xref: /freebsd/sys/kern/kern_sendfile.c (revision 5bf5ca772c6de2d53344a78cf461447cc322ccea)
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 if (pg->queue != PQ_INACTIVE)
171 					vm_page_deactivate(pg);
172 				else
173 					vm_page_requeue(pg);
174 			}
175 		}
176 	}
177 	vm_page_unlock(pg);
178 }
179 
180 static void
181 sendfile_free_mext(struct mbuf *m)
182 {
183 	struct sf_buf *sf;
184 	vm_page_t pg;
185 	bool nocache;
186 
187 	KASSERT(m->m_flags & M_EXT && m->m_ext.ext_type == EXT_SFBUF,
188 	    ("%s: m %p !M_EXT or !EXT_SFBUF", __func__, m));
189 
190 	sf = m->m_ext.ext_arg1;
191 	pg = sf_buf_page(sf);
192 	nocache = m->m_ext.ext_flags & EXT_FLAG_NOCACHE;
193 
194 	sf_buf_free(sf);
195 	sendfile_free_page(pg, nocache);
196 
197 	if (m->m_ext.ext_flags & EXT_FLAG_SYNC) {
198 		struct sendfile_sync *sfs = m->m_ext.ext_arg2;
199 
200 		mtx_lock(&sfs->mtx);
201 		KASSERT(sfs->count > 0, ("Sendfile sync botchup count == 0"));
202 		if (--sfs->count == 0)
203 			cv_signal(&sfs->cv);
204 		mtx_unlock(&sfs->mtx);
205 	}
206 }
207 
208 /*
209  * Helper function to calculate how much data to put into page i of n.
210  * Only first and last pages are special.
211  */
212 static inline off_t
213 xfsize(int i, int n, off_t off, off_t len)
214 {
215 
216 	if (i == 0)
217 		return (omin(PAGE_SIZE - (off & PAGE_MASK), len));
218 
219 	if (i == n - 1 && ((off + len) & PAGE_MASK) > 0)
220 		return ((off + len) & PAGE_MASK);
221 
222 	return (PAGE_SIZE);
223 }
224 
225 /*
226  * Helper function to get offset within object for i page.
227  */
228 static inline vm_ooffset_t
229 vmoff(int i, off_t off)
230 {
231 
232 	if (i == 0)
233 		return ((vm_ooffset_t)off);
234 
235 	return (trunc_page(off + i * PAGE_SIZE));
236 }
237 
238 /*
239  * Helper function used when allocation of a page or sf_buf failed.
240  * Pretend as if we don't have enough space, subtract xfsize() of
241  * all pages that failed.
242  */
243 static inline void
244 fixspace(int old, int new, off_t off, int *space)
245 {
246 
247 	KASSERT(old > new, ("%s: old %d new %d", __func__, old, new));
248 
249 	/* Subtract last one. */
250 	*space -= xfsize(old - 1, old, off, *space);
251 	old--;
252 
253 	if (new == old)
254 		/* There was only one page. */
255 		return;
256 
257 	/* Subtract first one. */
258 	if (new == 0) {
259 		*space -= xfsize(0, old, off, *space);
260 		new++;
261 	}
262 
263 	/* Rest of pages are full sized. */
264 	*space -= (old - new) * PAGE_SIZE;
265 
266 	KASSERT(*space >= 0, ("%s: space went backwards", __func__));
267 }
268 
269 /*
270  * I/O completion callback.
271  */
272 static void
273 sendfile_iodone(void *arg, vm_page_t *pg, int count, int error)
274 {
275 	struct sf_io *sfio = arg;
276 	struct socket *so = sfio->so;
277 
278 	for (int i = 0; i < count; i++)
279 		if (pg[i] != bogus_page)
280 			vm_page_xunbusy(pg[i]);
281 
282 	if (error)
283 		sfio->error = error;
284 
285 	if (!refcount_release(&sfio->nios))
286 		return;
287 
288 	CURVNET_SET(so->so_vnet);
289 	if (sfio->error) {
290 		struct mbuf *m;
291 
292 		/*
293 		 * I/O operation failed.  The state of data in the socket
294 		 * is now inconsistent, and all what we can do is to tear
295 		 * it down. Protocol abort method would tear down protocol
296 		 * state, free all ready mbufs and detach not ready ones.
297 		 * We will free the mbufs corresponding to this I/O manually.
298 		 *
299 		 * The socket would be marked with EIO and made available
300 		 * for read, so that application receives EIO on next
301 		 * syscall and eventually closes the socket.
302 		 */
303 		so->so_proto->pr_usrreqs->pru_abort(so);
304 		so->so_error = EIO;
305 
306 		m = sfio->m;
307 		for (int i = 0; i < sfio->npages; i++)
308 			m = m_free(m);
309 	} else
310 		(void )(so->so_proto->pr_usrreqs->pru_ready)(so, sfio->m,
311 		    sfio->npages);
312 
313 	SOCK_LOCK(so);
314 	sorele(so);
315 	CURVNET_RESTORE();
316 	free(sfio, M_TEMP);
317 }
318 
319 /*
320  * Iterate through pages vector and request paging for non-valid pages.
321  */
322 static int
323 sendfile_swapin(vm_object_t obj, struct sf_io *sfio, off_t off, off_t len,
324     int npages, int rhpages, int flags)
325 {
326 	vm_page_t *pa = sfio->pa;
327 	int grabbed, nios;
328 
329 	nios = 0;
330 	flags = (flags & SF_NODISKIO) ? VM_ALLOC_NOWAIT : 0;
331 
332 	/*
333 	 * First grab all the pages and wire them.  Note that we grab
334 	 * only required pages.  Readahead pages are dealt with later.
335 	 */
336 	VM_OBJECT_WLOCK(obj);
337 
338 	grabbed = vm_page_grab_pages(obj, OFF_TO_IDX(off),
339 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED | flags, pa, npages);
340 	if (grabbed < npages) {
341 		for (int i = grabbed; i < npages; i++)
342 			pa[i] = NULL;
343 		npages = grabbed;
344 		rhpages = 0;
345 	}
346 
347 	for (int i = 0; i < npages;) {
348 		int j, a, count, rv;
349 
350 		/* Skip valid pages. */
351 		if (vm_page_is_valid(pa[i], vmoff(i, off) & PAGE_MASK,
352 		    xfsize(i, npages, off, len))) {
353 			vm_page_xunbusy(pa[i]);
354 			SFSTAT_INC(sf_pages_valid);
355 			i++;
356 			continue;
357 		}
358 
359 		/*
360 		 * Next page is invalid.  Check if it belongs to pager.  It
361 		 * may not be there, which is a regular situation for shmem
362 		 * pager.  For vnode pager this happens only in case of
363 		 * a sparse file.
364 		 *
365 		 * Important feature of vm_pager_has_page() is the hint
366 		 * stored in 'a', about how many pages we can pagein after
367 		 * this page in a single I/O.
368 		 */
369 		if (!vm_pager_has_page(obj, OFF_TO_IDX(vmoff(i, off)), NULL,
370 		    &a)) {
371 			pmap_zero_page(pa[i]);
372 			pa[i]->valid = VM_PAGE_BITS_ALL;
373 			MPASS(pa[i]->dirty == 0);
374 			vm_page_xunbusy(pa[i]);
375 			i++;
376 			continue;
377 		}
378 
379 		/*
380 		 * We want to pagein as many pages as possible, limited only
381 		 * by the 'a' hint and actual request.
382 		 */
383 		count = min(a + 1, npages - i);
384 
385 		/*
386 		 * We should not pagein into a valid page, thus we first trim
387 		 * any valid pages off the end of request, and substitute
388 		 * to bogus_page those, that are in the middle.
389 		 */
390 		for (j = i + count - 1; j > i; j--) {
391 			if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
392 			    xfsize(j, npages, off, len))) {
393 				count--;
394 				rhpages = 0;
395 			} else
396 				break;
397 		}
398 		for (j = i + 1; j < i + count - 1; j++)
399 			if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
400 			    xfsize(j, npages, off, len))) {
401 				vm_page_xunbusy(pa[j]);
402 				SFSTAT_INC(sf_pages_valid);
403 				SFSTAT_INC(sf_pages_bogus);
404 				pa[j] = bogus_page;
405 			}
406 
407 		refcount_acquire(&sfio->nios);
408 		rv = vm_pager_get_pages_async(obj, pa + i, count, NULL,
409 		    i + count == npages ? &rhpages : NULL,
410 		    &sendfile_iodone, sfio);
411 		KASSERT(rv == VM_PAGER_OK, ("%s: pager fail obj %p page %p",
412 		    __func__, obj, pa[i]));
413 
414 		SFSTAT_INC(sf_iocnt);
415 		SFSTAT_ADD(sf_pages_read, count);
416 		if (i + count == npages)
417 			SFSTAT_ADD(sf_rhpages_read, rhpages);
418 
419 		/*
420 		 * Restore the valid page pointers.  They are already
421 		 * unbusied, but still wired.
422 		 */
423 		for (j = i; j < i + count; j++)
424 			if (pa[j] == bogus_page) {
425 				pa[j] = vm_page_lookup(obj,
426 				    OFF_TO_IDX(vmoff(j, off)));
427 				KASSERT(pa[j], ("%s: page %p[%d] disappeared",
428 				    __func__, pa, j));
429 
430 			}
431 		i += count;
432 		nios++;
433 	}
434 
435 	VM_OBJECT_WUNLOCK(obj);
436 
437 	if (nios == 0 && npages != 0)
438 		SFSTAT_INC(sf_noiocnt);
439 
440 	return (nios);
441 }
442 
443 static int
444 sendfile_getobj(struct thread *td, struct file *fp, vm_object_t *obj_res,
445     struct vnode **vp_res, struct shmfd **shmfd_res, off_t *obj_size,
446     int *bsize)
447 {
448 	struct vattr va;
449 	vm_object_t obj;
450 	struct vnode *vp;
451 	struct shmfd *shmfd;
452 	int error;
453 
454 	vp = *vp_res = NULL;
455 	obj = NULL;
456 	shmfd = *shmfd_res = NULL;
457 	*bsize = 0;
458 
459 	/*
460 	 * The file descriptor must be a regular file and have a
461 	 * backing VM object.
462 	 */
463 	if (fp->f_type == DTYPE_VNODE) {
464 		vp = fp->f_vnode;
465 		vn_lock(vp, LK_SHARED | LK_RETRY);
466 		if (vp->v_type != VREG) {
467 			error = EINVAL;
468 			goto out;
469 		}
470 		*bsize = vp->v_mount->mnt_stat.f_iosize;
471 		error = VOP_GETATTR(vp, &va, td->td_ucred);
472 		if (error != 0)
473 			goto out;
474 		*obj_size = va.va_size;
475 		obj = vp->v_object;
476 		if (obj == NULL) {
477 			error = EINVAL;
478 			goto out;
479 		}
480 	} else if (fp->f_type == DTYPE_SHM) {
481 		error = 0;
482 		shmfd = fp->f_data;
483 		obj = shmfd->shm_object;
484 		*obj_size = shmfd->shm_size;
485 	} else {
486 		error = EINVAL;
487 		goto out;
488 	}
489 
490 	VM_OBJECT_WLOCK(obj);
491 	if ((obj->flags & OBJ_DEAD) != 0) {
492 		VM_OBJECT_WUNLOCK(obj);
493 		error = EBADF;
494 		goto out;
495 	}
496 
497 	/*
498 	 * Temporarily increase the backing VM object's reference
499 	 * count so that a forced reclamation of its vnode does not
500 	 * immediately destroy it.
501 	 */
502 	vm_object_reference_locked(obj);
503 	VM_OBJECT_WUNLOCK(obj);
504 	*obj_res = obj;
505 	*vp_res = vp;
506 	*shmfd_res = shmfd;
507 
508 out:
509 	if (vp != NULL)
510 		VOP_UNLOCK(vp, 0);
511 	return (error);
512 }
513 
514 static int
515 sendfile_getsock(struct thread *td, int s, struct file **sock_fp,
516     struct socket **so)
517 {
518 	cap_rights_t rights;
519 	int error;
520 
521 	*sock_fp = NULL;
522 	*so = NULL;
523 
524 	/*
525 	 * The socket must be a stream socket and connected.
526 	 */
527 	error = getsock_cap(td, s, cap_rights_init(&rights, CAP_SEND),
528 	    sock_fp, NULL, NULL);
529 	if (error != 0)
530 		return (error);
531 	*so = (*sock_fp)->f_data;
532 	if ((*so)->so_type != SOCK_STREAM)
533 		return (EINVAL);
534 	return (0);
535 }
536 
537 int
538 vn_sendfile(struct file *fp, int sockfd, struct uio *hdr_uio,
539     struct uio *trl_uio, off_t offset, size_t nbytes, off_t *sent, int flags,
540     struct thread *td)
541 {
542 	struct file *sock_fp;
543 	struct vnode *vp;
544 	struct vm_object *obj;
545 	struct socket *so;
546 	struct mbuf *m, *mh, *mhtail;
547 	struct sf_buf *sf;
548 	struct shmfd *shmfd;
549 	struct sendfile_sync *sfs;
550 	struct vattr va;
551 	off_t off, sbytes, rem, obj_size;
552 	int error, softerr, bsize, hdrlen;
553 
554 	obj = NULL;
555 	so = NULL;
556 	m = mh = NULL;
557 	sfs = NULL;
558 	hdrlen = sbytes = 0;
559 	softerr = 0;
560 
561 	error = sendfile_getobj(td, fp, &obj, &vp, &shmfd, &obj_size, &bsize);
562 	if (error != 0)
563 		return (error);
564 
565 	error = sendfile_getsock(td, sockfd, &sock_fp, &so);
566 	if (error != 0)
567 		goto out;
568 
569 #ifdef MAC
570 	error = mac_socket_check_send(td->td_ucred, so);
571 	if (error != 0)
572 		goto out;
573 #endif
574 
575 	SFSTAT_INC(sf_syscalls);
576 	SFSTAT_ADD(sf_rhpages_requested, SF_READAHEAD(flags));
577 
578 	if (flags & SF_SYNC) {
579 		sfs = malloc(sizeof *sfs, M_TEMP, M_WAITOK | M_ZERO);
580 		mtx_init(&sfs->mtx, "sendfile", NULL, MTX_DEF);
581 		cv_init(&sfs->cv, "sendfile");
582 	}
583 
584 	rem = nbytes ? omin(nbytes, obj_size - offset) : obj_size - offset;
585 
586 	/*
587 	 * Protect against multiple writers to the socket.
588 	 *
589 	 * XXXRW: Historically this has assumed non-interruptibility, so now
590 	 * we implement that, but possibly shouldn't.
591 	 */
592 	(void)sblock(&so->so_snd, SBL_WAIT | SBL_NOINTR);
593 
594 	/*
595 	 * Loop through the pages of the file, starting with the requested
596 	 * offset. Get a file page (do I/O if necessary), map the file page
597 	 * into an sf_buf, attach an mbuf header to the sf_buf, and queue
598 	 * it on the socket.
599 	 * This is done in two loops.  The inner loop turns as many pages
600 	 * as it can, up to available socket buffer space, without blocking
601 	 * into mbufs to have it bulk delivered into the socket send buffer.
602 	 * The outer loop checks the state and available space of the socket
603 	 * and takes care of the overall progress.
604 	 */
605 	for (off = offset; rem > 0; ) {
606 		struct sf_io *sfio;
607 		vm_page_t *pa;
608 		struct mbuf *mtail;
609 		int nios, space, npages, rhpages;
610 
611 		mtail = NULL;
612 		/*
613 		 * Check the socket state for ongoing connection,
614 		 * no errors and space in socket buffer.
615 		 * If space is low allow for the remainder of the
616 		 * file to be processed if it fits the socket buffer.
617 		 * Otherwise block in waiting for sufficient space
618 		 * to proceed, or if the socket is nonblocking, return
619 		 * to userland with EAGAIN while reporting how far
620 		 * we've come.
621 		 * We wait until the socket buffer has significant free
622 		 * space to do bulk sends.  This makes good use of file
623 		 * system read ahead and allows packet segmentation
624 		 * offloading hardware to take over lots of work.  If
625 		 * we were not careful here we would send off only one
626 		 * sfbuf at a time.
627 		 */
628 		SOCKBUF_LOCK(&so->so_snd);
629 		if (so->so_snd.sb_lowat < so->so_snd.sb_hiwat / 2)
630 			so->so_snd.sb_lowat = so->so_snd.sb_hiwat / 2;
631 retry_space:
632 		if (so->so_snd.sb_state & SBS_CANTSENDMORE) {
633 			error = EPIPE;
634 			SOCKBUF_UNLOCK(&so->so_snd);
635 			goto done;
636 		} else if (so->so_error) {
637 			error = so->so_error;
638 			so->so_error = 0;
639 			SOCKBUF_UNLOCK(&so->so_snd);
640 			goto done;
641 		}
642 		if ((so->so_state & SS_ISCONNECTED) == 0) {
643 			SOCKBUF_UNLOCK(&so->so_snd);
644 			error = ENOTCONN;
645 			goto done;
646 		}
647 
648 		space = sbspace(&so->so_snd);
649 		if (space < rem &&
650 		    (space <= 0 ||
651 		     space < so->so_snd.sb_lowat)) {
652 			if (so->so_state & SS_NBIO) {
653 				SOCKBUF_UNLOCK(&so->so_snd);
654 				error = EAGAIN;
655 				goto done;
656 			}
657 			/*
658 			 * sbwait drops the lock while sleeping.
659 			 * When we loop back to retry_space the
660 			 * state may have changed and we retest
661 			 * for it.
662 			 */
663 			error = sbwait(&so->so_snd);
664 			/*
665 			 * An error from sbwait usually indicates that we've
666 			 * been interrupted by a signal. If we've sent anything
667 			 * then return bytes sent, otherwise return the error.
668 			 */
669 			if (error != 0) {
670 				SOCKBUF_UNLOCK(&so->so_snd);
671 				goto done;
672 			}
673 			goto retry_space;
674 		}
675 		SOCKBUF_UNLOCK(&so->so_snd);
676 
677 		/*
678 		 * At the beginning of the first loop check if any headers
679 		 * are specified and copy them into mbufs.  Reduce space in
680 		 * the socket buffer by the size of the header mbuf chain.
681 		 * Clear hdr_uio here and hdrlen at the end of the first loop.
682 		 */
683 		if (hdr_uio != NULL && hdr_uio->uio_resid > 0) {
684 			hdr_uio->uio_td = td;
685 			hdr_uio->uio_rw = UIO_WRITE;
686 			mh = m_uiotombuf(hdr_uio, M_WAITOK, space, 0, 0);
687 			hdrlen = m_length(mh, &mhtail);
688 			space -= hdrlen;
689 			/*
690 			 * If header consumed all the socket buffer space,
691 			 * don't waste CPU cycles and jump to the end.
692 			 */
693 			if (space == 0) {
694 				sfio = NULL;
695 				nios = 0;
696 				goto prepend_header;
697 			}
698 			hdr_uio = NULL;
699 		}
700 
701 		if (vp != NULL) {
702 			error = vn_lock(vp, LK_SHARED);
703 			if (error != 0)
704 				goto done;
705 			error = VOP_GETATTR(vp, &va, td->td_ucred);
706 			if (error != 0 || off >= va.va_size) {
707 				VOP_UNLOCK(vp, 0);
708 				goto done;
709 			}
710 			if (va.va_size != obj_size) {
711 				obj_size = va.va_size;
712 				rem = nbytes ?
713 				    omin(nbytes + offset, obj_size) : obj_size;
714 				rem -= off;
715 			}
716 		}
717 
718 		if (space > rem)
719 			space = rem;
720 
721 		npages = howmany(space + (off & PAGE_MASK), PAGE_SIZE);
722 
723 		/*
724 		 * Calculate maximum allowed number of pages for readahead
725 		 * at this iteration.  If SF_USER_READAHEAD was set, we don't
726 		 * do any heuristics and use exactly the value supplied by
727 		 * application.  Otherwise, we allow readahead up to "rem".
728 		 * If application wants more, let it be, but there is no
729 		 * reason to go above MAXPHYS.  Also check against "obj_size",
730 		 * since vm_pager_has_page() can hint beyond EOF.
731 		 */
732 		if (flags & SF_USER_READAHEAD) {
733 			rhpages = SF_READAHEAD(flags);
734 		} else {
735 			rhpages = howmany(rem + (off & PAGE_MASK), PAGE_SIZE) -
736 			    npages;
737 			rhpages += SF_READAHEAD(flags);
738 		}
739 		rhpages = min(howmany(MAXPHYS, PAGE_SIZE), rhpages);
740 		rhpages = min(howmany(obj_size - trunc_page(off), PAGE_SIZE) -
741 		    npages, rhpages);
742 
743 		sfio = malloc(sizeof(struct sf_io) +
744 		    npages * sizeof(vm_page_t), M_TEMP, M_WAITOK);
745 		refcount_init(&sfio->nios, 1);
746 		sfio->so = so;
747 		sfio->error = 0;
748 
749 		nios = sendfile_swapin(obj, sfio, off, space, npages, rhpages,
750 		    flags);
751 
752 		/*
753 		 * Loop and construct maximum sized mbuf chain to be bulk
754 		 * dumped into socket buffer.
755 		 */
756 		pa = sfio->pa;
757 		for (int i = 0; i < npages; i++) {
758 			struct mbuf *m0;
759 
760 			/*
761 			 * If a page wasn't grabbed successfully, then
762 			 * trim the array. Can happen only with SF_NODISKIO.
763 			 */
764 			if (pa[i] == NULL) {
765 				SFSTAT_INC(sf_busy);
766 				fixspace(npages, i, off, &space);
767 				npages = i;
768 				softerr = EBUSY;
769 				break;
770 			}
771 
772 			/*
773 			 * Get a sendfile buf.  When allocating the
774 			 * first buffer for mbuf chain, we usually
775 			 * wait as long as necessary, but this wait
776 			 * can be interrupted.  For consequent
777 			 * buffers, do not sleep, since several
778 			 * threads might exhaust the buffers and then
779 			 * deadlock.
780 			 */
781 			sf = sf_buf_alloc(pa[i],
782 			    m != NULL ? SFB_NOWAIT : SFB_CATCH);
783 			if (sf == NULL) {
784 				SFSTAT_INC(sf_allocfail);
785 				for (int j = i; j < npages; j++) {
786 					vm_page_lock(pa[j]);
787 					vm_page_unwire(pa[j], PQ_INACTIVE);
788 					vm_page_unlock(pa[j]);
789 				}
790 				if (m == NULL)
791 					softerr = ENOBUFS;
792 				fixspace(npages, i, off, &space);
793 				npages = i;
794 				break;
795 			}
796 
797 			m0 = m_get(M_WAITOK, MT_DATA);
798 			m0->m_ext.ext_buf = (char *)sf_buf_kva(sf);
799 			m0->m_ext.ext_size = PAGE_SIZE;
800 			m0->m_ext.ext_arg1 = sf;
801 			m0->m_ext.ext_type = EXT_SFBUF;
802 			m0->m_ext.ext_flags = EXT_FLAG_EMBREF;
803 			m0->m_ext.ext_free = sendfile_free_mext;
804 			/*
805 			 * SF_NOCACHE sets the page as being freed upon send.
806 			 * However, we ignore it for the last page in 'space',
807 			 * if the page is truncated, and we got more data to
808 			 * send (rem > space), or if we have readahead
809 			 * configured (rhpages > 0).
810 			 */
811 			if ((flags & SF_NOCACHE) &&
812 			    (i != npages - 1 ||
813 			    !((off + space) & PAGE_MASK) ||
814 			    !(rem > space || rhpages > 0)))
815 				m0->m_ext.ext_flags |= EXT_FLAG_NOCACHE;
816 			if (sfs != NULL) {
817 				m0->m_ext.ext_flags |= EXT_FLAG_SYNC;
818 				m0->m_ext.ext_arg2 = sfs;
819 				mtx_lock(&sfs->mtx);
820 				sfs->count++;
821 				mtx_unlock(&sfs->mtx);
822 			}
823 			m0->m_ext.ext_count = 1;
824 			m0->m_flags |= (M_EXT | M_RDONLY);
825 			if (nios)
826 				m0->m_flags |= M_NOTREADY;
827 			m0->m_data = (char *)sf_buf_kva(sf) +
828 			    (vmoff(i, off) & PAGE_MASK);
829 			m0->m_len = xfsize(i, npages, off, space);
830 
831 			if (i == 0)
832 				sfio->m = m0;
833 
834 			/* Append to mbuf chain. */
835 			if (mtail != NULL)
836 				mtail->m_next = m0;
837 			else
838 				m = m0;
839 			mtail = m0;
840 		}
841 
842 		if (vp != NULL)
843 			VOP_UNLOCK(vp, 0);
844 
845 		/* Keep track of bytes processed. */
846 		off += space;
847 		rem -= space;
848 
849 		/* Prepend header, if any. */
850 		if (hdrlen) {
851 prepend_header:
852 			mhtail->m_next = m;
853 			m = mh;
854 			mh = NULL;
855 		}
856 
857 		if (m == NULL) {
858 			KASSERT(softerr, ("%s: m NULL, no error", __func__));
859 			error = softerr;
860 			free(sfio, M_TEMP);
861 			goto done;
862 		}
863 
864 		/* Add the buffer chain to the socket buffer. */
865 		KASSERT(m_length(m, NULL) == space + hdrlen,
866 		    ("%s: mlen %u space %d hdrlen %d",
867 		    __func__, m_length(m, NULL), space, hdrlen));
868 
869 		CURVNET_SET(so->so_vnet);
870 		if (nios == 0) {
871 			/*
872 			 * If sendfile_swapin() didn't initiate any I/Os,
873 			 * which happens if all data is cached in VM, then
874 			 * we can send data right now without the
875 			 * PRUS_NOTREADY flag.
876 			 */
877 			free(sfio, M_TEMP);
878 			error = (*so->so_proto->pr_usrreqs->pru_send)
879 			    (so, 0, m, NULL, NULL, td);
880 		} else {
881 			sfio->npages = npages;
882 			soref(so);
883 			error = (*so->so_proto->pr_usrreqs->pru_send)
884 			    (so, PRUS_NOTREADY, m, NULL, NULL, td);
885 			sendfile_iodone(sfio, NULL, 0, 0);
886 		}
887 		CURVNET_RESTORE();
888 
889 		m = NULL;	/* pru_send always consumes */
890 		if (error)
891 			goto done;
892 		sbytes += space + hdrlen;
893 		if (hdrlen)
894 			hdrlen = 0;
895 		if (softerr) {
896 			error = softerr;
897 			goto done;
898 		}
899 	}
900 
901 	/*
902 	 * Send trailers. Wimp out and use writev(2).
903 	 */
904 	if (trl_uio != NULL) {
905 		sbunlock(&so->so_snd);
906 		error = kern_writev(td, sockfd, trl_uio);
907 		if (error == 0)
908 			sbytes += td->td_retval[0];
909 		goto out;
910 	}
911 
912 done:
913 	sbunlock(&so->so_snd);
914 out:
915 	/*
916 	 * If there was no error we have to clear td->td_retval[0]
917 	 * because it may have been set by writev.
918 	 */
919 	if (error == 0) {
920 		td->td_retval[0] = 0;
921 	}
922 	if (sent != NULL) {
923 		(*sent) = sbytes;
924 	}
925 	if (obj != NULL)
926 		vm_object_deallocate(obj);
927 	if (so)
928 		fdrop(sock_fp, td);
929 	if (m)
930 		m_freem(m);
931 	if (mh)
932 		m_freem(mh);
933 
934 	if (sfs != NULL) {
935 		mtx_lock(&sfs->mtx);
936 		if (sfs->count != 0)
937 			cv_wait(&sfs->cv, &sfs->mtx);
938 		KASSERT(sfs->count == 0, ("sendfile sync still busy"));
939 		cv_destroy(&sfs->cv);
940 		mtx_destroy(&sfs->mtx);
941 		free(sfs, M_TEMP);
942 	}
943 
944 	if (error == ERESTART)
945 		error = EINTR;
946 
947 	return (error);
948 }
949 
950 static int
951 sendfile(struct thread *td, struct sendfile_args *uap, int compat)
952 {
953 	struct sf_hdtr hdtr;
954 	struct uio *hdr_uio, *trl_uio;
955 	struct file *fp;
956 	cap_rights_t rights;
957 	off_t sbytes;
958 	int error;
959 
960 	/*
961 	 * File offset must be positive.  If it goes beyond EOF
962 	 * we send only the header/trailer and no payload data.
963 	 */
964 	if (uap->offset < 0)
965 		return (EINVAL);
966 
967 	sbytes = 0;
968 	hdr_uio = trl_uio = NULL;
969 
970 	if (uap->hdtr != NULL) {
971 		error = copyin(uap->hdtr, &hdtr, sizeof(hdtr));
972 		if (error != 0)
973 			goto out;
974 		if (hdtr.headers != NULL) {
975 			error = copyinuio(hdtr.headers, hdtr.hdr_cnt,
976 			    &hdr_uio);
977 			if (error != 0)
978 				goto out;
979 #ifdef COMPAT_FREEBSD4
980 			/*
981 			 * In FreeBSD < 5.0 the nbytes to send also included
982 			 * the header.  If compat is specified subtract the
983 			 * header size from nbytes.
984 			 */
985 			if (compat) {
986 				if (uap->nbytes > hdr_uio->uio_resid)
987 					uap->nbytes -= hdr_uio->uio_resid;
988 				else
989 					uap->nbytes = 0;
990 			}
991 #endif
992 		}
993 		if (hdtr.trailers != NULL) {
994 			error = copyinuio(hdtr.trailers, hdtr.trl_cnt,
995 			    &trl_uio);
996 			if (error != 0)
997 				goto out;
998 		}
999 	}
1000 
1001 	AUDIT_ARG_FD(uap->fd);
1002 
1003 	/*
1004 	 * sendfile(2) can start at any offset within a file so we require
1005 	 * CAP_READ+CAP_SEEK = CAP_PREAD.
1006 	 */
1007 	if ((error = fget_read(td, uap->fd,
1008 	    cap_rights_init(&rights, CAP_PREAD), &fp)) != 0) {
1009 		goto out;
1010 	}
1011 
1012 	error = fo_sendfile(fp, uap->s, hdr_uio, trl_uio, uap->offset,
1013 	    uap->nbytes, &sbytes, uap->flags, td);
1014 	fdrop(fp, td);
1015 
1016 	if (uap->sbytes != NULL)
1017 		copyout(&sbytes, uap->sbytes, sizeof(off_t));
1018 
1019 out:
1020 	free(hdr_uio, M_IOV);
1021 	free(trl_uio, M_IOV);
1022 	return (error);
1023 }
1024 
1025 /*
1026  * sendfile(2)
1027  *
1028  * int sendfile(int fd, int s, off_t offset, size_t nbytes,
1029  *       struct sf_hdtr *hdtr, off_t *sbytes, int flags)
1030  *
1031  * Send a file specified by 'fd' and starting at 'offset' to a socket
1032  * specified by 's'. Send only 'nbytes' of the file or until EOF if nbytes ==
1033  * 0.  Optionally add a header and/or trailer to the socket output.  If
1034  * specified, write the total number of bytes sent into *sbytes.
1035  */
1036 int
1037 sys_sendfile(struct thread *td, struct sendfile_args *uap)
1038 {
1039 
1040 	return (sendfile(td, uap, 0));
1041 }
1042 
1043 #ifdef COMPAT_FREEBSD4
1044 int
1045 freebsd4_sendfile(struct thread *td, struct freebsd4_sendfile_args *uap)
1046 {
1047 	struct sendfile_args args;
1048 
1049 	args.fd = uap->fd;
1050 	args.s = uap->s;
1051 	args.offset = uap->offset;
1052 	args.nbytes = uap->nbytes;
1053 	args.hdtr = uap->hdtr;
1054 	args.sbytes = uap->sbytes;
1055 	args.flags = uap->flags;
1056 
1057 	return (sendfile(td, &args, 1));
1058 }
1059 #endif /* COMPAT_FREEBSD4 */
1060