| 1 | /* $NetBSD: sys_pipe.c,v 1.140 2014/09/05 09:20:59 matt 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 1.140 2014/09/05 09:20:59 matt 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 *, bool); |
| 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 | if (flags & ~(O_CLOEXEC|O_NONBLOCK|O_NOSIGPIPE)) |
| 255 | return EINVAL; |
| 256 | p = curproc; |
| 257 | rpipe = wpipe = NULL; |
| 258 | if ((error = pipe_create(&rpipe, pipe_rd_cache)) || |
| 259 | (error = pipe_create(&wpipe, pipe_wr_cache))) { |
| 260 | goto free2; |
| 261 | } |
| 262 | rpipe->pipe_lock = mutex_obj_alloc(MUTEX_DEFAULT, IPL_NONE); |
| 263 | wpipe->pipe_lock = rpipe->pipe_lock; |
| 264 | mutex_obj_hold(wpipe->pipe_lock); |
| 265 | |
| 266 | error = fd_allocfile(&rf, &fd); |
| 267 | if (error) |
| 268 | goto free2; |
| 269 | retval[0] = fd; |
| 270 | |
| 271 | error = fd_allocfile(&wf, &fd); |
| 272 | if (error) |
| 273 | goto free3; |
| 274 | retval[1] = fd; |
| 275 | |
| 276 | rf->f_flag = FREAD | flags; |
| 277 | rf->f_type = DTYPE_PIPE; |
| 278 | rf->f_pipe = rpipe; |
| 279 | rf->f_ops = &pipeops; |
| 280 | fd_set_exclose(l, (int)retval[0], (flags & O_CLOEXEC) != 0); |
| 281 | |
| 282 | wf->f_flag = FWRITE | flags; |
| 283 | wf->f_type = DTYPE_PIPE; |
| 284 | wf->f_pipe = wpipe; |
| 285 | wf->f_ops = &pipeops; |
| 286 | fd_set_exclose(l, (int)retval[1], (flags & O_CLOEXEC) != 0); |
| 287 | |
| 288 | rpipe->pipe_peer = wpipe; |
| 289 | wpipe->pipe_peer = rpipe; |
| 290 | |
| 291 | fd_affix(p, rf, (int)retval[0]); |
| 292 | fd_affix(p, wf, (int)retval[1]); |
| 293 | return (0); |
| 294 | free3: |
| 295 | fd_abort(p, rf, (int)retval[0]); |
| 296 | free2: |
| 297 | pipeclose(wpipe); |
| 298 | pipeclose(rpipe); |
| 299 | |
| 300 | return (error); |
| 301 | } |
| 302 | |
| 303 | /* |
| 304 | * Allocate kva for pipe circular buffer, the space is pageable |
| 305 | * This routine will 'realloc' the size of a pipe safely, if it fails |
| 306 | * it will retain the old buffer. |
| 307 | * If it fails it will return ENOMEM. |
| 308 | */ |
| 309 | static int |
| 310 | pipespace(struct pipe *pipe, int size) |
| 311 | { |
| 312 | void *buffer; |
| 313 | |
| 314 | /* |
| 315 | * Allocate pageable virtual address space. Physical memory is |
| 316 | * allocated on demand. |
| 317 | */ |
| 318 | if (size == PIPE_SIZE && pipe->pipe_kmem != 0) { |
| 319 | buffer = (void *)pipe->pipe_kmem; |
| 320 | } else { |
| 321 | buffer = (void *)uvm_km_alloc(kernel_map, round_page(size), |
| 322 | 0, UVM_KMF_PAGEABLE); |
| 323 | if (buffer == NULL) |
| 324 | return (ENOMEM); |
| 325 | atomic_add_int(&amountpipekva, size); |
| 326 | } |
| 327 | |
| 328 | /* free old resources if we're resizing */ |
| 329 | pipe_free_kmem(pipe); |
| 330 | pipe->pipe_buffer.buffer = buffer; |
| 331 | pipe->pipe_buffer.size = size; |
| 332 | pipe->pipe_buffer.in = 0; |
| 333 | pipe->pipe_buffer.out = 0; |
| 334 | pipe->pipe_buffer.cnt = 0; |
| 335 | return (0); |
| 336 | } |
| 337 | |
| 338 | /* |
| 339 | * Initialize and allocate VM and memory for pipe. |
| 340 | */ |
| 341 | static int |
| 342 | pipe_create(struct pipe **pipep, pool_cache_t cache) |
| 343 | { |
| 344 | struct pipe *pipe; |
| 345 | int error; |
| 346 | |
| 347 | pipe = pool_cache_get(cache, PR_WAITOK); |
| 348 | KASSERT(pipe != NULL); |
| 349 | *pipep = pipe; |
| 350 | error = 0; |
| 351 | getnanotime(&pipe->pipe_btime); |
| 352 | pipe->pipe_atime = pipe->pipe_mtime = pipe->pipe_btime; |
| 353 | pipe->pipe_lock = NULL; |
| 354 | if (cache == pipe_rd_cache) { |
| 355 | error = pipespace(pipe, PIPE_SIZE); |
| 356 | } else { |
| 357 | pipe->pipe_buffer.buffer = NULL; |
| 358 | pipe->pipe_buffer.size = 0; |
| 359 | pipe->pipe_buffer.in = 0; |
| 360 | pipe->pipe_buffer.out = 0; |
| 361 | pipe->pipe_buffer.cnt = 0; |
| 362 | } |
| 363 | return error; |
| 364 | } |
| 365 | |
| 366 | /* |
| 367 | * Lock a pipe for I/O, blocking other access |
| 368 | * Called with pipe spin lock held. |
| 369 | */ |
| 370 | static int |
| 371 | pipelock(struct pipe *pipe, bool catch_p) |
| 372 | { |
| 373 | int error; |
| 374 | |
| 375 | KASSERT(mutex_owned(pipe->pipe_lock)); |
| 376 | |
| 377 | while (pipe->pipe_state & PIPE_LOCKFL) { |
| 378 | pipe->pipe_state |= PIPE_LWANT; |
| 379 | if (catch_p) { |
| 380 | error = cv_wait_sig(&pipe->pipe_lkcv, pipe->pipe_lock); |
| 381 | if (error != 0) |
| 382 | return error; |
| 383 | } else |
| 384 | cv_wait(&pipe->pipe_lkcv, pipe->pipe_lock); |
| 385 | } |
| 386 | |
| 387 | pipe->pipe_state |= PIPE_LOCKFL; |
| 388 | |
| 389 | return 0; |
| 390 | } |
| 391 | |
| 392 | /* |
| 393 | * unlock a pipe I/O lock |
| 394 | */ |
| 395 | static inline void |
| 396 | pipeunlock(struct pipe *pipe) |
| 397 | { |
| 398 | |
| 399 | KASSERT(pipe->pipe_state & PIPE_LOCKFL); |
| 400 | |
| 401 | pipe->pipe_state &= ~PIPE_LOCKFL; |
| 402 | if (pipe->pipe_state & PIPE_LWANT) { |
| 403 | pipe->pipe_state &= ~PIPE_LWANT; |
| 404 | cv_broadcast(&pipe->pipe_lkcv); |
| 405 | } |
| 406 | } |
| 407 | |
| 408 | /* |
| 409 | * Select/poll wakup. This also sends SIGIO to peer connected to |
| 410 | * 'sigpipe' side of pipe. |
| 411 | */ |
| 412 | static void |
| 413 | pipeselwakeup(struct pipe *selp, struct pipe *sigp, int code) |
| 414 | { |
| 415 | int band; |
| 416 | |
| 417 | switch (code) { |
| 418 | case POLL_IN: |
| 419 | band = POLLIN|POLLRDNORM; |
| 420 | break; |
| 421 | case POLL_OUT: |
| 422 | band = POLLOUT|POLLWRNORM; |
| 423 | break; |
| 424 | case POLL_HUP: |
| 425 | band = POLLHUP; |
| 426 | break; |
| 427 | case POLL_ERR: |
| 428 | band = POLLERR; |
| 429 | break; |
| 430 | default: |
| 431 | band = 0; |
| 432 | #ifdef DIAGNOSTIC |
| 433 | printf("bad siginfo code %d in pipe notification.\n" , code); |
| 434 | #endif |
| 435 | break; |
| 436 | } |
| 437 | |
| 438 | selnotify(&selp->pipe_sel, band, NOTE_SUBMIT); |
| 439 | |
| 440 | if (sigp == NULL || (sigp->pipe_state & PIPE_ASYNC) == 0) |
| 441 | return; |
| 442 | |
| 443 | fownsignal(sigp->pipe_pgid, SIGIO, code, band, selp); |
| 444 | } |
| 445 | |
| 446 | static int |
| 447 | pipe_read(file_t *fp, off_t *offset, struct uio *uio, kauth_cred_t cred, |
| 448 | int flags) |
| 449 | { |
| 450 | struct pipe *rpipe = fp->f_pipe; |
| 451 | struct pipebuf *bp = &rpipe->pipe_buffer; |
| 452 | kmutex_t *lock = rpipe->pipe_lock; |
| 453 | int error; |
| 454 | size_t nread = 0; |
| 455 | size_t size; |
| 456 | size_t ocnt; |
| 457 | unsigned int wakeup_state = 0; |
| 458 | |
| 459 | mutex_enter(lock); |
| 460 | ++rpipe->pipe_busy; |
| 461 | ocnt = bp->cnt; |
| 462 | |
| 463 | again: |
| 464 | error = pipelock(rpipe, true); |
| 465 | if (error) |
| 466 | goto unlocked_error; |
| 467 | |
| 468 | while (uio->uio_resid) { |
| 469 | /* |
| 470 | * Normal pipe buffer receive. |
| 471 | */ |
| 472 | if (bp->cnt > 0) { |
| 473 | size = bp->size - bp->out; |
| 474 | if (size > bp->cnt) |
| 475 | size = bp->cnt; |
| 476 | if (size > uio->uio_resid) |
| 477 | size = uio->uio_resid; |
| 478 | |
| 479 | mutex_exit(lock); |
| 480 | error = uiomove((char *)bp->buffer + bp->out, size, uio); |
| 481 | mutex_enter(lock); |
| 482 | if (error) |
| 483 | break; |
| 484 | |
| 485 | bp->out += size; |
| 486 | if (bp->out >= bp->size) |
| 487 | bp->out = 0; |
| 488 | |
| 489 | bp->cnt -= size; |
| 490 | |
| 491 | /* |
| 492 | * If there is no more to read in the pipe, reset |
| 493 | * its pointers to the beginning. This improves |
| 494 | * cache hit stats. |
| 495 | */ |
| 496 | if (bp->cnt == 0) { |
| 497 | bp->in = 0; |
| 498 | bp->out = 0; |
| 499 | } |
| 500 | nread += size; |
| 501 | continue; |
| 502 | } |
| 503 | |
| 504 | #ifndef PIPE_NODIRECT |
| 505 | if ((rpipe->pipe_state & PIPE_DIRECTR) != 0) { |
| 506 | struct pipemapping * const rmap = &rpipe->pipe_map; |
| 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 = rmap->cnt; |
| 516 | if (size > uio->uio_resid) |
| 517 | size = uio->uio_resid; |
| 518 | |
| 519 | va = (char *)rmap->kva + rmap->pos; |
| 520 | gen = rmap->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 | rmap->pos += size; |
| 534 | rmap->cnt -= size; |
| 535 | if (rmap->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 | struct pipemapping * const wmap = &wpipe->pipe_map; |
| 638 | const vsize_t len = ptoa(npages); |
| 639 | |
| 640 | atomic_add_int(&amountpipekva, len); |
| 641 | wmap->kva = uvm_km_alloc(kernel_map, len, 0, |
| 642 | UVM_KMF_COLORMATCH | UVM_KMF_VAONLY | UVM_KMF_WAITVA); |
| 643 | if (wmap->kva == 0) { |
| 644 | atomic_add_int(&amountpipekva, -len); |
| 645 | return (ENOMEM); |
| 646 | } |
| 647 | |
| 648 | wmap->npages = npages; |
| 649 | wmap->pgs = kmem_alloc(npages * sizeof(struct vm_page *), KM_SLEEP); |
| 650 | return (0); |
| 651 | } |
| 652 | |
| 653 | /* |
| 654 | * Free resources allocated for loan transfer. |
| 655 | */ |
| 656 | static void |
| 657 | pipe_loan_free(struct pipe *wpipe) |
| 658 | { |
| 659 | struct pipemapping * const wmap = &wpipe->pipe_map; |
| 660 | const vsize_t len = ptoa(wmap->npages); |
| 661 | |
| 662 | uvm_emap_remove(wmap->kva, len); /* XXX */ |
| 663 | uvm_km_free(kernel_map, wmap->kva, len, UVM_KMF_VAONLY); |
| 664 | wmap->kva = 0; |
| 665 | atomic_add_int(&amountpipekva, -len); |
| 666 | kmem_free(wmap->pgs, wmap->npages * sizeof(struct vm_page *)); |
| 667 | wmap->pgs = NULL; |
| 668 | #if 0 |
| 669 | wmap->npages = 0; |
| 670 | wmap->pos = 0; |
| 671 | wmap->cnt = 0; |
| 672 | #endif |
| 673 | } |
| 674 | |
| 675 | /* |
| 676 | * NetBSD direct write, using uvm_loan() mechanism. |
| 677 | * This implements the pipe buffer write mechanism. Note that only |
| 678 | * a direct write OR a normal pipe write can be pending at any given time. |
| 679 | * If there are any characters in the pipe buffer, the direct write will |
| 680 | * be deferred until the receiving process grabs all of the bytes from |
| 681 | * the pipe buffer. Then the direct mapping write is set-up. |
| 682 | * |
| 683 | * Called with the long-term pipe lock held. |
| 684 | */ |
| 685 | static int |
| 686 | pipe_direct_write(file_t *fp, struct pipe *wpipe, struct uio *uio) |
| 687 | { |
| 688 | struct pipemapping * const wmap = &wpipe->pipe_map; |
| 689 | kmutex_t * const lock = wpipe->pipe_lock; |
| 690 | struct vm_page **pgs; |
| 691 | vaddr_t bbase, base, bend; |
| 692 | vsize_t blen, bcnt; |
| 693 | int error, npages; |
| 694 | voff_t bpos; |
| 695 | u_int starting_color; |
| 696 | |
| 697 | KASSERT(mutex_owned(wpipe->pipe_lock)); |
| 698 | KASSERT(wmap->cnt == 0); |
| 699 | |
| 700 | mutex_exit(lock); |
| 701 | |
| 702 | /* |
| 703 | * Handle first PIPE_CHUNK_SIZE bytes of buffer. Deal with buffers |
| 704 | * not aligned to PAGE_SIZE. |
| 705 | */ |
| 706 | bbase = (vaddr_t)uio->uio_iov->iov_base; |
| 707 | base = trunc_page(bbase); |
| 708 | bend = round_page(bbase + uio->uio_iov->iov_len); |
| 709 | blen = bend - base; |
| 710 | bpos = bbase - base; |
| 711 | |
| 712 | if (blen > PIPE_DIRECT_CHUNK) { |
| 713 | blen = PIPE_DIRECT_CHUNK; |
| 714 | bend = base + blen; |
| 715 | bcnt = PIPE_DIRECT_CHUNK - bpos; |
| 716 | } else { |
| 717 | bcnt = uio->uio_iov->iov_len; |
| 718 | } |
| 719 | npages = atop(blen); |
| 720 | starting_color = atop(base) & uvmexp.colormask; |
| 721 | |
| 722 | /* |
| 723 | * Free the old kva if we need more pages than we have |
| 724 | * allocated. |
| 725 | */ |
| 726 | if (wmap->kva != 0 && starting_color + npages > wmap->npages) |
| 727 | pipe_loan_free(wpipe); |
| 728 | |
| 729 | /* Allocate new kva. */ |
| 730 | if (wmap->kva == 0) { |
| 731 | error = pipe_loan_alloc(wpipe, starting_color + npages); |
| 732 | if (error) { |
| 733 | mutex_enter(lock); |
| 734 | return (error); |
| 735 | } |
| 736 | } |
| 737 | |
| 738 | /* Loan the write buffer memory from writer process */ |
| 739 | pgs = wmap->pgs + starting_color; |
| 740 | error = uvm_loan(&uio->uio_vmspace->vm_map, base, blen, |
| 741 | pgs, UVM_LOAN_TOPAGE); |
| 742 | if (error) { |
| 743 | pipe_loan_free(wpipe); |
| 744 | mutex_enter(lock); |
| 745 | return (ENOMEM); /* so that caller fallback to ordinary write */ |
| 746 | } |
| 747 | |
| 748 | /* Enter the loaned pages to KVA, produce new emap generation number. */ |
| 749 | uvm_emap_enter(wmap->kva + ptoa(starting_color), pgs, npages); |
| 750 | wmap->egen = uvm_emap_produce(); |
| 751 | |
| 752 | /* Now we can put the pipe in direct write mode */ |
| 753 | wmap->pos = bpos + ptoa(starting_color); |
| 754 | wmap->cnt = bcnt; |
| 755 | |
| 756 | /* |
| 757 | * But before we can let someone do a direct read, we |
| 758 | * have to wait until the pipe is drained. Release the |
| 759 | * pipe lock while we wait. |
| 760 | */ |
| 761 | mutex_enter(lock); |
| 762 | wpipe->pipe_state |= PIPE_DIRECTW; |
| 763 | pipeunlock(wpipe); |
| 764 | |
| 765 | while (error == 0 && wpipe->pipe_buffer.cnt > 0) { |
| 766 | cv_broadcast(&wpipe->pipe_rcv); |
| 767 | error = cv_wait_sig(&wpipe->pipe_wcv, lock); |
| 768 | if (error == 0 && wpipe->pipe_state & PIPE_EOF) |
| 769 | error = EPIPE; |
| 770 | } |
| 771 | |
| 772 | /* Pipe is drained; next read will off the direct buffer */ |
| 773 | wpipe->pipe_state |= PIPE_DIRECTR; |
| 774 | |
| 775 | /* Wait until the reader is done */ |
| 776 | while (error == 0 && (wpipe->pipe_state & PIPE_DIRECTR)) { |
| 777 | cv_broadcast(&wpipe->pipe_rcv); |
| 778 | pipeselwakeup(wpipe, wpipe, POLL_IN); |
| 779 | error = cv_wait_sig(&wpipe->pipe_wcv, lock); |
| 780 | if (error == 0 && wpipe->pipe_state & PIPE_EOF) |
| 781 | error = EPIPE; |
| 782 | } |
| 783 | |
| 784 | /* Take pipe out of direct write mode */ |
| 785 | wpipe->pipe_state &= ~(PIPE_DIRECTW | PIPE_DIRECTR); |
| 786 | |
| 787 | /* Acquire the pipe lock and cleanup */ |
| 788 | (void)pipelock(wpipe, false); |
| 789 | mutex_exit(lock); |
| 790 | |
| 791 | if (pgs != NULL) { |
| 792 | /* XXX: uvm_emap_remove */ |
| 793 | uvm_unloan(pgs, npages, UVM_LOAN_TOPAGE); |
| 794 | } |
| 795 | if (error || amountpipekva > maxpipekva) |
| 796 | pipe_loan_free(wpipe); |
| 797 | |
| 798 | mutex_enter(lock); |
| 799 | if (error) { |
| 800 | pipeselwakeup(wpipe, wpipe, POLL_ERR); |
| 801 | |
| 802 | /* |
| 803 | * If nothing was read from what we offered, return error |
| 804 | * straight on. Otherwise update uio resid first. Caller |
| 805 | * will deal with the error condition, returning short |
| 806 | * write, error, or restarting the write(2) as appropriate. |
| 807 | */ |
| 808 | if (wmap->cnt == bcnt) { |
| 809 | wmap->cnt = 0; |
| 810 | cv_broadcast(&wpipe->pipe_wcv); |
| 811 | return (error); |
| 812 | } |
| 813 | |
| 814 | bcnt -= wpipe->cnt; |
| 815 | } |
| 816 | |
| 817 | uio->uio_resid -= bcnt; |
| 818 | /* uio_offset not updated, not set/used for write(2) */ |
| 819 | uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + bcnt; |
| 820 | uio->uio_iov->iov_len -= bcnt; |
| 821 | if (uio->uio_iov->iov_len == 0) { |
| 822 | uio->uio_iov++; |
| 823 | uio->uio_iovcnt--; |
| 824 | } |
| 825 | |
| 826 | wmap->cnt = 0; |
| 827 | return (error); |
| 828 | } |
| 829 | #endif /* !PIPE_NODIRECT */ |
| 830 | |
| 831 | static int |
| 832 | pipe_write(file_t *fp, off_t *offset, struct uio *uio, kauth_cred_t cred, |
| 833 | int flags) |
| 834 | { |
| 835 | struct pipe *wpipe, *rpipe; |
| 836 | struct pipebuf *bp; |
| 837 | kmutex_t *lock; |
| 838 | int error; |
| 839 | unsigned int wakeup_state = 0; |
| 840 | |
| 841 | /* We want to write to our peer */ |
| 842 | rpipe = fp->f_pipe; |
| 843 | lock = rpipe->pipe_lock; |
| 844 | error = 0; |
| 845 | |
| 846 | mutex_enter(lock); |
| 847 | wpipe = rpipe->pipe_peer; |
| 848 | |
| 849 | /* |
| 850 | * Detect loss of pipe read side, issue SIGPIPE if lost. |
| 851 | */ |
| 852 | if (wpipe == NULL || (wpipe->pipe_state & PIPE_EOF) != 0) { |
| 853 | mutex_exit(lock); |
| 854 | return EPIPE; |
| 855 | } |
| 856 | ++wpipe->pipe_busy; |
| 857 | |
| 858 | /* Aquire the long-term pipe lock */ |
| 859 | if ((error = pipelock(wpipe, true)) != 0) { |
| 860 | --wpipe->pipe_busy; |
| 861 | if (wpipe->pipe_busy == 0) { |
| 862 | wpipe->pipe_state &= ~PIPE_RESTART; |
| 863 | cv_broadcast(&wpipe->pipe_draincv); |
| 864 | } |
| 865 | mutex_exit(lock); |
| 866 | return (error); |
| 867 | } |
| 868 | |
| 869 | bp = &wpipe->pipe_buffer; |
| 870 | |
| 871 | /* |
| 872 | * If it is advantageous to resize the pipe buffer, do so. |
| 873 | */ |
| 874 | if ((uio->uio_resid > PIPE_SIZE) && |
| 875 | (nbigpipe < maxbigpipes) && |
| 876 | #ifndef PIPE_NODIRECT |
| 877 | (wpipe->pipe_state & PIPE_DIRECTW) == 0 && |
| 878 | #endif |
| 879 | (bp->size <= PIPE_SIZE) && (bp->cnt == 0)) { |
| 880 | |
| 881 | if (pipespace(wpipe, BIG_PIPE_SIZE) == 0) |
| 882 | atomic_inc_uint(&nbigpipe); |
| 883 | } |
| 884 | |
| 885 | while (uio->uio_resid) { |
| 886 | size_t space; |
| 887 | |
| 888 | #ifndef PIPE_NODIRECT |
| 889 | /* |
| 890 | * Pipe buffered writes cannot be coincidental with |
| 891 | * direct writes. Also, only one direct write can be |
| 892 | * in progress at any one time. We wait until the currently |
| 893 | * executing direct write is completed before continuing. |
| 894 | * |
| 895 | * We break out if a signal occurs or the reader goes away. |
| 896 | */ |
| 897 | while (error == 0 && wpipe->pipe_state & PIPE_DIRECTW) { |
| 898 | cv_broadcast(&wpipe->pipe_rcv); |
| 899 | pipeunlock(wpipe); |
| 900 | error = cv_wait_sig(&wpipe->pipe_wcv, lock); |
| 901 | (void)pipelock(wpipe, false); |
| 902 | if (wpipe->pipe_state & PIPE_EOF) |
| 903 | error = EPIPE; |
| 904 | } |
| 905 | if (error) |
| 906 | break; |
| 907 | |
| 908 | /* |
| 909 | * If the transfer is large, we can gain performance if |
| 910 | * we do process-to-process copies directly. |
| 911 | * If the write is non-blocking, we don't use the |
| 912 | * direct write mechanism. |
| 913 | * |
| 914 | * The direct write mechanism will detect the reader going |
| 915 | * away on us. |
| 916 | */ |
| 917 | if ((uio->uio_iov->iov_len >= PIPE_MINDIRECT) && |
| 918 | (fp->f_flag & FNONBLOCK) == 0 && |
| 919 | (wmap->kva || (amountpipekva < limitpipekva))) { |
| 920 | error = pipe_direct_write(fp, wpipe, uio); |
| 921 | |
| 922 | /* |
| 923 | * Break out if error occurred, unless it's ENOMEM. |
| 924 | * ENOMEM means we failed to allocate some resources |
| 925 | * for direct write, so we just fallback to ordinary |
| 926 | * write. If the direct write was successful, |
| 927 | * process rest of data via ordinary write. |
| 928 | */ |
| 929 | if (error == 0) |
| 930 | continue; |
| 931 | |
| 932 | if (error != ENOMEM) |
| 933 | break; |
| 934 | } |
| 935 | #endif /* PIPE_NODIRECT */ |
| 936 | |
| 937 | space = bp->size - bp->cnt; |
| 938 | |
| 939 | /* Writes of size <= PIPE_BUF must be atomic. */ |
| 940 | if ((space < uio->uio_resid) && (uio->uio_resid <= PIPE_BUF)) |
| 941 | space = 0; |
| 942 | |
| 943 | if (space > 0) { |
| 944 | int size; /* Transfer size */ |
| 945 | int segsize; /* first segment to transfer */ |
| 946 | |
| 947 | /* |
| 948 | * Transfer size is minimum of uio transfer |
| 949 | * and free space in pipe buffer. |
| 950 | */ |
| 951 | if (space > uio->uio_resid) |
| 952 | size = uio->uio_resid; |
| 953 | else |
| 954 | size = space; |
| 955 | /* |
| 956 | * First segment to transfer is minimum of |
| 957 | * transfer size and contiguous space in |
| 958 | * pipe buffer. If first segment to transfer |
| 959 | * is less than the transfer size, we've got |
| 960 | * a wraparound in the buffer. |
| 961 | */ |
| 962 | segsize = bp->size - bp->in; |
| 963 | if (segsize > size) |
| 964 | segsize = size; |
| 965 | |
| 966 | /* Transfer first segment */ |
| 967 | mutex_exit(lock); |
| 968 | error = uiomove((char *)bp->buffer + bp->in, segsize, |
| 969 | uio); |
| 970 | |
| 971 | if (error == 0 && segsize < size) { |
| 972 | /* |
| 973 | * Transfer remaining part now, to |
| 974 | * support atomic writes. Wraparound |
| 975 | * happened. |
| 976 | */ |
| 977 | KASSERT(bp->in + segsize == bp->size); |
| 978 | error = uiomove(bp->buffer, |
| 979 | size - segsize, uio); |
| 980 | } |
| 981 | mutex_enter(lock); |
| 982 | if (error) |
| 983 | break; |
| 984 | |
| 985 | bp->in += size; |
| 986 | if (bp->in >= bp->size) { |
| 987 | KASSERT(bp->in == size - segsize + bp->size); |
| 988 | bp->in = size - segsize; |
| 989 | } |
| 990 | |
| 991 | bp->cnt += size; |
| 992 | KASSERT(bp->cnt <= bp->size); |
| 993 | wakeup_state = 0; |
| 994 | } else { |
| 995 | /* |
| 996 | * If the "read-side" has been blocked, wake it up now. |
| 997 | */ |
| 998 | cv_broadcast(&wpipe->pipe_rcv); |
| 999 | |
| 1000 | /* |
| 1001 | * Don't block on non-blocking I/O. |
| 1002 | */ |
| 1003 | if (fp->f_flag & FNONBLOCK) { |
| 1004 | error = EAGAIN; |
| 1005 | break; |
| 1006 | } |
| 1007 | |
| 1008 | /* |
| 1009 | * We have no more space and have something to offer, |
| 1010 | * wake up select/poll. |
| 1011 | */ |
| 1012 | if (bp->cnt) |
| 1013 | pipeselwakeup(wpipe, wpipe, POLL_IN); |
| 1014 | |
| 1015 | if (wakeup_state & PIPE_RESTART) { |
| 1016 | error = ERESTART; |
| 1017 | break; |
| 1018 | } |
| 1019 | |
| 1020 | pipeunlock(wpipe); |
| 1021 | error = cv_wait_sig(&wpipe->pipe_wcv, lock); |
| 1022 | (void)pipelock(wpipe, false); |
| 1023 | if (error != 0) |
| 1024 | break; |
| 1025 | /* |
| 1026 | * If read side wants to go away, we just issue a signal |
| 1027 | * to ourselves. |
| 1028 | */ |
| 1029 | if (wpipe->pipe_state & PIPE_EOF) { |
| 1030 | error = EPIPE; |
| 1031 | break; |
| 1032 | } |
| 1033 | wakeup_state = wpipe->pipe_state; |
| 1034 | } |
| 1035 | } |
| 1036 | |
| 1037 | --wpipe->pipe_busy; |
| 1038 | if (wpipe->pipe_busy == 0) { |
| 1039 | wpipe->pipe_state &= ~PIPE_RESTART; |
| 1040 | cv_broadcast(&wpipe->pipe_draincv); |
| 1041 | } |
| 1042 | if (bp->cnt > 0) { |
| 1043 | cv_broadcast(&wpipe->pipe_rcv); |
| 1044 | } |
| 1045 | |
| 1046 | /* |
| 1047 | * Don't return EPIPE if I/O was successful |
| 1048 | */ |
| 1049 | if (error == EPIPE && bp->cnt == 0 && uio->uio_resid == 0) |
| 1050 | error = 0; |
| 1051 | |
| 1052 | if (error == 0) |
| 1053 | getnanotime(&wpipe->pipe_mtime); |
| 1054 | |
| 1055 | /* |
| 1056 | * We have something to offer, wake up select/poll. |
| 1057 | * wmap->cnt is always 0 in this point (direct write |
| 1058 | * is only done synchronously), so check only wpipe->pipe_buffer.cnt |
| 1059 | */ |
| 1060 | if (bp->cnt) |
| 1061 | pipeselwakeup(wpipe, wpipe, POLL_IN); |
| 1062 | |
| 1063 | /* |
| 1064 | * Arrange for next read(2) to do a signal. |
| 1065 | */ |
| 1066 | wpipe->pipe_state |= PIPE_SIGNALR; |
| 1067 | |
| 1068 | pipeunlock(wpipe); |
| 1069 | mutex_exit(lock); |
| 1070 | return (error); |
| 1071 | } |
| 1072 | |
| 1073 | /* |
| 1074 | * We implement a very minimal set of ioctls for compatibility with sockets. |
| 1075 | */ |
| 1076 | int |
| 1077 | pipe_ioctl(file_t *fp, u_long cmd, void *data) |
| 1078 | { |
| 1079 | struct pipe *pipe = fp->f_pipe; |
| 1080 | kmutex_t *lock = pipe->pipe_lock; |
| 1081 | |
| 1082 | switch (cmd) { |
| 1083 | |
| 1084 | case FIONBIO: |
| 1085 | return (0); |
| 1086 | |
| 1087 | case FIOASYNC: |
| 1088 | mutex_enter(lock); |
| 1089 | if (*(int *)data) { |
| 1090 | pipe->pipe_state |= PIPE_ASYNC; |
| 1091 | } else { |
| 1092 | pipe->pipe_state &= ~PIPE_ASYNC; |
| 1093 | } |
| 1094 | mutex_exit(lock); |
| 1095 | return (0); |
| 1096 | |
| 1097 | case FIONREAD: |
| 1098 | mutex_enter(lock); |
| 1099 | #ifndef PIPE_NODIRECT |
| 1100 | if (pipe->pipe_state & PIPE_DIRECTW) |
| 1101 | *(int *)data = pipe->pipe_map.cnt; |
| 1102 | else |
| 1103 | #endif |
| 1104 | *(int *)data = pipe->pipe_buffer.cnt; |
| 1105 | mutex_exit(lock); |
| 1106 | return (0); |
| 1107 | |
| 1108 | case FIONWRITE: |
| 1109 | /* Look at other side */ |
| 1110 | pipe = pipe->pipe_peer; |
| 1111 | mutex_enter(lock); |
| 1112 | #ifndef PIPE_NODIRECT |
| 1113 | if (pipe->pipe_state & PIPE_DIRECTW) |
| 1114 | *(int *)data = pipe->pipe_map.cnt; |
| 1115 | else |
| 1116 | #endif |
| 1117 | *(int *)data = pipe->pipe_buffer.cnt; |
| 1118 | mutex_exit(lock); |
| 1119 | return (0); |
| 1120 | |
| 1121 | case FIONSPACE: |
| 1122 | /* Look at other side */ |
| 1123 | pipe = pipe->pipe_peer; |
| 1124 | mutex_enter(lock); |
| 1125 | #ifndef PIPE_NODIRECT |
| 1126 | /* |
| 1127 | * If we're in direct-mode, we don't really have a |
| 1128 | * send queue, and any other write will block. Thus |
| 1129 | * zero seems like the best answer. |
| 1130 | */ |
| 1131 | if (pipe->pipe_state & PIPE_DIRECTW) |
| 1132 | *(int *)data = 0; |
| 1133 | else |
| 1134 | #endif |
| 1135 | *(int *)data = pipe->pipe_buffer.size - |
| 1136 | pipe->pipe_buffer.cnt; |
| 1137 | mutex_exit(lock); |
| 1138 | return (0); |
| 1139 | |
| 1140 | case TIOCSPGRP: |
| 1141 | case FIOSETOWN: |
| 1142 | return fsetown(&pipe->pipe_pgid, cmd, data); |
| 1143 | |
| 1144 | case TIOCGPGRP: |
| 1145 | case FIOGETOWN: |
| 1146 | return fgetown(pipe->pipe_pgid, cmd, data); |
| 1147 | |
| 1148 | } |
| 1149 | return (EPASSTHROUGH); |
| 1150 | } |
| 1151 | |
| 1152 | int |
| 1153 | pipe_poll(file_t *fp, int events) |
| 1154 | { |
| 1155 | struct pipe *rpipe = fp->f_pipe; |
| 1156 | struct pipe *wpipe; |
| 1157 | int eof = 0; |
| 1158 | int revents = 0; |
| 1159 | |
| 1160 | mutex_enter(rpipe->pipe_lock); |
| 1161 | wpipe = rpipe->pipe_peer; |
| 1162 | |
| 1163 | if (events & (POLLIN | POLLRDNORM)) |
| 1164 | if ((rpipe->pipe_buffer.cnt > 0) || |
| 1165 | #ifndef PIPE_NODIRECT |
| 1166 | (rpipe->pipe_state & PIPE_DIRECTR) || |
| 1167 | #endif |
| 1168 | (rpipe->pipe_state & PIPE_EOF)) |
| 1169 | revents |= events & (POLLIN | POLLRDNORM); |
| 1170 | |
| 1171 | eof |= (rpipe->pipe_state & PIPE_EOF); |
| 1172 | |
| 1173 | if (wpipe == NULL) |
| 1174 | revents |= events & (POLLOUT | POLLWRNORM); |
| 1175 | else { |
| 1176 | if (events & (POLLOUT | POLLWRNORM)) |
| 1177 | if ((wpipe->pipe_state & PIPE_EOF) || ( |
| 1178 | #ifndef PIPE_NODIRECT |
| 1179 | (wpipe->pipe_state & PIPE_DIRECTW) == 0 && |
| 1180 | #endif |
| 1181 | (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF)) |
| 1182 | revents |= events & (POLLOUT | POLLWRNORM); |
| 1183 | |
| 1184 | eof |= (wpipe->pipe_state & PIPE_EOF); |
| 1185 | } |
| 1186 | |
| 1187 | if (wpipe == NULL || eof) |
| 1188 | revents |= POLLHUP; |
| 1189 | |
| 1190 | if (revents == 0) { |
| 1191 | if (events & (POLLIN | POLLRDNORM)) |
| 1192 | selrecord(curlwp, &rpipe->pipe_sel); |
| 1193 | |
| 1194 | if (events & (POLLOUT | POLLWRNORM)) |
| 1195 | selrecord(curlwp, &wpipe->pipe_sel); |
| 1196 | } |
| 1197 | mutex_exit(rpipe->pipe_lock); |
| 1198 | |
| 1199 | return (revents); |
| 1200 | } |
| 1201 | |
| 1202 | static int |
| 1203 | pipe_stat(file_t *fp, struct stat *ub) |
| 1204 | { |
| 1205 | struct pipe *pipe = fp->f_pipe; |
| 1206 | |
| 1207 | mutex_enter(pipe->pipe_lock); |
| 1208 | memset(ub, 0, sizeof(*ub)); |
| 1209 | ub->st_mode = S_IFIFO | S_IRUSR | S_IWUSR; |
| 1210 | ub->st_blksize = pipe->pipe_buffer.size; |
| 1211 | if (ub->st_blksize == 0 && pipe->pipe_peer) |
| 1212 | ub->st_blksize = pipe->pipe_peer->pipe_buffer.size; |
| 1213 | ub->st_size = pipe->pipe_buffer.cnt; |
| 1214 | ub->st_blocks = (ub->st_size) ? 1 : 0; |
| 1215 | ub->st_atimespec = pipe->pipe_atime; |
| 1216 | ub->st_mtimespec = pipe->pipe_mtime; |
| 1217 | ub->st_ctimespec = ub->st_birthtimespec = pipe->pipe_btime; |
| 1218 | ub->st_uid = kauth_cred_geteuid(fp->f_cred); |
| 1219 | ub->st_gid = kauth_cred_getegid(fp->f_cred); |
| 1220 | |
| 1221 | /* |
| 1222 | * Left as 0: st_dev, st_ino, st_nlink, st_rdev, st_flags, st_gen. |
| 1223 | * XXX (st_dev, st_ino) should be unique. |
| 1224 | */ |
| 1225 | mutex_exit(pipe->pipe_lock); |
| 1226 | return 0; |
| 1227 | } |
| 1228 | |
| 1229 | static int |
| 1230 | pipe_close(file_t *fp) |
| 1231 | { |
| 1232 | struct pipe *pipe = fp->f_pipe; |
| 1233 | |
| 1234 | fp->f_pipe = NULL; |
| 1235 | pipeclose(pipe); |
| 1236 | return (0); |
| 1237 | } |
| 1238 | |
| 1239 | static void |
| 1240 | pipe_restart(file_t *fp) |
| 1241 | { |
| 1242 | struct pipe *pipe = fp->f_pipe; |
| 1243 | |
| 1244 | /* |
| 1245 | * Unblock blocked reads/writes in order to allow close() to complete. |
| 1246 | * System calls return ERESTART so that the fd is revalidated. |
| 1247 | * (Partial writes return the transfer length.) |
| 1248 | */ |
| 1249 | mutex_enter(pipe->pipe_lock); |
| 1250 | pipe->pipe_state |= PIPE_RESTART; |
| 1251 | /* Wakeup both cvs, maybe we only need one, but maybe there are some |
| 1252 | * other paths where wakeup is needed, and it saves deciding which! */ |
| 1253 | cv_broadcast(&pipe->pipe_rcv); |
| 1254 | cv_broadcast(&pipe->pipe_wcv); |
| 1255 | mutex_exit(pipe->pipe_lock); |
| 1256 | } |
| 1257 | |
| 1258 | static void |
| 1259 | pipe_free_kmem(struct pipe *pipe) |
| 1260 | { |
| 1261 | |
| 1262 | if (pipe->pipe_buffer.buffer != NULL) { |
| 1263 | if (pipe->pipe_buffer.size > PIPE_SIZE) { |
| 1264 | atomic_dec_uint(&nbigpipe); |
| 1265 | } |
| 1266 | if (pipe->pipe_buffer.buffer != (void *)pipe->pipe_kmem) { |
| 1267 | uvm_km_free(kernel_map, |
| 1268 | (vaddr_t)pipe->pipe_buffer.buffer, |
| 1269 | pipe->pipe_buffer.size, UVM_KMF_PAGEABLE); |
| 1270 | atomic_add_int(&amountpipekva, |
| 1271 | -pipe->pipe_buffer.size); |
| 1272 | } |
| 1273 | pipe->pipe_buffer.buffer = NULL; |
| 1274 | } |
| 1275 | #ifndef PIPE_NODIRECT |
| 1276 | if (pipe->pipe_map.kva != 0) { |
| 1277 | pipe_loan_free(pipe); |
| 1278 | pipe->pipe_map.cnt = 0; |
| 1279 | pipe->pipe_map.pos = 0; |
| 1280 | pipe->pipe_map.npages = 0; |
| 1281 | } |
| 1282 | #endif /* !PIPE_NODIRECT */ |
| 1283 | } |
| 1284 | |
| 1285 | /* |
| 1286 | * Shutdown the pipe. |
| 1287 | */ |
| 1288 | static void |
| 1289 | pipeclose(struct pipe *pipe) |
| 1290 | { |
| 1291 | kmutex_t *lock; |
| 1292 | struct pipe *ppipe; |
| 1293 | |
| 1294 | if (pipe == NULL) |
| 1295 | return; |
| 1296 | |
| 1297 | KASSERT(cv_is_valid(&pipe->pipe_rcv)); |
| 1298 | KASSERT(cv_is_valid(&pipe->pipe_wcv)); |
| 1299 | KASSERT(cv_is_valid(&pipe->pipe_draincv)); |
| 1300 | KASSERT(cv_is_valid(&pipe->pipe_lkcv)); |
| 1301 | |
| 1302 | lock = pipe->pipe_lock; |
| 1303 | if (lock == NULL) |
| 1304 | /* Must have failed during create */ |
| 1305 | goto free_resources; |
| 1306 | |
| 1307 | mutex_enter(lock); |
| 1308 | pipeselwakeup(pipe, pipe, POLL_HUP); |
| 1309 | |
| 1310 | /* |
| 1311 | * If the other side is blocked, wake it up saying that |
| 1312 | * we want to close it down. |
| 1313 | */ |
| 1314 | pipe->pipe_state |= PIPE_EOF; |
| 1315 | if (pipe->pipe_busy) { |
| 1316 | while (pipe->pipe_busy) { |
| 1317 | cv_broadcast(&pipe->pipe_wcv); |
| 1318 | cv_wait_sig(&pipe->pipe_draincv, lock); |
| 1319 | } |
| 1320 | } |
| 1321 | |
| 1322 | /* |
| 1323 | * Disconnect from peer. |
| 1324 | */ |
| 1325 | if ((ppipe = pipe->pipe_peer) != NULL) { |
| 1326 | pipeselwakeup(ppipe, ppipe, POLL_HUP); |
| 1327 | ppipe->pipe_state |= PIPE_EOF; |
| 1328 | cv_broadcast(&ppipe->pipe_rcv); |
| 1329 | ppipe->pipe_peer = NULL; |
| 1330 | } |
| 1331 | |
| 1332 | /* |
| 1333 | * Any knote objects still left in the list are |
| 1334 | * the one attached by peer. Since no one will |
| 1335 | * traverse this list, we just clear it. |
| 1336 | */ |
| 1337 | SLIST_INIT(&pipe->pipe_sel.sel_klist); |
| 1338 | |
| 1339 | KASSERT((pipe->pipe_state & PIPE_LOCKFL) == 0); |
| 1340 | mutex_exit(lock); |
| 1341 | mutex_obj_free(lock); |
| 1342 | |
| 1343 | /* |
| 1344 | * Free resources. |
| 1345 | */ |
| 1346 | free_resources: |
| 1347 | pipe->pipe_pgid = 0; |
| 1348 | pipe->pipe_state = PIPE_SIGNALR; |
| 1349 | pipe_free_kmem(pipe); |
| 1350 | if (pipe->pipe_kmem != 0) { |
| 1351 | pool_cache_put(pipe_rd_cache, pipe); |
| 1352 | } else { |
| 1353 | pool_cache_put(pipe_wr_cache, pipe); |
| 1354 | } |
| 1355 | } |
| 1356 | |
| 1357 | static void |
| 1358 | filt_pipedetach(struct knote *kn) |
| 1359 | { |
| 1360 | struct pipe *pipe; |
| 1361 | kmutex_t *lock; |
| 1362 | |
| 1363 | pipe = ((file_t *)kn->kn_obj)->f_pipe; |
| 1364 | lock = pipe->pipe_lock; |
| 1365 | |
| 1366 | mutex_enter(lock); |
| 1367 | |
| 1368 | switch(kn->kn_filter) { |
| 1369 | case EVFILT_WRITE: |
| 1370 | /* Need the peer structure, not our own. */ |
| 1371 | pipe = pipe->pipe_peer; |
| 1372 | |
| 1373 | /* If reader end already closed, just return. */ |
| 1374 | if (pipe == NULL) { |
| 1375 | mutex_exit(lock); |
| 1376 | return; |
| 1377 | } |
| 1378 | |
| 1379 | break; |
| 1380 | default: |
| 1381 | /* Nothing to do. */ |
| 1382 | break; |
| 1383 | } |
| 1384 | |
| 1385 | KASSERT(kn->kn_hook == pipe); |
| 1386 | SLIST_REMOVE(&pipe->pipe_sel.sel_klist, kn, knote, kn_selnext); |
| 1387 | mutex_exit(lock); |
| 1388 | } |
| 1389 | |
| 1390 | static int |
| 1391 | filt_piperead(struct knote *kn, long hint) |
| 1392 | { |
| 1393 | struct pipe *rpipe = ((file_t *)kn->kn_obj)->f_pipe; |
| 1394 | struct pipe *wpipe; |
| 1395 | |
| 1396 | if ((hint & NOTE_SUBMIT) == 0) { |
| 1397 | mutex_enter(rpipe->pipe_lock); |
| 1398 | } |
| 1399 | wpipe = rpipe->pipe_peer; |
| 1400 | kn->kn_data = rpipe->pipe_buffer.cnt; |
| 1401 | |
| 1402 | if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW)) |
| 1403 | kn->kn_data = rpipe->pipe_map.cnt; |
| 1404 | |
| 1405 | if ((rpipe->pipe_state & PIPE_EOF) || |
| 1406 | (wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) { |
| 1407 | kn->kn_flags |= EV_EOF; |
| 1408 | if ((hint & NOTE_SUBMIT) == 0) { |
| 1409 | mutex_exit(rpipe->pipe_lock); |
| 1410 | } |
| 1411 | return (1); |
| 1412 | } |
| 1413 | |
| 1414 | if ((hint & NOTE_SUBMIT) == 0) { |
| 1415 | mutex_exit(rpipe->pipe_lock); |
| 1416 | } |
| 1417 | return (kn->kn_data > 0); |
| 1418 | } |
| 1419 | |
| 1420 | static int |
| 1421 | filt_pipewrite(struct knote *kn, long hint) |
| 1422 | { |
| 1423 | struct pipe *rpipe = ((file_t *)kn->kn_obj)->f_pipe; |
| 1424 | struct pipe *wpipe; |
| 1425 | |
| 1426 | if ((hint & NOTE_SUBMIT) == 0) { |
| 1427 | mutex_enter(rpipe->pipe_lock); |
| 1428 | } |
| 1429 | wpipe = rpipe->pipe_peer; |
| 1430 | |
| 1431 | if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) { |
| 1432 | kn->kn_data = 0; |
| 1433 | kn->kn_flags |= EV_EOF; |
| 1434 | if ((hint & NOTE_SUBMIT) == 0) { |
| 1435 | mutex_exit(rpipe->pipe_lock); |
| 1436 | } |
| 1437 | return (1); |
| 1438 | } |
| 1439 | kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt; |
| 1440 | if (wpipe->pipe_state & PIPE_DIRECTW) |
| 1441 | kn->kn_data = 0; |
| 1442 | |
| 1443 | if ((hint & NOTE_SUBMIT) == 0) { |
| 1444 | mutex_exit(rpipe->pipe_lock); |
| 1445 | } |
| 1446 | return (kn->kn_data >= PIPE_BUF); |
| 1447 | } |
| 1448 | |
| 1449 | static const struct filterops pipe_rfiltops = |
| 1450 | { 1, NULL, filt_pipedetach, filt_piperead }; |
| 1451 | static const struct filterops pipe_wfiltops = |
| 1452 | { 1, NULL, filt_pipedetach, filt_pipewrite }; |
| 1453 | |
| 1454 | static int |
| 1455 | pipe_kqfilter(file_t *fp, struct knote *kn) |
| 1456 | { |
| 1457 | struct pipe *pipe; |
| 1458 | kmutex_t *lock; |
| 1459 | |
| 1460 | pipe = ((file_t *)kn->kn_obj)->f_pipe; |
| 1461 | lock = pipe->pipe_lock; |
| 1462 | |
| 1463 | mutex_enter(lock); |
| 1464 | |
| 1465 | switch (kn->kn_filter) { |
| 1466 | case EVFILT_READ: |
| 1467 | kn->kn_fop = &pipe_rfiltops; |
| 1468 | break; |
| 1469 | case EVFILT_WRITE: |
| 1470 | kn->kn_fop = &pipe_wfiltops; |
| 1471 | pipe = pipe->pipe_peer; |
| 1472 | if (pipe == NULL) { |
| 1473 | /* Other end of pipe has been closed. */ |
| 1474 | mutex_exit(lock); |
| 1475 | return (EBADF); |
| 1476 | } |
| 1477 | break; |
| 1478 | default: |
| 1479 | mutex_exit(lock); |
| 1480 | return (EINVAL); |
| 1481 | } |
| 1482 | |
| 1483 | kn->kn_hook = pipe; |
| 1484 | SLIST_INSERT_HEAD(&pipe->pipe_sel.sel_klist, kn, kn_selnext); |
| 1485 | mutex_exit(lock); |
| 1486 | |
| 1487 | return (0); |
| 1488 | } |
| 1489 | |
| 1490 | /* |
| 1491 | * Handle pipe sysctls. |
| 1492 | */ |
| 1493 | SYSCTL_SETUP(sysctl_kern_pipe_setup, "sysctl kern.pipe subtree setup" ) |
| 1494 | { |
| 1495 | |
| 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 | } |
| 1536 | |