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 return (0); 739 if (atomic_load_int(&rpipe->pipe_buffer.cnt) == 0 && 740 atomic_load_int(&rpipe->pipe_pages.cnt) == 0) 741 return (EAGAIN); 742 } 743 744 PIPE_LOCK(rpipe); 745 ++rpipe->pipe_busy; 746 error = pipelock(rpipe, 1); 747 if (error) 748 goto unlocked_error; 749 750 #ifdef MAC 751 error = mac_pipe_check_read(active_cred, rpipe->pipe_pair); 752 if (error) 753 goto locked_error; 754 #endif 755 if (amountpipekva > (3 * maxpipekva) / 4) { 756 if ((rpipe->pipe_state & PIPE_DIRECTW) == 0 && 757 rpipe->pipe_buffer.size > SMALL_PIPE_SIZE && 758 rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE && 759 piperesizeallowed == 1) { 760 PIPE_UNLOCK(rpipe); 761 pipespace(rpipe, SMALL_PIPE_SIZE); 762 PIPE_LOCK(rpipe); 763 } 764 } 765 766 while (uio->uio_resid) { 767 /* 768 * normal pipe buffer receive 769 */ 770 if (rpipe->pipe_buffer.cnt > 0) { 771 size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out; 772 if (size > rpipe->pipe_buffer.cnt) 773 size = rpipe->pipe_buffer.cnt; 774 if (size > uio->uio_resid) 775 size = uio->uio_resid; 776 777 PIPE_UNLOCK(rpipe); 778 error = uiomove( 779 &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out], 780 size, uio); 781 PIPE_LOCK(rpipe); 782 if (error) 783 break; 784 785 rpipe->pipe_buffer.out += size; 786 if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size) 787 rpipe->pipe_buffer.out = 0; 788 789 rpipe->pipe_buffer.cnt -= size; 790 791 /* 792 * If there is no more to read in the pipe, reset 793 * its pointers to the beginning. This improves 794 * cache hit stats. 795 */ 796 if (rpipe->pipe_buffer.cnt == 0) { 797 rpipe->pipe_buffer.in = 0; 798 rpipe->pipe_buffer.out = 0; 799 } 800 nread += size; 801 #ifndef PIPE_NODIRECT 802 /* 803 * Direct copy, bypassing a kernel buffer. 804 */ 805 } else if ((size = rpipe->pipe_pages.cnt) != 0) { 806 if (size > uio->uio_resid) 807 size = (u_int) uio->uio_resid; 808 PIPE_UNLOCK(rpipe); 809 error = uiomove_fromphys(rpipe->pipe_pages.ms, 810 rpipe->pipe_pages.pos, size, uio); 811 PIPE_LOCK(rpipe); 812 if (error) 813 break; 814 nread += size; 815 rpipe->pipe_pages.pos += size; 816 rpipe->pipe_pages.cnt -= size; 817 if (rpipe->pipe_pages.cnt == 0) { 818 rpipe->pipe_state &= ~PIPE_WANTW; 819 wakeup(rpipe); 820 } 821 #endif 822 } else { 823 /* 824 * detect EOF condition 825 * read returns 0 on EOF, no need to set error 826 */ 827 if (rpipe->pipe_state & PIPE_EOF) 828 break; 829 830 /* 831 * If the "write-side" has been blocked, wake it up now. 832 */ 833 if (rpipe->pipe_state & PIPE_WANTW) { 834 rpipe->pipe_state &= ~PIPE_WANTW; 835 wakeup(rpipe); 836 } 837 838 /* 839 * Break if some data was read. 840 */ 841 if (nread > 0) 842 break; 843 844 /* 845 * Unlock the pipe buffer for our remaining processing. 846 * We will either break out with an error or we will 847 * sleep and relock to loop. 848 */ 849 pipeunlock(rpipe); 850 851 /* 852 * Handle non-blocking mode operation or 853 * wait for more data. 854 */ 855 if (fp->f_flag & FNONBLOCK) { 856 error = EAGAIN; 857 } else { 858 rpipe->pipe_state |= PIPE_WANTR; 859 if ((error = msleep(rpipe, PIPE_MTX(rpipe), 860 PRIBIO | PCATCH, 861 "piperd", 0)) == 0) 862 error = pipelock(rpipe, 1); 863 } 864 if (error) 865 goto unlocked_error; 866 } 867 } 868 #ifdef MAC 869 locked_error: 870 #endif 871 pipeunlock(rpipe); 872 873 /* XXX: should probably do this before getting any locks. */ 874 if (error == 0) 875 pipe_timestamp(&rpipe->pipe_atime); 876 unlocked_error: 877 --rpipe->pipe_busy; 878 879 /* 880 * PIPE_WANT processing only makes sense if pipe_busy is 0. 881 */ 882 if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) { 883 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW); 884 wakeup(rpipe); 885 } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) { 886 /* 887 * Handle write blocking hysteresis. 888 */ 889 if (rpipe->pipe_state & PIPE_WANTW) { 890 rpipe->pipe_state &= ~PIPE_WANTW; 891 wakeup(rpipe); 892 } 893 } 894 895 /* 896 * Only wake up writers if there was actually something read. 897 * Otherwise, when calling read(2) at EOF, a spurious wakeup occurs. 898 */ 899 if (nread > 0 && 900 rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt >= PIPE_BUF) 901 pipeselwakeup(rpipe); 902 903 PIPE_UNLOCK(rpipe); 904 if (nread > 0) 905 td->td_ru.ru_msgrcv++; 906 return (error); 907 } 908 909 #ifndef PIPE_NODIRECT 910 /* 911 * Map the sending processes' buffer into kernel space and wire it. 912 * This is similar to a physical write operation. 913 */ 914 static int 915 pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio) 916 { 917 u_int size; 918 int i; 919 920 PIPE_LOCK_ASSERT(wpipe, MA_OWNED); 921 KASSERT((wpipe->pipe_state & PIPE_DIRECTW) == 0, 922 ("%s: PIPE_DIRECTW set on %p", __func__, wpipe)); 923 KASSERT(wpipe->pipe_pages.cnt == 0, 924 ("%s: pipe map for %p contains residual data", __func__, wpipe)); 925 926 if (uio->uio_iov->iov_len > wpipe->pipe_buffer.size) 927 size = wpipe->pipe_buffer.size; 928 else 929 size = uio->uio_iov->iov_len; 930 931 wpipe->pipe_state |= PIPE_DIRECTW; 932 PIPE_UNLOCK(wpipe); 933 i = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map, 934 (vm_offset_t)uio->uio_iov->iov_base, size, VM_PROT_READ, 935 wpipe->pipe_pages.ms, PIPENPAGES); 936 PIPE_LOCK(wpipe); 937 if (i < 0) { 938 wpipe->pipe_state &= ~PIPE_DIRECTW; 939 return (EFAULT); 940 } 941 942 wpipe->pipe_pages.npages = i; 943 wpipe->pipe_pages.pos = 944 ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK; 945 wpipe->pipe_pages.cnt = size; 946 947 uio->uio_iov->iov_len -= size; 948 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size; 949 if (uio->uio_iov->iov_len == 0) 950 uio->uio_iov++; 951 uio->uio_resid -= size; 952 uio->uio_offset += size; 953 return (0); 954 } 955 956 /* 957 * Unwire the process buffer. 958 */ 959 static void 960 pipe_destroy_write_buffer(struct pipe *wpipe) 961 { 962 963 PIPE_LOCK_ASSERT(wpipe, MA_OWNED); 964 KASSERT((wpipe->pipe_state & PIPE_DIRECTW) != 0, 965 ("%s: PIPE_DIRECTW not set on %p", __func__, wpipe)); 966 KASSERT(wpipe->pipe_pages.cnt == 0, 967 ("%s: pipe map for %p contains residual data", __func__, wpipe)); 968 969 wpipe->pipe_state &= ~PIPE_DIRECTW; 970 vm_page_unhold_pages(wpipe->pipe_pages.ms, wpipe->pipe_pages.npages); 971 wpipe->pipe_pages.npages = 0; 972 } 973 974 /* 975 * In the case of a signal, the writing process might go away. This 976 * code copies the data into the circular buffer so that the source 977 * pages can be freed without loss of data. 978 */ 979 static void 980 pipe_clone_write_buffer(struct pipe *wpipe) 981 { 982 struct uio uio; 983 struct iovec iov; 984 int size; 985 int pos; 986 987 PIPE_LOCK_ASSERT(wpipe, MA_OWNED); 988 KASSERT((wpipe->pipe_state & PIPE_DIRECTW) != 0, 989 ("%s: PIPE_DIRECTW not set on %p", __func__, wpipe)); 990 991 size = wpipe->pipe_pages.cnt; 992 pos = wpipe->pipe_pages.pos; 993 wpipe->pipe_pages.cnt = 0; 994 995 wpipe->pipe_buffer.in = size; 996 wpipe->pipe_buffer.out = 0; 997 wpipe->pipe_buffer.cnt = size; 998 999 PIPE_UNLOCK(wpipe); 1000 iov.iov_base = wpipe->pipe_buffer.buffer; 1001 iov.iov_len = size; 1002 uio.uio_iov = &iov; 1003 uio.uio_iovcnt = 1; 1004 uio.uio_offset = 0; 1005 uio.uio_resid = size; 1006 uio.uio_segflg = UIO_SYSSPACE; 1007 uio.uio_rw = UIO_READ; 1008 uio.uio_td = curthread; 1009 uiomove_fromphys(wpipe->pipe_pages.ms, pos, size, &uio); 1010 PIPE_LOCK(wpipe); 1011 pipe_destroy_write_buffer(wpipe); 1012 } 1013 1014 /* 1015 * This implements the pipe buffer write mechanism. Note that only 1016 * a direct write OR a normal pipe write can be pending at any given time. 1017 * If there are any characters in the pipe buffer, the direct write will 1018 * be deferred until the receiving process grabs all of the bytes from 1019 * the pipe buffer. Then the direct mapping write is set-up. 1020 */ 1021 static int 1022 pipe_direct_write(struct pipe *wpipe, struct uio *uio) 1023 { 1024 int error; 1025 1026 retry: 1027 PIPE_LOCK_ASSERT(wpipe, MA_OWNED); 1028 if ((wpipe->pipe_state & PIPE_EOF) != 0) { 1029 error = EPIPE; 1030 goto error1; 1031 } 1032 if (wpipe->pipe_state & PIPE_DIRECTW) { 1033 if (wpipe->pipe_state & PIPE_WANTR) { 1034 wpipe->pipe_state &= ~PIPE_WANTR; 1035 wakeup(wpipe); 1036 } 1037 pipeselwakeup(wpipe); 1038 wpipe->pipe_state |= PIPE_WANTW; 1039 pipeunlock(wpipe); 1040 error = msleep(wpipe, PIPE_MTX(wpipe), 1041 PRIBIO | PCATCH, "pipdww", 0); 1042 pipelock(wpipe, 0); 1043 if (error != 0) 1044 goto error1; 1045 goto retry; 1046 } 1047 if (wpipe->pipe_buffer.cnt > 0) { 1048 if (wpipe->pipe_state & PIPE_WANTR) { 1049 wpipe->pipe_state &= ~PIPE_WANTR; 1050 wakeup(wpipe); 1051 } 1052 pipeselwakeup(wpipe); 1053 wpipe->pipe_state |= PIPE_WANTW; 1054 pipeunlock(wpipe); 1055 error = msleep(wpipe, PIPE_MTX(wpipe), 1056 PRIBIO | PCATCH, "pipdwc", 0); 1057 pipelock(wpipe, 0); 1058 if (error != 0) 1059 goto error1; 1060 goto retry; 1061 } 1062 1063 error = pipe_build_write_buffer(wpipe, uio); 1064 if (error) { 1065 goto error1; 1066 } 1067 1068 while (wpipe->pipe_pages.cnt != 0 && 1069 (wpipe->pipe_state & PIPE_EOF) == 0) { 1070 if (wpipe->pipe_state & PIPE_WANTR) { 1071 wpipe->pipe_state &= ~PIPE_WANTR; 1072 wakeup(wpipe); 1073 } 1074 pipeselwakeup(wpipe); 1075 wpipe->pipe_state |= PIPE_WANTW; 1076 pipeunlock(wpipe); 1077 error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH, 1078 "pipdwt", 0); 1079 pipelock(wpipe, 0); 1080 if (error != 0) 1081 break; 1082 } 1083 1084 if ((wpipe->pipe_state & PIPE_EOF) != 0) { 1085 wpipe->pipe_pages.cnt = 0; 1086 pipe_destroy_write_buffer(wpipe); 1087 pipeselwakeup(wpipe); 1088 error = EPIPE; 1089 } else if (error == EINTR || error == ERESTART) { 1090 pipe_clone_write_buffer(wpipe); 1091 } else { 1092 pipe_destroy_write_buffer(wpipe); 1093 } 1094 KASSERT((wpipe->pipe_state & PIPE_DIRECTW) == 0, 1095 ("pipe %p leaked PIPE_DIRECTW", wpipe)); 1096 return (error); 1097 1098 error1: 1099 wakeup(wpipe); 1100 return (error); 1101 } 1102 #endif 1103 1104 static int 1105 pipe_write(struct file *fp, struct uio *uio, struct ucred *active_cred, 1106 int flags, struct thread *td) 1107 { 1108 struct pipe *wpipe, *rpipe; 1109 ssize_t orig_resid; 1110 int desiredsize, error; 1111 1112 rpipe = fp->f_data; 1113 wpipe = PIPE_PEER(rpipe); 1114 PIPE_LOCK(rpipe); 1115 error = pipelock(wpipe, 1); 1116 if (error) { 1117 PIPE_UNLOCK(rpipe); 1118 return (error); 1119 } 1120 /* 1121 * detect loss of pipe read side, issue SIGPIPE if lost. 1122 */ 1123 if (wpipe->pipe_present != PIPE_ACTIVE || 1124 (wpipe->pipe_state & PIPE_EOF)) { 1125 pipeunlock(wpipe); 1126 PIPE_UNLOCK(rpipe); 1127 return (EPIPE); 1128 } 1129 #ifdef MAC 1130 error = mac_pipe_check_write(active_cred, wpipe->pipe_pair); 1131 if (error) { 1132 pipeunlock(wpipe); 1133 PIPE_UNLOCK(rpipe); 1134 return (error); 1135 } 1136 #endif 1137 ++wpipe->pipe_busy; 1138 1139 /* Choose a larger size if it's advantageous */ 1140 desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size); 1141 while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) { 1142 if (piperesizeallowed != 1) 1143 break; 1144 if (amountpipekva > maxpipekva / 2) 1145 break; 1146 if (desiredsize == BIG_PIPE_SIZE) 1147 break; 1148 desiredsize = desiredsize * 2; 1149 } 1150 1151 /* Choose a smaller size if we're in a OOM situation */ 1152 if (amountpipekva > (3 * maxpipekva) / 4 && 1153 wpipe->pipe_buffer.size > SMALL_PIPE_SIZE && 1154 wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE && 1155 piperesizeallowed == 1) 1156 desiredsize = SMALL_PIPE_SIZE; 1157 1158 /* Resize if the above determined that a new size was necessary */ 1159 if (desiredsize != wpipe->pipe_buffer.size && 1160 (wpipe->pipe_state & PIPE_DIRECTW) == 0) { 1161 PIPE_UNLOCK(wpipe); 1162 pipespace(wpipe, desiredsize); 1163 PIPE_LOCK(wpipe); 1164 } 1165 MPASS(wpipe->pipe_buffer.size != 0); 1166 1167 orig_resid = uio->uio_resid; 1168 1169 while (uio->uio_resid) { 1170 int space; 1171 1172 if (wpipe->pipe_state & PIPE_EOF) { 1173 error = EPIPE; 1174 break; 1175 } 1176 #ifndef PIPE_NODIRECT 1177 /* 1178 * If the transfer is large, we can gain performance if 1179 * we do process-to-process copies directly. 1180 * If the write is non-blocking, we don't use the 1181 * direct write mechanism. 1182 * 1183 * The direct write mechanism will detect the reader going 1184 * away on us. 1185 */ 1186 if (uio->uio_segflg == UIO_USERSPACE && 1187 uio->uio_iov->iov_len >= pipe_mindirect && 1188 wpipe->pipe_buffer.size >= pipe_mindirect && 1189 (fp->f_flag & FNONBLOCK) == 0) { 1190 error = pipe_direct_write(wpipe, uio); 1191 if (error != 0) 1192 break; 1193 continue; 1194 } 1195 #endif 1196 1197 /* 1198 * Pipe buffered writes cannot be coincidental with 1199 * direct writes. We wait until the currently executing 1200 * direct write is completed before we start filling the 1201 * pipe buffer. We break out if a signal occurs or the 1202 * reader goes away. 1203 */ 1204 if (wpipe->pipe_pages.cnt != 0) { 1205 if (wpipe->pipe_state & PIPE_WANTR) { 1206 wpipe->pipe_state &= ~PIPE_WANTR; 1207 wakeup(wpipe); 1208 } 1209 pipeselwakeup(wpipe); 1210 wpipe->pipe_state |= PIPE_WANTW; 1211 pipeunlock(wpipe); 1212 error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH, 1213 "pipbww", 0); 1214 pipelock(wpipe, 0); 1215 if (error != 0) 1216 break; 1217 continue; 1218 } 1219 1220 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt; 1221 1222 /* Writes of size <= PIPE_BUF must be atomic. */ 1223 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF)) 1224 space = 0; 1225 1226 if (space > 0) { 1227 int size; /* Transfer size */ 1228 int segsize; /* first segment to transfer */ 1229 1230 /* 1231 * Transfer size is minimum of uio transfer 1232 * and free space in pipe buffer. 1233 */ 1234 if (space > uio->uio_resid) 1235 size = uio->uio_resid; 1236 else 1237 size = space; 1238 /* 1239 * First segment to transfer is minimum of 1240 * transfer size and contiguous space in 1241 * pipe buffer. If first segment to transfer 1242 * is less than the transfer size, we've got 1243 * a wraparound in the buffer. 1244 */ 1245 segsize = wpipe->pipe_buffer.size - 1246 wpipe->pipe_buffer.in; 1247 if (segsize > size) 1248 segsize = size; 1249 1250 /* Transfer first segment */ 1251 1252 PIPE_UNLOCK(rpipe); 1253 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in], 1254 segsize, uio); 1255 PIPE_LOCK(rpipe); 1256 1257 if (error == 0 && segsize < size) { 1258 KASSERT(wpipe->pipe_buffer.in + segsize == 1259 wpipe->pipe_buffer.size, 1260 ("Pipe buffer wraparound disappeared")); 1261 /* 1262 * Transfer remaining part now, to 1263 * support atomic writes. Wraparound 1264 * happened. 1265 */ 1266 1267 PIPE_UNLOCK(rpipe); 1268 error = uiomove( 1269 &wpipe->pipe_buffer.buffer[0], 1270 size - segsize, uio); 1271 PIPE_LOCK(rpipe); 1272 } 1273 if (error == 0) { 1274 wpipe->pipe_buffer.in += size; 1275 if (wpipe->pipe_buffer.in >= 1276 wpipe->pipe_buffer.size) { 1277 KASSERT(wpipe->pipe_buffer.in == 1278 size - segsize + 1279 wpipe->pipe_buffer.size, 1280 ("Expected wraparound bad")); 1281 wpipe->pipe_buffer.in = size - segsize; 1282 } 1283 1284 wpipe->pipe_buffer.cnt += size; 1285 KASSERT(wpipe->pipe_buffer.cnt <= 1286 wpipe->pipe_buffer.size, 1287 ("Pipe buffer overflow")); 1288 } 1289 if (error != 0) 1290 break; 1291 continue; 1292 } else { 1293 /* 1294 * If the "read-side" has been blocked, wake it up now. 1295 */ 1296 if (wpipe->pipe_state & PIPE_WANTR) { 1297 wpipe->pipe_state &= ~PIPE_WANTR; 1298 wakeup(wpipe); 1299 } 1300 1301 /* 1302 * don't block on non-blocking I/O 1303 */ 1304 if (fp->f_flag & FNONBLOCK) { 1305 error = EAGAIN; 1306 break; 1307 } 1308 1309 /* 1310 * We have no more space and have something to offer, 1311 * wake up select/poll. 1312 */ 1313 pipeselwakeup(wpipe); 1314 1315 wpipe->pipe_state |= PIPE_WANTW; 1316 pipeunlock(wpipe); 1317 error = msleep(wpipe, PIPE_MTX(rpipe), 1318 PRIBIO | PCATCH, "pipewr", 0); 1319 pipelock(wpipe, 0); 1320 if (error != 0) 1321 break; 1322 continue; 1323 } 1324 } 1325 1326 --wpipe->pipe_busy; 1327 1328 if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) { 1329 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR); 1330 wakeup(wpipe); 1331 } else if (wpipe->pipe_buffer.cnt > 0) { 1332 /* 1333 * If we have put any characters in the buffer, we wake up 1334 * the reader. 1335 */ 1336 if (wpipe->pipe_state & PIPE_WANTR) { 1337 wpipe->pipe_state &= ~PIPE_WANTR; 1338 wakeup(wpipe); 1339 } 1340 } 1341 1342 /* 1343 * Don't return EPIPE if any byte was written. 1344 * EINTR and other interrupts are handled by generic I/O layer. 1345 * Do not pretend that I/O succeeded for obvious user error 1346 * like EFAULT. 1347 */ 1348 if (uio->uio_resid != orig_resid && error == EPIPE) 1349 error = 0; 1350 1351 if (error == 0) 1352 pipe_timestamp(&wpipe->pipe_mtime); 1353 1354 /* 1355 * We have something to offer, 1356 * wake up select/poll. 1357 */ 1358 if (wpipe->pipe_buffer.cnt) 1359 pipeselwakeup(wpipe); 1360 1361 pipeunlock(wpipe); 1362 PIPE_UNLOCK(rpipe); 1363 if (uio->uio_resid != orig_resid) 1364 td->td_ru.ru_msgsnd++; 1365 return (error); 1366 } 1367 1368 /* ARGSUSED */ 1369 static int 1370 pipe_truncate(struct file *fp, off_t length, struct ucred *active_cred, 1371 struct thread *td) 1372 { 1373 struct pipe *cpipe; 1374 int error; 1375 1376 cpipe = fp->f_data; 1377 if (cpipe->pipe_type & PIPE_TYPE_NAMED) 1378 error = vnops.fo_truncate(fp, length, active_cred, td); 1379 else 1380 error = invfo_truncate(fp, length, active_cred, td); 1381 return (error); 1382 } 1383 1384 /* 1385 * we implement a very minimal set of ioctls for compatibility with sockets. 1386 */ 1387 static int 1388 pipe_ioctl(struct file *fp, u_long cmd, void *data, struct ucred *active_cred, 1389 struct thread *td) 1390 { 1391 struct pipe *mpipe = fp->f_data; 1392 int error; 1393 1394 PIPE_LOCK(mpipe); 1395 1396 #ifdef MAC 1397 error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data); 1398 if (error) { 1399 PIPE_UNLOCK(mpipe); 1400 return (error); 1401 } 1402 #endif 1403 1404 error = 0; 1405 switch (cmd) { 1406 case FIONBIO: 1407 break; 1408 1409 case FIOASYNC: 1410 if (*(int *)data) { 1411 mpipe->pipe_state |= PIPE_ASYNC; 1412 } else { 1413 mpipe->pipe_state &= ~PIPE_ASYNC; 1414 } 1415 break; 1416 1417 case FIONREAD: 1418 if (!(fp->f_flag & FREAD)) { 1419 *(int *)data = 0; 1420 PIPE_UNLOCK(mpipe); 1421 return (0); 1422 } 1423 if (mpipe->pipe_pages.cnt != 0) 1424 *(int *)data = mpipe->pipe_pages.cnt; 1425 else 1426 *(int *)data = mpipe->pipe_buffer.cnt; 1427 break; 1428 1429 case FIOSETOWN: 1430 PIPE_UNLOCK(mpipe); 1431 error = fsetown(*(int *)data, &mpipe->pipe_sigio); 1432 goto out_unlocked; 1433 1434 case FIOGETOWN: 1435 *(int *)data = fgetown(&mpipe->pipe_sigio); 1436 break; 1437 1438 /* This is deprecated, FIOSETOWN should be used instead. */ 1439 case TIOCSPGRP: 1440 PIPE_UNLOCK(mpipe); 1441 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio); 1442 goto out_unlocked; 1443 1444 /* This is deprecated, FIOGETOWN should be used instead. */ 1445 case TIOCGPGRP: 1446 *(int *)data = -fgetown(&mpipe->pipe_sigio); 1447 break; 1448 1449 default: 1450 error = ENOTTY; 1451 break; 1452 } 1453 PIPE_UNLOCK(mpipe); 1454 out_unlocked: 1455 return (error); 1456 } 1457 1458 static int 1459 pipe_poll(struct file *fp, int events, struct ucred *active_cred, 1460 struct thread *td) 1461 { 1462 struct pipe *rpipe; 1463 struct pipe *wpipe; 1464 int levents, revents; 1465 #ifdef MAC 1466 int error; 1467 #endif 1468 1469 revents = 0; 1470 rpipe = fp->f_data; 1471 wpipe = PIPE_PEER(rpipe); 1472 PIPE_LOCK(rpipe); 1473 #ifdef MAC 1474 error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair); 1475 if (error) 1476 goto locked_error; 1477 #endif 1478 if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM)) 1479 if (rpipe->pipe_pages.cnt > 0 || rpipe->pipe_buffer.cnt > 0) 1480 revents |= events & (POLLIN | POLLRDNORM); 1481 1482 if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM)) 1483 if (wpipe->pipe_present != PIPE_ACTIVE || 1484 (wpipe->pipe_state & PIPE_EOF) || 1485 ((wpipe->pipe_state & PIPE_DIRECTW) == 0 && 1486 ((wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF || 1487 wpipe->pipe_buffer.size == 0))) 1488 revents |= events & (POLLOUT | POLLWRNORM); 1489 1490 levents = events & 1491 (POLLIN | POLLINIGNEOF | POLLPRI | POLLRDNORM | POLLRDBAND); 1492 if (rpipe->pipe_type & PIPE_TYPE_NAMED && fp->f_flag & FREAD && levents && 1493 fp->f_pipegen == rpipe->pipe_wgen) 1494 events |= POLLINIGNEOF; 1495 1496 if ((events & POLLINIGNEOF) == 0) { 1497 if (rpipe->pipe_state & PIPE_EOF) { 1498 if (fp->f_flag & FREAD) 1499 revents |= (events & (POLLIN | POLLRDNORM)); 1500 if (wpipe->pipe_present != PIPE_ACTIVE || 1501 (wpipe->pipe_state & PIPE_EOF)) 1502 revents |= POLLHUP; 1503 } 1504 } 1505 1506 if (revents == 0) { 1507 /* 1508 * Add ourselves regardless of eventmask as we have to return 1509 * POLLHUP even if it was not asked for. 1510 */ 1511 if ((fp->f_flag & FREAD) != 0) { 1512 selrecord(td, &rpipe->pipe_sel); 1513 if (SEL_WAITING(&rpipe->pipe_sel)) 1514 rpipe->pipe_state |= PIPE_SEL; 1515 } 1516 1517 if ((fp->f_flag & FWRITE) != 0 && 1518 wpipe->pipe_present == PIPE_ACTIVE) { 1519 selrecord(td, &wpipe->pipe_sel); 1520 if (SEL_WAITING(&wpipe->pipe_sel)) 1521 wpipe->pipe_state |= PIPE_SEL; 1522 } 1523 } 1524 #ifdef MAC 1525 locked_error: 1526 #endif 1527 PIPE_UNLOCK(rpipe); 1528 1529 return (revents); 1530 } 1531 1532 /* 1533 * We shouldn't need locks here as we're doing a read and this should 1534 * be a natural race. 1535 */ 1536 static int 1537 pipe_stat(struct file *fp, struct stat *ub, struct ucred *active_cred) 1538 { 1539 struct pipe *pipe; 1540 #ifdef MAC 1541 int error; 1542 #endif 1543 1544 pipe = fp->f_data; 1545 #ifdef MAC 1546 if (mac_pipe_check_stat_enabled()) { 1547 PIPE_LOCK(pipe); 1548 error = mac_pipe_check_stat(active_cred, pipe->pipe_pair); 1549 PIPE_UNLOCK(pipe); 1550 if (error) { 1551 return (error); 1552 } 1553 } 1554 #endif 1555 1556 /* For named pipes ask the underlying filesystem. */ 1557 if (pipe->pipe_type & PIPE_TYPE_NAMED) { 1558 return (vnops.fo_stat(fp, ub, active_cred)); 1559 } 1560 1561 bzero(ub, sizeof(*ub)); 1562 ub->st_mode = S_IFIFO; 1563 ub->st_blksize = PAGE_SIZE; 1564 if (pipe->pipe_pages.cnt != 0) 1565 ub->st_size = pipe->pipe_pages.cnt; 1566 else 1567 ub->st_size = pipe->pipe_buffer.cnt; 1568 ub->st_blocks = howmany(ub->st_size, ub->st_blksize); 1569 ub->st_atim = pipe->pipe_atime; 1570 ub->st_mtim = pipe->pipe_mtime; 1571 ub->st_ctim = pipe->pipe_ctime; 1572 ub->st_uid = fp->f_cred->cr_uid; 1573 ub->st_gid = fp->f_cred->cr_gid; 1574 ub->st_dev = pipedev_ino; 1575 ub->st_ino = pipe->pipe_ino; 1576 /* 1577 * Left as 0: st_nlink, st_rdev, st_flags, st_gen. 1578 */ 1579 return (0); 1580 } 1581 1582 /* ARGSUSED */ 1583 static int 1584 pipe_close(struct file *fp, struct thread *td) 1585 { 1586 1587 if (fp->f_vnode != NULL) 1588 return vnops.fo_close(fp, td); 1589 fp->f_ops = &badfileops; 1590 pipe_dtor(fp->f_data); 1591 fp->f_data = NULL; 1592 return (0); 1593 } 1594 1595 static int 1596 pipe_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td) 1597 { 1598 struct pipe *cpipe; 1599 int error; 1600 1601 cpipe = fp->f_data; 1602 if (cpipe->pipe_type & PIPE_TYPE_NAMED) 1603 error = vn_chmod(fp, mode, active_cred, td); 1604 else 1605 error = invfo_chmod(fp, mode, active_cred, td); 1606 return (error); 1607 } 1608 1609 static int 1610 pipe_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred, 1611 struct thread *td) 1612 { 1613 struct pipe *cpipe; 1614 int error; 1615 1616 cpipe = fp->f_data; 1617 if (cpipe->pipe_type & PIPE_TYPE_NAMED) 1618 error = vn_chown(fp, uid, gid, active_cred, td); 1619 else 1620 error = invfo_chown(fp, uid, gid, active_cred, td); 1621 return (error); 1622 } 1623 1624 static int 1625 pipe_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp) 1626 { 1627 struct pipe *pi; 1628 1629 if (fp->f_type == DTYPE_FIFO) 1630 return (vn_fill_kinfo(fp, kif, fdp)); 1631 kif->kf_type = KF_TYPE_PIPE; 1632 pi = fp->f_data; 1633 kif->kf_un.kf_pipe.kf_pipe_addr = (uintptr_t)pi; 1634 kif->kf_un.kf_pipe.kf_pipe_peer = (uintptr_t)pi->pipe_peer; 1635 kif->kf_un.kf_pipe.kf_pipe_buffer_cnt = pi->pipe_buffer.cnt; 1636 kif->kf_un.kf_pipe.kf_pipe_buffer_in = pi->pipe_buffer.in; 1637 kif->kf_un.kf_pipe.kf_pipe_buffer_out = pi->pipe_buffer.out; 1638 kif->kf_un.kf_pipe.kf_pipe_buffer_size = pi->pipe_buffer.size; 1639 return (0); 1640 } 1641 1642 static void 1643 pipe_free_kmem(struct pipe *cpipe) 1644 { 1645 1646 KASSERT(!mtx_owned(PIPE_MTX(cpipe)), 1647 ("pipe_free_kmem: pipe mutex locked")); 1648 1649 if (cpipe->pipe_buffer.buffer != NULL) { 1650 atomic_subtract_long(&amountpipekva, cpipe->pipe_buffer.size); 1651 vm_map_remove(pipe_map, 1652 (vm_offset_t)cpipe->pipe_buffer.buffer, 1653 (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size); 1654 cpipe->pipe_buffer.buffer = NULL; 1655 } 1656 #ifndef PIPE_NODIRECT 1657 { 1658 cpipe->pipe_pages.cnt = 0; 1659 cpipe->pipe_pages.pos = 0; 1660 cpipe->pipe_pages.npages = 0; 1661 } 1662 #endif 1663 } 1664 1665 /* 1666 * shutdown the pipe 1667 */ 1668 static void 1669 pipeclose(struct pipe *cpipe) 1670 { 1671 #ifdef MAC 1672 struct pipepair *pp; 1673 #endif 1674 struct pipe *ppipe; 1675 1676 KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL")); 1677 1678 PIPE_LOCK(cpipe); 1679 pipelock(cpipe, 0); 1680 #ifdef MAC 1681 pp = cpipe->pipe_pair; 1682 #endif 1683 1684 /* 1685 * If the other side is blocked, wake it up saying that 1686 * we want to close it down. 1687 */ 1688 cpipe->pipe_state |= PIPE_EOF; 1689 while (cpipe->pipe_busy) { 1690 wakeup(cpipe); 1691 cpipe->pipe_state |= PIPE_WANT; 1692 pipeunlock(cpipe); 1693 msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0); 1694 pipelock(cpipe, 0); 1695 } 1696 1697 pipeselwakeup(cpipe); 1698 1699 /* 1700 * Disconnect from peer, if any. 1701 */ 1702 ppipe = cpipe->pipe_peer; 1703 if (ppipe->pipe_present == PIPE_ACTIVE) { 1704 ppipe->pipe_state |= PIPE_EOF; 1705 wakeup(ppipe); 1706 pipeselwakeup(ppipe); 1707 } 1708 1709 /* 1710 * Mark this endpoint as free. Release kmem resources. We 1711 * don't mark this endpoint as unused until we've finished 1712 * doing that, or the pipe might disappear out from under 1713 * us. 1714 */ 1715 PIPE_UNLOCK(cpipe); 1716 pipe_free_kmem(cpipe); 1717 PIPE_LOCK(cpipe); 1718 cpipe->pipe_present = PIPE_CLOSING; 1719 pipeunlock(cpipe); 1720 1721 /* 1722 * knlist_clear() may sleep dropping the PIPE_MTX. Set the 1723 * PIPE_FINALIZED, that allows other end to free the 1724 * pipe_pair, only after the knotes are completely dismantled. 1725 */ 1726 knlist_clear(&cpipe->pipe_sel.si_note, 1); 1727 cpipe->pipe_present = PIPE_FINALIZED; 1728 seldrain(&cpipe->pipe_sel); 1729 knlist_destroy(&cpipe->pipe_sel.si_note); 1730 1731 /* 1732 * If both endpoints are now closed, release the memory for the 1733 * pipe pair. If not, unlock. 1734 */ 1735 if (ppipe->pipe_present == PIPE_FINALIZED) { 1736 PIPE_UNLOCK(cpipe); 1737 #ifdef MAC 1738 mac_pipe_destroy(pp); 1739 #endif 1740 uma_zfree(pipe_zone, cpipe->pipe_pair); 1741 } else 1742 PIPE_UNLOCK(cpipe); 1743 } 1744 1745 /*ARGSUSED*/ 1746 static int 1747 pipe_kqfilter(struct file *fp, struct knote *kn) 1748 { 1749 struct pipe *cpipe; 1750 1751 /* 1752 * If a filter is requested that is not supported by this file 1753 * descriptor, don't return an error, but also don't ever generate an 1754 * event. 1755 */ 1756 if ((kn->kn_filter == EVFILT_READ) && !(fp->f_flag & FREAD)) { 1757 kn->kn_fop = &pipe_nfiltops; 1758 return (0); 1759 } 1760 if ((kn->kn_filter == EVFILT_WRITE) && !(fp->f_flag & FWRITE)) { 1761 kn->kn_fop = &pipe_nfiltops; 1762 return (0); 1763 } 1764 cpipe = fp->f_data; 1765 PIPE_LOCK(cpipe); 1766 switch (kn->kn_filter) { 1767 case EVFILT_READ: 1768 kn->kn_fop = &pipe_rfiltops; 1769 break; 1770 case EVFILT_WRITE: 1771 kn->kn_fop = &pipe_wfiltops; 1772 if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) { 1773 /* other end of pipe has been closed */ 1774 PIPE_UNLOCK(cpipe); 1775 return (EPIPE); 1776 } 1777 cpipe = PIPE_PEER(cpipe); 1778 break; 1779 default: 1780 if ((cpipe->pipe_type & PIPE_TYPE_NAMED) != 0) { 1781 PIPE_UNLOCK(cpipe); 1782 return (vnops.fo_kqfilter(fp, kn)); 1783 } 1784 PIPE_UNLOCK(cpipe); 1785 return (EINVAL); 1786 } 1787 1788 kn->kn_hook = cpipe; 1789 knlist_add(&cpipe->pipe_sel.si_note, kn, 1); 1790 PIPE_UNLOCK(cpipe); 1791 return (0); 1792 } 1793 1794 static void 1795 filt_pipedetach(struct knote *kn) 1796 { 1797 struct pipe *cpipe = kn->kn_hook; 1798 1799 PIPE_LOCK(cpipe); 1800 knlist_remove(&cpipe->pipe_sel.si_note, kn, 1); 1801 PIPE_UNLOCK(cpipe); 1802 } 1803 1804 /*ARGSUSED*/ 1805 static int 1806 filt_piperead(struct knote *kn, long hint) 1807 { 1808 struct file *fp = kn->kn_fp; 1809 struct pipe *rpipe = kn->kn_hook; 1810 1811 PIPE_LOCK_ASSERT(rpipe, MA_OWNED); 1812 kn->kn_data = rpipe->pipe_buffer.cnt; 1813 if (kn->kn_data == 0) 1814 kn->kn_data = rpipe->pipe_pages.cnt; 1815 1816 if ((rpipe->pipe_state & PIPE_EOF) != 0 && 1817 ((rpipe->pipe_type & PIPE_TYPE_NAMED) == 0 || 1818 fp->f_pipegen != rpipe->pipe_wgen)) { 1819 kn->kn_flags |= EV_EOF; 1820 return (1); 1821 } 1822 kn->kn_flags &= ~EV_EOF; 1823 return (kn->kn_data > 0); 1824 } 1825 1826 /*ARGSUSED*/ 1827 static int 1828 filt_pipewrite(struct knote *kn, long hint) 1829 { 1830 struct pipe *wpipe = kn->kn_hook; 1831 1832 /* 1833 * If this end of the pipe is closed, the knote was removed from the 1834 * knlist and the list lock (i.e., the pipe lock) is therefore not held. 1835 */ 1836 if (wpipe->pipe_present == PIPE_ACTIVE || 1837 (wpipe->pipe_type & PIPE_TYPE_NAMED) != 0) { 1838 PIPE_LOCK_ASSERT(wpipe, MA_OWNED); 1839 1840 if (wpipe->pipe_state & PIPE_DIRECTW) { 1841 kn->kn_data = 0; 1842 } else if (wpipe->pipe_buffer.size > 0) { 1843 kn->kn_data = wpipe->pipe_buffer.size - 1844 wpipe->pipe_buffer.cnt; 1845 } else { 1846 kn->kn_data = PIPE_BUF; 1847 } 1848 } 1849 1850 if (wpipe->pipe_present != PIPE_ACTIVE || 1851 (wpipe->pipe_state & PIPE_EOF)) { 1852 kn->kn_flags |= EV_EOF; 1853 return (1); 1854 } 1855 kn->kn_flags &= ~EV_EOF; 1856 return (kn->kn_data >= PIPE_BUF); 1857 } 1858 1859 static void 1860 filt_pipedetach_notsup(struct knote *kn) 1861 { 1862 1863 } 1864 1865 static int 1866 filt_pipenotsup(struct knote *kn, long hint) 1867 { 1868 1869 return (0); 1870 } 1871