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_kern_tls.h"
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/capsicum.h>
38 #include <sys/kernel.h>
39 #include <netinet/in.h>
42 #include <sys/mutex.h>
43 #include <sys/sysproto.h>
44 #include <sys/malloc.h>
47 #include <sys/mount.h>
49 #include <sys/protosw.h>
50 #include <sys/rwlock.h>
51 #include <sys/sf_buf.h>
52 #include <sys/socket.h>
53 #include <sys/socketvar.h>
54 #include <sys/syscallsubr.h>
55 #include <sys/sysctl.h>
56 #include <sys/vnode.h>
59 #include <netinet/tcp.h>
61 #include <security/audit/audit.h>
62 #include <security/mac/mac_framework.h>
65 #include <vm/vm_object.h>
66 #include <vm/vm_pager.h>
68 #define EXT_FLAG_SYNC EXT_FLAG_VENDOR1
69 #define EXT_FLAG_NOCACHE EXT_FLAG_VENDOR2
70 #define EXT_FLAG_CACHE_LAST EXT_FLAG_VENDOR3
73 * Structure describing a single sendfile(2) I/O, which may consist of
74 * several underlying pager I/Os.
76 * The syscall context allocates the structure and initializes 'nios'
77 * to 1. As sendfile_swapin() runs through pages and starts asynchronous
78 * paging operations, it increments 'nios'.
80 * Every I/O completion calls sendfile_iodone(), which decrements the 'nios',
81 * and the syscall also calls sendfile_iodone() after allocating all mbufs,
82 * linking them and sending to socket. Whoever reaches zero 'nios' is
83 * responsible to * call pru_ready on the socket, to notify it of readyness
94 struct ktls_session *tls;
100 * Structure used to track requests with SF_SYNC flag.
102 struct sendfile_sync {
108 counter_u64_t sfstat[sizeof(struct sfstat) / sizeof(uint64_t)];
111 sfstat_init(const void *unused)
114 COUNTER_ARRAY_ALLOC(sfstat, sizeof(struct sfstat) / sizeof(uint64_t),
117 SYSINIT(sfstat, SI_SUB_MBUF, SI_ORDER_FIRST, sfstat_init, NULL);
120 sfstat_sysctl(SYSCTL_HANDLER_ARGS)
124 COUNTER_ARRAY_COPY(sfstat, &s, sizeof(s) / sizeof(uint64_t));
126 COUNTER_ARRAY_ZERO(sfstat, sizeof(s) / sizeof(uint64_t));
127 return (SYSCTL_OUT(req, &s, sizeof(s)));
129 SYSCTL_PROC(_kern_ipc, OID_AUTO, sfstat, CTLTYPE_OPAQUE | CTLFLAG_RW,
130 NULL, 0, sfstat_sysctl, "I", "sendfile statistics");
133 sendfile_free_mext(struct mbuf *m)
139 KASSERT(m->m_flags & M_EXT && m->m_ext.ext_type == EXT_SFBUF,
140 ("%s: m %p !M_EXT or !EXT_SFBUF", __func__, m));
142 sf = m->m_ext.ext_arg1;
143 pg = sf_buf_page(sf);
144 flags = (m->m_ext.ext_flags & EXT_FLAG_NOCACHE) != 0 ? VPR_TRYFREE : 0;
147 vm_page_release(pg, flags);
149 if (m->m_ext.ext_flags & EXT_FLAG_SYNC) {
150 struct sendfile_sync *sfs = m->m_ext.ext_arg2;
153 KASSERT(sfs->count > 0, ("Sendfile sync botchup count == 0"));
154 if (--sfs->count == 0)
156 mtx_unlock(&sfs->mtx);
161 sendfile_free_mext_pg(struct mbuf *m)
163 struct mbuf_ext_pgs *ext_pgs;
168 KASSERT(m->m_flags & M_EXT && m->m_ext.ext_type == EXT_PGS,
169 ("%s: m %p !M_EXT or !EXT_PGS", __func__, m));
171 cache_last = m->m_ext.ext_flags & EXT_FLAG_CACHE_LAST;
172 ext_pgs = m->m_ext.ext_pgs;
173 flags = (m->m_ext.ext_flags & EXT_FLAG_NOCACHE) != 0 ? VPR_TRYFREE : 0;
175 for (i = 0; i < ext_pgs->npgs; i++) {
176 if (cache_last && i == ext_pgs->npgs - 1)
178 pg = PHYS_TO_VM_PAGE(ext_pgs->pa[i]);
179 vm_page_release(pg, flags);
182 if (m->m_ext.ext_flags & EXT_FLAG_SYNC) {
183 struct sendfile_sync *sfs = m->m_ext.ext_arg2;
186 KASSERT(sfs->count > 0, ("Sendfile sync botchup count == 0"));
187 if (--sfs->count == 0)
189 mtx_unlock(&sfs->mtx);
194 * Helper function to calculate how much data to put into page i of n.
195 * Only first and last pages are special.
198 xfsize(int i, int n, off_t off, off_t len)
202 return (omin(PAGE_SIZE - (off & PAGE_MASK), len));
204 if (i == n - 1 && ((off + len) & PAGE_MASK) > 0)
205 return ((off + len) & PAGE_MASK);
211 * Helper function to get offset within object for i page.
213 static inline vm_ooffset_t
214 vmoff(int i, off_t off)
218 return ((vm_ooffset_t)off);
220 return (trunc_page(off + i * PAGE_SIZE));
224 * Helper function used when allocation of a page or sf_buf failed.
225 * Pretend as if we don't have enough space, subtract xfsize() of
226 * all pages that failed.
229 fixspace(int old, int new, off_t off, int *space)
232 KASSERT(old > new, ("%s: old %d new %d", __func__, old, new));
234 /* Subtract last one. */
235 *space -= xfsize(old - 1, old, off, *space);
239 /* There was only one page. */
242 /* Subtract first one. */
244 *space -= xfsize(0, old, off, *space);
248 /* Rest of pages are full sized. */
249 *space -= (old - new) * PAGE_SIZE;
251 KASSERT(*space >= 0, ("%s: space went backwards", __func__));
255 * I/O completion callback.
258 sendfile_iodone(void *arg, vm_page_t *pg, int count, int error)
260 struct sf_io *sfio = arg;
261 struct socket *so = sfio->so;
263 for (int i = 0; i < count; i++)
264 if (pg[i] != bogus_page)
265 vm_page_xunbusy_unchecked(pg[i]);
270 if (!refcount_release(&sfio->nios))
273 vm_object_pip_wakeup(sfio->obj);
275 if (__predict_false(sfio->error && sfio->m == NULL)) {
277 * I/O operation failed, but pru_send hadn't been executed -
278 * nothing had been sent to the socket. The syscall has
279 * returned error to the user.
285 #if defined(KERN_TLS) && defined(INVARIANTS)
286 if ((sfio->m->m_flags & M_EXT) != 0 &&
287 sfio->m->m_ext.ext_type == EXT_PGS)
288 KASSERT(sfio->tls == sfio->m->m_ext.ext_pgs->tls,
289 ("TLS session mismatch"));
291 KASSERT(sfio->tls == NULL,
292 ("non-ext_pgs mbuf with TLS session"));
294 CURVNET_SET(so->so_vnet);
295 if (__predict_false(sfio->error)) {
297 * I/O operation failed. The state of data in the socket
298 * is now inconsistent, and all what we can do is to tear
299 * it down. Protocol abort method would tear down protocol
300 * state, free all ready mbufs and detach not ready ones.
301 * We will free the mbufs corresponding to this I/O manually.
303 * The socket would be marked with EIO and made available
304 * for read, so that application receives EIO on next
305 * syscall and eventually closes the socket.
307 so->so_proto->pr_usrreqs->pru_abort(so);
310 mb_free_notready(sfio->m, sfio->npages);
312 } else if (sfio->tls != NULL && sfio->tls->mode == TCP_TLS_MODE_SW) {
314 * I/O operation is complete, but we still need to
315 * encrypt. We cannot do this in the interrupt thread
316 * of the disk controller, so forward the mbufs to a
319 * Donate the socket reference from sfio to rather
320 * than explicitly invoking soref().
322 ktls_enqueue(sfio->m, so, sfio->npages);
326 (void)(so->so_proto->pr_usrreqs->pru_ready)(so, sfio->m,
339 * Iterate through pages vector and request paging for non-valid pages.
342 sendfile_swapin(vm_object_t obj, struct sf_io *sfio, int *nios, off_t off,
343 off_t len, int npages, int rhpages, int flags)
345 vm_page_t *pa = sfio->pa;
349 flags = (flags & SF_NODISKIO) ? VM_ALLOC_NOWAIT : 0;
352 * First grab all the pages and wire them. Note that we grab
353 * only required pages. Readahead pages are dealt with later.
355 VM_OBJECT_WLOCK(obj);
357 grabbed = vm_page_grab_pages(obj, OFF_TO_IDX(off),
358 VM_ALLOC_NORMAL | VM_ALLOC_WIRED | flags, pa, npages);
359 if (grabbed < npages) {
360 for (int i = grabbed; i < npages; i++)
366 for (int i = 0; i < npages;) {
369 /* Skip valid pages. */
370 if (vm_page_is_valid(pa[i], vmoff(i, off) & PAGE_MASK,
371 xfsize(i, npages, off, len))) {
372 vm_page_xunbusy(pa[i]);
373 SFSTAT_INC(sf_pages_valid);
379 * Next page is invalid. Check if it belongs to pager. It
380 * may not be there, which is a regular situation for shmem
381 * pager. For vnode pager this happens only in case of
384 * Important feature of vm_pager_has_page() is the hint
385 * stored in 'a', about how many pages we can pagein after
386 * this page in a single I/O.
388 if (!vm_pager_has_page(obj, OFF_TO_IDX(vmoff(i, off)), NULL,
390 pmap_zero_page(pa[i]);
391 vm_page_valid(pa[i]);
392 MPASS(pa[i]->dirty == 0);
393 vm_page_xunbusy(pa[i]);
399 * We want to pagein as many pages as possible, limited only
400 * by the 'a' hint and actual request.
402 count = min(a + 1, npages - i);
405 * We should not pagein into a valid page, thus we first trim
406 * any valid pages off the end of request, and substitute
407 * to bogus_page those, that are in the middle.
409 for (j = i + count - 1; j > i; j--) {
410 if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
411 xfsize(j, npages, off, len))) {
417 for (j = i + 1; j < i + count - 1; j++)
418 if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
419 xfsize(j, npages, off, len))) {
420 vm_page_xunbusy(pa[j]);
421 SFSTAT_INC(sf_pages_valid);
422 SFSTAT_INC(sf_pages_bogus);
426 refcount_acquire(&sfio->nios);
427 VM_OBJECT_WUNLOCK(obj);
428 rv = vm_pager_get_pages_async(obj, pa + i, count, NULL,
429 i + count == npages ? &rhpages : NULL,
430 &sendfile_iodone, sfio);
431 VM_OBJECT_WLOCK(obj);
432 if (__predict_false(rv != VM_PAGER_OK)) {
434 * Perform full pages recovery before returning EIO.
435 * Pages from 0 to npages are wired.
436 * Pages from i to npages are also busied.
437 * Pages from (i + 1) to (i + count - 1) may be
438 * substituted to bogus page, and not busied.
440 for (j = 0; j < npages; j++) {
441 if (j > i && j < i + count - 1 &&
443 pa[j] = vm_page_lookup(obj,
444 OFF_TO_IDX(vmoff(j, off)));
446 vm_page_xunbusy(pa[j]);
447 KASSERT(pa[j] != NULL && pa[j] != bogus_page,
448 ("%s: page %p[%d] I/O recovery failure",
450 vm_page_unwire(pa[j], PQ_INACTIVE);
452 VM_OBJECT_WUNLOCK(obj);
456 SFSTAT_INC(sf_iocnt);
457 SFSTAT_ADD(sf_pages_read, count);
458 if (i + count == npages)
459 SFSTAT_ADD(sf_rhpages_read, rhpages);
462 * Restore the valid page pointers. They are already
463 * unbusied, but still wired.
465 for (j = i; j < i + count; j++)
466 if (pa[j] == bogus_page) {
467 pa[j] = vm_page_lookup(obj,
468 OFF_TO_IDX(vmoff(j, off)));
469 KASSERT(pa[j], ("%s: page %p[%d] disappeared",
477 VM_OBJECT_WUNLOCK(obj);
479 if (*nios == 0 && npages != 0)
480 SFSTAT_INC(sf_noiocnt);
486 sendfile_getobj(struct thread *td, struct file *fp, vm_object_t *obj_res,
487 struct vnode **vp_res, struct shmfd **shmfd_res, off_t *obj_size,
498 shmfd = *shmfd_res = NULL;
502 * The file descriptor must be a regular file and have a
505 if (fp->f_type == DTYPE_VNODE) {
507 vn_lock(vp, LK_SHARED | LK_RETRY);
508 if (vp->v_type != VREG) {
512 *bsize = vp->v_mount->mnt_stat.f_iosize;
513 error = VOP_GETATTR(vp, &va, td->td_ucred);
516 *obj_size = va.va_size;
522 } else if (fp->f_type == DTYPE_SHM) {
525 obj = shmfd->shm_object;
526 *obj_size = shmfd->shm_size;
532 VM_OBJECT_WLOCK(obj);
533 if ((obj->flags & OBJ_DEAD) != 0) {
534 VM_OBJECT_WUNLOCK(obj);
540 * Temporarily increase the backing VM object's reference
541 * count so that a forced reclamation of its vnode does not
542 * immediately destroy it.
544 vm_object_reference_locked(obj);
545 VM_OBJECT_WUNLOCK(obj);
557 sendfile_getsock(struct thread *td, int s, struct file **sock_fp,
566 * The socket must be a stream socket and connected.
568 error = getsock_cap(td, s, &cap_send_rights,
569 sock_fp, NULL, NULL);
572 *so = (*sock_fp)->f_data;
573 if ((*so)->so_type != SOCK_STREAM)
575 if (SOLISTENING(*so))
581 vn_sendfile(struct file *fp, int sockfd, struct uio *hdr_uio,
582 struct uio *trl_uio, off_t offset, size_t nbytes, off_t *sent, int flags,
585 struct file *sock_fp;
587 struct vm_object *obj;
590 struct ktls_session *tls;
592 struct mbuf_ext_pgs *ext_pgs;
593 struct mbuf *m, *mh, *mhtail;
596 struct sendfile_sync *sfs;
598 off_t off, sbytes, rem, obj_size;
599 int bsize, error, ext_pgs_idx, hdrlen, max_pgs, softerr;
616 error = sendfile_getobj(td, fp, &obj, &vp, &shmfd, &obj_size, &bsize);
620 error = sendfile_getsock(td, sockfd, &sock_fp, &so);
625 error = mac_socket_check_send(td->td_ucred, so);
630 SFSTAT_INC(sf_syscalls);
631 SFSTAT_ADD(sf_rhpages_requested, SF_READAHEAD(flags));
633 if (flags & SF_SYNC) {
634 sfs = malloc(sizeof *sfs, M_TEMP, M_WAITOK | M_ZERO);
635 mtx_init(&sfs->mtx, "sendfile", NULL, MTX_DEF);
636 cv_init(&sfs->cv, "sendfile");
639 rem = nbytes ? omin(nbytes, obj_size - offset) : obj_size - offset;
642 * Protect against multiple writers to the socket.
644 * XXXRW: Historically this has assumed non-interruptibility, so now
645 * we implement that, but possibly shouldn't.
647 (void)sblock(&so->so_snd, SBL_WAIT | SBL_NOINTR);
649 tls = ktls_hold(so->so_snd.sb_tls_info);
653 * Loop through the pages of the file, starting with the requested
654 * offset. Get a file page (do I/O if necessary), map the file page
655 * into an sf_buf, attach an mbuf header to the sf_buf, and queue
657 * This is done in two loops. The inner loop turns as many pages
658 * as it can, up to available socket buffer space, without blocking
659 * into mbufs to have it bulk delivered into the socket send buffer.
660 * The outer loop checks the state and available space of the socket
661 * and takes care of the overall progress.
663 for (off = offset; rem > 0; ) {
667 int nios, space, npages, rhpages;
671 * Check the socket state for ongoing connection,
672 * no errors and space in socket buffer.
673 * If space is low allow for the remainder of the
674 * file to be processed if it fits the socket buffer.
675 * Otherwise block in waiting for sufficient space
676 * to proceed, or if the socket is nonblocking, return
677 * to userland with EAGAIN while reporting how far
679 * We wait until the socket buffer has significant free
680 * space to do bulk sends. This makes good use of file
681 * system read ahead and allows packet segmentation
682 * offloading hardware to take over lots of work. If
683 * we were not careful here we would send off only one
686 SOCKBUF_LOCK(&so->so_snd);
687 if (so->so_snd.sb_lowat < so->so_snd.sb_hiwat / 2)
688 so->so_snd.sb_lowat = so->so_snd.sb_hiwat / 2;
690 if (so->so_snd.sb_state & SBS_CANTSENDMORE) {
692 SOCKBUF_UNLOCK(&so->so_snd);
694 } else if (so->so_error) {
695 error = so->so_error;
697 SOCKBUF_UNLOCK(&so->so_snd);
700 if ((so->so_state & SS_ISCONNECTED) == 0) {
701 SOCKBUF_UNLOCK(&so->so_snd);
706 space = sbspace(&so->so_snd);
709 space < so->so_snd.sb_lowat)) {
710 if (so->so_state & SS_NBIO) {
711 SOCKBUF_UNLOCK(&so->so_snd);
716 * sbwait drops the lock while sleeping.
717 * When we loop back to retry_space the
718 * state may have changed and we retest
721 error = sbwait(&so->so_snd);
723 * An error from sbwait usually indicates that we've
724 * been interrupted by a signal. If we've sent anything
725 * then return bytes sent, otherwise return the error.
728 SOCKBUF_UNLOCK(&so->so_snd);
733 SOCKBUF_UNLOCK(&so->so_snd);
736 * At the beginning of the first loop check if any headers
737 * are specified and copy them into mbufs. Reduce space in
738 * the socket buffer by the size of the header mbuf chain.
739 * Clear hdr_uio here and hdrlen at the end of the first loop.
741 if (hdr_uio != NULL && hdr_uio->uio_resid > 0) {
742 hdr_uio->uio_td = td;
743 hdr_uio->uio_rw = UIO_WRITE;
746 mh = m_uiotombuf(hdr_uio, M_WAITOK, space,
747 tls->params.max_frame_len, M_NOMAP);
750 mh = m_uiotombuf(hdr_uio, M_WAITOK,
752 hdrlen = m_length(mh, &mhtail);
755 * If header consumed all the socket buffer space,
756 * don't waste CPU cycles and jump to the end.
767 error = vn_lock(vp, LK_SHARED);
770 error = VOP_GETATTR(vp, &va, td->td_ucred);
771 if (error != 0 || off >= va.va_size) {
775 if (va.va_size != obj_size) {
776 obj_size = va.va_size;
778 omin(nbytes + offset, obj_size) : obj_size;
785 else if (space > PAGE_SIZE) {
787 * Use page boundaries when possible for large
791 space -= (PAGE_SIZE - (off & PAGE_MASK));
792 space = trunc_page(space);
794 space += (PAGE_SIZE - (off & PAGE_MASK));
797 npages = howmany(space + (off & PAGE_MASK), PAGE_SIZE);
800 * Calculate maximum allowed number of pages for readahead
801 * at this iteration. If SF_USER_READAHEAD was set, we don't
802 * do any heuristics and use exactly the value supplied by
803 * application. Otherwise, we allow readahead up to "rem".
804 * If application wants more, let it be, but there is no
805 * reason to go above MAXPHYS. Also check against "obj_size",
806 * since vm_pager_has_page() can hint beyond EOF.
808 if (flags & SF_USER_READAHEAD) {
809 rhpages = SF_READAHEAD(flags);
811 rhpages = howmany(rem + (off & PAGE_MASK), PAGE_SIZE) -
813 rhpages += SF_READAHEAD(flags);
815 rhpages = min(howmany(MAXPHYS, PAGE_SIZE), rhpages);
816 rhpages = min(howmany(obj_size - trunc_page(off), PAGE_SIZE) -
819 sfio = malloc(sizeof(struct sf_io) +
820 npages * sizeof(vm_page_t), M_TEMP, M_WAITOK);
821 refcount_init(&sfio->nios, 1);
825 vm_object_pip_add(obj, 1);
829 * This doesn't use ktls_hold() because sfio->m will
830 * also have a reference on 'tls' that will be valid
831 * for all of sfio's lifetime.
836 error = sendfile_swapin(obj, sfio, &nios, off, space, npages,
842 sendfile_iodone(sfio, NULL, 0, error);
847 * Loop and construct maximum sized mbuf chain to be bulk
848 * dumped into socket buffer.
853 * Use unmapped mbufs if enabled for TCP. Unmapped
854 * bufs are restricted to TCP as that is what has been
855 * tested. In particular, unmapped mbufs have not
856 * been tested with UNIX-domain sockets.
858 * TLS frames always require unmapped mbufs.
860 if ((mb_use_ext_pgs &&
861 so->so_proto->pr_protocol == IPPROTO_TCP)
869 max_pgs = num_pages(tls->params.max_frame_len);
872 max_pgs = MBUF_PEXT_MAX_PGS;
874 /* Start at last index, to wrap on first use. */
875 ext_pgs_idx = max_pgs - 1;
878 for (int i = 0; i < npages; i++) {
882 * If a page wasn't grabbed successfully, then
883 * trim the array. Can happen only with SF_NODISKIO.
887 fixspace(npages, i, off, &space);
897 if (ext_pgs_idx == max_pgs) {
898 m0 = mb_alloc_ext_pgs(M_WAITOK, false,
899 sendfile_free_mext_pg);
901 if (flags & SF_NOCACHE) {
902 m0->m_ext.ext_flags |=
906 * See comment below regarding
907 * ignoring SF_NOCACHE for the
910 if ((npages - i <= max_pgs) &&
911 ((off + space) & PAGE_MASK) &&
912 (rem > space || rhpages > 0))
913 m0->m_ext.ext_flags |=
917 m0->m_ext.ext_flags |=
919 m0->m_ext.ext_arg2 = sfs;
922 mtx_unlock(&sfs->mtx);
924 ext_pgs = m0->m_ext.ext_pgs;
929 /* Append to mbuf chain. */
935 ext_pgs->first_pg_off =
936 vmoff(i, off) & PAGE_MASK;
939 mtail->m_flags |= M_NOTREADY;
943 ext_pgs->pa[ext_pgs_idx] = VM_PAGE_TO_PHYS(pa[i]);
945 xfs = xfsize(i, npages, off, space);
946 ext_pgs->last_pg_len = xfs;
947 MBUF_EXT_PGS_ASSERT_SANITY(ext_pgs);
949 mtail->m_ext.ext_size += PAGE_SIZE;
954 * Get a sendfile buf. When allocating the
955 * first buffer for mbuf chain, we usually
956 * wait as long as necessary, but this wait
957 * can be interrupted. For consequent
958 * buffers, do not sleep, since several
959 * threads might exhaust the buffers and then
962 sf = sf_buf_alloc(pa[i],
963 m != NULL ? SFB_NOWAIT : SFB_CATCH);
965 SFSTAT_INC(sf_allocfail);
966 for (int j = i; j < npages; j++)
967 vm_page_unwire(pa[j], PQ_INACTIVE);
970 fixspace(npages, i, off, &space);
975 m0 = m_get(M_WAITOK, MT_DATA);
976 m0->m_ext.ext_buf = (char *)sf_buf_kva(sf);
977 m0->m_ext.ext_size = PAGE_SIZE;
978 m0->m_ext.ext_arg1 = sf;
979 m0->m_ext.ext_type = EXT_SFBUF;
980 m0->m_ext.ext_flags = EXT_FLAG_EMBREF;
981 m0->m_ext.ext_free = sendfile_free_mext;
983 * SF_NOCACHE sets the page as being freed upon send.
984 * However, we ignore it for the last page in 'space',
985 * if the page is truncated, and we got more data to
986 * send (rem > space), or if we have readahead
987 * configured (rhpages > 0).
989 if ((flags & SF_NOCACHE) &&
991 !((off + space) & PAGE_MASK) ||
992 !(rem > space || rhpages > 0)))
993 m0->m_ext.ext_flags |= EXT_FLAG_NOCACHE;
995 m0->m_ext.ext_flags |= EXT_FLAG_SYNC;
996 m0->m_ext.ext_arg2 = sfs;
999 mtx_unlock(&sfs->mtx);
1001 m0->m_ext.ext_count = 1;
1002 m0->m_flags |= (M_EXT | M_RDONLY);
1004 m0->m_flags |= M_NOTREADY;
1005 m0->m_data = (char *)sf_buf_kva(sf) +
1006 (vmoff(i, off) & PAGE_MASK);
1007 m0->m_len = xfsize(i, npages, off, space);
1012 /* Append to mbuf chain. */
1023 /* Keep track of bytes processed. */
1027 /* Prepend header, if any. */
1036 KASSERT(softerr, ("%s: m NULL, no error", __func__));
1042 /* Add the buffer chain to the socket buffer. */
1043 KASSERT(m_length(m, NULL) == space + hdrlen,
1044 ("%s: mlen %u space %d hdrlen %d",
1045 __func__, m_length(m, NULL), space, hdrlen));
1047 CURVNET_SET(so->so_vnet);
1050 error = ktls_frame(m, tls, &tls_enq_cnt,
1058 * If sendfile_swapin() didn't initiate any I/Os,
1059 * which happens if all data is cached in VM, then
1060 * we can send data right now without the
1061 * PRUS_NOTREADY flag.
1064 vm_object_pip_wakeup(sfio->obj);
1068 if (tls != NULL && tls->mode == TCP_TLS_MODE_SW) {
1069 error = (*so->so_proto->pr_usrreqs->pru_send)
1070 (so, PRUS_NOTREADY, m, NULL, NULL, td);
1072 ktls_enqueue(m, so, tls_enq_cnt);
1075 error = (*so->so_proto->pr_usrreqs->pru_send)
1076 (so, 0, m, NULL, NULL, td);
1078 sfio->npages = npages;
1080 error = (*so->so_proto->pr_usrreqs->pru_send)
1081 (so, PRUS_NOTREADY, m, NULL, NULL, td);
1082 sendfile_iodone(sfio, NULL, 0, 0);
1086 m = NULL; /* pru_send always consumes */
1089 sbytes += space + hdrlen;
1099 * Send trailers. Wimp out and use writev(2).
1101 if (trl_uio != NULL) {
1102 sbunlock(&so->so_snd);
1103 error = kern_writev(td, sockfd, trl_uio);
1105 sbytes += td->td_retval[0];
1110 sbunlock(&so->so_snd);
1113 * If there was no error we have to clear td->td_retval[0]
1114 * because it may have been set by writev.
1117 td->td_retval[0] = 0;
1123 vm_object_deallocate(obj);
1132 mtx_lock(&sfs->mtx);
1133 if (sfs->count != 0)
1134 cv_wait(&sfs->cv, &sfs->mtx);
1135 KASSERT(sfs->count == 0, ("sendfile sync still busy"));
1136 cv_destroy(&sfs->cv);
1137 mtx_destroy(&sfs->mtx);
1145 if (error == ERESTART)
1152 sendfile(struct thread *td, struct sendfile_args *uap, int compat)
1154 struct sf_hdtr hdtr;
1155 struct uio *hdr_uio, *trl_uio;
1161 * File offset must be positive. If it goes beyond EOF
1162 * we send only the header/trailer and no payload data.
1164 if (uap->offset < 0)
1168 hdr_uio = trl_uio = NULL;
1170 if (uap->hdtr != NULL) {
1171 error = copyin(uap->hdtr, &hdtr, sizeof(hdtr));
1174 if (hdtr.headers != NULL) {
1175 error = copyinuio(hdtr.headers, hdtr.hdr_cnt,
1179 #ifdef COMPAT_FREEBSD4
1181 * In FreeBSD < 5.0 the nbytes to send also included
1182 * the header. If compat is specified subtract the
1183 * header size from nbytes.
1186 if (uap->nbytes > hdr_uio->uio_resid)
1187 uap->nbytes -= hdr_uio->uio_resid;
1193 if (hdtr.trailers != NULL) {
1194 error = copyinuio(hdtr.trailers, hdtr.trl_cnt,
1201 AUDIT_ARG_FD(uap->fd);
1204 * sendfile(2) can start at any offset within a file so we require
1205 * CAP_READ+CAP_SEEK = CAP_PREAD.
1207 if ((error = fget_read(td, uap->fd, &cap_pread_rights, &fp)) != 0)
1210 error = fo_sendfile(fp, uap->s, hdr_uio, trl_uio, uap->offset,
1211 uap->nbytes, &sbytes, uap->flags, td);
1214 if (uap->sbytes != NULL)
1215 copyout(&sbytes, uap->sbytes, sizeof(off_t));
1218 free(hdr_uio, M_IOV);
1219 free(trl_uio, M_IOV);
1226 * int sendfile(int fd, int s, off_t offset, size_t nbytes,
1227 * struct sf_hdtr *hdtr, off_t *sbytes, int flags)
1229 * Send a file specified by 'fd' and starting at 'offset' to a socket
1230 * specified by 's'. Send only 'nbytes' of the file or until EOF if nbytes ==
1231 * 0. Optionally add a header and/or trailer to the socket output. If
1232 * specified, write the total number of bytes sent into *sbytes.
1235 sys_sendfile(struct thread *td, struct sendfile_args *uap)
1238 return (sendfile(td, uap, 0));
1241 #ifdef COMPAT_FREEBSD4
1243 freebsd4_sendfile(struct thread *td, struct freebsd4_sendfile_args *uap)
1245 struct sendfile_args args;
1249 args.offset = uap->offset;
1250 args.nbytes = uap->nbytes;
1251 args.hdtr = uap->hdtr;
1252 args.sbytes = uap->sbytes;
1253 args.flags = uap->flags;
1255 return (sendfile(td, &args, 1));
1257 #endif /* COMPAT_FREEBSD4 */