xref: /freebsd/sys/kern/kern_sendfile.c (revision 60fde7ce5d7bf5d94290720ea53db5701ab406a8)
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 	bool freed;
132 
133 	vm_page_lock(pg);
134 	/*
135 	 * In either case check for the object going away on us.  This can
136 	 * happen since we don't hold a reference to it.  If so, we're
137 	 * responsible for freeing the page.  In 'noncache' case try to free
138 	 * the page, but only if it is cheap to.
139 	 */
140 	if (vm_page_unwire_noq(pg)) {
141 		vm_object_t obj;
142 
143 		if ((obj = pg->object) == NULL)
144 			vm_page_free(pg);
145 		else {
146 			freed = false;
147 			if (nocache && !vm_page_xbusied(pg) &&
148 			    VM_OBJECT_TRYWLOCK(obj)) {
149 				/* Only free unmapped pages. */
150 				if (obj->ref_count == 0 ||
151 				    !pmap_page_is_mapped(pg))
152 					/*
153 					 * The busy test before the object is
154 					 * locked cannot be relied upon.
155 					 */
156 					freed = vm_page_try_to_free(pg);
157 				VM_OBJECT_WUNLOCK(obj);
158 			}
159 			if (!freed) {
160 				/*
161 				 * If we were asked to not cache the page, place
162 				 * it near the head of the inactive queue so
163 				 * that it is reclaimed sooner.  Otherwise,
164 				 * maintain LRU.
165 				 */
166 				if (nocache)
167 					vm_page_deactivate_noreuse(pg);
168 				else if (pg->queue == PQ_ACTIVE)
169 					vm_page_reference(pg);
170 				else
171 					vm_page_deactivate(pg);
172 			}
173 		}
174 	}
175 	vm_page_unlock(pg);
176 }
177 
178 static void
179 sendfile_free_mext(struct mbuf *m)
180 {
181 	struct sf_buf *sf;
182 	vm_page_t pg;
183 	bool nocache;
184 
185 	KASSERT(m->m_flags & M_EXT && m->m_ext.ext_type == EXT_SFBUF,
186 	    ("%s: m %p !M_EXT or !EXT_SFBUF", __func__, m));
187 
188 	sf = m->m_ext.ext_arg1;
189 	pg = sf_buf_page(sf);
190 	nocache = m->m_ext.ext_flags & EXT_FLAG_NOCACHE;
191 
192 	sf_buf_free(sf);
193 	sendfile_free_page(pg, nocache);
194 
195 	if (m->m_ext.ext_flags & EXT_FLAG_SYNC) {
196 		struct sendfile_sync *sfs = m->m_ext.ext_arg2;
197 
198 		mtx_lock(&sfs->mtx);
199 		KASSERT(sfs->count > 0, ("Sendfile sync botchup count == 0"));
200 		if (--sfs->count == 0)
201 			cv_signal(&sfs->cv);
202 		mtx_unlock(&sfs->mtx);
203 	}
204 }
205 
206 /*
207  * Helper function to calculate how much data to put into page i of n.
208  * Only first and last pages are special.
209  */
210 static inline off_t
211 xfsize(int i, int n, off_t off, off_t len)
212 {
213 
214 	if (i == 0)
215 		return (omin(PAGE_SIZE - (off & PAGE_MASK), len));
216 
217 	if (i == n - 1 && ((off + len) & PAGE_MASK) > 0)
218 		return ((off + len) & PAGE_MASK);
219 
220 	return (PAGE_SIZE);
221 }
222 
223 /*
224  * Helper function to get offset within object for i page.
225  */
226 static inline vm_ooffset_t
227 vmoff(int i, off_t off)
228 {
229 
230 	if (i == 0)
231 		return ((vm_ooffset_t)off);
232 
233 	return (trunc_page(off + i * PAGE_SIZE));
234 }
235 
236 /*
237  * Helper function used when allocation of a page or sf_buf failed.
238  * Pretend as if we don't have enough space, subtract xfsize() of
239  * all pages that failed.
240  */
241 static inline void
242 fixspace(int old, int new, off_t off, int *space)
243 {
244 
245 	KASSERT(old > new, ("%s: old %d new %d", __func__, old, new));
246 
247 	/* Subtract last one. */
248 	*space -= xfsize(old - 1, old, off, *space);
249 	old--;
250 
251 	if (new == old)
252 		/* There was only one page. */
253 		return;
254 
255 	/* Subtract first one. */
256 	if (new == 0) {
257 		*space -= xfsize(0, old, off, *space);
258 		new++;
259 	}
260 
261 	/* Rest of pages are full sized. */
262 	*space -= (old - new) * PAGE_SIZE;
263 
264 	KASSERT(*space >= 0, ("%s: space went backwards", __func__));
265 }
266 
267 /*
268  * I/O completion callback.
269  */
270 static void
271 sendfile_iodone(void *arg, vm_page_t *pg, int count, int error)
272 {
273 	struct sf_io *sfio = arg;
274 	struct socket *so = sfio->so;
275 
276 	for (int i = 0; i < count; i++)
277 		if (pg[i] != bogus_page)
278 			vm_page_xunbusy(pg[i]);
279 
280 	if (error)
281 		sfio->error = error;
282 
283 	if (!refcount_release(&sfio->nios))
284 		return;
285 
286 	CURVNET_SET(so->so_vnet);
287 	if (sfio->error) {
288 		struct mbuf *m;
289 
290 		/*
291 		 * I/O operation failed.  The state of data in the socket
292 		 * is now inconsistent, and all what we can do is to tear
293 		 * it down. Protocol abort method would tear down protocol
294 		 * state, free all ready mbufs and detach not ready ones.
295 		 * We will free the mbufs corresponding to this I/O manually.
296 		 *
297 		 * The socket would be marked with EIO and made available
298 		 * for read, so that application receives EIO on next
299 		 * syscall and eventually closes the socket.
300 		 */
301 		so->so_proto->pr_usrreqs->pru_abort(so);
302 		so->so_error = EIO;
303 
304 		m = sfio->m;
305 		for (int i = 0; i < sfio->npages; i++)
306 			m = m_free(m);
307 	} else
308 		(void )(so->so_proto->pr_usrreqs->pru_ready)(so, sfio->m,
309 		    sfio->npages);
310 
311 	SOCK_LOCK(so);
312 	sorele(so);
313 	CURVNET_RESTORE();
314 	free(sfio, M_TEMP);
315 }
316 
317 /*
318  * Iterate through pages vector and request paging for non-valid pages.
319  */
320 static int
321 sendfile_swapin(vm_object_t obj, struct sf_io *sfio, off_t off, off_t len,
322     int npages, int rhpages, int flags)
323 {
324 	vm_page_t *pa = sfio->pa;
325 	int grabbed, nios;
326 
327 	nios = 0;
328 	flags = (flags & SF_NODISKIO) ? VM_ALLOC_NOWAIT : 0;
329 
330 	/*
331 	 * First grab all the pages and wire them.  Note that we grab
332 	 * only required pages.  Readahead pages are dealt with later.
333 	 */
334 	VM_OBJECT_WLOCK(obj);
335 
336 	grabbed = vm_page_grab_pages(obj, OFF_TO_IDX(off),
337 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED | flags, pa, npages);
338 	if (grabbed < npages) {
339 		for (int i = grabbed; i < npages; i++)
340 			pa[i] = NULL;
341 		npages = grabbed;
342 		rhpages = 0;
343 	}
344 
345 	for (int i = 0; i < npages;) {
346 		int j, a, count, rv;
347 
348 		/* Skip valid pages. */
349 		if (vm_page_is_valid(pa[i], vmoff(i, off) & PAGE_MASK,
350 		    xfsize(i, npages, off, len))) {
351 			vm_page_xunbusy(pa[i]);
352 			SFSTAT_INC(sf_pages_valid);
353 			i++;
354 			continue;
355 		}
356 
357 		/*
358 		 * Next page is invalid.  Check if it belongs to pager.  It
359 		 * may not be there, which is a regular situation for shmem
360 		 * pager.  For vnode pager this happens only in case of
361 		 * a sparse file.
362 		 *
363 		 * Important feature of vm_pager_has_page() is the hint
364 		 * stored in 'a', about how many pages we can pagein after
365 		 * this page in a single I/O.
366 		 */
367 		if (!vm_pager_has_page(obj, OFF_TO_IDX(vmoff(i, off)), NULL,
368 		    &a)) {
369 			pmap_zero_page(pa[i]);
370 			pa[i]->valid = VM_PAGE_BITS_ALL;
371 			MPASS(pa[i]->dirty == 0);
372 			vm_page_xunbusy(pa[i]);
373 			i++;
374 			continue;
375 		}
376 
377 		/*
378 		 * We want to pagein as many pages as possible, limited only
379 		 * by the 'a' hint and actual request.
380 		 */
381 		count = min(a + 1, npages - i);
382 
383 		/*
384 		 * We should not pagein into a valid page, thus we first trim
385 		 * any valid pages off the end of request, and substitute
386 		 * to bogus_page those, that are in the middle.
387 		 */
388 		for (j = i + count - 1; j > i; j--) {
389 			if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
390 			    xfsize(j, npages, off, len))) {
391 				count--;
392 				rhpages = 0;
393 			} else
394 				break;
395 		}
396 		for (j = i + 1; j < i + count - 1; j++)
397 			if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
398 			    xfsize(j, npages, off, len))) {
399 				vm_page_xunbusy(pa[j]);
400 				SFSTAT_INC(sf_pages_valid);
401 				SFSTAT_INC(sf_pages_bogus);
402 				pa[j] = bogus_page;
403 			}
404 
405 		refcount_acquire(&sfio->nios);
406 		rv = vm_pager_get_pages_async(obj, pa + i, count, NULL,
407 		    i + count == npages ? &rhpages : NULL,
408 		    &sendfile_iodone, sfio);
409 		KASSERT(rv == VM_PAGER_OK, ("%s: pager fail obj %p page %p",
410 		    __func__, obj, pa[i]));
411 
412 		SFSTAT_INC(sf_iocnt);
413 		SFSTAT_ADD(sf_pages_read, count);
414 		if (i + count == npages)
415 			SFSTAT_ADD(sf_rhpages_read, rhpages);
416 
417 		/*
418 		 * Restore the valid page pointers.  They are already
419 		 * unbusied, but still wired.
420 		 */
421 		for (j = i; j < i + count; j++)
422 			if (pa[j] == bogus_page) {
423 				pa[j] = vm_page_lookup(obj,
424 				    OFF_TO_IDX(vmoff(j, off)));
425 				KASSERT(pa[j], ("%s: page %p[%d] disappeared",
426 				    __func__, pa, j));
427 
428 			}
429 		i += count;
430 		nios++;
431 	}
432 
433 	VM_OBJECT_WUNLOCK(obj);
434 
435 	if (nios == 0 && npages != 0)
436 		SFSTAT_INC(sf_noiocnt);
437 
438 	return (nios);
439 }
440 
441 static int
442 sendfile_getobj(struct thread *td, struct file *fp, vm_object_t *obj_res,
443     struct vnode **vp_res, struct shmfd **shmfd_res, off_t *obj_size,
444     int *bsize)
445 {
446 	struct vattr va;
447 	vm_object_t obj;
448 	struct vnode *vp;
449 	struct shmfd *shmfd;
450 	int error;
451 
452 	vp = *vp_res = NULL;
453 	obj = NULL;
454 	shmfd = *shmfd_res = NULL;
455 	*bsize = 0;
456 
457 	/*
458 	 * The file descriptor must be a regular file and have a
459 	 * backing VM object.
460 	 */
461 	if (fp->f_type == DTYPE_VNODE) {
462 		vp = fp->f_vnode;
463 		vn_lock(vp, LK_SHARED | LK_RETRY);
464 		if (vp->v_type != VREG) {
465 			error = EINVAL;
466 			goto out;
467 		}
468 		*bsize = vp->v_mount->mnt_stat.f_iosize;
469 		error = VOP_GETATTR(vp, &va, td->td_ucred);
470 		if (error != 0)
471 			goto out;
472 		*obj_size = va.va_size;
473 		obj = vp->v_object;
474 		if (obj == NULL) {
475 			error = EINVAL;
476 			goto out;
477 		}
478 	} else if (fp->f_type == DTYPE_SHM) {
479 		error = 0;
480 		shmfd = fp->f_data;
481 		obj = shmfd->shm_object;
482 		*obj_size = shmfd->shm_size;
483 	} else {
484 		error = EINVAL;
485 		goto out;
486 	}
487 
488 	VM_OBJECT_WLOCK(obj);
489 	if ((obj->flags & OBJ_DEAD) != 0) {
490 		VM_OBJECT_WUNLOCK(obj);
491 		error = EBADF;
492 		goto out;
493 	}
494 
495 	/*
496 	 * Temporarily increase the backing VM object's reference
497 	 * count so that a forced reclamation of its vnode does not
498 	 * immediately destroy it.
499 	 */
500 	vm_object_reference_locked(obj);
501 	VM_OBJECT_WUNLOCK(obj);
502 	*obj_res = obj;
503 	*vp_res = vp;
504 	*shmfd_res = shmfd;
505 
506 out:
507 	if (vp != NULL)
508 		VOP_UNLOCK(vp, 0);
509 	return (error);
510 }
511 
512 static int
513 sendfile_getsock(struct thread *td, int s, struct file **sock_fp,
514     struct socket **so)
515 {
516 	cap_rights_t rights;
517 	int error;
518 
519 	*sock_fp = NULL;
520 	*so = NULL;
521 
522 	/*
523 	 * The socket must be a stream socket and connected.
524 	 */
525 	error = getsock_cap(td, s, cap_rights_init(&rights, CAP_SEND),
526 	    sock_fp, NULL, NULL);
527 	if (error != 0)
528 		return (error);
529 	*so = (*sock_fp)->f_data;
530 	if ((*so)->so_type != SOCK_STREAM)
531 		return (EINVAL);
532 	return (0);
533 }
534 
535 int
536 vn_sendfile(struct file *fp, int sockfd, struct uio *hdr_uio,
537     struct uio *trl_uio, off_t offset, size_t nbytes, off_t *sent, int flags,
538     struct thread *td)
539 {
540 	struct file *sock_fp;
541 	struct vnode *vp;
542 	struct vm_object *obj;
543 	struct socket *so;
544 	struct mbuf *m, *mh, *mhtail;
545 	struct sf_buf *sf;
546 	struct shmfd *shmfd;
547 	struct sendfile_sync *sfs;
548 	struct vattr va;
549 	off_t off, sbytes, rem, obj_size;
550 	int error, softerr, bsize, hdrlen;
551 
552 	obj = NULL;
553 	so = NULL;
554 	m = mh = NULL;
555 	sfs = NULL;
556 	hdrlen = sbytes = 0;
557 	softerr = 0;
558 
559 	error = sendfile_getobj(td, fp, &obj, &vp, &shmfd, &obj_size, &bsize);
560 	if (error != 0)
561 		return (error);
562 
563 	error = sendfile_getsock(td, sockfd, &sock_fp, &so);
564 	if (error != 0)
565 		goto out;
566 
567 #ifdef MAC
568 	error = mac_socket_check_send(td->td_ucred, so);
569 	if (error != 0)
570 		goto out;
571 #endif
572 
573 	SFSTAT_INC(sf_syscalls);
574 	SFSTAT_ADD(sf_rhpages_requested, SF_READAHEAD(flags));
575 
576 	if (flags & SF_SYNC) {
577 		sfs = malloc(sizeof *sfs, M_TEMP, M_WAITOK | M_ZERO);
578 		mtx_init(&sfs->mtx, "sendfile", NULL, MTX_DEF);
579 		cv_init(&sfs->cv, "sendfile");
580 	}
581 
582 	rem = nbytes ? omin(nbytes, obj_size - offset) : obj_size - offset;
583 
584 	/*
585 	 * Protect against multiple writers to the socket.
586 	 *
587 	 * XXXRW: Historically this has assumed non-interruptibility, so now
588 	 * we implement that, but possibly shouldn't.
589 	 */
590 	(void)sblock(&so->so_snd, SBL_WAIT | SBL_NOINTR);
591 
592 	/*
593 	 * Loop through the pages of the file, starting with the requested
594 	 * offset. Get a file page (do I/O if necessary), map the file page
595 	 * into an sf_buf, attach an mbuf header to the sf_buf, and queue
596 	 * it on the socket.
597 	 * This is done in two loops.  The inner loop turns as many pages
598 	 * as it can, up to available socket buffer space, without blocking
599 	 * into mbufs to have it bulk delivered into the socket send buffer.
600 	 * The outer loop checks the state and available space of the socket
601 	 * and takes care of the overall progress.
602 	 */
603 	for (off = offset; rem > 0; ) {
604 		struct sf_io *sfio;
605 		vm_page_t *pa;
606 		struct mbuf *mtail;
607 		int nios, space, npages, rhpages;
608 
609 		mtail = NULL;
610 		/*
611 		 * Check the socket state for ongoing connection,
612 		 * no errors and space in socket buffer.
613 		 * If space is low allow for the remainder of the
614 		 * file to be processed if it fits the socket buffer.
615 		 * Otherwise block in waiting for sufficient space
616 		 * to proceed, or if the socket is nonblocking, return
617 		 * to userland with EAGAIN while reporting how far
618 		 * we've come.
619 		 * We wait until the socket buffer has significant free
620 		 * space to do bulk sends.  This makes good use of file
621 		 * system read ahead and allows packet segmentation
622 		 * offloading hardware to take over lots of work.  If
623 		 * we were not careful here we would send off only one
624 		 * sfbuf at a time.
625 		 */
626 		SOCKBUF_LOCK(&so->so_snd);
627 		if (so->so_snd.sb_lowat < so->so_snd.sb_hiwat / 2)
628 			so->so_snd.sb_lowat = so->so_snd.sb_hiwat / 2;
629 retry_space:
630 		if (so->so_snd.sb_state & SBS_CANTSENDMORE) {
631 			error = EPIPE;
632 			SOCKBUF_UNLOCK(&so->so_snd);
633 			goto done;
634 		} else if (so->so_error) {
635 			error = so->so_error;
636 			so->so_error = 0;
637 			SOCKBUF_UNLOCK(&so->so_snd);
638 			goto done;
639 		}
640 		if ((so->so_state & SS_ISCONNECTED) == 0) {
641 			SOCKBUF_UNLOCK(&so->so_snd);
642 			error = ENOTCONN;
643 			goto done;
644 		}
645 
646 		space = sbspace(&so->so_snd);
647 		if (space < rem &&
648 		    (space <= 0 ||
649 		     space < so->so_snd.sb_lowat)) {
650 			if (so->so_state & SS_NBIO) {
651 				SOCKBUF_UNLOCK(&so->so_snd);
652 				error = EAGAIN;
653 				goto done;
654 			}
655 			/*
656 			 * sbwait drops the lock while sleeping.
657 			 * When we loop back to retry_space the
658 			 * state may have changed and we retest
659 			 * for it.
660 			 */
661 			error = sbwait(&so->so_snd);
662 			/*
663 			 * An error from sbwait usually indicates that we've
664 			 * been interrupted by a signal. If we've sent anything
665 			 * then return bytes sent, otherwise return the error.
666 			 */
667 			if (error != 0) {
668 				SOCKBUF_UNLOCK(&so->so_snd);
669 				goto done;
670 			}
671 			goto retry_space;
672 		}
673 		SOCKBUF_UNLOCK(&so->so_snd);
674 
675 		/*
676 		 * At the beginning of the first loop check if any headers
677 		 * are specified and copy them into mbufs.  Reduce space in
678 		 * the socket buffer by the size of the header mbuf chain.
679 		 * Clear hdr_uio here and hdrlen at the end of the first loop.
680 		 */
681 		if (hdr_uio != NULL && hdr_uio->uio_resid > 0) {
682 			hdr_uio->uio_td = td;
683 			hdr_uio->uio_rw = UIO_WRITE;
684 			mh = m_uiotombuf(hdr_uio, M_WAITOK, space, 0, 0);
685 			hdrlen = m_length(mh, &mhtail);
686 			space -= hdrlen;
687 			/*
688 			 * If header consumed all the socket buffer space,
689 			 * don't waste CPU cycles and jump to the end.
690 			 */
691 			if (space == 0) {
692 				sfio = NULL;
693 				nios = 0;
694 				goto prepend_header;
695 			}
696 			hdr_uio = NULL;
697 		}
698 
699 		if (vp != NULL) {
700 			error = vn_lock(vp, LK_SHARED);
701 			if (error != 0)
702 				goto done;
703 			error = VOP_GETATTR(vp, &va, td->td_ucred);
704 			if (error != 0 || off >= va.va_size) {
705 				VOP_UNLOCK(vp, 0);
706 				goto done;
707 			}
708 			if (va.va_size != obj_size) {
709 				obj_size = va.va_size;
710 				rem = nbytes ?
711 				    omin(nbytes + offset, obj_size) : obj_size;
712 				rem -= off;
713 			}
714 		}
715 
716 		if (space > rem)
717 			space = rem;
718 
719 		npages = howmany(space + (off & PAGE_MASK), PAGE_SIZE);
720 
721 		/*
722 		 * Calculate maximum allowed number of pages for readahead
723 		 * at this iteration.  If SF_USER_READAHEAD was set, we don't
724 		 * do any heuristics and use exactly the value supplied by
725 		 * application.  Otherwise, we allow readahead up to "rem".
726 		 * If application wants more, let it be, but there is no
727 		 * reason to go above MAXPHYS.  Also check against "obj_size",
728 		 * since vm_pager_has_page() can hint beyond EOF.
729 		 */
730 		if (flags & SF_USER_READAHEAD) {
731 			rhpages = SF_READAHEAD(flags);
732 		} else {
733 			rhpages = howmany(rem + (off & PAGE_MASK), PAGE_SIZE) -
734 			    npages;
735 			rhpages += SF_READAHEAD(flags);
736 		}
737 		rhpages = min(howmany(MAXPHYS, PAGE_SIZE), rhpages);
738 		rhpages = min(howmany(obj_size - trunc_page(off), PAGE_SIZE) -
739 		    npages, rhpages);
740 
741 		sfio = malloc(sizeof(struct sf_io) +
742 		    npages * sizeof(vm_page_t), M_TEMP, M_WAITOK);
743 		refcount_init(&sfio->nios, 1);
744 		sfio->so = so;
745 		sfio->error = 0;
746 
747 		nios = sendfile_swapin(obj, sfio, off, space, npages, rhpages,
748 		    flags);
749 
750 		/*
751 		 * Loop and construct maximum sized mbuf chain to be bulk
752 		 * dumped into socket buffer.
753 		 */
754 		pa = sfio->pa;
755 		for (int i = 0; i < npages; i++) {
756 			struct mbuf *m0;
757 
758 			/*
759 			 * If a page wasn't grabbed successfully, then
760 			 * trim the array. Can happen only with SF_NODISKIO.
761 			 */
762 			if (pa[i] == NULL) {
763 				SFSTAT_INC(sf_busy);
764 				fixspace(npages, i, off, &space);
765 				npages = i;
766 				softerr = EBUSY;
767 				break;
768 			}
769 
770 			/*
771 			 * Get a sendfile buf.  When allocating the
772 			 * first buffer for mbuf chain, we usually
773 			 * wait as long as necessary, but this wait
774 			 * can be interrupted.  For consequent
775 			 * buffers, do not sleep, since several
776 			 * threads might exhaust the buffers and then
777 			 * deadlock.
778 			 */
779 			sf = sf_buf_alloc(pa[i],
780 			    m != NULL ? SFB_NOWAIT : SFB_CATCH);
781 			if (sf == NULL) {
782 				SFSTAT_INC(sf_allocfail);
783 				for (int j = i; j < npages; j++) {
784 					vm_page_lock(pa[j]);
785 					vm_page_unwire(pa[j], PQ_INACTIVE);
786 					vm_page_unlock(pa[j]);
787 				}
788 				if (m == NULL)
789 					softerr = ENOBUFS;
790 				fixspace(npages, i, off, &space);
791 				npages = i;
792 				break;
793 			}
794 
795 			m0 = m_get(M_WAITOK, MT_DATA);
796 			m0->m_ext.ext_buf = (char *)sf_buf_kva(sf);
797 			m0->m_ext.ext_size = PAGE_SIZE;
798 			m0->m_ext.ext_arg1 = sf;
799 			m0->m_ext.ext_type = EXT_SFBUF;
800 			m0->m_ext.ext_flags = EXT_FLAG_EMBREF;
801 			m0->m_ext.ext_free = sendfile_free_mext;
802 			/*
803 			 * SF_NOCACHE sets the page as being freed upon send.
804 			 * However, we ignore it for the last page in 'space',
805 			 * if the page is truncated, and we got more data to
806 			 * send (rem > space), or if we have readahead
807 			 * configured (rhpages > 0).
808 			 */
809 			if ((flags & SF_NOCACHE) &&
810 			    (i != npages - 1 ||
811 			    !((off + space) & PAGE_MASK) ||
812 			    !(rem > space || rhpages > 0)))
813 				m0->m_ext.ext_flags |= EXT_FLAG_NOCACHE;
814 			if (sfs != NULL) {
815 				m0->m_ext.ext_flags |= EXT_FLAG_SYNC;
816 				m0->m_ext.ext_arg2 = sfs;
817 				mtx_lock(&sfs->mtx);
818 				sfs->count++;
819 				mtx_unlock(&sfs->mtx);
820 			}
821 			m0->m_ext.ext_count = 1;
822 			m0->m_flags |= (M_EXT | M_RDONLY);
823 			if (nios)
824 				m0->m_flags |= M_NOTREADY;
825 			m0->m_data = (char *)sf_buf_kva(sf) +
826 			    (vmoff(i, off) & PAGE_MASK);
827 			m0->m_len = xfsize(i, npages, off, space);
828 
829 			if (i == 0)
830 				sfio->m = m0;
831 
832 			/* Append to mbuf chain. */
833 			if (mtail != NULL)
834 				mtail->m_next = m0;
835 			else
836 				m = m0;
837 			mtail = m0;
838 		}
839 
840 		if (vp != NULL)
841 			VOP_UNLOCK(vp, 0);
842 
843 		/* Keep track of bytes processed. */
844 		off += space;
845 		rem -= space;
846 
847 		/* Prepend header, if any. */
848 		if (hdrlen) {
849 prepend_header:
850 			mhtail->m_next = m;
851 			m = mh;
852 			mh = NULL;
853 		}
854 
855 		if (m == NULL) {
856 			KASSERT(softerr, ("%s: m NULL, no error", __func__));
857 			error = softerr;
858 			free(sfio, M_TEMP);
859 			goto done;
860 		}
861 
862 		/* Add the buffer chain to the socket buffer. */
863 		KASSERT(m_length(m, NULL) == space + hdrlen,
864 		    ("%s: mlen %u space %d hdrlen %d",
865 		    __func__, m_length(m, NULL), space, hdrlen));
866 
867 		CURVNET_SET(so->so_vnet);
868 		if (nios == 0) {
869 			/*
870 			 * If sendfile_swapin() didn't initiate any I/Os,
871 			 * which happens if all data is cached in VM, then
872 			 * we can send data right now without the
873 			 * PRUS_NOTREADY flag.
874 			 */
875 			free(sfio, M_TEMP);
876 			error = (*so->so_proto->pr_usrreqs->pru_send)
877 			    (so, 0, m, NULL, NULL, td);
878 		} else {
879 			sfio->npages = npages;
880 			soref(so);
881 			error = (*so->so_proto->pr_usrreqs->pru_send)
882 			    (so, PRUS_NOTREADY, m, NULL, NULL, td);
883 			sendfile_iodone(sfio, NULL, 0, 0);
884 		}
885 		CURVNET_RESTORE();
886 
887 		m = NULL;	/* pru_send always consumes */
888 		if (error)
889 			goto done;
890 		sbytes += space + hdrlen;
891 		if (hdrlen)
892 			hdrlen = 0;
893 		if (softerr) {
894 			error = softerr;
895 			goto done;
896 		}
897 	}
898 
899 	/*
900 	 * Send trailers. Wimp out and use writev(2).
901 	 */
902 	if (trl_uio != NULL) {
903 		sbunlock(&so->so_snd);
904 		error = kern_writev(td, sockfd, trl_uio);
905 		if (error == 0)
906 			sbytes += td->td_retval[0];
907 		goto out;
908 	}
909 
910 done:
911 	sbunlock(&so->so_snd);
912 out:
913 	/*
914 	 * If there was no error we have to clear td->td_retval[0]
915 	 * because it may have been set by writev.
916 	 */
917 	if (error == 0) {
918 		td->td_retval[0] = 0;
919 	}
920 	if (sent != NULL) {
921 		(*sent) = sbytes;
922 	}
923 	if (obj != NULL)
924 		vm_object_deallocate(obj);
925 	if (so)
926 		fdrop(sock_fp, td);
927 	if (m)
928 		m_freem(m);
929 	if (mh)
930 		m_freem(mh);
931 
932 	if (sfs != NULL) {
933 		mtx_lock(&sfs->mtx);
934 		if (sfs->count != 0)
935 			cv_wait(&sfs->cv, &sfs->mtx);
936 		KASSERT(sfs->count == 0, ("sendfile sync still busy"));
937 		cv_destroy(&sfs->cv);
938 		mtx_destroy(&sfs->mtx);
939 		free(sfs, M_TEMP);
940 	}
941 
942 	if (error == ERESTART)
943 		error = EINTR;
944 
945 	return (error);
946 }
947 
948 static int
949 sendfile(struct thread *td, struct sendfile_args *uap, int compat)
950 {
951 	struct sf_hdtr hdtr;
952 	struct uio *hdr_uio, *trl_uio;
953 	struct file *fp;
954 	cap_rights_t rights;
955 	off_t sbytes;
956 	int error;
957 
958 	/*
959 	 * File offset must be positive.  If it goes beyond EOF
960 	 * we send only the header/trailer and no payload data.
961 	 */
962 	if (uap->offset < 0)
963 		return (EINVAL);
964 
965 	sbytes = 0;
966 	hdr_uio = trl_uio = NULL;
967 
968 	if (uap->hdtr != NULL) {
969 		error = copyin(uap->hdtr, &hdtr, sizeof(hdtr));
970 		if (error != 0)
971 			goto out;
972 		if (hdtr.headers != NULL) {
973 			error = copyinuio(hdtr.headers, hdtr.hdr_cnt,
974 			    &hdr_uio);
975 			if (error != 0)
976 				goto out;
977 #ifdef COMPAT_FREEBSD4
978 			/*
979 			 * In FreeBSD < 5.0 the nbytes to send also included
980 			 * the header.  If compat is specified subtract the
981 			 * header size from nbytes.
982 			 */
983 			if (compat) {
984 				if (uap->nbytes > hdr_uio->uio_resid)
985 					uap->nbytes -= hdr_uio->uio_resid;
986 				else
987 					uap->nbytes = 0;
988 			}
989 #endif
990 		}
991 		if (hdtr.trailers != NULL) {
992 			error = copyinuio(hdtr.trailers, hdtr.trl_cnt,
993 			    &trl_uio);
994 			if (error != 0)
995 				goto out;
996 		}
997 	}
998 
999 	AUDIT_ARG_FD(uap->fd);
1000 
1001 	/*
1002 	 * sendfile(2) can start at any offset within a file so we require
1003 	 * CAP_READ+CAP_SEEK = CAP_PREAD.
1004 	 */
1005 	if ((error = fget_read(td, uap->fd,
1006 	    cap_rights_init(&rights, CAP_PREAD), &fp)) != 0) {
1007 		goto out;
1008 	}
1009 
1010 	error = fo_sendfile(fp, uap->s, hdr_uio, trl_uio, uap->offset,
1011 	    uap->nbytes, &sbytes, uap->flags, td);
1012 	fdrop(fp, td);
1013 
1014 	if (uap->sbytes != NULL)
1015 		copyout(&sbytes, uap->sbytes, sizeof(off_t));
1016 
1017 out:
1018 	free(hdr_uio, M_IOV);
1019 	free(trl_uio, M_IOV);
1020 	return (error);
1021 }
1022 
1023 /*
1024  * sendfile(2)
1025  *
1026  * int sendfile(int fd, int s, off_t offset, size_t nbytes,
1027  *       struct sf_hdtr *hdtr, off_t *sbytes, int flags)
1028  *
1029  * Send a file specified by 'fd' and starting at 'offset' to a socket
1030  * specified by 's'. Send only 'nbytes' of the file or until EOF if nbytes ==
1031  * 0.  Optionally add a header and/or trailer to the socket output.  If
1032  * specified, write the total number of bytes sent into *sbytes.
1033  */
1034 int
1035 sys_sendfile(struct thread *td, struct sendfile_args *uap)
1036 {
1037 
1038 	return (sendfile(td, uap, 0));
1039 }
1040 
1041 #ifdef COMPAT_FREEBSD4
1042 int
1043 freebsd4_sendfile(struct thread *td, struct freebsd4_sendfile_args *uap)
1044 {
1045 	struct sendfile_args args;
1046 
1047 	args.fd = uap->fd;
1048 	args.s = uap->s;
1049 	args.offset = uap->offset;
1050 	args.nbytes = uap->nbytes;
1051 	args.hdtr = uap->hdtr;
1052 	args.sbytes = uap->sbytes;
1053 	args.flags = uap->flags;
1054 
1055 	return (sendfile(td, &args, 1));
1056 }
1057 #endif /* COMPAT_FREEBSD4 */
1058