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  * $FreeBSD$
  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
  29 /*
  30  * This code has two modes of operation, a small write mode and a large
  31  * write mode.  The small write mode acts like conventional pipes with
  32  * a kernel buffer.  If the buffer is less than PIPE_MINDIRECT, then the
  33  * "normal" pipe buffering is done.  If the buffer is between PIPE_MINDIRECT
  34  * and PIPE_SIZE in size, it is fully mapped and wired into the kernel, and
  35  * the receiving process can copy it directly from the pages in the sending
  36  * process.
  37  *
  38  * If the sending process receives a signal, it is possible that it will
  39  * go away, and certainly its address space can change, because control
  40  * is returned back to the user-mode side.  In that case, the pipe code
  41  * arranges to copy the buffer supplied by the user process, to a pageable
  42  * kernel buffer, and the receiving process will grab the data from the
  43  * pageable kernel buffer.  Since signals don't happen all that often,
  44  * the copy operation is normally eliminated.
  45  *
  46  * The constant PIPE_MINDIRECT is chosen to make sure that buffering will
  47  * happen for small transfers so that the system will not spend all of
  48  * its time context switching.  PIPE_SIZE is constrained by the
  49  * amount of kernel virtual memory.
  50  */
  51
  52 #include <sys/param.h>
  53 #include <sys/systm.h>
  54 #include <sys/proc.h>
  55 #include <sys/fcntl.h>
  56 #include <sys/file.h>
  57 #include <sys/filedesc.h>
  58 #include <sys/filio.h>
  59 #include <sys/ttycom.h>
  60 #include <sys/stat.h>
  61 #include <sys/poll.h>
  62 #include <sys/select.h>
  63 #include <sys/signalvar.h>
  64 #include <sys/sysproto.h>
  65 #include <sys/pipe.h>
  66 #include <sys/vnode.h>
  67 #include <sys/uio.h>
  68 #include <sys/event.h>
  69
  70 #include <vm/vm.h>
  71 #include <vm/vm_param.h>
  72 #include <sys/lock.h>
  73 #include <vm/vm_object.h>
  74 #include <vm/vm_kern.h>
  75 #include <vm/vm_extern.h>
  76 #include <vm/pmap.h>
  77 #include <vm/vm_map.h>
  78 #include <vm/vm_page.h>
  79 #include <vm/vm_zone.h>
  80
  81 /*
  82  * Use this define if you want to disable *fancy* VM things.  Expect an
  83  * approx 30% decrease in transfer rate.  This could be useful for
  84  * NetBSD or OpenBSD.
  85  */
  86 /* #define PIPE_NODIRECT */
  87
  88 /*
  89  * interfaces to the outside world
  90  */
  91 static int pipe_read __P((struct file *fp, struct uio *uio,
  92                 struct ucred *cred, int flags, struct proc *p));
  93 static int pipe_write __P((struct file *fp, struct uio *uio,
  94                 struct ucred *cred, int flags, struct proc *p));
  95 static int pipe_close __P((struct file *fp, struct proc *p));
  96 static int pipe_poll __P((struct file *fp, int events, struct ucred *cred,
  97                 struct proc *p));
  98 static int pipe_stat __P((struct file *fp, struct stat *sb, struct proc *p));
  99 static int pipe_ioctl __P((struct file *fp, u_long cmd, caddr_t data, struct proc *p));
 100
 101 static struct fileops pipeops =
 102     { pipe_read, pipe_write, pipe_ioctl, pipe_poll, pipe_stat, pipe_close };
 103
 104 static int      filt_pipeattach(struct knote *kn);
 105 static void     filt_pipedetach(struct knote *kn);
 106 static int      filt_piperead(struct knote *kn, long hint);
 107 static int      filt_pipewrite(struct knote *kn, long hint);
 108
 109 struct filterops pipe_rwfiltops[] = {
 110         { 1, filt_pipeattach, filt_pipedetach, filt_piperead },
 111         { 1, filt_pipeattach, filt_pipedetach, filt_pipewrite },
 112 };
 113
 114 /*
 115  * Default pipe buffer size(s), this can be kind-of large now because pipe
 116  * space is pageable.  The pipe code will try to maintain locality of
 117  * reference for performance reasons, so small amounts of outstanding I/O
 118  * will not wipe the cache.
 119  */
 120 #define MINPIPESIZE (PIPE_SIZE/3)
 121 #define MAXPIPESIZE (2*PIPE_SIZE/3)
 122
 123 /*
 124  * Maximum amount of kva for pipes -- this is kind-of a soft limit, but
 125  * is there so that on large systems, we don't exhaust it.
 126  */
 127 #define MAXPIPEKVA (8*1024*1024)
 128
 129 /*
 130  * Limit for direct transfers, we cannot, of course limit
 131  * the amount of kva for pipes in general though.
 132  */
 133 #define LIMITPIPEKVA (16*1024*1024)
 134
 135 /*
 136  * Limit the number of "big" pipes
 137  */
 138 #define LIMITBIGPIPES   32
 139 static int nbigpipe;
 140
 141 static int amountpipekva;
 142
 143 static void pipeclose __P((struct pipe *cpipe));
 144 static void pipeinit __P((struct pipe *cpipe));
 145 static __inline int pipelock __P((struct pipe *cpipe, int catch));
 146 static __inline void pipeunlock __P((struct pipe *cpipe));
 147 static __inline void pipeselwakeup __P((struct pipe *cpipe));
 148 #ifndef PIPE_NODIRECT
 149 static int pipe_build_write_buffer __P((struct pipe *wpipe, struct uio *uio));
 150 static void pipe_destroy_write_buffer __P((struct pipe *wpipe));
 151 static int pipe_direct_write __P((struct pipe *wpipe, struct uio *uio));
 152 static void pipe_clone_write_buffer __P((struct pipe *wpipe));
 153 #endif
 154 static void pipespace __P((struct pipe *cpipe));
 155
 156 static vm_zone_t pipe_zone;
 157
 158 /*
 159  * The pipe system call for the DTYPE_PIPE type of pipes
 160  */
 161
 162 /* ARGSUSED */
 163 int
 164 pipe(p, uap)
 165         struct proc *p;
 166         struct pipe_args /* {
 167                 int     dummy;
 168         } */ *uap;
 169 {
 170         register struct filedesc *fdp = p->p_fd;
 171         struct file *rf, *wf;
 172         struct pipe *rpipe, *wpipe;
 173         int fd, error;
 174
 175         if (pipe_zone == NULL)
 176                 pipe_zone = zinit("PIPE", sizeof (struct pipe), 0, 0, 4);
 177
 178         rpipe = zalloc( pipe_zone);
 179         pipeinit(rpipe);
 180         rpipe->pipe_state |= PIPE_DIRECTOK;
 181         wpipe = zalloc( pipe_zone);
 182         pipeinit(wpipe);
 183         wpipe->pipe_state |= PIPE_DIRECTOK;
 184
 185         error = falloc(p, &rf, &fd);
 186         if (error)
 187                 goto free2;
 188         p->p_retval[0] = fd;
 189         rf->f_flag = FREAD | FWRITE;
 190         rf->f_type = DTYPE_PIPE;
 191         rf->f_data = (caddr_t)rpipe;
 192         rf->f_ops = &pipeops;
 193         error = falloc(p, &wf, &fd);
 194         if (error)
 195                 goto free3;
 196         wf->f_flag = FREAD | FWRITE;
 197         wf->f_type = DTYPE_PIPE;
 198         wf->f_data = (caddr_t)wpipe;
 199         wf->f_ops = &pipeops;
 200         p->p_retval[1] = fd;
 201
 202         rpipe->pipe_peer = wpipe;
 203         wpipe->pipe_peer = rpipe;
 204
 205         return (0);
 206 free3:
 207         fdp->fd_ofiles[p->p_retval[0]] = 0;
 208         ffree(rf);
 209 free2:
 210         (void)pipeclose(wpipe);
 211         (void)pipeclose(rpipe);
 212         return (error);
 213 }
 214
 215 /*
 216  * Allocate kva for pipe circular buffer, the space is pageable
 217  */
 218 static void
 219 pipespace(cpipe)
 220         struct pipe *cpipe;
 221 {
 222         int npages, error;
 223
 224         npages = round_page(cpipe->pipe_buffer.size)/PAGE_SIZE;
 225         /*
 226          * Create an object, I don't like the idea of paging to/from
 227          * kernel_object.
 228          * XXX -- minor change needed here for NetBSD/OpenBSD VM systems.
 229          */
 230         cpipe->pipe_buffer.object = vm_object_allocate(OBJT_DEFAULT, npages);
 231         cpipe->pipe_buffer.buffer = (caddr_t) vm_map_min(kernel_map);
 232
 233         /*
 234          * Insert the object into the kernel map, and allocate kva for it.
 235          * The map entry is, by default, pageable.
 236          * XXX -- minor change needed here for NetBSD/OpenBSD VM systems.
 237          */
 238         error = vm_map_find(kernel_map, cpipe->pipe_buffer.object, 0,
 239                 (vm_offset_t *) &cpipe->pipe_buffer.buffer,
 240                 cpipe->pipe_buffer.size, 1,
 241                 VM_PROT_ALL, VM_PROT_ALL, 0);
 242
 243         if (error != KERN_SUCCESS)
 244                 panic("pipeinit: cannot allocate pipe -- out of kvm -- code = %d", error);
 245         amountpipekva += cpipe->pipe_buffer.size;
 246 }
 247
 248 /*
 249  * initialize and allocate VM and memory for pipe
 250  */
 251 static void
 252 pipeinit(cpipe)
 253         struct pipe *cpipe;
 254 {
 255
 256         cpipe->pipe_buffer.in = 0;
 257         cpipe->pipe_buffer.out = 0;
 258         cpipe->pipe_buffer.cnt = 0;
 259         cpipe->pipe_buffer.size = PIPE_SIZE;
 260
 261         /* Buffer kva gets dynamically allocated */
 262         cpipe->pipe_buffer.buffer = NULL;
 263         /* cpipe->pipe_buffer.object = invalid */
 264
 265         cpipe->pipe_state = 0;
 266         cpipe->pipe_peer = NULL;
 267         cpipe->pipe_busy = 0;
 268         vfs_timestamp(&cpipe->pipe_ctime);
 269         cpipe->pipe_atime = cpipe->pipe_ctime;
 270         cpipe->pipe_mtime = cpipe->pipe_ctime;
 271         bzero(&cpipe->pipe_sel, sizeof cpipe->pipe_sel);
 272
 273 #ifndef PIPE_NODIRECT
 274         /*
 275          * pipe data structure initializations to support direct pipe I/O
 276          */
 277         cpipe->pipe_map.cnt = 0;
 278         cpipe->pipe_map.kva = 0;
 279         cpipe->pipe_map.pos = 0;
 280         cpipe->pipe_map.npages = 0;
 281         /* cpipe->pipe_map.ms[] = invalid */
 282 #endif
 283 }
 284
 285
 286 /*
 287  * lock a pipe for I/O, blocking other access
 288  */
 289 static __inline int
 290 pipelock(cpipe, catch)
 291         struct pipe *cpipe;
 292         int catch;
 293 {
 294         int error;
 295         while (cpipe->pipe_state & PIPE_LOCK) {
 296                 cpipe->pipe_state |= PIPE_LWANT;
 297                 if ((error = tsleep( cpipe,
 298                         catch?(PRIBIO|PCATCH):PRIBIO, "pipelk", 0)) != 0) {
 299                         return error;
 300                 }
 301         }
 302         cpipe->pipe_state |= PIPE_LOCK;
 303         return 0;
 304 }
 305
 306 /*
 307  * unlock a pipe I/O lock
 308  */
 309 static __inline void
 310 pipeunlock(cpipe)
 311         struct pipe *cpipe;
 312 {
 313         cpipe->pipe_state &= ~PIPE_LOCK;
 314         if (cpipe->pipe_state & PIPE_LWANT) {
 315                 cpipe->pipe_state &= ~PIPE_LWANT;
 316                 wakeup(cpipe);
 317         }
 318 }
 319
 320 static __inline void
 321 pipeselwakeup(cpipe)
 322         struct pipe *cpipe;
 323 {
 324         if (cpipe->pipe_state & PIPE_SEL) {
 325                 cpipe->pipe_state &= ~PIPE_SEL;
 326                 selwakeup(&cpipe->pipe_sel);
 327         }
 328         if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
 329                 pgsigio(cpipe->pipe_sigio, SIGIO, 0);
 330         KNOTE(&cpipe->pipe_sel.si_note, 0);
 331 }
 332
 333 /* ARGSUSED */
 334 static int
 335 pipe_read(fp, uio, cred, flags, p)
 336         struct file *fp;
 337         struct uio *uio;
 338         struct ucred *cred;
 339         struct proc *p;
 340         int flags;
 341 {
 342
 343         struct pipe *rpipe = (struct pipe *) fp->f_data;
 344         int error;
 345         int nread = 0;
 346         u_int size;
 347
 348         ++rpipe->pipe_busy;
 349         error = pipelock(rpipe, 1);
 350         if (error)
 351                 goto unlocked_error;
 352
 353         while (uio->uio_resid) {
 354                 /*
 355                  * normal pipe buffer receive
 356                  */
 357                 if (rpipe->pipe_buffer.cnt > 0) {
 358                         size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
 359                         if (size > rpipe->pipe_buffer.cnt)
 360                                 size = rpipe->pipe_buffer.cnt;
 361                         if (size > (u_int) uio->uio_resid)
 362                                 size = (u_int) uio->uio_resid;
 363
 364                         error = uiomove(&rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
 365                                         size, uio);
 366                         if (error) {
 367                                 break;
 368                         }
 369                         rpipe->pipe_buffer.out += size;
 370                         if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
 371                                 rpipe->pipe_buffer.out = 0;
 372
 373                         rpipe->pipe_buffer.cnt -= size;
 374
 375                         /*
 376                          * If there is no more to read in the pipe, reset
 377                          * its pointers to the beginning.  This improves
 378                          * cache hit stats.
 379                          */
 380                         if (rpipe->pipe_buffer.cnt == 0) {
 381                                 rpipe->pipe_buffer.in = 0;
 382                                 rpipe->pipe_buffer.out = 0;
 383                         }
 384                         nread += size;
 385 #ifndef PIPE_NODIRECT
 386                 /*
 387                  * Direct copy, bypassing a kernel buffer.
 388                  */
 389                 } else if ((size = rpipe->pipe_map.cnt) &&
 390                            (rpipe->pipe_state & PIPE_DIRECTW)) {
 391                         caddr_t va;
 392                         if (size > (u_int) uio->uio_resid)
 393                                 size = (u_int) uio->uio_resid;
 394
 395                         va = (caddr_t) rpipe->pipe_map.kva + rpipe->pipe_map.pos;
 396                         error = uiomove(va, size, uio);
 397                         if (error)
 398                                 break;
 399                         nread += size;
 400                         rpipe->pipe_map.pos += size;
 401                         rpipe->pipe_map.cnt -= size;
 402                         if (rpipe->pipe_map.cnt == 0) {
 403                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
 404                                 wakeup(rpipe);
 405                         }
 406 #endif
 407                 } else {
 408                         /*
 409                          * detect EOF condition
 410                          */
 411                         if (rpipe->pipe_state & PIPE_EOF) {
 412                                 /* XXX error = ? */
 413                                 break;
 414                         }
 415
 416                         /*
 417                          * If the "write-side" has been blocked, wake it up now.
 418                          */
 419                         if (rpipe->pipe_state & PIPE_WANTW) {
 420                                 rpipe->pipe_state &= ~PIPE_WANTW;
 421                                 wakeup(rpipe);
 422                         }
 423
 424                         /*
 425                          * Break if some data was read.
 426                          */
 427                         if (nread > 0)
 428                                 break;
 429
 430                         /*
 431                          * Unlock the pipe buffer for our remaining processing.  We
 432                          * will either break out with an error or we will sleep and
 433                          * relock to loop.
 434                          */
 435                         pipeunlock(rpipe);
 436
 437                         /*
 438                          * Handle non-blocking mode operation or
 439                          * wait for more data.
 440                          */
 441                         if (fp->f_flag & FNONBLOCK)
 442                                 error = EAGAIN;
 443                         else {
 444                                 rpipe->pipe_state |= PIPE_WANTR;
 445                                 if ((error = tsleep(rpipe, PRIBIO|PCATCH, "piperd", 0)) == 0)
 446                                         error = pipelock(rpipe, 1);
 447                         }
 448                         if (error)
 449                                 goto unlocked_error;
 450                 }
 451         }
 452         pipeunlock(rpipe);
 453
 454         if (error == 0)
 455                 vfs_timestamp(&rpipe->pipe_atime);
 456 unlocked_error:
 457         --rpipe->pipe_busy;
 458
 459         /*
 460          * PIPE_WANT processing only makes sense if pipe_busy is 0.
 461          */
 462         if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
 463                 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
 464                 wakeup(rpipe);
 465         } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
 466                 /*
 467                  * Handle write blocking hysteresis.
 468                  */
 469                 if (rpipe->pipe_state & PIPE_WANTW) {
 470                         rpipe->pipe_state &= ~PIPE_WANTW;
 471                         wakeup(rpipe);
 472                 }
 473         }
 474
 475         if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
 476                 pipeselwakeup(rpipe);
 477
 478         return error;
 479 }
 480
 481 #ifndef PIPE_NODIRECT
 482 /*
 483  * Map the sending processes' buffer into kernel space and wire it.
 484  * This is similar to a physical write operation.
 485  */
 486 static int
 487 pipe_build_write_buffer(wpipe, uio)
 488         struct pipe *wpipe;
 489         struct uio *uio;
 490 {
 491         u_int size;
 492         int i;
 493         vm_offset_t addr, endaddr, paddr;
 494
 495         size = (u_int) uio->uio_iov->iov_len;
 496         if (size > wpipe->pipe_buffer.size)
 497                 size = wpipe->pipe_buffer.size;
 498
 499         endaddr = round_page((vm_offset_t)uio->uio_iov->iov_base + size);
 500         for(i = 0, addr = trunc_page((vm_offset_t)uio->uio_iov->iov_base);
 501                 addr < endaddr;
 502                 addr += PAGE_SIZE, i+=1) {
 503
 504                 vm_page_t m;
 505
 506                 if (vm_fault_quick((caddr_t)addr, VM_PROT_READ) < 0 ||
 507                     (paddr = pmap_kextract(addr)) == 0) {
 508                         int j;
 509                         for(j=0;j<i;j++)
 510                                 vm_page_unwire(wpipe->pipe_map.ms[j], 1);
 511                         return EFAULT;
 512                 }
 513
 514                 m = PHYS_TO_VM_PAGE(paddr);
 515                 vm_page_wire(m);
 516                 wpipe->pipe_map.ms[i] = m;
 517         }
 518
 519 /*
 520  * set up the control block
 521  */
 522         wpipe->pipe_map.npages = i;
 523         wpipe->pipe_map.pos = ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
 524         wpipe->pipe_map.cnt = size;
 525
 526 /*
 527  * and map the buffer
 528  */
 529         if (wpipe->pipe_map.kva == 0) {
 530                 /*
 531                  * We need to allocate space for an extra page because the
 532                  * address range might (will) span pages at times.
 533                  */
 534                 wpipe->pipe_map.kva = kmem_alloc_pageable(kernel_map,
 535                         wpipe->pipe_buffer.size + PAGE_SIZE);
 536                 amountpipekva += wpipe->pipe_buffer.size + PAGE_SIZE;
 537         }
 538         pmap_qenter(wpipe->pipe_map.kva, wpipe->pipe_map.ms,
 539                 wpipe->pipe_map.npages);
 540
 541 /*
 542  * and update the uio data
 543  */
 544
 545         uio->uio_iov->iov_len -= size;
 546         uio->uio_iov->iov_base += size;
 547         if (uio->uio_iov->iov_len == 0)
 548                 uio->uio_iov++;
 549         uio->uio_resid -= size;
 550         uio->uio_offset += size;
 551         return 0;
 552 }
 553
 554 /*
 555  * unmap and unwire the process buffer
 556  */
 557 static void
 558 pipe_destroy_write_buffer(wpipe)
 559 struct pipe *wpipe;
 560 {
 561         int i;
 562         if (wpipe->pipe_map.kva) {
 563                 pmap_qremove(wpipe->pipe_map.kva, wpipe->pipe_map.npages);
 564
 565                 if (amountpipekva > MAXPIPEKVA) {
 566                         vm_offset_t kva = wpipe->pipe_map.kva;
 567                         wpipe->pipe_map.kva = 0;
 568                         kmem_free(kernel_map, kva,
 569                                 wpipe->pipe_buffer.size + PAGE_SIZE);
 570                         amountpipekva -= wpipe->pipe_buffer.size + PAGE_SIZE;
 571                 }
 572         }
 573         for (i=0;i<wpipe->pipe_map.npages;i++)
 574                 vm_page_unwire(wpipe->pipe_map.ms[i], 1);
 575 }
 576
 577 /*
 578  * In the case of a signal, the writing process might go away.  This
 579  * code copies the data into the circular buffer so that the source
 580  * pages can be freed without loss of data.
 581  */
 582 static void
 583 pipe_clone_write_buffer(wpipe)
 584 struct pipe *wpipe;
 585 {
 586         int size;
 587         int pos;
 588
 589         size = wpipe->pipe_map.cnt;
 590         pos = wpipe->pipe_map.pos;
 591         bcopy((caddr_t) wpipe->pipe_map.kva+pos,
 592                         (caddr_t) wpipe->pipe_buffer.buffer,
 593                         size);
 594
 595         wpipe->pipe_buffer.in = size;
 596         wpipe->pipe_buffer.out = 0;
 597         wpipe->pipe_buffer.cnt = size;
 598         wpipe->pipe_state &= ~PIPE_DIRECTW;
 599
 600         pipe_destroy_write_buffer(wpipe);
 601 }
 602
 603 /*
 604  * This implements the pipe buffer write mechanism.  Note that only
 605  * a direct write OR a normal pipe write can be pending at any given time.
 606  * If there are any characters in the pipe buffer, the direct write will
 607  * be deferred until the receiving process grabs all of the bytes from
 608  * the pipe buffer.  Then the direct mapping write is set-up.
 609  */
 610 static int
 611 pipe_direct_write(wpipe, uio)
 612         struct pipe *wpipe;
 613         struct uio *uio;
 614 {
 615         int error;
 616 retry:
 617         while (wpipe->pipe_state & PIPE_DIRECTW) {
 618                 if ( wpipe->pipe_state & PIPE_WANTR) {
 619                         wpipe->pipe_state &= ~PIPE_WANTR;
 620                         wakeup(wpipe);
 621                 }
 622                 wpipe->pipe_state |= PIPE_WANTW;
 623                 error = tsleep(wpipe,
 624                                 PRIBIO|PCATCH, "pipdww", 0);
 625                 if (error)
 626                         goto error1;
 627                 if (wpipe->pipe_state & PIPE_EOF) {
 628                         error = EPIPE;
 629                         goto error1;
 630                 }
 631         }
 632         wpipe->pipe_map.cnt = 0;        /* transfer not ready yet */
 633         if (wpipe->pipe_buffer.cnt > 0) {
 634                 if ( wpipe->pipe_state & PIPE_WANTR) {
 635                         wpipe->pipe_state &= ~PIPE_WANTR;
 636                         wakeup(wpipe);
 637                 }
 638
 639                 wpipe->pipe_state |= PIPE_WANTW;
 640                 error = tsleep(wpipe,
 641                                 PRIBIO|PCATCH, "pipdwc", 0);
 642                 if (error)
 643                         goto error1;
 644                 if (wpipe->pipe_state & PIPE_EOF) {
 645                         error = EPIPE;
 646                         goto error1;
 647                 }
 648                 goto retry;
 649         }
 650
 651         wpipe->pipe_state |= PIPE_DIRECTW;
 652
 653         error = pipe_build_write_buffer(wpipe, uio);
 654         if (error) {
 655                 wpipe->pipe_state &= ~PIPE_DIRECTW;
 656                 goto error1;
 657         }
 658
 659         error = 0;
 660         while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
 661                 if (wpipe->pipe_state & PIPE_EOF) {
 662                         pipelock(wpipe, 0);
 663                         pipe_destroy_write_buffer(wpipe);
 664                         pipeunlock(wpipe);
 665                         pipeselwakeup(wpipe);
 666                         error = EPIPE;
 667                         goto error1;
 668                 }
 669                 if (wpipe->pipe_state & PIPE_WANTR) {
 670                         wpipe->pipe_state &= ~PIPE_WANTR;
 671                         wakeup(wpipe);
 672                 }
 673                 pipeselwakeup(wpipe);
 674                 error = tsleep(wpipe, PRIBIO|PCATCH, "pipdwt", 0);
 675         }
 676
 677         pipelock(wpipe,0);
 678         if (wpipe->pipe_state & PIPE_DIRECTW) {
 679                 /*
 680                  * this bit of trickery substitutes a kernel buffer for
 681                  * the process that might be going away.
 682                  */
 683                 pipe_clone_write_buffer(wpipe);
 684         } else {
 685                 pipe_destroy_write_buffer(wpipe);
 686         }
 687         pipeunlock(wpipe);
 688         return error;
 689
 690 error1:
 691         wakeup(wpipe);
 692         return error;
 693 }
 694 #endif
 695
 696 static int
 697 pipe_write(fp, uio, cred, flags, p)
 698         struct file *fp;
 699         struct uio *uio;
 700         struct ucred *cred;
 701         struct proc *p;
 702         int flags;
 703 {
 704         int error = 0;
 705         int orig_resid;
 706
 707         struct pipe *wpipe, *rpipe;
 708
 709         rpipe = (struct pipe *) fp->f_data;
 710         wpipe = rpipe->pipe_peer;
 711
 712         /*
 713          * detect loss of pipe read side, issue SIGPIPE if lost.
 714          */
 715         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
 716                 return EPIPE;
 717         }
 718
 719         /*
 720          * If it is advantageous to resize the pipe buffer, do
 721          * so.
 722          */
 723         if ((uio->uio_resid > PIPE_SIZE) &&
 724                 (nbigpipe < LIMITBIGPIPES) &&
 725                 (wpipe->pipe_state & PIPE_DIRECTW) == 0 &&
 726                 (wpipe->pipe_buffer.size <= PIPE_SIZE) &&
 727                 (wpipe->pipe_buffer.cnt == 0)) {
 728
 729                 if (wpipe->pipe_buffer.buffer) {
 730                         amountpipekva -= wpipe->pipe_buffer.size;
 731                         kmem_free(kernel_map,
 732                                 (vm_offset_t)wpipe->pipe_buffer.buffer,
 733                                 wpipe->pipe_buffer.size);
 734                 }
 735
 736 #ifndef PIPE_NODIRECT
 737                 if (wpipe->pipe_map.kva) {
 738                         amountpipekva -= wpipe->pipe_buffer.size + PAGE_SIZE;
 739                         kmem_free(kernel_map,
 740                                 wpipe->pipe_map.kva,
 741                                 wpipe->pipe_buffer.size + PAGE_SIZE);
 742                 }
 743 #endif
 744
 745                 wpipe->pipe_buffer.in = 0;
 746                 wpipe->pipe_buffer.out = 0;
 747                 wpipe->pipe_buffer.cnt = 0;
 748                 wpipe->pipe_buffer.size = BIG_PIPE_SIZE;
 749                 wpipe->pipe_buffer.buffer = NULL;
 750                 ++nbigpipe;
 751
 752 #ifndef PIPE_NODIRECT
 753                 wpipe->pipe_map.cnt = 0;
 754                 wpipe->pipe_map.kva = 0;
 755                 wpipe->pipe_map.pos = 0;
 756                 wpipe->pipe_map.npages = 0;
 757 #endif
 758
 759         }
 760
 761
 762         if( wpipe->pipe_buffer.buffer == NULL) {
 763                 if ((error = pipelock(wpipe,1)) == 0) {
 764                         pipespace(wpipe);
 765                         pipeunlock(wpipe);
 766                 } else {
 767                         return error;
 768                 }
 769         }
 770
 771         ++wpipe->pipe_busy;
 772         orig_resid = uio->uio_resid;
 773         while (uio->uio_resid) {
 774                 int space;
 775 #ifndef PIPE_NODIRECT
 776                 /*
 777                  * If the transfer is large, we can gain performance if
 778                  * we do process-to-process copies directly.
 779                  * If the write is non-blocking, we don't use the
 780                  * direct write mechanism.
 781                  *
 782                  * The direct write mechanism will detect the reader going
 783                  * away on us.
 784                  */
 785                 if ((uio->uio_iov->iov_len >= PIPE_MINDIRECT) &&
 786                     (fp->f_flag & FNONBLOCK) == 0 &&
 787                         (wpipe->pipe_map.kva || (amountpipekva < LIMITPIPEKVA)) &&
 788                         (uio->uio_iov->iov_len >= PIPE_MINDIRECT)) {
 789                         error = pipe_direct_write( wpipe, uio);
 790                         if (error) {
 791                                 break;
 792                         }
 793                         continue;
 794                 }
 795 #endif
 796
 797                 /*
 798                  * Pipe buffered writes cannot be coincidental with
 799                  * direct writes.  We wait until the currently executing
 800                  * direct write is completed before we start filling the
 801                  * pipe buffer.  We break out if a signal occurs or the
 802                  * reader goes away.
 803                  */
 804         retrywrite:
 805                 while (wpipe->pipe_state & PIPE_DIRECTW) {
 806                         if (wpipe->pipe_state & PIPE_WANTR) {
 807                                 wpipe->pipe_state &= ~PIPE_WANTR;
 808                                 wakeup(wpipe);
 809                         }
 810                         error = tsleep(wpipe, PRIBIO|PCATCH, "pipbww", 0);
 811                         if (wpipe->pipe_state & PIPE_EOF)
 812                                 break;
 813                         if (error)
 814                                 break;
 815                 }
 816                 if (wpipe->pipe_state & PIPE_EOF) {
 817                         error = EPIPE;
 818                         break;
 819                 }
 820
 821                 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
 822
 823                 /* Writes of size <= PIPE_BUF must be atomic. */
 824                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
 825                         space = 0;
 826
 827                 if (space > 0 && (wpipe->pipe_buffer.cnt < PIPE_SIZE)) {
 828                         if ((error = pipelock(wpipe,1)) == 0) {
 829                                 int size;       /* Transfer size */
 830                                 int segsize;    /* first segment to transfer */
 831                                 /*
 832                                  * It is possible for a direct write to
 833                                  * slip in on us... handle it here...
 834                                  */
 835                                 if (wpipe->pipe_state & PIPE_DIRECTW) {
 836                                         pipeunlock(wpipe);
 837                                         goto retrywrite;
 838                                 }
 839                                 /*
 840                                  * If a process blocked in uiomove, our
 841                                  * value for space might be bad.
 842                                  *
 843                                  * XXX will we be ok if the reader has gone
 844                                  * away here?
 845                                  */
 846                                 if (space > wpipe->pipe_buffer.size -
 847                                     wpipe->pipe_buffer.cnt) {
 848                                         pipeunlock(wpipe);
 849                                         goto retrywrite;
 850                                 }
 851
 852                                 /*
 853                                  * Transfer size is minimum of uio transfer
 854                                  * and free space in pipe buffer.
 855                                  */
 856                                 if (space > uio->uio_resid)
 857                                         size = uio->uio_resid;
 858                                 else
 859                                         size = space;
 860                                 /*
 861                                  * First segment to transfer is minimum of
 862                                  * transfer size and contiguous space in
 863                                  * pipe buffer.  If first segment to transfer
 864                                  * is less than the transfer size, we've got
 865                                  * a wraparound in the buffer.
 866                                  */
 867                                 segsize = wpipe->pipe_buffer.size -
 868                                         wpipe->pipe_buffer.in;
 869                                 if (segsize > size)
 870                                         segsize = size;
 871
 872                                 /* Transfer first segment */
 873
 874                                 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
 875                                                 segsize, uio);
 876
 877                                 if (error == 0 && segsize < size) {
 878                                         /*
 879                                          * Transfer remaining part now, to
 880                                          * support atomic writes.  Wraparound
 881                                          * happened.
 882                                          */
 883                                         if (wpipe->pipe_buffer.in + segsize !=
 884                                             wpipe->pipe_buffer.size)
 885                                                 panic("Expected pipe buffer wraparound disappeared");
 886
 887                                         error = uiomove(&wpipe->pipe_buffer.buffer[0],
 888                                                         size - segsize, uio);
 889                                 }
 890                                 if (error == 0) {
 891                                         wpipe->pipe_buffer.in += size;
 892                                         if (wpipe->pipe_buffer.in >=
 893                                             wpipe->pipe_buffer.size) {
 894                                                 if (wpipe->pipe_buffer.in != size - segsize + wpipe->pipe_buffer.size)
 895                                                         panic("Expected wraparound bad");
 896                                                 wpipe->pipe_buffer.in = size - segsize;
 897                                         }
 898
 899                                         wpipe->pipe_buffer.cnt += size;
 900                                         if (wpipe->pipe_buffer.cnt > wpipe->pipe_buffer.size)
 901                                                 panic("Pipe buffer overflow");
 902
 903                                 }
 904                                 pipeunlock(wpipe);
 905                         }
 906                         if (error)
 907                                 break;
 908
 909                 } else {
 910                         /*
 911                          * If the "read-side" has been blocked, wake it up now.
 912                          */
 913                         if (wpipe->pipe_state & PIPE_WANTR) {
 914                                 wpipe->pipe_state &= ~PIPE_WANTR;
 915                                 wakeup(wpipe);
 916                         }
 917
 918                         /*
 919                          * don't block on non-blocking I/O
 920                          */
 921                         if (fp->f_flag & FNONBLOCK) {
 922                                 error = EAGAIN;
 923                                 break;
 924                         }
 925
 926                         /*
 927                          * We have no more space and have something to offer,
 928                          * wake up select/poll.
 929                          */
 930                         pipeselwakeup(wpipe);
 931
 932                         wpipe->pipe_state |= PIPE_WANTW;
 933                         if ((error = tsleep(wpipe, (PRIBIO+1)|PCATCH, "pipewr", 0)) != 0) {
 934                                 break;
 935                         }
 936                         /*
 937                          * If read side wants to go away, we just issue a signal
 938                          * to ourselves.
 939                          */
 940                         if (wpipe->pipe_state & PIPE_EOF) {
 941                                 error = EPIPE;
 942                                 break;
 943                         }
 944                 }
 945         }
 946
 947         --wpipe->pipe_busy;
 948         if ((wpipe->pipe_busy == 0) &&
 949                 (wpipe->pipe_state & PIPE_WANT)) {
 950                 wpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTR);
 951                 wakeup(wpipe);
 952         } else if (wpipe->pipe_buffer.cnt > 0) {
 953                 /*
 954                  * If we have put any characters in the buffer, we wake up
 955                  * the reader.
 956                  */
 957                 if (wpipe->pipe_state & PIPE_WANTR) {
 958                         wpipe->pipe_state &= ~PIPE_WANTR;
 959                         wakeup(wpipe);
 960                 }
 961         }
 962
 963         /*
 964          * Don't return EPIPE if I/O was successful
 965          */
 966         if ((wpipe->pipe_buffer.cnt == 0) &&
 967                 (uio->uio_resid == 0) &&
 968                 (error == EPIPE))
 969                 error = 0;
 970
 971         if (error == 0)
 972                 vfs_timestamp(&wpipe->pipe_mtime);
 973
 974         /*
 975          * We have something to offer,
 976          * wake up select/poll.
 977          */
 978         if (wpipe->pipe_buffer.cnt)
 979                 pipeselwakeup(wpipe);
 980
 981         return error;
 982 }
 983
 984 /*
 985  * we implement a very minimal set of ioctls for compatibility with sockets.
 986  */
 987 int
 988 pipe_ioctl(fp, cmd, data, p)
 989         struct file *fp;
 990         u_long cmd;
 991         register caddr_t data;
 992         struct proc *p;
 993 {
 994         register struct pipe *mpipe = (struct pipe *)fp->f_data;
 995
 996         switch (cmd) {
 997
 998         case FIONBIO:
 999                 return (0);
1000
1001         case FIOASYNC:
1002                 if (*(int *)data) {
1003                         mpipe->pipe_state |= PIPE_ASYNC;
1004                 } else {
1005                         mpipe->pipe_state &= ~PIPE_ASYNC;
1006                 }
1007                 return (0);
1008
1009         case FIONREAD:
1010                 if (mpipe->pipe_state & PIPE_DIRECTW)
1011                         *(int *)data = mpipe->pipe_map.cnt;
1012                 else
1013                         *(int *)data = mpipe->pipe_buffer.cnt;
1014                 return (0);
1015
1016         case FIOSETOWN:
1017                 return (fsetown(*(int *)data, &mpipe->pipe_sigio));
1018
1019         case FIOGETOWN:
1020                 *(int *)data = fgetown(mpipe->pipe_sigio);
1021                 return (0);
1022
1023         /* This is deprecated, FIOSETOWN should be used instead. */
1024         case TIOCSPGRP:
1025                 return (fsetown(-(*(int *)data), &mpipe->pipe_sigio));
1026
1027         /* This is deprecated, FIOGETOWN should be used instead. */
1028         case TIOCGPGRP:
1029                 *(int *)data = -fgetown(mpipe->pipe_sigio);
1030                 return (0);
1031
1032         }
1033         return (ENOTTY);
1034 }
1035
1036 int
1037 pipe_poll(fp, events, cred, p)
1038         struct file *fp;
1039         int events;
1040         struct ucred *cred;
1041         struct proc *p;
1042 {
1043         register struct pipe *rpipe = (struct pipe *)fp->f_data;
1044         struct pipe *wpipe;
1045         int revents = 0;
1046
1047         wpipe = rpipe->pipe_peer;
1048         if (events & (POLLIN | POLLRDNORM))
1049                 if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1050                     (rpipe->pipe_buffer.cnt > 0) ||
1051                     (rpipe->pipe_state & PIPE_EOF))
1052                         revents |= events & (POLLIN | POLLRDNORM);
1053
1054         if (events & (POLLOUT | POLLWRNORM))
1055                 if (wpipe == NULL || (wpipe->pipe_state & PIPE_EOF) ||
1056                     (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1057                      (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF))
1058                         revents |= events & (POLLOUT | POLLWRNORM);
1059
1060         if ((rpipe->pipe_state & PIPE_EOF) ||
1061             (wpipe == NULL) ||
1062             (wpipe->pipe_state & PIPE_EOF))
1063                 revents |= POLLHUP;
1064
1065         if (revents == 0) {
1066                 if (events & (POLLIN | POLLRDNORM)) {
1067                         selrecord(p, &rpipe->pipe_sel);
1068                         rpipe->pipe_state |= PIPE_SEL;
1069                 }
1070
1071                 if (events & (POLLOUT | POLLWRNORM)) {
1072                         selrecord(p, &wpipe->pipe_sel);
1073                         wpipe->pipe_state |= PIPE_SEL;
1074                 }
1075         }
1076
1077         return (revents);
1078 }
1079
1080 static int
1081 pipe_stat(fp, ub, p)
1082         struct file *fp;
1083         struct stat *ub;
1084         struct proc *p;
1085 {
1086         struct pipe *pipe = (struct pipe *)fp->f_data;
1087
1088         bzero((caddr_t)ub, sizeof (*ub));
1089         ub->st_mode = S_IFIFO;
1090         ub->st_blksize = pipe->pipe_buffer.size;
1091         ub->st_size = pipe->pipe_buffer.cnt;
1092         ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1093         ub->st_atimespec = pipe->pipe_atime;
1094         ub->st_mtimespec = pipe->pipe_mtime;
1095         ub->st_ctimespec = pipe->pipe_ctime;
1096         ub->st_uid = fp->f_cred->cr_uid;
1097         ub->st_gid = fp->f_cred->cr_gid;
1098         /*
1099          * Left as 0: st_dev, st_ino, st_nlink, st_rdev, st_flags, st_gen.
1100          * XXX (st_dev, st_ino) should be unique.
1101          */
1102         return 0;
1103 }
1104
1105 /* ARGSUSED */
1106 static int
1107 pipe_close(fp, p)
1108         struct file *fp;
1109         struct proc *p;
1110 {
1111         struct pipe *cpipe = (struct pipe *)fp->f_data;
1112
1113         fp->f_ops = &badfileops;
1114         fp->f_data = NULL;
1115         funsetown(cpipe->pipe_sigio);
1116         pipeclose(cpipe);
1117         return 0;
1118 }
1119
1120 /*
1121  * shutdown the pipe
1122  */
1123 static void
1124 pipeclose(cpipe)
1125         struct pipe *cpipe;
1126 {
1127         struct pipe *ppipe;
1128         if (cpipe) {
1129
1130                 pipeselwakeup(cpipe);
1131
1132                 /*
1133                  * If the other side is blocked, wake it up saying that
1134                  * we want to close it down.
1135                  */
1136                 while (cpipe->pipe_busy) {
1137                         wakeup(cpipe);
1138                         cpipe->pipe_state |= PIPE_WANT|PIPE_EOF;
1139                         tsleep(cpipe, PRIBIO, "pipecl", 0);
1140                 }
1141
1142                 /*
1143                  * Disconnect from peer
1144                  */
1145                 if ((ppipe = cpipe->pipe_peer) != NULL) {
1146                         pipeselwakeup(ppipe);
1147
1148                         ppipe->pipe_state |= PIPE_EOF;
1149                         wakeup(ppipe);
1150                         ppipe->pipe_peer = NULL;
1151                 }
1152
1153                 /*
1154                  * free resources
1155                  */
1156                 if (cpipe->pipe_buffer.buffer) {
1157                         if (cpipe->pipe_buffer.size > PIPE_SIZE)
1158                                 --nbigpipe;
1159                         amountpipekva -= cpipe->pipe_buffer.size;
1160                         kmem_free(kernel_map,
1161                                 (vm_offset_t)cpipe->pipe_buffer.buffer,
1162                                 cpipe->pipe_buffer.size);
1163                 }
1164 #ifndef PIPE_NODIRECT
1165                 if (cpipe->pipe_map.kva) {
1166                         amountpipekva -= cpipe->pipe_buffer.size + PAGE_SIZE;
1167                         kmem_free(kernel_map,
1168                                 cpipe->pipe_map.kva,
1169                                 cpipe->pipe_buffer.size + PAGE_SIZE);
1170                 }
1171 #endif
1172                 zfree(pipe_zone, cpipe);
1173         }
1174 }
1175
1176 static int
1177 filt_pipeattach(struct knote *kn)
1178 {
1179         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1180
1181         SLIST_INSERT_HEAD(&rpipe->pipe_sel.si_note, kn, kn_selnext);
1182         return (0);
1183 }
1184
1185 static void
1186 filt_pipedetach(struct knote *kn)
1187 {
1188         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1189
1190         SLIST_REMOVE(&rpipe->pipe_sel.si_note, kn, knote, kn_selnext);
1191 }
1192
1193 /*ARGSUSED*/
1194 static int
1195 filt_piperead(struct knote *kn, long hint)
1196 {
1197         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1198         struct pipe *wpipe = rpipe->pipe_peer;
1199
1200         kn->kn_data = rpipe->pipe_buffer.cnt;
1201         if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
1202                 kn->kn_data = rpipe->pipe_map.cnt;
1203
1204         if ((rpipe->pipe_state & PIPE_EOF) ||
1205             (wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1206                 kn->kn_flags |= EV_EOF;
1207                 return (1);
1208         }
1209         return (kn->kn_data > 0);
1210 }
1211
1212 /*ARGSUSED*/
1213 static int
1214 filt_pipewrite(struct knote *kn, long hint)
1215 {
1216         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1217         struct pipe *wpipe = rpipe->pipe_peer;
1218
1219         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1220                 kn->kn_data = 0;
1221                 kn->kn_flags |= EV_EOF;
1222                 return (1);
1223         }
1224         kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1225         if ((wpipe->pipe_state & PIPE_DIRECTW) == 0)
1226                 kn->kn_data = 0;
1227
1228         return (kn->kn_data >= PIPE_BUF);
1229 }