kern_physio.c revision 1.80.2.6
1 /* $NetBSD: kern_physio.c,v 1.80.2.6 2007/08/24 23:28:39 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 1.80.2.6 2007/08/24 23:28:39 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 83 #include <uvm/uvm_extern.h> 84 85 ONCE_DECL(physio_initialized); 86 struct workqueue *physio_workqueue; 87 88 /* 89 * The routines implemented in this file are described in: 90 * Leffler, et al.: The Design and Implementation of the 4.3BSD 91 * UNIX Operating System (Addison Welley, 1989) 92 * on pages 231-233. 93 * 94 * The routines "getphysbuf" and "putphysbuf" steal and return a swap 95 * buffer. Leffler, et al., says that swap buffers are used to do the 96 * I/O, so raw I/O requests don't have to be single-threaded. Of course, 97 * NetBSD doesn't use "swap buffers" -- we have our own memory pool for 98 * buffer descriptors. 99 */ 100 101 /* #define PHYSIO_DEBUG */ 102 #if defined(PHYSIO_DEBUG) 103 #define DPRINTF(a) printf a 104 #else /* defined(PHYSIO_DEBUG) */ 105 #define DPRINTF(a) /* nothing */ 106 #endif /* defined(PHYSIO_DEBUG) */ 107 108 /* abuse these members/flags of struct buf */ 109 #define b_running b_freelistindex 110 #define b_endoffset b_lblkno 111 #define BC_DONTFREE BC_AGE 112 113 /* 114 * allocate a buffer structure for use in physical I/O. 115 */ 116 static struct buf * 117 getphysbuf(void) 118 { 119 struct buf *bp; 120 121 bp = getiobuf(NULL, true); 122 bp->b_error = 0; 123 bp->b_cflags = BC_BUSY; 124 return(bp); 125 } 126 127 /* 128 * get rid of a swap buffer structure which has been used in physical I/O. 129 */ 130 static void 131 putphysbuf(struct buf *bp) 132 { 133 134 if ((bp->b_cflags & BC_DONTFREE) != 0) { 135 return; 136 } 137 138 if (__predict_false((bp->b_cflags & BC_WANTED) != 0)) 139 panic("putphysbuf: private buf wanted"); 140 putiobuf(bp); 141 } 142 143 static void 144 physio_done(struct work *wk, void *dummy) 145 { 146 struct buf *bp = (void *)wk; 147 size_t todo = bp->b_bufsize; 148 size_t done = bp->b_bcount - bp->b_resid; 149 struct buf *mbp = bp->b_private; 150 151 KASSERT(&bp->b_work == wk); 152 KASSERT(bp->b_bcount <= todo); 153 KASSERT(bp->b_resid <= bp->b_bcount); 154 KASSERT((bp->b_flags & B_PHYS) != 0); 155 KASSERT(dummy == NULL); 156 157 vunmapbuf(bp, todo); 158 uvm_vsunlock(bp->b_proc->p_vmspace, bp->b_data, todo); 159 160 mutex_enter(mbp->b_objlock); 161 if (__predict_false(done != todo)) { 162 off_t endoffset = dbtob(bp->b_blkno) + done; 163 164 /* 165 * we got an error or hit EOM. 166 * 167 * we only care about the first one. 168 * ie. the one at the lowest offset. 169 */ 170 171 KASSERT(mbp->b_endoffset != endoffset); 172 DPRINTF(("%s: error=%d at %" PRIu64 " - %" PRIu64 173 ", blkno=%" PRIu64 ", bcount=%d, flags=0x%x\n", 174 __func__, bp->b_error, dbtob(bp->b_blkno), endoffset, 175 bp->b_blkno, bp->b_bcount, bp->b_flags)); 176 177 mbp->b_flags = EIO; 178 if (mbp->b_endoffset == -1 || endoffset < mbp->b_endoffset) { 179 DPRINTF(("%s: mbp=%p, error %d -> %d, endoff %" PRIu64 180 " -> %" PRIu64 "\n", 181 __func__, mbp, 182 mbp->b_error, bp->b_error, 183 mbp->b_endoffset, endoffset)); 184 185 mbp->b_endoffset = endoffset; 186 mbp->b_error = bp->b_error; 187 } 188 } else { 189 KASSERT(bp->b_error == 0); 190 } 191 192 mbp->b_running--; 193 cv_broadcast(&mbp->b_done); 194 mutex_exit(mbp->b_objlock); 195 196 putphysbuf(bp); 197 } 198 199 static void 200 physio_biodone(struct buf *bp) 201 { 202 #if defined(DIAGNOSTIC) 203 struct buf *mbp = bp->b_private; 204 size_t todo = bp->b_bufsize; 205 206 KASSERT(mbp->b_running > 0); 207 KASSERT(bp->b_bcount <= todo); 208 KASSERT(bp->b_resid <= bp->b_bcount); 209 #endif /* defined(DIAGNOSTIC) */ 210 211 workqueue_enqueue(physio_workqueue, &bp->b_work, NULL); 212 } 213 214 static void 215 physio_wait(struct buf *bp, int n) 216 { 217 218 KASSERT(mutex_owned(bp->b_objlock)); 219 220 while ((bp->b_oflags & BO_DONE) == 0 && bp->b_running > n) 221 cv_wait(&bp->b_done, bp->b_objlock); 222 } 223 224 static int 225 physio_init(void) 226 { 227 int error; 228 229 KASSERT(physio_workqueue == NULL); 230 231 error = workqueue_create(&physio_workqueue, "physiod", 232 physio_done, NULL, PRIBIO, IPL_NONE, 0); 233 234 return error; 235 } 236 237 #define PHYSIO_CONCURRENCY 16 /* XXX tune */ 238 239 /* 240 * Do "physical I/O" on behalf of a user. "Physical I/O" is I/O directly 241 * from the raw device to user buffers, and bypasses the buffer cache. 242 * 243 * Comments in brackets are from Leffler, et al.'s pseudo-code implementation. 244 */ 245 int 246 physio(void (*strategy)(struct buf *), struct buf *obp, dev_t dev, int flags, 247 void (*min_phys)(struct buf *), struct uio *uio) 248 { 249 struct iovec *iovp; 250 struct lwp *l = curlwp; 251 struct proc *p = l->l_proc; 252 int i; 253 int error; 254 struct buf *bp = NULL; 255 struct buf *mbp; 256 int concurrency = PHYSIO_CONCURRENCY - 1; 257 258 error = RUN_ONCE(&physio_initialized, physio_init); 259 if (__predict_false(error != 0)) { 260 return error; 261 } 262 263 DPRINTF(("%s: called: off=%" PRIu64 ", resid=%zu\n", 264 __func__, uio->uio_offset, uio->uio_resid)); 265 266 flags &= B_READ | B_WRITE; 267 268 /* Make sure we have a buffer, creating one if necessary. */ 269 if (obp != NULL) { 270 mutex_enter(&bufcache_lock); 271 while (bbusy(obp, false, 0) == EPASSTHROUGH) 272 ; 273 /* Mark it busy, so nobody else will use it. */ 274 obp->b_cflags = BC_DONTFREE; 275 mutex_exit(&bufcache_lock); 276 277 concurrency = 0; /* see "XXXkludge" comment below */ 278 } 279 280 mbp = getphysbuf(); 281 mbp->b_running = 0; 282 mbp->b_endoffset = -1; 283 284 uvm_lwp_hold(l); 285 286 for (i = 0; i < uio->uio_iovcnt; i++) { 287 bool sync = true; 288 289 iovp = &uio->uio_iov[i]; 290 while (iovp->iov_len > 0) { 291 size_t todo; 292 vaddr_t endp; 293 294 mutex_enter(mbp->b_objlock); 295 physio_wait(mbp, sync ? 0 : concurrency); 296 if (mbp->b_error != 0) { 297 goto done_locked; 298 } 299 mutex_exit(mbp->b_objlock); 300 if (obp != NULL) { 301 /* 302 * XXXkludge 303 * some drivers use "obp" as an identifier. 304 */ 305 bp = obp; 306 } else { 307 bp = getphysbuf(); 308 } 309 bp->b_dev = dev; 310 bp->b_proc = p; 311 bp->b_private = mbp; 312 bp->b_vp = NULL; 313 314 /* 315 * [mark the buffer busy for physical I/O] 316 * (i.e. set B_PHYS (because it's an I/O to user 317 * memory, and B_RAW, because B_RAW is to be 318 * "Set by physio for raw transfers.", in addition 319 * to the "busy" and read/write flag.) 320 */ 321 bp->b_oflags = 0; 322 bp->b_cflags = (bp->b_cflags & BC_DONTFREE) | BC_BUSY; 323 bp->b_flags = flags | B_PHYS | B_RAW; 324 bp->b_iodone = physio_biodone; 325 326 /* [set up the buffer for a maximum-sized transfer] */ 327 bp->b_blkno = btodb(uio->uio_offset); 328 if (dbtob(bp->b_blkno) != uio->uio_offset) { 329 error = EINVAL; 330 goto done; 331 } 332 bp->b_bcount = MIN(MAXPHYS, iovp->iov_len); 333 bp->b_data = iovp->iov_base; 334 335 /* 336 * [call minphys to bound the transfer size] 337 * and remember the amount of data to transfer, 338 * for later comparison. 339 */ 340 (*min_phys)(bp); 341 todo = bp->b_bufsize = bp->b_bcount; 342 #if defined(DIAGNOSTIC) 343 if (todo > MAXPHYS) 344 panic("todo(%zu) > MAXPHYS; minphys broken", 345 todo); 346 #endif /* defined(DIAGNOSTIC) */ 347 348 sync = false; 349 endp = (vaddr_t)bp->b_data + todo; 350 if (trunc_page(endp) != endp) { 351 /* 352 * following requests can overlap. 353 * note that uvm_vslock does round_page. 354 */ 355 sync = true; 356 } 357 358 /* 359 * [lock the part of the user address space involved 360 * in the transfer] 361 * Beware vmapbuf(); it clobbers b_data and 362 * saves it in b_saveaddr. However, vunmapbuf() 363 * restores it. 364 */ 365 error = uvm_vslock(p->p_vmspace, bp->b_data, todo, 366 (flags & B_READ) ? VM_PROT_WRITE : VM_PROT_READ); 367 if (error) { 368 goto done; 369 } 370 vmapbuf(bp, todo); 371 372 BIO_SETPRIO(bp, BPRIO_TIMECRITICAL); 373 374 mutex_enter(mbp->b_objlock); 375 mbp->b_running++; 376 mutex_exit(mbp->b_objlock); 377 378 /* [call strategy to start the transfer] */ 379 (*strategy)(bp); 380 bp = NULL; 381 382 iovp->iov_len -= todo; 383 iovp->iov_base = (char *)iovp->iov_base + todo; 384 uio->uio_offset += todo; 385 uio->uio_resid -= todo; 386 } 387 } 388 389 done: 390 mutex_enter(mbp->b_objlock); 391 done_locked: 392 physio_wait(mbp, 0); 393 mutex_exit(mbp->b_objlock); 394 395 if (mbp->b_error != 0) { 396 off_t delta; 397 398 delta = uio->uio_offset - mbp->b_endoffset; 399 KASSERT(delta > 0); 400 uio->uio_resid += delta; 401 /* uio->uio_offset = mbp->b_endoffset; */ 402 } else { 403 KASSERT(mbp->b_endoffset == -1); 404 } 405 if (bp != NULL) { 406 putphysbuf(bp); 407 } 408 if (error == 0) { 409 error = mbp->b_error; 410 } 411 putphysbuf(mbp); 412 413 /* 414 * [clean up the state of the buffer] 415 * Remember if somebody wants it, so we can wake them up below. 416 * Also, if we had to steal it, give it back. 417 */ 418 if (obp != NULL) { 419 KASSERT((obp->b_cflags & BC_BUSY) != 0); 420 KASSERT((obp->b_cflags & BC_DONTFREE) != 0); 421 422 /* 423 * [if another process is waiting for the raw I/O buffer, 424 * wake up processes waiting to do physical I/O; 425 */ 426 mutex_enter(&bufcache_lock); 427 obp->b_cflags &= ~(BC_DONTFREE | BC_BUSY); 428 obp->b_flags &= ~(B_PHYS | B_RAW); 429 obp->b_iodone = NULL; 430 cv_broadcast(&obp->b_busy); 431 mutex_exit(&bufcache_lock); 432 } 433 uvm_lwp_rele(l); 434 435 DPRINTF(("%s: done: off=%" PRIu64 ", resid=%zu\n", 436 __func__, uio->uio_offset, uio->uio_resid)); 437 438 return error; 439 } 440 441 /* 442 * Leffler, et al., says on p. 231: 443 * "The minphys() routine is called by physio() to adjust the 444 * size of each I/O transfer before the latter is passed to 445 * the strategy routine..." 446 * 447 * so, just adjust the buffer's count accounting to MAXPHYS here, 448 * and return the new count; 449 */ 450 void 451 minphys(struct buf *bp) 452 { 453 454 if (bp->b_bcount > MAXPHYS) 455 bp->b_bcount = MAXPHYS; 456 } 457
Indexes created Mon Sep 29 21:09:56 GMT 2025