xref: /freebsd/sys/kern/kern_sendfile.c (revision ae7e8a02e6e93455e026036132c4d053b2c12ad9)
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_kern_tls.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/ktls.h>
41 #include <sys/mutex.h>
42 #include <sys/malloc.h>
43 #include <sys/mman.h>
44 #include <sys/mount.h>
45 #include <sys/mbuf.h>
46 #include <sys/proc.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/sysproto.h>
55 #include <sys/vnode.h>
56 
57 #include <net/vnet.h>
58 #include <netinet/in.h>
59 #include <netinet/tcp.h>
60 
61 #include <security/audit/audit.h>
62 #include <security/mac/mac_framework.h>
63 
64 #include <vm/vm.h>
65 #include <vm/vm_object.h>
66 #include <vm/vm_pager.h>
67 
68 static MALLOC_DEFINE(M_SENDFILE, "sendfile", "sendfile dynamic memory");
69 
70 #define	EXT_FLAG_SYNC		EXT_FLAG_VENDOR1
71 #define	EXT_FLAG_NOCACHE	EXT_FLAG_VENDOR2
72 #define	EXT_FLAG_CACHE_LAST	EXT_FLAG_VENDOR3
73 
74 /*
75  * Structure describing a single sendfile(2) I/O, which may consist of
76  * several underlying pager I/Os.
77  *
78  * The syscall context allocates the structure and initializes 'nios'
79  * to 1.  As sendfile_swapin() runs through pages and starts asynchronous
80  * paging operations, it increments 'nios'.
81  *
82  * Every I/O completion calls sendfile_iodone(), which decrements the 'nios',
83  * and the syscall also calls sendfile_iodone() after allocating all mbufs,
84  * linking them and sending to socket.  Whoever reaches zero 'nios' is
85  * responsible to * call pru_ready on the socket, to notify it of readyness
86  * of the data.
87  */
88 struct sf_io {
89 	volatile u_int	nios;
90 	u_int		error;
91 	int		npages;
92 	struct socket	*so;
93 	struct mbuf	*m;
94 	vm_object_t	obj;
95 	vm_pindex_t	pindex0;
96 #ifdef KERN_TLS
97 	struct ktls_session *tls;
98 #endif
99 	vm_page_t	pa[];
100 };
101 
102 /*
103  * Structure used to track requests with SF_SYNC flag.
104  */
105 struct sendfile_sync {
106 	struct mtx	mtx;
107 	struct cv	cv;
108 	unsigned	count;
109 	bool		waiting;
110 };
111 
112 static void
113 sendfile_sync_destroy(struct sendfile_sync *sfs)
114 {
115 	KASSERT(sfs->count == 0, ("sendfile sync %p still busy", sfs));
116 
117 	cv_destroy(&sfs->cv);
118 	mtx_destroy(&sfs->mtx);
119 	free(sfs, M_SENDFILE);
120 }
121 
122 static void
123 sendfile_sync_signal(struct sendfile_sync *sfs)
124 {
125 	mtx_lock(&sfs->mtx);
126 	KASSERT(sfs->count > 0, ("sendfile sync %p not busy", sfs));
127 	if (--sfs->count == 0) {
128 		if (!sfs->waiting) {
129 			/* The sendfile() waiter was interrupted by a signal. */
130 			sendfile_sync_destroy(sfs);
131 			return;
132 		} else {
133 			cv_signal(&sfs->cv);
134 		}
135 	}
136 	mtx_unlock(&sfs->mtx);
137 }
138 
139 counter_u64_t sfstat[sizeof(struct sfstat) / sizeof(uint64_t)];
140 
141 static void
142 sfstat_init(const void *unused)
143 {
144 
145 	COUNTER_ARRAY_ALLOC(sfstat, sizeof(struct sfstat) / sizeof(uint64_t),
146 	    M_WAITOK);
147 }
148 SYSINIT(sfstat, SI_SUB_MBUF, SI_ORDER_FIRST, sfstat_init, NULL);
149 
150 static int
151 sfstat_sysctl(SYSCTL_HANDLER_ARGS)
152 {
153 	struct sfstat s;
154 
155 	COUNTER_ARRAY_COPY(sfstat, &s, sizeof(s) / sizeof(uint64_t));
156 	if (req->newptr)
157 		COUNTER_ARRAY_ZERO(sfstat, sizeof(s) / sizeof(uint64_t));
158 	return (SYSCTL_OUT(req, &s, sizeof(s)));
159 }
160 SYSCTL_PROC(_kern_ipc, OID_AUTO, sfstat,
161     CTLTYPE_OPAQUE | CTLFLAG_RW | CTLFLAG_NEEDGIANT, NULL, 0,
162     sfstat_sysctl, "I",
163     "sendfile statistics");
164 
165 static void
166 sendfile_free_mext(struct mbuf *m)
167 {
168 	struct sf_buf *sf;
169 	vm_page_t pg;
170 	int flags;
171 
172 	KASSERT(m->m_flags & M_EXT && m->m_ext.ext_type == EXT_SFBUF,
173 	    ("%s: m %p !M_EXT or !EXT_SFBUF", __func__, m));
174 
175 	sf = m->m_ext.ext_arg1;
176 	pg = sf_buf_page(sf);
177 	flags = (m->m_ext.ext_flags & EXT_FLAG_NOCACHE) != 0 ? VPR_TRYFREE : 0;
178 
179 	sf_buf_free(sf);
180 	vm_page_release(pg, flags);
181 
182 	if (m->m_ext.ext_flags & EXT_FLAG_SYNC) {
183 		struct sendfile_sync *sfs = m->m_ext.ext_arg2;
184 		sendfile_sync_signal(sfs);
185 	}
186 }
187 
188 static void
189 sendfile_free_mext_pg(struct mbuf *m)
190 {
191 	vm_page_t pg;
192 	int flags, i;
193 	bool cache_last;
194 
195 	M_ASSERTEXTPG(m);
196 
197 	cache_last = m->m_ext.ext_flags & EXT_FLAG_CACHE_LAST;
198 	flags = (m->m_ext.ext_flags & EXT_FLAG_NOCACHE) != 0 ? VPR_TRYFREE : 0;
199 
200 	for (i = 0; i < m->m_epg_npgs; i++) {
201 		if (cache_last && i == m->m_epg_npgs - 1)
202 			flags = 0;
203 		pg = PHYS_TO_VM_PAGE(m->m_epg_pa[i]);
204 		vm_page_release(pg, flags);
205 	}
206 
207 	if (m->m_ext.ext_flags & EXT_FLAG_SYNC) {
208 		struct sendfile_sync *sfs = m->m_ext.ext_arg1;
209 		sendfile_sync_signal(sfs);
210 	}
211 }
212 
213 /*
214  * Helper function to calculate how much data to put into page i of n.
215  * Only first and last pages are special.
216  */
217 static inline off_t
218 xfsize(int i, int n, off_t off, off_t len)
219 {
220 
221 	if (i == 0)
222 		return (omin(PAGE_SIZE - (off & PAGE_MASK), len));
223 
224 	if (i == n - 1 && ((off + len) & PAGE_MASK) > 0)
225 		return ((off + len) & PAGE_MASK);
226 
227 	return (PAGE_SIZE);
228 }
229 
230 /*
231  * Helper function to get offset within object for i page.
232  */
233 static inline vm_ooffset_t
234 vmoff(int i, off_t off)
235 {
236 
237 	if (i == 0)
238 		return ((vm_ooffset_t)off);
239 
240 	return (trunc_page(off + i * PAGE_SIZE));
241 }
242 
243 /*
244  * Helper function used when allocation of a page or sf_buf failed.
245  * Pretend as if we don't have enough space, subtract xfsize() of
246  * all pages that failed.
247  */
248 static inline void
249 fixspace(int old, int new, off_t off, int *space)
250 {
251 
252 	KASSERT(old > new, ("%s: old %d new %d", __func__, old, new));
253 
254 	/* Subtract last one. */
255 	*space -= xfsize(old - 1, old, off, *space);
256 	old--;
257 
258 	if (new == old)
259 		/* There was only one page. */
260 		return;
261 
262 	/* Subtract first one. */
263 	if (new == 0) {
264 		*space -= xfsize(0, old, off, *space);
265 		new++;
266 	}
267 
268 	/* Rest of pages are full sized. */
269 	*space -= (old - new) * PAGE_SIZE;
270 
271 	KASSERT(*space >= 0, ("%s: space went backwards", __func__));
272 }
273 
274 /*
275  * Wait for all in-flight ios to complete, we must not unwire pages
276  * under them.
277  */
278 static void
279 sendfile_iowait(struct sf_io *sfio, const char *wmesg)
280 {
281 	while (atomic_load_int(&sfio->nios) != 1)
282 		pause(wmesg, 1);
283 }
284 
285 /*
286  * I/O completion callback.
287  */
288 static void
289 sendfile_iodone(void *arg, vm_page_t *pa, int count, int error)
290 {
291 	struct sf_io *sfio = arg;
292 	struct socket *so;
293 	int i;
294 
295 	if (error != 0)
296 		sfio->error = error;
297 
298 	/*
299 	 * Restore the valid page pointers.  They are already
300 	 * unbusied, but still wired.
301 	 *
302 	 * XXXKIB since pages are only wired, and we do not
303 	 * own the object lock, other users might have
304 	 * invalidated them in meantime.  Similarly, after we
305 	 * unbusied the swapped-in pages, they can become
306 	 * invalid under us.
307 	 */
308 	MPASS(count == 0 || pa[0] != bogus_page);
309 	for (i = 0; i < count; i++) {
310 		if (pa[i] == bogus_page) {
311 			sfio->pa[(pa[0]->pindex - sfio->pindex0) + i] =
312 			    pa[i] = vm_page_relookup(sfio->obj,
313 			    pa[0]->pindex + i);
314 			KASSERT(pa[i] != NULL,
315 			    ("%s: page %p[%d] disappeared",
316 			    __func__, pa, i));
317 		} else {
318 			vm_page_xunbusy_unchecked(pa[i]);
319 		}
320 	}
321 
322 	if (!refcount_release(&sfio->nios))
323 		return;
324 
325 #ifdef INVARIANTS
326 	for (i = 1; i < sfio->npages; i++) {
327 		if (sfio->pa[i] == NULL)
328 			break;
329 		KASSERT(vm_page_wired(sfio->pa[i]),
330 		    ("sfio %p page %d %p not wired", sfio, i, sfio->pa[i]));
331 		if (i == 0)
332 			continue;
333 		KASSERT(sfio->pa[0]->object == sfio->pa[i]->object,
334 		    ("sfio %p page %d %p wrong owner %p %p", sfio, i,
335 		    sfio->pa[i], sfio->pa[0]->object, sfio->pa[i]->object));
336 		KASSERT(sfio->pa[0]->pindex + i == sfio->pa[i]->pindex,
337 		    ("sfio %p page %d %p wrong index %jx %jx", sfio, i,
338 		    sfio->pa[i], (uintmax_t)sfio->pa[0]->pindex,
339 		    (uintmax_t)sfio->pa[i]->pindex));
340 	}
341 #endif
342 
343 	vm_object_pip_wakeup(sfio->obj);
344 
345 	if (sfio->m == NULL) {
346 		/*
347 		 * Either I/O operation failed, or we failed to allocate
348 		 * buffers, or we bailed out on first busy page, or we
349 		 * succeeded filling the request without any I/Os. Anyway,
350 		 * pru_send hadn't been executed - nothing had been sent
351 		 * to the socket yet.
352 		 */
353 		MPASS((curthread->td_pflags & TDP_KTHREAD) == 0);
354 		free(sfio, M_SENDFILE);
355 		return;
356 	}
357 
358 #if defined(KERN_TLS) && defined(INVARIANTS)
359 	if ((sfio->m->m_flags & M_EXTPG) != 0)
360 		KASSERT(sfio->tls == sfio->m->m_epg_tls,
361 		    ("TLS session mismatch"));
362 	else
363 		KASSERT(sfio->tls == NULL,
364 		    ("non-ext_pgs mbuf with TLS session"));
365 #endif
366 	so = sfio->so;
367 	CURVNET_SET(so->so_vnet);
368 	if (__predict_false(sfio->error)) {
369 		/*
370 		 * I/O operation failed.  The state of data in the socket
371 		 * is now inconsistent, and all what we can do is to tear
372 		 * it down. Protocol abort method would tear down protocol
373 		 * state, free all ready mbufs and detach not ready ones.
374 		 * We will free the mbufs corresponding to this I/O manually.
375 		 *
376 		 * The socket would be marked with EIO and made available
377 		 * for read, so that application receives EIO on next
378 		 * syscall and eventually closes the socket.
379 		 */
380 		so->so_proto->pr_usrreqs->pru_abort(so);
381 		so->so_error = EIO;
382 
383 		mb_free_notready(sfio->m, sfio->npages);
384 #ifdef KERN_TLS
385 	} else if (sfio->tls != NULL && sfio->tls->mode == TCP_TLS_MODE_SW) {
386 		/*
387 		 * I/O operation is complete, but we still need to
388 		 * encrypt.  We cannot do this in the interrupt thread
389 		 * of the disk controller, so forward the mbufs to a
390 		 * different thread.
391 		 *
392 		 * Donate the socket reference from sfio to rather
393 		 * than explicitly invoking soref().
394 		 */
395 		ktls_enqueue(sfio->m, so, sfio->npages);
396 		goto out_with_ref;
397 #endif
398 	} else
399 		(void)(so->so_proto->pr_usrreqs->pru_ready)(so, sfio->m,
400 		    sfio->npages);
401 
402 	SOCK_LOCK(so);
403 	sorele(so);
404 #ifdef KERN_TLS
405 out_with_ref:
406 #endif
407 	CURVNET_RESTORE();
408 	free(sfio, M_SENDFILE);
409 }
410 
411 /*
412  * Iterate through pages vector and request paging for non-valid pages.
413  */
414 static int
415 sendfile_swapin(vm_object_t obj, struct sf_io *sfio, int *nios, off_t off,
416     off_t len, int rhpages, int flags)
417 {
418 	vm_page_t *pa;
419 	int a, count, count1, grabbed, i, j, npages, rv;
420 
421 	pa = sfio->pa;
422 	npages = sfio->npages;
423 	*nios = 0;
424 	flags = (flags & SF_NODISKIO) ? VM_ALLOC_NOWAIT : 0;
425 	sfio->pindex0 = OFF_TO_IDX(off);
426 
427 	/*
428 	 * First grab all the pages and wire them.  Note that we grab
429 	 * only required pages.  Readahead pages are dealt with later.
430 	 */
431 	grabbed = vm_page_grab_pages_unlocked(obj, OFF_TO_IDX(off),
432 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED | flags, pa, npages);
433 	if (grabbed < npages) {
434 		for (int i = grabbed; i < npages; i++)
435 			pa[i] = NULL;
436 		npages = grabbed;
437 		rhpages = 0;
438 	}
439 
440 	for (i = 0; i < npages;) {
441 		/* Skip valid pages. */
442 		if (vm_page_is_valid(pa[i], vmoff(i, off) & PAGE_MASK,
443 		    xfsize(i, npages, off, len))) {
444 			vm_page_xunbusy(pa[i]);
445 			SFSTAT_INC(sf_pages_valid);
446 			i++;
447 			continue;
448 		}
449 
450 		/*
451 		 * Next page is invalid.  Check if it belongs to pager.  It
452 		 * may not be there, which is a regular situation for shmem
453 		 * pager.  For vnode pager this happens only in case of
454 		 * a sparse file.
455 		 *
456 		 * Important feature of vm_pager_has_page() is the hint
457 		 * stored in 'a', about how many pages we can pagein after
458 		 * this page in a single I/O.
459 		 */
460 		VM_OBJECT_RLOCK(obj);
461 		if (!vm_pager_has_page(obj, OFF_TO_IDX(vmoff(i, off)), NULL,
462 		    &a)) {
463 			VM_OBJECT_RUNLOCK(obj);
464 			pmap_zero_page(pa[i]);
465 			vm_page_valid(pa[i]);
466 			MPASS(pa[i]->dirty == 0);
467 			vm_page_xunbusy(pa[i]);
468 			i++;
469 			continue;
470 		}
471 		VM_OBJECT_RUNLOCK(obj);
472 
473 		/*
474 		 * We want to pagein as many pages as possible, limited only
475 		 * by the 'a' hint and actual request.
476 		 */
477 		count = min(a + 1, npages - i);
478 
479 		/*
480 		 * We should not pagein into a valid page because
481 		 * there might be still unfinished write tracked by
482 		 * e.g. a buffer, thus we substitute any valid pages
483 		 * with the bogus one.
484 		 *
485 		 * We must not leave around xbusy pages which are not
486 		 * part of the run passed to vm_pager_getpages(),
487 		 * otherwise pager might deadlock waiting for the busy
488 		 * status of the page, e.g. if it constitues the
489 		 * buffer needed to validate other page.
490 		 *
491 		 * First trim the end of the run consisting of the
492 		 * valid pages, then replace the rest of the valid
493 		 * with bogus.
494 		 */
495 		count1 = count;
496 		for (j = i + count - 1; j > i; j--) {
497 			if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
498 			    xfsize(j, npages, off, len))) {
499 				vm_page_xunbusy(pa[j]);
500 				SFSTAT_INC(sf_pages_valid);
501 				count--;
502 			} else {
503 				break;
504 			}
505 		}
506 
507 		/*
508 		 * The last page in the run pa[i + count - 1] is
509 		 * guaranteed to be invalid by the trim above, so it
510 		 * is not replaced with bogus, thus -1 in the loop end
511 		 * condition.
512 		 */
513 		MPASS(pa[i + count - 1]->valid != VM_PAGE_BITS_ALL);
514 		for (j = i + 1; j < i + count - 1; j++) {
515 			if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
516 			    xfsize(j, npages, off, len))) {
517 				vm_page_xunbusy(pa[j]);
518 				SFSTAT_INC(sf_pages_valid);
519 				SFSTAT_INC(sf_pages_bogus);
520 				pa[j] = bogus_page;
521 			}
522 		}
523 
524 		refcount_acquire(&sfio->nios);
525 		rv = vm_pager_get_pages_async(obj, pa + i, count, NULL,
526 		    i + count == npages ? &rhpages : NULL,
527 		    &sendfile_iodone, sfio);
528 		if (__predict_false(rv != VM_PAGER_OK)) {
529 			sendfile_iowait(sfio, "sferrio");
530 
531 			/*
532 			 * Do remaining pages recovery before returning EIO.
533 			 * Pages from 0 to npages are wired.
534 			 * Pages from (i + count1) to npages are busied.
535 			 */
536 			for (j = 0; j < npages; j++) {
537 				if (j >= i + count1)
538 					vm_page_xunbusy(pa[j]);
539 				KASSERT(pa[j] != NULL && pa[j] != bogus_page,
540 				    ("%s: page %p[%d] I/O recovery failure",
541 				    __func__, pa, j));
542 				vm_page_unwire(pa[j], PQ_INACTIVE);
543 				pa[j] = NULL;
544 			}
545 			return (EIO);
546 		}
547 
548 		SFSTAT_INC(sf_iocnt);
549 		SFSTAT_ADD(sf_pages_read, count);
550 		if (i + count == npages)
551 			SFSTAT_ADD(sf_rhpages_read, rhpages);
552 
553 		i += count1;
554 		(*nios)++;
555 	}
556 
557 	if (*nios == 0 && npages != 0)
558 		SFSTAT_INC(sf_noiocnt);
559 
560 	return (0);
561 }
562 
563 static int
564 sendfile_getobj(struct thread *td, struct file *fp, vm_object_t *obj_res,
565     struct vnode **vp_res, struct shmfd **shmfd_res, off_t *obj_size,
566     int *bsize)
567 {
568 	struct vattr va;
569 	vm_object_t obj;
570 	struct vnode *vp;
571 	struct shmfd *shmfd;
572 	int error;
573 
574 	error = 0;
575 	vp = *vp_res = NULL;
576 	obj = NULL;
577 	shmfd = *shmfd_res = NULL;
578 	*bsize = 0;
579 
580 	/*
581 	 * The file descriptor must be a regular file and have a
582 	 * backing VM object.
583 	 */
584 	if (fp->f_type == DTYPE_VNODE) {
585 		vp = fp->f_vnode;
586 		vn_lock(vp, LK_SHARED | LK_RETRY);
587 		if (vp->v_type != VREG) {
588 			error = EINVAL;
589 			goto out;
590 		}
591 		*bsize = vp->v_mount->mnt_stat.f_iosize;
592 		obj = vp->v_object;
593 		if (obj == NULL) {
594 			error = EINVAL;
595 			goto out;
596 		}
597 
598 		/*
599 		 * Use the pager size when available to simplify synchronization
600 		 * with filesystems, which otherwise must atomically update both
601 		 * the vnode pager size and file size.
602 		 */
603 		if (obj->type == OBJT_VNODE) {
604 			VM_OBJECT_RLOCK(obj);
605 			*obj_size = obj->un_pager.vnp.vnp_size;
606 		} else {
607 			error = VOP_GETATTR(vp, &va, td->td_ucred);
608 			if (error != 0)
609 				goto out;
610 			*obj_size = va.va_size;
611 			VM_OBJECT_RLOCK(obj);
612 		}
613 	} else if (fp->f_type == DTYPE_SHM) {
614 		shmfd = fp->f_data;
615 		obj = shmfd->shm_object;
616 		VM_OBJECT_RLOCK(obj);
617 		*obj_size = shmfd->shm_size;
618 	} else {
619 		error = EINVAL;
620 		goto out;
621 	}
622 
623 	if ((obj->flags & OBJ_DEAD) != 0) {
624 		VM_OBJECT_RUNLOCK(obj);
625 		error = EBADF;
626 		goto out;
627 	}
628 
629 	/*
630 	 * Temporarily increase the backing VM object's reference
631 	 * count so that a forced reclamation of its vnode does not
632 	 * immediately destroy it.
633 	 */
634 	vm_object_reference_locked(obj);
635 	VM_OBJECT_RUNLOCK(obj);
636 	*obj_res = obj;
637 	*vp_res = vp;
638 	*shmfd_res = shmfd;
639 
640 out:
641 	if (vp != NULL)
642 		VOP_UNLOCK(vp);
643 	return (error);
644 }
645 
646 static int
647 sendfile_getsock(struct thread *td, int s, struct file **sock_fp,
648     struct socket **so)
649 {
650 	int error;
651 
652 	*sock_fp = NULL;
653 	*so = NULL;
654 
655 	/*
656 	 * The socket must be a stream socket and connected.
657 	 */
658 	error = getsock_cap(td, s, &cap_send_rights,
659 	    sock_fp, NULL, NULL);
660 	if (error != 0)
661 		return (error);
662 	*so = (*sock_fp)->f_data;
663 	if ((*so)->so_type != SOCK_STREAM)
664 		return (EINVAL);
665 	/*
666 	 * SCTP one-to-one style sockets currently don't work with
667 	 * sendfile(). So indicate EINVAL for now.
668 	 */
669 	if ((*so)->so_proto->pr_protocol == IPPROTO_SCTP)
670 		return (EINVAL);
671 	if (SOLISTENING(*so))
672 		return (ENOTCONN);
673 	return (0);
674 }
675 
676 int
677 vn_sendfile(struct file *fp, int sockfd, struct uio *hdr_uio,
678     struct uio *trl_uio, off_t offset, size_t nbytes, off_t *sent, int flags,
679     struct thread *td)
680 {
681 	struct file *sock_fp;
682 	struct vnode *vp;
683 	struct vm_object *obj;
684 	vm_page_t pga;
685 	struct socket *so;
686 #ifdef KERN_TLS
687 	struct ktls_session *tls;
688 #endif
689 	struct mbuf *m, *mh, *mhtail;
690 	struct sf_buf *sf;
691 	struct shmfd *shmfd;
692 	struct sendfile_sync *sfs;
693 	struct vattr va;
694 	off_t off, sbytes, rem, obj_size, nobj_size;
695 	int bsize, error, ext_pgs_idx, hdrlen, max_pgs, softerr;
696 #ifdef KERN_TLS
697 	int tls_enq_cnt;
698 #endif
699 	bool use_ext_pgs;
700 
701 	obj = NULL;
702 	so = NULL;
703 	m = mh = NULL;
704 	sfs = NULL;
705 #ifdef KERN_TLS
706 	tls = NULL;
707 #endif
708 	hdrlen = sbytes = 0;
709 	softerr = 0;
710 	use_ext_pgs = false;
711 
712 	error = sendfile_getobj(td, fp, &obj, &vp, &shmfd, &obj_size, &bsize);
713 	if (error != 0)
714 		return (error);
715 
716 	error = sendfile_getsock(td, sockfd, &sock_fp, &so);
717 	if (error != 0)
718 		goto out;
719 
720 #ifdef MAC
721 	error = mac_socket_check_send(td->td_ucred, so);
722 	if (error != 0)
723 		goto out;
724 #endif
725 
726 	SFSTAT_INC(sf_syscalls);
727 	SFSTAT_ADD(sf_rhpages_requested, SF_READAHEAD(flags));
728 
729 	if (flags & SF_SYNC) {
730 		sfs = malloc(sizeof(*sfs), M_SENDFILE, M_WAITOK | M_ZERO);
731 		mtx_init(&sfs->mtx, "sendfile", NULL, MTX_DEF);
732 		cv_init(&sfs->cv, "sendfile");
733 		sfs->waiting = true;
734 	}
735 
736 	rem = nbytes ? omin(nbytes, obj_size - offset) : obj_size - offset;
737 
738 	/*
739 	 * Protect against multiple writers to the socket.
740 	 *
741 	 * XXXRW: Historically this has assumed non-interruptibility, so now
742 	 * we implement that, but possibly shouldn't.
743 	 */
744 	(void)sblock(&so->so_snd, SBL_WAIT | SBL_NOINTR);
745 #ifdef KERN_TLS
746 	tls = ktls_hold(so->so_snd.sb_tls_info);
747 #endif
748 
749 	/*
750 	 * Loop through the pages of the file, starting with the requested
751 	 * offset. Get a file page (do I/O if necessary), map the file page
752 	 * into an sf_buf, attach an mbuf header to the sf_buf, and queue
753 	 * it on the socket.
754 	 * This is done in two loops.  The inner loop turns as many pages
755 	 * as it can, up to available socket buffer space, without blocking
756 	 * into mbufs to have it bulk delivered into the socket send buffer.
757 	 * The outer loop checks the state and available space of the socket
758 	 * and takes care of the overall progress.
759 	 */
760 	for (off = offset; rem > 0; ) {
761 		struct sf_io *sfio;
762 		vm_page_t *pa;
763 		struct mbuf *m0, *mtail;
764 		int nios, space, npages, rhpages;
765 
766 		mtail = NULL;
767 		/*
768 		 * Check the socket state for ongoing connection,
769 		 * no errors and space in socket buffer.
770 		 * If space is low allow for the remainder of the
771 		 * file to be processed if it fits the socket buffer.
772 		 * Otherwise block in waiting for sufficient space
773 		 * to proceed, or if the socket is nonblocking, return
774 		 * to userland with EAGAIN while reporting how far
775 		 * we've come.
776 		 * We wait until the socket buffer has significant free
777 		 * space to do bulk sends.  This makes good use of file
778 		 * system read ahead and allows packet segmentation
779 		 * offloading hardware to take over lots of work.  If
780 		 * we were not careful here we would send off only one
781 		 * sfbuf at a time.
782 		 */
783 		SOCKBUF_LOCK(&so->so_snd);
784 		if (so->so_snd.sb_lowat < so->so_snd.sb_hiwat / 2)
785 			so->so_snd.sb_lowat = so->so_snd.sb_hiwat / 2;
786 retry_space:
787 		if (so->so_snd.sb_state & SBS_CANTSENDMORE) {
788 			error = EPIPE;
789 			SOCKBUF_UNLOCK(&so->so_snd);
790 			goto done;
791 		} else if (so->so_error) {
792 			error = so->so_error;
793 			so->so_error = 0;
794 			SOCKBUF_UNLOCK(&so->so_snd);
795 			goto done;
796 		}
797 		if ((so->so_state & SS_ISCONNECTED) == 0) {
798 			SOCKBUF_UNLOCK(&so->so_snd);
799 			error = ENOTCONN;
800 			goto done;
801 		}
802 
803 		space = sbspace(&so->so_snd);
804 		if (space < rem &&
805 		    (space <= 0 ||
806 		     space < so->so_snd.sb_lowat)) {
807 			if (so->so_state & SS_NBIO) {
808 				SOCKBUF_UNLOCK(&so->so_snd);
809 				error = EAGAIN;
810 				goto done;
811 			}
812 			/*
813 			 * sbwait drops the lock while sleeping.
814 			 * When we loop back to retry_space the
815 			 * state may have changed and we retest
816 			 * for it.
817 			 */
818 			error = sbwait(&so->so_snd);
819 			/*
820 			 * An error from sbwait usually indicates that we've
821 			 * been interrupted by a signal. If we've sent anything
822 			 * then return bytes sent, otherwise return the error.
823 			 */
824 			if (error != 0) {
825 				SOCKBUF_UNLOCK(&so->so_snd);
826 				goto done;
827 			}
828 			goto retry_space;
829 		}
830 		SOCKBUF_UNLOCK(&so->so_snd);
831 
832 		/*
833 		 * At the beginning of the first loop check if any headers
834 		 * are specified and copy them into mbufs.  Reduce space in
835 		 * the socket buffer by the size of the header mbuf chain.
836 		 * Clear hdr_uio here and hdrlen at the end of the first loop.
837 		 */
838 		if (hdr_uio != NULL && hdr_uio->uio_resid > 0) {
839 			hdr_uio->uio_td = td;
840 			hdr_uio->uio_rw = UIO_WRITE;
841 #ifdef KERN_TLS
842 			if (tls != NULL)
843 				mh = m_uiotombuf(hdr_uio, M_WAITOK, space,
844 				    tls->params.max_frame_len, M_EXTPG);
845 			else
846 #endif
847 				mh = m_uiotombuf(hdr_uio, M_WAITOK,
848 				    space, 0, 0);
849 			hdrlen = m_length(mh, &mhtail);
850 			space -= hdrlen;
851 			/*
852 			 * If header consumed all the socket buffer space,
853 			 * don't waste CPU cycles and jump to the end.
854 			 */
855 			if (space == 0) {
856 				sfio = NULL;
857 				nios = 0;
858 				goto prepend_header;
859 			}
860 			hdr_uio = NULL;
861 		}
862 
863 		if (vp != NULL) {
864 			error = vn_lock(vp, LK_SHARED);
865 			if (error != 0)
866 				goto done;
867 
868 			/*
869 			 * Check to see if the file size has changed.
870 			 */
871 			if (obj->type == OBJT_VNODE) {
872 				VM_OBJECT_RLOCK(obj);
873 				nobj_size = obj->un_pager.vnp.vnp_size;
874 				VM_OBJECT_RUNLOCK(obj);
875 			} else {
876 				error = VOP_GETATTR(vp, &va, td->td_ucred);
877 				if (error != 0) {
878 					VOP_UNLOCK(vp);
879 					goto done;
880 				}
881 				nobj_size = va.va_size;
882 			}
883 			if (off >= nobj_size) {
884 				VOP_UNLOCK(vp);
885 				goto done;
886 			}
887 			if (nobj_size != obj_size) {
888 				obj_size = nobj_size;
889 				rem = nbytes ? omin(nbytes + offset, obj_size) :
890 				    obj_size;
891 				rem -= off;
892 			}
893 		}
894 
895 		if (space > rem)
896 			space = rem;
897 		else if (space > PAGE_SIZE) {
898 			/*
899 			 * Use page boundaries when possible for large
900 			 * requests.
901 			 */
902 			if (off & PAGE_MASK)
903 				space -= (PAGE_SIZE - (off & PAGE_MASK));
904 			space = trunc_page(space);
905 			if (off & PAGE_MASK)
906 				space += (PAGE_SIZE - (off & PAGE_MASK));
907 		}
908 
909 		npages = howmany(space + (off & PAGE_MASK), PAGE_SIZE);
910 
911 		/*
912 		 * Calculate maximum allowed number of pages for readahead
913 		 * at this iteration.  If SF_USER_READAHEAD was set, we don't
914 		 * do any heuristics and use exactly the value supplied by
915 		 * application.  Otherwise, we allow readahead up to "rem".
916 		 * If application wants more, let it be, but there is no
917 		 * reason to go above maxphys.  Also check against "obj_size",
918 		 * since vm_pager_has_page() can hint beyond EOF.
919 		 */
920 		if (flags & SF_USER_READAHEAD) {
921 			rhpages = SF_READAHEAD(flags);
922 		} else {
923 			rhpages = howmany(rem + (off & PAGE_MASK), PAGE_SIZE) -
924 			    npages;
925 			rhpages += SF_READAHEAD(flags);
926 		}
927 		rhpages = min(howmany(maxphys, PAGE_SIZE), rhpages);
928 		rhpages = min(howmany(obj_size - trunc_page(off), PAGE_SIZE) -
929 		    npages, rhpages);
930 
931 		sfio = malloc(sizeof(struct sf_io) +
932 		    npages * sizeof(vm_page_t), M_SENDFILE, M_WAITOK);
933 		refcount_init(&sfio->nios, 1);
934 		sfio->obj = obj;
935 		sfio->error = 0;
936 		sfio->m = NULL;
937 		sfio->npages = npages;
938 #ifdef KERN_TLS
939 		/*
940 		 * This doesn't use ktls_hold() because sfio->m will
941 		 * also have a reference on 'tls' that will be valid
942 		 * for all of sfio's lifetime.
943 		 */
944 		sfio->tls = tls;
945 #endif
946 		vm_object_pip_add(obj, 1);
947 		error = sendfile_swapin(obj, sfio, &nios, off, space, rhpages,
948 		    flags);
949 		if (error != 0) {
950 			if (vp != NULL)
951 				VOP_UNLOCK(vp);
952 			sendfile_iodone(sfio, NULL, 0, error);
953 			goto done;
954 		}
955 
956 		/*
957 		 * Loop and construct maximum sized mbuf chain to be bulk
958 		 * dumped into socket buffer.
959 		 */
960 		pa = sfio->pa;
961 
962 		/*
963 		 * Use unmapped mbufs if enabled for TCP.  Unmapped
964 		 * bufs are restricted to TCP as that is what has been
965 		 * tested.  In particular, unmapped mbufs have not
966 		 * been tested with UNIX-domain sockets.
967 		 *
968 		 * TLS frames always require unmapped mbufs.
969 		 */
970 		if ((mb_use_ext_pgs &&
971 		    so->so_proto->pr_protocol == IPPROTO_TCP)
972 #ifdef KERN_TLS
973 		    || tls != NULL
974 #endif
975 		    ) {
976 			use_ext_pgs = true;
977 #ifdef KERN_TLS
978 			if (tls != NULL)
979 				max_pgs = num_pages(tls->params.max_frame_len);
980 			else
981 #endif
982 				max_pgs = MBUF_PEXT_MAX_PGS;
983 
984 			/* Start at last index, to wrap on first use. */
985 			ext_pgs_idx = max_pgs - 1;
986 		}
987 
988 		for (int i = 0; i < npages; i++) {
989 			/*
990 			 * If a page wasn't grabbed successfully, then
991 			 * trim the array. Can happen only with SF_NODISKIO.
992 			 */
993 			if (pa[i] == NULL) {
994 				SFSTAT_INC(sf_busy);
995 				fixspace(npages, i, off, &space);
996 				sfio->npages = i;
997 				softerr = EBUSY;
998 				break;
999 			}
1000 			pga = pa[i];
1001 			if (pga == bogus_page)
1002 				pga = vm_page_relookup(obj, sfio->pindex0 + i);
1003 
1004 			if (use_ext_pgs) {
1005 				off_t xfs;
1006 
1007 				ext_pgs_idx++;
1008 				if (ext_pgs_idx == max_pgs) {
1009 					m0 = mb_alloc_ext_pgs(M_WAITOK,
1010 					    sendfile_free_mext_pg);
1011 
1012 					if (flags & SF_NOCACHE) {
1013 						m0->m_ext.ext_flags |=
1014 						    EXT_FLAG_NOCACHE;
1015 
1016 						/*
1017 						 * See comment below regarding
1018 						 * ignoring SF_NOCACHE for the
1019 						 * last page.
1020 						 */
1021 						if ((npages - i <= max_pgs) &&
1022 						    ((off + space) & PAGE_MASK) &&
1023 						    (rem > space || rhpages > 0))
1024 							m0->m_ext.ext_flags |=
1025 							    EXT_FLAG_CACHE_LAST;
1026 					}
1027 					if (sfs != NULL) {
1028 						m0->m_ext.ext_flags |=
1029 						    EXT_FLAG_SYNC;
1030 						m0->m_ext.ext_arg1 = sfs;
1031 						mtx_lock(&sfs->mtx);
1032 						sfs->count++;
1033 						mtx_unlock(&sfs->mtx);
1034 					}
1035 					ext_pgs_idx = 0;
1036 
1037 					/* Append to mbuf chain. */
1038 					if (mtail != NULL)
1039 						mtail->m_next = m0;
1040 					else
1041 						m = m0;
1042 					mtail = m0;
1043 					m0->m_epg_1st_off =
1044 					    vmoff(i, off) & PAGE_MASK;
1045 				}
1046 				if (nios) {
1047 					mtail->m_flags |= M_NOTREADY;
1048 					m0->m_epg_nrdy++;
1049 				}
1050 
1051 				m0->m_epg_pa[ext_pgs_idx] = VM_PAGE_TO_PHYS(pga);
1052 				m0->m_epg_npgs++;
1053 				xfs = xfsize(i, npages, off, space);
1054 				m0->m_epg_last_len = xfs;
1055 				MBUF_EXT_PGS_ASSERT_SANITY(m0);
1056 				mtail->m_len += xfs;
1057 				mtail->m_ext.ext_size += PAGE_SIZE;
1058 				continue;
1059 			}
1060 
1061 			/*
1062 			 * Get a sendfile buf.  When allocating the
1063 			 * first buffer for mbuf chain, we usually
1064 			 * wait as long as necessary, but this wait
1065 			 * can be interrupted.  For consequent
1066 			 * buffers, do not sleep, since several
1067 			 * threads might exhaust the buffers and then
1068 			 * deadlock.
1069 			 */
1070 			sf = sf_buf_alloc(pga,
1071 			    m != NULL ? SFB_NOWAIT : SFB_CATCH);
1072 			if (sf == NULL) {
1073 				SFSTAT_INC(sf_allocfail);
1074 				sendfile_iowait(sfio, "sfnosf");
1075 				for (int j = i; j < npages; j++) {
1076 					vm_page_unwire(pa[j], PQ_INACTIVE);
1077 					pa[j] = NULL;
1078 				}
1079 				if (m == NULL)
1080 					softerr = ENOBUFS;
1081 				fixspace(npages, i, off, &space);
1082 				sfio->npages = i;
1083 				break;
1084 			}
1085 
1086 			m0 = m_get(M_WAITOK, MT_DATA);
1087 			m0->m_ext.ext_buf = (char *)sf_buf_kva(sf);
1088 			m0->m_ext.ext_size = PAGE_SIZE;
1089 			m0->m_ext.ext_arg1 = sf;
1090 			m0->m_ext.ext_type = EXT_SFBUF;
1091 			m0->m_ext.ext_flags = EXT_FLAG_EMBREF;
1092 			m0->m_ext.ext_free = sendfile_free_mext;
1093 			/*
1094 			 * SF_NOCACHE sets the page as being freed upon send.
1095 			 * However, we ignore it for the last page in 'space',
1096 			 * if the page is truncated, and we got more data to
1097 			 * send (rem > space), or if we have readahead
1098 			 * configured (rhpages > 0).
1099 			 */
1100 			if ((flags & SF_NOCACHE) &&
1101 			    (i != npages - 1 ||
1102 			    !((off + space) & PAGE_MASK) ||
1103 			    !(rem > space || rhpages > 0)))
1104 				m0->m_ext.ext_flags |= EXT_FLAG_NOCACHE;
1105 			if (sfs != NULL) {
1106 				m0->m_ext.ext_flags |= EXT_FLAG_SYNC;
1107 				m0->m_ext.ext_arg2 = sfs;
1108 				mtx_lock(&sfs->mtx);
1109 				sfs->count++;
1110 				mtx_unlock(&sfs->mtx);
1111 			}
1112 			m0->m_ext.ext_count = 1;
1113 			m0->m_flags |= (M_EXT | M_RDONLY);
1114 			if (nios)
1115 				m0->m_flags |= M_NOTREADY;
1116 			m0->m_data = (char *)sf_buf_kva(sf) +
1117 			    (vmoff(i, off) & PAGE_MASK);
1118 			m0->m_len = xfsize(i, npages, off, space);
1119 
1120 			/* Append to mbuf chain. */
1121 			if (mtail != NULL)
1122 				mtail->m_next = m0;
1123 			else
1124 				m = m0;
1125 			mtail = m0;
1126 		}
1127 
1128 		if (vp != NULL)
1129 			VOP_UNLOCK(vp);
1130 
1131 		/* Keep track of bytes processed. */
1132 		off += space;
1133 		rem -= space;
1134 
1135 		/*
1136 		 * Prepend header, if any.  Save pointer to first mbuf
1137 		 * with a page.
1138 		 */
1139 		if (hdrlen) {
1140 prepend_header:
1141 			m0 = mhtail->m_next = m;
1142 			m = mh;
1143 			mh = NULL;
1144 		} else
1145 			m0 = m;
1146 
1147 		if (m == NULL) {
1148 			KASSERT(softerr, ("%s: m NULL, no error", __func__));
1149 			error = softerr;
1150 			sendfile_iodone(sfio, NULL, 0, 0);
1151 			goto done;
1152 		}
1153 
1154 		/* Add the buffer chain to the socket buffer. */
1155 		KASSERT(m_length(m, NULL) == space + hdrlen,
1156 		    ("%s: mlen %u space %d hdrlen %d",
1157 		    __func__, m_length(m, NULL), space, hdrlen));
1158 
1159 		CURVNET_SET(so->so_vnet);
1160 #ifdef KERN_TLS
1161 		if (tls != NULL)
1162 			ktls_frame(m, tls, &tls_enq_cnt, TLS_RLTYPE_APP);
1163 #endif
1164 		if (nios == 0) {
1165 			/*
1166 			 * If sendfile_swapin() didn't initiate any I/Os,
1167 			 * which happens if all data is cached in VM, or if
1168 			 * the header consumed all socket buffer space and
1169 			 * sfio is NULL, then we can send data right now
1170 			 * without the PRUS_NOTREADY flag.
1171 			 */
1172 			if (sfio != NULL)
1173 				sendfile_iodone(sfio, NULL, 0, 0);
1174 #ifdef KERN_TLS
1175 			if (tls != NULL && tls->mode == TCP_TLS_MODE_SW) {
1176 				error = (*so->so_proto->pr_usrreqs->pru_send)
1177 				    (so, PRUS_NOTREADY, m, NULL, NULL, td);
1178 				if (error != 0) {
1179 					m_freem(m);
1180 				} else {
1181 					soref(so);
1182 					ktls_enqueue(m, so, tls_enq_cnt);
1183 				}
1184 			} else
1185 #endif
1186 				error = (*so->so_proto->pr_usrreqs->pru_send)
1187 				    (so, 0, m, NULL, NULL, td);
1188 		} else {
1189 			sfio->so = so;
1190 			sfio->m = m0;
1191 			soref(so);
1192 			error = (*so->so_proto->pr_usrreqs->pru_send)
1193 			    (so, PRUS_NOTREADY, m, NULL, NULL, td);
1194 			sendfile_iodone(sfio, NULL, 0, error);
1195 		}
1196 		CURVNET_RESTORE();
1197 
1198 		m = NULL;
1199 		if (error)
1200 			goto done;
1201 		sbytes += space + hdrlen;
1202 		if (hdrlen)
1203 			hdrlen = 0;
1204 		if (softerr) {
1205 			error = softerr;
1206 			goto done;
1207 		}
1208 	}
1209 
1210 	/*
1211 	 * Send trailers. Wimp out and use writev(2).
1212 	 */
1213 	if (trl_uio != NULL) {
1214 		sbunlock(&so->so_snd);
1215 		error = kern_writev(td, sockfd, trl_uio);
1216 		if (error == 0)
1217 			sbytes += td->td_retval[0];
1218 		goto out;
1219 	}
1220 
1221 done:
1222 	sbunlock(&so->so_snd);
1223 out:
1224 	/*
1225 	 * If there was no error we have to clear td->td_retval[0]
1226 	 * because it may have been set by writev.
1227 	 */
1228 	if (error == 0) {
1229 		td->td_retval[0] = 0;
1230 	}
1231 	if (sent != NULL) {
1232 		(*sent) = sbytes;
1233 	}
1234 	if (obj != NULL)
1235 		vm_object_deallocate(obj);
1236 	if (so)
1237 		fdrop(sock_fp, td);
1238 	if (m)
1239 		m_freem(m);
1240 	if (mh)
1241 		m_freem(mh);
1242 
1243 	if (sfs != NULL) {
1244 		mtx_lock(&sfs->mtx);
1245 		if (sfs->count != 0)
1246 			error = cv_wait_sig(&sfs->cv, &sfs->mtx);
1247 		if (sfs->count == 0) {
1248 			sendfile_sync_destroy(sfs);
1249 		} else {
1250 			sfs->waiting = false;
1251 			mtx_unlock(&sfs->mtx);
1252 		}
1253 	}
1254 #ifdef KERN_TLS
1255 	if (tls != NULL)
1256 		ktls_free(tls);
1257 #endif
1258 
1259 	if (error == ERESTART)
1260 		error = EINTR;
1261 
1262 	return (error);
1263 }
1264 
1265 static int
1266 sendfile(struct thread *td, struct sendfile_args *uap, int compat)
1267 {
1268 	struct sf_hdtr hdtr;
1269 	struct uio *hdr_uio, *trl_uio;
1270 	struct file *fp;
1271 	off_t sbytes;
1272 	int error;
1273 
1274 	/*
1275 	 * File offset must be positive.  If it goes beyond EOF
1276 	 * we send only the header/trailer and no payload data.
1277 	 */
1278 	if (uap->offset < 0)
1279 		return (EINVAL);
1280 
1281 	sbytes = 0;
1282 	hdr_uio = trl_uio = NULL;
1283 
1284 	if (uap->hdtr != NULL) {
1285 		error = copyin(uap->hdtr, &hdtr, sizeof(hdtr));
1286 		if (error != 0)
1287 			goto out;
1288 		if (hdtr.headers != NULL) {
1289 			error = copyinuio(hdtr.headers, hdtr.hdr_cnt,
1290 			    &hdr_uio);
1291 			if (error != 0)
1292 				goto out;
1293 #ifdef COMPAT_FREEBSD4
1294 			/*
1295 			 * In FreeBSD < 5.0 the nbytes to send also included
1296 			 * the header.  If compat is specified subtract the
1297 			 * header size from nbytes.
1298 			 */
1299 			if (compat) {
1300 				if (uap->nbytes > hdr_uio->uio_resid)
1301 					uap->nbytes -= hdr_uio->uio_resid;
1302 				else
1303 					uap->nbytes = 0;
1304 			}
1305 #endif
1306 		}
1307 		if (hdtr.trailers != NULL) {
1308 			error = copyinuio(hdtr.trailers, hdtr.trl_cnt,
1309 			    &trl_uio);
1310 			if (error != 0)
1311 				goto out;
1312 		}
1313 	}
1314 
1315 	AUDIT_ARG_FD(uap->fd);
1316 
1317 	/*
1318 	 * sendfile(2) can start at any offset within a file so we require
1319 	 * CAP_READ+CAP_SEEK = CAP_PREAD.
1320 	 */
1321 	if ((error = fget_read(td, uap->fd, &cap_pread_rights, &fp)) != 0)
1322 		goto out;
1323 
1324 	error = fo_sendfile(fp, uap->s, hdr_uio, trl_uio, uap->offset,
1325 	    uap->nbytes, &sbytes, uap->flags, td);
1326 	fdrop(fp, td);
1327 
1328 	if (uap->sbytes != NULL)
1329 		copyout(&sbytes, uap->sbytes, sizeof(off_t));
1330 
1331 out:
1332 	free(hdr_uio, M_IOV);
1333 	free(trl_uio, M_IOV);
1334 	return (error);
1335 }
1336 
1337 /*
1338  * sendfile(2)
1339  *
1340  * int sendfile(int fd, int s, off_t offset, size_t nbytes,
1341  *       struct sf_hdtr *hdtr, off_t *sbytes, int flags)
1342  *
1343  * Send a file specified by 'fd' and starting at 'offset' to a socket
1344  * specified by 's'. Send only 'nbytes' of the file or until EOF if nbytes ==
1345  * 0.  Optionally add a header and/or trailer to the socket output.  If
1346  * specified, write the total number of bytes sent into *sbytes.
1347  */
1348 int
1349 sys_sendfile(struct thread *td, struct sendfile_args *uap)
1350 {
1351 
1352 	return (sendfile(td, uap, 0));
1353 }
1354 
1355 #ifdef COMPAT_FREEBSD4
1356 int
1357 freebsd4_sendfile(struct thread *td, struct freebsd4_sendfile_args *uap)
1358 {
1359 	struct sendfile_args args;
1360 
1361 	args.fd = uap->fd;
1362 	args.s = uap->s;
1363 	args.offset = uap->offset;
1364 	args.nbytes = uap->nbytes;
1365 	args.hdtr = uap->hdtr;
1366 	args.sbytes = uap->sbytes;
1367 	args.flags = uap->flags;
1368 
1369 	return (sendfile(td, &args, 1));
1370 }
1371 #endif /* COMPAT_FREEBSD4 */
1372