sys_pipe.c revision 70915
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 70915 2001-01-11 00:13:54Z dwmalone $ 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/selinfo.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 */ 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_stat __P((struct file *fp, struct stat *sb, struct proc *p)); 99static int pipe_ioctl __P((struct file *fp, u_long cmd, caddr_t data, struct proc *p)); 100 101static struct fileops pipeops = 102 { pipe_read, pipe_write, pipe_ioctl, pipe_poll, pipe_stat, pipe_close }; 103 104static int filt_pipeattach(struct knote *kn); 105static void filt_pipedetach(struct knote *kn); 106static int filt_piperead(struct knote *kn, long hint); 107static int filt_pipewrite(struct knote *kn, long hint); 108 109struct filterops pipe_rwfiltops[] = { 110 { 1, filt_pipeattach, filt_pipedetach, filt_piperead }, 111 { 1, filt_pipeattach, filt_pipedetach, filt_pipewrite }, 112}; 113 114/* 115 * Default pipe buffer size(s), this can be kind-of large now because pipe 116 * space is pageable. The pipe code will try to maintain locality of 117 * reference for performance reasons, so small amounts of outstanding I/O 118 * will not wipe the cache. 119 */ 120#define MINPIPESIZE (PIPE_SIZE/3) 121#define MAXPIPESIZE (2*PIPE_SIZE/3) 122 123/* 124 * Maximum amount of kva for pipes -- this is kind-of a soft limit, but 125 * is there so that on large systems, we don't exhaust it. 126 */ 127#define MAXPIPEKVA (8*1024*1024) 128 129/* 130 * Limit for direct transfers, we cannot, of course limit 131 * the amount of kva for pipes in general though. 132 */ 133#define LIMITPIPEKVA (16*1024*1024) 134 135/* 136 * Limit the number of "big" pipes 137 */ 138#define LIMITBIGPIPES 32 139static int nbigpipe; 140 141static int amountpipekva; 142 143static void pipeclose __P((struct pipe *cpipe)); 144static void pipeinit __P((struct pipe *cpipe)); 145static __inline int pipelock __P((struct pipe *cpipe, int catch)); 146static __inline void pipeunlock __P((struct pipe *cpipe)); 147static __inline void pipeselwakeup __P((struct pipe *cpipe)); 148#ifndef PIPE_NODIRECT 149static int pipe_build_write_buffer __P((struct pipe *wpipe, struct uio *uio)); 150static void pipe_destroy_write_buffer __P((struct pipe *wpipe)); 151static int pipe_direct_write __P((struct pipe *wpipe, struct uio *uio)); 152static void pipe_clone_write_buffer __P((struct pipe *wpipe)); 153#endif 154static void pipespace __P((struct pipe *cpipe)); 155 156static vm_zone_t pipe_zone; 157 158/* 159 * The pipe system call for the DTYPE_PIPE type of pipes 160 */ 161 162/* ARGSUSED */ 163int 164pipe(p, uap) 165 struct proc *p; 166 struct pipe_args /* { 167 int dummy; 168 } */ *uap; 169{ 170 register struct filedesc *fdp = p->p_fd; 171 struct file *rf, *wf; 172 struct pipe *rpipe, *wpipe; 173 int fd, error; 174 175 if (pipe_zone == NULL) 176 pipe_zone = zinit("PIPE", sizeof (struct pipe), 0, 0, 4); 177 178 rpipe = zalloc( pipe_zone); 179 pipeinit(rpipe); 180 rpipe->pipe_state |= PIPE_DIRECTOK; 181 wpipe = zalloc( pipe_zone); 182 pipeinit(wpipe); 183 wpipe->pipe_state |= PIPE_DIRECTOK; 184 185 error = falloc(p, &rf, &fd); 186 if (error) { 187 pipeclose(rpipe); 188 pipeclose(wpipe); 189 return (error); 190 } 191 fhold(rf); 192 p->p_retval[0] = fd; 193 194 /* 195 * Warning: once we've gotten past allocation of the fd for the 196 * read-side, we can only drop the read side via fdrop() in order 197 * to avoid races against processes which manage to dup() the read 198 * side while we are blocked trying to allocate the write side. 199 */ 200 rf->f_flag = FREAD | FWRITE; 201 rf->f_type = DTYPE_PIPE; 202 rf->f_data = (caddr_t)rpipe; 203 rf->f_ops = &pipeops; 204 error = falloc(p, &wf, &fd); 205 if (error) { 206 if (fdp->fd_ofiles[p->p_retval[0]] == rf) { 207 fdp->fd_ofiles[p->p_retval[0]] = NULL; 208 fdrop(rf, p); 209 } 210 fdrop(rf, p); 211 /* rpipe has been closed by fdrop(). */ 212 pipeclose(wpipe); 213 return (error); 214 } 215 wf->f_flag = FREAD | FWRITE; 216 wf->f_type = DTYPE_PIPE; 217 wf->f_data = (caddr_t)wpipe; 218 wf->f_ops = &pipeops; 219 p->p_retval[1] = fd; 220 221 rpipe->pipe_peer = wpipe; 222 wpipe->pipe_peer = rpipe; 223 fdrop(rf, p); 224 225 return (0); 226} 227 228/* 229 * Allocate kva for pipe circular buffer, the space is pageable 230 */ 231static void 232pipespace(cpipe) 233 struct pipe *cpipe; 234{ 235 int npages, error; 236 237 npages = round_page(cpipe->pipe_buffer.size)/PAGE_SIZE; 238 /* 239 * Create an object, I don't like the idea of paging to/from 240 * kernel_object. 241 * XXX -- minor change needed here for NetBSD/OpenBSD VM systems. 242 */ 243 cpipe->pipe_buffer.object = vm_object_allocate(OBJT_DEFAULT, npages); 244 cpipe->pipe_buffer.buffer = (caddr_t) vm_map_min(kernel_map); 245 246 /* 247 * Insert the object into the kernel map, and allocate kva for it. 248 * The map entry is, by default, pageable. 249 * XXX -- minor change needed here for NetBSD/OpenBSD VM systems. 250 */ 251 error = vm_map_find(kernel_map, cpipe->pipe_buffer.object, 0, 252 (vm_offset_t *) &cpipe->pipe_buffer.buffer, 253 cpipe->pipe_buffer.size, 1, 254 VM_PROT_ALL, VM_PROT_ALL, 0); 255 256 if (error != KERN_SUCCESS) 257 panic("pipeinit: cannot allocate pipe -- out of kvm -- code = %d", error); 258 amountpipekva += cpipe->pipe_buffer.size; 259} 260 261/* 262 * initialize and allocate VM and memory for pipe 263 */ 264static void 265pipeinit(cpipe) 266 struct pipe *cpipe; 267{ 268 269 cpipe->pipe_buffer.in = 0; 270 cpipe->pipe_buffer.out = 0; 271 cpipe->pipe_buffer.cnt = 0; 272 cpipe->pipe_buffer.size = PIPE_SIZE; 273 274 /* Buffer kva gets dynamically allocated */ 275 cpipe->pipe_buffer.buffer = NULL; 276 /* cpipe->pipe_buffer.object = invalid */ 277 278 cpipe->pipe_state = 0; 279 cpipe->pipe_peer = NULL; 280 cpipe->pipe_busy = 0; 281 vfs_timestamp(&cpipe->pipe_ctime); 282 cpipe->pipe_atime = cpipe->pipe_ctime; 283 cpipe->pipe_mtime = cpipe->pipe_ctime; 284 bzero(&cpipe->pipe_sel, sizeof cpipe->pipe_sel); 285 286#ifndef PIPE_NODIRECT 287 /* 288 * pipe data structure initializations to support direct pipe I/O 289 */ 290 cpipe->pipe_map.cnt = 0; 291 cpipe->pipe_map.kva = 0; 292 cpipe->pipe_map.pos = 0; 293 cpipe->pipe_map.npages = 0; 294 /* cpipe->pipe_map.ms[] = invalid */ 295#endif 296} 297 298 299/* 300 * lock a pipe for I/O, blocking other access 301 */ 302static __inline int 303pipelock(cpipe, catch) 304 struct pipe *cpipe; 305 int catch; 306{ 307 int error; 308 while (cpipe->pipe_state & PIPE_LOCK) { 309 cpipe->pipe_state |= PIPE_LWANT; 310 if ((error = tsleep( cpipe, 311 catch?(PRIBIO|PCATCH):PRIBIO, "pipelk", 0)) != 0) { 312 return error; 313 } 314 } 315 cpipe->pipe_state |= PIPE_LOCK; 316 return 0; 317} 318 319/* 320 * unlock a pipe I/O lock 321 */ 322static __inline void 323pipeunlock(cpipe) 324 struct pipe *cpipe; 325{ 326 cpipe->pipe_state &= ~PIPE_LOCK; 327 if (cpipe->pipe_state & PIPE_LWANT) { 328 cpipe->pipe_state &= ~PIPE_LWANT; 329 wakeup(cpipe); 330 } 331} 332 333static __inline void 334pipeselwakeup(cpipe) 335 struct pipe *cpipe; 336{ 337 if (cpipe->pipe_state & PIPE_SEL) { 338 cpipe->pipe_state &= ~PIPE_SEL; 339 selwakeup(&cpipe->pipe_sel); 340 } 341 if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio) 342 pgsigio(cpipe->pipe_sigio, SIGIO, 0); 343 KNOTE(&cpipe->pipe_sel.si_note, 0); 344} 345 346/* ARGSUSED */ 347static int 348pipe_read(fp, uio, cred, flags, p) 349 struct file *fp; 350 struct uio *uio; 351 struct ucred *cred; 352 struct proc *p; 353 int flags; 354{ 355 356 struct pipe *rpipe = (struct pipe *) fp->f_data; 357 int error; 358 int nread = 0; 359 u_int size; 360 361 ++rpipe->pipe_busy; 362 error = pipelock(rpipe, 1); 363 if (error) 364 goto unlocked_error; 365 366 while (uio->uio_resid) { 367 /* 368 * normal pipe buffer receive 369 */ 370 if (rpipe->pipe_buffer.cnt > 0) { 371 size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out; 372 if (size > rpipe->pipe_buffer.cnt) 373 size = rpipe->pipe_buffer.cnt; 374 if (size > (u_int) uio->uio_resid) 375 size = (u_int) uio->uio_resid; 376 377 error = uiomove(&rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out], 378 size, uio); 379 if (error) { 380 break; 381 } 382 rpipe->pipe_buffer.out += size; 383 if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size) 384 rpipe->pipe_buffer.out = 0; 385 386 rpipe->pipe_buffer.cnt -= size; 387 388 /* 389 * If there is no more to read in the pipe, reset 390 * its pointers to the beginning. This improves 391 * cache hit stats. 392 */ 393 if (rpipe->pipe_buffer.cnt == 0) { 394 rpipe->pipe_buffer.in = 0; 395 rpipe->pipe_buffer.out = 0; 396 } 397 nread += size; 398#ifndef PIPE_NODIRECT 399 /* 400 * Direct copy, bypassing a kernel buffer. 401 */ 402 } else if ((size = rpipe->pipe_map.cnt) && 403 (rpipe->pipe_state & PIPE_DIRECTW)) { 404 caddr_t va; 405 if (size > (u_int) uio->uio_resid) 406 size = (u_int) uio->uio_resid; 407 408 va = (caddr_t) rpipe->pipe_map.kva + rpipe->pipe_map.pos; 409 error = uiomove(va, size, uio); 410 if (error) 411 break; 412 nread += size; 413 rpipe->pipe_map.pos += size; 414 rpipe->pipe_map.cnt -= size; 415 if (rpipe->pipe_map.cnt == 0) { 416 rpipe->pipe_state &= ~PIPE_DIRECTW; 417 wakeup(rpipe); 418 } 419#endif 420 } else { 421 /* 422 * detect EOF condition 423 */ 424 if (rpipe->pipe_state & PIPE_EOF) { 425 /* XXX error = ? */ 426 break; 427 } 428 429 /* 430 * If the "write-side" has been blocked, wake it up now. 431 */ 432 if (rpipe->pipe_state & PIPE_WANTW) { 433 rpipe->pipe_state &= ~PIPE_WANTW; 434 wakeup(rpipe); 435 } 436 437 /* 438 * Break if some data was read. 439 */ 440 if (nread > 0) 441 break; 442 443 /* 444 * Unlock the pipe buffer for our remaining processing. We 445 * will either break out with an error or we will sleep and 446 * relock to loop. 447 */ 448 pipeunlock(rpipe); 449 450 /* 451 * Handle non-blocking mode operation or 452 * wait for more data. 453 */ 454 if (fp->f_flag & FNONBLOCK) 455 error = EAGAIN; 456 else { 457 rpipe->pipe_state |= PIPE_WANTR; 458 if ((error = tsleep(rpipe, PRIBIO|PCATCH, "piperd", 0)) == 0) 459 error = pipelock(rpipe, 1); 460 } 461 if (error) 462 goto unlocked_error; 463 } 464 } 465 pipeunlock(rpipe); 466 467 if (error == 0) 468 vfs_timestamp(&rpipe->pipe_atime); 469unlocked_error: 470 --rpipe->pipe_busy; 471 472 /* 473 * PIPE_WANT processing only makes sense if pipe_busy is 0. 474 */ 475 if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) { 476 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW); 477 wakeup(rpipe); 478 } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) { 479 /* 480 * Handle write blocking hysteresis. 481 */ 482 if (rpipe->pipe_state & PIPE_WANTW) { 483 rpipe->pipe_state &= ~PIPE_WANTW; 484 wakeup(rpipe); 485 } 486 } 487 488 if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF) 489 pipeselwakeup(rpipe); 490 491 return error; 492} 493 494#ifndef PIPE_NODIRECT 495/* 496 * Map the sending processes' buffer into kernel space and wire it. 497 * This is similar to a physical write operation. 498 */ 499static int 500pipe_build_write_buffer(wpipe, uio) 501 struct pipe *wpipe; 502 struct uio *uio; 503{ 504 u_int size; 505 int i; 506 vm_offset_t addr, endaddr, paddr; 507 508 size = (u_int) uio->uio_iov->iov_len; 509 if (size > wpipe->pipe_buffer.size) 510 size = wpipe->pipe_buffer.size; 511 512 endaddr = round_page((vm_offset_t)uio->uio_iov->iov_base + size); 513 for(i = 0, addr = trunc_page((vm_offset_t)uio->uio_iov->iov_base); 514 addr < endaddr; 515 addr += PAGE_SIZE, i+=1) { 516 517 vm_page_t m; 518 519 if (vm_fault_quick((caddr_t)addr, VM_PROT_READ) < 0 || 520 (paddr = pmap_kextract(addr)) == 0) { 521 int j; 522 for(j=0;j<i;j++) 523 vm_page_unwire(wpipe->pipe_map.ms[j], 1); 524 return EFAULT; 525 } 526 527 m = PHYS_TO_VM_PAGE(paddr); 528 vm_page_wire(m); 529 wpipe->pipe_map.ms[i] = m; 530 } 531 532/* 533 * set up the control block 534 */ 535 wpipe->pipe_map.npages = i; 536 wpipe->pipe_map.pos = ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK; 537 wpipe->pipe_map.cnt = size; 538 539/* 540 * and map the buffer 541 */ 542 if (wpipe->pipe_map.kva == 0) { 543 /* 544 * We need to allocate space for an extra page because the 545 * address range might (will) span pages at times. 546 */ 547 wpipe->pipe_map.kva = kmem_alloc_pageable(kernel_map, 548 wpipe->pipe_buffer.size + PAGE_SIZE); 549 amountpipekva += wpipe->pipe_buffer.size + PAGE_SIZE; 550 } 551 pmap_qenter(wpipe->pipe_map.kva, wpipe->pipe_map.ms, 552 wpipe->pipe_map.npages); 553 554/* 555 * and update the uio data 556 */ 557 558 uio->uio_iov->iov_len -= size; 559 uio->uio_iov->iov_base += size; 560 if (uio->uio_iov->iov_len == 0) 561 uio->uio_iov++; 562 uio->uio_resid -= size; 563 uio->uio_offset += size; 564 return 0; 565} 566 567/* 568 * unmap and unwire the process buffer 569 */ 570static void 571pipe_destroy_write_buffer(wpipe) 572struct pipe *wpipe; 573{ 574 int i; 575 if (wpipe->pipe_map.kva) { 576 pmap_qremove(wpipe->pipe_map.kva, wpipe->pipe_map.npages); 577 578 if (amountpipekva > MAXPIPEKVA) { 579 vm_offset_t kva = wpipe->pipe_map.kva; 580 wpipe->pipe_map.kva = 0; 581 kmem_free(kernel_map, kva, 582 wpipe->pipe_buffer.size + PAGE_SIZE); 583 amountpipekva -= wpipe->pipe_buffer.size + PAGE_SIZE; 584 } 585 } 586 for (i=0;i<wpipe->pipe_map.npages;i++) 587 vm_page_unwire(wpipe->pipe_map.ms[i], 1); 588} 589 590/* 591 * In the case of a signal, the writing process might go away. This 592 * code copies the data into the circular buffer so that the source 593 * pages can be freed without loss of data. 594 */ 595static void 596pipe_clone_write_buffer(wpipe) 597struct pipe *wpipe; 598{ 599 int size; 600 int pos; 601 602 size = wpipe->pipe_map.cnt; 603 pos = wpipe->pipe_map.pos; 604 bcopy((caddr_t) wpipe->pipe_map.kva+pos, 605 (caddr_t) wpipe->pipe_buffer.buffer, 606 size); 607 608 wpipe->pipe_buffer.in = size; 609 wpipe->pipe_buffer.out = 0; 610 wpipe->pipe_buffer.cnt = size; 611 wpipe->pipe_state &= ~PIPE_DIRECTW; 612 613 pipe_destroy_write_buffer(wpipe); 614} 615 616/* 617 * This implements the pipe buffer write mechanism. Note that only 618 * a direct write OR a normal pipe write can be pending at any given time. 619 * If there are any characters in the pipe buffer, the direct write will 620 * be deferred until the receiving process grabs all of the bytes from 621 * the pipe buffer. Then the direct mapping write is set-up. 622 */ 623static int 624pipe_direct_write(wpipe, uio) 625 struct pipe *wpipe; 626 struct uio *uio; 627{ 628 int error; 629retry: 630 while (wpipe->pipe_state & PIPE_DIRECTW) { 631 if ( wpipe->pipe_state & PIPE_WANTR) { 632 wpipe->pipe_state &= ~PIPE_WANTR; 633 wakeup(wpipe); 634 } 635 wpipe->pipe_state |= PIPE_WANTW; 636 error = tsleep(wpipe, 637 PRIBIO|PCATCH, "pipdww", 0); 638 if (error) 639 goto error1; 640 if (wpipe->pipe_state & PIPE_EOF) { 641 error = EPIPE; 642 goto error1; 643 } 644 } 645 wpipe->pipe_map.cnt = 0; /* transfer not ready yet */ 646 if (wpipe->pipe_buffer.cnt > 0) { 647 if ( wpipe->pipe_state & PIPE_WANTR) { 648 wpipe->pipe_state &= ~PIPE_WANTR; 649 wakeup(wpipe); 650 } 651 652 wpipe->pipe_state |= PIPE_WANTW; 653 error = tsleep(wpipe, 654 PRIBIO|PCATCH, "pipdwc", 0); 655 if (error) 656 goto error1; 657 if (wpipe->pipe_state & PIPE_EOF) { 658 error = EPIPE; 659 goto error1; 660 } 661 goto retry; 662 } 663 664 wpipe->pipe_state |= PIPE_DIRECTW; 665 666 error = pipe_build_write_buffer(wpipe, uio); 667 if (error) { 668 wpipe->pipe_state &= ~PIPE_DIRECTW; 669 goto error1; 670 } 671 672 error = 0; 673 while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) { 674 if (wpipe->pipe_state & PIPE_EOF) { 675 pipelock(wpipe, 0); 676 pipe_destroy_write_buffer(wpipe); 677 pipeunlock(wpipe); 678 pipeselwakeup(wpipe); 679 error = EPIPE; 680 goto error1; 681 } 682 if (wpipe->pipe_state & PIPE_WANTR) { 683 wpipe->pipe_state &= ~PIPE_WANTR; 684 wakeup(wpipe); 685 } 686 pipeselwakeup(wpipe); 687 error = tsleep(wpipe, PRIBIO|PCATCH, "pipdwt", 0); 688 } 689 690 pipelock(wpipe,0); 691 if (wpipe->pipe_state & PIPE_DIRECTW) { 692 /* 693 * this bit of trickery substitutes a kernel buffer for 694 * the process that might be going away. 695 */ 696 pipe_clone_write_buffer(wpipe); 697 } else { 698 pipe_destroy_write_buffer(wpipe); 699 } 700 pipeunlock(wpipe); 701 return error; 702 703error1: 704 wakeup(wpipe); 705 return error; 706} 707#endif 708 709static int 710pipe_write(fp, uio, cred, flags, p) 711 struct file *fp; 712 struct uio *uio; 713 struct ucred *cred; 714 struct proc *p; 715 int flags; 716{ 717 int error = 0; 718 int orig_resid; 719 720 struct pipe *wpipe, *rpipe; 721 722 rpipe = (struct pipe *) fp->f_data; 723 wpipe = rpipe->pipe_peer; 724 725 /* 726 * detect loss of pipe read side, issue SIGPIPE if lost. 727 */ 728 if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) { 729 return EPIPE; 730 } 731 732 /* 733 * If it is advantageous to resize the pipe buffer, do 734 * so. 735 */ 736 if ((uio->uio_resid > PIPE_SIZE) && 737 (nbigpipe < LIMITBIGPIPES) && 738 (wpipe->pipe_state & PIPE_DIRECTW) == 0 && 739 (wpipe->pipe_buffer.size <= PIPE_SIZE) && 740 (wpipe->pipe_buffer.cnt == 0)) { 741 742 if (wpipe->pipe_buffer.buffer) { 743 amountpipekva -= wpipe->pipe_buffer.size; 744 kmem_free(kernel_map, 745 (vm_offset_t)wpipe->pipe_buffer.buffer, 746 wpipe->pipe_buffer.size); 747 } 748 749#ifndef PIPE_NODIRECT 750 if (wpipe->pipe_map.kva) { 751 amountpipekva -= wpipe->pipe_buffer.size + PAGE_SIZE; 752 kmem_free(kernel_map, 753 wpipe->pipe_map.kva, 754 wpipe->pipe_buffer.size + PAGE_SIZE); 755 } 756#endif 757 758 wpipe->pipe_buffer.in = 0; 759 wpipe->pipe_buffer.out = 0; 760 wpipe->pipe_buffer.cnt = 0; 761 wpipe->pipe_buffer.size = BIG_PIPE_SIZE; 762 wpipe->pipe_buffer.buffer = NULL; 763 ++nbigpipe; 764 765#ifndef PIPE_NODIRECT 766 wpipe->pipe_map.cnt = 0; 767 wpipe->pipe_map.kva = 0; 768 wpipe->pipe_map.pos = 0; 769 wpipe->pipe_map.npages = 0; 770#endif 771 772 } 773 774 775 if( wpipe->pipe_buffer.buffer == NULL) { 776 if ((error = pipelock(wpipe,1)) == 0) { 777 pipespace(wpipe); 778 pipeunlock(wpipe); 779 } else { 780 return error; 781 } 782 } 783 784 ++wpipe->pipe_busy; 785 orig_resid = uio->uio_resid; 786 while (uio->uio_resid) { 787 int space; 788#ifndef PIPE_NODIRECT 789 /* 790 * If the transfer is large, we can gain performance if 791 * we do process-to-process copies directly. 792 * If the write is non-blocking, we don't use the 793 * direct write mechanism. 794 * 795 * The direct write mechanism will detect the reader going 796 * away on us. 797 */ 798 if ((uio->uio_iov->iov_len >= PIPE_MINDIRECT) && 799 (fp->f_flag & FNONBLOCK) == 0 && 800 (wpipe->pipe_map.kva || (amountpipekva < LIMITPIPEKVA)) && 801 (uio->uio_iov->iov_len >= PIPE_MINDIRECT)) { 802 error = pipe_direct_write( wpipe, uio); 803 if (error) { 804 break; 805 } 806 continue; 807 } 808#endif 809 810 /* 811 * Pipe buffered writes cannot be coincidental with 812 * direct writes. We wait until the currently executing 813 * direct write is completed before we start filling the 814 * pipe buffer. We break out if a signal occurs or the 815 * reader goes away. 816 */ 817 retrywrite: 818 while (wpipe->pipe_state & PIPE_DIRECTW) { 819 if (wpipe->pipe_state & PIPE_WANTR) { 820 wpipe->pipe_state &= ~PIPE_WANTR; 821 wakeup(wpipe); 822 } 823 error = tsleep(wpipe, PRIBIO|PCATCH, "pipbww", 0); 824 if (wpipe->pipe_state & PIPE_EOF) 825 break; 826 if (error) 827 break; 828 } 829 if (wpipe->pipe_state & PIPE_EOF) { 830 error = EPIPE; 831 break; 832 } 833 834 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt; 835 836 /* Writes of size <= PIPE_BUF must be atomic. */ 837 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF)) 838 space = 0; 839 840 if (space > 0 && (wpipe->pipe_buffer.cnt < PIPE_SIZE)) { 841 if ((error = pipelock(wpipe,1)) == 0) { 842 int size; /* Transfer size */ 843 int segsize; /* first segment to transfer */ 844 /* 845 * It is possible for a direct write to 846 * slip in on us... handle it here... 847 */ 848 if (wpipe->pipe_state & PIPE_DIRECTW) { 849 pipeunlock(wpipe); 850 goto retrywrite; 851 } 852 /* 853 * If a process blocked in uiomove, our 854 * value for space might be bad. 855 * 856 * XXX will we be ok if the reader has gone 857 * away here? 858 */ 859 if (space > wpipe->pipe_buffer.size - 860 wpipe->pipe_buffer.cnt) { 861 pipeunlock(wpipe); 862 goto retrywrite; 863 } 864 865 /* 866 * Transfer size is minimum of uio transfer 867 * and free space in pipe buffer. 868 */ 869 if (space > uio->uio_resid) 870 size = uio->uio_resid; 871 else 872 size = space; 873 /* 874 * First segment to transfer is minimum of 875 * transfer size and contiguous space in 876 * pipe buffer. If first segment to transfer 877 * is less than the transfer size, we've got 878 * a wraparound in the buffer. 879 */ 880 segsize = wpipe->pipe_buffer.size - 881 wpipe->pipe_buffer.in; 882 if (segsize > size) 883 segsize = size; 884 885 /* Transfer first segment */ 886 887 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in], 888 segsize, uio); 889 890 if (error == 0 && segsize < size) { 891 /* 892 * Transfer remaining part now, to 893 * support atomic writes. Wraparound 894 * happened. 895 */ 896 if (wpipe->pipe_buffer.in + segsize != 897 wpipe->pipe_buffer.size) 898 panic("Expected pipe buffer wraparound disappeared"); 899 900 error = uiomove(&wpipe->pipe_buffer.buffer[0], 901 size - segsize, uio); 902 } 903 if (error == 0) { 904 wpipe->pipe_buffer.in += size; 905 if (wpipe->pipe_buffer.in >= 906 wpipe->pipe_buffer.size) { 907 if (wpipe->pipe_buffer.in != size - segsize + wpipe->pipe_buffer.size) 908 panic("Expected wraparound bad"); 909 wpipe->pipe_buffer.in = size - segsize; 910 } 911 912 wpipe->pipe_buffer.cnt += size; 913 if (wpipe->pipe_buffer.cnt > wpipe->pipe_buffer.size) 914 panic("Pipe buffer overflow"); 915 916 } 917 pipeunlock(wpipe); 918 } 919 if (error) 920 break; 921 922 } else { 923 /* 924 * If the "read-side" has been blocked, wake it up now. 925 */ 926 if (wpipe->pipe_state & PIPE_WANTR) { 927 wpipe->pipe_state &= ~PIPE_WANTR; 928 wakeup(wpipe); 929 } 930 931 /* 932 * don't block on non-blocking I/O 933 */ 934 if (fp->f_flag & FNONBLOCK) { 935 error = EAGAIN; 936 break; 937 } 938 939 /* 940 * We have no more space and have something to offer, 941 * wake up select/poll. 942 */ 943 pipeselwakeup(wpipe); 944 945 wpipe->pipe_state |= PIPE_WANTW; 946 if ((error = tsleep(wpipe, (PRIBIO+1)|PCATCH, "pipewr", 0)) != 0) { 947 break; 948 } 949 /* 950 * If read side wants to go away, we just issue a signal 951 * to ourselves. 952 */ 953 if (wpipe->pipe_state & PIPE_EOF) { 954 error = EPIPE; 955 break; 956 } 957 } 958 } 959 960 --wpipe->pipe_busy; 961 if ((wpipe->pipe_busy == 0) && 962 (wpipe->pipe_state & PIPE_WANT)) { 963 wpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTR); 964 wakeup(wpipe); 965 } else if (wpipe->pipe_buffer.cnt > 0) { 966 /* 967 * If we have put any characters in the buffer, we wake up 968 * the reader. 969 */ 970 if (wpipe->pipe_state & PIPE_WANTR) { 971 wpipe->pipe_state &= ~PIPE_WANTR; 972 wakeup(wpipe); 973 } 974 } 975 976 /* 977 * Don't return EPIPE if I/O was successful 978 */ 979 if ((wpipe->pipe_buffer.cnt == 0) && 980 (uio->uio_resid == 0) && 981 (error == EPIPE)) 982 error = 0; 983 984 if (error == 0) 985 vfs_timestamp(&wpipe->pipe_mtime); 986 987 /* 988 * We have something to offer, 989 * wake up select/poll. 990 */ 991 if (wpipe->pipe_buffer.cnt) 992 pipeselwakeup(wpipe); 993 994 return error; 995} 996 997/* 998 * we implement a very minimal set of ioctls for compatibility with sockets. 999 */ 1000int 1001pipe_ioctl(fp, cmd, data, p) 1002 struct file *fp; 1003 u_long cmd; 1004 register caddr_t data; 1005 struct proc *p; 1006{ 1007 register struct pipe *mpipe = (struct pipe *)fp->f_data; 1008 1009 switch (cmd) { 1010 1011 case FIONBIO: 1012 return (0); 1013 1014 case FIOASYNC: 1015 if (*(int *)data) { 1016 mpipe->pipe_state |= PIPE_ASYNC; 1017 } else { 1018 mpipe->pipe_state &= ~PIPE_ASYNC; 1019 } 1020 return (0); 1021 1022 case FIONREAD: 1023 if (mpipe->pipe_state & PIPE_DIRECTW) 1024 *(int *)data = mpipe->pipe_map.cnt; 1025 else 1026 *(int *)data = mpipe->pipe_buffer.cnt; 1027 return (0); 1028 1029 case FIOSETOWN: 1030 return (fsetown(*(int *)data, &mpipe->pipe_sigio)); 1031 1032 case FIOGETOWN: 1033 *(int *)data = fgetown(mpipe->pipe_sigio); 1034 return (0); 1035 1036 /* This is deprecated, FIOSETOWN should be used instead. */ 1037 case TIOCSPGRP: 1038 return (fsetown(-(*(int *)data), &mpipe->pipe_sigio)); 1039 1040 /* This is deprecated, FIOGETOWN should be used instead. */ 1041 case TIOCGPGRP: 1042 *(int *)data = -fgetown(mpipe->pipe_sigio); 1043 return (0); 1044 1045 } 1046 return (ENOTTY); 1047} 1048 1049int 1050pipe_poll(fp, events, cred, p) 1051 struct file *fp; 1052 int events; 1053 struct ucred *cred; 1054 struct proc *p; 1055{ 1056 register struct pipe *rpipe = (struct pipe *)fp->f_data; 1057 struct pipe *wpipe; 1058 int revents = 0; 1059 1060 wpipe = rpipe->pipe_peer; 1061 if (events & (POLLIN | POLLRDNORM)) 1062 if ((rpipe->pipe_state & PIPE_DIRECTW) || 1063 (rpipe->pipe_buffer.cnt > 0) || 1064 (rpipe->pipe_state & PIPE_EOF)) 1065 revents |= events & (POLLIN | POLLRDNORM); 1066 1067 if (events & (POLLOUT | POLLWRNORM)) 1068 if (wpipe == NULL || (wpipe->pipe_state & PIPE_EOF) || 1069 (((wpipe->pipe_state & PIPE_DIRECTW) == 0) && 1070 (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF)) 1071 revents |= events & (POLLOUT | POLLWRNORM); 1072 1073 if ((rpipe->pipe_state & PIPE_EOF) || 1074 (wpipe == NULL) || 1075 (wpipe->pipe_state & PIPE_EOF)) 1076 revents |= POLLHUP; 1077 1078 if (revents == 0) { 1079 if (events & (POLLIN | POLLRDNORM)) { 1080 selrecord(p, &rpipe->pipe_sel); 1081 rpipe->pipe_state |= PIPE_SEL; 1082 } 1083 1084 if (events & (POLLOUT | POLLWRNORM)) { 1085 selrecord(p, &wpipe->pipe_sel); 1086 wpipe->pipe_state |= PIPE_SEL; 1087 } 1088 } 1089 1090 return (revents); 1091} 1092 1093static int 1094pipe_stat(fp, ub, p) 1095 struct file *fp; 1096 struct stat *ub; 1097 struct proc *p; 1098{ 1099 struct pipe *pipe = (struct pipe *)fp->f_data; 1100 1101 bzero((caddr_t)ub, sizeof (*ub)); 1102 ub->st_mode = S_IFIFO; 1103 ub->st_blksize = pipe->pipe_buffer.size; 1104 ub->st_size = pipe->pipe_buffer.cnt; 1105 ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize; 1106 ub->st_atimespec = pipe->pipe_atime; 1107 ub->st_mtimespec = pipe->pipe_mtime; 1108 ub->st_ctimespec = pipe->pipe_ctime; 1109 ub->st_uid = fp->f_cred->cr_uid; 1110 ub->st_gid = fp->f_cred->cr_gid; 1111 /* 1112 * Left as 0: st_dev, st_ino, st_nlink, st_rdev, st_flags, st_gen. 1113 * XXX (st_dev, st_ino) should be unique. 1114 */ 1115 return 0; 1116} 1117 1118/* ARGSUSED */ 1119static int 1120pipe_close(fp, p) 1121 struct file *fp; 1122 struct proc *p; 1123{ 1124 struct pipe *cpipe = (struct pipe *)fp->f_data; 1125 1126 fp->f_ops = &badfileops; 1127 fp->f_data = NULL; 1128 funsetown(cpipe->pipe_sigio); 1129 pipeclose(cpipe); 1130 return 0; 1131} 1132 1133/* 1134 * shutdown the pipe 1135 */ 1136static void 1137pipeclose(cpipe) 1138 struct pipe *cpipe; 1139{ 1140 struct pipe *ppipe; 1141 if (cpipe) { 1142 1143 pipeselwakeup(cpipe); 1144 1145 /* 1146 * If the other side is blocked, wake it up saying that 1147 * we want to close it down. 1148 */ 1149 while (cpipe->pipe_busy) { 1150 wakeup(cpipe); 1151 cpipe->pipe_state |= PIPE_WANT|PIPE_EOF; 1152 tsleep(cpipe, PRIBIO, "pipecl", 0); 1153 } 1154 1155 /* 1156 * Disconnect from peer 1157 */ 1158 if ((ppipe = cpipe->pipe_peer) != NULL) { 1159 pipeselwakeup(ppipe); 1160 1161 ppipe->pipe_state |= PIPE_EOF; 1162 wakeup(ppipe); 1163 ppipe->pipe_peer = NULL; 1164 } 1165 1166 /* 1167 * free resources 1168 */ 1169 if (cpipe->pipe_buffer.buffer) { 1170 if (cpipe->pipe_buffer.size > PIPE_SIZE) 1171 --nbigpipe; 1172 amountpipekva -= cpipe->pipe_buffer.size; 1173 kmem_free(kernel_map, 1174 (vm_offset_t)cpipe->pipe_buffer.buffer, 1175 cpipe->pipe_buffer.size); 1176 } 1177#ifndef PIPE_NODIRECT 1178 if (cpipe->pipe_map.kva) { 1179 amountpipekva -= cpipe->pipe_buffer.size + PAGE_SIZE; 1180 kmem_free(kernel_map, 1181 cpipe->pipe_map.kva, 1182 cpipe->pipe_buffer.size + PAGE_SIZE); 1183 } 1184#endif 1185 zfree(pipe_zone, cpipe); 1186 } 1187} 1188 1189static int 1190filt_pipeattach(struct knote *kn) 1191{ 1192 struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data; 1193 1194 SLIST_INSERT_HEAD(&rpipe->pipe_sel.si_note, kn, kn_selnext); 1195 return (0); 1196} 1197 1198static void 1199filt_pipedetach(struct knote *kn) 1200{ 1201 struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data; 1202 1203 SLIST_REMOVE(&rpipe->pipe_sel.si_note, kn, knote, kn_selnext); 1204} 1205 1206/*ARGSUSED*/ 1207static int 1208filt_piperead(struct knote *kn, long hint) 1209{ 1210 struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data; 1211 struct pipe *wpipe = rpipe->pipe_peer; 1212 1213 kn->kn_data = rpipe->pipe_buffer.cnt; 1214 if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW)) 1215 kn->kn_data = rpipe->pipe_map.cnt; 1216 1217 if ((rpipe->pipe_state & PIPE_EOF) || 1218 (wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) { 1219 kn->kn_flags |= EV_EOF; 1220 return (1); 1221 } 1222 return (kn->kn_data > 0); 1223} 1224 1225/*ARGSUSED*/ 1226static int 1227filt_pipewrite(struct knote *kn, long hint) 1228{ 1229 struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data; 1230 struct pipe *wpipe = rpipe->pipe_peer; 1231 1232 if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) { 1233 kn->kn_data = 0; 1234 kn->kn_flags |= EV_EOF; 1235 return (1); 1236 } 1237 kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt; 1238 if (wpipe->pipe_state & PIPE_DIRECTW) 1239 kn->kn_data = 0; 1240 1241 return (kn->kn_data >= PIPE_BUF); 1242} 1243