xref: /freebsd/sys/kern/kern_sendfile.c (revision 8c27c5541e95164a37ac1beb7bd4ef43429aed55)
1 /*-
2  * Copyright (c) 2013-2015 Gleb Smirnoff <glebius@FreeBSD.org>
3  * Copyright (c) 1998, David Greenman. All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  * 3. Neither the name of the University nor the names of its contributors
14  *    may be used to endorse or promote products derived from this software
15  *    without specific prior written permission.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
32 
33 #include "opt_kern_tls.h"
34 
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/capsicum.h>
38 #include <sys/kernel.h>
39 #include <netinet/in.h>
40 #include <sys/lock.h>
41 #include <sys/ktls.h>
42 #include <sys/mutex.h>
43 #include <sys/sysproto.h>
44 #include <sys/malloc.h>
45 #include <sys/proc.h>
46 #include <sys/mman.h>
47 #include <sys/mount.h>
48 #include <sys/mbuf.h>
49 #include <sys/protosw.h>
50 #include <sys/rwlock.h>
51 #include <sys/sf_buf.h>
52 #include <sys/socket.h>
53 #include <sys/socketvar.h>
54 #include <sys/syscallsubr.h>
55 #include <sys/sysctl.h>
56 #include <sys/vnode.h>
57 
58 #include <net/vnet.h>
59 #include <netinet/tcp.h>
60 
61 #include <security/audit/audit.h>
62 #include <security/mac/mac_framework.h>
63 
64 #include <vm/vm.h>
65 #include <vm/vm_object.h>
66 #include <vm/vm_pager.h>
67 
68 #define	EXT_FLAG_SYNC		EXT_FLAG_VENDOR1
69 #define	EXT_FLAG_NOCACHE	EXT_FLAG_VENDOR2
70 #define	EXT_FLAG_CACHE_LAST	EXT_FLAG_VENDOR3
71 
72 /*
73  * Structure describing a single sendfile(2) I/O, which may consist of
74  * several underlying pager I/Os.
75  *
76  * The syscall context allocates the structure and initializes 'nios'
77  * to 1.  As sendfile_swapin() runs through pages and starts asynchronous
78  * paging operations, it increments 'nios'.
79  *
80  * Every I/O completion calls sendfile_iodone(), which decrements the 'nios',
81  * and the syscall also calls sendfile_iodone() after allocating all mbufs,
82  * linking them and sending to socket.  Whoever reaches zero 'nios' is
83  * responsible to * call pru_ready on the socket, to notify it of readyness
84  * of the data.
85  */
86 struct sf_io {
87 	volatile u_int	nios;
88 	u_int		error;
89 	int		npages;
90 	struct socket	*so;
91 	struct mbuf	*m;
92 	vm_object_t	obj;
93 #ifdef KERN_TLS
94 	struct ktls_session *tls;
95 #endif
96 	vm_page_t	pa[];
97 };
98 
99 /*
100  * Structure used to track requests with SF_SYNC flag.
101  */
102 struct sendfile_sync {
103 	struct mtx	mtx;
104 	struct cv	cv;
105 	unsigned	count;
106 };
107 
108 counter_u64_t sfstat[sizeof(struct sfstat) / sizeof(uint64_t)];
109 
110 static void
111 sfstat_init(const void *unused)
112 {
113 
114 	COUNTER_ARRAY_ALLOC(sfstat, sizeof(struct sfstat) / sizeof(uint64_t),
115 	    M_WAITOK);
116 }
117 SYSINIT(sfstat, SI_SUB_MBUF, SI_ORDER_FIRST, sfstat_init, NULL);
118 
119 static int
120 sfstat_sysctl(SYSCTL_HANDLER_ARGS)
121 {
122 	struct sfstat s;
123 
124 	COUNTER_ARRAY_COPY(sfstat, &s, sizeof(s) / sizeof(uint64_t));
125 	if (req->newptr)
126 		COUNTER_ARRAY_ZERO(sfstat, sizeof(s) / sizeof(uint64_t));
127 	return (SYSCTL_OUT(req, &s, sizeof(s)));
128 }
129 SYSCTL_PROC(_kern_ipc, OID_AUTO, sfstat,
130     CTLTYPE_OPAQUE | CTLFLAG_RW | CTLFLAG_NEEDGIANT, NULL, 0,
131     sfstat_sysctl, "I",
132     "sendfile statistics");
133 
134 static void
135 sendfile_free_mext(struct mbuf *m)
136 {
137 	struct sf_buf *sf;
138 	vm_page_t pg;
139 	int flags;
140 
141 	KASSERT(m->m_flags & M_EXT && m->m_ext.ext_type == EXT_SFBUF,
142 	    ("%s: m %p !M_EXT or !EXT_SFBUF", __func__, m));
143 
144 	sf = m->m_ext.ext_arg1;
145 	pg = sf_buf_page(sf);
146 	flags = (m->m_ext.ext_flags & EXT_FLAG_NOCACHE) != 0 ? VPR_TRYFREE : 0;
147 
148 	sf_buf_free(sf);
149 	vm_page_release(pg, flags);
150 
151 	if (m->m_ext.ext_flags & EXT_FLAG_SYNC) {
152 		struct sendfile_sync *sfs = m->m_ext.ext_arg2;
153 
154 		mtx_lock(&sfs->mtx);
155 		KASSERT(sfs->count > 0, ("Sendfile sync botchup count == 0"));
156 		if (--sfs->count == 0)
157 			cv_signal(&sfs->cv);
158 		mtx_unlock(&sfs->mtx);
159 	}
160 }
161 
162 static void
163 sendfile_free_mext_pg(struct mbuf *m)
164 {
165 	struct mbuf_ext_pgs *ext_pgs;
166 	vm_page_t pg;
167 	int flags, i;
168 	bool cache_last;
169 
170 	KASSERT(m->m_flags & M_EXT && m->m_ext.ext_type == EXT_PGS,
171 	    ("%s: m %p !M_EXT or !EXT_PGS", __func__, m));
172 
173 	cache_last = m->m_ext.ext_flags & EXT_FLAG_CACHE_LAST;
174 	ext_pgs = m->m_ext.ext_pgs;
175 	flags = (m->m_ext.ext_flags & EXT_FLAG_NOCACHE) != 0 ? VPR_TRYFREE : 0;
176 
177 	for (i = 0; i < ext_pgs->npgs; i++) {
178 		if (cache_last && i == ext_pgs->npgs - 1)
179 			flags = 0;
180 		pg = PHYS_TO_VM_PAGE(ext_pgs->pa[i]);
181 		vm_page_release(pg, flags);
182 	}
183 
184 	if (m->m_ext.ext_flags & EXT_FLAG_SYNC) {
185 		struct sendfile_sync *sfs = m->m_ext.ext_arg2;
186 
187 		mtx_lock(&sfs->mtx);
188 		KASSERT(sfs->count > 0, ("Sendfile sync botchup count == 0"));
189 		if (--sfs->count == 0)
190 			cv_signal(&sfs->cv);
191 		mtx_unlock(&sfs->mtx);
192 	}
193 }
194 
195 /*
196  * Helper function to calculate how much data to put into page i of n.
197  * Only first and last pages are special.
198  */
199 static inline off_t
200 xfsize(int i, int n, off_t off, off_t len)
201 {
202 
203 	if (i == 0)
204 		return (omin(PAGE_SIZE - (off & PAGE_MASK), len));
205 
206 	if (i == n - 1 && ((off + len) & PAGE_MASK) > 0)
207 		return ((off + len) & PAGE_MASK);
208 
209 	return (PAGE_SIZE);
210 }
211 
212 /*
213  * Helper function to get offset within object for i page.
214  */
215 static inline vm_ooffset_t
216 vmoff(int i, off_t off)
217 {
218 
219 	if (i == 0)
220 		return ((vm_ooffset_t)off);
221 
222 	return (trunc_page(off + i * PAGE_SIZE));
223 }
224 
225 /*
226  * Helper function used when allocation of a page or sf_buf failed.
227  * Pretend as if we don't have enough space, subtract xfsize() of
228  * all pages that failed.
229  */
230 static inline void
231 fixspace(int old, int new, off_t off, int *space)
232 {
233 
234 	KASSERT(old > new, ("%s: old %d new %d", __func__, old, new));
235 
236 	/* Subtract last one. */
237 	*space -= xfsize(old - 1, old, off, *space);
238 	old--;
239 
240 	if (new == old)
241 		/* There was only one page. */
242 		return;
243 
244 	/* Subtract first one. */
245 	if (new == 0) {
246 		*space -= xfsize(0, old, off, *space);
247 		new++;
248 	}
249 
250 	/* Rest of pages are full sized. */
251 	*space -= (old - new) * PAGE_SIZE;
252 
253 	KASSERT(*space >= 0, ("%s: space went backwards", __func__));
254 }
255 
256 /*
257  * I/O completion callback.
258  */
259 static void
260 sendfile_iodone(void *arg, vm_page_t *pg, int count, int error)
261 {
262 	struct sf_io *sfio = arg;
263 	struct socket *so;
264 
265 	for (int i = 0; i < count; i++)
266 		if (pg[i] != bogus_page)
267 			vm_page_xunbusy_unchecked(pg[i]);
268 
269 	if (error)
270 		sfio->error = error;
271 
272 	if (!refcount_release(&sfio->nios))
273 		return;
274 
275 	vm_object_pip_wakeup(sfio->obj);
276 
277 	if (sfio->m == NULL) {
278 		/*
279 		 * Either I/O operation failed, or we failed to allocate
280 		 * buffers, or we bailed out on first busy page, or we
281 		 * succeeded filling the request without any I/Os. Anyway,
282 		 * pru_send hadn't been executed - nothing had been sent
283 		 * to the socket yet.
284 		 */
285 		MPASS((curthread->td_pflags & TDP_KTHREAD) == 0);
286 		free(sfio, M_TEMP);
287 		return;
288 	}
289 
290 #if defined(KERN_TLS) && defined(INVARIANTS)
291 	if ((sfio->m->m_flags & M_EXT) != 0 &&
292 	    sfio->m->m_ext.ext_type == EXT_PGS)
293 		KASSERT(sfio->tls == sfio->m->m_ext.ext_pgs->tls,
294 		    ("TLS session mismatch"));
295 	else
296 		KASSERT(sfio->tls == NULL,
297 		    ("non-ext_pgs mbuf with TLS session"));
298 #endif
299 	so = sfio->so;
300 	CURVNET_SET(so->so_vnet);
301 	if (__predict_false(sfio->error)) {
302 		/*
303 		 * I/O operation failed.  The state of data in the socket
304 		 * is now inconsistent, and all what we can do is to tear
305 		 * it down. Protocol abort method would tear down protocol
306 		 * state, free all ready mbufs and detach not ready ones.
307 		 * We will free the mbufs corresponding to this I/O manually.
308 		 *
309 		 * The socket would be marked with EIO and made available
310 		 * for read, so that application receives EIO on next
311 		 * syscall and eventually closes the socket.
312 		 */
313 		so->so_proto->pr_usrreqs->pru_abort(so);
314 		so->so_error = EIO;
315 
316 		mb_free_notready(sfio->m, sfio->npages);
317 #ifdef KERN_TLS
318 	} else if (sfio->tls != NULL && sfio->tls->mode == TCP_TLS_MODE_SW) {
319 		/*
320 		 * I/O operation is complete, but we still need to
321 		 * encrypt.  We cannot do this in the interrupt thread
322 		 * of the disk controller, so forward the mbufs to a
323 		 * different thread.
324 		 *
325 		 * Donate the socket reference from sfio to rather
326 		 * than explicitly invoking soref().
327 		 */
328 		ktls_enqueue(sfio->m, so, sfio->npages);
329 		goto out_with_ref;
330 #endif
331 	} else
332 		(void)(so->so_proto->pr_usrreqs->pru_ready)(so, sfio->m,
333 		    sfio->npages);
334 
335 	SOCK_LOCK(so);
336 	sorele(so);
337 #ifdef KERN_TLS
338 out_with_ref:
339 #endif
340 	CURVNET_RESTORE();
341 	free(sfio, M_TEMP);
342 }
343 
344 /*
345  * Iterate through pages vector and request paging for non-valid pages.
346  */
347 static int
348 sendfile_swapin(vm_object_t obj, struct sf_io *sfio, int *nios, off_t off,
349     off_t len, int npages, int rhpages, int flags)
350 {
351 	vm_page_t *pa = sfio->pa;
352 	int grabbed;
353 
354 	*nios = 0;
355 	flags = (flags & SF_NODISKIO) ? VM_ALLOC_NOWAIT : 0;
356 
357 	/*
358 	 * First grab all the pages and wire them.  Note that we grab
359 	 * only required pages.  Readahead pages are dealt with later.
360 	 */
361 	grabbed = vm_page_grab_pages_unlocked(obj, OFF_TO_IDX(off),
362 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED | flags, pa, npages);
363 	if (grabbed < npages) {
364 		for (int i = grabbed; i < npages; i++)
365 			pa[i] = NULL;
366 		npages = grabbed;
367 		rhpages = 0;
368 	}
369 
370 	for (int i = 0; i < npages;) {
371 		int j, a, count, rv;
372 
373 		/* Skip valid pages. */
374 		if (vm_page_is_valid(pa[i], vmoff(i, off) & PAGE_MASK,
375 		    xfsize(i, npages, off, len))) {
376 			vm_page_xunbusy(pa[i]);
377 			SFSTAT_INC(sf_pages_valid);
378 			i++;
379 			continue;
380 		}
381 
382 		/*
383 		 * Next page is invalid.  Check if it belongs to pager.  It
384 		 * may not be there, which is a regular situation for shmem
385 		 * pager.  For vnode pager this happens only in case of
386 		 * a sparse file.
387 		 *
388 		 * Important feature of vm_pager_has_page() is the hint
389 		 * stored in 'a', about how many pages we can pagein after
390 		 * this page in a single I/O.
391 		 */
392 		VM_OBJECT_RLOCK(obj);
393 		if (!vm_pager_has_page(obj, OFF_TO_IDX(vmoff(i, off)), NULL,
394 		    &a)) {
395 			VM_OBJECT_RUNLOCK(obj);
396 			pmap_zero_page(pa[i]);
397 			vm_page_valid(pa[i]);
398 			MPASS(pa[i]->dirty == 0);
399 			vm_page_xunbusy(pa[i]);
400 			i++;
401 			continue;
402 		}
403 		VM_OBJECT_RUNLOCK(obj);
404 
405 		/*
406 		 * We want to pagein as many pages as possible, limited only
407 		 * by the 'a' hint and actual request.
408 		 */
409 		count = min(a + 1, npages - i);
410 
411 		/*
412 		 * We should not pagein into a valid page, thus we first trim
413 		 * any valid pages off the end of request, and substitute
414 		 * to bogus_page those, that are in the middle.
415 		 */
416 		for (j = i + count - 1; j > i; j--) {
417 			if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
418 			    xfsize(j, npages, off, len))) {
419 				count--;
420 				rhpages = 0;
421 			} else
422 				break;
423 		}
424 		for (j = i + 1; j < i + count - 1; j++)
425 			if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK,
426 			    xfsize(j, npages, off, len))) {
427 				vm_page_xunbusy(pa[j]);
428 				SFSTAT_INC(sf_pages_valid);
429 				SFSTAT_INC(sf_pages_bogus);
430 				pa[j] = bogus_page;
431 			}
432 
433 		refcount_acquire(&sfio->nios);
434 		rv = vm_pager_get_pages_async(obj, pa + i, count, NULL,
435 		    i + count == npages ? &rhpages : NULL,
436 		    &sendfile_iodone, sfio);
437 		if (__predict_false(rv != VM_PAGER_OK)) {
438 			/*
439 			 * Perform full pages recovery before returning EIO.
440 			 * Pages from 0 to npages are wired.
441 			 * Pages from i to npages are also busied.
442 			 * Pages from (i + 1) to (i + count - 1) may be
443 			 * substituted to bogus page, and not busied.
444 			 */
445 			for (j = 0; j < npages; j++) {
446 				if (j > i && j < i + count - 1 &&
447 				    pa[j] == bogus_page)
448 					pa[j] = vm_page_relookup(obj,
449 					    OFF_TO_IDX(vmoff(j, off)));
450 				else if (j >= i)
451 					vm_page_xunbusy(pa[j]);
452 				KASSERT(pa[j] != NULL && pa[j] != bogus_page,
453 				    ("%s: page %p[%d] I/O recovery failure",
454 				    __func__, pa, j));
455 				vm_page_unwire(pa[j], PQ_INACTIVE);
456 			}
457 			refcount_release(&sfio->nios);
458 			return (EIO);
459 		}
460 
461 		SFSTAT_INC(sf_iocnt);
462 		SFSTAT_ADD(sf_pages_read, count);
463 		if (i + count == npages)
464 			SFSTAT_ADD(sf_rhpages_read, rhpages);
465 
466 		/*
467 		 * Restore the valid page pointers.  They are already
468 		 * unbusied, but still wired.
469 		 */
470 		for (j = i + 1; j < i + count - 1; j++)
471 			if (pa[j] == bogus_page) {
472 				pa[j] = vm_page_relookup(obj,
473 				    OFF_TO_IDX(vmoff(j, off)));
474 				KASSERT(pa[j], ("%s: page %p[%d] disappeared",
475 				    __func__, pa, j));
476 
477 			}
478 		i += count;
479 		(*nios)++;
480 	}
481 
482 	if (*nios == 0 && npages != 0)
483 		SFSTAT_INC(sf_noiocnt);
484 
485 	return (0);
486 }
487 
488 static int
489 sendfile_getobj(struct thread *td, struct file *fp, vm_object_t *obj_res,
490     struct vnode **vp_res, struct shmfd **shmfd_res, off_t *obj_size,
491     int *bsize)
492 {
493 	struct vattr va;
494 	vm_object_t obj;
495 	struct vnode *vp;
496 	struct shmfd *shmfd;
497 	int error;
498 
499 	vp = *vp_res = NULL;
500 	obj = NULL;
501 	shmfd = *shmfd_res = NULL;
502 	*bsize = 0;
503 
504 	/*
505 	 * The file descriptor must be a regular file and have a
506 	 * backing VM object.
507 	 */
508 	if (fp->f_type == DTYPE_VNODE) {
509 		vp = fp->f_vnode;
510 		vn_lock(vp, LK_SHARED | LK_RETRY);
511 		if (vp->v_type != VREG) {
512 			error = EINVAL;
513 			goto out;
514 		}
515 		*bsize = vp->v_mount->mnt_stat.f_iosize;
516 		error = VOP_GETATTR(vp, &va, td->td_ucred);
517 		if (error != 0)
518 			goto out;
519 		*obj_size = va.va_size;
520 		obj = vp->v_object;
521 		if (obj == NULL) {
522 			error = EINVAL;
523 			goto out;
524 		}
525 	} else if (fp->f_type == DTYPE_SHM) {
526 		error = 0;
527 		shmfd = fp->f_data;
528 		obj = shmfd->shm_object;
529 		*obj_size = shmfd->shm_size;
530 	} else {
531 		error = EINVAL;
532 		goto out;
533 	}
534 
535 	VM_OBJECT_WLOCK(obj);
536 	if ((obj->flags & OBJ_DEAD) != 0) {
537 		VM_OBJECT_WUNLOCK(obj);
538 		error = EBADF;
539 		goto out;
540 	}
541 
542 	/*
543 	 * Temporarily increase the backing VM object's reference
544 	 * count so that a forced reclamation of its vnode does not
545 	 * immediately destroy it.
546 	 */
547 	vm_object_reference_locked(obj);
548 	VM_OBJECT_WUNLOCK(obj);
549 	*obj_res = obj;
550 	*vp_res = vp;
551 	*shmfd_res = shmfd;
552 
553 out:
554 	if (vp != NULL)
555 		VOP_UNLOCK(vp);
556 	return (error);
557 }
558 
559 static int
560 sendfile_getsock(struct thread *td, int s, struct file **sock_fp,
561     struct socket **so)
562 {
563 	int error;
564 
565 	*sock_fp = NULL;
566 	*so = NULL;
567 
568 	/*
569 	 * The socket must be a stream socket and connected.
570 	 */
571 	error = getsock_cap(td, s, &cap_send_rights,
572 	    sock_fp, NULL, NULL);
573 	if (error != 0)
574 		return (error);
575 	*so = (*sock_fp)->f_data;
576 	if ((*so)->so_type != SOCK_STREAM)
577 		return (EINVAL);
578 	if (SOLISTENING(*so))
579 		return (ENOTCONN);
580 	return (0);
581 }
582 
583 int
584 vn_sendfile(struct file *fp, int sockfd, struct uio *hdr_uio,
585     struct uio *trl_uio, off_t offset, size_t nbytes, off_t *sent, int flags,
586     struct thread *td)
587 {
588 	struct file *sock_fp;
589 	struct vnode *vp;
590 	struct vm_object *obj;
591 	struct socket *so;
592 #ifdef KERN_TLS
593 	struct ktls_session *tls;
594 #endif
595 	struct mbuf_ext_pgs *ext_pgs;
596 	struct mbuf *m, *mh, *mhtail;
597 	struct sf_buf *sf;
598 	struct shmfd *shmfd;
599 	struct sendfile_sync *sfs;
600 	struct vattr va;
601 	off_t off, sbytes, rem, obj_size;
602 	int bsize, error, ext_pgs_idx, hdrlen, max_pgs, softerr;
603 #ifdef KERN_TLS
604 	int tls_enq_cnt;
605 #endif
606 	bool use_ext_pgs;
607 
608 	obj = NULL;
609 	so = NULL;
610 	m = mh = NULL;
611 	sfs = NULL;
612 #ifdef KERN_TLS
613 	tls = NULL;
614 #endif
615 	hdrlen = sbytes = 0;
616 	softerr = 0;
617 	use_ext_pgs = false;
618 
619 	error = sendfile_getobj(td, fp, &obj, &vp, &shmfd, &obj_size, &bsize);
620 	if (error != 0)
621 		return (error);
622 
623 	error = sendfile_getsock(td, sockfd, &sock_fp, &so);
624 	if (error != 0)
625 		goto out;
626 
627 #ifdef MAC
628 	error = mac_socket_check_send(td->td_ucred, so);
629 	if (error != 0)
630 		goto out;
631 #endif
632 
633 	SFSTAT_INC(sf_syscalls);
634 	SFSTAT_ADD(sf_rhpages_requested, SF_READAHEAD(flags));
635 
636 	if (flags & SF_SYNC) {
637 		sfs = malloc(sizeof *sfs, M_TEMP, M_WAITOK | M_ZERO);
638 		mtx_init(&sfs->mtx, "sendfile", NULL, MTX_DEF);
639 		cv_init(&sfs->cv, "sendfile");
640 	}
641 
642 	rem = nbytes ? omin(nbytes, obj_size - offset) : obj_size - offset;
643 
644 	/*
645 	 * Protect against multiple writers to the socket.
646 	 *
647 	 * XXXRW: Historically this has assumed non-interruptibility, so now
648 	 * we implement that, but possibly shouldn't.
649 	 */
650 	(void)sblock(&so->so_snd, SBL_WAIT | SBL_NOINTR);
651 #ifdef KERN_TLS
652 	tls = ktls_hold(so->so_snd.sb_tls_info);
653 #endif
654 
655 	/*
656 	 * Loop through the pages of the file, starting with the requested
657 	 * offset. Get a file page (do I/O if necessary), map the file page
658 	 * into an sf_buf, attach an mbuf header to the sf_buf, and queue
659 	 * it on the socket.
660 	 * This is done in two loops.  The inner loop turns as many pages
661 	 * as it can, up to available socket buffer space, without blocking
662 	 * into mbufs to have it bulk delivered into the socket send buffer.
663 	 * The outer loop checks the state and available space of the socket
664 	 * and takes care of the overall progress.
665 	 */
666 	for (off = offset; rem > 0; ) {
667 		struct sf_io *sfio;
668 		vm_page_t *pa;
669 		struct mbuf *m0, *mtail;
670 		int nios, space, npages, rhpages;
671 
672 		mtail = NULL;
673 		/*
674 		 * Check the socket state for ongoing connection,
675 		 * no errors and space in socket buffer.
676 		 * If space is low allow for the remainder of the
677 		 * file to be processed if it fits the socket buffer.
678 		 * Otherwise block in waiting for sufficient space
679 		 * to proceed, or if the socket is nonblocking, return
680 		 * to userland with EAGAIN while reporting how far
681 		 * we've come.
682 		 * We wait until the socket buffer has significant free
683 		 * space to do bulk sends.  This makes good use of file
684 		 * system read ahead and allows packet segmentation
685 		 * offloading hardware to take over lots of work.  If
686 		 * we were not careful here we would send off only one
687 		 * sfbuf at a time.
688 		 */
689 		SOCKBUF_LOCK(&so->so_snd);
690 		if (so->so_snd.sb_lowat < so->so_snd.sb_hiwat / 2)
691 			so->so_snd.sb_lowat = so->so_snd.sb_hiwat / 2;
692 retry_space:
693 		if (so->so_snd.sb_state & SBS_CANTSENDMORE) {
694 			error = EPIPE;
695 			SOCKBUF_UNLOCK(&so->so_snd);
696 			goto done;
697 		} else if (so->so_error) {
698 			error = so->so_error;
699 			so->so_error = 0;
700 			SOCKBUF_UNLOCK(&so->so_snd);
701 			goto done;
702 		}
703 		if ((so->so_state & SS_ISCONNECTED) == 0) {
704 			SOCKBUF_UNLOCK(&so->so_snd);
705 			error = ENOTCONN;
706 			goto done;
707 		}
708 
709 		space = sbspace(&so->so_snd);
710 		if (space < rem &&
711 		    (space <= 0 ||
712 		     space < so->so_snd.sb_lowat)) {
713 			if (so->so_state & SS_NBIO) {
714 				SOCKBUF_UNLOCK(&so->so_snd);
715 				error = EAGAIN;
716 				goto done;
717 			}
718 			/*
719 			 * sbwait drops the lock while sleeping.
720 			 * When we loop back to retry_space the
721 			 * state may have changed and we retest
722 			 * for it.
723 			 */
724 			error = sbwait(&so->so_snd);
725 			/*
726 			 * An error from sbwait usually indicates that we've
727 			 * been interrupted by a signal. If we've sent anything
728 			 * then return bytes sent, otherwise return the error.
729 			 */
730 			if (error != 0) {
731 				SOCKBUF_UNLOCK(&so->so_snd);
732 				goto done;
733 			}
734 			goto retry_space;
735 		}
736 		SOCKBUF_UNLOCK(&so->so_snd);
737 
738 		/*
739 		 * At the beginning of the first loop check if any headers
740 		 * are specified and copy them into mbufs.  Reduce space in
741 		 * the socket buffer by the size of the header mbuf chain.
742 		 * Clear hdr_uio here and hdrlen at the end of the first loop.
743 		 */
744 		if (hdr_uio != NULL && hdr_uio->uio_resid > 0) {
745 			hdr_uio->uio_td = td;
746 			hdr_uio->uio_rw = UIO_WRITE;
747 #ifdef KERN_TLS
748 			if (tls != NULL)
749 				mh = m_uiotombuf(hdr_uio, M_WAITOK, space,
750 				    tls->params.max_frame_len, M_NOMAP);
751 			else
752 #endif
753 				mh = m_uiotombuf(hdr_uio, M_WAITOK,
754 				    space, 0, 0);
755 			hdrlen = m_length(mh, &mhtail);
756 			space -= hdrlen;
757 			/*
758 			 * If header consumed all the socket buffer space,
759 			 * don't waste CPU cycles and jump to the end.
760 			 */
761 			if (space == 0) {
762 				sfio = NULL;
763 				nios = 0;
764 				goto prepend_header;
765 			}
766 			hdr_uio = NULL;
767 		}
768 
769 		if (vp != NULL) {
770 			error = vn_lock(vp, LK_SHARED);
771 			if (error != 0)
772 				goto done;
773 			error = VOP_GETATTR(vp, &va, td->td_ucred);
774 			if (error != 0 || off >= va.va_size) {
775 				VOP_UNLOCK(vp);
776 				goto done;
777 			}
778 			if (va.va_size != obj_size) {
779 				obj_size = va.va_size;
780 				rem = nbytes ?
781 				    omin(nbytes + offset, obj_size) : obj_size;
782 				rem -= off;
783 			}
784 		}
785 
786 		if (space > rem)
787 			space = rem;
788 		else if (space > PAGE_SIZE) {
789 			/*
790 			 * Use page boundaries when possible for large
791 			 * requests.
792 			 */
793 			if (off & PAGE_MASK)
794 				space -= (PAGE_SIZE - (off & PAGE_MASK));
795 			space = trunc_page(space);
796 			if (off & PAGE_MASK)
797 				space += (PAGE_SIZE - (off & PAGE_MASK));
798 		}
799 
800 		npages = howmany(space + (off & PAGE_MASK), PAGE_SIZE);
801 
802 		/*
803 		 * Calculate maximum allowed number of pages for readahead
804 		 * at this iteration.  If SF_USER_READAHEAD was set, we don't
805 		 * do any heuristics and use exactly the value supplied by
806 		 * application.  Otherwise, we allow readahead up to "rem".
807 		 * If application wants more, let it be, but there is no
808 		 * reason to go above MAXPHYS.  Also check against "obj_size",
809 		 * since vm_pager_has_page() can hint beyond EOF.
810 		 */
811 		if (flags & SF_USER_READAHEAD) {
812 			rhpages = SF_READAHEAD(flags);
813 		} else {
814 			rhpages = howmany(rem + (off & PAGE_MASK), PAGE_SIZE) -
815 			    npages;
816 			rhpages += SF_READAHEAD(flags);
817 		}
818 		rhpages = min(howmany(MAXPHYS, PAGE_SIZE), rhpages);
819 		rhpages = min(howmany(obj_size - trunc_page(off), PAGE_SIZE) -
820 		    npages, rhpages);
821 
822 		sfio = malloc(sizeof(struct sf_io) +
823 		    npages * sizeof(vm_page_t), M_TEMP, M_WAITOK);
824 		refcount_init(&sfio->nios, 1);
825 		sfio->obj = obj;
826 		sfio->error = 0;
827 		sfio->m = NULL;
828 #ifdef KERN_TLS
829 		/*
830 		 * This doesn't use ktls_hold() because sfio->m will
831 		 * also have a reference on 'tls' that will be valid
832 		 * for all of sfio's lifetime.
833 		 */
834 		sfio->tls = tls;
835 #endif
836 		vm_object_pip_add(obj, 1);
837 		error = sendfile_swapin(obj, sfio, &nios, off, space, npages,
838 		    rhpages, flags);
839 		if (error != 0) {
840 			if (vp != NULL)
841 				VOP_UNLOCK(vp);
842 			sendfile_iodone(sfio, NULL, 0, error);
843 			goto done;
844 		}
845 
846 		/*
847 		 * Loop and construct maximum sized mbuf chain to be bulk
848 		 * dumped into socket buffer.
849 		 */
850 		pa = sfio->pa;
851 
852 		/*
853 		 * Use unmapped mbufs if enabled for TCP.  Unmapped
854 		 * bufs are restricted to TCP as that is what has been
855 		 * tested.  In particular, unmapped mbufs have not
856 		 * been tested with UNIX-domain sockets.
857 		 *
858 		 * TLS frames always require unmapped mbufs.
859 		 */
860 		if ((mb_use_ext_pgs &&
861 		    so->so_proto->pr_protocol == IPPROTO_TCP)
862 #ifdef KERN_TLS
863 		    || tls != NULL
864 #endif
865 		    ) {
866 			use_ext_pgs = true;
867 #ifdef KERN_TLS
868 			if (tls != NULL)
869 				max_pgs = num_pages(tls->params.max_frame_len);
870 			else
871 #endif
872 				max_pgs = MBUF_PEXT_MAX_PGS;
873 
874 			/* Start at last index, to wrap on first use. */
875 			ext_pgs_idx = max_pgs - 1;
876 		}
877 
878 		for (int i = 0; i < npages; i++) {
879 			/*
880 			 * If a page wasn't grabbed successfully, then
881 			 * trim the array. Can happen only with SF_NODISKIO.
882 			 */
883 			if (pa[i] == NULL) {
884 				SFSTAT_INC(sf_busy);
885 				fixspace(npages, i, off, &space);
886 				npages = i;
887 				softerr = EBUSY;
888 				break;
889 			}
890 
891 			if (use_ext_pgs) {
892 				off_t xfs;
893 
894 				ext_pgs_idx++;
895 				if (ext_pgs_idx == max_pgs) {
896 					m0 = mb_alloc_ext_pgs(M_WAITOK, false,
897 					    sendfile_free_mext_pg);
898 
899 					if (flags & SF_NOCACHE) {
900 						m0->m_ext.ext_flags |=
901 						    EXT_FLAG_NOCACHE;
902 
903 						/*
904 						 * See comment below regarding
905 						 * ignoring SF_NOCACHE for the
906 						 * last page.
907 						 */
908 						if ((npages - i <= max_pgs) &&
909 						    ((off + space) & PAGE_MASK) &&
910 						    (rem > space || rhpages > 0))
911 							m0->m_ext.ext_flags |=
912 							    EXT_FLAG_CACHE_LAST;
913 					}
914 					if (sfs != NULL) {
915 						m0->m_ext.ext_flags |=
916 						    EXT_FLAG_SYNC;
917 						m0->m_ext.ext_arg2 = sfs;
918 						mtx_lock(&sfs->mtx);
919 						sfs->count++;
920 						mtx_unlock(&sfs->mtx);
921 					}
922 					ext_pgs = m0->m_ext.ext_pgs;
923 					ext_pgs_idx = 0;
924 
925 					/* Append to mbuf chain. */
926 					if (mtail != NULL)
927 						mtail->m_next = m0;
928 					else
929 						m = m0;
930 					mtail = m0;
931 					ext_pgs->first_pg_off =
932 					    vmoff(i, off) & PAGE_MASK;
933 				}
934 				if (nios) {
935 					mtail->m_flags |= M_NOTREADY;
936 					ext_pgs->nrdy++;
937 				}
938 
939 				ext_pgs->pa[ext_pgs_idx] = VM_PAGE_TO_PHYS(pa[i]);
940 				ext_pgs->npgs++;
941 				xfs = xfsize(i, npages, off, space);
942 				ext_pgs->last_pg_len = xfs;
943 				MBUF_EXT_PGS_ASSERT_SANITY(ext_pgs);
944 				mtail->m_len += xfs;
945 				mtail->m_ext.ext_size += PAGE_SIZE;
946 				continue;
947 			}
948 
949 			/*
950 			 * Get a sendfile buf.  When allocating the
951 			 * first buffer for mbuf chain, we usually
952 			 * wait as long as necessary, but this wait
953 			 * can be interrupted.  For consequent
954 			 * buffers, do not sleep, since several
955 			 * threads might exhaust the buffers and then
956 			 * deadlock.
957 			 */
958 			sf = sf_buf_alloc(pa[i],
959 			    m != NULL ? SFB_NOWAIT : SFB_CATCH);
960 			if (sf == NULL) {
961 				SFSTAT_INC(sf_allocfail);
962 				for (int j = i; j < npages; j++)
963 					vm_page_unwire(pa[j], PQ_INACTIVE);
964 				if (m == NULL)
965 					softerr = ENOBUFS;
966 				fixspace(npages, i, off, &space);
967 				npages = i;
968 				break;
969 			}
970 
971 			m0 = m_get(M_WAITOK, MT_DATA);
972 			m0->m_ext.ext_buf = (char *)sf_buf_kva(sf);
973 			m0->m_ext.ext_size = PAGE_SIZE;
974 			m0->m_ext.ext_arg1 = sf;
975 			m0->m_ext.ext_type = EXT_SFBUF;
976 			m0->m_ext.ext_flags = EXT_FLAG_EMBREF;
977 			m0->m_ext.ext_free = sendfile_free_mext;
978 			/*
979 			 * SF_NOCACHE sets the page as being freed upon send.
980 			 * However, we ignore it for the last page in 'space',
981 			 * if the page is truncated, and we got more data to
982 			 * send (rem > space), or if we have readahead
983 			 * configured (rhpages > 0).
984 			 */
985 			if ((flags & SF_NOCACHE) &&
986 			    (i != npages - 1 ||
987 			    !((off + space) & PAGE_MASK) ||
988 			    !(rem > space || rhpages > 0)))
989 				m0->m_ext.ext_flags |= EXT_FLAG_NOCACHE;
990 			if (sfs != NULL) {
991 				m0->m_ext.ext_flags |= EXT_FLAG_SYNC;
992 				m0->m_ext.ext_arg2 = sfs;
993 				mtx_lock(&sfs->mtx);
994 				sfs->count++;
995 				mtx_unlock(&sfs->mtx);
996 			}
997 			m0->m_ext.ext_count = 1;
998 			m0->m_flags |= (M_EXT | M_RDONLY);
999 			if (nios)
1000 				m0->m_flags |= M_NOTREADY;
1001 			m0->m_data = (char *)sf_buf_kva(sf) +
1002 			    (vmoff(i, off) & PAGE_MASK);
1003 			m0->m_len = xfsize(i, npages, off, space);
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);
1015 
1016 		/* Keep track of bytes processed. */
1017 		off += space;
1018 		rem -= space;
1019 
1020 		/*
1021 		 * Prepend header, if any.  Save pointer to first mbuf
1022 		 * with a page.
1023 		 */
1024 		if (hdrlen) {
1025 prepend_header:
1026 			m0 = mhtail->m_next = m;
1027 			m = mh;
1028 			mh = NULL;
1029 		} else
1030 			m0 = m;
1031 
1032 		if (m == NULL) {
1033 			KASSERT(softerr, ("%s: m NULL, no error", __func__));
1034 			error = softerr;
1035 			sendfile_iodone(sfio, NULL, 0, 0);
1036 			goto done;
1037 		}
1038 
1039 		/* Add the buffer chain to the socket buffer. */
1040 		KASSERT(m_length(m, NULL) == space + hdrlen,
1041 		    ("%s: mlen %u space %d hdrlen %d",
1042 		    __func__, m_length(m, NULL), space, hdrlen));
1043 
1044 		CURVNET_SET(so->so_vnet);
1045 #ifdef KERN_TLS
1046 		if (tls != NULL)
1047 			ktls_frame(m, tls, &tls_enq_cnt, TLS_RLTYPE_APP);
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, or if
1053 			 * the header consumed all socket buffer space and
1054 			 * sfio is NULL, then we can send data right now
1055 			 * without the PRUS_NOTREADY flag.
1056 			 */
1057 			if (sfio != NULL)
1058 				sendfile_iodone(sfio, NULL, 0, 0);
1059 #ifdef KERN_TLS
1060 			if (tls != NULL && tls->mode == TCP_TLS_MODE_SW) {
1061 				error = (*so->so_proto->pr_usrreqs->pru_send)
1062 				    (so, PRUS_NOTREADY, m, NULL, NULL, td);
1063 				soref(so);
1064 				ktls_enqueue(m, so, tls_enq_cnt);
1065 			} else
1066 #endif
1067 				error = (*so->so_proto->pr_usrreqs->pru_send)
1068 				    (so, 0, m, NULL, NULL, td);
1069 		} else {
1070 			sfio->so = so;
1071 			sfio->m = m0;
1072 			sfio->npages = npages;
1073 			soref(so);
1074 			error = (*so->so_proto->pr_usrreqs->pru_send)
1075 			    (so, PRUS_NOTREADY, m, NULL, NULL, td);
1076 			sendfile_iodone(sfio, NULL, 0, 0);
1077 		}
1078 		CURVNET_RESTORE();
1079 
1080 		m = NULL;	/* pru_send always consumes */
1081 		if (error)
1082 			goto done;
1083 		sbytes += space + hdrlen;
1084 		if (hdrlen)
1085 			hdrlen = 0;
1086 		if (softerr) {
1087 			error = softerr;
1088 			goto done;
1089 		}
1090 	}
1091 
1092 	/*
1093 	 * Send trailers. Wimp out and use writev(2).
1094 	 */
1095 	if (trl_uio != NULL) {
1096 		sbunlock(&so->so_snd);
1097 		error = kern_writev(td, sockfd, trl_uio);
1098 		if (error == 0)
1099 			sbytes += td->td_retval[0];
1100 		goto out;
1101 	}
1102 
1103 done:
1104 	sbunlock(&so->so_snd);
1105 out:
1106 	/*
1107 	 * If there was no error we have to clear td->td_retval[0]
1108 	 * because it may have been set by writev.
1109 	 */
1110 	if (error == 0) {
1111 		td->td_retval[0] = 0;
1112 	}
1113 	if (sent != NULL) {
1114 		(*sent) = sbytes;
1115 	}
1116 	if (obj != NULL)
1117 		vm_object_deallocate(obj);
1118 	if (so)
1119 		fdrop(sock_fp, td);
1120 	if (m)
1121 		m_freem(m);
1122 	if (mh)
1123 		m_freem(mh);
1124 
1125 	if (sfs != NULL) {
1126 		mtx_lock(&sfs->mtx);
1127 		if (sfs->count != 0)
1128 			cv_wait(&sfs->cv, &sfs->mtx);
1129 		KASSERT(sfs->count == 0, ("sendfile sync still busy"));
1130 		cv_destroy(&sfs->cv);
1131 		mtx_destroy(&sfs->mtx);
1132 		free(sfs, M_TEMP);
1133 	}
1134 #ifdef KERN_TLS
1135 	if (tls != NULL)
1136 		ktls_free(tls);
1137 #endif
1138 
1139 	if (error == ERESTART)
1140 		error = EINTR;
1141 
1142 	return (error);
1143 }
1144 
1145 static int
1146 sendfile(struct thread *td, struct sendfile_args *uap, int compat)
1147 {
1148 	struct sf_hdtr hdtr;
1149 	struct uio *hdr_uio, *trl_uio;
1150 	struct file *fp;
1151 	off_t sbytes;
1152 	int error;
1153 
1154 	/*
1155 	 * File offset must be positive.  If it goes beyond EOF
1156 	 * we send only the header/trailer and no payload data.
1157 	 */
1158 	if (uap->offset < 0)
1159 		return (EINVAL);
1160 
1161 	sbytes = 0;
1162 	hdr_uio = trl_uio = NULL;
1163 
1164 	if (uap->hdtr != NULL) {
1165 		error = copyin(uap->hdtr, &hdtr, sizeof(hdtr));
1166 		if (error != 0)
1167 			goto out;
1168 		if (hdtr.headers != NULL) {
1169 			error = copyinuio(hdtr.headers, hdtr.hdr_cnt,
1170 			    &hdr_uio);
1171 			if (error != 0)
1172 				goto out;
1173 #ifdef COMPAT_FREEBSD4
1174 			/*
1175 			 * In FreeBSD < 5.0 the nbytes to send also included
1176 			 * the header.  If compat is specified subtract the
1177 			 * header size from nbytes.
1178 			 */
1179 			if (compat) {
1180 				if (uap->nbytes > hdr_uio->uio_resid)
1181 					uap->nbytes -= hdr_uio->uio_resid;
1182 				else
1183 					uap->nbytes = 0;
1184 			}
1185 #endif
1186 		}
1187 		if (hdtr.trailers != NULL) {
1188 			error = copyinuio(hdtr.trailers, hdtr.trl_cnt,
1189 			    &trl_uio);
1190 			if (error != 0)
1191 				goto out;
1192 		}
1193 	}
1194 
1195 	AUDIT_ARG_FD(uap->fd);
1196 
1197 	/*
1198 	 * sendfile(2) can start at any offset within a file so we require
1199 	 * CAP_READ+CAP_SEEK = CAP_PREAD.
1200 	 */
1201 	if ((error = fget_read(td, uap->fd, &cap_pread_rights, &fp)) != 0)
1202 		goto out;
1203 
1204 	error = fo_sendfile(fp, uap->s, hdr_uio, trl_uio, uap->offset,
1205 	    uap->nbytes, &sbytes, uap->flags, td);
1206 	fdrop(fp, td);
1207 
1208 	if (uap->sbytes != NULL)
1209 		copyout(&sbytes, uap->sbytes, sizeof(off_t));
1210 
1211 out:
1212 	free(hdr_uio, M_IOV);
1213 	free(trl_uio, M_IOV);
1214 	return (error);
1215 }
1216 
1217 /*
1218  * sendfile(2)
1219  *
1220  * int sendfile(int fd, int s, off_t offset, size_t nbytes,
1221  *       struct sf_hdtr *hdtr, off_t *sbytes, int flags)
1222  *
1223  * Send a file specified by 'fd' and starting at 'offset' to a socket
1224  * specified by 's'. Send only 'nbytes' of the file or until EOF if nbytes ==
1225  * 0.  Optionally add a header and/or trailer to the socket output.  If
1226  * specified, write the total number of bytes sent into *sbytes.
1227  */
1228 int
1229 sys_sendfile(struct thread *td, struct sendfile_args *uap)
1230 {
1231 
1232 	return (sendfile(td, uap, 0));
1233 }
1234 
1235 #ifdef COMPAT_FREEBSD4
1236 int
1237 freebsd4_sendfile(struct thread *td, struct freebsd4_sendfile_args *uap)
1238 {
1239 	struct sendfile_args args;
1240 
1241 	args.fd = uap->fd;
1242 	args.s = uap->s;
1243 	args.offset = uap->offset;
1244 	args.nbytes = uap->nbytes;
1245 	args.hdtr = uap->hdtr;
1246 	args.sbytes = uap->sbytes;
1247 	args.flags = uap->flags;
1248 
1249 	return (sendfile(td, &args, 1));
1250 }
1251 #endif /* COMPAT_FREEBSD4 */
1252