1 /* $OpenBSD: uvm_page.c,v 1.40 2002/01/02 22:23:25 miod Exp $ */ 2 /* $NetBSD: uvm_page.c,v 1.44 2000/11/27 08:40:04 chs Exp $ */ 3 4 /* 5 * Copyright (c) 1997 Charles D. Cranor and Washington University. 6 * Copyright (c) 1991, 1993, The Regents of the University of California. 7 * 8 * All rights reserved. 9 * 10 * This code is derived from software contributed to Berkeley by 11 * The Mach Operating System project at Carnegie-Mellon University. 12 * 13 * Redistribution and use in source and binary forms, with or without 14 * modification, are permitted provided that the following conditions 15 * are met: 16 * 1. Redistributions of source code must retain the above copyright 17 * notice, this list of conditions and the following disclaimer. 18 * 2. Redistributions in binary form must reproduce the above copyright 19 * notice, this list of conditions and the following disclaimer in the 20 * documentation and/or other materials provided with the distribution. 21 * 3. All advertising materials mentioning features or use of this software 22 * must display the following acknowledgement: 23 * This product includes software developed by Charles D. Cranor, 24 * Washington University, the University of California, Berkeley and 25 * its contributors. 26 * 4. Neither the name of the University nor the names of its contributors 27 * may be used to endorse or promote products derived from this software 28 * without specific prior written permission. 29 * 30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 40 * SUCH DAMAGE. 41 * 42 * @(#)vm_page.c 8.3 (Berkeley) 3/21/94 43 * from: Id: uvm_page.c,v 1.1.2.18 1998/02/06 05:24:42 chs Exp 44 * 45 * 46 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 47 * All rights reserved. 48 * 49 * Permission to use, copy, modify and distribute this software and 50 * its documentation is hereby granted, provided that both the copyright 51 * notice and this permission notice appear in all copies of the 52 * software, derivative works or modified versions, and any portions 53 * thereof, and that both notices appear in supporting documentation. 54 * 55 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 56 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 57 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 58 * 59 * Carnegie Mellon requests users of this software to return to 60 * 61 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 62 * School of Computer Science 63 * Carnegie Mellon University 64 * Pittsburgh PA 15213-3890 65 * 66 * any improvements or extensions that they make and grant Carnegie the 67 * rights to redistribute these changes. 68 */ 69 70 /* 71 * uvm_page.c: page ops. 72 */ 73 74 #define UVM_PAGE /* pull in uvm_page.h functions */ 75 #include <sys/param.h> 76 #include <sys/systm.h> 77 #include <sys/malloc.h> 78 #include <sys/sched.h> 79 #include <sys/kernel.h> 80 #include <sys/vnode.h> 81 82 #include <uvm/uvm.h> 83 84 /* 85 * global vars... XXXCDC: move to uvm. structure. 86 */ 87 88 /* 89 * physical memory config is stored in vm_physmem. 90 */ 91 92 struct vm_physseg vm_physmem[VM_PHYSSEG_MAX]; /* XXXCDC: uvm.physmem */ 93 int vm_nphysseg = 0; /* XXXCDC: uvm.nphysseg */ 94 95 /* 96 * Some supported CPUs in a given architecture don't support all 97 * of the things necessary to do idle page zero'ing efficiently. 98 * We therefore provide a way to disable it from machdep code here. 99 */ 100 101 /* 102 * XXX disabled until we can find a way to do this without causing 103 * problems for either cpu caches or DMA latency. 104 */ 105 boolean_t vm_page_zero_enable = FALSE; 106 107 /* 108 * local variables 109 */ 110 111 /* 112 * these variables record the values returned by vm_page_bootstrap, 113 * for debugging purposes. The implementation of uvm_pageboot_alloc 114 * and pmap_startup here also uses them internally. 115 */ 116 117 static vaddr_t virtual_space_start; 118 static vaddr_t virtual_space_end; 119 120 /* 121 * we use a hash table with only one bucket during bootup. we will 122 * later rehash (resize) the hash table once the allocator is ready. 123 * we static allocate the one bootstrap bucket below... 124 */ 125 126 static struct pglist uvm_bootbucket; 127 128 /* 129 * local prototypes 130 */ 131 132 static void uvm_pageinsert __P((struct vm_page *)); 133 static void uvm_pageremove __P((struct vm_page *)); 134 135 /* 136 * inline functions 137 */ 138 139 /* 140 * uvm_pageinsert: insert a page in the object and the hash table 141 * 142 * => caller must lock object 143 * => caller must lock page queues 144 * => call should have already set pg's object and offset pointers 145 * and bumped the version counter 146 */ 147 148 __inline static void 149 uvm_pageinsert(pg) 150 struct vm_page *pg; 151 { 152 struct pglist *buck; 153 int s; 154 155 KASSERT((pg->flags & PG_TABLED) == 0); 156 buck = &uvm.page_hash[uvm_pagehash(pg->uobject,pg->offset)]; 157 s = splvm(); 158 simple_lock(&uvm.hashlock); 159 TAILQ_INSERT_TAIL(buck, pg, hashq); /* put in hash */ 160 simple_unlock(&uvm.hashlock); 161 splx(s); 162 163 TAILQ_INSERT_TAIL(&pg->uobject->memq, pg, listq); /* put in object */ 164 pg->flags |= PG_TABLED; 165 pg->uobject->uo_npages++; 166 } 167 168 /* 169 * uvm_page_remove: remove page from object and hash 170 * 171 * => caller must lock object 172 * => caller must lock page queues 173 */ 174 175 static __inline void 176 uvm_pageremove(pg) 177 struct vm_page *pg; 178 { 179 struct pglist *buck; 180 int s; 181 182 KASSERT(pg->flags & PG_TABLED); 183 buck = &uvm.page_hash[uvm_pagehash(pg->uobject,pg->offset)]; 184 s = splvm(); 185 simple_lock(&uvm.hashlock); 186 TAILQ_REMOVE(buck, pg, hashq); 187 simple_unlock(&uvm.hashlock); 188 splx(s); 189 190 #ifdef UBC 191 if (pg->uobject->pgops == &uvm_vnodeops) { 192 uvm_pgcnt_vnode--; 193 } 194 #endif 195 196 /* object should be locked */ 197 TAILQ_REMOVE(&pg->uobject->memq, pg, listq); 198 199 pg->flags &= ~PG_TABLED; 200 pg->uobject->uo_npages--; 201 pg->uobject = NULL; 202 pg->version++; 203 } 204 205 /* 206 * uvm_page_init: init the page system. called from uvm_init(). 207 * 208 * => we return the range of kernel virtual memory in kvm_startp/kvm_endp 209 */ 210 211 void 212 uvm_page_init(kvm_startp, kvm_endp) 213 vaddr_t *kvm_startp, *kvm_endp; 214 { 215 vsize_t freepages, pagecount, n; 216 vm_page_t pagearray; 217 int lcv, i; 218 paddr_t paddr; 219 220 /* 221 * init the page queues and page queue locks 222 */ 223 224 for (lcv = 0; lcv < VM_NFREELIST; lcv++) { 225 for (i = 0; i < PGFL_NQUEUES; i++) 226 TAILQ_INIT(&uvm.page_free[lcv].pgfl_queues[i]); 227 } 228 TAILQ_INIT(&uvm.page_active); 229 TAILQ_INIT(&uvm.page_inactive_swp); 230 TAILQ_INIT(&uvm.page_inactive_obj); 231 simple_lock_init(&uvm.pageqlock); 232 simple_lock_init(&uvm.fpageqlock); 233 234 /* 235 * init the <obj,offset> => <page> hash table. for now 236 * we just have one bucket (the bootstrap bucket). later on we 237 * will allocate new buckets as we dynamically resize the hash table. 238 */ 239 240 uvm.page_nhash = 1; /* 1 bucket */ 241 uvm.page_hashmask = 0; /* mask for hash function */ 242 uvm.page_hash = &uvm_bootbucket; /* install bootstrap bucket */ 243 TAILQ_INIT(uvm.page_hash); /* init hash table */ 244 simple_lock_init(&uvm.hashlock); /* init hash table lock */ 245 246 /* 247 * allocate vm_page structures. 248 */ 249 250 /* 251 * sanity check: 252 * before calling this function the MD code is expected to register 253 * some free RAM with the uvm_page_physload() function. our job 254 * now is to allocate vm_page structures for this memory. 255 */ 256 257 if (vm_nphysseg == 0) 258 panic("uvm_page_bootstrap: no memory pre-allocated"); 259 260 /* 261 * first calculate the number of free pages... 262 * 263 * note that we use start/end rather than avail_start/avail_end. 264 * this allows us to allocate extra vm_page structures in case we 265 * want to return some memory to the pool after booting. 266 */ 267 268 freepages = 0; 269 for (lcv = 0 ; lcv < vm_nphysseg ; lcv++) 270 freepages += (vm_physmem[lcv].end - vm_physmem[lcv].start); 271 272 /* 273 * we now know we have (PAGE_SIZE * freepages) bytes of memory we can 274 * use. for each page of memory we use we need a vm_page structure. 275 * thus, the total number of pages we can use is the total size of 276 * the memory divided by the PAGE_SIZE plus the size of the vm_page 277 * structure. we add one to freepages as a fudge factor to avoid 278 * truncation errors (since we can only allocate in terms of whole 279 * pages). 280 */ 281 282 pagecount = ((freepages + 1) << PAGE_SHIFT) / 283 (PAGE_SIZE + sizeof(struct vm_page)); 284 pagearray = (vm_page_t)uvm_pageboot_alloc(pagecount * 285 sizeof(struct vm_page)); 286 memset(pagearray, 0, pagecount * sizeof(struct vm_page)); 287 288 /* 289 * init the vm_page structures and put them in the correct place. 290 */ 291 292 for (lcv = 0 ; lcv < vm_nphysseg ; lcv++) { 293 n = vm_physmem[lcv].end - vm_physmem[lcv].start; 294 if (n > pagecount) { 295 printf("uvm_page_init: lost %ld page(s) in init\n", 296 (long)(n - pagecount)); 297 panic("uvm_page_init"); /* XXXCDC: shouldn't happen? */ 298 /* n = pagecount; */ 299 } 300 301 /* set up page array pointers */ 302 vm_physmem[lcv].pgs = pagearray; 303 pagearray += n; 304 pagecount -= n; 305 vm_physmem[lcv].lastpg = vm_physmem[lcv].pgs + (n - 1); 306 307 /* init and free vm_pages (we've already zeroed them) */ 308 paddr = ptoa(vm_physmem[lcv].start); 309 for (i = 0 ; i < n ; i++, paddr += PAGE_SIZE) { 310 vm_physmem[lcv].pgs[i].phys_addr = paddr; 311 if (atop(paddr) >= vm_physmem[lcv].avail_start && 312 atop(paddr) <= vm_physmem[lcv].avail_end) { 313 uvmexp.npages++; 314 /* add page to free pool */ 315 uvm_pagefree(&vm_physmem[lcv].pgs[i]); 316 } 317 } 318 } 319 320 /* 321 * pass up the values of virtual_space_start and 322 * virtual_space_end (obtained by uvm_pageboot_alloc) to the upper 323 * layers of the VM. 324 */ 325 326 *kvm_startp = round_page(virtual_space_start); 327 *kvm_endp = trunc_page(virtual_space_end); 328 329 /* 330 * init locks for kernel threads 331 */ 332 333 simple_lock_init(&uvm.pagedaemon_lock); 334 simple_lock_init(&uvm.aiodoned_lock); 335 336 /* 337 * init reserve thresholds 338 * XXXCDC - values may need adjusting 339 */ 340 uvmexp.reserve_pagedaemon = 4; 341 uvmexp.reserve_kernel = 6; 342 uvmexp.anonminpct = 10; 343 uvmexp.vnodeminpct = 10; 344 uvmexp.vtextminpct = 5; 345 uvmexp.anonmin = uvmexp.anonminpct * 256 / 100; 346 uvmexp.vnodemin = uvmexp.vnodeminpct * 256 / 100; 347 uvmexp.vtextmin = uvmexp.vtextminpct * 256 / 100; 348 349 /* 350 * determine if we should zero pages in the idle loop. 351 */ 352 353 uvm.page_idle_zero = vm_page_zero_enable; 354 355 /* 356 * done! 357 */ 358 359 uvm.page_init_done = TRUE; 360 } 361 362 /* 363 * uvm_setpagesize: set the page size 364 * 365 * => sets page_shift and page_mask from uvmexp.pagesize. 366 */ 367 368 void 369 uvm_setpagesize() 370 { 371 if (uvmexp.pagesize == 0) 372 uvmexp.pagesize = DEFAULT_PAGE_SIZE; 373 uvmexp.pagemask = uvmexp.pagesize - 1; 374 if ((uvmexp.pagemask & uvmexp.pagesize) != 0) 375 panic("uvm_setpagesize: page size not a power of two"); 376 for (uvmexp.pageshift = 0; ; uvmexp.pageshift++) 377 if ((1 << uvmexp.pageshift) == uvmexp.pagesize) 378 break; 379 } 380 381 /* 382 * uvm_pageboot_alloc: steal memory from physmem for bootstrapping 383 */ 384 385 vaddr_t 386 uvm_pageboot_alloc(size) 387 vsize_t size; 388 { 389 #if defined(PMAP_STEAL_MEMORY) 390 vaddr_t addr; 391 392 /* 393 * defer bootstrap allocation to MD code (it may want to allocate 394 * from a direct-mapped segment). pmap_steal_memory should round 395 * off virtual_space_start/virtual_space_end. 396 */ 397 398 addr = pmap_steal_memory(size, &virtual_space_start, 399 &virtual_space_end); 400 401 return(addr); 402 403 #else /* !PMAP_STEAL_MEMORY */ 404 405 static boolean_t initialized = FALSE; 406 vaddr_t addr, vaddr; 407 paddr_t paddr; 408 409 /* round to page size */ 410 size = round_page(size); 411 412 /* 413 * on first call to this function, initialize ourselves. 414 */ 415 if (initialized == FALSE) { 416 pmap_virtual_space(&virtual_space_start, &virtual_space_end); 417 418 /* round it the way we like it */ 419 virtual_space_start = round_page(virtual_space_start); 420 virtual_space_end = trunc_page(virtual_space_end); 421 422 initialized = TRUE; 423 } 424 425 /* 426 * allocate virtual memory for this request 427 */ 428 if (virtual_space_start == virtual_space_end || 429 (virtual_space_end - virtual_space_start) < size) 430 panic("uvm_pageboot_alloc: out of virtual space"); 431 432 addr = virtual_space_start; 433 434 #ifdef PMAP_GROWKERNEL 435 /* 436 * If the kernel pmap can't map the requested space, 437 * then allocate more resources for it. 438 */ 439 if (uvm_maxkaddr < (addr + size)) { 440 uvm_maxkaddr = pmap_growkernel(addr + size); 441 if (uvm_maxkaddr < (addr + size)) 442 panic("uvm_pageboot_alloc: pmap_growkernel() failed"); 443 } 444 #endif 445 446 virtual_space_start += size; 447 448 /* 449 * allocate and mapin physical pages to back new virtual pages 450 */ 451 452 for (vaddr = round_page(addr) ; vaddr < addr + size ; 453 vaddr += PAGE_SIZE) { 454 455 if (!uvm_page_physget(&paddr)) 456 panic("uvm_pageboot_alloc: out of memory"); 457 458 /* 459 * Note this memory is no longer managed, so using 460 * pmap_kenter is safe. 461 */ 462 pmap_kenter_pa(vaddr, paddr, VM_PROT_READ|VM_PROT_WRITE); 463 } 464 return(addr); 465 #endif /* PMAP_STEAL_MEMORY */ 466 } 467 468 #if !defined(PMAP_STEAL_MEMORY) 469 /* 470 * uvm_page_physget: "steal" one page from the vm_physmem structure. 471 * 472 * => attempt to allocate it off the end of a segment in which the "avail" 473 * values match the start/end values. if we can't do that, then we 474 * will advance both values (making them equal, and removing some 475 * vm_page structures from the non-avail area). 476 * => return false if out of memory. 477 */ 478 479 /* subroutine: try to allocate from memory chunks on the specified freelist */ 480 static boolean_t uvm_page_physget_freelist __P((paddr_t *, int)); 481 482 static boolean_t 483 uvm_page_physget_freelist(paddrp, freelist) 484 paddr_t *paddrp; 485 int freelist; 486 { 487 int lcv, x; 488 489 /* pass 1: try allocating from a matching end */ 490 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST) || \ 491 (VM_PHYSSEG_STRAT == VM_PSTRAT_BSEARCH) 492 for (lcv = vm_nphysseg - 1 ; lcv >= 0 ; lcv--) 493 #else 494 for (lcv = 0 ; lcv < vm_nphysseg ; lcv++) 495 #endif 496 { 497 498 if (uvm.page_init_done == TRUE) 499 panic("uvm_page_physget: called _after_ bootstrap"); 500 501 if (vm_physmem[lcv].free_list != freelist) 502 continue; 503 504 /* try from front */ 505 if (vm_physmem[lcv].avail_start == vm_physmem[lcv].start && 506 vm_physmem[lcv].avail_start < vm_physmem[lcv].avail_end) { 507 *paddrp = ptoa(vm_physmem[lcv].avail_start); 508 vm_physmem[lcv].avail_start++; 509 vm_physmem[lcv].start++; 510 /* nothing left? nuke it */ 511 if (vm_physmem[lcv].avail_start == 512 vm_physmem[lcv].end) { 513 if (vm_nphysseg == 1) 514 panic("vum_page_physget: out of memory!"); 515 vm_nphysseg--; 516 for (x = lcv ; x < vm_nphysseg ; x++) 517 /* structure copy */ 518 vm_physmem[x] = vm_physmem[x+1]; 519 } 520 return (TRUE); 521 } 522 523 /* try from rear */ 524 if (vm_physmem[lcv].avail_end == vm_physmem[lcv].end && 525 vm_physmem[lcv].avail_start < vm_physmem[lcv].avail_end) { 526 *paddrp = ptoa(vm_physmem[lcv].avail_end - 1); 527 vm_physmem[lcv].avail_end--; 528 vm_physmem[lcv].end--; 529 /* nothing left? nuke it */ 530 if (vm_physmem[lcv].avail_end == 531 vm_physmem[lcv].start) { 532 if (vm_nphysseg == 1) 533 panic("uvm_page_physget: out of memory!"); 534 vm_nphysseg--; 535 for (x = lcv ; x < vm_nphysseg ; x++) 536 /* structure copy */ 537 vm_physmem[x] = vm_physmem[x+1]; 538 } 539 return (TRUE); 540 } 541 } 542 543 /* pass2: forget about matching ends, just allocate something */ 544 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST) || \ 545 (VM_PHYSSEG_STRAT == VM_PSTRAT_BSEARCH) 546 for (lcv = vm_nphysseg - 1 ; lcv >= 0 ; lcv--) 547 #else 548 for (lcv = 0 ; lcv < vm_nphysseg ; lcv++) 549 #endif 550 { 551 552 /* any room in this bank? */ 553 if (vm_physmem[lcv].avail_start >= vm_physmem[lcv].avail_end) 554 continue; /* nope */ 555 556 *paddrp = ptoa(vm_physmem[lcv].avail_start); 557 vm_physmem[lcv].avail_start++; 558 /* truncate! */ 559 vm_physmem[lcv].start = vm_physmem[lcv].avail_start; 560 561 /* nothing left? nuke it */ 562 if (vm_physmem[lcv].avail_start == vm_physmem[lcv].end) { 563 if (vm_nphysseg == 1) 564 panic("uvm_page_physget: out of memory!"); 565 vm_nphysseg--; 566 for (x = lcv ; x < vm_nphysseg ; x++) 567 /* structure copy */ 568 vm_physmem[x] = vm_physmem[x+1]; 569 } 570 return (TRUE); 571 } 572 573 return (FALSE); /* whoops! */ 574 } 575 576 boolean_t 577 uvm_page_physget(paddrp) 578 paddr_t *paddrp; 579 { 580 int i; 581 582 /* try in the order of freelist preference */ 583 for (i = 0; i < VM_NFREELIST; i++) 584 if (uvm_page_physget_freelist(paddrp, i) == TRUE) 585 return (TRUE); 586 return (FALSE); 587 } 588 #endif /* PMAP_STEAL_MEMORY */ 589 590 /* 591 * uvm_page_physload: load physical memory into VM system 592 * 593 * => all args are PFs 594 * => all pages in start/end get vm_page structures 595 * => areas marked by avail_start/avail_end get added to the free page pool 596 * => we are limited to VM_PHYSSEG_MAX physical memory segments 597 */ 598 599 void 600 uvm_page_physload(start, end, avail_start, avail_end, free_list) 601 paddr_t start, end, avail_start, avail_end; 602 int free_list; 603 { 604 int preload, lcv; 605 psize_t npages; 606 struct vm_page *pgs; 607 struct vm_physseg *ps; 608 609 if (uvmexp.pagesize == 0) 610 panic("uvm_page_physload: page size not set!"); 611 612 if (free_list >= VM_NFREELIST || free_list < VM_FREELIST_DEFAULT) 613 panic("uvm_page_physload: bad free list %d\n", free_list); 614 615 if (start >= end) 616 panic("uvm_page_physload: start >= end"); 617 618 /* 619 * do we have room? 620 */ 621 if (vm_nphysseg == VM_PHYSSEG_MAX) { 622 printf("uvm_page_physload: unable to load physical memory " 623 "segment\n"); 624 printf("\t%d segments allocated, ignoring 0x%llx -> 0x%llx\n", 625 VM_PHYSSEG_MAX, (long long)start, (long long)end); 626 printf("\tincrease VM_PHYSSEG_MAX\n"); 627 return; 628 } 629 630 /* 631 * check to see if this is a "preload" (i.e. uvm_mem_init hasn't been 632 * called yet, so malloc is not available). 633 */ 634 for (lcv = 0 ; lcv < vm_nphysseg ; lcv++) { 635 if (vm_physmem[lcv].pgs) 636 break; 637 } 638 preload = (lcv == vm_nphysseg); 639 640 /* 641 * if VM is already running, attempt to malloc() vm_page structures 642 */ 643 if (!preload) { 644 #if defined(VM_PHYSSEG_NOADD) 645 panic("uvm_page_physload: tried to add RAM after vm_mem_init"); 646 #else 647 /* XXXCDC: need some sort of lockout for this case */ 648 paddr_t paddr; 649 npages = end - start; /* # of pages */ 650 pgs = malloc(sizeof(struct vm_page) * npages, 651 M_VMPAGE, M_NOWAIT); 652 if (pgs == NULL) { 653 printf("uvm_page_physload: can not malloc vm_page " 654 "structs for segment\n"); 655 printf("\tignoring 0x%lx -> 0x%lx\n", start, end); 656 return; 657 } 658 /* zero data, init phys_addr and free_list, and free pages */ 659 memset(pgs, 0, sizeof(struct vm_page) * npages); 660 for (lcv = 0, paddr = ptoa(start) ; 661 lcv < npages ; lcv++, paddr += PAGE_SIZE) { 662 pgs[lcv].phys_addr = paddr; 663 pgs[lcv].free_list = free_list; 664 if (atop(paddr) >= avail_start && 665 atop(paddr) <= avail_end) 666 uvm_pagefree(&pgs[lcv]); 667 } 668 /* XXXCDC: incomplete: need to update uvmexp.free, what else? */ 669 /* XXXCDC: need hook to tell pmap to rebuild pv_list, etc... */ 670 #endif 671 } else { 672 673 /* gcc complains if these don't get init'd */ 674 pgs = NULL; 675 npages = 0; 676 677 } 678 679 /* 680 * now insert us in the proper place in vm_physmem[] 681 */ 682 683 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_RANDOM) 684 685 /* random: put it at the end (easy!) */ 686 ps = &vm_physmem[vm_nphysseg]; 687 688 #elif (VM_PHYSSEG_STRAT == VM_PSTRAT_BSEARCH) 689 690 { 691 int x; 692 /* sort by address for binary search */ 693 for (lcv = 0 ; lcv < vm_nphysseg ; lcv++) 694 if (start < vm_physmem[lcv].start) 695 break; 696 ps = &vm_physmem[lcv]; 697 /* move back other entries, if necessary ... */ 698 for (x = vm_nphysseg ; x > lcv ; x--) 699 /* structure copy */ 700 vm_physmem[x] = vm_physmem[x - 1]; 701 } 702 703 #elif (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST) 704 705 { 706 int x; 707 /* sort by largest segment first */ 708 for (lcv = 0 ; lcv < vm_nphysseg ; lcv++) 709 if ((end - start) > 710 (vm_physmem[lcv].end - vm_physmem[lcv].start)) 711 break; 712 ps = &vm_physmem[lcv]; 713 /* move back other entries, if necessary ... */ 714 for (x = vm_nphysseg ; x > lcv ; x--) 715 /* structure copy */ 716 vm_physmem[x] = vm_physmem[x - 1]; 717 } 718 719 #else 720 721 panic("uvm_page_physload: unknown physseg strategy selected!"); 722 723 #endif 724 725 ps->start = start; 726 ps->end = end; 727 ps->avail_start = avail_start; 728 ps->avail_end = avail_end; 729 if (preload) { 730 ps->pgs = NULL; 731 } else { 732 ps->pgs = pgs; 733 ps->lastpg = pgs + npages - 1; 734 } 735 ps->free_list = free_list; 736 vm_nphysseg++; 737 738 /* 739 * done! 740 */ 741 742 if (!preload) 743 uvm_page_rehash(); 744 745 return; 746 } 747 748 /* 749 * uvm_page_rehash: reallocate hash table based on number of free pages. 750 */ 751 752 void 753 uvm_page_rehash() 754 { 755 int freepages, lcv, bucketcount, s, oldcount; 756 struct pglist *newbuckets, *oldbuckets; 757 struct vm_page *pg; 758 size_t newsize, oldsize; 759 760 /* 761 * compute number of pages that can go in the free pool 762 */ 763 764 freepages = 0; 765 for (lcv = 0 ; lcv < vm_nphysseg ; lcv++) 766 freepages += 767 (vm_physmem[lcv].avail_end - vm_physmem[lcv].avail_start); 768 769 /* 770 * compute number of buckets needed for this number of pages 771 */ 772 773 bucketcount = 1; 774 while (bucketcount < freepages) 775 bucketcount = bucketcount * 2; 776 777 /* 778 * compute the size of the current table and new table. 779 */ 780 781 oldbuckets = uvm.page_hash; 782 oldcount = uvm.page_nhash; 783 oldsize = round_page(sizeof(struct pglist) * oldcount); 784 newsize = round_page(sizeof(struct pglist) * bucketcount); 785 786 /* 787 * allocate the new buckets 788 */ 789 790 newbuckets = (struct pglist *) uvm_km_alloc(kernel_map, newsize); 791 if (newbuckets == NULL) { 792 printf("uvm_page_physrehash: WARNING: could not grow page " 793 "hash table\n"); 794 return; 795 } 796 for (lcv = 0 ; lcv < bucketcount ; lcv++) 797 TAILQ_INIT(&newbuckets[lcv]); 798 799 /* 800 * now replace the old buckets with the new ones and rehash everything 801 */ 802 803 s = splvm(); 804 simple_lock(&uvm.hashlock); 805 uvm.page_hash = newbuckets; 806 uvm.page_nhash = bucketcount; 807 uvm.page_hashmask = bucketcount - 1; /* power of 2 */ 808 809 /* ... and rehash */ 810 for (lcv = 0 ; lcv < oldcount ; lcv++) { 811 while ((pg = oldbuckets[lcv].tqh_first) != NULL) { 812 TAILQ_REMOVE(&oldbuckets[lcv], pg, hashq); 813 TAILQ_INSERT_TAIL( 814 &uvm.page_hash[uvm_pagehash(pg->uobject, pg->offset)], 815 pg, hashq); 816 } 817 } 818 simple_unlock(&uvm.hashlock); 819 splx(s); 820 821 /* 822 * free old bucket array if is not the boot-time table 823 */ 824 825 if (oldbuckets != &uvm_bootbucket) 826 uvm_km_free(kernel_map, (vaddr_t) oldbuckets, oldsize); 827 828 /* 829 * done 830 */ 831 return; 832 } 833 834 835 #if 1 /* XXXCDC: TMP TMP TMP DEBUG DEBUG DEBUG */ 836 837 void uvm_page_physdump __P((void)); /* SHUT UP GCC */ 838 839 /* call from DDB */ 840 void 841 uvm_page_physdump() 842 { 843 int lcv; 844 845 printf("rehash: physical memory config [segs=%d of %d]:\n", 846 vm_nphysseg, VM_PHYSSEG_MAX); 847 for (lcv = 0 ; lcv < vm_nphysseg ; lcv++) 848 printf("0x%llx->0x%llx [0x%llx->0x%llx]\n", 849 (long long)vm_physmem[lcv].start, 850 (long long)vm_physmem[lcv].end, 851 (long long)vm_physmem[lcv].avail_start, 852 (long long)vm_physmem[lcv].avail_end); 853 printf("STRATEGY = "); 854 switch (VM_PHYSSEG_STRAT) { 855 case VM_PSTRAT_RANDOM: printf("RANDOM\n"); break; 856 case VM_PSTRAT_BSEARCH: printf("BSEARCH\n"); break; 857 case VM_PSTRAT_BIGFIRST: printf("BIGFIRST\n"); break; 858 default: printf("<<UNKNOWN>>!!!!\n"); 859 } 860 printf("number of buckets = %d\n", uvm.page_nhash); 861 } 862 #endif 863 864 /* 865 * uvm_pagealloc_strat: allocate vm_page from a particular free list. 866 * 867 * => return null if no pages free 868 * => wake up pagedaemon if number of free pages drops below low water mark 869 * => if obj != NULL, obj must be locked (to put in hash) 870 * => if anon != NULL, anon must be locked (to put in anon) 871 * => only one of obj or anon can be non-null 872 * => caller must activate/deactivate page if it is not wired. 873 * => free_list is ignored if strat == UVM_PGA_STRAT_NORMAL. 874 * => policy decision: it is more important to pull a page off of the 875 * appropriate priority free list than it is to get a zero'd or 876 * unknown contents page. This is because we live with the 877 * consequences of a bad free list decision for the entire 878 * lifetime of the page, e.g. if the page comes from memory that 879 * is slower to access. 880 */ 881 882 struct vm_page * 883 uvm_pagealloc_strat(obj, off, anon, flags, strat, free_list) 884 struct uvm_object *obj; 885 voff_t off; 886 int flags; 887 struct vm_anon *anon; 888 int strat, free_list; 889 { 890 int lcv, try1, try2, s, zeroit = 0; 891 struct vm_page *pg; 892 struct pglist *freeq; 893 struct pgfreelist *pgfl; 894 boolean_t use_reserve; 895 896 KASSERT(obj == NULL || anon == NULL); 897 KASSERT(off == trunc_page(off)); 898 s = uvm_lock_fpageq(); 899 900 /* 901 * check to see if we need to generate some free pages waking 902 * the pagedaemon. 903 */ 904 905 #ifdef UBC 906 if (uvmexp.free + uvmexp.paging < uvmexp.freemin || 907 (uvmexp.free + uvmexp.paging < uvmexp.freetarg && 908 uvmexp.inactive < uvmexp.inactarg)) { 909 wakeup(&uvm.pagedaemon); 910 } 911 #else 912 if (uvmexp.free < uvmexp.freemin || (uvmexp.free < uvmexp.freetarg && 913 uvmexp.inactive < uvmexp.inactarg)) 914 wakeup(&uvm.pagedaemon); 915 #endif 916 917 /* 918 * fail if any of these conditions is true: 919 * [1] there really are no free pages, or 920 * [2] only kernel "reserved" pages remain and 921 * the page isn't being allocated to a kernel object. 922 * [3] only pagedaemon "reserved" pages remain and 923 * the requestor isn't the pagedaemon. 924 */ 925 926 use_reserve = (flags & UVM_PGA_USERESERVE) || 927 (obj && UVM_OBJ_IS_KERN_OBJECT(obj)); 928 if ((uvmexp.free <= uvmexp.reserve_kernel && !use_reserve) || 929 (uvmexp.free <= uvmexp.reserve_pagedaemon && 930 !(use_reserve && (curproc == uvm.pagedaemon_proc || 931 curproc == syncerproc)))) 932 goto fail; 933 934 #if PGFL_NQUEUES != 2 935 #error uvm_pagealloc_strat needs to be updated 936 #endif 937 938 /* 939 * If we want a zero'd page, try the ZEROS queue first, otherwise 940 * we try the UNKNOWN queue first. 941 */ 942 if (flags & UVM_PGA_ZERO) { 943 try1 = PGFL_ZEROS; 944 try2 = PGFL_UNKNOWN; 945 } else { 946 try1 = PGFL_UNKNOWN; 947 try2 = PGFL_ZEROS; 948 } 949 950 again: 951 switch (strat) { 952 case UVM_PGA_STRAT_NORMAL: 953 /* Check all freelists in descending priority order. */ 954 for (lcv = 0; lcv < VM_NFREELIST; lcv++) { 955 pgfl = &uvm.page_free[lcv]; 956 if ((pg = TAILQ_FIRST((freeq = 957 &pgfl->pgfl_queues[try1]))) != NULL || 958 (pg = TAILQ_FIRST((freeq = 959 &pgfl->pgfl_queues[try2]))) != NULL) 960 goto gotit; 961 } 962 963 /* No pages free! */ 964 goto fail; 965 966 case UVM_PGA_STRAT_ONLY: 967 case UVM_PGA_STRAT_FALLBACK: 968 /* Attempt to allocate from the specified free list. */ 969 KASSERT(free_list >= 0 && free_list < VM_NFREELIST); 970 pgfl = &uvm.page_free[free_list]; 971 if ((pg = TAILQ_FIRST((freeq = 972 &pgfl->pgfl_queues[try1]))) != NULL || 973 (pg = TAILQ_FIRST((freeq = 974 &pgfl->pgfl_queues[try2]))) != NULL) 975 goto gotit; 976 977 /* Fall back, if possible. */ 978 if (strat == UVM_PGA_STRAT_FALLBACK) { 979 strat = UVM_PGA_STRAT_NORMAL; 980 goto again; 981 } 982 983 /* No pages free! */ 984 goto fail; 985 986 default: 987 panic("uvm_pagealloc_strat: bad strat %d", strat); 988 /* NOTREACHED */ 989 } 990 991 gotit: 992 TAILQ_REMOVE(freeq, pg, pageq); 993 uvmexp.free--; 994 995 /* update zero'd page count */ 996 if (pg->flags & PG_ZERO) 997 uvmexp.zeropages--; 998 999 /* 1000 * update allocation statistics and remember if we have to 1001 * zero the page 1002 */ 1003 if (flags & UVM_PGA_ZERO) { 1004 if (pg->flags & PG_ZERO) { 1005 uvmexp.pga_zerohit++; 1006 zeroit = 0; 1007 } else { 1008 uvmexp.pga_zeromiss++; 1009 zeroit = 1; 1010 } 1011 } 1012 1013 uvm_unlock_fpageq(s); /* unlock free page queue */ 1014 1015 pg->offset = off; 1016 pg->uobject = obj; 1017 pg->uanon = anon; 1018 pg->flags = PG_BUSY|PG_CLEAN|PG_FAKE; 1019 pg->version++; 1020 if (anon) { 1021 anon->u.an_page = pg; 1022 pg->pqflags = PQ_ANON; 1023 #ifdef UBC 1024 uvm_pgcnt_anon++; 1025 #endif 1026 } else { 1027 if (obj) 1028 uvm_pageinsert(pg); 1029 pg->pqflags = 0; 1030 } 1031 #if defined(UVM_PAGE_TRKOWN) 1032 pg->owner_tag = NULL; 1033 #endif 1034 UVM_PAGE_OWN(pg, "new alloc"); 1035 1036 if (flags & UVM_PGA_ZERO) { 1037 /* 1038 * A zero'd page is not clean. If we got a page not already 1039 * zero'd, then we have to zero it ourselves. 1040 */ 1041 pg->flags &= ~PG_CLEAN; 1042 if (zeroit) 1043 pmap_zero_page(VM_PAGE_TO_PHYS(pg)); 1044 } 1045 1046 return(pg); 1047 1048 fail: 1049 uvm_unlock_fpageq(s); 1050 return (NULL); 1051 } 1052 1053 /* 1054 * uvm_pagerealloc: reallocate a page from one object to another 1055 * 1056 * => both objects must be locked 1057 */ 1058 1059 void 1060 uvm_pagerealloc(pg, newobj, newoff) 1061 struct vm_page *pg; 1062 struct uvm_object *newobj; 1063 voff_t newoff; 1064 { 1065 /* 1066 * remove it from the old object 1067 */ 1068 1069 if (pg->uobject) { 1070 uvm_pageremove(pg); 1071 } 1072 1073 /* 1074 * put it in the new object 1075 */ 1076 1077 if (newobj) { 1078 pg->uobject = newobj; 1079 pg->offset = newoff; 1080 pg->version++; 1081 uvm_pageinsert(pg); 1082 } 1083 } 1084 1085 1086 /* 1087 * uvm_pagefree: free page 1088 * 1089 * => erase page's identity (i.e. remove from hash/object) 1090 * => put page on free list 1091 * => caller must lock owning object (either anon or uvm_object) 1092 * => caller must lock page queues 1093 * => assumes all valid mappings of pg are gone 1094 */ 1095 1096 void 1097 uvm_pagefree(pg) 1098 struct vm_page *pg; 1099 { 1100 int s; 1101 int saved_loan_count = pg->loan_count; 1102 1103 #ifdef DEBUG 1104 if (pg->uobject == (void *)0xdeadbeef && 1105 pg->uanon == (void *)0xdeadbeef) { 1106 panic("uvm_pagefree: freeing free page %p\n", pg); 1107 } 1108 #endif 1109 1110 /* 1111 * if the page was an object page (and thus "TABLED"), remove it 1112 * from the object. 1113 */ 1114 1115 if (pg->flags & PG_TABLED) { 1116 1117 /* 1118 * if the object page is on loan we are going to drop ownership. 1119 * it is possible that an anon will take over as owner for this 1120 * page later on. the anon will want a !PG_CLEAN page so that 1121 * it knows it needs to allocate swap if it wants to page the 1122 * page out. 1123 */ 1124 1125 if (saved_loan_count) 1126 pg->flags &= ~PG_CLEAN; /* in case an anon takes over */ 1127 uvm_pageremove(pg); 1128 1129 /* 1130 * if our page was on loan, then we just lost control over it 1131 * (in fact, if it was loaned to an anon, the anon may have 1132 * already taken over ownership of the page by now and thus 1133 * changed the loan_count [e.g. in uvmfault_anonget()]) we just 1134 * return (when the last loan is dropped, then the page can be 1135 * freed by whatever was holding the last loan). 1136 */ 1137 1138 if (saved_loan_count) 1139 return; 1140 } else if (saved_loan_count && (pg->pqflags & PQ_ANON)) { 1141 1142 /* 1143 * if our page is owned by an anon and is loaned out to the 1144 * kernel then we just want to drop ownership and return. 1145 * the kernel must free the page when all its loans clear ... 1146 * note that the kernel can't change the loan status of our 1147 * page as long as we are holding PQ lock. 1148 */ 1149 1150 pg->pqflags &= ~PQ_ANON; 1151 pg->uanon = NULL; 1152 return; 1153 } 1154 KASSERT(saved_loan_count == 0); 1155 1156 /* 1157 * now remove the page from the queues 1158 */ 1159 1160 if (pg->pqflags & PQ_ACTIVE) { 1161 TAILQ_REMOVE(&uvm.page_active, pg, pageq); 1162 pg->pqflags &= ~PQ_ACTIVE; 1163 uvmexp.active--; 1164 } 1165 if (pg->pqflags & PQ_INACTIVE) { 1166 if (pg->pqflags & PQ_SWAPBACKED) 1167 TAILQ_REMOVE(&uvm.page_inactive_swp, pg, pageq); 1168 else 1169 TAILQ_REMOVE(&uvm.page_inactive_obj, pg, pageq); 1170 pg->pqflags &= ~PQ_INACTIVE; 1171 uvmexp.inactive--; 1172 } 1173 1174 /* 1175 * if the page was wired, unwire it now. 1176 */ 1177 1178 if (pg->wire_count) { 1179 pg->wire_count = 0; 1180 uvmexp.wired--; 1181 } 1182 #ifdef UBC 1183 if (pg->uanon) { 1184 uvm_pgcnt_anon--; 1185 } 1186 #endif 1187 1188 /* 1189 * and put on free queue 1190 */ 1191 1192 pg->flags &= ~PG_ZERO; 1193 1194 s = uvm_lock_fpageq(); 1195 TAILQ_INSERT_TAIL(&uvm.page_free[ 1196 uvm_page_lookup_freelist(pg)].pgfl_queues[PGFL_UNKNOWN], pg, pageq); 1197 pg->pqflags = PQ_FREE; 1198 #ifdef DEBUG 1199 pg->uobject = (void *)0xdeadbeef; 1200 pg->offset = 0xdeadbeef; 1201 pg->uanon = (void *)0xdeadbeef; 1202 #endif 1203 uvmexp.free++; 1204 1205 if (uvmexp.zeropages < UVM_PAGEZERO_TARGET) 1206 uvm.page_idle_zero = vm_page_zero_enable; 1207 1208 uvm_unlock_fpageq(s); 1209 } 1210 1211 /* 1212 * uvm_page_unbusy: unbusy an array of pages. 1213 * 1214 * => pages must either all belong to the same object, or all belong to anons. 1215 * => if pages are object-owned, object must be locked. 1216 * => if pages are anon-owned, anons must be unlockd and have 0 refcount. 1217 */ 1218 1219 void 1220 uvm_page_unbusy(pgs, npgs) 1221 struct vm_page **pgs; 1222 int npgs; 1223 { 1224 struct vm_page *pg; 1225 struct uvm_object *uobj; 1226 int i; 1227 UVMHIST_FUNC("uvm_page_unbusy"); UVMHIST_CALLED(ubchist); 1228 1229 for (i = 0; i < npgs; i++) { 1230 pg = pgs[i]; 1231 1232 if (pg == NULL) { 1233 continue; 1234 } 1235 if (pg->flags & PG_WANTED) { 1236 wakeup(pg); 1237 } 1238 if (pg->flags & PG_RELEASED) { 1239 UVMHIST_LOG(ubchist, "releasing pg %p", pg,0,0,0); 1240 uobj = pg->uobject; 1241 if (uobj != NULL) { 1242 uobj->pgops->pgo_releasepg(pg, NULL); 1243 } else { 1244 pg->flags &= ~(PG_BUSY); 1245 UVM_PAGE_OWN(pg, NULL); 1246 uvm_anfree(pg->uanon); 1247 } 1248 } else { 1249 UVMHIST_LOG(ubchist, "unbusying pg %p", pg,0,0,0); 1250 pg->flags &= ~(PG_WANTED|PG_BUSY); 1251 UVM_PAGE_OWN(pg, NULL); 1252 } 1253 } 1254 } 1255 1256 #if defined(UVM_PAGE_TRKOWN) 1257 /* 1258 * uvm_page_own: set or release page ownership 1259 * 1260 * => this is a debugging function that keeps track of who sets PG_BUSY 1261 * and where they do it. it can be used to track down problems 1262 * such a process setting "PG_BUSY" and never releasing it. 1263 * => page's object [if any] must be locked 1264 * => if "tag" is NULL then we are releasing page ownership 1265 */ 1266 void 1267 uvm_page_own(pg, tag) 1268 struct vm_page *pg; 1269 char *tag; 1270 { 1271 /* gain ownership? */ 1272 if (tag) { 1273 if (pg->owner_tag) { 1274 printf("uvm_page_own: page %p already owned " 1275 "by proc %d [%s]\n", pg, 1276 pg->owner, pg->owner_tag); 1277 panic("uvm_page_own"); 1278 } 1279 pg->owner = (curproc) ? curproc->p_pid : (pid_t) -1; 1280 pg->owner_tag = tag; 1281 return; 1282 } 1283 1284 /* drop ownership */ 1285 if (pg->owner_tag == NULL) { 1286 printf("uvm_page_own: dropping ownership of an non-owned " 1287 "page (%p)\n", pg); 1288 panic("uvm_page_own"); 1289 } 1290 pg->owner_tag = NULL; 1291 return; 1292 } 1293 #endif 1294 1295 /* 1296 * uvm_pageidlezero: zero free pages while the system is idle. 1297 * 1298 * => we do at least one iteration per call, if we are below the target. 1299 * => we loop until we either reach the target or whichqs indicates that 1300 * there is a process ready to run. 1301 */ 1302 void 1303 uvm_pageidlezero() 1304 { 1305 struct vm_page *pg; 1306 struct pgfreelist *pgfl; 1307 int free_list, s; 1308 1309 do { 1310 s = uvm_lock_fpageq(); 1311 1312 if (uvmexp.zeropages >= UVM_PAGEZERO_TARGET) { 1313 uvm.page_idle_zero = FALSE; 1314 uvm_unlock_fpageq(s); 1315 return; 1316 } 1317 1318 for (free_list = 0; free_list < VM_NFREELIST; free_list++) { 1319 pgfl = &uvm.page_free[free_list]; 1320 if ((pg = TAILQ_FIRST(&pgfl->pgfl_queues[ 1321 PGFL_UNKNOWN])) != NULL) 1322 break; 1323 } 1324 1325 if (pg == NULL) { 1326 /* 1327 * No non-zero'd pages; don't bother trying again 1328 * until we know we have non-zero'd pages free. 1329 */ 1330 uvm.page_idle_zero = FALSE; 1331 uvm_unlock_fpageq(s); 1332 return; 1333 } 1334 1335 TAILQ_REMOVE(&pgfl->pgfl_queues[PGFL_UNKNOWN], pg, pageq); 1336 uvmexp.free--; 1337 uvm_unlock_fpageq(s); 1338 1339 #ifdef PMAP_PAGEIDLEZERO 1340 if (PMAP_PAGEIDLEZERO(VM_PAGE_TO_PHYS(pg)) == FALSE) { 1341 /* 1342 * The machine-dependent code detected some 1343 * reason for us to abort zeroing pages, 1344 * probably because there is a process now 1345 * ready to run. 1346 */ 1347 s = uvm_lock_fpageq(); 1348 TAILQ_INSERT_HEAD(&pgfl->pgfl_queues[PGFL_UNKNOWN], 1349 pg, pageq); 1350 uvmexp.free++; 1351 uvmexp.zeroaborts++; 1352 uvm_unlock_fpageq(s); 1353 return; 1354 } 1355 #else 1356 /* 1357 * XXX This will toast the cache unless the pmap_zero_page() 1358 * XXX implementation does uncached access. 1359 */ 1360 pmap_zero_page(VM_PAGE_TO_PHYS(pg)); 1361 #endif 1362 pg->flags |= PG_ZERO; 1363 1364 s = uvm_lock_fpageq(); 1365 TAILQ_INSERT_HEAD(&pgfl->pgfl_queues[PGFL_ZEROS], pg, pageq); 1366 uvmexp.free++; 1367 uvmexp.zeropages++; 1368 uvm_unlock_fpageq(s); 1369 } while (whichqs == 0); 1370 } 1371