xref: /freebsd/sys/kern/kern_sendfile.c (revision c93b6e5fa24ba172ab271432c6692f9cc604e15a)
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 <sys/param.h>
34 #include <sys/systm.h>
35 #include <sys/capsicum.h>
36 #include <sys/kernel.h>
37 #include <netinet/in.h>
38 #include <sys/lock.h>
39 #include <sys/mutex.h>
40 #include <sys/sysproto.h>
41 #include <sys/malloc.h>
42 #include <sys/proc.h>
43 #include <sys/mman.h>
44 #include <sys/mount.h>
45 #include <sys/mbuf.h>
46 #include <sys/protosw.h>
47 #include <sys/rwlock.h>
48 #include <sys/sf_buf.h>
49 #include <sys/socket.h>
50 #include <sys/socketvar.h>
51 #include <sys/syscallsubr.h>
52 #include <sys/sysctl.h>
53 #include <sys/vnode.h>
54 
55 #include <net/vnet.h>
56 
57 #include <security/audit/audit.h>
58 #include <security/mac/mac_framework.h>
59 
60 #include <vm/vm.h>
61 #include <vm/vm_object.h>
62 #include <vm/vm_pager.h>
63 
64 #define	EXT_FLAG_SYNC		EXT_FLAG_VENDOR1
65 #define	EXT_FLAG_NOCACHE	EXT_FLAG_VENDOR2
66 #define	EXT_FLAG_CACHE_LAST	EXT_FLAG_VENDOR3
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 (vm_page_active(pg))
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 static void
207 sendfile_free_mext_pg(struct mbuf *m)
208 {
209 	struct mbuf_ext_pgs *ext_pgs;
210 	vm_page_t pg;
211 	int i;
212 	bool nocache, cache_last;
213 
214 	KASSERT(m->m_flags & M_EXT && m->m_ext.ext_type == EXT_PGS,
215 	    ("%s: m %p !M_EXT or !EXT_PGS", __func__, m));
216 
217 	nocache = m->m_ext.ext_flags & EXT_FLAG_NOCACHE;
218 	cache_last = m->m_ext.ext_flags & EXT_FLAG_CACHE_LAST;
219 	ext_pgs = m->m_ext.ext_pgs;
220 
221 	for (i = 0; i < ext_pgs->npgs; i++) {
222 		if (cache_last && i == ext_pgs->npgs - 1)
223 			nocache = false;
224 		pg = PHYS_TO_VM_PAGE(ext_pgs->pa[i]);
225 		sendfile_free_page(pg, nocache);
226 	}
227 
228 	if (m->m_ext.ext_flags & EXT_FLAG_SYNC) {
229 		struct sendfile_sync *sfs = m->m_ext.ext_arg2;
230 
231 		mtx_lock(&sfs->mtx);
232 		KASSERT(sfs->count > 0, ("Sendfile sync botchup count == 0"));
233 		if (--sfs->count == 0)
234 			cv_signal(&sfs->cv);
235 		mtx_unlock(&sfs->mtx);
236 	}
237 }
238 
239 /*
240  * Helper function to calculate how much data to put into page i of n.
241  * Only first and last pages are special.
242  */
243 static inline off_t
244 xfsize(int i, int n, off_t off, off_t len)
245 {
246 
247 	if (i == 0)
248 		return (omin(PAGE_SIZE - (off & PAGE_MASK), len));
249 
250 	if (i == n - 1 && ((off + len) & PAGE_MASK) > 0)
251 		return ((off + len) & PAGE_MASK);
252 
253 	return (PAGE_SIZE);
254 }
255 
256 /*
257  * Helper function to get offset within object for i page.
258  */
259 static inline vm_ooffset_t
260 vmoff(int i, off_t off)
261 {
262 
263 	if (i == 0)
264 		return ((vm_ooffset_t)off);
265 
266 	return (trunc_page(off + i * PAGE_SIZE));
267 }
268 
269 /*
270  * Helper function used when allocation of a page or sf_buf failed.
271  * Pretend as if we don't have enough space, subtract xfsize() of
272  * all pages that failed.
273  */
274 static inline void
275 fixspace(int old, int new, off_t off, int *space)
276 {
277 
278 	KASSERT(old > new, ("%s: old %d new %d", __func__, old, new));
279 
280 	/* Subtract last one. */
281 	*space -= xfsize(old - 1, old, off, *space);
282 	old--;
283 
284 	if (new == old)
285 		/* There was only one page. */
286 		return;
287 
288 	/* Subtract first one. */
289 	if (new == 0) {
290 		*space -= xfsize(0, old, off, *space);
291 		new++;
292 	}
293 
294 	/* Rest of pages are full sized. */
295 	*space -= (old - new) * PAGE_SIZE;
296 
297 	KASSERT(*space >= 0, ("%s: space went backwards", __func__));
298 }
299 
300 /*
301  * I/O completion callback.
302  */
303 static void
304 sendfile_iodone(void *arg, vm_page_t *pg, int count, int error)
305 {
306 	struct sf_io *sfio = arg;
307 	struct socket *so = sfio->so;
308 
309 	for (int i = 0; i < count; i++)
310 		if (pg[i] != bogus_page)
311 			vm_page_xunbusy(pg[i]);
312 
313 	if (error)
314 		sfio->error = error;
315 
316 	if (!refcount_release(&sfio->nios))
317 		return;
318 
319 	CURVNET_SET(so->so_vnet);
320 	if (sfio->error) {
321 		/*
322 		 * I/O operation failed.  The state of data in the socket
323 		 * is now inconsistent, and all what we can do is to tear
324 		 * it down. Protocol abort method would tear down protocol
325 		 * state, free all ready mbufs and detach not ready ones.
326 		 * We will free the mbufs corresponding to this I/O manually.
327 		 *
328 		 * The socket would be marked with EIO and made available
329 		 * for read, so that application receives EIO on next
330 		 * syscall and eventually closes the socket.
331 		 */
332 		so->so_proto->pr_usrreqs->pru_abort(so);
333 		so->so_error = EIO;
334 
335 		mb_free_notready(sfio->m, sfio->npages);
336 	} else
337 		(void)(so->so_proto->pr_usrreqs->pru_ready)(so, sfio->m,
338 		    sfio->npages);
339 
340 	SOCK_LOCK(so);
341 	sorele(so);
342 	CURVNET_RESTORE();
343 	free(sfio, M_TEMP);
344 }
345 
346 /*
347  * Iterate through pages vector and request paging for non-valid pages.
348  */
349 static int
350 sendfile_swapin(vm_object_t obj, struct sf_io *sfio, off_t off, off_t len,
351     int npages, int rhpages, int flags)
352 {
353 	vm_page_t *pa = sfio->pa;
354 	int grabbed, nios;
355 
356 	nios = 0;
357 	flags = (flags & SF_NODISKIO) ? VM_ALLOC_NOWAIT : 0;
358 
359 	/*
360 	 * First grab all the pages and wire them.  Note that we grab
361 	 * only required pages.  Readahead pages are dealt with later.
362 	 */
363 	VM_OBJECT_WLOCK(obj);
364 
365 	grabbed = vm_page_grab_pages(obj, OFF_TO_IDX(off),
366 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED | flags, pa, npages);
367 	if (grabbed < npages) {
368 		for (int i = grabbed; i < npages; i++)
369 			pa[i] = NULL;
370 		npages = grabbed;
371 		rhpages = 0;
372 	}
373 
374 	for (int i = 0; i < npages;) {
375 		int j, a, count, rv __unused;
376 
377 		/* Skip valid pages. */
378 		if (vm_page_is_valid(pa[i], vmoff(i, off) & PAGE_MASK,
379 		    xfsize(i, npages, off, len))) {
380 			vm_page_xunbusy(pa[i]);
381 			SFSTAT_INC(sf_pages_valid);
382 			i++;
383 			continue;
384 		}
385 
386 		/*
387 		 * Next page is invalid.  Check if it belongs to pager.  It
388 		 * may not be there, which is a regular situation for shmem
389 		 * pager.  For vnode pager this happens only in case of
390 		 * a sparse file.
391 		 *
392 		 * Important feature of vm_pager_has_page() is the hint
393 		 * stored in 'a', about how many pages we can pagein after
394 		 * this page in a single I/O.
395 		 */
396 		if (!vm_pager_has_page(obj, OFF_TO_IDX(vmoff(i, off)), NULL,
397 		    &a)) {
398 			pmap_zero_page(pa[i]);
399 			pa[i]->valid = VM_PAGE_BITS_ALL;
400 			MPASS(pa[i]->dirty == 0);
401 			vm_page_xunbusy(pa[i]);
402 			i++;
403 			continue;
404 		}
405 
406 		/*
407 		 * We want to pagein as many pages as possible, limited only
408 		 * by the 'a' hint and actual request.
409 		 */
410 		count = min(a + 1, npages - i);
411 
412 		/*
413 		 * We should not pagein into a valid page, thus we first trim
414 		 * any valid pages off the end of request, and substitute
415 		 * to bogus_page those, that are in the middle.
416 		 */
417 		for (j = i + count - 1; j > i; j--) {
418 			if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
419 			    xfsize(j, npages, off, len))) {
420 				count--;
421 				rhpages = 0;
422 			} else
423 				break;
424 		}
425 		for (j = i + 1; j < i + count - 1; j++)
426 			if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
427 			    xfsize(j, npages, off, len))) {
428 				vm_page_xunbusy(pa[j]);
429 				SFSTAT_INC(sf_pages_valid);
430 				SFSTAT_INC(sf_pages_bogus);
431 				pa[j] = bogus_page;
432 			}
433 
434 		refcount_acquire(&sfio->nios);
435 		rv = vm_pager_get_pages_async(obj, pa + i, count, NULL,
436 		    i + count == npages ? &rhpages : NULL,
437 		    &sendfile_iodone, sfio);
438 		KASSERT(rv == VM_PAGER_OK, ("%s: pager fail obj %p page %p",
439 		    __func__, obj, pa[i]));
440 
441 		SFSTAT_INC(sf_iocnt);
442 		SFSTAT_ADD(sf_pages_read, count);
443 		if (i + count == npages)
444 			SFSTAT_ADD(sf_rhpages_read, rhpages);
445 
446 		/*
447 		 * Restore the valid page pointers.  They are already
448 		 * unbusied, but still wired.
449 		 */
450 		for (j = i; j < i + count; j++)
451 			if (pa[j] == bogus_page) {
452 				pa[j] = vm_page_lookup(obj,
453 				    OFF_TO_IDX(vmoff(j, off)));
454 				KASSERT(pa[j], ("%s: page %p[%d] disappeared",
455 				    __func__, pa, j));
456 
457 			}
458 		i += count;
459 		nios++;
460 	}
461 
462 	VM_OBJECT_WUNLOCK(obj);
463 
464 	if (nios == 0 && npages != 0)
465 		SFSTAT_INC(sf_noiocnt);
466 
467 	return (nios);
468 }
469 
470 static int
471 sendfile_getobj(struct thread *td, struct file *fp, vm_object_t *obj_res,
472     struct vnode **vp_res, struct shmfd **shmfd_res, off_t *obj_size,
473     int *bsize)
474 {
475 	struct vattr va;
476 	vm_object_t obj;
477 	struct vnode *vp;
478 	struct shmfd *shmfd;
479 	int error;
480 
481 	vp = *vp_res = NULL;
482 	obj = NULL;
483 	shmfd = *shmfd_res = NULL;
484 	*bsize = 0;
485 
486 	/*
487 	 * The file descriptor must be a regular file and have a
488 	 * backing VM object.
489 	 */
490 	if (fp->f_type == DTYPE_VNODE) {
491 		vp = fp->f_vnode;
492 		vn_lock(vp, LK_SHARED | LK_RETRY);
493 		if (vp->v_type != VREG) {
494 			error = EINVAL;
495 			goto out;
496 		}
497 		*bsize = vp->v_mount->mnt_stat.f_iosize;
498 		error = VOP_GETATTR(vp, &va, td->td_ucred);
499 		if (error != 0)
500 			goto out;
501 		*obj_size = va.va_size;
502 		obj = vp->v_object;
503 		if (obj == NULL) {
504 			error = EINVAL;
505 			goto out;
506 		}
507 	} else if (fp->f_type == DTYPE_SHM) {
508 		error = 0;
509 		shmfd = fp->f_data;
510 		obj = shmfd->shm_object;
511 		*obj_size = shmfd->shm_size;
512 	} else {
513 		error = EINVAL;
514 		goto out;
515 	}
516 
517 	VM_OBJECT_WLOCK(obj);
518 	if ((obj->flags & OBJ_DEAD) != 0) {
519 		VM_OBJECT_WUNLOCK(obj);
520 		error = EBADF;
521 		goto out;
522 	}
523 
524 	/*
525 	 * Temporarily increase the backing VM object's reference
526 	 * count so that a forced reclamation of its vnode does not
527 	 * immediately destroy it.
528 	 */
529 	vm_object_reference_locked(obj);
530 	VM_OBJECT_WUNLOCK(obj);
531 	*obj_res = obj;
532 	*vp_res = vp;
533 	*shmfd_res = shmfd;
534 
535 out:
536 	if (vp != NULL)
537 		VOP_UNLOCK(vp, 0);
538 	return (error);
539 }
540 
541 static int
542 sendfile_getsock(struct thread *td, int s, struct file **sock_fp,
543     struct socket **so)
544 {
545 	int error;
546 
547 	*sock_fp = NULL;
548 	*so = NULL;
549 
550 	/*
551 	 * The socket must be a stream socket and connected.
552 	 */
553 	error = getsock_cap(td, s, &cap_send_rights,
554 	    sock_fp, NULL, NULL);
555 	if (error != 0)
556 		return (error);
557 	*so = (*sock_fp)->f_data;
558 	if ((*so)->so_type != SOCK_STREAM)
559 		return (EINVAL);
560 	if (SOLISTENING(*so))
561 		return (ENOTCONN);
562 	return (0);
563 }
564 
565 int
566 vn_sendfile(struct file *fp, int sockfd, struct uio *hdr_uio,
567     struct uio *trl_uio, off_t offset, size_t nbytes, off_t *sent, int flags,
568     struct thread *td)
569 {
570 	struct file *sock_fp;
571 	struct vnode *vp;
572 	struct vm_object *obj;
573 	struct socket *so;
574 	struct mbuf_ext_pgs *ext_pgs;
575 	struct mbuf *m, *mh, *mhtail;
576 	struct sf_buf *sf;
577 	struct shmfd *shmfd;
578 	struct sendfile_sync *sfs;
579 	struct vattr va;
580 	off_t off, sbytes, rem, obj_size;
581 	int bsize, error, ext_pgs_idx, hdrlen, max_pgs, softerr;
582 	bool use_ext_pgs;
583 
584 	obj = NULL;
585 	so = NULL;
586 	m = mh = NULL;
587 	sfs = NULL;
588 	hdrlen = sbytes = 0;
589 	softerr = 0;
590 	use_ext_pgs = false;
591 
592 	error = sendfile_getobj(td, fp, &obj, &vp, &shmfd, &obj_size, &bsize);
593 	if (error != 0)
594 		return (error);
595 
596 	error = sendfile_getsock(td, sockfd, &sock_fp, &so);
597 	if (error != 0)
598 		goto out;
599 
600 #ifdef MAC
601 	error = mac_socket_check_send(td->td_ucred, so);
602 	if (error != 0)
603 		goto out;
604 #endif
605 
606 	SFSTAT_INC(sf_syscalls);
607 	SFSTAT_ADD(sf_rhpages_requested, SF_READAHEAD(flags));
608 
609 	if (flags & SF_SYNC) {
610 		sfs = malloc(sizeof *sfs, M_TEMP, M_WAITOK | M_ZERO);
611 		mtx_init(&sfs->mtx, "sendfile", NULL, MTX_DEF);
612 		cv_init(&sfs->cv, "sendfile");
613 	}
614 
615 	rem = nbytes ? omin(nbytes, obj_size - offset) : obj_size - offset;
616 
617 	/*
618 	 * Protect against multiple writers to the socket.
619 	 *
620 	 * XXXRW: Historically this has assumed non-interruptibility, so now
621 	 * we implement that, but possibly shouldn't.
622 	 */
623 	(void)sblock(&so->so_snd, SBL_WAIT | SBL_NOINTR);
624 
625 	/*
626 	 * Loop through the pages of the file, starting with the requested
627 	 * offset. Get a file page (do I/O if necessary), map the file page
628 	 * into an sf_buf, attach an mbuf header to the sf_buf, and queue
629 	 * it on the socket.
630 	 * This is done in two loops.  The inner loop turns as many pages
631 	 * as it can, up to available socket buffer space, without blocking
632 	 * into mbufs to have it bulk delivered into the socket send buffer.
633 	 * The outer loop checks the state and available space of the socket
634 	 * and takes care of the overall progress.
635 	 */
636 	for (off = offset; rem > 0; ) {
637 		struct sf_io *sfio;
638 		vm_page_t *pa;
639 		struct mbuf *mtail;
640 		int nios, space, npages, rhpages;
641 
642 		mtail = NULL;
643 		/*
644 		 * Check the socket state for ongoing connection,
645 		 * no errors and space in socket buffer.
646 		 * If space is low allow for the remainder of the
647 		 * file to be processed if it fits the socket buffer.
648 		 * Otherwise block in waiting for sufficient space
649 		 * to proceed, or if the socket is nonblocking, return
650 		 * to userland with EAGAIN while reporting how far
651 		 * we've come.
652 		 * We wait until the socket buffer has significant free
653 		 * space to do bulk sends.  This makes good use of file
654 		 * system read ahead and allows packet segmentation
655 		 * offloading hardware to take over lots of work.  If
656 		 * we were not careful here we would send off only one
657 		 * sfbuf at a time.
658 		 */
659 		SOCKBUF_LOCK(&so->so_snd);
660 		if (so->so_snd.sb_lowat < so->so_snd.sb_hiwat / 2)
661 			so->so_snd.sb_lowat = so->so_snd.sb_hiwat / 2;
662 retry_space:
663 		if (so->so_snd.sb_state & SBS_CANTSENDMORE) {
664 			error = EPIPE;
665 			SOCKBUF_UNLOCK(&so->so_snd);
666 			goto done;
667 		} else if (so->so_error) {
668 			error = so->so_error;
669 			so->so_error = 0;
670 			SOCKBUF_UNLOCK(&so->so_snd);
671 			goto done;
672 		}
673 		if ((so->so_state & SS_ISCONNECTED) == 0) {
674 			SOCKBUF_UNLOCK(&so->so_snd);
675 			error = ENOTCONN;
676 			goto done;
677 		}
678 
679 		space = sbspace(&so->so_snd);
680 		if (space < rem &&
681 		    (space <= 0 ||
682 		     space < so->so_snd.sb_lowat)) {
683 			if (so->so_state & SS_NBIO) {
684 				SOCKBUF_UNLOCK(&so->so_snd);
685 				error = EAGAIN;
686 				goto done;
687 			}
688 			/*
689 			 * sbwait drops the lock while sleeping.
690 			 * When we loop back to retry_space the
691 			 * state may have changed and we retest
692 			 * for it.
693 			 */
694 			error = sbwait(&so->so_snd);
695 			/*
696 			 * An error from sbwait usually indicates that we've
697 			 * been interrupted by a signal. If we've sent anything
698 			 * then return bytes sent, otherwise return the error.
699 			 */
700 			if (error != 0) {
701 				SOCKBUF_UNLOCK(&so->so_snd);
702 				goto done;
703 			}
704 			goto retry_space;
705 		}
706 		SOCKBUF_UNLOCK(&so->so_snd);
707 
708 		/*
709 		 * At the beginning of the first loop check if any headers
710 		 * are specified and copy them into mbufs.  Reduce space in
711 		 * the socket buffer by the size of the header mbuf chain.
712 		 * Clear hdr_uio here and hdrlen at the end of the first loop.
713 		 */
714 		if (hdr_uio != NULL && hdr_uio->uio_resid > 0) {
715 			hdr_uio->uio_td = td;
716 			hdr_uio->uio_rw = UIO_WRITE;
717 			mh = m_uiotombuf(hdr_uio, M_WAITOK, space, 0, 0);
718 			hdrlen = m_length(mh, &mhtail);
719 			space -= hdrlen;
720 			/*
721 			 * If header consumed all the socket buffer space,
722 			 * don't waste CPU cycles and jump to the end.
723 			 */
724 			if (space == 0) {
725 				sfio = NULL;
726 				nios = 0;
727 				goto prepend_header;
728 			}
729 			hdr_uio = NULL;
730 		}
731 
732 		if (vp != NULL) {
733 			error = vn_lock(vp, LK_SHARED);
734 			if (error != 0)
735 				goto done;
736 			error = VOP_GETATTR(vp, &va, td->td_ucred);
737 			if (error != 0 || off >= va.va_size) {
738 				VOP_UNLOCK(vp, 0);
739 				goto done;
740 			}
741 			if (va.va_size != obj_size) {
742 				obj_size = va.va_size;
743 				rem = nbytes ?
744 				    omin(nbytes + offset, obj_size) : obj_size;
745 				rem -= off;
746 			}
747 		}
748 
749 		if (space > rem)
750 			space = rem;
751 		else if (space > PAGE_SIZE) {
752 			/*
753 			 * Use page boundaries when possible for large
754 			 * requests.
755 			 */
756 			if (off & PAGE_MASK)
757 				space -= (PAGE_SIZE - (off & PAGE_MASK));
758 			space = trunc_page(space);
759 			if (off & PAGE_MASK)
760 				space += (PAGE_SIZE - (off & PAGE_MASK));
761 		}
762 
763 		npages = howmany(space + (off & PAGE_MASK), PAGE_SIZE);
764 
765 		/*
766 		 * Calculate maximum allowed number of pages for readahead
767 		 * at this iteration.  If SF_USER_READAHEAD was set, we don't
768 		 * do any heuristics and use exactly the value supplied by
769 		 * application.  Otherwise, we allow readahead up to "rem".
770 		 * If application wants more, let it be, but there is no
771 		 * reason to go above MAXPHYS.  Also check against "obj_size",
772 		 * since vm_pager_has_page() can hint beyond EOF.
773 		 */
774 		if (flags & SF_USER_READAHEAD) {
775 			rhpages = SF_READAHEAD(flags);
776 		} else {
777 			rhpages = howmany(rem + (off & PAGE_MASK), PAGE_SIZE) -
778 			    npages;
779 			rhpages += SF_READAHEAD(flags);
780 		}
781 		rhpages = min(howmany(MAXPHYS, PAGE_SIZE), rhpages);
782 		rhpages = min(howmany(obj_size - trunc_page(off), PAGE_SIZE) -
783 		    npages, rhpages);
784 
785 		sfio = malloc(sizeof(struct sf_io) +
786 		    npages * sizeof(vm_page_t), M_TEMP, M_WAITOK);
787 		refcount_init(&sfio->nios, 1);
788 		sfio->so = so;
789 		sfio->error = 0;
790 
791 		nios = sendfile_swapin(obj, sfio, off, space, npages, rhpages,
792 		    flags);
793 
794 		/*
795 		 * Loop and construct maximum sized mbuf chain to be bulk
796 		 * dumped into socket buffer.
797 		 */
798 		pa = sfio->pa;
799 
800 		/*
801 		 * Use unmapped mbufs if enabled for TCP.  Unmapped
802 		 * bufs are restricted to TCP as that is what has been
803 		 * tested.  In particular, unmapped mbufs have not
804 		 * been tested with UNIX-domain sockets.
805 		 */
806 		if (mb_use_ext_pgs &&
807 		    so->so_proto->pr_protocol == IPPROTO_TCP) {
808 			use_ext_pgs = true;
809 			max_pgs = MBUF_PEXT_MAX_PGS;
810 
811 			/* Start at last index, to wrap on first use. */
812 			ext_pgs_idx = max_pgs - 1;
813 		}
814 
815 		for (int i = 0; i < npages; i++) {
816 			struct mbuf *m0;
817 
818 			/*
819 			 * If a page wasn't grabbed successfully, then
820 			 * trim the array. Can happen only with SF_NODISKIO.
821 			 */
822 			if (pa[i] == NULL) {
823 				SFSTAT_INC(sf_busy);
824 				fixspace(npages, i, off, &space);
825 				npages = i;
826 				softerr = EBUSY;
827 				break;
828 			}
829 
830 			if (use_ext_pgs) {
831 				off_t xfs;
832 
833 				ext_pgs_idx++;
834 				if (ext_pgs_idx == max_pgs) {
835 					m0 = mb_alloc_ext_pgs(M_WAITOK, false,
836 					    sendfile_free_mext_pg);
837 
838 					if (flags & SF_NOCACHE) {
839 						m0->m_ext.ext_flags |=
840 						    EXT_FLAG_NOCACHE;
841 
842 						/*
843 						 * See comment below regarding
844 						 * ignoring SF_NOCACHE for the
845 						 * last page.
846 						 */
847 						if ((npages - i <= max_pgs) &&
848 						    ((off + space) & PAGE_MASK) &&
849 						    (rem > space || rhpages > 0))
850 							m0->m_ext.ext_flags |=
851 							    EXT_FLAG_CACHE_LAST;
852 					}
853 					if (sfs != NULL) {
854 						m0->m_ext.ext_flags |=
855 						    EXT_FLAG_SYNC;
856 						m0->m_ext.ext_arg2 = sfs;
857 						mtx_lock(&sfs->mtx);
858 						sfs->count++;
859 						mtx_unlock(&sfs->mtx);
860 					}
861 					ext_pgs = m0->m_ext.ext_pgs;
862 					if (i == 0)
863 						sfio->m = m0;
864 					ext_pgs_idx = 0;
865 
866 					/* Append to mbuf chain. */
867 					if (mtail != NULL)
868 						mtail->m_next = m0;
869 					else
870 						m = m0;
871 					mtail = m0;
872 					ext_pgs->first_pg_off =
873 					    vmoff(i, off) & PAGE_MASK;
874 				}
875 				if (nios) {
876 					mtail->m_flags |= M_NOTREADY;
877 					ext_pgs->nrdy++;
878 				}
879 
880 				ext_pgs->pa[ext_pgs_idx] = VM_PAGE_TO_PHYS(pa[i]);
881 				ext_pgs->npgs++;
882 				xfs = xfsize(i, npages, off, space);
883 				ext_pgs->last_pg_len = xfs;
884 				MBUF_EXT_PGS_ASSERT_SANITY(ext_pgs);
885 				mtail->m_len += xfs;
886 				mtail->m_ext.ext_size += PAGE_SIZE;
887 				continue;
888 			}
889 
890 			/*
891 			 * Get a sendfile buf.  When allocating the
892 			 * first buffer for mbuf chain, we usually
893 			 * wait as long as necessary, but this wait
894 			 * can be interrupted.  For consequent
895 			 * buffers, do not sleep, since several
896 			 * threads might exhaust the buffers and then
897 			 * deadlock.
898 			 */
899 			sf = sf_buf_alloc(pa[i],
900 			    m != NULL ? SFB_NOWAIT : SFB_CATCH);
901 			if (sf == NULL) {
902 				SFSTAT_INC(sf_allocfail);
903 				for (int j = i; j < npages; j++) {
904 					vm_page_lock(pa[j]);
905 					vm_page_unwire(pa[j], PQ_INACTIVE);
906 					vm_page_unlock(pa[j]);
907 				}
908 				if (m == NULL)
909 					softerr = ENOBUFS;
910 				fixspace(npages, i, off, &space);
911 				npages = i;
912 				break;
913 			}
914 
915 			m0 = m_get(M_WAITOK, MT_DATA);
916 			m0->m_ext.ext_buf = (char *)sf_buf_kva(sf);
917 			m0->m_ext.ext_size = PAGE_SIZE;
918 			m0->m_ext.ext_arg1 = sf;
919 			m0->m_ext.ext_type = EXT_SFBUF;
920 			m0->m_ext.ext_flags = EXT_FLAG_EMBREF;
921 			m0->m_ext.ext_free = sendfile_free_mext;
922 			/*
923 			 * SF_NOCACHE sets the page as being freed upon send.
924 			 * However, we ignore it for the last page in 'space',
925 			 * if the page is truncated, and we got more data to
926 			 * send (rem > space), or if we have readahead
927 			 * configured (rhpages > 0).
928 			 */
929 			if ((flags & SF_NOCACHE) &&
930 			    (i != npages - 1 ||
931 			    !((off + space) & PAGE_MASK) ||
932 			    !(rem > space || rhpages > 0)))
933 				m0->m_ext.ext_flags |= EXT_FLAG_NOCACHE;
934 			if (sfs != NULL) {
935 				m0->m_ext.ext_flags |= EXT_FLAG_SYNC;
936 				m0->m_ext.ext_arg2 = sfs;
937 				mtx_lock(&sfs->mtx);
938 				sfs->count++;
939 				mtx_unlock(&sfs->mtx);
940 			}
941 			m0->m_ext.ext_count = 1;
942 			m0->m_flags |= (M_EXT | M_RDONLY);
943 			if (nios)
944 				m0->m_flags |= M_NOTREADY;
945 			m0->m_data = (char *)sf_buf_kva(sf) +
946 			    (vmoff(i, off) & PAGE_MASK);
947 			m0->m_len = xfsize(i, npages, off, space);
948 
949 			if (i == 0)
950 				sfio->m = m0;
951 
952 			/* Append to mbuf chain. */
953 			if (mtail != NULL)
954 				mtail->m_next = m0;
955 			else
956 				m = m0;
957 			mtail = m0;
958 		}
959 
960 		if (vp != NULL)
961 			VOP_UNLOCK(vp, 0);
962 
963 		/* Keep track of bytes processed. */
964 		off += space;
965 		rem -= space;
966 
967 		/* Prepend header, if any. */
968 		if (hdrlen) {
969 prepend_header:
970 			mhtail->m_next = m;
971 			m = mh;
972 			mh = NULL;
973 		}
974 
975 		if (m == NULL) {
976 			KASSERT(softerr, ("%s: m NULL, no error", __func__));
977 			error = softerr;
978 			free(sfio, M_TEMP);
979 			goto done;
980 		}
981 
982 		/* Add the buffer chain to the socket buffer. */
983 		KASSERT(m_length(m, NULL) == space + hdrlen,
984 		    ("%s: mlen %u space %d hdrlen %d",
985 		    __func__, m_length(m, NULL), space, hdrlen));
986 
987 		CURVNET_SET(so->so_vnet);
988 		if (nios == 0) {
989 			/*
990 			 * If sendfile_swapin() didn't initiate any I/Os,
991 			 * which happens if all data is cached in VM, then
992 			 * we can send data right now without the
993 			 * PRUS_NOTREADY flag.
994 			 */
995 			free(sfio, M_TEMP);
996 			error = (*so->so_proto->pr_usrreqs->pru_send)
997 			    (so, 0, m, NULL, NULL, td);
998 		} else {
999 			sfio->npages = npages;
1000 			soref(so);
1001 			error = (*so->so_proto->pr_usrreqs->pru_send)
1002 			    (so, PRUS_NOTREADY, m, NULL, NULL, td);
1003 			sendfile_iodone(sfio, NULL, 0, 0);
1004 		}
1005 		CURVNET_RESTORE();
1006 
1007 		m = NULL;	/* pru_send always consumes */
1008 		if (error)
1009 			goto done;
1010 		sbytes += space + hdrlen;
1011 		if (hdrlen)
1012 			hdrlen = 0;
1013 		if (softerr) {
1014 			error = softerr;
1015 			goto done;
1016 		}
1017 	}
1018 
1019 	/*
1020 	 * Send trailers. Wimp out and use writev(2).
1021 	 */
1022 	if (trl_uio != NULL) {
1023 		sbunlock(&so->so_snd);
1024 		error = kern_writev(td, sockfd, trl_uio);
1025 		if (error == 0)
1026 			sbytes += td->td_retval[0];
1027 		goto out;
1028 	}
1029 
1030 done:
1031 	sbunlock(&so->so_snd);
1032 out:
1033 	/*
1034 	 * If there was no error we have to clear td->td_retval[0]
1035 	 * because it may have been set by writev.
1036 	 */
1037 	if (error == 0) {
1038 		td->td_retval[0] = 0;
1039 	}
1040 	if (sent != NULL) {
1041 		(*sent) = sbytes;
1042 	}
1043 	if (obj != NULL)
1044 		vm_object_deallocate(obj);
1045 	if (so)
1046 		fdrop(sock_fp, td);
1047 	if (m)
1048 		m_freem(m);
1049 	if (mh)
1050 		m_freem(mh);
1051 
1052 	if (sfs != NULL) {
1053 		mtx_lock(&sfs->mtx);
1054 		if (sfs->count != 0)
1055 			cv_wait(&sfs->cv, &sfs->mtx);
1056 		KASSERT(sfs->count == 0, ("sendfile sync still busy"));
1057 		cv_destroy(&sfs->cv);
1058 		mtx_destroy(&sfs->mtx);
1059 		free(sfs, M_TEMP);
1060 	}
1061 
1062 	if (error == ERESTART)
1063 		error = EINTR;
1064 
1065 	return (error);
1066 }
1067 
1068 static int
1069 sendfile(struct thread *td, struct sendfile_args *uap, int compat)
1070 {
1071 	struct sf_hdtr hdtr;
1072 	struct uio *hdr_uio, *trl_uio;
1073 	struct file *fp;
1074 	off_t sbytes;
1075 	int error;
1076 
1077 	/*
1078 	 * File offset must be positive.  If it goes beyond EOF
1079 	 * we send only the header/trailer and no payload data.
1080 	 */
1081 	if (uap->offset < 0)
1082 		return (EINVAL);
1083 
1084 	sbytes = 0;
1085 	hdr_uio = trl_uio = NULL;
1086 
1087 	if (uap->hdtr != NULL) {
1088 		error = copyin(uap->hdtr, &hdtr, sizeof(hdtr));
1089 		if (error != 0)
1090 			goto out;
1091 		if (hdtr.headers != NULL) {
1092 			error = copyinuio(hdtr.headers, hdtr.hdr_cnt,
1093 			    &hdr_uio);
1094 			if (error != 0)
1095 				goto out;
1096 #ifdef COMPAT_FREEBSD4
1097 			/*
1098 			 * In FreeBSD < 5.0 the nbytes to send also included
1099 			 * the header.  If compat is specified subtract the
1100 			 * header size from nbytes.
1101 			 */
1102 			if (compat) {
1103 				if (uap->nbytes > hdr_uio->uio_resid)
1104 					uap->nbytes -= hdr_uio->uio_resid;
1105 				else
1106 					uap->nbytes = 0;
1107 			}
1108 #endif
1109 		}
1110 		if (hdtr.trailers != NULL) {
1111 			error = copyinuio(hdtr.trailers, hdtr.trl_cnt,
1112 			    &trl_uio);
1113 			if (error != 0)
1114 				goto out;
1115 		}
1116 	}
1117 
1118 	AUDIT_ARG_FD(uap->fd);
1119 
1120 	/*
1121 	 * sendfile(2) can start at any offset within a file so we require
1122 	 * CAP_READ+CAP_SEEK = CAP_PREAD.
1123 	 */
1124 	if ((error = fget_read(td, uap->fd, &cap_pread_rights, &fp)) != 0)
1125 		goto out;
1126 
1127 	error = fo_sendfile(fp, uap->s, hdr_uio, trl_uio, uap->offset,
1128 	    uap->nbytes, &sbytes, uap->flags, td);
1129 	fdrop(fp, td);
1130 
1131 	if (uap->sbytes != NULL)
1132 		copyout(&sbytes, uap->sbytes, sizeof(off_t));
1133 
1134 out:
1135 	free(hdr_uio, M_IOV);
1136 	free(trl_uio, M_IOV);
1137 	return (error);
1138 }
1139 
1140 /*
1141  * sendfile(2)
1142  *
1143  * int sendfile(int fd, int s, off_t offset, size_t nbytes,
1144  *       struct sf_hdtr *hdtr, off_t *sbytes, int flags)
1145  *
1146  * Send a file specified by 'fd' and starting at 'offset' to a socket
1147  * specified by 's'. Send only 'nbytes' of the file or until EOF if nbytes ==
1148  * 0.  Optionally add a header and/or trailer to the socket output.  If
1149  * specified, write the total number of bytes sent into *sbytes.
1150  */
1151 int
1152 sys_sendfile(struct thread *td, struct sendfile_args *uap)
1153 {
1154 
1155 	return (sendfile(td, uap, 0));
1156 }
1157 
1158 #ifdef COMPAT_FREEBSD4
1159 int
1160 freebsd4_sendfile(struct thread *td, struct freebsd4_sendfile_args *uap)
1161 {
1162 	struct sendfile_args args;
1163 
1164 	args.fd = uap->fd;
1165 	args.s = uap->s;
1166 	args.offset = uap->offset;
1167 	args.nbytes = uap->nbytes;
1168 	args.hdtr = uap->hdtr;
1169 	args.sbytes = uap->sbytes;
1170 	args.flags = uap->flags;
1171 
1172 	return (sendfile(td, &args, 1));
1173 }
1174 #endif /* COMPAT_FREEBSD4 */
1175