xref: /freebsd-12.1/sys/kern/kern_sendfile.c (revision 8cb553dc)
1 /*-
2  * Copyright (c) 2013-2015 Gleb Smirnoff <[email protected]>
3  * Copyright (c) 1998, David Greenman. All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  * 3. Neither the name of the University nor the names of its contributors
14  *    may be used to endorse or promote products derived from this software
15  *    without specific prior written permission.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
32 
33 #include <sys/param.h>
34 #include <sys/systm.h>
35 #include <sys/capsicum.h>
36 #include <sys/kernel.h>
37 #include <sys/lock.h>
38 #include <sys/mutex.h>
39 #include <sys/sysproto.h>
40 #include <sys/malloc.h>
41 #include <sys/proc.h>
42 #include <sys/mman.h>
43 #include <sys/mount.h>
44 #include <sys/mbuf.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/vnode.h>
53 
54 #include <net/vnet.h>
55 
56 #include <security/audit/audit.h>
57 #include <security/mac/mac_framework.h>
58 
59 #include <vm/vm.h>
60 #include <vm/vm_object.h>
61 #include <vm/vm_pager.h>
62 
63 #define	EXT_FLAG_SYNC		EXT_FLAG_VENDOR1
64 #define	EXT_FLAG_NOCACHE	EXT_FLAG_VENDOR2
65 
66 /*
67  * Structure describing a single sendfile(2) I/O, which may consist of
68  * several underlying pager I/Os.
69  *
70  * The syscall context allocates the structure and initializes 'nios'
71  * to 1.  As sendfile_swapin() runs through pages and starts asynchronous
72  * paging operations, it increments 'nios'.
73  *
74  * Every I/O completion calls sendfile_iodone(), which decrements the 'nios',
75  * and the syscall also calls sendfile_iodone() after allocating all mbufs,
76  * linking them and sending to socket.  Whoever reaches zero 'nios' is
77  * responsible to * call pru_ready on the socket, to notify it of readyness
78  * of the data.
79  */
80 struct sf_io {
81 	volatile u_int	nios;
82 	u_int		error;
83 	int		npages;
84 	struct socket	*so;
85 	struct mbuf	*m;
86 	vm_page_t	pa[];
87 };
88 
89 /*
90  * Structure used to track requests with SF_SYNC flag.
91  */
92 struct sendfile_sync {
93 	struct mtx	mtx;
94 	struct cv	cv;
95 	unsigned	count;
96 };
97 
98 counter_u64_t sfstat[sizeof(struct sfstat) / sizeof(uint64_t)];
99 
100 static void
sfstat_init(const void * unused)101 sfstat_init(const void *unused)
102 {
103 
104 	COUNTER_ARRAY_ALLOC(sfstat, sizeof(struct sfstat) / sizeof(uint64_t),
105 	    M_WAITOK);
106 }
107 SYSINIT(sfstat, SI_SUB_MBUF, SI_ORDER_FIRST, sfstat_init, NULL);
108 
109 static int
sfstat_sysctl(SYSCTL_HANDLER_ARGS)110 sfstat_sysctl(SYSCTL_HANDLER_ARGS)
111 {
112 	struct sfstat s;
113 
114 	COUNTER_ARRAY_COPY(sfstat, &s, sizeof(s) / sizeof(uint64_t));
115 	if (req->newptr)
116 		COUNTER_ARRAY_ZERO(sfstat, sizeof(s) / sizeof(uint64_t));
117 	return (SYSCTL_OUT(req, &s, sizeof(s)));
118 }
119 SYSCTL_PROC(_kern_ipc, OID_AUTO, sfstat, CTLTYPE_OPAQUE | CTLFLAG_RW,
120     NULL, 0, sfstat_sysctl, "I", "sendfile statistics");
121 
122 static void
sendfile_free_mext(struct mbuf * m)123 sendfile_free_mext(struct mbuf *m)
124 {
125 	struct sf_buf *sf;
126 	vm_page_t pg;
127 	int flags;
128 
129 	KASSERT(m->m_flags & M_EXT && m->m_ext.ext_type == EXT_SFBUF,
130 	    ("%s: m %p !M_EXT or !EXT_SFBUF", __func__, m));
131 
132 	sf = m->m_ext.ext_arg1;
133 	pg = sf_buf_page(sf);
134 	flags = (m->m_ext.ext_flags & EXT_FLAG_NOCACHE) != 0 ? VPR_TRYFREE : 0;
135 
136 	sf_buf_free(sf);
137 	vm_page_release(pg, flags);
138 
139 	if (m->m_ext.ext_flags & EXT_FLAG_SYNC) {
140 		struct sendfile_sync *sfs = m->m_ext.ext_arg2;
141 
142 		mtx_lock(&sfs->mtx);
143 		KASSERT(sfs->count > 0, ("Sendfile sync botchup count == 0"));
144 		if (--sfs->count == 0)
145 			cv_signal(&sfs->cv);
146 		mtx_unlock(&sfs->mtx);
147 	}
148 }
149 
150 /*
151  * Helper function to calculate how much data to put into page i of n.
152  * Only first and last pages are special.
153  */
154 static inline off_t
xfsize(int i,int n,off_t off,off_t len)155 xfsize(int i, int n, off_t off, off_t len)
156 {
157 
158 	if (i == 0)
159 		return (omin(PAGE_SIZE - (off & PAGE_MASK), len));
160 
161 	if (i == n - 1 && ((off + len) & PAGE_MASK) > 0)
162 		return ((off + len) & PAGE_MASK);
163 
164 	return (PAGE_SIZE);
165 }
166 
167 /*
168  * Helper function to get offset within object for i page.
169  */
170 static inline vm_ooffset_t
vmoff(int i,off_t off)171 vmoff(int i, off_t off)
172 {
173 
174 	if (i == 0)
175 		return ((vm_ooffset_t)off);
176 
177 	return (trunc_page(off + i * PAGE_SIZE));
178 }
179 
180 /*
181  * Helper function used when allocation of a page or sf_buf failed.
182  * Pretend as if we don't have enough space, subtract xfsize() of
183  * all pages that failed.
184  */
185 static inline void
fixspace(int old,int new,off_t off,int * space)186 fixspace(int old, int new, off_t off, int *space)
187 {
188 
189 	KASSERT(old > new, ("%s: old %d new %d", __func__, old, new));
190 
191 	/* Subtract last one. */
192 	*space -= xfsize(old - 1, old, off, *space);
193 	old--;
194 
195 	if (new == old)
196 		/* There was only one page. */
197 		return;
198 
199 	/* Subtract first one. */
200 	if (new == 0) {
201 		*space -= xfsize(0, old, off, *space);
202 		new++;
203 	}
204 
205 	/* Rest of pages are full sized. */
206 	*space -= (old - new) * PAGE_SIZE;
207 
208 	KASSERT(*space >= 0, ("%s: space went backwards", __func__));
209 }
210 
211 /*
212  * I/O completion callback.
213  */
214 static void
sendfile_iodone(void * arg,vm_page_t * pg,int count,int error)215 sendfile_iodone(void *arg, vm_page_t *pg, int count, int error)
216 {
217 	struct sf_io *sfio = arg;
218 	struct socket *so = sfio->so;
219 
220 	for (int i = 0; i < count; i++)
221 		if (pg[i] != bogus_page)
222 			vm_page_xunbusy(pg[i]);
223 
224 	if (error)
225 		sfio->error = error;
226 
227 	if (!refcount_release(&sfio->nios))
228 		return;
229 
230 	CURVNET_SET(so->so_vnet);
231 	if (sfio->error) {
232 		struct mbuf *m;
233 
234 		/*
235 		 * I/O operation failed.  The state of data in the socket
236 		 * is now inconsistent, and all what we can do is to tear
237 		 * it down. Protocol abort method would tear down protocol
238 		 * state, free all ready mbufs and detach not ready ones.
239 		 * We will free the mbufs corresponding to this I/O manually.
240 		 *
241 		 * The socket would be marked with EIO and made available
242 		 * for read, so that application receives EIO on next
243 		 * syscall and eventually closes the socket.
244 		 */
245 		so->so_proto->pr_usrreqs->pru_abort(so);
246 		so->so_error = EIO;
247 
248 		m = sfio->m;
249 		for (int i = 0; i < sfio->npages; i++)
250 			m = m_free(m);
251 	} else
252 		(void )(so->so_proto->pr_usrreqs->pru_ready)(so, sfio->m,
253 		    sfio->npages);
254 
255 	SOCK_LOCK(so);
256 	sorele(so);
257 	CURVNET_RESTORE();
258 	free(sfio, M_TEMP);
259 }
260 
261 /*
262  * Iterate through pages vector and request paging for non-valid pages.
263  */
264 static int
sendfile_swapin(vm_object_t obj,struct sf_io * sfio,int * nios,off_t off,off_t len,int npages,int rhpages,int flags)265 sendfile_swapin(vm_object_t obj, struct sf_io *sfio, int *nios, off_t off,
266     off_t len, int npages, int rhpages, int flags)
267 {
268 	vm_page_t *pa = sfio->pa;
269 	int grabbed;
270 
271 	*nios = 0;
272 	flags = (flags & SF_NODISKIO) ? VM_ALLOC_NOWAIT : 0;
273 
274 	/*
275 	 * First grab all the pages and wire them.  Note that we grab
276 	 * only required pages.  Readahead pages are dealt with later.
277 	 */
278 	VM_OBJECT_WLOCK(obj);
279 
280 	grabbed = vm_page_grab_pages(obj, OFF_TO_IDX(off),
281 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED | flags, pa, npages);
282 	if (grabbed < npages) {
283 		for (int i = grabbed; i < npages; i++)
284 			pa[i] = NULL;
285 		npages = grabbed;
286 		rhpages = 0;
287 	}
288 
289 	for (int i = 0; i < npages;) {
290 		int j, a, count, rv;
291 
292 		/* Skip valid pages. */
293 		if (vm_page_is_valid(pa[i], vmoff(i, off) & PAGE_MASK,
294 		    xfsize(i, npages, off, len))) {
295 			vm_page_xunbusy(pa[i]);
296 			SFSTAT_INC(sf_pages_valid);
297 			i++;
298 			continue;
299 		}
300 
301 		/*
302 		 * Next page is invalid.  Check if it belongs to pager.  It
303 		 * may not be there, which is a regular situation for shmem
304 		 * pager.  For vnode pager this happens only in case of
305 		 * a sparse file.
306 		 *
307 		 * Important feature of vm_pager_has_page() is the hint
308 		 * stored in 'a', about how many pages we can pagein after
309 		 * this page in a single I/O.
310 		 */
311 		if (!vm_pager_has_page(obj, OFF_TO_IDX(vmoff(i, off)), NULL,
312 		    &a)) {
313 			pmap_zero_page(pa[i]);
314 			pa[i]->valid = VM_PAGE_BITS_ALL;
315 			MPASS(pa[i]->dirty == 0);
316 			vm_page_xunbusy(pa[i]);
317 			i++;
318 			continue;
319 		}
320 
321 		/*
322 		 * We want to pagein as many pages as possible, limited only
323 		 * by the 'a' hint and actual request.
324 		 */
325 		count = min(a + 1, npages - i);
326 
327 		/*
328 		 * We should not pagein into a valid page, thus we first trim
329 		 * any valid pages off the end of request, and substitute
330 		 * to bogus_page those, that are in the middle.
331 		 */
332 		for (j = i + count - 1; j > i; j--) {
333 			if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
334 			    xfsize(j, npages, off, len))) {
335 				count--;
336 				rhpages = 0;
337 			} else
338 				break;
339 		}
340 		for (j = i + 1; j < i + count - 1; j++)
341 			if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
342 			    xfsize(j, npages, off, len))) {
343 				vm_page_xunbusy(pa[j]);
344 				SFSTAT_INC(sf_pages_valid);
345 				SFSTAT_INC(sf_pages_bogus);
346 				pa[j] = bogus_page;
347 			}
348 
349 		refcount_acquire(&sfio->nios);
350 		rv = vm_pager_get_pages_async(obj, pa + i, count, NULL,
351 		    i + count == npages ? &rhpages : NULL,
352 		    &sendfile_iodone, sfio);
353 		if (rv != VM_PAGER_OK) {
354 			for (j = i; j < i + count; j++) {
355 				if (pa[j] != bogus_page) {
356 					vm_page_lock(pa[j]);
357 					vm_page_unwire(pa[j], PQ_INACTIVE);
358 					vm_page_unlock(pa[j]);
359 				}
360 			}
361 			VM_OBJECT_WUNLOCK(obj);
362 			return (EIO);
363 		}
364 		KASSERT(rv == VM_PAGER_OK, ("%s: pager fail obj %p page %p",
365 		    __func__, obj, pa[i]));
366 
367 		SFSTAT_INC(sf_iocnt);
368 		SFSTAT_ADD(sf_pages_read, count);
369 		if (i + count == npages)
370 			SFSTAT_ADD(sf_rhpages_read, rhpages);
371 
372 		/*
373 		 * Restore the valid page pointers.  They are already
374 		 * unbusied, but still wired.
375 		 */
376 		for (j = i; j < i + count; j++)
377 			if (pa[j] == bogus_page) {
378 				pa[j] = vm_page_lookup(obj,
379 				    OFF_TO_IDX(vmoff(j, off)));
380 				KASSERT(pa[j], ("%s: page %p[%d] disappeared",
381 				    __func__, pa, j));
382 
383 			}
384 		i += count;
385 		(*nios)++;
386 	}
387 
388 	VM_OBJECT_WUNLOCK(obj);
389 
390 	if (*nios == 0 && npages != 0)
391 		SFSTAT_INC(sf_noiocnt);
392 
393 	return (0);
394 }
395 
396 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)397 sendfile_getobj(struct thread *td, struct file *fp, vm_object_t *obj_res,
398     struct vnode **vp_res, struct shmfd **shmfd_res, off_t *obj_size,
399     int *bsize)
400 {
401 	struct vattr va;
402 	vm_object_t obj;
403 	struct vnode *vp;
404 	struct shmfd *shmfd;
405 	int error;
406 
407 	vp = *vp_res = NULL;
408 	obj = NULL;
409 	shmfd = *shmfd_res = NULL;
410 	*bsize = 0;
411 
412 	/*
413 	 * The file descriptor must be a regular file and have a
414 	 * backing VM object.
415 	 */
416 	if (fp->f_type == DTYPE_VNODE) {
417 		vp = fp->f_vnode;
418 		vn_lock(vp, LK_SHARED | LK_RETRY);
419 		if (vp->v_type != VREG) {
420 			error = EINVAL;
421 			goto out;
422 		}
423 		*bsize = vp->v_mount->mnt_stat.f_iosize;
424 		error = VOP_GETATTR(vp, &va, td->td_ucred);
425 		if (error != 0)
426 			goto out;
427 		*obj_size = va.va_size;
428 		obj = vp->v_object;
429 		if (obj == NULL) {
430 			error = EINVAL;
431 			goto out;
432 		}
433 	} else if (fp->f_type == DTYPE_SHM) {
434 		error = 0;
435 		shmfd = fp->f_data;
436 		obj = shmfd->shm_object;
437 		*obj_size = shmfd->shm_size;
438 	} else {
439 		error = EINVAL;
440 		goto out;
441 	}
442 
443 	VM_OBJECT_WLOCK(obj);
444 	if ((obj->flags & OBJ_DEAD) != 0) {
445 		VM_OBJECT_WUNLOCK(obj);
446 		error = EBADF;
447 		goto out;
448 	}
449 
450 	/*
451 	 * Temporarily increase the backing VM object's reference
452 	 * count so that a forced reclamation of its vnode does not
453 	 * immediately destroy it.
454 	 */
455 	vm_object_reference_locked(obj);
456 	VM_OBJECT_WUNLOCK(obj);
457 	*obj_res = obj;
458 	*vp_res = vp;
459 	*shmfd_res = shmfd;
460 
461 out:
462 	if (vp != NULL)
463 		VOP_UNLOCK(vp, 0);
464 	return (error);
465 }
466 
467 static int
sendfile_getsock(struct thread * td,int s,struct file ** sock_fp,struct socket ** so)468 sendfile_getsock(struct thread *td, int s, struct file **sock_fp,
469     struct socket **so)
470 {
471 	int error;
472 
473 	*sock_fp = NULL;
474 	*so = NULL;
475 
476 	/*
477 	 * The socket must be a stream socket and connected.
478 	 */
479 	error = getsock_cap(td, s, &cap_send_rights,
480 	    sock_fp, NULL, NULL);
481 	if (error != 0)
482 		return (error);
483 	*so = (*sock_fp)->f_data;
484 	if ((*so)->so_type != SOCK_STREAM)
485 		return (EINVAL);
486 	if (SOLISTENING(*so))
487 		return (ENOTCONN);
488 	return (0);
489 }
490 
491 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)492 vn_sendfile(struct file *fp, int sockfd, struct uio *hdr_uio,
493     struct uio *trl_uio, off_t offset, size_t nbytes, off_t *sent, int flags,
494     struct thread *td)
495 {
496 	struct file *sock_fp;
497 	struct vnode *vp;
498 	struct vm_object *obj;
499 	struct socket *so;
500 	struct mbuf *m, *mh, *mhtail;
501 	struct sf_buf *sf;
502 	struct shmfd *shmfd;
503 	struct sendfile_sync *sfs;
504 	struct vattr va;
505 	off_t off, sbytes, rem, obj_size;
506 	int error, softerr, bsize, hdrlen;
507 
508 	obj = NULL;
509 	so = NULL;
510 	m = mh = NULL;
511 	sfs = NULL;
512 	hdrlen = sbytes = 0;
513 	softerr = 0;
514 
515 	error = sendfile_getobj(td, fp, &obj, &vp, &shmfd, &obj_size, &bsize);
516 	if (error != 0)
517 		return (error);
518 
519 	error = sendfile_getsock(td, sockfd, &sock_fp, &so);
520 	if (error != 0)
521 		goto out;
522 
523 #ifdef MAC
524 	error = mac_socket_check_send(td->td_ucred, so);
525 	if (error != 0)
526 		goto out;
527 #endif
528 
529 	SFSTAT_INC(sf_syscalls);
530 	SFSTAT_ADD(sf_rhpages_requested, SF_READAHEAD(flags));
531 
532 	if (flags & SF_SYNC) {
533 		sfs = malloc(sizeof *sfs, M_TEMP, M_WAITOK | M_ZERO);
534 		mtx_init(&sfs->mtx, "sendfile", NULL, MTX_DEF);
535 		cv_init(&sfs->cv, "sendfile");
536 	}
537 
538 	rem = nbytes ? omin(nbytes, obj_size - offset) : obj_size - offset;
539 
540 	/*
541 	 * Protect against multiple writers to the socket.
542 	 *
543 	 * XXXRW: Historically this has assumed non-interruptibility, so now
544 	 * we implement that, but possibly shouldn't.
545 	 */
546 	(void)sblock(&so->so_snd, SBL_WAIT | SBL_NOINTR);
547 
548 	/*
549 	 * Loop through the pages of the file, starting with the requested
550 	 * offset. Get a file page (do I/O if necessary), map the file page
551 	 * into an sf_buf, attach an mbuf header to the sf_buf, and queue
552 	 * it on the socket.
553 	 * This is done in two loops.  The inner loop turns as many pages
554 	 * as it can, up to available socket buffer space, without blocking
555 	 * into mbufs to have it bulk delivered into the socket send buffer.
556 	 * The outer loop checks the state and available space of the socket
557 	 * and takes care of the overall progress.
558 	 */
559 	for (off = offset; rem > 0; ) {
560 		struct sf_io *sfio;
561 		vm_page_t *pa;
562 		struct mbuf *mtail;
563 		int nios, space, npages, rhpages;
564 
565 		mtail = NULL;
566 		/*
567 		 * Check the socket state for ongoing connection,
568 		 * no errors and space in socket buffer.
569 		 * If space is low allow for the remainder of the
570 		 * file to be processed if it fits the socket buffer.
571 		 * Otherwise block in waiting for sufficient space
572 		 * to proceed, or if the socket is nonblocking, return
573 		 * to userland with EAGAIN while reporting how far
574 		 * we've come.
575 		 * We wait until the socket buffer has significant free
576 		 * space to do bulk sends.  This makes good use of file
577 		 * system read ahead and allows packet segmentation
578 		 * offloading hardware to take over lots of work.  If
579 		 * we were not careful here we would send off only one
580 		 * sfbuf at a time.
581 		 */
582 		SOCKBUF_LOCK(&so->so_snd);
583 		if (so->so_snd.sb_lowat < so->so_snd.sb_hiwat / 2)
584 			so->so_snd.sb_lowat = so->so_snd.sb_hiwat / 2;
585 retry_space:
586 		if (so->so_snd.sb_state & SBS_CANTSENDMORE) {
587 			error = EPIPE;
588 			SOCKBUF_UNLOCK(&so->so_snd);
589 			goto done;
590 		} else if (so->so_error) {
591 			error = so->so_error;
592 			so->so_error = 0;
593 			SOCKBUF_UNLOCK(&so->so_snd);
594 			goto done;
595 		}
596 		if ((so->so_state & SS_ISCONNECTED) == 0) {
597 			SOCKBUF_UNLOCK(&so->so_snd);
598 			error = ENOTCONN;
599 			goto done;
600 		}
601 
602 		space = sbspace(&so->so_snd);
603 		if (space < rem &&
604 		    (space <= 0 ||
605 		     space < so->so_snd.sb_lowat)) {
606 			if (so->so_state & SS_NBIO) {
607 				SOCKBUF_UNLOCK(&so->so_snd);
608 				error = EAGAIN;
609 				goto done;
610 			}
611 			/*
612 			 * sbwait drops the lock while sleeping.
613 			 * When we loop back to retry_space the
614 			 * state may have changed and we retest
615 			 * for it.
616 			 */
617 			error = sbwait(&so->so_snd);
618 			/*
619 			 * An error from sbwait usually indicates that we've
620 			 * been interrupted by a signal. If we've sent anything
621 			 * then return bytes sent, otherwise return the error.
622 			 */
623 			if (error != 0) {
624 				SOCKBUF_UNLOCK(&so->so_snd);
625 				goto done;
626 			}
627 			goto retry_space;
628 		}
629 		SOCKBUF_UNLOCK(&so->so_snd);
630 
631 		/*
632 		 * At the beginning of the first loop check if any headers
633 		 * are specified and copy them into mbufs.  Reduce space in
634 		 * the socket buffer by the size of the header mbuf chain.
635 		 * Clear hdr_uio here and hdrlen at the end of the first loop.
636 		 */
637 		if (hdr_uio != NULL && hdr_uio->uio_resid > 0) {
638 			hdr_uio->uio_td = td;
639 			hdr_uio->uio_rw = UIO_WRITE;
640 			mh = m_uiotombuf(hdr_uio, M_WAITOK, space, 0, 0);
641 			hdrlen = m_length(mh, &mhtail);
642 			space -= hdrlen;
643 			/*
644 			 * If header consumed all the socket buffer space,
645 			 * don't waste CPU cycles and jump to the end.
646 			 */
647 			if (space == 0) {
648 				sfio = NULL;
649 				nios = 0;
650 				goto prepend_header;
651 			}
652 			hdr_uio = NULL;
653 		}
654 
655 		if (vp != NULL) {
656 			error = vn_lock(vp, LK_SHARED);
657 			if (error != 0)
658 				goto done;
659 			error = VOP_GETATTR(vp, &va, td->td_ucred);
660 			if (error != 0 || off >= va.va_size) {
661 				VOP_UNLOCK(vp, 0);
662 				goto done;
663 			}
664 			if (va.va_size != obj_size) {
665 				obj_size = va.va_size;
666 				rem = nbytes ?
667 				    omin(nbytes + offset, obj_size) : obj_size;
668 				rem -= off;
669 			}
670 		}
671 
672 		if (space > rem)
673 			space = rem;
674 
675 		npages = howmany(space + (off & PAGE_MASK), PAGE_SIZE);
676 
677 		/*
678 		 * Calculate maximum allowed number of pages for readahead
679 		 * at this iteration.  If SF_USER_READAHEAD was set, we don't
680 		 * do any heuristics and use exactly the value supplied by
681 		 * application.  Otherwise, we allow readahead up to "rem".
682 		 * If application wants more, let it be, but there is no
683 		 * reason to go above MAXPHYS.  Also check against "obj_size",
684 		 * since vm_pager_has_page() can hint beyond EOF.
685 		 */
686 		if (flags & SF_USER_READAHEAD) {
687 			rhpages = SF_READAHEAD(flags);
688 		} else {
689 			rhpages = howmany(rem + (off & PAGE_MASK), PAGE_SIZE) -
690 			    npages;
691 			rhpages += SF_READAHEAD(flags);
692 		}
693 		rhpages = min(howmany(MAXPHYS, PAGE_SIZE), rhpages);
694 		rhpages = min(howmany(obj_size - trunc_page(off), PAGE_SIZE) -
695 		    npages, rhpages);
696 
697 		sfio = malloc(sizeof(struct sf_io) +
698 		    npages * sizeof(vm_page_t), M_TEMP, M_WAITOK);
699 		refcount_init(&sfio->nios, 1);
700 		sfio->so = so;
701 		sfio->error = 0;
702 
703 		error = sendfile_swapin(obj, sfio, &nios, off, space, npages,
704 		    rhpages, flags);
705 		if (error != 0) {
706 			if (vp != NULL)
707 				VOP_UNLOCK(vp, 0);
708 			free(sfio, M_TEMP);
709 			goto done;
710 		}
711 
712 		/*
713 		 * Loop and construct maximum sized mbuf chain to be bulk
714 		 * dumped into socket buffer.
715 		 */
716 		pa = sfio->pa;
717 		for (int i = 0; i < npages; i++) {
718 			struct mbuf *m0;
719 
720 			/*
721 			 * If a page wasn't grabbed successfully, then
722 			 * trim the array. Can happen only with SF_NODISKIO.
723 			 */
724 			if (pa[i] == NULL) {
725 				SFSTAT_INC(sf_busy);
726 				fixspace(npages, i, off, &space);
727 				npages = i;
728 				softerr = EBUSY;
729 				break;
730 			}
731 
732 			/*
733 			 * Get a sendfile buf.  When allocating the
734 			 * first buffer for mbuf chain, we usually
735 			 * wait as long as necessary, but this wait
736 			 * can be interrupted.  For consequent
737 			 * buffers, do not sleep, since several
738 			 * threads might exhaust the buffers and then
739 			 * deadlock.
740 			 */
741 			sf = sf_buf_alloc(pa[i],
742 			    m != NULL ? SFB_NOWAIT : SFB_CATCH);
743 			if (sf == NULL) {
744 				SFSTAT_INC(sf_allocfail);
745 				for (int j = i; j < npages; j++) {
746 					vm_page_lock(pa[j]);
747 					vm_page_unwire(pa[j], PQ_INACTIVE);
748 					vm_page_unlock(pa[j]);
749 				}
750 				if (m == NULL)
751 					softerr = ENOBUFS;
752 				fixspace(npages, i, off, &space);
753 				npages = i;
754 				break;
755 			}
756 
757 			m0 = m_get(M_WAITOK, MT_DATA);
758 			m0->m_ext.ext_buf = (char *)sf_buf_kva(sf);
759 			m0->m_ext.ext_size = PAGE_SIZE;
760 			m0->m_ext.ext_arg1 = sf;
761 			m0->m_ext.ext_type = EXT_SFBUF;
762 			m0->m_ext.ext_flags = EXT_FLAG_EMBREF;
763 			m0->m_ext.ext_free = sendfile_free_mext;
764 			/*
765 			 * SF_NOCACHE sets the page as being freed upon send.
766 			 * However, we ignore it for the last page in 'space',
767 			 * if the page is truncated, and we got more data to
768 			 * send (rem > space), or if we have readahead
769 			 * configured (rhpages > 0).
770 			 */
771 			if ((flags & SF_NOCACHE) &&
772 			    (i != npages - 1 ||
773 			    !((off + space) & PAGE_MASK) ||
774 			    !(rem > space || rhpages > 0)))
775 				m0->m_ext.ext_flags |= EXT_FLAG_NOCACHE;
776 			if (sfs != NULL) {
777 				m0->m_ext.ext_flags |= EXT_FLAG_SYNC;
778 				m0->m_ext.ext_arg2 = sfs;
779 				mtx_lock(&sfs->mtx);
780 				sfs->count++;
781 				mtx_unlock(&sfs->mtx);
782 			}
783 			m0->m_ext.ext_count = 1;
784 			m0->m_flags |= (M_EXT | M_RDONLY);
785 			if (nios)
786 				m0->m_flags |= M_NOTREADY;
787 			m0->m_data = (char *)sf_buf_kva(sf) +
788 			    (vmoff(i, off) & PAGE_MASK);
789 			m0->m_len = xfsize(i, npages, off, space);
790 
791 			if (i == 0)
792 				sfio->m = m0;
793 
794 			/* Append to mbuf chain. */
795 			if (mtail != NULL)
796 				mtail->m_next = m0;
797 			else
798 				m = m0;
799 			mtail = m0;
800 		}
801 
802 		if (vp != NULL)
803 			VOP_UNLOCK(vp, 0);
804 
805 		/* Keep track of bytes processed. */
806 		off += space;
807 		rem -= space;
808 
809 		/* Prepend header, if any. */
810 		if (hdrlen) {
811 prepend_header:
812 			mhtail->m_next = m;
813 			m = mh;
814 			mh = NULL;
815 		}
816 
817 		if (m == NULL) {
818 			KASSERT(softerr, ("%s: m NULL, no error", __func__));
819 			error = softerr;
820 			free(sfio, M_TEMP);
821 			goto done;
822 		}
823 
824 		/* Add the buffer chain to the socket buffer. */
825 		KASSERT(m_length(m, NULL) == space + hdrlen,
826 		    ("%s: mlen %u space %d hdrlen %d",
827 		    __func__, m_length(m, NULL), space, hdrlen));
828 
829 		CURVNET_SET(so->so_vnet);
830 		if (nios == 0) {
831 			/*
832 			 * If sendfile_swapin() didn't initiate any I/Os,
833 			 * which happens if all data is cached in VM, then
834 			 * we can send data right now without the
835 			 * PRUS_NOTREADY flag.
836 			 */
837 			free(sfio, M_TEMP);
838 			error = (*so->so_proto->pr_usrreqs->pru_send)
839 			    (so, 0, m, NULL, NULL, td);
840 		} else {
841 			sfio->npages = npages;
842 			soref(so);
843 			error = (*so->so_proto->pr_usrreqs->pru_send)
844 			    (so, PRUS_NOTREADY, m, NULL, NULL, td);
845 			sendfile_iodone(sfio, NULL, 0, 0);
846 		}
847 		CURVNET_RESTORE();
848 
849 		m = NULL;	/* pru_send always consumes */
850 		if (error)
851 			goto done;
852 		sbytes += space + hdrlen;
853 		if (hdrlen)
854 			hdrlen = 0;
855 		if (softerr) {
856 			error = softerr;
857 			goto done;
858 		}
859 	}
860 
861 	/*
862 	 * Send trailers. Wimp out and use writev(2).
863 	 */
864 	if (trl_uio != NULL) {
865 		sbunlock(&so->so_snd);
866 		error = kern_writev(td, sockfd, trl_uio);
867 		if (error == 0)
868 			sbytes += td->td_retval[0];
869 		goto out;
870 	}
871 
872 done:
873 	sbunlock(&so->so_snd);
874 out:
875 	/*
876 	 * If there was no error we have to clear td->td_retval[0]
877 	 * because it may have been set by writev.
878 	 */
879 	if (error == 0) {
880 		td->td_retval[0] = 0;
881 	}
882 	if (sent != NULL) {
883 		(*sent) = sbytes;
884 	}
885 	if (obj != NULL)
886 		vm_object_deallocate(obj);
887 	if (so)
888 		fdrop(sock_fp, td);
889 	if (m)
890 		m_freem(m);
891 	if (mh)
892 		m_freem(mh);
893 
894 	if (sfs != NULL) {
895 		mtx_lock(&sfs->mtx);
896 		if (sfs->count != 0)
897 			cv_wait(&sfs->cv, &sfs->mtx);
898 		KASSERT(sfs->count == 0, ("sendfile sync still busy"));
899 		cv_destroy(&sfs->cv);
900 		mtx_destroy(&sfs->mtx);
901 		free(sfs, M_TEMP);
902 	}
903 
904 	if (error == ERESTART)
905 		error = EINTR;
906 
907 	return (error);
908 }
909 
910 static int
sendfile(struct thread * td,struct sendfile_args * uap,int compat)911 sendfile(struct thread *td, struct sendfile_args *uap, int compat)
912 {
913 	struct sf_hdtr hdtr;
914 	struct uio *hdr_uio, *trl_uio;
915 	struct file *fp;
916 	off_t sbytes;
917 	int error;
918 
919 	/*
920 	 * File offset must be positive.  If it goes beyond EOF
921 	 * we send only the header/trailer and no payload data.
922 	 */
923 	if (uap->offset < 0)
924 		return (EINVAL);
925 
926 	sbytes = 0;
927 	hdr_uio = trl_uio = NULL;
928 
929 	if (uap->hdtr != NULL) {
930 		error = copyin(uap->hdtr, &hdtr, sizeof(hdtr));
931 		if (error != 0)
932 			goto out;
933 		if (hdtr.headers != NULL) {
934 			error = copyinuio(hdtr.headers, hdtr.hdr_cnt,
935 			    &hdr_uio);
936 			if (error != 0)
937 				goto out;
938 #ifdef COMPAT_FREEBSD4
939 			/*
940 			 * In FreeBSD < 5.0 the nbytes to send also included
941 			 * the header.  If compat is specified subtract the
942 			 * header size from nbytes.
943 			 */
944 			if (compat) {
945 				if (uap->nbytes > hdr_uio->uio_resid)
946 					uap->nbytes -= hdr_uio->uio_resid;
947 				else
948 					uap->nbytes = 0;
949 			}
950 #endif
951 		}
952 		if (hdtr.trailers != NULL) {
953 			error = copyinuio(hdtr.trailers, hdtr.trl_cnt,
954 			    &trl_uio);
955 			if (error != 0)
956 				goto out;
957 		}
958 	}
959 
960 	AUDIT_ARG_FD(uap->fd);
961 
962 	/*
963 	 * sendfile(2) can start at any offset within a file so we require
964 	 * CAP_READ+CAP_SEEK = CAP_PREAD.
965 	 */
966 	if ((error = fget_read(td, uap->fd, &cap_pread_rights, &fp)) != 0)
967 		goto out;
968 
969 	error = fo_sendfile(fp, uap->s, hdr_uio, trl_uio, uap->offset,
970 	    uap->nbytes, &sbytes, uap->flags, td);
971 	fdrop(fp, td);
972 
973 	if (uap->sbytes != NULL)
974 		copyout(&sbytes, uap->sbytes, sizeof(off_t));
975 
976 out:
977 	free(hdr_uio, M_IOV);
978 	free(trl_uio, M_IOV);
979 	return (error);
980 }
981 
982 /*
983  * sendfile(2)
984  *
985  * int sendfile(int fd, int s, off_t offset, size_t nbytes,
986  *       struct sf_hdtr *hdtr, off_t *sbytes, int flags)
987  *
988  * Send a file specified by 'fd' and starting at 'offset' to a socket
989  * specified by 's'. Send only 'nbytes' of the file or until EOF if nbytes ==
990  * 0.  Optionally add a header and/or trailer to the socket output.  If
991  * specified, write the total number of bytes sent into *sbytes.
992  */
993 int
sys_sendfile(struct thread * td,struct sendfile_args * uap)994 sys_sendfile(struct thread *td, struct sendfile_args *uap)
995 {
996 
997 	return (sendfile(td, uap, 0));
998 }
999 
1000 #ifdef COMPAT_FREEBSD4
1001 int
freebsd4_sendfile(struct thread * td,struct freebsd4_sendfile_args * uap)1002 freebsd4_sendfile(struct thread *td, struct freebsd4_sendfile_args *uap)
1003 {
1004 	struct sendfile_args args;
1005 
1006 	args.fd = uap->fd;
1007 	args.s = uap->s;
1008 	args.offset = uap->offset;
1009 	args.nbytes = uap->nbytes;
1010 	args.hdtr = uap->hdtr;
1011 	args.sbytes = uap->sbytes;
1012 	args.flags = uap->flags;
1013 
1014 	return (sendfile(td, &args, 1));
1015 }
1016 #endif /* COMPAT_FREEBSD4 */
1017