sys_pipe.c revision 216511
1/*- 2 * Copyright (c) 1996 John S. Dyson 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice immediately at the beginning of the file, without modification, 10 * this list of conditions, and the following disclaimer. 11 * 2. Redistributions in binary form must reproduce the above copyright 12 * notice, this list of conditions and the following disclaimer in the 13 * documentation and/or other materials provided with the distribution. 14 * 3. Absolutely no warranty of function or purpose is made by the author 15 * John S. Dyson. 16 * 4. Modifications may be freely made to this file if the above conditions 17 * are met. 18 */ 19 20/* 21 * This file contains a high-performance replacement for the socket-based 22 * pipes scheme originally used in FreeBSD/4.4Lite. It does not support 23 * all features of sockets, but does do everything that pipes normally 24 * do. 25 */ 26 27/* 28 * This code has two modes of operation, a small write mode and a large 29 * write mode. The small write mode acts like conventional pipes with 30 * a kernel buffer. If the buffer is less than PIPE_MINDIRECT, then the 31 * "normal" pipe buffering is done. If the buffer is between PIPE_MINDIRECT 32 * and PIPE_SIZE in size, it is fully mapped and wired into the kernel, and 33 * the receiving process can copy it directly from the pages in the sending 34 * process. 35 * 36 * If the sending process receives a signal, it is possible that it will 37 * go away, and certainly its address space can change, because control 38 * is returned back to the user-mode side. In that case, the pipe code 39 * arranges to copy the buffer supplied by the user process, to a pageable 40 * kernel buffer, and the receiving process will grab the data from the 41 * pageable kernel buffer. Since signals don't happen all that often, 42 * the copy operation is normally eliminated. 43 * 44 * The constant PIPE_MINDIRECT is chosen to make sure that buffering will 45 * happen for small transfers so that the system will not spend all of 46 * its time context switching. 47 * 48 * In order to limit the resource use of pipes, two sysctls exist: 49 * 50 * kern.ipc.maxpipekva - This is a hard limit on the amount of pageable 51 * address space available to us in pipe_map. This value is normally 52 * autotuned, but may also be loader tuned. 53 * 54 * kern.ipc.pipekva - This read-only sysctl tracks the current amount of 55 * memory in use by pipes. 56 * 57 * Based on how large pipekva is relative to maxpipekva, the following 58 * will happen: 59 * 60 * 0% - 50%: 61 * New pipes are given 16K of memory backing, pipes may dynamically 62 * grow to as large as 64K where needed. 63 * 50% - 75%: 64 * New pipes are given 4K (or PAGE_SIZE) of memory backing, 65 * existing pipes may NOT grow. 66 * 75% - 100%: 67 * New pipes are given 4K (or PAGE_SIZE) of memory backing, 68 * existing pipes will be shrunk down to 4K whenever possible. 69 * 70 * Resizing may be disabled by setting kern.ipc.piperesizeallowed=0. If 71 * that is set, the only resize that will occur is the 0 -> SMALL_PIPE_SIZE 72 * resize which MUST occur for reverse-direction pipes when they are 73 * first used. 74 * 75 * Additional information about the current state of pipes may be obtained 76 * from kern.ipc.pipes, kern.ipc.pipefragretry, kern.ipc.pipeallocfail, 77 * and kern.ipc.piperesizefail. 78 * 79 * Locking rules: There are two locks present here: A mutex, used via 80 * PIPE_LOCK, and a flag, used via pipelock(). All locking is done via 81 * the flag, as mutexes can not persist over uiomove. The mutex 82 * exists only to guard access to the flag, and is not in itself a 83 * locking mechanism. Also note that there is only a single mutex for 84 * both directions of a pipe. 85 * 86 * As pipelock() may have to sleep before it can acquire the flag, it 87 * is important to reread all data after a call to pipelock(); everything 88 * in the structure may have changed. 89 */ 90 91#include <sys/cdefs.h> 92__FBSDID("$FreeBSD: head/sys/kern/sys_pipe.c 216511 2010-12-17 22:41:22Z alc $"); 93 94#include <sys/param.h> 95#include <sys/systm.h> 96#include <sys/fcntl.h> 97#include <sys/file.h> 98#include <sys/filedesc.h> 99#include <sys/filio.h> 100#include <sys/kernel.h> 101#include <sys/lock.h> 102#include <sys/mutex.h> 103#include <sys/ttycom.h> 104#include <sys/stat.h> 105#include <sys/malloc.h> 106#include <sys/poll.h> 107#include <sys/selinfo.h> 108#include <sys/signalvar.h> 109#include <sys/syscallsubr.h> 110#include <sys/sysctl.h> 111#include <sys/sysproto.h> 112#include <sys/pipe.h> 113#include <sys/proc.h> 114#include <sys/vnode.h> 115#include <sys/uio.h> 116#include <sys/event.h> 117 118#include <security/mac/mac_framework.h> 119 120#include <vm/vm.h> 121#include <vm/vm_param.h> 122#include <vm/vm_object.h> 123#include <vm/vm_kern.h> 124#include <vm/vm_extern.h> 125#include <vm/pmap.h> 126#include <vm/vm_map.h> 127#include <vm/vm_page.h> 128#include <vm/uma.h> 129 130/* 131 * Use this define if you want to disable *fancy* VM things. Expect an 132 * approx 30% decrease in transfer rate. This could be useful for 133 * NetBSD or OpenBSD. 134 */ 135/* #define PIPE_NODIRECT */ 136 137/* 138 * interfaces to the outside world 139 */ 140static fo_rdwr_t pipe_read; 141static fo_rdwr_t pipe_write; 142static fo_truncate_t pipe_truncate; 143static fo_ioctl_t pipe_ioctl; 144static fo_poll_t pipe_poll; 145static fo_kqfilter_t pipe_kqfilter; 146static fo_stat_t pipe_stat; 147static fo_close_t pipe_close; 148 149static struct fileops pipeops = { 150 .fo_read = pipe_read, 151 .fo_write = pipe_write, 152 .fo_truncate = pipe_truncate, 153 .fo_ioctl = pipe_ioctl, 154 .fo_poll = pipe_poll, 155 .fo_kqfilter = pipe_kqfilter, 156 .fo_stat = pipe_stat, 157 .fo_close = pipe_close, 158 .fo_flags = DFLAG_PASSABLE 159}; 160 161static void filt_pipedetach(struct knote *kn); 162static int filt_piperead(struct knote *kn, long hint); 163static int filt_pipewrite(struct knote *kn, long hint); 164 165static struct filterops pipe_rfiltops = { 166 .f_isfd = 1, 167 .f_detach = filt_pipedetach, 168 .f_event = filt_piperead 169}; 170static struct filterops pipe_wfiltops = { 171 .f_isfd = 1, 172 .f_detach = filt_pipedetach, 173 .f_event = filt_pipewrite 174}; 175 176/* 177 * Default pipe buffer size(s), this can be kind-of large now because pipe 178 * space is pageable. The pipe code will try to maintain locality of 179 * reference for performance reasons, so small amounts of outstanding I/O 180 * will not wipe the cache. 181 */ 182#define MINPIPESIZE (PIPE_SIZE/3) 183#define MAXPIPESIZE (2*PIPE_SIZE/3) 184 185static long amountpipekva; 186static int pipefragretry; 187static int pipeallocfail; 188static int piperesizefail; 189static int piperesizeallowed = 1; 190 191SYSCTL_LONG(_kern_ipc, OID_AUTO, maxpipekva, CTLFLAG_RDTUN, 192 &maxpipekva, 0, "Pipe KVA limit"); 193SYSCTL_LONG(_kern_ipc, OID_AUTO, pipekva, CTLFLAG_RD, 194 &amountpipekva, 0, "Pipe KVA usage"); 195SYSCTL_INT(_kern_ipc, OID_AUTO, pipefragretry, CTLFLAG_RD, 196 &pipefragretry, 0, "Pipe allocation retries due to fragmentation"); 197SYSCTL_INT(_kern_ipc, OID_AUTO, pipeallocfail, CTLFLAG_RD, 198 &pipeallocfail, 0, "Pipe allocation failures"); 199SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizefail, CTLFLAG_RD, 200 &piperesizefail, 0, "Pipe resize failures"); 201SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizeallowed, CTLFLAG_RW, 202 &piperesizeallowed, 0, "Pipe resizing allowed"); 203 204static void pipeinit(void *dummy __unused); 205static void pipeclose(struct pipe *cpipe); 206static void pipe_free_kmem(struct pipe *cpipe); 207static int pipe_create(struct pipe *pipe, int backing); 208static __inline int pipelock(struct pipe *cpipe, int catch); 209static __inline void pipeunlock(struct pipe *cpipe); 210static __inline void pipeselwakeup(struct pipe *cpipe); 211#ifndef PIPE_NODIRECT 212static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio); 213static void pipe_destroy_write_buffer(struct pipe *wpipe); 214static int pipe_direct_write(struct pipe *wpipe, struct uio *uio); 215static void pipe_clone_write_buffer(struct pipe *wpipe); 216#endif 217static int pipespace(struct pipe *cpipe, int size); 218static int pipespace_new(struct pipe *cpipe, int size); 219 220static int pipe_zone_ctor(void *mem, int size, void *arg, int flags); 221static int pipe_zone_init(void *mem, int size, int flags); 222static void pipe_zone_fini(void *mem, int size); 223 224static uma_zone_t pipe_zone; 225 226SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL); 227 228static void 229pipeinit(void *dummy __unused) 230{ 231 232 pipe_zone = uma_zcreate("pipe", sizeof(struct pipepair), 233 pipe_zone_ctor, NULL, pipe_zone_init, pipe_zone_fini, 234 UMA_ALIGN_PTR, 0); 235 KASSERT(pipe_zone != NULL, ("pipe_zone not initialized")); 236} 237 238static int 239pipe_zone_ctor(void *mem, int size, void *arg, int flags) 240{ 241 struct pipepair *pp; 242 struct pipe *rpipe, *wpipe; 243 244 KASSERT(size == sizeof(*pp), ("pipe_zone_ctor: wrong size")); 245 246 pp = (struct pipepair *)mem; 247 248 /* 249 * We zero both pipe endpoints to make sure all the kmem pointers 250 * are NULL, flag fields are zero'd, etc. We timestamp both 251 * endpoints with the same time. 252 */ 253 rpipe = &pp->pp_rpipe; 254 bzero(rpipe, sizeof(*rpipe)); 255 vfs_timestamp(&rpipe->pipe_ctime); 256 rpipe->pipe_atime = rpipe->pipe_mtime = rpipe->pipe_ctime; 257 258 wpipe = &pp->pp_wpipe; 259 bzero(wpipe, sizeof(*wpipe)); 260 wpipe->pipe_ctime = rpipe->pipe_ctime; 261 wpipe->pipe_atime = wpipe->pipe_mtime = rpipe->pipe_ctime; 262 263 rpipe->pipe_peer = wpipe; 264 rpipe->pipe_pair = pp; 265 wpipe->pipe_peer = rpipe; 266 wpipe->pipe_pair = pp; 267 268 /* 269 * Mark both endpoints as present; they will later get free'd 270 * one at a time. When both are free'd, then the whole pair 271 * is released. 272 */ 273 rpipe->pipe_present = PIPE_ACTIVE; 274 wpipe->pipe_present = PIPE_ACTIVE; 275 276 /* 277 * Eventually, the MAC Framework may initialize the label 278 * in ctor or init, but for now we do it elswhere to avoid 279 * blocking in ctor or init. 280 */ 281 pp->pp_label = NULL; 282 283 return (0); 284} 285 286static int 287pipe_zone_init(void *mem, int size, int flags) 288{ 289 struct pipepair *pp; 290 291 KASSERT(size == sizeof(*pp), ("pipe_zone_init: wrong size")); 292 293 pp = (struct pipepair *)mem; 294 295 mtx_init(&pp->pp_mtx, "pipe mutex", NULL, MTX_DEF | MTX_RECURSE); 296 return (0); 297} 298 299static void 300pipe_zone_fini(void *mem, int size) 301{ 302 struct pipepair *pp; 303 304 KASSERT(size == sizeof(*pp), ("pipe_zone_fini: wrong size")); 305 306 pp = (struct pipepair *)mem; 307 308 mtx_destroy(&pp->pp_mtx); 309} 310 311/* 312 * The pipe system call for the DTYPE_PIPE type of pipes. If we fail, let 313 * the zone pick up the pieces via pipeclose(). 314 */ 315int 316kern_pipe(struct thread *td, int fildes[2]) 317{ 318 struct filedesc *fdp = td->td_proc->p_fd; 319 struct file *rf, *wf; 320 struct pipepair *pp; 321 struct pipe *rpipe, *wpipe; 322 int fd, error; 323 324 pp = uma_zalloc(pipe_zone, M_WAITOK); 325#ifdef MAC 326 /* 327 * The MAC label is shared between the connected endpoints. As a 328 * result mac_pipe_init() and mac_pipe_create() are called once 329 * for the pair, and not on the endpoints. 330 */ 331 mac_pipe_init(pp); 332 mac_pipe_create(td->td_ucred, pp); 333#endif 334 rpipe = &pp->pp_rpipe; 335 wpipe = &pp->pp_wpipe; 336 337 knlist_init_mtx(&rpipe->pipe_sel.si_note, PIPE_MTX(rpipe)); 338 knlist_init_mtx(&wpipe->pipe_sel.si_note, PIPE_MTX(wpipe)); 339 340 /* Only the forward direction pipe is backed by default */ 341 if ((error = pipe_create(rpipe, 1)) != 0 || 342 (error = pipe_create(wpipe, 0)) != 0) { 343 pipeclose(rpipe); 344 pipeclose(wpipe); 345 return (error); 346 } 347 348 rpipe->pipe_state |= PIPE_DIRECTOK; 349 wpipe->pipe_state |= PIPE_DIRECTOK; 350 351 error = falloc(td, &rf, &fd); 352 if (error) { 353 pipeclose(rpipe); 354 pipeclose(wpipe); 355 return (error); 356 } 357 /* An extra reference on `rf' has been held for us by falloc(). */ 358 fildes[0] = fd; 359 360 /* 361 * Warning: once we've gotten past allocation of the fd for the 362 * read-side, we can only drop the read side via fdrop() in order 363 * to avoid races against processes which manage to dup() the read 364 * side while we are blocked trying to allocate the write side. 365 */ 366 finit(rf, FREAD | FWRITE, DTYPE_PIPE, rpipe, &pipeops); 367 error = falloc(td, &wf, &fd); 368 if (error) { 369 fdclose(fdp, rf, fildes[0], td); 370 fdrop(rf, td); 371 /* rpipe has been closed by fdrop(). */ 372 pipeclose(wpipe); 373 return (error); 374 } 375 /* An extra reference on `wf' has been held for us by falloc(). */ 376 finit(wf, FREAD | FWRITE, DTYPE_PIPE, wpipe, &pipeops); 377 fdrop(wf, td); 378 fildes[1] = fd; 379 fdrop(rf, td); 380 381 return (0); 382} 383 384/* ARGSUSED */ 385int 386pipe(struct thread *td, struct pipe_args *uap) 387{ 388 int error; 389 int fildes[2]; 390 391 error = kern_pipe(td, fildes); 392 if (error) 393 return (error); 394 395 td->td_retval[0] = fildes[0]; 396 td->td_retval[1] = fildes[1]; 397 398 return (0); 399} 400 401/* 402 * Allocate kva for pipe circular buffer, the space is pageable 403 * This routine will 'realloc' the size of a pipe safely, if it fails 404 * it will retain the old buffer. 405 * If it fails it will return ENOMEM. 406 */ 407static int 408pipespace_new(cpipe, size) 409 struct pipe *cpipe; 410 int size; 411{ 412 caddr_t buffer; 413 int error, cnt, firstseg; 414 static int curfail = 0; 415 static struct timeval lastfail; 416 417 KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked")); 418 KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW), 419 ("pipespace: resize of direct writes not allowed")); 420retry: 421 cnt = cpipe->pipe_buffer.cnt; 422 if (cnt > size) 423 size = cnt; 424 425 size = round_page(size); 426 buffer = (caddr_t) vm_map_min(pipe_map); 427 428 error = vm_map_find(pipe_map, NULL, 0, 429 (vm_offset_t *) &buffer, size, 1, 430 VM_PROT_ALL, VM_PROT_ALL, 0); 431 if (error != KERN_SUCCESS) { 432 if ((cpipe->pipe_buffer.buffer == NULL) && 433 (size > SMALL_PIPE_SIZE)) { 434 size = SMALL_PIPE_SIZE; 435 pipefragretry++; 436 goto retry; 437 } 438 if (cpipe->pipe_buffer.buffer == NULL) { 439 pipeallocfail++; 440 if (ppsratecheck(&lastfail, &curfail, 1)) 441 printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n"); 442 } else { 443 piperesizefail++; 444 } 445 return (ENOMEM); 446 } 447 448 /* copy data, then free old resources if we're resizing */ 449 if (cnt > 0) { 450 if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) { 451 firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out; 452 bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out], 453 buffer, firstseg); 454 if ((cnt - firstseg) > 0) 455 bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg], 456 cpipe->pipe_buffer.in); 457 } else { 458 bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out], 459 buffer, cnt); 460 } 461 } 462 pipe_free_kmem(cpipe); 463 cpipe->pipe_buffer.buffer = buffer; 464 cpipe->pipe_buffer.size = size; 465 cpipe->pipe_buffer.in = cnt; 466 cpipe->pipe_buffer.out = 0; 467 cpipe->pipe_buffer.cnt = cnt; 468 atomic_add_long(&amountpipekva, cpipe->pipe_buffer.size); 469 return (0); 470} 471 472/* 473 * Wrapper for pipespace_new() that performs locking assertions. 474 */ 475static int 476pipespace(cpipe, size) 477 struct pipe *cpipe; 478 int size; 479{ 480 481 KASSERT(cpipe->pipe_state & PIPE_LOCKFL, 482 ("Unlocked pipe passed to pipespace")); 483 return (pipespace_new(cpipe, size)); 484} 485 486/* 487 * lock a pipe for I/O, blocking other access 488 */ 489static __inline int 490pipelock(cpipe, catch) 491 struct pipe *cpipe; 492 int catch; 493{ 494 int error; 495 496 PIPE_LOCK_ASSERT(cpipe, MA_OWNED); 497 while (cpipe->pipe_state & PIPE_LOCKFL) { 498 cpipe->pipe_state |= PIPE_LWANT; 499 error = msleep(cpipe, PIPE_MTX(cpipe), 500 catch ? (PRIBIO | PCATCH) : PRIBIO, 501 "pipelk", 0); 502 if (error != 0) 503 return (error); 504 } 505 cpipe->pipe_state |= PIPE_LOCKFL; 506 return (0); 507} 508 509/* 510 * unlock a pipe I/O lock 511 */ 512static __inline void 513pipeunlock(cpipe) 514 struct pipe *cpipe; 515{ 516 517 PIPE_LOCK_ASSERT(cpipe, MA_OWNED); 518 KASSERT(cpipe->pipe_state & PIPE_LOCKFL, 519 ("Unlocked pipe passed to pipeunlock")); 520 cpipe->pipe_state &= ~PIPE_LOCKFL; 521 if (cpipe->pipe_state & PIPE_LWANT) { 522 cpipe->pipe_state &= ~PIPE_LWANT; 523 wakeup(cpipe); 524 } 525} 526 527static __inline void 528pipeselwakeup(cpipe) 529 struct pipe *cpipe; 530{ 531 532 PIPE_LOCK_ASSERT(cpipe, MA_OWNED); 533 if (cpipe->pipe_state & PIPE_SEL) { 534 selwakeuppri(&cpipe->pipe_sel, PSOCK); 535 if (!SEL_WAITING(&cpipe->pipe_sel)) 536 cpipe->pipe_state &= ~PIPE_SEL; 537 } 538 if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio) 539 pgsigio(&cpipe->pipe_sigio, SIGIO, 0); 540 KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0); 541} 542 543/* 544 * Initialize and allocate VM and memory for pipe. The structure 545 * will start out zero'd from the ctor, so we just manage the kmem. 546 */ 547static int 548pipe_create(pipe, backing) 549 struct pipe *pipe; 550 int backing; 551{ 552 int error; 553 554 if (backing) { 555 if (amountpipekva > maxpipekva / 2) 556 error = pipespace_new(pipe, SMALL_PIPE_SIZE); 557 else 558 error = pipespace_new(pipe, PIPE_SIZE); 559 } else { 560 /* If we're not backing this pipe, no need to do anything. */ 561 error = 0; 562 } 563 return (error); 564} 565 566/* ARGSUSED */ 567static int 568pipe_read(fp, uio, active_cred, flags, td) 569 struct file *fp; 570 struct uio *uio; 571 struct ucred *active_cred; 572 struct thread *td; 573 int flags; 574{ 575 struct pipe *rpipe = fp->f_data; 576 int error; 577 int nread = 0; 578 u_int size; 579 580 PIPE_LOCK(rpipe); 581 ++rpipe->pipe_busy; 582 error = pipelock(rpipe, 1); 583 if (error) 584 goto unlocked_error; 585 586#ifdef MAC 587 error = mac_pipe_check_read(active_cred, rpipe->pipe_pair); 588 if (error) 589 goto locked_error; 590#endif 591 if (amountpipekva > (3 * maxpipekva) / 4) { 592 if (!(rpipe->pipe_state & PIPE_DIRECTW) && 593 (rpipe->pipe_buffer.size > SMALL_PIPE_SIZE) && 594 (rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) && 595 (piperesizeallowed == 1)) { 596 PIPE_UNLOCK(rpipe); 597 pipespace(rpipe, SMALL_PIPE_SIZE); 598 PIPE_LOCK(rpipe); 599 } 600 } 601 602 while (uio->uio_resid) { 603 /* 604 * normal pipe buffer receive 605 */ 606 if (rpipe->pipe_buffer.cnt > 0) { 607 size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out; 608 if (size > rpipe->pipe_buffer.cnt) 609 size = rpipe->pipe_buffer.cnt; 610 if (size > (u_int) uio->uio_resid) 611 size = (u_int) uio->uio_resid; 612 613 PIPE_UNLOCK(rpipe); 614 error = uiomove( 615 &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out], 616 size, uio); 617 PIPE_LOCK(rpipe); 618 if (error) 619 break; 620 621 rpipe->pipe_buffer.out += size; 622 if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size) 623 rpipe->pipe_buffer.out = 0; 624 625 rpipe->pipe_buffer.cnt -= size; 626 627 /* 628 * If there is no more to read in the pipe, reset 629 * its pointers to the beginning. This improves 630 * cache hit stats. 631 */ 632 if (rpipe->pipe_buffer.cnt == 0) { 633 rpipe->pipe_buffer.in = 0; 634 rpipe->pipe_buffer.out = 0; 635 } 636 nread += size; 637#ifndef PIPE_NODIRECT 638 /* 639 * Direct copy, bypassing a kernel buffer. 640 */ 641 } else if ((size = rpipe->pipe_map.cnt) && 642 (rpipe->pipe_state & PIPE_DIRECTW)) { 643 if (size > (u_int) uio->uio_resid) 644 size = (u_int) uio->uio_resid; 645 646 PIPE_UNLOCK(rpipe); 647 error = uiomove_fromphys(rpipe->pipe_map.ms, 648 rpipe->pipe_map.pos, size, uio); 649 PIPE_LOCK(rpipe); 650 if (error) 651 break; 652 nread += size; 653 rpipe->pipe_map.pos += size; 654 rpipe->pipe_map.cnt -= size; 655 if (rpipe->pipe_map.cnt == 0) { 656 rpipe->pipe_state &= ~PIPE_DIRECTW; 657 wakeup(rpipe); 658 } 659#endif 660 } else { 661 /* 662 * detect EOF condition 663 * read returns 0 on EOF, no need to set error 664 */ 665 if (rpipe->pipe_state & PIPE_EOF) 666 break; 667 668 /* 669 * If the "write-side" has been blocked, wake it up now. 670 */ 671 if (rpipe->pipe_state & PIPE_WANTW) { 672 rpipe->pipe_state &= ~PIPE_WANTW; 673 wakeup(rpipe); 674 } 675 676 /* 677 * Break if some data was read. 678 */ 679 if (nread > 0) 680 break; 681 682 /* 683 * Unlock the pipe buffer for our remaining processing. 684 * We will either break out with an error or we will 685 * sleep and relock to loop. 686 */ 687 pipeunlock(rpipe); 688 689 /* 690 * Handle non-blocking mode operation or 691 * wait for more data. 692 */ 693 if (fp->f_flag & FNONBLOCK) { 694 error = EAGAIN; 695 } else { 696 rpipe->pipe_state |= PIPE_WANTR; 697 if ((error = msleep(rpipe, PIPE_MTX(rpipe), 698 PRIBIO | PCATCH, 699 "piperd", 0)) == 0) 700 error = pipelock(rpipe, 1); 701 } 702 if (error) 703 goto unlocked_error; 704 } 705 } 706#ifdef MAC 707locked_error: 708#endif 709 pipeunlock(rpipe); 710 711 /* XXX: should probably do this before getting any locks. */ 712 if (error == 0) 713 vfs_timestamp(&rpipe->pipe_atime); 714unlocked_error: 715 --rpipe->pipe_busy; 716 717 /* 718 * PIPE_WANT processing only makes sense if pipe_busy is 0. 719 */ 720 if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) { 721 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW); 722 wakeup(rpipe); 723 } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) { 724 /* 725 * Handle write blocking hysteresis. 726 */ 727 if (rpipe->pipe_state & PIPE_WANTW) { 728 rpipe->pipe_state &= ~PIPE_WANTW; 729 wakeup(rpipe); 730 } 731 } 732 733 if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF) 734 pipeselwakeup(rpipe); 735 736 PIPE_UNLOCK(rpipe); 737 return (error); 738} 739 740#ifndef PIPE_NODIRECT 741/* 742 * Map the sending processes' buffer into kernel space and wire it. 743 * This is similar to a physical write operation. 744 */ 745static int 746pipe_build_write_buffer(wpipe, uio) 747 struct pipe *wpipe; 748 struct uio *uio; 749{ 750 pmap_t pmap; 751 u_int size; 752 int i; 753 vm_offset_t addr, endaddr; 754 755 PIPE_LOCK_ASSERT(wpipe, MA_NOTOWNED); 756 KASSERT(wpipe->pipe_state & PIPE_DIRECTW, 757 ("Clone attempt on non-direct write pipe!")); 758 759 size = (u_int) uio->uio_iov->iov_len; 760 if (size > wpipe->pipe_buffer.size) 761 size = wpipe->pipe_buffer.size; 762 763 pmap = vmspace_pmap(curproc->p_vmspace); 764 endaddr = round_page((vm_offset_t)uio->uio_iov->iov_base + size); 765 addr = trunc_page((vm_offset_t)uio->uio_iov->iov_base); 766 if (endaddr < addr) 767 return (EFAULT); 768 for (i = 0; addr < endaddr; addr += PAGE_SIZE, i++) { 769 /* 770 * vm_fault_quick() can sleep. 771 */ 772 race: 773 if (vm_fault_quick((caddr_t)addr, VM_PROT_READ) < 0) { 774 vm_page_unhold_pages(wpipe->pipe_map.ms, i); 775 return (EFAULT); 776 } 777 wpipe->pipe_map.ms[i] = pmap_extract_and_hold(pmap, addr, 778 VM_PROT_READ); 779 if (wpipe->pipe_map.ms[i] == NULL) 780 goto race; 781 } 782 783/* 784 * set up the control block 785 */ 786 wpipe->pipe_map.npages = i; 787 wpipe->pipe_map.pos = 788 ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK; 789 wpipe->pipe_map.cnt = size; 790 791/* 792 * and update the uio data 793 */ 794 795 uio->uio_iov->iov_len -= size; 796 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size; 797 if (uio->uio_iov->iov_len == 0) 798 uio->uio_iov++; 799 uio->uio_resid -= size; 800 uio->uio_offset += size; 801 return (0); 802} 803 804/* 805 * unmap and unwire the process buffer 806 */ 807static void 808pipe_destroy_write_buffer(wpipe) 809 struct pipe *wpipe; 810{ 811 812 PIPE_LOCK_ASSERT(wpipe, MA_OWNED); 813 vm_page_unhold_pages(wpipe->pipe_map.ms, wpipe->pipe_map.npages); 814 wpipe->pipe_map.npages = 0; 815} 816 817/* 818 * In the case of a signal, the writing process might go away. This 819 * code copies the data into the circular buffer so that the source 820 * pages can be freed without loss of data. 821 */ 822static void 823pipe_clone_write_buffer(wpipe) 824 struct pipe *wpipe; 825{ 826 struct uio uio; 827 struct iovec iov; 828 int size; 829 int pos; 830 831 PIPE_LOCK_ASSERT(wpipe, MA_OWNED); 832 size = wpipe->pipe_map.cnt; 833 pos = wpipe->pipe_map.pos; 834 835 wpipe->pipe_buffer.in = size; 836 wpipe->pipe_buffer.out = 0; 837 wpipe->pipe_buffer.cnt = size; 838 wpipe->pipe_state &= ~PIPE_DIRECTW; 839 840 PIPE_UNLOCK(wpipe); 841 iov.iov_base = wpipe->pipe_buffer.buffer; 842 iov.iov_len = size; 843 uio.uio_iov = &iov; 844 uio.uio_iovcnt = 1; 845 uio.uio_offset = 0; 846 uio.uio_resid = size; 847 uio.uio_segflg = UIO_SYSSPACE; 848 uio.uio_rw = UIO_READ; 849 uio.uio_td = curthread; 850 uiomove_fromphys(wpipe->pipe_map.ms, pos, size, &uio); 851 PIPE_LOCK(wpipe); 852 pipe_destroy_write_buffer(wpipe); 853} 854 855/* 856 * This implements the pipe buffer write mechanism. Note that only 857 * a direct write OR a normal pipe write can be pending at any given time. 858 * If there are any characters in the pipe buffer, the direct write will 859 * be deferred until the receiving process grabs all of the bytes from 860 * the pipe buffer. Then the direct mapping write is set-up. 861 */ 862static int 863pipe_direct_write(wpipe, uio) 864 struct pipe *wpipe; 865 struct uio *uio; 866{ 867 int error; 868 869retry: 870 PIPE_LOCK_ASSERT(wpipe, MA_OWNED); 871 error = pipelock(wpipe, 1); 872 if (wpipe->pipe_state & PIPE_EOF) 873 error = EPIPE; 874 if (error) { 875 pipeunlock(wpipe); 876 goto error1; 877 } 878 while (wpipe->pipe_state & PIPE_DIRECTW) { 879 if (wpipe->pipe_state & PIPE_WANTR) { 880 wpipe->pipe_state &= ~PIPE_WANTR; 881 wakeup(wpipe); 882 } 883 pipeselwakeup(wpipe); 884 wpipe->pipe_state |= PIPE_WANTW; 885 pipeunlock(wpipe); 886 error = msleep(wpipe, PIPE_MTX(wpipe), 887 PRIBIO | PCATCH, "pipdww", 0); 888 if (error) 889 goto error1; 890 else 891 goto retry; 892 } 893 wpipe->pipe_map.cnt = 0; /* transfer not ready yet */ 894 if (wpipe->pipe_buffer.cnt > 0) { 895 if (wpipe->pipe_state & PIPE_WANTR) { 896 wpipe->pipe_state &= ~PIPE_WANTR; 897 wakeup(wpipe); 898 } 899 pipeselwakeup(wpipe); 900 wpipe->pipe_state |= PIPE_WANTW; 901 pipeunlock(wpipe); 902 error = msleep(wpipe, PIPE_MTX(wpipe), 903 PRIBIO | PCATCH, "pipdwc", 0); 904 if (error) 905 goto error1; 906 else 907 goto retry; 908 } 909 910 wpipe->pipe_state |= PIPE_DIRECTW; 911 912 PIPE_UNLOCK(wpipe); 913 error = pipe_build_write_buffer(wpipe, uio); 914 PIPE_LOCK(wpipe); 915 if (error) { 916 wpipe->pipe_state &= ~PIPE_DIRECTW; 917 pipeunlock(wpipe); 918 goto error1; 919 } 920 921 error = 0; 922 while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) { 923 if (wpipe->pipe_state & PIPE_EOF) { 924 pipe_destroy_write_buffer(wpipe); 925 pipeselwakeup(wpipe); 926 pipeunlock(wpipe); 927 error = EPIPE; 928 goto error1; 929 } 930 if (wpipe->pipe_state & PIPE_WANTR) { 931 wpipe->pipe_state &= ~PIPE_WANTR; 932 wakeup(wpipe); 933 } 934 pipeselwakeup(wpipe); 935 pipeunlock(wpipe); 936 error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH, 937 "pipdwt", 0); 938 pipelock(wpipe, 0); 939 } 940 941 if (wpipe->pipe_state & PIPE_EOF) 942 error = EPIPE; 943 if (wpipe->pipe_state & PIPE_DIRECTW) { 944 /* 945 * this bit of trickery substitutes a kernel buffer for 946 * the process that might be going away. 947 */ 948 pipe_clone_write_buffer(wpipe); 949 } else { 950 pipe_destroy_write_buffer(wpipe); 951 } 952 pipeunlock(wpipe); 953 return (error); 954 955error1: 956 wakeup(wpipe); 957 return (error); 958} 959#endif 960 961static int 962pipe_write(fp, uio, active_cred, flags, td) 963 struct file *fp; 964 struct uio *uio; 965 struct ucred *active_cred; 966 struct thread *td; 967 int flags; 968{ 969 int error = 0; 970 int desiredsize, orig_resid; 971 struct pipe *wpipe, *rpipe; 972 973 rpipe = fp->f_data; 974 wpipe = rpipe->pipe_peer; 975 976 PIPE_LOCK(rpipe); 977 error = pipelock(wpipe, 1); 978 if (error) { 979 PIPE_UNLOCK(rpipe); 980 return (error); 981 } 982 /* 983 * detect loss of pipe read side, issue SIGPIPE if lost. 984 */ 985 if (wpipe->pipe_present != PIPE_ACTIVE || 986 (wpipe->pipe_state & PIPE_EOF)) { 987 pipeunlock(wpipe); 988 PIPE_UNLOCK(rpipe); 989 return (EPIPE); 990 } 991#ifdef MAC 992 error = mac_pipe_check_write(active_cred, wpipe->pipe_pair); 993 if (error) { 994 pipeunlock(wpipe); 995 PIPE_UNLOCK(rpipe); 996 return (error); 997 } 998#endif 999 ++wpipe->pipe_busy; 1000 1001 /* Choose a larger size if it's advantageous */ 1002 desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size); 1003 while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) { 1004 if (piperesizeallowed != 1) 1005 break; 1006 if (amountpipekva > maxpipekva / 2) 1007 break; 1008 if (desiredsize == BIG_PIPE_SIZE) 1009 break; 1010 desiredsize = desiredsize * 2; 1011 } 1012 1013 /* Choose a smaller size if we're in a OOM situation */ 1014 if ((amountpipekva > (3 * maxpipekva) / 4) && 1015 (wpipe->pipe_buffer.size > SMALL_PIPE_SIZE) && 1016 (wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) && 1017 (piperesizeallowed == 1)) 1018 desiredsize = SMALL_PIPE_SIZE; 1019 1020 /* Resize if the above determined that a new size was necessary */ 1021 if ((desiredsize != wpipe->pipe_buffer.size) && 1022 ((wpipe->pipe_state & PIPE_DIRECTW) == 0)) { 1023 PIPE_UNLOCK(wpipe); 1024 pipespace(wpipe, desiredsize); 1025 PIPE_LOCK(wpipe); 1026 } 1027 if (wpipe->pipe_buffer.size == 0) { 1028 /* 1029 * This can only happen for reverse direction use of pipes 1030 * in a complete OOM situation. 1031 */ 1032 error = ENOMEM; 1033 --wpipe->pipe_busy; 1034 pipeunlock(wpipe); 1035 PIPE_UNLOCK(wpipe); 1036 return (error); 1037 } 1038 1039 pipeunlock(wpipe); 1040 1041 orig_resid = uio->uio_resid; 1042 1043 while (uio->uio_resid) { 1044 int space; 1045 1046 pipelock(wpipe, 0); 1047 if (wpipe->pipe_state & PIPE_EOF) { 1048 pipeunlock(wpipe); 1049 error = EPIPE; 1050 break; 1051 } 1052#ifndef PIPE_NODIRECT 1053 /* 1054 * If the transfer is large, we can gain performance if 1055 * we do process-to-process copies directly. 1056 * If the write is non-blocking, we don't use the 1057 * direct write mechanism. 1058 * 1059 * The direct write mechanism will detect the reader going 1060 * away on us. 1061 */ 1062 if (uio->uio_segflg == UIO_USERSPACE && 1063 uio->uio_iov->iov_len >= PIPE_MINDIRECT && 1064 wpipe->pipe_buffer.size >= PIPE_MINDIRECT && 1065 (fp->f_flag & FNONBLOCK) == 0) { 1066 pipeunlock(wpipe); 1067 error = pipe_direct_write(wpipe, uio); 1068 if (error) 1069 break; 1070 continue; 1071 } 1072#endif 1073 1074 /* 1075 * Pipe buffered writes cannot be coincidental with 1076 * direct writes. We wait until the currently executing 1077 * direct write is completed before we start filling the 1078 * pipe buffer. We break out if a signal occurs or the 1079 * reader goes away. 1080 */ 1081 if (wpipe->pipe_state & PIPE_DIRECTW) { 1082 if (wpipe->pipe_state & PIPE_WANTR) { 1083 wpipe->pipe_state &= ~PIPE_WANTR; 1084 wakeup(wpipe); 1085 } 1086 pipeselwakeup(wpipe); 1087 wpipe->pipe_state |= PIPE_WANTW; 1088 pipeunlock(wpipe); 1089 error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH, 1090 "pipbww", 0); 1091 if (error) 1092 break; 1093 else 1094 continue; 1095 } 1096 1097 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt; 1098 1099 /* Writes of size <= PIPE_BUF must be atomic. */ 1100 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF)) 1101 space = 0; 1102 1103 if (space > 0) { 1104 int size; /* Transfer size */ 1105 int segsize; /* first segment to transfer */ 1106 1107 /* 1108 * Transfer size is minimum of uio transfer 1109 * and free space in pipe buffer. 1110 */ 1111 if (space > uio->uio_resid) 1112 size = uio->uio_resid; 1113 else 1114 size = space; 1115 /* 1116 * First segment to transfer is minimum of 1117 * transfer size and contiguous space in 1118 * pipe buffer. If first segment to transfer 1119 * is less than the transfer size, we've got 1120 * a wraparound in the buffer. 1121 */ 1122 segsize = wpipe->pipe_buffer.size - 1123 wpipe->pipe_buffer.in; 1124 if (segsize > size) 1125 segsize = size; 1126 1127 /* Transfer first segment */ 1128 1129 PIPE_UNLOCK(rpipe); 1130 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in], 1131 segsize, uio); 1132 PIPE_LOCK(rpipe); 1133 1134 if (error == 0 && segsize < size) { 1135 KASSERT(wpipe->pipe_buffer.in + segsize == 1136 wpipe->pipe_buffer.size, 1137 ("Pipe buffer wraparound disappeared")); 1138 /* 1139 * Transfer remaining part now, to 1140 * support atomic writes. Wraparound 1141 * happened. 1142 */ 1143 1144 PIPE_UNLOCK(rpipe); 1145 error = uiomove( 1146 &wpipe->pipe_buffer.buffer[0], 1147 size - segsize, uio); 1148 PIPE_LOCK(rpipe); 1149 } 1150 if (error == 0) { 1151 wpipe->pipe_buffer.in += size; 1152 if (wpipe->pipe_buffer.in >= 1153 wpipe->pipe_buffer.size) { 1154 KASSERT(wpipe->pipe_buffer.in == 1155 size - segsize + 1156 wpipe->pipe_buffer.size, 1157 ("Expected wraparound bad")); 1158 wpipe->pipe_buffer.in = size - segsize; 1159 } 1160 1161 wpipe->pipe_buffer.cnt += size; 1162 KASSERT(wpipe->pipe_buffer.cnt <= 1163 wpipe->pipe_buffer.size, 1164 ("Pipe buffer overflow")); 1165 } 1166 pipeunlock(wpipe); 1167 if (error != 0) 1168 break; 1169 } else { 1170 /* 1171 * If the "read-side" has been blocked, wake it up now. 1172 */ 1173 if (wpipe->pipe_state & PIPE_WANTR) { 1174 wpipe->pipe_state &= ~PIPE_WANTR; 1175 wakeup(wpipe); 1176 } 1177 1178 /* 1179 * don't block on non-blocking I/O 1180 */ 1181 if (fp->f_flag & FNONBLOCK) { 1182 error = EAGAIN; 1183 pipeunlock(wpipe); 1184 break; 1185 } 1186 1187 /* 1188 * We have no more space and have something to offer, 1189 * wake up select/poll. 1190 */ 1191 pipeselwakeup(wpipe); 1192 1193 wpipe->pipe_state |= PIPE_WANTW; 1194 pipeunlock(wpipe); 1195 error = msleep(wpipe, PIPE_MTX(rpipe), 1196 PRIBIO | PCATCH, "pipewr", 0); 1197 if (error != 0) 1198 break; 1199 } 1200 } 1201 1202 pipelock(wpipe, 0); 1203 --wpipe->pipe_busy; 1204 1205 if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) { 1206 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR); 1207 wakeup(wpipe); 1208 } else if (wpipe->pipe_buffer.cnt > 0) { 1209 /* 1210 * If we have put any characters in the buffer, we wake up 1211 * the reader. 1212 */ 1213 if (wpipe->pipe_state & PIPE_WANTR) { 1214 wpipe->pipe_state &= ~PIPE_WANTR; 1215 wakeup(wpipe); 1216 } 1217 } 1218 1219 /* 1220 * Don't return EPIPE if I/O was successful 1221 */ 1222 if ((wpipe->pipe_buffer.cnt == 0) && 1223 (uio->uio_resid == 0) && 1224 (error == EPIPE)) { 1225 error = 0; 1226 } 1227 1228 if (error == 0) 1229 vfs_timestamp(&wpipe->pipe_mtime); 1230 1231 /* 1232 * We have something to offer, 1233 * wake up select/poll. 1234 */ 1235 if (wpipe->pipe_buffer.cnt) 1236 pipeselwakeup(wpipe); 1237 1238 pipeunlock(wpipe); 1239 PIPE_UNLOCK(rpipe); 1240 return (error); 1241} 1242 1243/* ARGSUSED */ 1244static int 1245pipe_truncate(fp, length, active_cred, td) 1246 struct file *fp; 1247 off_t length; 1248 struct ucred *active_cred; 1249 struct thread *td; 1250{ 1251 1252 return (EINVAL); 1253} 1254 1255/* 1256 * we implement a very minimal set of ioctls for compatibility with sockets. 1257 */ 1258static int 1259pipe_ioctl(fp, cmd, data, active_cred, td) 1260 struct file *fp; 1261 u_long cmd; 1262 void *data; 1263 struct ucred *active_cred; 1264 struct thread *td; 1265{ 1266 struct pipe *mpipe = fp->f_data; 1267 int error; 1268 1269 PIPE_LOCK(mpipe); 1270 1271#ifdef MAC 1272 error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data); 1273 if (error) { 1274 PIPE_UNLOCK(mpipe); 1275 return (error); 1276 } 1277#endif 1278 1279 error = 0; 1280 switch (cmd) { 1281 1282 case FIONBIO: 1283 break; 1284 1285 case FIOASYNC: 1286 if (*(int *)data) { 1287 mpipe->pipe_state |= PIPE_ASYNC; 1288 } else { 1289 mpipe->pipe_state &= ~PIPE_ASYNC; 1290 } 1291 break; 1292 1293 case FIONREAD: 1294 if (mpipe->pipe_state & PIPE_DIRECTW) 1295 *(int *)data = mpipe->pipe_map.cnt; 1296 else 1297 *(int *)data = mpipe->pipe_buffer.cnt; 1298 break; 1299 1300 case FIOSETOWN: 1301 PIPE_UNLOCK(mpipe); 1302 error = fsetown(*(int *)data, &mpipe->pipe_sigio); 1303 goto out_unlocked; 1304 1305 case FIOGETOWN: 1306 *(int *)data = fgetown(&mpipe->pipe_sigio); 1307 break; 1308 1309 /* This is deprecated, FIOSETOWN should be used instead. */ 1310 case TIOCSPGRP: 1311 PIPE_UNLOCK(mpipe); 1312 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio); 1313 goto out_unlocked; 1314 1315 /* This is deprecated, FIOGETOWN should be used instead. */ 1316 case TIOCGPGRP: 1317 *(int *)data = -fgetown(&mpipe->pipe_sigio); 1318 break; 1319 1320 default: 1321 error = ENOTTY; 1322 break; 1323 } 1324 PIPE_UNLOCK(mpipe); 1325out_unlocked: 1326 return (error); 1327} 1328 1329static int 1330pipe_poll(fp, events, active_cred, td) 1331 struct file *fp; 1332 int events; 1333 struct ucred *active_cred; 1334 struct thread *td; 1335{ 1336 struct pipe *rpipe = fp->f_data; 1337 struct pipe *wpipe; 1338 int revents = 0; 1339#ifdef MAC 1340 int error; 1341#endif 1342 1343 wpipe = rpipe->pipe_peer; 1344 PIPE_LOCK(rpipe); 1345#ifdef MAC 1346 error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair); 1347 if (error) 1348 goto locked_error; 1349#endif 1350 if (events & (POLLIN | POLLRDNORM)) 1351 if ((rpipe->pipe_state & PIPE_DIRECTW) || 1352 (rpipe->pipe_buffer.cnt > 0)) 1353 revents |= events & (POLLIN | POLLRDNORM); 1354 1355 if (events & (POLLOUT | POLLWRNORM)) 1356 if (wpipe->pipe_present != PIPE_ACTIVE || 1357 (wpipe->pipe_state & PIPE_EOF) || 1358 (((wpipe->pipe_state & PIPE_DIRECTW) == 0) && 1359 (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF)) 1360 revents |= events & (POLLOUT | POLLWRNORM); 1361 1362 if ((events & POLLINIGNEOF) == 0) { 1363 if (rpipe->pipe_state & PIPE_EOF) { 1364 revents |= (events & (POLLIN | POLLRDNORM)); 1365 if (wpipe->pipe_present != PIPE_ACTIVE || 1366 (wpipe->pipe_state & PIPE_EOF)) 1367 revents |= POLLHUP; 1368 } 1369 } 1370 1371 if (revents == 0) { 1372 if (events & (POLLIN | POLLRDNORM)) { 1373 selrecord(td, &rpipe->pipe_sel); 1374 if (SEL_WAITING(&rpipe->pipe_sel)) 1375 rpipe->pipe_state |= PIPE_SEL; 1376 } 1377 1378 if (events & (POLLOUT | POLLWRNORM)) { 1379 selrecord(td, &wpipe->pipe_sel); 1380 if (SEL_WAITING(&wpipe->pipe_sel)) 1381 wpipe->pipe_state |= PIPE_SEL; 1382 } 1383 } 1384#ifdef MAC 1385locked_error: 1386#endif 1387 PIPE_UNLOCK(rpipe); 1388 1389 return (revents); 1390} 1391 1392/* 1393 * We shouldn't need locks here as we're doing a read and this should 1394 * be a natural race. 1395 */ 1396static int 1397pipe_stat(fp, ub, active_cred, td) 1398 struct file *fp; 1399 struct stat *ub; 1400 struct ucred *active_cred; 1401 struct thread *td; 1402{ 1403 struct pipe *pipe = fp->f_data; 1404#ifdef MAC 1405 int error; 1406 1407 PIPE_LOCK(pipe); 1408 error = mac_pipe_check_stat(active_cred, pipe->pipe_pair); 1409 PIPE_UNLOCK(pipe); 1410 if (error) 1411 return (error); 1412#endif 1413 bzero(ub, sizeof(*ub)); 1414 ub->st_mode = S_IFIFO; 1415 ub->st_blksize = PAGE_SIZE; 1416 if (pipe->pipe_state & PIPE_DIRECTW) 1417 ub->st_size = pipe->pipe_map.cnt; 1418 else 1419 ub->st_size = pipe->pipe_buffer.cnt; 1420 ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize; 1421 ub->st_atim = pipe->pipe_atime; 1422 ub->st_mtim = pipe->pipe_mtime; 1423 ub->st_ctim = pipe->pipe_ctime; 1424 ub->st_uid = fp->f_cred->cr_uid; 1425 ub->st_gid = fp->f_cred->cr_gid; 1426 /* 1427 * Left as 0: st_dev, st_ino, st_nlink, st_rdev, st_flags, st_gen. 1428 * XXX (st_dev, st_ino) should be unique. 1429 */ 1430 return (0); 1431} 1432 1433/* ARGSUSED */ 1434static int 1435pipe_close(fp, td) 1436 struct file *fp; 1437 struct thread *td; 1438{ 1439 struct pipe *cpipe = fp->f_data; 1440 1441 fp->f_ops = &badfileops; 1442 fp->f_data = NULL; 1443 funsetown(&cpipe->pipe_sigio); 1444 pipeclose(cpipe); 1445 return (0); 1446} 1447 1448static void 1449pipe_free_kmem(cpipe) 1450 struct pipe *cpipe; 1451{ 1452 1453 KASSERT(!mtx_owned(PIPE_MTX(cpipe)), 1454 ("pipe_free_kmem: pipe mutex locked")); 1455 1456 if (cpipe->pipe_buffer.buffer != NULL) { 1457 atomic_subtract_long(&amountpipekva, cpipe->pipe_buffer.size); 1458 vm_map_remove(pipe_map, 1459 (vm_offset_t)cpipe->pipe_buffer.buffer, 1460 (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size); 1461 cpipe->pipe_buffer.buffer = NULL; 1462 } 1463#ifndef PIPE_NODIRECT 1464 { 1465 cpipe->pipe_map.cnt = 0; 1466 cpipe->pipe_map.pos = 0; 1467 cpipe->pipe_map.npages = 0; 1468 } 1469#endif 1470} 1471 1472/* 1473 * shutdown the pipe 1474 */ 1475static void 1476pipeclose(cpipe) 1477 struct pipe *cpipe; 1478{ 1479 struct pipepair *pp; 1480 struct pipe *ppipe; 1481 1482 KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL")); 1483 1484 PIPE_LOCK(cpipe); 1485 pipelock(cpipe, 0); 1486 pp = cpipe->pipe_pair; 1487 1488 pipeselwakeup(cpipe); 1489 1490 /* 1491 * If the other side is blocked, wake it up saying that 1492 * we want to close it down. 1493 */ 1494 cpipe->pipe_state |= PIPE_EOF; 1495 while (cpipe->pipe_busy) { 1496 wakeup(cpipe); 1497 cpipe->pipe_state |= PIPE_WANT; 1498 pipeunlock(cpipe); 1499 msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0); 1500 pipelock(cpipe, 0); 1501 } 1502 1503 1504 /* 1505 * Disconnect from peer, if any. 1506 */ 1507 ppipe = cpipe->pipe_peer; 1508 if (ppipe->pipe_present == PIPE_ACTIVE) { 1509 pipeselwakeup(ppipe); 1510 1511 ppipe->pipe_state |= PIPE_EOF; 1512 wakeup(ppipe); 1513 KNOTE_LOCKED(&ppipe->pipe_sel.si_note, 0); 1514 } 1515 1516 /* 1517 * Mark this endpoint as free. Release kmem resources. We 1518 * don't mark this endpoint as unused until we've finished 1519 * doing that, or the pipe might disappear out from under 1520 * us. 1521 */ 1522 PIPE_UNLOCK(cpipe); 1523 pipe_free_kmem(cpipe); 1524 PIPE_LOCK(cpipe); 1525 cpipe->pipe_present = PIPE_CLOSING; 1526 pipeunlock(cpipe); 1527 1528 /* 1529 * knlist_clear() may sleep dropping the PIPE_MTX. Set the 1530 * PIPE_FINALIZED, that allows other end to free the 1531 * pipe_pair, only after the knotes are completely dismantled. 1532 */ 1533 knlist_clear(&cpipe->pipe_sel.si_note, 1); 1534 cpipe->pipe_present = PIPE_FINALIZED; 1535 knlist_destroy(&cpipe->pipe_sel.si_note); 1536 1537 /* 1538 * If both endpoints are now closed, release the memory for the 1539 * pipe pair. If not, unlock. 1540 */ 1541 if (ppipe->pipe_present == PIPE_FINALIZED) { 1542 PIPE_UNLOCK(cpipe); 1543#ifdef MAC 1544 mac_pipe_destroy(pp); 1545#endif 1546 uma_zfree(pipe_zone, cpipe->pipe_pair); 1547 } else 1548 PIPE_UNLOCK(cpipe); 1549} 1550 1551/*ARGSUSED*/ 1552static int 1553pipe_kqfilter(struct file *fp, struct knote *kn) 1554{ 1555 struct pipe *cpipe; 1556 1557 cpipe = kn->kn_fp->f_data; 1558 PIPE_LOCK(cpipe); 1559 switch (kn->kn_filter) { 1560 case EVFILT_READ: 1561 kn->kn_fop = &pipe_rfiltops; 1562 break; 1563 case EVFILT_WRITE: 1564 kn->kn_fop = &pipe_wfiltops; 1565 if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) { 1566 /* other end of pipe has been closed */ 1567 PIPE_UNLOCK(cpipe); 1568 return (EPIPE); 1569 } 1570 cpipe = cpipe->pipe_peer; 1571 break; 1572 default: 1573 PIPE_UNLOCK(cpipe); 1574 return (EINVAL); 1575 } 1576 1577 knlist_add(&cpipe->pipe_sel.si_note, kn, 1); 1578 PIPE_UNLOCK(cpipe); 1579 return (0); 1580} 1581 1582static void 1583filt_pipedetach(struct knote *kn) 1584{ 1585 struct pipe *cpipe = (struct pipe *)kn->kn_fp->f_data; 1586 1587 PIPE_LOCK(cpipe); 1588 if (kn->kn_filter == EVFILT_WRITE) 1589 cpipe = cpipe->pipe_peer; 1590 knlist_remove(&cpipe->pipe_sel.si_note, kn, 1); 1591 PIPE_UNLOCK(cpipe); 1592} 1593 1594/*ARGSUSED*/ 1595static int 1596filt_piperead(struct knote *kn, long hint) 1597{ 1598 struct pipe *rpipe = kn->kn_fp->f_data; 1599 struct pipe *wpipe = rpipe->pipe_peer; 1600 int ret; 1601 1602 PIPE_LOCK(rpipe); 1603 kn->kn_data = rpipe->pipe_buffer.cnt; 1604 if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW)) 1605 kn->kn_data = rpipe->pipe_map.cnt; 1606 1607 if ((rpipe->pipe_state & PIPE_EOF) || 1608 wpipe->pipe_present != PIPE_ACTIVE || 1609 (wpipe->pipe_state & PIPE_EOF)) { 1610 kn->kn_flags |= EV_EOF; 1611 PIPE_UNLOCK(rpipe); 1612 return (1); 1613 } 1614 ret = kn->kn_data > 0; 1615 PIPE_UNLOCK(rpipe); 1616 return ret; 1617} 1618 1619/*ARGSUSED*/ 1620static int 1621filt_pipewrite(struct knote *kn, long hint) 1622{ 1623 struct pipe *rpipe = kn->kn_fp->f_data; 1624 struct pipe *wpipe = rpipe->pipe_peer; 1625 1626 PIPE_LOCK(rpipe); 1627 if (wpipe->pipe_present != PIPE_ACTIVE || 1628 (wpipe->pipe_state & PIPE_EOF)) { 1629 kn->kn_data = 0; 1630 kn->kn_flags |= EV_EOF; 1631 PIPE_UNLOCK(rpipe); 1632 return (1); 1633 } 1634 kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt; 1635 if (wpipe->pipe_state & PIPE_DIRECTW) 1636 kn->kn_data = 0; 1637 1638 PIPE_UNLOCK(rpipe); 1639 return (kn->kn_data >= PIPE_BUF); 1640} 1641