sys_pipe.c revision 76754
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: head/sys/kern/sys_pipe.c 76754 2001-05-17 17:59:28Z alfred $ 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/fcntl.h> 55#include <sys/file.h> 56#include <sys/filedesc.h> 57#include <sys/filio.h> 58#include <sys/lock.h> 59#include <sys/ttycom.h> 60#include <sys/stat.h> 61#include <sys/poll.h> 62#include <sys/selinfo.h> 63#include <sys/signalvar.h> 64#include <sys/sysproto.h> 65#include <sys/pipe.h> 66#include <sys/proc.h> 67#include <sys/vnode.h> 68#include <sys/uio.h> 69#include <sys/event.h> 70 71#include <vm/vm.h> 72#include <vm/vm_param.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 */ 91static int pipe_read __P((struct file *fp, struct uio *uio, 92 struct ucred *cred, int flags, struct proc *p)); 93static int pipe_write __P((struct file *fp, struct uio *uio, 94 struct ucred *cred, int flags, struct proc *p)); 95static int pipe_close __P((struct file *fp, struct proc *p)); 96static int pipe_poll __P((struct file *fp, int events, struct ucred *cred, 97 struct proc *p)); 98static int pipe_kqfilter __P((struct file *fp, struct knote *kn)); 99static int pipe_stat __P((struct file *fp, struct stat *sb, struct proc *p)); 100static int pipe_ioctl __P((struct file *fp, u_long cmd, caddr_t data, struct proc *p)); 101 102static struct fileops pipeops = { 103 pipe_read, pipe_write, pipe_ioctl, pipe_poll, pipe_kqfilter, 104 pipe_stat, pipe_close 105}; 106 107static void filt_pipedetach(struct knote *kn); 108static int filt_piperead(struct knote *kn, long hint); 109static int filt_pipewrite(struct knote *kn, long hint); 110 111static struct filterops pipe_rfiltops = 112 { 1, NULL, filt_pipedetach, filt_piperead }; 113static 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 142static int nbigpipe; 143 144static int amountpipekva; 145 146static void pipeclose __P((struct pipe *cpipe)); 147static void pipe_free_kmem __P((struct pipe *cpipe)); 148static int pipe_create __P((struct pipe **cpipep)); 149static __inline int pipelock __P((struct pipe *cpipe, int catch)); 150static __inline void pipeunlock __P((struct pipe *cpipe)); 151static __inline void pipeselwakeup __P((struct pipe *cpipe)); 152#ifndef PIPE_NODIRECT 153static int pipe_build_write_buffer __P((struct pipe *wpipe, struct uio *uio)); 154static void pipe_destroy_write_buffer __P((struct pipe *wpipe)); 155static int pipe_direct_write __P((struct pipe *wpipe, struct uio *uio)); 156static void pipe_clone_write_buffer __P((struct pipe *wpipe)); 157#endif 158static int pipespace __P((struct pipe *cpipe, int size)); 159 160static vm_zone_t pipe_zone; 161 162/* 163 * The pipe system call for the DTYPE_PIPE type of pipes 164 */ 165 166/* ARGSUSED */ 167int 168pipe(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 if (pipe_create(&rpipe) || pipe_create(&wpipe)) { 183 pipeclose(rpipe); 184 pipeclose(wpipe); 185 return (ENFILE); 186 } 187 188 rpipe->pipe_state |= PIPE_DIRECTOK; 189 wpipe->pipe_state |= PIPE_DIRECTOK; 190 191 error = falloc(p, &rf, &fd); 192 if (error) { 193 pipeclose(rpipe); 194 pipeclose(wpipe); 195 return (error); 196 } 197 fhold(rf); 198 p->p_retval[0] = fd; 199 200 /* 201 * Warning: once we've gotten past allocation of the fd for the 202 * read-side, we can only drop the read side via fdrop() in order 203 * to avoid races against processes which manage to dup() the read 204 * side while we are blocked trying to allocate the write side. 205 */ 206 rf->f_flag = FREAD | FWRITE; 207 rf->f_type = DTYPE_PIPE; 208 rf->f_data = (caddr_t)rpipe; 209 rf->f_ops = &pipeops; 210 error = falloc(p, &wf, &fd); 211 if (error) { 212 if (fdp->fd_ofiles[p->p_retval[0]] == rf) { 213 fdp->fd_ofiles[p->p_retval[0]] = NULL; 214 fdrop(rf, p); 215 } 216 fdrop(rf, p); 217 /* rpipe has been closed by fdrop(). */ 218 pipeclose(wpipe); 219 return (error); 220 } 221 wf->f_flag = FREAD | FWRITE; 222 wf->f_type = DTYPE_PIPE; 223 wf->f_data = (caddr_t)wpipe; 224 wf->f_ops = &pipeops; 225 p->p_retval[1] = fd; 226 227 rpipe->pipe_peer = wpipe; 228 wpipe->pipe_peer = rpipe; 229 fdrop(rf, p); 230 231 return (0); 232} 233 234/* 235 * Allocate kva for pipe circular buffer, the space is pageable 236 * This routine will 'realloc' the size of a pipe safely, if it fails 237 * it will retain the old buffer. 238 * If it fails it will return ENOMEM. 239 */ 240static int 241pipespace(cpipe, size) 242 struct pipe *cpipe; 243 int size; 244{ 245 struct vm_object *object; 246 caddr_t buffer; 247 int npages, error; 248 249 npages = round_page(size)/PAGE_SIZE; 250 /* 251 * Create an object, I don't like the idea of paging to/from 252 * kernel_object. 253 * XXX -- minor change needed here for NetBSD/OpenBSD VM systems. 254 */ 255 object = vm_object_allocate(OBJT_DEFAULT, npages); 256 buffer = (caddr_t) vm_map_min(kernel_map); 257 258 /* 259 * Insert the object into the kernel map, and allocate kva for it. 260 * The map entry is, by default, pageable. 261 * XXX -- minor change needed here for NetBSD/OpenBSD VM systems. 262 */ 263 error = vm_map_find(kernel_map, object, 0, 264 (vm_offset_t *) &buffer, size, 1, 265 VM_PROT_ALL, VM_PROT_ALL, 0); 266 267 if (error != KERN_SUCCESS) { 268 vm_object_deallocate(object); 269 return (ENOMEM); 270 } 271 272 /* free old resources if we're resizing */ 273 pipe_free_kmem(cpipe); 274 cpipe->pipe_buffer.object = object; 275 cpipe->pipe_buffer.buffer = buffer; 276 cpipe->pipe_buffer.size = size; 277 cpipe->pipe_buffer.in = 0; 278 cpipe->pipe_buffer.out = 0; 279 cpipe->pipe_buffer.cnt = 0; 280 amountpipekva += cpipe->pipe_buffer.size; 281 return (0); 282} 283 284/* 285 * initialize and allocate VM and memory for pipe 286 */ 287static int 288pipe_create(cpipep) 289 struct pipe **cpipep; 290{ 291 struct pipe *cpipe; 292 int error; 293 294 *cpipep = zalloc(pipe_zone); 295 if (*cpipep == NULL) 296 return (ENOMEM); 297 298 cpipe = *cpipep; 299 300 /* so pipespace()->pipe_free_kmem() doesn't follow junk pointer */ 301 cpipe->pipe_buffer.object = NULL; 302#ifndef PIPE_NODIRECT 303 cpipe->pipe_map.kva = NULL; 304#endif 305 /* 306 * protect so pipeclose() doesn't follow a junk pointer 307 * if pipespace() fails. 308 */ 309 bzero(&cpipe->pipe_sel, sizeof(cpipe->pipe_sel)); 310 cpipe->pipe_state = 0; 311 cpipe->pipe_peer = NULL; 312 cpipe->pipe_busy = 0; 313 314#ifndef PIPE_NODIRECT 315 /* 316 * pipe data structure initializations to support direct pipe I/O 317 */ 318 cpipe->pipe_map.cnt = 0; 319 cpipe->pipe_map.kva = 0; 320 cpipe->pipe_map.pos = 0; 321 cpipe->pipe_map.npages = 0; 322 /* cpipe->pipe_map.ms[] = invalid */ 323#endif 324 325 error = pipespace(cpipe, PIPE_SIZE); 326 if (error) { 327 return (error); 328 } 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 */ 341static __inline int 342pipelock(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 if ((error = tsleep( cpipe, 351 catch?(PRIBIO|PCATCH):PRIBIO, "pipelk", 0)) != 0) { 352 return error; 353 } 354 } 355 cpipe->pipe_state |= PIPE_LOCK; 356 return 0; 357} 358 359/* 360 * unlock a pipe I/O lock 361 */ 362static __inline void 363pipeunlock(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 374static __inline void 375pipeselwakeup(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 */ 389static int 390pipe_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 + rpipe->pipe_map.pos; 450 error = uiomove(va, size, uio); 451 if (error) 452 break; 453 nread += size; 454 rpipe->pipe_map.pos += size; 455 rpipe->pipe_map.cnt -= size; 456 if (rpipe->pipe_map.cnt == 0) { 457 rpipe->pipe_state &= ~PIPE_DIRECTW; 458 wakeup(rpipe); 459 } 460#endif 461 } else { 462 /* 463 * detect EOF condition 464 */ 465 if (rpipe->pipe_state & PIPE_EOF) { 466 /* XXX error = ? */ 467 break; 468 } 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, "piperd", 0)) == 0) 500 error = pipelock(rpipe, 1); 501 } 502 if (error) 503 goto unlocked_error; 504 } 505 } 506 pipeunlock(rpipe); 507 508 if (error == 0) 509 vfs_timestamp(&rpipe->pipe_atime); 510unlocked_error: 511 --rpipe->pipe_busy; 512 513 /* 514 * PIPE_WANT processing only makes sense if pipe_busy is 0. 515 */ 516 if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) { 517 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW); 518 wakeup(rpipe); 519 } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) { 520 /* 521 * Handle write blocking hysteresis. 522 */ 523 if (rpipe->pipe_state & PIPE_WANTW) { 524 rpipe->pipe_state &= ~PIPE_WANTW; 525 wakeup(rpipe); 526 } 527 } 528 529 if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF) 530 pipeselwakeup(rpipe); 531 532 return error; 533} 534 535#ifndef PIPE_NODIRECT 536/* 537 * Map the sending processes' buffer into kernel space and wire it. 538 * This is similar to a physical write operation. 539 */ 540static int 541pipe_build_write_buffer(wpipe, uio) 542 struct pipe *wpipe; 543 struct uio *uio; 544{ 545 u_int size; 546 int i; 547 vm_offset_t addr, endaddr, paddr; 548 549 size = (u_int) uio->uio_iov->iov_len; 550 if (size > wpipe->pipe_buffer.size) 551 size = wpipe->pipe_buffer.size; 552 553 endaddr = round_page((vm_offset_t)uio->uio_iov->iov_base + size); 554 for(i = 0, addr = trunc_page((vm_offset_t)uio->uio_iov->iov_base); 555 addr < endaddr; 556 addr += PAGE_SIZE, i+=1) { 557 558 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 for(j=0;j<i;j++) 564 vm_page_unwire(wpipe->pipe_map.ms[j], 1); 565 return EFAULT; 566 } 567 568 m = PHYS_TO_VM_PAGE(paddr); 569 vm_page_wire(m); 570 wpipe->pipe_map.ms[i] = m; 571 } 572 573/* 574 * set up the control block 575 */ 576 wpipe->pipe_map.npages = i; 577 wpipe->pipe_map.pos = ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK; 578 wpipe->pipe_map.cnt = size; 579 580/* 581 * and map the buffer 582 */ 583 if (wpipe->pipe_map.kva == 0) { 584 /* 585 * We need to allocate space for an extra page because the 586 * address range might (will) span pages at times. 587 */ 588 wpipe->pipe_map.kva = kmem_alloc_pageable(kernel_map, 589 wpipe->pipe_buffer.size + PAGE_SIZE); 590 amountpipekva += wpipe->pipe_buffer.size + PAGE_SIZE; 591 } 592 pmap_qenter(wpipe->pipe_map.kva, wpipe->pipe_map.ms, 593 wpipe->pipe_map.npages); 594 595/* 596 * and update the uio data 597 */ 598 599 uio->uio_iov->iov_len -= size; 600 uio->uio_iov->iov_base += size; 601 if (uio->uio_iov->iov_len == 0) 602 uio->uio_iov++; 603 uio->uio_resid -= size; 604 uio->uio_offset += size; 605 return 0; 606} 607 608/* 609 * unmap and unwire the process buffer 610 */ 611static void 612pipe_destroy_write_buffer(wpipe) 613struct pipe *wpipe; 614{ 615 int i; 616 617 if (wpipe->pipe_map.kva) { 618 pmap_qremove(wpipe->pipe_map.kva, wpipe->pipe_map.npages); 619 620 if (amountpipekva > MAXPIPEKVA) { 621 vm_offset_t kva = wpipe->pipe_map.kva; 622 wpipe->pipe_map.kva = 0; 623 kmem_free(kernel_map, kva, 624 wpipe->pipe_buffer.size + PAGE_SIZE); 625 amountpipekva -= wpipe->pipe_buffer.size + PAGE_SIZE; 626 } 627 } 628 for (i=0;i<wpipe->pipe_map.npages;i++) 629 vm_page_unwire(wpipe->pipe_map.ms[i], 1); 630} 631 632/* 633 * In the case of a signal, the writing process might go away. This 634 * code copies the data into the circular buffer so that the source 635 * pages can be freed without loss of data. 636 */ 637static void 638pipe_clone_write_buffer(wpipe) 639 struct pipe *wpipe; 640{ 641 int size; 642 int pos; 643 644 size = wpipe->pipe_map.cnt; 645 pos = wpipe->pipe_map.pos; 646 bcopy((caddr_t) wpipe->pipe_map.kva+pos, 647 (caddr_t) wpipe->pipe_buffer.buffer, 648 size); 649 650 wpipe->pipe_buffer.in = size; 651 wpipe->pipe_buffer.out = 0; 652 wpipe->pipe_buffer.cnt = size; 653 wpipe->pipe_state &= ~PIPE_DIRECTW; 654 655 pipe_destroy_write_buffer(wpipe); 656} 657 658/* 659 * This implements the pipe buffer write mechanism. Note that only 660 * a direct write OR a normal pipe write can be pending at any given time. 661 * If there are any characters in the pipe buffer, the direct write will 662 * be deferred until the receiving process grabs all of the bytes from 663 * the pipe buffer. Then the direct mapping write is set-up. 664 */ 665static int 666pipe_direct_write(wpipe, uio) 667 struct pipe *wpipe; 668 struct uio *uio; 669{ 670 int error; 671 672retry: 673 while (wpipe->pipe_state & PIPE_DIRECTW) { 674 if ( wpipe->pipe_state & PIPE_WANTR) { 675 wpipe->pipe_state &= ~PIPE_WANTR; 676 wakeup(wpipe); 677 } 678 wpipe->pipe_state |= PIPE_WANTW; 679 error = tsleep(wpipe, 680 PRIBIO|PCATCH, "pipdww", 0); 681 if (error) 682 goto error1; 683 if (wpipe->pipe_state & PIPE_EOF) { 684 error = EPIPE; 685 goto error1; 686 } 687 } 688 wpipe->pipe_map.cnt = 0; /* transfer not ready yet */ 689 if (wpipe->pipe_buffer.cnt > 0) { 690 if ( wpipe->pipe_state & PIPE_WANTR) { 691 wpipe->pipe_state &= ~PIPE_WANTR; 692 wakeup(wpipe); 693 } 694 695 wpipe->pipe_state |= PIPE_WANTW; 696 error = tsleep(wpipe, 697 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 746error1: 747 wakeup(wpipe); 748 return error; 749} 750#endif 751 752static int 753pipe_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 } 773 774 /* 775 * If it is advantageous to resize the pipe buffer, do 776 * so. 777 */ 778 if ((uio->uio_resid > PIPE_SIZE) && 779 (nbigpipe < LIMITBIGPIPES) && 780 (wpipe->pipe_state & PIPE_DIRECTW) == 0 && 781 (wpipe->pipe_buffer.size <= PIPE_SIZE) && 782 (wpipe->pipe_buffer.cnt == 0)) { 783 784 if ((error = pipelock(wpipe,1)) == 0) { 785 if (pipespace(wpipe, BIG_PIPE_SIZE) == 0) 786 nbigpipe++; 787 pipeunlock(wpipe); 788 } else { 789 return error; 790 } 791 } 792 793 KASSERT(wpipe->pipe_buffer.buffer != NULL, ("pipe buffer gone")); 794 795 ++wpipe->pipe_busy; 796 orig_resid = uio->uio_resid; 797 while (uio->uio_resid) { 798 int space; 799#ifndef PIPE_NODIRECT 800 /* 801 * If the transfer is large, we can gain performance if 802 * we do process-to-process copies directly. 803 * If the write is non-blocking, we don't use the 804 * direct write mechanism. 805 * 806 * The direct write mechanism will detect the reader going 807 * away on us. 808 */ 809 if ((uio->uio_iov->iov_len >= PIPE_MINDIRECT) && 810 (fp->f_flag & FNONBLOCK) == 0 && 811 (wpipe->pipe_map.kva || (amountpipekva < LIMITPIPEKVA)) && 812 (uio->uio_iov->iov_len >= PIPE_MINDIRECT)) { 813 error = pipe_direct_write( wpipe, uio); 814 if (error) { 815 break; 816 } 817 continue; 818 } 819#endif 820 821 /* 822 * Pipe buffered writes cannot be coincidental with 823 * direct writes. We wait until the currently executing 824 * direct write is completed before we start filling the 825 * pipe buffer. We break out if a signal occurs or the 826 * reader goes away. 827 */ 828 retrywrite: 829 while (wpipe->pipe_state & PIPE_DIRECTW) { 830 if (wpipe->pipe_state & PIPE_WANTR) { 831 wpipe->pipe_state &= ~PIPE_WANTR; 832 wakeup(wpipe); 833 } 834 error = tsleep(wpipe, PRIBIO|PCATCH, "pipbww", 0); 835 if (wpipe->pipe_state & PIPE_EOF) 836 break; 837 if (error) 838 break; 839 } 840 if (wpipe->pipe_state & PIPE_EOF) { 841 error = EPIPE; 842 break; 843 } 844 845 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt; 846 847 /* Writes of size <= PIPE_BUF must be atomic. */ 848 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF)) 849 space = 0; 850 851 if (space > 0 && (wpipe->pipe_buffer.cnt < PIPE_SIZE)) { 852 if ((error = pipelock(wpipe,1)) == 0) { 853 int size; /* Transfer size */ 854 int segsize; /* first segment to transfer */ 855 /* 856 * It is possible for a direct write to 857 * slip in on us... handle it here... 858 */ 859 if (wpipe->pipe_state & PIPE_DIRECTW) { 860 pipeunlock(wpipe); 861 goto retrywrite; 862 } 863 /* 864 * If a process blocked in uiomove, our 865 * value for space might be bad. 866 * 867 * XXX will we be ok if the reader has gone 868 * away here? 869 */ 870 if (space > wpipe->pipe_buffer.size - 871 wpipe->pipe_buffer.cnt) { 872 pipeunlock(wpipe); 873 goto retrywrite; 874 } 875 876 /* 877 * Transfer size is minimum of uio transfer 878 * and free space in pipe buffer. 879 */ 880 if (space > uio->uio_resid) 881 size = uio->uio_resid; 882 else 883 size = space; 884 /* 885 * First segment to transfer is minimum of 886 * transfer size and contiguous space in 887 * pipe buffer. If first segment to transfer 888 * is less than the transfer size, we've got 889 * a wraparound in the buffer. 890 */ 891 segsize = wpipe->pipe_buffer.size - 892 wpipe->pipe_buffer.in; 893 if (segsize > size) 894 segsize = size; 895 896 /* Transfer first segment */ 897 898 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in], 899 segsize, uio); 900 901 if (error == 0 && segsize < size) { 902 /* 903 * Transfer remaining part now, to 904 * support atomic writes. Wraparound 905 * happened. 906 */ 907 if (wpipe->pipe_buffer.in + segsize != 908 wpipe->pipe_buffer.size) 909 panic("Expected pipe buffer wraparound disappeared"); 910 911 error = uiomove(&wpipe->pipe_buffer.buffer[0], 912 size - segsize, uio); 913 } 914 if (error == 0) { 915 wpipe->pipe_buffer.in += size; 916 if (wpipe->pipe_buffer.in >= 917 wpipe->pipe_buffer.size) { 918 if (wpipe->pipe_buffer.in != size - segsize + wpipe->pipe_buffer.size) 919 panic("Expected wraparound bad"); 920 wpipe->pipe_buffer.in = size - segsize; 921 } 922 923 wpipe->pipe_buffer.cnt += size; 924 if (wpipe->pipe_buffer.cnt > wpipe->pipe_buffer.size) 925 panic("Pipe buffer overflow"); 926 927 } 928 pipeunlock(wpipe); 929 } 930 if (error) 931 break; 932 933 } else { 934 /* 935 * If the "read-side" has been blocked, wake it up now. 936 */ 937 if (wpipe->pipe_state & PIPE_WANTR) { 938 wpipe->pipe_state &= ~PIPE_WANTR; 939 wakeup(wpipe); 940 } 941 942 /* 943 * don't block on non-blocking I/O 944 */ 945 if (fp->f_flag & FNONBLOCK) { 946 error = EAGAIN; 947 break; 948 } 949 950 /* 951 * We have no more space and have something to offer, 952 * wake up select/poll. 953 */ 954 pipeselwakeup(wpipe); 955 956 wpipe->pipe_state |= PIPE_WANTW; 957 if ((error = tsleep(wpipe, (PRIBIO+1)|PCATCH, "pipewr", 0)) != 0) { 958 break; 959 } 960 /* 961 * If read side wants to go away, we just issue a signal 962 * to ourselves. 963 */ 964 if (wpipe->pipe_state & PIPE_EOF) { 965 error = EPIPE; 966 break; 967 } 968 } 969 } 970 971 --wpipe->pipe_busy; 972 if ((wpipe->pipe_busy == 0) && 973 (wpipe->pipe_state & PIPE_WANT)) { 974 wpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTR); 975 wakeup(wpipe); 976 } else if (wpipe->pipe_buffer.cnt > 0) { 977 /* 978 * If we have put any characters in the buffer, we wake up 979 * the reader. 980 */ 981 if (wpipe->pipe_state & PIPE_WANTR) { 982 wpipe->pipe_state &= ~PIPE_WANTR; 983 wakeup(wpipe); 984 } 985 } 986 987 /* 988 * Don't return EPIPE if I/O was successful 989 */ 990 if ((wpipe->pipe_buffer.cnt == 0) && 991 (uio->uio_resid == 0) && 992 (error == EPIPE)) 993 error = 0; 994 995 if (error == 0) 996 vfs_timestamp(&wpipe->pipe_mtime); 997 998 /* 999 * We have something to offer, 1000 * wake up select/poll. 1001 */ 1002 if (wpipe->pipe_buffer.cnt) 1003 pipeselwakeup(wpipe); 1004 1005 return error; 1006} 1007 1008/* 1009 * we implement a very minimal set of ioctls for compatibility with sockets. 1010 */ 1011int 1012pipe_ioctl(fp, cmd, data, p) 1013 struct file *fp; 1014 u_long cmd; 1015 caddr_t data; 1016 struct proc *p; 1017{ 1018 struct pipe *mpipe = (struct pipe *)fp->f_data; 1019 1020 switch (cmd) { 1021 1022 case FIONBIO: 1023 return (0); 1024 1025 case FIOASYNC: 1026 if (*(int *)data) { 1027 mpipe->pipe_state |= PIPE_ASYNC; 1028 } else { 1029 mpipe->pipe_state &= ~PIPE_ASYNC; 1030 } 1031 return (0); 1032 1033 case FIONREAD: 1034 if (mpipe->pipe_state & PIPE_DIRECTW) 1035 *(int *)data = mpipe->pipe_map.cnt; 1036 else 1037 *(int *)data = mpipe->pipe_buffer.cnt; 1038 return (0); 1039 1040 case FIOSETOWN: 1041 return (fsetown(*(int *)data, &mpipe->pipe_sigio)); 1042 1043 case FIOGETOWN: 1044 *(int *)data = fgetown(mpipe->pipe_sigio); 1045 return (0); 1046 1047 /* This is deprecated, FIOSETOWN should be used instead. */ 1048 case TIOCSPGRP: 1049 return (fsetown(-(*(int *)data), &mpipe->pipe_sigio)); 1050 1051 /* This is deprecated, FIOGETOWN should be used instead. */ 1052 case TIOCGPGRP: 1053 *(int *)data = -fgetown(mpipe->pipe_sigio); 1054 return (0); 1055 1056 } 1057 return (ENOTTY); 1058} 1059 1060int 1061pipe_poll(fp, events, cred, p) 1062 struct file *fp; 1063 int events; 1064 struct ucred *cred; 1065 struct proc *p; 1066{ 1067 struct pipe *rpipe = (struct pipe *)fp->f_data; 1068 struct pipe *wpipe; 1069 int revents = 0; 1070 1071 wpipe = rpipe->pipe_peer; 1072 if (events & (POLLIN | POLLRDNORM)) 1073 if ((rpipe->pipe_state & PIPE_DIRECTW) || 1074 (rpipe->pipe_buffer.cnt > 0) || 1075 (rpipe->pipe_state & PIPE_EOF)) 1076 revents |= events & (POLLIN | POLLRDNORM); 1077 1078 if (events & (POLLOUT | POLLWRNORM)) 1079 if (wpipe == NULL || (wpipe->pipe_state & PIPE_EOF) || 1080 (((wpipe->pipe_state & PIPE_DIRECTW) == 0) && 1081 (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF)) 1082 revents |= events & (POLLOUT | POLLWRNORM); 1083 1084 if ((rpipe->pipe_state & PIPE_EOF) || 1085 (wpipe == NULL) || 1086 (wpipe->pipe_state & PIPE_EOF)) 1087 revents |= POLLHUP; 1088 1089 if (revents == 0) { 1090 if (events & (POLLIN | POLLRDNORM)) { 1091 selrecord(p, &rpipe->pipe_sel); 1092 rpipe->pipe_state |= PIPE_SEL; 1093 } 1094 1095 if (events & (POLLOUT | POLLWRNORM)) { 1096 selrecord(p, &wpipe->pipe_sel); 1097 wpipe->pipe_state |= PIPE_SEL; 1098 } 1099 } 1100 1101 return (revents); 1102} 1103 1104static int 1105pipe_stat(fp, ub, p) 1106 struct file *fp; 1107 struct stat *ub; 1108 struct proc *p; 1109{ 1110 struct pipe *pipe = (struct pipe *)fp->f_data; 1111 1112 bzero((caddr_t)ub, sizeof (*ub)); 1113 ub->st_mode = S_IFIFO; 1114 ub->st_blksize = pipe->pipe_buffer.size; 1115 ub->st_size = pipe->pipe_buffer.cnt; 1116 ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize; 1117 ub->st_atimespec = pipe->pipe_atime; 1118 ub->st_mtimespec = pipe->pipe_mtime; 1119 ub->st_ctimespec = pipe->pipe_ctime; 1120 ub->st_uid = fp->f_cred->cr_uid; 1121 ub->st_gid = fp->f_cred->cr_gid; 1122 /* 1123 * Left as 0: st_dev, st_ino, st_nlink, st_rdev, st_flags, st_gen. 1124 * XXX (st_dev, st_ino) should be unique. 1125 */ 1126 return 0; 1127} 1128 1129/* ARGSUSED */ 1130static int 1131pipe_close(fp, p) 1132 struct file *fp; 1133 struct proc *p; 1134{ 1135 struct pipe *cpipe = (struct pipe *)fp->f_data; 1136 1137 fp->f_ops = &badfileops; 1138 fp->f_data = NULL; 1139 funsetown(cpipe->pipe_sigio); 1140 pipeclose(cpipe); 1141 return 0; 1142} 1143 1144static void 1145pipe_free_kmem(cpipe) 1146 struct pipe *cpipe; 1147{ 1148 1149 if (cpipe->pipe_buffer.buffer != NULL) { 1150 if (cpipe->pipe_buffer.size > PIPE_SIZE) 1151 --nbigpipe; 1152 amountpipekva -= cpipe->pipe_buffer.size; 1153 kmem_free(kernel_map, 1154 (vm_offset_t)cpipe->pipe_buffer.buffer, 1155 cpipe->pipe_buffer.size); 1156 cpipe->pipe_buffer.buffer = NULL; 1157 } 1158#ifndef PIPE_NODIRECT 1159 if (cpipe->pipe_map.kva != NULL) { 1160 amountpipekva -= cpipe->pipe_buffer.size + PAGE_SIZE; 1161 kmem_free(kernel_map, 1162 cpipe->pipe_map.kva, 1163 cpipe->pipe_buffer.size + PAGE_SIZE); 1164 cpipe->pipe_map.cnt = 0; 1165 cpipe->pipe_map.kva = 0; 1166 cpipe->pipe_map.pos = 0; 1167 cpipe->pipe_map.npages = 0; 1168 } 1169#endif 1170} 1171 1172/* 1173 * shutdown the pipe 1174 */ 1175static void 1176pipeclose(cpipe) 1177 struct pipe *cpipe; 1178{ 1179 struct pipe *ppipe; 1180 1181 if (cpipe) { 1182 1183 pipeselwakeup(cpipe); 1184 1185 /* 1186 * If the other side is blocked, wake it up saying that 1187 * we want to close it down. 1188 */ 1189 while (cpipe->pipe_busy) { 1190 wakeup(cpipe); 1191 cpipe->pipe_state |= PIPE_WANT|PIPE_EOF; 1192 tsleep(cpipe, PRIBIO, "pipecl", 0); 1193 } 1194 1195 /* 1196 * Disconnect from peer 1197 */ 1198 if ((ppipe = cpipe->pipe_peer) != NULL) { 1199 pipeselwakeup(ppipe); 1200 1201 ppipe->pipe_state |= PIPE_EOF; 1202 wakeup(ppipe); 1203 ppipe->pipe_peer = NULL; 1204 } 1205 1206 /* 1207 * free resources 1208 */ 1209 pipe_free_kmem(cpipe); 1210 zfree(pipe_zone, cpipe); 1211 } 1212} 1213 1214/*ARGSUSED*/ 1215static int 1216pipe_kqfilter(struct file *fp, struct knote *kn) 1217{ 1218 struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data; 1219 1220 switch (kn->kn_filter) { 1221 case EVFILT_READ: 1222 kn->kn_fop = &pipe_rfiltops; 1223 break; 1224 case EVFILT_WRITE: 1225 kn->kn_fop = &pipe_wfiltops; 1226 break; 1227 default: 1228 return (1); 1229 } 1230 1231 SLIST_INSERT_HEAD(&rpipe->pipe_sel.si_note, kn, kn_selnext); 1232 return (0); 1233} 1234 1235static void 1236filt_pipedetach(struct knote *kn) 1237{ 1238 struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data; 1239 1240 SLIST_REMOVE(&rpipe->pipe_sel.si_note, kn, knote, kn_selnext); 1241} 1242 1243/*ARGSUSED*/ 1244static int 1245filt_piperead(struct knote *kn, long hint) 1246{ 1247 struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data; 1248 struct pipe *wpipe = rpipe->pipe_peer; 1249 1250 kn->kn_data = rpipe->pipe_buffer.cnt; 1251 if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW)) 1252 kn->kn_data = rpipe->pipe_map.cnt; 1253 1254 if ((rpipe->pipe_state & PIPE_EOF) || 1255 (wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) { 1256 kn->kn_flags |= EV_EOF; 1257 return (1); 1258 } 1259 return (kn->kn_data > 0); 1260} 1261 1262/*ARGSUSED*/ 1263static int 1264filt_pipewrite(struct knote *kn, long hint) 1265{ 1266 struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data; 1267 struct pipe *wpipe = rpipe->pipe_peer; 1268 1269 if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) { 1270 kn->kn_data = 0; 1271 kn->kn_flags |= EV_EOF; 1272 return (1); 1273 } 1274 kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt; 1275 if (wpipe->pipe_state & PIPE_DIRECTW) 1276 kn->kn_data = 0; 1277 1278 return (kn->kn_data >= PIPE_BUF); 1279} 1280