1 /*-
2 * Copyright (c) 2013-2015 Gleb Smirnoff <glebius@FreeBSD.org>
3 * Copyright (c) 1998, David Greenman. All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
13 * 3. Neither the name of the University nor the names of its contributors
14 * may be used to endorse or promote products derived from this software
15 * without specific prior written permission.
16 *
17 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27 * SUCH DAMAGE.
28 */
29
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
32
33 #include <sys/param.h>
34 #include <sys/systm.h>
35 #include <sys/capsicum.h>
36 #include <sys/kernel.h>
37 #include <sys/lock.h>
38 #include <sys/mutex.h>
39 #include <sys/malloc.h>
40 #include <sys/mman.h>
41 #include <sys/mount.h>
42 #include <sys/mbuf.h>
43 #include <sys/proc.h>
44 #include <sys/protosw.h>
45 #include <sys/rwlock.h>
46 #include <sys/sf_buf.h>
47 #include <sys/socket.h>
48 #include <sys/socketvar.h>
49 #include <sys/syscallsubr.h>
50 #include <sys/sysctl.h>
51 #include <sys/sysproto.h>
52 #include <sys/vnode.h>
53
54 #include <net/vnet.h>
55
56 #include <netinet/in.h>
57
58 #include <security/audit/audit.h>
59 #include <security/mac/mac_framework.h>
60
61 #include <vm/vm.h>
62 #include <vm/vm_object.h>
63 #include <vm/vm_pager.h>
64
65 static MALLOC_DEFINE(M_SENDFILE, "sendfile", "sendfile dynamic memory");
66
67 #define EXT_FLAG_SYNC EXT_FLAG_VENDOR1
68 #define EXT_FLAG_NOCACHE EXT_FLAG_VENDOR2
69
70 /*
71 * Structure describing a single sendfile(2) I/O, which may consist of
72 * several underlying pager I/Os.
73 *
74 * The syscall context allocates the structure and initializes 'nios'
75 * to 1. As sendfile_swapin() runs through pages and starts asynchronous
76 * paging operations, it increments 'nios'.
77 *
78 * Every I/O completion calls sendfile_iodone(), which decrements the 'nios',
79 * and the syscall also calls sendfile_iodone() after allocating all mbufs,
80 * linking them and sending to socket. Whoever reaches zero 'nios' is
81 * responsible to * call pru_ready on the socket, to notify it of readyness
82 * of the data.
83 */
84 struct sf_io {
85 volatile u_int nios;
86 u_int error;
87 int npages;
88 struct socket *so;
89 struct mbuf *m;
90 vm_page_t pa[];
91 };
92
93 /*
94 * Structure used to track requests with SF_SYNC flag.
95 */
96 struct sendfile_sync {
97 struct mtx mtx;
98 struct cv cv;
99 unsigned count;
100 bool waiting;
101 };
102
103 static void
104 sendfile_sync_destroy(struct sendfile_sync *sfs)
105 {
106 KASSERT(sfs->count == 0, ("sendfile sync %p still busy", sfs));
107
108 cv_destroy(&sfs->cv);
109 mtx_destroy(&sfs->mtx);
110 free(sfs, M_SENDFILE);
111 }
112
113 static void
114 sendfile_sync_signal(struct sendfile_sync *sfs)
115 {
116 mtx_lock(&sfs->mtx);
117 KASSERT(sfs->count > 0, ("sendfile sync %p not busy", sfs));
118 if (--sfs->count == 0) {
119 if (!sfs->waiting) {
120 /* The sendfile() waiter was interrupted by a signal. */
121 sendfile_sync_destroy(sfs);
122 return;
123 } else {
124 cv_signal(&sfs->cv);
125 }
126 }
127 mtx_unlock(&sfs->mtx);
128 }
129
130 counter_u64_t sfstat[sizeof(struct sfstat) / sizeof(uint64_t)];
131
132 static void
133 sfstat_init(const void *unused)
134 {
135
136 COUNTER_ARRAY_ALLOC(sfstat, sizeof(struct sfstat) / sizeof(uint64_t),
137 M_WAITOK);
138 }
139 SYSINIT(sfstat, SI_SUB_MBUF, SI_ORDER_FIRST, sfstat_init, NULL);
140
141 static int
142 sfstat_sysctl(SYSCTL_HANDLER_ARGS)
143 {
144 struct sfstat s;
145
146 COUNTER_ARRAY_COPY(sfstat, &s, sizeof(s) / sizeof(uint64_t));
147 if (req->newptr)
148 COUNTER_ARRAY_ZERO(sfstat, sizeof(s) / sizeof(uint64_t));
149 return (SYSCTL_OUT(req, &s, sizeof(s)));
150 }
151 SYSCTL_PROC(_kern_ipc, OID_AUTO, sfstat, CTLTYPE_OPAQUE | CTLFLAG_RW,
152 NULL, 0, sfstat_sysctl, "I", "sendfile statistics");
153
154 static void
155 sendfile_free_mext(struct mbuf *m)
156 {
157 struct sf_buf *sf;
158 vm_page_t pg;
159 int flags;
160
161 KASSERT(m->m_flags & M_EXT && m->m_ext.ext_type == EXT_SFBUF,
162 ("%s: m %p !M_EXT or !EXT_SFBUF", __func__, m));
163
164 sf = m->m_ext.ext_arg1;
165 pg = sf_buf_page(sf);
166 flags = (m->m_ext.ext_flags & EXT_FLAG_NOCACHE) != 0 ? VPR_TRYFREE : 0;
167
168 sf_buf_free(sf);
169 vm_page_release(pg, flags);
170
171 if (m->m_ext.ext_flags & EXT_FLAG_SYNC) {
172 struct sendfile_sync *sfs = m->m_ext.ext_arg2;
173 sendfile_sync_signal(sfs);
174 }
175 }
176
177 /*
178 * Helper function to calculate how much data to put into page i of n.
179 * Only first and last pages are special.
180 */
181 static inline off_t
182 xfsize(int i, int n, off_t off, off_t len)
183 {
184
185 if (i == 0)
186 return (omin(PAGE_SIZE - (off & PAGE_MASK), len));
187
188 if (i == n - 1 && ((off + len) & PAGE_MASK) > 0)
189 return ((off + len) & PAGE_MASK);
190
191 return (PAGE_SIZE);
192 }
193
194 /*
195 * Helper function to get offset within object for i page.
196 */
197 static inline vm_ooffset_t
198 vmoff(int i, off_t off)
199 {
200
201 if (i == 0)
202 return ((vm_ooffset_t)off);
203
204 return (trunc_page(off + i * PAGE_SIZE));
205 }
206
207 /*
208 * Helper function used when allocation of a page or sf_buf failed.
209 * Pretend as if we don't have enough space, subtract xfsize() of
210 * all pages that failed.
211 */
212 static inline void
213 fixspace(int old, int new, off_t off, int *space)
214 {
215
216 KASSERT(old > new, ("%s: old %d new %d", __func__, old, new));
217
218 /* Subtract last one. */
219 *space -= xfsize(old - 1, old, off, *space);
220 old--;
221
222 if (new == old)
223 /* There was only one page. */
224 return;
225
226 /* Subtract first one. */
227 if (new == 0) {
228 *space -= xfsize(0, old, off, *space);
229 new++;
230 }
231
232 /* Rest of pages are full sized. */
233 *space -= (old - new) * PAGE_SIZE;
234
235 KASSERT(*space >= 0, ("%s: space went backwards", __func__));
236 }
237
238 /*
239 * I/O completion callback.
240 */
241 static void
242 sendfile_iodone(void *arg, vm_page_t *pg, int count, int error)
243 {
244 struct sf_io *sfio = arg;
245 struct socket *so = sfio->so;
246
247 for (int i = 0; i < count; i++)
248 if (pg[i] != bogus_page)
249 vm_page_xunbusy(pg[i]);
250
251 if (error)
252 sfio->error = error;
253
254 if (!refcount_release(&sfio->nios))
255 return;
256
257 CURVNET_SET(so->so_vnet);
258 if (sfio->error) {
259 struct mbuf *m;
260
261 /*
262 * I/O operation failed. The state of data in the socket
263 * is now inconsistent, and all what we can do is to tear
264 * it down. Protocol abort method would tear down protocol
265 * state, free all ready mbufs and detach not ready ones.
266 * We will free the mbufs corresponding to this I/O manually.
267 *
268 * The socket would be marked with EIO and made available
269 * for read, so that application receives EIO on next
270 * syscall and eventually closes the socket.
271 */
272 so->so_proto->pr_usrreqs->pru_abort(so);
273 so->so_error = EIO;
274
275 m = sfio->m;
276 for (int i = 0; i < sfio->npages; i++)
277 m = m_free(m);
278 } else
279 (void )(so->so_proto->pr_usrreqs->pru_ready)(so, sfio->m,
280 sfio->npages);
281
282 SOCK_LOCK(so);
283 sorele(so);
284 CURVNET_RESTORE();
285 free(sfio, M_SENDFILE);
286 }
287
288 /*
289 * Iterate through pages vector and request paging for non-valid pages.
290 */
291 static int
292 sendfile_swapin(vm_object_t obj, struct sf_io *sfio, int *nios, off_t off,
293 off_t len, int npages, int rhpages, int flags)
294 {
295 vm_page_t *pa = sfio->pa;
296 int grabbed;
297
298 *nios = 0;
299 flags = (flags & SF_NODISKIO) ? VM_ALLOC_NOWAIT : 0;
300
301 /*
302 * First grab all the pages and wire them. Note that we grab
303 * only required pages. Readahead pages are dealt with later.
304 */
305 VM_OBJECT_WLOCK(obj);
306
307 grabbed = vm_page_grab_pages(obj, OFF_TO_IDX(off),
308 VM_ALLOC_NORMAL | VM_ALLOC_WIRED | flags, pa, npages);
309 if (grabbed < npages) {
310 for (int i = grabbed; i < npages; i++)
311 pa[i] = NULL;
312 npages = grabbed;
313 rhpages = 0;
314 }
315
316 for (int i = 0; i < npages;) {
317 int j, a, count, rv;
318
319 /* Skip valid pages. */
320 if (vm_page_is_valid(pa[i], vmoff(i, off) & PAGE_MASK,
321 xfsize(i, npages, off, len))) {
322 vm_page_xunbusy(pa[i]);
323 SFSTAT_INC(sf_pages_valid);
324 i++;
325 continue;
326 }
327
328 /*
329 * Next page is invalid. Check if it belongs to pager. It
330 * may not be there, which is a regular situation for shmem
331 * pager. For vnode pager this happens only in case of
332 * a sparse file.
333 *
334 * Important feature of vm_pager_has_page() is the hint
335 * stored in 'a', about how many pages we can pagein after
336 * this page in a single I/O.
337 */
338 if (!vm_pager_has_page(obj, OFF_TO_IDX(vmoff(i, off)), NULL,
339 &a)) {
340 pmap_zero_page(pa[i]);
341 pa[i]->valid = VM_PAGE_BITS_ALL;
342 MPASS(pa[i]->dirty == 0);
343 vm_page_xunbusy(pa[i]);
344 i++;
345 continue;
346 }
347
348 /*
349 * We want to pagein as many pages as possible, limited only
350 * by the 'a' hint and actual request.
351 */
352 count = min(a + 1, npages - i);
353
354 /*
355 * We should not pagein into a valid page, thus we first trim
356 * any valid pages off the end of request, and substitute
357 * to bogus_page those, that are in the middle.
358 */
359 for (j = i + count - 1; j > i; j--) {
360 if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
361 xfsize(j, npages, off, len))) {
362 count--;
363 rhpages = 0;
364 } else
365 break;
366 }
367 for (j = i + 1; j < i + count - 1; j++)
368 if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
369 xfsize(j, npages, off, len))) {
370 vm_page_xunbusy(pa[j]);
371 SFSTAT_INC(sf_pages_valid);
372 SFSTAT_INC(sf_pages_bogus);
373 pa[j] = bogus_page;
374 }
375
376 refcount_acquire(&sfio->nios);
377 rv = vm_pager_get_pages_async(obj, pa + i, count, NULL,
378 i + count == npages ? &rhpages : NULL,
379 &sendfile_iodone, sfio);
380 if (rv != VM_PAGER_OK) {
381 for (j = i; j < i + count; j++) {
382 if (pa[j] != bogus_page) {
383 vm_page_lock(pa[j]);
384 vm_page_unwire(pa[j], PQ_INACTIVE);
385 vm_page_unlock(pa[j]);
386 }
387 }
388 VM_OBJECT_WUNLOCK(obj);
389 return (EIO);
390 }
391 KASSERT(rv == VM_PAGER_OK, ("%s: pager fail obj %p page %p",
392 __func__, obj, pa[i]));
393
394 SFSTAT_INC(sf_iocnt);
395 SFSTAT_ADD(sf_pages_read, count);
396 if (i + count == npages)
397 SFSTAT_ADD(sf_rhpages_read, rhpages);
398
399 /*
400 * Restore the valid page pointers. They are already
401 * unbusied, but still wired.
402 */
403 for (j = i; j < i + count; j++)
404 if (pa[j] == bogus_page) {
405 pa[j] = vm_page_lookup(obj,
406 OFF_TO_IDX(vmoff(j, off)));
407 KASSERT(pa[j], ("%s: page %p[%d] disappeared",
408 __func__, pa, j));
409
410 }
411 i += count;
412 (*nios)++;
413 }
414
415 VM_OBJECT_WUNLOCK(obj);
416
417 if (*nios == 0 && npages != 0)
418 SFSTAT_INC(sf_noiocnt);
419
420 return (0);
421 }
422
423 static int
424 sendfile_getobj(struct thread *td, struct file *fp, vm_object_t *obj_res,
425 struct vnode **vp_res, struct shmfd **shmfd_res, off_t *obj_size,
426 int *bsize)
427 {
428 struct vattr va;
429 vm_object_t obj;
430 struct vnode *vp;
431 struct shmfd *shmfd;
432 int error;
433
434 vp = *vp_res = NULL;
435 obj = NULL;
436 shmfd = *shmfd_res = NULL;
437 *bsize = 0;
438
439 /*
440 * The file descriptor must be a regular file and have a
441 * backing VM object.
442 */
443 if (fp->f_type == DTYPE_VNODE) {
444 vp = fp->f_vnode;
445 vn_lock(vp, LK_SHARED | LK_RETRY);
446 if (vp->v_type != VREG) {
447 error = EINVAL;
448 goto out;
449 }
450 *bsize = vp->v_mount->mnt_stat.f_iosize;
451 error = VOP_GETATTR(vp, &va, td->td_ucred);
452 if (error != 0)
453 goto out;
454 *obj_size = va.va_size;
455 obj = vp->v_object;
456 if (obj == NULL) {
457 error = EINVAL;
458 goto out;
459 }
460 } else if (fp->f_type == DTYPE_SHM) {
461 error = 0;
462 shmfd = fp->f_data;
463 obj = shmfd->shm_object;
464 *obj_size = shmfd->shm_size;
465 } else {
466 error = EINVAL;
467 goto out;
468 }
469
470 VM_OBJECT_WLOCK(obj);
471 if ((obj->flags & OBJ_DEAD) != 0) {
472 VM_OBJECT_WUNLOCK(obj);
473 error = EBADF;
474 goto out;
475 }
476
477 /*
478 * Temporarily increase the backing VM object's reference
479 * count so that a forced reclamation of its vnode does not
480 * immediately destroy it.
481 */
482 vm_object_reference_locked(obj);
483 VM_OBJECT_WUNLOCK(obj);
484 *obj_res = obj;
485 *vp_res = vp;
486 *shmfd_res = shmfd;
487
488 out:
489 if (vp != NULL)
490 VOP_UNLOCK(vp, 0);
491 return (error);
492 }
493
494 static int
495 sendfile_getsock(struct thread *td, int s, struct file **sock_fp,
496 struct socket **so)
497 {
498 int error;
499
500 *sock_fp = NULL;
501 *so = NULL;
502
503 /*
504 * The socket must be a stream socket and connected.
505 */
506 error = getsock_cap(td, s, &cap_send_rights,
507 sock_fp, NULL, NULL);
508 if (error != 0)
509 return (error);
510 *so = (*sock_fp)->f_data;
511 if ((*so)->so_type != SOCK_STREAM)
512 return (EINVAL);
513 /*
514 * SCTP one-to-one style sockets currently don't work with
515 * sendfile(). So indicate EINVAL for now.
516 */
517 if ((*so)->so_proto->pr_protocol == IPPROTO_SCTP)
518 return (EINVAL);
519 if (SOLISTENING(*so))
520 return (ENOTCONN);
521 return (0);
522 }
523
524 int
525 vn_sendfile(struct file *fp, int sockfd, struct uio *hdr_uio,
526 struct uio *trl_uio, off_t offset, size_t nbytes, off_t *sent, int flags,
527 struct thread *td)
528 {
529 struct file *sock_fp;
530 struct vnode *vp;
531 struct vm_object *obj;
532 struct socket *so;
533 struct mbuf *m, *mh, *mhtail;
534 struct sf_buf *sf;
535 struct shmfd *shmfd;
536 struct sendfile_sync *sfs;
537 struct vattr va;
538 off_t off, sbytes, rem, obj_size;
539 int error, softerr, bsize, hdrlen;
540
541 obj = NULL;
542 so = NULL;
543 m = mh = NULL;
544 sfs = NULL;
545 hdrlen = sbytes = 0;
546 softerr = 0;
547
548 error = sendfile_getobj(td, fp, &obj, &vp, &shmfd, &obj_size, &bsize);
549 if (error != 0)
550 return (error);
551
552 error = sendfile_getsock(td, sockfd, &sock_fp, &so);
553 if (error != 0)
554 goto out;
555
556 #ifdef MAC
557 error = mac_socket_check_send(td->td_ucred, so);
558 if (error != 0)
559 goto out;
560 #endif
561
562 SFSTAT_INC(sf_syscalls);
563 SFSTAT_ADD(sf_rhpages_requested, SF_READAHEAD(flags));
564
565 if (flags & SF_SYNC) {
566 sfs = malloc(sizeof(*sfs), M_SENDFILE, M_WAITOK | M_ZERO);
567 mtx_init(&sfs->mtx, "sendfile", NULL, MTX_DEF);
568 cv_init(&sfs->cv, "sendfile");
569 sfs->waiting = true;
570 }
571
572 rem = nbytes ? omin(nbytes, obj_size - offset) : obj_size - offset;
573
574 /*
575 * Protect against multiple writers to the socket.
576 *
577 * XXXRW: Historically this has assumed non-interruptibility, so now
578 * we implement that, but possibly shouldn't.
579 */
580 (void)sblock(&so->so_snd, SBL_WAIT | SBL_NOINTR);
581
582 /*
583 * Loop through the pages of the file, starting with the requested
584 * offset. Get a file page (do I/O if necessary), map the file page
585 * into an sf_buf, attach an mbuf header to the sf_buf, and queue
586 * it on the socket.
587 * This is done in two loops. The inner loop turns as many pages
588 * as it can, up to available socket buffer space, without blocking
589 * into mbufs to have it bulk delivered into the socket send buffer.
590 * The outer loop checks the state and available space of the socket
591 * and takes care of the overall progress.
592 */
593 for (off = offset; rem > 0; ) {
594 struct sf_io *sfio;
595 vm_page_t *pa;
596 struct mbuf *mtail;
597 int nios, space, npages, rhpages;
598
599 mtail = NULL;
600 /*
601 * Check the socket state for ongoing connection,
602 * no errors and space in socket buffer.
603 * If space is low allow for the remainder of the
604 * file to be processed if it fits the socket buffer.
605 * Otherwise block in waiting for sufficient space
606 * to proceed, or if the socket is nonblocking, return
607 * to userland with EAGAIN while reporting how far
608 * we've come.
609 * We wait until the socket buffer has significant free
610 * space to do bulk sends. This makes good use of file
611 * system read ahead and allows packet segmentation
612 * offloading hardware to take over lots of work. If
613 * we were not careful here we would send off only one
614 * sfbuf at a time.
615 */
616 SOCKBUF_LOCK(&so->so_snd);
617 if (so->so_snd.sb_lowat < so->so_snd.sb_hiwat / 2)
618 so->so_snd.sb_lowat = so->so_snd.sb_hiwat / 2;
619 retry_space:
620 if (so->so_snd.sb_state & SBS_CANTSENDMORE) {
621 error = EPIPE;
622 SOCKBUF_UNLOCK(&so->so_snd);
623 goto done;
624 } else if (so->so_error) {
625 error = so->so_error;
626 so->so_error = 0;
627 SOCKBUF_UNLOCK(&so->so_snd);
628 goto done;
629 }
630 if ((so->so_state & SS_ISCONNECTED) == 0) {
631 SOCKBUF_UNLOCK(&so->so_snd);
632 error = ENOTCONN;
633 goto done;
634 }
635
636 space = sbspace(&so->so_snd);
637 if (space < rem &&
638 (space <= 0 ||
639 space < so->so_snd.sb_lowat)) {
640 if (so->so_state & SS_NBIO) {
641 SOCKBUF_UNLOCK(&so->so_snd);
642 error = EAGAIN;
643 goto done;
644 }
645 /*
646 * sbwait drops the lock while sleeping.
647 * When we loop back to retry_space the
648 * state may have changed and we retest
649 * for it.
650 */
651 error = sbwait(&so->so_snd);
652 /*
653 * An error from sbwait usually indicates that we've
654 * been interrupted by a signal. If we've sent anything
655 * then return bytes sent, otherwise return the error.
656 */
657 if (error != 0) {
658 SOCKBUF_UNLOCK(&so->so_snd);
659 goto done;
660 }
661 goto retry_space;
662 }
663 SOCKBUF_UNLOCK(&so->so_snd);
664
665 /*
666 * At the beginning of the first loop check if any headers
667 * are specified and copy them into mbufs. Reduce space in
668 * the socket buffer by the size of the header mbuf chain.
669 * Clear hdr_uio here and hdrlen at the end of the first loop.
670 */
671 if (hdr_uio != NULL && hdr_uio->uio_resid > 0) {
672 hdr_uio->uio_td = td;
673 hdr_uio->uio_rw = UIO_WRITE;
674 mh = m_uiotombuf(hdr_uio, M_WAITOK, space, 0, 0);
675 hdrlen = m_length(mh, &mhtail);
676 space -= hdrlen;
677 /*
678 * If header consumed all the socket buffer space,
679 * don't waste CPU cycles and jump to the end.
680 */
681 if (space == 0) {
682 sfio = NULL;
683 nios = 0;
684 goto prepend_header;
685 }
686 hdr_uio = NULL;
687 }
688
689 if (vp != NULL) {
690 error = vn_lock(vp, LK_SHARED);
691 if (error != 0)
692 goto done;
693 error = VOP_GETATTR(vp, &va, td->td_ucred);
694 if (error != 0 || off >= va.va_size) {
695 VOP_UNLOCK(vp, 0);
696 goto done;
697 }
698 if (va.va_size != obj_size) {
699 obj_size = va.va_size;
700 rem = nbytes ?
701 omin(nbytes + offset, obj_size) : obj_size;
702 rem -= off;
703 }
704 }
705
706 if (space > rem)
707 space = rem;
708
709 npages = howmany(space + (off & PAGE_MASK), PAGE_SIZE);
710
711 /*
712 * Calculate maximum allowed number of pages for readahead
713 * at this iteration. If SF_USER_READAHEAD was set, we don't
714 * do any heuristics and use exactly the value supplied by
715 * application. Otherwise, we allow readahead up to "rem".
716 * If application wants more, let it be, but there is no
717 * reason to go above MAXPHYS. Also check against "obj_size",
718 * since vm_pager_has_page() can hint beyond EOF.
719 */
720 if (flags & SF_USER_READAHEAD) {
721 rhpages = SF_READAHEAD(flags);
722 } else {
723 rhpages = howmany(rem + (off & PAGE_MASK), PAGE_SIZE) -
724 npages;
725 rhpages += SF_READAHEAD(flags);
726 }
727 rhpages = min(howmany(MAXPHYS, PAGE_SIZE), rhpages);
728 rhpages = min(howmany(obj_size - trunc_page(off), PAGE_SIZE) -
729 npages, rhpages);
730
731 sfio = malloc(sizeof(struct sf_io) +
732 npages * sizeof(vm_page_t), M_SENDFILE, M_WAITOK);
733 refcount_init(&sfio->nios, 1);
734 sfio->so = so;
735 sfio->error = 0;
736
737 error = sendfile_swapin(obj, sfio, &nios, off, space, npages,
738 rhpages, flags);
739 if (error != 0) {
740 if (vp != NULL)
741 VOP_UNLOCK(vp, 0);
742 free(sfio, M_SENDFILE);
743 goto done;
744 }
745
746 /*
747 * Loop and construct maximum sized mbuf chain to be bulk
748 * dumped into socket buffer.
749 */
750 pa = sfio->pa;
751 for (int i = 0; i < npages; i++) {
752 struct mbuf *m0;
753
754 /*
755 * If a page wasn't grabbed successfully, then
756 * trim the array. Can happen only with SF_NODISKIO.
757 */
758 if (pa[i] == NULL) {
759 SFSTAT_INC(sf_busy);
760 fixspace(npages, i, off, &space);
761 npages = i;
762 softerr = EBUSY;
763 break;
764 }
765
766 /*
767 * Get a sendfile buf. When allocating the
768 * first buffer for mbuf chain, we usually
769 * wait as long as necessary, but this wait
770 * can be interrupted. For consequent
771 * buffers, do not sleep, since several
772 * threads might exhaust the buffers and then
773 * deadlock.
774 */
775 sf = sf_buf_alloc(pa[i],
776 m != NULL ? SFB_NOWAIT : SFB_CATCH);
777 if (sf == NULL) {
778 SFSTAT_INC(sf_allocfail);
779 for (int j = i; j < npages; j++) {
780 vm_page_lock(pa[j]);
781 vm_page_unwire(pa[j], PQ_INACTIVE);
782 vm_page_unlock(pa[j]);
783 }
784 if (m == NULL)
785 softerr = ENOBUFS;
786 fixspace(npages, i, off, &space);
787 npages = i;
788 break;
789 }
790
791 m0 = m_get(M_WAITOK, MT_DATA);
792 m0->m_ext.ext_buf = (char *)sf_buf_kva(sf);
793 m0->m_ext.ext_size = PAGE_SIZE;
794 m0->m_ext.ext_arg1 = sf;
795 m0->m_ext.ext_type = EXT_SFBUF;
796 m0->m_ext.ext_flags = EXT_FLAG_EMBREF;
797 m0->m_ext.ext_free = sendfile_free_mext;
798 /*
799 * SF_NOCACHE sets the page as being freed upon send.
800 * However, we ignore it for the last page in 'space',
801 * if the page is truncated, and we got more data to
802 * send (rem > space), or if we have readahead
803 * configured (rhpages > 0).
804 */
805 if ((flags & SF_NOCACHE) &&
806 (i != npages - 1 ||
807 !((off + space) & PAGE_MASK) ||
808 !(rem > space || rhpages > 0)))
809 m0->m_ext.ext_flags |= EXT_FLAG_NOCACHE;
810 if (sfs != NULL) {
811 m0->m_ext.ext_flags |= EXT_FLAG_SYNC;
812 m0->m_ext.ext_arg2 = sfs;
813 mtx_lock(&sfs->mtx);
814 sfs->count++;
815 mtx_unlock(&sfs->mtx);
816 }
817 m0->m_ext.ext_count = 1;
818 m0->m_flags |= (M_EXT | M_RDONLY);
819 if (nios)
820 m0->m_flags |= M_NOTREADY;
821 m0->m_data = (char *)sf_buf_kva(sf) +
822 (vmoff(i, off) & PAGE_MASK);
823 m0->m_len = xfsize(i, npages, off, space);
824
825 if (i == 0)
826 sfio->m = m0;
827
828 /* Append to mbuf chain. */
829 if (mtail != NULL)
830 mtail->m_next = m0;
831 else
832 m = m0;
833 mtail = m0;
834 }
835
836 if (vp != NULL)
837 VOP_UNLOCK(vp, 0);
838
839 /* Keep track of bytes processed. */
840 off += space;
841 rem -= space;
842
843 /* Prepend header, if any. */
844 if (hdrlen) {
845 prepend_header:
846 mhtail->m_next = m;
847 m = mh;
848 mh = NULL;
849 }
850
851 if (m == NULL) {
852 KASSERT(softerr, ("%s: m NULL, no error", __func__));
853 error = softerr;
854 free(sfio, M_SENDFILE);
855 goto done;
856 }
857
858 /* Add the buffer chain to the socket buffer. */
859 KASSERT(m_length(m, NULL) == space + hdrlen,
860 ("%s: mlen %u space %d hdrlen %d",
861 __func__, m_length(m, NULL), space, hdrlen));
862
863 CURVNET_SET(so->so_vnet);
864 if (nios == 0) {
865 /*
866 * If sendfile_swapin() didn't initiate any I/Os,
867 * which happens if all data is cached in VM, then
868 * we can send data right now without the
869 * PRUS_NOTREADY flag.
870 */
871 free(sfio, M_SENDFILE);
872 error = (*so->so_proto->pr_usrreqs->pru_send)
873 (so, 0, m, NULL, NULL, td);
874 } else {
875 sfio->npages = npages;
876 soref(so);
877 error = (*so->so_proto->pr_usrreqs->pru_send)
878 (so, PRUS_NOTREADY, m, NULL, NULL, td);
879 sendfile_iodone(sfio, NULL, 0, 0);
880 }
881 CURVNET_RESTORE();
882
883 m = NULL; /* pru_send always consumes */
884 if (error)
885 goto done;
886 sbytes += space + hdrlen;
887 if (hdrlen)
888 hdrlen = 0;
889 if (softerr) {
890 error = softerr;
891 goto done;
892 }
893 }
894
895 /*
896 * Send trailers. Wimp out and use writev(2).
897 */
898 if (trl_uio != NULL) {
899 sbunlock(&so->so_snd);
900 error = kern_writev(td, sockfd, trl_uio);
901 if (error == 0)
902 sbytes += td->td_retval[0];
903 goto out;
904 }
905
906 done:
907 sbunlock(&so->so_snd);
908 out:
909 /*
910 * If there was no error we have to clear td->td_retval[0]
911 * because it may have been set by writev.
912 */
913 if (error == 0) {
914 td->td_retval[0] = 0;
915 }
916 if (sent != NULL) {
917 (*sent) = sbytes;
918 }
919 if (obj != NULL)
920 vm_object_deallocate(obj);
921 if (so)
922 fdrop(sock_fp, td);
923 if (m)
924 m_freem(m);
925 if (mh)
926 m_freem(mh);
927
928 if (sfs != NULL) {
929 mtx_lock(&sfs->mtx);
930 if (sfs->count != 0)
931 error = cv_wait_sig(&sfs->cv, &sfs->mtx);
932 if (sfs->count == 0) {
933 sendfile_sync_destroy(sfs);
934 } else {
935 sfs->waiting = false;
936 mtx_unlock(&sfs->mtx);
937 }
938 }
939
940 if (error == ERESTART)
941 error = EINTR;
942
943 return (error);
944 }
945
946 static int
947 sendfile(struct thread *td, struct sendfile_args *uap, int compat)
948 {
949 struct sf_hdtr hdtr;
950 struct uio *hdr_uio, *trl_uio;
951 struct file *fp;
952 off_t sbytes;
953 int error;
954
955 /*
956 * File offset must be positive. If it goes beyond EOF
957 * we send only the header/trailer and no payload data.
958 */
959 if (uap->offset < 0)
960 return (EINVAL);
961
962 sbytes = 0;
963 hdr_uio = trl_uio = NULL;
964
965 if (uap->hdtr != NULL) {
966 error = copyin(uap->hdtr, &hdtr, sizeof(hdtr));
967 if (error != 0)
968 goto out;
969 if (hdtr.headers != NULL) {
970 error = copyinuio(hdtr.headers, hdtr.hdr_cnt,
971 &hdr_uio);
972 if (error != 0)
973 goto out;
974 #ifdef COMPAT_FREEBSD4
975 /*
976 * In FreeBSD < 5.0 the nbytes to send also included
977 * the header. If compat is specified subtract the
978 * header size from nbytes.
979 */
980 if (compat) {
981 if (uap->nbytes > hdr_uio->uio_resid)
982 uap->nbytes -= hdr_uio->uio_resid;
983 else
984 uap->nbytes = 0;
985 }
986 #endif
987 }
988 if (hdtr.trailers != NULL) {
989 error = copyinuio(hdtr.trailers, hdtr.trl_cnt,
990 &trl_uio);
991 if (error != 0)
992 goto out;
993 }
994 }
995
996 AUDIT_ARG_FD(uap->fd);
997
998 /*
999 * sendfile(2) can start at any offset within a file so we require
1000 * CAP_READ+CAP_SEEK = CAP_PREAD.
1001 */
1002 if ((error = fget_read(td, uap->fd, &cap_pread_rights, &fp)) != 0)
1003 goto out;
1004
1005 error = fo_sendfile(fp, uap->s, hdr_uio, trl_uio, uap->offset,
1006 uap->nbytes, &sbytes, uap->flags, td);
1007 fdrop(fp, td);
1008
1009 if (uap->sbytes != NULL)
1010 copyout(&sbytes, uap->sbytes, sizeof(off_t));
1011
1012 out:
1013 free(hdr_uio, M_IOV);
1014 free(trl_uio, M_IOV);
1015 return (error);
1016 }
1017
1018 /*
1019 * sendfile(2)
1020 *
1021 * int sendfile(int fd, int s, off_t offset, size_t nbytes,
1022 * struct sf_hdtr *hdtr, off_t *sbytes, int flags)
1023 *
1024 * Send a file specified by 'fd' and starting at 'offset' to a socket
1025 * specified by 's'. Send only 'nbytes' of the file or until EOF if nbytes ==
1026 * 0. Optionally add a header and/or trailer to the socket output. If
1027 * specified, write the total number of bytes sent into *sbytes.
1028 */
1029 int
1030 sys_sendfile(struct thread *td, struct sendfile_args *uap)
1031 {
1032
1033 return (sendfile(td, uap, 0));
1034 }
1035
1036 #ifdef COMPAT_FREEBSD4
1037 int
1038 freebsd4_sendfile(struct thread *td, struct freebsd4_sendfile_args *uap)
1039 {
1040 struct sendfile_args args;
1041
1042 args.fd = uap->fd;
1043 args.s = uap->s;
1044 args.offset = uap->offset;
1045 args.nbytes = uap->nbytes;
1046 args.hdtr = uap->hdtr;
1047 args.sbytes = uap->sbytes;
1048 args.flags = uap->flags;
1049
1050 return (sendfile(td, &args, 1));
1051 }
1052 #endif /* COMPAT_FREEBSD4 */
Cache object: a60b869e811916d9ff77daed6546d054
|