kern_physio.c revision 1.87.16.1
1 /* $NetBSD: kern_physio.c,v 1.87.16.1 2008/10/19 22:17:27 haad 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 1.87.16.1 2008/10/19 22:17:27 haad 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 96 /* #define PHYSIO_DEBUG */ 97 #if defined(PHYSIO_DEBUG) 98 #define DPRINTF(a) printf a 99 #else /* defined(PHYSIO_DEBUG) */ 100 #define DPRINTF(a) /* nothing */ 101 #endif /* defined(PHYSIO_DEBUG) */ 102 103 struct physio_stat { 104 int ps_running; 105 int ps_error; 106 int ps_failed; 107 off_t ps_endoffset; 108 buf_t *ps_orig_bp; 109 kmutex_t ps_lock; 110 kcondvar_t ps_cv; 111 }; 112 113 static void 114 physio_done(struct work *wk, void *dummy) 115 { 116 struct buf *bp = (void *)wk; 117 size_t todo = bp->b_bufsize; 118 size_t done = bp->b_bcount - bp->b_resid; 119 struct physio_stat *ps = bp->b_private; 120 121 KASSERT(&bp->b_work == wk); 122 KASSERT(bp->b_bcount <= todo); 123 KASSERT(bp->b_resid <= bp->b_bcount); 124 KASSERT((bp->b_flags & B_PHYS) != 0); 125 KASSERT(dummy == NULL); 126 127 vunmapbuf(bp, todo); 128 uvm_vsunlock(bp->b_proc->p_vmspace, bp->b_data, todo); 129 130 mutex_enter(&ps->ps_lock); 131 if (__predict_false(done != todo)) { 132 off_t endoffset = dbtob(bp->b_blkno) + done; 133 134 /* 135 * we got an error or hit EOM. 136 * 137 * we only care about the first one. 138 * ie. the one at the lowest offset. 139 */ 140 141 KASSERT(ps->ps_endoffset != endoffset); 142 DPRINTF(("%s: error=%d at %" PRIu64 " - %" PRIu64 143 ", blkno=%" PRIu64 ", bcount=%d, flags=0x%x\n", 144 __func__, bp->b_error, dbtob(bp->b_blkno), endoffset, 145 bp->b_blkno, bp->b_bcount, bp->b_flags)); 146 147 if (ps->ps_endoffset == -1 || endoffset < ps->ps_endoffset) { 148 DPRINTF(("%s: ps=%p, error %d -> %d, endoff %" PRIu64 149 " -> %" PRIu64 "\n", 150 __func__, ps, 151 ps->ps_error, bp->b_error, 152 ps->ps_endoffset, endoffset)); 153 154 ps->ps_endoffset = endoffset; 155 ps->ps_error = bp->b_error; 156 } 157 ps->ps_failed++; 158 } else { 159 KASSERT(bp->b_error == 0); 160 } 161 162 ps->ps_running--; 163 cv_signal(&ps->ps_cv); 164 mutex_exit(&ps->ps_lock); 165 166 if (bp != ps->ps_orig_bp) 167 putiobuf(bp); 168 } 169 170 static void 171 physio_biodone(struct buf *bp) 172 { 173 #if defined(DIAGNOSTIC) 174 struct physio_stat *ps = bp->b_private; 175 size_t todo = bp->b_bufsize; 176 177 KASSERT(ps->ps_running > 0); 178 KASSERT(bp->b_bcount <= todo); 179 KASSERT(bp->b_resid <= bp->b_bcount); 180 #endif /* defined(DIAGNOSTIC) */ 181 182 workqueue_enqueue(physio_workqueue, &bp->b_work, NULL); 183 } 184 185 static void 186 physio_wait(struct physio_stat *ps, int n) 187 { 188 189 KASSERT(mutex_owned(&ps->ps_lock)); 190 191 while (ps->ps_running > n) 192 cv_wait(&ps->ps_cv, &ps->ps_lock); 193 } 194 195 static int 196 physio_init(void) 197 { 198 int error; 199 200 KASSERT(physio_workqueue == NULL); 201 202 error = workqueue_create(&physio_workqueue, "physiod", 203 physio_done, NULL, PRI_BIO, IPL_BIO, WQ_MPSAFE); 204 205 return error; 206 } 207 208 #define PHYSIO_CONCURRENCY 16 /* XXX tune */ 209 210 /* 211 * Do "physical I/O" on behalf of a user. "Physical I/O" is I/O directly 212 * from the raw device to user buffers, and bypasses the buffer cache. 213 * 214 * Comments in brackets are from Leffler, et al.'s pseudo-code implementation. 215 */ 216 int 217 physio(void (*strategy)(struct buf *), struct buf *obp, dev_t dev, int flags, 218 void (*min_phys)(struct buf *), struct uio *uio) 219 { 220 struct iovec *iovp; 221 struct lwp *l = curlwp; 222 struct proc *p = l->l_proc; 223 int i, error; 224 struct buf *bp = NULL; 225 struct physio_stat *ps; 226 int concurrency = PHYSIO_CONCURRENCY - 1; 227 228 error = RUN_ONCE(&physio_initialized, physio_init); 229 if (__predict_false(error != 0)) { 230 return error; 231 } 232 233 DPRINTF(("%s: called: off=%" PRIu64 ", resid=%zu\n", 234 __func__, uio->uio_offset, uio->uio_resid)); 235 236 flags &= B_READ | B_WRITE; 237 238 if ((ps = kmem_zalloc(sizeof(*ps), KM_SLEEP)) == NULL) 239 return ENOMEM; 240 /* ps->ps_running = 0; */ 241 /* ps->ps_error = 0; */ 242 /* ps->ps_failed = 0; */ 243 ps->ps_orig_bp = obp; 244 ps->ps_endoffset = -1; 245 mutex_init(&ps->ps_lock, MUTEX_DEFAULT, IPL_NONE); 246 cv_init(&ps->ps_cv, "physio"); 247 248 /* Make sure we have a buffer, creating one if necessary. */ 249 if (obp != NULL) { 250 /* [raise the processor priority level to splbio;] */ 251 mutex_enter(&bufcache_lock); 252 /* Mark it busy, so nobody else will use it. */ 253 while (bbusy(obp, false, 0, NULL) == EPASSTHROUGH) 254 ; 255 mutex_exit(&bufcache_lock); 256 concurrency = 0; /* see "XXXkludge" comment below */ 257 } 258 259 uvm_lwp_hold(l); 260 261 for (i = 0; i < uio->uio_iovcnt; i++) { 262 bool sync = true; 263 264 iovp = &uio->uio_iov[i]; 265 while (iovp->iov_len > 0) { 266 size_t todo; 267 vaddr_t endp; 268 269 mutex_enter(&ps->ps_lock); 270 if (ps->ps_failed != 0) { 271 goto done_locked; 272 } 273 physio_wait(ps, sync ? 0 : concurrency); 274 mutex_exit(&ps->ps_lock); 275 if (obp != NULL) { 276 /* 277 * XXXkludge 278 * some drivers use "obp" as an identifier. 279 */ 280 bp = obp; 281 } else { 282 bp = getiobuf(NULL, true); 283 bp->b_cflags = BC_BUSY; 284 } 285 bp->b_dev = dev; 286 bp->b_proc = p; 287 bp->b_private = ps; 288 289 /* 290 * [mark the buffer busy for physical I/O] 291 * (i.e. set B_PHYS (because it's an I/O to user 292 * memory, and B_RAW, because B_RAW is to be 293 * "Set by physio for raw transfers.", in addition 294 * to the "busy" and read/write flag.) 295 */ 296 bp->b_oflags = 0; 297 bp->b_cflags = BC_BUSY; 298 bp->b_flags = flags | B_PHYS | B_RAW; 299 bp->b_iodone = physio_biodone; 300 301 /* [set up the buffer for a maximum-sized transfer] */ 302 bp->b_blkno = btodb(uio->uio_offset); 303 if (dbtob(bp->b_blkno) != uio->uio_offset) { 304 error = EINVAL; 305 goto done; 306 } 307 bp->b_bcount = MIN(MAXPHYS, iovp->iov_len); 308 bp->b_data = iovp->iov_base; 309 310 /* 311 * [call minphys to bound the transfer size] 312 * and remember the amount of data to transfer, 313 * for later comparison. 314 */ 315 (*min_phys)(bp); 316 todo = bp->b_bufsize = bp->b_bcount; 317 #if defined(DIAGNOSTIC) 318 if (todo > MAXPHYS) 319 panic("todo(%zu) > MAXPHYS; minphys broken", 320 todo); 321 #endif /* defined(DIAGNOSTIC) */ 322 323 sync = false; 324 endp = (vaddr_t)bp->b_data + todo; 325 if (trunc_page(endp) != endp) { 326 /* 327 * following requests can overlap. 328 * note that uvm_vslock does round_page. 329 */ 330 sync = true; 331 } 332 333 /* 334 * [lock the part of the user address space involved 335 * in the transfer] 336 * Beware vmapbuf(); it clobbers b_data and 337 * saves it in b_saveaddr. However, vunmapbuf() 338 * restores it. 339 */ 340 error = uvm_vslock(p->p_vmspace, bp->b_data, todo, 341 (flags & B_READ) ? VM_PROT_WRITE : VM_PROT_READ); 342 if (error) { 343 goto done; 344 } 345 vmapbuf(bp, todo); 346 347 BIO_SETPRIO(bp, BPRIO_TIMECRITICAL); 348 349 mutex_enter(&ps->ps_lock); 350 ps->ps_running++; 351 mutex_exit(&ps->ps_lock); 352 353 /* [call strategy to start the transfer] */ 354 (*strategy)(bp); 355 bp = NULL; 356 357 iovp->iov_len -= todo; 358 iovp->iov_base = (char *)iovp->iov_base + todo; 359 uio->uio_offset += todo; 360 uio->uio_resid -= todo; 361 } 362 } 363 364 done: 365 mutex_enter(&ps->ps_lock); 366 done_locked: 367 physio_wait(ps, 0); 368 mutex_exit(&ps->ps_lock); 369 370 if (ps->ps_failed != 0) { 371 off_t delta; 372 373 delta = uio->uio_offset - ps->ps_endoffset; 374 KASSERT(delta > 0); 375 uio->uio_resid += delta; 376 /* uio->uio_offset = ps->ps_endoffset; */ 377 } else { 378 KASSERT(ps->ps_endoffset == -1); 379 } 380 if (bp != NULL && bp != obp) { 381 putiobuf(bp); 382 } 383 if (error == 0) { 384 error = ps->ps_error; 385 } 386 mutex_destroy(&ps->ps_lock); 387 cv_destroy(&ps->ps_cv); 388 kmem_free(ps, sizeof(*ps)); 389 390 /* 391 * [clean up the state of the buffer] 392 * Remember if somebody wants it, so we can wake them up below. 393 * Also, if we had to steal it, give it back. 394 */ 395 if (obp != NULL) { 396 KASSERT((obp->b_cflags & BC_BUSY) != 0); 397 398 /* 399 * [if another process is waiting for the raw I/O buffer, 400 * wake up processes waiting to do physical I/O; 401 */ 402 mutex_enter(&bufcache_lock); 403 obp->b_cflags &= ~(BC_BUSY | BC_WANTED); 404 obp->b_flags &= ~(B_PHYS | B_RAW); 405 obp->b_iodone = NULL; 406 cv_broadcast(&obp->b_busy); 407 mutex_exit(&bufcache_lock); 408 } 409 uvm_lwp_rele(l); 410 411 DPRINTF(("%s: done: off=%" PRIu64 ", resid=%zu\n", 412 __func__, uio->uio_offset, uio->uio_resid)); 413 414 return error; 415 } 416 417 /* 418 * Leffler, et al., says on p. 231: 419 * "The minphys() routine is called by physio() to adjust the 420 * size of each I/O transfer before the latter is passed to 421 * the strategy routine..." 422 * 423 * so, just adjust the buffer's count accounting to MAXPHYS here, 424 * and return the new count; 425 */ 426 void 427 minphys(struct buf *bp) 428 { 429 430 if (bp->b_bcount > MAXPHYS) 431 bp->b_bcount = MAXPHYS; 432 } 433
Indexes created Mon Sep 29 21:09:56 GMT 2025