xref: /freebsd/sys/kern/kern_sendfile.c (revision 193d9e768ba63fcfb187cfd17f461f7d41345048)
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 /*
66  * Structure describing a single sendfile(2) I/O, which may consist of
67  * several underlying pager I/Os.
68  *
69  * The syscall context allocates the structure and initializes 'nios'
70  * to 1.  As sendfile_swapin() runs through pages and starts asynchronous
71  * paging operations, it increments 'nios'.
72  *
73  * Every I/O completion calls sendfile_iodone(), which decrements the 'nios',
74  * and the syscall also calls sendfile_iodone() after allocating all mbufs,
75  * linking them and sending to socket.  Whoever reaches zero 'nios' is
76  * responsible to * call pru_ready on the socket, to notify it of readyness
77  * of the data.
78  */
79 struct sf_io {
80 	volatile u_int	nios;
81 	u_int		error;
82 	int		npages;
83 	struct file	*sock_fp;
84 	struct mbuf	*m;
85 	vm_page_t	pa[];
86 };
87 
88 /*
89  * Structure used to track requests with SF_SYNC flag.
90  */
91 struct sendfile_sync {
92 	struct mtx	mtx;
93 	struct cv	cv;
94 	unsigned	count;
95 };
96 
97 counter_u64_t sfstat[sizeof(struct sfstat) / sizeof(uint64_t)];
98 
99 static void
100 sfstat_init(const void *unused)
101 {
102 
103 	COUNTER_ARRAY_ALLOC(sfstat, sizeof(struct sfstat) / sizeof(uint64_t),
104 	    M_WAITOK);
105 }
106 SYSINIT(sfstat, SI_SUB_MBUF, SI_ORDER_FIRST, sfstat_init, NULL);
107 
108 static int
109 sfstat_sysctl(SYSCTL_HANDLER_ARGS)
110 {
111 	struct sfstat s;
112 
113 	COUNTER_ARRAY_COPY(sfstat, &s, sizeof(s) / sizeof(uint64_t));
114 	if (req->newptr)
115 		COUNTER_ARRAY_ZERO(sfstat, sizeof(s) / sizeof(uint64_t));
116 	return (SYSCTL_OUT(req, &s, sizeof(s)));
117 }
118 SYSCTL_PROC(_kern_ipc, OID_AUTO, sfstat, CTLTYPE_OPAQUE | CTLFLAG_RW,
119     NULL, 0, sfstat_sysctl, "I", "sendfile statistics");
120 
121 /*
122  * Detach mapped page and release resources back to the system.  Called
123  * by mbuf(9) code when last reference to a page is freed.
124  */
125 void
126 sf_ext_free(void *arg1, void *arg2)
127 {
128 	struct sf_buf *sf = arg1;
129 	struct sendfile_sync *sfs = arg2;
130 	vm_page_t pg = sf_buf_page(sf);
131 
132 	sf_buf_free(sf);
133 
134 	vm_page_lock(pg);
135 	/*
136 	 * Check for the object going away on us. This can
137 	 * happen since we don't hold a reference to it.
138 	 * If so, we're responsible for freeing the page.
139 	 */
140 	if (vm_page_unwire(pg, PQ_INACTIVE) && pg->object == NULL)
141 		vm_page_free(pg);
142 	vm_page_unlock(pg);
143 
144 	if (sfs != NULL) {
145 		mtx_lock(&sfs->mtx);
146 		KASSERT(sfs->count > 0, ("Sendfile sync botchup count == 0"));
147 		if (--sfs->count == 0)
148 			cv_signal(&sfs->cv);
149 		mtx_unlock(&sfs->mtx);
150 	}
151 }
152 
153 /*
154  * Same as above, but forces the page to be detached from the object
155  * and go into free pool.
156  */
157 void
158 sf_ext_free_nocache(void *arg1, void *arg2)
159 {
160 	struct sf_buf *sf = arg1;
161 	struct sendfile_sync *sfs = arg2;
162 	vm_page_t pg = sf_buf_page(sf);
163 
164 	sf_buf_free(sf);
165 
166 	vm_page_lock(pg);
167 	if (vm_page_unwire(pg, PQ_NONE)) {
168 		vm_object_t obj;
169 
170 		/* Try to free the page, but only if it is cheap to. */
171 		if ((obj = pg->object) == NULL)
172 			vm_page_free(pg);
173 		else if (!vm_page_xbusied(pg) && VM_OBJECT_TRYWLOCK(obj)) {
174 			vm_page_free(pg);
175 			VM_OBJECT_WUNLOCK(obj);
176 		} else
177 			vm_page_deactivate(pg);
178 	}
179 	vm_page_unlock(pg);
180 
181 	if (sfs != NULL) {
182 		mtx_lock(&sfs->mtx);
183 		KASSERT(sfs->count > 0, ("Sendfile sync botchup count == 0"));
184 		if (--sfs->count == 0)
185 			cv_signal(&sfs->cv);
186 		mtx_unlock(&sfs->mtx);
187 	}
188 }
189 
190 /*
191  * Helper function to calculate how much data to put into page i of n.
192  * Only first and last pages are special.
193  */
194 static inline off_t
195 xfsize(int i, int n, off_t off, off_t len)
196 {
197 
198 	if (i == 0)
199 		return (omin(PAGE_SIZE - (off & PAGE_MASK), len));
200 
201 	if (i == n - 1 && ((off + len) & PAGE_MASK) > 0)
202 		return ((off + len) & PAGE_MASK);
203 
204 	return (PAGE_SIZE);
205 }
206 
207 /*
208  * Helper function to get offset within object for i page.
209  */
210 static inline vm_offset_t
211 vmoff(int i, off_t off)
212 {
213 
214 	if (i == 0)
215 		return ((vm_offset_t)off);
216 
217 	return (trunc_page(off + i * PAGE_SIZE));
218 }
219 
220 /*
221  * Helper function used when allocation of a page or sf_buf failed.
222  * Pretend as if we don't have enough space, subtract xfsize() of
223  * all pages that failed.
224  */
225 static inline void
226 fixspace(int old, int new, off_t off, int *space)
227 {
228 
229 	KASSERT(old > new, ("%s: old %d new %d", __func__, old, new));
230 
231 	/* Subtract last one. */
232 	*space -= xfsize(old - 1, old, off, *space);
233 	old--;
234 
235 	if (new == old)
236 		/* There was only one page. */
237 		return;
238 
239 	/* Subtract first one. */
240 	if (new == 0) {
241 		*space -= xfsize(0, old, off, *space);
242 		new++;
243 	}
244 
245 	/* Rest of pages are full sized. */
246 	*space -= (old - new) * PAGE_SIZE;
247 
248 	KASSERT(*space >= 0, ("%s: space went backwards", __func__));
249 }
250 
251 /*
252  * I/O completion callback.
253  */
254 static void
255 sendfile_iodone(void *arg, vm_page_t *pg, int count, int error)
256 {
257 	struct sf_io *sfio = arg;
258 	struct socket *so;
259 
260 	for (int i = 0; i < count; i++)
261 		if (pg[i] != bogus_page)
262 			vm_page_xunbusy(pg[i]);
263 
264 	if (error)
265 		sfio->error = error;
266 
267 	if (!refcount_release(&sfio->nios))
268 		return;
269 
270 	so = sfio->sock_fp->f_data;
271 
272 	if (sfio->error) {
273 		struct mbuf *m;
274 
275 		/*
276 		 * I/O operation failed.  The state of data in the socket
277 		 * is now inconsistent, and all what we can do is to tear
278 		 * it down. Protocol abort method would tear down protocol
279 		 * state, free all ready mbufs and detach not ready ones.
280 		 * We will free the mbufs corresponding to this I/O manually.
281 		 *
282 		 * The socket would be marked with EIO and made available
283 		 * for read, so that application receives EIO on next
284 		 * syscall and eventually closes the socket.
285 		 */
286 		so->so_proto->pr_usrreqs->pru_abort(so);
287 		so->so_error = EIO;
288 
289 		m = sfio->m;
290 		for (int i = 0; i < sfio->npages; i++)
291 			m = m_free(m);
292 	} else {
293 		CURVNET_SET(so->so_vnet);
294 		(void )(so->so_proto->pr_usrreqs->pru_ready)(so, sfio->m,
295 		    sfio->npages);
296 		CURVNET_RESTORE();
297 	}
298 
299 	/* XXXGL: curthread */
300 	fdrop(sfio->sock_fp, curthread);
301 	free(sfio, M_TEMP);
302 }
303 
304 /*
305  * Iterate through pages vector and request paging for non-valid pages.
306  */
307 static int
308 sendfile_swapin(vm_object_t obj, struct sf_io *sfio, off_t off, off_t len,
309     int npages, int rhpages, int flags)
310 {
311 	vm_page_t *pa = sfio->pa;
312 	int nios;
313 
314 	nios = 0;
315 	flags = (flags & SF_NODISKIO) ? VM_ALLOC_NOWAIT : 0;
316 
317 	/*
318 	 * First grab all the pages and wire them.  Note that we grab
319 	 * only required pages.  Readahead pages are dealt with later.
320 	 */
321 	VM_OBJECT_WLOCK(obj);
322 	for (int i = 0; i < npages; i++) {
323 		pa[i] = vm_page_grab(obj, OFF_TO_IDX(vmoff(i, off)),
324 		    VM_ALLOC_WIRED | VM_ALLOC_NORMAL | flags);
325 		if (pa[i] == NULL) {
326 			npages = i;
327 			rhpages = 0;
328 			break;
329 		}
330 	}
331 
332 	for (int i = 0; i < npages;) {
333 		int j, a, count, rv;
334 
335 		/* Skip valid pages. */
336 		if (vm_page_is_valid(pa[i], vmoff(i, off) & PAGE_MASK,
337 		    xfsize(i, npages, off, len))) {
338 			vm_page_xunbusy(pa[i]);
339 			SFSTAT_INC(sf_pages_valid);
340 			i++;
341 			continue;
342 		}
343 
344 		/*
345 		 * Next page is invalid.  Check if it belongs to pager.  It
346 		 * may not be there, which is a regular situation for shmem
347 		 * pager.  For vnode pager this happens only in case of
348 		 * a sparse file.
349 		 *
350 		 * Important feature of vm_pager_has_page() is the hint
351 		 * stored in 'a', about how many pages we can pagein after
352 		 * this page in a single I/O.
353 		 */
354 		if (!vm_pager_has_page(obj, OFF_TO_IDX(vmoff(i, off)), NULL,
355 		    &a)) {
356 			pmap_zero_page(pa[i]);
357 			pa[i]->valid = VM_PAGE_BITS_ALL;
358 			pa[i]->dirty = 0;
359 			vm_page_xunbusy(pa[i]);
360 			i++;
361 			continue;
362 		}
363 
364 		/*
365 		 * We want to pagein as many pages as possible, limited only
366 		 * by the 'a' hint and actual request.
367 		 */
368 		count = min(a + 1, npages - i);
369 
370 		/*
371 		 * We should not pagein into a valid page, thus we first trim
372 		 * any valid pages off the end of request, and substitute
373 		 * to bogus_page those, that are in the middle.
374 		 */
375 		for (j = i + count - 1; j > i; j--) {
376 			if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
377 			    xfsize(j, npages, off, len))) {
378 				count--;
379 				rhpages = 0;
380 			} else
381 				break;
382 		}
383 		for (j = i + 1; j < i + count - 1; j++)
384 			if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
385 			    xfsize(j, npages, off, len))) {
386 				vm_page_xunbusy(pa[j]);
387 				SFSTAT_INC(sf_pages_valid);
388 				SFSTAT_INC(sf_pages_bogus);
389 				pa[j] = bogus_page;
390 			}
391 
392 		refcount_acquire(&sfio->nios);
393 		rv = vm_pager_get_pages_async(obj, pa + i, count, NULL,
394 		    i + count == npages ? &rhpages : NULL,
395 		    &sendfile_iodone, sfio);
396 		KASSERT(rv == VM_PAGER_OK, ("%s: pager fail obj %p page %p",
397 		    __func__, obj, pa[i]));
398 
399 		SFSTAT_INC(sf_iocnt);
400 		SFSTAT_ADD(sf_pages_read, count);
401 		if (i + count == npages)
402 			SFSTAT_ADD(sf_rhpages_read, rhpages);
403 
404 		/*
405 		 * Restore the valid page pointers.  They are already
406 		 * unbusied, but still wired.
407 		 */
408 		for (j = i; j < i + count; j++)
409 			if (pa[j] == bogus_page) {
410 				pa[j] = vm_page_lookup(obj,
411 				    OFF_TO_IDX(vmoff(j, off)));
412 				KASSERT(pa[j], ("%s: page %p[%d] disappeared",
413 				    __func__, pa, j));
414 
415 			}
416 		i += count;
417 		nios++;
418 	}
419 
420 	VM_OBJECT_WUNLOCK(obj);
421 
422 	if (nios == 0 && npages != 0)
423 		SFSTAT_INC(sf_noiocnt);
424 
425 	return (nios);
426 }
427 
428 static int
429 sendfile_getobj(struct thread *td, struct file *fp, vm_object_t *obj_res,
430     struct vnode **vp_res, struct shmfd **shmfd_res, off_t *obj_size,
431     int *bsize)
432 {
433 	struct vattr va;
434 	vm_object_t obj;
435 	struct vnode *vp;
436 	struct shmfd *shmfd;
437 	int error;
438 
439 	vp = *vp_res = NULL;
440 	obj = NULL;
441 	shmfd = *shmfd_res = NULL;
442 	*bsize = 0;
443 
444 	/*
445 	 * The file descriptor must be a regular file and have a
446 	 * backing VM object.
447 	 */
448 	if (fp->f_type == DTYPE_VNODE) {
449 		vp = fp->f_vnode;
450 		vn_lock(vp, LK_SHARED | LK_RETRY);
451 		if (vp->v_type != VREG) {
452 			error = EINVAL;
453 			goto out;
454 		}
455 		*bsize = vp->v_mount->mnt_stat.f_iosize;
456 		error = VOP_GETATTR(vp, &va, td->td_ucred);
457 		if (error != 0)
458 			goto out;
459 		*obj_size = va.va_size;
460 		obj = vp->v_object;
461 		if (obj == NULL) {
462 			error = EINVAL;
463 			goto out;
464 		}
465 	} else if (fp->f_type == DTYPE_SHM) {
466 		error = 0;
467 		shmfd = fp->f_data;
468 		obj = shmfd->shm_object;
469 		*obj_size = shmfd->shm_size;
470 	} else {
471 		error = EINVAL;
472 		goto out;
473 	}
474 
475 	VM_OBJECT_WLOCK(obj);
476 	if ((obj->flags & OBJ_DEAD) != 0) {
477 		VM_OBJECT_WUNLOCK(obj);
478 		error = EBADF;
479 		goto out;
480 	}
481 
482 	/*
483 	 * Temporarily increase the backing VM object's reference
484 	 * count so that a forced reclamation of its vnode does not
485 	 * immediately destroy it.
486 	 */
487 	vm_object_reference_locked(obj);
488 	VM_OBJECT_WUNLOCK(obj);
489 	*obj_res = obj;
490 	*vp_res = vp;
491 	*shmfd_res = shmfd;
492 
493 out:
494 	if (vp != NULL)
495 		VOP_UNLOCK(vp, 0);
496 	return (error);
497 }
498 
499 static int
500 sendfile_getsock(struct thread *td, int s, struct file **sock_fp,
501     struct socket **so)
502 {
503 	cap_rights_t rights;
504 	int error;
505 
506 	*sock_fp = NULL;
507 	*so = NULL;
508 
509 	/*
510 	 * The socket must be a stream socket and connected.
511 	 */
512 	error = getsock_cap(td, s, cap_rights_init(&rights, CAP_SEND),
513 	    sock_fp, NULL, NULL);
514 	if (error != 0)
515 		return (error);
516 	*so = (*sock_fp)->f_data;
517 	if ((*so)->so_type != SOCK_STREAM)
518 		return (EINVAL);
519 	if (((*so)->so_state & SS_ISCONNECTED) == 0)
520 		return (ENOTCONN);
521 	return (0);
522 }
523 
524 int
525 vn_sendfile(struct file *fp, int sockfd, struct uio *hdr_uio,
526     struct uio *trl_uio, off_t offset, size_t nbytes, off_t *sent, int flags,
527     struct thread *td)
528 {
529 	struct file *sock_fp;
530 	struct vnode *vp;
531 	struct vm_object *obj;
532 	struct socket *so;
533 	struct mbuf *m, *mh, *mhtail;
534 	struct sf_buf *sf;
535 	struct shmfd *shmfd;
536 	struct sendfile_sync *sfs;
537 	struct vattr va;
538 	off_t off, sbytes, rem, obj_size;
539 	int error, softerr, bsize, hdrlen;
540 
541 	obj = NULL;
542 	so = NULL;
543 	m = mh = NULL;
544 	sfs = NULL;
545 	hdrlen = sbytes = 0;
546 	softerr = 0;
547 
548 	error = sendfile_getobj(td, fp, &obj, &vp, &shmfd, &obj_size, &bsize);
549 	if (error != 0)
550 		return (error);
551 
552 	error = sendfile_getsock(td, sockfd, &sock_fp, &so);
553 	if (error != 0)
554 		goto out;
555 
556 #ifdef MAC
557 	error = mac_socket_check_send(td->td_ucred, so);
558 	if (error != 0)
559 		goto out;
560 #endif
561 
562 	SFSTAT_INC(sf_syscalls);
563 	SFSTAT_ADD(sf_rhpages_requested, SF_READAHEAD(flags));
564 
565 	if (flags & SF_SYNC) {
566 		sfs = malloc(sizeof *sfs, M_TEMP, M_WAITOK | M_ZERO);
567 		mtx_init(&sfs->mtx, "sendfile", NULL, MTX_DEF);
568 		cv_init(&sfs->cv, "sendfile");
569 	}
570 
571 	rem = nbytes ? omin(nbytes, obj_size - offset) : obj_size - offset;
572 
573 	/*
574 	 * Protect against multiple writers to the socket.
575 	 *
576 	 * XXXRW: Historically this has assumed non-interruptibility, so now
577 	 * we implement that, but possibly shouldn't.
578 	 */
579 	(void)sblock(&so->so_snd, SBL_WAIT | SBL_NOINTR);
580 
581 	/*
582 	 * Loop through the pages of the file, starting with the requested
583 	 * offset. Get a file page (do I/O if necessary), map the file page
584 	 * into an sf_buf, attach an mbuf header to the sf_buf, and queue
585 	 * it on the socket.
586 	 * This is done in two loops.  The inner loop turns as many pages
587 	 * as it can, up to available socket buffer space, without blocking
588 	 * into mbufs to have it bulk delivered into the socket send buffer.
589 	 * The outer loop checks the state and available space of the socket
590 	 * and takes care of the overall progress.
591 	 */
592 	for (off = offset; rem > 0; ) {
593 		struct sf_io *sfio;
594 		vm_page_t *pa;
595 		struct mbuf *mtail;
596 		int nios, space, npages, rhpages;
597 
598 		mtail = NULL;
599 		/*
600 		 * Check the socket state for ongoing connection,
601 		 * no errors and space in socket buffer.
602 		 * If space is low allow for the remainder of the
603 		 * file to be processed if it fits the socket buffer.
604 		 * Otherwise block in waiting for sufficient space
605 		 * to proceed, or if the socket is nonblocking, return
606 		 * to userland with EAGAIN while reporting how far
607 		 * we've come.
608 		 * We wait until the socket buffer has significant free
609 		 * space to do bulk sends.  This makes good use of file
610 		 * system read ahead and allows packet segmentation
611 		 * offloading hardware to take over lots of work.  If
612 		 * we were not careful here we would send off only one
613 		 * sfbuf at a time.
614 		 */
615 		SOCKBUF_LOCK(&so->so_snd);
616 		if (so->so_snd.sb_lowat < so->so_snd.sb_hiwat / 2)
617 			so->so_snd.sb_lowat = so->so_snd.sb_hiwat / 2;
618 retry_space:
619 		if (so->so_snd.sb_state & SBS_CANTSENDMORE) {
620 			error = EPIPE;
621 			SOCKBUF_UNLOCK(&so->so_snd);
622 			goto done;
623 		} else if (so->so_error) {
624 			error = so->so_error;
625 			so->so_error = 0;
626 			SOCKBUF_UNLOCK(&so->so_snd);
627 			goto done;
628 		}
629 		space = sbspace(&so->so_snd);
630 		if (space < rem &&
631 		    (space <= 0 ||
632 		     space < so->so_snd.sb_lowat)) {
633 			if (so->so_state & SS_NBIO) {
634 				SOCKBUF_UNLOCK(&so->so_snd);
635 				error = EAGAIN;
636 				goto done;
637 			}
638 			/*
639 			 * sbwait drops the lock while sleeping.
640 			 * When we loop back to retry_space the
641 			 * state may have changed and we retest
642 			 * for it.
643 			 */
644 			error = sbwait(&so->so_snd);
645 			/*
646 			 * An error from sbwait usually indicates that we've
647 			 * been interrupted by a signal. If we've sent anything
648 			 * then return bytes sent, otherwise return the error.
649 			 */
650 			if (error != 0) {
651 				SOCKBUF_UNLOCK(&so->so_snd);
652 				goto done;
653 			}
654 			goto retry_space;
655 		}
656 		SOCKBUF_UNLOCK(&so->so_snd);
657 
658 		/*
659 		 * At the beginning of the first loop check if any headers
660 		 * are specified and copy them into mbufs.  Reduce space in
661 		 * the socket buffer by the size of the header mbuf chain.
662 		 * Clear hdr_uio here and hdrlen at the end of the first loop.
663 		 */
664 		if (hdr_uio != NULL && hdr_uio->uio_resid > 0) {
665 			hdr_uio->uio_td = td;
666 			hdr_uio->uio_rw = UIO_WRITE;
667 			mh = m_uiotombuf(hdr_uio, M_WAITOK, space, 0, 0);
668 			hdrlen = m_length(mh, &mhtail);
669 			space -= hdrlen;
670 			/*
671 			 * If header consumed all the socket buffer space,
672 			 * don't waste CPU cycles and jump to the end.
673 			 */
674 			if (space == 0) {
675 				sfio = NULL;
676 				nios = 0;
677 				goto prepend_header;
678 			}
679 			hdr_uio = NULL;
680 		}
681 
682 		if (vp != NULL) {
683 			error = vn_lock(vp, LK_SHARED);
684 			if (error != 0)
685 				goto done;
686 			error = VOP_GETATTR(vp, &va, td->td_ucred);
687 			if (error != 0 || off >= va.va_size) {
688 				VOP_UNLOCK(vp, 0);
689 				goto done;
690 			}
691 			if (va.va_size != obj_size) {
692 				if (nbytes == 0)
693 					rem += va.va_size - obj_size;
694 				else if (offset + nbytes > va.va_size)
695 					rem -= (offset + nbytes - va.va_size);
696 				obj_size = va.va_size;
697 			}
698 		}
699 
700 		if (space > rem)
701 			space = rem;
702 
703 		npages = howmany(space + (off & PAGE_MASK), PAGE_SIZE);
704 
705 		/*
706 		 * Calculate maximum allowed number of pages for readahead
707 		 * at this iteration.  If SF_USER_READAHEAD was set, we don't
708 		 * do any heuristics and use exactly the value supplied by
709 		 * application.  Otherwise, we allow readahead up to "rem".
710 		 * If application wants more, let it be, but there is no
711 		 * reason to go above MAXPHYS.  Also check against "obj_size",
712 		 * since vm_pager_has_page() can hint beyond EOF.
713 		 */
714 		if (flags & SF_USER_READAHEAD) {
715 			rhpages = SF_READAHEAD(flags);
716 		} else {
717 			rhpages = howmany(rem + (off & PAGE_MASK), PAGE_SIZE) -
718 			    npages;
719 			rhpages += SF_READAHEAD(flags);
720 		}
721 		rhpages = min(howmany(MAXPHYS, PAGE_SIZE), rhpages);
722 		rhpages = min(howmany(obj_size - trunc_page(off), PAGE_SIZE) -
723 		    npages, rhpages);
724 
725 		sfio = malloc(sizeof(struct sf_io) +
726 		    npages * sizeof(vm_page_t), M_TEMP, M_WAITOK);
727 		refcount_init(&sfio->nios, 1);
728 		sfio->error = 0;
729 
730 		nios = sendfile_swapin(obj, sfio, off, space, npages, rhpages,
731 		    flags);
732 
733 		/*
734 		 * Loop and construct maximum sized mbuf chain to be bulk
735 		 * dumped into socket buffer.
736 		 */
737 		pa = sfio->pa;
738 		for (int i = 0; i < npages; i++) {
739 			struct mbuf *m0;
740 
741 			/*
742 			 * If a page wasn't grabbed successfully, then
743 			 * trim the array. Can happen only with SF_NODISKIO.
744 			 */
745 			if (pa[i] == NULL) {
746 				SFSTAT_INC(sf_busy);
747 				fixspace(npages, i, off, &space);
748 				npages = i;
749 				softerr = EBUSY;
750 				break;
751 			}
752 
753 			/*
754 			 * Get a sendfile buf.  When allocating the
755 			 * first buffer for mbuf chain, we usually
756 			 * wait as long as necessary, but this wait
757 			 * can be interrupted.  For consequent
758 			 * buffers, do not sleep, since several
759 			 * threads might exhaust the buffers and then
760 			 * deadlock.
761 			 */
762 			sf = sf_buf_alloc(pa[i],
763 			    m != NULL ? SFB_NOWAIT : SFB_CATCH);
764 			if (sf == NULL) {
765 				SFSTAT_INC(sf_allocfail);
766 				for (int j = i; j < npages; j++) {
767 					vm_page_lock(pa[j]);
768 					vm_page_unwire(pa[j], PQ_INACTIVE);
769 					vm_page_unlock(pa[j]);
770 				}
771 				if (m == NULL)
772 					softerr = ENOBUFS;
773 				fixspace(npages, i, off, &space);
774 				npages = i;
775 				break;
776 			}
777 
778 			m0 = m_get(M_WAITOK, MT_DATA);
779 			m0->m_ext.ext_buf = (char *)sf_buf_kva(sf);
780 			m0->m_ext.ext_size = PAGE_SIZE;
781 			m0->m_ext.ext_arg1 = sf;
782 			m0->m_ext.ext_arg2 = sfs;
783 			/*
784 			 * SF_NOCACHE sets the page as being freed upon send.
785 			 * However, we ignore it for the last page in 'space',
786 			 * if the page is truncated, and we got more data to
787 			 * send (rem > space), or if we have readahead
788 			 * configured (rhpages > 0).
789 			 */
790 			if ((flags & SF_NOCACHE) == 0 ||
791 			    (i == npages - 1 &&
792 			    ((off + space) & PAGE_MASK) &&
793 			    (rem > space || rhpages > 0)))
794 				m0->m_ext.ext_type = EXT_SFBUF;
795 			else
796 				m0->m_ext.ext_type = EXT_SFBUF_NOCACHE;
797 			m0->m_ext.ext_flags = EXT_FLAG_EMBREF;
798 			m0->m_ext.ext_count = 1;
799 			m0->m_flags |= (M_EXT | M_RDONLY);
800 			if (nios)
801 				m0->m_flags |= M_NOTREADY;
802 			m0->m_data = (char *)sf_buf_kva(sf) +
803 			    (vmoff(i, off) & PAGE_MASK);
804 			m0->m_len = xfsize(i, npages, off, space);
805 
806 			if (i == 0)
807 				sfio->m = m0;
808 
809 			/* Append to mbuf chain. */
810 			if (mtail != NULL)
811 				mtail->m_next = m0;
812 			else
813 				m = m0;
814 			mtail = m0;
815 
816 			if (sfs != NULL) {
817 				mtx_lock(&sfs->mtx);
818 				sfs->count++;
819 				mtx_unlock(&sfs->mtx);
820 			}
821 		}
822 
823 		if (vp != NULL)
824 			VOP_UNLOCK(vp, 0);
825 
826 		/* Keep track of bytes processed. */
827 		off += space;
828 		rem -= space;
829 
830 		/* Prepend header, if any. */
831 		if (hdrlen) {
832 prepend_header:
833 			mhtail->m_next = m;
834 			m = mh;
835 			mh = NULL;
836 		}
837 
838 		if (m == NULL) {
839 			KASSERT(softerr, ("%s: m NULL, no error", __func__));
840 			error = softerr;
841 			free(sfio, M_TEMP);
842 			goto done;
843 		}
844 
845 		/* Add the buffer chain to the socket buffer. */
846 		KASSERT(m_length(m, NULL) == space + hdrlen,
847 		    ("%s: mlen %u space %d hdrlen %d",
848 		    __func__, m_length(m, NULL), space, hdrlen));
849 
850 		CURVNET_SET(so->so_vnet);
851 		if (nios == 0) {
852 			/*
853 			 * If sendfile_swapin() didn't initiate any I/Os,
854 			 * which happens if all data is cached in VM, then
855 			 * we can send data right now without the
856 			 * PRUS_NOTREADY flag.
857 			 */
858 			free(sfio, M_TEMP);
859 			error = (*so->so_proto->pr_usrreqs->pru_send)
860 			    (so, 0, m, NULL, NULL, td);
861 		} else {
862 			sfio->sock_fp = sock_fp;
863 			sfio->npages = npages;
864 			fhold(sock_fp);
865 			error = (*so->so_proto->pr_usrreqs->pru_send)
866 			    (so, PRUS_NOTREADY, m, NULL, NULL, td);
867 			sendfile_iodone(sfio, NULL, 0, 0);
868 		}
869 		CURVNET_RESTORE();
870 
871 		m = NULL;	/* pru_send always consumes */
872 		if (error)
873 			goto done;
874 		sbytes += space + hdrlen;
875 		if (hdrlen)
876 			hdrlen = 0;
877 		if (softerr) {
878 			error = softerr;
879 			goto done;
880 		}
881 	}
882 
883 	/*
884 	 * Send trailers. Wimp out and use writev(2).
885 	 */
886 	if (trl_uio != NULL) {
887 		sbunlock(&so->so_snd);
888 		error = kern_writev(td, sockfd, trl_uio);
889 		if (error == 0)
890 			sbytes += td->td_retval[0];
891 		goto out;
892 	}
893 
894 done:
895 	sbunlock(&so->so_snd);
896 out:
897 	/*
898 	 * If there was no error we have to clear td->td_retval[0]
899 	 * because it may have been set by writev.
900 	 */
901 	if (error == 0) {
902 		td->td_retval[0] = 0;
903 	}
904 	if (sent != NULL) {
905 		(*sent) = sbytes;
906 	}
907 	if (obj != NULL)
908 		vm_object_deallocate(obj);
909 	if (so)
910 		fdrop(sock_fp, td);
911 	if (m)
912 		m_freem(m);
913 	if (mh)
914 		m_freem(mh);
915 
916 	if (sfs != NULL) {
917 		mtx_lock(&sfs->mtx);
918 		if (sfs->count != 0)
919 			cv_wait(&sfs->cv, &sfs->mtx);
920 		KASSERT(sfs->count == 0, ("sendfile sync still busy"));
921 		cv_destroy(&sfs->cv);
922 		mtx_destroy(&sfs->mtx);
923 		free(sfs, M_TEMP);
924 	}
925 
926 	if (error == ERESTART)
927 		error = EINTR;
928 
929 	return (error);
930 }
931 
932 static int
933 sendfile(struct thread *td, struct sendfile_args *uap, int compat)
934 {
935 	struct sf_hdtr hdtr;
936 	struct uio *hdr_uio, *trl_uio;
937 	struct file *fp;
938 	cap_rights_t rights;
939 	off_t sbytes;
940 	int error;
941 
942 	/*
943 	 * File offset must be positive.  If it goes beyond EOF
944 	 * we send only the header/trailer and no payload data.
945 	 */
946 	if (uap->offset < 0)
947 		return (EINVAL);
948 
949 	hdr_uio = trl_uio = NULL;
950 
951 	if (uap->hdtr != NULL) {
952 		error = copyin(uap->hdtr, &hdtr, sizeof(hdtr));
953 		if (error != 0)
954 			goto out;
955 		if (hdtr.headers != NULL) {
956 			error = copyinuio(hdtr.headers, hdtr.hdr_cnt,
957 			    &hdr_uio);
958 			if (error != 0)
959 				goto out;
960 #ifdef COMPAT_FREEBSD4
961 			/*
962 			 * In FreeBSD < 5.0 the nbytes to send also included
963 			 * the header.  If compat is specified subtract the
964 			 * header size from nbytes.
965 			 */
966 			if (compat) {
967 				if (uap->nbytes > hdr_uio->uio_resid)
968 					uap->nbytes -= hdr_uio->uio_resid;
969 				else
970 					uap->nbytes = 0;
971 			}
972 #endif
973 		}
974 		if (hdtr.trailers != NULL) {
975 			error = copyinuio(hdtr.trailers, hdtr.trl_cnt,
976 			    &trl_uio);
977 			if (error != 0)
978 				goto out;
979 		}
980 	}
981 
982 	AUDIT_ARG_FD(uap->fd);
983 
984 	/*
985 	 * sendfile(2) can start at any offset within a file so we require
986 	 * CAP_READ+CAP_SEEK = CAP_PREAD.
987 	 */
988 	if ((error = fget_read(td, uap->fd,
989 	    cap_rights_init(&rights, CAP_PREAD), &fp)) != 0) {
990 		goto out;
991 	}
992 
993 	error = fo_sendfile(fp, uap->s, hdr_uio, trl_uio, uap->offset,
994 	    uap->nbytes, &sbytes, uap->flags, td);
995 	fdrop(fp, td);
996 
997 	if (uap->sbytes != NULL)
998 		copyout(&sbytes, uap->sbytes, sizeof(off_t));
999 
1000 out:
1001 	free(hdr_uio, M_IOV);
1002 	free(trl_uio, M_IOV);
1003 	return (error);
1004 }
1005 
1006 /*
1007  * sendfile(2)
1008  *
1009  * int sendfile(int fd, int s, off_t offset, size_t nbytes,
1010  *       struct sf_hdtr *hdtr, off_t *sbytes, int flags)
1011  *
1012  * Send a file specified by 'fd' and starting at 'offset' to a socket
1013  * specified by 's'. Send only 'nbytes' of the file or until EOF if nbytes ==
1014  * 0.  Optionally add a header and/or trailer to the socket output.  If
1015  * specified, write the total number of bytes sent into *sbytes.
1016  */
1017 int
1018 sys_sendfile(struct thread *td, struct sendfile_args *uap)
1019 {
1020 
1021 	return (sendfile(td, uap, 0));
1022 }
1023 
1024 #ifdef COMPAT_FREEBSD4
1025 int
1026 freebsd4_sendfile(struct thread *td, struct freebsd4_sendfile_args *uap)
1027 {
1028 	struct sendfile_args args;
1029 
1030 	args.fd = uap->fd;
1031 	args.s = uap->s;
1032 	args.offset = uap->offset;
1033 	args.nbytes = uap->nbytes;
1034 	args.hdtr = uap->hdtr;
1035 	args.sbytes = uap->sbytes;
1036 	args.flags = uap->flags;
1037 
1038 	return (sendfile(td, &args, 1));
1039 }
1040 #endif /* COMPAT_FREEBSD4 */
1041