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 2015, Joyent, Inc. All rights reserved.
25 * Copyright (c) 2013, OmniTI Computer Consulting, Inc. All rights reserved.
26 * Copyright 2015 Nexenta Systems, Inc. All rights reserved.
27 * Copyright 2020 OmniOS Community Edition (OmniOSce) Association.
28 * Copyright 2022 Garrett D'Amore
29 */
30
31 #include <sys/types.h>
32 #include <sys/t_lock.h>
33 #include <sys/param.h>
34 #include <sys/systm.h>
35 #include <sys/buf.h>
36 #include <sys/conf.h>
37 #include <sys/cred.h>
38 #include <sys/kmem.h>
39 #include <sys/sysmacros.h>
40 #include <sys/vfs.h>
41 #include <sys/vnode.h>
42 #include <sys/debug.h>
43 #include <sys/errno.h>
44 #include <sys/time.h>
45 #include <sys/file.h>
46 #include <sys/user.h>
47 #include <sys/stream.h>
48 #include <sys/strsubr.h>
49 #include <sys/strsun.h>
50 #include <sys/sunddi.h>
51 #include <sys/esunddi.h>
52 #include <sys/flock.h>
53 #include <sys/modctl.h>
54 #include <sys/cmn_err.h>
55 #include <sys/vmsystm.h>
56 #include <sys/policy.h>
57 #include <sys/limits.h>
58
59 #include <sys/socket.h>
60 #include <sys/socketvar.h>
61
62 #include <sys/isa_defs.h>
63 #include <sys/inttypes.h>
64 #include <sys/systm.h>
65 #include <sys/cpuvar.h>
66 #include <sys/filio.h>
67 #include <sys/sendfile.h>
68 #include <sys/ddi.h>
69 #include <vm/seg.h>
70 #include <vm/seg_map.h>
71 #include <vm/seg_kpm.h>
72
73 #include <fs/sockfs/sockcommon.h>
74 #include <fs/sockfs/sockfilter_impl.h>
75 #include <fs/sockfs/socktpi.h>
76
77 #ifdef SOCK_TEST
78 int do_useracc = 1; /* Controlled by setting SO_DEBUG to 4 */
79 #else
80 #define do_useracc 1
81 #endif /* SOCK_TEST */
82
83 extern int xnet_truncate_print;
84
85 /*
86 * Kernel component of socket creation.
87 *
88 * The socket library determines which version number to use.
89 * First the library calls this with a NULL devpath. If this fails
90 * to find a transport (using solookup) the library will look in /etc/netconfig
91 * for the appropriate transport. If one is found it will pass in the
92 * devpath for the kernel to use.
93 */
94 int
so_socket(int family,int type_w_flags,int protocol,char * devpath,int version)95 so_socket(int family, int type_w_flags, int protocol, char *devpath,
96 int version)
97 {
98 struct sonode *so;
99 vnode_t *vp;
100 struct file *fp;
101 int fd;
102 int error;
103 int type;
104
105 type = type_w_flags & SOCK_TYPE_MASK;
106 type_w_flags &= ~SOCK_TYPE_MASK;
107 if (type_w_flags & ~(SOCK_CLOEXEC|SOCK_NDELAY|SOCK_NONBLOCK))
108 return (set_errno(EINVAL));
109
110 if (devpath != NULL) {
111 char *buf;
112 size_t kdevpathlen = 0;
113
114 buf = kmem_alloc(MAXPATHLEN, KM_SLEEP);
115 if ((error = copyinstr(devpath, buf,
116 MAXPATHLEN, &kdevpathlen)) != 0) {
117 kmem_free(buf, MAXPATHLEN);
118 return (set_errno(error));
119 }
120 so = socket_create(family, type, protocol, buf, NULL,
121 SOCKET_SLEEP, version, CRED(), &error);
122 kmem_free(buf, MAXPATHLEN);
123 } else {
124 so = socket_create(family, type, protocol, NULL, NULL,
125 SOCKET_SLEEP, version, CRED(), &error);
126 }
127 if (so == NULL)
128 return (set_errno(error));
129
130 /* Allocate a file descriptor for the socket */
131 vp = SOTOV(so);
132 error = falloc(vp, FWRITE|FREAD, &fp, &fd);
133 if (error != 0) {
134 (void) socket_close(so, 0, CRED());
135 socket_destroy(so);
136 return (set_errno(error));
137 }
138
139 /*
140 * Now fill in the entries that falloc reserved
141 */
142 if (type_w_flags & SOCK_NDELAY) {
143 so->so_state |= SS_NDELAY;
144 fp->f_flag |= FNDELAY;
145 }
146 if (type_w_flags & SOCK_NONBLOCK) {
147 so->so_state |= SS_NONBLOCK;
148 fp->f_flag |= FNONBLOCK;
149 }
150 mutex_exit(&fp->f_tlock);
151 setf(fd, fp);
152 if ((type_w_flags & SOCK_CLOEXEC) != 0) {
153 f_setfd(fd, FD_CLOEXEC);
154 }
155
156 return (fd);
157 }
158
159 /*
160 * Map from a file descriptor to a socket node.
161 * Returns with the file descriptor held i.e. the caller has to
162 * use releasef when done with the file descriptor.
163 */
164 struct sonode *
getsonode(int sock,int * errorp,file_t ** fpp)165 getsonode(int sock, int *errorp, file_t **fpp)
166 {
167 file_t *fp;
168 vnode_t *vp;
169 struct sonode *so;
170
171 if ((fp = getf(sock)) == NULL) {
172 *errorp = EBADF;
173 eprintline(*errorp);
174 return (NULL);
175 }
176 vp = fp->f_vnode;
177 /* Check if it is a socket */
178 if (vp->v_type != VSOCK) {
179 releasef(sock);
180 *errorp = ENOTSOCK;
181 eprintline(*errorp);
182 return (NULL);
183 }
184 /*
185 * Use the stream head to find the real socket vnode.
186 * This is needed when namefs sits above sockfs.
187 */
188 if (vp->v_stream) {
189 ASSERT(vp->v_stream->sd_vnode);
190 vp = vp->v_stream->sd_vnode;
191
192 so = VTOSO(vp);
193 if (so->so_version == SOV_STREAM) {
194 releasef(sock);
195 *errorp = ENOTSOCK;
196 eprintsoline(so, *errorp);
197 return (NULL);
198 }
199 } else {
200 so = VTOSO(vp);
201 }
202 if (fpp)
203 *fpp = fp;
204 return (so);
205 }
206
207 /*
208 * Allocate and copyin a sockaddr.
209 * Ensures NULL termination for AF_UNIX addresses by extending them
210 * with one NULL byte if need be. Verifies that the length is not
211 * excessive to prevent an application from consuming all of kernel
212 * memory. Returns NULL when an error occurred.
213 */
214 static struct sockaddr *
copyin_name(struct sonode * so,struct sockaddr * name,socklen_t * namelenp,int * errorp)215 copyin_name(struct sonode *so, struct sockaddr *name, socklen_t *namelenp,
216 int *errorp)
217 {
218 char *faddr;
219 size_t namelen = (size_t)*namelenp;
220
221 ASSERT(namelen != 0);
222 if (namelen > SO_MAXARGSIZE) {
223 *errorp = EINVAL;
224 eprintsoline(so, *errorp);
225 return (NULL);
226 }
227
228 faddr = (char *)kmem_alloc(namelen, KM_SLEEP);
229 if (copyin(name, faddr, namelen)) {
230 kmem_free(faddr, namelen);
231 *errorp = EFAULT;
232 eprintsoline(so, *errorp);
233 return (NULL);
234 }
235
236 /*
237 * Add space for NULL termination if needed.
238 * Do a quick check if the last byte is NUL.
239 */
240 if (so->so_family == AF_UNIX && faddr[namelen - 1] != '\0') {
241 /* Check if there is any NULL termination */
242 size_t i;
243 int foundnull = 0;
244
245 for (i = sizeof (name->sa_family); i < namelen; i++) {
246 if (faddr[i] == '\0') {
247 foundnull = 1;
248 break;
249 }
250 }
251 if (!foundnull) {
252 /* Add extra byte for NUL padding */
253 char *nfaddr;
254
255 nfaddr = (char *)kmem_alloc(namelen + 1, KM_SLEEP);
256 bcopy(faddr, nfaddr, namelen);
257 kmem_free(faddr, namelen);
258
259 /* NUL terminate */
260 nfaddr[namelen] = '\0';
261 namelen++;
262 ASSERT((socklen_t)namelen == namelen);
263 *namelenp = (socklen_t)namelen;
264 faddr = nfaddr;
265 }
266 }
267 return ((struct sockaddr *)faddr);
268 }
269
270 /*
271 * Copy from kaddr/klen to uaddr/ulen. Updates ulenp if non-NULL.
272 */
273 static int
copyout_arg(void * uaddr,socklen_t ulen,void * ulenp,void * kaddr,socklen_t klen)274 copyout_arg(void *uaddr, socklen_t ulen, void *ulenp, void *kaddr,
275 socklen_t klen)
276 {
277 if (uaddr != NULL) {
278 if (ulen > klen)
279 ulen = klen;
280
281 if (ulen != 0) {
282 if (copyout(kaddr, uaddr, ulen))
283 return (EFAULT);
284 }
285 } else
286 ulen = 0;
287
288 if (ulenp != NULL) {
289 if (copyout(&ulen, ulenp, sizeof (ulen)))
290 return (EFAULT);
291 }
292 return (0);
293 }
294
295 /*
296 * Copy from kaddr/klen to uaddr/ulen. Updates ulenp if non-NULL.
297 * If klen is greater than ulen it still uses the non-truncated
298 * klen to update ulenp.
299 */
300 static int
copyout_name(void * uaddr,socklen_t ulen,void * ulenp,void * kaddr,socklen_t klen)301 copyout_name(void *uaddr, socklen_t ulen, void *ulenp, void *kaddr,
302 socklen_t klen)
303 {
304 if (uaddr != NULL) {
305 if (ulen >= klen)
306 ulen = klen;
307 else if (ulen != 0 && xnet_truncate_print) {
308 printf("sockfs: truncating copyout of address using "
309 "XNET semantics for pid = %d. Lengths %d, %d\n",
310 curproc->p_pid, klen, ulen);
311 }
312
313 if (ulen != 0) {
314 if (copyout(kaddr, uaddr, ulen))
315 return (EFAULT);
316 } else
317 klen = 0;
318 } else
319 klen = 0;
320
321 if (ulenp != NULL) {
322 if (copyout(&klen, ulenp, sizeof (klen)))
323 return (EFAULT);
324 }
325 return (0);
326 }
327
328 /*
329 * The socketpair() code in libsocket creates two sockets (using
330 * the /etc/netconfig fallback if needed) before calling this routine
331 * to connect the two sockets together.
332 *
333 * For a SOCK_STREAM socketpair a listener is needed - in that case this
334 * routine will create a new file descriptor as part of accepting the
335 * connection. The library socketpair() will check if svs[2] has changed
336 * in which case it will close the changed fd.
337 *
338 * Note that this code could use the TPI feature of accepting the connection
339 * on the listening endpoint. However, that would require significant changes
340 * to soaccept.
341 */
342 int
so_socketpair(int sv[2])343 so_socketpair(int sv[2])
344 {
345 int svs[2];
346 struct sonode *so1, *so2;
347 int error;
348 int orig_flags;
349 struct sockaddr_ux *name;
350 size_t namelen;
351 sotpi_info_t *sti1;
352 sotpi_info_t *sti2;
353
354 dprint(1, ("so_socketpair(%p)\n", (void *)sv));
355
356 error = useracc(sv, sizeof (svs), B_WRITE);
357 if (error && do_useracc)
358 return (set_errno(EFAULT));
359
360 if (copyin(sv, svs, sizeof (svs)))
361 return (set_errno(EFAULT));
362
363 if ((so1 = getsonode(svs[0], &error, NULL)) == NULL)
364 return (set_errno(error));
365
366 if ((so2 = getsonode(svs[1], &error, NULL)) == NULL) {
367 releasef(svs[0]);
368 return (set_errno(error));
369 }
370
371 if (so1->so_family != AF_UNIX || so2->so_family != AF_UNIX) {
372 error = EOPNOTSUPP;
373 goto done;
374 }
375
376 sti1 = SOTOTPI(so1);
377 sti2 = SOTOTPI(so2);
378
379 /*
380 * The code below makes assumptions about the "sockfs" implementation.
381 * So make sure that the correct implementation is really used.
382 */
383 ASSERT(so1->so_ops == &sotpi_sonodeops);
384 ASSERT(so2->so_ops == &sotpi_sonodeops);
385
386 if (so1->so_type == SOCK_DGRAM) {
387 /*
388 * Bind both sockets and connect them with each other.
389 * Need to allocate name/namelen for soconnect.
390 */
391 error = socket_bind(so1, NULL, 0, _SOBIND_UNSPEC, CRED());
392 if (error) {
393 eprintsoline(so1, error);
394 goto done;
395 }
396 error = socket_bind(so2, NULL, 0, _SOBIND_UNSPEC, CRED());
397 if (error) {
398 eprintsoline(so2, error);
399 goto done;
400 }
401 namelen = sizeof (struct sockaddr_ux);
402 name = kmem_alloc(namelen, KM_SLEEP);
403 name->sou_family = AF_UNIX;
404 name->sou_addr = sti2->sti_ux_laddr;
405 error = socket_connect(so1,
406 (struct sockaddr *)name,
407 (socklen_t)namelen,
408 0, _SOCONNECT_NOXLATE, CRED());
409 if (error) {
410 kmem_free(name, namelen);
411 eprintsoline(so1, error);
412 goto done;
413 }
414 name->sou_addr = sti1->sti_ux_laddr;
415 error = socket_connect(so2,
416 (struct sockaddr *)name,
417 (socklen_t)namelen,
418 0, _SOCONNECT_NOXLATE, CRED());
419 kmem_free(name, namelen);
420 if (error) {
421 eprintsoline(so2, error);
422 goto done;
423 }
424 releasef(svs[0]);
425 releasef(svs[1]);
426 } else {
427 /*
428 * Bind both sockets, with so1 being a listener.
429 * Connect so2 to so1 - nonblocking to avoid waiting for
430 * soaccept to complete.
431 * Accept a connection on so1. Pass out the new fd as sv[0].
432 * The library will detect the changed fd and close
433 * the original one.
434 */
435 struct sonode *nso;
436 struct vnode *nvp;
437 struct file *nfp;
438 int nfd;
439
440 /*
441 * We could simply call socket_listen() here (which would do the
442 * binding automatically) if the code didn't rely on passing
443 * _SOBIND_NOXLATE to the TPI implementation of socket_bind().
444 */
445 error = socket_bind(so1, NULL, 0, _SOBIND_UNSPEC|
446 _SOBIND_NOXLATE|_SOBIND_LISTEN|_SOBIND_SOCKETPAIR,
447 CRED());
448 if (error) {
449 eprintsoline(so1, error);
450 goto done;
451 }
452 error = socket_bind(so2, NULL, 0, _SOBIND_UNSPEC, CRED());
453 if (error) {
454 eprintsoline(so2, error);
455 goto done;
456 }
457
458 namelen = sizeof (struct sockaddr_ux);
459 name = kmem_alloc(namelen, KM_SLEEP);
460 name->sou_family = AF_UNIX;
461 name->sou_addr = sti1->sti_ux_laddr;
462 error = socket_connect(so2,
463 (struct sockaddr *)name,
464 (socklen_t)namelen,
465 FNONBLOCK, _SOCONNECT_NOXLATE, CRED());
466 kmem_free(name, namelen);
467 if (error) {
468 if (error != EINPROGRESS) {
469 eprintsoline(so2, error); goto done;
470 }
471 }
472
473 error = socket_accept(so1, 0, CRED(), &nso);
474 if (error) {
475 eprintsoline(so1, error);
476 goto done;
477 }
478
479 /* wait for so2 being SS_CONNECTED ignoring signals */
480 mutex_enter(&so2->so_lock);
481 error = sowaitconnected(so2, 0, 1);
482 mutex_exit(&so2->so_lock);
483 if (error != 0) {
484 (void) socket_close(nso, 0, CRED());
485 socket_destroy(nso);
486 eprintsoline(so2, error);
487 goto done;
488 }
489
490 nvp = SOTOV(nso);
491 error = falloc(nvp, FWRITE|FREAD, &nfp, &nfd);
492 if (error != 0) {
493 (void) socket_close(nso, 0, CRED());
494 socket_destroy(nso);
495 eprintsoline(nso, error);
496 goto done;
497 }
498 /*
499 * copy over FNONBLOCK and FNDELAY flags should they exist
500 */
501 if (so1->so_state & SS_NONBLOCK)
502 nfp->f_flag |= FNONBLOCK;
503 if (so1->so_state & SS_NDELAY)
504 nfp->f_flag |= FNDELAY;
505
506 /*
507 * fill in the entries that falloc reserved
508 */
509 mutex_exit(&nfp->f_tlock);
510 setf(nfd, nfp);
511
512 /*
513 * get the original flags before we release
514 */
515 VERIFY(f_getfd_error(svs[0], &orig_flags) == 0);
516
517 releasef(svs[0]);
518 releasef(svs[1]);
519
520 /*
521 * If FD_CLOEXEC was set on the filedescriptor we're
522 * swapping out, we should set it on the new one too.
523 */
524 if (orig_flags & FD_CLOEXEC) {
525 f_setfd(nfd, FD_CLOEXEC);
526 }
527
528 /*
529 * The socketpair library routine will close the original
530 * svs[0] when this code passes out a different file
531 * descriptor.
532 */
533 svs[0] = nfd;
534
535 if (copyout(svs, sv, sizeof (svs))) {
536 (void) closeandsetf(nfd, NULL);
537 eprintline(EFAULT);
538 return (set_errno(EFAULT));
539 }
540 }
541 return (0);
542
543 done:
544 releasef(svs[0]);
545 releasef(svs[1]);
546 return (set_errno(error));
547 }
548
549 int
bind(int sock,struct sockaddr * name,socklen_t namelen,int version)550 bind(int sock, struct sockaddr *name, socklen_t namelen, int version)
551 {
552 struct sonode *so;
553 int error;
554
555 dprint(1, ("bind(%d, %p, %d)\n",
556 sock, (void *)name, namelen));
557
558 if ((so = getsonode(sock, &error, NULL)) == NULL)
559 return (set_errno(error));
560
561 /* Allocate and copyin name */
562 /*
563 * X/Open test does not expect EFAULT with NULL name and non-zero
564 * namelen.
565 */
566 if (name != NULL && namelen != 0) {
567 ASSERT(MUTEX_NOT_HELD(&so->so_lock));
568 name = copyin_name(so, name, &namelen, &error);
569 if (name == NULL) {
570 releasef(sock);
571 return (set_errno(error));
572 }
573 } else {
574 name = NULL;
575 namelen = 0;
576 }
577
578 switch (version) {
579 default:
580 error = socket_bind(so, name, namelen, 0, CRED());
581 break;
582 case SOV_XPG4_2:
583 error = socket_bind(so, name, namelen, _SOBIND_XPG4_2, CRED());
584 break;
585 case SOV_SOCKBSD:
586 error = socket_bind(so, name, namelen, _SOBIND_SOCKBSD, CRED());
587 break;
588 }
589 done:
590 releasef(sock);
591 if (name != NULL)
592 kmem_free(name, (size_t)namelen);
593
594 if (error)
595 return (set_errno(error));
596 return (0);
597 }
598
599 /* ARGSUSED2 */
600 int
listen(int sock,int backlog,int version)601 listen(int sock, int backlog, int version)
602 {
603 struct sonode *so;
604 int error;
605
606 dprint(1, ("listen(%d, %d)\n",
607 sock, backlog));
608
609 if ((so = getsonode(sock, &error, NULL)) == NULL)
610 return (set_errno(error));
611
612 error = socket_listen(so, backlog, CRED());
613
614 releasef(sock);
615 if (error)
616 return (set_errno(error));
617 return (0);
618 }
619
620 /*ARGSUSED3*/
621 int
accept(int sock,struct sockaddr * name,socklen_t * namelenp,int version,int flags)622 accept(int sock, struct sockaddr *name, socklen_t *namelenp, int version,
623 int flags)
624 {
625 struct sonode *so;
626 file_t *fp;
627 int error;
628 socklen_t namelen;
629 struct sonode *nso;
630 struct vnode *nvp;
631 struct file *nfp;
632 int nfd;
633 int ssflags;
634 struct sockaddr *addrp;
635 socklen_t addrlen;
636
637 dprint(1, ("accept(%d, %p, %p)\n",
638 sock, (void *)name, (void *)namelenp));
639
640 if (flags & ~(SOCK_CLOEXEC|SOCK_NONBLOCK|SOCK_NDELAY)) {
641 return (set_errno(EINVAL));
642 }
643
644 /* Translate SOCK_ flags to their SS_ variant */
645 ssflags = 0;
646 if (flags & SOCK_NONBLOCK)
647 ssflags |= SS_NONBLOCK;
648 if (flags & SOCK_NDELAY)
649 ssflags |= SS_NDELAY;
650
651 if ((so = getsonode(sock, &error, &fp)) == NULL)
652 return (set_errno(error));
653
654 if (name != NULL) {
655 ASSERT(MUTEX_NOT_HELD(&so->so_lock));
656 if (copyin(namelenp, &namelen, sizeof (namelen))) {
657 releasef(sock);
658 return (set_errno(EFAULT));
659 }
660 if (namelen != 0) {
661 error = useracc(name, (size_t)namelen, B_WRITE);
662 if (error && do_useracc) {
663 releasef(sock);
664 return (set_errno(EFAULT));
665 }
666 } else
667 name = NULL;
668 } else {
669 namelen = 0;
670 }
671
672 /*
673 * Allocate the user fd before socket_accept() in order to
674 * catch EMFILE errors before calling socket_accept().
675 */
676 if ((nfd = ufalloc(0)) == -1) {
677 eprintsoline(so, EMFILE);
678 releasef(sock);
679 return (set_errno(EMFILE));
680 }
681 error = socket_accept(so, fp->f_flag, CRED(), &nso);
682 if (error) {
683 setf(nfd, NULL);
684 releasef(sock);
685 return (set_errno(error));
686 }
687
688 nvp = SOTOV(nso);
689
690 ASSERT(MUTEX_NOT_HELD(&nso->so_lock));
691 if (namelen != 0) {
692 addrlen = so->so_max_addr_len;
693 addrp = (struct sockaddr *)kmem_alloc(addrlen, KM_SLEEP);
694
695 if ((error = socket_getpeername(nso, (struct sockaddr *)addrp,
696 &addrlen, B_TRUE, CRED())) == 0) {
697 error = copyout_name(name, namelen, namelenp,
698 addrp, addrlen);
699 } else {
700 ASSERT(error == EINVAL || error == ENOTCONN);
701 error = ECONNABORTED;
702 }
703 kmem_free(addrp, so->so_max_addr_len);
704 }
705
706 if (error) {
707 setf(nfd, NULL);
708 (void) socket_close(nso, 0, CRED());
709 socket_destroy(nso);
710 releasef(sock);
711 return (set_errno(error));
712 }
713 error = falloc(NULL, FWRITE|FREAD, &nfp, NULL);
714 if (error != 0) {
715 setf(nfd, NULL);
716 (void) socket_close(nso, 0, CRED());
717 socket_destroy(nso);
718 eprintsoline(so, error);
719 releasef(sock);
720 return (set_errno(error));
721 }
722 /*
723 * fill in the entries that falloc reserved
724 */
725 nfp->f_vnode = nvp;
726 mutex_exit(&nfp->f_tlock);
727 setf(nfd, nfp);
728
729 /*
730 * Act on SOCK_CLOEXEC from flags
731 */
732 if (flags & SOCK_CLOEXEC) {
733 f_setfd(nfd, FD_CLOEXEC);
734 }
735
736 /*
737 * Copy FNDELAY and FNONBLOCK from listener to acceptor
738 * and from ssflags
739 */
740 if ((ssflags | so->so_state) & (SS_NDELAY|SS_NONBLOCK)) {
741 uint_t oflag = nfp->f_flag;
742 int arg = 0;
743
744 if ((ssflags | so->so_state) & SS_NONBLOCK)
745 arg |= FNONBLOCK;
746 else if ((ssflags | so->so_state) & SS_NDELAY)
747 arg |= FNDELAY;
748
749 /*
750 * This code is a simplification of the F_SETFL code in fcntl()
751 * Ignore any errors from VOP_SETFL.
752 */
753 if ((error = VOP_SETFL(nvp, oflag, arg, nfp->f_cred, NULL))
754 != 0) {
755 eprintsoline(so, error);
756 error = 0;
757 } else {
758 mutex_enter(&nfp->f_tlock);
759 nfp->f_flag &= ~FMASK | (FREAD|FWRITE);
760 nfp->f_flag |= arg;
761 mutex_exit(&nfp->f_tlock);
762 }
763 }
764 releasef(sock);
765 return (nfd);
766 }
767
768 int
connect(int sock,struct sockaddr * name,socklen_t namelen,int version)769 connect(int sock, struct sockaddr *name, socklen_t namelen, int version)
770 {
771 struct sonode *so;
772 file_t *fp;
773 int error;
774
775 dprint(1, ("connect(%d, %p, %d)\n",
776 sock, (void *)name, namelen));
777
778 if ((so = getsonode(sock, &error, &fp)) == NULL)
779 return (set_errno(error));
780
781 /* Allocate and copyin name */
782 if (namelen != 0) {
783 ASSERT(MUTEX_NOT_HELD(&so->so_lock));
784 name = copyin_name(so, name, &namelen, &error);
785 if (name == NULL) {
786 releasef(sock);
787 return (set_errno(error));
788 }
789 } else
790 name = NULL;
791
792 error = socket_connect(so, name, namelen, fp->f_flag,
793 (version != SOV_XPG4_2) ? 0 : _SOCONNECT_XPG4_2, CRED());
794 releasef(sock);
795 if (name)
796 kmem_free(name, (size_t)namelen);
797 if (error)
798 return (set_errno(error));
799 return (0);
800 }
801
802 /*ARGSUSED2*/
803 int
shutdown(int sock,int how,int version)804 shutdown(int sock, int how, int version)
805 {
806 struct sonode *so;
807 int error;
808
809 dprint(1, ("shutdown(%d, %d)\n",
810 sock, how));
811
812 if ((so = getsonode(sock, &error, NULL)) == NULL)
813 return (set_errno(error));
814
815 error = socket_shutdown(so, how, CRED());
816
817 releasef(sock);
818 if (error)
819 return (set_errno(error));
820 return (0);
821 }
822
823 /*
824 * Common receive routine.
825 */
826 static ssize_t
recvit(int sock,struct nmsghdr * msg,struct uio * uiop,int flags,socklen_t * namelenp,socklen_t * controllenp,int * flagsp)827 recvit(int sock, struct nmsghdr *msg, struct uio *uiop, int flags,
828 socklen_t *namelenp, socklen_t *controllenp, int *flagsp)
829 {
830 struct sonode *so;
831 file_t *fp;
832 void *name;
833 socklen_t namelen;
834 void *control;
835 socklen_t controllen, free_controllen;
836 ssize_t len;
837 int error;
838
839 if ((so = getsonode(sock, &error, &fp)) == NULL)
840 return (set_errno(error));
841
842 len = uiop->uio_resid;
843 uiop->uio_fmode = fp->f_flag;
844 uiop->uio_extflg = UIO_COPY_CACHED;
845
846 name = msg->msg_name;
847 namelen = msg->msg_namelen;
848 control = msg->msg_control;
849 controllen = msg->msg_controllen;
850
851 msg->msg_flags = flags & (MSG_OOB | MSG_PEEK | MSG_WAITALL |
852 MSG_DONTWAIT | MSG_XPG4_2);
853
854 error = socket_recvmsg(so, msg, uiop, CRED());
855 if (error) {
856 releasef(sock);
857 return (set_errno(error));
858 }
859 lwp_stat_update(LWP_STAT_MSGRCV, 1);
860 releasef(sock);
861
862 free_controllen = msg->msg_controllen;
863
864 error = copyout_name(name, namelen, namelenp,
865 msg->msg_name, msg->msg_namelen);
866 if (error)
867 goto err;
868
869 if (flagsp != NULL) {
870 /*
871 * Clear internal flag.
872 */
873 msg->msg_flags &= ~MSG_XPG4_2;
874
875 /*
876 * Determine MSG_CTRUNC. sorecvmsg sets MSG_CTRUNC only
877 * when controllen is zero and there is control data to
878 * copy out.
879 */
880 if (controllen != 0 &&
881 (msg->msg_controllen > controllen || control == NULL)) {
882 dprint(1, ("recvit: CTRUNC %d %d %p\n",
883 msg->msg_controllen, controllen, control));
884
885 msg->msg_flags |= MSG_CTRUNC;
886 }
887 if (copyout(&msg->msg_flags, flagsp,
888 sizeof (msg->msg_flags))) {
889 error = EFAULT;
890 goto err;
891 }
892 }
893
894 if (controllen != 0) {
895 if (!(flags & MSG_XPG4_2)) {
896 /*
897 * Good old msg_accrights can only return a multiple
898 * of 4 bytes.
899 */
900 controllen &= ~((int)sizeof (uint32_t) - 1);
901 }
902
903 if (msg->msg_controllen > controllen || control == NULL) {
904 /*
905 * If the truncated part contains file descriptors,
906 * then they must be closed in the kernel as they
907 * will not be included in the data returned to
908 * user space. Close them now so that the header size
909 * can be safely adjusted prior to copyout. In case of
910 * an error during copyout, the remaining file
911 * descriptors will be closed in the error handler
912 * below.
913 */
914 so_closefds(msg->msg_control, msg->msg_controllen,
915 !(flags & MSG_XPG4_2),
916 control == NULL ? 0 : controllen);
917
918 /*
919 * In the case of a truncated control message, the last
920 * cmsg header that fits into the available buffer
921 * space must be adjusted to reflect the actual amount
922 * of associated data that will be returned. This only
923 * needs to be done for XPG4 messages as non-XPG4
924 * messages are not structured (they are just a
925 * buffer and a length - msg_accrights(len)).
926 */
927 if (control != NULL && (flags & MSG_XPG4_2)) {
928 so_truncatecmsg(msg->msg_control,
929 msg->msg_controllen, controllen);
930 msg->msg_controllen = controllen;
931 }
932 }
933
934 error = copyout_arg(control, controllen, controllenp,
935 msg->msg_control, msg->msg_controllen);
936
937 if (error)
938 goto err;
939
940 }
941 if (msg->msg_namelen != 0)
942 kmem_free(msg->msg_name, (size_t)msg->msg_namelen);
943 if (free_controllen != 0)
944 kmem_free(msg->msg_control, (size_t)free_controllen);
945 return (len - uiop->uio_resid);
946
947 err:
948 /*
949 * If we fail and the control part contains file descriptors
950 * we have to close them. For a truncated control message, the
951 * descriptors which were cut off have already been closed and the
952 * length adjusted so that they will not be closed again.
953 */
954 if (msg->msg_controllen != 0)
955 so_closefds(msg->msg_control, msg->msg_controllen,
956 !(flags & MSG_XPG4_2), 0);
957 if (msg->msg_namelen != 0)
958 kmem_free(msg->msg_name, (size_t)msg->msg_namelen);
959 if (free_controllen != 0)
960 kmem_free(msg->msg_control, (size_t)free_controllen);
961 return (set_errno(error));
962 }
963
964 /*
965 * Native system call
966 */
967 ssize_t
recv(int sock,void * buffer,size_t len,int flags)968 recv(int sock, void *buffer, size_t len, int flags)
969 {
970 struct nmsghdr lmsg;
971 struct uio auio;
972 struct iovec aiov[1];
973
974 dprint(1, ("recv(%d, %p, %ld, %d)\n",
975 sock, buffer, len, flags));
976
977 if ((ssize_t)len < 0) {
978 return (set_errno(EINVAL));
979 }
980
981 aiov[0].iov_base = buffer;
982 aiov[0].iov_len = len;
983 auio.uio_loffset = 0;
984 auio.uio_iov = aiov;
985 auio.uio_iovcnt = 1;
986 auio.uio_resid = len;
987 auio.uio_segflg = UIO_USERSPACE;
988 auio.uio_limit = 0;
989
990 lmsg.msg_namelen = 0;
991 lmsg.msg_controllen = 0;
992 lmsg.msg_flags = 0;
993 return (recvit(sock, &lmsg, &auio, flags, NULL, NULL, NULL));
994 }
995
996 ssize_t
recvfrom(int sock,void * buffer,size_t len,int flags,struct sockaddr * name,socklen_t * namelenp)997 recvfrom(int sock, void *buffer, size_t len, int flags, struct sockaddr *name,
998 socklen_t *namelenp)
999 {
1000 struct nmsghdr lmsg;
1001 struct uio auio;
1002 struct iovec aiov[1];
1003
1004 dprint(1, ("recvfrom(%d, %p, %ld, %d, %p, %p)\n",
1005 sock, buffer, len, flags, (void *)name, (void *)namelenp));
1006
1007 if ((ssize_t)len < 0) {
1008 return (set_errno(EINVAL));
1009 }
1010
1011 aiov[0].iov_base = buffer;
1012 aiov[0].iov_len = len;
1013 auio.uio_loffset = 0;
1014 auio.uio_iov = aiov;
1015 auio.uio_iovcnt = 1;
1016 auio.uio_resid = len;
1017 auio.uio_segflg = UIO_USERSPACE;
1018 auio.uio_limit = 0;
1019
1020 lmsg.msg_name = (char *)name;
1021 if (namelenp != NULL) {
1022 if (copyin(namelenp, &lmsg.msg_namelen,
1023 sizeof (lmsg.msg_namelen)))
1024 return (set_errno(EFAULT));
1025 } else {
1026 lmsg.msg_namelen = 0;
1027 }
1028 lmsg.msg_controllen = 0;
1029 lmsg.msg_flags = 0;
1030
1031 return (recvit(sock, &lmsg, &auio, flags, namelenp, NULL, NULL));
1032 }
1033
1034 /*
1035 * Uses the MSG_XPG4_2 flag to determine if the caller is using
1036 * struct omsghdr or struct nmsghdr.
1037 */
1038 ssize_t
recvmsg(int sock,struct nmsghdr * msg,int flags)1039 recvmsg(int sock, struct nmsghdr *msg, int flags)
1040 {
1041 STRUCT_DECL(nmsghdr, u_lmsg);
1042 STRUCT_HANDLE(nmsghdr, umsgptr);
1043 struct nmsghdr lmsg;
1044 struct uio auio;
1045 struct iovec buf[IOV_MAX_STACK], *aiov = buf;
1046 ssize_t iovsize = 0;
1047 int iovcnt;
1048 ssize_t len, rval;
1049 int i;
1050 int *flagsp;
1051 model_t model;
1052
1053 dprint(1, ("recvmsg(%d, %p, %d)\n",
1054 sock, (void *)msg, flags));
1055
1056 model = get_udatamodel();
1057 STRUCT_INIT(u_lmsg, model);
1058 STRUCT_SET_HANDLE(umsgptr, model, msg);
1059
1060 if (flags & MSG_XPG4_2) {
1061 if (copyin(msg, STRUCT_BUF(u_lmsg), STRUCT_SIZE(u_lmsg)))
1062 return (set_errno(EFAULT));
1063 flagsp = STRUCT_FADDR(umsgptr, msg_flags);
1064 } else {
1065 /*
1066 * Assumes that nmsghdr and omsghdr are identically shaped
1067 * except for the added msg_flags field.
1068 */
1069 if (copyin(msg, STRUCT_BUF(u_lmsg),
1070 SIZEOF_STRUCT(omsghdr, model)))
1071 return (set_errno(EFAULT));
1072 STRUCT_FSET(u_lmsg, msg_flags, 0);
1073 flagsp = NULL;
1074 }
1075
1076 /*
1077 * Code below us will kmem_alloc memory and hang it
1078 * off msg_control and msg_name fields. This forces
1079 * us to copy the structure to its native form.
1080 */
1081 lmsg.msg_name = STRUCT_FGETP(u_lmsg, msg_name);
1082 lmsg.msg_namelen = STRUCT_FGET(u_lmsg, msg_namelen);
1083 lmsg.msg_iov = STRUCT_FGETP(u_lmsg, msg_iov);
1084 lmsg.msg_iovlen = STRUCT_FGET(u_lmsg, msg_iovlen);
1085 lmsg.msg_control = STRUCT_FGETP(u_lmsg, msg_control);
1086 lmsg.msg_controllen = STRUCT_FGET(u_lmsg, msg_controllen);
1087 lmsg.msg_flags = STRUCT_FGET(u_lmsg, msg_flags);
1088
1089 iovcnt = lmsg.msg_iovlen;
1090
1091 if (iovcnt <= 0 || iovcnt > IOV_MAX) {
1092 return (set_errno(EMSGSIZE));
1093 }
1094
1095 if (iovcnt > IOV_MAX_STACK) {
1096 iovsize = iovcnt * sizeof (struct iovec);
1097 aiov = kmem_alloc(iovsize, KM_SLEEP);
1098 }
1099
1100 #ifdef _SYSCALL32_IMPL
1101 /*
1102 * 32-bit callers need to have their iovec expanded, while ensuring
1103 * that they can't move more than 2Gbytes of data in a single call.
1104 */
1105 if (model == DATAMODEL_ILP32) {
1106 struct iovec32 buf32[IOV_MAX_STACK], *aiov32 = buf32;
1107 ssize_t iov32size;
1108 ssize32_t count32;
1109
1110 iov32size = iovcnt * sizeof (struct iovec32);
1111 if (iovsize != 0)
1112 aiov32 = kmem_alloc(iov32size, KM_SLEEP);
1113
1114 if (copyin((struct iovec32 *)lmsg.msg_iov, aiov32, iov32size)) {
1115 if (iovsize != 0) {
1116 kmem_free(aiov32, iov32size);
1117 kmem_free(aiov, iovsize);
1118 }
1119
1120 return (set_errno(EFAULT));
1121 }
1122
1123 count32 = 0;
1124 for (i = 0; i < iovcnt; i++) {
1125 ssize32_t iovlen32;
1126
1127 iovlen32 = aiov32[i].iov_len;
1128 count32 += iovlen32;
1129 if (iovlen32 < 0 || count32 < 0) {
1130 if (iovsize != 0) {
1131 kmem_free(aiov32, iov32size);
1132 kmem_free(aiov, iovsize);
1133 }
1134
1135 return (set_errno(EINVAL));
1136 }
1137
1138 aiov[i].iov_len = iovlen32;
1139 aiov[i].iov_base =
1140 (caddr_t)(uintptr_t)aiov32[i].iov_base;
1141 }
1142
1143 if (iovsize != 0)
1144 kmem_free(aiov32, iov32size);
1145 } else
1146 #endif /* _SYSCALL32_IMPL */
1147 if (copyin(lmsg.msg_iov, aiov, iovcnt * sizeof (struct iovec))) {
1148 if (iovsize != 0)
1149 kmem_free(aiov, iovsize);
1150
1151 return (set_errno(EFAULT));
1152 }
1153 len = 0;
1154 for (i = 0; i < iovcnt; i++) {
1155 ssize_t iovlen = aiov[i].iov_len;
1156 len += iovlen;
1157 if (iovlen < 0 || len < 0) {
1158 if (iovsize != 0)
1159 kmem_free(aiov, iovsize);
1160
1161 return (set_errno(EINVAL));
1162 }
1163 }
1164 auio.uio_loffset = 0;
1165 auio.uio_iov = aiov;
1166 auio.uio_iovcnt = iovcnt;
1167 auio.uio_resid = len;
1168 auio.uio_segflg = UIO_USERSPACE;
1169 auio.uio_limit = 0;
1170
1171 if (lmsg.msg_control != NULL &&
1172 (do_useracc == 0 ||
1173 useracc(lmsg.msg_control, lmsg.msg_controllen,
1174 B_WRITE) != 0)) {
1175 if (iovsize != 0)
1176 kmem_free(aiov, iovsize);
1177
1178 return (set_errno(EFAULT));
1179 }
1180
1181 rval = recvit(sock, &lmsg, &auio, flags,
1182 STRUCT_FADDR(umsgptr, msg_namelen),
1183 STRUCT_FADDR(umsgptr, msg_controllen), flagsp);
1184
1185 if (iovsize != 0)
1186 kmem_free(aiov, iovsize);
1187
1188 return (rval);
1189 }
1190
1191 /*
1192 * Common send function.
1193 */
1194 static ssize_t
sendit(int sock,struct nmsghdr * msg,struct uio * uiop,int flags)1195 sendit(int sock, struct nmsghdr *msg, struct uio *uiop, int flags)
1196 {
1197 struct sonode *so;
1198 file_t *fp;
1199 void *name;
1200 socklen_t namelen;
1201 void *control;
1202 socklen_t controllen;
1203 ssize_t len;
1204 int error;
1205
1206 if ((so = getsonode(sock, &error, &fp)) == NULL)
1207 return (set_errno(error));
1208
1209 uiop->uio_fmode = fp->f_flag;
1210
1211 if (so->so_family == AF_UNIX)
1212 uiop->uio_extflg = UIO_COPY_CACHED;
1213 else
1214 uiop->uio_extflg = UIO_COPY_DEFAULT;
1215
1216 /* Allocate and copyin name and control */
1217 name = msg->msg_name;
1218 namelen = msg->msg_namelen;
1219 if (name != NULL && namelen != 0) {
1220 ASSERT(MUTEX_NOT_HELD(&so->so_lock));
1221 name = copyin_name(so,
1222 (struct sockaddr *)name,
1223 &namelen, &error);
1224 if (name == NULL)
1225 goto done3;
1226 /* copyin_name null terminates addresses for AF_UNIX */
1227 msg->msg_namelen = namelen;
1228 msg->msg_name = name;
1229 } else {
1230 msg->msg_name = name = NULL;
1231 msg->msg_namelen = namelen = 0;
1232 }
1233
1234 control = msg->msg_control;
1235 controllen = msg->msg_controllen;
1236 if ((control != NULL) && (controllen != 0)) {
1237 /*
1238 * Verify that the length is not excessive to prevent
1239 * an application from consuming all of kernel memory.
1240 */
1241 if (controllen > SO_MAXARGSIZE) {
1242 error = EINVAL;
1243 goto done2;
1244 }
1245 control = kmem_alloc(controllen, KM_SLEEP);
1246
1247 ASSERT(MUTEX_NOT_HELD(&so->so_lock));
1248 if (copyin(msg->msg_control, control, controllen)) {
1249 error = EFAULT;
1250 goto done1;
1251 }
1252 msg->msg_control = control;
1253 } else {
1254 msg->msg_control = control = NULL;
1255 msg->msg_controllen = controllen = 0;
1256 }
1257
1258 len = uiop->uio_resid;
1259 msg->msg_flags = flags;
1260
1261 error = socket_sendmsg(so, msg, uiop, CRED());
1262 done1:
1263 if (control != NULL)
1264 kmem_free(control, controllen);
1265 done2:
1266 if (name != NULL)
1267 kmem_free(name, namelen);
1268 done3:
1269 if (error != 0) {
1270 releasef(sock);
1271 return (set_errno(error));
1272 }
1273 lwp_stat_update(LWP_STAT_MSGSND, 1);
1274 releasef(sock);
1275 return (len - uiop->uio_resid);
1276 }
1277
1278 /*
1279 * Native system call
1280 */
1281 ssize_t
send(int sock,void * buffer,size_t len,int flags)1282 send(int sock, void *buffer, size_t len, int flags)
1283 {
1284 struct nmsghdr lmsg;
1285 struct uio auio;
1286 struct iovec aiov[1];
1287
1288 dprint(1, ("send(%d, %p, %ld, %d)\n",
1289 sock, buffer, len, flags));
1290
1291 if ((ssize_t)len < 0) {
1292 return (set_errno(EINVAL));
1293 }
1294
1295 aiov[0].iov_base = buffer;
1296 aiov[0].iov_len = len;
1297 auio.uio_loffset = 0;
1298 auio.uio_iov = aiov;
1299 auio.uio_iovcnt = 1;
1300 auio.uio_resid = len;
1301 auio.uio_segflg = UIO_USERSPACE;
1302 auio.uio_limit = 0;
1303
1304 lmsg.msg_name = NULL;
1305 lmsg.msg_control = NULL;
1306 if (!(flags & MSG_XPG4_2)) {
1307 /*
1308 * In order to be compatible with the libsocket/sockmod
1309 * implementation we set EOR for all send* calls.
1310 */
1311 flags |= MSG_EOR;
1312 }
1313 return (sendit(sock, &lmsg, &auio, flags));
1314 }
1315
1316 /*
1317 * Uses the MSG_XPG4_2 flag to determine if the caller is using
1318 * struct omsghdr or struct nmsghdr.
1319 */
1320 ssize_t
sendmsg(int sock,struct nmsghdr * msg,int flags)1321 sendmsg(int sock, struct nmsghdr *msg, int flags)
1322 {
1323 struct nmsghdr lmsg;
1324 STRUCT_DECL(nmsghdr, u_lmsg);
1325 struct uio auio;
1326 struct iovec buf[IOV_MAX_STACK], *aiov = buf;
1327 ssize_t iovsize = 0;
1328 int iovcnt;
1329 ssize_t len, rval;
1330 int i;
1331 model_t model;
1332
1333 dprint(1, ("sendmsg(%d, %p, %d)\n", sock, (void *)msg, flags));
1334
1335 model = get_udatamodel();
1336 STRUCT_INIT(u_lmsg, model);
1337
1338 if (flags & MSG_XPG4_2) {
1339 if (copyin(msg, (char *)STRUCT_BUF(u_lmsg),
1340 STRUCT_SIZE(u_lmsg)))
1341 return (set_errno(EFAULT));
1342 } else {
1343 /*
1344 * Assumes that nmsghdr and omsghdr are identically shaped
1345 * except for the added msg_flags field.
1346 */
1347 if (copyin(msg, (char *)STRUCT_BUF(u_lmsg),
1348 SIZEOF_STRUCT(omsghdr, model)))
1349 return (set_errno(EFAULT));
1350 /*
1351 * In order to be compatible with the libsocket/sockmod
1352 * implementation we set EOR for all send* calls.
1353 */
1354 flags |= MSG_EOR;
1355 }
1356
1357 /*
1358 * Code below us will kmem_alloc memory and hang it
1359 * off msg_control and msg_name fields. This forces
1360 * us to copy the structure to its native form.
1361 */
1362 lmsg.msg_name = STRUCT_FGETP(u_lmsg, msg_name);
1363 lmsg.msg_namelen = STRUCT_FGET(u_lmsg, msg_namelen);
1364 lmsg.msg_iov = STRUCT_FGETP(u_lmsg, msg_iov);
1365 lmsg.msg_iovlen = STRUCT_FGET(u_lmsg, msg_iovlen);
1366 lmsg.msg_control = STRUCT_FGETP(u_lmsg, msg_control);
1367 lmsg.msg_controllen = STRUCT_FGET(u_lmsg, msg_controllen);
1368 lmsg.msg_flags = STRUCT_FGET(u_lmsg, msg_flags);
1369
1370 iovcnt = lmsg.msg_iovlen;
1371
1372 if (iovcnt <= 0 || iovcnt > IOV_MAX) {
1373 /*
1374 * Unless this is XPG 4.2 we allow iovcnt == 0 to
1375 * be compatible with SunOS 4.X and 4.4BSD.
1376 */
1377 if (iovcnt != 0 || (flags & MSG_XPG4_2))
1378 return (set_errno(EMSGSIZE));
1379 }
1380
1381 if (iovcnt > IOV_MAX_STACK) {
1382 iovsize = iovcnt * sizeof (struct iovec);
1383 aiov = kmem_alloc(iovsize, KM_SLEEP);
1384 }
1385
1386 #ifdef _SYSCALL32_IMPL
1387 /*
1388 * 32-bit callers need to have their iovec expanded, while ensuring
1389 * that they can't move more than 2Gbytes of data in a single call.
1390 */
1391 if (model == DATAMODEL_ILP32) {
1392 struct iovec32 buf32[IOV_MAX_STACK], *aiov32 = buf32;
1393 ssize_t iov32size;
1394 ssize32_t count32;
1395
1396 iov32size = iovcnt * sizeof (struct iovec32);
1397 if (iovsize != 0)
1398 aiov32 = kmem_alloc(iov32size, KM_SLEEP);
1399
1400 if (iovcnt != 0 &&
1401 copyin((struct iovec32 *)lmsg.msg_iov, aiov32, iov32size)) {
1402 if (iovsize != 0) {
1403 kmem_free(aiov32, iov32size);
1404 kmem_free(aiov, iovsize);
1405 }
1406
1407 return (set_errno(EFAULT));
1408 }
1409
1410 count32 = 0;
1411 for (i = 0; i < iovcnt; i++) {
1412 ssize32_t iovlen32;
1413
1414 iovlen32 = aiov32[i].iov_len;
1415 count32 += iovlen32;
1416 if (iovlen32 < 0 || count32 < 0) {
1417 if (iovsize != 0) {
1418 kmem_free(aiov32, iov32size);
1419 kmem_free(aiov, iovsize);
1420 }
1421
1422 return (set_errno(EINVAL));
1423 }
1424
1425 aiov[i].iov_len = iovlen32;
1426 aiov[i].iov_base =
1427 (caddr_t)(uintptr_t)aiov32[i].iov_base;
1428 }
1429
1430 if (iovsize != 0)
1431 kmem_free(aiov32, iov32size);
1432 } else
1433 #endif /* _SYSCALL32_IMPL */
1434 if (iovcnt != 0 &&
1435 copyin(lmsg.msg_iov, aiov,
1436 (unsigned)iovcnt * sizeof (struct iovec))) {
1437 if (iovsize != 0)
1438 kmem_free(aiov, iovsize);
1439
1440 return (set_errno(EFAULT));
1441 }
1442 len = 0;
1443 for (i = 0; i < iovcnt; i++) {
1444 ssize_t iovlen = aiov[i].iov_len;
1445 len += iovlen;
1446 if (iovlen < 0 || len < 0) {
1447 if (iovsize != 0)
1448 kmem_free(aiov, iovsize);
1449
1450 return (set_errno(EINVAL));
1451 }
1452 }
1453 auio.uio_loffset = 0;
1454 auio.uio_iov = aiov;
1455 auio.uio_iovcnt = iovcnt;
1456 auio.uio_resid = len;
1457 auio.uio_segflg = UIO_USERSPACE;
1458 auio.uio_limit = 0;
1459
1460 rval = sendit(sock, &lmsg, &auio, flags);
1461
1462 if (iovsize != 0)
1463 kmem_free(aiov, iovsize);
1464
1465 return (rval);
1466 }
1467
1468 ssize_t
sendto(int sock,void * buffer,size_t len,int flags,struct sockaddr * name,socklen_t namelen)1469 sendto(int sock, void *buffer, size_t len, int flags,
1470 struct sockaddr *name, socklen_t namelen)
1471 {
1472 struct nmsghdr lmsg;
1473 struct uio auio;
1474 struct iovec aiov[1];
1475
1476 dprint(1, ("sendto(%d, %p, %ld, %d, %p, %d)\n",
1477 sock, buffer, len, flags, (void *)name, namelen));
1478
1479 if ((ssize_t)len < 0) {
1480 return (set_errno(EINVAL));
1481 }
1482
1483 aiov[0].iov_base = buffer;
1484 aiov[0].iov_len = len;
1485 auio.uio_loffset = 0;
1486 auio.uio_iov = aiov;
1487 auio.uio_iovcnt = 1;
1488 auio.uio_resid = len;
1489 auio.uio_segflg = UIO_USERSPACE;
1490 auio.uio_limit = 0;
1491
1492 lmsg.msg_name = (char *)name;
1493 lmsg.msg_namelen = namelen;
1494 lmsg.msg_control = NULL;
1495 if (!(flags & MSG_XPG4_2)) {
1496 /*
1497 * In order to be compatible with the libsocket/sockmod
1498 * implementation we set EOR for all send* calls.
1499 */
1500 flags |= MSG_EOR;
1501 }
1502 return (sendit(sock, &lmsg, &auio, flags));
1503 }
1504
1505 /*ARGSUSED3*/
1506 int
getpeername(int sock,struct sockaddr * name,socklen_t * namelenp,int version)1507 getpeername(int sock, struct sockaddr *name, socklen_t *namelenp, int version)
1508 {
1509 struct sonode *so;
1510 int error;
1511 socklen_t namelen;
1512 socklen_t sock_addrlen;
1513 struct sockaddr *sock_addrp;
1514
1515 dprint(1, ("getpeername(%d, %p, %p)\n",
1516 sock, (void *)name, (void *)namelenp));
1517
1518 if ((so = getsonode(sock, &error, NULL)) == NULL)
1519 goto bad;
1520
1521 ASSERT(MUTEX_NOT_HELD(&so->so_lock));
1522 if (copyin(namelenp, &namelen, sizeof (namelen)) ||
1523 (name == NULL && namelen != 0)) {
1524 error = EFAULT;
1525 goto rel_out;
1526 }
1527 sock_addrlen = so->so_max_addr_len;
1528 sock_addrp = (struct sockaddr *)kmem_alloc(sock_addrlen, KM_SLEEP);
1529
1530 if ((error = socket_getpeername(so, sock_addrp, &sock_addrlen,
1531 B_FALSE, CRED())) == 0) {
1532 ASSERT(sock_addrlen <= so->so_max_addr_len);
1533 error = copyout_name(name, namelen, namelenp,
1534 (void *)sock_addrp, sock_addrlen);
1535 }
1536 kmem_free(sock_addrp, so->so_max_addr_len);
1537 rel_out:
1538 releasef(sock);
1539 bad: return (error != 0 ? set_errno(error) : 0);
1540 }
1541
1542 /*ARGSUSED3*/
1543 int
getsockname(int sock,struct sockaddr * name,socklen_t * namelenp,int version)1544 getsockname(int sock, struct sockaddr *name, socklen_t *namelenp, int version)
1545 {
1546 struct sonode *so;
1547 int error;
1548 socklen_t namelen, sock_addrlen;
1549 struct sockaddr *sock_addrp;
1550
1551 dprint(1, ("getsockname(%d, %p, %p)\n",
1552 sock, (void *)name, (void *)namelenp));
1553
1554 if ((so = getsonode(sock, &error, NULL)) == NULL)
1555 goto bad;
1556
1557 ASSERT(MUTEX_NOT_HELD(&so->so_lock));
1558 if (copyin(namelenp, &namelen, sizeof (namelen)) ||
1559 (name == NULL && namelen != 0)) {
1560 error = EFAULT;
1561 goto rel_out;
1562 }
1563
1564 sock_addrlen = so->so_max_addr_len;
1565 sock_addrp = (struct sockaddr *)kmem_alloc(sock_addrlen, KM_SLEEP);
1566 if ((error = socket_getsockname(so, sock_addrp, &sock_addrlen,
1567 CRED())) == 0) {
1568 ASSERT(MUTEX_NOT_HELD(&so->so_lock));
1569 ASSERT(sock_addrlen <= so->so_max_addr_len);
1570 error = copyout_name(name, namelen, namelenp,
1571 (void *)sock_addrp, sock_addrlen);
1572 }
1573 kmem_free(sock_addrp, so->so_max_addr_len);
1574 rel_out:
1575 releasef(sock);
1576 bad: return (error != 0 ? set_errno(error) : 0);
1577 }
1578
1579 /*ARGSUSED5*/
1580 int
getsockopt(int sock,int level,int option_name,void * option_value,socklen_t * option_lenp,int version)1581 getsockopt(int sock, int level, int option_name, void *option_value,
1582 socklen_t *option_lenp, int version)
1583 {
1584 struct sonode *so;
1585 socklen_t optlen, optlen_res;
1586 void *optval;
1587 int error;
1588
1589 dprint(1, ("getsockopt(%d, %d, %d, %p, %p)\n",
1590 sock, level, option_name, option_value, (void *)option_lenp));
1591
1592 if ((so = getsonode(sock, &error, NULL)) == NULL)
1593 return (set_errno(error));
1594
1595 ASSERT(MUTEX_NOT_HELD(&so->so_lock));
1596 if (copyin(option_lenp, &optlen, sizeof (optlen))) {
1597 releasef(sock);
1598 return (set_errno(EFAULT));
1599 }
1600 /*
1601 * Verify that the length is not excessive to prevent
1602 * an application from consuming all of kernel memory.
1603 */
1604 if (optlen > SO_MAXARGSIZE) {
1605 error = EINVAL;
1606 releasef(sock);
1607 return (set_errno(error));
1608 }
1609 optval = kmem_alloc(optlen, KM_SLEEP);
1610 optlen_res = optlen;
1611 error = socket_getsockopt(so, level, option_name, optval,
1612 &optlen_res, (version != SOV_XPG4_2) ? 0 : _SOGETSOCKOPT_XPG4_2,
1613 CRED());
1614 releasef(sock);
1615 if (error) {
1616 kmem_free(optval, optlen);
1617 return (set_errno(error));
1618 }
1619 error = copyout_arg(option_value, optlen, option_lenp,
1620 optval, optlen_res);
1621 kmem_free(optval, optlen);
1622 if (error)
1623 return (set_errno(error));
1624 return (0);
1625 }
1626
1627 /*ARGSUSED5*/
1628 int
setsockopt(int sock,int level,int option_name,void * option_value,socklen_t option_len,int version)1629 setsockopt(int sock, int level, int option_name, void *option_value,
1630 socklen_t option_len, int version)
1631 {
1632 struct sonode *so;
1633 intptr_t buffer[2];
1634 void *optval = NULL;
1635 int error;
1636
1637 dprint(1, ("setsockopt(%d, %d, %d, %p, %d)\n",
1638 sock, level, option_name, option_value, option_len));
1639
1640 if ((so = getsonode(sock, &error, NULL)) == NULL)
1641 return (set_errno(error));
1642
1643 if (option_value != NULL) {
1644 if (option_len != 0) {
1645 /*
1646 * Verify that the length is not excessive to prevent
1647 * an application from consuming all of kernel memory.
1648 */
1649 if (option_len > SO_MAXARGSIZE) {
1650 error = EINVAL;
1651 goto done2;
1652 }
1653 optval = option_len <= sizeof (buffer) ?
1654 &buffer : kmem_alloc((size_t)option_len, KM_SLEEP);
1655 ASSERT(MUTEX_NOT_HELD(&so->so_lock));
1656 if (copyin(option_value, optval, (size_t)option_len)) {
1657 error = EFAULT;
1658 goto done1;
1659 }
1660 }
1661 } else
1662 option_len = 0;
1663
1664 error = socket_setsockopt(so, level, option_name, optval,
1665 (t_uscalar_t)option_len, CRED());
1666 done1:
1667 if (optval != buffer)
1668 kmem_free(optval, (size_t)option_len);
1669 done2:
1670 releasef(sock);
1671 if (error)
1672 return (set_errno(error));
1673 return (0);
1674 }
1675
1676 static int
sockconf_add_sock(int family,int type,int protocol,char * name)1677 sockconf_add_sock(int family, int type, int protocol, char *name)
1678 {
1679 int error = 0;
1680 char *kdevpath = NULL;
1681 char *kmodule = NULL;
1682 char *buf = NULL;
1683 size_t pathlen = 0;
1684 struct sockparams *sp;
1685
1686 if (name == NULL)
1687 return (EINVAL);
1688 /*
1689 * Copyin the name.
1690 * This also makes it possible to check for too long pathnames.
1691 * Compress the space needed for the name before passing it
1692 * to soconfig - soconfig will store the string until
1693 * the configuration is removed.
1694 */
1695 buf = kmem_alloc(MAXPATHLEN, KM_SLEEP);
1696 if ((error = copyinstr(name, buf, MAXPATHLEN, &pathlen)) != 0) {
1697 kmem_free(buf, MAXPATHLEN);
1698 return (error);
1699 }
1700 if (strncmp(buf, "/dev", strlen("/dev")) == 0) {
1701 /* For device */
1702 kdevpath = kmem_alloc(pathlen, KM_SLEEP);
1703 bcopy(buf, kdevpath, pathlen);
1704 kdevpath[pathlen - 1] = '\0';
1705 } else {
1706 /* For socket module */
1707 kmodule = kmem_alloc(pathlen, KM_SLEEP);
1708 bcopy(buf, kmodule, pathlen);
1709 kmodule[pathlen - 1] = '\0';
1710 pathlen = 0;
1711 }
1712 kmem_free(buf, MAXPATHLEN);
1713
1714 /* sockparams_create frees mod name and devpath upon failure */
1715 sp = sockparams_create(family, type, protocol, kmodule,
1716 kdevpath, pathlen, 0, KM_SLEEP, &error);
1717 if (sp != NULL) {
1718 error = sockparams_add(sp);
1719 if (error != 0)
1720 sockparams_destroy(sp);
1721 }
1722
1723 return (error);
1724 }
1725
1726 static int
sockconf_remove_sock(int family,int type,int protocol)1727 sockconf_remove_sock(int family, int type, int protocol)
1728 {
1729 return (sockparams_delete(family, type, protocol));
1730 }
1731
1732 static int
sockconfig_remove_filter(const char * uname)1733 sockconfig_remove_filter(const char *uname)
1734 {
1735 char kname[SOF_MAXNAMELEN];
1736 size_t len;
1737 int error;
1738 sof_entry_t *ent;
1739
1740 if ((error = copyinstr(uname, kname, SOF_MAXNAMELEN, &len)) != 0)
1741 return (error);
1742
1743 ent = sof_entry_remove_by_name(kname);
1744 if (ent == NULL)
1745 return (ENXIO);
1746
1747 mutex_enter(&ent->sofe_lock);
1748 ASSERT(!(ent->sofe_flags & SOFEF_CONDEMED));
1749 if (ent->sofe_refcnt == 0) {
1750 mutex_exit(&ent->sofe_lock);
1751 sof_entry_free(ent);
1752 } else {
1753 /* let the last socket free the filter */
1754 ent->sofe_flags |= SOFEF_CONDEMED;
1755 mutex_exit(&ent->sofe_lock);
1756 }
1757
1758 return (0);
1759 }
1760
1761 static int
sockconfig_add_filter(const char * uname,void * ufilpropp)1762 sockconfig_add_filter(const char *uname, void *ufilpropp)
1763 {
1764 struct sockconfig_filter_props filprop;
1765 sof_entry_t *ent;
1766 int error;
1767 size_t tuplesz, len;
1768 char hintbuf[SOF_MAXNAMELEN];
1769
1770 ent = kmem_zalloc(sizeof (sof_entry_t), KM_SLEEP);
1771 mutex_init(&ent->sofe_lock, NULL, MUTEX_DEFAULT, NULL);
1772
1773 if ((error = copyinstr(uname, ent->sofe_name, SOF_MAXNAMELEN,
1774 &len)) != 0) {
1775 sof_entry_free(ent);
1776 return (error);
1777 }
1778
1779 if (get_udatamodel() == DATAMODEL_NATIVE) {
1780 if (copyin(ufilpropp, &filprop, sizeof (filprop)) != 0) {
1781 sof_entry_free(ent);
1782 return (EFAULT);
1783 }
1784 }
1785 #ifdef _SYSCALL32_IMPL
1786 else {
1787 struct sockconfig_filter_props32 filprop32;
1788
1789 if (copyin(ufilpropp, &filprop32, sizeof (filprop32)) != 0) {
1790 sof_entry_free(ent);
1791 return (EFAULT);
1792 }
1793 filprop.sfp_modname = (char *)(uintptr_t)filprop32.sfp_modname;
1794 filprop.sfp_autoattach = filprop32.sfp_autoattach;
1795 filprop.sfp_hint = filprop32.sfp_hint;
1796 filprop.sfp_hintarg = (char *)(uintptr_t)filprop32.sfp_hintarg;
1797 filprop.sfp_socktuple_cnt = filprop32.sfp_socktuple_cnt;
1798 filprop.sfp_socktuple =
1799 (sof_socktuple_t *)(uintptr_t)filprop32.sfp_socktuple;
1800 }
1801 #endif /* _SYSCALL32_IMPL */
1802
1803 if ((error = copyinstr(filprop.sfp_modname, ent->sofe_modname,
1804 sizeof (ent->sofe_modname), &len)) != 0) {
1805 sof_entry_free(ent);
1806 return (error);
1807 }
1808
1809 /*
1810 * A filter must specify at least one socket tuple.
1811 */
1812 if (filprop.sfp_socktuple_cnt == 0 ||
1813 filprop.sfp_socktuple_cnt > SOF_MAXSOCKTUPLECNT) {
1814 sof_entry_free(ent);
1815 return (EINVAL);
1816 }
1817 ent->sofe_flags = filprop.sfp_autoattach ? SOFEF_AUTO : SOFEF_PROG;
1818 ent->sofe_hint = filprop.sfp_hint;
1819
1820 /*
1821 * Verify the hint, and copy in the hint argument, if necessary.
1822 */
1823 switch (ent->sofe_hint) {
1824 case SOF_HINT_BEFORE:
1825 case SOF_HINT_AFTER:
1826 if ((error = copyinstr(filprop.sfp_hintarg, hintbuf,
1827 sizeof (hintbuf), &len)) != 0) {
1828 sof_entry_free(ent);
1829 return (error);
1830 }
1831 ent->sofe_hintarg = kmem_alloc(len, KM_SLEEP);
1832 bcopy(hintbuf, ent->sofe_hintarg, len);
1833 /* FALLTHRU */
1834 case SOF_HINT_TOP:
1835 case SOF_HINT_BOTTOM:
1836 /* hints cannot be used with programmatic filters */
1837 if (ent->sofe_flags & SOFEF_PROG) {
1838 sof_entry_free(ent);
1839 return (EINVAL);
1840 }
1841 break;
1842 case SOF_HINT_NONE:
1843 break;
1844 default:
1845 /* bad hint value */
1846 sof_entry_free(ent);
1847 return (EINVAL);
1848 }
1849
1850 ent->sofe_socktuple_cnt = filprop.sfp_socktuple_cnt;
1851 tuplesz = sizeof (sof_socktuple_t) * ent->sofe_socktuple_cnt;
1852 ent->sofe_socktuple = kmem_alloc(tuplesz, KM_SLEEP);
1853
1854 if (get_udatamodel() == DATAMODEL_NATIVE) {
1855 if (copyin(filprop.sfp_socktuple, ent->sofe_socktuple,
1856 tuplesz)) {
1857 sof_entry_free(ent);
1858 return (EFAULT);
1859 }
1860 }
1861 #ifdef _SYSCALL32_IMPL
1862 else {
1863 int i;
1864 caddr_t data = (caddr_t)filprop.sfp_socktuple;
1865 sof_socktuple_t *tup = ent->sofe_socktuple;
1866 sof_socktuple32_t tup32;
1867
1868 tup = ent->sofe_socktuple;
1869 for (i = 0; i < ent->sofe_socktuple_cnt; i++, tup++) {
1870 ASSERT(tup < ent->sofe_socktuple + tuplesz);
1871
1872 if (copyin(data, &tup32, sizeof (tup32)) != 0) {
1873 sof_entry_free(ent);
1874 return (EFAULT);
1875 }
1876 tup->sofst_family = tup32.sofst_family;
1877 tup->sofst_type = tup32.sofst_type;
1878 tup->sofst_protocol = tup32.sofst_protocol;
1879
1880 data += sizeof (tup32);
1881 }
1882 }
1883 #endif /* _SYSCALL32_IMPL */
1884
1885 /* Sockets can start using the filter as soon as the filter is added */
1886 if ((error = sof_entry_add(ent)) != 0)
1887 sof_entry_free(ent);
1888
1889 return (error);
1890 }
1891
1892 /*
1893 * Socket configuration system call. It is used to add and remove
1894 * socket types.
1895 */
1896 int
sockconfig(int cmd,void * arg1,void * arg2,void * arg3,void * arg4)1897 sockconfig(int cmd, void *arg1, void *arg2, void *arg3, void *arg4)
1898 {
1899 int error = 0;
1900
1901 if (secpolicy_net_config(CRED(), B_FALSE) != 0)
1902 return (set_errno(EPERM));
1903
1904 switch (cmd) {
1905 case SOCKCONFIG_ADD_SOCK:
1906 error = sockconf_add_sock((int)(uintptr_t)arg1,
1907 (int)(uintptr_t)arg2, (int)(uintptr_t)arg3, arg4);
1908 break;
1909 case SOCKCONFIG_REMOVE_SOCK:
1910 error = sockconf_remove_sock((int)(uintptr_t)arg1,
1911 (int)(uintptr_t)arg2, (int)(uintptr_t)arg3);
1912 break;
1913 case SOCKCONFIG_ADD_FILTER:
1914 error = sockconfig_add_filter((const char *)arg1, arg2);
1915 break;
1916 case SOCKCONFIG_REMOVE_FILTER:
1917 error = sockconfig_remove_filter((const char *)arg1);
1918 break;
1919 case SOCKCONFIG_GET_SOCKTABLE:
1920 error = sockparams_copyout_socktable((int)(uintptr_t)arg1);
1921 break;
1922 default:
1923 #ifdef DEBUG
1924 cmn_err(CE_NOTE, "sockconfig: unkonwn subcommand %d", cmd);
1925 #endif
1926 error = EINVAL;
1927 break;
1928 }
1929
1930 if (error != 0) {
1931 eprintline(error);
1932 return (set_errno(error));
1933 }
1934 return (0);
1935 }
1936
1937
1938 /*
1939 * Sendfile is implemented through two schemes, direct I/O or by
1940 * caching in the filesystem page cache. We cache the input file by
1941 * default and use direct I/O only if sendfile_max_size is set
1942 * appropriately as explained below. Note that this logic is consistent
1943 * with other filesystems where caching is turned on by default
1944 * unless explicitly turned off by using the DIRECTIO ioctl.
1945 *
1946 * We choose a slightly different scheme here. One can turn off
1947 * caching by setting sendfile_max_size to 0. One can also enable
1948 * caching of files <= sendfile_max_size by setting sendfile_max_size
1949 * to an appropriate value. By default sendfile_max_size is set to the
1950 * maximum value so that all files are cached. In future, we may provide
1951 * better interfaces for caching the file.
1952 *
1953 * Sendfile through Direct I/O (Zero copy)
1954 * --------------------------------------
1955 *
1956 * As disks are normally slower than the network, we can't have a
1957 * single thread that reads the disk and writes to the network. We
1958 * need to have parallelism. This is done by having the sendfile
1959 * thread create another thread that reads from the filesystem
1960 * and queues it for network processing. In this scheme, the data
1961 * is never copied anywhere i.e it is zero copy unlike the other
1962 * scheme.
1963 *
1964 * We have a sendfile queue (snfq) where each sendfile
1965 * request (snf_req_t) is queued for processing by a thread. Number
1966 * of threads is dynamically allocated and they exit if they are idling
1967 * beyond a specified amount of time. When each request (snf_req_t) is
1968 * processed by a thread, it produces a number of mblk_t structures to
1969 * be consumed by the sendfile thread. snf_deque and snf_enque are
1970 * used for consuming and producing mblks. Size of the filesystem
1971 * read is determined by the tunable (sendfile_read_size). A single
1972 * mblk holds sendfile_read_size worth of data (except the last
1973 * read of the file) which is sent down as a whole to the network.
1974 * sendfile_read_size is set to 1 MB as this seems to be the optimal
1975 * value for the UFS filesystem backed by a striped storage array.
1976 *
1977 * Synchronisation between read (producer) and write (consumer) threads.
1978 * --------------------------------------------------------------------
1979 *
1980 * sr_lock protects sr_ib_head and sr_ib_tail. The lock is held while
1981 * adding and deleting items in this list. Error can happen anytime
1982 * during read or write. There could be unprocessed mblks in the
1983 * sr_ib_XXX list when a read or write error occurs. Whenever error
1984 * is encountered, we need two things to happen :
1985 *
1986 * a) One of the threads need to clean the mblks.
1987 * b) When one thread encounters an error, the other should stop.
1988 *
1989 * For (a), we don't want to penalize the reader thread as it could do
1990 * some useful work processing other requests. For (b), the error can
1991 * be detected by examining sr_read_error or sr_write_error.
1992 * sr_lock protects sr_read_error and sr_write_error. If both reader and
1993 * writer encounters error, we need to report the write error back to
1994 * the application as that's what would have happened if the operations
1995 * were done sequentially. With this in mind, following should work :
1996 *
1997 * - Check for errors before read or write.
1998 * - If the reader encounters error, set the error in sr_read_error.
1999 * Check sr_write_error, if it is set, send cv_signal as it is
2000 * waiting for reader to complete. If it is not set, the writer
2001 * is either running sinking data to the network or blocked
2002 * because of flow control. For handling the latter case, we
2003 * always send a signal. In any case, it will examine sr_read_error
2004 * and return. sr_read_error is marked with SR_READ_DONE to tell
2005 * the writer that the reader is done in all the cases.
2006 * - If the writer encounters error, set the error in sr_write_error.
2007 * The reader thread is either blocked because of flow control or
2008 * running reading data from the disk. For the former, we need to
2009 * wakeup the thread. Again to keep it simple, we always wake up
2010 * the reader thread. Then, wait for the read thread to complete
2011 * if it is not done yet. Cleanup and return.
2012 *
2013 * High and low water marks for the read thread.
2014 * --------------------------------------------
2015 *
2016 * If sendfile() is used to send data over a slow network, we need to
2017 * make sure that the read thread does not produce data at a faster
2018 * rate than the network. This can happen if the disk is faster than
2019 * the network. In such a case, we don't want to build a very large queue.
2020 * But we would still like to get all of the network throughput possible.
2021 * This implies that network should never block waiting for data.
2022 * As there are lot of disk throughput/network throughput combinations
2023 * possible, it is difficult to come up with an accurate number.
2024 * A typical 10K RPM disk has a max seek latency 17ms and rotational
2025 * latency of 3ms for reading a disk block. Thus, the total latency to
2026 * initiate a new read, transfer data from the disk and queue for
2027 * transmission would take about a max of 25ms. Todays max transfer rate
2028 * for network is 100MB/sec. If the thread is blocked because of flow
2029 * control, it would take 25ms to get new data ready for transmission.
2030 * We have to make sure that network is not idling, while we are initiating
2031 * new transfers. So, at 100MB/sec, to keep network busy we would need
2032 * 2.5MB of data. Rounding off, we keep the low water mark to be 3MB of data.
2033 * We need to pick a high water mark so that the woken up thread would
2034 * do considerable work before blocking again to prevent thrashing. Currently,
2035 * we pick this to be 10 times that of the low water mark.
2036 *
2037 * Sendfile with segmap caching (One copy from page cache to mblks).
2038 * ----------------------------------------------------------------
2039 *
2040 * We use the segmap cache for caching the file, if the size of file
2041 * is <= sendfile_max_size. In this case we don't use threads as VM
2042 * is reasonably fast enough to keep up with the network. If the underlying
2043 * transport allows, we call segmap_getmapflt() to map MAXBSIZE (8K) worth
2044 * of data into segmap space, and use the virtual address from segmap
2045 * directly through desballoc() to avoid copy. Once the transport is done
2046 * with the data, the mapping will be released through segmap_release()
2047 * called by the call-back routine.
2048 *
2049 * If zero-copy is not allowed by the transport, we simply call VOP_READ()
2050 * to copy the data from the filesystem into our temporary network buffer.
2051 *
2052 * To disable caching, set sendfile_max_size to 0.
2053 */
2054
2055 uint_t sendfile_read_size = 1024 * 1024;
2056 #define SENDFILE_REQ_LOWAT 3 * 1024 * 1024
2057 uint_t sendfile_req_lowat = SENDFILE_REQ_LOWAT;
2058 uint_t sendfile_req_hiwat = 10 * SENDFILE_REQ_LOWAT;
2059 struct sendfile_stats sf_stats;
2060 struct sendfile_queue *snfq;
2061 clock_t snfq_timeout;
2062 off64_t sendfile_max_size;
2063
2064 static void snf_enque(snf_req_t *, mblk_t *);
2065 static mblk_t *snf_deque(snf_req_t *);
2066
2067 void
sendfile_init(void)2068 sendfile_init(void)
2069 {
2070 snfq = kmem_zalloc(sizeof (struct sendfile_queue), KM_SLEEP);
2071
2072 mutex_init(&snfq->snfq_lock, NULL, MUTEX_DEFAULT, NULL);
2073 cv_init(&snfq->snfq_cv, NULL, CV_DEFAULT, NULL);
2074 snfq->snfq_max_threads = max_ncpus;
2075 snfq_timeout = SNFQ_TIMEOUT;
2076 /* Cache all files by default. */
2077 sendfile_max_size = MAXOFFSET_T;
2078 }
2079
2080 /*
2081 * Queues a mblk_t for network processing.
2082 */
2083 static void
snf_enque(snf_req_t * sr,mblk_t * mp)2084 snf_enque(snf_req_t *sr, mblk_t *mp)
2085 {
2086 mp->b_next = NULL;
2087 mutex_enter(&sr->sr_lock);
2088 if (sr->sr_mp_head == NULL) {
2089 sr->sr_mp_head = sr->sr_mp_tail = mp;
2090 cv_signal(&sr->sr_cv);
2091 } else {
2092 sr->sr_mp_tail->b_next = mp;
2093 sr->sr_mp_tail = mp;
2094 }
2095 sr->sr_qlen += MBLKL(mp);
2096 while ((sr->sr_qlen > sr->sr_hiwat) &&
2097 (sr->sr_write_error == 0)) {
2098 sf_stats.ss_full_waits++;
2099 cv_wait(&sr->sr_cv, &sr->sr_lock);
2100 }
2101 mutex_exit(&sr->sr_lock);
2102 }
2103
2104 /*
2105 * De-queues a mblk_t for network processing.
2106 */
2107 static mblk_t *
snf_deque(snf_req_t * sr)2108 snf_deque(snf_req_t *sr)
2109 {
2110 mblk_t *mp;
2111
2112 mutex_enter(&sr->sr_lock);
2113 /*
2114 * If we have encountered an error on read or read is
2115 * completed and no more mblks, return NULL.
2116 * We need to check for NULL sr_mp_head also as
2117 * the reads could have completed and there is
2118 * nothing more to come.
2119 */
2120 if (((sr->sr_read_error & ~SR_READ_DONE) != 0) ||
2121 ((sr->sr_read_error & SR_READ_DONE) &&
2122 sr->sr_mp_head == NULL)) {
2123 mutex_exit(&sr->sr_lock);
2124 return (NULL);
2125 }
2126 /*
2127 * To start with neither SR_READ_DONE is marked nor
2128 * the error is set. When we wake up from cv_wait,
2129 * following are the possibilities :
2130 *
2131 * a) sr_read_error is zero and mblks are queued.
2132 * b) sr_read_error is set to SR_READ_DONE
2133 * and mblks are queued.
2134 * c) sr_read_error is set to SR_READ_DONE
2135 * and no mblks.
2136 * d) sr_read_error is set to some error other
2137 * than SR_READ_DONE.
2138 */
2139
2140 while ((sr->sr_read_error == 0) && (sr->sr_mp_head == NULL)) {
2141 sf_stats.ss_empty_waits++;
2142 cv_wait(&sr->sr_cv, &sr->sr_lock);
2143 }
2144 /* Handle (a) and (b) first - the normal case. */
2145 if (((sr->sr_read_error & ~SR_READ_DONE) == 0) &&
2146 (sr->sr_mp_head != NULL)) {
2147 mp = sr->sr_mp_head;
2148 sr->sr_mp_head = mp->b_next;
2149 sr->sr_qlen -= MBLKL(mp);
2150 if (sr->sr_qlen < sr->sr_lowat)
2151 cv_signal(&sr->sr_cv);
2152 mutex_exit(&sr->sr_lock);
2153 mp->b_next = NULL;
2154 return (mp);
2155 }
2156 /* Handle (c) and (d). */
2157 mutex_exit(&sr->sr_lock);
2158 return (NULL);
2159 }
2160
2161 /*
2162 * Reads data from the filesystem and queues it for network processing.
2163 */
2164 void
snf_async_read(snf_req_t * sr)2165 snf_async_read(snf_req_t *sr)
2166 {
2167 size_t iosize;
2168 u_offset_t fileoff;
2169 u_offset_t size;
2170 int ret_size;
2171 int error;
2172 file_t *fp;
2173 mblk_t *mp;
2174 struct vnode *vp;
2175 int extra = 0;
2176 int maxblk = 0;
2177 int wroff = 0;
2178 struct sonode *so;
2179
2180 fp = sr->sr_fp;
2181 size = sr->sr_file_size;
2182 fileoff = sr->sr_file_off;
2183
2184 /*
2185 * Ignore the error for filesystems that doesn't support DIRECTIO.
2186 */
2187 (void) VOP_IOCTL(fp->f_vnode, _FIODIRECTIO, DIRECTIO_ON, 0,
2188 kcred, NULL, NULL);
2189
2190 vp = sr->sr_vp;
2191 if (vp->v_type == VSOCK) {
2192 stdata_t *stp;
2193
2194 /*
2195 * Get the extra space to insert a header and a trailer.
2196 */
2197 so = VTOSO(vp);
2198 stp = vp->v_stream;
2199 if (stp == NULL) {
2200 wroff = so->so_proto_props.sopp_wroff;
2201 maxblk = so->so_proto_props.sopp_maxblk;
2202 extra = wroff + so->so_proto_props.sopp_tail;
2203 } else {
2204 wroff = (int)(stp->sd_wroff);
2205 maxblk = (int)(stp->sd_maxblk);
2206 extra = wroff + (int)(stp->sd_tail);
2207 }
2208 }
2209
2210 while ((size != 0) && (sr->sr_write_error == 0)) {
2211
2212 iosize = (int)MIN(sr->sr_maxpsz, size);
2213
2214 /*
2215 * Socket filters can limit the mblk size,
2216 * so limit reads to maxblk if there are
2217 * filters present.
2218 */
2219 if (vp->v_type == VSOCK &&
2220 so->so_filter_active > 0 && maxblk != INFPSZ)
2221 iosize = (int)MIN(iosize, maxblk);
2222
2223 if (is_system_labeled()) {
2224 mp = allocb_cred(iosize + extra, CRED(),
2225 curproc->p_pid);
2226 } else {
2227 mp = allocb(iosize + extra, BPRI_MED);
2228 }
2229 if (mp == NULL) {
2230 error = EAGAIN;
2231 break;
2232 }
2233
2234 mp->b_rptr += wroff;
2235
2236 ret_size = soreadfile(fp, mp->b_rptr, fileoff, &error, iosize);
2237
2238 /* Error or Reached EOF ? */
2239 if ((error != 0) || (ret_size == 0)) {
2240 freeb(mp);
2241 break;
2242 }
2243 mp->b_wptr = mp->b_rptr + ret_size;
2244
2245 snf_enque(sr, mp);
2246 size -= ret_size;
2247 fileoff += ret_size;
2248 }
2249 (void) VOP_IOCTL(fp->f_vnode, _FIODIRECTIO, DIRECTIO_OFF, 0,
2250 kcred, NULL, NULL);
2251 mutex_enter(&sr->sr_lock);
2252 sr->sr_read_error = error;
2253 sr->sr_read_error |= SR_READ_DONE;
2254 cv_signal(&sr->sr_cv);
2255 mutex_exit(&sr->sr_lock);
2256 }
2257
2258 void
snf_async_thread(void)2259 snf_async_thread(void)
2260 {
2261 snf_req_t *sr;
2262 callb_cpr_t cprinfo;
2263 clock_t time_left = 1;
2264
2265 CALLB_CPR_INIT(&cprinfo, &snfq->snfq_lock, callb_generic_cpr, "snfq");
2266
2267 mutex_enter(&snfq->snfq_lock);
2268 for (;;) {
2269 /*
2270 * If we didn't find a entry, then block until woken up
2271 * again and then look through the queues again.
2272 */
2273 while ((sr = snfq->snfq_req_head) == NULL) {
2274 CALLB_CPR_SAFE_BEGIN(&cprinfo);
2275 if (time_left <= 0) {
2276 snfq->snfq_svc_threads--;
2277 CALLB_CPR_EXIT(&cprinfo);
2278 thread_exit();
2279 /* NOTREACHED */
2280 }
2281 snfq->snfq_idle_cnt++;
2282
2283 time_left = cv_reltimedwait(&snfq->snfq_cv,
2284 &snfq->snfq_lock, snfq_timeout, TR_CLOCK_TICK);
2285 snfq->snfq_idle_cnt--;
2286
2287 CALLB_CPR_SAFE_END(&cprinfo, &snfq->snfq_lock);
2288 }
2289 snfq->snfq_req_head = sr->sr_next;
2290 snfq->snfq_req_cnt--;
2291 mutex_exit(&snfq->snfq_lock);
2292 snf_async_read(sr);
2293 mutex_enter(&snfq->snfq_lock);
2294 }
2295 }
2296
2297
2298 snf_req_t *
create_thread(int operation,struct vnode * vp,file_t * fp,u_offset_t fileoff,u_offset_t size)2299 create_thread(int operation, struct vnode *vp, file_t *fp,
2300 u_offset_t fileoff, u_offset_t size)
2301 {
2302 snf_req_t *sr;
2303 stdata_t *stp;
2304
2305 sr = (snf_req_t *)kmem_zalloc(sizeof (snf_req_t), KM_SLEEP);
2306
2307 sr->sr_vp = vp;
2308 sr->sr_fp = fp;
2309 stp = vp->v_stream;
2310
2311 /*
2312 * store sd_qn_maxpsz into sr_maxpsz while we have stream head.
2313 * stream might be closed before thread returns from snf_async_read.
2314 */
2315 if (stp != NULL && stp->sd_qn_maxpsz > 0) {
2316 sr->sr_maxpsz = MIN(MAXBSIZE, stp->sd_qn_maxpsz);
2317 } else {
2318 sr->sr_maxpsz = MAXBSIZE;
2319 }
2320
2321 sr->sr_operation = operation;
2322 sr->sr_file_off = fileoff;
2323 sr->sr_file_size = size;
2324 sr->sr_hiwat = sendfile_req_hiwat;
2325 sr->sr_lowat = sendfile_req_lowat;
2326 mutex_init(&sr->sr_lock, NULL, MUTEX_DEFAULT, NULL);
2327 cv_init(&sr->sr_cv, NULL, CV_DEFAULT, NULL);
2328 /*
2329 * See whether we need another thread for servicing this
2330 * request. If there are already enough requests queued
2331 * for the threads, create one if not exceeding
2332 * snfq_max_threads.
2333 */
2334 mutex_enter(&snfq->snfq_lock);
2335 if (snfq->snfq_req_cnt >= snfq->snfq_idle_cnt &&
2336 snfq->snfq_svc_threads < snfq->snfq_max_threads) {
2337 (void) thread_create(NULL, 0, &snf_async_thread, 0, 0, &p0,
2338 TS_RUN, minclsyspri);
2339 snfq->snfq_svc_threads++;
2340 }
2341 if (snfq->snfq_req_head == NULL) {
2342 snfq->snfq_req_head = snfq->snfq_req_tail = sr;
2343 cv_signal(&snfq->snfq_cv);
2344 } else {
2345 snfq->snfq_req_tail->sr_next = sr;
2346 snfq->snfq_req_tail = sr;
2347 }
2348 snfq->snfq_req_cnt++;
2349 mutex_exit(&snfq->snfq_lock);
2350 return (sr);
2351 }
2352
2353 int
snf_direct_io(file_t * fp,file_t * rfp,u_offset_t fileoff,u_offset_t size,ssize_t * count)2354 snf_direct_io(file_t *fp, file_t *rfp, u_offset_t fileoff, u_offset_t size,
2355 ssize_t *count)
2356 {
2357 snf_req_t *sr;
2358 mblk_t *mp;
2359 int iosize;
2360 int error = 0;
2361 short fflag;
2362 struct vnode *vp;
2363 int ksize;
2364 struct nmsghdr msg;
2365
2366 ksize = 0;
2367 *count = 0;
2368 bzero(&msg, sizeof (msg));
2369
2370 vp = fp->f_vnode;
2371 fflag = fp->f_flag;
2372 if ((sr = create_thread(READ_OP, vp, rfp, fileoff, size)) == NULL)
2373 return (EAGAIN);
2374
2375 /*
2376 * We check for read error in snf_deque. It has to check
2377 * for successful READ_DONE and return NULL, and we might
2378 * as well make an additional check there.
2379 */
2380 while ((mp = snf_deque(sr)) != NULL) {
2381
2382 if (ISSIG(curthread, JUSTLOOKING)) {
2383 freeb(mp);
2384 error = EINTR;
2385 break;
2386 }
2387 iosize = MBLKL(mp);
2388
2389 error = socket_sendmblk(VTOSO(vp), &msg, fflag, CRED(), &mp);
2390
2391 if (error != 0) {
2392 if (mp != NULL)
2393 freeb(mp);
2394 break;
2395 }
2396 ksize += iosize;
2397 }
2398 *count = ksize;
2399
2400 mutex_enter(&sr->sr_lock);
2401 sr->sr_write_error = error;
2402 /* Look at the big comments on why we cv_signal here. */
2403 cv_signal(&sr->sr_cv);
2404
2405 /* Wait for the reader to complete always. */
2406 while (!(sr->sr_read_error & SR_READ_DONE)) {
2407 cv_wait(&sr->sr_cv, &sr->sr_lock);
2408 }
2409 /* If there is no write error, check for read error. */
2410 if (error == 0)
2411 error = (sr->sr_read_error & ~SR_READ_DONE);
2412
2413 if (error != 0) {
2414 mblk_t *next_mp;
2415
2416 mp = sr->sr_mp_head;
2417 while (mp != NULL) {
2418 next_mp = mp->b_next;
2419 mp->b_next = NULL;
2420 freeb(mp);
2421 mp = next_mp;
2422 }
2423 }
2424 mutex_exit(&sr->sr_lock);
2425 kmem_free(sr, sizeof (snf_req_t));
2426 return (error);
2427 }
2428
2429 /* Maximum no.of pages allocated by vpm for sendfile at a time */
2430 #define SNF_VPMMAXPGS (VPMMAXPGS/2)
2431
2432 /*
2433 * Maximum no.of elements in the list returned by vpm, including
2434 * NULL for the last entry
2435 */
2436 #define SNF_MAXVMAPS (SNF_VPMMAXPGS + 1)
2437
2438 typedef struct {
2439 unsigned int snfv_ref;
2440 frtn_t snfv_frtn;
2441 vnode_t *snfv_vp;
2442 struct vmap snfv_vml[SNF_MAXVMAPS];
2443 } snf_vmap_desbinfo;
2444
2445 typedef struct {
2446 frtn_t snfi_frtn;
2447 caddr_t snfi_base;
2448 uint_t snfi_mapoff;
2449 size_t snfi_len;
2450 vnode_t *snfi_vp;
2451 } snf_smap_desbinfo;
2452
2453 /*
2454 * The callback function used for vpm mapped mblks called when the last ref of
2455 * the mblk is dropped which normally occurs when TCP receives the ack. But it
2456 * can be the driver too due to lazy reclaim.
2457 */
2458 void
snf_vmap_desbfree(snf_vmap_desbinfo * snfv)2459 snf_vmap_desbfree(snf_vmap_desbinfo *snfv)
2460 {
2461 ASSERT(snfv->snfv_ref != 0);
2462 if (atomic_dec_32_nv(&snfv->snfv_ref) == 0) {
2463 vpm_unmap_pages(snfv->snfv_vml, S_READ);
2464 VN_RELE(snfv->snfv_vp);
2465 kmem_free(snfv, sizeof (snf_vmap_desbinfo));
2466 }
2467 }
2468
2469 /*
2470 * The callback function used for segmap'ped mblks called when the last ref of
2471 * the mblk is dropped which normally occurs when TCP receives the ack. But it
2472 * can be the driver too due to lazy reclaim.
2473 */
2474 void
snf_smap_desbfree(snf_smap_desbinfo * snfi)2475 snf_smap_desbfree(snf_smap_desbinfo *snfi)
2476 {
2477 if (! IS_KPM_ADDR(snfi->snfi_base)) {
2478 /*
2479 * We don't need to call segmap_fault(F_SOFTUNLOCK) for
2480 * segmap_kpm as long as the latter never falls back to
2481 * "use_segmap_range". (See segmap_getmapflt().)
2482 *
2483 * Using S_OTHER saves an redundant hat_setref() in
2484 * segmap_unlock()
2485 */
2486 (void) segmap_fault(kas.a_hat, segkmap,
2487 (caddr_t)(uintptr_t)(((uintptr_t)snfi->snfi_base +
2488 snfi->snfi_mapoff) & PAGEMASK), snfi->snfi_len,
2489 F_SOFTUNLOCK, S_OTHER);
2490 }
2491 (void) segmap_release(segkmap, snfi->snfi_base, SM_DONTNEED);
2492 VN_RELE(snfi->snfi_vp);
2493 kmem_free(snfi, sizeof (*snfi));
2494 }
2495
2496 /*
2497 * Use segmap or vpm instead of bcopy to send down a desballoca'ed, mblk.
2498 * When segmap is used, the mblk contains a segmap slot of no more
2499 * than MAXBSIZE.
2500 *
2501 * With vpm, a maximum of SNF_MAXVMAPS page-sized mappings can be obtained
2502 * in each iteration and sent by socket_sendmblk until an error occurs or
2503 * the requested size has been transferred. An mblk is esballoca'ed from
2504 * each mapped page and a chain of these mblk is sent to the transport layer.
2505 * vpm will be called to unmap the pages when all mblks have been freed by
2506 * free_func.
2507 *
2508 * At the end of the whole sendfile() operation, we wait till the data from
2509 * the last mblk is ack'ed by the transport before returning so that the
2510 * caller of sendfile() can safely modify the file content.
2511 *
2512 * The caller of this function should make sure that total_size does not exceed
2513 * the actual file size of fvp.
2514 */
2515 int
snf_segmap(file_t * fp,vnode_t * fvp,u_offset_t fileoff,u_offset_t total_size,ssize_t * count,boolean_t nowait)2516 snf_segmap(file_t *fp, vnode_t *fvp, u_offset_t fileoff, u_offset_t total_size,
2517 ssize_t *count, boolean_t nowait)
2518 {
2519 caddr_t base;
2520 int mapoff;
2521 vnode_t *vp;
2522 mblk_t *mp = NULL;
2523 int chain_size;
2524 int error;
2525 clock_t deadlk_wait;
2526 short fflag;
2527 int ksize;
2528 struct vattr va;
2529 boolean_t dowait = B_FALSE;
2530 struct nmsghdr msg;
2531
2532 vp = fp->f_vnode;
2533 fflag = fp->f_flag;
2534 ksize = 0;
2535 bzero(&msg, sizeof (msg));
2536
2537 for (;;) {
2538 if (ISSIG(curthread, JUSTLOOKING)) {
2539 error = EINTR;
2540 break;
2541 }
2542
2543 if (vpm_enable) {
2544 snf_vmap_desbinfo *snfv;
2545 mblk_t *nmp;
2546 int mblk_size;
2547 int maxsize;
2548 int i;
2549
2550 mapoff = fileoff & PAGEOFFSET;
2551 maxsize = MIN((SNF_VPMMAXPGS * PAGESIZE), total_size);
2552
2553 snfv = kmem_zalloc(sizeof (snf_vmap_desbinfo),
2554 KM_SLEEP);
2555
2556 /*
2557 * Get vpm mappings for maxsize with read access.
2558 * If the pages aren't available yet, we get
2559 * DEADLK, so wait and try again a little later using
2560 * an increasing wait. We might be here a long time.
2561 *
2562 * If delay_sig returns EINTR, be sure to exit and
2563 * pass it up to the caller.
2564 */
2565 deadlk_wait = 0;
2566 while ((error = vpm_map_pages(fvp, fileoff,
2567 (size_t)maxsize, (VPM_FETCHPAGE), snfv->snfv_vml,
2568 SNF_MAXVMAPS, NULL, S_READ)) == EDEADLK) {
2569 deadlk_wait += (deadlk_wait < 5) ? 1 : 4;
2570 if ((error = delay_sig(deadlk_wait)) != 0) {
2571 break;
2572 }
2573 }
2574 if (error != 0) {
2575 kmem_free(snfv, sizeof (snf_vmap_desbinfo));
2576 error = (error == EINTR) ? EINTR : EIO;
2577 goto out;
2578 }
2579 snfv->snfv_frtn.free_func = snf_vmap_desbfree;
2580 snfv->snfv_frtn.free_arg = (caddr_t)snfv;
2581
2582 /* Construct the mblk chain from the page mappings */
2583 chain_size = 0;
2584 for (i = 0; (snfv->snfv_vml[i].vs_addr != NULL) &&
2585 total_size > 0; i++) {
2586 ASSERT(chain_size < maxsize);
2587 mblk_size = MIN(snfv->snfv_vml[i].vs_len -
2588 mapoff, total_size);
2589 nmp = esballoca(
2590 (uchar_t *)snfv->snfv_vml[i].vs_addr +
2591 mapoff, mblk_size, BPRI_HI,
2592 &snfv->snfv_frtn);
2593
2594 /*
2595 * We return EAGAIN after unmapping the pages
2596 * if we cannot allocate the the head of the
2597 * chain. Otherwise, we continue sending the
2598 * mblks constructed so far.
2599 */
2600 if (nmp == NULL) {
2601 if (i == 0) {
2602 vpm_unmap_pages(snfv->snfv_vml,
2603 S_READ);
2604 kmem_free(snfv,
2605 sizeof (snf_vmap_desbinfo));
2606 error = EAGAIN;
2607 goto out;
2608 }
2609 break;
2610 }
2611 /* Mark this dblk with the zero-copy flag */
2612 nmp->b_datap->db_struioflag |= STRUIO_ZC;
2613 nmp->b_wptr += mblk_size;
2614 chain_size += mblk_size;
2615 fileoff += mblk_size;
2616 total_size -= mblk_size;
2617 snfv->snfv_ref++;
2618 mapoff = 0;
2619 if (i > 0)
2620 linkb(mp, nmp);
2621 else
2622 mp = nmp;
2623 }
2624 VN_HOLD(fvp);
2625 snfv->snfv_vp = fvp;
2626 } else {
2627 /* vpm not supported. fallback to segmap */
2628 snf_smap_desbinfo *snfi;
2629
2630 mapoff = fileoff & MAXBOFFSET;
2631 chain_size = MAXBSIZE - mapoff;
2632 if (chain_size > total_size)
2633 chain_size = total_size;
2634 /*
2635 * we don't forcefault because we'll call
2636 * segmap_fault(F_SOFTLOCK) next.
2637 *
2638 * S_READ will get the ref bit set (by either
2639 * segmap_getmapflt() or segmap_fault()) and page
2640 * shared locked.
2641 */
2642 base = segmap_getmapflt(segkmap, fvp, fileoff,
2643 chain_size, segmap_kpm ? SM_FAULT : 0, S_READ);
2644
2645 snfi = kmem_alloc(sizeof (*snfi), KM_SLEEP);
2646 snfi->snfi_len = (size_t)roundup(mapoff+chain_size,
2647 PAGESIZE)- (mapoff & PAGEMASK);
2648 /*
2649 * We must call segmap_fault() even for segmap_kpm
2650 * because that's how error gets returned.
2651 * (segmap_getmapflt() never fails but segmap_fault()
2652 * does.)
2653 *
2654 * If the pages aren't available yet, we get
2655 * DEADLK, so wait and try again a little later using
2656 * an increasing wait. We might be here a long time.
2657 *
2658 * If delay_sig returns EINTR, be sure to exit and
2659 * pass it up to the caller.
2660 */
2661 deadlk_wait = 0;
2662 while ((error = FC_ERRNO(segmap_fault(kas.a_hat,
2663 segkmap, (caddr_t)(uintptr_t)(((uintptr_t)base +
2664 mapoff) & PAGEMASK), snfi->snfi_len, F_SOFTLOCK,
2665 S_READ))) == EDEADLK) {
2666 deadlk_wait += (deadlk_wait < 5) ? 1 : 4;
2667 if ((error = delay_sig(deadlk_wait)) != 0) {
2668 break;
2669 }
2670 }
2671 if (error != 0) {
2672 (void) segmap_release(segkmap, base, 0);
2673 kmem_free(snfi, sizeof (*snfi));
2674 error = (error == EINTR) ? EINTR : EIO;
2675 goto out;
2676 }
2677 snfi->snfi_frtn.free_func = snf_smap_desbfree;
2678 snfi->snfi_frtn.free_arg = (caddr_t)snfi;
2679 snfi->snfi_base = base;
2680 snfi->snfi_mapoff = mapoff;
2681 mp = esballoca((uchar_t *)base + mapoff, chain_size,
2682 BPRI_HI, &snfi->snfi_frtn);
2683
2684 if (mp == NULL) {
2685 (void) segmap_fault(kas.a_hat, segkmap,
2686 (caddr_t)(uintptr_t)(((uintptr_t)base +
2687 mapoff) & PAGEMASK), snfi->snfi_len,
2688 F_SOFTUNLOCK, S_OTHER);
2689 (void) segmap_release(segkmap, base, 0);
2690 kmem_free(snfi, sizeof (*snfi));
2691 freemsg(mp);
2692 error = EAGAIN;
2693 goto out;
2694 }
2695 VN_HOLD(fvp);
2696 snfi->snfi_vp = fvp;
2697 mp->b_wptr += chain_size;
2698
2699 /* Mark this dblk with the zero-copy flag */
2700 mp->b_datap->db_struioflag |= STRUIO_ZC;
2701 fileoff += chain_size;
2702 total_size -= chain_size;
2703 }
2704
2705 if (total_size == 0 && !nowait) {
2706 ASSERT(!dowait);
2707 dowait = B_TRUE;
2708 mp->b_datap->db_struioflag |= STRUIO_ZCNOTIFY;
2709 }
2710 VOP_RWUNLOCK(fvp, V_WRITELOCK_FALSE, NULL);
2711 error = socket_sendmblk(VTOSO(vp), &msg, fflag, CRED(), &mp);
2712 if (error != 0) {
2713 /*
2714 * mp contains the mblks that were not sent by
2715 * socket_sendmblk. Use its size to update *count
2716 */
2717 *count = ksize + (chain_size - msgdsize(mp));
2718 if (mp != NULL)
2719 freemsg(mp);
2720 return (error);
2721 }
2722 ksize += chain_size;
2723 if (total_size == 0)
2724 goto done;
2725
2726 (void) VOP_RWLOCK(fvp, V_WRITELOCK_FALSE, NULL);
2727 va.va_mask = AT_SIZE;
2728 error = VOP_GETATTR(fvp, &va, 0, kcred, NULL);
2729 if (error)
2730 break;
2731 /* Read as much as possible. */
2732 if (fileoff >= va.va_size)
2733 break;
2734 if (total_size + fileoff > va.va_size)
2735 total_size = va.va_size - fileoff;
2736 }
2737 out:
2738 VOP_RWUNLOCK(fvp, V_WRITELOCK_FALSE, NULL);
2739 done:
2740 *count = ksize;
2741 if (dowait) {
2742 stdata_t *stp;
2743
2744 stp = vp->v_stream;
2745 if (stp == NULL) {
2746 struct sonode *so;
2747 so = VTOSO(vp);
2748 error = so_zcopy_wait(so);
2749 } else {
2750 mutex_enter(&stp->sd_lock);
2751 while (!(stp->sd_flag & STZCNOTIFY)) {
2752 if (cv_wait_sig(&stp->sd_zcopy_wait,
2753 &stp->sd_lock) == 0) {
2754 error = EINTR;
2755 break;
2756 }
2757 }
2758 stp->sd_flag &= ~STZCNOTIFY;
2759 mutex_exit(&stp->sd_lock);
2760 }
2761 }
2762 return (error);
2763 }
2764
2765 int
snf_cache(file_t * fp,vnode_t * fvp,u_offset_t fileoff,u_offset_t size,uint_t maxpsz,ssize_t * count)2766 snf_cache(file_t *fp, vnode_t *fvp, u_offset_t fileoff, u_offset_t size,
2767 uint_t maxpsz, ssize_t *count)
2768 {
2769 struct vnode *vp;
2770 mblk_t *mp;
2771 int iosize;
2772 int extra = 0;
2773 int error;
2774 short fflag;
2775 int ksize;
2776 int ioflag;
2777 struct uio auio;
2778 struct iovec aiov;
2779 struct vattr va;
2780 int maxblk = 0;
2781 int wroff = 0;
2782 struct sonode *so;
2783 struct nmsghdr msg;
2784
2785 vp = fp->f_vnode;
2786 if (vp->v_type == VSOCK) {
2787 stdata_t *stp;
2788
2789 /*
2790 * Get the extra space to insert a header and a trailer.
2791 */
2792 so = VTOSO(vp);
2793 stp = vp->v_stream;
2794 if (stp == NULL) {
2795 wroff = so->so_proto_props.sopp_wroff;
2796 maxblk = so->so_proto_props.sopp_maxblk;
2797 extra = wroff + so->so_proto_props.sopp_tail;
2798 } else {
2799 wroff = (int)(stp->sd_wroff);
2800 maxblk = (int)(stp->sd_maxblk);
2801 extra = wroff + (int)(stp->sd_tail);
2802 }
2803 }
2804 bzero(&msg, sizeof (msg));
2805 fflag = fp->f_flag;
2806 ksize = 0;
2807 auio.uio_iov = &aiov;
2808 auio.uio_iovcnt = 1;
2809 auio.uio_segflg = UIO_SYSSPACE;
2810 auio.uio_llimit = MAXOFFSET_T;
2811 auio.uio_fmode = fflag;
2812 auio.uio_extflg = UIO_COPY_CACHED;
2813 ioflag = auio.uio_fmode & (FSYNC|FDSYNC|FRSYNC);
2814 /* If read sync is not asked for, filter sync flags */
2815 if ((ioflag & FRSYNC) == 0)
2816 ioflag &= ~(FSYNC|FDSYNC);
2817 for (;;) {
2818 if (ISSIG(curthread, JUSTLOOKING)) {
2819 error = EINTR;
2820 break;
2821 }
2822 iosize = (int)MIN(maxpsz, size);
2823
2824 /*
2825 * Socket filters can limit the mblk size,
2826 * so limit reads to maxblk if there are
2827 * filters present.
2828 */
2829 if (vp->v_type == VSOCK &&
2830 so->so_filter_active > 0 && maxblk != INFPSZ)
2831 iosize = (int)MIN(iosize, maxblk);
2832
2833 if (is_system_labeled()) {
2834 mp = allocb_cred(iosize + extra, CRED(),
2835 curproc->p_pid);
2836 } else {
2837 mp = allocb(iosize + extra, BPRI_MED);
2838 }
2839 if (mp == NULL) {
2840 error = EAGAIN;
2841 break;
2842 }
2843
2844 mp->b_rptr += wroff;
2845
2846 aiov.iov_base = (caddr_t)mp->b_rptr;
2847 aiov.iov_len = iosize;
2848 auio.uio_loffset = fileoff;
2849 auio.uio_resid = iosize;
2850
2851 error = VOP_READ(fvp, &auio, ioflag, fp->f_cred, NULL);
2852 iosize -= auio.uio_resid;
2853
2854 if (error == EINTR && iosize != 0)
2855 error = 0;
2856
2857 if (error != 0 || iosize == 0) {
2858 freeb(mp);
2859 break;
2860 }
2861 mp->b_wptr = mp->b_rptr + iosize;
2862
2863 VOP_RWUNLOCK(fvp, V_WRITELOCK_FALSE, NULL);
2864
2865 error = socket_sendmblk(VTOSO(vp), &msg, fflag, CRED(), &mp);
2866
2867 if (error != 0) {
2868 *count = ksize;
2869 if (mp != NULL)
2870 freeb(mp);
2871 return (error);
2872 }
2873 ksize += iosize;
2874 size -= iosize;
2875 if (size == 0)
2876 goto done;
2877
2878 fileoff += iosize;
2879 (void) VOP_RWLOCK(fvp, V_WRITELOCK_FALSE, NULL);
2880 va.va_mask = AT_SIZE;
2881 error = VOP_GETATTR(fvp, &va, 0, kcred, NULL);
2882 if (error)
2883 break;
2884 /* Read as much as possible. */
2885 if (fileoff >= va.va_size)
2886 size = 0;
2887 else if (size + fileoff > va.va_size)
2888 size = va.va_size - fileoff;
2889 }
2890 VOP_RWUNLOCK(fvp, V_WRITELOCK_FALSE, NULL);
2891 done:
2892 *count = ksize;
2893 return (error);
2894 }
2895
2896 #if defined(_SYSCALL32_IMPL) || defined(_ILP32)
2897 /*
2898 * Largefile support for 32 bit applications only.
2899 */
2900 int
sosendfile64(file_t * fp,file_t * rfp,const struct ksendfilevec64 * sfv,ssize32_t * count32)2901 sosendfile64(file_t *fp, file_t *rfp, const struct ksendfilevec64 *sfv,
2902 ssize32_t *count32)
2903 {
2904 ssize32_t sfv_len;
2905 u_offset_t sfv_off, va_size;
2906 struct vnode *vp, *fvp, *realvp;
2907 struct vattr va;
2908 stdata_t *stp;
2909 ssize_t count = 0;
2910 int error = 0;
2911 boolean_t dozcopy = B_FALSE;
2912 uint_t maxpsz;
2913
2914 sfv_len = (ssize32_t)sfv->sfv_len;
2915 if (sfv_len < 0) {
2916 error = EINVAL;
2917 goto out;
2918 }
2919
2920 if (sfv_len == 0) goto out;
2921
2922 sfv_off = (u_offset_t)sfv->sfv_off;
2923
2924 /* Same checks as in pread */
2925 if (sfv_off > MAXOFFSET_T) {
2926 error = EINVAL;
2927 goto out;
2928 }
2929 if (sfv_off + sfv_len > MAXOFFSET_T)
2930 sfv_len = (ssize32_t)(MAXOFFSET_T - sfv_off);
2931
2932 /*
2933 * There are no more checks on sfv_len. So, we cast it to
2934 * u_offset_t and share the snf_direct_io/snf_cache code between
2935 * 32 bit and 64 bit.
2936 *
2937 * TODO: should do nbl_need_check() like read()?
2938 */
2939 if (sfv_len > sendfile_max_size) {
2940 sf_stats.ss_file_not_cached++;
2941 error = snf_direct_io(fp, rfp, sfv_off, (u_offset_t)sfv_len,
2942 &count);
2943 goto out;
2944 }
2945 fvp = rfp->f_vnode;
2946 if (VOP_REALVP(fvp, &realvp, NULL) == 0)
2947 fvp = realvp;
2948 /*
2949 * Grab the lock as a reader to prevent the file size
2950 * from changing underneath.
2951 */
2952 (void) VOP_RWLOCK(fvp, V_WRITELOCK_FALSE, NULL);
2953 va.va_mask = AT_SIZE;
2954 error = VOP_GETATTR(fvp, &va, 0, kcred, NULL);
2955 va_size = va.va_size;
2956 if ((error != 0) || (va_size == 0) || (sfv_off >= va_size)) {
2957 VOP_RWUNLOCK(fvp, V_WRITELOCK_FALSE, NULL);
2958 goto out;
2959 }
2960 /* Read as much as possible. */
2961 if (sfv_off + sfv_len > va_size)
2962 sfv_len = va_size - sfv_off;
2963
2964 vp = fp->f_vnode;
2965 stp = vp->v_stream;
2966 /*
2967 * When the NOWAIT flag is not set, we enable zero-copy only if the
2968 * transfer size is large enough. This prevents performance loss
2969 * when the caller sends the file piece by piece.
2970 */
2971 if (sfv_len >= MAXBSIZE && (sfv_len >= (va_size >> 1) ||
2972 (sfv->sfv_flag & SFV_NOWAIT) || sfv_len >= 0x1000000) &&
2973 !vn_has_flocks(fvp) && !(fvp->v_flag & VNOMAP)) {
2974 uint_t copyflag;
2975 copyflag = stp != NULL ? stp->sd_copyflag :
2976 VTOSO(vp)->so_proto_props.sopp_zcopyflag;
2977 if ((copyflag & (STZCVMSAFE|STZCVMUNSAFE)) == 0) {
2978 int on = 1;
2979
2980 if (socket_setsockopt(VTOSO(vp), SOL_SOCKET,
2981 SO_SND_COPYAVOID, &on, sizeof (on), CRED()) == 0)
2982 dozcopy = B_TRUE;
2983 } else {
2984 dozcopy = copyflag & STZCVMSAFE;
2985 }
2986 }
2987 if (dozcopy) {
2988 sf_stats.ss_file_segmap++;
2989 error = snf_segmap(fp, fvp, sfv_off, (u_offset_t)sfv_len,
2990 &count, ((sfv->sfv_flag & SFV_NOWAIT) != 0));
2991 } else {
2992 if (vp->v_type == VSOCK && stp == NULL) {
2993 sonode_t *so = VTOSO(vp);
2994 maxpsz = so->so_proto_props.sopp_maxpsz;
2995 } else if (stp != NULL) {
2996 maxpsz = stp->sd_qn_maxpsz;
2997 } else {
2998 maxpsz = maxphys;
2999 }
3000
3001 if (maxpsz == INFPSZ)
3002 maxpsz = maxphys;
3003 else
3004 maxpsz = roundup(maxpsz, MAXBSIZE);
3005 sf_stats.ss_file_cached++;
3006 error = snf_cache(fp, fvp, sfv_off, (u_offset_t)sfv_len,
3007 maxpsz, &count);
3008 }
3009 out:
3010 releasef(sfv->sfv_fd);
3011 *count32 = (ssize32_t)count;
3012 return (error);
3013 }
3014 #endif
3015
3016 #ifdef _SYSCALL32_IMPL
3017 /*
3018 * recv32(), recvfrom32(), send32(), sendto32(): intentionally return a
3019 * ssize_t rather than ssize32_t; see the comments above read32 for details.
3020 */
3021
3022 ssize_t
recv32(int32_t sock,caddr32_t buffer,size32_t len,int32_t flags)3023 recv32(int32_t sock, caddr32_t buffer, size32_t len, int32_t flags)
3024 {
3025 return (recv(sock, (void *)(uintptr_t)buffer, (ssize32_t)len, flags));
3026 }
3027
3028 ssize_t
recvfrom32(int32_t sock,caddr32_t buffer,size32_t len,int32_t flags,caddr32_t name,caddr32_t namelenp)3029 recvfrom32(int32_t sock, caddr32_t buffer, size32_t len, int32_t flags,
3030 caddr32_t name, caddr32_t namelenp)
3031 {
3032 return (recvfrom(sock, (void *)(uintptr_t)buffer, (ssize32_t)len, flags,
3033 (void *)(uintptr_t)name, (void *)(uintptr_t)namelenp));
3034 }
3035
3036 ssize_t
send32(int32_t sock,caddr32_t buffer,size32_t len,int32_t flags)3037 send32(int32_t sock, caddr32_t buffer, size32_t len, int32_t flags)
3038 {
3039 return (send(sock, (void *)(uintptr_t)buffer, (ssize32_t)len, flags));
3040 }
3041
3042 ssize_t
sendto32(int32_t sock,caddr32_t buffer,size32_t len,int32_t flags,caddr32_t name,socklen_t namelen)3043 sendto32(int32_t sock, caddr32_t buffer, size32_t len, int32_t flags,
3044 caddr32_t name, socklen_t namelen)
3045 {
3046 return (sendto(sock, (void *)(uintptr_t)buffer, (ssize32_t)len, flags,
3047 (void *)(uintptr_t)name, namelen));
3048 }
3049 #endif /* _SYSCALL32_IMPL */
3050
3051 /*
3052 * Function wrappers (mostly around the sonode switch) for
3053 * backward compatibility.
3054 */
3055
3056 int
soaccept(struct sonode * so,int fflag,struct sonode ** nsop)3057 soaccept(struct sonode *so, int fflag, struct sonode **nsop)
3058 {
3059 return (socket_accept(so, fflag, CRED(), nsop));
3060 }
3061
3062 int
sobind(struct sonode * so,struct sockaddr * name,socklen_t namelen,int backlog,int flags)3063 sobind(struct sonode *so, struct sockaddr *name, socklen_t namelen,
3064 int backlog, int flags)
3065 {
3066 int error;
3067
3068 error = socket_bind(so, name, namelen, flags, CRED());
3069 if (error == 0 && backlog != 0)
3070 return (socket_listen(so, backlog, CRED()));
3071
3072 return (error);
3073 }
3074
3075 int
solisten(struct sonode * so,int backlog)3076 solisten(struct sonode *so, int backlog)
3077 {
3078 return (socket_listen(so, backlog, CRED()));
3079 }
3080
3081 int
soconnect(struct sonode * so,struct sockaddr * name,socklen_t namelen,int fflag,int flags)3082 soconnect(struct sonode *so, struct sockaddr *name, socklen_t namelen,
3083 int fflag, int flags)
3084 {
3085 return (socket_connect(so, name, namelen, fflag, flags, CRED()));
3086 }
3087
3088 int
sorecvmsg(struct sonode * so,struct nmsghdr * msg,struct uio * uiop)3089 sorecvmsg(struct sonode *so, struct nmsghdr *msg, struct uio *uiop)
3090 {
3091 return (socket_recvmsg(so, msg, uiop, CRED()));
3092 }
3093
3094 int
sosendmsg(struct sonode * so,struct nmsghdr * msg,struct uio * uiop)3095 sosendmsg(struct sonode *so, struct nmsghdr *msg, struct uio *uiop)
3096 {
3097 return (socket_sendmsg(so, msg, uiop, CRED()));
3098 }
3099
3100 int
soshutdown(struct sonode * so,int how)3101 soshutdown(struct sonode *so, int how)
3102 {
3103 return (socket_shutdown(so, how, CRED()));
3104 }
3105
3106 int
sogetsockopt(struct sonode * so,int level,int option_name,void * optval,socklen_t * optlenp,int flags)3107 sogetsockopt(struct sonode *so, int level, int option_name, void *optval,
3108 socklen_t *optlenp, int flags)
3109 {
3110 return (socket_getsockopt(so, level, option_name, optval, optlenp,
3111 flags, CRED()));
3112 }
3113
3114 int
sosetsockopt(struct sonode * so,int level,int option_name,const void * optval,t_uscalar_t optlen)3115 sosetsockopt(struct sonode *so, int level, int option_name, const void *optval,
3116 t_uscalar_t optlen)
3117 {
3118 return (socket_setsockopt(so, level, option_name, optval, optlen,
3119 CRED()));
3120 }
3121
3122 /*
3123 * Because this is backward compatibility interface it only needs to be
3124 * able to handle the creation of TPI sockfs sockets.
3125 */
3126 struct sonode *
socreate(struct sockparams * sp,int family,int type,int protocol,int version,int * errorp)3127 socreate(struct sockparams *sp, int family, int type, int protocol, int version,
3128 int *errorp)
3129 {
3130 struct sonode *so;
3131
3132 ASSERT(sp != NULL);
3133
3134 so = sp->sp_smod_info->smod_sock_create_func(sp, family, type, protocol,
3135 version, SOCKET_SLEEP, errorp, CRED());
3136 if (so == NULL) {
3137 SOCKPARAMS_DEC_REF(sp);
3138 } else {
3139 if ((*errorp = SOP_INIT(so, NULL, CRED(), SOCKET_SLEEP)) == 0) {
3140 /* Cannot fail, only bumps so_count */
3141 (void) VOP_OPEN(&SOTOV(so), FREAD|FWRITE, CRED(), NULL);
3142 } else {
3143 socket_destroy(so);
3144 so = NULL;
3145 }
3146 }
3147 return (so);
3148 }
3149