xref: /freebsd/sys/kern/kern_sendfile.c (revision 60eddb209b5ad13a549ca74a41b7cb38a31da5ef)
1 /*-
2  * Copyright (c) 2013-2015 Gleb Smirnoff <glebius@FreeBSD.org>
3  * Copyright (c) 1998, David Greenman. All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  * 3. Neither the name of the University nor the names of its contributors
14  *    may be used to endorse or promote products derived from this software
15  *    without specific prior written permission.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
32 
33 #include "opt_compat.h"
34 
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/capsicum.h>
38 #include <sys/kernel.h>
39 #include <sys/lock.h>
40 #include <sys/mutex.h>
41 #include <sys/sysproto.h>
42 #include <sys/malloc.h>
43 #include <sys/proc.h>
44 #include <sys/mman.h>
45 #include <sys/mount.h>
46 #include <sys/mbuf.h>
47 #include <sys/protosw.h>
48 #include <sys/rwlock.h>
49 #include <sys/sf_buf.h>
50 #include <sys/socket.h>
51 #include <sys/socketvar.h>
52 #include <sys/syscallsubr.h>
53 #include <sys/sysctl.h>
54 #include <sys/vnode.h>
55 
56 #include <net/vnet.h>
57 
58 #include <security/audit/audit.h>
59 #include <security/mac/mac_framework.h>
60 
61 #include <vm/vm.h>
62 #include <vm/vm_object.h>
63 #include <vm/vm_pager.h>
64 
65 #define	EXT_FLAG_SYNC		EXT_FLAG_VENDOR1
66 #define	EXT_FLAG_NOCACHE	EXT_FLAG_VENDOR2
67 
68 /*
69  * Structure describing a single sendfile(2) I/O, which may consist of
70  * several underlying pager I/Os.
71  *
72  * The syscall context allocates the structure and initializes 'nios'
73  * to 1.  As sendfile_swapin() runs through pages and starts asynchronous
74  * paging operations, it increments 'nios'.
75  *
76  * Every I/O completion calls sendfile_iodone(), which decrements the 'nios',
77  * and the syscall also calls sendfile_iodone() after allocating all mbufs,
78  * linking them and sending to socket.  Whoever reaches zero 'nios' is
79  * responsible to * call pru_ready on the socket, to notify it of readyness
80  * of the data.
81  */
82 struct sf_io {
83 	volatile u_int	nios;
84 	u_int		error;
85 	int		npages;
86 	struct socket	*so;
87 	struct mbuf	*m;
88 	vm_page_t	pa[];
89 };
90 
91 /*
92  * Structure used to track requests with SF_SYNC flag.
93  */
94 struct sendfile_sync {
95 	struct mtx	mtx;
96 	struct cv	cv;
97 	unsigned	count;
98 };
99 
100 counter_u64_t sfstat[sizeof(struct sfstat) / sizeof(uint64_t)];
101 
102 static void
103 sfstat_init(const void *unused)
104 {
105 
106 	COUNTER_ARRAY_ALLOC(sfstat, sizeof(struct sfstat) / sizeof(uint64_t),
107 	    M_WAITOK);
108 }
109 SYSINIT(sfstat, SI_SUB_MBUF, SI_ORDER_FIRST, sfstat_init, NULL);
110 
111 static int
112 sfstat_sysctl(SYSCTL_HANDLER_ARGS)
113 {
114 	struct sfstat s;
115 
116 	COUNTER_ARRAY_COPY(sfstat, &s, sizeof(s) / sizeof(uint64_t));
117 	if (req->newptr)
118 		COUNTER_ARRAY_ZERO(sfstat, sizeof(s) / sizeof(uint64_t));
119 	return (SYSCTL_OUT(req, &s, sizeof(s)));
120 }
121 SYSCTL_PROC(_kern_ipc, OID_AUTO, sfstat, CTLTYPE_OPAQUE | CTLFLAG_RW,
122     NULL, 0, sfstat_sysctl, "I", "sendfile statistics");
123 
124 /*
125  * Detach mapped page and release resources back to the system.  Called
126  * by mbuf(9) code when last reference to a page is freed.
127  */
128 static void
129 sendfile_free_page(vm_page_t pg, bool nocache)
130 {
131 
132 	vm_page_lock(pg);
133 	/*
134 	 * In either case check for the object going away on us.  This can
135 	 * happen since we don't hold a reference to it.  If so, we're
136 	 * responsible for freeing the page.  In 'noncache' case try to free
137 	 * the page, but only if it is cheap to.
138 	 */
139 	if (vm_page_unwire(pg, nocache ? PQ_NONE : PQ_INACTIVE)) {
140 		vm_object_t obj;
141 
142 		if ((obj = pg->object) == NULL)
143 			vm_page_free(pg);
144 		else if (nocache) {
145 			if (!vm_page_xbusied(pg) && VM_OBJECT_TRYWLOCK(obj)) {
146 				bool freed;
147 
148 				/* Only free unmapped pages. */
149 				if (obj->ref_count == 0 ||
150 				    !pmap_page_is_mapped(pg))
151 					/*
152 					 * The busy test before the object is
153 					 * locked cannot be relied upon.
154 					 */
155 					freed = vm_page_try_to_free(pg);
156 				else
157 					freed = false;
158 				VM_OBJECT_WUNLOCK(obj);
159 				if (!freed)
160 					vm_page_deactivate_noreuse(pg);
161 			} else
162 				vm_page_deactivate_noreuse(pg);
163 		}
164 	}
165 	vm_page_unlock(pg);
166 }
167 
168 static void
169 sendfile_free_mext(struct mbuf *m)
170 {
171 	struct sf_buf *sf;
172 	vm_page_t pg;
173 	bool nocache;
174 
175 	KASSERT(m->m_flags & M_EXT && m->m_ext.ext_type == EXT_SFBUF,
176 	    ("%s: m %p !M_EXT or !EXT_SFBUF", __func__, m));
177 
178 	sf = m->m_ext.ext_arg1;
179 	pg = sf_buf_page(sf);
180 	nocache = m->m_ext.ext_flags & EXT_FLAG_NOCACHE;
181 
182 	sf_buf_free(sf);
183 	sendfile_free_page(pg, nocache);
184 
185 	if (m->m_ext.ext_flags & EXT_FLAG_SYNC) {
186 		struct sendfile_sync *sfs = m->m_ext.ext_arg2;
187 
188 		mtx_lock(&sfs->mtx);
189 		KASSERT(sfs->count > 0, ("Sendfile sync botchup count == 0"));
190 		if (--sfs->count == 0)
191 			cv_signal(&sfs->cv);
192 		mtx_unlock(&sfs->mtx);
193 	}
194 }
195 
196 /*
197  * Helper function to calculate how much data to put into page i of n.
198  * Only first and last pages are special.
199  */
200 static inline off_t
201 xfsize(int i, int n, off_t off, off_t len)
202 {
203 
204 	if (i == 0)
205 		return (omin(PAGE_SIZE - (off & PAGE_MASK), len));
206 
207 	if (i == n - 1 && ((off + len) & PAGE_MASK) > 0)
208 		return ((off + len) & PAGE_MASK);
209 
210 	return (PAGE_SIZE);
211 }
212 
213 /*
214  * Helper function to get offset within object for i page.
215  */
216 static inline vm_ooffset_t
217 vmoff(int i, off_t off)
218 {
219 
220 	if (i == 0)
221 		return ((vm_ooffset_t)off);
222 
223 	return (trunc_page(off + i * PAGE_SIZE));
224 }
225 
226 /*
227  * Helper function used when allocation of a page or sf_buf failed.
228  * Pretend as if we don't have enough space, subtract xfsize() of
229  * all pages that failed.
230  */
231 static inline void
232 fixspace(int old, int new, off_t off, int *space)
233 {
234 
235 	KASSERT(old > new, ("%s: old %d new %d", __func__, old, new));
236 
237 	/* Subtract last one. */
238 	*space -= xfsize(old - 1, old, off, *space);
239 	old--;
240 
241 	if (new == old)
242 		/* There was only one page. */
243 		return;
244 
245 	/* Subtract first one. */
246 	if (new == 0) {
247 		*space -= xfsize(0, old, off, *space);
248 		new++;
249 	}
250 
251 	/* Rest of pages are full sized. */
252 	*space -= (old - new) * PAGE_SIZE;
253 
254 	KASSERT(*space >= 0, ("%s: space went backwards", __func__));
255 }
256 
257 /*
258  * I/O completion callback.
259  */
260 static void
261 sendfile_iodone(void *arg, vm_page_t *pg, int count, int error)
262 {
263 	struct sf_io *sfio = arg;
264 	struct socket *so = sfio->so;
265 
266 	for (int i = 0; i < count; i++)
267 		if (pg[i] != bogus_page)
268 			vm_page_xunbusy(pg[i]);
269 
270 	if (error)
271 		sfio->error = error;
272 
273 	if (!refcount_release(&sfio->nios))
274 		return;
275 
276 	CURVNET_SET(so->so_vnet);
277 	if (sfio->error) {
278 		struct mbuf *m;
279 
280 		/*
281 		 * I/O operation failed.  The state of data in the socket
282 		 * is now inconsistent, and all what we can do is to tear
283 		 * it down. Protocol abort method would tear down protocol
284 		 * state, free all ready mbufs and detach not ready ones.
285 		 * We will free the mbufs corresponding to this I/O manually.
286 		 *
287 		 * The socket would be marked with EIO and made available
288 		 * for read, so that application receives EIO on next
289 		 * syscall and eventually closes the socket.
290 		 */
291 		so->so_proto->pr_usrreqs->pru_abort(so);
292 		so->so_error = EIO;
293 
294 		m = sfio->m;
295 		for (int i = 0; i < sfio->npages; i++)
296 			m = m_free(m);
297 	} else
298 		(void )(so->so_proto->pr_usrreqs->pru_ready)(so, sfio->m,
299 		    sfio->npages);
300 
301 	SOCK_LOCK(so);
302 	sorele(so);
303 	CURVNET_RESTORE();
304 	free(sfio, M_TEMP);
305 }
306 
307 /*
308  * Iterate through pages vector and request paging for non-valid pages.
309  */
310 static int
311 sendfile_swapin(vm_object_t obj, struct sf_io *sfio, off_t off, off_t len,
312     int npages, int rhpages, int flags)
313 {
314 	vm_page_t *pa = sfio->pa;
315 	int grabbed, nios;
316 
317 	nios = 0;
318 	flags = (flags & SF_NODISKIO) ? VM_ALLOC_NOWAIT : 0;
319 
320 	/*
321 	 * First grab all the pages and wire them.  Note that we grab
322 	 * only required pages.  Readahead pages are dealt with later.
323 	 */
324 	VM_OBJECT_WLOCK(obj);
325 
326 	grabbed = vm_page_grab_pages(obj, OFF_TO_IDX(off),
327 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED | flags, pa, npages);
328 	if (grabbed < npages) {
329 		for (int i = grabbed; i < npages; i++)
330 			pa[i] = NULL;
331 		npages = grabbed;
332 		rhpages = 0;
333 	}
334 
335 	for (int i = 0; i < npages;) {
336 		int j, a, count, rv;
337 
338 		/* Skip valid pages. */
339 		if (vm_page_is_valid(pa[i], vmoff(i, off) & PAGE_MASK,
340 		    xfsize(i, npages, off, len))) {
341 			vm_page_xunbusy(pa[i]);
342 			SFSTAT_INC(sf_pages_valid);
343 			i++;
344 			continue;
345 		}
346 
347 		/*
348 		 * Next page is invalid.  Check if it belongs to pager.  It
349 		 * may not be there, which is a regular situation for shmem
350 		 * pager.  For vnode pager this happens only in case of
351 		 * a sparse file.
352 		 *
353 		 * Important feature of vm_pager_has_page() is the hint
354 		 * stored in 'a', about how many pages we can pagein after
355 		 * this page in a single I/O.
356 		 */
357 		if (!vm_pager_has_page(obj, OFF_TO_IDX(vmoff(i, off)), NULL,
358 		    &a)) {
359 			pmap_zero_page(pa[i]);
360 			pa[i]->valid = VM_PAGE_BITS_ALL;
361 			MPASS(pa[i]->dirty == 0);
362 			vm_page_xunbusy(pa[i]);
363 			i++;
364 			continue;
365 		}
366 
367 		/*
368 		 * We want to pagein as many pages as possible, limited only
369 		 * by the 'a' hint and actual request.
370 		 */
371 		count = min(a + 1, npages - i);
372 
373 		/*
374 		 * We should not pagein into a valid page, thus we first trim
375 		 * any valid pages off the end of request, and substitute
376 		 * to bogus_page those, that are in the middle.
377 		 */
378 		for (j = i + count - 1; j > i; j--) {
379 			if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
380 			    xfsize(j, npages, off, len))) {
381 				count--;
382 				rhpages = 0;
383 			} else
384 				break;
385 		}
386 		for (j = i + 1; j < i + count - 1; j++)
387 			if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
388 			    xfsize(j, npages, off, len))) {
389 				vm_page_xunbusy(pa[j]);
390 				SFSTAT_INC(sf_pages_valid);
391 				SFSTAT_INC(sf_pages_bogus);
392 				pa[j] = bogus_page;
393 			}
394 
395 		refcount_acquire(&sfio->nios);
396 		rv = vm_pager_get_pages_async(obj, pa + i, count, NULL,
397 		    i + count == npages ? &rhpages : NULL,
398 		    &sendfile_iodone, sfio);
399 		KASSERT(rv == VM_PAGER_OK, ("%s: pager fail obj %p page %p",
400 		    __func__, obj, pa[i]));
401 
402 		SFSTAT_INC(sf_iocnt);
403 		SFSTAT_ADD(sf_pages_read, count);
404 		if (i + count == npages)
405 			SFSTAT_ADD(sf_rhpages_read, rhpages);
406 
407 		/*
408 		 * Restore the valid page pointers.  They are already
409 		 * unbusied, but still wired.
410 		 */
411 		for (j = i; j < i + count; j++)
412 			if (pa[j] == bogus_page) {
413 				pa[j] = vm_page_lookup(obj,
414 				    OFF_TO_IDX(vmoff(j, off)));
415 				KASSERT(pa[j], ("%s: page %p[%d] disappeared",
416 				    __func__, pa, j));
417 
418 			}
419 		i += count;
420 		nios++;
421 	}
422 
423 	VM_OBJECT_WUNLOCK(obj);
424 
425 	if (nios == 0 && npages != 0)
426 		SFSTAT_INC(sf_noiocnt);
427 
428 	return (nios);
429 }
430 
431 static int
432 sendfile_getobj(struct thread *td, struct file *fp, vm_object_t *obj_res,
433     struct vnode **vp_res, struct shmfd **shmfd_res, off_t *obj_size,
434     int *bsize)
435 {
436 	struct vattr va;
437 	vm_object_t obj;
438 	struct vnode *vp;
439 	struct shmfd *shmfd;
440 	int error;
441 
442 	vp = *vp_res = NULL;
443 	obj = NULL;
444 	shmfd = *shmfd_res = NULL;
445 	*bsize = 0;
446 
447 	/*
448 	 * The file descriptor must be a regular file and have a
449 	 * backing VM object.
450 	 */
451 	if (fp->f_type == DTYPE_VNODE) {
452 		vp = fp->f_vnode;
453 		vn_lock(vp, LK_SHARED | LK_RETRY);
454 		if (vp->v_type != VREG) {
455 			error = EINVAL;
456 			goto out;
457 		}
458 		*bsize = vp->v_mount->mnt_stat.f_iosize;
459 		error = VOP_GETATTR(vp, &va, td->td_ucred);
460 		if (error != 0)
461 			goto out;
462 		*obj_size = va.va_size;
463 		obj = vp->v_object;
464 		if (obj == NULL) {
465 			error = EINVAL;
466 			goto out;
467 		}
468 	} else if (fp->f_type == DTYPE_SHM) {
469 		error = 0;
470 		shmfd = fp->f_data;
471 		obj = shmfd->shm_object;
472 		*obj_size = shmfd->shm_size;
473 	} else {
474 		error = EINVAL;
475 		goto out;
476 	}
477 
478 	VM_OBJECT_WLOCK(obj);
479 	if ((obj->flags & OBJ_DEAD) != 0) {
480 		VM_OBJECT_WUNLOCK(obj);
481 		error = EBADF;
482 		goto out;
483 	}
484 
485 	/*
486 	 * Temporarily increase the backing VM object's reference
487 	 * count so that a forced reclamation of its vnode does not
488 	 * immediately destroy it.
489 	 */
490 	vm_object_reference_locked(obj);
491 	VM_OBJECT_WUNLOCK(obj);
492 	*obj_res = obj;
493 	*vp_res = vp;
494 	*shmfd_res = shmfd;
495 
496 out:
497 	if (vp != NULL)
498 		VOP_UNLOCK(vp, 0);
499 	return (error);
500 }
501 
502 static int
503 sendfile_getsock(struct thread *td, int s, struct file **sock_fp,
504     struct socket **so)
505 {
506 	cap_rights_t rights;
507 	int error;
508 
509 	*sock_fp = NULL;
510 	*so = NULL;
511 
512 	/*
513 	 * The socket must be a stream socket and connected.
514 	 */
515 	error = getsock_cap(td, s, cap_rights_init(&rights, CAP_SEND),
516 	    sock_fp, NULL, NULL);
517 	if (error != 0)
518 		return (error);
519 	*so = (*sock_fp)->f_data;
520 	if ((*so)->so_type != SOCK_STREAM)
521 		return (EINVAL);
522 	return (0);
523 }
524 
525 int
526 vn_sendfile(struct file *fp, int sockfd, struct uio *hdr_uio,
527     struct uio *trl_uio, off_t offset, size_t nbytes, off_t *sent, int flags,
528     struct thread *td)
529 {
530 	struct file *sock_fp;
531 	struct vnode *vp;
532 	struct vm_object *obj;
533 	struct socket *so;
534 	struct mbuf *m, *mh, *mhtail;
535 	struct sf_buf *sf;
536 	struct shmfd *shmfd;
537 	struct sendfile_sync *sfs;
538 	struct vattr va;
539 	off_t off, sbytes, rem, obj_size;
540 	int error, softerr, bsize, hdrlen;
541 
542 	obj = NULL;
543 	so = NULL;
544 	m = mh = NULL;
545 	sfs = NULL;
546 	hdrlen = sbytes = 0;
547 	softerr = 0;
548 
549 	error = sendfile_getobj(td, fp, &obj, &vp, &shmfd, &obj_size, &bsize);
550 	if (error != 0)
551 		return (error);
552 
553 	error = sendfile_getsock(td, sockfd, &sock_fp, &so);
554 	if (error != 0)
555 		goto out;
556 
557 #ifdef MAC
558 	error = mac_socket_check_send(td->td_ucred, so);
559 	if (error != 0)
560 		goto out;
561 #endif
562 
563 	SFSTAT_INC(sf_syscalls);
564 	SFSTAT_ADD(sf_rhpages_requested, SF_READAHEAD(flags));
565 
566 	if (flags & SF_SYNC) {
567 		sfs = malloc(sizeof *sfs, M_TEMP, M_WAITOK | M_ZERO);
568 		mtx_init(&sfs->mtx, "sendfile", NULL, MTX_DEF);
569 		cv_init(&sfs->cv, "sendfile");
570 	}
571 
572 	rem = nbytes ? omin(nbytes, obj_size - offset) : obj_size - offset;
573 
574 	/*
575 	 * Protect against multiple writers to the socket.
576 	 *
577 	 * XXXRW: Historically this has assumed non-interruptibility, so now
578 	 * we implement that, but possibly shouldn't.
579 	 */
580 	(void)sblock(&so->so_snd, SBL_WAIT | SBL_NOINTR);
581 
582 	/*
583 	 * Loop through the pages of the file, starting with the requested
584 	 * offset. Get a file page (do I/O if necessary), map the file page
585 	 * into an sf_buf, attach an mbuf header to the sf_buf, and queue
586 	 * it on the socket.
587 	 * This is done in two loops.  The inner loop turns as many pages
588 	 * as it can, up to available socket buffer space, without blocking
589 	 * into mbufs to have it bulk delivered into the socket send buffer.
590 	 * The outer loop checks the state and available space of the socket
591 	 * and takes care of the overall progress.
592 	 */
593 	for (off = offset; rem > 0; ) {
594 		struct sf_io *sfio;
595 		vm_page_t *pa;
596 		struct mbuf *mtail;
597 		int nios, space, npages, rhpages;
598 
599 		mtail = NULL;
600 		/*
601 		 * Check the socket state for ongoing connection,
602 		 * no errors and space in socket buffer.
603 		 * If space is low allow for the remainder of the
604 		 * file to be processed if it fits the socket buffer.
605 		 * Otherwise block in waiting for sufficient space
606 		 * to proceed, or if the socket is nonblocking, return
607 		 * to userland with EAGAIN while reporting how far
608 		 * we've come.
609 		 * We wait until the socket buffer has significant free
610 		 * space to do bulk sends.  This makes good use of file
611 		 * system read ahead and allows packet segmentation
612 		 * offloading hardware to take over lots of work.  If
613 		 * we were not careful here we would send off only one
614 		 * sfbuf at a time.
615 		 */
616 		SOCKBUF_LOCK(&so->so_snd);
617 		if (so->so_snd.sb_lowat < so->so_snd.sb_hiwat / 2)
618 			so->so_snd.sb_lowat = so->so_snd.sb_hiwat / 2;
619 retry_space:
620 		if (so->so_snd.sb_state & SBS_CANTSENDMORE) {
621 			error = EPIPE;
622 			SOCKBUF_UNLOCK(&so->so_snd);
623 			goto done;
624 		} else if (so->so_error) {
625 			error = so->so_error;
626 			so->so_error = 0;
627 			SOCKBUF_UNLOCK(&so->so_snd);
628 			goto done;
629 		}
630 		if ((so->so_state & SS_ISCONNECTED) == 0) {
631 			SOCKBUF_UNLOCK(&so->so_snd);
632 			error = ENOTCONN;
633 			goto done;
634 		}
635 
636 		space = sbspace(&so->so_snd);
637 		if (space < rem &&
638 		    (space <= 0 ||
639 		     space < so->so_snd.sb_lowat)) {
640 			if (so->so_state & SS_NBIO) {
641 				SOCKBUF_UNLOCK(&so->so_snd);
642 				error = EAGAIN;
643 				goto done;
644 			}
645 			/*
646 			 * sbwait drops the lock while sleeping.
647 			 * When we loop back to retry_space the
648 			 * state may have changed and we retest
649 			 * for it.
650 			 */
651 			error = sbwait(&so->so_snd);
652 			/*
653 			 * An error from sbwait usually indicates that we've
654 			 * been interrupted by a signal. If we've sent anything
655 			 * then return bytes sent, otherwise return the error.
656 			 */
657 			if (error != 0) {
658 				SOCKBUF_UNLOCK(&so->so_snd);
659 				goto done;
660 			}
661 			goto retry_space;
662 		}
663 		SOCKBUF_UNLOCK(&so->so_snd);
664 
665 		/*
666 		 * At the beginning of the first loop check if any headers
667 		 * are specified and copy them into mbufs.  Reduce space in
668 		 * the socket buffer by the size of the header mbuf chain.
669 		 * Clear hdr_uio here and hdrlen at the end of the first loop.
670 		 */
671 		if (hdr_uio != NULL && hdr_uio->uio_resid > 0) {
672 			hdr_uio->uio_td = td;
673 			hdr_uio->uio_rw = UIO_WRITE;
674 			mh = m_uiotombuf(hdr_uio, M_WAITOK, space, 0, 0);
675 			hdrlen = m_length(mh, &mhtail);
676 			space -= hdrlen;
677 			/*
678 			 * If header consumed all the socket buffer space,
679 			 * don't waste CPU cycles and jump to the end.
680 			 */
681 			if (space == 0) {
682 				sfio = NULL;
683 				nios = 0;
684 				goto prepend_header;
685 			}
686 			hdr_uio = NULL;
687 		}
688 
689 		if (vp != NULL) {
690 			error = vn_lock(vp, LK_SHARED);
691 			if (error != 0)
692 				goto done;
693 			error = VOP_GETATTR(vp, &va, td->td_ucred);
694 			if (error != 0 || off >= va.va_size) {
695 				VOP_UNLOCK(vp, 0);
696 				goto done;
697 			}
698 			if (va.va_size != obj_size) {
699 				obj_size = va.va_size;
700 				rem = nbytes ?
701 				    omin(nbytes + offset, obj_size) : obj_size;
702 				rem -= off;
703 			}
704 		}
705 
706 		if (space > rem)
707 			space = rem;
708 
709 		npages = howmany(space + (off & PAGE_MASK), PAGE_SIZE);
710 
711 		/*
712 		 * Calculate maximum allowed number of pages for readahead
713 		 * at this iteration.  If SF_USER_READAHEAD was set, we don't
714 		 * do any heuristics and use exactly the value supplied by
715 		 * application.  Otherwise, we allow readahead up to "rem".
716 		 * If application wants more, let it be, but there is no
717 		 * reason to go above MAXPHYS.  Also check against "obj_size",
718 		 * since vm_pager_has_page() can hint beyond EOF.
719 		 */
720 		if (flags & SF_USER_READAHEAD) {
721 			rhpages = SF_READAHEAD(flags);
722 		} else {
723 			rhpages = howmany(rem + (off & PAGE_MASK), PAGE_SIZE) -
724 			    npages;
725 			rhpages += SF_READAHEAD(flags);
726 		}
727 		rhpages = min(howmany(MAXPHYS, PAGE_SIZE), rhpages);
728 		rhpages = min(howmany(obj_size - trunc_page(off), PAGE_SIZE) -
729 		    npages, rhpages);
730 
731 		sfio = malloc(sizeof(struct sf_io) +
732 		    npages * sizeof(vm_page_t), M_TEMP, M_WAITOK);
733 		refcount_init(&sfio->nios, 1);
734 		sfio->so = so;
735 		sfio->error = 0;
736 
737 		nios = sendfile_swapin(obj, sfio, off, space, npages, rhpages,
738 		    flags);
739 
740 		/*
741 		 * Loop and construct maximum sized mbuf chain to be bulk
742 		 * dumped into socket buffer.
743 		 */
744 		pa = sfio->pa;
745 		for (int i = 0; i < npages; i++) {
746 			struct mbuf *m0;
747 
748 			/*
749 			 * If a page wasn't grabbed successfully, then
750 			 * trim the array. Can happen only with SF_NODISKIO.
751 			 */
752 			if (pa[i] == NULL) {
753 				SFSTAT_INC(sf_busy);
754 				fixspace(npages, i, off, &space);
755 				npages = i;
756 				softerr = EBUSY;
757 				break;
758 			}
759 
760 			/*
761 			 * Get a sendfile buf.  When allocating the
762 			 * first buffer for mbuf chain, we usually
763 			 * wait as long as necessary, but this wait
764 			 * can be interrupted.  For consequent
765 			 * buffers, do not sleep, since several
766 			 * threads might exhaust the buffers and then
767 			 * deadlock.
768 			 */
769 			sf = sf_buf_alloc(pa[i],
770 			    m != NULL ? SFB_NOWAIT : SFB_CATCH);
771 			if (sf == NULL) {
772 				SFSTAT_INC(sf_allocfail);
773 				for (int j = i; j < npages; j++) {
774 					vm_page_lock(pa[j]);
775 					vm_page_unwire(pa[j], PQ_INACTIVE);
776 					vm_page_unlock(pa[j]);
777 				}
778 				if (m == NULL)
779 					softerr = ENOBUFS;
780 				fixspace(npages, i, off, &space);
781 				npages = i;
782 				break;
783 			}
784 
785 			m0 = m_get(M_WAITOK, MT_DATA);
786 			m0->m_ext.ext_buf = (char *)sf_buf_kva(sf);
787 			m0->m_ext.ext_size = PAGE_SIZE;
788 			m0->m_ext.ext_arg1 = sf;
789 			m0->m_ext.ext_type = EXT_SFBUF;
790 			m0->m_ext.ext_flags = EXT_FLAG_EMBREF;
791 			m0->m_ext.ext_free = sendfile_free_mext;
792 			/*
793 			 * SF_NOCACHE sets the page as being freed upon send.
794 			 * However, we ignore it for the last page in 'space',
795 			 * if the page is truncated, and we got more data to
796 			 * send (rem > space), or if we have readahead
797 			 * configured (rhpages > 0).
798 			 */
799 			if ((flags & SF_NOCACHE) &&
800 			    (i != npages - 1 ||
801 			    !((off + space) & PAGE_MASK) ||
802 			    !(rem > space || rhpages > 0)))
803 				m0->m_ext.ext_flags |= EXT_FLAG_NOCACHE;
804 			if (sfs != NULL) {
805 				m0->m_ext.ext_flags |= EXT_FLAG_SYNC;
806 				m0->m_ext.ext_arg2 = sfs;
807 				mtx_lock(&sfs->mtx);
808 				sfs->count++;
809 				mtx_unlock(&sfs->mtx);
810 			}
811 			m0->m_ext.ext_count = 1;
812 			m0->m_flags |= (M_EXT | M_RDONLY);
813 			if (nios)
814 				m0->m_flags |= M_NOTREADY;
815 			m0->m_data = (char *)sf_buf_kva(sf) +
816 			    (vmoff(i, off) & PAGE_MASK);
817 			m0->m_len = xfsize(i, npages, off, space);
818 
819 			if (i == 0)
820 				sfio->m = m0;
821 
822 			/* Append to mbuf chain. */
823 			if (mtail != NULL)
824 				mtail->m_next = m0;
825 			else
826 				m = m0;
827 			mtail = m0;
828 		}
829 
830 		if (vp != NULL)
831 			VOP_UNLOCK(vp, 0);
832 
833 		/* Keep track of bytes processed. */
834 		off += space;
835 		rem -= space;
836 
837 		/* Prepend header, if any. */
838 		if (hdrlen) {
839 prepend_header:
840 			mhtail->m_next = m;
841 			m = mh;
842 			mh = NULL;
843 		}
844 
845 		if (m == NULL) {
846 			KASSERT(softerr, ("%s: m NULL, no error", __func__));
847 			error = softerr;
848 			free(sfio, M_TEMP);
849 			goto done;
850 		}
851 
852 		/* Add the buffer chain to the socket buffer. */
853 		KASSERT(m_length(m, NULL) == space + hdrlen,
854 		    ("%s: mlen %u space %d hdrlen %d",
855 		    __func__, m_length(m, NULL), space, hdrlen));
856 
857 		CURVNET_SET(so->so_vnet);
858 		if (nios == 0) {
859 			/*
860 			 * If sendfile_swapin() didn't initiate any I/Os,
861 			 * which happens if all data is cached in VM, then
862 			 * we can send data right now without the
863 			 * PRUS_NOTREADY flag.
864 			 */
865 			free(sfio, M_TEMP);
866 			error = (*so->so_proto->pr_usrreqs->pru_send)
867 			    (so, 0, m, NULL, NULL, td);
868 		} else {
869 			sfio->npages = npages;
870 			soref(so);
871 			error = (*so->so_proto->pr_usrreqs->pru_send)
872 			    (so, PRUS_NOTREADY, m, NULL, NULL, td);
873 			sendfile_iodone(sfio, NULL, 0, 0);
874 		}
875 		CURVNET_RESTORE();
876 
877 		m = NULL;	/* pru_send always consumes */
878 		if (error)
879 			goto done;
880 		sbytes += space + hdrlen;
881 		if (hdrlen)
882 			hdrlen = 0;
883 		if (softerr) {
884 			error = softerr;
885 			goto done;
886 		}
887 	}
888 
889 	/*
890 	 * Send trailers. Wimp out and use writev(2).
891 	 */
892 	if (trl_uio != NULL) {
893 		sbunlock(&so->so_snd);
894 		error = kern_writev(td, sockfd, trl_uio);
895 		if (error == 0)
896 			sbytes += td->td_retval[0];
897 		goto out;
898 	}
899 
900 done:
901 	sbunlock(&so->so_snd);
902 out:
903 	/*
904 	 * If there was no error we have to clear td->td_retval[0]
905 	 * because it may have been set by writev.
906 	 */
907 	if (error == 0) {
908 		td->td_retval[0] = 0;
909 	}
910 	if (sent != NULL) {
911 		(*sent) = sbytes;
912 	}
913 	if (obj != NULL)
914 		vm_object_deallocate(obj);
915 	if (so)
916 		fdrop(sock_fp, td);
917 	if (m)
918 		m_freem(m);
919 	if (mh)
920 		m_freem(mh);
921 
922 	if (sfs != NULL) {
923 		mtx_lock(&sfs->mtx);
924 		if (sfs->count != 0)
925 			cv_wait(&sfs->cv, &sfs->mtx);
926 		KASSERT(sfs->count == 0, ("sendfile sync still busy"));
927 		cv_destroy(&sfs->cv);
928 		mtx_destroy(&sfs->mtx);
929 		free(sfs, M_TEMP);
930 	}
931 
932 	if (error == ERESTART)
933 		error = EINTR;
934 
935 	return (error);
936 }
937 
938 static int
939 sendfile(struct thread *td, struct sendfile_args *uap, int compat)
940 {
941 	struct sf_hdtr hdtr;
942 	struct uio *hdr_uio, *trl_uio;
943 	struct file *fp;
944 	cap_rights_t rights;
945 	off_t sbytes;
946 	int error;
947 
948 	/*
949 	 * File offset must be positive.  If it goes beyond EOF
950 	 * we send only the header/trailer and no payload data.
951 	 */
952 	if (uap->offset < 0)
953 		return (EINVAL);
954 
955 	sbytes = 0;
956 	hdr_uio = trl_uio = NULL;
957 
958 	if (uap->hdtr != NULL) {
959 		error = copyin(uap->hdtr, &hdtr, sizeof(hdtr));
960 		if (error != 0)
961 			goto out;
962 		if (hdtr.headers != NULL) {
963 			error = copyinuio(hdtr.headers, hdtr.hdr_cnt,
964 			    &hdr_uio);
965 			if (error != 0)
966 				goto out;
967 #ifdef COMPAT_FREEBSD4
968 			/*
969 			 * In FreeBSD < 5.0 the nbytes to send also included
970 			 * the header.  If compat is specified subtract the
971 			 * header size from nbytes.
972 			 */
973 			if (compat) {
974 				if (uap->nbytes > hdr_uio->uio_resid)
975 					uap->nbytes -= hdr_uio->uio_resid;
976 				else
977 					uap->nbytes = 0;
978 			}
979 #endif
980 		}
981 		if (hdtr.trailers != NULL) {
982 			error = copyinuio(hdtr.trailers, hdtr.trl_cnt,
983 			    &trl_uio);
984 			if (error != 0)
985 				goto out;
986 		}
987 	}
988 
989 	AUDIT_ARG_FD(uap->fd);
990 
991 	/*
992 	 * sendfile(2) can start at any offset within a file so we require
993 	 * CAP_READ+CAP_SEEK = CAP_PREAD.
994 	 */
995 	if ((error = fget_read(td, uap->fd,
996 	    cap_rights_init(&rights, CAP_PREAD), &fp)) != 0) {
997 		goto out;
998 	}
999 
1000 	error = fo_sendfile(fp, uap->s, hdr_uio, trl_uio, uap->offset,
1001 	    uap->nbytes, &sbytes, uap->flags, td);
1002 	fdrop(fp, td);
1003 
1004 	if (uap->sbytes != NULL)
1005 		copyout(&sbytes, uap->sbytes, sizeof(off_t));
1006 
1007 out:
1008 	free(hdr_uio, M_IOV);
1009 	free(trl_uio, M_IOV);
1010 	return (error);
1011 }
1012 
1013 /*
1014  * sendfile(2)
1015  *
1016  * int sendfile(int fd, int s, off_t offset, size_t nbytes,
1017  *       struct sf_hdtr *hdtr, off_t *sbytes, int flags)
1018  *
1019  * Send a file specified by 'fd' and starting at 'offset' to a socket
1020  * specified by 's'. Send only 'nbytes' of the file or until EOF if nbytes ==
1021  * 0.  Optionally add a header and/or trailer to the socket output.  If
1022  * specified, write the total number of bytes sent into *sbytes.
1023  */
1024 int
1025 sys_sendfile(struct thread *td, struct sendfile_args *uap)
1026 {
1027 
1028 	return (sendfile(td, uap, 0));
1029 }
1030 
1031 #ifdef COMPAT_FREEBSD4
1032 int
1033 freebsd4_sendfile(struct thread *td, struct freebsd4_sendfile_args *uap)
1034 {
1035 	struct sendfile_args args;
1036 
1037 	args.fd = uap->fd;
1038 	args.s = uap->s;
1039 	args.offset = uap->offset;
1040 	args.nbytes = uap->nbytes;
1041 	args.hdtr = uap->hdtr;
1042 	args.sbytes = uap->sbytes;
1043 	args.flags = uap->flags;
1044 
1045 	return (sendfile(td, &args, 1));
1046 }
1047 #endif /* COMPAT_FREEBSD4 */
1048