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