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_sendfile = invfo_sendfile,
168 .fo_flags = DFLAG_PASSABLE
171 static void filt_pipedetach(struct knote *kn);
172 static void filt_pipedetach_notsup(struct knote *kn);
173 static int filt_pipenotsup(struct knote *kn, long hint);
174 static int filt_piperead(struct knote *kn, long hint);
175 static int filt_pipewrite(struct knote *kn, long hint);
177 static struct filterops pipe_nfiltops = {
179 .f_detach = filt_pipedetach_notsup,
180 .f_event = filt_pipenotsup
182 static struct filterops pipe_rfiltops = {
184 .f_detach = filt_pipedetach,
185 .f_event = filt_piperead
187 static struct filterops pipe_wfiltops = {
189 .f_detach = filt_pipedetach,
190 .f_event = filt_pipewrite
194 * Default pipe buffer size(s), this can be kind-of large now because pipe
195 * space is pageable. The pipe code will try to maintain locality of
196 * reference for performance reasons, so small amounts of outstanding I/O
197 * will not wipe the cache.
199 #define MINPIPESIZE (PIPE_SIZE/3)
200 #define MAXPIPESIZE (2*PIPE_SIZE/3)
202 static long amountpipekva;
203 static int pipefragretry;
204 static int pipeallocfail;
205 static int piperesizefail;
206 static int piperesizeallowed = 1;
208 SYSCTL_LONG(_kern_ipc, OID_AUTO, maxpipekva, CTLFLAG_RDTUN,
209 &maxpipekva, 0, "Pipe KVA limit");
210 SYSCTL_LONG(_kern_ipc, OID_AUTO, pipekva, CTLFLAG_RD,
211 &amountpipekva, 0, "Pipe KVA usage");
212 SYSCTL_INT(_kern_ipc, OID_AUTO, pipefragretry, CTLFLAG_RD,
213 &pipefragretry, 0, "Pipe allocation retries due to fragmentation");
214 SYSCTL_INT(_kern_ipc, OID_AUTO, pipeallocfail, CTLFLAG_RD,
215 &pipeallocfail, 0, "Pipe allocation failures");
216 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizefail, CTLFLAG_RD,
217 &piperesizefail, 0, "Pipe resize failures");
218 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizeallowed, CTLFLAG_RW,
219 &piperesizeallowed, 0, "Pipe resizing allowed");
221 static void pipeinit(void *dummy __unused);
222 static void pipeclose(struct pipe *cpipe);
223 static void pipe_free_kmem(struct pipe *cpipe);
224 static void pipe_create(struct pipe *pipe, int backing);
225 static void pipe_paircreate(struct thread *td, struct pipepair **p_pp);
226 static __inline int pipelock(struct pipe *cpipe, int catch);
227 static __inline void pipeunlock(struct pipe *cpipe);
228 #ifndef PIPE_NODIRECT
229 static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio);
230 static void pipe_destroy_write_buffer(struct pipe *wpipe);
231 static int pipe_direct_write(struct pipe *wpipe, struct uio *uio);
232 static void pipe_clone_write_buffer(struct pipe *wpipe);
234 static int pipespace(struct pipe *cpipe, int size);
235 static int pipespace_new(struct pipe *cpipe, int size);
237 static int pipe_zone_ctor(void *mem, int size, void *arg, int flags);
238 static int pipe_zone_init(void *mem, int size, int flags);
239 static void pipe_zone_fini(void *mem, int size);
241 static uma_zone_t pipe_zone;
242 static struct unrhdr *pipeino_unr;
243 static dev_t pipedev_ino;
245 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL);
248 pipeinit(void *dummy __unused)
251 pipe_zone = uma_zcreate("pipe", sizeof(struct pipepair),
252 pipe_zone_ctor, NULL, pipe_zone_init, pipe_zone_fini,
254 KASSERT(pipe_zone != NULL, ("pipe_zone not initialized"));
255 pipeino_unr = new_unrhdr(1, INT32_MAX, NULL);
256 KASSERT(pipeino_unr != NULL, ("pipe fake inodes not initialized"));
257 pipedev_ino = devfs_alloc_cdp_inode();
258 KASSERT(pipedev_ino > 0, ("pipe dev inode not initialized"));
262 pipe_zone_ctor(void *mem, int size, void *arg, int flags)
265 struct pipe *rpipe, *wpipe;
267 KASSERT(size == sizeof(*pp), ("pipe_zone_ctor: wrong size"));
269 pp = (struct pipepair *)mem;
272 * We zero both pipe endpoints to make sure all the kmem pointers
273 * are NULL, flag fields are zero'd, etc. We timestamp both
274 * endpoints with the same time.
276 rpipe = &pp->pp_rpipe;
277 bzero(rpipe, sizeof(*rpipe));
278 vfs_timestamp(&rpipe->pipe_ctime);
279 rpipe->pipe_atime = rpipe->pipe_mtime = rpipe->pipe_ctime;
281 wpipe = &pp->pp_wpipe;
282 bzero(wpipe, sizeof(*wpipe));
283 wpipe->pipe_ctime = rpipe->pipe_ctime;
284 wpipe->pipe_atime = wpipe->pipe_mtime = rpipe->pipe_ctime;
286 rpipe->pipe_peer = wpipe;
287 rpipe->pipe_pair = pp;
288 wpipe->pipe_peer = rpipe;
289 wpipe->pipe_pair = pp;
292 * Mark both endpoints as present; they will later get free'd
293 * one at a time. When both are free'd, then the whole pair
296 rpipe->pipe_present = PIPE_ACTIVE;
297 wpipe->pipe_present = PIPE_ACTIVE;
300 * Eventually, the MAC Framework may initialize the label
301 * in ctor or init, but for now we do it elswhere to avoid
302 * blocking in ctor or init.
310 pipe_zone_init(void *mem, int size, int flags)
314 KASSERT(size == sizeof(*pp), ("pipe_zone_init: wrong size"));
316 pp = (struct pipepair *)mem;
318 mtx_init(&pp->pp_mtx, "pipe mutex", NULL, MTX_DEF);
323 pipe_zone_fini(void *mem, int size)
327 KASSERT(size == sizeof(*pp), ("pipe_zone_fini: wrong size"));
329 pp = (struct pipepair *)mem;
331 mtx_destroy(&pp->pp_mtx);
335 pipe_paircreate(struct thread *td, struct pipepair **p_pp)
338 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 pipe_create(rpipe, 1);
358 pipe_create(wpipe, 0);
360 rpipe->pipe_state |= PIPE_DIRECTOK;
361 wpipe->pipe_state |= PIPE_DIRECTOK;
365 pipe_named_ctor(struct pipe **ppipe, struct thread *td)
369 pipe_paircreate(td, &pp);
370 pp->pp_rpipe.pipe_state |= PIPE_NAMED;
371 *ppipe = &pp->pp_rpipe;
375 pipe_dtor(struct pipe *dpipe)
380 ino = dpipe->pipe_ino;
381 peer = (dpipe->pipe_state & PIPE_NAMED) != 0 ? dpipe->pipe_peer : NULL;
382 funsetown(&dpipe->pipe_sigio);
385 funsetown(&peer->pipe_sigio);
388 if (ino != 0 && ino != (ino_t)-1)
389 free_unr(pipeino_unr, ino);
393 * The pipe system call for the DTYPE_PIPE type of pipes. If we fail, let
394 * the zone pick up the pieces via pipeclose().
397 kern_pipe(struct thread *td, int fildes[2])
400 return (kern_pipe2(td, fildes, 0));
404 kern_pipe2(struct thread *td, int fildes[2], int flags)
406 struct filedesc *fdp;
407 struct file *rf, *wf;
408 struct pipe *rpipe, *wpipe;
410 int fd, fflags, error;
412 fdp = td->td_proc->p_fd;
413 pipe_paircreate(td, &pp);
414 rpipe = &pp->pp_rpipe;
415 wpipe = &pp->pp_wpipe;
416 error = falloc(td, &rf, &fd, flags);
422 /* An extra reference on `rf' has been held for us by falloc(). */
425 fflags = FREAD | FWRITE;
426 if ((flags & O_NONBLOCK) != 0)
430 * Warning: once we've gotten past allocation of the fd for the
431 * read-side, we can only drop the read side via fdrop() in order
432 * to avoid races against processes which manage to dup() the read
433 * side while we are blocked trying to allocate the write side.
435 finit(rf, fflags, DTYPE_PIPE, rpipe, &pipeops);
436 error = falloc(td, &wf, &fd, flags);
438 fdclose(fdp, rf, fildes[0], td);
440 /* rpipe has been closed by fdrop(). */
444 /* An extra reference on `wf' has been held for us by falloc(). */
445 finit(wf, fflags, DTYPE_PIPE, wpipe, &pipeops);
455 sys_pipe(struct thread *td, struct pipe_args *uap)
460 error = kern_pipe(td, fildes);
464 td->td_retval[0] = fildes[0];
465 td->td_retval[1] = fildes[1];
471 sys_pipe2(struct thread *td, struct pipe2_args *uap)
473 int error, fildes[2];
475 if (uap->flags & ~(O_CLOEXEC | O_NONBLOCK))
477 error = kern_pipe2(td, fildes, uap->flags);
480 error = copyout(fildes, uap->fildes, 2 * sizeof(int));
482 (void)kern_close(td, fildes[0]);
483 (void)kern_close(td, fildes[1]);
489 * Allocate kva for pipe circular buffer, the space is pageable
490 * This routine will 'realloc' the size of a pipe safely, if it fails
491 * it will retain the old buffer.
492 * If it fails it will return ENOMEM.
495 pipespace_new(cpipe, size)
500 int error, cnt, firstseg;
501 static int curfail = 0;
502 static struct timeval lastfail;
504 KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked"));
505 KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW),
506 ("pipespace: resize of direct writes not allowed"));
508 cnt = cpipe->pipe_buffer.cnt;
512 size = round_page(size);
513 buffer = (caddr_t) vm_map_min(pipe_map);
515 error = vm_map_find(pipe_map, NULL, 0,
516 (vm_offset_t *) &buffer, size, 0, VMFS_ANY_SPACE,
517 VM_PROT_ALL, VM_PROT_ALL, 0);
518 if (error != KERN_SUCCESS) {
519 if ((cpipe->pipe_buffer.buffer == NULL) &&
520 (size > SMALL_PIPE_SIZE)) {
521 size = SMALL_PIPE_SIZE;
525 if (cpipe->pipe_buffer.buffer == NULL) {
527 if (ppsratecheck(&lastfail, &curfail, 1))
528 printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n");
535 /* copy data, then free old resources if we're resizing */
537 if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) {
538 firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out;
539 bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
541 if ((cnt - firstseg) > 0)
542 bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg],
543 cpipe->pipe_buffer.in);
545 bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
549 pipe_free_kmem(cpipe);
550 cpipe->pipe_buffer.buffer = buffer;
551 cpipe->pipe_buffer.size = size;
552 cpipe->pipe_buffer.in = cnt;
553 cpipe->pipe_buffer.out = 0;
554 cpipe->pipe_buffer.cnt = cnt;
555 atomic_add_long(&amountpipekva, cpipe->pipe_buffer.size);
560 * Wrapper for pipespace_new() that performs locking assertions.
563 pipespace(cpipe, size)
568 KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
569 ("Unlocked pipe passed to pipespace"));
570 return (pipespace_new(cpipe, size));
574 * lock a pipe for I/O, blocking other access
577 pipelock(cpipe, catch)
583 PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
584 while (cpipe->pipe_state & PIPE_LOCKFL) {
585 cpipe->pipe_state |= PIPE_LWANT;
586 error = msleep(cpipe, PIPE_MTX(cpipe),
587 catch ? (PRIBIO | PCATCH) : PRIBIO,
592 cpipe->pipe_state |= PIPE_LOCKFL;
597 * unlock a pipe I/O lock
604 PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
605 KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
606 ("Unlocked pipe passed to pipeunlock"));
607 cpipe->pipe_state &= ~PIPE_LOCKFL;
608 if (cpipe->pipe_state & PIPE_LWANT) {
609 cpipe->pipe_state &= ~PIPE_LWANT;
619 PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
620 if (cpipe->pipe_state & PIPE_SEL) {
621 selwakeuppri(&cpipe->pipe_sel, PSOCK);
622 if (!SEL_WAITING(&cpipe->pipe_sel))
623 cpipe->pipe_state &= ~PIPE_SEL;
625 if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
626 pgsigio(&cpipe->pipe_sigio, SIGIO, 0);
627 KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0);
631 * Initialize and allocate VM and memory for pipe. The structure
632 * will start out zero'd from the ctor, so we just manage the kmem.
635 pipe_create(pipe, backing)
642 * Note that these functions can fail if pipe map is exhausted
643 * (as a result of too many pipes created), but we ignore the
644 * error as it is not fatal and could be provoked by
645 * unprivileged users. The only consequence is worse performance
648 if (amountpipekva > maxpipekva / 2)
649 (void)pipespace_new(pipe, SMALL_PIPE_SIZE);
651 (void)pipespace_new(pipe, PIPE_SIZE);
659 pipe_read(fp, uio, active_cred, flags, td)
662 struct ucred *active_cred;
674 error = pipelock(rpipe, 1);
679 error = mac_pipe_check_read(active_cred, rpipe->pipe_pair);
683 if (amountpipekva > (3 * maxpipekva) / 4) {
684 if (!(rpipe->pipe_state & PIPE_DIRECTW) &&
685 (rpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
686 (rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
687 (piperesizeallowed == 1)) {
689 pipespace(rpipe, SMALL_PIPE_SIZE);
694 while (uio->uio_resid) {
696 * normal pipe buffer receive
698 if (rpipe->pipe_buffer.cnt > 0) {
699 size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
700 if (size > rpipe->pipe_buffer.cnt)
701 size = rpipe->pipe_buffer.cnt;
702 if (size > uio->uio_resid)
703 size = uio->uio_resid;
707 &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
713 rpipe->pipe_buffer.out += size;
714 if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
715 rpipe->pipe_buffer.out = 0;
717 rpipe->pipe_buffer.cnt -= size;
720 * If there is no more to read in the pipe, reset
721 * its pointers to the beginning. This improves
724 if (rpipe->pipe_buffer.cnt == 0) {
725 rpipe->pipe_buffer.in = 0;
726 rpipe->pipe_buffer.out = 0;
729 #ifndef PIPE_NODIRECT
731 * Direct copy, bypassing a kernel buffer.
733 } else if ((size = rpipe->pipe_map.cnt) &&
734 (rpipe->pipe_state & PIPE_DIRECTW)) {
735 if (size > uio->uio_resid)
736 size = (u_int) uio->uio_resid;
739 error = uiomove_fromphys(rpipe->pipe_map.ms,
740 rpipe->pipe_map.pos, size, uio);
745 rpipe->pipe_map.pos += size;
746 rpipe->pipe_map.cnt -= size;
747 if (rpipe->pipe_map.cnt == 0) {
748 rpipe->pipe_state &= ~(PIPE_DIRECTW|PIPE_WANTW);
754 * detect EOF condition
755 * read returns 0 on EOF, no need to set error
757 if (rpipe->pipe_state & PIPE_EOF)
761 * If the "write-side" has been blocked, wake it up now.
763 if (rpipe->pipe_state & PIPE_WANTW) {
764 rpipe->pipe_state &= ~PIPE_WANTW;
769 * Break if some data was read.
775 * Unlock the pipe buffer for our remaining processing.
776 * We will either break out with an error or we will
777 * sleep and relock to loop.
782 * Handle non-blocking mode operation or
783 * wait for more data.
785 if (fp->f_flag & FNONBLOCK) {
788 rpipe->pipe_state |= PIPE_WANTR;
789 if ((error = msleep(rpipe, PIPE_MTX(rpipe),
792 error = pipelock(rpipe, 1);
803 /* XXX: should probably do this before getting any locks. */
805 vfs_timestamp(&rpipe->pipe_atime);
810 * PIPE_WANT processing only makes sense if pipe_busy is 0.
812 if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
813 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
815 } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
817 * Handle write blocking hysteresis.
819 if (rpipe->pipe_state & PIPE_WANTW) {
820 rpipe->pipe_state &= ~PIPE_WANTW;
825 if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
826 pipeselwakeup(rpipe);
832 #ifndef PIPE_NODIRECT
834 * Map the sending processes' buffer into kernel space and wire it.
835 * This is similar to a physical write operation.
838 pipe_build_write_buffer(wpipe, uio)
845 PIPE_LOCK_ASSERT(wpipe, MA_NOTOWNED);
846 KASSERT(wpipe->pipe_state & PIPE_DIRECTW,
847 ("Clone attempt on non-direct write pipe!"));
849 if (uio->uio_iov->iov_len > wpipe->pipe_buffer.size)
850 size = wpipe->pipe_buffer.size;
852 size = uio->uio_iov->iov_len;
854 if ((i = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
855 (vm_offset_t)uio->uio_iov->iov_base, size, VM_PROT_READ,
856 wpipe->pipe_map.ms, PIPENPAGES)) < 0)
860 * set up the control block
862 wpipe->pipe_map.npages = i;
863 wpipe->pipe_map.pos =
864 ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
865 wpipe->pipe_map.cnt = size;
868 * and update the uio data
871 uio->uio_iov->iov_len -= size;
872 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size;
873 if (uio->uio_iov->iov_len == 0)
875 uio->uio_resid -= size;
876 uio->uio_offset += size;
881 * unmap and unwire the process buffer
884 pipe_destroy_write_buffer(wpipe)
888 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
889 vm_page_unhold_pages(wpipe->pipe_map.ms, wpipe->pipe_map.npages);
890 wpipe->pipe_map.npages = 0;
894 * In the case of a signal, the writing process might go away. This
895 * code copies the data into the circular buffer so that the source
896 * pages can be freed without loss of data.
899 pipe_clone_write_buffer(wpipe)
907 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
908 size = wpipe->pipe_map.cnt;
909 pos = wpipe->pipe_map.pos;
911 wpipe->pipe_buffer.in = size;
912 wpipe->pipe_buffer.out = 0;
913 wpipe->pipe_buffer.cnt = size;
914 wpipe->pipe_state &= ~PIPE_DIRECTW;
917 iov.iov_base = wpipe->pipe_buffer.buffer;
922 uio.uio_resid = size;
923 uio.uio_segflg = UIO_SYSSPACE;
924 uio.uio_rw = UIO_READ;
925 uio.uio_td = curthread;
926 uiomove_fromphys(wpipe->pipe_map.ms, pos, size, &uio);
928 pipe_destroy_write_buffer(wpipe);
932 * This implements the pipe buffer write mechanism. Note that only
933 * a direct write OR a normal pipe write can be pending at any given time.
934 * If there are any characters in the pipe buffer, the direct write will
935 * be deferred until the receiving process grabs all of the bytes from
936 * the pipe buffer. Then the direct mapping write is set-up.
939 pipe_direct_write(wpipe, uio)
946 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
947 error = pipelock(wpipe, 1);
950 if ((wpipe->pipe_state & PIPE_EOF) != 0) {
955 while (wpipe->pipe_state & PIPE_DIRECTW) {
956 if (wpipe->pipe_state & PIPE_WANTR) {
957 wpipe->pipe_state &= ~PIPE_WANTR;
960 pipeselwakeup(wpipe);
961 wpipe->pipe_state |= PIPE_WANTW;
963 error = msleep(wpipe, PIPE_MTX(wpipe),
964 PRIBIO | PCATCH, "pipdww", 0);
970 wpipe->pipe_map.cnt = 0; /* transfer not ready yet */
971 if (wpipe->pipe_buffer.cnt > 0) {
972 if (wpipe->pipe_state & PIPE_WANTR) {
973 wpipe->pipe_state &= ~PIPE_WANTR;
976 pipeselwakeup(wpipe);
977 wpipe->pipe_state |= PIPE_WANTW;
979 error = msleep(wpipe, PIPE_MTX(wpipe),
980 PRIBIO | PCATCH, "pipdwc", 0);
987 wpipe->pipe_state |= PIPE_DIRECTW;
990 error = pipe_build_write_buffer(wpipe, uio);
993 wpipe->pipe_state &= ~PIPE_DIRECTW;
999 while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
1000 if (wpipe->pipe_state & PIPE_EOF) {
1001 pipe_destroy_write_buffer(wpipe);
1002 pipeselwakeup(wpipe);
1007 if (wpipe->pipe_state & PIPE_WANTR) {
1008 wpipe->pipe_state &= ~PIPE_WANTR;
1011 pipeselwakeup(wpipe);
1012 wpipe->pipe_state |= PIPE_WANTW;
1014 error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH,
1019 if (wpipe->pipe_state & PIPE_EOF)
1021 if (wpipe->pipe_state & PIPE_DIRECTW) {
1023 * this bit of trickery substitutes a kernel buffer for
1024 * the process that might be going away.
1026 pipe_clone_write_buffer(wpipe);
1028 pipe_destroy_write_buffer(wpipe);
1040 pipe_write(fp, uio, active_cred, flags, td)
1043 struct ucred *active_cred;
1050 struct pipe *wpipe, *rpipe;
1053 wpipe = PIPE_PEER(rpipe);
1055 error = pipelock(wpipe, 1);
1061 * detect loss of pipe read side, issue SIGPIPE if lost.
1063 if (wpipe->pipe_present != PIPE_ACTIVE ||
1064 (wpipe->pipe_state & PIPE_EOF)) {
1070 error = mac_pipe_check_write(active_cred, wpipe->pipe_pair);
1079 /* Choose a larger size if it's advantageous */
1080 desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size);
1081 while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) {
1082 if (piperesizeallowed != 1)
1084 if (amountpipekva > maxpipekva / 2)
1086 if (desiredsize == BIG_PIPE_SIZE)
1088 desiredsize = desiredsize * 2;
1091 /* Choose a smaller size if we're in a OOM situation */
1092 if ((amountpipekva > (3 * maxpipekva) / 4) &&
1093 (wpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
1094 (wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
1095 (piperesizeallowed == 1))
1096 desiredsize = SMALL_PIPE_SIZE;
1098 /* Resize if the above determined that a new size was necessary */
1099 if ((desiredsize != wpipe->pipe_buffer.size) &&
1100 ((wpipe->pipe_state & PIPE_DIRECTW) == 0)) {
1102 pipespace(wpipe, desiredsize);
1105 if (wpipe->pipe_buffer.size == 0) {
1107 * This can only happen for reverse direction use of pipes
1108 * in a complete OOM situation.
1119 orig_resid = uio->uio_resid;
1121 while (uio->uio_resid) {
1125 if (wpipe->pipe_state & PIPE_EOF) {
1130 #ifndef PIPE_NODIRECT
1132 * If the transfer is large, we can gain performance if
1133 * we do process-to-process copies directly.
1134 * If the write is non-blocking, we don't use the
1135 * direct write mechanism.
1137 * The direct write mechanism will detect the reader going
1140 if (uio->uio_segflg == UIO_USERSPACE &&
1141 uio->uio_iov->iov_len >= PIPE_MINDIRECT &&
1142 wpipe->pipe_buffer.size >= PIPE_MINDIRECT &&
1143 (fp->f_flag & FNONBLOCK) == 0) {
1145 error = pipe_direct_write(wpipe, uio);
1153 * Pipe buffered writes cannot be coincidental with
1154 * direct writes. We wait until the currently executing
1155 * direct write is completed before we start filling the
1156 * pipe buffer. We break out if a signal occurs or the
1159 if (wpipe->pipe_state & PIPE_DIRECTW) {
1160 if (wpipe->pipe_state & PIPE_WANTR) {
1161 wpipe->pipe_state &= ~PIPE_WANTR;
1164 pipeselwakeup(wpipe);
1165 wpipe->pipe_state |= PIPE_WANTW;
1167 error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH,
1175 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1177 /* Writes of size <= PIPE_BUF must be atomic. */
1178 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
1182 int size; /* Transfer size */
1183 int segsize; /* first segment to transfer */
1186 * Transfer size is minimum of uio transfer
1187 * and free space in pipe buffer.
1189 if (space > uio->uio_resid)
1190 size = uio->uio_resid;
1194 * First segment to transfer is minimum of
1195 * transfer size and contiguous space in
1196 * pipe buffer. If first segment to transfer
1197 * is less than the transfer size, we've got
1198 * a wraparound in the buffer.
1200 segsize = wpipe->pipe_buffer.size -
1201 wpipe->pipe_buffer.in;
1205 /* Transfer first segment */
1208 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1212 if (error == 0 && segsize < size) {
1213 KASSERT(wpipe->pipe_buffer.in + segsize ==
1214 wpipe->pipe_buffer.size,
1215 ("Pipe buffer wraparound disappeared"));
1217 * Transfer remaining part now, to
1218 * support atomic writes. Wraparound
1224 &wpipe->pipe_buffer.buffer[0],
1225 size - segsize, uio);
1229 wpipe->pipe_buffer.in += size;
1230 if (wpipe->pipe_buffer.in >=
1231 wpipe->pipe_buffer.size) {
1232 KASSERT(wpipe->pipe_buffer.in ==
1234 wpipe->pipe_buffer.size,
1235 ("Expected wraparound bad"));
1236 wpipe->pipe_buffer.in = size - segsize;
1239 wpipe->pipe_buffer.cnt += size;
1240 KASSERT(wpipe->pipe_buffer.cnt <=
1241 wpipe->pipe_buffer.size,
1242 ("Pipe buffer overflow"));
1249 * If the "read-side" has been blocked, wake it up now.
1251 if (wpipe->pipe_state & PIPE_WANTR) {
1252 wpipe->pipe_state &= ~PIPE_WANTR;
1257 * don't block on non-blocking I/O
1259 if (fp->f_flag & FNONBLOCK) {
1266 * We have no more space and have something to offer,
1267 * wake up select/poll.
1269 pipeselwakeup(wpipe);
1271 wpipe->pipe_state |= PIPE_WANTW;
1273 error = msleep(wpipe, PIPE_MTX(rpipe),
1274 PRIBIO | PCATCH, "pipewr", 0);
1283 if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1284 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1286 } else if (wpipe->pipe_buffer.cnt > 0) {
1288 * If we have put any characters in the buffer, we wake up
1291 if (wpipe->pipe_state & PIPE_WANTR) {
1292 wpipe->pipe_state &= ~PIPE_WANTR;
1298 * Don't return EPIPE if any byte was written.
1299 * EINTR and other interrupts are handled by generic I/O layer.
1300 * Do not pretend that I/O succeeded for obvious user error
1303 if (uio->uio_resid != orig_resid && error == EPIPE)
1307 vfs_timestamp(&wpipe->pipe_mtime);
1310 * We have something to offer,
1311 * wake up select/poll.
1313 if (wpipe->pipe_buffer.cnt)
1314 pipeselwakeup(wpipe);
1323 pipe_truncate(fp, length, active_cred, td)
1326 struct ucred *active_cred;
1330 /* For named pipes call the vnode operation. */
1331 if (fp->f_vnode != NULL)
1332 return (vnops.fo_truncate(fp, length, active_cred, td));
1337 * we implement a very minimal set of ioctls for compatibility with sockets.
1340 pipe_ioctl(fp, cmd, data, active_cred, td)
1344 struct ucred *active_cred;
1347 struct pipe *mpipe = fp->f_data;
1353 error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data);
1368 mpipe->pipe_state |= PIPE_ASYNC;
1370 mpipe->pipe_state &= ~PIPE_ASYNC;
1375 if (!(fp->f_flag & FREAD)) {
1380 if (mpipe->pipe_state & PIPE_DIRECTW)
1381 *(int *)data = mpipe->pipe_map.cnt;
1383 *(int *)data = mpipe->pipe_buffer.cnt;
1388 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1392 *(int *)data = fgetown(&mpipe->pipe_sigio);
1395 /* This is deprecated, FIOSETOWN should be used instead. */
1398 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1401 /* This is deprecated, FIOGETOWN should be used instead. */
1403 *(int *)data = -fgetown(&mpipe->pipe_sigio);
1416 pipe_poll(fp, events, active_cred, td)
1419 struct ucred *active_cred;
1424 int levents, revents;
1431 wpipe = PIPE_PEER(rpipe);
1434 error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair);
1438 if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM))
1439 if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1440 (rpipe->pipe_buffer.cnt > 0))
1441 revents |= events & (POLLIN | POLLRDNORM);
1443 if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM))
1444 if (wpipe->pipe_present != PIPE_ACTIVE ||
1445 (wpipe->pipe_state & PIPE_EOF) ||
1446 (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1447 ((wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF ||
1448 wpipe->pipe_buffer.size == 0)))
1449 revents |= events & (POLLOUT | POLLWRNORM);
1452 (POLLIN | POLLINIGNEOF | POLLPRI | POLLRDNORM | POLLRDBAND);
1453 if (rpipe->pipe_state & PIPE_NAMED && fp->f_flag & FREAD && levents &&
1454 fp->f_seqcount == rpipe->pipe_wgen)
1455 events |= POLLINIGNEOF;
1457 if ((events & POLLINIGNEOF) == 0) {
1458 if (rpipe->pipe_state & PIPE_EOF) {
1459 revents |= (events & (POLLIN | POLLRDNORM));
1460 if (wpipe->pipe_present != PIPE_ACTIVE ||
1461 (wpipe->pipe_state & PIPE_EOF))
1467 if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM)) {
1468 selrecord(td, &rpipe->pipe_sel);
1469 if (SEL_WAITING(&rpipe->pipe_sel))
1470 rpipe->pipe_state |= PIPE_SEL;
1473 if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM)) {
1474 selrecord(td, &wpipe->pipe_sel);
1475 if (SEL_WAITING(&wpipe->pipe_sel))
1476 wpipe->pipe_state |= PIPE_SEL;
1488 * We shouldn't need locks here as we're doing a read and this should
1489 * be a natural race.
1492 pipe_stat(fp, ub, active_cred, td)
1495 struct ucred *active_cred;
1507 error = mac_pipe_check_stat(active_cred, pipe->pipe_pair);
1514 /* For named pipes ask the underlying filesystem. */
1515 if (pipe->pipe_state & PIPE_NAMED) {
1517 return (vnops.fo_stat(fp, ub, active_cred, td));
1521 * Lazily allocate an inode number for the pipe. Most pipe
1522 * users do not call fstat(2) on the pipe, which means that
1523 * postponing the inode allocation until it is must be
1524 * returned to userland is useful. If alloc_unr failed,
1525 * assign st_ino zero instead of returning an error.
1526 * Special pipe_ino values:
1527 * -1 - not yet initialized;
1528 * 0 - alloc_unr failed, return 0 as st_ino forever.
1530 if (pipe->pipe_ino == (ino_t)-1) {
1531 new_unr = alloc_unr(pipeino_unr);
1533 pipe->pipe_ino = new_unr;
1539 bzero(ub, sizeof(*ub));
1540 ub->st_mode = S_IFIFO;
1541 ub->st_blksize = PAGE_SIZE;
1542 if (pipe->pipe_state & PIPE_DIRECTW)
1543 ub->st_size = pipe->pipe_map.cnt;
1545 ub->st_size = pipe->pipe_buffer.cnt;
1546 ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1547 ub->st_atim = pipe->pipe_atime;
1548 ub->st_mtim = pipe->pipe_mtime;
1549 ub->st_ctim = pipe->pipe_ctime;
1550 ub->st_uid = fp->f_cred->cr_uid;
1551 ub->st_gid = fp->f_cred->cr_gid;
1552 ub->st_dev = pipedev_ino;
1553 ub->st_ino = pipe->pipe_ino;
1555 * Left as 0: st_nlink, st_rdev, st_flags, st_gen.
1567 if (fp->f_vnode != NULL)
1568 return vnops.fo_close(fp, td);
1569 fp->f_ops = &badfileops;
1570 pipe_dtor(fp->f_data);
1576 pipe_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td)
1582 if (cpipe->pipe_state & PIPE_NAMED)
1583 error = vn_chmod(fp, mode, active_cred, td);
1585 error = invfo_chmod(fp, mode, active_cred, td);
1590 pipe_chown(fp, uid, gid, active_cred, td)
1594 struct ucred *active_cred;
1601 if (cpipe->pipe_state & PIPE_NAMED)
1602 error = vn_chown(fp, uid, gid, active_cred, td);
1604 error = invfo_chown(fp, uid, gid, active_cred, td);
1609 pipe_free_kmem(cpipe)
1613 KASSERT(!mtx_owned(PIPE_MTX(cpipe)),
1614 ("pipe_free_kmem: pipe mutex locked"));
1616 if (cpipe->pipe_buffer.buffer != NULL) {
1617 atomic_subtract_long(&amountpipekva, cpipe->pipe_buffer.size);
1618 vm_map_remove(pipe_map,
1619 (vm_offset_t)cpipe->pipe_buffer.buffer,
1620 (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size);
1621 cpipe->pipe_buffer.buffer = NULL;
1623 #ifndef PIPE_NODIRECT
1625 cpipe->pipe_map.cnt = 0;
1626 cpipe->pipe_map.pos = 0;
1627 cpipe->pipe_map.npages = 0;
1639 struct pipepair *pp;
1642 KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL"));
1646 pp = cpipe->pipe_pair;
1648 pipeselwakeup(cpipe);
1651 * If the other side is blocked, wake it up saying that
1652 * we want to close it down.
1654 cpipe->pipe_state |= PIPE_EOF;
1655 while (cpipe->pipe_busy) {
1657 cpipe->pipe_state |= PIPE_WANT;
1659 msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0);
1665 * Disconnect from peer, if any.
1667 ppipe = cpipe->pipe_peer;
1668 if (ppipe->pipe_present == PIPE_ACTIVE) {
1669 pipeselwakeup(ppipe);
1671 ppipe->pipe_state |= PIPE_EOF;
1673 KNOTE_LOCKED(&ppipe->pipe_sel.si_note, 0);
1677 * Mark this endpoint as free. Release kmem resources. We
1678 * don't mark this endpoint as unused until we've finished
1679 * doing that, or the pipe might disappear out from under
1683 pipe_free_kmem(cpipe);
1685 cpipe->pipe_present = PIPE_CLOSING;
1689 * knlist_clear() may sleep dropping the PIPE_MTX. Set the
1690 * PIPE_FINALIZED, that allows other end to free the
1691 * pipe_pair, only after the knotes are completely dismantled.
1693 knlist_clear(&cpipe->pipe_sel.si_note, 1);
1694 cpipe->pipe_present = PIPE_FINALIZED;
1695 seldrain(&cpipe->pipe_sel);
1696 knlist_destroy(&cpipe->pipe_sel.si_note);
1699 * If both endpoints are now closed, release the memory for the
1700 * pipe pair. If not, unlock.
1702 if (ppipe->pipe_present == PIPE_FINALIZED) {
1705 mac_pipe_destroy(pp);
1707 uma_zfree(pipe_zone, cpipe->pipe_pair);
1714 pipe_kqfilter(struct file *fp, struct knote *kn)
1719 * If a filter is requested that is not supported by this file
1720 * descriptor, don't return an error, but also don't ever generate an
1723 if ((kn->kn_filter == EVFILT_READ) && !(fp->f_flag & FREAD)) {
1724 kn->kn_fop = &pipe_nfiltops;
1727 if ((kn->kn_filter == EVFILT_WRITE) && !(fp->f_flag & FWRITE)) {
1728 kn->kn_fop = &pipe_nfiltops;
1733 switch (kn->kn_filter) {
1735 kn->kn_fop = &pipe_rfiltops;
1738 kn->kn_fop = &pipe_wfiltops;
1739 if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) {
1740 /* other end of pipe has been closed */
1744 cpipe = PIPE_PEER(cpipe);
1751 kn->kn_hook = cpipe;
1752 knlist_add(&cpipe->pipe_sel.si_note, kn, 1);
1758 filt_pipedetach(struct knote *kn)
1760 struct pipe *cpipe = kn->kn_hook;
1763 knlist_remove(&cpipe->pipe_sel.si_note, kn, 1);
1769 filt_piperead(struct knote *kn, long hint)
1771 struct pipe *rpipe = kn->kn_hook;
1772 struct pipe *wpipe = rpipe->pipe_peer;
1775 PIPE_LOCK_ASSERT(rpipe, MA_OWNED);
1776 kn->kn_data = rpipe->pipe_buffer.cnt;
1777 if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
1778 kn->kn_data = rpipe->pipe_map.cnt;
1780 if ((rpipe->pipe_state & PIPE_EOF) ||
1781 wpipe->pipe_present != PIPE_ACTIVE ||
1782 (wpipe->pipe_state & PIPE_EOF)) {
1783 kn->kn_flags |= EV_EOF;
1786 ret = kn->kn_data > 0;
1792 filt_pipewrite(struct knote *kn, long hint)
1796 wpipe = kn->kn_hook;
1797 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
1798 if (wpipe->pipe_present != PIPE_ACTIVE ||
1799 (wpipe->pipe_state & PIPE_EOF)) {
1801 kn->kn_flags |= EV_EOF;
1804 kn->kn_data = (wpipe->pipe_buffer.size > 0) ?
1805 (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) : PIPE_BUF;
1806 if (wpipe->pipe_state & PIPE_DIRECTW)
1809 return (kn->kn_data >= PIPE_BUF);
1813 filt_pipedetach_notsup(struct knote *kn)
1819 filt_pipenotsup(struct knote *kn, long hint)