2 * Copyright (c) 1996 John S. Dyson
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice immediately at the beginning of the file, without modification,
10 * this list of conditions, and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
14 * 3. Absolutely no warranty of function or purpose is made by the author
16 * 4. Modifications may be freely made to this file if the above conditions
21 * This file contains a high-performance replacement for the socket-based
22 * pipes scheme originally used in FreeBSD/4.4Lite. It does not support
23 * all features of sockets, but does do everything that pipes normally
28 * This code has two modes of operation, a small write mode and a large
29 * write mode. The small write mode acts like conventional pipes with
30 * a kernel buffer. If the buffer is less than PIPE_MINDIRECT, then the
31 * "normal" pipe buffering is done. If the buffer is between PIPE_MINDIRECT
32 * and PIPE_SIZE in size, the sending process pins the underlying pages in
33 * memory, and the receiving process copies directly from these pinned pages
34 * in the sending process.
36 * If the sending process receives a signal, it is possible that it will
37 * go away, and certainly its address space can change, because control
38 * is returned back to the user-mode side. In that case, the pipe code
39 * arranges to copy the buffer supplied by the user process, to a pageable
40 * kernel buffer, and the receiving process will grab the data from the
41 * pageable kernel buffer. Since signals don't happen all that often,
42 * the copy operation is normally eliminated.
44 * The constant PIPE_MINDIRECT is chosen to make sure that buffering will
45 * happen for small transfers so that the system will not spend all of
46 * its time context switching.
48 * In order to limit the resource use of pipes, two sysctls exist:
50 * kern.ipc.maxpipekva - This is a hard limit on the amount of pageable
51 * address space available to us in pipe_map. This value is normally
52 * autotuned, but may also be loader tuned.
54 * kern.ipc.pipekva - This read-only sysctl tracks the current amount of
55 * memory in use by pipes.
57 * Based on how large pipekva is relative to maxpipekva, the following
61 * New pipes are given 16K of memory backing, pipes may dynamically
62 * grow to as large as 64K where needed.
64 * New pipes are given 4K (or PAGE_SIZE) of memory backing,
65 * existing pipes may NOT grow.
67 * New pipes are given 4K (or PAGE_SIZE) of memory backing,
68 * existing pipes will be shrunk down to 4K whenever possible.
70 * Resizing may be disabled by setting kern.ipc.piperesizeallowed=0. If
71 * that is set, the only resize that will occur is the 0 -> SMALL_PIPE_SIZE
72 * resize which MUST occur for reverse-direction pipes when they are
75 * Additional information about the current state of pipes may be obtained
76 * from kern.ipc.pipes, kern.ipc.pipefragretry, kern.ipc.pipeallocfail,
77 * and kern.ipc.piperesizefail.
79 * Locking rules: There are two locks present here: A mutex, used via
80 * PIPE_LOCK, and a flag, used via pipelock(). All locking is done via
81 * the flag, as mutexes can not persist over uiomove. The mutex
82 * exists only to guard access to the flag, and is not in itself a
83 * locking mechanism. Also note that there is only a single mutex for
84 * both directions of a pipe.
86 * As pipelock() may have to sleep before it can acquire the flag, it
87 * is important to reread all data after a call to pipelock(); everything
88 * in the structure may have changed.
91 #include <sys/cdefs.h>
92 __FBSDID("$FreeBSD$");
94 #include <sys/param.h>
95 #include <sys/systm.h>
97 #include <sys/fcntl.h>
99 #include <sys/filedesc.h>
100 #include <sys/filio.h>
101 #include <sys/kernel.h>
102 #include <sys/lock.h>
103 #include <sys/mutex.h>
104 #include <sys/ttycom.h>
105 #include <sys/stat.h>
106 #include <sys/malloc.h>
107 #include <sys/poll.h>
108 #include <sys/selinfo.h>
109 #include <sys/signalvar.h>
110 #include <sys/syscallsubr.h>
111 #include <sys/sysctl.h>
112 #include <sys/sysproto.h>
113 #include <sys/pipe.h>
114 #include <sys/proc.h>
115 #include <sys/vnode.h>
117 #include <sys/event.h>
119 #include <security/mac/mac_framework.h>
122 #include <vm/vm_param.h>
123 #include <vm/vm_object.h>
124 #include <vm/vm_kern.h>
125 #include <vm/vm_extern.h>
127 #include <vm/vm_map.h>
128 #include <vm/vm_page.h>
132 int do_pipe(struct thread *td, int fildes[2], int flags);
135 * Use this define if you want to disable *fancy* VM things. Expect an
136 * approx 30% decrease in transfer rate. This could be useful for
139 /* #define PIPE_NODIRECT */
142 * interfaces to the outside world
144 static fo_rdwr_t pipe_read;
145 static fo_rdwr_t pipe_write;
146 static fo_truncate_t pipe_truncate;
147 static fo_ioctl_t pipe_ioctl;
148 static fo_poll_t pipe_poll;
149 static fo_kqfilter_t pipe_kqfilter;
150 static fo_stat_t pipe_stat;
151 static fo_close_t pipe_close;
153 static struct fileops pipeops = {
154 .fo_read = pipe_read,
155 .fo_write = pipe_write,
156 .fo_truncate = pipe_truncate,
157 .fo_ioctl = pipe_ioctl,
158 .fo_poll = pipe_poll,
159 .fo_kqfilter = pipe_kqfilter,
160 .fo_stat = pipe_stat,
161 .fo_close = pipe_close,
162 .fo_chmod = invfo_chmod,
163 .fo_chown = invfo_chown,
164 .fo_flags = DFLAG_PASSABLE
167 static void filt_pipedetach(struct knote *kn);
168 static int filt_piperead(struct knote *kn, long hint);
169 static int filt_pipewrite(struct knote *kn, long hint);
171 static struct filterops pipe_rfiltops = {
173 .f_detach = filt_pipedetach,
174 .f_event = filt_piperead
176 static struct filterops pipe_wfiltops = {
178 .f_detach = filt_pipedetach,
179 .f_event = filt_pipewrite
183 * Default pipe buffer size(s), this can be kind-of large now because pipe
184 * space is pageable. The pipe code will try to maintain locality of
185 * reference for performance reasons, so small amounts of outstanding I/O
186 * will not wipe the cache.
188 #define MINPIPESIZE (PIPE_SIZE/3)
189 #define MAXPIPESIZE (2*PIPE_SIZE/3)
191 static long amountpipekva;
192 static int pipefragretry;
193 static int pipeallocfail;
194 static int piperesizefail;
195 static int piperesizeallowed = 1;
197 SYSCTL_LONG(_kern_ipc, OID_AUTO, maxpipekva, CTLFLAG_RDTUN,
198 &maxpipekva, 0, "Pipe KVA limit");
199 SYSCTL_LONG(_kern_ipc, OID_AUTO, pipekva, CTLFLAG_RD,
200 &amountpipekva, 0, "Pipe KVA usage");
201 SYSCTL_INT(_kern_ipc, OID_AUTO, pipefragretry, CTLFLAG_RD,
202 &pipefragretry, 0, "Pipe allocation retries due to fragmentation");
203 SYSCTL_INT(_kern_ipc, OID_AUTO, pipeallocfail, CTLFLAG_RD,
204 &pipeallocfail, 0, "Pipe allocation failures");
205 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizefail, CTLFLAG_RD,
206 &piperesizefail, 0, "Pipe resize failures");
207 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizeallowed, CTLFLAG_RW,
208 &piperesizeallowed, 0, "Pipe resizing allowed");
210 static void pipeinit(void *dummy __unused);
211 static void pipeclose(struct pipe *cpipe);
212 static void pipe_free_kmem(struct pipe *cpipe);
213 static int pipe_create(struct pipe *pipe, int backing);
214 static __inline int pipelock(struct pipe *cpipe, int catch);
215 static __inline void pipeunlock(struct pipe *cpipe);
216 static __inline void pipeselwakeup(struct pipe *cpipe);
217 #ifndef PIPE_NODIRECT
218 static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio);
219 static void pipe_destroy_write_buffer(struct pipe *wpipe);
220 static int pipe_direct_write(struct pipe *wpipe, struct uio *uio);
221 static void pipe_clone_write_buffer(struct pipe *wpipe);
223 static int pipespace(struct pipe *cpipe, int size);
224 static int pipespace_new(struct pipe *cpipe, int size);
226 static int pipe_zone_ctor(void *mem, int size, void *arg, int flags);
227 static int pipe_zone_init(void *mem, int size, int flags);
228 static void pipe_zone_fini(void *mem, int size);
230 static uma_zone_t pipe_zone;
231 static struct unrhdr *pipeino_unr;
232 static dev_t pipedev_ino;
234 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL);
237 pipeinit(void *dummy __unused)
240 pipe_zone = uma_zcreate("pipe", sizeof(struct pipepair),
241 pipe_zone_ctor, NULL, pipe_zone_init, pipe_zone_fini,
243 KASSERT(pipe_zone != NULL, ("pipe_zone not initialized"));
244 pipeino_unr = new_unrhdr(1, INT32_MAX, NULL);
245 KASSERT(pipeino_unr != NULL, ("pipe fake inodes not initialized"));
246 pipedev_ino = devfs_alloc_cdp_inode();
247 KASSERT(pipedev_ino > 0, ("pipe dev inode not initialized"));
251 pipe_zone_ctor(void *mem, int size, void *arg, int flags)
254 struct pipe *rpipe, *wpipe;
256 KASSERT(size == sizeof(*pp), ("pipe_zone_ctor: wrong size"));
258 pp = (struct pipepair *)mem;
261 * We zero both pipe endpoints to make sure all the kmem pointers
262 * are NULL, flag fields are zero'd, etc. We timestamp both
263 * endpoints with the same time.
265 rpipe = &pp->pp_rpipe;
266 bzero(rpipe, sizeof(*rpipe));
267 vfs_timestamp(&rpipe->pipe_ctime);
268 rpipe->pipe_atime = rpipe->pipe_mtime = rpipe->pipe_ctime;
270 wpipe = &pp->pp_wpipe;
271 bzero(wpipe, sizeof(*wpipe));
272 wpipe->pipe_ctime = rpipe->pipe_ctime;
273 wpipe->pipe_atime = wpipe->pipe_mtime = rpipe->pipe_ctime;
275 rpipe->pipe_peer = wpipe;
276 rpipe->pipe_pair = pp;
277 wpipe->pipe_peer = rpipe;
278 wpipe->pipe_pair = pp;
281 * Mark both endpoints as present; they will later get free'd
282 * one at a time. When both are free'd, then the whole pair
285 rpipe->pipe_present = PIPE_ACTIVE;
286 wpipe->pipe_present = PIPE_ACTIVE;
289 * Eventually, the MAC Framework may initialize the label
290 * in ctor or init, but for now we do it elswhere to avoid
291 * blocking in ctor or init.
299 pipe_zone_init(void *mem, int size, int flags)
303 KASSERT(size == sizeof(*pp), ("pipe_zone_init: wrong size"));
305 pp = (struct pipepair *)mem;
307 mtx_init(&pp->pp_mtx, "pipe mutex", NULL, MTX_DEF | MTX_RECURSE);
312 pipe_zone_fini(void *mem, int size)
316 KASSERT(size == sizeof(*pp), ("pipe_zone_fini: wrong size"));
318 pp = (struct pipepair *)mem;
320 mtx_destroy(&pp->pp_mtx);
324 * The pipe system call for the DTYPE_PIPE type of pipes. If we fail, let
325 * the zone pick up the pieces via pipeclose().
328 kern_pipe(struct thread *td, int fildes[2])
331 return (do_pipe(td, fildes, 0));
335 do_pipe(struct thread *td, int fildes[2], int flags)
337 struct filedesc *fdp = td->td_proc->p_fd;
338 struct file *rf, *wf;
340 struct pipe *rpipe, *wpipe;
341 int fd, fflags, error;
343 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;
370 error = falloc(td, &rf, &fd, flags);
376 /* An extra reference on `rf' has been held for us by falloc(). */
379 fflags = FREAD | FWRITE;
380 if ((flags & O_NONBLOCK) != 0)
384 * Warning: once we've gotten past allocation of the fd for the
385 * read-side, we can only drop the read side via fdrop() in order
386 * to avoid races against processes which manage to dup() the read
387 * side while we are blocked trying to allocate the write side.
389 finit(rf, fflags, DTYPE_PIPE, rpipe, &pipeops);
390 error = falloc(td, &wf, &fd, flags);
392 fdclose(fdp, rf, fildes[0], td);
394 /* rpipe has been closed by fdrop(). */
398 /* An extra reference on `wf' has been held for us by falloc(). */
399 finit(wf, fflags, DTYPE_PIPE, wpipe, &pipeops);
409 sys_pipe(struct thread *td, struct pipe_args *uap)
414 error = kern_pipe(td, fildes);
418 td->td_retval[0] = fildes[0];
419 td->td_retval[1] = fildes[1];
425 * Allocate kva for pipe circular buffer, the space is pageable
426 * This routine will 'realloc' the size of a pipe safely, if it fails
427 * it will retain the old buffer.
428 * If it fails it will return ENOMEM.
431 pipespace_new(cpipe, size)
436 int error, cnt, firstseg;
437 static int curfail = 0;
438 static struct timeval lastfail;
440 KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked"));
441 KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW),
442 ("pipespace: resize of direct writes not allowed"));
444 cnt = cpipe->pipe_buffer.cnt;
448 size = round_page(size);
449 buffer = (caddr_t) vm_map_min(pipe_map);
451 error = vm_map_find(pipe_map, NULL, 0,
452 (vm_offset_t *) &buffer, size, 1,
453 VM_PROT_ALL, VM_PROT_ALL, 0);
454 if (error != KERN_SUCCESS) {
455 if ((cpipe->pipe_buffer.buffer == NULL) &&
456 (size > SMALL_PIPE_SIZE)) {
457 size = SMALL_PIPE_SIZE;
461 if (cpipe->pipe_buffer.buffer == NULL) {
463 if (ppsratecheck(&lastfail, &curfail, 1))
464 printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n");
471 /* copy data, then free old resources if we're resizing */
473 if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) {
474 firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out;
475 bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
477 if ((cnt - firstseg) > 0)
478 bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg],
479 cpipe->pipe_buffer.in);
481 bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
485 pipe_free_kmem(cpipe);
486 cpipe->pipe_buffer.buffer = buffer;
487 cpipe->pipe_buffer.size = size;
488 cpipe->pipe_buffer.in = cnt;
489 cpipe->pipe_buffer.out = 0;
490 cpipe->pipe_buffer.cnt = cnt;
491 atomic_add_long(&amountpipekva, cpipe->pipe_buffer.size);
496 * Wrapper for pipespace_new() that performs locking assertions.
499 pipespace(cpipe, size)
504 KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
505 ("Unlocked pipe passed to pipespace"));
506 return (pipespace_new(cpipe, size));
510 * lock a pipe for I/O, blocking other access
513 pipelock(cpipe, catch)
519 PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
520 while (cpipe->pipe_state & PIPE_LOCKFL) {
521 cpipe->pipe_state |= PIPE_LWANT;
522 error = msleep(cpipe, PIPE_MTX(cpipe),
523 catch ? (PRIBIO | PCATCH) : PRIBIO,
528 cpipe->pipe_state |= PIPE_LOCKFL;
533 * unlock a pipe I/O lock
540 PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
541 KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
542 ("Unlocked pipe passed to pipeunlock"));
543 cpipe->pipe_state &= ~PIPE_LOCKFL;
544 if (cpipe->pipe_state & PIPE_LWANT) {
545 cpipe->pipe_state &= ~PIPE_LWANT;
555 PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
556 if (cpipe->pipe_state & PIPE_SEL) {
557 selwakeuppri(&cpipe->pipe_sel, PSOCK);
558 if (!SEL_WAITING(&cpipe->pipe_sel))
559 cpipe->pipe_state &= ~PIPE_SEL;
561 if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
562 pgsigio(&cpipe->pipe_sigio, SIGIO, 0);
563 KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0);
567 * Initialize and allocate VM and memory for pipe. The structure
568 * will start out zero'd from the ctor, so we just manage the kmem.
571 pipe_create(pipe, backing)
578 if (amountpipekva > maxpipekva / 2)
579 error = pipespace_new(pipe, SMALL_PIPE_SIZE);
581 error = pipespace_new(pipe, PIPE_SIZE);
583 /* If we're not backing this pipe, no need to do anything. */
592 pipe_read(fp, uio, active_cred, flags, td)
595 struct ucred *active_cred;
599 struct pipe *rpipe = fp->f_data;
606 error = pipelock(rpipe, 1);
611 error = mac_pipe_check_read(active_cred, rpipe->pipe_pair);
615 if (amountpipekva > (3 * maxpipekva) / 4) {
616 if (!(rpipe->pipe_state & PIPE_DIRECTW) &&
617 (rpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
618 (rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
619 (piperesizeallowed == 1)) {
621 pipespace(rpipe, SMALL_PIPE_SIZE);
626 while (uio->uio_resid) {
628 * normal pipe buffer receive
630 if (rpipe->pipe_buffer.cnt > 0) {
631 size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
632 if (size > rpipe->pipe_buffer.cnt)
633 size = rpipe->pipe_buffer.cnt;
634 if (size > uio->uio_resid)
635 size = uio->uio_resid;
639 &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
645 rpipe->pipe_buffer.out += size;
646 if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
647 rpipe->pipe_buffer.out = 0;
649 rpipe->pipe_buffer.cnt -= size;
652 * If there is no more to read in the pipe, reset
653 * its pointers to the beginning. This improves
656 if (rpipe->pipe_buffer.cnt == 0) {
657 rpipe->pipe_buffer.in = 0;
658 rpipe->pipe_buffer.out = 0;
661 #ifndef PIPE_NODIRECT
663 * Direct copy, bypassing a kernel buffer.
665 } else if ((size = rpipe->pipe_map.cnt) &&
666 (rpipe->pipe_state & PIPE_DIRECTW)) {
667 if (size > uio->uio_resid)
668 size = (u_int) uio->uio_resid;
671 error = uiomove_fromphys(rpipe->pipe_map.ms,
672 rpipe->pipe_map.pos, size, uio);
677 rpipe->pipe_map.pos += size;
678 rpipe->pipe_map.cnt -= size;
679 if (rpipe->pipe_map.cnt == 0) {
680 rpipe->pipe_state &= ~PIPE_DIRECTW;
686 * detect EOF condition
687 * read returns 0 on EOF, no need to set error
689 if (rpipe->pipe_state & PIPE_EOF)
693 * If the "write-side" has been blocked, wake it up now.
695 if (rpipe->pipe_state & PIPE_WANTW) {
696 rpipe->pipe_state &= ~PIPE_WANTW;
701 * Break if some data was read.
707 * Unlock the pipe buffer for our remaining processing.
708 * We will either break out with an error or we will
709 * sleep and relock to loop.
714 * Handle non-blocking mode operation or
715 * wait for more data.
717 if (fp->f_flag & FNONBLOCK) {
720 rpipe->pipe_state |= PIPE_WANTR;
721 if ((error = msleep(rpipe, PIPE_MTX(rpipe),
724 error = pipelock(rpipe, 1);
735 /* XXX: should probably do this before getting any locks. */
737 vfs_timestamp(&rpipe->pipe_atime);
742 * PIPE_WANT processing only makes sense if pipe_busy is 0.
744 if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
745 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
747 } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
749 * Handle write blocking hysteresis.
751 if (rpipe->pipe_state & PIPE_WANTW) {
752 rpipe->pipe_state &= ~PIPE_WANTW;
757 if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
758 pipeselwakeup(rpipe);
764 #ifndef PIPE_NODIRECT
766 * Map the sending processes' buffer into kernel space and wire it.
767 * This is similar to a physical write operation.
770 pipe_build_write_buffer(wpipe, uio)
777 PIPE_LOCK_ASSERT(wpipe, MA_NOTOWNED);
778 KASSERT(wpipe->pipe_state & PIPE_DIRECTW,
779 ("Clone attempt on non-direct write pipe!"));
781 if (uio->uio_iov->iov_len > wpipe->pipe_buffer.size)
782 size = wpipe->pipe_buffer.size;
784 size = uio->uio_iov->iov_len;
786 if ((i = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
787 (vm_offset_t)uio->uio_iov->iov_base, size, VM_PROT_READ,
788 wpipe->pipe_map.ms, PIPENPAGES)) < 0)
792 * set up the control block
794 wpipe->pipe_map.npages = i;
795 wpipe->pipe_map.pos =
796 ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
797 wpipe->pipe_map.cnt = size;
800 * and update the uio data
803 uio->uio_iov->iov_len -= size;
804 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size;
805 if (uio->uio_iov->iov_len == 0)
807 uio->uio_resid -= size;
808 uio->uio_offset += size;
813 * unmap and unwire the process buffer
816 pipe_destroy_write_buffer(wpipe)
820 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
821 vm_page_unhold_pages(wpipe->pipe_map.ms, wpipe->pipe_map.npages);
822 wpipe->pipe_map.npages = 0;
826 * In the case of a signal, the writing process might go away. This
827 * code copies the data into the circular buffer so that the source
828 * pages can be freed without loss of data.
831 pipe_clone_write_buffer(wpipe)
839 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
840 size = wpipe->pipe_map.cnt;
841 pos = wpipe->pipe_map.pos;
843 wpipe->pipe_buffer.in = size;
844 wpipe->pipe_buffer.out = 0;
845 wpipe->pipe_buffer.cnt = size;
846 wpipe->pipe_state &= ~PIPE_DIRECTW;
849 iov.iov_base = wpipe->pipe_buffer.buffer;
854 uio.uio_resid = size;
855 uio.uio_segflg = UIO_SYSSPACE;
856 uio.uio_rw = UIO_READ;
857 uio.uio_td = curthread;
858 uiomove_fromphys(wpipe->pipe_map.ms, pos, size, &uio);
860 pipe_destroy_write_buffer(wpipe);
864 * This implements the pipe buffer write mechanism. Note that only
865 * a direct write OR a normal pipe write can be pending at any given time.
866 * If there are any characters in the pipe buffer, the direct write will
867 * be deferred until the receiving process grabs all of the bytes from
868 * the pipe buffer. Then the direct mapping write is set-up.
871 pipe_direct_write(wpipe, uio)
878 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
879 error = pipelock(wpipe, 1);
880 if (wpipe->pipe_state & PIPE_EOF)
886 while (wpipe->pipe_state & PIPE_DIRECTW) {
887 if (wpipe->pipe_state & PIPE_WANTR) {
888 wpipe->pipe_state &= ~PIPE_WANTR;
891 pipeselwakeup(wpipe);
892 wpipe->pipe_state |= PIPE_WANTW;
894 error = msleep(wpipe, PIPE_MTX(wpipe),
895 PRIBIO | PCATCH, "pipdww", 0);
901 wpipe->pipe_map.cnt = 0; /* transfer not ready yet */
902 if (wpipe->pipe_buffer.cnt > 0) {
903 if (wpipe->pipe_state & PIPE_WANTR) {
904 wpipe->pipe_state &= ~PIPE_WANTR;
907 pipeselwakeup(wpipe);
908 wpipe->pipe_state |= PIPE_WANTW;
910 error = msleep(wpipe, PIPE_MTX(wpipe),
911 PRIBIO | PCATCH, "pipdwc", 0);
918 wpipe->pipe_state |= PIPE_DIRECTW;
921 error = pipe_build_write_buffer(wpipe, uio);
924 wpipe->pipe_state &= ~PIPE_DIRECTW;
930 while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
931 if (wpipe->pipe_state & PIPE_EOF) {
932 pipe_destroy_write_buffer(wpipe);
933 pipeselwakeup(wpipe);
938 if (wpipe->pipe_state & PIPE_WANTR) {
939 wpipe->pipe_state &= ~PIPE_WANTR;
942 pipeselwakeup(wpipe);
944 error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH,
949 if (wpipe->pipe_state & PIPE_EOF)
951 if (wpipe->pipe_state & PIPE_DIRECTW) {
953 * this bit of trickery substitutes a kernel buffer for
954 * the process that might be going away.
956 pipe_clone_write_buffer(wpipe);
958 pipe_destroy_write_buffer(wpipe);
970 pipe_write(fp, uio, active_cred, flags, td)
973 struct ucred *active_cred;
980 struct pipe *wpipe, *rpipe;
983 wpipe = rpipe->pipe_peer;
986 error = pipelock(wpipe, 1);
992 * detect loss of pipe read side, issue SIGPIPE if lost.
994 if (wpipe->pipe_present != PIPE_ACTIVE ||
995 (wpipe->pipe_state & PIPE_EOF)) {
1001 error = mac_pipe_check_write(active_cred, wpipe->pipe_pair);
1010 /* Choose a larger size if it's advantageous */
1011 desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size);
1012 while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) {
1013 if (piperesizeallowed != 1)
1015 if (amountpipekva > maxpipekva / 2)
1017 if (desiredsize == BIG_PIPE_SIZE)
1019 desiredsize = desiredsize * 2;
1022 /* Choose a smaller size if we're in a OOM situation */
1023 if ((amountpipekva > (3 * maxpipekva) / 4) &&
1024 (wpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
1025 (wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
1026 (piperesizeallowed == 1))
1027 desiredsize = SMALL_PIPE_SIZE;
1029 /* Resize if the above determined that a new size was necessary */
1030 if ((desiredsize != wpipe->pipe_buffer.size) &&
1031 ((wpipe->pipe_state & PIPE_DIRECTW) == 0)) {
1033 pipespace(wpipe, desiredsize);
1036 if (wpipe->pipe_buffer.size == 0) {
1038 * This can only happen for reverse direction use of pipes
1039 * in a complete OOM situation.
1050 orig_resid = uio->uio_resid;
1052 while (uio->uio_resid) {
1056 if (wpipe->pipe_state & PIPE_EOF) {
1061 #ifndef PIPE_NODIRECT
1063 * If the transfer is large, we can gain performance if
1064 * we do process-to-process copies directly.
1065 * If the write is non-blocking, we don't use the
1066 * direct write mechanism.
1068 * The direct write mechanism will detect the reader going
1071 if (uio->uio_segflg == UIO_USERSPACE &&
1072 uio->uio_iov->iov_len >= PIPE_MINDIRECT &&
1073 wpipe->pipe_buffer.size >= PIPE_MINDIRECT &&
1074 (fp->f_flag & FNONBLOCK) == 0) {
1076 error = pipe_direct_write(wpipe, uio);
1084 * Pipe buffered writes cannot be coincidental with
1085 * direct writes. We wait until the currently executing
1086 * direct write is completed before we start filling the
1087 * pipe buffer. We break out if a signal occurs or the
1090 if (wpipe->pipe_state & PIPE_DIRECTW) {
1091 if (wpipe->pipe_state & PIPE_WANTR) {
1092 wpipe->pipe_state &= ~PIPE_WANTR;
1095 pipeselwakeup(wpipe);
1096 wpipe->pipe_state |= PIPE_WANTW;
1098 error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH,
1106 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1108 /* Writes of size <= PIPE_BUF must be atomic. */
1109 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
1113 int size; /* Transfer size */
1114 int segsize; /* first segment to transfer */
1117 * Transfer size is minimum of uio transfer
1118 * and free space in pipe buffer.
1120 if (space > uio->uio_resid)
1121 size = uio->uio_resid;
1125 * First segment to transfer is minimum of
1126 * transfer size and contiguous space in
1127 * pipe buffer. If first segment to transfer
1128 * is less than the transfer size, we've got
1129 * a wraparound in the buffer.
1131 segsize = wpipe->pipe_buffer.size -
1132 wpipe->pipe_buffer.in;
1136 /* Transfer first segment */
1139 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1143 if (error == 0 && segsize < size) {
1144 KASSERT(wpipe->pipe_buffer.in + segsize ==
1145 wpipe->pipe_buffer.size,
1146 ("Pipe buffer wraparound disappeared"));
1148 * Transfer remaining part now, to
1149 * support atomic writes. Wraparound
1155 &wpipe->pipe_buffer.buffer[0],
1156 size - segsize, uio);
1160 wpipe->pipe_buffer.in += size;
1161 if (wpipe->pipe_buffer.in >=
1162 wpipe->pipe_buffer.size) {
1163 KASSERT(wpipe->pipe_buffer.in ==
1165 wpipe->pipe_buffer.size,
1166 ("Expected wraparound bad"));
1167 wpipe->pipe_buffer.in = size - segsize;
1170 wpipe->pipe_buffer.cnt += size;
1171 KASSERT(wpipe->pipe_buffer.cnt <=
1172 wpipe->pipe_buffer.size,
1173 ("Pipe buffer overflow"));
1180 * If the "read-side" has been blocked, wake it up now.
1182 if (wpipe->pipe_state & PIPE_WANTR) {
1183 wpipe->pipe_state &= ~PIPE_WANTR;
1188 * don't block on non-blocking I/O
1190 if (fp->f_flag & FNONBLOCK) {
1197 * We have no more space and have something to offer,
1198 * wake up select/poll.
1200 pipeselwakeup(wpipe);
1202 wpipe->pipe_state |= PIPE_WANTW;
1204 error = msleep(wpipe, PIPE_MTX(rpipe),
1205 PRIBIO | PCATCH, "pipewr", 0);
1214 if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1215 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1217 } else if (wpipe->pipe_buffer.cnt > 0) {
1219 * If we have put any characters in the buffer, we wake up
1222 if (wpipe->pipe_state & PIPE_WANTR) {
1223 wpipe->pipe_state &= ~PIPE_WANTR;
1229 * Don't return EPIPE if I/O was successful
1231 if ((wpipe->pipe_buffer.cnt == 0) &&
1232 (uio->uio_resid == 0) &&
1238 vfs_timestamp(&wpipe->pipe_mtime);
1241 * We have something to offer,
1242 * wake up select/poll.
1244 if (wpipe->pipe_buffer.cnt)
1245 pipeselwakeup(wpipe);
1254 pipe_truncate(fp, length, active_cred, td)
1257 struct ucred *active_cred;
1265 * we implement a very minimal set of ioctls for compatibility with sockets.
1268 pipe_ioctl(fp, cmd, data, active_cred, td)
1272 struct ucred *active_cred;
1275 struct pipe *mpipe = fp->f_data;
1281 error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data);
1296 mpipe->pipe_state |= PIPE_ASYNC;
1298 mpipe->pipe_state &= ~PIPE_ASYNC;
1303 if (mpipe->pipe_state & PIPE_DIRECTW)
1304 *(int *)data = mpipe->pipe_map.cnt;
1306 *(int *)data = mpipe->pipe_buffer.cnt;
1311 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1315 *(int *)data = fgetown(&mpipe->pipe_sigio);
1318 /* This is deprecated, FIOSETOWN should be used instead. */
1321 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1324 /* This is deprecated, FIOGETOWN should be used instead. */
1326 *(int *)data = -fgetown(&mpipe->pipe_sigio);
1339 pipe_poll(fp, events, active_cred, td)
1342 struct ucred *active_cred;
1345 struct pipe *rpipe = fp->f_data;
1352 wpipe = rpipe->pipe_peer;
1355 error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair);
1359 if (events & (POLLIN | POLLRDNORM))
1360 if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1361 (rpipe->pipe_buffer.cnt > 0))
1362 revents |= events & (POLLIN | POLLRDNORM);
1364 if (events & (POLLOUT | POLLWRNORM))
1365 if (wpipe->pipe_present != PIPE_ACTIVE ||
1366 (wpipe->pipe_state & PIPE_EOF) ||
1367 (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1368 ((wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF ||
1369 wpipe->pipe_buffer.size == 0)))
1370 revents |= events & (POLLOUT | POLLWRNORM);
1372 if ((events & POLLINIGNEOF) == 0) {
1373 if (rpipe->pipe_state & PIPE_EOF) {
1374 revents |= (events & (POLLIN | POLLRDNORM));
1375 if (wpipe->pipe_present != PIPE_ACTIVE ||
1376 (wpipe->pipe_state & PIPE_EOF))
1382 if (events & (POLLIN | POLLRDNORM)) {
1383 selrecord(td, &rpipe->pipe_sel);
1384 if (SEL_WAITING(&rpipe->pipe_sel))
1385 rpipe->pipe_state |= PIPE_SEL;
1388 if (events & (POLLOUT | POLLWRNORM)) {
1389 selrecord(td, &wpipe->pipe_sel);
1390 if (SEL_WAITING(&wpipe->pipe_sel))
1391 wpipe->pipe_state |= PIPE_SEL;
1403 * We shouldn't need locks here as we're doing a read and this should
1404 * be a natural race.
1407 pipe_stat(fp, ub, active_cred, td)
1410 struct ucred *active_cred;
1422 error = mac_pipe_check_stat(active_cred, pipe->pipe_pair);
1429 * Lazily allocate an inode number for the pipe. Most pipe
1430 * users do not call fstat(2) on the pipe, which means that
1431 * postponing the inode allocation until it is must be
1432 * returned to userland is useful. If alloc_unr failed,
1433 * assign st_ino zero instead of returning an error.
1434 * Special pipe_ino values:
1435 * -1 - not yet initialized;
1436 * 0 - alloc_unr failed, return 0 as st_ino forever.
1438 if (pipe->pipe_ino == (ino_t)-1) {
1439 new_unr = alloc_unr(pipeino_unr);
1441 pipe->pipe_ino = new_unr;
1447 bzero(ub, sizeof(*ub));
1448 ub->st_mode = S_IFIFO;
1449 ub->st_blksize = PAGE_SIZE;
1450 if (pipe->pipe_state & PIPE_DIRECTW)
1451 ub->st_size = pipe->pipe_map.cnt;
1453 ub->st_size = pipe->pipe_buffer.cnt;
1454 ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1455 ub->st_atim = pipe->pipe_atime;
1456 ub->st_mtim = pipe->pipe_mtime;
1457 ub->st_ctim = pipe->pipe_ctime;
1458 ub->st_uid = fp->f_cred->cr_uid;
1459 ub->st_gid = fp->f_cred->cr_gid;
1460 ub->st_dev = pipedev_ino;
1461 ub->st_ino = pipe->pipe_ino;
1463 * Left as 0: st_nlink, st_rdev, st_flags, st_gen.
1474 struct pipe *cpipe = fp->f_data;
1476 fp->f_ops = &badfileops;
1478 funsetown(&cpipe->pipe_sigio);
1484 pipe_free_kmem(cpipe)
1488 KASSERT(!mtx_owned(PIPE_MTX(cpipe)),
1489 ("pipe_free_kmem: pipe mutex locked"));
1491 if (cpipe->pipe_buffer.buffer != NULL) {
1492 atomic_subtract_long(&amountpipekva, cpipe->pipe_buffer.size);
1493 vm_map_remove(pipe_map,
1494 (vm_offset_t)cpipe->pipe_buffer.buffer,
1495 (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size);
1496 cpipe->pipe_buffer.buffer = NULL;
1498 #ifndef PIPE_NODIRECT
1500 cpipe->pipe_map.cnt = 0;
1501 cpipe->pipe_map.pos = 0;
1502 cpipe->pipe_map.npages = 0;
1514 struct pipepair *pp;
1518 KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL"));
1522 pp = cpipe->pipe_pair;
1524 pipeselwakeup(cpipe);
1527 * If the other side is blocked, wake it up saying that
1528 * we want to close it down.
1530 cpipe->pipe_state |= PIPE_EOF;
1531 while (cpipe->pipe_busy) {
1533 cpipe->pipe_state |= PIPE_WANT;
1535 msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0);
1541 * Disconnect from peer, if any.
1543 ppipe = cpipe->pipe_peer;
1544 if (ppipe->pipe_present == PIPE_ACTIVE) {
1545 pipeselwakeup(ppipe);
1547 ppipe->pipe_state |= PIPE_EOF;
1549 KNOTE_LOCKED(&ppipe->pipe_sel.si_note, 0);
1553 * Mark this endpoint as free. Release kmem resources. We
1554 * don't mark this endpoint as unused until we've finished
1555 * doing that, or the pipe might disappear out from under
1559 pipe_free_kmem(cpipe);
1561 cpipe->pipe_present = PIPE_CLOSING;
1565 * knlist_clear() may sleep dropping the PIPE_MTX. Set the
1566 * PIPE_FINALIZED, that allows other end to free the
1567 * pipe_pair, only after the knotes are completely dismantled.
1569 knlist_clear(&cpipe->pipe_sel.si_note, 1);
1570 cpipe->pipe_present = PIPE_FINALIZED;
1571 seldrain(&cpipe->pipe_sel);
1572 knlist_destroy(&cpipe->pipe_sel.si_note);
1575 * Postpone the destroy of the fake inode number allocated for
1576 * our end, until pipe mtx is unlocked.
1578 ino = cpipe->pipe_ino;
1581 * If both endpoints are now closed, release the memory for the
1582 * pipe pair. If not, unlock.
1584 if (ppipe->pipe_present == PIPE_FINALIZED) {
1587 mac_pipe_destroy(pp);
1589 uma_zfree(pipe_zone, cpipe->pipe_pair);
1593 if (ino != 0 && ino != (ino_t)-1)
1594 free_unr(pipeino_unr, ino);
1599 pipe_kqfilter(struct file *fp, struct knote *kn)
1603 cpipe = kn->kn_fp->f_data;
1605 switch (kn->kn_filter) {
1607 kn->kn_fop = &pipe_rfiltops;
1610 kn->kn_fop = &pipe_wfiltops;
1611 if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) {
1612 /* other end of pipe has been closed */
1616 cpipe = cpipe->pipe_peer;
1623 knlist_add(&cpipe->pipe_sel.si_note, kn, 1);
1629 filt_pipedetach(struct knote *kn)
1631 struct pipe *cpipe = (struct pipe *)kn->kn_fp->f_data;
1634 if (kn->kn_filter == EVFILT_WRITE)
1635 cpipe = cpipe->pipe_peer;
1636 knlist_remove(&cpipe->pipe_sel.si_note, kn, 1);
1642 filt_piperead(struct knote *kn, long hint)
1644 struct pipe *rpipe = kn->kn_fp->f_data;
1645 struct pipe *wpipe = rpipe->pipe_peer;
1649 kn->kn_data = rpipe->pipe_buffer.cnt;
1650 if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
1651 kn->kn_data = rpipe->pipe_map.cnt;
1653 if ((rpipe->pipe_state & PIPE_EOF) ||
1654 wpipe->pipe_present != PIPE_ACTIVE ||
1655 (wpipe->pipe_state & PIPE_EOF)) {
1656 kn->kn_flags |= EV_EOF;
1660 ret = kn->kn_data > 0;
1667 filt_pipewrite(struct knote *kn, long hint)
1669 struct pipe *rpipe = kn->kn_fp->f_data;
1670 struct pipe *wpipe = rpipe->pipe_peer;
1673 if (wpipe->pipe_present != PIPE_ACTIVE ||
1674 (wpipe->pipe_state & PIPE_EOF)) {
1676 kn->kn_flags |= EV_EOF;
1680 kn->kn_data = (wpipe->pipe_buffer.size > 0) ?
1681 (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) : PIPE_BUF;
1682 if (wpipe->pipe_state & PIPE_DIRECTW)
1686 return (kn->kn_data >= PIPE_BUF);