1 /* 2 * Copyright (c) 1996 John S. Dyson 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice immediately at the beginning of the file, without modification, 10 * this list of conditions, and the following disclaimer. 11 * 2. Redistributions in binary form must reproduce the above copyright 12 * notice, this list of conditions and the following disclaimer in the 13 * documentation and/or other materials provided with the distribution. 14 * 3. Absolutely no warranty of function or purpose is made by the author 15 * John S. Dyson. 16 * 4. Modifications may be freely made to this file if the above conditions 17 * are met. 18 * 19 * $FreeBSD: src/sys/kern/sys_pipe.c,v 1.60.2.13 2002/08/05 15:05:15 des Exp $ 20 * $DragonFly: src/sys/kern/sys_pipe.c,v 1.50 2008/09/09 04:06:13 dillon Exp $ 21 */ 22 23 /* 24 * This file contains a high-performance replacement for the socket-based 25 * pipes scheme originally used in FreeBSD/4.4Lite. It does not support 26 * all features of sockets, but does do everything that pipes normally 27 * do. 28 */ 29 #include <sys/param.h> 30 #include <sys/systm.h> 31 #include <sys/kernel.h> 32 #include <sys/proc.h> 33 #include <sys/fcntl.h> 34 #include <sys/file.h> 35 #include <sys/filedesc.h> 36 #include <sys/filio.h> 37 #include <sys/ttycom.h> 38 #include <sys/stat.h> 39 #include <sys/poll.h> 40 #include <sys/select.h> 41 #include <sys/signalvar.h> 42 #include <sys/sysproto.h> 43 #include <sys/pipe.h> 44 #include <sys/vnode.h> 45 #include <sys/uio.h> 46 #include <sys/event.h> 47 #include <sys/globaldata.h> 48 #include <sys/module.h> 49 #include <sys/malloc.h> 50 #include <sys/sysctl.h> 51 #include <sys/socket.h> 52 53 #include <vm/vm.h> 54 #include <vm/vm_param.h> 55 #include <sys/lock.h> 56 #include <vm/vm_object.h> 57 #include <vm/vm_kern.h> 58 #include <vm/vm_extern.h> 59 #include <vm/pmap.h> 60 #include <vm/vm_map.h> 61 #include <vm/vm_page.h> 62 #include <vm/vm_zone.h> 63 64 #include <sys/file2.h> 65 66 #include <machine/cpufunc.h> 67 68 /* 69 * interfaces to the outside world 70 */ 71 static int pipe_read (struct file *fp, struct uio *uio, 72 struct ucred *cred, int flags); 73 static int pipe_write (struct file *fp, struct uio *uio, 74 struct ucred *cred, int flags); 75 static int pipe_close (struct file *fp); 76 static int pipe_shutdown (struct file *fp, int how); 77 static int pipe_poll (struct file *fp, int events, struct ucred *cred); 78 static int pipe_kqfilter (struct file *fp, struct knote *kn); 79 static int pipe_stat (struct file *fp, struct stat *sb, struct ucred *cred); 80 static int pipe_ioctl (struct file *fp, u_long cmd, caddr_t data, struct ucred *cred); 81 82 static struct fileops pipeops = { 83 .fo_read = pipe_read, 84 .fo_write = pipe_write, 85 .fo_ioctl = pipe_ioctl, 86 .fo_poll = pipe_poll, 87 .fo_kqfilter = pipe_kqfilter, 88 .fo_stat = pipe_stat, 89 .fo_close = pipe_close, 90 .fo_shutdown = pipe_shutdown 91 }; 92 93 static void filt_pipedetach(struct knote *kn); 94 static int filt_piperead(struct knote *kn, long hint); 95 static int filt_pipewrite(struct knote *kn, long hint); 96 97 static struct filterops pipe_rfiltops = 98 { 1, NULL, filt_pipedetach, filt_piperead }; 99 static struct filterops pipe_wfiltops = 100 { 1, NULL, filt_pipedetach, filt_pipewrite }; 101 102 MALLOC_DEFINE(M_PIPE, "pipe", "pipe structures"); 103 104 /* 105 * Default pipe buffer size(s), this can be kind-of large now because pipe 106 * space is pageable. The pipe code will try to maintain locality of 107 * reference for performance reasons, so small amounts of outstanding I/O 108 * will not wipe the cache. 109 */ 110 #define MINPIPESIZE (PIPE_SIZE/3) 111 #define MAXPIPESIZE (2*PIPE_SIZE/3) 112 113 /* 114 * Limit the number of "big" pipes 115 */ 116 #define LIMITBIGPIPES 64 117 #define PIPEQ_MAX_CACHE 16 /* per-cpu pipe structure cache */ 118 119 static int pipe_maxbig = LIMITBIGPIPES; 120 static int pipe_maxcache = PIPEQ_MAX_CACHE; 121 static int pipe_bigcount; 122 static int pipe_nbig; 123 static int pipe_bcache_alloc; 124 static int pipe_bkmem_alloc; 125 static int pipe_rblocked_count; 126 static int pipe_wblocked_count; 127 128 SYSCTL_NODE(_kern, OID_AUTO, pipe, CTLFLAG_RW, 0, "Pipe operation"); 129 SYSCTL_INT(_kern_pipe, OID_AUTO, nbig, 130 CTLFLAG_RD, &pipe_nbig, 0, "numer of big pipes allocated"); 131 SYSCTL_INT(_kern_pipe, OID_AUTO, bigcount, 132 CTLFLAG_RW, &pipe_bigcount, 0, "number of times pipe expanded"); 133 SYSCTL_INT(_kern_pipe, OID_AUTO, rblocked, 134 CTLFLAG_RW, &pipe_rblocked_count, 0, "number of times pipe expanded"); 135 SYSCTL_INT(_kern_pipe, OID_AUTO, wblocked, 136 CTLFLAG_RW, &pipe_wblocked_count, 0, "number of times pipe expanded"); 137 SYSCTL_INT(_kern_pipe, OID_AUTO, maxcache, 138 CTLFLAG_RW, &pipe_maxcache, 0, "max pipes cached per-cpu"); 139 SYSCTL_INT(_kern_pipe, OID_AUTO, maxbig, 140 CTLFLAG_RW, &pipe_maxbig, 0, "max number of big pipes"); 141 #ifdef SMP 142 static int pipe_delay = 5000; /* 5uS default */ 143 SYSCTL_INT(_kern_pipe, OID_AUTO, delay, 144 CTLFLAG_RW, &pipe_delay, 0, "SMP delay optimization in ns"); 145 static int pipe_mpsafe = 0; 146 SYSCTL_INT(_kern_pipe, OID_AUTO, mpsafe, 147 CTLFLAG_RW, &pipe_mpsafe, 0, ""); 148 #endif 149 #if !defined(NO_PIPE_SYSCTL_STATS) 150 SYSCTL_INT(_kern_pipe, OID_AUTO, bcache_alloc, 151 CTLFLAG_RW, &pipe_bcache_alloc, 0, "pipe buffer from pcpu cache"); 152 SYSCTL_INT(_kern_pipe, OID_AUTO, bkmem_alloc, 153 CTLFLAG_RW, &pipe_bkmem_alloc, 0, "pipe buffer from kmem"); 154 #endif 155 156 static void pipeclose (struct pipe *cpipe); 157 static void pipe_free_kmem (struct pipe *cpipe); 158 static int pipe_create (struct pipe **cpipep); 159 static __inline void pipeselwakeup (struct pipe *cpipe); 160 static int pipespace (struct pipe *cpipe, int size); 161 162 static __inline void 163 pipeselwakeup(struct pipe *cpipe) 164 { 165 if (cpipe->pipe_state & PIPE_SEL) { 166 get_mplock(); 167 cpipe->pipe_state &= ~PIPE_SEL; 168 selwakeup(&cpipe->pipe_sel); 169 rel_mplock(); 170 } 171 if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio) { 172 get_mplock(); 173 pgsigio(cpipe->pipe_sigio, SIGIO, 0); 174 rel_mplock(); 175 } 176 if (SLIST_FIRST(&cpipe->pipe_sel.si_note)) { 177 get_mplock(); 178 KNOTE(&cpipe->pipe_sel.si_note, 0); 179 rel_mplock(); 180 } 181 } 182 183 /* 184 * These routines are called before and after a UIO. The UIO 185 * may block, causing our held tokens to be lost temporarily. 186 * 187 * We use these routines to serialize reads against other reads 188 * and writes against other writes. 189 * 190 * The read token is held on entry so *ipp does not race. 191 */ 192 static __inline int 193 pipe_start_uio(struct pipe *cpipe, u_int *ipp) 194 { 195 int error; 196 197 while (*ipp) { 198 *ipp = -1; 199 error = tsleep(ipp, PCATCH, "pipexx", 0); 200 if (error) 201 return (error); 202 } 203 *ipp = 1; 204 return (0); 205 } 206 207 static __inline void 208 pipe_end_uio(struct pipe *cpipe, u_int *ipp) 209 { 210 if (*ipp < 0) { 211 *ipp = 0; 212 wakeup(ipp); 213 } else { 214 *ipp = 0; 215 } 216 } 217 218 static __inline void 219 pipe_get_mplock(int *save) 220 { 221 #ifdef SMP 222 if (pipe_mpsafe == 0) { 223 get_mplock(); 224 *save = 1; 225 } else 226 #endif 227 { 228 *save = 0; 229 } 230 } 231 232 static __inline void 233 pipe_rel_mplock(int *save) 234 { 235 #ifdef SMP 236 if (*save) 237 rel_mplock(); 238 #endif 239 } 240 241 242 /* 243 * The pipe system call for the DTYPE_PIPE type of pipes 244 * 245 * pipe_ARgs(int dummy) 246 */ 247 248 /* ARGSUSED */ 249 int 250 sys_pipe(struct pipe_args *uap) 251 { 252 struct thread *td = curthread; 253 struct proc *p = td->td_proc; 254 struct file *rf, *wf; 255 struct pipe *rpipe, *wpipe; 256 int fd1, fd2, error; 257 258 KKASSERT(p); 259 260 rpipe = wpipe = NULL; 261 if (pipe_create(&rpipe) || pipe_create(&wpipe)) { 262 pipeclose(rpipe); 263 pipeclose(wpipe); 264 return (ENFILE); 265 } 266 267 error = falloc(p, &rf, &fd1); 268 if (error) { 269 pipeclose(rpipe); 270 pipeclose(wpipe); 271 return (error); 272 } 273 uap->sysmsg_fds[0] = fd1; 274 275 /* 276 * Warning: once we've gotten past allocation of the fd for the 277 * read-side, we can only drop the read side via fdrop() in order 278 * to avoid races against processes which manage to dup() the read 279 * side while we are blocked trying to allocate the write side. 280 */ 281 rf->f_type = DTYPE_PIPE; 282 rf->f_flag = FREAD | FWRITE; 283 rf->f_ops = &pipeops; 284 rf->f_data = rpipe; 285 error = falloc(p, &wf, &fd2); 286 if (error) { 287 fsetfd(p, NULL, fd1); 288 fdrop(rf); 289 /* rpipe has been closed by fdrop(). */ 290 pipeclose(wpipe); 291 return (error); 292 } 293 wf->f_type = DTYPE_PIPE; 294 wf->f_flag = FREAD | FWRITE; 295 wf->f_ops = &pipeops; 296 wf->f_data = wpipe; 297 uap->sysmsg_fds[1] = fd2; 298 299 rpipe->pipe_slock = kmalloc(sizeof(struct lock), 300 M_PIPE, M_WAITOK|M_ZERO); 301 wpipe->pipe_slock = rpipe->pipe_slock; 302 rpipe->pipe_peer = wpipe; 303 wpipe->pipe_peer = rpipe; 304 lockinit(rpipe->pipe_slock, "pipecl", 0, 0); 305 306 /* 307 * Once activated the peer relationship remains valid until 308 * both sides are closed. 309 */ 310 fsetfd(p, rf, fd1); 311 fsetfd(p, wf, fd2); 312 fdrop(rf); 313 fdrop(wf); 314 315 return (0); 316 } 317 318 /* 319 * Allocate kva for pipe circular buffer, the space is pageable 320 * This routine will 'realloc' the size of a pipe safely, if it fails 321 * it will retain the old buffer. 322 * If it fails it will return ENOMEM. 323 */ 324 static int 325 pipespace(struct pipe *cpipe, int size) 326 { 327 struct vm_object *object; 328 caddr_t buffer; 329 int npages, error; 330 331 npages = round_page(size) / PAGE_SIZE; 332 object = cpipe->pipe_buffer.object; 333 334 /* 335 * [re]create the object if necessary and reserve space for it 336 * in the kernel_map. The object and memory are pageable. On 337 * success, free the old resources before assigning the new 338 * ones. 339 */ 340 if (object == NULL || object->size != npages) { 341 get_mplock(); 342 object = vm_object_allocate(OBJT_DEFAULT, npages); 343 buffer = (caddr_t)vm_map_min(&kernel_map); 344 345 error = vm_map_find(&kernel_map, object, 0, 346 (vm_offset_t *)&buffer, size, 347 1, 348 VM_MAPTYPE_NORMAL, 349 VM_PROT_ALL, VM_PROT_ALL, 350 0); 351 352 if (error != KERN_SUCCESS) { 353 vm_object_deallocate(object); 354 rel_mplock(); 355 return (ENOMEM); 356 } 357 pipe_free_kmem(cpipe); 358 rel_mplock(); 359 cpipe->pipe_buffer.object = object; 360 cpipe->pipe_buffer.buffer = buffer; 361 cpipe->pipe_buffer.size = size; 362 ++pipe_bkmem_alloc; 363 } else { 364 ++pipe_bcache_alloc; 365 } 366 cpipe->pipe_buffer.rindex = 0; 367 cpipe->pipe_buffer.windex = 0; 368 return (0); 369 } 370 371 /* 372 * Initialize and allocate VM and memory for pipe, pulling the pipe from 373 * our per-cpu cache if possible. For now make sure it is sized for the 374 * smaller PIPE_SIZE default. 375 */ 376 static int 377 pipe_create(struct pipe **cpipep) 378 { 379 globaldata_t gd = mycpu; 380 struct pipe *cpipe; 381 int error; 382 383 if ((cpipe = gd->gd_pipeq) != NULL) { 384 gd->gd_pipeq = cpipe->pipe_peer; 385 --gd->gd_pipeqcount; 386 cpipe->pipe_peer = NULL; 387 cpipe->pipe_wantwcnt = 0; 388 } else { 389 cpipe = kmalloc(sizeof(struct pipe), M_PIPE, M_WAITOK|M_ZERO); 390 } 391 *cpipep = cpipe; 392 if ((error = pipespace(cpipe, PIPE_SIZE)) != 0) 393 return (error); 394 vfs_timestamp(&cpipe->pipe_ctime); 395 cpipe->pipe_atime = cpipe->pipe_ctime; 396 cpipe->pipe_mtime = cpipe->pipe_ctime; 397 lwkt_token_init(&cpipe->pipe_rlock); 398 lwkt_token_init(&cpipe->pipe_wlock); 399 return (0); 400 } 401 402 /* 403 * MPALMOSTSAFE (acquires mplock) 404 */ 405 static int 406 pipe_read(struct file *fp, struct uio *uio, struct ucred *cred, int fflags) 407 { 408 struct pipe *rpipe; 409 int error; 410 int orig_resid; 411 int nread = 0; 412 int nbio; 413 u_int size; /* total bytes available */ 414 u_int nsize; /* total bytes to read */ 415 u_int rindex; /* contiguous bytes available */ 416 int notify_writer; 417 lwkt_tokref rlock; 418 lwkt_tokref wlock; 419 int mpsave; 420 421 /* 422 * Degenerate case 423 */ 424 orig_resid = uio->uio_resid; 425 if (orig_resid == 0) 426 return(0); 427 428 /* 429 * Setup locks, calculate nbio 430 */ 431 pipe_get_mplock(&mpsave); 432 rpipe = (struct pipe *)fp->f_data; 433 lwkt_gettoken(&rlock, &rpipe->pipe_rlock); 434 435 if (fflags & O_FBLOCKING) 436 nbio = 0; 437 else if (fflags & O_FNONBLOCKING) 438 nbio = 1; 439 else if (fp->f_flag & O_NONBLOCK) 440 nbio = 1; 441 else 442 nbio = 0; 443 444 /* 445 * Reads are serialized. Note howeverthat pipe_buffer.buffer and 446 * pipe_buffer.size can change out from under us when the number 447 * of bytes in the buffer are zero due to the write-side doing a 448 * pipespace(). 449 */ 450 error = pipe_start_uio(rpipe, &rpipe->pipe_rip); 451 if (error) { 452 pipe_rel_mplock(&mpsave); 453 lwkt_reltoken(&rlock); 454 return (error); 455 } 456 notify_writer = 0; 457 while (uio->uio_resid) { 458 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex; 459 cpu_lfence(); 460 if (size) { 461 rindex = rpipe->pipe_buffer.rindex & 462 (rpipe->pipe_buffer.size - 1); 463 nsize = size; 464 if (nsize > rpipe->pipe_buffer.size - rindex) 465 nsize = rpipe->pipe_buffer.size - rindex; 466 if (nsize > (u_int)uio->uio_resid) 467 nsize = (u_int)uio->uio_resid; 468 469 error = uiomove(&rpipe->pipe_buffer.buffer[rindex], 470 nsize, uio); 471 if (error) 472 break; 473 cpu_mfence(); 474 rpipe->pipe_buffer.rindex += nsize; 475 nread += nsize; 476 477 /* 478 * If the FIFO is still over half full just continue 479 * and do not try to notify the writer yet. 480 */ 481 if (size - nsize >= (rpipe->pipe_buffer.size >> 1)) { 482 notify_writer = 0; 483 continue; 484 } 485 486 /* 487 * When the FIFO is less then half full notify any 488 * waiting writer. WANTW can be checked while 489 * holding just the rlock. 490 */ 491 notify_writer = 1; 492 if ((rpipe->pipe_state & PIPE_WANTW) == 0) 493 continue; 494 } 495 496 /* 497 * If the "write-side" was blocked we wake it up. This code 498 * is reached either when the buffer is completely emptied 499 * or if it becomes more then half-empty. 500 * 501 * Pipe_state can only be modified if both the rlock and 502 * wlock are held. 503 */ 504 if (rpipe->pipe_state & PIPE_WANTW) { 505 lwkt_gettoken(&wlock, &rpipe->pipe_wlock); 506 if (rpipe->pipe_state & PIPE_WANTW) { 507 notify_writer = 0; 508 rpipe->pipe_state &= ~PIPE_WANTW; 509 lwkt_reltoken(&wlock); 510 wakeup(rpipe); 511 } else { 512 lwkt_reltoken(&wlock); 513 } 514 } 515 516 /* 517 * Pick up our copy loop again if the writer sent data to 518 * us while we were messing around. 519 * 520 * On a SMP box poll up to pipe_delay nanoseconds for new 521 * data. Typically a value of 2000 to 4000 is sufficient 522 * to eradicate most IPIs/tsleeps/wakeups when a pipe 523 * is used for synchronous communications with small packets, 524 * and 8000 or so (8uS) will pipeline large buffer xfers 525 * between cpus over a pipe. 526 * 527 * For synchronous communications a hit means doing a 528 * full Awrite-Bread-Bwrite-Aread cycle in less then 2uS, 529 * where as miss requiring a tsleep/wakeup sequence 530 * will take 7uS or more. 531 */ 532 if (rpipe->pipe_buffer.windex != rpipe->pipe_buffer.rindex) 533 continue; 534 535 #if defined(SMP) && defined(_RDTSC_SUPPORTED_) 536 if (pipe_delay) { 537 int64_t tsc_target; 538 int good = 0; 539 540 tsc_target = tsc_get_target(pipe_delay); 541 while (tsc_test_target(tsc_target) == 0) { 542 if (rpipe->pipe_buffer.windex != 543 rpipe->pipe_buffer.rindex) { 544 good = 1; 545 break; 546 } 547 } 548 if (good) 549 continue; 550 } 551 #endif 552 553 /* 554 * Detect EOF condition, do not set error. 555 */ 556 if (rpipe->pipe_state & PIPE_REOF) 557 break; 558 559 /* 560 * Break if some data was read, or if this was a non-blocking 561 * read. 562 */ 563 if (nread > 0) 564 break; 565 566 if (nbio) { 567 error = EAGAIN; 568 break; 569 } 570 571 /* 572 * Last chance, interlock with WANTR. 573 */ 574 lwkt_gettoken(&wlock, &rpipe->pipe_wlock); 575 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex; 576 if (size) { 577 lwkt_reltoken(&wlock); 578 continue; 579 } 580 581 /* 582 * If there is no more to read in the pipe, reset its 583 * pointers to the beginning. This improves cache hit 584 * stats. 585 * 586 * We need both locks to modify both pointers, and there 587 * must also not be a write in progress or the uiomove() 588 * in the write might block and temporarily release 589 * its wlock, then reacquire and update windex. We are 590 * only serialized against reads, not writes. 591 * 592 * XXX should we even bother resetting the indices? It 593 * might actually be more cache efficient not to. 594 */ 595 if (rpipe->pipe_buffer.rindex == rpipe->pipe_buffer.windex && 596 rpipe->pipe_wip == 0) { 597 rpipe->pipe_buffer.rindex = 0; 598 rpipe->pipe_buffer.windex = 0; 599 } 600 601 /* 602 * Wait for more data. 603 * 604 * Pipe_state can only be set if both the rlock and wlock 605 * are held. 606 */ 607 rpipe->pipe_state |= PIPE_WANTR; 608 crit_enter(); 609 tsleep_interlock(rpipe); 610 lwkt_reltoken(&wlock); 611 error = tsleep(rpipe, PCATCH, "piperd", 0); 612 crit_exit(); 613 ++pipe_rblocked_count; 614 if (error) 615 break; 616 } 617 pipe_end_uio(rpipe, &rpipe->pipe_rip); 618 619 /* 620 * Uptime last access time 621 */ 622 if (error == 0 && nread) 623 vfs_timestamp(&rpipe->pipe_atime); 624 625 /* 626 * If we drained the FIFO more then half way then handle 627 * write blocking hysteresis. 628 * 629 * Note that PIPE_WANTW cannot be set by the writer without 630 * it holding both rlock and wlock, so we can test it 631 * while holding just rlock. 632 */ 633 if (notify_writer) { 634 if (rpipe->pipe_state & PIPE_WANTW) { 635 lwkt_gettoken(&wlock, &rpipe->pipe_wlock); 636 if (rpipe->pipe_state & PIPE_WANTW) { 637 rpipe->pipe_state &= ~PIPE_WANTW; 638 lwkt_reltoken(&wlock); 639 wakeup(rpipe); 640 } else { 641 lwkt_reltoken(&wlock); 642 } 643 } 644 } 645 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex; 646 lwkt_reltoken(&rlock); 647 648 /* 649 * If enough space is available in buffer then wakeup sel writers? 650 */ 651 if ((rpipe->pipe_buffer.size - size) >= PIPE_BUF) 652 pipeselwakeup(rpipe); 653 pipe_rel_mplock(&mpsave); 654 return (error); 655 } 656 657 /* 658 * MPALMOSTSAFE - acquires mplock 659 */ 660 static int 661 pipe_write(struct file *fp, struct uio *uio, struct ucred *cred, int fflags) 662 { 663 int error; 664 int orig_resid; 665 int nbio; 666 struct pipe *wpipe, *rpipe; 667 lwkt_tokref rlock; 668 lwkt_tokref wlock; 669 u_int windex; 670 u_int space; 671 u_int wcount; 672 int mpsave; 673 674 pipe_get_mplock(&mpsave); 675 676 /* 677 * Writes go to the peer. The peer will always exist. 678 */ 679 rpipe = (struct pipe *) fp->f_data; 680 wpipe = rpipe->pipe_peer; 681 lwkt_gettoken(&wlock, &wpipe->pipe_wlock); 682 if (wpipe->pipe_state & PIPE_WEOF) { 683 pipe_rel_mplock(&mpsave); 684 lwkt_reltoken(&wlock); 685 return (EPIPE); 686 } 687 688 /* 689 * Degenerate case (EPIPE takes prec) 690 */ 691 if (uio->uio_resid == 0) { 692 pipe_rel_mplock(&mpsave); 693 lwkt_reltoken(&wlock); 694 return(0); 695 } 696 697 /* 698 * Writes are serialized (start_uio must be called with wlock) 699 */ 700 error = pipe_start_uio(wpipe, &wpipe->pipe_wip); 701 if (error) { 702 pipe_rel_mplock(&mpsave); 703 lwkt_reltoken(&wlock); 704 return (error); 705 } 706 707 if (fflags & O_FBLOCKING) 708 nbio = 0; 709 else if (fflags & O_FNONBLOCKING) 710 nbio = 1; 711 else if (fp->f_flag & O_NONBLOCK) 712 nbio = 1; 713 else 714 nbio = 0; 715 716 /* 717 * If it is advantageous to resize the pipe buffer, do 718 * so. We are write-serialized so we can block safely. 719 */ 720 if ((wpipe->pipe_buffer.size <= PIPE_SIZE) && 721 (pipe_nbig < pipe_maxbig) && 722 wpipe->pipe_wantwcnt > 4 && 723 (wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex)) { 724 /* 725 * Recheck after lock. 726 */ 727 lwkt_gettoken(&rlock, &wpipe->pipe_rlock); 728 if ((wpipe->pipe_buffer.size <= PIPE_SIZE) && 729 (pipe_nbig < pipe_maxbig) && 730 (wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex)) { 731 atomic_add_int(&pipe_nbig, 1); 732 if (pipespace(wpipe, BIG_PIPE_SIZE) == 0) 733 ++pipe_bigcount; 734 else 735 atomic_subtract_int(&pipe_nbig, 1); 736 } 737 lwkt_reltoken(&rlock); 738 } 739 740 orig_resid = uio->uio_resid; 741 wcount = 0; 742 743 while (uio->uio_resid) { 744 if (wpipe->pipe_state & PIPE_WEOF) { 745 error = EPIPE; 746 break; 747 } 748 749 windex = wpipe->pipe_buffer.windex & 750 (wpipe->pipe_buffer.size - 1); 751 space = wpipe->pipe_buffer.size - 752 (wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex); 753 cpu_lfence(); 754 755 /* Writes of size <= PIPE_BUF must be atomic. */ 756 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF)) 757 space = 0; 758 759 /* 760 * Write to fill, read size handles write hysteresis. Also 761 * additional restrictions can cause select-based non-blocking 762 * writes to spin. 763 */ 764 if (space > 0) { 765 u_int segsize; 766 767 /* 768 * Transfer size is minimum of uio transfer 769 * and free space in pipe buffer. 770 * 771 * Limit each uiocopy to no more then PIPE_SIZE 772 * so we can keep the gravy train going on a 773 * SMP box. This doubles the performance for 774 * write sizes > 16K. Otherwise large writes 775 * wind up doing an inefficient synchronous 776 * ping-pong. 777 */ 778 if (space > (u_int)uio->uio_resid) 779 space = (u_int)uio->uio_resid; 780 if (space > PIPE_SIZE) 781 space = PIPE_SIZE; 782 783 /* 784 * First segment to transfer is minimum of 785 * transfer size and contiguous space in 786 * pipe buffer. If first segment to transfer 787 * is less than the transfer size, we've got 788 * a wraparound in the buffer. 789 */ 790 segsize = wpipe->pipe_buffer.size - windex; 791 if (segsize > space) 792 segsize = space; 793 794 #ifdef SMP 795 /* 796 * If this is the first loop and the reader is 797 * blocked, do a preemptive wakeup of the reader. 798 * 799 * On SMP the IPI latency plus the wlock interlock 800 * on the reader side is the fastest way to get the 801 * reader going. (The scheduler will hard loop on 802 * lock tokens). 803 * 804 * NOTE: We can't clear WANTR here without acquiring 805 * the rlock, which we don't want to do here! 806 */ 807 if ((wpipe->pipe_state & PIPE_WANTR) && pipe_mpsafe > 1) 808 wakeup(wpipe); 809 #endif 810 811 /* 812 * Transfer segment, which may include a wrap-around. 813 * Update windex to account for both all in one go 814 * so the reader can read() the data atomically. 815 */ 816 error = uiomove(&wpipe->pipe_buffer.buffer[windex], 817 segsize, uio); 818 if (error == 0 && segsize < space) { 819 segsize = space - segsize; 820 error = uiomove(&wpipe->pipe_buffer.buffer[0], 821 segsize, uio); 822 } 823 if (error) 824 break; 825 cpu_mfence(); 826 wpipe->pipe_buffer.windex += space; 827 wcount += space; 828 continue; 829 } 830 831 /* 832 * We need both the rlock and the wlock to interlock against 833 * the EOF, WANTW, and size checks, and to modify pipe_state. 834 * 835 * These are token locks so we do not have to worry about 836 * deadlocks. 837 */ 838 lwkt_gettoken(&rlock, &wpipe->pipe_rlock); 839 840 /* 841 * If the "read-side" has been blocked, wake it up now 842 * and yield to let it drain synchronously rather 843 * then block. 844 */ 845 if (wpipe->pipe_state & PIPE_WANTR) { 846 wpipe->pipe_state &= ~PIPE_WANTR; 847 wakeup(wpipe); 848 } 849 850 /* 851 * don't block on non-blocking I/O 852 */ 853 if (nbio) { 854 lwkt_reltoken(&rlock); 855 error = EAGAIN; 856 break; 857 } 858 859 /* 860 * re-test whether we have to block in the writer after 861 * acquiring both locks, in case the reader opened up 862 * some space. 863 */ 864 space = wpipe->pipe_buffer.size - 865 (wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex); 866 cpu_lfence(); 867 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF)) 868 space = 0; 869 870 /* 871 * We have no more space and have something to offer, 872 * wake up select/poll. 873 */ 874 if (space == 0) { 875 pipeselwakeup(wpipe); 876 ++wpipe->pipe_wantwcnt; 877 wpipe->pipe_state |= PIPE_WANTW; 878 error = tsleep(wpipe, PCATCH, "pipewr", 0); 879 ++pipe_wblocked_count; 880 } 881 lwkt_reltoken(&rlock); 882 883 /* 884 * Break out if we errored or the read side wants us to go 885 * away. 886 */ 887 if (error) 888 break; 889 if (wpipe->pipe_state & PIPE_WEOF) { 890 error = EPIPE; 891 break; 892 } 893 } 894 pipe_end_uio(wpipe, &wpipe->pipe_wip); 895 896 /* 897 * If we have put any characters in the buffer, we wake up 898 * the reader. 899 * 900 * Both rlock and wlock are required to be able to modify pipe_state. 901 */ 902 if (wpipe->pipe_buffer.windex != wpipe->pipe_buffer.rindex) { 903 if (wpipe->pipe_state & PIPE_WANTR) { 904 lwkt_gettoken(&rlock, &wpipe->pipe_rlock); 905 if (wpipe->pipe_state & PIPE_WANTR) { 906 wpipe->pipe_state &= ~PIPE_WANTR; 907 lwkt_reltoken(&rlock); 908 wakeup(wpipe); 909 } else { 910 lwkt_reltoken(&rlock); 911 } 912 } 913 } 914 915 /* 916 * Don't return EPIPE if I/O was successful 917 */ 918 if ((wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex) && 919 (uio->uio_resid == 0) && 920 (error == EPIPE)) { 921 error = 0; 922 } 923 924 if (error == 0) 925 vfs_timestamp(&wpipe->pipe_mtime); 926 927 /* 928 * We have something to offer, 929 * wake up select/poll. 930 */ 931 space = wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex; 932 lwkt_reltoken(&wlock); 933 if (space) 934 pipeselwakeup(wpipe); 935 pipe_rel_mplock(&mpsave); 936 return (error); 937 } 938 939 /* 940 * MPALMOSTSAFE - acquires mplock 941 * 942 * we implement a very minimal set of ioctls for compatibility with sockets. 943 */ 944 int 945 pipe_ioctl(struct file *fp, u_long cmd, caddr_t data, struct ucred *cred) 946 { 947 struct pipe *mpipe; 948 lwkt_tokref rlock; 949 lwkt_tokref wlock; 950 int error; 951 int mpsave; 952 953 pipe_get_mplock(&mpsave); 954 mpipe = (struct pipe *)fp->f_data; 955 956 lwkt_gettoken(&rlock, &mpipe->pipe_rlock); 957 lwkt_gettoken(&wlock, &mpipe->pipe_wlock); 958 959 switch (cmd) { 960 case FIOASYNC: 961 if (*(int *)data) { 962 mpipe->pipe_state |= PIPE_ASYNC; 963 } else { 964 mpipe->pipe_state &= ~PIPE_ASYNC; 965 } 966 error = 0; 967 break; 968 case FIONREAD: 969 *(int *)data = mpipe->pipe_buffer.windex - 970 mpipe->pipe_buffer.rindex; 971 error = 0; 972 break; 973 case FIOSETOWN: 974 get_mplock(); 975 error = fsetown(*(int *)data, &mpipe->pipe_sigio); 976 rel_mplock(); 977 break; 978 case FIOGETOWN: 979 *(int *)data = fgetown(mpipe->pipe_sigio); 980 error = 0; 981 break; 982 case TIOCSPGRP: 983 /* This is deprecated, FIOSETOWN should be used instead. */ 984 get_mplock(); 985 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio); 986 rel_mplock(); 987 break; 988 989 case TIOCGPGRP: 990 /* This is deprecated, FIOGETOWN should be used instead. */ 991 *(int *)data = -fgetown(mpipe->pipe_sigio); 992 error = 0; 993 break; 994 default: 995 error = ENOTTY; 996 break; 997 } 998 lwkt_reltoken(&rlock); 999 lwkt_reltoken(&wlock); 1000 pipe_rel_mplock(&mpsave); 1001 1002 return (error); 1003 } 1004 1005 /* 1006 * MPALMOSTSAFE - acquires mplock 1007 */ 1008 int 1009 pipe_poll(struct file *fp, int events, struct ucred *cred) 1010 { 1011 struct pipe *rpipe; 1012 struct pipe *wpipe; 1013 int revents = 0; 1014 u_int space; 1015 int mpsave; 1016 1017 pipe_get_mplock(&mpsave); 1018 rpipe = (struct pipe *)fp->f_data; 1019 wpipe = rpipe->pipe_peer; 1020 if (events & (POLLIN | POLLRDNORM)) { 1021 if ((rpipe->pipe_buffer.windex != rpipe->pipe_buffer.rindex) || 1022 (rpipe->pipe_state & PIPE_REOF)) { 1023 revents |= events & (POLLIN | POLLRDNORM); 1024 } 1025 } 1026 1027 if (events & (POLLOUT | POLLWRNORM)) { 1028 if (wpipe == NULL || (wpipe->pipe_state & PIPE_WEOF)) { 1029 revents |= events & (POLLOUT | POLLWRNORM); 1030 } else { 1031 space = wpipe->pipe_buffer.windex - 1032 wpipe->pipe_buffer.rindex; 1033 space = wpipe->pipe_buffer.size - space; 1034 if (space >= PIPE_BUF) 1035 revents |= events & (POLLOUT | POLLWRNORM); 1036 } 1037 } 1038 1039 if ((rpipe->pipe_state & PIPE_REOF) || 1040 (wpipe == NULL) || 1041 (wpipe->pipe_state & PIPE_WEOF)) 1042 revents |= POLLHUP; 1043 1044 if (revents == 0) { 1045 if (events & (POLLIN | POLLRDNORM)) { 1046 selrecord(curthread, &rpipe->pipe_sel); 1047 rpipe->pipe_state |= PIPE_SEL; 1048 } 1049 1050 if (events & (POLLOUT | POLLWRNORM)) { 1051 selrecord(curthread, &wpipe->pipe_sel); 1052 wpipe->pipe_state |= PIPE_SEL; 1053 } 1054 } 1055 pipe_rel_mplock(&mpsave); 1056 return (revents); 1057 } 1058 1059 /* 1060 * MPSAFE 1061 */ 1062 static int 1063 pipe_stat(struct file *fp, struct stat *ub, struct ucred *cred) 1064 { 1065 struct pipe *pipe; 1066 int mpsave; 1067 1068 pipe_get_mplock(&mpsave); 1069 pipe = (struct pipe *)fp->f_data; 1070 1071 bzero((caddr_t)ub, sizeof(*ub)); 1072 ub->st_mode = S_IFIFO; 1073 ub->st_blksize = pipe->pipe_buffer.size; 1074 ub->st_size = pipe->pipe_buffer.windex - pipe->pipe_buffer.rindex; 1075 ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize; 1076 ub->st_atimespec = pipe->pipe_atime; 1077 ub->st_mtimespec = pipe->pipe_mtime; 1078 ub->st_ctimespec = pipe->pipe_ctime; 1079 /* 1080 * Left as 0: st_dev, st_ino, st_nlink, st_uid, st_gid, st_rdev, 1081 * st_flags, st_gen. 1082 * XXX (st_dev, st_ino) should be unique. 1083 */ 1084 pipe_rel_mplock(&mpsave); 1085 return (0); 1086 } 1087 1088 /* 1089 * MPALMOSTSAFE - acquires mplock 1090 */ 1091 static int 1092 pipe_close(struct file *fp) 1093 { 1094 struct pipe *cpipe; 1095 1096 get_mplock(); 1097 cpipe = (struct pipe *)fp->f_data; 1098 fp->f_ops = &badfileops; 1099 fp->f_data = NULL; 1100 funsetown(cpipe->pipe_sigio); 1101 pipeclose(cpipe); 1102 rel_mplock(); 1103 return (0); 1104 } 1105 1106 /* 1107 * Shutdown one or both directions of a full-duplex pipe. 1108 * 1109 * MPALMOSTSAFE - acquires mplock 1110 */ 1111 static int 1112 pipe_shutdown(struct file *fp, int how) 1113 { 1114 struct pipe *rpipe; 1115 struct pipe *wpipe; 1116 int error = EPIPE; 1117 lwkt_tokref rpipe_rlock; 1118 lwkt_tokref rpipe_wlock; 1119 lwkt_tokref wpipe_rlock; 1120 lwkt_tokref wpipe_wlock; 1121 int mpsave; 1122 1123 pipe_get_mplock(&mpsave); 1124 rpipe = (struct pipe *)fp->f_data; 1125 wpipe = rpipe->pipe_peer; 1126 1127 /* 1128 * We modify pipe_state on both pipes, which means we need 1129 * all four tokens! 1130 */ 1131 lwkt_gettoken(&rpipe_rlock, &rpipe->pipe_rlock); 1132 lwkt_gettoken(&rpipe_wlock, &rpipe->pipe_wlock); 1133 lwkt_gettoken(&wpipe_rlock, &wpipe->pipe_rlock); 1134 lwkt_gettoken(&wpipe_wlock, &wpipe->pipe_wlock); 1135 1136 switch(how) { 1137 case SHUT_RDWR: 1138 case SHUT_RD: 1139 rpipe->pipe_state |= PIPE_REOF; 1140 wpipe->pipe_state |= PIPE_WEOF; 1141 if (rpipe->pipe_state & PIPE_WANTR) { 1142 rpipe->pipe_state &= ~PIPE_WANTR; 1143 wakeup(rpipe); 1144 } 1145 if (wpipe->pipe_state & PIPE_WANTW) { 1146 wpipe->pipe_state &= ~PIPE_WANTW; 1147 wakeup(wpipe); 1148 } 1149 pipeselwakeup(rpipe); 1150 error = 0; 1151 if (how == SHUT_RD) 1152 break; 1153 /* fall through */ 1154 case SHUT_WR: 1155 wpipe->pipe_state |= PIPE_WEOF; 1156 rpipe->pipe_state |= PIPE_REOF; 1157 if (wpipe->pipe_state & PIPE_WANTW) { 1158 wpipe->pipe_state &= ~PIPE_WANTW; 1159 wakeup(wpipe); 1160 } 1161 if (rpipe->pipe_state & PIPE_WANTR) { 1162 rpipe->pipe_state &= ~PIPE_WANTR; 1163 wakeup(rpipe); 1164 } 1165 pipeselwakeup(wpipe); 1166 error = 0; 1167 break; 1168 } 1169 1170 lwkt_reltoken(&rpipe_rlock); 1171 lwkt_reltoken(&rpipe_wlock); 1172 lwkt_reltoken(&wpipe_rlock); 1173 lwkt_reltoken(&wpipe_wlock); 1174 1175 pipe_rel_mplock(&mpsave); 1176 return (error); 1177 } 1178 1179 static void 1180 pipe_free_kmem(struct pipe *cpipe) 1181 { 1182 if (cpipe->pipe_buffer.buffer != NULL) { 1183 if (cpipe->pipe_buffer.size > PIPE_SIZE) 1184 atomic_subtract_int(&pipe_nbig, 1); 1185 kmem_free(&kernel_map, 1186 (vm_offset_t)cpipe->pipe_buffer.buffer, 1187 cpipe->pipe_buffer.size); 1188 cpipe->pipe_buffer.buffer = NULL; 1189 cpipe->pipe_buffer.object = NULL; 1190 } 1191 } 1192 1193 /* 1194 * Close the pipe. The slock must be held to interlock against simultanious 1195 * closes. The rlock and wlock must be held to adjust the pipe_state. 1196 */ 1197 static void 1198 pipeclose(struct pipe *cpipe) 1199 { 1200 globaldata_t gd; 1201 struct pipe *ppipe; 1202 lwkt_tokref cpipe_rlock; 1203 lwkt_tokref cpipe_wlock; 1204 lwkt_tokref ppipe_rlock; 1205 lwkt_tokref ppipe_wlock; 1206 1207 if (cpipe == NULL) 1208 return; 1209 1210 /* 1211 * The slock may not have been allocated yet (close during 1212 * initialization) 1213 * 1214 * We need both the read and write tokens to modify pipe_state. 1215 */ 1216 if (cpipe->pipe_slock) 1217 lockmgr(cpipe->pipe_slock, LK_EXCLUSIVE); 1218 lwkt_gettoken(&cpipe_rlock, &cpipe->pipe_rlock); 1219 lwkt_gettoken(&cpipe_wlock, &cpipe->pipe_wlock); 1220 1221 /* 1222 * Set our state, wakeup anyone waiting in select, and 1223 * wakeup anyone blocked on our pipe. 1224 */ 1225 cpipe->pipe_state |= PIPE_CLOSED | PIPE_REOF | PIPE_WEOF; 1226 pipeselwakeup(cpipe); 1227 if (cpipe->pipe_state & (PIPE_WANTR | PIPE_WANTW)) { 1228 cpipe->pipe_state &= ~(PIPE_WANTR | PIPE_WANTW); 1229 wakeup(cpipe); 1230 } 1231 1232 /* 1233 * Disconnect from peer 1234 */ 1235 if ((ppipe = cpipe->pipe_peer) != NULL) { 1236 lwkt_gettoken(&ppipe_rlock, &ppipe->pipe_rlock); 1237 lwkt_gettoken(&ppipe_wlock, &ppipe->pipe_wlock); 1238 ppipe->pipe_state |= PIPE_REOF; 1239 pipeselwakeup(ppipe); 1240 if (ppipe->pipe_state & (PIPE_WANTR | PIPE_WANTW)) { 1241 ppipe->pipe_state &= ~(PIPE_WANTR | PIPE_WANTW); 1242 wakeup(ppipe); 1243 } 1244 if (SLIST_FIRST(&ppipe->pipe_sel.si_note)) { 1245 get_mplock(); 1246 KNOTE(&ppipe->pipe_sel.si_note, 0); 1247 rel_mplock(); 1248 } 1249 lwkt_reltoken(&ppipe_rlock); 1250 lwkt_reltoken(&ppipe_wlock); 1251 } 1252 1253 /* 1254 * If the peer is also closed we can free resources for both 1255 * sides, otherwise we leave our side intact to deal with any 1256 * races (since we only have the slock). 1257 */ 1258 if (ppipe && (ppipe->pipe_state & PIPE_CLOSED)) { 1259 cpipe->pipe_peer = NULL; 1260 ppipe->pipe_peer = NULL; 1261 ppipe->pipe_slock = NULL; /* we will free the slock */ 1262 pipeclose(ppipe); 1263 ppipe = NULL; 1264 } 1265 1266 lwkt_reltoken(&cpipe_rlock); 1267 lwkt_reltoken(&cpipe_wlock); 1268 if (cpipe->pipe_slock) 1269 lockmgr(cpipe->pipe_slock, LK_RELEASE); 1270 1271 /* 1272 * If we disassociated from our peer we can free resources 1273 */ 1274 if (ppipe == NULL) { 1275 gd = mycpu; 1276 if (cpipe->pipe_slock) { 1277 kfree(cpipe->pipe_slock, M_PIPE); 1278 cpipe->pipe_slock = NULL; 1279 } 1280 if (gd->gd_pipeqcount >= pipe_maxcache || 1281 cpipe->pipe_buffer.size != PIPE_SIZE 1282 ) { 1283 pipe_free_kmem(cpipe); 1284 kfree(cpipe, M_PIPE); 1285 } else { 1286 cpipe->pipe_state = 0; 1287 cpipe->pipe_peer = gd->gd_pipeq; 1288 gd->gd_pipeq = cpipe; 1289 ++gd->gd_pipeqcount; 1290 } 1291 } 1292 } 1293 1294 /* 1295 * MPALMOSTSAFE - acquires mplock 1296 */ 1297 static int 1298 pipe_kqfilter(struct file *fp, struct knote *kn) 1299 { 1300 struct pipe *cpipe; 1301 1302 get_mplock(); 1303 cpipe = (struct pipe *)kn->kn_fp->f_data; 1304 1305 switch (kn->kn_filter) { 1306 case EVFILT_READ: 1307 kn->kn_fop = &pipe_rfiltops; 1308 break; 1309 case EVFILT_WRITE: 1310 kn->kn_fop = &pipe_wfiltops; 1311 cpipe = cpipe->pipe_peer; 1312 if (cpipe == NULL) { 1313 /* other end of pipe has been closed */ 1314 rel_mplock(); 1315 return (EPIPE); 1316 } 1317 break; 1318 default: 1319 return (1); 1320 } 1321 kn->kn_hook = (caddr_t)cpipe; 1322 1323 SLIST_INSERT_HEAD(&cpipe->pipe_sel.si_note, kn, kn_selnext); 1324 rel_mplock(); 1325 return (0); 1326 } 1327 1328 static void 1329 filt_pipedetach(struct knote *kn) 1330 { 1331 struct pipe *cpipe = (struct pipe *)kn->kn_hook; 1332 1333 SLIST_REMOVE(&cpipe->pipe_sel.si_note, kn, knote, kn_selnext); 1334 } 1335 1336 /*ARGSUSED*/ 1337 static int 1338 filt_piperead(struct knote *kn, long hint) 1339 { 1340 struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data; 1341 1342 kn->kn_data = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex; 1343 1344 /* XXX RACE */ 1345 if (rpipe->pipe_state & PIPE_REOF) { 1346 kn->kn_flags |= EV_EOF; 1347 return (1); 1348 } 1349 return (kn->kn_data > 0); 1350 } 1351 1352 /*ARGSUSED*/ 1353 static int 1354 filt_pipewrite(struct knote *kn, long hint) 1355 { 1356 struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data; 1357 struct pipe *wpipe = rpipe->pipe_peer; 1358 u_int32_t space; 1359 1360 /* XXX RACE */ 1361 if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_WEOF)) { 1362 kn->kn_data = 0; 1363 kn->kn_flags |= EV_EOF; 1364 return (1); 1365 } 1366 space = wpipe->pipe_buffer.windex - 1367 wpipe->pipe_buffer.rindex; 1368 space = wpipe->pipe_buffer.size - space; 1369 kn->kn_data = space; 1370 return (kn->kn_data >= PIPE_BUF); 1371 } 1372