xref: /freebsd/sys/kern/sys_pipe.c (revision 36138969847b231cd98f48272e2bdf88a8dc08dd)
1 /*-
2  * Copyright (c) 1996 John S. Dyson
3  * Copyright (c) 2012 Giovanni Trematerra
4  * All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice immediately at the beginning of the file, without modification,
11  *    this list of conditions, and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  * 3. Absolutely no warranty of function or purpose is made by the author
16  *    John S. Dyson.
17  * 4. Modifications may be freely made to this file if the above conditions
18  *    are met.
19  */
20 
21 /*
22  * This file contains a high-performance replacement for the socket-based
23  * pipes scheme originally used in FreeBSD/4.4Lite.  It does not support
24  * all features of sockets, but does do everything that pipes normally
25  * do.
26  */
27 
28 /*
29  * This code has two modes of operation, a small write mode and a large
30  * write mode.  The small write mode acts like conventional pipes with
31  * a kernel buffer.  If the buffer is less than PIPE_MINDIRECT, then the
32  * "normal" pipe buffering is done.  If the buffer is between PIPE_MINDIRECT
33  * and PIPE_SIZE in size, the sending process pins the underlying pages in
34  * memory, and the receiving process copies directly from these pinned pages
35  * in the sending process.
36  *
37  * If the sending process receives a signal, it is possible that it will
38  * go away, and certainly its address space can change, because control
39  * is returned back to the user-mode side.  In that case, the pipe code
40  * arranges to copy the buffer supplied by the user process, to a pageable
41  * kernel buffer, and the receiving process will grab the data from the
42  * pageable kernel buffer.  Since signals don't happen all that often,
43  * the copy operation is normally eliminated.
44  *
45  * The constant PIPE_MINDIRECT is chosen to make sure that buffering will
46  * happen for small transfers so that the system will not spend all of
47  * its time context switching.
48  *
49  * In order to limit the resource use of pipes, two sysctls exist:
50  *
51  * kern.ipc.maxpipekva - This is a hard limit on the amount of pageable
52  * address space available to us in pipe_map. This value is normally
53  * autotuned, but may also be loader tuned.
54  *
55  * kern.ipc.pipekva - This read-only sysctl tracks the current amount of
56  * memory in use by pipes.
57  *
58  * Based on how large pipekva is relative to maxpipekva, the following
59  * will happen:
60  *
61  * 0% - 50%:
62  *     New pipes are given 16K of memory backing, pipes may dynamically
63  *     grow to as large as 64K where needed.
64  * 50% - 75%:
65  *     New pipes are given 4K (or PAGE_SIZE) of memory backing,
66  *     existing pipes may NOT grow.
67  * 75% - 100%:
68  *     New pipes are given 4K (or PAGE_SIZE) of memory backing,
69  *     existing pipes will be shrunk down to 4K whenever possible.
70  *
71  * Resizing may be disabled by setting kern.ipc.piperesizeallowed=0.  If
72  * that is set,  the only resize that will occur is the 0 -> SMALL_PIPE_SIZE
73  * resize which MUST occur for reverse-direction pipes when they are
74  * first used.
75  *
76  * Additional information about the current state of pipes may be obtained
77  * from kern.ipc.pipes, kern.ipc.pipefragretry, kern.ipc.pipeallocfail,
78  * and kern.ipc.piperesizefail.
79  *
80  * Locking rules:  There are two locks present here:  A mutex, used via
81  * PIPE_LOCK, and a flag, used via pipelock().  All locking is done via
82  * the flag, as mutexes can not persist over uiomove.  The mutex
83  * exists only to guard access to the flag, and is not in itself a
84  * locking mechanism.  Also note that there is only a single mutex for
85  * both directions of a pipe.
86  *
87  * As pipelock() may have to sleep before it can acquire the flag, it
88  * is important to reread all data after a call to pipelock(); everything
89  * in the structure may have changed.
90  */
91 
92 #include <sys/param.h>
93 #include <sys/systm.h>
94 #include <sys/conf.h>
95 #include <sys/fcntl.h>
96 #include <sys/file.h>
97 #include <sys/filedesc.h>
98 #include <sys/filio.h>
99 #include <sys/kernel.h>
100 #include <sys/lock.h>
101 #include <sys/mutex.h>
102 #include <sys/ttycom.h>
103 #include <sys/stat.h>
104 #include <sys/malloc.h>
105 #include <sys/poll.h>
106 #include <sys/priv.h>
107 #include <sys/selinfo.h>
108 #include <sys/signalvar.h>
109 #include <sys/syscallsubr.h>
110 #include <sys/sysctl.h>
111 #include <sys/sysproto.h>
112 #include <sys/pipe.h>
113 #include <sys/proc.h>
114 #include <sys/vnode.h>
115 #include <sys/uio.h>
116 #include <sys/user.h>
117 #include <sys/event.h>
118 
119 #include <security/mac/mac_framework.h>
120 
121 #include <vm/vm.h>
122 #include <vm/vm_param.h>
123 #include <vm/vm_object.h>
124 #include <vm/vm_kern.h>
125 #include <vm/vm_extern.h>
126 #include <vm/pmap.h>
127 #include <vm/vm_map.h>
128 #include <vm/vm_page.h>
129 #include <vm/uma.h>
130 
131 /*
132  * Use this define if you want to disable *fancy* VM things.  Expect an
133  * approx 30% decrease in transfer rate.  This could be useful for
134  * NetBSD or OpenBSD.
135  */
136 /* #define PIPE_NODIRECT */
137 
138 #define PIPE_PEER(pipe)	\
139 	(((pipe)->pipe_type & PIPE_TYPE_NAMED) ? (pipe) : ((pipe)->pipe_peer))
140 
141 /*
142  * interfaces to the outside world
143  */
144 static fo_rdwr_t	pipe_read;
145 static fo_rdwr_t	pipe_write;
146 static fo_truncate_t	pipe_truncate;
147 static fo_ioctl_t	pipe_ioctl;
148 static fo_poll_t	pipe_poll;
149 static fo_kqfilter_t	pipe_kqfilter;
150 static fo_stat_t	pipe_stat;
151 static fo_close_t	pipe_close;
152 static fo_chmod_t	pipe_chmod;
153 static fo_chown_t	pipe_chown;
154 static fo_fill_kinfo_t	pipe_fill_kinfo;
155 
156 const struct fileops pipeops = {
157 	.fo_read = pipe_read,
158 	.fo_write = pipe_write,
159 	.fo_truncate = pipe_truncate,
160 	.fo_ioctl = pipe_ioctl,
161 	.fo_poll = pipe_poll,
162 	.fo_kqfilter = pipe_kqfilter,
163 	.fo_stat = pipe_stat,
164 	.fo_close = pipe_close,
165 	.fo_chmod = pipe_chmod,
166 	.fo_chown = pipe_chown,
167 	.fo_sendfile = invfo_sendfile,
168 	.fo_fill_kinfo = pipe_fill_kinfo,
169 	.fo_cmp = file_kcmp_generic,
170 	.fo_flags = DFLAG_PASSABLE
171 };
172 
173 static void	filt_pipedetach(struct knote *kn);
174 static void	filt_pipedetach_notsup(struct knote *kn);
175 static int	filt_pipenotsup(struct knote *kn, long hint);
176 static int	filt_piperead(struct knote *kn, long hint);
177 static int	filt_pipewrite(struct knote *kn, long hint);
178 static int	filt_pipedump(struct proc *p, struct knote *kn,
179     struct kinfo_knote *kin);
180 
181 static const struct filterops pipe_nfiltops = {
182 	.f_isfd = 1,
183 	.f_detach = filt_pipedetach_notsup,
184 	.f_event = filt_pipenotsup,
185 	/* no userdump */
186 	.f_copy = knote_triv_copy,
187 };
188 static const struct filterops pipe_rfiltops = {
189 	.f_isfd = 1,
190 	.f_detach = filt_pipedetach,
191 	.f_event = filt_piperead,
192 	.f_userdump = filt_pipedump,
193 	.f_copy = knote_triv_copy,
194 };
195 static const struct filterops pipe_wfiltops = {
196 	.f_isfd = 1,
197 	.f_detach = filt_pipedetach,
198 	.f_event = filt_pipewrite,
199 	.f_userdump = filt_pipedump,
200 	.f_copy = knote_triv_copy,
201 };
202 
203 /*
204  * Default pipe buffer size(s), this can be kind-of large now because pipe
205  * space is pageable.  The pipe code will try to maintain locality of
206  * reference for performance reasons, so small amounts of outstanding I/O
207  * will not wipe the cache.
208  */
209 #define MINPIPESIZE (PIPE_SIZE/3)
210 #define MAXPIPESIZE (2*PIPE_SIZE/3)
211 
212 static long amountpipekva;
213 static int pipefragretry;
214 static int pipeallocfail;
215 static int piperesizefail;
216 static int piperesizeallowed = 1;
217 static long pipe_mindirect = PIPE_MINDIRECT;
218 static int pipebuf_reserv = 2;
219 
220 SYSCTL_LONG(_kern_ipc, OID_AUTO, maxpipekva, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
221 	   &maxpipekva, 0, "Pipe KVA limit");
222 SYSCTL_LONG(_kern_ipc, OID_AUTO, pipekva, CTLFLAG_RD,
223 	   &amountpipekva, 0, "Pipe KVA usage");
224 SYSCTL_INT(_kern_ipc, OID_AUTO, pipefragretry, CTLFLAG_RD,
225 	  &pipefragretry, 0, "Pipe allocation retries due to fragmentation");
226 SYSCTL_INT(_kern_ipc, OID_AUTO, pipeallocfail, CTLFLAG_RD,
227 	  &pipeallocfail, 0, "Pipe allocation failures");
228 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizefail, CTLFLAG_RD,
229 	  &piperesizefail, 0, "Pipe resize failures");
230 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizeallowed, CTLFLAG_RW,
231 	  &piperesizeallowed, 0, "Pipe resizing allowed");
232 SYSCTL_INT(_kern_ipc, OID_AUTO, pipebuf_reserv, CTLFLAG_RW,
233     &pipebuf_reserv, 0,
234     "Superuser-reserved percentage of the pipe buffers space");
235 
236 static void pipeinit(void *dummy __unused);
237 static void pipeclose(struct pipe *cpipe);
238 static void pipe_free_kmem(struct pipe *cpipe);
239 static int pipe_create(struct pipe *pipe, bool backing);
240 static void pipe_destroy(struct pipe *pipe);
241 static int pipe_paircreate(struct thread *td, struct pipepair **p_pp);
242 static __inline int pipelock(struct pipe *cpipe, bool catch);
243 static __inline void pipeunlock(struct pipe *cpipe);
244 static void pipe_timestamp(struct timespec *tsp);
245 #ifndef PIPE_NODIRECT
246 static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio);
247 static void pipe_destroy_write_buffer(struct pipe *wpipe);
248 static int pipe_direct_write(struct pipe *wpipe, struct uio *uio);
249 static void pipe_clone_write_buffer(struct pipe *wpipe);
250 #endif
251 static int pipespace(struct pipe *cpipe, int size);
252 static int pipespace_new(struct pipe *cpipe, int size);
253 
254 static int	pipe_zone_ctor(void *mem, int size, void *arg, int flags);
255 static int	pipe_zone_init(void *mem, int size, int flags);
256 static void	pipe_zone_fini(void *mem, int size);
257 
258 static uma_zone_t pipe_zone;
259 static struct unrhdr64 pipeino_unr;
260 static dev_t pipedev_ino;
261 
262 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL);
263 
264 static void
pipeinit(void * dummy __unused)265 pipeinit(void *dummy __unused)
266 {
267 
268 	pipe_zone = uma_zcreate("pipe", sizeof(struct pipepair),
269 	    pipe_zone_ctor, NULL, pipe_zone_init, pipe_zone_fini,
270 	    UMA_ALIGN_PTR, 0);
271 	KASSERT(pipe_zone != NULL, ("pipe_zone not initialized"));
272 	new_unrhdr64(&pipeino_unr, 1);
273 	pipedev_ino = devfs_alloc_cdp_inode();
274 	KASSERT(pipedev_ino > 0, ("pipe dev inode not initialized"));
275 }
276 
277 static int
sysctl_handle_pipe_mindirect(SYSCTL_HANDLER_ARGS)278 sysctl_handle_pipe_mindirect(SYSCTL_HANDLER_ARGS)
279 {
280 	int error = 0;
281 	long tmp_pipe_mindirect = pipe_mindirect;
282 
283 	error = sysctl_handle_long(oidp, &tmp_pipe_mindirect, arg2, req);
284 	if (error != 0 || req->newptr == NULL)
285 		return (error);
286 
287 	/*
288 	 * Don't allow pipe_mindirect to be set so low that we violate
289 	 * atomicity requirements.
290 	 */
291 	if (tmp_pipe_mindirect <= PIPE_BUF)
292 		return (EINVAL);
293 	pipe_mindirect = tmp_pipe_mindirect;
294 	return (0);
295 }
296 SYSCTL_OID(_kern_ipc, OID_AUTO, pipe_mindirect, CTLTYPE_LONG | CTLFLAG_RW,
297     &pipe_mindirect, 0, sysctl_handle_pipe_mindirect, "L",
298     "Minimum write size triggering VM optimization");
299 
300 static int
pipe_zone_ctor(void * mem,int size,void * arg,int flags)301 pipe_zone_ctor(void *mem, int size, void *arg, int flags)
302 {
303 	struct pipepair *pp;
304 	struct pipe *rpipe, *wpipe;
305 
306 	KASSERT(size == sizeof(*pp), ("pipe_zone_ctor: wrong size"));
307 
308 	pp = (struct pipepair *)mem;
309 
310 	/*
311 	 * We zero both pipe endpoints to make sure all the kmem pointers
312 	 * are NULL, flag fields are zero'd, etc.  We timestamp both
313 	 * endpoints with the same time.
314 	 */
315 	rpipe = &pp->pp_rpipe;
316 	bzero(rpipe, sizeof(*rpipe));
317 	pipe_timestamp(&rpipe->pipe_ctime);
318 	rpipe->pipe_atime = rpipe->pipe_mtime = rpipe->pipe_ctime;
319 
320 	wpipe = &pp->pp_wpipe;
321 	bzero(wpipe, sizeof(*wpipe));
322 	wpipe->pipe_ctime = rpipe->pipe_ctime;
323 	wpipe->pipe_atime = wpipe->pipe_mtime = rpipe->pipe_ctime;
324 
325 	rpipe->pipe_peer = wpipe;
326 	rpipe->pipe_pair = pp;
327 	wpipe->pipe_peer = rpipe;
328 	wpipe->pipe_pair = pp;
329 
330 	/*
331 	 * Mark both endpoints as present; they will later get free'd
332 	 * one at a time.  When both are free'd, then the whole pair
333 	 * is released.
334 	 */
335 	rpipe->pipe_present = PIPE_ACTIVE;
336 	wpipe->pipe_present = PIPE_ACTIVE;
337 
338 	/*
339 	 * Eventually, the MAC Framework may initialize the label
340 	 * in ctor or init, but for now we do it elswhere to avoid
341 	 * blocking in ctor or init.
342 	 */
343 	pp->pp_label = NULL;
344 
345 	return (0);
346 }
347 
348 static int
pipe_zone_init(void * mem,int size,int flags)349 pipe_zone_init(void *mem, int size, int flags)
350 {
351 	struct pipepair *pp;
352 
353 	KASSERT(size == sizeof(*pp), ("pipe_zone_init: wrong size"));
354 
355 	pp = (struct pipepair *)mem;
356 
357 	mtx_init(&pp->pp_mtx, "pipe mutex", NULL, MTX_DEF | MTX_NEW);
358 	return (0);
359 }
360 
361 static void
pipe_zone_fini(void * mem,int size)362 pipe_zone_fini(void *mem, int size)
363 {
364 	struct pipepair *pp;
365 
366 	KASSERT(size == sizeof(*pp), ("pipe_zone_fini: wrong size"));
367 
368 	pp = (struct pipepair *)mem;
369 
370 	mtx_destroy(&pp->pp_mtx);
371 }
372 
373 static int
pipe_paircreate(struct thread * td,struct pipepair ** p_pp)374 pipe_paircreate(struct thread *td, struct pipepair **p_pp)
375 {
376 	struct pipepair *pp;
377 	struct pipe *rpipe, *wpipe;
378 	int error;
379 
380 	*p_pp = pp = uma_zalloc(pipe_zone, M_WAITOK);
381 #ifdef MAC
382 	/*
383 	 * The MAC label is shared between the connected endpoints.  As a
384 	 * result mac_pipe_init() and mac_pipe_create() are called once
385 	 * for the pair, and not on the endpoints.
386 	 */
387 	mac_pipe_init(pp);
388 	mac_pipe_create(td->td_ucred, pp);
389 #endif
390 	rpipe = &pp->pp_rpipe;
391 	wpipe = &pp->pp_wpipe;
392 	pp->pp_owner = crhold(td->td_ucred);
393 
394 	knlist_init_mtx(&rpipe->pipe_sel.si_note, PIPE_MTX(rpipe));
395 	knlist_init_mtx(&wpipe->pipe_sel.si_note, PIPE_MTX(wpipe));
396 
397 	/*
398 	 * Only the forward direction pipe is backed by big buffer by
399 	 * default.
400 	 */
401 	error = pipe_create(rpipe, true);
402 	if (error != 0)
403 		goto fail;
404 	error = pipe_create(wpipe, false);
405 	if (error != 0) {
406 		pipe_destroy(rpipe);
407 		goto fail;
408 	}
409 
410 	rpipe->pipe_state |= PIPE_DIRECTOK;
411 	wpipe->pipe_state |= PIPE_DIRECTOK;
412 	return (0);
413 
414 fail:
415 	knlist_destroy(&rpipe->pipe_sel.si_note);
416 	knlist_destroy(&wpipe->pipe_sel.si_note);
417 	crfree(pp->pp_owner);
418 #ifdef MAC
419 	mac_pipe_destroy(pp);
420 #endif
421 	uma_zfree(pipe_zone, pp);
422 	return (error);
423 }
424 
425 int
pipe_named_ctor(struct pipe ** ppipe,struct thread * td)426 pipe_named_ctor(struct pipe **ppipe, struct thread *td)
427 {
428 	struct pipepair *pp;
429 	int error;
430 
431 	error = pipe_paircreate(td, &pp);
432 	if (error != 0)
433 		return (error);
434 	pp->pp_rpipe.pipe_type |= PIPE_TYPE_NAMED;
435 	*ppipe = &pp->pp_rpipe;
436 	return (0);
437 }
438 
439 void
pipe_dtor(struct pipe * dpipe)440 pipe_dtor(struct pipe *dpipe)
441 {
442 	struct pipe *peer;
443 
444 	peer = (dpipe->pipe_type & PIPE_TYPE_NAMED) != 0 ? dpipe->pipe_peer : NULL;
445 	funsetown(&dpipe->pipe_sigio);
446 	pipeclose(dpipe);
447 	if (peer != NULL) {
448 		funsetown(&peer->pipe_sigio);
449 		pipeclose(peer);
450 	}
451 }
452 
453 /*
454  * Get a timestamp.
455  *
456  * This used to be vfs_timestamp but the higher precision is unnecessary and
457  * can very negatively affect performance in virtualized environments (e.g., on
458  * vms running on amd64 when using the rdtscp instruction).
459  */
460 static void
pipe_timestamp(struct timespec * tsp)461 pipe_timestamp(struct timespec *tsp)
462 {
463 
464 	getnanotime(tsp);
465 }
466 
467 /*
468  * The pipe system call for the DTYPE_PIPE type of pipes.  If we fail, let
469  * the zone pick up the pieces via pipeclose().
470  */
471 int
kern_pipe(struct thread * td,int fildes[2],int flags,struct filecaps * fcaps1,struct filecaps * fcaps2)472 kern_pipe(struct thread *td, int fildes[2], int flags, struct filecaps *fcaps1,
473     struct filecaps *fcaps2)
474 {
475 	struct file *rf, *wf;
476 	struct pipe *rpipe, *wpipe;
477 	struct pipepair *pp;
478 	int fd, fflags, error;
479 
480 	error = pipe_paircreate(td, &pp);
481 	if (error != 0)
482 		return (error);
483 	rpipe = &pp->pp_rpipe;
484 	wpipe = &pp->pp_wpipe;
485 	error = falloc_caps(td, &rf, &fd, flags, fcaps1);
486 	if (error) {
487 		pipeclose(rpipe);
488 		pipeclose(wpipe);
489 		return (error);
490 	}
491 	/* An extra reference on `rf' has been held for us by falloc_caps(). */
492 	fildes[0] = fd;
493 
494 	fflags = FREAD | FWRITE;
495 	if ((flags & O_NONBLOCK) != 0)
496 		fflags |= FNONBLOCK;
497 
498 	/*
499 	 * Warning: once we've gotten past allocation of the fd for the
500 	 * read-side, we can only drop the read side via fdrop() in order
501 	 * to avoid races against processes which manage to dup() the read
502 	 * side while we are blocked trying to allocate the write side.
503 	 */
504 	finit(rf, fflags, DTYPE_PIPE, rpipe, &pipeops);
505 	error = falloc_caps(td, &wf, &fd, flags, fcaps2);
506 	if (error) {
507 		fdclose(td, rf, fildes[0]);
508 		fdrop(rf, td);
509 		/* rpipe has been closed by fdrop(). */
510 		pipeclose(wpipe);
511 		return (error);
512 	}
513 	/* An extra reference on `wf' has been held for us by falloc_caps(). */
514 	finit(wf, fflags, DTYPE_PIPE, wpipe, &pipeops);
515 	fdrop(wf, td);
516 	fildes[1] = fd;
517 	fdrop(rf, td);
518 
519 	return (0);
520 }
521 
522 #ifdef COMPAT_FREEBSD10
523 /* ARGSUSED */
524 int
freebsd10_pipe(struct thread * td,struct freebsd10_pipe_args * uap __unused)525 freebsd10_pipe(struct thread *td, struct freebsd10_pipe_args *uap __unused)
526 {
527 	int error;
528 	int fildes[2];
529 
530 	error = kern_pipe(td, fildes, 0, NULL, NULL);
531 	if (error)
532 		return (error);
533 
534 	td->td_retval[0] = fildes[0];
535 	td->td_retval[1] = fildes[1];
536 
537 	return (0);
538 }
539 #endif
540 
541 int
sys_pipe2(struct thread * td,struct pipe2_args * uap)542 sys_pipe2(struct thread *td, struct pipe2_args *uap)
543 {
544 	int error, fildes[2];
545 
546 	if ((uap->flags & ~(O_CLOEXEC | O_CLOFORK | O_NONBLOCK)) != 0)
547 		return (EINVAL);
548 	error = kern_pipe(td, fildes, uap->flags, NULL, NULL);
549 	if (error)
550 		return (error);
551 	error = copyout(fildes, uap->fildes, 2 * sizeof(int));
552 	if (error) {
553 		(void)kern_close(td, fildes[0]);
554 		(void)kern_close(td, fildes[1]);
555 	}
556 	return (error);
557 }
558 
559 /*
560  * Allocate kva for pipe circular buffer, the space is pageable
561  * This routine will 'realloc' the size of a pipe safely, if it fails
562  * it will retain the old buffer.
563  * If it fails it will return ENOMEM.
564  */
565 static int
pipespace_new(struct pipe * cpipe,int size)566 pipespace_new(struct pipe *cpipe, int size)
567 {
568 	caddr_t buffer;
569 	int error, cnt, firstseg;
570 	static int curfail = 0;
571 	static struct timeval lastfail;
572 
573 	PIPE_LOCK_ASSERT(cpipe, MA_NOTOWNED);
574 	KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW),
575 		("pipespace: resize of direct writes not allowed"));
576 retry:
577 	cnt = cpipe->pipe_buffer.cnt;
578 	if (cnt > size)
579 		size = cnt;
580 
581 	size = round_page(size);
582 	buffer = (caddr_t) vm_map_min(pipe_map);
583 
584 	if (!chgpipecnt(cpipe->pipe_pair->pp_owner->cr_ruidinfo,
585 	    size, lim_cur(curthread, RLIMIT_PIPEBUF))) {
586 		if (cpipe->pipe_buffer.buffer == NULL &&
587 		    size > SMALL_PIPE_SIZE) {
588 			size = SMALL_PIPE_SIZE;
589 			goto retry;
590 		}
591 		return (ENOMEM);
592 	}
593 
594 	vm_map_lock(pipe_map);
595 	if (priv_check(curthread, PRIV_PIPEBUF) != 0 && maxpipekva / 100 *
596 	    (100 - pipebuf_reserv) < amountpipekva + size) {
597 		vm_map_unlock(pipe_map);
598 		chgpipecnt(cpipe->pipe_pair->pp_owner->cr_ruidinfo, -size, 0);
599 		if (cpipe->pipe_buffer.buffer == NULL &&
600 		    size > SMALL_PIPE_SIZE) {
601 			size = SMALL_PIPE_SIZE;
602 			pipefragretry++;
603 			goto retry;
604 		}
605 		return (ENOMEM);
606 	}
607 	error = vm_map_find_locked(pipe_map, NULL, 0, (vm_offset_t *)&buffer,
608 	    size, 0, VMFS_ANY_SPACE, VM_PROT_RW, VM_PROT_RW, 0);
609 	vm_map_unlock(pipe_map);
610 	if (error != KERN_SUCCESS) {
611 		chgpipecnt(cpipe->pipe_pair->pp_owner->cr_ruidinfo, -size, 0);
612 		if (cpipe->pipe_buffer.buffer == NULL &&
613 		    size > SMALL_PIPE_SIZE) {
614 			size = SMALL_PIPE_SIZE;
615 			pipefragretry++;
616 			goto retry;
617 		}
618 		if (cpipe->pipe_buffer.buffer == NULL) {
619 			pipeallocfail++;
620 			if (ppsratecheck(&lastfail, &curfail, 1))
621 				printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n");
622 		} else {
623 			piperesizefail++;
624 		}
625 		return (ENOMEM);
626 	}
627 
628 	/* copy data, then free old resources if we're resizing */
629 	if (cnt > 0) {
630 		if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) {
631 			firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out;
632 			bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
633 				buffer, firstseg);
634 			if ((cnt - firstseg) > 0)
635 				bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg],
636 					cpipe->pipe_buffer.in);
637 		} else {
638 			bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
639 				buffer, cnt);
640 		}
641 	}
642 	pipe_free_kmem(cpipe);
643 	cpipe->pipe_buffer.buffer = buffer;
644 	cpipe->pipe_buffer.size = size;
645 	cpipe->pipe_buffer.in = cnt;
646 	cpipe->pipe_buffer.out = 0;
647 	cpipe->pipe_buffer.cnt = cnt;
648 	atomic_add_long(&amountpipekva, cpipe->pipe_buffer.size);
649 	return (0);
650 }
651 
652 /*
653  * Wrapper for pipespace_new() that performs locking assertions.
654  */
655 static int
pipespace(struct pipe * cpipe,int size)656 pipespace(struct pipe *cpipe, int size)
657 {
658 
659 	KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
660 	    ("Unlocked pipe passed to pipespace"));
661 	return (pipespace_new(cpipe, size));
662 }
663 
664 /*
665  * lock a pipe for I/O, blocking other access
666  */
667 static __inline int
pipelock(struct pipe * cpipe,bool catch)668 pipelock(struct pipe *cpipe, bool catch)
669 {
670 	int error, prio;
671 
672 	PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
673 
674 	prio = PRIBIO;
675 	if (catch)
676 		prio |= PCATCH;
677 	while (cpipe->pipe_state & PIPE_LOCKFL) {
678 		KASSERT(cpipe->pipe_waiters >= 0,
679 		    ("%s: bad waiter count %d", __func__,
680 		    cpipe->pipe_waiters));
681 		cpipe->pipe_waiters++;
682 		error = msleep(&cpipe->pipe_waiters, PIPE_MTX(cpipe), prio,
683 		    "pipelk", 0);
684 		cpipe->pipe_waiters--;
685 		if (error != 0)
686 			return (error);
687 	}
688 	cpipe->pipe_state |= PIPE_LOCKFL;
689 	return (0);
690 }
691 
692 /*
693  * unlock a pipe I/O lock
694  */
695 static __inline void
pipeunlock(struct pipe * cpipe)696 pipeunlock(struct pipe *cpipe)
697 {
698 
699 	PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
700 	KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
701 		("Unlocked pipe passed to pipeunlock"));
702 	KASSERT(cpipe->pipe_waiters >= 0,
703 	    ("%s: bad waiter count %d", __func__,
704 	    cpipe->pipe_waiters));
705 	cpipe->pipe_state &= ~PIPE_LOCKFL;
706 	if (cpipe->pipe_waiters > 0)
707 		wakeup_one(&cpipe->pipe_waiters);
708 }
709 
710 void
pipeselwakeup(struct pipe * cpipe)711 pipeselwakeup(struct pipe *cpipe)
712 {
713 
714 	PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
715 	if (cpipe->pipe_state & PIPE_SEL) {
716 		selwakeuppri(&cpipe->pipe_sel, PSOCK);
717 		if (!SEL_WAITING(&cpipe->pipe_sel))
718 			cpipe->pipe_state &= ~PIPE_SEL;
719 	}
720 	if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
721 		pgsigio(&cpipe->pipe_sigio, SIGIO, 0);
722 	KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0);
723 }
724 
725 /*
726  * Initialize and allocate VM and memory for pipe.  The structure
727  * will start out zero'd from the ctor, so we just manage the kmem.
728  */
729 static int
pipe_create(struct pipe * pipe,bool large_backing)730 pipe_create(struct pipe *pipe, bool large_backing)
731 {
732 	int error;
733 
734 	error = pipespace_new(pipe, !large_backing || amountpipekva >
735 	    maxpipekva / 2 ? SMALL_PIPE_SIZE : PIPE_SIZE);
736 	if (error == 0)
737 		pipe->pipe_ino = alloc_unr64(&pipeino_unr);
738 	return (error);
739 }
740 
741 static void
pipe_destroy(struct pipe * pipe)742 pipe_destroy(struct pipe *pipe)
743 {
744 	pipe_free_kmem(pipe);
745 	/*
746 	 * Note: we "leak" pipe_ino -- by design the alloc_unr64 mechanism does
747 	 * not undo allocations.
748 	 */
749 }
750 
751 /* ARGSUSED */
752 static int
pipe_read(struct file * fp,struct uio * uio,struct ucred * active_cred,int flags,struct thread * td)753 pipe_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
754     int flags, struct thread *td)
755 {
756 	struct pipe *rpipe;
757 	int error;
758 	int nread = 0;
759 	int size;
760 
761 	rpipe = fp->f_data;
762 
763 	/*
764 	 * Try to avoid locking the pipe if we have nothing to do.
765 	 *
766 	 * There are programs which share one pipe amongst multiple processes
767 	 * and perform non-blocking reads in parallel, even if the pipe is
768 	 * empty.  This in particular is the case with BSD make, which when
769 	 * spawned with a high -j number can find itself with over half of the
770 	 * calls failing to find anything.
771 	 */
772 	if ((fp->f_flag & FNONBLOCK) != 0 && !mac_pipe_check_read_enabled()) {
773 		if (__predict_false(uio->uio_resid == 0))
774 			return (0);
775 		if ((atomic_load_short(&rpipe->pipe_state) & PIPE_EOF) == 0 &&
776 		    atomic_load_int(&rpipe->pipe_buffer.cnt) == 0 &&
777 		    atomic_load_int(&rpipe->pipe_pages.cnt) == 0)
778 			return (EAGAIN);
779 	}
780 
781 	PIPE_LOCK(rpipe);
782 	++rpipe->pipe_busy;
783 	error = pipelock(rpipe, true);
784 	if (error)
785 		goto unlocked_error;
786 
787 #ifdef MAC
788 	error = mac_pipe_check_read(active_cred, rpipe->pipe_pair);
789 	if (error)
790 		goto locked_error;
791 #endif
792 	if (amountpipekva > (3 * maxpipekva) / 4) {
793 		if ((rpipe->pipe_state & PIPE_DIRECTW) == 0 &&
794 		    rpipe->pipe_buffer.size > SMALL_PIPE_SIZE &&
795 		    rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE &&
796 		    piperesizeallowed == 1) {
797 			PIPE_UNLOCK(rpipe);
798 			pipespace(rpipe, SMALL_PIPE_SIZE);
799 			PIPE_LOCK(rpipe);
800 		}
801 	}
802 
803 	while (uio->uio_resid) {
804 		/*
805 		 * normal pipe buffer receive
806 		 */
807 		if (rpipe->pipe_buffer.cnt > 0) {
808 			size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
809 			if (size > rpipe->pipe_buffer.cnt)
810 				size = rpipe->pipe_buffer.cnt;
811 			if (size > uio->uio_resid)
812 				size = uio->uio_resid;
813 
814 			PIPE_UNLOCK(rpipe);
815 			error = uiomove(
816 			    &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
817 			    size, uio);
818 			PIPE_LOCK(rpipe);
819 			if (error)
820 				break;
821 
822 			rpipe->pipe_buffer.out += size;
823 			if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
824 				rpipe->pipe_buffer.out = 0;
825 
826 			rpipe->pipe_buffer.cnt -= size;
827 
828 			/*
829 			 * If there is no more to read in the pipe, reset
830 			 * its pointers to the beginning.  This improves
831 			 * cache hit stats.
832 			 */
833 			if (rpipe->pipe_buffer.cnt == 0) {
834 				rpipe->pipe_buffer.in = 0;
835 				rpipe->pipe_buffer.out = 0;
836 			}
837 			nread += size;
838 #ifndef PIPE_NODIRECT
839 		/*
840 		 * Direct copy, bypassing a kernel buffer.
841 		 */
842 		} else if ((size = rpipe->pipe_pages.cnt) != 0) {
843 			if (size > uio->uio_resid)
844 				size = (u_int) uio->uio_resid;
845 			PIPE_UNLOCK(rpipe);
846 			error = uiomove_fromphys(rpipe->pipe_pages.ms,
847 			    rpipe->pipe_pages.pos, size, uio);
848 			PIPE_LOCK(rpipe);
849 			if (error)
850 				break;
851 			nread += size;
852 			rpipe->pipe_pages.pos += size;
853 			rpipe->pipe_pages.cnt -= size;
854 			if (rpipe->pipe_pages.cnt == 0) {
855 				rpipe->pipe_state &= ~PIPE_WANTW;
856 				wakeup(rpipe);
857 			}
858 #endif
859 		} else {
860 			/*
861 			 * detect EOF condition
862 			 * read returns 0 on EOF, no need to set error
863 			 */
864 			if (rpipe->pipe_state & PIPE_EOF)
865 				break;
866 
867 			/*
868 			 * If the "write-side" has been blocked, wake it up now.
869 			 */
870 			if (rpipe->pipe_state & PIPE_WANTW) {
871 				rpipe->pipe_state &= ~PIPE_WANTW;
872 				wakeup(rpipe);
873 			}
874 
875 			/*
876 			 * Break if some data was read.
877 			 */
878 			if (nread > 0)
879 				break;
880 
881 			/*
882 			 * Unlock the pipe buffer for our remaining processing.
883 			 * We will either break out with an error or we will
884 			 * sleep and relock to loop.
885 			 */
886 			pipeunlock(rpipe);
887 
888 			/*
889 			 * Handle non-blocking mode operation or
890 			 * wait for more data.
891 			 */
892 			if (fp->f_flag & FNONBLOCK) {
893 				error = EAGAIN;
894 			} else {
895 				rpipe->pipe_state |= PIPE_WANTR;
896 				if ((error = msleep(rpipe, PIPE_MTX(rpipe),
897 				    PRIBIO | PCATCH,
898 				    "piperd", 0)) == 0)
899 					error = pipelock(rpipe, true);
900 			}
901 			if (error)
902 				goto unlocked_error;
903 		}
904 	}
905 #ifdef MAC
906 locked_error:
907 #endif
908 	pipeunlock(rpipe);
909 
910 	/* XXX: should probably do this before getting any locks. */
911 	if (error == 0)
912 		pipe_timestamp(&rpipe->pipe_atime);
913 unlocked_error:
914 	--rpipe->pipe_busy;
915 
916 	/*
917 	 * PIPE_WANT processing only makes sense if pipe_busy is 0.
918 	 */
919 	if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
920 		rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
921 		wakeup(rpipe);
922 	} else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
923 		/*
924 		 * Handle write blocking hysteresis.
925 		 */
926 		if (rpipe->pipe_state & PIPE_WANTW) {
927 			rpipe->pipe_state &= ~PIPE_WANTW;
928 			wakeup(rpipe);
929 		}
930 	}
931 
932 	/*
933 	 * Only wake up writers if there was actually something read.
934 	 * Otherwise, when calling read(2) at EOF, a spurious wakeup occurs.
935 	 */
936 	if (nread > 0 &&
937 	    rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt >= PIPE_BUF)
938 		pipeselwakeup(rpipe);
939 
940 	PIPE_UNLOCK(rpipe);
941 	if (nread > 0)
942 		td->td_ru.ru_msgrcv++;
943 	return (error);
944 }
945 
946 #ifndef PIPE_NODIRECT
947 /*
948  * Map the sending processes' buffer into kernel space and wire it.
949  * This is similar to a physical write operation.
950  */
951 static int
pipe_build_write_buffer(struct pipe * wpipe,struct uio * uio)952 pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio)
953 {
954 	u_int size;
955 	int i;
956 
957 	PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
958 	KASSERT((wpipe->pipe_state & PIPE_DIRECTW) == 0,
959 	    ("%s: PIPE_DIRECTW set on %p", __func__, wpipe));
960 	KASSERT(wpipe->pipe_pages.cnt == 0,
961 	    ("%s: pipe map for %p contains residual data", __func__, wpipe));
962 
963 	if (uio->uio_iov->iov_len > wpipe->pipe_buffer.size)
964                 size = wpipe->pipe_buffer.size;
965 	else
966                 size = uio->uio_iov->iov_len;
967 
968 	wpipe->pipe_state |= PIPE_DIRECTW;
969 	PIPE_UNLOCK(wpipe);
970 	i = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
971 	    (vm_offset_t)uio->uio_iov->iov_base, size, VM_PROT_READ,
972 	    wpipe->pipe_pages.ms, PIPENPAGES);
973 	PIPE_LOCK(wpipe);
974 	if (i < 0) {
975 		wpipe->pipe_state &= ~PIPE_DIRECTW;
976 		return (EFAULT);
977 	}
978 
979 	wpipe->pipe_pages.npages = i;
980 	wpipe->pipe_pages.pos =
981 	    ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
982 	wpipe->pipe_pages.cnt = size;
983 
984 	uio->uio_iov->iov_len -= size;
985 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size;
986 	if (uio->uio_iov->iov_len == 0) {
987 		uio->uio_iov++;
988 		uio->uio_iovcnt--;
989 	}
990 	uio->uio_resid -= size;
991 	uio->uio_offset += size;
992 	return (0);
993 }
994 
995 /*
996  * Unwire the process buffer.
997  */
998 static void
pipe_destroy_write_buffer(struct pipe * wpipe)999 pipe_destroy_write_buffer(struct pipe *wpipe)
1000 {
1001 
1002 	PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
1003 	KASSERT((wpipe->pipe_state & PIPE_DIRECTW) != 0,
1004 	    ("%s: PIPE_DIRECTW not set on %p", __func__, wpipe));
1005 	KASSERT(wpipe->pipe_pages.cnt == 0,
1006 	    ("%s: pipe map for %p contains residual data", __func__, wpipe));
1007 
1008 	wpipe->pipe_state &= ~PIPE_DIRECTW;
1009 	vm_page_unhold_pages(wpipe->pipe_pages.ms, wpipe->pipe_pages.npages);
1010 	wpipe->pipe_pages.npages = 0;
1011 }
1012 
1013 /*
1014  * In the case of a signal, the writing process might go away.  This
1015  * code copies the data into the circular buffer so that the source
1016  * pages can be freed without loss of data.
1017  */
1018 static void
pipe_clone_write_buffer(struct pipe * wpipe)1019 pipe_clone_write_buffer(struct pipe *wpipe)
1020 {
1021 	struct uio uio;
1022 	struct iovec iov;
1023 	int size;
1024 	int pos;
1025 
1026 	PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
1027 	KASSERT((wpipe->pipe_state & PIPE_DIRECTW) != 0,
1028 	    ("%s: PIPE_DIRECTW not set on %p", __func__, wpipe));
1029 
1030 	size = wpipe->pipe_pages.cnt;
1031 	pos = wpipe->pipe_pages.pos;
1032 	wpipe->pipe_pages.cnt = 0;
1033 
1034 	wpipe->pipe_buffer.in = size;
1035 	wpipe->pipe_buffer.out = 0;
1036 	wpipe->pipe_buffer.cnt = size;
1037 
1038 	PIPE_UNLOCK(wpipe);
1039 	iov.iov_base = wpipe->pipe_buffer.buffer;
1040 	iov.iov_len = size;
1041 	uio.uio_iov = &iov;
1042 	uio.uio_iovcnt = 1;
1043 	uio.uio_offset = 0;
1044 	uio.uio_resid = size;
1045 	uio.uio_segflg = UIO_SYSSPACE;
1046 	uio.uio_rw = UIO_READ;
1047 	uio.uio_td = curthread;
1048 	uiomove_fromphys(wpipe->pipe_pages.ms, pos, size, &uio);
1049 	PIPE_LOCK(wpipe);
1050 	pipe_destroy_write_buffer(wpipe);
1051 }
1052 
1053 /*
1054  * This implements the pipe buffer write mechanism.  Note that only
1055  * a direct write OR a normal pipe write can be pending at any given time.
1056  * If there are any characters in the pipe buffer, the direct write will
1057  * be deferred until the receiving process grabs all of the bytes from
1058  * the pipe buffer.  Then the direct mapping write is set-up.
1059  */
1060 static int
pipe_direct_write(struct pipe * wpipe,struct uio * uio)1061 pipe_direct_write(struct pipe *wpipe, struct uio *uio)
1062 {
1063 	int error;
1064 
1065 retry:
1066 	PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
1067 	if ((wpipe->pipe_state & PIPE_EOF) != 0) {
1068 		error = EPIPE;
1069 		goto error1;
1070 	}
1071 	if (wpipe->pipe_state & PIPE_DIRECTW) {
1072 		if (wpipe->pipe_state & PIPE_WANTR) {
1073 			wpipe->pipe_state &= ~PIPE_WANTR;
1074 			wakeup(wpipe);
1075 		}
1076 		pipeselwakeup(wpipe);
1077 		wpipe->pipe_state |= PIPE_WANTW;
1078 		pipeunlock(wpipe);
1079 		error = msleep(wpipe, PIPE_MTX(wpipe),
1080 		    PRIBIO | PCATCH, "pipdww", 0);
1081 		pipelock(wpipe, false);
1082 		if (error != 0)
1083 			goto error1;
1084 		goto retry;
1085 	}
1086 	if (wpipe->pipe_buffer.cnt > 0) {
1087 		if (wpipe->pipe_state & PIPE_WANTR) {
1088 			wpipe->pipe_state &= ~PIPE_WANTR;
1089 			wakeup(wpipe);
1090 		}
1091 		pipeselwakeup(wpipe);
1092 		wpipe->pipe_state |= PIPE_WANTW;
1093 		pipeunlock(wpipe);
1094 		error = msleep(wpipe, PIPE_MTX(wpipe),
1095 		    PRIBIO | PCATCH, "pipdwc", 0);
1096 		pipelock(wpipe, false);
1097 		if (error != 0)
1098 			goto error1;
1099 		goto retry;
1100 	}
1101 
1102 	error = pipe_build_write_buffer(wpipe, uio);
1103 	if (error) {
1104 		goto error1;
1105 	}
1106 
1107 	while (wpipe->pipe_pages.cnt != 0 &&
1108 	    (wpipe->pipe_state & PIPE_EOF) == 0) {
1109 		if (wpipe->pipe_state & PIPE_WANTR) {
1110 			wpipe->pipe_state &= ~PIPE_WANTR;
1111 			wakeup(wpipe);
1112 		}
1113 		pipeselwakeup(wpipe);
1114 		wpipe->pipe_state |= PIPE_WANTW;
1115 		pipeunlock(wpipe);
1116 		error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH,
1117 		    "pipdwt", 0);
1118 		pipelock(wpipe, false);
1119 		if (error != 0)
1120 			break;
1121 	}
1122 
1123 	if ((wpipe->pipe_state & PIPE_EOF) != 0) {
1124 		wpipe->pipe_pages.cnt = 0;
1125 		pipe_destroy_write_buffer(wpipe);
1126 		pipeselwakeup(wpipe);
1127 		error = EPIPE;
1128 	} else if (error == EINTR || error == ERESTART) {
1129 		pipe_clone_write_buffer(wpipe);
1130 	} else {
1131 		pipe_destroy_write_buffer(wpipe);
1132 	}
1133 	KASSERT((wpipe->pipe_state & PIPE_DIRECTW) == 0,
1134 	    ("pipe %p leaked PIPE_DIRECTW", wpipe));
1135 	return (error);
1136 
1137 error1:
1138 	wakeup(wpipe);
1139 	return (error);
1140 }
1141 #endif
1142 
1143 static int
pipe_write(struct file * fp,struct uio * uio,struct ucred * active_cred,int flags,struct thread * td)1144 pipe_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
1145     int flags, struct thread *td)
1146 {
1147 	struct pipe *wpipe, *rpipe;
1148 	ssize_t orig_resid;
1149 	int desiredsize, error;
1150 
1151 	rpipe = fp->f_data;
1152 	wpipe = PIPE_PEER(rpipe);
1153 	PIPE_LOCK(rpipe);
1154 	error = pipelock(wpipe, true);
1155 	if (error) {
1156 		PIPE_UNLOCK(rpipe);
1157 		return (error);
1158 	}
1159 	/*
1160 	 * detect loss of pipe read side, issue SIGPIPE if lost.
1161 	 */
1162 	if (wpipe->pipe_present != PIPE_ACTIVE ||
1163 	    (wpipe->pipe_state & PIPE_EOF)) {
1164 		pipeunlock(wpipe);
1165 		PIPE_UNLOCK(rpipe);
1166 		return (EPIPE);
1167 	}
1168 #ifdef MAC
1169 	error = mac_pipe_check_write(active_cred, wpipe->pipe_pair);
1170 	if (error) {
1171 		pipeunlock(wpipe);
1172 		PIPE_UNLOCK(rpipe);
1173 		return (error);
1174 	}
1175 #endif
1176 	++wpipe->pipe_busy;
1177 
1178 	/* Choose a larger size if it's advantageous */
1179 	desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size);
1180 	while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) {
1181 		if (piperesizeallowed != 1)
1182 			break;
1183 		if (amountpipekva > maxpipekva / 2)
1184 			break;
1185 		if (desiredsize == BIG_PIPE_SIZE)
1186 			break;
1187 		desiredsize = desiredsize * 2;
1188 	}
1189 
1190 	/* Choose a smaller size if we're in a OOM situation */
1191 	if (amountpipekva > (3 * maxpipekva) / 4 &&
1192 	    wpipe->pipe_buffer.size > SMALL_PIPE_SIZE &&
1193 	    wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE &&
1194 	    piperesizeallowed == 1)
1195 		desiredsize = SMALL_PIPE_SIZE;
1196 
1197 	/* Resize if the above determined that a new size was necessary */
1198 	if (desiredsize != wpipe->pipe_buffer.size &&
1199 	    (wpipe->pipe_state & PIPE_DIRECTW) == 0) {
1200 		PIPE_UNLOCK(wpipe);
1201 		pipespace(wpipe, desiredsize);
1202 		PIPE_LOCK(wpipe);
1203 	}
1204 	MPASS(wpipe->pipe_buffer.size != 0);
1205 
1206 	orig_resid = uio->uio_resid;
1207 
1208 	while (uio->uio_resid) {
1209 		int space;
1210 
1211 		if (wpipe->pipe_state & PIPE_EOF) {
1212 			error = EPIPE;
1213 			break;
1214 		}
1215 #ifndef PIPE_NODIRECT
1216 		/*
1217 		 * If the transfer is large, we can gain performance if
1218 		 * we do process-to-process copies directly.
1219 		 * If the write is non-blocking, we don't use the
1220 		 * direct write mechanism.
1221 		 *
1222 		 * The direct write mechanism will detect the reader going
1223 		 * away on us.
1224 		 */
1225 		if (uio->uio_segflg == UIO_USERSPACE &&
1226 		    uio->uio_iov->iov_len >= pipe_mindirect &&
1227 		    wpipe->pipe_buffer.size >= pipe_mindirect &&
1228 		    (fp->f_flag & FNONBLOCK) == 0) {
1229 			error = pipe_direct_write(wpipe, uio);
1230 			if (error != 0)
1231 				break;
1232 			continue;
1233 		}
1234 #endif
1235 
1236 		/*
1237 		 * Pipe buffered writes cannot be coincidental with
1238 		 * direct writes.  We wait until the currently executing
1239 		 * direct write is completed before we start filling the
1240 		 * pipe buffer.  We break out if a signal occurs or the
1241 		 * reader goes away.
1242 		 */
1243 		if (wpipe->pipe_pages.cnt != 0) {
1244 			if (wpipe->pipe_state & PIPE_WANTR) {
1245 				wpipe->pipe_state &= ~PIPE_WANTR;
1246 				wakeup(wpipe);
1247 			}
1248 			pipeselwakeup(wpipe);
1249 			wpipe->pipe_state |= PIPE_WANTW;
1250 			pipeunlock(wpipe);
1251 			error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH,
1252 			    "pipbww", 0);
1253 			pipelock(wpipe, false);
1254 			if (error != 0)
1255 				break;
1256 			continue;
1257 		}
1258 
1259 		space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1260 
1261 		/* Writes of size <= PIPE_BUF must be atomic. */
1262 		if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
1263 			space = 0;
1264 
1265 		if (space > 0) {
1266 			int size;	/* Transfer size */
1267 			int segsize;	/* first segment to transfer */
1268 
1269 			/*
1270 			 * Transfer size is minimum of uio transfer
1271 			 * and free space in pipe buffer.
1272 			 */
1273 			if (space > uio->uio_resid)
1274 				size = uio->uio_resid;
1275 			else
1276 				size = space;
1277 			/*
1278 			 * First segment to transfer is minimum of
1279 			 * transfer size and contiguous space in
1280 			 * pipe buffer.  If first segment to transfer
1281 			 * is less than the transfer size, we've got
1282 			 * a wraparound in the buffer.
1283 			 */
1284 			segsize = wpipe->pipe_buffer.size -
1285 				wpipe->pipe_buffer.in;
1286 			if (segsize > size)
1287 				segsize = size;
1288 
1289 			/* Transfer first segment */
1290 
1291 			PIPE_UNLOCK(rpipe);
1292 			error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1293 					segsize, uio);
1294 			PIPE_LOCK(rpipe);
1295 
1296 			if (error == 0 && segsize < size) {
1297 				KASSERT(wpipe->pipe_buffer.in + segsize ==
1298 					wpipe->pipe_buffer.size,
1299 					("Pipe buffer wraparound disappeared"));
1300 				/*
1301 				 * Transfer remaining part now, to
1302 				 * support atomic writes.  Wraparound
1303 				 * happened.
1304 				 */
1305 
1306 				PIPE_UNLOCK(rpipe);
1307 				error = uiomove(
1308 				    &wpipe->pipe_buffer.buffer[0],
1309 				    size - segsize, uio);
1310 				PIPE_LOCK(rpipe);
1311 			}
1312 			if (error == 0) {
1313 				wpipe->pipe_buffer.in += size;
1314 				if (wpipe->pipe_buffer.in >=
1315 				    wpipe->pipe_buffer.size) {
1316 					KASSERT(wpipe->pipe_buffer.in ==
1317 						size - segsize +
1318 						wpipe->pipe_buffer.size,
1319 						("Expected wraparound bad"));
1320 					wpipe->pipe_buffer.in = size - segsize;
1321 				}
1322 
1323 				wpipe->pipe_buffer.cnt += size;
1324 				KASSERT(wpipe->pipe_buffer.cnt <=
1325 					wpipe->pipe_buffer.size,
1326 					("Pipe buffer overflow"));
1327 			}
1328 			if (error != 0)
1329 				break;
1330 			continue;
1331 		} else {
1332 			/*
1333 			 * If the "read-side" has been blocked, wake it up now.
1334 			 */
1335 			if (wpipe->pipe_state & PIPE_WANTR) {
1336 				wpipe->pipe_state &= ~PIPE_WANTR;
1337 				wakeup(wpipe);
1338 			}
1339 
1340 			/*
1341 			 * don't block on non-blocking I/O
1342 			 */
1343 			if (fp->f_flag & FNONBLOCK) {
1344 				error = EAGAIN;
1345 				break;
1346 			}
1347 
1348 			/*
1349 			 * We have no more space and have something to offer,
1350 			 * wake up select/poll.
1351 			 */
1352 			pipeselwakeup(wpipe);
1353 
1354 			wpipe->pipe_state |= PIPE_WANTW;
1355 			pipeunlock(wpipe);
1356 			error = msleep(wpipe, PIPE_MTX(rpipe),
1357 			    PRIBIO | PCATCH, "pipewr", 0);
1358 			pipelock(wpipe, false);
1359 			if (error != 0)
1360 				break;
1361 			continue;
1362 		}
1363 	}
1364 
1365 	--wpipe->pipe_busy;
1366 
1367 	if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1368 		wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1369 		wakeup(wpipe);
1370 	} else if (wpipe->pipe_buffer.cnt > 0) {
1371 		/*
1372 		 * If we have put any characters in the buffer, we wake up
1373 		 * the reader.
1374 		 */
1375 		if (wpipe->pipe_state & PIPE_WANTR) {
1376 			wpipe->pipe_state &= ~PIPE_WANTR;
1377 			wakeup(wpipe);
1378 		}
1379 	}
1380 
1381 	/*
1382 	 * Don't return EPIPE if any byte was written.
1383 	 * EINTR and other interrupts are handled by generic I/O layer.
1384 	 * Do not pretend that I/O succeeded for obvious user error
1385 	 * like EFAULT.
1386 	 */
1387 	if (uio->uio_resid != orig_resid && error == EPIPE)
1388 		error = 0;
1389 
1390 	if (error == 0)
1391 		pipe_timestamp(&wpipe->pipe_mtime);
1392 
1393 	/*
1394 	 * We have something to offer,
1395 	 * wake up select/poll.
1396 	 */
1397 	if (wpipe->pipe_buffer.cnt)
1398 		pipeselwakeup(wpipe);
1399 
1400 	pipeunlock(wpipe);
1401 	PIPE_UNLOCK(rpipe);
1402 	if (uio->uio_resid != orig_resid)
1403 		td->td_ru.ru_msgsnd++;
1404 	return (error);
1405 }
1406 
1407 /* ARGSUSED */
1408 static int
pipe_truncate(struct file * fp,off_t length,struct ucred * active_cred,struct thread * td)1409 pipe_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1410     struct thread *td)
1411 {
1412 	struct pipe *cpipe;
1413 	int error;
1414 
1415 	cpipe = fp->f_data;
1416 	if (cpipe->pipe_type & PIPE_TYPE_NAMED)
1417 		error = vnops.fo_truncate(fp, length, active_cred, td);
1418 	else
1419 		error = invfo_truncate(fp, length, active_cred, td);
1420 	return (error);
1421 }
1422 
1423 /*
1424  * we implement a very minimal set of ioctls for compatibility with sockets.
1425  */
1426 static int
pipe_ioctl(struct file * fp,u_long cmd,void * data,struct ucred * active_cred,struct thread * td)1427 pipe_ioctl(struct file *fp, u_long cmd, void *data, struct ucred *active_cred,
1428     struct thread *td)
1429 {
1430 	struct pipe *mpipe = fp->f_data;
1431 	int error;
1432 
1433 	PIPE_LOCK(mpipe);
1434 
1435 #ifdef MAC
1436 	error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data);
1437 	if (error) {
1438 		PIPE_UNLOCK(mpipe);
1439 		return (error);
1440 	}
1441 #endif
1442 
1443 	error = 0;
1444 	switch (cmd) {
1445 	case FIONBIO:
1446 		break;
1447 
1448 	case FIOASYNC:
1449 		if (*(int *)data) {
1450 			mpipe->pipe_state |= PIPE_ASYNC;
1451 		} else {
1452 			mpipe->pipe_state &= ~PIPE_ASYNC;
1453 		}
1454 		break;
1455 
1456 	case FIONREAD:
1457 		if (!(fp->f_flag & FREAD)) {
1458 			*(int *)data = 0;
1459 			PIPE_UNLOCK(mpipe);
1460 			return (0);
1461 		}
1462 		if (mpipe->pipe_pages.cnt != 0)
1463 			*(int *)data = mpipe->pipe_pages.cnt;
1464 		else
1465 			*(int *)data = mpipe->pipe_buffer.cnt;
1466 		break;
1467 
1468 	case FIOSETOWN:
1469 		PIPE_UNLOCK(mpipe);
1470 		error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1471 		goto out_unlocked;
1472 
1473 	case FIOGETOWN:
1474 		*(int *)data = fgetown(&mpipe->pipe_sigio);
1475 		break;
1476 
1477 	/* This is deprecated, FIOSETOWN should be used instead. */
1478 	case TIOCSPGRP:
1479 		PIPE_UNLOCK(mpipe);
1480 		error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1481 		goto out_unlocked;
1482 
1483 	/* This is deprecated, FIOGETOWN should be used instead. */
1484 	case TIOCGPGRP:
1485 		*(int *)data = -fgetown(&mpipe->pipe_sigio);
1486 		break;
1487 
1488 	default:
1489 		error = ENOTTY;
1490 		break;
1491 	}
1492 	PIPE_UNLOCK(mpipe);
1493 out_unlocked:
1494 	return (error);
1495 }
1496 
1497 static int
pipe_poll(struct file * fp,int events,struct ucred * active_cred,struct thread * td)1498 pipe_poll(struct file *fp, int events, struct ucred *active_cred,
1499     struct thread *td)
1500 {
1501 	struct pipe *rpipe;
1502 	struct pipe *wpipe;
1503 	int levents, revents;
1504 #ifdef MAC
1505 	int error;
1506 #endif
1507 
1508 	revents = 0;
1509 	rpipe = fp->f_data;
1510 	wpipe = PIPE_PEER(rpipe);
1511 	PIPE_LOCK(rpipe);
1512 #ifdef MAC
1513 	error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair);
1514 	if (error)
1515 		goto locked_error;
1516 #endif
1517 	if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM))
1518 		if (rpipe->pipe_pages.cnt > 0 || rpipe->pipe_buffer.cnt > 0)
1519 			revents |= events & (POLLIN | POLLRDNORM);
1520 
1521 	if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM))
1522 		if (wpipe->pipe_present != PIPE_ACTIVE ||
1523 		    (wpipe->pipe_state & PIPE_EOF) ||
1524 		    ((wpipe->pipe_state & PIPE_DIRECTW) == 0 &&
1525 		     ((wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF ||
1526 			 wpipe->pipe_buffer.size == 0)))
1527 			revents |= events & (POLLOUT | POLLWRNORM);
1528 
1529 	levents = events &
1530 	    (POLLIN | POLLINIGNEOF | POLLPRI | POLLRDNORM | POLLRDBAND);
1531 	if (rpipe->pipe_type & PIPE_TYPE_NAMED && fp->f_flag & FREAD && levents &&
1532 	    fp->f_pipegen == rpipe->pipe_wgen)
1533 		events |= POLLINIGNEOF;
1534 
1535 	if ((events & POLLINIGNEOF) == 0) {
1536 		if (rpipe->pipe_state & PIPE_EOF) {
1537 			if (fp->f_flag & FREAD)
1538 				revents |= (events & (POLLIN | POLLRDNORM));
1539 			if (wpipe->pipe_present != PIPE_ACTIVE ||
1540 			    (wpipe->pipe_state & PIPE_EOF))
1541 				revents |= POLLHUP;
1542 		}
1543 	}
1544 
1545 	if (revents == 0) {
1546 		/*
1547 		 * Add ourselves regardless of eventmask as we have to return
1548 		 * POLLHUP even if it was not asked for.
1549 		 */
1550 		if ((fp->f_flag & FREAD) != 0) {
1551 			selrecord(td, &rpipe->pipe_sel);
1552 			if (SEL_WAITING(&rpipe->pipe_sel))
1553 				rpipe->pipe_state |= PIPE_SEL;
1554 		}
1555 
1556 		if ((fp->f_flag & FWRITE) != 0 &&
1557 		    wpipe->pipe_present == PIPE_ACTIVE) {
1558 			selrecord(td, &wpipe->pipe_sel);
1559 			if (SEL_WAITING(&wpipe->pipe_sel))
1560 				wpipe->pipe_state |= PIPE_SEL;
1561 		}
1562 	}
1563 #ifdef MAC
1564 locked_error:
1565 #endif
1566 	PIPE_UNLOCK(rpipe);
1567 
1568 	return (revents);
1569 }
1570 
1571 /*
1572  * We shouldn't need locks here as we're doing a read and this should
1573  * be a natural race.
1574  */
1575 static int
pipe_stat(struct file * fp,struct stat * ub,struct ucred * active_cred)1576 pipe_stat(struct file *fp, struct stat *ub, struct ucred *active_cred)
1577 {
1578 	struct pipe *pipe;
1579 #ifdef MAC
1580 	int error;
1581 #endif
1582 
1583 	pipe = fp->f_data;
1584 #ifdef MAC
1585 	if (mac_pipe_check_stat_enabled()) {
1586 		PIPE_LOCK(pipe);
1587 		error = mac_pipe_check_stat(active_cred, pipe->pipe_pair);
1588 		PIPE_UNLOCK(pipe);
1589 		if (error) {
1590 			return (error);
1591 		}
1592 	}
1593 #endif
1594 
1595 	/* For named pipes ask the underlying filesystem. */
1596 	if (pipe->pipe_type & PIPE_TYPE_NAMED) {
1597 		return (vnops.fo_stat(fp, ub, active_cred));
1598 	}
1599 
1600 	bzero(ub, sizeof(*ub));
1601 	ub->st_mode = S_IFIFO;
1602 	ub->st_blksize = PAGE_SIZE;
1603 	if (pipe->pipe_pages.cnt != 0)
1604 		ub->st_size = pipe->pipe_pages.cnt;
1605 	else
1606 		ub->st_size = pipe->pipe_buffer.cnt;
1607 	ub->st_blocks = howmany(ub->st_size, ub->st_blksize);
1608 	ub->st_atim = pipe->pipe_atime;
1609 	ub->st_mtim = pipe->pipe_mtime;
1610 	ub->st_ctim = pipe->pipe_ctime;
1611 	ub->st_uid = fp->f_cred->cr_uid;
1612 	ub->st_gid = fp->f_cred->cr_gid;
1613 	ub->st_dev = pipedev_ino;
1614 	ub->st_ino = pipe->pipe_ino;
1615 	/*
1616 	 * Left as 0: st_nlink, st_rdev, st_flags, st_gen.
1617 	 */
1618 	return (0);
1619 }
1620 
1621 /* ARGSUSED */
1622 static int
pipe_close(struct file * fp,struct thread * td)1623 pipe_close(struct file *fp, struct thread *td)
1624 {
1625 
1626 	if (fp->f_vnode != NULL)
1627 		return vnops.fo_close(fp, td);
1628 	fp->f_ops = &badfileops;
1629 	pipe_dtor(fp->f_data);
1630 	fp->f_data = NULL;
1631 	return (0);
1632 }
1633 
1634 static int
pipe_chmod(struct file * fp,mode_t mode,struct ucred * active_cred,struct thread * td)1635 pipe_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td)
1636 {
1637 	struct pipe *cpipe;
1638 	int error;
1639 
1640 	cpipe = fp->f_data;
1641 	if (cpipe->pipe_type & PIPE_TYPE_NAMED)
1642 		error = vn_chmod(fp, mode, active_cred, td);
1643 	else
1644 		error = invfo_chmod(fp, mode, active_cred, td);
1645 	return (error);
1646 }
1647 
1648 static int
pipe_chown(struct file * fp,uid_t uid,gid_t gid,struct ucred * active_cred,struct thread * td)1649 pipe_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
1650     struct thread *td)
1651 {
1652 	struct pipe *cpipe;
1653 	int error;
1654 
1655 	cpipe = fp->f_data;
1656 	if (cpipe->pipe_type & PIPE_TYPE_NAMED)
1657 		error = vn_chown(fp, uid, gid, active_cred, td);
1658 	else
1659 		error = invfo_chown(fp, uid, gid, active_cred, td);
1660 	return (error);
1661 }
1662 
1663 static int
pipe_fill_kinfo(struct file * fp,struct kinfo_file * kif,struct filedesc * fdp)1664 pipe_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
1665 {
1666 	struct pipe *pi;
1667 
1668 	if (fp->f_type == DTYPE_FIFO)
1669 		return (vn_fill_kinfo(fp, kif, fdp));
1670 	kif->kf_type = KF_TYPE_PIPE;
1671 	pi = fp->f_data;
1672 	kif->kf_un.kf_pipe.kf_pipe_addr = (uintptr_t)pi;
1673 	kif->kf_un.kf_pipe.kf_pipe_peer = (uintptr_t)pi->pipe_peer;
1674 	kif->kf_un.kf_pipe.kf_pipe_buffer_cnt = pi->pipe_buffer.cnt;
1675 	kif->kf_un.kf_pipe.kf_pipe_buffer_in = pi->pipe_buffer.in;
1676 	kif->kf_un.kf_pipe.kf_pipe_buffer_out = pi->pipe_buffer.out;
1677 	kif->kf_un.kf_pipe.kf_pipe_buffer_size = pi->pipe_buffer.size;
1678 	return (0);
1679 }
1680 
1681 static void
pipe_free_kmem(struct pipe * cpipe)1682 pipe_free_kmem(struct pipe *cpipe)
1683 {
1684 
1685 	PIPE_LOCK_ASSERT(cpipe, MA_NOTOWNED);
1686 
1687 	if (cpipe->pipe_buffer.buffer != NULL) {
1688 		atomic_subtract_long(&amountpipekva, cpipe->pipe_buffer.size);
1689 		chgpipecnt(cpipe->pipe_pair->pp_owner->cr_ruidinfo,
1690 		    -cpipe->pipe_buffer.size, 0);
1691 		vm_map_remove(pipe_map,
1692 		    (vm_offset_t)cpipe->pipe_buffer.buffer,
1693 		    (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size);
1694 		cpipe->pipe_buffer.buffer = NULL;
1695 	}
1696 #ifndef PIPE_NODIRECT
1697 	{
1698 		cpipe->pipe_pages.cnt = 0;
1699 		cpipe->pipe_pages.pos = 0;
1700 		cpipe->pipe_pages.npages = 0;
1701 	}
1702 #endif
1703 }
1704 
1705 /*
1706  * shutdown the pipe
1707  */
1708 static void
pipeclose(struct pipe * cpipe)1709 pipeclose(struct pipe *cpipe)
1710 {
1711 #ifdef MAC
1712 	struct pipepair *pp;
1713 #endif
1714 	struct pipe *ppipe;
1715 
1716 	KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL"));
1717 
1718 	PIPE_LOCK(cpipe);
1719 	pipelock(cpipe, false);
1720 #ifdef MAC
1721 	pp = cpipe->pipe_pair;
1722 #endif
1723 
1724 	/*
1725 	 * If the other side is blocked, wake it up saying that
1726 	 * we want to close it down.
1727 	 */
1728 	cpipe->pipe_state |= PIPE_EOF;
1729 	while (cpipe->pipe_busy) {
1730 		wakeup(cpipe);
1731 		cpipe->pipe_state |= PIPE_WANT;
1732 		pipeunlock(cpipe);
1733 		msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0);
1734 		pipelock(cpipe, false);
1735 	}
1736 
1737 	pipeselwakeup(cpipe);
1738 
1739 	/*
1740 	 * Disconnect from peer, if any.
1741 	 */
1742 	ppipe = cpipe->pipe_peer;
1743 	if (ppipe->pipe_present == PIPE_ACTIVE) {
1744 		ppipe->pipe_state |= PIPE_EOF;
1745 		wakeup(ppipe);
1746 		pipeselwakeup(ppipe);
1747 	}
1748 
1749 	/*
1750 	 * Mark this endpoint as free.  Release kmem resources.  We
1751 	 * don't mark this endpoint as unused until we've finished
1752 	 * doing that, or the pipe might disappear out from under
1753 	 * us.
1754 	 */
1755 	PIPE_UNLOCK(cpipe);
1756 	pipe_free_kmem(cpipe);
1757 	PIPE_LOCK(cpipe);
1758 	cpipe->pipe_present = PIPE_CLOSING;
1759 	pipeunlock(cpipe);
1760 
1761 	/*
1762 	 * knlist_clear() may sleep dropping the PIPE_MTX. Set the
1763 	 * PIPE_FINALIZED, that allows other end to free the
1764 	 * pipe_pair, only after the knotes are completely dismantled.
1765 	 */
1766 	knlist_clear(&cpipe->pipe_sel.si_note, 1);
1767 	cpipe->pipe_present = PIPE_FINALIZED;
1768 	seldrain(&cpipe->pipe_sel);
1769 	knlist_destroy(&cpipe->pipe_sel.si_note);
1770 
1771 	/*
1772 	 * If both endpoints are now closed, release the memory for the
1773 	 * pipe pair.  If not, unlock.
1774 	 */
1775 	if (ppipe->pipe_present == PIPE_FINALIZED) {
1776 		PIPE_UNLOCK(cpipe);
1777 		crfree(cpipe->pipe_pair->pp_owner);
1778 #ifdef MAC
1779 		mac_pipe_destroy(pp);
1780 #endif
1781 		uma_zfree(pipe_zone, cpipe->pipe_pair);
1782 	} else
1783 		PIPE_UNLOCK(cpipe);
1784 }
1785 
1786 /*ARGSUSED*/
1787 static int
pipe_kqfilter(struct file * fp,struct knote * kn)1788 pipe_kqfilter(struct file *fp, struct knote *kn)
1789 {
1790 	struct pipe *cpipe;
1791 
1792 	/*
1793 	 * If a filter is requested that is not supported by this file
1794 	 * descriptor, don't return an error, but also don't ever generate an
1795 	 * event.
1796 	 */
1797 	if ((kn->kn_filter == EVFILT_READ) && !(fp->f_flag & FREAD)) {
1798 		kn->kn_fop = &pipe_nfiltops;
1799 		return (0);
1800 	}
1801 	if ((kn->kn_filter == EVFILT_WRITE) && !(fp->f_flag & FWRITE)) {
1802 		kn->kn_fop = &pipe_nfiltops;
1803 		return (0);
1804 	}
1805 	cpipe = fp->f_data;
1806 	PIPE_LOCK(cpipe);
1807 	switch (kn->kn_filter) {
1808 	case EVFILT_READ:
1809 		kn->kn_fop = &pipe_rfiltops;
1810 		break;
1811 	case EVFILT_WRITE:
1812 		kn->kn_fop = &pipe_wfiltops;
1813 		if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) {
1814 			/* other end of pipe has been closed */
1815 			PIPE_UNLOCK(cpipe);
1816 			return (EPIPE);
1817 		}
1818 		cpipe = PIPE_PEER(cpipe);
1819 		break;
1820 	default:
1821 		if ((cpipe->pipe_type & PIPE_TYPE_NAMED) != 0) {
1822 			PIPE_UNLOCK(cpipe);
1823 			return (vnops.fo_kqfilter(fp, kn));
1824 		}
1825 		PIPE_UNLOCK(cpipe);
1826 		return (EINVAL);
1827 	}
1828 
1829 	kn->kn_hook = cpipe;
1830 	knlist_add(&cpipe->pipe_sel.si_note, kn, 1);
1831 	PIPE_UNLOCK(cpipe);
1832 	return (0);
1833 }
1834 
1835 static void
filt_pipedetach(struct knote * kn)1836 filt_pipedetach(struct knote *kn)
1837 {
1838 	struct pipe *cpipe = kn->kn_hook;
1839 
1840 	PIPE_LOCK(cpipe);
1841 	knlist_remove(&cpipe->pipe_sel.si_note, kn, 1);
1842 	PIPE_UNLOCK(cpipe);
1843 }
1844 
1845 /*ARGSUSED*/
1846 static int
filt_piperead(struct knote * kn,long hint)1847 filt_piperead(struct knote *kn, long hint)
1848 {
1849 	struct file *fp = kn->kn_fp;
1850 	struct pipe *rpipe = kn->kn_hook;
1851 
1852 	PIPE_LOCK_ASSERT(rpipe, MA_OWNED);
1853 	kn->kn_data = rpipe->pipe_buffer.cnt;
1854 	if (kn->kn_data == 0)
1855 		kn->kn_data = rpipe->pipe_pages.cnt;
1856 
1857 	if ((rpipe->pipe_state & PIPE_EOF) != 0 &&
1858 	    ((rpipe->pipe_type & PIPE_TYPE_NAMED) == 0 ||
1859 	    fp->f_pipegen != rpipe->pipe_wgen)) {
1860 		kn->kn_flags |= EV_EOF;
1861 		return (1);
1862 	}
1863 	kn->kn_flags &= ~EV_EOF;
1864 	return (kn->kn_data > 0);
1865 }
1866 
1867 /*ARGSUSED*/
1868 static int
filt_pipewrite(struct knote * kn,long hint)1869 filt_pipewrite(struct knote *kn, long hint)
1870 {
1871 	struct pipe *wpipe = kn->kn_hook;
1872 
1873 	/*
1874 	 * If this end of the pipe is closed, the knote was removed from the
1875 	 * knlist and the list lock (i.e., the pipe lock) is therefore not held.
1876 	 */
1877 	if (wpipe->pipe_present == PIPE_ACTIVE ||
1878 	    (wpipe->pipe_type & PIPE_TYPE_NAMED) != 0) {
1879 		PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
1880 
1881 		if (wpipe->pipe_state & PIPE_DIRECTW) {
1882 			kn->kn_data = 0;
1883 		} else if (wpipe->pipe_buffer.size > 0) {
1884 			kn->kn_data = wpipe->pipe_buffer.size -
1885 			    wpipe->pipe_buffer.cnt;
1886 		} else {
1887 			kn->kn_data = PIPE_BUF;
1888 		}
1889 	}
1890 
1891 	if (wpipe->pipe_present != PIPE_ACTIVE ||
1892 	    (wpipe->pipe_state & PIPE_EOF)) {
1893 		kn->kn_flags |= EV_EOF;
1894 		return (1);
1895 	}
1896 	kn->kn_flags &= ~EV_EOF;
1897 	return (kn->kn_data >= PIPE_BUF);
1898 }
1899 
1900 static void
filt_pipedetach_notsup(struct knote * kn)1901 filt_pipedetach_notsup(struct knote *kn)
1902 {
1903 
1904 }
1905 
1906 static int
filt_pipenotsup(struct knote * kn,long hint)1907 filt_pipenotsup(struct knote *kn, long hint)
1908 {
1909 
1910 	return (0);
1911 }
1912 
1913 static int
filt_pipedump(struct proc * p,struct knote * kn,struct kinfo_knote * kin)1914 filt_pipedump(struct proc *p, struct knote *kn,
1915     struct kinfo_knote *kin)
1916 {
1917 	struct pipe *pipe = kn->kn_hook;
1918 
1919 	kin->knt_extdata = KNOTE_EXTDATA_PIPE;
1920 	kin->knt_pipe.knt_pipe_ino = pipe->pipe_ino;
1921 	return (0);
1922 }
1923