1 /* $NetBSD: sys_pipe.c,v 1.125 2009/12/13 20:02:23 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 1.125 2009/12/13 20:02:23 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 int slept = 0; 453 454 mutex_enter(lock); 455 ++rpipe->pipe_busy; 456 ocnt = bp->cnt; 457 458 again: 459 error = pipelock(rpipe, 1); 460 if (error) 461 goto unlocked_error; 462 463 while (uio->uio_resid) { 464 /* 465 * Normal pipe buffer receive. 466 */ 467 if (bp->cnt > 0) { 468 size = bp->size - bp->out; 469 if (size > bp->cnt) 470 size = bp->cnt; 471 if (size > uio->uio_resid) 472 size = uio->uio_resid; 473 474 mutex_exit(lock); 475 error = uiomove((char *)bp->buffer + bp->out, size, uio); 476 mutex_enter(lock); 477 if (error) 478 break; 479 480 bp->out += size; 481 if (bp->out >= bp->size) 482 bp->out = 0; 483 484 bp->cnt -= size; 485 486 /* 487 * If there is no more to read in the pipe, reset 488 * its pointers to the beginning. This improves 489 * cache hit stats. 490 */ 491 if (bp->cnt == 0) { 492 bp->in = 0; 493 bp->out = 0; 494 } 495 nread += size; 496 continue; 497 } 498 499 #ifndef PIPE_NODIRECT 500 if ((rpipe->pipe_state & PIPE_DIRECTR) != 0) { 501 /* 502 * Direct copy, bypassing a kernel buffer. 503 */ 504 void *va; 505 u_int gen; 506 507 KASSERT(rpipe->pipe_state & PIPE_DIRECTW); 508 509 size = rpipe->pipe_map.cnt; 510 if (size > uio->uio_resid) 511 size = uio->uio_resid; 512 513 va = (char *)rpipe->pipe_map.kva + rpipe->pipe_map.pos; 514 gen = rpipe->pipe_map.egen; 515 mutex_exit(lock); 516 517 /* 518 * Consume emap and read the data from loaned pages. 519 */ 520 uvm_emap_consume(gen); 521 error = uiomove(va, size, uio); 522 523 mutex_enter(lock); 524 if (error) 525 break; 526 nread += size; 527 rpipe->pipe_map.pos += size; 528 rpipe->pipe_map.cnt -= size; 529 if (rpipe->pipe_map.cnt == 0) { 530 rpipe->pipe_state &= ~PIPE_DIRECTR; 531 cv_broadcast(&rpipe->pipe_wcv); 532 } 533 continue; 534 } 535 #endif 536 /* 537 * Break if some data was read. 538 */ 539 if (nread > 0) 540 break; 541 542 /* 543 * Detect EOF condition. 544 * Read returns 0 on EOF, no need to set error. 545 */ 546 if (rpipe->pipe_state & PIPE_EOF) 547 break; 548 549 /* 550 * Don't block on non-blocking I/O. 551 */ 552 if (fp->f_flag & FNONBLOCK) { 553 error = EAGAIN; 554 break; 555 } 556 557 /* 558 * Unlock the pipe buffer for our remaining processing. 559 * We will either break out with an error or we will 560 * sleep and relock to loop. 561 */ 562 pipeunlock(rpipe); 563 564 /* 565 * Re-check to see if more direct writes are pending. 566 */ 567 if ((rpipe->pipe_state & PIPE_DIRECTR) != 0) 568 goto again; 569 570 #if 1 /* XXX (dsl) I'm sure these aren't needed here ... */ 571 /* 572 * We want to read more, wake up select/poll. 573 */ 574 pipeselwakeup(rpipe, rpipe->pipe_peer, POLL_OUT); 575 576 /* 577 * If the "write-side" is blocked, wake it up now. 578 */ 579 cv_broadcast(&rpipe->pipe_wcv); 580 #endif 581 582 if (slept) { 583 error = ERESTART; 584 goto unlocked_error; 585 } 586 587 /* Now wait until the pipe is filled */ 588 error = cv_wait_sig(&rpipe->pipe_rcv, lock); 589 if (error != 0) 590 goto unlocked_error; 591 slept = 1; 592 goto again; 593 } 594 595 if (error == 0) 596 getnanotime(&rpipe->pipe_atime); 597 pipeunlock(rpipe); 598 599 unlocked_error: 600 --rpipe->pipe_busy; 601 if (rpipe->pipe_busy == 0) { 602 cv_broadcast(&rpipe->pipe_draincv); 603 } 604 if (bp->cnt < MINPIPESIZE) { 605 cv_broadcast(&rpipe->pipe_wcv); 606 } 607 608 /* 609 * If anything was read off the buffer, signal to the writer it's 610 * possible to write more data. Also send signal if we are here for the 611 * first time after last write. 612 */ 613 if ((bp->size - bp->cnt) >= PIPE_BUF 614 && (ocnt != bp->cnt || (rpipe->pipe_state & PIPE_SIGNALR))) { 615 pipeselwakeup(rpipe, rpipe->pipe_peer, POLL_OUT); 616 rpipe->pipe_state &= ~PIPE_SIGNALR; 617 } 618 619 mutex_exit(lock); 620 return (error); 621 } 622 623 #ifndef PIPE_NODIRECT 624 /* 625 * Allocate structure for loan transfer. 626 */ 627 static int 628 pipe_loan_alloc(struct pipe *wpipe, int npages) 629 { 630 vsize_t len; 631 632 len = (vsize_t)npages << PAGE_SHIFT; 633 atomic_add_int(&amountpipekva, len); 634 wpipe->pipe_map.kva = uvm_km_alloc(kernel_map, len, 0, 635 UVM_KMF_VAONLY | UVM_KMF_WAITVA); 636 if (wpipe->pipe_map.kva == 0) { 637 atomic_add_int(&amountpipekva, -len); 638 return (ENOMEM); 639 } 640 641 wpipe->pipe_map.npages = npages; 642 wpipe->pipe_map.pgs = kmem_alloc(npages * sizeof(struct vm_page *), 643 KM_SLEEP); 644 return (0); 645 } 646 647 /* 648 * Free resources allocated for loan transfer. 649 */ 650 static void 651 pipe_loan_free(struct pipe *wpipe) 652 { 653 vsize_t len; 654 655 len = (vsize_t)wpipe->pipe_map.npages << PAGE_SHIFT; 656 uvm_emap_remove(wpipe->pipe_map.kva, len); /* XXX */ 657 uvm_km_free(kernel_map, wpipe->pipe_map.kva, len, UVM_KMF_VAONLY); 658 wpipe->pipe_map.kva = 0; 659 atomic_add_int(&amountpipekva, -len); 660 kmem_free(wpipe->pipe_map.pgs, 661 wpipe->pipe_map.npages * sizeof(struct vm_page *)); 662 wpipe->pipe_map.pgs = NULL; 663 } 664 665 /* 666 * NetBSD direct write, using uvm_loan() mechanism. 667 * This implements the pipe buffer write mechanism. Note that only 668 * a direct write OR a normal pipe write can be pending at any given time. 669 * If there are any characters in the pipe buffer, the direct write will 670 * be deferred until the receiving process grabs all of the bytes from 671 * the pipe buffer. Then the direct mapping write is set-up. 672 * 673 * Called with the long-term pipe lock held. 674 */ 675 static int 676 pipe_direct_write(file_t *fp, struct pipe *wpipe, struct uio *uio) 677 { 678 struct vm_page **pgs; 679 vaddr_t bbase, base, bend; 680 vsize_t blen, bcnt; 681 int error, npages; 682 voff_t bpos; 683 kmutex_t *lock = wpipe->pipe_lock; 684 685 KASSERT(mutex_owned(wpipe->pipe_lock)); 686 KASSERT(wpipe->pipe_map.cnt == 0); 687 688 mutex_exit(lock); 689 690 /* 691 * Handle first PIPE_CHUNK_SIZE bytes of buffer. Deal with buffers 692 * not aligned to PAGE_SIZE. 693 */ 694 bbase = (vaddr_t)uio->uio_iov->iov_base; 695 base = trunc_page(bbase); 696 bend = round_page(bbase + uio->uio_iov->iov_len); 697 blen = bend - base; 698 bpos = bbase - base; 699 700 if (blen > PIPE_DIRECT_CHUNK) { 701 blen = PIPE_DIRECT_CHUNK; 702 bend = base + blen; 703 bcnt = PIPE_DIRECT_CHUNK - bpos; 704 } else { 705 bcnt = uio->uio_iov->iov_len; 706 } 707 npages = blen >> PAGE_SHIFT; 708 709 /* 710 * Free the old kva if we need more pages than we have 711 * allocated. 712 */ 713 if (wpipe->pipe_map.kva != 0 && npages > wpipe->pipe_map.npages) 714 pipe_loan_free(wpipe); 715 716 /* Allocate new kva. */ 717 if (wpipe->pipe_map.kva == 0) { 718 error = pipe_loan_alloc(wpipe, npages); 719 if (error) { 720 mutex_enter(lock); 721 return (error); 722 } 723 } 724 725 /* Loan the write buffer memory from writer process */ 726 pgs = wpipe->pipe_map.pgs; 727 error = uvm_loan(&uio->uio_vmspace->vm_map, base, blen, 728 pgs, UVM_LOAN_TOPAGE); 729 if (error) { 730 pipe_loan_free(wpipe); 731 mutex_enter(lock); 732 return (ENOMEM); /* so that caller fallback to ordinary write */ 733 } 734 735 /* Enter the loaned pages to KVA, produce new emap generation number. */ 736 uvm_emap_enter(wpipe->pipe_map.kva, pgs, npages); 737 wpipe->pipe_map.egen = uvm_emap_produce(); 738 739 /* Now we can put the pipe in direct write mode */ 740 wpipe->pipe_map.pos = bpos; 741 wpipe->pipe_map.cnt = bcnt; 742 743 /* 744 * But before we can let someone do a direct read, we 745 * have to wait until the pipe is drained. Release the 746 * pipe lock while we wait. 747 */ 748 mutex_enter(lock); 749 wpipe->pipe_state |= PIPE_DIRECTW; 750 pipeunlock(wpipe); 751 752 while (error == 0 && wpipe->pipe_buffer.cnt > 0) { 753 cv_broadcast(&wpipe->pipe_rcv); 754 error = cv_wait_sig(&wpipe->pipe_wcv, lock); 755 if (error == 0 && wpipe->pipe_state & PIPE_EOF) 756 error = EPIPE; 757 } 758 759 /* Pipe is drained; next read will off the direct buffer */ 760 wpipe->pipe_state |= PIPE_DIRECTR; 761 762 /* Wait until the reader is done */ 763 while (error == 0 && (wpipe->pipe_state & PIPE_DIRECTR)) { 764 cv_broadcast(&wpipe->pipe_rcv); 765 pipeselwakeup(wpipe, wpipe, POLL_IN); 766 error = cv_wait_sig(&wpipe->pipe_wcv, lock); 767 if (error == 0 && wpipe->pipe_state & PIPE_EOF) 768 error = EPIPE; 769 } 770 771 /* Take pipe out of direct write mode */ 772 wpipe->pipe_state &= ~(PIPE_DIRECTW | PIPE_DIRECTR); 773 774 /* Acquire the pipe lock and cleanup */ 775 (void)pipelock(wpipe, 0); 776 mutex_exit(lock); 777 778 if (pgs != NULL) { 779 /* XXX: uvm_emap_remove */ 780 uvm_unloan(pgs, npages, UVM_LOAN_TOPAGE); 781 } 782 if (error || amountpipekva > maxpipekva) 783 pipe_loan_free(wpipe); 784 785 mutex_enter(lock); 786 if (error) { 787 pipeselwakeup(wpipe, wpipe, POLL_ERR); 788 789 /* 790 * If nothing was read from what we offered, return error 791 * straight on. Otherwise update uio resid first. Caller 792 * will deal with the error condition, returning short 793 * write, error, or restarting the write(2) as appropriate. 794 */ 795 if (wpipe->pipe_map.cnt == bcnt) { 796 wpipe->pipe_map.cnt = 0; 797 cv_broadcast(&wpipe->pipe_wcv); 798 return (error); 799 } 800 801 bcnt -= wpipe->pipe_map.cnt; 802 } 803 804 uio->uio_resid -= bcnt; 805 /* uio_offset not updated, not set/used for write(2) */ 806 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + bcnt; 807 uio->uio_iov->iov_len -= bcnt; 808 if (uio->uio_iov->iov_len == 0) { 809 uio->uio_iov++; 810 uio->uio_iovcnt--; 811 } 812 813 wpipe->pipe_map.cnt = 0; 814 return (error); 815 } 816 #endif /* !PIPE_NODIRECT */ 817 818 static int 819 pipe_write(file_t *fp, off_t *offset, struct uio *uio, kauth_cred_t cred, 820 int flags) 821 { 822 struct pipe *wpipe, *rpipe; 823 struct pipebuf *bp; 824 kmutex_t *lock; 825 int error; 826 int slept = 0; 827 828 /* We want to write to our peer */ 829 rpipe = (struct pipe *) fp->f_data; 830 lock = rpipe->pipe_lock; 831 error = 0; 832 833 mutex_enter(lock); 834 wpipe = rpipe->pipe_peer; 835 836 /* 837 * Detect loss of pipe read side, issue SIGPIPE if lost. 838 */ 839 if (wpipe == NULL || (wpipe->pipe_state & PIPE_EOF) != 0) { 840 mutex_exit(lock); 841 return EPIPE; 842 } 843 ++wpipe->pipe_busy; 844 845 /* Aquire the long-term pipe lock */ 846 if ((error = pipelock(wpipe, 1)) != 0) { 847 --wpipe->pipe_busy; 848 if (wpipe->pipe_busy == 0) { 849 cv_broadcast(&wpipe->pipe_draincv); 850 } 851 mutex_exit(lock); 852 return (error); 853 } 854 855 bp = &wpipe->pipe_buffer; 856 857 /* 858 * If it is advantageous to resize the pipe buffer, do so. 859 */ 860 if ((uio->uio_resid > PIPE_SIZE) && 861 (nbigpipe < maxbigpipes) && 862 #ifndef PIPE_NODIRECT 863 (wpipe->pipe_state & PIPE_DIRECTW) == 0 && 864 #endif 865 (bp->size <= PIPE_SIZE) && (bp->cnt == 0)) { 866 867 if (pipespace(wpipe, BIG_PIPE_SIZE) == 0) 868 atomic_inc_uint(&nbigpipe); 869 } 870 871 while (uio->uio_resid) { 872 size_t space; 873 874 #ifndef PIPE_NODIRECT 875 /* 876 * Pipe buffered writes cannot be coincidental with 877 * direct writes. Also, only one direct write can be 878 * in progress at any one time. We wait until the currently 879 * executing direct write is completed before continuing. 880 * 881 * We break out if a signal occurs or the reader goes away. 882 */ 883 while (error == 0 && wpipe->pipe_state & PIPE_DIRECTW) { 884 cv_broadcast(&wpipe->pipe_rcv); 885 pipeunlock(wpipe); 886 error = cv_wait_sig(&wpipe->pipe_wcv, lock); 887 (void)pipelock(wpipe, 0); 888 if (wpipe->pipe_state & PIPE_EOF) 889 error = EPIPE; 890 } 891 if (error) 892 break; 893 894 /* 895 * If the transfer is large, we can gain performance if 896 * we do process-to-process copies directly. 897 * If the write is non-blocking, we don't use the 898 * direct write mechanism. 899 * 900 * The direct write mechanism will detect the reader going 901 * away on us. 902 */ 903 if ((uio->uio_iov->iov_len >= PIPE_MINDIRECT) && 904 (fp->f_flag & FNONBLOCK) == 0 && 905 (wpipe->pipe_map.kva || (amountpipekva < limitpipekva))) { 906 error = pipe_direct_write(fp, wpipe, uio); 907 908 /* 909 * Break out if error occurred, unless it's ENOMEM. 910 * ENOMEM means we failed to allocate some resources 911 * for direct write, so we just fallback to ordinary 912 * write. If the direct write was successful, 913 * process rest of data via ordinary write. 914 */ 915 if (error == 0) 916 continue; 917 918 if (error != ENOMEM) 919 break; 920 } 921 #endif /* PIPE_NODIRECT */ 922 923 space = bp->size - bp->cnt; 924 925 /* Writes of size <= PIPE_BUF must be atomic. */ 926 if ((space < uio->uio_resid) && (uio->uio_resid <= PIPE_BUF)) 927 space = 0; 928 929 if (space > 0) { 930 int size; /* Transfer size */ 931 int segsize; /* first segment to transfer */ 932 933 /* 934 * Transfer size is minimum of uio transfer 935 * and free space in pipe buffer. 936 */ 937 if (space > uio->uio_resid) 938 size = uio->uio_resid; 939 else 940 size = space; 941 /* 942 * First segment to transfer is minimum of 943 * transfer size and contiguous space in 944 * pipe buffer. If first segment to transfer 945 * is less than the transfer size, we've got 946 * a wraparound in the buffer. 947 */ 948 segsize = bp->size - bp->in; 949 if (segsize > size) 950 segsize = size; 951 952 /* Transfer first segment */ 953 mutex_exit(lock); 954 error = uiomove((char *)bp->buffer + bp->in, segsize, 955 uio); 956 957 if (error == 0 && segsize < size) { 958 /* 959 * Transfer remaining part now, to 960 * support atomic writes. Wraparound 961 * happened. 962 */ 963 KASSERT(bp->in + segsize == bp->size); 964 error = uiomove(bp->buffer, 965 size - segsize, uio); 966 } 967 mutex_enter(lock); 968 if (error) 969 break; 970 971 bp->in += size; 972 if (bp->in >= bp->size) { 973 KASSERT(bp->in == size - segsize + bp->size); 974 bp->in = size - segsize; 975 } 976 977 bp->cnt += size; 978 KASSERT(bp->cnt <= bp->size); 979 } else { 980 /* 981 * If the "read-side" has been blocked, wake it up now. 982 */ 983 cv_broadcast(&wpipe->pipe_rcv); 984 985 /* 986 * Don't block on non-blocking I/O. 987 */ 988 if (fp->f_flag & FNONBLOCK) { 989 error = EAGAIN; 990 break; 991 } 992 993 /* 994 * We have no more space and have something to offer, 995 * wake up select/poll. 996 */ 997 if (bp->cnt) 998 pipeselwakeup(wpipe, wpipe, POLL_IN); 999 1000 if (slept) { 1001 error = ERESTART; 1002 break; 1003 } 1004 1005 pipeunlock(wpipe); 1006 error = cv_wait_sig(&wpipe->pipe_wcv, lock); 1007 (void)pipelock(wpipe, 0); 1008 if (error != 0) 1009 break; 1010 /* 1011 * If read side wants to go away, we just issue a signal 1012 * to ourselves. 1013 */ 1014 if (wpipe->pipe_state & PIPE_EOF) { 1015 error = EPIPE; 1016 break; 1017 } 1018 slept = 1; 1019 } 1020 } 1021 1022 --wpipe->pipe_busy; 1023 if (wpipe->pipe_busy == 0) { 1024 cv_broadcast(&wpipe->pipe_draincv); 1025 } 1026 if (bp->cnt > 0) { 1027 cv_broadcast(&wpipe->pipe_rcv); 1028 } 1029 1030 /* 1031 * Don't return EPIPE if I/O was successful 1032 */ 1033 if (error == EPIPE && bp->cnt == 0 && uio->uio_resid == 0) 1034 error = 0; 1035 1036 if (error == 0) 1037 getnanotime(&wpipe->pipe_mtime); 1038 1039 /* 1040 * We have something to offer, wake up select/poll. 1041 * wpipe->pipe_map.cnt is always 0 in this point (direct write 1042 * is only done synchronously), so check only wpipe->pipe_buffer.cnt 1043 */ 1044 if (bp->cnt) 1045 pipeselwakeup(wpipe, wpipe, POLL_IN); 1046 1047 /* 1048 * Arrange for next read(2) to do a signal. 1049 */ 1050 wpipe->pipe_state |= PIPE_SIGNALR; 1051 1052 pipeunlock(wpipe); 1053 mutex_exit(lock); 1054 return (error); 1055 } 1056 1057 /* 1058 * We implement a very minimal set of ioctls for compatibility with sockets. 1059 */ 1060 int 1061 pipe_ioctl(file_t *fp, u_long cmd, void *data) 1062 { 1063 struct pipe *pipe = fp->f_data; 1064 kmutex_t *lock = pipe->pipe_lock; 1065 1066 switch (cmd) { 1067 1068 case FIONBIO: 1069 return (0); 1070 1071 case FIOASYNC: 1072 mutex_enter(lock); 1073 if (*(int *)data) { 1074 pipe->pipe_state |= PIPE_ASYNC; 1075 } else { 1076 pipe->pipe_state &= ~PIPE_ASYNC; 1077 } 1078 mutex_exit(lock); 1079 return (0); 1080 1081 case FIONREAD: 1082 mutex_enter(lock); 1083 #ifndef PIPE_NODIRECT 1084 if (pipe->pipe_state & PIPE_DIRECTW) 1085 *(int *)data = pipe->pipe_map.cnt; 1086 else 1087 #endif 1088 *(int *)data = pipe->pipe_buffer.cnt; 1089 mutex_exit(lock); 1090 return (0); 1091 1092 case FIONWRITE: 1093 /* Look at other side */ 1094 pipe = pipe->pipe_peer; 1095 mutex_enter(lock); 1096 #ifndef PIPE_NODIRECT 1097 if (pipe->pipe_state & PIPE_DIRECTW) 1098 *(int *)data = pipe->pipe_map.cnt; 1099 else 1100 #endif 1101 *(int *)data = pipe->pipe_buffer.cnt; 1102 mutex_exit(lock); 1103 return (0); 1104 1105 case FIONSPACE: 1106 /* Look at other side */ 1107 pipe = pipe->pipe_peer; 1108 mutex_enter(lock); 1109 #ifndef PIPE_NODIRECT 1110 /* 1111 * If we're in direct-mode, we don't really have a 1112 * send queue, and any other write will block. Thus 1113 * zero seems like the best answer. 1114 */ 1115 if (pipe->pipe_state & PIPE_DIRECTW) 1116 *(int *)data = 0; 1117 else 1118 #endif 1119 *(int *)data = pipe->pipe_buffer.size - 1120 pipe->pipe_buffer.cnt; 1121 mutex_exit(lock); 1122 return (0); 1123 1124 case TIOCSPGRP: 1125 case FIOSETOWN: 1126 return fsetown(&pipe->pipe_pgid, cmd, data); 1127 1128 case TIOCGPGRP: 1129 case FIOGETOWN: 1130 return fgetown(pipe->pipe_pgid, cmd, data); 1131 1132 } 1133 return (EPASSTHROUGH); 1134 } 1135 1136 int 1137 pipe_poll(file_t *fp, int events) 1138 { 1139 struct pipe *rpipe = fp->f_data; 1140 struct pipe *wpipe; 1141 int eof = 0; 1142 int revents = 0; 1143 1144 mutex_enter(rpipe->pipe_lock); 1145 wpipe = rpipe->pipe_peer; 1146 1147 if (events & (POLLIN | POLLRDNORM)) 1148 if ((rpipe->pipe_buffer.cnt > 0) || 1149 #ifndef PIPE_NODIRECT 1150 (rpipe->pipe_state & PIPE_DIRECTR) || 1151 #endif 1152 (rpipe->pipe_state & PIPE_EOF)) 1153 revents |= events & (POLLIN | POLLRDNORM); 1154 1155 eof |= (rpipe->pipe_state & PIPE_EOF); 1156 1157 if (wpipe == NULL) 1158 revents |= events & (POLLOUT | POLLWRNORM); 1159 else { 1160 if (events & (POLLOUT | POLLWRNORM)) 1161 if ((wpipe->pipe_state & PIPE_EOF) || ( 1162 #ifndef PIPE_NODIRECT 1163 (wpipe->pipe_state & PIPE_DIRECTW) == 0 && 1164 #endif 1165 (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF)) 1166 revents |= events & (POLLOUT | POLLWRNORM); 1167 1168 eof |= (wpipe->pipe_state & PIPE_EOF); 1169 } 1170 1171 if (wpipe == NULL || eof) 1172 revents |= POLLHUP; 1173 1174 if (revents == 0) { 1175 if (events & (POLLIN | POLLRDNORM)) 1176 selrecord(curlwp, &rpipe->pipe_sel); 1177 1178 if (events & (POLLOUT | POLLWRNORM)) 1179 selrecord(curlwp, &wpipe->pipe_sel); 1180 } 1181 mutex_exit(rpipe->pipe_lock); 1182 1183 return (revents); 1184 } 1185 1186 static int 1187 pipe_stat(file_t *fp, struct stat *ub) 1188 { 1189 struct pipe *pipe = fp->f_data; 1190 1191 mutex_enter(pipe->pipe_lock); 1192 memset(ub, 0, sizeof(*ub)); 1193 ub->st_mode = S_IFIFO | S_IRUSR | S_IWUSR; 1194 ub->st_blksize = pipe->pipe_buffer.size; 1195 if (ub->st_blksize == 0 && pipe->pipe_peer) 1196 ub->st_blksize = pipe->pipe_peer->pipe_buffer.size; 1197 ub->st_size = pipe->pipe_buffer.cnt; 1198 ub->st_blocks = (ub->st_size) ? 1 : 0; 1199 ub->st_atimespec = pipe->pipe_atime; 1200 ub->st_mtimespec = pipe->pipe_mtime; 1201 ub->st_ctimespec = ub->st_birthtimespec = pipe->pipe_btime; 1202 ub->st_uid = kauth_cred_geteuid(fp->f_cred); 1203 ub->st_gid = kauth_cred_getegid(fp->f_cred); 1204 1205 /* 1206 * Left as 0: st_dev, st_ino, st_nlink, st_rdev, st_flags, st_gen. 1207 * XXX (st_dev, st_ino) should be unique. 1208 */ 1209 mutex_exit(pipe->pipe_lock); 1210 return 0; 1211 } 1212 1213 static int 1214 pipe_close(file_t *fp) 1215 { 1216 struct pipe *pipe = fp->f_data; 1217 1218 fp->f_data = NULL; 1219 pipeclose(pipe); 1220 return (0); 1221 } 1222 1223 static void 1224 pipe_abort(file_t *fp) 1225 { 1226 struct pipe *pipe = fp->f_data; 1227 1228 /* 1229 * Unblock blocked reads/writes in order to allow close() to complete. 1230 * This isn't going to work yet! 1231 * The underlying problem is that only the 'fd' in question needs 1232 * its operations terminating, the pipe itself my be open via 1233 * other fd. 1234 */ 1235 mutex_enter(pipe->pipe_lock); 1236 cv_broadcast(&pipe->pipe_rcv); 1237 cv_broadcast(&pipe->pipe_wcv); 1238 mutex_exit(pipe->pipe_lock); 1239 } 1240 1241 static void 1242 pipe_free_kmem(struct pipe *pipe) 1243 { 1244 1245 if (pipe->pipe_buffer.buffer != NULL) { 1246 if (pipe->pipe_buffer.size > PIPE_SIZE) { 1247 atomic_dec_uint(&nbigpipe); 1248 } 1249 if (pipe->pipe_buffer.buffer != (void *)pipe->pipe_kmem) { 1250 uvm_km_free(kernel_map, 1251 (vaddr_t)pipe->pipe_buffer.buffer, 1252 pipe->pipe_buffer.size, UVM_KMF_PAGEABLE); 1253 atomic_add_int(&amountpipekva, 1254 -pipe->pipe_buffer.size); 1255 } 1256 pipe->pipe_buffer.buffer = NULL; 1257 } 1258 #ifndef PIPE_NODIRECT 1259 if (pipe->pipe_map.kva != 0) { 1260 pipe_loan_free(pipe); 1261 pipe->pipe_map.cnt = 0; 1262 pipe->pipe_map.kva = 0; 1263 pipe->pipe_map.pos = 0; 1264 pipe->pipe_map.npages = 0; 1265 } 1266 #endif /* !PIPE_NODIRECT */ 1267 } 1268 1269 /* 1270 * Shutdown the pipe. 1271 */ 1272 static void 1273 pipeclose(struct pipe *pipe) 1274 { 1275 kmutex_t *lock; 1276 struct pipe *ppipe; 1277 1278 if (pipe == NULL) 1279 return; 1280 1281 KASSERT(cv_is_valid(&pipe->pipe_rcv)); 1282 KASSERT(cv_is_valid(&pipe->pipe_wcv)); 1283 KASSERT(cv_is_valid(&pipe->pipe_draincv)); 1284 KASSERT(cv_is_valid(&pipe->pipe_lkcv)); 1285 1286 lock = pipe->pipe_lock; 1287 if (lock == NULL) 1288 /* Must have failed during create */ 1289 goto free_resources; 1290 1291 mutex_enter(lock); 1292 pipeselwakeup(pipe, pipe, POLL_HUP); 1293 1294 /* 1295 * If the other side is blocked, wake it up saying that 1296 * we want to close it down. 1297 */ 1298 pipe->pipe_state |= PIPE_EOF; 1299 if (pipe->pipe_busy) { 1300 while (pipe->pipe_busy) { 1301 cv_broadcast(&pipe->pipe_wcv); 1302 cv_wait_sig(&pipe->pipe_draincv, lock); 1303 } 1304 } 1305 1306 /* 1307 * Disconnect from peer. 1308 */ 1309 if ((ppipe = pipe->pipe_peer) != NULL) { 1310 pipeselwakeup(ppipe, ppipe, POLL_HUP); 1311 ppipe->pipe_state |= PIPE_EOF; 1312 cv_broadcast(&ppipe->pipe_rcv); 1313 ppipe->pipe_peer = NULL; 1314 } 1315 1316 /* 1317 * Any knote objects still left in the list are 1318 * the one attached by peer. Since no one will 1319 * traverse this list, we just clear it. 1320 */ 1321 SLIST_INIT(&pipe->pipe_sel.sel_klist); 1322 1323 KASSERT((pipe->pipe_state & PIPE_LOCKFL) == 0); 1324 mutex_exit(lock); 1325 mutex_obj_free(lock); 1326 1327 /* 1328 * Free resources. 1329 */ 1330 free_resources: 1331 pipe->pipe_pgid = 0; 1332 pipe->pipe_state = PIPE_SIGNALR; 1333 pipe_free_kmem(pipe); 1334 if (pipe->pipe_kmem != 0) { 1335 pool_cache_put(pipe_rd_cache, pipe); 1336 } else { 1337 pool_cache_put(pipe_wr_cache, pipe); 1338 } 1339 } 1340 1341 static void 1342 filt_pipedetach(struct knote *kn) 1343 { 1344 struct pipe *pipe; 1345 kmutex_t *lock; 1346 1347 pipe = ((file_t *)kn->kn_obj)->f_data; 1348 lock = pipe->pipe_lock; 1349 1350 mutex_enter(lock); 1351 1352 switch(kn->kn_filter) { 1353 case EVFILT_WRITE: 1354 /* Need the peer structure, not our own. */ 1355 pipe = pipe->pipe_peer; 1356 1357 /* If reader end already closed, just return. */ 1358 if (pipe == NULL) { 1359 mutex_exit(lock); 1360 return; 1361 } 1362 1363 break; 1364 default: 1365 /* Nothing to do. */ 1366 break; 1367 } 1368 1369 KASSERT(kn->kn_hook == pipe); 1370 SLIST_REMOVE(&pipe->pipe_sel.sel_klist, kn, knote, kn_selnext); 1371 mutex_exit(lock); 1372 } 1373 1374 static int 1375 filt_piperead(struct knote *kn, long hint) 1376 { 1377 struct pipe *rpipe = ((file_t *)kn->kn_obj)->f_data; 1378 struct pipe *wpipe; 1379 1380 if ((hint & NOTE_SUBMIT) == 0) { 1381 mutex_enter(rpipe->pipe_lock); 1382 } 1383 wpipe = rpipe->pipe_peer; 1384 kn->kn_data = rpipe->pipe_buffer.cnt; 1385 1386 if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW)) 1387 kn->kn_data = rpipe->pipe_map.cnt; 1388 1389 if ((rpipe->pipe_state & PIPE_EOF) || 1390 (wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) { 1391 kn->kn_flags |= EV_EOF; 1392 if ((hint & NOTE_SUBMIT) == 0) { 1393 mutex_exit(rpipe->pipe_lock); 1394 } 1395 return (1); 1396 } 1397 1398 if ((hint & NOTE_SUBMIT) == 0) { 1399 mutex_exit(rpipe->pipe_lock); 1400 } 1401 return (kn->kn_data > 0); 1402 } 1403 1404 static int 1405 filt_pipewrite(struct knote *kn, long hint) 1406 { 1407 struct pipe *rpipe = ((file_t *)kn->kn_obj)->f_data; 1408 struct pipe *wpipe; 1409 1410 if ((hint & NOTE_SUBMIT) == 0) { 1411 mutex_enter(rpipe->pipe_lock); 1412 } 1413 wpipe = rpipe->pipe_peer; 1414 1415 if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) { 1416 kn->kn_data = 0; 1417 kn->kn_flags |= EV_EOF; 1418 if ((hint & NOTE_SUBMIT) == 0) { 1419 mutex_exit(rpipe->pipe_lock); 1420 } 1421 return (1); 1422 } 1423 kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt; 1424 if (wpipe->pipe_state & PIPE_DIRECTW) 1425 kn->kn_data = 0; 1426 1427 if ((hint & NOTE_SUBMIT) == 0) { 1428 mutex_exit(rpipe->pipe_lock); 1429 } 1430 return (kn->kn_data >= PIPE_BUF); 1431 } 1432 1433 static const struct filterops pipe_rfiltops = 1434 { 1, NULL, filt_pipedetach, filt_piperead }; 1435 static const struct filterops pipe_wfiltops = 1436 { 1, NULL, filt_pipedetach, filt_pipewrite }; 1437 1438 static int 1439 pipe_kqfilter(file_t *fp, struct knote *kn) 1440 { 1441 struct pipe *pipe; 1442 kmutex_t *lock; 1443 1444 pipe = ((file_t *)kn->kn_obj)->f_data; 1445 lock = pipe->pipe_lock; 1446 1447 mutex_enter(lock); 1448 1449 switch (kn->kn_filter) { 1450 case EVFILT_READ: 1451 kn->kn_fop = &pipe_rfiltops; 1452 break; 1453 case EVFILT_WRITE: 1454 kn->kn_fop = &pipe_wfiltops; 1455 pipe = pipe->pipe_peer; 1456 if (pipe == NULL) { 1457 /* Other end of pipe has been closed. */ 1458 mutex_exit(lock); 1459 return (EBADF); 1460 } 1461 break; 1462 default: 1463 mutex_exit(lock); 1464 return (EINVAL); 1465 } 1466 1467 kn->kn_hook = pipe; 1468 SLIST_INSERT_HEAD(&pipe->pipe_sel.sel_klist, kn, kn_selnext); 1469 mutex_exit(lock); 1470 1471 return (0); 1472 } 1473 1474 /* 1475 * Handle pipe sysctls. 1476 */ 1477 SYSCTL_SETUP(sysctl_kern_pipe_setup, "sysctl kern.pipe subtree setup") 1478 { 1479 1480 sysctl_createv(clog, 0, NULL, NULL, 1481 CTLFLAG_PERMANENT, 1482 CTLTYPE_NODE, "kern", NULL, 1483 NULL, 0, NULL, 0, 1484 CTL_KERN, CTL_EOL); 1485 sysctl_createv(clog, 0, NULL, NULL, 1486 CTLFLAG_PERMANENT, 1487 CTLTYPE_NODE, "pipe", 1488 SYSCTL_DESCR("Pipe settings"), 1489 NULL, 0, NULL, 0, 1490 CTL_KERN, KERN_PIPE, CTL_EOL); 1491 1492 sysctl_createv(clog, 0, NULL, NULL, 1493 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 1494 CTLTYPE_INT, "maxkvasz", 1495 SYSCTL_DESCR("Maximum amount of kernel memory to be " 1496 "used for pipes"), 1497 NULL, 0, &maxpipekva, 0, 1498 CTL_KERN, KERN_PIPE, KERN_PIPE_MAXKVASZ, CTL_EOL); 1499 sysctl_createv(clog, 0, NULL, NULL, 1500 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 1501 CTLTYPE_INT, "maxloankvasz", 1502 SYSCTL_DESCR("Limit for direct transfers via page loan"), 1503 NULL, 0, &limitpipekva, 0, 1504 CTL_KERN, KERN_PIPE, KERN_PIPE_LIMITKVA, CTL_EOL); 1505 sysctl_createv(clog, 0, NULL, NULL, 1506 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 1507 CTLTYPE_INT, "maxbigpipes", 1508 SYSCTL_DESCR("Maximum number of \"big\" pipes"), 1509 NULL, 0, &maxbigpipes, 0, 1510 CTL_KERN, KERN_PIPE, KERN_PIPE_MAXBIGPIPES, CTL_EOL); 1511 sysctl_createv(clog, 0, NULL, NULL, 1512 CTLFLAG_PERMANENT, 1513 CTLTYPE_INT, "nbigpipes", 1514 SYSCTL_DESCR("Number of \"big\" pipes"), 1515 NULL, 0, &nbigpipe, 0, 1516 CTL_KERN, KERN_PIPE, KERN_PIPE_NBIGPIPES, CTL_EOL); 1517 sysctl_createv(clog, 0, NULL, NULL, 1518 CTLFLAG_PERMANENT, 1519 CTLTYPE_INT, "kvasize", 1520 SYSCTL_DESCR("Amount of kernel memory consumed by pipe " 1521 "buffers"), 1522 NULL, 0, &amountpipekva, 0, 1523 CTL_KERN, KERN_PIPE, KERN_PIPE_KVASIZE, CTL_EOL); 1524 } 1525