The Design and Implementation of the FreeBSD Operating System, Second Edition
Now available: The Design and Implementation of the FreeBSD Operating System (Second Edition)


[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]

FreeBSD/Linux Kernel Cross Reference
sys/kern/sys_pipe.c

Version: -  FREEBSD  -  FREEBSD-12-STABLE  -  FREEBSD-12-0  -  FREEBSD-11-STABLE  -  FREEBSD-11-2  -  FREEBSD-11-1  -  FREEBSD-11-0  -  FREEBSD-10-STABLE  -  FREEBSD-10-4  -  FREEBSD-10-3  -  FREEBSD-10-2  -  FREEBSD-10-1  -  FREEBSD-10-0  -  FREEBSD-9-STABLE  -  FREEBSD-9-3  -  FREEBSD-9-2  -  FREEBSD-9-1  -  FREEBSD-9-0  -  FREEBSD-8-STABLE  -  FREEBSD-8-4  -  FREEBSD-8-3  -  FREEBSD-8-2  -  FREEBSD-8-1  -  FREEBSD-8-0  -  FREEBSD-7-STABLE  -  FREEBSD-7-4  -  FREEBSD-7-3  -  FREEBSD-7-2  -  FREEBSD-7-1  -  FREEBSD-7-0  -  FREEBSD-6-STABLE  -  FREEBSD-6-4  -  FREEBSD-6-3  -  FREEBSD-6-2  -  FREEBSD-6-1  -  FREEBSD-6-0  -  FREEBSD-5-STABLE  -  FREEBSD-5-5  -  FREEBSD-5-4  -  FREEBSD-5-3  -  FREEBSD-5-2  -  FREEBSD-5-1  -  FREEBSD-5-0  -  FREEBSD-4-STABLE  -  FREEBSD-3-STABLE  -  FREEBSD22  -  linux-2.6  -  linux-2.4.22  -  MK83  -  MK84  -  PLAN9  -  DFBSD  -  NETBSD  -  NETBSD5  -  NETBSD4  -  NETBSD3  -  NETBSD20  -  OPENBSD  -  xnu-517  -  xnu-792  -  xnu-792.6.70  -  xnu-1228  -  xnu-1456.1.26  -  xnu-1699.24.8  -  xnu-2050.18.24  -  OPENSOLARIS  -  minix-3-1-1 
SearchContext: -  none  -  3  -  10 

    1 /*      $NetBSD: sys_pipe.c,v 1.130 2011/04/10 15:45:33 christos Exp $  */
    2 
    3 /*-
    4  * Copyright (c) 2003, 2007, 2008, 2009 The NetBSD Foundation, Inc.
    5  * All rights reserved.
    6  *
    7  * This code is derived from software contributed to The NetBSD Foundation
    8  * by Paul Kranenburg, and by Andrew Doran.
    9  *
   10  * Redistribution and use in source and binary forms, with or without
   11  * modification, are permitted provided that the following conditions
   12  * are met:
   13  * 1. Redistributions of source code must retain the above copyright
   14  *    notice, this list of conditions and the following disclaimer.
   15  * 2. Redistributions in binary form must reproduce the above copyright
   16  *    notice, this list of conditions and the following disclaimer in the
   17  *    documentation and/or other materials provided with the distribution.
   18  *
   19  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
   20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
   21  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
   22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
   23  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
   24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
   25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
   26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
   27  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
   28  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
   29  * POSSIBILITY OF SUCH DAMAGE.
   30  */
   31 
   32 /*
   33  * Copyright (c) 1996 John S. Dyson
   34  * All rights reserved.
   35  *
   36  * Redistribution and use in source and binary forms, with or without
   37  * modification, are permitted provided that the following conditions
   38  * are met:
   39  * 1. Redistributions of source code must retain the above copyright
   40  *    notice immediately at the beginning of the file, without modification,
   41  *    this list of conditions, and the following disclaimer.
   42  * 2. Redistributions in binary form must reproduce the above copyright
   43  *    notice, this list of conditions and the following disclaimer in the
   44  *    documentation and/or other materials provided with the distribution.
   45  * 3. Absolutely no warranty of function or purpose is made by the author
   46  *    John S. Dyson.
   47  * 4. Modifications may be freely made to this file if the above conditions
   48  *    are met.
   49  */
   50 
   51 /*
   52  * This file contains a high-performance replacement for the socket-based
   53  * pipes scheme originally used.  It does not support all features of
   54  * sockets, but does do everything that pipes normally do.
   55  *
   56  * This code has two modes of operation, a small write mode and a large
   57  * write mode.  The small write mode acts like conventional pipes with
   58  * a kernel buffer.  If the buffer is less than PIPE_MINDIRECT, then the
   59  * "normal" pipe buffering is done.  If the buffer is between PIPE_MINDIRECT
   60  * and PIPE_SIZE in size it is mapped read-only into the kernel address space
   61  * using the UVM page loan facility from where the receiving process can copy
   62  * the data directly from the pages in the sending process.
   63  *
   64  * The constant PIPE_MINDIRECT is chosen to make sure that buffering will
   65  * happen for small transfers so that the system will not spend all of
   66  * its time context switching.  PIPE_SIZE is constrained by the
   67  * amount of kernel virtual memory.
   68  */
   69 
   70 #include <sys/cdefs.h>
   71 __KERNEL_RCSID(0, "$NetBSD: sys_pipe.c,v 1.130 2011/04/10 15:45:33 christos Exp $");
   72 
   73 #include <sys/param.h>
   74 #include <sys/systm.h>
   75 #include <sys/proc.h>
   76 #include <sys/fcntl.h>
   77 #include <sys/file.h>
   78 #include <sys/filedesc.h>
   79 #include <sys/filio.h>
   80 #include <sys/kernel.h>
   81 #include <sys/ttycom.h>
   82 #include <sys/stat.h>
   83 #include <sys/poll.h>
   84 #include <sys/signalvar.h>
   85 #include <sys/vnode.h>
   86 #include <sys/uio.h>
   87 #include <sys/select.h>
   88 #include <sys/mount.h>
   89 #include <sys/syscallargs.h>
   90 #include <sys/sysctl.h>
   91 #include <sys/kauth.h>
   92 #include <sys/atomic.h>
   93 #include <sys/pipe.h>
   94 
   95 #include <uvm/uvm_extern.h>
   96 
   97 /*
   98  * Use this to disable direct I/O and decrease the code size:
   99  * #define PIPE_NODIRECT
  100  */
  101 
  102 /* XXX Disabled for now; rare hangs switching between direct/buffered */        
  103 #define PIPE_NODIRECT
  104 
  105 static int      pipe_read(file_t *, off_t *, struct uio *, kauth_cred_t, int);
  106 static int      pipe_write(file_t *, off_t *, struct uio *, kauth_cred_t, int);
  107 static int      pipe_close(file_t *);
  108 static int      pipe_poll(file_t *, int);
  109 static int      pipe_kqfilter(file_t *, struct knote *);
  110 static int      pipe_stat(file_t *, struct stat *);
  111 static int      pipe_ioctl(file_t *, u_long, void *);
  112 static void     pipe_restart(file_t *);
  113 
  114 static const struct fileops pipeops = {
  115         .fo_read = pipe_read,
  116         .fo_write = pipe_write,
  117         .fo_ioctl = pipe_ioctl,
  118         .fo_fcntl = fnullop_fcntl,
  119         .fo_poll = pipe_poll,
  120         .fo_stat = pipe_stat,
  121         .fo_close = pipe_close,
  122         .fo_kqfilter = pipe_kqfilter,
  123         .fo_restart = pipe_restart,
  124 };
  125 
  126 /*
  127  * Default pipe buffer size(s), this can be kind-of large now because pipe
  128  * space is pageable.  The pipe code will try to maintain locality of
  129  * reference for performance reasons, so small amounts of outstanding I/O
  130  * will not wipe the cache.
  131  */
  132 #define MINPIPESIZE     (PIPE_SIZE / 3)
  133 #define MAXPIPESIZE     (2 * PIPE_SIZE / 3)
  134 
  135 /*
  136  * Maximum amount of kva for pipes -- this is kind-of a soft limit, but
  137  * is there so that on large systems, we don't exhaust it.
  138  */
  139 #define MAXPIPEKVA      (8 * 1024 * 1024)
  140 static u_int    maxpipekva = MAXPIPEKVA;
  141 
  142 /*
  143  * Limit for direct transfers, we cannot, of course limit
  144  * the amount of kva for pipes in general though.
  145  */
  146 #define LIMITPIPEKVA    (16 * 1024 * 1024)
  147 static u_int    limitpipekva = LIMITPIPEKVA;
  148 
  149 /*
  150  * Limit the number of "big" pipes
  151  */
  152 #define LIMITBIGPIPES   32
  153 static u_int    maxbigpipes = LIMITBIGPIPES;
  154 static u_int    nbigpipe = 0;
  155 
  156 /*
  157  * Amount of KVA consumed by pipe buffers.
  158  */
  159 static u_int    amountpipekva = 0;
  160 
  161 static void     pipeclose(struct pipe *);
  162 static void     pipe_free_kmem(struct pipe *);
  163 static int      pipe_create(struct pipe **, pool_cache_t);
  164 static int      pipelock(struct pipe *, int);
  165 static inline void pipeunlock(struct pipe *);
  166 static void     pipeselwakeup(struct pipe *, struct pipe *, int);
  167 #ifndef PIPE_NODIRECT
  168 static int      pipe_direct_write(file_t *, struct pipe *, struct uio *);
  169 #endif
  170 static int      pipespace(struct pipe *, int);
  171 static int      pipe_ctor(void *, void *, int);
  172 static void     pipe_dtor(void *, void *);
  173 
  174 #ifndef PIPE_NODIRECT
  175 static int      pipe_loan_alloc(struct pipe *, int);
  176 static void     pipe_loan_free(struct pipe *);
  177 #endif /* PIPE_NODIRECT */
  178 
  179 static pool_cache_t     pipe_wr_cache;
  180 static pool_cache_t     pipe_rd_cache;
  181 
  182 void
  183 pipe_init(void)
  184 {
  185 
  186         /* Writer side is not automatically allocated KVA. */
  187         pipe_wr_cache = pool_cache_init(sizeof(struct pipe), 0, 0, 0, "pipewr",
  188             NULL, IPL_NONE, pipe_ctor, pipe_dtor, NULL);
  189         KASSERT(pipe_wr_cache != NULL);
  190 
  191         /* Reader side gets preallocated KVA. */
  192         pipe_rd_cache = pool_cache_init(sizeof(struct pipe), 0, 0, 0, "piperd",
  193             NULL, IPL_NONE, pipe_ctor, pipe_dtor, (void *)1);
  194         KASSERT(pipe_rd_cache != NULL);
  195 }
  196 
  197 static int
  198 pipe_ctor(void *arg, void *obj, int flags)
  199 {
  200         struct pipe *pipe;
  201         vaddr_t va;
  202 
  203         pipe = obj;
  204 
  205         memset(pipe, 0, sizeof(struct pipe));
  206         if (arg != NULL) {
  207                 /* Preallocate space. */
  208                 va = uvm_km_alloc(kernel_map, PIPE_SIZE, 0,
  209                     UVM_KMF_PAGEABLE | UVM_KMF_WAITVA);
  210                 KASSERT(va != 0);
  211                 pipe->pipe_kmem = va;
  212                 atomic_add_int(&amountpipekva, PIPE_SIZE);
  213         }
  214         cv_init(&pipe->pipe_rcv, "pipe_rd");
  215         cv_init(&pipe->pipe_wcv, "pipe_wr");
  216         cv_init(&pipe->pipe_draincv, "pipe_drn");
  217         cv_init(&pipe->pipe_lkcv, "pipe_lk");
  218         selinit(&pipe->pipe_sel);
  219         pipe->pipe_state = PIPE_SIGNALR;
  220 
  221         return 0;
  222 }
  223 
  224 static void
  225 pipe_dtor(void *arg, void *obj)
  226 {
  227         struct pipe *pipe;
  228 
  229         pipe = obj;
  230 
  231         cv_destroy(&pipe->pipe_rcv);
  232         cv_destroy(&pipe->pipe_wcv);
  233         cv_destroy(&pipe->pipe_draincv);
  234         cv_destroy(&pipe->pipe_lkcv);
  235         seldestroy(&pipe->pipe_sel);
  236         if (pipe->pipe_kmem != 0) {
  237                 uvm_km_free(kernel_map, pipe->pipe_kmem, PIPE_SIZE,
  238                     UVM_KMF_PAGEABLE);
  239                 atomic_add_int(&amountpipekva, -PIPE_SIZE);
  240         }
  241 }
  242 
  243 /*
  244  * The pipe system call for the DTYPE_PIPE type of pipes
  245  */
  246 int
  247 pipe1(struct lwp *l, register_t *retval, int flags)
  248 {
  249         struct pipe *rpipe, *wpipe;
  250         file_t *rf, *wf;
  251         int fd, error;
  252         proc_t *p;
  253 
  254         p = curproc;
  255         rpipe = wpipe = NULL;
  256         if (pipe_create(&rpipe, pipe_rd_cache) ||
  257             pipe_create(&wpipe, pipe_wr_cache)) {
  258                 error = ENOMEM;
  259                 goto free2;
  260         }
  261         rpipe->pipe_lock = mutex_obj_alloc(MUTEX_DEFAULT, IPL_NONE);
  262         wpipe->pipe_lock = rpipe->pipe_lock;
  263         mutex_obj_hold(wpipe->pipe_lock);
  264 
  265         error = fd_allocfile(&rf, &fd);
  266         if (error)
  267                 goto free2;
  268         retval[0] = fd;
  269         rf->f_flag = FREAD | flags;
  270         rf->f_type = DTYPE_PIPE;
  271         rf->f_data = (void *)rpipe;
  272         rf->f_ops = &pipeops;
  273 
  274         error = fd_allocfile(&wf, &fd);
  275         if (error)
  276                 goto free3;
  277         retval[1] = fd;
  278         wf->f_flag = FWRITE | flags;
  279         wf->f_type = DTYPE_PIPE;
  280         wf->f_data = (void *)wpipe;
  281         wf->f_ops = &pipeops;
  282 
  283         rpipe->pipe_peer = wpipe;
  284         wpipe->pipe_peer = rpipe;
  285 
  286         fd_affix(p, rf, (int)retval[0]);
  287         fd_affix(p, wf, (int)retval[1]);
  288         return (0);
  289 free3:
  290         fd_abort(p, rf, (int)retval[0]);
  291 free2:
  292         pipeclose(wpipe);
  293         pipeclose(rpipe);
  294 
  295         return (error);
  296 }
  297 
  298 int
  299 sys_pipe(struct lwp *l, const void *v, register_t *retval)
  300 {
  301         return pipe1(l, retval, 0);
  302 }
  303 
  304 /*
  305  * Allocate kva for pipe circular buffer, the space is pageable
  306  * This routine will 'realloc' the size of a pipe safely, if it fails
  307  * it will retain the old buffer.
  308  * If it fails it will return ENOMEM.
  309  */
  310 static int
  311 pipespace(struct pipe *pipe, int size)
  312 {
  313         void *buffer;
  314 
  315         /*
  316          * Allocate pageable virtual address space.  Physical memory is
  317          * allocated on demand.
  318          */
  319         if (size == PIPE_SIZE && pipe->pipe_kmem != 0) {
  320                 buffer = (void *)pipe->pipe_kmem;
  321         } else {
  322                 buffer = (void *)uvm_km_alloc(kernel_map, round_page(size),
  323                     0, UVM_KMF_PAGEABLE);
  324                 if (buffer == NULL)
  325                         return (ENOMEM);
  326                 atomic_add_int(&amountpipekva, size);
  327         }
  328 
  329         /* free old resources if we're resizing */
  330         pipe_free_kmem(pipe);
  331         pipe->pipe_buffer.buffer = buffer;
  332         pipe->pipe_buffer.size = size;
  333         pipe->pipe_buffer.in = 0;
  334         pipe->pipe_buffer.out = 0;
  335         pipe->pipe_buffer.cnt = 0;
  336         return (0);
  337 }
  338 
  339 /*
  340  * Initialize and allocate VM and memory for pipe.
  341  */
  342 static int
  343 pipe_create(struct pipe **pipep, pool_cache_t cache)
  344 {
  345         struct pipe *pipe;
  346         int error;
  347 
  348         pipe = pool_cache_get(cache, PR_WAITOK);
  349         KASSERT(pipe != NULL);
  350         *pipep = pipe;
  351         error = 0;
  352         getnanotime(&pipe->pipe_btime);
  353         pipe->pipe_atime = pipe->pipe_mtime = pipe->pipe_btime;
  354         pipe->pipe_lock = NULL;
  355         if (cache == pipe_rd_cache) {
  356                 error = pipespace(pipe, PIPE_SIZE);
  357         } else {
  358                 pipe->pipe_buffer.buffer = NULL;
  359                 pipe->pipe_buffer.size = 0;
  360                 pipe->pipe_buffer.in = 0;
  361                 pipe->pipe_buffer.out = 0;
  362                 pipe->pipe_buffer.cnt = 0;
  363         }
  364         return error;
  365 }
  366 
  367 /*
  368  * Lock a pipe for I/O, blocking other access
  369  * Called with pipe spin lock held.
  370  */
  371 static int
  372 pipelock(struct pipe *pipe, int catch)
  373 {
  374         int error;
  375 
  376         KASSERT(mutex_owned(pipe->pipe_lock));
  377 
  378         while (pipe->pipe_state & PIPE_LOCKFL) {
  379                 pipe->pipe_state |= PIPE_LWANT;
  380                 if (catch) {
  381                         error = cv_wait_sig(&pipe->pipe_lkcv, pipe->pipe_lock);
  382                         if (error != 0)
  383                                 return error;
  384                 } else
  385                         cv_wait(&pipe->pipe_lkcv, pipe->pipe_lock);
  386         }
  387 
  388         pipe->pipe_state |= PIPE_LOCKFL;
  389 
  390         return 0;
  391 }
  392 
  393 /*
  394  * unlock a pipe I/O lock
  395  */
  396 static inline void
  397 pipeunlock(struct pipe *pipe)
  398 {
  399 
  400         KASSERT(pipe->pipe_state & PIPE_LOCKFL);
  401 
  402         pipe->pipe_state &= ~PIPE_LOCKFL;
  403         if (pipe->pipe_state & PIPE_LWANT) {
  404                 pipe->pipe_state &= ~PIPE_LWANT;
  405                 cv_broadcast(&pipe->pipe_lkcv);
  406         }
  407 }
  408 
  409 /*
  410  * Select/poll wakup. This also sends SIGIO to peer connected to
  411  * 'sigpipe' side of pipe.
  412  */
  413 static void
  414 pipeselwakeup(struct pipe *selp, struct pipe *sigp, int code)
  415 {
  416         int band;
  417 
  418         switch (code) {
  419         case POLL_IN:
  420                 band = POLLIN|POLLRDNORM;
  421                 break;
  422         case POLL_OUT:
  423                 band = POLLOUT|POLLWRNORM;
  424                 break;
  425         case POLL_HUP:
  426                 band = POLLHUP;
  427                 break;
  428         case POLL_ERR:
  429                 band = POLLERR;
  430                 break;
  431         default:
  432                 band = 0;
  433 #ifdef DIAGNOSTIC
  434                 printf("bad siginfo code %d in pipe notification.\n", code);
  435 #endif
  436                 break;
  437         }
  438 
  439         selnotify(&selp->pipe_sel, band, NOTE_SUBMIT);
  440 
  441         if (sigp == NULL || (sigp->pipe_state & PIPE_ASYNC) == 0)
  442                 return;
  443 
  444         fownsignal(sigp->pipe_pgid, SIGIO, code, band, selp);
  445 }
  446 
  447 static int
  448 pipe_read(file_t *fp, off_t *offset, struct uio *uio, kauth_cred_t cred,
  449     int flags)
  450 {
  451         struct pipe *rpipe = (struct pipe *) fp->f_data;
  452         struct pipebuf *bp = &rpipe->pipe_buffer;
  453         kmutex_t *lock = rpipe->pipe_lock;
  454         int error;
  455         size_t nread = 0;
  456         size_t size;
  457         size_t ocnt;
  458         unsigned int wakeup_state = 0;
  459 
  460         mutex_enter(lock);
  461         ++rpipe->pipe_busy;
  462         ocnt = bp->cnt;
  463 
  464 again:
  465         error = pipelock(rpipe, 1);
  466         if (error)
  467                 goto unlocked_error;
  468 
  469         while (uio->uio_resid) {
  470                 /*
  471                  * Normal pipe buffer receive.
  472                  */
  473                 if (bp->cnt > 0) {
  474                         size = bp->size - bp->out;
  475                         if (size > bp->cnt)
  476                                 size = bp->cnt;
  477                         if (size > uio->uio_resid)
  478                                 size = uio->uio_resid;
  479 
  480                         mutex_exit(lock);
  481                         error = uiomove((char *)bp->buffer + bp->out, size, uio);
  482                         mutex_enter(lock);
  483                         if (error)
  484                                 break;
  485 
  486                         bp->out += size;
  487                         if (bp->out >= bp->size)
  488                                 bp->out = 0;
  489 
  490                         bp->cnt -= size;
  491 
  492                         /*
  493                          * If there is no more to read in the pipe, reset
  494                          * its pointers to the beginning.  This improves
  495                          * cache hit stats.
  496                          */
  497                         if (bp->cnt == 0) {
  498                                 bp->in = 0;
  499                                 bp->out = 0;
  500                         }
  501                         nread += size;
  502                         continue;
  503                 }
  504 
  505 #ifndef PIPE_NODIRECT
  506                 if ((rpipe->pipe_state & PIPE_DIRECTR) != 0) {
  507                         /*
  508                          * Direct copy, bypassing a kernel buffer.
  509                          */
  510                         void *va;
  511                         u_int gen;
  512 
  513                         KASSERT(rpipe->pipe_state & PIPE_DIRECTW);
  514 
  515                         size = rpipe->pipe_map.cnt;
  516                         if (size > uio->uio_resid)
  517                                 size = uio->uio_resid;
  518 
  519                         va = (char *)rpipe->pipe_map.kva + rpipe->pipe_map.pos;
  520                         gen = rpipe->pipe_map.egen;
  521                         mutex_exit(lock);
  522 
  523                         /*
  524                          * Consume emap and read the data from loaned pages.
  525                          */
  526                         uvm_emap_consume(gen);
  527                         error = uiomove(va, size, uio);
  528 
  529                         mutex_enter(lock);
  530                         if (error)
  531                                 break;
  532                         nread += size;
  533                         rpipe->pipe_map.pos += size;
  534                         rpipe->pipe_map.cnt -= size;
  535                         if (rpipe->pipe_map.cnt == 0) {
  536                                 rpipe->pipe_state &= ~PIPE_DIRECTR;
  537                                 cv_broadcast(&rpipe->pipe_wcv);
  538                         }
  539                         continue;
  540                 }
  541 #endif
  542                 /*
  543                  * Break if some data was read.
  544                  */
  545                 if (nread > 0)
  546                         break;
  547 
  548                 /*
  549                  * Detect EOF condition.
  550                  * Read returns 0 on EOF, no need to set error.
  551                  */
  552                 if (rpipe->pipe_state & PIPE_EOF)
  553                         break;
  554 
  555                 /*
  556                  * Don't block on non-blocking I/O.
  557                  */
  558                 if (fp->f_flag & FNONBLOCK) {
  559                         error = EAGAIN;
  560                         break;
  561                 }
  562 
  563                 /*
  564                  * Unlock the pipe buffer for our remaining processing.
  565                  * We will either break out with an error or we will
  566                  * sleep and relock to loop.
  567                  */
  568                 pipeunlock(rpipe);
  569 
  570                 /*
  571                  * Re-check to see if more direct writes are pending.
  572                  */
  573                 if ((rpipe->pipe_state & PIPE_DIRECTR) != 0)
  574                         goto again;
  575 
  576 #if 1   /* XXX (dsl) I'm sure these aren't needed here ... */
  577                 /*
  578                  * We want to read more, wake up select/poll.
  579                  */
  580                 pipeselwakeup(rpipe, rpipe->pipe_peer, POLL_OUT);
  581 
  582                 /*
  583                  * If the "write-side" is blocked, wake it up now.
  584                  */
  585                 cv_broadcast(&rpipe->pipe_wcv);
  586 #endif
  587 
  588                 if (wakeup_state & PIPE_RESTART) {
  589                         error = ERESTART;
  590                         goto unlocked_error;
  591                 }
  592 
  593                 /* Now wait until the pipe is filled */
  594                 error = cv_wait_sig(&rpipe->pipe_rcv, lock);
  595                 if (error != 0)
  596                         goto unlocked_error;
  597                 wakeup_state = rpipe->pipe_state;
  598                 goto again;
  599         }
  600 
  601         if (error == 0)
  602                 getnanotime(&rpipe->pipe_atime);
  603         pipeunlock(rpipe);
  604 
  605 unlocked_error:
  606         --rpipe->pipe_busy;
  607         if (rpipe->pipe_busy == 0) {
  608                 rpipe->pipe_state &= ~PIPE_RESTART;
  609                 cv_broadcast(&rpipe->pipe_draincv);
  610         }
  611         if (bp->cnt < MINPIPESIZE) {
  612                 cv_broadcast(&rpipe->pipe_wcv);
  613         }
  614 
  615         /*
  616          * If anything was read off the buffer, signal to the writer it's
  617          * possible to write more data. Also send signal if we are here for the
  618          * first time after last write.
  619          */
  620         if ((bp->size - bp->cnt) >= PIPE_BUF
  621             && (ocnt != bp->cnt || (rpipe->pipe_state & PIPE_SIGNALR))) {
  622                 pipeselwakeup(rpipe, rpipe->pipe_peer, POLL_OUT);
  623                 rpipe->pipe_state &= ~PIPE_SIGNALR;
  624         }
  625 
  626         mutex_exit(lock);
  627         return (error);
  628 }
  629 
  630 #ifndef PIPE_NODIRECT
  631 /*
  632  * Allocate structure for loan transfer.
  633  */
  634 static int
  635 pipe_loan_alloc(struct pipe *wpipe, int npages)
  636 {
  637         vsize_t len;
  638 
  639         len = (vsize_t)npages << PAGE_SHIFT;
  640         atomic_add_int(&amountpipekva, len);
  641         wpipe->pipe_map.kva = uvm_km_alloc(kernel_map, len, 0,
  642             UVM_KMF_VAONLY | UVM_KMF_WAITVA);
  643         if (wpipe->pipe_map.kva == 0) {
  644                 atomic_add_int(&amountpipekva, -len);
  645                 return (ENOMEM);
  646         }
  647 
  648         wpipe->pipe_map.npages = npages;
  649         wpipe->pipe_map.pgs = kmem_alloc(npages * sizeof(struct vm_page *),
  650             KM_SLEEP);
  651         return (0);
  652 }
  653 
  654 /*
  655  * Free resources allocated for loan transfer.
  656  */
  657 static void
  658 pipe_loan_free(struct pipe *wpipe)
  659 {
  660         vsize_t len;
  661 
  662         len = (vsize_t)wpipe->pipe_map.npages << PAGE_SHIFT;
  663         uvm_emap_remove(wpipe->pipe_map.kva, len);      /* XXX */
  664         uvm_km_free(kernel_map, wpipe->pipe_map.kva, len, UVM_KMF_VAONLY);
  665         wpipe->pipe_map.kva = 0;
  666         atomic_add_int(&amountpipekva, -len);
  667         kmem_free(wpipe->pipe_map.pgs,
  668             wpipe->pipe_map.npages * sizeof(struct vm_page *));
  669         wpipe->pipe_map.pgs = NULL;
  670 }
  671 
  672 /*
  673  * NetBSD direct write, using uvm_loan() mechanism.
  674  * This implements the pipe buffer write mechanism.  Note that only
  675  * a direct write OR a normal pipe write can be pending at any given time.
  676  * If there are any characters in the pipe buffer, the direct write will
  677  * be deferred until the receiving process grabs all of the bytes from
  678  * the pipe buffer.  Then the direct mapping write is set-up.
  679  *
  680  * Called with the long-term pipe lock held.
  681  */
  682 static int
  683 pipe_direct_write(file_t *fp, struct pipe *wpipe, struct uio *uio)
  684 {
  685         struct vm_page **pgs;
  686         vaddr_t bbase, base, bend;
  687         vsize_t blen, bcnt;
  688         int error, npages;
  689         voff_t bpos;
  690         kmutex_t *lock = wpipe->pipe_lock;
  691 
  692         KASSERT(mutex_owned(wpipe->pipe_lock));
  693         KASSERT(wpipe->pipe_map.cnt == 0);
  694 
  695         mutex_exit(lock);
  696 
  697         /*
  698          * Handle first PIPE_CHUNK_SIZE bytes of buffer. Deal with buffers
  699          * not aligned to PAGE_SIZE.
  700          */
  701         bbase = (vaddr_t)uio->uio_iov->iov_base;
  702         base = trunc_page(bbase);
  703         bend = round_page(bbase + uio->uio_iov->iov_len);
  704         blen = bend - base;
  705         bpos = bbase - base;
  706 
  707         if (blen > PIPE_DIRECT_CHUNK) {
  708                 blen = PIPE_DIRECT_CHUNK;
  709                 bend = base + blen;
  710                 bcnt = PIPE_DIRECT_CHUNK - bpos;
  711         } else {
  712                 bcnt = uio->uio_iov->iov_len;
  713         }
  714         npages = blen >> PAGE_SHIFT;
  715 
  716         /*
  717          * Free the old kva if we need more pages than we have
  718          * allocated.
  719          */
  720         if (wpipe->pipe_map.kva != 0 && npages > wpipe->pipe_map.npages)
  721                 pipe_loan_free(wpipe);
  722 
  723         /* Allocate new kva. */
  724         if (wpipe->pipe_map.kva == 0) {
  725                 error = pipe_loan_alloc(wpipe, npages);
  726                 if (error) {
  727                         mutex_enter(lock);
  728                         return (error);
  729                 }
  730         }
  731 
  732         /* Loan the write buffer memory from writer process */
  733         pgs = wpipe->pipe_map.pgs;
  734         error = uvm_loan(&uio->uio_vmspace->vm_map, base, blen,
  735                          pgs, UVM_LOAN_TOPAGE);
  736         if (error) {
  737                 pipe_loan_free(wpipe);
  738                 mutex_enter(lock);
  739                 return (ENOMEM); /* so that caller fallback to ordinary write */
  740         }
  741 
  742         /* Enter the loaned pages to KVA, produce new emap generation number. */
  743         uvm_emap_enter(wpipe->pipe_map.kva, pgs, npages);
  744         wpipe->pipe_map.egen = uvm_emap_produce();
  745 
  746         /* Now we can put the pipe in direct write mode */
  747         wpipe->pipe_map.pos = bpos;
  748         wpipe->pipe_map.cnt = bcnt;
  749 
  750         /*
  751          * But before we can let someone do a direct read, we
  752          * have to wait until the pipe is drained.  Release the
  753          * pipe lock while we wait.
  754          */
  755         mutex_enter(lock);
  756         wpipe->pipe_state |= PIPE_DIRECTW;
  757         pipeunlock(wpipe);
  758 
  759         while (error == 0 && wpipe->pipe_buffer.cnt > 0) {
  760                 cv_broadcast(&wpipe->pipe_rcv);
  761                 error = cv_wait_sig(&wpipe->pipe_wcv, lock);
  762                 if (error == 0 && wpipe->pipe_state & PIPE_EOF)
  763                         error = EPIPE;
  764         }
  765 
  766         /* Pipe is drained; next read will off the direct buffer */
  767         wpipe->pipe_state |= PIPE_DIRECTR;
  768 
  769         /* Wait until the reader is done */
  770         while (error == 0 && (wpipe->pipe_state & PIPE_DIRECTR)) {
  771                 cv_broadcast(&wpipe->pipe_rcv);
  772                 pipeselwakeup(wpipe, wpipe, POLL_IN);
  773                 error = cv_wait_sig(&wpipe->pipe_wcv, lock);
  774                 if (error == 0 && wpipe->pipe_state & PIPE_EOF)
  775                         error = EPIPE;
  776         }
  777 
  778         /* Take pipe out of direct write mode */
  779         wpipe->pipe_state &= ~(PIPE_DIRECTW | PIPE_DIRECTR);
  780 
  781         /* Acquire the pipe lock and cleanup */
  782         (void)pipelock(wpipe, 0);
  783         mutex_exit(lock);
  784 
  785         if (pgs != NULL) {
  786                 /* XXX: uvm_emap_remove */
  787                 uvm_unloan(pgs, npages, UVM_LOAN_TOPAGE);
  788         }
  789         if (error || amountpipekva > maxpipekva)
  790                 pipe_loan_free(wpipe);
  791 
  792         mutex_enter(lock);
  793         if (error) {
  794                 pipeselwakeup(wpipe, wpipe, POLL_ERR);
  795 
  796                 /*
  797                  * If nothing was read from what we offered, return error
  798                  * straight on. Otherwise update uio resid first. Caller
  799                  * will deal with the error condition, returning short
  800                  * write, error, or restarting the write(2) as appropriate.
  801                  */
  802                 if (wpipe->pipe_map.cnt == bcnt) {
  803                         wpipe->pipe_map.cnt = 0;
  804                         cv_broadcast(&wpipe->pipe_wcv);
  805                         return (error);
  806                 }
  807 
  808                 bcnt -= wpipe->pipe_map.cnt;
  809         }
  810 
  811         uio->uio_resid -= bcnt;
  812         /* uio_offset not updated, not set/used for write(2) */
  813         uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + bcnt;
  814         uio->uio_iov->iov_len -= bcnt;
  815         if (uio->uio_iov->iov_len == 0) {
  816                 uio->uio_iov++;
  817                 uio->uio_iovcnt--;
  818         }
  819 
  820         wpipe->pipe_map.cnt = 0;
  821         return (error);
  822 }
  823 #endif /* !PIPE_NODIRECT */
  824 
  825 static int
  826 pipe_write(file_t *fp, off_t *offset, struct uio *uio, kauth_cred_t cred,
  827     int flags)
  828 {
  829         struct pipe *wpipe, *rpipe;
  830         struct pipebuf *bp;
  831         kmutex_t *lock;
  832         int error;
  833         unsigned int wakeup_state = 0;
  834 
  835         /* We want to write to our peer */
  836         rpipe = (struct pipe *) fp->f_data;
  837         lock = rpipe->pipe_lock;
  838         error = 0;
  839 
  840         mutex_enter(lock);
  841         wpipe = rpipe->pipe_peer;
  842 
  843         /*
  844          * Detect loss of pipe read side, issue SIGPIPE if lost.
  845          */
  846         if (wpipe == NULL || (wpipe->pipe_state & PIPE_EOF) != 0) {
  847                 mutex_exit(lock);
  848                 return EPIPE;
  849         }
  850         ++wpipe->pipe_busy;
  851 
  852         /* Aquire the long-term pipe lock */
  853         if ((error = pipelock(wpipe, 1)) != 0) {
  854                 --wpipe->pipe_busy;
  855                 if (wpipe->pipe_busy == 0) {
  856                         wpipe->pipe_state &= ~PIPE_RESTART;
  857                         cv_broadcast(&wpipe->pipe_draincv);
  858                 }
  859                 mutex_exit(lock);
  860                 return (error);
  861         }
  862 
  863         bp = &wpipe->pipe_buffer;
  864 
  865         /*
  866          * If it is advantageous to resize the pipe buffer, do so.
  867          */
  868         if ((uio->uio_resid > PIPE_SIZE) &&
  869             (nbigpipe < maxbigpipes) &&
  870 #ifndef PIPE_NODIRECT
  871             (wpipe->pipe_state & PIPE_DIRECTW) == 0 &&
  872 #endif
  873             (bp->size <= PIPE_SIZE) && (bp->cnt == 0)) {
  874 
  875                 if (pipespace(wpipe, BIG_PIPE_SIZE) == 0)
  876                         atomic_inc_uint(&nbigpipe);
  877         }
  878 
  879         while (uio->uio_resid) {
  880                 size_t space;
  881 
  882 #ifndef PIPE_NODIRECT
  883                 /*
  884                  * Pipe buffered writes cannot be coincidental with
  885                  * direct writes.  Also, only one direct write can be
  886                  * in progress at any one time.  We wait until the currently
  887                  * executing direct write is completed before continuing.
  888                  *
  889                  * We break out if a signal occurs or the reader goes away.
  890                  */
  891                 while (error == 0 && wpipe->pipe_state & PIPE_DIRECTW) {
  892                         cv_broadcast(&wpipe->pipe_rcv);
  893                         pipeunlock(wpipe);
  894                         error = cv_wait_sig(&wpipe->pipe_wcv, lock);
  895                         (void)pipelock(wpipe, 0);
  896                         if (wpipe->pipe_state & PIPE_EOF)
  897                                 error = EPIPE;
  898                 }
  899                 if (error)
  900                         break;
  901 
  902                 /*
  903                  * If the transfer is large, we can gain performance if
  904                  * we do process-to-process copies directly.
  905                  * If the write is non-blocking, we don't use the
  906                  * direct write mechanism.
  907                  *
  908                  * The direct write mechanism will detect the reader going
  909                  * away on us.
  910                  */
  911                 if ((uio->uio_iov->iov_len >= PIPE_MINDIRECT) &&
  912                     (fp->f_flag & FNONBLOCK) == 0 &&
  913                     (wpipe->pipe_map.kva || (amountpipekva < limitpipekva))) {
  914                         error = pipe_direct_write(fp, wpipe, uio);
  915 
  916                         /*
  917                          * Break out if error occurred, unless it's ENOMEM.
  918                          * ENOMEM means we failed to allocate some resources
  919                          * for direct write, so we just fallback to ordinary
  920                          * write. If the direct write was successful,
  921                          * process rest of data via ordinary write.
  922                          */
  923                         if (error == 0)
  924                                 continue;
  925 
  926                         if (error != ENOMEM)
  927                                 break;
  928                 }
  929 #endif /* PIPE_NODIRECT */
  930 
  931                 space = bp->size - bp->cnt;
  932 
  933                 /* Writes of size <= PIPE_BUF must be atomic. */
  934                 if ((space < uio->uio_resid) && (uio->uio_resid <= PIPE_BUF))
  935                         space = 0;
  936 
  937                 if (space > 0) {
  938                         int size;       /* Transfer size */
  939                         int segsize;    /* first segment to transfer */
  940 
  941                         /*
  942                          * Transfer size is minimum of uio transfer
  943                          * and free space in pipe buffer.
  944                          */
  945                         if (space > uio->uio_resid)
  946                                 size = uio->uio_resid;
  947                         else
  948                                 size = space;
  949                         /*
  950                          * First segment to transfer is minimum of
  951                          * transfer size and contiguous space in
  952                          * pipe buffer.  If first segment to transfer
  953                          * is less than the transfer size, we've got
  954                          * a wraparound in the buffer.
  955                          */
  956                         segsize = bp->size - bp->in;
  957                         if (segsize > size)
  958                                 segsize = size;
  959 
  960                         /* Transfer first segment */
  961                         mutex_exit(lock);
  962                         error = uiomove((char *)bp->buffer + bp->in, segsize,
  963                             uio);
  964 
  965                         if (error == 0 && segsize < size) {
  966                                 /*
  967                                  * Transfer remaining part now, to
  968                                  * support atomic writes.  Wraparound
  969                                  * happened.
  970                                  */
  971                                 KASSERT(bp->in + segsize == bp->size);
  972                                 error = uiomove(bp->buffer,
  973                                     size - segsize, uio);
  974                         }
  975                         mutex_enter(lock);
  976                         if (error)
  977                                 break;
  978 
  979                         bp->in += size;
  980                         if (bp->in >= bp->size) {
  981                                 KASSERT(bp->in == size - segsize + bp->size);
  982                                 bp->in = size - segsize;
  983                         }
  984 
  985                         bp->cnt += size;
  986                         KASSERT(bp->cnt <= bp->size);
  987                         wakeup_state = 0;
  988                 } else {
  989                         /*
  990                          * If the "read-side" has been blocked, wake it up now.
  991                          */
  992                         cv_broadcast(&wpipe->pipe_rcv);
  993 
  994                         /*
  995                          * Don't block on non-blocking I/O.
  996                          */
  997                         if (fp->f_flag & FNONBLOCK) {
  998                                 error = EAGAIN;
  999                                 break;
 1000                         }
 1001 
 1002                         /*
 1003                          * We have no more space and have something to offer,
 1004                          * wake up select/poll.
 1005                          */
 1006                         if (bp->cnt)
 1007                                 pipeselwakeup(wpipe, wpipe, POLL_IN);
 1008 
 1009                         if (wakeup_state & PIPE_RESTART) {
 1010                                 error = ERESTART;
 1011                                 break;
 1012                         }
 1013 
 1014                         pipeunlock(wpipe);
 1015                         error = cv_wait_sig(&wpipe->pipe_wcv, lock);
 1016                         (void)pipelock(wpipe, 0);
 1017                         if (error != 0)
 1018                                 break;
 1019                         /*
 1020                          * If read side wants to go away, we just issue a signal
 1021                          * to ourselves.
 1022                          */
 1023                         if (wpipe->pipe_state & PIPE_EOF) {
 1024                                 error = EPIPE;
 1025                                 break;
 1026                         }
 1027                         wakeup_state = wpipe->pipe_state;
 1028                 }
 1029         }
 1030 
 1031         --wpipe->pipe_busy;
 1032         if (wpipe->pipe_busy == 0) {
 1033                 wpipe->pipe_state &= ~PIPE_RESTART;
 1034                 cv_broadcast(&wpipe->pipe_draincv);
 1035         }
 1036         if (bp->cnt > 0) {
 1037                 cv_broadcast(&wpipe->pipe_rcv);
 1038         }
 1039 
 1040         /*
 1041          * Don't return EPIPE if I/O was successful
 1042          */
 1043         if (error == EPIPE && bp->cnt == 0 && uio->uio_resid == 0)
 1044                 error = 0;
 1045 
 1046         if (error == 0)
 1047                 getnanotime(&wpipe->pipe_mtime);
 1048 
 1049         /*
 1050          * We have something to offer, wake up select/poll.
 1051          * wpipe->pipe_map.cnt is always 0 in this point (direct write
 1052          * is only done synchronously), so check only wpipe->pipe_buffer.cnt
 1053          */
 1054         if (bp->cnt)
 1055                 pipeselwakeup(wpipe, wpipe, POLL_IN);
 1056 
 1057         /*
 1058          * Arrange for next read(2) to do a signal.
 1059          */
 1060         wpipe->pipe_state |= PIPE_SIGNALR;
 1061 
 1062         pipeunlock(wpipe);
 1063         mutex_exit(lock);
 1064         return (error);
 1065 }
 1066 
 1067 /*
 1068  * We implement a very minimal set of ioctls for compatibility with sockets.
 1069  */
 1070 int
 1071 pipe_ioctl(file_t *fp, u_long cmd, void *data)
 1072 {
 1073         struct pipe *pipe = fp->f_data;
 1074         kmutex_t *lock = pipe->pipe_lock;
 1075 
 1076         switch (cmd) {
 1077 
 1078         case FIONBIO:
 1079                 return (0);
 1080 
 1081         case FIOASYNC:
 1082                 mutex_enter(lock);
 1083                 if (*(int *)data) {
 1084                         pipe->pipe_state |= PIPE_ASYNC;
 1085                 } else {
 1086                         pipe->pipe_state &= ~PIPE_ASYNC;
 1087                 }
 1088                 mutex_exit(lock);
 1089                 return (0);
 1090 
 1091         case FIONREAD:
 1092                 mutex_enter(lock);
 1093 #ifndef PIPE_NODIRECT
 1094                 if (pipe->pipe_state & PIPE_DIRECTW)
 1095                         *(int *)data = pipe->pipe_map.cnt;
 1096                 else
 1097 #endif
 1098                         *(int *)data = pipe->pipe_buffer.cnt;
 1099                 mutex_exit(lock);
 1100                 return (0);
 1101 
 1102         case FIONWRITE:
 1103                 /* Look at other side */
 1104                 pipe = pipe->pipe_peer;
 1105                 mutex_enter(lock);
 1106 #ifndef PIPE_NODIRECT
 1107                 if (pipe->pipe_state & PIPE_DIRECTW)
 1108                         *(int *)data = pipe->pipe_map.cnt;
 1109                 else
 1110 #endif
 1111                         *(int *)data = pipe->pipe_buffer.cnt;
 1112                 mutex_exit(lock);
 1113                 return (0);
 1114 
 1115         case FIONSPACE:
 1116                 /* Look at other side */
 1117                 pipe = pipe->pipe_peer;
 1118                 mutex_enter(lock);
 1119 #ifndef PIPE_NODIRECT
 1120                 /*
 1121                  * If we're in direct-mode, we don't really have a
 1122                  * send queue, and any other write will block. Thus
 1123                  * zero seems like the best answer.
 1124                  */
 1125                 if (pipe->pipe_state & PIPE_DIRECTW)
 1126                         *(int *)data = 0;
 1127                 else
 1128 #endif
 1129                         *(int *)data = pipe->pipe_buffer.size -
 1130                             pipe->pipe_buffer.cnt;
 1131                 mutex_exit(lock);
 1132                 return (0);
 1133 
 1134         case TIOCSPGRP:
 1135         case FIOSETOWN:
 1136                 return fsetown(&pipe->pipe_pgid, cmd, data);
 1137 
 1138         case TIOCGPGRP:
 1139         case FIOGETOWN:
 1140                 return fgetown(pipe->pipe_pgid, cmd, data);
 1141 
 1142         }
 1143         return (EPASSTHROUGH);
 1144 }
 1145 
 1146 int
 1147 pipe_poll(file_t *fp, int events)
 1148 {
 1149         struct pipe *rpipe = fp->f_data;
 1150         struct pipe *wpipe;
 1151         int eof = 0;
 1152         int revents = 0;
 1153 
 1154         mutex_enter(rpipe->pipe_lock);
 1155         wpipe = rpipe->pipe_peer;
 1156 
 1157         if (events & (POLLIN | POLLRDNORM))
 1158                 if ((rpipe->pipe_buffer.cnt > 0) ||
 1159 #ifndef PIPE_NODIRECT
 1160                     (rpipe->pipe_state & PIPE_DIRECTR) ||
 1161 #endif
 1162                     (rpipe->pipe_state & PIPE_EOF))
 1163                         revents |= events & (POLLIN | POLLRDNORM);
 1164 
 1165         eof |= (rpipe->pipe_state & PIPE_EOF);
 1166 
 1167         if (wpipe == NULL)
 1168                 revents |= events & (POLLOUT | POLLWRNORM);
 1169         else {
 1170                 if (events & (POLLOUT | POLLWRNORM))
 1171                         if ((wpipe->pipe_state & PIPE_EOF) || (
 1172 #ifndef PIPE_NODIRECT
 1173                              (wpipe->pipe_state & PIPE_DIRECTW) == 0 &&
 1174 #endif
 1175                              (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF))
 1176                                 revents |= events & (POLLOUT | POLLWRNORM);
 1177 
 1178                 eof |= (wpipe->pipe_state & PIPE_EOF);
 1179         }
 1180 
 1181         if (wpipe == NULL || eof)
 1182                 revents |= POLLHUP;
 1183 
 1184         if (revents == 0) {
 1185                 if (events & (POLLIN | POLLRDNORM))
 1186                         selrecord(curlwp, &rpipe->pipe_sel);
 1187 
 1188                 if (events & (POLLOUT | POLLWRNORM))
 1189                         selrecord(curlwp, &wpipe->pipe_sel);
 1190         }
 1191         mutex_exit(rpipe->pipe_lock);
 1192 
 1193         return (revents);
 1194 }
 1195 
 1196 static int
 1197 pipe_stat(file_t *fp, struct stat *ub)
 1198 {
 1199         struct pipe *pipe = fp->f_data;
 1200 
 1201         mutex_enter(pipe->pipe_lock);
 1202         memset(ub, 0, sizeof(*ub));
 1203         ub->st_mode = S_IFIFO | S_IRUSR | S_IWUSR;
 1204         ub->st_blksize = pipe->pipe_buffer.size;
 1205         if (ub->st_blksize == 0 && pipe->pipe_peer)
 1206                 ub->st_blksize = pipe->pipe_peer->pipe_buffer.size;
 1207         ub->st_size = pipe->pipe_buffer.cnt;
 1208         ub->st_blocks = (ub->st_size) ? 1 : 0;
 1209         ub->st_atimespec = pipe->pipe_atime;
 1210         ub->st_mtimespec = pipe->pipe_mtime;
 1211         ub->st_ctimespec = ub->st_birthtimespec = pipe->pipe_btime;
 1212         ub->st_uid = kauth_cred_geteuid(fp->f_cred);
 1213         ub->st_gid = kauth_cred_getegid(fp->f_cred);
 1214 
 1215         /*
 1216          * Left as 0: st_dev, st_ino, st_nlink, st_rdev, st_flags, st_gen.
 1217          * XXX (st_dev, st_ino) should be unique.
 1218          */
 1219         mutex_exit(pipe->pipe_lock);
 1220         return 0;
 1221 }
 1222 
 1223 static int
 1224 pipe_close(file_t *fp)
 1225 {
 1226         struct pipe *pipe = fp->f_data;
 1227 
 1228         fp->f_data = NULL;
 1229         pipeclose(pipe);
 1230         return (0);
 1231 }
 1232 
 1233 static void
 1234 pipe_restart(file_t *fp)
 1235 {
 1236         struct pipe *pipe = fp->f_data;
 1237 
 1238         /*
 1239          * Unblock blocked reads/writes in order to allow close() to complete.
 1240          * System calls return ERESTART so that the fd is revalidated.
 1241          * (Partial writes return the transfer length.)
 1242          */
 1243         mutex_enter(pipe->pipe_lock);
 1244         pipe->pipe_state |= PIPE_RESTART;
 1245         /* Wakeup both cvs, maybe we only need one, but maybe there are some
 1246          * other paths where wakeup is needed, and it saves deciding which! */
 1247         cv_broadcast(&pipe->pipe_rcv);
 1248         cv_broadcast(&pipe->pipe_wcv);
 1249         mutex_exit(pipe->pipe_lock);
 1250 }
 1251 
 1252 static void
 1253 pipe_free_kmem(struct pipe *pipe)
 1254 {
 1255 
 1256         if (pipe->pipe_buffer.buffer != NULL) {
 1257                 if (pipe->pipe_buffer.size > PIPE_SIZE) {
 1258                         atomic_dec_uint(&nbigpipe);
 1259                 }
 1260                 if (pipe->pipe_buffer.buffer != (void *)pipe->pipe_kmem) {
 1261                         uvm_km_free(kernel_map,
 1262                             (vaddr_t)pipe->pipe_buffer.buffer,
 1263                             pipe->pipe_buffer.size, UVM_KMF_PAGEABLE);
 1264                         atomic_add_int(&amountpipekva,
 1265                             -pipe->pipe_buffer.size);
 1266                 }
 1267                 pipe->pipe_buffer.buffer = NULL;
 1268         }
 1269 #ifndef PIPE_NODIRECT
 1270         if (pipe->pipe_map.kva != 0) {
 1271                 pipe_loan_free(pipe);
 1272                 pipe->pipe_map.cnt = 0;
 1273                 pipe->pipe_map.kva = 0;
 1274                 pipe->pipe_map.pos = 0;
 1275                 pipe->pipe_map.npages = 0;
 1276         }
 1277 #endif /* !PIPE_NODIRECT */
 1278 }
 1279 
 1280 /*
 1281  * Shutdown the pipe.
 1282  */
 1283 static void
 1284 pipeclose(struct pipe *pipe)
 1285 {
 1286         kmutex_t *lock;
 1287         struct pipe *ppipe;
 1288 
 1289         if (pipe == NULL)
 1290                 return;
 1291 
 1292         KASSERT(cv_is_valid(&pipe->pipe_rcv));
 1293         KASSERT(cv_is_valid(&pipe->pipe_wcv));
 1294         KASSERT(cv_is_valid(&pipe->pipe_draincv));
 1295         KASSERT(cv_is_valid(&pipe->pipe_lkcv));
 1296 
 1297         lock = pipe->pipe_lock;
 1298         if (lock == NULL)
 1299                 /* Must have failed during create */
 1300                 goto free_resources;
 1301 
 1302         mutex_enter(lock);
 1303         pipeselwakeup(pipe, pipe, POLL_HUP);
 1304 
 1305         /*
 1306          * If the other side is blocked, wake it up saying that
 1307          * we want to close it down.
 1308          */
 1309         pipe->pipe_state |= PIPE_EOF;
 1310         if (pipe->pipe_busy) {
 1311                 while (pipe->pipe_busy) {
 1312                         cv_broadcast(&pipe->pipe_wcv);
 1313                         cv_wait_sig(&pipe->pipe_draincv, lock);
 1314                 }
 1315         }
 1316 
 1317         /*
 1318          * Disconnect from peer.
 1319          */
 1320         if ((ppipe = pipe->pipe_peer) != NULL) {
 1321                 pipeselwakeup(ppipe, ppipe, POLL_HUP);
 1322                 ppipe->pipe_state |= PIPE_EOF;
 1323                 cv_broadcast(&ppipe->pipe_rcv);
 1324                 ppipe->pipe_peer = NULL;
 1325         }
 1326 
 1327         /*
 1328          * Any knote objects still left in the list are
 1329          * the one attached by peer.  Since no one will
 1330          * traverse this list, we just clear it.
 1331          */
 1332         SLIST_INIT(&pipe->pipe_sel.sel_klist);
 1333 
 1334         KASSERT((pipe->pipe_state & PIPE_LOCKFL) == 0);
 1335         mutex_exit(lock);
 1336         mutex_obj_free(lock);
 1337 
 1338         /*
 1339          * Free resources.
 1340          */
 1341     free_resources:
 1342         pipe->pipe_pgid = 0;
 1343         pipe->pipe_state = PIPE_SIGNALR;
 1344         pipe_free_kmem(pipe);
 1345         if (pipe->pipe_kmem != 0) {
 1346                 pool_cache_put(pipe_rd_cache, pipe);
 1347         } else {
 1348                 pool_cache_put(pipe_wr_cache, pipe);
 1349         }
 1350 }
 1351 
 1352 static void
 1353 filt_pipedetach(struct knote *kn)
 1354 {
 1355         struct pipe *pipe;
 1356         kmutex_t *lock;
 1357 
 1358         pipe = ((file_t *)kn->kn_obj)->f_data;
 1359         lock = pipe->pipe_lock;
 1360 
 1361         mutex_enter(lock);
 1362 
 1363         switch(kn->kn_filter) {
 1364         case EVFILT_WRITE:
 1365                 /* Need the peer structure, not our own. */
 1366                 pipe = pipe->pipe_peer;
 1367 
 1368                 /* If reader end already closed, just return. */
 1369                 if (pipe == NULL) {
 1370                         mutex_exit(lock);
 1371                         return;
 1372                 }
 1373 
 1374                 break;
 1375         default:
 1376                 /* Nothing to do. */
 1377                 break;
 1378         }
 1379 
 1380         KASSERT(kn->kn_hook == pipe);
 1381         SLIST_REMOVE(&pipe->pipe_sel.sel_klist, kn, knote, kn_selnext);
 1382         mutex_exit(lock);
 1383 }
 1384 
 1385 static int
 1386 filt_piperead(struct knote *kn, long hint)
 1387 {
 1388         struct pipe *rpipe = ((file_t *)kn->kn_obj)->f_data;
 1389         struct pipe *wpipe;
 1390 
 1391         if ((hint & NOTE_SUBMIT) == 0) {
 1392                 mutex_enter(rpipe->pipe_lock);
 1393         }
 1394         wpipe = rpipe->pipe_peer;
 1395         kn->kn_data = rpipe->pipe_buffer.cnt;
 1396 
 1397         if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
 1398                 kn->kn_data = rpipe->pipe_map.cnt;
 1399 
 1400         if ((rpipe->pipe_state & PIPE_EOF) ||
 1401             (wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
 1402                 kn->kn_flags |= EV_EOF;
 1403                 if ((hint & NOTE_SUBMIT) == 0) {
 1404                         mutex_exit(rpipe->pipe_lock);
 1405                 }
 1406                 return (1);
 1407         }
 1408 
 1409         if ((hint & NOTE_SUBMIT) == 0) {
 1410                 mutex_exit(rpipe->pipe_lock);
 1411         }
 1412         return (kn->kn_data > 0);
 1413 }
 1414 
 1415 static int
 1416 filt_pipewrite(struct knote *kn, long hint)
 1417 {
 1418         struct pipe *rpipe = ((file_t *)kn->kn_obj)->f_data;
 1419         struct pipe *wpipe;
 1420 
 1421         if ((hint & NOTE_SUBMIT) == 0) {
 1422                 mutex_enter(rpipe->pipe_lock);
 1423         }
 1424         wpipe = rpipe->pipe_peer;
 1425 
 1426         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
 1427                 kn->kn_data = 0;
 1428                 kn->kn_flags |= EV_EOF;
 1429                 if ((hint & NOTE_SUBMIT) == 0) {
 1430                         mutex_exit(rpipe->pipe_lock);
 1431                 }
 1432                 return (1);
 1433         }
 1434         kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
 1435         if (wpipe->pipe_state & PIPE_DIRECTW)
 1436                 kn->kn_data = 0;
 1437 
 1438         if ((hint & NOTE_SUBMIT) == 0) {
 1439                 mutex_exit(rpipe->pipe_lock);
 1440         }
 1441         return (kn->kn_data >= PIPE_BUF);
 1442 }
 1443 
 1444 static const struct filterops pipe_rfiltops =
 1445         { 1, NULL, filt_pipedetach, filt_piperead };
 1446 static const struct filterops pipe_wfiltops =
 1447         { 1, NULL, filt_pipedetach, filt_pipewrite };
 1448 
 1449 static int
 1450 pipe_kqfilter(file_t *fp, struct knote *kn)
 1451 {
 1452         struct pipe *pipe;
 1453         kmutex_t *lock;
 1454 
 1455         pipe = ((file_t *)kn->kn_obj)->f_data;
 1456         lock = pipe->pipe_lock;
 1457 
 1458         mutex_enter(lock);
 1459 
 1460         switch (kn->kn_filter) {
 1461         case EVFILT_READ:
 1462                 kn->kn_fop = &pipe_rfiltops;
 1463                 break;
 1464         case EVFILT_WRITE:
 1465                 kn->kn_fop = &pipe_wfiltops;
 1466                 pipe = pipe->pipe_peer;
 1467                 if (pipe == NULL) {
 1468                         /* Other end of pipe has been closed. */
 1469                         mutex_exit(lock);
 1470                         return (EBADF);
 1471                 }
 1472                 break;
 1473         default:
 1474                 mutex_exit(lock);
 1475                 return (EINVAL);
 1476         }
 1477 
 1478         kn->kn_hook = pipe;
 1479         SLIST_INSERT_HEAD(&pipe->pipe_sel.sel_klist, kn, kn_selnext);
 1480         mutex_exit(lock);
 1481 
 1482         return (0);
 1483 }
 1484 
 1485 /*
 1486  * Handle pipe sysctls.
 1487  */
 1488 SYSCTL_SETUP(sysctl_kern_pipe_setup, "sysctl kern.pipe subtree setup")
 1489 {
 1490 
 1491         sysctl_createv(clog, 0, NULL, NULL,
 1492                        CTLFLAG_PERMANENT,
 1493                        CTLTYPE_NODE, "kern", NULL,
 1494                        NULL, 0, NULL, 0,
 1495                        CTL_KERN, CTL_EOL);
 1496         sysctl_createv(clog, 0, NULL, NULL,
 1497                        CTLFLAG_PERMANENT,
 1498                        CTLTYPE_NODE, "pipe",
 1499                        SYSCTL_DESCR("Pipe settings"),
 1500                        NULL, 0, NULL, 0,
 1501                        CTL_KERN, KERN_PIPE, CTL_EOL);
 1502 
 1503         sysctl_createv(clog, 0, NULL, NULL,
 1504                        CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
 1505                        CTLTYPE_INT, "maxkvasz",
 1506                        SYSCTL_DESCR("Maximum amount of kernel memory to be "
 1507                                     "used for pipes"),
 1508                        NULL, 0, &maxpipekva, 0,
 1509                        CTL_KERN, KERN_PIPE, KERN_PIPE_MAXKVASZ, CTL_EOL);
 1510         sysctl_createv(clog, 0, NULL, NULL,
 1511                        CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
 1512                        CTLTYPE_INT, "maxloankvasz",
 1513                        SYSCTL_DESCR("Limit for direct transfers via page loan"),
 1514                        NULL, 0, &limitpipekva, 0,
 1515                        CTL_KERN, KERN_PIPE, KERN_PIPE_LIMITKVA, CTL_EOL);
 1516         sysctl_createv(clog, 0, NULL, NULL,
 1517                        CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
 1518                        CTLTYPE_INT, "maxbigpipes",
 1519                        SYSCTL_DESCR("Maximum number of \"big\" pipes"),
 1520                        NULL, 0, &maxbigpipes, 0,
 1521                        CTL_KERN, KERN_PIPE, KERN_PIPE_MAXBIGPIPES, CTL_EOL);
 1522         sysctl_createv(clog, 0, NULL, NULL,
 1523                        CTLFLAG_PERMANENT,
 1524                        CTLTYPE_INT, "nbigpipes",
 1525                        SYSCTL_DESCR("Number of \"big\" pipes"),
 1526                        NULL, 0, &nbigpipe, 0,
 1527                        CTL_KERN, KERN_PIPE, KERN_PIPE_NBIGPIPES, CTL_EOL);
 1528         sysctl_createv(clog, 0, NULL, NULL,
 1529                        CTLFLAG_PERMANENT,
 1530                        CTLTYPE_INT, "kvasize",
 1531                        SYSCTL_DESCR("Amount of kernel memory consumed by pipe "
 1532                                     "buffers"),
 1533                        NULL, 0, &amountpipekva, 0,
 1534                        CTL_KERN, KERN_PIPE, KERN_PIPE_KVASIZE, CTL_EOL);
 1535 }

Cache object: 6868220c29ca70808586cd9b317587ff


[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]


This page is part of the FreeBSD/Linux Linux Kernel Cross-Reference, and was automatically generated using a modified version of the LXR engine.