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 <sys/param.h>
34 #include <sys/systm.h>
35 #include <sys/capsicum.h>
36 #include <sys/kernel.h>
38 #include <sys/mutex.h>
39 #include <sys/sysproto.h>
40 #include <sys/malloc.h>
43 #include <sys/mount.h>
45 #include <sys/protosw.h>
46 #include <sys/rwlock.h>
48 #include <sys/sf_buf.h>
49 #include <sys/socket.h>
50 #include <sys/socketvar.h>
51 #include <sys/syscallsubr.h>
52 #include <sys/sysctl.h>
53 #include <sys/vnode.h>
57 #include <security/audit/audit.h>
58 #include <security/mac/mac_framework.h>
61 #include <vm/vm_object.h>
62 #include <vm/vm_pager.h>
64 #define EXT_FLAG_SYNC EXT_FLAG_VENDOR1
65 #define EXT_FLAG_NOCACHE EXT_FLAG_VENDOR2
67 SDT_PROVIDER_DECLARE(vfs);
70 * Structure describing a single sendfile(2) I/O, which may consist of
71 * several underlying pager I/Os.
73 * The syscall context allocates the structure and initializes 'nios'
74 * to 1. As sendfile_swapin() runs through pages and starts asynchronous
75 * paging operations, it increments 'nios'.
77 * Every I/O completion calls sendfile_iodone(), which decrements the 'nios',
78 * and the syscall also calls sendfile_iodone() after allocating all mbufs,
79 * linking them and sending to socket. Whoever reaches zero 'nios' is
80 * responsible to * call pru_ready on the socket, to notify it of readyness
93 * Structure used to track requests with SF_SYNC flag.
95 struct sendfile_sync {
101 counter_u64_t sfstat[sizeof(struct sfstat) / sizeof(uint64_t)];
104 sfstat_init(const void *unused)
107 COUNTER_ARRAY_ALLOC(sfstat, sizeof(struct sfstat) / sizeof(uint64_t),
110 SYSINIT(sfstat, SI_SUB_MBUF, SI_ORDER_FIRST, sfstat_init, NULL);
113 sfstat_sysctl(SYSCTL_HANDLER_ARGS)
117 COUNTER_ARRAY_COPY(sfstat, &s, sizeof(s) / sizeof(uint64_t));
119 COUNTER_ARRAY_ZERO(sfstat, sizeof(s) / sizeof(uint64_t));
120 return (SYSCTL_OUT(req, &s, sizeof(s)));
122 SYSCTL_PROC(_kern_ipc, OID_AUTO, sfstat, CTLTYPE_OPAQUE | CTLFLAG_RW,
123 NULL, 0, sfstat_sysctl, "I", "sendfile statistics");
126 * Detach mapped page and release resources back to the system. Called
127 * by mbuf(9) code when last reference to a page is freed.
130 sendfile_free_page(vm_page_t pg, bool nocache)
136 * In either case check for the object going away on us. This can
137 * happen since we don't hold a reference to it. If so, we're
138 * responsible for freeing the page. In 'noncache' case try to free
139 * the page, but only if it is cheap to.
141 if (vm_page_unwire_noq(pg)) {
144 if ((obj = pg->object) == NULL)
148 if (nocache && !vm_page_xbusied(pg) &&
149 VM_OBJECT_TRYWLOCK(obj)) {
150 /* Only free unmapped pages. */
151 if (obj->ref_count == 0 ||
152 !pmap_page_is_mapped(pg))
154 * The busy test before the object is
155 * locked cannot be relied upon.
157 freed = vm_page_try_to_free(pg);
158 VM_OBJECT_WUNLOCK(obj);
162 * If we were asked to not cache the page, place
163 * it near the head of the inactive queue so
164 * that it is reclaimed sooner. Otherwise,
168 vm_page_deactivate_noreuse(pg);
169 else if (vm_page_active(pg))
170 vm_page_reference(pg);
172 vm_page_deactivate(pg);
180 sendfile_free_mext(struct mbuf *m)
186 KASSERT(m->m_flags & M_EXT && m->m_ext.ext_type == EXT_SFBUF,
187 ("%s: m %p !M_EXT or !EXT_SFBUF", __func__, m));
189 sf = m->m_ext.ext_arg1;
190 pg = sf_buf_page(sf);
191 nocache = m->m_ext.ext_flags & EXT_FLAG_NOCACHE;
194 sendfile_free_page(pg, nocache);
196 if (m->m_ext.ext_flags & EXT_FLAG_SYNC) {
197 struct sendfile_sync *sfs = m->m_ext.ext_arg2;
200 KASSERT(sfs->count > 0, ("Sendfile sync botchup count == 0"));
201 if (--sfs->count == 0)
203 mtx_unlock(&sfs->mtx);
208 * Helper function to calculate how much data to put into page i of n.
209 * Only first and last pages are special.
212 xfsize(int i, int n, off_t off, off_t len)
216 return (omin(PAGE_SIZE - (off & PAGE_MASK), len));
218 if (i == n - 1 && ((off + len) & PAGE_MASK) > 0)
219 return ((off + len) & PAGE_MASK);
225 * Helper function to get offset within object for i page.
227 static inline vm_ooffset_t
228 vmoff(int i, off_t off)
232 return ((vm_ooffset_t)off);
234 return (trunc_page(off + i * PAGE_SIZE));
238 * Helper function used when allocation of a page or sf_buf failed.
239 * Pretend as if we don't have enough space, subtract xfsize() of
240 * all pages that failed.
243 fixspace(int old, int new, off_t off, int *space)
246 KASSERT(old > new, ("%s: old %d new %d", __func__, old, new));
248 /* Subtract last one. */
249 *space -= xfsize(old - 1, old, off, *space);
253 /* There was only one page. */
256 /* Subtract first one. */
258 *space -= xfsize(0, old, off, *space);
262 /* Rest of pages are full sized. */
263 *space -= (old - new) * PAGE_SIZE;
265 KASSERT(*space >= 0, ("%s: space went backwards", __func__));
269 * I/O completion callback.
272 sendfile_iodone(void *arg, vm_page_t *pg, int count, int error)
274 struct sf_io *sfio = arg;
275 struct socket *so = sfio->so;
277 for (int i = 0; i < count; i++)
278 if (pg[i] != bogus_page)
279 vm_page_xunbusy(pg[i]);
284 if (!refcount_release(&sfio->nios))
287 CURVNET_SET(so->so_vnet);
292 * I/O operation failed. The state of data in the socket
293 * is now inconsistent, and all what we can do is to tear
294 * it down. Protocol abort method would tear down protocol
295 * state, free all ready mbufs and detach not ready ones.
296 * We will free the mbufs corresponding to this I/O manually.
298 * The socket would be marked with EIO and made available
299 * for read, so that application receives EIO on next
300 * syscall and eventually closes the socket.
302 so->so_proto->pr_usrreqs->pru_abort(so);
306 for (int i = 0; i < sfio->npages; i++)
309 (void)(so->so_proto->pr_usrreqs->pru_ready)(so, sfio->m,
318 SDT_PROBE_DEFINE1(vfs, sendfile, swapin, pager_error, "int");
320 * Iterate through pages vector and request paging for non-valid pages.
323 sendfile_swapin(vm_object_t obj, struct sf_io *sfio, int *nios, off_t off,
324 off_t len, int npages, int rhpages, int flags)
326 vm_page_t *pa = sfio->pa;
330 flags = (flags & SF_NODISKIO) ? VM_ALLOC_NOWAIT : 0;
333 * First grab all the pages and wire them. Note that we grab
334 * only required pages. Readahead pages are dealt with later.
336 VM_OBJECT_WLOCK(obj);
338 grabbed = vm_page_grab_pages(obj, OFF_TO_IDX(off),
339 VM_ALLOC_NORMAL | VM_ALLOC_WIRED | flags, pa, npages);
340 if (grabbed < npages) {
341 for (int i = grabbed; i < npages; i++)
347 for (int i = 0; i < npages;) {
350 /* Skip valid pages. */
351 if (vm_page_is_valid(pa[i], vmoff(i, off) & PAGE_MASK,
352 xfsize(i, npages, off, len))) {
353 vm_page_xunbusy(pa[i]);
354 SFSTAT_INC(sf_pages_valid);
360 * Next page is invalid. Check if it belongs to pager. It
361 * may not be there, which is a regular situation for shmem
362 * pager. For vnode pager this happens only in case of
365 * Important feature of vm_pager_has_page() is the hint
366 * stored in 'a', about how many pages we can pagein after
367 * this page in a single I/O.
369 if (!vm_pager_has_page(obj, OFF_TO_IDX(vmoff(i, off)), NULL,
371 pmap_zero_page(pa[i]);
372 pa[i]->valid = VM_PAGE_BITS_ALL;
373 MPASS(pa[i]->dirty == 0);
374 vm_page_xunbusy(pa[i]);
380 * We want to pagein as many pages as possible, limited only
381 * by the 'a' hint and actual request.
383 count = min(a + 1, npages - i);
386 * We should not pagein into a valid page, thus we first trim
387 * any valid pages off the end of request, and substitute
388 * to bogus_page those, that are in the middle.
390 for (j = i + count - 1; j > i; j--) {
391 if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
392 xfsize(j, npages, off, len))) {
398 for (j = i + 1; j < i + count - 1; j++)
399 if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
400 xfsize(j, npages, off, len))) {
401 vm_page_xunbusy(pa[j]);
402 SFSTAT_INC(sf_pages_valid);
403 SFSTAT_INC(sf_pages_bogus);
407 refcount_acquire(&sfio->nios);
408 rv = vm_pager_get_pages_async(obj, pa + i, count, NULL,
409 i + count == npages ? &rhpages : NULL,
410 &sendfile_iodone, sfio);
411 if (rv != VM_PAGER_OK) {
412 SDT_PROBE1(vfs, sendfile, swapin, pager_error, rv);
413 for (j = 0; j < count; j++) {
414 vm_page_lock(*(pa + i + j));
415 vm_page_unwire(*(pa + i + j), PQ_INACTIVE);
416 vm_page_unlock(*(pa + i + j));
418 VM_OBJECT_WUNLOCK(obj);
421 KASSERT(rv == VM_PAGER_OK, ("%s: pager fail obj %p page %p",
422 __func__, obj, pa[i]));
424 SFSTAT_INC(sf_iocnt);
425 SFSTAT_ADD(sf_pages_read, count);
426 if (i + count == npages)
427 SFSTAT_ADD(sf_rhpages_read, rhpages);
430 * Restore the valid page pointers. They are already
431 * unbusied, but still wired.
433 for (j = i; j < i + count; j++)
434 if (pa[j] == bogus_page) {
435 pa[j] = vm_page_lookup(obj,
436 OFF_TO_IDX(vmoff(j, off)));
437 KASSERT(pa[j], ("%s: page %p[%d] disappeared",
445 VM_OBJECT_WUNLOCK(obj);
447 if (*nios == 0 && npages != 0)
448 SFSTAT_INC(sf_noiocnt);
454 sendfile_getobj(struct thread *td, struct file *fp, vm_object_t *obj_res,
455 struct vnode **vp_res, struct shmfd **shmfd_res, off_t *obj_size,
466 shmfd = *shmfd_res = NULL;
470 * The file descriptor must be a regular file and have a
473 if (fp->f_type == DTYPE_VNODE) {
475 vn_lock(vp, LK_SHARED | LK_RETRY);
476 if (vp->v_type != VREG) {
480 *bsize = vp->v_mount->mnt_stat.f_iosize;
481 error = VOP_GETATTR(vp, &va, td->td_ucred);
484 *obj_size = va.va_size;
490 } else if (fp->f_type == DTYPE_SHM) {
493 obj = shmfd->shm_object;
494 *obj_size = shmfd->shm_size;
500 VM_OBJECT_WLOCK(obj);
501 if ((obj->flags & OBJ_DEAD) != 0) {
502 VM_OBJECT_WUNLOCK(obj);
508 * Temporarily increase the backing VM object's reference
509 * count so that a forced reclamation of its vnode does not
510 * immediately destroy it.
512 vm_object_reference_locked(obj);
513 VM_OBJECT_WUNLOCK(obj);
525 sendfile_getsock(struct thread *td, int s, struct file **sock_fp,
534 * The socket must be a stream socket and connected.
536 error = getsock_cap(td, s, &cap_send_rights,
537 sock_fp, NULL, NULL);
540 *so = (*sock_fp)->f_data;
541 if ((*so)->so_type != SOCK_STREAM)
543 if (SOLISTENING(*so))
549 vn_sendfile(struct file *fp, int sockfd, struct uio *hdr_uio,
550 struct uio *trl_uio, off_t offset, size_t nbytes, off_t *sent, int flags,
553 struct file *sock_fp;
555 struct vm_object *obj;
557 struct mbuf *m, *mh, *mhtail;
560 struct sendfile_sync *sfs;
562 off_t off, sbytes, rem, obj_size;
563 int error, softerr, bsize, hdrlen;
572 error = sendfile_getobj(td, fp, &obj, &vp, &shmfd, &obj_size, &bsize);
576 error = sendfile_getsock(td, sockfd, &sock_fp, &so);
581 error = mac_socket_check_send(td->td_ucred, so);
586 SFSTAT_INC(sf_syscalls);
587 SFSTAT_ADD(sf_rhpages_requested, SF_READAHEAD(flags));
589 if (flags & SF_SYNC) {
590 sfs = malloc(sizeof *sfs, M_TEMP, M_WAITOK | M_ZERO);
591 mtx_init(&sfs->mtx, "sendfile", NULL, MTX_DEF);
592 cv_init(&sfs->cv, "sendfile");
595 rem = nbytes ? omin(nbytes, obj_size - offset) : obj_size - offset;
598 * Protect against multiple writers to the socket.
600 * XXXRW: Historically this has assumed non-interruptibility, so now
601 * we implement that, but possibly shouldn't.
603 (void)sblock(&so->so_snd, SBL_WAIT | SBL_NOINTR);
606 * Loop through the pages of the file, starting with the requested
607 * offset. Get a file page (do I/O if necessary), map the file page
608 * into an sf_buf, attach an mbuf header to the sf_buf, and queue
610 * This is done in two loops. The inner loop turns as many pages
611 * as it can, up to available socket buffer space, without blocking
612 * into mbufs to have it bulk delivered into the socket send buffer.
613 * The outer loop checks the state and available space of the socket
614 * and takes care of the overall progress.
616 for (off = offset; rem > 0; ) {
620 int nios, space, npages, rhpages;
624 * Check the socket state for ongoing connection,
625 * no errors and space in socket buffer.
626 * If space is low allow for the remainder of the
627 * file to be processed if it fits the socket buffer.
628 * Otherwise block in waiting for sufficient space
629 * to proceed, or if the socket is nonblocking, return
630 * to userland with EAGAIN while reporting how far
632 * We wait until the socket buffer has significant free
633 * space to do bulk sends. This makes good use of file
634 * system read ahead and allows packet segmentation
635 * offloading hardware to take over lots of work. If
636 * we were not careful here we would send off only one
639 SOCKBUF_LOCK(&so->so_snd);
640 if (so->so_snd.sb_lowat < so->so_snd.sb_hiwat / 2)
641 so->so_snd.sb_lowat = so->so_snd.sb_hiwat / 2;
643 if (so->so_snd.sb_state & SBS_CANTSENDMORE) {
645 SOCKBUF_UNLOCK(&so->so_snd);
647 } else if (so->so_error) {
648 error = so->so_error;
650 SOCKBUF_UNLOCK(&so->so_snd);
653 if ((so->so_state & SS_ISCONNECTED) == 0) {
654 SOCKBUF_UNLOCK(&so->so_snd);
659 space = sbspace(&so->so_snd);
662 space < so->so_snd.sb_lowat)) {
663 if (so->so_state & SS_NBIO) {
664 SOCKBUF_UNLOCK(&so->so_snd);
669 * sbwait drops the lock while sleeping.
670 * When we loop back to retry_space the
671 * state may have changed and we retest
674 error = sbwait(&so->so_snd);
676 * An error from sbwait usually indicates that we've
677 * been interrupted by a signal. If we've sent anything
678 * then return bytes sent, otherwise return the error.
681 SOCKBUF_UNLOCK(&so->so_snd);
686 SOCKBUF_UNLOCK(&so->so_snd);
689 * At the beginning of the first loop check if any headers
690 * are specified and copy them into mbufs. Reduce space in
691 * the socket buffer by the size of the header mbuf chain.
692 * Clear hdr_uio here and hdrlen at the end of the first loop.
694 if (hdr_uio != NULL && hdr_uio->uio_resid > 0) {
695 hdr_uio->uio_td = td;
696 hdr_uio->uio_rw = UIO_WRITE;
697 mh = m_uiotombuf(hdr_uio, M_WAITOK, space, 0, 0);
698 hdrlen = m_length(mh, &mhtail);
701 * If header consumed all the socket buffer space,
702 * don't waste CPU cycles and jump to the end.
713 error = vn_lock(vp, LK_SHARED);
716 error = VOP_GETATTR(vp, &va, td->td_ucred);
717 if (error != 0 || off >= va.va_size) {
721 if (va.va_size != obj_size) {
722 obj_size = va.va_size;
724 omin(nbytes + offset, obj_size) : obj_size;
732 npages = howmany(space + (off & PAGE_MASK), PAGE_SIZE);
735 * Calculate maximum allowed number of pages for readahead
736 * at this iteration. If SF_USER_READAHEAD was set, we don't
737 * do any heuristics and use exactly the value supplied by
738 * application. Otherwise, we allow readahead up to "rem".
739 * If application wants more, let it be, but there is no
740 * reason to go above MAXPHYS. Also check against "obj_size",
741 * since vm_pager_has_page() can hint beyond EOF.
743 if (flags & SF_USER_READAHEAD) {
744 rhpages = SF_READAHEAD(flags);
746 rhpages = howmany(rem + (off & PAGE_MASK), PAGE_SIZE) -
748 rhpages += SF_READAHEAD(flags);
750 rhpages = min(howmany(MAXPHYS, PAGE_SIZE), rhpages);
751 rhpages = min(howmany(obj_size - trunc_page(off), PAGE_SIZE) -
754 sfio = malloc(sizeof(struct sf_io) +
755 npages * sizeof(vm_page_t), M_TEMP, M_WAITOK);
756 refcount_init(&sfio->nios, 1);
760 error = sendfile_swapin(obj, sfio, &nios, off, space, npages,
770 * Loop and construct maximum sized mbuf chain to be bulk
771 * dumped into socket buffer.
774 for (int i = 0; i < npages; i++) {
778 * If a page wasn't grabbed successfully, then
779 * trim the array. Can happen only with SF_NODISKIO.
783 fixspace(npages, i, off, &space);
790 * Get a sendfile buf. When allocating the
791 * first buffer for mbuf chain, we usually
792 * wait as long as necessary, but this wait
793 * can be interrupted. For consequent
794 * buffers, do not sleep, since several
795 * threads might exhaust the buffers and then
798 sf = sf_buf_alloc(pa[i],
799 m != NULL ? SFB_NOWAIT : SFB_CATCH);
801 SFSTAT_INC(sf_allocfail);
802 for (int j = i; j < npages; j++) {
804 vm_page_unwire(pa[j], PQ_INACTIVE);
805 vm_page_unlock(pa[j]);
809 fixspace(npages, i, off, &space);
814 m0 = m_get(M_WAITOK, MT_DATA);
815 m0->m_ext.ext_buf = (char *)sf_buf_kva(sf);
816 m0->m_ext.ext_size = PAGE_SIZE;
817 m0->m_ext.ext_arg1 = sf;
818 m0->m_ext.ext_type = EXT_SFBUF;
819 m0->m_ext.ext_flags = EXT_FLAG_EMBREF;
820 m0->m_ext.ext_free = sendfile_free_mext;
822 * SF_NOCACHE sets the page as being freed upon send.
823 * However, we ignore it for the last page in 'space',
824 * if the page is truncated, and we got more data to
825 * send (rem > space), or if we have readahead
826 * configured (rhpages > 0).
828 if ((flags & SF_NOCACHE) &&
830 !((off + space) & PAGE_MASK) ||
831 !(rem > space || rhpages > 0)))
832 m0->m_ext.ext_flags |= EXT_FLAG_NOCACHE;
834 m0->m_ext.ext_flags |= EXT_FLAG_SYNC;
835 m0->m_ext.ext_arg2 = sfs;
838 mtx_unlock(&sfs->mtx);
840 m0->m_ext.ext_count = 1;
841 m0->m_flags |= (M_EXT | M_RDONLY);
843 m0->m_flags |= M_NOTREADY;
844 m0->m_data = (char *)sf_buf_kva(sf) +
845 (vmoff(i, off) & PAGE_MASK);
846 m0->m_len = xfsize(i, npages, off, space);
851 /* Append to mbuf chain. */
862 /* Keep track of bytes processed. */
866 /* Prepend header, if any. */
875 KASSERT(softerr, ("%s: m NULL, no error", __func__));
881 /* Add the buffer chain to the socket buffer. */
882 KASSERT(m_length(m, NULL) == space + hdrlen,
883 ("%s: mlen %u space %d hdrlen %d",
884 __func__, m_length(m, NULL), space, hdrlen));
886 CURVNET_SET(so->so_vnet);
889 * If sendfile_swapin() didn't initiate any I/Os,
890 * which happens if all data is cached in VM, then
891 * we can send data right now without the
892 * PRUS_NOTREADY flag.
895 error = (*so->so_proto->pr_usrreqs->pru_send)
896 (so, 0, m, NULL, NULL, td);
898 sfio->npages = npages;
900 error = (*so->so_proto->pr_usrreqs->pru_send)
901 (so, PRUS_NOTREADY, m, NULL, NULL, td);
902 sendfile_iodone(sfio, NULL, 0, 0);
906 m = NULL; /* pru_send always consumes */
909 sbytes += space + hdrlen;
919 * Send trailers. Wimp out and use writev(2).
921 if (trl_uio != NULL) {
922 sbunlock(&so->so_snd);
923 error = kern_writev(td, sockfd, trl_uio);
925 sbytes += td->td_retval[0];
930 sbunlock(&so->so_snd);
933 * If there was no error we have to clear td->td_retval[0]
934 * because it may have been set by writev.
937 td->td_retval[0] = 0;
943 vm_object_deallocate(obj);
954 cv_wait(&sfs->cv, &sfs->mtx);
955 KASSERT(sfs->count == 0, ("sendfile sync still busy"));
956 cv_destroy(&sfs->cv);
957 mtx_destroy(&sfs->mtx);
961 if (error == ERESTART)
968 sendfile(struct thread *td, struct sendfile_args *uap, int compat)
971 struct uio *hdr_uio, *trl_uio;
977 * File offset must be positive. If it goes beyond EOF
978 * we send only the header/trailer and no payload data.
984 hdr_uio = trl_uio = NULL;
986 if (uap->hdtr != NULL) {
987 error = copyin(uap->hdtr, &hdtr, sizeof(hdtr));
990 if (hdtr.headers != NULL) {
991 error = copyinuio(hdtr.headers, hdtr.hdr_cnt,
995 #ifdef COMPAT_FREEBSD4
997 * In FreeBSD < 5.0 the nbytes to send also included
998 * the header. If compat is specified subtract the
999 * header size from nbytes.
1002 if (uap->nbytes > hdr_uio->uio_resid)
1003 uap->nbytes -= hdr_uio->uio_resid;
1009 if (hdtr.trailers != NULL) {
1010 error = copyinuio(hdtr.trailers, hdtr.trl_cnt,
1017 AUDIT_ARG_FD(uap->fd);
1020 * sendfile(2) can start at any offset within a file so we require
1021 * CAP_READ+CAP_SEEK = CAP_PREAD.
1023 if ((error = fget_read(td, uap->fd, &cap_pread_rights, &fp)) != 0)
1026 error = fo_sendfile(fp, uap->s, hdr_uio, trl_uio, uap->offset,
1027 uap->nbytes, &sbytes, uap->flags, td);
1030 if (uap->sbytes != NULL)
1031 copyout(&sbytes, uap->sbytes, sizeof(off_t));
1034 free(hdr_uio, M_IOV);
1035 free(trl_uio, M_IOV);
1042 * int sendfile(int fd, int s, off_t offset, size_t nbytes,
1043 * struct sf_hdtr *hdtr, off_t *sbytes, int flags)
1045 * Send a file specified by 'fd' and starting at 'offset' to a socket
1046 * specified by 's'. Send only 'nbytes' of the file or until EOF if nbytes ==
1047 * 0. Optionally add a header and/or trailer to the socket output. If
1048 * specified, write the total number of bytes sent into *sbytes.
1051 sys_sendfile(struct thread *td, struct sendfile_args *uap)
1054 return (sendfile(td, uap, 0));
1057 #ifdef COMPAT_FREEBSD4
1059 freebsd4_sendfile(struct thread *td, struct freebsd4_sendfile_args *uap)
1061 struct sendfile_args args;
1065 args.offset = uap->offset;
1066 args.nbytes = uap->nbytes;
1067 args.hdtr = uap->hdtr;
1068 args.sbytes = uap->sbytes;
1069 args.flags = uap->flags;
1071 return (sendfile(td, &args, 1));
1073 #endif /* COMPAT_FREEBSD4 */