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