sys/kern/sys_pipe.c

   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 1.5 2003/07/19 21:14:39 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
  30 /*
  31  * This code has two modes of operation, a small write mode and a large
  32  * write mode.  The small write mode acts like conventional pipes with
  33  * a kernel buffer.  If the buffer is less than PIPE_MINDIRECT, then the
  34  * "normal" pipe buffering is done.  If the buffer is between PIPE_MINDIRECT
  35  * and PIPE_SIZE in size, it is fully mapped and wired into the kernel, and
  36  * the receiving process can copy it directly from the pages in the sending
  37  * process.
  38  *
  39  * If the sending process receives a signal, it is possible that it will
  40  * go away, and certainly its address space can change, because control
  41  * is returned back to the user-mode side.  In that case, the pipe code
  42  * arranges to copy the buffer supplied by the user process, to a pageable
  43  * kernel buffer, and the receiving process will grab the data from the
  44  * pageable kernel buffer.  Since signals don't happen all that often,
  45  * the copy operation is normally eliminated.
  46  *
  47  * The constant PIPE_MINDIRECT is chosen to make sure that buffering will
  48  * happen for small transfers so that the system will not spend all of
  49  * its time context switching.  PIPE_SIZE is constrained by the
  50  * amount of kernel virtual memory.
  51  */
  52
  53 #include <sys/param.h>
  54 #include <sys/systm.h>
  55 #include <sys/proc.h>
  56 #include <sys/fcntl.h>
  57 #include <sys/file.h>
  58 #include <sys/filedesc.h>
  59 #include <sys/filio.h>
  60 #include <sys/ttycom.h>
  61 #include <sys/stat.h>
  62 #include <sys/poll.h>
  63 #include <sys/select.h>
  64 #include <sys/signalvar.h>
  65 #include <sys/sysproto.h>
  66 #include <sys/pipe.h>
  67 #include <sys/vnode.h>
  68 #include <sys/uio.h>
  69 #include <sys/event.h>
  70
  71 #include <vm/vm.h>
  72 #include <vm/vm_param.h>
  73 #include <sys/lock.h>
  74 #include <vm/vm_object.h>
  75 #include <vm/vm_kern.h>
  76 #include <vm/vm_extern.h>
  77 #include <vm/pmap.h>
  78 #include <vm/vm_map.h>
  79 #include <vm/vm_page.h>
  80 #include <vm/vm_zone.h>
  81
  82 #include <sys/file2.h>
  83
  84 /*
  85  * Use this define if you want to disable *fancy* VM things.  Expect an
  86  * approx 30% decrease in transfer rate.  This could be useful for
  87  * NetBSD or OpenBSD.
  88  */
  89 /* #define PIPE_NODIRECT */
  90
  91 /*
  92  * interfaces to the outside world
  93  */
  94 static int pipe_read __P((struct file *fp, struct uio *uio,
  95                 struct ucred *cred, int flags, struct thread *td));
  96 static int pipe_write __P((struct file *fp, struct uio *uio,
  97                 struct ucred *cred, int flags, struct thread *td));
  98 static int pipe_close __P((struct file *fp, struct thread *td));
  99 static int pipe_poll __P((struct file *fp, int events, struct ucred *cred,
 100                 struct thread *td));
 101 static int pipe_kqfilter __P((struct file *fp, struct knote *kn));
 102 static int pipe_stat __P((struct file *fp, struct stat *sb, struct thread *td));
 103 static int pipe_ioctl __P((struct file *fp, u_long cmd, caddr_t data, struct thread *td));
 104
 105 static struct fileops pipeops = {
 106         pipe_read, pipe_write, pipe_ioctl, pipe_poll, pipe_kqfilter,
 107         pipe_stat, pipe_close
 108 };
 109
 110 static void     filt_pipedetach(struct knote *kn);
 111 static int      filt_piperead(struct knote *kn, long hint);
 112 static int      filt_pipewrite(struct knote *kn, long hint);
 113
 114 static struct filterops pipe_rfiltops =
 115         { 1, NULL, filt_pipedetach, filt_piperead };
 116 static struct filterops pipe_wfiltops =
 117         { 1, NULL, filt_pipedetach, filt_pipewrite };
 118
 119
 120 /*
 121  * Default pipe buffer size(s), this can be kind-of large now because pipe
 122  * space is pageable.  The pipe code will try to maintain locality of
 123  * reference for performance reasons, so small amounts of outstanding I/O
 124  * will not wipe the cache.
 125  */
 126 #define MINPIPESIZE (PIPE_SIZE/3)
 127 #define MAXPIPESIZE (2*PIPE_SIZE/3)
 128
 129 /*
 130  * Maximum amount of kva for pipes -- this is kind-of a soft limit, but
 131  * is there so that on large systems, we don't exhaust it.
 132  */
 133 #define MAXPIPEKVA (8*1024*1024)
 134
 135 /*
 136  * Limit for direct transfers, we cannot, of course limit
 137  * the amount of kva for pipes in general though.
 138  */
 139 #define LIMITPIPEKVA (16*1024*1024)
 140
 141 /*
 142  * Limit the number of "big" pipes
 143  */
 144 #define LIMITBIGPIPES   32
 145 static int nbigpipe;
 146
 147 static int amountpipekva;
 148
 149 static void pipeclose __P((struct pipe *cpipe));
 150 static void pipe_free_kmem __P((struct pipe *cpipe));
 151 static int pipe_create __P((struct pipe **cpipep));
 152 static __inline int pipelock __P((struct pipe *cpipe, int catch));
 153 static __inline void pipeunlock __P((struct pipe *cpipe));
 154 static __inline void pipeselwakeup __P((struct pipe *cpipe));
 155 #ifndef PIPE_NODIRECT
 156 static int pipe_build_write_buffer __P((struct pipe *wpipe, struct uio *uio));
 157 static void pipe_destroy_write_buffer __P((struct pipe *wpipe));
 158 static int pipe_direct_write __P((struct pipe *wpipe, struct uio *uio));
 159 static void pipe_clone_write_buffer __P((struct pipe *wpipe));
 160 #endif
 161 static int pipespace __P((struct pipe *cpipe, int size));
 162
 163 static vm_zone_t pipe_zone;
 164
 165 /*
 166  * The pipe system call for the DTYPE_PIPE type of pipes
 167  *
 168  * pipe_ARgs(int dummy)
 169  */
 170
 171 /* ARGSUSED */
 172 int
 173 pipe(struct pipe_args *uap)
 174 {
 175         struct thread *td = curthread;
 176         struct proc *p = td->td_proc;
 177         struct filedesc *fdp;
 178         struct file *rf, *wf;
 179         struct pipe *rpipe, *wpipe;
 180         int fd, error;
 181
 182         KKASSERT(p);
 183         fdp = p->p_fd;
 184
 185         if (pipe_zone == NULL)
 186                 pipe_zone = zinit("PIPE", sizeof(struct pipe), 0, 0, 4);
 187
 188         rpipe = wpipe = NULL;
 189         if (pipe_create(&rpipe) || pipe_create(&wpipe)) {
 190                 pipeclose(rpipe);
 191                 pipeclose(wpipe);
 192                 return (ENFILE);
 193         }
 194
 195         rpipe->pipe_state |= PIPE_DIRECTOK;
 196         wpipe->pipe_state |= PIPE_DIRECTOK;
 197
 198         error = falloc(p, &rf, &fd);
 199         if (error) {
 200                 pipeclose(rpipe);
 201                 pipeclose(wpipe);
 202                 return (error);
 203         }
 204         fhold(rf);
 205         p->p_retval[0] = fd;
 206
 207         /*
 208          * Warning: once we've gotten past allocation of the fd for the
 209          * read-side, we can only drop the read side via fdrop() in order
 210          * to avoid races against processes which manage to dup() the read
 211          * side while we are blocked trying to allocate the write side.
 212          */
 213         rf->f_flag = FREAD | FWRITE;
 214         rf->f_type = DTYPE_PIPE;
 215         rf->f_data = (caddr_t)rpipe;
 216         rf->f_ops = &pipeops;
 217         error = falloc(p, &wf, &fd);
 218         if (error) {
 219                 if (fdp->fd_ofiles[p->p_retval[0]] == rf) {
 220                         fdp->fd_ofiles[p->p_retval[0]] = NULL;
 221                         fdrop(rf, td);
 222                 }
 223                 fdrop(rf, td);
 224                 /* rpipe has been closed by fdrop(). */
 225                 pipeclose(wpipe);
 226                 return (error);
 227         }
 228         wf->f_flag = FREAD | FWRITE;
 229         wf->f_type = DTYPE_PIPE;
 230         wf->f_data = (caddr_t)wpipe;
 231         wf->f_ops = &pipeops;
 232         p->p_retval[1] = fd;
 233
 234         rpipe->pipe_peer = wpipe;
 235         wpipe->pipe_peer = rpipe;
 236         fdrop(rf, td);
 237
 238         return (0);
 239 }
 240
 241 /*
 242  * Allocate kva for pipe circular buffer, the space is pageable
 243  * This routine will 'realloc' the size of a pipe safely, if it fails
 244  * it will retain the old buffer.
 245  * If it fails it will return ENOMEM.
 246  */
 247 static int
 248 pipespace(cpipe, size)
 249         struct pipe *cpipe;
 250         int size;
 251 {
 252         struct vm_object *object;
 253         caddr_t buffer;
 254         int npages, error;
 255
 256         npages = round_page(size)/PAGE_SIZE;
 257         /*
 258          * Create an object, I don't like the idea of paging to/from
 259          * kernel_object.
 260          * XXX -- minor change needed here for NetBSD/OpenBSD VM systems.
 261          */
 262         object = vm_object_allocate(OBJT_DEFAULT, npages);
 263         buffer = (caddr_t) vm_map_min(kernel_map);
 264
 265         /*
 266          * Insert the object into the kernel map, and allocate kva for it.
 267          * The map entry is, by default, pageable.
 268          * XXX -- minor change needed here for NetBSD/OpenBSD VM systems.
 269          */
 270         error = vm_map_find(kernel_map, object, 0,
 271                 (vm_offset_t *) &buffer, size, 1,
 272                 VM_PROT_ALL, VM_PROT_ALL, 0);
 273
 274         if (error != KERN_SUCCESS) {
 275                 vm_object_deallocate(object);
 276                 return (ENOMEM);
 277         }
 278
 279         /* free old resources if we're resizing */
 280         pipe_free_kmem(cpipe);
 281         cpipe->pipe_buffer.object = object;
 282         cpipe->pipe_buffer.buffer = buffer;
 283         cpipe->pipe_buffer.size = size;
 284         cpipe->pipe_buffer.in = 0;
 285         cpipe->pipe_buffer.out = 0;
 286         cpipe->pipe_buffer.cnt = 0;
 287         amountpipekva += cpipe->pipe_buffer.size;
 288         return (0);
 289 }
 290
 291 /*
 292  * initialize and allocate VM and memory for pipe
 293  */
 294 static int
 295 pipe_create(cpipep)
 296         struct pipe **cpipep;
 297 {
 298         struct pipe *cpipe;
 299         int error;
 300
 301         *cpipep = zalloc(pipe_zone);
 302         if (*cpipep == NULL)
 303                 return (ENOMEM);
 304
 305         cpipe = *cpipep;
 306
 307         /* so pipespace()->pipe_free_kmem() doesn't follow junk pointer */
 308         cpipe->pipe_buffer.object = NULL;
 309 #ifndef PIPE_NODIRECT
 310         cpipe->pipe_map.kva = NULL;
 311 #endif
 312         /*
 313          * protect so pipeclose() doesn't follow a junk pointer
 314          * if pipespace() fails.
 315          */
 316         bzero(&cpipe->pipe_sel, sizeof(cpipe->pipe_sel));
 317         cpipe->pipe_state = 0;
 318         cpipe->pipe_peer = NULL;
 319         cpipe->pipe_busy = 0;
 320
 321 #ifndef PIPE_NODIRECT
 322         /*
 323          * pipe data structure initializations to support direct pipe I/O
 324          */
 325         cpipe->pipe_map.cnt = 0;
 326         cpipe->pipe_map.kva = 0;
 327         cpipe->pipe_map.pos = 0;
 328         cpipe->pipe_map.npages = 0;
 329         /* cpipe->pipe_map.ms[] = invalid */
 330 #endif
 331
 332         error = pipespace(cpipe, PIPE_SIZE);
 333         if (error)
 334                 return (error);
 335
 336         vfs_timestamp(&cpipe->pipe_ctime);
 337         cpipe->pipe_atime = cpipe->pipe_ctime;
 338         cpipe->pipe_mtime = cpipe->pipe_ctime;
 339
 340         return (0);
 341 }
 342
 343
 344 /*
 345  * lock a pipe for I/O, blocking other access
 346  */
 347 static __inline int
 348 pipelock(cpipe, catch)
 349         struct pipe *cpipe;
 350         int catch;
 351 {
 352         int error;
 353
 354         while (cpipe->pipe_state & PIPE_LOCK) {
 355                 cpipe->pipe_state |= PIPE_LWANT;
 356                 error = tsleep(cpipe, (catch ? PCATCH : 0), "pipelk", 0);
 357                 if (error != 0)
 358                         return (error);
 359         }
 360         cpipe->pipe_state |= PIPE_LOCK;
 361         return (0);
 362 }
 363
 364 /*
 365  * unlock a pipe I/O lock
 366  */
 367 static __inline void
 368 pipeunlock(cpipe)
 369         struct pipe *cpipe;
 370 {
 371
 372         cpipe->pipe_state &= ~PIPE_LOCK;
 373         if (cpipe->pipe_state & PIPE_LWANT) {
 374                 cpipe->pipe_state &= ~PIPE_LWANT;
 375                 wakeup(cpipe);
 376         }
 377 }
 378
 379 static __inline void
 380 pipeselwakeup(cpipe)
 381         struct pipe *cpipe;
 382 {
 383
 384         if (cpipe->pipe_state & PIPE_SEL) {
 385                 cpipe->pipe_state &= ~PIPE_SEL;
 386                 selwakeup(&cpipe->pipe_sel);
 387         }
 388         if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
 389                 pgsigio(cpipe->pipe_sigio, SIGIO, 0);
 390         KNOTE(&cpipe->pipe_sel.si_note, 0);
 391 }
 392
 393 /* ARGSUSED */
 394 static int
 395 pipe_read(struct file *fp, struct uio *uio, struct ucred *cred,
 396         int flags, struct thread *td)
 397 {
 398         struct pipe *rpipe = (struct pipe *) fp->f_data;
 399         int error;
 400         int nread = 0;
 401         u_int size;
 402
 403         ++rpipe->pipe_busy;
 404         error = pipelock(rpipe, 1);
 405         if (error)
 406                 goto unlocked_error;
 407
 408         while (uio->uio_resid) {
 409                 /*
 410                  * normal pipe buffer receive
 411                  */
 412                 if (rpipe->pipe_buffer.cnt > 0) {
 413                         size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
 414                         if (size > rpipe->pipe_buffer.cnt)
 415                                 size = rpipe->pipe_buffer.cnt;
 416                         if (size > (u_int) uio->uio_resid)
 417                                 size = (u_int) uio->uio_resid;
 418
 419                         error = uiomove(&rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
 420                                         size, uio);
 421                         if (error)
 422                                 break;
 423
 424                         rpipe->pipe_buffer.out += size;
 425                         if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
 426                                 rpipe->pipe_buffer.out = 0;
 427
 428                         rpipe->pipe_buffer.cnt -= size;
 429
 430                         /*
 431                          * If there is no more to read in the pipe, reset
 432                          * its pointers to the beginning.  This improves
 433                          * cache hit stats.
 434                          */
 435                         if (rpipe->pipe_buffer.cnt == 0) {
 436                                 rpipe->pipe_buffer.in = 0;
 437                                 rpipe->pipe_buffer.out = 0;
 438                         }
 439                         nread += size;
 440 #ifndef PIPE_NODIRECT
 441                 /*
 442                  * Direct copy, bypassing a kernel buffer.
 443                  */
 444                 } else if ((size = rpipe->pipe_map.cnt) &&
 445                            (rpipe->pipe_state & PIPE_DIRECTW)) {
 446                         caddr_t va;
 447                         if (size > (u_int) uio->uio_resid)
 448                                 size = (u_int) uio->uio_resid;
 449
 450                         va = (caddr_t) rpipe->pipe_map.kva +
 451                             rpipe->pipe_map.pos;
 452                         error = uiomove(va, size, uio);
 453                         if (error)
 454                                 break;
 455                         nread += size;
 456                         rpipe->pipe_map.pos += size;
 457                         rpipe->pipe_map.cnt -= size;
 458                         if (rpipe->pipe_map.cnt == 0) {
 459                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
 460                                 wakeup(rpipe);
 461                         }
 462 #endif
 463                 } else {
 464                         /*
 465                          * detect EOF condition
 466                          * read returns 0 on EOF, no need to set error
 467                          */
 468                         if (rpipe->pipe_state & PIPE_EOF)
 469                                 break;
 470
 471                         /*
 472                          * If the "write-side" has been blocked, wake it up now.
 473                          */
 474                         if (rpipe->pipe_state & PIPE_WANTW) {
 475                                 rpipe->pipe_state &= ~PIPE_WANTW;
 476                                 wakeup(rpipe);
 477                         }
 478
 479                         /*
 480                          * Break if some data was read.
 481                          */
 482                         if (nread > 0)
 483                                 break;
 484
 485                         /*
 486                          * Unlock the pipe buffer for our remaining processing.  We
 487                          * will either break out with an error or we will sleep and
 488                          * relock to loop.
 489                          */
 490                         pipeunlock(rpipe);
 491
 492                         /*
 493                          * Handle non-blocking mode operation or
 494                          * wait for more data.
 495                          */
 496                         if (fp->f_flag & FNONBLOCK) {
 497                                 error = EAGAIN;
 498                         } else {
 499                                 rpipe->pipe_state |= PIPE_WANTR;
 500                                 if ((error = tsleep(rpipe, PCATCH,
 501                                     "piperd", 0)) == 0) {
 502                                         error = pipelock(rpipe, 1);
 503                                 }
 504                         }
 505                         if (error)
 506                                 goto unlocked_error;
 507                 }
 508         }
 509         pipeunlock(rpipe);
 510
 511         if (error == 0)
 512                 vfs_timestamp(&rpipe->pipe_atime);
 513 unlocked_error:
 514         --rpipe->pipe_busy;
 515
 516         /*
 517          * PIPE_WANT processing only makes sense if pipe_busy is 0.
 518          */
 519         if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
 520                 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
 521                 wakeup(rpipe);
 522         } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
 523                 /*
 524                  * Handle write blocking hysteresis.
 525                  */
 526                 if (rpipe->pipe_state & PIPE_WANTW) {
 527                         rpipe->pipe_state &= ~PIPE_WANTW;
 528                         wakeup(rpipe);
 529                 }
 530         }
 531
 532         if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
 533                 pipeselwakeup(rpipe);
 534
 535         return (error);
 536 }
 537
 538 #ifndef PIPE_NODIRECT
 539 /*
 540  * Map the sending processes' buffer into kernel space and wire it.
 541  * This is similar to a physical write operation.
 542  */
 543 static int
 544 pipe_build_write_buffer(wpipe, uio)
 545         struct pipe *wpipe;
 546         struct uio *uio;
 547 {
 548         u_int size;
 549         int i;
 550         vm_offset_t addr, endaddr, paddr;
 551
 552         size = (u_int) uio->uio_iov->iov_len;
 553         if (size > wpipe->pipe_buffer.size)
 554                 size = wpipe->pipe_buffer.size;
 555
 556         endaddr = round_page((vm_offset_t)uio->uio_iov->iov_base + size);
 557         addr = trunc_page((vm_offset_t)uio->uio_iov->iov_base);
 558         for (i = 0; addr < endaddr; addr += PAGE_SIZE, i++) {
 559                 vm_page_t m;
 560
 561                 if (vm_fault_quick((caddr_t)addr, VM_PROT_READ) < 0 ||
 562                     (paddr = pmap_kextract(addr)) == 0) {
 563                         int j;
 564
 565                         for (j = 0; j < i; j++)
 566                                 vm_page_unwire(wpipe->pipe_map.ms[j], 1);
 567                         return (EFAULT);
 568                 }
 569
 570                 m = PHYS_TO_VM_PAGE(paddr);
 571                 vm_page_wire(m);
 572                 wpipe->pipe_map.ms[i] = m;
 573         }
 574
 575 /*
 576  * set up the control block
 577  */
 578         wpipe->pipe_map.npages = i;
 579         wpipe->pipe_map.pos =
 580             ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
 581         wpipe->pipe_map.cnt = size;
 582
 583 /*
 584  * and map the buffer
 585  */
 586         if (wpipe->pipe_map.kva == 0) {
 587                 /*
 588                  * We need to allocate space for an extra page because the
 589                  * address range might (will) span pages at times.
 590                  */
 591                 wpipe->pipe_map.kva = kmem_alloc_pageable(kernel_map,
 592                         wpipe->pipe_buffer.size + PAGE_SIZE);
 593                 amountpipekva += wpipe->pipe_buffer.size + PAGE_SIZE;
 594         }
 595         pmap_qenter(wpipe->pipe_map.kva, wpipe->pipe_map.ms,
 596                 wpipe->pipe_map.npages);
 597
 598 /*
 599  * and update the uio data
 600  */
 601
 602         uio->uio_iov->iov_len -= size;
 603         uio->uio_iov->iov_base += size;
 604         if (uio->uio_iov->iov_len == 0)
 605                 uio->uio_iov++;
 606         uio->uio_resid -= size;
 607         uio->uio_offset += size;
 608         return (0);
 609 }
 610
 611 /*
 612  * unmap and unwire the process buffer
 613  */
 614 static void
 615 pipe_destroy_write_buffer(wpipe)
 616         struct pipe *wpipe;
 617 {
 618         int i;
 619
 620         if (wpipe->pipe_map.kva) {
 621                 pmap_qremove(wpipe->pipe_map.kva, wpipe->pipe_map.npages);
 622
 623                 if (amountpipekva > MAXPIPEKVA) {
 624                         vm_offset_t kva = wpipe->pipe_map.kva;
 625                         wpipe->pipe_map.kva = 0;
 626                         kmem_free(kernel_map, kva,
 627                                 wpipe->pipe_buffer.size + PAGE_SIZE);
 628                         amountpipekva -= wpipe->pipe_buffer.size + PAGE_SIZE;
 629                 }
 630         }
 631         for (i = 0; i < wpipe->pipe_map.npages; i++)
 632                 vm_page_unwire(wpipe->pipe_map.ms[i], 1);
 633         wpipe->pipe_map.npages = 0;
 634 }
 635
 636 /*
 637  * In the case of a signal, the writing process might go away.  This
 638  * code copies the data into the circular buffer so that the source
 639  * pages can be freed without loss of data.
 640  */
 641 static void
 642 pipe_clone_write_buffer(wpipe)
 643         struct pipe *wpipe;
 644 {
 645         int size;
 646         int pos;
 647
 648         size = wpipe->pipe_map.cnt;
 649         pos = wpipe->pipe_map.pos;
 650         bcopy((caddr_t) wpipe->pipe_map.kva + pos,
 651             (caddr_t) wpipe->pipe_buffer.buffer, size);
 652
 653         wpipe->pipe_buffer.in = size;
 654         wpipe->pipe_buffer.out = 0;
 655         wpipe->pipe_buffer.cnt = size;
 656         wpipe->pipe_state &= ~PIPE_DIRECTW;
 657
 658         pipe_destroy_write_buffer(wpipe);
 659 }
 660
 661 /*
 662  * This implements the pipe buffer write mechanism.  Note that only
 663  * a direct write OR a normal pipe write can be pending at any given time.
 664  * If there are any characters in the pipe buffer, the direct write will
 665  * be deferred until the receiving process grabs all of the bytes from
 666  * the pipe buffer.  Then the direct mapping write is set-up.
 667  */
 668 static int
 669 pipe_direct_write(wpipe, uio)
 670         struct pipe *wpipe;
 671         struct uio *uio;
 672 {
 673         int error;
 674
 675 retry:
 676         while (wpipe->pipe_state & PIPE_DIRECTW) {
 677                 if (wpipe->pipe_state & PIPE_WANTR) {
 678                         wpipe->pipe_state &= ~PIPE_WANTR;
 679                         wakeup(wpipe);
 680                 }
 681                 wpipe->pipe_state |= PIPE_WANTW;
 682                 error = tsleep(wpipe, PCATCH, "pipdww", 0);
 683                 if (error)
 684                         goto error1;
 685                 if (wpipe->pipe_state & PIPE_EOF) {
 686                         error = EPIPE;
 687                         goto error1;
 688                 }
 689         }
 690         wpipe->pipe_map.cnt = 0;        /* transfer not ready yet */
 691         if (wpipe->pipe_buffer.cnt > 0) {
 692                 if (wpipe->pipe_state & PIPE_WANTR) {
 693                         wpipe->pipe_state &= ~PIPE_WANTR;
 694                         wakeup(wpipe);
 695                 }
 696
 697                 wpipe->pipe_state |= PIPE_WANTW;
 698                 error = tsleep(wpipe, PCATCH, "pipdwc", 0);
 699                 if (error)
 700                         goto error1;
 701                 if (wpipe->pipe_state & PIPE_EOF) {
 702                         error = EPIPE;
 703                         goto error1;
 704                 }
 705                 goto retry;
 706         }
 707
 708         wpipe->pipe_state |= PIPE_DIRECTW;
 709
 710         error = pipe_build_write_buffer(wpipe, uio);
 711         if (error) {
 712                 wpipe->pipe_state &= ~PIPE_DIRECTW;
 713                 goto error1;
 714         }
 715
 716         error = 0;
 717         while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
 718                 if (wpipe->pipe_state & PIPE_EOF) {
 719                         pipelock(wpipe, 0);
 720                         pipe_destroy_write_buffer(wpipe);
 721                         pipeunlock(wpipe);
 722                         pipeselwakeup(wpipe);
 723                         error = EPIPE;
 724                         goto error1;
 725                 }
 726                 if (wpipe->pipe_state & PIPE_WANTR) {
 727                         wpipe->pipe_state &= ~PIPE_WANTR;
 728                         wakeup(wpipe);
 729                 }
 730                 pipeselwakeup(wpipe);
 731                 error = tsleep(wpipe, PCATCH, "pipdwt", 0);
 732         }
 733
 734         pipelock(wpipe,0);
 735         if (wpipe->pipe_state & PIPE_DIRECTW) {
 736                 /*
 737                  * this bit of trickery substitutes a kernel buffer for
 738                  * the process that might be going away.
 739                  */
 740                 pipe_clone_write_buffer(wpipe);
 741         } else {
 742                 pipe_destroy_write_buffer(wpipe);
 743         }
 744         pipeunlock(wpipe);
 745         return (error);
 746
 747 error1:
 748         wakeup(wpipe);
 749         return (error);
 750 }
 751 #endif
 752
 753 static int
 754 pipe_write(struct file *fp, struct uio *uio, struct ucred *cred,
 755         int flags, struct thread *td)
 756 {
 757         int error = 0;
 758         int orig_resid;
 759         struct pipe *wpipe, *rpipe;
 760
 761         rpipe = (struct pipe *) fp->f_data;
 762         wpipe = rpipe->pipe_peer;
 763
 764         /*
 765          * detect loss of pipe read side, issue SIGPIPE if lost.
 766          */
 767         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
 768                 return (EPIPE);
 769         }
 770         ++wpipe->pipe_busy;
 771
 772         /*
 773          * If it is advantageous to resize the pipe buffer, do
 774          * so.
 775          */
 776         if ((uio->uio_resid > PIPE_SIZE) &&
 777                 (nbigpipe < LIMITBIGPIPES) &&
 778                 (wpipe->pipe_state & PIPE_DIRECTW) == 0 &&
 779                 (wpipe->pipe_buffer.size <= PIPE_SIZE) &&
 780                 (wpipe->pipe_buffer.cnt == 0)) {
 781
 782                 if ((error = pipelock(wpipe,1)) == 0) {
 783                         if (pipespace(wpipe, BIG_PIPE_SIZE) == 0)
 784                                 nbigpipe++;
 785                         pipeunlock(wpipe);
 786                 }
 787         }
 788
 789         /*
 790          * If an early error occured unbusy and return, waking up any pending
 791          * readers.
 792          */
 793         if (error) {
 794                 --wpipe->pipe_busy;
 795                 if ((wpipe->pipe_busy == 0) &&
 796                     (wpipe->pipe_state & PIPE_WANT)) {
 797                         wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
 798                         wakeup(wpipe);
 799                 }
 800                 return(error);
 801         }
 802
 803         KASSERT(wpipe->pipe_buffer.buffer != NULL, ("pipe buffer gone"));
 804
 805         orig_resid = uio->uio_resid;
 806
 807         while (uio->uio_resid) {
 808                 int space;
 809
 810 #ifndef PIPE_NODIRECT
 811                 /*
 812                  * If the transfer is large, we can gain performance if
 813                  * we do process-to-process copies directly.
 814                  * If the write is non-blocking, we don't use the
 815                  * direct write mechanism.
 816                  *
 817                  * The direct write mechanism will detect the reader going
 818                  * away on us.
 819                  */
 820                 if ((uio->uio_iov->iov_len >= PIPE_MINDIRECT) &&
 821                     (fp->f_flag & FNONBLOCK) == 0 &&
 822                         (wpipe->pipe_map.kva || (amountpipekva < LIMITPIPEKVA)) &&
 823                         (uio->uio_iov->iov_len >= PIPE_MINDIRECT)) {
 824                         error = pipe_direct_write( wpipe, uio);
 825                         if (error)
 826                                 break;
 827                         continue;
 828                 }
 829 #endif
 830
 831                 /*
 832                  * Pipe buffered writes cannot be coincidental with
 833                  * direct writes.  We wait until the currently executing
 834                  * direct write is completed before we start filling the
 835                  * pipe buffer.  We break out if a signal occurs or the
 836                  * reader goes away.
 837                  */
 838         retrywrite:
 839                 while (wpipe->pipe_state & PIPE_DIRECTW) {
 840                         if (wpipe->pipe_state & PIPE_WANTR) {
 841                                 wpipe->pipe_state &= ~PIPE_WANTR;
 842                                 wakeup(wpipe);
 843                         }
 844                         error = tsleep(wpipe, PCATCH, "pipbww", 0);
 845                         if (wpipe->pipe_state & PIPE_EOF)
 846                                 break;
 847                         if (error)
 848                                 break;
 849                 }
 850                 if (wpipe->pipe_state & PIPE_EOF) {
 851                         error = EPIPE;
 852                         break;
 853                 }
 854
 855                 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
 856
 857                 /* Writes of size <= PIPE_BUF must be atomic. */
 858                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
 859                         space = 0;
 860
 861                 if (space > 0 && (wpipe->pipe_buffer.cnt < PIPE_SIZE)) {
 862                         if ((error = pipelock(wpipe,1)) == 0) {
 863                                 int size;       /* Transfer size */
 864                                 int segsize;    /* first segment to transfer */
 865
 866                                 /*
 867                                  * It is possible for a direct write to
 868                                  * slip in on us... handle it here...
 869                                  */
 870                                 if (wpipe->pipe_state & PIPE_DIRECTW) {
 871                                         pipeunlock(wpipe);
 872                                         goto retrywrite;
 873                                 }
 874                                 /*
 875                                  * If a process blocked in uiomove, our
 876                                  * value for space might be bad.
 877                                  *
 878                                  * XXX will we be ok if the reader has gone
 879                                  * away here?
 880                                  */
 881                                 if (space > wpipe->pipe_buffer.size -
 882                                     wpipe->pipe_buffer.cnt) {
 883                                         pipeunlock(wpipe);
 884                                         goto retrywrite;
 885                                 }
 886
 887                                 /*
 888                                  * Transfer size is minimum of uio transfer
 889                                  * and free space in pipe buffer.
 890                                  */
 891                                 if (space > uio->uio_resid)
 892                                         size = uio->uio_resid;
 893                                 else
 894                                         size = space;
 895                                 /*
 896                                  * First segment to transfer is minimum of
 897                                  * transfer size and contiguous space in
 898                                  * pipe buffer.  If first segment to transfer
 899                                  * is less than the transfer size, we've got
 900                                  * a wraparound in the buffer.
 901                                  */
 902                                 segsize = wpipe->pipe_buffer.size -
 903                                         wpipe->pipe_buffer.in;
 904                                 if (segsize > size)
 905                                         segsize = size;
 906
 907                                 /* Transfer first segment */
 908
 909                                 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
 910                                                 segsize, uio);
 911
 912                                 if (error == 0 && segsize < size) {
 913                                         /*
 914                                          * Transfer remaining part now, to
 915                                          * support atomic writes.  Wraparound
 916                                          * happened.
 917                                          */
 918                                         if (wpipe->pipe_buffer.in + segsize !=
 919                                             wpipe->pipe_buffer.size)
 920                                                 panic("Expected pipe buffer wraparound disappeared");
 921
 922                                         error = uiomove(&wpipe->pipe_buffer.buffer[0],
 923                                                         size - segsize, uio);
 924                                 }
 925                                 if (error == 0) {
 926                                         wpipe->pipe_buffer.in += size;
 927                                         if (wpipe->pipe_buffer.in >=
 928                                             wpipe->pipe_buffer.size) {
 929                                                 if (wpipe->pipe_buffer.in != size - segsize + wpipe->pipe_buffer.size)
 930                                                         panic("Expected wraparound bad");
 931                                                 wpipe->pipe_buffer.in = size - segsize;
 932                                         }
 933
 934                                         wpipe->pipe_buffer.cnt += size;
 935                                         if (wpipe->pipe_buffer.cnt > wpipe->pipe_buffer.size)
 936                                                 panic("Pipe buffer overflow");
 937
 938                                 }
 939                                 pipeunlock(wpipe);
 940                         }
 941                         if (error)
 942                                 break;
 943
 944                 } else {
 945                         /*
 946                          * If the "read-side" has been blocked, wake it up now.
 947                          */
 948                         if (wpipe->pipe_state & PIPE_WANTR) {
 949                                 wpipe->pipe_state &= ~PIPE_WANTR;
 950                                 wakeup(wpipe);
 951                         }
 952
 953                         /*
 954                          * don't block on non-blocking I/O
 955                          */
 956                         if (fp->f_flag & FNONBLOCK) {
 957                                 error = EAGAIN;
 958                                 break;
 959                         }
 960
 961                         /*
 962                          * We have no more space and have something to offer,
 963                          * wake up select/poll.
 964                          */
 965                         pipeselwakeup(wpipe);
 966
 967                         wpipe->pipe_state |= PIPE_WANTW;
 968                         error = tsleep(wpipe, PCATCH, "pipewr", 0);
 969                         if (error != 0)
 970                                 break;
 971                         /*
 972                          * If read side wants to go away, we just issue a signal
 973                          * to ourselves.
 974                          */
 975                         if (wpipe->pipe_state & PIPE_EOF) {
 976                                 error = EPIPE;
 977                                 break;
 978                         }
 979                 }
 980         }
 981
 982         --wpipe->pipe_busy;
 983
 984         if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
 985                 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
 986                 wakeup(wpipe);
 987         } else if (wpipe->pipe_buffer.cnt > 0) {
 988                 /*
 989                  * If we have put any characters in the buffer, we wake up
 990                  * the reader.
 991                  */
 992                 if (wpipe->pipe_state & PIPE_WANTR) {
 993                         wpipe->pipe_state &= ~PIPE_WANTR;
 994                         wakeup(wpipe);
 995                 }
 996         }
 997
 998         /*
 999          * Don't return EPIPE if I/O was successful
1000          */
1001         if ((wpipe->pipe_buffer.cnt == 0) &&
1002             (uio->uio_resid == 0) &&
1003             (error == EPIPE)) {
1004                 error = 0;
1005         }
1006
1007         if (error == 0)
1008                 vfs_timestamp(&wpipe->pipe_mtime);
1009
1010         /*
1011          * We have something to offer,
1012          * wake up select/poll.
1013          */
1014         if (wpipe->pipe_buffer.cnt)
1015                 pipeselwakeup(wpipe);
1016
1017         return (error);
1018 }
1019
1020 /*
1021  * we implement a very minimal set of ioctls for compatibility with sockets.
1022  */
1023 int
1024 pipe_ioctl(struct file *fp, u_long cmd, caddr_t data, struct thread *td)
1025 {
1026         struct pipe *mpipe = (struct pipe *)fp->f_data;
1027
1028         switch (cmd) {
1029
1030         case FIONBIO:
1031                 return (0);
1032
1033         case FIOASYNC:
1034                 if (*(int *)data) {
1035                         mpipe->pipe_state |= PIPE_ASYNC;
1036                 } else {
1037                         mpipe->pipe_state &= ~PIPE_ASYNC;
1038                 }
1039                 return (0);
1040
1041         case FIONREAD:
1042                 if (mpipe->pipe_state & PIPE_DIRECTW)
1043                         *(int *)data = mpipe->pipe_map.cnt;
1044                 else
1045                         *(int *)data = mpipe->pipe_buffer.cnt;
1046                 return (0);
1047
1048         case FIOSETOWN:
1049                 return (fsetown(*(int *)data, &mpipe->pipe_sigio));
1050
1051         case FIOGETOWN:
1052                 *(int *)data = fgetown(mpipe->pipe_sigio);
1053                 return (0);
1054
1055         /* This is deprecated, FIOSETOWN should be used instead. */
1056         case TIOCSPGRP:
1057                 return (fsetown(-(*(int *)data), &mpipe->pipe_sigio));
1058
1059         /* This is deprecated, FIOGETOWN should be used instead. */
1060         case TIOCGPGRP:
1061                 *(int *)data = -fgetown(mpipe->pipe_sigio);
1062                 return (0);
1063
1064         }
1065         return (ENOTTY);
1066 }
1067
1068 int
1069 pipe_poll(struct file *fp, int events, struct ucred *cred, struct thread *td)
1070 {
1071         struct pipe *rpipe = (struct pipe *)fp->f_data;
1072         struct pipe *wpipe;
1073         int revents = 0;
1074
1075         wpipe = rpipe->pipe_peer;
1076         if (events & (POLLIN | POLLRDNORM))
1077                 if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1078                     (rpipe->pipe_buffer.cnt > 0) ||
1079                     (rpipe->pipe_state & PIPE_EOF))
1080                         revents |= events & (POLLIN | POLLRDNORM);
1081
1082         if (events & (POLLOUT | POLLWRNORM))
1083                 if (wpipe == NULL || (wpipe->pipe_state & PIPE_EOF) ||
1084                     (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1085                      (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF))
1086                         revents |= events & (POLLOUT | POLLWRNORM);
1087
1088         if ((rpipe->pipe_state & PIPE_EOF) ||
1089             (wpipe == NULL) ||
1090             (wpipe->pipe_state & PIPE_EOF))
1091                 revents |= POLLHUP;
1092
1093         if (revents == 0) {
1094                 if (events & (POLLIN | POLLRDNORM)) {
1095                         selrecord(td, &rpipe->pipe_sel);
1096                         rpipe->pipe_state |= PIPE_SEL;
1097                 }
1098
1099                 if (events & (POLLOUT | POLLWRNORM)) {
1100                         selrecord(td, &wpipe->pipe_sel);
1101                         wpipe->pipe_state |= PIPE_SEL;
1102                 }
1103         }
1104
1105         return (revents);
1106 }
1107
1108 static int
1109 pipe_stat(struct file *fp, struct stat *ub, struct thread *td)
1110 {
1111         struct pipe *pipe = (struct pipe *)fp->f_data;
1112
1113         bzero((caddr_t)ub, sizeof(*ub));
1114         ub->st_mode = S_IFIFO;
1115         ub->st_blksize = pipe->pipe_buffer.size;
1116         ub->st_size = pipe->pipe_buffer.cnt;
1117         ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1118         ub->st_atimespec = pipe->pipe_atime;
1119         ub->st_mtimespec = pipe->pipe_mtime;
1120         ub->st_ctimespec = pipe->pipe_ctime;
1121         /*
1122          * Left as 0: st_dev, st_ino, st_nlink, st_uid, st_gid, st_rdev,
1123          * st_flags, st_gen.
1124          * XXX (st_dev, st_ino) should be unique.
1125          */
1126         return (0);
1127 }
1128
1129 /* ARGSUSED */
1130 static int
1131 pipe_close(struct file *fp, struct thread *td)
1132 {
1133         struct pipe *cpipe = (struct pipe *)fp->f_data;
1134
1135         fp->f_ops = &badfileops;
1136         fp->f_data = NULL;
1137         funsetown(cpipe->pipe_sigio);
1138         pipeclose(cpipe);
1139         return (0);
1140 }
1141
1142 static void
1143 pipe_free_kmem(struct pipe *cpipe)
1144 {
1145
1146         if (cpipe->pipe_buffer.buffer != NULL) {
1147                 if (cpipe->pipe_buffer.size > PIPE_SIZE)
1148                         --nbigpipe;
1149                 amountpipekva -= cpipe->pipe_buffer.size;
1150                 kmem_free(kernel_map,
1151                         (vm_offset_t)cpipe->pipe_buffer.buffer,
1152                         cpipe->pipe_buffer.size);
1153                 cpipe->pipe_buffer.buffer = NULL;
1154         }
1155 #ifndef PIPE_NODIRECT
1156         if (cpipe->pipe_map.kva != NULL) {
1157                 amountpipekva -= cpipe->pipe_buffer.size + PAGE_SIZE;
1158                 kmem_free(kernel_map,
1159                         cpipe->pipe_map.kva,
1160                         cpipe->pipe_buffer.size + PAGE_SIZE);
1161                 cpipe->pipe_map.cnt = 0;
1162                 cpipe->pipe_map.kva = 0;
1163                 cpipe->pipe_map.pos = 0;
1164                 cpipe->pipe_map.npages = 0;
1165         }
1166 #endif
1167 }
1168
1169 /*
1170  * shutdown the pipe
1171  */
1172 static void
1173 pipeclose(struct pipe *cpipe)
1174 {
1175         struct pipe *ppipe;
1176
1177         if (cpipe) {
1178
1179                 pipeselwakeup(cpipe);
1180
1181                 /*
1182                  * If the other side is blocked, wake it up saying that
1183                  * we want to close it down.
1184                  */
1185                 while (cpipe->pipe_busy) {
1186                         wakeup(cpipe);
1187                         cpipe->pipe_state |= PIPE_WANT | PIPE_EOF;
1188                         tsleep(cpipe, 0, "pipecl", 0);
1189                 }
1190
1191                 /*
1192                  * Disconnect from peer
1193                  */
1194                 if ((ppipe = cpipe->pipe_peer) != NULL) {
1195                         pipeselwakeup(ppipe);
1196
1197                         ppipe->pipe_state |= PIPE_EOF;
1198                         wakeup(ppipe);
1199                         KNOTE(&ppipe->pipe_sel.si_note, 0);
1200                         ppipe->pipe_peer = NULL;
1201                 }
1202                 /*
1203                  * free resources
1204                  */
1205                 pipe_free_kmem(cpipe);
1206                 zfree(pipe_zone, cpipe);
1207         }
1208 }
1209
1210 /*ARGSUSED*/
1211 static int
1212 pipe_kqfilter(struct file *fp, struct knote *kn)
1213 {
1214         struct pipe *cpipe = (struct pipe *)kn->kn_fp->f_data;
1215
1216         switch (kn->kn_filter) {
1217         case EVFILT_READ:
1218                 kn->kn_fop = &pipe_rfiltops;
1219                 break;
1220         case EVFILT_WRITE:
1221                 kn->kn_fop = &pipe_wfiltops;
1222                 cpipe = cpipe->pipe_peer;
1223                 if (cpipe == NULL)
1224                         /* other end of pipe has been closed */
1225                         return (EBADF);
1226                 break;
1227         default:
1228                 return (1);
1229         }
1230         kn->kn_hook = (caddr_t)cpipe;
1231
1232         SLIST_INSERT_HEAD(&cpipe->pipe_sel.si_note, kn, kn_selnext);
1233         return (0);
1234 }
1235
1236 static void
1237 filt_pipedetach(struct knote *kn)
1238 {
1239         struct pipe *cpipe = (struct pipe *)kn->kn_hook;
1240
1241         SLIST_REMOVE(&cpipe->pipe_sel.si_note, kn, knote, kn_selnext);
1242 }
1243
1244 /*ARGSUSED*/
1245 static int
1246 filt_piperead(struct knote *kn, long hint)
1247 {
1248         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1249         struct pipe *wpipe = rpipe->pipe_peer;
1250
1251         kn->kn_data = rpipe->pipe_buffer.cnt;
1252         if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
1253                 kn->kn_data = rpipe->pipe_map.cnt;
1254
1255         if ((rpipe->pipe_state & PIPE_EOF) ||
1256             (wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1257                 kn->kn_flags |= EV_EOF;
1258                 return (1);
1259         }
1260         return (kn->kn_data > 0);
1261 }
1262
1263 /*ARGSUSED*/
1264 static int
1265 filt_pipewrite(struct knote *kn, long hint)
1266 {
1267         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1268         struct pipe *wpipe = rpipe->pipe_peer;
1269
1270         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1271                 kn->kn_data = 0;
1272                 kn->kn_flags |= EV_EOF;
1273                 return (1);
1274         }
1275         kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1276         if (wpipe->pipe_state & PIPE_DIRECTW)
1277                 kn->kn_data = 0;
1278
1279         return (kn->kn_data >= PIPE_BUF);
1280 }