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/select.h> 40 #include <sys/signalvar.h> 41 #include <sys/sysproto.h> 42 #include <sys/pipe.h> 43 #include <sys/vnode.h> 44 #include <sys/uio.h> 45 #include <sys/event.h> 46 #include <sys/globaldata.h> 47 #include <sys/module.h> 48 #include <sys/malloc.h> 49 #include <sys/sysctl.h> 50 #include <sys/socket.h> 51 52 #include <vm/vm.h> 53 #include <vm/vm_param.h> 54 #include <sys/lock.h> 55 #include <vm/vm_object.h> 56 #include <vm/vm_kern.h> 57 #include <vm/vm_extern.h> 58 #include <vm/pmap.h> 59 #include <vm/vm_map.h> 60 #include <vm/vm_page.h> 61 #include <vm/vm_zone.h> 62 63 #include <sys/file2.h> 64 #include <sys/signal2.h> 65 #include <sys/mplock2.h> 66 67 #include <machine/cpufunc.h> 68 69 /* 70 * interfaces to the outside world 71 */ 72 static int pipe_read (struct file *fp, struct uio *uio, 73 struct ucred *cred, int flags); 74 static int pipe_write (struct file *fp, struct uio *uio, 75 struct ucred *cred, int flags); 76 static int pipe_close (struct file *fp); 77 static int pipe_shutdown (struct file *fp, int how); 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, 81 struct ucred *cred, struct sysmsg *msg); 82 83 static struct fileops pipeops = { 84 .fo_read = pipe_read, 85 .fo_write = pipe_write, 86 .fo_ioctl = pipe_ioctl, 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 = 1; 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 int 163 pipeseltest(struct pipe *cpipe) 164 { 165 return ((cpipe->pipe_state & PIPE_SEL) || 166 ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio) || 167 SLIST_FIRST(&cpipe->pipe_sel.si_note)); 168 } 169 170 static __inline void 171 pipeselwakeup(struct pipe *cpipe) 172 { 173 if (cpipe->pipe_state & PIPE_SEL) { 174 get_mplock(); 175 cpipe->pipe_state &= ~PIPE_SEL; 176 selwakeup(&cpipe->pipe_sel); 177 rel_mplock(); 178 } 179 if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio) { 180 get_mplock(); 181 pgsigio(cpipe->pipe_sigio, SIGIO, 0); 182 rel_mplock(); 183 } 184 if (SLIST_FIRST(&cpipe->pipe_sel.si_note)) { 185 get_mplock(); 186 KNOTE(&cpipe->pipe_sel.si_note, 0); 187 rel_mplock(); 188 } 189 } 190 191 /* 192 * These routines are called before and after a UIO. The UIO 193 * may block, causing our held tokens to be lost temporarily. 194 * 195 * We use these routines to serialize reads against other reads 196 * and writes against other writes. 197 * 198 * The read token is held on entry so *ipp does not race. 199 */ 200 static __inline int 201 pipe_start_uio(struct pipe *cpipe, int *ipp) 202 { 203 int error; 204 205 while (*ipp) { 206 *ipp = -1; 207 error = tsleep(ipp, PCATCH, "pipexx", 0); 208 if (error) 209 return (error); 210 } 211 *ipp = 1; 212 return (0); 213 } 214 215 static __inline void 216 pipe_end_uio(struct pipe *cpipe, int *ipp) 217 { 218 if (*ipp < 0) { 219 *ipp = 0; 220 wakeup(ipp); 221 } else { 222 KKASSERT(*ipp > 0); 223 *ipp = 0; 224 } 225 } 226 227 static __inline void 228 pipe_get_mplock(int *save) 229 { 230 #ifdef SMP 231 if (pipe_mpsafe == 0) { 232 get_mplock(); 233 *save = 1; 234 } else 235 #endif 236 { 237 *save = 0; 238 } 239 } 240 241 static __inline void 242 pipe_rel_mplock(int *save) 243 { 244 #ifdef SMP 245 if (*save) 246 rel_mplock(); 247 #endif 248 } 249 250 251 /* 252 * The pipe system call for the DTYPE_PIPE type of pipes 253 * 254 * pipe_args(int dummy) 255 * 256 * MPSAFE 257 */ 258 int 259 sys_pipe(struct pipe_args *uap) 260 { 261 struct thread *td = curthread; 262 struct filedesc *fdp = td->td_proc->p_fd; 263 struct file *rf, *wf; 264 struct pipe *rpipe, *wpipe; 265 int fd1, fd2, error; 266 267 rpipe = wpipe = NULL; 268 if (pipe_create(&rpipe) || pipe_create(&wpipe)) { 269 pipeclose(rpipe); 270 pipeclose(wpipe); 271 return (ENFILE); 272 } 273 274 error = falloc(td->td_lwp, &rf, &fd1); 275 if (error) { 276 pipeclose(rpipe); 277 pipeclose(wpipe); 278 return (error); 279 } 280 uap->sysmsg_fds[0] = fd1; 281 282 /* 283 * Warning: once we've gotten past allocation of the fd for the 284 * read-side, we can only drop the read side via fdrop() in order 285 * to avoid races against processes which manage to dup() the read 286 * side while we are blocked trying to allocate the write side. 287 */ 288 rf->f_type = DTYPE_PIPE; 289 rf->f_flag = FREAD | FWRITE; 290 rf->f_ops = &pipeops; 291 rf->f_data = rpipe; 292 error = falloc(td->td_lwp, &wf, &fd2); 293 if (error) { 294 fsetfd(fdp, NULL, fd1); 295 fdrop(rf); 296 /* rpipe has been closed by fdrop(). */ 297 pipeclose(wpipe); 298 return (error); 299 } 300 wf->f_type = DTYPE_PIPE; 301 wf->f_flag = FREAD | FWRITE; 302 wf->f_ops = &pipeops; 303 wf->f_data = wpipe; 304 uap->sysmsg_fds[1] = fd2; 305 306 rpipe->pipe_slock = kmalloc(sizeof(struct lock), 307 M_PIPE, M_WAITOK|M_ZERO); 308 wpipe->pipe_slock = rpipe->pipe_slock; 309 rpipe->pipe_peer = wpipe; 310 wpipe->pipe_peer = rpipe; 311 lockinit(rpipe->pipe_slock, "pipecl", 0, 0); 312 313 /* 314 * Once activated the peer relationship remains valid until 315 * both sides are closed. 316 */ 317 fsetfd(fdp, rf, fd1); 318 fsetfd(fdp, wf, fd2); 319 fdrop(rf); 320 fdrop(wf); 321 322 return (0); 323 } 324 325 /* 326 * Allocate kva for pipe circular buffer, the space is pageable 327 * This routine will 'realloc' the size of a pipe safely, if it fails 328 * it will retain the old buffer. 329 * If it fails it will return ENOMEM. 330 */ 331 static int 332 pipespace(struct pipe *cpipe, int size) 333 { 334 struct vm_object *object; 335 caddr_t buffer; 336 int npages, error; 337 338 npages = round_page(size) / PAGE_SIZE; 339 object = cpipe->pipe_buffer.object; 340 341 /* 342 * [re]create the object if necessary and reserve space for it 343 * in the kernel_map. The object and memory are pageable. On 344 * success, free the old resources before assigning the new 345 * ones. 346 */ 347 if (object == NULL || object->size != npages) { 348 get_mplock(); 349 object = vm_object_allocate(OBJT_DEFAULT, npages); 350 buffer = (caddr_t)vm_map_min(&kernel_map); 351 352 error = vm_map_find(&kernel_map, object, 0, 353 (vm_offset_t *)&buffer, 354 size, PAGE_SIZE, 355 1, VM_MAPTYPE_NORMAL, 356 VM_PROT_ALL, VM_PROT_ALL, 357 0); 358 359 if (error != KERN_SUCCESS) { 360 vm_object_deallocate(object); 361 rel_mplock(); 362 return (ENOMEM); 363 } 364 pipe_free_kmem(cpipe); 365 rel_mplock(); 366 cpipe->pipe_buffer.object = object; 367 cpipe->pipe_buffer.buffer = buffer; 368 cpipe->pipe_buffer.size = size; 369 ++pipe_bkmem_alloc; 370 } else { 371 ++pipe_bcache_alloc; 372 } 373 cpipe->pipe_buffer.rindex = 0; 374 cpipe->pipe_buffer.windex = 0; 375 return (0); 376 } 377 378 /* 379 * Initialize and allocate VM and memory for pipe, pulling the pipe from 380 * our per-cpu cache if possible. For now make sure it is sized for the 381 * smaller PIPE_SIZE default. 382 */ 383 static int 384 pipe_create(struct pipe **cpipep) 385 { 386 globaldata_t gd = mycpu; 387 struct pipe *cpipe; 388 int error; 389 390 if ((cpipe = gd->gd_pipeq) != NULL) { 391 gd->gd_pipeq = cpipe->pipe_peer; 392 --gd->gd_pipeqcount; 393 cpipe->pipe_peer = NULL; 394 cpipe->pipe_wantwcnt = 0; 395 } else { 396 cpipe = kmalloc(sizeof(struct pipe), M_PIPE, M_WAITOK|M_ZERO); 397 } 398 *cpipep = cpipe; 399 if ((error = pipespace(cpipe, PIPE_SIZE)) != 0) 400 return (error); 401 vfs_timestamp(&cpipe->pipe_ctime); 402 cpipe->pipe_atime = cpipe->pipe_ctime; 403 cpipe->pipe_mtime = cpipe->pipe_ctime; 404 lwkt_token_init(&cpipe->pipe_rlock, 1); 405 lwkt_token_init(&cpipe->pipe_wlock, 1); 406 return (0); 407 } 408 409 /* 410 * MPALMOSTSAFE (acquires mplock) 411 */ 412 static int 413 pipe_read(struct file *fp, struct uio *uio, struct ucred *cred, int fflags) 414 { 415 struct pipe *rpipe; 416 int error; 417 size_t nread = 0; 418 int nbio; 419 u_int size; /* total bytes available */ 420 u_int nsize; /* total bytes to read */ 421 u_int rindex; /* contiguous bytes available */ 422 int notify_writer; 423 int mpsave; 424 int bigread; 425 int bigcount; 426 427 if (uio->uio_resid == 0) 428 return(0); 429 430 /* 431 * Setup locks, calculate nbio 432 */ 433 pipe_get_mplock(&mpsave); 434 rpipe = (struct pipe *)fp->f_data; 435 lwkt_gettoken(&rpipe->pipe_rlock); 436 437 if (fflags & O_FBLOCKING) 438 nbio = 0; 439 else if (fflags & O_FNONBLOCKING) 440 nbio = 1; 441 else if (fp->f_flag & O_NONBLOCK) 442 nbio = 1; 443 else 444 nbio = 0; 445 446 /* 447 * Reads are serialized. Note howeverthat pipe_buffer.buffer and 448 * pipe_buffer.size can change out from under us when the number 449 * of bytes in the buffer are zero due to the write-side doing a 450 * pipespace(). 451 */ 452 error = pipe_start_uio(rpipe, &rpipe->pipe_rip); 453 if (error) { 454 pipe_rel_mplock(&mpsave); 455 lwkt_reltoken(&rpipe->pipe_rlock); 456 return (error); 457 } 458 notify_writer = 0; 459 460 bigread = (uio->uio_resid > 10 * 1024 * 1024); 461 bigcount = 10; 462 463 while (uio->uio_resid) { 464 /* 465 * Don't hog the cpu. 466 */ 467 if (bigread && --bigcount == 0) { 468 lwkt_user_yield(); 469 bigcount = 10; 470 if (CURSIG(curthread->td_lwp)) { 471 error = EINTR; 472 break; 473 } 474 } 475 476 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex; 477 cpu_lfence(); 478 if (size) { 479 rindex = rpipe->pipe_buffer.rindex & 480 (rpipe->pipe_buffer.size - 1); 481 nsize = size; 482 if (nsize > rpipe->pipe_buffer.size - rindex) 483 nsize = rpipe->pipe_buffer.size - rindex; 484 nsize = szmin(nsize, uio->uio_resid); 485 486 error = uiomove(&rpipe->pipe_buffer.buffer[rindex], 487 nsize, uio); 488 if (error) 489 break; 490 cpu_mfence(); 491 rpipe->pipe_buffer.rindex += nsize; 492 nread += nsize; 493 494 /* 495 * If the FIFO is still over half full just continue 496 * and do not try to notify the writer yet. 497 */ 498 if (size - nsize >= (rpipe->pipe_buffer.size >> 1)) { 499 notify_writer = 0; 500 continue; 501 } 502 503 /* 504 * When the FIFO is less then half full notify any 505 * waiting writer. WANTW can be checked while 506 * holding just the rlock. 507 */ 508 notify_writer = 1; 509 if ((rpipe->pipe_state & PIPE_WANTW) == 0) 510 continue; 511 } 512 513 /* 514 * If the "write-side" was blocked we wake it up. This code 515 * is reached either when the buffer is completely emptied 516 * or if it becomes more then half-empty. 517 * 518 * Pipe_state can only be modified if both the rlock and 519 * wlock are held. 520 */ 521 if (rpipe->pipe_state & PIPE_WANTW) { 522 lwkt_gettoken(&rpipe->pipe_wlock); 523 if (rpipe->pipe_state & PIPE_WANTW) { 524 notify_writer = 0; 525 rpipe->pipe_state &= ~PIPE_WANTW; 526 lwkt_reltoken(&rpipe->pipe_wlock); 527 wakeup(rpipe); 528 } else { 529 lwkt_reltoken(&rpipe->pipe_wlock); 530 } 531 } 532 533 /* 534 * Pick up our copy loop again if the writer sent data to 535 * us while we were messing around. 536 * 537 * On a SMP box poll up to pipe_delay nanoseconds for new 538 * data. Typically a value of 2000 to 4000 is sufficient 539 * to eradicate most IPIs/tsleeps/wakeups when a pipe 540 * is used for synchronous communications with small packets, 541 * and 8000 or so (8uS) will pipeline large buffer xfers 542 * between cpus over a pipe. 543 * 544 * For synchronous communications a hit means doing a 545 * full Awrite-Bread-Bwrite-Aread cycle in less then 2uS, 546 * where as miss requiring a tsleep/wakeup sequence 547 * will take 7uS or more. 548 */ 549 if (rpipe->pipe_buffer.windex != rpipe->pipe_buffer.rindex) 550 continue; 551 552 #if defined(SMP) && defined(_RDTSC_SUPPORTED_) 553 if (pipe_delay) { 554 int64_t tsc_target; 555 int good = 0; 556 557 tsc_target = tsc_get_target(pipe_delay); 558 while (tsc_test_target(tsc_target) == 0) { 559 if (rpipe->pipe_buffer.windex != 560 rpipe->pipe_buffer.rindex) { 561 good = 1; 562 break; 563 } 564 } 565 if (good) 566 continue; 567 } 568 #endif 569 570 /* 571 * Detect EOF condition, do not set error. 572 */ 573 if (rpipe->pipe_state & PIPE_REOF) 574 break; 575 576 /* 577 * Break if some data was read, or if this was a non-blocking 578 * read. 579 */ 580 if (nread > 0) 581 break; 582 583 if (nbio) { 584 error = EAGAIN; 585 break; 586 } 587 588 /* 589 * Last chance, interlock with WANTR. 590 */ 591 lwkt_gettoken(&rpipe->pipe_wlock); 592 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex; 593 if (size) { 594 lwkt_reltoken(&rpipe->pipe_wlock); 595 continue; 596 } 597 598 /* 599 * Retest EOF - acquiring a new token can temporarily release 600 * tokens already held. 601 */ 602 if (rpipe->pipe_state & PIPE_REOF) { 603 lwkt_reltoken(&rpipe->pipe_wlock); 604 break; 605 } 606 607 /* 608 * If there is no more to read in the pipe, reset its 609 * pointers to the beginning. This improves cache hit 610 * stats. 611 * 612 * We need both locks to modify both pointers, and there 613 * must also not be a write in progress or the uiomove() 614 * in the write might block and temporarily release 615 * its wlock, then reacquire and update windex. We are 616 * only serialized against reads, not writes. 617 * 618 * XXX should we even bother resetting the indices? It 619 * might actually be more cache efficient not to. 620 */ 621 if (rpipe->pipe_buffer.rindex == rpipe->pipe_buffer.windex && 622 rpipe->pipe_wip == 0) { 623 rpipe->pipe_buffer.rindex = 0; 624 rpipe->pipe_buffer.windex = 0; 625 } 626 627 /* 628 * Wait for more data. 629 * 630 * Pipe_state can only be set if both the rlock and wlock 631 * are held. 632 */ 633 rpipe->pipe_state |= PIPE_WANTR; 634 tsleep_interlock(rpipe, PCATCH); 635 lwkt_reltoken(&rpipe->pipe_wlock); 636 error = tsleep(rpipe, PCATCH | PINTERLOCKED, "piperd", 0); 637 ++pipe_rblocked_count; 638 if (error) 639 break; 640 } 641 pipe_end_uio(rpipe, &rpipe->pipe_rip); 642 643 /* 644 * Uptime last access time 645 */ 646 if (error == 0 && nread) 647 vfs_timestamp(&rpipe->pipe_atime); 648 649 /* 650 * If we drained the FIFO more then half way then handle 651 * write blocking hysteresis. 652 * 653 * Note that PIPE_WANTW cannot be set by the writer without 654 * it holding both rlock and wlock, so we can test it 655 * while holding just rlock. 656 */ 657 if (notify_writer) { 658 if (rpipe->pipe_state & PIPE_WANTW) { 659 lwkt_gettoken(&rpipe->pipe_wlock); 660 if (rpipe->pipe_state & PIPE_WANTW) { 661 rpipe->pipe_state &= ~PIPE_WANTW; 662 lwkt_reltoken(&rpipe->pipe_wlock); 663 wakeup(rpipe); 664 } else { 665 lwkt_reltoken(&rpipe->pipe_wlock); 666 } 667 } 668 if (pipeseltest(rpipe)) { 669 lwkt_gettoken(&rpipe->pipe_wlock); 670 pipeselwakeup(rpipe); 671 lwkt_reltoken(&rpipe->pipe_wlock); 672 } 673 } 674 /*size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;*/ 675 lwkt_reltoken(&rpipe->pipe_rlock); 676 677 pipe_rel_mplock(&mpsave); 678 return (error); 679 } 680 681 /* 682 * MPALMOSTSAFE - acquires mplock 683 */ 684 static int 685 pipe_write(struct file *fp, struct uio *uio, struct ucred *cred, int fflags) 686 { 687 int error; 688 int orig_resid; 689 int nbio; 690 struct pipe *wpipe, *rpipe; 691 u_int windex; 692 u_int space; 693 u_int wcount; 694 int mpsave; 695 int bigwrite; 696 int bigcount; 697 698 pipe_get_mplock(&mpsave); 699 700 /* 701 * Writes go to the peer. The peer will always exist. 702 */ 703 rpipe = (struct pipe *) fp->f_data; 704 wpipe = rpipe->pipe_peer; 705 lwkt_gettoken(&wpipe->pipe_wlock); 706 if (wpipe->pipe_state & PIPE_WEOF) { 707 pipe_rel_mplock(&mpsave); 708 lwkt_reltoken(&wpipe->pipe_wlock); 709 return (EPIPE); 710 } 711 712 /* 713 * Degenerate case (EPIPE takes prec) 714 */ 715 if (uio->uio_resid == 0) { 716 pipe_rel_mplock(&mpsave); 717 lwkt_reltoken(&wpipe->pipe_wlock); 718 return(0); 719 } 720 721 /* 722 * Writes are serialized (start_uio must be called with wlock) 723 */ 724 error = pipe_start_uio(wpipe, &wpipe->pipe_wip); 725 if (error) { 726 pipe_rel_mplock(&mpsave); 727 lwkt_reltoken(&wpipe->pipe_wlock); 728 return (error); 729 } 730 731 if (fflags & O_FBLOCKING) 732 nbio = 0; 733 else if (fflags & O_FNONBLOCKING) 734 nbio = 1; 735 else if (fp->f_flag & O_NONBLOCK) 736 nbio = 1; 737 else 738 nbio = 0; 739 740 /* 741 * If it is advantageous to resize the pipe buffer, do 742 * so. We are write-serialized so we can block safely. 743 */ 744 if ((wpipe->pipe_buffer.size <= PIPE_SIZE) && 745 (pipe_nbig < pipe_maxbig) && 746 wpipe->pipe_wantwcnt > 4 && 747 (wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex)) { 748 /* 749 * Recheck after lock. 750 */ 751 lwkt_gettoken(&wpipe->pipe_rlock); 752 if ((wpipe->pipe_buffer.size <= PIPE_SIZE) && 753 (pipe_nbig < pipe_maxbig) && 754 (wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex)) { 755 atomic_add_int(&pipe_nbig, 1); 756 if (pipespace(wpipe, BIG_PIPE_SIZE) == 0) 757 ++pipe_bigcount; 758 else 759 atomic_subtract_int(&pipe_nbig, 1); 760 } 761 lwkt_reltoken(&wpipe->pipe_rlock); 762 } 763 764 orig_resid = uio->uio_resid; 765 wcount = 0; 766 767 bigwrite = (uio->uio_resid > 10 * 1024 * 1024); 768 bigcount = 10; 769 770 while (uio->uio_resid) { 771 if (wpipe->pipe_state & PIPE_WEOF) { 772 error = EPIPE; 773 break; 774 } 775 776 /* 777 * Don't hog the cpu. 778 */ 779 if (bigwrite && --bigcount == 0) { 780 lwkt_user_yield(); 781 bigcount = 10; 782 if (CURSIG(curthread->td_lwp)) { 783 error = EINTR; 784 break; 785 } 786 } 787 788 windex = wpipe->pipe_buffer.windex & 789 (wpipe->pipe_buffer.size - 1); 790 space = wpipe->pipe_buffer.size - 791 (wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex); 792 cpu_lfence(); 793 794 /* Writes of size <= PIPE_BUF must be atomic. */ 795 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF)) 796 space = 0; 797 798 /* 799 * Write to fill, read size handles write hysteresis. Also 800 * additional restrictions can cause select-based non-blocking 801 * writes to spin. 802 */ 803 if (space > 0) { 804 u_int segsize; 805 806 /* 807 * Transfer size is minimum of uio transfer 808 * and free space in pipe buffer. 809 * 810 * Limit each uiocopy to no more then PIPE_SIZE 811 * so we can keep the gravy train going on a 812 * SMP box. This doubles the performance for 813 * write sizes > 16K. Otherwise large writes 814 * wind up doing an inefficient synchronous 815 * ping-pong. 816 */ 817 space = szmin(space, uio->uio_resid); 818 if (space > PIPE_SIZE) 819 space = PIPE_SIZE; 820 821 /* 822 * First segment to transfer is minimum of 823 * transfer size and contiguous space in 824 * pipe buffer. If first segment to transfer 825 * is less than the transfer size, we've got 826 * a wraparound in the buffer. 827 */ 828 segsize = wpipe->pipe_buffer.size - windex; 829 if (segsize > space) 830 segsize = space; 831 832 #ifdef SMP 833 /* 834 * If this is the first loop and the reader is 835 * blocked, do a preemptive wakeup of the reader. 836 * 837 * On SMP the IPI latency plus the wlock interlock 838 * on the reader side is the fastest way to get the 839 * reader going. (The scheduler will hard loop on 840 * lock tokens). 841 * 842 * NOTE: We can't clear WANTR here without acquiring 843 * the rlock, which we don't want to do here! 844 */ 845 if ((wpipe->pipe_state & PIPE_WANTR) && pipe_mpsafe > 1) 846 wakeup(wpipe); 847 #endif 848 849 /* 850 * Transfer segment, which may include a wrap-around. 851 * Update windex to account for both all in one go 852 * so the reader can read() the data atomically. 853 */ 854 error = uiomove(&wpipe->pipe_buffer.buffer[windex], 855 segsize, uio); 856 if (error == 0 && segsize < space) { 857 segsize = space - segsize; 858 error = uiomove(&wpipe->pipe_buffer.buffer[0], 859 segsize, uio); 860 } 861 if (error) 862 break; 863 cpu_mfence(); 864 wpipe->pipe_buffer.windex += space; 865 wcount += space; 866 continue; 867 } 868 869 /* 870 * We need both the rlock and the wlock to interlock against 871 * the EOF, WANTW, and size checks, and to modify pipe_state. 872 * 873 * These are token locks so we do not have to worry about 874 * deadlocks. 875 */ 876 lwkt_gettoken(&wpipe->pipe_rlock); 877 878 /* 879 * If the "read-side" has been blocked, wake it up now 880 * and yield to let it drain synchronously rather 881 * then block. 882 */ 883 if (wpipe->pipe_state & PIPE_WANTR) { 884 wpipe->pipe_state &= ~PIPE_WANTR; 885 wakeup(wpipe); 886 } 887 888 /* 889 * don't block on non-blocking I/O 890 */ 891 if (nbio) { 892 lwkt_reltoken(&wpipe->pipe_rlock); 893 error = EAGAIN; 894 break; 895 } 896 897 /* 898 * re-test whether we have to block in the writer after 899 * acquiring both locks, in case the reader opened up 900 * some space. 901 */ 902 space = wpipe->pipe_buffer.size - 903 (wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex); 904 cpu_lfence(); 905 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF)) 906 space = 0; 907 908 /* 909 * Retest EOF - acquiring a new token can temporarily release 910 * tokens already held. 911 */ 912 if (wpipe->pipe_state & PIPE_WEOF) { 913 lwkt_reltoken(&wpipe->pipe_rlock); 914 error = EPIPE; 915 break; 916 } 917 918 /* 919 * We have no more space and have something to offer, 920 * wake up select/poll. 921 */ 922 if (space == 0) { 923 wpipe->pipe_state |= PIPE_WANTW; 924 ++wpipe->pipe_wantwcnt; 925 pipeselwakeup(wpipe); 926 if (wpipe->pipe_state & PIPE_WANTW) 927 error = tsleep(wpipe, PCATCH, "pipewr", 0); 928 ++pipe_wblocked_count; 929 } 930 lwkt_reltoken(&wpipe->pipe_rlock); 931 932 /* 933 * Break out if we errored or the read side wants us to go 934 * away. 935 */ 936 if (error) 937 break; 938 if (wpipe->pipe_state & PIPE_WEOF) { 939 error = EPIPE; 940 break; 941 } 942 } 943 pipe_end_uio(wpipe, &wpipe->pipe_wip); 944 945 /* 946 * If we have put any characters in the buffer, we wake up 947 * the reader. 948 * 949 * Both rlock and wlock are required to be able to modify pipe_state. 950 */ 951 if (wpipe->pipe_buffer.windex != wpipe->pipe_buffer.rindex) { 952 if (wpipe->pipe_state & PIPE_WANTR) { 953 lwkt_gettoken(&wpipe->pipe_rlock); 954 if (wpipe->pipe_state & PIPE_WANTR) { 955 wpipe->pipe_state &= ~PIPE_WANTR; 956 lwkt_reltoken(&wpipe->pipe_rlock); 957 wakeup(wpipe); 958 } else { 959 lwkt_reltoken(&wpipe->pipe_rlock); 960 } 961 } 962 if (pipeseltest(wpipe)) { 963 lwkt_gettoken(&wpipe->pipe_rlock); 964 pipeselwakeup(wpipe); 965 lwkt_reltoken(&wpipe->pipe_rlock); 966 } 967 } 968 969 /* 970 * Don't return EPIPE if I/O was successful 971 */ 972 if ((wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex) && 973 (uio->uio_resid == 0) && 974 (error == EPIPE)) { 975 error = 0; 976 } 977 978 if (error == 0) 979 vfs_timestamp(&wpipe->pipe_mtime); 980 981 /* 982 * We have something to offer, 983 * wake up select/poll. 984 */ 985 /*space = wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex;*/ 986 lwkt_reltoken(&wpipe->pipe_wlock); 987 pipe_rel_mplock(&mpsave); 988 return (error); 989 } 990 991 /* 992 * MPALMOSTSAFE - acquires mplock 993 * 994 * we implement a very minimal set of ioctls for compatibility with sockets. 995 */ 996 int 997 pipe_ioctl(struct file *fp, u_long cmd, caddr_t data, 998 struct ucred *cred, struct sysmsg *msg) 999 { 1000 struct pipe *mpipe; 1001 int error; 1002 int mpsave; 1003 1004 pipe_get_mplock(&mpsave); 1005 mpipe = (struct pipe *)fp->f_data; 1006 1007 lwkt_gettoken(&mpipe->pipe_rlock); 1008 lwkt_gettoken(&mpipe->pipe_wlock); 1009 1010 switch (cmd) { 1011 case FIOASYNC: 1012 if (*(int *)data) { 1013 mpipe->pipe_state |= PIPE_ASYNC; 1014 } else { 1015 mpipe->pipe_state &= ~PIPE_ASYNC; 1016 } 1017 error = 0; 1018 break; 1019 case FIONREAD: 1020 *(int *)data = mpipe->pipe_buffer.windex - 1021 mpipe->pipe_buffer.rindex; 1022 error = 0; 1023 break; 1024 case FIOSETOWN: 1025 get_mplock(); 1026 error = fsetown(*(int *)data, &mpipe->pipe_sigio); 1027 rel_mplock(); 1028 break; 1029 case FIOGETOWN: 1030 *(int *)data = fgetown(mpipe->pipe_sigio); 1031 error = 0; 1032 break; 1033 case TIOCSPGRP: 1034 /* This is deprecated, FIOSETOWN should be used instead. */ 1035 get_mplock(); 1036 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio); 1037 rel_mplock(); 1038 break; 1039 1040 case TIOCGPGRP: 1041 /* This is deprecated, FIOGETOWN should be used instead. */ 1042 *(int *)data = -fgetown(mpipe->pipe_sigio); 1043 error = 0; 1044 break; 1045 default: 1046 error = ENOTTY; 1047 break; 1048 } 1049 lwkt_reltoken(&mpipe->pipe_wlock); 1050 lwkt_reltoken(&mpipe->pipe_rlock); 1051 pipe_rel_mplock(&mpsave); 1052 1053 return (error); 1054 } 1055 1056 /* 1057 * MPSAFE 1058 */ 1059 static int 1060 pipe_stat(struct file *fp, struct stat *ub, struct ucred *cred) 1061 { 1062 struct pipe *pipe; 1063 int mpsave; 1064 1065 pipe_get_mplock(&mpsave); 1066 pipe = (struct pipe *)fp->f_data; 1067 1068 bzero((caddr_t)ub, sizeof(*ub)); 1069 ub->st_mode = S_IFIFO; 1070 ub->st_blksize = pipe->pipe_buffer.size; 1071 ub->st_size = pipe->pipe_buffer.windex - pipe->pipe_buffer.rindex; 1072 ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize; 1073 ub->st_atimespec = pipe->pipe_atime; 1074 ub->st_mtimespec = pipe->pipe_mtime; 1075 ub->st_ctimespec = pipe->pipe_ctime; 1076 /* 1077 * Left as 0: st_dev, st_ino, st_nlink, st_uid, st_gid, st_rdev, 1078 * st_flags, st_gen. 1079 * XXX (st_dev, st_ino) should be unique. 1080 */ 1081 pipe_rel_mplock(&mpsave); 1082 return (0); 1083 } 1084 1085 /* 1086 * MPALMOSTSAFE - acquires mplock 1087 */ 1088 static int 1089 pipe_close(struct file *fp) 1090 { 1091 struct pipe *cpipe; 1092 1093 get_mplock(); 1094 cpipe = (struct pipe *)fp->f_data; 1095 fp->f_ops = &badfileops; 1096 fp->f_data = NULL; 1097 funsetown(cpipe->pipe_sigio); 1098 pipeclose(cpipe); 1099 rel_mplock(); 1100 return (0); 1101 } 1102 1103 /* 1104 * Shutdown one or both directions of a full-duplex pipe. 1105 * 1106 * MPALMOSTSAFE - acquires mplock 1107 */ 1108 static int 1109 pipe_shutdown(struct file *fp, int how) 1110 { 1111 struct pipe *rpipe; 1112 struct pipe *wpipe; 1113 int error = EPIPE; 1114 int mpsave; 1115 1116 pipe_get_mplock(&mpsave); 1117 rpipe = (struct pipe *)fp->f_data; 1118 wpipe = rpipe->pipe_peer; 1119 1120 /* 1121 * We modify pipe_state on both pipes, which means we need 1122 * all four tokens! 1123 */ 1124 lwkt_gettoken(&rpipe->pipe_rlock); 1125 lwkt_gettoken(&rpipe->pipe_wlock); 1126 lwkt_gettoken(&wpipe->pipe_rlock); 1127 lwkt_gettoken(&wpipe->pipe_wlock); 1128 1129 switch(how) { 1130 case SHUT_RDWR: 1131 case SHUT_RD: 1132 rpipe->pipe_state |= PIPE_REOF; /* my reads */ 1133 rpipe->pipe_state |= PIPE_WEOF; /* peer writes */ 1134 if (rpipe->pipe_state & PIPE_WANTR) { 1135 rpipe->pipe_state &= ~PIPE_WANTR; 1136 wakeup(rpipe); 1137 } 1138 if (rpipe->pipe_state & PIPE_WANTW) { 1139 rpipe->pipe_state &= ~PIPE_WANTW; 1140 wakeup(rpipe); 1141 } 1142 error = 0; 1143 if (how == SHUT_RD) 1144 break; 1145 /* fall through */ 1146 case SHUT_WR: 1147 wpipe->pipe_state |= PIPE_REOF; /* peer reads */ 1148 wpipe->pipe_state |= PIPE_WEOF; /* my writes */ 1149 if (wpipe->pipe_state & PIPE_WANTR) { 1150 wpipe->pipe_state &= ~PIPE_WANTR; 1151 wakeup(wpipe); 1152 } 1153 if (wpipe->pipe_state & PIPE_WANTW) { 1154 wpipe->pipe_state &= ~PIPE_WANTW; 1155 wakeup(wpipe); 1156 } 1157 error = 0; 1158 break; 1159 } 1160 pipeselwakeup(rpipe); 1161 pipeselwakeup(wpipe); 1162 1163 lwkt_reltoken(&wpipe->pipe_wlock); 1164 lwkt_reltoken(&wpipe->pipe_rlock); 1165 lwkt_reltoken(&rpipe->pipe_wlock); 1166 lwkt_reltoken(&rpipe->pipe_rlock); 1167 1168 pipe_rel_mplock(&mpsave); 1169 return (error); 1170 } 1171 1172 static void 1173 pipe_free_kmem(struct pipe *cpipe) 1174 { 1175 if (cpipe->pipe_buffer.buffer != NULL) { 1176 if (cpipe->pipe_buffer.size > PIPE_SIZE) 1177 atomic_subtract_int(&pipe_nbig, 1); 1178 kmem_free(&kernel_map, 1179 (vm_offset_t)cpipe->pipe_buffer.buffer, 1180 cpipe->pipe_buffer.size); 1181 cpipe->pipe_buffer.buffer = NULL; 1182 cpipe->pipe_buffer.object = NULL; 1183 } 1184 } 1185 1186 /* 1187 * Close the pipe. The slock must be held to interlock against simultanious 1188 * closes. The rlock and wlock must be held to adjust the pipe_state. 1189 */ 1190 static void 1191 pipeclose(struct pipe *cpipe) 1192 { 1193 globaldata_t gd; 1194 struct pipe *ppipe; 1195 1196 if (cpipe == NULL) 1197 return; 1198 1199 /* 1200 * The slock may not have been allocated yet (close during 1201 * initialization) 1202 * 1203 * We need both the read and write tokens to modify pipe_state. 1204 */ 1205 if (cpipe->pipe_slock) 1206 lockmgr(cpipe->pipe_slock, LK_EXCLUSIVE); 1207 lwkt_gettoken(&cpipe->pipe_rlock); 1208 lwkt_gettoken(&cpipe->pipe_wlock); 1209 1210 /* 1211 * Set our state, wakeup anyone waiting in select, and 1212 * wakeup anyone blocked on our pipe. 1213 */ 1214 cpipe->pipe_state |= PIPE_CLOSED | PIPE_REOF | PIPE_WEOF; 1215 pipeselwakeup(cpipe); 1216 if (cpipe->pipe_state & (PIPE_WANTR | PIPE_WANTW)) { 1217 cpipe->pipe_state &= ~(PIPE_WANTR | PIPE_WANTW); 1218 wakeup(cpipe); 1219 } 1220 1221 /* 1222 * Disconnect from peer. 1223 */ 1224 if ((ppipe = cpipe->pipe_peer) != NULL) { 1225 lwkt_gettoken(&ppipe->pipe_rlock); 1226 lwkt_gettoken(&ppipe->pipe_wlock); 1227 ppipe->pipe_state |= PIPE_REOF | PIPE_WEOF; 1228 pipeselwakeup(ppipe); 1229 if (ppipe->pipe_state & (PIPE_WANTR | PIPE_WANTW)) { 1230 ppipe->pipe_state &= ~(PIPE_WANTR | PIPE_WANTW); 1231 wakeup(ppipe); 1232 } 1233 if (SLIST_FIRST(&ppipe->pipe_sel.si_note)) { 1234 get_mplock(); 1235 KNOTE(&ppipe->pipe_sel.si_note, 0); 1236 rel_mplock(); 1237 } 1238 lwkt_reltoken(&ppipe->pipe_wlock); 1239 lwkt_reltoken(&ppipe->pipe_rlock); 1240 } 1241 1242 /* 1243 * If the peer is also closed we can free resources for both 1244 * sides, otherwise we leave our side intact to deal with any 1245 * races (since we only have the slock). 1246 */ 1247 if (ppipe && (ppipe->pipe_state & PIPE_CLOSED)) { 1248 cpipe->pipe_peer = NULL; 1249 ppipe->pipe_peer = NULL; 1250 ppipe->pipe_slock = NULL; /* we will free the slock */ 1251 pipeclose(ppipe); 1252 ppipe = NULL; 1253 } 1254 1255 lwkt_reltoken(&cpipe->pipe_wlock); 1256 lwkt_reltoken(&cpipe->pipe_rlock); 1257 if (cpipe->pipe_slock) 1258 lockmgr(cpipe->pipe_slock, LK_RELEASE); 1259 1260 /* 1261 * If we disassociated from our peer we can free resources 1262 */ 1263 if (ppipe == NULL) { 1264 gd = mycpu; 1265 if (cpipe->pipe_slock) { 1266 kfree(cpipe->pipe_slock, M_PIPE); 1267 cpipe->pipe_slock = NULL; 1268 } 1269 if (gd->gd_pipeqcount >= pipe_maxcache || 1270 cpipe->pipe_buffer.size != PIPE_SIZE 1271 ) { 1272 pipe_free_kmem(cpipe); 1273 kfree(cpipe, M_PIPE); 1274 } else { 1275 cpipe->pipe_state = 0; 1276 cpipe->pipe_peer = gd->gd_pipeq; 1277 gd->gd_pipeq = cpipe; 1278 ++gd->gd_pipeqcount; 1279 } 1280 } 1281 } 1282 1283 /* 1284 * MPALMOSTSAFE - acquires mplock 1285 */ 1286 static int 1287 pipe_kqfilter(struct file *fp, struct knote *kn) 1288 { 1289 struct pipe *cpipe; 1290 1291 get_mplock(); 1292 cpipe = (struct pipe *)kn->kn_fp->f_data; 1293 1294 switch (kn->kn_filter) { 1295 case EVFILT_READ: 1296 kn->kn_fop = &pipe_rfiltops; 1297 break; 1298 case EVFILT_WRITE: 1299 kn->kn_fop = &pipe_wfiltops; 1300 cpipe = cpipe->pipe_peer; 1301 if (cpipe == NULL) { 1302 /* other end of pipe has been closed */ 1303 rel_mplock(); 1304 return (EPIPE); 1305 } 1306 break; 1307 default: 1308 rel_mplock(); 1309 return (EOPNOTSUPP); 1310 } 1311 kn->kn_hook = (caddr_t)cpipe; 1312 1313 SLIST_INSERT_HEAD(&cpipe->pipe_sel.si_note, kn, kn_selnext); 1314 rel_mplock(); 1315 return (0); 1316 } 1317 1318 static void 1319 filt_pipedetach(struct knote *kn) 1320 { 1321 struct pipe *cpipe = (struct pipe *)kn->kn_hook; 1322 1323 SLIST_REMOVE(&cpipe->pipe_sel.si_note, kn, knote, kn_selnext); 1324 } 1325 1326 /*ARGSUSED*/ 1327 static int 1328 filt_piperead(struct knote *kn, long hint) 1329 { 1330 struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data; 1331 1332 kn->kn_data = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex; 1333 1334 /* XXX RACE */ 1335 if (rpipe->pipe_state & PIPE_REOF) { 1336 kn->kn_flags |= EV_EOF; 1337 return (1); 1338 } 1339 return (kn->kn_data > 0); 1340 } 1341 1342 /*ARGSUSED*/ 1343 static int 1344 filt_pipewrite(struct knote *kn, long hint) 1345 { 1346 struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data; 1347 struct pipe *wpipe = rpipe->pipe_peer; 1348 u_int32_t space; 1349 1350 /* XXX RACE */ 1351 if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_WEOF)) { 1352 kn->kn_data = 0; 1353 kn->kn_flags |= EV_EOF; 1354 return (1); 1355 } 1356 space = wpipe->pipe_buffer.windex - 1357 wpipe->pipe_buffer.rindex; 1358 space = wpipe->pipe_buffer.size - space; 1359 kn->kn_data = space; 1360 return (kn->kn_data >= PIPE_BUF); 1361 } 1362