2 * Copyright (c) 1996 John S. Dyson
3 * Copyright (c) 2012 Giovanni Trematerra
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
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
17 * 4. Modifications may be freely made to this file if the above conditions
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
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.
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.
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.
49 * In order to limit the resource use of pipes, two sysctls exist:
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.
55 * kern.ipc.pipekva - This read-only sysctl tracks the current amount of
56 * memory in use by pipes.
58 * Based on how large pipekva is relative to maxpipekva, the following
62 * New pipes are given 16K of memory backing, pipes may dynamically
63 * grow to as large as 64K where needed.
65 * New pipes are given 4K (or PAGE_SIZE) of memory backing,
66 * existing pipes may NOT grow.
68 * New pipes are given 4K (or PAGE_SIZE) of memory backing,
69 * existing pipes will be shrunk down to 4K whenever possible.
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
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.
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.
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.
92 #include <sys/cdefs.h>
93 __FBSDID("$FreeBSD$");
95 #include <sys/param.h>
96 #include <sys/systm.h>
98 #include <sys/fcntl.h>
100 #include <sys/filedesc.h>
101 #include <sys/filio.h>
102 #include <sys/kernel.h>
103 #include <sys/lock.h>
104 #include <sys/mutex.h>
105 #include <sys/ttycom.h>
106 #include <sys/stat.h>
107 #include <sys/malloc.h>
108 #include <sys/poll.h>
109 #include <sys/selinfo.h>
110 #include <sys/signalvar.h>
111 #include <sys/syscallsubr.h>
112 #include <sys/sysctl.h>
113 #include <sys/sysproto.h>
114 #include <sys/pipe.h>
115 #include <sys/proc.h>
116 #include <sys/vnode.h>
118 #include <sys/event.h>
120 #include <security/mac/mac_framework.h>
123 #include <vm/vm_param.h>
124 #include <vm/vm_object.h>
125 #include <vm/vm_kern.h>
126 #include <vm/vm_extern.h>
128 #include <vm/vm_map.h>
129 #include <vm/vm_page.h>
133 * Use this define if you want to disable *fancy* VM things. Expect an
134 * approx 30% decrease in transfer rate. This could be useful for
137 /* #define PIPE_NODIRECT */
139 #define PIPE_PEER(pipe) \
140 (((pipe)->pipe_state & PIPE_NAMED) ? (pipe) : ((pipe)->pipe_peer))
143 * interfaces to the outside world
145 static fo_rdwr_t pipe_read;
146 static fo_rdwr_t pipe_write;
147 static fo_truncate_t pipe_truncate;
148 static fo_ioctl_t pipe_ioctl;
149 static fo_poll_t pipe_poll;
150 static fo_kqfilter_t pipe_kqfilter;
151 static fo_stat_t pipe_stat;
152 static fo_close_t pipe_close;
153 static fo_chmod_t pipe_chmod;
154 static fo_chown_t pipe_chown;
156 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_flags = DFLAG_PASSABLE
170 static void filt_pipedetach(struct knote *kn);
171 static void filt_pipedetach_notsup(struct knote *kn);
172 static int filt_pipenotsup(struct knote *kn, long hint);
173 static int filt_piperead(struct knote *kn, long hint);
174 static int filt_pipewrite(struct knote *kn, long hint);
176 static struct filterops pipe_nfiltops = {
178 .f_detach = filt_pipedetach_notsup,
179 .f_event = filt_pipenotsup
181 static struct filterops pipe_rfiltops = {
183 .f_detach = filt_pipedetach,
184 .f_event = filt_piperead
186 static struct filterops pipe_wfiltops = {
188 .f_detach = filt_pipedetach,
189 .f_event = filt_pipewrite
193 * Default pipe buffer size(s), this can be kind-of large now because pipe
194 * space is pageable. The pipe code will try to maintain locality of
195 * reference for performance reasons, so small amounts of outstanding I/O
196 * will not wipe the cache.
198 #define MINPIPESIZE (PIPE_SIZE/3)
199 #define MAXPIPESIZE (2*PIPE_SIZE/3)
201 static long amountpipekva;
202 static int pipefragretry;
203 static int pipeallocfail;
204 static int piperesizefail;
205 static int piperesizeallowed = 1;
207 SYSCTL_LONG(_kern_ipc, OID_AUTO, maxpipekva, CTLFLAG_RDTUN,
208 &maxpipekva, 0, "Pipe KVA limit");
209 SYSCTL_LONG(_kern_ipc, OID_AUTO, pipekva, CTLFLAG_RD,
210 &amountpipekva, 0, "Pipe KVA usage");
211 SYSCTL_INT(_kern_ipc, OID_AUTO, pipefragretry, CTLFLAG_RD,
212 &pipefragretry, 0, "Pipe allocation retries due to fragmentation");
213 SYSCTL_INT(_kern_ipc, OID_AUTO, pipeallocfail, CTLFLAG_RD,
214 &pipeallocfail, 0, "Pipe allocation failures");
215 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizefail, CTLFLAG_RD,
216 &piperesizefail, 0, "Pipe resize failures");
217 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizeallowed, CTLFLAG_RW,
218 &piperesizeallowed, 0, "Pipe resizing allowed");
220 static void pipeinit(void *dummy __unused);
221 static void pipeclose(struct pipe *cpipe);
222 static void pipe_free_kmem(struct pipe *cpipe);
223 static int pipe_create(struct pipe *pipe, int backing);
224 static int pipe_paircreate(struct thread *td, struct pipepair **p_pp);
225 static __inline int pipelock(struct pipe *cpipe, int catch);
226 static __inline void pipeunlock(struct pipe *cpipe);
227 #ifndef PIPE_NODIRECT
228 static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio);
229 static void pipe_destroy_write_buffer(struct pipe *wpipe);
230 static int pipe_direct_write(struct pipe *wpipe, struct uio *uio);
231 static void pipe_clone_write_buffer(struct pipe *wpipe);
233 static int pipespace(struct pipe *cpipe, int size);
234 static int pipespace_new(struct pipe *cpipe, int size);
236 static int pipe_zone_ctor(void *mem, int size, void *arg, int flags);
237 static int pipe_zone_init(void *mem, int size, int flags);
238 static void pipe_zone_fini(void *mem, int size);
240 static uma_zone_t pipe_zone;
241 static struct unrhdr *pipeino_unr;
242 static dev_t pipedev_ino;
244 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL);
247 pipeinit(void *dummy __unused)
250 pipe_zone = uma_zcreate("pipe", sizeof(struct pipepair),
251 pipe_zone_ctor, NULL, pipe_zone_init, pipe_zone_fini,
253 KASSERT(pipe_zone != NULL, ("pipe_zone not initialized"));
254 pipeino_unr = new_unrhdr(1, INT32_MAX, NULL);
255 KASSERT(pipeino_unr != NULL, ("pipe fake inodes not initialized"));
256 pipedev_ino = devfs_alloc_cdp_inode();
257 KASSERT(pipedev_ino > 0, ("pipe dev inode not initialized"));
261 pipe_zone_ctor(void *mem, int size, void *arg, int flags)
264 struct pipe *rpipe, *wpipe;
266 KASSERT(size == sizeof(*pp), ("pipe_zone_ctor: wrong size"));
268 pp = (struct pipepair *)mem;
271 * We zero both pipe endpoints to make sure all the kmem pointers
272 * are NULL, flag fields are zero'd, etc. We timestamp both
273 * endpoints with the same time.
275 rpipe = &pp->pp_rpipe;
276 bzero(rpipe, sizeof(*rpipe));
277 vfs_timestamp(&rpipe->pipe_ctime);
278 rpipe->pipe_atime = rpipe->pipe_mtime = rpipe->pipe_ctime;
280 wpipe = &pp->pp_wpipe;
281 bzero(wpipe, sizeof(*wpipe));
282 wpipe->pipe_ctime = rpipe->pipe_ctime;
283 wpipe->pipe_atime = wpipe->pipe_mtime = rpipe->pipe_ctime;
285 rpipe->pipe_peer = wpipe;
286 rpipe->pipe_pair = pp;
287 wpipe->pipe_peer = rpipe;
288 wpipe->pipe_pair = pp;
291 * Mark both endpoints as present; they will later get free'd
292 * one at a time. When both are free'd, then the whole pair
295 rpipe->pipe_present = PIPE_ACTIVE;
296 wpipe->pipe_present = PIPE_ACTIVE;
299 * Eventually, the MAC Framework may initialize the label
300 * in ctor or init, but for now we do it elswhere to avoid
301 * blocking in ctor or init.
309 pipe_zone_init(void *mem, int size, int flags)
313 KASSERT(size == sizeof(*pp), ("pipe_zone_init: wrong size"));
315 pp = (struct pipepair *)mem;
317 mtx_init(&pp->pp_mtx, "pipe mutex", NULL, MTX_DEF | MTX_RECURSE);
322 pipe_zone_fini(void *mem, int size)
326 KASSERT(size == sizeof(*pp), ("pipe_zone_fini: wrong size"));
328 pp = (struct pipepair *)mem;
330 mtx_destroy(&pp->pp_mtx);
334 pipe_paircreate(struct thread *td, struct pipepair **p_pp)
337 struct pipe *rpipe, *wpipe;
340 *p_pp = pp = uma_zalloc(pipe_zone, M_WAITOK);
343 * The MAC label is shared between the connected endpoints. As a
344 * result mac_pipe_init() and mac_pipe_create() are called once
345 * for the pair, and not on the endpoints.
348 mac_pipe_create(td->td_ucred, pp);
350 rpipe = &pp->pp_rpipe;
351 wpipe = &pp->pp_wpipe;
353 knlist_init_mtx(&rpipe->pipe_sel.si_note, PIPE_MTX(rpipe));
354 knlist_init_mtx(&wpipe->pipe_sel.si_note, PIPE_MTX(wpipe));
356 /* Only the forward direction pipe is backed by default */
357 if ((error = pipe_create(rpipe, 1)) != 0 ||
358 (error = pipe_create(wpipe, 0)) != 0) {
364 rpipe->pipe_state |= PIPE_DIRECTOK;
365 wpipe->pipe_state |= PIPE_DIRECTOK;
370 pipe_named_ctor(struct pipe **ppipe, struct thread *td)
375 error = pipe_paircreate(td, &pp);
378 pp->pp_rpipe.pipe_state |= PIPE_NAMED;
379 *ppipe = &pp->pp_rpipe;
384 pipe_dtor(struct pipe *dpipe)
388 ino = dpipe->pipe_ino;
389 funsetown(&dpipe->pipe_sigio);
391 if (dpipe->pipe_state & PIPE_NAMED) {
392 dpipe = dpipe->pipe_peer;
393 funsetown(&dpipe->pipe_sigio);
396 if (ino != 0 && ino != (ino_t)-1)
397 free_unr(pipeino_unr, ino);
401 * The pipe system call for the DTYPE_PIPE type of pipes. If we fail, let
402 * the zone pick up the pieces via pipeclose().
405 kern_pipe(struct thread *td, int fildes[2])
408 return (kern_pipe2(td, fildes, 0));
412 kern_pipe2(struct thread *td, int fildes[2], int flags)
414 struct filedesc *fdp;
415 struct file *rf, *wf;
416 struct pipe *rpipe, *wpipe;
418 int fd, fflags, error;
420 fdp = td->td_proc->p_fd;
421 error = pipe_paircreate(td, &pp);
424 rpipe = &pp->pp_rpipe;
425 wpipe = &pp->pp_wpipe;
426 error = falloc(td, &rf, &fd, flags);
432 /* An extra reference on `rf' has been held for us by falloc(). */
435 fflags = FREAD | FWRITE;
436 if ((flags & O_NONBLOCK) != 0)
440 * Warning: once we've gotten past allocation of the fd for the
441 * read-side, we can only drop the read side via fdrop() in order
442 * to avoid races against processes which manage to dup() the read
443 * side while we are blocked trying to allocate the write side.
445 finit(rf, fflags, DTYPE_PIPE, rpipe, &pipeops);
446 error = falloc(td, &wf, &fd, flags);
448 fdclose(fdp, rf, fildes[0], td);
450 /* rpipe has been closed by fdrop(). */
454 /* An extra reference on `wf' has been held for us by falloc(). */
455 finit(wf, fflags, DTYPE_PIPE, wpipe, &pipeops);
465 sys_pipe(struct thread *td, struct pipe_args *uap)
470 error = kern_pipe(td, fildes);
474 td->td_retval[0] = fildes[0];
475 td->td_retval[1] = fildes[1];
481 * Allocate kva for pipe circular buffer, the space is pageable
482 * This routine will 'realloc' the size of a pipe safely, if it fails
483 * it will retain the old buffer.
484 * If it fails it will return ENOMEM.
487 pipespace_new(cpipe, size)
492 int error, cnt, firstseg;
493 static int curfail = 0;
494 static struct timeval lastfail;
496 KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked"));
497 KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW),
498 ("pipespace: resize of direct writes not allowed"));
500 cnt = cpipe->pipe_buffer.cnt;
504 size = round_page(size);
505 buffer = (caddr_t) vm_map_min(pipe_map);
507 error = vm_map_find(pipe_map, NULL, 0,
508 (vm_offset_t *) &buffer, size, 1,
509 VM_PROT_ALL, VM_PROT_ALL, 0);
510 if (error != KERN_SUCCESS) {
511 if ((cpipe->pipe_buffer.buffer == NULL) &&
512 (size > SMALL_PIPE_SIZE)) {
513 size = SMALL_PIPE_SIZE;
517 if (cpipe->pipe_buffer.buffer == NULL) {
519 if (ppsratecheck(&lastfail, &curfail, 1))
520 printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n");
527 /* copy data, then free old resources if we're resizing */
529 if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) {
530 firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out;
531 bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
533 if ((cnt - firstseg) > 0)
534 bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg],
535 cpipe->pipe_buffer.in);
537 bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
541 pipe_free_kmem(cpipe);
542 cpipe->pipe_buffer.buffer = buffer;
543 cpipe->pipe_buffer.size = size;
544 cpipe->pipe_buffer.in = cnt;
545 cpipe->pipe_buffer.out = 0;
546 cpipe->pipe_buffer.cnt = cnt;
547 atomic_add_long(&amountpipekva, cpipe->pipe_buffer.size);
552 * Wrapper for pipespace_new() that performs locking assertions.
555 pipespace(cpipe, size)
560 KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
561 ("Unlocked pipe passed to pipespace"));
562 return (pipespace_new(cpipe, size));
566 * lock a pipe for I/O, blocking other access
569 pipelock(cpipe, catch)
575 PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
576 while (cpipe->pipe_state & PIPE_LOCKFL) {
577 cpipe->pipe_state |= PIPE_LWANT;
578 error = msleep(cpipe, PIPE_MTX(cpipe),
579 catch ? (PRIBIO | PCATCH) : PRIBIO,
584 cpipe->pipe_state |= PIPE_LOCKFL;
589 * unlock a pipe I/O lock
596 PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
597 KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
598 ("Unlocked pipe passed to pipeunlock"));
599 cpipe->pipe_state &= ~PIPE_LOCKFL;
600 if (cpipe->pipe_state & PIPE_LWANT) {
601 cpipe->pipe_state &= ~PIPE_LWANT;
611 PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
612 if (cpipe->pipe_state & PIPE_SEL) {
613 selwakeuppri(&cpipe->pipe_sel, PSOCK);
614 if (!SEL_WAITING(&cpipe->pipe_sel))
615 cpipe->pipe_state &= ~PIPE_SEL;
617 if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
618 pgsigio(&cpipe->pipe_sigio, SIGIO, 0);
619 KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0);
623 * Initialize and allocate VM and memory for pipe. The structure
624 * will start out zero'd from the ctor, so we just manage the kmem.
627 pipe_create(pipe, backing)
634 if (amountpipekva > maxpipekva / 2)
635 error = pipespace_new(pipe, SMALL_PIPE_SIZE);
637 error = pipespace_new(pipe, PIPE_SIZE);
639 /* If we're not backing this pipe, no need to do anything. */
648 pipe_read(fp, uio, active_cred, flags, td)
651 struct ucred *active_cred;
663 error = pipelock(rpipe, 1);
668 error = mac_pipe_check_read(active_cred, rpipe->pipe_pair);
672 if (amountpipekva > (3 * maxpipekva) / 4) {
673 if (!(rpipe->pipe_state & PIPE_DIRECTW) &&
674 (rpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
675 (rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
676 (piperesizeallowed == 1)) {
678 pipespace(rpipe, SMALL_PIPE_SIZE);
683 while (uio->uio_resid) {
685 * normal pipe buffer receive
687 if (rpipe->pipe_buffer.cnt > 0) {
688 size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
689 if (size > rpipe->pipe_buffer.cnt)
690 size = rpipe->pipe_buffer.cnt;
691 if (size > uio->uio_resid)
692 size = uio->uio_resid;
696 &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
702 rpipe->pipe_buffer.out += size;
703 if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
704 rpipe->pipe_buffer.out = 0;
706 rpipe->pipe_buffer.cnt -= size;
709 * If there is no more to read in the pipe, reset
710 * its pointers to the beginning. This improves
713 if (rpipe->pipe_buffer.cnt == 0) {
714 rpipe->pipe_buffer.in = 0;
715 rpipe->pipe_buffer.out = 0;
718 #ifndef PIPE_NODIRECT
720 * Direct copy, bypassing a kernel buffer.
722 } else if ((size = rpipe->pipe_map.cnt) &&
723 (rpipe->pipe_state & PIPE_DIRECTW)) {
724 if (size > uio->uio_resid)
725 size = (u_int) uio->uio_resid;
728 error = uiomove_fromphys(rpipe->pipe_map.ms,
729 rpipe->pipe_map.pos, size, uio);
734 rpipe->pipe_map.pos += size;
735 rpipe->pipe_map.cnt -= size;
736 if (rpipe->pipe_map.cnt == 0) {
737 rpipe->pipe_state &= ~(PIPE_DIRECTW|PIPE_WANTW);
743 * detect EOF condition
744 * read returns 0 on EOF, no need to set error
746 if (rpipe->pipe_state & PIPE_EOF)
750 * If the "write-side" has been blocked, wake it up now.
752 if (rpipe->pipe_state & PIPE_WANTW) {
753 rpipe->pipe_state &= ~PIPE_WANTW;
758 * Break if some data was read.
764 * Unlock the pipe buffer for our remaining processing.
765 * We will either break out with an error or we will
766 * sleep and relock to loop.
771 * Handle non-blocking mode operation or
772 * wait for more data.
774 if (fp->f_flag & FNONBLOCK) {
777 rpipe->pipe_state |= PIPE_WANTR;
778 if ((error = msleep(rpipe, PIPE_MTX(rpipe),
781 error = pipelock(rpipe, 1);
792 /* XXX: should probably do this before getting any locks. */
794 vfs_timestamp(&rpipe->pipe_atime);
799 * PIPE_WANT processing only makes sense if pipe_busy is 0.
801 if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
802 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
804 } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
806 * Handle write blocking hysteresis.
808 if (rpipe->pipe_state & PIPE_WANTW) {
809 rpipe->pipe_state &= ~PIPE_WANTW;
814 if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
815 pipeselwakeup(rpipe);
821 #ifndef PIPE_NODIRECT
823 * Map the sending processes' buffer into kernel space and wire it.
824 * This is similar to a physical write operation.
827 pipe_build_write_buffer(wpipe, uio)
834 PIPE_LOCK_ASSERT(wpipe, MA_NOTOWNED);
835 KASSERT(wpipe->pipe_state & PIPE_DIRECTW,
836 ("Clone attempt on non-direct write pipe!"));
838 if (uio->uio_iov->iov_len > wpipe->pipe_buffer.size)
839 size = wpipe->pipe_buffer.size;
841 size = uio->uio_iov->iov_len;
843 if ((i = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
844 (vm_offset_t)uio->uio_iov->iov_base, size, VM_PROT_READ,
845 wpipe->pipe_map.ms, PIPENPAGES)) < 0)
849 * set up the control block
851 wpipe->pipe_map.npages = i;
852 wpipe->pipe_map.pos =
853 ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
854 wpipe->pipe_map.cnt = size;
857 * and update the uio data
860 uio->uio_iov->iov_len -= size;
861 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size;
862 if (uio->uio_iov->iov_len == 0)
864 uio->uio_resid -= size;
865 uio->uio_offset += size;
870 * unmap and unwire the process buffer
873 pipe_destroy_write_buffer(wpipe)
877 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
878 vm_page_unhold_pages(wpipe->pipe_map.ms, wpipe->pipe_map.npages);
879 wpipe->pipe_map.npages = 0;
883 * In the case of a signal, the writing process might go away. This
884 * code copies the data into the circular buffer so that the source
885 * pages can be freed without loss of data.
888 pipe_clone_write_buffer(wpipe)
896 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
897 size = wpipe->pipe_map.cnt;
898 pos = wpipe->pipe_map.pos;
900 wpipe->pipe_buffer.in = size;
901 wpipe->pipe_buffer.out = 0;
902 wpipe->pipe_buffer.cnt = size;
903 wpipe->pipe_state &= ~PIPE_DIRECTW;
906 iov.iov_base = wpipe->pipe_buffer.buffer;
911 uio.uio_resid = size;
912 uio.uio_segflg = UIO_SYSSPACE;
913 uio.uio_rw = UIO_READ;
914 uio.uio_td = curthread;
915 uiomove_fromphys(wpipe->pipe_map.ms, pos, size, &uio);
917 pipe_destroy_write_buffer(wpipe);
921 * This implements the pipe buffer write mechanism. Note that only
922 * a direct write OR a normal pipe write can be pending at any given time.
923 * If there are any characters in the pipe buffer, the direct write will
924 * be deferred until the receiving process grabs all of the bytes from
925 * the pipe buffer. Then the direct mapping write is set-up.
928 pipe_direct_write(wpipe, uio)
935 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
936 error = pipelock(wpipe, 1);
937 if (wpipe->pipe_state & PIPE_EOF)
943 while (wpipe->pipe_state & PIPE_DIRECTW) {
944 if (wpipe->pipe_state & PIPE_WANTR) {
945 wpipe->pipe_state &= ~PIPE_WANTR;
948 pipeselwakeup(wpipe);
949 wpipe->pipe_state |= PIPE_WANTW;
951 error = msleep(wpipe, PIPE_MTX(wpipe),
952 PRIBIO | PCATCH, "pipdww", 0);
958 wpipe->pipe_map.cnt = 0; /* transfer not ready yet */
959 if (wpipe->pipe_buffer.cnt > 0) {
960 if (wpipe->pipe_state & PIPE_WANTR) {
961 wpipe->pipe_state &= ~PIPE_WANTR;
964 pipeselwakeup(wpipe);
965 wpipe->pipe_state |= PIPE_WANTW;
967 error = msleep(wpipe, PIPE_MTX(wpipe),
968 PRIBIO | PCATCH, "pipdwc", 0);
975 wpipe->pipe_state |= PIPE_DIRECTW;
978 error = pipe_build_write_buffer(wpipe, uio);
981 wpipe->pipe_state &= ~PIPE_DIRECTW;
987 while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
988 if (wpipe->pipe_state & PIPE_EOF) {
989 pipe_destroy_write_buffer(wpipe);
990 pipeselwakeup(wpipe);
995 if (wpipe->pipe_state & PIPE_WANTR) {
996 wpipe->pipe_state &= ~PIPE_WANTR;
999 pipeselwakeup(wpipe);
1000 wpipe->pipe_state |= PIPE_WANTW;
1002 error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH,
1007 if (wpipe->pipe_state & PIPE_EOF)
1009 if (wpipe->pipe_state & PIPE_DIRECTW) {
1011 * this bit of trickery substitutes a kernel buffer for
1012 * the process that might be going away.
1014 pipe_clone_write_buffer(wpipe);
1016 pipe_destroy_write_buffer(wpipe);
1028 pipe_write(fp, uio, active_cred, flags, td)
1031 struct ucred *active_cred;
1038 struct pipe *wpipe, *rpipe;
1041 wpipe = PIPE_PEER(rpipe);
1043 error = pipelock(wpipe, 1);
1049 * detect loss of pipe read side, issue SIGPIPE if lost.
1051 if (wpipe->pipe_present != PIPE_ACTIVE ||
1052 (wpipe->pipe_state & PIPE_EOF)) {
1058 error = mac_pipe_check_write(active_cred, wpipe->pipe_pair);
1067 /* Choose a larger size if it's advantageous */
1068 desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size);
1069 while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) {
1070 if (piperesizeallowed != 1)
1072 if (amountpipekva > maxpipekva / 2)
1074 if (desiredsize == BIG_PIPE_SIZE)
1076 desiredsize = desiredsize * 2;
1079 /* Choose a smaller size if we're in a OOM situation */
1080 if ((amountpipekva > (3 * maxpipekva) / 4) &&
1081 (wpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
1082 (wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
1083 (piperesizeallowed == 1))
1084 desiredsize = SMALL_PIPE_SIZE;
1086 /* Resize if the above determined that a new size was necessary */
1087 if ((desiredsize != wpipe->pipe_buffer.size) &&
1088 ((wpipe->pipe_state & PIPE_DIRECTW) == 0)) {
1090 pipespace(wpipe, desiredsize);
1093 if (wpipe->pipe_buffer.size == 0) {
1095 * This can only happen for reverse direction use of pipes
1096 * in a complete OOM situation.
1107 orig_resid = uio->uio_resid;
1109 while (uio->uio_resid) {
1113 if (wpipe->pipe_state & PIPE_EOF) {
1118 #ifndef PIPE_NODIRECT
1120 * If the transfer is large, we can gain performance if
1121 * we do process-to-process copies directly.
1122 * If the write is non-blocking, we don't use the
1123 * direct write mechanism.
1125 * The direct write mechanism will detect the reader going
1128 if (uio->uio_segflg == UIO_USERSPACE &&
1129 uio->uio_iov->iov_len >= PIPE_MINDIRECT &&
1130 wpipe->pipe_buffer.size >= PIPE_MINDIRECT &&
1131 (fp->f_flag & FNONBLOCK) == 0) {
1133 error = pipe_direct_write(wpipe, uio);
1141 * Pipe buffered writes cannot be coincidental with
1142 * direct writes. We wait until the currently executing
1143 * direct write is completed before we start filling the
1144 * pipe buffer. We break out if a signal occurs or the
1147 if (wpipe->pipe_state & PIPE_DIRECTW) {
1148 if (wpipe->pipe_state & PIPE_WANTR) {
1149 wpipe->pipe_state &= ~PIPE_WANTR;
1152 pipeselwakeup(wpipe);
1153 wpipe->pipe_state |= PIPE_WANTW;
1155 error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH,
1163 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1165 /* Writes of size <= PIPE_BUF must be atomic. */
1166 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
1170 int size; /* Transfer size */
1171 int segsize; /* first segment to transfer */
1174 * Transfer size is minimum of uio transfer
1175 * and free space in pipe buffer.
1177 if (space > uio->uio_resid)
1178 size = uio->uio_resid;
1182 * First segment to transfer is minimum of
1183 * transfer size and contiguous space in
1184 * pipe buffer. If first segment to transfer
1185 * is less than the transfer size, we've got
1186 * a wraparound in the buffer.
1188 segsize = wpipe->pipe_buffer.size -
1189 wpipe->pipe_buffer.in;
1193 /* Transfer first segment */
1196 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1200 if (error == 0 && segsize < size) {
1201 KASSERT(wpipe->pipe_buffer.in + segsize ==
1202 wpipe->pipe_buffer.size,
1203 ("Pipe buffer wraparound disappeared"));
1205 * Transfer remaining part now, to
1206 * support atomic writes. Wraparound
1212 &wpipe->pipe_buffer.buffer[0],
1213 size - segsize, uio);
1217 wpipe->pipe_buffer.in += size;
1218 if (wpipe->pipe_buffer.in >=
1219 wpipe->pipe_buffer.size) {
1220 KASSERT(wpipe->pipe_buffer.in ==
1222 wpipe->pipe_buffer.size,
1223 ("Expected wraparound bad"));
1224 wpipe->pipe_buffer.in = size - segsize;
1227 wpipe->pipe_buffer.cnt += size;
1228 KASSERT(wpipe->pipe_buffer.cnt <=
1229 wpipe->pipe_buffer.size,
1230 ("Pipe buffer overflow"));
1237 * If the "read-side" has been blocked, wake it up now.
1239 if (wpipe->pipe_state & PIPE_WANTR) {
1240 wpipe->pipe_state &= ~PIPE_WANTR;
1245 * don't block on non-blocking I/O
1247 if (fp->f_flag & FNONBLOCK) {
1254 * We have no more space and have something to offer,
1255 * wake up select/poll.
1257 pipeselwakeup(wpipe);
1259 wpipe->pipe_state |= PIPE_WANTW;
1261 error = msleep(wpipe, PIPE_MTX(rpipe),
1262 PRIBIO | PCATCH, "pipewr", 0);
1271 if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1272 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1274 } else if (wpipe->pipe_buffer.cnt > 0) {
1276 * If we have put any characters in the buffer, we wake up
1279 if (wpipe->pipe_state & PIPE_WANTR) {
1280 wpipe->pipe_state &= ~PIPE_WANTR;
1286 * Don't return EPIPE if I/O was successful
1288 if ((wpipe->pipe_buffer.cnt == 0) &&
1289 (uio->uio_resid == 0) &&
1295 vfs_timestamp(&wpipe->pipe_mtime);
1298 * We have something to offer,
1299 * wake up select/poll.
1301 if (wpipe->pipe_buffer.cnt)
1302 pipeselwakeup(wpipe);
1311 pipe_truncate(fp, length, active_cred, td)
1314 struct ucred *active_cred;
1318 /* For named pipes call the vnode operation. */
1319 if (fp->f_vnode != NULL)
1320 return (vnops.fo_truncate(fp, length, active_cred, td));
1325 * we implement a very minimal set of ioctls for compatibility with sockets.
1328 pipe_ioctl(fp, cmd, data, active_cred, td)
1332 struct ucred *active_cred;
1335 struct pipe *mpipe = fp->f_data;
1341 error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data);
1356 mpipe->pipe_state |= PIPE_ASYNC;
1358 mpipe->pipe_state &= ~PIPE_ASYNC;
1363 if (!(fp->f_flag & FREAD)) {
1368 if (mpipe->pipe_state & PIPE_DIRECTW)
1369 *(int *)data = mpipe->pipe_map.cnt;
1371 *(int *)data = mpipe->pipe_buffer.cnt;
1376 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1380 *(int *)data = fgetown(&mpipe->pipe_sigio);
1383 /* This is deprecated, FIOSETOWN should be used instead. */
1386 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1389 /* This is deprecated, FIOGETOWN should be used instead. */
1391 *(int *)data = -fgetown(&mpipe->pipe_sigio);
1404 pipe_poll(fp, events, active_cred, td)
1407 struct ucred *active_cred;
1412 int levents, revents;
1419 wpipe = PIPE_PEER(rpipe);
1422 error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair);
1426 if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM))
1427 if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1428 (rpipe->pipe_buffer.cnt > 0))
1429 revents |= events & (POLLIN | POLLRDNORM);
1431 if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM))
1432 if (wpipe->pipe_present != PIPE_ACTIVE ||
1433 (wpipe->pipe_state & PIPE_EOF) ||
1434 (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1435 ((wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF ||
1436 wpipe->pipe_buffer.size == 0)))
1437 revents |= events & (POLLOUT | POLLWRNORM);
1440 (POLLIN | POLLINIGNEOF | POLLPRI | POLLRDNORM | POLLRDBAND);
1441 if (rpipe->pipe_state & PIPE_NAMED && fp->f_flag & FREAD && levents &&
1442 fp->f_seqcount == rpipe->pipe_wgen)
1443 events |= POLLINIGNEOF;
1445 if ((events & POLLINIGNEOF) == 0) {
1446 if (rpipe->pipe_state & PIPE_EOF) {
1447 revents |= (events & (POLLIN | POLLRDNORM));
1448 if (wpipe->pipe_present != PIPE_ACTIVE ||
1449 (wpipe->pipe_state & PIPE_EOF))
1455 if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM)) {
1456 selrecord(td, &rpipe->pipe_sel);
1457 if (SEL_WAITING(&rpipe->pipe_sel))
1458 rpipe->pipe_state |= PIPE_SEL;
1461 if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM)) {
1462 selrecord(td, &wpipe->pipe_sel);
1463 if (SEL_WAITING(&wpipe->pipe_sel))
1464 wpipe->pipe_state |= PIPE_SEL;
1476 * We shouldn't need locks here as we're doing a read and this should
1477 * be a natural race.
1480 pipe_stat(fp, ub, active_cred, td)
1483 struct ucred *active_cred;
1495 error = mac_pipe_check_stat(active_cred, pipe->pipe_pair);
1502 /* For named pipes ask the underlying filesystem. */
1503 if (pipe->pipe_state & PIPE_NAMED) {
1505 return (vnops.fo_stat(fp, ub, active_cred, td));
1509 * Lazily allocate an inode number for the pipe. Most pipe
1510 * users do not call fstat(2) on the pipe, which means that
1511 * postponing the inode allocation until it is must be
1512 * returned to userland is useful. If alloc_unr failed,
1513 * assign st_ino zero instead of returning an error.
1514 * Special pipe_ino values:
1515 * -1 - not yet initialized;
1516 * 0 - alloc_unr failed, return 0 as st_ino forever.
1518 if (pipe->pipe_ino == (ino_t)-1) {
1519 new_unr = alloc_unr(pipeino_unr);
1521 pipe->pipe_ino = new_unr;
1527 bzero(ub, sizeof(*ub));
1528 ub->st_mode = S_IFIFO;
1529 ub->st_blksize = PAGE_SIZE;
1530 if (pipe->pipe_state & PIPE_DIRECTW)
1531 ub->st_size = pipe->pipe_map.cnt;
1533 ub->st_size = pipe->pipe_buffer.cnt;
1534 ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1535 ub->st_atim = pipe->pipe_atime;
1536 ub->st_mtim = pipe->pipe_mtime;
1537 ub->st_ctim = pipe->pipe_ctime;
1538 ub->st_uid = fp->f_cred->cr_uid;
1539 ub->st_gid = fp->f_cred->cr_gid;
1540 ub->st_dev = pipedev_ino;
1541 ub->st_ino = pipe->pipe_ino;
1543 * Left as 0: st_nlink, st_rdev, st_flags, st_gen.
1555 if (fp->f_vnode != NULL)
1556 return vnops.fo_close(fp, td);
1557 fp->f_ops = &badfileops;
1558 pipe_dtor(fp->f_data);
1564 pipe_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td)
1570 if (cpipe->pipe_state & PIPE_NAMED)
1571 error = vn_chmod(fp, mode, active_cred, td);
1573 error = invfo_chmod(fp, mode, active_cred, td);
1578 pipe_chown(fp, uid, gid, active_cred, td)
1582 struct ucred *active_cred;
1589 if (cpipe->pipe_state & PIPE_NAMED)
1590 error = vn_chown(fp, uid, gid, active_cred, td);
1592 error = invfo_chown(fp, uid, gid, active_cred, td);
1597 pipe_free_kmem(cpipe)
1601 KASSERT(!mtx_owned(PIPE_MTX(cpipe)),
1602 ("pipe_free_kmem: pipe mutex locked"));
1604 if (cpipe->pipe_buffer.buffer != NULL) {
1605 atomic_subtract_long(&amountpipekva, cpipe->pipe_buffer.size);
1606 vm_map_remove(pipe_map,
1607 (vm_offset_t)cpipe->pipe_buffer.buffer,
1608 (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size);
1609 cpipe->pipe_buffer.buffer = NULL;
1611 #ifndef PIPE_NODIRECT
1613 cpipe->pipe_map.cnt = 0;
1614 cpipe->pipe_map.pos = 0;
1615 cpipe->pipe_map.npages = 0;
1627 struct pipepair *pp;
1630 KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL"));
1634 pp = cpipe->pipe_pair;
1636 pipeselwakeup(cpipe);
1639 * If the other side is blocked, wake it up saying that
1640 * we want to close it down.
1642 cpipe->pipe_state |= PIPE_EOF;
1643 while (cpipe->pipe_busy) {
1645 cpipe->pipe_state |= PIPE_WANT;
1647 msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0);
1653 * Disconnect from peer, if any.
1655 ppipe = cpipe->pipe_peer;
1656 if (ppipe->pipe_present == PIPE_ACTIVE) {
1657 pipeselwakeup(ppipe);
1659 ppipe->pipe_state |= PIPE_EOF;
1661 KNOTE_LOCKED(&ppipe->pipe_sel.si_note, 0);
1665 * Mark this endpoint as free. Release kmem resources. We
1666 * don't mark this endpoint as unused until we've finished
1667 * doing that, or the pipe might disappear out from under
1671 pipe_free_kmem(cpipe);
1673 cpipe->pipe_present = PIPE_CLOSING;
1677 * knlist_clear() may sleep dropping the PIPE_MTX. Set the
1678 * PIPE_FINALIZED, that allows other end to free the
1679 * pipe_pair, only after the knotes are completely dismantled.
1681 knlist_clear(&cpipe->pipe_sel.si_note, 1);
1682 cpipe->pipe_present = PIPE_FINALIZED;
1683 seldrain(&cpipe->pipe_sel);
1684 knlist_destroy(&cpipe->pipe_sel.si_note);
1687 * If both endpoints are now closed, release the memory for the
1688 * pipe pair. If not, unlock.
1690 if (ppipe->pipe_present == PIPE_FINALIZED) {
1693 mac_pipe_destroy(pp);
1695 uma_zfree(pipe_zone, cpipe->pipe_pair);
1702 pipe_kqfilter(struct file *fp, struct knote *kn)
1707 * If a filter is requested that is not supported by this file
1708 * descriptor, don't return an error, but also don't ever generate an
1711 if ((kn->kn_filter == EVFILT_READ) && !(fp->f_flag & FREAD)) {
1712 kn->kn_fop = &pipe_nfiltops;
1715 if ((kn->kn_filter == EVFILT_WRITE) && !(fp->f_flag & FWRITE)) {
1716 kn->kn_fop = &pipe_nfiltops;
1721 switch (kn->kn_filter) {
1723 kn->kn_fop = &pipe_rfiltops;
1726 kn->kn_fop = &pipe_wfiltops;
1727 if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) {
1728 /* other end of pipe has been closed */
1732 cpipe = PIPE_PEER(cpipe);
1739 kn->kn_hook = cpipe;
1740 knlist_add(&cpipe->pipe_sel.si_note, kn, 1);
1746 filt_pipedetach(struct knote *kn)
1748 struct pipe *cpipe = kn->kn_hook;
1751 knlist_remove(&cpipe->pipe_sel.si_note, kn, 1);
1757 filt_piperead(struct knote *kn, long hint)
1759 struct pipe *rpipe = kn->kn_hook;
1760 struct pipe *wpipe = rpipe->pipe_peer;
1764 kn->kn_data = rpipe->pipe_buffer.cnt;
1765 if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
1766 kn->kn_data = rpipe->pipe_map.cnt;
1768 if ((rpipe->pipe_state & PIPE_EOF) ||
1769 wpipe->pipe_present != PIPE_ACTIVE ||
1770 (wpipe->pipe_state & PIPE_EOF)) {
1771 kn->kn_flags |= EV_EOF;
1775 ret = kn->kn_data > 0;
1782 filt_pipewrite(struct knote *kn, long hint)
1786 wpipe = kn->kn_hook;
1788 if (wpipe->pipe_present != PIPE_ACTIVE ||
1789 (wpipe->pipe_state & PIPE_EOF)) {
1791 kn->kn_flags |= EV_EOF;
1795 kn->kn_data = (wpipe->pipe_buffer.size > 0) ?
1796 (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) : PIPE_BUF;
1797 if (wpipe->pipe_state & PIPE_DIRECTW)
1801 return (kn->kn_data >= PIPE_BUF);
1805 filt_pipedetach_notsup(struct knote *kn)
1811 filt_pipenotsup(struct knote *kn, long hint)