1 /* $NetBSD: uvm_page.c,v 1.200 2019/09/20 11:09:43 maxv 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 1.200 2019/09/20 11:09:43 maxv Exp $"); 70 71 #include "opt_ddb.h" 72 #include "opt_uvm.h" 73 #include "opt_uvmhist.h" 74 #include "opt_readahead.h" 75 76 #include <sys/param.h> 77 #include <sys/systm.h> 78 #include <sys/sched.h> 79 #include <sys/kernel.h> 80 #include <sys/vnode.h> 81 #include <sys/proc.h> 82 #include <sys/atomic.h> 83 #include <sys/cpu.h> 84 #include <sys/extent.h> 85 86 #include <uvm/uvm.h> 87 #include <uvm/uvm_ddb.h> 88 #include <uvm/uvm_pdpolicy.h> 89 90 /* 91 * Some supported CPUs in a given architecture don't support all 92 * of the things necessary to do idle page zero'ing efficiently. 93 * We therefore provide a way to enable it from machdep code here. 94 */ 95 bool vm_page_zero_enable = false; 96 97 /* 98 * number of pages per-CPU to reserve for the kernel. 99 */ 100 #ifndef UVM_RESERVED_PAGES_PER_CPU 101 #define UVM_RESERVED_PAGES_PER_CPU 5 102 #endif 103 int vm_page_reserve_kernel = UVM_RESERVED_PAGES_PER_CPU; 104 105 /* 106 * physical memory size; 107 */ 108 psize_t physmem; 109 110 /* 111 * local variables 112 */ 113 114 /* 115 * these variables record the values returned by vm_page_bootstrap, 116 * for debugging purposes. The implementation of uvm_pageboot_alloc 117 * and pmap_startup here also uses them internally. 118 */ 119 120 static vaddr_t virtual_space_start; 121 static vaddr_t virtual_space_end; 122 123 /* 124 * we allocate an initial number of page colors in uvm_page_init(), 125 * and remember them. We may re-color pages as cache sizes are 126 * discovered during the autoconfiguration phase. But we can never 127 * free the initial set of buckets, since they are allocated using 128 * uvm_pageboot_alloc(). 129 */ 130 131 static size_t recolored_pages_memsize /* = 0 */; 132 133 #ifdef DEBUG 134 vaddr_t uvm_zerocheckkva; 135 #endif /* DEBUG */ 136 137 /* 138 * These functions are reserved for uvm(9) internal use and are not 139 * exported in the header file uvm_physseg.h 140 * 141 * Thus they are redefined here. 142 */ 143 void uvm_physseg_init_seg(uvm_physseg_t, struct vm_page *); 144 void uvm_physseg_seg_chomp_slab(uvm_physseg_t, struct vm_page *, size_t); 145 146 /* returns a pgs array */ 147 struct vm_page *uvm_physseg_seg_alloc_from_slab(uvm_physseg_t, size_t); 148 149 /* 150 * local prototypes 151 */ 152 153 static void uvm_pageinsert(struct uvm_object *, struct vm_page *); 154 static void uvm_pageremove(struct uvm_object *, struct vm_page *); 155 156 /* 157 * per-object tree of pages 158 */ 159 160 static signed int 161 uvm_page_compare_nodes(void *ctx, const void *n1, const void *n2) 162 { 163 const struct vm_page *pg1 = n1; 164 const struct vm_page *pg2 = n2; 165 const voff_t a = pg1->offset; 166 const voff_t b = pg2->offset; 167 168 if (a < b) 169 return -1; 170 if (a > b) 171 return 1; 172 return 0; 173 } 174 175 static signed int 176 uvm_page_compare_key(void *ctx, const void *n, const void *key) 177 { 178 const struct vm_page *pg = n; 179 const voff_t a = pg->offset; 180 const voff_t b = *(const voff_t *)key; 181 182 if (a < b) 183 return -1; 184 if (a > b) 185 return 1; 186 return 0; 187 } 188 189 const rb_tree_ops_t uvm_page_tree_ops = { 190 .rbto_compare_nodes = uvm_page_compare_nodes, 191 .rbto_compare_key = uvm_page_compare_key, 192 .rbto_node_offset = offsetof(struct vm_page, rb_node), 193 .rbto_context = NULL 194 }; 195 196 /* 197 * inline functions 198 */ 199 200 /* 201 * uvm_pageinsert: insert a page in the object. 202 * 203 * => caller must lock object 204 * => caller must lock page queues 205 * => call should have already set pg's object and offset pointers 206 * and bumped the version counter 207 */ 208 209 static inline void 210 uvm_pageinsert_list(struct uvm_object *uobj, struct vm_page *pg, 211 struct vm_page *where) 212 { 213 214 KASSERT(uobj == pg->uobject); 215 KASSERT(mutex_owned(uobj->vmobjlock)); 216 KASSERT((pg->flags & PG_TABLED) == 0); 217 KASSERT(where == NULL || (where->flags & PG_TABLED)); 218 KASSERT(where == NULL || (where->uobject == uobj)); 219 220 if (UVM_OBJ_IS_VNODE(uobj)) { 221 if (uobj->uo_npages == 0) { 222 struct vnode *vp = (struct vnode *)uobj; 223 224 vholdl(vp); 225 } 226 if (UVM_OBJ_IS_VTEXT(uobj)) { 227 atomic_inc_uint(&uvmexp.execpages); 228 } else { 229 atomic_inc_uint(&uvmexp.filepages); 230 } 231 } else if (UVM_OBJ_IS_AOBJ(uobj)) { 232 atomic_inc_uint(&uvmexp.anonpages); 233 } 234 235 if (where) 236 TAILQ_INSERT_AFTER(&uobj->memq, where, pg, listq.queue); 237 else 238 TAILQ_INSERT_TAIL(&uobj->memq, pg, listq.queue); 239 pg->flags |= PG_TABLED; 240 uobj->uo_npages++; 241 } 242 243 244 static inline void 245 uvm_pageinsert_tree(struct uvm_object *uobj, struct vm_page *pg) 246 { 247 struct vm_page *ret __diagused; 248 249 KASSERT(uobj == pg->uobject); 250 ret = rb_tree_insert_node(&uobj->rb_tree, pg); 251 KASSERT(ret == pg); 252 } 253 254 static inline void 255 uvm_pageinsert(struct uvm_object *uobj, struct vm_page *pg) 256 { 257 258 KDASSERT(uobj != NULL); 259 uvm_pageinsert_tree(uobj, pg); 260 uvm_pageinsert_list(uobj, pg, NULL); 261 } 262 263 /* 264 * uvm_page_remove: remove page from object. 265 * 266 * => caller must lock object 267 * => caller must lock page queues 268 */ 269 270 static inline void 271 uvm_pageremove_list(struct uvm_object *uobj, struct vm_page *pg) 272 { 273 274 KASSERT(uobj == pg->uobject); 275 KASSERT(mutex_owned(uobj->vmobjlock)); 276 KASSERT(pg->flags & PG_TABLED); 277 278 if (UVM_OBJ_IS_VNODE(uobj)) { 279 if (uobj->uo_npages == 1) { 280 struct vnode *vp = (struct vnode *)uobj; 281 282 holdrelel(vp); 283 } 284 if (UVM_OBJ_IS_VTEXT(uobj)) { 285 atomic_dec_uint(&uvmexp.execpages); 286 } else { 287 atomic_dec_uint(&uvmexp.filepages); 288 } 289 } else if (UVM_OBJ_IS_AOBJ(uobj)) { 290 atomic_dec_uint(&uvmexp.anonpages); 291 } 292 293 /* object should be locked */ 294 uobj->uo_npages--; 295 TAILQ_REMOVE(&uobj->memq, pg, listq.queue); 296 pg->flags &= ~PG_TABLED; 297 pg->uobject = NULL; 298 } 299 300 static inline void 301 uvm_pageremove_tree(struct uvm_object *uobj, struct vm_page *pg) 302 { 303 304 KASSERT(uobj == pg->uobject); 305 rb_tree_remove_node(&uobj->rb_tree, pg); 306 } 307 308 static inline void 309 uvm_pageremove(struct uvm_object *uobj, struct vm_page *pg) 310 { 311 312 KDASSERT(uobj != NULL); 313 uvm_pageremove_tree(uobj, pg); 314 uvm_pageremove_list(uobj, pg); 315 } 316 317 static void 318 uvm_page_init_buckets(struct pgfreelist *pgfl) 319 { 320 int color, i; 321 322 for (color = 0; color < uvmexp.ncolors; color++) { 323 for (i = 0; i < PGFL_NQUEUES; i++) { 324 LIST_INIT(&pgfl->pgfl_buckets[color].pgfl_queues[i]); 325 } 326 } 327 } 328 329 /* 330 * uvm_page_init: init the page system. called from uvm_init(). 331 * 332 * => we return the range of kernel virtual memory in kvm_startp/kvm_endp 333 */ 334 335 void 336 uvm_page_init(vaddr_t *kvm_startp, vaddr_t *kvm_endp) 337 { 338 static struct uvm_cpu boot_cpu; 339 psize_t freepages, pagecount, bucketcount, n; 340 struct pgflbucket *bucketarray, *cpuarray; 341 struct vm_page *pagearray; 342 uvm_physseg_t bank; 343 int lcv; 344 345 KASSERT(ncpu <= 1); 346 CTASSERT(sizeof(pagearray->offset) >= sizeof(struct uvm_cpu *)); 347 348 /* 349 * init the page queues and page queue locks, except the free 350 * list; we allocate that later (with the initial vm_page 351 * structures). 352 */ 353 354 uvm.cpus[0] = &boot_cpu; 355 curcpu()->ci_data.cpu_uvm = &boot_cpu; 356 uvmpdpol_init(); 357 mutex_init(&uvm_pageqlock, MUTEX_DRIVER, IPL_NONE); 358 mutex_init(&uvm_fpageqlock, MUTEX_DRIVER, IPL_VM); 359 360 /* 361 * allocate vm_page structures. 362 */ 363 364 /* 365 * sanity check: 366 * before calling this function the MD code is expected to register 367 * some free RAM with the uvm_page_physload() function. our job 368 * now is to allocate vm_page structures for this memory. 369 */ 370 371 if (uvm_physseg_get_last() == UVM_PHYSSEG_TYPE_INVALID) 372 panic("uvm_page_bootstrap: no memory pre-allocated"); 373 374 /* 375 * first calculate the number of free pages... 376 * 377 * note that we use start/end rather than avail_start/avail_end. 378 * this allows us to allocate extra vm_page structures in case we 379 * want to return some memory to the pool after booting. 380 */ 381 382 freepages = 0; 383 384 for (bank = uvm_physseg_get_first(); 385 uvm_physseg_valid_p(bank) ; 386 bank = uvm_physseg_get_next(bank)) { 387 freepages += (uvm_physseg_get_end(bank) - uvm_physseg_get_start(bank)); 388 } 389 390 /* 391 * Let MD code initialize the number of colors, or default 392 * to 1 color if MD code doesn't care. 393 */ 394 if (uvmexp.ncolors == 0) 395 uvmexp.ncolors = 1; 396 uvmexp.colormask = uvmexp.ncolors - 1; 397 KASSERT((uvmexp.colormask & uvmexp.ncolors) == 0); 398 399 /* 400 * we now know we have (PAGE_SIZE * freepages) bytes of memory we can 401 * use. for each page of memory we use we need a vm_page structure. 402 * thus, the total number of pages we can use is the total size of 403 * the memory divided by the PAGE_SIZE plus the size of the vm_page 404 * structure. we add one to freepages as a fudge factor to avoid 405 * truncation errors (since we can only allocate in terms of whole 406 * pages). 407 */ 408 409 bucketcount = uvmexp.ncolors * VM_NFREELIST; 410 pagecount = ((freepages + 1) << PAGE_SHIFT) / 411 (PAGE_SIZE + sizeof(struct vm_page)); 412 413 bucketarray = (void *)uvm_pageboot_alloc((bucketcount * 414 sizeof(struct pgflbucket) * 2) + (pagecount * 415 sizeof(struct vm_page))); 416 cpuarray = bucketarray + bucketcount; 417 pagearray = (struct vm_page *)(bucketarray + bucketcount * 2); 418 419 for (lcv = 0; lcv < VM_NFREELIST; lcv++) { 420 uvm.page_free[lcv].pgfl_buckets = 421 (bucketarray + (lcv * uvmexp.ncolors)); 422 uvm_page_init_buckets(&uvm.page_free[lcv]); 423 uvm.cpus[0]->page_free[lcv].pgfl_buckets = 424 (cpuarray + (lcv * uvmexp.ncolors)); 425 uvm_page_init_buckets(&uvm.cpus[0]->page_free[lcv]); 426 } 427 memset(pagearray, 0, pagecount * sizeof(struct vm_page)); 428 429 /* 430 * init the vm_page structures and put them in the correct place. 431 */ 432 /* First init the extent */ 433 434 for (bank = uvm_physseg_get_first(), 435 uvm_physseg_seg_chomp_slab(bank, pagearray, pagecount); 436 uvm_physseg_valid_p(bank); 437 bank = uvm_physseg_get_next(bank)) { 438 439 n = uvm_physseg_get_end(bank) - uvm_physseg_get_start(bank); 440 uvm_physseg_seg_alloc_from_slab(bank, n); 441 uvm_physseg_init_seg(bank, pagearray); 442 443 /* set up page array pointers */ 444 pagearray += n; 445 pagecount -= n; 446 } 447 448 /* 449 * pass up the values of virtual_space_start and 450 * virtual_space_end (obtained by uvm_pageboot_alloc) to the upper 451 * layers of the VM. 452 */ 453 454 *kvm_startp = round_page(virtual_space_start); 455 *kvm_endp = trunc_page(virtual_space_end); 456 #ifdef DEBUG 457 /* 458 * steal kva for uvm_pagezerocheck(). 459 */ 460 uvm_zerocheckkva = *kvm_startp; 461 *kvm_startp += PAGE_SIZE; 462 #endif /* DEBUG */ 463 464 /* 465 * init various thresholds. 466 */ 467 468 uvmexp.reserve_pagedaemon = 1; 469 uvmexp.reserve_kernel = vm_page_reserve_kernel; 470 471 /* 472 * determine if we should zero pages in the idle loop. 473 */ 474 475 uvm.cpus[0]->page_idle_zero = vm_page_zero_enable; 476 477 /* 478 * done! 479 */ 480 481 uvm.page_init_done = true; 482 } 483 484 /* 485 * uvm_setpagesize: set the page size 486 * 487 * => sets page_shift and page_mask from uvmexp.pagesize. 488 */ 489 490 void 491 uvm_setpagesize(void) 492 { 493 494 /* 495 * If uvmexp.pagesize is 0 at this point, we expect PAGE_SIZE 496 * to be a constant (indicated by being a non-zero value). 497 */ 498 if (uvmexp.pagesize == 0) { 499 if (PAGE_SIZE == 0) 500 panic("uvm_setpagesize: uvmexp.pagesize not set"); 501 uvmexp.pagesize = PAGE_SIZE; 502 } 503 uvmexp.pagemask = uvmexp.pagesize - 1; 504 if ((uvmexp.pagemask & uvmexp.pagesize) != 0) 505 panic("uvm_setpagesize: page size %u (%#x) not a power of two", 506 uvmexp.pagesize, uvmexp.pagesize); 507 for (uvmexp.pageshift = 0; ; uvmexp.pageshift++) 508 if ((1 << uvmexp.pageshift) == uvmexp.pagesize) 509 break; 510 } 511 512 /* 513 * uvm_pageboot_alloc: steal memory from physmem for bootstrapping 514 */ 515 516 vaddr_t 517 uvm_pageboot_alloc(vsize_t size) 518 { 519 static bool initialized = false; 520 vaddr_t addr; 521 #if !defined(PMAP_STEAL_MEMORY) 522 vaddr_t vaddr; 523 paddr_t paddr; 524 #endif 525 526 /* 527 * on first call to this function, initialize ourselves. 528 */ 529 if (initialized == false) { 530 pmap_virtual_space(&virtual_space_start, &virtual_space_end); 531 532 /* round it the way we like it */ 533 virtual_space_start = round_page(virtual_space_start); 534 virtual_space_end = trunc_page(virtual_space_end); 535 536 initialized = true; 537 } 538 539 /* round to page size */ 540 size = round_page(size); 541 uvmexp.bootpages += atop(size); 542 543 #if defined(PMAP_STEAL_MEMORY) 544 545 /* 546 * defer bootstrap allocation to MD code (it may want to allocate 547 * from a direct-mapped segment). pmap_steal_memory should adjust 548 * virtual_space_start/virtual_space_end if necessary. 549 */ 550 551 addr = pmap_steal_memory(size, &virtual_space_start, 552 &virtual_space_end); 553 554 return(addr); 555 556 #else /* !PMAP_STEAL_MEMORY */ 557 558 /* 559 * allocate virtual memory for this request 560 */ 561 if (virtual_space_start == virtual_space_end || 562 (virtual_space_end - virtual_space_start) < size) 563 panic("uvm_pageboot_alloc: out of virtual space"); 564 565 addr = virtual_space_start; 566 567 #ifdef PMAP_GROWKERNEL 568 /* 569 * If the kernel pmap can't map the requested space, 570 * then allocate more resources for it. 571 */ 572 if (uvm_maxkaddr < (addr + size)) { 573 uvm_maxkaddr = pmap_growkernel(addr + size); 574 if (uvm_maxkaddr < (addr + size)) 575 panic("uvm_pageboot_alloc: pmap_growkernel() failed"); 576 } 577 #endif 578 579 virtual_space_start += size; 580 581 /* 582 * allocate and mapin physical pages to back new virtual pages 583 */ 584 585 for (vaddr = round_page(addr) ; vaddr < addr + size ; 586 vaddr += PAGE_SIZE) { 587 588 if (!uvm_page_physget(&paddr)) 589 panic("uvm_pageboot_alloc: out of memory"); 590 591 /* 592 * Note this memory is no longer managed, so using 593 * pmap_kenter is safe. 594 */ 595 pmap_kenter_pa(vaddr, paddr, VM_PROT_READ|VM_PROT_WRITE, 0); 596 } 597 pmap_update(pmap_kernel()); 598 return(addr); 599 #endif /* PMAP_STEAL_MEMORY */ 600 } 601 602 #if !defined(PMAP_STEAL_MEMORY) 603 /* 604 * uvm_page_physget: "steal" one page from the vm_physmem structure. 605 * 606 * => attempt to allocate it off the end of a segment in which the "avail" 607 * values match the start/end values. if we can't do that, then we 608 * will advance both values (making them equal, and removing some 609 * vm_page structures from the non-avail area). 610 * => return false if out of memory. 611 */ 612 613 /* subroutine: try to allocate from memory chunks on the specified freelist */ 614 static bool uvm_page_physget_freelist(paddr_t *, int); 615 616 static bool 617 uvm_page_physget_freelist(paddr_t *paddrp, int freelist) 618 { 619 uvm_physseg_t lcv; 620 621 /* pass 1: try allocating from a matching end */ 622 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST) 623 for (lcv = uvm_physseg_get_last(); uvm_physseg_valid_p(lcv); lcv = uvm_physseg_get_prev(lcv)) 624 #else 625 for (lcv = uvm_physseg_get_first(); uvm_physseg_valid_p(lcv); lcv = uvm_physseg_get_next(lcv)) 626 #endif 627 { 628 if (uvm.page_init_done == true) 629 panic("uvm_page_physget: called _after_ bootstrap"); 630 631 /* Try to match at front or back on unused segment */ 632 if (uvm_page_physunload(lcv, freelist, paddrp)) 633 return true; 634 } 635 636 /* pass2: forget about matching ends, just allocate something */ 637 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST) 638 for (lcv = uvm_physseg_get_last(); uvm_physseg_valid_p(lcv); lcv = uvm_physseg_get_prev(lcv)) 639 #else 640 for (lcv = uvm_physseg_get_first(); uvm_physseg_valid_p(lcv); lcv = uvm_physseg_get_next(lcv)) 641 #endif 642 { 643 /* Try the front regardless. */ 644 if (uvm_page_physunload_force(lcv, freelist, paddrp)) 645 return true; 646 } 647 return false; 648 } 649 650 bool 651 uvm_page_physget(paddr_t *paddrp) 652 { 653 int i; 654 655 /* try in the order of freelist preference */ 656 for (i = 0; i < VM_NFREELIST; i++) 657 if (uvm_page_physget_freelist(paddrp, i) == true) 658 return (true); 659 return (false); 660 } 661 #endif /* PMAP_STEAL_MEMORY */ 662 663 /* 664 * PHYS_TO_VM_PAGE: find vm_page for a PA. used by MI code to get vm_pages 665 * back from an I/O mapping (ugh!). used in some MD code as well. 666 */ 667 struct vm_page * 668 uvm_phys_to_vm_page(paddr_t pa) 669 { 670 paddr_t pf = atop(pa); 671 paddr_t off; 672 uvm_physseg_t upm; 673 674 upm = uvm_physseg_find(pf, &off); 675 if (upm != UVM_PHYSSEG_TYPE_INVALID) 676 return uvm_physseg_get_pg(upm, off); 677 return(NULL); 678 } 679 680 paddr_t 681 uvm_vm_page_to_phys(const struct vm_page *pg) 682 { 683 684 return pg->phys_addr; 685 } 686 687 /* 688 * uvm_page_recolor: Recolor the pages if the new bucket count is 689 * larger than the old one. 690 */ 691 692 void 693 uvm_page_recolor(int newncolors) 694 { 695 struct pgflbucket *bucketarray, *cpuarray, *oldbucketarray; 696 struct pgfreelist gpgfl, pgfl; 697 struct vm_page *pg; 698 vsize_t bucketcount; 699 size_t bucketmemsize, oldbucketmemsize; 700 int color, i, ocolors; 701 int lcv; 702 struct uvm_cpu *ucpu; 703 704 KASSERT(((newncolors - 1) & newncolors) == 0); 705 706 if (newncolors <= uvmexp.ncolors) 707 return; 708 709 if (uvm.page_init_done == false) { 710 uvmexp.ncolors = newncolors; 711 return; 712 } 713 714 bucketcount = newncolors * VM_NFREELIST; 715 bucketmemsize = bucketcount * sizeof(struct pgflbucket) * 2; 716 bucketarray = kmem_alloc(bucketmemsize, KM_SLEEP); 717 cpuarray = bucketarray + bucketcount; 718 719 mutex_spin_enter(&uvm_fpageqlock); 720 721 /* Make sure we should still do this. */ 722 if (newncolors <= uvmexp.ncolors) { 723 mutex_spin_exit(&uvm_fpageqlock); 724 kmem_free(bucketarray, bucketmemsize); 725 return; 726 } 727 728 oldbucketarray = uvm.page_free[0].pgfl_buckets; 729 ocolors = uvmexp.ncolors; 730 731 uvmexp.ncolors = newncolors; 732 uvmexp.colormask = uvmexp.ncolors - 1; 733 734 ucpu = curcpu()->ci_data.cpu_uvm; 735 for (lcv = 0; lcv < VM_NFREELIST; lcv++) { 736 gpgfl.pgfl_buckets = (bucketarray + (lcv * newncolors)); 737 pgfl.pgfl_buckets = (cpuarray + (lcv * uvmexp.ncolors)); 738 uvm_page_init_buckets(&gpgfl); 739 uvm_page_init_buckets(&pgfl); 740 for (color = 0; color < ocolors; color++) { 741 for (i = 0; i < PGFL_NQUEUES; i++) { 742 while ((pg = LIST_FIRST(&uvm.page_free[ 743 lcv].pgfl_buckets[color].pgfl_queues[i])) 744 != NULL) { 745 LIST_REMOVE(pg, pageq.list); /* global */ 746 LIST_REMOVE(pg, listq.list); /* cpu */ 747 LIST_INSERT_HEAD(&gpgfl.pgfl_buckets[ 748 VM_PGCOLOR_BUCKET(pg)].pgfl_queues[ 749 i], pg, pageq.list); 750 LIST_INSERT_HEAD(&pgfl.pgfl_buckets[ 751 VM_PGCOLOR_BUCKET(pg)].pgfl_queues[ 752 i], pg, listq.list); 753 } 754 } 755 } 756 uvm.page_free[lcv].pgfl_buckets = gpgfl.pgfl_buckets; 757 ucpu->page_free[lcv].pgfl_buckets = pgfl.pgfl_buckets; 758 } 759 760 oldbucketmemsize = recolored_pages_memsize; 761 762 recolored_pages_memsize = bucketmemsize; 763 mutex_spin_exit(&uvm_fpageqlock); 764 765 if (oldbucketmemsize) { 766 kmem_free(oldbucketarray, oldbucketmemsize); 767 } 768 769 /* 770 * this calls uvm_km_alloc() which may want to hold 771 * uvm_fpageqlock. 772 */ 773 uvm_pager_realloc_emerg(); 774 } 775 776 /* 777 * uvm_cpu_attach: initialize per-CPU data structures. 778 */ 779 780 void 781 uvm_cpu_attach(struct cpu_info *ci) 782 { 783 struct pgflbucket *bucketarray; 784 struct pgfreelist pgfl; 785 struct uvm_cpu *ucpu; 786 vsize_t bucketcount; 787 int lcv; 788 789 if (CPU_IS_PRIMARY(ci)) { 790 /* Already done in uvm_page_init(). */ 791 goto attachrnd; 792 } 793 794 /* Add more reserve pages for this CPU. */ 795 uvmexp.reserve_kernel += vm_page_reserve_kernel; 796 797 /* Configure this CPU's free lists. */ 798 bucketcount = uvmexp.ncolors * VM_NFREELIST; 799 bucketarray = kmem_alloc(bucketcount * sizeof(struct pgflbucket), 800 KM_SLEEP); 801 ucpu = kmem_zalloc(sizeof(*ucpu), KM_SLEEP); 802 uvm.cpus[cpu_index(ci)] = ucpu; 803 ci->ci_data.cpu_uvm = ucpu; 804 for (lcv = 0; lcv < VM_NFREELIST; lcv++) { 805 pgfl.pgfl_buckets = (bucketarray + (lcv * uvmexp.ncolors)); 806 uvm_page_init_buckets(&pgfl); 807 ucpu->page_free[lcv].pgfl_buckets = pgfl.pgfl_buckets; 808 } 809 810 attachrnd: 811 /* 812 * Attach RNG source for this CPU's VM events 813 */ 814 rnd_attach_source(&uvm.cpus[cpu_index(ci)]->rs, 815 ci->ci_data.cpu_name, RND_TYPE_VM, 816 RND_FLAG_COLLECT_TIME|RND_FLAG_COLLECT_VALUE| 817 RND_FLAG_ESTIMATE_VALUE); 818 819 } 820 821 /* 822 * uvm_pagealloc_pgfl: helper routine for uvm_pagealloc_strat 823 */ 824 825 static struct vm_page * 826 uvm_pagealloc_pgfl(struct uvm_cpu *ucpu, int flist, int try1, int try2, 827 int *trycolorp) 828 { 829 struct pgflist *freeq; 830 struct vm_page *pg; 831 int color, trycolor = *trycolorp; 832 struct pgfreelist *gpgfl, *pgfl; 833 834 KASSERT(mutex_owned(&uvm_fpageqlock)); 835 836 color = trycolor; 837 pgfl = &ucpu->page_free[flist]; 838 gpgfl = &uvm.page_free[flist]; 839 do { 840 /* cpu, try1 */ 841 if ((pg = LIST_FIRST((freeq = 842 &pgfl->pgfl_buckets[color].pgfl_queues[try1]))) != NULL) { 843 KASSERT(pg->pqflags & PQ_FREE); 844 KASSERT(try1 == PGFL_ZEROS || !(pg->flags & PG_ZERO)); 845 KASSERT(try1 == PGFL_UNKNOWN || (pg->flags & PG_ZERO)); 846 KASSERT(ucpu == VM_FREE_PAGE_TO_CPU(pg)); 847 VM_FREE_PAGE_TO_CPU(pg)->pages[try1]--; 848 uvmexp.cpuhit++; 849 goto gotit; 850 } 851 /* global, try1 */ 852 if ((pg = LIST_FIRST((freeq = 853 &gpgfl->pgfl_buckets[color].pgfl_queues[try1]))) != NULL) { 854 KASSERT(pg->pqflags & PQ_FREE); 855 KASSERT(try1 == PGFL_ZEROS || !(pg->flags & PG_ZERO)); 856 KASSERT(try1 == PGFL_UNKNOWN || (pg->flags & PG_ZERO)); 857 KASSERT(ucpu != VM_FREE_PAGE_TO_CPU(pg)); 858 VM_FREE_PAGE_TO_CPU(pg)->pages[try1]--; 859 uvmexp.cpumiss++; 860 goto gotit; 861 } 862 /* cpu, try2 */ 863 if ((pg = LIST_FIRST((freeq = 864 &pgfl->pgfl_buckets[color].pgfl_queues[try2]))) != NULL) { 865 KASSERT(pg->pqflags & PQ_FREE); 866 KASSERT(try2 == PGFL_ZEROS || !(pg->flags & PG_ZERO)); 867 KASSERT(try2 == PGFL_UNKNOWN || (pg->flags & PG_ZERO)); 868 KASSERT(ucpu == VM_FREE_PAGE_TO_CPU(pg)); 869 VM_FREE_PAGE_TO_CPU(pg)->pages[try2]--; 870 uvmexp.cpuhit++; 871 goto gotit; 872 } 873 /* global, try2 */ 874 if ((pg = LIST_FIRST((freeq = 875 &gpgfl->pgfl_buckets[color].pgfl_queues[try2]))) != NULL) { 876 KASSERT(pg->pqflags & PQ_FREE); 877 KASSERT(try2 == PGFL_ZEROS || !(pg->flags & PG_ZERO)); 878 KASSERT(try2 == PGFL_UNKNOWN || (pg->flags & PG_ZERO)); 879 KASSERT(ucpu != VM_FREE_PAGE_TO_CPU(pg)); 880 VM_FREE_PAGE_TO_CPU(pg)->pages[try2]--; 881 uvmexp.cpumiss++; 882 goto gotit; 883 } 884 color = (color + 1) & uvmexp.colormask; 885 } while (color != trycolor); 886 887 return (NULL); 888 889 gotit: 890 LIST_REMOVE(pg, pageq.list); /* global list */ 891 LIST_REMOVE(pg, listq.list); /* per-cpu list */ 892 uvmexp.free--; 893 894 /* update zero'd page count */ 895 if (pg->flags & PG_ZERO) 896 uvmexp.zeropages--; 897 898 if (color == trycolor) 899 uvmexp.colorhit++; 900 else { 901 uvmexp.colormiss++; 902 *trycolorp = color; 903 } 904 905 return (pg); 906 } 907 908 /* 909 * uvm_pagealloc_strat: allocate vm_page from a particular free list. 910 * 911 * => return null if no pages free 912 * => wake up pagedaemon if number of free pages drops below low water mark 913 * => if obj != NULL, obj must be locked (to put in obj's tree) 914 * => if anon != NULL, anon must be locked (to put in anon) 915 * => only one of obj or anon can be non-null 916 * => caller must activate/deactivate page if it is not wired. 917 * => free_list is ignored if strat == UVM_PGA_STRAT_NORMAL. 918 * => policy decision: it is more important to pull a page off of the 919 * appropriate priority free list than it is to get a zero'd or 920 * unknown contents page. This is because we live with the 921 * consequences of a bad free list decision for the entire 922 * lifetime of the page, e.g. if the page comes from memory that 923 * is slower to access. 924 */ 925 926 struct vm_page * 927 uvm_pagealloc_strat(struct uvm_object *obj, voff_t off, struct vm_anon *anon, 928 int flags, int strat, int free_list) 929 { 930 int try1, try2, zeroit = 0, color; 931 int lcv; 932 struct uvm_cpu *ucpu; 933 struct vm_page *pg; 934 lwp_t *l; 935 936 KASSERT(obj == NULL || anon == NULL); 937 KASSERT(anon == NULL || (flags & UVM_FLAG_COLORMATCH) || off == 0); 938 KASSERT(off == trunc_page(off)); 939 KASSERT(obj == NULL || mutex_owned(obj->vmobjlock)); 940 KASSERT(anon == NULL || anon->an_lock == NULL || 941 mutex_owned(anon->an_lock)); 942 943 mutex_spin_enter(&uvm_fpageqlock); 944 945 /* 946 * This implements a global round-robin page coloring 947 * algorithm. 948 */ 949 950 ucpu = curcpu()->ci_data.cpu_uvm; 951 if (flags & UVM_FLAG_COLORMATCH) { 952 color = atop(off) & uvmexp.colormask; 953 } else { 954 color = ucpu->page_free_nextcolor; 955 } 956 957 /* 958 * check to see if we need to generate some free pages waking 959 * the pagedaemon. 960 */ 961 962 uvm_kick_pdaemon(); 963 964 /* 965 * fail if any of these conditions is true: 966 * [1] there really are no free pages, or 967 * [2] only kernel "reserved" pages remain and 968 * reserved pages have not been requested. 969 * [3] only pagedaemon "reserved" pages remain and 970 * the requestor isn't the pagedaemon. 971 * we make kernel reserve pages available if called by a 972 * kernel thread or a realtime thread. 973 */ 974 l = curlwp; 975 if (__predict_true(l != NULL) && lwp_eprio(l) >= PRI_KTHREAD) { 976 flags |= UVM_PGA_USERESERVE; 977 } 978 if ((uvmexp.free <= uvmexp.reserve_kernel && 979 (flags & UVM_PGA_USERESERVE) == 0) || 980 (uvmexp.free <= uvmexp.reserve_pagedaemon && 981 curlwp != uvm.pagedaemon_lwp)) 982 goto fail; 983 984 #if PGFL_NQUEUES != 2 985 #error uvm_pagealloc_strat needs to be updated 986 #endif 987 988 /* 989 * If we want a zero'd page, try the ZEROS queue first, otherwise 990 * we try the UNKNOWN queue first. 991 */ 992 if (flags & UVM_PGA_ZERO) { 993 try1 = PGFL_ZEROS; 994 try2 = PGFL_UNKNOWN; 995 } else { 996 try1 = PGFL_UNKNOWN; 997 try2 = PGFL_ZEROS; 998 } 999 1000 again: 1001 switch (strat) { 1002 case UVM_PGA_STRAT_NORMAL: 1003 /* Check freelists: descending priority (ascending id) order */ 1004 for (lcv = 0; lcv < VM_NFREELIST; lcv++) { 1005 pg = uvm_pagealloc_pgfl(ucpu, lcv, 1006 try1, try2, &color); 1007 if (pg != NULL) 1008 goto gotit; 1009 } 1010 1011 /* No pages free! */ 1012 goto fail; 1013 1014 case UVM_PGA_STRAT_ONLY: 1015 case UVM_PGA_STRAT_FALLBACK: 1016 /* Attempt to allocate from the specified free list. */ 1017 KASSERT(free_list >= 0 && free_list < VM_NFREELIST); 1018 pg = uvm_pagealloc_pgfl(ucpu, free_list, 1019 try1, try2, &color); 1020 if (pg != NULL) 1021 goto gotit; 1022 1023 /* Fall back, if possible. */ 1024 if (strat == UVM_PGA_STRAT_FALLBACK) { 1025 strat = UVM_PGA_STRAT_NORMAL; 1026 goto again; 1027 } 1028 1029 /* No pages free! */ 1030 goto fail; 1031 1032 default: 1033 panic("uvm_pagealloc_strat: bad strat %d", strat); 1034 /* NOTREACHED */ 1035 } 1036 1037 gotit: 1038 /* 1039 * We now know which color we actually allocated from; set 1040 * the next color accordingly. 1041 */ 1042 1043 ucpu->page_free_nextcolor = (color + 1) & uvmexp.colormask; 1044 1045 /* 1046 * update allocation statistics and remember if we have to 1047 * zero the page 1048 */ 1049 1050 if (flags & UVM_PGA_ZERO) { 1051 if (pg->flags & PG_ZERO) { 1052 uvmexp.pga_zerohit++; 1053 zeroit = 0; 1054 } else { 1055 uvmexp.pga_zeromiss++; 1056 zeroit = 1; 1057 } 1058 if (ucpu->pages[PGFL_ZEROS] < ucpu->pages[PGFL_UNKNOWN]) { 1059 ucpu->page_idle_zero = vm_page_zero_enable; 1060 } 1061 } 1062 KASSERT(pg->pqflags == PQ_FREE); 1063 1064 pg->offset = off; 1065 pg->uobject = obj; 1066 pg->uanon = anon; 1067 pg->flags = PG_BUSY|PG_CLEAN|PG_FAKE; 1068 if (anon) { 1069 anon->an_page = pg; 1070 pg->pqflags = PQ_ANON; 1071 atomic_inc_uint(&uvmexp.anonpages); 1072 } else { 1073 if (obj) { 1074 uvm_pageinsert(obj, pg); 1075 } 1076 pg->pqflags = 0; 1077 } 1078 mutex_spin_exit(&uvm_fpageqlock); 1079 1080 #if defined(UVM_PAGE_TRKOWN) 1081 pg->owner_tag = NULL; 1082 #endif 1083 UVM_PAGE_OWN(pg, "new alloc"); 1084 1085 if (flags & UVM_PGA_ZERO) { 1086 /* 1087 * A zero'd page is not clean. If we got a page not already 1088 * zero'd, then we have to zero it ourselves. 1089 */ 1090 pg->flags &= ~PG_CLEAN; 1091 if (zeroit) 1092 pmap_zero_page(VM_PAGE_TO_PHYS(pg)); 1093 } 1094 1095 return(pg); 1096 1097 fail: 1098 mutex_spin_exit(&uvm_fpageqlock); 1099 return (NULL); 1100 } 1101 1102 /* 1103 * uvm_pagereplace: replace a page with another 1104 * 1105 * => object must be locked 1106 */ 1107 1108 void 1109 uvm_pagereplace(struct vm_page *oldpg, struct vm_page *newpg) 1110 { 1111 struct uvm_object *uobj = oldpg->uobject; 1112 1113 KASSERT((oldpg->flags & PG_TABLED) != 0); 1114 KASSERT(uobj != NULL); 1115 KASSERT((newpg->flags & PG_TABLED) == 0); 1116 KASSERT(newpg->uobject == NULL); 1117 KASSERT(mutex_owned(uobj->vmobjlock)); 1118 1119 newpg->uobject = uobj; 1120 newpg->offset = oldpg->offset; 1121 1122 uvm_pageremove_tree(uobj, oldpg); 1123 uvm_pageinsert_tree(uobj, newpg); 1124 uvm_pageinsert_list(uobj, newpg, oldpg); 1125 uvm_pageremove_list(uobj, oldpg); 1126 } 1127 1128 /* 1129 * uvm_pagerealloc: reallocate a page from one object to another 1130 * 1131 * => both objects must be locked 1132 */ 1133 1134 void 1135 uvm_pagerealloc(struct vm_page *pg, struct uvm_object *newobj, voff_t newoff) 1136 { 1137 /* 1138 * remove it from the old object 1139 */ 1140 1141 if (pg->uobject) { 1142 uvm_pageremove(pg->uobject, pg); 1143 } 1144 1145 /* 1146 * put it in the new object 1147 */ 1148 1149 if (newobj) { 1150 pg->uobject = newobj; 1151 pg->offset = newoff; 1152 uvm_pageinsert(newobj, pg); 1153 } 1154 } 1155 1156 #ifdef DEBUG 1157 /* 1158 * check if page is zero-filled 1159 * 1160 * - called with free page queue lock held. 1161 */ 1162 void 1163 uvm_pagezerocheck(struct vm_page *pg) 1164 { 1165 int *p, *ep; 1166 1167 KASSERT(uvm_zerocheckkva != 0); 1168 KASSERT(mutex_owned(&uvm_fpageqlock)); 1169 1170 /* 1171 * XXX assuming pmap_kenter_pa and pmap_kremove never call 1172 * uvm page allocator. 1173 * 1174 * it might be better to have "CPU-local temporary map" pmap interface. 1175 */ 1176 pmap_kenter_pa(uvm_zerocheckkva, VM_PAGE_TO_PHYS(pg), VM_PROT_READ, 0); 1177 p = (int *)uvm_zerocheckkva; 1178 ep = (int *)((char *)p + PAGE_SIZE); 1179 pmap_update(pmap_kernel()); 1180 while (p < ep) { 1181 if (*p != 0) 1182 panic("PG_ZERO page isn't zero-filled"); 1183 p++; 1184 } 1185 pmap_kremove(uvm_zerocheckkva, PAGE_SIZE); 1186 /* 1187 * pmap_update() is not necessary here because no one except us 1188 * uses this VA. 1189 */ 1190 } 1191 #endif /* DEBUG */ 1192 1193 /* 1194 * uvm_pagefree: free page 1195 * 1196 * => erase page's identity (i.e. remove from object) 1197 * => put page on free list 1198 * => caller must lock owning object (either anon or uvm_object) 1199 * => caller must lock page queues 1200 * => assumes all valid mappings of pg are gone 1201 */ 1202 1203 void 1204 uvm_pagefree(struct vm_page *pg) 1205 { 1206 struct pgflist *pgfl; 1207 struct uvm_cpu *ucpu; 1208 int index, color, queue; 1209 bool iszero; 1210 1211 #ifdef DEBUG 1212 if (pg->uobject == (void *)0xdeadbeef && 1213 pg->uanon == (void *)0xdeadbeef) { 1214 panic("uvm_pagefree: freeing free page %p", pg); 1215 } 1216 #endif /* DEBUG */ 1217 1218 KASSERT((pg->flags & PG_PAGEOUT) == 0); 1219 KASSERT(!(pg->pqflags & PQ_FREE)); 1220 //KASSERT(mutex_owned(&uvm_pageqlock) || !uvmpdpol_pageisqueued_p(pg)); 1221 KASSERT(pg->uobject == NULL || mutex_owned(pg->uobject->vmobjlock)); 1222 KASSERT(pg->uobject != NULL || pg->uanon == NULL || 1223 mutex_owned(pg->uanon->an_lock)); 1224 1225 /* 1226 * if the page is loaned, resolve the loan instead of freeing. 1227 */ 1228 1229 if (pg->loan_count) { 1230 KASSERT(pg->wire_count == 0); 1231 1232 /* 1233 * if the page is owned by an anon then we just want to 1234 * drop anon ownership. the kernel will free the page when 1235 * it is done with it. if the page is owned by an object, 1236 * remove it from the object and mark it dirty for the benefit 1237 * of possible anon owners. 1238 * 1239 * regardless of previous ownership, wakeup any waiters, 1240 * unbusy the page, and we're done. 1241 */ 1242 1243 if (pg->uobject != NULL) { 1244 uvm_pageremove(pg->uobject, pg); 1245 pg->flags &= ~PG_CLEAN; 1246 } else if (pg->uanon != NULL) { 1247 if ((pg->pqflags & PQ_ANON) == 0) { 1248 pg->loan_count--; 1249 } else { 1250 pg->pqflags &= ~PQ_ANON; 1251 atomic_dec_uint(&uvmexp.anonpages); 1252 } 1253 pg->uanon->an_page = NULL; 1254 pg->uanon = NULL; 1255 } 1256 if (pg->flags & PG_WANTED) { 1257 wakeup(pg); 1258 } 1259 pg->flags &= ~(PG_WANTED|PG_BUSY|PG_RELEASED|PG_PAGER1); 1260 #ifdef UVM_PAGE_TRKOWN 1261 pg->owner_tag = NULL; 1262 #endif 1263 if (pg->loan_count) { 1264 KASSERT(pg->uobject == NULL); 1265 if (pg->uanon == NULL) { 1266 KASSERT(mutex_owned(&uvm_pageqlock)); 1267 uvm_pagedequeue(pg); 1268 } 1269 return; 1270 } 1271 } 1272 1273 /* 1274 * remove page from its object or anon. 1275 */ 1276 1277 if (pg->uobject != NULL) { 1278 uvm_pageremove(pg->uobject, pg); 1279 } else if (pg->uanon != NULL) { 1280 pg->uanon->an_page = NULL; 1281 atomic_dec_uint(&uvmexp.anonpages); 1282 } 1283 1284 /* 1285 * now remove the page from the queues. 1286 */ 1287 if (uvmpdpol_pageisqueued_p(pg)) { 1288 KASSERT(mutex_owned(&uvm_pageqlock)); 1289 uvm_pagedequeue(pg); 1290 } 1291 1292 /* 1293 * if the page was wired, unwire it now. 1294 */ 1295 1296 if (pg->wire_count) { 1297 pg->wire_count = 0; 1298 uvmexp.wired--; 1299 } 1300 1301 /* 1302 * and put on free queue 1303 */ 1304 1305 iszero = (pg->flags & PG_ZERO); 1306 index = uvm_page_lookup_freelist(pg); 1307 color = VM_PGCOLOR_BUCKET(pg); 1308 queue = (iszero ? PGFL_ZEROS : PGFL_UNKNOWN); 1309 1310 #ifdef DEBUG 1311 pg->uobject = (void *)0xdeadbeef; 1312 pg->uanon = (void *)0xdeadbeef; 1313 #endif 1314 1315 mutex_spin_enter(&uvm_fpageqlock); 1316 pg->pqflags = PQ_FREE; 1317 1318 #ifdef DEBUG 1319 if (iszero) 1320 uvm_pagezerocheck(pg); 1321 #endif /* DEBUG */ 1322 1323 1324 /* global list */ 1325 pgfl = &uvm.page_free[index].pgfl_buckets[color].pgfl_queues[queue]; 1326 LIST_INSERT_HEAD(pgfl, pg, pageq.list); 1327 uvmexp.free++; 1328 if (iszero) { 1329 uvmexp.zeropages++; 1330 } 1331 1332 /* per-cpu list */ 1333 ucpu = curcpu()->ci_data.cpu_uvm; 1334 pg->offset = (uintptr_t)ucpu; 1335 pgfl = &ucpu->page_free[index].pgfl_buckets[color].pgfl_queues[queue]; 1336 LIST_INSERT_HEAD(pgfl, pg, listq.list); 1337 ucpu->pages[queue]++; 1338 if (ucpu->pages[PGFL_ZEROS] < ucpu->pages[PGFL_UNKNOWN]) { 1339 ucpu->page_idle_zero = vm_page_zero_enable; 1340 } 1341 1342 mutex_spin_exit(&uvm_fpageqlock); 1343 } 1344 1345 /* 1346 * uvm_page_unbusy: unbusy an array of pages. 1347 * 1348 * => pages must either all belong to the same object, or all belong to anons. 1349 * => if pages are object-owned, object must be locked. 1350 * => if pages are anon-owned, anons must be locked. 1351 * => caller must lock page queues if pages may be released. 1352 * => caller must make sure that anon-owned pages are not PG_RELEASED. 1353 */ 1354 1355 void 1356 uvm_page_unbusy(struct vm_page **pgs, int npgs) 1357 { 1358 struct vm_page *pg; 1359 int i; 1360 UVMHIST_FUNC("uvm_page_unbusy"); UVMHIST_CALLED(ubchist); 1361 1362 for (i = 0; i < npgs; i++) { 1363 pg = pgs[i]; 1364 if (pg == NULL || pg == PGO_DONTCARE) { 1365 continue; 1366 } 1367 1368 KASSERT(uvm_page_locked_p(pg)); 1369 KASSERT(pg->flags & PG_BUSY); 1370 KASSERT((pg->flags & PG_PAGEOUT) == 0); 1371 if (pg->flags & PG_WANTED) { 1372 wakeup(pg); 1373 } 1374 if (pg->flags & PG_RELEASED) { 1375 UVMHIST_LOG(ubchist, "releasing pg %#jx", 1376 (uintptr_t)pg, 0, 0, 0); 1377 KASSERT(pg->uobject != NULL || 1378 (pg->uanon != NULL && pg->uanon->an_ref > 0)); 1379 pg->flags &= ~PG_RELEASED; 1380 uvm_pagefree(pg); 1381 } else { 1382 UVMHIST_LOG(ubchist, "unbusying pg %#jx", 1383 (uintptr_t)pg, 0, 0, 0); 1384 KASSERT((pg->flags & PG_FAKE) == 0); 1385 pg->flags &= ~(PG_WANTED|PG_BUSY); 1386 UVM_PAGE_OWN(pg, NULL); 1387 } 1388 } 1389 } 1390 1391 #if defined(UVM_PAGE_TRKOWN) 1392 /* 1393 * uvm_page_own: set or release page ownership 1394 * 1395 * => this is a debugging function that keeps track of who sets PG_BUSY 1396 * and where they do it. it can be used to track down problems 1397 * such a process setting "PG_BUSY" and never releasing it. 1398 * => page's object [if any] must be locked 1399 * => if "tag" is NULL then we are releasing page ownership 1400 */ 1401 void 1402 uvm_page_own(struct vm_page *pg, const char *tag) 1403 { 1404 1405 KASSERT((pg->flags & (PG_PAGEOUT|PG_RELEASED)) == 0); 1406 KASSERT((pg->flags & PG_WANTED) == 0); 1407 KASSERT(uvm_page_locked_p(pg)); 1408 1409 /* gain ownership? */ 1410 if (tag) { 1411 KASSERT((pg->flags & PG_BUSY) != 0); 1412 if (pg->owner_tag) { 1413 printf("uvm_page_own: page %p already owned " 1414 "by proc %d [%s]\n", pg, 1415 pg->owner, pg->owner_tag); 1416 panic("uvm_page_own"); 1417 } 1418 pg->owner = curproc->p_pid; 1419 pg->lowner = curlwp->l_lid; 1420 pg->owner_tag = tag; 1421 return; 1422 } 1423 1424 /* drop ownership */ 1425 KASSERT((pg->flags & PG_BUSY) == 0); 1426 if (pg->owner_tag == NULL) { 1427 printf("uvm_page_own: dropping ownership of an non-owned " 1428 "page (%p)\n", pg); 1429 panic("uvm_page_own"); 1430 } 1431 if (!uvmpdpol_pageisqueued_p(pg)) { 1432 KASSERT((pg->uanon == NULL && pg->uobject == NULL) || 1433 pg->wire_count > 0); 1434 } else { 1435 KASSERT(pg->wire_count == 0); 1436 } 1437 pg->owner_tag = NULL; 1438 } 1439 #endif 1440 1441 /* 1442 * uvm_pageidlezero: zero free pages while the system is idle. 1443 * 1444 * => try to complete one color bucket at a time, to reduce our impact 1445 * on the CPU cache. 1446 * => we loop until we either reach the target or there is a lwp ready 1447 * to run, or MD code detects a reason to break early. 1448 */ 1449 void 1450 uvm_pageidlezero(void) 1451 { 1452 struct vm_page *pg; 1453 struct pgfreelist *pgfl, *gpgfl; 1454 struct uvm_cpu *ucpu; 1455 int free_list, firstbucket, nextbucket; 1456 bool lcont = false; 1457 1458 ucpu = curcpu()->ci_data.cpu_uvm; 1459 if (!ucpu->page_idle_zero || 1460 ucpu->pages[PGFL_UNKNOWN] < uvmexp.ncolors) { 1461 ucpu->page_idle_zero = false; 1462 return; 1463 } 1464 if (!mutex_tryenter(&uvm_fpageqlock)) { 1465 /* Contention: let other CPUs to use the lock. */ 1466 return; 1467 } 1468 firstbucket = ucpu->page_free_nextcolor; 1469 nextbucket = firstbucket; 1470 do { 1471 for (free_list = 0; free_list < VM_NFREELIST; free_list++) { 1472 if (sched_curcpu_runnable_p()) { 1473 goto quit; 1474 } 1475 pgfl = &ucpu->page_free[free_list]; 1476 gpgfl = &uvm.page_free[free_list]; 1477 while ((pg = LIST_FIRST(&pgfl->pgfl_buckets[ 1478 nextbucket].pgfl_queues[PGFL_UNKNOWN])) != NULL) { 1479 if (lcont || sched_curcpu_runnable_p()) { 1480 goto quit; 1481 } 1482 LIST_REMOVE(pg, pageq.list); /* global list */ 1483 LIST_REMOVE(pg, listq.list); /* per-cpu list */ 1484 ucpu->pages[PGFL_UNKNOWN]--; 1485 uvmexp.free--; 1486 KASSERT(pg->pqflags == PQ_FREE); 1487 pg->pqflags = 0; 1488 mutex_spin_exit(&uvm_fpageqlock); 1489 #ifdef PMAP_PAGEIDLEZERO 1490 if (!PMAP_PAGEIDLEZERO(VM_PAGE_TO_PHYS(pg))) { 1491 1492 /* 1493 * The machine-dependent code detected 1494 * some reason for us to abort zeroing 1495 * pages, probably because there is a 1496 * process now ready to run. 1497 */ 1498 1499 mutex_spin_enter(&uvm_fpageqlock); 1500 pg->pqflags = PQ_FREE; 1501 LIST_INSERT_HEAD(&gpgfl->pgfl_buckets[ 1502 nextbucket].pgfl_queues[ 1503 PGFL_UNKNOWN], pg, pageq.list); 1504 LIST_INSERT_HEAD(&pgfl->pgfl_buckets[ 1505 nextbucket].pgfl_queues[ 1506 PGFL_UNKNOWN], pg, listq.list); 1507 ucpu->pages[PGFL_UNKNOWN]++; 1508 uvmexp.free++; 1509 uvmexp.zeroaborts++; 1510 goto quit; 1511 } 1512 #else 1513 pmap_zero_page(VM_PAGE_TO_PHYS(pg)); 1514 #endif /* PMAP_PAGEIDLEZERO */ 1515 pg->flags |= PG_ZERO; 1516 1517 if (!mutex_tryenter(&uvm_fpageqlock)) { 1518 lcont = true; 1519 mutex_spin_enter(&uvm_fpageqlock); 1520 } else { 1521 lcont = false; 1522 } 1523 pg->pqflags = PQ_FREE; 1524 LIST_INSERT_HEAD(&gpgfl->pgfl_buckets[ 1525 nextbucket].pgfl_queues[PGFL_ZEROS], 1526 pg, pageq.list); 1527 LIST_INSERT_HEAD(&pgfl->pgfl_buckets[ 1528 nextbucket].pgfl_queues[PGFL_ZEROS], 1529 pg, listq.list); 1530 ucpu->pages[PGFL_ZEROS]++; 1531 uvmexp.free++; 1532 uvmexp.zeropages++; 1533 } 1534 } 1535 if (ucpu->pages[PGFL_UNKNOWN] < uvmexp.ncolors) { 1536 break; 1537 } 1538 nextbucket = (nextbucket + 1) & uvmexp.colormask; 1539 } while (nextbucket != firstbucket); 1540 ucpu->page_idle_zero = false; 1541 quit: 1542 mutex_spin_exit(&uvm_fpageqlock); 1543 } 1544 1545 /* 1546 * uvm_pagelookup: look up a page 1547 * 1548 * => caller should lock object to keep someone from pulling the page 1549 * out from under it 1550 */ 1551 1552 struct vm_page * 1553 uvm_pagelookup(struct uvm_object *obj, voff_t off) 1554 { 1555 struct vm_page *pg; 1556 1557 KASSERT(mutex_owned(obj->vmobjlock)); 1558 1559 pg = rb_tree_find_node(&obj->rb_tree, &off); 1560 1561 KASSERT(pg == NULL || obj->uo_npages != 0); 1562 KASSERT(pg == NULL || (pg->flags & (PG_RELEASED|PG_PAGEOUT)) == 0 || 1563 (pg->flags & PG_BUSY) != 0); 1564 return pg; 1565 } 1566 1567 /* 1568 * uvm_pagewire: wire the page, thus removing it from the daemon's grasp 1569 * 1570 * => caller must lock page queues 1571 */ 1572 1573 void 1574 uvm_pagewire(struct vm_page *pg) 1575 { 1576 KASSERT(mutex_owned(&uvm_pageqlock)); 1577 #if defined(READAHEAD_STATS) 1578 if ((pg->pqflags & PQ_READAHEAD) != 0) { 1579 uvm_ra_hit.ev_count++; 1580 pg->pqflags &= ~PQ_READAHEAD; 1581 } 1582 #endif /* defined(READAHEAD_STATS) */ 1583 if (pg->wire_count == 0) { 1584 uvm_pagedequeue(pg); 1585 uvmexp.wired++; 1586 } 1587 pg->wire_count++; 1588 KASSERT(pg->wire_count > 0); /* detect wraparound */ 1589 } 1590 1591 /* 1592 * uvm_pageunwire: unwire the page. 1593 * 1594 * => activate if wire count goes to zero. 1595 * => caller must lock page queues 1596 */ 1597 1598 void 1599 uvm_pageunwire(struct vm_page *pg) 1600 { 1601 KASSERT(mutex_owned(&uvm_pageqlock)); 1602 KASSERT(pg->wire_count != 0); 1603 pg->wire_count--; 1604 if (pg->wire_count == 0) { 1605 uvm_pageactivate(pg); 1606 KASSERT(uvmexp.wired != 0); 1607 uvmexp.wired--; 1608 } 1609 } 1610 1611 /* 1612 * uvm_pagedeactivate: deactivate page 1613 * 1614 * => caller must lock page queues 1615 * => caller must check to make sure page is not wired 1616 * => object that page belongs to must be locked (so we can adjust pg->flags) 1617 * => caller must clear the reference on the page before calling 1618 */ 1619 1620 void 1621 uvm_pagedeactivate(struct vm_page *pg) 1622 { 1623 1624 KASSERT(mutex_owned(&uvm_pageqlock)); 1625 KASSERT(uvm_page_locked_p(pg)); 1626 KASSERT(pg->wire_count != 0 || uvmpdpol_pageisqueued_p(pg)); 1627 uvmpdpol_pagedeactivate(pg); 1628 } 1629 1630 /* 1631 * uvm_pageactivate: activate page 1632 * 1633 * => caller must lock page queues 1634 */ 1635 1636 void 1637 uvm_pageactivate(struct vm_page *pg) 1638 { 1639 1640 KASSERT(mutex_owned(&uvm_pageqlock)); 1641 KASSERT(uvm_page_locked_p(pg)); 1642 #if defined(READAHEAD_STATS) 1643 if ((pg->pqflags & PQ_READAHEAD) != 0) { 1644 uvm_ra_hit.ev_count++; 1645 pg->pqflags &= ~PQ_READAHEAD; 1646 } 1647 #endif /* defined(READAHEAD_STATS) */ 1648 if (pg->wire_count != 0) { 1649 return; 1650 } 1651 uvmpdpol_pageactivate(pg); 1652 } 1653 1654 /* 1655 * uvm_pagedequeue: remove a page from any paging queue 1656 */ 1657 1658 void 1659 uvm_pagedequeue(struct vm_page *pg) 1660 { 1661 1662 if (uvmpdpol_pageisqueued_p(pg)) { 1663 KASSERT(mutex_owned(&uvm_pageqlock)); 1664 } 1665 1666 uvmpdpol_pagedequeue(pg); 1667 } 1668 1669 /* 1670 * uvm_pageenqueue: add a page to a paging queue without activating. 1671 * used where a page is not really demanded (yet). eg. read-ahead 1672 */ 1673 1674 void 1675 uvm_pageenqueue(struct vm_page *pg) 1676 { 1677 1678 KASSERT(mutex_owned(&uvm_pageqlock)); 1679 if (pg->wire_count != 0) { 1680 return; 1681 } 1682 uvmpdpol_pageenqueue(pg); 1683 } 1684 1685 /* 1686 * uvm_pagezero: zero fill a page 1687 * 1688 * => if page is part of an object then the object should be locked 1689 * to protect pg->flags. 1690 */ 1691 1692 void 1693 uvm_pagezero(struct vm_page *pg) 1694 { 1695 pg->flags &= ~PG_CLEAN; 1696 pmap_zero_page(VM_PAGE_TO_PHYS(pg)); 1697 } 1698 1699 /* 1700 * uvm_pagecopy: copy a page 1701 * 1702 * => if page is part of an object then the object should be locked 1703 * to protect pg->flags. 1704 */ 1705 1706 void 1707 uvm_pagecopy(struct vm_page *src, struct vm_page *dst) 1708 { 1709 1710 dst->flags &= ~PG_CLEAN; 1711 pmap_copy_page(VM_PAGE_TO_PHYS(src), VM_PAGE_TO_PHYS(dst)); 1712 } 1713 1714 /* 1715 * uvm_pageismanaged: test it see that a page (specified by PA) is managed. 1716 */ 1717 1718 bool 1719 uvm_pageismanaged(paddr_t pa) 1720 { 1721 1722 return (uvm_physseg_find(atop(pa), NULL) != UVM_PHYSSEG_TYPE_INVALID); 1723 } 1724 1725 /* 1726 * uvm_page_lookup_freelist: look up the free list for the specified page 1727 */ 1728 1729 int 1730 uvm_page_lookup_freelist(struct vm_page *pg) 1731 { 1732 uvm_physseg_t upm; 1733 1734 upm = uvm_physseg_find(atop(VM_PAGE_TO_PHYS(pg)), NULL); 1735 KASSERT(upm != UVM_PHYSSEG_TYPE_INVALID); 1736 return uvm_physseg_get_free_list(upm); 1737 } 1738 1739 /* 1740 * uvm_page_locked_p: return true if object associated with page is 1741 * locked. this is a weak check for runtime assertions only. 1742 */ 1743 1744 bool 1745 uvm_page_locked_p(struct vm_page *pg) 1746 { 1747 1748 if (pg->uobject != NULL) { 1749 return mutex_owned(pg->uobject->vmobjlock); 1750 } 1751 if (pg->uanon != NULL) { 1752 return mutex_owned(pg->uanon->an_lock); 1753 } 1754 return true; 1755 } 1756 1757 #ifdef PMAP_DIRECT 1758 /* 1759 * Call pmap to translate physical address into a virtual and to run a callback 1760 * for it. Used to avoid actually mapping the pages, pmap most likely uses direct map 1761 * or equivalent. 1762 */ 1763 int 1764 uvm_direct_process(struct vm_page **pgs, u_int npages, voff_t off, vsize_t len, 1765 int (*process)(void *, size_t, void *), void *arg) 1766 { 1767 int error = 0; 1768 paddr_t pa; 1769 size_t todo; 1770 voff_t pgoff = (off & PAGE_MASK); 1771 struct vm_page *pg; 1772 1773 KASSERT(npages > 0 && len > 0); 1774 1775 for (int i = 0; i < npages; i++) { 1776 pg = pgs[i]; 1777 1778 KASSERT(len > 0); 1779 1780 /* 1781 * Caller is responsible for ensuring all the pages are 1782 * available. 1783 */ 1784 KASSERT(pg != NULL && pg != PGO_DONTCARE); 1785 1786 pa = VM_PAGE_TO_PHYS(pg); 1787 todo = MIN(len, PAGE_SIZE - pgoff); 1788 1789 error = pmap_direct_process(pa, pgoff, todo, process, arg); 1790 if (error) 1791 break; 1792 1793 pgoff = 0; 1794 len -= todo; 1795 } 1796 1797 KASSERTMSG(error != 0 || len == 0, "len %lu != 0 for non-error", len); 1798 return error; 1799 } 1800 #endif /* PMAP_DIRECT */ 1801 1802 #if defined(DDB) || defined(DEBUGPRINT) 1803 1804 /* 1805 * uvm_page_printit: actually print the page 1806 */ 1807 1808 static const char page_flagbits[] = UVM_PGFLAGBITS; 1809 static const char page_pqflagbits[] = UVM_PQFLAGBITS; 1810 1811 void 1812 uvm_page_printit(struct vm_page *pg, bool full, 1813 void (*pr)(const char *, ...)) 1814 { 1815 struct vm_page *tpg; 1816 struct uvm_object *uobj; 1817 struct pgflist *pgl; 1818 char pgbuf[128]; 1819 char pqbuf[128]; 1820 1821 (*pr)("PAGE %p:\n", pg); 1822 snprintb(pgbuf, sizeof(pgbuf), page_flagbits, pg->flags); 1823 snprintb(pqbuf, sizeof(pqbuf), page_pqflagbits, pg->pqflags); 1824 (*pr)(" flags=%s, pqflags=%s, wire_count=%d, pa=0x%lx\n", 1825 pgbuf, pqbuf, pg->wire_count, (long)VM_PAGE_TO_PHYS(pg)); 1826 (*pr)(" uobject=%p, uanon=%p, offset=0x%llx loan_count=%d\n", 1827 pg->uobject, pg->uanon, (long long)pg->offset, pg->loan_count); 1828 #if defined(UVM_PAGE_TRKOWN) 1829 if (pg->flags & PG_BUSY) 1830 (*pr)(" owning process = %d, tag=%s\n", 1831 pg->owner, pg->owner_tag); 1832 else 1833 (*pr)(" page not busy, no owner\n"); 1834 #else 1835 (*pr)(" [page ownership tracking disabled]\n"); 1836 #endif 1837 1838 if (!full) 1839 return; 1840 1841 /* cross-verify object/anon */ 1842 if ((pg->pqflags & PQ_FREE) == 0) { 1843 if (pg->pqflags & PQ_ANON) { 1844 if (pg->uanon == NULL || pg->uanon->an_page != pg) 1845 (*pr)(" >>> ANON DOES NOT POINT HERE <<< (%p)\n", 1846 (pg->uanon) ? pg->uanon->an_page : NULL); 1847 else 1848 (*pr)(" anon backpointer is OK\n"); 1849 } else { 1850 uobj = pg->uobject; 1851 if (uobj) { 1852 (*pr)(" checking object list\n"); 1853 TAILQ_FOREACH(tpg, &uobj->memq, listq.queue) { 1854 if (tpg == pg) { 1855 break; 1856 } 1857 } 1858 if (tpg) 1859 (*pr)(" page found on object list\n"); 1860 else 1861 (*pr)(" >>> PAGE NOT FOUND ON OBJECT LIST! <<<\n"); 1862 } 1863 } 1864 } 1865 1866 /* cross-verify page queue */ 1867 if (pg->pqflags & PQ_FREE) { 1868 int fl = uvm_page_lookup_freelist(pg); 1869 int color = VM_PGCOLOR_BUCKET(pg); 1870 pgl = &uvm.page_free[fl].pgfl_buckets[color].pgfl_queues[ 1871 ((pg)->flags & PG_ZERO) ? PGFL_ZEROS : PGFL_UNKNOWN]; 1872 } else { 1873 pgl = NULL; 1874 } 1875 1876 if (pgl) { 1877 (*pr)(" checking pageq list\n"); 1878 LIST_FOREACH(tpg, pgl, pageq.list) { 1879 if (tpg == pg) { 1880 break; 1881 } 1882 } 1883 if (tpg) 1884 (*pr)(" page found on pageq list\n"); 1885 else 1886 (*pr)(" >>> PAGE NOT FOUND ON PAGEQ LIST! <<<\n"); 1887 } 1888 } 1889 1890 /* 1891 * uvm_pages_printthem - print a summary of all managed pages 1892 */ 1893 1894 void 1895 uvm_page_printall(void (*pr)(const char *, ...)) 1896 { 1897 uvm_physseg_t i; 1898 paddr_t pfn; 1899 struct vm_page *pg; 1900 1901 (*pr)("%18s %4s %4s %18s %18s" 1902 #ifdef UVM_PAGE_TRKOWN 1903 " OWNER" 1904 #endif 1905 "\n", "PAGE", "FLAG", "PQ", "UOBJECT", "UANON"); 1906 for (i = uvm_physseg_get_first(); 1907 uvm_physseg_valid_p(i); 1908 i = uvm_physseg_get_next(i)) { 1909 for (pfn = uvm_physseg_get_start(i); 1910 pfn < uvm_physseg_get_end(i); 1911 pfn++) { 1912 pg = PHYS_TO_VM_PAGE(ptoa(pfn)); 1913 1914 (*pr)("%18p %04x %04x %18p %18p", 1915 pg, pg->flags, pg->pqflags, pg->uobject, 1916 pg->uanon); 1917 #ifdef UVM_PAGE_TRKOWN 1918 if (pg->flags & PG_BUSY) 1919 (*pr)(" %d [%s]", pg->owner, pg->owner_tag); 1920 #endif 1921 (*pr)("\n"); 1922 } 1923 } 1924 } 1925 1926 #endif /* DDB || DEBUGPRINT */ 1927