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 */ 21 22 /* 23 * This file contains a high-performance replacement for the socket-based 24 * pipes scheme originally used in FreeBSD/4.4Lite. It does not support 25 * all features of sockets, but does do everything that pipes normally 26 * do. 27 */ 28 #include <sys/param.h> 29 #include <sys/systm.h> 30 #include <sys/kernel.h> 31 #include <sys/proc.h> 32 #include <sys/fcntl.h> 33 #include <sys/file.h> 34 #include <sys/filedesc.h> 35 #include <sys/filio.h> 36 #include <sys/ttycom.h> 37 #include <sys/stat.h> 38 #include <sys/signalvar.h> 39 #include <sys/sysproto.h> 40 #include <sys/pipe.h> 41 #include <sys/vnode.h> 42 #include <sys/uio.h> 43 #include <sys/event.h> 44 #include <sys/globaldata.h> 45 #include <sys/module.h> 46 #include <sys/malloc.h> 47 #include <sys/sysctl.h> 48 #include <sys/socket.h> 49 50 #include <vm/vm.h> 51 #include <vm/vm_param.h> 52 #include <sys/lock.h> 53 #include <vm/vm_object.h> 54 #include <vm/vm_kern.h> 55 #include <vm/vm_extern.h> 56 #include <vm/pmap.h> 57 #include <vm/vm_map.h> 58 #include <vm/vm_page.h> 59 #include <vm/vm_zone.h> 60 61 #include <sys/file2.h> 62 #include <sys/signal2.h> 63 64 #include <machine/cpufunc.h> 65 66 /* 67 * interfaces to the outside world 68 */ 69 static int pipe_read (struct file *fp, struct uio *uio, 70 struct ucred *cred, int flags); 71 static int pipe_write (struct file *fp, struct uio *uio, 72 struct ucred *cred, int flags); 73 static int pipe_close (struct file *fp); 74 static int pipe_shutdown (struct file *fp, int how); 75 static int pipe_kqfilter (struct file *fp, struct knote *kn); 76 static int pipe_stat (struct file *fp, struct stat *sb, struct ucred *cred); 77 static int pipe_ioctl (struct file *fp, u_long cmd, caddr_t data, 78 struct ucred *cred, struct sysmsg *msg); 79 80 static struct fileops pipeops = { 81 .fo_read = pipe_read, 82 .fo_write = pipe_write, 83 .fo_ioctl = pipe_ioctl, 84 .fo_kqfilter = pipe_kqfilter, 85 .fo_stat = pipe_stat, 86 .fo_close = pipe_close, 87 .fo_shutdown = pipe_shutdown 88 }; 89 90 static void filt_pipedetach(struct knote *kn); 91 static int filt_piperead(struct knote *kn, long hint); 92 static int filt_pipewrite(struct knote *kn, long hint); 93 94 static struct filterops pipe_rfiltops = 95 { FILTEROP_ISFD|FILTEROP_MPSAFE, NULL, filt_pipedetach, filt_piperead }; 96 static struct filterops pipe_wfiltops = 97 { FILTEROP_ISFD|FILTEROP_MPSAFE, NULL, filt_pipedetach, filt_pipewrite }; 98 99 MALLOC_DEFINE(M_PIPE, "pipe", "pipe structures"); 100 101 /* 102 * Default pipe buffer size(s), this can be kind-of large now because pipe 103 * space is pageable. The pipe code will try to maintain locality of 104 * reference for performance reasons, so small amounts of outstanding I/O 105 * will not wipe the cache. 106 */ 107 #define MINPIPESIZE (PIPE_SIZE/3) 108 #define MAXPIPESIZE (2*PIPE_SIZE/3) 109 110 /* 111 * Limit the number of "big" pipes 112 */ 113 #define LIMITBIGPIPES 64 114 #define PIPEQ_MAX_CACHE 16 /* per-cpu pipe structure cache */ 115 116 static int pipe_maxbig = LIMITBIGPIPES; 117 static int pipe_maxcache = PIPEQ_MAX_CACHE; 118 static int pipe_bigcount; 119 static int pipe_nbig; 120 static int pipe_bcache_alloc; 121 static int pipe_bkmem_alloc; 122 static int pipe_rblocked_count; 123 static int pipe_wblocked_count; 124 125 SYSCTL_NODE(_kern, OID_AUTO, pipe, CTLFLAG_RW, 0, "Pipe operation"); 126 SYSCTL_INT(_kern_pipe, OID_AUTO, nbig, 127 CTLFLAG_RD, &pipe_nbig, 0, "number of big pipes allocated"); 128 SYSCTL_INT(_kern_pipe, OID_AUTO, bigcount, 129 CTLFLAG_RW, &pipe_bigcount, 0, "number of times pipe expanded"); 130 SYSCTL_INT(_kern_pipe, OID_AUTO, rblocked, 131 CTLFLAG_RW, &pipe_rblocked_count, 0, "number of times pipe expanded"); 132 SYSCTL_INT(_kern_pipe, OID_AUTO, wblocked, 133 CTLFLAG_RW, &pipe_wblocked_count, 0, "number of times pipe expanded"); 134 SYSCTL_INT(_kern_pipe, OID_AUTO, maxcache, 135 CTLFLAG_RW, &pipe_maxcache, 0, "max pipes cached per-cpu"); 136 SYSCTL_INT(_kern_pipe, OID_AUTO, maxbig, 137 CTLFLAG_RW, &pipe_maxbig, 0, "max number of big pipes"); 138 static int pipe_delay = 5000; /* 5uS default */ 139 SYSCTL_INT(_kern_pipe, OID_AUTO, delay, 140 CTLFLAG_RW, &pipe_delay, 0, "SMP delay optimization in ns"); 141 #if !defined(NO_PIPE_SYSCTL_STATS) 142 SYSCTL_INT(_kern_pipe, OID_AUTO, bcache_alloc, 143 CTLFLAG_RW, &pipe_bcache_alloc, 0, "pipe buffer from pcpu cache"); 144 SYSCTL_INT(_kern_pipe, OID_AUTO, bkmem_alloc, 145 CTLFLAG_RW, &pipe_bkmem_alloc, 0, "pipe buffer from kmem"); 146 #endif 147 148 /* 149 * Auto-size pipe cache to reduce kmem allocations and frees. 150 */ 151 static 152 void 153 pipeinit(void *dummy) 154 { 155 size_t mbytes = kmem_lim_size(); 156 157 if (pipe_maxbig == LIMITBIGPIPES) { 158 if (mbytes >= 7 * 1024) 159 pipe_maxbig *= 2; 160 if (mbytes >= 15 * 1024) 161 pipe_maxbig *= 2; 162 } 163 if (pipe_maxcache == PIPEQ_MAX_CACHE) { 164 if (mbytes >= 7 * 1024) 165 pipe_maxcache *= 2; 166 if (mbytes >= 15 * 1024) 167 pipe_maxcache *= 2; 168 } 169 } 170 SYSINIT(kmem, SI_BOOT2_MACHDEP, SI_ORDER_ANY, pipeinit, NULL) 171 172 static void pipeclose (struct pipe *cpipe); 173 static void pipe_free_kmem (struct pipe *cpipe); 174 static int pipe_create (struct pipe **cpipep); 175 static int pipespace (struct pipe *cpipe, int size); 176 177 static __inline void 178 pipewakeup(struct pipe *cpipe, int dosigio) 179 { 180 if (dosigio && (cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio) { 181 lwkt_gettoken(&proc_token); 182 pgsigio(cpipe->pipe_sigio, SIGIO, 0); 183 lwkt_reltoken(&proc_token); 184 } 185 KNOTE(&cpipe->pipe_kq.ki_note, 0); 186 } 187 188 /* 189 * These routines are called before and after a UIO. The UIO 190 * may block, causing our held tokens to be lost temporarily. 191 * 192 * We use these routines to serialize reads against other reads 193 * and writes against other writes. 194 * 195 * The read token is held on entry so *ipp does not race. 196 */ 197 static __inline int 198 pipe_start_uio(struct pipe *cpipe, int *ipp) 199 { 200 int error; 201 202 while (*ipp) { 203 *ipp = -1; 204 error = tsleep(ipp, PCATCH, "pipexx", 0); 205 if (error) 206 return (error); 207 } 208 *ipp = 1; 209 return (0); 210 } 211 212 static __inline void 213 pipe_end_uio(struct pipe *cpipe, int *ipp) 214 { 215 if (*ipp < 0) { 216 *ipp = 0; 217 wakeup(ipp); 218 } else { 219 KKASSERT(*ipp > 0); 220 *ipp = 0; 221 } 222 } 223 224 /* 225 * The pipe system call for the DTYPE_PIPE type of pipes 226 * 227 * pipe_args(int dummy) 228 * 229 * MPSAFE 230 */ 231 int 232 sys_pipe(struct pipe_args *uap) 233 { 234 struct thread *td = curthread; 235 struct filedesc *fdp = td->td_proc->p_fd; 236 struct file *rf, *wf; 237 struct pipe *rpipe, *wpipe; 238 int fd1, fd2, error; 239 240 rpipe = wpipe = NULL; 241 if (pipe_create(&rpipe) || pipe_create(&wpipe)) { 242 pipeclose(rpipe); 243 pipeclose(wpipe); 244 return (ENFILE); 245 } 246 247 error = falloc(td->td_lwp, &rf, &fd1); 248 if (error) { 249 pipeclose(rpipe); 250 pipeclose(wpipe); 251 return (error); 252 } 253 uap->sysmsg_fds[0] = fd1; 254 255 /* 256 * Warning: once we've gotten past allocation of the fd for the 257 * read-side, we can only drop the read side via fdrop() in order 258 * to avoid races against processes which manage to dup() the read 259 * side while we are blocked trying to allocate the write side. 260 */ 261 rf->f_type = DTYPE_PIPE; 262 rf->f_flag = FREAD | FWRITE; 263 rf->f_ops = &pipeops; 264 rf->f_data = rpipe; 265 error = falloc(td->td_lwp, &wf, &fd2); 266 if (error) { 267 fsetfd(fdp, NULL, fd1); 268 fdrop(rf); 269 /* rpipe has been closed by fdrop(). */ 270 pipeclose(wpipe); 271 return (error); 272 } 273 wf->f_type = DTYPE_PIPE; 274 wf->f_flag = FREAD | FWRITE; 275 wf->f_ops = &pipeops; 276 wf->f_data = wpipe; 277 uap->sysmsg_fds[1] = fd2; 278 279 rpipe->pipe_slock = kmalloc(sizeof(struct lock), 280 M_PIPE, M_WAITOK|M_ZERO); 281 wpipe->pipe_slock = rpipe->pipe_slock; 282 rpipe->pipe_peer = wpipe; 283 wpipe->pipe_peer = rpipe; 284 lockinit(rpipe->pipe_slock, "pipecl", 0, 0); 285 286 /* 287 * Once activated the peer relationship remains valid until 288 * both sides are closed. 289 */ 290 fsetfd(fdp, rf, fd1); 291 fsetfd(fdp, wf, fd2); 292 fdrop(rf); 293 fdrop(wf); 294 295 return (0); 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 *cpipe, int size) 306 { 307 struct vm_object *object; 308 caddr_t buffer; 309 int npages, error; 310 311 npages = round_page(size) / PAGE_SIZE; 312 object = cpipe->pipe_buffer.object; 313 314 /* 315 * [re]create the object if necessary and reserve space for it 316 * in the kernel_map. The object and memory are pageable. On 317 * success, free the old resources before assigning the new 318 * ones. 319 */ 320 if (object == NULL || object->size != npages) { 321 object = vm_object_allocate(OBJT_DEFAULT, npages); 322 buffer = (caddr_t)vm_map_min(&kernel_map); 323 324 error = vm_map_find(&kernel_map, object, 0, 325 (vm_offset_t *)&buffer, 326 size, PAGE_SIZE, 327 1, VM_MAPTYPE_NORMAL, 328 VM_PROT_ALL, VM_PROT_ALL, 329 0); 330 331 if (error != KERN_SUCCESS) { 332 vm_object_deallocate(object); 333 return (ENOMEM); 334 } 335 pipe_free_kmem(cpipe); 336 cpipe->pipe_buffer.object = object; 337 cpipe->pipe_buffer.buffer = buffer; 338 cpipe->pipe_buffer.size = size; 339 ++pipe_bkmem_alloc; 340 } else { 341 ++pipe_bcache_alloc; 342 } 343 cpipe->pipe_buffer.rindex = 0; 344 cpipe->pipe_buffer.windex = 0; 345 return (0); 346 } 347 348 /* 349 * Initialize and allocate VM and memory for pipe, pulling the pipe from 350 * our per-cpu cache if possible. For now make sure it is sized for the 351 * smaller PIPE_SIZE default. 352 */ 353 static int 354 pipe_create(struct pipe **cpipep) 355 { 356 globaldata_t gd = mycpu; 357 struct pipe *cpipe; 358 int error; 359 360 if ((cpipe = gd->gd_pipeq) != NULL) { 361 gd->gd_pipeq = cpipe->pipe_peer; 362 --gd->gd_pipeqcount; 363 cpipe->pipe_peer = NULL; 364 cpipe->pipe_wantwcnt = 0; 365 } else { 366 cpipe = kmalloc(sizeof(struct pipe), M_PIPE, M_WAITOK|M_ZERO); 367 } 368 *cpipep = cpipe; 369 if ((error = pipespace(cpipe, PIPE_SIZE)) != 0) 370 return (error); 371 vfs_timestamp(&cpipe->pipe_ctime); 372 cpipe->pipe_atime = cpipe->pipe_ctime; 373 cpipe->pipe_mtime = cpipe->pipe_ctime; 374 lwkt_token_init(&cpipe->pipe_rlock, "piper"); 375 lwkt_token_init(&cpipe->pipe_wlock, "pipew"); 376 return (0); 377 } 378 379 static int 380 pipe_read(struct file *fp, struct uio *uio, struct ucred *cred, int fflags) 381 { 382 struct pipe *rpipe; 383 struct pipe *wpipe; 384 int error; 385 size_t nread = 0; 386 int nbio; 387 u_int size; /* total bytes available */ 388 u_int nsize; /* total bytes to read */ 389 u_int rindex; /* contiguous bytes available */ 390 int notify_writer; 391 int bigread; 392 int bigcount; 393 394 atomic_set_int(&curthread->td_mpflags, TDF_MP_BATCH_DEMARC); 395 396 if (uio->uio_resid == 0) 397 return(0); 398 399 /* 400 * Setup locks, calculate nbio 401 */ 402 rpipe = (struct pipe *)fp->f_data; 403 wpipe = rpipe->pipe_peer; 404 lwkt_gettoken(&rpipe->pipe_rlock); 405 406 if (fflags & O_FBLOCKING) 407 nbio = 0; 408 else if (fflags & O_FNONBLOCKING) 409 nbio = 1; 410 else if (fp->f_flag & O_NONBLOCK) 411 nbio = 1; 412 else 413 nbio = 0; 414 415 /* 416 * Reads are serialized. Note however that pipe_buffer.buffer and 417 * pipe_buffer.size can change out from under us when the number 418 * of bytes in the buffer are zero due to the write-side doing a 419 * pipespace(). 420 */ 421 error = pipe_start_uio(rpipe, &rpipe->pipe_rip); 422 if (error) { 423 lwkt_reltoken(&rpipe->pipe_rlock); 424 return (error); 425 } 426 notify_writer = 0; 427 428 bigread = (uio->uio_resid > 10 * 1024 * 1024); 429 bigcount = 10; 430 431 while (uio->uio_resid) { 432 /* 433 * Don't hog the cpu. 434 */ 435 if (bigread && --bigcount == 0) { 436 lwkt_user_yield(); 437 bigcount = 10; 438 if (CURSIG(curthread->td_lwp)) { 439 error = EINTR; 440 break; 441 } 442 } 443 444 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex; 445 cpu_lfence(); 446 if (size) { 447 rindex = rpipe->pipe_buffer.rindex & 448 (rpipe->pipe_buffer.size - 1); 449 nsize = size; 450 if (nsize > rpipe->pipe_buffer.size - rindex) 451 nsize = rpipe->pipe_buffer.size - rindex; 452 nsize = szmin(nsize, uio->uio_resid); 453 454 error = uiomove(&rpipe->pipe_buffer.buffer[rindex], 455 nsize, uio); 456 if (error) 457 break; 458 cpu_mfence(); 459 rpipe->pipe_buffer.rindex += nsize; 460 nread += nsize; 461 462 /* 463 * If the FIFO is still over half full just continue 464 * and do not try to notify the writer yet. 465 */ 466 if (size - nsize >= (rpipe->pipe_buffer.size >> 1)) { 467 notify_writer = 0; 468 continue; 469 } 470 471 /* 472 * When the FIFO is less then half full notify any 473 * waiting writer. WANTW can be checked while 474 * holding just the rlock. 475 */ 476 notify_writer = 1; 477 if ((rpipe->pipe_state & PIPE_WANTW) == 0) 478 continue; 479 } 480 481 /* 482 * If the "write-side" was blocked we wake it up. This code 483 * is reached either when the buffer is completely emptied 484 * or if it becomes more then half-empty. 485 * 486 * Pipe_state can only be modified if both the rlock and 487 * wlock are held. 488 */ 489 if (rpipe->pipe_state & PIPE_WANTW) { 490 lwkt_gettoken(&rpipe->pipe_wlock); 491 if (rpipe->pipe_state & PIPE_WANTW) { 492 rpipe->pipe_state &= ~PIPE_WANTW; 493 lwkt_reltoken(&rpipe->pipe_wlock); 494 wakeup(rpipe); 495 } else { 496 lwkt_reltoken(&rpipe->pipe_wlock); 497 } 498 } 499 500 /* 501 * Pick up our copy loop again if the writer sent data to 502 * us while we were messing around. 503 * 504 * On a SMP box poll up to pipe_delay nanoseconds for new 505 * data. Typically a value of 2000 to 4000 is sufficient 506 * to eradicate most IPIs/tsleeps/wakeups when a pipe 507 * is used for synchronous communications with small packets, 508 * and 8000 or so (8uS) will pipeline large buffer xfers 509 * between cpus over a pipe. 510 * 511 * For synchronous communications a hit means doing a 512 * full Awrite-Bread-Bwrite-Aread cycle in less then 2uS, 513 * where as miss requiring a tsleep/wakeup sequence 514 * will take 7uS or more. 515 */ 516 if (rpipe->pipe_buffer.windex != rpipe->pipe_buffer.rindex) 517 continue; 518 519 #ifdef _RDTSC_SUPPORTED_ 520 if (pipe_delay) { 521 int64_t tsc_target; 522 int good = 0; 523 524 tsc_target = tsc_get_target(pipe_delay); 525 while (tsc_test_target(tsc_target) == 0) { 526 if (rpipe->pipe_buffer.windex != 527 rpipe->pipe_buffer.rindex) { 528 good = 1; 529 break; 530 } 531 } 532 if (good) 533 continue; 534 } 535 #endif 536 537 /* 538 * Detect EOF condition, do not set error. 539 */ 540 if (rpipe->pipe_state & PIPE_REOF) 541 break; 542 543 /* 544 * Break if some data was read, or if this was a non-blocking 545 * read. 546 */ 547 if (nread > 0) 548 break; 549 550 if (nbio) { 551 error = EAGAIN; 552 break; 553 } 554 555 /* 556 * Last chance, interlock with WANTR. 557 */ 558 lwkt_gettoken(&rpipe->pipe_wlock); 559 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex; 560 if (size) { 561 lwkt_reltoken(&rpipe->pipe_wlock); 562 continue; 563 } 564 565 /* 566 * Retest EOF - acquiring a new token can temporarily release 567 * tokens already held. 568 */ 569 if (rpipe->pipe_state & PIPE_REOF) { 570 lwkt_reltoken(&rpipe->pipe_wlock); 571 break; 572 } 573 574 /* 575 * If there is no more to read in the pipe, reset its 576 * pointers to the beginning. This improves cache hit 577 * stats. 578 * 579 * We need both locks to modify both pointers, and there 580 * must also not be a write in progress or the uiomove() 581 * in the write might block and temporarily release 582 * its wlock, then reacquire and update windex. We are 583 * only serialized against reads, not writes. 584 * 585 * XXX should we even bother resetting the indices? It 586 * might actually be more cache efficient not to. 587 */ 588 if (rpipe->pipe_buffer.rindex == rpipe->pipe_buffer.windex && 589 rpipe->pipe_wip == 0) { 590 rpipe->pipe_buffer.rindex = 0; 591 rpipe->pipe_buffer.windex = 0; 592 } 593 594 /* 595 * Wait for more data. 596 * 597 * Pipe_state can only be set if both the rlock and wlock 598 * are held. 599 */ 600 rpipe->pipe_state |= PIPE_WANTR; 601 tsleep_interlock(rpipe, PCATCH); 602 lwkt_reltoken(&rpipe->pipe_wlock); 603 error = tsleep(rpipe, PCATCH | PINTERLOCKED, "piperd", 0); 604 ++pipe_rblocked_count; 605 if (error) 606 break; 607 } 608 pipe_end_uio(rpipe, &rpipe->pipe_rip); 609 610 /* 611 * Uptime last access time 612 */ 613 if (error == 0 && nread) 614 vfs_timestamp(&rpipe->pipe_atime); 615 616 /* 617 * If we drained the FIFO more then half way then handle 618 * write blocking hysteresis. 619 * 620 * Note that PIPE_WANTW cannot be set by the writer without 621 * it holding both rlock and wlock, so we can test it 622 * while holding just rlock. 623 */ 624 if (notify_writer) { 625 /* 626 * Synchronous blocking is done on the pipe involved 627 */ 628 if (rpipe->pipe_state & PIPE_WANTW) { 629 lwkt_gettoken(&rpipe->pipe_wlock); 630 if (rpipe->pipe_state & PIPE_WANTW) { 631 rpipe->pipe_state &= ~PIPE_WANTW; 632 lwkt_reltoken(&rpipe->pipe_wlock); 633 wakeup(rpipe); 634 } else { 635 lwkt_reltoken(&rpipe->pipe_wlock); 636 } 637 } 638 639 /* 640 * But we may also have to deal with a kqueue which is 641 * stored on the same pipe as its descriptor, so a 642 * EVFILT_WRITE event waiting for our side to drain will 643 * be on the other side. 644 */ 645 lwkt_gettoken(&wpipe->pipe_wlock); 646 pipewakeup(wpipe, 0); 647 lwkt_reltoken(&wpipe->pipe_wlock); 648 } 649 /*size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;*/ 650 lwkt_reltoken(&rpipe->pipe_rlock); 651 652 return (error); 653 } 654 655 static int 656 pipe_write(struct file *fp, struct uio *uio, struct ucred *cred, int fflags) 657 { 658 int error; 659 int orig_resid; 660 int nbio; 661 struct pipe *wpipe; 662 struct pipe *rpipe; 663 u_int windex; 664 u_int space; 665 u_int wcount; 666 int bigwrite; 667 int bigcount; 668 669 /* 670 * Writes go to the peer. The peer will always exist. 671 */ 672 rpipe = (struct pipe *) fp->f_data; 673 wpipe = rpipe->pipe_peer; 674 lwkt_gettoken(&wpipe->pipe_wlock); 675 if (wpipe->pipe_state & PIPE_WEOF) { 676 lwkt_reltoken(&wpipe->pipe_wlock); 677 return (EPIPE); 678 } 679 680 /* 681 * Degenerate case (EPIPE takes prec) 682 */ 683 if (uio->uio_resid == 0) { 684 lwkt_reltoken(&wpipe->pipe_wlock); 685 return(0); 686 } 687 688 /* 689 * Writes are serialized (start_uio must be called with wlock) 690 */ 691 error = pipe_start_uio(wpipe, &wpipe->pipe_wip); 692 if (error) { 693 lwkt_reltoken(&wpipe->pipe_wlock); 694 return (error); 695 } 696 697 if (fflags & O_FBLOCKING) 698 nbio = 0; 699 else if (fflags & O_FNONBLOCKING) 700 nbio = 1; 701 else if (fp->f_flag & O_NONBLOCK) 702 nbio = 1; 703 else 704 nbio = 0; 705 706 /* 707 * If it is advantageous to resize the pipe buffer, do 708 * so. We are write-serialized so we can block safely. 709 */ 710 if ((wpipe->pipe_buffer.size <= PIPE_SIZE) && 711 (pipe_nbig < pipe_maxbig) && 712 wpipe->pipe_wantwcnt > 4 && 713 (wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex)) { 714 /* 715 * Recheck after lock. 716 */ 717 lwkt_gettoken(&wpipe->pipe_rlock); 718 if ((wpipe->pipe_buffer.size <= PIPE_SIZE) && 719 (pipe_nbig < pipe_maxbig) && 720 (wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex)) { 721 atomic_add_int(&pipe_nbig, 1); 722 if (pipespace(wpipe, BIG_PIPE_SIZE) == 0) 723 ++pipe_bigcount; 724 else 725 atomic_subtract_int(&pipe_nbig, 1); 726 } 727 lwkt_reltoken(&wpipe->pipe_rlock); 728 } 729 730 orig_resid = uio->uio_resid; 731 wcount = 0; 732 733 bigwrite = (uio->uio_resid > 10 * 1024 * 1024); 734 bigcount = 10; 735 736 while (uio->uio_resid) { 737 if (wpipe->pipe_state & PIPE_WEOF) { 738 error = EPIPE; 739 break; 740 } 741 742 /* 743 * Don't hog the cpu. 744 */ 745 if (bigwrite && --bigcount == 0) { 746 lwkt_user_yield(); 747 bigcount = 10; 748 if (CURSIG(curthread->td_lwp)) { 749 error = EINTR; 750 break; 751 } 752 } 753 754 windex = wpipe->pipe_buffer.windex & 755 (wpipe->pipe_buffer.size - 1); 756 space = wpipe->pipe_buffer.size - 757 (wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex); 758 cpu_lfence(); 759 760 /* Writes of size <= PIPE_BUF must be atomic. */ 761 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF)) 762 space = 0; 763 764 /* 765 * Write to fill, read size handles write hysteresis. Also 766 * additional restrictions can cause select-based non-blocking 767 * writes to spin. 768 */ 769 if (space > 0) { 770 u_int segsize; 771 772 /* 773 * Transfer size is minimum of uio transfer 774 * and free space in pipe buffer. 775 * 776 * Limit each uiocopy to no more then PIPE_SIZE 777 * so we can keep the gravy train going on a 778 * SMP box. This doubles the performance for 779 * write sizes > 16K. Otherwise large writes 780 * wind up doing an inefficient synchronous 781 * ping-pong. 782 */ 783 space = szmin(space, uio->uio_resid); 784 if (space > PIPE_SIZE) 785 space = PIPE_SIZE; 786 787 /* 788 * First segment to transfer is minimum of 789 * transfer size and contiguous space in 790 * pipe buffer. If first segment to transfer 791 * is less than the transfer size, we've got 792 * a wraparound in the buffer. 793 */ 794 segsize = wpipe->pipe_buffer.size - windex; 795 if (segsize > space) 796 segsize = space; 797 798 /* 799 * If this is the first loop and the reader is 800 * blocked, do a preemptive wakeup of the reader. 801 * 802 * On SMP the IPI latency plus the wlock interlock 803 * on the reader side is the fastest way to get the 804 * reader going. (The scheduler will hard loop on 805 * lock tokens). 806 * 807 * NOTE: We can't clear WANTR here without acquiring 808 * the rlock, which we don't want to do here! 809 */ 810 if ((wpipe->pipe_state & PIPE_WANTR)) 811 wakeup(wpipe); 812 813 /* 814 * Transfer segment, which may include a wrap-around. 815 * Update windex to account for both all in one go 816 * so the reader can read() the data atomically. 817 */ 818 error = uiomove(&wpipe->pipe_buffer.buffer[windex], 819 segsize, uio); 820 if (error == 0 && segsize < space) { 821 segsize = space - segsize; 822 error = uiomove(&wpipe->pipe_buffer.buffer[0], 823 segsize, uio); 824 } 825 if (error) 826 break; 827 cpu_mfence(); 828 wpipe->pipe_buffer.windex += space; 829 wcount += space; 830 continue; 831 } 832 833 /* 834 * We need both the rlock and the wlock to interlock against 835 * the EOF, WANTW, and size checks, and to modify pipe_state. 836 * 837 * These are token locks so we do not have to worry about 838 * deadlocks. 839 */ 840 lwkt_gettoken(&wpipe->pipe_rlock); 841 842 /* 843 * If the "read-side" has been blocked, wake it up now 844 * and yield to let it drain synchronously rather 845 * then block. 846 */ 847 if (wpipe->pipe_state & PIPE_WANTR) { 848 wpipe->pipe_state &= ~PIPE_WANTR; 849 wakeup(wpipe); 850 } 851 852 /* 853 * don't block on non-blocking I/O 854 */ 855 if (nbio) { 856 lwkt_reltoken(&wpipe->pipe_rlock); 857 error = EAGAIN; 858 break; 859 } 860 861 /* 862 * re-test whether we have to block in the writer after 863 * acquiring both locks, in case the reader opened up 864 * some space. 865 */ 866 space = wpipe->pipe_buffer.size - 867 (wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex); 868 cpu_lfence(); 869 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF)) 870 space = 0; 871 872 /* 873 * Retest EOF - acquiring a new token can temporarily release 874 * tokens already held. 875 */ 876 if (wpipe->pipe_state & PIPE_WEOF) { 877 lwkt_reltoken(&wpipe->pipe_rlock); 878 error = EPIPE; 879 break; 880 } 881 882 /* 883 * We have no more space and have something to offer, 884 * wake up select/poll/kq. 885 */ 886 if (space == 0) { 887 wpipe->pipe_state |= PIPE_WANTW; 888 ++wpipe->pipe_wantwcnt; 889 pipewakeup(wpipe, 1); 890 if (wpipe->pipe_state & PIPE_WANTW) 891 error = tsleep(wpipe, PCATCH, "pipewr", 0); 892 ++pipe_wblocked_count; 893 } 894 lwkt_reltoken(&wpipe->pipe_rlock); 895 896 /* 897 * Break out if we errored or the read side wants us to go 898 * away. 899 */ 900 if (error) 901 break; 902 if (wpipe->pipe_state & PIPE_WEOF) { 903 error = EPIPE; 904 break; 905 } 906 } 907 pipe_end_uio(wpipe, &wpipe->pipe_wip); 908 909 /* 910 * If we have put any characters in the buffer, we wake up 911 * the reader. 912 * 913 * Both rlock and wlock are required to be able to modify pipe_state. 914 */ 915 if (wpipe->pipe_buffer.windex != wpipe->pipe_buffer.rindex) { 916 if (wpipe->pipe_state & PIPE_WANTR) { 917 lwkt_gettoken(&wpipe->pipe_rlock); 918 if (wpipe->pipe_state & PIPE_WANTR) { 919 wpipe->pipe_state &= ~PIPE_WANTR; 920 lwkt_reltoken(&wpipe->pipe_rlock); 921 wakeup(wpipe); 922 } else { 923 lwkt_reltoken(&wpipe->pipe_rlock); 924 } 925 } 926 lwkt_gettoken(&wpipe->pipe_rlock); 927 pipewakeup(wpipe, 1); 928 lwkt_reltoken(&wpipe->pipe_rlock); 929 } 930 931 /* 932 * Don't return EPIPE if I/O was successful 933 */ 934 if ((wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex) && 935 (uio->uio_resid == 0) && 936 (error == EPIPE)) { 937 error = 0; 938 } 939 940 if (error == 0) 941 vfs_timestamp(&wpipe->pipe_mtime); 942 943 /* 944 * We have something to offer, 945 * wake up select/poll/kq. 946 */ 947 /*space = wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex;*/ 948 lwkt_reltoken(&wpipe->pipe_wlock); 949 return (error); 950 } 951 952 /* 953 * we implement a very minimal set of ioctls for compatibility with sockets. 954 */ 955 int 956 pipe_ioctl(struct file *fp, u_long cmd, caddr_t data, 957 struct ucred *cred, struct sysmsg *msg) 958 { 959 struct pipe *mpipe; 960 int error; 961 962 mpipe = (struct pipe *)fp->f_data; 963 964 lwkt_gettoken(&mpipe->pipe_rlock); 965 lwkt_gettoken(&mpipe->pipe_wlock); 966 967 switch (cmd) { 968 case FIOASYNC: 969 if (*(int *)data) { 970 mpipe->pipe_state |= PIPE_ASYNC; 971 } else { 972 mpipe->pipe_state &= ~PIPE_ASYNC; 973 } 974 error = 0; 975 break; 976 case FIONREAD: 977 *(int *)data = mpipe->pipe_buffer.windex - 978 mpipe->pipe_buffer.rindex; 979 error = 0; 980 break; 981 case FIOSETOWN: 982 error = fsetown(*(int *)data, &mpipe->pipe_sigio); 983 break; 984 case FIOGETOWN: 985 *(int *)data = fgetown(&mpipe->pipe_sigio); 986 error = 0; 987 break; 988 case TIOCSPGRP: 989 /* This is deprecated, FIOSETOWN should be used instead. */ 990 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio); 991 break; 992 993 case TIOCGPGRP: 994 /* This is deprecated, FIOGETOWN should be used instead. */ 995 *(int *)data = -fgetown(&mpipe->pipe_sigio); 996 error = 0; 997 break; 998 default: 999 error = ENOTTY; 1000 break; 1001 } 1002 lwkt_reltoken(&mpipe->pipe_wlock); 1003 lwkt_reltoken(&mpipe->pipe_rlock); 1004 1005 return (error); 1006 } 1007 1008 /* 1009 * MPSAFE 1010 */ 1011 static int 1012 pipe_stat(struct file *fp, struct stat *ub, struct ucred *cred) 1013 { 1014 struct pipe *pipe; 1015 1016 pipe = (struct pipe *)fp->f_data; 1017 1018 bzero((caddr_t)ub, sizeof(*ub)); 1019 ub->st_mode = S_IFIFO; 1020 ub->st_blksize = pipe->pipe_buffer.size; 1021 ub->st_size = pipe->pipe_buffer.windex - pipe->pipe_buffer.rindex; 1022 ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize; 1023 ub->st_atimespec = pipe->pipe_atime; 1024 ub->st_mtimespec = pipe->pipe_mtime; 1025 ub->st_ctimespec = pipe->pipe_ctime; 1026 /* 1027 * Left as 0: st_dev, st_ino, st_nlink, st_uid, st_gid, st_rdev, 1028 * st_flags, st_gen. 1029 * XXX (st_dev, st_ino) should be unique. 1030 */ 1031 return (0); 1032 } 1033 1034 static int 1035 pipe_close(struct file *fp) 1036 { 1037 struct pipe *cpipe; 1038 1039 cpipe = (struct pipe *)fp->f_data; 1040 fp->f_ops = &badfileops; 1041 fp->f_data = NULL; 1042 funsetown(&cpipe->pipe_sigio); 1043 pipeclose(cpipe); 1044 return (0); 1045 } 1046 1047 /* 1048 * Shutdown one or both directions of a full-duplex pipe. 1049 */ 1050 static int 1051 pipe_shutdown(struct file *fp, int how) 1052 { 1053 struct pipe *rpipe; 1054 struct pipe *wpipe; 1055 int error = EPIPE; 1056 1057 rpipe = (struct pipe *)fp->f_data; 1058 wpipe = rpipe->pipe_peer; 1059 1060 /* 1061 * We modify pipe_state on both pipes, which means we need 1062 * all four tokens! 1063 */ 1064 lwkt_gettoken(&rpipe->pipe_rlock); 1065 lwkt_gettoken(&rpipe->pipe_wlock); 1066 lwkt_gettoken(&wpipe->pipe_rlock); 1067 lwkt_gettoken(&wpipe->pipe_wlock); 1068 1069 switch(how) { 1070 case SHUT_RDWR: 1071 case SHUT_RD: 1072 rpipe->pipe_state |= PIPE_REOF; /* my reads */ 1073 rpipe->pipe_state |= PIPE_WEOF; /* peer writes */ 1074 if (rpipe->pipe_state & PIPE_WANTR) { 1075 rpipe->pipe_state &= ~PIPE_WANTR; 1076 wakeup(rpipe); 1077 } 1078 if (rpipe->pipe_state & PIPE_WANTW) { 1079 rpipe->pipe_state &= ~PIPE_WANTW; 1080 wakeup(rpipe); 1081 } 1082 error = 0; 1083 if (how == SHUT_RD) 1084 break; 1085 /* fall through */ 1086 case SHUT_WR: 1087 wpipe->pipe_state |= PIPE_REOF; /* peer reads */ 1088 wpipe->pipe_state |= PIPE_WEOF; /* my writes */ 1089 if (wpipe->pipe_state & PIPE_WANTR) { 1090 wpipe->pipe_state &= ~PIPE_WANTR; 1091 wakeup(wpipe); 1092 } 1093 if (wpipe->pipe_state & PIPE_WANTW) { 1094 wpipe->pipe_state &= ~PIPE_WANTW; 1095 wakeup(wpipe); 1096 } 1097 error = 0; 1098 break; 1099 } 1100 pipewakeup(rpipe, 1); 1101 pipewakeup(wpipe, 1); 1102 1103 lwkt_reltoken(&wpipe->pipe_wlock); 1104 lwkt_reltoken(&wpipe->pipe_rlock); 1105 lwkt_reltoken(&rpipe->pipe_wlock); 1106 lwkt_reltoken(&rpipe->pipe_rlock); 1107 1108 return (error); 1109 } 1110 1111 static void 1112 pipe_free_kmem(struct pipe *cpipe) 1113 { 1114 if (cpipe->pipe_buffer.buffer != NULL) { 1115 if (cpipe->pipe_buffer.size > PIPE_SIZE) 1116 atomic_subtract_int(&pipe_nbig, 1); 1117 kmem_free(&kernel_map, 1118 (vm_offset_t)cpipe->pipe_buffer.buffer, 1119 cpipe->pipe_buffer.size); 1120 cpipe->pipe_buffer.buffer = NULL; 1121 cpipe->pipe_buffer.object = NULL; 1122 } 1123 } 1124 1125 /* 1126 * Close the pipe. The slock must be held to interlock against simultanious 1127 * closes. The rlock and wlock must be held to adjust the pipe_state. 1128 */ 1129 static void 1130 pipeclose(struct pipe *cpipe) 1131 { 1132 globaldata_t gd; 1133 struct pipe *ppipe; 1134 1135 if (cpipe == NULL) 1136 return; 1137 1138 /* 1139 * The slock may not have been allocated yet (close during 1140 * initialization) 1141 * 1142 * We need both the read and write tokens to modify pipe_state. 1143 */ 1144 if (cpipe->pipe_slock) 1145 lockmgr(cpipe->pipe_slock, LK_EXCLUSIVE); 1146 lwkt_gettoken(&cpipe->pipe_rlock); 1147 lwkt_gettoken(&cpipe->pipe_wlock); 1148 1149 /* 1150 * Set our state, wakeup anyone waiting in select/poll/kq, and 1151 * wakeup anyone blocked on our pipe. 1152 */ 1153 cpipe->pipe_state |= PIPE_CLOSED | PIPE_REOF | PIPE_WEOF; 1154 pipewakeup(cpipe, 1); 1155 if (cpipe->pipe_state & (PIPE_WANTR | PIPE_WANTW)) { 1156 cpipe->pipe_state &= ~(PIPE_WANTR | PIPE_WANTW); 1157 wakeup(cpipe); 1158 } 1159 1160 /* 1161 * Disconnect from peer. 1162 */ 1163 if ((ppipe = cpipe->pipe_peer) != NULL) { 1164 lwkt_gettoken(&ppipe->pipe_rlock); 1165 lwkt_gettoken(&ppipe->pipe_wlock); 1166 ppipe->pipe_state |= PIPE_REOF | PIPE_WEOF; 1167 pipewakeup(ppipe, 1); 1168 if (ppipe->pipe_state & (PIPE_WANTR | PIPE_WANTW)) { 1169 ppipe->pipe_state &= ~(PIPE_WANTR | PIPE_WANTW); 1170 wakeup(ppipe); 1171 } 1172 if (SLIST_FIRST(&ppipe->pipe_kq.ki_note)) 1173 KNOTE(&ppipe->pipe_kq.ki_note, 0); 1174 lwkt_reltoken(&ppipe->pipe_wlock); 1175 lwkt_reltoken(&ppipe->pipe_rlock); 1176 } 1177 1178 /* 1179 * If the peer is also closed we can free resources for both 1180 * sides, otherwise we leave our side intact to deal with any 1181 * races (since we only have the slock). 1182 */ 1183 if (ppipe && (ppipe->pipe_state & PIPE_CLOSED)) { 1184 cpipe->pipe_peer = NULL; 1185 ppipe->pipe_peer = NULL; 1186 ppipe->pipe_slock = NULL; /* we will free the slock */ 1187 pipeclose(ppipe); 1188 ppipe = NULL; 1189 } 1190 1191 lwkt_reltoken(&cpipe->pipe_wlock); 1192 lwkt_reltoken(&cpipe->pipe_rlock); 1193 if (cpipe->pipe_slock) 1194 lockmgr(cpipe->pipe_slock, LK_RELEASE); 1195 1196 /* 1197 * If we disassociated from our peer we can free resources 1198 */ 1199 if (ppipe == NULL) { 1200 gd = mycpu; 1201 if (cpipe->pipe_slock) { 1202 kfree(cpipe->pipe_slock, M_PIPE); 1203 cpipe->pipe_slock = NULL; 1204 } 1205 if (gd->gd_pipeqcount >= pipe_maxcache || 1206 cpipe->pipe_buffer.size != PIPE_SIZE 1207 ) { 1208 pipe_free_kmem(cpipe); 1209 kfree(cpipe, M_PIPE); 1210 } else { 1211 cpipe->pipe_state = 0; 1212 cpipe->pipe_peer = gd->gd_pipeq; 1213 gd->gd_pipeq = cpipe; 1214 ++gd->gd_pipeqcount; 1215 } 1216 } 1217 } 1218 1219 static int 1220 pipe_kqfilter(struct file *fp, struct knote *kn) 1221 { 1222 struct pipe *cpipe; 1223 1224 cpipe = (struct pipe *)kn->kn_fp->f_data; 1225 1226 switch (kn->kn_filter) { 1227 case EVFILT_READ: 1228 kn->kn_fop = &pipe_rfiltops; 1229 break; 1230 case EVFILT_WRITE: 1231 kn->kn_fop = &pipe_wfiltops; 1232 if (cpipe->pipe_peer == NULL) { 1233 /* other end of pipe has been closed */ 1234 return (EPIPE); 1235 } 1236 break; 1237 default: 1238 return (EOPNOTSUPP); 1239 } 1240 kn->kn_hook = (caddr_t)cpipe; 1241 1242 knote_insert(&cpipe->pipe_kq.ki_note, kn); 1243 1244 return (0); 1245 } 1246 1247 static void 1248 filt_pipedetach(struct knote *kn) 1249 { 1250 struct pipe *cpipe = (struct pipe *)kn->kn_hook; 1251 1252 knote_remove(&cpipe->pipe_kq.ki_note, kn); 1253 } 1254 1255 /*ARGSUSED*/ 1256 static int 1257 filt_piperead(struct knote *kn, long hint) 1258 { 1259 struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data; 1260 int ready = 0; 1261 1262 lwkt_gettoken(&rpipe->pipe_rlock); 1263 lwkt_gettoken(&rpipe->pipe_wlock); 1264 1265 kn->kn_data = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex; 1266 1267 if (rpipe->pipe_state & PIPE_REOF) { 1268 /* 1269 * Only set NODATA if all data has been exhausted 1270 */ 1271 if (kn->kn_data == 0) 1272 kn->kn_flags |= EV_NODATA; 1273 kn->kn_flags |= EV_EOF; 1274 ready = 1; 1275 } 1276 1277 lwkt_reltoken(&rpipe->pipe_wlock); 1278 lwkt_reltoken(&rpipe->pipe_rlock); 1279 1280 if (!ready) 1281 ready = kn->kn_data > 0; 1282 1283 return (ready); 1284 } 1285 1286 /*ARGSUSED*/ 1287 static int 1288 filt_pipewrite(struct knote *kn, long hint) 1289 { 1290 struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data; 1291 struct pipe *wpipe = rpipe->pipe_peer; 1292 int ready = 0; 1293 1294 kn->kn_data = 0; 1295 if (wpipe == NULL) { 1296 kn->kn_flags |= (EV_EOF | EV_NODATA); 1297 return (1); 1298 } 1299 1300 lwkt_gettoken(&wpipe->pipe_rlock); 1301 lwkt_gettoken(&wpipe->pipe_wlock); 1302 1303 if (wpipe->pipe_state & PIPE_WEOF) { 1304 kn->kn_flags |= (EV_EOF | EV_NODATA); 1305 ready = 1; 1306 } 1307 1308 if (!ready) 1309 kn->kn_data = wpipe->pipe_buffer.size - 1310 (wpipe->pipe_buffer.windex - 1311 wpipe->pipe_buffer.rindex); 1312 1313 lwkt_reltoken(&wpipe->pipe_wlock); 1314 lwkt_reltoken(&wpipe->pipe_rlock); 1315 1316 if (!ready) 1317 ready = kn->kn_data >= PIPE_BUF; 1318 1319 return (ready); 1320 } 1321