sys_pipe.c revision 173750
1108287Stjr/*- 2108287Stjr * Copyright (c) 1996 John S. Dyson 3108287Stjr * All rights reserved. 4108287Stjr * 5108287Stjr * Redistribution and use in source and binary forms, with or without 6108287Stjr * modification, are permitted provided that the following conditions 7108287Stjr * are met: 8108287Stjr * 1. Redistributions of source code must retain the above copyright 9108287Stjr * notice immediately at the beginning of the file, without modification, 10108287Stjr * this list of conditions, and the following disclaimer. 11108287Stjr * 2. Redistributions in binary form must reproduce the above copyright 12108287Stjr * notice, this list of conditions and the following disclaimer in the 13108287Stjr * documentation and/or other materials provided with the distribution. 14108287Stjr * 3. Absolutely no warranty of function or purpose is made by the author 15108287Stjr * John S. Dyson. 16108287Stjr * 4. Modifications may be freely made to this file if the above conditions 17108287Stjr * are met. 18108287Stjr */ 19108287Stjr 20108287Stjr/* 21108287Stjr * This file contains a high-performance replacement for the socket-based 22108287Stjr * pipes scheme originally used in FreeBSD/4.4Lite. It does not support 23108287Stjr * all features of sockets, but does do everything that pipes normally 24108287Stjr * do. 25108287Stjr */ 26108287Stjr 27108287Stjr/* 28108287Stjr * This code has two modes of operation, a small write mode and a large 29108287Stjr * write mode. The small write mode acts like conventional pipes with 30108287Stjr * a kernel buffer. If the buffer is less than PIPE_MINDIRECT, then the 31108287Stjr * "normal" pipe buffering is done. If the buffer is between PIPE_MINDIRECT 32108287Stjr * and PIPE_SIZE in size, it is fully mapped and wired into the kernel, and 33108287Stjr * the receiving process can copy it directly from the pages in the sending 34108287Stjr * process. 35108287Stjr * 36108287Stjr * If the sending process receives a signal, it is possible that it will 37108287Stjr * go away, and certainly its address space can change, because control 38108287Stjr * is returned back to the user-mode side. In that case, the pipe code 39108287Stjr * arranges to copy the buffer supplied by the user process, to a pageable 40108287Stjr * kernel buffer, and the receiving process will grab the data from the 41108287Stjr * pageable kernel buffer. Since signals don't happen all that often, 42108287Stjr * the copy operation is normally eliminated. 43108287Stjr * 44108287Stjr * The constant PIPE_MINDIRECT is chosen to make sure that buffering will 45108287Stjr * happen for small transfers so that the system will not spend all of 46108287Stjr * its time context switching. 47108287Stjr * 48108287Stjr * In order to limit the resource use of pipes, two sysctls exist: 49108287Stjr * 50108287Stjr * kern.ipc.maxpipekva - This is a hard limit on the amount of pageable 51108287Stjr * address space available to us in pipe_map. This value is normally 52108287Stjr * autotuned, but may also be loader tuned. 53108287Stjr * 54108287Stjr * kern.ipc.pipekva - This read-only sysctl tracks the current amount of 55108287Stjr * memory in use by pipes. 56108287Stjr * 57108287Stjr * Based on how large pipekva is relative to maxpipekva, the following 58108287Stjr * will happen: 59108287Stjr * 60108287Stjr * 0% - 50%: 61108287Stjr * New pipes are given 16K of memory backing, pipes may dynamically 62108287Stjr * grow to as large as 64K where needed. 63108287Stjr * 50% - 75%: 64108287Stjr * New pipes are given 4K (or PAGE_SIZE) of memory backing, 65108287Stjr * existing pipes may NOT grow. 66108287Stjr * 75% - 100%: 67108287Stjr * New pipes are given 4K (or PAGE_SIZE) of memory backing, 68108287Stjr * existing pipes will be shrunk down to 4K whenever possible. 69108287Stjr * 70108287Stjr * Resizing may be disabled by setting kern.ipc.piperesizeallowed=0. If 71108290Stjr * that is set, the only resize that will occur is the 0 -> SMALL_PIPE_SIZE 72108287Stjr * resize which MUST occur for reverse-direction pipes when they are 73108287Stjr * first used. 74108287Stjr * 75108287Stjr * 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 173750 2007-11-19 15:05:20Z dumbbell $"); 93 94#include "opt_mac.h" 95 96#include <sys/param.h> 97#include <sys/systm.h> 98#include <sys/fcntl.h> 99#include <sys/file.h> 100#include <sys/filedesc.h> 101#include <sys/filio.h> 102#include <sys/kernel.h> 103#include <sys/lock.h> 104#include <sys/mutex.h> 105#include <sys/ttycom.h> 106#include <sys/stat.h> 107#include <sys/malloc.h> 108#include <sys/poll.h> 109#include <sys/selinfo.h> 110#include <sys/signalvar.h> 111#include <sys/sysctl.h> 112#include <sys/sysproto.h> 113#include <sys/pipe.h> 114#include <sys/proc.h> 115#include <sys/vnode.h> 116#include <sys/uio.h> 117#include <sys/event.h> 118 119#include <security/mac/mac_framework.h> 120 121#include <vm/vm.h> 122#include <vm/vm_param.h> 123#include <vm/vm_object.h> 124#include <vm/vm_kern.h> 125#include <vm/vm_extern.h> 126#include <vm/pmap.h> 127#include <vm/vm_map.h> 128#include <vm/vm_page.h> 129#include <vm/uma.h> 130 131/* 132 * Use this define if you want to disable *fancy* VM things. Expect an 133 * approx 30% decrease in transfer rate. This could be useful for 134 * NetBSD or OpenBSD. 135 */ 136/* #define PIPE_NODIRECT */ 137 138/* 139 * interfaces to the outside world 140 */ 141static fo_rdwr_t pipe_read; 142static fo_rdwr_t pipe_write; 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_ioctl = pipe_ioctl, 153 .fo_poll = pipe_poll, 154 .fo_kqfilter = pipe_kqfilter, 155 .fo_stat = pipe_stat, 156 .fo_close = pipe_close, 157 .fo_flags = DFLAG_PASSABLE 158}; 159 160static void filt_pipedetach(struct knote *kn); 161static int filt_piperead(struct knote *kn, long hint); 162static int filt_pipewrite(struct knote *kn, long hint); 163 164static struct filterops pipe_rfiltops = 165 { 1, NULL, filt_pipedetach, filt_piperead }; 166static struct filterops pipe_wfiltops = 167 { 1, NULL, filt_pipedetach, filt_pipewrite }; 168 169/* 170 * Default pipe buffer size(s), this can be kind-of large now because pipe 171 * space is pageable. The pipe code will try to maintain locality of 172 * reference for performance reasons, so small amounts of outstanding I/O 173 * will not wipe the cache. 174 */ 175#define MINPIPESIZE (PIPE_SIZE/3) 176#define MAXPIPESIZE (2*PIPE_SIZE/3) 177 178static int amountpipekva; 179static int pipefragretry; 180static int pipeallocfail; 181static int piperesizefail; 182static int piperesizeallowed = 1; 183 184SYSCTL_INT(_kern_ipc, OID_AUTO, maxpipekva, CTLFLAG_RDTUN, 185 &maxpipekva, 0, "Pipe KVA limit"); 186SYSCTL_INT(_kern_ipc, OID_AUTO, pipekva, CTLFLAG_RD, 187 &amountpipekva, 0, "Pipe KVA usage"); 188SYSCTL_INT(_kern_ipc, OID_AUTO, pipefragretry, CTLFLAG_RD, 189 &pipefragretry, 0, "Pipe allocation retries due to fragmentation"); 190SYSCTL_INT(_kern_ipc, OID_AUTO, pipeallocfail, CTLFLAG_RD, 191 &pipeallocfail, 0, "Pipe allocation failures"); 192SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizefail, CTLFLAG_RD, 193 &piperesizefail, 0, "Pipe resize failures"); 194SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizeallowed, CTLFLAG_RW, 195 &piperesizeallowed, 0, "Pipe resizing allowed"); 196 197static void pipeinit(void *dummy __unused); 198static void pipeclose(struct pipe *cpipe); 199static void pipe_free_kmem(struct pipe *cpipe); 200static int pipe_create(struct pipe *pipe, int backing); 201static __inline int pipelock(struct pipe *cpipe, int catch); 202static __inline void pipeunlock(struct pipe *cpipe); 203static __inline void pipeselwakeup(struct pipe *cpipe); 204#ifndef PIPE_NODIRECT 205static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio); 206static void pipe_destroy_write_buffer(struct pipe *wpipe); 207static int pipe_direct_write(struct pipe *wpipe, struct uio *uio); 208static void pipe_clone_write_buffer(struct pipe *wpipe); 209#endif 210static int pipespace(struct pipe *cpipe, int size); 211static int pipespace_new(struct pipe *cpipe, int size); 212 213static int pipe_zone_ctor(void *mem, int size, void *arg, int flags); 214static int pipe_zone_init(void *mem, int size, int flags); 215static void pipe_zone_fini(void *mem, int size); 216 217static uma_zone_t pipe_zone; 218 219SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL); 220 221static void 222pipeinit(void *dummy __unused) 223{ 224 225 pipe_zone = uma_zcreate("pipe", sizeof(struct pipepair), 226 pipe_zone_ctor, NULL, pipe_zone_init, pipe_zone_fini, 227 UMA_ALIGN_PTR, 0); 228 KASSERT(pipe_zone != NULL, ("pipe_zone not initialized")); 229} 230 231static int 232pipe_zone_ctor(void *mem, int size, void *arg, int flags) 233{ 234 struct pipepair *pp; 235 struct pipe *rpipe, *wpipe; 236 237 KASSERT(size == sizeof(*pp), ("pipe_zone_ctor: wrong size")); 238 239 pp = (struct pipepair *)mem; 240 241 /* 242 * We zero both pipe endpoints to make sure all the kmem pointers 243 * are NULL, flag fields are zero'd, etc. We timestamp both 244 * endpoints with the same time. 245 */ 246 rpipe = &pp->pp_rpipe; 247 bzero(rpipe, sizeof(*rpipe)); 248 vfs_timestamp(&rpipe->pipe_ctime); 249 rpipe->pipe_atime = rpipe->pipe_mtime = rpipe->pipe_ctime; 250 251 wpipe = &pp->pp_wpipe; 252 bzero(wpipe, sizeof(*wpipe)); 253 wpipe->pipe_ctime = rpipe->pipe_ctime; 254 wpipe->pipe_atime = wpipe->pipe_mtime = rpipe->pipe_ctime; 255 256 rpipe->pipe_peer = wpipe; 257 rpipe->pipe_pair = pp; 258 wpipe->pipe_peer = rpipe; 259 wpipe->pipe_pair = pp; 260 261 /* 262 * Mark both endpoints as present; they will later get free'd 263 * one at a time. When both are free'd, then the whole pair 264 * is released. 265 */ 266 rpipe->pipe_present = 1; 267 wpipe->pipe_present = 1; 268 269 /* 270 * Eventually, the MAC Framework may initialize the label 271 * in ctor or init, but for now we do it elswhere to avoid 272 * blocking in ctor or init. 273 */ 274 pp->pp_label = NULL; 275 276 return (0); 277} 278 279static int 280pipe_zone_init(void *mem, int size, int flags) 281{ 282 struct pipepair *pp; 283 284 KASSERT(size == sizeof(*pp), ("pipe_zone_init: wrong size")); 285 286 pp = (struct pipepair *)mem; 287 288 mtx_init(&pp->pp_mtx, "pipe mutex", NULL, MTX_DEF | MTX_RECURSE); 289 return (0); 290} 291 292static void 293pipe_zone_fini(void *mem, int size) 294{ 295 struct pipepair *pp; 296 297 KASSERT(size == sizeof(*pp), ("pipe_zone_fini: wrong size")); 298 299 pp = (struct pipepair *)mem; 300 301 mtx_destroy(&pp->pp_mtx); 302} 303 304/* 305 * The pipe system call for the DTYPE_PIPE type of pipes. If we fail, let 306 * the zone pick up the pieces via pipeclose(). 307 */ 308/* ARGSUSED */ 309int 310pipe(td, uap) 311 struct thread *td; 312 struct pipe_args /* { 313 int dummy; 314 } */ *uap; 315{ 316 struct filedesc *fdp = td->td_proc->p_fd; 317 struct file *rf, *wf; 318 struct pipepair *pp; 319 struct pipe *rpipe, *wpipe; 320 int fd, error; 321 322 pp = uma_zalloc(pipe_zone, M_WAITOK); 323#ifdef MAC 324 /* 325 * The MAC label is shared between the connected endpoints. As a 326 * result mac_pipe_init() and mac_pipe_create() are called once 327 * for the pair, and not on the endpoints. 328 */ 329 mac_pipe_init(pp); 330 mac_pipe_create(td->td_ucred, pp); 331#endif 332 rpipe = &pp->pp_rpipe; 333 wpipe = &pp->pp_wpipe; 334 335 knlist_init(&rpipe->pipe_sel.si_note, PIPE_MTX(rpipe), NULL, NULL, 336 NULL); 337 knlist_init(&wpipe->pipe_sel.si_note, PIPE_MTX(wpipe), NULL, NULL, 338 NULL); 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 td->td_retval[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 FILE_LOCK(rf); 367 rf->f_flag = FREAD | FWRITE; 368 rf->f_type = DTYPE_PIPE; 369 rf->f_data = rpipe; 370 rf->f_ops = &pipeops; 371 FILE_UNLOCK(rf); 372 error = falloc(td, &wf, &fd); 373 if (error) { 374 fdclose(fdp, rf, td->td_retval[0], td); 375 fdrop(rf, td); 376 /* rpipe has been closed by fdrop(). */ 377 pipeclose(wpipe); 378 return (error); 379 } 380 /* An extra reference on `wf' has been held for us by falloc(). */ 381 FILE_LOCK(wf); 382 wf->f_flag = FREAD | FWRITE; 383 wf->f_type = DTYPE_PIPE; 384 wf->f_data = wpipe; 385 wf->f_ops = &pipeops; 386 FILE_UNLOCK(wf); 387 fdrop(wf, td); 388 td->td_retval[1] = fd; 389 fdrop(rf, td); 390 391 return (0); 392} 393 394/* 395 * Allocate kva for pipe circular buffer, the space is pageable 396 * This routine will 'realloc' the size of a pipe safely, if it fails 397 * it will retain the old buffer. 398 * If it fails it will return ENOMEM. 399 */ 400static int 401pipespace_new(cpipe, size) 402 struct pipe *cpipe; 403 int size; 404{ 405 caddr_t buffer; 406 int error, cnt, firstseg; 407 static int curfail = 0; 408 static struct timeval lastfail; 409 410 KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked")); 411 KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW), 412 ("pipespace: resize of direct writes not allowed")); 413retry: 414 cnt = cpipe->pipe_buffer.cnt; 415 if (cnt > size) 416 size = cnt; 417 418 size = round_page(size); 419 buffer = (caddr_t) vm_map_min(pipe_map); 420 421 error = vm_map_find(pipe_map, NULL, 0, 422 (vm_offset_t *) &buffer, size, 1, 423 VM_PROT_ALL, VM_PROT_ALL, 0); 424 if (error != KERN_SUCCESS) { 425 if ((cpipe->pipe_buffer.buffer == NULL) && 426 (size > SMALL_PIPE_SIZE)) { 427 size = SMALL_PIPE_SIZE; 428 pipefragretry++; 429 goto retry; 430 } 431 if (cpipe->pipe_buffer.buffer == NULL) { 432 pipeallocfail++; 433 if (ppsratecheck(&lastfail, &curfail, 1)) 434 printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n"); 435 } else { 436 piperesizefail++; 437 } 438 return (ENOMEM); 439 } 440 441 /* copy data, then free old resources if we're resizing */ 442 if (cnt > 0) { 443 if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) { 444 firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out; 445 bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out], 446 buffer, firstseg); 447 if ((cnt - firstseg) > 0) 448 bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg], 449 cpipe->pipe_buffer.in); 450 } else { 451 bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out], 452 buffer, cnt); 453 } 454 } 455 pipe_free_kmem(cpipe); 456 cpipe->pipe_buffer.buffer = buffer; 457 cpipe->pipe_buffer.size = size; 458 cpipe->pipe_buffer.in = cnt; 459 cpipe->pipe_buffer.out = 0; 460 cpipe->pipe_buffer.cnt = cnt; 461 atomic_add_int(&amountpipekva, cpipe->pipe_buffer.size); 462 return (0); 463} 464 465/* 466 * Wrapper for pipespace_new() that performs locking assertions. 467 */ 468static int 469pipespace(cpipe, size) 470 struct pipe *cpipe; 471 int size; 472{ 473 474 KASSERT(cpipe->pipe_state & PIPE_LOCKFL, 475 ("Unlocked pipe passed to pipespace")); 476 return (pipespace_new(cpipe, size)); 477} 478 479/* 480 * lock a pipe for I/O, blocking other access 481 */ 482static __inline int 483pipelock(cpipe, catch) 484 struct pipe *cpipe; 485 int catch; 486{ 487 int error; 488 489 PIPE_LOCK_ASSERT(cpipe, MA_OWNED); 490 while (cpipe->pipe_state & PIPE_LOCKFL) { 491 cpipe->pipe_state |= PIPE_LWANT; 492 error = msleep(cpipe, PIPE_MTX(cpipe), 493 catch ? (PRIBIO | PCATCH) : PRIBIO, 494 "pipelk", 0); 495 if (error != 0) 496 return (error); 497 } 498 cpipe->pipe_state |= PIPE_LOCKFL; 499 return (0); 500} 501 502/* 503 * unlock a pipe I/O lock 504 */ 505static __inline void 506pipeunlock(cpipe) 507 struct pipe *cpipe; 508{ 509 510 PIPE_LOCK_ASSERT(cpipe, MA_OWNED); 511 KASSERT(cpipe->pipe_state & PIPE_LOCKFL, 512 ("Unlocked pipe passed to pipeunlock")); 513 cpipe->pipe_state &= ~PIPE_LOCKFL; 514 if (cpipe->pipe_state & PIPE_LWANT) { 515 cpipe->pipe_state &= ~PIPE_LWANT; 516 wakeup(cpipe); 517 } 518} 519 520static __inline void 521pipeselwakeup(cpipe) 522 struct pipe *cpipe; 523{ 524 525 PIPE_LOCK_ASSERT(cpipe, MA_OWNED); 526 if (cpipe->pipe_state & PIPE_SEL) { 527 cpipe->pipe_state &= ~PIPE_SEL; 528 selwakeuppri(&cpipe->pipe_sel, PSOCK); 529 } 530 if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio) 531 pgsigio(&cpipe->pipe_sigio, SIGIO, 0); 532 KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0); 533} 534 535/* 536 * Initialize and allocate VM and memory for pipe. The structure 537 * will start out zero'd from the ctor, so we just manage the kmem. 538 */ 539static int 540pipe_create(pipe, backing) 541 struct pipe *pipe; 542 int backing; 543{ 544 int error; 545 546 if (backing) { 547 if (amountpipekva > maxpipekva / 2) 548 error = pipespace_new(pipe, SMALL_PIPE_SIZE); 549 else 550 error = pipespace_new(pipe, PIPE_SIZE); 551 } else { 552 /* If we're not backing this pipe, no need to do anything. */ 553 error = 0; 554 } 555 return (error); 556} 557 558/* ARGSUSED */ 559static int 560pipe_read(fp, uio, active_cred, flags, td) 561 struct file *fp; 562 struct uio *uio; 563 struct ucred *active_cred; 564 struct thread *td; 565 int flags; 566{ 567 struct pipe *rpipe = fp->f_data; 568 int error; 569 int nread = 0; 570 u_int size; 571 572 PIPE_LOCK(rpipe); 573 ++rpipe->pipe_busy; 574 error = pipelock(rpipe, 1); 575 if (error) 576 goto unlocked_error; 577 578#ifdef MAC 579 error = mac_pipe_check_read(active_cred, rpipe->pipe_pair); 580 if (error) 581 goto locked_error; 582#endif 583 if (amountpipekva > (3 * maxpipekva) / 4) { 584 if (!(rpipe->pipe_state & PIPE_DIRECTW) && 585 (rpipe->pipe_buffer.size > SMALL_PIPE_SIZE) && 586 (rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) && 587 (piperesizeallowed == 1)) { 588 PIPE_UNLOCK(rpipe); 589 pipespace(rpipe, SMALL_PIPE_SIZE); 590 PIPE_LOCK(rpipe); 591 } 592 } 593 594 while (uio->uio_resid) { 595 /* 596 * normal pipe buffer receive 597 */ 598 if (rpipe->pipe_buffer.cnt > 0) { 599 size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out; 600 if (size > rpipe->pipe_buffer.cnt) 601 size = rpipe->pipe_buffer.cnt; 602 if (size > (u_int) uio->uio_resid) 603 size = (u_int) uio->uio_resid; 604 605 PIPE_UNLOCK(rpipe); 606 error = uiomove( 607 &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out], 608 size, uio); 609 PIPE_LOCK(rpipe); 610 if (error) 611 break; 612 613 rpipe->pipe_buffer.out += size; 614 if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size) 615 rpipe->pipe_buffer.out = 0; 616 617 rpipe->pipe_buffer.cnt -= size; 618 619 /* 620 * If there is no more to read in the pipe, reset 621 * its pointers to the beginning. This improves 622 * cache hit stats. 623 */ 624 if (rpipe->pipe_buffer.cnt == 0) { 625 rpipe->pipe_buffer.in = 0; 626 rpipe->pipe_buffer.out = 0; 627 } 628 nread += size; 629#ifndef PIPE_NODIRECT 630 /* 631 * Direct copy, bypassing a kernel buffer. 632 */ 633 } else if ((size = rpipe->pipe_map.cnt) && 634 (rpipe->pipe_state & PIPE_DIRECTW)) { 635 if (size > (u_int) uio->uio_resid) 636 size = (u_int) uio->uio_resid; 637 638 PIPE_UNLOCK(rpipe); 639 error = uiomove_fromphys(rpipe->pipe_map.ms, 640 rpipe->pipe_map.pos, size, uio); 641 PIPE_LOCK(rpipe); 642 if (error) 643 break; 644 nread += size; 645 rpipe->pipe_map.pos += size; 646 rpipe->pipe_map.cnt -= size; 647 if (rpipe->pipe_map.cnt == 0) { 648 rpipe->pipe_state &= ~PIPE_DIRECTW; 649 wakeup(rpipe); 650 } 651#endif 652 } else { 653 /* 654 * detect EOF condition 655 * read returns 0 on EOF, no need to set error 656 */ 657 if (rpipe->pipe_state & PIPE_EOF) 658 break; 659 660 /* 661 * If the "write-side" has been blocked, wake it up now. 662 */ 663 if (rpipe->pipe_state & PIPE_WANTW) { 664 rpipe->pipe_state &= ~PIPE_WANTW; 665 wakeup(rpipe); 666 } 667 668 /* 669 * Break if some data was read. 670 */ 671 if (nread > 0) 672 break; 673 674 /* 675 * Unlock the pipe buffer for our remaining processing. 676 * We will either break out with an error or we will 677 * sleep and relock to loop. 678 */ 679 pipeunlock(rpipe); 680 681 /* 682 * Handle non-blocking mode operation or 683 * wait for more data. 684 */ 685 if (fp->f_flag & FNONBLOCK) { 686 error = EAGAIN; 687 } else { 688 rpipe->pipe_state |= PIPE_WANTR; 689 if ((error = msleep(rpipe, PIPE_MTX(rpipe), 690 PRIBIO | PCATCH, 691 "piperd", 0)) == 0) 692 error = pipelock(rpipe, 1); 693 } 694 if (error) 695 goto unlocked_error; 696 } 697 } 698#ifdef MAC 699locked_error: 700#endif 701 pipeunlock(rpipe); 702 703 /* XXX: should probably do this before getting any locks. */ 704 if (error == 0) 705 vfs_timestamp(&rpipe->pipe_atime); 706unlocked_error: 707 --rpipe->pipe_busy; 708 709 /* 710 * PIPE_WANT processing only makes sense if pipe_busy is 0. 711 */ 712 if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) { 713 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW); 714 wakeup(rpipe); 715 } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) { 716 /* 717 * Handle write blocking hysteresis. 718 */ 719 if (rpipe->pipe_state & PIPE_WANTW) { 720 rpipe->pipe_state &= ~PIPE_WANTW; 721 wakeup(rpipe); 722 } 723 } 724 725 if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF) 726 pipeselwakeup(rpipe); 727 728 PIPE_UNLOCK(rpipe); 729 return (error); 730} 731 732#ifndef PIPE_NODIRECT 733/* 734 * Map the sending processes' buffer into kernel space and wire it. 735 * This is similar to a physical write operation. 736 */ 737static int 738pipe_build_write_buffer(wpipe, uio) 739 struct pipe *wpipe; 740 struct uio *uio; 741{ 742 pmap_t pmap; 743 u_int size; 744 int i, j; 745 vm_offset_t addr, endaddr; 746 747 PIPE_LOCK_ASSERT(wpipe, MA_NOTOWNED); 748 KASSERT(wpipe->pipe_state & PIPE_DIRECTW, 749 ("Clone attempt on non-direct write pipe!")); 750 751 size = (u_int) uio->uio_iov->iov_len; 752 if (size > wpipe->pipe_buffer.size) 753 size = wpipe->pipe_buffer.size; 754 755 pmap = vmspace_pmap(curproc->p_vmspace); 756 endaddr = round_page((vm_offset_t)uio->uio_iov->iov_base + size); 757 addr = trunc_page((vm_offset_t)uio->uio_iov->iov_base); 758 for (i = 0; addr < endaddr; addr += PAGE_SIZE, i++) { 759 /* 760 * vm_fault_quick() can sleep. Consequently, 761 * vm_page_lock_queue() and vm_page_unlock_queue() 762 * should not be performed outside of this loop. 763 */ 764 race: 765 if (vm_fault_quick((caddr_t)addr, VM_PROT_READ) < 0) { 766 vm_page_lock_queues(); 767 for (j = 0; j < i; j++) 768 vm_page_unhold(wpipe->pipe_map.ms[j]); 769 vm_page_unlock_queues(); 770 return (EFAULT); 771 } 772 wpipe->pipe_map.ms[i] = pmap_extract_and_hold(pmap, addr, 773 VM_PROT_READ); 774 if (wpipe->pipe_map.ms[i] == NULL) 775 goto race; 776 } 777 778/* 779 * set up the control block 780 */ 781 wpipe->pipe_map.npages = i; 782 wpipe->pipe_map.pos = 783 ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK; 784 wpipe->pipe_map.cnt = size; 785 786/* 787 * and update the uio data 788 */ 789 790 uio->uio_iov->iov_len -= size; 791 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size; 792 if (uio->uio_iov->iov_len == 0) 793 uio->uio_iov++; 794 uio->uio_resid -= size; 795 uio->uio_offset += size; 796 return (0); 797} 798 799/* 800 * unmap and unwire the process buffer 801 */ 802static void 803pipe_destroy_write_buffer(wpipe) 804 struct pipe *wpipe; 805{ 806 int i; 807 808 PIPE_LOCK_ASSERT(wpipe, MA_OWNED); 809 vm_page_lock_queues(); 810 for (i = 0; i < wpipe->pipe_map.npages; i++) { 811 vm_page_unhold(wpipe->pipe_map.ms[i]); 812 } 813 vm_page_unlock_queues(); 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) || (wpipe->pipe_state & PIPE_EOF)) { 986 pipeunlock(wpipe); 987 PIPE_UNLOCK(rpipe); 988 return (EPIPE); 989 } 990#ifdef MAC 991 error = mac_pipe_check_write(active_cred, wpipe->pipe_pair); 992 if (error) { 993 pipeunlock(wpipe); 994 PIPE_UNLOCK(rpipe); 995 return (error); 996 } 997#endif 998 ++wpipe->pipe_busy; 999 1000 /* Choose a larger size if it's advantageous */ 1001 desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size); 1002 while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) { 1003 if (piperesizeallowed != 1) 1004 break; 1005 if (amountpipekva > maxpipekva / 2) 1006 break; 1007 if (desiredsize == BIG_PIPE_SIZE) 1008 break; 1009 desiredsize = desiredsize * 2; 1010 } 1011 1012 /* Choose a smaller size if we're in a OOM situation */ 1013 if ((amountpipekva > (3 * maxpipekva) / 4) && 1014 (wpipe->pipe_buffer.size > SMALL_PIPE_SIZE) && 1015 (wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) && 1016 (piperesizeallowed == 1)) 1017 desiredsize = SMALL_PIPE_SIZE; 1018 1019 /* Resize if the above determined that a new size was necessary */ 1020 if ((desiredsize != wpipe->pipe_buffer.size) && 1021 ((wpipe->pipe_state & PIPE_DIRECTW) == 0)) { 1022 PIPE_UNLOCK(wpipe); 1023 pipespace(wpipe, desiredsize); 1024 PIPE_LOCK(wpipe); 1025 } 1026 if (wpipe->pipe_buffer.size == 0) { 1027 /* 1028 * This can only happen for reverse direction use of pipes 1029 * in a complete OOM situation. 1030 */ 1031 error = ENOMEM; 1032 --wpipe->pipe_busy; 1033 pipeunlock(wpipe); 1034 PIPE_UNLOCK(wpipe); 1035 return (error); 1036 } 1037 1038 pipeunlock(wpipe); 1039 1040 orig_resid = uio->uio_resid; 1041 1042 while (uio->uio_resid) { 1043 int space; 1044 1045 pipelock(wpipe, 0); 1046 if (wpipe->pipe_state & PIPE_EOF) { 1047 pipeunlock(wpipe); 1048 error = EPIPE; 1049 break; 1050 } 1051#ifndef PIPE_NODIRECT 1052 /* 1053 * If the transfer is large, we can gain performance if 1054 * we do process-to-process copies directly. 1055 * If the write is non-blocking, we don't use the 1056 * direct write mechanism. 1057 * 1058 * The direct write mechanism will detect the reader going 1059 * away on us. 1060 */ 1061 if (uio->uio_segflg == UIO_USERSPACE && 1062 uio->uio_iov->iov_len >= PIPE_MINDIRECT && 1063 wpipe->pipe_buffer.size >= PIPE_MINDIRECT && 1064 (fp->f_flag & FNONBLOCK) == 0) { 1065 pipeunlock(wpipe); 1066 error = pipe_direct_write(wpipe, uio); 1067 if (error) 1068 break; 1069 continue; 1070 } 1071#endif 1072 1073 /* 1074 * Pipe buffered writes cannot be coincidental with 1075 * direct writes. We wait until the currently executing 1076 * direct write is completed before we start filling the 1077 * pipe buffer. We break out if a signal occurs or the 1078 * reader goes away. 1079 */ 1080 if (wpipe->pipe_state & PIPE_DIRECTW) { 1081 if (wpipe->pipe_state & PIPE_WANTR) { 1082 wpipe->pipe_state &= ~PIPE_WANTR; 1083 wakeup(wpipe); 1084 } 1085 pipeselwakeup(wpipe); 1086 wpipe->pipe_state |= PIPE_WANTW; 1087 pipeunlock(wpipe); 1088 error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH, 1089 "pipbww", 0); 1090 if (error) 1091 break; 1092 else 1093 continue; 1094 } 1095 1096 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt; 1097 1098 /* Writes of size <= PIPE_BUF must be atomic. */ 1099 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF)) 1100 space = 0; 1101 1102 if (space > 0) { 1103 int size; /* Transfer size */ 1104 int segsize; /* first segment to transfer */ 1105 1106 /* 1107 * Transfer size is minimum of uio transfer 1108 * and free space in pipe buffer. 1109 */ 1110 if (space > uio->uio_resid) 1111 size = uio->uio_resid; 1112 else 1113 size = space; 1114 /* 1115 * First segment to transfer is minimum of 1116 * transfer size and contiguous space in 1117 * pipe buffer. If first segment to transfer 1118 * is less than the transfer size, we've got 1119 * a wraparound in the buffer. 1120 */ 1121 segsize = wpipe->pipe_buffer.size - 1122 wpipe->pipe_buffer.in; 1123 if (segsize > size) 1124 segsize = size; 1125 1126 /* Transfer first segment */ 1127 1128 PIPE_UNLOCK(rpipe); 1129 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in], 1130 segsize, uio); 1131 PIPE_LOCK(rpipe); 1132 1133 if (error == 0 && segsize < size) { 1134 KASSERT(wpipe->pipe_buffer.in + segsize == 1135 wpipe->pipe_buffer.size, 1136 ("Pipe buffer wraparound disappeared")); 1137 /* 1138 * Transfer remaining part now, to 1139 * support atomic writes. Wraparound 1140 * happened. 1141 */ 1142 1143 PIPE_UNLOCK(rpipe); 1144 error = uiomove( 1145 &wpipe->pipe_buffer.buffer[0], 1146 size - segsize, uio); 1147 PIPE_LOCK(rpipe); 1148 } 1149 if (error == 0) { 1150 wpipe->pipe_buffer.in += size; 1151 if (wpipe->pipe_buffer.in >= 1152 wpipe->pipe_buffer.size) { 1153 KASSERT(wpipe->pipe_buffer.in == 1154 size - segsize + 1155 wpipe->pipe_buffer.size, 1156 ("Expected wraparound bad")); 1157 wpipe->pipe_buffer.in = size - segsize; 1158 } 1159 1160 wpipe->pipe_buffer.cnt += size; 1161 KASSERT(wpipe->pipe_buffer.cnt <= 1162 wpipe->pipe_buffer.size, 1163 ("Pipe buffer overflow")); 1164 } 1165 pipeunlock(wpipe); 1166 if (error != 0) 1167 break; 1168 } else { 1169 /* 1170 * If the "read-side" has been blocked, wake it up now. 1171 */ 1172 if (wpipe->pipe_state & PIPE_WANTR) { 1173 wpipe->pipe_state &= ~PIPE_WANTR; 1174 wakeup(wpipe); 1175 } 1176 1177 /* 1178 * don't block on non-blocking I/O 1179 */ 1180 if (fp->f_flag & FNONBLOCK) { 1181 error = EAGAIN; 1182 pipeunlock(wpipe); 1183 break; 1184 } 1185 1186 /* 1187 * We have no more space and have something to offer, 1188 * wake up select/poll. 1189 */ 1190 pipeselwakeup(wpipe); 1191 1192 wpipe->pipe_state |= PIPE_WANTW; 1193 pipeunlock(wpipe); 1194 error = msleep(wpipe, PIPE_MTX(rpipe), 1195 PRIBIO | PCATCH, "pipewr", 0); 1196 if (error != 0) 1197 break; 1198 } 1199 } 1200 1201 pipelock(wpipe, 0); 1202 --wpipe->pipe_busy; 1203 1204 if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) { 1205 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR); 1206 wakeup(wpipe); 1207 } else if (wpipe->pipe_buffer.cnt > 0) { 1208 /* 1209 * If we have put any characters in the buffer, we wake up 1210 * the reader. 1211 */ 1212 if (wpipe->pipe_state & PIPE_WANTR) { 1213 wpipe->pipe_state &= ~PIPE_WANTR; 1214 wakeup(wpipe); 1215 } 1216 } 1217 1218 /* 1219 * Don't return EPIPE if I/O was successful 1220 */ 1221 if ((wpipe->pipe_buffer.cnt == 0) && 1222 (uio->uio_resid == 0) && 1223 (error == EPIPE)) { 1224 error = 0; 1225 } 1226 1227 if (error == 0) 1228 vfs_timestamp(&wpipe->pipe_mtime); 1229 1230 /* 1231 * We have something to offer, 1232 * wake up select/poll. 1233 */ 1234 if (wpipe->pipe_buffer.cnt) 1235 pipeselwakeup(wpipe); 1236 1237 pipeunlock(wpipe); 1238 PIPE_UNLOCK(rpipe); 1239 return (error); 1240} 1241 1242/* 1243 * we implement a very minimal set of ioctls for compatibility with sockets. 1244 */ 1245static int 1246pipe_ioctl(fp, cmd, data, active_cred, td) 1247 struct file *fp; 1248 u_long cmd; 1249 void *data; 1250 struct ucred *active_cred; 1251 struct thread *td; 1252{ 1253 struct pipe *mpipe = fp->f_data; 1254 int error; 1255 1256 PIPE_LOCK(mpipe); 1257 1258#ifdef MAC 1259 error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data); 1260 if (error) { 1261 PIPE_UNLOCK(mpipe); 1262 return (error); 1263 } 1264#endif 1265 1266 error = 0; 1267 switch (cmd) { 1268 1269 case FIONBIO: 1270 break; 1271 1272 case FIOASYNC: 1273 if (*(int *)data) { 1274 mpipe->pipe_state |= PIPE_ASYNC; 1275 } else { 1276 mpipe->pipe_state &= ~PIPE_ASYNC; 1277 } 1278 break; 1279 1280 case FIONREAD: 1281 if (mpipe->pipe_state & PIPE_DIRECTW) 1282 *(int *)data = mpipe->pipe_map.cnt; 1283 else 1284 *(int *)data = mpipe->pipe_buffer.cnt; 1285 break; 1286 1287 case FIOSETOWN: 1288 PIPE_UNLOCK(mpipe); 1289 error = fsetown(*(int *)data, &mpipe->pipe_sigio); 1290 goto out_unlocked; 1291 1292 case FIOGETOWN: 1293 *(int *)data = fgetown(&mpipe->pipe_sigio); 1294 break; 1295 1296 /* This is deprecated, FIOSETOWN should be used instead. */ 1297 case TIOCSPGRP: 1298 PIPE_UNLOCK(mpipe); 1299 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio); 1300 goto out_unlocked; 1301 1302 /* This is deprecated, FIOGETOWN should be used instead. */ 1303 case TIOCGPGRP: 1304 *(int *)data = -fgetown(&mpipe->pipe_sigio); 1305 break; 1306 1307 default: 1308 error = ENOTTY; 1309 break; 1310 } 1311 PIPE_UNLOCK(mpipe); 1312out_unlocked: 1313 return (error); 1314} 1315 1316static int 1317pipe_poll(fp, events, active_cred, td) 1318 struct file *fp; 1319 int events; 1320 struct ucred *active_cred; 1321 struct thread *td; 1322{ 1323 struct pipe *rpipe = fp->f_data; 1324 struct pipe *wpipe; 1325 int revents = 0; 1326#ifdef MAC 1327 int error; 1328#endif 1329 1330 wpipe = rpipe->pipe_peer; 1331 PIPE_LOCK(rpipe); 1332#ifdef MAC 1333 error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair); 1334 if (error) 1335 goto locked_error; 1336#endif 1337 if (events & (POLLIN | POLLRDNORM)) 1338 if ((rpipe->pipe_state & PIPE_DIRECTW) || 1339 (rpipe->pipe_buffer.cnt > 0) || 1340 (rpipe->pipe_state & PIPE_EOF)) 1341 revents |= events & (POLLIN | POLLRDNORM); 1342 1343 if (events & (POLLOUT | POLLWRNORM)) 1344 if (!wpipe->pipe_present || (wpipe->pipe_state & PIPE_EOF) || 1345 (((wpipe->pipe_state & PIPE_DIRECTW) == 0) && 1346 (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF)) 1347 revents |= events & (POLLOUT | POLLWRNORM); 1348 1349 if ((rpipe->pipe_state & PIPE_EOF) || 1350 (!wpipe->pipe_present) || 1351 (wpipe->pipe_state & PIPE_EOF)) 1352 revents |= POLLHUP; 1353 1354 if (revents == 0) { 1355 if (events & (POLLIN | POLLRDNORM)) { 1356 selrecord(td, &rpipe->pipe_sel); 1357 rpipe->pipe_state |= PIPE_SEL; 1358 } 1359 1360 if (events & (POLLOUT | POLLWRNORM)) { 1361 selrecord(td, &wpipe->pipe_sel); 1362 wpipe->pipe_state |= PIPE_SEL; 1363 } 1364 } 1365#ifdef MAC 1366locked_error: 1367#endif 1368 PIPE_UNLOCK(rpipe); 1369 1370 return (revents); 1371} 1372 1373/* 1374 * We shouldn't need locks here as we're doing a read and this should 1375 * be a natural race. 1376 */ 1377static int 1378pipe_stat(fp, ub, active_cred, td) 1379 struct file *fp; 1380 struct stat *ub; 1381 struct ucred *active_cred; 1382 struct thread *td; 1383{ 1384 struct pipe *pipe = fp->f_data; 1385#ifdef MAC 1386 int error; 1387 1388 PIPE_LOCK(pipe); 1389 error = mac_pipe_check_stat(active_cred, pipe->pipe_pair); 1390 PIPE_UNLOCK(pipe); 1391 if (error) 1392 return (error); 1393#endif 1394 bzero(ub, sizeof(*ub)); 1395 ub->st_mode = S_IFIFO; 1396 ub->st_blksize = PAGE_SIZE; 1397 if (pipe->pipe_state & PIPE_DIRECTW) 1398 ub->st_size = pipe->pipe_map.cnt; 1399 else 1400 ub->st_size = pipe->pipe_buffer.cnt; 1401 ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize; 1402 ub->st_atimespec = pipe->pipe_atime; 1403 ub->st_mtimespec = pipe->pipe_mtime; 1404 ub->st_ctimespec = pipe->pipe_ctime; 1405 ub->st_uid = fp->f_cred->cr_uid; 1406 ub->st_gid = fp->f_cred->cr_gid; 1407 /* 1408 * Left as 0: st_dev, st_ino, st_nlink, st_rdev, st_flags, st_gen. 1409 * XXX (st_dev, st_ino) should be unique. 1410 */ 1411 return (0); 1412} 1413 1414/* ARGSUSED */ 1415static int 1416pipe_close(fp, td) 1417 struct file *fp; 1418 struct thread *td; 1419{ 1420 struct pipe *cpipe = fp->f_data; 1421 1422 fp->f_ops = &badfileops; 1423 fp->f_data = NULL; 1424 funsetown(&cpipe->pipe_sigio); 1425 pipeclose(cpipe); 1426 return (0); 1427} 1428 1429static void 1430pipe_free_kmem(cpipe) 1431 struct pipe *cpipe; 1432{ 1433 1434 KASSERT(!mtx_owned(PIPE_MTX(cpipe)), 1435 ("pipe_free_kmem: pipe mutex locked")); 1436 1437 if (cpipe->pipe_buffer.buffer != NULL) { 1438 atomic_subtract_int(&amountpipekva, cpipe->pipe_buffer.size); 1439 vm_map_remove(pipe_map, 1440 (vm_offset_t)cpipe->pipe_buffer.buffer, 1441 (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size); 1442 cpipe->pipe_buffer.buffer = NULL; 1443 } 1444#ifndef PIPE_NODIRECT 1445 { 1446 cpipe->pipe_map.cnt = 0; 1447 cpipe->pipe_map.pos = 0; 1448 cpipe->pipe_map.npages = 0; 1449 } 1450#endif 1451} 1452 1453/* 1454 * shutdown the pipe 1455 */ 1456static void 1457pipeclose(cpipe) 1458 struct pipe *cpipe; 1459{ 1460 struct pipepair *pp; 1461 struct pipe *ppipe; 1462 1463 KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL")); 1464 1465 PIPE_LOCK(cpipe); 1466 pipelock(cpipe, 0); 1467 pp = cpipe->pipe_pair; 1468 1469 pipeselwakeup(cpipe); 1470 1471 /* 1472 * If the other side is blocked, wake it up saying that 1473 * we want to close it down. 1474 */ 1475 cpipe->pipe_state |= PIPE_EOF; 1476 while (cpipe->pipe_busy) { 1477 wakeup(cpipe); 1478 cpipe->pipe_state |= PIPE_WANT; 1479 pipeunlock(cpipe); 1480 msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0); 1481 pipelock(cpipe, 0); 1482 } 1483 1484 1485 /* 1486 * Disconnect from peer, if any. 1487 */ 1488 ppipe = cpipe->pipe_peer; 1489 if (ppipe->pipe_present != 0) { 1490 pipeselwakeup(ppipe); 1491 1492 ppipe->pipe_state |= PIPE_EOF; 1493 wakeup(ppipe); 1494 KNOTE_LOCKED(&ppipe->pipe_sel.si_note, 0); 1495 } 1496 1497 /* 1498 * Mark this endpoint as free. Release kmem resources. We 1499 * don't mark this endpoint as unused until we've finished 1500 * doing that, or the pipe might disappear out from under 1501 * us. 1502 */ 1503 PIPE_UNLOCK(cpipe); 1504 pipe_free_kmem(cpipe); 1505 PIPE_LOCK(cpipe); 1506 cpipe->pipe_present = 0; 1507 pipeunlock(cpipe); 1508 knlist_clear(&cpipe->pipe_sel.si_note, 1); 1509 knlist_destroy(&cpipe->pipe_sel.si_note); 1510 1511 /* 1512 * If both endpoints are now closed, release the memory for the 1513 * pipe pair. If not, unlock. 1514 */ 1515 if (ppipe->pipe_present == 0) { 1516 PIPE_UNLOCK(cpipe); 1517#ifdef MAC 1518 mac_pipe_destroy(pp); 1519#endif 1520 uma_zfree(pipe_zone, cpipe->pipe_pair); 1521 } else 1522 PIPE_UNLOCK(cpipe); 1523} 1524 1525/*ARGSUSED*/ 1526static int 1527pipe_kqfilter(struct file *fp, struct knote *kn) 1528{ 1529 struct pipe *cpipe; 1530 1531 cpipe = kn->kn_fp->f_data; 1532 PIPE_LOCK(cpipe); 1533 switch (kn->kn_filter) { 1534 case EVFILT_READ: 1535 kn->kn_fop = &pipe_rfiltops; 1536 break; 1537 case EVFILT_WRITE: 1538 kn->kn_fop = &pipe_wfiltops; 1539 if (!cpipe->pipe_peer->pipe_present) { 1540 /* other end of pipe has been closed */ 1541 PIPE_UNLOCK(cpipe); 1542 return (EPIPE); 1543 } 1544 cpipe = cpipe->pipe_peer; 1545 break; 1546 default: 1547 PIPE_UNLOCK(cpipe); 1548 return (EINVAL); 1549 } 1550 1551 knlist_add(&cpipe->pipe_sel.si_note, kn, 1); 1552 PIPE_UNLOCK(cpipe); 1553 return (0); 1554} 1555 1556static void 1557filt_pipedetach(struct knote *kn) 1558{ 1559 struct pipe *cpipe = (struct pipe *)kn->kn_fp->f_data; 1560 1561 PIPE_LOCK(cpipe); 1562 if (kn->kn_filter == EVFILT_WRITE) { 1563 if (!cpipe->pipe_peer->pipe_present) { 1564 PIPE_UNLOCK(cpipe); 1565 return; 1566 } 1567 cpipe = cpipe->pipe_peer; 1568 } 1569 knlist_remove(&cpipe->pipe_sel.si_note, kn, 1); 1570 PIPE_UNLOCK(cpipe); 1571} 1572 1573/*ARGSUSED*/ 1574static int 1575filt_piperead(struct knote *kn, long hint) 1576{ 1577 struct pipe *rpipe = kn->kn_fp->f_data; 1578 struct pipe *wpipe = rpipe->pipe_peer; 1579 int ret; 1580 1581 PIPE_LOCK(rpipe); 1582 kn->kn_data = rpipe->pipe_buffer.cnt; 1583 if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW)) 1584 kn->kn_data = rpipe->pipe_map.cnt; 1585 1586 if ((rpipe->pipe_state & PIPE_EOF) || 1587 (!wpipe->pipe_present) || (wpipe->pipe_state & PIPE_EOF)) { 1588 kn->kn_flags |= EV_EOF; 1589 PIPE_UNLOCK(rpipe); 1590 return (1); 1591 } 1592 ret = kn->kn_data > 0; 1593 PIPE_UNLOCK(rpipe); 1594 return ret; 1595} 1596 1597/*ARGSUSED*/ 1598static int 1599filt_pipewrite(struct knote *kn, long hint) 1600{ 1601 struct pipe *rpipe = kn->kn_fp->f_data; 1602 struct pipe *wpipe = rpipe->pipe_peer; 1603 1604 PIPE_LOCK(rpipe); 1605 if ((!wpipe->pipe_present) || (wpipe->pipe_state & PIPE_EOF)) { 1606 kn->kn_data = 0; 1607 kn->kn_flags |= EV_EOF; 1608 PIPE_UNLOCK(rpipe); 1609 return (1); 1610 } 1611 kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt; 1612 if (wpipe->pipe_state & PIPE_DIRECTW) 1613 kn->kn_data = 0; 1614 1615 PIPE_UNLOCK(rpipe); 1616 return (kn->kn_data >= PIPE_BUF); 1617} 1618