xref: /freebsd/sys/kern/kern_sendfile.c (revision 480093f4440d54b30b3025afeac24b48f2ba7a2e)
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_kern_tls.h"
34 
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/capsicum.h>
38 #include <sys/kernel.h>
39 #include <netinet/in.h>
40 #include <sys/lock.h>
41 #include <sys/ktls.h>
42 #include <sys/mutex.h>
43 #include <sys/sysproto.h>
44 #include <sys/malloc.h>
45 #include <sys/proc.h>
46 #include <sys/mman.h>
47 #include <sys/mount.h>
48 #include <sys/mbuf.h>
49 #include <sys/protosw.h>
50 #include <sys/rwlock.h>
51 #include <sys/sf_buf.h>
52 #include <sys/socket.h>
53 #include <sys/socketvar.h>
54 #include <sys/syscallsubr.h>
55 #include <sys/sysctl.h>
56 #include <sys/vnode.h>
57 
58 #include <net/vnet.h>
59 #include <netinet/tcp.h>
60 
61 #include <security/audit/audit.h>
62 #include <security/mac/mac_framework.h>
63 
64 #include <vm/vm.h>
65 #include <vm/vm_object.h>
66 #include <vm/vm_pager.h>
67 
68 #define	EXT_FLAG_SYNC		EXT_FLAG_VENDOR1
69 #define	EXT_FLAG_NOCACHE	EXT_FLAG_VENDOR2
70 #define	EXT_FLAG_CACHE_LAST	EXT_FLAG_VENDOR3
71 
72 /*
73  * Structure describing a single sendfile(2) I/O, which may consist of
74  * several underlying pager I/Os.
75  *
76  * The syscall context allocates the structure and initializes 'nios'
77  * to 1.  As sendfile_swapin() runs through pages and starts asynchronous
78  * paging operations, it increments 'nios'.
79  *
80  * Every I/O completion calls sendfile_iodone(), which decrements the 'nios',
81  * and the syscall also calls sendfile_iodone() after allocating all mbufs,
82  * linking them and sending to socket.  Whoever reaches zero 'nios' is
83  * responsible to * call pru_ready on the socket, to notify it of readyness
84  * of the data.
85  */
86 struct sf_io {
87 	volatile u_int	nios;
88 	u_int		error;
89 	int		npages;
90 	struct socket	*so;
91 	struct mbuf	*m;
92 	vm_object_t	obj;
93 #ifdef KERN_TLS
94 	struct ktls_session *tls;
95 #endif
96 	vm_page_t	pa[];
97 };
98 
99 /*
100  * Structure used to track requests with SF_SYNC flag.
101  */
102 struct sendfile_sync {
103 	struct mtx	mtx;
104 	struct cv	cv;
105 	unsigned	count;
106 };
107 
108 counter_u64_t sfstat[sizeof(struct sfstat) / sizeof(uint64_t)];
109 
110 static void
111 sfstat_init(const void *unused)
112 {
113 
114 	COUNTER_ARRAY_ALLOC(sfstat, sizeof(struct sfstat) / sizeof(uint64_t),
115 	    M_WAITOK);
116 }
117 SYSINIT(sfstat, SI_SUB_MBUF, SI_ORDER_FIRST, sfstat_init, NULL);
118 
119 static int
120 sfstat_sysctl(SYSCTL_HANDLER_ARGS)
121 {
122 	struct sfstat s;
123 
124 	COUNTER_ARRAY_COPY(sfstat, &s, sizeof(s) / sizeof(uint64_t));
125 	if (req->newptr)
126 		COUNTER_ARRAY_ZERO(sfstat, sizeof(s) / sizeof(uint64_t));
127 	return (SYSCTL_OUT(req, &s, sizeof(s)));
128 }
129 SYSCTL_PROC(_kern_ipc, OID_AUTO, sfstat, CTLTYPE_OPAQUE | CTLFLAG_RW,
130     NULL, 0, sfstat_sysctl, "I", "sendfile statistics");
131 
132 static void
133 sendfile_free_mext(struct mbuf *m)
134 {
135 	struct sf_buf *sf;
136 	vm_page_t pg;
137 	int flags;
138 
139 	KASSERT(m->m_flags & M_EXT && m->m_ext.ext_type == EXT_SFBUF,
140 	    ("%s: m %p !M_EXT or !EXT_SFBUF", __func__, m));
141 
142 	sf = m->m_ext.ext_arg1;
143 	pg = sf_buf_page(sf);
144 	flags = (m->m_ext.ext_flags & EXT_FLAG_NOCACHE) != 0 ? VPR_TRYFREE : 0;
145 
146 	sf_buf_free(sf);
147 	vm_page_release(pg, flags);
148 
149 	if (m->m_ext.ext_flags & EXT_FLAG_SYNC) {
150 		struct sendfile_sync *sfs = m->m_ext.ext_arg2;
151 
152 		mtx_lock(&sfs->mtx);
153 		KASSERT(sfs->count > 0, ("Sendfile sync botchup count == 0"));
154 		if (--sfs->count == 0)
155 			cv_signal(&sfs->cv);
156 		mtx_unlock(&sfs->mtx);
157 	}
158 }
159 
160 static void
161 sendfile_free_mext_pg(struct mbuf *m)
162 {
163 	struct mbuf_ext_pgs *ext_pgs;
164 	vm_page_t pg;
165 	int flags, i;
166 	bool cache_last;
167 
168 	KASSERT(m->m_flags & M_EXT && m->m_ext.ext_type == EXT_PGS,
169 	    ("%s: m %p !M_EXT or !EXT_PGS", __func__, m));
170 
171 	cache_last = m->m_ext.ext_flags & EXT_FLAG_CACHE_LAST;
172 	ext_pgs = m->m_ext.ext_pgs;
173 	flags = (m->m_ext.ext_flags & EXT_FLAG_NOCACHE) != 0 ? VPR_TRYFREE : 0;
174 
175 	for (i = 0; i < ext_pgs->npgs; i++) {
176 		if (cache_last && i == ext_pgs->npgs - 1)
177 			flags = 0;
178 		pg = PHYS_TO_VM_PAGE(ext_pgs->pa[i]);
179 		vm_page_release(pg, flags);
180 	}
181 
182 	if (m->m_ext.ext_flags & EXT_FLAG_SYNC) {
183 		struct sendfile_sync *sfs = m->m_ext.ext_arg2;
184 
185 		mtx_lock(&sfs->mtx);
186 		KASSERT(sfs->count > 0, ("Sendfile sync botchup count == 0"));
187 		if (--sfs->count == 0)
188 			cv_signal(&sfs->cv);
189 		mtx_unlock(&sfs->mtx);
190 	}
191 }
192 
193 /*
194  * Helper function to calculate how much data to put into page i of n.
195  * Only first and last pages are special.
196  */
197 static inline off_t
198 xfsize(int i, int n, off_t off, off_t len)
199 {
200 
201 	if (i == 0)
202 		return (omin(PAGE_SIZE - (off & PAGE_MASK), len));
203 
204 	if (i == n - 1 && ((off + len) & PAGE_MASK) > 0)
205 		return ((off + len) & PAGE_MASK);
206 
207 	return (PAGE_SIZE);
208 }
209 
210 /*
211  * Helper function to get offset within object for i page.
212  */
213 static inline vm_ooffset_t
214 vmoff(int i, off_t off)
215 {
216 
217 	if (i == 0)
218 		return ((vm_ooffset_t)off);
219 
220 	return (trunc_page(off + i * PAGE_SIZE));
221 }
222 
223 /*
224  * Helper function used when allocation of a page or sf_buf failed.
225  * Pretend as if we don't have enough space, subtract xfsize() of
226  * all pages that failed.
227  */
228 static inline void
229 fixspace(int old, int new, off_t off, int *space)
230 {
231 
232 	KASSERT(old > new, ("%s: old %d new %d", __func__, old, new));
233 
234 	/* Subtract last one. */
235 	*space -= xfsize(old - 1, old, off, *space);
236 	old--;
237 
238 	if (new == old)
239 		/* There was only one page. */
240 		return;
241 
242 	/* Subtract first one. */
243 	if (new == 0) {
244 		*space -= xfsize(0, old, off, *space);
245 		new++;
246 	}
247 
248 	/* Rest of pages are full sized. */
249 	*space -= (old - new) * PAGE_SIZE;
250 
251 	KASSERT(*space >= 0, ("%s: space went backwards", __func__));
252 }
253 
254 /*
255  * I/O completion callback.
256  */
257 static void
258 sendfile_iodone(void *arg, vm_page_t *pg, int count, int error)
259 {
260 	struct sf_io *sfio = arg;
261 	struct socket *so = sfio->so;
262 
263 	for (int i = 0; i < count; i++)
264 		if (pg[i] != bogus_page)
265 			vm_page_xunbusy_unchecked(pg[i]);
266 
267 	if (error)
268 		sfio->error = error;
269 
270 	if (!refcount_release(&sfio->nios))
271 		return;
272 
273 	vm_object_pip_wakeup(sfio->obj);
274 
275 	if (__predict_false(sfio->error && sfio->m == NULL)) {
276 		/*
277 		 * I/O operation failed, but pru_send hadn't been executed -
278 		 * nothing had been sent to the socket.  The syscall has
279 		 * returned error to the user.
280 		 */
281 		free(sfio, M_TEMP);
282 		return;
283 	}
284 
285 #if defined(KERN_TLS) && defined(INVARIANTS)
286 	if ((sfio->m->m_flags & M_EXT) != 0 &&
287 	    sfio->m->m_ext.ext_type == EXT_PGS)
288 		KASSERT(sfio->tls == sfio->m->m_ext.ext_pgs->tls,
289 		    ("TLS session mismatch"));
290 	else
291 		KASSERT(sfio->tls == NULL,
292 		    ("non-ext_pgs mbuf with TLS session"));
293 #endif
294 	CURVNET_SET(so->so_vnet);
295 	if (__predict_false(sfio->error)) {
296 		/*
297 		 * I/O operation failed.  The state of data in the socket
298 		 * is now inconsistent, and all what we can do is to tear
299 		 * it down. Protocol abort method would tear down protocol
300 		 * state, free all ready mbufs and detach not ready ones.
301 		 * We will free the mbufs corresponding to this I/O manually.
302 		 *
303 		 * The socket would be marked with EIO and made available
304 		 * for read, so that application receives EIO on next
305 		 * syscall and eventually closes the socket.
306 		 */
307 		so->so_proto->pr_usrreqs->pru_abort(so);
308 		so->so_error = EIO;
309 
310 		mb_free_notready(sfio->m, sfio->npages);
311 #ifdef KERN_TLS
312 	} else if (sfio->tls != NULL && sfio->tls->mode == TCP_TLS_MODE_SW) {
313 		/*
314 		 * I/O operation is complete, but we still need to
315 		 * encrypt.  We cannot do this in the interrupt thread
316 		 * of the disk controller, so forward the mbufs to a
317 		 * different thread.
318 		 *
319 		 * Donate the socket reference from sfio to rather
320 		 * than explicitly invoking soref().
321 		 */
322 		ktls_enqueue(sfio->m, so, sfio->npages);
323 		goto out_with_ref;
324 #endif
325 	} else
326 		(void)(so->so_proto->pr_usrreqs->pru_ready)(so, sfio->m,
327 		    sfio->npages);
328 
329 	SOCK_LOCK(so);
330 	sorele(so);
331 #ifdef KERN_TLS
332 out_with_ref:
333 #endif
334 	CURVNET_RESTORE();
335 	free(sfio, M_TEMP);
336 }
337 
338 /*
339  * Iterate through pages vector and request paging for non-valid pages.
340  */
341 static int
342 sendfile_swapin(vm_object_t obj, struct sf_io *sfio, int *nios, off_t off,
343     off_t len, int npages, int rhpages, int flags)
344 {
345 	vm_page_t *pa = sfio->pa;
346 	int grabbed;
347 
348 	*nios = 0;
349 	flags = (flags & SF_NODISKIO) ? VM_ALLOC_NOWAIT : 0;
350 
351 	/*
352 	 * First grab all the pages and wire them.  Note that we grab
353 	 * only required pages.  Readahead pages are dealt with later.
354 	 */
355 	VM_OBJECT_WLOCK(obj);
356 
357 	grabbed = vm_page_grab_pages(obj, OFF_TO_IDX(off),
358 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED | flags, pa, npages);
359 	if (grabbed < npages) {
360 		for (int i = grabbed; i < npages; i++)
361 			pa[i] = NULL;
362 		npages = grabbed;
363 		rhpages = 0;
364 	}
365 
366 	for (int i = 0; i < npages;) {
367 		int j, a, count, rv;
368 
369 		/* Skip valid pages. */
370 		if (vm_page_is_valid(pa[i], vmoff(i, off) & PAGE_MASK,
371 		    xfsize(i, npages, off, len))) {
372 			vm_page_xunbusy(pa[i]);
373 			SFSTAT_INC(sf_pages_valid);
374 			i++;
375 			continue;
376 		}
377 
378 		/*
379 		 * Next page is invalid.  Check if it belongs to pager.  It
380 		 * may not be there, which is a regular situation for shmem
381 		 * pager.  For vnode pager this happens only in case of
382 		 * a sparse file.
383 		 *
384 		 * Important feature of vm_pager_has_page() is the hint
385 		 * stored in 'a', about how many pages we can pagein after
386 		 * this page in a single I/O.
387 		 */
388 		if (!vm_pager_has_page(obj, OFF_TO_IDX(vmoff(i, off)), NULL,
389 		    &a)) {
390 			pmap_zero_page(pa[i]);
391 			pa[i]->valid = VM_PAGE_BITS_ALL;
392 			MPASS(pa[i]->dirty == 0);
393 			vm_page_xunbusy(pa[i]);
394 			i++;
395 			continue;
396 		}
397 
398 		/*
399 		 * We want to pagein as many pages as possible, limited only
400 		 * by the 'a' hint and actual request.
401 		 */
402 		count = min(a + 1, npages - i);
403 
404 		/*
405 		 * We should not pagein into a valid page, thus we first trim
406 		 * any valid pages off the end of request, and substitute
407 		 * to bogus_page those, that are in the middle.
408 		 */
409 		for (j = i + count - 1; j > i; j--) {
410 			if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
411 			    xfsize(j, npages, off, len))) {
412 				count--;
413 				rhpages = 0;
414 			} else
415 				break;
416 		}
417 		for (j = i + 1; j < i + count - 1; j++)
418 			if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
419 			    xfsize(j, npages, off, len))) {
420 				vm_page_xunbusy(pa[j]);
421 				SFSTAT_INC(sf_pages_valid);
422 				SFSTAT_INC(sf_pages_bogus);
423 				pa[j] = bogus_page;
424 			}
425 
426 		refcount_acquire(&sfio->nios);
427 		VM_OBJECT_WUNLOCK(obj);
428 		rv = vm_pager_get_pages_async(obj, pa + i, count, NULL,
429 		    i + count == npages ? &rhpages : NULL,
430 		    &sendfile_iodone, sfio);
431 		VM_OBJECT_WLOCK(obj);
432 		if (__predict_false(rv != VM_PAGER_OK)) {
433 			/*
434 			 * Perform full pages recovery before returning EIO.
435 			 * Pages from 0 to npages are wired.
436 			 * Pages from i to npages are also busied.
437 			 * Pages from (i + 1) to (i + count - 1) may be
438 			 * substituted to bogus page, and not busied.
439 			 */
440 			for (j = 0; j < npages; j++) {
441 				if (j > i && j < i + count - 1 &&
442 				    pa[j] == bogus_page)
443 					pa[j] = vm_page_lookup(obj,
444 					    OFF_TO_IDX(vmoff(j, off)));
445 				else if (j >= i)
446 					vm_page_xunbusy(pa[j]);
447 				KASSERT(pa[j] != NULL && pa[j] != bogus_page,
448 				    ("%s: page %p[%d] I/O recovery failure",
449 				    __func__, pa, j));
450 				vm_page_unwire(pa[j], PQ_INACTIVE);
451 			}
452 			VM_OBJECT_WUNLOCK(obj);
453 			return (EIO);
454 		}
455 
456 		SFSTAT_INC(sf_iocnt);
457 		SFSTAT_ADD(sf_pages_read, count);
458 		if (i + count == npages)
459 			SFSTAT_ADD(sf_rhpages_read, rhpages);
460 
461 		/*
462 		 * Restore the valid page pointers.  They are already
463 		 * unbusied, but still wired.
464 		 */
465 		for (j = i; j < i + count; j++)
466 			if (pa[j] == bogus_page) {
467 				pa[j] = vm_page_lookup(obj,
468 				    OFF_TO_IDX(vmoff(j, off)));
469 				KASSERT(pa[j], ("%s: page %p[%d] disappeared",
470 				    __func__, pa, j));
471 
472 			}
473 		i += count;
474 		(*nios)++;
475 	}
476 
477 	VM_OBJECT_WUNLOCK(obj);
478 
479 	if (*nios == 0 && npages != 0)
480 		SFSTAT_INC(sf_noiocnt);
481 
482 	return (0);
483 }
484 
485 static int
486 sendfile_getobj(struct thread *td, struct file *fp, vm_object_t *obj_res,
487     struct vnode **vp_res, struct shmfd **shmfd_res, off_t *obj_size,
488     int *bsize)
489 {
490 	struct vattr va;
491 	vm_object_t obj;
492 	struct vnode *vp;
493 	struct shmfd *shmfd;
494 	int error;
495 
496 	vp = *vp_res = NULL;
497 	obj = NULL;
498 	shmfd = *shmfd_res = NULL;
499 	*bsize = 0;
500 
501 	/*
502 	 * The file descriptor must be a regular file and have a
503 	 * backing VM object.
504 	 */
505 	if (fp->f_type == DTYPE_VNODE) {
506 		vp = fp->f_vnode;
507 		vn_lock(vp, LK_SHARED | LK_RETRY);
508 		if (vp->v_type != VREG) {
509 			error = EINVAL;
510 			goto out;
511 		}
512 		*bsize = vp->v_mount->mnt_stat.f_iosize;
513 		error = VOP_GETATTR(vp, &va, td->td_ucred);
514 		if (error != 0)
515 			goto out;
516 		*obj_size = va.va_size;
517 		obj = vp->v_object;
518 		if (obj == NULL) {
519 			error = EINVAL;
520 			goto out;
521 		}
522 	} else if (fp->f_type == DTYPE_SHM) {
523 		error = 0;
524 		shmfd = fp->f_data;
525 		obj = shmfd->shm_object;
526 		*obj_size = shmfd->shm_size;
527 	} else {
528 		error = EINVAL;
529 		goto out;
530 	}
531 
532 	VM_OBJECT_WLOCK(obj);
533 	if ((obj->flags & OBJ_DEAD) != 0) {
534 		VM_OBJECT_WUNLOCK(obj);
535 		error = EBADF;
536 		goto out;
537 	}
538 
539 	/*
540 	 * Temporarily increase the backing VM object's reference
541 	 * count so that a forced reclamation of its vnode does not
542 	 * immediately destroy it.
543 	 */
544 	vm_object_reference_locked(obj);
545 	VM_OBJECT_WUNLOCK(obj);
546 	*obj_res = obj;
547 	*vp_res = vp;
548 	*shmfd_res = shmfd;
549 
550 out:
551 	if (vp != NULL)
552 		VOP_UNLOCK(vp);
553 	return (error);
554 }
555 
556 static int
557 sendfile_getsock(struct thread *td, int s, struct file **sock_fp,
558     struct socket **so)
559 {
560 	int error;
561 
562 	*sock_fp = NULL;
563 	*so = NULL;
564 
565 	/*
566 	 * The socket must be a stream socket and connected.
567 	 */
568 	error = getsock_cap(td, s, &cap_send_rights,
569 	    sock_fp, NULL, NULL);
570 	if (error != 0)
571 		return (error);
572 	*so = (*sock_fp)->f_data;
573 	if ((*so)->so_type != SOCK_STREAM)
574 		return (EINVAL);
575 	if (SOLISTENING(*so))
576 		return (ENOTCONN);
577 	return (0);
578 }
579 
580 int
581 vn_sendfile(struct file *fp, int sockfd, struct uio *hdr_uio,
582     struct uio *trl_uio, off_t offset, size_t nbytes, off_t *sent, int flags,
583     struct thread *td)
584 {
585 	struct file *sock_fp;
586 	struct vnode *vp;
587 	struct vm_object *obj;
588 	struct socket *so;
589 #ifdef KERN_TLS
590 	struct ktls_session *tls;
591 #endif
592 	struct mbuf_ext_pgs *ext_pgs;
593 	struct mbuf *m, *mh, *mhtail;
594 	struct sf_buf *sf;
595 	struct shmfd *shmfd;
596 	struct sendfile_sync *sfs;
597 	struct vattr va;
598 	off_t off, sbytes, rem, obj_size;
599 	int bsize, error, ext_pgs_idx, hdrlen, max_pgs, softerr;
600 #ifdef KERN_TLS
601 	int tls_enq_cnt;
602 #endif
603 	bool use_ext_pgs;
604 
605 	obj = NULL;
606 	so = NULL;
607 	m = mh = NULL;
608 	sfs = NULL;
609 #ifdef KERN_TLS
610 	tls = NULL;
611 #endif
612 	hdrlen = sbytes = 0;
613 	softerr = 0;
614 	use_ext_pgs = false;
615 
616 	error = sendfile_getobj(td, fp, &obj, &vp, &shmfd, &obj_size, &bsize);
617 	if (error != 0)
618 		return (error);
619 
620 	error = sendfile_getsock(td, sockfd, &sock_fp, &so);
621 	if (error != 0)
622 		goto out;
623 
624 #ifdef MAC
625 	error = mac_socket_check_send(td->td_ucred, so);
626 	if (error != 0)
627 		goto out;
628 #endif
629 
630 	SFSTAT_INC(sf_syscalls);
631 	SFSTAT_ADD(sf_rhpages_requested, SF_READAHEAD(flags));
632 
633 	if (flags & SF_SYNC) {
634 		sfs = malloc(sizeof *sfs, M_TEMP, M_WAITOK | M_ZERO);
635 		mtx_init(&sfs->mtx, "sendfile", NULL, MTX_DEF);
636 		cv_init(&sfs->cv, "sendfile");
637 	}
638 
639 	rem = nbytes ? omin(nbytes, obj_size - offset) : obj_size - offset;
640 
641 	/*
642 	 * Protect against multiple writers to the socket.
643 	 *
644 	 * XXXRW: Historically this has assumed non-interruptibility, so now
645 	 * we implement that, but possibly shouldn't.
646 	 */
647 	(void)sblock(&so->so_snd, SBL_WAIT | SBL_NOINTR);
648 #ifdef KERN_TLS
649 	tls = ktls_hold(so->so_snd.sb_tls_info);
650 #endif
651 
652 	/*
653 	 * Loop through the pages of the file, starting with the requested
654 	 * offset. Get a file page (do I/O if necessary), map the file page
655 	 * into an sf_buf, attach an mbuf header to the sf_buf, and queue
656 	 * it on the socket.
657 	 * This is done in two loops.  The inner loop turns as many pages
658 	 * as it can, up to available socket buffer space, without blocking
659 	 * into mbufs to have it bulk delivered into the socket send buffer.
660 	 * The outer loop checks the state and available space of the socket
661 	 * and takes care of the overall progress.
662 	 */
663 	for (off = offset; rem > 0; ) {
664 		struct sf_io *sfio;
665 		vm_page_t *pa;
666 		struct mbuf *mtail;
667 		int nios, space, npages, rhpages;
668 
669 		mtail = NULL;
670 		/*
671 		 * Check the socket state for ongoing connection,
672 		 * no errors and space in socket buffer.
673 		 * If space is low allow for the remainder of the
674 		 * file to be processed if it fits the socket buffer.
675 		 * Otherwise block in waiting for sufficient space
676 		 * to proceed, or if the socket is nonblocking, return
677 		 * to userland with EAGAIN while reporting how far
678 		 * we've come.
679 		 * We wait until the socket buffer has significant free
680 		 * space to do bulk sends.  This makes good use of file
681 		 * system read ahead and allows packet segmentation
682 		 * offloading hardware to take over lots of work.  If
683 		 * we were not careful here we would send off only one
684 		 * sfbuf at a time.
685 		 */
686 		SOCKBUF_LOCK(&so->so_snd);
687 		if (so->so_snd.sb_lowat < so->so_snd.sb_hiwat / 2)
688 			so->so_snd.sb_lowat = so->so_snd.sb_hiwat / 2;
689 retry_space:
690 		if (so->so_snd.sb_state & SBS_CANTSENDMORE) {
691 			error = EPIPE;
692 			SOCKBUF_UNLOCK(&so->so_snd);
693 			goto done;
694 		} else if (so->so_error) {
695 			error = so->so_error;
696 			so->so_error = 0;
697 			SOCKBUF_UNLOCK(&so->so_snd);
698 			goto done;
699 		}
700 		if ((so->so_state & SS_ISCONNECTED) == 0) {
701 			SOCKBUF_UNLOCK(&so->so_snd);
702 			error = ENOTCONN;
703 			goto done;
704 		}
705 
706 		space = sbspace(&so->so_snd);
707 		if (space < rem &&
708 		    (space <= 0 ||
709 		     space < so->so_snd.sb_lowat)) {
710 			if (so->so_state & SS_NBIO) {
711 				SOCKBUF_UNLOCK(&so->so_snd);
712 				error = EAGAIN;
713 				goto done;
714 			}
715 			/*
716 			 * sbwait drops the lock while sleeping.
717 			 * When we loop back to retry_space the
718 			 * state may have changed and we retest
719 			 * for it.
720 			 */
721 			error = sbwait(&so->so_snd);
722 			/*
723 			 * An error from sbwait usually indicates that we've
724 			 * been interrupted by a signal. If we've sent anything
725 			 * then return bytes sent, otherwise return the error.
726 			 */
727 			if (error != 0) {
728 				SOCKBUF_UNLOCK(&so->so_snd);
729 				goto done;
730 			}
731 			goto retry_space;
732 		}
733 		SOCKBUF_UNLOCK(&so->so_snd);
734 
735 		/*
736 		 * At the beginning of the first loop check if any headers
737 		 * are specified and copy them into mbufs.  Reduce space in
738 		 * the socket buffer by the size of the header mbuf chain.
739 		 * Clear hdr_uio here and hdrlen at the end of the first loop.
740 		 */
741 		if (hdr_uio != NULL && hdr_uio->uio_resid > 0) {
742 			hdr_uio->uio_td = td;
743 			hdr_uio->uio_rw = UIO_WRITE;
744 #ifdef KERN_TLS
745 			if (tls != NULL)
746 				mh = m_uiotombuf(hdr_uio, M_WAITOK, space,
747 				    tls->params.max_frame_len, M_NOMAP);
748 			else
749 #endif
750 				mh = m_uiotombuf(hdr_uio, M_WAITOK,
751 				    space, 0, 0);
752 			hdrlen = m_length(mh, &mhtail);
753 			space -= hdrlen;
754 			/*
755 			 * If header consumed all the socket buffer space,
756 			 * don't waste CPU cycles and jump to the end.
757 			 */
758 			if (space == 0) {
759 				sfio = NULL;
760 				nios = 0;
761 				goto prepend_header;
762 			}
763 			hdr_uio = NULL;
764 		}
765 
766 		if (vp != NULL) {
767 			error = vn_lock(vp, LK_SHARED);
768 			if (error != 0)
769 				goto done;
770 			error = VOP_GETATTR(vp, &va, td->td_ucred);
771 			if (error != 0 || off >= va.va_size) {
772 				VOP_UNLOCK(vp);
773 				goto done;
774 			}
775 			if (va.va_size != obj_size) {
776 				obj_size = va.va_size;
777 				rem = nbytes ?
778 				    omin(nbytes + offset, obj_size) : obj_size;
779 				rem -= off;
780 			}
781 		}
782 
783 		if (space > rem)
784 			space = rem;
785 		else if (space > PAGE_SIZE) {
786 			/*
787 			 * Use page boundaries when possible for large
788 			 * requests.
789 			 */
790 			if (off & PAGE_MASK)
791 				space -= (PAGE_SIZE - (off & PAGE_MASK));
792 			space = trunc_page(space);
793 			if (off & PAGE_MASK)
794 				space += (PAGE_SIZE - (off & PAGE_MASK));
795 		}
796 
797 		npages = howmany(space + (off & PAGE_MASK), PAGE_SIZE);
798 
799 		/*
800 		 * Calculate maximum allowed number of pages for readahead
801 		 * at this iteration.  If SF_USER_READAHEAD was set, we don't
802 		 * do any heuristics and use exactly the value supplied by
803 		 * application.  Otherwise, we allow readahead up to "rem".
804 		 * If application wants more, let it be, but there is no
805 		 * reason to go above MAXPHYS.  Also check against "obj_size",
806 		 * since vm_pager_has_page() can hint beyond EOF.
807 		 */
808 		if (flags & SF_USER_READAHEAD) {
809 			rhpages = SF_READAHEAD(flags);
810 		} else {
811 			rhpages = howmany(rem + (off & PAGE_MASK), PAGE_SIZE) -
812 			    npages;
813 			rhpages += SF_READAHEAD(flags);
814 		}
815 		rhpages = min(howmany(MAXPHYS, PAGE_SIZE), rhpages);
816 		rhpages = min(howmany(obj_size - trunc_page(off), PAGE_SIZE) -
817 		    npages, rhpages);
818 
819 		sfio = malloc(sizeof(struct sf_io) +
820 		    npages * sizeof(vm_page_t), M_TEMP, M_WAITOK);
821 		refcount_init(&sfio->nios, 1);
822 		sfio->so = so;
823 		sfio->obj = obj;
824 		sfio->error = 0;
825 		vm_object_pip_add(obj, 1);
826 
827 #ifdef KERN_TLS
828 		/*
829 		 * This doesn't use ktls_hold() because sfio->m will
830 		 * also have a reference on 'tls' that will be valid
831 		 * for all of sfio's lifetime.
832 		 */
833 		sfio->tls = tls;
834 #endif
835 
836 		error = sendfile_swapin(obj, sfio, &nios, off, space, npages,
837 		    rhpages, flags);
838 		if (error != 0) {
839 			if (vp != NULL)
840 				VOP_UNLOCK(vp);
841 			sfio->m = NULL;
842 			sendfile_iodone(sfio, NULL, 0, error);
843 			goto done;
844 		}
845 
846 		/*
847 		 * Loop and construct maximum sized mbuf chain to be bulk
848 		 * dumped into socket buffer.
849 		 */
850 		pa = sfio->pa;
851 
852 		/*
853 		 * Use unmapped mbufs if enabled for TCP.  Unmapped
854 		 * bufs are restricted to TCP as that is what has been
855 		 * tested.  In particular, unmapped mbufs have not
856 		 * been tested with UNIX-domain sockets.
857 		 *
858 		 * TLS frames always require unmapped mbufs.
859 		 */
860 		if ((mb_use_ext_pgs &&
861 		    so->so_proto->pr_protocol == IPPROTO_TCP)
862 #ifdef KERN_TLS
863 		    || tls != NULL
864 #endif
865 		    ) {
866 			use_ext_pgs = true;
867 #ifdef KERN_TLS
868 			if (tls != NULL)
869 				max_pgs = num_pages(tls->params.max_frame_len);
870 			else
871 #endif
872 				max_pgs = MBUF_PEXT_MAX_PGS;
873 
874 			/* Start at last index, to wrap on first use. */
875 			ext_pgs_idx = max_pgs - 1;
876 		}
877 
878 		for (int i = 0; i < npages; i++) {
879 			struct mbuf *m0;
880 
881 			/*
882 			 * If a page wasn't grabbed successfully, then
883 			 * trim the array. Can happen only with SF_NODISKIO.
884 			 */
885 			if (pa[i] == NULL) {
886 				SFSTAT_INC(sf_busy);
887 				fixspace(npages, i, off, &space);
888 				npages = i;
889 				softerr = EBUSY;
890 				break;
891 			}
892 
893 			if (use_ext_pgs) {
894 				off_t xfs;
895 
896 				ext_pgs_idx++;
897 				if (ext_pgs_idx == max_pgs) {
898 					m0 = mb_alloc_ext_pgs(M_WAITOK, false,
899 					    sendfile_free_mext_pg);
900 
901 					if (flags & SF_NOCACHE) {
902 						m0->m_ext.ext_flags |=
903 						    EXT_FLAG_NOCACHE;
904 
905 						/*
906 						 * See comment below regarding
907 						 * ignoring SF_NOCACHE for the
908 						 * last page.
909 						 */
910 						if ((npages - i <= max_pgs) &&
911 						    ((off + space) & PAGE_MASK) &&
912 						    (rem > space || rhpages > 0))
913 							m0->m_ext.ext_flags |=
914 							    EXT_FLAG_CACHE_LAST;
915 					}
916 					if (sfs != NULL) {
917 						m0->m_ext.ext_flags |=
918 						    EXT_FLAG_SYNC;
919 						m0->m_ext.ext_arg2 = sfs;
920 						mtx_lock(&sfs->mtx);
921 						sfs->count++;
922 						mtx_unlock(&sfs->mtx);
923 					}
924 					ext_pgs = m0->m_ext.ext_pgs;
925 					if (i == 0)
926 						sfio->m = m0;
927 					ext_pgs_idx = 0;
928 
929 					/* Append to mbuf chain. */
930 					if (mtail != NULL)
931 						mtail->m_next = m0;
932 					else
933 						m = m0;
934 					mtail = m0;
935 					ext_pgs->first_pg_off =
936 					    vmoff(i, off) & PAGE_MASK;
937 				}
938 				if (nios) {
939 					mtail->m_flags |= M_NOTREADY;
940 					ext_pgs->nrdy++;
941 				}
942 
943 				ext_pgs->pa[ext_pgs_idx] = VM_PAGE_TO_PHYS(pa[i]);
944 				ext_pgs->npgs++;
945 				xfs = xfsize(i, npages, off, space);
946 				ext_pgs->last_pg_len = xfs;
947 				MBUF_EXT_PGS_ASSERT_SANITY(ext_pgs);
948 				mtail->m_len += xfs;
949 				mtail->m_ext.ext_size += PAGE_SIZE;
950 				continue;
951 			}
952 
953 			/*
954 			 * Get a sendfile buf.  When allocating the
955 			 * first buffer for mbuf chain, we usually
956 			 * wait as long as necessary, but this wait
957 			 * can be interrupted.  For consequent
958 			 * buffers, do not sleep, since several
959 			 * threads might exhaust the buffers and then
960 			 * deadlock.
961 			 */
962 			sf = sf_buf_alloc(pa[i],
963 			    m != NULL ? SFB_NOWAIT : SFB_CATCH);
964 			if (sf == NULL) {
965 				SFSTAT_INC(sf_allocfail);
966 				for (int j = i; j < npages; j++)
967 					vm_page_unwire(pa[j], PQ_INACTIVE);
968 				if (m == NULL)
969 					softerr = ENOBUFS;
970 				fixspace(npages, i, off, &space);
971 				npages = i;
972 				break;
973 			}
974 
975 			m0 = m_get(M_WAITOK, MT_DATA);
976 			m0->m_ext.ext_buf = (char *)sf_buf_kva(sf);
977 			m0->m_ext.ext_size = PAGE_SIZE;
978 			m0->m_ext.ext_arg1 = sf;
979 			m0->m_ext.ext_type = EXT_SFBUF;
980 			m0->m_ext.ext_flags = EXT_FLAG_EMBREF;
981 			m0->m_ext.ext_free = sendfile_free_mext;
982 			/*
983 			 * SF_NOCACHE sets the page as being freed upon send.
984 			 * However, we ignore it for the last page in 'space',
985 			 * if the page is truncated, and we got more data to
986 			 * send (rem > space), or if we have readahead
987 			 * configured (rhpages > 0).
988 			 */
989 			if ((flags & SF_NOCACHE) &&
990 			    (i != npages - 1 ||
991 			    !((off + space) & PAGE_MASK) ||
992 			    !(rem > space || rhpages > 0)))
993 				m0->m_ext.ext_flags |= EXT_FLAG_NOCACHE;
994 			if (sfs != NULL) {
995 				m0->m_ext.ext_flags |= EXT_FLAG_SYNC;
996 				m0->m_ext.ext_arg2 = sfs;
997 				mtx_lock(&sfs->mtx);
998 				sfs->count++;
999 				mtx_unlock(&sfs->mtx);
1000 			}
1001 			m0->m_ext.ext_count = 1;
1002 			m0->m_flags |= (M_EXT | M_RDONLY);
1003 			if (nios)
1004 				m0->m_flags |= M_NOTREADY;
1005 			m0->m_data = (char *)sf_buf_kva(sf) +
1006 			    (vmoff(i, off) & PAGE_MASK);
1007 			m0->m_len = xfsize(i, npages, off, space);
1008 
1009 			if (i == 0)
1010 				sfio->m = m0;
1011 
1012 			/* Append to mbuf chain. */
1013 			if (mtail != NULL)
1014 				mtail->m_next = m0;
1015 			else
1016 				m = m0;
1017 			mtail = m0;
1018 		}
1019 
1020 		if (vp != NULL)
1021 			VOP_UNLOCK(vp);
1022 
1023 		/* Keep track of bytes processed. */
1024 		off += space;
1025 		rem -= space;
1026 
1027 		/* Prepend header, if any. */
1028 		if (hdrlen) {
1029 prepend_header:
1030 			mhtail->m_next = m;
1031 			m = mh;
1032 			mh = NULL;
1033 		}
1034 
1035 		if (m == NULL) {
1036 			KASSERT(softerr, ("%s: m NULL, no error", __func__));
1037 			error = softerr;
1038 			free(sfio, M_TEMP);
1039 			goto done;
1040 		}
1041 
1042 		/* Add the buffer chain to the socket buffer. */
1043 		KASSERT(m_length(m, NULL) == space + hdrlen,
1044 		    ("%s: mlen %u space %d hdrlen %d",
1045 		    __func__, m_length(m, NULL), space, hdrlen));
1046 
1047 		CURVNET_SET(so->so_vnet);
1048 #ifdef KERN_TLS
1049 		if (tls != NULL) {
1050 			error = ktls_frame(m, tls, &tls_enq_cnt,
1051 			    TLS_RLTYPE_APP);
1052 			if (error != 0)
1053 				goto done;
1054 		}
1055 #endif
1056 		if (nios == 0) {
1057 			/*
1058 			 * If sendfile_swapin() didn't initiate any I/Os,
1059 			 * which happens if all data is cached in VM, then
1060 			 * we can send data right now without the
1061 			 * PRUS_NOTREADY flag.
1062 			 */
1063 			vm_object_pip_wakeup(sfio->obj);
1064 			free(sfio, M_TEMP);
1065 #ifdef KERN_TLS
1066 			if (tls != NULL && tls->mode == TCP_TLS_MODE_SW) {
1067 				error = (*so->so_proto->pr_usrreqs->pru_send)
1068 				    (so, PRUS_NOTREADY, m, NULL, NULL, td);
1069 				soref(so);
1070 				ktls_enqueue(m, so, tls_enq_cnt);
1071 			} else
1072 #endif
1073 				error = (*so->so_proto->pr_usrreqs->pru_send)
1074 				    (so, 0, m, NULL, NULL, td);
1075 		} else {
1076 			sfio->npages = npages;
1077 			soref(so);
1078 			error = (*so->so_proto->pr_usrreqs->pru_send)
1079 			    (so, PRUS_NOTREADY, m, NULL, NULL, td);
1080 			sendfile_iodone(sfio, NULL, 0, 0);
1081 		}
1082 		CURVNET_RESTORE();
1083 
1084 		m = NULL;	/* pru_send always consumes */
1085 		if (error)
1086 			goto done;
1087 		sbytes += space + hdrlen;
1088 		if (hdrlen)
1089 			hdrlen = 0;
1090 		if (softerr) {
1091 			error = softerr;
1092 			goto done;
1093 		}
1094 	}
1095 
1096 	/*
1097 	 * Send trailers. Wimp out and use writev(2).
1098 	 */
1099 	if (trl_uio != NULL) {
1100 		sbunlock(&so->so_snd);
1101 		error = kern_writev(td, sockfd, trl_uio);
1102 		if (error == 0)
1103 			sbytes += td->td_retval[0];
1104 		goto out;
1105 	}
1106 
1107 done:
1108 	sbunlock(&so->so_snd);
1109 out:
1110 	/*
1111 	 * If there was no error we have to clear td->td_retval[0]
1112 	 * because it may have been set by writev.
1113 	 */
1114 	if (error == 0) {
1115 		td->td_retval[0] = 0;
1116 	}
1117 	if (sent != NULL) {
1118 		(*sent) = sbytes;
1119 	}
1120 	if (obj != NULL)
1121 		vm_object_deallocate(obj);
1122 	if (so)
1123 		fdrop(sock_fp, td);
1124 	if (m)
1125 		m_freem(m);
1126 	if (mh)
1127 		m_freem(mh);
1128 
1129 	if (sfs != NULL) {
1130 		mtx_lock(&sfs->mtx);
1131 		if (sfs->count != 0)
1132 			cv_wait(&sfs->cv, &sfs->mtx);
1133 		KASSERT(sfs->count == 0, ("sendfile sync still busy"));
1134 		cv_destroy(&sfs->cv);
1135 		mtx_destroy(&sfs->mtx);
1136 		free(sfs, M_TEMP);
1137 	}
1138 #ifdef KERN_TLS
1139 	if (tls != NULL)
1140 		ktls_free(tls);
1141 #endif
1142 
1143 	if (error == ERESTART)
1144 		error = EINTR;
1145 
1146 	return (error);
1147 }
1148 
1149 static int
1150 sendfile(struct thread *td, struct sendfile_args *uap, int compat)
1151 {
1152 	struct sf_hdtr hdtr;
1153 	struct uio *hdr_uio, *trl_uio;
1154 	struct file *fp;
1155 	off_t sbytes;
1156 	int error;
1157 
1158 	/*
1159 	 * File offset must be positive.  If it goes beyond EOF
1160 	 * we send only the header/trailer and no payload data.
1161 	 */
1162 	if (uap->offset < 0)
1163 		return (EINVAL);
1164 
1165 	sbytes = 0;
1166 	hdr_uio = trl_uio = NULL;
1167 
1168 	if (uap->hdtr != NULL) {
1169 		error = copyin(uap->hdtr, &hdtr, sizeof(hdtr));
1170 		if (error != 0)
1171 			goto out;
1172 		if (hdtr.headers != NULL) {
1173 			error = copyinuio(hdtr.headers, hdtr.hdr_cnt,
1174 			    &hdr_uio);
1175 			if (error != 0)
1176 				goto out;
1177 #ifdef COMPAT_FREEBSD4
1178 			/*
1179 			 * In FreeBSD < 5.0 the nbytes to send also included
1180 			 * the header.  If compat is specified subtract the
1181 			 * header size from nbytes.
1182 			 */
1183 			if (compat) {
1184 				if (uap->nbytes > hdr_uio->uio_resid)
1185 					uap->nbytes -= hdr_uio->uio_resid;
1186 				else
1187 					uap->nbytes = 0;
1188 			}
1189 #endif
1190 		}
1191 		if (hdtr.trailers != NULL) {
1192 			error = copyinuio(hdtr.trailers, hdtr.trl_cnt,
1193 			    &trl_uio);
1194 			if (error != 0)
1195 				goto out;
1196 		}
1197 	}
1198 
1199 	AUDIT_ARG_FD(uap->fd);
1200 
1201 	/*
1202 	 * sendfile(2) can start at any offset within a file so we require
1203 	 * CAP_READ+CAP_SEEK = CAP_PREAD.
1204 	 */
1205 	if ((error = fget_read(td, uap->fd, &cap_pread_rights, &fp)) != 0)
1206 		goto out;
1207 
1208 	error = fo_sendfile(fp, uap->s, hdr_uio, trl_uio, uap->offset,
1209 	    uap->nbytes, &sbytes, uap->flags, td);
1210 	fdrop(fp, td);
1211 
1212 	if (uap->sbytes != NULL)
1213 		copyout(&sbytes, uap->sbytes, sizeof(off_t));
1214 
1215 out:
1216 	free(hdr_uio, M_IOV);
1217 	free(trl_uio, M_IOV);
1218 	return (error);
1219 }
1220 
1221 /*
1222  * sendfile(2)
1223  *
1224  * int sendfile(int fd, int s, off_t offset, size_t nbytes,
1225  *       struct sf_hdtr *hdtr, off_t *sbytes, int flags)
1226  *
1227  * Send a file specified by 'fd' and starting at 'offset' to a socket
1228  * specified by 's'. Send only 'nbytes' of the file or until EOF if nbytes ==
1229  * 0.  Optionally add a header and/or trailer to the socket output.  If
1230  * specified, write the total number of bytes sent into *sbytes.
1231  */
1232 int
1233 sys_sendfile(struct thread *td, struct sendfile_args *uap)
1234 {
1235 
1236 	return (sendfile(td, uap, 0));
1237 }
1238 
1239 #ifdef COMPAT_FREEBSD4
1240 int
1241 freebsd4_sendfile(struct thread *td, struct freebsd4_sendfile_args *uap)
1242 {
1243 	struct sendfile_args args;
1244 
1245 	args.fd = uap->fd;
1246 	args.s = uap->s;
1247 	args.offset = uap->offset;
1248 	args.nbytes = uap->nbytes;
1249 	args.hdtr = uap->hdtr;
1250 	args.sbytes = uap->sbytes;
1251 	args.flags = uap->flags;
1252 
1253 	return (sendfile(td, &args, 1));
1254 }
1255 #endif /* COMPAT_FREEBSD4 */
1256