xref: /freebsd/sys/kern/kern_sendfile.c (revision d8a0fe102c0cfdfcd5b818f850eff09d8536c9bc)
1 /*-
2  * Copyright (c) 2013-2015 Gleb Smirnoff <glebius@FreeBSD.org>
3  * Copyright (c) 1998, David Greenman. All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  * 3. Neither the name of the University nor the names of its contributors
14  *    may be used to endorse or promote products derived from this software
15  *    without specific prior written permission.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 #include <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 #define	EXT_FLAG_SYNC		EXT_FLAG_VENDOR1
66 #define	EXT_FLAG_NOCACHE	EXT_FLAG_VENDOR2
67 
68 /*
69  * Structure describing a single sendfile(2) I/O, which may consist of
70  * several underlying pager I/Os.
71  *
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'.
75  *
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
80  * of the data.
81  */
82 struct sf_io {
83 	volatile u_int	nios;
84 	u_int		error;
85 	int		npages;
86 	struct socket	*so;
87 	struct mbuf	*m;
88 	vm_page_t	pa[];
89 };
90 
91 /*
92  * Structure used to track requests with SF_SYNC flag.
93  */
94 struct sendfile_sync {
95 	struct mtx	mtx;
96 	struct cv	cv;
97 	unsigned	count;
98 };
99 
100 counter_u64_t sfstat[sizeof(struct sfstat) / sizeof(uint64_t)];
101 
102 static void
103 sfstat_init(const void *unused)
104 {
105 
106 	COUNTER_ARRAY_ALLOC(sfstat, sizeof(struct sfstat) / sizeof(uint64_t),
107 	    M_WAITOK);
108 }
109 SYSINIT(sfstat, SI_SUB_MBUF, SI_ORDER_FIRST, sfstat_init, NULL);
110 
111 static int
112 sfstat_sysctl(SYSCTL_HANDLER_ARGS)
113 {
114 	struct sfstat s;
115 
116 	COUNTER_ARRAY_COPY(sfstat, &s, sizeof(s) / sizeof(uint64_t));
117 	if (req->newptr)
118 		COUNTER_ARRAY_ZERO(sfstat, sizeof(s) / sizeof(uint64_t));
119 	return (SYSCTL_OUT(req, &s, sizeof(s)));
120 }
121 SYSCTL_PROC(_kern_ipc, OID_AUTO, sfstat, CTLTYPE_OPAQUE | CTLFLAG_RW,
122     NULL, 0, sfstat_sysctl, "I", "sendfile statistics");
123 
124 /*
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.
127  */
128 static void
129 sendfile_free_page(vm_page_t pg, bool nocache)
130 {
131 
132 	vm_page_lock(pg);
133 	/*
134 	 * In either case check for the object going away on us.  This can
135 	 * happen since we don't hold a reference to it.  If so, we're
136 	 * responsible for freeing the page.  In 'noncache' case try to free
137 	 * the page, but only if it is cheap to.
138 	 */
139 	if (vm_page_unwire(pg, nocache ? PQ_NONE : PQ_INACTIVE)) {
140 		vm_object_t obj;
141 
142 		if ((obj = pg->object) == NULL)
143 			vm_page_free(pg);
144 		else if (nocache) {
145 			if (!vm_page_xbusied(pg) && VM_OBJECT_TRYWLOCK(obj)) {
146 				bool freed;
147 
148 				/* Only free unmapped pages. */
149 				if (obj->ref_count == 0 ||
150 				    !pmap_page_is_mapped(pg))
151 					/*
152 					 * The busy test before the object is
153 					 * locked cannot be relied upon.
154 					 */
155 					freed = vm_page_try_to_free(pg);
156 				else
157 					freed = false;
158 				VM_OBJECT_WUNLOCK(obj);
159 				if (!freed)
160 					vm_page_deactivate_noreuse(pg);
161 			} else
162 				vm_page_deactivate_noreuse(pg);
163 		}
164 	}
165 	vm_page_unlock(pg);
166 }
167 
168 static void
169 sendfile_free_mext(struct mbuf *m)
170 {
171 	struct sf_buf *sf;
172 	vm_page_t pg;
173 	bool nocache;
174 
175 	KASSERT(m->m_flags & M_EXT && m->m_ext.ext_type == EXT_SFBUF,
176 	    ("%s: m %p !M_EXT or !EXT_SFBUF", __func__, m));
177 
178 	sf = m->m_ext.ext_arg1;
179 	pg = sf_buf_page(sf);
180 	nocache = m->m_ext.ext_flags & EXT_FLAG_NOCACHE;
181 
182 	sf_buf_free(sf);
183 	sendfile_free_page(pg, nocache);
184 
185 	if (m->m_ext.ext_flags & EXT_FLAG_SYNC) {
186 		struct sendfile_sync *sfs = m->m_ext.ext_arg2;
187 
188 		mtx_lock(&sfs->mtx);
189 		KASSERT(sfs->count > 0, ("Sendfile sync botchup count == 0"));
190 		if (--sfs->count == 0)
191 			cv_signal(&sfs->cv);
192 		mtx_unlock(&sfs->mtx);
193 	}
194 }
195 
196 /*
197  * Helper function to calculate how much data to put into page i of n.
198  * Only first and last pages are special.
199  */
200 static inline off_t
201 xfsize(int i, int n, off_t off, off_t len)
202 {
203 
204 	if (i == 0)
205 		return (omin(PAGE_SIZE - (off & PAGE_MASK), len));
206 
207 	if (i == n - 1 && ((off + len) & PAGE_MASK) > 0)
208 		return ((off + len) & PAGE_MASK);
209 
210 	return (PAGE_SIZE);
211 }
212 
213 /*
214  * Helper function to get offset within object for i page.
215  */
216 static inline vm_ooffset_t
217 vmoff(int i, off_t off)
218 {
219 
220 	if (i == 0)
221 		return ((vm_ooffset_t)off);
222 
223 	return (trunc_page(off + i * PAGE_SIZE));
224 }
225 
226 /*
227  * Helper function used when allocation of a page or sf_buf failed.
228  * Pretend as if we don't have enough space, subtract xfsize() of
229  * all pages that failed.
230  */
231 static inline void
232 fixspace(int old, int new, off_t off, int *space)
233 {
234 
235 	KASSERT(old > new, ("%s: old %d new %d", __func__, old, new));
236 
237 	/* Subtract last one. */
238 	*space -= xfsize(old - 1, old, off, *space);
239 	old--;
240 
241 	if (new == old)
242 		/* There was only one page. */
243 		return;
244 
245 	/* Subtract first one. */
246 	if (new == 0) {
247 		*space -= xfsize(0, old, off, *space);
248 		new++;
249 	}
250 
251 	/* Rest of pages are full sized. */
252 	*space -= (old - new) * PAGE_SIZE;
253 
254 	KASSERT(*space >= 0, ("%s: space went backwards", __func__));
255 }
256 
257 /*
258  * I/O completion callback.
259  */
260 static void
261 sendfile_iodone(void *arg, vm_page_t *pg, int count, int error)
262 {
263 	struct sf_io *sfio = arg;
264 	struct socket *so = sfio->so;
265 
266 	for (int i = 0; i < count; i++)
267 		if (pg[i] != bogus_page)
268 			vm_page_xunbusy(pg[i]);
269 
270 	if (error)
271 		sfio->error = error;
272 
273 	if (!refcount_release(&sfio->nios))
274 		return;
275 
276 	if (sfio->error) {
277 		struct mbuf *m;
278 
279 		/*
280 		 * I/O operation failed.  The state of data in the socket
281 		 * is now inconsistent, and all what we can do is to tear
282 		 * it down. Protocol abort method would tear down protocol
283 		 * state, free all ready mbufs and detach not ready ones.
284 		 * We will free the mbufs corresponding to this I/O manually.
285 		 *
286 		 * The socket would be marked with EIO and made available
287 		 * for read, so that application receives EIO on next
288 		 * syscall and eventually closes the socket.
289 		 */
290 		so->so_proto->pr_usrreqs->pru_abort(so);
291 		so->so_error = EIO;
292 
293 		m = sfio->m;
294 		for (int i = 0; i < sfio->npages; i++)
295 			m = m_free(m);
296 	} else {
297 		CURVNET_SET(so->so_vnet);
298 		(void )(so->so_proto->pr_usrreqs->pru_ready)(so, sfio->m,
299 		    sfio->npages);
300 		CURVNET_RESTORE();
301 	}
302 
303 	SOCK_LOCK(so);
304 	sorele(so);
305 	free(sfio, M_TEMP);
306 }
307 
308 /*
309  * Iterate through pages vector and request paging for non-valid pages.
310  */
311 static int
312 sendfile_swapin(vm_object_t obj, struct sf_io *sfio, off_t off, off_t len,
313     int npages, int rhpages, int flags)
314 {
315 	vm_page_t *pa = sfio->pa;
316 	int grabbed, nios;
317 
318 	nios = 0;
319 	flags = (flags & SF_NODISKIO) ? VM_ALLOC_NOWAIT : 0;
320 
321 	/*
322 	 * First grab all the pages and wire them.  Note that we grab
323 	 * only required pages.  Readahead pages are dealt with later.
324 	 */
325 	VM_OBJECT_WLOCK(obj);
326 
327 	grabbed = vm_page_grab_pages(obj, OFF_TO_IDX(off),
328 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED | flags, pa, npages);
329 	if (grabbed < npages) {
330 		for (int i = grabbed; i < npages; i++)
331 			pa[i] = NULL;
332 		npages = grabbed;
333 		rhpages = 0;
334 	}
335 
336 	for (int i = 0; i < npages;) {
337 		int j, a, count, rv;
338 
339 		/* Skip valid pages. */
340 		if (vm_page_is_valid(pa[i], vmoff(i, off) & PAGE_MASK,
341 		    xfsize(i, npages, off, len))) {
342 			vm_page_xunbusy(pa[i]);
343 			SFSTAT_INC(sf_pages_valid);
344 			i++;
345 			continue;
346 		}
347 
348 		/*
349 		 * Next page is invalid.  Check if it belongs to pager.  It
350 		 * may not be there, which is a regular situation for shmem
351 		 * pager.  For vnode pager this happens only in case of
352 		 * a sparse file.
353 		 *
354 		 * Important feature of vm_pager_has_page() is the hint
355 		 * stored in 'a', about how many pages we can pagein after
356 		 * this page in a single I/O.
357 		 */
358 		if (!vm_pager_has_page(obj, OFF_TO_IDX(vmoff(i, off)), NULL,
359 		    &a)) {
360 			pmap_zero_page(pa[i]);
361 			pa[i]->valid = VM_PAGE_BITS_ALL;
362 			MPASS(pa[i]->dirty == 0);
363 			vm_page_xunbusy(pa[i]);
364 			i++;
365 			continue;
366 		}
367 
368 		/*
369 		 * We want to pagein as many pages as possible, limited only
370 		 * by the 'a' hint and actual request.
371 		 */
372 		count = min(a + 1, npages - i);
373 
374 		/*
375 		 * We should not pagein into a valid page, thus we first trim
376 		 * any valid pages off the end of request, and substitute
377 		 * to bogus_page those, that are in the middle.
378 		 */
379 		for (j = i + count - 1; j > i; j--) {
380 			if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
381 			    xfsize(j, npages, off, len))) {
382 				count--;
383 				rhpages = 0;
384 			} else
385 				break;
386 		}
387 		for (j = i + 1; j < i + count - 1; j++)
388 			if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
389 			    xfsize(j, npages, off, len))) {
390 				vm_page_xunbusy(pa[j]);
391 				SFSTAT_INC(sf_pages_valid);
392 				SFSTAT_INC(sf_pages_bogus);
393 				pa[j] = bogus_page;
394 			}
395 
396 		refcount_acquire(&sfio->nios);
397 		rv = vm_pager_get_pages_async(obj, pa + i, count, NULL,
398 		    i + count == npages ? &rhpages : NULL,
399 		    &sendfile_iodone, sfio);
400 		KASSERT(rv == VM_PAGER_OK, ("%s: pager fail obj %p page %p",
401 		    __func__, obj, pa[i]));
402 
403 		SFSTAT_INC(sf_iocnt);
404 		SFSTAT_ADD(sf_pages_read, count);
405 		if (i + count == npages)
406 			SFSTAT_ADD(sf_rhpages_read, rhpages);
407 
408 		/*
409 		 * Restore the valid page pointers.  They are already
410 		 * unbusied, but still wired.
411 		 */
412 		for (j = i; j < i + count; j++)
413 			if (pa[j] == bogus_page) {
414 				pa[j] = vm_page_lookup(obj,
415 				    OFF_TO_IDX(vmoff(j, off)));
416 				KASSERT(pa[j], ("%s: page %p[%d] disappeared",
417 				    __func__, pa, j));
418 
419 			}
420 		i += count;
421 		nios++;
422 	}
423 
424 	VM_OBJECT_WUNLOCK(obj);
425 
426 	if (nios == 0 && npages != 0)
427 		SFSTAT_INC(sf_noiocnt);
428 
429 	return (nios);
430 }
431 
432 static int
433 sendfile_getobj(struct thread *td, struct file *fp, vm_object_t *obj_res,
434     struct vnode **vp_res, struct shmfd **shmfd_res, off_t *obj_size,
435     int *bsize)
436 {
437 	struct vattr va;
438 	vm_object_t obj;
439 	struct vnode *vp;
440 	struct shmfd *shmfd;
441 	int error;
442 
443 	vp = *vp_res = NULL;
444 	obj = NULL;
445 	shmfd = *shmfd_res = NULL;
446 	*bsize = 0;
447 
448 	/*
449 	 * The file descriptor must be a regular file and have a
450 	 * backing VM object.
451 	 */
452 	if (fp->f_type == DTYPE_VNODE) {
453 		vp = fp->f_vnode;
454 		vn_lock(vp, LK_SHARED | LK_RETRY);
455 		if (vp->v_type != VREG) {
456 			error = EINVAL;
457 			goto out;
458 		}
459 		*bsize = vp->v_mount->mnt_stat.f_iosize;
460 		error = VOP_GETATTR(vp, &va, td->td_ucred);
461 		if (error != 0)
462 			goto out;
463 		*obj_size = va.va_size;
464 		obj = vp->v_object;
465 		if (obj == NULL) {
466 			error = EINVAL;
467 			goto out;
468 		}
469 	} else if (fp->f_type == DTYPE_SHM) {
470 		error = 0;
471 		shmfd = fp->f_data;
472 		obj = shmfd->shm_object;
473 		*obj_size = shmfd->shm_size;
474 	} else {
475 		error = EINVAL;
476 		goto out;
477 	}
478 
479 	VM_OBJECT_WLOCK(obj);
480 	if ((obj->flags & OBJ_DEAD) != 0) {
481 		VM_OBJECT_WUNLOCK(obj);
482 		error = EBADF;
483 		goto out;
484 	}
485 
486 	/*
487 	 * Temporarily increase the backing VM object's reference
488 	 * count so that a forced reclamation of its vnode does not
489 	 * immediately destroy it.
490 	 */
491 	vm_object_reference_locked(obj);
492 	VM_OBJECT_WUNLOCK(obj);
493 	*obj_res = obj;
494 	*vp_res = vp;
495 	*shmfd_res = shmfd;
496 
497 out:
498 	if (vp != NULL)
499 		VOP_UNLOCK(vp, 0);
500 	return (error);
501 }
502 
503 static int
504 sendfile_getsock(struct thread *td, int s, struct file **sock_fp,
505     struct socket **so)
506 {
507 	cap_rights_t rights;
508 	int error;
509 
510 	*sock_fp = NULL;
511 	*so = NULL;
512 
513 	/*
514 	 * The socket must be a stream socket and connected.
515 	 */
516 	error = getsock_cap(td, s, cap_rights_init(&rights, CAP_SEND),
517 	    sock_fp, NULL, NULL);
518 	if (error != 0)
519 		return (error);
520 	*so = (*sock_fp)->f_data;
521 	if ((*so)->so_type != SOCK_STREAM)
522 		return (EINVAL);
523 	return (0);
524 }
525 
526 int
527 vn_sendfile(struct file *fp, int sockfd, struct uio *hdr_uio,
528     struct uio *trl_uio, off_t offset, size_t nbytes, off_t *sent, int flags,
529     struct thread *td)
530 {
531 	struct file *sock_fp;
532 	struct vnode *vp;
533 	struct vm_object *obj;
534 	struct socket *so;
535 	struct mbuf *m, *mh, *mhtail;
536 	struct sf_buf *sf;
537 	struct shmfd *shmfd;
538 	struct sendfile_sync *sfs;
539 	struct vattr va;
540 	off_t off, sbytes, rem, obj_size;
541 	int error, softerr, bsize, hdrlen;
542 
543 	obj = NULL;
544 	so = NULL;
545 	m = mh = NULL;
546 	sfs = NULL;
547 	hdrlen = sbytes = 0;
548 	softerr = 0;
549 
550 	error = sendfile_getobj(td, fp, &obj, &vp, &shmfd, &obj_size, &bsize);
551 	if (error != 0)
552 		return (error);
553 
554 	error = sendfile_getsock(td, sockfd, &sock_fp, &so);
555 	if (error != 0)
556 		goto out;
557 
558 #ifdef MAC
559 	error = mac_socket_check_send(td->td_ucred, so);
560 	if (error != 0)
561 		goto out;
562 #endif
563 
564 	SFSTAT_INC(sf_syscalls);
565 	SFSTAT_ADD(sf_rhpages_requested, SF_READAHEAD(flags));
566 
567 	if (flags & SF_SYNC) {
568 		sfs = malloc(sizeof *sfs, M_TEMP, M_WAITOK | M_ZERO);
569 		mtx_init(&sfs->mtx, "sendfile", NULL, MTX_DEF);
570 		cv_init(&sfs->cv, "sendfile");
571 	}
572 
573 	rem = nbytes ? omin(nbytes, obj_size - offset) : obj_size - offset;
574 
575 	/*
576 	 * Protect against multiple writers to the socket.
577 	 *
578 	 * XXXRW: Historically this has assumed non-interruptibility, so now
579 	 * we implement that, but possibly shouldn't.
580 	 */
581 	(void)sblock(&so->so_snd, SBL_WAIT | SBL_NOINTR);
582 
583 	/*
584 	 * Loop through the pages of the file, starting with the requested
585 	 * offset. Get a file page (do I/O if necessary), map the file page
586 	 * into an sf_buf, attach an mbuf header to the sf_buf, and queue
587 	 * it on the socket.
588 	 * This is done in two loops.  The inner loop turns as many pages
589 	 * as it can, up to available socket buffer space, without blocking
590 	 * into mbufs to have it bulk delivered into the socket send buffer.
591 	 * The outer loop checks the state and available space of the socket
592 	 * and takes care of the overall progress.
593 	 */
594 	for (off = offset; rem > 0; ) {
595 		struct sf_io *sfio;
596 		vm_page_t *pa;
597 		struct mbuf *mtail;
598 		int nios, space, npages, rhpages;
599 
600 		mtail = NULL;
601 		/*
602 		 * Check the socket state for ongoing connection,
603 		 * no errors and space in socket buffer.
604 		 * If space is low allow for the remainder of the
605 		 * file to be processed if it fits the socket buffer.
606 		 * Otherwise block in waiting for sufficient space
607 		 * to proceed, or if the socket is nonblocking, return
608 		 * to userland with EAGAIN while reporting how far
609 		 * we've come.
610 		 * We wait until the socket buffer has significant free
611 		 * space to do bulk sends.  This makes good use of file
612 		 * system read ahead and allows packet segmentation
613 		 * offloading hardware to take over lots of work.  If
614 		 * we were not careful here we would send off only one
615 		 * sfbuf at a time.
616 		 */
617 		SOCKBUF_LOCK(&so->so_snd);
618 		if (so->so_snd.sb_lowat < so->so_snd.sb_hiwat / 2)
619 			so->so_snd.sb_lowat = so->so_snd.sb_hiwat / 2;
620 retry_space:
621 		if (so->so_snd.sb_state & SBS_CANTSENDMORE) {
622 			error = EPIPE;
623 			SOCKBUF_UNLOCK(&so->so_snd);
624 			goto done;
625 		} else if (so->so_error) {
626 			error = so->so_error;
627 			so->so_error = 0;
628 			SOCKBUF_UNLOCK(&so->so_snd);
629 			goto done;
630 		}
631 		if ((so->so_state & SS_ISCONNECTED) == 0) {
632 			SOCKBUF_UNLOCK(&so->so_snd);
633 			error = ENOTCONN;
634 			goto done;
635 		}
636 
637 		space = sbspace(&so->so_snd);
638 		if (space < rem &&
639 		    (space <= 0 ||
640 		     space < so->so_snd.sb_lowat)) {
641 			if (so->so_state & SS_NBIO) {
642 				SOCKBUF_UNLOCK(&so->so_snd);
643 				error = EAGAIN;
644 				goto done;
645 			}
646 			/*
647 			 * sbwait drops the lock while sleeping.
648 			 * When we loop back to retry_space the
649 			 * state may have changed and we retest
650 			 * for it.
651 			 */
652 			error = sbwait(&so->so_snd);
653 			/*
654 			 * An error from sbwait usually indicates that we've
655 			 * been interrupted by a signal. If we've sent anything
656 			 * then return bytes sent, otherwise return the error.
657 			 */
658 			if (error != 0) {
659 				SOCKBUF_UNLOCK(&so->so_snd);
660 				goto done;
661 			}
662 			goto retry_space;
663 		}
664 		SOCKBUF_UNLOCK(&so->so_snd);
665 
666 		/*
667 		 * At the beginning of the first loop check if any headers
668 		 * are specified and copy them into mbufs.  Reduce space in
669 		 * the socket buffer by the size of the header mbuf chain.
670 		 * Clear hdr_uio here and hdrlen at the end of the first loop.
671 		 */
672 		if (hdr_uio != NULL && hdr_uio->uio_resid > 0) {
673 			hdr_uio->uio_td = td;
674 			hdr_uio->uio_rw = UIO_WRITE;
675 			mh = m_uiotombuf(hdr_uio, M_WAITOK, space, 0, 0);
676 			hdrlen = m_length(mh, &mhtail);
677 			space -= hdrlen;
678 			/*
679 			 * If header consumed all the socket buffer space,
680 			 * don't waste CPU cycles and jump to the end.
681 			 */
682 			if (space == 0) {
683 				sfio = NULL;
684 				nios = 0;
685 				goto prepend_header;
686 			}
687 			hdr_uio = NULL;
688 		}
689 
690 		if (vp != NULL) {
691 			error = vn_lock(vp, LK_SHARED);
692 			if (error != 0)
693 				goto done;
694 			error = VOP_GETATTR(vp, &va, td->td_ucred);
695 			if (error != 0 || off >= va.va_size) {
696 				VOP_UNLOCK(vp, 0);
697 				goto done;
698 			}
699 			if (va.va_size != obj_size) {
700 				obj_size = va.va_size;
701 				rem = nbytes ?
702 				    omin(nbytes + offset, obj_size) : obj_size;
703 				rem -= off;
704 			}
705 		}
706 
707 		if (space > rem)
708 			space = rem;
709 
710 		npages = howmany(space + (off & PAGE_MASK), PAGE_SIZE);
711 
712 		/*
713 		 * Calculate maximum allowed number of pages for readahead
714 		 * at this iteration.  If SF_USER_READAHEAD was set, we don't
715 		 * do any heuristics and use exactly the value supplied by
716 		 * application.  Otherwise, we allow readahead up to "rem".
717 		 * If application wants more, let it be, but there is no
718 		 * reason to go above MAXPHYS.  Also check against "obj_size",
719 		 * since vm_pager_has_page() can hint beyond EOF.
720 		 */
721 		if (flags & SF_USER_READAHEAD) {
722 			rhpages = SF_READAHEAD(flags);
723 		} else {
724 			rhpages = howmany(rem + (off & PAGE_MASK), PAGE_SIZE) -
725 			    npages;
726 			rhpages += SF_READAHEAD(flags);
727 		}
728 		rhpages = min(howmany(MAXPHYS, PAGE_SIZE), rhpages);
729 		rhpages = min(howmany(obj_size - trunc_page(off), PAGE_SIZE) -
730 		    npages, rhpages);
731 
732 		sfio = malloc(sizeof(struct sf_io) +
733 		    npages * sizeof(vm_page_t), M_TEMP, M_WAITOK);
734 		refcount_init(&sfio->nios, 1);
735 		sfio->so = so;
736 		sfio->error = 0;
737 
738 		nios = sendfile_swapin(obj, sfio, off, space, npages, rhpages,
739 		    flags);
740 
741 		/*
742 		 * Loop and construct maximum sized mbuf chain to be bulk
743 		 * dumped into socket buffer.
744 		 */
745 		pa = sfio->pa;
746 		for (int i = 0; i < npages; i++) {
747 			struct mbuf *m0;
748 
749 			/*
750 			 * If a page wasn't grabbed successfully, then
751 			 * trim the array. Can happen only with SF_NODISKIO.
752 			 */
753 			if (pa[i] == NULL) {
754 				SFSTAT_INC(sf_busy);
755 				fixspace(npages, i, off, &space);
756 				npages = i;
757 				softerr = EBUSY;
758 				break;
759 			}
760 
761 			/*
762 			 * Get a sendfile buf.  When allocating the
763 			 * first buffer for mbuf chain, we usually
764 			 * wait as long as necessary, but this wait
765 			 * can be interrupted.  For consequent
766 			 * buffers, do not sleep, since several
767 			 * threads might exhaust the buffers and then
768 			 * deadlock.
769 			 */
770 			sf = sf_buf_alloc(pa[i],
771 			    m != NULL ? SFB_NOWAIT : SFB_CATCH);
772 			if (sf == NULL) {
773 				SFSTAT_INC(sf_allocfail);
774 				for (int j = i; j < npages; j++) {
775 					vm_page_lock(pa[j]);
776 					vm_page_unwire(pa[j], PQ_INACTIVE);
777 					vm_page_unlock(pa[j]);
778 				}
779 				if (m == NULL)
780 					softerr = ENOBUFS;
781 				fixspace(npages, i, off, &space);
782 				npages = i;
783 				break;
784 			}
785 
786 			m0 = m_get(M_WAITOK, MT_DATA);
787 			m0->m_ext.ext_buf = (char *)sf_buf_kva(sf);
788 			m0->m_ext.ext_size = PAGE_SIZE;
789 			m0->m_ext.ext_arg1 = sf;
790 			m0->m_ext.ext_type = EXT_SFBUF;
791 			m0->m_ext.ext_flags = EXT_FLAG_EMBREF;
792 			m0->m_ext.ext_free = sendfile_free_mext;
793 			/*
794 			 * SF_NOCACHE sets the page as being freed upon send.
795 			 * However, we ignore it for the last page in 'space',
796 			 * if the page is truncated, and we got more data to
797 			 * send (rem > space), or if we have readahead
798 			 * configured (rhpages > 0).
799 			 */
800 			if ((flags & SF_NOCACHE) &&
801 			    (i != npages - 1 ||
802 			    !((off + space) & PAGE_MASK) ||
803 			    !(rem > space || rhpages > 0)))
804 				m0->m_ext.ext_flags |= EXT_FLAG_NOCACHE;
805 			if (sfs != NULL) {
806 				m0->m_ext.ext_flags |= EXT_FLAG_SYNC;
807 				m0->m_ext.ext_arg2 = sfs;
808 				mtx_lock(&sfs->mtx);
809 				sfs->count++;
810 				mtx_unlock(&sfs->mtx);
811 			}
812 			m0->m_ext.ext_count = 1;
813 			m0->m_flags |= (M_EXT | M_RDONLY);
814 			if (nios)
815 				m0->m_flags |= M_NOTREADY;
816 			m0->m_data = (char *)sf_buf_kva(sf) +
817 			    (vmoff(i, off) & PAGE_MASK);
818 			m0->m_len = xfsize(i, npages, off, space);
819 
820 			if (i == 0)
821 				sfio->m = m0;
822 
823 			/* Append to mbuf chain. */
824 			if (mtail != NULL)
825 				mtail->m_next = m0;
826 			else
827 				m = m0;
828 			mtail = m0;
829 		}
830 
831 		if (vp != NULL)
832 			VOP_UNLOCK(vp, 0);
833 
834 		/* Keep track of bytes processed. */
835 		off += space;
836 		rem -= space;
837 
838 		/* Prepend header, if any. */
839 		if (hdrlen) {
840 prepend_header:
841 			mhtail->m_next = m;
842 			m = mh;
843 			mh = NULL;
844 		}
845 
846 		if (m == NULL) {
847 			KASSERT(softerr, ("%s: m NULL, no error", __func__));
848 			error = softerr;
849 			free(sfio, M_TEMP);
850 			goto done;
851 		}
852 
853 		/* Add the buffer chain to the socket buffer. */
854 		KASSERT(m_length(m, NULL) == space + hdrlen,
855 		    ("%s: mlen %u space %d hdrlen %d",
856 		    __func__, m_length(m, NULL), space, hdrlen));
857 
858 		CURVNET_SET(so->so_vnet);
859 		if (nios == 0) {
860 			/*
861 			 * If sendfile_swapin() didn't initiate any I/Os,
862 			 * which happens if all data is cached in VM, then
863 			 * we can send data right now without the
864 			 * PRUS_NOTREADY flag.
865 			 */
866 			free(sfio, M_TEMP);
867 			error = (*so->so_proto->pr_usrreqs->pru_send)
868 			    (so, 0, m, NULL, NULL, td);
869 		} else {
870 			sfio->npages = npages;
871 			soref(so);
872 			error = (*so->so_proto->pr_usrreqs->pru_send)
873 			    (so, PRUS_NOTREADY, m, NULL, NULL, td);
874 			sendfile_iodone(sfio, NULL, 0, 0);
875 		}
876 		CURVNET_RESTORE();
877 
878 		m = NULL;	/* pru_send always consumes */
879 		if (error)
880 			goto done;
881 		sbytes += space + hdrlen;
882 		if (hdrlen)
883 			hdrlen = 0;
884 		if (softerr) {
885 			error = softerr;
886 			goto done;
887 		}
888 	}
889 
890 	/*
891 	 * Send trailers. Wimp out and use writev(2).
892 	 */
893 	if (trl_uio != NULL) {
894 		sbunlock(&so->so_snd);
895 		error = kern_writev(td, sockfd, trl_uio);
896 		if (error == 0)
897 			sbytes += td->td_retval[0];
898 		goto out;
899 	}
900 
901 done:
902 	sbunlock(&so->so_snd);
903 out:
904 	/*
905 	 * If there was no error we have to clear td->td_retval[0]
906 	 * because it may have been set by writev.
907 	 */
908 	if (error == 0) {
909 		td->td_retval[0] = 0;
910 	}
911 	if (sent != NULL) {
912 		(*sent) = sbytes;
913 	}
914 	if (obj != NULL)
915 		vm_object_deallocate(obj);
916 	if (so)
917 		fdrop(sock_fp, td);
918 	if (m)
919 		m_freem(m);
920 	if (mh)
921 		m_freem(mh);
922 
923 	if (sfs != NULL) {
924 		mtx_lock(&sfs->mtx);
925 		if (sfs->count != 0)
926 			cv_wait(&sfs->cv, &sfs->mtx);
927 		KASSERT(sfs->count == 0, ("sendfile sync still busy"));
928 		cv_destroy(&sfs->cv);
929 		mtx_destroy(&sfs->mtx);
930 		free(sfs, M_TEMP);
931 	}
932 
933 	if (error == ERESTART)
934 		error = EINTR;
935 
936 	return (error);
937 }
938 
939 static int
940 sendfile(struct thread *td, struct sendfile_args *uap, int compat)
941 {
942 	struct sf_hdtr hdtr;
943 	struct uio *hdr_uio, *trl_uio;
944 	struct file *fp;
945 	cap_rights_t rights;
946 	off_t sbytes;
947 	int error;
948 
949 	/*
950 	 * File offset must be positive.  If it goes beyond EOF
951 	 * we send only the header/trailer and no payload data.
952 	 */
953 	if (uap->offset < 0)
954 		return (EINVAL);
955 
956 	sbytes = 0;
957 	hdr_uio = trl_uio = NULL;
958 
959 	if (uap->hdtr != NULL) {
960 		error = copyin(uap->hdtr, &hdtr, sizeof(hdtr));
961 		if (error != 0)
962 			goto out;
963 		if (hdtr.headers != NULL) {
964 			error = copyinuio(hdtr.headers, hdtr.hdr_cnt,
965 			    &hdr_uio);
966 			if (error != 0)
967 				goto out;
968 #ifdef COMPAT_FREEBSD4
969 			/*
970 			 * In FreeBSD < 5.0 the nbytes to send also included
971 			 * the header.  If compat is specified subtract the
972 			 * header size from nbytes.
973 			 */
974 			if (compat) {
975 				if (uap->nbytes > hdr_uio->uio_resid)
976 					uap->nbytes -= hdr_uio->uio_resid;
977 				else
978 					uap->nbytes = 0;
979 			}
980 #endif
981 		}
982 		if (hdtr.trailers != NULL) {
983 			error = copyinuio(hdtr.trailers, hdtr.trl_cnt,
984 			    &trl_uio);
985 			if (error != 0)
986 				goto out;
987 		}
988 	}
989 
990 	AUDIT_ARG_FD(uap->fd);
991 
992 	/*
993 	 * sendfile(2) can start at any offset within a file so we require
994 	 * CAP_READ+CAP_SEEK = CAP_PREAD.
995 	 */
996 	if ((error = fget_read(td, uap->fd,
997 	    cap_rights_init(&rights, CAP_PREAD), &fp)) != 0) {
998 		goto out;
999 	}
1000 
1001 	error = fo_sendfile(fp, uap->s, hdr_uio, trl_uio, uap->offset,
1002 	    uap->nbytes, &sbytes, uap->flags, td);
1003 	fdrop(fp, td);
1004 
1005 	if (uap->sbytes != NULL)
1006 		copyout(&sbytes, uap->sbytes, sizeof(off_t));
1007 
1008 out:
1009 	free(hdr_uio, M_IOV);
1010 	free(trl_uio, M_IOV);
1011 	return (error);
1012 }
1013 
1014 /*
1015  * sendfile(2)
1016  *
1017  * int sendfile(int fd, int s, off_t offset, size_t nbytes,
1018  *       struct sf_hdtr *hdtr, off_t *sbytes, int flags)
1019  *
1020  * Send a file specified by 'fd' and starting at 'offset' to a socket
1021  * specified by 's'. Send only 'nbytes' of the file or until EOF if nbytes ==
1022  * 0.  Optionally add a header and/or trailer to the socket output.  If
1023  * specified, write the total number of bytes sent into *sbytes.
1024  */
1025 int
1026 sys_sendfile(struct thread *td, struct sendfile_args *uap)
1027 {
1028 
1029 	return (sendfile(td, uap, 0));
1030 }
1031 
1032 #ifdef COMPAT_FREEBSD4
1033 int
1034 freebsd4_sendfile(struct thread *td, struct freebsd4_sendfile_args *uap)
1035 {
1036 	struct sendfile_args args;
1037 
1038 	args.fd = uap->fd;
1039 	args.s = uap->s;
1040 	args.offset = uap->offset;
1041 	args.nbytes = uap->nbytes;
1042 	args.hdtr = uap->hdtr;
1043 	args.sbytes = uap->sbytes;
1044 	args.flags = uap->flags;
1045 
1046 	return (sendfile(td, &args, 1));
1047 }
1048 #endif /* COMPAT_FREEBSD4 */
1049