1 /* $NetBSD: sys_pipe.c,v 1.123 2009/12/12 21:28:04 dsl Exp $ */ 2 3 /*- 4 * Copyright (c) 2003, 2007, 2008, 2009 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, and by Andrew Doran. 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 * 19 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS 20 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 21 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 22 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS 23 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 24 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 25 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 26 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 27 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 28 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 29 * POSSIBILITY OF SUCH DAMAGE. 30 */ 31 32 /* 33 * Copyright (c) 1996 John S. Dyson 34 * All rights reserved. 35 * 36 * Redistribution and use in source and binary forms, with or without 37 * modification, are permitted provided that the following conditions 38 * are met: 39 * 1. Redistributions of source code must retain the above copyright 40 * notice immediately at the beginning of the file, without modification, 41 * this list of conditions, and the following disclaimer. 42 * 2. Redistributions in binary form must reproduce the above copyright 43 * notice, this list of conditions and the following disclaimer in the 44 * documentation and/or other materials provided with the distribution. 45 * 3. Absolutely no warranty of function or purpose is made by the author 46 * John S. Dyson. 47 * 4. Modifications may be freely made to this file if the above conditions 48 * are met. 49 */ 50 51 /* 52 * This file contains a high-performance replacement for the socket-based 53 * pipes scheme originally used. It does not support all features of 54 * sockets, but does do everything that pipes normally do. 55 * 56 * This code has two modes of operation, a small write mode and a large 57 * write mode. The small write mode acts like conventional pipes with 58 * a kernel buffer. If the buffer is less than PIPE_MINDIRECT, then the 59 * "normal" pipe buffering is done. If the buffer is between PIPE_MINDIRECT 60 * and PIPE_SIZE in size it is mapped read-only into the kernel address space 61 * using the UVM page loan facility from where the receiving process can copy 62 * the data directly from the pages in the sending process. 63 * 64 * The constant PIPE_MINDIRECT is chosen to make sure that buffering will 65 * happen for small transfers so that the system will not spend all of 66 * its time context switching. PIPE_SIZE is constrained by the 67 * amount of kernel virtual memory. 68 */ 69 70 #include <sys/cdefs.h> 71 __KERNEL_RCSID(0, "$NetBSD: sys_pipe.c,v 1.123 2009/12/12 21:28:04 dsl Exp $"); 72 73 #include <sys/param.h> 74 #include <sys/systm.h> 75 #include <sys/proc.h> 76 #include <sys/fcntl.h> 77 #include <sys/file.h> 78 #include <sys/filedesc.h> 79 #include <sys/filio.h> 80 #include <sys/kernel.h> 81 #include <sys/ttycom.h> 82 #include <sys/stat.h> 83 #include <sys/poll.h> 84 #include <sys/signalvar.h> 85 #include <sys/vnode.h> 86 #include <sys/uio.h> 87 #include <sys/select.h> 88 #include <sys/mount.h> 89 #include <sys/syscallargs.h> 90 #include <sys/sysctl.h> 91 #include <sys/kauth.h> 92 #include <sys/atomic.h> 93 #include <sys/pipe.h> 94 95 #include <uvm/uvm.h> 96 97 /* 98 * Use this to disable direct I/O and decrease the code size: 99 * #define PIPE_NODIRECT 100 */ 101 102 /* XXX Disabled for now; rare hangs switching between direct/buffered */ 103 #define PIPE_NODIRECT 104 105 static int pipe_read(file_t *, off_t *, struct uio *, kauth_cred_t, int); 106 static int pipe_write(file_t *, off_t *, struct uio *, kauth_cred_t, int); 107 static int pipe_close(file_t *); 108 static int pipe_poll(file_t *, int); 109 static int pipe_kqfilter(file_t *, struct knote *); 110 static int pipe_stat(file_t *, struct stat *); 111 static int pipe_ioctl(file_t *, u_long, void *); 112 static void pipe_abort(file_t *); 113 114 static const struct fileops pipeops = { 115 .fo_read = pipe_read, 116 .fo_write = pipe_write, 117 .fo_ioctl = pipe_ioctl, 118 .fo_fcntl = fnullop_fcntl, 119 .fo_poll = pipe_poll, 120 .fo_stat = pipe_stat, 121 .fo_close = pipe_close, 122 .fo_kqfilter = pipe_kqfilter, 123 .fo_abort = pipe_abort, 124 }; 125 126 /* 127 * Default pipe buffer size(s), this can be kind-of large now because pipe 128 * space is pageable. The pipe code will try to maintain locality of 129 * reference for performance reasons, so small amounts of outstanding I/O 130 * will not wipe the cache. 131 */ 132 #define MINPIPESIZE (PIPE_SIZE / 3) 133 #define MAXPIPESIZE (2 * PIPE_SIZE / 3) 134 135 /* 136 * Maximum amount of kva for pipes -- this is kind-of a soft limit, but 137 * is there so that on large systems, we don't exhaust it. 138 */ 139 #define MAXPIPEKVA (8 * 1024 * 1024) 140 static u_int maxpipekva = MAXPIPEKVA; 141 142 /* 143 * Limit for direct transfers, we cannot, of course limit 144 * the amount of kva for pipes in general though. 145 */ 146 #define LIMITPIPEKVA (16 * 1024 * 1024) 147 static u_int limitpipekva = LIMITPIPEKVA; 148 149 /* 150 * Limit the number of "big" pipes 151 */ 152 #define LIMITBIGPIPES 32 153 static u_int maxbigpipes = LIMITBIGPIPES; 154 static u_int nbigpipe = 0; 155 156 /* 157 * Amount of KVA consumed by pipe buffers. 158 */ 159 static u_int amountpipekva = 0; 160 161 static void pipeclose(struct pipe *); 162 static void pipe_free_kmem(struct pipe *); 163 static int pipe_create(struct pipe **, pool_cache_t); 164 static int pipelock(struct pipe *, int); 165 static inline void pipeunlock(struct pipe *); 166 static void pipeselwakeup(struct pipe *, struct pipe *, int); 167 #ifndef PIPE_NODIRECT 168 static int pipe_direct_write(file_t *, struct pipe *, struct uio *); 169 #endif 170 static int pipespace(struct pipe *, int); 171 static int pipe_ctor(void *, void *, int); 172 static void pipe_dtor(void *, void *); 173 174 #ifndef PIPE_NODIRECT 175 static int pipe_loan_alloc(struct pipe *, int); 176 static void pipe_loan_free(struct pipe *); 177 #endif /* PIPE_NODIRECT */ 178 179 static pool_cache_t pipe_wr_cache; 180 static pool_cache_t pipe_rd_cache; 181 182 void 183 pipe_init(void) 184 { 185 186 /* Writer side is not automatically allocated KVA. */ 187 pipe_wr_cache = pool_cache_init(sizeof(struct pipe), 0, 0, 0, "pipewr", 188 NULL, IPL_NONE, pipe_ctor, pipe_dtor, NULL); 189 KASSERT(pipe_wr_cache != NULL); 190 191 /* Reader side gets preallocated KVA. */ 192 pipe_rd_cache = pool_cache_init(sizeof(struct pipe), 0, 0, 0, "piperd", 193 NULL, IPL_NONE, pipe_ctor, pipe_dtor, (void *)1); 194 KASSERT(pipe_rd_cache != NULL); 195 } 196 197 static int 198 pipe_ctor(void *arg, void *obj, int flags) 199 { 200 struct pipe *pipe; 201 vaddr_t va; 202 203 pipe = obj; 204 205 memset(pipe, 0, sizeof(struct pipe)); 206 if (arg != NULL) { 207 /* Preallocate space. */ 208 va = uvm_km_alloc(kernel_map, PIPE_SIZE, 0, 209 UVM_KMF_PAGEABLE | UVM_KMF_WAITVA); 210 KASSERT(va != 0); 211 pipe->pipe_kmem = va; 212 atomic_add_int(&amountpipekva, PIPE_SIZE); 213 } 214 cv_init(&pipe->pipe_rcv, "piperd"); 215 cv_init(&pipe->pipe_wcv, "pipewr"); 216 cv_init(&pipe->pipe_draincv, "pipedrain"); 217 cv_init(&pipe->pipe_lkcv, "pipelk"); 218 selinit(&pipe->pipe_sel); 219 pipe->pipe_state = PIPE_SIGNALR; 220 221 return 0; 222 } 223 224 static void 225 pipe_dtor(void *arg, void *obj) 226 { 227 struct pipe *pipe; 228 229 pipe = obj; 230 231 cv_destroy(&pipe->pipe_rcv); 232 cv_destroy(&pipe->pipe_wcv); 233 cv_destroy(&pipe->pipe_draincv); 234 cv_destroy(&pipe->pipe_lkcv); 235 seldestroy(&pipe->pipe_sel); 236 if (pipe->pipe_kmem != 0) { 237 uvm_km_free(kernel_map, pipe->pipe_kmem, PIPE_SIZE, 238 UVM_KMF_PAGEABLE); 239 atomic_add_int(&amountpipekva, -PIPE_SIZE); 240 } 241 } 242 243 /* 244 * The pipe system call for the DTYPE_PIPE type of pipes 245 */ 246 int 247 sys_pipe(struct lwp *l, const void *v, register_t *retval) 248 { 249 struct pipe *rpipe, *wpipe; 250 file_t *rf, *wf; 251 int fd, error; 252 proc_t *p; 253 254 p = curproc; 255 rpipe = wpipe = NULL; 256 if (pipe_create(&rpipe, pipe_rd_cache) || 257 pipe_create(&wpipe, pipe_wr_cache)) { 258 error = ENOMEM; 259 goto free2; 260 } 261 rpipe->pipe_lock = mutex_obj_alloc(MUTEX_DEFAULT, IPL_NONE); 262 wpipe->pipe_lock = rpipe->pipe_lock; 263 mutex_obj_hold(wpipe->pipe_lock); 264 265 error = fd_allocfile(&rf, &fd); 266 if (error) 267 goto free2; 268 retval[0] = fd; 269 rf->f_flag = FREAD; 270 rf->f_type = DTYPE_PIPE; 271 rf->f_data = (void *)rpipe; 272 rf->f_ops = &pipeops; 273 274 error = fd_allocfile(&wf, &fd); 275 if (error) 276 goto free3; 277 retval[1] = fd; 278 wf->f_flag = FWRITE; 279 wf->f_type = DTYPE_PIPE; 280 wf->f_data = (void *)wpipe; 281 wf->f_ops = &pipeops; 282 283 rpipe->pipe_peer = wpipe; 284 wpipe->pipe_peer = rpipe; 285 286 fd_affix(p, rf, (int)retval[0]); 287 fd_affix(p, wf, (int)retval[1]); 288 return (0); 289 free3: 290 fd_abort(p, rf, (int)retval[0]); 291 free2: 292 pipeclose(wpipe); 293 pipeclose(rpipe); 294 295 return (error); 296 } 297 298 /* 299 * Allocate kva for pipe circular buffer, the space is pageable 300 * This routine will 'realloc' the size of a pipe safely, if it fails 301 * it will retain the old buffer. 302 * If it fails it will return ENOMEM. 303 */ 304 static int 305 pipespace(struct pipe *pipe, int size) 306 { 307 void *buffer; 308 309 /* 310 * Allocate pageable virtual address space. Physical memory is 311 * allocated on demand. 312 */ 313 if (size == PIPE_SIZE && pipe->pipe_kmem != 0) { 314 buffer = (void *)pipe->pipe_kmem; 315 } else { 316 buffer = (void *)uvm_km_alloc(kernel_map, round_page(size), 317 0, UVM_KMF_PAGEABLE); 318 if (buffer == NULL) 319 return (ENOMEM); 320 atomic_add_int(&amountpipekva, size); 321 } 322 323 /* free old resources if we're resizing */ 324 pipe_free_kmem(pipe); 325 pipe->pipe_buffer.buffer = buffer; 326 pipe->pipe_buffer.size = size; 327 pipe->pipe_buffer.in = 0; 328 pipe->pipe_buffer.out = 0; 329 pipe->pipe_buffer.cnt = 0; 330 return (0); 331 } 332 333 /* 334 * Initialize and allocate VM and memory for pipe. 335 */ 336 static int 337 pipe_create(struct pipe **pipep, pool_cache_t cache) 338 { 339 struct pipe *pipe; 340 int error; 341 342 pipe = pool_cache_get(cache, PR_WAITOK); 343 KASSERT(pipe != NULL); 344 *pipep = pipe; 345 error = 0; 346 getnanotime(&pipe->pipe_btime); 347 pipe->pipe_atime = pipe->pipe_mtime = pipe->pipe_btime; 348 pipe->pipe_lock = NULL; 349 if (cache == pipe_rd_cache) { 350 error = pipespace(pipe, PIPE_SIZE); 351 } else { 352 pipe->pipe_buffer.buffer = NULL; 353 pipe->pipe_buffer.size = 0; 354 pipe->pipe_buffer.in = 0; 355 pipe->pipe_buffer.out = 0; 356 pipe->pipe_buffer.cnt = 0; 357 } 358 return error; 359 } 360 361 /* 362 * Lock a pipe for I/O, blocking other access 363 * Called with pipe spin lock held. 364 */ 365 static int 366 pipelock(struct pipe *pipe, int catch) 367 { 368 int error; 369 370 KASSERT(mutex_owned(pipe->pipe_lock)); 371 372 while (pipe->pipe_state & PIPE_LOCKFL) { 373 pipe->pipe_state |= PIPE_LWANT; 374 if (catch) { 375 error = cv_wait_sig(&pipe->pipe_lkcv, pipe->pipe_lock); 376 if (error != 0) 377 return error; 378 } else 379 cv_wait(&pipe->pipe_lkcv, pipe->pipe_lock); 380 } 381 382 pipe->pipe_state |= PIPE_LOCKFL; 383 384 return 0; 385 } 386 387 /* 388 * unlock a pipe I/O lock 389 */ 390 static inline void 391 pipeunlock(struct pipe *pipe) 392 { 393 394 KASSERT(pipe->pipe_state & PIPE_LOCKFL); 395 396 pipe->pipe_state &= ~PIPE_LOCKFL; 397 if (pipe->pipe_state & PIPE_LWANT) { 398 pipe->pipe_state &= ~PIPE_LWANT; 399 cv_broadcast(&pipe->pipe_lkcv); 400 } 401 } 402 403 /* 404 * Select/poll wakup. This also sends SIGIO to peer connected to 405 * 'sigpipe' side of pipe. 406 */ 407 static void 408 pipeselwakeup(struct pipe *selp, struct pipe *sigp, int code) 409 { 410 int band; 411 412 switch (code) { 413 case POLL_IN: 414 band = POLLIN|POLLRDNORM; 415 break; 416 case POLL_OUT: 417 band = POLLOUT|POLLWRNORM; 418 break; 419 case POLL_HUP: 420 band = POLLHUP; 421 break; 422 case POLL_ERR: 423 band = POLLERR; 424 break; 425 default: 426 band = 0; 427 #ifdef DIAGNOSTIC 428 printf("bad siginfo code %d in pipe notification.\n", code); 429 #endif 430 break; 431 } 432 433 selnotify(&selp->pipe_sel, band, NOTE_SUBMIT); 434 435 if (sigp == NULL || (sigp->pipe_state & PIPE_ASYNC) == 0) 436 return; 437 438 fownsignal(sigp->pipe_pgid, SIGIO, code, band, selp); 439 } 440 441 static int 442 pipe_read(file_t *fp, off_t *offset, struct uio *uio, kauth_cred_t cred, 443 int flags) 444 { 445 struct pipe *rpipe = (struct pipe *) fp->f_data; 446 struct pipebuf *bp = &rpipe->pipe_buffer; 447 kmutex_t *lock = rpipe->pipe_lock; 448 int error; 449 size_t nread = 0; 450 size_t size; 451 size_t ocnt; 452 453 mutex_enter(lock); 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 mutex_exit(lock); 474 error = uiomove((char *)bp->buffer + bp->out, size, uio); 475 mutex_enter(lock); 476 if (error) 477 break; 478 479 bp->out += size; 480 if (bp->out >= bp->size) 481 bp->out = 0; 482 483 bp->cnt -= size; 484 485 /* 486 * If there is no more to read in the pipe, reset 487 * its pointers to the beginning. This improves 488 * cache hit stats. 489 */ 490 if (bp->cnt == 0) { 491 bp->in = 0; 492 bp->out = 0; 493 } 494 nread += size; 495 continue; 496 } 497 498 #ifndef PIPE_NODIRECT 499 if ((rpipe->pipe_state & PIPE_DIRECTR) != 0) { 500 /* 501 * Direct copy, bypassing a kernel buffer. 502 */ 503 void *va; 504 u_int gen; 505 506 KASSERT(rpipe->pipe_state & PIPE_DIRECTW); 507 508 size = rpipe->pipe_map.cnt; 509 if (size > uio->uio_resid) 510 size = uio->uio_resid; 511 512 va = (char *)rpipe->pipe_map.kva + rpipe->pipe_map.pos; 513 gen = rpipe->pipe_map.egen; 514 mutex_exit(lock); 515 516 /* 517 * Consume emap and read the data from loaned pages. 518 */ 519 uvm_emap_consume(gen); 520 error = uiomove(va, size, uio); 521 522 mutex_enter(lock); 523 if (error) 524 break; 525 nread += size; 526 rpipe->pipe_map.pos += size; 527 rpipe->pipe_map.cnt -= size; 528 if (rpipe->pipe_map.cnt == 0) { 529 rpipe->pipe_state &= ~PIPE_DIRECTR; 530 cv_broadcast(&rpipe->pipe_wcv); 531 } 532 continue; 533 } 534 #endif 535 /* 536 * Break if some data was read. 537 */ 538 if (nread > 0) 539 break; 540 541 /* 542 * Detect EOF condition. 543 * Read returns 0 on EOF, no need to set error. 544 */ 545 if (rpipe->pipe_state & (PIPE_EOF | PIPE_ABORTED)) { 546 if (rpipe->pipe_state & PIPE_ABORTED) 547 /* Another thread has called close() */ 548 error = EBADF; 549 break; 550 } 551 552 /* 553 * Don't block on non-blocking I/O. 554 */ 555 if (fp->f_flag & FNONBLOCK) { 556 error = EAGAIN; 557 break; 558 } 559 560 /* 561 * Unlock the pipe buffer for our remaining processing. 562 * We will either break out with an error or we will 563 * sleep and relock to loop. 564 */ 565 pipeunlock(rpipe); 566 567 /* 568 * Re-check to see if more direct writes are pending. 569 */ 570 if ((rpipe->pipe_state & PIPE_DIRECTR) != 0) 571 goto again; 572 573 /* 574 * We want to read more, wake up select/poll. 575 */ 576 pipeselwakeup(rpipe, rpipe->pipe_peer, POLL_OUT); 577 578 /* 579 * If the "write-side" is blocked, wake it up now. 580 */ 581 cv_broadcast(&rpipe->pipe_wcv); 582 583 /* Now wait until the pipe is filled */ 584 error = cv_wait_sig(&rpipe->pipe_rcv, lock); 585 if (error != 0) 586 goto unlocked_error; 587 goto again; 588 } 589 590 if (error == 0) 591 getnanotime(&rpipe->pipe_atime); 592 pipeunlock(rpipe); 593 594 unlocked_error: 595 --rpipe->pipe_busy; 596 if (rpipe->pipe_busy == 0) { 597 cv_broadcast(&rpipe->pipe_draincv); 598 } 599 if (bp->cnt < MINPIPESIZE) { 600 cv_broadcast(&rpipe->pipe_wcv); 601 } 602 603 /* 604 * If anything was read off the buffer, signal to the writer it's 605 * possible to write more data. Also send signal if we are here for the 606 * first time after last write. 607 */ 608 if ((bp->size - bp->cnt) >= PIPE_BUF 609 && (ocnt != bp->cnt || (rpipe->pipe_state & PIPE_SIGNALR))) { 610 pipeselwakeup(rpipe, rpipe->pipe_peer, POLL_OUT); 611 rpipe->pipe_state &= ~PIPE_SIGNALR; 612 } 613 614 mutex_exit(lock); 615 return (error); 616 } 617 618 #ifndef PIPE_NODIRECT 619 /* 620 * Allocate structure for loan transfer. 621 */ 622 static int 623 pipe_loan_alloc(struct pipe *wpipe, int npages) 624 { 625 vsize_t len; 626 627 len = (vsize_t)npages << PAGE_SHIFT; 628 atomic_add_int(&amountpipekva, len); 629 wpipe->pipe_map.kva = uvm_km_alloc(kernel_map, len, 0, 630 UVM_KMF_VAONLY | UVM_KMF_WAITVA); 631 if (wpipe->pipe_map.kva == 0) { 632 atomic_add_int(&amountpipekva, -len); 633 return (ENOMEM); 634 } 635 636 wpipe->pipe_map.npages = npages; 637 wpipe->pipe_map.pgs = kmem_alloc(npages * sizeof(struct vm_page *), 638 KM_SLEEP); 639 return (0); 640 } 641 642 /* 643 * Free resources allocated for loan transfer. 644 */ 645 static void 646 pipe_loan_free(struct pipe *wpipe) 647 { 648 vsize_t len; 649 650 len = (vsize_t)wpipe->pipe_map.npages << PAGE_SHIFT; 651 uvm_emap_remove(wpipe->pipe_map.kva, len); /* XXX */ 652 uvm_km_free(kernel_map, wpipe->pipe_map.kva, len, UVM_KMF_VAONLY); 653 wpipe->pipe_map.kva = 0; 654 atomic_add_int(&amountpipekva, -len); 655 kmem_free(wpipe->pipe_map.pgs, 656 wpipe->pipe_map.npages * sizeof(struct vm_page *)); 657 wpipe->pipe_map.pgs = NULL; 658 } 659 660 /* 661 * NetBSD direct write, using uvm_loan() mechanism. 662 * This implements the pipe buffer write mechanism. Note that only 663 * a direct write OR a normal pipe write can be pending at any given time. 664 * If there are any characters in the pipe buffer, the direct write will 665 * be deferred until the receiving process grabs all of the bytes from 666 * the pipe buffer. Then the direct mapping write is set-up. 667 * 668 * Called with the long-term pipe lock held. 669 */ 670 static int 671 pipe_direct_write(file_t *fp, struct pipe *wpipe, struct uio *uio) 672 { 673 struct vm_page **pgs; 674 vaddr_t bbase, base, bend; 675 vsize_t blen, bcnt; 676 int error, npages; 677 voff_t bpos; 678 kmutex_t *lock = wpipe->pipe_lock; 679 680 KASSERT(mutex_owned(wpipe->pipe_lock)); 681 KASSERT(wpipe->pipe_map.cnt == 0); 682 683 mutex_exit(lock); 684 685 /* 686 * Handle first PIPE_CHUNK_SIZE bytes of buffer. Deal with buffers 687 * not aligned to PAGE_SIZE. 688 */ 689 bbase = (vaddr_t)uio->uio_iov->iov_base; 690 base = trunc_page(bbase); 691 bend = round_page(bbase + uio->uio_iov->iov_len); 692 blen = bend - base; 693 bpos = bbase - base; 694 695 if (blen > PIPE_DIRECT_CHUNK) { 696 blen = PIPE_DIRECT_CHUNK; 697 bend = base + blen; 698 bcnt = PIPE_DIRECT_CHUNK - bpos; 699 } else { 700 bcnt = uio->uio_iov->iov_len; 701 } 702 npages = blen >> PAGE_SHIFT; 703 704 /* 705 * Free the old kva if we need more pages than we have 706 * allocated. 707 */ 708 if (wpipe->pipe_map.kva != 0 && npages > wpipe->pipe_map.npages) 709 pipe_loan_free(wpipe); 710 711 /* Allocate new kva. */ 712 if (wpipe->pipe_map.kva == 0) { 713 error = pipe_loan_alloc(wpipe, npages); 714 if (error) { 715 mutex_enter(lock); 716 return (error); 717 } 718 } 719 720 /* Loan the write buffer memory from writer process */ 721 pgs = wpipe->pipe_map.pgs; 722 error = uvm_loan(&uio->uio_vmspace->vm_map, base, blen, 723 pgs, UVM_LOAN_TOPAGE); 724 if (error) { 725 pipe_loan_free(wpipe); 726 mutex_enter(lock); 727 return (ENOMEM); /* so that caller fallback to ordinary write */ 728 } 729 730 /* Enter the loaned pages to KVA, produce new emap generation number. */ 731 uvm_emap_enter(wpipe->pipe_map.kva, pgs, npages); 732 wpipe->pipe_map.egen = uvm_emap_produce(); 733 734 /* Now we can put the pipe in direct write mode */ 735 wpipe->pipe_map.pos = bpos; 736 wpipe->pipe_map.cnt = bcnt; 737 738 /* 739 * But before we can let someone do a direct read, we 740 * have to wait until the pipe is drained. Release the 741 * pipe lock while we wait. 742 */ 743 mutex_enter(lock); 744 wpipe->pipe_state |= PIPE_DIRECTW; 745 pipeunlock(wpipe); 746 747 while (error == 0 && wpipe->pipe_buffer.cnt > 0) { 748 cv_broadcast(&wpipe->pipe_rcv); 749 error = cv_wait_sig(&wpipe->pipe_wcv, lock); 750 if (error == 0 && wpipe->pipe_state & PIPE_EOF) 751 error = EPIPE; 752 } 753 754 /* Pipe is drained; next read will off the direct buffer */ 755 wpipe->pipe_state |= PIPE_DIRECTR; 756 757 /* Wait until the reader is done */ 758 while (error == 0 && (wpipe->pipe_state & PIPE_DIRECTR)) { 759 cv_broadcast(&wpipe->pipe_rcv); 760 pipeselwakeup(wpipe, wpipe, POLL_IN); 761 error = cv_wait_sig(&wpipe->pipe_wcv, lock); 762 if (error == 0 && wpipe->pipe_state & PIPE_EOF) 763 error = EPIPE; 764 } 765 766 /* Take pipe out of direct write mode */ 767 wpipe->pipe_state &= ~(PIPE_DIRECTW | PIPE_DIRECTR); 768 769 /* Acquire the pipe lock and cleanup */ 770 (void)pipelock(wpipe, 0); 771 mutex_exit(lock); 772 773 if (pgs != NULL) { 774 /* XXX: uvm_emap_remove */ 775 uvm_unloan(pgs, npages, UVM_LOAN_TOPAGE); 776 } 777 if (error || amountpipekva > maxpipekva) 778 pipe_loan_free(wpipe); 779 780 mutex_enter(lock); 781 if (error) { 782 pipeselwakeup(wpipe, wpipe, POLL_ERR); 783 784 /* 785 * If nothing was read from what we offered, return error 786 * straight on. Otherwise update uio resid first. Caller 787 * will deal with the error condition, returning short 788 * write, error, or restarting the write(2) as appropriate. 789 */ 790 if (wpipe->pipe_map.cnt == bcnt) { 791 wpipe->pipe_map.cnt = 0; 792 cv_broadcast(&wpipe->pipe_wcv); 793 return (error); 794 } 795 796 bcnt -= wpipe->pipe_map.cnt; 797 } 798 799 uio->uio_resid -= bcnt; 800 /* uio_offset not updated, not set/used for write(2) */ 801 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + bcnt; 802 uio->uio_iov->iov_len -= bcnt; 803 if (uio->uio_iov->iov_len == 0) { 804 uio->uio_iov++; 805 uio->uio_iovcnt--; 806 } 807 808 wpipe->pipe_map.cnt = 0; 809 return (error); 810 } 811 #endif /* !PIPE_NODIRECT */ 812 813 static int 814 pipe_write(file_t *fp, off_t *offset, struct uio *uio, kauth_cred_t cred, 815 int flags) 816 { 817 struct pipe *wpipe, *rpipe; 818 struct pipebuf *bp; 819 kmutex_t *lock; 820 int error; 821 822 /* We want to write to our peer */ 823 rpipe = (struct pipe *) fp->f_data; 824 lock = rpipe->pipe_lock; 825 error = 0; 826 827 mutex_enter(lock); 828 wpipe = rpipe->pipe_peer; 829 830 /* 831 * Detect loss of pipe read side, issue SIGPIPE if lost. 832 */ 833 if (wpipe == NULL || (wpipe->pipe_state & PIPE_EOF) != 0) { 834 mutex_exit(lock); 835 return EPIPE; 836 } 837 ++wpipe->pipe_busy; 838 839 /* Aquire the long-term pipe lock */ 840 if ((error = pipelock(wpipe, 1)) != 0) { 841 --wpipe->pipe_busy; 842 if (wpipe->pipe_busy == 0) { 843 cv_broadcast(&wpipe->pipe_draincv); 844 } 845 mutex_exit(lock); 846 return (error); 847 } 848 849 bp = &wpipe->pipe_buffer; 850 851 /* 852 * If it is advantageous to resize the pipe buffer, do so. 853 */ 854 if ((uio->uio_resid > PIPE_SIZE) && 855 (nbigpipe < maxbigpipes) && 856 #ifndef PIPE_NODIRECT 857 (wpipe->pipe_state & PIPE_DIRECTW) == 0 && 858 #endif 859 (bp->size <= PIPE_SIZE) && (bp->cnt == 0)) { 860 861 if (pipespace(wpipe, BIG_PIPE_SIZE) == 0) 862 atomic_inc_uint(&nbigpipe); 863 } 864 865 while (uio->uio_resid) { 866 size_t space; 867 868 #ifndef PIPE_NODIRECT 869 /* 870 * Pipe buffered writes cannot be coincidental with 871 * direct writes. Also, only one direct write can be 872 * in progress at any one time. We wait until the currently 873 * executing direct write is completed before continuing. 874 * 875 * We break out if a signal occurs or the reader goes away. 876 */ 877 while (error == 0 && wpipe->pipe_state & PIPE_DIRECTW) { 878 cv_broadcast(&wpipe->pipe_rcv); 879 pipeunlock(wpipe); 880 error = cv_wait_sig(&wpipe->pipe_wcv, lock); 881 (void)pipelock(wpipe, 0); 882 if (wpipe->pipe_state & PIPE_EOF) 883 error = EPIPE; 884 } 885 if (error) 886 break; 887 888 /* 889 * If the transfer is large, we can gain performance if 890 * we do process-to-process copies directly. 891 * If the write is non-blocking, we don't use the 892 * direct write mechanism. 893 * 894 * The direct write mechanism will detect the reader going 895 * away on us. 896 */ 897 if ((uio->uio_iov->iov_len >= PIPE_MINDIRECT) && 898 (fp->f_flag & FNONBLOCK) == 0 && 899 (wpipe->pipe_map.kva || (amountpipekva < limitpipekva))) { 900 error = pipe_direct_write(fp, wpipe, uio); 901 902 /* 903 * Break out if error occurred, unless it's ENOMEM. 904 * ENOMEM means we failed to allocate some resources 905 * for direct write, so we just fallback to ordinary 906 * write. If the direct write was successful, 907 * process rest of data via ordinary write. 908 */ 909 if (error == 0) 910 continue; 911 912 if (error != ENOMEM) 913 break; 914 } 915 #endif /* PIPE_NODIRECT */ 916 917 space = bp->size - bp->cnt; 918 919 /* Writes of size <= PIPE_BUF must be atomic. */ 920 if ((space < uio->uio_resid) && (uio->uio_resid <= PIPE_BUF)) 921 space = 0; 922 923 if (space > 0) { 924 int size; /* Transfer size */ 925 int segsize; /* first segment to transfer */ 926 927 /* 928 * Transfer size is minimum of uio transfer 929 * and free space in pipe buffer. 930 */ 931 if (space > uio->uio_resid) 932 size = uio->uio_resid; 933 else 934 size = space; 935 /* 936 * First segment to transfer is minimum of 937 * transfer size and contiguous space in 938 * pipe buffer. If first segment to transfer 939 * is less than the transfer size, we've got 940 * a wraparound in the buffer. 941 */ 942 segsize = bp->size - bp->in; 943 if (segsize > size) 944 segsize = size; 945 946 /* Transfer first segment */ 947 mutex_exit(lock); 948 error = uiomove((char *)bp->buffer + bp->in, segsize, 949 uio); 950 951 if (error == 0 && segsize < size) { 952 /* 953 * Transfer remaining part now, to 954 * support atomic writes. Wraparound 955 * happened. 956 */ 957 KASSERT(bp->in + segsize == bp->size); 958 error = uiomove(bp->buffer, 959 size - segsize, uio); 960 } 961 mutex_enter(lock); 962 if (error) 963 break; 964 965 bp->in += size; 966 if (bp->in >= bp->size) { 967 KASSERT(bp->in == size - segsize + bp->size); 968 bp->in = size - segsize; 969 } 970 971 bp->cnt += size; 972 KASSERT(bp->cnt <= bp->size); 973 } else { 974 /* 975 * If the "read-side" has been blocked, wake it up now. 976 */ 977 cv_broadcast(&wpipe->pipe_rcv); 978 979 /* 980 * Don't block on non-blocking I/O. 981 */ 982 if (fp->f_flag & FNONBLOCK) { 983 error = EAGAIN; 984 break; 985 } 986 987 if (wpipe->pipe_state & PIPE_ABORTED) { 988 /* Another thread has called close() */ 989 if (uio->uio_resid == 0) 990 error = EBADF; 991 break; 992 } 993 994 /* 995 * We have no more space and have something to offer, 996 * wake up select/poll. 997 */ 998 if (bp->cnt) 999 pipeselwakeup(wpipe, wpipe, POLL_IN); 1000 1001 pipeunlock(wpipe); 1002 error = cv_wait_sig(&wpipe->pipe_wcv, lock); 1003 (void)pipelock(wpipe, 0); 1004 if (error != 0) 1005 break; 1006 /* 1007 * If read side wants to go away, we just issue a signal 1008 * to ourselves. 1009 */ 1010 if (wpipe->pipe_state & PIPE_EOF) { 1011 error = EPIPE; 1012 break; 1013 } 1014 } 1015 } 1016 1017 --wpipe->pipe_busy; 1018 if (wpipe->pipe_busy == 0) { 1019 cv_broadcast(&wpipe->pipe_draincv); 1020 } 1021 if (bp->cnt > 0) { 1022 cv_broadcast(&wpipe->pipe_rcv); 1023 } 1024 1025 /* 1026 * Don't return EPIPE if I/O was successful 1027 */ 1028 if (error == EPIPE && bp->cnt == 0 && uio->uio_resid == 0) 1029 error = 0; 1030 1031 if (error == 0) 1032 getnanotime(&wpipe->pipe_mtime); 1033 1034 /* 1035 * We have something to offer, wake up select/poll. 1036 * wpipe->pipe_map.cnt is always 0 in this point (direct write 1037 * is only done synchronously), so check only wpipe->pipe_buffer.cnt 1038 */ 1039 if (bp->cnt) 1040 pipeselwakeup(wpipe, wpipe, POLL_IN); 1041 1042 /* 1043 * Arrange for next read(2) to do a signal. 1044 */ 1045 wpipe->pipe_state |= PIPE_SIGNALR; 1046 1047 pipeunlock(wpipe); 1048 mutex_exit(lock); 1049 return (error); 1050 } 1051 1052 /* 1053 * We implement a very minimal set of ioctls for compatibility with sockets. 1054 */ 1055 int 1056 pipe_ioctl(file_t *fp, u_long cmd, void *data) 1057 { 1058 struct pipe *pipe = fp->f_data; 1059 kmutex_t *lock = pipe->pipe_lock; 1060 1061 switch (cmd) { 1062 1063 case FIONBIO: 1064 return (0); 1065 1066 case FIOASYNC: 1067 mutex_enter(lock); 1068 if (*(int *)data) { 1069 pipe->pipe_state |= PIPE_ASYNC; 1070 } else { 1071 pipe->pipe_state &= ~PIPE_ASYNC; 1072 } 1073 mutex_exit(lock); 1074 return (0); 1075 1076 case FIONREAD: 1077 mutex_enter(lock); 1078 #ifndef PIPE_NODIRECT 1079 if (pipe->pipe_state & PIPE_DIRECTW) 1080 *(int *)data = pipe->pipe_map.cnt; 1081 else 1082 #endif 1083 *(int *)data = pipe->pipe_buffer.cnt; 1084 mutex_exit(lock); 1085 return (0); 1086 1087 case FIONWRITE: 1088 /* Look at other side */ 1089 pipe = pipe->pipe_peer; 1090 mutex_enter(lock); 1091 #ifndef PIPE_NODIRECT 1092 if (pipe->pipe_state & PIPE_DIRECTW) 1093 *(int *)data = pipe->pipe_map.cnt; 1094 else 1095 #endif 1096 *(int *)data = pipe->pipe_buffer.cnt; 1097 mutex_exit(lock); 1098 return (0); 1099 1100 case FIONSPACE: 1101 /* Look at other side */ 1102 pipe = pipe->pipe_peer; 1103 mutex_enter(lock); 1104 #ifndef PIPE_NODIRECT 1105 /* 1106 * If we're in direct-mode, we don't really have a 1107 * send queue, and any other write will block. Thus 1108 * zero seems like the best answer. 1109 */ 1110 if (pipe->pipe_state & PIPE_DIRECTW) 1111 *(int *)data = 0; 1112 else 1113 #endif 1114 *(int *)data = pipe->pipe_buffer.size - 1115 pipe->pipe_buffer.cnt; 1116 mutex_exit(lock); 1117 return (0); 1118 1119 case TIOCSPGRP: 1120 case FIOSETOWN: 1121 return fsetown(&pipe->pipe_pgid, cmd, data); 1122 1123 case TIOCGPGRP: 1124 case FIOGETOWN: 1125 return fgetown(pipe->pipe_pgid, cmd, data); 1126 1127 } 1128 return (EPASSTHROUGH); 1129 } 1130 1131 int 1132 pipe_poll(file_t *fp, int events) 1133 { 1134 struct pipe *rpipe = fp->f_data; 1135 struct pipe *wpipe; 1136 int eof = 0; 1137 int revents = 0; 1138 1139 mutex_enter(rpipe->pipe_lock); 1140 wpipe = rpipe->pipe_peer; 1141 1142 if (events & (POLLIN | POLLRDNORM)) 1143 if ((rpipe->pipe_buffer.cnt > 0) || 1144 #ifndef PIPE_NODIRECT 1145 (rpipe->pipe_state & PIPE_DIRECTR) || 1146 #endif 1147 (rpipe->pipe_state & PIPE_EOF)) 1148 revents |= events & (POLLIN | POLLRDNORM); 1149 1150 eof |= (rpipe->pipe_state & PIPE_EOF); 1151 1152 if (wpipe == NULL) 1153 revents |= events & (POLLOUT | POLLWRNORM); 1154 else { 1155 if (events & (POLLOUT | POLLWRNORM)) 1156 if ((wpipe->pipe_state & PIPE_EOF) || ( 1157 #ifndef PIPE_NODIRECT 1158 (wpipe->pipe_state & PIPE_DIRECTW) == 0 && 1159 #endif 1160 (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF)) 1161 revents |= events & (POLLOUT | POLLWRNORM); 1162 1163 eof |= (wpipe->pipe_state & PIPE_EOF); 1164 } 1165 1166 if (wpipe == NULL || eof) 1167 revents |= POLLHUP; 1168 1169 if (revents == 0) { 1170 if (events & (POLLIN | POLLRDNORM)) 1171 selrecord(curlwp, &rpipe->pipe_sel); 1172 1173 if (events & (POLLOUT | POLLWRNORM)) 1174 selrecord(curlwp, &wpipe->pipe_sel); 1175 } 1176 mutex_exit(rpipe->pipe_lock); 1177 1178 return (revents); 1179 } 1180 1181 static int 1182 pipe_stat(file_t *fp, struct stat *ub) 1183 { 1184 struct pipe *pipe = fp->f_data; 1185 1186 mutex_enter(pipe->pipe_lock); 1187 memset(ub, 0, sizeof(*ub)); 1188 ub->st_mode = S_IFIFO | S_IRUSR | S_IWUSR; 1189 ub->st_blksize = pipe->pipe_buffer.size; 1190 if (ub->st_blksize == 0 && pipe->pipe_peer) 1191 ub->st_blksize = pipe->pipe_peer->pipe_buffer.size; 1192 ub->st_size = pipe->pipe_buffer.cnt; 1193 ub->st_blocks = (ub->st_size) ? 1 : 0; 1194 ub->st_atimespec = pipe->pipe_atime; 1195 ub->st_mtimespec = pipe->pipe_mtime; 1196 ub->st_ctimespec = ub->st_birthtimespec = pipe->pipe_btime; 1197 ub->st_uid = kauth_cred_geteuid(fp->f_cred); 1198 ub->st_gid = kauth_cred_getegid(fp->f_cred); 1199 1200 /* 1201 * Left as 0: st_dev, st_ino, st_nlink, st_rdev, st_flags, st_gen. 1202 * XXX (st_dev, st_ino) should be unique. 1203 */ 1204 mutex_exit(pipe->pipe_lock); 1205 return 0; 1206 } 1207 1208 static int 1209 pipe_close(file_t *fp) 1210 { 1211 struct pipe *pipe = fp->f_data; 1212 1213 fp->f_data = NULL; 1214 pipeclose(pipe); 1215 return (0); 1216 } 1217 1218 static void 1219 pipe_abort(file_t *fp) 1220 { 1221 struct pipe *pipe = fp->f_data; 1222 1223 /* Unblock blocked reads/writes - they will return EBADF. */ 1224 mutex_enter(pipe->pipe_lock); 1225 pipe->pipe_state |= PIPE_ABORTED; 1226 cv_broadcast(&pipe->pipe_rcv); 1227 cv_broadcast(&pipe->pipe_wcv); 1228 mutex_exit(pipe->pipe_lock); 1229 } 1230 1231 static void 1232 pipe_free_kmem(struct pipe *pipe) 1233 { 1234 1235 if (pipe->pipe_buffer.buffer != NULL) { 1236 if (pipe->pipe_buffer.size > PIPE_SIZE) { 1237 atomic_dec_uint(&nbigpipe); 1238 } 1239 if (pipe->pipe_buffer.buffer != (void *)pipe->pipe_kmem) { 1240 uvm_km_free(kernel_map, 1241 (vaddr_t)pipe->pipe_buffer.buffer, 1242 pipe->pipe_buffer.size, UVM_KMF_PAGEABLE); 1243 atomic_add_int(&amountpipekva, 1244 -pipe->pipe_buffer.size); 1245 } 1246 pipe->pipe_buffer.buffer = NULL; 1247 } 1248 #ifndef PIPE_NODIRECT 1249 if (pipe->pipe_map.kva != 0) { 1250 pipe_loan_free(pipe); 1251 pipe->pipe_map.cnt = 0; 1252 pipe->pipe_map.kva = 0; 1253 pipe->pipe_map.pos = 0; 1254 pipe->pipe_map.npages = 0; 1255 } 1256 #endif /* !PIPE_NODIRECT */ 1257 } 1258 1259 /* 1260 * Shutdown the pipe. 1261 */ 1262 static void 1263 pipeclose(struct pipe *pipe) 1264 { 1265 kmutex_t *lock; 1266 struct pipe *ppipe; 1267 1268 if (pipe == NULL) 1269 return; 1270 1271 KASSERT(cv_is_valid(&pipe->pipe_rcv)); 1272 KASSERT(cv_is_valid(&pipe->pipe_wcv)); 1273 KASSERT(cv_is_valid(&pipe->pipe_draincv)); 1274 KASSERT(cv_is_valid(&pipe->pipe_lkcv)); 1275 1276 lock = pipe->pipe_lock; 1277 if (lock == NULL) 1278 /* Must have failed during create */ 1279 goto free_resources; 1280 1281 mutex_enter(lock); 1282 pipeselwakeup(pipe, pipe, POLL_HUP); 1283 1284 /* 1285 * If the other side is blocked, wake it up saying that 1286 * we want to close it down. 1287 */ 1288 pipe->pipe_state |= PIPE_EOF; 1289 if (pipe->pipe_busy) { 1290 while (pipe->pipe_busy) { 1291 cv_broadcast(&pipe->pipe_wcv); 1292 cv_wait_sig(&pipe->pipe_draincv, lock); 1293 } 1294 } 1295 1296 /* 1297 * Disconnect from peer. 1298 */ 1299 if ((ppipe = pipe->pipe_peer) != NULL) { 1300 pipeselwakeup(ppipe, ppipe, POLL_HUP); 1301 ppipe->pipe_state |= PIPE_EOF; 1302 cv_broadcast(&ppipe->pipe_rcv); 1303 ppipe->pipe_peer = NULL; 1304 } 1305 1306 /* 1307 * Any knote objects still left in the list are 1308 * the one attached by peer. Since no one will 1309 * traverse this list, we just clear it. 1310 */ 1311 SLIST_INIT(&pipe->pipe_sel.sel_klist); 1312 1313 KASSERT((pipe->pipe_state & PIPE_LOCKFL) == 0); 1314 mutex_exit(lock); 1315 mutex_obj_free(lock); 1316 1317 /* 1318 * Free resources. 1319 */ 1320 free_resources: 1321 pipe->pipe_pgid = 0; 1322 pipe->pipe_state = PIPE_SIGNALR; 1323 pipe_free_kmem(pipe); 1324 if (pipe->pipe_kmem != 0) { 1325 pool_cache_put(pipe_rd_cache, pipe); 1326 } else { 1327 pool_cache_put(pipe_wr_cache, pipe); 1328 } 1329 } 1330 1331 static void 1332 filt_pipedetach(struct knote *kn) 1333 { 1334 struct pipe *pipe; 1335 kmutex_t *lock; 1336 1337 pipe = ((file_t *)kn->kn_obj)->f_data; 1338 lock = pipe->pipe_lock; 1339 1340 mutex_enter(lock); 1341 1342 switch(kn->kn_filter) { 1343 case EVFILT_WRITE: 1344 /* Need the peer structure, not our own. */ 1345 pipe = pipe->pipe_peer; 1346 1347 /* If reader end already closed, just return. */ 1348 if (pipe == NULL) { 1349 mutex_exit(lock); 1350 return; 1351 } 1352 1353 break; 1354 default: 1355 /* Nothing to do. */ 1356 break; 1357 } 1358 1359 KASSERT(kn->kn_hook == pipe); 1360 SLIST_REMOVE(&pipe->pipe_sel.sel_klist, kn, knote, kn_selnext); 1361 mutex_exit(lock); 1362 } 1363 1364 static int 1365 filt_piperead(struct knote *kn, long hint) 1366 { 1367 struct pipe *rpipe = ((file_t *)kn->kn_obj)->f_data; 1368 struct pipe *wpipe; 1369 1370 if ((hint & NOTE_SUBMIT) == 0) { 1371 mutex_enter(rpipe->pipe_lock); 1372 } 1373 wpipe = rpipe->pipe_peer; 1374 kn->kn_data = rpipe->pipe_buffer.cnt; 1375 1376 if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW)) 1377 kn->kn_data = rpipe->pipe_map.cnt; 1378 1379 if ((rpipe->pipe_state & PIPE_EOF) || 1380 (wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) { 1381 kn->kn_flags |= EV_EOF; 1382 if ((hint & NOTE_SUBMIT) == 0) { 1383 mutex_exit(rpipe->pipe_lock); 1384 } 1385 return (1); 1386 } 1387 1388 if ((hint & NOTE_SUBMIT) == 0) { 1389 mutex_exit(rpipe->pipe_lock); 1390 } 1391 return (kn->kn_data > 0); 1392 } 1393 1394 static int 1395 filt_pipewrite(struct knote *kn, long hint) 1396 { 1397 struct pipe *rpipe = ((file_t *)kn->kn_obj)->f_data; 1398 struct pipe *wpipe; 1399 1400 if ((hint & NOTE_SUBMIT) == 0) { 1401 mutex_enter(rpipe->pipe_lock); 1402 } 1403 wpipe = rpipe->pipe_peer; 1404 1405 if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) { 1406 kn->kn_data = 0; 1407 kn->kn_flags |= EV_EOF; 1408 if ((hint & NOTE_SUBMIT) == 0) { 1409 mutex_exit(rpipe->pipe_lock); 1410 } 1411 return (1); 1412 } 1413 kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt; 1414 if (wpipe->pipe_state & PIPE_DIRECTW) 1415 kn->kn_data = 0; 1416 1417 if ((hint & NOTE_SUBMIT) == 0) { 1418 mutex_exit(rpipe->pipe_lock); 1419 } 1420 return (kn->kn_data >= PIPE_BUF); 1421 } 1422 1423 static const struct filterops pipe_rfiltops = 1424 { 1, NULL, filt_pipedetach, filt_piperead }; 1425 static const struct filterops pipe_wfiltops = 1426 { 1, NULL, filt_pipedetach, filt_pipewrite }; 1427 1428 static int 1429 pipe_kqfilter(file_t *fp, struct knote *kn) 1430 { 1431 struct pipe *pipe; 1432 kmutex_t *lock; 1433 1434 pipe = ((file_t *)kn->kn_obj)->f_data; 1435 lock = pipe->pipe_lock; 1436 1437 mutex_enter(lock); 1438 1439 switch (kn->kn_filter) { 1440 case EVFILT_READ: 1441 kn->kn_fop = &pipe_rfiltops; 1442 break; 1443 case EVFILT_WRITE: 1444 kn->kn_fop = &pipe_wfiltops; 1445 pipe = pipe->pipe_peer; 1446 if (pipe == NULL) { 1447 /* Other end of pipe has been closed. */ 1448 mutex_exit(lock); 1449 return (EBADF); 1450 } 1451 break; 1452 default: 1453 mutex_exit(lock); 1454 return (EINVAL); 1455 } 1456 1457 kn->kn_hook = pipe; 1458 SLIST_INSERT_HEAD(&pipe->pipe_sel.sel_klist, kn, kn_selnext); 1459 mutex_exit(lock); 1460 1461 return (0); 1462 } 1463 1464 /* 1465 * Handle pipe sysctls. 1466 */ 1467 SYSCTL_SETUP(sysctl_kern_pipe_setup, "sysctl kern.pipe subtree setup") 1468 { 1469 1470 sysctl_createv(clog, 0, NULL, NULL, 1471 CTLFLAG_PERMANENT, 1472 CTLTYPE_NODE, "kern", NULL, 1473 NULL, 0, NULL, 0, 1474 CTL_KERN, CTL_EOL); 1475 sysctl_createv(clog, 0, NULL, NULL, 1476 CTLFLAG_PERMANENT, 1477 CTLTYPE_NODE, "pipe", 1478 SYSCTL_DESCR("Pipe settings"), 1479 NULL, 0, NULL, 0, 1480 CTL_KERN, KERN_PIPE, CTL_EOL); 1481 1482 sysctl_createv(clog, 0, NULL, NULL, 1483 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 1484 CTLTYPE_INT, "maxkvasz", 1485 SYSCTL_DESCR("Maximum amount of kernel memory to be " 1486 "used for pipes"), 1487 NULL, 0, &maxpipekva, 0, 1488 CTL_KERN, KERN_PIPE, KERN_PIPE_MAXKVASZ, CTL_EOL); 1489 sysctl_createv(clog, 0, NULL, NULL, 1490 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 1491 CTLTYPE_INT, "maxloankvasz", 1492 SYSCTL_DESCR("Limit for direct transfers via page loan"), 1493 NULL, 0, &limitpipekva, 0, 1494 CTL_KERN, KERN_PIPE, KERN_PIPE_LIMITKVA, CTL_EOL); 1495 sysctl_createv(clog, 0, NULL, NULL, 1496 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 1497 CTLTYPE_INT, "maxbigpipes", 1498 SYSCTL_DESCR("Maximum number of \"big\" pipes"), 1499 NULL, 0, &maxbigpipes, 0, 1500 CTL_KERN, KERN_PIPE, KERN_PIPE_MAXBIGPIPES, CTL_EOL); 1501 sysctl_createv(clog, 0, NULL, NULL, 1502 CTLFLAG_PERMANENT, 1503 CTLTYPE_INT, "nbigpipes", 1504 SYSCTL_DESCR("Number of \"big\" pipes"), 1505 NULL, 0, &nbigpipe, 0, 1506 CTL_KERN, KERN_PIPE, KERN_PIPE_NBIGPIPES, CTL_EOL); 1507 sysctl_createv(clog, 0, NULL, NULL, 1508 CTLFLAG_PERMANENT, 1509 CTLTYPE_INT, "kvasize", 1510 SYSCTL_DESCR("Amount of kernel memory consumed by pipe " 1511 "buffers"), 1512 NULL, 0, &amountpipekva, 0, 1513 CTL_KERN, KERN_PIPE, KERN_PIPE_KVASIZE, CTL_EOL); 1514 } 1515