xref: /freebsd/sys/kern/kern_sendfile.c (revision d17cbe46983c08ba0567b0ffd71434abd35aa205)
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 #include "opt_kern_tls.h"
32 
33 #include <sys/param.h>
34 #include <sys/systm.h>
35 #include <sys/capsicum.h>
36 #include <sys/kernel.h>
37 #include <sys/lock.h>
38 #include <sys/ktls.h>
39 #include <sys/mutex.h>
40 #include <sys/malloc.h>
41 #include <sys/mman.h>
42 #include <sys/mount.h>
43 #include <sys/mbuf.h>
44 #include <sys/proc.h>
45 #include <sys/protosw.h>
46 #include <sys/rwlock.h>
47 #include <sys/sf_buf.h>
48 #include <sys/socket.h>
49 #include <sys/socketvar.h>
50 #include <sys/syscallsubr.h>
51 #include <sys/sysctl.h>
52 #include <sys/sysproto.h>
53 #include <sys/vnode.h>
54 
55 #include <net/vnet.h>
56 #include <netinet/in.h>
57 #include <netinet/tcp.h>
58 #include <netinet/in_pcb.h>
59 #include <netinet/tcp_var.h>
60 #include <netinet/tcp_log_buf.h>
61 
62 #include <security/audit/audit.h>
63 #include <security/mac/mac_framework.h>
64 
65 #include <vm/vm.h>
66 #include <vm/vm_object.h>
67 #include <vm/vm_pager.h>
68 
69 static MALLOC_DEFINE(M_SENDFILE, "sendfile", "sendfile dynamic memory");
70 
71 #define	EXT_FLAG_SYNC		EXT_FLAG_VENDOR1
72 #define	EXT_FLAG_NOCACHE	EXT_FLAG_VENDOR2
73 #define	EXT_FLAG_CACHE_LAST	EXT_FLAG_VENDOR3
74 
75 /*
76  * Structure describing a single sendfile(2) I/O, which may consist of
77  * several underlying pager I/Os.
78  *
79  * The syscall context allocates the structure and initializes 'nios'
80  * to 1.  As sendfile_swapin() runs through pages and starts asynchronous
81  * paging operations, it increments 'nios'.
82  *
83  * Every I/O completion calls sendfile_iodone(), which decrements the 'nios',
84  * and the syscall also calls sendfile_iodone() after allocating all mbufs,
85  * linking them and sending to socket.  Whoever reaches zero 'nios' is
86  * responsible to call pr_ready() on the socket, to notify it of readyness
87  * of the data.
88  */
89 struct sf_io {
90 	volatile u_int	nios;
91 	u_int		error;
92 	int		npages;
93 	struct socket	*so;
94 	struct mbuf	*m;
95 	vm_object_t	obj;
96 	vm_pindex_t	pindex0;
97 #ifdef KERN_TLS
98 	struct ktls_session *tls;
99 #endif
100 	vm_page_t	pa[];
101 };
102 
103 /*
104  * Structure used to track requests with SF_SYNC flag.
105  */
106 struct sendfile_sync {
107 	struct mtx	mtx;
108 	struct cv	cv;
109 	unsigned	count;
110 	bool		waiting;
111 };
112 
113 static void
sendfile_sync_destroy(struct sendfile_sync * sfs)114 sendfile_sync_destroy(struct sendfile_sync *sfs)
115 {
116 	KASSERT(sfs->count == 0, ("sendfile sync %p still busy", sfs));
117 
118 	cv_destroy(&sfs->cv);
119 	mtx_destroy(&sfs->mtx);
120 	free(sfs, M_SENDFILE);
121 }
122 
123 static void
sendfile_sync_signal(struct sendfile_sync * sfs)124 sendfile_sync_signal(struct sendfile_sync *sfs)
125 {
126 	mtx_lock(&sfs->mtx);
127 	KASSERT(sfs->count > 0, ("sendfile sync %p not busy", sfs));
128 	if (--sfs->count == 0) {
129 		if (!sfs->waiting) {
130 			/* The sendfile() waiter was interrupted by a signal. */
131 			sendfile_sync_destroy(sfs);
132 			return;
133 		} else {
134 			cv_signal(&sfs->cv);
135 		}
136 	}
137 	mtx_unlock(&sfs->mtx);
138 }
139 
140 counter_u64_t sfstat[sizeof(struct sfstat) / sizeof(uint64_t)];
141 
142 static void
sfstat_init(const void * unused)143 sfstat_init(const void *unused)
144 {
145 
146 	COUNTER_ARRAY_ALLOC(sfstat, sizeof(struct sfstat) / sizeof(uint64_t),
147 	    M_WAITOK);
148 }
149 SYSINIT(sfstat, SI_SUB_MBUF, SI_ORDER_FIRST, sfstat_init, NULL);
150 
151 static int
sfstat_sysctl(SYSCTL_HANDLER_ARGS)152 sfstat_sysctl(SYSCTL_HANDLER_ARGS)
153 {
154 	struct sfstat s;
155 
156 	COUNTER_ARRAY_COPY(sfstat, &s, sizeof(s) / sizeof(uint64_t));
157 	if (req->newptr)
158 		COUNTER_ARRAY_ZERO(sfstat, sizeof(s) / sizeof(uint64_t));
159 	return (SYSCTL_OUT(req, &s, sizeof(s)));
160 }
161 SYSCTL_PROC(_kern_ipc, OID_AUTO, sfstat,
162     CTLTYPE_OPAQUE | CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 0,
163     sfstat_sysctl, "I",
164     "sendfile statistics");
165 
166 static void
sendfile_free_mext(struct mbuf * m)167 sendfile_free_mext(struct mbuf *m)
168 {
169 	struct sf_buf *sf;
170 	vm_page_t pg;
171 	int flags;
172 
173 	KASSERT(m->m_flags & M_EXT && m->m_ext.ext_type == EXT_SFBUF,
174 	    ("%s: m %p !M_EXT or !EXT_SFBUF", __func__, m));
175 
176 	sf = m->m_ext.ext_arg1;
177 	pg = sf_buf_page(sf);
178 	flags = (m->m_ext.ext_flags & EXT_FLAG_NOCACHE) != 0 ? VPR_TRYFREE : 0;
179 
180 	sf_buf_free(sf);
181 	vm_page_release(pg, flags);
182 
183 	if (m->m_ext.ext_flags & EXT_FLAG_SYNC) {
184 		struct sendfile_sync *sfs = m->m_ext.ext_arg2;
185 		sendfile_sync_signal(sfs);
186 	}
187 }
188 
189 static void
sendfile_free_mext_pg(struct mbuf * m)190 sendfile_free_mext_pg(struct mbuf *m)
191 {
192 	vm_page_t pg;
193 	int flags, i;
194 	bool cache_last;
195 
196 	M_ASSERTEXTPG(m);
197 
198 	cache_last = m->m_ext.ext_flags & EXT_FLAG_CACHE_LAST;
199 	flags = (m->m_ext.ext_flags & EXT_FLAG_NOCACHE) != 0 ? VPR_TRYFREE : 0;
200 
201 	for (i = 0; i < m->m_epg_npgs; i++) {
202 		if (cache_last && i == m->m_epg_npgs - 1)
203 			flags = 0;
204 		pg = PHYS_TO_VM_PAGE(m->m_epg_pa[i]);
205 		vm_page_release(pg, flags);
206 	}
207 
208 	if (m->m_ext.ext_flags & EXT_FLAG_SYNC) {
209 		struct sendfile_sync *sfs = m->m_ext.ext_arg1;
210 		sendfile_sync_signal(sfs);
211 	}
212 }
213 
214 /*
215  * Helper function to calculate how much data to put into page i of n.
216  * Only first and last pages are special.
217  */
218 static inline off_t
xfsize(int i,int n,off_t off,off_t len)219 xfsize(int i, int n, off_t off, off_t len)
220 {
221 
222 	if (i == 0)
223 		return (omin(PAGE_SIZE - (off & PAGE_MASK), len));
224 
225 	if (i == n - 1 && ((off + len) & PAGE_MASK) > 0)
226 		return ((off + len) & PAGE_MASK);
227 
228 	return (PAGE_SIZE);
229 }
230 
231 /*
232  * Helper function to get offset within object for i page.
233  */
234 static inline vm_ooffset_t
vmoff(int i,off_t off)235 vmoff(int i, off_t off)
236 {
237 
238 	if (i == 0)
239 		return ((vm_ooffset_t)off);
240 
241 	return (trunc_page(off + i * PAGE_SIZE));
242 }
243 
244 /*
245  * Helper function used when allocation of a page or sf_buf failed.
246  * Pretend as if we don't have enough space, subtract xfsize() of
247  * all pages that failed.
248  */
249 static inline void
fixspace(int old,int new,off_t off,int * space)250 fixspace(int old, int new, off_t off, int *space)
251 {
252 
253 	KASSERT(old > new, ("%s: old %d new %d", __func__, old, new));
254 
255 	/* Subtract last one. */
256 	*space -= xfsize(old - 1, old, off, *space);
257 	old--;
258 
259 	if (new == old)
260 		/* There was only one page. */
261 		return;
262 
263 	/* Subtract first one. */
264 	if (new == 0) {
265 		*space -= xfsize(0, old, off, *space);
266 		new++;
267 	}
268 
269 	/* Rest of pages are full sized. */
270 	*space -= (old - new) * PAGE_SIZE;
271 
272 	KASSERT(*space >= 0, ("%s: space went backwards", __func__));
273 }
274 
275 /*
276  * Wait for all in-flight ios to complete, we must not unwire pages
277  * under them.
278  */
279 static void
sendfile_iowait(struct sf_io * sfio,const char * wmesg)280 sendfile_iowait(struct sf_io *sfio, const char *wmesg)
281 {
282 	while (atomic_load_int(&sfio->nios) != 1)
283 		pause(wmesg, 1);
284 }
285 
286 /*
287  * I/O completion callback.
288  *
289  * When called via I/O path, the curvnet is not set and should be obtained
290  * from the socket.  When called synchronously from vn_sendfile(), usually
291  * to report error or just release the reference (all pages are valid), then
292  * curvnet shall be already set.
293  */
294 static void
sendfile_iodone(void * arg,vm_page_t * pa,int count,int error)295 sendfile_iodone(void *arg, vm_page_t *pa, int count, int error)
296 {
297 	struct sf_io *sfio = arg;
298 	struct socket *so;
299 	int i;
300 
301 	if (error != 0)
302 		sfio->error = error;
303 
304 	/*
305 	 * Restore the valid page pointers.  They are already
306 	 * unbusied, but still wired.
307 	 *
308 	 * XXXKIB since pages are only wired, and we do not
309 	 * own the object lock, other users might have
310 	 * invalidated them in meantime.  Similarly, after we
311 	 * unbusied the swapped-in pages, they can become
312 	 * invalid under us.
313 	 */
314 	MPASS(count == 0 || pa[0] != bogus_page);
315 	for (i = 0; i < count; i++) {
316 		if (pa[i] == bogus_page) {
317 			sfio->pa[(pa[0]->pindex - sfio->pindex0) + i] =
318 			    pa[i] = vm_page_relookup(sfio->obj,
319 			    pa[0]->pindex + i);
320 			KASSERT(pa[i] != NULL,
321 			    ("%s: page %p[%d] disappeared",
322 			    __func__, pa, i));
323 		} else {
324 			vm_page_xunbusy_unchecked(pa[i]);
325 		}
326 	}
327 
328 	if (!refcount_release(&sfio->nios))
329 		return;
330 
331 #ifdef INVARIANTS
332 	for (i = 1; i < sfio->npages; i++) {
333 		if (sfio->pa[i] == NULL)
334 			break;
335 		KASSERT(vm_page_wired(sfio->pa[i]),
336 		    ("sfio %p page %d %p not wired", sfio, i, sfio->pa[i]));
337 		if (i == 0)
338 			continue;
339 		KASSERT(sfio->pa[0]->object == sfio->pa[i]->object,
340 		    ("sfio %p page %d %p wrong owner %p %p", sfio, i,
341 		    sfio->pa[i], sfio->pa[0]->object, sfio->pa[i]->object));
342 		KASSERT(sfio->pa[0]->pindex + i == sfio->pa[i]->pindex,
343 		    ("sfio %p page %d %p wrong index %jx %jx", sfio, i,
344 		    sfio->pa[i], (uintmax_t)sfio->pa[0]->pindex,
345 		    (uintmax_t)sfio->pa[i]->pindex));
346 	}
347 #endif
348 
349 	vm_object_pip_wakeup(sfio->obj);
350 
351 	if (sfio->m == NULL) {
352 		/*
353 		 * Either I/O operation failed, or we failed to allocate
354 		 * buffers, or we bailed out on first busy page, or we
355 		 * succeeded filling the request without any I/Os. Anyway,
356 		 * pr_send() hadn't been executed - nothing had been sent
357 		 * to the socket yet.
358 		 */
359 		MPASS((curthread->td_pflags & TDP_KTHREAD) == 0);
360 		free(sfio, M_SENDFILE);
361 		return;
362 	}
363 
364 #if defined(KERN_TLS) && defined(INVARIANTS)
365 	if ((sfio->m->m_flags & M_EXTPG) != 0)
366 		KASSERT(sfio->tls == sfio->m->m_epg_tls,
367 		    ("TLS session mismatch"));
368 	else
369 		KASSERT(sfio->tls == NULL,
370 		    ("non-ext_pgs mbuf with TLS session"));
371 #endif
372 	so = sfio->so;
373 	CURVNET_SET_QUIET(so->so_vnet);
374 	if (__predict_false(sfio->error)) {
375 		/*
376 		 * I/O operation failed.  The state of data in the socket
377 		 * is now inconsistent, and all what we can do is to tear
378 		 * it down. Protocol abort method would tear down protocol
379 		 * state, free all ready mbufs and detach not ready ones.
380 		 * We will free the mbufs corresponding to this I/O manually.
381 		 *
382 		 * The socket would be marked with EIO and made available
383 		 * for read, so that application receives EIO on next
384 		 * syscall and eventually closes the socket.
385 		 */
386 		so->so_proto->pr_abort(so);
387 		so->so_error = EIO;
388 
389 		mb_free_notready(sfio->m, sfio->npages);
390 #ifdef KERN_TLS
391 	} else if (sfio->tls != NULL && sfio->tls->mode == TCP_TLS_MODE_SW) {
392 		/*
393 		 * I/O operation is complete, but we still need to
394 		 * encrypt.  We cannot do this in the interrupt thread
395 		 * of the disk controller, so forward the mbufs to a
396 		 * different thread.
397 		 *
398 		 * Donate the socket reference from sfio to rather
399 		 * than explicitly invoking soref().
400 		 */
401 		ktls_enqueue(sfio->m, so, sfio->npages);
402 		goto out_with_ref;
403 #endif
404 	} else
405 		(void)so->so_proto->pr_ready(so, sfio->m, sfio->npages);
406 
407 	sorele(so);
408 #ifdef KERN_TLS
409 out_with_ref:
410 #endif
411 	CURVNET_RESTORE();
412 	free(sfio, M_SENDFILE);
413 }
414 
415 /*
416  * Iterate through pages vector and request paging for non-valid pages.
417  */
418 static int
sendfile_swapin(vm_object_t obj,struct sf_io * sfio,int * nios,off_t off,off_t len,int rhpages,int flags)419 sendfile_swapin(vm_object_t obj, struct sf_io *sfio, int *nios, off_t off,
420     off_t len, int rhpages, int flags)
421 {
422 	vm_page_t *pa;
423 	int a, count, count1, grabbed, i, j, npages, rv;
424 
425 	pa = sfio->pa;
426 	npages = sfio->npages;
427 	*nios = 0;
428 	flags = (flags & SF_NODISKIO) ? VM_ALLOC_NOWAIT : 0;
429 	sfio->pindex0 = OFF_TO_IDX(off);
430 
431 	/*
432 	 * First grab all the pages and wire them.  Note that we grab
433 	 * only required pages.  Readahead pages are dealt with later.
434 	 */
435 	grabbed = vm_page_grab_pages_unlocked(obj, OFF_TO_IDX(off),
436 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED | flags, pa, npages);
437 	if (grabbed < npages) {
438 		for (int i = grabbed; i < npages; i++)
439 			pa[i] = NULL;
440 		npages = grabbed;
441 		rhpages = 0;
442 	}
443 
444 	for (i = 0; i < npages;) {
445 		/* Skip valid pages. */
446 		if (vm_page_is_valid(pa[i], vmoff(i, off) & PAGE_MASK,
447 		    xfsize(i, npages, off, len))) {
448 			vm_page_xunbusy(pa[i]);
449 			SFSTAT_INC(sf_pages_valid);
450 			i++;
451 			continue;
452 		}
453 
454 		/*
455 		 * Next page is invalid.  Check if it belongs to pager.  It
456 		 * may not be there, which is a regular situation for shmem
457 		 * pager.  For vnode pager this happens only in case of
458 		 * a sparse file.
459 		 *
460 		 * Important feature of vm_pager_has_page() is the hint
461 		 * stored in 'a', about how many pages we can pagein after
462 		 * this page in a single I/O.
463 		 */
464 		VM_OBJECT_RLOCK(obj);
465 		if (!vm_pager_has_page(obj, OFF_TO_IDX(vmoff(i, off)), NULL,
466 		    &a)) {
467 			VM_OBJECT_RUNLOCK(obj);
468 			pmap_zero_page(pa[i]);
469 			vm_page_valid(pa[i]);
470 			MPASS(pa[i]->dirty == 0);
471 			vm_page_xunbusy(pa[i]);
472 			i++;
473 			continue;
474 		}
475 		VM_OBJECT_RUNLOCK(obj);
476 
477 		/*
478 		 * We want to pagein as many pages as possible, limited only
479 		 * by the 'a' hint and actual request.
480 		 */
481 		count = min(a + 1, npages - i);
482 
483 		/*
484 		 * We should not pagein into a valid page because
485 		 * there might be still unfinished write tracked by
486 		 * e.g. a buffer, thus we substitute any valid pages
487 		 * with the bogus one.
488 		 *
489 		 * We must not leave around xbusy pages which are not
490 		 * part of the run passed to vm_pager_getpages(),
491 		 * otherwise pager might deadlock waiting for the busy
492 		 * status of the page, e.g. if it constitues the
493 		 * buffer needed to validate other page.
494 		 *
495 		 * First trim the end of the run consisting of the
496 		 * valid pages, then replace the rest of the valid
497 		 * with bogus.
498 		 */
499 		count1 = count;
500 		for (j = i + count - 1; j > i; j--) {
501 			if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
502 			    xfsize(j, npages, off, len))) {
503 				vm_page_xunbusy(pa[j]);
504 				SFSTAT_INC(sf_pages_valid);
505 				count--;
506 			} else {
507 				break;
508 			}
509 		}
510 
511 		/*
512 		 * The last page in the run pa[i + count - 1] is
513 		 * guaranteed to be invalid by the trim above, so it
514 		 * is not replaced with bogus, thus -1 in the loop end
515 		 * condition.
516 		 */
517 		MPASS(pa[i + count - 1]->valid != VM_PAGE_BITS_ALL);
518 		for (j = i + 1; j < i + count - 1; j++) {
519 			if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
520 			    xfsize(j, npages, off, len))) {
521 				vm_page_xunbusy(pa[j]);
522 				SFSTAT_INC(sf_pages_valid);
523 				SFSTAT_INC(sf_pages_bogus);
524 				pa[j] = bogus_page;
525 			}
526 		}
527 
528 		refcount_acquire(&sfio->nios);
529 		rv = vm_pager_get_pages_async(obj, pa + i, count, NULL,
530 		    i + count == npages ? &rhpages : NULL,
531 		    &sendfile_iodone, sfio);
532 		if (__predict_false(rv != VM_PAGER_OK)) {
533 			sendfile_iowait(sfio, "sferrio");
534 
535 			/*
536 			 * Do remaining pages recovery before returning EIO.
537 			 * Pages from 0 to npages are wired.
538 			 * Pages from (i + count1) to npages are busied.
539 			 */
540 			for (j = 0; j < npages; j++) {
541 				if (j >= i + count1)
542 					vm_page_xunbusy(pa[j]);
543 				KASSERT(pa[j] != NULL && pa[j] != bogus_page,
544 				    ("%s: page %p[%d] I/O recovery failure",
545 				    __func__, pa, j));
546 				vm_page_unwire(pa[j], PQ_INACTIVE);
547 				pa[j] = NULL;
548 			}
549 			return (EIO);
550 		}
551 
552 		SFSTAT_INC(sf_iocnt);
553 		SFSTAT_ADD(sf_pages_read, count);
554 		if (i + count == npages)
555 			SFSTAT_ADD(sf_rhpages_read, rhpages);
556 
557 		i += count1;
558 		(*nios)++;
559 	}
560 
561 	if (*nios == 0 && npages != 0)
562 		SFSTAT_INC(sf_noiocnt);
563 
564 	return (0);
565 }
566 
567 static int
sendfile_getobj(struct thread * td,struct file * fp,vm_object_t * obj_res,struct vnode ** vp_res,struct shmfd ** shmfd_res,off_t * obj_size,int * bsize)568 sendfile_getobj(struct thread *td, struct file *fp, vm_object_t *obj_res,
569     struct vnode **vp_res, struct shmfd **shmfd_res, off_t *obj_size,
570     int *bsize)
571 {
572 	vm_object_t obj;
573 	struct vnode *vp;
574 	struct shmfd *shmfd;
575 	int error;
576 
577 	error = 0;
578 	vp = *vp_res = NULL;
579 	obj = NULL;
580 	shmfd = *shmfd_res = NULL;
581 	*bsize = 0;
582 
583 	/*
584 	 * The file descriptor must be a regular file and have a
585 	 * backing VM object.
586 	 */
587 	if (fp->f_type == DTYPE_VNODE) {
588 		vp = fp->f_vnode;
589 		vn_lock(vp, LK_SHARED | LK_RETRY);
590 		if (vp->v_type != VREG) {
591 			error = EINVAL;
592 			goto out;
593 		}
594 		*bsize = vp->v_mount->mnt_stat.f_iosize;
595 		obj = vp->v_object;
596 		if (obj == NULL) {
597 			error = EINVAL;
598 			goto out;
599 		}
600 
601 		/*
602 		 * Use the pager size when available to simplify synchronization
603 		 * with filesystems, which otherwise must atomically update both
604 		 * the vnode pager size and file size.
605 		 */
606 		if (obj->type == OBJT_VNODE) {
607 			VM_OBJECT_RLOCK(obj);
608 			*obj_size = obj->un_pager.vnp.vnp_size;
609 		} else {
610 			error = vn_getsize_locked(vp, obj_size, td->td_ucred);
611 			if (error != 0)
612 				goto out;
613 			VM_OBJECT_RLOCK(obj);
614 		}
615 	} else if (fp->f_type == DTYPE_SHM) {
616 		shmfd = fp->f_data;
617 		obj = shmfd->shm_object;
618 		VM_OBJECT_RLOCK(obj);
619 		*obj_size = shmfd->shm_size;
620 	} else {
621 		error = EINVAL;
622 		goto out;
623 	}
624 
625 	if ((obj->flags & OBJ_DEAD) != 0) {
626 		VM_OBJECT_RUNLOCK(obj);
627 		error = EBADF;
628 		goto out;
629 	}
630 
631 	/*
632 	 * Temporarily increase the backing VM object's reference
633 	 * count so that a forced reclamation of its vnode does not
634 	 * immediately destroy it.
635 	 */
636 	vm_object_reference_locked(obj);
637 	VM_OBJECT_RUNLOCK(obj);
638 	*obj_res = obj;
639 	*vp_res = vp;
640 	*shmfd_res = shmfd;
641 
642 out:
643 	if (vp != NULL)
644 		VOP_UNLOCK(vp);
645 	return (error);
646 }
647 
648 static int
sendfile_getsock(struct thread * td,int s,struct file ** sock_fp,struct socket ** so)649 sendfile_getsock(struct thread *td, int s, struct file **sock_fp,
650     struct socket **so)
651 {
652 	int error;
653 
654 	*sock_fp = NULL;
655 	*so = NULL;
656 
657 	/*
658 	 * The socket must be a stream socket and connected.
659 	 */
660 	error = getsock(td, s, &cap_send_rights, sock_fp);
661 	if (error != 0)
662 		return (error);
663 	*so = (*sock_fp)->f_data;
664 	if ((*so)->so_type != SOCK_STREAM)
665 		return (EINVAL);
666 	/*
667 	 * SCTP one-to-one style sockets currently don't work with
668 	 * sendfile(). So indicate EINVAL for now.
669 	 */
670 	if ((*so)->so_proto->pr_protocol == IPPROTO_SCTP)
671 		return (EINVAL);
672 	return (0);
673 }
674 
675 /*
676  * Check socket state and wait (or EAGAIN) for needed amount of space.
677  */
678 int
sendfile_wait_generic(struct socket * so,off_t need,int * space)679 sendfile_wait_generic(struct socket *so, off_t need, int *space)
680 {
681 	int error;
682 
683 	MPASS(need > 0);
684 	MPASS(space != NULL);
685 
686 	/*
687 	 * XXXGL: the hack with sb_lowat originates from d99b0dd2c5297.  To
688 	 * achieve high performance sending with sendfile(2) a non-blocking
689 	 * socket needs a fairly high low watermark.  Otherwise, the socket
690 	 * will be reported as writable too early, and sendfile(2) will send
691 	 * just a few bytes each time.  It is important to understand that
692 	 * we are changing sb_lowat not for the current invocation of the
693 	 * syscall, but for the *next* syscall. So there is no way to
694 	 * workaround the problem, e.g. provide a special version of sbspace().
695 	 * Since this hack has been in the kernel for a long time, we
696 	 * anticipate that there is a lot of software that will suffer if we
697 	 * remove it.  See also b21104487324.
698 	 */
699 	error = 0;
700 	SOCK_SENDBUF_LOCK(so);
701 	if (so->so_snd.sb_lowat < so->so_snd.sb_hiwat / 2)
702 		so->so_snd.sb_lowat = so->so_snd.sb_hiwat / 2;
703 	if (so->so_snd.sb_lowat < PAGE_SIZE && so->so_snd.sb_hiwat >= PAGE_SIZE)
704 		so->so_snd.sb_lowat = PAGE_SIZE;
705 retry_space:
706 	if (so->so_snd.sb_state & SBS_CANTSENDMORE) {
707 		error = EPIPE;
708 		goto done;
709 	} else if (so->so_error) {
710 		error = so->so_error;
711 		so->so_error = 0;
712 		goto done;
713 	}
714 	if ((so->so_state & SS_ISCONNECTED) == 0) {
715 		error = ENOTCONN;
716 		goto done;
717 	}
718 
719 	*space = sbspace(&so->so_snd);
720 	if (*space < need && (*space <= 0 || *space < so->so_snd.sb_lowat)) {
721 		if (so->so_state & SS_NBIO) {
722 			error = EAGAIN;
723 			goto done;
724 		}
725 		/*
726 		 * sbwait() drops the lock while sleeping.  When we loop back
727 		 * to retry_space the state may have changed and we retest
728 		 * for it.
729 		 */
730 		error = sbwait(so, SO_SND);
731 		/*
732 		 * An error from sbwait() usually indicates that we've been
733 		 * interrupted by a signal.  If we've sent anything then return
734 		 * bytes sent, otherwise return the error.
735 		 */
736 		if (error != 0)
737 			goto done;
738 		goto retry_space;
739 	}
740 done:
741 	SOCK_SENDBUF_UNLOCK(so);
742 
743 	return (error);
744 }
745 
746 int
vn_sendfile(struct file * fp,int sockfd,struct uio * hdr_uio,struct uio * trl_uio,off_t offset,size_t nbytes,off_t * sent,int flags,struct thread * td)747 vn_sendfile(struct file *fp, int sockfd, struct uio *hdr_uio,
748     struct uio *trl_uio, off_t offset, size_t nbytes, off_t *sent, int flags,
749     struct thread *td)
750 {
751 	struct file *sock_fp;
752 	struct vnode *vp;
753 	struct vm_object *obj;
754 	vm_page_t pga;
755 	struct socket *so;
756 	const struct protosw *pr;
757 #ifdef KERN_TLS
758 	struct ktls_session *tls;
759 #endif
760 	struct mbuf *m, *mh, *mhtail;
761 	struct sf_buf *sf;
762 	struct shmfd *shmfd;
763 	struct sendfile_sync *sfs;
764 	struct vattr va;
765 	off_t off, sbytes, rem, obj_size, nobj_size;
766 	int bsize, error, ext_pgs_idx, hdrlen, max_pgs, softerr;
767 #ifdef KERN_TLS
768 	int tls_enq_cnt;
769 #endif
770 	bool use_ext_pgs;
771 
772 	obj = NULL;
773 	so = NULL;
774 	m = mh = NULL;
775 	sfs = NULL;
776 #ifdef KERN_TLS
777 	tls = NULL;
778 #endif
779 	hdrlen = sbytes = 0;
780 	softerr = 0;
781 	use_ext_pgs = false;
782 
783 	error = sendfile_getobj(td, fp, &obj, &vp, &shmfd, &obj_size, &bsize);
784 	if (error != 0)
785 		return (error);
786 
787 	error = sendfile_getsock(td, sockfd, &sock_fp, &so);
788 	if (error != 0)
789 		goto out;
790 	pr = so->so_proto;
791 
792 #ifdef MAC
793 	error = mac_socket_check_send(td->td_ucred, so);
794 	if (error != 0)
795 		goto out;
796 #endif
797 
798 	SFSTAT_INC(sf_syscalls);
799 	SFSTAT_ADD(sf_rhpages_requested, SF_READAHEAD(flags));
800 
801 	if (__predict_false(flags & SF_SYNC)) {
802 		gone_in(16, "Warning! %s[%u] uses SF_SYNC sendfile(2) flag. "
803 		    "Please follow up to https://bugs.freebsd.org/"
804 		    "bugzilla/show_bug.cgi?id=287348. ",
805 		    td->td_proc->p_comm, td->td_proc->p_pid);
806 		sfs = malloc(sizeof(*sfs), M_SENDFILE, M_WAITOK | M_ZERO);
807 		mtx_init(&sfs->mtx, "sendfile", NULL, MTX_DEF);
808 		cv_init(&sfs->cv, "sendfile");
809 		sfs->waiting = true;
810 	}
811 
812 	rem = nbytes ? omin(nbytes, obj_size - offset) : obj_size - offset;
813 
814 	/*
815 	 * Protect against multiple writers to the socket.
816 	 *
817 	 * XXXRW: Historically this has assumed non-interruptibility, so now
818 	 * we implement that, but possibly shouldn't.
819 	 */
820 	error = SOCK_IO_SEND_LOCK(so, SBL_WAIT | SBL_NOINTR);
821 	if (error != 0)
822 		goto out;
823 	CURVNET_SET(so->so_vnet);
824 #ifdef KERN_TLS
825 	tls = ktls_hold(so->so_snd.sb_tls_info);
826 #endif
827 
828 	/*
829 	 * Loop through the pages of the file, starting with the requested
830 	 * offset. Get a file page (do I/O if necessary), map the file page
831 	 * into an sf_buf, attach an mbuf header to the sf_buf, and queue
832 	 * it on the socket.
833 	 * This is done in two loops.  The inner loop turns as many pages
834 	 * as it can, up to available socket buffer space, without blocking
835 	 * into mbufs to have it bulk delivered into the socket send buffer.
836 	 * The outer loop checks the state and available space of the socket
837 	 * and takes care of the overall progress.
838 	 */
839 	for (off = offset; rem > 0; ) {
840 		struct sf_io *sfio;
841 		vm_page_t *pa;
842 		struct mbuf *m0, *mtail;
843 		int nios, space, npages, rhpages;
844 
845 		mtail = NULL;
846 		if ((error = pr->pr_sendfile_wait(so, rem, &space)) != 0)
847 			goto done;
848 		/*
849 		 * At the beginning of the first loop check if any headers
850 		 * are specified and copy them into mbufs.  Reduce space in
851 		 * the socket buffer by the size of the header mbuf chain.
852 		 * Clear hdr_uio here and hdrlen at the end of the first loop.
853 		 */
854 		if (hdr_uio != NULL && hdr_uio->uio_resid > 0) {
855 			hdr_uio->uio_td = td;
856 			hdr_uio->uio_rw = UIO_WRITE;
857 #ifdef KERN_TLS
858 			if (tls != NULL)
859 				mh = m_uiotombuf(hdr_uio, M_WAITOK, space,
860 				    tls->params.max_frame_len, M_EXTPG);
861 			else
862 #endif
863 				mh = m_uiotombuf(hdr_uio, M_WAITOK,
864 				    space, 0, 0);
865 			hdrlen = m_length(mh, &mhtail);
866 			space -= hdrlen;
867 			/*
868 			 * If header consumed all the socket buffer space,
869 			 * don't waste CPU cycles and jump to the end.
870 			 */
871 			if (space == 0) {
872 				sfio = NULL;
873 				nios = 0;
874 				goto prepend_header;
875 			}
876 			hdr_uio = NULL;
877 		}
878 
879 		if (vp != NULL) {
880 			error = vn_lock(vp, LK_SHARED);
881 			if (error != 0)
882 				goto done;
883 
884 			/*
885 			 * Check to see if the file size has changed.
886 			 */
887 			if (obj->type == OBJT_VNODE) {
888 				VM_OBJECT_RLOCK(obj);
889 				nobj_size = obj->un_pager.vnp.vnp_size;
890 				VM_OBJECT_RUNLOCK(obj);
891 			} else {
892 				error = VOP_GETATTR(vp, &va, td->td_ucred);
893 				if (error != 0) {
894 					VOP_UNLOCK(vp);
895 					goto done;
896 				}
897 				nobj_size = va.va_size;
898 			}
899 			if (off >= nobj_size) {
900 				VOP_UNLOCK(vp);
901 				goto done;
902 			}
903 			if (nobj_size != obj_size) {
904 				obj_size = nobj_size;
905 				rem = nbytes ? omin(nbytes + offset, obj_size) :
906 				    obj_size;
907 				rem -= off;
908 			}
909 		}
910 
911 		if (space > rem)
912 			space = rem;
913 		else if (space > PAGE_SIZE) {
914 			/*
915 			 * Use page boundaries when possible for large
916 			 * requests.
917 			 */
918 			if (off & PAGE_MASK)
919 				space -= (PAGE_SIZE - (off & PAGE_MASK));
920 			space = trunc_page(space);
921 			if (off & PAGE_MASK)
922 				space += (PAGE_SIZE - (off & PAGE_MASK));
923 		}
924 
925 		npages = howmany(space + (off & PAGE_MASK), PAGE_SIZE);
926 
927 		/*
928 		 * Calculate maximum allowed number of pages for readahead
929 		 * at this iteration.  If SF_USER_READAHEAD was set, we don't
930 		 * do any heuristics and use exactly the value supplied by
931 		 * application.  Otherwise, we allow readahead up to "rem".
932 		 * If application wants more, let it be, but there is no
933 		 * reason to go above maxphys.  Also check against "obj_size",
934 		 * since vm_pager_has_page() can hint beyond EOF.
935 		 */
936 		if (flags & SF_USER_READAHEAD) {
937 			rhpages = SF_READAHEAD(flags);
938 		} else {
939 			rhpages = howmany(rem + (off & PAGE_MASK), PAGE_SIZE) -
940 			    npages;
941 			rhpages += SF_READAHEAD(flags);
942 		}
943 		rhpages = min(howmany(maxphys, PAGE_SIZE), rhpages);
944 		rhpages = min(howmany(obj_size - trunc_page(off), PAGE_SIZE) -
945 		    npages, rhpages);
946 
947 		sfio = malloc(sizeof(struct sf_io) +
948 		    npages * sizeof(vm_page_t), M_SENDFILE, M_WAITOK);
949 		refcount_init(&sfio->nios, 1);
950 		sfio->obj = obj;
951 		sfio->error = 0;
952 		sfio->m = NULL;
953 		sfio->npages = npages;
954 #ifdef KERN_TLS
955 		/*
956 		 * This doesn't use ktls_hold() because sfio->m will
957 		 * also have a reference on 'tls' that will be valid
958 		 * for all of sfio's lifetime.
959 		 */
960 		sfio->tls = tls;
961 #endif
962 		vm_object_pip_add(obj, 1);
963 		error = sendfile_swapin(obj, sfio, &nios, off, space, rhpages,
964 		    flags);
965 		if (error != 0) {
966 			if (vp != NULL)
967 				VOP_UNLOCK(vp);
968 			sendfile_iodone(sfio, NULL, 0, error);
969 			goto done;
970 		}
971 
972 		/*
973 		 * Loop and construct maximum sized mbuf chain to be bulk
974 		 * dumped into socket buffer.
975 		 */
976 		pa = sfio->pa;
977 
978 		/*
979 		 * Use unmapped mbufs if enabled for TCP.  Unmapped
980 		 * bufs are restricted to TCP as that is what has been
981 		 * tested.  In particular, unmapped mbufs have not
982 		 * been tested with UNIX-domain sockets.
983 		 *
984 		 * TLS frames always require unmapped mbufs.
985 		 */
986 		if ((mb_use_ext_pgs && pr->pr_protocol == IPPROTO_TCP)
987 #ifdef KERN_TLS
988 		    || tls != NULL
989 #endif
990 		    ) {
991 			use_ext_pgs = true;
992 #ifdef KERN_TLS
993 			if (tls != NULL)
994 				max_pgs = num_pages(tls->params.max_frame_len);
995 			else
996 #endif
997 				max_pgs = MBUF_PEXT_MAX_PGS;
998 
999 			/* Start at last index, to wrap on first use. */
1000 			ext_pgs_idx = max_pgs - 1;
1001 		}
1002 
1003 		for (int i = 0; i < npages; i++) {
1004 			/*
1005 			 * If a page wasn't grabbed successfully, then
1006 			 * trim the array. Can happen only with SF_NODISKIO.
1007 			 */
1008 			if (pa[i] == NULL) {
1009 				SFSTAT_INC(sf_busy);
1010 				fixspace(npages, i, off, &space);
1011 				sfio->npages = i;
1012 				softerr = EBUSY;
1013 				break;
1014 			}
1015 			pga = pa[i];
1016 			if (pga == bogus_page)
1017 				pga = vm_page_relookup(obj, sfio->pindex0 + i);
1018 
1019 			if (use_ext_pgs) {
1020 				off_t xfs;
1021 
1022 				ext_pgs_idx++;
1023 				if (ext_pgs_idx == max_pgs) {
1024 					m0 = mb_alloc_ext_pgs(M_WAITOK,
1025 					    sendfile_free_mext_pg, M_RDONLY);
1026 
1027 					if (flags & SF_NOCACHE) {
1028 						m0->m_ext.ext_flags |=
1029 						    EXT_FLAG_NOCACHE;
1030 
1031 						/*
1032 						 * See comment below regarding
1033 						 * ignoring SF_NOCACHE for the
1034 						 * last page.
1035 						 */
1036 						if ((npages - i <= max_pgs) &&
1037 						    ((off + space) & PAGE_MASK) &&
1038 						    (rem > space || rhpages > 0))
1039 							m0->m_ext.ext_flags |=
1040 							    EXT_FLAG_CACHE_LAST;
1041 					}
1042 					if (sfs != NULL) {
1043 						m0->m_ext.ext_flags |=
1044 						    EXT_FLAG_SYNC;
1045 						m0->m_ext.ext_arg1 = sfs;
1046 						mtx_lock(&sfs->mtx);
1047 						sfs->count++;
1048 						mtx_unlock(&sfs->mtx);
1049 					}
1050 					ext_pgs_idx = 0;
1051 
1052 					/* Append to mbuf chain. */
1053 					if (mtail != NULL)
1054 						mtail->m_next = m0;
1055 					else
1056 						m = m0;
1057 					mtail = m0;
1058 					m0->m_epg_1st_off =
1059 					    vmoff(i, off) & PAGE_MASK;
1060 				}
1061 				if (nios) {
1062 					mtail->m_flags |= M_NOTREADY;
1063 					m0->m_epg_nrdy++;
1064 				}
1065 
1066 				m0->m_epg_pa[ext_pgs_idx] = VM_PAGE_TO_PHYS(pga);
1067 				m0->m_epg_npgs++;
1068 				xfs = xfsize(i, npages, off, space);
1069 				m0->m_epg_last_len = xfs;
1070 				MBUF_EXT_PGS_ASSERT_SANITY(m0);
1071 				mtail->m_len += xfs;
1072 				mtail->m_ext.ext_size += PAGE_SIZE;
1073 				continue;
1074 			}
1075 
1076 			/*
1077 			 * Get a sendfile buf.  When allocating the
1078 			 * first buffer for mbuf chain, we usually
1079 			 * wait as long as necessary, but this wait
1080 			 * can be interrupted.  For consequent
1081 			 * buffers, do not sleep, since several
1082 			 * threads might exhaust the buffers and then
1083 			 * deadlock.
1084 			 */
1085 			sf = sf_buf_alloc(pga,
1086 			    m != NULL ? SFB_NOWAIT : SFB_CATCH);
1087 			if (sf == NULL) {
1088 				SFSTAT_INC(sf_allocfail);
1089 				sendfile_iowait(sfio, "sfnosf");
1090 				for (int j = i; j < npages; j++) {
1091 					vm_page_unwire(pa[j], PQ_INACTIVE);
1092 					pa[j] = NULL;
1093 				}
1094 				if (m == NULL)
1095 					softerr = ENOBUFS;
1096 				fixspace(npages, i, off, &space);
1097 				sfio->npages = i;
1098 				break;
1099 			}
1100 
1101 			m0 = m_get(M_WAITOK, MT_DATA);
1102 			m0->m_ext.ext_buf = (char *)sf_buf_kva(sf);
1103 			m0->m_ext.ext_size = PAGE_SIZE;
1104 			m0->m_ext.ext_arg1 = sf;
1105 			m0->m_ext.ext_type = EXT_SFBUF;
1106 			m0->m_ext.ext_flags = EXT_FLAG_EMBREF;
1107 			m0->m_ext.ext_free = sendfile_free_mext;
1108 			/*
1109 			 * SF_NOCACHE sets the page as being freed upon send.
1110 			 * However, we ignore it for the last page in 'space',
1111 			 * if the page is truncated, and we got more data to
1112 			 * send (rem > space), or if we have readahead
1113 			 * configured (rhpages > 0).
1114 			 */
1115 			if ((flags & SF_NOCACHE) &&
1116 			    (i != npages - 1 ||
1117 			    !((off + space) & PAGE_MASK) ||
1118 			    !(rem > space || rhpages > 0)))
1119 				m0->m_ext.ext_flags |= EXT_FLAG_NOCACHE;
1120 			if (sfs != NULL) {
1121 				m0->m_ext.ext_flags |= EXT_FLAG_SYNC;
1122 				m0->m_ext.ext_arg2 = sfs;
1123 				mtx_lock(&sfs->mtx);
1124 				sfs->count++;
1125 				mtx_unlock(&sfs->mtx);
1126 			}
1127 			m0->m_ext.ext_count = 1;
1128 			m0->m_flags |= (M_EXT | M_RDONLY);
1129 			if (nios)
1130 				m0->m_flags |= M_NOTREADY;
1131 			m0->m_data = (char *)sf_buf_kva(sf) +
1132 			    (vmoff(i, off) & PAGE_MASK);
1133 			m0->m_len = xfsize(i, npages, off, space);
1134 
1135 			/* Append to mbuf chain. */
1136 			if (mtail != NULL)
1137 				mtail->m_next = m0;
1138 			else
1139 				m = m0;
1140 			mtail = m0;
1141 		}
1142 
1143 		if (vp != NULL)
1144 			VOP_UNLOCK(vp);
1145 
1146 		/* Keep track of bytes processed. */
1147 		off += space;
1148 		rem -= space;
1149 
1150 		/*
1151 		 * Prepend header, if any.  Save pointer to first mbuf
1152 		 * with a page.
1153 		 */
1154 		if (hdrlen) {
1155 prepend_header:
1156 			m0 = mhtail->m_next = m;
1157 			m = mh;
1158 			mh = NULL;
1159 		} else
1160 			m0 = m;
1161 
1162 		if (m == NULL) {
1163 			KASSERT(softerr, ("%s: m NULL, no error", __func__));
1164 			error = softerr;
1165 			sendfile_iodone(sfio, NULL, 0, 0);
1166 			goto done;
1167 		}
1168 
1169 		/* Add the buffer chain to the socket buffer. */
1170 		KASSERT(m_length(m, NULL) == space + hdrlen,
1171 		    ("%s: mlen %u space %d hdrlen %d",
1172 		    __func__, m_length(m, NULL), space, hdrlen));
1173 
1174 #ifdef KERN_TLS
1175 		if (tls != NULL)
1176 			ktls_frame(m, tls, &tls_enq_cnt, TLS_RLTYPE_APP);
1177 #endif
1178 		if (nios == 0) {
1179 			/*
1180 			 * If sendfile_swapin() didn't initiate any I/Os,
1181 			 * which happens if all data is cached in VM, or if
1182 			 * the header consumed all socket buffer space and
1183 			 * sfio is NULL, then we can send data right now
1184 			 * without the PRUS_NOTREADY flag.
1185 			 */
1186 			if (sfio != NULL)
1187 				sendfile_iodone(sfio, NULL, 0, 0);
1188 #ifdef KERN_TLS
1189 			if (tls != NULL && tls->mode == TCP_TLS_MODE_SW) {
1190 				error = pr->pr_send(so, PRUS_NOTREADY, m, NULL,
1191 				    NULL, td);
1192 				if (error != 0) {
1193 					m_freem(m);
1194 				} else {
1195 					soref(so);
1196 					ktls_enqueue(m, so, tls_enq_cnt);
1197 				}
1198 			} else
1199 #endif
1200 				error = pr->pr_send(so, 0, m, NULL, NULL, td);
1201 		} else {
1202 			sfio->so = so;
1203 			sfio->m = m0;
1204 			soref(so);
1205 			error = pr->pr_send(so, PRUS_NOTREADY, m, NULL, NULL,
1206 			    td);
1207 			sendfile_iodone(sfio, NULL, 0, error);
1208 		}
1209 #ifdef TCP_REQUEST_TRK
1210 		if (so->so_proto->pr_protocol == IPPROTO_TCP) {
1211 			/* log the sendfile call to the TCP log, if enabled */
1212 			tcp_log_sendfile(so, offset, nbytes, flags);
1213 		}
1214 #endif
1215 		m = NULL;
1216 		if (error)
1217 			goto done;
1218 		sbytes += space + hdrlen;
1219 		if (hdrlen)
1220 			hdrlen = 0;
1221 		if (softerr) {
1222 			error = softerr;
1223 			goto done;
1224 		}
1225 	}
1226 
1227 	/*
1228 	 * Send trailers. Wimp out and use writev(2).
1229 	 */
1230 	if (trl_uio != NULL) {
1231 		SOCK_IO_SEND_UNLOCK(so);
1232 		CURVNET_RESTORE();
1233 		error = kern_writev(td, sockfd, trl_uio);
1234 		if (error == 0)
1235 			sbytes += td->td_retval[0];
1236 		goto out;
1237 	}
1238 
1239 done:
1240 	SOCK_IO_SEND_UNLOCK(so);
1241 	CURVNET_RESTORE();
1242 out:
1243 	/*
1244 	 * If there was no error we have to clear td->td_retval[0]
1245 	 * because it may have been set by writev.
1246 	 */
1247 	if (error == 0) {
1248 		td->td_retval[0] = 0;
1249 	}
1250 	if (sent != NULL) {
1251 		(*sent) = sbytes;
1252 	}
1253 	if (obj != NULL)
1254 		vm_object_deallocate(obj);
1255 	if (so)
1256 		fdrop(sock_fp, td);
1257 	if (m)
1258 		m_freem(m);
1259 	if (mh)
1260 		m_freem(mh);
1261 
1262 	if (sfs != NULL) {
1263 		mtx_lock(&sfs->mtx);
1264 		if (sfs->count != 0)
1265 			error = cv_wait_sig(&sfs->cv, &sfs->mtx);
1266 		if (sfs->count == 0) {
1267 			sendfile_sync_destroy(sfs);
1268 		} else {
1269 			sfs->waiting = false;
1270 			mtx_unlock(&sfs->mtx);
1271 		}
1272 	}
1273 #ifdef KERN_TLS
1274 	if (tls != NULL)
1275 		ktls_free(tls);
1276 #endif
1277 	if (error == ERESTART)
1278 		error = EINTR;
1279 
1280 	return (error);
1281 }
1282 
1283 static int
sendfile(struct thread * td,struct sendfile_args * uap,int compat)1284 sendfile(struct thread *td, struct sendfile_args *uap, int compat)
1285 {
1286 	struct sf_hdtr hdtr;
1287 	struct uio *hdr_uio, *trl_uio;
1288 	struct file *fp;
1289 	off_t sbytes;
1290 	int error;
1291 
1292 	/*
1293 	 * File offset must be positive.  If it goes beyond EOF
1294 	 * we send only the header/trailer and no payload data.
1295 	 */
1296 	if (uap->offset < 0)
1297 		return (EINVAL);
1298 
1299 	sbytes = 0;
1300 	hdr_uio = trl_uio = NULL;
1301 
1302 	if (uap->hdtr != NULL) {
1303 		error = copyin(uap->hdtr, &hdtr, sizeof(hdtr));
1304 		if (error != 0)
1305 			goto out;
1306 		if (hdtr.headers != NULL) {
1307 			error = copyinuio(hdtr.headers, hdtr.hdr_cnt,
1308 			    &hdr_uio);
1309 			if (error != 0)
1310 				goto out;
1311 #ifdef COMPAT_FREEBSD4
1312 			/*
1313 			 * In FreeBSD < 5.0 the nbytes to send also included
1314 			 * the header.  If compat is specified subtract the
1315 			 * header size from nbytes.
1316 			 */
1317 			if (compat) {
1318 				if (uap->nbytes > hdr_uio->uio_resid)
1319 					uap->nbytes -= hdr_uio->uio_resid;
1320 				else
1321 					uap->nbytes = 0;
1322 			}
1323 #endif
1324 		}
1325 		if (hdtr.trailers != NULL) {
1326 			error = copyinuio(hdtr.trailers, hdtr.trl_cnt,
1327 			    &trl_uio);
1328 			if (error != 0)
1329 				goto out;
1330 		}
1331 	}
1332 
1333 	AUDIT_ARG_FD(uap->fd);
1334 
1335 	/*
1336 	 * sendfile(2) can start at any offset within a file so we require
1337 	 * CAP_READ+CAP_SEEK = CAP_PREAD.
1338 	 */
1339 	if ((error = fget_read(td, uap->fd, &cap_pread_rights, &fp)) != 0)
1340 		goto out;
1341 
1342 	error = fo_sendfile(fp, uap->s, hdr_uio, trl_uio, uap->offset,
1343 	    uap->nbytes, &sbytes, uap->flags, td);
1344 	fdrop(fp, td);
1345 
1346 	if (uap->sbytes != NULL)
1347 		(void)copyout(&sbytes, uap->sbytes, sizeof(off_t));
1348 
1349 out:
1350 	freeuio(hdr_uio);
1351 	freeuio(trl_uio);
1352 	return (error);
1353 }
1354 
1355 /*
1356  * sendfile(2)
1357  *
1358  * int sendfile(int fd, int s, off_t offset, size_t nbytes,
1359  *       struct sf_hdtr *hdtr, off_t *sbytes, int flags)
1360  *
1361  * Send a file specified by 'fd' and starting at 'offset' to a socket
1362  * specified by 's'. Send only 'nbytes' of the file or until EOF if nbytes ==
1363  * 0.  Optionally add a header and/or trailer to the socket output.  If
1364  * specified, write the total number of bytes sent into *sbytes.
1365  */
1366 int
sys_sendfile(struct thread * td,struct sendfile_args * uap)1367 sys_sendfile(struct thread *td, struct sendfile_args *uap)
1368 {
1369 
1370 	return (sendfile(td, uap, 0));
1371 }
1372 
1373 #ifdef COMPAT_FREEBSD4
1374 int
freebsd4_sendfile(struct thread * td,struct freebsd4_sendfile_args * uap)1375 freebsd4_sendfile(struct thread *td, struct freebsd4_sendfile_args *uap)
1376 {
1377 	struct sendfile_args args;
1378 
1379 	args.fd = uap->fd;
1380 	args.s = uap->s;
1381 	args.offset = uap->offset;
1382 	args.nbytes = uap->nbytes;
1383 	args.hdtr = uap->hdtr;
1384 	args.sbytes = uap->sbytes;
1385 	args.flags = uap->flags;
1386 
1387 	return (sendfile(td, &args, 1));
1388 }
1389 #endif /* COMPAT_FREEBSD4 */
1390