1 /* $NetBSD: kern_physio.c,v 1.104 2026/01/04 01:36:43 riastradh 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 1.104 2026/01/04 01:36:43 riastradh Exp $"); 75 76 #include <sys/param.h> 77 #include <sys/types.h> 78 79 #include <sys/buf.h> 80 #include <sys/conf.h> 81 #include <sys/kmem.h> 82 #include <sys/once.h> 83 #include <sys/proc.h> 84 #include <sys/sdt.h> 85 #include <sys/systm.h> 86 #include <sys/workqueue.h> 87 88 #include <uvm/uvm_extern.h> 89 90 ONCE_DECL(physio_initialized); 91 struct workqueue *physio_workqueue; 92 93 int physio_concurrency = 16; 94 95 /* #define PHYSIO_DEBUG */ 96 #if defined(PHYSIO_DEBUG) 97 #define DPRINTF(a) printf a 98 #else /* defined(PHYSIO_DEBUG) */ 99 #define DPRINTF(a) /* nothing */ 100 #endif /* defined(PHYSIO_DEBUG) */ 101 102 struct physio_stat { 103 int ps_running; 104 int ps_error; 105 int ps_failed; 106 off_t ps_endoffset; 107 size_t ps_resid; 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 bool is_iobuf; 121 122 KASSERT(&bp->b_work == wk); 123 KASSERT(bp->b_bcount <= todo); 124 KASSERT(bp->b_resid <= bp->b_bcount); 125 KASSERT((bp->b_flags & B_PHYS) != 0); 126 KASSERT(dummy == NULL); 127 128 vunmapbuf(bp, todo); 129 uvm_vsunlock(bp->b_proc->p_vmspace, bp->b_data, todo); 130 131 mutex_enter(&ps->ps_lock); 132 is_iobuf = (bp != ps->ps_orig_bp); 133 if (__predict_false(done != todo)) { 134 off_t endoffset = dbtob(bp->b_blkno) + done; 135 136 /* 137 * we got an error or hit EOM. 138 * 139 * we only care about the first one. 140 * ie. the one at the lowest offset. 141 */ 142 143 KASSERT(ps->ps_endoffset != endoffset); 144 DPRINTF(("%s: error=%d at %" PRIu64 " - %" PRIu64 145 ", blkno=%" PRIu64 ", bcount=%d, flags=0x%x\n", 146 __func__, bp->b_error, dbtob(bp->b_blkno), endoffset, 147 bp->b_blkno, bp->b_bcount, bp->b_flags)); 148 149 if (ps->ps_endoffset == -1 || endoffset < ps->ps_endoffset) { 150 DPRINTF(("%s: ps=%p, error %d -> %d, endoff %" PRIu64 151 " -> %" PRIu64 "\n", 152 __func__, ps, 153 ps->ps_error, bp->b_error, 154 ps->ps_endoffset, endoffset)); 155 156 ps->ps_endoffset = endoffset; 157 ps->ps_error = bp->b_error; 158 } 159 ps->ps_failed++; 160 161 ps->ps_resid += todo - done; 162 } else { 163 KASSERT(bp->b_error == 0); 164 } 165 166 ps->ps_running--; 167 cv_signal(&ps->ps_cv); 168 mutex_exit(&ps->ps_lock); 169 170 if (is_iobuf) 171 putiobuf(bp); 172 } 173 174 static void 175 physio_biodone(struct buf *bp) 176 { 177 #if defined(DIAGNOSTIC) 178 struct physio_stat *ps = bp->b_private; 179 size_t todo = bp->b_bufsize; 180 size_t done = bp->b_bcount - bp->b_resid; 181 182 KASSERT(ps->ps_running > 0); 183 KASSERT(bp->b_bcount <= todo); 184 KASSERT(bp->b_resid <= bp->b_bcount); 185 if (done == todo) 186 KASSERTMSG(bp->b_error == 0, "error=%d", bp->b_error); 187 #endif /* defined(DIAGNOSTIC) */ 188 189 workqueue_enqueue(physio_workqueue, &bp->b_work, NULL); 190 } 191 192 static void 193 physio_wait(struct physio_stat *ps, int n) 194 { 195 196 KASSERT(mutex_owned(&ps->ps_lock)); 197 198 while (ps->ps_running > n) 199 cv_wait(&ps->ps_cv, &ps->ps_lock); 200 } 201 202 static int 203 physio_init(void) 204 { 205 int error; 206 207 KASSERT(physio_workqueue == NULL); 208 209 error = workqueue_create(&physio_workqueue, "physiod", 210 physio_done, NULL, PRI_BIO, IPL_BIO, WQ_MPSAFE); 211 212 return error; 213 } 214 215 /* 216 * Do "physical I/O" on behalf of a user. "Physical I/O" is I/O directly 217 * from the raw device to user buffers, and bypasses the buffer cache. 218 */ 219 int 220 physio(void (*strategy)(struct buf *), struct buf *obp, dev_t dev, int flags, 221 void (*min_phys)(struct buf *), struct uio *uio) 222 { 223 struct iovec *iovp; 224 struct lwp *l = curlwp; 225 struct proc *p = l->l_proc; 226 int i, error; 227 struct buf *bp = NULL; 228 struct physio_stat *ps; 229 int concurrency = physio_concurrency - 1; 230 int isdisk; 231 232 error = RUN_ONCE(&physio_initialized, physio_init); 233 if (__predict_false(error != 0)) { 234 return error; 235 } 236 237 DPRINTF(("%s: called: off=%" PRIu64 ", resid=%zu\n", 238 __func__, uio->uio_offset, uio->uio_resid)); 239 240 flags &= B_READ | B_WRITE; 241 242 ps = kmem_zalloc(sizeof(*ps), KM_SLEEP); 243 /* ps->ps_running = 0; */ 244 /* ps->ps_error = 0; */ 245 /* ps->ps_failed = 0; */ 246 ps->ps_orig_bp = obp; 247 ps->ps_endoffset = -1; 248 ps->ps_resid = 0; 249 mutex_init(&ps->ps_lock, MUTEX_DEFAULT, IPL_NONE); 250 cv_init(&ps->ps_cv, "physio"); 251 252 /* Allow concurrent I/O only for disks */ 253 isdisk = cdev_type(dev) == D_DISK; 254 if (!isdisk) 255 concurrency = 0; 256 257 /* Make sure we have a buffer, creating one if necessary. */ 258 if (obp != NULL) { 259 mutex_enter(&bufcache_lock); 260 /* Mark it busy, so nobody else will use it. */ 261 while (bbusy(obp, false, 0, NULL) == EPASSTHROUGH) 262 ; 263 mutex_exit(&bufcache_lock); 264 concurrency = 0; /* see "XXXkludge" comment below */ 265 } 266 267 for (i = 0; i < uio->uio_iovcnt; i++) { 268 bool sync = true; 269 270 iovp = &uio->uio_iov[i]; 271 while (iovp->iov_len > 0) { 272 size_t todo; 273 vaddr_t endp; 274 275 mutex_enter(&ps->ps_lock); 276 if (ps->ps_failed != 0) { 277 goto done_locked; 278 } 279 physio_wait(ps, sync ? 0 : concurrency); 280 mutex_exit(&ps->ps_lock); 281 if (obp != NULL) { 282 /* 283 * XXXkludge 284 * some drivers use "obp" as an identifier. 285 */ 286 bp = obp; 287 } else { 288 bp = getiobuf(NULL, true); 289 bp->b_cflags |= BC_BUSY; 290 } 291 bp->b_dev = dev; 292 bp->b_proc = p; 293 bp->b_private = ps; 294 295 /* 296 * Mrk the buffer busy for physical I/O. Also set 297 * B_PHYS because it's an I/O to user memory, and 298 * B_RAW because B_RAW is to be "set by physio for 299 * raw transfers". 300 */ 301 bp->b_oflags = 0; 302 bp->b_cflags |= BC_BUSY; 303 bp->b_flags = flags | B_PHYS | B_RAW; 304 bp->b_iodone = physio_biodone; 305 306 /* Set up the buffer for a maximum-sized transfer. */ 307 bp->b_blkno = btodb(uio->uio_offset); 308 if (isdisk) { 309 /* 310 * For disks, check that offsets are at least block 311 * aligned, the block addresses are used to track 312 * errors of finished requests. 313 */ 314 if (uio->uio_offset & (DEV_BSIZE - 1)) { 315 error = SET_ERROR(EINVAL); 316 goto done; 317 } 318 /* 319 * Split request into MAXPHYS chunks 320 */ 321 bp->b_bcount = MIN(MAXPHYS, iovp->iov_len); 322 } else { 323 bp->b_bcount = MIN(INT_MAX, iovp->iov_len); 324 } 325 bp->b_data = iovp->iov_base; 326 327 /* 328 * Call minphys to bound the transfer size, 329 * and remember the amount of data to transfer, 330 * for later comparison. 331 */ 332 (*min_phys)(bp); 333 todo = bp->b_bufsize = bp->b_bcount; 334 #if defined(DIAGNOSTIC) 335 if (todo > MAXPHYS) 336 panic("todo(%zu) > MAXPHYS; minphys broken", 337 todo); 338 #endif /* defined(DIAGNOSTIC) */ 339 340 sync = false; 341 endp = (vaddr_t)bp->b_data + todo; 342 if (trunc_page(endp) != endp) { 343 /* 344 * Following requests can overlap. 345 * note that uvm_vslock does round_page. 346 */ 347 sync = true; 348 } 349 350 /* 351 * Lock the part of the user address space involved 352 * in the transfer. 353 */ 354 error = uvm_vslock(p->p_vmspace, bp->b_data, todo, 355 (flags & B_READ) ? VM_PROT_WRITE : VM_PROT_READ); 356 if (error) { 357 goto done; 358 } 359 360 /* 361 * Beware vmapbuf(); if successful it clobbers 362 * b_data and saves it in b_saveaddr. 363 * However, vunmapbuf() restores b_data. 364 */ 365 if ((error = vmapbuf(bp, todo)) != 0) { 366 uvm_vsunlock(p->p_vmspace, bp->b_data, todo); 367 goto done; 368 } 369 370 BIO_SETPRIO(bp, BPRIO_TIMECRITICAL); 371 372 mutex_enter(&ps->ps_lock); 373 ps->ps_running++; 374 mutex_exit(&ps->ps_lock); 375 376 /* Call strategy to start the transfer. */ 377 (*strategy)(bp); 378 bp = NULL; 379 380 iovp->iov_len -= todo; 381 iovp->iov_base = (char *)iovp->iov_base + todo; 382 uio->uio_offset += todo; 383 uio->uio_resid -= todo; 384 } 385 } 386 387 done: 388 mutex_enter(&ps->ps_lock); 389 done_locked: 390 physio_wait(ps, 0); 391 mutex_exit(&ps->ps_lock); 392 393 KASSERT(ps->ps_failed || ps->ps_endoffset == -1); 394 395 /* 396 * Compute residual, for disks adjust for the 397 * lowest numbered block that returned an error. 398 */ 399 if (isdisk) { 400 if (ps->ps_failed != 0) { 401 off_t delta; 402 403 delta = uio->uio_offset - ps->ps_endoffset; 404 KASSERT(delta > 0); 405 uio->uio_resid += delta; 406 /* uio->uio_offset = ps->ps_endoffset; */ 407 } 408 } else { 409 uio->uio_resid += ps->ps_resid; 410 } 411 412 if (bp != NULL && bp != obp) { 413 putiobuf(bp); 414 } 415 if (error == 0) { 416 error = ps->ps_error; 417 } 418 mutex_destroy(&ps->ps_lock); 419 cv_destroy(&ps->ps_cv); 420 kmem_free(ps, sizeof(*ps)); 421 422 /* 423 * Clean up the state of the buffer. Remember if somebody wants 424 * it, so we can wake them up below. Also, if we had to steal it, 425 * give it back. 426 */ 427 if (obp != NULL) { 428 KASSERT((obp->b_cflags & BC_BUSY) != 0); 429 430 /* 431 * If another process is waiting for the raw I/O buffer, 432 * wake up processes waiting to do physical I/O; 433 */ 434 mutex_enter(&bufcache_lock); 435 obp->b_cflags &= ~(BC_BUSY | BC_WANTED); 436 obp->b_flags &= ~(B_PHYS | B_RAW); 437 obp->b_iodone = NULL; 438 cv_broadcast(&obp->b_busy); 439 mutex_exit(&bufcache_lock); 440 } 441 442 DPRINTF(("%s: done: off=%" PRIu64 ", resid=%zu\n", 443 __func__, uio->uio_offset, uio->uio_resid)); 444 445 return error; 446 } 447 448 /* 449 * A minphys() routine is called by physio() to adjust the size of each 450 * I/O transfer before the latter is passed to the strategy routine. 451 * 452 * This minphys() is a default that must be called to enforce limits 453 * that are applicable to all devices, because of limitations in the 454 * kernel or the hardware platform. 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 Sat Feb 21 16:20:20 UTC 2026