sys_pipe.c revision 76756
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 76756 2001-05-17 18:22:58Z 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 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 */ 241static int 242pipespace(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 */ 288static int 289pipe_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 } 330 331 vfs_timestamp(&cpipe->pipe_ctime); 332 cpipe->pipe_atime = cpipe->pipe_ctime; 333 cpipe->pipe_mtime = cpipe->pipe_ctime; 334 335 return (0); 336} 337 338 339/* 340 * lock a pipe for I/O, blocking other access 341 */ 342static __inline int 343pipelock(cpipe, catch) 344 struct pipe *cpipe; 345 int catch; 346{ 347 int error; 348 349 while (cpipe->pipe_state & PIPE_LOCK) { 350 cpipe->pipe_state |= PIPE_LWANT; 351 if ((error = tsleep( cpipe, 352 catch?(PRIBIO|PCATCH):PRIBIO, "pipelk", 0)) != 0) { 353 return error; 354 } 355 } 356 cpipe->pipe_state |= PIPE_LOCK; 357 return 0; 358} 359 360/* 361 * unlock a pipe I/O lock 362 */ 363static __inline void 364pipeunlock(cpipe) 365 struct pipe *cpipe; 366{ 367 368 cpipe->pipe_state &= ~PIPE_LOCK; 369 if (cpipe->pipe_state & PIPE_LWANT) { 370 cpipe->pipe_state &= ~PIPE_LWANT; 371 wakeup(cpipe); 372 } 373} 374 375static __inline void 376pipeselwakeup(cpipe) 377 struct pipe *cpipe; 378{ 379 380 if (cpipe->pipe_state & PIPE_SEL) { 381 cpipe->pipe_state &= ~PIPE_SEL; 382 selwakeup(&cpipe->pipe_sel); 383 } 384 if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio) 385 pgsigio(cpipe->pipe_sigio, SIGIO, 0); 386 KNOTE(&cpipe->pipe_sel.si_note, 0); 387} 388 389/* ARGSUSED */ 390static int 391pipe_read(fp, uio, cred, flags, p) 392 struct file *fp; 393 struct uio *uio; 394 struct ucred *cred; 395 struct proc *p; 396 int flags; 397{ 398 struct pipe *rpipe = (struct pipe *) fp->f_data; 399 int error; 400 int nread = 0; 401 u_int size; 402 403 ++rpipe->pipe_busy; 404 error = pipelock(rpipe, 1); 405 if (error) 406 goto unlocked_error; 407 408 while (uio->uio_resid) { 409 /* 410 * normal pipe buffer receive 411 */ 412 if (rpipe->pipe_buffer.cnt > 0) { 413 size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out; 414 if (size > rpipe->pipe_buffer.cnt) 415 size = rpipe->pipe_buffer.cnt; 416 if (size > (u_int) uio->uio_resid) 417 size = (u_int) uio->uio_resid; 418 419 error = uiomove(&rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out], 420 size, uio); 421 if (error) { 422 break; 423 } 424 rpipe->pipe_buffer.out += size; 425 if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size) 426 rpipe->pipe_buffer.out = 0; 427 428 rpipe->pipe_buffer.cnt -= size; 429 430 /* 431 * If there is no more to read in the pipe, reset 432 * its pointers to the beginning. This improves 433 * cache hit stats. 434 */ 435 if (rpipe->pipe_buffer.cnt == 0) { 436 rpipe->pipe_buffer.in = 0; 437 rpipe->pipe_buffer.out = 0; 438 } 439 nread += size; 440#ifndef PIPE_NODIRECT 441 /* 442 * Direct copy, bypassing a kernel buffer. 443 */ 444 } else if ((size = rpipe->pipe_map.cnt) && 445 (rpipe->pipe_state & PIPE_DIRECTW)) { 446 caddr_t va; 447 if (size > (u_int) uio->uio_resid) 448 size = (u_int) uio->uio_resid; 449 450 va = (caddr_t) rpipe->pipe_map.kva + 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 */ 466 if (rpipe->pipe_state & PIPE_EOF) { 467 /* XXX error = ? */ 468 break; 469 } 470 471 /* 472 * If the "write-side" has been blocked, wake it up now. 473 */ 474 if (rpipe->pipe_state & PIPE_WANTW) { 475 rpipe->pipe_state &= ~PIPE_WANTW; 476 wakeup(rpipe); 477 } 478 479 /* 480 * Break if some data was read. 481 */ 482 if (nread > 0) 483 break; 484 485 /* 486 * Unlock the pipe buffer for our remaining processing. We 487 * will either break out with an error or we will sleep and 488 * relock to loop. 489 */ 490 pipeunlock(rpipe); 491 492 /* 493 * Handle non-blocking mode operation or 494 * wait for more data. 495 */ 496 if (fp->f_flag & FNONBLOCK) 497 error = EAGAIN; 498 else { 499 rpipe->pipe_state |= PIPE_WANTR; 500 if ((error = tsleep(rpipe, PRIBIO|PCATCH, "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); 511unlocked_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 */ 541static int 542pipe_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, paddr; 549 550 size = (u_int) uio->uio_iov->iov_len; 551 if (size > wpipe->pipe_buffer.size) 552 size = wpipe->pipe_buffer.size; 553 554 endaddr = round_page((vm_offset_t)uio->uio_iov->iov_base + size); 555 for(i = 0, addr = trunc_page((vm_offset_t)uio->uio_iov->iov_base); 556 addr < endaddr; 557 addr += PAGE_SIZE, i+=1) { 558 559 vm_page_t m; 560 561 if (vm_fault_quick((caddr_t)addr, VM_PROT_READ) < 0 || 562 (paddr = pmap_kextract(addr)) == 0) { 563 int j; 564 for(j=0;j<i;j++) 565 vm_page_unwire(wpipe->pipe_map.ms[j], 1); 566 return EFAULT; 567 } 568 569 m = PHYS_TO_VM_PAGE(paddr); 570 vm_page_wire(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 = ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK; 579 wpipe->pipe_map.cnt = size; 580 581/* 582 * and map the buffer 583 */ 584 if (wpipe->pipe_map.kva == 0) { 585 /* 586 * We need to allocate space for an extra page because the 587 * address range might (will) span pages at times. 588 */ 589 wpipe->pipe_map.kva = kmem_alloc_pageable(kernel_map, 590 wpipe->pipe_buffer.size + PAGE_SIZE); 591 amountpipekva += wpipe->pipe_buffer.size + PAGE_SIZE; 592 } 593 pmap_qenter(wpipe->pipe_map.kva, wpipe->pipe_map.ms, 594 wpipe->pipe_map.npages); 595 596/* 597 * and update the uio data 598 */ 599 600 uio->uio_iov->iov_len -= size; 601 uio->uio_iov->iov_base += size; 602 if (uio->uio_iov->iov_len == 0) 603 uio->uio_iov++; 604 uio->uio_resid -= size; 605 uio->uio_offset += size; 606 return 0; 607} 608 609/* 610 * unmap and unwire the process buffer 611 */ 612static void 613pipe_destroy_write_buffer(wpipe) 614struct pipe *wpipe; 615{ 616 int i; 617 618 if (wpipe->pipe_map.kva) { 619 pmap_qremove(wpipe->pipe_map.kva, wpipe->pipe_map.npages); 620 621 if (amountpipekva > MAXPIPEKVA) { 622 vm_offset_t kva = wpipe->pipe_map.kva; 623 wpipe->pipe_map.kva = 0; 624 kmem_free(kernel_map, kva, 625 wpipe->pipe_buffer.size + PAGE_SIZE); 626 amountpipekva -= wpipe->pipe_buffer.size + PAGE_SIZE; 627 } 628 } 629 for (i=0;i<wpipe->pipe_map.npages;i++) 630 vm_page_unwire(wpipe->pipe_map.ms[i], 1); 631} 632 633/* 634 * In the case of a signal, the writing process might go away. This 635 * code copies the data into the circular buffer so that the source 636 * pages can be freed without loss of data. 637 */ 638static void 639pipe_clone_write_buffer(wpipe) 640 struct pipe *wpipe; 641{ 642 int size; 643 int pos; 644 645 size = wpipe->pipe_map.cnt; 646 pos = wpipe->pipe_map.pos; 647 bcopy((caddr_t) wpipe->pipe_map.kva+pos, 648 (caddr_t) wpipe->pipe_buffer.buffer, 649 size); 650 651 wpipe->pipe_buffer.in = size; 652 wpipe->pipe_buffer.out = 0; 653 wpipe->pipe_buffer.cnt = size; 654 wpipe->pipe_state &= ~PIPE_DIRECTW; 655 656 pipe_destroy_write_buffer(wpipe); 657} 658 659/* 660 * This implements the pipe buffer write mechanism. Note that only 661 * a direct write OR a normal pipe write can be pending at any given time. 662 * If there are any characters in the pipe buffer, the direct write will 663 * be deferred until the receiving process grabs all of the bytes from 664 * the pipe buffer. Then the direct mapping write is set-up. 665 */ 666static int 667pipe_direct_write(wpipe, uio) 668 struct pipe *wpipe; 669 struct uio *uio; 670{ 671 int error; 672 673retry: 674 while (wpipe->pipe_state & PIPE_DIRECTW) { 675 if ( wpipe->pipe_state & PIPE_WANTR) { 676 wpipe->pipe_state &= ~PIPE_WANTR; 677 wakeup(wpipe); 678 } 679 wpipe->pipe_state |= PIPE_WANTW; 680 error = tsleep(wpipe, 681 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, 698 PRIBIO|PCATCH, "pipdwc", 0); 699 if (error) 700 goto error1; 701 if (wpipe->pipe_state & PIPE_EOF) { 702 error = EPIPE; 703 goto error1; 704 } 705 goto retry; 706 } 707 708 wpipe->pipe_state |= PIPE_DIRECTW; 709 710 error = pipe_build_write_buffer(wpipe, uio); 711 if (error) { 712 wpipe->pipe_state &= ~PIPE_DIRECTW; 713 goto error1; 714 } 715 716 error = 0; 717 while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) { 718 if (wpipe->pipe_state & PIPE_EOF) { 719 pipelock(wpipe, 0); 720 pipe_destroy_write_buffer(wpipe); 721 pipeunlock(wpipe); 722 pipeselwakeup(wpipe); 723 error = EPIPE; 724 goto error1; 725 } 726 if (wpipe->pipe_state & PIPE_WANTR) { 727 wpipe->pipe_state &= ~PIPE_WANTR; 728 wakeup(wpipe); 729 } 730 pipeselwakeup(wpipe); 731 error = tsleep(wpipe, PRIBIO|PCATCH, "pipdwt", 0); 732 } 733 734 pipelock(wpipe,0); 735 if (wpipe->pipe_state & PIPE_DIRECTW) { 736 /* 737 * this bit of trickery substitutes a kernel buffer for 738 * the process that might be going away. 739 */ 740 pipe_clone_write_buffer(wpipe); 741 } else { 742 pipe_destroy_write_buffer(wpipe); 743 } 744 pipeunlock(wpipe); 745 return error; 746 747error1: 748 wakeup(wpipe); 749 return error; 750} 751#endif 752 753static int 754pipe_write(fp, uio, cred, flags, p) 755 struct file *fp; 756 struct uio *uio; 757 struct ucred *cred; 758 struct proc *p; 759 int flags; 760{ 761 int error = 0; 762 int orig_resid; 763 struct pipe *wpipe, *rpipe; 764 765 rpipe = (struct pipe *) fp->f_data; 766 wpipe = rpipe->pipe_peer; 767 768 /* 769 * detect loss of pipe read side, issue SIGPIPE if lost. 770 */ 771 if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) { 772 return EPIPE; 773 } 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 } else { 790 return error; 791 } 792 } 793 794 KASSERT(wpipe->pipe_buffer.buffer != NULL, ("pipe buffer gone")); 795 796 ++wpipe->pipe_busy; 797 orig_resid = uio->uio_resid; 798 while (uio->uio_resid) { 799 int space; 800#ifndef PIPE_NODIRECT 801 /* 802 * If the transfer is large, we can gain performance if 803 * we do process-to-process copies directly. 804 * If the write is non-blocking, we don't use the 805 * direct write mechanism. 806 * 807 * The direct write mechanism will detect the reader going 808 * away on us. 809 */ 810 if ((uio->uio_iov->iov_len >= PIPE_MINDIRECT) && 811 (fp->f_flag & FNONBLOCK) == 0 && 812 (wpipe->pipe_map.kva || (amountpipekva < LIMITPIPEKVA)) && 813 (uio->uio_iov->iov_len >= PIPE_MINDIRECT)) { 814 error = pipe_direct_write( wpipe, uio); 815 if (error) { 816 break; 817 } 818 continue; 819 } 820#endif 821 822 /* 823 * Pipe buffered writes cannot be coincidental with 824 * direct writes. We wait until the currently executing 825 * direct write is completed before we start filling the 826 * pipe buffer. We break out if a signal occurs or the 827 * reader goes away. 828 */ 829 retrywrite: 830 while (wpipe->pipe_state & PIPE_DIRECTW) { 831 if (wpipe->pipe_state & PIPE_WANTR) { 832 wpipe->pipe_state &= ~PIPE_WANTR; 833 wakeup(wpipe); 834 } 835 error = tsleep(wpipe, PRIBIO|PCATCH, "pipbww", 0); 836 if (wpipe->pipe_state & PIPE_EOF) 837 break; 838 if (error) 839 break; 840 } 841 if (wpipe->pipe_state & PIPE_EOF) { 842 error = EPIPE; 843 break; 844 } 845 846 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt; 847 848 /* Writes of size <= PIPE_BUF must be atomic. */ 849 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF)) 850 space = 0; 851 852 if (space > 0 && (wpipe->pipe_buffer.cnt < PIPE_SIZE)) { 853 if ((error = pipelock(wpipe,1)) == 0) { 854 int size; /* Transfer size */ 855 int segsize; /* first segment to transfer */ 856 /* 857 * It is possible for a direct write to 858 * slip in on us... handle it here... 859 */ 860 if (wpipe->pipe_state & PIPE_DIRECTW) { 861 pipeunlock(wpipe); 862 goto retrywrite; 863 } 864 /* 865 * If a process blocked in uiomove, our 866 * value for space might be bad. 867 * 868 * XXX will we be ok if the reader has gone 869 * away here? 870 */ 871 if (space > wpipe->pipe_buffer.size - 872 wpipe->pipe_buffer.cnt) { 873 pipeunlock(wpipe); 874 goto retrywrite; 875 } 876 877 /* 878 * Transfer size is minimum of uio transfer 879 * and free space in pipe buffer. 880 */ 881 if (space > uio->uio_resid) 882 size = uio->uio_resid; 883 else 884 size = space; 885 /* 886 * First segment to transfer is minimum of 887 * transfer size and contiguous space in 888 * pipe buffer. If first segment to transfer 889 * is less than the transfer size, we've got 890 * a wraparound in the buffer. 891 */ 892 segsize = wpipe->pipe_buffer.size - 893 wpipe->pipe_buffer.in; 894 if (segsize > size) 895 segsize = size; 896 897 /* Transfer first segment */ 898 899 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in], 900 segsize, uio); 901 902 if (error == 0 && segsize < size) { 903 /* 904 * Transfer remaining part now, to 905 * support atomic writes. Wraparound 906 * happened. 907 */ 908 if (wpipe->pipe_buffer.in + segsize != 909 wpipe->pipe_buffer.size) 910 panic("Expected pipe buffer wraparound disappeared"); 911 912 error = uiomove(&wpipe->pipe_buffer.buffer[0], 913 size - segsize, uio); 914 } 915 if (error == 0) { 916 wpipe->pipe_buffer.in += size; 917 if (wpipe->pipe_buffer.in >= 918 wpipe->pipe_buffer.size) { 919 if (wpipe->pipe_buffer.in != size - segsize + wpipe->pipe_buffer.size) 920 panic("Expected wraparound bad"); 921 wpipe->pipe_buffer.in = size - segsize; 922 } 923 924 wpipe->pipe_buffer.cnt += size; 925 if (wpipe->pipe_buffer.cnt > wpipe->pipe_buffer.size) 926 panic("Pipe buffer overflow"); 927 928 } 929 pipeunlock(wpipe); 930 } 931 if (error) 932 break; 933 934 } else { 935 /* 936 * If the "read-side" has been blocked, wake it up now. 937 */ 938 if (wpipe->pipe_state & PIPE_WANTR) { 939 wpipe->pipe_state &= ~PIPE_WANTR; 940 wakeup(wpipe); 941 } 942 943 /* 944 * don't block on non-blocking I/O 945 */ 946 if (fp->f_flag & FNONBLOCK) { 947 error = EAGAIN; 948 break; 949 } 950 951 /* 952 * We have no more space and have something to offer, 953 * wake up select/poll. 954 */ 955 pipeselwakeup(wpipe); 956 957 wpipe->pipe_state |= PIPE_WANTW; 958 if ((error = tsleep(wpipe, (PRIBIO+1)|PCATCH, "pipewr", 0)) != 0) { 959 break; 960 } 961 /* 962 * If read side wants to go away, we just issue a signal 963 * to ourselves. 964 */ 965 if (wpipe->pipe_state & PIPE_EOF) { 966 error = EPIPE; 967 break; 968 } 969 } 970 } 971 972 --wpipe->pipe_busy; 973 if ((wpipe->pipe_busy == 0) && 974 (wpipe->pipe_state & PIPE_WANT)) { 975 wpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTR); 976 wakeup(wpipe); 977 } else if (wpipe->pipe_buffer.cnt > 0) { 978 /* 979 * If we have put any characters in the buffer, we wake up 980 * the reader. 981 */ 982 if (wpipe->pipe_state & PIPE_WANTR) { 983 wpipe->pipe_state &= ~PIPE_WANTR; 984 wakeup(wpipe); 985 } 986 } 987 988 /* 989 * Don't return EPIPE if I/O was successful 990 */ 991 if ((wpipe->pipe_buffer.cnt == 0) && 992 (uio->uio_resid == 0) && 993 (error == EPIPE)) 994 error = 0; 995 996 if (error == 0) 997 vfs_timestamp(&wpipe->pipe_mtime); 998 999 /* 1000 * We have something to offer, 1001 * wake up select/poll. 1002 */ 1003 if (wpipe->pipe_buffer.cnt) 1004 pipeselwakeup(wpipe); 1005 1006 return error; 1007} 1008 1009/* 1010 * we implement a very minimal set of ioctls for compatibility with sockets. 1011 */ 1012int 1013pipe_ioctl(fp, cmd, data, p) 1014 struct file *fp; 1015 u_long cmd; 1016 caddr_t data; 1017 struct proc *p; 1018{ 1019 struct pipe *mpipe = (struct pipe *)fp->f_data; 1020 1021 switch (cmd) { 1022 1023 case FIONBIO: 1024 return (0); 1025 1026 case FIOASYNC: 1027 if (*(int *)data) { 1028 mpipe->pipe_state |= PIPE_ASYNC; 1029 } else { 1030 mpipe->pipe_state &= ~PIPE_ASYNC; 1031 } 1032 return (0); 1033 1034 case FIONREAD: 1035 if (mpipe->pipe_state & PIPE_DIRECTW) 1036 *(int *)data = mpipe->pipe_map.cnt; 1037 else 1038 *(int *)data = mpipe->pipe_buffer.cnt; 1039 return (0); 1040 1041 case FIOSETOWN: 1042 return (fsetown(*(int *)data, &mpipe->pipe_sigio)); 1043 1044 case FIOGETOWN: 1045 *(int *)data = fgetown(mpipe->pipe_sigio); 1046 return (0); 1047 1048 /* This is deprecated, FIOSETOWN should be used instead. */ 1049 case TIOCSPGRP: 1050 return (fsetown(-(*(int *)data), &mpipe->pipe_sigio)); 1051 1052 /* This is deprecated, FIOGETOWN should be used instead. */ 1053 case TIOCGPGRP: 1054 *(int *)data = -fgetown(mpipe->pipe_sigio); 1055 return (0); 1056 1057 } 1058 return (ENOTTY); 1059} 1060 1061int 1062pipe_poll(fp, events, cred, p) 1063 struct file *fp; 1064 int events; 1065 struct ucred *cred; 1066 struct proc *p; 1067{ 1068 struct pipe *rpipe = (struct pipe *)fp->f_data; 1069 struct pipe *wpipe; 1070 int revents = 0; 1071 1072 wpipe = rpipe->pipe_peer; 1073 if (events & (POLLIN | POLLRDNORM)) 1074 if ((rpipe->pipe_state & PIPE_DIRECTW) || 1075 (rpipe->pipe_buffer.cnt > 0) || 1076 (rpipe->pipe_state & PIPE_EOF)) 1077 revents |= events & (POLLIN | POLLRDNORM); 1078 1079 if (events & (POLLOUT | POLLWRNORM)) 1080 if (wpipe == NULL || (wpipe->pipe_state & PIPE_EOF) || 1081 (((wpipe->pipe_state & PIPE_DIRECTW) == 0) && 1082 (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF)) 1083 revents |= events & (POLLOUT | POLLWRNORM); 1084 1085 if ((rpipe->pipe_state & PIPE_EOF) || 1086 (wpipe == NULL) || 1087 (wpipe->pipe_state & PIPE_EOF)) 1088 revents |= POLLHUP; 1089 1090 if (revents == 0) { 1091 if (events & (POLLIN | POLLRDNORM)) { 1092 selrecord(p, &rpipe->pipe_sel); 1093 rpipe->pipe_state |= PIPE_SEL; 1094 } 1095 1096 if (events & (POLLOUT | POLLWRNORM)) { 1097 selrecord(p, &wpipe->pipe_sel); 1098 wpipe->pipe_state |= PIPE_SEL; 1099 } 1100 } 1101 1102 return (revents); 1103} 1104 1105static int 1106pipe_stat(fp, ub, p) 1107 struct file *fp; 1108 struct stat *ub; 1109 struct proc *p; 1110{ 1111 struct pipe *pipe = (struct pipe *)fp->f_data; 1112 1113 bzero((caddr_t)ub, sizeof (*ub)); 1114 ub->st_mode = S_IFIFO; 1115 ub->st_blksize = pipe->pipe_buffer.size; 1116 ub->st_size = pipe->pipe_buffer.cnt; 1117 ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize; 1118 ub->st_atimespec = pipe->pipe_atime; 1119 ub->st_mtimespec = pipe->pipe_mtime; 1120 ub->st_ctimespec = pipe->pipe_ctime; 1121 ub->st_uid = fp->f_cred->cr_uid; 1122 ub->st_gid = fp->f_cred->cr_gid; 1123 /* 1124 * Left as 0: st_dev, st_ino, st_nlink, st_rdev, st_flags, st_gen. 1125 * XXX (st_dev, st_ino) should be unique. 1126 */ 1127 return 0; 1128} 1129 1130/* ARGSUSED */ 1131static int 1132pipe_close(fp, p) 1133 struct file *fp; 1134 struct proc *p; 1135{ 1136 struct pipe *cpipe = (struct pipe *)fp->f_data; 1137 1138 fp->f_ops = &badfileops; 1139 fp->f_data = NULL; 1140 funsetown(cpipe->pipe_sigio); 1141 pipeclose(cpipe); 1142 return 0; 1143} 1144 1145static void 1146pipe_free_kmem(cpipe) 1147 struct pipe *cpipe; 1148{ 1149 1150 if (cpipe->pipe_buffer.buffer != NULL) { 1151 if (cpipe->pipe_buffer.size > PIPE_SIZE) 1152 --nbigpipe; 1153 amountpipekva -= cpipe->pipe_buffer.size; 1154 kmem_free(kernel_map, 1155 (vm_offset_t)cpipe->pipe_buffer.buffer, 1156 cpipe->pipe_buffer.size); 1157 cpipe->pipe_buffer.buffer = NULL; 1158 } 1159#ifndef PIPE_NODIRECT 1160 if (cpipe->pipe_map.kva != NULL) { 1161 amountpipekva -= cpipe->pipe_buffer.size + PAGE_SIZE; 1162 kmem_free(kernel_map, 1163 cpipe->pipe_map.kva, 1164 cpipe->pipe_buffer.size + PAGE_SIZE); 1165 cpipe->pipe_map.cnt = 0; 1166 cpipe->pipe_map.kva = 0; 1167 cpipe->pipe_map.pos = 0; 1168 cpipe->pipe_map.npages = 0; 1169 } 1170#endif 1171} 1172 1173/* 1174 * shutdown the pipe 1175 */ 1176static void 1177pipeclose(cpipe) 1178 struct pipe *cpipe; 1179{ 1180 struct pipe *ppipe; 1181 1182 if (cpipe) { 1183 1184 pipeselwakeup(cpipe); 1185 1186 /* 1187 * If the other side is blocked, wake it up saying that 1188 * we want to close it down. 1189 */ 1190 while (cpipe->pipe_busy) { 1191 wakeup(cpipe); 1192 cpipe->pipe_state |= PIPE_WANT|PIPE_EOF; 1193 tsleep(cpipe, PRIBIO, "pipecl", 0); 1194 } 1195 1196 /* 1197 * Disconnect from peer 1198 */ 1199 if ((ppipe = cpipe->pipe_peer) != NULL) { 1200 pipeselwakeup(ppipe); 1201 1202 ppipe->pipe_state |= PIPE_EOF; 1203 wakeup(ppipe); 1204 ppipe->pipe_peer = NULL; 1205 } 1206 1207 /* 1208 * free resources 1209 */ 1210 pipe_free_kmem(cpipe); 1211 zfree(pipe_zone, cpipe); 1212 } 1213} 1214 1215/*ARGSUSED*/ 1216static int 1217pipe_kqfilter(struct file *fp, struct knote *kn) 1218{ 1219 struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data; 1220 1221 switch (kn->kn_filter) { 1222 case EVFILT_READ: 1223 kn->kn_fop = &pipe_rfiltops; 1224 break; 1225 case EVFILT_WRITE: 1226 kn->kn_fop = &pipe_wfiltops; 1227 break; 1228 default: 1229 return (1); 1230 } 1231 1232 SLIST_INSERT_HEAD(&rpipe->pipe_sel.si_note, kn, kn_selnext); 1233 return (0); 1234} 1235 1236static void 1237filt_pipedetach(struct knote *kn) 1238{ 1239 struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data; 1240 1241 SLIST_REMOVE(&rpipe->pipe_sel.si_note, kn, knote, kn_selnext); 1242} 1243 1244/*ARGSUSED*/ 1245static int 1246filt_piperead(struct knote *kn, long hint) 1247{ 1248 struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data; 1249 struct pipe *wpipe = rpipe->pipe_peer; 1250 1251 kn->kn_data = rpipe->pipe_buffer.cnt; 1252 if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW)) 1253 kn->kn_data = rpipe->pipe_map.cnt; 1254 1255 if ((rpipe->pipe_state & PIPE_EOF) || 1256 (wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) { 1257 kn->kn_flags |= EV_EOF; 1258 return (1); 1259 } 1260 return (kn->kn_data > 0); 1261} 1262 1263/*ARGSUSED*/ 1264static int 1265filt_pipewrite(struct knote *kn, long hint) 1266{ 1267 struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data; 1268 struct pipe *wpipe = rpipe->pipe_peer; 1269 1270 if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) { 1271 kn->kn_data = 0; 1272 kn->kn_flags |= EV_EOF; 1273 return (1); 1274 } 1275 kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt; 1276 if (wpipe->pipe_state & PIPE_DIRECTW) 1277 kn->kn_data = 0; 1278 1279 return (kn->kn_data >= PIPE_BUF); 1280} 1281