Now available: The Design and Implementation of the FreeBSD Operating System (Second Edition)
[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]
FreeBSD/Linux Kernel Cross Reference
sys/kern/sys_pipe.c
Version:
- FREEBSD - FREEBSD-13-STABLE - FREEBSD-13-0 - FREEBSD-12-STABLE - FREEBSD-12-0 - FREEBSD-11-STABLE - FREEBSD-11-0 - FREEBSD-10-STABLE - FREEBSD-10-0 - FREEBSD-9-STABLE - FREEBSD-9-0 - FREEBSD-8-STABLE - FREEBSD-8-0 - FREEBSD-7-STABLE - FREEBSD-7-0 - FREEBSD-6-STABLE - FREEBSD-6-0 - FREEBSD-5-STABLE - FREEBSD-5-0 - FREEBSD-4-STABLE - FREEBSD-3-STABLE - FREEBSD22 - l41 - OPENBSD - linux-2.6 - MK84 - PLAN9 - xnu-8792
SearchContext: - none - 3 - 10
SearchContext: - none - 3 - 10
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 * $FreeBSD$ 20 */ 21 22 /* 23 * This file contains a high-performance replacement for the socket-based 24 * pipes scheme originally used in FreeBSD/4.4Lite. It does not support 25 * all features of sockets, but does do everything that pipes normally 26 * do. 27 */ 28 29 /* 30 * This code has two modes of operation, a small write mode and a large 31 * write mode. The small write mode acts like conventional pipes with 32 * a kernel buffer. If the buffer is less than PIPE_MINDIRECT, then the 33 * "normal" pipe buffering is done. If the buffer is between PIPE_MINDIRECT 34 * and PIPE_SIZE in size, it is fully mapped and wired into the kernel, and 35 * the receiving process can copy it directly from the pages in the sending 36 * process. 37 * 38 * If the sending process receives a signal, it is possible that it will 39 * go away, and certainly its address space can change, because control 40 * is returned back to the user-mode side. In that case, the pipe code 41 * arranges to copy the buffer supplied by the user process, to a pageable 42 * kernel buffer, and the receiving process will grab the data from the 43 * pageable kernel buffer. Since signals don't happen all that often, 44 * the copy operation is normally eliminated. 45 * 46 * The constant PIPE_MINDIRECT is chosen to make sure that buffering will 47 * happen for small transfers so that the system will not spend all of 48 * its time context switching. PIPE_SIZE is constrained by the 49 * amount of kernel virtual memory. 50 */ 51 52 #include <sys/param.h> 53 #include <sys/systm.h> 54 #include <sys/proc.h> 55 #include <sys/fcntl.h> 56 #include <sys/file.h> 57 #include <sys/filedesc.h> 58 #include <sys/filio.h> 59 #include <sys/ttycom.h> 60 #include <sys/stat.h> 61 #include <sys/poll.h> 62 #include <sys/select.h> 63 #include <sys/signalvar.h> 64 #include <sys/sysproto.h> 65 #include <sys/pipe.h> 66 #include <sys/vnode.h> 67 #include <sys/uio.h> 68 #include <sys/event.h> 69 70 #include <vm/vm.h> 71 #include <vm/vm_param.h> 72 #include <sys/lock.h> 73 #include <vm/vm_object.h> 74 #include <vm/vm_kern.h> 75 #include <vm/vm_extern.h> 76 #include <vm/pmap.h> 77 #include <vm/vm_map.h> 78 #include <vm/vm_page.h> 79 #include <vm/vm_zone.h> 80 81 /* 82 * Use this define if you want to disable *fancy* VM things. Expect an 83 * approx 30% decrease in transfer rate. This could be useful for 84 * NetBSD or OpenBSD. 85 */ 86 /* #define PIPE_NODIRECT */ 87 88 /* 89 * interfaces to the outside world 90 */ 91 static int pipe_read __P((struct file *fp, struct uio *uio, 92 struct ucred *cred, int flags, struct proc *p)); 93 static int pipe_write __P((struct file *fp, struct uio *uio, 94 struct ucred *cred, int flags, struct proc *p)); 95 static int pipe_close __P((struct file *fp, struct proc *p)); 96 static int pipe_poll __P((struct file *fp, int events, struct ucred *cred, 97 struct proc *p)); 98 static int pipe_kqfilter __P((struct file *fp, struct knote *kn)); 99 static int pipe_stat __P((struct file *fp, struct stat *sb, struct proc *p)); 100 static int pipe_ioctl __P((struct file *fp, u_long cmd, caddr_t data, struct proc *p)); 101 102 static struct fileops pipeops = { 103 pipe_read, pipe_write, pipe_ioctl, pipe_poll, pipe_kqfilter, 104 pipe_stat, pipe_close 105 }; 106 107 static void filt_pipedetach(struct knote *kn); 108 static int filt_piperead(struct knote *kn, long hint); 109 static int filt_pipewrite(struct knote *kn, long hint); 110 111 static struct filterops pipe_rfiltops = 112 { 1, NULL, filt_pipedetach, filt_piperead }; 113 static struct filterops pipe_wfiltops = 114 { 1, NULL, filt_pipedetach, filt_pipewrite }; 115 116 117 /* 118 * Default pipe buffer size(s), this can be kind-of large now because pipe 119 * space is pageable. The pipe code will try to maintain locality of 120 * reference for performance reasons, so small amounts of outstanding I/O 121 * will not wipe the cache. 122 */ 123 #define MINPIPESIZE (PIPE_SIZE/3) 124 #define MAXPIPESIZE (2*PIPE_SIZE/3) 125 126 /* 127 * Maximum amount of kva for pipes -- this is kind-of a soft limit, but 128 * is there so that on large systems, we don't exhaust it. 129 */ 130 #define MAXPIPEKVA (8*1024*1024) 131 132 /* 133 * Limit for direct transfers, we cannot, of course limit 134 * the amount of kva for pipes in general though. 135 */ 136 #define LIMITPIPEKVA (16*1024*1024) 137 138 /* 139 * Limit the number of "big" pipes 140 */ 141 #define LIMITBIGPIPES 32 142 static int nbigpipe; 143 144 static int amountpipekva; 145 146 static void pipeclose __P((struct pipe *cpipe)); 147 static void pipe_free_kmem __P((struct pipe *cpipe)); 148 static int pipe_create __P((struct pipe **cpipep)); 149 static __inline int pipelock __P((struct pipe *cpipe, int catch)); 150 static __inline void pipeunlock __P((struct pipe *cpipe)); 151 static __inline void pipeselwakeup __P((struct pipe *cpipe)); 152 #ifndef PIPE_NODIRECT 153 static int pipe_build_write_buffer __P((struct pipe *wpipe, struct uio *uio)); 154 static void pipe_destroy_write_buffer __P((struct pipe *wpipe)); 155 static int pipe_direct_write __P((struct pipe *wpipe, struct uio *uio)); 156 static void pipe_clone_write_buffer __P((struct pipe *wpipe)); 157 #endif 158 static int pipespace __P((struct pipe *cpipe, int size)); 159 160 static vm_zone_t pipe_zone; 161 162 /* 163 * The pipe system call for the DTYPE_PIPE type of pipes 164 */ 165 166 /* ARGSUSED */ 167 int 168 pipe(p, uap) 169 struct proc *p; 170 struct pipe_args /* { 171 int dummy; 172 } */ *uap; 173 { 174 struct filedesc *fdp = p->p_fd; 175 struct file *rf, *wf; 176 struct pipe *rpipe, *wpipe; 177 int fd, error; 178 179 if (pipe_zone == NULL) 180 pipe_zone = zinit("PIPE", sizeof(struct pipe), 0, 0, 4); 181 182 rpipe = wpipe = NULL; 183 if (pipe_create(&rpipe) || pipe_create(&wpipe)) { 184 pipeclose(rpipe); 185 pipeclose(wpipe); 186 return (ENFILE); 187 } 188 189 rpipe->pipe_state |= PIPE_DIRECTOK; 190 wpipe->pipe_state |= PIPE_DIRECTOK; 191 192 error = falloc(p, &rf, &fd); 193 if (error) { 194 pipeclose(rpipe); 195 pipeclose(wpipe); 196 return (error); 197 } 198 fhold(rf); 199 p->p_retval[0] = fd; 200 201 /* 202 * Warning: once we've gotten past allocation of the fd for the 203 * read-side, we can only drop the read side via fdrop() in order 204 * to avoid races against processes which manage to dup() the read 205 * side while we are blocked trying to allocate the write side. 206 */ 207 rf->f_flag = FREAD | FWRITE; 208 rf->f_type = DTYPE_PIPE; 209 rf->f_data = (caddr_t)rpipe; 210 rf->f_ops = &pipeops; 211 error = falloc(p, &wf, &fd); 212 if (error) { 213 if (fdp->fd_ofiles[p->p_retval[0]] == rf) { 214 fdp->fd_ofiles[p->p_retval[0]] = NULL; 215 fdrop(rf, p); 216 } 217 fdrop(rf, p); 218 /* rpipe has been closed by fdrop(). */ 219 pipeclose(wpipe); 220 return (error); 221 } 222 wf->f_flag = FREAD | FWRITE; 223 wf->f_type = DTYPE_PIPE; 224 wf->f_data = (caddr_t)wpipe; 225 wf->f_ops = &pipeops; 226 p->p_retval[1] = fd; 227 228 rpipe->pipe_peer = wpipe; 229 wpipe->pipe_peer = rpipe; 230 fdrop(rf, p); 231 232 return (0); 233 } 234 235 /* 236 * Allocate kva for pipe circular buffer, the space is pageable 237 * This routine will 'realloc' the size of a pipe safely, if it fails 238 * it will retain the old buffer. 239 * If it fails it will return ENOMEM. 240 */ 241 static int 242 pipespace(cpipe, size) 243 struct pipe *cpipe; 244 int size; 245 { 246 struct vm_object *object; 247 caddr_t buffer; 248 int npages, error; 249 250 npages = round_page(size)/PAGE_SIZE; 251 /* 252 * Create an object, I don't like the idea of paging to/from 253 * kernel_object. 254 * XXX -- minor change needed here for NetBSD/OpenBSD VM systems. 255 */ 256 object = vm_object_allocate(OBJT_DEFAULT, npages); 257 buffer = (caddr_t) vm_map_min(kernel_map); 258 259 /* 260 * Insert the object into the kernel map, and allocate kva for it. 261 * The map entry is, by default, pageable. 262 * XXX -- minor change needed here for NetBSD/OpenBSD VM systems. 263 */ 264 error = vm_map_find(kernel_map, object, 0, 265 (vm_offset_t *) &buffer, size, 1, 266 VM_PROT_ALL, VM_PROT_ALL, 0); 267 268 if (error != KERN_SUCCESS) { 269 vm_object_deallocate(object); 270 return (ENOMEM); 271 } 272 273 /* free old resources if we're resizing */ 274 pipe_free_kmem(cpipe); 275 cpipe->pipe_buffer.object = object; 276 cpipe->pipe_buffer.buffer = buffer; 277 cpipe->pipe_buffer.size = size; 278 cpipe->pipe_buffer.in = 0; 279 cpipe->pipe_buffer.out = 0; 280 cpipe->pipe_buffer.cnt = 0; 281 amountpipekva += cpipe->pipe_buffer.size; 282 return (0); 283 } 284 285 /* 286 * initialize and allocate VM and memory for pipe 287 */ 288 static int 289 pipe_create(cpipep) 290 struct pipe **cpipep; 291 { 292 struct pipe *cpipe; 293 int error; 294 295 *cpipep = zalloc(pipe_zone); 296 if (*cpipep == NULL) 297 return (ENOMEM); 298 299 cpipe = *cpipep; 300 301 /* so pipespace()->pipe_free_kmem() doesn't follow junk pointer */ 302 cpipe->pipe_buffer.object = NULL; 303 #ifndef PIPE_NODIRECT 304 cpipe->pipe_map.kva = NULL; 305 #endif 306 /* 307 * protect so pipeclose() doesn't follow a junk pointer 308 * if pipespace() fails. 309 */ 310 bzero(&cpipe->pipe_sel, sizeof(cpipe->pipe_sel)); 311 cpipe->pipe_state = 0; 312 cpipe->pipe_peer = NULL; 313 cpipe->pipe_busy = 0; 314 315 #ifndef PIPE_NODIRECT 316 /* 317 * pipe data structure initializations to support direct pipe I/O 318 */ 319 cpipe->pipe_map.cnt = 0; 320 cpipe->pipe_map.kva = 0; 321 cpipe->pipe_map.pos = 0; 322 cpipe->pipe_map.npages = 0; 323 /* cpipe->pipe_map.ms[] = invalid */ 324 #endif 325 326 error = pipespace(cpipe, PIPE_SIZE); 327 if (error) 328 return (error); 329 330 vfs_timestamp(&cpipe->pipe_ctime); 331 cpipe->pipe_atime = cpipe->pipe_ctime; 332 cpipe->pipe_mtime = cpipe->pipe_ctime; 333 334 return (0); 335 } 336 337 338 /* 339 * lock a pipe for I/O, blocking other access 340 */ 341 static __inline int 342 pipelock(cpipe, catch) 343 struct pipe *cpipe; 344 int catch; 345 { 346 int error; 347 348 while (cpipe->pipe_state & PIPE_LOCK) { 349 cpipe->pipe_state |= PIPE_LWANT; 350 error = tsleep(cpipe, catch ? (PRIBIO | PCATCH) : PRIBIO, 351 "pipelk", 0); 352 if (error != 0) 353 return (error); 354 } 355 cpipe->pipe_state |= PIPE_LOCK; 356 return (0); 357 } 358 359 /* 360 * unlock a pipe I/O lock 361 */ 362 static __inline void 363 pipeunlock(cpipe) 364 struct pipe *cpipe; 365 { 366 367 cpipe->pipe_state &= ~PIPE_LOCK; 368 if (cpipe->pipe_state & PIPE_LWANT) { 369 cpipe->pipe_state &= ~PIPE_LWANT; 370 wakeup(cpipe); 371 } 372 } 373 374 static __inline void 375 pipeselwakeup(cpipe) 376 struct pipe *cpipe; 377 { 378 379 if (cpipe->pipe_state & PIPE_SEL) { 380 cpipe->pipe_state &= ~PIPE_SEL; 381 selwakeup(&cpipe->pipe_sel); 382 } 383 if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio) 384 pgsigio(cpipe->pipe_sigio, SIGIO, 0); 385 KNOTE(&cpipe->pipe_sel.si_note, 0); 386 } 387 388 /* ARGSUSED */ 389 static int 390 pipe_read(fp, uio, cred, flags, p) 391 struct file *fp; 392 struct uio *uio; 393 struct ucred *cred; 394 struct proc *p; 395 int flags; 396 { 397 struct pipe *rpipe = (struct pipe *) fp->f_data; 398 int error; 399 int nread = 0; 400 u_int size; 401 402 ++rpipe->pipe_busy; 403 error = pipelock(rpipe, 1); 404 if (error) 405 goto unlocked_error; 406 407 while (uio->uio_resid) { 408 /* 409 * normal pipe buffer receive 410 */ 411 if (rpipe->pipe_buffer.cnt > 0) { 412 size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out; 413 if (size > rpipe->pipe_buffer.cnt) 414 size = rpipe->pipe_buffer.cnt; 415 if (size > (u_int) uio->uio_resid) 416 size = (u_int) uio->uio_resid; 417 418 error = uiomove(&rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out], 419 size, uio); 420 if (error) 421 break; 422 423 rpipe->pipe_buffer.out += size; 424 if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size) 425 rpipe->pipe_buffer.out = 0; 426 427 rpipe->pipe_buffer.cnt -= size; 428 429 /* 430 * If there is no more to read in the pipe, reset 431 * its pointers to the beginning. This improves 432 * cache hit stats. 433 */ 434 if (rpipe->pipe_buffer.cnt == 0) { 435 rpipe->pipe_buffer.in = 0; 436 rpipe->pipe_buffer.out = 0; 437 } 438 nread += size; 439 #ifndef PIPE_NODIRECT 440 /* 441 * Direct copy, bypassing a kernel buffer. 442 */ 443 } else if ((size = rpipe->pipe_map.cnt) && 444 (rpipe->pipe_state & PIPE_DIRECTW)) { 445 caddr_t va; 446 if (size > (u_int) uio->uio_resid) 447 size = (u_int) uio->uio_resid; 448 449 va = (caddr_t) rpipe->pipe_map.kva + 450 rpipe->pipe_map.pos; 451 error = uiomove(va, size, uio); 452 if (error) 453 break; 454 nread += size; 455 rpipe->pipe_map.pos += size; 456 rpipe->pipe_map.cnt -= size; 457 if (rpipe->pipe_map.cnt == 0) { 458 rpipe->pipe_state &= ~PIPE_DIRECTW; 459 wakeup(rpipe); 460 } 461 #endif 462 } else { 463 /* 464 * detect EOF condition 465 * read returns 0 on EOF, no need to set error 466 */ 467 if (rpipe->pipe_state & PIPE_EOF) 468 break; 469 470 /* 471 * If the "write-side" has been blocked, wake it up now. 472 */ 473 if (rpipe->pipe_state & PIPE_WANTW) { 474 rpipe->pipe_state &= ~PIPE_WANTW; 475 wakeup(rpipe); 476 } 477 478 /* 479 * Break if some data was read. 480 */ 481 if (nread > 0) 482 break; 483 484 /* 485 * Unlock the pipe buffer for our remaining processing. We 486 * will either break out with an error or we will sleep and 487 * relock to loop. 488 */ 489 pipeunlock(rpipe); 490 491 /* 492 * Handle non-blocking mode operation or 493 * wait for more data. 494 */ 495 if (fp->f_flag & FNONBLOCK) { 496 error = EAGAIN; 497 } else { 498 rpipe->pipe_state |= PIPE_WANTR; 499 if ((error = tsleep(rpipe, PRIBIO | PCATCH, 500 "piperd", 0)) == 0) 501 error = pipelock(rpipe, 1); 502 } 503 if (error) 504 goto unlocked_error; 505 } 506 } 507 pipeunlock(rpipe); 508 509 if (error == 0) 510 vfs_timestamp(&rpipe->pipe_atime); 511 unlocked_error: 512 --rpipe->pipe_busy; 513 514 /* 515 * PIPE_WANT processing only makes sense if pipe_busy is 0. 516 */ 517 if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) { 518 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW); 519 wakeup(rpipe); 520 } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) { 521 /* 522 * Handle write blocking hysteresis. 523 */ 524 if (rpipe->pipe_state & PIPE_WANTW) { 525 rpipe->pipe_state &= ~PIPE_WANTW; 526 wakeup(rpipe); 527 } 528 } 529 530 if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF) 531 pipeselwakeup(rpipe); 532 533 return (error); 534 } 535 536 #ifndef PIPE_NODIRECT 537 /* 538 * Map the sending processes' buffer into kernel space and wire it. 539 * This is similar to a physical write operation. 540 */ 541 static int 542 pipe_build_write_buffer(wpipe, uio) 543 struct pipe *wpipe; 544 struct uio *uio; 545 { 546 u_int size; 547 int i; 548 vm_offset_t addr, endaddr; 549 vm_paddr_t paddr; 550 551 size = (u_int) uio->uio_iov->iov_len; 552 if (size > wpipe->pipe_buffer.size) 553 size = wpipe->pipe_buffer.size; 554 555 endaddr = round_page((vm_offset_t)uio->uio_iov->iov_base + size); 556 addr = trunc_page((vm_offset_t)uio->uio_iov->iov_base); 557 for (i = 0; addr < endaddr; addr += PAGE_SIZE, i++) { 558 vm_page_t m; 559 560 if (vm_fault_quick((caddr_t)addr, VM_PROT_READ) < 0 || 561 (paddr = pmap_kextract(addr)) == 0) { 562 int j; 563 564 for (j = 0; j < i; j++) 565 vm_page_unhold(wpipe->pipe_map.ms[j]); 566 return (EFAULT); 567 } 568 569 m = PHYS_TO_VM_PAGE(paddr); 570 vm_page_hold(m); 571 wpipe->pipe_map.ms[i] = m; 572 } 573 574 /* 575 * set up the control block 576 */ 577 wpipe->pipe_map.npages = i; 578 wpipe->pipe_map.pos = 579 ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK; 580 wpipe->pipe_map.cnt = size; 581 582 /* 583 * and map the buffer 584 */ 585 if (wpipe->pipe_map.kva == 0) { 586 /* 587 * We need to allocate space for an extra page because the 588 * address range might (will) span pages at times. 589 */ 590 wpipe->pipe_map.kva = kmem_alloc_pageable(kernel_map, 591 wpipe->pipe_buffer.size + PAGE_SIZE); 592 amountpipekva += wpipe->pipe_buffer.size + PAGE_SIZE; 593 } 594 pmap_qenter(wpipe->pipe_map.kva, wpipe->pipe_map.ms, 595 wpipe->pipe_map.npages); 596 597 /* 598 * and update the uio data 599 */ 600 601 uio->uio_iov->iov_len -= size; 602 uio->uio_iov->iov_base += size; 603 if (uio->uio_iov->iov_len == 0) 604 uio->uio_iov++; 605 uio->uio_resid -= size; 606 uio->uio_offset += size; 607 return (0); 608 } 609 610 /* 611 * unmap and unwire the process buffer 612 */ 613 static void 614 pipe_destroy_write_buffer(wpipe) 615 struct pipe *wpipe; 616 { 617 int i; 618 619 if (wpipe->pipe_map.kva) { 620 pmap_qremove(wpipe->pipe_map.kva, wpipe->pipe_map.npages); 621 622 if (amountpipekva > MAXPIPEKVA) { 623 vm_offset_t kva = wpipe->pipe_map.kva; 624 wpipe->pipe_map.kva = 0; 625 kmem_free(kernel_map, kva, 626 wpipe->pipe_buffer.size + PAGE_SIZE); 627 amountpipekva -= wpipe->pipe_buffer.size + PAGE_SIZE; 628 } 629 } 630 for (i = 0; i < wpipe->pipe_map.npages; i++) 631 vm_page_unhold(wpipe->pipe_map.ms[i]); 632 wpipe->pipe_map.npages = 0; 633 } 634 635 /* 636 * In the case of a signal, the writing process might go away. This 637 * code copies the data into the circular buffer so that the source 638 * pages can be freed without loss of data. 639 */ 640 static void 641 pipe_clone_write_buffer(wpipe) 642 struct pipe *wpipe; 643 { 644 int size; 645 int pos; 646 647 size = wpipe->pipe_map.cnt; 648 pos = wpipe->pipe_map.pos; 649 bcopy((caddr_t) wpipe->pipe_map.kva + pos, 650 (caddr_t) wpipe->pipe_buffer.buffer, size); 651 652 wpipe->pipe_buffer.in = size; 653 wpipe->pipe_buffer.out = 0; 654 wpipe->pipe_buffer.cnt = size; 655 wpipe->pipe_state &= ~PIPE_DIRECTW; 656 657 pipe_destroy_write_buffer(wpipe); 658 } 659 660 /* 661 * This implements the pipe buffer write mechanism. Note that only 662 * a direct write OR a normal pipe write can be pending at any given time. 663 * If there are any characters in the pipe buffer, the direct write will 664 * be deferred until the receiving process grabs all of the bytes from 665 * the pipe buffer. Then the direct mapping write is set-up. 666 */ 667 static int 668 pipe_direct_write(wpipe, uio) 669 struct pipe *wpipe; 670 struct uio *uio; 671 { 672 int error; 673 674 retry: 675 while (wpipe->pipe_state & PIPE_DIRECTW) { 676 if (wpipe->pipe_state & PIPE_WANTR) { 677 wpipe->pipe_state &= ~PIPE_WANTR; 678 wakeup(wpipe); 679 } 680 wpipe->pipe_state |= PIPE_WANTW; 681 error = tsleep(wpipe, PRIBIO | PCATCH, "pipdww", 0); 682 if (error) 683 goto error1; 684 if (wpipe->pipe_state & PIPE_EOF) { 685 error = EPIPE; 686 goto error1; 687 } 688 } 689 wpipe->pipe_map.cnt = 0; /* transfer not ready yet */ 690 if (wpipe->pipe_buffer.cnt > 0) { 691 if (wpipe->pipe_state & PIPE_WANTR) { 692 wpipe->pipe_state &= ~PIPE_WANTR; 693 wakeup(wpipe); 694 } 695 696 wpipe->pipe_state |= PIPE_WANTW; 697 error = tsleep(wpipe, PRIBIO | PCATCH, "pipdwc", 0); 698 if (error) 699 goto error1; 700 if (wpipe->pipe_state & PIPE_EOF) { 701 error = EPIPE; 702 goto error1; 703 } 704 goto retry; 705 } 706 707 wpipe->pipe_state |= PIPE_DIRECTW; 708 709 error = pipe_build_write_buffer(wpipe, uio); 710 if (error) { 711 wpipe->pipe_state &= ~PIPE_DIRECTW; 712 goto error1; 713 } 714 715 error = 0; 716 while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) { 717 if (wpipe->pipe_state & PIPE_EOF) { 718 pipelock(wpipe, 0); 719 pipe_destroy_write_buffer(wpipe); 720 pipeunlock(wpipe); 721 pipeselwakeup(wpipe); 722 error = EPIPE; 723 goto error1; 724 } 725 if (wpipe->pipe_state & PIPE_WANTR) { 726 wpipe->pipe_state &= ~PIPE_WANTR; 727 wakeup(wpipe); 728 } 729 pipeselwakeup(wpipe); 730 error = tsleep(wpipe, PRIBIO | PCATCH, "pipdwt", 0); 731 } 732 733 pipelock(wpipe,0); 734 if (wpipe->pipe_state & PIPE_DIRECTW) { 735 /* 736 * this bit of trickery substitutes a kernel buffer for 737 * the process that might be going away. 738 */ 739 pipe_clone_write_buffer(wpipe); 740 } else { 741 pipe_destroy_write_buffer(wpipe); 742 } 743 pipeunlock(wpipe); 744 return (error); 745 746 error1: 747 wakeup(wpipe); 748 return (error); 749 } 750 #endif 751 752 static int 753 pipe_write(fp, uio, cred, flags, p) 754 struct file *fp; 755 struct uio *uio; 756 struct ucred *cred; 757 struct proc *p; 758 int flags; 759 { 760 int error = 0; 761 int orig_resid; 762 struct pipe *wpipe, *rpipe; 763 764 rpipe = (struct pipe *) fp->f_data; 765 wpipe = rpipe->pipe_peer; 766 767 /* 768 * detect loss of pipe read side, issue SIGPIPE if lost. 769 */ 770 if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) { 771 return (EPIPE); 772 } 773 ++wpipe->pipe_busy; 774 775 /* 776 * If it is advantageous to resize the pipe buffer, do 777 * so. 778 */ 779 if ((uio->uio_resid > PIPE_SIZE) && 780 (nbigpipe < LIMITBIGPIPES) && 781 (wpipe->pipe_state & PIPE_DIRECTW) == 0 && 782 (wpipe->pipe_buffer.size <= PIPE_SIZE) && 783 (wpipe->pipe_buffer.cnt == 0)) { 784 785 if ((error = pipelock(wpipe,1)) == 0) { 786 if (pipespace(wpipe, BIG_PIPE_SIZE) == 0) 787 nbigpipe++; 788 pipeunlock(wpipe); 789 } 790 } 791 792 /* 793 * If an early error occured unbusy and return, waking up any pending 794 * readers. 795 */ 796 if (error) { 797 --wpipe->pipe_busy; 798 if ((wpipe->pipe_busy == 0) && 799 (wpipe->pipe_state & PIPE_WANT)) { 800 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR); 801 wakeup(wpipe); 802 } 803 return(error); 804 } 805 806 KASSERT(wpipe->pipe_buffer.buffer != NULL, ("pipe buffer gone")); 807 808 orig_resid = uio->uio_resid; 809 810 while (uio->uio_resid) { 811 int space; 812 813 #ifndef PIPE_NODIRECT 814 /* 815 * If the transfer is large, we can gain performance if 816 * we do process-to-process copies directly. 817 * If the write is non-blocking, we don't use the 818 * direct write mechanism. 819 * 820 * The direct write mechanism will detect the reader going 821 * away on us. 822 */ 823 if ((uio->uio_iov->iov_len >= PIPE_MINDIRECT) && 824 (fp->f_flag & FNONBLOCK) == 0 && 825 (wpipe->pipe_map.kva || (amountpipekva < LIMITPIPEKVA)) && 826 (uio->uio_iov->iov_len >= PIPE_MINDIRECT)) { 827 error = pipe_direct_write( wpipe, uio); 828 if (error) 829 break; 830 continue; 831 } 832 #endif 833 834 /* 835 * Pipe buffered writes cannot be coincidental with 836 * direct writes. We wait until the currently executing 837 * direct write is completed before we start filling the 838 * pipe buffer. We break out if a signal occurs or the 839 * reader goes away. 840 */ 841 retrywrite: 842 while (wpipe->pipe_state & PIPE_DIRECTW) { 843 if (wpipe->pipe_state & PIPE_WANTR) { 844 wpipe->pipe_state &= ~PIPE_WANTR; 845 wakeup(wpipe); 846 } 847 error = tsleep(wpipe, PRIBIO | PCATCH, "pipbww", 0); 848 if (wpipe->pipe_state & PIPE_EOF) 849 break; 850 if (error) 851 break; 852 } 853 if (wpipe->pipe_state & PIPE_EOF) { 854 error = EPIPE; 855 break; 856 } 857 858 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt; 859 860 /* Writes of size <= PIPE_BUF must be atomic. */ 861 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF)) 862 space = 0; 863 864 if (space > 0) { 865 if ((error = pipelock(wpipe,1)) == 0) { 866 int size; /* Transfer size */ 867 int segsize; /* first segment to transfer */ 868 869 /* 870 * It is possible for a direct write to 871 * slip in on us... handle it here... 872 */ 873 if (wpipe->pipe_state & PIPE_DIRECTW) { 874 pipeunlock(wpipe); 875 goto retrywrite; 876 } 877 /* 878 * If a process blocked in uiomove, our 879 * value for space might be bad. 880 * 881 * XXX will we be ok if the reader has gone 882 * away here? 883 */ 884 if (space > wpipe->pipe_buffer.size - 885 wpipe->pipe_buffer.cnt) { 886 pipeunlock(wpipe); 887 goto retrywrite; 888 } 889 890 /* 891 * Transfer size is minimum of uio transfer 892 * and free space in pipe buffer. 893 */ 894 if (space > uio->uio_resid) 895 size = uio->uio_resid; 896 else 897 size = space; 898 /* 899 * First segment to transfer is minimum of 900 * transfer size and contiguous space in 901 * pipe buffer. If first segment to transfer 902 * is less than the transfer size, we've got 903 * a wraparound in the buffer. 904 */ 905 segsize = wpipe->pipe_buffer.size - 906 wpipe->pipe_buffer.in; 907 if (segsize > size) 908 segsize = size; 909 910 /* Transfer first segment */ 911 912 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in], 913 segsize, uio); 914 915 if (error == 0 && segsize < size) { 916 /* 917 * Transfer remaining part now, to 918 * support atomic writes. Wraparound 919 * happened. 920 */ 921 if (wpipe->pipe_buffer.in + segsize != 922 wpipe->pipe_buffer.size) 923 panic("Expected pipe buffer wraparound disappeared"); 924 925 error = uiomove(&wpipe->pipe_buffer.buffer[0], 926 size - segsize, uio); 927 } 928 if (error == 0) { 929 wpipe->pipe_buffer.in += size; 930 if (wpipe->pipe_buffer.in >= 931 wpipe->pipe_buffer.size) { 932 if (wpipe->pipe_buffer.in != size - segsize + wpipe->pipe_buffer.size) 933 panic("Expected wraparound bad"); 934 wpipe->pipe_buffer.in = size - segsize; 935 } 936 937 wpipe->pipe_buffer.cnt += size; 938 if (wpipe->pipe_buffer.cnt > wpipe->pipe_buffer.size) 939 panic("Pipe buffer overflow"); 940 941 } 942 pipeunlock(wpipe); 943 } 944 if (error) 945 break; 946 947 } else { 948 /* 949 * If the "read-side" has been blocked, wake it up now. 950 */ 951 if (wpipe->pipe_state & PIPE_WANTR) { 952 wpipe->pipe_state &= ~PIPE_WANTR; 953 wakeup(wpipe); 954 } 955 956 /* 957 * don't block on non-blocking I/O 958 */ 959 if (fp->f_flag & FNONBLOCK) { 960 error = EAGAIN; 961 break; 962 } 963 964 /* 965 * We have no more space and have something to offer, 966 * wake up select/poll. 967 */ 968 pipeselwakeup(wpipe); 969 970 wpipe->pipe_state |= PIPE_WANTW; 971 error = tsleep(wpipe, PRIBIO | PCATCH, "pipewr", 0); 972 if (error != 0) 973 break; 974 /* 975 * If read side wants to go away, we just issue a signal 976 * to ourselves. 977 */ 978 if (wpipe->pipe_state & PIPE_EOF) { 979 error = EPIPE; 980 break; 981 } 982 } 983 } 984 985 --wpipe->pipe_busy; 986 987 if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) { 988 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR); 989 wakeup(wpipe); 990 } else if (wpipe->pipe_buffer.cnt > 0) { 991 /* 992 * If we have put any characters in the buffer, we wake up 993 * the reader. 994 */ 995 if (wpipe->pipe_state & PIPE_WANTR) { 996 wpipe->pipe_state &= ~PIPE_WANTR; 997 wakeup(wpipe); 998 } 999 } 1000 1001 /* 1002 * Don't return EPIPE if I/O was successful 1003 */ 1004 if ((wpipe->pipe_buffer.cnt == 0) && 1005 (uio->uio_resid == 0) && 1006 (error == EPIPE)) { 1007 error = 0; 1008 } 1009 1010 if (error == 0) 1011 vfs_timestamp(&wpipe->pipe_mtime); 1012 1013 /* 1014 * We have something to offer, 1015 * wake up select/poll. 1016 */ 1017 if (wpipe->pipe_buffer.cnt) 1018 pipeselwakeup(wpipe); 1019 1020 return (error); 1021 } 1022 1023 /* 1024 * we implement a very minimal set of ioctls for compatibility with sockets. 1025 */ 1026 int 1027 pipe_ioctl(fp, cmd, data, p) 1028 struct file *fp; 1029 u_long cmd; 1030 caddr_t data; 1031 struct proc *p; 1032 { 1033 struct pipe *mpipe = (struct pipe *)fp->f_data; 1034 1035 switch (cmd) { 1036 1037 case FIONBIO: 1038 return (0); 1039 1040 case FIOASYNC: 1041 if (*(int *)data) { 1042 mpipe->pipe_state |= PIPE_ASYNC; 1043 } else { 1044 mpipe->pipe_state &= ~PIPE_ASYNC; 1045 } 1046 return (0); 1047 1048 case FIONREAD: 1049 if (mpipe->pipe_state & PIPE_DIRECTW) 1050 *(int *)data = mpipe->pipe_map.cnt; 1051 else 1052 *(int *)data = mpipe->pipe_buffer.cnt; 1053 return (0); 1054 1055 case FIOSETOWN: 1056 return (fsetown(*(int *)data, &mpipe->pipe_sigio)); 1057 1058 case FIOGETOWN: 1059 *(int *)data = fgetown(mpipe->pipe_sigio); 1060 return (0); 1061 1062 /* This is deprecated, FIOSETOWN should be used instead. */ 1063 case TIOCSPGRP: 1064 return (fsetown(-(*(int *)data), &mpipe->pipe_sigio)); 1065 1066 /* This is deprecated, FIOGETOWN should be used instead. */ 1067 case TIOCGPGRP: 1068 *(int *)data = -fgetown(mpipe->pipe_sigio); 1069 return (0); 1070 1071 } 1072 return (ENOTTY); 1073 } 1074 1075 int 1076 pipe_poll(fp, events, cred, p) 1077 struct file *fp; 1078 int events; 1079 struct ucred *cred; 1080 struct proc *p; 1081 { 1082 struct pipe *rpipe = (struct pipe *)fp->f_data; 1083 struct pipe *wpipe; 1084 int revents = 0; 1085 1086 wpipe = rpipe->pipe_peer; 1087 if (events & (POLLIN | POLLRDNORM)) 1088 if ((rpipe->pipe_state & PIPE_DIRECTW) || 1089 (rpipe->pipe_buffer.cnt > 0) || 1090 (rpipe->pipe_state & PIPE_EOF)) 1091 revents |= events & (POLLIN | POLLRDNORM); 1092 1093 if (events & (POLLOUT | POLLWRNORM)) 1094 if (wpipe == NULL || (wpipe->pipe_state & PIPE_EOF) || 1095 (((wpipe->pipe_state & PIPE_DIRECTW) == 0) && 1096 (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF)) 1097 revents |= events & (POLLOUT | POLLWRNORM); 1098 1099 if ((rpipe->pipe_state & PIPE_EOF) || 1100 (wpipe == NULL) || 1101 (wpipe->pipe_state & PIPE_EOF)) 1102 revents |= POLLHUP; 1103 1104 if (revents == 0) { 1105 if (events & (POLLIN | POLLRDNORM)) { 1106 selrecord(p, &rpipe->pipe_sel); 1107 rpipe->pipe_state |= PIPE_SEL; 1108 } 1109 1110 if (events & (POLLOUT | POLLWRNORM)) { 1111 selrecord(p, &wpipe->pipe_sel); 1112 wpipe->pipe_state |= PIPE_SEL; 1113 } 1114 } 1115 1116 return (revents); 1117 } 1118 1119 static int 1120 pipe_stat(fp, ub, p) 1121 struct file *fp; 1122 struct stat *ub; 1123 struct proc *p; 1124 { 1125 struct pipe *pipe = (struct pipe *)fp->f_data; 1126 1127 bzero((caddr_t)ub, sizeof(*ub)); 1128 ub->st_mode = S_IFIFO; 1129 ub->st_blksize = pipe->pipe_buffer.size; 1130 ub->st_size = pipe->pipe_buffer.cnt; 1131 ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize; 1132 ub->st_atimespec = pipe->pipe_atime; 1133 ub->st_mtimespec = pipe->pipe_mtime; 1134 ub->st_ctimespec = pipe->pipe_ctime; 1135 /* 1136 * Left as 0: st_dev, st_ino, st_nlink, st_uid, st_gid, st_rdev, 1137 * st_flags, st_gen. 1138 * XXX (st_dev, st_ino) should be unique. 1139 */ 1140 return (0); 1141 } 1142 1143 /* ARGSUSED */ 1144 static int 1145 pipe_close(fp, p) 1146 struct file *fp; 1147 struct proc *p; 1148 { 1149 struct pipe *cpipe = (struct pipe *)fp->f_data; 1150 1151 fp->f_ops = &badfileops; 1152 fp->f_data = NULL; 1153 funsetown(cpipe->pipe_sigio); 1154 pipeclose(cpipe); 1155 return (0); 1156 } 1157 1158 static void 1159 pipe_free_kmem(cpipe) 1160 struct pipe *cpipe; 1161 { 1162 1163 if (cpipe->pipe_buffer.buffer != NULL) { 1164 if (cpipe->pipe_buffer.size > PIPE_SIZE) 1165 --nbigpipe; 1166 amountpipekva -= cpipe->pipe_buffer.size; 1167 kmem_free(kernel_map, 1168 (vm_offset_t)cpipe->pipe_buffer.buffer, 1169 cpipe->pipe_buffer.size); 1170 cpipe->pipe_buffer.buffer = NULL; 1171 } 1172 #ifndef PIPE_NODIRECT 1173 if (cpipe->pipe_map.kva != NULL) { 1174 amountpipekva -= cpipe->pipe_buffer.size + PAGE_SIZE; 1175 kmem_free(kernel_map, 1176 cpipe->pipe_map.kva, 1177 cpipe->pipe_buffer.size + PAGE_SIZE); 1178 cpipe->pipe_map.cnt = 0; 1179 cpipe->pipe_map.kva = 0; 1180 cpipe->pipe_map.pos = 0; 1181 cpipe->pipe_map.npages = 0; 1182 } 1183 #endif 1184 } 1185 1186 /* 1187 * shutdown the pipe 1188 */ 1189 static void 1190 pipeclose(cpipe) 1191 struct pipe *cpipe; 1192 { 1193 struct pipe *ppipe; 1194 1195 if (cpipe) { 1196 1197 pipeselwakeup(cpipe); 1198 1199 /* 1200 * If the other side is blocked, wake it up saying that 1201 * we want to close it down. 1202 */ 1203 while (cpipe->pipe_busy) { 1204 wakeup(cpipe); 1205 cpipe->pipe_state |= PIPE_WANT | PIPE_EOF; 1206 tsleep(cpipe, PRIBIO, "pipecl", 0); 1207 } 1208 1209 /* 1210 * Disconnect from peer 1211 */ 1212 if ((ppipe = cpipe->pipe_peer) != NULL) { 1213 pipeselwakeup(ppipe); 1214 1215 ppipe->pipe_state |= PIPE_EOF; 1216 wakeup(ppipe); 1217 KNOTE(&ppipe->pipe_sel.si_note, 0); 1218 ppipe->pipe_peer = NULL; 1219 } 1220 /* 1221 * free resources 1222 */ 1223 pipe_free_kmem(cpipe); 1224 zfree(pipe_zone, cpipe); 1225 } 1226 } 1227 1228 /*ARGSUSED*/ 1229 static int 1230 pipe_kqfilter(struct file *fp, struct knote *kn) 1231 { 1232 struct pipe *cpipe = (struct pipe *)kn->kn_fp->f_data; 1233 1234 switch (kn->kn_filter) { 1235 case EVFILT_READ: 1236 kn->kn_fop = &pipe_rfiltops; 1237 break; 1238 case EVFILT_WRITE: 1239 kn->kn_fop = &pipe_wfiltops; 1240 cpipe = cpipe->pipe_peer; 1241 if (cpipe == NULL) 1242 /* other end of pipe has been closed */ 1243 return (EPIPE); 1244 break; 1245 default: 1246 return (1); 1247 } 1248 1249 SLIST_INSERT_HEAD(&cpipe->pipe_sel.si_note, kn, kn_selnext); 1250 return (0); 1251 } 1252 1253 static void 1254 filt_pipedetach(struct knote *kn) 1255 { 1256 struct pipe *cpipe = (struct pipe *)kn->kn_fp->f_data; 1257 1258 if (kn->kn_filter == EVFILT_WRITE) { 1259 if (cpipe->pipe_peer == NULL) 1260 return; 1261 cpipe = cpipe->pipe_peer; 1262 } 1263 1264 SLIST_REMOVE(&cpipe->pipe_sel.si_note, kn, knote, kn_selnext); 1265 } 1266 1267 /*ARGSUSED*/ 1268 static int 1269 filt_piperead(struct knote *kn, long hint) 1270 { 1271 struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data; 1272 struct pipe *wpipe = rpipe->pipe_peer; 1273 1274 kn->kn_data = rpipe->pipe_buffer.cnt; 1275 if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW)) 1276 kn->kn_data = rpipe->pipe_map.cnt; 1277 1278 if ((rpipe->pipe_state & PIPE_EOF) || 1279 (wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) { 1280 kn->kn_flags |= EV_EOF; 1281 return (1); 1282 } 1283 return (kn->kn_data > 0); 1284 } 1285 1286 /*ARGSUSED*/ 1287 static int 1288 filt_pipewrite(struct knote *kn, long hint) 1289 { 1290 struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data; 1291 struct pipe *wpipe = rpipe->pipe_peer; 1292 1293 if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) { 1294 kn->kn_data = 0; 1295 kn->kn_flags |= EV_EOF; 1296 return (1); 1297 } 1298 kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt; 1299 if (wpipe->pipe_state & PIPE_DIRECTW) 1300 kn->kn_data = 0; 1301 1302 return (kn->kn_data >= PIPE_BUF); 1303 }
Cache object: 96066b36d8fdb095c1db76859e5d3333
[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]