2 * Copyright (c) 2013-2015 Gleb Smirnoff <glebius@FreeBSD.org>
3 * Copyright (c) 1998, David Greenman. All rights reserved.
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
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.
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
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
33 #include "opt_compat.h"
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/capsicum.h>
38 #include <sys/kernel.h>
40 #include <sys/mutex.h>
41 #include <sys/sysproto.h>
42 #include <sys/malloc.h>
45 #include <sys/mount.h>
47 #include <sys/protosw.h>
48 #include <sys/rwlock.h>
49 #include <sys/sf_buf.h>
50 #include <sys/socket.h>
51 #include <sys/socketvar.h>
52 #include <sys/syscallsubr.h>
53 #include <sys/sysctl.h>
54 #include <sys/vnode.h>
58 #include <security/audit/audit.h>
59 #include <security/mac/mac_framework.h>
62 #include <vm/vm_object.h>
63 #include <vm/vm_pager.h>
65 #define EXT_FLAG_SYNC EXT_FLAG_VENDOR1
66 #define EXT_FLAG_NOCACHE EXT_FLAG_VENDOR2
69 * Structure describing a single sendfile(2) I/O, which may consist of
70 * several underlying pager I/Os.
72 * The syscall context allocates the structure and initializes 'nios'
73 * to 1. As sendfile_swapin() runs through pages and starts asynchronous
74 * paging operations, it increments 'nios'.
76 * Every I/O completion calls sendfile_iodone(), which decrements the 'nios',
77 * and the syscall also calls sendfile_iodone() after allocating all mbufs,
78 * linking them and sending to socket. Whoever reaches zero 'nios' is
79 * responsible to * call pru_ready on the socket, to notify it of readyness
92 * Structure used to track requests with SF_SYNC flag.
94 struct sendfile_sync {
100 counter_u64_t sfstat[sizeof(struct sfstat) / sizeof(uint64_t)];
103 sfstat_init(const void *unused)
106 COUNTER_ARRAY_ALLOC(sfstat, sizeof(struct sfstat) / sizeof(uint64_t),
109 SYSINIT(sfstat, SI_SUB_MBUF, SI_ORDER_FIRST, sfstat_init, NULL);
112 sfstat_sysctl(SYSCTL_HANDLER_ARGS)
116 COUNTER_ARRAY_COPY(sfstat, &s, sizeof(s) / sizeof(uint64_t));
118 COUNTER_ARRAY_ZERO(sfstat, sizeof(s) / sizeof(uint64_t));
119 return (SYSCTL_OUT(req, &s, sizeof(s)));
121 SYSCTL_PROC(_kern_ipc, OID_AUTO, sfstat, CTLTYPE_OPAQUE | CTLFLAG_RW,
122 NULL, 0, sfstat_sysctl, "I", "sendfile statistics");
125 * Detach mapped page and release resources back to the system. Called
126 * by mbuf(9) code when last reference to a page is freed.
129 sendfile_free_page(vm_page_t pg, bool nocache)
135 * In either case check for the object going away on us. This can
136 * happen since we don't hold a reference to it. If so, we're
137 * responsible for freeing the page. In 'noncache' case try to free
138 * the page, but only if it is cheap to.
140 if (vm_page_unwire_noq(pg)) {
143 if ((obj = pg->object) == NULL)
147 if (nocache && !vm_page_xbusied(pg) &&
148 VM_OBJECT_TRYWLOCK(obj)) {
149 /* Only free unmapped pages. */
150 if (obj->ref_count == 0 ||
151 !pmap_page_is_mapped(pg))
153 * The busy test before the object is
154 * locked cannot be relied upon.
156 freed = vm_page_try_to_free(pg);
157 VM_OBJECT_WUNLOCK(obj);
161 * If we were asked to not cache the page, place
162 * it near the head of the inactive queue so
163 * that it is reclaimed sooner. Otherwise,
167 vm_page_deactivate_noreuse(pg);
168 else if (pg->queue == PQ_ACTIVE)
169 vm_page_reference(pg);
171 vm_page_deactivate(pg);
179 sendfile_free_mext(struct mbuf *m)
185 KASSERT(m->m_flags & M_EXT && m->m_ext.ext_type == EXT_SFBUF,
186 ("%s: m %p !M_EXT or !EXT_SFBUF", __func__, m));
188 sf = m->m_ext.ext_arg1;
189 pg = sf_buf_page(sf);
190 nocache = m->m_ext.ext_flags & EXT_FLAG_NOCACHE;
193 sendfile_free_page(pg, nocache);
195 if (m->m_ext.ext_flags & EXT_FLAG_SYNC) {
196 struct sendfile_sync *sfs = m->m_ext.ext_arg2;
199 KASSERT(sfs->count > 0, ("Sendfile sync botchup count == 0"));
200 if (--sfs->count == 0)
202 mtx_unlock(&sfs->mtx);
207 * Helper function to calculate how much data to put into page i of n.
208 * Only first and last pages are special.
211 xfsize(int i, int n, off_t off, off_t len)
215 return (omin(PAGE_SIZE - (off & PAGE_MASK), len));
217 if (i == n - 1 && ((off + len) & PAGE_MASK) > 0)
218 return ((off + len) & PAGE_MASK);
224 * Helper function to get offset within object for i page.
226 static inline vm_ooffset_t
227 vmoff(int i, off_t off)
231 return ((vm_ooffset_t)off);
233 return (trunc_page(off + i * PAGE_SIZE));
237 * Helper function used when allocation of a page or sf_buf failed.
238 * Pretend as if we don't have enough space, subtract xfsize() of
239 * all pages that failed.
242 fixspace(int old, int new, off_t off, int *space)
245 KASSERT(old > new, ("%s: old %d new %d", __func__, old, new));
247 /* Subtract last one. */
248 *space -= xfsize(old - 1, old, off, *space);
252 /* There was only one page. */
255 /* Subtract first one. */
257 *space -= xfsize(0, old, off, *space);
261 /* Rest of pages are full sized. */
262 *space -= (old - new) * PAGE_SIZE;
264 KASSERT(*space >= 0, ("%s: space went backwards", __func__));
268 * I/O completion callback.
271 sendfile_iodone(void *arg, vm_page_t *pg, int count, int error)
273 struct sf_io *sfio = arg;
274 struct socket *so = sfio->so;
276 for (int i = 0; i < count; i++)
277 if (pg[i] != bogus_page)
278 vm_page_xunbusy(pg[i]);
283 if (!refcount_release(&sfio->nios))
286 CURVNET_SET(so->so_vnet);
291 * I/O operation failed. The state of data in the socket
292 * is now inconsistent, and all what we can do is to tear
293 * it down. Protocol abort method would tear down protocol
294 * state, free all ready mbufs and detach not ready ones.
295 * We will free the mbufs corresponding to this I/O manually.
297 * The socket would be marked with EIO and made available
298 * for read, so that application receives EIO on next
299 * syscall and eventually closes the socket.
301 so->so_proto->pr_usrreqs->pru_abort(so);
305 for (int i = 0; i < sfio->npages; i++)
308 (void )(so->so_proto->pr_usrreqs->pru_ready)(so, sfio->m,
318 * Iterate through pages vector and request paging for non-valid pages.
321 sendfile_swapin(vm_object_t obj, struct sf_io *sfio, off_t off, off_t len,
322 int npages, int rhpages, int flags)
324 vm_page_t *pa = sfio->pa;
328 flags = (flags & SF_NODISKIO) ? VM_ALLOC_NOWAIT : 0;
331 * First grab all the pages and wire them. Note that we grab
332 * only required pages. Readahead pages are dealt with later.
334 VM_OBJECT_WLOCK(obj);
336 grabbed = vm_page_grab_pages(obj, OFF_TO_IDX(off),
337 VM_ALLOC_NORMAL | VM_ALLOC_WIRED | flags, pa, npages);
338 if (grabbed < npages) {
339 for (int i = grabbed; i < npages; i++)
345 for (int i = 0; i < npages;) {
348 /* Skip valid pages. */
349 if (vm_page_is_valid(pa[i], vmoff(i, off) & PAGE_MASK,
350 xfsize(i, npages, off, len))) {
351 vm_page_xunbusy(pa[i]);
352 SFSTAT_INC(sf_pages_valid);
358 * Next page is invalid. Check if it belongs to pager. It
359 * may not be there, which is a regular situation for shmem
360 * pager. For vnode pager this happens only in case of
363 * Important feature of vm_pager_has_page() is the hint
364 * stored in 'a', about how many pages we can pagein after
365 * this page in a single I/O.
367 if (!vm_pager_has_page(obj, OFF_TO_IDX(vmoff(i, off)), NULL,
369 pmap_zero_page(pa[i]);
370 pa[i]->valid = VM_PAGE_BITS_ALL;
371 MPASS(pa[i]->dirty == 0);
372 vm_page_xunbusy(pa[i]);
378 * We want to pagein as many pages as possible, limited only
379 * by the 'a' hint and actual request.
381 count = min(a + 1, npages - i);
384 * We should not pagein into a valid page, thus we first trim
385 * any valid pages off the end of request, and substitute
386 * to bogus_page those, that are in the middle.
388 for (j = i + count - 1; j > i; j--) {
389 if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
390 xfsize(j, npages, off, len))) {
396 for (j = i + 1; j < i + count - 1; j++)
397 if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
398 xfsize(j, npages, off, len))) {
399 vm_page_xunbusy(pa[j]);
400 SFSTAT_INC(sf_pages_valid);
401 SFSTAT_INC(sf_pages_bogus);
405 refcount_acquire(&sfio->nios);
406 rv = vm_pager_get_pages_async(obj, pa + i, count, NULL,
407 i + count == npages ? &rhpages : NULL,
408 &sendfile_iodone, sfio);
409 KASSERT(rv == VM_PAGER_OK, ("%s: pager fail obj %p page %p",
410 __func__, obj, pa[i]));
412 SFSTAT_INC(sf_iocnt);
413 SFSTAT_ADD(sf_pages_read, count);
414 if (i + count == npages)
415 SFSTAT_ADD(sf_rhpages_read, rhpages);
418 * Restore the valid page pointers. They are already
419 * unbusied, but still wired.
421 for (j = i; j < i + count; j++)
422 if (pa[j] == bogus_page) {
423 pa[j] = vm_page_lookup(obj,
424 OFF_TO_IDX(vmoff(j, off)));
425 KASSERT(pa[j], ("%s: page %p[%d] disappeared",
433 VM_OBJECT_WUNLOCK(obj);
435 if (nios == 0 && npages != 0)
436 SFSTAT_INC(sf_noiocnt);
442 sendfile_getobj(struct thread *td, struct file *fp, vm_object_t *obj_res,
443 struct vnode **vp_res, struct shmfd **shmfd_res, off_t *obj_size,
454 shmfd = *shmfd_res = NULL;
458 * The file descriptor must be a regular file and have a
461 if (fp->f_type == DTYPE_VNODE) {
463 vn_lock(vp, LK_SHARED | LK_RETRY);
464 if (vp->v_type != VREG) {
468 *bsize = vp->v_mount->mnt_stat.f_iosize;
469 error = VOP_GETATTR(vp, &va, td->td_ucred);
472 *obj_size = va.va_size;
478 } else if (fp->f_type == DTYPE_SHM) {
481 obj = shmfd->shm_object;
482 *obj_size = shmfd->shm_size;
488 VM_OBJECT_WLOCK(obj);
489 if ((obj->flags & OBJ_DEAD) != 0) {
490 VM_OBJECT_WUNLOCK(obj);
496 * Temporarily increase the backing VM object's reference
497 * count so that a forced reclamation of its vnode does not
498 * immediately destroy it.
500 vm_object_reference_locked(obj);
501 VM_OBJECT_WUNLOCK(obj);
513 sendfile_getsock(struct thread *td, int s, struct file **sock_fp,
523 * The socket must be a stream socket and connected.
525 error = getsock_cap(td, s, cap_rights_init(&rights, CAP_SEND),
526 sock_fp, NULL, NULL);
529 *so = (*sock_fp)->f_data;
530 if ((*so)->so_type != SOCK_STREAM)
536 vn_sendfile(struct file *fp, int sockfd, struct uio *hdr_uio,
537 struct uio *trl_uio, off_t offset, size_t nbytes, off_t *sent, int flags,
540 struct file *sock_fp;
542 struct vm_object *obj;
544 struct mbuf *m, *mh, *mhtail;
547 struct sendfile_sync *sfs;
549 off_t off, sbytes, rem, obj_size;
550 int error, softerr, bsize, hdrlen;
559 error = sendfile_getobj(td, fp, &obj, &vp, &shmfd, &obj_size, &bsize);
563 error = sendfile_getsock(td, sockfd, &sock_fp, &so);
568 error = mac_socket_check_send(td->td_ucred, so);
573 SFSTAT_INC(sf_syscalls);
574 SFSTAT_ADD(sf_rhpages_requested, SF_READAHEAD(flags));
576 if (flags & SF_SYNC) {
577 sfs = malloc(sizeof *sfs, M_TEMP, M_WAITOK | M_ZERO);
578 mtx_init(&sfs->mtx, "sendfile", NULL, MTX_DEF);
579 cv_init(&sfs->cv, "sendfile");
582 rem = nbytes ? omin(nbytes, obj_size - offset) : obj_size - offset;
585 * Protect against multiple writers to the socket.
587 * XXXRW: Historically this has assumed non-interruptibility, so now
588 * we implement that, but possibly shouldn't.
590 (void)sblock(&so->so_snd, SBL_WAIT | SBL_NOINTR);
593 * Loop through the pages of the file, starting with the requested
594 * offset. Get a file page (do I/O if necessary), map the file page
595 * into an sf_buf, attach an mbuf header to the sf_buf, and queue
597 * This is done in two loops. The inner loop turns as many pages
598 * as it can, up to available socket buffer space, without blocking
599 * into mbufs to have it bulk delivered into the socket send buffer.
600 * The outer loop checks the state and available space of the socket
601 * and takes care of the overall progress.
603 for (off = offset; rem > 0; ) {
607 int nios, space, npages, rhpages;
611 * Check the socket state for ongoing connection,
612 * no errors and space in socket buffer.
613 * If space is low allow for the remainder of the
614 * file to be processed if it fits the socket buffer.
615 * Otherwise block in waiting for sufficient space
616 * to proceed, or if the socket is nonblocking, return
617 * to userland with EAGAIN while reporting how far
619 * We wait until the socket buffer has significant free
620 * space to do bulk sends. This makes good use of file
621 * system read ahead and allows packet segmentation
622 * offloading hardware to take over lots of work. If
623 * we were not careful here we would send off only one
626 SOCKBUF_LOCK(&so->so_snd);
627 if (so->so_snd.sb_lowat < so->so_snd.sb_hiwat / 2)
628 so->so_snd.sb_lowat = so->so_snd.sb_hiwat / 2;
630 if (so->so_snd.sb_state & SBS_CANTSENDMORE) {
632 SOCKBUF_UNLOCK(&so->so_snd);
634 } else if (so->so_error) {
635 error = so->so_error;
637 SOCKBUF_UNLOCK(&so->so_snd);
640 if ((so->so_state & SS_ISCONNECTED) == 0) {
641 SOCKBUF_UNLOCK(&so->so_snd);
646 space = sbspace(&so->so_snd);
649 space < so->so_snd.sb_lowat)) {
650 if (so->so_state & SS_NBIO) {
651 SOCKBUF_UNLOCK(&so->so_snd);
656 * sbwait drops the lock while sleeping.
657 * When we loop back to retry_space the
658 * state may have changed and we retest
661 error = sbwait(&so->so_snd);
663 * An error from sbwait usually indicates that we've
664 * been interrupted by a signal. If we've sent anything
665 * then return bytes sent, otherwise return the error.
668 SOCKBUF_UNLOCK(&so->so_snd);
673 SOCKBUF_UNLOCK(&so->so_snd);
676 * At the beginning of the first loop check if any headers
677 * are specified and copy them into mbufs. Reduce space in
678 * the socket buffer by the size of the header mbuf chain.
679 * Clear hdr_uio here and hdrlen at the end of the first loop.
681 if (hdr_uio != NULL && hdr_uio->uio_resid > 0) {
682 hdr_uio->uio_td = td;
683 hdr_uio->uio_rw = UIO_WRITE;
684 mh = m_uiotombuf(hdr_uio, M_WAITOK, space, 0, 0);
685 hdrlen = m_length(mh, &mhtail);
688 * If header consumed all the socket buffer space,
689 * don't waste CPU cycles and jump to the end.
700 error = vn_lock(vp, LK_SHARED);
703 error = VOP_GETATTR(vp, &va, td->td_ucred);
704 if (error != 0 || off >= va.va_size) {
708 if (va.va_size != obj_size) {
709 obj_size = va.va_size;
711 omin(nbytes + offset, obj_size) : obj_size;
719 npages = howmany(space + (off & PAGE_MASK), PAGE_SIZE);
722 * Calculate maximum allowed number of pages for readahead
723 * at this iteration. If SF_USER_READAHEAD was set, we don't
724 * do any heuristics and use exactly the value supplied by
725 * application. Otherwise, we allow readahead up to "rem".
726 * If application wants more, let it be, but there is no
727 * reason to go above MAXPHYS. Also check against "obj_size",
728 * since vm_pager_has_page() can hint beyond EOF.
730 if (flags & SF_USER_READAHEAD) {
731 rhpages = SF_READAHEAD(flags);
733 rhpages = howmany(rem + (off & PAGE_MASK), PAGE_SIZE) -
735 rhpages += SF_READAHEAD(flags);
737 rhpages = min(howmany(MAXPHYS, PAGE_SIZE), rhpages);
738 rhpages = min(howmany(obj_size - trunc_page(off), PAGE_SIZE) -
741 sfio = malloc(sizeof(struct sf_io) +
742 npages * sizeof(vm_page_t), M_TEMP, M_WAITOK);
743 refcount_init(&sfio->nios, 1);
747 nios = sendfile_swapin(obj, sfio, off, space, npages, rhpages,
751 * Loop and construct maximum sized mbuf chain to be bulk
752 * dumped into socket buffer.
755 for (int i = 0; i < npages; i++) {
759 * If a page wasn't grabbed successfully, then
760 * trim the array. Can happen only with SF_NODISKIO.
764 fixspace(npages, i, off, &space);
771 * Get a sendfile buf. When allocating the
772 * first buffer for mbuf chain, we usually
773 * wait as long as necessary, but this wait
774 * can be interrupted. For consequent
775 * buffers, do not sleep, since several
776 * threads might exhaust the buffers and then
779 sf = sf_buf_alloc(pa[i],
780 m != NULL ? SFB_NOWAIT : SFB_CATCH);
782 SFSTAT_INC(sf_allocfail);
783 for (int j = i; j < npages; j++) {
785 vm_page_unwire(pa[j], PQ_INACTIVE);
786 vm_page_unlock(pa[j]);
790 fixspace(npages, i, off, &space);
795 m0 = m_get(M_WAITOK, MT_DATA);
796 m0->m_ext.ext_buf = (char *)sf_buf_kva(sf);
797 m0->m_ext.ext_size = PAGE_SIZE;
798 m0->m_ext.ext_arg1 = sf;
799 m0->m_ext.ext_type = EXT_SFBUF;
800 m0->m_ext.ext_flags = EXT_FLAG_EMBREF;
801 m0->m_ext.ext_free = sendfile_free_mext;
803 * SF_NOCACHE sets the page as being freed upon send.
804 * However, we ignore it for the last page in 'space',
805 * if the page is truncated, and we got more data to
806 * send (rem > space), or if we have readahead
807 * configured (rhpages > 0).
809 if ((flags & SF_NOCACHE) &&
811 !((off + space) & PAGE_MASK) ||
812 !(rem > space || rhpages > 0)))
813 m0->m_ext.ext_flags |= EXT_FLAG_NOCACHE;
815 m0->m_ext.ext_flags |= EXT_FLAG_SYNC;
816 m0->m_ext.ext_arg2 = sfs;
819 mtx_unlock(&sfs->mtx);
821 m0->m_ext.ext_count = 1;
822 m0->m_flags |= (M_EXT | M_RDONLY);
824 m0->m_flags |= M_NOTREADY;
825 m0->m_data = (char *)sf_buf_kva(sf) +
826 (vmoff(i, off) & PAGE_MASK);
827 m0->m_len = xfsize(i, npages, off, space);
832 /* Append to mbuf chain. */
843 /* Keep track of bytes processed. */
847 /* Prepend header, if any. */
856 KASSERT(softerr, ("%s: m NULL, no error", __func__));
862 /* Add the buffer chain to the socket buffer. */
863 KASSERT(m_length(m, NULL) == space + hdrlen,
864 ("%s: mlen %u space %d hdrlen %d",
865 __func__, m_length(m, NULL), space, hdrlen));
867 CURVNET_SET(so->so_vnet);
870 * If sendfile_swapin() didn't initiate any I/Os,
871 * which happens if all data is cached in VM, then
872 * we can send data right now without the
873 * PRUS_NOTREADY flag.
876 error = (*so->so_proto->pr_usrreqs->pru_send)
877 (so, 0, m, NULL, NULL, td);
879 sfio->npages = npages;
881 error = (*so->so_proto->pr_usrreqs->pru_send)
882 (so, PRUS_NOTREADY, m, NULL, NULL, td);
883 sendfile_iodone(sfio, NULL, 0, 0);
887 m = NULL; /* pru_send always consumes */
890 sbytes += space + hdrlen;
900 * Send trailers. Wimp out and use writev(2).
902 if (trl_uio != NULL) {
903 sbunlock(&so->so_snd);
904 error = kern_writev(td, sockfd, trl_uio);
906 sbytes += td->td_retval[0];
911 sbunlock(&so->so_snd);
914 * If there was no error we have to clear td->td_retval[0]
915 * because it may have been set by writev.
918 td->td_retval[0] = 0;
924 vm_object_deallocate(obj);
935 cv_wait(&sfs->cv, &sfs->mtx);
936 KASSERT(sfs->count == 0, ("sendfile sync still busy"));
937 cv_destroy(&sfs->cv);
938 mtx_destroy(&sfs->mtx);
942 if (error == ERESTART)
949 sendfile(struct thread *td, struct sendfile_args *uap, int compat)
952 struct uio *hdr_uio, *trl_uio;
959 * File offset must be positive. If it goes beyond EOF
960 * we send only the header/trailer and no payload data.
966 hdr_uio = trl_uio = NULL;
968 if (uap->hdtr != NULL) {
969 error = copyin(uap->hdtr, &hdtr, sizeof(hdtr));
972 if (hdtr.headers != NULL) {
973 error = copyinuio(hdtr.headers, hdtr.hdr_cnt,
977 #ifdef COMPAT_FREEBSD4
979 * In FreeBSD < 5.0 the nbytes to send also included
980 * the header. If compat is specified subtract the
981 * header size from nbytes.
984 if (uap->nbytes > hdr_uio->uio_resid)
985 uap->nbytes -= hdr_uio->uio_resid;
991 if (hdtr.trailers != NULL) {
992 error = copyinuio(hdtr.trailers, hdtr.trl_cnt,
999 AUDIT_ARG_FD(uap->fd);
1002 * sendfile(2) can start at any offset within a file so we require
1003 * CAP_READ+CAP_SEEK = CAP_PREAD.
1005 if ((error = fget_read(td, uap->fd,
1006 cap_rights_init(&rights, CAP_PREAD), &fp)) != 0) {
1010 error = fo_sendfile(fp, uap->s, hdr_uio, trl_uio, uap->offset,
1011 uap->nbytes, &sbytes, uap->flags, td);
1014 if (uap->sbytes != NULL)
1015 copyout(&sbytes, uap->sbytes, sizeof(off_t));
1018 free(hdr_uio, M_IOV);
1019 free(trl_uio, M_IOV);
1026 * int sendfile(int fd, int s, off_t offset, size_t nbytes,
1027 * struct sf_hdtr *hdtr, off_t *sbytes, int flags)
1029 * Send a file specified by 'fd' and starting at 'offset' to a socket
1030 * specified by 's'. Send only 'nbytes' of the file or until EOF if nbytes ==
1031 * 0. Optionally add a header and/or trailer to the socket output. If
1032 * specified, write the total number of bytes sent into *sbytes.
1035 sys_sendfile(struct thread *td, struct sendfile_args *uap)
1038 return (sendfile(td, uap, 0));
1041 #ifdef COMPAT_FREEBSD4
1043 freebsd4_sendfile(struct thread *td, struct freebsd4_sendfile_args *uap)
1045 struct sendfile_args args;
1049 args.offset = uap->offset;
1050 args.nbytes = uap->nbytes;
1051 args.hdtr = uap->hdtr;
1052 args.sbytes = uap->sbytes;
1053 args.flags = uap->flags;
1055 return (sendfile(td, &args, 1));
1057 #endif /* COMPAT_FREEBSD4 */