Now available: The Design and Implementation of the FreeBSD Operating System (Second Edition)
[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]
FreeBSD/Linux Kernel Cross Reference
sys/kern/sys_pipe.c
Version:
- FREEBSD - FREEBSD-13-STABLE - FREEBSD-13-0 - FREEBSD-12-STABLE - FREEBSD-12-0 - FREEBSD-11-STABLE - FREEBSD-11-0 - FREEBSD-10-STABLE - FREEBSD-10-0 - FREEBSD-9-STABLE - FREEBSD-9-0 - FREEBSD-8-STABLE - FREEBSD-8-0 - FREEBSD-7-STABLE - FREEBSD-7-0 - FREEBSD-6-STABLE - FREEBSD-6-0 - FREEBSD-5-STABLE - FREEBSD-5-0 - FREEBSD-4-STABLE - FREEBSD-3-STABLE - FREEBSD22 - l41 - OPENBSD - linux-2.6 - MK84 - PLAN9 - xnu-8792
SearchContext: - none - 3 - 10
SearchContext: - none - 3 - 10
1 /*- 2 * SPDX-License-Identifier: BSD-4-Clause 3 * 4 * Copyright (c) 1996 John S. Dyson 5 * Copyright (c) 2012 Giovanni Trematerra 6 * All rights reserved. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice immediately at the beginning of the file, without modification, 13 * this list of conditions, and the following disclaimer. 14 * 2. Redistributions in binary form must reproduce the above copyright 15 * notice, this list of conditions and the following disclaimer in the 16 * documentation and/or other materials provided with the distribution. 17 * 3. Absolutely no warranty of function or purpose is made by the author 18 * John S. Dyson. 19 * 4. Modifications may be freely made to this file if the above conditions 20 * are met. 21 */ 22 23 /* 24 * This file contains a high-performance replacement for the socket-based 25 * pipes scheme originally used in FreeBSD/4.4Lite. It does not support 26 * all features of sockets, but does do everything that pipes normally 27 * do. 28 */ 29 30 /* 31 * This code has two modes of operation, a small write mode and a large 32 * write mode. The small write mode acts like conventional pipes with 33 * a kernel buffer. If the buffer is less than PIPE_MINDIRECT, then the 34 * "normal" pipe buffering is done. If the buffer is between PIPE_MINDIRECT 35 * and PIPE_SIZE in size, the sending process pins the underlying pages in 36 * memory, and the receiving process copies directly from these pinned pages 37 * in the sending process. 38 * 39 * If the sending process receives a signal, it is possible that it will 40 * go away, and certainly its address space can change, because control 41 * is returned back to the user-mode side. In that case, the pipe code 42 * arranges to copy the buffer supplied by the user process, to a pageable 43 * kernel buffer, and the receiving process will grab the data from the 44 * pageable kernel buffer. Since signals don't happen all that often, 45 * the copy operation is normally eliminated. 46 * 47 * The constant PIPE_MINDIRECT is chosen to make sure that buffering will 48 * happen for small transfers so that the system will not spend all of 49 * its time context switching. 50 * 51 * In order to limit the resource use of pipes, two sysctls exist: 52 * 53 * kern.ipc.maxpipekva - This is a hard limit on the amount of pageable 54 * address space available to us in pipe_map. This value is normally 55 * autotuned, but may also be loader tuned. 56 * 57 * kern.ipc.pipekva - This read-only sysctl tracks the current amount of 58 * memory in use by pipes. 59 * 60 * Based on how large pipekva is relative to maxpipekva, the following 61 * will happen: 62 * 63 * 0% - 50%: 64 * New pipes are given 16K of memory backing, pipes may dynamically 65 * grow to as large as 64K where needed. 66 * 50% - 75%: 67 * New pipes are given 4K (or PAGE_SIZE) of memory backing, 68 * existing pipes may NOT grow. 69 * 75% - 100%: 70 * New pipes are given 4K (or PAGE_SIZE) of memory backing, 71 * existing pipes will be shrunk down to 4K whenever possible. 72 * 73 * Resizing may be disabled by setting kern.ipc.piperesizeallowed=0. If 74 * that is set, the only resize that will occur is the 0 -> SMALL_PIPE_SIZE 75 * resize which MUST occur for reverse-direction pipes when they are 76 * first used. 77 * 78 * Additional information about the current state of pipes may be obtained 79 * from kern.ipc.pipes, kern.ipc.pipefragretry, kern.ipc.pipeallocfail, 80 * and kern.ipc.piperesizefail. 81 * 82 * Locking rules: There are two locks present here: A mutex, used via 83 * PIPE_LOCK, and a flag, used via pipelock(). All locking is done via 84 * the flag, as mutexes can not persist over uiomove. The mutex 85 * exists only to guard access to the flag, and is not in itself a 86 * locking mechanism. Also note that there is only a single mutex for 87 * both directions of a pipe. 88 * 89 * As pipelock() may have to sleep before it can acquire the flag, it 90 * is important to reread all data after a call to pipelock(); everything 91 * in the structure may have changed. 92 */ 93 94 #include <sys/cdefs.h> 95 __FBSDID("$FreeBSD$"); 96 97 #include <sys/param.h> 98 #include <sys/systm.h> 99 #include <sys/conf.h> 100 #include <sys/fcntl.h> 101 #include <sys/file.h> 102 #include <sys/filedesc.h> 103 #include <sys/filio.h> 104 #include <sys/kernel.h> 105 #include <sys/lock.h> 106 #include <sys/mutex.h> 107 #include <sys/ttycom.h> 108 #include <sys/stat.h> 109 #include <sys/malloc.h> 110 #include <sys/poll.h> 111 #include <sys/selinfo.h> 112 #include <sys/signalvar.h> 113 #include <sys/syscallsubr.h> 114 #include <sys/sysctl.h> 115 #include <sys/sysproto.h> 116 #include <sys/pipe.h> 117 #include <sys/proc.h> 118 #include <sys/vnode.h> 119 #include <sys/uio.h> 120 #include <sys/user.h> 121 #include <sys/event.h> 122 123 #include <security/mac/mac_framework.h> 124 125 #include <vm/vm.h> 126 #include <vm/vm_param.h> 127 #include <vm/vm_object.h> 128 #include <vm/vm_kern.h> 129 #include <vm/vm_extern.h> 130 #include <vm/pmap.h> 131 #include <vm/vm_map.h> 132 #include <vm/vm_page.h> 133 #include <vm/uma.h> 134 135 /* 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 138 * NetBSD or OpenBSD. 139 */ 140 /* #define PIPE_NODIRECT */ 141 142 #define PIPE_PEER(pipe) \ 143 (((pipe)->pipe_type & PIPE_TYPE_NAMED) ? (pipe) : ((pipe)->pipe_peer)) 144 145 /* 146 * interfaces to the outside world 147 */ 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; 158 static fo_fill_kinfo_t pipe_fill_kinfo; 159 160 struct fileops pipeops = { 161 .fo_read = pipe_read, 162 .fo_write = pipe_write, 163 .fo_truncate = pipe_truncate, 164 .fo_ioctl = pipe_ioctl, 165 .fo_poll = pipe_poll, 166 .fo_kqfilter = pipe_kqfilter, 167 .fo_stat = pipe_stat, 168 .fo_close = pipe_close, 169 .fo_chmod = pipe_chmod, 170 .fo_chown = pipe_chown, 171 .fo_sendfile = invfo_sendfile, 172 .fo_fill_kinfo = pipe_fill_kinfo, 173 .fo_flags = DFLAG_PASSABLE 174 }; 175 176 static void filt_pipedetach(struct knote *kn); 177 static void filt_pipedetach_notsup(struct knote *kn); 178 static int filt_pipenotsup(struct knote *kn, long hint); 179 static int filt_piperead(struct knote *kn, long hint); 180 static int filt_pipewrite(struct knote *kn, long hint); 181 182 static struct filterops pipe_nfiltops = { 183 .f_isfd = 1, 184 .f_detach = filt_pipedetach_notsup, 185 .f_event = filt_pipenotsup 186 }; 187 static struct filterops pipe_rfiltops = { 188 .f_isfd = 1, 189 .f_detach = filt_pipedetach, 190 .f_event = filt_piperead 191 }; 192 static struct filterops pipe_wfiltops = { 193 .f_isfd = 1, 194 .f_detach = filt_pipedetach, 195 .f_event = filt_pipewrite 196 }; 197 198 /* 199 * Default pipe buffer size(s), this can be kind-of large now because pipe 200 * space is pageable. The pipe code will try to maintain locality of 201 * reference for performance reasons, so small amounts of outstanding I/O 202 * will not wipe the cache. 203 */ 204 #define MINPIPESIZE (PIPE_SIZE/3) 205 #define MAXPIPESIZE (2*PIPE_SIZE/3) 206 207 static long amountpipekva; 208 static int pipefragretry; 209 static int pipeallocfail; 210 static int piperesizefail; 211 static int piperesizeallowed = 1; 212 static long pipe_mindirect = PIPE_MINDIRECT; 213 214 SYSCTL_LONG(_kern_ipc, OID_AUTO, maxpipekva, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, 215 &maxpipekva, 0, "Pipe KVA limit"); 216 SYSCTL_LONG(_kern_ipc, OID_AUTO, pipekva, CTLFLAG_RD, 217 &amountpipekva, 0, "Pipe KVA usage"); 218 SYSCTL_INT(_kern_ipc, OID_AUTO, pipefragretry, CTLFLAG_RD, 219 &pipefragretry, 0, "Pipe allocation retries due to fragmentation"); 220 SYSCTL_INT(_kern_ipc, OID_AUTO, pipeallocfail, CTLFLAG_RD, 221 &pipeallocfail, 0, "Pipe allocation failures"); 222 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizefail, CTLFLAG_RD, 223 &piperesizefail, 0, "Pipe resize failures"); 224 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizeallowed, CTLFLAG_RW, 225 &piperesizeallowed, 0, "Pipe resizing allowed"); 226 SYSCTL_LONG(_kern_ipc, OID_AUTO, pipe_mindirect, CTLFLAG_RWTUN, 227 &pipe_mindirect, 0, "Pipe minimum I/OO size for VM optimization"); 228 229 static void pipeinit(void *dummy __unused); 230 static void pipeclose(struct pipe *cpipe); 231 static void pipe_free_kmem(struct pipe *cpipe); 232 static int pipe_create(struct pipe *pipe, bool backing); 233 static int pipe_paircreate(struct thread *td, struct pipepair **p_pp); 234 static __inline int pipelock(struct pipe *cpipe, int catch); 235 static __inline void pipeunlock(struct pipe *cpipe); 236 static void pipe_timestamp(struct timespec *tsp); 237 #ifndef PIPE_NODIRECT 238 static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio); 239 static void pipe_destroy_write_buffer(struct pipe *wpipe); 240 static int pipe_direct_write(struct pipe *wpipe, struct uio *uio); 241 static void pipe_clone_write_buffer(struct pipe *wpipe); 242 #endif 243 static int pipespace(struct pipe *cpipe, int size); 244 static int pipespace_new(struct pipe *cpipe, int size); 245 246 static int pipe_zone_ctor(void *mem, int size, void *arg, int flags); 247 static int pipe_zone_init(void *mem, int size, int flags); 248 static void pipe_zone_fini(void *mem, int size); 249 250 static uma_zone_t pipe_zone; 251 static struct unrhdr64 pipeino_unr; 252 static dev_t pipedev_ino; 253 254 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL); 255 256 static void 257 pipeinit(void *dummy __unused) 258 { 259 260 pipe_zone = uma_zcreate("pipe", sizeof(struct pipepair), 261 pipe_zone_ctor, NULL, pipe_zone_init, pipe_zone_fini, 262 UMA_ALIGN_PTR, 0); 263 KASSERT(pipe_zone != NULL, ("pipe_zone not initialized")); 264 new_unrhdr64(&pipeino_unr, 1); 265 pipedev_ino = devfs_alloc_cdp_inode(); 266 KASSERT(pipedev_ino > 0, ("pipe dev inode not initialized")); 267 } 268 269 static int 270 pipe_zone_ctor(void *mem, int size, void *arg, int flags) 271 { 272 struct pipepair *pp; 273 struct pipe *rpipe, *wpipe; 274 275 KASSERT(size == sizeof(*pp), ("pipe_zone_ctor: wrong size")); 276 277 pp = (struct pipepair *)mem; 278 279 /* 280 * We zero both pipe endpoints to make sure all the kmem pointers 281 * are NULL, flag fields are zero'd, etc. We timestamp both 282 * endpoints with the same time. 283 */ 284 rpipe = &pp->pp_rpipe; 285 bzero(rpipe, sizeof(*rpipe)); 286 pipe_timestamp(&rpipe->pipe_ctime); 287 rpipe->pipe_atime = rpipe->pipe_mtime = rpipe->pipe_ctime; 288 289 wpipe = &pp->pp_wpipe; 290 bzero(wpipe, sizeof(*wpipe)); 291 wpipe->pipe_ctime = rpipe->pipe_ctime; 292 wpipe->pipe_atime = wpipe->pipe_mtime = rpipe->pipe_ctime; 293 294 rpipe->pipe_peer = wpipe; 295 rpipe->pipe_pair = pp; 296 wpipe->pipe_peer = rpipe; 297 wpipe->pipe_pair = pp; 298 299 /* 300 * Mark both endpoints as present; they will later get free'd 301 * one at a time. When both are free'd, then the whole pair 302 * is released. 303 */ 304 rpipe->pipe_present = PIPE_ACTIVE; 305 wpipe->pipe_present = PIPE_ACTIVE; 306 307 /* 308 * Eventually, the MAC Framework may initialize the label 309 * in ctor or init, but for now we do it elswhere to avoid 310 * blocking in ctor or init. 311 */ 312 pp->pp_label = NULL; 313 314 return (0); 315 } 316 317 static int 318 pipe_zone_init(void *mem, int size, int flags) 319 { 320 struct pipepair *pp; 321 322 KASSERT(size == sizeof(*pp), ("pipe_zone_init: wrong size")); 323 324 pp = (struct pipepair *)mem; 325 326 mtx_init(&pp->pp_mtx, "pipe mutex", NULL, MTX_DEF | MTX_NEW); 327 return (0); 328 } 329 330 static void 331 pipe_zone_fini(void *mem, int size) 332 { 333 struct pipepair *pp; 334 335 KASSERT(size == sizeof(*pp), ("pipe_zone_fini: wrong size")); 336 337 pp = (struct pipepair *)mem; 338 339 mtx_destroy(&pp->pp_mtx); 340 } 341 342 static int 343 pipe_paircreate(struct thread *td, struct pipepair **p_pp) 344 { 345 struct pipepair *pp; 346 struct pipe *rpipe, *wpipe; 347 int error; 348 349 *p_pp = pp = uma_zalloc(pipe_zone, M_WAITOK); 350 #ifdef MAC 351 /* 352 * The MAC label is shared between the connected endpoints. As a 353 * result mac_pipe_init() and mac_pipe_create() are called once 354 * for the pair, and not on the endpoints. 355 */ 356 mac_pipe_init(pp); 357 mac_pipe_create(td->td_ucred, pp); 358 #endif 359 rpipe = &pp->pp_rpipe; 360 wpipe = &pp->pp_wpipe; 361 362 knlist_init_mtx(&rpipe->pipe_sel.si_note, PIPE_MTX(rpipe)); 363 knlist_init_mtx(&wpipe->pipe_sel.si_note, PIPE_MTX(wpipe)); 364 365 /* 366 * Only the forward direction pipe is backed by big buffer by 367 * default. 368 */ 369 error = pipe_create(rpipe, true); 370 if (error != 0) 371 goto fail; 372 error = pipe_create(wpipe, false); 373 if (error != 0) { 374 /* 375 * This cleanup leaves the pipe inode number for rpipe 376 * still allocated, but never used. We do not free 377 * inode numbers for opened pipes, which is required 378 * for correctness because numbers must be unique. 379 * But also it avoids any memory use by the unr 380 * allocator, so stashing away the transient inode 381 * number is reasonable. 382 */ 383 pipe_free_kmem(rpipe); 384 goto fail; 385 } 386 387 rpipe->pipe_state |= PIPE_DIRECTOK; 388 wpipe->pipe_state |= PIPE_DIRECTOK; 389 return (0); 390 391 fail: 392 knlist_destroy(&rpipe->pipe_sel.si_note); 393 knlist_destroy(&wpipe->pipe_sel.si_note); 394 #ifdef MAC 395 mac_pipe_destroy(pp); 396 #endif 397 uma_zfree(pipe_zone, pp); 398 return (error); 399 } 400 401 int 402 pipe_named_ctor(struct pipe **ppipe, struct thread *td) 403 { 404 struct pipepair *pp; 405 int error; 406 407 error = pipe_paircreate(td, &pp); 408 if (error != 0) 409 return (error); 410 pp->pp_rpipe.pipe_type |= PIPE_TYPE_NAMED; 411 *ppipe = &pp->pp_rpipe; 412 return (0); 413 } 414 415 void 416 pipe_dtor(struct pipe *dpipe) 417 { 418 struct pipe *peer; 419 420 peer = (dpipe->pipe_type & PIPE_TYPE_NAMED) != 0 ? dpipe->pipe_peer : NULL; 421 funsetown(&dpipe->pipe_sigio); 422 pipeclose(dpipe); 423 if (peer != NULL) { 424 funsetown(&peer->pipe_sigio); 425 pipeclose(peer); 426 } 427 } 428 429 /* 430 * Get a timestamp. 431 * 432 * This used to be vfs_timestamp but the higher precision is unnecessary and 433 * can very negatively affect performance in virtualized environments (e.g., on 434 * vms running on amd64 when using the rdtscp instruction). 435 */ 436 static void 437 pipe_timestamp(struct timespec *tsp) 438 { 439 440 getnanotime(tsp); 441 } 442 443 /* 444 * The pipe system call for the DTYPE_PIPE type of pipes. If we fail, let 445 * the zone pick up the pieces via pipeclose(). 446 */ 447 int 448 kern_pipe(struct thread *td, int fildes[2], int flags, struct filecaps *fcaps1, 449 struct filecaps *fcaps2) 450 { 451 struct file *rf, *wf; 452 struct pipe *rpipe, *wpipe; 453 struct pipepair *pp; 454 int fd, fflags, error; 455 456 error = pipe_paircreate(td, &pp); 457 if (error != 0) 458 return (error); 459 rpipe = &pp->pp_rpipe; 460 wpipe = &pp->pp_wpipe; 461 error = falloc_caps(td, &rf, &fd, flags, fcaps1); 462 if (error) { 463 pipeclose(rpipe); 464 pipeclose(wpipe); 465 return (error); 466 } 467 /* An extra reference on `rf' has been held for us by falloc_caps(). */ 468 fildes[0] = fd; 469 470 fflags = FREAD | FWRITE; 471 if ((flags & O_NONBLOCK) != 0) 472 fflags |= FNONBLOCK; 473 474 /* 475 * Warning: once we've gotten past allocation of the fd for the 476 * read-side, we can only drop the read side via fdrop() in order 477 * to avoid races against processes which manage to dup() the read 478 * side while we are blocked trying to allocate the write side. 479 */ 480 finit(rf, fflags, DTYPE_PIPE, rpipe, &pipeops); 481 error = falloc_caps(td, &wf, &fd, flags, fcaps2); 482 if (error) { 483 fdclose(td, rf, fildes[0]); 484 fdrop(rf, td); 485 /* rpipe has been closed by fdrop(). */ 486 pipeclose(wpipe); 487 return (error); 488 } 489 /* An extra reference on `wf' has been held for us by falloc_caps(). */ 490 finit(wf, fflags, DTYPE_PIPE, wpipe, &pipeops); 491 fdrop(wf, td); 492 fildes[1] = fd; 493 fdrop(rf, td); 494 495 return (0); 496 } 497 498 #ifdef COMPAT_FREEBSD10 499 /* ARGSUSED */ 500 int 501 freebsd10_pipe(struct thread *td, struct freebsd10_pipe_args *uap __unused) 502 { 503 int error; 504 int fildes[2]; 505 506 error = kern_pipe(td, fildes, 0, NULL, NULL); 507 if (error) 508 return (error); 509 510 td->td_retval[0] = fildes[0]; 511 td->td_retval[1] = fildes[1]; 512 513 return (0); 514 } 515 #endif 516 517 int 518 sys_pipe2(struct thread *td, struct pipe2_args *uap) 519 { 520 int error, fildes[2]; 521 522 if (uap->flags & ~(O_CLOEXEC | O_NONBLOCK)) 523 return (EINVAL); 524 error = kern_pipe(td, fildes, uap->flags, NULL, NULL); 525 if (error) 526 return (error); 527 error = copyout(fildes, uap->fildes, 2 * sizeof(int)); 528 if (error) { 529 (void)kern_close(td, fildes[0]); 530 (void)kern_close(td, fildes[1]); 531 } 532 return (error); 533 } 534 535 /* 536 * Allocate kva for pipe circular buffer, the space is pageable 537 * This routine will 'realloc' the size of a pipe safely, if it fails 538 * it will retain the old buffer. 539 * If it fails it will return ENOMEM. 540 */ 541 static int 542 pipespace_new(struct pipe *cpipe, int size) 543 { 544 caddr_t buffer; 545 int error, cnt, firstseg; 546 static int curfail = 0; 547 static struct timeval lastfail; 548 549 KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked")); 550 KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW), 551 ("pipespace: resize of direct writes not allowed")); 552 retry: 553 cnt = cpipe->pipe_buffer.cnt; 554 if (cnt > size) 555 size = cnt; 556 557 size = round_page(size); 558 buffer = (caddr_t) vm_map_min(pipe_map); 559 560 error = vm_map_find(pipe_map, NULL, 0, (vm_offset_t *)&buffer, size, 0, 561 VMFS_ANY_SPACE, VM_PROT_RW, VM_PROT_RW, 0); 562 if (error != KERN_SUCCESS) { 563 if (cpipe->pipe_buffer.buffer == NULL && 564 size > SMALL_PIPE_SIZE) { 565 size = SMALL_PIPE_SIZE; 566 pipefragretry++; 567 goto retry; 568 } 569 if (cpipe->pipe_buffer.buffer == NULL) { 570 pipeallocfail++; 571 if (ppsratecheck(&lastfail, &curfail, 1)) 572 printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n"); 573 } else { 574 piperesizefail++; 575 } 576 return (ENOMEM); 577 } 578 579 /* copy data, then free old resources if we're resizing */ 580 if (cnt > 0) { 581 if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) { 582 firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out; 583 bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out], 584 buffer, firstseg); 585 if ((cnt - firstseg) > 0) 586 bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg], 587 cpipe->pipe_buffer.in); 588 } else { 589 bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out], 590 buffer, cnt); 591 } 592 } 593 pipe_free_kmem(cpipe); 594 cpipe->pipe_buffer.buffer = buffer; 595 cpipe->pipe_buffer.size = size; 596 cpipe->pipe_buffer.in = cnt; 597 cpipe->pipe_buffer.out = 0; 598 cpipe->pipe_buffer.cnt = cnt; 599 atomic_add_long(&amountpipekva, cpipe->pipe_buffer.size); 600 return (0); 601 } 602 603 /* 604 * Wrapper for pipespace_new() that performs locking assertions. 605 */ 606 static int 607 pipespace(struct pipe *cpipe, int size) 608 { 609 610 KASSERT(cpipe->pipe_state & PIPE_LOCKFL, 611 ("Unlocked pipe passed to pipespace")); 612 return (pipespace_new(cpipe, size)); 613 } 614 615 /* 616 * lock a pipe for I/O, blocking other access 617 */ 618 static __inline int 619 pipelock(struct pipe *cpipe, int catch) 620 { 621 int error, prio; 622 623 PIPE_LOCK_ASSERT(cpipe, MA_OWNED); 624 625 prio = PRIBIO; 626 if (catch) 627 prio |= PCATCH; 628 while (cpipe->pipe_state & PIPE_LOCKFL) { 629 KASSERT(cpipe->pipe_waiters >= 0, 630 ("%s: bad waiter count %d", __func__, 631 cpipe->pipe_waiters)); 632 cpipe->pipe_waiters++; 633 error = msleep(cpipe, PIPE_MTX(cpipe), 634 prio, "pipelk", 0); 635 cpipe->pipe_waiters--; 636 if (error != 0) 637 return (error); 638 } 639 cpipe->pipe_state |= PIPE_LOCKFL; 640 return (0); 641 } 642 643 /* 644 * unlock a pipe I/O lock 645 */ 646 static __inline void 647 pipeunlock(struct pipe *cpipe) 648 { 649 650 PIPE_LOCK_ASSERT(cpipe, MA_OWNED); 651 KASSERT(cpipe->pipe_state & PIPE_LOCKFL, 652 ("Unlocked pipe passed to pipeunlock")); 653 KASSERT(cpipe->pipe_waiters >= 0, 654 ("%s: bad waiter count %d", __func__, 655 cpipe->pipe_waiters)); 656 cpipe->pipe_state &= ~PIPE_LOCKFL; 657 if (cpipe->pipe_waiters > 0) { 658 wakeup_one(cpipe); 659 } 660 } 661 662 void 663 pipeselwakeup(struct pipe *cpipe) 664 { 665 666 PIPE_LOCK_ASSERT(cpipe, MA_OWNED); 667 if (cpipe->pipe_state & PIPE_SEL) { 668 selwakeuppri(&cpipe->pipe_sel, PSOCK); 669 if (!SEL_WAITING(&cpipe->pipe_sel)) 670 cpipe->pipe_state &= ~PIPE_SEL; 671 } 672 if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio) 673 pgsigio(&cpipe->pipe_sigio, SIGIO, 0); 674 KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0); 675 } 676 677 /* 678 * Initialize and allocate VM and memory for pipe. The structure 679 * will start out zero'd from the ctor, so we just manage the kmem. 680 */ 681 static int 682 pipe_create(struct pipe *pipe, bool large_backing) 683 { 684 int error; 685 686 error = pipespace_new(pipe, !large_backing || amountpipekva > 687 maxpipekva / 2 ? SMALL_PIPE_SIZE : PIPE_SIZE); 688 if (error == 0) 689 pipe->pipe_ino = alloc_unr64(&pipeino_unr); 690 return (error); 691 } 692 693 /* ARGSUSED */ 694 static int 695 pipe_read(struct file *fp, struct uio *uio, struct ucred *active_cred, 696 int flags, struct thread *td) 697 { 698 struct pipe *rpipe; 699 int error; 700 int nread = 0; 701 int size; 702 703 rpipe = fp->f_data; 704 PIPE_LOCK(rpipe); 705 ++rpipe->pipe_busy; 706 error = pipelock(rpipe, 1); 707 if (error) 708 goto unlocked_error; 709 710 #ifdef MAC 711 error = mac_pipe_check_read(active_cred, rpipe->pipe_pair); 712 if (error) 713 goto locked_error; 714 #endif 715 if (amountpipekva > (3 * maxpipekva) / 4) { 716 if ((rpipe->pipe_state & PIPE_DIRECTW) == 0 && 717 rpipe->pipe_buffer.size > SMALL_PIPE_SIZE && 718 rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE && 719 piperesizeallowed == 1) { 720 PIPE_UNLOCK(rpipe); 721 pipespace(rpipe, SMALL_PIPE_SIZE); 722 PIPE_LOCK(rpipe); 723 } 724 } 725 726 while (uio->uio_resid) { 727 /* 728 * normal pipe buffer receive 729 */ 730 if (rpipe->pipe_buffer.cnt > 0) { 731 size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out; 732 if (size > rpipe->pipe_buffer.cnt) 733 size = rpipe->pipe_buffer.cnt; 734 if (size > uio->uio_resid) 735 size = uio->uio_resid; 736 737 PIPE_UNLOCK(rpipe); 738 error = uiomove( 739 &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out], 740 size, uio); 741 PIPE_LOCK(rpipe); 742 if (error) 743 break; 744 745 rpipe->pipe_buffer.out += size; 746 if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size) 747 rpipe->pipe_buffer.out = 0; 748 749 rpipe->pipe_buffer.cnt -= size; 750 751 /* 752 * If there is no more to read in the pipe, reset 753 * its pointers to the beginning. This improves 754 * cache hit stats. 755 */ 756 if (rpipe->pipe_buffer.cnt == 0) { 757 rpipe->pipe_buffer.in = 0; 758 rpipe->pipe_buffer.out = 0; 759 } 760 nread += size; 761 #ifndef PIPE_NODIRECT 762 /* 763 * Direct copy, bypassing a kernel buffer. 764 */ 765 } else if ((size = rpipe->pipe_pages.cnt) != 0) { 766 if (size > uio->uio_resid) 767 size = (u_int) uio->uio_resid; 768 PIPE_UNLOCK(rpipe); 769 error = uiomove_fromphys(rpipe->pipe_pages.ms, 770 rpipe->pipe_pages.pos, size, uio); 771 PIPE_LOCK(rpipe); 772 if (error) 773 break; 774 nread += size; 775 rpipe->pipe_pages.pos += size; 776 rpipe->pipe_pages.cnt -= size; 777 if (rpipe->pipe_pages.cnt == 0) { 778 rpipe->pipe_state &= ~PIPE_WANTW; 779 wakeup(rpipe); 780 } 781 #endif 782 } else { 783 /* 784 * detect EOF condition 785 * read returns 0 on EOF, no need to set error 786 */ 787 if (rpipe->pipe_state & PIPE_EOF) 788 break; 789 790 /* 791 * If the "write-side" has been blocked, wake it up now. 792 */ 793 if (rpipe->pipe_state & PIPE_WANTW) { 794 rpipe->pipe_state &= ~PIPE_WANTW; 795 wakeup(rpipe); 796 } 797 798 /* 799 * Break if some data was read. 800 */ 801 if (nread > 0) 802 break; 803 804 /* 805 * Unlock the pipe buffer for our remaining processing. 806 * We will either break out with an error or we will 807 * sleep and relock to loop. 808 */ 809 pipeunlock(rpipe); 810 811 /* 812 * Handle non-blocking mode operation or 813 * wait for more data. 814 */ 815 if (fp->f_flag & FNONBLOCK) { 816 error = EAGAIN; 817 } else { 818 rpipe->pipe_state |= PIPE_WANTR; 819 if ((error = msleep(rpipe, PIPE_MTX(rpipe), 820 PRIBIO | PCATCH, 821 "piperd", 0)) == 0) 822 error = pipelock(rpipe, 1); 823 } 824 if (error) 825 goto unlocked_error; 826 } 827 } 828 #ifdef MAC 829 locked_error: 830 #endif 831 pipeunlock(rpipe); 832 833 /* XXX: should probably do this before getting any locks. */ 834 if (error == 0) 835 pipe_timestamp(&rpipe->pipe_atime); 836 unlocked_error: 837 --rpipe->pipe_busy; 838 839 /* 840 * PIPE_WANT processing only makes sense if pipe_busy is 0. 841 */ 842 if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) { 843 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW); 844 wakeup(rpipe); 845 } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) { 846 /* 847 * Handle write blocking hysteresis. 848 */ 849 if (rpipe->pipe_state & PIPE_WANTW) { 850 rpipe->pipe_state &= ~PIPE_WANTW; 851 wakeup(rpipe); 852 } 853 } 854 855 /* 856 * Only wake up writers if there was actually something read. 857 * Otherwise, when calling read(2) at EOF, a spurious wakeup occurs. 858 */ 859 if (nread > 0 && 860 rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt >= PIPE_BUF) 861 pipeselwakeup(rpipe); 862 863 PIPE_UNLOCK(rpipe); 864 if (nread > 0) 865 td->td_ru.ru_msgrcv++; 866 return (error); 867 } 868 869 #ifndef PIPE_NODIRECT 870 /* 871 * Map the sending processes' buffer into kernel space and wire it. 872 * This is similar to a physical write operation. 873 */ 874 static int 875 pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio) 876 { 877 u_int size; 878 int i; 879 880 PIPE_LOCK_ASSERT(wpipe, MA_OWNED); 881 KASSERT((wpipe->pipe_state & PIPE_DIRECTW) == 0, 882 ("%s: PIPE_DIRECTW set on %p", __func__, wpipe)); 883 KASSERT(wpipe->pipe_pages.cnt == 0, 884 ("%s: pipe map for %p contains residual data", __func__, wpipe)); 885 886 if (uio->uio_iov->iov_len > wpipe->pipe_buffer.size) 887 size = wpipe->pipe_buffer.size; 888 else 889 size = uio->uio_iov->iov_len; 890 891 wpipe->pipe_state |= PIPE_DIRECTW; 892 PIPE_UNLOCK(wpipe); 893 i = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map, 894 (vm_offset_t)uio->uio_iov->iov_base, size, VM_PROT_READ, 895 wpipe->pipe_pages.ms, PIPENPAGES); 896 PIPE_LOCK(wpipe); 897 if (i < 0) { 898 wpipe->pipe_state &= ~PIPE_DIRECTW; 899 return (EFAULT); 900 } 901 902 wpipe->pipe_pages.npages = i; 903 wpipe->pipe_pages.pos = 904 ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK; 905 wpipe->pipe_pages.cnt = size; 906 907 uio->uio_iov->iov_len -= size; 908 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size; 909 if (uio->uio_iov->iov_len == 0) 910 uio->uio_iov++; 911 uio->uio_resid -= size; 912 uio->uio_offset += size; 913 return (0); 914 } 915 916 /* 917 * Unwire the process buffer. 918 */ 919 static void 920 pipe_destroy_write_buffer(struct pipe *wpipe) 921 { 922 923 PIPE_LOCK_ASSERT(wpipe, MA_OWNED); 924 KASSERT((wpipe->pipe_state & PIPE_DIRECTW) != 0, 925 ("%s: PIPE_DIRECTW not set on %p", __func__, wpipe)); 926 KASSERT(wpipe->pipe_pages.cnt == 0, 927 ("%s: pipe map for %p contains residual data", __func__, wpipe)); 928 929 wpipe->pipe_state &= ~PIPE_DIRECTW; 930 vm_page_unhold_pages(wpipe->pipe_pages.ms, wpipe->pipe_pages.npages); 931 wpipe->pipe_pages.npages = 0; 932 } 933 934 /* 935 * In the case of a signal, the writing process might go away. This 936 * code copies the data into the circular buffer so that the source 937 * pages can be freed without loss of data. 938 */ 939 static void 940 pipe_clone_write_buffer(struct pipe *wpipe) 941 { 942 struct uio uio; 943 struct iovec iov; 944 int size; 945 int pos; 946 947 PIPE_LOCK_ASSERT(wpipe, MA_OWNED); 948 KASSERT((wpipe->pipe_state & PIPE_DIRECTW) != 0, 949 ("%s: PIPE_DIRECTW not set on %p", __func__, wpipe)); 950 951 size = wpipe->pipe_pages.cnt; 952 pos = wpipe->pipe_pages.pos; 953 wpipe->pipe_pages.cnt = 0; 954 955 wpipe->pipe_buffer.in = size; 956 wpipe->pipe_buffer.out = 0; 957 wpipe->pipe_buffer.cnt = size; 958 959 PIPE_UNLOCK(wpipe); 960 iov.iov_base = wpipe->pipe_buffer.buffer; 961 iov.iov_len = size; 962 uio.uio_iov = &iov; 963 uio.uio_iovcnt = 1; 964 uio.uio_offset = 0; 965 uio.uio_resid = size; 966 uio.uio_segflg = UIO_SYSSPACE; 967 uio.uio_rw = UIO_READ; 968 uio.uio_td = curthread; 969 uiomove_fromphys(wpipe->pipe_pages.ms, pos, size, &uio); 970 PIPE_LOCK(wpipe); 971 pipe_destroy_write_buffer(wpipe); 972 } 973 974 /* 975 * This implements the pipe buffer write mechanism. Note that only 976 * a direct write OR a normal pipe write can be pending at any given time. 977 * If there are any characters in the pipe buffer, the direct write will 978 * be deferred until the receiving process grabs all of the bytes from 979 * the pipe buffer. Then the direct mapping write is set-up. 980 */ 981 static int 982 pipe_direct_write(struct pipe *wpipe, struct uio *uio) 983 { 984 int error; 985 986 retry: 987 PIPE_LOCK_ASSERT(wpipe, MA_OWNED); 988 if ((wpipe->pipe_state & PIPE_EOF) != 0) { 989 error = EPIPE; 990 goto error1; 991 } 992 if (wpipe->pipe_state & PIPE_DIRECTW) { 993 if (wpipe->pipe_state & PIPE_WANTR) { 994 wpipe->pipe_state &= ~PIPE_WANTR; 995 wakeup(wpipe); 996 } 997 pipeselwakeup(wpipe); 998 wpipe->pipe_state |= PIPE_WANTW; 999 pipeunlock(wpipe); 1000 error = msleep(wpipe, PIPE_MTX(wpipe), 1001 PRIBIO | PCATCH, "pipdww", 0); 1002 pipelock(wpipe, 0); 1003 if (error != 0) 1004 goto error1; 1005 goto retry; 1006 } 1007 if (wpipe->pipe_buffer.cnt > 0) { 1008 if (wpipe->pipe_state & PIPE_WANTR) { 1009 wpipe->pipe_state &= ~PIPE_WANTR; 1010 wakeup(wpipe); 1011 } 1012 pipeselwakeup(wpipe); 1013 wpipe->pipe_state |= PIPE_WANTW; 1014 pipeunlock(wpipe); 1015 error = msleep(wpipe, PIPE_MTX(wpipe), 1016 PRIBIO | PCATCH, "pipdwc", 0); 1017 pipelock(wpipe, 0); 1018 if (error != 0) 1019 goto error1; 1020 goto retry; 1021 } 1022 1023 error = pipe_build_write_buffer(wpipe, uio); 1024 if (error) { 1025 goto error1; 1026 } 1027 1028 while (wpipe->pipe_pages.cnt != 0 && 1029 (wpipe->pipe_state & PIPE_EOF) == 0) { 1030 if (wpipe->pipe_state & PIPE_WANTR) { 1031 wpipe->pipe_state &= ~PIPE_WANTR; 1032 wakeup(wpipe); 1033 } 1034 pipeselwakeup(wpipe); 1035 wpipe->pipe_state |= PIPE_WANTW; 1036 pipeunlock(wpipe); 1037 error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH, 1038 "pipdwt", 0); 1039 pipelock(wpipe, 0); 1040 if (error != 0) 1041 break; 1042 } 1043 1044 if ((wpipe->pipe_state & PIPE_EOF) != 0) { 1045 wpipe->pipe_pages.cnt = 0; 1046 pipe_destroy_write_buffer(wpipe); 1047 pipeselwakeup(wpipe); 1048 error = EPIPE; 1049 } else if (error == EINTR || error == ERESTART) { 1050 pipe_clone_write_buffer(wpipe); 1051 } else { 1052 pipe_destroy_write_buffer(wpipe); 1053 } 1054 KASSERT((wpipe->pipe_state & PIPE_DIRECTW) == 0, 1055 ("pipe %p leaked PIPE_DIRECTW", wpipe)); 1056 return (error); 1057 1058 error1: 1059 wakeup(wpipe); 1060 return (error); 1061 } 1062 #endif 1063 1064 static int 1065 pipe_write(struct file *fp, struct uio *uio, struct ucred *active_cred, 1066 int flags, struct thread *td) 1067 { 1068 struct pipe *wpipe, *rpipe; 1069 ssize_t orig_resid; 1070 int desiredsize, error; 1071 1072 rpipe = fp->f_data; 1073 wpipe = PIPE_PEER(rpipe); 1074 PIPE_LOCK(rpipe); 1075 error = pipelock(wpipe, 1); 1076 if (error) { 1077 PIPE_UNLOCK(rpipe); 1078 return (error); 1079 } 1080 /* 1081 * detect loss of pipe read side, issue SIGPIPE if lost. 1082 */ 1083 if (wpipe->pipe_present != PIPE_ACTIVE || 1084 (wpipe->pipe_state & PIPE_EOF)) { 1085 pipeunlock(wpipe); 1086 PIPE_UNLOCK(rpipe); 1087 return (EPIPE); 1088 } 1089 #ifdef MAC 1090 error = mac_pipe_check_write(active_cred, wpipe->pipe_pair); 1091 if (error) { 1092 pipeunlock(wpipe); 1093 PIPE_UNLOCK(rpipe); 1094 return (error); 1095 } 1096 #endif 1097 ++wpipe->pipe_busy; 1098 1099 /* Choose a larger size if it's advantageous */ 1100 desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size); 1101 while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) { 1102 if (piperesizeallowed != 1) 1103 break; 1104 if (amountpipekva > maxpipekva / 2) 1105 break; 1106 if (desiredsize == BIG_PIPE_SIZE) 1107 break; 1108 desiredsize = desiredsize * 2; 1109 } 1110 1111 /* Choose a smaller size if we're in a OOM situation */ 1112 if (amountpipekva > (3 * maxpipekva) / 4 && 1113 wpipe->pipe_buffer.size > SMALL_PIPE_SIZE && 1114 wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE && 1115 piperesizeallowed == 1) 1116 desiredsize = SMALL_PIPE_SIZE; 1117 1118 /* Resize if the above determined that a new size was necessary */ 1119 if (desiredsize != wpipe->pipe_buffer.size && 1120 (wpipe->pipe_state & PIPE_DIRECTW) == 0) { 1121 PIPE_UNLOCK(wpipe); 1122 pipespace(wpipe, desiredsize); 1123 PIPE_LOCK(wpipe); 1124 } 1125 MPASS(wpipe->pipe_buffer.size != 0); 1126 1127 orig_resid = uio->uio_resid; 1128 1129 while (uio->uio_resid) { 1130 int space; 1131 1132 if (wpipe->pipe_state & PIPE_EOF) { 1133 error = EPIPE; 1134 break; 1135 } 1136 #ifndef PIPE_NODIRECT 1137 /* 1138 * If the transfer is large, we can gain performance if 1139 * we do process-to-process copies directly. 1140 * If the write is non-blocking, we don't use the 1141 * direct write mechanism. 1142 * 1143 * The direct write mechanism will detect the reader going 1144 * away on us. 1145 */ 1146 if (uio->uio_segflg == UIO_USERSPACE && 1147 uio->uio_iov->iov_len >= pipe_mindirect && 1148 wpipe->pipe_buffer.size >= pipe_mindirect && 1149 (fp->f_flag & FNONBLOCK) == 0) { 1150 error = pipe_direct_write(wpipe, uio); 1151 if (error != 0) 1152 break; 1153 continue; 1154 } 1155 #endif 1156 1157 /* 1158 * Pipe buffered writes cannot be coincidental with 1159 * direct writes. We wait until the currently executing 1160 * direct write is completed before we start filling the 1161 * pipe buffer. We break out if a signal occurs or the 1162 * reader goes away. 1163 */ 1164 if (wpipe->pipe_pages.cnt != 0) { 1165 if (wpipe->pipe_state & PIPE_WANTR) { 1166 wpipe->pipe_state &= ~PIPE_WANTR; 1167 wakeup(wpipe); 1168 } 1169 pipeselwakeup(wpipe); 1170 wpipe->pipe_state |= PIPE_WANTW; 1171 pipeunlock(wpipe); 1172 error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH, 1173 "pipbww", 0); 1174 pipelock(wpipe, 0); 1175 if (error != 0) 1176 break; 1177 continue; 1178 } 1179 1180 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt; 1181 1182 /* Writes of size <= PIPE_BUF must be atomic. */ 1183 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF)) 1184 space = 0; 1185 1186 if (space > 0) { 1187 int size; /* Transfer size */ 1188 int segsize; /* first segment to transfer */ 1189 1190 /* 1191 * Transfer size is minimum of uio transfer 1192 * and free space in pipe buffer. 1193 */ 1194 if (space > uio->uio_resid) 1195 size = uio->uio_resid; 1196 else 1197 size = space; 1198 /* 1199 * First segment to transfer is minimum of 1200 * transfer size and contiguous space in 1201 * pipe buffer. If first segment to transfer 1202 * is less than the transfer size, we've got 1203 * a wraparound in the buffer. 1204 */ 1205 segsize = wpipe->pipe_buffer.size - 1206 wpipe->pipe_buffer.in; 1207 if (segsize > size) 1208 segsize = size; 1209 1210 /* Transfer first segment */ 1211 1212 PIPE_UNLOCK(rpipe); 1213 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in], 1214 segsize, uio); 1215 PIPE_LOCK(rpipe); 1216 1217 if (error == 0 && segsize < size) { 1218 KASSERT(wpipe->pipe_buffer.in + segsize == 1219 wpipe->pipe_buffer.size, 1220 ("Pipe buffer wraparound disappeared")); 1221 /* 1222 * Transfer remaining part now, to 1223 * support atomic writes. Wraparound 1224 * happened. 1225 */ 1226 1227 PIPE_UNLOCK(rpipe); 1228 error = uiomove( 1229 &wpipe->pipe_buffer.buffer[0], 1230 size - segsize, uio); 1231 PIPE_LOCK(rpipe); 1232 } 1233 if (error == 0) { 1234 wpipe->pipe_buffer.in += size; 1235 if (wpipe->pipe_buffer.in >= 1236 wpipe->pipe_buffer.size) { 1237 KASSERT(wpipe->pipe_buffer.in == 1238 size - segsize + 1239 wpipe->pipe_buffer.size, 1240 ("Expected wraparound bad")); 1241 wpipe->pipe_buffer.in = size - segsize; 1242 } 1243 1244 wpipe->pipe_buffer.cnt += size; 1245 KASSERT(wpipe->pipe_buffer.cnt <= 1246 wpipe->pipe_buffer.size, 1247 ("Pipe buffer overflow")); 1248 } 1249 if (error != 0) 1250 break; 1251 continue; 1252 } else { 1253 /* 1254 * If the "read-side" has been blocked, wake it up now. 1255 */ 1256 if (wpipe->pipe_state & PIPE_WANTR) { 1257 wpipe->pipe_state &= ~PIPE_WANTR; 1258 wakeup(wpipe); 1259 } 1260 1261 /* 1262 * don't block on non-blocking I/O 1263 */ 1264 if (fp->f_flag & FNONBLOCK) { 1265 error = EAGAIN; 1266 break; 1267 } 1268 1269 /* 1270 * We have no more space and have something to offer, 1271 * wake up select/poll. 1272 */ 1273 pipeselwakeup(wpipe); 1274 1275 wpipe->pipe_state |= PIPE_WANTW; 1276 pipeunlock(wpipe); 1277 error = msleep(wpipe, PIPE_MTX(rpipe), 1278 PRIBIO | PCATCH, "pipewr", 0); 1279 pipelock(wpipe, 0); 1280 if (error != 0) 1281 break; 1282 continue; 1283 } 1284 } 1285 1286 --wpipe->pipe_busy; 1287 1288 if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) { 1289 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR); 1290 wakeup(wpipe); 1291 } else if (wpipe->pipe_buffer.cnt > 0) { 1292 /* 1293 * If we have put any characters in the buffer, we wake up 1294 * the reader. 1295 */ 1296 if (wpipe->pipe_state & PIPE_WANTR) { 1297 wpipe->pipe_state &= ~PIPE_WANTR; 1298 wakeup(wpipe); 1299 } 1300 } 1301 1302 /* 1303 * Don't return EPIPE if any byte was written. 1304 * EINTR and other interrupts are handled by generic I/O layer. 1305 * Do not pretend that I/O succeeded for obvious user error 1306 * like EFAULT. 1307 */ 1308 if (uio->uio_resid != orig_resid && error == EPIPE) 1309 error = 0; 1310 1311 if (error == 0) 1312 pipe_timestamp(&wpipe->pipe_mtime); 1313 1314 /* 1315 * We have something to offer, 1316 * wake up select/poll. 1317 */ 1318 if (wpipe->pipe_buffer.cnt) 1319 pipeselwakeup(wpipe); 1320 1321 pipeunlock(wpipe); 1322 PIPE_UNLOCK(rpipe); 1323 if (uio->uio_resid != orig_resid) 1324 td->td_ru.ru_msgsnd++; 1325 return (error); 1326 } 1327 1328 /* ARGSUSED */ 1329 static int 1330 pipe_truncate(struct file *fp, off_t length, struct ucred *active_cred, 1331 struct thread *td) 1332 { 1333 struct pipe *cpipe; 1334 int error; 1335 1336 cpipe = fp->f_data; 1337 if (cpipe->pipe_type & PIPE_TYPE_NAMED) 1338 error = vnops.fo_truncate(fp, length, active_cred, td); 1339 else 1340 error = invfo_truncate(fp, length, active_cred, td); 1341 return (error); 1342 } 1343 1344 /* 1345 * we implement a very minimal set of ioctls for compatibility with sockets. 1346 */ 1347 static int 1348 pipe_ioctl(struct file *fp, u_long cmd, void *data, struct ucred *active_cred, 1349 struct thread *td) 1350 { 1351 struct pipe *mpipe = fp->f_data; 1352 int error; 1353 1354 PIPE_LOCK(mpipe); 1355 1356 #ifdef MAC 1357 error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data); 1358 if (error) { 1359 PIPE_UNLOCK(mpipe); 1360 return (error); 1361 } 1362 #endif 1363 1364 error = 0; 1365 switch (cmd) { 1366 case FIONBIO: 1367 break; 1368 1369 case FIOASYNC: 1370 if (*(int *)data) { 1371 mpipe->pipe_state |= PIPE_ASYNC; 1372 } else { 1373 mpipe->pipe_state &= ~PIPE_ASYNC; 1374 } 1375 break; 1376 1377 case FIONREAD: 1378 if (!(fp->f_flag & FREAD)) { 1379 *(int *)data = 0; 1380 PIPE_UNLOCK(mpipe); 1381 return (0); 1382 } 1383 if (mpipe->pipe_pages.cnt != 0) 1384 *(int *)data = mpipe->pipe_pages.cnt; 1385 else 1386 *(int *)data = mpipe->pipe_buffer.cnt; 1387 break; 1388 1389 case FIOSETOWN: 1390 PIPE_UNLOCK(mpipe); 1391 error = fsetown(*(int *)data, &mpipe->pipe_sigio); 1392 goto out_unlocked; 1393 1394 case FIOGETOWN: 1395 *(int *)data = fgetown(&mpipe->pipe_sigio); 1396 break; 1397 1398 /* This is deprecated, FIOSETOWN should be used instead. */ 1399 case TIOCSPGRP: 1400 PIPE_UNLOCK(mpipe); 1401 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio); 1402 goto out_unlocked; 1403 1404 /* This is deprecated, FIOGETOWN should be used instead. */ 1405 case TIOCGPGRP: 1406 *(int *)data = -fgetown(&mpipe->pipe_sigio); 1407 break; 1408 1409 default: 1410 error = ENOTTY; 1411 break; 1412 } 1413 PIPE_UNLOCK(mpipe); 1414 out_unlocked: 1415 return (error); 1416 } 1417 1418 static int 1419 pipe_poll(struct file *fp, int events, struct ucred *active_cred, 1420 struct thread *td) 1421 { 1422 struct pipe *rpipe; 1423 struct pipe *wpipe; 1424 int levents, revents; 1425 #ifdef MAC 1426 int error; 1427 #endif 1428 1429 revents = 0; 1430 rpipe = fp->f_data; 1431 wpipe = PIPE_PEER(rpipe); 1432 PIPE_LOCK(rpipe); 1433 #ifdef MAC 1434 error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair); 1435 if (error) 1436 goto locked_error; 1437 #endif 1438 if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM)) 1439 if (rpipe->pipe_pages.cnt > 0 || rpipe->pipe_buffer.cnt > 0) 1440 revents |= events & (POLLIN | POLLRDNORM); 1441 1442 if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM)) 1443 if (wpipe->pipe_present != PIPE_ACTIVE || 1444 (wpipe->pipe_state & PIPE_EOF) || 1445 ((wpipe->pipe_state & PIPE_DIRECTW) == 0 && 1446 ((wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF || 1447 wpipe->pipe_buffer.size == 0))) 1448 revents |= events & (POLLOUT | POLLWRNORM); 1449 1450 levents = events & 1451 (POLLIN | POLLINIGNEOF | POLLPRI | POLLRDNORM | POLLRDBAND); 1452 if (rpipe->pipe_type & PIPE_TYPE_NAMED && fp->f_flag & FREAD && levents && 1453 fp->f_pipegen == rpipe->pipe_wgen) 1454 events |= POLLINIGNEOF; 1455 1456 if ((events & POLLINIGNEOF) == 0) { 1457 if (rpipe->pipe_state & PIPE_EOF) { 1458 if (fp->f_flag & FREAD) 1459 revents |= (events & (POLLIN | POLLRDNORM)); 1460 if (wpipe->pipe_present != PIPE_ACTIVE || 1461 (wpipe->pipe_state & PIPE_EOF)) 1462 revents |= POLLHUP; 1463 } 1464 } 1465 1466 if (revents == 0) { 1467 /* 1468 * Add ourselves regardless of eventmask as we have to return 1469 * POLLHUP even if it was not asked for. 1470 */ 1471 if ((fp->f_flag & FREAD) != 0) { 1472 selrecord(td, &rpipe->pipe_sel); 1473 if (SEL_WAITING(&rpipe->pipe_sel)) 1474 rpipe->pipe_state |= PIPE_SEL; 1475 } 1476 1477 if ((fp->f_flag & FWRITE) != 0 && 1478 wpipe->pipe_present == PIPE_ACTIVE) { 1479 selrecord(td, &wpipe->pipe_sel); 1480 if (SEL_WAITING(&wpipe->pipe_sel)) 1481 wpipe->pipe_state |= PIPE_SEL; 1482 } 1483 } 1484 #ifdef MAC 1485 locked_error: 1486 #endif 1487 PIPE_UNLOCK(rpipe); 1488 1489 return (revents); 1490 } 1491 1492 /* 1493 * We shouldn't need locks here as we're doing a read and this should 1494 * be a natural race. 1495 */ 1496 static int 1497 pipe_stat(struct file *fp, struct stat *ub, struct ucred *active_cred, 1498 struct thread *td) 1499 { 1500 struct pipe *pipe; 1501 #ifdef MAC 1502 int error; 1503 #endif 1504 1505 pipe = fp->f_data; 1506 #ifdef MAC 1507 if (mac_pipe_check_stat_enabled()) { 1508 PIPE_LOCK(pipe); 1509 error = mac_pipe_check_stat(active_cred, pipe->pipe_pair); 1510 PIPE_UNLOCK(pipe); 1511 if (error) { 1512 return (error); 1513 } 1514 } 1515 #endif 1516 1517 /* For named pipes ask the underlying filesystem. */ 1518 if (pipe->pipe_type & PIPE_TYPE_NAMED) { 1519 return (vnops.fo_stat(fp, ub, active_cred, td)); 1520 } 1521 1522 bzero(ub, sizeof(*ub)); 1523 ub->st_mode = S_IFIFO; 1524 ub->st_blksize = PAGE_SIZE; 1525 if (pipe->pipe_pages.cnt != 0) 1526 ub->st_size = pipe->pipe_pages.cnt; 1527 else 1528 ub->st_size = pipe->pipe_buffer.cnt; 1529 ub->st_blocks = howmany(ub->st_size, ub->st_blksize); 1530 ub->st_atim = pipe->pipe_atime; 1531 ub->st_mtim = pipe->pipe_mtime; 1532 ub->st_ctim = pipe->pipe_ctime; 1533 ub->st_uid = fp->f_cred->cr_uid; 1534 ub->st_gid = fp->f_cred->cr_gid; 1535 ub->st_dev = pipedev_ino; 1536 ub->st_ino = pipe->pipe_ino; 1537 /* 1538 * Left as 0: st_nlink, st_rdev, st_flags, st_gen. 1539 */ 1540 return (0); 1541 } 1542 1543 /* ARGSUSED */ 1544 static int 1545 pipe_close(struct file *fp, struct thread *td) 1546 { 1547 1548 if (fp->f_vnode != NULL) 1549 return vnops.fo_close(fp, td); 1550 fp->f_ops = &badfileops; 1551 pipe_dtor(fp->f_data); 1552 fp->f_data = NULL; 1553 return (0); 1554 } 1555 1556 static int 1557 pipe_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td) 1558 { 1559 struct pipe *cpipe; 1560 int error; 1561 1562 cpipe = fp->f_data; 1563 if (cpipe->pipe_type & PIPE_TYPE_NAMED) 1564 error = vn_chmod(fp, mode, active_cred, td); 1565 else 1566 error = invfo_chmod(fp, mode, active_cred, td); 1567 return (error); 1568 } 1569 1570 static int 1571 pipe_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred, 1572 struct thread *td) 1573 { 1574 struct pipe *cpipe; 1575 int error; 1576 1577 cpipe = fp->f_data; 1578 if (cpipe->pipe_type & PIPE_TYPE_NAMED) 1579 error = vn_chown(fp, uid, gid, active_cred, td); 1580 else 1581 error = invfo_chown(fp, uid, gid, active_cred, td); 1582 return (error); 1583 } 1584 1585 static int 1586 pipe_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp) 1587 { 1588 struct pipe *pi; 1589 1590 if (fp->f_type == DTYPE_FIFO) 1591 return (vn_fill_kinfo(fp, kif, fdp)); 1592 kif->kf_type = KF_TYPE_PIPE; 1593 pi = fp->f_data; 1594 kif->kf_un.kf_pipe.kf_pipe_addr = (uintptr_t)pi; 1595 kif->kf_un.kf_pipe.kf_pipe_peer = (uintptr_t)pi->pipe_peer; 1596 kif->kf_un.kf_pipe.kf_pipe_buffer_cnt = pi->pipe_buffer.cnt; 1597 return (0); 1598 } 1599 1600 static void 1601 pipe_free_kmem(struct pipe *cpipe) 1602 { 1603 1604 KASSERT(!mtx_owned(PIPE_MTX(cpipe)), 1605 ("pipe_free_kmem: pipe mutex locked")); 1606 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; 1613 } 1614 #ifndef PIPE_NODIRECT 1615 { 1616 cpipe->pipe_pages.cnt = 0; 1617 cpipe->pipe_pages.pos = 0; 1618 cpipe->pipe_pages.npages = 0; 1619 } 1620 #endif 1621 } 1622 1623 /* 1624 * shutdown the pipe 1625 */ 1626 static void 1627 pipeclose(struct pipe *cpipe) 1628 { 1629 struct pipepair *pp; 1630 struct pipe *ppipe; 1631 1632 KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL")); 1633 1634 PIPE_LOCK(cpipe); 1635 pipelock(cpipe, 0); 1636 pp = cpipe->pipe_pair; 1637 1638 /* 1639 * If the other side is blocked, wake it up saying that 1640 * we want to close it down. 1641 */ 1642 cpipe->pipe_state |= PIPE_EOF; 1643 while (cpipe->pipe_busy) { 1644 wakeup(cpipe); 1645 cpipe->pipe_state |= PIPE_WANT; 1646 pipeunlock(cpipe); 1647 msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0); 1648 pipelock(cpipe, 0); 1649 } 1650 1651 pipeselwakeup(cpipe); 1652 1653 /* 1654 * Disconnect from peer, if any. 1655 */ 1656 ppipe = cpipe->pipe_peer; 1657 if (ppipe->pipe_present == PIPE_ACTIVE) { 1658 ppipe->pipe_state |= PIPE_EOF; 1659 wakeup(ppipe); 1660 pipeselwakeup(ppipe); 1661 } 1662 1663 /* 1664 * Mark this endpoint as free. Release kmem resources. We 1665 * don't mark this endpoint as unused until we've finished 1666 * doing that, or the pipe might disappear out from under 1667 * us. 1668 */ 1669 PIPE_UNLOCK(cpipe); 1670 pipe_free_kmem(cpipe); 1671 PIPE_LOCK(cpipe); 1672 cpipe->pipe_present = PIPE_CLOSING; 1673 pipeunlock(cpipe); 1674 1675 /* 1676 * knlist_clear() may sleep dropping the PIPE_MTX. Set the 1677 * PIPE_FINALIZED, that allows other end to free the 1678 * pipe_pair, only after the knotes are completely dismantled. 1679 */ 1680 knlist_clear(&cpipe->pipe_sel.si_note, 1); 1681 cpipe->pipe_present = PIPE_FINALIZED; 1682 seldrain(&cpipe->pipe_sel); 1683 knlist_destroy(&cpipe->pipe_sel.si_note); 1684 1685 /* 1686 * If both endpoints are now closed, release the memory for the 1687 * pipe pair. If not, unlock. 1688 */ 1689 if (ppipe->pipe_present == PIPE_FINALIZED) { 1690 PIPE_UNLOCK(cpipe); 1691 #ifdef MAC 1692 mac_pipe_destroy(pp); 1693 #endif 1694 uma_zfree(pipe_zone, cpipe->pipe_pair); 1695 } else 1696 PIPE_UNLOCK(cpipe); 1697 } 1698 1699 /*ARGSUSED*/ 1700 static int 1701 pipe_kqfilter(struct file *fp, struct knote *kn) 1702 { 1703 struct pipe *cpipe; 1704 1705 /* 1706 * If a filter is requested that is not supported by this file 1707 * descriptor, don't return an error, but also don't ever generate an 1708 * event. 1709 */ 1710 if ((kn->kn_filter == EVFILT_READ) && !(fp->f_flag & FREAD)) { 1711 kn->kn_fop = &pipe_nfiltops; 1712 return (0); 1713 } 1714 if ((kn->kn_filter == EVFILT_WRITE) && !(fp->f_flag & FWRITE)) { 1715 kn->kn_fop = &pipe_nfiltops; 1716 return (0); 1717 } 1718 cpipe = fp->f_data; 1719 PIPE_LOCK(cpipe); 1720 switch (kn->kn_filter) { 1721 case EVFILT_READ: 1722 kn->kn_fop = &pipe_rfiltops; 1723 break; 1724 case EVFILT_WRITE: 1725 kn->kn_fop = &pipe_wfiltops; 1726 if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) { 1727 /* other end of pipe has been closed */ 1728 PIPE_UNLOCK(cpipe); 1729 return (EPIPE); 1730 } 1731 cpipe = PIPE_PEER(cpipe); 1732 break; 1733 default: 1734 PIPE_UNLOCK(cpipe); 1735 return (EINVAL); 1736 } 1737 1738 kn->kn_hook = cpipe; 1739 knlist_add(&cpipe->pipe_sel.si_note, kn, 1); 1740 PIPE_UNLOCK(cpipe); 1741 return (0); 1742 } 1743 1744 static void 1745 filt_pipedetach(struct knote *kn) 1746 { 1747 struct pipe *cpipe = kn->kn_hook; 1748 1749 PIPE_LOCK(cpipe); 1750 knlist_remove(&cpipe->pipe_sel.si_note, kn, 1); 1751 PIPE_UNLOCK(cpipe); 1752 } 1753 1754 /*ARGSUSED*/ 1755 static int 1756 filt_piperead(struct knote *kn, long hint) 1757 { 1758 struct file *fp = kn->kn_fp; 1759 struct pipe *rpipe = kn->kn_hook; 1760 1761 PIPE_LOCK_ASSERT(rpipe, MA_OWNED); 1762 kn->kn_data = rpipe->pipe_buffer.cnt; 1763 if (kn->kn_data == 0) 1764 kn->kn_data = rpipe->pipe_pages.cnt; 1765 1766 if ((rpipe->pipe_state & PIPE_EOF) != 0 && 1767 ((rpipe->pipe_type & PIPE_TYPE_NAMED) == 0 || 1768 fp->f_pipegen != rpipe->pipe_wgen)) { 1769 kn->kn_flags |= EV_EOF; 1770 return (1); 1771 } 1772 kn->kn_flags &= ~EV_EOF; 1773 return (kn->kn_data > 0); 1774 } 1775 1776 /*ARGSUSED*/ 1777 static int 1778 filt_pipewrite(struct knote *kn, long hint) 1779 { 1780 struct pipe *wpipe = kn->kn_hook; 1781 1782 /* 1783 * If this end of the pipe is closed, the knote was removed from the 1784 * knlist and the list lock (i.e., the pipe lock) is therefore not held. 1785 */ 1786 if (wpipe->pipe_present == PIPE_ACTIVE || 1787 (wpipe->pipe_type & PIPE_TYPE_NAMED) != 0) { 1788 PIPE_LOCK_ASSERT(wpipe, MA_OWNED); 1789 1790 if (wpipe->pipe_state & PIPE_DIRECTW) { 1791 kn->kn_data = 0; 1792 } else if (wpipe->pipe_buffer.size > 0) { 1793 kn->kn_data = wpipe->pipe_buffer.size - 1794 wpipe->pipe_buffer.cnt; 1795 } else { 1796 kn->kn_data = PIPE_BUF; 1797 } 1798 } 1799 1800 if (wpipe->pipe_present != PIPE_ACTIVE || 1801 (wpipe->pipe_state & PIPE_EOF)) { 1802 kn->kn_flags |= EV_EOF; 1803 return (1); 1804 } 1805 kn->kn_flags &= ~EV_EOF; 1806 return (kn->kn_data >= PIPE_BUF); 1807 } 1808 1809 static void 1810 filt_pipedetach_notsup(struct knote *kn) 1811 { 1812 1813 } 1814 1815 static int 1816 filt_pipenotsup(struct knote *kn, long hint) 1817 { 1818 1819 return (0); 1820 }
Cache object: bb21d3e82577dde6557e04a8196a2a5b
[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]