1 /* $NetBSD: sys_pipe.c,v 1.65 2005/04/01 11:59:37 yamt Exp $ */ 2 3 /*- 4 * Copyright (c) 2003 The NetBSD Foundation, Inc. 5 * All rights reserved. 6 * 7 * This code is derived from software contributed to The NetBSD Foundation 8 * by Paul Kranenburg. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 3. All advertising materials mentioning features or use of this software 19 * must display the following acknowledgement: 20 * This product includes software developed by the NetBSD 21 * Foundation, Inc. and its contributors. 22 * 4. Neither the name of The NetBSD Foundation nor the names of its 23 * contributors may be used to endorse or promote products derived 24 * from this software without specific prior written permission. 25 * 26 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS 27 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 28 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 29 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS 30 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 31 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 32 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 33 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 34 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 35 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 36 * POSSIBILITY OF SUCH DAMAGE. 37 */ 38 39 /* 40 * Copyright (c) 1996 John S. Dyson 41 * All rights reserved. 42 * 43 * Redistribution and use in source and binary forms, with or without 44 * modification, are permitted provided that the following conditions 45 * are met: 46 * 1. Redistributions of source code must retain the above copyright 47 * notice immediately at the beginning of the file, without modification, 48 * this list of conditions, and the following disclaimer. 49 * 2. Redistributions in binary form must reproduce the above copyright 50 * notice, this list of conditions and the following disclaimer in the 51 * documentation and/or other materials provided with the distribution. 52 * 3. Absolutely no warranty of function or purpose is made by the author 53 * John S. Dyson. 54 * 4. Modifications may be freely made to this file if the above conditions 55 * are met. 56 * 57 * $FreeBSD: src/sys/kern/sys_pipe.c,v 1.95 2002/03/09 22:06:31 alfred Exp $ 58 */ 59 60 /* 61 * This file contains a high-performance replacement for the socket-based 62 * pipes scheme originally used in FreeBSD/4.4Lite. It does not support 63 * all features of sockets, but does do everything that pipes normally 64 * do. 65 * 66 * Adaption for NetBSD UVM, including uvm_loan() based direct write, was 67 * written by Jaromir Dolecek. 68 */ 69 70 /* 71 * This code has two modes of operation, a small write mode and a large 72 * write mode. The small write mode acts like conventional pipes with 73 * a kernel buffer. If the buffer is less than PIPE_MINDIRECT, then the 74 * "normal" pipe buffering is done. If the buffer is between PIPE_MINDIRECT 75 * and PIPE_SIZE in size it is mapped read-only into the kernel address space 76 * using the UVM page loan facility from where the receiving process can copy 77 * the data directly from the pages in the sending process. 78 * 79 * The constant PIPE_MINDIRECT is chosen to make sure that buffering will 80 * happen for small transfers so that the system will not spend all of 81 * its time context switching. PIPE_SIZE is constrained by the 82 * amount of kernel virtual memory. 83 */ 84 85 #include <sys/cdefs.h> 86 __KERNEL_RCSID(0, "$NetBSD: sys_pipe.c,v 1.65 2005/04/01 11:59:37 yamt Exp $"); 87 88 #include <sys/param.h> 89 #include <sys/systm.h> 90 #include <sys/proc.h> 91 #include <sys/fcntl.h> 92 #include <sys/file.h> 93 #include <sys/filedesc.h> 94 #include <sys/filio.h> 95 #include <sys/kernel.h> 96 #include <sys/lock.h> 97 #include <sys/ttycom.h> 98 #include <sys/stat.h> 99 #include <sys/malloc.h> 100 #include <sys/poll.h> 101 #include <sys/signalvar.h> 102 #include <sys/vnode.h> 103 #include <sys/uio.h> 104 #include <sys/lock.h> 105 #include <sys/select.h> 106 #include <sys/mount.h> 107 #include <sys/sa.h> 108 #include <sys/syscallargs.h> 109 #include <uvm/uvm.h> 110 #include <sys/sysctl.h> 111 #include <sys/kernel.h> 112 113 #include <sys/pipe.h> 114 115 /* 116 * Avoid microtime(9), it's slow. We don't guard the read from time(9) 117 * with splclock(9) since we don't actually need to be THAT sure the access 118 * is atomic. 119 */ 120 #define PIPE_TIMESTAMP(tvp) (*(tvp) = time) 121 122 123 /* 124 * Use this define if you want to disable *fancy* VM things. Expect an 125 * approx 30% decrease in transfer rate. 126 */ 127 /* #define PIPE_NODIRECT */ 128 129 /* 130 * interfaces to the outside world 131 */ 132 static int pipe_read(struct file *fp, off_t *offset, struct uio *uio, 133 struct ucred *cred, int flags); 134 static int pipe_write(struct file *fp, off_t *offset, struct uio *uio, 135 struct ucred *cred, int flags); 136 static int pipe_close(struct file *fp, struct proc *p); 137 static int pipe_poll(struct file *fp, int events, struct proc *p); 138 static int pipe_kqfilter(struct file *fp, struct knote *kn); 139 static int pipe_stat(struct file *fp, struct stat *sb, struct proc *p); 140 static int pipe_ioctl(struct file *fp, u_long cmd, void *data, 141 struct proc *p); 142 143 static const struct fileops pipeops = { 144 pipe_read, pipe_write, pipe_ioctl, fnullop_fcntl, pipe_poll, 145 pipe_stat, pipe_close, pipe_kqfilter 146 }; 147 148 /* 149 * Default pipe buffer size(s), this can be kind-of large now because pipe 150 * space is pageable. The pipe code will try to maintain locality of 151 * reference for performance reasons, so small amounts of outstanding I/O 152 * will not wipe the cache. 153 */ 154 #define MINPIPESIZE (PIPE_SIZE/3) 155 #define MAXPIPESIZE (2*PIPE_SIZE/3) 156 157 /* 158 * Maximum amount of kva for pipes -- this is kind-of a soft limit, but 159 * is there so that on large systems, we don't exhaust it. 160 */ 161 #define MAXPIPEKVA (8*1024*1024) 162 static int maxpipekva = MAXPIPEKVA; 163 164 /* 165 * Limit for direct transfers, we cannot, of course limit 166 * the amount of kva for pipes in general though. 167 */ 168 #define LIMITPIPEKVA (16*1024*1024) 169 static int limitpipekva = LIMITPIPEKVA; 170 171 /* 172 * Limit the number of "big" pipes 173 */ 174 #define LIMITBIGPIPES 32 175 static int maxbigpipes = LIMITBIGPIPES; 176 static int nbigpipe = 0; 177 178 /* 179 * Amount of KVA consumed by pipe buffers. 180 */ 181 static int amountpipekva = 0; 182 183 MALLOC_DEFINE(M_PIPE, "pipe", "Pipe structures"); 184 185 static void pipeclose(struct file *fp, struct pipe *pipe); 186 static void pipe_free_kmem(struct pipe *pipe); 187 static int pipe_create(struct pipe **pipep, int allockva); 188 static int pipelock(struct pipe *pipe, int catch); 189 static __inline void pipeunlock(struct pipe *pipe); 190 static void pipeselwakeup(struct pipe *pipe, struct pipe *sigp, void *data, 191 int code); 192 #ifndef PIPE_NODIRECT 193 static int pipe_direct_write(struct file *fp, struct pipe *wpipe, 194 struct uio *uio); 195 #endif 196 static int pipespace(struct pipe *pipe, int size); 197 198 #ifndef PIPE_NODIRECT 199 static int pipe_loan_alloc(struct pipe *, int); 200 static void pipe_loan_free(struct pipe *); 201 #endif /* PIPE_NODIRECT */ 202 203 static POOL_INIT(pipe_pool, sizeof(struct pipe), 0, 0, 0, "pipepl", 204 &pool_allocator_nointr); 205 206 /* 207 * The pipe system call for the DTYPE_PIPE type of pipes 208 */ 209 210 /* ARGSUSED */ 211 int 212 sys_pipe(l, v, retval) 213 struct lwp *l; 214 void *v; 215 register_t *retval; 216 { 217 struct file *rf, *wf; 218 struct pipe *rpipe, *wpipe; 219 int fd, error; 220 struct proc *p; 221 222 p = l->l_proc; 223 rpipe = wpipe = NULL; 224 if (pipe_create(&rpipe, 1) || pipe_create(&wpipe, 0)) { 225 pipeclose(NULL, rpipe); 226 pipeclose(NULL, wpipe); 227 return (ENFILE); 228 } 229 230 /* 231 * Note: the file structure returned from falloc() is marked 232 * as 'larval' initially. Unless we mark it as 'mature' by 233 * FILE_SET_MATURE(), any attempt to do anything with it would 234 * return EBADF, including e.g. dup(2) or close(2). This avoids 235 * file descriptor races if we block in the second falloc(). 236 */ 237 238 error = falloc(p, &rf, &fd); 239 if (error) 240 goto free2; 241 retval[0] = fd; 242 rf->f_flag = FREAD; 243 rf->f_type = DTYPE_PIPE; 244 rf->f_data = (caddr_t)rpipe; 245 rf->f_ops = &pipeops; 246 247 error = falloc(p, &wf, &fd); 248 if (error) 249 goto free3; 250 retval[1] = fd; 251 wf->f_flag = FWRITE; 252 wf->f_type = DTYPE_PIPE; 253 wf->f_data = (caddr_t)wpipe; 254 wf->f_ops = &pipeops; 255 256 rpipe->pipe_peer = wpipe; 257 wpipe->pipe_peer = rpipe; 258 259 FILE_SET_MATURE(rf); 260 FILE_SET_MATURE(wf); 261 FILE_UNUSE(rf, p); 262 FILE_UNUSE(wf, p); 263 return (0); 264 free3: 265 FILE_UNUSE(rf, p); 266 ffree(rf); 267 fdremove(p->p_fd, retval[0]); 268 free2: 269 pipeclose(NULL, wpipe); 270 pipeclose(NULL, rpipe); 271 272 return (error); 273 } 274 275 /* 276 * Allocate kva for pipe circular buffer, the space is pageable 277 * This routine will 'realloc' the size of a pipe safely, if it fails 278 * it will retain the old buffer. 279 * If it fails it will return ENOMEM. 280 */ 281 static int 282 pipespace(pipe, size) 283 struct pipe *pipe; 284 int size; 285 { 286 caddr_t buffer; 287 /* 288 * Allocate pageable virtual address space. Physical memory is 289 * allocated on demand. 290 */ 291 buffer = (caddr_t) uvm_km_alloc(kernel_map, round_page(size), 0, 292 UVM_KMF_PAGEABLE); 293 if (buffer == NULL) 294 return (ENOMEM); 295 296 /* free old resources if we're resizing */ 297 pipe_free_kmem(pipe); 298 pipe->pipe_buffer.buffer = buffer; 299 pipe->pipe_buffer.size = size; 300 pipe->pipe_buffer.in = 0; 301 pipe->pipe_buffer.out = 0; 302 pipe->pipe_buffer.cnt = 0; 303 amountpipekva += pipe->pipe_buffer.size; 304 return (0); 305 } 306 307 /* 308 * Initialize and allocate VM and memory for pipe. 309 */ 310 static int 311 pipe_create(pipep, allockva) 312 struct pipe **pipep; 313 int allockva; 314 { 315 struct pipe *pipe; 316 int error; 317 318 pipe = *pipep = pool_get(&pipe_pool, PR_WAITOK); 319 320 /* Initialize */ 321 memset(pipe, 0, sizeof(struct pipe)); 322 pipe->pipe_state = PIPE_SIGNALR; 323 324 PIPE_TIMESTAMP(&pipe->pipe_ctime); 325 pipe->pipe_atime = pipe->pipe_ctime; 326 pipe->pipe_mtime = pipe->pipe_ctime; 327 simple_lock_init(&pipe->pipe_slock); 328 lockinit(&pipe->pipe_lock, PSOCK | PCATCH, "pipelk", 0, 0); 329 330 if (allockva && (error = pipespace(pipe, PIPE_SIZE))) 331 return (error); 332 333 return (0); 334 } 335 336 337 /* 338 * Lock a pipe for I/O, blocking other access 339 * Called with pipe spin lock held. 340 * Return with pipe spin lock released on success. 341 */ 342 static int 343 pipelock(pipe, catch) 344 struct pipe *pipe; 345 int catch; 346 { 347 int error; 348 349 LOCK_ASSERT(simple_lock_held(&pipe->pipe_slock)); 350 351 while (1) { 352 error = lockmgr(&pipe->pipe_lock, LK_EXCLUSIVE | LK_INTERLOCK, 353 &pipe->pipe_slock); 354 if (error == 0) 355 break; 356 357 simple_lock(&pipe->pipe_slock); 358 if (catch || (error != EINTR && error != ERESTART)) 359 break; 360 /* 361 * XXX XXX XXX 362 * The pipe lock is initialised with PCATCH on and we cannot 363 * override this in a lockmgr() call. Thus a pending signal 364 * will cause lockmgr() to return with EINTR or ERESTART. 365 * We cannot simply re-enter lockmgr() at this point since 366 * the pending signals have not yet been posted and would 367 * cause an immediate EINTR/ERESTART return again. 368 * As a workaround we pause for a while here, giving the lock 369 * a chance to drain, before trying again. 370 * XXX XXX XXX 371 * 372 * NOTE: Consider dropping PCATCH from this lock; in practice 373 * it is never held for long enough periods for having it 374 * interruptable at the start of pipe_read/pipe_write to be 375 * beneficial. 376 */ 377 (void) ltsleep(&lbolt, PSOCK, "rstrtpipelock", hz, 378 &pipe->pipe_slock); 379 } 380 return (error); 381 } 382 383 /* 384 * unlock a pipe I/O lock 385 */ 386 static __inline void 387 pipeunlock(pipe) 388 struct pipe *pipe; 389 { 390 391 lockmgr(&pipe->pipe_lock, LK_RELEASE, NULL); 392 } 393 394 /* 395 * Select/poll wakup. This also sends SIGIO to peer connected to 396 * 'sigpipe' side of pipe. 397 */ 398 static void 399 pipeselwakeup(selp, sigp, data, code) 400 struct pipe *selp, *sigp; 401 void *data; 402 int code; 403 { 404 int band; 405 406 selnotify(&selp->pipe_sel, NOTE_SUBMIT); 407 408 if (sigp == NULL || (sigp->pipe_state & PIPE_ASYNC) == 0) 409 return; 410 411 switch (code) { 412 case POLL_IN: 413 band = POLLIN|POLLRDNORM; 414 break; 415 case POLL_OUT: 416 band = POLLOUT|POLLWRNORM; 417 break; 418 case POLL_HUP: 419 band = POLLHUP; 420 break; 421 #if POLL_HUP != POLL_ERR 422 case POLL_ERR: 423 band = POLLERR; 424 break; 425 #endif 426 default: 427 band = 0; 428 #ifdef DIAGNOSTIC 429 printf("bad siginfo code %d in pipe notification.\n", code); 430 #endif 431 break; 432 } 433 434 fownsignal(sigp->pipe_pgid, SIGIO, code, band, selp); 435 } 436 437 /* ARGSUSED */ 438 static int 439 pipe_read(fp, offset, uio, cred, flags) 440 struct file *fp; 441 off_t *offset; 442 struct uio *uio; 443 struct ucred *cred; 444 int flags; 445 { 446 struct pipe *rpipe = (struct pipe *) fp->f_data; 447 struct pipebuf *bp = &rpipe->pipe_buffer; 448 int error; 449 size_t nread = 0; 450 size_t size; 451 size_t ocnt; 452 453 PIPE_LOCK(rpipe); 454 ++rpipe->pipe_busy; 455 ocnt = bp->cnt; 456 457 again: 458 error = pipelock(rpipe, 1); 459 if (error) 460 goto unlocked_error; 461 462 while (uio->uio_resid) { 463 /* 464 * normal pipe buffer receive 465 */ 466 if (bp->cnt > 0) { 467 size = bp->size - bp->out; 468 if (size > bp->cnt) 469 size = bp->cnt; 470 if (size > uio->uio_resid) 471 size = uio->uio_resid; 472 473 error = uiomove(&bp->buffer[bp->out], size, uio); 474 if (error) 475 break; 476 477 bp->out += size; 478 if (bp->out >= bp->size) 479 bp->out = 0; 480 481 bp->cnt -= size; 482 483 /* 484 * If there is no more to read in the pipe, reset 485 * its pointers to the beginning. This improves 486 * cache hit stats. 487 */ 488 if (bp->cnt == 0) { 489 bp->in = 0; 490 bp->out = 0; 491 } 492 nread += size; 493 #ifndef PIPE_NODIRECT 494 } else if ((rpipe->pipe_state & PIPE_DIRECTR) != 0) { 495 /* 496 * Direct copy, bypassing a kernel buffer. 497 */ 498 caddr_t va; 499 500 KASSERT(rpipe->pipe_state & PIPE_DIRECTW); 501 502 size = rpipe->pipe_map.cnt; 503 if (size > uio->uio_resid) 504 size = uio->uio_resid; 505 506 va = (caddr_t) rpipe->pipe_map.kva + 507 rpipe->pipe_map.pos; 508 error = uiomove(va, size, uio); 509 if (error) 510 break; 511 nread += size; 512 rpipe->pipe_map.pos += size; 513 rpipe->pipe_map.cnt -= size; 514 if (rpipe->pipe_map.cnt == 0) { 515 PIPE_LOCK(rpipe); 516 rpipe->pipe_state &= ~PIPE_DIRECTR; 517 wakeup(rpipe); 518 PIPE_UNLOCK(rpipe); 519 } 520 #endif 521 } else { 522 /* 523 * Break if some data was read. 524 */ 525 if (nread > 0) 526 break; 527 528 PIPE_LOCK(rpipe); 529 530 /* 531 * detect EOF condition 532 * read returns 0 on EOF, no need to set error 533 */ 534 if (rpipe->pipe_state & PIPE_EOF) { 535 PIPE_UNLOCK(rpipe); 536 break; 537 } 538 539 /* 540 * don't block on non-blocking I/O 541 */ 542 if (fp->f_flag & FNONBLOCK) { 543 PIPE_UNLOCK(rpipe); 544 error = EAGAIN; 545 break; 546 } 547 548 /* 549 * Unlock the pipe buffer for our remaining processing. 550 * We will either break out with an error or we will 551 * sleep and relock to loop. 552 */ 553 pipeunlock(rpipe); 554 555 /* 556 * The PIPE_DIRECTR flag is not under the control 557 * of the long-term lock (see pipe_direct_write()), 558 * so re-check now while holding the spin lock. 559 */ 560 if ((rpipe->pipe_state & PIPE_DIRECTR) != 0) 561 goto again; 562 563 /* 564 * We want to read more, wake up select/poll. 565 */ 566 pipeselwakeup(rpipe, rpipe->pipe_peer, fp->f_data, 567 POLL_IN); 568 569 /* 570 * If the "write-side" is blocked, wake it up now. 571 */ 572 if (rpipe->pipe_state & PIPE_WANTW) { 573 rpipe->pipe_state &= ~PIPE_WANTW; 574 wakeup(rpipe); 575 } 576 577 /* Now wait until the pipe is filled */ 578 rpipe->pipe_state |= PIPE_WANTR; 579 error = ltsleep(rpipe, PSOCK | PCATCH, 580 "piperd", 0, &rpipe->pipe_slock); 581 if (error != 0) 582 goto unlocked_error; 583 goto again; 584 } 585 } 586 587 if (error == 0) 588 PIPE_TIMESTAMP(&rpipe->pipe_atime); 589 590 PIPE_LOCK(rpipe); 591 pipeunlock(rpipe); 592 593 unlocked_error: 594 --rpipe->pipe_busy; 595 596 /* 597 * PIPE_WANTCLOSE processing only makes sense if pipe_busy is 0. 598 */ 599 if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANTCLOSE)) { 600 rpipe->pipe_state &= ~(PIPE_WANTCLOSE|PIPE_WANTW); 601 wakeup(rpipe); 602 } else if (bp->cnt < MINPIPESIZE) { 603 /* 604 * Handle write blocking hysteresis. 605 */ 606 if (rpipe->pipe_state & PIPE_WANTW) { 607 rpipe->pipe_state &= ~PIPE_WANTW; 608 wakeup(rpipe); 609 } 610 } 611 612 /* 613 * If anything was read off the buffer, signal to the writer it's 614 * possible to write more data. Also send signal if we are here for the 615 * first time after last write. 616 */ 617 if ((bp->size - bp->cnt) >= PIPE_BUF 618 && (ocnt != bp->cnt || (rpipe->pipe_state & PIPE_SIGNALR))) { 619 pipeselwakeup(rpipe, rpipe->pipe_peer, fp->f_data, POLL_OUT); 620 rpipe->pipe_state &= ~PIPE_SIGNALR; 621 } 622 623 PIPE_UNLOCK(rpipe); 624 return (error); 625 } 626 627 #ifndef PIPE_NODIRECT 628 /* 629 * Allocate structure for loan transfer. 630 */ 631 static int 632 pipe_loan_alloc(wpipe, npages) 633 struct pipe *wpipe; 634 int npages; 635 { 636 vsize_t len; 637 638 len = (vsize_t)npages << PAGE_SHIFT; 639 wpipe->pipe_map.kva = uvm_km_alloc(kernel_map, len, 0, 640 UVM_KMF_VAONLY | UVM_KMF_WAITVA); 641 if (wpipe->pipe_map.kva == 0) 642 return (ENOMEM); 643 644 amountpipekva += len; 645 wpipe->pipe_map.npages = npages; 646 wpipe->pipe_map.pgs = malloc(npages * sizeof(struct vm_page *), M_PIPE, 647 M_WAITOK); 648 return (0); 649 } 650 651 /* 652 * Free resources allocated for loan transfer. 653 */ 654 static void 655 pipe_loan_free(wpipe) 656 struct pipe *wpipe; 657 { 658 vsize_t len; 659 660 len = (vsize_t)wpipe->pipe_map.npages << PAGE_SHIFT; 661 uvm_km_free(kernel_map, wpipe->pipe_map.kva, len, UVM_KMF_VAONLY); 662 wpipe->pipe_map.kva = 0; 663 amountpipekva -= len; 664 free(wpipe->pipe_map.pgs, M_PIPE); 665 wpipe->pipe_map.pgs = NULL; 666 } 667 668 /* 669 * NetBSD direct write, using uvm_loan() mechanism. 670 * This implements the pipe buffer write mechanism. Note that only 671 * a direct write OR a normal pipe write can be pending at any given time. 672 * If there are any characters in the pipe buffer, the direct write will 673 * be deferred until the receiving process grabs all of the bytes from 674 * the pipe buffer. Then the direct mapping write is set-up. 675 * 676 * Called with the long-term pipe lock held. 677 */ 678 static int 679 pipe_direct_write(fp, wpipe, uio) 680 struct file *fp; 681 struct pipe *wpipe; 682 struct uio *uio; 683 { 684 int error, npages, j; 685 struct vm_page **pgs; 686 vaddr_t bbase, kva, base, bend; 687 vsize_t blen, bcnt; 688 voff_t bpos; 689 690 KASSERT(wpipe->pipe_map.cnt == 0); 691 692 /* 693 * Handle first PIPE_CHUNK_SIZE bytes of buffer. Deal with buffers 694 * not aligned to PAGE_SIZE. 695 */ 696 bbase = (vaddr_t)uio->uio_iov->iov_base; 697 base = trunc_page(bbase); 698 bend = round_page(bbase + uio->uio_iov->iov_len); 699 blen = bend - base; 700 bpos = bbase - base; 701 702 if (blen > PIPE_DIRECT_CHUNK) { 703 blen = PIPE_DIRECT_CHUNK; 704 bend = base + blen; 705 bcnt = PIPE_DIRECT_CHUNK - bpos; 706 } else { 707 bcnt = uio->uio_iov->iov_len; 708 } 709 npages = blen >> PAGE_SHIFT; 710 711 /* 712 * Free the old kva if we need more pages than we have 713 * allocated. 714 */ 715 if (wpipe->pipe_map.kva != 0 && npages > wpipe->pipe_map.npages) 716 pipe_loan_free(wpipe); 717 718 /* Allocate new kva. */ 719 if (wpipe->pipe_map.kva == 0) { 720 error = pipe_loan_alloc(wpipe, npages); 721 if (error) 722 return (error); 723 } 724 725 /* Loan the write buffer memory from writer process */ 726 pgs = wpipe->pipe_map.pgs; 727 error = uvm_loan(&uio->uio_procp->p_vmspace->vm_map, base, blen, 728 pgs, UVM_LOAN_TOPAGE); 729 if (error) { 730 pipe_loan_free(wpipe); 731 return (ENOMEM); /* so that caller fallback to ordinary write */ 732 } 733 734 /* Enter the loaned pages to kva */ 735 kva = wpipe->pipe_map.kva; 736 for (j = 0; j < npages; j++, kva += PAGE_SIZE) { 737 pmap_kenter_pa(kva, VM_PAGE_TO_PHYS(pgs[j]), VM_PROT_READ); 738 } 739 pmap_update(pmap_kernel()); 740 741 /* Now we can put the pipe in direct write mode */ 742 wpipe->pipe_map.pos = bpos; 743 wpipe->pipe_map.cnt = bcnt; 744 wpipe->pipe_state |= PIPE_DIRECTW; 745 746 /* 747 * But before we can let someone do a direct read, 748 * we have to wait until the pipe is drained. 749 */ 750 751 /* Relase the pipe lock while we wait */ 752 PIPE_LOCK(wpipe); 753 pipeunlock(wpipe); 754 755 while (error == 0 && wpipe->pipe_buffer.cnt > 0) { 756 if (wpipe->pipe_state & PIPE_WANTR) { 757 wpipe->pipe_state &= ~PIPE_WANTR; 758 wakeup(wpipe); 759 } 760 761 wpipe->pipe_state |= PIPE_WANTW; 762 error = ltsleep(wpipe, PSOCK | PCATCH, "pipdwc", 0, 763 &wpipe->pipe_slock); 764 if (error == 0 && wpipe->pipe_state & PIPE_EOF) 765 error = EPIPE; 766 } 767 768 /* Pipe is drained; next read will off the direct buffer */ 769 wpipe->pipe_state |= PIPE_DIRECTR; 770 771 /* Wait until the reader is done */ 772 while (error == 0 && (wpipe->pipe_state & PIPE_DIRECTR)) { 773 if (wpipe->pipe_state & PIPE_WANTR) { 774 wpipe->pipe_state &= ~PIPE_WANTR; 775 wakeup(wpipe); 776 } 777 pipeselwakeup(wpipe, wpipe, fp->f_data, POLL_IN); 778 error = ltsleep(wpipe, PSOCK | PCATCH, "pipdwt", 0, 779 &wpipe->pipe_slock); 780 if (error == 0 && wpipe->pipe_state & PIPE_EOF) 781 error = EPIPE; 782 } 783 784 /* Take pipe out of direct write mode */ 785 wpipe->pipe_state &= ~(PIPE_DIRECTW | PIPE_DIRECTR); 786 787 /* Acquire the pipe lock and cleanup */ 788 (void)pipelock(wpipe, 0); 789 if (pgs != NULL) { 790 pmap_kremove(wpipe->pipe_map.kva, blen); 791 uvm_unloan(pgs, npages, UVM_LOAN_TOPAGE); 792 } 793 if (error || amountpipekva > maxpipekva) 794 pipe_loan_free(wpipe); 795 796 if (error) { 797 pipeselwakeup(wpipe, wpipe, fp->f_data, POLL_ERR); 798 799 /* 800 * If nothing was read from what we offered, return error 801 * straight on. Otherwise update uio resid first. Caller 802 * will deal with the error condition, returning short 803 * write, error, or restarting the write(2) as appropriate. 804 */ 805 if (wpipe->pipe_map.cnt == bcnt) { 806 wpipe->pipe_map.cnt = 0; 807 wakeup(wpipe); 808 return (error); 809 } 810 811 bcnt -= wpipe->pipe_map.cnt; 812 } 813 814 uio->uio_resid -= bcnt; 815 /* uio_offset not updated, not set/used for write(2) */ 816 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + bcnt; 817 uio->uio_iov->iov_len -= bcnt; 818 if (uio->uio_iov->iov_len == 0) { 819 uio->uio_iov++; 820 uio->uio_iovcnt--; 821 } 822 823 wpipe->pipe_map.cnt = 0; 824 return (error); 825 } 826 #endif /* !PIPE_NODIRECT */ 827 828 static int 829 pipe_write(fp, offset, uio, cred, flags) 830 struct file *fp; 831 off_t *offset; 832 struct uio *uio; 833 struct ucred *cred; 834 int flags; 835 { 836 struct pipe *wpipe, *rpipe; 837 struct pipebuf *bp; 838 int error; 839 840 /* We want to write to our peer */ 841 rpipe = (struct pipe *) fp->f_data; 842 843 retry: 844 error = 0; 845 PIPE_LOCK(rpipe); 846 wpipe = rpipe->pipe_peer; 847 848 /* 849 * Detect loss of pipe read side, issue SIGPIPE if lost. 850 */ 851 if (wpipe == NULL) 852 error = EPIPE; 853 else if (simple_lock_try(&wpipe->pipe_slock) == 0) { 854 /* Deal with race for peer */ 855 PIPE_UNLOCK(rpipe); 856 goto retry; 857 } else if ((wpipe->pipe_state & PIPE_EOF) != 0) { 858 PIPE_UNLOCK(wpipe); 859 error = EPIPE; 860 } 861 862 PIPE_UNLOCK(rpipe); 863 if (error != 0) 864 return (error); 865 866 ++wpipe->pipe_busy; 867 868 /* Aquire the long-term pipe lock */ 869 if ((error = pipelock(wpipe,1)) != 0) { 870 --wpipe->pipe_busy; 871 if (wpipe->pipe_busy == 0 872 && (wpipe->pipe_state & PIPE_WANTCLOSE)) { 873 wpipe->pipe_state &= ~(PIPE_WANTCLOSE | PIPE_WANTR); 874 wakeup(wpipe); 875 } 876 PIPE_UNLOCK(wpipe); 877 return (error); 878 } 879 880 bp = &wpipe->pipe_buffer; 881 882 /* 883 * If it is advantageous to resize the pipe buffer, do so. 884 */ 885 if ((uio->uio_resid > PIPE_SIZE) && 886 (nbigpipe < maxbigpipes) && 887 #ifndef PIPE_NODIRECT 888 (wpipe->pipe_state & PIPE_DIRECTW) == 0 && 889 #endif 890 (bp->size <= PIPE_SIZE) && (bp->cnt == 0)) { 891 892 if (pipespace(wpipe, BIG_PIPE_SIZE) == 0) 893 nbigpipe++; 894 } 895 896 while (uio->uio_resid) { 897 size_t space; 898 899 #ifndef PIPE_NODIRECT 900 /* 901 * Pipe buffered writes cannot be coincidental with 902 * direct writes. Also, only one direct write can be 903 * in progress at any one time. We wait until the currently 904 * executing direct write is completed before continuing. 905 * 906 * We break out if a signal occurs or the reader goes away. 907 */ 908 while (error == 0 && wpipe->pipe_state & PIPE_DIRECTW) { 909 PIPE_LOCK(wpipe); 910 if (wpipe->pipe_state & PIPE_WANTR) { 911 wpipe->pipe_state &= ~PIPE_WANTR; 912 wakeup(wpipe); 913 } 914 pipeunlock(wpipe); 915 error = ltsleep(wpipe, PSOCK | PCATCH, 916 "pipbww", 0, &wpipe->pipe_slock); 917 918 (void)pipelock(wpipe, 0); 919 if (wpipe->pipe_state & PIPE_EOF) 920 error = EPIPE; 921 } 922 if (error) 923 break; 924 925 /* 926 * If the transfer is large, we can gain performance if 927 * we do process-to-process copies directly. 928 * If the write is non-blocking, we don't use the 929 * direct write mechanism. 930 * 931 * The direct write mechanism will detect the reader going 932 * away on us. 933 */ 934 if ((uio->uio_iov->iov_len >= PIPE_MINDIRECT) && 935 (fp->f_flag & FNONBLOCK) == 0 && 936 (wpipe->pipe_map.kva || (amountpipekva < limitpipekva))) { 937 error = pipe_direct_write(fp, wpipe, uio); 938 939 /* 940 * Break out if error occurred, unless it's ENOMEM. 941 * ENOMEM means we failed to allocate some resources 942 * for direct write, so we just fallback to ordinary 943 * write. If the direct write was successful, 944 * process rest of data via ordinary write. 945 */ 946 if (error == 0) 947 continue; 948 949 if (error != ENOMEM) 950 break; 951 } 952 #endif /* PIPE_NODIRECT */ 953 954 space = bp->size - bp->cnt; 955 956 /* Writes of size <= PIPE_BUF must be atomic. */ 957 if ((space < uio->uio_resid) && (uio->uio_resid <= PIPE_BUF)) 958 space = 0; 959 960 if (space > 0) { 961 int size; /* Transfer size */ 962 int segsize; /* first segment to transfer */ 963 964 /* 965 * Transfer size is minimum of uio transfer 966 * and free space in pipe buffer. 967 */ 968 if (space > uio->uio_resid) 969 size = uio->uio_resid; 970 else 971 size = space; 972 /* 973 * First segment to transfer is minimum of 974 * transfer size and contiguous space in 975 * pipe buffer. If first segment to transfer 976 * is less than the transfer size, we've got 977 * a wraparound in the buffer. 978 */ 979 segsize = bp->size - bp->in; 980 if (segsize > size) 981 segsize = size; 982 983 /* Transfer first segment */ 984 error = uiomove(&bp->buffer[bp->in], segsize, uio); 985 986 if (error == 0 && segsize < size) { 987 /* 988 * Transfer remaining part now, to 989 * support atomic writes. Wraparound 990 * happened. 991 */ 992 #ifdef DEBUG 993 if (bp->in + segsize != bp->size) 994 panic("Expected pipe buffer wraparound disappeared"); 995 #endif 996 997 error = uiomove(&bp->buffer[0], 998 size - segsize, uio); 999 } 1000 if (error) 1001 break; 1002 1003 bp->in += size; 1004 if (bp->in >= bp->size) { 1005 #ifdef DEBUG 1006 if (bp->in != size - segsize + bp->size) 1007 panic("Expected wraparound bad"); 1008 #endif 1009 bp->in = size - segsize; 1010 } 1011 1012 bp->cnt += size; 1013 #ifdef DEBUG 1014 if (bp->cnt > bp->size) 1015 panic("Pipe buffer overflow"); 1016 #endif 1017 } else { 1018 /* 1019 * If the "read-side" has been blocked, wake it up now. 1020 */ 1021 PIPE_LOCK(wpipe); 1022 if (wpipe->pipe_state & PIPE_WANTR) { 1023 wpipe->pipe_state &= ~PIPE_WANTR; 1024 wakeup(wpipe); 1025 } 1026 PIPE_UNLOCK(wpipe); 1027 1028 /* 1029 * don't block on non-blocking I/O 1030 */ 1031 if (fp->f_flag & FNONBLOCK) { 1032 error = EAGAIN; 1033 break; 1034 } 1035 1036 /* 1037 * We have no more space and have something to offer, 1038 * wake up select/poll. 1039 */ 1040 if (bp->cnt) 1041 pipeselwakeup(wpipe, wpipe, fp->f_data, 1042 POLL_OUT); 1043 1044 PIPE_LOCK(wpipe); 1045 pipeunlock(wpipe); 1046 wpipe->pipe_state |= PIPE_WANTW; 1047 error = ltsleep(wpipe, PSOCK | PCATCH, "pipewr", 0, 1048 &wpipe->pipe_slock); 1049 (void)pipelock(wpipe, 0); 1050 if (error != 0) 1051 break; 1052 /* 1053 * If read side wants to go away, we just issue a signal 1054 * to ourselves. 1055 */ 1056 if (wpipe->pipe_state & PIPE_EOF) { 1057 error = EPIPE; 1058 break; 1059 } 1060 } 1061 } 1062 1063 PIPE_LOCK(wpipe); 1064 --wpipe->pipe_busy; 1065 if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANTCLOSE)) { 1066 wpipe->pipe_state &= ~(PIPE_WANTCLOSE | PIPE_WANTR); 1067 wakeup(wpipe); 1068 } else if (bp->cnt > 0) { 1069 /* 1070 * If we have put any characters in the buffer, we wake up 1071 * the reader. 1072 */ 1073 if (wpipe->pipe_state & PIPE_WANTR) { 1074 wpipe->pipe_state &= ~PIPE_WANTR; 1075 wakeup(wpipe); 1076 } 1077 } 1078 1079 /* 1080 * Don't return EPIPE if I/O was successful 1081 */ 1082 if (error == EPIPE && bp->cnt == 0 && uio->uio_resid == 0) 1083 error = 0; 1084 1085 if (error == 0) 1086 PIPE_TIMESTAMP(&wpipe->pipe_mtime); 1087 1088 /* 1089 * We have something to offer, wake up select/poll. 1090 * wpipe->pipe_map.cnt is always 0 in this point (direct write 1091 * is only done synchronously), so check only wpipe->pipe_buffer.cnt 1092 */ 1093 if (bp->cnt) 1094 pipeselwakeup(wpipe, wpipe, fp->f_data, POLL_OUT); 1095 1096 /* 1097 * Arrange for next read(2) to do a signal. 1098 */ 1099 wpipe->pipe_state |= PIPE_SIGNALR; 1100 1101 pipeunlock(wpipe); 1102 PIPE_UNLOCK(wpipe); 1103 return (error); 1104 } 1105 1106 /* 1107 * we implement a very minimal set of ioctls for compatibility with sockets. 1108 */ 1109 int 1110 pipe_ioctl(fp, cmd, data, p) 1111 struct file *fp; 1112 u_long cmd; 1113 void *data; 1114 struct proc *p; 1115 { 1116 struct pipe *pipe = (struct pipe *)fp->f_data; 1117 1118 switch (cmd) { 1119 1120 case FIONBIO: 1121 return (0); 1122 1123 case FIOASYNC: 1124 PIPE_LOCK(pipe); 1125 if (*(int *)data) { 1126 pipe->pipe_state |= PIPE_ASYNC; 1127 } else { 1128 pipe->pipe_state &= ~PIPE_ASYNC; 1129 } 1130 PIPE_UNLOCK(pipe); 1131 return (0); 1132 1133 case FIONREAD: 1134 PIPE_LOCK(pipe); 1135 #ifndef PIPE_NODIRECT 1136 if (pipe->pipe_state & PIPE_DIRECTW) 1137 *(int *)data = pipe->pipe_map.cnt; 1138 else 1139 #endif 1140 *(int *)data = pipe->pipe_buffer.cnt; 1141 PIPE_UNLOCK(pipe); 1142 return (0); 1143 1144 case FIONWRITE: 1145 /* Look at other side */ 1146 pipe = pipe->pipe_peer; 1147 PIPE_LOCK(pipe); 1148 #ifndef PIPE_NODIRECT 1149 if (pipe->pipe_state & PIPE_DIRECTW) 1150 *(int *)data = pipe->pipe_map.cnt; 1151 else 1152 #endif 1153 *(int *)data = pipe->pipe_buffer.cnt; 1154 PIPE_UNLOCK(pipe); 1155 return (0); 1156 1157 case FIONSPACE: 1158 /* Look at other side */ 1159 pipe = pipe->pipe_peer; 1160 PIPE_LOCK(pipe); 1161 #ifndef PIPE_NODIRECT 1162 /* 1163 * If we're in direct-mode, we don't really have a 1164 * send queue, and any other write will block. Thus 1165 * zero seems like the best answer. 1166 */ 1167 if (pipe->pipe_state & PIPE_DIRECTW) 1168 *(int *)data = 0; 1169 else 1170 #endif 1171 *(int *)data = pipe->pipe_buffer.size - 1172 pipe->pipe_buffer.cnt; 1173 PIPE_UNLOCK(pipe); 1174 return (0); 1175 1176 case TIOCSPGRP: 1177 case FIOSETOWN: 1178 return fsetown(p, &pipe->pipe_pgid, cmd, data); 1179 1180 case TIOCGPGRP: 1181 case FIOGETOWN: 1182 return fgetown(p, pipe->pipe_pgid, cmd, data); 1183 1184 } 1185 return (EPASSTHROUGH); 1186 } 1187 1188 int 1189 pipe_poll(fp, events, td) 1190 struct file *fp; 1191 int events; 1192 struct proc *td; 1193 { 1194 struct pipe *rpipe = (struct pipe *)fp->f_data; 1195 struct pipe *wpipe; 1196 int eof = 0; 1197 int revents = 0; 1198 1199 retry: 1200 PIPE_LOCK(rpipe); 1201 wpipe = rpipe->pipe_peer; 1202 if (wpipe != NULL && simple_lock_try(&wpipe->pipe_slock) == 0) { 1203 /* Deal with race for peer */ 1204 PIPE_UNLOCK(rpipe); 1205 goto retry; 1206 } 1207 1208 if (events & (POLLIN | POLLRDNORM)) 1209 if ((rpipe->pipe_buffer.cnt > 0) || 1210 #ifndef PIPE_NODIRECT 1211 (rpipe->pipe_state & PIPE_DIRECTR) || 1212 #endif 1213 (rpipe->pipe_state & PIPE_EOF)) 1214 revents |= events & (POLLIN | POLLRDNORM); 1215 1216 eof |= (rpipe->pipe_state & PIPE_EOF); 1217 PIPE_UNLOCK(rpipe); 1218 1219 if (wpipe == NULL) 1220 revents |= events & (POLLOUT | POLLWRNORM); 1221 else { 1222 if (events & (POLLOUT | POLLWRNORM)) 1223 if ((wpipe->pipe_state & PIPE_EOF) || ( 1224 #ifndef PIPE_NODIRECT 1225 (wpipe->pipe_state & PIPE_DIRECTW) == 0 && 1226 #endif 1227 (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF)) 1228 revents |= events & (POLLOUT | POLLWRNORM); 1229 1230 eof |= (wpipe->pipe_state & PIPE_EOF); 1231 PIPE_UNLOCK(wpipe); 1232 } 1233 1234 if (wpipe == NULL || eof) 1235 revents |= POLLHUP; 1236 1237 if (revents == 0) { 1238 if (events & (POLLIN | POLLRDNORM)) 1239 selrecord(td, &rpipe->pipe_sel); 1240 1241 if (events & (POLLOUT | POLLWRNORM)) 1242 selrecord(td, &wpipe->pipe_sel); 1243 } 1244 1245 return (revents); 1246 } 1247 1248 static int 1249 pipe_stat(fp, ub, td) 1250 struct file *fp; 1251 struct stat *ub; 1252 struct proc *td; 1253 { 1254 struct pipe *pipe = (struct pipe *)fp->f_data; 1255 1256 memset((caddr_t)ub, 0, sizeof(*ub)); 1257 ub->st_mode = S_IFIFO | S_IRUSR | S_IWUSR; 1258 ub->st_blksize = pipe->pipe_buffer.size; 1259 if (ub->st_blksize == 0 && pipe->pipe_peer) 1260 ub->st_blksize = pipe->pipe_peer->pipe_buffer.size; 1261 ub->st_size = pipe->pipe_buffer.cnt; 1262 ub->st_blocks = (ub->st_size) ? 1 : 0; 1263 TIMEVAL_TO_TIMESPEC(&pipe->pipe_atime, &ub->st_atimespec); 1264 TIMEVAL_TO_TIMESPEC(&pipe->pipe_mtime, &ub->st_mtimespec); 1265 TIMEVAL_TO_TIMESPEC(&pipe->pipe_ctime, &ub->st_ctimespec); 1266 ub->st_uid = fp->f_cred->cr_uid; 1267 ub->st_gid = fp->f_cred->cr_gid; 1268 /* 1269 * Left as 0: st_dev, st_ino, st_nlink, st_rdev, st_flags, st_gen. 1270 * XXX (st_dev, st_ino) should be unique. 1271 */ 1272 return (0); 1273 } 1274 1275 /* ARGSUSED */ 1276 static int 1277 pipe_close(fp, td) 1278 struct file *fp; 1279 struct proc *td; 1280 { 1281 struct pipe *pipe = (struct pipe *)fp->f_data; 1282 1283 fp->f_data = NULL; 1284 pipeclose(fp, pipe); 1285 return (0); 1286 } 1287 1288 static void 1289 pipe_free_kmem(pipe) 1290 struct pipe *pipe; 1291 { 1292 1293 if (pipe->pipe_buffer.buffer != NULL) { 1294 if (pipe->pipe_buffer.size > PIPE_SIZE) 1295 --nbigpipe; 1296 amountpipekva -= pipe->pipe_buffer.size; 1297 uvm_km_free(kernel_map, 1298 (vaddr_t)pipe->pipe_buffer.buffer, 1299 pipe->pipe_buffer.size, UVM_KMF_PAGEABLE); 1300 pipe->pipe_buffer.buffer = NULL; 1301 } 1302 #ifndef PIPE_NODIRECT 1303 if (pipe->pipe_map.kva != 0) { 1304 pipe_loan_free(pipe); 1305 pipe->pipe_map.cnt = 0; 1306 pipe->pipe_map.kva = 0; 1307 pipe->pipe_map.pos = 0; 1308 pipe->pipe_map.npages = 0; 1309 } 1310 #endif /* !PIPE_NODIRECT */ 1311 } 1312 1313 /* 1314 * shutdown the pipe 1315 */ 1316 static void 1317 pipeclose(fp, pipe) 1318 struct file *fp; 1319 struct pipe *pipe; 1320 { 1321 struct pipe *ppipe; 1322 1323 if (pipe == NULL) 1324 return; 1325 1326 retry: 1327 PIPE_LOCK(pipe); 1328 1329 if (fp) 1330 pipeselwakeup(pipe, pipe, fp->f_data, POLL_HUP); 1331 1332 /* 1333 * If the other side is blocked, wake it up saying that 1334 * we want to close it down. 1335 */ 1336 while (pipe->pipe_busy) { 1337 wakeup(pipe); 1338 pipe->pipe_state |= PIPE_WANTCLOSE | PIPE_EOF; 1339 ltsleep(pipe, PSOCK, "pipecl", 0, &pipe->pipe_slock); 1340 } 1341 1342 /* 1343 * Disconnect from peer 1344 */ 1345 if ((ppipe = pipe->pipe_peer) != NULL) { 1346 /* Deal with race for peer */ 1347 if (simple_lock_try(&ppipe->pipe_slock) == 0) { 1348 PIPE_UNLOCK(pipe); 1349 goto retry; 1350 } 1351 if (fp) 1352 pipeselwakeup(ppipe, ppipe, fp->f_data, POLL_HUP); 1353 1354 ppipe->pipe_state |= PIPE_EOF; 1355 wakeup(ppipe); 1356 ppipe->pipe_peer = NULL; 1357 PIPE_UNLOCK(ppipe); 1358 } 1359 1360 (void)lockmgr(&pipe->pipe_lock, LK_DRAIN | LK_INTERLOCK, 1361 &pipe->pipe_slock); 1362 1363 /* 1364 * free resources 1365 */ 1366 pipe_free_kmem(pipe); 1367 pool_put(&pipe_pool, pipe); 1368 } 1369 1370 static void 1371 filt_pipedetach(struct knote *kn) 1372 { 1373 struct pipe *pipe = (struct pipe *)kn->kn_fp->f_data; 1374 1375 switch(kn->kn_filter) { 1376 case EVFILT_WRITE: 1377 /* need the peer structure, not our own */ 1378 pipe = pipe->pipe_peer; 1379 /* XXXSMP: race for peer */ 1380 1381 /* if reader end already closed, just return */ 1382 if (pipe == NULL) 1383 return; 1384 1385 break; 1386 default: 1387 /* nothing to do */ 1388 break; 1389 } 1390 1391 #ifdef DIAGNOSTIC 1392 if (kn->kn_hook != pipe) 1393 panic("filt_pipedetach: inconsistent knote"); 1394 #endif 1395 1396 PIPE_LOCK(pipe); 1397 SLIST_REMOVE(&pipe->pipe_sel.sel_klist, kn, knote, kn_selnext); 1398 PIPE_UNLOCK(pipe); 1399 } 1400 1401 /*ARGSUSED*/ 1402 static int 1403 filt_piperead(struct knote *kn, long hint) 1404 { 1405 struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data; 1406 struct pipe *wpipe = rpipe->pipe_peer; 1407 1408 if ((hint & NOTE_SUBMIT) == 0) 1409 PIPE_LOCK(rpipe); 1410 kn->kn_data = rpipe->pipe_buffer.cnt; 1411 if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW)) 1412 kn->kn_data = rpipe->pipe_map.cnt; 1413 1414 /* XXXSMP: race for peer */ 1415 if ((rpipe->pipe_state & PIPE_EOF) || 1416 (wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) { 1417 kn->kn_flags |= EV_EOF; 1418 if ((hint & NOTE_SUBMIT) == 0) 1419 PIPE_UNLOCK(rpipe); 1420 return (1); 1421 } 1422 if ((hint & NOTE_SUBMIT) == 0) 1423 PIPE_UNLOCK(rpipe); 1424 return (kn->kn_data > 0); 1425 } 1426 1427 /*ARGSUSED*/ 1428 static int 1429 filt_pipewrite(struct knote *kn, long hint) 1430 { 1431 struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data; 1432 struct pipe *wpipe = rpipe->pipe_peer; 1433 1434 if ((hint & NOTE_SUBMIT) == 0) 1435 PIPE_LOCK(rpipe); 1436 /* XXXSMP: race for peer */ 1437 if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) { 1438 kn->kn_data = 0; 1439 kn->kn_flags |= EV_EOF; 1440 if ((hint & NOTE_SUBMIT) == 0) 1441 PIPE_UNLOCK(rpipe); 1442 return (1); 1443 } 1444 kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt; 1445 if (wpipe->pipe_state & PIPE_DIRECTW) 1446 kn->kn_data = 0; 1447 1448 if ((hint & NOTE_SUBMIT) == 0) 1449 PIPE_UNLOCK(rpipe); 1450 return (kn->kn_data >= PIPE_BUF); 1451 } 1452 1453 static const struct filterops pipe_rfiltops = 1454 { 1, NULL, filt_pipedetach, filt_piperead }; 1455 static const struct filterops pipe_wfiltops = 1456 { 1, NULL, filt_pipedetach, filt_pipewrite }; 1457 1458 /*ARGSUSED*/ 1459 static int 1460 pipe_kqfilter(struct file *fp, struct knote *kn) 1461 { 1462 struct pipe *pipe; 1463 1464 pipe = (struct pipe *)kn->kn_fp->f_data; 1465 switch (kn->kn_filter) { 1466 case EVFILT_READ: 1467 kn->kn_fop = &pipe_rfiltops; 1468 break; 1469 case EVFILT_WRITE: 1470 kn->kn_fop = &pipe_wfiltops; 1471 /* XXXSMP: race for peer */ 1472 pipe = pipe->pipe_peer; 1473 if (pipe == NULL) { 1474 /* other end of pipe has been closed */ 1475 return (EBADF); 1476 } 1477 break; 1478 default: 1479 return (1); 1480 } 1481 kn->kn_hook = pipe; 1482 1483 PIPE_LOCK(pipe); 1484 SLIST_INSERT_HEAD(&pipe->pipe_sel.sel_klist, kn, kn_selnext); 1485 PIPE_UNLOCK(pipe); 1486 return (0); 1487 } 1488 1489 /* 1490 * Handle pipe sysctls. 1491 */ 1492 SYSCTL_SETUP(sysctl_kern_pipe_setup, "sysctl kern.pipe subtree setup") 1493 { 1494 1495 sysctl_createv(clog, 0, NULL, NULL, 1496 CTLFLAG_PERMANENT, 1497 CTLTYPE_NODE, "kern", NULL, 1498 NULL, 0, NULL, 0, 1499 CTL_KERN, CTL_EOL); 1500 sysctl_createv(clog, 0, NULL, NULL, 1501 CTLFLAG_PERMANENT, 1502 CTLTYPE_NODE, "pipe", 1503 SYSCTL_DESCR("Pipe settings"), 1504 NULL, 0, NULL, 0, 1505 CTL_KERN, KERN_PIPE, CTL_EOL); 1506 1507 sysctl_createv(clog, 0, NULL, NULL, 1508 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 1509 CTLTYPE_INT, "maxkvasz", 1510 SYSCTL_DESCR("Maximum amount of kernel memory to be " 1511 "used for pipes"), 1512 NULL, 0, &maxpipekva, 0, 1513 CTL_KERN, KERN_PIPE, KERN_PIPE_MAXKVASZ, CTL_EOL); 1514 sysctl_createv(clog, 0, NULL, NULL, 1515 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 1516 CTLTYPE_INT, "maxloankvasz", 1517 SYSCTL_DESCR("Limit for direct transfers via page loan"), 1518 NULL, 0, &limitpipekva, 0, 1519 CTL_KERN, KERN_PIPE, KERN_PIPE_LIMITKVA, CTL_EOL); 1520 sysctl_createv(clog, 0, NULL, NULL, 1521 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 1522 CTLTYPE_INT, "maxbigpipes", 1523 SYSCTL_DESCR("Maximum number of \"big\" pipes"), 1524 NULL, 0, &maxbigpipes, 0, 1525 CTL_KERN, KERN_PIPE, KERN_PIPE_MAXBIGPIPES, CTL_EOL); 1526 sysctl_createv(clog, 0, NULL, NULL, 1527 CTLFLAG_PERMANENT, 1528 CTLTYPE_INT, "nbigpipes", 1529 SYSCTL_DESCR("Number of \"big\" pipes"), 1530 NULL, 0, &nbigpipe, 0, 1531 CTL_KERN, KERN_PIPE, KERN_PIPE_NBIGPIPES, CTL_EOL); 1532 sysctl_createv(clog, 0, NULL, NULL, 1533 CTLFLAG_PERMANENT, 1534 CTLTYPE_INT, "kvasize", 1535 SYSCTL_DESCR("Amount of kernel memory consumed by pipe " 1536 "buffers"), 1537 NULL, 0, &amountpipekva, 0, 1538 CTL_KERN, KERN_PIPE, KERN_PIPE_KVASIZE, CTL_EOL); 1539 } 1540