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 (c) 1995, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Copyright (c) 2013, OmniTI Computer Consulting, Inc. All rights reserved.
25 * Copyright 2015, Joyent, Inc. All rights reserved.
26 */
27
28 /* Copyright (c) 2013, OmniTI Computer Consulting, Inc. All rights reserved. */
29 /*
30 * Copyright 2015 Nexenta Systems, Inc. All rights reserved.
31 */
32
33 #include <sys/types.h>
34 #include <sys/t_lock.h>
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/buf.h>
38 #include <sys/conf.h>
39 #include <sys/cred.h>
40 #include <sys/kmem.h>
41 #include <sys/sysmacros.h>
42 #include <sys/vfs.h>
43 #include <sys/vnode.h>
44 #include <sys/debug.h>
45 #include <sys/errno.h>
46 #include <sys/time.h>
47 #include <sys/file.h>
48 #include <sys/user.h>
49 #include <sys/stream.h>
50 #include <sys/strsubr.h>
51 #include <sys/strsun.h>
52 #include <sys/sunddi.h>
53 #include <sys/esunddi.h>
54 #include <sys/flock.h>
55 #include <sys/modctl.h>
56 #include <sys/cmn_err.h>
57 #include <sys/vmsystm.h>
58 #include <sys/policy.h>
59 #include <sys/limits.h>
60
61 #include <sys/socket.h>
62 #include <sys/socketvar.h>
63
64 #include <sys/isa_defs.h>
65 #include <sys/inttypes.h>
66 #include <sys/systm.h>
67 #include <sys/cpuvar.h>
68 #include <sys/filio.h>
69 #include <sys/sendfile.h>
70 #include <sys/ddi.h>
71 #include <vm/seg.h>
72 #include <vm/seg_map.h>
73 #include <vm/seg_kpm.h>
74
75 #include <fs/sockfs/nl7c.h>
76 #include <fs/sockfs/sockcommon.h>
77 #include <fs/sockfs/sockfilter_impl.h>
78 #include <fs/sockfs/socktpi.h>
79
80 #ifdef SOCK_TEST
81 int do_useracc = 1; /* Controlled by setting SO_DEBUG to 4 */
82 #else
83 #define do_useracc 1
84 #endif /* SOCK_TEST */
85
86 extern int xnet_truncate_print;
87
88 extern void nl7c_init(void);
89 extern int sockfs_defer_nl7c_init;
90
91 /*
92 * Kernel component of socket creation.
93 *
94 * The socket library determines which version number to use.
95 * First the library calls this with a NULL devpath. If this fails
96 * to find a transport (using solookup) the library will look in /etc/netconfig
97 * for the appropriate transport. If one is found it will pass in the
98 * devpath for the kernel to use.
99 */
100 int
101 so_socket(int family, int type_w_flags, int protocol, char *devpath,
102 int version)
103 {
104 struct sonode *so;
105 vnode_t *vp;
106 struct file *fp;
107 int fd;
108 int error;
109 int type;
110
111 type = type_w_flags & SOCK_TYPE_MASK;
112 type_w_flags &= ~SOCK_TYPE_MASK;
113 if (type_w_flags & ~(SOCK_CLOEXEC|SOCK_NDELAY|SOCK_NONBLOCK))
114 return (set_errno(EINVAL));
115
116 if (devpath != NULL) {
117 char *buf;
118 size_t kdevpathlen = 0;
119
120 buf = kmem_alloc(MAXPATHLEN, KM_SLEEP);
121 if ((error = copyinstr(devpath, buf,
122 MAXPATHLEN, &kdevpathlen)) != 0) {
123 kmem_free(buf, MAXPATHLEN);
124 return (set_errno(error));
125 }
126 so = socket_create(family, type, protocol, buf, NULL,
127 SOCKET_SLEEP, version, CRED(), &error);
128 kmem_free(buf, MAXPATHLEN);
129 } else {
130 so = socket_create(family, type, protocol, NULL, NULL,
131 SOCKET_SLEEP, version, CRED(), &error);
132 }
133 if (so == NULL)
134 return (set_errno(error));
135
136 /* Allocate a file descriptor for the socket */
137 vp = SOTOV(so);
138 if (error = falloc(vp, FWRITE|FREAD, &fp, &fd)) {
139 (void) socket_close(so, 0, CRED());
140 socket_destroy(so);
141 return (set_errno(error));
142 }
143
144 /*
145 * Now fill in the entries that falloc reserved
146 */
147 if (type_w_flags & SOCK_NDELAY) {
148 so->so_state |= SS_NDELAY;
149 fp->f_flag |= FNDELAY;
150 }
151 if (type_w_flags & SOCK_NONBLOCK) {
152 so->so_state |= SS_NONBLOCK;
153 fp->f_flag |= FNONBLOCK;
154 }
155 mutex_exit(&fp->f_tlock);
156 setf(fd, fp);
157 if ((type_w_flags & SOCK_CLOEXEC) != 0) {
158 f_setfd(fd, FD_CLOEXEC);
159 }
160
161 return (fd);
162 }
163
164 /*
165 * Map from a file descriptor to a socket node.
166 * Returns with the file descriptor held i.e. the caller has to
167 * use releasef when done with the file descriptor.
168 */
169 struct sonode *
170 getsonode(int sock, int *errorp, file_t **fpp)
171 {
172 file_t *fp;
173 vnode_t *vp;
174 struct sonode *so;
175
176 if ((fp = getf(sock)) == NULL) {
177 *errorp = EBADF;
178 eprintline(*errorp);
179 return (NULL);
180 }
181 vp = fp->f_vnode;
182 /* Check if it is a socket */
183 if (vp->v_type != VSOCK) {
184 releasef(sock);
185 *errorp = ENOTSOCK;
186 eprintline(*errorp);
187 return (NULL);
188 }
189 /*
190 * Use the stream head to find the real socket vnode.
191 * This is needed when namefs sits above sockfs.
192 */
193 if (vp->v_stream) {
194 ASSERT(vp->v_stream->sd_vnode);
195 vp = vp->v_stream->sd_vnode;
196
197 so = VTOSO(vp);
198 if (so->so_version == SOV_STREAM) {
199 releasef(sock);
200 *errorp = ENOTSOCK;
201 eprintsoline(so, *errorp);
202 return (NULL);
203 }
204 } else {
205 so = VTOSO(vp);
206 }
207 if (fpp)
208 *fpp = fp;
209 return (so);
210 }
211
212 /*
213 * Allocate and copyin a sockaddr.
214 * Ensures NULL termination for AF_UNIX addresses by extending them
215 * with one NULL byte if need be. Verifies that the length is not
216 * excessive to prevent an application from consuming all of kernel
217 * memory. Returns NULL when an error occurred.
218 */
219 static struct sockaddr *
220 copyin_name(struct sonode *so, struct sockaddr *name, socklen_t *namelenp,
221 int *errorp)
222 {
223 char *faddr;
224 size_t namelen = (size_t)*namelenp;
225
226 ASSERT(namelen != 0);
227 if (namelen > SO_MAXARGSIZE) {
228 *errorp = EINVAL;
229 eprintsoline(so, *errorp);
230 return (NULL);
231 }
232
233 faddr = (char *)kmem_alloc(namelen, KM_SLEEP);
234 if (copyin(name, faddr, namelen)) {
235 kmem_free(faddr, namelen);
236 *errorp = EFAULT;
237 eprintsoline(so, *errorp);
238 return (NULL);
239 }
240
241 /*
242 * Add space for NULL termination if needed.
243 * Do a quick check if the last byte is NUL.
244 */
245 if (so->so_family == AF_UNIX && faddr[namelen - 1] != '\0') {
246 /* Check if there is any NULL termination */
247 size_t i;
248 int foundnull = 0;
249
250 for (i = sizeof (name->sa_family); i < namelen; i++) {
251 if (faddr[i] == '\0') {
252 foundnull = 1;
253 break;
254 }
255 }
256 if (!foundnull) {
257 /* Add extra byte for NUL padding */
258 char *nfaddr;
259
260 nfaddr = (char *)kmem_alloc(namelen + 1, KM_SLEEP);
261 bcopy(faddr, nfaddr, namelen);
262 kmem_free(faddr, namelen);
263
264 /* NUL terminate */
265 nfaddr[namelen] = '\0';
266 namelen++;
267 ASSERT((socklen_t)namelen == namelen);
268 *namelenp = (socklen_t)namelen;
269 faddr = nfaddr;
270 }
271 }
272 return ((struct sockaddr *)faddr);
273 }
274
275 /*
276 * Copy from kaddr/klen to uaddr/ulen. Updates ulenp if non-NULL.
277 */
278 static int
279 copyout_arg(void *uaddr, socklen_t ulen, void *ulenp,
280 void *kaddr, socklen_t klen)
281 {
282 if (uaddr != NULL) {
283 if (ulen > klen)
284 ulen = klen;
285
286 if (ulen != 0) {
287 if (copyout(kaddr, uaddr, ulen))
288 return (EFAULT);
289 }
290 } else
291 ulen = 0;
292
293 if (ulenp != NULL) {
294 if (copyout(&ulen, ulenp, sizeof (ulen)))
295 return (EFAULT);
296 }
297 return (0);
298 }
299
300 /*
301 * Copy from kaddr/klen to uaddr/ulen. Updates ulenp if non-NULL.
302 * If klen is greater than ulen it still uses the non-truncated
303 * klen to update ulenp.
304 */
305 static int
306 copyout_name(void *uaddr, socklen_t ulen, void *ulenp,
307 void *kaddr, socklen_t klen)
308 {
309 if (uaddr != NULL) {
310 if (ulen >= klen)
311 ulen = klen;
312 else if (ulen != 0 && xnet_truncate_print) {
313 printf("sockfs: truncating copyout of address using "
314 "XNET semantics for pid = %d. Lengths %d, %d\n",
315 curproc->p_pid, klen, ulen);
316 }
317
318 if (ulen != 0) {
319 if (copyout(kaddr, uaddr, ulen))
320 return (EFAULT);
321 } else
322 klen = 0;
323 } else
324 klen = 0;
325
326 if (ulenp != NULL) {
327 if (copyout(&klen, ulenp, sizeof (klen)))
328 return (EFAULT);
329 }
330 return (0);
331 }
332
333 /*
334 * The socketpair() code in libsocket creates two sockets (using
335 * the /etc/netconfig fallback if needed) before calling this routine
336 * to connect the two sockets together.
337 *
338 * For a SOCK_STREAM socketpair a listener is needed - in that case this
339 * routine will create a new file descriptor as part of accepting the
340 * connection. The library socketpair() will check if svs[2] has changed
341 * in which case it will close the changed fd.
342 *
343 * Note that this code could use the TPI feature of accepting the connection
344 * on the listening endpoint. However, that would require significant changes
345 * to soaccept.
346 */
347 int
348 so_socketpair(int sv[2])
349 {
350 int svs[2];
351 struct sonode *so1, *so2;
352 int error;
353 int orig_flags;
354 struct sockaddr_ux *name;
355 size_t namelen;
356 sotpi_info_t *sti1;
357 sotpi_info_t *sti2;
358
359 dprint(1, ("so_socketpair(%p)\n", (void *)sv));
360
361 error = useracc(sv, sizeof (svs), B_WRITE);
362 if (error && do_useracc)
363 return (set_errno(EFAULT));
364
365 if (copyin(sv, svs, sizeof (svs)))
366 return (set_errno(EFAULT));
367
368 if ((so1 = getsonode(svs[0], &error, NULL)) == NULL)
369 return (set_errno(error));
370
371 if ((so2 = getsonode(svs[1], &error, NULL)) == NULL) {
372 releasef(svs[0]);
373 return (set_errno(error));
374 }
375
376 if (so1->so_family != AF_UNIX || so2->so_family != AF_UNIX) {
377 error = EOPNOTSUPP;
378 goto done;
379 }
380
381 sti1 = SOTOTPI(so1);
382 sti2 = SOTOTPI(so2);
383
384 /*
385 * The code below makes assumptions about the "sockfs" implementation.
386 * So make sure that the correct implementation is really used.
387 */
388 ASSERT(so1->so_ops == &sotpi_sonodeops);
389 ASSERT(so2->so_ops == &sotpi_sonodeops);
390
391 if (so1->so_type == SOCK_DGRAM) {
392 /*
393 * Bind both sockets and connect them with each other.
394 * Need to allocate name/namelen for soconnect.
395 */
396 error = socket_bind(so1, NULL, 0, _SOBIND_UNSPEC, CRED());
397 if (error) {
398 eprintsoline(so1, error);
399 goto done;
400 }
401 error = socket_bind(so2, NULL, 0, _SOBIND_UNSPEC, CRED());
402 if (error) {
403 eprintsoline(so2, error);
404 goto done;
405 }
406 namelen = sizeof (struct sockaddr_ux);
407 name = kmem_alloc(namelen, KM_SLEEP);
408 name->sou_family = AF_UNIX;
409 name->sou_addr = sti2->sti_ux_laddr;
410 error = socket_connect(so1,
411 (struct sockaddr *)name,
412 (socklen_t)namelen,
413 0, _SOCONNECT_NOXLATE, CRED());
414 if (error) {
415 kmem_free(name, namelen);
416 eprintsoline(so1, error);
417 goto done;
418 }
419 name->sou_addr = sti1->sti_ux_laddr;
420 error = socket_connect(so2,
421 (struct sockaddr *)name,
422 (socklen_t)namelen,
423 0, _SOCONNECT_NOXLATE, CRED());
424 kmem_free(name, namelen);
425 if (error) {
426 eprintsoline(so2, error);
427 goto done;
428 }
429 releasef(svs[0]);
430 releasef(svs[1]);
431 } else {
432 /*
433 * Bind both sockets, with so1 being a listener.
434 * Connect so2 to so1 - nonblocking to avoid waiting for
435 * soaccept to complete.
436 * Accept a connection on so1. Pass out the new fd as sv[0].
437 * The library will detect the changed fd and close
438 * the original one.
439 */
440 struct sonode *nso;
441 struct vnode *nvp;
442 struct file *nfp;
443 int nfd;
444
445 /*
446 * We could simply call socket_listen() here (which would do the
447 * binding automatically) if the code didn't rely on passing
448 * _SOBIND_NOXLATE to the TPI implementation of socket_bind().
449 */
450 error = socket_bind(so1, NULL, 0, _SOBIND_UNSPEC|
451 _SOBIND_NOXLATE|_SOBIND_LISTEN|_SOBIND_SOCKETPAIR,
452 CRED());
453 if (error) {
454 eprintsoline(so1, error);
455 goto done;
456 }
457 error = socket_bind(so2, NULL, 0, _SOBIND_UNSPEC, CRED());
458 if (error) {
459 eprintsoline(so2, error);
460 goto done;
461 }
462
463 namelen = sizeof (struct sockaddr_ux);
464 name = kmem_alloc(namelen, KM_SLEEP);
465 name->sou_family = AF_UNIX;
466 name->sou_addr = sti1->sti_ux_laddr;
467 error = socket_connect(so2,
468 (struct sockaddr *)name,
469 (socklen_t)namelen,
470 FNONBLOCK, _SOCONNECT_NOXLATE, CRED());
471 kmem_free(name, namelen);
472 if (error) {
473 if (error != EINPROGRESS) {
474 eprintsoline(so2, error); goto done;
475 }
476 }
477
478 error = socket_accept(so1, 0, CRED(), &nso);
479 if (error) {
480 eprintsoline(so1, error);
481 goto done;
482 }
483
484 /* wait for so2 being SS_CONNECTED ignoring signals */
485 mutex_enter(&so2->so_lock);
486 error = sowaitconnected(so2, 0, 1);
487 mutex_exit(&so2->so_lock);
488 if (error != 0) {
489 (void) socket_close(nso, 0, CRED());
490 socket_destroy(nso);
491 eprintsoline(so2, error);
492 goto done;
493 }
494
495 nvp = SOTOV(nso);
496 if (error = falloc(nvp, FWRITE|FREAD, &nfp, &nfd)) {
497 (void) socket_close(nso, 0, CRED());
498 socket_destroy(nso);
499 eprintsoline(nso, error);
500 goto done;
501 }
502 /*
503 * copy over FNONBLOCK and FNDELAY flags should they exist
504 */
505 if (so1->so_state & SS_NONBLOCK)
506 nfp->f_flag |= FNONBLOCK;
507 if (so1->so_state & SS_NDELAY)
508 nfp->f_flag |= FNDELAY;
509
510 /*
511 * fill in the entries that falloc reserved
512 */
513 mutex_exit(&nfp->f_tlock);
514 setf(nfd, nfp);
515
516 /*
517 * get the original flags before we release
518 */
519 VERIFY(f_getfd_error(svs[0], &orig_flags) == 0);
520
521 releasef(svs[0]);
522 releasef(svs[1]);
523
524 /*
525 * If FD_CLOEXEC was set on the filedescriptor we're
526 * swapping out, we should set it on the new one too.
527 */
528 if (orig_flags & FD_CLOEXEC) {
529 f_setfd(nfd, FD_CLOEXEC);
530 }
531
532 /*
533 * The socketpair library routine will close the original
534 * svs[0] when this code passes out a different file
535 * descriptor.
536 */
537 svs[0] = nfd;
538
539 if (copyout(svs, sv, sizeof (svs))) {
540 (void) closeandsetf(nfd, NULL);
541 eprintline(EFAULT);
542 return (set_errno(EFAULT));
543 }
544 }
545 return (0);
546
547 done:
548 releasef(svs[0]);
549 releasef(svs[1]);
550 return (set_errno(error));
551 }
552
553 int
554 bind(int sock, struct sockaddr *name, socklen_t namelen, int version)
555 {
556 struct sonode *so;
557 int error;
558
559 dprint(1, ("bind(%d, %p, %d)\n",
560 sock, (void *)name, namelen));
561
562 if ((so = getsonode(sock, &error, NULL)) == NULL)
563 return (set_errno(error));
564
565 /* Allocate and copyin name */
566 /*
567 * X/Open test does not expect EFAULT with NULL name and non-zero
568 * namelen.
569 */
570 if (name != NULL && namelen != 0) {
571 ASSERT(MUTEX_NOT_HELD(&so->so_lock));
572 name = copyin_name(so, name, &namelen, &error);
573 if (name == NULL) {
574 releasef(sock);
575 return (set_errno(error));
576 }
577 } else {
578 name = NULL;
579 namelen = 0;
580 }
581
582 switch (version) {
583 default:
584 error = socket_bind(so, name, namelen, 0, CRED());
585 break;
586 case SOV_XPG4_2:
587 error = socket_bind(so, name, namelen, _SOBIND_XPG4_2, CRED());
588 break;
589 case SOV_SOCKBSD:
590 error = socket_bind(so, name, namelen, _SOBIND_SOCKBSD, CRED());
591 break;
592 }
593 done:
594 releasef(sock);
595 if (name != NULL)
596 kmem_free(name, (size_t)namelen);
597
598 if (error)
599 return (set_errno(error));
600 return (0);
601 }
602
603 /* ARGSUSED2 */
604 int
605 listen(int sock, int backlog, int version)
606 {
607 struct sonode *so;
608 int error;
609
610 dprint(1, ("listen(%d, %d)\n",
611 sock, backlog));
612
613 if ((so = getsonode(sock, &error, NULL)) == NULL)
614 return (set_errno(error));
615
616 error = socket_listen(so, backlog, CRED());
617
618 releasef(sock);
619 if (error)
620 return (set_errno(error));
621 return (0);
622 }
623
624 /*ARGSUSED3*/
625 int
626 accept(int sock, struct sockaddr *name, socklen_t *namelenp, int version,
627 int flags)
628 {
629 struct sonode *so;
630 file_t *fp;
631 int error;
632 socklen_t namelen;
633 struct sonode *nso;
634 struct vnode *nvp;
635 struct file *nfp;
636 int nfd;
637 int ssflags;
638 struct sockaddr *addrp;
639 socklen_t addrlen;
640
641 dprint(1, ("accept(%d, %p, %p)\n",
642 sock, (void *)name, (void *)namelenp));
643
644 if (flags & ~(SOCK_CLOEXEC|SOCK_NONBLOCK|SOCK_NDELAY)) {
645 return (set_errno(EINVAL));
646 }
647
648 /* Translate SOCK_ flags to their SS_ variant */
649 ssflags = 0;
650 if (flags & SOCK_NONBLOCK)
651 ssflags |= SS_NONBLOCK;
652 if (flags & SOCK_NDELAY)
653 ssflags |= SS_NDELAY;
654
655 if ((so = getsonode(sock, &error, &fp)) == NULL)
656 return (set_errno(error));
657
658 if (name != NULL) {
659 ASSERT(MUTEX_NOT_HELD(&so->so_lock));
660 if (copyin(namelenp, &namelen, sizeof (namelen))) {
661 releasef(sock);
662 return (set_errno(EFAULT));
663 }
664 if (namelen != 0) {
665 error = useracc(name, (size_t)namelen, B_WRITE);
666 if (error && do_useracc) {
667 releasef(sock);
668 return (set_errno(EFAULT));
669 }
670 } else
671 name = NULL;
672 } else {
673 namelen = 0;
674 }
675
676 /*
677 * Allocate the user fd before socket_accept() in order to
678 * catch EMFILE errors before calling socket_accept().
679 */
680 if ((nfd = ufalloc(0)) == -1) {
681 eprintsoline(so, EMFILE);
682 releasef(sock);
683 return (set_errno(EMFILE));
684 }
685 error = socket_accept(so, fp->f_flag, CRED(), &nso);
686 if (error) {
687 setf(nfd, NULL);
688 releasef(sock);
689 return (set_errno(error));
690 }
691
692 nvp = SOTOV(nso);
693
694 ASSERT(MUTEX_NOT_HELD(&nso->so_lock));
695 if (namelen != 0) {
696 addrlen = so->so_max_addr_len;
697 addrp = (struct sockaddr *)kmem_alloc(addrlen, KM_SLEEP);
698
699 if ((error = socket_getpeername(nso, (struct sockaddr *)addrp,
700 &addrlen, B_TRUE, CRED())) == 0) {
701 error = copyout_name(name, namelen, namelenp,
702 addrp, addrlen);
703 } else {
704 ASSERT(error == EINVAL || error == ENOTCONN);
705 error = ECONNABORTED;
706 }
707 kmem_free(addrp, so->so_max_addr_len);
708 }
709
710 if (error) {
711 setf(nfd, NULL);
712 (void) socket_close(nso, 0, CRED());
713 socket_destroy(nso);
714 releasef(sock);
715 return (set_errno(error));
716 }
717 if (error = falloc(NULL, FWRITE|FREAD, &nfp, NULL)) {
718 setf(nfd, NULL);
719 (void) socket_close(nso, 0, CRED());
720 socket_destroy(nso);
721 eprintsoline(so, error);
722 releasef(sock);
723 return (set_errno(error));
724 }
725 /*
726 * fill in the entries that falloc reserved
727 */
728 nfp->f_vnode = nvp;
729 mutex_exit(&nfp->f_tlock);
730 setf(nfd, nfp);
731
732 /*
733 * Act on SOCK_CLOEXEC from flags
734 */
735 if (flags & SOCK_CLOEXEC) {
736 f_setfd(nfd, FD_CLOEXEC);
737 }
738
739 /*
740 * Copy FNDELAY and FNONBLOCK from listener to acceptor
741 * and from ssflags
742 */
743 if ((ssflags | so->so_state) & (SS_NDELAY|SS_NONBLOCK)) {
744 uint_t oflag = nfp->f_flag;
745 int arg = 0;
746
747 if ((ssflags | so->so_state) & SS_NONBLOCK)
748 arg |= FNONBLOCK;
749 else if ((ssflags | so->so_state) & SS_NDELAY)
750 arg |= FNDELAY;
751
752 /*
753 * This code is a simplification of the F_SETFL code in fcntl()
754 * Ignore any errors from VOP_SETFL.
755 */
756 if ((error = VOP_SETFL(nvp, oflag, arg, nfp->f_cred, NULL))
757 != 0) {
758 eprintsoline(so, error);
759 error = 0;
760 } else {
761 mutex_enter(&nfp->f_tlock);
762 nfp->f_flag &= ~FMASK | (FREAD|FWRITE);
763 nfp->f_flag |= arg;
764 mutex_exit(&nfp->f_tlock);
765 }
766 }
767 releasef(sock);
768 return (nfd);
769 }
770
771 int
772 connect(int sock, struct sockaddr *name, socklen_t namelen, int version)
773 {
774 struct sonode *so;
775 file_t *fp;
776 int error;
777
778 dprint(1, ("connect(%d, %p, %d)\n",
779 sock, (void *)name, namelen));
780
781 if ((so = getsonode(sock, &error, &fp)) == NULL)
782 return (set_errno(error));
783
784 /* Allocate and copyin name */
785 if (namelen != 0) {
786 ASSERT(MUTEX_NOT_HELD(&so->so_lock));
787 name = copyin_name(so, name, &namelen, &error);
788 if (name == NULL) {
789 releasef(sock);
790 return (set_errno(error));
791 }
792 } else
793 name = NULL;
794
795 error = socket_connect(so, name, namelen, fp->f_flag,
796 (version != SOV_XPG4_2) ? 0 : _SOCONNECT_XPG4_2, CRED());
797 releasef(sock);
798 if (name)
799 kmem_free(name, (size_t)namelen);
800 if (error)
801 return (set_errno(error));
802 return (0);
803 }
804
805 /*ARGSUSED2*/
806 int
807 shutdown(int sock, int how, int version)
808 {
809 struct sonode *so;
810 int error;
811
812 dprint(1, ("shutdown(%d, %d)\n",
813 sock, how));
814
815 if ((so = getsonode(sock, &error, NULL)) == NULL)
816 return (set_errno(error));
817
818 error = socket_shutdown(so, how, CRED());
819
820 releasef(sock);
821 if (error)
822 return (set_errno(error));
823 return (0);
824 }
825
826 /*
827 * Common receive routine.
828 */
829 static ssize_t
830 recvit(int sock,
831 struct nmsghdr *msg,
832 struct uio *uiop,
833 int flags,
834 socklen_t *namelenp,
835 socklen_t *controllenp,
836 int *flagsp)
837 {
838 struct sonode *so;
839 file_t *fp;
840 void *name;
841 socklen_t namelen;
842 void *control;
843 socklen_t controllen;
844 ssize_t len;
845 int error;
846
847 if ((so = getsonode(sock, &error, &fp)) == NULL)
848 return (set_errno(error));
849
850 len = uiop->uio_resid;
851 uiop->uio_fmode = fp->f_flag;
852 uiop->uio_extflg = UIO_COPY_CACHED;
853
854 name = msg->msg_name;
855 namelen = msg->msg_namelen;
856 control = msg->msg_control;
857 controllen = msg->msg_controllen;
858
859 msg->msg_flags = flags & (MSG_OOB | MSG_PEEK | MSG_WAITALL |
860 MSG_DONTWAIT | MSG_XPG4_2);
861
862 error = socket_recvmsg(so, msg, uiop, CRED());
863 if (error) {
864 releasef(sock);
865 return (set_errno(error));
866 }
867 lwp_stat_update(LWP_STAT_MSGRCV, 1);
868 releasef(sock);
869
870 error = copyout_name(name, namelen, namelenp,
871 msg->msg_name, msg->msg_namelen);
872 if (error)
873 goto err;
874
875 if (flagsp != NULL) {
876 /*
877 * Clear internal flag.
878 */
879 msg->msg_flags &= ~MSG_XPG4_2;
880
881 /*
882 * Determine MSG_CTRUNC. sorecvmsg sets MSG_CTRUNC only
883 * when controllen is zero and there is control data to
884 * copy out.
885 */
886 if (controllen != 0 &&
887 (msg->msg_controllen > controllen || control == NULL)) {
888 dprint(1, ("recvit: CTRUNC %d %d %p\n",
889 msg->msg_controllen, controllen, control));
890
891 msg->msg_flags |= MSG_CTRUNC;
892 }
893 if (copyout(&msg->msg_flags, flagsp,
894 sizeof (msg->msg_flags))) {
895 error = EFAULT;
896 goto err;
897 }
898 }
899 /*
900 * Note: This MUST be done last. There can be no "goto err" after this
901 * point since it could make so_closefds run twice on some part
902 * of the file descriptor array.
903 */
904 if (controllen != 0) {
905 if (!(flags & MSG_XPG4_2)) {
906 /*
907 * Good old msg_accrights can only return a multiple
908 * of 4 bytes.
909 */
910 controllen &= ~((int)sizeof (uint32_t) - 1);
911 }
912 error = copyout_arg(control, controllen, controllenp,
913 msg->msg_control, msg->msg_controllen);
914 if (error)
915 goto err;
916
917 if (msg->msg_controllen > controllen || control == NULL) {
918 if (control == NULL)
919 controllen = 0;
920 so_closefds(msg->msg_control, msg->msg_controllen,
921 !(flags & MSG_XPG4_2), controllen);
922 }
923 }
924 if (msg->msg_namelen != 0)
925 kmem_free(msg->msg_name, (size_t)msg->msg_namelen);
926 if (msg->msg_controllen != 0)
927 kmem_free(msg->msg_control, (size_t)msg->msg_controllen);
928 return (len - uiop->uio_resid);
929
930 err:
931 /*
932 * If we fail and the control part contains file descriptors
933 * we have to close the fd's.
934 */
935 if (msg->msg_controllen != 0)
936 so_closefds(msg->msg_control, msg->msg_controllen,
937 !(flags & MSG_XPG4_2), 0);
938 if (msg->msg_namelen != 0)
939 kmem_free(msg->msg_name, (size_t)msg->msg_namelen);
940 if (msg->msg_controllen != 0)
941 kmem_free(msg->msg_control, (size_t)msg->msg_controllen);
942 return (set_errno(error));
943 }
944
945 /*
946 * Native system call
947 */
948 ssize_t
949 recv(int sock, void *buffer, size_t len, int flags)
950 {
951 struct nmsghdr lmsg;
952 struct uio auio;
953 struct iovec aiov[1];
954
955 dprint(1, ("recv(%d, %p, %ld, %d)\n",
956 sock, buffer, len, flags));
957
958 if ((ssize_t)len < 0) {
959 return (set_errno(EINVAL));
960 }
961
962 aiov[0].iov_base = buffer;
963 aiov[0].iov_len = len;
964 auio.uio_loffset = 0;
965 auio.uio_iov = aiov;
966 auio.uio_iovcnt = 1;
967 auio.uio_resid = len;
968 auio.uio_segflg = UIO_USERSPACE;
969 auio.uio_limit = 0;
970
971 lmsg.msg_namelen = 0;
972 lmsg.msg_controllen = 0;
973 lmsg.msg_flags = 0;
974 return (recvit(sock, &lmsg, &auio, flags, NULL, NULL, NULL));
975 }
976
977 ssize_t
978 recvfrom(int sock, void *buffer, size_t len, int flags,
979 struct sockaddr *name, socklen_t *namelenp)
980 {
981 struct nmsghdr lmsg;
982 struct uio auio;
983 struct iovec aiov[1];
984
985 dprint(1, ("recvfrom(%d, %p, %ld, %d, %p, %p)\n",
986 sock, buffer, len, flags, (void *)name, (void *)namelenp));
987
988 if ((ssize_t)len < 0) {
989 return (set_errno(EINVAL));
990 }
991
992 aiov[0].iov_base = buffer;
993 aiov[0].iov_len = len;
994 auio.uio_loffset = 0;
995 auio.uio_iov = aiov;
996 auio.uio_iovcnt = 1;
997 auio.uio_resid = len;
998 auio.uio_segflg = UIO_USERSPACE;
999 auio.uio_limit = 0;
1000
1001 lmsg.msg_name = (char *)name;
1002 if (namelenp != NULL) {
1003 if (copyin(namelenp, &lmsg.msg_namelen,
1004 sizeof (lmsg.msg_namelen)))
1005 return (set_errno(EFAULT));
1006 } else {
1007 lmsg.msg_namelen = 0;
1008 }
1009 lmsg.msg_controllen = 0;
1010 lmsg.msg_flags = 0;
1011
1012 return (recvit(sock, &lmsg, &auio, flags, namelenp, NULL, NULL));
1013 }
1014
1015 /*
1016 * Uses the MSG_XPG4_2 flag to determine if the caller is using
1017 * struct omsghdr or struct nmsghdr.
1018 */
1019 ssize_t
1020 recvmsg(int sock, struct nmsghdr *msg, int flags)
1021 {
1022 STRUCT_DECL(nmsghdr, u_lmsg);
1023 STRUCT_HANDLE(nmsghdr, umsgptr);
1024 struct nmsghdr lmsg;
1025 struct uio auio;
1026 struct iovec buf[IOV_MAX_STACK], *aiov = buf;
1027 ssize_t iovsize = 0;
1028 int iovcnt;
1029 ssize_t len, rval;
1030 int i;
1031 int *flagsp;
1032 model_t model;
1033
1034 dprint(1, ("recvmsg(%d, %p, %d)\n",
1035 sock, (void *)msg, flags));
1036
1037 model = get_udatamodel();
1038 STRUCT_INIT(u_lmsg, model);
1039 STRUCT_SET_HANDLE(umsgptr, model, msg);
1040
1041 if (flags & MSG_XPG4_2) {
1042 if (copyin(msg, STRUCT_BUF(u_lmsg), STRUCT_SIZE(u_lmsg)))
1043 return (set_errno(EFAULT));
1044 flagsp = STRUCT_FADDR(umsgptr, msg_flags);
1045 } else {
1046 /*
1047 * Assumes that nmsghdr and omsghdr are identically shaped
1048 * except for the added msg_flags field.
1049 */
1050 if (copyin(msg, STRUCT_BUF(u_lmsg),
1051 SIZEOF_STRUCT(omsghdr, model)))
1052 return (set_errno(EFAULT));
1053 STRUCT_FSET(u_lmsg, msg_flags, 0);
1054 flagsp = NULL;
1055 }
1056
1057 /*
1058 * Code below us will kmem_alloc memory and hang it
1059 * off msg_control and msg_name fields. This forces
1060 * us to copy the structure to its native form.
1061 */
1062 lmsg.msg_name = STRUCT_FGETP(u_lmsg, msg_name);
1063 lmsg.msg_namelen = STRUCT_FGET(u_lmsg, msg_namelen);
1064 lmsg.msg_iov = STRUCT_FGETP(u_lmsg, msg_iov);
1065 lmsg.msg_iovlen = STRUCT_FGET(u_lmsg, msg_iovlen);
1066 lmsg.msg_control = STRUCT_FGETP(u_lmsg, msg_control);
1067 lmsg.msg_controllen = STRUCT_FGET(u_lmsg, msg_controllen);
1068 lmsg.msg_flags = STRUCT_FGET(u_lmsg, msg_flags);
1069
1070 iovcnt = lmsg.msg_iovlen;
1071
1072 if (iovcnt <= 0 || iovcnt > IOV_MAX) {
1073 return (set_errno(EMSGSIZE));
1074 }
1075
1076 if (iovcnt > IOV_MAX_STACK) {
1077 iovsize = iovcnt * sizeof (struct iovec);
1078 aiov = kmem_alloc(iovsize, KM_SLEEP);
1079 }
1080
1081 #ifdef _SYSCALL32_IMPL
1082 /*
1083 * 32-bit callers need to have their iovec expanded, while ensuring
1084 * that they can't move more than 2Gbytes of data in a single call.
1085 */
1086 if (model == DATAMODEL_ILP32) {
1087 struct iovec32 buf32[IOV_MAX_STACK], *aiov32 = buf32;
1088 ssize_t iov32size;
1089 ssize32_t count32;
1090
1091 iov32size = iovcnt * sizeof (struct iovec32);
1092 if (iovsize != 0)
1093 aiov32 = kmem_alloc(iov32size, KM_SLEEP);
1094
1095 if (copyin((struct iovec32 *)lmsg.msg_iov, aiov32, iov32size)) {
1096 if (iovsize != 0) {
1097 kmem_free(aiov32, iov32size);
1098 kmem_free(aiov, iovsize);
1099 }
1100
1101 return (set_errno(EFAULT));
1102 }
1103
1104 count32 = 0;
1105 for (i = 0; i < iovcnt; i++) {
1106 ssize32_t iovlen32;
1107
1108 iovlen32 = aiov32[i].iov_len;
1109 count32 += iovlen32;
1110 if (iovlen32 < 0 || count32 < 0) {
1111 if (iovsize != 0) {
1112 kmem_free(aiov32, iov32size);
1113 kmem_free(aiov, iovsize);
1114 }
1115
1116 return (set_errno(EINVAL));
1117 }
1118
1119 aiov[i].iov_len = iovlen32;
1120 aiov[i].iov_base =
1121 (caddr_t)(uintptr_t)aiov32[i].iov_base;
1122 }
1123
1124 if (iovsize != 0)
1125 kmem_free(aiov32, iov32size);
1126 } else
1127 #endif /* _SYSCALL32_IMPL */
1128 if (copyin(lmsg.msg_iov, aiov, iovcnt * sizeof (struct iovec))) {
1129 if (iovsize != 0)
1130 kmem_free(aiov, iovsize);
1131
1132 return (set_errno(EFAULT));
1133 }
1134 len = 0;
1135 for (i = 0; i < iovcnt; i++) {
1136 ssize_t iovlen = aiov[i].iov_len;
1137 len += iovlen;
1138 if (iovlen < 0 || len < 0) {
1139 if (iovsize != 0)
1140 kmem_free(aiov, iovsize);
1141
1142 return (set_errno(EINVAL));
1143 }
1144 }
1145 auio.uio_loffset = 0;
1146 auio.uio_iov = aiov;
1147 auio.uio_iovcnt = iovcnt;
1148 auio.uio_resid = len;
1149 auio.uio_segflg = UIO_USERSPACE;
1150 auio.uio_limit = 0;
1151
1152 if (lmsg.msg_control != NULL &&
1153 (do_useracc == 0 ||
1154 useracc(lmsg.msg_control, lmsg.msg_controllen,
1155 B_WRITE) != 0)) {
1156 if (iovsize != 0)
1157 kmem_free(aiov, iovsize);
1158
1159 return (set_errno(EFAULT));
1160 }
1161
1162 rval = recvit(sock, &lmsg, &auio, flags,
1163 STRUCT_FADDR(umsgptr, msg_namelen),
1164 STRUCT_FADDR(umsgptr, msg_controllen), flagsp);
1165
1166 if (iovsize != 0)
1167 kmem_free(aiov, iovsize);
1168
1169 return (rval);
1170 }
1171
1172 /*
1173 * Common send function.
1174 */
1175 static ssize_t
1176 sendit(int sock, struct nmsghdr *msg, struct uio *uiop, int flags)
1177 {
1178 struct sonode *so;
1179 file_t *fp;
1180 void *name;
1181 socklen_t namelen;
1182 void *control;
1183 socklen_t controllen;
1184 ssize_t len;
1185 int error;
1186
1187 if ((so = getsonode(sock, &error, &fp)) == NULL)
1188 return (set_errno(error));
1189
1190 uiop->uio_fmode = fp->f_flag;
1191
1192 if (so->so_family == AF_UNIX)
1193 uiop->uio_extflg = UIO_COPY_CACHED;
1194 else
1195 uiop->uio_extflg = UIO_COPY_DEFAULT;
1196
1197 /* Allocate and copyin name and control */
1198 name = msg->msg_name;
1199 namelen = msg->msg_namelen;
1200 if (name != NULL && namelen != 0) {
1201 ASSERT(MUTEX_NOT_HELD(&so->so_lock));
1202 name = copyin_name(so,
1203 (struct sockaddr *)name,
1204 &namelen, &error);
1205 if (name == NULL)
1206 goto done3;
1207 /* copyin_name null terminates addresses for AF_UNIX */
1208 msg->msg_namelen = namelen;
1209 msg->msg_name = name;
1210 } else {
1211 msg->msg_name = name = NULL;
1212 msg->msg_namelen = namelen = 0;
1213 }
1214
1215 control = msg->msg_control;
1216 controllen = msg->msg_controllen;
1217 if ((control != NULL) && (controllen != 0)) {
1218 /*
1219 * Verify that the length is not excessive to prevent
1220 * an application from consuming all of kernel memory.
1221 */
1222 if (controllen > SO_MAXARGSIZE) {
1223 error = EINVAL;
1224 goto done2;
1225 }
1226 control = kmem_alloc(controllen, KM_SLEEP);
1227
1228 ASSERT(MUTEX_NOT_HELD(&so->so_lock));
1229 if (copyin(msg->msg_control, control, controllen)) {
1230 error = EFAULT;
1231 goto done1;
1232 }
1233 msg->msg_control = control;
1234 } else {
1235 msg->msg_control = control = NULL;
1236 msg->msg_controllen = controllen = 0;
1237 }
1238
1239 len = uiop->uio_resid;
1240 msg->msg_flags = flags;
1241
1242 error = socket_sendmsg(so, msg, uiop, CRED());
1243 done1:
1244 if (control != NULL)
1245 kmem_free(control, controllen);
1246 done2:
1247 if (name != NULL)
1248 kmem_free(name, namelen);
1249 done3:
1250 if (error != 0) {
1251 releasef(sock);
1252 return (set_errno(error));
1253 }
1254 lwp_stat_update(LWP_STAT_MSGSND, 1);
1255 releasef(sock);
1256 return (len - uiop->uio_resid);
1257 }
1258
1259 /*
1260 * Native system call
1261 */
1262 ssize_t
1263 send(int sock, void *buffer, size_t len, int flags)
1264 {
1265 struct nmsghdr lmsg;
1266 struct uio auio;
1267 struct iovec aiov[1];
1268
1269 dprint(1, ("send(%d, %p, %ld, %d)\n",
1270 sock, buffer, len, flags));
1271
1272 if ((ssize_t)len < 0) {
1273 return (set_errno(EINVAL));
1274 }
1275
1276 aiov[0].iov_base = buffer;
1277 aiov[0].iov_len = len;
1278 auio.uio_loffset = 0;
1279 auio.uio_iov = aiov;
1280 auio.uio_iovcnt = 1;
1281 auio.uio_resid = len;
1282 auio.uio_segflg = UIO_USERSPACE;
1283 auio.uio_limit = 0;
1284
1285 lmsg.msg_name = NULL;
1286 lmsg.msg_control = NULL;
1287 if (!(flags & MSG_XPG4_2)) {
1288 /*
1289 * In order to be compatible with the libsocket/sockmod
1290 * implementation we set EOR for all send* calls.
1291 */
1292 flags |= MSG_EOR;
1293 }
1294 return (sendit(sock, &lmsg, &auio, flags));
1295 }
1296
1297 /*
1298 * Uses the MSG_XPG4_2 flag to determine if the caller is using
1299 * struct omsghdr or struct nmsghdr.
1300 */
1301 ssize_t
1302 sendmsg(int sock, struct nmsghdr *msg, int flags)
1303 {
1304 struct nmsghdr lmsg;
1305 STRUCT_DECL(nmsghdr, u_lmsg);
1306 struct uio auio;
1307 struct iovec buf[IOV_MAX_STACK], *aiov = buf;
1308 ssize_t iovsize = 0;
1309 int iovcnt;
1310 ssize_t len, rval;
1311 int i;
1312 model_t model;
1313
1314 dprint(1, ("sendmsg(%d, %p, %d)\n", sock, (void *)msg, flags));
1315
1316 model = get_udatamodel();
1317 STRUCT_INIT(u_lmsg, model);
1318
1319 if (flags & MSG_XPG4_2) {
1320 if (copyin(msg, (char *)STRUCT_BUF(u_lmsg),
1321 STRUCT_SIZE(u_lmsg)))
1322 return (set_errno(EFAULT));
1323 } else {
1324 /*
1325 * Assumes that nmsghdr and omsghdr are identically shaped
1326 * except for the added msg_flags field.
1327 */
1328 if (copyin(msg, (char *)STRUCT_BUF(u_lmsg),
1329 SIZEOF_STRUCT(omsghdr, model)))
1330 return (set_errno(EFAULT));
1331 /*
1332 * In order to be compatible with the libsocket/sockmod
1333 * implementation we set EOR for all send* calls.
1334 */
1335 flags |= MSG_EOR;
1336 }
1337
1338 /*
1339 * Code below us will kmem_alloc memory and hang it
1340 * off msg_control and msg_name fields. This forces
1341 * us to copy the structure to its native form.
1342 */
1343 lmsg.msg_name = STRUCT_FGETP(u_lmsg, msg_name);
1344 lmsg.msg_namelen = STRUCT_FGET(u_lmsg, msg_namelen);
1345 lmsg.msg_iov = STRUCT_FGETP(u_lmsg, msg_iov);
1346 lmsg.msg_iovlen = STRUCT_FGET(u_lmsg, msg_iovlen);
1347 lmsg.msg_control = STRUCT_FGETP(u_lmsg, msg_control);
1348 lmsg.msg_controllen = STRUCT_FGET(u_lmsg, msg_controllen);
1349 lmsg.msg_flags = STRUCT_FGET(u_lmsg, msg_flags);
1350
1351 iovcnt = lmsg.msg_iovlen;
1352
1353 if (iovcnt <= 0 || iovcnt > IOV_MAX) {
1354 /*
1355 * Unless this is XPG 4.2 we allow iovcnt == 0 to
1356 * be compatible with SunOS 4.X and 4.4BSD.
1357 */
1358 if (iovcnt != 0 || (flags & MSG_XPG4_2))
1359 return (set_errno(EMSGSIZE));
1360 }
1361
1362 if (iovcnt > IOV_MAX_STACK) {
1363 iovsize = iovcnt * sizeof (struct iovec);
1364 aiov = kmem_alloc(iovsize, KM_SLEEP);
1365 }
1366
1367 #ifdef _SYSCALL32_IMPL
1368 /*
1369 * 32-bit callers need to have their iovec expanded, while ensuring
1370 * that they can't move more than 2Gbytes of data in a single call.
1371 */
1372 if (model == DATAMODEL_ILP32) {
1373 struct iovec32 buf32[IOV_MAX_STACK], *aiov32 = buf32;
1374 ssize_t iov32size;
1375 ssize32_t count32;
1376
1377 iov32size = iovcnt * sizeof (struct iovec32);
1378 if (iovsize != 0)
1379 aiov32 = kmem_alloc(iov32size, KM_SLEEP);
1380
1381 if (iovcnt != 0 &&
1382 copyin((struct iovec32 *)lmsg.msg_iov, aiov32, iov32size)) {
1383 if (iovsize != 0) {
1384 kmem_free(aiov32, iov32size);
1385 kmem_free(aiov, iovsize);
1386 }
1387
1388 return (set_errno(EFAULT));
1389 }
1390
1391 count32 = 0;
1392 for (i = 0; i < iovcnt; i++) {
1393 ssize32_t iovlen32;
1394
1395 iovlen32 = aiov32[i].iov_len;
1396 count32 += iovlen32;
1397 if (iovlen32 < 0 || count32 < 0) {
1398 if (iovsize != 0) {
1399 kmem_free(aiov32, iov32size);
1400 kmem_free(aiov, iovsize);
1401 }
1402
1403 return (set_errno(EINVAL));
1404 }
1405
1406 aiov[i].iov_len = iovlen32;
1407 aiov[i].iov_base =
1408 (caddr_t)(uintptr_t)aiov32[i].iov_base;
1409 }
1410
1411 if (iovsize != 0)
1412 kmem_free(aiov32, iov32size);
1413 } else
1414 #endif /* _SYSCALL32_IMPL */
1415 if (iovcnt != 0 &&
1416 copyin(lmsg.msg_iov, aiov,
1417 (unsigned)iovcnt * sizeof (struct iovec))) {
1418 if (iovsize != 0)
1419 kmem_free(aiov, iovsize);
1420
1421 return (set_errno(EFAULT));
1422 }
1423 len = 0;
1424 for (i = 0; i < iovcnt; i++) {
1425 ssize_t iovlen = aiov[i].iov_len;
1426 len += iovlen;
1427 if (iovlen < 0 || len < 0) {
1428 if (iovsize != 0)
1429 kmem_free(aiov, iovsize);
1430
1431 return (set_errno(EINVAL));
1432 }
1433 }
1434 auio.uio_loffset = 0;
1435 auio.uio_iov = aiov;
1436 auio.uio_iovcnt = iovcnt;
1437 auio.uio_resid = len;
1438 auio.uio_segflg = UIO_USERSPACE;
1439 auio.uio_limit = 0;
1440
1441 rval = sendit(sock, &lmsg, &auio, flags);
1442
1443 if (iovsize != 0)
1444 kmem_free(aiov, iovsize);
1445
1446 return (rval);
1447 }
1448
1449 ssize_t
1450 sendto(int sock, void *buffer, size_t len, int flags,
1451 struct sockaddr *name, socklen_t namelen)
1452 {
1453 struct nmsghdr lmsg;
1454 struct uio auio;
1455 struct iovec aiov[1];
1456
1457 dprint(1, ("sendto(%d, %p, %ld, %d, %p, %d)\n",
1458 sock, buffer, len, flags, (void *)name, namelen));
1459
1460 if ((ssize_t)len < 0) {
1461 return (set_errno(EINVAL));
1462 }
1463
1464 aiov[0].iov_base = buffer;
1465 aiov[0].iov_len = len;
1466 auio.uio_loffset = 0;
1467 auio.uio_iov = aiov;
1468 auio.uio_iovcnt = 1;
1469 auio.uio_resid = len;
1470 auio.uio_segflg = UIO_USERSPACE;
1471 auio.uio_limit = 0;
1472
1473 lmsg.msg_name = (char *)name;
1474 lmsg.msg_namelen = namelen;
1475 lmsg.msg_control = NULL;
1476 if (!(flags & MSG_XPG4_2)) {
1477 /*
1478 * In order to be compatible with the libsocket/sockmod
1479 * implementation we set EOR for all send* calls.
1480 */
1481 flags |= MSG_EOR;
1482 }
1483 return (sendit(sock, &lmsg, &auio, flags));
1484 }
1485
1486 /*ARGSUSED3*/
1487 int
1488 getpeername(int sock, struct sockaddr *name, socklen_t *namelenp, int version)
1489 {
1490 struct sonode *so;
1491 int error;
1492 socklen_t namelen;
1493 socklen_t sock_addrlen;
1494 struct sockaddr *sock_addrp;
1495
1496 dprint(1, ("getpeername(%d, %p, %p)\n",
1497 sock, (void *)name, (void *)namelenp));
1498
1499 if ((so = getsonode(sock, &error, NULL)) == NULL)
1500 goto bad;
1501
1502 ASSERT(MUTEX_NOT_HELD(&so->so_lock));
1503 if (copyin(namelenp, &namelen, sizeof (namelen)) ||
1504 (name == NULL && namelen != 0)) {
1505 error = EFAULT;
1506 goto rel_out;
1507 }
1508 sock_addrlen = so->so_max_addr_len;
1509 sock_addrp = (struct sockaddr *)kmem_alloc(sock_addrlen, KM_SLEEP);
1510
1511 if ((error = socket_getpeername(so, sock_addrp, &sock_addrlen,
1512 B_FALSE, CRED())) == 0) {
1513 ASSERT(sock_addrlen <= so->so_max_addr_len);
1514 error = copyout_name(name, namelen, namelenp,
1515 (void *)sock_addrp, sock_addrlen);
1516 }
1517 kmem_free(sock_addrp, so->so_max_addr_len);
1518 rel_out:
1519 releasef(sock);
1520 bad: return (error != 0 ? set_errno(error) : 0);
1521 }
1522
1523 /*ARGSUSED3*/
1524 int
1525 getsockname(int sock, struct sockaddr *name,
1526 socklen_t *namelenp, int version)
1527 {
1528 struct sonode *so;
1529 int error;
1530 socklen_t namelen, sock_addrlen;
1531 struct sockaddr *sock_addrp;
1532
1533 dprint(1, ("getsockname(%d, %p, %p)\n",
1534 sock, (void *)name, (void *)namelenp));
1535
1536 if ((so = getsonode(sock, &error, NULL)) == NULL)
1537 goto bad;
1538
1539 ASSERT(MUTEX_NOT_HELD(&so->so_lock));
1540 if (copyin(namelenp, &namelen, sizeof (namelen)) ||
1541 (name == NULL && namelen != 0)) {
1542 error = EFAULT;
1543 goto rel_out;
1544 }
1545
1546 sock_addrlen = so->so_max_addr_len;
1547 sock_addrp = (struct sockaddr *)kmem_alloc(sock_addrlen, KM_SLEEP);
1548 if ((error = socket_getsockname(so, sock_addrp, &sock_addrlen,
1549 CRED())) == 0) {
1550 ASSERT(MUTEX_NOT_HELD(&so->so_lock));
1551 ASSERT(sock_addrlen <= so->so_max_addr_len);
1552 error = copyout_name(name, namelen, namelenp,
1553 (void *)sock_addrp, sock_addrlen);
1554 }
1555 kmem_free(sock_addrp, so->so_max_addr_len);
1556 rel_out:
1557 releasef(sock);
1558 bad: return (error != 0 ? set_errno(error) : 0);
1559 }
1560
1561 /*ARGSUSED5*/
1562 int
1563 getsockopt(int sock,
1564 int level,
1565 int option_name,
1566 void *option_value,
1567 socklen_t *option_lenp,
1568 int version)
1569 {
1570 struct sonode *so;
1571 socklen_t optlen, optlen_res;
1572 void *optval;
1573 int error;
1574
1575 dprint(1, ("getsockopt(%d, %d, %d, %p, %p)\n",
1576 sock, level, option_name, option_value, (void *)option_lenp));
1577
1578 if ((so = getsonode(sock, &error, NULL)) == NULL)
1579 return (set_errno(error));
1580
1581 ASSERT(MUTEX_NOT_HELD(&so->so_lock));
1582 if (copyin(option_lenp, &optlen, sizeof (optlen))) {
1583 releasef(sock);
1584 return (set_errno(EFAULT));
1585 }
1586 /*
1587 * Verify that the length is not excessive to prevent
1588 * an application from consuming all of kernel memory.
1589 */
1590 if (optlen > SO_MAXARGSIZE) {
1591 error = EINVAL;
1592 releasef(sock);
1593 return (set_errno(error));
1594 }
1595 optval = kmem_alloc(optlen, KM_SLEEP);
1596 optlen_res = optlen;
1597 error = socket_getsockopt(so, level, option_name, optval,
1598 &optlen_res, (version != SOV_XPG4_2) ? 0 : _SOGETSOCKOPT_XPG4_2,
1599 CRED());
1600 releasef(sock);
1601 if (error) {
1602 kmem_free(optval, optlen);
1603 return (set_errno(error));
1604 }
1605 error = copyout_arg(option_value, optlen, option_lenp,
1606 optval, optlen_res);
1607 kmem_free(optval, optlen);
1608 if (error)
1609 return (set_errno(error));
1610 return (0);
1611 }
1612
1613 /*ARGSUSED5*/
1614 int
1615 setsockopt(int sock,
1616 int level,
1617 int option_name,
1618 void *option_value,
1619 socklen_t option_len,
1620 int version)
1621 {
1622 struct sonode *so;
1623 intptr_t buffer[2];
1624 void *optval = NULL;
1625 int error;
1626
1627 dprint(1, ("setsockopt(%d, %d, %d, %p, %d)\n",
1628 sock, level, option_name, option_value, option_len));
1629
1630 if ((so = getsonode(sock, &error, NULL)) == NULL)
1631 return (set_errno(error));
1632
1633 if (option_value != NULL) {
1634 if (option_len != 0) {
1635 /*
1636 * Verify that the length is not excessive to prevent
1637 * an application from consuming all of kernel memory.
1638 */
1639 if (option_len > SO_MAXARGSIZE) {
1640 error = EINVAL;
1641 goto done2;
1642 }
1643 optval = option_len <= sizeof (buffer) ?
1644 &buffer : kmem_alloc((size_t)option_len, KM_SLEEP);
1645 ASSERT(MUTEX_NOT_HELD(&so->so_lock));
1646 if (copyin(option_value, optval, (size_t)option_len)) {
1647 error = EFAULT;
1648 goto done1;
1649 }
1650 }
1651 } else
1652 option_len = 0;
1653
1654 error = socket_setsockopt(so, level, option_name, optval,
1655 (t_uscalar_t)option_len, CRED());
1656 done1:
1657 if (optval != buffer)
1658 kmem_free(optval, (size_t)option_len);
1659 done2:
1660 releasef(sock);
1661 if (error)
1662 return (set_errno(error));
1663 return (0);
1664 }
1665
1666 static int
1667 sockconf_add_sock(int family, int type, int protocol, char *name)
1668 {
1669 int error = 0;
1670 char *kdevpath = NULL;
1671 char *kmodule = NULL;
1672 char *buf = NULL;
1673 size_t pathlen = 0;
1674 struct sockparams *sp;
1675
1676 if (name == NULL)
1677 return (EINVAL);
1678 /*
1679 * Copyin the name.
1680 * This also makes it possible to check for too long pathnames.
1681 * Compress the space needed for the name before passing it
1682 * to soconfig - soconfig will store the string until
1683 * the configuration is removed.
1684 */
1685 buf = kmem_alloc(MAXPATHLEN, KM_SLEEP);
1686 if ((error = copyinstr(name, buf, MAXPATHLEN, &pathlen)) != 0) {
1687 kmem_free(buf, MAXPATHLEN);
1688 return (error);
1689 }
1690 if (strncmp(buf, "/dev", strlen("/dev")) == 0) {
1691 /* For device */
1692
1693 /*
1694 * Special handling for NCA:
1695 *
1696 * DEV_NCA is never opened even if an application
1697 * requests for AF_NCA. The device opened is instead a
1698 * predefined AF_INET transport (NCA_INET_DEV).
1699 *
1700 * Prior to Volo (PSARC/2007/587) NCA would determine
1701 * the device using a lookup, which worked then because
1702 * all protocols were based on TPI. Since TPI is no
1703 * longer the default, we have to explicitly state
1704 * which device to use.
1705 */
1706 if (strcmp(buf, NCA_DEV) == 0) {
1707 /* only support entry <28, 2, 0> */
1708 if (family != AF_NCA || type != SOCK_STREAM ||
1709 protocol != 0) {
1710 kmem_free(buf, MAXPATHLEN);
1711 return (EINVAL);
1712 }
1713
1714 pathlen = strlen(NCA_INET_DEV) + 1;
1715 kdevpath = kmem_alloc(pathlen, KM_SLEEP);
1716 bcopy(NCA_INET_DEV, kdevpath, pathlen);
1717 kdevpath[pathlen - 1] = '\0';
1718 } else {
1719 kdevpath = kmem_alloc(pathlen, KM_SLEEP);
1720 bcopy(buf, kdevpath, pathlen);
1721 kdevpath[pathlen - 1] = '\0';
1722 }
1723 } else {
1724 /* For socket module */
1725 kmodule = kmem_alloc(pathlen, KM_SLEEP);
1726 bcopy(buf, kmodule, pathlen);
1727 kmodule[pathlen - 1] = '\0';
1728 pathlen = 0;
1729 }
1730 kmem_free(buf, MAXPATHLEN);
1731
1732 /* sockparams_create frees mod name and devpath upon failure */
1733 sp = sockparams_create(family, type, protocol, kmodule,
1734 kdevpath, pathlen, 0, KM_SLEEP, &error);
1735 if (sp != NULL) {
1736 error = sockparams_add(sp);
1737 if (error != 0)
1738 sockparams_destroy(sp);
1739 }
1740
1741 return (error);
1742 }
1743
1744 static int
1745 sockconf_remove_sock(int family, int type, int protocol)
1746 {
1747 return (sockparams_delete(family, type, protocol));
1748 }
1749
1750 static int
1751 sockconfig_remove_filter(const char *uname)
1752 {
1753 char kname[SOF_MAXNAMELEN];
1754 size_t len;
1755 int error;
1756 sof_entry_t *ent;
1757
1758 if ((error = copyinstr(uname, kname, SOF_MAXNAMELEN, &len)) != 0)
1759 return (error);
1760
1761 ent = sof_entry_remove_by_name(kname);
1762 if (ent == NULL)
1763 return (ENXIO);
1764
1765 mutex_enter(&ent->sofe_lock);
1766 ASSERT(!(ent->sofe_flags & SOFEF_CONDEMED));
1767 if (ent->sofe_refcnt == 0) {
1768 mutex_exit(&ent->sofe_lock);
1769 sof_entry_free(ent);
1770 } else {
1771 /* let the last socket free the filter */
1772 ent->sofe_flags |= SOFEF_CONDEMED;
1773 mutex_exit(&ent->sofe_lock);
1774 }
1775
1776 return (0);
1777 }
1778
1779 static int
1780 sockconfig_add_filter(const char *uname, void *ufilpropp)
1781 {
1782 struct sockconfig_filter_props filprop;
1783 sof_entry_t *ent;
1784 int error;
1785 size_t tuplesz, len;
1786 char hintbuf[SOF_MAXNAMELEN];
1787
1788 ent = kmem_zalloc(sizeof (sof_entry_t), KM_SLEEP);
1789 mutex_init(&ent->sofe_lock, NULL, MUTEX_DEFAULT, NULL);
1790
1791 if ((error = copyinstr(uname, ent->sofe_name, SOF_MAXNAMELEN,
1792 &len)) != 0) {
1793 sof_entry_free(ent);
1794 return (error);
1795 }
1796
1797 if (get_udatamodel() == DATAMODEL_NATIVE) {
1798 if (copyin(ufilpropp, &filprop, sizeof (filprop)) != 0) {
1799 sof_entry_free(ent);
1800 return (EFAULT);
1801 }
1802 }
1803 #ifdef _SYSCALL32_IMPL
1804 else {
1805 struct sockconfig_filter_props32 filprop32;
1806
1807 if (copyin(ufilpropp, &filprop32, sizeof (filprop32)) != 0) {
1808 sof_entry_free(ent);
1809 return (EFAULT);
1810 }
1811 filprop.sfp_modname = (char *)(uintptr_t)filprop32.sfp_modname;
1812 filprop.sfp_autoattach = filprop32.sfp_autoattach;
1813 filprop.sfp_hint = filprop32.sfp_hint;
1814 filprop.sfp_hintarg = (char *)(uintptr_t)filprop32.sfp_hintarg;
1815 filprop.sfp_socktuple_cnt = filprop32.sfp_socktuple_cnt;
1816 filprop.sfp_socktuple =
1817 (sof_socktuple_t *)(uintptr_t)filprop32.sfp_socktuple;
1818 }
1819 #endif /* _SYSCALL32_IMPL */
1820
1821 if ((error = copyinstr(filprop.sfp_modname, ent->sofe_modname,
1822 sizeof (ent->sofe_modname), &len)) != 0) {
1823 sof_entry_free(ent);
1824 return (error);
1825 }
1826
1827 /*
1828 * A filter must specify at least one socket tuple.
1829 */
1830 if (filprop.sfp_socktuple_cnt == 0 ||
1831 filprop.sfp_socktuple_cnt > SOF_MAXSOCKTUPLECNT) {
1832 sof_entry_free(ent);
1833 return (EINVAL);
1834 }
1835 ent->sofe_flags = filprop.sfp_autoattach ? SOFEF_AUTO : SOFEF_PROG;
1836 ent->sofe_hint = filprop.sfp_hint;
1837
1838 /*
1839 * Verify the hint, and copy in the hint argument, if necessary.
1840 */
1841 switch (ent->sofe_hint) {
1842 case SOF_HINT_BEFORE:
1843 case SOF_HINT_AFTER:
1844 if ((error = copyinstr(filprop.sfp_hintarg, hintbuf,
1845 sizeof (hintbuf), &len)) != 0) {
1846 sof_entry_free(ent);
1847 return (error);
1848 }
1849 ent->sofe_hintarg = kmem_alloc(len, KM_SLEEP);
1850 bcopy(hintbuf, ent->sofe_hintarg, len);
1851 /* FALLTHRU */
1852 case SOF_HINT_TOP:
1853 case SOF_HINT_BOTTOM:
1854 /* hints cannot be used with programmatic filters */
1855 if (ent->sofe_flags & SOFEF_PROG) {
1856 sof_entry_free(ent);
1857 return (EINVAL);
1858 }
1859 break;
1860 case SOF_HINT_NONE:
1861 break;
1862 default:
1863 /* bad hint value */
1864 sof_entry_free(ent);
1865 return (EINVAL);
1866 }
1867
1868 ent->sofe_socktuple_cnt = filprop.sfp_socktuple_cnt;
1869 tuplesz = sizeof (sof_socktuple_t) * ent->sofe_socktuple_cnt;
1870 ent->sofe_socktuple = kmem_alloc(tuplesz, KM_SLEEP);
1871
1872 if (get_udatamodel() == DATAMODEL_NATIVE) {
1873 if (copyin(filprop.sfp_socktuple, ent->sofe_socktuple,
1874 tuplesz)) {
1875 sof_entry_free(ent);
1876 return (EFAULT);
1877 }
1878 }
1879 #ifdef _SYSCALL32_IMPL
1880 else {
1881 int i;
1882 caddr_t data = (caddr_t)filprop.sfp_socktuple;
1883 sof_socktuple_t *tup = ent->sofe_socktuple;
1884 sof_socktuple32_t tup32;
1885
1886 tup = ent->sofe_socktuple;
1887 for (i = 0; i < ent->sofe_socktuple_cnt; i++, tup++) {
1888 ASSERT(tup < ent->sofe_socktuple + tuplesz);
1889
1890 if (copyin(data, &tup32, sizeof (tup32)) != 0) {
1891 sof_entry_free(ent);
1892 return (EFAULT);
1893 }
1894 tup->sofst_family = tup32.sofst_family;
1895 tup->sofst_type = tup32.sofst_type;
1896 tup->sofst_protocol = tup32.sofst_protocol;
1897
1898 data += sizeof (tup32);
1899 }
1900 }
1901 #endif /* _SYSCALL32_IMPL */
1902
1903 /* Sockets can start using the filter as soon as the filter is added */
1904 if ((error = sof_entry_add(ent)) != 0)
1905 sof_entry_free(ent);
1906
1907 return (error);
1908 }
1909
1910 /*
1911 * Socket configuration system call. It is used to add and remove
1912 * socket types.
1913 */
1914 int
1915 sockconfig(int cmd, void *arg1, void *arg2, void *arg3, void *arg4)
1916 {
1917 int error = 0;
1918
1919 if (secpolicy_net_config(CRED(), B_FALSE) != 0)
1920 return (set_errno(EPERM));
1921
1922 if (sockfs_defer_nl7c_init) {
1923 nl7c_init();
1924 sockfs_defer_nl7c_init = 0;
1925 }
1926
1927 switch (cmd) {
1928 case SOCKCONFIG_ADD_SOCK:
1929 error = sockconf_add_sock((int)(uintptr_t)arg1,
1930 (int)(uintptr_t)arg2, (int)(uintptr_t)arg3, arg4);
1931 break;
1932 case SOCKCONFIG_REMOVE_SOCK:
1933 error = sockconf_remove_sock((int)(uintptr_t)arg1,
1934 (int)(uintptr_t)arg2, (int)(uintptr_t)arg3);
1935 break;
1936 case SOCKCONFIG_ADD_FILTER:
1937 error = sockconfig_add_filter((const char *)arg1, arg2);
1938 break;
1939 case SOCKCONFIG_REMOVE_FILTER:
1940 error = sockconfig_remove_filter((const char *)arg1);
1941 break;
1942 case SOCKCONFIG_GET_SOCKTABLE:
1943 error = sockparams_copyout_socktable((int)(uintptr_t)arg1);
1944 break;
1945 default:
1946 #ifdef DEBUG
1947 cmn_err(CE_NOTE, "sockconfig: unkonwn subcommand %d", cmd);
1948 #endif
1949 error = EINVAL;
1950 break;
1951 }
1952
1953 if (error != 0) {
1954 eprintline(error);
1955 return (set_errno(error));
1956 }
1957 return (0);
1958 }
1959
1960
1961 /*
1962 * Sendfile is implemented through two schemes, direct I/O or by
1963 * caching in the filesystem page cache. We cache the input file by
1964 * default and use direct I/O only if sendfile_max_size is set
1965 * appropriately as explained below. Note that this logic is consistent
1966 * with other filesystems where caching is turned on by default
1967 * unless explicitly turned off by using the DIRECTIO ioctl.
1968 *
1969 * We choose a slightly different scheme here. One can turn off
1970 * caching by setting sendfile_max_size to 0. One can also enable
1971 * caching of files <= sendfile_max_size by setting sendfile_max_size
1972 * to an appropriate value. By default sendfile_max_size is set to the
1973 * maximum value so that all files are cached. In future, we may provide
1974 * better interfaces for caching the file.
1975 *
1976 * Sendfile through Direct I/O (Zero copy)
1977 * --------------------------------------
1978 *
1979 * As disks are normally slower than the network, we can't have a
1980 * single thread that reads the disk and writes to the network. We
1981 * need to have parallelism. This is done by having the sendfile
1982 * thread create another thread that reads from the filesystem
1983 * and queues it for network processing. In this scheme, the data
1984 * is never copied anywhere i.e it is zero copy unlike the other
1985 * scheme.
1986 *
1987 * We have a sendfile queue (snfq) where each sendfile
1988 * request (snf_req_t) is queued for processing by a thread. Number
1989 * of threads is dynamically allocated and they exit if they are idling
1990 * beyond a specified amount of time. When each request (snf_req_t) is
1991 * processed by a thread, it produces a number of mblk_t structures to
1992 * be consumed by the sendfile thread. snf_deque and snf_enque are
1993 * used for consuming and producing mblks. Size of the filesystem
1994 * read is determined by the tunable (sendfile_read_size). A single
1995 * mblk holds sendfile_read_size worth of data (except the last
1996 * read of the file) which is sent down as a whole to the network.
1997 * sendfile_read_size is set to 1 MB as this seems to be the optimal
1998 * value for the UFS filesystem backed by a striped storage array.
1999 *
2000 * Synchronisation between read (producer) and write (consumer) threads.
2001 * --------------------------------------------------------------------
2002 *
2003 * sr_lock protects sr_ib_head and sr_ib_tail. The lock is held while
2004 * adding and deleting items in this list. Error can happen anytime
2005 * during read or write. There could be unprocessed mblks in the
2006 * sr_ib_XXX list when a read or write error occurs. Whenever error
2007 * is encountered, we need two things to happen :
2008 *
2009 * a) One of the threads need to clean the mblks.
2010 * b) When one thread encounters an error, the other should stop.
2011 *
2012 * For (a), we don't want to penalize the reader thread as it could do
2013 * some useful work processing other requests. For (b), the error can
2014 * be detected by examining sr_read_error or sr_write_error.
2015 * sr_lock protects sr_read_error and sr_write_error. If both reader and
2016 * writer encounters error, we need to report the write error back to
2017 * the application as that's what would have happened if the operations
2018 * were done sequentially. With this in mind, following should work :
2019 *
2020 * - Check for errors before read or write.
2021 * - If the reader encounters error, set the error in sr_read_error.
2022 * Check sr_write_error, if it is set, send cv_signal as it is
2023 * waiting for reader to complete. If it is not set, the writer
2024 * is either running sinking data to the network or blocked
2025 * because of flow control. For handling the latter case, we
2026 * always send a signal. In any case, it will examine sr_read_error
2027 * and return. sr_read_error is marked with SR_READ_DONE to tell
2028 * the writer that the reader is done in all the cases.
2029 * - If the writer encounters error, set the error in sr_write_error.
2030 * The reader thread is either blocked because of flow control or
2031 * running reading data from the disk. For the former, we need to
2032 * wakeup the thread. Again to keep it simple, we always wake up
2033 * the reader thread. Then, wait for the read thread to complete
2034 * if it is not done yet. Cleanup and return.
2035 *
2036 * High and low water marks for the read thread.
2037 * --------------------------------------------
2038 *
2039 * If sendfile() is used to send data over a slow network, we need to
2040 * make sure that the read thread does not produce data at a faster
2041 * rate than the network. This can happen if the disk is faster than
2042 * the network. In such a case, we don't want to build a very large queue.
2043 * But we would still like to get all of the network throughput possible.
2044 * This implies that network should never block waiting for data.
2045 * As there are lot of disk throughput/network throughput combinations
2046 * possible, it is difficult to come up with an accurate number.
2047 * A typical 10K RPM disk has a max seek latency 17ms and rotational
2048 * latency of 3ms for reading a disk block. Thus, the total latency to
2049 * initiate a new read, transfer data from the disk and queue for
2050 * transmission would take about a max of 25ms. Todays max transfer rate
2051 * for network is 100MB/sec. If the thread is blocked because of flow
2052 * control, it would take 25ms to get new data ready for transmission.
2053 * We have to make sure that network is not idling, while we are initiating
2054 * new transfers. So, at 100MB/sec, to keep network busy we would need
2055 * 2.5MB of data. Rounding off, we keep the low water mark to be 3MB of data.
2056 * We need to pick a high water mark so that the woken up thread would
2057 * do considerable work before blocking again to prevent thrashing. Currently,
2058 * we pick this to be 10 times that of the low water mark.
2059 *
2060 * Sendfile with segmap caching (One copy from page cache to mblks).
2061 * ----------------------------------------------------------------
2062 *
2063 * We use the segmap cache for caching the file, if the size of file
2064 * is <= sendfile_max_size. In this case we don't use threads as VM
2065 * is reasonably fast enough to keep up with the network. If the underlying
2066 * transport allows, we call segmap_getmapflt() to map MAXBSIZE (8K) worth
2067 * of data into segmap space, and use the virtual address from segmap
2068 * directly through desballoc() to avoid copy. Once the transport is done
2069 * with the data, the mapping will be released through segmap_release()
2070 * called by the call-back routine.
2071 *
2072 * If zero-copy is not allowed by the transport, we simply call VOP_READ()
2073 * to copy the data from the filesystem into our temporary network buffer.
2074 *
2075 * To disable caching, set sendfile_max_size to 0.
2076 */
2077
2078 uint_t sendfile_read_size = 1024 * 1024;
2079 #define SENDFILE_REQ_LOWAT 3 * 1024 * 1024
2080 uint_t sendfile_req_lowat = SENDFILE_REQ_LOWAT;
2081 uint_t sendfile_req_hiwat = 10 * SENDFILE_REQ_LOWAT;
2082 struct sendfile_stats sf_stats;
2083 struct sendfile_queue *snfq;
2084 clock_t snfq_timeout;
2085 off64_t sendfile_max_size;
2086
2087 static void snf_enque(snf_req_t *, mblk_t *);
2088 static mblk_t *snf_deque(snf_req_t *);
2089
2090 void
2091 sendfile_init(void)
2092 {
2093 snfq = kmem_zalloc(sizeof (struct sendfile_queue), KM_SLEEP);
2094
2095 mutex_init(&snfq->snfq_lock, NULL, MUTEX_DEFAULT, NULL);
2096 cv_init(&snfq->snfq_cv, NULL, CV_DEFAULT, NULL);
2097 snfq->snfq_max_threads = max_ncpus;
2098 snfq_timeout = SNFQ_TIMEOUT;
2099 /* Cache all files by default. */
2100 sendfile_max_size = MAXOFFSET_T;
2101 }
2102
2103 /*
2104 * Queues a mblk_t for network processing.
2105 */
2106 static void
2107 snf_enque(snf_req_t *sr, mblk_t *mp)
2108 {
2109 mp->b_next = NULL;
2110 mutex_enter(&sr->sr_lock);
2111 if (sr->sr_mp_head == NULL) {
2112 sr->sr_mp_head = sr->sr_mp_tail = mp;
2113 cv_signal(&sr->sr_cv);
2114 } else {
2115 sr->sr_mp_tail->b_next = mp;
2116 sr->sr_mp_tail = mp;
2117 }
2118 sr->sr_qlen += MBLKL(mp);
2119 while ((sr->sr_qlen > sr->sr_hiwat) &&
2120 (sr->sr_write_error == 0)) {
2121 sf_stats.ss_full_waits++;
2122 cv_wait(&sr->sr_cv, &sr->sr_lock);
2123 }
2124 mutex_exit(&sr->sr_lock);
2125 }
2126
2127 /*
2128 * De-queues a mblk_t for network processing.
2129 */
2130 static mblk_t *
2131 snf_deque(snf_req_t *sr)
2132 {
2133 mblk_t *mp;
2134
2135 mutex_enter(&sr->sr_lock);
2136 /*
2137 * If we have encountered an error on read or read is
2138 * completed and no more mblks, return NULL.
2139 * We need to check for NULL sr_mp_head also as
2140 * the reads could have completed and there is
2141 * nothing more to come.
2142 */
2143 if (((sr->sr_read_error & ~SR_READ_DONE) != 0) ||
2144 ((sr->sr_read_error & SR_READ_DONE) &&
2145 sr->sr_mp_head == NULL)) {
2146 mutex_exit(&sr->sr_lock);
2147 return (NULL);
2148 }
2149 /*
2150 * To start with neither SR_READ_DONE is marked nor
2151 * the error is set. When we wake up from cv_wait,
2152 * following are the possibilities :
2153 *
2154 * a) sr_read_error is zero and mblks are queued.
2155 * b) sr_read_error is set to SR_READ_DONE
2156 * and mblks are queued.
2157 * c) sr_read_error is set to SR_READ_DONE
2158 * and no mblks.
2159 * d) sr_read_error is set to some error other
2160 * than SR_READ_DONE.
2161 */
2162
2163 while ((sr->sr_read_error == 0) && (sr->sr_mp_head == NULL)) {
2164 sf_stats.ss_empty_waits++;
2165 cv_wait(&sr->sr_cv, &sr->sr_lock);
2166 }
2167 /* Handle (a) and (b) first - the normal case. */
2168 if (((sr->sr_read_error & ~SR_READ_DONE) == 0) &&
2169 (sr->sr_mp_head != NULL)) {
2170 mp = sr->sr_mp_head;
2171 sr->sr_mp_head = mp->b_next;
2172 sr->sr_qlen -= MBLKL(mp);
2173 if (sr->sr_qlen < sr->sr_lowat)
2174 cv_signal(&sr->sr_cv);
2175 mutex_exit(&sr->sr_lock);
2176 mp->b_next = NULL;
2177 return (mp);
2178 }
2179 /* Handle (c) and (d). */
2180 mutex_exit(&sr->sr_lock);
2181 return (NULL);
2182 }
2183
2184 /*
2185 * Reads data from the filesystem and queues it for network processing.
2186 */
2187 void
2188 snf_async_read(snf_req_t *sr)
2189 {
2190 size_t iosize;
2191 u_offset_t fileoff;
2192 u_offset_t size;
2193 int ret_size;
2194 int error;
2195 file_t *fp;
2196 mblk_t *mp;
2197 struct vnode *vp;
2198 int extra = 0;
2199 int maxblk = 0;
2200 int wroff = 0;
2201 struct sonode *so;
2202
2203 fp = sr->sr_fp;
2204 size = sr->sr_file_size;
2205 fileoff = sr->sr_file_off;
2206
2207 /*
2208 * Ignore the error for filesystems that doesn't support DIRECTIO.
2209 */
2210 (void) VOP_IOCTL(fp->f_vnode, _FIODIRECTIO, DIRECTIO_ON, 0,
2211 kcred, NULL, NULL);
2212
2213 vp = sr->sr_vp;
2214 if (vp->v_type == VSOCK) {
2215 stdata_t *stp;
2216
2217 /*
2218 * Get the extra space to insert a header and a trailer.
2219 */
2220 so = VTOSO(vp);
2221 stp = vp->v_stream;
2222 if (stp == NULL) {
2223 wroff = so->so_proto_props.sopp_wroff;
2224 maxblk = so->so_proto_props.sopp_maxblk;
2225 extra = wroff + so->so_proto_props.sopp_tail;
2226 } else {
2227 wroff = (int)(stp->sd_wroff);
2228 maxblk = (int)(stp->sd_maxblk);
2229 extra = wroff + (int)(stp->sd_tail);
2230 }
2231 }
2232
2233 while ((size != 0) && (sr->sr_write_error == 0)) {
2234
2235 iosize = (int)MIN(sr->sr_maxpsz, size);
2236
2237 /*
2238 * Socket filters can limit the mblk size,
2239 * so limit reads to maxblk if there are
2240 * filters present.
2241 */
2242 if (vp->v_type == VSOCK &&
2243 so->so_filter_active > 0 && maxblk != INFPSZ)
2244 iosize = (int)MIN(iosize, maxblk);
2245
2246 if (is_system_labeled()) {
2247 mp = allocb_cred(iosize + extra, CRED(),
2248 curproc->p_pid);
2249 } else {
2250 mp = allocb(iosize + extra, BPRI_MED);
2251 }
2252 if (mp == NULL) {
2253 error = EAGAIN;
2254 break;
2255 }
2256
2257 mp->b_rptr += wroff;
2258
2259 ret_size = soreadfile(fp, mp->b_rptr, fileoff, &error, iosize);
2260
2261 /* Error or Reached EOF ? */
2262 if ((error != 0) || (ret_size == 0)) {
2263 freeb(mp);
2264 break;
2265 }
2266 mp->b_wptr = mp->b_rptr + ret_size;
2267
2268 snf_enque(sr, mp);
2269 size -= ret_size;
2270 fileoff += ret_size;
2271 }
2272 (void) VOP_IOCTL(fp->f_vnode, _FIODIRECTIO, DIRECTIO_OFF, 0,
2273 kcred, NULL, NULL);
2274 mutex_enter(&sr->sr_lock);
2275 sr->sr_read_error = error;
2276 sr->sr_read_error |= SR_READ_DONE;
2277 cv_signal(&sr->sr_cv);
2278 mutex_exit(&sr->sr_lock);
2279 }
2280
2281 void
2282 snf_async_thread(void)
2283 {
2284 snf_req_t *sr;
2285 callb_cpr_t cprinfo;
2286 clock_t time_left = 1;
2287
2288 CALLB_CPR_INIT(&cprinfo, &snfq->snfq_lock, callb_generic_cpr, "snfq");
2289
2290 mutex_enter(&snfq->snfq_lock);
2291 for (;;) {
2292 /*
2293 * If we didn't find a entry, then block until woken up
2294 * again and then look through the queues again.
2295 */
2296 while ((sr = snfq->snfq_req_head) == NULL) {
2297 CALLB_CPR_SAFE_BEGIN(&cprinfo);
2298 if (time_left <= 0) {
2299 snfq->snfq_svc_threads--;
2300 CALLB_CPR_EXIT(&cprinfo);
2301 thread_exit();
2302 /* NOTREACHED */
2303 }
2304 snfq->snfq_idle_cnt++;
2305
2306 time_left = cv_reltimedwait(&snfq->snfq_cv,
2307 &snfq->snfq_lock, snfq_timeout, TR_CLOCK_TICK);
2308 snfq->snfq_idle_cnt--;
2309
2310 CALLB_CPR_SAFE_END(&cprinfo, &snfq->snfq_lock);
2311 }
2312 snfq->snfq_req_head = sr->sr_next;
2313 snfq->snfq_req_cnt--;
2314 mutex_exit(&snfq->snfq_lock);
2315 snf_async_read(sr);
2316 mutex_enter(&snfq->snfq_lock);
2317 }
2318 }
2319
2320
2321 snf_req_t *
2322 create_thread(int operation, struct vnode *vp, file_t *fp,
2323 u_offset_t fileoff, u_offset_t size)
2324 {
2325 snf_req_t *sr;
2326 stdata_t *stp;
2327
2328 sr = (snf_req_t *)kmem_zalloc(sizeof (snf_req_t), KM_SLEEP);
2329
2330 sr->sr_vp = vp;
2331 sr->sr_fp = fp;
2332 stp = vp->v_stream;
2333
2334 /*
2335 * store sd_qn_maxpsz into sr_maxpsz while we have stream head.
2336 * stream might be closed before thread returns from snf_async_read.
2337 */
2338 if (stp != NULL && stp->sd_qn_maxpsz > 0) {
2339 sr->sr_maxpsz = MIN(MAXBSIZE, stp->sd_qn_maxpsz);
2340 } else {
2341 sr->sr_maxpsz = MAXBSIZE;
2342 }
2343
2344 sr->sr_operation = operation;
2345 sr->sr_file_off = fileoff;
2346 sr->sr_file_size = size;
2347 sr->sr_hiwat = sendfile_req_hiwat;
2348 sr->sr_lowat = sendfile_req_lowat;
2349 mutex_init(&sr->sr_lock, NULL, MUTEX_DEFAULT, NULL);
2350 cv_init(&sr->sr_cv, NULL, CV_DEFAULT, NULL);
2351 /*
2352 * See whether we need another thread for servicing this
2353 * request. If there are already enough requests queued
2354 * for the threads, create one if not exceeding
2355 * snfq_max_threads.
2356 */
2357 mutex_enter(&snfq->snfq_lock);
2358 if (snfq->snfq_req_cnt >= snfq->snfq_idle_cnt &&
2359 snfq->snfq_svc_threads < snfq->snfq_max_threads) {
2360 (void) thread_create(NULL, 0, &snf_async_thread, 0, 0, &p0,
2361 TS_RUN, minclsyspri);
2362 snfq->snfq_svc_threads++;
2363 }
2364 if (snfq->snfq_req_head == NULL) {
2365 snfq->snfq_req_head = snfq->snfq_req_tail = sr;
2366 cv_signal(&snfq->snfq_cv);
2367 } else {
2368 snfq->snfq_req_tail->sr_next = sr;
2369 snfq->snfq_req_tail = sr;
2370 }
2371 snfq->snfq_req_cnt++;
2372 mutex_exit(&snfq->snfq_lock);
2373 return (sr);
2374 }
2375
2376 int
2377 snf_direct_io(file_t *fp, file_t *rfp, u_offset_t fileoff, u_offset_t size,
2378 ssize_t *count)
2379 {
2380 snf_req_t *sr;
2381 mblk_t *mp;
2382 int iosize;
2383 int error = 0;
2384 short fflag;
2385 struct vnode *vp;
2386 int ksize;
2387 struct nmsghdr msg;
2388
2389 ksize = 0;
2390 *count = 0;
2391 bzero(&msg, sizeof (msg));
2392
2393 vp = fp->f_vnode;
2394 fflag = fp->f_flag;
2395 if ((sr = create_thread(READ_OP, vp, rfp, fileoff, size)) == NULL)
2396 return (EAGAIN);
2397
2398 /*
2399 * We check for read error in snf_deque. It has to check
2400 * for successful READ_DONE and return NULL, and we might
2401 * as well make an additional check there.
2402 */
2403 while ((mp = snf_deque(sr)) != NULL) {
2404
2405 if (ISSIG(curthread, JUSTLOOKING)) {
2406 freeb(mp);
2407 error = EINTR;
2408 break;
2409 }
2410 iosize = MBLKL(mp);
2411
2412 error = socket_sendmblk(VTOSO(vp), &msg, fflag, CRED(), &mp);
2413
2414 if (error != 0) {
2415 if (mp != NULL)
2416 freeb(mp);
2417 break;
2418 }
2419 ksize += iosize;
2420 }
2421 *count = ksize;
2422
2423 mutex_enter(&sr->sr_lock);
2424 sr->sr_write_error = error;
2425 /* Look at the big comments on why we cv_signal here. */
2426 cv_signal(&sr->sr_cv);
2427
2428 /* Wait for the reader to complete always. */
2429 while (!(sr->sr_read_error & SR_READ_DONE)) {
2430 cv_wait(&sr->sr_cv, &sr->sr_lock);
2431 }
2432 /* If there is no write error, check for read error. */
2433 if (error == 0)
2434 error = (sr->sr_read_error & ~SR_READ_DONE);
2435
2436 if (error != 0) {
2437 mblk_t *next_mp;
2438
2439 mp = sr->sr_mp_head;
2440 while (mp != NULL) {
2441 next_mp = mp->b_next;
2442 mp->b_next = NULL;
2443 freeb(mp);
2444 mp = next_mp;
2445 }
2446 }
2447 mutex_exit(&sr->sr_lock);
2448 kmem_free(sr, sizeof (snf_req_t));
2449 return (error);
2450 }
2451
2452 /* Maximum no.of pages allocated by vpm for sendfile at a time */
2453 #define SNF_VPMMAXPGS (VPMMAXPGS/2)
2454
2455 /*
2456 * Maximum no.of elements in the list returned by vpm, including
2457 * NULL for the last entry
2458 */
2459 #define SNF_MAXVMAPS (SNF_VPMMAXPGS + 1)
2460
2461 typedef struct {
2462 unsigned int snfv_ref;
2463 frtn_t snfv_frtn;
2464 vnode_t *snfv_vp;
2465 struct vmap snfv_vml[SNF_MAXVMAPS];
2466 } snf_vmap_desbinfo;
2467
2468 typedef struct {
2469 frtn_t snfi_frtn;
2470 caddr_t snfi_base;
2471 uint_t snfi_mapoff;
2472 size_t snfi_len;
2473 vnode_t *snfi_vp;
2474 } snf_smap_desbinfo;
2475
2476 /*
2477 * The callback function used for vpm mapped mblks called when the last ref of
2478 * the mblk is dropped which normally occurs when TCP receives the ack. But it
2479 * can be the driver too due to lazy reclaim.
2480 */
2481 void
2482 snf_vmap_desbfree(snf_vmap_desbinfo *snfv)
2483 {
2484 ASSERT(snfv->snfv_ref != 0);
2485 if (atomic_dec_32_nv(&snfv->snfv_ref) == 0) {
2486 vpm_unmap_pages(snfv->snfv_vml, S_READ);
2487 VN_RELE(snfv->snfv_vp);
2488 kmem_free(snfv, sizeof (snf_vmap_desbinfo));
2489 }
2490 }
2491
2492 /*
2493 * The callback function used for segmap'ped mblks called when the last ref of
2494 * the mblk is dropped which normally occurs when TCP receives the ack. But it
2495 * can be the driver too due to lazy reclaim.
2496 */
2497 void
2498 snf_smap_desbfree(snf_smap_desbinfo *snfi)
2499 {
2500 if (! IS_KPM_ADDR(snfi->snfi_base)) {
2501 /*
2502 * We don't need to call segmap_fault(F_SOFTUNLOCK) for
2503 * segmap_kpm as long as the latter never falls back to
2504 * "use_segmap_range". (See segmap_getmapflt().)
2505 *
2506 * Using S_OTHER saves an redundant hat_setref() in
2507 * segmap_unlock()
2508 */
2509 (void) segmap_fault(kas.a_hat, segkmap,
2510 (caddr_t)(uintptr_t)(((uintptr_t)snfi->snfi_base +
2511 snfi->snfi_mapoff) & PAGEMASK), snfi->snfi_len,
2512 F_SOFTUNLOCK, S_OTHER);
2513 }
2514 (void) segmap_release(segkmap, snfi->snfi_base, SM_DONTNEED);
2515 VN_RELE(snfi->snfi_vp);
2516 kmem_free(snfi, sizeof (*snfi));
2517 }
2518
2519 /*
2520 * Use segmap or vpm instead of bcopy to send down a desballoca'ed, mblk.
2521 * When segmap is used, the mblk contains a segmap slot of no more
2522 * than MAXBSIZE.
2523 *
2524 * With vpm, a maximum of SNF_MAXVMAPS page-sized mappings can be obtained
2525 * in each iteration and sent by socket_sendmblk until an error occurs or
2526 * the requested size has been transferred. An mblk is esballoca'ed from
2527 * each mapped page and a chain of these mblk is sent to the transport layer.
2528 * vpm will be called to unmap the pages when all mblks have been freed by
2529 * free_func.
2530 *
2531 * At the end of the whole sendfile() operation, we wait till the data from
2532 * the last mblk is ack'ed by the transport before returning so that the
2533 * caller of sendfile() can safely modify the file content.
2534 */
2535 int
2536 snf_segmap(file_t *fp, vnode_t *fvp, u_offset_t fileoff, u_offset_t total_size,
2537 ssize_t *count, boolean_t nowait)
2538 {
2539 caddr_t base;
2540 int mapoff;
2541 vnode_t *vp;
2542 mblk_t *mp = NULL;
2543 int chain_size;
2544 int error;
2545 clock_t deadlk_wait;
2546 short fflag;
2547 int ksize;
2548 struct vattr va;
2549 boolean_t dowait = B_FALSE;
2550 struct nmsghdr msg;
2551
2552 vp = fp->f_vnode;
2553 fflag = fp->f_flag;
2554 ksize = 0;
2555 bzero(&msg, sizeof (msg));
2556
2557 for (;;) {
2558 if (ISSIG(curthread, JUSTLOOKING)) {
2559 error = EINTR;
2560 break;
2561 }
2562
2563 if (vpm_enable) {
2564 snf_vmap_desbinfo *snfv;
2565 mblk_t *nmp;
2566 int mblk_size;
2567 int maxsize;
2568 int i;
2569
2570 mapoff = fileoff & PAGEOFFSET;
2571 maxsize = MIN((SNF_VPMMAXPGS * PAGESIZE), total_size);
2572
2573 snfv = kmem_zalloc(sizeof (snf_vmap_desbinfo),
2574 KM_SLEEP);
2575
2576 /*
2577 * Get vpm mappings for maxsize with read access.
2578 * If the pages aren't available yet, we get
2579 * DEADLK, so wait and try again a little later using
2580 * an increasing wait. We might be here a long time.
2581 *
2582 * If delay_sig returns EINTR, be sure to exit and
2583 * pass it up to the caller.
2584 */
2585 deadlk_wait = 0;
2586 while ((error = vpm_map_pages(fvp, fileoff,
2587 (size_t)maxsize, (VPM_FETCHPAGE), snfv->snfv_vml,
2588 SNF_MAXVMAPS, NULL, S_READ)) == EDEADLK) {
2589 deadlk_wait += (deadlk_wait < 5) ? 1 : 4;
2590 if ((error = delay_sig(deadlk_wait)) != 0) {
2591 break;
2592 }
2593 }
2594 if (error != 0) {
2595 kmem_free(snfv, sizeof (snf_vmap_desbinfo));
2596 error = (error == EINTR) ? EINTR : EIO;
2597 goto out;
2598 }
2599 snfv->snfv_frtn.free_func = snf_vmap_desbfree;
2600 snfv->snfv_frtn.free_arg = (caddr_t)snfv;
2601
2602 /* Construct the mblk chain from the page mappings */
2603 chain_size = 0;
2604 for (i = 0; (snfv->snfv_vml[i].vs_addr != NULL) &&
2605 total_size > 0; i++) {
2606 ASSERT(chain_size < maxsize);
2607 mblk_size = MIN(snfv->snfv_vml[i].vs_len -
2608 mapoff, total_size);
2609 nmp = esballoca(
2610 (uchar_t *)snfv->snfv_vml[i].vs_addr +
2611 mapoff, mblk_size, BPRI_HI,
2612 &snfv->snfv_frtn);
2613
2614 /*
2615 * We return EAGAIN after unmapping the pages
2616 * if we cannot allocate the the head of the
2617 * chain. Otherwise, we continue sending the
2618 * mblks constructed so far.
2619 */
2620 if (nmp == NULL) {
2621 if (i == 0) {
2622 vpm_unmap_pages(snfv->snfv_vml,
2623 S_READ);
2624 kmem_free(snfv,
2625 sizeof (snf_vmap_desbinfo));
2626 error = EAGAIN;
2627 goto out;
2628 }
2629 break;
2630 }
2631 /* Mark this dblk with the zero-copy flag */
2632 nmp->b_datap->db_struioflag |= STRUIO_ZC;
2633 nmp->b_wptr += mblk_size;
2634 chain_size += mblk_size;
2635 fileoff += mblk_size;
2636 total_size -= mblk_size;
2637 snfv->snfv_ref++;
2638 mapoff = 0;
2639 if (i > 0)
2640 linkb(mp, nmp);
2641 else
2642 mp = nmp;
2643 }
2644 VN_HOLD(fvp);
2645 snfv->snfv_vp = fvp;
2646 } else {
2647 /* vpm not supported. fallback to segmap */
2648 snf_smap_desbinfo *snfi;
2649
2650 mapoff = fileoff & MAXBOFFSET;
2651 chain_size = MAXBSIZE - mapoff;
2652 if (chain_size > total_size)
2653 chain_size = total_size;
2654 /*
2655 * we don't forcefault because we'll call
2656 * segmap_fault(F_SOFTLOCK) next.
2657 *
2658 * S_READ will get the ref bit set (by either
2659 * segmap_getmapflt() or segmap_fault()) and page
2660 * shared locked.
2661 */
2662 base = segmap_getmapflt(segkmap, fvp, fileoff,
2663 chain_size, segmap_kpm ? SM_FAULT : 0, S_READ);
2664
2665 snfi = kmem_alloc(sizeof (*snfi), KM_SLEEP);
2666 snfi->snfi_len = (size_t)roundup(mapoff+chain_size,
2667 PAGESIZE)- (mapoff & PAGEMASK);
2668 /*
2669 * We must call segmap_fault() even for segmap_kpm
2670 * because that's how error gets returned.
2671 * (segmap_getmapflt() never fails but segmap_fault()
2672 * does.)
2673 *
2674 * If the pages aren't available yet, we get
2675 * DEADLK, so wait and try again a little later using
2676 * an increasing wait. We might be here a long time.
2677 *
2678 * If delay_sig returns EINTR, be sure to exit and
2679 * pass it up to the caller.
2680 */
2681 deadlk_wait = 0;
2682 while ((error = FC_ERRNO(segmap_fault(kas.a_hat,
2683 segkmap, (caddr_t)(uintptr_t)(((uintptr_t)base +
2684 mapoff) & PAGEMASK), snfi->snfi_len, F_SOFTLOCK,
2685 S_READ))) == EDEADLK) {
2686 deadlk_wait += (deadlk_wait < 5) ? 1 : 4;
2687 if ((error = delay_sig(deadlk_wait)) != 0) {
2688 break;
2689 }
2690 }
2691 if (error != 0) {
2692 (void) segmap_release(segkmap, base, 0);
2693 kmem_free(snfi, sizeof (*snfi));
2694 error = (error == EINTR) ? EINTR : EIO;
2695 goto out;
2696 }
2697 snfi->snfi_frtn.free_func = snf_smap_desbfree;
2698 snfi->snfi_frtn.free_arg = (caddr_t)snfi;
2699 snfi->snfi_base = base;
2700 snfi->snfi_mapoff = mapoff;
2701 mp = esballoca((uchar_t *)base + mapoff, chain_size,
2702 BPRI_HI, &snfi->snfi_frtn);
2703
2704 if (mp == NULL) {
2705 (void) segmap_fault(kas.a_hat, segkmap,
2706 (caddr_t)(uintptr_t)(((uintptr_t)base +
2707 mapoff) & PAGEMASK), snfi->snfi_len,
2708 F_SOFTUNLOCK, S_OTHER);
2709 (void) segmap_release(segkmap, base, 0);
2710 kmem_free(snfi, sizeof (*snfi));
2711 freemsg(mp);
2712 error = EAGAIN;
2713 goto out;
2714 }
2715 VN_HOLD(fvp);
2716 snfi->snfi_vp = fvp;
2717 mp->b_wptr += chain_size;
2718
2719 /* Mark this dblk with the zero-copy flag */
2720 mp->b_datap->db_struioflag |= STRUIO_ZC;
2721 fileoff += chain_size;
2722 total_size -= chain_size;
2723 }
2724
2725 if (total_size == 0 && !nowait) {
2726 ASSERT(!dowait);
2727 dowait = B_TRUE;
2728 mp->b_datap->db_struioflag |= STRUIO_ZCNOTIFY;
2729 }
2730 VOP_RWUNLOCK(fvp, V_WRITELOCK_FALSE, NULL);
2731 error = socket_sendmblk(VTOSO(vp), &msg, fflag, CRED(), &mp);
2732 if (error != 0) {
2733 /*
2734 * mp contains the mblks that were not sent by
2735 * socket_sendmblk. Use its size to update *count
2736 */
2737 *count = ksize + (chain_size - msgdsize(mp));
2738 if (mp != NULL)
2739 freemsg(mp);
2740 return (error);
2741 }
2742 ksize += chain_size;
2743 if (total_size == 0)
2744 goto done;
2745
2746 (void) VOP_RWLOCK(fvp, V_WRITELOCK_FALSE, NULL);
2747 va.va_mask = AT_SIZE;
2748 error = VOP_GETATTR(fvp, &va, 0, kcred, NULL);
2749 if (error)
2750 break;
2751 /* Read as much as possible. */
2752 if (fileoff >= va.va_size)
2753 break;
2754 if (total_size + fileoff > va.va_size)
2755 total_size = va.va_size - fileoff;
2756 }
2757 out:
2758 VOP_RWUNLOCK(fvp, V_WRITELOCK_FALSE, NULL);
2759 done:
2760 *count = ksize;
2761 if (dowait) {
2762 stdata_t *stp;
2763
2764 stp = vp->v_stream;
2765 if (stp == NULL) {
2766 struct sonode *so;
2767 so = VTOSO(vp);
2768 error = so_zcopy_wait(so);
2769 } else {
2770 mutex_enter(&stp->sd_lock);
2771 while (!(stp->sd_flag & STZCNOTIFY)) {
2772 if (cv_wait_sig(&stp->sd_zcopy_wait,
2773 &stp->sd_lock) == 0) {
2774 error = EINTR;
2775 break;
2776 }
2777 }
2778 stp->sd_flag &= ~STZCNOTIFY;
2779 mutex_exit(&stp->sd_lock);
2780 }
2781 }
2782 return (error);
2783 }
2784
2785 int
2786 snf_cache(file_t *fp, vnode_t *fvp, u_offset_t fileoff, u_offset_t size,
2787 uint_t maxpsz, ssize_t *count)
2788 {
2789 struct vnode *vp;
2790 mblk_t *mp;
2791 int iosize;
2792 int extra = 0;
2793 int error;
2794 short fflag;
2795 int ksize;
2796 int ioflag;
2797 struct uio auio;
2798 struct iovec aiov;
2799 struct vattr va;
2800 int maxblk = 0;
2801 int wroff = 0;
2802 struct sonode *so;
2803 struct nmsghdr msg;
2804
2805 vp = fp->f_vnode;
2806 if (vp->v_type == VSOCK) {
2807 stdata_t *stp;
2808
2809 /*
2810 * Get the extra space to insert a header and a trailer.
2811 */
2812 so = VTOSO(vp);
2813 stp = vp->v_stream;
2814 if (stp == NULL) {
2815 wroff = so->so_proto_props.sopp_wroff;
2816 maxblk = so->so_proto_props.sopp_maxblk;
2817 extra = wroff + so->so_proto_props.sopp_tail;
2818 } else {
2819 wroff = (int)(stp->sd_wroff);
2820 maxblk = (int)(stp->sd_maxblk);
2821 extra = wroff + (int)(stp->sd_tail);
2822 }
2823 }
2824 bzero(&msg, sizeof (msg));
2825 fflag = fp->f_flag;
2826 ksize = 0;
2827 auio.uio_iov = &aiov;
2828 auio.uio_iovcnt = 1;
2829 auio.uio_segflg = UIO_SYSSPACE;
2830 auio.uio_llimit = MAXOFFSET_T;
2831 auio.uio_fmode = fflag;
2832 auio.uio_extflg = UIO_COPY_CACHED;
2833 ioflag = auio.uio_fmode & (FSYNC|FDSYNC|FRSYNC);
2834 /* If read sync is not asked for, filter sync flags */
2835 if ((ioflag & FRSYNC) == 0)
2836 ioflag &= ~(FSYNC|FDSYNC);
2837 for (;;) {
2838 if (ISSIG(curthread, JUSTLOOKING)) {
2839 error = EINTR;
2840 break;
2841 }
2842 iosize = (int)MIN(maxpsz, size);
2843
2844 /*
2845 * Socket filters can limit the mblk size,
2846 * so limit reads to maxblk if there are
2847 * filters present.
2848 */
2849 if (vp->v_type == VSOCK &&
2850 so->so_filter_active > 0 && maxblk != INFPSZ)
2851 iosize = (int)MIN(iosize, maxblk);
2852
2853 if (is_system_labeled()) {
2854 mp = allocb_cred(iosize + extra, CRED(),
2855 curproc->p_pid);
2856 } else {
2857 mp = allocb(iosize + extra, BPRI_MED);
2858 }
2859 if (mp == NULL) {
2860 error = EAGAIN;
2861 break;
2862 }
2863
2864 mp->b_rptr += wroff;
2865
2866 aiov.iov_base = (caddr_t)mp->b_rptr;
2867 aiov.iov_len = iosize;
2868 auio.uio_loffset = fileoff;
2869 auio.uio_resid = iosize;
2870
2871 error = VOP_READ(fvp, &auio, ioflag, fp->f_cred, NULL);
2872 iosize -= auio.uio_resid;
2873
2874 if (error == EINTR && iosize != 0)
2875 error = 0;
2876
2877 if (error != 0 || iosize == 0) {
2878 freeb(mp);
2879 break;
2880 }
2881 mp->b_wptr = mp->b_rptr + iosize;
2882
2883 VOP_RWUNLOCK(fvp, V_WRITELOCK_FALSE, NULL);
2884
2885 error = socket_sendmblk(VTOSO(vp), &msg, fflag, CRED(), &mp);
2886
2887 if (error != 0) {
2888 *count = ksize;
2889 if (mp != NULL)
2890 freeb(mp);
2891 return (error);
2892 }
2893 ksize += iosize;
2894 size -= iosize;
2895 if (size == 0)
2896 goto done;
2897
2898 fileoff += iosize;
2899 (void) VOP_RWLOCK(fvp, V_WRITELOCK_FALSE, NULL);
2900 va.va_mask = AT_SIZE;
2901 error = VOP_GETATTR(fvp, &va, 0, kcred, NULL);
2902 if (error)
2903 break;
2904 /* Read as much as possible. */
2905 if (fileoff >= va.va_size)
2906 size = 0;
2907 else if (size + fileoff > va.va_size)
2908 size = va.va_size - fileoff;
2909 }
2910 VOP_RWUNLOCK(fvp, V_WRITELOCK_FALSE, NULL);
2911 done:
2912 *count = ksize;
2913 return (error);
2914 }
2915
2916 #if defined(_SYSCALL32_IMPL) || defined(_ILP32)
2917 /*
2918 * Largefile support for 32 bit applications only.
2919 */
2920 int
2921 sosendfile64(file_t *fp, file_t *rfp, const struct ksendfilevec64 *sfv,
2922 ssize32_t *count32)
2923 {
2924 ssize32_t sfv_len;
2925 u_offset_t sfv_off, va_size;
2926 struct vnode *vp, *fvp, *realvp;
2927 struct vattr va;
2928 stdata_t *stp;
2929 ssize_t count = 0;
2930 int error = 0;
2931 boolean_t dozcopy = B_FALSE;
2932 uint_t maxpsz;
2933
2934 sfv_len = (ssize32_t)sfv->sfv_len;
2935 if (sfv_len < 0) {
2936 error = EINVAL;
2937 goto out;
2938 }
2939
2940 if (sfv_len == 0) goto out;
2941
2942 sfv_off = (u_offset_t)sfv->sfv_off;
2943
2944 /* Same checks as in pread */
2945 if (sfv_off > MAXOFFSET_T) {
2946 error = EINVAL;
2947 goto out;
2948 }
2949 if (sfv_off + sfv_len > MAXOFFSET_T)
2950 sfv_len = (ssize32_t)(MAXOFFSET_T - sfv_off);
2951
2952 /*
2953 * There are no more checks on sfv_len. So, we cast it to
2954 * u_offset_t and share the snf_direct_io/snf_cache code between
2955 * 32 bit and 64 bit.
2956 *
2957 * TODO: should do nbl_need_check() like read()?
2958 */
2959 if (sfv_len > sendfile_max_size) {
2960 sf_stats.ss_file_not_cached++;
2961 error = snf_direct_io(fp, rfp, sfv_off, (u_offset_t)sfv_len,
2962 &count);
2963 goto out;
2964 }
2965 fvp = rfp->f_vnode;
2966 if (VOP_REALVP(fvp, &realvp, NULL) == 0)
2967 fvp = realvp;
2968 /*
2969 * Grab the lock as a reader to prevent the file size
2970 * from changing underneath.
2971 */
2972 (void) VOP_RWLOCK(fvp, V_WRITELOCK_FALSE, NULL);
2973 va.va_mask = AT_SIZE;
2974 error = VOP_GETATTR(fvp, &va, 0, kcred, NULL);
2975 va_size = va.va_size;
2976 if ((error != 0) || (va_size == 0) || (sfv_off >= va_size)) {
2977 VOP_RWUNLOCK(fvp, V_WRITELOCK_FALSE, NULL);
2978 goto out;
2979 }
2980 /* Read as much as possible. */
2981 if (sfv_off + sfv_len > va_size)
2982 sfv_len = va_size - sfv_off;
2983
2984 vp = fp->f_vnode;
2985 stp = vp->v_stream;
2986 /*
2987 * When the NOWAIT flag is not set, we enable zero-copy only if the
2988 * transfer size is large enough. This prevents performance loss
2989 * when the caller sends the file piece by piece.
2990 */
2991 if (sfv_len >= MAXBSIZE && (sfv_len >= (va_size >> 1) ||
2992 (sfv->sfv_flag & SFV_NOWAIT) || sfv_len >= 0x1000000) &&
2993 !vn_has_flocks(fvp) && !(fvp->v_flag & VNOMAP)) {
2994 uint_t copyflag;
2995 copyflag = stp != NULL ? stp->sd_copyflag :
2996 VTOSO(vp)->so_proto_props.sopp_zcopyflag;
2997 if ((copyflag & (STZCVMSAFE|STZCVMUNSAFE)) == 0) {
2998 int on = 1;
2999
3000 if (socket_setsockopt(VTOSO(vp), SOL_SOCKET,
3001 SO_SND_COPYAVOID, &on, sizeof (on), CRED()) == 0)
3002 dozcopy = B_TRUE;
3003 } else {
3004 dozcopy = copyflag & STZCVMSAFE;
3005 }
3006 }
3007 if (dozcopy) {
3008 sf_stats.ss_file_segmap++;
3009 error = snf_segmap(fp, fvp, sfv_off, (u_offset_t)sfv_len,
3010 &count, ((sfv->sfv_flag & SFV_NOWAIT) != 0));
3011 } else {
3012 if (vp->v_type == VSOCK && stp == NULL) {
3013 sonode_t *so = VTOSO(vp);
3014 maxpsz = so->so_proto_props.sopp_maxpsz;
3015 } else if (stp != NULL) {
3016 maxpsz = stp->sd_qn_maxpsz;
3017 } else {
3018 maxpsz = maxphys;
3019 }
3020
3021 if (maxpsz == INFPSZ)
3022 maxpsz = maxphys;
3023 else
3024 maxpsz = roundup(maxpsz, MAXBSIZE);
3025 sf_stats.ss_file_cached++;
3026 error = snf_cache(fp, fvp, sfv_off, (u_offset_t)sfv_len,
3027 maxpsz, &count);
3028 }
3029 out:
3030 releasef(sfv->sfv_fd);
3031 *count32 = (ssize32_t)count;
3032 return (error);
3033 }
3034 #endif
3035
3036 #ifdef _SYSCALL32_IMPL
3037 /*
3038 * recv32(), recvfrom32(), send32(), sendto32(): intentionally return a
3039 * ssize_t rather than ssize32_t; see the comments above read32 for details.
3040 */
3041
3042 ssize_t
3043 recv32(int32_t sock, caddr32_t buffer, size32_t len, int32_t flags)
3044 {
3045 return (recv(sock, (void *)(uintptr_t)buffer, (ssize32_t)len, flags));
3046 }
3047
3048 ssize_t
3049 recvfrom32(int32_t sock, caddr32_t buffer, size32_t len, int32_t flags,
3050 caddr32_t name, caddr32_t namelenp)
3051 {
3052 return (recvfrom(sock, (void *)(uintptr_t)buffer, (ssize32_t)len, flags,
3053 (void *)(uintptr_t)name, (void *)(uintptr_t)namelenp));
3054 }
3055
3056 ssize_t
3057 send32(int32_t sock, caddr32_t buffer, size32_t len, int32_t flags)
3058 {
3059 return (send(sock, (void *)(uintptr_t)buffer, (ssize32_t)len, flags));
3060 }
3061
3062 ssize_t
3063 sendto32(int32_t sock, caddr32_t buffer, size32_t len, int32_t flags,
3064 caddr32_t name, socklen_t namelen)
3065 {
3066 return (sendto(sock, (void *)(uintptr_t)buffer, (ssize32_t)len, flags,
3067 (void *)(uintptr_t)name, namelen));
3068 }
3069 #endif /* _SYSCALL32_IMPL */
3070
3071 /*
3072 * Function wrappers (mostly around the sonode switch) for
3073 * backward compatibility.
3074 */
3075
3076 int
3077 soaccept(struct sonode *so, int fflag, struct sonode **nsop)
3078 {
3079 return (socket_accept(so, fflag, CRED(), nsop));
3080 }
3081
3082 int
3083 sobind(struct sonode *so, struct sockaddr *name, socklen_t namelen,
3084 int backlog, int flags)
3085 {
3086 int error;
3087
3088 error = socket_bind(so, name, namelen, flags, CRED());
3089 if (error == 0 && backlog != 0)
3090 return (socket_listen(so, backlog, CRED()));
3091
3092 return (error);
3093 }
3094
3095 int
3096 solisten(struct sonode *so, int backlog)
3097 {
3098 return (socket_listen(so, backlog, CRED()));
3099 }
3100
3101 int
3102 soconnect(struct sonode *so, struct sockaddr *name, socklen_t namelen,
3103 int fflag, int flags)
3104 {
3105 return (socket_connect(so, name, namelen, fflag, flags, CRED()));
3106 }
3107
3108 int
3109 sorecvmsg(struct sonode *so, struct nmsghdr *msg, struct uio *uiop)
3110 {
3111 return (socket_recvmsg(so, msg, uiop, CRED()));
3112 }
3113
3114 int
3115 sosendmsg(struct sonode *so, struct nmsghdr *msg, struct uio *uiop)
3116 {
3117 return (socket_sendmsg(so, msg, uiop, CRED()));
3118 }
3119
3120 int
3121 soshutdown(struct sonode *so, int how)
3122 {
3123 return (socket_shutdown(so, how, CRED()));
3124 }
3125
3126 int
3127 sogetsockopt(struct sonode *so, int level, int option_name, void *optval,
3128 socklen_t *optlenp, int flags)
3129 {
3130 return (socket_getsockopt(so, level, option_name, optval, optlenp,
3131 flags, CRED()));
3132 }
3133
3134 int
3135 sosetsockopt(struct sonode *so, int level, int option_name, const void *optval,
3136 t_uscalar_t optlen)
3137 {
3138 return (socket_setsockopt(so, level, option_name, optval, optlen,
3139 CRED()));
3140 }
3141
3142 /*
3143 * Because this is backward compatibility interface it only needs to be
3144 * able to handle the creation of TPI sockfs sockets.
3145 */
3146 struct sonode *
3147 socreate(struct sockparams *sp, int family, int type, int protocol, int version,
3148 int *errorp)
3149 {
3150 struct sonode *so;
3151
3152 ASSERT(sp != NULL);
3153
3154 so = sp->sp_smod_info->smod_sock_create_func(sp, family, type, protocol,
3155 version, SOCKET_SLEEP, errorp, CRED());
3156 if (so == NULL) {
3157 SOCKPARAMS_DEC_REF(sp);
3158 } else {
3159 if ((*errorp = SOP_INIT(so, NULL, CRED(), SOCKET_SLEEP)) == 0) {
3160 /* Cannot fail, only bumps so_count */
3161 (void) VOP_OPEN(&SOTOV(so), FREAD|FWRITE, CRED(), NULL);
3162 } else {
3163 socket_destroy(so);
3164 so = NULL;
3165 }
3166 }
3167 return (so);
3168 }