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 int do_pipe(struct thread *td, int fildes[2], int flags);
136 * Use this define if you want to disable *fancy* VM things. Expect an
137 * approx 30% decrease in transfer rate. This could be useful for
140 /* #define PIPE_NODIRECT */
142 #define PIPE_PEER(pipe) \
143 (((pipe)->pipe_state & PIPE_NAMED) ? (pipe) : ((pipe)->pipe_peer))
146 * interfaces to the outside world
148 static fo_rdwr_t pipe_read;
149 static fo_rdwr_t pipe_write;
150 static fo_truncate_t pipe_truncate;
151 static fo_ioctl_t pipe_ioctl;
152 static fo_poll_t pipe_poll;
153 static fo_kqfilter_t pipe_kqfilter;
154 static fo_stat_t pipe_stat;
155 static fo_close_t pipe_close;
156 static fo_chmod_t pipe_chmod;
157 static fo_chown_t pipe_chown;
159 struct fileops pipeops = {
160 .fo_read = pipe_read,
161 .fo_write = pipe_write,
162 .fo_truncate = pipe_truncate,
163 .fo_ioctl = pipe_ioctl,
164 .fo_poll = pipe_poll,
165 .fo_kqfilter = pipe_kqfilter,
166 .fo_stat = pipe_stat,
167 .fo_close = pipe_close,
168 .fo_chmod = pipe_chmod,
169 .fo_chown = pipe_chown,
170 .fo_flags = DFLAG_PASSABLE
173 static void filt_pipedetach(struct knote *kn);
174 static void filt_pipedetach_notsup(struct knote *kn);
175 static int filt_pipenotsup(struct knote *kn, long hint);
176 static int filt_piperead(struct knote *kn, long hint);
177 static int filt_pipewrite(struct knote *kn, long hint);
179 static struct filterops pipe_nfiltops = {
181 .f_detach = filt_pipedetach_notsup,
182 .f_event = filt_pipenotsup
184 static struct filterops pipe_rfiltops = {
186 .f_detach = filt_pipedetach,
187 .f_event = filt_piperead
189 static struct filterops pipe_wfiltops = {
191 .f_detach = filt_pipedetach,
192 .f_event = filt_pipewrite
196 * Default pipe buffer size(s), this can be kind-of large now because pipe
197 * space is pageable. The pipe code will try to maintain locality of
198 * reference for performance reasons, so small amounts of outstanding I/O
199 * will not wipe the cache.
201 #define MINPIPESIZE (PIPE_SIZE/3)
202 #define MAXPIPESIZE (2*PIPE_SIZE/3)
204 static long amountpipekva;
205 static int pipefragretry;
206 static int pipeallocfail;
207 static int piperesizefail;
208 static int piperesizeallowed = 1;
210 SYSCTL_LONG(_kern_ipc, OID_AUTO, maxpipekva, CTLFLAG_RDTUN,
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");
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, int backing);
227 static int pipe_paircreate(struct thread *td, struct pipepair **p_pp);
228 static __inline int pipelock(struct pipe *cpipe, int catch);
229 static __inline void pipeunlock(struct pipe *cpipe);
230 #ifndef PIPE_NODIRECT
231 static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio);
232 static void pipe_destroy_write_buffer(struct pipe *wpipe);
233 static int pipe_direct_write(struct pipe *wpipe, struct uio *uio);
234 static void pipe_clone_write_buffer(struct pipe *wpipe);
236 static int pipespace(struct pipe *cpipe, int size);
237 static int pipespace_new(struct pipe *cpipe, int size);
239 static int pipe_zone_ctor(void *mem, int size, void *arg, int flags);
240 static int pipe_zone_init(void *mem, int size, int flags);
241 static void pipe_zone_fini(void *mem, int size);
243 static uma_zone_t pipe_zone;
244 static struct unrhdr *pipeino_unr;
245 static dev_t pipedev_ino;
247 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL);
250 pipeinit(void *dummy __unused)
253 pipe_zone = uma_zcreate("pipe", sizeof(struct pipepair),
254 pipe_zone_ctor, NULL, pipe_zone_init, pipe_zone_fini,
256 KASSERT(pipe_zone != NULL, ("pipe_zone not initialized"));
257 pipeino_unr = new_unrhdr(1, INT32_MAX, NULL);
258 KASSERT(pipeino_unr != NULL, ("pipe fake inodes not initialized"));
259 pipedev_ino = devfs_alloc_cdp_inode();
260 KASSERT(pipedev_ino > 0, ("pipe dev inode not initialized"));
264 pipe_zone_ctor(void *mem, int size, void *arg, int flags)
267 struct pipe *rpipe, *wpipe;
269 KASSERT(size == sizeof(*pp), ("pipe_zone_ctor: wrong size"));
271 pp = (struct pipepair *)mem;
274 * We zero both pipe endpoints to make sure all the kmem pointers
275 * are NULL, flag fields are zero'd, etc. We timestamp both
276 * endpoints with the same time.
278 rpipe = &pp->pp_rpipe;
279 bzero(rpipe, sizeof(*rpipe));
280 vfs_timestamp(&rpipe->pipe_ctime);
281 rpipe->pipe_atime = rpipe->pipe_mtime = rpipe->pipe_ctime;
283 wpipe = &pp->pp_wpipe;
284 bzero(wpipe, sizeof(*wpipe));
285 wpipe->pipe_ctime = rpipe->pipe_ctime;
286 wpipe->pipe_atime = wpipe->pipe_mtime = rpipe->pipe_ctime;
288 rpipe->pipe_peer = wpipe;
289 rpipe->pipe_pair = pp;
290 wpipe->pipe_peer = rpipe;
291 wpipe->pipe_pair = pp;
294 * Mark both endpoints as present; they will later get free'd
295 * one at a time. When both are free'd, then the whole pair
298 rpipe->pipe_present = PIPE_ACTIVE;
299 wpipe->pipe_present = PIPE_ACTIVE;
302 * Eventually, the MAC Framework may initialize the label
303 * in ctor or init, but for now we do it elswhere to avoid
304 * blocking in ctor or init.
312 pipe_zone_init(void *mem, int size, int flags)
316 KASSERT(size == sizeof(*pp), ("pipe_zone_init: wrong size"));
318 pp = (struct pipepair *)mem;
320 mtx_init(&pp->pp_mtx, "pipe mutex", NULL, MTX_DEF | MTX_RECURSE);
325 pipe_zone_fini(void *mem, int size)
329 KASSERT(size == sizeof(*pp), ("pipe_zone_fini: wrong size"));
331 pp = (struct pipepair *)mem;
333 mtx_destroy(&pp->pp_mtx);
337 pipe_paircreate(struct thread *td, struct pipepair **p_pp)
340 struct pipe *rpipe, *wpipe;
343 *p_pp = pp = uma_zalloc(pipe_zone, M_WAITOK);
346 * The MAC label is shared between the connected endpoints. As a
347 * result mac_pipe_init() and mac_pipe_create() are called once
348 * for the pair, and not on the endpoints.
351 mac_pipe_create(td->td_ucred, pp);
353 rpipe = &pp->pp_rpipe;
354 wpipe = &pp->pp_wpipe;
356 knlist_init_mtx(&rpipe->pipe_sel.si_note, PIPE_MTX(rpipe));
357 knlist_init_mtx(&wpipe->pipe_sel.si_note, PIPE_MTX(wpipe));
359 /* Only the forward direction pipe is backed by default */
360 if ((error = pipe_create(rpipe, 1)) != 0 ||
361 (error = pipe_create(wpipe, 0)) != 0) {
367 rpipe->pipe_state |= PIPE_DIRECTOK;
368 wpipe->pipe_state |= PIPE_DIRECTOK;
373 pipe_named_ctor(struct pipe **ppipe, struct thread *td)
378 error = pipe_paircreate(td, &pp);
381 pp->pp_rpipe.pipe_state |= PIPE_NAMED;
382 *ppipe = &pp->pp_rpipe;
387 pipe_dtor(struct pipe *dpipe)
391 ino = dpipe->pipe_ino;
392 funsetown(&dpipe->pipe_sigio);
394 if (dpipe->pipe_state & PIPE_NAMED) {
395 dpipe = dpipe->pipe_peer;
396 funsetown(&dpipe->pipe_sigio);
399 if (ino != 0 && ino != (ino_t)-1)
400 free_unr(pipeino_unr, ino);
404 * The pipe system call for the DTYPE_PIPE type of pipes. If we fail, let
405 * the zone pick up the pieces via pipeclose().
408 kern_pipe(struct thread *td, int fildes[2])
411 return (do_pipe(td, fildes, 0));
415 do_pipe(struct thread *td, int fildes[2], int flags)
417 struct filedesc *fdp;
418 struct file *rf, *wf;
419 struct pipe *rpipe, *wpipe;
421 int fd, fflags, error;
423 fdp = td->td_proc->p_fd;
424 error = pipe_paircreate(td, &pp);
427 rpipe = &pp->pp_rpipe;
428 wpipe = &pp->pp_wpipe;
429 error = falloc(td, &rf, &fd, flags);
435 /* An extra reference on `rf' has been held for us by falloc(). */
438 fflags = FREAD | FWRITE;
439 if ((flags & O_NONBLOCK) != 0)
443 * Warning: once we've gotten past allocation of the fd for the
444 * read-side, we can only drop the read side via fdrop() in order
445 * to avoid races against processes which manage to dup() the read
446 * side while we are blocked trying to allocate the write side.
448 finit(rf, fflags, DTYPE_PIPE, rpipe, &pipeops);
449 error = falloc(td, &wf, &fd, flags);
451 fdclose(fdp, rf, fildes[0], td);
453 /* rpipe has been closed by fdrop(). */
457 /* An extra reference on `wf' has been held for us by falloc(). */
458 finit(wf, fflags, DTYPE_PIPE, wpipe, &pipeops);
468 sys_pipe(struct thread *td, struct pipe_args *uap)
473 error = kern_pipe(td, fildes);
477 td->td_retval[0] = fildes[0];
478 td->td_retval[1] = fildes[1];
484 * Allocate kva for pipe circular buffer, the space is pageable
485 * This routine will 'realloc' the size of a pipe safely, if it fails
486 * it will retain the old buffer.
487 * If it fails it will return ENOMEM.
490 pipespace_new(cpipe, size)
495 int error, cnt, firstseg;
496 static int curfail = 0;
497 static struct timeval lastfail;
499 KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked"));
500 KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW),
501 ("pipespace: resize of direct writes not allowed"));
503 cnt = cpipe->pipe_buffer.cnt;
507 size = round_page(size);
508 buffer = (caddr_t) vm_map_min(pipe_map);
510 error = vm_map_find(pipe_map, NULL, 0,
511 (vm_offset_t *) &buffer, size, 1,
512 VM_PROT_ALL, VM_PROT_ALL, 0);
513 if (error != KERN_SUCCESS) {
514 if ((cpipe->pipe_buffer.buffer == NULL) &&
515 (size > SMALL_PIPE_SIZE)) {
516 size = SMALL_PIPE_SIZE;
520 if (cpipe->pipe_buffer.buffer == NULL) {
522 if (ppsratecheck(&lastfail, &curfail, 1))
523 printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n");
530 /* copy data, then free old resources if we're resizing */
532 if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) {
533 firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out;
534 bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
536 if ((cnt - firstseg) > 0)
537 bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg],
538 cpipe->pipe_buffer.in);
540 bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
544 pipe_free_kmem(cpipe);
545 cpipe->pipe_buffer.buffer = buffer;
546 cpipe->pipe_buffer.size = size;
547 cpipe->pipe_buffer.in = cnt;
548 cpipe->pipe_buffer.out = 0;
549 cpipe->pipe_buffer.cnt = cnt;
550 atomic_add_long(&amountpipekva, cpipe->pipe_buffer.size);
555 * Wrapper for pipespace_new() that performs locking assertions.
558 pipespace(cpipe, size)
563 KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
564 ("Unlocked pipe passed to pipespace"));
565 return (pipespace_new(cpipe, size));
569 * lock a pipe for I/O, blocking other access
572 pipelock(cpipe, catch)
578 PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
579 while (cpipe->pipe_state & PIPE_LOCKFL) {
580 cpipe->pipe_state |= PIPE_LWANT;
581 error = msleep(cpipe, PIPE_MTX(cpipe),
582 catch ? (PRIBIO | PCATCH) : PRIBIO,
587 cpipe->pipe_state |= PIPE_LOCKFL;
592 * unlock a pipe I/O lock
599 PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
600 KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
601 ("Unlocked pipe passed to pipeunlock"));
602 cpipe->pipe_state &= ~PIPE_LOCKFL;
603 if (cpipe->pipe_state & PIPE_LWANT) {
604 cpipe->pipe_state &= ~PIPE_LWANT;
614 PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
615 if (cpipe->pipe_state & PIPE_SEL) {
616 selwakeuppri(&cpipe->pipe_sel, PSOCK);
617 if (!SEL_WAITING(&cpipe->pipe_sel))
618 cpipe->pipe_state &= ~PIPE_SEL;
620 if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
621 pgsigio(&cpipe->pipe_sigio, SIGIO, 0);
622 KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0);
626 * Initialize and allocate VM and memory for pipe. The structure
627 * will start out zero'd from the ctor, so we just manage the kmem.
630 pipe_create(pipe, backing)
637 if (amountpipekva > maxpipekva / 2)
638 error = pipespace_new(pipe, SMALL_PIPE_SIZE);
640 error = pipespace_new(pipe, PIPE_SIZE);
642 /* If we're not backing this pipe, no need to do anything. */
651 pipe_read(fp, uio, active_cred, flags, td)
654 struct ucred *active_cred;
666 error = pipelock(rpipe, 1);
671 error = mac_pipe_check_read(active_cred, rpipe->pipe_pair);
675 if (amountpipekva > (3 * maxpipekva) / 4) {
676 if (!(rpipe->pipe_state & PIPE_DIRECTW) &&
677 (rpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
678 (rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
679 (piperesizeallowed == 1)) {
681 pipespace(rpipe, SMALL_PIPE_SIZE);
686 while (uio->uio_resid) {
688 * normal pipe buffer receive
690 if (rpipe->pipe_buffer.cnt > 0) {
691 size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
692 if (size > rpipe->pipe_buffer.cnt)
693 size = rpipe->pipe_buffer.cnt;
694 if (size > uio->uio_resid)
695 size = uio->uio_resid;
699 &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
705 rpipe->pipe_buffer.out += size;
706 if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
707 rpipe->pipe_buffer.out = 0;
709 rpipe->pipe_buffer.cnt -= size;
712 * If there is no more to read in the pipe, reset
713 * its pointers to the beginning. This improves
716 if (rpipe->pipe_buffer.cnt == 0) {
717 rpipe->pipe_buffer.in = 0;
718 rpipe->pipe_buffer.out = 0;
721 #ifndef PIPE_NODIRECT
723 * Direct copy, bypassing a kernel buffer.
725 } else if ((size = rpipe->pipe_map.cnt) &&
726 (rpipe->pipe_state & PIPE_DIRECTW)) {
727 if (size > uio->uio_resid)
728 size = (u_int) uio->uio_resid;
731 error = uiomove_fromphys(rpipe->pipe_map.ms,
732 rpipe->pipe_map.pos, size, uio);
737 rpipe->pipe_map.pos += size;
738 rpipe->pipe_map.cnt -= size;
739 if (rpipe->pipe_map.cnt == 0) {
740 rpipe->pipe_state &= ~(PIPE_DIRECTW|PIPE_WANTW);
746 * detect EOF condition
747 * read returns 0 on EOF, no need to set error
749 if (rpipe->pipe_state & PIPE_EOF)
753 * If the "write-side" has been blocked, wake it up now.
755 if (rpipe->pipe_state & PIPE_WANTW) {
756 rpipe->pipe_state &= ~PIPE_WANTW;
761 * Break if some data was read.
767 * Unlock the pipe buffer for our remaining processing.
768 * We will either break out with an error or we will
769 * sleep and relock to loop.
774 * Handle non-blocking mode operation or
775 * wait for more data.
777 if (fp->f_flag & FNONBLOCK) {
780 rpipe->pipe_state |= PIPE_WANTR;
781 if ((error = msleep(rpipe, PIPE_MTX(rpipe),
784 error = pipelock(rpipe, 1);
795 /* XXX: should probably do this before getting any locks. */
797 vfs_timestamp(&rpipe->pipe_atime);
802 * PIPE_WANT processing only makes sense if pipe_busy is 0.
804 if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
805 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
807 } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
809 * Handle write blocking hysteresis.
811 if (rpipe->pipe_state & PIPE_WANTW) {
812 rpipe->pipe_state &= ~PIPE_WANTW;
817 if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
818 pipeselwakeup(rpipe);
824 #ifndef PIPE_NODIRECT
826 * Map the sending processes' buffer into kernel space and wire it.
827 * This is similar to a physical write operation.
830 pipe_build_write_buffer(wpipe, uio)
837 PIPE_LOCK_ASSERT(wpipe, MA_NOTOWNED);
838 KASSERT(wpipe->pipe_state & PIPE_DIRECTW,
839 ("Clone attempt on non-direct write pipe!"));
841 if (uio->uio_iov->iov_len > wpipe->pipe_buffer.size)
842 size = wpipe->pipe_buffer.size;
844 size = uio->uio_iov->iov_len;
846 if ((i = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
847 (vm_offset_t)uio->uio_iov->iov_base, size, VM_PROT_READ,
848 wpipe->pipe_map.ms, PIPENPAGES)) < 0)
852 * set up the control block
854 wpipe->pipe_map.npages = i;
855 wpipe->pipe_map.pos =
856 ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
857 wpipe->pipe_map.cnt = size;
860 * and update the uio data
863 uio->uio_iov->iov_len -= size;
864 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size;
865 if (uio->uio_iov->iov_len == 0)
867 uio->uio_resid -= size;
868 uio->uio_offset += size;
873 * unmap and unwire the process buffer
876 pipe_destroy_write_buffer(wpipe)
880 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
881 vm_page_unhold_pages(wpipe->pipe_map.ms, wpipe->pipe_map.npages);
882 wpipe->pipe_map.npages = 0;
886 * In the case of a signal, the writing process might go away. This
887 * code copies the data into the circular buffer so that the source
888 * pages can be freed without loss of data.
891 pipe_clone_write_buffer(wpipe)
899 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
900 size = wpipe->pipe_map.cnt;
901 pos = wpipe->pipe_map.pos;
903 wpipe->pipe_buffer.in = size;
904 wpipe->pipe_buffer.out = 0;
905 wpipe->pipe_buffer.cnt = size;
906 wpipe->pipe_state &= ~PIPE_DIRECTW;
909 iov.iov_base = wpipe->pipe_buffer.buffer;
914 uio.uio_resid = size;
915 uio.uio_segflg = UIO_SYSSPACE;
916 uio.uio_rw = UIO_READ;
917 uio.uio_td = curthread;
918 uiomove_fromphys(wpipe->pipe_map.ms, pos, size, &uio);
920 pipe_destroy_write_buffer(wpipe);
924 * This implements the pipe buffer write mechanism. Note that only
925 * a direct write OR a normal pipe write can be pending at any given time.
926 * If there are any characters in the pipe buffer, the direct write will
927 * be deferred until the receiving process grabs all of the bytes from
928 * the pipe buffer. Then the direct mapping write is set-up.
931 pipe_direct_write(wpipe, uio)
938 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
939 error = pipelock(wpipe, 1);
940 if (wpipe->pipe_state & PIPE_EOF)
946 while (wpipe->pipe_state & PIPE_DIRECTW) {
947 if (wpipe->pipe_state & PIPE_WANTR) {
948 wpipe->pipe_state &= ~PIPE_WANTR;
951 pipeselwakeup(wpipe);
952 wpipe->pipe_state |= PIPE_WANTW;
954 error = msleep(wpipe, PIPE_MTX(wpipe),
955 PRIBIO | PCATCH, "pipdww", 0);
961 wpipe->pipe_map.cnt = 0; /* transfer not ready yet */
962 if (wpipe->pipe_buffer.cnt > 0) {
963 if (wpipe->pipe_state & PIPE_WANTR) {
964 wpipe->pipe_state &= ~PIPE_WANTR;
967 pipeselwakeup(wpipe);
968 wpipe->pipe_state |= PIPE_WANTW;
970 error = msleep(wpipe, PIPE_MTX(wpipe),
971 PRIBIO | PCATCH, "pipdwc", 0);
978 wpipe->pipe_state |= PIPE_DIRECTW;
981 error = pipe_build_write_buffer(wpipe, uio);
984 wpipe->pipe_state &= ~PIPE_DIRECTW;
990 while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
991 if (wpipe->pipe_state & PIPE_EOF) {
992 pipe_destroy_write_buffer(wpipe);
993 pipeselwakeup(wpipe);
998 if (wpipe->pipe_state & PIPE_WANTR) {
999 wpipe->pipe_state &= ~PIPE_WANTR;
1002 pipeselwakeup(wpipe);
1003 wpipe->pipe_state |= PIPE_WANTW;
1005 error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH,
1010 if (wpipe->pipe_state & PIPE_EOF)
1012 if (wpipe->pipe_state & PIPE_DIRECTW) {
1014 * this bit of trickery substitutes a kernel buffer for
1015 * the process that might be going away.
1017 pipe_clone_write_buffer(wpipe);
1019 pipe_destroy_write_buffer(wpipe);
1031 pipe_write(fp, uio, active_cred, flags, td)
1034 struct ucred *active_cred;
1041 struct pipe *wpipe, *rpipe;
1044 wpipe = PIPE_PEER(rpipe);
1046 error = pipelock(wpipe, 1);
1052 * detect loss of pipe read side, issue SIGPIPE if lost.
1054 if (wpipe->pipe_present != PIPE_ACTIVE ||
1055 (wpipe->pipe_state & PIPE_EOF)) {
1061 error = mac_pipe_check_write(active_cred, wpipe->pipe_pair);
1070 /* Choose a larger size if it's advantageous */
1071 desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size);
1072 while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) {
1073 if (piperesizeallowed != 1)
1075 if (amountpipekva > maxpipekva / 2)
1077 if (desiredsize == BIG_PIPE_SIZE)
1079 desiredsize = desiredsize * 2;
1082 /* Choose a smaller size if we're in a OOM situation */
1083 if ((amountpipekva > (3 * maxpipekva) / 4) &&
1084 (wpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
1085 (wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
1086 (piperesizeallowed == 1))
1087 desiredsize = SMALL_PIPE_SIZE;
1089 /* Resize if the above determined that a new size was necessary */
1090 if ((desiredsize != wpipe->pipe_buffer.size) &&
1091 ((wpipe->pipe_state & PIPE_DIRECTW) == 0)) {
1093 pipespace(wpipe, desiredsize);
1096 if (wpipe->pipe_buffer.size == 0) {
1098 * This can only happen for reverse direction use of pipes
1099 * in a complete OOM situation.
1110 orig_resid = uio->uio_resid;
1112 while (uio->uio_resid) {
1116 if (wpipe->pipe_state & PIPE_EOF) {
1121 #ifndef PIPE_NODIRECT
1123 * If the transfer is large, we can gain performance if
1124 * we do process-to-process copies directly.
1125 * If the write is non-blocking, we don't use the
1126 * direct write mechanism.
1128 * The direct write mechanism will detect the reader going
1131 if (uio->uio_segflg == UIO_USERSPACE &&
1132 uio->uio_iov->iov_len >= PIPE_MINDIRECT &&
1133 wpipe->pipe_buffer.size >= PIPE_MINDIRECT &&
1134 (fp->f_flag & FNONBLOCK) == 0) {
1136 error = pipe_direct_write(wpipe, uio);
1144 * Pipe buffered writes cannot be coincidental with
1145 * direct writes. We wait until the currently executing
1146 * direct write is completed before we start filling the
1147 * pipe buffer. We break out if a signal occurs or the
1150 if (wpipe->pipe_state & PIPE_DIRECTW) {
1151 if (wpipe->pipe_state & PIPE_WANTR) {
1152 wpipe->pipe_state &= ~PIPE_WANTR;
1155 pipeselwakeup(wpipe);
1156 wpipe->pipe_state |= PIPE_WANTW;
1158 error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH,
1166 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1168 /* Writes of size <= PIPE_BUF must be atomic. */
1169 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
1173 int size; /* Transfer size */
1174 int segsize; /* first segment to transfer */
1177 * Transfer size is minimum of uio transfer
1178 * and free space in pipe buffer.
1180 if (space > uio->uio_resid)
1181 size = uio->uio_resid;
1185 * First segment to transfer is minimum of
1186 * transfer size and contiguous space in
1187 * pipe buffer. If first segment to transfer
1188 * is less than the transfer size, we've got
1189 * a wraparound in the buffer.
1191 segsize = wpipe->pipe_buffer.size -
1192 wpipe->pipe_buffer.in;
1196 /* Transfer first segment */
1199 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1203 if (error == 0 && segsize < size) {
1204 KASSERT(wpipe->pipe_buffer.in + segsize ==
1205 wpipe->pipe_buffer.size,
1206 ("Pipe buffer wraparound disappeared"));
1208 * Transfer remaining part now, to
1209 * support atomic writes. Wraparound
1215 &wpipe->pipe_buffer.buffer[0],
1216 size - segsize, uio);
1220 wpipe->pipe_buffer.in += size;
1221 if (wpipe->pipe_buffer.in >=
1222 wpipe->pipe_buffer.size) {
1223 KASSERT(wpipe->pipe_buffer.in ==
1225 wpipe->pipe_buffer.size,
1226 ("Expected wraparound bad"));
1227 wpipe->pipe_buffer.in = size - segsize;
1230 wpipe->pipe_buffer.cnt += size;
1231 KASSERT(wpipe->pipe_buffer.cnt <=
1232 wpipe->pipe_buffer.size,
1233 ("Pipe buffer overflow"));
1240 * If the "read-side" has been blocked, wake it up now.
1242 if (wpipe->pipe_state & PIPE_WANTR) {
1243 wpipe->pipe_state &= ~PIPE_WANTR;
1248 * don't block on non-blocking I/O
1250 if (fp->f_flag & FNONBLOCK) {
1257 * We have no more space and have something to offer,
1258 * wake up select/poll.
1260 pipeselwakeup(wpipe);
1262 wpipe->pipe_state |= PIPE_WANTW;
1264 error = msleep(wpipe, PIPE_MTX(rpipe),
1265 PRIBIO | PCATCH, "pipewr", 0);
1274 if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1275 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1277 } else if (wpipe->pipe_buffer.cnt > 0) {
1279 * If we have put any characters in the buffer, we wake up
1282 if (wpipe->pipe_state & PIPE_WANTR) {
1283 wpipe->pipe_state &= ~PIPE_WANTR;
1289 * Don't return EPIPE if I/O was successful
1291 if ((wpipe->pipe_buffer.cnt == 0) &&
1292 (uio->uio_resid == 0) &&
1298 vfs_timestamp(&wpipe->pipe_mtime);
1301 * We have something to offer,
1302 * wake up select/poll.
1304 if (wpipe->pipe_buffer.cnt)
1305 pipeselwakeup(wpipe);
1314 pipe_truncate(fp, length, active_cred, td)
1317 struct ucred *active_cred;
1321 /* For named pipes call the vnode operation. */
1322 if (fp->f_vnode != NULL)
1323 return (vnops.fo_truncate(fp, length, active_cred, td));
1328 * we implement a very minimal set of ioctls for compatibility with sockets.
1331 pipe_ioctl(fp, cmd, data, active_cred, td)
1335 struct ucred *active_cred;
1338 struct pipe *mpipe = fp->f_data;
1344 error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data);
1359 mpipe->pipe_state |= PIPE_ASYNC;
1361 mpipe->pipe_state &= ~PIPE_ASYNC;
1366 if (!(fp->f_flag & FREAD)) {
1371 if (mpipe->pipe_state & PIPE_DIRECTW)
1372 *(int *)data = mpipe->pipe_map.cnt;
1374 *(int *)data = mpipe->pipe_buffer.cnt;
1379 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1383 *(int *)data = fgetown(&mpipe->pipe_sigio);
1386 /* This is deprecated, FIOSETOWN should be used instead. */
1389 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1392 /* This is deprecated, FIOGETOWN should be used instead. */
1394 *(int *)data = -fgetown(&mpipe->pipe_sigio);
1407 pipe_poll(fp, events, active_cred, td)
1410 struct ucred *active_cred;
1415 int levents, revents;
1422 wpipe = PIPE_PEER(rpipe);
1425 error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair);
1429 if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM))
1430 if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1431 (rpipe->pipe_buffer.cnt > 0))
1432 revents |= events & (POLLIN | POLLRDNORM);
1434 if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM))
1435 if (wpipe->pipe_present != PIPE_ACTIVE ||
1436 (wpipe->pipe_state & PIPE_EOF) ||
1437 (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1438 ((wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF ||
1439 wpipe->pipe_buffer.size == 0)))
1440 revents |= events & (POLLOUT | POLLWRNORM);
1443 (POLLIN | POLLINIGNEOF | POLLPRI | POLLRDNORM | POLLRDBAND);
1444 if (rpipe->pipe_state & PIPE_NAMED && fp->f_flag & FREAD && levents &&
1445 fp->f_seqcount == rpipe->pipe_wgen)
1446 events |= POLLINIGNEOF;
1448 if ((events & POLLINIGNEOF) == 0) {
1449 if (rpipe->pipe_state & PIPE_EOF) {
1450 revents |= (events & (POLLIN | POLLRDNORM));
1451 if (wpipe->pipe_present != PIPE_ACTIVE ||
1452 (wpipe->pipe_state & PIPE_EOF))
1458 if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM)) {
1459 selrecord(td, &rpipe->pipe_sel);
1460 if (SEL_WAITING(&rpipe->pipe_sel))
1461 rpipe->pipe_state |= PIPE_SEL;
1464 if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM)) {
1465 selrecord(td, &wpipe->pipe_sel);
1466 if (SEL_WAITING(&wpipe->pipe_sel))
1467 wpipe->pipe_state |= PIPE_SEL;
1479 * We shouldn't need locks here as we're doing a read and this should
1480 * be a natural race.
1483 pipe_stat(fp, ub, active_cred, td)
1486 struct ucred *active_cred;
1498 error = mac_pipe_check_stat(active_cred, pipe->pipe_pair);
1505 /* For named pipes ask the underlying filesystem. */
1506 if (pipe->pipe_state & PIPE_NAMED) {
1508 return (vnops.fo_stat(fp, ub, active_cred, td));
1512 * Lazily allocate an inode number for the pipe. Most pipe
1513 * users do not call fstat(2) on the pipe, which means that
1514 * postponing the inode allocation until it is must be
1515 * returned to userland is useful. If alloc_unr failed,
1516 * assign st_ino zero instead of returning an error.
1517 * Special pipe_ino values:
1518 * -1 - not yet initialized;
1519 * 0 - alloc_unr failed, return 0 as st_ino forever.
1521 if (pipe->pipe_ino == (ino_t)-1) {
1522 new_unr = alloc_unr(pipeino_unr);
1524 pipe->pipe_ino = new_unr;
1530 bzero(ub, sizeof(*ub));
1531 ub->st_mode = S_IFIFO;
1532 ub->st_blksize = PAGE_SIZE;
1533 if (pipe->pipe_state & PIPE_DIRECTW)
1534 ub->st_size = pipe->pipe_map.cnt;
1536 ub->st_size = pipe->pipe_buffer.cnt;
1537 ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1538 ub->st_atim = pipe->pipe_atime;
1539 ub->st_mtim = pipe->pipe_mtime;
1540 ub->st_ctim = pipe->pipe_ctime;
1541 ub->st_uid = fp->f_cred->cr_uid;
1542 ub->st_gid = fp->f_cred->cr_gid;
1543 ub->st_dev = pipedev_ino;
1544 ub->st_ino = pipe->pipe_ino;
1546 * Left as 0: st_nlink, st_rdev, st_flags, st_gen.
1558 if (fp->f_vnode != NULL)
1559 return vnops.fo_close(fp, td);
1560 fp->f_ops = &badfileops;
1561 pipe_dtor(fp->f_data);
1567 pipe_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td)
1573 if (cpipe->pipe_state & PIPE_NAMED)
1574 error = vn_chmod(fp, mode, active_cred, td);
1576 error = invfo_chmod(fp, mode, active_cred, td);
1581 pipe_chown(fp, uid, gid, active_cred, td)
1585 struct ucred *active_cred;
1592 if (cpipe->pipe_state & PIPE_NAMED)
1593 error = vn_chown(fp, uid, gid, active_cred, td);
1595 error = invfo_chown(fp, uid, gid, active_cred, td);
1600 pipe_free_kmem(cpipe)
1604 KASSERT(!mtx_owned(PIPE_MTX(cpipe)),
1605 ("pipe_free_kmem: pipe mutex locked"));
1607 if (cpipe->pipe_buffer.buffer != NULL) {
1608 atomic_subtract_long(&amountpipekva, cpipe->pipe_buffer.size);
1609 vm_map_remove(pipe_map,
1610 (vm_offset_t)cpipe->pipe_buffer.buffer,
1611 (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size);
1612 cpipe->pipe_buffer.buffer = NULL;
1614 #ifndef PIPE_NODIRECT
1616 cpipe->pipe_map.cnt = 0;
1617 cpipe->pipe_map.pos = 0;
1618 cpipe->pipe_map.npages = 0;
1630 struct pipepair *pp;
1633 KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL"));
1637 pp = cpipe->pipe_pair;
1639 pipeselwakeup(cpipe);
1642 * If the other side is blocked, wake it up saying that
1643 * we want to close it down.
1645 cpipe->pipe_state |= PIPE_EOF;
1646 while (cpipe->pipe_busy) {
1648 cpipe->pipe_state |= PIPE_WANT;
1650 msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0);
1656 * Disconnect from peer, if any.
1658 ppipe = cpipe->pipe_peer;
1659 if (ppipe->pipe_present == PIPE_ACTIVE) {
1660 pipeselwakeup(ppipe);
1662 ppipe->pipe_state |= PIPE_EOF;
1664 KNOTE_LOCKED(&ppipe->pipe_sel.si_note, 0);
1668 * Mark this endpoint as free. Release kmem resources. We
1669 * don't mark this endpoint as unused until we've finished
1670 * doing that, or the pipe might disappear out from under
1674 pipe_free_kmem(cpipe);
1676 cpipe->pipe_present = PIPE_CLOSING;
1680 * knlist_clear() may sleep dropping the PIPE_MTX. Set the
1681 * PIPE_FINALIZED, that allows other end to free the
1682 * pipe_pair, only after the knotes are completely dismantled.
1684 knlist_clear(&cpipe->pipe_sel.si_note, 1);
1685 cpipe->pipe_present = PIPE_FINALIZED;
1686 seldrain(&cpipe->pipe_sel);
1687 knlist_destroy(&cpipe->pipe_sel.si_note);
1690 * If both endpoints are now closed, release the memory for the
1691 * pipe pair. If not, unlock.
1693 if (ppipe->pipe_present == PIPE_FINALIZED) {
1696 mac_pipe_destroy(pp);
1698 uma_zfree(pipe_zone, cpipe->pipe_pair);
1705 pipe_kqfilter(struct file *fp, struct knote *kn)
1710 * If a filter is requested that is not supported by this file
1711 * descriptor, don't return an error, but also don't ever generate an
1714 if ((kn->kn_filter == EVFILT_READ) && !(fp->f_flag & FREAD)) {
1715 kn->kn_fop = &pipe_nfiltops;
1718 if ((kn->kn_filter == EVFILT_WRITE) && !(fp->f_flag & FWRITE)) {
1719 kn->kn_fop = &pipe_nfiltops;
1724 switch (kn->kn_filter) {
1726 kn->kn_fop = &pipe_rfiltops;
1729 kn->kn_fop = &pipe_wfiltops;
1730 if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) {
1731 /* other end of pipe has been closed */
1735 cpipe = PIPE_PEER(cpipe);
1742 kn->kn_hook = cpipe;
1743 knlist_add(&cpipe->pipe_sel.si_note, kn, 1);
1749 filt_pipedetach(struct knote *kn)
1751 struct pipe *cpipe = kn->kn_hook;
1754 knlist_remove(&cpipe->pipe_sel.si_note, kn, 1);
1760 filt_piperead(struct knote *kn, long hint)
1762 struct pipe *rpipe = kn->kn_hook;
1763 struct pipe *wpipe = rpipe->pipe_peer;
1767 kn->kn_data = rpipe->pipe_buffer.cnt;
1768 if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
1769 kn->kn_data = rpipe->pipe_map.cnt;
1771 if ((rpipe->pipe_state & PIPE_EOF) ||
1772 wpipe->pipe_present != PIPE_ACTIVE ||
1773 (wpipe->pipe_state & PIPE_EOF)) {
1774 kn->kn_flags |= EV_EOF;
1778 ret = kn->kn_data > 0;
1785 filt_pipewrite(struct knote *kn, long hint)
1789 wpipe = kn->kn_hook;
1791 if (wpipe->pipe_present != PIPE_ACTIVE ||
1792 (wpipe->pipe_state & PIPE_EOF)) {
1794 kn->kn_flags |= EV_EOF;
1798 kn->kn_data = (wpipe->pipe_buffer.size > 0) ?
1799 (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) : PIPE_BUF;
1800 if (wpipe->pipe_state & PIPE_DIRECTW)
1804 return (kn->kn_data >= PIPE_BUF);
1808 filt_pipedetach_notsup(struct knote *kn)
1814 filt_pipenotsup(struct knote *kn, long hint)