1 /*- 2 * Copyright (c) 1996 John S. Dyson 3 * Copyright (c) 2012 Giovanni Trematerra 4 * All rights reserved. 5 * 6 * Redistribution and use in source and binary forms, with or without 7 * modification, are permitted provided that the following conditions 8 * are met: 9 * 1. Redistributions of source code must retain the above copyright 10 * notice immediately at the beginning of the file, without modification, 11 * this list of conditions, and the following disclaimer. 12 * 2. Redistributions in binary form must reproduce the above copyright 13 * notice, this list of conditions and the following disclaimer in the 14 * documentation and/or other materials provided with the distribution. 15 * 3. Absolutely no warranty of function or purpose is made by the author 16 * John S. Dyson. 17 * 4. Modifications may be freely made to this file if the above conditions 18 * are met. 19 */ 20 21 /* 22 * This file contains a high-performance replacement for the socket-based 23 * pipes scheme originally used in FreeBSD/4.4Lite. It does not support 24 * all features of sockets, but does do everything that pipes normally 25 * do. 26 */ 27 28 /* 29 * This code has two modes of operation, a small write mode and a large 30 * write mode. The small write mode acts like conventional pipes with 31 * a kernel buffer. If the buffer is less than PIPE_MINDIRECT, then the 32 * "normal" pipe buffering is done. If the buffer is between PIPE_MINDIRECT 33 * and PIPE_SIZE in size, the sending process pins the underlying pages in 34 * memory, and the receiving process copies directly from these pinned pages 35 * in the sending process. 36 * 37 * If the sending process receives a signal, it is possible that it will 38 * go away, and certainly its address space can change, because control 39 * is returned back to the user-mode side. In that case, the pipe code 40 * arranges to copy the buffer supplied by the user process, to a pageable 41 * kernel buffer, and the receiving process will grab the data from the 42 * pageable kernel buffer. Since signals don't happen all that often, 43 * the copy operation is normally eliminated. 44 * 45 * The constant PIPE_MINDIRECT is chosen to make sure that buffering will 46 * happen for small transfers so that the system will not spend all of 47 * its time context switching. 48 * 49 * In order to limit the resource use of pipes, two sysctls exist: 50 * 51 * kern.ipc.maxpipekva - This is a hard limit on the amount of pageable 52 * address space available to us in pipe_map. This value is normally 53 * autotuned, but may also be loader tuned. 54 * 55 * kern.ipc.pipekva - This read-only sysctl tracks the current amount of 56 * memory in use by pipes. 57 * 58 * Based on how large pipekva is relative to maxpipekva, the following 59 * will happen: 60 * 61 * 0% - 50%: 62 * New pipes are given 16K of memory backing, pipes may dynamically 63 * grow to as large as 64K where needed. 64 * 50% - 75%: 65 * New pipes are given 4K (or PAGE_SIZE) of memory backing, 66 * existing pipes may NOT grow. 67 * 75% - 100%: 68 * New pipes are given 4K (or PAGE_SIZE) of memory backing, 69 * existing pipes will be shrunk down to 4K whenever possible. 70 * 71 * Resizing may be disabled by setting kern.ipc.piperesizeallowed=0. If 72 * that is set, the only resize that will occur is the 0 -> SMALL_PIPE_SIZE 73 * resize which MUST occur for reverse-direction pipes when they are 74 * first used. 75 * 76 * Additional information about the current state of pipes may be obtained 77 * from kern.ipc.pipes, kern.ipc.pipefragretry, kern.ipc.pipeallocfail, 78 * and kern.ipc.piperesizefail. 79 * 80 * Locking rules: There are two locks present here: A mutex, used via 81 * PIPE_LOCK, and a flag, used via pipelock(). All locking is done via 82 * the flag, as mutexes can not persist over uiomove. The mutex 83 * exists only to guard access to the flag, and is not in itself a 84 * locking mechanism. Also note that there is only a single mutex for 85 * both directions of a pipe. 86 * 87 * As pipelock() may have to sleep before it can acquire the flag, it 88 * is important to reread all data after a call to pipelock(); everything 89 * in the structure may have changed. 90 */ 91 92 #include <sys/cdefs.h> 93 __FBSDID("$FreeBSD$"); 94 95 #include <sys/param.h> 96 #include <sys/systm.h> 97 #include <sys/conf.h> 98 #include <sys/fcntl.h> 99 #include <sys/file.h> 100 #include <sys/filedesc.h> 101 #include <sys/filio.h> 102 #include <sys/kernel.h> 103 #include <sys/lock.h> 104 #include <sys/mutex.h> 105 #include <sys/ttycom.h> 106 #include <sys/stat.h> 107 #include <sys/malloc.h> 108 #include <sys/poll.h> 109 #include <sys/selinfo.h> 110 #include <sys/signalvar.h> 111 #include <sys/syscallsubr.h> 112 #include <sys/sysctl.h> 113 #include <sys/sysproto.h> 114 #include <sys/pipe.h> 115 #include <sys/proc.h> 116 #include <sys/vnode.h> 117 #include <sys/uio.h> 118 #include <sys/event.h> 119 120 #include <security/mac/mac_framework.h> 121 122 #include <vm/vm.h> 123 #include <vm/vm_param.h> 124 #include <vm/vm_object.h> 125 #include <vm/vm_kern.h> 126 #include <vm/vm_extern.h> 127 #include <vm/pmap.h> 128 #include <vm/vm_map.h> 129 #include <vm/vm_page.h> 130 #include <vm/uma.h> 131 132 /* XXX */ 133 int do_pipe(struct thread *td, int fildes[2], int flags); 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_state & PIPE_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; 158 159 struct fileops pipeops = { 160 .fo_read = pipe_read, 161 .fo_write = pipe_write, 162 .fo_truncate = pipe_truncate, 163 .fo_ioctl = pipe_ioctl, 164 .fo_poll = pipe_poll, 165 .fo_kqfilter = pipe_kqfilter, 166 .fo_stat = pipe_stat, 167 .fo_close = pipe_close, 168 .fo_chmod = pipe_chmod, 169 .fo_chown = pipe_chown, 170 .fo_flags = DFLAG_PASSABLE 171 }; 172 173 static void filt_pipedetach(struct knote *kn); 174 static void filt_pipedetach_notsup(struct knote *kn); 175 static int filt_pipenotsup(struct knote *kn, long hint); 176 static int filt_piperead(struct knote *kn, long hint); 177 static int filt_pipewrite(struct knote *kn, long hint); 178 179 static struct filterops pipe_nfiltops = { 180 .f_isfd = 1, 181 .f_detach = filt_pipedetach_notsup, 182 .f_event = filt_pipenotsup 183 }; 184 static struct filterops pipe_rfiltops = { 185 .f_isfd = 1, 186 .f_detach = filt_pipedetach, 187 .f_event = filt_piperead 188 }; 189 static struct filterops pipe_wfiltops = { 190 .f_isfd = 1, 191 .f_detach = filt_pipedetach, 192 .f_event = filt_pipewrite 193 }; 194 195 /* 196 * Default pipe buffer size(s), this can be kind-of large now because pipe 197 * space is pageable. The pipe code will try to maintain locality of 198 * reference for performance reasons, so small amounts of outstanding I/O 199 * will not wipe the cache. 200 */ 201 #define MINPIPESIZE (PIPE_SIZE/3) 202 #define MAXPIPESIZE (2*PIPE_SIZE/3) 203 204 static long amountpipekva; 205 static int pipefragretry; 206 static int pipeallocfail; 207 static int piperesizefail; 208 static int piperesizeallowed = 1; 209 210 SYSCTL_LONG(_kern_ipc, OID_AUTO, maxpipekva, CTLFLAG_RDTUN, 211 &maxpipekva, 0, "Pipe KVA limit"); 212 SYSCTL_LONG(_kern_ipc, OID_AUTO, pipekva, CTLFLAG_RD, 213 &amountpipekva, 0, "Pipe KVA usage"); 214 SYSCTL_INT(_kern_ipc, OID_AUTO, pipefragretry, CTLFLAG_RD, 215 &pipefragretry, 0, "Pipe allocation retries due to fragmentation"); 216 SYSCTL_INT(_kern_ipc, OID_AUTO, pipeallocfail, CTLFLAG_RD, 217 &pipeallocfail, 0, "Pipe allocation failures"); 218 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizefail, CTLFLAG_RD, 219 &piperesizefail, 0, "Pipe resize failures"); 220 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizeallowed, CTLFLAG_RW, 221 &piperesizeallowed, 0, "Pipe resizing allowed"); 222 223 static void pipeinit(void *dummy __unused); 224 static void pipeclose(struct pipe *cpipe); 225 static void pipe_free_kmem(struct pipe *cpipe); 226 static int pipe_create(struct pipe *pipe, int backing); 227 static int pipe_paircreate(struct thread *td, struct pipepair **p_pp); 228 static __inline int pipelock(struct pipe *cpipe, int catch); 229 static __inline void pipeunlock(struct pipe *cpipe); 230 static __inline void pipeselwakeup(struct pipe *cpipe); 231 #ifndef PIPE_NODIRECT 232 static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio); 233 static void pipe_destroy_write_buffer(struct pipe *wpipe); 234 static int pipe_direct_write(struct pipe *wpipe, struct uio *uio); 235 static void pipe_clone_write_buffer(struct pipe *wpipe); 236 #endif 237 static int pipespace(struct pipe *cpipe, int size); 238 static int pipespace_new(struct pipe *cpipe, int size); 239 240 static int pipe_zone_ctor(void *mem, int size, void *arg, int flags); 241 static int pipe_zone_init(void *mem, int size, int flags); 242 static void pipe_zone_fini(void *mem, int size); 243 244 static uma_zone_t pipe_zone; 245 static struct unrhdr *pipeino_unr; 246 static dev_t pipedev_ino; 247 248 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL); 249 250 static void 251 pipeinit(void *dummy __unused) 252 { 253 254 pipe_zone = uma_zcreate("pipe", sizeof(struct pipepair), 255 pipe_zone_ctor, NULL, pipe_zone_init, pipe_zone_fini, 256 UMA_ALIGN_PTR, 0); 257 KASSERT(pipe_zone != NULL, ("pipe_zone not initialized")); 258 pipeino_unr = new_unrhdr(1, INT32_MAX, NULL); 259 KASSERT(pipeino_unr != NULL, ("pipe fake inodes not initialized")); 260 pipedev_ino = devfs_alloc_cdp_inode(); 261 KASSERT(pipedev_ino > 0, ("pipe dev inode not initialized")); 262 } 263 264 static int 265 pipe_zone_ctor(void *mem, int size, void *arg, int flags) 266 { 267 struct pipepair *pp; 268 struct pipe *rpipe, *wpipe; 269 270 KASSERT(size == sizeof(*pp), ("pipe_zone_ctor: wrong size")); 271 272 pp = (struct pipepair *)mem; 273 274 /* 275 * We zero both pipe endpoints to make sure all the kmem pointers 276 * are NULL, flag fields are zero'd, etc. We timestamp both 277 * endpoints with the same time. 278 */ 279 rpipe = &pp->pp_rpipe; 280 bzero(rpipe, sizeof(*rpipe)); 281 vfs_timestamp(&rpipe->pipe_ctime); 282 rpipe->pipe_atime = rpipe->pipe_mtime = rpipe->pipe_ctime; 283 284 wpipe = &pp->pp_wpipe; 285 bzero(wpipe, sizeof(*wpipe)); 286 wpipe->pipe_ctime = rpipe->pipe_ctime; 287 wpipe->pipe_atime = wpipe->pipe_mtime = rpipe->pipe_ctime; 288 289 rpipe->pipe_peer = wpipe; 290 rpipe->pipe_pair = pp; 291 wpipe->pipe_peer = rpipe; 292 wpipe->pipe_pair = pp; 293 294 /* 295 * Mark both endpoints as present; they will later get free'd 296 * one at a time. When both are free'd, then the whole pair 297 * is released. 298 */ 299 rpipe->pipe_present = PIPE_ACTIVE; 300 wpipe->pipe_present = PIPE_ACTIVE; 301 302 /* 303 * Eventually, the MAC Framework may initialize the label 304 * in ctor or init, but for now we do it elswhere to avoid 305 * blocking in ctor or init. 306 */ 307 pp->pp_label = NULL; 308 309 return (0); 310 } 311 312 static int 313 pipe_zone_init(void *mem, int size, int flags) 314 { 315 struct pipepair *pp; 316 317 KASSERT(size == sizeof(*pp), ("pipe_zone_init: wrong size")); 318 319 pp = (struct pipepair *)mem; 320 321 mtx_init(&pp->pp_mtx, "pipe mutex", NULL, MTX_DEF | MTX_RECURSE); 322 return (0); 323 } 324 325 static void 326 pipe_zone_fini(void *mem, int size) 327 { 328 struct pipepair *pp; 329 330 KASSERT(size == sizeof(*pp), ("pipe_zone_fini: wrong size")); 331 332 pp = (struct pipepair *)mem; 333 334 mtx_destroy(&pp->pp_mtx); 335 } 336 337 static int 338 pipe_paircreate(struct thread *td, struct pipepair **p_pp) 339 { 340 struct pipepair *pp; 341 struct pipe *rpipe, *wpipe; 342 int error; 343 344 *p_pp = pp = uma_zalloc(pipe_zone, M_WAITOK); 345 #ifdef MAC 346 /* 347 * The MAC label is shared between the connected endpoints. As a 348 * result mac_pipe_init() and mac_pipe_create() are called once 349 * for the pair, and not on the endpoints. 350 */ 351 mac_pipe_init(pp); 352 mac_pipe_create(td->td_ucred, pp); 353 #endif 354 rpipe = &pp->pp_rpipe; 355 wpipe = &pp->pp_wpipe; 356 357 knlist_init_mtx(&rpipe->pipe_sel.si_note, PIPE_MTX(rpipe)); 358 knlist_init_mtx(&wpipe->pipe_sel.si_note, PIPE_MTX(wpipe)); 359 360 /* Only the forward direction pipe is backed by default */ 361 if ((error = pipe_create(rpipe, 1)) != 0 || 362 (error = pipe_create(wpipe, 0)) != 0) { 363 pipeclose(rpipe); 364 pipeclose(wpipe); 365 return (error); 366 } 367 368 rpipe->pipe_state |= PIPE_DIRECTOK; 369 wpipe->pipe_state |= PIPE_DIRECTOK; 370 return (0); 371 } 372 373 int 374 pipe_named_ctor(struct pipe **ppipe, struct thread *td) 375 { 376 struct pipepair *pp; 377 int error; 378 379 error = pipe_paircreate(td, &pp); 380 if (error != 0) 381 return (error); 382 pp->pp_rpipe.pipe_state |= PIPE_NAMED; 383 *ppipe = &pp->pp_rpipe; 384 return (0); 385 } 386 387 void 388 pipe_dtor(struct pipe *dpipe) 389 { 390 ino_t ino; 391 392 ino = dpipe->pipe_ino; 393 funsetown(&dpipe->pipe_sigio); 394 pipeclose(dpipe); 395 if (dpipe->pipe_state & PIPE_NAMED) { 396 dpipe = dpipe->pipe_peer; 397 funsetown(&dpipe->pipe_sigio); 398 pipeclose(dpipe); 399 } 400 if (ino != 0 && ino != (ino_t)-1) 401 free_unr(pipeino_unr, ino); 402 } 403 404 /* 405 * The pipe system call for the DTYPE_PIPE type of pipes. If we fail, let 406 * the zone pick up the pieces via pipeclose(). 407 */ 408 int 409 kern_pipe(struct thread *td, int fildes[2]) 410 { 411 412 return (do_pipe(td, fildes, 0)); 413 } 414 415 int 416 do_pipe(struct thread *td, int fildes[2], int flags) 417 { 418 struct filedesc *fdp; 419 struct file *rf, *wf; 420 struct pipe *rpipe, *wpipe; 421 struct pipepair *pp; 422 int fd, fflags, error; 423 424 fdp = td->td_proc->p_fd; 425 error = pipe_paircreate(td, &pp); 426 if (error != 0) 427 return (error); 428 rpipe = &pp->pp_rpipe; 429 wpipe = &pp->pp_wpipe; 430 error = falloc(td, &rf, &fd, flags); 431 if (error) { 432 pipeclose(rpipe); 433 pipeclose(wpipe); 434 return (error); 435 } 436 /* An extra reference on `rf' has been held for us by falloc(). */ 437 fildes[0] = fd; 438 439 fflags = FREAD | FWRITE; 440 if ((flags & O_NONBLOCK) != 0) 441 fflags |= FNONBLOCK; 442 443 /* 444 * Warning: once we've gotten past allocation of the fd for the 445 * read-side, we can only drop the read side via fdrop() in order 446 * to avoid races against processes which manage to dup() the read 447 * side while we are blocked trying to allocate the write side. 448 */ 449 finit(rf, fflags, DTYPE_PIPE, rpipe, &pipeops); 450 error = falloc(td, &wf, &fd, flags); 451 if (error) { 452 fdclose(fdp, rf, fildes[0], td); 453 fdrop(rf, td); 454 /* rpipe has been closed by fdrop(). */ 455 pipeclose(wpipe); 456 return (error); 457 } 458 /* An extra reference on `wf' has been held for us by falloc(). */ 459 finit(wf, fflags, DTYPE_PIPE, wpipe, &pipeops); 460 fdrop(wf, td); 461 fildes[1] = fd; 462 fdrop(rf, td); 463 464 return (0); 465 } 466 467 /* ARGSUSED */ 468 int 469 sys_pipe(struct thread *td, struct pipe_args *uap) 470 { 471 int error; 472 int fildes[2]; 473 474 error = kern_pipe(td, fildes); 475 if (error) 476 return (error); 477 478 td->td_retval[0] = fildes[0]; 479 td->td_retval[1] = fildes[1]; 480 481 return (0); 482 } 483 484 /* 485 * Allocate kva for pipe circular buffer, the space is pageable 486 * This routine will 'realloc' the size of a pipe safely, if it fails 487 * it will retain the old buffer. 488 * If it fails it will return ENOMEM. 489 */ 490 static int 491 pipespace_new(cpipe, size) 492 struct pipe *cpipe; 493 int size; 494 { 495 caddr_t buffer; 496 int error, cnt, firstseg; 497 static int curfail = 0; 498 static struct timeval lastfail; 499 500 KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked")); 501 KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW), 502 ("pipespace: resize of direct writes not allowed")); 503 retry: 504 cnt = cpipe->pipe_buffer.cnt; 505 if (cnt > size) 506 size = cnt; 507 508 size = round_page(size); 509 buffer = (caddr_t) vm_map_min(pipe_map); 510 511 error = vm_map_find(pipe_map, NULL, 0, 512 (vm_offset_t *) &buffer, size, 1, 513 VM_PROT_ALL, VM_PROT_ALL, 0); 514 if (error != KERN_SUCCESS) { 515 if ((cpipe->pipe_buffer.buffer == NULL) && 516 (size > SMALL_PIPE_SIZE)) { 517 size = SMALL_PIPE_SIZE; 518 pipefragretry++; 519 goto retry; 520 } 521 if (cpipe->pipe_buffer.buffer == NULL) { 522 pipeallocfail++; 523 if (ppsratecheck(&lastfail, &curfail, 1)) 524 printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n"); 525 } else { 526 piperesizefail++; 527 } 528 return (ENOMEM); 529 } 530 531 /* copy data, then free old resources if we're resizing */ 532 if (cnt > 0) { 533 if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) { 534 firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out; 535 bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out], 536 buffer, firstseg); 537 if ((cnt - firstseg) > 0) 538 bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg], 539 cpipe->pipe_buffer.in); 540 } else { 541 bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out], 542 buffer, cnt); 543 } 544 } 545 pipe_free_kmem(cpipe); 546 cpipe->pipe_buffer.buffer = buffer; 547 cpipe->pipe_buffer.size = size; 548 cpipe->pipe_buffer.in = cnt; 549 cpipe->pipe_buffer.out = 0; 550 cpipe->pipe_buffer.cnt = cnt; 551 atomic_add_long(&amountpipekva, cpipe->pipe_buffer.size); 552 return (0); 553 } 554 555 /* 556 * Wrapper for pipespace_new() that performs locking assertions. 557 */ 558 static int 559 pipespace(cpipe, size) 560 struct pipe *cpipe; 561 int size; 562 { 563 564 KASSERT(cpipe->pipe_state & PIPE_LOCKFL, 565 ("Unlocked pipe passed to pipespace")); 566 return (pipespace_new(cpipe, size)); 567 } 568 569 /* 570 * lock a pipe for I/O, blocking other access 571 */ 572 static __inline int 573 pipelock(cpipe, catch) 574 struct pipe *cpipe; 575 int catch; 576 { 577 int error; 578 579 PIPE_LOCK_ASSERT(cpipe, MA_OWNED); 580 while (cpipe->pipe_state & PIPE_LOCKFL) { 581 cpipe->pipe_state |= PIPE_LWANT; 582 error = msleep(cpipe, PIPE_MTX(cpipe), 583 catch ? (PRIBIO | PCATCH) : PRIBIO, 584 "pipelk", 0); 585 if (error != 0) 586 return (error); 587 } 588 cpipe->pipe_state |= PIPE_LOCKFL; 589 return (0); 590 } 591 592 /* 593 * unlock a pipe I/O lock 594 */ 595 static __inline void 596 pipeunlock(cpipe) 597 struct pipe *cpipe; 598 { 599 600 PIPE_LOCK_ASSERT(cpipe, MA_OWNED); 601 KASSERT(cpipe->pipe_state & PIPE_LOCKFL, 602 ("Unlocked pipe passed to pipeunlock")); 603 cpipe->pipe_state &= ~PIPE_LOCKFL; 604 if (cpipe->pipe_state & PIPE_LWANT) { 605 cpipe->pipe_state &= ~PIPE_LWANT; 606 wakeup(cpipe); 607 } 608 } 609 610 static __inline void 611 pipeselwakeup(cpipe) 612 struct pipe *cpipe; 613 { 614 615 PIPE_LOCK_ASSERT(cpipe, MA_OWNED); 616 if (cpipe->pipe_state & PIPE_SEL) { 617 selwakeuppri(&cpipe->pipe_sel, PSOCK); 618 if (!SEL_WAITING(&cpipe->pipe_sel)) 619 cpipe->pipe_state &= ~PIPE_SEL; 620 } 621 if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio) 622 pgsigio(&cpipe->pipe_sigio, SIGIO, 0); 623 KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0); 624 } 625 626 /* 627 * Initialize and allocate VM and memory for pipe. The structure 628 * will start out zero'd from the ctor, so we just manage the kmem. 629 */ 630 static int 631 pipe_create(pipe, backing) 632 struct pipe *pipe; 633 int backing; 634 { 635 int error; 636 637 if (backing) { 638 if (amountpipekva > maxpipekva / 2) 639 error = pipespace_new(pipe, SMALL_PIPE_SIZE); 640 else 641 error = pipespace_new(pipe, PIPE_SIZE); 642 } else { 643 /* If we're not backing this pipe, no need to do anything. */ 644 error = 0; 645 } 646 pipe->pipe_ino = -1; 647 return (error); 648 } 649 650 /* ARGSUSED */ 651 static int 652 pipe_read(fp, uio, active_cred, flags, td) 653 struct file *fp; 654 struct uio *uio; 655 struct ucred *active_cred; 656 struct thread *td; 657 int flags; 658 { 659 struct pipe *rpipe; 660 int error; 661 int nread = 0; 662 int size; 663 664 rpipe = fp->f_data; 665 PIPE_LOCK(rpipe); 666 ++rpipe->pipe_busy; 667 error = pipelock(rpipe, 1); 668 if (error) 669 goto unlocked_error; 670 671 #ifdef MAC 672 error = mac_pipe_check_read(active_cred, rpipe->pipe_pair); 673 if (error) 674 goto locked_error; 675 #endif 676 if (amountpipekva > (3 * maxpipekva) / 4) { 677 if (!(rpipe->pipe_state & PIPE_DIRECTW) && 678 (rpipe->pipe_buffer.size > SMALL_PIPE_SIZE) && 679 (rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) && 680 (piperesizeallowed == 1)) { 681 PIPE_UNLOCK(rpipe); 682 pipespace(rpipe, SMALL_PIPE_SIZE); 683 PIPE_LOCK(rpipe); 684 } 685 } 686 687 while (uio->uio_resid) { 688 /* 689 * normal pipe buffer receive 690 */ 691 if (rpipe->pipe_buffer.cnt > 0) { 692 size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out; 693 if (size > rpipe->pipe_buffer.cnt) 694 size = rpipe->pipe_buffer.cnt; 695 if (size > uio->uio_resid) 696 size = uio->uio_resid; 697 698 PIPE_UNLOCK(rpipe); 699 error = uiomove( 700 &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out], 701 size, uio); 702 PIPE_LOCK(rpipe); 703 if (error) 704 break; 705 706 rpipe->pipe_buffer.out += size; 707 if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size) 708 rpipe->pipe_buffer.out = 0; 709 710 rpipe->pipe_buffer.cnt -= size; 711 712 /* 713 * If there is no more to read in the pipe, reset 714 * its pointers to the beginning. This improves 715 * cache hit stats. 716 */ 717 if (rpipe->pipe_buffer.cnt == 0) { 718 rpipe->pipe_buffer.in = 0; 719 rpipe->pipe_buffer.out = 0; 720 } 721 nread += size; 722 #ifndef PIPE_NODIRECT 723 /* 724 * Direct copy, bypassing a kernel buffer. 725 */ 726 } else if ((size = rpipe->pipe_map.cnt) && 727 (rpipe->pipe_state & PIPE_DIRECTW)) { 728 if (size > uio->uio_resid) 729 size = (u_int) uio->uio_resid; 730 731 PIPE_UNLOCK(rpipe); 732 error = uiomove_fromphys(rpipe->pipe_map.ms, 733 rpipe->pipe_map.pos, size, uio); 734 PIPE_LOCK(rpipe); 735 if (error) 736 break; 737 nread += size; 738 rpipe->pipe_map.pos += size; 739 rpipe->pipe_map.cnt -= size; 740 if (rpipe->pipe_map.cnt == 0) { 741 rpipe->pipe_state &= ~PIPE_DIRECTW; 742 wakeup(rpipe); 743 } 744 #endif 745 } else { 746 /* 747 * detect EOF condition 748 * read returns 0 on EOF, no need to set error 749 */ 750 if (rpipe->pipe_state & PIPE_EOF) 751 break; 752 753 /* 754 * If the "write-side" has been blocked, wake it up now. 755 */ 756 if (rpipe->pipe_state & PIPE_WANTW) { 757 rpipe->pipe_state &= ~PIPE_WANTW; 758 wakeup(rpipe); 759 } 760 761 /* 762 * Break if some data was read. 763 */ 764 if (nread > 0) 765 break; 766 767 /* 768 * Unlock the pipe buffer for our remaining processing. 769 * We will either break out with an error or we will 770 * sleep and relock to loop. 771 */ 772 pipeunlock(rpipe); 773 774 /* 775 * Handle non-blocking mode operation or 776 * wait for more data. 777 */ 778 if (fp->f_flag & FNONBLOCK) { 779 error = EAGAIN; 780 } else { 781 rpipe->pipe_state |= PIPE_WANTR; 782 if ((error = msleep(rpipe, PIPE_MTX(rpipe), 783 PRIBIO | PCATCH, 784 "piperd", 0)) == 0) 785 error = pipelock(rpipe, 1); 786 } 787 if (error) 788 goto unlocked_error; 789 } 790 } 791 #ifdef MAC 792 locked_error: 793 #endif 794 pipeunlock(rpipe); 795 796 /* XXX: should probably do this before getting any locks. */ 797 if (error == 0) 798 vfs_timestamp(&rpipe->pipe_atime); 799 unlocked_error: 800 --rpipe->pipe_busy; 801 802 /* 803 * PIPE_WANT processing only makes sense if pipe_busy is 0. 804 */ 805 if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) { 806 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW); 807 wakeup(rpipe); 808 } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) { 809 /* 810 * Handle write blocking hysteresis. 811 */ 812 if (rpipe->pipe_state & PIPE_WANTW) { 813 rpipe->pipe_state &= ~PIPE_WANTW; 814 wakeup(rpipe); 815 } 816 } 817 818 if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF) 819 pipeselwakeup(rpipe); 820 821 PIPE_UNLOCK(rpipe); 822 return (error); 823 } 824 825 #ifndef PIPE_NODIRECT 826 /* 827 * Map the sending processes' buffer into kernel space and wire it. 828 * This is similar to a physical write operation. 829 */ 830 static int 831 pipe_build_write_buffer(wpipe, uio) 832 struct pipe *wpipe; 833 struct uio *uio; 834 { 835 u_int size; 836 int i; 837 838 PIPE_LOCK_ASSERT(wpipe, MA_NOTOWNED); 839 KASSERT(wpipe->pipe_state & PIPE_DIRECTW, 840 ("Clone attempt on non-direct write pipe!")); 841 842 if (uio->uio_iov->iov_len > wpipe->pipe_buffer.size) 843 size = wpipe->pipe_buffer.size; 844 else 845 size = uio->uio_iov->iov_len; 846 847 if ((i = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map, 848 (vm_offset_t)uio->uio_iov->iov_base, size, VM_PROT_READ, 849 wpipe->pipe_map.ms, PIPENPAGES)) < 0) 850 return (EFAULT); 851 852 /* 853 * set up the control block 854 */ 855 wpipe->pipe_map.npages = i; 856 wpipe->pipe_map.pos = 857 ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK; 858 wpipe->pipe_map.cnt = size; 859 860 /* 861 * and update the uio data 862 */ 863 864 uio->uio_iov->iov_len -= size; 865 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size; 866 if (uio->uio_iov->iov_len == 0) 867 uio->uio_iov++; 868 uio->uio_resid -= size; 869 uio->uio_offset += size; 870 return (0); 871 } 872 873 /* 874 * unmap and unwire the process buffer 875 */ 876 static void 877 pipe_destroy_write_buffer(wpipe) 878 struct pipe *wpipe; 879 { 880 881 PIPE_LOCK_ASSERT(wpipe, MA_OWNED); 882 vm_page_unhold_pages(wpipe->pipe_map.ms, wpipe->pipe_map.npages); 883 wpipe->pipe_map.npages = 0; 884 } 885 886 /* 887 * In the case of a signal, the writing process might go away. This 888 * code copies the data into the circular buffer so that the source 889 * pages can be freed without loss of data. 890 */ 891 static void 892 pipe_clone_write_buffer(wpipe) 893 struct pipe *wpipe; 894 { 895 struct uio uio; 896 struct iovec iov; 897 int size; 898 int pos; 899 900 PIPE_LOCK_ASSERT(wpipe, MA_OWNED); 901 size = wpipe->pipe_map.cnt; 902 pos = wpipe->pipe_map.pos; 903 904 wpipe->pipe_buffer.in = size; 905 wpipe->pipe_buffer.out = 0; 906 wpipe->pipe_buffer.cnt = size; 907 wpipe->pipe_state &= ~PIPE_DIRECTW; 908 909 PIPE_UNLOCK(wpipe); 910 iov.iov_base = wpipe->pipe_buffer.buffer; 911 iov.iov_len = size; 912 uio.uio_iov = &iov; 913 uio.uio_iovcnt = 1; 914 uio.uio_offset = 0; 915 uio.uio_resid = size; 916 uio.uio_segflg = UIO_SYSSPACE; 917 uio.uio_rw = UIO_READ; 918 uio.uio_td = curthread; 919 uiomove_fromphys(wpipe->pipe_map.ms, pos, size, &uio); 920 PIPE_LOCK(wpipe); 921 pipe_destroy_write_buffer(wpipe); 922 } 923 924 /* 925 * This implements the pipe buffer write mechanism. Note that only 926 * a direct write OR a normal pipe write can be pending at any given time. 927 * If there are any characters in the pipe buffer, the direct write will 928 * be deferred until the receiving process grabs all of the bytes from 929 * the pipe buffer. Then the direct mapping write is set-up. 930 */ 931 static int 932 pipe_direct_write(wpipe, uio) 933 struct pipe *wpipe; 934 struct uio *uio; 935 { 936 int error; 937 938 retry: 939 PIPE_LOCK_ASSERT(wpipe, MA_OWNED); 940 error = pipelock(wpipe, 1); 941 if (wpipe->pipe_state & PIPE_EOF) 942 error = EPIPE; 943 if (error) { 944 pipeunlock(wpipe); 945 goto error1; 946 } 947 while (wpipe->pipe_state & PIPE_DIRECTW) { 948 if (wpipe->pipe_state & PIPE_WANTR) { 949 wpipe->pipe_state &= ~PIPE_WANTR; 950 wakeup(wpipe); 951 } 952 pipeselwakeup(wpipe); 953 wpipe->pipe_state |= PIPE_WANTW; 954 pipeunlock(wpipe); 955 error = msleep(wpipe, PIPE_MTX(wpipe), 956 PRIBIO | PCATCH, "pipdww", 0); 957 if (error) 958 goto error1; 959 else 960 goto retry; 961 } 962 wpipe->pipe_map.cnt = 0; /* transfer not ready yet */ 963 if (wpipe->pipe_buffer.cnt > 0) { 964 if (wpipe->pipe_state & PIPE_WANTR) { 965 wpipe->pipe_state &= ~PIPE_WANTR; 966 wakeup(wpipe); 967 } 968 pipeselwakeup(wpipe); 969 wpipe->pipe_state |= PIPE_WANTW; 970 pipeunlock(wpipe); 971 error = msleep(wpipe, PIPE_MTX(wpipe), 972 PRIBIO | PCATCH, "pipdwc", 0); 973 if (error) 974 goto error1; 975 else 976 goto retry; 977 } 978 979 wpipe->pipe_state |= PIPE_DIRECTW; 980 981 PIPE_UNLOCK(wpipe); 982 error = pipe_build_write_buffer(wpipe, uio); 983 PIPE_LOCK(wpipe); 984 if (error) { 985 wpipe->pipe_state &= ~PIPE_DIRECTW; 986 pipeunlock(wpipe); 987 goto error1; 988 } 989 990 error = 0; 991 while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) { 992 if (wpipe->pipe_state & PIPE_EOF) { 993 pipe_destroy_write_buffer(wpipe); 994 pipeselwakeup(wpipe); 995 pipeunlock(wpipe); 996 error = EPIPE; 997 goto error1; 998 } 999 if (wpipe->pipe_state & PIPE_WANTR) { 1000 wpipe->pipe_state &= ~PIPE_WANTR; 1001 wakeup(wpipe); 1002 } 1003 pipeselwakeup(wpipe); 1004 pipeunlock(wpipe); 1005 error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH, 1006 "pipdwt", 0); 1007 pipelock(wpipe, 0); 1008 } 1009 1010 if (wpipe->pipe_state & PIPE_EOF) 1011 error = EPIPE; 1012 if (wpipe->pipe_state & PIPE_DIRECTW) { 1013 /* 1014 * this bit of trickery substitutes a kernel buffer for 1015 * the process that might be going away. 1016 */ 1017 pipe_clone_write_buffer(wpipe); 1018 } else { 1019 pipe_destroy_write_buffer(wpipe); 1020 } 1021 pipeunlock(wpipe); 1022 return (error); 1023 1024 error1: 1025 wakeup(wpipe); 1026 return (error); 1027 } 1028 #endif 1029 1030 static int 1031 pipe_write(fp, uio, active_cred, flags, td) 1032 struct file *fp; 1033 struct uio *uio; 1034 struct ucred *active_cred; 1035 struct thread *td; 1036 int flags; 1037 { 1038 int error = 0; 1039 int desiredsize; 1040 ssize_t orig_resid; 1041 struct pipe *wpipe, *rpipe; 1042 1043 rpipe = fp->f_data; 1044 wpipe = PIPE_PEER(rpipe); 1045 PIPE_LOCK(rpipe); 1046 error = pipelock(wpipe, 1); 1047 if (error) { 1048 PIPE_UNLOCK(rpipe); 1049 return (error); 1050 } 1051 /* 1052 * detect loss of pipe read side, issue SIGPIPE if lost. 1053 */ 1054 if (wpipe->pipe_present != PIPE_ACTIVE || 1055 (wpipe->pipe_state & PIPE_EOF)) { 1056 pipeunlock(wpipe); 1057 PIPE_UNLOCK(rpipe); 1058 return (EPIPE); 1059 } 1060 #ifdef MAC 1061 error = mac_pipe_check_write(active_cred, wpipe->pipe_pair); 1062 if (error) { 1063 pipeunlock(wpipe); 1064 PIPE_UNLOCK(rpipe); 1065 return (error); 1066 } 1067 #endif 1068 ++wpipe->pipe_busy; 1069 1070 /* Choose a larger size if it's advantageous */ 1071 desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size); 1072 while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) { 1073 if (piperesizeallowed != 1) 1074 break; 1075 if (amountpipekva > maxpipekva / 2) 1076 break; 1077 if (desiredsize == BIG_PIPE_SIZE) 1078 break; 1079 desiredsize = desiredsize * 2; 1080 } 1081 1082 /* Choose a smaller size if we're in a OOM situation */ 1083 if ((amountpipekva > (3 * maxpipekva) / 4) && 1084 (wpipe->pipe_buffer.size > SMALL_PIPE_SIZE) && 1085 (wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) && 1086 (piperesizeallowed == 1)) 1087 desiredsize = SMALL_PIPE_SIZE; 1088 1089 /* Resize if the above determined that a new size was necessary */ 1090 if ((desiredsize != wpipe->pipe_buffer.size) && 1091 ((wpipe->pipe_state & PIPE_DIRECTW) == 0)) { 1092 PIPE_UNLOCK(wpipe); 1093 pipespace(wpipe, desiredsize); 1094 PIPE_LOCK(wpipe); 1095 } 1096 if (wpipe->pipe_buffer.size == 0) { 1097 /* 1098 * This can only happen for reverse direction use of pipes 1099 * in a complete OOM situation. 1100 */ 1101 error = ENOMEM; 1102 --wpipe->pipe_busy; 1103 pipeunlock(wpipe); 1104 PIPE_UNLOCK(wpipe); 1105 return (error); 1106 } 1107 1108 pipeunlock(wpipe); 1109 1110 orig_resid = uio->uio_resid; 1111 1112 while (uio->uio_resid) { 1113 int space; 1114 1115 pipelock(wpipe, 0); 1116 if (wpipe->pipe_state & PIPE_EOF) { 1117 pipeunlock(wpipe); 1118 error = EPIPE; 1119 break; 1120 } 1121 #ifndef PIPE_NODIRECT 1122 /* 1123 * If the transfer is large, we can gain performance if 1124 * we do process-to-process copies directly. 1125 * If the write is non-blocking, we don't use the 1126 * direct write mechanism. 1127 * 1128 * The direct write mechanism will detect the reader going 1129 * away on us. 1130 */ 1131 if (uio->uio_segflg == UIO_USERSPACE && 1132 uio->uio_iov->iov_len >= PIPE_MINDIRECT && 1133 wpipe->pipe_buffer.size >= PIPE_MINDIRECT && 1134 (fp->f_flag & FNONBLOCK) == 0) { 1135 pipeunlock(wpipe); 1136 error = pipe_direct_write(wpipe, uio); 1137 if (error) 1138 break; 1139 continue; 1140 } 1141 #endif 1142 1143 /* 1144 * Pipe buffered writes cannot be coincidental with 1145 * direct writes. We wait until the currently executing 1146 * direct write is completed before we start filling the 1147 * pipe buffer. We break out if a signal occurs or the 1148 * reader goes away. 1149 */ 1150 if (wpipe->pipe_state & PIPE_DIRECTW) { 1151 if (wpipe->pipe_state & PIPE_WANTR) { 1152 wpipe->pipe_state &= ~PIPE_WANTR; 1153 wakeup(wpipe); 1154 } 1155 pipeselwakeup(wpipe); 1156 wpipe->pipe_state |= PIPE_WANTW; 1157 pipeunlock(wpipe); 1158 error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH, 1159 "pipbww", 0); 1160 if (error) 1161 break; 1162 else 1163 continue; 1164 } 1165 1166 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt; 1167 1168 /* Writes of size <= PIPE_BUF must be atomic. */ 1169 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF)) 1170 space = 0; 1171 1172 if (space > 0) { 1173 int size; /* Transfer size */ 1174 int segsize; /* first segment to transfer */ 1175 1176 /* 1177 * Transfer size is minimum of uio transfer 1178 * and free space in pipe buffer. 1179 */ 1180 if (space > uio->uio_resid) 1181 size = uio->uio_resid; 1182 else 1183 size = space; 1184 /* 1185 * First segment to transfer is minimum of 1186 * transfer size and contiguous space in 1187 * pipe buffer. If first segment to transfer 1188 * is less than the transfer size, we've got 1189 * a wraparound in the buffer. 1190 */ 1191 segsize = wpipe->pipe_buffer.size - 1192 wpipe->pipe_buffer.in; 1193 if (segsize > size) 1194 segsize = size; 1195 1196 /* Transfer first segment */ 1197 1198 PIPE_UNLOCK(rpipe); 1199 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in], 1200 segsize, uio); 1201 PIPE_LOCK(rpipe); 1202 1203 if (error == 0 && segsize < size) { 1204 KASSERT(wpipe->pipe_buffer.in + segsize == 1205 wpipe->pipe_buffer.size, 1206 ("Pipe buffer wraparound disappeared")); 1207 /* 1208 * Transfer remaining part now, to 1209 * support atomic writes. Wraparound 1210 * happened. 1211 */ 1212 1213 PIPE_UNLOCK(rpipe); 1214 error = uiomove( 1215 &wpipe->pipe_buffer.buffer[0], 1216 size - segsize, uio); 1217 PIPE_LOCK(rpipe); 1218 } 1219 if (error == 0) { 1220 wpipe->pipe_buffer.in += size; 1221 if (wpipe->pipe_buffer.in >= 1222 wpipe->pipe_buffer.size) { 1223 KASSERT(wpipe->pipe_buffer.in == 1224 size - segsize + 1225 wpipe->pipe_buffer.size, 1226 ("Expected wraparound bad")); 1227 wpipe->pipe_buffer.in = size - segsize; 1228 } 1229 1230 wpipe->pipe_buffer.cnt += size; 1231 KASSERT(wpipe->pipe_buffer.cnt <= 1232 wpipe->pipe_buffer.size, 1233 ("Pipe buffer overflow")); 1234 } 1235 pipeunlock(wpipe); 1236 if (error != 0) 1237 break; 1238 } else { 1239 /* 1240 * If the "read-side" has been blocked, wake it up now. 1241 */ 1242 if (wpipe->pipe_state & PIPE_WANTR) { 1243 wpipe->pipe_state &= ~PIPE_WANTR; 1244 wakeup(wpipe); 1245 } 1246 1247 /* 1248 * don't block on non-blocking I/O 1249 */ 1250 if (fp->f_flag & FNONBLOCK) { 1251 error = EAGAIN; 1252 pipeunlock(wpipe); 1253 break; 1254 } 1255 1256 /* 1257 * We have no more space and have something to offer, 1258 * wake up select/poll. 1259 */ 1260 pipeselwakeup(wpipe); 1261 1262 wpipe->pipe_state |= PIPE_WANTW; 1263 pipeunlock(wpipe); 1264 error = msleep(wpipe, PIPE_MTX(rpipe), 1265 PRIBIO | PCATCH, "pipewr", 0); 1266 if (error != 0) 1267 break; 1268 } 1269 } 1270 1271 pipelock(wpipe, 0); 1272 --wpipe->pipe_busy; 1273 1274 if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) { 1275 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR); 1276 wakeup(wpipe); 1277 } else if (wpipe->pipe_buffer.cnt > 0) { 1278 /* 1279 * If we have put any characters in the buffer, we wake up 1280 * the reader. 1281 */ 1282 if (wpipe->pipe_state & PIPE_WANTR) { 1283 wpipe->pipe_state &= ~PIPE_WANTR; 1284 wakeup(wpipe); 1285 } 1286 } 1287 1288 /* 1289 * Don't return EPIPE if I/O was successful 1290 */ 1291 if ((wpipe->pipe_buffer.cnt == 0) && 1292 (uio->uio_resid == 0) && 1293 (error == EPIPE)) { 1294 error = 0; 1295 } 1296 1297 if (error == 0) 1298 vfs_timestamp(&wpipe->pipe_mtime); 1299 1300 /* 1301 * We have something to offer, 1302 * wake up select/poll. 1303 */ 1304 if (wpipe->pipe_buffer.cnt) 1305 pipeselwakeup(wpipe); 1306 1307 pipeunlock(wpipe); 1308 PIPE_UNLOCK(rpipe); 1309 return (error); 1310 } 1311 1312 /* ARGSUSED */ 1313 static int 1314 pipe_truncate(fp, length, active_cred, td) 1315 struct file *fp; 1316 off_t length; 1317 struct ucred *active_cred; 1318 struct thread *td; 1319 { 1320 1321 /* For named pipes call the vnode operation. */ 1322 if (fp->f_vnode != NULL) 1323 return (vnops.fo_truncate(fp, length, active_cred, td)); 1324 return (EINVAL); 1325 } 1326 1327 /* 1328 * we implement a very minimal set of ioctls for compatibility with sockets. 1329 */ 1330 static int 1331 pipe_ioctl(fp, cmd, data, active_cred, td) 1332 struct file *fp; 1333 u_long cmd; 1334 void *data; 1335 struct ucred *active_cred; 1336 struct thread *td; 1337 { 1338 struct pipe *mpipe = fp->f_data; 1339 int error; 1340 1341 PIPE_LOCK(mpipe); 1342 1343 #ifdef MAC 1344 error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data); 1345 if (error) { 1346 PIPE_UNLOCK(mpipe); 1347 return (error); 1348 } 1349 #endif 1350 1351 error = 0; 1352 switch (cmd) { 1353 1354 case FIONBIO: 1355 break; 1356 1357 case FIOASYNC: 1358 if (*(int *)data) { 1359 mpipe->pipe_state |= PIPE_ASYNC; 1360 } else { 1361 mpipe->pipe_state &= ~PIPE_ASYNC; 1362 } 1363 break; 1364 1365 case FIONREAD: 1366 if (!(fp->f_flag & FREAD)) { 1367 *(int *)data = 0; 1368 PIPE_UNLOCK(mpipe); 1369 return (0); 1370 } 1371 if (mpipe->pipe_state & PIPE_DIRECTW) 1372 *(int *)data = mpipe->pipe_map.cnt; 1373 else 1374 *(int *)data = mpipe->pipe_buffer.cnt; 1375 break; 1376 1377 case FIOSETOWN: 1378 PIPE_UNLOCK(mpipe); 1379 error = fsetown(*(int *)data, &mpipe->pipe_sigio); 1380 goto out_unlocked; 1381 1382 case FIOGETOWN: 1383 *(int *)data = fgetown(&mpipe->pipe_sigio); 1384 break; 1385 1386 /* This is deprecated, FIOSETOWN should be used instead. */ 1387 case TIOCSPGRP: 1388 PIPE_UNLOCK(mpipe); 1389 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio); 1390 goto out_unlocked; 1391 1392 /* This is deprecated, FIOGETOWN should be used instead. */ 1393 case TIOCGPGRP: 1394 *(int *)data = -fgetown(&mpipe->pipe_sigio); 1395 break; 1396 1397 default: 1398 error = ENOTTY; 1399 break; 1400 } 1401 PIPE_UNLOCK(mpipe); 1402 out_unlocked: 1403 return (error); 1404 } 1405 1406 static int 1407 pipe_poll(fp, events, active_cred, td) 1408 struct file *fp; 1409 int events; 1410 struct ucred *active_cred; 1411 struct thread *td; 1412 { 1413 struct pipe *rpipe; 1414 struct pipe *wpipe; 1415 int levents, revents; 1416 #ifdef MAC 1417 int error; 1418 #endif 1419 1420 revents = 0; 1421 rpipe = fp->f_data; 1422 wpipe = PIPE_PEER(rpipe); 1423 PIPE_LOCK(rpipe); 1424 #ifdef MAC 1425 error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair); 1426 if (error) 1427 goto locked_error; 1428 #endif 1429 if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM)) 1430 if ((rpipe->pipe_state & PIPE_DIRECTW) || 1431 (rpipe->pipe_buffer.cnt > 0)) 1432 revents |= events & (POLLIN | POLLRDNORM); 1433 1434 if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM)) 1435 if (wpipe->pipe_present != PIPE_ACTIVE || 1436 (wpipe->pipe_state & PIPE_EOF) || 1437 (((wpipe->pipe_state & PIPE_DIRECTW) == 0) && 1438 ((wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF || 1439 wpipe->pipe_buffer.size == 0))) 1440 revents |= events & (POLLOUT | POLLWRNORM); 1441 1442 levents = events & 1443 (POLLIN | POLLINIGNEOF | POLLPRI | POLLRDNORM | POLLRDBAND); 1444 if (rpipe->pipe_state & PIPE_NAMED && fp->f_flag & FREAD && levents && 1445 rpipe->pipe_state & PIPE_SAMEWGEN) 1446 events |= POLLINIGNEOF; 1447 1448 if ((events & POLLINIGNEOF) == 0) { 1449 if (rpipe->pipe_state & PIPE_EOF) { 1450 revents |= (events & (POLLIN | POLLRDNORM)); 1451 if (wpipe->pipe_present != PIPE_ACTIVE || 1452 (wpipe->pipe_state & PIPE_EOF)) 1453 revents |= POLLHUP; 1454 } 1455 } 1456 1457 if (revents == 0) { 1458 if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM)) { 1459 selrecord(td, &rpipe->pipe_sel); 1460 if (SEL_WAITING(&rpipe->pipe_sel)) 1461 rpipe->pipe_state |= PIPE_SEL; 1462 } 1463 1464 if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM)) { 1465 selrecord(td, &wpipe->pipe_sel); 1466 if (SEL_WAITING(&wpipe->pipe_sel)) 1467 wpipe->pipe_state |= PIPE_SEL; 1468 } 1469 } 1470 #ifdef MAC 1471 locked_error: 1472 #endif 1473 PIPE_UNLOCK(rpipe); 1474 1475 return (revents); 1476 } 1477 1478 /* 1479 * We shouldn't need locks here as we're doing a read and this should 1480 * be a natural race. 1481 */ 1482 static int 1483 pipe_stat(fp, ub, active_cred, td) 1484 struct file *fp; 1485 struct stat *ub; 1486 struct ucred *active_cred; 1487 struct thread *td; 1488 { 1489 struct pipe *pipe; 1490 int new_unr; 1491 #ifdef MAC 1492 int error; 1493 #endif 1494 1495 pipe = fp->f_data; 1496 PIPE_LOCK(pipe); 1497 #ifdef MAC 1498 error = mac_pipe_check_stat(active_cred, pipe->pipe_pair); 1499 if (error) { 1500 PIPE_UNLOCK(pipe); 1501 return (error); 1502 } 1503 #endif 1504 1505 /* For named pipes ask the underlying filesystem. */ 1506 if (pipe->pipe_state & PIPE_NAMED) { 1507 PIPE_UNLOCK(pipe); 1508 return (vnops.fo_stat(fp, ub, active_cred, td)); 1509 } 1510 1511 /* 1512 * Lazily allocate an inode number for the pipe. Most pipe 1513 * users do not call fstat(2) on the pipe, which means that 1514 * postponing the inode allocation until it is must be 1515 * returned to userland is useful. If alloc_unr failed, 1516 * assign st_ino zero instead of returning an error. 1517 * Special pipe_ino values: 1518 * -1 - not yet initialized; 1519 * 0 - alloc_unr failed, return 0 as st_ino forever. 1520 */ 1521 if (pipe->pipe_ino == (ino_t)-1) { 1522 new_unr = alloc_unr(pipeino_unr); 1523 if (new_unr != -1) 1524 pipe->pipe_ino = new_unr; 1525 else 1526 pipe->pipe_ino = 0; 1527 } 1528 PIPE_UNLOCK(pipe); 1529 1530 bzero(ub, sizeof(*ub)); 1531 ub->st_mode = S_IFIFO; 1532 ub->st_blksize = PAGE_SIZE; 1533 if (pipe->pipe_state & PIPE_DIRECTW) 1534 ub->st_size = pipe->pipe_map.cnt; 1535 else 1536 ub->st_size = pipe->pipe_buffer.cnt; 1537 ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize; 1538 ub->st_atim = pipe->pipe_atime; 1539 ub->st_mtim = pipe->pipe_mtime; 1540 ub->st_ctim = pipe->pipe_ctime; 1541 ub->st_uid = fp->f_cred->cr_uid; 1542 ub->st_gid = fp->f_cred->cr_gid; 1543 ub->st_dev = pipedev_ino; 1544 ub->st_ino = pipe->pipe_ino; 1545 /* 1546 * Left as 0: st_nlink, st_rdev, st_flags, st_gen. 1547 */ 1548 return (0); 1549 } 1550 1551 /* ARGSUSED */ 1552 static int 1553 pipe_close(fp, td) 1554 struct file *fp; 1555 struct thread *td; 1556 { 1557 1558 if (fp->f_vnode != NULL) 1559 return vnops.fo_close(fp, td); 1560 fp->f_ops = &badfileops; 1561 pipe_dtor(fp->f_data); 1562 fp->f_data = NULL; 1563 return (0); 1564 } 1565 1566 static int 1567 pipe_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td) 1568 { 1569 struct pipe *cpipe; 1570 int error; 1571 1572 cpipe = fp->f_data; 1573 if (cpipe->pipe_state & PIPE_NAMED) 1574 error = vn_chmod(fp, mode, active_cred, td); 1575 else 1576 error = invfo_chmod(fp, mode, active_cred, td); 1577 return (error); 1578 } 1579 1580 static int 1581 pipe_chown(fp, uid, gid, active_cred, td) 1582 struct file *fp; 1583 uid_t uid; 1584 gid_t gid; 1585 struct ucred *active_cred; 1586 struct thread *td; 1587 { 1588 struct pipe *cpipe; 1589 int error; 1590 1591 cpipe = fp->f_data; 1592 if (cpipe->pipe_state & PIPE_NAMED) 1593 error = vn_chown(fp, uid, gid, active_cred, td); 1594 else 1595 error = invfo_chown(fp, uid, gid, active_cred, td); 1596 return (error); 1597 } 1598 1599 static void 1600 pipe_free_kmem(cpipe) 1601 struct pipe *cpipe; 1602 { 1603 1604 KASSERT(!mtx_owned(PIPE_MTX(cpipe)), 1605 ("pipe_free_kmem: pipe mutex locked")); 1606 1607 if (cpipe->pipe_buffer.buffer != NULL) { 1608 atomic_subtract_long(&amountpipekva, cpipe->pipe_buffer.size); 1609 vm_map_remove(pipe_map, 1610 (vm_offset_t)cpipe->pipe_buffer.buffer, 1611 (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size); 1612 cpipe->pipe_buffer.buffer = NULL; 1613 } 1614 #ifndef PIPE_NODIRECT 1615 { 1616 cpipe->pipe_map.cnt = 0; 1617 cpipe->pipe_map.pos = 0; 1618 cpipe->pipe_map.npages = 0; 1619 } 1620 #endif 1621 } 1622 1623 /* 1624 * shutdown the pipe 1625 */ 1626 static void 1627 pipeclose(cpipe) 1628 struct pipe *cpipe; 1629 { 1630 struct pipepair *pp; 1631 struct pipe *ppipe; 1632 1633 KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL")); 1634 1635 PIPE_LOCK(cpipe); 1636 pipelock(cpipe, 0); 1637 pp = cpipe->pipe_pair; 1638 1639 pipeselwakeup(cpipe); 1640 1641 /* 1642 * If the other side is blocked, wake it up saying that 1643 * we want to close it down. 1644 */ 1645 cpipe->pipe_state |= PIPE_EOF; 1646 while (cpipe->pipe_busy) { 1647 wakeup(cpipe); 1648 cpipe->pipe_state |= PIPE_WANT; 1649 pipeunlock(cpipe); 1650 msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0); 1651 pipelock(cpipe, 0); 1652 } 1653 1654 1655 /* 1656 * Disconnect from peer, if any. 1657 */ 1658 ppipe = cpipe->pipe_peer; 1659 if (ppipe->pipe_present == PIPE_ACTIVE) { 1660 pipeselwakeup(ppipe); 1661 1662 ppipe->pipe_state |= PIPE_EOF; 1663 wakeup(ppipe); 1664 KNOTE_LOCKED(&ppipe->pipe_sel.si_note, 0); 1665 } 1666 1667 /* 1668 * Mark this endpoint as free. Release kmem resources. We 1669 * don't mark this endpoint as unused until we've finished 1670 * doing that, or the pipe might disappear out from under 1671 * us. 1672 */ 1673 PIPE_UNLOCK(cpipe); 1674 pipe_free_kmem(cpipe); 1675 PIPE_LOCK(cpipe); 1676 cpipe->pipe_present = PIPE_CLOSING; 1677 pipeunlock(cpipe); 1678 1679 /* 1680 * knlist_clear() may sleep dropping the PIPE_MTX. Set the 1681 * PIPE_FINALIZED, that allows other end to free the 1682 * pipe_pair, only after the knotes are completely dismantled. 1683 */ 1684 knlist_clear(&cpipe->pipe_sel.si_note, 1); 1685 cpipe->pipe_present = PIPE_FINALIZED; 1686 seldrain(&cpipe->pipe_sel); 1687 knlist_destroy(&cpipe->pipe_sel.si_note); 1688 1689 /* 1690 * If both endpoints are now closed, release the memory for the 1691 * pipe pair. If not, unlock. 1692 */ 1693 if (ppipe->pipe_present == PIPE_FINALIZED) { 1694 PIPE_UNLOCK(cpipe); 1695 #ifdef MAC 1696 mac_pipe_destroy(pp); 1697 #endif 1698 uma_zfree(pipe_zone, cpipe->pipe_pair); 1699 } else 1700 PIPE_UNLOCK(cpipe); 1701 } 1702 1703 /*ARGSUSED*/ 1704 static int 1705 pipe_kqfilter(struct file *fp, struct knote *kn) 1706 { 1707 struct pipe *cpipe; 1708 1709 /* 1710 * If a filter is requested that is not supported by this file 1711 * descriptor, don't return an error, but also don't ever generate an 1712 * event. 1713 */ 1714 if ((kn->kn_filter == EVFILT_READ) && !(fp->f_flag & FREAD)) { 1715 kn->kn_fop = &pipe_nfiltops; 1716 return (0); 1717 } 1718 if ((kn->kn_filter == EVFILT_WRITE) && !(fp->f_flag & FWRITE)) { 1719 kn->kn_fop = &pipe_nfiltops; 1720 return (0); 1721 } 1722 cpipe = fp->f_data; 1723 PIPE_LOCK(cpipe); 1724 switch (kn->kn_filter) { 1725 case EVFILT_READ: 1726 kn->kn_fop = &pipe_rfiltops; 1727 break; 1728 case EVFILT_WRITE: 1729 kn->kn_fop = &pipe_wfiltops; 1730 if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) { 1731 /* other end of pipe has been closed */ 1732 PIPE_UNLOCK(cpipe); 1733 return (EPIPE); 1734 } 1735 cpipe = PIPE_PEER(cpipe); 1736 break; 1737 default: 1738 PIPE_UNLOCK(cpipe); 1739 return (EINVAL); 1740 } 1741 1742 kn->kn_hook = cpipe; 1743 knlist_add(&cpipe->pipe_sel.si_note, kn, 1); 1744 PIPE_UNLOCK(cpipe); 1745 return (0); 1746 } 1747 1748 static void 1749 filt_pipedetach(struct knote *kn) 1750 { 1751 struct pipe *cpipe = kn->kn_hook; 1752 1753 PIPE_LOCK(cpipe); 1754 knlist_remove(&cpipe->pipe_sel.si_note, kn, 1); 1755 PIPE_UNLOCK(cpipe); 1756 } 1757 1758 /*ARGSUSED*/ 1759 static int 1760 filt_piperead(struct knote *kn, long hint) 1761 { 1762 struct pipe *rpipe = kn->kn_hook; 1763 struct pipe *wpipe = rpipe->pipe_peer; 1764 int ret; 1765 1766 PIPE_LOCK(rpipe); 1767 kn->kn_data = rpipe->pipe_buffer.cnt; 1768 if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW)) 1769 kn->kn_data = rpipe->pipe_map.cnt; 1770 1771 if ((rpipe->pipe_state & PIPE_EOF) || 1772 wpipe->pipe_present != PIPE_ACTIVE || 1773 (wpipe->pipe_state & PIPE_EOF)) { 1774 kn->kn_flags |= EV_EOF; 1775 PIPE_UNLOCK(rpipe); 1776 return (1); 1777 } 1778 ret = kn->kn_data > 0; 1779 PIPE_UNLOCK(rpipe); 1780 return ret; 1781 } 1782 1783 /*ARGSUSED*/ 1784 static int 1785 filt_pipewrite(struct knote *kn, long hint) 1786 { 1787 struct pipe *wpipe; 1788 1789 wpipe = kn->kn_hook; 1790 PIPE_LOCK(wpipe); 1791 if (wpipe->pipe_present != PIPE_ACTIVE || 1792 (wpipe->pipe_state & PIPE_EOF)) { 1793 kn->kn_data = 0; 1794 kn->kn_flags |= EV_EOF; 1795 PIPE_UNLOCK(wpipe); 1796 return (1); 1797 } 1798 kn->kn_data = (wpipe->pipe_buffer.size > 0) ? 1799 (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) : PIPE_BUF; 1800 if (wpipe->pipe_state & PIPE_DIRECTW) 1801 kn->kn_data = 0; 1802 1803 PIPE_UNLOCK(wpipe); 1804 return (kn->kn_data >= PIPE_BUF); 1805 } 1806 1807 static void 1808 filt_pipedetach_notsup(struct knote *kn) 1809 { 1810 1811 } 1812 1813 static int 1814 filt_pipenotsup(struct knote *kn, long hint) 1815 { 1816 1817 return (0); 1818 } 1819