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 /*
23 * Copyright 2010 Sun Microsystems, Inc. All rights reserved.
24 * Use is subject to license terms.
25 * Copyright 2021 Joyent, Inc.
26 */
27
28 #include <sys/types.h>
29 #include <sys/systm.h>
30 #include <sys/schedctl.h>
31 #include <sys/proc.h>
32 #include <sys/thread.h>
33 #include <sys/class.h>
34 #include <sys/cred.h>
35 #include <sys/kmem.h>
36 #include <sys/cmn_err.h>
37 #include <sys/stack.h>
38 #include <sys/debug.h>
39 #include <sys/cpuvar.h>
40 #include <sys/sobject.h>
41 #include <sys/door.h>
42 #include <sys/modctl.h>
43 #include <sys/syscall.h>
44 #include <sys/sysmacros.h>
45 #include <sys/vmsystm.h>
46 #include <sys/mman.h>
47 #include <sys/vnode.h>
48 #include <sys/swap.h>
49 #include <sys/lwp.h>
50 #include <sys/bitmap.h>
51 #include <sys/atomic.h>
52 #include <sys/fcntl.h>
53 #include <vm/seg_kp.h>
54 #include <vm/seg_vn.h>
55 #include <vm/as.h>
56 #include <fs/fs_subr.h>
57
58 /*
59 * Page handling structures. This is set up as a list of per-page
60 * control structures (sc_page_ctl), with p->p_pagep pointing to
61 * the first. The per-page structures point to the actual pages
62 * and contain pointers to the user address for each mapped page.
63 *
64 * All data is protected by p->p_sc_lock. Since this lock is
65 * held while waiting for memory, schedctl_shared_alloc() should
66 * not be called while holding p_lock.
67 */
68
69 typedef struct sc_page_ctl {
70 struct sc_page_ctl *spc_next;
71 sc_shared_t *spc_base; /* base of kernel page */
72 sc_shared_t *spc_end; /* end of usable space */
73 ulong_t *spc_map; /* bitmap of allocated space on page */
74 size_t spc_space; /* amount of space on page */
75 caddr_t spc_uaddr; /* user-level address of the page */
76 struct anon_map *spc_amp; /* anonymous memory structure */
77 } sc_page_ctl_t;
78
79 static size_t sc_pagesize; /* size of usable space on page */
80 static size_t sc_bitmap_len; /* # of bits in allocation bitmap */
81 static size_t sc_bitmap_words; /* # of words in allocation bitmap */
82
83 /* Context ops */
84 static void schedctl_save(sc_shared_t *);
85 static void schedctl_restore(sc_shared_t *);
86 static void schedctl_fork(kthread_t *, kthread_t *);
87
88 /* Functions for handling shared pages */
89 static int schedctl_shared_alloc(sc_shared_t **, uintptr_t *);
90 static sc_page_ctl_t *schedctl_page_lookup(sc_shared_t *);
91 static int schedctl_map(struct anon_map *, caddr_t *, caddr_t);
92 static int schedctl_getpage(struct anon_map **, caddr_t *);
93 static void schedctl_freepage(struct anon_map *, caddr_t);
94
95 /*
96 * System call interface to scheduler activations.
97 * This always operates on the current lwp.
98 */
99 caddr_t
100 schedctl(void)
101 {
102 kthread_t *t = curthread;
103 sc_shared_t *ssp;
104 uintptr_t uaddr;
105 int error;
106
107 if (t->t_schedctl == NULL) {
108 /*
109 * Allocate and initialize the shared structure.
110 */
111 if ((error = schedctl_shared_alloc(&ssp, &uaddr)) != 0)
112 return ((caddr_t)(uintptr_t)set_errno(error));
113 bzero(ssp, sizeof (*ssp));
114
115 installctx(t, ssp, schedctl_save, schedctl_restore,
116 schedctl_fork, NULL, NULL, NULL, NULL);
117
118 thread_lock(t); /* protect against ts_tick and ts_update */
119 t->t_schedctl = ssp;
120 t->t_sc_uaddr = uaddr;
121 ssp->sc_cid = t->t_cid;
122 ssp->sc_cpri = t->t_cpri;
123 ssp->sc_priority = DISP_PRIO(t);
124 thread_unlock(t);
125 }
126
127 return ((caddr_t)t->t_sc_uaddr);
128 }
129
130
131 /*
132 * Clean up scheduler activations state associated with an exiting
133 * (or execing) lwp. t is always the current thread.
134 */
135 void
136 schedctl_lwp_cleanup(kthread_t *t)
137 {
138 sc_shared_t *ssp = t->t_schedctl;
139 proc_t *p = ttoproc(t);
140 sc_page_ctl_t *pagep;
141 index_t index;
142
143 ASSERT(MUTEX_NOT_HELD(&p->p_lock));
144
145 thread_lock(t); /* protect against ts_tick and ts_update */
146 t->t_schedctl = NULL;
147 t->t_sc_uaddr = 0;
148 thread_unlock(t);
149
150 /*
151 * Remove the context op to avoid the final call to
152 * schedctl_save when switching away from this lwp.
153 */
154 (void) removectx(t, ssp, schedctl_save, schedctl_restore,
155 schedctl_fork, NULL, NULL, NULL);
156
157 /*
158 * Do not unmap the shared page until the process exits.
159 * User-level library code relies on this for adaptive mutex locking.
160 */
161 mutex_enter(&p->p_sc_lock);
162 ssp->sc_state = SC_FREE;
163 pagep = schedctl_page_lookup(ssp);
164 index = (index_t)(ssp - pagep->spc_base);
165 BT_CLEAR(pagep->spc_map, index);
166 pagep->spc_space += sizeof (sc_shared_t);
167 mutex_exit(&p->p_sc_lock);
168 }
169
170
171 /*
172 * Cleanup the list of schedctl shared pages for the process.
173 * Called from exec() and exit() system calls.
174 */
175 void
176 schedctl_proc_cleanup(void)
177 {
178 proc_t *p = curproc;
179 sc_page_ctl_t *pagep;
180 sc_page_ctl_t *next;
181
182 ASSERT(p->p_lwpcnt == 1); /* we are single-threaded now */
183 ASSERT(curthread->t_schedctl == NULL);
184
185 /*
186 * Since we are single-threaded, we don't have to hold p->p_sc_lock.
187 */
188 pagep = p->p_pagep;
189 p->p_pagep = NULL;
190 while (pagep != NULL) {
191 ASSERT(pagep->spc_space == sc_pagesize);
192 next = pagep->spc_next;
193 /*
194 * Unmap the user space and free the mapping structure.
195 */
196 (void) as_unmap(p->p_as, pagep->spc_uaddr, PAGESIZE);
197 schedctl_freepage(pagep->spc_amp, (caddr_t)(pagep->spc_base));
198 kmem_free(pagep->spc_map, sizeof (ulong_t) * sc_bitmap_words);
199 kmem_free(pagep, sizeof (sc_page_ctl_t));
200 pagep = next;
201 }
202 }
203
204
205 /*
206 * Called by resume just before switching away from the current thread.
207 * Save new thread state.
208 */
209 static void
210 schedctl_save(sc_shared_t *ssp)
211 {
212 ssp->sc_state = curthread->t_state;
213 }
214
215
216 /*
217 * Called by resume after switching to the current thread.
218 * Save new thread state and CPU.
219 */
220 static void
221 schedctl_restore(sc_shared_t *ssp)
222 {
223 ssp->sc_state = SC_ONPROC;
224 ssp->sc_cpu = CPU->cpu_id;
225 }
226
227
228 /*
229 * On fork, remove inherited mappings from the child's address space.
230 * The child's threads must call schedctl() to get new shared mappings.
231 */
232 static void
233 schedctl_fork(kthread_t *pt, kthread_t *ct)
234 {
235 proc_t *pp = ttoproc(pt);
236 proc_t *cp = ttoproc(ct);
237 sc_page_ctl_t *pagep;
238
239 ASSERT(ct->t_schedctl == NULL);
240
241 /*
242 * Do this only once, whether we are doing fork1() or forkall().
243 * Don't do it at all if the child process is a child of vfork()
244 * because a child of vfork() borrows the parent's address space.
245 */
246 if (pt != curthread || (cp->p_flag & SVFORK))
247 return;
248
249 mutex_enter(&pp->p_sc_lock);
250 for (pagep = pp->p_pagep; pagep != NULL; pagep = pagep->spc_next)
251 (void) as_unmap(cp->p_as, pagep->spc_uaddr, PAGESIZE);
252 mutex_exit(&pp->p_sc_lock);
253 }
254
255
256 /*
257 * Returns non-zero if the specified thread shouldn't be preempted at this time.
258 * Called by ts_preempt(), ts_tick(), and ts_update().
259 */
260 int
261 schedctl_get_nopreempt(kthread_t *t)
262 {
263 ASSERT(THREAD_LOCK_HELD(t));
264 return (t->t_schedctl->sc_preemptctl.sc_nopreempt);
265 }
266
267
268 /*
269 * Sets the value of the nopreempt field for the specified thread.
270 * Called by ts_preempt() to clear the field on preemption.
271 */
272 void
273 schedctl_set_nopreempt(kthread_t *t, short val)
274 {
275 ASSERT(THREAD_LOCK_HELD(t));
276 t->t_schedctl->sc_preemptctl.sc_nopreempt = val;
277 }
278
279
280 /*
281 * Sets the value of the yield field for the specified thread.
282 * Called by ts_preempt() and ts_tick() to set the field, and
283 * ts_yield() to clear it.
284 * The kernel never looks at this field so we don't need a
285 * schedctl_get_yield() function.
286 */
287 void
288 schedctl_set_yield(kthread_t *t, short val)
289 {
290 ASSERT(THREAD_LOCK_HELD(t));
291 t->t_schedctl->sc_preemptctl.sc_yield = val;
292 }
293
294
295 /*
296 * Sets the values of the cid and priority fields for the specified thread.
297 * Called from thread_change_pri(), thread_change_epri(), THREAD_CHANGE_PRI().
298 * Called following calls to CL_FORKRET() and CL_ENTERCLASS().
299 */
300 void
301 schedctl_set_cidpri(kthread_t *t)
302 {
303 sc_shared_t *tdp = t->t_schedctl;
304
305 if (tdp != NULL) {
306 tdp->sc_cid = t->t_cid;
307 tdp->sc_cpri = t->t_cpri;
308 tdp->sc_priority = DISP_PRIO(t);
309 }
310 }
311
312
313 /*
314 * Returns non-zero if the specified thread has requested that all
315 * signals be blocked. Called by signal-related code that tests
316 * the signal mask of a thread that may not be the current thread
317 * and where the process's p_lock cannot be acquired.
318 */
319 int
320 schedctl_sigblock(kthread_t *t)
321 {
322 sc_shared_t *tdp = t->t_schedctl;
323
324 if (tdp != NULL)
325 return (tdp->sc_sigblock);
326 return (0);
327 }
328
329
330 /*
331 * If the sc_sigblock field is set for the specified thread, set its signal
332 * mask to block all maskable signals, then clear the sc_sigblock field. This
333 * accomplishes what user-level code requested to be done when it set
334 * tdp->sc_shared->sc_sigblock non-zero.
335 *
336 * This is generally called by signal-related code in the current thread. In
337 * order to call against a thread other than curthread, p_lock for the
338 * containing process must be held. Even then, the caller is not protected
339 * from races with the thread in question updating its own fields. It is the
340 * responsibility of the caller to perform additional synchronization.
341 *
342 */
343 void
344 schedctl_finish_sigblock(kthread_t *t)
345 {
346 sc_shared_t *tdp = t->t_schedctl;
347
348 ASSERT(t == curthread || MUTEX_HELD(&ttoproc(t)->p_lock));
349
350 if (tdp != NULL && tdp->sc_sigblock) {
351 t->t_hold.__sigbits[0] = FILLSET0 & ~CANTMASK0;
352 t->t_hold.__sigbits[1] = FILLSET1 & ~CANTMASK1;
353 t->t_hold.__sigbits[2] = FILLSET2 & ~CANTMASK2;
354 tdp->sc_sigblock = 0;
355 }
356 }
357
358
359 /*
360 * Return non-zero if the current thread has declared that it has
361 * a cancellation pending and that cancellation is not disabled.
362 * If SIGCANCEL is blocked, we must be going over the wire in an
363 * NFS transaction (sigintr() was called); return zero in this case.
364 */
365 int
366 schedctl_cancel_pending(void)
367 {
368 sc_shared_t *tdp = curthread->t_schedctl;
369
370 if (tdp != NULL &&
371 (tdp->sc_flgs & SC_CANCEL_FLG) &&
372 !tdp->sc_sigblock &&
373 !sigismember(&curthread->t_hold, SIGCANCEL))
374 return (1);
375 return (0);
376 }
377
378
379 /*
380 * Inform libc that the kernel returned EINTR from some system call
381 * due to there being a cancellation pending (SC_CANCEL_FLG set or
382 * we received an SI_LWP SIGCANCEL while in a system call), rather
383 * than because of some other signal. User-level code can try to
384 * recover from receiving other signals, but it can't recover from
385 * being cancelled.
386 */
387 void
388 schedctl_cancel_eintr(void)
389 {
390 sc_shared_t *tdp = curthread->t_schedctl;
391
392 if (tdp != NULL)
393 tdp->sc_flgs |= SC_EINTR_FLG;
394 }
395
396
397 /*
398 * Return non-zero if the current thread has declared that
399 * it is calling into the kernel to park, else return zero.
400 */
401 int
402 schedctl_is_park(void)
403 {
404 sc_shared_t *tdp = curthread->t_schedctl;
405
406 if (tdp != NULL)
407 return ((tdp->sc_flgs & SC_PARK_FLG) != 0);
408 /*
409 * If we're here and there is no shared memory (how could
410 * that happen?) then just assume we really are here to park.
411 */
412 return (1);
413 }
414
415
416 /*
417 * Declare thread is parking.
418 *
419 * libc will set "sc_flgs |= SC_PARK_FLG" before calling lwpsys_park(0, tid)
420 * in order to declare that the thread is calling into the kernel to park.
421 *
422 * This interface exists ONLY to support older versions of libthread which
423 * are not aware of the SC_PARK_FLG flag.
424 *
425 * Older versions of libthread which are not aware of the SC_PARK_FLG flag
426 * need to be modified or emulated to call lwpsys_park(4, ...) instead of
427 * lwpsys_park(0, ...). This will invoke schedctl_set_park() before
428 * lwp_park() to declare that the thread is parking.
429 */
430 void
431 schedctl_set_park(void)
432 {
433 sc_shared_t *tdp = curthread->t_schedctl;
434 if (tdp != NULL)
435 tdp->sc_flgs |= SC_PARK_FLG;
436 }
437
438
439 /*
440 * Clear the parking flag on return from parking in the kernel.
441 */
442 void
443 schedctl_unpark(void)
444 {
445 sc_shared_t *tdp = curthread->t_schedctl;
446
447 if (tdp != NULL)
448 tdp->sc_flgs &= ~SC_PARK_FLG;
449 }
450
451
452 /*
453 * Page handling code.
454 */
455
456 void
457 schedctl_init(void)
458 {
459 /*
460 * Amount of page that can hold sc_shared_t structures. If
461 * sizeof (sc_shared_t) is a power of 2, this should just be
462 * PAGESIZE.
463 */
464 sc_pagesize = PAGESIZE - (PAGESIZE % sizeof (sc_shared_t));
465
466 /*
467 * Allocation bitmap is one bit per struct on a page.
468 */
469 sc_bitmap_len = sc_pagesize / sizeof (sc_shared_t);
470 sc_bitmap_words = howmany(sc_bitmap_len, BT_NBIPUL);
471 }
472
473
474 static int
475 schedctl_shared_alloc(sc_shared_t **kaddrp, uintptr_t *uaddrp)
476 {
477 proc_t *p = curproc;
478 sc_page_ctl_t *pagep;
479 sc_shared_t *ssp;
480 caddr_t base;
481 index_t index;
482 int error;
483
484 ASSERT(MUTEX_NOT_HELD(&p->p_lock));
485 mutex_enter(&p->p_sc_lock);
486
487 /*
488 * Try to find space for the new data in existing pages
489 * within the process's list of shared pages.
490 */
491 for (pagep = p->p_pagep; pagep != NULL; pagep = pagep->spc_next)
492 if (pagep->spc_space != 0)
493 break;
494
495 if (pagep != NULL)
496 base = pagep->spc_uaddr;
497 else {
498 struct anon_map *amp;
499 caddr_t kaddr;
500
501 /*
502 * No room, need to allocate a new page. Also set up
503 * a mapping to the kernel address space for the new
504 * page and lock it in memory.
505 */
506 if ((error = schedctl_getpage(&, &kaddr)) != 0) {
507 mutex_exit(&p->p_sc_lock);
508 return (error);
509 }
510 if ((error = schedctl_map(amp, &base, kaddr)) != 0) {
511 schedctl_freepage(amp, kaddr);
512 mutex_exit(&p->p_sc_lock);
513 return (error);
514 }
515
516 /*
517 * Allocate and initialize the page control structure.
518 */
519 pagep = kmem_alloc(sizeof (sc_page_ctl_t), KM_SLEEP);
520 pagep->spc_amp = amp;
521 pagep->spc_base = (sc_shared_t *)kaddr;
522 pagep->spc_end = (sc_shared_t *)(kaddr + sc_pagesize);
523 pagep->spc_uaddr = base;
524
525 pagep->spc_map = kmem_zalloc(sizeof (ulong_t) * sc_bitmap_words,
526 KM_SLEEP);
527 pagep->spc_space = sc_pagesize;
528
529 pagep->spc_next = p->p_pagep;
530 p->p_pagep = pagep;
531 }
532
533 /*
534 * Got a page, now allocate space for the data. There should
535 * be space unless something's wrong.
536 */
537 ASSERT(pagep != NULL && pagep->spc_space >= sizeof (sc_shared_t));
538 index = bt_availbit(pagep->spc_map, sc_bitmap_len);
539 ASSERT(index != -1);
540
541 /*
542 * Get location with pointer arithmetic. spc_base is of type
543 * sc_shared_t *. Mark as allocated.
544 */
545 ssp = pagep->spc_base + index;
546 BT_SET(pagep->spc_map, index);
547 pagep->spc_space -= sizeof (sc_shared_t);
548
549 mutex_exit(&p->p_sc_lock);
550
551 /*
552 * Return kernel and user addresses.
553 */
554 *kaddrp = ssp;
555 *uaddrp = (uintptr_t)base + ((uintptr_t)ssp & PAGEOFFSET);
556 return (0);
557 }
558
559
560 /*
561 * Find the page control structure corresponding to a kernel address.
562 */
563 static sc_page_ctl_t *
564 schedctl_page_lookup(sc_shared_t *ssp)
565 {
566 proc_t *p = curproc;
567 sc_page_ctl_t *pagep;
568
569 ASSERT(MUTEX_HELD(&p->p_sc_lock));
570 for (pagep = p->p_pagep; pagep != NULL; pagep = pagep->spc_next) {
571 if (ssp >= pagep->spc_base && ssp < pagep->spc_end)
572 return (pagep);
573 }
574 return (NULL); /* This "can't happen". Should we panic? */
575 }
576
577
578 /*
579 * This function is called when a page needs to be mapped into a
580 * process's address space. Allocate the user address space and
581 * set up the mapping to the page. Assumes the page has already
582 * been allocated and locked in memory via schedctl_getpage.
583 */
584 static int
585 schedctl_map(struct anon_map *amp, caddr_t *uaddrp, caddr_t kaddr)
586 {
587 caddr_t addr = NULL;
588 struct as *as = curproc->p_as;
589 struct segvn_crargs vn_a;
590 int error;
591
592 as_rangelock(as);
593 /* pass address of kernel mapping as offset to avoid VAC conflicts */
594 map_addr(&addr, PAGESIZE, (offset_t)(uintptr_t)kaddr, 1, 0);
595 if (addr == NULL) {
596 as_rangeunlock(as);
597 return (ENOMEM);
598 }
599
600 /*
601 * Use segvn to set up the mapping to the page.
602 */
603 vn_a.vp = NULL;
604 vn_a.offset = 0;
605 vn_a.cred = NULL;
606 vn_a.type = MAP_SHARED;
607 vn_a.prot = vn_a.maxprot = PROT_ALL;
608 vn_a.flags = 0;
609 vn_a.amp = amp;
610 vn_a.szc = 0;
611 vn_a.lgrp_mem_policy_flags = 0;
612 error = as_map(as, addr, PAGESIZE, segvn_create, &vn_a);
613 as_rangeunlock(as);
614
615 if (error)
616 return (error);
617
618 *uaddrp = addr;
619 return (0);
620 }
621
622
623 /*
624 * Allocate a new page from anonymous memory. Also, create a kernel
625 * mapping to the page and lock the page in memory.
626 */
627 static int
628 schedctl_getpage(struct anon_map **newamp, caddr_t *newaddr)
629 {
630 struct anon_map *amp;
631 caddr_t kaddr;
632
633 /*
634 * Set up anonymous memory struct. No swap reservation is
635 * needed since the page will be locked into memory.
636 */
637 amp = anonmap_alloc(PAGESIZE, 0, ANON_SLEEP);
638
639 /*
640 * Allocate the page.
641 */
642 kaddr = segkp_get_withanonmap(segkp, PAGESIZE,
643 KPD_NO_ANON | KPD_LOCKED | KPD_ZERO, amp);
644 if (kaddr == NULL) {
645 amp->refcnt--;
646 anonmap_free(amp);
647 return (ENOMEM);
648 }
649
650 /*
651 * The page is left SE_SHARED locked so that it won't be
652 * paged out or relocated (KPD_LOCKED above).
653 */
654
655 *newamp = amp;
656 *newaddr = kaddr;
657 return (0);
658 }
659
660
661 /*
662 * Take the necessary steps to allow a page to be released.
663 * This is called when the process is doing exit() or exec().
664 * There should be no accesses to the page after this.
665 * The kernel mapping of the page is released and the page is unlocked.
666 */
667 static void
668 schedctl_freepage(struct anon_map *amp, caddr_t kaddr)
669 {
670 /*
671 * Release the lock on the page and remove the kernel mapping.
672 */
673 ANON_LOCK_ENTER(&->a_rwlock, RW_WRITER);
674 segkp_release(segkp, kaddr);
675
676 /*
677 * Decrement the refcnt so the anon_map structure will be freed.
678 */
679 if (--amp->refcnt == 0) {
680 /*
681 * The current process no longer has the page mapped, so
682 * we have to free everything rather than letting as_free
683 * do the work.
684 */
685 anonmap_purge(amp);
686 anon_free(amp->ahp, 0, PAGESIZE);
687 ANON_LOCK_EXIT(&->a_rwlock);
688 anonmap_free(amp);
689 } else {
690 ANON_LOCK_EXIT(&->a_rwlock);
691 }
692 }