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 "opt_kern_tls.h"
31
32 #include <sys/param.h>
33 #include <sys/systm.h>
34 #include <sys/capsicum.h>
35 #include <sys/inotify.h>
36 #include <sys/kernel.h>
37 #include <sys/lock.h>
38 #include <sys/ktls.h>
39 #include <sys/mutex.h>
40 #include <sys/malloc.h>
41 #include <sys/mman.h>
42 #include <sys/mount.h>
43 #include <sys/mbuf.h>
44 #include <sys/proc.h>
45 #include <sys/protosw.h>
46 #include <sys/rwlock.h>
47 #include <sys/sf_buf.h>
48 #include <sys/socket.h>
49 #include <sys/socketvar.h>
50 #include <sys/syscallsubr.h>
51 #include <sys/sysctl.h>
52 #include <sys/sysproto.h>
53 #include <sys/vnode.h>
54
55 #include <net/vnet.h>
56 #include <netinet/in.h>
57 #include <netinet/tcp.h>
58 #include <netinet/in_pcb.h>
59 #include <netinet/tcp_var.h>
60 #include <netinet/tcp_log_buf.h>
61
62 #include <security/audit/audit.h>
63 #include <security/mac/mac_framework.h>
64
65 #include <vm/vm.h>
66 #include <vm/vm_object.h>
67 #include <vm/vm_pager.h>
68
69 static MALLOC_DEFINE(M_SENDFILE, "sendfile", "sendfile dynamic memory");
70
71 #define EXT_FLAG_SYNC EXT_FLAG_VENDOR1
72 #define EXT_FLAG_NOCACHE EXT_FLAG_VENDOR2
73 #define EXT_FLAG_CACHE_LAST EXT_FLAG_VENDOR3
74
75 /*
76 * Structure describing a single sendfile(2) I/O, which may consist of
77 * several underlying pager I/Os.
78 *
79 * The syscall context allocates the structure and initializes 'nios'
80 * to 1. As sendfile_swapin() runs through pages and starts asynchronous
81 * paging operations, it increments 'nios'.
82 *
83 * Every I/O completion calls sendfile_iodone(), which decrements the 'nios',
84 * and the syscall also calls sendfile_iodone() after allocating all mbufs,
85 * linking them and sending to socket. Whoever reaches zero 'nios' is
86 * responsible to call pr_ready() on the socket, to notify it of readyness
87 * of the data.
88 */
89 struct sf_io {
90 volatile u_int nios;
91 u_int error;
92 int npages;
93 struct socket *so;
94 struct mbuf *m;
95 vm_object_t obj;
96 vm_pindex_t pindex0;
97 #ifdef KERN_TLS
98 struct ktls_session *tls;
99 #endif
100 vm_page_t pa[];
101 };
102
103 /*
104 * Structure used to track requests with SF_SYNC flag.
105 */
106 struct sendfile_sync {
107 struct mtx mtx;
108 struct cv cv;
109 unsigned count;
110 bool waiting;
111 };
112
113 static void
sendfile_sync_destroy(struct sendfile_sync * sfs)114 sendfile_sync_destroy(struct sendfile_sync *sfs)
115 {
116 KASSERT(sfs->count == 0, ("sendfile sync %p still busy", sfs));
117
118 cv_destroy(&sfs->cv);
119 mtx_destroy(&sfs->mtx);
120 free(sfs, M_SENDFILE);
121 }
122
123 static void
sendfile_sync_signal(struct sendfile_sync * sfs)124 sendfile_sync_signal(struct sendfile_sync *sfs)
125 {
126 mtx_lock(&sfs->mtx);
127 KASSERT(sfs->count > 0, ("sendfile sync %p not busy", sfs));
128 if (--sfs->count == 0) {
129 if (!sfs->waiting) {
130 /* The sendfile() waiter was interrupted by a signal. */
131 sendfile_sync_destroy(sfs);
132 return;
133 } else {
134 cv_signal(&sfs->cv);
135 }
136 }
137 mtx_unlock(&sfs->mtx);
138 }
139
140 counter_u64_t sfstat[sizeof(struct sfstat) / sizeof(uint64_t)];
141
142 static void
sfstat_init(const void * unused)143 sfstat_init(const void *unused)
144 {
145
146 COUNTER_ARRAY_ALLOC(sfstat, sizeof(struct sfstat) / sizeof(uint64_t),
147 M_WAITOK);
148 }
149 SYSINIT(sfstat, SI_SUB_MBUF, SI_ORDER_FIRST, sfstat_init, NULL);
150
151 static int
sfstat_sysctl(SYSCTL_HANDLER_ARGS)152 sfstat_sysctl(SYSCTL_HANDLER_ARGS)
153 {
154 struct sfstat s;
155
156 COUNTER_ARRAY_COPY(sfstat, &s, sizeof(s) / sizeof(uint64_t));
157 if (req->newptr)
158 COUNTER_ARRAY_ZERO(sfstat, sizeof(s) / sizeof(uint64_t));
159 return (SYSCTL_OUT(req, &s, sizeof(s)));
160 }
161 SYSCTL_PROC(_kern_ipc, OID_AUTO, sfstat,
162 CTLTYPE_OPAQUE | CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 0,
163 sfstat_sysctl, "I",
164 "sendfile statistics");
165
166 static void
sendfile_free_mext(struct mbuf * m)167 sendfile_free_mext(struct mbuf *m)
168 {
169 struct sf_buf *sf;
170 vm_page_t pg;
171 int flags;
172
173 KASSERT(m->m_flags & M_EXT && m->m_ext.ext_type == EXT_SFBUF,
174 ("%s: m %p !M_EXT or !EXT_SFBUF", __func__, m));
175
176 sf = m->m_ext.ext_arg1;
177 pg = sf_buf_page(sf);
178 flags = (m->m_ext.ext_flags & EXT_FLAG_NOCACHE) != 0 ? VPR_TRYFREE : 0;
179
180 sf_buf_free(sf);
181 vm_page_release(pg, flags);
182
183 if (m->m_ext.ext_flags & EXT_FLAG_SYNC) {
184 struct sendfile_sync *sfs = m->m_ext.ext_arg2;
185 sendfile_sync_signal(sfs);
186 }
187 }
188
189 static void
sendfile_free_mext_pg(struct mbuf * m)190 sendfile_free_mext_pg(struct mbuf *m)
191 {
192 vm_page_t pg;
193 int flags, i;
194 bool cache_last;
195
196 M_ASSERTEXTPG(m);
197
198 cache_last = m->m_ext.ext_flags & EXT_FLAG_CACHE_LAST;
199 flags = (m->m_ext.ext_flags & EXT_FLAG_NOCACHE) != 0 ? VPR_TRYFREE : 0;
200
201 for (i = 0; i < m->m_epg_npgs; i++) {
202 if (cache_last && i == m->m_epg_npgs - 1)
203 flags = 0;
204 pg = PHYS_TO_VM_PAGE(m->m_epg_pa[i]);
205 vm_page_release(pg, flags);
206 }
207
208 if (m->m_ext.ext_flags & EXT_FLAG_SYNC) {
209 struct sendfile_sync *sfs = m->m_ext.ext_arg1;
210 sendfile_sync_signal(sfs);
211 }
212 }
213
214 /*
215 * Helper function to calculate how much data to put into page i of n.
216 * Only first and last pages are special.
217 */
218 static inline off_t
xfsize(int i,int n,off_t off,off_t len)219 xfsize(int i, int n, off_t off, off_t len)
220 {
221
222 if (i == 0)
223 return (omin(PAGE_SIZE - (off & PAGE_MASK), len));
224
225 if (i == n - 1 && ((off + len) & PAGE_MASK) > 0)
226 return ((off + len) & PAGE_MASK);
227
228 return (PAGE_SIZE);
229 }
230
231 /*
232 * Helper function to get offset within object for i page.
233 */
234 static inline vm_ooffset_t
vmoff(int i,off_t off)235 vmoff(int i, off_t off)
236 {
237
238 if (i == 0)
239 return ((vm_ooffset_t)off);
240
241 return (trunc_page(off + i * PAGE_SIZE));
242 }
243
244 /*
245 * Helper function used when allocation of a page or sf_buf failed.
246 * Pretend as if we don't have enough space, subtract xfsize() of
247 * all pages that failed.
248 */
249 static inline void
fixspace(int old,int new,off_t off,int * space)250 fixspace(int old, int new, off_t off, int *space)
251 {
252
253 KASSERT(old > new, ("%s: old %d new %d", __func__, old, new));
254
255 /* Subtract last one. */
256 *space -= xfsize(old - 1, old, off, *space);
257 old--;
258
259 if (new == old)
260 /* There was only one page. */
261 return;
262
263 /* Subtract first one. */
264 if (new == 0) {
265 *space -= xfsize(0, old, off, *space);
266 new++;
267 }
268
269 /* Rest of pages are full sized. */
270 *space -= (old - new) * PAGE_SIZE;
271
272 KASSERT(*space >= 0, ("%s: space went backwards", __func__));
273 }
274
275 /*
276 * Wait for all in-flight ios to complete, we must not unwire pages
277 * under them.
278 */
279 static void
sendfile_iowait(struct sf_io * sfio,const char * wmesg)280 sendfile_iowait(struct sf_io *sfio, const char *wmesg)
281 {
282 while (atomic_load_int(&sfio->nios) != 1)
283 pause(wmesg, 1);
284 }
285
286 /*
287 * I/O completion callback.
288 *
289 * When called via I/O path, the curvnet is not set and should be obtained
290 * from the socket. When called synchronously from vn_sendfile(), usually
291 * to report error or just release the reference (all pages are valid), then
292 * curvnet shall be already set.
293 */
294 static void
sendfile_iodone(void * arg,vm_page_t * pa,int count,int error)295 sendfile_iodone(void *arg, vm_page_t *pa, int count, int error)
296 {
297 struct sf_io *sfio = arg;
298 struct socket *so;
299 int i;
300
301 if (error != 0)
302 sfio->error = error;
303
304 /*
305 * Restore the valid page pointers. They are already
306 * unbusied, but still wired.
307 *
308 * XXXKIB since pages are only wired, and we do not
309 * own the object lock, other users might have
310 * invalidated them in meantime. Similarly, after we
311 * unbusied the swapped-in pages, they can become
312 * invalid under us.
313 */
314 MPASS(count == 0 || pa[0] != bogus_page);
315 for (i = 0; i < count; i++) {
316 if (pa[i] == bogus_page) {
317 sfio->pa[(pa[0]->pindex - sfio->pindex0) + i] =
318 pa[i] = vm_page_relookup(sfio->obj,
319 pa[0]->pindex + i);
320 KASSERT(pa[i] != NULL,
321 ("%s: page %p[%d] disappeared",
322 __func__, pa, i));
323 } else {
324 vm_page_xunbusy_unchecked(pa[i]);
325 }
326 }
327
328 if (!refcount_release(&sfio->nios))
329 return;
330
331 #ifdef INVARIANTS
332 for (i = 1; i < sfio->npages; i++) {
333 if (sfio->pa[i] == NULL)
334 break;
335 KASSERT(vm_page_wired(sfio->pa[i]),
336 ("sfio %p page %d %p not wired", sfio, i, sfio->pa[i]));
337 if (i == 0)
338 continue;
339 KASSERT(sfio->pa[0]->object == sfio->pa[i]->object,
340 ("sfio %p page %d %p wrong owner %p %p", sfio, i,
341 sfio->pa[i], sfio->pa[0]->object, sfio->pa[i]->object));
342 KASSERT(sfio->pa[0]->pindex + i == sfio->pa[i]->pindex,
343 ("sfio %p page %d %p wrong index %jx %jx", sfio, i,
344 sfio->pa[i], (uintmax_t)sfio->pa[0]->pindex,
345 (uintmax_t)sfio->pa[i]->pindex));
346 }
347 #endif
348
349 vm_object_pip_wakeup(sfio->obj);
350
351 if (sfio->m == NULL) {
352 /*
353 * Either I/O operation failed, or we failed to allocate
354 * buffers, or we bailed out on first busy page, or we
355 * succeeded filling the request without any I/Os. Anyway,
356 * pr_send() hadn't been executed - nothing had been sent
357 * to the socket yet.
358 */
359 MPASS((curthread->td_pflags & TDP_KTHREAD) == 0);
360 free(sfio, M_SENDFILE);
361 return;
362 }
363
364 #if defined(KERN_TLS) && defined(INVARIANTS)
365 if ((sfio->m->m_flags & M_EXTPG) != 0)
366 KASSERT(sfio->tls == sfio->m->m_epg_tls,
367 ("TLS session mismatch"));
368 else
369 KASSERT(sfio->tls == NULL,
370 ("non-ext_pgs mbuf with TLS session"));
371 #endif
372 so = sfio->so;
373 CURVNET_SET_QUIET(so->so_vnet);
374 if (__predict_false(sfio->error)) {
375 /*
376 * I/O operation failed. The state of data in the socket
377 * is now inconsistent, and all what we can do is to tear
378 * it down. Protocol abort method would tear down protocol
379 * state, free all ready mbufs and detach not ready ones.
380 * We will free the mbufs corresponding to this I/O manually.
381 *
382 * The socket would be marked with EIO and made available
383 * for read, so that application receives EIO on next
384 * syscall and eventually closes the socket.
385 */
386 so->so_proto->pr_abort(so);
387 so->so_error = EIO;
388
389 mb_free_notready(sfio->m, sfio->npages);
390 #ifdef KERN_TLS
391 } else if (sfio->tls != NULL && sfio->tls->mode == TCP_TLS_MODE_SW) {
392 /*
393 * I/O operation is complete, but we still need to
394 * encrypt. We cannot do this in the interrupt thread
395 * of the disk controller, so forward the mbufs to a
396 * different thread.
397 *
398 * Donate the socket reference from sfio to rather
399 * than explicitly invoking soref().
400 */
401 ktls_enqueue(sfio->m, so, sfio->npages);
402 goto out_with_ref;
403 #endif
404 } else
405 (void)so->so_proto->pr_ready(so, sfio->m, sfio->npages);
406
407 sorele(so);
408 #ifdef KERN_TLS
409 out_with_ref:
410 #endif
411 CURVNET_RESTORE();
412 free(sfio, M_SENDFILE);
413 }
414
415 /*
416 * Iterate through pages vector and request paging for non-valid pages.
417 */
418 static int
sendfile_swapin(vm_object_t obj,struct sf_io * sfio,int * nios,off_t off,off_t len,int rhpages,int flags)419 sendfile_swapin(vm_object_t obj, struct sf_io *sfio, int *nios, off_t off,
420 off_t len, int rhpages, int flags)
421 {
422 vm_page_t *pa;
423 int a, count, count1, grabbed, i, j, npages, rv;
424
425 pa = sfio->pa;
426 npages = sfio->npages;
427 *nios = 0;
428 flags = (flags & SF_NODISKIO) ? VM_ALLOC_NOWAIT : 0;
429 sfio->pindex0 = OFF_TO_IDX(off);
430
431 /*
432 * First grab all the pages and wire them. Note that we grab
433 * only required pages. Readahead pages are dealt with later.
434 */
435 grabbed = vm_page_grab_pages_unlocked(obj, OFF_TO_IDX(off),
436 VM_ALLOC_NORMAL | VM_ALLOC_WIRED | flags, pa, npages);
437 if (grabbed < npages) {
438 for (int i = grabbed; i < npages; i++)
439 pa[i] = NULL;
440 npages = grabbed;
441 rhpages = 0;
442 }
443
444 for (i = 0; i < npages;) {
445 /* Skip valid pages. */
446 if (vm_page_is_valid(pa[i], vmoff(i, off) & PAGE_MASK,
447 xfsize(i, npages, off, len))) {
448 vm_page_xunbusy(pa[i]);
449 SFSTAT_INC(sf_pages_valid);
450 i++;
451 continue;
452 }
453
454 /*
455 * Next page is invalid. Check if it belongs to pager. It
456 * may not be there, which is a regular situation for shmem
457 * pager. For vnode pager this happens only in case of
458 * a sparse file.
459 *
460 * Important feature of vm_pager_has_page() is the hint
461 * stored in 'a', about how many pages we can pagein after
462 * this page in a single I/O.
463 */
464 VM_OBJECT_RLOCK(obj);
465 if (!vm_pager_has_page(obj, OFF_TO_IDX(vmoff(i, off)), NULL,
466 &a)) {
467 VM_OBJECT_RUNLOCK(obj);
468 pmap_zero_page(pa[i]);
469 vm_page_valid(pa[i]);
470 MPASS(pa[i]->dirty == 0);
471 vm_page_xunbusy(pa[i]);
472 i++;
473 continue;
474 }
475 VM_OBJECT_RUNLOCK(obj);
476
477 /*
478 * We want to pagein as many pages as possible, limited only
479 * by the 'a' hint and actual request.
480 */
481 count = min(a + 1, npages - i);
482
483 /*
484 * We should not pagein into a valid page because
485 * there might be still unfinished write tracked by
486 * e.g. a buffer, thus we substitute any valid pages
487 * with the bogus one.
488 *
489 * We must not leave around xbusy pages which are not
490 * part of the run passed to vm_pager_getpages(),
491 * otherwise pager might deadlock waiting for the busy
492 * status of the page, e.g. if it constitues the
493 * buffer needed to validate other page.
494 *
495 * First trim the end of the run consisting of the
496 * valid pages, then replace the rest of the valid
497 * with bogus.
498 */
499 count1 = count;
500 for (j = i + count - 1; j > i; j--) {
501 if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
502 xfsize(j, npages, off, len))) {
503 vm_page_xunbusy(pa[j]);
504 SFSTAT_INC(sf_pages_valid);
505 count--;
506 } else {
507 break;
508 }
509 }
510
511 /*
512 * The last page in the run pa[i + count - 1] is
513 * guaranteed to be invalid by the trim above, so it
514 * is not replaced with bogus, thus -1 in the loop end
515 * condition.
516 */
517 MPASS(pa[i + count - 1]->valid != VM_PAGE_BITS_ALL);
518 for (j = i + 1; j < i + count - 1; j++) {
519 if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
520 xfsize(j, npages, off, len))) {
521 vm_page_xunbusy(pa[j]);
522 SFSTAT_INC(sf_pages_valid);
523 SFSTAT_INC(sf_pages_bogus);
524 pa[j] = bogus_page;
525 }
526 }
527
528 refcount_acquire(&sfio->nios);
529 rv = vm_pager_get_pages_async(obj, pa + i, count, NULL,
530 i + count == npages ? &rhpages : NULL,
531 &sendfile_iodone, sfio);
532 if (__predict_false(rv != VM_PAGER_OK)) {
533 sendfile_iowait(sfio, "sferrio");
534
535 /*
536 * Do remaining pages recovery before returning EIO.
537 * Pages from 0 to npages are wired.
538 * Pages from (i + count1) to npages are busied.
539 */
540 for (j = 0; j < npages; j++) {
541 if (j >= i + count1)
542 vm_page_xunbusy(pa[j]);
543 KASSERT(pa[j] != NULL && pa[j] != bogus_page,
544 ("%s: page %p[%d] I/O recovery failure",
545 __func__, pa, j));
546 vm_page_unwire(pa[j], PQ_INACTIVE);
547 pa[j] = NULL;
548 }
549 return (EIO);
550 }
551
552 SFSTAT_INC(sf_iocnt);
553 SFSTAT_ADD(sf_pages_read, count);
554 if (i + count == npages)
555 SFSTAT_ADD(sf_rhpages_read, rhpages);
556
557 i += count1;
558 (*nios)++;
559 }
560
561 if (*nios == 0 && npages != 0)
562 SFSTAT_INC(sf_noiocnt);
563
564 return (0);
565 }
566
567 static int
sendfile_getobj(struct thread * td,struct file * fp,vm_object_t * obj_res,struct vnode ** vp_res,struct shmfd ** shmfd_res,off_t * obj_size,int * bsize)568 sendfile_getobj(struct thread *td, struct file *fp, vm_object_t *obj_res,
569 struct vnode **vp_res, struct shmfd **shmfd_res, off_t *obj_size,
570 int *bsize)
571 {
572 vm_object_t obj;
573 struct vnode *vp;
574 struct shmfd *shmfd;
575 int error;
576
577 error = 0;
578 vp = *vp_res = NULL;
579 obj = NULL;
580 shmfd = *shmfd_res = NULL;
581 *bsize = 0;
582
583 /*
584 * The file descriptor must be a regular file and have a
585 * backing VM object.
586 */
587 if (fp->f_type == DTYPE_VNODE) {
588 vp = fp->f_vnode;
589 vn_lock(vp, LK_SHARED | LK_RETRY);
590 if (vp->v_type != VREG) {
591 error = EINVAL;
592 goto out;
593 }
594 *bsize = vp->v_mount->mnt_stat.f_iosize;
595 obj = vp->v_object;
596 if (obj == NULL) {
597 error = EINVAL;
598 goto out;
599 }
600
601 /*
602 * Use the pager size when available to simplify synchronization
603 * with filesystems, which otherwise must atomically update both
604 * the vnode pager size and file size.
605 */
606 if (obj->type == OBJT_VNODE) {
607 VM_OBJECT_RLOCK(obj);
608 *obj_size = obj->un_pager.vnp.vnp_size;
609 } else {
610 error = vn_getsize_locked(vp, obj_size, td->td_ucred);
611 if (error != 0)
612 goto out;
613 VM_OBJECT_RLOCK(obj);
614 }
615 } else if (fp->f_type == DTYPE_SHM) {
616 shmfd = fp->f_data;
617 obj = shmfd->shm_object;
618 VM_OBJECT_RLOCK(obj);
619 *obj_size = shmfd->shm_size;
620 } else {
621 error = EINVAL;
622 goto out;
623 }
624
625 if ((obj->flags & OBJ_DEAD) != 0) {
626 VM_OBJECT_RUNLOCK(obj);
627 error = EBADF;
628 goto out;
629 }
630
631 /*
632 * Temporarily increase the backing VM object's reference
633 * count so that a forced reclamation of its vnode does not
634 * immediately destroy it.
635 */
636 vm_object_reference_locked(obj);
637 VM_OBJECT_RUNLOCK(obj);
638 *obj_res = obj;
639 *vp_res = vp;
640 *shmfd_res = shmfd;
641
642 out:
643 if (vp != NULL)
644 VOP_UNLOCK(vp);
645 return (error);
646 }
647
648 static int
sendfile_getsock(struct thread * td,int s,struct file ** sock_fp,struct socket ** so)649 sendfile_getsock(struct thread *td, int s, struct file **sock_fp,
650 struct socket **so)
651 {
652 int error;
653
654 *sock_fp = NULL;
655 *so = NULL;
656
657 /*
658 * The socket must be a stream socket and connected.
659 */
660 error = getsock(td, s, &cap_send_rights, sock_fp);
661 if (error != 0)
662 return (error);
663 *so = (*sock_fp)->f_data;
664 if ((*so)->so_type != SOCK_STREAM)
665 return (EINVAL);
666 /*
667 * SCTP one-to-one style sockets currently don't work with
668 * sendfile(). So indicate EINVAL for now.
669 */
670 if ((*so)->so_proto->pr_protocol == IPPROTO_SCTP)
671 return (EINVAL);
672 return (0);
673 }
674
675 /*
676 * Check socket state and wait (or EAGAIN) for needed amount of space.
677 */
678 int
sendfile_wait_generic(struct socket * so,off_t need,int * space)679 sendfile_wait_generic(struct socket *so, off_t need, int *space)
680 {
681 int error;
682
683 MPASS(need > 0);
684 MPASS(space != NULL);
685
686 /*
687 * XXXGL: the hack with sb_lowat originates from d99b0dd2c5297. To
688 * achieve high performance sending with sendfile(2) a non-blocking
689 * socket needs a fairly high low watermark. Otherwise, the socket
690 * will be reported as writable too early, and sendfile(2) will send
691 * just a few bytes each time. It is important to understand that
692 * we are changing sb_lowat not for the current invocation of the
693 * syscall, but for the *next* syscall. So there is no way to
694 * workaround the problem, e.g. provide a special version of sbspace().
695 * Since this hack has been in the kernel for a long time, we
696 * anticipate that there is a lot of software that will suffer if we
697 * remove it. See also b21104487324.
698 */
699 error = 0;
700 SOCK_SENDBUF_LOCK(so);
701 if (so->so_snd.sb_flags & SB_AUTOLOWAT) {
702 if (so->so_snd.sb_lowat < so->so_snd.sb_hiwat / 2)
703 so->so_snd.sb_lowat = so->so_snd.sb_hiwat / 2;
704 if (so->so_snd.sb_lowat < PAGE_SIZE &&
705 so->so_snd.sb_hiwat >= PAGE_SIZE)
706 so->so_snd.sb_lowat = PAGE_SIZE;
707 }
708 retry_space:
709 if (so->so_snd.sb_state & SBS_CANTSENDMORE) {
710 error = EPIPE;
711 goto done;
712 } else if (so->so_error) {
713 error = so->so_error;
714 so->so_error = 0;
715 goto done;
716 }
717 if ((so->so_state & SS_ISCONNECTED) == 0) {
718 error = ENOTCONN;
719 goto done;
720 }
721
722 *space = sbspace(&so->so_snd);
723 if (*space < need && (*space <= 0 || *space < so->so_snd.sb_lowat)) {
724 if (so->so_state & SS_NBIO) {
725 error = EAGAIN;
726 goto done;
727 }
728 /*
729 * sbwait() drops the lock while sleeping. When we loop back
730 * to retry_space the state may have changed and we retest
731 * for it.
732 */
733 error = sbwait(so, SO_SND);
734 /*
735 * An error from sbwait() usually indicates that we've been
736 * interrupted by a signal. If we've sent anything then return
737 * bytes sent, otherwise return the error.
738 */
739 if (error != 0)
740 goto done;
741 goto retry_space;
742 }
743 done:
744 SOCK_SENDBUF_UNLOCK(so);
745
746 return (error);
747 }
748
749 int
vn_sendfile(struct file * fp,int sockfd,struct uio * hdr_uio,struct uio * trl_uio,off_t offset,size_t nbytes,off_t * sent,int flags,struct thread * td)750 vn_sendfile(struct file *fp, int sockfd, struct uio *hdr_uio,
751 struct uio *trl_uio, off_t offset, size_t nbytes, off_t *sent, int flags,
752 struct thread *td)
753 {
754 struct file *sock_fp;
755 struct vnode *vp;
756 struct vm_object *obj;
757 vm_page_t pga;
758 struct socket *so;
759 const struct protosw *pr;
760 #ifdef KERN_TLS
761 struct ktls_session *tls;
762 #endif
763 struct mbuf *m, *mh, *mhtail;
764 struct sf_buf *sf;
765 struct shmfd *shmfd;
766 struct sendfile_sync *sfs;
767 struct vattr va;
768 off_t off, sbytes, rem, obj_size, nobj_size;
769 int bsize, error, ext_pgs_idx, hdrlen, max_pgs, softerr;
770 #ifdef KERN_TLS
771 int tls_enq_cnt;
772 #endif
773 bool use_ext_pgs;
774
775 obj = NULL;
776 so = NULL;
777 m = mh = NULL;
778 sfs = NULL;
779 #ifdef KERN_TLS
780 tls = NULL;
781 #endif
782 hdrlen = sbytes = 0;
783 softerr = 0;
784 use_ext_pgs = false;
785
786 error = sendfile_getobj(td, fp, &obj, &vp, &shmfd, &obj_size, &bsize);
787 if (error != 0)
788 return (error);
789
790 error = sendfile_getsock(td, sockfd, &sock_fp, &so);
791 if (error != 0)
792 goto out;
793 pr = so->so_proto;
794
795 #ifdef MAC
796 error = mac_socket_check_send(td->td_ucred, so);
797 if (error != 0)
798 goto out;
799 #endif
800
801 SFSTAT_INC(sf_syscalls);
802 SFSTAT_ADD(sf_rhpages_requested, SF_READAHEAD(flags));
803
804 if (__predict_false(flags & SF_SYNC)) {
805 gone_in(16, "Warning! %s[%u] uses SF_SYNC sendfile(2) flag. "
806 "Please follow up to https://bugs.freebsd.org/"
807 "bugzilla/show_bug.cgi?id=287348. ",
808 td->td_proc->p_comm, td->td_proc->p_pid);
809 sfs = malloc(sizeof(*sfs), M_SENDFILE, M_WAITOK | M_ZERO);
810 mtx_init(&sfs->mtx, "sendfile", NULL, MTX_DEF);
811 cv_init(&sfs->cv, "sendfile");
812 sfs->waiting = true;
813 }
814
815 rem = nbytes ? omin(nbytes, obj_size - offset) : obj_size - offset;
816
817 /*
818 * Protect against multiple writers to the socket.
819 *
820 * XXXRW: Historically this has assumed non-interruptibility, so now
821 * we implement that, but possibly shouldn't.
822 */
823 error = SOCK_IO_SEND_LOCK(so, SBL_WAIT | SBL_NOINTR);
824 if (error != 0)
825 goto out;
826 CURVNET_SET(so->so_vnet);
827 #ifdef KERN_TLS
828 tls = ktls_hold(so->so_snd.sb_tls_info);
829 #endif
830
831 /*
832 * Loop through the pages of the file, starting with the requested
833 * offset. Get a file page (do I/O if necessary), map the file page
834 * into an sf_buf, attach an mbuf header to the sf_buf, and queue
835 * it on the socket.
836 * This is done in two loops. The inner loop turns as many pages
837 * as it can, up to available socket buffer space, without blocking
838 * into mbufs to have it bulk delivered into the socket send buffer.
839 * The outer loop checks the state and available space of the socket
840 * and takes care of the overall progress.
841 */
842 for (off = offset; rem > 0; ) {
843 struct sf_io *sfio;
844 vm_page_t *pa;
845 struct mbuf *m0, *mtail;
846 int nios, space, npages, rhpages;
847
848 mtail = NULL;
849 if ((error = pr->pr_sendfile_wait(so, rem, &space)) != 0)
850 goto done;
851 /*
852 * At the beginning of the first loop check if any headers
853 * are specified and copy them into mbufs. Reduce space in
854 * the socket buffer by the size of the header mbuf chain.
855 * Clear hdr_uio here and hdrlen at the end of the first loop.
856 */
857 if (hdr_uio != NULL && hdr_uio->uio_resid > 0) {
858 hdr_uio->uio_td = td;
859 hdr_uio->uio_rw = UIO_WRITE;
860 #ifdef KERN_TLS
861 if (tls != NULL)
862 mh = m_uiotombuf(hdr_uio, M_WAITOK, space,
863 tls->params.max_frame_len, M_EXTPG);
864 else
865 #endif
866 mh = m_uiotombuf(hdr_uio, M_WAITOK,
867 space, 0, 0);
868 hdrlen = m_length(mh, &mhtail);
869 space -= hdrlen;
870 /*
871 * If header consumed all the socket buffer space,
872 * don't waste CPU cycles and jump to the end.
873 */
874 if (space == 0) {
875 sfio = NULL;
876 nios = 0;
877 goto prepend_header;
878 }
879 hdr_uio = NULL;
880 }
881
882 if (vp != NULL) {
883 error = vn_lock(vp, LK_SHARED);
884 if (error != 0)
885 goto done;
886
887 /*
888 * Check to see if the file size has changed.
889 */
890 if (obj->type == OBJT_VNODE) {
891 VM_OBJECT_RLOCK(obj);
892 nobj_size = obj->un_pager.vnp.vnp_size;
893 VM_OBJECT_RUNLOCK(obj);
894 } else {
895 error = VOP_GETATTR(vp, &va, td->td_ucred);
896 if (error != 0) {
897 VOP_UNLOCK(vp);
898 goto done;
899 }
900 nobj_size = va.va_size;
901 }
902 if (off >= nobj_size) {
903 VOP_UNLOCK(vp);
904 goto done;
905 }
906 if (nobj_size != obj_size) {
907 obj_size = nobj_size;
908 rem = nbytes ? omin(nbytes + offset, obj_size) :
909 obj_size;
910 rem -= off;
911 }
912 }
913
914 if (space > rem)
915 space = rem;
916 else if (space > PAGE_SIZE) {
917 /*
918 * Use page boundaries when possible for large
919 * requests.
920 */
921 if (off & PAGE_MASK)
922 space -= (PAGE_SIZE - (off & PAGE_MASK));
923 space = trunc_page(space);
924 if (off & PAGE_MASK)
925 space += (PAGE_SIZE - (off & PAGE_MASK));
926 }
927
928 npages = howmany(space + (off & PAGE_MASK), PAGE_SIZE);
929
930 /*
931 * Calculate maximum allowed number of pages for readahead
932 * at this iteration. If SF_USER_READAHEAD was set, we don't
933 * do any heuristics and use exactly the value supplied by
934 * application. Otherwise, we allow readahead up to "rem".
935 * If application wants more, let it be, but there is no
936 * reason to go above maxphys. Also check against "obj_size",
937 * since vm_pager_has_page() can hint beyond EOF.
938 */
939 if (flags & SF_USER_READAHEAD) {
940 rhpages = SF_READAHEAD(flags);
941 } else {
942 rhpages = howmany(rem + (off & PAGE_MASK), PAGE_SIZE) -
943 npages;
944 rhpages += SF_READAHEAD(flags);
945 }
946 rhpages = min(howmany(maxphys, PAGE_SIZE), rhpages);
947 rhpages = min(howmany(obj_size - trunc_page(off), PAGE_SIZE) -
948 npages, rhpages);
949
950 sfio = malloc(sizeof(struct sf_io) +
951 npages * sizeof(vm_page_t), M_SENDFILE, M_WAITOK);
952 refcount_init(&sfio->nios, 1);
953 sfio->obj = obj;
954 sfio->error = 0;
955 sfio->m = NULL;
956 sfio->npages = npages;
957 #ifdef KERN_TLS
958 /*
959 * This doesn't use ktls_hold() because sfio->m will
960 * also have a reference on 'tls' that will be valid
961 * for all of sfio's lifetime.
962 */
963 sfio->tls = tls;
964 #endif
965 vm_object_pip_add(obj, 1);
966 error = sendfile_swapin(obj, sfio, &nios, off, space, rhpages,
967 flags);
968 if (error != 0) {
969 if (vp != NULL)
970 VOP_UNLOCK(vp);
971 sendfile_iodone(sfio, NULL, 0, error);
972 goto done;
973 }
974
975 /*
976 * Loop and construct maximum sized mbuf chain to be bulk
977 * dumped into socket buffer.
978 */
979 pa = sfio->pa;
980
981 /*
982 * Use unmapped mbufs if enabled for TCP. Unmapped
983 * bufs are restricted to TCP as that is what has been
984 * tested. In particular, unmapped mbufs have not
985 * been tested with UNIX-domain sockets.
986 *
987 * TLS frames always require unmapped mbufs.
988 */
989 if ((mb_use_ext_pgs && pr->pr_protocol == IPPROTO_TCP)
990 #ifdef KERN_TLS
991 || tls != NULL
992 #endif
993 ) {
994 use_ext_pgs = true;
995 #ifdef KERN_TLS
996 if (tls != NULL)
997 max_pgs = num_pages(tls->params.max_frame_len);
998 else
999 #endif
1000 max_pgs = MBUF_PEXT_MAX_PGS;
1001
1002 /* Start at last index, to wrap on first use. */
1003 ext_pgs_idx = max_pgs - 1;
1004 }
1005
1006 for (int i = 0; i < npages; i++) {
1007 /*
1008 * If a page wasn't grabbed successfully, then
1009 * trim the array. Can happen only with SF_NODISKIO.
1010 */
1011 if (pa[i] == NULL) {
1012 SFSTAT_INC(sf_busy);
1013 fixspace(npages, i, off, &space);
1014 sfio->npages = i;
1015 softerr = EBUSY;
1016 break;
1017 }
1018 pga = pa[i];
1019 if (pga == bogus_page)
1020 pga = vm_page_relookup(obj, sfio->pindex0 + i);
1021
1022 if (use_ext_pgs) {
1023 off_t xfs;
1024
1025 ext_pgs_idx++;
1026 if (ext_pgs_idx == max_pgs) {
1027 m0 = mb_alloc_ext_pgs(M_WAITOK,
1028 sendfile_free_mext_pg, M_RDONLY);
1029
1030 if (flags & SF_NOCACHE) {
1031 m0->m_ext.ext_flags |=
1032 EXT_FLAG_NOCACHE;
1033
1034 /*
1035 * See comment below regarding
1036 * ignoring SF_NOCACHE for the
1037 * last page.
1038 */
1039 if ((npages - i <= max_pgs) &&
1040 ((off + space) & PAGE_MASK) &&
1041 (rem > space || rhpages > 0))
1042 m0->m_ext.ext_flags |=
1043 EXT_FLAG_CACHE_LAST;
1044 }
1045 if (sfs != NULL) {
1046 m0->m_ext.ext_flags |=
1047 EXT_FLAG_SYNC;
1048 m0->m_ext.ext_arg1 = sfs;
1049 mtx_lock(&sfs->mtx);
1050 sfs->count++;
1051 mtx_unlock(&sfs->mtx);
1052 }
1053 ext_pgs_idx = 0;
1054
1055 /* Append to mbuf chain. */
1056 if (mtail != NULL)
1057 mtail->m_next = m0;
1058 else
1059 m = m0;
1060 mtail = m0;
1061 m0->m_epg_1st_off =
1062 vmoff(i, off) & PAGE_MASK;
1063 }
1064 if (nios) {
1065 mtail->m_flags |= M_NOTREADY;
1066 m0->m_epg_nrdy++;
1067 }
1068
1069 m0->m_epg_pa[ext_pgs_idx] = VM_PAGE_TO_PHYS(pga);
1070 m0->m_epg_npgs++;
1071 xfs = xfsize(i, npages, off, space);
1072 m0->m_epg_last_len = xfs;
1073 MBUF_EXT_PGS_ASSERT_SANITY(m0);
1074 mtail->m_len += xfs;
1075 mtail->m_ext.ext_size += PAGE_SIZE;
1076 continue;
1077 }
1078
1079 /*
1080 * Get a sendfile buf. When allocating the
1081 * first buffer for mbuf chain, we usually
1082 * wait as long as necessary, but this wait
1083 * can be interrupted. For consequent
1084 * buffers, do not sleep, since several
1085 * threads might exhaust the buffers and then
1086 * deadlock.
1087 */
1088 sf = sf_buf_alloc(pga,
1089 m != NULL ? SFB_NOWAIT : SFB_CATCH);
1090 if (sf == NULL) {
1091 SFSTAT_INC(sf_allocfail);
1092 sendfile_iowait(sfio, "sfnosf");
1093 for (int j = i; j < npages; j++) {
1094 vm_page_unwire(pa[j], PQ_INACTIVE);
1095 pa[j] = NULL;
1096 }
1097 if (m == NULL)
1098 softerr = ENOBUFS;
1099 fixspace(npages, i, off, &space);
1100 sfio->npages = i;
1101 break;
1102 }
1103
1104 m0 = m_get(M_WAITOK, MT_DATA);
1105 m0->m_ext.ext_buf = (char *)sf_buf_kva(sf);
1106 m0->m_ext.ext_size = PAGE_SIZE;
1107 m0->m_ext.ext_arg1 = sf;
1108 m0->m_ext.ext_type = EXT_SFBUF;
1109 m0->m_ext.ext_flags = EXT_FLAG_EMBREF;
1110 m0->m_ext.ext_free = sendfile_free_mext;
1111 /*
1112 * SF_NOCACHE sets the page as being freed upon send.
1113 * However, we ignore it for the last page in 'space',
1114 * if the page is truncated, and we got more data to
1115 * send (rem > space), or if we have readahead
1116 * configured (rhpages > 0).
1117 */
1118 if ((flags & SF_NOCACHE) &&
1119 (i != npages - 1 ||
1120 !((off + space) & PAGE_MASK) ||
1121 !(rem > space || rhpages > 0)))
1122 m0->m_ext.ext_flags |= EXT_FLAG_NOCACHE;
1123 if (sfs != NULL) {
1124 m0->m_ext.ext_flags |= EXT_FLAG_SYNC;
1125 m0->m_ext.ext_arg2 = sfs;
1126 mtx_lock(&sfs->mtx);
1127 sfs->count++;
1128 mtx_unlock(&sfs->mtx);
1129 }
1130 m0->m_ext.ext_count = 1;
1131 m0->m_flags |= (M_EXT | M_RDONLY);
1132 if (nios)
1133 m0->m_flags |= M_NOTREADY;
1134 m0->m_data = (char *)sf_buf_kva(sf) +
1135 (vmoff(i, off) & PAGE_MASK);
1136 m0->m_len = xfsize(i, npages, off, space);
1137
1138 /* Append to mbuf chain. */
1139 if (mtail != NULL)
1140 mtail->m_next = m0;
1141 else
1142 m = m0;
1143 mtail = m0;
1144 }
1145
1146 if (vp != NULL)
1147 VOP_UNLOCK(vp);
1148
1149 /* Keep track of bytes processed. */
1150 off += space;
1151 rem -= space;
1152
1153 /*
1154 * Prepend header, if any. Save pointer to first mbuf
1155 * with a page.
1156 */
1157 if (hdrlen) {
1158 prepend_header:
1159 m0 = mhtail->m_next = m;
1160 m = mh;
1161 mh = NULL;
1162 } else
1163 m0 = m;
1164
1165 if (m == NULL) {
1166 KASSERT(softerr, ("%s: m NULL, no error", __func__));
1167 error = softerr;
1168 sendfile_iodone(sfio, NULL, 0, 0);
1169 goto done;
1170 }
1171
1172 /* Add the buffer chain to the socket buffer. */
1173 KASSERT(m_length(m, NULL) == space + hdrlen,
1174 ("%s: mlen %u space %d hdrlen %d",
1175 __func__, m_length(m, NULL), space, hdrlen));
1176
1177 #ifdef KERN_TLS
1178 if (tls != NULL)
1179 ktls_frame(m, tls, &tls_enq_cnt, TLS_RLTYPE_APP);
1180 #endif
1181 if (nios == 0) {
1182 /*
1183 * If sendfile_swapin() didn't initiate any I/Os,
1184 * which happens if all data is cached in VM, or if
1185 * the header consumed all socket buffer space and
1186 * sfio is NULL, then we can send data right now
1187 * without the PRUS_NOTREADY flag.
1188 */
1189 if (sfio != NULL)
1190 sendfile_iodone(sfio, NULL, 0, 0);
1191 #ifdef KERN_TLS
1192 if (tls != NULL && tls->mode == TCP_TLS_MODE_SW) {
1193 error = pr->pr_send(so, PRUS_NOTREADY, m, NULL,
1194 NULL, td);
1195 if (error != 0) {
1196 m_freem(m);
1197 } else {
1198 soref(so);
1199 ktls_enqueue(m, so, tls_enq_cnt);
1200 }
1201 } else
1202 #endif
1203 error = pr->pr_send(so, 0, m, NULL, NULL, td);
1204 } else {
1205 sfio->so = so;
1206 sfio->m = m0;
1207 soref(so);
1208 error = pr->pr_send(so, PRUS_NOTREADY, m, NULL, NULL,
1209 td);
1210 sendfile_iodone(sfio, NULL, 0, error);
1211 }
1212 #ifdef TCP_REQUEST_TRK
1213 if (so->so_proto->pr_protocol == IPPROTO_TCP) {
1214 /* log the sendfile call to the TCP log, if enabled */
1215 tcp_log_sendfile(so, offset, nbytes, flags);
1216 }
1217 #endif
1218 m = NULL;
1219 if (error)
1220 goto done;
1221 sbytes += space + hdrlen;
1222 if (hdrlen)
1223 hdrlen = 0;
1224 if (softerr) {
1225 error = softerr;
1226 goto done;
1227 }
1228 }
1229
1230 /*
1231 * Send trailers. Wimp out and use writev(2).
1232 */
1233 if (trl_uio != NULL) {
1234 SOCK_IO_SEND_UNLOCK(so);
1235 CURVNET_RESTORE();
1236 error = kern_writev(td, sockfd, trl_uio);
1237 if (error == 0)
1238 sbytes += td->td_retval[0];
1239 goto out;
1240 }
1241
1242 done:
1243 SOCK_IO_SEND_UNLOCK(so);
1244 CURVNET_RESTORE();
1245 out:
1246 /*
1247 * If there was no error we have to clear td->td_retval[0]
1248 * because it may have been set by writev.
1249 */
1250 if (error == 0) {
1251 td->td_retval[0] = 0;
1252 if (sbytes > 0 && vp != NULL)
1253 INOTIFY(vp, IN_ACCESS);
1254 }
1255 if (sent != NULL) {
1256 (*sent) = sbytes;
1257 }
1258 if (obj != NULL)
1259 vm_object_deallocate(obj);
1260 if (so)
1261 fdrop(sock_fp, td);
1262 if (m)
1263 m_freem(m);
1264 if (mh)
1265 m_freem(mh);
1266
1267 if (sfs != NULL) {
1268 mtx_lock(&sfs->mtx);
1269 if (sfs->count != 0)
1270 error = cv_wait_sig(&sfs->cv, &sfs->mtx);
1271 if (sfs->count == 0) {
1272 sendfile_sync_destroy(sfs);
1273 } else {
1274 sfs->waiting = false;
1275 mtx_unlock(&sfs->mtx);
1276 }
1277 }
1278 #ifdef KERN_TLS
1279 if (tls != NULL)
1280 ktls_free(tls);
1281 #endif
1282 if (error == ERESTART)
1283 error = EINTR;
1284
1285 return (error);
1286 }
1287
1288 static int
sendfile(struct thread * td,struct sendfile_args * uap,int compat)1289 sendfile(struct thread *td, struct sendfile_args *uap, int compat)
1290 {
1291 struct sf_hdtr hdtr;
1292 struct uio *hdr_uio, *trl_uio;
1293 struct file *fp;
1294 off_t sbytes;
1295 int error;
1296
1297 /*
1298 * File offset must be positive. If it goes beyond EOF
1299 * we send only the header/trailer and no payload data.
1300 */
1301 if (uap->offset < 0)
1302 return (EINVAL);
1303
1304 sbytes = 0;
1305 hdr_uio = trl_uio = NULL;
1306
1307 if (uap->hdtr != NULL) {
1308 error = copyin(uap->hdtr, &hdtr, sizeof(hdtr));
1309 if (error != 0)
1310 goto out;
1311 if (hdtr.headers != NULL) {
1312 error = copyinuio(hdtr.headers, hdtr.hdr_cnt,
1313 &hdr_uio);
1314 if (error != 0)
1315 goto out;
1316 #ifdef COMPAT_FREEBSD4
1317 /*
1318 * In FreeBSD < 5.0 the nbytes to send also included
1319 * the header. If compat is specified subtract the
1320 * header size from nbytes.
1321 */
1322 if (compat) {
1323 if (uap->nbytes > hdr_uio->uio_resid)
1324 uap->nbytes -= hdr_uio->uio_resid;
1325 else
1326 uap->nbytes = 0;
1327 }
1328 #endif
1329 }
1330 if (hdtr.trailers != NULL) {
1331 error = copyinuio(hdtr.trailers, hdtr.trl_cnt,
1332 &trl_uio);
1333 if (error != 0)
1334 goto out;
1335 }
1336 }
1337
1338 AUDIT_ARG_FD(uap->fd);
1339
1340 /*
1341 * sendfile(2) can start at any offset within a file so we require
1342 * CAP_READ+CAP_SEEK = CAP_PREAD.
1343 */
1344 if ((error = fget_read(td, uap->fd, &cap_pread_rights, &fp)) != 0)
1345 goto out;
1346
1347 error = fo_sendfile(fp, uap->s, hdr_uio, trl_uio, uap->offset,
1348 uap->nbytes, &sbytes, uap->flags, td);
1349 fdrop(fp, td);
1350
1351 if (uap->sbytes != NULL)
1352 (void)copyout(&sbytes, uap->sbytes, sizeof(off_t));
1353
1354 out:
1355 freeuio(hdr_uio);
1356 freeuio(trl_uio);
1357 return (error);
1358 }
1359
1360 /*
1361 * sendfile(2)
1362 *
1363 * int sendfile(int fd, int s, off_t offset, size_t nbytes,
1364 * struct sf_hdtr *hdtr, off_t *sbytes, int flags)
1365 *
1366 * Send a file specified by 'fd' and starting at 'offset' to a socket
1367 * specified by 's'. Send only 'nbytes' of the file or until EOF if nbytes ==
1368 * 0. Optionally add a header and/or trailer to the socket output. If
1369 * specified, write the total number of bytes sent into *sbytes.
1370 */
1371 int
sys_sendfile(struct thread * td,struct sendfile_args * uap)1372 sys_sendfile(struct thread *td, struct sendfile_args *uap)
1373 {
1374
1375 return (sendfile(td, uap, 0));
1376 }
1377
1378 #ifdef COMPAT_FREEBSD4
1379 int
freebsd4_sendfile(struct thread * td,struct freebsd4_sendfile_args * uap)1380 freebsd4_sendfile(struct thread *td, struct freebsd4_sendfile_args *uap)
1381 {
1382 struct sendfile_args args;
1383
1384 args.fd = uap->fd;
1385 args.s = uap->s;
1386 args.offset = uap->offset;
1387 args.nbytes = uap->nbytes;
1388 args.hdtr = uap->hdtr;
1389 args.sbytes = uap->sbytes;
1390 args.flags = uap->flags;
1391
1392 return (sendfile(td, &args, 1));
1393 }
1394 #endif /* COMPAT_FREEBSD4 */
1395