,
 949         3 * 8,
 950         4 * 8,          5 * 8,          6 * 8,          7 * 8,
 951         4 * 16,         5 * 16,         6 * 16,         7 * 16,
 952         4 * 32,         5 * 32,         6 * 32,         7 * 32,
 953         4 * 64,         5 * 64,         6 * 64,         7 * 64,
 954         4 * 128,        5 * 128,        6 * 128,        7 * 128,
 955         P2ALIGN(8192 / 7, 64),
 956         P2ALIGN(8192 / 6, 64),
 957         P2ALIGN(8192 / 5, 64),
 958         P2ALIGN(8192 / 4, 64),
 959         P2ALIGN(8192 / 3, 64),
 960         P2ALIGN(8192 / 2, 64),
 961 };
 962 
 963 static const int kmem_big_alloc_sizes[] = {
 964         2 * 4096,       3 * 4096,
 965         2 * 8192,       3 * 8192,
 966         4 * 8192,       5 * 8192,       6 * 8192,       7 * 8192,
 967         8 * 8192,       9 * 8192,       10 * 8192,      11 * 8192,
 968         12 * 8192,      13 * 8192,      14 * 8192,      15 * 8192,
 969         16 * 8192
 970 };
 971 
 972 #define KMEM_MAXBUF             4096
 973 #define KMEM_BIG_MAXBUF_32BIT   32768
 974 #define KMEM_BIG_MAXBUF         131072
 975 
 976 #define KMEM_BIG_MULTIPLE       4096    /* big_alloc_sizes must be a multiple */
 977 #define KMEM_BIG_SHIFT          12      /* lg(KMEM_BIG_MULTIPLE) */
 978 
 979 static kmem_cache_t *kmem_alloc_table[KMEM_MAXBUF >> KMEM_ALIGN_SHIFT];
 980 static kmem_cache_t *kmem_big_alloc_table[KMEM_BIG_MAXBUF >> KMEM_BIG_SHIFT];
 981 
 982 #define KMEM_ALLOC_TABLE_MAX    (KMEM_MAXBUF >> KMEM_ALIGN_SHIFT)
 983 static size_t kmem_big_alloc_table_max = 0;     /* # of filled elements */
 984 
 985 static kmem_magtype_t kmem_magtype[] = {
 986         { 1,    8,      3200,   65536   },
 987         { 3,    16,     256,    32768   },
 988         { 7,    32,     64,     16384   },
 989         { 15,   64,     0,      8192    },
 990         { 31,   64,     0,      4096    },
 991         { 47,   64,     0,      2048    },
 992         { 63,   64,     0,      1024    },
 993         { 95,   64,     0,      512     },
 994         { 143,  64,     0,      0       },
 995 };
 996 
 997 static uint32_t kmem_reaping;
 998 static uint32_t kmem_reaping_idspace;
 999 
1000 /*
1001  * kmem tunables
1002  */
1003 clock_t kmem_reap_interval;     /* cache reaping rate [15 * HZ ticks] */
1004 int kmem_depot_contention = 3;  /* max failed tryenters per real interval */
1005 pgcnt_t kmem_reapahead = 0;     /* start reaping N pages before pageout */
1006 int kmem_panic = 1;             /* whether to panic on error */
1007 int kmem_logging = 1;           /* kmem_log_enter() override */
1008 uint32_t kmem_mtbf = 0;         /* mean time between failures [default: off] */
1009 size_t kmem_transaction_log_size; /* transaction log size [2% of memory] */
1010 size_t kmem_content_log_size;   /* content log size [2% of memory] */
1011 size_t kmem_failure_log_size;   /* failure log [4 pages per CPU] */
1012 size_t kmem_slab_log_size;      /* slab create log [4 pages per CPU] */
1013 size_t kmem_content_maxsave = 256; /* KMF_CONTENTS max bytes to log */
1014 size_t kmem_lite_minsize = 0;   /* minimum buffer size for KMF_LITE */
1015 size_t kmem_lite_maxalign = 1024; /* maximum buffer alignment for KMF_LITE */
1016 int kmem_lite_pcs = 4;          /* number of PCs to store in KMF_LITE mode */
1017 size_t kmem_maxverify;          /* maximum bytes to inspect in debug routines */
1018 size_t kmem_minfirewall;        /* hardware-enforced redzone threshold */
1019 
1020 #ifdef _LP64
1021 size_t  kmem_max_cached = KMEM_BIG_MAXBUF;      /* maximum kmem_alloc cache */
1022 #else
1023 size_t  kmem_max_cached = KMEM_BIG_MAXBUF_32BIT; /* maximum kmem_alloc cache */
1024 #endif
1025 
1026 #ifdef DEBUG
1027 int kmem_flags = KMF_AUDIT | KMF_DEADBEEF | KMF_REDZONE | KMF_CONTENTS;
1028 #else
1029 int kmem_flags = 0;
1030 #endif
1031 int kmem_ready;
1032 
1033 static kmem_cache_t     *kmem_slab_cache;
1034 static kmem_cache_t     *kmem_bufctl_cache;
1035 static kmem_cache_t     *kmem_bufctl_audit_cache;
1036 
1037 static kmutex_t         kmem_cache_lock;        /* inter-cache linkage only */
1038 static list_t           kmem_caches;
1039 
1040 static taskq_t          *kmem_taskq;
1041 static kmutex_t         kmem_flags_lock;
1042 static vmem_t           *kmem_metadata_arena;
1043 static vmem_t           *kmem_msb_arena;        /* arena for metadata caches */
1044 static vmem_t           *kmem_cache_arena;
1045 static vmem_t           *kmem_hash_arena;
1046 static vmem_t           *kmem_log_arena;
1047 static vmem_t           *kmem_oversize_arena;
1048 static vmem_t           *kmem_va_arena;
1049 static vmem_t           *kmem_default_arena;
1050 static vmem_t           *kmem_firewall_va_arena;
1051 static vmem_t           *kmem_firewall_arena;
1052 
1053 /*
1054  * kmem slab consolidator thresholds (tunables)
1055  */
1056 size_t kmem_frag_minslabs = 101;        /* minimum total slabs */
1057 size_t kmem_frag_numer = 1;             /* free buffers (numerator) */
1058 size_t kmem_frag_denom = KMEM_VOID_FRACTION; /* buffers (denominator) */
1059 /*
1060  * Maximum number of slabs from which to move buffers during a single
1061  * maintenance interval while the system is not low on memory.
1062  */
1063 size_t kmem_reclaim_max_slabs = 1;
1064 /*
1065  * Number of slabs to scan backwards from the end of the partial slab list
1066  * when searching for buffers to relocate.
1067  */
1068 size_t kmem_reclaim_scan_range = 12;
1069 
1070 /* consolidator knobs */
1071 boolean_t kmem_move_noreap;
1072 boolean_t kmem_move_blocked;
1073 boolean_t kmem_move_fulltilt;
1074 boolean_t kmem_move_any_partial;
1075 
1076 #ifdef  DEBUG
1077 /*
1078  * kmem consolidator debug tunables:
1079  * Ensure code coverage by occasionally running the consolidator even when the
1080  * caches are not fragmented (they may never be). These intervals are mean time
1081  * in cache maintenance intervals (kmem_cache_update).
1082  */
1083 uint32_t kmem_mtb_move = 60;    /* defrag 1 slab (~15min) */
1084 uint32_t kmem_mtb_reap = 1800;  /* defrag all slabs (~7.5hrs) */
1085 #endif  /* DEBUG */
1086 
1087 static kmem_cache_t     *kmem_defrag_cache;
1088 static kmem_cache_t     *kmem_move_cache;
1089 static taskq_t          *kmem_move_taskq;
1090 
1091 static void kmem_cache_scan(kmem_cache_t *);
1092 static void kmem_cache_defrag(kmem_cache_t *);
1093 static void kmem_slab_prefill(kmem_cache_t *, kmem_slab_t *);
1094 
1095 
1096 kmem_log_header_t       *kmem_transaction_log;
1097 kmem_log_header_t       *kmem_content_log;
1098 kmem_log_header_t       *kmem_failure_log;
1099 kmem_log_header_t       *kmem_slab_log;
1100 
1101 static int              kmem_lite_count; /* # of PCs in kmem_buftag_lite_t */
1102 
1103 #define KMEM_BUFTAG_LITE_ENTER(bt, count, caller)                       \
1104         if ((count) > 0) {                                           \
1105                 pc_t *_s = ((kmem_buftag_lite_t *)(bt))->bt_history; \
1106                 pc_t *_e;                                               \
1107                 /* memmove() the old entries down one notch */          \
1108                 for (_e = &_s[(count) - 1]; _e > _s; _e--)               \
1109                         *_e = *(_e - 1);                                \
1110                 *_s = (uintptr_t)(caller);                              \
1111         }
1112 
1113 #define KMERR_MODIFIED  0       /* buffer modified while on freelist */
1114 #define KMERR_REDZONE   1       /* redzone violation (write past end of buf) */
1115 #define KMERR_DUPFREE   2       /* freed a buffer twice */
1116 #define KMERR_BADADDR   3       /* freed a bad (unallocated) address */
1117 #define KMERR_BADBUFTAG 4       /* buftag corrupted */
1118 #define KMERR_BADBUFCTL 5       /* bufctl corrupted */
1119 #define KMERR_BADCACHE  6       /* freed a buffer to the wrong cache */
1120 #define KMERR_BADSIZE   7       /* alloc size != free size */
1121 #define KMERR_BADBASE   8       /* buffer base address wrong */
1122 
1123 struct {
1124         hrtime_t        kmp_timestamp;  /* timestamp of panic */
1125         int             kmp_error;      /* type of kmem error */
1126         void            *kmp_buffer;    /* buffer that induced panic */
1127         void            *kmp_realbuf;   /* real start address for buffer */
1128         kmem_cache_t    *kmp_cache;     /* buffer's cache according to client */
1129         kmem_cache_t    *kmp_realcache; /* actual cache containing buffer */
1130         kmem_slab_t     *kmp_slab;      /* slab accoring to kmem_findslab() */
1131         kmem_bufctl_t   *kmp_bufctl;    /* bufctl */
1132 } kmem_panic_info;
1133 
1134 
1135 static void
1136 copy_pattern(uint64_t pattern, void *buf_arg, size_t size)
1137 {
1138         uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1139         uint64_t *buf = buf_arg;
1140 
1141         while (buf < bufend)
1142                 *buf++ = pattern;
1143 }
1144 
1145 static void *
1146 verify_pattern(uint64_t pattern, void *buf_arg, size_t size)
1147 {
1148         uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1149         uint64_t *buf;
1150 
1151         for (buf = buf_arg; buf < bufend; buf++)
1152                 if (*buf != pattern)
1153                         return (buf);
1154         return (NULL);
1155 }
1156 
1157 static void *
1158 verify_and_copy_pattern(uint64_t old, uint64_t new, void *buf_arg, size_t size)
1159 {
1160         uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1161         uint64_t *buf;
1162 
1163         for (buf = buf_arg; buf < bufend; buf++) {
1164                 if (*buf != old) {
1165                         copy_pattern(old, buf_arg,
1166                             (char *)buf - (char *)buf_arg);
1167                         return (buf);
1168                 }
1169                 *buf = new;
1170         }
1171 
1172         return (NULL);
1173 }
1174 
1175 static void
1176 kmem_cache_applyall(void (*func)(kmem_cache_t *), taskq_t *tq, int tqflag)
1177 {
1178         kmem_cache_t *cp;
1179 
1180         mutex_enter(&kmem_cache_lock);
1181         for (cp = list_head(&kmem_caches); cp != NULL;
1182             cp = list_next(&kmem_caches, cp))
1183                 if (tq != NULL)
1184                         (void) taskq_dispatch(tq, (task_func_t *)func, cp,
1185                             tqflag);
1186                 else
1187                         func(cp);
1188         mutex_exit(&kmem_cache_lock);
1189 }
1190 
1191 static void
1192 kmem_cache_applyall_id(void (*func)(kmem_cache_t *), taskq_t *tq, int tqflag)
1193 {
1194         kmem_cache_t *cp;
1195 
1196         mutex_enter(&kmem_cache_lock);
1197         for (cp = list_head(&kmem_caches); cp != NULL;
1198             cp = list_next(&kmem_caches, cp)) {
1199                 if (!(cp->cache_cflags & KMC_IDENTIFIER))
1200                         continue;
1201                 if (tq != NULL)
1202                         (void) taskq_dispatch(tq, (task_func_t *)func, cp,
1203                             tqflag);
1204                 else
1205                         func(cp);
1206         }
1207         mutex_exit(&kmem_cache_lock);
1208 }
1209 
1210 /*
1211  * Debugging support.  Given a buffer address, find its slab.
1212  */
1213 static kmem_slab_t *
1214 kmem_findslab(kmem_cache_t *cp, void *buf)
1215 {
1216         kmem_slab_t *sp;
1217 
1218         mutex_enter(&cp->cache_lock);
1219         for (sp = list_head(&cp->cache_complete_slabs); sp != NULL;
1220             sp = list_next(&cp->cache_complete_slabs, sp)) {
1221                 if (KMEM_SLAB_MEMBER(sp, buf)) {
1222                         mutex_exit(&cp->cache_lock);
1223                         return (sp);
1224                 }
1225         }
1226         for (sp = avl_first(&cp->cache_partial_slabs); sp != NULL;
1227             sp = AVL_NEXT(&cp->cache_partial_slabs, sp)) {
1228                 if (KMEM_SLAB_MEMBER(sp, buf)) {
1229                         mutex_exit(&cp->cache_lock);
1230                         return (sp);
1231                 }
1232         }
1233         mutex_exit(&cp->cache_lock);
1234 
1235         return (NULL);
1236 }
1237 
1238 static void
1239 kmem_error(int error, kmem_cache_t *cparg, void *bufarg)
1240 {
1241         kmem_buftag_t *btp = NULL;
1242         kmem_bufctl_t *bcp = NULL;
1243         kmem_cache_t *cp = cparg;
1244         kmem_slab_t *sp;
1245         uint64_t *off;
1246         void *buf = bufarg;
1247 
1248         kmem_logging = 0;       /* stop logging when a bad thing happens */
1249 
1250         kmem_panic_info.kmp_timestamp = gethrtime();
1251 
1252         sp = kmem_findslab(cp, buf);
1253         if (sp == NULL) {
1254                 for (cp = list_tail(&kmem_caches); cp != NULL;
1255                     cp = list_prev(&kmem_caches, cp)) {
1256                         if ((sp = kmem_findslab(cp, buf)) != NULL)
1257                                 break;
1258                 }
1259         }
1260 
1261         if (sp == NULL) {
1262                 cp = NULL;
1263                 error = KMERR_BADADDR;
1264         } else {
1265                 if (cp != cparg)
1266                         error = KMERR_BADCACHE;
1267                 else
1268                         buf = (char *)bufarg - ((uintptr_t)bufarg -
1269                             (uintptr_t)sp->slab_base) % cp->cache_chunksize;
1270                 if (buf != bufarg)
1271                         error = KMERR_BADBASE;
1272                 if (cp->cache_flags & KMF_BUFTAG)
1273                         btp = KMEM_BUFTAG(cp, buf);
1274                 if (cp->cache_flags & KMF_HASH) {
1275                         mutex_enter(&cp->cache_lock);
1276                         for (bcp = *KMEM_HASH(cp, buf); bcp; bcp = bcp->bc_next)
1277                                 if (bcp->bc_addr == buf)
1278                                         break;
1279                         mutex_exit(&cp->cache_lock);
1280                         if (bcp == NULL && btp != NULL)
1281                                 bcp = btp->bt_bufctl;
1282                         if (kmem_findslab(cp->cache_bufctl_cache, bcp) ==
1283                             NULL || P2PHASE((uintptr_t)bcp, KMEM_ALIGN) ||
1284                             bcp->bc_addr != buf) {
1285                                 error = KMERR_BADBUFCTL;
1286                                 bcp = NULL;
1287                         }
1288                 }
1289         }
1290 
1291         kmem_panic_info.kmp_error = error;
1292         kmem_panic_info.kmp_buffer = bufarg;
1293         kmem_panic_info.kmp_realbuf = buf;
1294         kmem_panic_info.kmp_cache = cparg;
1295         kmem_panic_info.kmp_realcache = cp;
1296         kmem_panic_info.kmp_slab = sp;
1297         kmem_panic_info.kmp_bufctl = bcp;
1298 
1299         printf("kernel memory allocator: ");
1300 
1301         switch (error) {
1302 
1303         case KMERR_MODIFIED:
1304                 printf("buffer modified after being freed\n");
1305                 off = verify_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1306                 if (off == NULL)        /* shouldn't happen */
1307                         off = buf;
1308                 printf("modification occurred at offset 0x%lx "
1309                     "(0x%llx replaced by 0x%llx)\n",
1310                     (uintptr_t)off - (uintptr_t)buf,
1311                     (longlong_t)KMEM_FREE_PATTERN, (longlong_t)*off);
1312                 break;
1313 
1314         case KMERR_REDZONE:
1315                 printf("redzone violation: write past end of buffer\n");
1316                 break;
1317 
1318         case KMERR_BADADDR:
1319                 printf("invalid free: buffer not in cache\n");
1320                 break;
1321 
1322         case KMERR_DUPFREE:
1323                 printf("duplicate free: buffer freed twice\n");
1324                 break;
1325 
1326         case KMERR_BADBUFTAG:
1327                 printf("boundary tag corrupted\n");
1328                 printf("bcp ^ bxstat = %lx, should be %lx\n",
1329                     (intptr_t)btp->bt_bufctl ^ btp->bt_bxstat,
1330                     KMEM_BUFTAG_FREE);
1331                 break;
1332 
1333         case KMERR_BADBUFCTL:
1334                 printf("bufctl corrupted\n");
1335                 break;
1336 
1337         case KMERR_BADCACHE:
1338                 printf("buffer freed to wrong cache\n");
1339                 printf("buffer was allocated from %s,\n", cp->cache_name);
1340                 printf("caller attempting free to %s.\n", cparg->cache_name);
1341                 break;
1342 
1343         case KMERR_BADSIZE:
1344                 printf("bad free: free size (%u) != alloc size (%u)\n",
1345                     KMEM_SIZE_DECODE(((uint32_t *)btp)[0]),
1346                     KMEM_SIZE_DECODE(((uint32_t *)btp)[1]));
1347                 break;
1348 
1349         case KMERR_BADBASE:
1350                 printf("bad free: free address (%p) != alloc address (%p)\n",
1351                     bufarg, buf);
1352                 break;
1353         }
1354 
1355         printf("buffer=%p  bufctl=%p  cache: %s\n",
1356             bufarg, (void *)bcp, cparg->cache_name);
1357 
1358         if (bcp != NULL && (cp->cache_flags & KMF_AUDIT) &&
1359             error != KMERR_BADBUFCTL) {
1360                 int d;
1361                 timestruc_t ts;
1362                 kmem_bufctl_audit_t *bcap = (kmem_bufctl_audit_t *)bcp;
1363 
1364                 hrt2ts(kmem_panic_info.kmp_timestamp - bcap->bc_timestamp, &ts);
1365                 printf("previous transaction on buffer %p:\n", buf);
1366                 printf("thread=%p  time=T-%ld.%09ld  slab=%p  cache: %s\n",
1367                     (void *)bcap->bc_thread, ts.tv_sec, ts.tv_nsec,
1368                     (void *)sp, cp->cache_name);
1369                 for (d = 0; d < MIN(bcap->bc_depth, KMEM_STACK_DEPTH); d++) {
1370                         ulong_t off;
1371                         char *sym = kobj_getsymname(bcap->bc_stack[d], &off);
1372                         printf("%s+%lx\n", sym ? sym : "?", off);
1373                 }
1374         }
1375         if (kmem_panic > 0)
1376                 panic("kernel heap corruption detected");
1377         if (kmem_panic == 0)
1378                 debug_enter(NULL);
1379         kmem_logging = 1;       /* resume logging */
1380 }
1381 
1382 static kmem_log_header_t *
1383 kmem_log_init(size_t logsize)
1384 {
1385         kmem_log_header_t *lhp;
1386         int nchunks = 4 * max_ncpus;
1387         size_t lhsize = (size_t)&((kmem_log_header_t *)0)->lh_cpu[max_ncpus];
1388         int i;
1389 
1390         /*
1391          * Make sure that lhp->lh_cpu[] is nicely aligned
1392          * to prevent false sharing of cache lines.
1393          */
1394         lhsize = P2ROUNDUP(lhsize, KMEM_ALIGN);
1395         lhp = vmem_xalloc(kmem_log_arena, lhsize, 64, P2NPHASE(lhsize, 64), 0,
1396             NULL, NULL, VM_SLEEP);
1397         bzero(lhp, lhsize);
1398 
1399         mutex_init(&lhp->lh_lock, NULL, MUTEX_DEFAULT, NULL);
1400         lhp->lh_nchunks = nchunks;
1401         lhp->lh_chunksize = P2ROUNDUP(logsize / nchunks + 1, PAGESIZE);
1402         lhp->lh_base = vmem_alloc(kmem_log_arena,
1403             lhp->lh_chunksize * nchunks, VM_SLEEP);
1404         lhp->lh_free = vmem_alloc(kmem_log_arena,
1405             nchunks * sizeof (int), VM_SLEEP);
1406         bzero(lhp->lh_base, lhp->lh_chunksize * nchunks);
1407 
1408         for (i = 0; i < max_ncpus; i++) {
1409                 kmem_cpu_log_header_t *clhp = &lhp->lh_cpu[i];
1410                 mutex_init(&clhp->clh_lock, NULL, MUTEX_DEFAULT, NULL);
1411                 clhp->clh_chunk = i;
1412         }
1413 
1414         for (i = max_ncpus; i < nchunks; i++)
1415                 lhp->lh_free[i] = i;
1416 
1417         lhp->lh_head = max_ncpus;
1418         lhp->lh_tail = 0;
1419 
1420         return (lhp);
1421 }
1422 
1423 static void *
1424 kmem_log_enter(kmem_log_header_t *lhp, void *data, size_t size)
1425 {
1426         void *logspace;
1427         kmem_cpu_log_header_t *clhp = &lhp->lh_cpu[CPU->cpu_seqid];
1428 
1429         if (lhp == NULL || kmem_logging == 0 || panicstr)
1430                 return (NULL);
1431 
1432         mutex_enter(&clhp->clh_lock);
1433         clhp->clh_hits++;
1434         if (size > clhp->clh_avail) {
1435                 mutex_enter(&lhp->lh_lock);
1436                 lhp->lh_hits++;
1437                 lhp->lh_free[lhp->lh_tail] = clhp->clh_chunk;
1438                 lhp->lh_tail = (lhp->lh_tail + 1) % lhp->lh_nchunks;
1439                 clhp->clh_chunk = lhp->lh_free[lhp->lh_head];
1440                 lhp->lh_head = (lhp->lh_head + 1) % lhp->lh_nchunks;
1441                 clhp->clh_current = lhp->lh_base +
1442                     clhp->clh_chunk * lhp->lh_chunksize;
1443                 clhp->clh_avail = lhp->lh_chunksize;
1444                 if (size > lhp->lh_chunksize)
1445                         size = lhp->lh_chunksize;
1446                 mutex_exit(&lhp->lh_lock);
1447         }
1448         logspace = clhp->clh_current;
1449         clhp->clh_current += size;
1450         clhp->clh_avail -= size;
1451         bcopy(data, logspace, size);
1452         mutex_exit(&clhp->clh_lock);
1453         return (logspace);
1454 }
1455 
1456 #define KMEM_AUDIT(lp, cp, bcp)                                         \
1457 {                                                                       \
1458         kmem_bufctl_audit_t *_bcp = (kmem_bufctl_audit_t *)(bcp);       \
1459         _bcp->bc_timestamp = gethrtime();                            \
1460         _bcp->bc_thread = curthread;                                 \
1461         _bcp->bc_depth = getpcstack(_bcp->bc_stack, KMEM_STACK_DEPTH);    \
1462         _bcp->bc_lastlog = kmem_log_enter((lp), _bcp, sizeof (*_bcp));       \
1463 }
1464 
1465 static void
1466 kmem_log_event(kmem_log_header_t *lp, kmem_cache_t *cp,
1467     kmem_slab_t *sp, void *addr)
1468 {
1469         kmem_bufctl_audit_t bca;
1470 
1471         bzero(&bca, sizeof (kmem_bufctl_audit_t));
1472         bca.bc_addr = addr;
1473         bca.bc_slab = sp;
1474         bca.bc_cache = cp;
1475         KMEM_AUDIT(lp, cp, &bca);
1476 }
1477 
1478 /*
1479  * Create a new slab for cache cp.
1480  */
1481 static kmem_slab_t *
1482 kmem_slab_create(kmem_cache_t *cp, int kmflag)
1483 {
1484         size_t slabsize = cp->cache_slabsize;
1485         size_t chunksize = cp->cache_chunksize;
1486         int cache_flags = cp->cache_flags;
1487         size_t color, chunks;
1488         char *buf, *slab;
1489         kmem_slab_t *sp;
1490         kmem_bufctl_t *bcp;
1491         vmem_t *vmp = cp->cache_arena;
1492 
1493         ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
1494 
1495         color = cp->cache_color + cp->cache_align;
1496         if (color > cp->cache_maxcolor)
1497                 color = cp->cache_mincolor;
1498         cp->cache_color = color;
1499 
1500         slab = vmem_alloc(vmp, slabsize, kmflag & KM_VMFLAGS);
1501 
1502         if (slab == NULL)
1503                 goto vmem_alloc_failure;
1504 
1505         ASSERT(P2PHASE((uintptr_t)slab, vmp->vm_quantum) == 0);
1506 
1507         /*
1508          * Reverify what was already checked in kmem_cache_set_move(), since the
1509          * consolidator depends (for correctness) on slabs being initialized
1510          * with the 0xbaddcafe memory pattern (setting a low order bit usable by
1511          * clients to distinguish uninitialized memory from known objects).
1512          */
1513         ASSERT((cp->cache_move == NULL) || !(cp->cache_cflags & KMC_NOTOUCH));
1514         if (!(cp->cache_cflags & KMC_NOTOUCH))
1515                 copy_pattern(KMEM_UNINITIALIZED_PATTERN, slab, slabsize);
1516 
1517         if (cache_flags & KMF_HASH) {
1518                 if ((sp = kmem_cache_alloc(kmem_slab_cache, kmflag)) == NULL)
1519                         goto slab_alloc_failure;
1520                 chunks = (slabsize - color) / chunksize;
1521         } else {
1522                 sp = KMEM_SLAB(cp, slab);
1523                 chunks = (slabsize - sizeof (kmem_slab_t) - color) / chunksize;
1524         }
1525 
1526         sp->slab_cache       = cp;
1527         sp->slab_head        = NULL;
1528         sp->slab_refcnt      = 0;
1529         sp->slab_base        = buf = slab + color;
1530         sp->slab_chunks      = chunks;
1531         sp->slab_stuck_offset = (uint32_t)-1;
1532         sp->slab_later_count = 0;
1533         sp->slab_flags = 0;
1534 
1535         ASSERT(chunks > 0);
1536         while (chunks-- != 0) {
1537                 if (cache_flags & KMF_HASH) {
1538                         bcp = kmem_cache_alloc(cp->cache_bufctl_cache, kmflag);
1539                         if (bcp == NULL)
1540                                 goto bufctl_alloc_failure;
1541                         if (cache_flags & KMF_AUDIT) {
1542                                 kmem_bufctl_audit_t *bcap =
1543                                     (kmem_bufctl_audit_t *)bcp;
1544                                 bzero(bcap, sizeof (kmem_bufctl_audit_t));
1545                                 bcap->bc_cache = cp;
1546                         }
1547                         bcp->bc_addr = buf;
1548                         bcp->bc_slab = sp;
1549                 } else {
1550                         bcp = KMEM_BUFCTL(cp, buf);
1551                 }
1552                 if (cache_flags & KMF_BUFTAG) {
1553                         kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1554                         btp->bt_redzone = KMEM_REDZONE_PATTERN;
1555                         btp->bt_bufctl = bcp;
1556                         btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
1557                         if (cache_flags & KMF_DEADBEEF) {
1558                                 copy_pattern(KMEM_FREE_PATTERN, buf,
1559                                     cp->cache_verify);
1560                         }
1561                 }
1562                 bcp->bc_next = sp->slab_head;
1563                 sp->slab_head = bcp;
1564                 buf += chunksize;
1565         }
1566 
1567         kmem_log_event(kmem_slab_log, cp, sp, slab);
1568 
1569         return (sp);
1570 
1571 bufctl_alloc_failure:
1572 
1573         while ((bcp = sp->slab_head) != NULL) {
1574                 sp->slab_head = bcp->bc_next;
1575                 kmem_cache_free(cp->cache_bufctl_cache, bcp);
1576         }
1577         kmem_cache_free(kmem_slab_cache, sp);
1578 
1579 slab_alloc_failure:
1580 
1581         vmem_free(vmp, slab, slabsize);
1582 
1583 vmem_alloc_failure:
1584 
1585         kmem_log_event(kmem_failure_log, cp, NULL, NULL);
1586         atomic_inc_64(&cp->cache_alloc_fail);
1587 
1588         return (NULL);
1589 }
1590 
1591 /*
1592  * Destroy a slab.
1593  */
1594 static void
1595 kmem_slab_destroy(kmem_cache_t *cp, kmem_slab_t *sp)
1596 {
1597         vmem_t *vmp = cp->cache_arena;
1598         void *slab = (void *)P2ALIGN((uintptr_t)sp->slab_base, vmp->vm_quantum);
1599 
1600         ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
1601         ASSERT(sp->slab_refcnt == 0);
1602 
1603         if (cp->cache_flags & KMF_HASH) {
1604                 kmem_bufctl_t *bcp;
1605                 while ((bcp = sp->slab_head) != NULL) {
1606                         sp->slab_head = bcp->bc_next;
1607                         kmem_cache_free(cp->cache_bufctl_cache, bcp);
1608                 }
1609                 kmem_cache_free(kmem_slab_cache, sp);
1610         }
1611         vmem_free(vmp, slab, cp->cache_slabsize);
1612 }
1613 
1614 static void *
1615 kmem_slab_alloc_impl(kmem_cache_t *cp, kmem_slab_t *sp, boolean_t prefill)
1616 {
1617         kmem_bufctl_t *bcp, **hash_bucket;
1618         void *buf;
1619         boolean_t new_slab = (sp->slab_refcnt == 0);
1620 
1621         ASSERT(MUTEX_HELD(&cp->cache_lock));
1622         /*
1623          * kmem_slab_alloc() drops cache_lock when it creates a new slab, so we
1624          * can't ASSERT(avl_is_empty(&cp->cache_partial_slabs)) here when the
1625          * slab is newly created.
1626          */
1627         ASSERT(new_slab || (KMEM_SLAB_IS_PARTIAL(sp) &&
1628             (sp == avl_first(&cp->cache_partial_slabs))));
1629         ASSERT(sp->slab_cache == cp);
1630 
1631         cp->cache_slab_alloc++;
1632         cp->cache_bufslab--;
1633         sp->slab_refcnt++;
1634 
1635         bcp = sp->slab_head;
1636         sp->slab_head = bcp->bc_next;
1637 
1638         if (cp->cache_flags & KMF_HASH) {
1639                 /*
1640                  * Add buffer to allocated-address hash table.
1641                  */
1642                 buf = bcp->bc_addr;
1643                 hash_bucket = KMEM_HASH(cp, buf);
1644                 bcp->bc_next = *hash_bucket;
1645                 *hash_bucket = bcp;
1646                 if ((cp->cache_flags & (KMF_AUDIT | KMF_BUFTAG)) == KMF_AUDIT) {
1647                         KMEM_AUDIT(kmem_transaction_log, cp, bcp);
1648                 }
1649         } else {
1650                 buf = KMEM_BUF(cp, bcp);
1651         }
1652 
1653         ASSERT(KMEM_SLAB_MEMBER(sp, buf));
1654 
1655         if (sp->slab_head == NULL) {
1656                 ASSERT(KMEM_SLAB_IS_ALL_USED(sp));
1657                 if (new_slab) {
1658                         ASSERT(sp->slab_chunks == 1);
1659                 } else {
1660                         ASSERT(sp->slab_chunks > 1); /* the slab was partial */
1661                         avl_remove(&cp->cache_partial_slabs, sp);
1662                         sp->slab_later_count = 0; /* clear history */
1663                         sp->slab_flags &= ~KMEM_SLAB_NOMOVE;
1664                         sp->slab_stuck_offset = (uint32_t)-1;
1665                 }
1666                 list_insert_head(&cp->cache_complete_slabs, sp);
1667                 cp->cache_complete_slab_count++;
1668                 return (buf);
1669         }
1670 
1671         ASSERT(KMEM_SLAB_IS_PARTIAL(sp));
1672         /*
1673          * Peek to see if the magazine layer is enabled before
1674          * we prefill.  We're not holding the cpu cache lock,
1675          * so the peek could be wrong, but there's no harm in it.
1676          */
1677         if (new_slab && prefill && (cp->cache_flags & KMF_PREFILL) &&
1678             (KMEM_CPU_CACHE(cp)->cc_magsize != 0))  {
1679                 kmem_slab_prefill(cp, sp);
1680                 return (buf);
1681         }
1682 
1683         if (new_slab) {
1684                 avl_add(&cp->cache_partial_slabs, sp);
1685                 return (buf);
1686         }
1687 
1688         /*
1689          * The slab is now more allocated than it was, so the
1690          * order remains unchanged.
1691          */
1692         ASSERT(!avl_update(&cp->cache_partial_slabs, sp));
1693         return (buf);
1694 }
1695 
1696 /*
1697  * Allocate a raw (unconstructed) buffer from cp's slab layer.
1698  */
1699 static void *
1700 kmem_slab_alloc(kmem_cache_t *cp, int kmflag)
1701 {
1702         kmem_slab_t *sp;
1703         void *buf;
1704         boolean_t test_destructor;
1705 
1706         mutex_enter(&cp->cache_lock);
1707         test_destructor = (cp->cache_slab_alloc == 0);
1708         sp = avl_first(&cp->cache_partial_slabs);
1709         if (sp == NULL) {
1710                 ASSERT(cp->cache_bufslab == 0);
1711 
1712                 /*
1713                  * The freelist is empty.  Create a new slab.
1714                  */
1715                 mutex_exit(&cp->cache_lock);
1716                 if ((sp = kmem_slab_create(cp, kmflag)) == NULL) {
1717                         return (NULL);
1718                 }
1719                 mutex_enter(&cp->cache_lock);
1720                 cp->cache_slab_create++;
1721                 if ((cp->cache_buftotal += sp->slab_chunks) > cp->cache_bufmax)
1722                         cp->cache_bufmax = cp->cache_buftotal;
1723                 cp->cache_bufslab += sp->slab_chunks;
1724         }
1725 
1726         buf = kmem_slab_alloc_impl(cp, sp, B_TRUE);
1727         ASSERT((cp->cache_slab_create - cp->cache_slab_destroy) ==
1728             (cp->cache_complete_slab_count +
1729             avl_numnodes(&cp->cache_partial_slabs) +
1730             (cp->cache_defrag == NULL ? 0 : cp->cache_defrag->kmd_deadcount)));
1731         mutex_exit(&cp->cache_lock);
1732 
1733         if (test_destructor && cp->cache_destructor != NULL) {
1734                 /*
1735                  * On the first kmem_slab_alloc(), assert that it is valid to
1736                  * call the destructor on a newly constructed object without any
1737                  * client involvement.
1738                  */
1739                 if ((cp->cache_constructor == NULL) ||
1740                     cp->cache_constructor(buf, cp->cache_private,
1741                     kmflag) == 0) {
1742                         cp->cache_destructor(buf, cp->cache_private);
1743                 }
1744                 copy_pattern(KMEM_UNINITIALIZED_PATTERN, buf,
1745                     cp->cache_bufsize);
1746                 if (cp->cache_flags & KMF_DEADBEEF) {
1747                         copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1748                 }
1749         }
1750 
1751         return (buf);
1752 }
1753 
1754 static void kmem_slab_move_yes(kmem_cache_t *, kmem_slab_t *, void *);
1755 
1756 /*
1757  * Free a raw (unconstructed) buffer to cp's slab layer.
1758  */
1759 static void
1760 kmem_slab_free(kmem_cache_t *cp, void *buf)
1761 {
1762         kmem_slab_t *sp;
1763         kmem_bufctl_t *bcp, **prev_bcpp;
1764 
1765         ASSERT(buf != NULL);
1766 
1767         mutex_enter(&cp->cache_lock);
1768         cp->cache_slab_free++;
1769 
1770         if (cp->cache_flags & KMF_HASH) {
1771                 /*
1772                  * Look up buffer in allocated-address hash table.
1773                  */
1774                 prev_bcpp = KMEM_HASH(cp, buf);
1775                 while ((bcp = *prev_bcpp) != NULL) {
1776                         if (bcp->bc_addr == buf) {
1777                                 *prev_bcpp = bcp->bc_next;
1778                                 sp = bcp->bc_slab;
1779                                 break;
1780                         }
1781                         cp->cache_lookup_depth++;
1782                         prev_bcpp = &bcp->bc_next;
1783                 }
1784         } else {
1785                 bcp = KMEM_BUFCTL(cp, buf);
1786                 sp = KMEM_SLAB(cp, buf);
1787         }
1788 
1789         if (bcp == NULL || sp->slab_cache != cp || !KMEM_SLAB_MEMBER(sp, buf)) {
1790                 mutex_exit(&cp->cache_lock);
1791                 kmem_error(KMERR_BADADDR, cp, buf);
1792                 return;
1793         }
1794 
1795         if (KMEM_SLAB_OFFSET(sp, buf) == sp->slab_stuck_offset) {
1796                 /*
1797                  * If this is the buffer that prevented the consolidator from
1798                  * clearing the slab, we can reset the slab flags now that the
1799                  * buffer is freed. (It makes sense to do this in
1800                  * kmem_cache_free(), where the client gives up ownership of the
1801                  * buffer, but on the hot path the test is too expensive.)
1802                  */
1803                 kmem_slab_move_yes(cp, sp, buf);
1804         }
1805 
1806         if ((cp->cache_flags & (KMF_AUDIT | KMF_BUFTAG)) == KMF_AUDIT) {
1807                 if (cp->cache_flags & KMF_CONTENTS)
1808                         ((kmem_bufctl_audit_t *)bcp)->bc_contents =
1809                             kmem_log_enter(kmem_content_log, buf,
1810                             cp->cache_contents);
1811                 KMEM_AUDIT(kmem_transaction_log, cp, bcp);
1812         }
1813 
1814         bcp->bc_next = sp->slab_head;
1815         sp->slab_head = bcp;
1816 
1817         cp->cache_bufslab++;
1818         ASSERT(sp->slab_refcnt >= 1);
1819 
1820         if (--sp->slab_refcnt == 0) {
1821                 /*
1822                  * There are no outstanding allocations from this slab,
1823                  * so we can reclaim the memory.
1824                  */
1825                 if (sp->slab_chunks == 1) {
1826                         list_remove(&cp->cache_complete_slabs, sp);
1827                         cp->cache_complete_slab_count--;
1828                 } else {
1829                         avl_remove(&cp->cache_partial_slabs, sp);
1830                 }
1831 
1832                 cp->cache_buftotal -= sp->slab_chunks;
1833                 cp->cache_bufslab -= sp->slab_chunks;
1834                 /*
1835                  * Defer releasing the slab to the virtual memory subsystem
1836                  * while there is a pending move callback, since we guarantee
1837                  * that buffers passed to the move callback have only been
1838                  * touched by kmem or by the client itself. Since the memory
1839                  * patterns baddcafe (uninitialized) and deadbeef (freed) both
1840                  * set at least one of the two lowest order bits, the client can
1841                  * test those bits in the move callback to determine whether or
1842                  * not it knows about the buffer (assuming that the client also
1843                  * sets one of those low order bits whenever it frees a buffer).
1844                  */
1845                 if (cp->cache_defrag == NULL ||
1846                     (avl_is_empty(&cp->cache_defrag->kmd_moves_pending) &&
1847                     !(sp->slab_flags & KMEM_SLAB_MOVE_PENDING))) {
1848                         cp->cache_slab_destroy++;
1849                         mutex_exit(&cp->cache_lock);
1850                         kmem_slab_destroy(cp, sp);
1851                 } else {
1852                         list_t *deadlist = &cp->cache_defrag->kmd_deadlist;
1853                         /*
1854                          * Slabs are inserted at both ends of the deadlist to
1855                          * distinguish between slabs freed while move callbacks
1856                          * are pending (list head) and a slab freed while the
1857                          * lock is dropped in kmem_move_buffers() (list tail) so
1858                          * that in both cases slab_destroy() is called from the
1859                          * right context.
1860                          */
1861                         if (sp->slab_flags & KMEM_SLAB_MOVE_PENDING) {
1862                                 list_insert_tail(deadlist, sp);
1863                         } else {
1864                                 list_insert_head(deadlist, sp);
1865                         }
1866                         cp->cache_defrag->kmd_deadcount++;
1867                         mutex_exit(&cp->cache_lock);
1868                 }
1869                 return;
1870         }
1871 
1872         if (bcp->bc_next == NULL) {
1873                 /* Transition the slab from completely allocated to partial. */
1874                 ASSERT(sp->slab_refcnt == (sp->slab_chunks - 1));
1875                 ASSERT(sp->slab_chunks > 1);
1876                 list_remove(&cp->cache_complete_slabs, sp);
1877                 cp->cache_complete_slab_count--;
1878                 avl_add(&cp->cache_partial_slabs, sp);
1879         } else {
1880                 (void) avl_update_gt(&cp->cache_partial_slabs, sp);
1881         }
1882 
1883         ASSERT((cp->cache_slab_create - cp->cache_slab_destroy) ==
1884             (cp->cache_complete_slab_count +
1885             avl_numnodes(&cp->cache_partial_slabs) +
1886             (cp->cache_defrag == NULL ? 0 : cp->cache_defrag->kmd_deadcount)));
1887         mutex_exit(&cp->cache_lock);
1888 }
1889 
1890 /*
1891  * Return -1 if kmem_error, 1 if constructor fails, 0 if successful.
1892  */
1893 static int
1894 kmem_cache_alloc_debug(kmem_cache_t *cp, void *buf, int kmflag, int construct,
1895     caddr_t caller)
1896 {
1897         kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1898         kmem_bufctl_audit_t *bcp = (kmem_bufctl_audit_t *)btp->bt_bufctl;
1899         uint32_t mtbf;
1900 
1901         if (btp->bt_bxstat != ((intptr_t)bcp ^ KMEM_BUFTAG_FREE)) {
1902                 kmem_error(KMERR_BADBUFTAG, cp, buf);
1903                 return (-1);
1904         }
1905 
1906         btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_ALLOC;
1907 
1908         if ((cp->cache_flags & KMF_HASH) && bcp->bc_addr != buf) {
1909                 kmem_error(KMERR_BADBUFCTL, cp, buf);
1910                 return (-1);
1911         }
1912 
1913         if (cp->cache_flags & KMF_DEADBEEF) {
1914                 if (!construct && (cp->cache_flags & KMF_LITE)) {
1915                         if (*(uint64_t *)buf != KMEM_FREE_PATTERN) {
1916                                 kmem_error(KMERR_MODIFIED, cp, buf);
1917                                 return (-1);
1918                         }
1919                         if (cp->cache_constructor != NULL)
1920                                 *(uint64_t *)buf = btp->bt_redzone;
1921                         else
1922                                 *(uint64_t *)buf = KMEM_UNINITIALIZED_PATTERN;
1923                 } else {
1924                         construct = 1;
1925                         if (verify_and_copy_pattern(KMEM_FREE_PATTERN,
1926                             KMEM_UNINITIALIZED_PATTERN, buf,
1927                             cp->cache_verify)) {
1928                                 kmem_error(KMERR_MODIFIED, cp, buf);
1929                                 return (-1);
1930                         }
1931                 }
1932         }
1933         btp->bt_redzone = KMEM_REDZONE_PATTERN;
1934 
1935         if ((mtbf = kmem_mtbf | cp->cache_mtbf) != 0 &&
1936             gethrtime() % mtbf == 0 &&
1937             (kmflag & (KM_NOSLEEP | KM_PANIC)) == KM_NOSLEEP) {
1938                 kmem_log_event(kmem_failure_log, cp, NULL, NULL);
1939                 if (!construct && cp->cache_destructor != NULL)
1940                         cp->cache_destructor(buf, cp->cache_private);
1941         } else {
1942                 mtbf = 0;
1943         }
1944 
1945         if (mtbf || (construct && cp->cache_constructor != NULL &&
1946             cp->cache_constructor(buf, cp->cache_private, kmflag) != 0)) {
1947                 atomic_inc_64(&cp->cache_alloc_fail);
1948                 btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
1949                 if (cp->cache_flags & KMF_DEADBEEF)
1950                         copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1951                 kmem_slab_free(cp, buf);
1952                 return (1);
1953         }
1954 
1955         if (cp->cache_flags & KMF_AUDIT) {
1956                 KMEM_AUDIT(kmem_transaction_log, cp, bcp);
1957         }
1958 
1959         if ((cp->cache_flags & KMF_LITE) &&
1960             !(cp->cache_cflags & KMC_KMEM_ALLOC)) {
1961                 KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller);
1962         }
1963 
1964         return (0);
1965 }
1966 
1967 static int
1968 kmem_cache_free_debug(kmem_cache_t *cp, void *buf, caddr_t caller)
1969 {
1970         kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1971         kmem_bufctl_audit_t *bcp = (kmem_bufctl_audit_t *)btp->bt_bufctl;
1972         kmem_slab_t *sp;
1973 
1974         if (btp->bt_bxstat != ((intptr_t)bcp ^ KMEM_BUFTAG_ALLOC)) {
1975                 if (btp->bt_bxstat == ((intptr_t)bcp ^ KMEM_BUFTAG_FREE)) {
1976                         kmem_error(KMERR_DUPFREE, cp, buf);
1977                         return (-1);
1978                 }
1979                 sp = kmem_findslab(cp, buf);
1980                 if (sp == NULL || sp->slab_cache != cp)
1981                         kmem_error(KMERR_BADADDR, cp, buf);
1982                 else
1983                         kmem_error(KMERR_REDZONE, cp, buf);
1984                 return (-1);
1985         }
1986 
1987         btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
1988 
1989         if ((cp->cache_flags & KMF_HASH) && bcp->bc_addr != buf) {
1990                 kmem_error(KMERR_BADBUFCTL, cp, buf);
1991                 return (-1);
1992         }
1993 
1994         if (btp->bt_redzone != KMEM_REDZONE_PATTERN) {
1995                 kmem_error(KMERR_REDZONE, cp, buf);
1996                 return (-1);
1997         }
1998 
1999         if (cp->cache_flags & KMF_AUDIT) {
2000                 if (cp->cache_flags & KMF_CONTENTS)
2001                         bcp->bc_contents = kmem_log_enter(kmem_content_log,
2002                             buf, cp->cache_contents);
2003                 KMEM_AUDIT(kmem_transaction_log, cp, bcp);
2004         }
2005 
2006         if ((cp->cache_flags & KMF_LITE) &&
2007             !(cp->cache_cflags & KMC_KMEM_ALLOC)) {
2008                 KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller);
2009         }
2010 
2011         if (cp->cache_flags & KMF_DEADBEEF) {
2012                 if (cp->cache_flags & KMF_LITE)
2013                         btp->bt_redzone = *(uint64_t *)buf;
2014                 else if (cp->cache_destructor != NULL)
2015                         cp->cache_destructor(buf, cp->cache_private);
2016 
2017                 copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
2018         }
2019 
2020         return (0);
2021 }
2022 
2023 /*
2024  * Free each object in magazine mp to cp's slab layer, and free mp itself.
2025  */
2026 static void
2027 kmem_magazine_destroy(kmem_cache_t *cp, kmem_magazine_t *mp, int nrounds)
2028 {
2029         int round;
2030 
2031         ASSERT(!list_link_active(&cp->cache_link) ||
2032             taskq_member(kmem_taskq, curthread));
2033 
2034         for (round = 0; round < nrounds; round++) {
2035                 void *buf = mp->mag_round[round];
2036 
2037                 if (cp->cache_flags & KMF_DEADBEEF) {
2038                         if (verify_pattern(KMEM_FREE_PATTERN, buf,
2039                             cp->cache_verify) != NULL) {
2040                                 kmem_error(KMERR_MODIFIED, cp, buf);
2041                                 continue;
2042                         }
2043                         if ((cp->cache_flags & KMF_LITE) &&
2044                             cp->cache_destructor != NULL) {
2045                                 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2046                                 *(uint64_t *)buf = btp->bt_redzone;
2047                                 cp->cache_destructor(buf, cp->cache_private);
2048                                 *(uint64_t *)buf = KMEM_FREE_PATTERN;
2049                         }
2050                 } else if (cp->cache_destructor != NULL) {
2051                         cp->cache_destructor(buf, cp->cache_private);
2052                 }
2053 
2054                 kmem_slab_free(cp, buf);
2055         }
2056         ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2057         kmem_cache_free(cp->cache_magtype->mt_cache, mp);
2058 }
2059 
2060 /*
2061  * Allocate a magazine from the depot.
2062  */
2063 static kmem_magazine_t *
2064 kmem_depot_alloc(kmem_cache_t *cp, kmem_maglist_t *mlp)
2065 {
2066         kmem_magazine_t *mp;
2067 
2068         /*
2069          * If we can't get the depot lock without contention,
2070          * update our contention count.  We use the depot
2071          * contention rate to determine whether we need to
2072          * increase the magazine size for better scalability.
2073          */
2074         if (!mutex_tryenter(&cp->cache_depot_lock)) {
2075                 mutex_enter(&cp->cache_depot_lock);
2076                 cp->cache_depot_contention++;
2077         }
2078 
2079         if ((mp = mlp->ml_list) != NULL) {
2080                 ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2081                 mlp->ml_list = mp->mag_next;
2082                 if (--mlp->ml_total < mlp->ml_min)
2083                         mlp->ml_min = mlp->ml_total;
2084                 mlp->ml_alloc++;
2085         }
2086 
2087         mutex_exit(&cp->cache_depot_lock);
2088 
2089         return (mp);
2090 }
2091 
2092 /*
2093  * Free a magazine to the depot.
2094  */
2095 static void
2096 kmem_depot_free(kmem_cache_t *cp, kmem_maglist_t *mlp, kmem_magazine_t *mp)
2097 {
2098         mutex_enter(&cp->cache_depot_lock);
2099         ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2100         mp->mag_next = mlp->ml_list;
2101         mlp->ml_list = mp;
2102         mlp->ml_total++;
2103         mutex_exit(&cp->cache_depot_lock);
2104 }
2105 
2106 /*
2107  * Update the working set statistics for cp's depot.
2108  */
2109 static void
2110 kmem_depot_ws_update(kmem_cache_t *cp)
2111 {
2112         mutex_enter(&cp->cache_depot_lock);
2113         cp->cache_full.ml_reaplimit = cp->cache_full.ml_min;
2114         cp->cache_full.ml_min = cp->cache_full.ml_total;
2115         cp->cache_empty.ml_reaplimit = cp->cache_empty.ml_min;
2116         cp->cache_empty.ml_min = cp->cache_empty.ml_total;
2117         mutex_exit(&cp->cache_depot_lock);
2118 }
2119 
2120 /*
2121  * Set the working set statistics for cp's depot to zero.  (Everything is
2122  * eligible for reaping.)
2123  */
2124 static void
2125 kmem_depot_ws_zero(kmem_cache_t *cp)
2126 {
2127         mutex_enter(&cp->cache_depot_lock);
2128         cp->cache_full.ml_reaplimit = cp->cache_full.ml_total;
2129         cp->cache_full.ml_min = cp->cache_full.ml_total;
2130         cp->cache_empty.ml_reaplimit = cp->cache_empty.ml_total;
2131         cp->cache_empty.ml_min = cp->cache_empty.ml_total;
2132         mutex_exit(&cp->cache_depot_lock);
2133 }
2134 
2135 /*
2136  * The number of bytes to reap before we call kpreempt(). The default (1MB)
2137  * causes us to preempt reaping up to hundreds of times per second. Using a
2138  * larger value (1GB) causes this to have virtually no effect.
2139  */
2140 size_t kmem_reap_preempt_bytes = 1024 * 1024;
2141 
2142 /*
2143  * Reap all magazines that have fallen out of the depot's working set.
2144  */
2145 static void
2146 kmem_depot_ws_reap(kmem_cache_t *cp)
2147 {
2148         size_t bytes = 0;
2149         long reap;
2150         kmem_magazine_t *mp;
2151 
2152         ASSERT(!list_link_active(&cp->cache_link) ||
2153             taskq_member(kmem_taskq, curthread));
2154 
2155         reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
2156         while (reap-- &&
2157             (mp = kmem_depot_alloc(cp, &cp->cache_full)) != NULL) {
2158                 kmem_magazine_destroy(cp, mp, cp->cache_magtype->mt_magsize);
2159                 bytes += cp->cache_magtype->mt_magsize * cp->cache_bufsize;
2160                 if (bytes > kmem_reap_preempt_bytes) {
2161                         kpreempt(KPREEMPT_SYNC);
2162                         bytes = 0;
2163                 }
2164         }
2165 
2166         reap = MIN(cp->cache_empty.ml_reaplimit, cp->cache_empty.ml_min);
2167         while (reap-- &&
2168             (mp = kmem_depot_alloc(cp, &cp->cache_empty)) != NULL) {
2169                 kmem_magazine_destroy(cp, mp, 0);
2170                 bytes += cp->cache_magtype->mt_magsize * cp->cache_bufsize;
2171                 if (bytes > kmem_reap_preempt_bytes) {
2172                         kpreempt(KPREEMPT_SYNC);
2173                         bytes = 0;
2174                 }
2175         }
2176 }
2177 
2178 static void
2179 kmem_cpu_reload(kmem_cpu_cache_t *ccp, kmem_magazine_t *mp, int rounds)
2180 {
2181         ASSERT((ccp->cc_loaded == NULL && ccp->cc_rounds == -1) ||
2182             (ccp->cc_loaded && ccp->cc_rounds + rounds == ccp->cc_magsize));
2183         ASSERT(ccp->cc_magsize > 0);
2184 
2185         ccp->cc_ploaded = ccp->cc_loaded;
2186         ccp->cc_prounds = ccp->cc_rounds;
2187         ccp->cc_loaded = mp;
2188         ccp->cc_rounds = rounds;
2189 }
2190 
2191 /*
2192  * Intercept kmem alloc/free calls during crash dump in order to avoid
2193  * changing kmem state while memory is being saved to the dump device.
2194  * Otherwise, ::kmem_verify will report "corrupt buffers".  Note that
2195  * there are no locks because only one CPU calls kmem during a crash
2196  * dump. To enable this feature, first create the associated vmem
2197  * arena with VMC_DUMPSAFE.
2198  */
2199 static void *kmem_dump_start;   /* start of pre-reserved heap */
2200 static void *kmem_dump_end;     /* end of heap area */
2201 static void *kmem_dump_curr;    /* current free heap pointer */
2202 static size_t kmem_dump_size;   /* size of heap area */
2203 
2204 /* append to each buf created in the pre-reserved heap */
2205 typedef struct kmem_dumpctl {
2206         void    *kdc_next;      /* cache dump free list linkage */
2207 } kmem_dumpctl_t;
2208 
2209 #define KMEM_DUMPCTL(cp, buf)   \
2210         ((kmem_dumpctl_t *)P2ROUNDUP((uintptr_t)(buf) + (cp)->cache_bufsize, \
2211             sizeof (void *)))
2212 
2213 /* Keep some simple stats. */
2214 #define KMEM_DUMP_LOGS  (100)
2215 
2216 typedef struct kmem_dump_log {
2217         kmem_cache_t    *kdl_cache;
2218         uint_t          kdl_allocs;             /* # of dump allocations */
2219         uint_t          kdl_frees;              /* # of dump frees */
2220         uint_t          kdl_alloc_fails;        /* # of allocation failures */
2221         uint_t          kdl_free_nondump;       /* # of non-dump frees */
2222         uint_t          kdl_unsafe;             /* cache was used, but unsafe */
2223 } kmem_dump_log_t;
2224 
2225 static kmem_dump_log_t *kmem_dump_log;
2226 static int kmem_dump_log_idx;
2227 
2228 #define KDI_LOG(cp, stat) {                                             \
2229         kmem_dump_log_t *kdl;                                           \
2230         if ((kdl = (kmem_dump_log_t *)((cp)->cache_dumplog)) != NULL) {      \
2231                 kdl->stat++;                                         \
2232         } else if (kmem_dump_log_idx < KMEM_DUMP_LOGS) {             \
2233                 kdl = &kmem_dump_log[kmem_dump_log_idx++];          \
2234                 kdl->stat++;                                         \
2235                 kdl->kdl_cache = (cp);                                       \
2236                 (cp)->cache_dumplog = kdl;                           \
2237         }                                                               \
2238 }
2239 
2240 /* set non zero for full report */
2241 uint_t kmem_dump_verbose = 0;
2242 
2243 /* stats for overize heap */
2244 uint_t kmem_dump_oversize_allocs = 0;
2245 uint_t kmem_dump_oversize_max = 0;
2246 
2247 static void
2248 kmem_dumppr(char **pp, char *e, const char *format, ...)
2249 {
2250         char *p = *pp;
2251 
2252         if (p < e) {
2253                 int n;
2254                 va_list ap;
2255 
2256                 va_start(ap, format);
2257                 n = vsnprintf(p, e - p, format, ap);
2258                 va_end(ap);
2259                 *pp = p + n;
2260         }
2261 }
2262 
2263 /*
2264  * Called when dumpadm(1M) configures dump parameters.
2265  */
2266 void
2267 kmem_dump_init(size_t size)
2268 {
2269         if (kmem_dump_start != NULL)
2270                 kmem_free(kmem_dump_start, kmem_dump_size);
2271 
2272         if (kmem_dump_log == NULL)
2273                 kmem_dump_log = (kmem_dump_log_t *)kmem_zalloc(KMEM_DUMP_LOGS *
2274                     sizeof (kmem_dump_log_t), KM_SLEEP);
2275 
2276         kmem_dump_start = kmem_alloc(size, KM_SLEEP);
2277 
2278         if (kmem_dump_start != NULL) {
2279                 kmem_dump_size = size;
2280                 kmem_dump_curr = kmem_dump_start;
2281                 kmem_dump_end = (void *)((char *)kmem_dump_start + size);
2282                 copy_pattern(KMEM_UNINITIALIZED_PATTERN, kmem_dump_start, size);
2283         } else {
2284                 kmem_dump_size = 0;
2285                 kmem_dump_curr = NULL;
2286                 kmem_dump_end = NULL;
2287         }
2288 }
2289 
2290 /*
2291  * Set flag for each kmem_cache_t if is safe to use alternate dump
2292  * memory. Called just before panic crash dump starts. Set the flag
2293  * for the calling CPU.
2294  */
2295 void
2296 kmem_dump_begin(void)
2297 {
2298         ASSERT(panicstr != NULL);
2299         if (kmem_dump_start != NULL) {
2300                 kmem_cache_t *cp;
2301 
2302                 for (cp = list_head(&kmem_caches); cp != NULL;
2303                     cp = list_next(&kmem_caches, cp)) {
2304                         kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2305 
2306                         if (cp->cache_arena->vm_cflags & VMC_DUMPSAFE) {
2307                                 cp->cache_flags |= KMF_DUMPDIVERT;
2308                                 ccp->cc_flags |= KMF_DUMPDIVERT;
2309                                 ccp->cc_dump_rounds = ccp->cc_rounds;
2310                                 ccp->cc_dump_prounds = ccp->cc_prounds;
2311                                 ccp->cc_rounds = ccp->cc_prounds = -1;
2312                         } else {
2313                                 cp->cache_flags |= KMF_DUMPUNSAFE;
2314                                 ccp->cc_flags |= KMF_DUMPUNSAFE;
2315                         }
2316                 }
2317         }
2318 }
2319 
2320 /*
2321  * finished dump intercept
2322  * print any warnings on the console
2323  * return verbose information to dumpsys() in the given buffer
2324  */
2325 size_t
2326 kmem_dump_finish(char *buf, size_t size)
2327 {
2328         int kdi_idx;
2329         int kdi_end = kmem_dump_log_idx;
2330         int percent = 0;
2331         int header = 0;
2332         int warn = 0;
2333         size_t used;
2334         kmem_cache_t *cp;
2335         kmem_dump_log_t *kdl;
2336         char *e = buf + size;
2337         char *p = buf;
2338 
2339         if (kmem_dump_size == 0 || kmem_dump_verbose == 0)
2340                 return (0);
2341 
2342         used = (char *)kmem_dump_curr - (char *)kmem_dump_start;
2343         percent = (used * 100) / kmem_dump_size;
2344 
2345         kmem_dumppr(&p, e, "%% heap used,%d\n", percent);
2346         kmem_dumppr(&p, e, "used bytes,%ld\n", used);
2347         kmem_dumppr(&p, e, "heap size,%ld\n", kmem_dump_size);
2348         kmem_dumppr(&p, e, "Oversize allocs,%d\n",
2349             kmem_dump_oversize_allocs);
2350         kmem_dumppr(&p, e, "Oversize max size,%ld\n",
2351             kmem_dump_oversize_max);
2352 
2353         for (kdi_idx = 0; kdi_idx < kdi_end; kdi_idx++) {
2354                 kdl = &kmem_dump_log[kdi_idx];
2355                 cp = kdl->kdl_cache;
2356                 if (cp == NULL)
2357                         break;
2358                 if (kdl->kdl_alloc_fails)
2359                         ++warn;
2360                 if (header == 0) {
2361                         kmem_dumppr(&p, e,
2362                             "Cache Name,Allocs,Frees,Alloc Fails,"
2363                             "Nondump Frees,Unsafe Allocs/Frees\n");
2364                         header = 1;
2365                 }
2366                 kmem_dumppr(&p, e, "%s,%d,%d,%d,%d,%d\n",
2367                     cp->cache_name, kdl->kdl_allocs, kdl->kdl_frees,
2368                     kdl->kdl_alloc_fails, kdl->kdl_free_nondump,
2369                     kdl->kdl_unsafe);
2370         }
2371 
2372         /* return buffer size used */
2373         if (p < e)
2374                 bzero(p, e - p);
2375         return (p - buf);
2376 }
2377 
2378 /*
2379  * Allocate a constructed object from alternate dump memory.
2380  */
2381 void *
2382 kmem_cache_alloc_dump(kmem_cache_t *cp, int kmflag)
2383 {
2384         void *buf;
2385         void *curr;
2386         char *bufend;
2387 
2388         /* return a constructed object */
2389         if ((buf = cp->cache_dumpfreelist) != NULL) {
2390                 cp->cache_dumpfreelist = KMEM_DUMPCTL(cp, buf)->kdc_next;
2391                 KDI_LOG(cp, kdl_allocs);
2392                 return (buf);
2393         }
2394 
2395         /* create a new constructed object */
2396         curr = kmem_dump_curr;
2397         buf = (void *)P2ROUNDUP((uintptr_t)curr, cp->cache_align);
2398         bufend = (char *)KMEM_DUMPCTL(cp, buf) + sizeof (kmem_dumpctl_t);
2399 
2400         /* hat layer objects cannot cross a page boundary */
2401         if (cp->cache_align < PAGESIZE) {
2402                 char *page = (char *)P2ROUNDUP((uintptr_t)buf, PAGESIZE);
2403                 if (bufend > page) {
2404                         bufend += page - (char *)buf;
2405                         buf = (void *)page;
2406                 }
2407         }
2408 
2409         /* fall back to normal alloc if reserved area is used up */
2410         if (bufend > (char *)kmem_dump_end) {
2411                 kmem_dump_curr = kmem_dump_end;
2412                 KDI_LOG(cp, kdl_alloc_fails);
2413                 return (NULL);
2414         }
2415 
2416         /*
2417          * Must advance curr pointer before calling a constructor that
2418          * may also allocate memory.
2419          */
2420         kmem_dump_curr = bufend;
2421 
2422         /* run constructor */
2423         if (cp->cache_constructor != NULL &&
2424             cp->cache_constructor(buf, cp->cache_private, kmflag)
2425             != 0) {
2426 #ifdef DEBUG
2427                 printf("name='%s' cache=0x%p: kmem cache constructor failed\n",
2428                     cp->cache_name, (void *)cp);
2429 #endif
2430                 /* reset curr pointer iff no allocs were done */
2431                 if (kmem_dump_curr == bufend)
2432                         kmem_dump_curr = curr;
2433 
2434                 /* fall back to normal alloc if the constructor fails */
2435                 KDI_LOG(cp, kdl_alloc_fails);
2436                 return (NULL);
2437         }
2438 
2439         KDI_LOG(cp, kdl_allocs);
2440         return (buf);
2441 }
2442 
2443 /*
2444  * Free a constructed object in alternate dump memory.
2445  */
2446 int
2447 kmem_cache_free_dump(kmem_cache_t *cp, void *buf)
2448 {
2449         /* save constructed buffers for next time */
2450         if ((char *)buf >= (char *)kmem_dump_start &&
2451             (char *)buf < (char *)kmem_dump_end) {
2452                 KMEM_DUMPCTL(cp, buf)->kdc_next = cp->cache_dumpfreelist;
2453                 cp->cache_dumpfreelist = buf;
2454                 KDI_LOG(cp, kdl_frees);
2455                 return (0);
2456         }
2457 
2458         /* count all non-dump buf frees */
2459         KDI_LOG(cp, kdl_free_nondump);
2460 
2461         /* just drop buffers that were allocated before dump started */
2462         if (kmem_dump_curr < kmem_dump_end)
2463                 return (0);
2464 
2465         /* fall back to normal free if reserved area is used up */
2466         return (1);
2467 }
2468 
2469 /*
2470  * Allocate a constructed object from cache cp.
2471  */
2472 void *
2473 kmem_cache_alloc(kmem_cache_t *cp, int kmflag)
2474 {
2475         kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2476         kmem_magazine_t *fmp;
2477         void *buf;
2478 
2479         mutex_enter(&ccp->cc_lock);
2480         for (;;) {
2481                 /*
2482                  * If there's an object available in the current CPU's
2483                  * loaded magazine, just take it and return.
2484                  */
2485                 if (ccp->cc_rounds > 0) {
2486                         buf = ccp->cc_loaded->mag_round[--ccp->cc_rounds];
2487                         ccp->cc_alloc++;
2488                         mutex_exit(&ccp->cc_lock);
2489                         if (ccp->cc_flags & (KMF_BUFTAG | KMF_DUMPUNSAFE)) {
2490                                 if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2491                                         ASSERT(!(ccp->cc_flags &
2492                                             KMF_DUMPDIVERT));
2493                                         KDI_LOG(cp, kdl_unsafe);
2494                                 }
2495                                 if ((ccp->cc_flags & KMF_BUFTAG) &&
2496                                     kmem_cache_alloc_debug(cp, buf, kmflag, 0,
2497                                     caller()) != 0) {
2498                                         if (kmflag & KM_NOSLEEP)
2499                                                 return (NULL);
2500                                         mutex_enter(&ccp->cc_lock);
2501                                         continue;
2502                                 }
2503                         }
2504                         return (buf);
2505                 }
2506 
2507                 /*
2508                  * The loaded magazine is empty.  If the previously loaded
2509                  * magazine was full, exchange them and try again.
2510                  */
2511                 if (ccp->cc_prounds > 0) {
2512                         kmem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
2513                         continue;
2514                 }
2515 
2516                 /*
2517                  * Return an alternate buffer at dump time to preserve
2518                  * the heap.
2519                  */
2520                 if (ccp->cc_flags & (KMF_DUMPDIVERT | KMF_DUMPUNSAFE)) {
2521                         if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2522                                 ASSERT(!(ccp->cc_flags & KMF_DUMPDIVERT));
2523                                 /* log it so that we can warn about it */
2524                                 KDI_LOG(cp, kdl_unsafe);
2525                         } else {
2526                                 if ((buf = kmem_cache_alloc_dump(cp, kmflag)) !=
2527                                     NULL) {
2528                                         mutex_exit(&ccp->cc_lock);
2529                                         return (buf);
2530                                 }
2531                                 break;          /* fall back to slab layer */
2532                         }
2533                 }
2534 
2535                 /*
2536                  * If the magazine layer is disabled, break out now.
2537                  */
2538                 if (ccp->cc_magsize == 0)
2539                         break;
2540 
2541                 /*
2542                  * Try to get a full magazine from the depot.
2543                  */
2544                 fmp = kmem_depot_alloc(cp, &cp->cache_full);
2545                 if (fmp != NULL) {
2546                         if (ccp->cc_ploaded != NULL)
2547                                 kmem_depot_free(cp, &cp->cache_empty,
2548                                     ccp->cc_ploaded);
2549                         kmem_cpu_reload(ccp, fmp, ccp->cc_magsize);
2550                         continue;
2551                 }
2552 
2553                 /*
2554                  * There are no full magazines in the depot,
2555                  * so fall through to the slab layer.
2556                  */
2557                 break;
2558         }
2559         mutex_exit(&ccp->cc_lock);
2560 
2561         /*
2562          * We couldn't allocate a constructed object from the magazine layer,
2563          * so get a raw buffer from the slab layer and apply its constructor.
2564          */
2565         buf = kmem_slab_alloc(cp, kmflag);
2566 
2567         if (buf == NULL)
2568                 return (NULL);
2569 
2570         if (cp->cache_flags & KMF_BUFTAG) {
2571                 /*
2572                  * Make kmem_cache_alloc_debug() apply the constructor for us.
2573                  */
2574                 int rc = kmem_cache_alloc_debug(cp, buf, kmflag, 1, caller());
2575                 if (rc != 0) {
2576                         if (kmflag & KM_NOSLEEP)
2577                                 return (NULL);
2578                         /*
2579                          * kmem_cache_alloc_debug() detected corruption
2580                          * but didn't panic (kmem_panic <= 0). We should not be
2581                          * here because the constructor failed (indicated by a
2582                          * return code of 1). Try again.
2583                          */
2584                         ASSERT(rc == -1);
2585                         return (kmem_cache_alloc(cp, kmflag));
2586                 }
2587                 return (buf);
2588         }
2589 
2590         if (cp->cache_constructor != NULL &&
2591             cp->cache_constructor(buf, cp->cache_private, kmflag) != 0) {
2592                 atomic_inc_64(&cp->cache_alloc_fail);
2593                 kmem_slab_free(cp, buf);
2594                 return (NULL);
2595         }
2596 
2597         return (buf);
2598 }
2599 
2600 /*
2601  * The freed argument tells whether or not kmem_cache_free_debug() has already
2602  * been called so that we can avoid the duplicate free error. For example, a
2603  * buffer on a magazine has already been freed by the client but is still
2604  * constructed.
2605  */
2606 static void
2607 kmem_slab_free_constructed(kmem_cache_t *cp, void *buf, boolean_t freed)
2608 {
2609         if (!freed && (cp->cache_flags & KMF_BUFTAG))
2610                 if (kmem_cache_free_debug(cp, buf, caller()) == -1)
2611                         return;
2612 
2613         /*
2614          * Note that if KMF_DEADBEEF is in effect and KMF_LITE is not,
2615          * kmem_cache_free_debug() will have already applied the destructor.
2616          */
2617         if ((cp->cache_flags & (KMF_DEADBEEF | KMF_LITE)) != KMF_DEADBEEF &&
2618             cp->cache_destructor != NULL) {
2619                 if (cp->cache_flags & KMF_DEADBEEF) {    /* KMF_LITE implied */
2620                         kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2621                         *(uint64_t *)buf = btp->bt_redzone;
2622                         cp->cache_destructor(buf, cp->cache_private);
2623                         *(uint64_t *)buf = KMEM_FREE_PATTERN;
2624                 } else {
2625                         cp->cache_destructor(buf, cp->cache_private);
2626                 }
2627         }
2628 
2629         kmem_slab_free(cp, buf);
2630 }
2631 
2632 /*
2633  * Used when there's no room to free a buffer to the per-CPU cache.
2634  * Drops and re-acquires &ccp->cc_lock, and returns non-zero if the
2635  * caller should try freeing to the per-CPU cache again.
2636  * Note that we don't directly install the magazine in the cpu cache,
2637  * since its state may have changed wildly while the lock was dropped.
2638  */
2639 static int
2640 kmem_cpucache_magazine_alloc(kmem_cpu_cache_t *ccp, kmem_cache_t *cp)
2641 {
2642         kmem_magazine_t *emp;
2643         kmem_magtype_t *mtp;
2644 
2645         ASSERT(MUTEX_HELD(&ccp->cc_lock));
2646         ASSERT(((uint_t)ccp->cc_rounds == ccp->cc_magsize ||
2647             ((uint_t)ccp->cc_rounds == -1)) &&
2648             ((uint_t)ccp->cc_prounds == ccp->cc_magsize ||
2649             ((uint_t)ccp->cc_prounds == -1)));
2650 
2651         emp = kmem_depot_alloc(cp, &cp->cache_empty);
2652         if (emp != NULL) {
2653                 if (ccp->cc_ploaded != NULL)
2654                         kmem_depot_free(cp, &cp->cache_full,
2655                             ccp->cc_ploaded);
2656                 kmem_cpu_reload(ccp, emp, 0);
2657                 return (1);
2658         }
2659         /*
2660          * There are no empty magazines in the depot,
2661          * so try to allocate a new one.  We must drop all locks
2662          * across kmem_cache_alloc() because lower layers may
2663          * attempt to allocate from this cache.
2664          */
2665         mtp = cp->cache_magtype;
2666         mutex_exit(&ccp->cc_lock);
2667         emp = kmem_cache_alloc(mtp->mt_cache, KM_NOSLEEP);
2668         mutex_enter(&ccp->cc_lock);
2669 
2670         if (emp != NULL) {
2671                 /*
2672                  * We successfully allocated an empty magazine.
2673                  * However, we had to drop ccp->cc_lock to do it,
2674                  * so the cache's magazine size may have changed.
2675                  * If so, free the magazine and try again.
2676                  */
2677                 if (ccp->cc_magsize != mtp->mt_magsize) {
2678                         mutex_exit(&ccp->cc_lock);
2679                         kmem_cache_free(mtp->mt_cache, emp);
2680                         mutex_enter(&ccp->cc_lock);
2681                         return (1);
2682                 }
2683 
2684                 /*
2685                  * We got a magazine of the right size.  Add it to
2686                  * the depot and try the whole dance again.
2687                  */
2688                 kmem_depot_free(cp, &cp->cache_empty, emp);
2689                 return (1);
2690         }
2691 
2692         /*
2693          * We couldn't allocate an empty magazine,
2694          * so fall through to the slab layer.
2695          */
2696         return (0);
2697 }
2698 
2699 /*
2700  * Free a constructed object to cache cp.
2701  */
2702 void
2703 kmem_cache_free(kmem_cache_t *cp, void *buf)
2704 {
2705         kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2706 
2707         /*
2708          * The client must not free either of the buffers passed to the move
2709          * callback function.
2710          */
2711         ASSERT(cp->cache_defrag == NULL ||
2712             cp->cache_defrag->kmd_thread != curthread ||
2713             (buf != cp->cache_defrag->kmd_from_buf &&
2714             buf != cp->cache_defrag->kmd_to_buf));
2715 
2716         if (ccp->cc_flags & (KMF_BUFTAG | KMF_DUMPDIVERT | KMF_DUMPUNSAFE)) {
2717                 if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2718                         ASSERT(!(ccp->cc_flags & KMF_DUMPDIVERT));
2719                         /* log it so that we can warn about it */
2720                         KDI_LOG(cp, kdl_unsafe);
2721                 } else if (KMEM_DUMPCC(ccp) && !kmem_cache_free_dump(cp, buf)) {
2722                         return;
2723                 }
2724                 if (ccp->cc_flags & KMF_BUFTAG) {
2725                         if (kmem_cache_free_debug(cp, buf, caller()) == -1)
2726                                 return;
2727                 }
2728         }
2729 
2730         mutex_enter(&ccp->cc_lock);
2731         /*
2732          * Any changes to this logic should be reflected in kmem_slab_prefill()
2733          */
2734         for (;;) {
2735                 /*
2736                  * If there's a slot available in the current CPU's
2737                  * loaded magazine, just put the object there and return.
2738                  */
2739                 if ((uint_t)ccp->cc_rounds < ccp->cc_magsize) {
2740                         ccp->cc_loaded->mag_round[ccp->cc_rounds++] = buf;
2741                         ccp->cc_free++;
2742                         mutex_exit(&ccp->cc_lock);
2743                         return;
2744                 }
2745 
2746                 /*
2747                  * The loaded magazine is full.  If the previously loaded
2748                  * magazine was empty, exchange them and try again.
2749                  */
2750                 if (ccp->cc_prounds == 0) {
2751                         kmem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
2752                         continue;
2753                 }
2754 
2755                 /*
2756                  * If the magazine layer is disabled, break out now.
2757                  */
2758                 if (ccp->cc_magsize == 0)
2759                         break;
2760 
2761                 if (!kmem_cpucache_magazine_alloc(ccp, cp)) {
2762                         /*
2763                          * We couldn't free our constructed object to the
2764                          * magazine layer, so apply its destructor and free it
2765                          * to the slab layer.
2766                          */
2767                         break;
2768                 }
2769         }
2770         mutex_exit(&ccp->cc_lock);
2771         kmem_slab_free_constructed(cp, buf, B_TRUE);
2772 }
2773 
2774 static void
2775 kmem_slab_prefill(kmem_cache_t *cp, kmem_slab_t *sp)
2776 {
2777         kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2778         int cache_flags = cp->cache_flags;
2779 
2780         kmem_bufctl_t *next, *head;
2781         size_t nbufs;
2782 
2783         /*
2784          * Completely allocate the newly created slab and put the pre-allocated
2785          * buffers in magazines. Any of the buffers that cannot be put in
2786          * magazines must be returned to the slab.
2787          */
2788         ASSERT(MUTEX_HELD(&cp->cache_lock));
2789         ASSERT((cache_flags & (KMF_PREFILL|KMF_BUFTAG)) == KMF_PREFILL);
2790         ASSERT(cp->cache_constructor == NULL);
2791         ASSERT(sp->slab_cache == cp);
2792         ASSERT(sp->slab_refcnt == 1);
2793         ASSERT(sp->slab_head != NULL && sp->slab_chunks > sp->slab_refcnt);
2794         ASSERT(avl_find(&cp->cache_partial_slabs, sp, NULL) == NULL);
2795 
2796         head = sp->slab_head;
2797         nbufs = (sp->slab_chunks - sp->slab_refcnt);
2798         sp->slab_head = NULL;
2799         sp->slab_refcnt += nbufs;
2800         cp->cache_bufslab -= nbufs;
2801         cp->cache_slab_alloc += nbufs;
2802         list_insert_head(&cp->cache_complete_slabs, sp);
2803         cp->cache_complete_slab_count++;
2804         mutex_exit(&cp->cache_lock);
2805         mutex_enter(&ccp->cc_lock);
2806 
2807         while (head != NULL) {
2808                 void *buf = KMEM_BUF(cp, head);
2809                 /*
2810                  * If there's a slot available in the current CPU's
2811                  * loaded magazine, just put the object there and
2812                  * continue.
2813                  */
2814                 if ((uint_t)ccp->cc_rounds < ccp->cc_magsize) {
2815                         ccp->cc_loaded->mag_round[ccp->cc_rounds++] =
2816                             buf;
2817                         ccp->cc_free++;
2818                         nbufs--;
2819                         head = head->bc_next;
2820                         continue;
2821                 }
2822 
2823                 /*
2824                  * The loaded magazine is full.  If the previously
2825                  * loaded magazine was empty, exchange them and try
2826                  * again.
2827                  */
2828                 if (ccp->cc_prounds == 0) {
2829                         kmem_cpu_reload(ccp, ccp->cc_ploaded,
2830                             ccp->cc_prounds);
2831                         continue;
2832                 }
2833 
2834                 /*
2835                  * If the magazine layer is disabled, break out now.
2836                  */
2837 
2838                 if (ccp->cc_magsize == 0) {
2839                         break;
2840                 }
2841 
2842                 if (!kmem_cpucache_magazine_alloc(ccp, cp))
2843                         break;
2844         }
2845         mutex_exit(&ccp->cc_lock);
2846         if (nbufs != 0) {
2847                 ASSERT(head != NULL);
2848 
2849                 /*
2850                  * If there was a failure, return remaining objects to
2851                  * the slab
2852                  */
2853                 while (head != NULL) {
2854                         ASSERT(nbufs != 0);
2855                         next = head->bc_next;
2856                         head->bc_next = NULL;
2857                         kmem_slab_free(cp, KMEM_BUF(cp, head));
2858                         head = next;
2859                         nbufs--;
2860                 }
2861         }
2862         ASSERT(head == NULL);
2863         ASSERT(nbufs == 0);
2864         mutex_enter(&cp->cache_lock);
2865 }
2866 
2867 void *
2868 kmem_zalloc(size_t size, int kmflag)
2869 {
2870         size_t index;
2871         void *buf;
2872 
2873         if ((index = ((size - 1) >> KMEM_ALIGN_SHIFT)) < KMEM_ALLOC_TABLE_MAX) {
2874                 kmem_cache_t *cp = kmem_alloc_table[index];
2875                 buf = kmem_cache_alloc(cp, kmflag);
2876                 if (buf != NULL) {
2877                         if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp)) {
2878                                 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2879                                 ((uint8_t *)buf)[size] = KMEM_REDZONE_BYTE;
2880                                 ((uint32_t *)btp)[1] = KMEM_SIZE_ENCODE(size);
2881 
2882                                 if (cp->cache_flags & KMF_LITE) {
2883                                         KMEM_BUFTAG_LITE_ENTER(btp,
2884                                             kmem_lite_count, caller());
2885                                 }
2886                         }
2887                         bzero(buf, size);
2888                 }
2889         } else {
2890                 buf = kmem_alloc(size, kmflag);
2891                 if (buf != NULL)
2892                         bzero(buf, size);
2893         }
2894         return (buf);
2895 }
2896 
2897 void *
2898 kmem_alloc(size_t size, int kmflag)
2899 {
2900         size_t index;
2901         kmem_cache_t *cp;
2902         void *buf;
2903 
2904         if ((index = ((size - 1) >> KMEM_ALIGN_SHIFT)) < KMEM_ALLOC_TABLE_MAX) {
2905                 cp = kmem_alloc_table[index];
2906                 /* fall through to kmem_cache_alloc() */
2907 
2908         } else if ((index = ((size - 1) >> KMEM_BIG_SHIFT)) <
2909             kmem_big_alloc_table_max) {
2910                 cp = kmem_big_alloc_table[index];
2911                 /* fall through to kmem_cache_alloc() */
2912 
2913         } else {
2914                 if (size == 0)
2915                         return (NULL);
2916 
2917                 buf = vmem_alloc(kmem_oversize_arena, size,
2918                     kmflag & KM_VMFLAGS);
2919                 if (buf == NULL)
2920                         kmem_log_event(kmem_failure_log, NULL, NULL,
2921                             (void *)size);
2922                 else if (KMEM_DUMP(kmem_slab_cache)) {
2923                         /* stats for dump intercept */
2924                         kmem_dump_oversize_allocs++;
2925                         if (size > kmem_dump_oversize_max)
2926                                 kmem_dump_oversize_max = size;
2927                 }
2928                 return (buf);
2929         }
2930 
2931         buf = kmem_cache_alloc(cp, kmflag);
2932         if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp) && buf != NULL) {
2933                 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2934                 ((uint8_t *)buf)[size] = KMEM_REDZONE_BYTE;
2935                 ((uint32_t *)btp)[1] = KMEM_SIZE_ENCODE(size);
2936 
2937                 if (cp->cache_flags & KMF_LITE) {
2938                         KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller());
2939                 }
2940         }
2941         return (buf);
2942 }
2943 
2944 void
2945 kmem_free(void *buf, size_t size)
2946 {
2947         size_t index;
2948         kmem_cache_t *cp;
2949 
2950         if ((index = (size - 1) >> KMEM_ALIGN_SHIFT) < KMEM_ALLOC_TABLE_MAX) {
2951                 cp = kmem_alloc_table[index];
2952                 /* fall through to kmem_cache_free() */
2953 
2954         } else if ((index = ((size - 1) >> KMEM_BIG_SHIFT)) <
2955             kmem_big_alloc_table_max) {
2956                 cp = kmem_big_alloc_table[index];
2957                 /* fall through to kmem_cache_free() */
2958 
2959         } else {
2960                 EQUIV(buf == NULL, size == 0);
2961                 if (buf == NULL && size == 0)
2962                         return;
2963                 vmem_free(kmem_oversize_arena, buf, size);
2964                 return;
2965         }
2966 
2967         if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp)) {
2968                 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2969                 uint32_t *ip = (uint32_t *)btp;
2970                 if (ip[1] != KMEM_SIZE_ENCODE(size)) {
2971                         if (*(uint64_t *)buf == KMEM_FREE_PATTERN) {
2972                                 kmem_error(KMERR_DUPFREE, cp, buf);
2973                                 return;
2974                         }
2975                         if (KMEM_SIZE_VALID(ip[1])) {
2976                                 ip[0] = KMEM_SIZE_ENCODE(size);
2977                                 kmem_error(KMERR_BADSIZE, cp, buf);
2978                         } else {
2979                                 kmem_error(KMERR_REDZONE, cp, buf);
2980                         }
2981                         return;
2982                 }
2983                 if (((uint8_t *)buf)[size] != KMEM_REDZONE_BYTE) {
2984                         kmem_error(KMERR_REDZONE, cp, buf);
2985                         return;
2986                 }
2987                 btp->bt_redzone = KMEM_REDZONE_PATTERN;
2988                 if (cp->cache_flags & KMF_LITE) {
2989                         KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count,
2990                             caller());
2991                 }
2992         }
2993         kmem_cache_free(cp, buf);
2994 }
2995 
2996 void *
2997 kmem_firewall_va_alloc(vmem_t *vmp, size_t size, int vmflag)
2998 {
2999         size_t realsize = size + vmp->vm_quantum;
3000         void *addr;
3001 
3002         /*
3003          * Annoying edge case: if 'size' is just shy of ULONG_MAX, adding
3004          * vm_quantum will cause integer wraparound.  Check for this, and
3005          * blow off the firewall page in this case.  Note that such a
3006          * giant allocation (the entire kernel address space) can never
3007          * be satisfied, so it will either fail immediately (VM_NOSLEEP)
3008          * or sleep forever (VM_SLEEP).  Thus, there is no need for a
3009          * corresponding check in kmem_firewall_va_free().
3010          */
3011         if (realsize < size)
3012                 realsize = size;
3013 
3014         /*
3015          * While boot still owns resource management, make sure that this
3016          * redzone virtual address allocation is properly accounted for in
3017          * OBPs "virtual-memory" "available" lists because we're
3018          * effectively claiming them for a red zone.  If we don't do this,
3019          * the available lists become too fragmented and too large for the
3020          * current boot/kernel memory list interface.
3021          */
3022         addr = vmem_alloc(vmp, realsize, vmflag | VM_NEXTFIT);
3023 
3024         if (addr != NULL && kvseg.s_base == NULL && realsize != size)
3025                 (void) boot_virt_alloc((char *)addr + size, vmp->vm_quantum);
3026 
3027         return (addr);
3028 }
3029 
3030 void
3031 kmem_firewall_va_free(vmem_t *vmp, void *addr, size_t size)
3032 {
3033         ASSERT((kvseg.s_base == NULL ?
3034             va_to_pfn((char *)addr + size) :
3035             hat_getpfnum(kas.a_hat, (caddr_t)addr + size)) == PFN_INVALID);
3036 
3037         vmem_free(vmp, addr, size + vmp->vm_quantum);
3038 }
3039 
3040 /*
3041  * Try to allocate at least `size' bytes of memory without sleeping or
3042  * panicking. Return actual allocated size in `asize'. If allocation failed,
3043  * try final allocation with sleep or panic allowed.
3044  */
3045 void *
3046 kmem_alloc_tryhard(size_t size, size_t *asize, int kmflag)
3047 {
3048         void *p;
3049 
3050         *asize = P2ROUNDUP(size, KMEM_ALIGN);
3051         do {
3052                 p = kmem_alloc(*asize, (kmflag | KM_NOSLEEP) & ~KM_PANIC);
3053                 if (p != NULL)
3054                         return (p);
3055                 *asize += KMEM_ALIGN;
3056         } while (*asize <= PAGESIZE);
3057 
3058         *asize = P2ROUNDUP(size, KMEM_ALIGN);
3059         return (kmem_alloc(*asize, kmflag));
3060 }
3061 
3062 /*
3063  * Reclaim all unused memory from a cache.
3064  */
3065 static void
3066 kmem_cache_reap(kmem_cache_t *cp)
3067 {
3068         ASSERT(taskq_member(kmem_taskq, curthread));
3069         cp->cache_reap++;
3070 
3071         /*
3072          * Ask the cache's owner to free some memory if possible.
3073          * The idea is to handle things like the inode cache, which
3074          * typically sits on a bunch of memory that it doesn't truly
3075          * *need*.  Reclaim policy is entirely up to the owner; this
3076          * callback is just an advisory plea for help.
3077          */
3078         if (cp->cache_reclaim != NULL) {
3079                 long delta;
3080 
3081                 /*
3082                  * Reclaimed memory should be reapable (not included in the
3083                  * depot's working set).
3084                  */
3085                 delta = cp->cache_full.ml_total;
3086                 cp->cache_reclaim(cp->cache_private);
3087                 delta = cp->cache_full.ml_total - delta;
3088                 if (delta > 0) {
3089                         mutex_enter(&cp->cache_depot_lock);
3090                         cp->cache_full.ml_reaplimit += delta;
3091                         cp->cache_full.ml_min += delta;
3092                         mutex_exit(&cp->cache_depot_lock);
3093                 }
3094         }
3095 
3096         kmem_depot_ws_reap(cp);
3097 
3098         if (cp->cache_defrag != NULL && !kmem_move_noreap) {
3099                 kmem_cache_defrag(cp);
3100         }
3101 }
3102 
3103 static void
3104 kmem_reap_timeout(void *flag_arg)
3105 {
3106         uint32_t *flag = (uint32_t *)flag_arg;
3107 
3108         ASSERT(flag == &kmem_reaping || flag == &kmem_reaping_idspace);
3109         *flag = 0;
3110 }
3111 
3112 static void
3113 kmem_reap_done(void *flag)
3114 {
3115         if (!callout_init_done) {
3116                 /* can't schedule a timeout at this point */
3117                 kmem_reap_timeout(flag);
3118         } else {
3119                 (void) timeout(kmem_reap_timeout, flag, kmem_reap_interval);
3120         }
3121 }
3122 
3123 static void
3124 kmem_reap_start(void *flag)
3125 {
3126         ASSERT(flag == &kmem_reaping || flag == &kmem_reaping_idspace);
3127 
3128         if (flag == &kmem_reaping) {
3129                 kmem_cache_applyall(kmem_cache_reap, kmem_taskq, TQ_NOSLEEP);
3130                 /*
3131                  * if we have segkp under heap, reap segkp cache.
3132                  */
3133                 if (segkp_fromheap)
3134                         segkp_cache_free();
3135         }
3136         else
3137                 kmem_cache_applyall_id(kmem_cache_reap, kmem_taskq, TQ_NOSLEEP);
3138 
3139         /*
3140          * We use taskq_dispatch() to schedule a timeout to clear
3141          * the flag so that kmem_reap() becomes self-throttling:
3142          * we won't reap again until the current reap completes *and*
3143          * at least kmem_reap_interval ticks have elapsed.
3144          */
3145         if (!taskq_dispatch(kmem_taskq, kmem_reap_done, flag, TQ_NOSLEEP))
3146                 kmem_reap_done(flag);
3147 }
3148 
3149 static void
3150 kmem_reap_common(void *flag_arg)
3151 {
3152         uint32_t *flag = (uint32_t *)flag_arg;
3153 
3154         if (MUTEX_HELD(&kmem_cache_lock) || kmem_taskq == NULL ||
3155             atomic_cas_32(flag, 0, 1) != 0)
3156                 return;
3157 
3158         /*
3159          * It may not be kosher to do memory allocation when a reap is called
3160          * (for example, if vmem_populate() is in the call chain).  So we
3161          * start the reap going with a TQ_NOALLOC dispatch.  If the dispatch
3162          * fails, we reset the flag, and the next reap will try again.
3163          */
3164         if (!taskq_dispatch(kmem_taskq, kmem_reap_start, flag, TQ_NOALLOC))
3165                 *flag = 0;
3166 }
3167 
3168 /*
3169  * Reclaim all unused memory from all caches.  Called from the VM system
3170  * when memory gets tight.
3171  */
3172 void
3173 kmem_reap(void)
3174 {
3175         kmem_reap_common(&kmem_reaping);
3176 }
3177 
3178 /*
3179  * Reclaim all unused memory from identifier arenas, called when a vmem
3180  * arena not back by memory is exhausted.  Since reaping memory-backed caches
3181  * cannot help with identifier exhaustion, we avoid both a large amount of
3182  * work and unwanted side-effects from reclaim callbacks.
3183  */
3184 void
3185 kmem_reap_idspace(void)
3186 {
3187         kmem_reap_common(&kmem_reaping_idspace);
3188 }
3189 
3190 /*
3191  * Purge all magazines from a cache and set its magazine limit to zero.
3192  * All calls are serialized by the kmem_taskq lock, except for the final
3193  * call from kmem_cache_destroy().
3194  */
3195 static void
3196 kmem_cache_magazine_purge(kmem_cache_t *cp)
3197 {
3198         kmem_cpu_cache_t *ccp;
3199         kmem_magazine_t *mp, *pmp;
3200         int rounds, prounds, cpu_seqid;
3201 
3202         ASSERT(!list_link_active(&cp->cache_link) ||
3203             taskq_member(kmem_taskq, curthread));
3204         ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
3205 
3206         for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3207                 ccp = &cp->cache_cpu[cpu_seqid];
3208 
3209                 mutex_enter(&ccp->cc_lock);
3210                 mp = ccp->cc_loaded;
3211                 pmp = ccp->cc_ploaded;
3212                 rounds = ccp->cc_rounds;
3213                 prounds = ccp->cc_prounds;
3214                 ccp->cc_loaded = NULL;
3215                 ccp->cc_ploaded = NULL;
3216                 ccp->cc_rounds = -1;
3217                 ccp->cc_prounds = -1;
3218                 ccp->cc_magsize = 0;
3219                 mutex_exit(&ccp->cc_lock);
3220 
3221                 if (mp)
3222                         kmem_magazine_destroy(cp, mp, rounds);
3223                 if (pmp)
3224                         kmem_magazine_destroy(cp, pmp, prounds);
3225         }
3226 
3227         kmem_depot_ws_zero(cp);
3228         kmem_depot_ws_reap(cp);
3229 }
3230 
3231 /*
3232  * Enable per-cpu magazines on a cache.
3233  */
3234 static void
3235 kmem_cache_magazine_enable(kmem_cache_t *cp)
3236 {
3237         int cpu_seqid;
3238 
3239         if (cp->cache_flags & KMF_NOMAGAZINE)
3240                 return;
3241 
3242         for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3243                 kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
3244                 mutex_enter(&ccp->cc_lock);
3245                 ccp->cc_magsize = cp->cache_magtype->mt_magsize;
3246                 mutex_exit(&ccp->cc_lock);
3247         }
3248 
3249 }
3250 
3251 /*
3252  * Reap (almost) everything right now.
3253  */
3254 void
3255 kmem_cache_reap_now(kmem_cache_t *cp)
3256 {
3257         ASSERT(list_link_active(&cp->cache_link));
3258 
3259         kmem_depot_ws_zero(cp);
3260 
3261         (void) taskq_dispatch(kmem_taskq,
3262             (task_func_t *)kmem_depot_ws_reap, cp, TQ_SLEEP);
3263         taskq_wait(kmem_taskq);
3264 }
3265 
3266 /*
3267  * Recompute a cache's magazine size.  The trade-off is that larger magazines
3268  * provide a higher transfer rate with the depot, while smaller magazines
3269  * reduce memory consumption.  Magazine resizing is an expensive operation;
3270  * it should not be done frequently.
3271  *
3272  * Changes to the magazine size are serialized by the kmem_taskq lock.
3273  *
3274  * Note: at present this only grows the magazine size.  It might be useful
3275  * to allow shrinkage too.
3276  */
3277 static void
3278 kmem_cache_magazine_resize(kmem_cache_t *cp)
3279 {
3280         kmem_magtype_t *mtp = cp->cache_magtype;
3281 
3282         ASSERT(taskq_member(kmem_taskq, curthread));
3283 
3284         if (cp->cache_chunksize < mtp->mt_maxbuf) {
3285                 kmem_cache_magazine_purge(cp);
3286                 mutex_enter(&cp->cache_depot_lock);
3287                 cp->cache_magtype = ++mtp;
3288                 cp->cache_depot_contention_prev =
3289                     cp->cache_depot_contention + INT_MAX;
3290                 mutex_exit(&cp->cache_depot_lock);
3291                 kmem_cache_magazine_enable(cp);
3292         }
3293 }
3294 
3295 /*
3296  * Rescale a cache's hash table, so that the table size is roughly the
3297  * cache size.  We want the average lookup time to be extremely small.
3298  */
3299 static void
3300 kmem_hash_rescale(kmem_cache_t *cp)
3301 {
3302         kmem_bufctl_t **old_table, **new_table, *bcp;
3303         size_t old_size, new_size, h;
3304 
3305         ASSERT(taskq_member(kmem_taskq, curthread));
3306 
3307         new_size = MAX(KMEM_HASH_INITIAL,
3308             1 << (highbit(3 * cp->cache_buftotal + 4) - 2));
3309         old_size = cp->cache_hash_mask + 1;
3310 
3311         if ((old_size >> 1) <= new_size && new_size <= (old_size << 1))
3312                 return;
3313 
3314         new_table = vmem_alloc(kmem_hash_arena, new_size * sizeof (void *),
3315             VM_NOSLEEP);
3316         if (new_table == NULL)
3317                 return;
3318         bzero(new_table, new_size * sizeof (void *));
3319 
3320         mutex_enter(&cp->cache_lock);
3321 
3322         old_size = cp->cache_hash_mask + 1;
3323         old_table = cp->cache_hash_table;
3324 
3325         cp->cache_hash_mask = new_size - 1;
3326         cp->cache_hash_table = new_table;
3327         cp->cache_rescale++;
3328 
3329         for (h = 0; h < old_size; h++) {
3330                 bcp = old_table[h];
3331                 while (bcp != NULL) {
3332                         void *addr = bcp->bc_addr;
3333                         kmem_bufctl_t *next_bcp = bcp->bc_next;
3334                         kmem_bufctl_t **hash_bucket = KMEM_HASH(cp, addr);
3335                         bcp->bc_next = *hash_bucket;
3336                         *hash_bucket = bcp;
3337                         bcp = next_bcp;
3338                 }
3339         }
3340 
3341         mutex_exit(&cp->cache_lock);
3342 
3343         vmem_free(kmem_hash_arena, old_table, old_size * sizeof (void *));
3344 }
3345 
3346 /*
3347  * Perform periodic maintenance on a cache: hash rescaling, depot working-set
3348  * update, magazine resizing, and slab consolidation.
3349  */
3350 static void
3351 kmem_cache_update(kmem_cache_t *cp)
3352 {
3353         int need_hash_rescale = 0;
3354         int need_magazine_resize = 0;
3355 
3356         ASSERT(MUTEX_HELD(&kmem_cache_lock));
3357 
3358         /*
3359          * If the cache has become much larger or smaller than its hash table,
3360          * fire off a request to rescale the hash table.
3361          */
3362         mutex_enter(&cp->cache_lock);
3363 
3364         if ((cp->cache_flags & KMF_HASH) &&
3365             (cp->cache_buftotal > (cp->cache_hash_mask << 1) ||
3366             (cp->cache_buftotal < (cp->cache_hash_mask >> 1) &&
3367             cp->cache_hash_mask > KMEM_HASH_INITIAL)))
3368                 need_hash_rescale = 1;
3369 
3370         mutex_exit(&cp->cache_lock);
3371 
3372         /*
3373          * Update the depot working set statistics.
3374          */
3375         kmem_depot_ws_update(cp);
3376 
3377         /*
3378          * If there's a lot of contention in the depot,
3379          * increase the magazine size.
3380          */
3381         mutex_enter(&cp->cache_depot_lock);
3382 
3383         if (cp->cache_chunksize < cp->cache_magtype->mt_maxbuf &&
3384             (int)(cp->cache_depot_contention -
3385             cp->cache_depot_contention_prev) > kmem_depot_contention)
3386                 need_magazine_resize = 1;
3387 
3388         cp->cache_depot_contention_prev = cp->cache_depot_contention;
3389 
3390         mutex_exit(&cp->cache_depot_lock);
3391 
3392         if (need_hash_rescale)
3393                 (void) taskq_dispatch(kmem_taskq,
3394                     (task_func_t *)kmem_hash_rescale, cp, TQ_NOSLEEP);
3395 
3396         if (need_magazine_resize)
3397                 (void) taskq_dispatch(kmem_taskq,
3398                     (task_func_t *)kmem_cache_magazine_resize, cp, TQ_NOSLEEP);
3399 
3400         if (cp->cache_defrag != NULL)
3401                 (void) taskq_dispatch(kmem_taskq,
3402                     (task_func_t *)kmem_cache_scan, cp, TQ_NOSLEEP);
3403 }
3404 
3405 static void kmem_update(void *);
3406 
3407 static void
3408 kmem_update_timeout(void *dummy)
3409 {
3410         (void) timeout(kmem_update, dummy, kmem_reap_interval);
3411 }
3412 
3413 static void
3414 kmem_update(void *dummy)
3415 {
3416         kmem_cache_applyall(kmem_cache_update, NULL, TQ_NOSLEEP);
3417 
3418         /*
3419          * We use taskq_dispatch() to reschedule the timeout so that
3420          * kmem_update() becomes self-throttling: it won't schedule
3421          * new tasks until all previous tasks have completed.
3422          */
3423         if (!taskq_dispatch(kmem_taskq, kmem_update_timeout, dummy, TQ_NOSLEEP))
3424                 kmem_update_timeout(NULL);
3425 }
3426 
3427 static int
3428 kmem_cache_kstat_update(kstat_t *ksp, int rw)
3429 {
3430         struct kmem_cache_kstat *kmcp = &kmem_cache_kstat;
3431         kmem_cache_t *cp = ksp->ks_private;
3432         uint64_t cpu_buf_avail;
3433         uint64_t buf_avail = 0;
3434         int cpu_seqid;
3435         long reap;
3436 
3437         ASSERT(MUTEX_HELD(&kmem_cache_kstat_lock));
3438 
3439         if (rw == KSTAT_WRITE)
3440                 return (EACCES);
3441 
3442         mutex_enter(&cp->cache_lock);
3443 
3444         kmcp->kmc_alloc_fail.value.ui64              = cp->cache_alloc_fail;
3445         kmcp->kmc_alloc.value.ui64           = cp->cache_slab_alloc;
3446         kmcp->kmc_free.value.ui64            = cp->cache_slab_free;
3447         kmcp->kmc_slab_alloc.value.ui64              = cp->cache_slab_alloc;
3448         kmcp->kmc_slab_free.value.ui64               = cp->cache_slab_free;
3449 
3450         for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3451                 kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
3452 
3453                 mutex_enter(&ccp->cc_lock);
3454 
3455                 cpu_buf_avail = 0;
3456                 if (ccp->cc_rounds > 0)
3457                         cpu_buf_avail += ccp->cc_rounds;
3458                 if (ccp->cc_prounds > 0)
3459                         cpu_buf_avail += ccp->cc_prounds;
3460 
3461                 kmcp->kmc_alloc.value.ui64   += ccp->cc_alloc;
3462                 kmcp->kmc_free.value.ui64    += ccp->cc_free;
3463                 buf_avail                       += cpu_buf_avail;
3464 
3465                 mutex_exit(&ccp->cc_lock);
3466         }
3467 
3468         mutex_enter(&cp->cache_depot_lock);
3469 
3470         kmcp->kmc_depot_alloc.value.ui64     = cp->cache_full.ml_alloc;
3471         kmcp->kmc_depot_free.value.ui64              = cp->cache_empty.ml_alloc;
3472         kmcp->kmc_depot_contention.value.ui64        = cp->cache_depot_contention;
3473         kmcp->kmc_full_magazines.value.ui64  = cp->cache_full.ml_total;
3474         kmcp->kmc_empty_magazines.value.ui64 = cp->cache_empty.ml_total;
3475         kmcp->kmc_magazine_size.value.ui64   =
3476             (cp->cache_flags & KMF_NOMAGAZINE) ?
3477             0 : cp->cache_magtype->mt_magsize;
3478 
3479         kmcp->kmc_alloc.value.ui64           += cp->cache_full.ml_alloc;
3480         kmcp->kmc_free.value.ui64            += cp->cache_empty.ml_alloc;
3481         buf_avail += cp->cache_full.ml_total * cp->cache_magtype->mt_magsize;
3482 
3483         reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
3484         reap = MIN(reap, cp->cache_full.ml_total);
3485 
3486         mutex_exit(&cp->cache_depot_lock);
3487 
3488         kmcp->kmc_buf_size.value.ui64        = cp->cache_bufsize;
3489         kmcp->kmc_align.value.ui64   = cp->cache_align;
3490         kmcp->kmc_chunk_size.value.ui64      = cp->cache_chunksize;
3491         kmcp->kmc_slab_size.value.ui64       = cp->cache_slabsize;
3492         kmcp->kmc_buf_constructed.value.ui64 = buf_avail;
3493         buf_avail += cp->cache_bufslab;
3494         kmcp->kmc_buf_avail.value.ui64       = buf_avail;
3495         kmcp->kmc_buf_inuse.value.ui64       = cp->cache_buftotal - buf_avail;
3496         kmcp->kmc_buf_total.value.ui64       = cp->cache_buftotal;
3497         kmcp->kmc_buf_max.value.ui64 = cp->cache_bufmax;
3498         kmcp->kmc_slab_create.value.ui64     = cp->cache_slab_create;
3499         kmcp->kmc_slab_destroy.value.ui64    = cp->cache_slab_destroy;
3500         kmcp->kmc_hash_size.value.ui64       = (cp->cache_flags & KMF_HASH) ?
3501             cp->cache_hash_mask + 1 : 0;
3502         kmcp->kmc_hash_lookup_depth.value.ui64       = cp->cache_lookup_depth;
3503         kmcp->kmc_hash_rescale.value.ui64    = cp->cache_rescale;
3504         kmcp->kmc_vmem_source.value.ui64     = cp->cache_arena->vm_id;
3505         kmcp->kmc_reap.value.ui64    = cp->cache_reap;
3506 
3507         if (cp->cache_defrag == NULL) {
3508                 kmcp->kmc_move_callbacks.value.ui64  = 0;
3509                 kmcp->kmc_move_yes.value.ui64                = 0;
3510                 kmcp->kmc_move_no.value.ui64         = 0;
3511                 kmcp->kmc_move_later.value.ui64              = 0;
3512                 kmcp->kmc_move_dont_need.value.ui64  = 0;
3513                 kmcp->kmc_move_dont_know.value.ui64  = 0;
3514                 kmcp->kmc_move_hunt_found.value.ui64 = 0;
3515                 kmcp->kmc_move_slabs_freed.value.ui64        = 0;
3516                 kmcp->kmc_defrag.value.ui64          = 0;
3517                 kmcp->kmc_scan.value.ui64            = 0;
3518                 kmcp->kmc_move_reclaimable.value.ui64        = 0;
3519         } else {
3520                 int64_t reclaimable;
3521 
3522                 kmem_defrag_t *kd = cp->cache_defrag;
3523                 kmcp->kmc_move_callbacks.value.ui64  = kd->kmd_callbacks;
3524                 kmcp->kmc_move_yes.value.ui64                = kd->kmd_yes;
3525                 kmcp->kmc_move_no.value.ui64         = kd->kmd_no;
3526                 kmcp->kmc_move_later.value.ui64              = kd->kmd_later;
3527                 kmcp->kmc_move_dont_need.value.ui64  = kd->kmd_dont_need;
3528                 kmcp->kmc_move_dont_know.value.ui64  = kd->kmd_dont_know;
3529                 kmcp->kmc_move_hunt_found.value.ui64 = 0;
3530                 kmcp->kmc_move_slabs_freed.value.ui64        = kd->kmd_slabs_freed;
3531                 kmcp->kmc_defrag.value.ui64          = kd->kmd_defrags;
3532                 kmcp->kmc_scan.value.ui64            = kd->kmd_scans;
3533 
3534                 reclaimable = cp->cache_bufslab - (cp->cache_maxchunks - 1);
3535                 reclaimable = MAX(reclaimable, 0);
3536                 reclaimable += ((uint64_t)reap * cp->cache_magtype->mt_magsize);
3537                 kmcp->kmc_move_reclaimable.value.ui64        = reclaimable;
3538         }
3539 
3540         mutex_exit(&cp->cache_lock);
3541         return (0);
3542 }
3543 
3544 /*
3545  * Return a named statistic about a particular cache.
3546  * This shouldn't be called very often, so it's currently designed for
3547  * simplicity (leverages existing kstat support) rather than efficiency.
3548  */
3549 uint64_t
3550 kmem_cache_stat(kmem_cache_t *cp, char *name)
3551 {
3552         int i;
3553         kstat_t *ksp = cp->cache_kstat;
3554         kstat_named_t *knp = (kstat_named_t *)&kmem_cache_kstat;
3555         uint64_t value = 0;
3556 
3557         if (ksp != NULL) {
3558                 mutex_enter(&kmem_cache_kstat_lock);
3559                 (void) kmem_cache_kstat_update(ksp, KSTAT_READ);
3560                 for (i = 0; i < ksp->ks_ndata; i++) {
3561                         if (strcmp(knp[i].name, name) == 0) {
3562                                 value = knp[i].value.ui64;
3563                                 break;
3564                         }
3565                 }
3566                 mutex_exit(&kmem_cache_kstat_lock);
3567         }
3568         return (value);
3569 }
3570 
3571 /*
3572  * Return an estimate of currently available kernel heap memory.
3573  * On 32-bit systems, physical memory may exceed virtual memory,
3574  * we just truncate the result at 1GB.
3575  */
3576 size_t
3577 kmem_avail(void)
3578 {
3579         spgcnt_t rmem = availrmem - tune.t_minarmem;
3580         spgcnt_t fmem = freemem - minfree;
3581 
3582         return ((size_t)ptob(MIN(MAX(MIN(rmem, fmem), 0),
3583             1 << (30 - PAGESHIFT))));
3584 }
3585 
3586 /*
3587  * Return the maximum amount of memory that is (in theory) allocatable
3588  * from the heap. This may be used as an estimate only since there
3589  * is no guarentee this space will still be available when an allocation
3590  * request is made, nor that the space may be allocated in one big request
3591  * due to kernel heap fragmentation.
3592  */
3593 size_t
3594 kmem_maxavail(void)
3595 {
3596         spgcnt_t pmem = availrmem - tune.t_minarmem;
3597         spgcnt_t vmem = btop(vmem_size(heap_arena, VMEM_FREE));
3598 
3599         return ((size_t)ptob(MAX(MIN(pmem, vmem), 0)));
3600 }
3601 
3602 /*
3603  * Indicate whether memory-intensive kmem debugging is enabled.
3604  */
3605 int
3606 kmem_debugging(void)
3607 {
3608         return (kmem_flags & (KMF_AUDIT | KMF_REDZONE));
3609 }
3610 
3611 /* binning function, sorts finely at the two extremes */
3612 #define KMEM_PARTIAL_SLAB_WEIGHT(sp, binshift)                          \
3613         ((((sp)->slab_refcnt <= (binshift)) ||                            \
3614             (((sp)->slab_chunks - (sp)->slab_refcnt) <= (binshift)))   \
3615             ? -(sp)->slab_refcnt                                     \
3616             : -((binshift) + ((sp)->slab_refcnt >> (binshift))))
3617 
3618 /*
3619  * Minimizing the number of partial slabs on the freelist minimizes
3620  * fragmentation (the ratio of unused buffers held by the slab layer). There are
3621  * two ways to get a slab off of the freelist: 1) free all the buffers on the
3622  * slab, and 2) allocate all the buffers on the slab. It follows that we want
3623  * the most-used slabs at the front of the list where they have the best chance
3624  * of being completely allocated, and the least-used slabs at a safe distance
3625  * from the front to improve the odds that the few remaining buffers will all be
3626  * freed before another allocation can tie up the slab. For that reason a slab
3627  * with a higher slab_refcnt sorts less than than a slab with a lower
3628  * slab_refcnt.
3629  *
3630  * However, if a slab has at least one buffer that is deemed unfreeable, we
3631  * would rather have that slab at the front of the list regardless of
3632  * slab_refcnt, since even one unfreeable buffer makes the entire slab
3633  * unfreeable. If the client returns KMEM_CBRC_NO in response to a cache_move()
3634  * callback, the slab is marked unfreeable for as long as it remains on the
3635  * freelist.
3636  */
3637 static int
3638 kmem_partial_slab_cmp(const void *p0, const void *p1)
3639 {
3640         const kmem_cache_t *cp;
3641         const kmem_slab_t *s0 = p0;
3642         const kmem_slab_t *s1 = p1;
3643         int w0, w1;
3644         size_t binshift;
3645 
3646         ASSERT(KMEM_SLAB_IS_PARTIAL(s0));
3647         ASSERT(KMEM_SLAB_IS_PARTIAL(s1));
3648         ASSERT(s0->slab_cache == s1->slab_cache);
3649         cp = s1->slab_cache;
3650         ASSERT(MUTEX_HELD(&cp->cache_lock));
3651         binshift = cp->cache_partial_binshift;
3652 
3653         /* weight of first slab */
3654         w0 = KMEM_PARTIAL_SLAB_WEIGHT(s0, binshift);
3655         if (s0->slab_flags & KMEM_SLAB_NOMOVE) {
3656                 w0 -= cp->cache_maxchunks;
3657         }
3658 
3659         /* weight of second slab */
3660         w1 = KMEM_PARTIAL_SLAB_WEIGHT(s1, binshift);
3661         if (s1->slab_flags & KMEM_SLAB_NOMOVE) {
3662                 w1 -= cp->cache_maxchunks;
3663         }
3664 
3665         if (w0 < w1)
3666                 return (-1);
3667         if (w0 > w1)
3668                 return (1);
3669 
3670         /* compare pointer values */
3671         if ((uintptr_t)s0 < (uintptr_t)s1)
3672                 return (-1);
3673         if ((uintptr_t)s0 > (uintptr_t)s1)
3674                 return (1);
3675 
3676         return (0);
3677 }
3678 
3679 /*
3680  * It must be valid to call the destructor (if any) on a newly created object.
3681  * That is, the constructor (if any) must leave the object in a valid state for
3682  * the destructor.
3683  */
3684 kmem_cache_t *
3685 kmem_cache_create(
3686         char *name,             /* descriptive name for this cache */
3687         size_t bufsize,         /* size of the objects it manages */
3688         size_t align,           /* required object alignment */
3689         int (*constructor)(void *, void *, int), /* object constructor */
3690         void (*destructor)(void *, void *),     /* object destructor */
3691         void (*reclaim)(void *), /* memory reclaim callback */
3692         void *private,          /* pass-thru arg for constr/destr/reclaim */
3693         vmem_t *vmp,            /* vmem source for slab allocation */
3694         int cflags)             /* cache creation flags */
3695 {
3696         int cpu_seqid;
3697         size_t chunksize;
3698         kmem_cache_t *cp;
3699         kmem_magtype_t *mtp;
3700         size_t csize = KMEM_CACHE_SIZE(max_ncpus);
3701 
3702 #ifdef  DEBUG
3703         /*
3704          * Cache names should conform to the rules for valid C identifiers
3705          */
3706         if (!strident_valid(name)) {
3707                 cmn_err(CE_CONT,
3708                     "kmem_cache_create: '%s' is an invalid cache name\n"
3709                     "cache names must conform to the rules for "
3710                     "C identifiers\n", name);
3711         }
3712 #endif  /* DEBUG */
3713 
3714         if (vmp == NULL)
3715                 vmp = kmem_default_arena;
3716 
3717         /*
3718          * If this kmem cache has an identifier vmem arena as its source, mark
3719          * it such to allow kmem_reap_idspace().
3720          */
3721         ASSERT(!(cflags & KMC_IDENTIFIER));   /* consumer should not set this */
3722         if (vmp->vm_cflags & VMC_IDENTIFIER)
3723                 cflags |= KMC_IDENTIFIER;
3724 
3725         /*
3726          * Get a kmem_cache structure.  We arrange that cp->cache_cpu[]
3727          * is aligned on a KMEM_CPU_CACHE_SIZE boundary to prevent
3728          * false sharing of per-CPU data.
3729          */
3730         cp = vmem_xalloc(kmem_cache_arena, csize, KMEM_CPU_CACHE_SIZE,
3731             P2NPHASE(csize, KMEM_CPU_CACHE_SIZE), 0, NULL, NULL, VM_SLEEP);
3732         bzero(cp, csize);
3733         list_link_init(&cp->cache_link);
3734 
3735         if (align == 0)
3736                 align = KMEM_ALIGN;
3737 
3738         /*
3739          * If we're not at least KMEM_ALIGN aligned, we can't use free
3740          * memory to hold bufctl information (because we can't safely
3741          * perform word loads and stores on it).
3742          */
3743         if (align < KMEM_ALIGN)
3744                 cflags |= KMC_NOTOUCH;
3745 
3746         if (!ISP2(align) || align > vmp->vm_quantum)
3747                 panic("kmem_cache_create: bad alignment %lu", align);
3748 
3749         mutex_enter(&kmem_flags_lock);
3750         if (kmem_flags & KMF_RANDOMIZE)
3751                 kmem_flags = (((kmem_flags | ~KMF_RANDOM) + 1) & KMF_RANDOM) |
3752                     KMF_RANDOMIZE;
3753         cp->cache_flags = (kmem_flags | cflags) & KMF_DEBUG;
3754         mutex_exit(&kmem_flags_lock);
3755 
3756         /*
3757          * Make sure all the various flags are reasonable.
3758          */
3759         ASSERT(!(cflags & KMC_NOHASH) || !(cflags & KMC_NOTOUCH));
3760 
3761         if (cp->cache_flags & KMF_LITE) {
3762                 if (bufsize >= kmem_lite_minsize &&
3763                     align <= kmem_lite_maxalign &&
3764                     P2PHASE(bufsize, kmem_lite_maxalign) != 0) {
3765                         cp->cache_flags |= KMF_BUFTAG;
3766                         cp->cache_flags &= ~(KMF_AUDIT | KMF_FIREWALL);
3767                 } else {
3768                         cp->cache_flags &= ~KMF_DEBUG;
3769                 }
3770         }
3771 
3772         if (cp->cache_flags & KMF_DEADBEEF)
3773                 cp->cache_flags |= KMF_REDZONE;
3774 
3775         if ((cflags & KMC_QCACHE) && (cp->cache_flags & KMF_AUDIT))
3776                 cp->cache_flags |= KMF_NOMAGAZINE;
3777 
3778         if (cflags & KMC_NODEBUG)
3779                 cp->cache_flags &= ~KMF_DEBUG;
3780 
3781         if (cflags & KMC_NOTOUCH)
3782                 cp->cache_flags &= ~KMF_TOUCH;
3783 
3784         if (cflags & KMC_PREFILL)
3785                 cp->cache_flags |= KMF_PREFILL;
3786 
3787         if (cflags & KMC_NOHASH)
3788                 cp->cache_flags &= ~(KMF_AUDIT | KMF_FIREWALL);
3789 
3790         if (cflags & KMC_NOMAGAZINE)
3791                 cp->cache_flags |= KMF_NOMAGAZINE;
3792 
3793         if ((cp->cache_flags & KMF_AUDIT) && !(cflags & KMC_NOTOUCH))
3794                 cp->cache_flags |= KMF_REDZONE;
3795 
3796         if (!(cp->cache_flags & KMF_AUDIT))
3797                 cp->cache_flags &= ~KMF_CONTENTS;
3798 
3799         if ((cp->cache_flags & KMF_BUFTAG) && bufsize >= kmem_minfirewall &&
3800             !(cp->cache_flags & KMF_LITE) && !(cflags & KMC_NOHASH))
3801                 cp->cache_flags |= KMF_FIREWALL;
3802 
3803         if (vmp != kmem_default_arena || kmem_firewall_arena == NULL)
3804                 cp->cache_flags &= ~KMF_FIREWALL;
3805 
3806         if (cp->cache_flags & KMF_FIREWALL) {
3807                 cp->cache_flags &= ~KMF_BUFTAG;
3808                 cp->cache_flags |= KMF_NOMAGAZINE;
3809                 ASSERT(vmp == kmem_default_arena);
3810                 vmp = kmem_firewall_arena;
3811         }
3812 
3813         /*
3814          * Set cache properties.
3815          */
3816         (void) strncpy(cp->cache_name, name, KMEM_CACHE_NAMELEN);
3817         strident_canon(cp->cache_name, KMEM_CACHE_NAMELEN + 1);
3818         cp->cache_bufsize = bufsize;
3819         cp->cache_align = align;
3820         cp->cache_constructor = constructor;
3821         cp->cache_destructor = destructor;
3822         cp->cache_reclaim = reclaim;
3823         cp->cache_private = private;
3824         cp->cache_arena = vmp;
3825         cp->cache_cflags = cflags;
3826 
3827         /*
3828          * Determine the chunk size.
3829          */
3830         chunksize = bufsize;
3831 
3832         if (align >= KMEM_ALIGN) {
3833                 chunksize = P2ROUNDUP(chunksize, KMEM_ALIGN);
3834                 cp->cache_bufctl = chunksize - KMEM_ALIGN;
3835         }
3836 
3837         if (cp->cache_flags & KMF_BUFTAG) {
3838                 cp->cache_bufctl = chunksize;
3839                 cp->cache_buftag = chunksize;
3840                 if (cp->cache_flags & KMF_LITE)
3841                         chunksize += KMEM_BUFTAG_LITE_SIZE(kmem_lite_count);
3842                 else
3843                         chunksize += sizeof (kmem_buftag_t);
3844         }
3845 
3846         if (cp->cache_flags & KMF_DEADBEEF) {
3847                 cp->cache_verify = MIN(cp->cache_buftag, kmem_maxverify);
3848                 if (cp->cache_flags & KMF_LITE)
3849                         cp->cache_verify = sizeof (uint64_t);
3850         }
3851 
3852         cp->cache_contents = MIN(cp->cache_bufctl, kmem_content_maxsave);
3853 
3854         cp->cache_chunksize = chunksize = P2ROUNDUP(chunksize, align);
3855 
3856         /*
3857          * Now that we know the chunk size, determine the optimal slab size.
3858          */
3859         if (vmp == kmem_firewall_arena) {
3860                 cp->cache_slabsize = P2ROUNDUP(chunksize, vmp->vm_quantum);
3861                 cp->cache_mincolor = cp->cache_slabsize - chunksize;
3862                 cp->cache_maxcolor = cp->cache_mincolor;
3863                 cp->cache_flags |= KMF_HASH;
3864                 ASSERT(!(cp->cache_flags & KMF_BUFTAG));
3865         } else if ((cflags & KMC_NOHASH) || (!(cflags & KMC_NOTOUCH) &&
3866             !(cp->cache_flags & KMF_AUDIT) &&
3867             chunksize < vmp->vm_quantum / KMEM_VOID_FRACTION)) {
3868                 cp->cache_slabsize = vmp->vm_quantum;
3869                 cp->cache_mincolor = 0;
3870                 cp->cache_maxcolor =
3871                     (cp->cache_slabsize - sizeof (kmem_slab_t)) % chunksize;
3872                 ASSERT(chunksize + sizeof (kmem_slab_t) <= cp->cache_slabsize);
3873                 ASSERT(!(cp->cache_flags & KMF_AUDIT));
3874         } else {
3875                 size_t chunks, bestfit, waste, slabsize;
3876                 size_t minwaste = LONG_MAX;
3877 
3878                 for (chunks = 1; chunks <= KMEM_VOID_FRACTION; chunks++) {
3879                         slabsize = P2ROUNDUP(chunksize * chunks,
3880                             vmp->vm_quantum);
3881                         chunks = slabsize / chunksize;
3882                         waste = (slabsize % chunksize) / chunks;
3883                         if (waste < minwaste) {
3884                                 minwaste = waste;
3885                                 bestfit = slabsize;
3886                         }
3887                 }
3888                 if (cflags & KMC_QCACHE)
3889                         bestfit = VMEM_QCACHE_SLABSIZE(vmp->vm_qcache_max);
3890                 cp->cache_slabsize = bestfit;
3891                 cp->cache_mincolor = 0;
3892                 cp->cache_maxcolor = bestfit % chunksize;
3893                 cp->cache_flags |= KMF_HASH;
3894         }
3895 
3896         cp->cache_maxchunks = (cp->cache_slabsize / cp->cache_chunksize);
3897         cp->cache_partial_binshift = highbit(cp->cache_maxchunks / 16) + 1;
3898 
3899         /*
3900          * Disallowing prefill when either the DEBUG or HASH flag is set or when
3901          * there is a constructor avoids some tricky issues with debug setup
3902          * that may be revisited later. We cannot allow prefill in a
3903          * metadata cache because of potential recursion.
3904          */
3905         if (vmp == kmem_msb_arena ||
3906             cp->cache_flags & (KMF_HASH | KMF_BUFTAG) ||
3907             cp->cache_constructor != NULL)
3908                 cp->cache_flags &= ~KMF_PREFILL;
3909 
3910         if (cp->cache_flags & KMF_HASH) {
3911                 ASSERT(!(cflags & KMC_NOHASH));
3912                 cp->cache_bufctl_cache = (cp->cache_flags & KMF_AUDIT) ?
3913                     kmem_bufctl_audit_cache : kmem_bufctl_cache;
3914         }
3915 
3916         if (cp->cache_maxcolor >= vmp->vm_quantum)
3917                 cp->cache_maxcolor = vmp->vm_quantum - 1;
3918 
3919         cp->cache_color = cp->cache_mincolor;
3920 
3921         /*
3922          * Initialize the rest of the slab layer.
3923          */
3924         mutex_init(&cp->cache_lock, NULL, MUTEX_DEFAULT, NULL);
3925 
3926         avl_create(&cp->cache_partial_slabs, kmem_partial_slab_cmp,
3927             sizeof (kmem_slab_t), offsetof(kmem_slab_t, slab_link));
3928         /* LINTED: E_TRUE_LOGICAL_EXPR */
3929         ASSERT(sizeof (list_node_t) <= sizeof (avl_node_t));
3930         /* reuse partial slab AVL linkage for complete slab list linkage */
3931         list_create(&cp->cache_complete_slabs,
3932             sizeof (kmem_slab_t), offsetof(kmem_slab_t, slab_link));
3933 
3934         if (cp->cache_flags & KMF_HASH) {
3935                 cp->cache_hash_table = vmem_alloc(kmem_hash_arena,
3936                     KMEM_HASH_INITIAL * sizeof (void *), VM_SLEEP);
3937                 bzero(cp->cache_hash_table,
3938                     KMEM_HASH_INITIAL * sizeof (void *));
3939                 cp->cache_hash_mask = KMEM_HASH_INITIAL - 1;
3940                 cp->cache_hash_shift = highbit((ulong_t)chunksize) - 1;
3941         }
3942 
3943         /*
3944          * Initialize the depot.
3945          */
3946         mutex_init(&cp->cache_depot_lock, NULL, MUTEX_DEFAULT, NULL);
3947 
3948         for (mtp = kmem_magtype; chunksize <= mtp->mt_minbuf; mtp++)
3949                 continue;
3950 
3951         cp->cache_magtype = mtp;
3952 
3953         /*
3954          * Initialize the CPU layer.
3955          */
3956         for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3957                 kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
3958                 mutex_init(&ccp->cc_lock, NULL, MUTEX_DEFAULT, NULL);
3959                 ccp->cc_flags = cp->cache_flags;
3960                 ccp->cc_rounds = -1;
3961                 ccp->cc_prounds = -1;
3962         }
3963 
3964         /*
3965          * Create the cache's kstats.
3966          */
3967         if ((cp->cache_kstat = kstat_create("unix", 0, cp->cache_name,
3968             "kmem_cache", KSTAT_TYPE_NAMED,
3969             sizeof (kmem_cache_kstat) / sizeof (kstat_named_t),
3970             KSTAT_FLAG_VIRTUAL)) != NULL) {
3971                 cp->cache_kstat->ks_data = &kmem_cache_kstat;
3972                 cp->cache_kstat->ks_update = kmem_cache_kstat_update;
3973                 cp->cache_kstat->ks_private = cp;
3974                 cp->cache_kstat->ks_lock = &kmem_cache_kstat_lock;
3975                 kstat_install(cp->cache_kstat);
3976         }
3977 
3978         /*
3979          * Add the cache to the global list.  This makes it visible
3980          * to kmem_update(), so the cache must be ready for business.
3981          */
3982         mutex_enter(&kmem_cache_lock);
3983         list_insert_tail(&kmem_caches, cp);
3984         mutex_exit(&kmem_cache_lock);
3985 
3986         if (kmem_ready)
3987                 kmem_cache_magazine_enable(cp);
3988 
3989         return (cp);
3990 }
3991 
3992 static int
3993 kmem_move_cmp(const void *buf, const void *p)
3994 {
3995         const kmem_move_t *kmm = p;
3996         uintptr_t v1 = (uintptr_t)buf;
3997         uintptr_t v2 = (uintptr_t)kmm->kmm_from_buf;
3998         return (v1 < v2 ? -1 : (v1 > v2 ? 1 : 0));
3999 }
4000 
4001 static void
4002 kmem_reset_reclaim_threshold(kmem_defrag_t *kmd)
4003 {
4004         kmd->kmd_reclaim_numer = 1;
4005 }
4006 
4007 /*
4008  * Initially, when choosing candidate slabs for buffers to move, we want to be
4009  * very selective and take only slabs that are less than
4010  * (1 / KMEM_VOID_FRACTION) allocated. If we have difficulty finding candidate
4011  * slabs, then we raise the allocation ceiling incrementally. The reclaim
4012  * threshold is reset to (1 / KMEM_VOID_FRACTION) as soon as the cache is no
4013  * longer fragmented.
4014  */
4015 static void
4016 kmem_adjust_reclaim_threshold(kmem_defrag_t *kmd, int direction)
4017 {
4018         if (direction > 0) {
4019                 /* make it easier to find a candidate slab */
4020                 if (kmd->kmd_reclaim_numer < (KMEM_VOID_FRACTION - 1)) {
4021                         kmd->kmd_reclaim_numer++;
4022                 }
4023         } else {
4024                 /* be more selective */
4025                 if (kmd->kmd_reclaim_numer > 1) {
4026                         kmd->kmd_reclaim_numer--;
4027                 }
4028         }
4029 }
4030 
4031 void
4032 kmem_cache_set_move(kmem_cache_t *cp,
4033     kmem_cbrc_t (*move)(void *, void *, size_t, void *))
4034 {
4035         kmem_defrag_t *defrag;
4036 
4037         ASSERT(move != NULL);
4038         /*
4039          * The consolidator does not support NOTOUCH caches because kmem cannot
4040          * initialize their slabs with the 0xbaddcafe memory pattern, which sets
4041          * a low order bit usable by clients to distinguish uninitialized memory
4042          * from known objects (see kmem_slab_create).
4043          */
4044         ASSERT(!(cp->cache_cflags & KMC_NOTOUCH));
4045         ASSERT(!(cp->cache_cflags & KMC_IDENTIFIER));
4046 
4047         /*
4048          * We should not be holding anyone's cache lock when calling
4049          * kmem_cache_alloc(), so allocate in all cases before acquiring the
4050          * lock.
4051          */
4052         defrag = kmem_cache_alloc(kmem_defrag_cache, KM_SLEEP);
4053 
4054         mutex_enter(&cp->cache_lock);
4055 
4056         if (KMEM_IS_MOVABLE(cp)) {
4057                 if (cp->cache_move == NULL) {
4058                         ASSERT(cp->cache_slab_alloc == 0);
4059 
4060                         cp->cache_defrag = defrag;
4061                         defrag = NULL; /* nothing to free */
4062                         bzero(cp->cache_defrag, sizeof (kmem_defrag_t));
4063                         avl_create(&cp->cache_defrag->kmd_moves_pending,
4064                             kmem_move_cmp, sizeof (kmem_move_t),
4065                             offsetof(kmem_move_t, kmm_entry));
4066                         /* LINTED: E_TRUE_LOGICAL_EXPR */
4067                         ASSERT(sizeof (list_node_t) <= sizeof (avl_node_t));
4068                         /* reuse the slab's AVL linkage for deadlist linkage */
4069                         list_create(&cp->cache_defrag->kmd_deadlist,
4070                             sizeof (kmem_slab_t),
4071                             offsetof(kmem_slab_t, slab_link));
4072                         kmem_reset_reclaim_threshold(cp->cache_defrag);
4073                 }
4074                 cp->cache_move = move;
4075         }
4076 
4077         mutex_exit(&cp->cache_lock);
4078 
4079         if (defrag != NULL) {
4080                 kmem_cache_free(kmem_defrag_cache, defrag); /* unused */
4081         }
4082 }
4083 
4084 void
4085 kmem_cache_destroy(kmem_cache_t *cp)
4086 {
4087         int cpu_seqid;
4088 
4089         /*
4090          * Remove the cache from the global cache list so that no one else
4091          * can schedule tasks on its behalf, wait for any pending tasks to
4092          * complete, purge the cache, and then destroy it.
4093          */
4094         mutex_enter(&kmem_cache_lock);
4095         list_remove(&kmem_caches, cp);
4096         mutex_exit(&kmem_cache_lock);
4097 
4098         if (kmem_taskq != NULL)
4099                 taskq_wait(kmem_taskq);
4100 
4101         if (kmem_move_taskq != NULL && cp->cache_defrag != NULL)
4102                 taskq_wait(kmem_move_taskq);
4103 
4104         kmem_cache_magazine_purge(cp);
4105 
4106         mutex_enter(&cp->cache_lock);
4107         if (cp->cache_buftotal != 0)
4108                 cmn_err(CE_WARN, "kmem_cache_destroy: '%s' (%p) not empty",
4109                     cp->cache_name, (void *)cp);
4110         if (cp->cache_defrag != NULL) {
4111                 avl_destroy(&cp->cache_defrag->kmd_moves_pending);
4112                 list_destroy(&cp->cache_defrag->kmd_deadlist);
4113                 kmem_cache_free(kmem_defrag_cache, cp->cache_defrag);
4114                 cp->cache_defrag = NULL;
4115         }
4116         /*
4117          * The cache is now dead.  There should be no further activity.  We
4118          * enforce this by setting land mines in the constructor, destructor,
4119          * reclaim, and move routines that induce a kernel text fault if
4120          * invoked.
4121          */
4122         cp->cache_constructor = (int (*)(void *, void *, int))1;
4123         cp->cache_destructor = (void (*)(void *, void *))2;
4124         cp->cache_reclaim = (void (*)(void *))3;
4125         cp->cache_move = (kmem_cbrc_t (*)(void *, void *, size_t, void *))4;
4126         mutex_exit(&cp->cache_lock);
4127 
4128         kstat_delete(cp->cache_kstat);
4129 
4130         if (cp->cache_hash_table != NULL)
4131                 vmem_free(kmem_hash_arena, cp->cache_hash_table,
4132                     (cp->cache_hash_mask + 1) * sizeof (void *));
4133 
4134         for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++)
4135                 mutex_destroy(&cp->cache_cpu[cpu_seqid].cc_lock);
4136 
4137         mutex_destroy(&cp->cache_depot_lock);
4138         mutex_destroy(&cp->cache_lock);
4139 
4140         vmem_free(kmem_cache_arena, cp, KMEM_CACHE_SIZE(max_ncpus));
4141 }
4142 
4143 /*ARGSUSED*/
4144 static int
4145 kmem_cpu_setup(cpu_setup_t what, int id, void *arg)
4146 {
4147         ASSERT(MUTEX_HELD(&cpu_lock));
4148         if (what == CPU_UNCONFIG) {
4149                 kmem_cache_applyall(kmem_cache_magazine_purge,
4150                     kmem_taskq, TQ_SLEEP);
4151                 kmem_cache_applyall(kmem_cache_magazine_enable,
4152                     kmem_taskq, TQ_SLEEP);
4153         }
4154         return (0);
4155 }
4156 
4157 static void
4158 kmem_alloc_caches_create(const int *array, size_t count,
4159     kmem_cache_t **alloc_table, size_t maxbuf, uint_t shift)
4160 {
4161         char name[KMEM_CACHE_NAMELEN + 1];
4162         size_t table_unit = (1 << shift); /* range of one alloc_table entry */
4163         size_t size = table_unit;
4164         int i;
4165 
4166         for (i = 0; i < count; i++) {
4167                 size_t cache_size = array[i];
4168                 size_t align = KMEM_ALIGN;
4169                 kmem_cache_t *cp;
4170 
4171                 /* if the table has an entry for maxbuf, we're done */
4172                 if (size > maxbuf)
4173                         break;
4174 
4175                 /* cache size must be a multiple of the table unit */
4176                 ASSERT(P2PHASE(cache_size, table_unit) == 0);
4177 
4178                 /*
4179                  * If they allocate a multiple of the coherency granularity,
4180                  * they get a coherency-granularity-aligned address.
4181                  */
4182                 if (IS_P2ALIGNED(cache_size, 64))
4183                         align = 64;
4184                 if (IS_P2ALIGNED(cache_size, PAGESIZE))
4185                         align = PAGESIZE;
4186                 (void) snprintf(name, sizeof (name),
4187                     "kmem_alloc_%lu", cache_size);
4188                 cp = kmem_cache_create(name, cache_size, align,
4189                     NULL, NULL, NULL, NULL, NULL, KMC_KMEM_ALLOC);
4190 
4191                 while (size <= cache_size) {
4192                         alloc_table[(size - 1) >> shift] = cp;
4193                         size += table_unit;
4194                 }
4195         }
4196 
4197         ASSERT(size > maxbuf);               /* i.e. maxbuf <= max(cache_size) */
4198 }
4199 
4200 static void
4201 kmem_cache_init(int pass, int use_large_pages)
4202 {
4203         int i;
4204         size_t maxbuf;
4205         kmem_magtype_t *mtp;
4206 
4207         for (i = 0; i < sizeof (kmem_magtype) / sizeof (*mtp); i++) {
4208                 char name[KMEM_CACHE_NAMELEN + 1];
4209 
4210                 mtp = &kmem_magtype[i];
4211                 (void) sprintf(name, "kmem_magazine_%d", mtp->mt_magsize);
4212                 mtp->mt_cache = kmem_cache_create(name,
4213                     (mtp->mt_magsize + 1) * sizeof (void *),
4214                     mtp->mt_align, NULL, NULL, NULL, NULL,
4215                     kmem_msb_arena, KMC_NOHASH);
4216         }
4217 
4218         kmem_slab_cache = kmem_cache_create("kmem_slab_cache",
4219             sizeof (kmem_slab_t), 0, NULL, NULL, NULL, NULL,
4220             kmem_msb_arena, KMC_NOHASH);
4221 
4222         kmem_bufctl_cache = kmem_cache_create("kmem_bufctl_cache",
4223             sizeof (kmem_bufctl_t), 0, NULL, NULL, NULL, NULL,
4224             kmem_msb_arena, KMC_NOHASH);
4225 
4226         kmem_bufctl_audit_cache = kmem_cache_create("kmem_bufctl_audit_cache",
4227             sizeof (kmem_bufctl_audit_t), 0, NULL, NULL, NULL, NULL,
4228             kmem_msb_arena, KMC_NOHASH);
4229 
4230         if (pass == 2) {
4231                 kmem_va_arena = vmem_create("kmem_va",
4232                     NULL, 0, PAGESIZE,
4233                     vmem_alloc, vmem_free, heap_arena,
4234                     8 * PAGESIZE, VM_SLEEP);
4235 
4236                 if (use_large_pages) {
4237                         kmem_default_arena = vmem_xcreate("kmem_default",
4238                             NULL, 0, PAGESIZE,
4239                             segkmem_alloc_lp, segkmem_free_lp, kmem_va_arena,
4240                             0, VMC_DUMPSAFE | VM_SLEEP);
4241                 } else {
4242                         kmem_default_arena = vmem_create("kmem_default",
4243                             NULL, 0, PAGESIZE,
4244                             segkmem_alloc, segkmem_free, kmem_va_arena,
4245                             0, VMC_DUMPSAFE | VM_SLEEP);
4246                 }
4247 
4248                 /* Figure out what our maximum cache size is */
4249                 maxbuf = kmem_max_cached;
4250                 if (maxbuf <= KMEM_MAXBUF) {
4251                         maxbuf = 0;
4252                         kmem_max_cached = KMEM_MAXBUF;
4253                 } else {
4254                         size_t size = 0;
4255                         size_t max =
4256                             sizeof (kmem_big_alloc_sizes) / sizeof (int);
4257                         /*
4258                          * Round maxbuf up to an existing cache size.  If maxbuf
4259                          * is larger than the largest cache, we truncate it to
4260                          * the largest cache's size.
4261                          */
4262                         for (i = 0; i < max; i++) {
4263                                 size = kmem_big_alloc_sizes[i];
4264                                 if (maxbuf <= size)
4265                                         break;
4266                         }
4267                         kmem_max_cached = maxbuf = size;
4268                 }
4269 
4270                 /*
4271                  * The big alloc table may not be completely overwritten, so
4272                  * we clear out any stale cache pointers from the first pass.
4273                  */
4274                 bzero(kmem_big_alloc_table, sizeof (kmem_big_alloc_table));
4275         } else {
4276                 /*
4277                  * During the first pass, the kmem_alloc_* caches
4278                  * are treated as metadata.
4279                  */
4280                 kmem_default_arena = kmem_msb_arena;
4281                 maxbuf = KMEM_BIG_MAXBUF_32BIT;
4282         }
4283 
4284         /*
4285          * Set up the default caches to back kmem_alloc()
4286          */
4287         kmem_alloc_caches_create(
4288             kmem_alloc_sizes, sizeof (kmem_alloc_sizes) / sizeof (int),
4289             kmem_alloc_table, KMEM_MAXBUF, KMEM_ALIGN_SHIFT);
4290 
4291         kmem_alloc_caches_create(
4292             kmem_big_alloc_sizes, sizeof (kmem_big_alloc_sizes) / sizeof (int),
4293             kmem_big_alloc_table, maxbuf, KMEM_BIG_SHIFT);
4294 
4295         kmem_big_alloc_table_max = maxbuf >> KMEM_BIG_SHIFT;
4296 }
4297 
4298 void
4299 kmem_init(void)
4300 {
4301         kmem_cache_t *cp;
4302         int old_kmem_flags = kmem_flags;
4303         int use_large_pages = 0;
4304         size_t maxverify, minfirewall;
4305 
4306         kstat_init();
4307 
4308         /*
4309          * Don't do firewalled allocations if the heap is less than 1TB
4310          * (i.e. on a 32-bit kernel)
4311          * The resulting VM_NEXTFIT allocations would create too much
4312          * fragmentation in a small heap.
4313          */
4314 #if defined(_LP64)
4315         maxverify = minfirewall = PAGESIZE / 2;
4316 #else
4317         maxverify = minfirewall = ULONG_MAX;
4318 #endif
4319 
4320         /* LINTED */
4321         ASSERT(sizeof (kmem_cpu_cache_t) == KMEM_CPU_CACHE_SIZE);
4322 
4323         list_create(&kmem_caches, sizeof (kmem_cache_t),
4324             offsetof(kmem_cache_t, cache_link));
4325 
4326         kmem_metadata_arena = vmem_create("kmem_metadata", NULL, 0, PAGESIZE,
4327             vmem_alloc, vmem_free, heap_arena, 8 * PAGESIZE,
4328             VM_SLEEP | VMC_NO_QCACHE);
4329 
4330         kmem_msb_arena = vmem_create("kmem_msb", NULL, 0,
4331             PAGESIZE, segkmem_alloc, segkmem_free, kmem_metadata_arena, 0,
4332             VMC_DUMPSAFE | VM_SLEEP);
4333 
4334         kmem_cache_arena = vmem_create("kmem_cache", NULL, 0, KMEM_ALIGN,
4335             segkmem_alloc, segkmem_free, kmem_metadata_arena, 0, VM_SLEEP);
4336 
4337         kmem_hash_arena = vmem_create("kmem_hash", NULL, 0, KMEM_ALIGN,
4338             segkmem_alloc, segkmem_free, kmem_metadata_arena, 0, VM_SLEEP);
4339 
4340         kmem_log_arena = vmem_create("kmem_log", NULL, 0, KMEM_ALIGN,
4341             segkmem_alloc, segkmem_free, heap_arena, 0, VM_SLEEP);
4342 
4343         kmem_firewall_va_arena = vmem_create("kmem_firewall_va",
4344             NULL, 0, PAGESIZE,
4345             kmem_firewall_va_alloc, kmem_firewall_va_free, heap_arena,
4346             0, VM_SLEEP);
4347 
4348         kmem_firewall_arena = vmem_create("kmem_firewall", NULL, 0, PAGESIZE,
4349             segkmem_alloc, segkmem_free, kmem_firewall_va_arena, 0,
4350             VMC_DUMPSAFE | VM_SLEEP);
4351 
4352         /* temporary oversize arena for mod_read_system_file */
4353         kmem_oversize_arena = vmem_create("kmem_oversize", NULL, 0, PAGESIZE,
4354             segkmem_alloc, segkmem_free, heap_arena, 0, VM_SLEEP);
4355 
4356         kmem_reap_interval = 15 * hz;
4357 
4358         /*
4359          * Read /etc/system.  This is a chicken-and-egg problem because
4360          * kmem_flags may be set in /etc/system, but mod_read_system_file()
4361          * needs to use the allocator.  The simplest solution is to create
4362          * all the standard kmem caches, read /etc/system, destroy all the
4363          * caches we just created, and then create them all again in light
4364          * of the (possibly) new kmem_flags and other kmem tunables.
4365          */
4366         kmem_cache_init(1, 0);
4367 
4368         mod_read_system_file(boothowto & RB_ASKNAME);
4369 
4370         while ((cp = list_tail(&kmem_caches)) != NULL)
4371                 kmem_cache_destroy(cp);
4372 
4373         vmem_destroy(kmem_oversize_arena);
4374 
4375         if (old_kmem_flags & KMF_STICKY)
4376                 kmem_flags = old_kmem_flags;
4377 
4378         if (!(kmem_flags & KMF_AUDIT))
4379                 vmem_seg_size = offsetof(vmem_seg_t, vs_thread);
4380 
4381         if (kmem_maxverify == 0)
4382                 kmem_maxverify = maxverify;
4383 
4384         if (kmem_minfirewall == 0)
4385                 kmem_minfirewall = minfirewall;
4386 
4387         /*
4388          * give segkmem a chance to figure out if we are using large pages
4389          * for the kernel heap
4390          */
4391         use_large_pages = segkmem_lpsetup();
4392 
4393         /*
4394          * To protect against corruption, we keep the actual number of callers
4395          * KMF_LITE records seperate from the tunable.  We arbitrarily clamp
4396          * to 16, since the overhead for small buffers quickly gets out of
4397          * hand.
4398          *
4399          * The real limit would depend on the needs of the largest KMC_NOHASH
4400          * cache.
4401          */
4402         kmem_lite_count = MIN(MAX(0, kmem_lite_pcs), 16);
4403         kmem_lite_pcs = kmem_lite_count;
4404 
4405         /*
4406          * Normally, we firewall oversized allocations when possible, but
4407          * if we are using large pages for kernel memory, and we don't have
4408          * any non-LITE debugging flags set, we want to allocate oversized
4409          * buffers from large pages, and so skip the firewalling.
4410          */
4411         if (use_large_pages &&
4412             ((kmem_flags & KMF_LITE) || !(kmem_flags & KMF_DEBUG))) {
4413                 kmem_oversize_arena = vmem_xcreate("kmem_oversize", NULL, 0,
4414                     PAGESIZE, segkmem_alloc_lp, segkmem_free_lp, heap_arena,
4415                     0, VMC_DUMPSAFE | VM_SLEEP);
4416         } else {
4417                 kmem_oversize_arena = vmem_create("kmem_oversize",
4418                     NULL, 0, PAGESIZE,
4419                     segkmem_alloc, segkmem_free, kmem_minfirewall < ULONG_MAX?
4420                     kmem_firewall_va_arena : heap_arena, 0, VMC_DUMPSAFE |
4421                     VM_SLEEP);
4422         }
4423 
4424         kmem_cache_init(2, use_large_pages);
4425 
4426         if (kmem_flags & (KMF_AUDIT | KMF_RANDOMIZE)) {
4427                 if (kmem_transaction_log_size == 0)
4428                         kmem_transaction_log_size = kmem_maxavail() / 50;
4429                 kmem_transaction_log = kmem_log_init(kmem_transaction_log_size);
4430         }
4431 
4432         if (kmem_flags & (KMF_CONTENTS | KMF_RANDOMIZE)) {
4433                 if (kmem_content_log_size == 0)
4434                         kmem_content_log_size = kmem_maxavail() / 50;
4435                 kmem_content_log = kmem_log_init(kmem_content_log_size);
4436         }
4437 
4438         kmem_failure_log = kmem_log_init(kmem_failure_log_size);
4439 
4440         kmem_slab_log = kmem_log_init(kmem_slab_log_size);
4441 
4442         /*
4443          * Initialize STREAMS message caches so allocb() is available.
4444          * This allows us to initialize the logging framework (cmn_err(9F),
4445          * strlog(9F), etc) so we can start recording messages.
4446          */
4447         streams_msg_init();
4448 
4449         /*
4450          * Initialize the ZSD framework in Zones so modules loaded henceforth
4451          * can register their callbacks.
4452          */
4453         zone_zsd_init();
4454 
4455         log_init();
4456         taskq_init();
4457 
4458         /*
4459          * Warn about invalid or dangerous values of kmem_flags.
4460          * Always warn about unsupported values.
4461          */
4462         if (((kmem_flags & ~(KMF_AUDIT | KMF_DEADBEEF | KMF_REDZONE |
4463             KMF_CONTENTS | KMF_LITE)) != 0) ||
4464             ((kmem_flags & KMF_LITE) && kmem_flags != KMF_LITE))
4465                 cmn_err(CE_WARN, "kmem_flags set to unsupported value 0x%x. "
4466                     "See the Solaris Tunable Parameters Reference Manual.",
4467                     kmem_flags);
4468 
4469 #ifdef DEBUG
4470         if ((kmem_flags & KMF_DEBUG) == 0)
4471                 cmn_err(CE_NOTE, "kmem debugging disabled.");
4472 #else
4473         /*
4474          * For non-debug kernels, the only "normal" flags are 0, KMF_LITE,
4475          * KMF_REDZONE, and KMF_CONTENTS (the last because it is only enabled
4476          * if KMF_AUDIT is set). We should warn the user about the performance
4477          * penalty of KMF_AUDIT or KMF_DEADBEEF if they are set and KMF_LITE
4478          * isn't set (since that disables AUDIT).
4479          */
4480         if (!(kmem_flags & KMF_LITE) &&
4481             (kmem_flags & (KMF_AUDIT | KMF_DEADBEEF)) != 0)
4482                 cmn_err(CE_WARN, "High-overhead kmem debugging features "
4483                     "enabled (kmem_flags = 0x%x).  Performance degradation "
4484                     "and large memory overhead possible. See the Solaris "
4485                     "Tunable Parameters Reference Manual.", kmem_flags);
4486 #endif /* not DEBUG */
4487 
4488         kmem_cache_applyall(kmem_cache_magazine_enable, NULL, TQ_SLEEP);
4489 
4490         kmem_ready = 1;
4491 
4492         /*
4493          * Initialize the platform-specific aligned/DMA memory allocator.
4494          */
4495         ka_init();
4496 
4497         /*
4498          * Initialize 32-bit ID cache.
4499          */
4500         id32_init();
4501 
4502         /*
4503          * Initialize the networking stack so modules loaded can
4504          * register their callbacks.
4505          */
4506         netstack_init();
4507 }
4508 
4509 static void
4510 kmem_move_init(void)
4511 {
4512         kmem_defrag_cache = kmem_cache_create("kmem_defrag_cache",
4513             sizeof (kmem_defrag_t), 0, NULL, NULL, NULL, NULL,
4514             kmem_msb_arena, KMC_NOHASH);
4515         kmem_move_cache = kmem_cache_create("kmem_move_cache",
4516             sizeof (kmem_move_t), 0, NULL, NULL, NULL, NULL,
4517             kmem_msb_arena, KMC_NOHASH);
4518 
4519         /*
4520          * kmem guarantees that move callbacks are sequential and that even
4521          * across multiple caches no two moves ever execute simultaneously.
4522          * Move callbacks are processed on a separate taskq so that client code
4523          * does not interfere with internal maintenance tasks.
4524          */
4525         kmem_move_taskq = taskq_create_instance("kmem_move_taskq", 0, 1,
4526             minclsyspri, 100, INT_MAX, TASKQ_PREPOPULATE);
4527 }
4528 
4529 void
4530 kmem_thread_init(void)
4531 {
4532         kmem_move_init();
4533         kmem_taskq = taskq_create_instance("kmem_taskq", 0, 1, minclsyspri,
4534             300, INT_MAX, TASKQ_PREPOPULATE);
4535 }
4536 
4537 void
4538 kmem_mp_init(void)
4539 {
4540         mutex_enter(&cpu_lock);
4541         register_cpu_setup_func(kmem_cpu_setup, NULL);
4542         mutex_exit(&cpu_lock);
4543 
4544         kmem_update_timeout(NULL);
4545 
4546         taskq_mp_init();
4547 }
4548 
4549 /*
4550  * Return the slab of the allocated buffer, or NULL if the buffer is not
4551  * allocated. This function may be called with a known slab address to determine
4552  * whether or not the buffer is allocated, or with a NULL slab address to obtain
4553  * an allocated buffer's slab.
4554  */
4555 static kmem_slab_t *
4556 kmem_slab_allocated(kmem_cache_t *cp, kmem_slab_t *sp, void *buf)
4557 {
4558         kmem_bufctl_t *bcp, *bufbcp;
4559 
4560         ASSERT(MUTEX_HELD(&cp->cache_lock));
4561         ASSERT(sp == NULL || KMEM_SLAB_MEMBER(sp, buf));
4562 
4563         if (cp->cache_flags & KMF_HASH) {
4564                 for (bcp = *KMEM_HASH(cp, buf);
4565                     (bcp != NULL) && (bcp->bc_addr != buf);
4566                     bcp = bcp->bc_next) {
4567                         continue;
4568                 }
4569                 ASSERT(sp != NULL && bcp != NULL ? sp == bcp->bc_slab : 1);
4570                 return (bcp == NULL ? NULL : bcp->bc_slab);
4571         }
4572 
4573         if (sp == NULL) {
4574                 sp = KMEM_SLAB(cp, buf);
4575         }
4576         bufbcp = KMEM_BUFCTL(cp, buf);
4577         for (bcp = sp->slab_head;
4578             (bcp != NULL) && (bcp != bufbcp);
4579             bcp = bcp->bc_next) {
4580                 continue;
4581         }
4582         return (bcp == NULL ? sp : NULL);
4583 }
4584 
4585 static boolean_t
4586 kmem_slab_is_reclaimable(kmem_cache_t *cp, kmem_slab_t *sp, int flags)
4587 {
4588         long refcnt = sp->slab_refcnt;
4589 
4590         ASSERT(cp->cache_defrag != NULL);
4591 
4592         /*
4593          * For code coverage we want to be able to move an object within the
4594          * same slab (the only partial slab) even if allocating the destination
4595          * buffer resulted in a completely allocated slab.
4596          */
4597         if (flags & KMM_DEBUG) {
4598                 return ((flags & KMM_DESPERATE) ||
4599                     ((sp->slab_flags & KMEM_SLAB_NOMOVE) == 0));
4600         }
4601 
4602         /* If we're desperate, we don't care if the client said NO. */
4603         if (flags & KMM_DESPERATE) {
4604                 return (refcnt < sp->slab_chunks); /* any partial */
4605         }
4606 
4607         if (sp->slab_flags & KMEM_SLAB_NOMOVE) {
4608                 return (B_FALSE);
4609         }
4610 
4611         if ((refcnt == 1) || kmem_move_any_partial) {
4612                 return (refcnt < sp->slab_chunks);
4613         }
4614 
4615         /*
4616          * The reclaim threshold is adjusted at each kmem_cache_scan() so that
4617          * slabs with a progressively higher percentage of used buffers can be
4618          * reclaimed until the cache as a whole is no longer fragmented.
4619          *
4620          *      sp->slab_refcnt   kmd_reclaim_numer
4621          *      --------------- < ------------------
4622          *      sp->slab_chunks   KMEM_VOID_FRACTION
4623          */
4624         return ((refcnt * KMEM_VOID_FRACTION) <
4625             (sp->slab_chunks * cp->cache_defrag->kmd_reclaim_numer));
4626 }
4627 
4628 /*
4629  * May be called from the kmem_move_taskq, from kmem_cache_move_notify_task(),
4630  * or when the buffer is freed.
4631  */
4632 static void
4633 kmem_slab_move_yes(kmem_cache_t *cp, kmem_slab_t *sp, void *from_buf)
4634 {
4635         ASSERT(MUTEX_HELD(&cp->cache_lock));
4636         ASSERT(KMEM_SLAB_MEMBER(sp, from_buf));
4637 
4638         if (!KMEM_SLAB_IS_PARTIAL(sp)) {
4639                 return;
4640         }
4641 
4642         if (sp->slab_flags & KMEM_SLAB_NOMOVE) {
4643                 if (KMEM_SLAB_OFFSET(sp, from_buf) == sp->slab_stuck_offset) {
4644                         avl_remove(&cp->cache_partial_slabs, sp);
4645                         sp->slab_flags &= ~KMEM_SLAB_NOMOVE;
4646                         sp->slab_stuck_offset = (uint32_t)-1;
4647                         avl_add(&cp->cache_partial_slabs, sp);
4648                 }
4649         } else {
4650                 sp->slab_later_count = 0;
4651                 sp->slab_stuck_offset = (uint32_t)-1;
4652         }
4653 }
4654 
4655 static void
4656 kmem_slab_move_no(kmem_cache_t *cp, kmem_slab_t *sp, void *from_buf)
4657 {
4658         ASSERT(taskq_member(kmem_move_taskq, curthread));
4659         ASSERT(MUTEX_HELD(&cp->cache_lock));
4660         ASSERT(KMEM_SLAB_MEMBER(sp, from_buf));
4661 
4662         if (!KMEM_SLAB_IS_PARTIAL(sp)) {
4663                 return;
4664         }
4665 
4666         avl_remove(&cp->cache_partial_slabs, sp);
4667         sp->slab_later_count = 0;
4668         sp->slab_flags |= KMEM_SLAB_NOMOVE;
4669         sp->slab_stuck_offset = KMEM_SLAB_OFFSET(sp, from_buf);
4670         avl_add(&cp->cache_partial_slabs, sp);
4671 }
4672 
4673 static void kmem_move_end(kmem_cache_t *, kmem_move_t *);
4674 
4675 /*
4676  * The move callback takes two buffer addresses, the buffer to be moved, and a
4677  * newly allocated and constructed buffer selected by kmem as the destination.
4678  * It also takes the size of the buffer and an optional user argument specified
4679  * at cache creation time. kmem guarantees that the buffer to be moved has not
4680  * been unmapped by the virtual memory subsystem. Beyond that, it cannot
4681  * guarantee the present whereabouts of the buffer to be moved, so it is up to
4682  * the client to safely determine whether or not it is still using the buffer.
4683  * The client must not free either of the buffers passed to the move callback,
4684  * since kmem wants to free them directly to the slab layer. The client response
4685  * tells kmem which of the two buffers to free:
4686  *
4687  * YES          kmem frees the old buffer (the move was successful)
4688  * NO           kmem frees the new buffer, marks the slab of the old buffer
4689  *              non-reclaimable to avoid bothering the client again
4690  * LATER        kmem frees the new buffer, increments slab_later_count
4691  * DONT_KNOW    kmem frees the new buffer
4692  * DONT_NEED    kmem frees both the old buffer and the new buffer
4693  *
4694  * The pending callback argument now being processed contains both of the
4695  * buffers (old and new) passed to the move callback function, the slab of the
4696  * old buffer, and flags related to the move request, such as whether or not the
4697  * system was desperate for memory.
4698  *
4699  * Slabs are not freed while there is a pending callback, but instead are kept
4700  * on a deadlist, which is drained after the last callback completes. This means
4701  * that slabs are safe to access until kmem_move_end(), no matter how many of
4702  * their buffers have been freed. Once slab_refcnt reaches zero, it stays at
4703  * zero for as long as the slab remains on the deadlist and until the slab is
4704  * freed.
4705  */
4706 static void
4707 kmem_move_buffer(kmem_move_t *callback)
4708 {
4709         kmem_cbrc_t response;
4710         kmem_slab_t *sp = callback->kmm_from_slab;
4711         kmem_cache_t *cp = sp->slab_cache;
4712         boolean_t free_on_slab;
4713 
4714         ASSERT(taskq_member(kmem_move_taskq, curthread));
4715         ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
4716         ASSERT(KMEM_SLAB_MEMBER(sp, callback->kmm_from_buf));
4717 
4718         /*
4719          * The number of allocated buffers on the slab may have changed since we
4720          * last checked the slab's reclaimability (when the pending move was
4721          * enqueued), or the client may have responded NO when asked to move
4722          * another buffer on the same slab.
4723          */
4724         if (!kmem_slab_is_reclaimable(cp, sp, callback->kmm_flags)) {
4725                 kmem_slab_free(cp, callback->kmm_to_buf);
4726                 kmem_move_end(cp, callback);
4727                 return;
4728         }
4729 
4730         /*
4731          * Checking the slab layer is easy, so we might as well do that here
4732          * in case we can avoid bothering the client.
4733          */
4734         mutex_enter(&cp->cache_lock);
4735         free_on_slab = (kmem_slab_allocated(cp, sp,
4736             callback->kmm_from_buf) == NULL);
4737         mutex_exit(&cp->cache_lock);
4738 
4739         if (free_on_slab) {
4740                 kmem_slab_free(cp, callback->kmm_to_buf);
4741                 kmem_move_end(cp, callback);
4742                 return;
4743         }
4744 
4745         if (cp->cache_flags & KMF_BUFTAG) {
4746                 /*
4747                  * Make kmem_cache_alloc_debug() apply the constructor for us.
4748                  */
4749                 if (kmem_cache_alloc_debug(cp, callback->kmm_to_buf,
4750                     KM_NOSLEEP, 1, caller()) != 0) {
4751                         kmem_move_end(cp, callback);
4752                         return;
4753                 }
4754         } else if (cp->cache_constructor != NULL &&
4755             cp->cache_constructor(callback->kmm_to_buf, cp->cache_private,
4756             KM_NOSLEEP) != 0) {
4757                 atomic_inc_64(&cp->cache_alloc_fail);
4758                 kmem_slab_free(cp, callback->kmm_to_buf);
4759                 kmem_move_end(cp, callback);
4760                 return;
4761         }
4762 
4763         cp->cache_defrag->kmd_callbacks++;
4764         cp->cache_defrag->kmd_thread = curthread;
4765         cp->cache_defrag->kmd_from_buf = callback->kmm_from_buf;
4766         cp->cache_defrag->kmd_to_buf = callback->kmm_to_buf;
4767         DTRACE_PROBE2(kmem__move__start, kmem_cache_t *, cp, kmem_move_t *,
4768             callback);
4769 
4770         response = cp->cache_move(callback->kmm_from_buf,
4771             callback->kmm_to_buf, cp->cache_bufsize, cp->cache_private);
4772 
4773         DTRACE_PROBE3(kmem__move__end, kmem_cache_t *, cp, kmem_move_t *,
4774             callback, kmem_cbrc_t, response);
4775         cp->cache_defrag->kmd_thread = NULL;
4776         cp->cache_defrag->kmd_from_buf = NULL;
4777         cp->cache_defrag->kmd_to_buf = NULL;
4778 
4779         if (response == KMEM_CBRC_YES) {
4780                 cp->cache_defrag->kmd_yes++;
4781                 kmem_slab_free_constructed(cp, callback->kmm_from_buf, B_FALSE);
4782                 /* slab safe to access until kmem_move_end() */
4783                 if (sp->slab_refcnt == 0)
4784                         cp->cache_defrag->kmd_slabs_freed++;
4785                 mutex_enter(&cp->cache_lock);
4786                 kmem_slab_move_yes(cp, sp, callback->kmm_from_buf);
4787                 mutex_exit(&cp->cache_lock);
4788                 kmem_move_end(cp, callback);
4789                 return;
4790         }
4791 
4792         switch (response) {
4793         case KMEM_CBRC_NO:
4794                 cp->cache_defrag->kmd_no++;
4795                 mutex_enter(&cp->cache_lock);
4796                 kmem_slab_move_no(cp, sp, callback->kmm_from_buf);
4797                 mutex_exit(&cp->cache_lock);
4798                 break;
4799         case KMEM_CBRC_LATER:
4800                 cp->cache_defrag->kmd_later++;
4801                 mutex_enter(&cp->cache_lock);
4802                 if (!KMEM_SLAB_IS_PARTIAL(sp)) {
4803                         mutex_exit(&cp->cache_lock);
4804                         break;
4805                 }
4806 
4807                 if (++sp->slab_later_count >= KMEM_DISBELIEF) {
4808                         kmem_slab_move_no(cp, sp, callback->kmm_from_buf);
4809                 } else if (!(sp->slab_flags & KMEM_SLAB_NOMOVE)) {
4810                         sp->slab_stuck_offset = KMEM_SLAB_OFFSET(sp,
4811                             callback->kmm_from_buf);
4812                 }
4813                 mutex_exit(&cp->cache_lock);
4814                 break;
4815         case KMEM_CBRC_DONT_NEED:
4816                 cp->cache_defrag->kmd_dont_need++;
4817                 kmem_slab_free_constructed(cp, callback->kmm_from_buf, B_FALSE);
4818                 if (sp->slab_refcnt == 0)
4819                         cp->cache_defrag->kmd_slabs_freed++;
4820                 mutex_enter(&cp->cache_lock);
4821                 kmem_slab_move_yes(cp, sp, callback->kmm_from_buf);
4822                 mutex_exit(&cp->cache_lock);
4823                 break;
4824         case KMEM_CBRC_DONT_KNOW:
4825                 /*
4826                  * If we don't know if we can move this buffer or not, we'll
4827                  * just assume that we can't:  if the buffer is in fact free,
4828                  * then it is sitting in one of the per-CPU magazines or in
4829                  * a full magazine in the depot layer.  Either way, because
4830                  * defrag is induced in the same logic that reaps a cache,
4831                  * it's likely that full magazines will be returned to the
4832                  * system soon (thereby accomplishing what we're trying to
4833                  * accomplish here: return those magazines to their slabs).
4834                  * Given this, any work that we might do now to locate a buffer
4835                  * in a magazine is wasted (and expensive!) work; we bump
4836                  * a counter in this case and otherwise assume that we can't
4837                  * move it.
4838                  */
4839                 cp->cache_defrag->kmd_dont_know++;
4840                 break;
4841         default:
4842                 panic("'%s' (%p) unexpected move callback response %d\n",
4843                     cp->cache_name, (void *)cp, response);
4844         }
4845 
4846         kmem_slab_free_constructed(cp, callback->kmm_to_buf, B_FALSE);
4847         kmem_move_end(cp, callback);
4848 }
4849 
4850 /* Return B_FALSE if there is insufficient memory for the move request. */
4851 static boolean_t
4852 kmem_move_begin(kmem_cache_t *cp, kmem_slab_t *sp, void *buf, int flags)
4853 {
4854         void *to_buf;
4855         avl_index_t index;
4856         kmem_move_t *callback, *pending;
4857         ulong_t n;
4858 
4859         ASSERT(taskq_member(kmem_taskq, curthread));
4860         ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
4861         ASSERT(sp->slab_flags & KMEM_SLAB_MOVE_PENDING);
4862 
4863         callback = kmem_cache_alloc(kmem_move_cache, KM_NOSLEEP);
4864 
4865         if (callback == NULL)
4866                 return (B_FALSE);
4867 
4868         callback->kmm_from_slab = sp;
4869         callback->kmm_from_buf = buf;
4870         callback->kmm_flags = flags;
4871 
4872         mutex_enter(&cp->cache_lock);
4873 
4874         n = avl_numnodes(&cp->cache_partial_slabs);
4875         if ((n == 0) || ((n == 1) && !(flags & KMM_DEBUG))) {
4876                 mutex_exit(&cp->cache_lock);
4877                 kmem_cache_free(kmem_move_cache, callback);
4878                 return (B_TRUE); /* there is no need for the move request */
4879         }
4880 
4881         pending = avl_find(&cp->cache_defrag->kmd_moves_pending, buf, &index);
4882         if (pending != NULL) {
4883                 /*
4884                  * If the move is already pending and we're desperate now,
4885                  * update the move flags.
4886                  */
4887                 if (flags & KMM_DESPERATE) {
4888                         pending->kmm_flags |= KMM_DESPERATE;
4889                 }
4890                 mutex_exit(&cp->cache_lock);
4891                 kmem_cache_free(kmem_move_cache, callback);
4892                 return (B_TRUE);
4893         }
4894 
4895         to_buf = kmem_slab_alloc_impl(cp, avl_first(&cp->cache_partial_slabs),
4896             B_FALSE);
4897         callback->kmm_to_buf = to_buf;
4898         avl_insert(&cp->cache_defrag->kmd_moves_pending, callback, index);
4899 
4900         mutex_exit(&cp->cache_lock);
4901 
4902         if (!taskq_dispatch(kmem_move_taskq, (task_func_t *)kmem_move_buffer,
4903             callback, TQ_NOSLEEP)) {
4904                 mutex_enter(&cp->cache_lock);
4905                 avl_remove(&cp->cache_defrag->kmd_moves_pending, callback);
4906                 mutex_exit(&cp->cache_lock);
4907                 kmem_slab_free(cp, to_buf);
4908                 kmem_cache_free(kmem_move_cache, callback);
4909                 return (B_FALSE);
4910         }
4911 
4912         return (B_TRUE);
4913 }
4914 
4915 static void
4916 kmem_move_end(kmem_cache_t *cp, kmem_move_t *callback)
4917 {
4918         avl_index_t index;
4919 
4920         ASSERT(cp->cache_defrag != NULL);
4921         ASSERT(taskq_member(kmem_move_taskq, curthread));
4922         ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
4923 
4924         mutex_enter(&cp->cache_lock);
4925         VERIFY(avl_find(&cp->cache_defrag->kmd_moves_pending,
4926             callback->kmm_from_buf, &index) != NULL);
4927         avl_remove(&cp->cache_defrag->kmd_moves_pending, callback);
4928         if (avl_is_empty(&cp->cache_defrag->kmd_moves_pending)) {
4929                 list_t *deadlist = &cp->cache_defrag->kmd_deadlist;
4930                 kmem_slab_t *sp;
4931 
4932                 /*
4933                  * The last pending move completed. Release all slabs from the
4934                  * front of the dead list except for any slab at the tail that
4935                  * needs to be released from the context of kmem_move_buffers().
4936                  * kmem deferred unmapping the buffers on these slabs in order
4937                  * to guarantee that buffers passed to the move callback have
4938                  * been touched only by kmem or by the client itself.
4939                  */
4940                 while ((sp = list_remove_head(deadlist)) != NULL) {
4941                         if (sp->slab_flags & KMEM_SLAB_MOVE_PENDING) {
4942                                 list_insert_tail(deadlist, sp);
4943                                 break;
4944                         }
4945                         cp->cache_defrag->kmd_deadcount--;
4946                         cp->cache_slab_destroy++;
4947                         mutex_exit(&cp->cache_lock);
4948                         kmem_slab_destroy(cp, sp);
4949                         mutex_enter(&cp->cache_lock);
4950                 }
4951         }
4952         mutex_exit(&cp->cache_lock);
4953         kmem_cache_free(kmem_move_cache, callback);
4954 }
4955 
4956 /*
4957  * Move buffers from least used slabs first by scanning backwards from the end
4958  * of the partial slab list. Scan at most max_scan candidate slabs and move
4959  * buffers from at most max_slabs slabs (0 for all partial slabs in both cases).
4960  * If desperate to reclaim memory, move buffers from any partial slab, otherwise
4961  * skip slabs with a ratio of allocated buffers at or above the current
4962  * threshold. Return the number of unskipped slabs (at most max_slabs, -1 if the
4963  * scan is aborted) so that the caller can adjust the reclaimability threshold
4964  * depending on how many reclaimable slabs it finds.
4965  *
4966  * kmem_move_buffers() drops and reacquires cache_lock every time it issues a
4967  * move request, since it is not valid for kmem_move_begin() to call
4968  * kmem_cache_alloc() or taskq_dispatch() with cache_lock held.
4969  */
4970 static int
4971 kmem_move_buffers(kmem_cache_t *cp, size_t max_scan, size_t max_slabs,
4972     int flags)
4973 {
4974         kmem_slab_t *sp;
4975         void *buf;
4976         int i, j; /* slab index, buffer index */
4977         int s; /* reclaimable slabs */
4978         int b; /* allocated (movable) buffers on reclaimable slab */
4979         boolean_t success;
4980         int refcnt;
4981         int nomove;
4982 
4983         ASSERT(taskq_member(kmem_taskq, curthread));
4984         ASSERT(MUTEX_HELD(&cp->cache_lock));
4985         ASSERT(kmem_move_cache != NULL);
4986         ASSERT(cp->cache_move != NULL && cp->cache_defrag != NULL);
4987         ASSERT((flags & KMM_DEBUG) ? !avl_is_empty(&cp->cache_partial_slabs) :
4988             avl_numnodes(&cp->cache_partial_slabs) > 1);
4989 
4990         if (kmem_move_blocked) {
4991                 return (0);
4992         }
4993 
4994         if (kmem_move_fulltilt) {
4995                 flags |= KMM_DESPERATE;
4996         }
4997 
4998         if (max_scan == 0 || (flags & KMM_DESPERATE)) {
4999                 /*
5000                  * Scan as many slabs as needed to find the desired number of
5001                  * candidate slabs.
5002                  */
5003                 max_scan = (size_t)-1;
5004         }
5005 
5006         if (max_slabs == 0 || (flags & KMM_DESPERATE)) {
5007                 /* Find as many candidate slabs as possible. */
5008                 max_slabs = (size_t)-1;
5009         }
5010 
5011         sp = avl_last(&cp->cache_partial_slabs);
5012         ASSERT(KMEM_SLAB_IS_PARTIAL(sp));
5013         for (i = 0, s = 0; (i < max_scan) && (s < max_slabs) && (sp != NULL) &&
5014             ((sp != avl_first(&cp->cache_partial_slabs)) ||
5015             (flags & KMM_DEBUG));
5016             sp = AVL_PREV(&cp->cache_partial_slabs, sp), i++) {
5017 
5018                 if (!kmem_slab_is_reclaimable(cp, sp, flags)) {
5019                         continue;
5020                 }
5021                 s++;
5022 
5023                 /* Look for allocated buffers to move. */
5024                 for (j = 0, b = 0, buf = sp->slab_base;
5025                     (j < sp->slab_chunks) && (b < sp->slab_refcnt);
5026                     buf = (((char *)buf) + cp->cache_chunksize), j++) {
5027 
5028                         if (kmem_slab_allocated(cp, sp, buf) == NULL) {
5029                                 continue;
5030                         }
5031 
5032                         b++;
5033 
5034                         /*
5035                          * Prevent the slab from being destroyed while we drop
5036                          * cache_lock and while the pending move is not yet
5037                          * registered. Flag the pending move while
5038                          * kmd_moves_pending may still be empty, since we can't
5039                          * yet rely on a non-zero pending move count to prevent
5040                          * the slab from being destroyed.
5041                          */
5042                         ASSERT(!(sp->slab_flags & KMEM_SLAB_MOVE_PENDING));
5043                         sp->slab_flags |= KMEM_SLAB_MOVE_PENDING;
5044                         /*
5045                          * Recheck refcnt and nomove after reacquiring the lock,
5046                          * since these control the order of partial slabs, and
5047                          * we want to know if we can pick up the scan where we
5048                          * left off.
5049                          */
5050                         refcnt = sp->slab_refcnt;
5051                         nomove = (sp->slab_flags & KMEM_SLAB_NOMOVE);
5052                         mutex_exit(&cp->cache_lock);
5053 
5054                         success = kmem_move_begin(cp, sp, buf, flags);
5055 
5056                         /*
5057                          * Now, before the lock is reacquired, kmem could
5058                          * process all pending move requests and purge the
5059                          * deadlist, so that upon reacquiring the lock, sp has
5060                          * been remapped. Or, the client may free all the
5061                          * objects on the slab while the pending moves are still
5062                          * on the taskq. Therefore, the KMEM_SLAB_MOVE_PENDING
5063                          * flag causes the slab to be put at the end of the
5064                          * deadlist and prevents it from being destroyed, since
5065                          * we plan to destroy it here after reacquiring the
5066                          * lock.
5067                          */
5068                         mutex_enter(&cp->cache_lock);
5069                         ASSERT(sp->slab_flags & KMEM_SLAB_MOVE_PENDING);
5070                         sp->slab_flags &= ~KMEM_SLAB_MOVE_PENDING;
5071 
5072                         if (sp->slab_refcnt == 0) {
5073                                 list_t *deadlist =
5074                                     &cp->cache_defrag->kmd_deadlist;
5075                                 list_remove(deadlist, sp);
5076 
5077                                 if (!avl_is_empty(
5078                                     &cp->cache_defrag->kmd_moves_pending)) {
5079                                         /*
5080                                          * A pending move makes it unsafe to
5081                                          * destroy the slab, because even though
5082                                          * the move is no longer needed, the
5083                                          * context where that is determined
5084                                          * requires the slab to exist.
5085                                          * Fortunately, a pending move also
5086                                          * means we don't need to destroy the
5087                                          * slab here, since it will get
5088                                          * destroyed along with any other slabs
5089                                          * on the deadlist after the last
5090                                          * pending move completes.
5091                                          */
5092                                         list_insert_head(deadlist, sp);
5093                                         return (-1);
5094                                 }
5095 
5096                                 /*
5097                                  * Destroy the slab now if it was completely
5098                                  * freed while we dropped cache_lock and there
5099                                  * are no pending moves. Since slab_refcnt
5100                                  * cannot change once it reaches zero, no new
5101                                  * pending moves from that slab are possible.
5102                                  */
5103                                 cp->cache_defrag->kmd_deadcount--;
5104                                 cp->cache_slab_destroy++;
5105                                 mutex_exit(&cp->cache_lock);
5106                                 kmem_slab_destroy(cp, sp);
5107                                 mutex_enter(&cp->cache_lock);
5108                                 /*
5109                                  * Since we can't pick up the scan where we left
5110                                  * off, abort the scan and say nothing about the
5111                                  * number of reclaimable slabs.
5112                                  */
5113                                 return (-1);
5114                         }
5115 
5116                         if (!success) {
5117                                 /*
5118                                  * Abort the scan if there is not enough memory
5119                                  * for the request and say nothing about the
5120                                  * number of reclaimable slabs.
5121                                  */
5122                                 return (-1);
5123                         }
5124 
5125                         /*
5126                          * The slab's position changed while the lock was
5127                          * dropped, so we don't know where we are in the
5128                          * sequence any more.
5129                          */
5130                         if (sp->slab_refcnt != refcnt) {
5131                                 /*
5132                                  * If this is a KMM_DEBUG move, the slab_refcnt
5133                                  * may have changed because we allocated a
5134                                  * destination buffer on the same slab. In that
5135                                  * case, we're not interested in counting it.
5136                                  */
5137                                 return (-1);
5138                         }
5139                         if ((sp->slab_flags & KMEM_SLAB_NOMOVE) != nomove)
5140                                 return (-1);
5141 
5142                         /*
5143                          * Generating a move request allocates a destination
5144                          * buffer from the slab layer, bumping the first partial
5145                          * slab if it is completely allocated. If the current
5146                          * slab becomes the first partial slab as a result, we
5147                          * can't continue to scan backwards.
5148                          *
5149                          * If this is a KMM_DEBUG move and we allocated the
5150                          * destination buffer from the last partial slab, then
5151                          * the buffer we're moving is on the same slab and our
5152                          * slab_refcnt has changed, causing us to return before
5153                          * reaching here if there are no partial slabs left.
5154                          */
5155                         ASSERT(!avl_is_empty(&cp->cache_partial_slabs));
5156                         if (sp == avl_first(&cp->cache_partial_slabs)) {
5157                                 /*
5158                                  * We're not interested in a second KMM_DEBUG
5159                                  * move.
5160                                  */
5161                                 goto end_scan;
5162                         }
5163                 }
5164         }
5165 end_scan:
5166 
5167         return (s);
5168 }
5169 
5170 typedef struct kmem_move_notify_args {
5171         kmem_cache_t *kmna_cache;
5172         void *kmna_buf;
5173 } kmem_move_notify_args_t;
5174 
5175 static void
5176 kmem_cache_move_notify_task(void *arg)
5177 {
5178         kmem_move_notify_args_t *args = arg;
5179         kmem_cache_t *cp = args->kmna_cache;
5180         void *buf = args->kmna_buf;
5181         kmem_slab_t *sp;
5182 
5183         ASSERT(taskq_member(kmem_taskq, curthread));
5184         ASSERT(list_link_active(&cp->cache_link));
5185 
5186         kmem_free(args, sizeof (kmem_move_notify_args_t));
5187         mutex_enter(&cp->cache_lock);
5188         sp = kmem_slab_allocated(cp, NULL, buf);
5189 
5190         /* Ignore the notification if the buffer is no longer allocated. */
5191         if (sp == NULL) {
5192                 mutex_exit(&cp->cache_lock);
5193                 return;
5194         }
5195 
5196         /* Ignore the notification if there's no reason to move the buffer. */
5197         if (avl_numnodes(&cp->cache_partial_slabs) > 1) {
5198                 /*
5199                  * So far the notification is not ignored. Ignore the
5200                  * notification if the slab is not marked by an earlier refusal
5201                  * to move a buffer.
5202                  */
5203                 if (!(sp->slab_flags & KMEM_SLAB_NOMOVE) &&
5204                     (sp->slab_later_count == 0)) {
5205                         mutex_exit(&cp->cache_lock);
5206                         return;
5207                 }
5208 
5209                 kmem_slab_move_yes(cp, sp, buf);
5210                 ASSERT(!(sp->slab_flags & KMEM_SLAB_MOVE_PENDING));
5211                 sp->slab_flags |= KMEM_SLAB_MOVE_PENDING;
5212                 mutex_exit(&cp->cache_lock);
5213                 /* see kmem_move_buffers() about dropping the lock */
5214                 (void) kmem_move_begin(cp, sp, buf, KMM_NOTIFY);
5215                 mutex_enter(&cp->cache_lock);
5216                 ASSERT(sp->slab_flags & KMEM_SLAB_MOVE_PENDING);
5217                 sp->slab_flags &= ~KMEM_SLAB_MOVE_PENDING;
5218                 if (sp->slab_refcnt == 0) {
5219                         list_t *deadlist = &cp->cache_defrag->kmd_deadlist;
5220                         list_remove(deadlist, sp);
5221 
5222                         if (!avl_is_empty(
5223                             &cp->cache_defrag->kmd_moves_pending)) {
5224                                 list_insert_head(deadlist, sp);
5225                                 mutex_exit(&cp->cache_lock);
5226                                 return;
5227                         }
5228 
5229                         cp->cache_defrag->kmd_deadcount--;
5230                         cp->cache_slab_destroy++;
5231                         mutex_exit(&cp->cache_lock);
5232                         kmem_slab_destroy(cp, sp);
5233                         return;
5234                 }
5235         } else {
5236                 kmem_slab_move_yes(cp, sp, buf);
5237         }
5238         mutex_exit(&cp->cache_lock);
5239 }
5240 
5241 void
5242 kmem_cache_move_notify(kmem_cache_t *cp, void *buf)
5243 {
5244         kmem_move_notify_args_t *args;
5245 
5246         args = kmem_alloc(sizeof (kmem_move_notify_args_t), KM_NOSLEEP);
5247         if (args != NULL) {
5248                 args->kmna_cache = cp;
5249                 args->kmna_buf = buf;
5250                 if (!taskq_dispatch(kmem_taskq,
5251                     (task_func_t *)kmem_cache_move_notify_task, args,
5252                     TQ_NOSLEEP))
5253                         kmem_free(args, sizeof (kmem_move_notify_args_t));
5254         }
5255 }
5256 
5257 static void
5258 kmem_cache_defrag(kmem_cache_t *cp)
5259 {
5260         size_t n;
5261 
5262         ASSERT(cp->cache_defrag != NULL);
5263 
5264         mutex_enter(&cp->cache_lock);
5265         n = avl_numnodes(&cp->cache_partial_slabs);
5266         if (n > 1) {
5267                 /* kmem_move_buffers() drops and reacquires cache_lock */
5268                 cp->cache_defrag->kmd_defrags++;
5269                 (void) kmem_move_buffers(cp, n, 0, KMM_DESPERATE);
5270         }
5271         mutex_exit(&cp->cache_lock);
5272 }
5273 
5274 /* Is this cache above the fragmentation threshold? */
5275 static boolean_t
5276 kmem_cache_frag_threshold(kmem_cache_t *cp, uint64_t nfree)
5277 {
5278         /*
5279          *      nfree           kmem_frag_numer
5280          * ------------------ > ---------------
5281          * cp->cache_buftotal        kmem_frag_denom
5282          */
5283         return ((nfree * kmem_frag_denom) >
5284             (cp->cache_buftotal * kmem_frag_numer));
5285 }
5286 
5287 static boolean_t
5288 kmem_cache_is_fragmented(kmem_cache_t *cp, boolean_t *doreap)
5289 {
5290         boolean_t fragmented;
5291         uint64_t nfree;
5292 
5293         ASSERT(MUTEX_HELD(&cp->cache_lock));
5294         *doreap = B_FALSE;
5295 
5296         if (kmem_move_fulltilt) {
5297                 if (avl_numnodes(&cp->cache_partial_slabs) > 1) {
5298                         return (B_TRUE);
5299                 }
5300         } else {
5301                 if ((cp->cache_complete_slab_count + avl_numnodes(
5302                     &cp->cache_partial_slabs)) < kmem_frag_minslabs) {
5303                         return (B_FALSE);
5304                 }
5305         }
5306 
5307         nfree = cp->cache_bufslab;
5308         fragmented = ((avl_numnodes(&cp->cache_partial_slabs) > 1) &&
5309             kmem_cache_frag_threshold(cp, nfree));
5310 
5311         /*
5312          * Free buffers in the magazine layer appear allocated from the point of
5313          * view of the slab layer. We want to know if the slab layer would
5314          * appear fragmented if we included free buffers from magazines that
5315          * have fallen out of the working set.
5316          */
5317         if (!fragmented) {
5318                 long reap;
5319 
5320                 mutex_enter(&cp->cache_depot_lock);
5321                 reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
5322                 reap = MIN(reap, cp->cache_full.ml_total);
5323                 mutex_exit(&cp->cache_depot_lock);
5324 
5325                 nfree += ((uint64_t)reap * cp->cache_magtype->mt_magsize);
5326                 if (kmem_cache_frag_threshold(cp, nfree)) {
5327                         *doreap = B_TRUE;
5328                 }
5329         }
5330 
5331         return (fragmented);
5332 }
5333 
5334 /* Called periodically from kmem_taskq */
5335 static void
5336 kmem_cache_scan(kmem_cache_t *cp)
5337 {
5338         boolean_t reap = B_FALSE;
5339         kmem_defrag_t *kmd;
5340 
5341         ASSERT(taskq_member(kmem_taskq, curthread));
5342 
5343         mutex_enter(&cp->cache_lock);
5344 
5345         kmd = cp->cache_defrag;
5346         if (kmd->kmd_consolidate > 0) {
5347                 kmd->kmd_consolidate--;
5348                 mutex_exit(&cp->cache_lock);
5349                 kmem_cache_reap(cp);
5350                 return;
5351         }
5352 
5353         if (kmem_cache_is_fragmented(cp, &reap)) {
5354                 size_t slabs_found;
5355 
5356                 /*
5357                  * Consolidate reclaimable slabs from the end of the partial
5358                  * slab list (scan at most kmem_reclaim_scan_range slabs to find
5359                  * reclaimable slabs). Keep track of how many candidate slabs we
5360                  * looked for and how many we actually found so we can adjust
5361                  * the definition of a candidate slab if we're having trouble
5362                  * finding them.
5363                  *
5364                  * kmem_move_buffers() drops and reacquires cache_lock.
5365                  */
5366                 kmd->kmd_scans++;
5367                 slabs_found = kmem_move_buffers(cp, kmem_reclaim_scan_range,
5368                     kmem_reclaim_max_slabs, 0);
5369                 if (slabs_found >= 0) {
5370                         kmd->kmd_slabs_sought += kmem_reclaim_max_slabs;
5371                         kmd->kmd_slabs_found += slabs_found;
5372                 }
5373 
5374                 if (++kmd->kmd_tries >= kmem_reclaim_scan_range) {
5375                         kmd->kmd_tries = 0;
5376 
5377                         /*
5378                          * If we had difficulty finding candidate slabs in
5379                          * previous scans, adjust the threshold so that
5380                          * candidates are easier to find.
5381                          */
5382                         if (kmd->kmd_slabs_found == kmd->kmd_slabs_sought) {
5383                                 kmem_adjust_reclaim_threshold(kmd, -1);
5384                         } else if ((kmd->kmd_slabs_found * 2) <
5385                             kmd->kmd_slabs_sought) {
5386                                 kmem_adjust_reclaim_threshold(kmd, 1);
5387                         }
5388                         kmd->kmd_slabs_sought = 0;
5389                         kmd->kmd_slabs_found = 0;
5390                 }
5391         } else {
5392                 kmem_reset_reclaim_threshold(cp->cache_defrag);
5393 #ifdef  DEBUG
5394                 if (!avl_is_empty(&cp->cache_partial_slabs)) {
5395                         /*
5396                          * In a debug kernel we want the consolidator to
5397                          * run occasionally even when there is plenty of
5398                          * memory.
5399                          */
5400                         uint16_t debug_rand;
5401 
5402                         (void) random_get_bytes((uint8_t *)&debug_rand, 2);
5403                         if (!kmem_move_noreap &&
5404                             ((debug_rand % kmem_mtb_reap) == 0)) {
5405                                 mutex_exit(&cp->cache_lock);
5406                                 kmem_cache_reap(cp);
5407                                 return;
5408                         } else if ((debug_rand % kmem_mtb_move) == 0) {
5409                                 kmd->kmd_scans++;
5410                                 (void) kmem_move_buffers(cp,
5411                                     kmem_reclaim_scan_range, 1, KMM_DEBUG);
5412                         }
5413                 }
5414 #endif  /* DEBUG */
5415         }
5416 
5417         mutex_exit(&cp->cache_lock);
5418 
5419         if (reap)
5420                 kmem_depot_ws_reap(cp);
5421 }