uvm_bio.c revision 1.33
1 /* $NetBSD: uvm_bio.c,v 1.33 2005/01/09 16:42:44 chs Exp $ */ 2 3 /* 4 * Copyright (c) 1998 Chuck Silvers. 5 * All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice, this list of conditions and the following disclaimer. 12 * 2. Redistributions in binary form must reproduce the above copyright 13 * notice, this list of conditions and the following disclaimer in the 14 * documentation and/or other materials provided with the distribution. 15 * 3. The name of the author may not be used to endorse or promote products 16 * derived from this software without specific prior written permission. 17 * 18 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 19 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 20 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 21 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 22 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, 23 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 24 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 25 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 26 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 27 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 28 * SUCH DAMAGE. 29 * 30 */ 31 32 /* 33 * uvm_bio.c: buffered i/o object mapping cache 34 */ 35 36 #include <sys/cdefs.h> 37 __KERNEL_RCSID(0, "$NetBSD: uvm_bio.c,v 1.33 2005/01/09 16:42:44 chs Exp $"); 38 39 #include "opt_uvmhist.h" 40 41 #include <sys/param.h> 42 #include <sys/systm.h> 43 #include <sys/malloc.h> 44 #include <sys/kernel.h> 45 #include <sys/vnode.h> 46 #include <sys/proc.h> 47 48 #include <uvm/uvm.h> 49 50 /* 51 * global data structures 52 */ 53 54 /* 55 * local functions 56 */ 57 58 int ubc_fault(struct uvm_faultinfo *, vaddr_t, struct vm_page **, int, 59 int, vm_fault_t, vm_prot_t, int); 60 struct ubc_map *ubc_find_mapping(struct uvm_object *, voff_t); 61 62 /* 63 * local data structues 64 */ 65 66 #define UBC_HASH(uobj, offset) \ 67 (((((u_long)(uobj)) >> 8) + (((u_long)(offset)) >> PAGE_SHIFT)) & \ 68 ubc_object.hashmask) 69 70 #define UBC_QUEUE(offset) \ 71 (&ubc_object.inactive[(((u_long)(offset)) >> ubc_winshift) & \ 72 (UBC_NQUEUES - 1)]) 73 74 #define UBC_UMAP_ADDR(u) \ 75 (vaddr_t)(ubc_object.kva + (((u) - ubc_object.umap) << ubc_winshift)) 76 77 78 #define UMAP_PAGES_LOCKED 0x0001 79 #define UMAP_MAPPING_CACHED 0x0002 80 81 struct ubc_map 82 { 83 struct uvm_object * uobj; /* mapped object */ 84 voff_t offset; /* offset into uobj */ 85 voff_t writeoff; /* write offset */ 86 vsize_t writelen; /* write len */ 87 int refcount; /* refcount on mapping */ 88 int flags; /* extra state */ 89 90 LIST_ENTRY(ubc_map) hash; /* hash table */ 91 TAILQ_ENTRY(ubc_map) inactive; /* inactive queue */ 92 }; 93 94 static struct ubc_object 95 { 96 struct uvm_object uobj; /* glue for uvm_map() */ 97 char *kva; /* where ubc_object is mapped */ 98 struct ubc_map *umap; /* array of ubc_map's */ 99 100 LIST_HEAD(, ubc_map) *hash; /* hashtable for cached ubc_map's */ 101 u_long hashmask; /* mask for hashtable */ 102 103 TAILQ_HEAD(ubc_inactive_head, ubc_map) *inactive; 104 /* inactive queues for ubc_map's */ 105 106 } ubc_object; 107 108 struct uvm_pagerops ubc_pager = 109 { 110 NULL, /* init */ 111 NULL, /* reference */ 112 NULL, /* detach */ 113 ubc_fault, /* fault */ 114 /* ... rest are NULL */ 115 }; 116 117 int ubc_nwins = UBC_NWINS; 118 int ubc_winshift = UBC_WINSHIFT; 119 int ubc_winsize; 120 #if defined(PMAP_PREFER) 121 int ubc_nqueues; 122 #define UBC_NQUEUES ubc_nqueues 123 #else 124 #define UBC_NQUEUES 1 125 #endif 126 127 /* 128 * ubc_init 129 * 130 * init pager private data structures. 131 */ 132 133 void 134 ubc_init(void) 135 { 136 struct ubc_map *umap; 137 vaddr_t va; 138 int i; 139 140 /* 141 * Make sure ubc_winshift is sane. 142 */ 143 if (ubc_winshift < PAGE_SHIFT) 144 ubc_winshift = PAGE_SHIFT; 145 146 /* 147 * init ubc_object. 148 * alloc and init ubc_map's. 149 * init inactive queues. 150 * alloc and init hashtable. 151 * map in ubc_object. 152 */ 153 154 simple_lock_init(&ubc_object.uobj.vmobjlock); 155 ubc_object.uobj.pgops = &ubc_pager; 156 TAILQ_INIT(&ubc_object.uobj.memq); 157 ubc_object.uobj.uo_npages = 0; 158 ubc_object.uobj.uo_refs = UVM_OBJ_KERN; 159 160 ubc_object.umap = malloc(ubc_nwins * sizeof(struct ubc_map), 161 M_TEMP, M_NOWAIT); 162 if (ubc_object.umap == NULL) 163 panic("ubc_init: failed to allocate ubc_map"); 164 memset(ubc_object.umap, 0, ubc_nwins * sizeof(struct ubc_map)); 165 166 if (ubc_winshift < PAGE_SHIFT) { 167 ubc_winshift = PAGE_SHIFT; 168 } 169 va = (vaddr_t)1L; 170 #ifdef PMAP_PREFER 171 PMAP_PREFER(0, &va); 172 ubc_nqueues = va >> ubc_winshift; 173 if (ubc_nqueues == 0) { 174 ubc_nqueues = 1; 175 } 176 #endif 177 ubc_winsize = 1 << ubc_winshift; 178 ubc_object.inactive = malloc(UBC_NQUEUES * 179 sizeof(struct ubc_inactive_head), M_TEMP, M_NOWAIT); 180 if (ubc_object.inactive == NULL) 181 panic("ubc_init: failed to allocate inactive queue heads"); 182 for (i = 0; i < UBC_NQUEUES; i++) { 183 TAILQ_INIT(&ubc_object.inactive[i]); 184 } 185 for (i = 0; i < ubc_nwins; i++) { 186 umap = &ubc_object.umap[i]; 187 TAILQ_INSERT_TAIL(&ubc_object.inactive[i & (UBC_NQUEUES - 1)], 188 umap, inactive); 189 } 190 191 ubc_object.hash = hashinit(ubc_nwins, HASH_LIST, M_TEMP, M_NOWAIT, 192 &ubc_object.hashmask); 193 for (i = 0; i <= ubc_object.hashmask; i++) { 194 LIST_INIT(&ubc_object.hash[i]); 195 } 196 197 if (uvm_map(kernel_map, (vaddr_t *)&ubc_object.kva, 198 ubc_nwins << ubc_winshift, &ubc_object.uobj, 0, (vsize_t)va, 199 UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE, 200 UVM_ADV_RANDOM, UVM_FLAG_NOMERGE)) != 0) { 201 panic("ubc_init: failed to map ubc_object"); 202 } 203 UVMHIST_INIT(ubchist, 300); 204 } 205 206 /* 207 * ubc_fault: fault routine for ubc mapping 208 */ 209 210 int 211 ubc_fault(ufi, ign1, ign2, ign3, ign4, fault_type, access_type, flags) 212 struct uvm_faultinfo *ufi; 213 vaddr_t ign1; 214 struct vm_page **ign2; 215 int ign3, ign4; 216 vm_fault_t fault_type; 217 vm_prot_t access_type; 218 int flags; 219 { 220 struct uvm_object *uobj; 221 struct ubc_map *umap; 222 vaddr_t va, eva, ubc_offset, slot_offset; 223 int i, error, npages; 224 struct vm_page *pgs[ubc_winsize >> PAGE_SHIFT], *pg; 225 vm_prot_t prot; 226 UVMHIST_FUNC("ubc_fault"); UVMHIST_CALLED(ubchist); 227 228 /* 229 * no need to try with PGO_LOCKED... 230 * we don't need to have the map locked since we know that 231 * no one will mess with it until our reference is released. 232 */ 233 234 if (flags & PGO_LOCKED) { 235 uvmfault_unlockall(ufi, NULL, &ubc_object.uobj, NULL); 236 flags &= ~PGO_LOCKED; 237 } 238 239 va = ufi->orig_rvaddr; 240 ubc_offset = va - (vaddr_t)ubc_object.kva; 241 242 UVMHIST_LOG(ubchist, "va 0x%lx ubc_offset 0x%lx at %d", 243 va, ubc_offset, access_type, 0); 244 245 umap = &ubc_object.umap[ubc_offset >> ubc_winshift]; 246 KASSERT(umap->refcount != 0); 247 slot_offset = ubc_offset & (ubc_winsize - 1); 248 249 #ifdef DIAGNOSTIC 250 if ((access_type & VM_PROT_WRITE) != 0) { 251 if (slot_offset < trunc_page(umap->writeoff) || 252 umap->writeoff + umap->writelen <= slot_offset) { 253 panic("ubc_fault: out of range write"); 254 } 255 } 256 #endif 257 258 /* no umap locking needed since we have a ref on the umap */ 259 uobj = umap->uobj; 260 261 if ((access_type & VM_PROT_WRITE) == 0) { 262 npages = (ubc_winsize - slot_offset) >> PAGE_SHIFT; 263 } else { 264 npages = (round_page(umap->offset + umap->writeoff + 265 umap->writelen) - (umap->offset + slot_offset)) 266 >> PAGE_SHIFT; 267 flags |= PGO_PASTEOF; 268 } 269 270 again: 271 memset(pgs, 0, sizeof (pgs)); 272 simple_lock(&uobj->vmobjlock); 273 274 UVMHIST_LOG(ubchist, "slot_offset 0x%x writeoff 0x%x writelen 0x%x ", 275 slot_offset, umap->writeoff, umap->writelen, 0); 276 UVMHIST_LOG(ubchist, "getpages uobj %p offset 0x%x npages %d", 277 uobj, umap->offset + slot_offset, npages, 0); 278 279 error = (*uobj->pgops->pgo_get)(uobj, umap->offset + slot_offset, pgs, 280 &npages, 0, access_type, 0, flags); 281 UVMHIST_LOG(ubchist, "getpages error %d npages %d", error, npages, 0, 282 0); 283 284 if (error == EAGAIN) { 285 tsleep(&lbolt, PVM, "ubc_fault", 0); 286 goto again; 287 } 288 if (error) { 289 return error; 290 } 291 292 va = ufi->orig_rvaddr; 293 eva = ufi->orig_rvaddr + (npages << PAGE_SHIFT); 294 295 UVMHIST_LOG(ubchist, "va 0x%lx eva 0x%lx", va, eva, 0, 0); 296 simple_lock(&uobj->vmobjlock); 297 uvm_lock_pageq(); 298 for (i = 0; va < eva; i++, va += PAGE_SIZE) { 299 /* 300 * for virtually-indexed, virtually-tagged caches we should 301 * avoid creating writable mappings when we don't absolutely 302 * need them, since the "compatible alias" trick doesn't work 303 * on such caches. otherwise, we can always map the pages 304 * writable. 305 */ 306 307 #ifdef PMAP_CACHE_VIVT 308 prot = VM_PROT_READ | access_type; 309 #else 310 prot = VM_PROT_READ | VM_PROT_WRITE; 311 #endif 312 UVMHIST_LOG(ubchist, "pgs[%d] = %p", i, pgs[i], 0, 0); 313 pg = pgs[i]; 314 315 if (pg == NULL || pg == PGO_DONTCARE) { 316 continue; 317 } 318 if (pg->flags & PG_WANTED) { 319 wakeup(pg); 320 } 321 KASSERT((pg->flags & PG_FAKE) == 0); 322 if (pg->flags & PG_RELEASED) { 323 uvm_pagefree(pg); 324 continue; 325 } 326 if (pg->loan_count != 0) { 327 /* 328 * avoid unneeded loan break if possible. 329 */ 330 if ((access_type & VM_PROT_WRITE) == 0) 331 prot &= ~VM_PROT_WRITE; 332 333 if (prot & VM_PROT_WRITE) { 334 uvm_unlock_pageq(); 335 pg = uvm_loanbreak(pg); 336 uvm_lock_pageq(); 337 if (pg == NULL) 338 continue; /* will re-fault */ 339 } 340 } 341 KASSERT(access_type == VM_PROT_READ || 342 (pg->flags & PG_RDONLY) == 0); 343 pmap_enter(ufi->orig_map->pmap, va, VM_PAGE_TO_PHYS(pg), 344 (pg->flags & PG_RDONLY) ? prot & ~VM_PROT_WRITE : prot, 345 access_type); 346 uvm_pageactivate(pg); 347 pg->flags &= ~(PG_BUSY); 348 UVM_PAGE_OWN(pg, NULL); 349 } 350 uvm_unlock_pageq(); 351 simple_unlock(&uobj->vmobjlock); 352 pmap_update(ufi->orig_map->pmap); 353 return 0; 354 } 355 356 /* 357 * local functions 358 */ 359 360 struct ubc_map * 361 ubc_find_mapping(uobj, offset) 362 struct uvm_object *uobj; 363 voff_t offset; 364 { 365 struct ubc_map *umap; 366 367 LIST_FOREACH(umap, &ubc_object.hash[UBC_HASH(uobj, offset)], hash) { 368 if (umap->uobj == uobj && umap->offset == offset) { 369 return umap; 370 } 371 } 372 return NULL; 373 } 374 375 376 /* 377 * ubc interface functions 378 */ 379 380 /* 381 * ubc_alloc: allocate a file mapping window 382 */ 383 384 void * 385 ubc_alloc(uobj, offset, lenp, flags) 386 struct uvm_object *uobj; 387 voff_t offset; 388 vsize_t *lenp; 389 int flags; 390 { 391 vaddr_t slot_offset, va; 392 struct ubc_map *umap; 393 voff_t umap_offset; 394 int error; 395 UVMHIST_FUNC("ubc_alloc"); UVMHIST_CALLED(ubchist); 396 397 UVMHIST_LOG(ubchist, "uobj %p offset 0x%lx len 0x%lx", 398 uobj, offset, *lenp, 0); 399 400 umap_offset = (offset & ~((voff_t)ubc_winsize - 1)); 401 slot_offset = (vaddr_t)(offset & ((voff_t)ubc_winsize - 1)); 402 *lenp = MIN(*lenp, ubc_winsize - slot_offset); 403 404 /* 405 * the object is always locked here, so we don't need to add a ref. 406 */ 407 408 again: 409 simple_lock(&ubc_object.uobj.vmobjlock); 410 umap = ubc_find_mapping(uobj, umap_offset); 411 if (umap == NULL) { 412 umap = TAILQ_FIRST(UBC_QUEUE(offset)); 413 if (umap == NULL) { 414 simple_unlock(&ubc_object.uobj.vmobjlock); 415 tsleep(&lbolt, PVM, "ubc_alloc", 0); 416 goto again; 417 } 418 419 /* 420 * remove from old hash (if any), add to new hash. 421 */ 422 423 if (umap->uobj != NULL) { 424 LIST_REMOVE(umap, hash); 425 } 426 umap->uobj = uobj; 427 umap->offset = umap_offset; 428 LIST_INSERT_HEAD(&ubc_object.hash[UBC_HASH(uobj, umap_offset)], 429 umap, hash); 430 va = UBC_UMAP_ADDR(umap); 431 if (umap->flags & UMAP_MAPPING_CACHED) { 432 umap->flags &= ~UMAP_MAPPING_CACHED; 433 pmap_remove(pmap_kernel(), va, va + ubc_winsize); 434 pmap_update(pmap_kernel()); 435 } 436 } else { 437 va = UBC_UMAP_ADDR(umap); 438 } 439 440 if (umap->refcount == 0) { 441 TAILQ_REMOVE(UBC_QUEUE(offset), umap, inactive); 442 } 443 444 #ifdef DIAGNOSTIC 445 if ((flags & UBC_WRITE) && (umap->writeoff || umap->writelen)) { 446 panic("ubc_alloc: concurrent writes uobj %p", uobj); 447 } 448 #endif 449 if (flags & UBC_WRITE) { 450 umap->writeoff = slot_offset; 451 umap->writelen = *lenp; 452 } 453 454 umap->refcount++; 455 simple_unlock(&ubc_object.uobj.vmobjlock); 456 UVMHIST_LOG(ubchist, "umap %p refs %d va %p flags 0x%x", 457 umap, umap->refcount, va, flags); 458 459 if (flags & UBC_FAULTBUSY) { 460 int npages = (*lenp + PAGE_SIZE - 1) >> PAGE_SHIFT; 461 struct vm_page *pgs[npages]; 462 int gpflags = PGO_SYNCIO|PGO_OVERWRITE|PGO_PASTEOF; 463 int i; 464 KDASSERT(flags & UBC_WRITE); 465 466 if (umap->flags & UMAP_MAPPING_CACHED) { 467 umap->flags &= ~UMAP_MAPPING_CACHED; 468 pmap_remove(pmap_kernel(), va, va + ubc_winsize); 469 } 470 memset(pgs, 0, sizeof(pgs)); 471 simple_lock(&uobj->vmobjlock); 472 error = (*uobj->pgops->pgo_get)(uobj, trunc_page(offset), pgs, 473 &npages, 0, VM_PROT_READ | VM_PROT_WRITE, 0, gpflags); 474 UVMHIST_LOG(ubchist, "faultbusy getpages %d", error, 0, 0, 0); 475 if (error) { 476 goto out; 477 } 478 for (i = 0; i < npages; i++) { 479 pmap_kenter_pa(va + slot_offset + (i << PAGE_SHIFT), 480 VM_PAGE_TO_PHYS(pgs[i]), 481 VM_PROT_READ | VM_PROT_WRITE); 482 } 483 pmap_update(pmap_kernel()); 484 umap->flags |= UMAP_PAGES_LOCKED; 485 } 486 487 out: 488 return (void *)(va + slot_offset); 489 } 490 491 /* 492 * ubc_release: free a file mapping window. 493 */ 494 495 void 496 ubc_release(va, flags) 497 void *va; 498 int flags; 499 { 500 struct ubc_map *umap; 501 struct uvm_object *uobj; 502 vaddr_t umapva; 503 boolean_t unmapped; 504 UVMHIST_FUNC("ubc_release"); UVMHIST_CALLED(ubchist); 505 506 UVMHIST_LOG(ubchist, "va %p", va, 0, 0, 0); 507 umap = &ubc_object.umap[((char *)va - ubc_object.kva) >> ubc_winshift]; 508 umapva = UBC_UMAP_ADDR(umap); 509 uobj = umap->uobj; 510 KASSERT(uobj != NULL); 511 512 if (umap->flags & UMAP_PAGES_LOCKED) { 513 int slot_offset = umap->writeoff; 514 int endoff = umap->writeoff + umap->writelen; 515 int zerolen = round_page(endoff) - endoff; 516 int npages = (int)(round_page(umap->writeoff + umap->writelen) 517 - trunc_page(umap->writeoff)) >> PAGE_SHIFT; 518 struct vm_page *pgs[npages]; 519 paddr_t pa; 520 int i; 521 boolean_t rv; 522 523 if (zerolen) { 524 memset((char *)umapva + endoff, 0, zerolen); 525 } 526 umap->flags &= ~UMAP_PAGES_LOCKED; 527 uvm_lock_pageq(); 528 for (i = 0; i < npages; i++) { 529 rv = pmap_extract(pmap_kernel(), 530 umapva + slot_offset + (i << PAGE_SHIFT), &pa); 531 KASSERT(rv); 532 pgs[i] = PHYS_TO_VM_PAGE(pa); 533 pgs[i]->flags &= ~(PG_FAKE|PG_CLEAN); 534 KASSERT(pgs[i]->loan_count == 0); 535 uvm_pageactivate(pgs[i]); 536 } 537 uvm_unlock_pageq(); 538 pmap_kremove(umapva, ubc_winsize); 539 pmap_update(pmap_kernel()); 540 simple_lock(&uobj->vmobjlock); 541 uvm_page_unbusy(pgs, npages); 542 simple_unlock(&uobj->vmobjlock); 543 unmapped = TRUE; 544 } else { 545 unmapped = FALSE; 546 } 547 548 simple_lock(&ubc_object.uobj.vmobjlock); 549 umap->writeoff = 0; 550 umap->writelen = 0; 551 umap->refcount--; 552 if (umap->refcount == 0) { 553 if (flags & UBC_UNMAP) { 554 555 /* 556 * Invalidate any cached mappings if requested. 557 * This is typically used to avoid leaving 558 * incompatible cache aliases around indefinitely. 559 */ 560 561 pmap_remove(pmap_kernel(), umapva, 562 umapva + ubc_winsize); 563 umap->flags &= ~UMAP_MAPPING_CACHED; 564 pmap_update(pmap_kernel()); 565 LIST_REMOVE(umap, hash); 566 umap->uobj = NULL; 567 TAILQ_INSERT_HEAD(UBC_QUEUE(umap->offset), umap, 568 inactive); 569 } else { 570 if (!unmapped) { 571 umap->flags |= UMAP_MAPPING_CACHED; 572 } 573 TAILQ_INSERT_TAIL(UBC_QUEUE(umap->offset), umap, 574 inactive); 575 } 576 } 577 UVMHIST_LOG(ubchist, "umap %p refs %d", umap, umap->refcount, 0, 0); 578 simple_unlock(&ubc_object.uobj.vmobjlock); 579 } 580 581 582 #if 0 /* notused */ 583 /* 584 * removing a range of mappings from the ubc mapping cache. 585 */ 586 587 void 588 ubc_flush(uobj, start, end) 589 struct uvm_object *uobj; 590 voff_t start, end; 591 { 592 struct ubc_map *umap; 593 vaddr_t va; 594 UVMHIST_FUNC("ubc_flush"); UVMHIST_CALLED(ubchist); 595 596 UVMHIST_LOG(ubchist, "uobj %p start 0x%lx end 0x%lx", 597 uobj, start, end, 0); 598 599 simple_lock(&ubc_object.uobj.vmobjlock); 600 for (umap = ubc_object.umap; 601 umap < &ubc_object.umap[ubc_nwins]; 602 umap++) { 603 604 if (umap->uobj != uobj || umap->offset < start || 605 (umap->offset >= end && end != 0) || 606 umap->refcount > 0) { 607 continue; 608 } 609 610 /* 611 * remove from hash, 612 * move to head of inactive queue. 613 */ 614 615 va = (vaddr_t)(ubc_object.kva + 616 ((umap - ubc_object.umap) << ubc_winshift)); 617 pmap_remove(pmap_kernel(), va, va + ubc_winsize); 618 619 LIST_REMOVE(umap, hash); 620 umap->uobj = NULL; 621 TAILQ_REMOVE(UBC_QUEUE(umap->offset), umap, inactive); 622 TAILQ_INSERT_HEAD(UBC_QUEUE(umap->offset), umap, inactive); 623 } 624 pmap_update(pmap_kernel()); 625 simple_unlock(&ubc_object.uobj.vmobjlock); 626 } 627 #endif /* notused */ 628
Indexes created Mon Sep 29 21:09:56 GMT 2025