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"); 226 227 static void pipeinit(void *dummy __unused); 228 static void pipeclose(struct pipe *cpipe); 229 static void pipe_free_kmem(struct pipe *cpipe); 230 static int pipe_create(struct pipe *pipe, bool backing); 231 static int pipe_paircreate(struct thread *td, struct pipepair **p_pp); 232 static __inline int pipelock(struct pipe *cpipe, int catch); 233 static __inline void pipeunlock(struct pipe *cpipe); 234 static void pipe_timestamp(struct timespec *tsp); 235 #ifndef PIPE_NODIRECT 236 static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio); 237 static void pipe_destroy_write_buffer(struct pipe *wpipe); 238 static int pipe_direct_write(struct pipe *wpipe, struct uio *uio); 239 static void pipe_clone_write_buffer(struct pipe *wpipe); 240 #endif 241 static int pipespace(struct pipe *cpipe, int size); 242 static int pipespace_new(struct pipe *cpipe, int size); 243 244 static int pipe_zone_ctor(void *mem, int size, void *arg, int flags); 245 static int pipe_zone_init(void *mem, int size, int flags); 246 static void pipe_zone_fini(void *mem, int size); 247 248 static uma_zone_t pipe_zone; 249 static struct unrhdr64 pipeino_unr; 250 static dev_t pipedev_ino; 251 252 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL); 253 254 static void 255 pipeinit(void *dummy __unused) 256 { 257 258 pipe_zone = uma_zcreate("pipe", sizeof(struct pipepair), 259 pipe_zone_ctor, NULL, pipe_zone_init, pipe_zone_fini, 260 UMA_ALIGN_PTR, 0); 261 KASSERT(pipe_zone != NULL, ("pipe_zone not initialized")); 262 new_unrhdr64(&pipeino_unr, 1); 263 pipedev_ino = devfs_alloc_cdp_inode(); 264 KASSERT(pipedev_ino > 0, ("pipe dev inode not initialized")); 265 } 266 267 static int 268 sysctl_handle_pipe_mindirect(SYSCTL_HANDLER_ARGS) 269 { 270 int error = 0; 271 long tmp_pipe_mindirect = pipe_mindirect; 272 273 error = sysctl_handle_long(oidp, &tmp_pipe_mindirect, arg2, req); 274 if (error != 0 || req->newptr == NULL) 275 return (error); 276 277 /* 278 * Don't allow pipe_mindirect to be set so low that we violate 279 * atomicity requirements. 280 */ 281 if (tmp_pipe_mindirect <= PIPE_BUF) 282 return (EINVAL); 283 pipe_mindirect = tmp_pipe_mindirect; 284 return (0); 285 } 286 SYSCTL_OID(_kern_ipc, OID_AUTO, pipe_mindirect, CTLTYPE_LONG | CTLFLAG_RW, 287 &pipe_mindirect, 0, sysctl_handle_pipe_mindirect, "L", 288 "Minimum write size triggering VM optimization"); 289 290 static int 291 pipe_zone_ctor(void *mem, int size, void *arg, int flags) 292 { 293 struct pipepair *pp; 294 struct pipe *rpipe, *wpipe; 295 296 KASSERT(size == sizeof(*pp), ("pipe_zone_ctor: wrong size")); 297 298 pp = (struct pipepair *)mem; 299 300 /* 301 * We zero both pipe endpoints to make sure all the kmem pointers 302 * are NULL, flag fields are zero'd, etc. We timestamp both 303 * endpoints with the same time. 304 */ 305 rpipe = &pp->pp_rpipe; 306 bzero(rpipe, sizeof(*rpipe)); 307 pipe_timestamp(&rpipe->pipe_ctime); 308 rpipe->pipe_atime = rpipe->pipe_mtime = rpipe->pipe_ctime; 309 310 wpipe = &pp->pp_wpipe; 311 bzero(wpipe, sizeof(*wpipe)); 312 wpipe->pipe_ctime = rpipe->pipe_ctime; 313 wpipe->pipe_atime = wpipe->pipe_mtime = rpipe->pipe_ctime; 314 315 rpipe->pipe_peer = wpipe; 316 rpipe->pipe_pair = pp; 317 wpipe->pipe_peer = rpipe; 318 wpipe->pipe_pair = pp; 319 320 /* 321 * Mark both endpoints as present; they will later get free'd 322 * one at a time. When both are free'd, then the whole pair 323 * is released. 324 */ 325 rpipe->pipe_present = PIPE_ACTIVE; 326 wpipe->pipe_present = PIPE_ACTIVE; 327 328 /* 329 * Eventually, the MAC Framework may initialize the label 330 * in ctor or init, but for now we do it elswhere to avoid 331 * blocking in ctor or init. 332 */ 333 pp->pp_label = NULL; 334 335 return (0); 336 } 337 338 static int 339 pipe_zone_init(void *mem, int size, int flags) 340 { 341 struct pipepair *pp; 342 343 KASSERT(size == sizeof(*pp), ("pipe_zone_init: wrong size")); 344 345 pp = (struct pipepair *)mem; 346 347 mtx_init(&pp->pp_mtx, "pipe mutex", NULL, MTX_DEF | MTX_NEW); 348 return (0); 349 } 350 351 static void 352 pipe_zone_fini(void *mem, int size) 353 { 354 struct pipepair *pp; 355 356 KASSERT(size == sizeof(*pp), ("pipe_zone_fini: wrong size")); 357 358 pp = (struct pipepair *)mem; 359 360 mtx_destroy(&pp->pp_mtx); 361 } 362 363 static int 364 pipe_paircreate(struct thread *td, struct pipepair **p_pp) 365 { 366 struct pipepair *pp; 367 struct pipe *rpipe, *wpipe; 368 int error; 369 370 *p_pp = pp = uma_zalloc(pipe_zone, M_WAITOK); 371 #ifdef MAC 372 /* 373 * The MAC label is shared between the connected endpoints. As a 374 * result mac_pipe_init() and mac_pipe_create() are called once 375 * for the pair, and not on the endpoints. 376 */ 377 mac_pipe_init(pp); 378 mac_pipe_create(td->td_ucred, pp); 379 #endif 380 rpipe = &pp->pp_rpipe; 381 wpipe = &pp->pp_wpipe; 382 383 knlist_init_mtx(&rpipe->pipe_sel.si_note, PIPE_MTX(rpipe)); 384 knlist_init_mtx(&wpipe->pipe_sel.si_note, PIPE_MTX(wpipe)); 385 386 /* 387 * Only the forward direction pipe is backed by big buffer by 388 * default. 389 */ 390 error = pipe_create(rpipe, true); 391 if (error != 0) 392 goto fail; 393 error = pipe_create(wpipe, false); 394 if (error != 0) { 395 /* 396 * This cleanup leaves the pipe inode number for rpipe 397 * still allocated, but never used. We do not free 398 * inode numbers for opened pipes, which is required 399 * for correctness because numbers must be unique. 400 * But also it avoids any memory use by the unr 401 * allocator, so stashing away the transient inode 402 * number is reasonable. 403 */ 404 pipe_free_kmem(rpipe); 405 goto fail; 406 } 407 408 rpipe->pipe_state |= PIPE_DIRECTOK; 409 wpipe->pipe_state |= PIPE_DIRECTOK; 410 return (0); 411 412 fail: 413 knlist_destroy(&rpipe->pipe_sel.si_note); 414 knlist_destroy(&wpipe->pipe_sel.si_note); 415 #ifdef MAC 416 mac_pipe_destroy(pp); 417 #endif 418 uma_zfree(pipe_zone, pp); 419 return (error); 420 } 421 422 int 423 pipe_named_ctor(struct pipe **ppipe, struct thread *td) 424 { 425 struct pipepair *pp; 426 int error; 427 428 error = pipe_paircreate(td, &pp); 429 if (error != 0) 430 return (error); 431 pp->pp_rpipe.pipe_type |= PIPE_TYPE_NAMED; 432 *ppipe = &pp->pp_rpipe; 433 return (0); 434 } 435 436 void 437 pipe_dtor(struct pipe *dpipe) 438 { 439 struct pipe *peer; 440 441 peer = (dpipe->pipe_type & PIPE_TYPE_NAMED) != 0 ? dpipe->pipe_peer : NULL; 442 funsetown(&dpipe->pipe_sigio); 443 pipeclose(dpipe); 444 if (peer != NULL) { 445 funsetown(&peer->pipe_sigio); 446 pipeclose(peer); 447 } 448 } 449 450 /* 451 * Get a timestamp. 452 * 453 * This used to be vfs_timestamp but the higher precision is unnecessary and 454 * can very negatively affect performance in virtualized environments (e.g., on 455 * vms running on amd64 when using the rdtscp instruction). 456 */ 457 static void 458 pipe_timestamp(struct timespec *tsp) 459 { 460 461 getnanotime(tsp); 462 } 463 464 /* 465 * The pipe system call for the DTYPE_PIPE type of pipes. If we fail, let 466 * the zone pick up the pieces via pipeclose(). 467 */ 468 int 469 kern_pipe(struct thread *td, int fildes[2], int flags, struct filecaps *fcaps1, 470 struct filecaps *fcaps2) 471 { 472 struct file *rf, *wf; 473 struct pipe *rpipe, *wpipe; 474 struct pipepair *pp; 475 int fd, fflags, error; 476 477 error = pipe_paircreate(td, &pp); 478 if (error != 0) 479 return (error); 480 rpipe = &pp->pp_rpipe; 481 wpipe = &pp->pp_wpipe; 482 error = falloc_caps(td, &rf, &fd, flags, fcaps1); 483 if (error) { 484 pipeclose(rpipe); 485 pipeclose(wpipe); 486 return (error); 487 } 488 /* An extra reference on `rf' has been held for us by falloc_caps(). */ 489 fildes[0] = fd; 490 491 fflags = FREAD | FWRITE; 492 if ((flags & O_NONBLOCK) != 0) 493 fflags |= FNONBLOCK; 494 495 /* 496 * Warning: once we've gotten past allocation of the fd for the 497 * read-side, we can only drop the read side via fdrop() in order 498 * to avoid races against processes which manage to dup() the read 499 * side while we are blocked trying to allocate the write side. 500 */ 501 finit(rf, fflags, DTYPE_PIPE, rpipe, &pipeops); 502 error = falloc_caps(td, &wf, &fd, flags, fcaps2); 503 if (error) { 504 fdclose(td, rf, fildes[0]); 505 fdrop(rf, td); 506 /* rpipe has been closed by fdrop(). */ 507 pipeclose(wpipe); 508 return (error); 509 } 510 /* An extra reference on `wf' has been held for us by falloc_caps(). */ 511 finit(wf, fflags, DTYPE_PIPE, wpipe, &pipeops); 512 fdrop(wf, td); 513 fildes[1] = fd; 514 fdrop(rf, td); 515 516 return (0); 517 } 518 519 #ifdef COMPAT_FREEBSD10 520 /* ARGSUSED */ 521 int 522 freebsd10_pipe(struct thread *td, struct freebsd10_pipe_args *uap __unused) 523 { 524 int error; 525 int fildes[2]; 526 527 error = kern_pipe(td, fildes, 0, NULL, NULL); 528 if (error) 529 return (error); 530 531 td->td_retval[0] = fildes[0]; 532 td->td_retval[1] = fildes[1]; 533 534 return (0); 535 } 536 #endif 537 538 int 539 sys_pipe2(struct thread *td, struct pipe2_args *uap) 540 { 541 int error, fildes[2]; 542 543 if (uap->flags & ~(O_CLOEXEC | O_NONBLOCK)) 544 return (EINVAL); 545 error = kern_pipe(td, fildes, uap->flags, NULL, NULL); 546 if (error) 547 return (error); 548 error = copyout(fildes, uap->fildes, 2 * sizeof(int)); 549 if (error) { 550 (void)kern_close(td, fildes[0]); 551 (void)kern_close(td, fildes[1]); 552 } 553 return (error); 554 } 555 556 /* 557 * Allocate kva for pipe circular buffer, the space is pageable 558 * This routine will 'realloc' the size of a pipe safely, if it fails 559 * it will retain the old buffer. 560 * If it fails it will return ENOMEM. 561 */ 562 static int 563 pipespace_new(struct pipe *cpipe, int size) 564 { 565 caddr_t buffer; 566 int error, cnt, firstseg; 567 static int curfail = 0; 568 static struct timeval lastfail; 569 570 KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked")); 571 KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW), 572 ("pipespace: resize of direct writes not allowed")); 573 retry: 574 cnt = cpipe->pipe_buffer.cnt; 575 if (cnt > size) 576 size = cnt; 577 578 size = round_page(size); 579 buffer = (caddr_t) vm_map_min(pipe_map); 580 581 error = vm_map_find(pipe_map, NULL, 0, (vm_offset_t *)&buffer, size, 0, 582 VMFS_ANY_SPACE, VM_PROT_RW, VM_PROT_RW, 0); 583 if (error != KERN_SUCCESS) { 584 if (cpipe->pipe_buffer.buffer == NULL && 585 size > SMALL_PIPE_SIZE) { 586 size = SMALL_PIPE_SIZE; 587 pipefragretry++; 588 goto retry; 589 } 590 if (cpipe->pipe_buffer.buffer == NULL) { 591 pipeallocfail++; 592 if (ppsratecheck(&lastfail, &curfail, 1)) 593 printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n"); 594 } else { 595 piperesizefail++; 596 } 597 return (ENOMEM); 598 } 599 600 /* copy data, then free old resources if we're resizing */ 601 if (cnt > 0) { 602 if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) { 603 firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out; 604 bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out], 605 buffer, firstseg); 606 if ((cnt - firstseg) > 0) 607 bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg], 608 cpipe->pipe_buffer.in); 609 } else { 610 bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out], 611 buffer, cnt); 612 } 613 } 614 pipe_free_kmem(cpipe); 615 cpipe->pipe_buffer.buffer = buffer; 616 cpipe->pipe_buffer.size = size; 617 cpipe->pipe_buffer.in = cnt; 618 cpipe->pipe_buffer.out = 0; 619 cpipe->pipe_buffer.cnt = cnt; 620 atomic_add_long(&amountpipekva, cpipe->pipe_buffer.size); 621 return (0); 622 } 623 624 /* 625 * Wrapper for pipespace_new() that performs locking assertions. 626 */ 627 static int 628 pipespace(struct pipe *cpipe, int size) 629 { 630 631 KASSERT(cpipe->pipe_state & PIPE_LOCKFL, 632 ("Unlocked pipe passed to pipespace")); 633 return (pipespace_new(cpipe, size)); 634 } 635 636 /* 637 * lock a pipe for I/O, blocking other access 638 */ 639 static __inline int 640 pipelock(struct pipe *cpipe, int catch) 641 { 642 int error, prio; 643 644 PIPE_LOCK_ASSERT(cpipe, MA_OWNED); 645 646 prio = PRIBIO; 647 if (catch) 648 prio |= PCATCH; 649 while (cpipe->pipe_state & PIPE_LOCKFL) { 650 KASSERT(cpipe->pipe_waiters >= 0, 651 ("%s: bad waiter count %d", __func__, 652 cpipe->pipe_waiters)); 653 cpipe->pipe_waiters++; 654 error = msleep(&cpipe->pipe_waiters, PIPE_MTX(cpipe), prio, 655 "pipelk", 0); 656 cpipe->pipe_waiters--; 657 if (error != 0) 658 return (error); 659 } 660 cpipe->pipe_state |= PIPE_LOCKFL; 661 return (0); 662 } 663 664 /* 665 * unlock a pipe I/O lock 666 */ 667 static __inline void 668 pipeunlock(struct pipe *cpipe) 669 { 670 671 PIPE_LOCK_ASSERT(cpipe, MA_OWNED); 672 KASSERT(cpipe->pipe_state & PIPE_LOCKFL, 673 ("Unlocked pipe passed to pipeunlock")); 674 KASSERT(cpipe->pipe_waiters >= 0, 675 ("%s: bad waiter count %d", __func__, 676 cpipe->pipe_waiters)); 677 cpipe->pipe_state &= ~PIPE_LOCKFL; 678 if (cpipe->pipe_waiters > 0) 679 wakeup_one(&cpipe->pipe_waiters); 680 } 681 682 void 683 pipeselwakeup(struct pipe *cpipe) 684 { 685 686 PIPE_LOCK_ASSERT(cpipe, MA_OWNED); 687 if (cpipe->pipe_state & PIPE_SEL) { 688 selwakeuppri(&cpipe->pipe_sel, PSOCK); 689 if (!SEL_WAITING(&cpipe->pipe_sel)) 690 cpipe->pipe_state &= ~PIPE_SEL; 691 } 692 if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio) 693 pgsigio(&cpipe->pipe_sigio, SIGIO, 0); 694 KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0); 695 } 696 697 /* 698 * Initialize and allocate VM and memory for pipe. The structure 699 * will start out zero'd from the ctor, so we just manage the kmem. 700 */ 701 static int 702 pipe_create(struct pipe *pipe, bool large_backing) 703 { 704 int error; 705 706 error = pipespace_new(pipe, !large_backing || amountpipekva > 707 maxpipekva / 2 ? SMALL_PIPE_SIZE : PIPE_SIZE); 708 if (error == 0) 709 pipe->pipe_ino = alloc_unr64(&pipeino_unr); 710 return (error); 711 } 712 713 /* ARGSUSED */ 714 static int 715 pipe_read(struct file *fp, struct uio *uio, struct ucred *active_cred, 716 int flags, struct thread *td) 717 { 718 struct pipe *rpipe; 719 int error; 720 int nread = 0; 721 int size; 722 723 rpipe = fp->f_data; 724 725 /* 726 * Try to avoid locking the pipe if we have nothing to do. 727 * 728 * There are programs which share one pipe amongst multiple processes 729 * and perform non-blocking reads in parallel, even if the pipe is 730 * empty. This in particular is the case with BSD make, which when 731 * spawned with a high -j number can find itself with over half of the 732 * calls failing to find anything. 733 */ 734 if ((fp->f_flag & FNONBLOCK) != 0 && !mac_pipe_check_read_enabled()) { 735 if (__predict_false(uio->uio_resid == 0)) 736 return (0); 737 if ((atomic_load_short(&rpipe->pipe_state) & PIPE_EOF) == 0 && 738 atomic_load_int(&rpipe->pipe_buffer.cnt) == 0 && 739 atomic_load_int(&rpipe->pipe_pages.cnt) == 0) 740 return (EAGAIN); 741 } 742 743 PIPE_LOCK(rpipe); 744 ++rpipe->pipe_busy; 745 error = pipelock(rpipe, 1); 746 if (error) 747 goto unlocked_error; 748 749 #ifdef MAC 750 error = mac_pipe_check_read(active_cred, rpipe->pipe_pair); 751 if (error) 752 goto locked_error; 753 #endif 754 if (amountpipekva > (3 * maxpipekva) / 4) { 755 if ((rpipe->pipe_state & PIPE_DIRECTW) == 0 && 756 rpipe->pipe_buffer.size > SMALL_PIPE_SIZE && 757 rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE && 758 piperesizeallowed == 1) { 759 PIPE_UNLOCK(rpipe); 760 pipespace(rpipe, SMALL_PIPE_SIZE); 761 PIPE_LOCK(rpipe); 762 } 763 } 764 765 while (uio->uio_resid) { 766 /* 767 * normal pipe buffer receive 768 */ 769 if (rpipe->pipe_buffer.cnt > 0) { 770 size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out; 771 if (size > rpipe->pipe_buffer.cnt) 772 size = rpipe->pipe_buffer.cnt; 773 if (size > uio->uio_resid) 774 size = uio->uio_resid; 775 776 PIPE_UNLOCK(rpipe); 777 error = uiomove( 778 &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out], 779 size, uio); 780 PIPE_LOCK(rpipe); 781 if (error) 782 break; 783 784 rpipe->pipe_buffer.out += size; 785 if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size) 786 rpipe->pipe_buffer.out = 0; 787 788 rpipe->pipe_buffer.cnt -= size; 789 790 /* 791 * If there is no more to read in the pipe, reset 792 * its pointers to the beginning. This improves 793 * cache hit stats. 794 */ 795 if (rpipe->pipe_buffer.cnt == 0) { 796 rpipe->pipe_buffer.in = 0; 797 rpipe->pipe_buffer.out = 0; 798 } 799 nread += size; 800 #ifndef PIPE_NODIRECT 801 /* 802 * Direct copy, bypassing a kernel buffer. 803 */ 804 } else if ((size = rpipe->pipe_pages.cnt) != 0) { 805 if (size > uio->uio_resid) 806 size = (u_int) uio->uio_resid; 807 PIPE_UNLOCK(rpipe); 808 error = uiomove_fromphys(rpipe->pipe_pages.ms, 809 rpipe->pipe_pages.pos, size, uio); 810 PIPE_LOCK(rpipe); 811 if (error) 812 break; 813 nread += size; 814 rpipe->pipe_pages.pos += size; 815 rpipe->pipe_pages.cnt -= size; 816 if (rpipe->pipe_pages.cnt == 0) { 817 rpipe->pipe_state &= ~PIPE_WANTW; 818 wakeup(rpipe); 819 } 820 #endif 821 } else { 822 /* 823 * detect EOF condition 824 * read returns 0 on EOF, no need to set error 825 */ 826 if (rpipe->pipe_state & PIPE_EOF) 827 break; 828 829 /* 830 * If the "write-side" has been blocked, wake it up now. 831 */ 832 if (rpipe->pipe_state & PIPE_WANTW) { 833 rpipe->pipe_state &= ~PIPE_WANTW; 834 wakeup(rpipe); 835 } 836 837 /* 838 * Break if some data was read. 839 */ 840 if (nread > 0) 841 break; 842 843 /* 844 * Unlock the pipe buffer for our remaining processing. 845 * We will either break out with an error or we will 846 * sleep and relock to loop. 847 */ 848 pipeunlock(rpipe); 849 850 /* 851 * Handle non-blocking mode operation or 852 * wait for more data. 853 */ 854 if (fp->f_flag & FNONBLOCK) { 855 error = EAGAIN; 856 } else { 857 rpipe->pipe_state |= PIPE_WANTR; 858 if ((error = msleep(rpipe, PIPE_MTX(rpipe), 859 PRIBIO | PCATCH, 860 "piperd", 0)) == 0) 861 error = pipelock(rpipe, 1); 862 } 863 if (error) 864 goto unlocked_error; 865 } 866 } 867 #ifdef MAC 868 locked_error: 869 #endif 870 pipeunlock(rpipe); 871 872 /* XXX: should probably do this before getting any locks. */ 873 if (error == 0) 874 pipe_timestamp(&rpipe->pipe_atime); 875 unlocked_error: 876 --rpipe->pipe_busy; 877 878 /* 879 * PIPE_WANT processing only makes sense if pipe_busy is 0. 880 */ 881 if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) { 882 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW); 883 wakeup(rpipe); 884 } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) { 885 /* 886 * Handle write blocking hysteresis. 887 */ 888 if (rpipe->pipe_state & PIPE_WANTW) { 889 rpipe->pipe_state &= ~PIPE_WANTW; 890 wakeup(rpipe); 891 } 892 } 893 894 /* 895 * Only wake up writers if there was actually something read. 896 * Otherwise, when calling read(2) at EOF, a spurious wakeup occurs. 897 */ 898 if (nread > 0 && 899 rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt >= PIPE_BUF) 900 pipeselwakeup(rpipe); 901 902 PIPE_UNLOCK(rpipe); 903 if (nread > 0) 904 td->td_ru.ru_msgrcv++; 905 return (error); 906 } 907 908 #ifndef PIPE_NODIRECT 909 /* 910 * Map the sending processes' buffer into kernel space and wire it. 911 * This is similar to a physical write operation. 912 */ 913 static int 914 pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio) 915 { 916 u_int size; 917 int i; 918 919 PIPE_LOCK_ASSERT(wpipe, MA_OWNED); 920 KASSERT((wpipe->pipe_state & PIPE_DIRECTW) == 0, 921 ("%s: PIPE_DIRECTW set on %p", __func__, wpipe)); 922 KASSERT(wpipe->pipe_pages.cnt == 0, 923 ("%s: pipe map for %p contains residual data", __func__, wpipe)); 924 925 if (uio->uio_iov->iov_len > wpipe->pipe_buffer.size) 926 size = wpipe->pipe_buffer.size; 927 else 928 size = uio->uio_iov->iov_len; 929 930 wpipe->pipe_state |= PIPE_DIRECTW; 931 PIPE_UNLOCK(wpipe); 932 i = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map, 933 (vm_offset_t)uio->uio_iov->iov_base, size, VM_PROT_READ, 934 wpipe->pipe_pages.ms, PIPENPAGES); 935 PIPE_LOCK(wpipe); 936 if (i < 0) { 937 wpipe->pipe_state &= ~PIPE_DIRECTW; 938 return (EFAULT); 939 } 940 941 wpipe->pipe_pages.npages = i; 942 wpipe->pipe_pages.pos = 943 ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK; 944 wpipe->pipe_pages.cnt = size; 945 946 uio->uio_iov->iov_len -= size; 947 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size; 948 if (uio->uio_iov->iov_len == 0) 949 uio->uio_iov++; 950 uio->uio_resid -= size; 951 uio->uio_offset += size; 952 return (0); 953 } 954 955 /* 956 * Unwire the process buffer. 957 */ 958 static void 959 pipe_destroy_write_buffer(struct pipe *wpipe) 960 { 961 962 PIPE_LOCK_ASSERT(wpipe, MA_OWNED); 963 KASSERT((wpipe->pipe_state & PIPE_DIRECTW) != 0, 964 ("%s: PIPE_DIRECTW not set on %p", __func__, wpipe)); 965 KASSERT(wpipe->pipe_pages.cnt == 0, 966 ("%s: pipe map for %p contains residual data", __func__, wpipe)); 967 968 wpipe->pipe_state &= ~PIPE_DIRECTW; 969 vm_page_unhold_pages(wpipe->pipe_pages.ms, wpipe->pipe_pages.npages); 970 wpipe->pipe_pages.npages = 0; 971 } 972 973 /* 974 * In the case of a signal, the writing process might go away. This 975 * code copies the data into the circular buffer so that the source 976 * pages can be freed without loss of data. 977 */ 978 static void 979 pipe_clone_write_buffer(struct pipe *wpipe) 980 { 981 struct uio uio; 982 struct iovec iov; 983 int size; 984 int pos; 985 986 PIPE_LOCK_ASSERT(wpipe, MA_OWNED); 987 KASSERT((wpipe->pipe_state & PIPE_DIRECTW) != 0, 988 ("%s: PIPE_DIRECTW not set on %p", __func__, wpipe)); 989 990 size = wpipe->pipe_pages.cnt; 991 pos = wpipe->pipe_pages.pos; 992 wpipe->pipe_pages.cnt = 0; 993 994 wpipe->pipe_buffer.in = size; 995 wpipe->pipe_buffer.out = 0; 996 wpipe->pipe_buffer.cnt = size; 997 998 PIPE_UNLOCK(wpipe); 999 iov.iov_base = wpipe->pipe_buffer.buffer; 1000 iov.iov_len = size; 1001 uio.uio_iov = &iov; 1002 uio.uio_iovcnt = 1; 1003 uio.uio_offset = 0; 1004 uio.uio_resid = size; 1005 uio.uio_segflg = UIO_SYSSPACE; 1006 uio.uio_rw = UIO_READ; 1007 uio.uio_td = curthread; 1008 uiomove_fromphys(wpipe->pipe_pages.ms, pos, size, &uio); 1009 PIPE_LOCK(wpipe); 1010 pipe_destroy_write_buffer(wpipe); 1011 } 1012 1013 /* 1014 * This implements the pipe buffer write mechanism. Note that only 1015 * a direct write OR a normal pipe write can be pending at any given time. 1016 * If there are any characters in the pipe buffer, the direct write will 1017 * be deferred until the receiving process grabs all of the bytes from 1018 * the pipe buffer. Then the direct mapping write is set-up. 1019 */ 1020 static int 1021 pipe_direct_write(struct pipe *wpipe, struct uio *uio) 1022 { 1023 int error; 1024 1025 retry: 1026 PIPE_LOCK_ASSERT(wpipe, MA_OWNED); 1027 if ((wpipe->pipe_state & PIPE_EOF) != 0) { 1028 error = EPIPE; 1029 goto error1; 1030 } 1031 if (wpipe->pipe_state & PIPE_DIRECTW) { 1032 if (wpipe->pipe_state & PIPE_WANTR) { 1033 wpipe->pipe_state &= ~PIPE_WANTR; 1034 wakeup(wpipe); 1035 } 1036 pipeselwakeup(wpipe); 1037 wpipe->pipe_state |= PIPE_WANTW; 1038 pipeunlock(wpipe); 1039 error = msleep(wpipe, PIPE_MTX(wpipe), 1040 PRIBIO | PCATCH, "pipdww", 0); 1041 pipelock(wpipe, 0); 1042 if (error != 0) 1043 goto error1; 1044 goto retry; 1045 } 1046 if (wpipe->pipe_buffer.cnt > 0) { 1047 if (wpipe->pipe_state & PIPE_WANTR) { 1048 wpipe->pipe_state &= ~PIPE_WANTR; 1049 wakeup(wpipe); 1050 } 1051 pipeselwakeup(wpipe); 1052 wpipe->pipe_state |= PIPE_WANTW; 1053 pipeunlock(wpipe); 1054 error = msleep(wpipe, PIPE_MTX(wpipe), 1055 PRIBIO | PCATCH, "pipdwc", 0); 1056 pipelock(wpipe, 0); 1057 if (error != 0) 1058 goto error1; 1059 goto retry; 1060 } 1061 1062 error = pipe_build_write_buffer(wpipe, uio); 1063 if (error) { 1064 goto error1; 1065 } 1066 1067 while (wpipe->pipe_pages.cnt != 0 && 1068 (wpipe->pipe_state & PIPE_EOF) == 0) { 1069 if (wpipe->pipe_state & PIPE_WANTR) { 1070 wpipe->pipe_state &= ~PIPE_WANTR; 1071 wakeup(wpipe); 1072 } 1073 pipeselwakeup(wpipe); 1074 wpipe->pipe_state |= PIPE_WANTW; 1075 pipeunlock(wpipe); 1076 error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH, 1077 "pipdwt", 0); 1078 pipelock(wpipe, 0); 1079 if (error != 0) 1080 break; 1081 } 1082 1083 if ((wpipe->pipe_state & PIPE_EOF) != 0) { 1084 wpipe->pipe_pages.cnt = 0; 1085 pipe_destroy_write_buffer(wpipe); 1086 pipeselwakeup(wpipe); 1087 error = EPIPE; 1088 } else if (error == EINTR || error == ERESTART) { 1089 pipe_clone_write_buffer(wpipe); 1090 } else { 1091 pipe_destroy_write_buffer(wpipe); 1092 } 1093 KASSERT((wpipe->pipe_state & PIPE_DIRECTW) == 0, 1094 ("pipe %p leaked PIPE_DIRECTW", wpipe)); 1095 return (error); 1096 1097 error1: 1098 wakeup(wpipe); 1099 return (error); 1100 } 1101 #endif 1102 1103 static int 1104 pipe_write(struct file *fp, struct uio *uio, struct ucred *active_cred, 1105 int flags, struct thread *td) 1106 { 1107 struct pipe *wpipe, *rpipe; 1108 ssize_t orig_resid; 1109 int desiredsize, error; 1110 1111 rpipe = fp->f_data; 1112 wpipe = PIPE_PEER(rpipe); 1113 PIPE_LOCK(rpipe); 1114 error = pipelock(wpipe, 1); 1115 if (error) { 1116 PIPE_UNLOCK(rpipe); 1117 return (error); 1118 } 1119 /* 1120 * detect loss of pipe read side, issue SIGPIPE if lost. 1121 */ 1122 if (wpipe->pipe_present != PIPE_ACTIVE || 1123 (wpipe->pipe_state & PIPE_EOF)) { 1124 pipeunlock(wpipe); 1125 PIPE_UNLOCK(rpipe); 1126 return (EPIPE); 1127 } 1128 #ifdef MAC 1129 error = mac_pipe_check_write(active_cred, wpipe->pipe_pair); 1130 if (error) { 1131 pipeunlock(wpipe); 1132 PIPE_UNLOCK(rpipe); 1133 return (error); 1134 } 1135 #endif 1136 ++wpipe->pipe_busy; 1137 1138 /* Choose a larger size if it's advantageous */ 1139 desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size); 1140 while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) { 1141 if (piperesizeallowed != 1) 1142 break; 1143 if (amountpipekva > maxpipekva / 2) 1144 break; 1145 if (desiredsize == BIG_PIPE_SIZE) 1146 break; 1147 desiredsize = desiredsize * 2; 1148 } 1149 1150 /* Choose a smaller size if we're in a OOM situation */ 1151 if (amountpipekva > (3 * maxpipekva) / 4 && 1152 wpipe->pipe_buffer.size > SMALL_PIPE_SIZE && 1153 wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE && 1154 piperesizeallowed == 1) 1155 desiredsize = SMALL_PIPE_SIZE; 1156 1157 /* Resize if the above determined that a new size was necessary */ 1158 if (desiredsize != wpipe->pipe_buffer.size && 1159 (wpipe->pipe_state & PIPE_DIRECTW) == 0) { 1160 PIPE_UNLOCK(wpipe); 1161 pipespace(wpipe, desiredsize); 1162 PIPE_LOCK(wpipe); 1163 } 1164 MPASS(wpipe->pipe_buffer.size != 0); 1165 1166 orig_resid = uio->uio_resid; 1167 1168 while (uio->uio_resid) { 1169 int space; 1170 1171 if (wpipe->pipe_state & PIPE_EOF) { 1172 error = EPIPE; 1173 break; 1174 } 1175 #ifndef PIPE_NODIRECT 1176 /* 1177 * If the transfer is large, we can gain performance if 1178 * we do process-to-process copies directly. 1179 * If the write is non-blocking, we don't use the 1180 * direct write mechanism. 1181 * 1182 * The direct write mechanism will detect the reader going 1183 * away on us. 1184 */ 1185 if (uio->uio_segflg == UIO_USERSPACE && 1186 uio->uio_iov->iov_len >= pipe_mindirect && 1187 wpipe->pipe_buffer.size >= pipe_mindirect && 1188 (fp->f_flag & FNONBLOCK) == 0) { 1189 error = pipe_direct_write(wpipe, uio); 1190 if (error != 0) 1191 break; 1192 continue; 1193 } 1194 #endif 1195 1196 /* 1197 * Pipe buffered writes cannot be coincidental with 1198 * direct writes. We wait until the currently executing 1199 * direct write is completed before we start filling the 1200 * pipe buffer. We break out if a signal occurs or the 1201 * reader goes away. 1202 */ 1203 if (wpipe->pipe_pages.cnt != 0) { 1204 if (wpipe->pipe_state & PIPE_WANTR) { 1205 wpipe->pipe_state &= ~PIPE_WANTR; 1206 wakeup(wpipe); 1207 } 1208 pipeselwakeup(wpipe); 1209 wpipe->pipe_state |= PIPE_WANTW; 1210 pipeunlock(wpipe); 1211 error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH, 1212 "pipbww", 0); 1213 pipelock(wpipe, 0); 1214 if (error != 0) 1215 break; 1216 continue; 1217 } 1218 1219 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt; 1220 1221 /* Writes of size <= PIPE_BUF must be atomic. */ 1222 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF)) 1223 space = 0; 1224 1225 if (space > 0) { 1226 int size; /* Transfer size */ 1227 int segsize; /* first segment to transfer */ 1228 1229 /* 1230 * Transfer size is minimum of uio transfer 1231 * and free space in pipe buffer. 1232 */ 1233 if (space > uio->uio_resid) 1234 size = uio->uio_resid; 1235 else 1236 size = space; 1237 /* 1238 * First segment to transfer is minimum of 1239 * transfer size and contiguous space in 1240 * pipe buffer. If first segment to transfer 1241 * is less than the transfer size, we've got 1242 * a wraparound in the buffer. 1243 */ 1244 segsize = wpipe->pipe_buffer.size - 1245 wpipe->pipe_buffer.in; 1246 if (segsize > size) 1247 segsize = size; 1248 1249 /* Transfer first segment */ 1250 1251 PIPE_UNLOCK(rpipe); 1252 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in], 1253 segsize, uio); 1254 PIPE_LOCK(rpipe); 1255 1256 if (error == 0 && segsize < size) { 1257 KASSERT(wpipe->pipe_buffer.in + segsize == 1258 wpipe->pipe_buffer.size, 1259 ("Pipe buffer wraparound disappeared")); 1260 /* 1261 * Transfer remaining part now, to 1262 * support atomic writes. Wraparound 1263 * happened. 1264 */ 1265 1266 PIPE_UNLOCK(rpipe); 1267 error = uiomove( 1268 &wpipe->pipe_buffer.buffer[0], 1269 size - segsize, uio); 1270 PIPE_LOCK(rpipe); 1271 } 1272 if (error == 0) { 1273 wpipe->pipe_buffer.in += size; 1274 if (wpipe->pipe_buffer.in >= 1275 wpipe->pipe_buffer.size) { 1276 KASSERT(wpipe->pipe_buffer.in == 1277 size - segsize + 1278 wpipe->pipe_buffer.size, 1279 ("Expected wraparound bad")); 1280 wpipe->pipe_buffer.in = size - segsize; 1281 } 1282 1283 wpipe->pipe_buffer.cnt += size; 1284 KASSERT(wpipe->pipe_buffer.cnt <= 1285 wpipe->pipe_buffer.size, 1286 ("Pipe buffer overflow")); 1287 } 1288 if (error != 0) 1289 break; 1290 continue; 1291 } else { 1292 /* 1293 * If the "read-side" has been blocked, wake it up now. 1294 */ 1295 if (wpipe->pipe_state & PIPE_WANTR) { 1296 wpipe->pipe_state &= ~PIPE_WANTR; 1297 wakeup(wpipe); 1298 } 1299 1300 /* 1301 * don't block on non-blocking I/O 1302 */ 1303 if (fp->f_flag & FNONBLOCK) { 1304 error = EAGAIN; 1305 break; 1306 } 1307 1308 /* 1309 * We have no more space and have something to offer, 1310 * wake up select/poll. 1311 */ 1312 pipeselwakeup(wpipe); 1313 1314 wpipe->pipe_state |= PIPE_WANTW; 1315 pipeunlock(wpipe); 1316 error = msleep(wpipe, PIPE_MTX(rpipe), 1317 PRIBIO | PCATCH, "pipewr", 0); 1318 pipelock(wpipe, 0); 1319 if (error != 0) 1320 break; 1321 continue; 1322 } 1323 } 1324 1325 --wpipe->pipe_busy; 1326 1327 if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) { 1328 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR); 1329 wakeup(wpipe); 1330 } else if (wpipe->pipe_buffer.cnt > 0) { 1331 /* 1332 * If we have put any characters in the buffer, we wake up 1333 * the reader. 1334 */ 1335 if (wpipe->pipe_state & PIPE_WANTR) { 1336 wpipe->pipe_state &= ~PIPE_WANTR; 1337 wakeup(wpipe); 1338 } 1339 } 1340 1341 /* 1342 * Don't return EPIPE if any byte was written. 1343 * EINTR and other interrupts are handled by generic I/O layer. 1344 * Do not pretend that I/O succeeded for obvious user error 1345 * like EFAULT. 1346 */ 1347 if (uio->uio_resid != orig_resid && error == EPIPE) 1348 error = 0; 1349 1350 if (error == 0) 1351 pipe_timestamp(&wpipe->pipe_mtime); 1352 1353 /* 1354 * We have something to offer, 1355 * wake up select/poll. 1356 */ 1357 if (wpipe->pipe_buffer.cnt) 1358 pipeselwakeup(wpipe); 1359 1360 pipeunlock(wpipe); 1361 PIPE_UNLOCK(rpipe); 1362 if (uio->uio_resid != orig_resid) 1363 td->td_ru.ru_msgsnd++; 1364 return (error); 1365 } 1366 1367 /* ARGSUSED */ 1368 static int 1369 pipe_truncate(struct file *fp, off_t length, struct ucred *active_cred, 1370 struct thread *td) 1371 { 1372 struct pipe *cpipe; 1373 int error; 1374 1375 cpipe = fp->f_data; 1376 if (cpipe->pipe_type & PIPE_TYPE_NAMED) 1377 error = vnops.fo_truncate(fp, length, active_cred, td); 1378 else 1379 error = invfo_truncate(fp, length, active_cred, td); 1380 return (error); 1381 } 1382 1383 /* 1384 * we implement a very minimal set of ioctls for compatibility with sockets. 1385 */ 1386 static int 1387 pipe_ioctl(struct file *fp, u_long cmd, void *data, struct ucred *active_cred, 1388 struct thread *td) 1389 { 1390 struct pipe *mpipe = fp->f_data; 1391 int error; 1392 1393 PIPE_LOCK(mpipe); 1394 1395 #ifdef MAC 1396 error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data); 1397 if (error) { 1398 PIPE_UNLOCK(mpipe); 1399 return (error); 1400 } 1401 #endif 1402 1403 error = 0; 1404 switch (cmd) { 1405 case FIONBIO: 1406 break; 1407 1408 case FIOASYNC: 1409 if (*(int *)data) { 1410 mpipe->pipe_state |= PIPE_ASYNC; 1411 } else { 1412 mpipe->pipe_state &= ~PIPE_ASYNC; 1413 } 1414 break; 1415 1416 case FIONREAD: 1417 if (!(fp->f_flag & FREAD)) { 1418 *(int *)data = 0; 1419 PIPE_UNLOCK(mpipe); 1420 return (0); 1421 } 1422 if (mpipe->pipe_pages.cnt != 0) 1423 *(int *)data = mpipe->pipe_pages.cnt; 1424 else 1425 *(int *)data = mpipe->pipe_buffer.cnt; 1426 break; 1427 1428 case FIOSETOWN: 1429 PIPE_UNLOCK(mpipe); 1430 error = fsetown(*(int *)data, &mpipe->pipe_sigio); 1431 goto out_unlocked; 1432 1433 case FIOGETOWN: 1434 *(int *)data = fgetown(&mpipe->pipe_sigio); 1435 break; 1436 1437 /* This is deprecated, FIOSETOWN should be used instead. */ 1438 case TIOCSPGRP: 1439 PIPE_UNLOCK(mpipe); 1440 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio); 1441 goto out_unlocked; 1442 1443 /* This is deprecated, FIOGETOWN should be used instead. */ 1444 case TIOCGPGRP: 1445 *(int *)data = -fgetown(&mpipe->pipe_sigio); 1446 break; 1447 1448 default: 1449 error = ENOTTY; 1450 break; 1451 } 1452 PIPE_UNLOCK(mpipe); 1453 out_unlocked: 1454 return (error); 1455 } 1456 1457 static int 1458 pipe_poll(struct file *fp, int events, struct ucred *active_cred, 1459 struct thread *td) 1460 { 1461 struct pipe *rpipe; 1462 struct pipe *wpipe; 1463 int levents, revents; 1464 #ifdef MAC 1465 int error; 1466 #endif 1467 1468 revents = 0; 1469 rpipe = fp->f_data; 1470 wpipe = PIPE_PEER(rpipe); 1471 PIPE_LOCK(rpipe); 1472 #ifdef MAC 1473 error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair); 1474 if (error) 1475 goto locked_error; 1476 #endif 1477 if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM)) 1478 if (rpipe->pipe_pages.cnt > 0 || rpipe->pipe_buffer.cnt > 0) 1479 revents |= events & (POLLIN | POLLRDNORM); 1480 1481 if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM)) 1482 if (wpipe->pipe_present != PIPE_ACTIVE || 1483 (wpipe->pipe_state & PIPE_EOF) || 1484 ((wpipe->pipe_state & PIPE_DIRECTW) == 0 && 1485 ((wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF || 1486 wpipe->pipe_buffer.size == 0))) 1487 revents |= events & (POLLOUT | POLLWRNORM); 1488 1489 levents = events & 1490 (POLLIN | POLLINIGNEOF | POLLPRI | POLLRDNORM | POLLRDBAND); 1491 if (rpipe->pipe_type & PIPE_TYPE_NAMED && fp->f_flag & FREAD && levents && 1492 fp->f_pipegen == rpipe->pipe_wgen) 1493 events |= POLLINIGNEOF; 1494 1495 if ((events & POLLINIGNEOF) == 0) { 1496 if (rpipe->pipe_state & PIPE_EOF) { 1497 if (fp->f_flag & FREAD) 1498 revents |= (events & (POLLIN | POLLRDNORM)); 1499 if (wpipe->pipe_present != PIPE_ACTIVE || 1500 (wpipe->pipe_state & PIPE_EOF)) 1501 revents |= POLLHUP; 1502 } 1503 } 1504 1505 if (revents == 0) { 1506 /* 1507 * Add ourselves regardless of eventmask as we have to return 1508 * POLLHUP even if it was not asked for. 1509 */ 1510 if ((fp->f_flag & FREAD) != 0) { 1511 selrecord(td, &rpipe->pipe_sel); 1512 if (SEL_WAITING(&rpipe->pipe_sel)) 1513 rpipe->pipe_state |= PIPE_SEL; 1514 } 1515 1516 if ((fp->f_flag & FWRITE) != 0 && 1517 wpipe->pipe_present == PIPE_ACTIVE) { 1518 selrecord(td, &wpipe->pipe_sel); 1519 if (SEL_WAITING(&wpipe->pipe_sel)) 1520 wpipe->pipe_state |= PIPE_SEL; 1521 } 1522 } 1523 #ifdef MAC 1524 locked_error: 1525 #endif 1526 PIPE_UNLOCK(rpipe); 1527 1528 return (revents); 1529 } 1530 1531 /* 1532 * We shouldn't need locks here as we're doing a read and this should 1533 * be a natural race. 1534 */ 1535 static int 1536 pipe_stat(struct file *fp, struct stat *ub, struct ucred *active_cred) 1537 { 1538 struct pipe *pipe; 1539 #ifdef MAC 1540 int error; 1541 #endif 1542 1543 pipe = fp->f_data; 1544 #ifdef MAC 1545 if (mac_pipe_check_stat_enabled()) { 1546 PIPE_LOCK(pipe); 1547 error = mac_pipe_check_stat(active_cred, pipe->pipe_pair); 1548 PIPE_UNLOCK(pipe); 1549 if (error) { 1550 return (error); 1551 } 1552 } 1553 #endif 1554 1555 /* For named pipes ask the underlying filesystem. */ 1556 if (pipe->pipe_type & PIPE_TYPE_NAMED) { 1557 return (vnops.fo_stat(fp, ub, active_cred)); 1558 } 1559 1560 bzero(ub, sizeof(*ub)); 1561 ub->st_mode = S_IFIFO; 1562 ub->st_blksize = PAGE_SIZE; 1563 if (pipe->pipe_pages.cnt != 0) 1564 ub->st_size = pipe->pipe_pages.cnt; 1565 else 1566 ub->st_size = pipe->pipe_buffer.cnt; 1567 ub->st_blocks = howmany(ub->st_size, ub->st_blksize); 1568 ub->st_atim = pipe->pipe_atime; 1569 ub->st_mtim = pipe->pipe_mtime; 1570 ub->st_ctim = pipe->pipe_ctime; 1571 ub->st_uid = fp->f_cred->cr_uid; 1572 ub->st_gid = fp->f_cred->cr_gid; 1573 ub->st_dev = pipedev_ino; 1574 ub->st_ino = pipe->pipe_ino; 1575 /* 1576 * Left as 0: st_nlink, st_rdev, st_flags, st_gen. 1577 */ 1578 return (0); 1579 } 1580 1581 /* ARGSUSED */ 1582 static int 1583 pipe_close(struct file *fp, struct thread *td) 1584 { 1585 1586 if (fp->f_vnode != NULL) 1587 return vnops.fo_close(fp, td); 1588 fp->f_ops = &badfileops; 1589 pipe_dtor(fp->f_data); 1590 fp->f_data = NULL; 1591 return (0); 1592 } 1593 1594 static int 1595 pipe_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td) 1596 { 1597 struct pipe *cpipe; 1598 int error; 1599 1600 cpipe = fp->f_data; 1601 if (cpipe->pipe_type & PIPE_TYPE_NAMED) 1602 error = vn_chmod(fp, mode, active_cred, td); 1603 else 1604 error = invfo_chmod(fp, mode, active_cred, td); 1605 return (error); 1606 } 1607 1608 static int 1609 pipe_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred, 1610 struct thread *td) 1611 { 1612 struct pipe *cpipe; 1613 int error; 1614 1615 cpipe = fp->f_data; 1616 if (cpipe->pipe_type & PIPE_TYPE_NAMED) 1617 error = vn_chown(fp, uid, gid, active_cred, td); 1618 else 1619 error = invfo_chown(fp, uid, gid, active_cred, td); 1620 return (error); 1621 } 1622 1623 static int 1624 pipe_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp) 1625 { 1626 struct pipe *pi; 1627 1628 if (fp->f_type == DTYPE_FIFO) 1629 return (vn_fill_kinfo(fp, kif, fdp)); 1630 kif->kf_type = KF_TYPE_PIPE; 1631 pi = fp->f_data; 1632 kif->kf_un.kf_pipe.kf_pipe_addr = (uintptr_t)pi; 1633 kif->kf_un.kf_pipe.kf_pipe_peer = (uintptr_t)pi->pipe_peer; 1634 kif->kf_un.kf_pipe.kf_pipe_buffer_cnt = pi->pipe_buffer.cnt; 1635 kif->kf_un.kf_pipe.kf_pipe_buffer_in = pi->pipe_buffer.in; 1636 kif->kf_un.kf_pipe.kf_pipe_buffer_out = pi->pipe_buffer.out; 1637 kif->kf_un.kf_pipe.kf_pipe_buffer_size = pi->pipe_buffer.size; 1638 return (0); 1639 } 1640 1641 static void 1642 pipe_free_kmem(struct pipe *cpipe) 1643 { 1644 1645 KASSERT(!mtx_owned(PIPE_MTX(cpipe)), 1646 ("pipe_free_kmem: pipe mutex locked")); 1647 1648 if (cpipe->pipe_buffer.buffer != NULL) { 1649 atomic_subtract_long(&amountpipekva, cpipe->pipe_buffer.size); 1650 vm_map_remove(pipe_map, 1651 (vm_offset_t)cpipe->pipe_buffer.buffer, 1652 (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size); 1653 cpipe->pipe_buffer.buffer = NULL; 1654 } 1655 #ifndef PIPE_NODIRECT 1656 { 1657 cpipe->pipe_pages.cnt = 0; 1658 cpipe->pipe_pages.pos = 0; 1659 cpipe->pipe_pages.npages = 0; 1660 } 1661 #endif 1662 } 1663 1664 /* 1665 * shutdown the pipe 1666 */ 1667 static void 1668 pipeclose(struct pipe *cpipe) 1669 { 1670 #ifdef MAC 1671 struct pipepair *pp; 1672 #endif 1673 struct pipe *ppipe; 1674 1675 KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL")); 1676 1677 PIPE_LOCK(cpipe); 1678 pipelock(cpipe, 0); 1679 #ifdef MAC 1680 pp = cpipe->pipe_pair; 1681 #endif 1682 1683 /* 1684 * If the other side is blocked, wake it up saying that 1685 * we want to close it down. 1686 */ 1687 cpipe->pipe_state |= PIPE_EOF; 1688 while (cpipe->pipe_busy) { 1689 wakeup(cpipe); 1690 cpipe->pipe_state |= PIPE_WANT; 1691 pipeunlock(cpipe); 1692 msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0); 1693 pipelock(cpipe, 0); 1694 } 1695 1696 pipeselwakeup(cpipe); 1697 1698 /* 1699 * Disconnect from peer, if any. 1700 */ 1701 ppipe = cpipe->pipe_peer; 1702 if (ppipe->pipe_present == PIPE_ACTIVE) { 1703 ppipe->pipe_state |= PIPE_EOF; 1704 wakeup(ppipe); 1705 pipeselwakeup(ppipe); 1706 } 1707 1708 /* 1709 * Mark this endpoint as free. Release kmem resources. We 1710 * don't mark this endpoint as unused until we've finished 1711 * doing that, or the pipe might disappear out from under 1712 * us. 1713 */ 1714 PIPE_UNLOCK(cpipe); 1715 pipe_free_kmem(cpipe); 1716 PIPE_LOCK(cpipe); 1717 cpipe->pipe_present = PIPE_CLOSING; 1718 pipeunlock(cpipe); 1719 1720 /* 1721 * knlist_clear() may sleep dropping the PIPE_MTX. Set the 1722 * PIPE_FINALIZED, that allows other end to free the 1723 * pipe_pair, only after the knotes are completely dismantled. 1724 */ 1725 knlist_clear(&cpipe->pipe_sel.si_note, 1); 1726 cpipe->pipe_present = PIPE_FINALIZED; 1727 seldrain(&cpipe->pipe_sel); 1728 knlist_destroy(&cpipe->pipe_sel.si_note); 1729 1730 /* 1731 * If both endpoints are now closed, release the memory for the 1732 * pipe pair. If not, unlock. 1733 */ 1734 if (ppipe->pipe_present == PIPE_FINALIZED) { 1735 PIPE_UNLOCK(cpipe); 1736 #ifdef MAC 1737 mac_pipe_destroy(pp); 1738 #endif 1739 uma_zfree(pipe_zone, cpipe->pipe_pair); 1740 } else 1741 PIPE_UNLOCK(cpipe); 1742 } 1743 1744 /*ARGSUSED*/ 1745 static int 1746 pipe_kqfilter(struct file *fp, struct knote *kn) 1747 { 1748 struct pipe *cpipe; 1749 1750 /* 1751 * If a filter is requested that is not supported by this file 1752 * descriptor, don't return an error, but also don't ever generate an 1753 * event. 1754 */ 1755 if ((kn->kn_filter == EVFILT_READ) && !(fp->f_flag & FREAD)) { 1756 kn->kn_fop = &pipe_nfiltops; 1757 return (0); 1758 } 1759 if ((kn->kn_filter == EVFILT_WRITE) && !(fp->f_flag & FWRITE)) { 1760 kn->kn_fop = &pipe_nfiltops; 1761 return (0); 1762 } 1763 cpipe = fp->f_data; 1764 PIPE_LOCK(cpipe); 1765 switch (kn->kn_filter) { 1766 case EVFILT_READ: 1767 kn->kn_fop = &pipe_rfiltops; 1768 break; 1769 case EVFILT_WRITE: 1770 kn->kn_fop = &pipe_wfiltops; 1771 if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) { 1772 /* other end of pipe has been closed */ 1773 PIPE_UNLOCK(cpipe); 1774 return (EPIPE); 1775 } 1776 cpipe = PIPE_PEER(cpipe); 1777 break; 1778 default: 1779 if ((cpipe->pipe_type & PIPE_TYPE_NAMED) != 0) { 1780 PIPE_UNLOCK(cpipe); 1781 return (vnops.fo_kqfilter(fp, kn)); 1782 } 1783 PIPE_UNLOCK(cpipe); 1784 return (EINVAL); 1785 } 1786 1787 kn->kn_hook = cpipe; 1788 knlist_add(&cpipe->pipe_sel.si_note, kn, 1); 1789 PIPE_UNLOCK(cpipe); 1790 return (0); 1791 } 1792 1793 static void 1794 filt_pipedetach(struct knote *kn) 1795 { 1796 struct pipe *cpipe = kn->kn_hook; 1797 1798 PIPE_LOCK(cpipe); 1799 knlist_remove(&cpipe->pipe_sel.si_note, kn, 1); 1800 PIPE_UNLOCK(cpipe); 1801 } 1802 1803 /*ARGSUSED*/ 1804 static int 1805 filt_piperead(struct knote *kn, long hint) 1806 { 1807 struct file *fp = kn->kn_fp; 1808 struct pipe *rpipe = kn->kn_hook; 1809 1810 PIPE_LOCK_ASSERT(rpipe, MA_OWNED); 1811 kn->kn_data = rpipe->pipe_buffer.cnt; 1812 if (kn->kn_data == 0) 1813 kn->kn_data = rpipe->pipe_pages.cnt; 1814 1815 if ((rpipe->pipe_state & PIPE_EOF) != 0 && 1816 ((rpipe->pipe_type & PIPE_TYPE_NAMED) == 0 || 1817 fp->f_pipegen != rpipe->pipe_wgen)) { 1818 kn->kn_flags |= EV_EOF; 1819 return (1); 1820 } 1821 kn->kn_flags &= ~EV_EOF; 1822 return (kn->kn_data > 0); 1823 } 1824 1825 /*ARGSUSED*/ 1826 static int 1827 filt_pipewrite(struct knote *kn, long hint) 1828 { 1829 struct pipe *wpipe = kn->kn_hook; 1830 1831 /* 1832 * If this end of the pipe is closed, the knote was removed from the 1833 * knlist and the list lock (i.e., the pipe lock) is therefore not held. 1834 */ 1835 if (wpipe->pipe_present == PIPE_ACTIVE || 1836 (wpipe->pipe_type & PIPE_TYPE_NAMED) != 0) { 1837 PIPE_LOCK_ASSERT(wpipe, MA_OWNED); 1838 1839 if (wpipe->pipe_state & PIPE_DIRECTW) { 1840 kn->kn_data = 0; 1841 } else if (wpipe->pipe_buffer.size > 0) { 1842 kn->kn_data = wpipe->pipe_buffer.size - 1843 wpipe->pipe_buffer.cnt; 1844 } else { 1845 kn->kn_data = PIPE_BUF; 1846 } 1847 } 1848 1849 if (wpipe->pipe_present != PIPE_ACTIVE || 1850 (wpipe->pipe_state & PIPE_EOF)) { 1851 kn->kn_flags |= EV_EOF; 1852 return (1); 1853 } 1854 kn->kn_flags &= ~EV_EOF; 1855 return (kn->kn_data >= PIPE_BUF); 1856 } 1857 1858 static void 1859 filt_pipedetach_notsup(struct knote *kn) 1860 { 1861 1862 } 1863 1864 static int 1865 filt_pipenotsup(struct knote *kn, long hint) 1866 { 1867 1868 return (0); 1869 } 1870