1 /* $NetBSD: uvm_page.c,v 1.186 2014/09/05 05:36:21 matt Exp $ */ 2 3 /* 4 * Copyright (c) 1997 Charles D. Cranor and Washington University. 5 * Copyright (c) 1991, 1993, The Regents of the University of California. 6 * 7 * All rights reserved. 8 * 9 * This code is derived from software contributed to Berkeley by 10 * The Mach Operating System project at Carnegie-Mellon University. 11 * 12 * Redistribution and use in source and binary forms, with or without 13 * modification, are permitted provided that the following conditions 14 * are met: 15 * 1. Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 2. Redistributions in binary form must reproduce the above copyright 18 * notice, this list of conditions and the following disclaimer in the 19 * documentation and/or other materials provided with the distribution. 20 * 3. Neither the name of the University nor the names of its contributors 21 * may be used to endorse or promote products derived from this software 22 * without specific prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 * 36 * @(#)vm_page.c 8.3 (Berkeley) 3/21/94 37 * from: Id: uvm_page.c,v 1.1.2.18 1998/02/06 05:24:42 chs Exp 38 * 39 * 40 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 41 * All rights reserved. 42 * 43 * Permission to use, copy, modify and distribute this software and 44 * its documentation is hereby granted, provided that both the copyright 45 * notice and this permission notice appear in all copies of the 46 * software, derivative works or modified versions, and any portions 47 * thereof, and that both notices appear in supporting documentation. 48 * 49 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 50 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 51 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 52 * 53 * Carnegie Mellon requests users of this software to return to 54 * 55 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 56 * School of Computer Science 57 * Carnegie Mellon University 58 * Pittsburgh PA 15213-3890 59 * 60 * any improvements or extensions that they make and grant Carnegie the 61 * rights to redistribute these changes. 62 */ 63 64 /* 65 * uvm_page.c: page ops. 66 */ 67 68 #include <sys/cdefs.h> 69 __KERNEL_RCSID(0, "$NetBSD: uvm_page.c,v 1.186 2014/09/05 05:36:21 matt Exp $"); 70 71 #include "opt_ddb.h" 72 #include "opt_uvmhist.h" 73 #include "opt_readahead.h" 74 75 #include <sys/param.h> 76 #include <sys/systm.h> 77 #include <sys/sched.h> 78 #include <sys/kernel.h> 79 #include <sys/vnode.h> 80 #include <sys/proc.h> 81 #include <sys/atomic.h> 82 #include <sys/cpu.h> 83 84 #include <uvm/uvm.h> 85 #include <uvm/uvm_ddb.h> 86 #include <uvm/uvm_pdpolicy.h> 87 88 /* 89 * global vars... XXXCDC: move to uvm. structure. 90 */ 91 92 /* 93 * physical memory config is stored in vm_physmem. 94 */ 95 96 struct vm_physseg vm_physmem[VM_PHYSSEG_MAX]; /* XXXCDC: uvm.physmem */ 97 int vm_nphysseg = 0; /* XXXCDC: uvm.nphysseg */ 98 #define vm_nphysmem vm_nphysseg 99 100 /* 101 * Some supported CPUs in a given architecture don't support all 102 * of the things necessary to do idle page zero'ing efficiently. 103 * We therefore provide a way to enable it from machdep code here. 104 */ 105 bool vm_page_zero_enable = false; 106 107 /* 108 * number of pages per-CPU to reserve for the kernel. 109 */ 110 int vm_page_reserve_kernel = 5; 111 112 /* 113 * physical memory size; 114 */ 115 int physmem; 116 117 /* 118 * local variables 119 */ 120 121 /* 122 * these variables record the values returned by vm_page_bootstrap, 123 * for debugging purposes. The implementation of uvm_pageboot_alloc 124 * and pmap_startup here also uses them internally. 125 */ 126 127 static vaddr_t virtual_space_start; 128 static vaddr_t virtual_space_end; 129 130 /* 131 * we allocate an initial number of page colors in uvm_page_init(), 132 * and remember them. We may re-color pages as cache sizes are 133 * discovered during the autoconfiguration phase. But we can never 134 * free the initial set of buckets, since they are allocated using 135 * uvm_pageboot_alloc(). 136 */ 137 138 static size_t recolored_pages_memsize /* = 0 */; 139 140 #ifdef DEBUG 141 vaddr_t uvm_zerocheckkva; 142 #endif /* DEBUG */ 143 144 /* 145 * local prototypes 146 */ 147 148 static void uvm_pageinsert(struct uvm_object *, struct vm_page *); 149 static void uvm_pageremove(struct uvm_object *, struct vm_page *); 150 151 /* 152 * per-object tree of pages 153 */ 154 155 static signed int 156 uvm_page_compare_nodes(void *ctx, const void *n1, const void *n2) 157 { 158 const struct vm_page *pg1 = n1; 159 const struct vm_page *pg2 = n2; 160 const voff_t a = pg1->offset; 161 const voff_t b = pg2->offset; 162 163 if (a < b) 164 return -1; 165 if (a > b) 166 return 1; 167 return 0; 168 } 169 170 static signed int 171 uvm_page_compare_key(void *ctx, const void *n, const void *key) 172 { 173 const struct vm_page *pg = n; 174 const voff_t a = pg->offset; 175 const voff_t b = *(const voff_t *)key; 176 177 if (a < b) 178 return -1; 179 if (a > b) 180 return 1; 181 return 0; 182 } 183 184 const rb_tree_ops_t uvm_page_tree_ops = { 185 .rbto_compare_nodes = uvm_page_compare_nodes, 186 .rbto_compare_key = uvm_page_compare_key, 187 .rbto_node_offset = offsetof(struct vm_page, rb_node), 188 .rbto_context = NULL 189 }; 190 191 /* 192 * inline functions 193 */ 194 195 /* 196 * uvm_pageinsert: insert a page in the object. 197 * 198 * => caller must lock object 199 * => caller must lock page queues 200 * => call should have already set pg's object and offset pointers 201 * and bumped the version counter 202 */ 203 204 static inline void 205 uvm_pageinsert_list(struct uvm_object *uobj, struct vm_page *pg, 206 struct vm_page *where) 207 { 208 209 KASSERT(uobj == pg->uobject); 210 KASSERT(mutex_owned(uobj->vmobjlock)); 211 KASSERT((pg->flags & PG_TABLED) == 0); 212 KASSERT(where == NULL || (where->flags & PG_TABLED)); 213 KASSERT(where == NULL || (where->uobject == uobj)); 214 215 if (UVM_OBJ_IS_VNODE(uobj)) { 216 if (uobj->uo_npages == 0) { 217 struct vnode *vp = (struct vnode *)uobj; 218 219 vholdl(vp); 220 } 221 if (UVM_OBJ_IS_VTEXT(uobj)) { 222 atomic_inc_uint(&uvmexp.execpages); 223 } else { 224 atomic_inc_uint(&uvmexp.filepages); 225 } 226 } else if (UVM_OBJ_IS_AOBJ(uobj)) { 227 atomic_inc_uint(&uvmexp.anonpages); 228 } 229 230 if (where) 231 TAILQ_INSERT_AFTER(&uobj->memq, where, pg, listq.queue); 232 else 233 TAILQ_INSERT_TAIL(&uobj->memq, pg, listq.queue); 234 pg->flags |= PG_TABLED; 235 uobj->uo_npages++; 236 } 237 238 239 static inline void 240 uvm_pageinsert_tree(struct uvm_object *uobj, struct vm_page *pg) 241 { 242 struct vm_page *ret __diagused; 243 244 KASSERT(uobj == pg->uobject); 245 ret = rb_tree_insert_node(&uobj->rb_tree, pg); 246 KASSERT(ret == pg); 247 } 248 249 static inline void 250 uvm_pageinsert(struct uvm_object *uobj, struct vm_page *pg) 251 { 252 253 KDASSERT(uobj != NULL); 254 uvm_pageinsert_tree(uobj, pg); 255 uvm_pageinsert_list(uobj, pg, NULL); 256 } 257 258 /* 259 * uvm_page_remove: remove page from object. 260 * 261 * => caller must lock object 262 * => caller must lock page queues 263 */ 264 265 static inline void 266 uvm_pageremove_list(struct uvm_object *uobj, struct vm_page *pg) 267 { 268 269 KASSERT(uobj == pg->uobject); 270 KASSERT(mutex_owned(uobj->vmobjlock)); 271 KASSERT(pg->flags & PG_TABLED); 272 273 if (UVM_OBJ_IS_VNODE(uobj)) { 274 if (uobj->uo_npages == 1) { 275 struct vnode *vp = (struct vnode *)uobj; 276 277 holdrelel(vp); 278 } 279 if (UVM_OBJ_IS_VTEXT(uobj)) { 280 atomic_dec_uint(&uvmexp.execpages); 281 } else { 282 atomic_dec_uint(&uvmexp.filepages); 283 } 284 } else if (UVM_OBJ_IS_AOBJ(uobj)) { 285 atomic_dec_uint(&uvmexp.anonpages); 286 } 287 288 /* object should be locked */ 289 uobj->uo_npages--; 290 TAILQ_REMOVE(&uobj->memq, pg, listq.queue); 291 pg->flags &= ~PG_TABLED; 292 pg->uobject = NULL; 293 } 294 295 static inline void 296 uvm_pageremove_tree(struct uvm_object *uobj, struct vm_page *pg) 297 { 298 299 KASSERT(uobj == pg->uobject); 300 rb_tree_remove_node(&uobj->rb_tree, pg); 301 } 302 303 static inline void 304 uvm_pageremove(struct uvm_object *uobj, struct vm_page *pg) 305 { 306 307 KDASSERT(uobj != NULL); 308 uvm_pageremove_tree(uobj, pg); 309 uvm_pageremove_list(uobj, pg); 310 } 311 312 static void 313 uvm_page_init_buckets(struct pgfreelist *pgfl) 314 { 315 int color, i; 316 317 for (color = 0; color < uvmexp.ncolors; color++) { 318 for (i = 0; i < PGFL_NQUEUES; i++) { 319 LIST_INIT(&pgfl->pgfl_buckets[color].pgfl_queues[i]); 320 } 321 } 322 } 323 324 /* 325 * uvm_page_init: init the page system. called from uvm_init(). 326 * 327 * => we return the range of kernel virtual memory in kvm_startp/kvm_endp 328 */ 329 330 void 331 uvm_page_init(vaddr_t *kvm_startp, vaddr_t *kvm_endp) 332 { 333 static struct uvm_cpu boot_cpu; 334 psize_t freepages, pagecount, bucketcount, n; 335 struct pgflbucket *bucketarray, *cpuarray; 336 struct vm_physseg *seg; 337 struct vm_page *pagearray; 338 int lcv; 339 u_int i; 340 paddr_t paddr; 341 342 KASSERT(ncpu <= 1); 343 CTASSERT(sizeof(pagearray->offset) >= sizeof(struct uvm_cpu *)); 344 345 /* 346 * init the page queues and page queue locks, except the free 347 * list; we allocate that later (with the initial vm_page 348 * structures). 349 */ 350 351 uvm.cpus[0] = &boot_cpu; 352 curcpu()->ci_data.cpu_uvm = &boot_cpu; 353 uvmpdpol_init(); 354 mutex_init(&uvm_pageqlock, MUTEX_DRIVER, IPL_NONE); 355 mutex_init(&uvm_fpageqlock, MUTEX_DRIVER, IPL_VM); 356 357 /* 358 * allocate vm_page structures. 359 */ 360 361 /* 362 * sanity check: 363 * before calling this function the MD code is expected to register 364 * some free RAM with the uvm_page_physload() function. our job 365 * now is to allocate vm_page structures for this memory. 366 */ 367 368 if (vm_nphysmem == 0) 369 panic("uvm_page_bootstrap: no memory pre-allocated"); 370 371 /* 372 * first calculate the number of free pages... 373 * 374 * note that we use start/end rather than avail_start/avail_end. 375 * this allows us to allocate extra vm_page structures in case we 376 * want to return some memory to the pool after booting. 377 */ 378 379 freepages = 0; 380 for (lcv = 0 ; lcv < vm_nphysmem ; lcv++) { 381 seg = VM_PHYSMEM_PTR(lcv); 382 freepages += (seg->end - seg->start); 383 } 384 385 /* 386 * Let MD code initialize the number of colors, or default 387 * to 1 color if MD code doesn't care. 388 */ 389 if (uvmexp.ncolors == 0) 390 uvmexp.ncolors = 1; 391 uvmexp.colormask = uvmexp.ncolors - 1; 392 KASSERT((uvmexp.colormask & uvmexp.ncolors) == 0); 393 394 /* 395 * we now know we have (PAGE_SIZE * freepages) bytes of memory we can 396 * use. for each page of memory we use we need a vm_page structure. 397 * thus, the total number of pages we can use is the total size of 398 * the memory divided by the PAGE_SIZE plus the size of the vm_page 399 * structure. we add one to freepages as a fudge factor to avoid 400 * truncation errors (since we can only allocate in terms of whole 401 * pages). 402 */ 403 404 bucketcount = uvmexp.ncolors * VM_NFREELIST; 405 pagecount = ((freepages + 1) << PAGE_SHIFT) / 406 (PAGE_SIZE + sizeof(struct vm_page)); 407 408 bucketarray = (void *)uvm_pageboot_alloc((bucketcount * 409 sizeof(struct pgflbucket) * 2) + (pagecount * 410 sizeof(struct vm_page))); 411 cpuarray = bucketarray + bucketcount; 412 pagearray = (struct vm_page *)(bucketarray + bucketcount * 2); 413 414 for (lcv = 0; lcv < VM_NFREELIST; lcv++) { 415 uvm.page_free[lcv].pgfl_buckets = 416 (bucketarray + (lcv * uvmexp.ncolors)); 417 uvm_page_init_buckets(&uvm.page_free[lcv]); 418 uvm.cpus[0]->page_free[lcv].pgfl_buckets = 419 (cpuarray + (lcv * uvmexp.ncolors)); 420 uvm_page_init_buckets(&uvm.cpus[0]->page_free[lcv]); 421 } 422 memset(pagearray, 0, pagecount * sizeof(struct vm_page)); 423 424 /* 425 * init the vm_page structures and put them in the correct place. 426 */ 427 428 for (lcv = 0 ; lcv < vm_nphysmem ; lcv++) { 429 seg = VM_PHYSMEM_PTR(lcv); 430 n = seg->end - seg->start; 431 432 /* set up page array pointers */ 433 seg->pgs = pagearray; 434 pagearray += n; 435 pagecount -= n; 436 seg->lastpg = seg->pgs + n; 437 438 /* init and free vm_pages (we've already zeroed them) */ 439 paddr = ctob(seg->start); 440 for (i = 0 ; i < n ; i++, paddr += PAGE_SIZE) { 441 seg->pgs[i].phys_addr = paddr; 442 #ifdef __HAVE_VM_PAGE_MD 443 VM_MDPAGE_INIT(&seg->pgs[i]); 444 #endif 445 if (atop(paddr) >= seg->avail_start && 446 atop(paddr) < seg->avail_end) { 447 uvmexp.npages++; 448 /* add page to free pool */ 449 uvm_pagefree(&seg->pgs[i]); 450 } 451 } 452 } 453 454 /* 455 * pass up the values of virtual_space_start and 456 * virtual_space_end (obtained by uvm_pageboot_alloc) to the upper 457 * layers of the VM. 458 */ 459 460 *kvm_startp = round_page(virtual_space_start); 461 *kvm_endp = trunc_page(virtual_space_end); 462 #ifdef DEBUG 463 /* 464 * steal kva for uvm_pagezerocheck(). 465 */ 466 uvm_zerocheckkva = *kvm_startp; 467 *kvm_startp += PAGE_SIZE; 468 #endif /* DEBUG */ 469 470 /* 471 * init various thresholds. 472 */ 473 474 uvmexp.reserve_pagedaemon = 1; 475 uvmexp.reserve_kernel = vm_page_reserve_kernel; 476 477 /* 478 * determine if we should zero pages in the idle loop. 479 */ 480 481 uvm.cpus[0]->page_idle_zero = vm_page_zero_enable; 482 483 /* 484 * done! 485 */ 486 487 uvm.page_init_done = true; 488 } 489 490 /* 491 * uvm_setpagesize: set the page size 492 * 493 * => sets page_shift and page_mask from uvmexp.pagesize. 494 */ 495 496 void 497 uvm_setpagesize(void) 498 { 499 500 /* 501 * If uvmexp.pagesize is 0 at this point, we expect PAGE_SIZE 502 * to be a constant (indicated by being a non-zero value). 503 */ 504 if (uvmexp.pagesize == 0) { 505 if (PAGE_SIZE == 0) 506 panic("uvm_setpagesize: uvmexp.pagesize not set"); 507 uvmexp.pagesize = PAGE_SIZE; 508 } 509 uvmexp.pagemask = uvmexp.pagesize - 1; 510 if ((uvmexp.pagemask & uvmexp.pagesize) != 0) 511 panic("uvm_setpagesize: page size %u (%#x) not a power of two", 512 uvmexp.pagesize, uvmexp.pagesize); 513 for (uvmexp.pageshift = 0; ; uvmexp.pageshift++) 514 if ((1 << uvmexp.pageshift) == uvmexp.pagesize) 515 break; 516 } 517 518 /* 519 * uvm_pageboot_alloc: steal memory from physmem for bootstrapping 520 */ 521 522 vaddr_t 523 uvm_pageboot_alloc(vsize_t size) 524 { 525 static bool initialized = false; 526 vaddr_t addr; 527 #if !defined(PMAP_STEAL_MEMORY) 528 vaddr_t vaddr; 529 paddr_t paddr; 530 #endif 531 532 /* 533 * on first call to this function, initialize ourselves. 534 */ 535 if (initialized == false) { 536 pmap_virtual_space(&virtual_space_start, &virtual_space_end); 537 538 /* round it the way we like it */ 539 virtual_space_start = round_page(virtual_space_start); 540 virtual_space_end = trunc_page(virtual_space_end); 541 542 initialized = true; 543 } 544 545 /* round to page size */ 546 size = round_page(size); 547 548 #if defined(PMAP_STEAL_MEMORY) 549 550 /* 551 * defer bootstrap allocation to MD code (it may want to allocate 552 * from a direct-mapped segment). pmap_steal_memory should adjust 553 * virtual_space_start/virtual_space_end if necessary. 554 */ 555 556 addr = pmap_steal_memory(size, &virtual_space_start, 557 &virtual_space_end); 558 559 return(addr); 560 561 #else /* !PMAP_STEAL_MEMORY */ 562 563 /* 564 * allocate virtual memory for this request 565 */ 566 if (virtual_space_start == virtual_space_end || 567 (virtual_space_end - virtual_space_start) < size) 568 panic("uvm_pageboot_alloc: out of virtual space"); 569 570 addr = virtual_space_start; 571 572 #ifdef PMAP_GROWKERNEL 573 /* 574 * If the kernel pmap can't map the requested space, 575 * then allocate more resources for it. 576 */ 577 if (uvm_maxkaddr < (addr + size)) { 578 uvm_maxkaddr = pmap_growkernel(addr + size); 579 if (uvm_maxkaddr < (addr + size)) 580 panic("uvm_pageboot_alloc: pmap_growkernel() failed"); 581 } 582 #endif 583 584 virtual_space_start += size; 585 586 /* 587 * allocate and mapin physical pages to back new virtual pages 588 */ 589 590 for (vaddr = round_page(addr) ; vaddr < addr + size ; 591 vaddr += PAGE_SIZE) { 592 593 if (!uvm_page_physget(&paddr)) 594 panic("uvm_pageboot_alloc: out of memory"); 595 596 /* 597 * Note this memory is no longer managed, so using 598 * pmap_kenter is safe. 599 */ 600 pmap_kenter_pa(vaddr, paddr, VM_PROT_READ|VM_PROT_WRITE, 0); 601 } 602 pmap_update(pmap_kernel()); 603 return(addr); 604 #endif /* PMAP_STEAL_MEMORY */ 605 } 606 607 #if !defined(PMAP_STEAL_MEMORY) 608 /* 609 * uvm_page_physget: "steal" one page from the vm_physmem structure. 610 * 611 * => attempt to allocate it off the end of a segment in which the "avail" 612 * values match the start/end values. if we can't do that, then we 613 * will advance both values (making them equal, and removing some 614 * vm_page structures from the non-avail area). 615 * => return false if out of memory. 616 */ 617 618 /* subroutine: try to allocate from memory chunks on the specified freelist */ 619 static bool uvm_page_physget_freelist(paddr_t *, int); 620 621 static bool 622 uvm_page_physget_freelist(paddr_t *paddrp, int freelist) 623 { 624 struct vm_physseg *seg; 625 int lcv, x; 626 627 /* pass 1: try allocating from a matching end */ 628 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST) 629 for (lcv = vm_nphysmem - 1 ; lcv >= 0 ; lcv--) 630 #else 631 for (lcv = 0 ; lcv < vm_nphysmem ; lcv++) 632 #endif 633 { 634 seg = VM_PHYSMEM_PTR(lcv); 635 636 if (uvm.page_init_done == true) 637 panic("uvm_page_physget: called _after_ bootstrap"); 638 639 if (seg->free_list != freelist) 640 continue; 641 642 /* try from front */ 643 if (seg->avail_start == seg->start && 644 seg->avail_start < seg->avail_end) { 645 *paddrp = ctob(seg->avail_start); 646 seg->avail_start++; 647 seg->start++; 648 /* nothing left? nuke it */ 649 if (seg->avail_start == seg->end) { 650 if (vm_nphysmem == 1) 651 panic("uvm_page_physget: out of memory!"); 652 vm_nphysmem--; 653 for (x = lcv ; x < vm_nphysmem ; x++) 654 /* structure copy */ 655 VM_PHYSMEM_PTR_SWAP(x, x + 1); 656 } 657 return (true); 658 } 659 660 /* try from rear */ 661 if (seg->avail_end == seg->end && 662 seg->avail_start < seg->avail_end) { 663 *paddrp = ctob(seg->avail_end - 1); 664 seg->avail_end--; 665 seg->end--; 666 /* nothing left? nuke it */ 667 if (seg->avail_end == seg->start) { 668 if (vm_nphysmem == 1) 669 panic("uvm_page_physget: out of memory!"); 670 vm_nphysmem--; 671 for (x = lcv ; x < vm_nphysmem ; x++) 672 /* structure copy */ 673 VM_PHYSMEM_PTR_SWAP(x, x + 1); 674 } 675 return (true); 676 } 677 } 678 679 /* pass2: forget about matching ends, just allocate something */ 680 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST) 681 for (lcv = vm_nphysmem - 1 ; lcv >= 0 ; lcv--) 682 #else 683 for (lcv = 0 ; lcv < vm_nphysmem ; lcv++) 684 #endif 685 { 686 seg = VM_PHYSMEM_PTR(lcv); 687 688 /* any room in this bank? */ 689 if (seg->avail_start >= seg->avail_end) 690 continue; /* nope */ 691 692 *paddrp = ctob(seg->avail_start); 693 seg->avail_start++; 694 /* truncate! */ 695 seg->start = seg->avail_start; 696 697 /* nothing left? nuke it */ 698 if (seg->avail_start == seg->end) { 699 if (vm_nphysmem == 1) 700 panic("uvm_page_physget: out of memory!"); 701 vm_nphysmem--; 702 for (x = lcv ; x < vm_nphysmem ; x++) 703 /* structure copy */ 704 VM_PHYSMEM_PTR_SWAP(x, x + 1); 705 } 706 return (true); 707 } 708 709 return (false); /* whoops! */ 710 } 711 712 bool 713 uvm_page_physget(paddr_t *paddrp) 714 { 715 int i; 716 717 /* try in the order of freelist preference */ 718 for (i = 0; i < VM_NFREELIST; i++) 719 if (uvm_page_physget_freelist(paddrp, i) == true) 720 return (true); 721 return (false); 722 } 723 #endif /* PMAP_STEAL_MEMORY */ 724 725 /* 726 * uvm_page_physload: load physical memory into VM system 727 * 728 * => all args are PFs 729 * => all pages in start/end get vm_page structures 730 * => areas marked by avail_start/avail_end get added to the free page pool 731 * => we are limited to VM_PHYSSEG_MAX physical memory segments 732 */ 733 734 void 735 uvm_page_physload(paddr_t start, paddr_t end, paddr_t avail_start, 736 paddr_t avail_end, int free_list) 737 { 738 int preload, lcv; 739 psize_t npages; 740 struct vm_page *pgs; 741 struct vm_physseg *ps; 742 743 if (uvmexp.pagesize == 0) 744 panic("uvm_page_physload: page size not set!"); 745 if (free_list >= VM_NFREELIST || free_list < VM_FREELIST_DEFAULT) 746 panic("uvm_page_physload: bad free list %d", free_list); 747 if (start >= end) 748 panic("uvm_page_physload: start >= end"); 749 750 /* 751 * do we have room? 752 */ 753 754 if (vm_nphysmem == VM_PHYSSEG_MAX) { 755 printf("uvm_page_physload: unable to load physical memory " 756 "segment\n"); 757 printf("\t%d segments allocated, ignoring 0x%llx -> 0x%llx\n", 758 VM_PHYSSEG_MAX, (long long)start, (long long)end); 759 printf("\tincrease VM_PHYSSEG_MAX\n"); 760 return; 761 } 762 763 /* 764 * check to see if this is a "preload" (i.e. uvm_page_init hasn't been 765 * called yet, so kmem is not available). 766 */ 767 768 for (lcv = 0 ; lcv < vm_nphysmem ; lcv++) { 769 if (VM_PHYSMEM_PTR(lcv)->pgs) 770 break; 771 } 772 preload = (lcv == vm_nphysmem); 773 774 /* 775 * if VM is already running, attempt to kmem_alloc vm_page structures 776 */ 777 778 if (!preload) { 779 panic("uvm_page_physload: tried to add RAM after vm_mem_init"); 780 } else { 781 pgs = NULL; 782 npages = 0; 783 } 784 785 /* 786 * now insert us in the proper place in vm_physmem[] 787 */ 788 789 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_RANDOM) 790 /* random: put it at the end (easy!) */ 791 ps = VM_PHYSMEM_PTR(vm_nphysmem); 792 #elif (VM_PHYSSEG_STRAT == VM_PSTRAT_BSEARCH) 793 { 794 int x; 795 /* sort by address for binary search */ 796 for (lcv = 0 ; lcv < vm_nphysmem ; lcv++) 797 if (start < VM_PHYSMEM_PTR(lcv)->start) 798 break; 799 ps = VM_PHYSMEM_PTR(lcv); 800 /* move back other entries, if necessary ... */ 801 for (x = vm_nphysmem ; x > lcv ; x--) 802 /* structure copy */ 803 VM_PHYSMEM_PTR_SWAP(x, x - 1); 804 } 805 #elif (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST) 806 { 807 int x; 808 /* sort by largest segment first */ 809 for (lcv = 0 ; lcv < vm_nphysmem ; lcv++) 810 if ((end - start) > 811 (VM_PHYSMEM_PTR(lcv)->end - VM_PHYSMEM_PTR(lcv)->start)) 812 break; 813 ps = VM_PHYSMEM_PTR(lcv); 814 /* move back other entries, if necessary ... */ 815 for (x = vm_nphysmem ; x > lcv ; x--) 816 /* structure copy */ 817 VM_PHYSMEM_PTR_SWAP(x, x - 1); 818 } 819 #else 820 panic("uvm_page_physload: unknown physseg strategy selected!"); 821 #endif 822 823 ps->start = start; 824 ps->end = end; 825 ps->avail_start = avail_start; 826 ps->avail_end = avail_end; 827 if (preload) { 828 ps->pgs = NULL; 829 } else { 830 ps->pgs = pgs; 831 ps->lastpg = pgs + npages; 832 } 833 ps->free_list = free_list; 834 vm_nphysmem++; 835 836 if (!preload) { 837 uvmpdpol_reinit(); 838 } 839 } 840 841 /* 842 * when VM_PHYSSEG_MAX is 1, we can simplify these functions 843 */ 844 845 #if VM_PHYSSEG_MAX == 1 846 static inline int vm_physseg_find_contig(struct vm_physseg *, int, paddr_t, int *); 847 #elif (VM_PHYSSEG_STRAT == VM_PSTRAT_BSEARCH) 848 static inline int vm_physseg_find_bsearch(struct vm_physseg *, int, paddr_t, int *); 849 #else 850 static inline int vm_physseg_find_linear(struct vm_physseg *, int, paddr_t, int *); 851 #endif 852 853 /* 854 * vm_physseg_find: find vm_physseg structure that belongs to a PA 855 */ 856 int 857 vm_physseg_find(paddr_t pframe, int *offp) 858 { 859 860 #if VM_PHYSSEG_MAX == 1 861 return vm_physseg_find_contig(vm_physmem, vm_nphysseg, pframe, offp); 862 #elif (VM_PHYSSEG_STRAT == VM_PSTRAT_BSEARCH) 863 return vm_physseg_find_bsearch(vm_physmem, vm_nphysseg, pframe, offp); 864 #else 865 return vm_physseg_find_linear(vm_physmem, vm_nphysseg, pframe, offp); 866 #endif 867 } 868 869 #if VM_PHYSSEG_MAX == 1 870 static inline int 871 vm_physseg_find_contig(struct vm_physseg *segs, int nsegs, paddr_t pframe, int *offp) 872 { 873 874 /* 'contig' case */ 875 if (pframe >= segs[0].start && pframe < segs[0].end) { 876 if (offp) 877 *offp = pframe - segs[0].start; 878 return(0); 879 } 880 return(-1); 881 } 882 883 #elif (VM_PHYSSEG_STRAT == VM_PSTRAT_BSEARCH) 884 885 static inline int 886 vm_physseg_find_bsearch(struct vm_physseg *segs, int nsegs, paddr_t pframe, int *offp) 887 { 888 /* binary search for it */ 889 u_int start, len, guess; 890 891 /* 892 * if try is too large (thus target is less than try) we reduce 893 * the length to trunc(len/2) [i.e. everything smaller than "try"] 894 * 895 * if the try is too small (thus target is greater than try) then 896 * we set the new start to be (try + 1). this means we need to 897 * reduce the length to (round(len/2) - 1). 898 * 899 * note "adjust" below which takes advantage of the fact that 900 * (round(len/2) - 1) == trunc((len - 1) / 2) 901 * for any value of len we may have 902 */ 903 904 for (start = 0, len = nsegs ; len != 0 ; len = len / 2) { 905 guess = start + (len / 2); /* try in the middle */ 906 907 /* start past our try? */ 908 if (pframe >= segs[guess].start) { 909 /* was try correct? */ 910 if (pframe < segs[guess].end) { 911 if (offp) 912 *offp = pframe - segs[guess].start; 913 return guess; /* got it */ 914 } 915 start = guess + 1; /* next time, start here */ 916 len--; /* "adjust" */ 917 } else { 918 /* 919 * pframe before try, just reduce length of 920 * region, done in "for" loop 921 */ 922 } 923 } 924 return(-1); 925 } 926 927 #else 928 929 static inline int 930 vm_physseg_find_linear(struct vm_physseg *segs, int nsegs, paddr_t pframe, int *offp) 931 { 932 /* linear search for it */ 933 int lcv; 934 935 for (lcv = 0; lcv < nsegs; lcv++) { 936 if (pframe >= segs[lcv].start && 937 pframe < segs[lcv].end) { 938 if (offp) 939 *offp = pframe - segs[lcv].start; 940 return(lcv); /* got it */ 941 } 942 } 943 return(-1); 944 } 945 #endif 946 947 /* 948 * PHYS_TO_VM_PAGE: find vm_page for a PA. used by MI code to get vm_pages 949 * back from an I/O mapping (ugh!). used in some MD code as well. 950 */ 951 struct vm_page * 952 uvm_phys_to_vm_page(paddr_t pa) 953 { 954 paddr_t pf = atop(pa); 955 int off; 956 int psi; 957 958 psi = vm_physseg_find(pf, &off); 959 if (psi != -1) 960 return(&VM_PHYSMEM_PTR(psi)->pgs[off]); 961 return(NULL); 962 } 963 964 paddr_t 965 uvm_vm_page_to_phys(const struct vm_page *pg) 966 { 967 968 return pg->phys_addr; 969 } 970 971 /* 972 * uvm_page_recolor: Recolor the pages if the new bucket count is 973 * larger than the old one. 974 */ 975 976 void 977 uvm_page_recolor(int newncolors) 978 { 979 struct pgflbucket *bucketarray, *cpuarray, *oldbucketarray; 980 struct pgfreelist gpgfl, pgfl; 981 struct vm_page *pg; 982 vsize_t bucketcount; 983 size_t bucketmemsize, oldbucketmemsize; 984 int lcv, color, i, ocolors; 985 struct uvm_cpu *ucpu; 986 987 KASSERT(((newncolors - 1) & newncolors) == 0); 988 989 if (newncolors <= uvmexp.ncolors) 990 return; 991 992 if (uvm.page_init_done == false) { 993 uvmexp.ncolors = newncolors; 994 return; 995 } 996 997 bucketcount = newncolors * VM_NFREELIST; 998 bucketmemsize = bucketcount * sizeof(struct pgflbucket) * 2; 999 bucketarray = kmem_alloc(bucketmemsize, KM_SLEEP); 1000 cpuarray = bucketarray + bucketcount; 1001 if (bucketarray == NULL) { 1002 printf("WARNING: unable to allocate %ld page color buckets\n", 1003 (long) bucketcount); 1004 return; 1005 } 1006 1007 mutex_spin_enter(&uvm_fpageqlock); 1008 1009 /* Make sure we should still do this. */ 1010 if (newncolors <= uvmexp.ncolors) { 1011 mutex_spin_exit(&uvm_fpageqlock); 1012 kmem_free(bucketarray, bucketmemsize); 1013 return; 1014 } 1015 1016 oldbucketarray = uvm.page_free[0].pgfl_buckets; 1017 ocolors = uvmexp.ncolors; 1018 1019 uvmexp.ncolors = newncolors; 1020 uvmexp.colormask = uvmexp.ncolors - 1; 1021 1022 ucpu = curcpu()->ci_data.cpu_uvm; 1023 for (lcv = 0; lcv < VM_NFREELIST; lcv++) { 1024 gpgfl.pgfl_buckets = (bucketarray + (lcv * newncolors)); 1025 pgfl.pgfl_buckets = (cpuarray + (lcv * uvmexp.ncolors)); 1026 uvm_page_init_buckets(&gpgfl); 1027 uvm_page_init_buckets(&pgfl); 1028 for (color = 0; color < ocolors; color++) { 1029 for (i = 0; i < PGFL_NQUEUES; i++) { 1030 while ((pg = LIST_FIRST(&uvm.page_free[ 1031 lcv].pgfl_buckets[color].pgfl_queues[i])) 1032 != NULL) { 1033 LIST_REMOVE(pg, pageq.list); /* global */ 1034 LIST_REMOVE(pg, listq.list); /* cpu */ 1035 LIST_INSERT_HEAD(&gpgfl.pgfl_buckets[ 1036 VM_PGCOLOR_BUCKET(pg)].pgfl_queues[ 1037 i], pg, pageq.list); 1038 LIST_INSERT_HEAD(&pgfl.pgfl_buckets[ 1039 VM_PGCOLOR_BUCKET(pg)].pgfl_queues[ 1040 i], pg, listq.list); 1041 } 1042 } 1043 } 1044 uvm.page_free[lcv].pgfl_buckets = gpgfl.pgfl_buckets; 1045 ucpu->page_free[lcv].pgfl_buckets = pgfl.pgfl_buckets; 1046 } 1047 1048 oldbucketmemsize = recolored_pages_memsize; 1049 1050 recolored_pages_memsize = bucketmemsize; 1051 mutex_spin_exit(&uvm_fpageqlock); 1052 1053 if (oldbucketmemsize) { 1054 kmem_free(oldbucketarray, recolored_pages_memsize); 1055 } 1056 1057 /* 1058 * this calls uvm_km_alloc() which may want to hold 1059 * uvm_fpageqlock. 1060 */ 1061 uvm_pager_realloc_emerg(); 1062 } 1063 1064 /* 1065 * uvm_cpu_attach: initialize per-CPU data structures. 1066 */ 1067 1068 void 1069 uvm_cpu_attach(struct cpu_info *ci) 1070 { 1071 struct pgflbucket *bucketarray; 1072 struct pgfreelist pgfl; 1073 struct uvm_cpu *ucpu; 1074 vsize_t bucketcount; 1075 int lcv; 1076 1077 if (CPU_IS_PRIMARY(ci)) { 1078 /* Already done in uvm_page_init(). */ 1079 goto attachrnd; 1080 } 1081 1082 /* Add more reserve pages for this CPU. */ 1083 uvmexp.reserve_kernel += vm_page_reserve_kernel; 1084 1085 /* Configure this CPU's free lists. */ 1086 bucketcount = uvmexp.ncolors * VM_NFREELIST; 1087 bucketarray = kmem_alloc(bucketcount * sizeof(struct pgflbucket), 1088 KM_SLEEP); 1089 ucpu = kmem_zalloc(sizeof(*ucpu), KM_SLEEP); 1090 uvm.cpus[cpu_index(ci)] = ucpu; 1091 ci->ci_data.cpu_uvm = ucpu; 1092 for (lcv = 0; lcv < VM_NFREELIST; lcv++) { 1093 pgfl.pgfl_buckets = (bucketarray + (lcv * uvmexp.ncolors)); 1094 uvm_page_init_buckets(&pgfl); 1095 ucpu->page_free[lcv].pgfl_buckets = pgfl.pgfl_buckets; 1096 } 1097 1098 attachrnd: 1099 /* 1100 * Attach RNG source for this CPU's VM events 1101 */ 1102 rnd_attach_source(&uvm.cpus[cpu_index(ci)]->rs, 1103 ci->ci_data.cpu_name, RND_TYPE_VM, 1104 RND_FLAG_COLLECT_TIME|RND_FLAG_COLLECT_VALUE| 1105 RND_FLAG_ESTIMATE_VALUE); 1106 1107 } 1108 1109 /* 1110 * uvm_pagealloc_pgfl: helper routine for uvm_pagealloc_strat 1111 */ 1112 1113 static struct vm_page * 1114 uvm_pagealloc_pgfl(struct uvm_cpu *ucpu, int flist, int try1, int try2, 1115 int *trycolorp) 1116 { 1117 struct pgflist *freeq; 1118 struct vm_page *pg; 1119 int color, trycolor = *trycolorp; 1120 struct pgfreelist *gpgfl, *pgfl; 1121 1122 KASSERT(mutex_owned(&uvm_fpageqlock)); 1123 1124 color = trycolor; 1125 pgfl = &ucpu->page_free[flist]; 1126 gpgfl = &uvm.page_free[flist]; 1127 do { 1128 /* cpu, try1 */ 1129 if ((pg = LIST_FIRST((freeq = 1130 &pgfl->pgfl_buckets[color].pgfl_queues[try1]))) != NULL) { 1131 KASSERT(pg->pqflags & PQ_FREE); 1132 KASSERT(try1 == PGFL_ZEROS || !(pg->flags & PG_ZERO)); 1133 KASSERT(try1 == PGFL_UNKNOWN || (pg->flags & PG_ZERO)); 1134 KASSERT(ucpu == VM_FREE_PAGE_TO_CPU(pg)); 1135 VM_FREE_PAGE_TO_CPU(pg)->pages[try1]--; 1136 uvmexp.cpuhit++; 1137 goto gotit; 1138 } 1139 /* global, try1 */ 1140 if ((pg = LIST_FIRST((freeq = 1141 &gpgfl->pgfl_buckets[color].pgfl_queues[try1]))) != NULL) { 1142 KASSERT(pg->pqflags & PQ_FREE); 1143 KASSERT(try1 == PGFL_ZEROS || !(pg->flags & PG_ZERO)); 1144 KASSERT(try1 == PGFL_UNKNOWN || (pg->flags & PG_ZERO)); 1145 KASSERT(ucpu != VM_FREE_PAGE_TO_CPU(pg)); 1146 VM_FREE_PAGE_TO_CPU(pg)->pages[try1]--; 1147 uvmexp.cpumiss++; 1148 goto gotit; 1149 } 1150 /* cpu, try2 */ 1151 if ((pg = LIST_FIRST((freeq = 1152 &pgfl->pgfl_buckets[color].pgfl_queues[try2]))) != NULL) { 1153 KASSERT(pg->pqflags & PQ_FREE); 1154 KASSERT(try2 == PGFL_ZEROS || !(pg->flags & PG_ZERO)); 1155 KASSERT(try2 == PGFL_UNKNOWN || (pg->flags & PG_ZERO)); 1156 KASSERT(ucpu == VM_FREE_PAGE_TO_CPU(pg)); 1157 VM_FREE_PAGE_TO_CPU(pg)->pages[try2]--; 1158 uvmexp.cpuhit++; 1159 goto gotit; 1160 } 1161 /* global, try2 */ 1162 if ((pg = LIST_FIRST((freeq = 1163 &gpgfl->pgfl_buckets[color].pgfl_queues[try2]))) != NULL) { 1164 KASSERT(pg->pqflags & PQ_FREE); 1165 KASSERT(try2 == PGFL_ZEROS || !(pg->flags & PG_ZERO)); 1166 KASSERT(try2 == PGFL_UNKNOWN || (pg->flags & PG_ZERO)); 1167 KASSERT(ucpu != VM_FREE_PAGE_TO_CPU(pg)); 1168 VM_FREE_PAGE_TO_CPU(pg)->pages[try2]--; 1169 uvmexp.cpumiss++; 1170 goto gotit; 1171 } 1172 color = (color + 1) & uvmexp.colormask; 1173 } while (color != trycolor); 1174 1175 return (NULL); 1176 1177 gotit: 1178 LIST_REMOVE(pg, pageq.list); /* global list */ 1179 LIST_REMOVE(pg, listq.list); /* per-cpu list */ 1180 uvmexp.free--; 1181 1182 /* update zero'd page count */ 1183 if (pg->flags & PG_ZERO) 1184 uvmexp.zeropages--; 1185 1186 if (color == trycolor) 1187 uvmexp.colorhit++; 1188 else { 1189 uvmexp.colormiss++; 1190 *trycolorp = color; 1191 } 1192 1193 return (pg); 1194 } 1195 1196 /* 1197 * uvm_pagealloc_strat: allocate vm_page from a particular free list. 1198 * 1199 * => return null if no pages free 1200 * => wake up pagedaemon if number of free pages drops below low water mark 1201 * => if obj != NULL, obj must be locked (to put in obj's tree) 1202 * => if anon != NULL, anon must be locked (to put in anon) 1203 * => only one of obj or anon can be non-null 1204 * => caller must activate/deactivate page if it is not wired. 1205 * => free_list is ignored if strat == UVM_PGA_STRAT_NORMAL. 1206 * => policy decision: it is more important to pull a page off of the 1207 * appropriate priority free list than it is to get a zero'd or 1208 * unknown contents page. This is because we live with the 1209 * consequences of a bad free list decision for the entire 1210 * lifetime of the page, e.g. if the page comes from memory that 1211 * is slower to access. 1212 */ 1213 1214 struct vm_page * 1215 uvm_pagealloc_strat(struct uvm_object *obj, voff_t off, struct vm_anon *anon, 1216 int flags, int strat, int free_list) 1217 { 1218 int lcv, try1, try2, zeroit = 0, color; 1219 struct uvm_cpu *ucpu; 1220 struct vm_page *pg; 1221 lwp_t *l; 1222 1223 KASSERT(obj == NULL || anon == NULL); 1224 KASSERT(anon == NULL || (flags & UVM_FLAG_COLORMATCH) || off == 0); 1225 KASSERT(off == trunc_page(off)); 1226 KASSERT(obj == NULL || mutex_owned(obj->vmobjlock)); 1227 KASSERT(anon == NULL || anon->an_lock == NULL || 1228 mutex_owned(anon->an_lock)); 1229 1230 mutex_spin_enter(&uvm_fpageqlock); 1231 1232 /* 1233 * This implements a global round-robin page coloring 1234 * algorithm. 1235 */ 1236 1237 ucpu = curcpu()->ci_data.cpu_uvm; 1238 if (flags & UVM_FLAG_COLORMATCH) { 1239 color = atop(off) & uvmexp.colormask; 1240 } else { 1241 color = ucpu->page_free_nextcolor; 1242 } 1243 1244 /* 1245 * check to see if we need to generate some free pages waking 1246 * the pagedaemon. 1247 */ 1248 1249 uvm_kick_pdaemon(); 1250 1251 /* 1252 * fail if any of these conditions is true: 1253 * [1] there really are no free pages, or 1254 * [2] only kernel "reserved" pages remain and 1255 * reserved pages have not been requested. 1256 * [3] only pagedaemon "reserved" pages remain and 1257 * the requestor isn't the pagedaemon. 1258 * we make kernel reserve pages available if called by a 1259 * kernel thread or a realtime thread. 1260 */ 1261 l = curlwp; 1262 if (__predict_true(l != NULL) && lwp_eprio(l) >= PRI_KTHREAD) { 1263 flags |= UVM_PGA_USERESERVE; 1264 } 1265 if ((uvmexp.free <= uvmexp.reserve_kernel && 1266 (flags & UVM_PGA_USERESERVE) == 0) || 1267 (uvmexp.free <= uvmexp.reserve_pagedaemon && 1268 curlwp != uvm.pagedaemon_lwp)) 1269 goto fail; 1270 1271 #if PGFL_NQUEUES != 2 1272 #error uvm_pagealloc_strat needs to be updated 1273 #endif 1274 1275 /* 1276 * If we want a zero'd page, try the ZEROS queue first, otherwise 1277 * we try the UNKNOWN queue first. 1278 */ 1279 if (flags & UVM_PGA_ZERO) { 1280 try1 = PGFL_ZEROS; 1281 try2 = PGFL_UNKNOWN; 1282 } else { 1283 try1 = PGFL_UNKNOWN; 1284 try2 = PGFL_ZEROS; 1285 } 1286 1287 again: 1288 switch (strat) { 1289 case UVM_PGA_STRAT_NORMAL: 1290 /* Check freelists: descending priority (ascending id) order */ 1291 for (lcv = 0; lcv < VM_NFREELIST; lcv++) { 1292 pg = uvm_pagealloc_pgfl(ucpu, lcv, 1293 try1, try2, &color); 1294 if (pg != NULL) 1295 goto gotit; 1296 } 1297 1298 /* No pages free! */ 1299 goto fail; 1300 1301 case UVM_PGA_STRAT_ONLY: 1302 case UVM_PGA_STRAT_FALLBACK: 1303 /* Attempt to allocate from the specified free list. */ 1304 KASSERT(free_list >= 0 && free_list < VM_NFREELIST); 1305 pg = uvm_pagealloc_pgfl(ucpu, free_list, 1306 try1, try2, &color); 1307 if (pg != NULL) 1308 goto gotit; 1309 1310 /* Fall back, if possible. */ 1311 if (strat == UVM_PGA_STRAT_FALLBACK) { 1312 strat = UVM_PGA_STRAT_NORMAL; 1313 goto again; 1314 } 1315 1316 /* No pages free! */ 1317 goto fail; 1318 1319 default: 1320 panic("uvm_pagealloc_strat: bad strat %d", strat); 1321 /* NOTREACHED */ 1322 } 1323 1324 gotit: 1325 /* 1326 * We now know which color we actually allocated from; set 1327 * the next color accordingly. 1328 */ 1329 1330 ucpu->page_free_nextcolor = (color + 1) & uvmexp.colormask; 1331 1332 /* 1333 * update allocation statistics and remember if we have to 1334 * zero the page 1335 */ 1336 1337 if (flags & UVM_PGA_ZERO) { 1338 if (pg->flags & PG_ZERO) { 1339 uvmexp.pga_zerohit++; 1340 zeroit = 0; 1341 } else { 1342 uvmexp.pga_zeromiss++; 1343 zeroit = 1; 1344 } 1345 if (ucpu->pages[PGFL_ZEROS] < ucpu->pages[PGFL_UNKNOWN]) { 1346 ucpu->page_idle_zero = vm_page_zero_enable; 1347 } 1348 } 1349 KASSERT(pg->pqflags == PQ_FREE); 1350 1351 pg->offset = off; 1352 pg->uobject = obj; 1353 pg->uanon = anon; 1354 pg->flags = PG_BUSY|PG_CLEAN|PG_FAKE; 1355 if (anon) { 1356 anon->an_page = pg; 1357 pg->pqflags = PQ_ANON; 1358 atomic_inc_uint(&uvmexp.anonpages); 1359 } else { 1360 if (obj) { 1361 uvm_pageinsert(obj, pg); 1362 } 1363 pg->pqflags = 0; 1364 } 1365 mutex_spin_exit(&uvm_fpageqlock); 1366 1367 #if defined(UVM_PAGE_TRKOWN) 1368 pg->owner_tag = NULL; 1369 #endif 1370 UVM_PAGE_OWN(pg, "new alloc"); 1371 1372 if (flags & UVM_PGA_ZERO) { 1373 /* 1374 * A zero'd page is not clean. If we got a page not already 1375 * zero'd, then we have to zero it ourselves. 1376 */ 1377 pg->flags &= ~PG_CLEAN; 1378 if (zeroit) 1379 pmap_zero_page(VM_PAGE_TO_PHYS(pg)); 1380 } 1381 1382 return(pg); 1383 1384 fail: 1385 mutex_spin_exit(&uvm_fpageqlock); 1386 return (NULL); 1387 } 1388 1389 /* 1390 * uvm_pagereplace: replace a page with another 1391 * 1392 * => object must be locked 1393 */ 1394 1395 void 1396 uvm_pagereplace(struct vm_page *oldpg, struct vm_page *newpg) 1397 { 1398 struct uvm_object *uobj = oldpg->uobject; 1399 1400 KASSERT((oldpg->flags & PG_TABLED) != 0); 1401 KASSERT(uobj != NULL); 1402 KASSERT((newpg->flags & PG_TABLED) == 0); 1403 KASSERT(newpg->uobject == NULL); 1404 KASSERT(mutex_owned(uobj->vmobjlock)); 1405 1406 newpg->uobject = uobj; 1407 newpg->offset = oldpg->offset; 1408 1409 uvm_pageremove_tree(uobj, oldpg); 1410 uvm_pageinsert_tree(uobj, newpg); 1411 uvm_pageinsert_list(uobj, newpg, oldpg); 1412 uvm_pageremove_list(uobj, oldpg); 1413 } 1414 1415 /* 1416 * uvm_pagerealloc: reallocate a page from one object to another 1417 * 1418 * => both objects must be locked 1419 */ 1420 1421 void 1422 uvm_pagerealloc(struct vm_page *pg, struct uvm_object *newobj, voff_t newoff) 1423 { 1424 /* 1425 * remove it from the old object 1426 */ 1427 1428 if (pg->uobject) { 1429 uvm_pageremove(pg->uobject, pg); 1430 } 1431 1432 /* 1433 * put it in the new object 1434 */ 1435 1436 if (newobj) { 1437 pg->uobject = newobj; 1438 pg->offset = newoff; 1439 uvm_pageinsert(newobj, pg); 1440 } 1441 } 1442 1443 #ifdef DEBUG 1444 /* 1445 * check if page is zero-filled 1446 * 1447 * - called with free page queue lock held. 1448 */ 1449 void 1450 uvm_pagezerocheck(struct vm_page *pg) 1451 { 1452 int *p, *ep; 1453 1454 KASSERT(uvm_zerocheckkva != 0); 1455 KASSERT(mutex_owned(&uvm_fpageqlock)); 1456 1457 /* 1458 * XXX assuming pmap_kenter_pa and pmap_kremove never call 1459 * uvm page allocator. 1460 * 1461 * it might be better to have "CPU-local temporary map" pmap interface. 1462 */ 1463 pmap_kenter_pa(uvm_zerocheckkva, VM_PAGE_TO_PHYS(pg), VM_PROT_READ, 0); 1464 p = (int *)uvm_zerocheckkva; 1465 ep = (int *)((char *)p + PAGE_SIZE); 1466 pmap_update(pmap_kernel()); 1467 while (p < ep) { 1468 if (*p != 0) 1469 panic("PG_ZERO page isn't zero-filled"); 1470 p++; 1471 } 1472 pmap_kremove(uvm_zerocheckkva, PAGE_SIZE); 1473 /* 1474 * pmap_update() is not necessary here because no one except us 1475 * uses this VA. 1476 */ 1477 } 1478 #endif /* DEBUG */ 1479 1480 /* 1481 * uvm_pagefree: free page 1482 * 1483 * => erase page's identity (i.e. remove from object) 1484 * => put page on free list 1485 * => caller must lock owning object (either anon or uvm_object) 1486 * => caller must lock page queues 1487 * => assumes all valid mappings of pg are gone 1488 */ 1489 1490 void 1491 uvm_pagefree(struct vm_page *pg) 1492 { 1493 struct pgflist *pgfl; 1494 struct uvm_cpu *ucpu; 1495 int index, color, queue; 1496 bool iszero; 1497 1498 #ifdef DEBUG 1499 if (pg->uobject == (void *)0xdeadbeef && 1500 pg->uanon == (void *)0xdeadbeef) { 1501 panic("uvm_pagefree: freeing free page %p", pg); 1502 } 1503 #endif /* DEBUG */ 1504 1505 KASSERT((pg->flags & PG_PAGEOUT) == 0); 1506 KASSERT(!(pg->pqflags & PQ_FREE)); 1507 //KASSERT(mutex_owned(&uvm_pageqlock) || !uvmpdpol_pageisqueued_p(pg)); 1508 KASSERT(pg->uobject == NULL || mutex_owned(pg->uobject->vmobjlock)); 1509 KASSERT(pg->uobject != NULL || pg->uanon == NULL || 1510 mutex_owned(pg->uanon->an_lock)); 1511 1512 /* 1513 * if the page is loaned, resolve the loan instead of freeing. 1514 */ 1515 1516 if (pg->loan_count) { 1517 KASSERT(pg->wire_count == 0); 1518 1519 /* 1520 * if the page is owned by an anon then we just want to 1521 * drop anon ownership. the kernel will free the page when 1522 * it is done with it. if the page is owned by an object, 1523 * remove it from the object and mark it dirty for the benefit 1524 * of possible anon owners. 1525 * 1526 * regardless of previous ownership, wakeup any waiters, 1527 * unbusy the page, and we're done. 1528 */ 1529 1530 if (pg->uobject != NULL) { 1531 uvm_pageremove(pg->uobject, pg); 1532 pg->flags &= ~PG_CLEAN; 1533 } else if (pg->uanon != NULL) { 1534 if ((pg->pqflags & PQ_ANON) == 0) { 1535 pg->loan_count--; 1536 } else { 1537 pg->pqflags &= ~PQ_ANON; 1538 atomic_dec_uint(&uvmexp.anonpages); 1539 } 1540 pg->uanon->an_page = NULL; 1541 pg->uanon = NULL; 1542 } 1543 if (pg->flags & PG_WANTED) { 1544 wakeup(pg); 1545 } 1546 pg->flags &= ~(PG_WANTED|PG_BUSY|PG_RELEASED|PG_PAGER1); 1547 #ifdef UVM_PAGE_TRKOWN 1548 pg->owner_tag = NULL; 1549 #endif 1550 if (pg->loan_count) { 1551 KASSERT(pg->uobject == NULL); 1552 if (pg->uanon == NULL) { 1553 KASSERT(mutex_owned(&uvm_pageqlock)); 1554 uvm_pagedequeue(pg); 1555 } 1556 return; 1557 } 1558 } 1559 1560 /* 1561 * remove page from its object or anon. 1562 */ 1563 1564 if (pg->uobject != NULL) { 1565 uvm_pageremove(pg->uobject, pg); 1566 } else if (pg->uanon != NULL) { 1567 pg->uanon->an_page = NULL; 1568 atomic_dec_uint(&uvmexp.anonpages); 1569 } 1570 1571 /* 1572 * now remove the page from the queues. 1573 */ 1574 if (uvmpdpol_pageisqueued_p(pg)) { 1575 KASSERT(mutex_owned(&uvm_pageqlock)); 1576 uvm_pagedequeue(pg); 1577 } 1578 1579 /* 1580 * if the page was wired, unwire it now. 1581 */ 1582 1583 if (pg->wire_count) { 1584 pg->wire_count = 0; 1585 uvmexp.wired--; 1586 } 1587 1588 /* 1589 * and put on free queue 1590 */ 1591 1592 iszero = (pg->flags & PG_ZERO); 1593 index = uvm_page_lookup_freelist(pg); 1594 color = VM_PGCOLOR_BUCKET(pg); 1595 queue = (iszero ? PGFL_ZEROS : PGFL_UNKNOWN); 1596 1597 #ifdef DEBUG 1598 pg->uobject = (void *)0xdeadbeef; 1599 pg->uanon = (void *)0xdeadbeef; 1600 #endif 1601 1602 mutex_spin_enter(&uvm_fpageqlock); 1603 pg->pqflags = PQ_FREE; 1604 1605 #ifdef DEBUG 1606 if (iszero) 1607 uvm_pagezerocheck(pg); 1608 #endif /* DEBUG */ 1609 1610 1611 /* global list */ 1612 pgfl = &uvm.page_free[index].pgfl_buckets[color].pgfl_queues[queue]; 1613 LIST_INSERT_HEAD(pgfl, pg, pageq.list); 1614 uvmexp.free++; 1615 if (iszero) { 1616 uvmexp.zeropages++; 1617 } 1618 1619 /* per-cpu list */ 1620 ucpu = curcpu()->ci_data.cpu_uvm; 1621 pg->offset = (uintptr_t)ucpu; 1622 pgfl = &ucpu->page_free[index].pgfl_buckets[color].pgfl_queues[queue]; 1623 LIST_INSERT_HEAD(pgfl, pg, listq.list); 1624 ucpu->pages[queue]++; 1625 if (ucpu->pages[PGFL_ZEROS] < ucpu->pages[PGFL_UNKNOWN]) { 1626 ucpu->page_idle_zero = vm_page_zero_enable; 1627 } 1628 1629 mutex_spin_exit(&uvm_fpageqlock); 1630 } 1631 1632 /* 1633 * uvm_page_unbusy: unbusy an array of pages. 1634 * 1635 * => pages must either all belong to the same object, or all belong to anons. 1636 * => if pages are object-owned, object must be locked. 1637 * => if pages are anon-owned, anons must be locked. 1638 * => caller must lock page queues if pages may be released. 1639 * => caller must make sure that anon-owned pages are not PG_RELEASED. 1640 */ 1641 1642 void 1643 uvm_page_unbusy(struct vm_page **pgs, int npgs) 1644 { 1645 struct vm_page *pg; 1646 int i; 1647 UVMHIST_FUNC("uvm_page_unbusy"); UVMHIST_CALLED(ubchist); 1648 1649 for (i = 0; i < npgs; i++) { 1650 pg = pgs[i]; 1651 if (pg == NULL || pg == PGO_DONTCARE) { 1652 continue; 1653 } 1654 1655 KASSERT(uvm_page_locked_p(pg)); 1656 KASSERT(pg->flags & PG_BUSY); 1657 KASSERT((pg->flags & PG_PAGEOUT) == 0); 1658 if (pg->flags & PG_WANTED) { 1659 wakeup(pg); 1660 } 1661 if (pg->flags & PG_RELEASED) { 1662 UVMHIST_LOG(ubchist, "releasing pg %p", pg,0,0,0); 1663 KASSERT(pg->uobject != NULL || 1664 (pg->uanon != NULL && pg->uanon->an_ref > 0)); 1665 pg->flags &= ~PG_RELEASED; 1666 uvm_pagefree(pg); 1667 } else { 1668 UVMHIST_LOG(ubchist, "unbusying pg %p", pg,0,0,0); 1669 KASSERT((pg->flags & PG_FAKE) == 0); 1670 pg->flags &= ~(PG_WANTED|PG_BUSY); 1671 UVM_PAGE_OWN(pg, NULL); 1672 } 1673 } 1674 } 1675 1676 #if defined(UVM_PAGE_TRKOWN) 1677 /* 1678 * uvm_page_own: set or release page ownership 1679 * 1680 * => this is a debugging function that keeps track of who sets PG_BUSY 1681 * and where they do it. it can be used to track down problems 1682 * such a process setting "PG_BUSY" and never releasing it. 1683 * => page's object [if any] must be locked 1684 * => if "tag" is NULL then we are releasing page ownership 1685 */ 1686 void 1687 uvm_page_own(struct vm_page *pg, const char *tag) 1688 { 1689 1690 KASSERT((pg->flags & (PG_PAGEOUT|PG_RELEASED)) == 0); 1691 KASSERT((pg->flags & PG_WANTED) == 0); 1692 KASSERT(uvm_page_locked_p(pg)); 1693 1694 /* gain ownership? */ 1695 if (tag) { 1696 KASSERT((pg->flags & PG_BUSY) != 0); 1697 if (pg->owner_tag) { 1698 printf("uvm_page_own: page %p already owned " 1699 "by proc %d [%s]\n", pg, 1700 pg->owner, pg->owner_tag); 1701 panic("uvm_page_own"); 1702 } 1703 pg->owner = curproc->p_pid; 1704 pg->lowner = curlwp->l_lid; 1705 pg->owner_tag = tag; 1706 return; 1707 } 1708 1709 /* drop ownership */ 1710 KASSERT((pg->flags & PG_BUSY) == 0); 1711 if (pg->owner_tag == NULL) { 1712 printf("uvm_page_own: dropping ownership of an non-owned " 1713 "page (%p)\n", pg); 1714 panic("uvm_page_own"); 1715 } 1716 if (!uvmpdpol_pageisqueued_p(pg)) { 1717 KASSERT((pg->uanon == NULL && pg->uobject == NULL) || 1718 pg->wire_count > 0); 1719 } else { 1720 KASSERT(pg->wire_count == 0); 1721 } 1722 pg->owner_tag = NULL; 1723 } 1724 #endif 1725 1726 /* 1727 * uvm_pageidlezero: zero free pages while the system is idle. 1728 * 1729 * => try to complete one color bucket at a time, to reduce our impact 1730 * on the CPU cache. 1731 * => we loop until we either reach the target or there is a lwp ready 1732 * to run, or MD code detects a reason to break early. 1733 */ 1734 void 1735 uvm_pageidlezero(void) 1736 { 1737 struct vm_page *pg; 1738 struct pgfreelist *pgfl, *gpgfl; 1739 struct uvm_cpu *ucpu; 1740 int free_list, firstbucket, nextbucket; 1741 bool lcont = false; 1742 1743 ucpu = curcpu()->ci_data.cpu_uvm; 1744 if (!ucpu->page_idle_zero || 1745 ucpu->pages[PGFL_UNKNOWN] < uvmexp.ncolors) { 1746 ucpu->page_idle_zero = false; 1747 return; 1748 } 1749 if (!mutex_tryenter(&uvm_fpageqlock)) { 1750 /* Contention: let other CPUs to use the lock. */ 1751 return; 1752 } 1753 firstbucket = ucpu->page_free_nextcolor; 1754 nextbucket = firstbucket; 1755 do { 1756 for (free_list = 0; free_list < VM_NFREELIST; free_list++) { 1757 if (sched_curcpu_runnable_p()) { 1758 goto quit; 1759 } 1760 pgfl = &ucpu->page_free[free_list]; 1761 gpgfl = &uvm.page_free[free_list]; 1762 while ((pg = LIST_FIRST(&pgfl->pgfl_buckets[ 1763 nextbucket].pgfl_queues[PGFL_UNKNOWN])) != NULL) { 1764 if (lcont || sched_curcpu_runnable_p()) { 1765 goto quit; 1766 } 1767 LIST_REMOVE(pg, pageq.list); /* global list */ 1768 LIST_REMOVE(pg, listq.list); /* per-cpu list */ 1769 ucpu->pages[PGFL_UNKNOWN]--; 1770 uvmexp.free--; 1771 KASSERT(pg->pqflags == PQ_FREE); 1772 pg->pqflags = 0; 1773 mutex_spin_exit(&uvm_fpageqlock); 1774 #ifdef PMAP_PAGEIDLEZERO 1775 if (!PMAP_PAGEIDLEZERO(VM_PAGE_TO_PHYS(pg))) { 1776 1777 /* 1778 * The machine-dependent code detected 1779 * some reason for us to abort zeroing 1780 * pages, probably because there is a 1781 * process now ready to run. 1782 */ 1783 1784 mutex_spin_enter(&uvm_fpageqlock); 1785 pg->pqflags = PQ_FREE; 1786 LIST_INSERT_HEAD(&gpgfl->pgfl_buckets[ 1787 nextbucket].pgfl_queues[ 1788 PGFL_UNKNOWN], pg, pageq.list); 1789 LIST_INSERT_HEAD(&pgfl->pgfl_buckets[ 1790 nextbucket].pgfl_queues[ 1791 PGFL_UNKNOWN], pg, listq.list); 1792 ucpu->pages[PGFL_UNKNOWN]++; 1793 uvmexp.free++; 1794 uvmexp.zeroaborts++; 1795 goto quit; 1796 } 1797 #else 1798 pmap_zero_page(VM_PAGE_TO_PHYS(pg)); 1799 #endif /* PMAP_PAGEIDLEZERO */ 1800 pg->flags |= PG_ZERO; 1801 1802 if (!mutex_tryenter(&uvm_fpageqlock)) { 1803 lcont = true; 1804 mutex_spin_enter(&uvm_fpageqlock); 1805 } else { 1806 lcont = false; 1807 } 1808 pg->pqflags = PQ_FREE; 1809 LIST_INSERT_HEAD(&gpgfl->pgfl_buckets[ 1810 nextbucket].pgfl_queues[PGFL_ZEROS], 1811 pg, pageq.list); 1812 LIST_INSERT_HEAD(&pgfl->pgfl_buckets[ 1813 nextbucket].pgfl_queues[PGFL_ZEROS], 1814 pg, listq.list); 1815 ucpu->pages[PGFL_ZEROS]++; 1816 uvmexp.free++; 1817 uvmexp.zeropages++; 1818 } 1819 } 1820 if (ucpu->pages[PGFL_UNKNOWN] < uvmexp.ncolors) { 1821 break; 1822 } 1823 nextbucket = (nextbucket + 1) & uvmexp.colormask; 1824 } while (nextbucket != firstbucket); 1825 ucpu->page_idle_zero = false; 1826 quit: 1827 mutex_spin_exit(&uvm_fpageqlock); 1828 } 1829 1830 /* 1831 * uvm_pagelookup: look up a page 1832 * 1833 * => caller should lock object to keep someone from pulling the page 1834 * out from under it 1835 */ 1836 1837 struct vm_page * 1838 uvm_pagelookup(struct uvm_object *obj, voff_t off) 1839 { 1840 struct vm_page *pg; 1841 1842 KASSERT(mutex_owned(obj->vmobjlock)); 1843 1844 pg = rb_tree_find_node(&obj->rb_tree, &off); 1845 1846 KASSERT(pg == NULL || obj->uo_npages != 0); 1847 KASSERT(pg == NULL || (pg->flags & (PG_RELEASED|PG_PAGEOUT)) == 0 || 1848 (pg->flags & PG_BUSY) != 0); 1849 return pg; 1850 } 1851 1852 /* 1853 * uvm_pagewire: wire the page, thus removing it from the daemon's grasp 1854 * 1855 * => caller must lock page queues 1856 */ 1857 1858 void 1859 uvm_pagewire(struct vm_page *pg) 1860 { 1861 KASSERT(mutex_owned(&uvm_pageqlock)); 1862 #if defined(READAHEAD_STATS) 1863 if ((pg->pqflags & PQ_READAHEAD) != 0) { 1864 uvm_ra_hit.ev_count++; 1865 pg->pqflags &= ~PQ_READAHEAD; 1866 } 1867 #endif /* defined(READAHEAD_STATS) */ 1868 if (pg->wire_count == 0) { 1869 uvm_pagedequeue(pg); 1870 uvmexp.wired++; 1871 } 1872 pg->wire_count++; 1873 } 1874 1875 /* 1876 * uvm_pageunwire: unwire the page. 1877 * 1878 * => activate if wire count goes to zero. 1879 * => caller must lock page queues 1880 */ 1881 1882 void 1883 uvm_pageunwire(struct vm_page *pg) 1884 { 1885 KASSERT(mutex_owned(&uvm_pageqlock)); 1886 pg->wire_count--; 1887 if (pg->wire_count == 0) { 1888 uvm_pageactivate(pg); 1889 uvmexp.wired--; 1890 } 1891 } 1892 1893 /* 1894 * uvm_pagedeactivate: deactivate page 1895 * 1896 * => caller must lock page queues 1897 * => caller must check to make sure page is not wired 1898 * => object that page belongs to must be locked (so we can adjust pg->flags) 1899 * => caller must clear the reference on the page before calling 1900 */ 1901 1902 void 1903 uvm_pagedeactivate(struct vm_page *pg) 1904 { 1905 1906 KASSERT(mutex_owned(&uvm_pageqlock)); 1907 KASSERT(uvm_page_locked_p(pg)); 1908 KASSERT(pg->wire_count != 0 || uvmpdpol_pageisqueued_p(pg)); 1909 uvmpdpol_pagedeactivate(pg); 1910 } 1911 1912 /* 1913 * uvm_pageactivate: activate page 1914 * 1915 * => caller must lock page queues 1916 */ 1917 1918 void 1919 uvm_pageactivate(struct vm_page *pg) 1920 { 1921 1922 KASSERT(mutex_owned(&uvm_pageqlock)); 1923 KASSERT(uvm_page_locked_p(pg)); 1924 #if defined(READAHEAD_STATS) 1925 if ((pg->pqflags & PQ_READAHEAD) != 0) { 1926 uvm_ra_hit.ev_count++; 1927 pg->pqflags &= ~PQ_READAHEAD; 1928 } 1929 #endif /* defined(READAHEAD_STATS) */ 1930 if (pg->wire_count != 0) { 1931 return; 1932 } 1933 uvmpdpol_pageactivate(pg); 1934 } 1935 1936 /* 1937 * uvm_pagedequeue: remove a page from any paging queue 1938 */ 1939 1940 void 1941 uvm_pagedequeue(struct vm_page *pg) 1942 { 1943 1944 if (uvmpdpol_pageisqueued_p(pg)) { 1945 KASSERT(mutex_owned(&uvm_pageqlock)); 1946 } 1947 1948 uvmpdpol_pagedequeue(pg); 1949 } 1950 1951 /* 1952 * uvm_pageenqueue: add a page to a paging queue without activating. 1953 * used where a page is not really demanded (yet). eg. read-ahead 1954 */ 1955 1956 void 1957 uvm_pageenqueue(struct vm_page *pg) 1958 { 1959 1960 KASSERT(mutex_owned(&uvm_pageqlock)); 1961 if (pg->wire_count != 0) { 1962 return; 1963 } 1964 uvmpdpol_pageenqueue(pg); 1965 } 1966 1967 /* 1968 * uvm_pagezero: zero fill a page 1969 * 1970 * => if page is part of an object then the object should be locked 1971 * to protect pg->flags. 1972 */ 1973 1974 void 1975 uvm_pagezero(struct vm_page *pg) 1976 { 1977 pg->flags &= ~PG_CLEAN; 1978 pmap_zero_page(VM_PAGE_TO_PHYS(pg)); 1979 } 1980 1981 /* 1982 * uvm_pagecopy: copy a page 1983 * 1984 * => if page is part of an object then the object should be locked 1985 * to protect pg->flags. 1986 */ 1987 1988 void 1989 uvm_pagecopy(struct vm_page *src, struct vm_page *dst) 1990 { 1991 1992 dst->flags &= ~PG_CLEAN; 1993 pmap_copy_page(VM_PAGE_TO_PHYS(src), VM_PAGE_TO_PHYS(dst)); 1994 } 1995 1996 /* 1997 * uvm_pageismanaged: test it see that a page (specified by PA) is managed. 1998 */ 1999 2000 bool 2001 uvm_pageismanaged(paddr_t pa) 2002 { 2003 2004 return (vm_physseg_find(atop(pa), NULL) != -1); 2005 } 2006 2007 /* 2008 * uvm_page_lookup_freelist: look up the free list for the specified page 2009 */ 2010 2011 int 2012 uvm_page_lookup_freelist(struct vm_page *pg) 2013 { 2014 int lcv; 2015 2016 lcv = vm_physseg_find(atop(VM_PAGE_TO_PHYS(pg)), NULL); 2017 KASSERT(lcv != -1); 2018 return (VM_PHYSMEM_PTR(lcv)->free_list); 2019 } 2020 2021 /* 2022 * uvm_page_locked_p: return true if object associated with page is 2023 * locked. this is a weak check for runtime assertions only. 2024 */ 2025 2026 bool 2027 uvm_page_locked_p(struct vm_page *pg) 2028 { 2029 2030 if (pg->uobject != NULL) { 2031 return mutex_owned(pg->uobject->vmobjlock); 2032 } 2033 if (pg->uanon != NULL) { 2034 return mutex_owned(pg->uanon->an_lock); 2035 } 2036 return true; 2037 } 2038 2039 #if defined(DDB) || defined(DEBUGPRINT) 2040 2041 /* 2042 * uvm_page_printit: actually print the page 2043 */ 2044 2045 static const char page_flagbits[] = UVM_PGFLAGBITS; 2046 static const char page_pqflagbits[] = UVM_PQFLAGBITS; 2047 2048 void 2049 uvm_page_printit(struct vm_page *pg, bool full, 2050 void (*pr)(const char *, ...)) 2051 { 2052 struct vm_page *tpg; 2053 struct uvm_object *uobj; 2054 struct pgflist *pgl; 2055 char pgbuf[128]; 2056 char pqbuf[128]; 2057 2058 (*pr)("PAGE %p:\n", pg); 2059 snprintb(pgbuf, sizeof(pgbuf), page_flagbits, pg->flags); 2060 snprintb(pqbuf, sizeof(pqbuf), page_pqflagbits, pg->pqflags); 2061 (*pr)(" flags=%s, pqflags=%s, wire_count=%d, pa=0x%lx\n", 2062 pgbuf, pqbuf, pg->wire_count, (long)VM_PAGE_TO_PHYS(pg)); 2063 (*pr)(" uobject=%p, uanon=%p, offset=0x%llx loan_count=%d\n", 2064 pg->uobject, pg->uanon, (long long)pg->offset, pg->loan_count); 2065 #if defined(UVM_PAGE_TRKOWN) 2066 if (pg->flags & PG_BUSY) 2067 (*pr)(" owning process = %d, tag=%s\n", 2068 pg->owner, pg->owner_tag); 2069 else 2070 (*pr)(" page not busy, no owner\n"); 2071 #else 2072 (*pr)(" [page ownership tracking disabled]\n"); 2073 #endif 2074 2075 if (!full) 2076 return; 2077 2078 /* cross-verify object/anon */ 2079 if ((pg->pqflags & PQ_FREE) == 0) { 2080 if (pg->pqflags & PQ_ANON) { 2081 if (pg->uanon == NULL || pg->uanon->an_page != pg) 2082 (*pr)(" >>> ANON DOES NOT POINT HERE <<< (%p)\n", 2083 (pg->uanon) ? pg->uanon->an_page : NULL); 2084 else 2085 (*pr)(" anon backpointer is OK\n"); 2086 } else { 2087 uobj = pg->uobject; 2088 if (uobj) { 2089 (*pr)(" checking object list\n"); 2090 TAILQ_FOREACH(tpg, &uobj->memq, listq.queue) { 2091 if (tpg == pg) { 2092 break; 2093 } 2094 } 2095 if (tpg) 2096 (*pr)(" page found on object list\n"); 2097 else 2098 (*pr)(" >>> PAGE NOT FOUND ON OBJECT LIST! <<<\n"); 2099 } 2100 } 2101 } 2102 2103 /* cross-verify page queue */ 2104 if (pg->pqflags & PQ_FREE) { 2105 int fl = uvm_page_lookup_freelist(pg); 2106 int color = VM_PGCOLOR_BUCKET(pg); 2107 pgl = &uvm.page_free[fl].pgfl_buckets[color].pgfl_queues[ 2108 ((pg)->flags & PG_ZERO) ? PGFL_ZEROS : PGFL_UNKNOWN]; 2109 } else { 2110 pgl = NULL; 2111 } 2112 2113 if (pgl) { 2114 (*pr)(" checking pageq list\n"); 2115 LIST_FOREACH(tpg, pgl, pageq.list) { 2116 if (tpg == pg) { 2117 break; 2118 } 2119 } 2120 if (tpg) 2121 (*pr)(" page found on pageq list\n"); 2122 else 2123 (*pr)(" >>> PAGE NOT FOUND ON PAGEQ LIST! <<<\n"); 2124 } 2125 } 2126 2127 /* 2128 * uvm_pages_printthem - print a summary of all managed pages 2129 */ 2130 2131 void 2132 uvm_page_printall(void (*pr)(const char *, ...)) 2133 { 2134 unsigned i; 2135 struct vm_page *pg; 2136 2137 (*pr)("%18s %4s %4s %18s %18s" 2138 #ifdef UVM_PAGE_TRKOWN 2139 " OWNER" 2140 #endif 2141 "\n", "PAGE", "FLAG", "PQ", "UOBJECT", "UANON"); 2142 for (i = 0; i < vm_nphysmem; i++) { 2143 for (pg = VM_PHYSMEM_PTR(i)->pgs; pg < VM_PHYSMEM_PTR(i)->lastpg; pg++) { 2144 (*pr)("%18p %04x %04x %18p %18p", 2145 pg, pg->flags, pg->pqflags, pg->uobject, 2146 pg->uanon); 2147 #ifdef UVM_PAGE_TRKOWN 2148 if (pg->flags & PG_BUSY) 2149 (*pr)(" %d [%s]", pg->owner, pg->owner_tag); 2150 #endif 2151 (*pr)("\n"); 2152 } 2153 } 2154 } 2155 2156 #endif /* DDB || DEBUGPRINT */ 2157