xref: /freebsd/sys/kern/kern_sendfile.c (revision a70cba95822f662d3f9da5119b6a0c433e8f70af)
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  * 4. Neither the name of the University nor the names of its contributors
14  *    may be used to endorse or promote products derived from this software
15  *    without specific prior written permission.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
32 
33 #include "opt_compat.h"
34 
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/capsicum.h>
38 #include <sys/kernel.h>
39 #include <sys/lock.h>
40 #include <sys/mutex.h>
41 #include <sys/sysproto.h>
42 #include <sys/malloc.h>
43 #include <sys/proc.h>
44 #include <sys/mman.h>
45 #include <sys/mount.h>
46 #include <sys/mbuf.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>
55 
56 #include <net/vnet.h>
57 
58 #include <security/audit/audit.h>
59 #include <security/mac/mac_framework.h>
60 
61 #include <vm/vm.h>
62 #include <vm/vm_object.h>
63 #include <vm/vm_pager.h>
64 
65 /*
66  * Structure describing a single sendfile(2) I/O, which may consist of
67  * several underlying pager I/Os.
68  *
69  * The syscall context allocates the structure and initializes 'nios'
70  * to 1.  As sendfile_swapin() runs through pages and starts asynchronous
71  * paging operations, it increments 'nios'.
72  *
73  * Every I/O completion calls sendfile_iodone(), which decrements the 'nios',
74  * and the syscall also calls sendfile_iodone() after allocating all mbufs,
75  * linking them and sending to socket.  Whoever reaches zero 'nios' is
76  * responsible to * call pru_ready on the socket, to notify it of readyness
77  * of the data.
78  */
79 struct sf_io {
80 	volatile u_int	nios;
81 	u_int		error;
82 	int		npages;
83 	struct file	*sock_fp;
84 	struct mbuf	*m;
85 	vm_page_t	pa[];
86 };
87 
88 /*
89  * Structure used to track requests with SF_SYNC flag.
90  */
91 struct sendfile_sync {
92 	struct mtx	mtx;
93 	struct cv	cv;
94 	unsigned	count;
95 };
96 
97 counter_u64_t sfstat[sizeof(struct sfstat) / sizeof(uint64_t)];
98 
99 static void
100 sfstat_init(const void *unused)
101 {
102 
103 	COUNTER_ARRAY_ALLOC(sfstat, sizeof(struct sfstat) / sizeof(uint64_t),
104 	    M_WAITOK);
105 }
106 SYSINIT(sfstat, SI_SUB_MBUF, SI_ORDER_FIRST, sfstat_init, NULL);
107 
108 static int
109 sfstat_sysctl(SYSCTL_HANDLER_ARGS)
110 {
111 	struct sfstat s;
112 
113 	COUNTER_ARRAY_COPY(sfstat, &s, sizeof(s) / sizeof(uint64_t));
114 	if (req->newptr)
115 		COUNTER_ARRAY_ZERO(sfstat, sizeof(s) / sizeof(uint64_t));
116 	return (SYSCTL_OUT(req, &s, sizeof(s)));
117 }
118 SYSCTL_PROC(_kern_ipc, OID_AUTO, sfstat, CTLTYPE_OPAQUE | CTLFLAG_RW,
119     NULL, 0, sfstat_sysctl, "I", "sendfile statistics");
120 
121 /*
122  * Add more references to a vm_page + sf_buf + sendfile_sync.  Called
123  * by mbuf(9) code to add extra references to a page.
124  */
125 void
126 sf_ext_ref(void *arg1, void *arg2)
127 {
128 	struct sf_buf *sf = arg1;
129 	struct sendfile_sync *sfs = arg2;
130 	vm_page_t pg = sf_buf_page(sf);
131 
132 	sf_buf_ref(sf);
133 
134 	vm_page_lock(pg);
135 	vm_page_wire(pg);
136 	vm_page_unlock(pg);
137 
138 	if (sfs != NULL) {
139 		mtx_lock(&sfs->mtx);
140 		KASSERT(sfs->count > 0, ("Sendfile sync botchup count == 0"));
141 		sfs->count++;
142 		mtx_unlock(&sfs->mtx);
143 	}
144 }
145 
146 /*
147  * Detach mapped page and release resources back to the system.  Called
148  * by mbuf(9) code when last reference to a page is freed.
149  */
150 void
151 sf_ext_free(void *arg1, void *arg2)
152 {
153 	struct sf_buf *sf = arg1;
154 	struct sendfile_sync *sfs = arg2;
155 	vm_page_t pg = sf_buf_page(sf);
156 
157 	sf_buf_free(sf);
158 
159 	vm_page_lock(pg);
160 	/*
161 	 * Check for the object going away on us. This can
162 	 * happen since we don't hold a reference to it.
163 	 * If so, we're responsible for freeing the page.
164 	 */
165 	if (vm_page_unwire(pg, PQ_INACTIVE) && pg->object == NULL)
166 		vm_page_free(pg);
167 	vm_page_unlock(pg);
168 
169 	if (sfs != NULL) {
170 		mtx_lock(&sfs->mtx);
171 		KASSERT(sfs->count > 0, ("Sendfile sync botchup count == 0"));
172 		if (--sfs->count == 0)
173 			cv_signal(&sfs->cv);
174 		mtx_unlock(&sfs->mtx);
175 	}
176 }
177 
178 /*
179  * Same as above, but forces the page to be detached from the object
180  * and go into free pool.
181  */
182 void
183 sf_ext_free_nocache(void *arg1, void *arg2)
184 {
185 	struct sf_buf *sf = arg1;
186 	struct sendfile_sync *sfs = arg2;
187 	vm_page_t pg = sf_buf_page(sf);
188 
189 	sf_buf_free(sf);
190 
191 	vm_page_lock(pg);
192 	if (vm_page_unwire(pg, PQ_NONE)) {
193 		vm_object_t obj;
194 
195 		/* Try to free the page, but only if it is cheap to. */
196 		if ((obj = pg->object) == NULL)
197 			vm_page_free(pg);
198 		else if (!vm_page_xbusied(pg) && VM_OBJECT_TRYWLOCK(obj)) {
199 			vm_page_free(pg);
200 			VM_OBJECT_WUNLOCK(obj);
201 		} else
202 			vm_page_deactivate(pg);
203 	}
204 	vm_page_unlock(pg);
205 
206 	if (sfs != NULL) {
207 		mtx_lock(&sfs->mtx);
208 		KASSERT(sfs->count > 0, ("Sendfile sync botchup count == 0"));
209 		if (--sfs->count == 0)
210 			cv_signal(&sfs->cv);
211 		mtx_unlock(&sfs->mtx);
212 	}
213 }
214 
215 /*
216  * Helper function to calculate how much data to put into page i of n.
217  * Only first and last pages are special.
218  */
219 static inline off_t
220 xfsize(int i, int n, off_t off, off_t len)
221 {
222 
223 	if (i == 0)
224 		return (omin(PAGE_SIZE - (off & PAGE_MASK), len));
225 
226 	if (i == n - 1 && ((off + len) & PAGE_MASK) > 0)
227 		return ((off + len) & PAGE_MASK);
228 
229 	return (PAGE_SIZE);
230 }
231 
232 /*
233  * Helper function to get offset within object for i page.
234  */
235 static inline vm_offset_t
236 vmoff(int i, off_t off)
237 {
238 
239 	if (i == 0)
240 		return ((vm_offset_t)off);
241 
242 	return (trunc_page(off + i * PAGE_SIZE));
243 }
244 
245 /*
246  * Helper function used when allocation of a page or sf_buf failed.
247  * Pretend as if we don't have enough space, subtract xfsize() of
248  * all pages that failed.
249  */
250 static inline void
251 fixspace(int old, int new, off_t off, int *space)
252 {
253 
254 	KASSERT(old > new, ("%s: old %d new %d", __func__, old, new));
255 
256 	/* Subtract last one. */
257 	*space -= xfsize(old - 1, old, off, *space);
258 	old--;
259 
260 	if (new == old)
261 		/* There was only one page. */
262 		return;
263 
264 	/* Subtract first one. */
265 	if (new == 0) {
266 		*space -= xfsize(0, old, off, *space);
267 		new++;
268 	}
269 
270 	/* Rest of pages are full sized. */
271 	*space -= (old - new) * PAGE_SIZE;
272 
273 	KASSERT(*space >= 0, ("%s: space went backwards", __func__));
274 }
275 
276 /*
277  * I/O completion callback.
278  */
279 static void
280 sendfile_iodone(void *arg, vm_page_t *pg, int count, int error)
281 {
282 	struct sf_io *sfio = arg;
283 	struct socket *so;
284 
285 	for (int i = 0; i < count; i++)
286 		vm_page_xunbusy(pg[i]);
287 
288 	if (error)
289 		sfio->error = error;
290 
291 	if (!refcount_release(&sfio->nios))
292 		return;
293 
294 	so = sfio->sock_fp->f_data;
295 
296 	if (sfio->error) {
297 		struct mbuf *m;
298 
299 		/*
300 		 * I/O operation failed.  The state of data in the socket
301 		 * is now inconsistent, and all what we can do is to tear
302 		 * it down. Protocol abort method would tear down protocol
303 		 * state, free all ready mbufs and detach not ready ones.
304 		 * We will free the mbufs corresponding to this I/O manually.
305 		 *
306 		 * The socket would be marked with EIO and made available
307 		 * for read, so that application receives EIO on next
308 		 * syscall and eventually closes the socket.
309 		 */
310 		so->so_proto->pr_usrreqs->pru_abort(so);
311 		so->so_error = EIO;
312 
313 		m = sfio->m;
314 		for (int i = 0; i < sfio->npages; i++)
315 			m = m_free(m);
316 	} else {
317 		CURVNET_SET(so->so_vnet);
318 		(void )(so->so_proto->pr_usrreqs->pru_ready)(so, sfio->m,
319 		    sfio->npages);
320 		CURVNET_RESTORE();
321 	}
322 
323 	/* XXXGL: curthread */
324 	fdrop(sfio->sock_fp, curthread);
325 	free(sfio, M_TEMP);
326 }
327 
328 /*
329  * Iterate through pages vector and request paging for non-valid pages.
330  */
331 static int
332 sendfile_swapin(vm_object_t obj, struct sf_io *sfio, off_t off, off_t len,
333     int npages, int rhpages, int flags)
334 {
335 	vm_page_t *pa = sfio->pa;
336 	int nios;
337 
338 	nios = 0;
339 	flags = (flags & SF_NODISKIO) ? VM_ALLOC_NOWAIT : 0;
340 
341 	/*
342 	 * First grab all the pages and wire them.  Note that we grab
343 	 * only required pages.  Readahead pages are dealt with later.
344 	 */
345 	VM_OBJECT_WLOCK(obj);
346 	for (int i = 0; i < npages; i++) {
347 		pa[i] = vm_page_grab(obj, OFF_TO_IDX(vmoff(i, off)),
348 		    VM_ALLOC_WIRED | VM_ALLOC_NORMAL | flags);
349 		if (pa[i] == NULL) {
350 			npages = i;
351 			rhpages = 0;
352 			break;
353 		}
354 	}
355 
356 	for (int i = 0; i < npages;) {
357 		int j, a, count, rv;
358 
359 		/* Skip valid pages. */
360 		if (vm_page_is_valid(pa[i], vmoff(i, off) & PAGE_MASK,
361 		    xfsize(i, npages, off, len))) {
362 			vm_page_xunbusy(pa[i]);
363 			SFSTAT_INC(sf_pages_valid);
364 			i++;
365 			continue;
366 		}
367 
368 		/*
369 		 * Now 'i' points to first invalid page, iterate further
370 		 * to make 'j' point at first valid after a bunch of
371 		 * invalid ones.
372 		 */
373 		for (j = i + 1; j < npages; j++)
374 			if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
375 			    xfsize(j, npages, off, len))) {
376 				SFSTAT_INC(sf_pages_valid);
377 				break;
378 			}
379 
380 		/*
381 		 * Now we got region of invalid pages between 'i' and 'j'.
382 		 * Check that they belong to pager.  They may not be there,
383 		 * which is a regular situation for shmem pager.  For vnode
384 		 * pager this happens only in case of sparse file.
385 		 *
386 		 * Important feature of vm_pager_has_page() is the hint
387 		 * stored in 'a', about how many pages we can pagein after
388 		 * this page in a single I/O.
389 		 */
390 		while (!vm_pager_has_page(obj, OFF_TO_IDX(vmoff(i, off)),
391 		    NULL, &a) && i < j) {
392 			pmap_zero_page(pa[i]);
393 			pa[i]->valid = VM_PAGE_BITS_ALL;
394 			pa[i]->dirty = 0;
395 			vm_page_xunbusy(pa[i]);
396 			i++;
397 		}
398 		if (i == j)
399 			continue;
400 
401 		/*
402 		 * We want to pagein as many pages as possible, limited only
403 		 * by the 'a' hint and actual request.
404 		 *
405 		 * We should not pagein into already valid page, thus if
406 		 * 'j' didn't reach last page, trim by that page.
407 		 *
408 		 * When the pagein fulfils the request, also specify readahead.
409 		 */
410 		if (j < npages)
411 			a = min(a, j - i - 1);
412 		count = min(a + 1, npages - i);
413 
414 		refcount_acquire(&sfio->nios);
415 		rv = vm_pager_get_pages_async(obj, pa + i, count, NULL,
416 		    i + count == npages ? &rhpages : NULL,
417 		    &sendfile_iodone, sfio);
418 		KASSERT(rv == VM_PAGER_OK, ("%s: pager fail obj %p page %p",
419 		    __func__, obj, pa[i]));
420 
421 		SFSTAT_INC(sf_iocnt);
422 		SFSTAT_ADD(sf_pages_read, count);
423 		if (i + count == npages)
424 			SFSTAT_ADD(sf_rhpages_read, rhpages);
425 
426 #ifdef INVARIANTS
427 		for (j = i; j < i + count && j < npages; j++)
428 			KASSERT(pa[j] == vm_page_lookup(obj,
429 			    OFF_TO_IDX(vmoff(j, off))),
430 			    ("pa[j] %p lookup %p\n", pa[j],
431 			    vm_page_lookup(obj, OFF_TO_IDX(vmoff(j, off)))));
432 #endif
433 		i += count;
434 		nios++;
435 	}
436 
437 	VM_OBJECT_WUNLOCK(obj);
438 
439 	if (nios == 0 && npages != 0)
440 		SFSTAT_INC(sf_noiocnt);
441 
442 	return (nios);
443 }
444 
445 static int
446 sendfile_getobj(struct thread *td, struct file *fp, vm_object_t *obj_res,
447     struct vnode **vp_res, struct shmfd **shmfd_res, off_t *obj_size,
448     int *bsize)
449 {
450 	struct vattr va;
451 	vm_object_t obj;
452 	struct vnode *vp;
453 	struct shmfd *shmfd;
454 	int error;
455 
456 	vp = *vp_res = NULL;
457 	obj = NULL;
458 	shmfd = *shmfd_res = NULL;
459 	*bsize = 0;
460 
461 	/*
462 	 * The file descriptor must be a regular file and have a
463 	 * backing VM object.
464 	 */
465 	if (fp->f_type == DTYPE_VNODE) {
466 		vp = fp->f_vnode;
467 		vn_lock(vp, LK_SHARED | LK_RETRY);
468 		if (vp->v_type != VREG) {
469 			error = EINVAL;
470 			goto out;
471 		}
472 		*bsize = vp->v_mount->mnt_stat.f_iosize;
473 		error = VOP_GETATTR(vp, &va, td->td_ucred);
474 		if (error != 0)
475 			goto out;
476 		*obj_size = va.va_size;
477 		obj = vp->v_object;
478 		if (obj == NULL) {
479 			error = EINVAL;
480 			goto out;
481 		}
482 	} else if (fp->f_type == DTYPE_SHM) {
483 		error = 0;
484 		shmfd = fp->f_data;
485 		obj = shmfd->shm_object;
486 		*obj_size = shmfd->shm_size;
487 	} else {
488 		error = EINVAL;
489 		goto out;
490 	}
491 
492 	VM_OBJECT_WLOCK(obj);
493 	if ((obj->flags & OBJ_DEAD) != 0) {
494 		VM_OBJECT_WUNLOCK(obj);
495 		error = EBADF;
496 		goto out;
497 	}
498 
499 	/*
500 	 * Temporarily increase the backing VM object's reference
501 	 * count so that a forced reclamation of its vnode does not
502 	 * immediately destroy it.
503 	 */
504 	vm_object_reference_locked(obj);
505 	VM_OBJECT_WUNLOCK(obj);
506 	*obj_res = obj;
507 	*vp_res = vp;
508 	*shmfd_res = shmfd;
509 
510 out:
511 	if (vp != NULL)
512 		VOP_UNLOCK(vp, 0);
513 	return (error);
514 }
515 
516 static int
517 sendfile_getsock(struct thread *td, int s, struct file **sock_fp,
518     struct socket **so)
519 {
520 	cap_rights_t rights;
521 	int error;
522 
523 	*sock_fp = NULL;
524 	*so = NULL;
525 
526 	/*
527 	 * The socket must be a stream socket and connected.
528 	 */
529 	error = getsock_cap(td, s, cap_rights_init(&rights, CAP_SEND),
530 	    sock_fp, NULL);
531 	if (error != 0)
532 		return (error);
533 	*so = (*sock_fp)->f_data;
534 	if ((*so)->so_type != SOCK_STREAM)
535 		return (EINVAL);
536 	if (((*so)->so_state & SS_ISCONNECTED) == 0)
537 		return (ENOTCONN);
538 	return (0);
539 }
540 
541 int
542 vn_sendfile(struct file *fp, int sockfd, struct uio *hdr_uio,
543     struct uio *trl_uio, off_t offset, size_t nbytes, off_t *sent, int flags,
544     int kflags, struct thread *td)
545 {
546 	struct file *sock_fp;
547 	struct vnode *vp;
548 	struct vm_object *obj;
549 	struct socket *so;
550 	struct mbuf *m, *mh, *mhtail;
551 	struct sf_buf *sf;
552 	struct shmfd *shmfd;
553 	struct sendfile_sync *sfs;
554 	struct vattr va;
555 	off_t off, sbytes, rem, obj_size;
556 	int error, softerr, bsize, hdrlen;
557 
558 	obj = NULL;
559 	so = NULL;
560 	m = mh = NULL;
561 	sfs = NULL;
562 	sbytes = 0;
563 	softerr = 0;
564 
565 	error = sendfile_getobj(td, fp, &obj, &vp, &shmfd, &obj_size, &bsize);
566 	if (error != 0)
567 		return (error);
568 
569 	error = sendfile_getsock(td, sockfd, &sock_fp, &so);
570 	if (error != 0)
571 		goto out;
572 
573 #ifdef MAC
574 	error = mac_socket_check_send(td->td_ucred, so);
575 	if (error != 0)
576 		goto out;
577 #endif
578 
579 	SFSTAT_INC(sf_syscalls);
580 	SFSTAT_ADD(sf_rhpages_requested, SF_READAHEAD(flags));
581 
582 	if (flags & SF_SYNC) {
583 		sfs = malloc(sizeof *sfs, M_TEMP, M_WAITOK | M_ZERO);
584 		mtx_init(&sfs->mtx, "sendfile", NULL, MTX_DEF);
585 		cv_init(&sfs->cv, "sendfile");
586 	}
587 
588 	/* If headers are specified copy them into mbufs. */
589 	if (hdr_uio != NULL && hdr_uio->uio_resid > 0) {
590 		hdr_uio->uio_td = td;
591 		hdr_uio->uio_rw = UIO_WRITE;
592 		/*
593 		 * In FBSD < 5.0 the nbytes to send also included
594 		 * the header.  If compat is specified subtract the
595 		 * header size from nbytes.
596 		 */
597 		if (kflags & SFK_COMPAT) {
598 			if (nbytes > hdr_uio->uio_resid)
599 				nbytes -= hdr_uio->uio_resid;
600 			else
601 				nbytes = 0;
602 		}
603 		mh = m_uiotombuf(hdr_uio, M_WAITOK, 0, 0, 0);
604 		hdrlen = m_length(mh, &mhtail);
605 	} else
606 		hdrlen = 0;
607 
608 	rem = nbytes ? omin(nbytes, obj_size - offset) : obj_size - offset;
609 
610 	/*
611 	 * Protect against multiple writers to the socket.
612 	 *
613 	 * XXXRW: Historically this has assumed non-interruptibility, so now
614 	 * we implement that, but possibly shouldn't.
615 	 */
616 	(void)sblock(&so->so_snd, SBL_WAIT | SBL_NOINTR);
617 
618 	/*
619 	 * Loop through the pages of the file, starting with the requested
620 	 * offset. Get a file page (do I/O if necessary), map the file page
621 	 * into an sf_buf, attach an mbuf header to the sf_buf, and queue
622 	 * it on the socket.
623 	 * This is done in two loops.  The inner loop turns as many pages
624 	 * as it can, up to available socket buffer space, without blocking
625 	 * into mbufs to have it bulk delivered into the socket send buffer.
626 	 * The outer loop checks the state and available space of the socket
627 	 * and takes care of the overall progress.
628 	 */
629 	for (off = offset; rem > 0; ) {
630 		struct sf_io *sfio;
631 		vm_page_t *pa;
632 		struct mbuf *mtail;
633 		int nios, space, npages, rhpages;
634 
635 		mtail = NULL;
636 		/*
637 		 * Check the socket state for ongoing connection,
638 		 * no errors and space in socket buffer.
639 		 * If space is low allow for the remainder of the
640 		 * file to be processed if it fits the socket buffer.
641 		 * Otherwise block in waiting for sufficient space
642 		 * to proceed, or if the socket is nonblocking, return
643 		 * to userland with EAGAIN while reporting how far
644 		 * we've come.
645 		 * We wait until the socket buffer has significant free
646 		 * space to do bulk sends.  This makes good use of file
647 		 * system read ahead and allows packet segmentation
648 		 * offloading hardware to take over lots of work.  If
649 		 * we were not careful here we would send off only one
650 		 * sfbuf at a time.
651 		 */
652 		SOCKBUF_LOCK(&so->so_snd);
653 		if (so->so_snd.sb_lowat < so->so_snd.sb_hiwat / 2)
654 			so->so_snd.sb_lowat = so->so_snd.sb_hiwat / 2;
655 retry_space:
656 		if (so->so_snd.sb_state & SBS_CANTSENDMORE) {
657 			error = EPIPE;
658 			SOCKBUF_UNLOCK(&so->so_snd);
659 			goto done;
660 		} else if (so->so_error) {
661 			error = so->so_error;
662 			so->so_error = 0;
663 			SOCKBUF_UNLOCK(&so->so_snd);
664 			goto done;
665 		}
666 		space = sbspace(&so->so_snd);
667 		if (space < rem &&
668 		    (space <= 0 ||
669 		     space < so->so_snd.sb_lowat)) {
670 			if (so->so_state & SS_NBIO) {
671 				SOCKBUF_UNLOCK(&so->so_snd);
672 				error = EAGAIN;
673 				goto done;
674 			}
675 			/*
676 			 * sbwait drops the lock while sleeping.
677 			 * When we loop back to retry_space the
678 			 * state may have changed and we retest
679 			 * for it.
680 			 */
681 			error = sbwait(&so->so_snd);
682 			/*
683 			 * An error from sbwait usually indicates that we've
684 			 * been interrupted by a signal. If we've sent anything
685 			 * then return bytes sent, otherwise return the error.
686 			 */
687 			if (error != 0) {
688 				SOCKBUF_UNLOCK(&so->so_snd);
689 				goto done;
690 			}
691 			goto retry_space;
692 		}
693 		SOCKBUF_UNLOCK(&so->so_snd);
694 
695 		/*
696 		 * Reduce space in the socket buffer by the size of
697 		 * the header mbuf chain.
698 		 * hdrlen is set to 0 after the first loop.
699 		 */
700 		space -= hdrlen;
701 
702 		if (vp != NULL) {
703 			error = vn_lock(vp, LK_SHARED);
704 			if (error != 0)
705 				goto done;
706 			error = VOP_GETATTR(vp, &va, td->td_ucred);
707 			if (error != 0 || off >= va.va_size) {
708 				VOP_UNLOCK(vp, 0);
709 				goto done;
710 			}
711 			if (va.va_size != obj_size) {
712 				if (nbytes == 0)
713 					rem += va.va_size - obj_size;
714 				else if (offset + nbytes > va.va_size)
715 					rem -= (offset + nbytes - va.va_size);
716 				obj_size = va.va_size;
717 			}
718 		}
719 
720 		if (space > rem)
721 			space = rem;
722 
723 		npages = howmany(space + (off & PAGE_MASK), PAGE_SIZE);
724 
725 		/*
726 		 * Calculate maximum allowed number of pages for readahead
727 		 * at this iteration.  First, we allow readahead up to "rem".
728 		 * If application wants more, let it be, but there is no
729 		 * reason to go above MAXPHYS.  Also check against "obj_size",
730 		 * since vm_pager_has_page() can hint beyond EOF.
731 		 */
732 		rhpages = howmany(rem + (off & PAGE_MASK), PAGE_SIZE) - npages;
733 		rhpages += SF_READAHEAD(flags);
734 		rhpages = min(howmany(MAXPHYS, PAGE_SIZE), rhpages);
735 		rhpages = min(howmany(obj_size - trunc_page(off), PAGE_SIZE) -
736 		    npages, rhpages);
737 
738 		sfio = malloc(sizeof(struct sf_io) +
739 		    npages * sizeof(vm_page_t), M_TEMP, M_WAITOK);
740 		refcount_init(&sfio->nios, 1);
741 		sfio->error = 0;
742 
743 		nios = sendfile_swapin(obj, sfio, off, space, npages, rhpages,
744 		    flags);
745 
746 		/*
747 		 * Loop and construct maximum sized mbuf chain to be bulk
748 		 * dumped into socket buffer.
749 		 */
750 		pa = sfio->pa;
751 		for (int i = 0; i < npages; i++) {
752 			struct mbuf *m0;
753 
754 			/*
755 			 * If a page wasn't grabbed successfully, then
756 			 * trim the array. Can happen only with SF_NODISKIO.
757 			 */
758 			if (pa[i] == NULL) {
759 				SFSTAT_INC(sf_busy);
760 				fixspace(npages, i, off, &space);
761 				npages = i;
762 				softerr = EBUSY;
763 				break;
764 			}
765 
766 			/*
767 			 * Get a sendfile buf.  When allocating the
768 			 * first buffer for mbuf chain, we usually
769 			 * wait as long as necessary, but this wait
770 			 * can be interrupted.  For consequent
771 			 * buffers, do not sleep, since several
772 			 * threads might exhaust the buffers and then
773 			 * deadlock.
774 			 */
775 			sf = sf_buf_alloc(pa[i],
776 			    m != NULL ? SFB_NOWAIT : SFB_CATCH);
777 			if (sf == NULL) {
778 				SFSTAT_INC(sf_allocfail);
779 				for (int j = i; j < npages; j++) {
780 					vm_page_lock(pa[j]);
781 					vm_page_unwire(pa[j], PQ_INACTIVE);
782 					vm_page_unlock(pa[j]);
783 				}
784 				if (m == NULL)
785 					softerr = ENOBUFS;
786 				fixspace(npages, i, off, &space);
787 				npages = i;
788 				break;
789 			}
790 
791 			m0 = m_get(M_WAITOK, MT_DATA);
792 			m0->m_ext.ext_buf = (char *)sf_buf_kva(sf);
793 			m0->m_ext.ext_size = PAGE_SIZE;
794 			m0->m_ext.ext_arg1 = sf;
795 			m0->m_ext.ext_arg2 = sfs;
796 			/*
797 			 * SF_NOCACHE sets the page as being freed upon send.
798 			 * However, we ignore it for the last page in 'space',
799 			 * if the page is truncated, and we got more data to
800 			 * send (rem > space), or if we have readahead
801 			 * configured (rhpages > 0).
802 			 */
803 			if ((flags & SF_NOCACHE) == 0 ||
804 			    (i == npages - 1 &&
805 			    ((off + space) & PAGE_MASK) &&
806 			    (rem > space || rhpages > 0)))
807 				m0->m_ext.ext_type = EXT_SFBUF;
808 			else
809 				m0->m_ext.ext_type = EXT_SFBUF_NOCACHE;
810 			m0->m_ext.ext_flags = 0;
811 			m0->m_flags |= (M_EXT | M_RDONLY);
812 			if (nios)
813 				m0->m_flags |= M_NOTREADY;
814 			m0->m_data = (char *)sf_buf_kva(sf) +
815 			    (vmoff(i, off) & PAGE_MASK);
816 			m0->m_len = xfsize(i, npages, off, space);
817 
818 			if (i == 0)
819 				sfio->m = m0;
820 
821 			/* Append to mbuf chain. */
822 			if (mtail != NULL)
823 				mtail->m_next = m0;
824 			else
825 				m = m0;
826 			mtail = m0;
827 
828 			if (sfs != NULL) {
829 				mtx_lock(&sfs->mtx);
830 				sfs->count++;
831 				mtx_unlock(&sfs->mtx);
832 			}
833 		}
834 
835 		if (vp != NULL)
836 			VOP_UNLOCK(vp, 0);
837 
838 		/* Keep track of bytes processed. */
839 		off += space;
840 		rem -= space;
841 
842 		/* Prepend header, if any. */
843 		if (hdrlen) {
844 			mhtail->m_next = m;
845 			m = mh;
846 			mh = NULL;
847 		}
848 
849 		if (m == NULL) {
850 			KASSERT(softerr, ("%s: m NULL, no error", __func__));
851 			error = softerr;
852 			free(sfio, M_TEMP);
853 			goto done;
854 		}
855 
856 		/* Add the buffer chain to the socket buffer. */
857 		KASSERT(m_length(m, NULL) == space + hdrlen,
858 		    ("%s: mlen %u space %d hdrlen %d",
859 		    __func__, m_length(m, NULL), space, hdrlen));
860 
861 		CURVNET_SET(so->so_vnet);
862 		if (nios == 0) {
863 			/*
864 			 * If sendfile_swapin() didn't initiate any I/Os,
865 			 * which happens if all data is cached in VM, then
866 			 * we can send data right now without the
867 			 * PRUS_NOTREADY flag.
868 			 */
869 			free(sfio, M_TEMP);
870 			error = (*so->so_proto->pr_usrreqs->pru_send)
871 			    (so, 0, m, NULL, NULL, td);
872 		} else {
873 			sfio->sock_fp = sock_fp;
874 			sfio->npages = npages;
875 			fhold(sock_fp);
876 			error = (*so->so_proto->pr_usrreqs->pru_send)
877 			    (so, PRUS_NOTREADY, m, NULL, NULL, td);
878 			sendfile_iodone(sfio, NULL, 0, 0);
879 		}
880 		CURVNET_RESTORE();
881 
882 		m = NULL;	/* pru_send always consumes */
883 		if (error)
884 			goto done;
885 		sbytes += space + hdrlen;
886 		if (hdrlen)
887 			hdrlen = 0;
888 		if (softerr) {
889 			error = softerr;
890 			goto done;
891 		}
892 	}
893 
894 	/*
895 	 * Send trailers. Wimp out and use writev(2).
896 	 */
897 	if (trl_uio != NULL) {
898 		sbunlock(&so->so_snd);
899 		error = kern_writev(td, sockfd, trl_uio);
900 		if (error == 0)
901 			sbytes += td->td_retval[0];
902 		goto out;
903 	}
904 
905 done:
906 	sbunlock(&so->so_snd);
907 out:
908 	/*
909 	 * If there was no error we have to clear td->td_retval[0]
910 	 * because it may have been set by writev.
911 	 */
912 	if (error == 0) {
913 		td->td_retval[0] = 0;
914 	}
915 	if (sent != NULL) {
916 		(*sent) = sbytes;
917 	}
918 	if (obj != NULL)
919 		vm_object_deallocate(obj);
920 	if (so)
921 		fdrop(sock_fp, td);
922 	if (m)
923 		m_freem(m);
924 	if (mh)
925 		m_freem(mh);
926 
927 	if (sfs != NULL) {
928 		mtx_lock(&sfs->mtx);
929 		if (sfs->count != 0)
930 			cv_wait(&sfs->cv, &sfs->mtx);
931 		KASSERT(sfs->count == 0, ("sendfile sync still busy"));
932 		cv_destroy(&sfs->cv);
933 		mtx_destroy(&sfs->mtx);
934 		free(sfs, M_TEMP);
935 	}
936 
937 	if (error == ERESTART)
938 		error = EINTR;
939 
940 	return (error);
941 }
942 
943 static int
944 sendfile(struct thread *td, struct sendfile_args *uap, int compat)
945 {
946 	struct sf_hdtr hdtr;
947 	struct uio *hdr_uio, *trl_uio;
948 	struct file *fp;
949 	cap_rights_t rights;
950 	off_t sbytes;
951 	int error;
952 
953 	/*
954 	 * File offset must be positive.  If it goes beyond EOF
955 	 * we send only the header/trailer and no payload data.
956 	 */
957 	if (uap->offset < 0)
958 		return (EINVAL);
959 
960 	hdr_uio = trl_uio = NULL;
961 
962 	if (uap->hdtr != NULL) {
963 		error = copyin(uap->hdtr, &hdtr, sizeof(hdtr));
964 		if (error != 0)
965 			goto out;
966 		if (hdtr.headers != NULL) {
967 			error = copyinuio(hdtr.headers, hdtr.hdr_cnt,
968 			    &hdr_uio);
969 			if (error != 0)
970 				goto out;
971 		}
972 		if (hdtr.trailers != NULL) {
973 			error = copyinuio(hdtr.trailers, hdtr.trl_cnt,
974 			    &trl_uio);
975 			if (error != 0)
976 				goto out;
977 		}
978 	}
979 
980 	AUDIT_ARG_FD(uap->fd);
981 
982 	/*
983 	 * sendfile(2) can start at any offset within a file so we require
984 	 * CAP_READ+CAP_SEEK = CAP_PREAD.
985 	 */
986 	if ((error = fget_read(td, uap->fd,
987 	    cap_rights_init(&rights, CAP_PREAD), &fp)) != 0) {
988 		goto out;
989 	}
990 
991 	error = fo_sendfile(fp, uap->s, hdr_uio, trl_uio, uap->offset,
992 	    uap->nbytes, &sbytes, uap->flags, compat ? SFK_COMPAT : 0, td);
993 	fdrop(fp, td);
994 
995 	if (uap->sbytes != NULL)
996 		copyout(&sbytes, uap->sbytes, sizeof(off_t));
997 
998 out:
999 	free(hdr_uio, M_IOV);
1000 	free(trl_uio, M_IOV);
1001 	return (error);
1002 }
1003 
1004 /*
1005  * sendfile(2)
1006  *
1007  * int sendfile(int fd, int s, off_t offset, size_t nbytes,
1008  *       struct sf_hdtr *hdtr, off_t *sbytes, int flags)
1009  *
1010  * Send a file specified by 'fd' and starting at 'offset' to a socket
1011  * specified by 's'. Send only 'nbytes' of the file or until EOF if nbytes ==
1012  * 0.  Optionally add a header and/or trailer to the socket output.  If
1013  * specified, write the total number of bytes sent into *sbytes.
1014  */
1015 int
1016 sys_sendfile(struct thread *td, struct sendfile_args *uap)
1017 {
1018 
1019 	return (sendfile(td, uap, 0));
1020 }
1021 
1022 #ifdef COMPAT_FREEBSD4
1023 int
1024 freebsd4_sendfile(struct thread *td, struct freebsd4_sendfile_args *uap)
1025 {
1026 	struct sendfile_args args;
1027 
1028 	args.fd = uap->fd;
1029 	args.s = uap->s;
1030 	args.offset = uap->offset;
1031 	args.nbytes = uap->nbytes;
1032 	args.hdtr = uap->hdtr;
1033 	args.sbytes = uap->sbytes;
1034 	args.flags = uap->flags;
1035 
1036 	return (sendfile(td, &args, 1));
1037 }
1038 #endif /* COMPAT_FREEBSD4 */
1039