1 /* $NetBSD: uvm_bio.c,v 1.39 2005/06/27 02:19:48 thorpej 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 1.39 2005/06/27 02:19:48 thorpej 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 46 #include <uvm/uvm.h> 47 48 /* 49 * global data structures 50 */ 51 52 /* 53 * local functions 54 */ 55 56 static int ubc_fault(struct uvm_faultinfo *, vaddr_t, struct vm_page **, 57 int, int, vm_fault_t, vm_prot_t, int); 58 static struct ubc_map *ubc_find_mapping(struct uvm_object *, voff_t); 59 60 /* 61 * local data structues 62 */ 63 64 #define UBC_HASH(uobj, offset) \ 65 (((((u_long)(uobj)) >> 8) + (((u_long)(offset)) >> PAGE_SHIFT)) & \ 66 ubc_object.hashmask) 67 68 #define UBC_QUEUE(offset) \ 69 (&ubc_object.inactive[(((u_long)(offset)) >> ubc_winshift) & \ 70 (UBC_NQUEUES - 1)]) 71 72 #define UBC_UMAP_ADDR(u) \ 73 (vaddr_t)(ubc_object.kva + (((u) - ubc_object.umap) << ubc_winshift)) 74 75 76 #define UMAP_PAGES_LOCKED 0x0001 77 #define UMAP_MAPPING_CACHED 0x0002 78 79 struct ubc_map 80 { 81 struct uvm_object * uobj; /* mapped object */ 82 voff_t offset; /* offset into uobj */ 83 voff_t writeoff; /* write offset */ 84 vsize_t writelen; /* write len */ 85 int refcount; /* refcount on mapping */ 86 int flags; /* extra state */ 87 88 LIST_ENTRY(ubc_map) hash; /* hash table */ 89 TAILQ_ENTRY(ubc_map) inactive; /* inactive queue */ 90 }; 91 92 static struct ubc_object 93 { 94 struct uvm_object uobj; /* glue for uvm_map() */ 95 char *kva; /* where ubc_object is mapped */ 96 struct ubc_map *umap; /* array of ubc_map's */ 97 98 LIST_HEAD(, ubc_map) *hash; /* hashtable for cached ubc_map's */ 99 u_long hashmask; /* mask for hashtable */ 100 101 TAILQ_HEAD(ubc_inactive_head, ubc_map) *inactive; 102 /* inactive queues for ubc_map's */ 103 104 } ubc_object; 105 106 struct uvm_pagerops ubc_pager = 107 { 108 NULL, /* init */ 109 NULL, /* reference */ 110 NULL, /* detach */ 111 ubc_fault, /* fault */ 112 /* ... rest are NULL */ 113 }; 114 115 int ubc_nwins = UBC_NWINS; 116 int ubc_winshift = UBC_WINSHIFT; 117 int ubc_winsize; 118 #if defined(PMAP_PREFER) 119 int ubc_nqueues; 120 #define UBC_NQUEUES ubc_nqueues 121 #else 122 #define UBC_NQUEUES 1 123 #endif 124 125 /* 126 * ubc_init 127 * 128 * init pager private data structures. 129 */ 130 131 void 132 ubc_init(void) 133 { 134 struct ubc_map *umap; 135 vaddr_t va; 136 int i; 137 138 /* 139 * Make sure ubc_winshift is sane. 140 */ 141 if (ubc_winshift < PAGE_SHIFT) 142 ubc_winshift = PAGE_SHIFT; 143 144 /* 145 * init ubc_object. 146 * alloc and init ubc_map's. 147 * init inactive queues. 148 * alloc and init hashtable. 149 * map in ubc_object. 150 */ 151 152 UVM_OBJ_INIT(&ubc_object.uobj, &ubc_pager, UVM_OBJ_KERN); 153 154 ubc_object.umap = malloc(ubc_nwins * sizeof(struct ubc_map), 155 M_TEMP, M_NOWAIT); 156 if (ubc_object.umap == NULL) 157 panic("ubc_init: failed to allocate ubc_map"); 158 memset(ubc_object.umap, 0, ubc_nwins * sizeof(struct ubc_map)); 159 160 if (ubc_winshift < PAGE_SHIFT) { 161 ubc_winshift = PAGE_SHIFT; 162 } 163 va = (vaddr_t)1L; 164 #ifdef PMAP_PREFER 165 PMAP_PREFER(0, &va, 0, 0); /* kernel is never topdown */ 166 ubc_nqueues = va >> ubc_winshift; 167 if (ubc_nqueues == 0) { 168 ubc_nqueues = 1; 169 } 170 #endif 171 ubc_winsize = 1 << ubc_winshift; 172 ubc_object.inactive = malloc(UBC_NQUEUES * 173 sizeof(struct ubc_inactive_head), M_TEMP, M_NOWAIT); 174 if (ubc_object.inactive == NULL) 175 panic("ubc_init: failed to allocate inactive queue heads"); 176 for (i = 0; i < UBC_NQUEUES; i++) { 177 TAILQ_INIT(&ubc_object.inactive[i]); 178 } 179 for (i = 0; i < ubc_nwins; i++) { 180 umap = &ubc_object.umap[i]; 181 TAILQ_INSERT_TAIL(&ubc_object.inactive[i & (UBC_NQUEUES - 1)], 182 umap, inactive); 183 } 184 185 ubc_object.hash = hashinit(ubc_nwins, HASH_LIST, M_TEMP, M_NOWAIT, 186 &ubc_object.hashmask); 187 for (i = 0; i <= ubc_object.hashmask; i++) { 188 LIST_INIT(&ubc_object.hash[i]); 189 } 190 191 if (uvm_map(kernel_map, (vaddr_t *)&ubc_object.kva, 192 ubc_nwins << ubc_winshift, &ubc_object.uobj, 0, (vsize_t)va, 193 UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE, 194 UVM_ADV_RANDOM, UVM_FLAG_NOMERGE)) != 0) { 195 panic("ubc_init: failed to map ubc_object"); 196 } 197 UVMHIST_INIT(ubchist, 300); 198 } 199 200 /* 201 * ubc_fault: fault routine for ubc mapping 202 */ 203 204 static int 205 ubc_fault(struct uvm_faultinfo *ufi, vaddr_t ign1, struct vm_page **ign2, 206 int ign3, int ign4, vm_fault_t fault_type, vm_prot_t access_type, 207 int flags) 208 { 209 struct uvm_object *uobj; 210 struct ubc_map *umap; 211 vaddr_t va, eva, ubc_offset, slot_offset; 212 int i, error, npages; 213 struct vm_page *pgs[ubc_winsize >> PAGE_SHIFT], *pg; 214 vm_prot_t prot; 215 UVMHIST_FUNC("ubc_fault"); UVMHIST_CALLED(ubchist); 216 217 /* 218 * no need to try with PGO_LOCKED... 219 * we don't need to have the map locked since we know that 220 * no one will mess with it until our reference is released. 221 */ 222 223 if (flags & PGO_LOCKED) { 224 uvmfault_unlockall(ufi, NULL, &ubc_object.uobj, NULL); 225 flags &= ~PGO_LOCKED; 226 } 227 228 va = ufi->orig_rvaddr; 229 ubc_offset = va - (vaddr_t)ubc_object.kva; 230 umap = &ubc_object.umap[ubc_offset >> ubc_winshift]; 231 KASSERT(umap->refcount != 0); 232 slot_offset = ubc_offset & (ubc_winsize - 1); 233 234 /* 235 * some platforms cannot write to individual bytes atomically, so 236 * software has to do read/modify/write of larger quantities instead. 237 * this means that the access_type for "write" operations 238 * can be VM_PROT_READ, which confuses us mightily. 239 * 240 * deal with this by resetting access_type based on the info 241 * that ubc_alloc() stores for us. 242 */ 243 244 access_type = umap->writelen ? VM_PROT_WRITE : VM_PROT_READ; 245 UVMHIST_LOG(ubchist, "va 0x%lx ubc_offset 0x%lx access_type %d", 246 va, ubc_offset, access_type, 0); 247 248 #ifdef DIAGNOSTIC 249 if ((access_type & VM_PROT_WRITE) != 0) { 250 if (slot_offset < trunc_page(umap->writeoff) || 251 umap->writeoff + umap->writelen <= slot_offset) { 252 panic("ubc_fault: out of range write"); 253 } 254 } 255 #endif 256 257 /* no umap locking needed since we have a ref on the umap */ 258 uobj = umap->uobj; 259 260 if ((access_type & VM_PROT_WRITE) == 0) { 261 npages = (ubc_winsize - slot_offset) >> PAGE_SHIFT; 262 } else { 263 npages = (round_page(umap->offset + umap->writeoff + 264 umap->writelen) - (umap->offset + slot_offset)) 265 >> PAGE_SHIFT; 266 flags |= PGO_PASTEOF; 267 } 268 269 again: 270 memset(pgs, 0, sizeof (pgs)); 271 simple_lock(&uobj->vmobjlock); 272 273 UVMHIST_LOG(ubchist, "slot_offset 0x%x writeoff 0x%x writelen 0x%x ", 274 slot_offset, umap->writeoff, umap->writelen, 0); 275 UVMHIST_LOG(ubchist, "getpages uobj %p offset 0x%x npages %d", 276 uobj, umap->offset + slot_offset, npages, 0); 277 278 error = (*uobj->pgops->pgo_get)(uobj, umap->offset + slot_offset, pgs, 279 &npages, 0, access_type, 0, flags); 280 UVMHIST_LOG(ubchist, "getpages error %d npages %d", error, npages, 0, 281 0); 282 283 if (error == EAGAIN) { 284 tsleep(&lbolt, PVM, "ubc_fault", 0); 285 goto again; 286 } 287 if (error) { 288 return error; 289 } 290 291 va = ufi->orig_rvaddr; 292 eva = ufi->orig_rvaddr + (npages << PAGE_SHIFT); 293 294 UVMHIST_LOG(ubchist, "va 0x%lx eva 0x%lx", va, eva, 0, 0); 295 simple_lock(&uobj->vmobjlock); 296 uvm_lock_pageq(); 297 for (i = 0; va < eva; i++, va += PAGE_SIZE) { 298 299 /* 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 /* 329 * avoid unneeded loan break if possible. 330 */ 331 332 if ((access_type & VM_PROT_WRITE) == 0) 333 prot &= ~VM_PROT_WRITE; 334 335 if (prot & VM_PROT_WRITE) { 336 uvm_unlock_pageq(); 337 pg = uvm_loanbreak(pg); 338 uvm_lock_pageq(); 339 if (pg == NULL) 340 continue; /* will re-fault */ 341 } 342 } 343 KASSERT(access_type == VM_PROT_READ || 344 (pg->flags & PG_RDONLY) == 0); 345 pmap_enter(ufi->orig_map->pmap, va, VM_PAGE_TO_PHYS(pg), 346 (pg->flags & PG_RDONLY) ? prot & ~VM_PROT_WRITE : prot, 347 access_type); 348 uvm_pageactivate(pg); 349 pg->flags &= ~(PG_BUSY); 350 UVM_PAGE_OWN(pg, NULL); 351 } 352 uvm_unlock_pageq(); 353 simple_unlock(&uobj->vmobjlock); 354 pmap_update(ufi->orig_map->pmap); 355 return 0; 356 } 357 358 /* 359 * local functions 360 */ 361 362 static struct ubc_map * 363 ubc_find_mapping(struct uvm_object *uobj, 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(struct uvm_object *uobj, voff_t offset, vsize_t *lenp, int flags) 386 { 387 vaddr_t slot_offset, va; 388 struct ubc_map *umap; 389 voff_t umap_offset; 390 int error; 391 UVMHIST_FUNC("ubc_alloc"); UVMHIST_CALLED(ubchist); 392 393 UVMHIST_LOG(ubchist, "uobj %p offset 0x%lx len 0x%lx", 394 uobj, offset, *lenp, 0); 395 396 KASSERT(*lenp > 0); 397 umap_offset = (offset & ~((voff_t)ubc_winsize - 1)); 398 slot_offset = (vaddr_t)(offset & ((voff_t)ubc_winsize - 1)); 399 *lenp = MIN(*lenp, ubc_winsize - slot_offset); 400 401 /* 402 * the object is always locked here, so we don't need to add a ref. 403 */ 404 405 again: 406 simple_lock(&ubc_object.uobj.vmobjlock); 407 umap = ubc_find_mapping(uobj, umap_offset); 408 if (umap == NULL) { 409 umap = TAILQ_FIRST(UBC_QUEUE(offset)); 410 if (umap == NULL) { 411 simple_unlock(&ubc_object.uobj.vmobjlock); 412 tsleep(&lbolt, PVM, "ubc_alloc", 0); 413 goto again; 414 } 415 416 /* 417 * remove from old hash (if any), add to new hash. 418 */ 419 420 if (umap->uobj != NULL) { 421 LIST_REMOVE(umap, hash); 422 } 423 umap->uobj = uobj; 424 umap->offset = umap_offset; 425 LIST_INSERT_HEAD(&ubc_object.hash[UBC_HASH(uobj, umap_offset)], 426 umap, hash); 427 va = UBC_UMAP_ADDR(umap); 428 if (umap->flags & UMAP_MAPPING_CACHED) { 429 umap->flags &= ~UMAP_MAPPING_CACHED; 430 pmap_remove(pmap_kernel(), va, va + ubc_winsize); 431 pmap_update(pmap_kernel()); 432 } 433 } else { 434 va = UBC_UMAP_ADDR(umap); 435 } 436 437 if (umap->refcount == 0) { 438 TAILQ_REMOVE(UBC_QUEUE(offset), umap, inactive); 439 } 440 441 #ifdef DIAGNOSTIC 442 if ((flags & UBC_WRITE) && (umap->writeoff || umap->writelen)) { 443 panic("ubc_alloc: concurrent writes uobj %p", uobj); 444 } 445 #endif 446 if (flags & UBC_WRITE) { 447 umap->writeoff = slot_offset; 448 umap->writelen = *lenp; 449 } 450 451 umap->refcount++; 452 simple_unlock(&ubc_object.uobj.vmobjlock); 453 UVMHIST_LOG(ubchist, "umap %p refs %d va %p flags 0x%x", 454 umap, umap->refcount, va, flags); 455 456 if (flags & UBC_FAULTBUSY) { 457 int npages = (*lenp + PAGE_SIZE - 1) >> PAGE_SHIFT; 458 struct vm_page *pgs[npages]; 459 int gpflags = PGO_SYNCIO|PGO_OVERWRITE|PGO_PASTEOF; 460 int i; 461 KDASSERT(flags & UBC_WRITE); 462 463 if (umap->flags & UMAP_MAPPING_CACHED) { 464 umap->flags &= ~UMAP_MAPPING_CACHED; 465 pmap_remove(pmap_kernel(), va, va + ubc_winsize); 466 } 467 memset(pgs, 0, sizeof(pgs)); 468 simple_lock(&uobj->vmobjlock); 469 error = (*uobj->pgops->pgo_get)(uobj, trunc_page(offset), pgs, 470 &npages, 0, VM_PROT_READ | VM_PROT_WRITE, 0, gpflags); 471 UVMHIST_LOG(ubchist, "faultbusy getpages %d", error, 0, 0, 0); 472 if (error) { 473 goto out; 474 } 475 for (i = 0; i < npages; i++) { 476 pmap_kenter_pa(va + slot_offset + (i << PAGE_SHIFT), 477 VM_PAGE_TO_PHYS(pgs[i]), 478 VM_PROT_READ | VM_PROT_WRITE); 479 } 480 pmap_update(pmap_kernel()); 481 umap->flags |= UMAP_PAGES_LOCKED; 482 } 483 484 out: 485 return (void *)(va + slot_offset); 486 } 487 488 /* 489 * ubc_release: free a file mapping window. 490 */ 491 492 void 493 ubc_release(void *va, int flags) 494 { 495 struct ubc_map *umap; 496 struct uvm_object *uobj; 497 vaddr_t umapva; 498 boolean_t unmapped; 499 UVMHIST_FUNC("ubc_release"); UVMHIST_CALLED(ubchist); 500 501 UVMHIST_LOG(ubchist, "va %p", va, 0, 0, 0); 502 umap = &ubc_object.umap[((char *)va - ubc_object.kva) >> ubc_winshift]; 503 umapva = UBC_UMAP_ADDR(umap); 504 uobj = umap->uobj; 505 KASSERT(uobj != NULL); 506 507 if (umap->flags & UMAP_PAGES_LOCKED) { 508 int slot_offset = umap->writeoff; 509 int endoff = umap->writeoff + umap->writelen; 510 int zerolen = round_page(endoff) - endoff; 511 int npages = (int)(round_page(umap->writeoff + umap->writelen) 512 - trunc_page(umap->writeoff)) >> PAGE_SHIFT; 513 struct vm_page *pgs[npages]; 514 paddr_t pa; 515 int i; 516 boolean_t rv; 517 518 if (zerolen) { 519 memset((char *)umapva + endoff, 0, zerolen); 520 } 521 umap->flags &= ~UMAP_PAGES_LOCKED; 522 uvm_lock_pageq(); 523 for (i = 0; i < npages; i++) { 524 rv = pmap_extract(pmap_kernel(), 525 umapva + slot_offset + (i << PAGE_SHIFT), &pa); 526 KASSERT(rv); 527 pgs[i] = PHYS_TO_VM_PAGE(pa); 528 pgs[i]->flags &= ~(PG_FAKE|PG_CLEAN); 529 KASSERT(pgs[i]->loan_count == 0); 530 uvm_pageactivate(pgs[i]); 531 } 532 uvm_unlock_pageq(); 533 pmap_kremove(umapva, ubc_winsize); 534 pmap_update(pmap_kernel()); 535 simple_lock(&uobj->vmobjlock); 536 uvm_page_unbusy(pgs, npages); 537 simple_unlock(&uobj->vmobjlock); 538 unmapped = TRUE; 539 } else { 540 unmapped = FALSE; 541 } 542 543 simple_lock(&ubc_object.uobj.vmobjlock); 544 umap->writeoff = 0; 545 umap->writelen = 0; 546 umap->refcount--; 547 if (umap->refcount == 0) { 548 if (flags & UBC_UNMAP) { 549 550 /* 551 * Invalidate any cached mappings if requested. 552 * This is typically used to avoid leaving 553 * incompatible cache aliases around indefinitely. 554 */ 555 556 pmap_remove(pmap_kernel(), umapva, 557 umapva + ubc_winsize); 558 umap->flags &= ~UMAP_MAPPING_CACHED; 559 pmap_update(pmap_kernel()); 560 LIST_REMOVE(umap, hash); 561 umap->uobj = NULL; 562 TAILQ_INSERT_HEAD(UBC_QUEUE(umap->offset), umap, 563 inactive); 564 } else { 565 if (!unmapped) { 566 umap->flags |= UMAP_MAPPING_CACHED; 567 } 568 TAILQ_INSERT_TAIL(UBC_QUEUE(umap->offset), umap, 569 inactive); 570 } 571 } 572 UVMHIST_LOG(ubchist, "umap %p refs %d", umap, umap->refcount, 0, 0); 573 simple_unlock(&ubc_object.uobj.vmobjlock); 574 } 575 576 577 #if 0 /* notused */ 578 /* 579 * removing a range of mappings from the ubc mapping cache. 580 */ 581 582 void 583 ubc_flush(struct uvm_object *uobj, voff_t start, voff_t end) 584 { 585 struct ubc_map *umap; 586 vaddr_t va; 587 UVMHIST_FUNC("ubc_flush"); UVMHIST_CALLED(ubchist); 588 589 UVMHIST_LOG(ubchist, "uobj %p start 0x%lx end 0x%lx", 590 uobj, start, end, 0); 591 592 simple_lock(&ubc_object.uobj.vmobjlock); 593 for (umap = ubc_object.umap; 594 umap < &ubc_object.umap[ubc_nwins]; 595 umap++) { 596 597 if (umap->uobj != uobj || umap->offset < start || 598 (umap->offset >= end && end != 0) || 599 umap->refcount > 0) { 600 continue; 601 } 602 603 /* 604 * remove from hash, 605 * move to head of inactive queue. 606 */ 607 608 va = (vaddr_t)(ubc_object.kva + 609 ((umap - ubc_object.umap) << ubc_winshift)); 610 pmap_remove(pmap_kernel(), va, va + ubc_winsize); 611 612 LIST_REMOVE(umap, hash); 613 umap->uobj = NULL; 614 TAILQ_REMOVE(UBC_QUEUE(umap->offset), umap, inactive); 615 TAILQ_INSERT_HEAD(UBC_QUEUE(umap->offset), umap, inactive); 616 } 617 pmap_update(pmap_kernel()); 618 simple_unlock(&ubc_object.uobj.vmobjlock); 619 } 620 #endif /* notused */ 621