1 /* $NetBSD: uvm_pager.c,v 1.97 2009/11/07 07:27:50 cegger 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 1.97 2009/11/07 07:27:50 cegger 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 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 single-buf i/os 258 * or the top-level buf of a nested-buf i/o. 259 */ 260 261 void 262 uvm_aio_biodone(struct buf *bp) 263 { 264 /* reset b_iodone for when this is a single-buf i/o. */ 265 bp->b_iodone = uvm_aio_aiodone; 266 267 workqueue_enqueue(uvm.aiodone_queue, &bp->b_work, NULL); 268 } 269 270 void 271 uvm_aio_aiodone_pages(struct vm_page **pgs, int npages, bool write, int error) 272 { 273 struct uvm_object *uobj; 274 struct vm_page *pg; 275 kmutex_t *slock; 276 int pageout_done; 277 int swslot; 278 int i; 279 bool swap; 280 UVMHIST_FUNC("uvm_aio_aiodone_pages"); UVMHIST_CALLED(ubchist); 281 282 swslot = 0; 283 pageout_done = 0; 284 slock = NULL; 285 uobj = NULL; 286 pg = pgs[0]; 287 swap = (pg->uanon != NULL && pg->uobject == NULL) || 288 (pg->pqflags & PQ_AOBJ) != 0; 289 if (!swap) { 290 uobj = pg->uobject; 291 slock = &uobj->vmobjlock; 292 mutex_enter(slock); 293 mutex_enter(&uvm_pageqlock); 294 } else { 295 #if defined(VMSWAP) 296 if (error) { 297 if (pg->uobject != NULL) { 298 swslot = uao_find_swslot(pg->uobject, 299 pg->offset >> PAGE_SHIFT); 300 } else { 301 KASSERT(pg->uanon != NULL); 302 swslot = pg->uanon->an_swslot; 303 } 304 KASSERT(swslot); 305 } 306 #else /* defined(VMSWAP) */ 307 panic("%s: swap", __func__); 308 #endif /* defined(VMSWAP) */ 309 } 310 for (i = 0; i < npages; i++) { 311 pg = pgs[i]; 312 KASSERT(swap || pg->uobject == uobj); 313 UVMHIST_LOG(ubchist, "pg %p", pg, 0,0,0); 314 315 #if defined(VMSWAP) 316 /* 317 * for swap i/os, lock each page's object (or anon) 318 * individually since each page may need a different lock. 319 */ 320 321 if (swap) { 322 if (pg->uobject != NULL) { 323 slock = &pg->uobject->vmobjlock; 324 } else { 325 slock = &pg->uanon->an_lock; 326 } 327 mutex_enter(slock); 328 mutex_enter(&uvm_pageqlock); 329 } 330 #endif /* defined(VMSWAP) */ 331 332 /* 333 * process errors. for reads, just mark the page to be freed. 334 * for writes, if the error was ENOMEM, we assume this was 335 * a transient failure so we mark the page dirty so that 336 * we'll try to write it again later. for all other write 337 * errors, we assume the error is permanent, thus the data 338 * in the page is lost. bummer. 339 */ 340 341 if (error) { 342 int slot; 343 if (!write) { 344 pg->flags |= PG_RELEASED; 345 continue; 346 } else if (error == ENOMEM) { 347 if (pg->flags & PG_PAGEOUT) { 348 pg->flags &= ~PG_PAGEOUT; 349 pageout_done++; 350 } 351 pg->flags &= ~PG_CLEAN; 352 uvm_pageactivate(pg); 353 slot = 0; 354 } else 355 slot = SWSLOT_BAD; 356 357 #if defined(VMSWAP) 358 if (swap) { 359 if (pg->uobject != NULL) { 360 int oldslot; 361 oldslot = uao_set_swslot(pg->uobject, 362 pg->offset >> PAGE_SHIFT, slot); 363 KASSERT(oldslot == swslot + i); 364 } else { 365 KASSERT(pg->uanon->an_swslot == 366 swslot + i); 367 pg->uanon->an_swslot = slot; 368 } 369 } 370 #endif /* defined(VMSWAP) */ 371 } 372 373 /* 374 * if the page is PG_FAKE, this must have been a read to 375 * initialize the page. clear PG_FAKE and activate the page. 376 * we must also clear the pmap "modified" flag since it may 377 * still be set from the page's previous identity. 378 */ 379 380 if (pg->flags & PG_FAKE) { 381 KASSERT(!write); 382 pg->flags &= ~PG_FAKE; 383 #if defined(READAHEAD_STATS) 384 pg->pqflags |= PQ_READAHEAD; 385 uvm_ra_total.ev_count++; 386 #endif /* defined(READAHEAD_STATS) */ 387 KASSERT((pg->flags & PG_CLEAN) != 0); 388 uvm_pageenqueue(pg); 389 pmap_clear_modify(pg); 390 } 391 392 /* 393 * do accounting for pagedaemon i/o and arrange to free 394 * the pages instead of just unbusying them. 395 */ 396 397 if (pg->flags & PG_PAGEOUT) { 398 pg->flags &= ~PG_PAGEOUT; 399 pageout_done++; 400 uvmexp.pdfreed++; 401 pg->flags |= PG_RELEASED; 402 } 403 404 #if defined(VMSWAP) 405 /* 406 * for swap pages, unlock everything for this page now. 407 */ 408 409 if (swap) { 410 if (pg->uobject == NULL && pg->uanon->an_ref == 0 && 411 (pg->flags & PG_RELEASED) != 0) { 412 mutex_exit(&uvm_pageqlock); 413 uvm_anon_release(pg->uanon); 414 } else { 415 uvm_page_unbusy(&pg, 1); 416 mutex_exit(&uvm_pageqlock); 417 mutex_exit(slock); 418 } 419 } 420 #endif /* defined(VMSWAP) */ 421 } 422 uvm_pageout_done(pageout_done); 423 if (!swap) { 424 uvm_page_unbusy(pgs, npages); 425 mutex_exit(&uvm_pageqlock); 426 mutex_exit(slock); 427 } else { 428 #if defined(VMSWAP) 429 KASSERT(write); 430 431 /* these pages are now only in swap. */ 432 mutex_enter(&uvm_swap_data_lock); 433 KASSERT(uvmexp.swpgonly + npages <= uvmexp.swpginuse); 434 if (error != ENOMEM) 435 uvmexp.swpgonly += npages; 436 mutex_exit(&uvm_swap_data_lock); 437 if (error) { 438 if (error != ENOMEM) 439 uvm_swap_markbad(swslot, npages); 440 else 441 uvm_swap_free(swslot, npages); 442 } 443 uvmexp.pdpending--; 444 #endif /* defined(VMSWAP) */ 445 } 446 } 447 448 /* 449 * uvm_aio_aiodone: do iodone processing for async i/os. 450 * this should be called in thread context, not interrupt context. 451 */ 452 453 void 454 uvm_aio_aiodone(struct buf *bp) 455 { 456 int npages = bp->b_bufsize >> PAGE_SHIFT; 457 struct vm_page *pgs[npages]; 458 int i, error; 459 bool write; 460 UVMHIST_FUNC("uvm_aio_aiodone"); UVMHIST_CALLED(ubchist); 461 UVMHIST_LOG(ubchist, "bp %p", bp, 0,0,0); 462 463 error = bp->b_error; 464 write = (bp->b_flags & B_READ) == 0; 465 466 for (i = 0; i < npages; i++) { 467 pgs[i] = uvm_pageratop((vaddr_t)bp->b_data + (i << PAGE_SHIFT)); 468 UVMHIST_LOG(ubchist, "pgs[%d] = %p", i, pgs[i],0,0); 469 } 470 uvm_pagermapout((vaddr_t)bp->b_data, npages); 471 472 uvm_aio_aiodone_pages(pgs, npages, write, error); 473 474 if (write && (bp->b_cflags & BC_AGE) != 0) { 475 mutex_enter(bp->b_objlock); 476 vwakeup(bp); 477 mutex_exit(bp->b_objlock); 478 } 479 putiobuf(bp); 480 } 481 482 /* 483 * uvm_pageratop: convert KVAs in the pager map back to their page 484 * structures. 485 */ 486 487 struct vm_page * 488 uvm_pageratop(vaddr_t kva) 489 { 490 struct vm_page *pg; 491 paddr_t pa; 492 bool rv; 493 494 rv = pmap_extract(pmap_kernel(), kva, &pa); 495 KASSERT(rv); 496 pg = PHYS_TO_VM_PAGE(pa); 497 KASSERT(pg != NULL); 498 return (pg); 499 } 500