kern_physio.c revision 1.80.2.7
1 /* $NetBSD: kern_physio.c,v 1.80.2.7 2007/10/09 13:44:27 ad Exp $ */ 2 3 /*- 4 * Copyright (c) 1982, 1986, 1990, 1993 5 * The Regents of the University of California. All rights reserved. 6 * (c) UNIX System Laboratories, Inc. 7 * All or some portions of this file are derived from material licensed 8 * to the University of California by American Telephone and Telegraph 9 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 10 * the permission of UNIX System Laboratories, Inc. 11 * 12 * Redistribution and use in source and binary forms, with or without 13 * modification, are permitted provided that the following conditions 14 * are met: 15 * 1. Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 2. Redistributions in binary form must reproduce the above copyright 18 * notice, this list of conditions and the following disclaimer in the 19 * documentation and/or other materials provided with the distribution. 20 * 3. Neither the name of the University nor the names of its contributors 21 * may be used to endorse or promote products derived from this software 22 * without specific prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 * 36 * @(#)kern_physio.c 8.1 (Berkeley) 6/10/93 37 */ 38 39 /*- 40 * Copyright (c) 1994 Christopher G. Demetriou 41 * 42 * Redistribution and use in source and binary forms, with or without 43 * modification, are permitted provided that the following conditions 44 * are met: 45 * 1. Redistributions of source code must retain the above copyright 46 * notice, this list of conditions and the following disclaimer. 47 * 2. Redistributions in binary form must reproduce the above copyright 48 * notice, this list of conditions and the following disclaimer in the 49 * documentation and/or other materials provided with the distribution. 50 * 3. All advertising materials mentioning features or use of this software 51 * must display the following acknowledgement: 52 * This product includes software developed by the University of 53 * California, Berkeley and its contributors. 54 * 4. Neither the name of the University nor the names of its contributors 55 * may be used to endorse or promote products derived from this software 56 * without specific prior written permission. 57 * 58 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 59 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 60 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 61 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 62 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 63 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 64 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 65 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 66 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 67 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 68 * SUCH DAMAGE. 69 * 70 * @(#)kern_physio.c 8.1 (Berkeley) 6/10/93 71 */ 72 73 #include <sys/cdefs.h> 74 __KERNEL_RCSID(0, "$NetBSD: kern_physio.c,v 1.80.2.7 2007/10/09 13:44:27 ad Exp $"); 75 76 #include <sys/param.h> 77 #include <sys/systm.h> 78 #include <sys/buf.h> 79 #include <sys/proc.h> 80 #include <sys/once.h> 81 #include <sys/workqueue.h> 82 #include <sys/kmem.h> 83 84 #include <uvm/uvm_extern.h> 85 86 ONCE_DECL(physio_initialized); 87 struct workqueue *physio_workqueue; 88 89 /* 90 * The routines implemented in this file are described in: 91 * Leffler, et al.: The Design and Implementation of the 4.3BSD 92 * UNIX Operating System (Addison Welley, 1989) 93 * on pages 231-233. 94 * 95 * The routines "getphysbuf" and "putphysbuf" steal and return a swap 96 * buffer. Leffler, et al., says that swap buffers are used to do the 97 * I/O, so raw I/O requests don't have to be single-threaded. Of course, 98 * NetBSD doesn't use "swap buffers" -- we have our own memory pool for 99 * buffer descriptors. 100 */ 101 102 /* #define PHYSIO_DEBUG */ 103 #if defined(PHYSIO_DEBUG) 104 #define DPRINTF(a) printf a 105 #else /* defined(PHYSIO_DEBUG) */ 106 #define DPRINTF(a) /* nothing */ 107 #endif /* defined(PHYSIO_DEBUG) */ 108 109 /* abuse these members/flags of struct buf */ 110 #define b_running b_freelistindex 111 #define b_endoffset b_lblkno 112 #define BC_DONTFREE BC_AGE 113 114 /* 115 * allocate a buffer structure for use in physical I/O. 116 */ 117 static struct buf * 118 getphysbuf(void) 119 { 120 struct buf *bp; 121 122 bp = getiobuf(NULL, true); 123 bp->b_error = 0; 124 bp->b_cflags = BC_BUSY; 125 return(bp); 126 } 127 128 /* 129 * get rid of a swap buffer structure which has been used in physical I/O. 130 */ 131 static void 132 putphysbuf(struct buf *bp) 133 { 134 135 if ((bp->b_cflags & BC_DONTFREE) != 0) { 136 return; 137 } 138 139 if (__predict_false((bp->b_cflags & BC_WANTED) != 0)) 140 panic("putphysbuf: private buf wanted"); 141 putiobuf(bp); 142 } 143 144 static void 145 physio_done(struct work *wk, void *dummy) 146 { 147 struct buf *bp = (void *)wk; 148 size_t todo = bp->b_bufsize; 149 size_t done = bp->b_bcount - bp->b_resid; 150 struct physio_stat *ps = bp->b_private; 151 152 KASSERT(&bp->b_work == wk); 153 KASSERT(bp->b_bcount <= todo); 154 KASSERT(bp->b_resid <= bp->b_bcount); 155 KASSERT((bp->b_flags & B_PHYS) != 0); 156 KASSERT(dummy == NULL); 157 158 vunmapbuf(bp, todo); 159 uvm_vsunlock(bp->b_proc->p_vmspace, bp->b_data, todo); 160 161 mutex_enter(mbp->b_objlock); 162 if (__predict_false(done != todo)) { 163 off_t endoffset = dbtob(bp->b_blkno) + done; 164 165 /* 166 * we got an error or hit EOM. 167 * 168 * we only care about the first one. 169 * ie. the one at the lowest offset. 170 */ 171 172 KASSERT(mbp->b_endoffset != endoffset); 173 DPRINTF(("%s: error=%d at %" PRIu64 " - %" PRIu64 174 ", blkno=%" PRIu64 ", bcount=%d, flags=0x%x\n", 175 __func__, bp->b_error, dbtob(bp->b_blkno), endoffset, 176 bp->b_blkno, bp->b_bcount, bp->b_flags)); 177 178 mbp->b_flags = EIO; 179 if (mbp->b_endoffset == -1 || endoffset < mbp->b_endoffset) { 180 DPRINTF(("%s: mbp=%p, error %d -> %d, endoff %" PRIu64 181 " -> %" PRIu64 "\n", 182 __func__, mbp, 183 mbp->b_error, bp->b_error, 184 mbp->b_endoffset, endoffset)); 185 186 mbp->b_endoffset = endoffset; 187 mbp->b_error = bp->b_error; 188 } 189 } else { 190 KASSERT(bp->b_error == 0); 191 } 192 193 mbp->b_running--; 194 cv_broadcast(&mbp->b_done); 195 mutex_exit(mbp->b_objlock); 196 197 putphysbuf(bp); 198 } 199 200 static void 201 physio_biodone(struct buf *bp) 202 { 203 #if defined(DIAGNOSTIC) 204 struct physio_stat *ps = bp->b_private; 205 size_t todo = bp->b_bufsize; 206 207 KASSERT(ps->ps_running > 0); 208 KASSERT(bp->b_bcount <= todo); 209 KASSERT(bp->b_resid <= bp->b_bcount); 210 #endif /* defined(DIAGNOSTIC) */ 211 212 workqueue_enqueue(physio_workqueue, &bp->b_work, NULL); 213 } 214 215 static void 216 physio_wait(struct buf *bp, int n) 217 { 218 219 KASSERT(mutex_owned(bp->b_objlock)); 220 221 while ((bp->b_oflags & BO_DONE) == 0 && bp->b_running > n) 222 cv_wait(&bp->b_done, bp->b_objlock); 223 } 224 225 static int 226 physio_init(void) 227 { 228 int error; 229 230 KASSERT(physio_workqueue == NULL); 231 232 error = workqueue_create(&physio_workqueue, "physiod", 233 physio_done, NULL, PRIBIO, IPL_NONE, 0); 234 235 return error; 236 } 237 238 #define PHYSIO_CONCURRENCY 16 /* XXX tune */ 239 240 /* 241 * Do "physical I/O" on behalf of a user. "Physical I/O" is I/O directly 242 * from the raw device to user buffers, and bypasses the buffer cache. 243 * 244 * Comments in brackets are from Leffler, et al.'s pseudo-code implementation. 245 */ 246 int 247 physio(void (*strategy)(struct buf *), struct buf *obp, dev_t dev, int flags, 248 void (*min_phys)(struct buf *), struct uio *uio) 249 { 250 struct iovec *iovp; 251 struct lwp *l = curlwp; 252 struct proc *p = l->l_proc; 253 int i; 254 int error; 255 struct buf *bp = NULL; 256 struct buf *mbp; 257 int concurrency = PHYSIO_CONCURRENCY - 1; 258 259 error = RUN_ONCE(&physio_initialized, physio_init); 260 if (__predict_false(error != 0)) { 261 return error; 262 } 263 264 DPRINTF(("%s: called: off=%" PRIu64 ", resid=%zu\n", 265 __func__, uio->uio_offset, uio->uio_resid)); 266 267 flags &= B_READ | B_WRITE; 268 269 if ((ps = kmem_zalloc(sizeof(*ps), KM_SLEEP)) == NULL) 270 return ENOMEM; 271 /* ps->ps_running = 0; */ 272 /* ps->ps_error = 0; */ 273 /* ps->ps_failed = 0; */ 274 ps->ps_endoffset = -1; 275 mutex_init(&ps->ps_lock, MUTEX_DEFAULT, IPL_NONE); 276 cv_init(&ps->ps_cv, "physio"); 277 278 /* Make sure we have a buffer, creating one if necessary. */ 279 if (obp != NULL) { 280 mutex_enter(&bufcache_lock); 281 while (bbusy(obp, false, 0) == EPASSTHROUGH) 282 ; 283 /* Mark it busy, so nobody else will use it. */ 284 obp->b_cflags = BC_DONTFREE; 285 mutex_exit(&bufcache_lock); 286 287 concurrency = 0; /* see "XXXkludge" comment below */ 288 } 289 290 uvm_lwp_hold(l); 291 292 for (i = 0; i < uio->uio_iovcnt; i++) { 293 bool sync = true; 294 295 iovp = &uio->uio_iov[i]; 296 while (iovp->iov_len > 0) { 297 size_t todo; 298 vaddr_t endp; 299 300 mutex_enter(mbp->b_objlock); 301 physio_wait(mbp, sync ? 0 : concurrency); 302 if (mbp->b_error != 0) { 303 goto done_locked; 304 } 305 mutex_exit(mbp->b_objlock); 306 if (obp != NULL) { 307 /* 308 * XXXkludge 309 * some drivers use "obp" as an identifier. 310 */ 311 bp = obp; 312 } else { 313 bp = getphysbuf(); 314 } 315 bp->b_dev = dev; 316 bp->b_proc = p; 317 bp->b_private = ps; 318 bp->b_vp = NULL; 319 320 /* 321 * [mark the buffer busy for physical I/O] 322 * (i.e. set B_PHYS (because it's an I/O to user 323 * memory, and B_RAW, because B_RAW is to be 324 * "Set by physio for raw transfers.", in addition 325 * to the "busy" and read/write flag.) 326 */ 327 bp->b_oflags = 0; 328 bp->b_cflags = (bp->b_cflags & BC_DONTFREE) | BC_BUSY; 329 bp->b_flags = flags | B_PHYS | B_RAW; 330 bp->b_iodone = physio_biodone; 331 332 /* [set up the buffer for a maximum-sized transfer] */ 333 bp->b_blkno = btodb(uio->uio_offset); 334 if (dbtob(bp->b_blkno) != uio->uio_offset) { 335 error = EINVAL; 336 goto done; 337 } 338 bp->b_bcount = MIN(MAXPHYS, iovp->iov_len); 339 bp->b_data = iovp->iov_base; 340 341 /* 342 * [call minphys to bound the transfer size] 343 * and remember the amount of data to transfer, 344 * for later comparison. 345 */ 346 (*min_phys)(bp); 347 todo = bp->b_bufsize = bp->b_bcount; 348 #if defined(DIAGNOSTIC) 349 if (todo > MAXPHYS) 350 panic("todo(%zu) > MAXPHYS; minphys broken", 351 todo); 352 #endif /* defined(DIAGNOSTIC) */ 353 354 sync = false; 355 endp = (vaddr_t)bp->b_data + todo; 356 if (trunc_page(endp) != endp) { 357 /* 358 * following requests can overlap. 359 * note that uvm_vslock does round_page. 360 */ 361 sync = true; 362 } 363 364 /* 365 * [lock the part of the user address space involved 366 * in the transfer] 367 * Beware vmapbuf(); it clobbers b_data and 368 * saves it in b_saveaddr. However, vunmapbuf() 369 * restores it. 370 */ 371 error = uvm_vslock(p->p_vmspace, bp->b_data, todo, 372 (flags & B_READ) ? VM_PROT_WRITE : VM_PROT_READ); 373 if (error) { 374 goto done; 375 } 376 vmapbuf(bp, todo); 377 378 BIO_SETPRIO(bp, BPRIO_TIMECRITICAL); 379 380 mutex_enter(mbp->b_objlock); 381 mbp->b_running++; 382 mutex_exit(mbp->b_objlock); 383 384 /* [call strategy to start the transfer] */ 385 (*strategy)(bp); 386 bp = NULL; 387 388 iovp->iov_len -= todo; 389 iovp->iov_base = (char *)iovp->iov_base + todo; 390 uio->uio_offset += todo; 391 uio->uio_resid -= todo; 392 } 393 } 394 395 done: 396 mutex_enter(mbp->b_objlock); 397 done_locked: 398 physio_wait(mbp, 0); 399 mutex_exit(mbp->b_objlock); 400 401 if (mbp->b_error != 0) { 402 off_t delta; 403 404 delta = uio->uio_offset - mbp->b_endoffset; 405 KASSERT(delta > 0); 406 uio->uio_resid += delta; 407 /* uio->uio_offset = mbp->b_endoffset; */ 408 } else { 409 KASSERT(mbp->b_endoffset == -1); 410 } 411 if (bp != NULL) { 412 putphysbuf(bp); 413 } 414 if (error == 0) { 415 error = mbp->b_error; 416 } 417 putphysbuf(mbp); 418 419 /* 420 * [clean up the state of the buffer] 421 * Remember if somebody wants it, so we can wake them up below. 422 * Also, if we had to steal it, give it back. 423 */ 424 if (obp != NULL) { 425 KASSERT((obp->b_cflags & BC_BUSY) != 0); 426 KASSERT((obp->b_cflags & BC_DONTFREE) != 0); 427 428 /* 429 * [if another process is waiting for the raw I/O buffer, 430 * wake up processes waiting to do physical I/O; 431 */ 432 mutex_enter(&bufcache_lock); 433 obp->b_cflags &= ~(BC_DONTFREE | BC_BUSY); 434 obp->b_flags &= ~(B_PHYS | B_RAW); 435 obp->b_iodone = NULL; 436 cv_broadcast(&obp->b_busy); 437 mutex_exit(&bufcache_lock); 438 } 439 uvm_lwp_rele(l); 440 441 DPRINTF(("%s: done: off=%" PRIu64 ", resid=%zu\n", 442 __func__, uio->uio_offset, uio->uio_resid)); 443 444 return error; 445 } 446 447 /* 448 * Leffler, et al., says on p. 231: 449 * "The minphys() routine is called by physio() to adjust the 450 * size of each I/O transfer before the latter is passed to 451 * the strategy routine..." 452 * 453 * so, just adjust the buffer's count accounting to MAXPHYS here, 454 * and return the new count; 455 */ 456 void 457 minphys(struct buf *bp) 458 { 459 460 if (bp->b_bcount > MAXPHYS) 461 bp->b_bcount = MAXPHYS; 462 } 463
Indexes created Mon Sep 29 21:09:56 GMT 2025