sys_pipe.c revision 123915
1/* 2 * Copyright (c) 1996 John S. Dyson 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice immediately at the beginning of the file, without modification, 10 * this list of conditions, and the following disclaimer. 11 * 2. Redistributions in binary form must reproduce the above copyright 12 * notice, this list of conditions and the following disclaimer in the 13 * documentation and/or other materials provided with the distribution. 14 * 3. Absolutely no warranty of function or purpose is made by the author 15 * John S. Dyson. 16 * 4. Modifications may be freely made to this file if the above conditions 17 * are met. 18 */ 19 20/* 21 * This file contains a high-performance replacement for the socket-based 22 * pipes scheme originally used in FreeBSD/4.4Lite. It does not support 23 * all features of sockets, but does do everything that pipes normally 24 * do. 25 */ 26 27/* 28 * This code has two modes of operation, a small write mode and a large 29 * write mode. The small write mode acts like conventional pipes with 30 * a kernel buffer. If the buffer is less than PIPE_MINDIRECT, then the 31 * "normal" pipe buffering is done. If the buffer is between PIPE_MINDIRECT 32 * and PIPE_SIZE in size, it is fully mapped and wired into the kernel, and 33 * the receiving process can copy it directly from the pages in the sending 34 * process. 35 * 36 * If the sending process receives a signal, it is possible that it will 37 * go away, and certainly its address space can change, because control 38 * is returned back to the user-mode side. In that case, the pipe code 39 * arranges to copy the buffer supplied by the user process, to a pageable 40 * kernel buffer, and the receiving process will grab the data from the 41 * pageable kernel buffer. Since signals don't happen all that often, 42 * the copy operation is normally eliminated. 43 * 44 * The constant PIPE_MINDIRECT is chosen to make sure that buffering will 45 * happen for small transfers so that the system will not spend all of 46 * its time context switching. 47 * 48 * In order to limit the resource use of pipes, two sysctls exist: 49 * 50 * kern.ipc.maxpipekva - This is a hard limit on the amount of pageable 51 * address space available to us in pipe_map. Whenever the amount in use 52 * exceeds half of this value, all new pipes will be created with size 53 * SMALL_PIPE_SIZE, rather than PIPE_SIZE. Big pipe creation will be limited 54 * as well. This value is loader tunable only. 55 * 56 * kern.ipc.maxpipekvawired - This value limits the amount of memory that may 57 * be wired in order to facilitate direct copies using page flipping. 58 * Whenever this value is exceeded, pipes will fall back to using regular 59 * copies. This value is sysctl controllable at all times. 60 * 61 * These values are autotuned in subr_param.c. 62 * 63 * Memory usage may be monitored through the sysctls 64 * kern.ipc.pipes, kern.ipc.pipekva and kern.ipc.pipekvawired. 65 * 66 */ 67 68#include <sys/cdefs.h> 69__FBSDID("$FreeBSD: head/sys/kern/sys_pipe.c 123915 2003-12-28 01:19:58Z silby $"); 70 71#include "opt_mac.h" 72 73#include <sys/param.h> 74#include <sys/systm.h> 75#include <sys/fcntl.h> 76#include <sys/file.h> 77#include <sys/filedesc.h> 78#include <sys/filio.h> 79#include <sys/kernel.h> 80#include <sys/lock.h> 81#include <sys/mac.h> 82#include <sys/mutex.h> 83#include <sys/ttycom.h> 84#include <sys/stat.h> 85#include <sys/malloc.h> 86#include <sys/poll.h> 87#include <sys/selinfo.h> 88#include <sys/signalvar.h> 89#include <sys/sysctl.h> 90#include <sys/sysproto.h> 91#include <sys/pipe.h> 92#include <sys/proc.h> 93#include <sys/vnode.h> 94#include <sys/uio.h> 95#include <sys/event.h> 96 97#include <vm/vm.h> 98#include <vm/vm_param.h> 99#include <vm/vm_object.h> 100#include <vm/vm_kern.h> 101#include <vm/vm_extern.h> 102#include <vm/pmap.h> 103#include <vm/vm_map.h> 104#include <vm/vm_page.h> 105#include <vm/uma.h> 106 107/* 108 * Use this define if you want to disable *fancy* VM things. Expect an 109 * approx 30% decrease in transfer rate. This could be useful for 110 * NetBSD or OpenBSD. 111 */ 112/* #define PIPE_NODIRECT */ 113 114/* 115 * interfaces to the outside world 116 */ 117static fo_rdwr_t pipe_read; 118static fo_rdwr_t pipe_write; 119static fo_ioctl_t pipe_ioctl; 120static fo_poll_t pipe_poll; 121static fo_kqfilter_t pipe_kqfilter; 122static fo_stat_t pipe_stat; 123static fo_close_t pipe_close; 124 125static struct fileops pipeops = { 126 .fo_read = pipe_read, 127 .fo_write = pipe_write, 128 .fo_ioctl = pipe_ioctl, 129 .fo_poll = pipe_poll, 130 .fo_kqfilter = pipe_kqfilter, 131 .fo_stat = pipe_stat, 132 .fo_close = pipe_close, 133 .fo_flags = DFLAG_PASSABLE 134}; 135 136static void filt_pipedetach(struct knote *kn); 137static int filt_piperead(struct knote *kn, long hint); 138static int filt_pipewrite(struct knote *kn, long hint); 139 140static struct filterops pipe_rfiltops = 141 { 1, NULL, filt_pipedetach, filt_piperead }; 142static struct filterops pipe_wfiltops = 143 { 1, NULL, filt_pipedetach, filt_pipewrite }; 144 145/* 146 * Default pipe buffer size(s), this can be kind-of large now because pipe 147 * space is pageable. The pipe code will try to maintain locality of 148 * reference for performance reasons, so small amounts of outstanding I/O 149 * will not wipe the cache. 150 */ 151#define MINPIPESIZE (PIPE_SIZE/3) 152#define MAXPIPESIZE (2*PIPE_SIZE/3) 153 154/* 155 * Limit the number of "big" pipes 156 */ 157#define LIMITBIGPIPES 32 158static int nbigpipe; 159 160static int amountpipes; 161static int amountpipekva; 162static int amountpipekvawired; 163 164SYSCTL_DECL(_kern_ipc); 165 166SYSCTL_INT(_kern_ipc, OID_AUTO, maxpipekva, CTLFLAG_RDTUN, 167 &maxpipekva, 0, "Pipe KVA limit"); 168SYSCTL_INT(_kern_ipc, OID_AUTO, maxpipekvawired, CTLFLAG_RW, 169 &maxpipekvawired, 0, "Pipe KVA wired limit"); 170SYSCTL_INT(_kern_ipc, OID_AUTO, pipes, CTLFLAG_RD, 171 &amountpipes, 0, "Current # of pipes"); 172SYSCTL_INT(_kern_ipc, OID_AUTO, bigpipes, CTLFLAG_RD, 173 &nbigpipe, 0, "Current # of big pipes"); 174SYSCTL_INT(_kern_ipc, OID_AUTO, pipekva, CTLFLAG_RD, 175 &amountpipekva, 0, "Pipe KVA usage"); 176SYSCTL_INT(_kern_ipc, OID_AUTO, pipekvawired, CTLFLAG_RD, 177 &amountpipekvawired, 0, "Pipe wired KVA usage"); 178 179static void pipeinit(void *dummy __unused); 180static void pipeclose(struct pipe *cpipe); 181static void pipe_free_kmem(struct pipe *cpipe); 182static int pipe_create(struct pipe **cpipep); 183static __inline int pipelock(struct pipe *cpipe, int catch); 184static __inline void pipeunlock(struct pipe *cpipe); 185static __inline void pipeselwakeup(struct pipe *cpipe); 186#ifndef PIPE_NODIRECT 187static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio); 188static void pipe_destroy_write_buffer(struct pipe *wpipe); 189static int pipe_direct_write(struct pipe *wpipe, struct uio *uio); 190static void pipe_clone_write_buffer(struct pipe *wpipe); 191#endif 192static int pipespace(struct pipe *cpipe, int size); 193 194static uma_zone_t pipe_zone; 195 196SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL); 197 198static void 199pipeinit(void *dummy __unused) 200{ 201 202 pipe_zone = uma_zcreate("PIPE", sizeof(struct pipe), NULL, 203 NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 204 KASSERT(pipe_zone != NULL, ("pipe_zone not initialized")); 205} 206 207/* 208 * The pipe system call for the DTYPE_PIPE type of pipes 209 */ 210 211/* ARGSUSED */ 212int 213pipe(td, uap) 214 struct thread *td; 215 struct pipe_args /* { 216 int dummy; 217 } */ *uap; 218{ 219 struct filedesc *fdp = td->td_proc->p_fd; 220 struct file *rf, *wf; 221 struct pipe *rpipe, *wpipe; 222 struct mtx *pmtx; 223 int fd, error; 224 225 rpipe = wpipe = NULL; 226 if (pipe_create(&rpipe) || pipe_create(&wpipe)) { 227 pipeclose(rpipe); 228 pipeclose(wpipe); 229 return (ENFILE); 230 } 231 232 rpipe->pipe_state |= PIPE_DIRECTOK; 233 wpipe->pipe_state |= PIPE_DIRECTOK; 234 235 error = falloc(td, &rf, &fd); 236 if (error) { 237 pipeclose(rpipe); 238 pipeclose(wpipe); 239 return (error); 240 } 241 /* An extra reference on `rf' has been held for us by falloc(). */ 242 td->td_retval[0] = fd; 243 244 /* 245 * Warning: once we've gotten past allocation of the fd for the 246 * read-side, we can only drop the read side via fdrop() in order 247 * to avoid races against processes which manage to dup() the read 248 * side while we are blocked trying to allocate the write side. 249 */ 250 FILE_LOCK(rf); 251 rf->f_flag = FREAD | FWRITE; 252 rf->f_type = DTYPE_PIPE; 253 rf->f_data = rpipe; 254 rf->f_ops = &pipeops; 255 FILE_UNLOCK(rf); 256 error = falloc(td, &wf, &fd); 257 if (error) { 258 FILEDESC_LOCK(fdp); 259 if (fdp->fd_ofiles[td->td_retval[0]] == rf) { 260 fdp->fd_ofiles[td->td_retval[0]] = NULL; 261 FILEDESC_UNLOCK(fdp); 262 fdrop(rf, td); 263 } else 264 FILEDESC_UNLOCK(fdp); 265 fdrop(rf, td); 266 /* rpipe has been closed by fdrop(). */ 267 pipeclose(wpipe); 268 return (error); 269 } 270 /* An extra reference on `wf' has been held for us by falloc(). */ 271 FILE_LOCK(wf); 272 wf->f_flag = FREAD | FWRITE; 273 wf->f_type = DTYPE_PIPE; 274 wf->f_data = wpipe; 275 wf->f_ops = &pipeops; 276 FILE_UNLOCK(wf); 277 fdrop(wf, td); 278 td->td_retval[1] = fd; 279 rpipe->pipe_peer = wpipe; 280 wpipe->pipe_peer = rpipe; 281#ifdef MAC 282 /* 283 * struct pipe represents a pipe endpoint. The MAC label is shared 284 * between the connected endpoints. As a result mac_init_pipe() and 285 * mac_create_pipe() should only be called on one of the endpoints 286 * after they have been connected. 287 */ 288 mac_init_pipe(rpipe); 289 mac_create_pipe(td->td_ucred, rpipe); 290#endif 291 pmtx = malloc(sizeof(*pmtx), M_TEMP, M_WAITOK | M_ZERO); 292 mtx_init(pmtx, "pipe mutex", NULL, MTX_DEF | MTX_RECURSE); 293 rpipe->pipe_mtxp = wpipe->pipe_mtxp = pmtx; 294 fdrop(rf, td); 295 296 return (0); 297} 298 299/* 300 * Allocate kva for pipe circular buffer, the space is pageable 301 * This routine will 'realloc' the size of a pipe safely, if it fails 302 * it will retain the old buffer. 303 * If it fails it will return ENOMEM. 304 */ 305static int 306pipespace(cpipe, size) 307 struct pipe *cpipe; 308 int size; 309{ 310 caddr_t buffer; 311 int error; 312 static int curfail = 0; 313 static struct timeval lastfail; 314 315 KASSERT(cpipe->pipe_mtxp == NULL || !mtx_owned(PIPE_MTX(cpipe)), 316 ("pipespace: pipe mutex locked")); 317 318 size = round_page(size); 319 /* 320 * XXX -- minor change needed here for NetBSD/OpenBSD VM systems. 321 */ 322 buffer = (caddr_t) vm_map_min(pipe_map); 323 324 /* 325 * The map entry is, by default, pageable. 326 * XXX -- minor change needed here for NetBSD/OpenBSD VM systems. 327 */ 328 error = vm_map_find(pipe_map, NULL, 0, 329 (vm_offset_t *) &buffer, size, 1, 330 VM_PROT_ALL, VM_PROT_ALL, 0); 331 if (error != KERN_SUCCESS) { 332 if (ppsratecheck(&lastfail, &curfail, 1)) 333 printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n"); 334 return (ENOMEM); 335 } 336 337 /* free old resources if we're resizing */ 338 pipe_free_kmem(cpipe); 339 cpipe->pipe_buffer.buffer = buffer; 340 cpipe->pipe_buffer.size = size; 341 cpipe->pipe_buffer.in = 0; 342 cpipe->pipe_buffer.out = 0; 343 cpipe->pipe_buffer.cnt = 0; 344 atomic_add_int(&amountpipes, 1); 345 atomic_add_int(&amountpipekva, cpipe->pipe_buffer.size); 346 return (0); 347} 348 349/* 350 * initialize and allocate VM and memory for pipe 351 */ 352static int 353pipe_create(cpipep) 354 struct pipe **cpipep; 355{ 356 struct pipe *cpipe; 357 int error; 358 359 *cpipep = uma_zalloc(pipe_zone, M_WAITOK); 360 if (*cpipep == NULL) 361 return (ENOMEM); 362 363 cpipe = *cpipep; 364 365 /* 366 * protect so pipeclose() doesn't follow a junk pointer 367 * if pipespace() fails. 368 */ 369 bzero(&cpipe->pipe_sel, sizeof(cpipe->pipe_sel)); 370 cpipe->pipe_state = 0; 371 cpipe->pipe_peer = NULL; 372 cpipe->pipe_busy = 0; 373 374#ifndef PIPE_NODIRECT 375 /* 376 * pipe data structure initializations to support direct pipe I/O 377 */ 378 cpipe->pipe_map.cnt = 0; 379 cpipe->pipe_map.kva = 0; 380 cpipe->pipe_map.pos = 0; 381 cpipe->pipe_map.npages = 0; 382 /* cpipe->pipe_map.ms[] = invalid */ 383#endif 384 385 cpipe->pipe_mtxp = NULL; /* avoid pipespace assertion */ 386 /* 387 * Reduce to 1/4th pipe size if we're over our global max. 388 */ 389 if (amountpipekva > maxpipekva / 2) 390 error = pipespace(cpipe, SMALL_PIPE_SIZE); 391 else 392 error = pipespace(cpipe, PIPE_SIZE); 393 if (error) 394 return (error); 395 396 vfs_timestamp(&cpipe->pipe_ctime); 397 cpipe->pipe_atime = cpipe->pipe_ctime; 398 cpipe->pipe_mtime = cpipe->pipe_ctime; 399 400 return (0); 401} 402 403 404/* 405 * lock a pipe for I/O, blocking other access 406 */ 407static __inline int 408pipelock(cpipe, catch) 409 struct pipe *cpipe; 410 int catch; 411{ 412 int error; 413 414 PIPE_LOCK_ASSERT(cpipe, MA_OWNED); 415 while (cpipe->pipe_state & PIPE_LOCKFL) { 416 cpipe->pipe_state |= PIPE_LWANT; 417 error = msleep(cpipe, PIPE_MTX(cpipe), 418 catch ? (PRIBIO | PCATCH) : PRIBIO, 419 "pipelk", 0); 420 if (error != 0) 421 return (error); 422 } 423 cpipe->pipe_state |= PIPE_LOCKFL; 424 return (0); 425} 426 427/* 428 * unlock a pipe I/O lock 429 */ 430static __inline void 431pipeunlock(cpipe) 432 struct pipe *cpipe; 433{ 434 435 PIPE_LOCK_ASSERT(cpipe, MA_OWNED); 436 cpipe->pipe_state &= ~PIPE_LOCKFL; 437 if (cpipe->pipe_state & PIPE_LWANT) { 438 cpipe->pipe_state &= ~PIPE_LWANT; 439 wakeup(cpipe); 440 } 441} 442 443static __inline void 444pipeselwakeup(cpipe) 445 struct pipe *cpipe; 446{ 447 448 if (cpipe->pipe_state & PIPE_SEL) { 449 cpipe->pipe_state &= ~PIPE_SEL; 450 selwakeuppri(&cpipe->pipe_sel, PSOCK); 451 } 452 if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio) 453 pgsigio(&cpipe->pipe_sigio, SIGIO, 0); 454 KNOTE(&cpipe->pipe_sel.si_note, 0); 455} 456 457/* ARGSUSED */ 458static int 459pipe_read(fp, uio, active_cred, flags, td) 460 struct file *fp; 461 struct uio *uio; 462 struct ucred *active_cred; 463 struct thread *td; 464 int flags; 465{ 466 struct pipe *rpipe = fp->f_data; 467 int error; 468 int nread = 0; 469 u_int size; 470 471 PIPE_LOCK(rpipe); 472 ++rpipe->pipe_busy; 473 error = pipelock(rpipe, 1); 474 if (error) 475 goto unlocked_error; 476 477#ifdef MAC 478 error = mac_check_pipe_read(active_cred, rpipe); 479 if (error) 480 goto locked_error; 481#endif 482 483 while (uio->uio_resid) { 484 /* 485 * normal pipe buffer receive 486 */ 487 if (rpipe->pipe_buffer.cnt > 0) { 488 size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out; 489 if (size > rpipe->pipe_buffer.cnt) 490 size = rpipe->pipe_buffer.cnt; 491 if (size > (u_int) uio->uio_resid) 492 size = (u_int) uio->uio_resid; 493 494 PIPE_UNLOCK(rpipe); 495 error = uiomove( 496 &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out], 497 size, uio); 498 PIPE_LOCK(rpipe); 499 if (error) 500 break; 501 502 rpipe->pipe_buffer.out += size; 503 if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size) 504 rpipe->pipe_buffer.out = 0; 505 506 rpipe->pipe_buffer.cnt -= size; 507 508 /* 509 * If there is no more to read in the pipe, reset 510 * its pointers to the beginning. This improves 511 * cache hit stats. 512 */ 513 if (rpipe->pipe_buffer.cnt == 0) { 514 rpipe->pipe_buffer.in = 0; 515 rpipe->pipe_buffer.out = 0; 516 } 517 nread += size; 518#ifndef PIPE_NODIRECT 519 /* 520 * Direct copy, bypassing a kernel buffer. 521 */ 522 } else if ((size = rpipe->pipe_map.cnt) && 523 (rpipe->pipe_state & PIPE_DIRECTW)) { 524 caddr_t va; 525 if (size > (u_int) uio->uio_resid) 526 size = (u_int) uio->uio_resid; 527 528 va = (caddr_t) rpipe->pipe_map.kva + 529 rpipe->pipe_map.pos; 530 PIPE_UNLOCK(rpipe); 531 error = uiomove(va, size, uio); 532 PIPE_LOCK(rpipe); 533 if (error) 534 break; 535 nread += size; 536 rpipe->pipe_map.pos += size; 537 rpipe->pipe_map.cnt -= size; 538 if (rpipe->pipe_map.cnt == 0) { 539 rpipe->pipe_state &= ~PIPE_DIRECTW; 540 wakeup(rpipe); 541 } 542#endif 543 } else { 544 /* 545 * detect EOF condition 546 * read returns 0 on EOF, no need to set error 547 */ 548 if (rpipe->pipe_state & PIPE_EOF) 549 break; 550 551 /* 552 * If the "write-side" has been blocked, wake it up now. 553 */ 554 if (rpipe->pipe_state & PIPE_WANTW) { 555 rpipe->pipe_state &= ~PIPE_WANTW; 556 wakeup(rpipe); 557 } 558 559 /* 560 * Break if some data was read. 561 */ 562 if (nread > 0) 563 break; 564 565 /* 566 * Unlock the pipe buffer for our remaining processing. 567 * We will either break out with an error or we will 568 * sleep and relock to loop. 569 */ 570 pipeunlock(rpipe); 571 572 /* 573 * Handle non-blocking mode operation or 574 * wait for more data. 575 */ 576 if (fp->f_flag & FNONBLOCK) { 577 error = EAGAIN; 578 } else { 579 rpipe->pipe_state |= PIPE_WANTR; 580 if ((error = msleep(rpipe, PIPE_MTX(rpipe), 581 PRIBIO | PCATCH, 582 "piperd", 0)) == 0) 583 error = pipelock(rpipe, 1); 584 } 585 if (error) 586 goto unlocked_error; 587 } 588 } 589#ifdef MAC 590locked_error: 591#endif 592 pipeunlock(rpipe); 593 594 /* XXX: should probably do this before getting any locks. */ 595 if (error == 0) 596 vfs_timestamp(&rpipe->pipe_atime); 597unlocked_error: 598 --rpipe->pipe_busy; 599 600 /* 601 * PIPE_WANT processing only makes sense if pipe_busy is 0. 602 */ 603 if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) { 604 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW); 605 wakeup(rpipe); 606 } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) { 607 /* 608 * Handle write blocking hysteresis. 609 */ 610 if (rpipe->pipe_state & PIPE_WANTW) { 611 rpipe->pipe_state &= ~PIPE_WANTW; 612 wakeup(rpipe); 613 } 614 } 615 616 if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF) 617 pipeselwakeup(rpipe); 618 619 PIPE_UNLOCK(rpipe); 620 return (error); 621} 622 623#ifndef PIPE_NODIRECT 624/* 625 * Map the sending processes' buffer into kernel space and wire it. 626 * This is similar to a physical write operation. 627 */ 628static int 629pipe_build_write_buffer(wpipe, uio) 630 struct pipe *wpipe; 631 struct uio *uio; 632{ 633 pmap_t pmap; 634 u_int size; 635 int i, j; 636 vm_offset_t addr, endaddr; 637 638 PIPE_LOCK_ASSERT(wpipe, MA_NOTOWNED); 639 640 size = (u_int) uio->uio_iov->iov_len; 641 if (size > wpipe->pipe_buffer.size) 642 size = wpipe->pipe_buffer.size; 643 644 pmap = vmspace_pmap(curproc->p_vmspace); 645 endaddr = round_page((vm_offset_t)uio->uio_iov->iov_base + size); 646 addr = trunc_page((vm_offset_t)uio->uio_iov->iov_base); 647 for (i = 0; addr < endaddr; addr += PAGE_SIZE, i++) { 648 /* 649 * vm_fault_quick() can sleep. Consequently, 650 * vm_page_lock_queue() and vm_page_unlock_queue() 651 * should not be performed outside of this loop. 652 */ 653 race: 654 if (vm_fault_quick((caddr_t)addr, VM_PROT_READ) < 0) { 655 vm_page_lock_queues(); 656 for (j = 0; j < i; j++) 657 vm_page_unhold(wpipe->pipe_map.ms[j]); 658 vm_page_unlock_queues(); 659 return (EFAULT); 660 } 661 wpipe->pipe_map.ms[i] = pmap_extract_and_hold(pmap, addr, 662 VM_PROT_READ); 663 if (wpipe->pipe_map.ms[i] == NULL) 664 goto race; 665 } 666 667/* 668 * set up the control block 669 */ 670 wpipe->pipe_map.npages = i; 671 wpipe->pipe_map.pos = 672 ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK; 673 wpipe->pipe_map.cnt = size; 674 675/* 676 * and map the buffer 677 */ 678 if (wpipe->pipe_map.kva == 0) { 679 /* 680 * We need to allocate space for an extra page because the 681 * address range might (will) span pages at times. 682 */ 683 wpipe->pipe_map.kva = kmem_alloc_nofault(kernel_map, 684 wpipe->pipe_buffer.size + PAGE_SIZE); 685 atomic_add_int(&amountpipekvawired, 686 wpipe->pipe_buffer.size + PAGE_SIZE); 687 } 688 pmap_qenter(wpipe->pipe_map.kva, wpipe->pipe_map.ms, 689 wpipe->pipe_map.npages); 690 691/* 692 * and update the uio data 693 */ 694 695 uio->uio_iov->iov_len -= size; 696 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size; 697 if (uio->uio_iov->iov_len == 0) 698 uio->uio_iov++; 699 uio->uio_resid -= size; 700 uio->uio_offset += size; 701 return (0); 702} 703 704/* 705 * unmap and unwire the process buffer 706 */ 707static void 708pipe_destroy_write_buffer(wpipe) 709 struct pipe *wpipe; 710{ 711 int i; 712 713 PIPE_LOCK_ASSERT(wpipe, MA_NOTOWNED); 714 if (wpipe->pipe_map.kva) { 715 pmap_qremove(wpipe->pipe_map.kva, wpipe->pipe_map.npages); 716 717 if (amountpipekvawired > maxpipekvawired / 2) { 718 /* Conserve address space */ 719 vm_offset_t kva = wpipe->pipe_map.kva; 720 wpipe->pipe_map.kva = 0; 721 kmem_free(kernel_map, kva, 722 wpipe->pipe_buffer.size + PAGE_SIZE); 723 atomic_subtract_int(&amountpipekvawired, 724 wpipe->pipe_buffer.size + PAGE_SIZE); 725 } 726 } 727 vm_page_lock_queues(); 728 for (i = 0; i < wpipe->pipe_map.npages; i++) { 729 vm_page_unhold(wpipe->pipe_map.ms[i]); 730 } 731 vm_page_unlock_queues(); 732 wpipe->pipe_map.npages = 0; 733} 734 735/* 736 * In the case of a signal, the writing process might go away. This 737 * code copies the data into the circular buffer so that the source 738 * pages can be freed without loss of data. 739 */ 740static void 741pipe_clone_write_buffer(wpipe) 742 struct pipe *wpipe; 743{ 744 int size; 745 int pos; 746 747 PIPE_LOCK_ASSERT(wpipe, MA_OWNED); 748 size = wpipe->pipe_map.cnt; 749 pos = wpipe->pipe_map.pos; 750 751 wpipe->pipe_buffer.in = size; 752 wpipe->pipe_buffer.out = 0; 753 wpipe->pipe_buffer.cnt = size; 754 wpipe->pipe_state &= ~PIPE_DIRECTW; 755 756 PIPE_UNLOCK(wpipe); 757 bcopy((caddr_t) wpipe->pipe_map.kva + pos, 758 wpipe->pipe_buffer.buffer, size); 759 pipe_destroy_write_buffer(wpipe); 760 PIPE_LOCK(wpipe); 761} 762 763/* 764 * This implements the pipe buffer write mechanism. Note that only 765 * a direct write OR a normal pipe write can be pending at any given time. 766 * If there are any characters in the pipe buffer, the direct write will 767 * be deferred until the receiving process grabs all of the bytes from 768 * the pipe buffer. Then the direct mapping write is set-up. 769 */ 770static int 771pipe_direct_write(wpipe, uio) 772 struct pipe *wpipe; 773 struct uio *uio; 774{ 775 int error; 776 777retry: 778 PIPE_LOCK_ASSERT(wpipe, MA_OWNED); 779 while (wpipe->pipe_state & PIPE_DIRECTW) { 780 if (wpipe->pipe_state & PIPE_WANTR) { 781 wpipe->pipe_state &= ~PIPE_WANTR; 782 wakeup(wpipe); 783 } 784 wpipe->pipe_state |= PIPE_WANTW; 785 error = msleep(wpipe, PIPE_MTX(wpipe), 786 PRIBIO | PCATCH, "pipdww", 0); 787 if (error) 788 goto error1; 789 if (wpipe->pipe_state & PIPE_EOF) { 790 error = EPIPE; 791 goto error1; 792 } 793 } 794 wpipe->pipe_map.cnt = 0; /* transfer not ready yet */ 795 if (wpipe->pipe_buffer.cnt > 0) { 796 if (wpipe->pipe_state & PIPE_WANTR) { 797 wpipe->pipe_state &= ~PIPE_WANTR; 798 wakeup(wpipe); 799 } 800 801 wpipe->pipe_state |= PIPE_WANTW; 802 error = msleep(wpipe, PIPE_MTX(wpipe), 803 PRIBIO | PCATCH, "pipdwc", 0); 804 if (error) 805 goto error1; 806 if (wpipe->pipe_state & PIPE_EOF) { 807 error = EPIPE; 808 goto error1; 809 } 810 goto retry; 811 } 812 813 wpipe->pipe_state |= PIPE_DIRECTW; 814 815 pipelock(wpipe, 0); 816 PIPE_UNLOCK(wpipe); 817 error = pipe_build_write_buffer(wpipe, uio); 818 PIPE_LOCK(wpipe); 819 pipeunlock(wpipe); 820 if (error) { 821 wpipe->pipe_state &= ~PIPE_DIRECTW; 822 goto error1; 823 } 824 825 error = 0; 826 while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) { 827 if (wpipe->pipe_state & PIPE_EOF) { 828 pipelock(wpipe, 0); 829 PIPE_UNLOCK(wpipe); 830 pipe_destroy_write_buffer(wpipe); 831 PIPE_LOCK(wpipe); 832 pipeselwakeup(wpipe); 833 pipeunlock(wpipe); 834 error = EPIPE; 835 goto error1; 836 } 837 if (wpipe->pipe_state & PIPE_WANTR) { 838 wpipe->pipe_state &= ~PIPE_WANTR; 839 wakeup(wpipe); 840 } 841 pipeselwakeup(wpipe); 842 error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH, 843 "pipdwt", 0); 844 } 845 846 pipelock(wpipe,0); 847 if (wpipe->pipe_state & PIPE_DIRECTW) { 848 /* 849 * this bit of trickery substitutes a kernel buffer for 850 * the process that might be going away. 851 */ 852 pipe_clone_write_buffer(wpipe); 853 } else { 854 PIPE_UNLOCK(wpipe); 855 pipe_destroy_write_buffer(wpipe); 856 PIPE_LOCK(wpipe); 857 } 858 pipeunlock(wpipe); 859 return (error); 860 861error1: 862 wakeup(wpipe); 863 return (error); 864} 865#endif 866 867static int 868pipe_write(fp, uio, active_cred, flags, td) 869 struct file *fp; 870 struct uio *uio; 871 struct ucred *active_cred; 872 struct thread *td; 873 int flags; 874{ 875 int error = 0; 876 int orig_resid; 877 struct pipe *wpipe, *rpipe; 878 879 rpipe = fp->f_data; 880 wpipe = rpipe->pipe_peer; 881 882 PIPE_LOCK(rpipe); 883 /* 884 * detect loss of pipe read side, issue SIGPIPE if lost. 885 */ 886 if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) { 887 PIPE_UNLOCK(rpipe); 888 return (EPIPE); 889 } 890#ifdef MAC 891 error = mac_check_pipe_write(active_cred, wpipe); 892 if (error) { 893 PIPE_UNLOCK(rpipe); 894 return (error); 895 } 896#endif 897 ++wpipe->pipe_busy; 898 899 /* 900 * If it is advantageous to resize the pipe buffer, do 901 * so. 902 */ 903 if ((uio->uio_resid > PIPE_SIZE) && 904 (amountpipekva < maxpipekva / 2) && 905 (nbigpipe < LIMITBIGPIPES) && 906 (wpipe->pipe_state & PIPE_DIRECTW) == 0 && 907 (wpipe->pipe_buffer.size <= PIPE_SIZE) && 908 (wpipe->pipe_buffer.cnt == 0)) { 909 910 if ((error = pipelock(wpipe, 1)) == 0) { 911 PIPE_UNLOCK(wpipe); 912 if (pipespace(wpipe, BIG_PIPE_SIZE) == 0) 913 atomic_add_int(&nbigpipe, 1); 914 PIPE_LOCK(wpipe); 915 pipeunlock(wpipe); 916 } 917 } 918 919 /* 920 * If an early error occured unbusy and return, waking up any pending 921 * readers. 922 */ 923 if (error) { 924 --wpipe->pipe_busy; 925 if ((wpipe->pipe_busy == 0) && 926 (wpipe->pipe_state & PIPE_WANT)) { 927 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR); 928 wakeup(wpipe); 929 } 930 PIPE_UNLOCK(rpipe); 931 return(error); 932 } 933 934 orig_resid = uio->uio_resid; 935 936 while (uio->uio_resid) { 937 int space; 938 939#ifndef PIPE_NODIRECT 940 /* 941 * If the transfer is large, we can gain performance if 942 * we do process-to-process copies directly. 943 * If the write is non-blocking, we don't use the 944 * direct write mechanism. 945 * 946 * The direct write mechanism will detect the reader going 947 * away on us. 948 */ 949 if ((uio->uio_iov->iov_len >= PIPE_MINDIRECT) && 950 (fp->f_flag & FNONBLOCK) == 0 && 951 amountpipekvawired + uio->uio_resid < maxpipekvawired) { 952 error = pipe_direct_write(wpipe, uio); 953 if (error) 954 break; 955 continue; 956 } 957#endif 958 959 /* 960 * Pipe buffered writes cannot be coincidental with 961 * direct writes. We wait until the currently executing 962 * direct write is completed before we start filling the 963 * pipe buffer. We break out if a signal occurs or the 964 * reader goes away. 965 */ 966 retrywrite: 967 while (wpipe->pipe_state & PIPE_DIRECTW) { 968 if (wpipe->pipe_state & PIPE_WANTR) { 969 wpipe->pipe_state &= ~PIPE_WANTR; 970 wakeup(wpipe); 971 } 972 error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH, 973 "pipbww", 0); 974 if (wpipe->pipe_state & PIPE_EOF) 975 break; 976 if (error) 977 break; 978 } 979 if (wpipe->pipe_state & PIPE_EOF) { 980 error = EPIPE; 981 break; 982 } 983 984 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt; 985 986 /* Writes of size <= PIPE_BUF must be atomic. */ 987 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF)) 988 space = 0; 989 990 if (space > 0) { 991 if ((error = pipelock(wpipe,1)) == 0) { 992 int size; /* Transfer size */ 993 int segsize; /* first segment to transfer */ 994 995 /* 996 * It is possible for a direct write to 997 * slip in on us... handle it here... 998 */ 999 if (wpipe->pipe_state & PIPE_DIRECTW) { 1000 pipeunlock(wpipe); 1001 goto retrywrite; 1002 } 1003 /* 1004 * If a process blocked in uiomove, our 1005 * value for space might be bad. 1006 * 1007 * XXX will we be ok if the reader has gone 1008 * away here? 1009 */ 1010 if (space > wpipe->pipe_buffer.size - 1011 wpipe->pipe_buffer.cnt) { 1012 pipeunlock(wpipe); 1013 goto retrywrite; 1014 } 1015 1016 /* 1017 * Transfer size is minimum of uio transfer 1018 * and free space in pipe buffer. 1019 */ 1020 if (space > uio->uio_resid) 1021 size = uio->uio_resid; 1022 else 1023 size = space; 1024 /* 1025 * First segment to transfer is minimum of 1026 * transfer size and contiguous space in 1027 * pipe buffer. If first segment to transfer 1028 * is less than the transfer size, we've got 1029 * a wraparound in the buffer. 1030 */ 1031 segsize = wpipe->pipe_buffer.size - 1032 wpipe->pipe_buffer.in; 1033 if (segsize > size) 1034 segsize = size; 1035 1036 /* Transfer first segment */ 1037 1038 PIPE_UNLOCK(rpipe); 1039 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in], 1040 segsize, uio); 1041 PIPE_LOCK(rpipe); 1042 1043 if (error == 0 && segsize < size) { 1044 /* 1045 * Transfer remaining part now, to 1046 * support atomic writes. Wraparound 1047 * happened. 1048 */ 1049 if (wpipe->pipe_buffer.in + segsize != 1050 wpipe->pipe_buffer.size) 1051 panic("Expected pipe buffer " 1052 "wraparound disappeared"); 1053 1054 PIPE_UNLOCK(rpipe); 1055 error = uiomove( 1056 &wpipe->pipe_buffer.buffer[0], 1057 size - segsize, uio); 1058 PIPE_LOCK(rpipe); 1059 } 1060 if (error == 0) { 1061 wpipe->pipe_buffer.in += size; 1062 if (wpipe->pipe_buffer.in >= 1063 wpipe->pipe_buffer.size) { 1064 if (wpipe->pipe_buffer.in != 1065 size - segsize + 1066 wpipe->pipe_buffer.size) 1067 panic("Expected " 1068 "wraparound bad"); 1069 wpipe->pipe_buffer.in = size - 1070 segsize; 1071 } 1072 1073 wpipe->pipe_buffer.cnt += size; 1074 if (wpipe->pipe_buffer.cnt > 1075 wpipe->pipe_buffer.size) 1076 panic("Pipe buffer overflow"); 1077 1078 } 1079 pipeunlock(wpipe); 1080 } 1081 if (error) 1082 break; 1083 1084 } else { 1085 /* 1086 * If the "read-side" has been blocked, wake it up now. 1087 */ 1088 if (wpipe->pipe_state & PIPE_WANTR) { 1089 wpipe->pipe_state &= ~PIPE_WANTR; 1090 wakeup(wpipe); 1091 } 1092 1093 /* 1094 * don't block on non-blocking I/O 1095 */ 1096 if (fp->f_flag & FNONBLOCK) { 1097 error = EAGAIN; 1098 break; 1099 } 1100 1101 /* 1102 * We have no more space and have something to offer, 1103 * wake up select/poll. 1104 */ 1105 pipeselwakeup(wpipe); 1106 1107 wpipe->pipe_state |= PIPE_WANTW; 1108 error = msleep(wpipe, PIPE_MTX(rpipe), 1109 PRIBIO | PCATCH, "pipewr", 0); 1110 if (error != 0) 1111 break; 1112 /* 1113 * If read side wants to go away, we just issue a signal 1114 * to ourselves. 1115 */ 1116 if (wpipe->pipe_state & PIPE_EOF) { 1117 error = EPIPE; 1118 break; 1119 } 1120 } 1121 } 1122 1123 --wpipe->pipe_busy; 1124 1125 if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) { 1126 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR); 1127 wakeup(wpipe); 1128 } else if (wpipe->pipe_buffer.cnt > 0) { 1129 /* 1130 * If we have put any characters in the buffer, we wake up 1131 * the reader. 1132 */ 1133 if (wpipe->pipe_state & PIPE_WANTR) { 1134 wpipe->pipe_state &= ~PIPE_WANTR; 1135 wakeup(wpipe); 1136 } 1137 } 1138 1139 /* 1140 * Don't return EPIPE if I/O was successful 1141 */ 1142 if ((wpipe->pipe_buffer.cnt == 0) && 1143 (uio->uio_resid == 0) && 1144 (error == EPIPE)) { 1145 error = 0; 1146 } 1147 1148 if (error == 0) 1149 vfs_timestamp(&wpipe->pipe_mtime); 1150 1151 /* 1152 * We have something to offer, 1153 * wake up select/poll. 1154 */ 1155 if (wpipe->pipe_buffer.cnt) 1156 pipeselwakeup(wpipe); 1157 1158 PIPE_UNLOCK(rpipe); 1159 return (error); 1160} 1161 1162/* 1163 * we implement a very minimal set of ioctls for compatibility with sockets. 1164 */ 1165static int 1166pipe_ioctl(fp, cmd, data, active_cred, td) 1167 struct file *fp; 1168 u_long cmd; 1169 void *data; 1170 struct ucred *active_cred; 1171 struct thread *td; 1172{ 1173 struct pipe *mpipe = fp->f_data; 1174#ifdef MAC 1175 int error; 1176#endif 1177 1178 PIPE_LOCK(mpipe); 1179 1180#ifdef MAC 1181 error = mac_check_pipe_ioctl(active_cred, mpipe, cmd, data); 1182 if (error) { 1183 PIPE_UNLOCK(mpipe); 1184 return (error); 1185 } 1186#endif 1187 1188 switch (cmd) { 1189 1190 case FIONBIO: 1191 PIPE_UNLOCK(mpipe); 1192 return (0); 1193 1194 case FIOASYNC: 1195 if (*(int *)data) { 1196 mpipe->pipe_state |= PIPE_ASYNC; 1197 } else { 1198 mpipe->pipe_state &= ~PIPE_ASYNC; 1199 } 1200 PIPE_UNLOCK(mpipe); 1201 return (0); 1202 1203 case FIONREAD: 1204 if (mpipe->pipe_state & PIPE_DIRECTW) 1205 *(int *)data = mpipe->pipe_map.cnt; 1206 else 1207 *(int *)data = mpipe->pipe_buffer.cnt; 1208 PIPE_UNLOCK(mpipe); 1209 return (0); 1210 1211 case FIOSETOWN: 1212 PIPE_UNLOCK(mpipe); 1213 return (fsetown(*(int *)data, &mpipe->pipe_sigio)); 1214 1215 case FIOGETOWN: 1216 PIPE_UNLOCK(mpipe); 1217 *(int *)data = fgetown(&mpipe->pipe_sigio); 1218 return (0); 1219 1220 /* This is deprecated, FIOSETOWN should be used instead. */ 1221 case TIOCSPGRP: 1222 PIPE_UNLOCK(mpipe); 1223 return (fsetown(-(*(int *)data), &mpipe->pipe_sigio)); 1224 1225 /* This is deprecated, FIOGETOWN should be used instead. */ 1226 case TIOCGPGRP: 1227 PIPE_UNLOCK(mpipe); 1228 *(int *)data = -fgetown(&mpipe->pipe_sigio); 1229 return (0); 1230 1231 } 1232 PIPE_UNLOCK(mpipe); 1233 return (ENOTTY); 1234} 1235 1236static int 1237pipe_poll(fp, events, active_cred, td) 1238 struct file *fp; 1239 int events; 1240 struct ucred *active_cred; 1241 struct thread *td; 1242{ 1243 struct pipe *rpipe = fp->f_data; 1244 struct pipe *wpipe; 1245 int revents = 0; 1246#ifdef MAC 1247 int error; 1248#endif 1249 1250 wpipe = rpipe->pipe_peer; 1251 PIPE_LOCK(rpipe); 1252#ifdef MAC 1253 error = mac_check_pipe_poll(active_cred, rpipe); 1254 if (error) 1255 goto locked_error; 1256#endif 1257 if (events & (POLLIN | POLLRDNORM)) 1258 if ((rpipe->pipe_state & PIPE_DIRECTW) || 1259 (rpipe->pipe_buffer.cnt > 0) || 1260 (rpipe->pipe_state & PIPE_EOF)) 1261 revents |= events & (POLLIN | POLLRDNORM); 1262 1263 if (events & (POLLOUT | POLLWRNORM)) 1264 if (wpipe == NULL || (wpipe->pipe_state & PIPE_EOF) || 1265 (((wpipe->pipe_state & PIPE_DIRECTW) == 0) && 1266 (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF)) 1267 revents |= events & (POLLOUT | POLLWRNORM); 1268 1269 if ((rpipe->pipe_state & PIPE_EOF) || 1270 (wpipe == NULL) || 1271 (wpipe->pipe_state & PIPE_EOF)) 1272 revents |= POLLHUP; 1273 1274 if (revents == 0) { 1275 if (events & (POLLIN | POLLRDNORM)) { 1276 selrecord(td, &rpipe->pipe_sel); 1277 rpipe->pipe_state |= PIPE_SEL; 1278 } 1279 1280 if (events & (POLLOUT | POLLWRNORM)) { 1281 selrecord(td, &wpipe->pipe_sel); 1282 wpipe->pipe_state |= PIPE_SEL; 1283 } 1284 } 1285#ifdef MAC 1286locked_error: 1287#endif 1288 PIPE_UNLOCK(rpipe); 1289 1290 return (revents); 1291} 1292 1293/* 1294 * We shouldn't need locks here as we're doing a read and this should 1295 * be a natural race. 1296 */ 1297static int 1298pipe_stat(fp, ub, active_cred, td) 1299 struct file *fp; 1300 struct stat *ub; 1301 struct ucred *active_cred; 1302 struct thread *td; 1303{ 1304 struct pipe *pipe = fp->f_data; 1305#ifdef MAC 1306 int error; 1307 1308 PIPE_LOCK(pipe); 1309 error = mac_check_pipe_stat(active_cred, pipe); 1310 PIPE_UNLOCK(pipe); 1311 if (error) 1312 return (error); 1313#endif 1314 bzero(ub, sizeof(*ub)); 1315 ub->st_mode = S_IFIFO; 1316 ub->st_blksize = pipe->pipe_buffer.size; 1317 ub->st_size = pipe->pipe_buffer.cnt; 1318 ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize; 1319 ub->st_atimespec = pipe->pipe_atime; 1320 ub->st_mtimespec = pipe->pipe_mtime; 1321 ub->st_ctimespec = pipe->pipe_ctime; 1322 ub->st_uid = fp->f_cred->cr_uid; 1323 ub->st_gid = fp->f_cred->cr_gid; 1324 /* 1325 * Left as 0: st_dev, st_ino, st_nlink, st_rdev, st_flags, st_gen. 1326 * XXX (st_dev, st_ino) should be unique. 1327 */ 1328 return (0); 1329} 1330 1331/* ARGSUSED */ 1332static int 1333pipe_close(fp, td) 1334 struct file *fp; 1335 struct thread *td; 1336{ 1337 struct pipe *cpipe = fp->f_data; 1338 1339 fp->f_ops = &badfileops; 1340 fp->f_data = NULL; 1341 funsetown(&cpipe->pipe_sigio); 1342 pipeclose(cpipe); 1343 return (0); 1344} 1345 1346static void 1347pipe_free_kmem(cpipe) 1348 struct pipe *cpipe; 1349{ 1350 1351 KASSERT(cpipe->pipe_mtxp == NULL || !mtx_owned(PIPE_MTX(cpipe)), 1352 ("pipespace: pipe mutex locked")); 1353 1354 if (cpipe->pipe_buffer.buffer != NULL) { 1355 if (cpipe->pipe_buffer.size > PIPE_SIZE) 1356 atomic_subtract_int(&nbigpipe, 1); 1357 atomic_subtract_int(&amountpipekva, cpipe->pipe_buffer.size); 1358 atomic_subtract_int(&amountpipes, 1); 1359 vm_map_remove(pipe_map, 1360 (vm_offset_t)cpipe->pipe_buffer.buffer, 1361 (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size); 1362 cpipe->pipe_buffer.buffer = NULL; 1363 } 1364#ifndef PIPE_NODIRECT 1365 if (cpipe->pipe_map.kva != 0) { 1366 atomic_subtract_int(&amountpipekvawired, 1367 cpipe->pipe_buffer.size + PAGE_SIZE); 1368 kmem_free(kernel_map, 1369 cpipe->pipe_map.kva, 1370 cpipe->pipe_buffer.size + PAGE_SIZE); 1371 cpipe->pipe_map.cnt = 0; 1372 cpipe->pipe_map.kva = 0; 1373 cpipe->pipe_map.pos = 0; 1374 cpipe->pipe_map.npages = 0; 1375 } 1376#endif 1377} 1378 1379/* 1380 * shutdown the pipe 1381 */ 1382static void 1383pipeclose(cpipe) 1384 struct pipe *cpipe; 1385{ 1386 struct pipe *ppipe; 1387 int hadpeer; 1388 1389 if (cpipe == NULL) 1390 return; 1391 1392 hadpeer = 0; 1393 1394 /* partially created pipes won't have a valid mutex. */ 1395 if (PIPE_MTX(cpipe) != NULL) 1396 PIPE_LOCK(cpipe); 1397 1398 pipeselwakeup(cpipe); 1399 1400 /* 1401 * If the other side is blocked, wake it up saying that 1402 * we want to close it down. 1403 */ 1404 while (cpipe->pipe_busy) { 1405 wakeup(cpipe); 1406 cpipe->pipe_state |= PIPE_WANT | PIPE_EOF; 1407 msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0); 1408 } 1409 1410#ifdef MAC 1411 if (cpipe->pipe_label != NULL && cpipe->pipe_peer == NULL) 1412 mac_destroy_pipe(cpipe); 1413#endif 1414 1415 /* 1416 * Disconnect from peer 1417 */ 1418 if ((ppipe = cpipe->pipe_peer) != NULL) { 1419 hadpeer++; 1420 pipeselwakeup(ppipe); 1421 1422 ppipe->pipe_state |= PIPE_EOF; 1423 wakeup(ppipe); 1424 KNOTE(&ppipe->pipe_sel.si_note, 0); 1425 ppipe->pipe_peer = NULL; 1426 } 1427 /* 1428 * free resources 1429 */ 1430 if (PIPE_MTX(cpipe) != NULL) { 1431 PIPE_UNLOCK(cpipe); 1432 if (!hadpeer) { 1433 mtx_destroy(PIPE_MTX(cpipe)); 1434 free(PIPE_MTX(cpipe), M_TEMP); 1435 } 1436 } 1437 pipe_free_kmem(cpipe); 1438 uma_zfree(pipe_zone, cpipe); 1439} 1440 1441/*ARGSUSED*/ 1442static int 1443pipe_kqfilter(struct file *fp, struct knote *kn) 1444{ 1445 struct pipe *cpipe; 1446 1447 cpipe = kn->kn_fp->f_data; 1448 switch (kn->kn_filter) { 1449 case EVFILT_READ: 1450 kn->kn_fop = &pipe_rfiltops; 1451 break; 1452 case EVFILT_WRITE: 1453 kn->kn_fop = &pipe_wfiltops; 1454 cpipe = cpipe->pipe_peer; 1455 if (cpipe == NULL) 1456 /* other end of pipe has been closed */ 1457 return (EPIPE); 1458 break; 1459 default: 1460 return (1); 1461 } 1462 1463 PIPE_LOCK(cpipe); 1464 SLIST_INSERT_HEAD(&cpipe->pipe_sel.si_note, kn, kn_selnext); 1465 PIPE_UNLOCK(cpipe); 1466 return (0); 1467} 1468 1469static void 1470filt_pipedetach(struct knote *kn) 1471{ 1472 struct pipe *cpipe = (struct pipe *)kn->kn_fp->f_data; 1473 1474 if (kn->kn_filter == EVFILT_WRITE) { 1475 if (cpipe->pipe_peer == NULL) 1476 return; 1477 cpipe = cpipe->pipe_peer; 1478 } 1479 1480 PIPE_LOCK(cpipe); 1481 SLIST_REMOVE(&cpipe->pipe_sel.si_note, kn, knote, kn_selnext); 1482 PIPE_UNLOCK(cpipe); 1483} 1484 1485/*ARGSUSED*/ 1486static int 1487filt_piperead(struct knote *kn, long hint) 1488{ 1489 struct pipe *rpipe = kn->kn_fp->f_data; 1490 struct pipe *wpipe = rpipe->pipe_peer; 1491 1492 PIPE_LOCK(rpipe); 1493 kn->kn_data = rpipe->pipe_buffer.cnt; 1494 if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW)) 1495 kn->kn_data = rpipe->pipe_map.cnt; 1496 1497 if ((rpipe->pipe_state & PIPE_EOF) || 1498 (wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) { 1499 kn->kn_flags |= EV_EOF; 1500 PIPE_UNLOCK(rpipe); 1501 return (1); 1502 } 1503 PIPE_UNLOCK(rpipe); 1504 return (kn->kn_data > 0); 1505} 1506 1507/*ARGSUSED*/ 1508static int 1509filt_pipewrite(struct knote *kn, long hint) 1510{ 1511 struct pipe *rpipe = kn->kn_fp->f_data; 1512 struct pipe *wpipe = rpipe->pipe_peer; 1513 1514 PIPE_LOCK(rpipe); 1515 if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) { 1516 kn->kn_data = 0; 1517 kn->kn_flags |= EV_EOF; 1518 PIPE_UNLOCK(rpipe); 1519 return (1); 1520 } 1521 kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt; 1522 if (wpipe->pipe_state & PIPE_DIRECTW) 1523 kn->kn_data = 0; 1524 1525 PIPE_UNLOCK(rpipe); 1526 return (kn->kn_data >= PIPE_BUF); 1527} 1528