1 /*- 2 * Copyright (c) 1996 John S. Dyson 3 * Copyright (c) 2012 Giovanni Trematerra 4 * All rights reserved. 5 * 6 * Redistribution and use in source and binary forms, with or without 7 * modification, are permitted provided that the following conditions 8 * are met: 9 * 1. Redistributions of source code must retain the above copyright 10 * notice immediately at the beginning of the file, without modification, 11 * this list of conditions, and the following disclaimer. 12 * 2. Redistributions in binary form must reproduce the above copyright 13 * notice, this list of conditions and the following disclaimer in the 14 * documentation and/or other materials provided with the distribution. 15 * 3. Absolutely no warranty of function or purpose is made by the author 16 * John S. Dyson. 17 * 4. Modifications may be freely made to this file if the above conditions 18 * are met. 19 */ 20 21 /* 22 * This file contains a high-performance replacement for the socket-based 23 * pipes scheme originally used in FreeBSD/4.4Lite. It does not support 24 * all features of sockets, but does do everything that pipes normally 25 * do. 26 */ 27 28 /* 29 * This code has two modes of operation, a small write mode and a large 30 * write mode. The small write mode acts like conventional pipes with 31 * a kernel buffer. If the buffer is less than PIPE_MINDIRECT, then the 32 * "normal" pipe buffering is done. If the buffer is between PIPE_MINDIRECT 33 * and PIPE_SIZE in size, the sending process pins the underlying pages in 34 * memory, and the receiving process copies directly from these pinned pages 35 * in the sending process. 36 * 37 * If the sending process receives a signal, it is possible that it will 38 * go away, and certainly its address space can change, because control 39 * is returned back to the user-mode side. In that case, the pipe code 40 * arranges to copy the buffer supplied by the user process, to a pageable 41 * kernel buffer, and the receiving process will grab the data from the 42 * pageable kernel buffer. Since signals don't happen all that often, 43 * the copy operation is normally eliminated. 44 * 45 * The constant PIPE_MINDIRECT is chosen to make sure that buffering will 46 * happen for small transfers so that the system will not spend all of 47 * its time context switching. 48 * 49 * In order to limit the resource use of pipes, two sysctls exist: 50 * 51 * kern.ipc.maxpipekva - This is a hard limit on the amount of pageable 52 * address space available to us in pipe_map. This value is normally 53 * autotuned, but may also be loader tuned. 54 * 55 * kern.ipc.pipekva - This read-only sysctl tracks the current amount of 56 * memory in use by pipes. 57 * 58 * Based on how large pipekva is relative to maxpipekva, the following 59 * will happen: 60 * 61 * 0% - 50%: 62 * New pipes are given 16K of memory backing, pipes may dynamically 63 * grow to as large as 64K where needed. 64 * 50% - 75%: 65 * New pipes are given 4K (or PAGE_SIZE) of memory backing, 66 * existing pipes may NOT grow. 67 * 75% - 100%: 68 * New pipes are given 4K (or PAGE_SIZE) of memory backing, 69 * existing pipes will be shrunk down to 4K whenever possible. 70 * 71 * Resizing may be disabled by setting kern.ipc.piperesizeallowed=0. If 72 * that is set, the only resize that will occur is the 0 -> SMALL_PIPE_SIZE 73 * resize which MUST occur for reverse-direction pipes when they are 74 * first used. 75 * 76 * Additional information about the current state of pipes may be obtained 77 * from kern.ipc.pipes, kern.ipc.pipefragretry, kern.ipc.pipeallocfail, 78 * and kern.ipc.piperesizefail. 79 * 80 * Locking rules: There are two locks present here: A mutex, used via 81 * PIPE_LOCK, and a flag, used via pipelock(). All locking is done via 82 * the flag, as mutexes can not persist over uiomove. The mutex 83 * exists only to guard access to the flag, and is not in itself a 84 * locking mechanism. Also note that there is only a single mutex for 85 * both directions of a pipe. 86 * 87 * As pipelock() may have to sleep before it can acquire the flag, it 88 * is important to reread all data after a call to pipelock(); everything 89 * in the structure may have changed. 90 */ 91 92 #include <sys/cdefs.h> 93 __FBSDID("$FreeBSD$"); 94 95 #include <sys/param.h> 96 #include <sys/systm.h> 97 #include <sys/conf.h> 98 #include <sys/fcntl.h> 99 #include <sys/file.h> 100 #include <sys/filedesc.h> 101 #include <sys/filio.h> 102 #include <sys/kernel.h> 103 #include <sys/lock.h> 104 #include <sys/mutex.h> 105 #include <sys/ttycom.h> 106 #include <sys/stat.h> 107 #include <sys/malloc.h> 108 #include <sys/poll.h> 109 #include <sys/selinfo.h> 110 #include <sys/signalvar.h> 111 #include <sys/syscallsubr.h> 112 #include <sys/sysctl.h> 113 #include <sys/sysproto.h> 114 #include <sys/pipe.h> 115 #include <sys/proc.h> 116 #include <sys/vnode.h> 117 #include <sys/uio.h> 118 #include <sys/user.h> 119 #include <sys/event.h> 120 121 #include <security/mac/mac_framework.h> 122 123 #include <vm/vm.h> 124 #include <vm/vm_param.h> 125 #include <vm/vm_object.h> 126 #include <vm/vm_kern.h> 127 #include <vm/vm_extern.h> 128 #include <vm/pmap.h> 129 #include <vm/vm_map.h> 130 #include <vm/vm_page.h> 131 #include <vm/uma.h> 132 133 /* 134 * Use this define if you want to disable *fancy* VM things. Expect an 135 * approx 30% decrease in transfer rate. This could be useful for 136 * NetBSD or OpenBSD. 137 */ 138 /* #define PIPE_NODIRECT */ 139 140 #define PIPE_PEER(pipe) \ 141 (((pipe)->pipe_type & PIPE_TYPE_NAMED) ? (pipe) : ((pipe)->pipe_peer)) 142 143 /* 144 * interfaces to the outside world 145 */ 146 static fo_rdwr_t pipe_read; 147 static fo_rdwr_t pipe_write; 148 static fo_truncate_t pipe_truncate; 149 static fo_ioctl_t pipe_ioctl; 150 static fo_poll_t pipe_poll; 151 static fo_kqfilter_t pipe_kqfilter; 152 static fo_stat_t pipe_stat; 153 static fo_close_t pipe_close; 154 static fo_chmod_t pipe_chmod; 155 static fo_chown_t pipe_chown; 156 static fo_fill_kinfo_t pipe_fill_kinfo; 157 158 struct fileops pipeops = { 159 .fo_read = pipe_read, 160 .fo_write = pipe_write, 161 .fo_truncate = pipe_truncate, 162 .fo_ioctl = pipe_ioctl, 163 .fo_poll = pipe_poll, 164 .fo_kqfilter = pipe_kqfilter, 165 .fo_stat = pipe_stat, 166 .fo_close = pipe_close, 167 .fo_chmod = pipe_chmod, 168 .fo_chown = pipe_chown, 169 .fo_sendfile = invfo_sendfile, 170 .fo_fill_kinfo = pipe_fill_kinfo, 171 .fo_flags = DFLAG_PASSABLE 172 }; 173 174 static void filt_pipedetach(struct knote *kn); 175 static void filt_pipedetach_notsup(struct knote *kn); 176 static int filt_pipenotsup(struct knote *kn, long hint); 177 static int filt_piperead(struct knote *kn, long hint); 178 static int filt_pipewrite(struct knote *kn, long hint); 179 180 static struct filterops pipe_nfiltops = { 181 .f_isfd = 1, 182 .f_detach = filt_pipedetach_notsup, 183 .f_event = filt_pipenotsup 184 }; 185 static struct filterops pipe_rfiltops = { 186 .f_isfd = 1, 187 .f_detach = filt_pipedetach, 188 .f_event = filt_piperead 189 }; 190 static struct filterops pipe_wfiltops = { 191 .f_isfd = 1, 192 .f_detach = filt_pipedetach, 193 .f_event = filt_pipewrite 194 }; 195 196 /* 197 * Default pipe buffer size(s), this can be kind-of large now because pipe 198 * space is pageable. The pipe code will try to maintain locality of 199 * reference for performance reasons, so small amounts of outstanding I/O 200 * will not wipe the cache. 201 */ 202 #define MINPIPESIZE (PIPE_SIZE/3) 203 #define MAXPIPESIZE (2*PIPE_SIZE/3) 204 205 static long amountpipekva; 206 static int pipefragretry; 207 static int pipeallocfail; 208 static int piperesizefail; 209 static int piperesizeallowed = 1; 210 static long pipe_mindirect = PIPE_MINDIRECT; 211 212 SYSCTL_LONG(_kern_ipc, OID_AUTO, maxpipekva, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, 213 &maxpipekva, 0, "Pipe KVA limit"); 214 SYSCTL_LONG(_kern_ipc, OID_AUTO, pipekva, CTLFLAG_RD, 215 &amountpipekva, 0, "Pipe KVA usage"); 216 SYSCTL_INT(_kern_ipc, OID_AUTO, pipefragretry, CTLFLAG_RD, 217 &pipefragretry, 0, "Pipe allocation retries due to fragmentation"); 218 SYSCTL_INT(_kern_ipc, OID_AUTO, pipeallocfail, CTLFLAG_RD, 219 &pipeallocfail, 0, "Pipe allocation failures"); 220 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizefail, CTLFLAG_RD, 221 &piperesizefail, 0, "Pipe resize failures"); 222 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizeallowed, CTLFLAG_RW, 223 &piperesizeallowed, 0, "Pipe resizing allowed"); 224 225 static void pipeinit(void *dummy __unused); 226 static void pipeclose(struct pipe *cpipe); 227 static void pipe_free_kmem(struct pipe *cpipe); 228 static int pipe_create(struct pipe *pipe, bool backing); 229 static int pipe_paircreate(struct thread *td, struct pipepair **p_pp); 230 static __inline int pipelock(struct pipe *cpipe, int catch); 231 static __inline void pipeunlock(struct pipe *cpipe); 232 static void pipe_timestamp(struct timespec *tsp); 233 #ifndef PIPE_NODIRECT 234 static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio); 235 static void pipe_destroy_write_buffer(struct pipe *wpipe); 236 static int pipe_direct_write(struct pipe *wpipe, struct uio *uio); 237 static void pipe_clone_write_buffer(struct pipe *wpipe); 238 #endif 239 static int pipespace(struct pipe *cpipe, int size); 240 static int pipespace_new(struct pipe *cpipe, int size); 241 242 static int pipe_zone_ctor(void *mem, int size, void *arg, int flags); 243 static int pipe_zone_init(void *mem, int size, int flags); 244 static void pipe_zone_fini(void *mem, int size); 245 246 static uma_zone_t pipe_zone; 247 static struct unrhdr64 pipeino_unr; 248 static dev_t pipedev_ino; 249 250 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL); 251 252 static void 253 pipeinit(void *dummy __unused) 254 { 255 256 pipe_zone = uma_zcreate("pipe", sizeof(struct pipepair), 257 pipe_zone_ctor, NULL, pipe_zone_init, pipe_zone_fini, 258 UMA_ALIGN_PTR, 0); 259 KASSERT(pipe_zone != NULL, ("pipe_zone not initialized")); 260 new_unrhdr64(&pipeino_unr, 1); 261 pipedev_ino = devfs_alloc_cdp_inode(); 262 KASSERT(pipedev_ino > 0, ("pipe dev inode not initialized")); 263 } 264 265 static int 266 sysctl_handle_pipe_mindirect(SYSCTL_HANDLER_ARGS) 267 { 268 int error = 0; 269 long tmp_pipe_mindirect = pipe_mindirect; 270 271 error = sysctl_handle_long(oidp, &tmp_pipe_mindirect, arg2, req); 272 if (error != 0 || req->newptr == NULL) 273 return (error); 274 275 /* 276 * Don't allow pipe_mindirect to be set so low that we violate 277 * atomicity requirements. 278 */ 279 if (tmp_pipe_mindirect <= PIPE_BUF) 280 return (EINVAL); 281 pipe_mindirect = tmp_pipe_mindirect; 282 return (0); 283 } 284 SYSCTL_OID(_kern_ipc, OID_AUTO, pipe_mindirect, CTLTYPE_LONG | CTLFLAG_RW, 285 &pipe_mindirect, 0, sysctl_handle_pipe_mindirect, "L", 286 "Minimum write size triggering VM optimization"); 287 288 static int 289 pipe_zone_ctor(void *mem, int size, void *arg, int flags) 290 { 291 struct pipepair *pp; 292 struct pipe *rpipe, *wpipe; 293 294 KASSERT(size == sizeof(*pp), ("pipe_zone_ctor: wrong size")); 295 296 pp = (struct pipepair *)mem; 297 298 /* 299 * We zero both pipe endpoints to make sure all the kmem pointers 300 * are NULL, flag fields are zero'd, etc. We timestamp both 301 * endpoints with the same time. 302 */ 303 rpipe = &pp->pp_rpipe; 304 bzero(rpipe, sizeof(*rpipe)); 305 pipe_timestamp(&rpipe->pipe_ctime); 306 rpipe->pipe_atime = rpipe->pipe_mtime = rpipe->pipe_ctime; 307 308 wpipe = &pp->pp_wpipe; 309 bzero(wpipe, sizeof(*wpipe)); 310 wpipe->pipe_ctime = rpipe->pipe_ctime; 311 wpipe->pipe_atime = wpipe->pipe_mtime = rpipe->pipe_ctime; 312 313 rpipe->pipe_peer = wpipe; 314 rpipe->pipe_pair = pp; 315 wpipe->pipe_peer = rpipe; 316 wpipe->pipe_pair = pp; 317 318 /* 319 * Mark both endpoints as present; they will later get free'd 320 * one at a time. When both are free'd, then the whole pair 321 * is released. 322 */ 323 rpipe->pipe_present = PIPE_ACTIVE; 324 wpipe->pipe_present = PIPE_ACTIVE; 325 326 /* 327 * Eventually, the MAC Framework may initialize the label 328 * in ctor or init, but for now we do it elswhere to avoid 329 * blocking in ctor or init. 330 */ 331 pp->pp_label = NULL; 332 333 return (0); 334 } 335 336 static int 337 pipe_zone_init(void *mem, int size, int flags) 338 { 339 struct pipepair *pp; 340 341 KASSERT(size == sizeof(*pp), ("pipe_zone_init: wrong size")); 342 343 pp = (struct pipepair *)mem; 344 345 mtx_init(&pp->pp_mtx, "pipe mutex", NULL, MTX_DEF | MTX_NEW); 346 return (0); 347 } 348 349 static void 350 pipe_zone_fini(void *mem, int size) 351 { 352 struct pipepair *pp; 353 354 KASSERT(size == sizeof(*pp), ("pipe_zone_fini: wrong size")); 355 356 pp = (struct pipepair *)mem; 357 358 mtx_destroy(&pp->pp_mtx); 359 } 360 361 static int 362 pipe_paircreate(struct thread *td, struct pipepair **p_pp) 363 { 364 struct pipepair *pp; 365 struct pipe *rpipe, *wpipe; 366 int error; 367 368 *p_pp = pp = uma_zalloc(pipe_zone, M_WAITOK); 369 #ifdef MAC 370 /* 371 * The MAC label is shared between the connected endpoints. As a 372 * result mac_pipe_init() and mac_pipe_create() are called once 373 * for the pair, and not on the endpoints. 374 */ 375 mac_pipe_init(pp); 376 mac_pipe_create(td->td_ucred, pp); 377 #endif 378 rpipe = &pp->pp_rpipe; 379 wpipe = &pp->pp_wpipe; 380 381 knlist_init_mtx(&rpipe->pipe_sel.si_note, PIPE_MTX(rpipe)); 382 knlist_init_mtx(&wpipe->pipe_sel.si_note, PIPE_MTX(wpipe)); 383 384 /* 385 * Only the forward direction pipe is backed by big buffer by 386 * default. 387 */ 388 error = pipe_create(rpipe, true); 389 if (error != 0) 390 goto fail; 391 error = pipe_create(wpipe, false); 392 if (error != 0) { 393 /* 394 * This cleanup leaves the pipe inode number for rpipe 395 * still allocated, but never used. We do not free 396 * inode numbers for opened pipes, which is required 397 * for correctness because numbers must be unique. 398 * But also it avoids any memory use by the unr 399 * allocator, so stashing away the transient inode 400 * number is reasonable. 401 */ 402 pipe_free_kmem(rpipe); 403 goto fail; 404 } 405 406 rpipe->pipe_state |= PIPE_DIRECTOK; 407 wpipe->pipe_state |= PIPE_DIRECTOK; 408 return (0); 409 410 fail: 411 knlist_destroy(&rpipe->pipe_sel.si_note); 412 knlist_destroy(&wpipe->pipe_sel.si_note); 413 #ifdef MAC 414 mac_pipe_destroy(pp); 415 #endif 416 uma_zfree(pipe_zone, pp); 417 return (error); 418 } 419 420 int 421 pipe_named_ctor(struct pipe **ppipe, struct thread *td) 422 { 423 struct pipepair *pp; 424 int error; 425 426 error = pipe_paircreate(td, &pp); 427 if (error != 0) 428 return (error); 429 pp->pp_rpipe.pipe_type |= PIPE_TYPE_NAMED; 430 *ppipe = &pp->pp_rpipe; 431 return (0); 432 } 433 434 void 435 pipe_dtor(struct pipe *dpipe) 436 { 437 struct pipe *peer; 438 439 peer = (dpipe->pipe_type & PIPE_TYPE_NAMED) != 0 ? dpipe->pipe_peer : NULL; 440 funsetown(&dpipe->pipe_sigio); 441 pipeclose(dpipe); 442 if (peer != NULL) { 443 funsetown(&peer->pipe_sigio); 444 pipeclose(peer); 445 } 446 } 447 448 /* 449 * Get a timestamp. 450 * 451 * This used to be vfs_timestamp but the higher precision is unnecessary and 452 * can very negatively affect performance in virtualized environments (e.g., on 453 * vms running on amd64 when using the rdtscp instruction). 454 */ 455 static void 456 pipe_timestamp(struct timespec *tsp) 457 { 458 459 getnanotime(tsp); 460 } 461 462 /* 463 * The pipe system call for the DTYPE_PIPE type of pipes. If we fail, let 464 * the zone pick up the pieces via pipeclose(). 465 */ 466 int 467 kern_pipe(struct thread *td, int fildes[2], int flags, struct filecaps *fcaps1, 468 struct filecaps *fcaps2) 469 { 470 struct file *rf, *wf; 471 struct pipe *rpipe, *wpipe; 472 struct pipepair *pp; 473 int fd, fflags, error; 474 475 error = pipe_paircreate(td, &pp); 476 if (error != 0) 477 return (error); 478 rpipe = &pp->pp_rpipe; 479 wpipe = &pp->pp_wpipe; 480 error = falloc_caps(td, &rf, &fd, flags, fcaps1); 481 if (error) { 482 pipeclose(rpipe); 483 pipeclose(wpipe); 484 return (error); 485 } 486 /* An extra reference on `rf' has been held for us by falloc_caps(). */ 487 fildes[0] = fd; 488 489 fflags = FREAD | FWRITE; 490 if ((flags & O_NONBLOCK) != 0) 491 fflags |= FNONBLOCK; 492 493 /* 494 * Warning: once we've gotten past allocation of the fd for the 495 * read-side, we can only drop the read side via fdrop() in order 496 * to avoid races against processes which manage to dup() the read 497 * side while we are blocked trying to allocate the write side. 498 */ 499 finit(rf, fflags, DTYPE_PIPE, rpipe, &pipeops); 500 error = falloc_caps(td, &wf, &fd, flags, fcaps2); 501 if (error) { 502 fdclose(td, rf, fildes[0]); 503 fdrop(rf, td); 504 /* rpipe has been closed by fdrop(). */ 505 pipeclose(wpipe); 506 return (error); 507 } 508 /* An extra reference on `wf' has been held for us by falloc_caps(). */ 509 finit(wf, fflags, DTYPE_PIPE, wpipe, &pipeops); 510 fdrop(wf, td); 511 fildes[1] = fd; 512 fdrop(rf, td); 513 514 return (0); 515 } 516 517 #ifdef COMPAT_FREEBSD10 518 /* ARGSUSED */ 519 int 520 freebsd10_pipe(struct thread *td, struct freebsd10_pipe_args *uap __unused) 521 { 522 int error; 523 int fildes[2]; 524 525 error = kern_pipe(td, fildes, 0, NULL, NULL); 526 if (error) 527 return (error); 528 529 td->td_retval[0] = fildes[0]; 530 td->td_retval[1] = fildes[1]; 531 532 return (0); 533 } 534 #endif 535 536 int 537 sys_pipe2(struct thread *td, struct pipe2_args *uap) 538 { 539 int error, fildes[2]; 540 541 if (uap->flags & ~(O_CLOEXEC | O_NONBLOCK)) 542 return (EINVAL); 543 error = kern_pipe(td, fildes, uap->flags, NULL, NULL); 544 if (error) 545 return (error); 546 error = copyout(fildes, uap->fildes, 2 * sizeof(int)); 547 if (error) { 548 (void)kern_close(td, fildes[0]); 549 (void)kern_close(td, fildes[1]); 550 } 551 return (error); 552 } 553 554 /* 555 * Allocate kva for pipe circular buffer, the space is pageable 556 * This routine will 'realloc' the size of a pipe safely, if it fails 557 * it will retain the old buffer. 558 * If it fails it will return ENOMEM. 559 */ 560 static int 561 pipespace_new(struct pipe *cpipe, int size) 562 { 563 caddr_t buffer; 564 int error, cnt, firstseg; 565 static int curfail = 0; 566 static struct timeval lastfail; 567 568 KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked")); 569 KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW), 570 ("pipespace: resize of direct writes not allowed")); 571 retry: 572 cnt = cpipe->pipe_buffer.cnt; 573 if (cnt > size) 574 size = cnt; 575 576 size = round_page(size); 577 buffer = (caddr_t) vm_map_min(pipe_map); 578 579 error = vm_map_find(pipe_map, NULL, 0, (vm_offset_t *)&buffer, size, 0, 580 VMFS_ANY_SPACE, VM_PROT_RW, VM_PROT_RW, 0); 581 if (error != KERN_SUCCESS) { 582 if (cpipe->pipe_buffer.buffer == NULL && 583 size > SMALL_PIPE_SIZE) { 584 size = SMALL_PIPE_SIZE; 585 pipefragretry++; 586 goto retry; 587 } 588 if (cpipe->pipe_buffer.buffer == NULL) { 589 pipeallocfail++; 590 if (ppsratecheck(&lastfail, &curfail, 1)) 591 printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n"); 592 } else { 593 piperesizefail++; 594 } 595 return (ENOMEM); 596 } 597 598 /* copy data, then free old resources if we're resizing */ 599 if (cnt > 0) { 600 if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) { 601 firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out; 602 bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out], 603 buffer, firstseg); 604 if ((cnt - firstseg) > 0) 605 bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg], 606 cpipe->pipe_buffer.in); 607 } else { 608 bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out], 609 buffer, cnt); 610 } 611 } 612 pipe_free_kmem(cpipe); 613 cpipe->pipe_buffer.buffer = buffer; 614 cpipe->pipe_buffer.size = size; 615 cpipe->pipe_buffer.in = cnt; 616 cpipe->pipe_buffer.out = 0; 617 cpipe->pipe_buffer.cnt = cnt; 618 atomic_add_long(&amountpipekva, cpipe->pipe_buffer.size); 619 return (0); 620 } 621 622 /* 623 * Wrapper for pipespace_new() that performs locking assertions. 624 */ 625 static int 626 pipespace(struct pipe *cpipe, int size) 627 { 628 629 KASSERT(cpipe->pipe_state & PIPE_LOCKFL, 630 ("Unlocked pipe passed to pipespace")); 631 return (pipespace_new(cpipe, size)); 632 } 633 634 /* 635 * lock a pipe for I/O, blocking other access 636 */ 637 static __inline int 638 pipelock(struct pipe *cpipe, int catch) 639 { 640 int error, prio; 641 642 PIPE_LOCK_ASSERT(cpipe, MA_OWNED); 643 644 prio = PRIBIO; 645 if (catch) 646 prio |= PCATCH; 647 while (cpipe->pipe_state & PIPE_LOCKFL) { 648 KASSERT(cpipe->pipe_waiters >= 0, 649 ("%s: bad waiter count %d", __func__, 650 cpipe->pipe_waiters)); 651 cpipe->pipe_waiters++; 652 error = msleep(&cpipe->pipe_waiters, PIPE_MTX(cpipe), prio, 653 "pipelk", 0); 654 cpipe->pipe_waiters--; 655 if (error != 0) 656 return (error); 657 } 658 cpipe->pipe_state |= PIPE_LOCKFL; 659 return (0); 660 } 661 662 /* 663 * unlock a pipe I/O lock 664 */ 665 static __inline void 666 pipeunlock(struct pipe *cpipe) 667 { 668 669 PIPE_LOCK_ASSERT(cpipe, MA_OWNED); 670 KASSERT(cpipe->pipe_state & PIPE_LOCKFL, 671 ("Unlocked pipe passed to pipeunlock")); 672 KASSERT(cpipe->pipe_waiters >= 0, 673 ("%s: bad waiter count %d", __func__, 674 cpipe->pipe_waiters)); 675 cpipe->pipe_state &= ~PIPE_LOCKFL; 676 if (cpipe->pipe_waiters > 0) 677 wakeup_one(&cpipe->pipe_waiters); 678 } 679 680 void 681 pipeselwakeup(struct pipe *cpipe) 682 { 683 684 PIPE_LOCK_ASSERT(cpipe, MA_OWNED); 685 if (cpipe->pipe_state & PIPE_SEL) { 686 selwakeuppri(&cpipe->pipe_sel, PSOCK); 687 if (!SEL_WAITING(&cpipe->pipe_sel)) 688 cpipe->pipe_state &= ~PIPE_SEL; 689 } 690 if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio) 691 pgsigio(&cpipe->pipe_sigio, SIGIO, 0); 692 KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0); 693 } 694 695 /* 696 * Initialize and allocate VM and memory for pipe. The structure 697 * will start out zero'd from the ctor, so we just manage the kmem. 698 */ 699 static int 700 pipe_create(struct pipe *pipe, bool large_backing) 701 { 702 int error; 703 704 error = pipespace_new(pipe, !large_backing || amountpipekva > 705 maxpipekva / 2 ? SMALL_PIPE_SIZE : PIPE_SIZE); 706 if (error == 0) 707 pipe->pipe_ino = alloc_unr64(&pipeino_unr); 708 return (error); 709 } 710 711 /* ARGSUSED */ 712 static int 713 pipe_read(struct file *fp, struct uio *uio, struct ucred *active_cred, 714 int flags, struct thread *td) 715 { 716 struct pipe *rpipe; 717 int error; 718 int nread = 0; 719 int size; 720 721 rpipe = fp->f_data; 722 723 /* 724 * Try to avoid locking the pipe if we have nothing to do. 725 * 726 * There are programs which share one pipe amongst multiple processes 727 * and perform non-blocking reads in parallel, even if the pipe is 728 * empty. This in particular is the case with BSD make, which when 729 * spawned with a high -j number can find itself with over half of the 730 * calls failing to find anything. 731 */ 732 if ((fp->f_flag & FNONBLOCK) != 0 && !mac_pipe_check_read_enabled()) { 733 if (__predict_false(uio->uio_resid == 0)) 734 return (0); 735 if ((atomic_load_short(&rpipe->pipe_state) & PIPE_EOF) == 0 && 736 atomic_load_int(&rpipe->pipe_buffer.cnt) == 0 && 737 atomic_load_int(&rpipe->pipe_pages.cnt) == 0) 738 return (EAGAIN); 739 } 740 741 PIPE_LOCK(rpipe); 742 ++rpipe->pipe_busy; 743 error = pipelock(rpipe, 1); 744 if (error) 745 goto unlocked_error; 746 747 #ifdef MAC 748 error = mac_pipe_check_read(active_cred, rpipe->pipe_pair); 749 if (error) 750 goto locked_error; 751 #endif 752 if (amountpipekva > (3 * maxpipekva) / 4) { 753 if ((rpipe->pipe_state & PIPE_DIRECTW) == 0 && 754 rpipe->pipe_buffer.size > SMALL_PIPE_SIZE && 755 rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE && 756 piperesizeallowed == 1) { 757 PIPE_UNLOCK(rpipe); 758 pipespace(rpipe, SMALL_PIPE_SIZE); 759 PIPE_LOCK(rpipe); 760 } 761 } 762 763 while (uio->uio_resid) { 764 /* 765 * normal pipe buffer receive 766 */ 767 if (rpipe->pipe_buffer.cnt > 0) { 768 size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out; 769 if (size > rpipe->pipe_buffer.cnt) 770 size = rpipe->pipe_buffer.cnt; 771 if (size > uio->uio_resid) 772 size = uio->uio_resid; 773 774 PIPE_UNLOCK(rpipe); 775 error = uiomove( 776 &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out], 777 size, uio); 778 PIPE_LOCK(rpipe); 779 if (error) 780 break; 781 782 rpipe->pipe_buffer.out += size; 783 if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size) 784 rpipe->pipe_buffer.out = 0; 785 786 rpipe->pipe_buffer.cnt -= size; 787 788 /* 789 * If there is no more to read in the pipe, reset 790 * its pointers to the beginning. This improves 791 * cache hit stats. 792 */ 793 if (rpipe->pipe_buffer.cnt == 0) { 794 rpipe->pipe_buffer.in = 0; 795 rpipe->pipe_buffer.out = 0; 796 } 797 nread += size; 798 #ifndef PIPE_NODIRECT 799 /* 800 * Direct copy, bypassing a kernel buffer. 801 */ 802 } else if ((size = rpipe->pipe_pages.cnt) != 0) { 803 if (size > uio->uio_resid) 804 size = (u_int) uio->uio_resid; 805 PIPE_UNLOCK(rpipe); 806 error = uiomove_fromphys(rpipe->pipe_pages.ms, 807 rpipe->pipe_pages.pos, size, uio); 808 PIPE_LOCK(rpipe); 809 if (error) 810 break; 811 nread += size; 812 rpipe->pipe_pages.pos += size; 813 rpipe->pipe_pages.cnt -= size; 814 if (rpipe->pipe_pages.cnt == 0) { 815 rpipe->pipe_state &= ~PIPE_WANTW; 816 wakeup(rpipe); 817 } 818 #endif 819 } else { 820 /* 821 * detect EOF condition 822 * read returns 0 on EOF, no need to set error 823 */ 824 if (rpipe->pipe_state & PIPE_EOF) 825 break; 826 827 /* 828 * If the "write-side" has been blocked, wake it up now. 829 */ 830 if (rpipe->pipe_state & PIPE_WANTW) { 831 rpipe->pipe_state &= ~PIPE_WANTW; 832 wakeup(rpipe); 833 } 834 835 /* 836 * Break if some data was read. 837 */ 838 if (nread > 0) 839 break; 840 841 /* 842 * Unlock the pipe buffer for our remaining processing. 843 * We will either break out with an error or we will 844 * sleep and relock to loop. 845 */ 846 pipeunlock(rpipe); 847 848 /* 849 * Handle non-blocking mode operation or 850 * wait for more data. 851 */ 852 if (fp->f_flag & FNONBLOCK) { 853 error = EAGAIN; 854 } else { 855 rpipe->pipe_state |= PIPE_WANTR; 856 if ((error = msleep(rpipe, PIPE_MTX(rpipe), 857 PRIBIO | PCATCH, 858 "piperd", 0)) == 0) 859 error = pipelock(rpipe, 1); 860 } 861 if (error) 862 goto unlocked_error; 863 } 864 } 865 #ifdef MAC 866 locked_error: 867 #endif 868 pipeunlock(rpipe); 869 870 /* XXX: should probably do this before getting any locks. */ 871 if (error == 0) 872 pipe_timestamp(&rpipe->pipe_atime); 873 unlocked_error: 874 --rpipe->pipe_busy; 875 876 /* 877 * PIPE_WANT processing only makes sense if pipe_busy is 0. 878 */ 879 if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) { 880 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW); 881 wakeup(rpipe); 882 } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) { 883 /* 884 * Handle write blocking hysteresis. 885 */ 886 if (rpipe->pipe_state & PIPE_WANTW) { 887 rpipe->pipe_state &= ~PIPE_WANTW; 888 wakeup(rpipe); 889 } 890 } 891 892 /* 893 * Only wake up writers if there was actually something read. 894 * Otherwise, when calling read(2) at EOF, a spurious wakeup occurs. 895 */ 896 if (nread > 0 && 897 rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt >= PIPE_BUF) 898 pipeselwakeup(rpipe); 899 900 PIPE_UNLOCK(rpipe); 901 if (nread > 0) 902 td->td_ru.ru_msgrcv++; 903 return (error); 904 } 905 906 #ifndef PIPE_NODIRECT 907 /* 908 * Map the sending processes' buffer into kernel space and wire it. 909 * This is similar to a physical write operation. 910 */ 911 static int 912 pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio) 913 { 914 u_int size; 915 int i; 916 917 PIPE_LOCK_ASSERT(wpipe, MA_OWNED); 918 KASSERT((wpipe->pipe_state & PIPE_DIRECTW) == 0, 919 ("%s: PIPE_DIRECTW set on %p", __func__, wpipe)); 920 KASSERT(wpipe->pipe_pages.cnt == 0, 921 ("%s: pipe map for %p contains residual data", __func__, wpipe)); 922 923 if (uio->uio_iov->iov_len > wpipe->pipe_buffer.size) 924 size = wpipe->pipe_buffer.size; 925 else 926 size = uio->uio_iov->iov_len; 927 928 wpipe->pipe_state |= PIPE_DIRECTW; 929 PIPE_UNLOCK(wpipe); 930 i = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map, 931 (vm_offset_t)uio->uio_iov->iov_base, size, VM_PROT_READ, 932 wpipe->pipe_pages.ms, PIPENPAGES); 933 PIPE_LOCK(wpipe); 934 if (i < 0) { 935 wpipe->pipe_state &= ~PIPE_DIRECTW; 936 return (EFAULT); 937 } 938 939 wpipe->pipe_pages.npages = i; 940 wpipe->pipe_pages.pos = 941 ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK; 942 wpipe->pipe_pages.cnt = size; 943 944 uio->uio_iov->iov_len -= size; 945 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size; 946 if (uio->uio_iov->iov_len == 0) 947 uio->uio_iov++; 948 uio->uio_resid -= size; 949 uio->uio_offset += size; 950 return (0); 951 } 952 953 /* 954 * Unwire the process buffer. 955 */ 956 static void 957 pipe_destroy_write_buffer(struct pipe *wpipe) 958 { 959 960 PIPE_LOCK_ASSERT(wpipe, MA_OWNED); 961 KASSERT((wpipe->pipe_state & PIPE_DIRECTW) != 0, 962 ("%s: PIPE_DIRECTW not set on %p", __func__, wpipe)); 963 KASSERT(wpipe->pipe_pages.cnt == 0, 964 ("%s: pipe map for %p contains residual data", __func__, wpipe)); 965 966 wpipe->pipe_state &= ~PIPE_DIRECTW; 967 vm_page_unhold_pages(wpipe->pipe_pages.ms, wpipe->pipe_pages.npages); 968 wpipe->pipe_pages.npages = 0; 969 } 970 971 /* 972 * In the case of a signal, the writing process might go away. This 973 * code copies the data into the circular buffer so that the source 974 * pages can be freed without loss of data. 975 */ 976 static void 977 pipe_clone_write_buffer(struct pipe *wpipe) 978 { 979 struct uio uio; 980 struct iovec iov; 981 int size; 982 int pos; 983 984 PIPE_LOCK_ASSERT(wpipe, MA_OWNED); 985 KASSERT((wpipe->pipe_state & PIPE_DIRECTW) != 0, 986 ("%s: PIPE_DIRECTW not set on %p", __func__, wpipe)); 987 988 size = wpipe->pipe_pages.cnt; 989 pos = wpipe->pipe_pages.pos; 990 wpipe->pipe_pages.cnt = 0; 991 992 wpipe->pipe_buffer.in = size; 993 wpipe->pipe_buffer.out = 0; 994 wpipe->pipe_buffer.cnt = size; 995 996 PIPE_UNLOCK(wpipe); 997 iov.iov_base = wpipe->pipe_buffer.buffer; 998 iov.iov_len = size; 999 uio.uio_iov = &iov; 1000 uio.uio_iovcnt = 1; 1001 uio.uio_offset = 0; 1002 uio.uio_resid = size; 1003 uio.uio_segflg = UIO_SYSSPACE; 1004 uio.uio_rw = UIO_READ; 1005 uio.uio_td = curthread; 1006 uiomove_fromphys(wpipe->pipe_pages.ms, pos, size, &uio); 1007 PIPE_LOCK(wpipe); 1008 pipe_destroy_write_buffer(wpipe); 1009 } 1010 1011 /* 1012 * This implements the pipe buffer write mechanism. Note that only 1013 * a direct write OR a normal pipe write can be pending at any given time. 1014 * If there are any characters in the pipe buffer, the direct write will 1015 * be deferred until the receiving process grabs all of the bytes from 1016 * the pipe buffer. Then the direct mapping write is set-up. 1017 */ 1018 static int 1019 pipe_direct_write(struct pipe *wpipe, struct uio *uio) 1020 { 1021 int error; 1022 1023 retry: 1024 PIPE_LOCK_ASSERT(wpipe, MA_OWNED); 1025 if ((wpipe->pipe_state & PIPE_EOF) != 0) { 1026 error = EPIPE; 1027 goto error1; 1028 } 1029 if (wpipe->pipe_state & PIPE_DIRECTW) { 1030 if (wpipe->pipe_state & PIPE_WANTR) { 1031 wpipe->pipe_state &= ~PIPE_WANTR; 1032 wakeup(wpipe); 1033 } 1034 pipeselwakeup(wpipe); 1035 wpipe->pipe_state |= PIPE_WANTW; 1036 pipeunlock(wpipe); 1037 error = msleep(wpipe, PIPE_MTX(wpipe), 1038 PRIBIO | PCATCH, "pipdww", 0); 1039 pipelock(wpipe, 0); 1040 if (error != 0) 1041 goto error1; 1042 goto retry; 1043 } 1044 if (wpipe->pipe_buffer.cnt > 0) { 1045 if (wpipe->pipe_state & PIPE_WANTR) { 1046 wpipe->pipe_state &= ~PIPE_WANTR; 1047 wakeup(wpipe); 1048 } 1049 pipeselwakeup(wpipe); 1050 wpipe->pipe_state |= PIPE_WANTW; 1051 pipeunlock(wpipe); 1052 error = msleep(wpipe, PIPE_MTX(wpipe), 1053 PRIBIO | PCATCH, "pipdwc", 0); 1054 pipelock(wpipe, 0); 1055 if (error != 0) 1056 goto error1; 1057 goto retry; 1058 } 1059 1060 error = pipe_build_write_buffer(wpipe, uio); 1061 if (error) { 1062 goto error1; 1063 } 1064 1065 while (wpipe->pipe_pages.cnt != 0 && 1066 (wpipe->pipe_state & PIPE_EOF) == 0) { 1067 if (wpipe->pipe_state & PIPE_WANTR) { 1068 wpipe->pipe_state &= ~PIPE_WANTR; 1069 wakeup(wpipe); 1070 } 1071 pipeselwakeup(wpipe); 1072 wpipe->pipe_state |= PIPE_WANTW; 1073 pipeunlock(wpipe); 1074 error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH, 1075 "pipdwt", 0); 1076 pipelock(wpipe, 0); 1077 if (error != 0) 1078 break; 1079 } 1080 1081 if ((wpipe->pipe_state & PIPE_EOF) != 0) { 1082 wpipe->pipe_pages.cnt = 0; 1083 pipe_destroy_write_buffer(wpipe); 1084 pipeselwakeup(wpipe); 1085 error = EPIPE; 1086 } else if (error == EINTR || error == ERESTART) { 1087 pipe_clone_write_buffer(wpipe); 1088 } else { 1089 pipe_destroy_write_buffer(wpipe); 1090 } 1091 KASSERT((wpipe->pipe_state & PIPE_DIRECTW) == 0, 1092 ("pipe %p leaked PIPE_DIRECTW", wpipe)); 1093 return (error); 1094 1095 error1: 1096 wakeup(wpipe); 1097 return (error); 1098 } 1099 #endif 1100 1101 static int 1102 pipe_write(struct file *fp, struct uio *uio, struct ucred *active_cred, 1103 int flags, struct thread *td) 1104 { 1105 struct pipe *wpipe, *rpipe; 1106 ssize_t orig_resid; 1107 int desiredsize, error; 1108 1109 rpipe = fp->f_data; 1110 wpipe = PIPE_PEER(rpipe); 1111 PIPE_LOCK(rpipe); 1112 error = pipelock(wpipe, 1); 1113 if (error) { 1114 PIPE_UNLOCK(rpipe); 1115 return (error); 1116 } 1117 /* 1118 * detect loss of pipe read side, issue SIGPIPE if lost. 1119 */ 1120 if (wpipe->pipe_present != PIPE_ACTIVE || 1121 (wpipe->pipe_state & PIPE_EOF)) { 1122 pipeunlock(wpipe); 1123 PIPE_UNLOCK(rpipe); 1124 return (EPIPE); 1125 } 1126 #ifdef MAC 1127 error = mac_pipe_check_write(active_cred, wpipe->pipe_pair); 1128 if (error) { 1129 pipeunlock(wpipe); 1130 PIPE_UNLOCK(rpipe); 1131 return (error); 1132 } 1133 #endif 1134 ++wpipe->pipe_busy; 1135 1136 /* Choose a larger size if it's advantageous */ 1137 desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size); 1138 while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) { 1139 if (piperesizeallowed != 1) 1140 break; 1141 if (amountpipekva > maxpipekva / 2) 1142 break; 1143 if (desiredsize == BIG_PIPE_SIZE) 1144 break; 1145 desiredsize = desiredsize * 2; 1146 } 1147 1148 /* Choose a smaller size if we're in a OOM situation */ 1149 if (amountpipekva > (3 * maxpipekva) / 4 && 1150 wpipe->pipe_buffer.size > SMALL_PIPE_SIZE && 1151 wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE && 1152 piperesizeallowed == 1) 1153 desiredsize = SMALL_PIPE_SIZE; 1154 1155 /* Resize if the above determined that a new size was necessary */ 1156 if (desiredsize != wpipe->pipe_buffer.size && 1157 (wpipe->pipe_state & PIPE_DIRECTW) == 0) { 1158 PIPE_UNLOCK(wpipe); 1159 pipespace(wpipe, desiredsize); 1160 PIPE_LOCK(wpipe); 1161 } 1162 MPASS(wpipe->pipe_buffer.size != 0); 1163 1164 orig_resid = uio->uio_resid; 1165 1166 while (uio->uio_resid) { 1167 int space; 1168 1169 if (wpipe->pipe_state & PIPE_EOF) { 1170 error = EPIPE; 1171 break; 1172 } 1173 #ifndef PIPE_NODIRECT 1174 /* 1175 * If the transfer is large, we can gain performance if 1176 * we do process-to-process copies directly. 1177 * If the write is non-blocking, we don't use the 1178 * direct write mechanism. 1179 * 1180 * The direct write mechanism will detect the reader going 1181 * away on us. 1182 */ 1183 if (uio->uio_segflg == UIO_USERSPACE && 1184 uio->uio_iov->iov_len >= pipe_mindirect && 1185 wpipe->pipe_buffer.size >= pipe_mindirect && 1186 (fp->f_flag & FNONBLOCK) == 0) { 1187 error = pipe_direct_write(wpipe, uio); 1188 if (error != 0) 1189 break; 1190 continue; 1191 } 1192 #endif 1193 1194 /* 1195 * Pipe buffered writes cannot be coincidental with 1196 * direct writes. We wait until the currently executing 1197 * direct write is completed before we start filling the 1198 * pipe buffer. We break out if a signal occurs or the 1199 * reader goes away. 1200 */ 1201 if (wpipe->pipe_pages.cnt != 0) { 1202 if (wpipe->pipe_state & PIPE_WANTR) { 1203 wpipe->pipe_state &= ~PIPE_WANTR; 1204 wakeup(wpipe); 1205 } 1206 pipeselwakeup(wpipe); 1207 wpipe->pipe_state |= PIPE_WANTW; 1208 pipeunlock(wpipe); 1209 error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH, 1210 "pipbww", 0); 1211 pipelock(wpipe, 0); 1212 if (error != 0) 1213 break; 1214 continue; 1215 } 1216 1217 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt; 1218 1219 /* Writes of size <= PIPE_BUF must be atomic. */ 1220 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF)) 1221 space = 0; 1222 1223 if (space > 0) { 1224 int size; /* Transfer size */ 1225 int segsize; /* first segment to transfer */ 1226 1227 /* 1228 * Transfer size is minimum of uio transfer 1229 * and free space in pipe buffer. 1230 */ 1231 if (space > uio->uio_resid) 1232 size = uio->uio_resid; 1233 else 1234 size = space; 1235 /* 1236 * First segment to transfer is minimum of 1237 * transfer size and contiguous space in 1238 * pipe buffer. If first segment to transfer 1239 * is less than the transfer size, we've got 1240 * a wraparound in the buffer. 1241 */ 1242 segsize = wpipe->pipe_buffer.size - 1243 wpipe->pipe_buffer.in; 1244 if (segsize > size) 1245 segsize = size; 1246 1247 /* Transfer first segment */ 1248 1249 PIPE_UNLOCK(rpipe); 1250 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in], 1251 segsize, uio); 1252 PIPE_LOCK(rpipe); 1253 1254 if (error == 0 && segsize < size) { 1255 KASSERT(wpipe->pipe_buffer.in + segsize == 1256 wpipe->pipe_buffer.size, 1257 ("Pipe buffer wraparound disappeared")); 1258 /* 1259 * Transfer remaining part now, to 1260 * support atomic writes. Wraparound 1261 * happened. 1262 */ 1263 1264 PIPE_UNLOCK(rpipe); 1265 error = uiomove( 1266 &wpipe->pipe_buffer.buffer[0], 1267 size - segsize, uio); 1268 PIPE_LOCK(rpipe); 1269 } 1270 if (error == 0) { 1271 wpipe->pipe_buffer.in += size; 1272 if (wpipe->pipe_buffer.in >= 1273 wpipe->pipe_buffer.size) { 1274 KASSERT(wpipe->pipe_buffer.in == 1275 size - segsize + 1276 wpipe->pipe_buffer.size, 1277 ("Expected wraparound bad")); 1278 wpipe->pipe_buffer.in = size - segsize; 1279 } 1280 1281 wpipe->pipe_buffer.cnt += size; 1282 KASSERT(wpipe->pipe_buffer.cnt <= 1283 wpipe->pipe_buffer.size, 1284 ("Pipe buffer overflow")); 1285 } 1286 if (error != 0) 1287 break; 1288 continue; 1289 } else { 1290 /* 1291 * If the "read-side" has been blocked, wake it up now. 1292 */ 1293 if (wpipe->pipe_state & PIPE_WANTR) { 1294 wpipe->pipe_state &= ~PIPE_WANTR; 1295 wakeup(wpipe); 1296 } 1297 1298 /* 1299 * don't block on non-blocking I/O 1300 */ 1301 if (fp->f_flag & FNONBLOCK) { 1302 error = EAGAIN; 1303 break; 1304 } 1305 1306 /* 1307 * We have no more space and have something to offer, 1308 * wake up select/poll. 1309 */ 1310 pipeselwakeup(wpipe); 1311 1312 wpipe->pipe_state |= PIPE_WANTW; 1313 pipeunlock(wpipe); 1314 error = msleep(wpipe, PIPE_MTX(rpipe), 1315 PRIBIO | PCATCH, "pipewr", 0); 1316 pipelock(wpipe, 0); 1317 if (error != 0) 1318 break; 1319 continue; 1320 } 1321 } 1322 1323 --wpipe->pipe_busy; 1324 1325 if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) { 1326 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR); 1327 wakeup(wpipe); 1328 } else if (wpipe->pipe_buffer.cnt > 0) { 1329 /* 1330 * If we have put any characters in the buffer, we wake up 1331 * the reader. 1332 */ 1333 if (wpipe->pipe_state & PIPE_WANTR) { 1334 wpipe->pipe_state &= ~PIPE_WANTR; 1335 wakeup(wpipe); 1336 } 1337 } 1338 1339 /* 1340 * Don't return EPIPE if any byte was written. 1341 * EINTR and other interrupts are handled by generic I/O layer. 1342 * Do not pretend that I/O succeeded for obvious user error 1343 * like EFAULT. 1344 */ 1345 if (uio->uio_resid != orig_resid && error == EPIPE) 1346 error = 0; 1347 1348 if (error == 0) 1349 pipe_timestamp(&wpipe->pipe_mtime); 1350 1351 /* 1352 * We have something to offer, 1353 * wake up select/poll. 1354 */ 1355 if (wpipe->pipe_buffer.cnt) 1356 pipeselwakeup(wpipe); 1357 1358 pipeunlock(wpipe); 1359 PIPE_UNLOCK(rpipe); 1360 if (uio->uio_resid != orig_resid) 1361 td->td_ru.ru_msgsnd++; 1362 return (error); 1363 } 1364 1365 /* ARGSUSED */ 1366 static int 1367 pipe_truncate(struct file *fp, off_t length, struct ucred *active_cred, 1368 struct thread *td) 1369 { 1370 struct pipe *cpipe; 1371 int error; 1372 1373 cpipe = fp->f_data; 1374 if (cpipe->pipe_type & PIPE_TYPE_NAMED) 1375 error = vnops.fo_truncate(fp, length, active_cred, td); 1376 else 1377 error = invfo_truncate(fp, length, active_cred, td); 1378 return (error); 1379 } 1380 1381 /* 1382 * we implement a very minimal set of ioctls for compatibility with sockets. 1383 */ 1384 static int 1385 pipe_ioctl(struct file *fp, u_long cmd, void *data, struct ucred *active_cred, 1386 struct thread *td) 1387 { 1388 struct pipe *mpipe = fp->f_data; 1389 int error; 1390 1391 PIPE_LOCK(mpipe); 1392 1393 #ifdef MAC 1394 error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data); 1395 if (error) { 1396 PIPE_UNLOCK(mpipe); 1397 return (error); 1398 } 1399 #endif 1400 1401 error = 0; 1402 switch (cmd) { 1403 case FIONBIO: 1404 break; 1405 1406 case FIOASYNC: 1407 if (*(int *)data) { 1408 mpipe->pipe_state |= PIPE_ASYNC; 1409 } else { 1410 mpipe->pipe_state &= ~PIPE_ASYNC; 1411 } 1412 break; 1413 1414 case FIONREAD: 1415 if (!(fp->f_flag & FREAD)) { 1416 *(int *)data = 0; 1417 PIPE_UNLOCK(mpipe); 1418 return (0); 1419 } 1420 if (mpipe->pipe_pages.cnt != 0) 1421 *(int *)data = mpipe->pipe_pages.cnt; 1422 else 1423 *(int *)data = mpipe->pipe_buffer.cnt; 1424 break; 1425 1426 case FIOSETOWN: 1427 PIPE_UNLOCK(mpipe); 1428 error = fsetown(*(int *)data, &mpipe->pipe_sigio); 1429 goto out_unlocked; 1430 1431 case FIOGETOWN: 1432 *(int *)data = fgetown(&mpipe->pipe_sigio); 1433 break; 1434 1435 /* This is deprecated, FIOSETOWN should be used instead. */ 1436 case TIOCSPGRP: 1437 PIPE_UNLOCK(mpipe); 1438 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio); 1439 goto out_unlocked; 1440 1441 /* This is deprecated, FIOGETOWN should be used instead. */ 1442 case TIOCGPGRP: 1443 *(int *)data = -fgetown(&mpipe->pipe_sigio); 1444 break; 1445 1446 default: 1447 error = ENOTTY; 1448 break; 1449 } 1450 PIPE_UNLOCK(mpipe); 1451 out_unlocked: 1452 return (error); 1453 } 1454 1455 static int 1456 pipe_poll(struct file *fp, int events, struct ucred *active_cred, 1457 struct thread *td) 1458 { 1459 struct pipe *rpipe; 1460 struct pipe *wpipe; 1461 int levents, revents; 1462 #ifdef MAC 1463 int error; 1464 #endif 1465 1466 revents = 0; 1467 rpipe = fp->f_data; 1468 wpipe = PIPE_PEER(rpipe); 1469 PIPE_LOCK(rpipe); 1470 #ifdef MAC 1471 error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair); 1472 if (error) 1473 goto locked_error; 1474 #endif 1475 if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM)) 1476 if (rpipe->pipe_pages.cnt > 0 || rpipe->pipe_buffer.cnt > 0) 1477 revents |= events & (POLLIN | POLLRDNORM); 1478 1479 if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM)) 1480 if (wpipe->pipe_present != PIPE_ACTIVE || 1481 (wpipe->pipe_state & PIPE_EOF) || 1482 ((wpipe->pipe_state & PIPE_DIRECTW) == 0 && 1483 ((wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF || 1484 wpipe->pipe_buffer.size == 0))) 1485 revents |= events & (POLLOUT | POLLWRNORM); 1486 1487 levents = events & 1488 (POLLIN | POLLINIGNEOF | POLLPRI | POLLRDNORM | POLLRDBAND); 1489 if (rpipe->pipe_type & PIPE_TYPE_NAMED && fp->f_flag & FREAD && levents && 1490 fp->f_pipegen == rpipe->pipe_wgen) 1491 events |= POLLINIGNEOF; 1492 1493 if ((events & POLLINIGNEOF) == 0) { 1494 if (rpipe->pipe_state & PIPE_EOF) { 1495 if (fp->f_flag & FREAD) 1496 revents |= (events & (POLLIN | POLLRDNORM)); 1497 if (wpipe->pipe_present != PIPE_ACTIVE || 1498 (wpipe->pipe_state & PIPE_EOF)) 1499 revents |= POLLHUP; 1500 } 1501 } 1502 1503 if (revents == 0) { 1504 /* 1505 * Add ourselves regardless of eventmask as we have to return 1506 * POLLHUP even if it was not asked for. 1507 */ 1508 if ((fp->f_flag & FREAD) != 0) { 1509 selrecord(td, &rpipe->pipe_sel); 1510 if (SEL_WAITING(&rpipe->pipe_sel)) 1511 rpipe->pipe_state |= PIPE_SEL; 1512 } 1513 1514 if ((fp->f_flag & FWRITE) != 0 && 1515 wpipe->pipe_present == PIPE_ACTIVE) { 1516 selrecord(td, &wpipe->pipe_sel); 1517 if (SEL_WAITING(&wpipe->pipe_sel)) 1518 wpipe->pipe_state |= PIPE_SEL; 1519 } 1520 } 1521 #ifdef MAC 1522 locked_error: 1523 #endif 1524 PIPE_UNLOCK(rpipe); 1525 1526 return (revents); 1527 } 1528 1529 /* 1530 * We shouldn't need locks here as we're doing a read and this should 1531 * be a natural race. 1532 */ 1533 static int 1534 pipe_stat(struct file *fp, struct stat *ub, struct ucred *active_cred) 1535 { 1536 struct pipe *pipe; 1537 #ifdef MAC 1538 int error; 1539 #endif 1540 1541 pipe = fp->f_data; 1542 #ifdef MAC 1543 if (mac_pipe_check_stat_enabled()) { 1544 PIPE_LOCK(pipe); 1545 error = mac_pipe_check_stat(active_cred, pipe->pipe_pair); 1546 PIPE_UNLOCK(pipe); 1547 if (error) { 1548 return (error); 1549 } 1550 } 1551 #endif 1552 1553 /* For named pipes ask the underlying filesystem. */ 1554 if (pipe->pipe_type & PIPE_TYPE_NAMED) { 1555 return (vnops.fo_stat(fp, ub, active_cred)); 1556 } 1557 1558 bzero(ub, sizeof(*ub)); 1559 ub->st_mode = S_IFIFO; 1560 ub->st_blksize = PAGE_SIZE; 1561 if (pipe->pipe_pages.cnt != 0) 1562 ub->st_size = pipe->pipe_pages.cnt; 1563 else 1564 ub->st_size = pipe->pipe_buffer.cnt; 1565 ub->st_blocks = howmany(ub->st_size, ub->st_blksize); 1566 ub->st_atim = pipe->pipe_atime; 1567 ub->st_mtim = pipe->pipe_mtime; 1568 ub->st_ctim = pipe->pipe_ctime; 1569 ub->st_uid = fp->f_cred->cr_uid; 1570 ub->st_gid = fp->f_cred->cr_gid; 1571 ub->st_dev = pipedev_ino; 1572 ub->st_ino = pipe->pipe_ino; 1573 /* 1574 * Left as 0: st_nlink, st_rdev, st_flags, st_gen. 1575 */ 1576 return (0); 1577 } 1578 1579 /* ARGSUSED */ 1580 static int 1581 pipe_close(struct file *fp, struct thread *td) 1582 { 1583 1584 if (fp->f_vnode != NULL) 1585 return vnops.fo_close(fp, td); 1586 fp->f_ops = &badfileops; 1587 pipe_dtor(fp->f_data); 1588 fp->f_data = NULL; 1589 return (0); 1590 } 1591 1592 static int 1593 pipe_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td) 1594 { 1595 struct pipe *cpipe; 1596 int error; 1597 1598 cpipe = fp->f_data; 1599 if (cpipe->pipe_type & PIPE_TYPE_NAMED) 1600 error = vn_chmod(fp, mode, active_cred, td); 1601 else 1602 error = invfo_chmod(fp, mode, active_cred, td); 1603 return (error); 1604 } 1605 1606 static int 1607 pipe_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred, 1608 struct thread *td) 1609 { 1610 struct pipe *cpipe; 1611 int error; 1612 1613 cpipe = fp->f_data; 1614 if (cpipe->pipe_type & PIPE_TYPE_NAMED) 1615 error = vn_chown(fp, uid, gid, active_cred, td); 1616 else 1617 error = invfo_chown(fp, uid, gid, active_cred, td); 1618 return (error); 1619 } 1620 1621 static int 1622 pipe_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp) 1623 { 1624 struct pipe *pi; 1625 1626 if (fp->f_type == DTYPE_FIFO) 1627 return (vn_fill_kinfo(fp, kif, fdp)); 1628 kif->kf_type = KF_TYPE_PIPE; 1629 pi = fp->f_data; 1630 kif->kf_un.kf_pipe.kf_pipe_addr = (uintptr_t)pi; 1631 kif->kf_un.kf_pipe.kf_pipe_peer = (uintptr_t)pi->pipe_peer; 1632 kif->kf_un.kf_pipe.kf_pipe_buffer_cnt = pi->pipe_buffer.cnt; 1633 kif->kf_un.kf_pipe.kf_pipe_buffer_in = pi->pipe_buffer.in; 1634 kif->kf_un.kf_pipe.kf_pipe_buffer_out = pi->pipe_buffer.out; 1635 kif->kf_un.kf_pipe.kf_pipe_buffer_size = pi->pipe_buffer.size; 1636 return (0); 1637 } 1638 1639 static void 1640 pipe_free_kmem(struct pipe *cpipe) 1641 { 1642 1643 KASSERT(!mtx_owned(PIPE_MTX(cpipe)), 1644 ("pipe_free_kmem: pipe mutex locked")); 1645 1646 if (cpipe->pipe_buffer.buffer != NULL) { 1647 atomic_subtract_long(&amountpipekva, cpipe->pipe_buffer.size); 1648 vm_map_remove(pipe_map, 1649 (vm_offset_t)cpipe->pipe_buffer.buffer, 1650 (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size); 1651 cpipe->pipe_buffer.buffer = NULL; 1652 } 1653 #ifndef PIPE_NODIRECT 1654 { 1655 cpipe->pipe_pages.cnt = 0; 1656 cpipe->pipe_pages.pos = 0; 1657 cpipe->pipe_pages.npages = 0; 1658 } 1659 #endif 1660 } 1661 1662 /* 1663 * shutdown the pipe 1664 */ 1665 static void 1666 pipeclose(struct pipe *cpipe) 1667 { 1668 #ifdef MAC 1669 struct pipepair *pp; 1670 #endif 1671 struct pipe *ppipe; 1672 1673 KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL")); 1674 1675 PIPE_LOCK(cpipe); 1676 pipelock(cpipe, 0); 1677 #ifdef MAC 1678 pp = cpipe->pipe_pair; 1679 #endif 1680 1681 /* 1682 * If the other side is blocked, wake it up saying that 1683 * we want to close it down. 1684 */ 1685 cpipe->pipe_state |= PIPE_EOF; 1686 while (cpipe->pipe_busy) { 1687 wakeup(cpipe); 1688 cpipe->pipe_state |= PIPE_WANT; 1689 pipeunlock(cpipe); 1690 msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0); 1691 pipelock(cpipe, 0); 1692 } 1693 1694 pipeselwakeup(cpipe); 1695 1696 /* 1697 * Disconnect from peer, if any. 1698 */ 1699 ppipe = cpipe->pipe_peer; 1700 if (ppipe->pipe_present == PIPE_ACTIVE) { 1701 ppipe->pipe_state |= PIPE_EOF; 1702 wakeup(ppipe); 1703 pipeselwakeup(ppipe); 1704 } 1705 1706 /* 1707 * Mark this endpoint as free. Release kmem resources. We 1708 * don't mark this endpoint as unused until we've finished 1709 * doing that, or the pipe might disappear out from under 1710 * us. 1711 */ 1712 PIPE_UNLOCK(cpipe); 1713 pipe_free_kmem(cpipe); 1714 PIPE_LOCK(cpipe); 1715 cpipe->pipe_present = PIPE_CLOSING; 1716 pipeunlock(cpipe); 1717 1718 /* 1719 * knlist_clear() may sleep dropping the PIPE_MTX. Set the 1720 * PIPE_FINALIZED, that allows other end to free the 1721 * pipe_pair, only after the knotes are completely dismantled. 1722 */ 1723 knlist_clear(&cpipe->pipe_sel.si_note, 1); 1724 cpipe->pipe_present = PIPE_FINALIZED; 1725 seldrain(&cpipe->pipe_sel); 1726 knlist_destroy(&cpipe->pipe_sel.si_note); 1727 1728 /* 1729 * If both endpoints are now closed, release the memory for the 1730 * pipe pair. If not, unlock. 1731 */ 1732 if (ppipe->pipe_present == PIPE_FINALIZED) { 1733 PIPE_UNLOCK(cpipe); 1734 #ifdef MAC 1735 mac_pipe_destroy(pp); 1736 #endif 1737 uma_zfree(pipe_zone, cpipe->pipe_pair); 1738 } else 1739 PIPE_UNLOCK(cpipe); 1740 } 1741 1742 /*ARGSUSED*/ 1743 static int 1744 pipe_kqfilter(struct file *fp, struct knote *kn) 1745 { 1746 struct pipe *cpipe; 1747 1748 /* 1749 * If a filter is requested that is not supported by this file 1750 * descriptor, don't return an error, but also don't ever generate an 1751 * event. 1752 */ 1753 if ((kn->kn_filter == EVFILT_READ) && !(fp->f_flag & FREAD)) { 1754 kn->kn_fop = &pipe_nfiltops; 1755 return (0); 1756 } 1757 if ((kn->kn_filter == EVFILT_WRITE) && !(fp->f_flag & FWRITE)) { 1758 kn->kn_fop = &pipe_nfiltops; 1759 return (0); 1760 } 1761 cpipe = fp->f_data; 1762 PIPE_LOCK(cpipe); 1763 switch (kn->kn_filter) { 1764 case EVFILT_READ: 1765 kn->kn_fop = &pipe_rfiltops; 1766 break; 1767 case EVFILT_WRITE: 1768 kn->kn_fop = &pipe_wfiltops; 1769 if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) { 1770 /* other end of pipe has been closed */ 1771 PIPE_UNLOCK(cpipe); 1772 return (EPIPE); 1773 } 1774 cpipe = PIPE_PEER(cpipe); 1775 break; 1776 default: 1777 if ((cpipe->pipe_type & PIPE_TYPE_NAMED) != 0) { 1778 PIPE_UNLOCK(cpipe); 1779 return (vnops.fo_kqfilter(fp, kn)); 1780 } 1781 PIPE_UNLOCK(cpipe); 1782 return (EINVAL); 1783 } 1784 1785 kn->kn_hook = cpipe; 1786 knlist_add(&cpipe->pipe_sel.si_note, kn, 1); 1787 PIPE_UNLOCK(cpipe); 1788 return (0); 1789 } 1790 1791 static void 1792 filt_pipedetach(struct knote *kn) 1793 { 1794 struct pipe *cpipe = kn->kn_hook; 1795 1796 PIPE_LOCK(cpipe); 1797 knlist_remove(&cpipe->pipe_sel.si_note, kn, 1); 1798 PIPE_UNLOCK(cpipe); 1799 } 1800 1801 /*ARGSUSED*/ 1802 static int 1803 filt_piperead(struct knote *kn, long hint) 1804 { 1805 struct file *fp = kn->kn_fp; 1806 struct pipe *rpipe = kn->kn_hook; 1807 1808 PIPE_LOCK_ASSERT(rpipe, MA_OWNED); 1809 kn->kn_data = rpipe->pipe_buffer.cnt; 1810 if (kn->kn_data == 0) 1811 kn->kn_data = rpipe->pipe_pages.cnt; 1812 1813 if ((rpipe->pipe_state & PIPE_EOF) != 0 && 1814 ((rpipe->pipe_type & PIPE_TYPE_NAMED) == 0 || 1815 fp->f_pipegen != rpipe->pipe_wgen)) { 1816 kn->kn_flags |= EV_EOF; 1817 return (1); 1818 } 1819 kn->kn_flags &= ~EV_EOF; 1820 return (kn->kn_data > 0); 1821 } 1822 1823 /*ARGSUSED*/ 1824 static int 1825 filt_pipewrite(struct knote *kn, long hint) 1826 { 1827 struct pipe *wpipe = kn->kn_hook; 1828 1829 /* 1830 * If this end of the pipe is closed, the knote was removed from the 1831 * knlist and the list lock (i.e., the pipe lock) is therefore not held. 1832 */ 1833 if (wpipe->pipe_present == PIPE_ACTIVE || 1834 (wpipe->pipe_type & PIPE_TYPE_NAMED) != 0) { 1835 PIPE_LOCK_ASSERT(wpipe, MA_OWNED); 1836 1837 if (wpipe->pipe_state & PIPE_DIRECTW) { 1838 kn->kn_data = 0; 1839 } else if (wpipe->pipe_buffer.size > 0) { 1840 kn->kn_data = wpipe->pipe_buffer.size - 1841 wpipe->pipe_buffer.cnt; 1842 } else { 1843 kn->kn_data = PIPE_BUF; 1844 } 1845 } 1846 1847 if (wpipe->pipe_present != PIPE_ACTIVE || 1848 (wpipe->pipe_state & PIPE_EOF)) { 1849 kn->kn_flags |= EV_EOF; 1850 return (1); 1851 } 1852 kn->kn_flags &= ~EV_EOF; 1853 return (kn->kn_data >= PIPE_BUF); 1854 } 1855 1856 static void 1857 filt_pipedetach_notsup(struct knote *kn) 1858 { 1859 1860 } 1861 1862 static int 1863 filt_pipenotsup(struct knote *kn, long hint) 1864 { 1865 1866 return (0); 1867 } 1868