xref: /freebsd/sys/kern/kern_sendfile.c (revision f374ba41f55c1a127303d92d830dd58eef2f5243)
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_MPSAFE, 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_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_ready(so, sfio->m, sfio->npages);
400 
401 	sorele(so);
402 #ifdef KERN_TLS
403 out_with_ref:
404 #endif
405 	CURVNET_RESTORE();
406 	free(sfio, M_SENDFILE);
407 }
408 
409 /*
410  * Iterate through pages vector and request paging for non-valid pages.
411  */
412 static int
413 sendfile_swapin(vm_object_t obj, struct sf_io *sfio, int *nios, off_t off,
414     off_t len, int rhpages, int flags)
415 {
416 	vm_page_t *pa;
417 	int a, count, count1, grabbed, i, j, npages, rv;
418 
419 	pa = sfio->pa;
420 	npages = sfio->npages;
421 	*nios = 0;
422 	flags = (flags & SF_NODISKIO) ? VM_ALLOC_NOWAIT : 0;
423 	sfio->pindex0 = OFF_TO_IDX(off);
424 
425 	/*
426 	 * First grab all the pages and wire them.  Note that we grab
427 	 * only required pages.  Readahead pages are dealt with later.
428 	 */
429 	grabbed = vm_page_grab_pages_unlocked(obj, OFF_TO_IDX(off),
430 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED | flags, pa, npages);
431 	if (grabbed < npages) {
432 		for (int i = grabbed; i < npages; i++)
433 			pa[i] = NULL;
434 		npages = grabbed;
435 		rhpages = 0;
436 	}
437 
438 	for (i = 0; i < npages;) {
439 		/* Skip valid pages. */
440 		if (vm_page_is_valid(pa[i], vmoff(i, off) & PAGE_MASK,
441 		    xfsize(i, npages, off, len))) {
442 			vm_page_xunbusy(pa[i]);
443 			SFSTAT_INC(sf_pages_valid);
444 			i++;
445 			continue;
446 		}
447 
448 		/*
449 		 * Next page is invalid.  Check if it belongs to pager.  It
450 		 * may not be there, which is a regular situation for shmem
451 		 * pager.  For vnode pager this happens only in case of
452 		 * a sparse file.
453 		 *
454 		 * Important feature of vm_pager_has_page() is the hint
455 		 * stored in 'a', about how many pages we can pagein after
456 		 * this page in a single I/O.
457 		 */
458 		VM_OBJECT_RLOCK(obj);
459 		if (!vm_pager_has_page(obj, OFF_TO_IDX(vmoff(i, off)), NULL,
460 		    &a)) {
461 			VM_OBJECT_RUNLOCK(obj);
462 			pmap_zero_page(pa[i]);
463 			vm_page_valid(pa[i]);
464 			MPASS(pa[i]->dirty == 0);
465 			vm_page_xunbusy(pa[i]);
466 			i++;
467 			continue;
468 		}
469 		VM_OBJECT_RUNLOCK(obj);
470 
471 		/*
472 		 * We want to pagein as many pages as possible, limited only
473 		 * by the 'a' hint and actual request.
474 		 */
475 		count = min(a + 1, npages - i);
476 
477 		/*
478 		 * We should not pagein into a valid page because
479 		 * there might be still unfinished write tracked by
480 		 * e.g. a buffer, thus we substitute any valid pages
481 		 * with the bogus one.
482 		 *
483 		 * We must not leave around xbusy pages which are not
484 		 * part of the run passed to vm_pager_getpages(),
485 		 * otherwise pager might deadlock waiting for the busy
486 		 * status of the page, e.g. if it constitues the
487 		 * buffer needed to validate other page.
488 		 *
489 		 * First trim the end of the run consisting of the
490 		 * valid pages, then replace the rest of the valid
491 		 * with bogus.
492 		 */
493 		count1 = count;
494 		for (j = i + count - 1; j > i; j--) {
495 			if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
496 			    xfsize(j, npages, off, len))) {
497 				vm_page_xunbusy(pa[j]);
498 				SFSTAT_INC(sf_pages_valid);
499 				count--;
500 			} else {
501 				break;
502 			}
503 		}
504 
505 		/*
506 		 * The last page in the run pa[i + count - 1] is
507 		 * guaranteed to be invalid by the trim above, so it
508 		 * is not replaced with bogus, thus -1 in the loop end
509 		 * condition.
510 		 */
511 		MPASS(pa[i + count - 1]->valid != VM_PAGE_BITS_ALL);
512 		for (j = i + 1; j < i + count - 1; j++) {
513 			if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
514 			    xfsize(j, npages, off, len))) {
515 				vm_page_xunbusy(pa[j]);
516 				SFSTAT_INC(sf_pages_valid);
517 				SFSTAT_INC(sf_pages_bogus);
518 				pa[j] = bogus_page;
519 			}
520 		}
521 
522 		refcount_acquire(&sfio->nios);
523 		rv = vm_pager_get_pages_async(obj, pa + i, count, NULL,
524 		    i + count == npages ? &rhpages : NULL,
525 		    &sendfile_iodone, sfio);
526 		if (__predict_false(rv != VM_PAGER_OK)) {
527 			sendfile_iowait(sfio, "sferrio");
528 
529 			/*
530 			 * Do remaining pages recovery before returning EIO.
531 			 * Pages from 0 to npages are wired.
532 			 * Pages from (i + count1) to npages are busied.
533 			 */
534 			for (j = 0; j < npages; j++) {
535 				if (j >= i + count1)
536 					vm_page_xunbusy(pa[j]);
537 				KASSERT(pa[j] != NULL && pa[j] != bogus_page,
538 				    ("%s: page %p[%d] I/O recovery failure",
539 				    __func__, pa, j));
540 				vm_page_unwire(pa[j], PQ_INACTIVE);
541 				pa[j] = NULL;
542 			}
543 			return (EIO);
544 		}
545 
546 		SFSTAT_INC(sf_iocnt);
547 		SFSTAT_ADD(sf_pages_read, count);
548 		if (i + count == npages)
549 			SFSTAT_ADD(sf_rhpages_read, rhpages);
550 
551 		i += count1;
552 		(*nios)++;
553 	}
554 
555 	if (*nios == 0 && npages != 0)
556 		SFSTAT_INC(sf_noiocnt);
557 
558 	return (0);
559 }
560 
561 static int
562 sendfile_getobj(struct thread *td, struct file *fp, vm_object_t *obj_res,
563     struct vnode **vp_res, struct shmfd **shmfd_res, off_t *obj_size,
564     int *bsize)
565 {
566 	vm_object_t obj;
567 	struct vnode *vp;
568 	struct shmfd *shmfd;
569 	int error;
570 
571 	error = 0;
572 	vp = *vp_res = NULL;
573 	obj = NULL;
574 	shmfd = *shmfd_res = NULL;
575 	*bsize = 0;
576 
577 	/*
578 	 * The file descriptor must be a regular file and have a
579 	 * backing VM object.
580 	 */
581 	if (fp->f_type == DTYPE_VNODE) {
582 		vp = fp->f_vnode;
583 		vn_lock(vp, LK_SHARED | LK_RETRY);
584 		if (vp->v_type != VREG) {
585 			error = EINVAL;
586 			goto out;
587 		}
588 		*bsize = vp->v_mount->mnt_stat.f_iosize;
589 		obj = vp->v_object;
590 		if (obj == NULL) {
591 			error = EINVAL;
592 			goto out;
593 		}
594 
595 		/*
596 		 * Use the pager size when available to simplify synchronization
597 		 * with filesystems, which otherwise must atomically update both
598 		 * the vnode pager size and file size.
599 		 */
600 		if (obj->type == OBJT_VNODE) {
601 			VM_OBJECT_RLOCK(obj);
602 			*obj_size = obj->un_pager.vnp.vnp_size;
603 		} else {
604 			error = vn_getsize_locked(vp, obj_size, td->td_ucred);
605 			if (error != 0)
606 				goto out;
607 			VM_OBJECT_RLOCK(obj);
608 		}
609 	} else if (fp->f_type == DTYPE_SHM) {
610 		shmfd = fp->f_data;
611 		obj = shmfd->shm_object;
612 		VM_OBJECT_RLOCK(obj);
613 		*obj_size = shmfd->shm_size;
614 	} else {
615 		error = EINVAL;
616 		goto out;
617 	}
618 
619 	if ((obj->flags & OBJ_DEAD) != 0) {
620 		VM_OBJECT_RUNLOCK(obj);
621 		error = EBADF;
622 		goto out;
623 	}
624 
625 	/*
626 	 * Temporarily increase the backing VM object's reference
627 	 * count so that a forced reclamation of its vnode does not
628 	 * immediately destroy it.
629 	 */
630 	vm_object_reference_locked(obj);
631 	VM_OBJECT_RUNLOCK(obj);
632 	*obj_res = obj;
633 	*vp_res = vp;
634 	*shmfd_res = shmfd;
635 
636 out:
637 	if (vp != NULL)
638 		VOP_UNLOCK(vp);
639 	return (error);
640 }
641 
642 static int
643 sendfile_getsock(struct thread *td, int s, struct file **sock_fp,
644     struct socket **so)
645 {
646 	int error;
647 
648 	*sock_fp = NULL;
649 	*so = NULL;
650 
651 	/*
652 	 * The socket must be a stream socket and connected.
653 	 */
654 	error = getsock(td, s, &cap_send_rights, sock_fp);
655 	if (error != 0)
656 		return (error);
657 	*so = (*sock_fp)->f_data;
658 	if ((*so)->so_type != SOCK_STREAM)
659 		return (EINVAL);
660 	/*
661 	 * SCTP one-to-one style sockets currently don't work with
662 	 * sendfile(). So indicate EINVAL for now.
663 	 */
664 	if ((*so)->so_proto->pr_protocol == IPPROTO_SCTP)
665 		return (EINVAL);
666 	return (0);
667 }
668 
669 int
670 vn_sendfile(struct file *fp, int sockfd, struct uio *hdr_uio,
671     struct uio *trl_uio, off_t offset, size_t nbytes, off_t *sent, int flags,
672     struct thread *td)
673 {
674 	struct file *sock_fp;
675 	struct vnode *vp;
676 	struct vm_object *obj;
677 	vm_page_t pga;
678 	struct socket *so;
679 #ifdef KERN_TLS
680 	struct ktls_session *tls;
681 #endif
682 	struct mbuf *m, *mh, *mhtail;
683 	struct sf_buf *sf;
684 	struct shmfd *shmfd;
685 	struct sendfile_sync *sfs;
686 	struct vattr va;
687 	off_t off, sbytes, rem, obj_size, nobj_size;
688 	int bsize, error, ext_pgs_idx, hdrlen, max_pgs, softerr;
689 #ifdef KERN_TLS
690 	int tls_enq_cnt;
691 #endif
692 	bool use_ext_pgs;
693 
694 	obj = NULL;
695 	so = NULL;
696 	m = mh = NULL;
697 	sfs = NULL;
698 #ifdef KERN_TLS
699 	tls = NULL;
700 #endif
701 	hdrlen = sbytes = 0;
702 	softerr = 0;
703 	use_ext_pgs = false;
704 
705 	error = sendfile_getobj(td, fp, &obj, &vp, &shmfd, &obj_size, &bsize);
706 	if (error != 0)
707 		return (error);
708 
709 	error = sendfile_getsock(td, sockfd, &sock_fp, &so);
710 	if (error != 0)
711 		goto out;
712 
713 #ifdef MAC
714 	error = mac_socket_check_send(td->td_ucred, so);
715 	if (error != 0)
716 		goto out;
717 #endif
718 
719 	SFSTAT_INC(sf_syscalls);
720 	SFSTAT_ADD(sf_rhpages_requested, SF_READAHEAD(flags));
721 
722 	if (flags & SF_SYNC) {
723 		sfs = malloc(sizeof(*sfs), M_SENDFILE, M_WAITOK | M_ZERO);
724 		mtx_init(&sfs->mtx, "sendfile", NULL, MTX_DEF);
725 		cv_init(&sfs->cv, "sendfile");
726 		sfs->waiting = true;
727 	}
728 
729 	rem = nbytes ? omin(nbytes, obj_size - offset) : obj_size - offset;
730 
731 	/*
732 	 * Protect against multiple writers to the socket.
733 	 *
734 	 * XXXRW: Historically this has assumed non-interruptibility, so now
735 	 * we implement that, but possibly shouldn't.
736 	 */
737 	error = SOCK_IO_SEND_LOCK(so, SBL_WAIT | SBL_NOINTR);
738 	if (error != 0)
739 		goto out;
740 #ifdef KERN_TLS
741 	tls = ktls_hold(so->so_snd.sb_tls_info);
742 #endif
743 
744 	/*
745 	 * Loop through the pages of the file, starting with the requested
746 	 * offset. Get a file page (do I/O if necessary), map the file page
747 	 * into an sf_buf, attach an mbuf header to the sf_buf, and queue
748 	 * it on the socket.
749 	 * This is done in two loops.  The inner loop turns as many pages
750 	 * as it can, up to available socket buffer space, without blocking
751 	 * into mbufs to have it bulk delivered into the socket send buffer.
752 	 * The outer loop checks the state and available space of the socket
753 	 * and takes care of the overall progress.
754 	 */
755 	for (off = offset; rem > 0; ) {
756 		struct sf_io *sfio;
757 		vm_page_t *pa;
758 		struct mbuf *m0, *mtail;
759 		int nios, space, npages, rhpages;
760 
761 		mtail = NULL;
762 		/*
763 		 * Check the socket state for ongoing connection,
764 		 * no errors and space in socket buffer.
765 		 * If space is low allow for the remainder of the
766 		 * file to be processed if it fits the socket buffer.
767 		 * Otherwise block in waiting for sufficient space
768 		 * to proceed, or if the socket is nonblocking, return
769 		 * to userland with EAGAIN while reporting how far
770 		 * we've come.
771 		 * We wait until the socket buffer has significant free
772 		 * space to do bulk sends.  This makes good use of file
773 		 * system read ahead and allows packet segmentation
774 		 * offloading hardware to take over lots of work.  If
775 		 * we were not careful here we would send off only one
776 		 * sfbuf at a time.
777 		 */
778 		SOCKBUF_LOCK(&so->so_snd);
779 		if (so->so_snd.sb_lowat < so->so_snd.sb_hiwat / 2)
780 			so->so_snd.sb_lowat = so->so_snd.sb_hiwat / 2;
781 retry_space:
782 		if (so->so_snd.sb_state & SBS_CANTSENDMORE) {
783 			error = EPIPE;
784 			SOCKBUF_UNLOCK(&so->so_snd);
785 			goto done;
786 		} else if (so->so_error) {
787 			error = so->so_error;
788 			so->so_error = 0;
789 			SOCKBUF_UNLOCK(&so->so_snd);
790 			goto done;
791 		}
792 		if ((so->so_state & SS_ISCONNECTED) == 0) {
793 			SOCKBUF_UNLOCK(&so->so_snd);
794 			error = ENOTCONN;
795 			goto done;
796 		}
797 
798 		space = sbspace(&so->so_snd);
799 		if (space < rem &&
800 		    (space <= 0 ||
801 		     space < so->so_snd.sb_lowat)) {
802 			if (so->so_state & SS_NBIO) {
803 				SOCKBUF_UNLOCK(&so->so_snd);
804 				error = EAGAIN;
805 				goto done;
806 			}
807 			/*
808 			 * sbwait drops the lock while sleeping.
809 			 * When we loop back to retry_space the
810 			 * state may have changed and we retest
811 			 * for it.
812 			 */
813 			error = sbwait(so, SO_SND);
814 			/*
815 			 * An error from sbwait usually indicates that we've
816 			 * been interrupted by a signal. If we've sent anything
817 			 * then return bytes sent, otherwise return the error.
818 			 */
819 			if (error != 0) {
820 				SOCKBUF_UNLOCK(&so->so_snd);
821 				goto done;
822 			}
823 			goto retry_space;
824 		}
825 		SOCKBUF_UNLOCK(&so->so_snd);
826 
827 		/*
828 		 * At the beginning of the first loop check if any headers
829 		 * are specified and copy them into mbufs.  Reduce space in
830 		 * the socket buffer by the size of the header mbuf chain.
831 		 * Clear hdr_uio here and hdrlen at the end of the first loop.
832 		 */
833 		if (hdr_uio != NULL && hdr_uio->uio_resid > 0) {
834 			hdr_uio->uio_td = td;
835 			hdr_uio->uio_rw = UIO_WRITE;
836 #ifdef KERN_TLS
837 			if (tls != NULL)
838 				mh = m_uiotombuf(hdr_uio, M_WAITOK, space,
839 				    tls->params.max_frame_len, M_EXTPG);
840 			else
841 #endif
842 				mh = m_uiotombuf(hdr_uio, M_WAITOK,
843 				    space, 0, 0);
844 			hdrlen = m_length(mh, &mhtail);
845 			space -= hdrlen;
846 			/*
847 			 * If header consumed all the socket buffer space,
848 			 * don't waste CPU cycles and jump to the end.
849 			 */
850 			if (space == 0) {
851 				sfio = NULL;
852 				nios = 0;
853 				goto prepend_header;
854 			}
855 			hdr_uio = NULL;
856 		}
857 
858 		if (vp != NULL) {
859 			error = vn_lock(vp, LK_SHARED);
860 			if (error != 0)
861 				goto done;
862 
863 			/*
864 			 * Check to see if the file size has changed.
865 			 */
866 			if (obj->type == OBJT_VNODE) {
867 				VM_OBJECT_RLOCK(obj);
868 				nobj_size = obj->un_pager.vnp.vnp_size;
869 				VM_OBJECT_RUNLOCK(obj);
870 			} else {
871 				error = VOP_GETATTR(vp, &va, td->td_ucred);
872 				if (error != 0) {
873 					VOP_UNLOCK(vp);
874 					goto done;
875 				}
876 				nobj_size = va.va_size;
877 			}
878 			if (off >= nobj_size) {
879 				VOP_UNLOCK(vp);
880 				goto done;
881 			}
882 			if (nobj_size != obj_size) {
883 				obj_size = nobj_size;
884 				rem = nbytes ? omin(nbytes + offset, obj_size) :
885 				    obj_size;
886 				rem -= off;
887 			}
888 		}
889 
890 		if (space > rem)
891 			space = rem;
892 		else if (space > PAGE_SIZE) {
893 			/*
894 			 * Use page boundaries when possible for large
895 			 * requests.
896 			 */
897 			if (off & PAGE_MASK)
898 				space -= (PAGE_SIZE - (off & PAGE_MASK));
899 			space = trunc_page(space);
900 			if (off & PAGE_MASK)
901 				space += (PAGE_SIZE - (off & PAGE_MASK));
902 		}
903 
904 		npages = howmany(space + (off & PAGE_MASK), PAGE_SIZE);
905 
906 		/*
907 		 * Calculate maximum allowed number of pages for readahead
908 		 * at this iteration.  If SF_USER_READAHEAD was set, we don't
909 		 * do any heuristics and use exactly the value supplied by
910 		 * application.  Otherwise, we allow readahead up to "rem".
911 		 * If application wants more, let it be, but there is no
912 		 * reason to go above maxphys.  Also check against "obj_size",
913 		 * since vm_pager_has_page() can hint beyond EOF.
914 		 */
915 		if (flags & SF_USER_READAHEAD) {
916 			rhpages = SF_READAHEAD(flags);
917 		} else {
918 			rhpages = howmany(rem + (off & PAGE_MASK), PAGE_SIZE) -
919 			    npages;
920 			rhpages += SF_READAHEAD(flags);
921 		}
922 		rhpages = min(howmany(maxphys, PAGE_SIZE), rhpages);
923 		rhpages = min(howmany(obj_size - trunc_page(off), PAGE_SIZE) -
924 		    npages, rhpages);
925 
926 		sfio = malloc(sizeof(struct sf_io) +
927 		    npages * sizeof(vm_page_t), M_SENDFILE, M_WAITOK);
928 		refcount_init(&sfio->nios, 1);
929 		sfio->obj = obj;
930 		sfio->error = 0;
931 		sfio->m = NULL;
932 		sfio->npages = npages;
933 #ifdef KERN_TLS
934 		/*
935 		 * This doesn't use ktls_hold() because sfio->m will
936 		 * also have a reference on 'tls' that will be valid
937 		 * for all of sfio's lifetime.
938 		 */
939 		sfio->tls = tls;
940 #endif
941 		vm_object_pip_add(obj, 1);
942 		error = sendfile_swapin(obj, sfio, &nios, off, space, rhpages,
943 		    flags);
944 		if (error != 0) {
945 			if (vp != NULL)
946 				VOP_UNLOCK(vp);
947 			sendfile_iodone(sfio, NULL, 0, error);
948 			goto done;
949 		}
950 
951 		/*
952 		 * Loop and construct maximum sized mbuf chain to be bulk
953 		 * dumped into socket buffer.
954 		 */
955 		pa = sfio->pa;
956 
957 		/*
958 		 * Use unmapped mbufs if enabled for TCP.  Unmapped
959 		 * bufs are restricted to TCP as that is what has been
960 		 * tested.  In particular, unmapped mbufs have not
961 		 * been tested with UNIX-domain sockets.
962 		 *
963 		 * TLS frames always require unmapped mbufs.
964 		 */
965 		if ((mb_use_ext_pgs &&
966 		    so->so_proto->pr_protocol == IPPROTO_TCP)
967 #ifdef KERN_TLS
968 		    || tls != NULL
969 #endif
970 		    ) {
971 			use_ext_pgs = true;
972 #ifdef KERN_TLS
973 			if (tls != NULL)
974 				max_pgs = num_pages(tls->params.max_frame_len);
975 			else
976 #endif
977 				max_pgs = MBUF_PEXT_MAX_PGS;
978 
979 			/* Start at last index, to wrap on first use. */
980 			ext_pgs_idx = max_pgs - 1;
981 		}
982 
983 		for (int i = 0; i < npages; i++) {
984 			/*
985 			 * If a page wasn't grabbed successfully, then
986 			 * trim the array. Can happen only with SF_NODISKIO.
987 			 */
988 			if (pa[i] == NULL) {
989 				SFSTAT_INC(sf_busy);
990 				fixspace(npages, i, off, &space);
991 				sfio->npages = i;
992 				softerr = EBUSY;
993 				break;
994 			}
995 			pga = pa[i];
996 			if (pga == bogus_page)
997 				pga = vm_page_relookup(obj, sfio->pindex0 + i);
998 
999 			if (use_ext_pgs) {
1000 				off_t xfs;
1001 
1002 				ext_pgs_idx++;
1003 				if (ext_pgs_idx == max_pgs) {
1004 					m0 = mb_alloc_ext_pgs(M_WAITOK,
1005 					    sendfile_free_mext_pg);
1006 
1007 					if (flags & SF_NOCACHE) {
1008 						m0->m_ext.ext_flags |=
1009 						    EXT_FLAG_NOCACHE;
1010 
1011 						/*
1012 						 * See comment below regarding
1013 						 * ignoring SF_NOCACHE for the
1014 						 * last page.
1015 						 */
1016 						if ((npages - i <= max_pgs) &&
1017 						    ((off + space) & PAGE_MASK) &&
1018 						    (rem > space || rhpages > 0))
1019 							m0->m_ext.ext_flags |=
1020 							    EXT_FLAG_CACHE_LAST;
1021 					}
1022 					if (sfs != NULL) {
1023 						m0->m_ext.ext_flags |=
1024 						    EXT_FLAG_SYNC;
1025 						m0->m_ext.ext_arg1 = sfs;
1026 						mtx_lock(&sfs->mtx);
1027 						sfs->count++;
1028 						mtx_unlock(&sfs->mtx);
1029 					}
1030 					ext_pgs_idx = 0;
1031 
1032 					/* Append to mbuf chain. */
1033 					if (mtail != NULL)
1034 						mtail->m_next = m0;
1035 					else
1036 						m = m0;
1037 					mtail = m0;
1038 					m0->m_epg_1st_off =
1039 					    vmoff(i, off) & PAGE_MASK;
1040 				}
1041 				if (nios) {
1042 					mtail->m_flags |= M_NOTREADY;
1043 					m0->m_epg_nrdy++;
1044 				}
1045 
1046 				m0->m_epg_pa[ext_pgs_idx] = VM_PAGE_TO_PHYS(pga);
1047 				m0->m_epg_npgs++;
1048 				xfs = xfsize(i, npages, off, space);
1049 				m0->m_epg_last_len = xfs;
1050 				MBUF_EXT_PGS_ASSERT_SANITY(m0);
1051 				mtail->m_len += xfs;
1052 				mtail->m_ext.ext_size += PAGE_SIZE;
1053 				continue;
1054 			}
1055 
1056 			/*
1057 			 * Get a sendfile buf.  When allocating the
1058 			 * first buffer for mbuf chain, we usually
1059 			 * wait as long as necessary, but this wait
1060 			 * can be interrupted.  For consequent
1061 			 * buffers, do not sleep, since several
1062 			 * threads might exhaust the buffers and then
1063 			 * deadlock.
1064 			 */
1065 			sf = sf_buf_alloc(pga,
1066 			    m != NULL ? SFB_NOWAIT : SFB_CATCH);
1067 			if (sf == NULL) {
1068 				SFSTAT_INC(sf_allocfail);
1069 				sendfile_iowait(sfio, "sfnosf");
1070 				for (int j = i; j < npages; j++) {
1071 					vm_page_unwire(pa[j], PQ_INACTIVE);
1072 					pa[j] = NULL;
1073 				}
1074 				if (m == NULL)
1075 					softerr = ENOBUFS;
1076 				fixspace(npages, i, off, &space);
1077 				sfio->npages = i;
1078 				break;
1079 			}
1080 
1081 			m0 = m_get(M_WAITOK, MT_DATA);
1082 			m0->m_ext.ext_buf = (char *)sf_buf_kva(sf);
1083 			m0->m_ext.ext_size = PAGE_SIZE;
1084 			m0->m_ext.ext_arg1 = sf;
1085 			m0->m_ext.ext_type = EXT_SFBUF;
1086 			m0->m_ext.ext_flags = EXT_FLAG_EMBREF;
1087 			m0->m_ext.ext_free = sendfile_free_mext;
1088 			/*
1089 			 * SF_NOCACHE sets the page as being freed upon send.
1090 			 * However, we ignore it for the last page in 'space',
1091 			 * if the page is truncated, and we got more data to
1092 			 * send (rem > space), or if we have readahead
1093 			 * configured (rhpages > 0).
1094 			 */
1095 			if ((flags & SF_NOCACHE) &&
1096 			    (i != npages - 1 ||
1097 			    !((off + space) & PAGE_MASK) ||
1098 			    !(rem > space || rhpages > 0)))
1099 				m0->m_ext.ext_flags |= EXT_FLAG_NOCACHE;
1100 			if (sfs != NULL) {
1101 				m0->m_ext.ext_flags |= EXT_FLAG_SYNC;
1102 				m0->m_ext.ext_arg2 = sfs;
1103 				mtx_lock(&sfs->mtx);
1104 				sfs->count++;
1105 				mtx_unlock(&sfs->mtx);
1106 			}
1107 			m0->m_ext.ext_count = 1;
1108 			m0->m_flags |= (M_EXT | M_RDONLY);
1109 			if (nios)
1110 				m0->m_flags |= M_NOTREADY;
1111 			m0->m_data = (char *)sf_buf_kva(sf) +
1112 			    (vmoff(i, off) & PAGE_MASK);
1113 			m0->m_len = xfsize(i, npages, off, space);
1114 
1115 			/* Append to mbuf chain. */
1116 			if (mtail != NULL)
1117 				mtail->m_next = m0;
1118 			else
1119 				m = m0;
1120 			mtail = m0;
1121 		}
1122 
1123 		if (vp != NULL)
1124 			VOP_UNLOCK(vp);
1125 
1126 		/* Keep track of bytes processed. */
1127 		off += space;
1128 		rem -= space;
1129 
1130 		/*
1131 		 * Prepend header, if any.  Save pointer to first mbuf
1132 		 * with a page.
1133 		 */
1134 		if (hdrlen) {
1135 prepend_header:
1136 			m0 = mhtail->m_next = m;
1137 			m = mh;
1138 			mh = NULL;
1139 		} else
1140 			m0 = m;
1141 
1142 		if (m == NULL) {
1143 			KASSERT(softerr, ("%s: m NULL, no error", __func__));
1144 			error = softerr;
1145 			sendfile_iodone(sfio, NULL, 0, 0);
1146 			goto done;
1147 		}
1148 
1149 		/* Add the buffer chain to the socket buffer. */
1150 		KASSERT(m_length(m, NULL) == space + hdrlen,
1151 		    ("%s: mlen %u space %d hdrlen %d",
1152 		    __func__, m_length(m, NULL), space, hdrlen));
1153 
1154 		CURVNET_SET(so->so_vnet);
1155 #ifdef KERN_TLS
1156 		if (tls != NULL)
1157 			ktls_frame(m, tls, &tls_enq_cnt, TLS_RLTYPE_APP);
1158 #endif
1159 		if (nios == 0) {
1160 			/*
1161 			 * If sendfile_swapin() didn't initiate any I/Os,
1162 			 * which happens if all data is cached in VM, or if
1163 			 * the header consumed all socket buffer space and
1164 			 * sfio is NULL, then we can send data right now
1165 			 * without the PRUS_NOTREADY flag.
1166 			 */
1167 			if (sfio != NULL)
1168 				sendfile_iodone(sfio, NULL, 0, 0);
1169 #ifdef KERN_TLS
1170 			if (tls != NULL && tls->mode == TCP_TLS_MODE_SW) {
1171 				error = so->so_proto->pr_send(so,
1172 				    PRUS_NOTREADY, m, NULL, NULL, td);
1173 				if (error != 0) {
1174 					m_freem(m);
1175 				} else {
1176 					soref(so);
1177 					ktls_enqueue(m, so, tls_enq_cnt);
1178 				}
1179 			} else
1180 #endif
1181 				error = so->so_proto->pr_send(so, 0, m, NULL,
1182 				    NULL, td);
1183 		} else {
1184 			sfio->so = so;
1185 			sfio->m = m0;
1186 			soref(so);
1187 			error = so->so_proto->pr_send(so, PRUS_NOTREADY, m,
1188 			    NULL, NULL, td);
1189 			sendfile_iodone(sfio, NULL, 0, error);
1190 		}
1191 		CURVNET_RESTORE();
1192 
1193 		m = NULL;
1194 		if (error)
1195 			goto done;
1196 		sbytes += space + hdrlen;
1197 		if (hdrlen)
1198 			hdrlen = 0;
1199 		if (softerr) {
1200 			error = softerr;
1201 			goto done;
1202 		}
1203 	}
1204 
1205 	/*
1206 	 * Send trailers. Wimp out and use writev(2).
1207 	 */
1208 	if (trl_uio != NULL) {
1209 		SOCK_IO_SEND_UNLOCK(so);
1210 		error = kern_writev(td, sockfd, trl_uio);
1211 		if (error == 0)
1212 			sbytes += td->td_retval[0];
1213 		goto out;
1214 	}
1215 
1216 done:
1217 	SOCK_IO_SEND_UNLOCK(so);
1218 out:
1219 	/*
1220 	 * If there was no error we have to clear td->td_retval[0]
1221 	 * because it may have been set by writev.
1222 	 */
1223 	if (error == 0) {
1224 		td->td_retval[0] = 0;
1225 	}
1226 	if (sent != NULL) {
1227 		(*sent) = sbytes;
1228 	}
1229 	if (obj != NULL)
1230 		vm_object_deallocate(obj);
1231 	if (so)
1232 		fdrop(sock_fp, td);
1233 	if (m)
1234 		m_freem(m);
1235 	if (mh)
1236 		m_freem(mh);
1237 
1238 	if (sfs != NULL) {
1239 		mtx_lock(&sfs->mtx);
1240 		if (sfs->count != 0)
1241 			error = cv_wait_sig(&sfs->cv, &sfs->mtx);
1242 		if (sfs->count == 0) {
1243 			sendfile_sync_destroy(sfs);
1244 		} else {
1245 			sfs->waiting = false;
1246 			mtx_unlock(&sfs->mtx);
1247 		}
1248 	}
1249 #ifdef KERN_TLS
1250 	if (tls != NULL)
1251 		ktls_free(tls);
1252 #endif
1253 
1254 	if (error == ERESTART)
1255 		error = EINTR;
1256 
1257 	return (error);
1258 }
1259 
1260 static int
1261 sendfile(struct thread *td, struct sendfile_args *uap, int compat)
1262 {
1263 	struct sf_hdtr hdtr;
1264 	struct uio *hdr_uio, *trl_uio;
1265 	struct file *fp;
1266 	off_t sbytes;
1267 	int error;
1268 
1269 	/*
1270 	 * File offset must be positive.  If it goes beyond EOF
1271 	 * we send only the header/trailer and no payload data.
1272 	 */
1273 	if (uap->offset < 0)
1274 		return (EINVAL);
1275 
1276 	sbytes = 0;
1277 	hdr_uio = trl_uio = NULL;
1278 
1279 	if (uap->hdtr != NULL) {
1280 		error = copyin(uap->hdtr, &hdtr, sizeof(hdtr));
1281 		if (error != 0)
1282 			goto out;
1283 		if (hdtr.headers != NULL) {
1284 			error = copyinuio(hdtr.headers, hdtr.hdr_cnt,
1285 			    &hdr_uio);
1286 			if (error != 0)
1287 				goto out;
1288 #ifdef COMPAT_FREEBSD4
1289 			/*
1290 			 * In FreeBSD < 5.0 the nbytes to send also included
1291 			 * the header.  If compat is specified subtract the
1292 			 * header size from nbytes.
1293 			 */
1294 			if (compat) {
1295 				if (uap->nbytes > hdr_uio->uio_resid)
1296 					uap->nbytes -= hdr_uio->uio_resid;
1297 				else
1298 					uap->nbytes = 0;
1299 			}
1300 #endif
1301 		}
1302 		if (hdtr.trailers != NULL) {
1303 			error = copyinuio(hdtr.trailers, hdtr.trl_cnt,
1304 			    &trl_uio);
1305 			if (error != 0)
1306 				goto out;
1307 		}
1308 	}
1309 
1310 	AUDIT_ARG_FD(uap->fd);
1311 
1312 	/*
1313 	 * sendfile(2) can start at any offset within a file so we require
1314 	 * CAP_READ+CAP_SEEK = CAP_PREAD.
1315 	 */
1316 	if ((error = fget_read(td, uap->fd, &cap_pread_rights, &fp)) != 0)
1317 		goto out;
1318 
1319 	error = fo_sendfile(fp, uap->s, hdr_uio, trl_uio, uap->offset,
1320 	    uap->nbytes, &sbytes, uap->flags, td);
1321 	fdrop(fp, td);
1322 
1323 	if (uap->sbytes != NULL)
1324 		copyout(&sbytes, uap->sbytes, sizeof(off_t));
1325 
1326 out:
1327 	free(hdr_uio, M_IOV);
1328 	free(trl_uio, M_IOV);
1329 	return (error);
1330 }
1331 
1332 /*
1333  * sendfile(2)
1334  *
1335  * int sendfile(int fd, int s, off_t offset, size_t nbytes,
1336  *       struct sf_hdtr *hdtr, off_t *sbytes, int flags)
1337  *
1338  * Send a file specified by 'fd' and starting at 'offset' to a socket
1339  * specified by 's'. Send only 'nbytes' of the file or until EOF if nbytes ==
1340  * 0.  Optionally add a header and/or trailer to the socket output.  If
1341  * specified, write the total number of bytes sent into *sbytes.
1342  */
1343 int
1344 sys_sendfile(struct thread *td, struct sendfile_args *uap)
1345 {
1346 
1347 	return (sendfile(td, uap, 0));
1348 }
1349 
1350 #ifdef COMPAT_FREEBSD4
1351 int
1352 freebsd4_sendfile(struct thread *td, struct freebsd4_sendfile_args *uap)
1353 {
1354 	struct sendfile_args args;
1355 
1356 	args.fd = uap->fd;
1357 	args.s = uap->s;
1358 	args.offset = uap->offset;
1359 	args.nbytes = uap->nbytes;
1360 	args.hdtr = uap->hdtr;
1361 	args.sbytes = uap->sbytes;
1362 	args.flags = uap->flags;
1363 
1364 	return (sendfile(td, &args, 1));
1365 }
1366 #endif /* COMPAT_FREEBSD4 */
1367