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