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("$FreeBSD$"); 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 */ 117 static fo_rdwr_t pipe_read; 118 static fo_rdwr_t pipe_write; 119 static fo_ioctl_t pipe_ioctl; 120 static fo_poll_t pipe_poll; 121 static fo_kqfilter_t pipe_kqfilter; 122 static fo_stat_t pipe_stat; 123 static fo_close_t pipe_close; 124 125 static 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 }; 135 136 static void filt_pipedetach(struct knote *kn); 137 static int filt_piperead(struct knote *kn, long hint); 138 static int filt_pipewrite(struct knote *kn, long hint); 139 140 static struct filterops pipe_rfiltops = 141 { 1, NULL, filt_pipedetach, filt_piperead }; 142 static 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 32 158 static int nbigpipe; 159 160 static int amountpipes; 161 static int amountpipekva; 162 static int amountpipekvawired; 163 164 SYSCTL_DECL(_kern_ipc); 165 166 SYSCTL_INT(_kern_ipc, OID_AUTO, maxpipekva, CTLFLAG_RDTUN, 167 &maxpipekva, 0, "Pipe KVA limit"); 168 SYSCTL_INT(_kern_ipc, OID_AUTO, maxpipekvawired, CTLFLAG_RW, 169 &maxpipekvawired, 0, "Pipe KVA wired limit"); 170 SYSCTL_INT(_kern_ipc, OID_AUTO, pipes, CTLFLAG_RD, 171 &amountpipes, 0, "Current # of pipes"); 172 SYSCTL_INT(_kern_ipc, OID_AUTO, bigpipes, CTLFLAG_RD, 173 &nbigpipe, 0, "Current # of big pipes"); 174 SYSCTL_INT(_kern_ipc, OID_AUTO, pipekva, CTLFLAG_RD, 175 &amountpipekva, 0, "Pipe KVA usage"); 176 SYSCTL_INT(_kern_ipc, OID_AUTO, pipekvawired, CTLFLAG_RD, 177 &amountpipekvawired, 0, "Pipe wired KVA usage"); 178 179 static void pipeinit(void *dummy __unused); 180 static void pipeclose(struct pipe *cpipe); 181 static void pipe_free_kmem(struct pipe *cpipe); 182 static int pipe_create(struct pipe **cpipep); 183 static __inline int pipelock(struct pipe *cpipe, int catch); 184 static __inline void pipeunlock(struct pipe *cpipe); 185 static __inline void pipeselwakeup(struct pipe *cpipe); 186 #ifndef PIPE_NODIRECT 187 static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio); 188 static void pipe_destroy_write_buffer(struct pipe *wpipe); 189 static int pipe_direct_write(struct pipe *wpipe, struct uio *uio); 190 static void pipe_clone_write_buffer(struct pipe *wpipe); 191 #endif 192 static int pipespace(struct pipe *cpipe, int size); 193 194 static uma_zone_t pipe_zone; 195 196 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL); 197 198 static void 199 pipeinit(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 */ 212 int 213 pipe(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 fdunused(fdp, td->td_retval[0]); 262 FILEDESC_UNLOCK(fdp); 263 fdrop(rf, td); 264 } else { 265 FILEDESC_UNLOCK(fdp); 266 } 267 fdrop(rf, td); 268 /* rpipe has been closed by fdrop(). */ 269 pipeclose(wpipe); 270 return (error); 271 } 272 /* An extra reference on `wf' has been held for us by falloc(). */ 273 FILE_LOCK(wf); 274 wf->f_flag = FREAD | FWRITE; 275 wf->f_type = DTYPE_PIPE; 276 wf->f_data = wpipe; 277 wf->f_ops = &pipeops; 278 FILE_UNLOCK(wf); 279 fdrop(wf, td); 280 td->td_retval[1] = fd; 281 rpipe->pipe_peer = wpipe; 282 wpipe->pipe_peer = rpipe; 283 #ifdef MAC 284 /* 285 * struct pipe represents a pipe endpoint. The MAC label is shared 286 * between the connected endpoints. As a result mac_init_pipe() and 287 * mac_create_pipe() should only be called on one of the endpoints 288 * after they have been connected. 289 */ 290 mac_init_pipe(rpipe); 291 mac_create_pipe(td->td_ucred, rpipe); 292 #endif 293 pmtx = malloc(sizeof(*pmtx), M_TEMP, M_WAITOK | M_ZERO); 294 mtx_init(pmtx, "pipe mutex", NULL, MTX_DEF | MTX_RECURSE); 295 rpipe->pipe_mtxp = wpipe->pipe_mtxp = pmtx; 296 fdrop(rf, td); 297 298 return (0); 299 } 300 301 /* 302 * Allocate kva for pipe circular buffer, the space is pageable 303 * This routine will 'realloc' the size of a pipe safely, if it fails 304 * it will retain the old buffer. 305 * If it fails it will return ENOMEM. 306 */ 307 static int 308 pipespace(cpipe, size) 309 struct pipe *cpipe; 310 int size; 311 { 312 caddr_t buffer; 313 int error; 314 static int curfail = 0; 315 static struct timeval lastfail; 316 317 KASSERT(cpipe->pipe_mtxp == NULL || !mtx_owned(PIPE_MTX(cpipe)), 318 ("pipespace: pipe mutex locked")); 319 320 size = round_page(size); 321 /* 322 * XXX -- minor change needed here for NetBSD/OpenBSD VM systems. 323 */ 324 buffer = (caddr_t) vm_map_min(pipe_map); 325 326 /* 327 * The map entry is, by default, pageable. 328 * XXX -- minor change needed here for NetBSD/OpenBSD VM systems. 329 */ 330 error = vm_map_find(pipe_map, NULL, 0, 331 (vm_offset_t *) &buffer, size, 1, 332 VM_PROT_ALL, VM_PROT_ALL, 0); 333 if (error != KERN_SUCCESS) { 334 if (ppsratecheck(&lastfail, &curfail, 1)) 335 printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n"); 336 return (ENOMEM); 337 } 338 339 /* free old resources if we're resizing */ 340 pipe_free_kmem(cpipe); 341 cpipe->pipe_buffer.buffer = buffer; 342 cpipe->pipe_buffer.size = size; 343 cpipe->pipe_buffer.in = 0; 344 cpipe->pipe_buffer.out = 0; 345 cpipe->pipe_buffer.cnt = 0; 346 atomic_add_int(&amountpipes, 1); 347 atomic_add_int(&amountpipekva, cpipe->pipe_buffer.size); 348 return (0); 349 } 350 351 /* 352 * initialize and allocate VM and memory for pipe 353 */ 354 static int 355 pipe_create(cpipep) 356 struct pipe **cpipep; 357 { 358 struct pipe *cpipe; 359 int error; 360 361 *cpipep = uma_zalloc(pipe_zone, M_WAITOK); 362 if (*cpipep == NULL) 363 return (ENOMEM); 364 365 cpipe = *cpipep; 366 367 /* 368 * protect so pipeclose() doesn't follow a junk pointer 369 * if pipespace() fails. 370 */ 371 bzero(&cpipe->pipe_sel, sizeof(cpipe->pipe_sel)); 372 cpipe->pipe_state = 0; 373 cpipe->pipe_peer = NULL; 374 cpipe->pipe_busy = 0; 375 376 #ifndef PIPE_NODIRECT 377 /* 378 * pipe data structure initializations to support direct pipe I/O 379 */ 380 cpipe->pipe_map.cnt = 0; 381 cpipe->pipe_map.kva = 0; 382 cpipe->pipe_map.pos = 0; 383 cpipe->pipe_map.npages = 0; 384 /* cpipe->pipe_map.ms[] = invalid */ 385 #endif 386 387 cpipe->pipe_mtxp = NULL; /* avoid pipespace assertion */ 388 /* 389 * Reduce to 1/4th pipe size if we're over our global max. 390 */ 391 if (amountpipekva > maxpipekva / 2) 392 error = pipespace(cpipe, SMALL_PIPE_SIZE); 393 else 394 error = pipespace(cpipe, PIPE_SIZE); 395 if (error) 396 return (error); 397 398 vfs_timestamp(&cpipe->pipe_ctime); 399 cpipe->pipe_atime = cpipe->pipe_ctime; 400 cpipe->pipe_mtime = cpipe->pipe_ctime; 401 402 return (0); 403 } 404 405 406 /* 407 * lock a pipe for I/O, blocking other access 408 */ 409 static __inline int 410 pipelock(cpipe, catch) 411 struct pipe *cpipe; 412 int catch; 413 { 414 int error; 415 416 PIPE_LOCK_ASSERT(cpipe, MA_OWNED); 417 while (cpipe->pipe_state & PIPE_LOCKFL) { 418 cpipe->pipe_state |= PIPE_LWANT; 419 error = msleep(cpipe, PIPE_MTX(cpipe), 420 catch ? (PRIBIO | PCATCH) : PRIBIO, 421 "pipelk", 0); 422 if (error != 0) 423 return (error); 424 } 425 cpipe->pipe_state |= PIPE_LOCKFL; 426 return (0); 427 } 428 429 /* 430 * unlock a pipe I/O lock 431 */ 432 static __inline void 433 pipeunlock(cpipe) 434 struct pipe *cpipe; 435 { 436 437 PIPE_LOCK_ASSERT(cpipe, MA_OWNED); 438 cpipe->pipe_state &= ~PIPE_LOCKFL; 439 if (cpipe->pipe_state & PIPE_LWANT) { 440 cpipe->pipe_state &= ~PIPE_LWANT; 441 wakeup(cpipe); 442 } 443 } 444 445 static __inline void 446 pipeselwakeup(cpipe) 447 struct pipe *cpipe; 448 { 449 450 if (cpipe->pipe_state & PIPE_SEL) { 451 cpipe->pipe_state &= ~PIPE_SEL; 452 selwakeuppri(&cpipe->pipe_sel, PSOCK); 453 } 454 if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio) 455 pgsigio(&cpipe->pipe_sigio, SIGIO, 0); 456 KNOTE(&cpipe->pipe_sel.si_note, 0); 457 } 458 459 /* ARGSUSED */ 460 static int 461 pipe_read(fp, uio, active_cred, flags, td) 462 struct file *fp; 463 struct uio *uio; 464 struct ucred *active_cred; 465 struct thread *td; 466 int flags; 467 { 468 struct pipe *rpipe = fp->f_data; 469 int error; 470 int nread = 0; 471 u_int size; 472 473 PIPE_LOCK(rpipe); 474 ++rpipe->pipe_busy; 475 error = pipelock(rpipe, 1); 476 if (error) 477 goto unlocked_error; 478 479 #ifdef MAC 480 error = mac_check_pipe_read(active_cred, rpipe); 481 if (error) 482 goto locked_error; 483 #endif 484 485 while (uio->uio_resid) { 486 /* 487 * normal pipe buffer receive 488 */ 489 if (rpipe->pipe_buffer.cnt > 0) { 490 size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out; 491 if (size > rpipe->pipe_buffer.cnt) 492 size = rpipe->pipe_buffer.cnt; 493 if (size > (u_int) uio->uio_resid) 494 size = (u_int) uio->uio_resid; 495 496 PIPE_UNLOCK(rpipe); 497 error = uiomove( 498 &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out], 499 size, uio); 500 PIPE_LOCK(rpipe); 501 if (error) 502 break; 503 504 rpipe->pipe_buffer.out += size; 505 if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size) 506 rpipe->pipe_buffer.out = 0; 507 508 rpipe->pipe_buffer.cnt -= size; 509 510 /* 511 * If there is no more to read in the pipe, reset 512 * its pointers to the beginning. This improves 513 * cache hit stats. 514 */ 515 if (rpipe->pipe_buffer.cnt == 0) { 516 rpipe->pipe_buffer.in = 0; 517 rpipe->pipe_buffer.out = 0; 518 } 519 nread += size; 520 #ifndef PIPE_NODIRECT 521 /* 522 * Direct copy, bypassing a kernel buffer. 523 */ 524 } else if ((size = rpipe->pipe_map.cnt) && 525 (rpipe->pipe_state & PIPE_DIRECTW)) { 526 caddr_t va; 527 if (size > (u_int) uio->uio_resid) 528 size = (u_int) uio->uio_resid; 529 530 va = (caddr_t) rpipe->pipe_map.kva + 531 rpipe->pipe_map.pos; 532 PIPE_UNLOCK(rpipe); 533 error = uiomove(va, size, uio); 534 PIPE_LOCK(rpipe); 535 if (error) 536 break; 537 nread += size; 538 rpipe->pipe_map.pos += size; 539 rpipe->pipe_map.cnt -= size; 540 if (rpipe->pipe_map.cnt == 0) { 541 rpipe->pipe_state &= ~PIPE_DIRECTW; 542 wakeup(rpipe); 543 } 544 #endif 545 } else { 546 /* 547 * detect EOF condition 548 * read returns 0 on EOF, no need to set error 549 */ 550 if (rpipe->pipe_state & PIPE_EOF) 551 break; 552 553 /* 554 * If the "write-side" has been blocked, wake it up now. 555 */ 556 if (rpipe->pipe_state & PIPE_WANTW) { 557 rpipe->pipe_state &= ~PIPE_WANTW; 558 wakeup(rpipe); 559 } 560 561 /* 562 * Break if some data was read. 563 */ 564 if (nread > 0) 565 break; 566 567 /* 568 * Unlock the pipe buffer for our remaining processing. 569 * We will either break out with an error or we will 570 * sleep and relock to loop. 571 */ 572 pipeunlock(rpipe); 573 574 /* 575 * Handle non-blocking mode operation or 576 * wait for more data. 577 */ 578 if (fp->f_flag & FNONBLOCK) { 579 error = EAGAIN; 580 } else { 581 rpipe->pipe_state |= PIPE_WANTR; 582 if ((error = msleep(rpipe, PIPE_MTX(rpipe), 583 PRIBIO | PCATCH, 584 "piperd", 0)) == 0) 585 error = pipelock(rpipe, 1); 586 } 587 if (error) 588 goto unlocked_error; 589 } 590 } 591 #ifdef MAC 592 locked_error: 593 #endif 594 pipeunlock(rpipe); 595 596 /* XXX: should probably do this before getting any locks. */ 597 if (error == 0) 598 vfs_timestamp(&rpipe->pipe_atime); 599 unlocked_error: 600 --rpipe->pipe_busy; 601 602 /* 603 * PIPE_WANT processing only makes sense if pipe_busy is 0. 604 */ 605 if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) { 606 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW); 607 wakeup(rpipe); 608 } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) { 609 /* 610 * Handle write blocking hysteresis. 611 */ 612 if (rpipe->pipe_state & PIPE_WANTW) { 613 rpipe->pipe_state &= ~PIPE_WANTW; 614 wakeup(rpipe); 615 } 616 } 617 618 if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF) 619 pipeselwakeup(rpipe); 620 621 PIPE_UNLOCK(rpipe); 622 return (error); 623 } 624 625 #ifndef PIPE_NODIRECT 626 /* 627 * Map the sending processes' buffer into kernel space and wire it. 628 * This is similar to a physical write operation. 629 */ 630 static int 631 pipe_build_write_buffer(wpipe, uio) 632 struct pipe *wpipe; 633 struct uio *uio; 634 { 635 pmap_t pmap; 636 u_int size; 637 int i, j; 638 vm_offset_t addr, endaddr; 639 640 PIPE_LOCK_ASSERT(wpipe, MA_NOTOWNED); 641 642 size = (u_int) uio->uio_iov->iov_len; 643 if (size > wpipe->pipe_buffer.size) 644 size = wpipe->pipe_buffer.size; 645 646 pmap = vmspace_pmap(curproc->p_vmspace); 647 endaddr = round_page((vm_offset_t)uio->uio_iov->iov_base + size); 648 addr = trunc_page((vm_offset_t)uio->uio_iov->iov_base); 649 for (i = 0; addr < endaddr; addr += PAGE_SIZE, i++) { 650 /* 651 * vm_fault_quick() can sleep. Consequently, 652 * vm_page_lock_queue() and vm_page_unlock_queue() 653 * should not be performed outside of this loop. 654 */ 655 race: 656 if (vm_fault_quick((caddr_t)addr, VM_PROT_READ) < 0) { 657 vm_page_lock_queues(); 658 for (j = 0; j < i; j++) 659 vm_page_unhold(wpipe->pipe_map.ms[j]); 660 vm_page_unlock_queues(); 661 return (EFAULT); 662 } 663 wpipe->pipe_map.ms[i] = pmap_extract_and_hold(pmap, addr, 664 VM_PROT_READ); 665 if (wpipe->pipe_map.ms[i] == NULL) 666 goto race; 667 } 668 669 /* 670 * set up the control block 671 */ 672 wpipe->pipe_map.npages = i; 673 wpipe->pipe_map.pos = 674 ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK; 675 wpipe->pipe_map.cnt = size; 676 677 /* 678 * and map the buffer 679 */ 680 if (wpipe->pipe_map.kva == 0) { 681 /* 682 * We need to allocate space for an extra page because the 683 * address range might (will) span pages at times. 684 */ 685 wpipe->pipe_map.kva = kmem_alloc_nofault(kernel_map, 686 wpipe->pipe_buffer.size + PAGE_SIZE); 687 atomic_add_int(&amountpipekvawired, 688 wpipe->pipe_buffer.size + PAGE_SIZE); 689 } 690 pmap_qenter(wpipe->pipe_map.kva, wpipe->pipe_map.ms, 691 wpipe->pipe_map.npages); 692 693 /* 694 * and update the uio data 695 */ 696 697 uio->uio_iov->iov_len -= size; 698 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size; 699 if (uio->uio_iov->iov_len == 0) 700 uio->uio_iov++; 701 uio->uio_resid -= size; 702 uio->uio_offset += size; 703 return (0); 704 } 705 706 /* 707 * unmap and unwire the process buffer 708 */ 709 static void 710 pipe_destroy_write_buffer(wpipe) 711 struct pipe *wpipe; 712 { 713 int i; 714 715 PIPE_LOCK_ASSERT(wpipe, MA_NOTOWNED); 716 if (wpipe->pipe_map.kva) { 717 pmap_qremove(wpipe->pipe_map.kva, wpipe->pipe_map.npages); 718 719 if (amountpipekvawired > maxpipekvawired / 2) { 720 /* Conserve address space */ 721 vm_offset_t kva = wpipe->pipe_map.kva; 722 wpipe->pipe_map.kva = 0; 723 kmem_free(kernel_map, kva, 724 wpipe->pipe_buffer.size + PAGE_SIZE); 725 atomic_subtract_int(&amountpipekvawired, 726 wpipe->pipe_buffer.size + PAGE_SIZE); 727 } 728 } 729 vm_page_lock_queues(); 730 for (i = 0; i < wpipe->pipe_map.npages; i++) { 731 vm_page_unhold(wpipe->pipe_map.ms[i]); 732 } 733 vm_page_unlock_queues(); 734 wpipe->pipe_map.npages = 0; 735 } 736 737 /* 738 * In the case of a signal, the writing process might go away. This 739 * code copies the data into the circular buffer so that the source 740 * pages can be freed without loss of data. 741 */ 742 static void 743 pipe_clone_write_buffer(wpipe) 744 struct pipe *wpipe; 745 { 746 int size; 747 int pos; 748 749 PIPE_LOCK_ASSERT(wpipe, MA_OWNED); 750 size = wpipe->pipe_map.cnt; 751 pos = wpipe->pipe_map.pos; 752 753 wpipe->pipe_buffer.in = size; 754 wpipe->pipe_buffer.out = 0; 755 wpipe->pipe_buffer.cnt = size; 756 wpipe->pipe_state &= ~PIPE_DIRECTW; 757 758 PIPE_UNLOCK(wpipe); 759 bcopy((caddr_t) wpipe->pipe_map.kva + pos, 760 wpipe->pipe_buffer.buffer, size); 761 pipe_destroy_write_buffer(wpipe); 762 PIPE_LOCK(wpipe); 763 } 764 765 /* 766 * This implements the pipe buffer write mechanism. Note that only 767 * a direct write OR a normal pipe write can be pending at any given time. 768 * If there are any characters in the pipe buffer, the direct write will 769 * be deferred until the receiving process grabs all of the bytes from 770 * the pipe buffer. Then the direct mapping write is set-up. 771 */ 772 static int 773 pipe_direct_write(wpipe, uio) 774 struct pipe *wpipe; 775 struct uio *uio; 776 { 777 int error; 778 779 retry: 780 PIPE_LOCK_ASSERT(wpipe, MA_OWNED); 781 while (wpipe->pipe_state & PIPE_DIRECTW) { 782 if (wpipe->pipe_state & PIPE_WANTR) { 783 wpipe->pipe_state &= ~PIPE_WANTR; 784 wakeup(wpipe); 785 } 786 wpipe->pipe_state |= PIPE_WANTW; 787 error = msleep(wpipe, PIPE_MTX(wpipe), 788 PRIBIO | PCATCH, "pipdww", 0); 789 if (error) 790 goto error1; 791 if (wpipe->pipe_state & PIPE_EOF) { 792 error = EPIPE; 793 goto error1; 794 } 795 } 796 wpipe->pipe_map.cnt = 0; /* transfer not ready yet */ 797 if (wpipe->pipe_buffer.cnt > 0) { 798 if (wpipe->pipe_state & PIPE_WANTR) { 799 wpipe->pipe_state &= ~PIPE_WANTR; 800 wakeup(wpipe); 801 } 802 803 wpipe->pipe_state |= PIPE_WANTW; 804 error = msleep(wpipe, PIPE_MTX(wpipe), 805 PRIBIO | PCATCH, "pipdwc", 0); 806 if (error) 807 goto error1; 808 if (wpipe->pipe_state & PIPE_EOF) { 809 error = EPIPE; 810 goto error1; 811 } 812 goto retry; 813 } 814 815 wpipe->pipe_state |= PIPE_DIRECTW; 816 817 pipelock(wpipe, 0); 818 PIPE_UNLOCK(wpipe); 819 error = pipe_build_write_buffer(wpipe, uio); 820 PIPE_LOCK(wpipe); 821 pipeunlock(wpipe); 822 if (error) { 823 wpipe->pipe_state &= ~PIPE_DIRECTW; 824 goto error1; 825 } 826 827 error = 0; 828 while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) { 829 if (wpipe->pipe_state & PIPE_EOF) { 830 pipelock(wpipe, 0); 831 PIPE_UNLOCK(wpipe); 832 pipe_destroy_write_buffer(wpipe); 833 PIPE_LOCK(wpipe); 834 pipeselwakeup(wpipe); 835 pipeunlock(wpipe); 836 error = EPIPE; 837 goto error1; 838 } 839 if (wpipe->pipe_state & PIPE_WANTR) { 840 wpipe->pipe_state &= ~PIPE_WANTR; 841 wakeup(wpipe); 842 } 843 pipeselwakeup(wpipe); 844 error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH, 845 "pipdwt", 0); 846 } 847 848 pipelock(wpipe,0); 849 if (wpipe->pipe_state & PIPE_DIRECTW) { 850 /* 851 * this bit of trickery substitutes a kernel buffer for 852 * the process that might be going away. 853 */ 854 pipe_clone_write_buffer(wpipe); 855 } else { 856 PIPE_UNLOCK(wpipe); 857 pipe_destroy_write_buffer(wpipe); 858 PIPE_LOCK(wpipe); 859 } 860 pipeunlock(wpipe); 861 return (error); 862 863 error1: 864 wakeup(wpipe); 865 return (error); 866 } 867 #endif 868 869 static int 870 pipe_write(fp, uio, active_cred, flags, td) 871 struct file *fp; 872 struct uio *uio; 873 struct ucred *active_cred; 874 struct thread *td; 875 int flags; 876 { 877 int error = 0; 878 int orig_resid; 879 struct pipe *wpipe, *rpipe; 880 881 rpipe = fp->f_data; 882 wpipe = rpipe->pipe_peer; 883 884 PIPE_LOCK(rpipe); 885 /* 886 * detect loss of pipe read side, issue SIGPIPE if lost. 887 */ 888 if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) { 889 PIPE_UNLOCK(rpipe); 890 return (EPIPE); 891 } 892 #ifdef MAC 893 error = mac_check_pipe_write(active_cred, wpipe); 894 if (error) { 895 PIPE_UNLOCK(rpipe); 896 return (error); 897 } 898 #endif 899 ++wpipe->pipe_busy; 900 901 /* 902 * If it is advantageous to resize the pipe buffer, do 903 * so. 904 */ 905 if ((uio->uio_resid > PIPE_SIZE) && 906 (amountpipekva < maxpipekva / 2) && 907 (nbigpipe < LIMITBIGPIPES) && 908 (wpipe->pipe_state & PIPE_DIRECTW) == 0 && 909 (wpipe->pipe_buffer.size <= PIPE_SIZE) && 910 (wpipe->pipe_buffer.cnt == 0)) { 911 912 if ((error = pipelock(wpipe, 1)) == 0) { 913 PIPE_UNLOCK(wpipe); 914 if (pipespace(wpipe, BIG_PIPE_SIZE) == 0) 915 atomic_add_int(&nbigpipe, 1); 916 PIPE_LOCK(wpipe); 917 pipeunlock(wpipe); 918 } 919 } 920 921 /* 922 * If an early error occured unbusy and return, waking up any pending 923 * readers. 924 */ 925 if (error) { 926 --wpipe->pipe_busy; 927 if ((wpipe->pipe_busy == 0) && 928 (wpipe->pipe_state & PIPE_WANT)) { 929 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR); 930 wakeup(wpipe); 931 } 932 PIPE_UNLOCK(rpipe); 933 return(error); 934 } 935 936 orig_resid = uio->uio_resid; 937 938 while (uio->uio_resid) { 939 int space; 940 941 #ifndef PIPE_NODIRECT 942 /* 943 * If the transfer is large, we can gain performance if 944 * we do process-to-process copies directly. 945 * If the write is non-blocking, we don't use the 946 * direct write mechanism. 947 * 948 * The direct write mechanism will detect the reader going 949 * away on us. 950 */ 951 if ((uio->uio_iov->iov_len >= PIPE_MINDIRECT) && 952 (fp->f_flag & FNONBLOCK) == 0 && 953 amountpipekvawired + uio->uio_resid < maxpipekvawired) { 954 error = pipe_direct_write(wpipe, uio); 955 if (error) 956 break; 957 continue; 958 } 959 #endif 960 961 /* 962 * Pipe buffered writes cannot be coincidental with 963 * direct writes. We wait until the currently executing 964 * direct write is completed before we start filling the 965 * pipe buffer. We break out if a signal occurs or the 966 * reader goes away. 967 */ 968 retrywrite: 969 while (wpipe->pipe_state & PIPE_DIRECTW) { 970 if (wpipe->pipe_state & PIPE_WANTR) { 971 wpipe->pipe_state &= ~PIPE_WANTR; 972 wakeup(wpipe); 973 } 974 error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH, 975 "pipbww", 0); 976 if (wpipe->pipe_state & PIPE_EOF) 977 break; 978 if (error) 979 break; 980 } 981 if (wpipe->pipe_state & PIPE_EOF) { 982 error = EPIPE; 983 break; 984 } 985 986 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt; 987 988 /* Writes of size <= PIPE_BUF must be atomic. */ 989 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF)) 990 space = 0; 991 992 if (space > 0) { 993 if ((error = pipelock(wpipe,1)) == 0) { 994 int size; /* Transfer size */ 995 int segsize; /* first segment to transfer */ 996 997 /* 998 * It is possible for a direct write to 999 * slip in on us... handle it here... 1000 */ 1001 if (wpipe->pipe_state & PIPE_DIRECTW) { 1002 pipeunlock(wpipe); 1003 goto retrywrite; 1004 } 1005 /* 1006 * If a process blocked in uiomove, our 1007 * value for space might be bad. 1008 * 1009 * XXX will we be ok if the reader has gone 1010 * away here? 1011 */ 1012 if (space > wpipe->pipe_buffer.size - 1013 wpipe->pipe_buffer.cnt) { 1014 pipeunlock(wpipe); 1015 goto retrywrite; 1016 } 1017 1018 /* 1019 * Transfer size is minimum of uio transfer 1020 * and free space in pipe buffer. 1021 */ 1022 if (space > uio->uio_resid) 1023 size = uio->uio_resid; 1024 else 1025 size = space; 1026 /* 1027 * First segment to transfer is minimum of 1028 * transfer size and contiguous space in 1029 * pipe buffer. If first segment to transfer 1030 * is less than the transfer size, we've got 1031 * a wraparound in the buffer. 1032 */ 1033 segsize = wpipe->pipe_buffer.size - 1034 wpipe->pipe_buffer.in; 1035 if (segsize > size) 1036 segsize = size; 1037 1038 /* Transfer first segment */ 1039 1040 PIPE_UNLOCK(rpipe); 1041 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in], 1042 segsize, uio); 1043 PIPE_LOCK(rpipe); 1044 1045 if (error == 0 && segsize < size) { 1046 /* 1047 * Transfer remaining part now, to 1048 * support atomic writes. Wraparound 1049 * happened. 1050 */ 1051 if (wpipe->pipe_buffer.in + segsize != 1052 wpipe->pipe_buffer.size) 1053 panic("Expected pipe buffer " 1054 "wraparound disappeared"); 1055 1056 PIPE_UNLOCK(rpipe); 1057 error = uiomove( 1058 &wpipe->pipe_buffer.buffer[0], 1059 size - segsize, uio); 1060 PIPE_LOCK(rpipe); 1061 } 1062 if (error == 0) { 1063 wpipe->pipe_buffer.in += size; 1064 if (wpipe->pipe_buffer.in >= 1065 wpipe->pipe_buffer.size) { 1066 if (wpipe->pipe_buffer.in != 1067 size - segsize + 1068 wpipe->pipe_buffer.size) 1069 panic("Expected " 1070 "wraparound bad"); 1071 wpipe->pipe_buffer.in = size - 1072 segsize; 1073 } 1074 1075 wpipe->pipe_buffer.cnt += size; 1076 if (wpipe->pipe_buffer.cnt > 1077 wpipe->pipe_buffer.size) 1078 panic("Pipe buffer overflow"); 1079 1080 } 1081 pipeunlock(wpipe); 1082 } 1083 if (error) 1084 break; 1085 1086 } else { 1087 /* 1088 * If the "read-side" has been blocked, wake it up now. 1089 */ 1090 if (wpipe->pipe_state & PIPE_WANTR) { 1091 wpipe->pipe_state &= ~PIPE_WANTR; 1092 wakeup(wpipe); 1093 } 1094 1095 /* 1096 * don't block on non-blocking I/O 1097 */ 1098 if (fp->f_flag & FNONBLOCK) { 1099 error = EAGAIN; 1100 break; 1101 } 1102 1103 /* 1104 * We have no more space and have something to offer, 1105 * wake up select/poll. 1106 */ 1107 pipeselwakeup(wpipe); 1108 1109 wpipe->pipe_state |= PIPE_WANTW; 1110 error = msleep(wpipe, PIPE_MTX(rpipe), 1111 PRIBIO | PCATCH, "pipewr", 0); 1112 if (error != 0) 1113 break; 1114 /* 1115 * If read side wants to go away, we just issue a signal 1116 * to ourselves. 1117 */ 1118 if (wpipe->pipe_state & PIPE_EOF) { 1119 error = EPIPE; 1120 break; 1121 } 1122 } 1123 } 1124 1125 --wpipe->pipe_busy; 1126 1127 if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) { 1128 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR); 1129 wakeup(wpipe); 1130 } else if (wpipe->pipe_buffer.cnt > 0) { 1131 /* 1132 * If we have put any characters in the buffer, we wake up 1133 * the reader. 1134 */ 1135 if (wpipe->pipe_state & PIPE_WANTR) { 1136 wpipe->pipe_state &= ~PIPE_WANTR; 1137 wakeup(wpipe); 1138 } 1139 } 1140 1141 /* 1142 * Don't return EPIPE if I/O was successful 1143 */ 1144 if ((wpipe->pipe_buffer.cnt == 0) && 1145 (uio->uio_resid == 0) && 1146 (error == EPIPE)) { 1147 error = 0; 1148 } 1149 1150 if (error == 0) 1151 vfs_timestamp(&wpipe->pipe_mtime); 1152 1153 /* 1154 * We have something to offer, 1155 * wake up select/poll. 1156 */ 1157 if (wpipe->pipe_buffer.cnt) 1158 pipeselwakeup(wpipe); 1159 1160 PIPE_UNLOCK(rpipe); 1161 return (error); 1162 } 1163 1164 /* 1165 * we implement a very minimal set of ioctls for compatibility with sockets. 1166 */ 1167 static int 1168 pipe_ioctl(fp, cmd, data, active_cred, td) 1169 struct file *fp; 1170 u_long cmd; 1171 void *data; 1172 struct ucred *active_cred; 1173 struct thread *td; 1174 { 1175 struct pipe *mpipe = fp->f_data; 1176 #ifdef MAC 1177 int error; 1178 #endif 1179 1180 PIPE_LOCK(mpipe); 1181 1182 #ifdef MAC 1183 error = mac_check_pipe_ioctl(active_cred, mpipe, cmd, data); 1184 if (error) { 1185 PIPE_UNLOCK(mpipe); 1186 return (error); 1187 } 1188 #endif 1189 1190 switch (cmd) { 1191 1192 case FIONBIO: 1193 PIPE_UNLOCK(mpipe); 1194 return (0); 1195 1196 case FIOASYNC: 1197 if (*(int *)data) { 1198 mpipe->pipe_state |= PIPE_ASYNC; 1199 } else { 1200 mpipe->pipe_state &= ~PIPE_ASYNC; 1201 } 1202 PIPE_UNLOCK(mpipe); 1203 return (0); 1204 1205 case FIONREAD: 1206 if (mpipe->pipe_state & PIPE_DIRECTW) 1207 *(int *)data = mpipe->pipe_map.cnt; 1208 else 1209 *(int *)data = mpipe->pipe_buffer.cnt; 1210 PIPE_UNLOCK(mpipe); 1211 return (0); 1212 1213 case FIOSETOWN: 1214 PIPE_UNLOCK(mpipe); 1215 return (fsetown(*(int *)data, &mpipe->pipe_sigio)); 1216 1217 case FIOGETOWN: 1218 PIPE_UNLOCK(mpipe); 1219 *(int *)data = fgetown(&mpipe->pipe_sigio); 1220 return (0); 1221 1222 /* This is deprecated, FIOSETOWN should be used instead. */ 1223 case TIOCSPGRP: 1224 PIPE_UNLOCK(mpipe); 1225 return (fsetown(-(*(int *)data), &mpipe->pipe_sigio)); 1226 1227 /* This is deprecated, FIOGETOWN should be used instead. */ 1228 case TIOCGPGRP: 1229 PIPE_UNLOCK(mpipe); 1230 *(int *)data = -fgetown(&mpipe->pipe_sigio); 1231 return (0); 1232 1233 } 1234 PIPE_UNLOCK(mpipe); 1235 return (ENOTTY); 1236 } 1237 1238 static int 1239 pipe_poll(fp, events, active_cred, td) 1240 struct file *fp; 1241 int events; 1242 struct ucred *active_cred; 1243 struct thread *td; 1244 { 1245 struct pipe *rpipe = fp->f_data; 1246 struct pipe *wpipe; 1247 int revents = 0; 1248 #ifdef MAC 1249 int error; 1250 #endif 1251 1252 wpipe = rpipe->pipe_peer; 1253 PIPE_LOCK(rpipe); 1254 #ifdef MAC 1255 error = mac_check_pipe_poll(active_cred, rpipe); 1256 if (error) 1257 goto locked_error; 1258 #endif 1259 if (events & (POLLIN | POLLRDNORM)) 1260 if ((rpipe->pipe_state & PIPE_DIRECTW) || 1261 (rpipe->pipe_buffer.cnt > 0) || 1262 (rpipe->pipe_state & PIPE_EOF)) 1263 revents |= events & (POLLIN | POLLRDNORM); 1264 1265 if (events & (POLLOUT | POLLWRNORM)) 1266 if (wpipe == NULL || (wpipe->pipe_state & PIPE_EOF) || 1267 (((wpipe->pipe_state & PIPE_DIRECTW) == 0) && 1268 (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF)) 1269 revents |= events & (POLLOUT | POLLWRNORM); 1270 1271 if ((rpipe->pipe_state & PIPE_EOF) || 1272 (wpipe == NULL) || 1273 (wpipe->pipe_state & PIPE_EOF)) 1274 revents |= POLLHUP; 1275 1276 if (revents == 0) { 1277 if (events & (POLLIN | POLLRDNORM)) { 1278 selrecord(td, &rpipe->pipe_sel); 1279 rpipe->pipe_state |= PIPE_SEL; 1280 } 1281 1282 if (events & (POLLOUT | POLLWRNORM)) { 1283 selrecord(td, &wpipe->pipe_sel); 1284 wpipe->pipe_state |= PIPE_SEL; 1285 } 1286 } 1287 #ifdef MAC 1288 locked_error: 1289 #endif 1290 PIPE_UNLOCK(rpipe); 1291 1292 return (revents); 1293 } 1294 1295 /* 1296 * We shouldn't need locks here as we're doing a read and this should 1297 * be a natural race. 1298 */ 1299 static int 1300 pipe_stat(fp, ub, active_cred, td) 1301 struct file *fp; 1302 struct stat *ub; 1303 struct ucred *active_cred; 1304 struct thread *td; 1305 { 1306 struct pipe *pipe = fp->f_data; 1307 #ifdef MAC 1308 int error; 1309 1310 PIPE_LOCK(pipe); 1311 error = mac_check_pipe_stat(active_cred, pipe); 1312 PIPE_UNLOCK(pipe); 1313 if (error) 1314 return (error); 1315 #endif 1316 bzero(ub, sizeof(*ub)); 1317 ub->st_mode = S_IFIFO; 1318 ub->st_blksize = pipe->pipe_buffer.size; 1319 ub->st_size = pipe->pipe_buffer.cnt; 1320 ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize; 1321 ub->st_atimespec = pipe->pipe_atime; 1322 ub->st_mtimespec = pipe->pipe_mtime; 1323 ub->st_ctimespec = pipe->pipe_ctime; 1324 ub->st_uid = fp->f_cred->cr_uid; 1325 ub->st_gid = fp->f_cred->cr_gid; 1326 /* 1327 * Left as 0: st_dev, st_ino, st_nlink, st_rdev, st_flags, st_gen. 1328 * XXX (st_dev, st_ino) should be unique. 1329 */ 1330 return (0); 1331 } 1332 1333 /* ARGSUSED */ 1334 static int 1335 pipe_close(fp, td) 1336 struct file *fp; 1337 struct thread *td; 1338 { 1339 struct pipe *cpipe = fp->f_data; 1340 1341 fp->f_ops = &badfileops; 1342 fp->f_data = NULL; 1343 funsetown(&cpipe->pipe_sigio); 1344 pipeclose(cpipe); 1345 return (0); 1346 } 1347 1348 static void 1349 pipe_free_kmem(cpipe) 1350 struct pipe *cpipe; 1351 { 1352 1353 KASSERT(cpipe->pipe_mtxp == NULL || !mtx_owned(PIPE_MTX(cpipe)), 1354 ("pipespace: pipe mutex locked")); 1355 1356 if (cpipe->pipe_buffer.buffer != NULL) { 1357 if (cpipe->pipe_buffer.size > PIPE_SIZE) 1358 atomic_subtract_int(&nbigpipe, 1); 1359 atomic_subtract_int(&amountpipekva, cpipe->pipe_buffer.size); 1360 atomic_subtract_int(&amountpipes, 1); 1361 vm_map_remove(pipe_map, 1362 (vm_offset_t)cpipe->pipe_buffer.buffer, 1363 (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size); 1364 cpipe->pipe_buffer.buffer = NULL; 1365 } 1366 #ifndef PIPE_NODIRECT 1367 if (cpipe->pipe_map.kva != 0) { 1368 atomic_subtract_int(&amountpipekvawired, 1369 cpipe->pipe_buffer.size + PAGE_SIZE); 1370 kmem_free(kernel_map, 1371 cpipe->pipe_map.kva, 1372 cpipe->pipe_buffer.size + PAGE_SIZE); 1373 cpipe->pipe_map.cnt = 0; 1374 cpipe->pipe_map.kva = 0; 1375 cpipe->pipe_map.pos = 0; 1376 cpipe->pipe_map.npages = 0; 1377 } 1378 #endif 1379 } 1380 1381 /* 1382 * shutdown the pipe 1383 */ 1384 static void 1385 pipeclose(cpipe) 1386 struct pipe *cpipe; 1387 { 1388 struct pipe *ppipe; 1389 int hadpeer; 1390 1391 if (cpipe == NULL) 1392 return; 1393 1394 hadpeer = 0; 1395 1396 /* partially created pipes won't have a valid mutex. */ 1397 if (PIPE_MTX(cpipe) != NULL) 1398 PIPE_LOCK(cpipe); 1399 1400 pipeselwakeup(cpipe); 1401 1402 /* 1403 * If the other side is blocked, wake it up saying that 1404 * we want to close it down. 1405 */ 1406 while (cpipe->pipe_busy) { 1407 wakeup(cpipe); 1408 cpipe->pipe_state |= PIPE_WANT | PIPE_EOF; 1409 msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0); 1410 } 1411 1412 #ifdef MAC 1413 if (cpipe->pipe_label != NULL && cpipe->pipe_peer == NULL) 1414 mac_destroy_pipe(cpipe); 1415 #endif 1416 1417 /* 1418 * Disconnect from peer 1419 */ 1420 if ((ppipe = cpipe->pipe_peer) != NULL) { 1421 hadpeer++; 1422 pipeselwakeup(ppipe); 1423 1424 ppipe->pipe_state |= PIPE_EOF; 1425 wakeup(ppipe); 1426 KNOTE(&ppipe->pipe_sel.si_note, 0); 1427 ppipe->pipe_peer = NULL; 1428 } 1429 /* 1430 * free resources 1431 */ 1432 if (PIPE_MTX(cpipe) != NULL) { 1433 PIPE_UNLOCK(cpipe); 1434 if (!hadpeer) { 1435 mtx_destroy(PIPE_MTX(cpipe)); 1436 free(PIPE_MTX(cpipe), M_TEMP); 1437 } 1438 } 1439 pipe_free_kmem(cpipe); 1440 uma_zfree(pipe_zone, cpipe); 1441 } 1442 1443 /*ARGSUSED*/ 1444 static int 1445 pipe_kqfilter(struct file *fp, struct knote *kn) 1446 { 1447 struct pipe *cpipe; 1448 1449 cpipe = kn->kn_fp->f_data; 1450 switch (kn->kn_filter) { 1451 case EVFILT_READ: 1452 kn->kn_fop = &pipe_rfiltops; 1453 break; 1454 case EVFILT_WRITE: 1455 kn->kn_fop = &pipe_wfiltops; 1456 cpipe = cpipe->pipe_peer; 1457 if (cpipe == NULL) 1458 /* other end of pipe has been closed */ 1459 return (EPIPE); 1460 break; 1461 default: 1462 return (1); 1463 } 1464 1465 PIPE_LOCK(cpipe); 1466 SLIST_INSERT_HEAD(&cpipe->pipe_sel.si_note, kn, kn_selnext); 1467 PIPE_UNLOCK(cpipe); 1468 return (0); 1469 } 1470 1471 static void 1472 filt_pipedetach(struct knote *kn) 1473 { 1474 struct pipe *cpipe = (struct pipe *)kn->kn_fp->f_data; 1475 1476 if (kn->kn_filter == EVFILT_WRITE) { 1477 if (cpipe->pipe_peer == NULL) 1478 return; 1479 cpipe = cpipe->pipe_peer; 1480 } 1481 1482 PIPE_LOCK(cpipe); 1483 SLIST_REMOVE(&cpipe->pipe_sel.si_note, kn, knote, kn_selnext); 1484 PIPE_UNLOCK(cpipe); 1485 } 1486 1487 /*ARGSUSED*/ 1488 static int 1489 filt_piperead(struct knote *kn, long hint) 1490 { 1491 struct pipe *rpipe = kn->kn_fp->f_data; 1492 struct pipe *wpipe = rpipe->pipe_peer; 1493 1494 PIPE_LOCK(rpipe); 1495 kn->kn_data = rpipe->pipe_buffer.cnt; 1496 if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW)) 1497 kn->kn_data = rpipe->pipe_map.cnt; 1498 1499 if ((rpipe->pipe_state & PIPE_EOF) || 1500 (wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) { 1501 kn->kn_flags |= EV_EOF; 1502 PIPE_UNLOCK(rpipe); 1503 return (1); 1504 } 1505 PIPE_UNLOCK(rpipe); 1506 return (kn->kn_data > 0); 1507 } 1508 1509 /*ARGSUSED*/ 1510 static int 1511 filt_pipewrite(struct knote *kn, long hint) 1512 { 1513 struct pipe *rpipe = kn->kn_fp->f_data; 1514 struct pipe *wpipe = rpipe->pipe_peer; 1515 1516 PIPE_LOCK(rpipe); 1517 if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) { 1518 kn->kn_data = 0; 1519 kn->kn_flags |= EV_EOF; 1520 PIPE_UNLOCK(rpipe); 1521 return (1); 1522 } 1523 kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt; 1524 if (wpipe->pipe_state & PIPE_DIRECTW) 1525 kn->kn_data = 0; 1526 1527 PIPE_UNLOCK(rpipe); 1528 return (kn->kn_data >= PIPE_BUF); 1529 } 1530