xref: /freebsd/sys/kern/kern_sendfile.c (revision 0b57cec536236d46e3dba9bd041533462f33dbb7)
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 #ifdef KERN_TLS
93 	struct ktls_session *tls;
94 #endif
95 	vm_page_t	pa[];
96 };
97 
98 /*
99  * Structure used to track requests with SF_SYNC flag.
100  */
101 struct sendfile_sync {
102 	struct mtx	mtx;
103 	struct cv	cv;
104 	unsigned	count;
105 };
106 
107 counter_u64_t sfstat[sizeof(struct sfstat) / sizeof(uint64_t)];
108 
109 static void
110 sfstat_init(const void *unused)
111 {
112 
113 	COUNTER_ARRAY_ALLOC(sfstat, sizeof(struct sfstat) / sizeof(uint64_t),
114 	    M_WAITOK);
115 }
116 SYSINIT(sfstat, SI_SUB_MBUF, SI_ORDER_FIRST, sfstat_init, NULL);
117 
118 static int
119 sfstat_sysctl(SYSCTL_HANDLER_ARGS)
120 {
121 	struct sfstat s;
122 
123 	COUNTER_ARRAY_COPY(sfstat, &s, sizeof(s) / sizeof(uint64_t));
124 	if (req->newptr)
125 		COUNTER_ARRAY_ZERO(sfstat, sizeof(s) / sizeof(uint64_t));
126 	return (SYSCTL_OUT(req, &s, sizeof(s)));
127 }
128 SYSCTL_PROC(_kern_ipc, OID_AUTO, sfstat, CTLTYPE_OPAQUE | CTLFLAG_RW,
129     NULL, 0, sfstat_sysctl, "I", "sendfile statistics");
130 
131 static void
132 sendfile_free_mext(struct mbuf *m)
133 {
134 	struct sf_buf *sf;
135 	vm_page_t pg;
136 	int flags;
137 
138 	KASSERT(m->m_flags & M_EXT && m->m_ext.ext_type == EXT_SFBUF,
139 	    ("%s: m %p !M_EXT or !EXT_SFBUF", __func__, m));
140 
141 	sf = m->m_ext.ext_arg1;
142 	pg = sf_buf_page(sf);
143 	flags = (m->m_ext.ext_flags & EXT_FLAG_NOCACHE) != 0 ? VPR_TRYFREE : 0;
144 
145 	sf_buf_free(sf);
146 	vm_page_release(pg, flags);
147 
148 	if (m->m_ext.ext_flags & EXT_FLAG_SYNC) {
149 		struct sendfile_sync *sfs = m->m_ext.ext_arg2;
150 
151 		mtx_lock(&sfs->mtx);
152 		KASSERT(sfs->count > 0, ("Sendfile sync botchup count == 0"));
153 		if (--sfs->count == 0)
154 			cv_signal(&sfs->cv);
155 		mtx_unlock(&sfs->mtx);
156 	}
157 }
158 
159 static void
160 sendfile_free_mext_pg(struct mbuf *m)
161 {
162 	struct mbuf_ext_pgs *ext_pgs;
163 	vm_page_t pg;
164 	int flags, i;
165 	bool cache_last;
166 
167 	KASSERT(m->m_flags & M_EXT && m->m_ext.ext_type == EXT_PGS,
168 	    ("%s: m %p !M_EXT or !EXT_PGS", __func__, m));
169 
170 	cache_last = m->m_ext.ext_flags & EXT_FLAG_CACHE_LAST;
171 	ext_pgs = m->m_ext.ext_pgs;
172 	flags = (m->m_ext.ext_flags & EXT_FLAG_NOCACHE) != 0 ? VPR_TRYFREE : 0;
173 
174 	for (i = 0; i < ext_pgs->npgs; i++) {
175 		if (cache_last && i == ext_pgs->npgs - 1)
176 			flags = 0;
177 		pg = PHYS_TO_VM_PAGE(ext_pgs->pa[i]);
178 		vm_page_release(pg, flags);
179 	}
180 
181 	if (m->m_ext.ext_flags & EXT_FLAG_SYNC) {
182 		struct sendfile_sync *sfs = m->m_ext.ext_arg2;
183 
184 		mtx_lock(&sfs->mtx);
185 		KASSERT(sfs->count > 0, ("Sendfile sync botchup count == 0"));
186 		if (--sfs->count == 0)
187 			cv_signal(&sfs->cv);
188 		mtx_unlock(&sfs->mtx);
189 	}
190 }
191 
192 /*
193  * Helper function to calculate how much data to put into page i of n.
194  * Only first and last pages are special.
195  */
196 static inline off_t
197 xfsize(int i, int n, off_t off, off_t len)
198 {
199 
200 	if (i == 0)
201 		return (omin(PAGE_SIZE - (off & PAGE_MASK), len));
202 
203 	if (i == n - 1 && ((off + len) & PAGE_MASK) > 0)
204 		return ((off + len) & PAGE_MASK);
205 
206 	return (PAGE_SIZE);
207 }
208 
209 /*
210  * Helper function to get offset within object for i page.
211  */
212 static inline vm_ooffset_t
213 vmoff(int i, off_t off)
214 {
215 
216 	if (i == 0)
217 		return ((vm_ooffset_t)off);
218 
219 	return (trunc_page(off + i * PAGE_SIZE));
220 }
221 
222 /*
223  * Helper function used when allocation of a page or sf_buf failed.
224  * Pretend as if we don't have enough space, subtract xfsize() of
225  * all pages that failed.
226  */
227 static inline void
228 fixspace(int old, int new, off_t off, int *space)
229 {
230 
231 	KASSERT(old > new, ("%s: old %d new %d", __func__, old, new));
232 
233 	/* Subtract last one. */
234 	*space -= xfsize(old - 1, old, off, *space);
235 	old--;
236 
237 	if (new == old)
238 		/* There was only one page. */
239 		return;
240 
241 	/* Subtract first one. */
242 	if (new == 0) {
243 		*space -= xfsize(0, old, off, *space);
244 		new++;
245 	}
246 
247 	/* Rest of pages are full sized. */
248 	*space -= (old - new) * PAGE_SIZE;
249 
250 	KASSERT(*space >= 0, ("%s: space went backwards", __func__));
251 }
252 
253 /*
254  * I/O completion callback.
255  */
256 static void
257 sendfile_iodone(void *arg, vm_page_t *pg, int count, int error)
258 {
259 	struct sf_io *sfio = arg;
260 	struct socket *so = sfio->so;
261 
262 	for (int i = 0; i < count; i++)
263 		if (pg[i] != bogus_page)
264 			vm_page_xunbusy_unchecked(pg[i]);
265 
266 	if (error)
267 		sfio->error = error;
268 
269 	if (!refcount_release(&sfio->nios))
270 		return;
271 
272 	if (__predict_false(sfio->error && sfio->m == NULL)) {
273 		/*
274 		 * I/O operation failed, but pru_send hadn't been executed -
275 		 * nothing had been sent to the socket.  The syscall has
276 		 * returned error to the user.
277 		 */
278 		free(sfio, M_TEMP);
279 		return;
280 	}
281 
282 #if defined(KERN_TLS) && defined(INVARIANTS)
283 	if ((sfio->m->m_flags & M_EXT) != 0 &&
284 	    sfio->m->m_ext.ext_type == EXT_PGS)
285 		KASSERT(sfio->tls == sfio->m->m_ext.ext_pgs->tls,
286 		    ("TLS session mismatch"));
287 	else
288 		KASSERT(sfio->tls == NULL,
289 		    ("non-ext_pgs mbuf with TLS session"));
290 #endif
291 	CURVNET_SET(so->so_vnet);
292 	if (__predict_false(sfio->error)) {
293 		/*
294 		 * I/O operation failed.  The state of data in the socket
295 		 * is now inconsistent, and all what we can do is to tear
296 		 * it down. Protocol abort method would tear down protocol
297 		 * state, free all ready mbufs and detach not ready ones.
298 		 * We will free the mbufs corresponding to this I/O manually.
299 		 *
300 		 * The socket would be marked with EIO and made available
301 		 * for read, so that application receives EIO on next
302 		 * syscall and eventually closes the socket.
303 		 */
304 		so->so_proto->pr_usrreqs->pru_abort(so);
305 		so->so_error = EIO;
306 
307 		mb_free_notready(sfio->m, sfio->npages);
308 #ifdef KERN_TLS
309 	} else if (sfio->tls != NULL && sfio->tls->mode == TCP_TLS_MODE_SW) {
310 		/*
311 		 * I/O operation is complete, but we still need to
312 		 * encrypt.  We cannot do this in the interrupt thread
313 		 * of the disk controller, so forward the mbufs to a
314 		 * different thread.
315 		 *
316 		 * Donate the socket reference from sfio to rather
317 		 * than explicitly invoking soref().
318 		 */
319 		ktls_enqueue(sfio->m, so, sfio->npages);
320 		goto out_with_ref;
321 #endif
322 	} else
323 		(void)(so->so_proto->pr_usrreqs->pru_ready)(so, sfio->m,
324 		    sfio->npages);
325 
326 	SOCK_LOCK(so);
327 	sorele(so);
328 #ifdef KERN_TLS
329 out_with_ref:
330 #endif
331 	CURVNET_RESTORE();
332 	free(sfio, M_TEMP);
333 }
334 
335 /*
336  * Iterate through pages vector and request paging for non-valid pages.
337  */
338 static int
339 sendfile_swapin(vm_object_t obj, struct sf_io *sfio, int *nios, off_t off,
340     off_t len, int npages, int rhpages, int flags)
341 {
342 	vm_page_t *pa = sfio->pa;
343 	int grabbed;
344 
345 	*nios = 0;
346 	flags = (flags & SF_NODISKIO) ? VM_ALLOC_NOWAIT : 0;
347 
348 	/*
349 	 * First grab all the pages and wire them.  Note that we grab
350 	 * only required pages.  Readahead pages are dealt with later.
351 	 */
352 	VM_OBJECT_WLOCK(obj);
353 
354 	grabbed = vm_page_grab_pages(obj, OFF_TO_IDX(off),
355 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED | flags, pa, npages);
356 	if (grabbed < npages) {
357 		for (int i = grabbed; i < npages; i++)
358 			pa[i] = NULL;
359 		npages = grabbed;
360 		rhpages = 0;
361 	}
362 
363 	for (int i = 0; i < npages;) {
364 		int j, a, count, rv;
365 
366 		/* Skip valid pages. */
367 		if (vm_page_is_valid(pa[i], vmoff(i, off) & PAGE_MASK,
368 		    xfsize(i, npages, off, len))) {
369 			vm_page_xunbusy(pa[i]);
370 			SFSTAT_INC(sf_pages_valid);
371 			i++;
372 			continue;
373 		}
374 
375 		/*
376 		 * Next page is invalid.  Check if it belongs to pager.  It
377 		 * may not be there, which is a regular situation for shmem
378 		 * pager.  For vnode pager this happens only in case of
379 		 * a sparse file.
380 		 *
381 		 * Important feature of vm_pager_has_page() is the hint
382 		 * stored in 'a', about how many pages we can pagein after
383 		 * this page in a single I/O.
384 		 */
385 		if (!vm_pager_has_page(obj, OFF_TO_IDX(vmoff(i, off)), NULL,
386 		    &a)) {
387 			pmap_zero_page(pa[i]);
388 			pa[i]->valid = VM_PAGE_BITS_ALL;
389 			MPASS(pa[i]->dirty == 0);
390 			vm_page_xunbusy(pa[i]);
391 			i++;
392 			continue;
393 		}
394 
395 		/*
396 		 * We want to pagein as many pages as possible, limited only
397 		 * by the 'a' hint and actual request.
398 		 */
399 		count = min(a + 1, npages - i);
400 
401 		/*
402 		 * We should not pagein into a valid page, thus we first trim
403 		 * any valid pages off the end of request, and substitute
404 		 * to bogus_page those, that are in the middle.
405 		 */
406 		for (j = i + count - 1; j > i; j--) {
407 			if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
408 			    xfsize(j, npages, off, len))) {
409 				count--;
410 				rhpages = 0;
411 			} else
412 				break;
413 		}
414 		for (j = i + 1; j < i + count - 1; j++)
415 			if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
416 			    xfsize(j, npages, off, len))) {
417 				vm_page_xunbusy(pa[j]);
418 				SFSTAT_INC(sf_pages_valid);
419 				SFSTAT_INC(sf_pages_bogus);
420 				pa[j] = bogus_page;
421 			}
422 
423 		refcount_acquire(&sfio->nios);
424 		rv = vm_pager_get_pages_async(obj, pa + i, count, NULL,
425 		    i + count == npages ? &rhpages : NULL,
426 		    &sendfile_iodone, sfio);
427 		if (__predict_false(rv != VM_PAGER_OK)) {
428 			/*
429 			 * Perform full pages recovery before returning EIO.
430 			 * Pages from 0 to npages are wired.
431 			 * Pages from i to npages are also busied.
432 			 * Pages from (i + 1) to (i + count - 1) may be
433 			 * substituted to bogus page, and not busied.
434 			 */
435 			for (j = 0; j < npages; j++) {
436 				if (j > i && j < i + count - 1 &&
437 				    pa[j] == bogus_page)
438 					pa[j] = vm_page_lookup(obj,
439 					    OFF_TO_IDX(vmoff(j, off)));
440 				else if (j >= i)
441 					vm_page_xunbusy(pa[j]);
442 				KASSERT(pa[j] != NULL && pa[j] != bogus_page,
443 				    ("%s: page %p[%d] I/O recovery failure",
444 				    __func__, pa, j));
445 				vm_page_unwire(pa[j], PQ_INACTIVE);
446 			}
447 			VM_OBJECT_WUNLOCK(obj);
448 			return (EIO);
449 		}
450 
451 		SFSTAT_INC(sf_iocnt);
452 		SFSTAT_ADD(sf_pages_read, count);
453 		if (i + count == npages)
454 			SFSTAT_ADD(sf_rhpages_read, rhpages);
455 
456 		/*
457 		 * Restore the valid page pointers.  They are already
458 		 * unbusied, but still wired.
459 		 */
460 		for (j = i; j < i + count; j++)
461 			if (pa[j] == bogus_page) {
462 				pa[j] = vm_page_lookup(obj,
463 				    OFF_TO_IDX(vmoff(j, off)));
464 				KASSERT(pa[j], ("%s: page %p[%d] disappeared",
465 				    __func__, pa, j));
466 
467 			}
468 		i += count;
469 		(*nios)++;
470 	}
471 
472 	VM_OBJECT_WUNLOCK(obj);
473 
474 	if (*nios == 0 && npages != 0)
475 		SFSTAT_INC(sf_noiocnt);
476 
477 	return (0);
478 }
479 
480 static int
481 sendfile_getobj(struct thread *td, struct file *fp, vm_object_t *obj_res,
482     struct vnode **vp_res, struct shmfd **shmfd_res, off_t *obj_size,
483     int *bsize)
484 {
485 	struct vattr va;
486 	vm_object_t obj;
487 	struct vnode *vp;
488 	struct shmfd *shmfd;
489 	int error;
490 
491 	vp = *vp_res = NULL;
492 	obj = NULL;
493 	shmfd = *shmfd_res = NULL;
494 	*bsize = 0;
495 
496 	/*
497 	 * The file descriptor must be a regular file and have a
498 	 * backing VM object.
499 	 */
500 	if (fp->f_type == DTYPE_VNODE) {
501 		vp = fp->f_vnode;
502 		vn_lock(vp, LK_SHARED | LK_RETRY);
503 		if (vp->v_type != VREG) {
504 			error = EINVAL;
505 			goto out;
506 		}
507 		*bsize = vp->v_mount->mnt_stat.f_iosize;
508 		error = VOP_GETATTR(vp, &va, td->td_ucred);
509 		if (error != 0)
510 			goto out;
511 		*obj_size = va.va_size;
512 		obj = vp->v_object;
513 		if (obj == NULL) {
514 			error = EINVAL;
515 			goto out;
516 		}
517 	} else if (fp->f_type == DTYPE_SHM) {
518 		error = 0;
519 		shmfd = fp->f_data;
520 		obj = shmfd->shm_object;
521 		*obj_size = shmfd->shm_size;
522 	} else {
523 		error = EINVAL;
524 		goto out;
525 	}
526 
527 	VM_OBJECT_WLOCK(obj);
528 	if ((obj->flags & OBJ_DEAD) != 0) {
529 		VM_OBJECT_WUNLOCK(obj);
530 		error = EBADF;
531 		goto out;
532 	}
533 
534 	/*
535 	 * Temporarily increase the backing VM object's reference
536 	 * count so that a forced reclamation of its vnode does not
537 	 * immediately destroy it.
538 	 */
539 	vm_object_reference_locked(obj);
540 	VM_OBJECT_WUNLOCK(obj);
541 	*obj_res = obj;
542 	*vp_res = vp;
543 	*shmfd_res = shmfd;
544 
545 out:
546 	if (vp != NULL)
547 		VOP_UNLOCK(vp, 0);
548 	return (error);
549 }
550 
551 static int
552 sendfile_getsock(struct thread *td, int s, struct file **sock_fp,
553     struct socket **so)
554 {
555 	int error;
556 
557 	*sock_fp = NULL;
558 	*so = NULL;
559 
560 	/*
561 	 * The socket must be a stream socket and connected.
562 	 */
563 	error = getsock_cap(td, s, &cap_send_rights,
564 	    sock_fp, NULL, NULL);
565 	if (error != 0)
566 		return (error);
567 	*so = (*sock_fp)->f_data;
568 	if ((*so)->so_type != SOCK_STREAM)
569 		return (EINVAL);
570 	if (SOLISTENING(*so))
571 		return (ENOTCONN);
572 	return (0);
573 }
574 
575 int
576 vn_sendfile(struct file *fp, int sockfd, struct uio *hdr_uio,
577     struct uio *trl_uio, off_t offset, size_t nbytes, off_t *sent, int flags,
578     struct thread *td)
579 {
580 	struct file *sock_fp;
581 	struct vnode *vp;
582 	struct vm_object *obj;
583 	struct socket *so;
584 #ifdef KERN_TLS
585 	struct ktls_session *tls;
586 #endif
587 	struct mbuf_ext_pgs *ext_pgs;
588 	struct mbuf *m, *mh, *mhtail;
589 	struct sf_buf *sf;
590 	struct shmfd *shmfd;
591 	struct sendfile_sync *sfs;
592 	struct vattr va;
593 	off_t off, sbytes, rem, obj_size;
594 	int bsize, error, ext_pgs_idx, hdrlen, max_pgs, softerr;
595 #ifdef KERN_TLS
596 	int tls_enq_cnt;
597 #endif
598 	bool use_ext_pgs;
599 
600 	obj = NULL;
601 	so = NULL;
602 	m = mh = NULL;
603 	sfs = NULL;
604 #ifdef KERN_TLS
605 	tls = NULL;
606 #endif
607 	hdrlen = sbytes = 0;
608 	softerr = 0;
609 	use_ext_pgs = false;
610 
611 	error = sendfile_getobj(td, fp, &obj, &vp, &shmfd, &obj_size, &bsize);
612 	if (error != 0)
613 		return (error);
614 
615 	error = sendfile_getsock(td, sockfd, &sock_fp, &so);
616 	if (error != 0)
617 		goto out;
618 
619 #ifdef MAC
620 	error = mac_socket_check_send(td->td_ucred, so);
621 	if (error != 0)
622 		goto out;
623 #endif
624 
625 	SFSTAT_INC(sf_syscalls);
626 	SFSTAT_ADD(sf_rhpages_requested, SF_READAHEAD(flags));
627 
628 	if (flags & SF_SYNC) {
629 		sfs = malloc(sizeof *sfs, M_TEMP, M_WAITOK | M_ZERO);
630 		mtx_init(&sfs->mtx, "sendfile", NULL, MTX_DEF);
631 		cv_init(&sfs->cv, "sendfile");
632 	}
633 
634 	rem = nbytes ? omin(nbytes, obj_size - offset) : obj_size - offset;
635 
636 	/*
637 	 * Protect against multiple writers to the socket.
638 	 *
639 	 * XXXRW: Historically this has assumed non-interruptibility, so now
640 	 * we implement that, but possibly shouldn't.
641 	 */
642 	(void)sblock(&so->so_snd, SBL_WAIT | SBL_NOINTR);
643 #ifdef KERN_TLS
644 	tls = ktls_hold(so->so_snd.sb_tls_info);
645 #endif
646 
647 	/*
648 	 * Loop through the pages of the file, starting with the requested
649 	 * offset. Get a file page (do I/O if necessary), map the file page
650 	 * into an sf_buf, attach an mbuf header to the sf_buf, and queue
651 	 * it on the socket.
652 	 * This is done in two loops.  The inner loop turns as many pages
653 	 * as it can, up to available socket buffer space, without blocking
654 	 * into mbufs to have it bulk delivered into the socket send buffer.
655 	 * The outer loop checks the state and available space of the socket
656 	 * and takes care of the overall progress.
657 	 */
658 	for (off = offset; rem > 0; ) {
659 		struct sf_io *sfio;
660 		vm_page_t *pa;
661 		struct mbuf *mtail;
662 		int nios, space, npages, rhpages;
663 
664 		mtail = NULL;
665 		/*
666 		 * Check the socket state for ongoing connection,
667 		 * no errors and space in socket buffer.
668 		 * If space is low allow for the remainder of the
669 		 * file to be processed if it fits the socket buffer.
670 		 * Otherwise block in waiting for sufficient space
671 		 * to proceed, or if the socket is nonblocking, return
672 		 * to userland with EAGAIN while reporting how far
673 		 * we've come.
674 		 * We wait until the socket buffer has significant free
675 		 * space to do bulk sends.  This makes good use of file
676 		 * system read ahead and allows packet segmentation
677 		 * offloading hardware to take over lots of work.  If
678 		 * we were not careful here we would send off only one
679 		 * sfbuf at a time.
680 		 */
681 		SOCKBUF_LOCK(&so->so_snd);
682 		if (so->so_snd.sb_lowat < so->so_snd.sb_hiwat / 2)
683 			so->so_snd.sb_lowat = so->so_snd.sb_hiwat / 2;
684 retry_space:
685 		if (so->so_snd.sb_state & SBS_CANTSENDMORE) {
686 			error = EPIPE;
687 			SOCKBUF_UNLOCK(&so->so_snd);
688 			goto done;
689 		} else if (so->so_error) {
690 			error = so->so_error;
691 			so->so_error = 0;
692 			SOCKBUF_UNLOCK(&so->so_snd);
693 			goto done;
694 		}
695 		if ((so->so_state & SS_ISCONNECTED) == 0) {
696 			SOCKBUF_UNLOCK(&so->so_snd);
697 			error = ENOTCONN;
698 			goto done;
699 		}
700 
701 		space = sbspace(&so->so_snd);
702 		if (space < rem &&
703 		    (space <= 0 ||
704 		     space < so->so_snd.sb_lowat)) {
705 			if (so->so_state & SS_NBIO) {
706 				SOCKBUF_UNLOCK(&so->so_snd);
707 				error = EAGAIN;
708 				goto done;
709 			}
710 			/*
711 			 * sbwait drops the lock while sleeping.
712 			 * When we loop back to retry_space the
713 			 * state may have changed and we retest
714 			 * for it.
715 			 */
716 			error = sbwait(&so->so_snd);
717 			/*
718 			 * An error from sbwait usually indicates that we've
719 			 * been interrupted by a signal. If we've sent anything
720 			 * then return bytes sent, otherwise return the error.
721 			 */
722 			if (error != 0) {
723 				SOCKBUF_UNLOCK(&so->so_snd);
724 				goto done;
725 			}
726 			goto retry_space;
727 		}
728 		SOCKBUF_UNLOCK(&so->so_snd);
729 
730 		/*
731 		 * At the beginning of the first loop check if any headers
732 		 * are specified and copy them into mbufs.  Reduce space in
733 		 * the socket buffer by the size of the header mbuf chain.
734 		 * Clear hdr_uio here and hdrlen at the end of the first loop.
735 		 */
736 		if (hdr_uio != NULL && hdr_uio->uio_resid > 0) {
737 			hdr_uio->uio_td = td;
738 			hdr_uio->uio_rw = UIO_WRITE;
739 #ifdef KERN_TLS
740 			if (tls != NULL)
741 				mh = m_uiotombuf(hdr_uio, M_WAITOK, space,
742 				    tls->params.max_frame_len, M_NOMAP);
743 			else
744 #endif
745 				mh = m_uiotombuf(hdr_uio, M_WAITOK,
746 				    space, 0, 0);
747 			hdrlen = m_length(mh, &mhtail);
748 			space -= hdrlen;
749 			/*
750 			 * If header consumed all the socket buffer space,
751 			 * don't waste CPU cycles and jump to the end.
752 			 */
753 			if (space == 0) {
754 				sfio = NULL;
755 				nios = 0;
756 				goto prepend_header;
757 			}
758 			hdr_uio = NULL;
759 		}
760 
761 		if (vp != NULL) {
762 			error = vn_lock(vp, LK_SHARED);
763 			if (error != 0)
764 				goto done;
765 			error = VOP_GETATTR(vp, &va, td->td_ucred);
766 			if (error != 0 || off >= va.va_size) {
767 				VOP_UNLOCK(vp, 0);
768 				goto done;
769 			}
770 			if (va.va_size != obj_size) {
771 				obj_size = va.va_size;
772 				rem = nbytes ?
773 				    omin(nbytes + offset, obj_size) : obj_size;
774 				rem -= off;
775 			}
776 		}
777 
778 		if (space > rem)
779 			space = rem;
780 		else if (space > PAGE_SIZE) {
781 			/*
782 			 * Use page boundaries when possible for large
783 			 * requests.
784 			 */
785 			if (off & PAGE_MASK)
786 				space -= (PAGE_SIZE - (off & PAGE_MASK));
787 			space = trunc_page(space);
788 			if (off & PAGE_MASK)
789 				space += (PAGE_SIZE - (off & PAGE_MASK));
790 		}
791 
792 		npages = howmany(space + (off & PAGE_MASK), PAGE_SIZE);
793 
794 		/*
795 		 * Calculate maximum allowed number of pages for readahead
796 		 * at this iteration.  If SF_USER_READAHEAD was set, we don't
797 		 * do any heuristics and use exactly the value supplied by
798 		 * application.  Otherwise, we allow readahead up to "rem".
799 		 * If application wants more, let it be, but there is no
800 		 * reason to go above MAXPHYS.  Also check against "obj_size",
801 		 * since vm_pager_has_page() can hint beyond EOF.
802 		 */
803 		if (flags & SF_USER_READAHEAD) {
804 			rhpages = SF_READAHEAD(flags);
805 		} else {
806 			rhpages = howmany(rem + (off & PAGE_MASK), PAGE_SIZE) -
807 			    npages;
808 			rhpages += SF_READAHEAD(flags);
809 		}
810 		rhpages = min(howmany(MAXPHYS, PAGE_SIZE), rhpages);
811 		rhpages = min(howmany(obj_size - trunc_page(off), PAGE_SIZE) -
812 		    npages, rhpages);
813 
814 		sfio = malloc(sizeof(struct sf_io) +
815 		    npages * sizeof(vm_page_t), M_TEMP, M_WAITOK);
816 		refcount_init(&sfio->nios, 1);
817 		sfio->so = so;
818 		sfio->error = 0;
819 
820 #ifdef KERN_TLS
821 		/*
822 		 * This doesn't use ktls_hold() because sfio->m will
823 		 * also have a reference on 'tls' that will be valid
824 		 * for all of sfio's lifetime.
825 		 */
826 		sfio->tls = tls;
827 #endif
828 
829 		error = sendfile_swapin(obj, sfio, &nios, off, space, npages,
830 		    rhpages, flags);
831 		if (error != 0) {
832 			if (vp != NULL)
833 				VOP_UNLOCK(vp, 0);
834 			sfio->m = NULL;
835 			sendfile_iodone(sfio, NULL, 0, error);
836 			goto done;
837 		}
838 
839 		/*
840 		 * Loop and construct maximum sized mbuf chain to be bulk
841 		 * dumped into socket buffer.
842 		 */
843 		pa = sfio->pa;
844 
845 		/*
846 		 * Use unmapped mbufs if enabled for TCP.  Unmapped
847 		 * bufs are restricted to TCP as that is what has been
848 		 * tested.  In particular, unmapped mbufs have not
849 		 * been tested with UNIX-domain sockets.
850 		 *
851 		 * TLS frames always require unmapped mbufs.
852 		 */
853 		if ((mb_use_ext_pgs &&
854 		    so->so_proto->pr_protocol == IPPROTO_TCP)
855 #ifdef KERN_TLS
856 		    || tls != NULL
857 #endif
858 		    ) {
859 			use_ext_pgs = true;
860 #ifdef KERN_TLS
861 			if (tls != NULL)
862 				max_pgs = num_pages(tls->params.max_frame_len);
863 			else
864 #endif
865 				max_pgs = MBUF_PEXT_MAX_PGS;
866 
867 			/* Start at last index, to wrap on first use. */
868 			ext_pgs_idx = max_pgs - 1;
869 		}
870 
871 		for (int i = 0; i < npages; i++) {
872 			struct mbuf *m0;
873 
874 			/*
875 			 * If a page wasn't grabbed successfully, then
876 			 * trim the array. Can happen only with SF_NODISKIO.
877 			 */
878 			if (pa[i] == NULL) {
879 				SFSTAT_INC(sf_busy);
880 				fixspace(npages, i, off, &space);
881 				npages = i;
882 				softerr = EBUSY;
883 				break;
884 			}
885 
886 			if (use_ext_pgs) {
887 				off_t xfs;
888 
889 				ext_pgs_idx++;
890 				if (ext_pgs_idx == max_pgs) {
891 					m0 = mb_alloc_ext_pgs(M_WAITOK, false,
892 					    sendfile_free_mext_pg);
893 
894 					if (flags & SF_NOCACHE) {
895 						m0->m_ext.ext_flags |=
896 						    EXT_FLAG_NOCACHE;
897 
898 						/*
899 						 * See comment below regarding
900 						 * ignoring SF_NOCACHE for the
901 						 * last page.
902 						 */
903 						if ((npages - i <= max_pgs) &&
904 						    ((off + space) & PAGE_MASK) &&
905 						    (rem > space || rhpages > 0))
906 							m0->m_ext.ext_flags |=
907 							    EXT_FLAG_CACHE_LAST;
908 					}
909 					if (sfs != NULL) {
910 						m0->m_ext.ext_flags |=
911 						    EXT_FLAG_SYNC;
912 						m0->m_ext.ext_arg2 = sfs;
913 						mtx_lock(&sfs->mtx);
914 						sfs->count++;
915 						mtx_unlock(&sfs->mtx);
916 					}
917 					ext_pgs = m0->m_ext.ext_pgs;
918 					if (i == 0)
919 						sfio->m = m0;
920 					ext_pgs_idx = 0;
921 
922 					/* Append to mbuf chain. */
923 					if (mtail != NULL)
924 						mtail->m_next = m0;
925 					else
926 						m = m0;
927 					mtail = m0;
928 					ext_pgs->first_pg_off =
929 					    vmoff(i, off) & PAGE_MASK;
930 				}
931 				if (nios) {
932 					mtail->m_flags |= M_NOTREADY;
933 					ext_pgs->nrdy++;
934 				}
935 
936 				ext_pgs->pa[ext_pgs_idx] = VM_PAGE_TO_PHYS(pa[i]);
937 				ext_pgs->npgs++;
938 				xfs = xfsize(i, npages, off, space);
939 				ext_pgs->last_pg_len = xfs;
940 				MBUF_EXT_PGS_ASSERT_SANITY(ext_pgs);
941 				mtail->m_len += xfs;
942 				mtail->m_ext.ext_size += PAGE_SIZE;
943 				continue;
944 			}
945 
946 			/*
947 			 * Get a sendfile buf.  When allocating the
948 			 * first buffer for mbuf chain, we usually
949 			 * wait as long as necessary, but this wait
950 			 * can be interrupted.  For consequent
951 			 * buffers, do not sleep, since several
952 			 * threads might exhaust the buffers and then
953 			 * deadlock.
954 			 */
955 			sf = sf_buf_alloc(pa[i],
956 			    m != NULL ? SFB_NOWAIT : SFB_CATCH);
957 			if (sf == NULL) {
958 				SFSTAT_INC(sf_allocfail);
959 				for (int j = i; j < npages; j++)
960 					vm_page_unwire(pa[j], PQ_INACTIVE);
961 				if (m == NULL)
962 					softerr = ENOBUFS;
963 				fixspace(npages, i, off, &space);
964 				npages = i;
965 				break;
966 			}
967 
968 			m0 = m_get(M_WAITOK, MT_DATA);
969 			m0->m_ext.ext_buf = (char *)sf_buf_kva(sf);
970 			m0->m_ext.ext_size = PAGE_SIZE;
971 			m0->m_ext.ext_arg1 = sf;
972 			m0->m_ext.ext_type = EXT_SFBUF;
973 			m0->m_ext.ext_flags = EXT_FLAG_EMBREF;
974 			m0->m_ext.ext_free = sendfile_free_mext;
975 			/*
976 			 * SF_NOCACHE sets the page as being freed upon send.
977 			 * However, we ignore it for the last page in 'space',
978 			 * if the page is truncated, and we got more data to
979 			 * send (rem > space), or if we have readahead
980 			 * configured (rhpages > 0).
981 			 */
982 			if ((flags & SF_NOCACHE) &&
983 			    (i != npages - 1 ||
984 			    !((off + space) & PAGE_MASK) ||
985 			    !(rem > space || rhpages > 0)))
986 				m0->m_ext.ext_flags |= EXT_FLAG_NOCACHE;
987 			if (sfs != NULL) {
988 				m0->m_ext.ext_flags |= EXT_FLAG_SYNC;
989 				m0->m_ext.ext_arg2 = sfs;
990 				mtx_lock(&sfs->mtx);
991 				sfs->count++;
992 				mtx_unlock(&sfs->mtx);
993 			}
994 			m0->m_ext.ext_count = 1;
995 			m0->m_flags |= (M_EXT | M_RDONLY);
996 			if (nios)
997 				m0->m_flags |= M_NOTREADY;
998 			m0->m_data = (char *)sf_buf_kva(sf) +
999 			    (vmoff(i, off) & PAGE_MASK);
1000 			m0->m_len = xfsize(i, npages, off, space);
1001 
1002 			if (i == 0)
1003 				sfio->m = m0;
1004 
1005 			/* Append to mbuf chain. */
1006 			if (mtail != NULL)
1007 				mtail->m_next = m0;
1008 			else
1009 				m = m0;
1010 			mtail = m0;
1011 		}
1012 
1013 		if (vp != NULL)
1014 			VOP_UNLOCK(vp, 0);
1015 
1016 		/* Keep track of bytes processed. */
1017 		off += space;
1018 		rem -= space;
1019 
1020 		/* Prepend header, if any. */
1021 		if (hdrlen) {
1022 prepend_header:
1023 			mhtail->m_next = m;
1024 			m = mh;
1025 			mh = NULL;
1026 		}
1027 
1028 		if (m == NULL) {
1029 			KASSERT(softerr, ("%s: m NULL, no error", __func__));
1030 			error = softerr;
1031 			free(sfio, M_TEMP);
1032 			goto done;
1033 		}
1034 
1035 		/* Add the buffer chain to the socket buffer. */
1036 		KASSERT(m_length(m, NULL) == space + hdrlen,
1037 		    ("%s: mlen %u space %d hdrlen %d",
1038 		    __func__, m_length(m, NULL), space, hdrlen));
1039 
1040 		CURVNET_SET(so->so_vnet);
1041 #ifdef KERN_TLS
1042 		if (tls != NULL) {
1043 			error = ktls_frame(m, tls, &tls_enq_cnt,
1044 			    TLS_RLTYPE_APP);
1045 			if (error != 0)
1046 				goto done;
1047 		}
1048 #endif
1049 		if (nios == 0) {
1050 			/*
1051 			 * If sendfile_swapin() didn't initiate any I/Os,
1052 			 * which happens if all data is cached in VM, then
1053 			 * we can send data right now without the
1054 			 * PRUS_NOTREADY flag.
1055 			 */
1056 			free(sfio, M_TEMP);
1057 #ifdef KERN_TLS
1058 			if (tls != NULL && tls->mode == TCP_TLS_MODE_SW) {
1059 				error = (*so->so_proto->pr_usrreqs->pru_send)
1060 				    (so, PRUS_NOTREADY, m, NULL, NULL, td);
1061 				soref(so);
1062 				ktls_enqueue(m, so, tls_enq_cnt);
1063 			} else
1064 #endif
1065 				error = (*so->so_proto->pr_usrreqs->pru_send)
1066 				    (so, 0, m, NULL, NULL, td);
1067 		} else {
1068 			sfio->npages = npages;
1069 			soref(so);
1070 			error = (*so->so_proto->pr_usrreqs->pru_send)
1071 			    (so, PRUS_NOTREADY, m, NULL, NULL, td);
1072 			sendfile_iodone(sfio, NULL, 0, 0);
1073 		}
1074 		CURVNET_RESTORE();
1075 
1076 		m = NULL;	/* pru_send always consumes */
1077 		if (error)
1078 			goto done;
1079 		sbytes += space + hdrlen;
1080 		if (hdrlen)
1081 			hdrlen = 0;
1082 		if (softerr) {
1083 			error = softerr;
1084 			goto done;
1085 		}
1086 	}
1087 
1088 	/*
1089 	 * Send trailers. Wimp out and use writev(2).
1090 	 */
1091 	if (trl_uio != NULL) {
1092 		sbunlock(&so->so_snd);
1093 		error = kern_writev(td, sockfd, trl_uio);
1094 		if (error == 0)
1095 			sbytes += td->td_retval[0];
1096 		goto out;
1097 	}
1098 
1099 done:
1100 	sbunlock(&so->so_snd);
1101 out:
1102 	/*
1103 	 * If there was no error we have to clear td->td_retval[0]
1104 	 * because it may have been set by writev.
1105 	 */
1106 	if (error == 0) {
1107 		td->td_retval[0] = 0;
1108 	}
1109 	if (sent != NULL) {
1110 		(*sent) = sbytes;
1111 	}
1112 	if (obj != NULL)
1113 		vm_object_deallocate(obj);
1114 	if (so)
1115 		fdrop(sock_fp, td);
1116 	if (m)
1117 		m_freem(m);
1118 	if (mh)
1119 		m_freem(mh);
1120 
1121 	if (sfs != NULL) {
1122 		mtx_lock(&sfs->mtx);
1123 		if (sfs->count != 0)
1124 			cv_wait(&sfs->cv, &sfs->mtx);
1125 		KASSERT(sfs->count == 0, ("sendfile sync still busy"));
1126 		cv_destroy(&sfs->cv);
1127 		mtx_destroy(&sfs->mtx);
1128 		free(sfs, M_TEMP);
1129 	}
1130 #ifdef KERN_TLS
1131 	if (tls != NULL)
1132 		ktls_free(tls);
1133 #endif
1134 
1135 	if (error == ERESTART)
1136 		error = EINTR;
1137 
1138 	return (error);
1139 }
1140 
1141 static int
1142 sendfile(struct thread *td, struct sendfile_args *uap, int compat)
1143 {
1144 	struct sf_hdtr hdtr;
1145 	struct uio *hdr_uio, *trl_uio;
1146 	struct file *fp;
1147 	off_t sbytes;
1148 	int error;
1149 
1150 	/*
1151 	 * File offset must be positive.  If it goes beyond EOF
1152 	 * we send only the header/trailer and no payload data.
1153 	 */
1154 	if (uap->offset < 0)
1155 		return (EINVAL);
1156 
1157 	sbytes = 0;
1158 	hdr_uio = trl_uio = NULL;
1159 
1160 	if (uap->hdtr != NULL) {
1161 		error = copyin(uap->hdtr, &hdtr, sizeof(hdtr));
1162 		if (error != 0)
1163 			goto out;
1164 		if (hdtr.headers != NULL) {
1165 			error = copyinuio(hdtr.headers, hdtr.hdr_cnt,
1166 			    &hdr_uio);
1167 			if (error != 0)
1168 				goto out;
1169 #ifdef COMPAT_FREEBSD4
1170 			/*
1171 			 * In FreeBSD < 5.0 the nbytes to send also included
1172 			 * the header.  If compat is specified subtract the
1173 			 * header size from nbytes.
1174 			 */
1175 			if (compat) {
1176 				if (uap->nbytes > hdr_uio->uio_resid)
1177 					uap->nbytes -= hdr_uio->uio_resid;
1178 				else
1179 					uap->nbytes = 0;
1180 			}
1181 #endif
1182 		}
1183 		if (hdtr.trailers != NULL) {
1184 			error = copyinuio(hdtr.trailers, hdtr.trl_cnt,
1185 			    &trl_uio);
1186 			if (error != 0)
1187 				goto out;
1188 		}
1189 	}
1190 
1191 	AUDIT_ARG_FD(uap->fd);
1192 
1193 	/*
1194 	 * sendfile(2) can start at any offset within a file so we require
1195 	 * CAP_READ+CAP_SEEK = CAP_PREAD.
1196 	 */
1197 	if ((error = fget_read(td, uap->fd, &cap_pread_rights, &fp)) != 0)
1198 		goto out;
1199 
1200 	error = fo_sendfile(fp, uap->s, hdr_uio, trl_uio, uap->offset,
1201 	    uap->nbytes, &sbytes, uap->flags, td);
1202 	fdrop(fp, td);
1203 
1204 	if (uap->sbytes != NULL)
1205 		copyout(&sbytes, uap->sbytes, sizeof(off_t));
1206 
1207 out:
1208 	free(hdr_uio, M_IOV);
1209 	free(trl_uio, M_IOV);
1210 	return (error);
1211 }
1212 
1213 /*
1214  * sendfile(2)
1215  *
1216  * int sendfile(int fd, int s, off_t offset, size_t nbytes,
1217  *       struct sf_hdtr *hdtr, off_t *sbytes, int flags)
1218  *
1219  * Send a file specified by 'fd' and starting at 'offset' to a socket
1220  * specified by 's'. Send only 'nbytes' of the file or until EOF if nbytes ==
1221  * 0.  Optionally add a header and/or trailer to the socket output.  If
1222  * specified, write the total number of bytes sent into *sbytes.
1223  */
1224 int
1225 sys_sendfile(struct thread *td, struct sendfile_args *uap)
1226 {
1227 
1228 	return (sendfile(td, uap, 0));
1229 }
1230 
1231 #ifdef COMPAT_FREEBSD4
1232 int
1233 freebsd4_sendfile(struct thread *td, struct freebsd4_sendfile_args *uap)
1234 {
1235 	struct sendfile_args args;
1236 
1237 	args.fd = uap->fd;
1238 	args.s = uap->s;
1239 	args.offset = uap->offset;
1240 	args.nbytes = uap->nbytes;
1241 	args.hdtr = uap->hdtr;
1242 	args.sbytes = uap->sbytes;
1243 	args.flags = uap->flags;
1244 
1245 	return (sendfile(td, &args, 1));
1246 }
1247 #endif /* COMPAT_FREEBSD4 */
1248