1 /* $NetBSD: uvm_pager.c,v 1.98 2010/06/22 18:34:50 rmind 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 1.98 2010/06/22 18:34:50 rmind 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, 0); 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 pmap_update(pmap_kernel()); 234 235 if (kva == emergva) { 236 mutex_enter(&pager_map_wanted_lock); 237 emerginuse = false; 238 wakeup(&emergva); 239 mutex_exit(&pager_map_wanted_lock); 240 return; 241 } 242 243 vm_map_lock(pager_map); 244 uvm_unmap_remove(pager_map, kva, kva + size, &entries, NULL, 0); 245 mutex_enter(&pager_map_wanted_lock); 246 if (pager_map_wanted) { 247 pager_map_wanted = false; 248 wakeup(pager_map); 249 } 250 mutex_exit(&pager_map_wanted_lock); 251 vm_map_unlock(pager_map); 252 if (entries) 253 uvm_unmap_detach(entries, 0); 254 UVMHIST_LOG(maphist,"<- done",0,0,0,0); 255 } 256 257 /* 258 * interrupt-context iodone handler for single-buf i/os 259 * or the top-level buf of a nested-buf i/o. 260 */ 261 262 void 263 uvm_aio_biodone(struct buf *bp) 264 { 265 /* reset b_iodone for when this is a single-buf i/o. */ 266 bp->b_iodone = uvm_aio_aiodone; 267 268 workqueue_enqueue(uvm.aiodone_queue, &bp->b_work, NULL); 269 } 270 271 void 272 uvm_aio_aiodone_pages(struct vm_page **pgs, int npages, bool write, int error) 273 { 274 struct uvm_object *uobj; 275 struct vm_page *pg; 276 kmutex_t *slock; 277 int pageout_done; 278 int swslot; 279 int i; 280 bool swap; 281 UVMHIST_FUNC("uvm_aio_aiodone_pages"); UVMHIST_CALLED(ubchist); 282 283 swslot = 0; 284 pageout_done = 0; 285 slock = NULL; 286 uobj = NULL; 287 pg = pgs[0]; 288 swap = (pg->uanon != NULL && pg->uobject == NULL) || 289 (pg->pqflags & PQ_AOBJ) != 0; 290 if (!swap) { 291 uobj = pg->uobject; 292 slock = &uobj->vmobjlock; 293 mutex_enter(slock); 294 mutex_enter(&uvm_pageqlock); 295 } else { 296 #if defined(VMSWAP) 297 if (error) { 298 if (pg->uobject != NULL) { 299 swslot = uao_find_swslot(pg->uobject, 300 pg->offset >> PAGE_SHIFT); 301 } else { 302 KASSERT(pg->uanon != NULL); 303 swslot = pg->uanon->an_swslot; 304 } 305 KASSERT(swslot); 306 } 307 #else /* defined(VMSWAP) */ 308 panic("%s: swap", __func__); 309 #endif /* defined(VMSWAP) */ 310 } 311 for (i = 0; i < npages; i++) { 312 pg = pgs[i]; 313 KASSERT(swap || pg->uobject == uobj); 314 UVMHIST_LOG(ubchist, "pg %p", pg, 0,0,0); 315 316 #if defined(VMSWAP) 317 /* 318 * for swap i/os, lock each page's object (or anon) 319 * individually since each page may need a different lock. 320 */ 321 322 if (swap) { 323 if (pg->uobject != NULL) { 324 slock = &pg->uobject->vmobjlock; 325 } else { 326 slock = &pg->uanon->an_lock; 327 } 328 mutex_enter(slock); 329 mutex_enter(&uvm_pageqlock); 330 } 331 #endif /* defined(VMSWAP) */ 332 333 /* 334 * process errors. for reads, just mark the page to be freed. 335 * for writes, if the error was ENOMEM, we assume this was 336 * a transient failure so we mark the page dirty so that 337 * we'll try to write it again later. for all other write 338 * errors, we assume the error is permanent, thus the data 339 * in the page is lost. bummer. 340 */ 341 342 if (error) { 343 int slot; 344 if (!write) { 345 pg->flags |= PG_RELEASED; 346 continue; 347 } else if (error == ENOMEM) { 348 if (pg->flags & PG_PAGEOUT) { 349 pg->flags &= ~PG_PAGEOUT; 350 pageout_done++; 351 } 352 pg->flags &= ~PG_CLEAN; 353 uvm_pageactivate(pg); 354 slot = 0; 355 } else 356 slot = SWSLOT_BAD; 357 358 #if defined(VMSWAP) 359 if (swap) { 360 if (pg->uobject != NULL) { 361 int oldslot; 362 oldslot = uao_set_swslot(pg->uobject, 363 pg->offset >> PAGE_SHIFT, slot); 364 KASSERT(oldslot == swslot + i); 365 } else { 366 KASSERT(pg->uanon->an_swslot == 367 swslot + i); 368 pg->uanon->an_swslot = slot; 369 } 370 } 371 #endif /* defined(VMSWAP) */ 372 } 373 374 /* 375 * if the page is PG_FAKE, this must have been a read to 376 * initialize the page. clear PG_FAKE and activate the page. 377 * we must also clear the pmap "modified" flag since it may 378 * still be set from the page's previous identity. 379 */ 380 381 if (pg->flags & PG_FAKE) { 382 KASSERT(!write); 383 pg->flags &= ~PG_FAKE; 384 #if defined(READAHEAD_STATS) 385 pg->pqflags |= PQ_READAHEAD; 386 uvm_ra_total.ev_count++; 387 #endif /* defined(READAHEAD_STATS) */ 388 KASSERT((pg->flags & PG_CLEAN) != 0); 389 uvm_pageenqueue(pg); 390 pmap_clear_modify(pg); 391 } 392 393 /* 394 * do accounting for pagedaemon i/o and arrange to free 395 * the pages instead of just unbusying them. 396 */ 397 398 if (pg->flags & PG_PAGEOUT) { 399 pg->flags &= ~PG_PAGEOUT; 400 pageout_done++; 401 uvmexp.pdfreed++; 402 pg->flags |= PG_RELEASED; 403 } 404 405 #if defined(VMSWAP) 406 /* 407 * for swap pages, unlock everything for this page now. 408 */ 409 410 if (swap) { 411 if (pg->uobject == NULL && pg->uanon->an_ref == 0 && 412 (pg->flags & PG_RELEASED) != 0) { 413 mutex_exit(&uvm_pageqlock); 414 uvm_anon_release(pg->uanon); 415 } else { 416 uvm_page_unbusy(&pg, 1); 417 mutex_exit(&uvm_pageqlock); 418 mutex_exit(slock); 419 } 420 } 421 #endif /* defined(VMSWAP) */ 422 } 423 uvm_pageout_done(pageout_done); 424 if (!swap) { 425 uvm_page_unbusy(pgs, npages); 426 mutex_exit(&uvm_pageqlock); 427 mutex_exit(slock); 428 } else { 429 #if defined(VMSWAP) 430 KASSERT(write); 431 432 /* these pages are now only in swap. */ 433 mutex_enter(&uvm_swap_data_lock); 434 KASSERT(uvmexp.swpgonly + npages <= uvmexp.swpginuse); 435 if (error != ENOMEM) 436 uvmexp.swpgonly += npages; 437 mutex_exit(&uvm_swap_data_lock); 438 if (error) { 439 if (error != ENOMEM) 440 uvm_swap_markbad(swslot, npages); 441 else 442 uvm_swap_free(swslot, npages); 443 } 444 uvmexp.pdpending--; 445 #endif /* defined(VMSWAP) */ 446 } 447 } 448 449 /* 450 * uvm_aio_aiodone: do iodone processing for async i/os. 451 * this should be called in thread context, not interrupt context. 452 */ 453 454 void 455 uvm_aio_aiodone(struct buf *bp) 456 { 457 int npages = bp->b_bufsize >> PAGE_SHIFT; 458 struct vm_page *pgs[npages]; 459 int i, error; 460 bool write; 461 UVMHIST_FUNC("uvm_aio_aiodone"); UVMHIST_CALLED(ubchist); 462 UVMHIST_LOG(ubchist, "bp %p", bp, 0,0,0); 463 464 error = bp->b_error; 465 write = (bp->b_flags & B_READ) == 0; 466 467 for (i = 0; i < npages; i++) { 468 pgs[i] = uvm_pageratop((vaddr_t)bp->b_data + (i << PAGE_SHIFT)); 469 UVMHIST_LOG(ubchist, "pgs[%d] = %p", i, pgs[i],0,0); 470 } 471 uvm_pagermapout((vaddr_t)bp->b_data, npages); 472 473 uvm_aio_aiodone_pages(pgs, npages, write, error); 474 475 if (write && (bp->b_cflags & BC_AGE) != 0) { 476 mutex_enter(bp->b_objlock); 477 vwakeup(bp); 478 mutex_exit(bp->b_objlock); 479 } 480 putiobuf(bp); 481 } 482 483 /* 484 * uvm_pageratop: convert KVAs in the pager map back to their page 485 * structures. 486 */ 487 488 struct vm_page * 489 uvm_pageratop(vaddr_t kva) 490 { 491 struct vm_page *pg; 492 paddr_t pa; 493 bool rv; 494 495 rv = pmap_extract(pmap_kernel(), kva, &pa); 496 KASSERT(rv); 497 pg = PHYS_TO_VM_PAGE(pa); 498 KASSERT(pg != NULL); 499 return (pg); 500 } 501