kern_physio.c revision 1.80.2.5
1 /* $NetBSD: kern_physio.c,v 1.80.2.5 2007/08/19 19:24:53 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 1.80.2.5 2007/08/19 19:24:53 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 B_DONTFREE B_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(); 122 bp->b_error = 0; 123 bp->b_flags = B_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_flags & B_DONTFREE) != 0) { 135 return; 136 } 137 138 if (__predict_false(bp->b_flags & B_WANTED)) 139 panic("putphysbuf: private buf B_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_interlock); 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 if ((mbp->b_flags & B_WANTED) != 0) { 194 mbp->b_flags &= ~B_WANTED; 195 wakeup(mbp); 196 } 197 mutex_exit(&mbp->b_interlock); 198 199 putphysbuf(bp); 200 } 201 202 static void 203 physio_biodone(struct buf *bp) 204 { 205 #if defined(DIAGNOSTIC) 206 struct buf *mbp = bp->b_private; 207 size_t todo = bp->b_bufsize; 208 209 KASSERT(mbp->b_running > 0); 210 KASSERT(bp->b_bcount <= todo); 211 KASSERT(bp->b_resid <= bp->b_bcount); 212 #endif /* defined(DIAGNOSTIC) */ 213 214 workqueue_enqueue(physio_workqueue, &bp->b_work, NULL); 215 } 216 217 static int 218 physio_wait(struct buf *bp, int n, const char *wchan) 219 { 220 int error = 0; 221 222 KASSERT(mutex_owned(&bp->b_interlock)); 223 224 while (bp->b_running > n) { 225 bp->b_flags |= B_WANTED; 226 error = mtsleep(bp, PRIBIO + 1, wchan, 0, &bp->b_interlock); 227 if (error) { 228 break; 229 } 230 } 231 232 return error; 233 } 234 235 static int 236 physio_init(void) 237 { 238 int error; 239 240 KASSERT(physio_workqueue == NULL); 241 242 error = workqueue_create(&physio_workqueue, "physiod", 243 physio_done, NULL, PRIBIO, IPL_NONE, 0); 244 245 return error; 246 } 247 248 #define PHYSIO_CONCURRENCY 16 /* XXX tune */ 249 250 /* 251 * Do "physical I/O" on behalf of a user. "Physical I/O" is I/O directly 252 * from the raw device to user buffers, and bypasses the buffer cache. 253 * 254 * Comments in brackets are from Leffler, et al.'s pseudo-code implementation. 255 */ 256 int 257 physio(void (*strategy)(struct buf *), struct buf *obp, dev_t dev, int flags, 258 void (*min_phys)(struct buf *), struct uio *uio) 259 { 260 struct iovec *iovp; 261 struct lwp *l = curlwp; 262 struct proc *p = l->l_proc; 263 int i; 264 int error; 265 int error2; 266 struct buf *bp = NULL; 267 struct buf *mbp; 268 int concurrency = PHYSIO_CONCURRENCY - 1; 269 270 error = RUN_ONCE(&physio_initialized, physio_init); 271 if (__predict_false(error != 0)) { 272 return error; 273 } 274 275 DPRINTF(("%s: called: off=%" PRIu64 ", resid=%zu\n", 276 __func__, uio->uio_offset, uio->uio_resid)); 277 278 flags &= B_READ | B_WRITE; 279 280 /* Make sure we have a buffer, creating one if necessary. */ 281 if (obp != NULL) { 282 mutex_enter(&obp->b_interlock); 283 284 /* [while the buffer is marked busy] */ 285 while (obp->b_flags & B_BUSY) { 286 /* [mark the buffer wanted] */ 287 obp->b_flags |= B_WANTED; 288 /* [wait until the buffer is available] */ 289 mtsleep(obp, PRIBIO+1, "physbuf", 0, &obp->b_interlock); 290 } 291 292 /* Mark it busy, so nobody else will use it. */ 293 obp->b_flags = B_BUSY | B_DONTFREE; 294 295 /* [lower the priority level] */ 296 mutex_exit(&obp->b_interlock); 297 298 concurrency = 0; /* see "XXXkludge" comment below */ 299 } 300 301 mbp = getphysbuf(); 302 mbp->b_running = 0; 303 mbp->b_endoffset = -1; 304 305 uvm_lwp_hold(l); 306 307 for (i = 0; i < uio->uio_iovcnt; i++) { 308 bool sync = true; 309 310 iovp = &uio->uio_iov[i]; 311 while (iovp->iov_len > 0) { 312 size_t todo; 313 vaddr_t endp; 314 315 mutex_enter(&mbp->b_interlock); 316 if (mbp->b_error != 0) { 317 goto done_locked; 318 } 319 error = physio_wait(mbp, sync ? 0 : concurrency, 320 "physio1"); 321 if (error) { 322 goto done_locked; 323 } 324 mutex_exit(&mbp->b_interlock); 325 if (obp != NULL) { 326 /* 327 * XXXkludge 328 * some drivers use "obp" as an identifier. 329 */ 330 bp = obp; 331 } else { 332 bp = getphysbuf(); 333 } 334 bp->b_dev = dev; 335 bp->b_proc = p; 336 bp->b_private = mbp; 337 bp->b_vp = NULL; 338 339 /* 340 * [mark the buffer busy for physical I/O] 341 * (i.e. set B_PHYS (because it's an I/O to user 342 * memory, and B_RAW, because B_RAW is to be 343 * "Set by physio for raw transfers.", in addition 344 * to the "busy" and read/write flag.) 345 */ 346 bp->b_flags = (bp->b_flags & B_DONTFREE) | 347 B_BUSY | B_PHYS | B_RAW | B_CALL | flags; 348 bp->b_iodone = physio_biodone; 349 350 /* [set up the buffer for a maximum-sized transfer] */ 351 bp->b_blkno = btodb(uio->uio_offset); 352 if (dbtob(bp->b_blkno) != uio->uio_offset) { 353 error = EINVAL; 354 goto done; 355 } 356 bp->b_bcount = MIN(MAXPHYS, iovp->iov_len); 357 bp->b_data = iovp->iov_base; 358 359 /* 360 * [call minphys to bound the transfer size] 361 * and remember the amount of data to transfer, 362 * for later comparison. 363 */ 364 (*min_phys)(bp); 365 todo = bp->b_bufsize = bp->b_bcount; 366 #if defined(DIAGNOSTIC) 367 if (todo > MAXPHYS) 368 panic("todo(%zu) > MAXPHYS; minphys broken", 369 todo); 370 #endif /* defined(DIAGNOSTIC) */ 371 372 sync = false; 373 endp = (vaddr_t)bp->b_data + todo; 374 if (trunc_page(endp) != endp) { 375 /* 376 * following requests can overlap. 377 * note that uvm_vslock does round_page. 378 */ 379 sync = true; 380 } 381 382 /* 383 * [lock the part of the user address space involved 384 * in the transfer] 385 * Beware vmapbuf(); it clobbers b_data and 386 * saves it in b_saveaddr. However, vunmapbuf() 387 * restores it. 388 */ 389 error = uvm_vslock(p->p_vmspace, bp->b_data, todo, 390 (flags & B_READ) ? VM_PROT_WRITE : VM_PROT_READ); 391 if (error) { 392 goto done; 393 } 394 vmapbuf(bp, todo); 395 396 BIO_SETPRIO(bp, BPRIO_TIMECRITICAL); 397 398 mutex_enter(&mbp->b_interlock); 399 mbp->b_running++; 400 mutex_exit(&mbp->b_interlock); 401 402 /* [call strategy to start the transfer] */ 403 (*strategy)(bp); 404 bp = NULL; 405 406 iovp->iov_len -= todo; 407 iovp->iov_base = (char *)iovp->iov_base + todo; 408 uio->uio_offset += todo; 409 uio->uio_resid -= todo; 410 } 411 } 412 413 done: 414 mutex_enter(&mbp->b_interlock); 415 done_locked: 416 error2 = physio_wait(mbp, 0, "physio2"); 417 if (error == 0) { 418 error = error2; 419 } 420 mutex_exit(&mbp->b_interlock); 421 422 if (mbp->b_error != 0) { 423 off_t delta; 424 425 delta = uio->uio_offset - mbp->b_endoffset; 426 KASSERT(delta > 0); 427 uio->uio_resid += delta; 428 /* uio->uio_offset = mbp->b_endoffset; */ 429 } else { 430 KASSERT(mbp->b_endoffset == -1); 431 } 432 if (bp != NULL) { 433 putphysbuf(bp); 434 } 435 if (error == 0) { 436 error = mbp->b_error; 437 } 438 putphysbuf(mbp); 439 440 /* 441 * [clean up the state of the buffer] 442 * Remember if somebody wants it, so we can wake them up below. 443 * Also, if we had to steal it, give it back. 444 */ 445 if (obp != NULL) { 446 KASSERT((obp->b_flags & B_BUSY) != 0); 447 KASSERT((obp->b_flags & B_DONTFREE) != 0); 448 449 /* 450 * [if another process is waiting for the raw I/O buffer, 451 * wake up processes waiting to do physical I/O; 452 */ 453 mutex_enter(&obp->b_interlock); 454 obp->b_flags &= 455 ~(B_BUSY | B_PHYS | B_RAW | B_CALL | B_DONTFREE); 456 if ((obp->b_flags & B_WANTED) != 0) { 457 obp->b_flags &= ~B_WANTED; 458 wakeup(obp); 459 } 460 mutex_exit(&obp->b_interlock); 461 } 462 uvm_lwp_rele(l); 463 464 DPRINTF(("%s: done: off=%" PRIu64 ", resid=%zu\n", 465 __func__, uio->uio_offset, uio->uio_resid)); 466 467 return error; 468 } 469 470 /* 471 * Leffler, et al., says on p. 231: 472 * "The minphys() routine is called by physio() to adjust the 473 * size of each I/O transfer before the latter is passed to 474 * the strategy routine..." 475 * 476 * so, just adjust the buffer's count accounting to MAXPHYS here, 477 * and return the new count; 478 */ 479 void 480 minphys(struct buf *bp) 481 { 482 483 if (bp->b_bcount > MAXPHYS) 484 bp->b_bcount = MAXPHYS; 485 } 486
Indexes created Mon Sep 29 21:09:56 GMT 2025