1 /* $NetBSD: uvm_pager.c,v 1.95 2009/03/30 16:36:36 yamt Exp $ */ 2 3 /* 4 * 5 * Copyright (c) 1997 Charles D. Cranor and Washington University. 6 * All rights reserved. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 3. All advertising materials mentioning features or use of this software 17 * must display the following acknowledgement: 18 * This product includes software developed by Charles D. Cranor and 19 * Washington University. 20 * 4. The name of the author may not be used to endorse or promote products 21 * derived from this software without specific prior written permission. 22 * 23 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 24 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 25 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 26 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 27 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 28 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 29 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 30 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 31 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 32 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 33 * 34 * from: Id: uvm_pager.c,v 1.1.2.23 1998/02/02 20:38:06 chuck Exp 35 */ 36 37 /* 38 * uvm_pager.c: generic functions used to assist the pagers. 39 */ 40 41 #include <sys/cdefs.h> 42 __KERNEL_RCSID(0, "$NetBSD: uvm_pager.c,v 1.95 2009/03/30 16:36:36 yamt Exp $"); 43 44 #include "opt_uvmhist.h" 45 #include "opt_readahead.h" 46 #include "opt_pagermap.h" 47 48 #include <sys/param.h> 49 #include <sys/systm.h> 50 #include <sys/proc.h> 51 #include <sys/malloc.h> 52 #include <sys/vnode.h> 53 #include <sys/buf.h> 54 55 #include <uvm/uvm.h> 56 57 /* 58 * XXX 59 * this is needed until the device strategy interface 60 * is changed to do physically-addressed i/o. 61 */ 62 63 #ifndef PAGER_MAP_DEFAULT_SIZE 64 #define PAGER_MAP_DEFAULT_SIZE (16 * 1024 * 1024) 65 #endif 66 67 #ifndef PAGER_MAP_SIZE 68 #define PAGER_MAP_SIZE PAGER_MAP_DEFAULT_SIZE 69 #endif 70 71 size_t pager_map_size = PAGER_MAP_SIZE; 72 73 /* 74 * list of uvm pagers in the system 75 */ 76 77 const struct uvm_pagerops * const uvmpagerops[] = { 78 &aobj_pager, 79 &uvm_deviceops, 80 &uvm_vnodeops, 81 &ubc_pager, 82 }; 83 84 /* 85 * the pager map: provides KVA for I/O 86 */ 87 88 struct vm_map *pager_map; /* XXX */ 89 kmutex_t pager_map_wanted_lock; 90 bool pager_map_wanted; /* locked by pager map */ 91 static vaddr_t emergva; 92 static bool emerginuse; 93 94 /* 95 * uvm_pager_init: init pagers (at boot time) 96 */ 97 98 void 99 uvm_pager_init(void) 100 { 101 u_int lcv; 102 vaddr_t sva, eva; 103 104 /* 105 * init pager map 106 */ 107 108 sva = 0; 109 pager_map = uvm_km_suballoc(kernel_map, &sva, &eva, pager_map_size, 0, 110 false, NULL); 111 mutex_init(&pager_map_wanted_lock, MUTEX_DEFAULT, IPL_NONE); 112 pager_map_wanted = false; 113 emergva = uvm_km_alloc(kernel_map, round_page(MAXPHYS), 0, 114 UVM_KMF_VAONLY); 115 #if defined(DEBUG) 116 if (emergva == 0) 117 panic("emergva"); 118 #endif 119 emerginuse = false; 120 121 /* 122 * init ASYNC I/O queue 123 */ 124 125 TAILQ_INIT(&uvm.aio_done); 126 127 /* 128 * call pager init functions 129 */ 130 for (lcv = 0 ; lcv < __arraycount(uvmpagerops); lcv++) { 131 if (uvmpagerops[lcv]->pgo_init) 132 uvmpagerops[lcv]->pgo_init(); 133 } 134 } 135 136 /* 137 * uvm_pagermapin: map pages into KVA (pager_map) for I/O that needs mappings 138 * 139 * we basically just map in a blank map entry to reserve the space in the 140 * map and then use pmap_enter() to put the mappings in by hand. 141 */ 142 143 vaddr_t 144 uvm_pagermapin(struct vm_page **pps, int npages, int flags) 145 { 146 vsize_t size; 147 vaddr_t kva; 148 vaddr_t cva; 149 struct vm_page *pp; 150 vm_prot_t prot; 151 const bool pdaemon = curlwp == uvm.pagedaemon_lwp; 152 UVMHIST_FUNC("uvm_pagermapin"); UVMHIST_CALLED(maphist); 153 154 UVMHIST_LOG(maphist,"(pps=0x%x, npages=%d)", pps, npages,0,0); 155 156 /* 157 * compute protection. outgoing I/O only needs read 158 * access to the page, whereas incoming needs read/write. 159 */ 160 161 prot = VM_PROT_READ; 162 if (flags & UVMPAGER_MAPIN_READ) 163 prot |= VM_PROT_WRITE; 164 165 ReStart: 166 size = npages << PAGE_SHIFT; 167 kva = 0; /* let system choose VA */ 168 169 if (uvm_map(pager_map, &kva, size, NULL, UVM_UNKNOWN_OFFSET, 0, 170 UVM_FLAG_NOMERGE | (pdaemon ? UVM_FLAG_NOWAIT : 0)) != 0) { 171 if (pdaemon) { 172 mutex_enter(&pager_map_wanted_lock); 173 if (emerginuse) { 174 UVM_UNLOCK_AND_WAIT(&emergva, 175 &pager_map_wanted_lock, false, 176 "emergva", 0); 177 goto ReStart; 178 } 179 emerginuse = true; 180 mutex_exit(&pager_map_wanted_lock); 181 kva = emergva; 182 /* The shift implicitly truncates to PAGE_SIZE */ 183 KASSERT(npages <= (MAXPHYS >> PAGE_SHIFT)); 184 goto enter; 185 } 186 if ((flags & UVMPAGER_MAPIN_WAITOK) == 0) { 187 UVMHIST_LOG(maphist,"<- NOWAIT failed", 0,0,0,0); 188 return(0); 189 } 190 mutex_enter(&pager_map_wanted_lock); 191 pager_map_wanted = true; 192 UVMHIST_LOG(maphist, " SLEEPING on pager_map",0,0,0,0); 193 UVM_UNLOCK_AND_WAIT(pager_map, &pager_map_wanted_lock, false, 194 "pager_map", 0); 195 goto ReStart; 196 } 197 198 enter: 199 /* got it */ 200 for (cva = kva ; size != 0 ; size -= PAGE_SIZE, cva += PAGE_SIZE) { 201 pp = *pps++; 202 KASSERT(pp); 203 KASSERT(pp->flags & PG_BUSY); 204 pmap_kenter_pa(cva, VM_PAGE_TO_PHYS(pp), prot); 205 } 206 pmap_update(vm_map_pmap(pager_map)); 207 208 UVMHIST_LOG(maphist, "<- done (KVA=0x%x)", kva,0,0,0); 209 return(kva); 210 } 211 212 /* 213 * uvm_pagermapout: remove pager_map mapping 214 * 215 * we remove our mappings by hand and then remove the mapping (waking 216 * up anyone wanting space). 217 */ 218 219 void 220 uvm_pagermapout(vaddr_t kva, int npages) 221 { 222 vsize_t size = npages << PAGE_SHIFT; 223 struct vm_map_entry *entries; 224 UVMHIST_FUNC("uvm_pagermapout"); UVMHIST_CALLED(maphist); 225 226 UVMHIST_LOG(maphist, " (kva=0x%x, npages=%d)", kva, npages,0,0); 227 228 /* 229 * duplicate uvm_unmap, but add in pager_map_wanted handling. 230 */ 231 232 pmap_kremove(kva, npages << PAGE_SHIFT); 233 if (kva == emergva) { 234 mutex_enter(&pager_map_wanted_lock); 235 emerginuse = false; 236 wakeup(&emergva); 237 mutex_exit(&pager_map_wanted_lock); 238 return; 239 } 240 241 vm_map_lock(pager_map); 242 uvm_unmap_remove(pager_map, kva, kva + size, &entries, NULL, 0); 243 mutex_enter(&pager_map_wanted_lock); 244 if (pager_map_wanted) { 245 pager_map_wanted = false; 246 wakeup(pager_map); 247 } 248 mutex_exit(&pager_map_wanted_lock); 249 vm_map_unlock(pager_map); 250 if (entries) 251 uvm_unmap_detach(entries, 0); 252 pmap_update(pmap_kernel()); 253 UVMHIST_LOG(maphist,"<- done",0,0,0,0); 254 } 255 256 /* 257 * interrupt-context iodone handler for nested i/o bufs. 258 * 259 * => the buffer is private so need not be locked here 260 */ 261 262 void 263 uvm_aio_biodone1(struct buf *bp) 264 { 265 struct buf *mbp = bp->b_private; 266 267 KASSERT(mbp != bp); 268 if (bp->b_error != 0) { 269 mbp->b_error = bp->b_error; 270 } 271 mbp->b_resid -= bp->b_bcount; 272 putiobuf(bp); 273 if (mbp->b_resid == 0) { 274 biodone(mbp); 275 } 276 } 277 278 /* 279 * interrupt-context iodone handler for single-buf i/os 280 * or the top-level buf of a nested-buf i/o. 281 */ 282 283 void 284 uvm_aio_biodone(struct buf *bp) 285 { 286 /* reset b_iodone for when this is a single-buf i/o. */ 287 bp->b_iodone = uvm_aio_aiodone; 288 289 workqueue_enqueue(uvm.aiodone_queue, &bp->b_work, NULL); 290 } 291 292 void 293 uvm_aio_aiodone_pages(struct vm_page **pgs, int npages, bool write, int error) 294 { 295 struct uvm_object *uobj; 296 struct vm_page *pg; 297 kmutex_t *slock; 298 int pageout_done; 299 int swslot; 300 int i; 301 bool swap; 302 UVMHIST_FUNC("uvm_aio_aiodone_pages"); UVMHIST_CALLED(ubchist); 303 304 swslot = 0; 305 pageout_done = 0; 306 slock = NULL; 307 uobj = NULL; 308 pg = pgs[0]; 309 swap = (pg->uanon != NULL && pg->uobject == NULL) || 310 (pg->pqflags & PQ_AOBJ) != 0; 311 if (!swap) { 312 uobj = pg->uobject; 313 slock = &uobj->vmobjlock; 314 mutex_enter(slock); 315 mutex_enter(&uvm_pageqlock); 316 } else { 317 #if defined(VMSWAP) 318 if (error) { 319 if (pg->uobject != NULL) { 320 swslot = uao_find_swslot(pg->uobject, 321 pg->offset >> PAGE_SHIFT); 322 } else { 323 KASSERT(pg->uanon != NULL); 324 swslot = pg->uanon->an_swslot; 325 } 326 KASSERT(swslot); 327 } 328 #else /* defined(VMSWAP) */ 329 panic("%s: swap", __func__); 330 #endif /* defined(VMSWAP) */ 331 } 332 for (i = 0; i < npages; i++) { 333 pg = pgs[i]; 334 KASSERT(swap || pg->uobject == uobj); 335 UVMHIST_LOG(ubchist, "pg %p", pg, 0,0,0); 336 337 #if defined(VMSWAP) 338 /* 339 * for swap i/os, lock each page's object (or anon) 340 * individually since each page may need a different lock. 341 */ 342 343 if (swap) { 344 if (pg->uobject != NULL) { 345 slock = &pg->uobject->vmobjlock; 346 } else { 347 slock = &pg->uanon->an_lock; 348 } 349 mutex_enter(slock); 350 mutex_enter(&uvm_pageqlock); 351 } 352 #endif /* defined(VMSWAP) */ 353 354 /* 355 * process errors. for reads, just mark the page to be freed. 356 * for writes, if the error was ENOMEM, we assume this was 357 * a transient failure so we mark the page dirty so that 358 * we'll try to write it again later. for all other write 359 * errors, we assume the error is permanent, thus the data 360 * in the page is lost. bummer. 361 */ 362 363 if (error) { 364 int slot; 365 if (!write) { 366 pg->flags |= PG_RELEASED; 367 continue; 368 } else if (error == ENOMEM) { 369 if (pg->flags & PG_PAGEOUT) { 370 pg->flags &= ~PG_PAGEOUT; 371 pageout_done++; 372 } 373 pg->flags &= ~PG_CLEAN; 374 uvm_pageactivate(pg); 375 slot = 0; 376 } else 377 slot = SWSLOT_BAD; 378 379 #if defined(VMSWAP) 380 if (swap) { 381 if (pg->uobject != NULL) { 382 int oldslot; 383 oldslot = uao_set_swslot(pg->uobject, 384 pg->offset >> PAGE_SHIFT, slot); 385 KASSERT(oldslot == swslot + i); 386 } else { 387 KASSERT(pg->uanon->an_swslot == 388 swslot + i); 389 pg->uanon->an_swslot = slot; 390 } 391 } 392 #endif /* defined(VMSWAP) */ 393 } 394 395 /* 396 * if the page is PG_FAKE, this must have been a read to 397 * initialize the page. clear PG_FAKE and activate the page. 398 * we must also clear the pmap "modified" flag since it may 399 * still be set from the page's previous identity. 400 */ 401 402 if (pg->flags & PG_FAKE) { 403 KASSERT(!write); 404 pg->flags &= ~PG_FAKE; 405 #if defined(READAHEAD_STATS) 406 pg->pqflags |= PQ_READAHEAD; 407 uvm_ra_total.ev_count++; 408 #endif /* defined(READAHEAD_STATS) */ 409 KASSERT((pg->flags & PG_CLEAN) != 0); 410 uvm_pageenqueue(pg); 411 pmap_clear_modify(pg); 412 } 413 414 /* 415 * do accounting for pagedaemon i/o and arrange to free 416 * the pages instead of just unbusying them. 417 */ 418 419 if (pg->flags & PG_PAGEOUT) { 420 pg->flags &= ~PG_PAGEOUT; 421 pageout_done++; 422 uvmexp.pdfreed++; 423 pg->flags |= PG_RELEASED; 424 } 425 426 #if defined(VMSWAP) 427 /* 428 * for swap pages, unlock everything for this page now. 429 */ 430 431 if (swap) { 432 if (pg->uobject == NULL && pg->uanon->an_ref == 0 && 433 (pg->flags & PG_RELEASED) != 0) { 434 mutex_exit(&uvm_pageqlock); 435 uvm_anon_release(pg->uanon); 436 } else { 437 uvm_page_unbusy(&pg, 1); 438 mutex_exit(&uvm_pageqlock); 439 mutex_exit(slock); 440 } 441 } 442 #endif /* defined(VMSWAP) */ 443 } 444 uvm_pageout_done(pageout_done); 445 if (!swap) { 446 uvm_page_unbusy(pgs, npages); 447 mutex_exit(&uvm_pageqlock); 448 mutex_exit(slock); 449 } else { 450 #if defined(VMSWAP) 451 KASSERT(write); 452 453 /* these pages are now only in swap. */ 454 mutex_enter(&uvm_swap_data_lock); 455 KASSERT(uvmexp.swpgonly + npages <= uvmexp.swpginuse); 456 if (error != ENOMEM) 457 uvmexp.swpgonly += npages; 458 mutex_exit(&uvm_swap_data_lock); 459 if (error) { 460 if (error != ENOMEM) 461 uvm_swap_markbad(swslot, npages); 462 else 463 uvm_swap_free(swslot, npages); 464 } 465 uvmexp.pdpending--; 466 #endif /* defined(VMSWAP) */ 467 } 468 } 469 470 /* 471 * uvm_aio_aiodone: do iodone processing for async i/os. 472 * this should be called in thread context, not interrupt context. 473 */ 474 475 void 476 uvm_aio_aiodone(struct buf *bp) 477 { 478 int npages = bp->b_bufsize >> PAGE_SHIFT; 479 struct vm_page *pgs[npages]; 480 int i, error; 481 bool write; 482 UVMHIST_FUNC("uvm_aio_aiodone"); UVMHIST_CALLED(ubchist); 483 UVMHIST_LOG(ubchist, "bp %p", bp, 0,0,0); 484 485 error = bp->b_error; 486 write = (bp->b_flags & B_READ) == 0; 487 488 for (i = 0; i < npages; i++) { 489 pgs[i] = uvm_pageratop((vaddr_t)bp->b_data + (i << PAGE_SHIFT)); 490 UVMHIST_LOG(ubchist, "pgs[%d] = %p", i, pgs[i],0,0); 491 } 492 uvm_pagermapout((vaddr_t)bp->b_data, npages); 493 494 uvm_aio_aiodone_pages(pgs, npages, write, error); 495 496 if (write && (bp->b_cflags & BC_AGE) != 0) { 497 mutex_enter(bp->b_objlock); 498 vwakeup(bp); 499 mutex_exit(bp->b_objlock); 500 } 501 putiobuf(bp); 502 } 503 504 /* 505 * uvm_pageratop: convert KVAs in the pager map back to their page 506 * structures. 507 */ 508 509 struct vm_page * 510 uvm_pageratop(vaddr_t kva) 511 { 512 struct vm_page *pg; 513 paddr_t pa; 514 bool rv; 515 516 rv = pmap_extract(pmap_kernel(), kva, &pa); 517 KASSERT(rv); 518 pg = PHYS_TO_VM_PAGE(pa); 519 KASSERT(pg != NULL); 520 return (pg); 521 } 522