1 /* 2 * Copyright (c) 1991 Regents of the University of California. 3 * All rights reserved. 4 * Copyright (c) 2003-2011 The DragonFly Project. All rights reserved. 5 * 6 * This code is derived from software contributed to Berkeley by 7 * The Mach Operating System project at Carnegie-Mellon University. 8 * 9 * This code is derived from software contributed to The DragonFly Project 10 * by Matthew Dillon <dillon@backplane.com> 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 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91 37 * $FreeBSD: src/sys/vm/vm_page.c,v 1.147.2.18 2002/03/10 05:03:19 alc Exp $ 38 */ 39 40 /* 41 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 42 * All rights reserved. 43 * 44 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 45 * 46 * Permission to use, copy, modify and distribute this software and 47 * its documentation is hereby granted, provided that both the copyright 48 * notice and this permission notice appear in all copies of the 49 * software, derivative works or modified versions, and any portions 50 * thereof, and that both notices appear in supporting documentation. 51 * 52 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 53 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 54 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 55 * 56 * Carnegie Mellon requests users of this software to return to 57 * 58 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 59 * School of Computer Science 60 * Carnegie Mellon University 61 * Pittsburgh PA 15213-3890 62 * 63 * any improvements or extensions that they make and grant Carnegie the 64 * rights to redistribute these changes. 65 */ 66 /* 67 * Resident memory management module. The module manipulates 'VM pages'. 68 * A VM page is the core building block for memory management. 69 */ 70 71 #include <sys/param.h> 72 #include <sys/systm.h> 73 #include <sys/malloc.h> 74 #include <sys/proc.h> 75 #include <sys/vmmeter.h> 76 #include <sys/vnode.h> 77 #include <sys/kernel.h> 78 #include <sys/alist.h> 79 #include <sys/sysctl.h> 80 #include <sys/cpu_topology.h> 81 82 #include <vm/vm.h> 83 #include <vm/vm_param.h> 84 #include <sys/lock.h> 85 #include <vm/vm_kern.h> 86 #include <vm/pmap.h> 87 #include <vm/vm_map.h> 88 #include <vm/vm_object.h> 89 #include <vm/vm_page.h> 90 #include <vm/vm_pageout.h> 91 #include <vm/vm_pager.h> 92 #include <vm/vm_extern.h> 93 #include <vm/swap_pager.h> 94 95 #include <machine/inttypes.h> 96 #include <machine/md_var.h> 97 #include <machine/specialreg.h> 98 #include <machine/bus_dma.h> 99 100 #include <vm/vm_page2.h> 101 #include <sys/spinlock2.h> 102 103 /* 104 * SET - Minimum required set associative size, must be a power of 2. We 105 * want this to match or exceed the set-associativeness of the cpu. 106 * 107 * GRP - A larger set that allows bleed-over into the domains of other 108 * nearby cpus. Also must be a power of 2. Used by the page zeroing 109 * code to smooth things out a bit. 110 */ 111 #define PQ_SET_ASSOC 16 112 #define PQ_SET_ASSOC_MASK (PQ_SET_ASSOC - 1) 113 114 #define PQ_GRP_ASSOC (PQ_SET_ASSOC * 2) 115 #define PQ_GRP_ASSOC_MASK (PQ_GRP_ASSOC - 1) 116 117 static void vm_page_queue_init(void); 118 static void vm_page_free_wakeup(void); 119 static vm_page_t vm_page_select_cache(u_short pg_color); 120 static vm_page_t _vm_page_list_find2(int basequeue, int index); 121 static void _vm_page_deactivate_locked(vm_page_t m, int athead); 122 123 /* 124 * Array of tailq lists 125 */ 126 __cachealign struct vpgqueues vm_page_queues[PQ_COUNT]; 127 128 static volatile int vm_pages_waiting; 129 static struct alist vm_contig_alist; 130 static struct almeta vm_contig_ameta[ALIST_RECORDS_65536]; 131 static struct spinlock vm_contig_spin = SPINLOCK_INITIALIZER(&vm_contig_spin, "vm_contig_spin"); 132 133 static u_long vm_dma_reserved = 0; 134 TUNABLE_ULONG("vm.dma_reserved", &vm_dma_reserved); 135 SYSCTL_ULONG(_vm, OID_AUTO, dma_reserved, CTLFLAG_RD, &vm_dma_reserved, 0, 136 "Memory reserved for DMA"); 137 SYSCTL_UINT(_vm, OID_AUTO, dma_free_pages, CTLFLAG_RD, 138 &vm_contig_alist.bl_free, 0, "Memory reserved for DMA"); 139 140 static int vm_contig_verbose = 0; 141 TUNABLE_INT("vm.contig_verbose", &vm_contig_verbose); 142 143 RB_GENERATE2(vm_page_rb_tree, vm_page, rb_entry, rb_vm_page_compare, 144 vm_pindex_t, pindex); 145 146 static void 147 vm_page_queue_init(void) 148 { 149 int i; 150 151 for (i = 0; i < PQ_L2_SIZE; i++) 152 vm_page_queues[PQ_FREE+i].cnt_offset = 153 offsetof(struct vmstats, v_free_count); 154 for (i = 0; i < PQ_L2_SIZE; i++) 155 vm_page_queues[PQ_CACHE+i].cnt_offset = 156 offsetof(struct vmstats, v_cache_count); 157 for (i = 0; i < PQ_L2_SIZE; i++) 158 vm_page_queues[PQ_INACTIVE+i].cnt_offset = 159 offsetof(struct vmstats, v_inactive_count); 160 for (i = 0; i < PQ_L2_SIZE; i++) 161 vm_page_queues[PQ_ACTIVE+i].cnt_offset = 162 offsetof(struct vmstats, v_active_count); 163 for (i = 0; i < PQ_L2_SIZE; i++) 164 vm_page_queues[PQ_HOLD+i].cnt_offset = 165 offsetof(struct vmstats, v_active_count); 166 /* PQ_NONE has no queue */ 167 168 for (i = 0; i < PQ_COUNT; i++) { 169 TAILQ_INIT(&vm_page_queues[i].pl); 170 spin_init(&vm_page_queues[i].spin, "vm_page_queue_init"); 171 } 172 } 173 174 /* 175 * note: place in initialized data section? Is this necessary? 176 */ 177 vm_pindex_t first_page = 0; 178 vm_pindex_t vm_page_array_size = 0; 179 vm_page_t vm_page_array = NULL; 180 vm_paddr_t vm_low_phys_reserved; 181 182 /* 183 * (low level boot) 184 * 185 * Sets the page size, perhaps based upon the memory size. 186 * Must be called before any use of page-size dependent functions. 187 */ 188 void 189 vm_set_page_size(void) 190 { 191 if (vmstats.v_page_size == 0) 192 vmstats.v_page_size = PAGE_SIZE; 193 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0) 194 panic("vm_set_page_size: page size not a power of two"); 195 } 196 197 /* 198 * (low level boot) 199 * 200 * Add a new page to the freelist for use by the system. New pages 201 * are added to both the head and tail of the associated free page 202 * queue in a bottom-up fashion, so both zero'd and non-zero'd page 203 * requests pull 'recent' adds (higher physical addresses) first. 204 * 205 * Beware that the page zeroing daemon will also be running soon after 206 * boot, moving pages from the head to the tail of the PQ_FREE queues. 207 * 208 * Must be called in a critical section. 209 */ 210 static void 211 vm_add_new_page(vm_paddr_t pa) 212 { 213 struct vpgqueues *vpq; 214 vm_page_t m; 215 216 m = PHYS_TO_VM_PAGE(pa); 217 m->phys_addr = pa; 218 m->flags = 0; 219 m->pat_mode = PAT_WRITE_BACK; 220 m->pc = (pa >> PAGE_SHIFT); 221 222 /* 223 * Twist for cpu localization in addition to page coloring, so 224 * different cpus selecting by m->queue get different page colors. 225 */ 226 m->pc ^= ((pa >> PAGE_SHIFT) / PQ_L2_SIZE); 227 m->pc ^= ((pa >> PAGE_SHIFT) / (PQ_L2_SIZE * PQ_L2_SIZE)); 228 m->pc &= PQ_L2_MASK; 229 230 /* 231 * Reserve a certain number of contiguous low memory pages for 232 * contigmalloc() to use. 233 */ 234 if (pa < vm_low_phys_reserved) { 235 atomic_add_long(&vmstats.v_page_count, 1); 236 atomic_add_long(&vmstats.v_dma_pages, 1); 237 m->queue = PQ_NONE; 238 m->wire_count = 1; 239 atomic_add_long(&vmstats.v_wire_count, 1); 240 alist_free(&vm_contig_alist, pa >> PAGE_SHIFT, 1); 241 return; 242 } 243 244 /* 245 * General page 246 */ 247 m->queue = m->pc + PQ_FREE; 248 KKASSERT(m->dirty == 0); 249 250 atomic_add_long(&vmstats.v_page_count, 1); 251 atomic_add_long(&vmstats.v_free_count, 1); 252 vpq = &vm_page_queues[m->queue]; 253 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq); 254 ++vpq->lcnt; 255 } 256 257 /* 258 * (low level boot) 259 * 260 * Initializes the resident memory module. 261 * 262 * Preallocates memory for critical VM structures and arrays prior to 263 * kernel_map becoming available. 264 * 265 * Memory is allocated from (virtual2_start, virtual2_end) if available, 266 * otherwise memory is allocated from (virtual_start, virtual_end). 267 * 268 * On x86-64 (virtual_start, virtual_end) is only 2GB and may not be 269 * large enough to hold vm_page_array & other structures for machines with 270 * large amounts of ram, so we want to use virtual2* when available. 271 */ 272 void 273 vm_page_startup(void) 274 { 275 vm_offset_t vaddr = virtual2_start ? virtual2_start : virtual_start; 276 vm_offset_t mapped; 277 vm_pindex_t npages; 278 vm_paddr_t page_range; 279 vm_paddr_t new_end; 280 int i; 281 vm_paddr_t pa; 282 vm_paddr_t last_pa; 283 vm_paddr_t end; 284 vm_paddr_t biggestone, biggestsize; 285 vm_paddr_t total; 286 vm_page_t m; 287 288 total = 0; 289 biggestsize = 0; 290 biggestone = 0; 291 vaddr = round_page(vaddr); 292 293 /* 294 * Make sure ranges are page-aligned. 295 */ 296 for (i = 0; phys_avail[i].phys_end; ++i) { 297 phys_avail[i].phys_beg = round_page64(phys_avail[i].phys_beg); 298 phys_avail[i].phys_end = trunc_page64(phys_avail[i].phys_end); 299 if (phys_avail[i].phys_end < phys_avail[i].phys_beg) 300 phys_avail[i].phys_end = phys_avail[i].phys_beg; 301 } 302 303 /* 304 * Locate largest block 305 */ 306 for (i = 0; phys_avail[i].phys_end; ++i) { 307 vm_paddr_t size = phys_avail[i].phys_end - 308 phys_avail[i].phys_beg; 309 310 if (size > biggestsize) { 311 biggestone = i; 312 biggestsize = size; 313 } 314 total += size; 315 } 316 --i; /* adjust to last entry for use down below */ 317 318 end = phys_avail[biggestone].phys_end; 319 end = trunc_page(end); 320 321 /* 322 * Initialize the queue headers for the free queue, the active queue 323 * and the inactive queue. 324 */ 325 vm_page_queue_init(); 326 327 #if !defined(_KERNEL_VIRTUAL) 328 /* 329 * VKERNELs don't support minidumps and as such don't need 330 * vm_page_dump 331 * 332 * Allocate a bitmap to indicate that a random physical page 333 * needs to be included in a minidump. 334 * 335 * The amd64 port needs this to indicate which direct map pages 336 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 337 * 338 * However, x86 still needs this workspace internally within the 339 * minidump code. In theory, they are not needed on x86, but are 340 * included should the sf_buf code decide to use them. 341 */ 342 page_range = phys_avail[i].phys_end / PAGE_SIZE; 343 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 344 end -= vm_page_dump_size; 345 vm_page_dump = (void *)pmap_map(&vaddr, end, end + vm_page_dump_size, 346 VM_PROT_READ | VM_PROT_WRITE); 347 bzero((void *)vm_page_dump, vm_page_dump_size); 348 #endif 349 /* 350 * Compute the number of pages of memory that will be available for 351 * use (taking into account the overhead of a page structure per 352 * page). 353 */ 354 first_page = phys_avail[0].phys_beg / PAGE_SIZE; 355 page_range = phys_avail[i].phys_end / PAGE_SIZE - first_page; 356 npages = (total - (page_range * sizeof(struct vm_page))) / PAGE_SIZE; 357 358 #ifndef _KERNEL_VIRTUAL 359 /* 360 * (only applies to real kernels) 361 * 362 * Reserve a large amount of low memory for potential 32-bit DMA 363 * space allocations. Once device initialization is complete we 364 * release most of it, but keep (vm_dma_reserved) memory reserved 365 * for later use. Typically for X / graphics. Through trial and 366 * error we find that GPUs usually requires ~60-100MB or so. 367 * 368 * By default, 128M is left in reserve on machines with 2G+ of ram. 369 */ 370 vm_low_phys_reserved = (vm_paddr_t)65536 << PAGE_SHIFT; 371 if (vm_low_phys_reserved > total / 4) 372 vm_low_phys_reserved = total / 4; 373 if (vm_dma_reserved == 0) { 374 vm_dma_reserved = 128 * 1024 * 1024; /* 128MB */ 375 if (vm_dma_reserved > total / 16) 376 vm_dma_reserved = total / 16; 377 } 378 #endif 379 alist_init(&vm_contig_alist, 65536, vm_contig_ameta, 380 ALIST_RECORDS_65536); 381 382 /* 383 * Initialize the mem entry structures now, and put them in the free 384 * queue. 385 */ 386 if (bootverbose && ctob(physmem) >= 400LL*1024*1024*1024) 387 kprintf("initializing vm_page_array "); 388 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 389 mapped = pmap_map(&vaddr, new_end, end, VM_PROT_READ | VM_PROT_WRITE); 390 vm_page_array = (vm_page_t)mapped; 391 392 #if defined(__x86_64__) && !defined(_KERNEL_VIRTUAL) 393 /* 394 * since pmap_map on amd64 returns stuff out of a direct-map region, 395 * we have to manually add these pages to the minidump tracking so 396 * that they can be dumped, including the vm_page_array. 397 */ 398 for (pa = new_end; 399 pa < phys_avail[biggestone].phys_end; 400 pa += PAGE_SIZE) { 401 dump_add_page(pa); 402 } 403 #endif 404 405 /* 406 * Clear all of the page structures, run basic initialization so 407 * PHYS_TO_VM_PAGE() operates properly even on pages not in the 408 * map. 409 */ 410 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 411 vm_page_array_size = page_range; 412 if (bootverbose && ctob(physmem) >= 400LL*1024*1024*1024) 413 kprintf("size = 0x%zx\n", vm_page_array_size); 414 415 m = &vm_page_array[0]; 416 pa = ptoa(first_page); 417 for (i = 0; i < page_range; ++i) { 418 spin_init(&m->spin, "vm_page"); 419 m->phys_addr = pa; 420 pa += PAGE_SIZE; 421 ++m; 422 } 423 424 /* 425 * Construct the free queue(s) in ascending order (by physical 426 * address) so that the first 16MB of physical memory is allocated 427 * last rather than first. On large-memory machines, this avoids 428 * the exhaustion of low physical memory before isa_dma_init has run. 429 */ 430 vmstats.v_page_count = 0; 431 vmstats.v_free_count = 0; 432 for (i = 0; phys_avail[i].phys_end && npages > 0; ++i) { 433 pa = phys_avail[i].phys_beg; 434 if (i == biggestone) 435 last_pa = new_end; 436 else 437 last_pa = phys_avail[i].phys_end; 438 while (pa < last_pa && npages-- > 0) { 439 vm_add_new_page(pa); 440 pa += PAGE_SIZE; 441 } 442 } 443 if (virtual2_start) 444 virtual2_start = vaddr; 445 else 446 virtual_start = vaddr; 447 mycpu->gd_vmstats = vmstats; 448 } 449 450 /* 451 * Reorganize VM pages based on numa data. May be called as many times as 452 * necessary. Will reorganize the vm_page_t page color and related queue(s) 453 * to allow vm_page_alloc() to choose pages based on socket affinity. 454 * 455 * NOTE: This function is only called while we are still in UP mode, so 456 * we only need a critical section to protect the queues (which 457 * saves a lot of time, there are likely a ton of pages). 458 */ 459 void 460 vm_numa_organize(vm_paddr_t ran_beg, vm_paddr_t bytes, int physid) 461 { 462 vm_paddr_t scan_beg; 463 vm_paddr_t scan_end; 464 vm_paddr_t ran_end; 465 struct vpgqueues *vpq; 466 vm_page_t m; 467 vm_page_t mend; 468 int i; 469 int socket_mod; 470 int socket_value; 471 472 /* 473 * Check if no physical information, or there was only one socket 474 * (so don't waste time doing nothing!). 475 */ 476 if (cpu_topology_phys_ids <= 1 || 477 cpu_topology_core_ids == 0) { 478 return; 479 } 480 481 /* 482 * Setup for our iteration. Note that ACPI may iterate CPU 483 * sockets starting at 0 or 1 or some other number. The 484 * cpu_topology code mod's it against the socket count. 485 */ 486 ran_end = ran_beg + bytes; 487 physid %= cpu_topology_phys_ids; 488 489 socket_mod = PQ_L2_SIZE / cpu_topology_phys_ids; 490 socket_value = physid * socket_mod; 491 mend = &vm_page_array[vm_page_array_size]; 492 493 crit_enter(); 494 495 /* 496 * Adjust vm_page->pc and requeue all affected pages. The 497 * allocator will then be able to localize memory allocations 498 * to some degree. 499 */ 500 for (i = 0; phys_avail[i].phys_end; ++i) { 501 scan_beg = phys_avail[i].phys_beg; 502 scan_end = phys_avail[i].phys_end; 503 if (scan_end <= ran_beg) 504 continue; 505 if (scan_beg >= ran_end) 506 continue; 507 if (scan_beg < ran_beg) 508 scan_beg = ran_beg; 509 if (scan_end > ran_end) 510 scan_end = ran_end; 511 if (atop(scan_end) > first_page + vm_page_array_size) 512 scan_end = ptoa(first_page + vm_page_array_size); 513 514 m = PHYS_TO_VM_PAGE(scan_beg); 515 while (scan_beg < scan_end) { 516 KKASSERT(m < mend); 517 if (m->queue != PQ_NONE) { 518 vpq = &vm_page_queues[m->queue]; 519 TAILQ_REMOVE(&vpq->pl, m, pageq); 520 --vpq->lcnt; 521 /* queue doesn't change, no need to adj cnt */ 522 m->queue -= m->pc; 523 m->pc %= socket_mod; 524 m->pc += socket_value; 525 m->pc &= PQ_L2_MASK; 526 m->queue += m->pc; 527 vpq = &vm_page_queues[m->queue]; 528 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq); 529 ++vpq->lcnt; 530 /* queue doesn't change, no need to adj cnt */ 531 } else { 532 m->pc %= socket_mod; 533 m->pc += socket_value; 534 m->pc &= PQ_L2_MASK; 535 } 536 scan_beg += PAGE_SIZE; 537 ++m; 538 } 539 } 540 crit_exit(); 541 } 542 543 /* 544 * We tended to reserve a ton of memory for contigmalloc(). Now that most 545 * drivers have initialized we want to return most the remaining free 546 * reserve back to the VM page queues so they can be used for normal 547 * allocations. 548 * 549 * We leave vm_dma_reserved bytes worth of free pages in the reserve pool. 550 */ 551 static void 552 vm_page_startup_finish(void *dummy __unused) 553 { 554 alist_blk_t blk; 555 alist_blk_t rblk; 556 alist_blk_t count; 557 alist_blk_t xcount; 558 alist_blk_t bfree; 559 vm_page_t m; 560 561 spin_lock(&vm_contig_spin); 562 for (;;) { 563 bfree = alist_free_info(&vm_contig_alist, &blk, &count); 564 if (bfree <= vm_dma_reserved / PAGE_SIZE) 565 break; 566 if (count == 0) 567 break; 568 569 /* 570 * Figure out how much of the initial reserve we have to 571 * free in order to reach our target. 572 */ 573 bfree -= vm_dma_reserved / PAGE_SIZE; 574 if (count > bfree) { 575 blk += count - bfree; 576 count = bfree; 577 } 578 579 /* 580 * Calculate the nearest power of 2 <= count. 581 */ 582 for (xcount = 1; xcount <= count; xcount <<= 1) 583 ; 584 xcount >>= 1; 585 blk += count - xcount; 586 count = xcount; 587 588 /* 589 * Allocate the pages from the alist, then free them to 590 * the normal VM page queues. 591 * 592 * Pages allocated from the alist are wired. We have to 593 * busy, unwire, and free them. We must also adjust 594 * vm_low_phys_reserved before freeing any pages to prevent 595 * confusion. 596 */ 597 rblk = alist_alloc(&vm_contig_alist, blk, count); 598 if (rblk != blk) { 599 kprintf("vm_page_startup_finish: Unable to return " 600 "dma space @0x%08x/%d -> 0x%08x\n", 601 blk, count, rblk); 602 break; 603 } 604 atomic_add_long(&vmstats.v_dma_pages, -(long)count); 605 spin_unlock(&vm_contig_spin); 606 607 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT); 608 vm_low_phys_reserved = VM_PAGE_TO_PHYS(m); 609 while (count) { 610 vm_page_busy_wait(m, FALSE, "cpgfr"); 611 vm_page_unwire(m, 0); 612 vm_page_free(m); 613 --count; 614 ++m; 615 } 616 spin_lock(&vm_contig_spin); 617 } 618 spin_unlock(&vm_contig_spin); 619 620 /* 621 * Print out how much DMA space drivers have already allocated and 622 * how much is left over. 623 */ 624 kprintf("DMA space used: %jdk, remaining available: %jdk\n", 625 (intmax_t)(vmstats.v_dma_pages - vm_contig_alist.bl_free) * 626 (PAGE_SIZE / 1024), 627 (intmax_t)vm_contig_alist.bl_free * (PAGE_SIZE / 1024)); 628 } 629 SYSINIT(vm_pgend, SI_SUB_PROC0_POST, SI_ORDER_ANY, 630 vm_page_startup_finish, NULL); 631 632 633 /* 634 * Scan comparison function for Red-Black tree scans. An inclusive 635 * (start,end) is expected. Other fields are not used. 636 */ 637 int 638 rb_vm_page_scancmp(struct vm_page *p, void *data) 639 { 640 struct rb_vm_page_scan_info *info = data; 641 642 if (p->pindex < info->start_pindex) 643 return(-1); 644 if (p->pindex > info->end_pindex) 645 return(1); 646 return(0); 647 } 648 649 int 650 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2) 651 { 652 if (p1->pindex < p2->pindex) 653 return(-1); 654 if (p1->pindex > p2->pindex) 655 return(1); 656 return(0); 657 } 658 659 void 660 vm_page_init(vm_page_t m) 661 { 662 /* do nothing for now. Called from pmap_page_init() */ 663 } 664 665 /* 666 * Each page queue has its own spin lock, which is fairly optimal for 667 * allocating and freeing pages at least. 668 * 669 * The caller must hold the vm_page_spin_lock() before locking a vm_page's 670 * queue spinlock via this function. Also note that m->queue cannot change 671 * unless both the page and queue are locked. 672 */ 673 static __inline 674 void 675 _vm_page_queue_spin_lock(vm_page_t m) 676 { 677 u_short queue; 678 679 queue = m->queue; 680 if (queue != PQ_NONE) { 681 spin_lock(&vm_page_queues[queue].spin); 682 KKASSERT(queue == m->queue); 683 } 684 } 685 686 static __inline 687 void 688 _vm_page_queue_spin_unlock(vm_page_t m) 689 { 690 u_short queue; 691 692 queue = m->queue; 693 cpu_ccfence(); 694 if (queue != PQ_NONE) 695 spin_unlock(&vm_page_queues[queue].spin); 696 } 697 698 static __inline 699 void 700 _vm_page_queues_spin_lock(u_short queue) 701 { 702 cpu_ccfence(); 703 if (queue != PQ_NONE) 704 spin_lock(&vm_page_queues[queue].spin); 705 } 706 707 708 static __inline 709 void 710 _vm_page_queues_spin_unlock(u_short queue) 711 { 712 cpu_ccfence(); 713 if (queue != PQ_NONE) 714 spin_unlock(&vm_page_queues[queue].spin); 715 } 716 717 void 718 vm_page_queue_spin_lock(vm_page_t m) 719 { 720 _vm_page_queue_spin_lock(m); 721 } 722 723 void 724 vm_page_queues_spin_lock(u_short queue) 725 { 726 _vm_page_queues_spin_lock(queue); 727 } 728 729 void 730 vm_page_queue_spin_unlock(vm_page_t m) 731 { 732 _vm_page_queue_spin_unlock(m); 733 } 734 735 void 736 vm_page_queues_spin_unlock(u_short queue) 737 { 738 _vm_page_queues_spin_unlock(queue); 739 } 740 741 /* 742 * This locks the specified vm_page and its queue in the proper order 743 * (page first, then queue). The queue may change so the caller must 744 * recheck on return. 745 */ 746 static __inline 747 void 748 _vm_page_and_queue_spin_lock(vm_page_t m) 749 { 750 vm_page_spin_lock(m); 751 _vm_page_queue_spin_lock(m); 752 } 753 754 static __inline 755 void 756 _vm_page_and_queue_spin_unlock(vm_page_t m) 757 { 758 _vm_page_queues_spin_unlock(m->queue); 759 vm_page_spin_unlock(m); 760 } 761 762 void 763 vm_page_and_queue_spin_unlock(vm_page_t m) 764 { 765 _vm_page_and_queue_spin_unlock(m); 766 } 767 768 void 769 vm_page_and_queue_spin_lock(vm_page_t m) 770 { 771 _vm_page_and_queue_spin_lock(m); 772 } 773 774 /* 775 * Helper function removes vm_page from its current queue. 776 * Returns the base queue the page used to be on. 777 * 778 * The vm_page and the queue must be spinlocked. 779 * This function will unlock the queue but leave the page spinlocked. 780 */ 781 static __inline u_short 782 _vm_page_rem_queue_spinlocked(vm_page_t m) 783 { 784 struct vpgqueues *pq; 785 u_short queue; 786 u_short oqueue; 787 long *cnt; 788 789 queue = m->queue; 790 if (queue != PQ_NONE) { 791 pq = &vm_page_queues[queue]; 792 TAILQ_REMOVE(&pq->pl, m, pageq); 793 794 /* 795 * Adjust our pcpu stats. In order for the nominal low-memory 796 * algorithms to work properly we don't let any pcpu stat get 797 * too negative before we force it to be rolled-up into the 798 * global stats. Otherwise our pageout and vm_wait tests 799 * will fail badly. 800 * 801 * The idea here is to reduce unnecessary SMP cache 802 * mastership changes in the global vmstats, which can be 803 * particularly bad in multi-socket systems. 804 */ 805 cnt = (long *)((char *)&mycpu->gd_vmstats_adj + pq->cnt_offset); 806 atomic_add_long(cnt, -1); 807 if (*cnt < -VMMETER_SLOP_COUNT) { 808 u_long copy = atomic_swap_long(cnt, 0); 809 cnt = (long *)((char *)&vmstats + pq->cnt_offset); 810 atomic_add_long(cnt, copy); 811 cnt = (long *)((char *)&mycpu->gd_vmstats + 812 pq->cnt_offset); 813 atomic_add_long(cnt, copy); 814 } 815 pq->lcnt--; 816 m->queue = PQ_NONE; 817 oqueue = queue; 818 queue -= m->pc; 819 vm_page_queues_spin_unlock(oqueue); /* intended */ 820 } 821 return queue; 822 } 823 824 /* 825 * Helper function places the vm_page on the specified queue. Generally 826 * speaking only PQ_FREE pages are placed at the head, to allow them to 827 * be allocated sooner rather than later on the assumption that they 828 * are cache-hot. 829 * 830 * The vm_page must be spinlocked. 831 * This function will return with both the page and the queue locked. 832 */ 833 static __inline void 834 _vm_page_add_queue_spinlocked(vm_page_t m, u_short queue, int athead) 835 { 836 struct vpgqueues *pq; 837 u_long *cnt; 838 839 KKASSERT(m->queue == PQ_NONE); 840 841 if (queue != PQ_NONE) { 842 vm_page_queues_spin_lock(queue); 843 pq = &vm_page_queues[queue]; 844 ++pq->lcnt; 845 846 /* 847 * Adjust our pcpu stats. If a system entity really needs 848 * to incorporate the count it will call vmstats_rollup() 849 * to roll it all up into the global vmstats strufture. 850 */ 851 cnt = (long *)((char *)&mycpu->gd_vmstats_adj + pq->cnt_offset); 852 atomic_add_long(cnt, 1); 853 854 /* 855 * PQ_FREE is always handled LIFO style to try to provide 856 * cache-hot pages to programs. 857 */ 858 m->queue = queue; 859 if (queue - m->pc == PQ_FREE) { 860 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 861 } else if (athead) { 862 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 863 } else { 864 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 865 } 866 /* leave the queue spinlocked */ 867 } 868 } 869 870 /* 871 * Wait until page is no longer BUSY. If also_m_busy is TRUE we wait 872 * until the page is no longer BUSY or SBUSY (busy_count field is 0). 873 * 874 * Returns TRUE if it had to sleep, FALSE if we did not. Only one sleep 875 * call will be made before returning. 876 * 877 * This function does NOT busy the page and on return the page is not 878 * guaranteed to be available. 879 */ 880 void 881 vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg) 882 { 883 u_int32_t busy_count; 884 885 for (;;) { 886 busy_count = m->busy_count; 887 cpu_ccfence(); 888 889 if ((busy_count & PBUSY_LOCKED) == 0 && 890 (also_m_busy == 0 || (busy_count & PBUSY_MASK) == 0)) { 891 break; 892 } 893 tsleep_interlock(m, 0); 894 if (atomic_cmpset_int(&m->busy_count, busy_count, 895 busy_count | PBUSY_WANTED)) { 896 atomic_set_int(&m->flags, PG_REFERENCED); 897 tsleep(m, PINTERLOCKED, msg, 0); 898 break; 899 } 900 } 901 } 902 903 /* 904 * This calculates and returns a page color given an optional VM object and 905 * either a pindex or an iterator. We attempt to return a cpu-localized 906 * pg_color that is still roughly 16-way set-associative. The CPU topology 907 * is used if it was probed. 908 * 909 * The caller may use the returned value to index into e.g. PQ_FREE when 910 * allocating a page in order to nominally obtain pages that are hopefully 911 * already localized to the requesting cpu. This function is not able to 912 * provide any sort of guarantee of this, but does its best to improve 913 * hardware cache management performance. 914 * 915 * WARNING! The caller must mask the returned value with PQ_L2_MASK. 916 */ 917 u_short 918 vm_get_pg_color(int cpuid, vm_object_t object, vm_pindex_t pindex) 919 { 920 u_short pg_color; 921 int phys_id; 922 int core_id; 923 int object_pg_color; 924 925 phys_id = get_cpu_phys_id(cpuid); 926 core_id = get_cpu_core_id(cpuid); 927 object_pg_color = object ? object->pg_color : 0; 928 929 if (cpu_topology_phys_ids && cpu_topology_core_ids) { 930 int grpsize; 931 932 /* 933 * Break us down by socket and cpu 934 */ 935 pg_color = phys_id * PQ_L2_SIZE / cpu_topology_phys_ids; 936 pg_color += core_id * PQ_L2_SIZE / 937 (cpu_topology_core_ids * cpu_topology_phys_ids); 938 939 /* 940 * Calculate remaining component for object/queue color 941 */ 942 grpsize = PQ_L2_SIZE / (cpu_topology_core_ids * 943 cpu_topology_phys_ids); 944 if (grpsize >= 8) { 945 pg_color += (pindex + object_pg_color) % grpsize; 946 } else { 947 if (grpsize <= 2) { 948 grpsize = 8; 949 } else { 950 /* 3->9, 4->8, 5->10, 6->12, 7->14 */ 951 grpsize += grpsize; 952 if (grpsize < 8) 953 grpsize += grpsize; 954 } 955 pg_color += (pindex + object_pg_color) % grpsize; 956 } 957 } else { 958 /* 959 * Unknown topology, distribute things evenly. 960 */ 961 pg_color = cpuid * PQ_L2_SIZE / ncpus; 962 pg_color += pindex + object_pg_color; 963 } 964 return (pg_color & PQ_L2_MASK); 965 } 966 967 /* 968 * Wait until BUSY can be set, then set it. If also_m_busy is TRUE we 969 * also wait for m->busy_count to become 0 before setting PBUSY_LOCKED. 970 */ 971 void 972 VM_PAGE_DEBUG_EXT(vm_page_busy_wait)(vm_page_t m, 973 int also_m_busy, const char *msg 974 VM_PAGE_DEBUG_ARGS) 975 { 976 u_int32_t busy_count; 977 978 for (;;) { 979 busy_count = m->busy_count; 980 cpu_ccfence(); 981 if (busy_count & PBUSY_LOCKED) { 982 tsleep_interlock(m, 0); 983 if (atomic_cmpset_int(&m->busy_count, busy_count, 984 busy_count | PBUSY_WANTED)) { 985 atomic_set_int(&m->flags, PG_REFERENCED); 986 tsleep(m, PINTERLOCKED, msg, 0); 987 } 988 } else if (also_m_busy && busy_count) { 989 tsleep_interlock(m, 0); 990 if (atomic_cmpset_int(&m->busy_count, busy_count, 991 busy_count | PBUSY_WANTED)) { 992 atomic_set_int(&m->flags, PG_REFERENCED); 993 tsleep(m, PINTERLOCKED, msg, 0); 994 } 995 } else { 996 if (atomic_cmpset_int(&m->busy_count, busy_count, 997 busy_count | PBUSY_LOCKED)) { 998 #ifdef VM_PAGE_DEBUG 999 m->busy_func = func; 1000 m->busy_line = lineno; 1001 #endif 1002 break; 1003 } 1004 } 1005 } 1006 } 1007 1008 /* 1009 * Attempt to set BUSY. If also_m_busy is TRUE we only succeed if 1010 * m->busy_count is also 0. 1011 * 1012 * Returns non-zero on failure. 1013 */ 1014 int 1015 VM_PAGE_DEBUG_EXT(vm_page_busy_try)(vm_page_t m, int also_m_busy 1016 VM_PAGE_DEBUG_ARGS) 1017 { 1018 u_int32_t busy_count; 1019 1020 for (;;) { 1021 busy_count = m->busy_count; 1022 cpu_ccfence(); 1023 if (busy_count & PBUSY_LOCKED) 1024 return TRUE; 1025 if (also_m_busy && (busy_count & PBUSY_MASK) != 0) 1026 return TRUE; 1027 if (atomic_cmpset_int(&m->busy_count, busy_count, 1028 busy_count | PBUSY_LOCKED)) { 1029 #ifdef VM_PAGE_DEBUG 1030 m->busy_func = func; 1031 m->busy_line = lineno; 1032 #endif 1033 return FALSE; 1034 } 1035 } 1036 } 1037 1038 /* 1039 * Clear the BUSY flag and return non-zero to indicate to the caller 1040 * that a wakeup() should be performed. 1041 * 1042 * The vm_page must be spinlocked and will remain spinlocked on return. 1043 * The related queue must NOT be spinlocked (which could deadlock us). 1044 * 1045 * (inline version) 1046 */ 1047 static __inline 1048 int 1049 _vm_page_wakeup(vm_page_t m) 1050 { 1051 u_int32_t busy_count; 1052 1053 for (;;) { 1054 busy_count = m->busy_count; 1055 cpu_ccfence(); 1056 if (atomic_cmpset_int(&m->busy_count, busy_count, 1057 busy_count & 1058 ~(PBUSY_LOCKED | PBUSY_WANTED))) { 1059 break; 1060 } 1061 } 1062 return((int)(busy_count & PBUSY_WANTED)); 1063 } 1064 1065 /* 1066 * Clear the BUSY flag and wakeup anyone waiting for the page. This 1067 * is typically the last call you make on a page before moving onto 1068 * other things. 1069 */ 1070 void 1071 vm_page_wakeup(vm_page_t m) 1072 { 1073 KASSERT(m->busy_count & PBUSY_LOCKED, 1074 ("vm_page_wakeup: page not busy!!!")); 1075 vm_page_spin_lock(m); 1076 if (_vm_page_wakeup(m)) { 1077 vm_page_spin_unlock(m); 1078 wakeup(m); 1079 } else { 1080 vm_page_spin_unlock(m); 1081 } 1082 } 1083 1084 /* 1085 * Holding a page keeps it from being reused. Other parts of the system 1086 * can still disassociate the page from its current object and free it, or 1087 * perform read or write I/O on it and/or otherwise manipulate the page, 1088 * but if the page is held the VM system will leave the page and its data 1089 * intact and not reuse the page for other purposes until the last hold 1090 * reference is released. (see vm_page_wire() if you want to prevent the 1091 * page from being disassociated from its object too). 1092 * 1093 * The caller must still validate the contents of the page and, if necessary, 1094 * wait for any pending I/O (e.g. vm_page_sleep_busy() loop) to complete 1095 * before manipulating the page. 1096 * 1097 * XXX get vm_page_spin_lock() here and move FREE->HOLD if necessary 1098 */ 1099 void 1100 vm_page_hold(vm_page_t m) 1101 { 1102 vm_page_spin_lock(m); 1103 atomic_add_int(&m->hold_count, 1); 1104 if (m->queue - m->pc == PQ_FREE) { 1105 _vm_page_queue_spin_lock(m); 1106 _vm_page_rem_queue_spinlocked(m); 1107 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0); 1108 _vm_page_queue_spin_unlock(m); 1109 } 1110 vm_page_spin_unlock(m); 1111 } 1112 1113 /* 1114 * The opposite of vm_page_hold(). If the page is on the HOLD queue 1115 * it was freed while held and must be moved back to the FREE queue. 1116 */ 1117 void 1118 vm_page_unhold(vm_page_t m) 1119 { 1120 KASSERT(m->hold_count > 0 && m->queue - m->pc != PQ_FREE, 1121 ("vm_page_unhold: pg %p illegal hold_count (%d) or on FREE queue (%d)", 1122 m, m->hold_count, m->queue - m->pc)); 1123 vm_page_spin_lock(m); 1124 atomic_add_int(&m->hold_count, -1); 1125 if (m->hold_count == 0 && m->queue - m->pc == PQ_HOLD) { 1126 _vm_page_queue_spin_lock(m); 1127 _vm_page_rem_queue_spinlocked(m); 1128 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 1); 1129 _vm_page_queue_spin_unlock(m); 1130 } 1131 vm_page_spin_unlock(m); 1132 } 1133 1134 /* 1135 * vm_page_getfake: 1136 * 1137 * Create a fictitious page with the specified physical address and 1138 * memory attribute. The memory attribute is the only the machine- 1139 * dependent aspect of a fictitious page that must be initialized. 1140 */ 1141 1142 void 1143 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1144 { 1145 1146 if ((m->flags & PG_FICTITIOUS) != 0) { 1147 /* 1148 * The page's memattr might have changed since the 1149 * previous initialization. Update the pmap to the 1150 * new memattr. 1151 */ 1152 goto memattr; 1153 } 1154 m->phys_addr = paddr; 1155 m->queue = PQ_NONE; 1156 /* Fictitious pages don't use "segind". */ 1157 /* Fictitious pages don't use "order" or "pool". */ 1158 m->flags = PG_FICTITIOUS | PG_UNMANAGED; 1159 m->busy_count = PBUSY_LOCKED; 1160 m->wire_count = 1; 1161 spin_init(&m->spin, "fake_page"); 1162 pmap_page_init(m); 1163 memattr: 1164 pmap_page_set_memattr(m, memattr); 1165 } 1166 1167 /* 1168 * Inserts the given vm_page into the object and object list. 1169 * 1170 * The pagetables are not updated but will presumably fault the page 1171 * in if necessary, or if a kernel page the caller will at some point 1172 * enter the page into the kernel's pmap. We are not allowed to block 1173 * here so we *can't* do this anyway. 1174 * 1175 * This routine may not block. 1176 * This routine must be called with the vm_object held. 1177 * This routine must be called with a critical section held. 1178 * 1179 * This routine returns TRUE if the page was inserted into the object 1180 * successfully, and FALSE if the page already exists in the object. 1181 */ 1182 int 1183 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 1184 { 1185 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(object)); 1186 if (m->object != NULL) 1187 panic("vm_page_insert: already inserted"); 1188 1189 atomic_add_int(&object->generation, 1); 1190 1191 /* 1192 * Record the object/offset pair in this page and add the 1193 * pv_list_count of the page to the object. 1194 * 1195 * The vm_page spin lock is required for interactions with the pmap. 1196 */ 1197 vm_page_spin_lock(m); 1198 m->object = object; 1199 m->pindex = pindex; 1200 if (vm_page_rb_tree_RB_INSERT(&object->rb_memq, m)) { 1201 m->object = NULL; 1202 m->pindex = 0; 1203 vm_page_spin_unlock(m); 1204 return FALSE; 1205 } 1206 ++object->resident_page_count; 1207 ++mycpu->gd_vmtotal.t_rm; 1208 vm_page_spin_unlock(m); 1209 1210 /* 1211 * Since we are inserting a new and possibly dirty page, 1212 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 1213 */ 1214 if ((m->valid & m->dirty) || 1215 (m->flags & (PG_WRITEABLE | PG_NEED_COMMIT))) 1216 vm_object_set_writeable_dirty(object); 1217 1218 /* 1219 * Checks for a swap assignment and sets PG_SWAPPED if appropriate. 1220 */ 1221 swap_pager_page_inserted(m); 1222 return TRUE; 1223 } 1224 1225 /* 1226 * Removes the given vm_page_t from the (object,index) table 1227 * 1228 * The underlying pmap entry (if any) is NOT removed here. 1229 * This routine may not block. 1230 * 1231 * The page must be BUSY and will remain BUSY on return. 1232 * No other requirements. 1233 * 1234 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave 1235 * it busy. 1236 */ 1237 void 1238 vm_page_remove(vm_page_t m) 1239 { 1240 vm_object_t object; 1241 1242 if (m->object == NULL) { 1243 return; 1244 } 1245 1246 if ((m->busy_count & PBUSY_LOCKED) == 0) 1247 panic("vm_page_remove: page not busy"); 1248 1249 object = m->object; 1250 1251 vm_object_hold(object); 1252 1253 /* 1254 * Remove the page from the object and update the object. 1255 * 1256 * The vm_page spin lock is required for interactions with the pmap. 1257 */ 1258 vm_page_spin_lock(m); 1259 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m); 1260 --object->resident_page_count; 1261 --mycpu->gd_vmtotal.t_rm; 1262 m->object = NULL; 1263 atomic_add_int(&object->generation, 1); 1264 vm_page_spin_unlock(m); 1265 1266 vm_object_drop(object); 1267 } 1268 1269 /* 1270 * Locate and return the page at (object, pindex), or NULL if the 1271 * page could not be found. 1272 * 1273 * The caller must hold the vm_object token. 1274 */ 1275 vm_page_t 1276 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1277 { 1278 vm_page_t m; 1279 1280 /* 1281 * Search the hash table for this object/offset pair 1282 */ 1283 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1284 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1285 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex)); 1286 return(m); 1287 } 1288 1289 vm_page_t 1290 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_wait)(struct vm_object *object, 1291 vm_pindex_t pindex, 1292 int also_m_busy, const char *msg 1293 VM_PAGE_DEBUG_ARGS) 1294 { 1295 u_int32_t busy_count; 1296 vm_page_t m; 1297 1298 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1299 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1300 while (m) { 1301 KKASSERT(m->object == object && m->pindex == pindex); 1302 busy_count = m->busy_count; 1303 cpu_ccfence(); 1304 if (busy_count & PBUSY_LOCKED) { 1305 tsleep_interlock(m, 0); 1306 if (atomic_cmpset_int(&m->busy_count, busy_count, 1307 busy_count | PBUSY_WANTED)) { 1308 atomic_set_int(&m->flags, PG_REFERENCED); 1309 tsleep(m, PINTERLOCKED, msg, 0); 1310 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, 1311 pindex); 1312 } 1313 } else if (also_m_busy && busy_count) { 1314 tsleep_interlock(m, 0); 1315 if (atomic_cmpset_int(&m->busy_count, busy_count, 1316 busy_count | PBUSY_WANTED)) { 1317 atomic_set_int(&m->flags, PG_REFERENCED); 1318 tsleep(m, PINTERLOCKED, msg, 0); 1319 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, 1320 pindex); 1321 } 1322 } else if (atomic_cmpset_int(&m->busy_count, busy_count, 1323 busy_count | PBUSY_LOCKED)) { 1324 #ifdef VM_PAGE_DEBUG 1325 m->busy_func = func; 1326 m->busy_line = lineno; 1327 #endif 1328 break; 1329 } 1330 } 1331 return m; 1332 } 1333 1334 /* 1335 * Attempt to lookup and busy a page. 1336 * 1337 * Returns NULL if the page could not be found 1338 * 1339 * Returns a vm_page and error == TRUE if the page exists but could not 1340 * be busied. 1341 * 1342 * Returns a vm_page and error == FALSE on success. 1343 */ 1344 vm_page_t 1345 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_try)(struct vm_object *object, 1346 vm_pindex_t pindex, 1347 int also_m_busy, int *errorp 1348 VM_PAGE_DEBUG_ARGS) 1349 { 1350 u_int32_t busy_count; 1351 vm_page_t m; 1352 1353 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1354 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1355 *errorp = FALSE; 1356 while (m) { 1357 KKASSERT(m->object == object && m->pindex == pindex); 1358 busy_count = m->busy_count; 1359 cpu_ccfence(); 1360 if (busy_count & PBUSY_LOCKED) { 1361 *errorp = TRUE; 1362 break; 1363 } 1364 if (also_m_busy && busy_count) { 1365 *errorp = TRUE; 1366 break; 1367 } 1368 if (atomic_cmpset_int(&m->busy_count, busy_count, 1369 busy_count | PBUSY_LOCKED)) { 1370 #ifdef VM_PAGE_DEBUG 1371 m->busy_func = func; 1372 m->busy_line = lineno; 1373 #endif 1374 break; 1375 } 1376 } 1377 return m; 1378 } 1379 1380 /* 1381 * Returns a page that is only soft-busied for use by the caller in 1382 * a read-only fashion. Returns NULL if the page could not be found, 1383 * the soft busy could not be obtained, or the page data is invalid. 1384 */ 1385 vm_page_t 1386 vm_page_lookup_sbusy_try(struct vm_object *object, vm_pindex_t pindex, 1387 int pgoff, int pgbytes) 1388 { 1389 vm_page_t m; 1390 1391 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1392 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1393 if (m) { 1394 if ((m->valid != VM_PAGE_BITS_ALL && 1395 !vm_page_is_valid(m, pgoff, pgbytes)) || 1396 (m->flags & PG_FICTITIOUS)) { 1397 m = NULL; 1398 } else if (vm_page_sbusy_try(m)) { 1399 m = NULL; 1400 } else if ((m->valid != VM_PAGE_BITS_ALL && 1401 !vm_page_is_valid(m, pgoff, pgbytes)) || 1402 (m->flags & PG_FICTITIOUS)) { 1403 vm_page_sbusy_drop(m); 1404 m = NULL; 1405 } 1406 } 1407 return m; 1408 } 1409 1410 /* 1411 * Caller must hold the related vm_object 1412 */ 1413 vm_page_t 1414 vm_page_next(vm_page_t m) 1415 { 1416 vm_page_t next; 1417 1418 next = vm_page_rb_tree_RB_NEXT(m); 1419 if (next && next->pindex != m->pindex + 1) 1420 next = NULL; 1421 return (next); 1422 } 1423 1424 /* 1425 * vm_page_rename() 1426 * 1427 * Move the given vm_page from its current object to the specified 1428 * target object/offset. The page must be busy and will remain so 1429 * on return. 1430 * 1431 * new_object must be held. 1432 * This routine might block. XXX ? 1433 * 1434 * NOTE: Swap associated with the page must be invalidated by the move. We 1435 * have to do this for several reasons: (1) we aren't freeing the 1436 * page, (2) we are dirtying the page, (3) the VM system is probably 1437 * moving the page from object A to B, and will then later move 1438 * the backing store from A to B and we can't have a conflict. 1439 * 1440 * NOTE: We *always* dirty the page. It is necessary both for the 1441 * fact that we moved it, and because we may be invalidating 1442 * swap. If the page is on the cache, we have to deactivate it 1443 * or vm_page_dirty() will panic. Dirty pages are not allowed 1444 * on the cache. 1445 */ 1446 void 1447 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1448 { 1449 KKASSERT(m->busy_count & PBUSY_LOCKED); 1450 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(new_object)); 1451 if (m->object) { 1452 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(m->object)); 1453 vm_page_remove(m); 1454 } 1455 if (vm_page_insert(m, new_object, new_pindex) == FALSE) { 1456 panic("vm_page_rename: target exists (%p,%"PRIu64")", 1457 new_object, new_pindex); 1458 } 1459 if (m->queue - m->pc == PQ_CACHE) 1460 vm_page_deactivate(m); 1461 vm_page_dirty(m); 1462 } 1463 1464 /* 1465 * vm_page_unqueue() without any wakeup. This routine is used when a page 1466 * is to remain BUSYied by the caller. 1467 * 1468 * This routine may not block. 1469 */ 1470 void 1471 vm_page_unqueue_nowakeup(vm_page_t m) 1472 { 1473 vm_page_and_queue_spin_lock(m); 1474 (void)_vm_page_rem_queue_spinlocked(m); 1475 vm_page_spin_unlock(m); 1476 } 1477 1478 /* 1479 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon 1480 * if necessary. 1481 * 1482 * This routine may not block. 1483 */ 1484 void 1485 vm_page_unqueue(vm_page_t m) 1486 { 1487 u_short queue; 1488 1489 vm_page_and_queue_spin_lock(m); 1490 queue = _vm_page_rem_queue_spinlocked(m); 1491 if (queue == PQ_FREE || queue == PQ_CACHE) { 1492 vm_page_spin_unlock(m); 1493 pagedaemon_wakeup(); 1494 } else { 1495 vm_page_spin_unlock(m); 1496 } 1497 } 1498 1499 /* 1500 * vm_page_list_find() 1501 * 1502 * Find a page on the specified queue with color optimization. 1503 * 1504 * The page coloring optimization attempts to locate a page that does 1505 * not overload other nearby pages in the object in the cpu's L1 or L2 1506 * caches. We need this optimization because cpu caches tend to be 1507 * physical caches, while object spaces tend to be virtual. 1508 * 1509 * The page coloring optimization also, very importantly, tries to localize 1510 * memory to cpus and physical sockets. 1511 * 1512 * On MP systems each PQ_FREE and PQ_CACHE color queue has its own spinlock 1513 * and the algorithm is adjusted to localize allocations on a per-core basis. 1514 * This is done by 'twisting' the colors. 1515 * 1516 * The page is returned spinlocked and removed from its queue (it will 1517 * be on PQ_NONE), or NULL. The page is not BUSY'd. The caller 1518 * is responsible for dealing with the busy-page case (usually by 1519 * deactivating the page and looping). 1520 * 1521 * NOTE: This routine is carefully inlined. A non-inlined version 1522 * is available for outside callers but the only critical path is 1523 * from within this source file. 1524 * 1525 * NOTE: This routine assumes that the vm_pages found in PQ_CACHE and PQ_FREE 1526 * represent stable storage, allowing us to order our locks vm_page 1527 * first, then queue. 1528 */ 1529 static __inline 1530 vm_page_t 1531 _vm_page_list_find(int basequeue, int index) 1532 { 1533 vm_page_t m; 1534 1535 for (;;) { 1536 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl); 1537 if (m == NULL) { 1538 m = _vm_page_list_find2(basequeue, index); 1539 return(m); 1540 } 1541 vm_page_and_queue_spin_lock(m); 1542 if (m->queue == basequeue + index) { 1543 _vm_page_rem_queue_spinlocked(m); 1544 /* vm_page_t spin held, no queue spin */ 1545 break; 1546 } 1547 vm_page_and_queue_spin_unlock(m); 1548 } 1549 return(m); 1550 } 1551 1552 /* 1553 * If we could not find the page in the desired queue try to find it in 1554 * a nearby queue. 1555 */ 1556 static vm_page_t 1557 _vm_page_list_find2(int basequeue, int index) 1558 { 1559 struct vpgqueues *pq; 1560 vm_page_t m = NULL; 1561 int pqmask = PQ_SET_ASSOC_MASK >> 1; 1562 int pqi; 1563 int i; 1564 1565 index &= PQ_L2_MASK; 1566 pq = &vm_page_queues[basequeue]; 1567 1568 /* 1569 * Run local sets of 16, 32, 64, 128, and the whole queue if all 1570 * else fails (PQ_L2_MASK which is 255). 1571 */ 1572 do { 1573 pqmask = (pqmask << 1) | 1; 1574 for (i = 0; i <= pqmask; ++i) { 1575 pqi = (index & ~pqmask) | ((index + i) & pqmask); 1576 m = TAILQ_FIRST(&pq[pqi].pl); 1577 if (m) { 1578 _vm_page_and_queue_spin_lock(m); 1579 if (m->queue == basequeue + pqi) { 1580 _vm_page_rem_queue_spinlocked(m); 1581 return(m); 1582 } 1583 _vm_page_and_queue_spin_unlock(m); 1584 --i; 1585 continue; 1586 } 1587 } 1588 } while (pqmask != PQ_L2_MASK); 1589 1590 return(m); 1591 } 1592 1593 /* 1594 * Returns a vm_page candidate for allocation. The page is not busied so 1595 * it can move around. The caller must busy the page (and typically 1596 * deactivate it if it cannot be busied!) 1597 * 1598 * Returns a spinlocked vm_page that has been removed from its queue. 1599 */ 1600 vm_page_t 1601 vm_page_list_find(int basequeue, int index) 1602 { 1603 return(_vm_page_list_find(basequeue, index)); 1604 } 1605 1606 /* 1607 * Find a page on the cache queue with color optimization, remove it 1608 * from the queue, and busy it. The returned page will not be spinlocked. 1609 * 1610 * A candidate failure will be deactivated. Candidates can fail due to 1611 * being busied by someone else, in which case they will be deactivated. 1612 * 1613 * This routine may not block. 1614 * 1615 */ 1616 static vm_page_t 1617 vm_page_select_cache(u_short pg_color) 1618 { 1619 vm_page_t m; 1620 1621 for (;;) { 1622 m = _vm_page_list_find(PQ_CACHE, pg_color & PQ_L2_MASK); 1623 if (m == NULL) 1624 break; 1625 /* 1626 * (m) has been removed from its queue and spinlocked 1627 */ 1628 if (vm_page_busy_try(m, TRUE)) { 1629 _vm_page_deactivate_locked(m, 0); 1630 vm_page_spin_unlock(m); 1631 } else { 1632 /* 1633 * We successfully busied the page 1634 */ 1635 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) == 0 && 1636 m->hold_count == 0 && 1637 m->wire_count == 0 && 1638 (m->dirty & m->valid) == 0) { 1639 vm_page_spin_unlock(m); 1640 pagedaemon_wakeup(); 1641 return(m); 1642 } 1643 1644 /* 1645 * The page cannot be recycled, deactivate it. 1646 */ 1647 _vm_page_deactivate_locked(m, 0); 1648 if (_vm_page_wakeup(m)) { 1649 vm_page_spin_unlock(m); 1650 wakeup(m); 1651 } else { 1652 vm_page_spin_unlock(m); 1653 } 1654 } 1655 } 1656 return (m); 1657 } 1658 1659 /* 1660 * Find a free page. We attempt to inline the nominal case and fall back 1661 * to _vm_page_select_free() otherwise. A busied page is removed from 1662 * the queue and returned. 1663 * 1664 * This routine may not block. 1665 */ 1666 static __inline vm_page_t 1667 vm_page_select_free(u_short pg_color) 1668 { 1669 vm_page_t m; 1670 1671 for (;;) { 1672 m = _vm_page_list_find(PQ_FREE, pg_color & PQ_L2_MASK); 1673 if (m == NULL) 1674 break; 1675 if (vm_page_busy_try(m, TRUE)) { 1676 /* 1677 * Various mechanisms such as a pmap_collect can 1678 * result in a busy page on the free queue. We 1679 * have to move the page out of the way so we can 1680 * retry the allocation. If the other thread is not 1681 * allocating the page then m->valid will remain 0 and 1682 * the pageout daemon will free the page later on. 1683 * 1684 * Since we could not busy the page, however, we 1685 * cannot make assumptions as to whether the page 1686 * will be allocated by the other thread or not, 1687 * so all we can do is deactivate it to move it out 1688 * of the way. In particular, if the other thread 1689 * wires the page it may wind up on the inactive 1690 * queue and the pageout daemon will have to deal 1691 * with that case too. 1692 */ 1693 _vm_page_deactivate_locked(m, 0); 1694 vm_page_spin_unlock(m); 1695 } else { 1696 /* 1697 * Theoretically if we are able to busy the page 1698 * atomic with the queue removal (using the vm_page 1699 * lock) nobody else should be able to mess with the 1700 * page before us. 1701 */ 1702 KKASSERT((m->flags & (PG_UNMANAGED | 1703 PG_NEED_COMMIT)) == 0); 1704 KASSERT(m->hold_count == 0, ("m->hold_count is not zero " 1705 "pg %p q=%d flags=%08x hold=%d wire=%d", 1706 m, m->queue, m->flags, m->hold_count, m->wire_count)); 1707 KKASSERT(m->wire_count == 0); 1708 vm_page_spin_unlock(m); 1709 pagedaemon_wakeup(); 1710 1711 /* return busied and removed page */ 1712 return(m); 1713 } 1714 } 1715 return(m); 1716 } 1717 1718 /* 1719 * vm_page_alloc() 1720 * 1721 * Allocate and return a memory cell associated with this VM object/offset 1722 * pair. If object is NULL an unassociated page will be allocated. 1723 * 1724 * The returned page will be busied and removed from its queues. This 1725 * routine can block and may return NULL if a race occurs and the page 1726 * is found to already exist at the specified (object, pindex). 1727 * 1728 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain 1729 * VM_ALLOC_QUICK like normal but cannot use cache 1730 * VM_ALLOC_SYSTEM greater free drain 1731 * VM_ALLOC_INTERRUPT allow free list to be completely drained 1732 * VM_ALLOC_ZERO advisory request for pre-zero'd page only 1733 * VM_ALLOC_FORCE_ZERO advisory request for pre-zero'd page only 1734 * VM_ALLOC_NULL_OK ok to return NULL on insertion collision 1735 * (see vm_page_grab()) 1736 * VM_ALLOC_USE_GD ok to use per-gd cache 1737 * 1738 * VM_ALLOC_CPU(n) allocate using specified cpu localization 1739 * 1740 * The object must be held if not NULL 1741 * This routine may not block 1742 * 1743 * Additional special handling is required when called from an interrupt 1744 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache 1745 * in this case. 1746 */ 1747 vm_page_t 1748 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req) 1749 { 1750 globaldata_t gd; 1751 vm_object_t obj; 1752 vm_page_t m; 1753 u_short pg_color; 1754 int cpuid_local; 1755 1756 #if 0 1757 /* 1758 * Special per-cpu free VM page cache. The pages are pre-busied 1759 * and pre-zerod for us. 1760 */ 1761 if (gd->gd_vmpg_count && (page_req & VM_ALLOC_USE_GD)) { 1762 crit_enter_gd(gd); 1763 if (gd->gd_vmpg_count) { 1764 m = gd->gd_vmpg_array[--gd->gd_vmpg_count]; 1765 crit_exit_gd(gd); 1766 goto done; 1767 } 1768 crit_exit_gd(gd); 1769 } 1770 #endif 1771 m = NULL; 1772 1773 /* 1774 * CPU LOCALIZATION 1775 * 1776 * CPU localization algorithm. Break the page queues up by physical 1777 * id and core id (note that two cpu threads will have the same core 1778 * id, and core_id != gd_cpuid). 1779 * 1780 * This is nowhere near perfect, for example the last pindex in a 1781 * subgroup will overflow into the next cpu or package. But this 1782 * should get us good page reuse locality in heavy mixed loads. 1783 * 1784 * (may be executed before the APs are started, so other GDs might 1785 * not exist!) 1786 */ 1787 if (page_req & VM_ALLOC_CPU_SPEC) 1788 cpuid_local = VM_ALLOC_GETCPU(page_req); 1789 else 1790 cpuid_local = mycpu->gd_cpuid; 1791 1792 pg_color = vm_get_pg_color(cpuid_local, object, pindex); 1793 1794 KKASSERT(page_req & 1795 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK| 1796 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 1797 1798 /* 1799 * Certain system threads (pageout daemon, buf_daemon's) are 1800 * allowed to eat deeper into the free page list. 1801 */ 1802 if (curthread->td_flags & TDF_SYSTHREAD) 1803 page_req |= VM_ALLOC_SYSTEM; 1804 1805 /* 1806 * Impose various limitations. Note that the v_free_reserved test 1807 * must match the opposite of vm_page_count_target() to avoid 1808 * livelocks, be careful. 1809 */ 1810 loop: 1811 gd = mycpu; 1812 if (gd->gd_vmstats.v_free_count >= gd->gd_vmstats.v_free_reserved || 1813 ((page_req & VM_ALLOC_INTERRUPT) && 1814 gd->gd_vmstats.v_free_count > 0) || 1815 ((page_req & VM_ALLOC_SYSTEM) && 1816 gd->gd_vmstats.v_cache_count == 0 && 1817 gd->gd_vmstats.v_free_count > 1818 gd->gd_vmstats.v_interrupt_free_min) 1819 ) { 1820 /* 1821 * The free queue has sufficient free pages to take one out. 1822 */ 1823 m = vm_page_select_free(pg_color); 1824 } else if (page_req & VM_ALLOC_NORMAL) { 1825 /* 1826 * Allocatable from the cache (non-interrupt only). On 1827 * success, we must free the page and try again, thus 1828 * ensuring that vmstats.v_*_free_min counters are replenished. 1829 */ 1830 #ifdef INVARIANTS 1831 if (curthread->td_preempted) { 1832 kprintf("vm_page_alloc(): warning, attempt to allocate" 1833 " cache page from preempting interrupt\n"); 1834 m = NULL; 1835 } else { 1836 m = vm_page_select_cache(pg_color); 1837 } 1838 #else 1839 m = vm_page_select_cache(pg_color); 1840 #endif 1841 /* 1842 * On success move the page into the free queue and loop. 1843 * 1844 * Only do this if we can safely acquire the vm_object lock, 1845 * because this is effectively a random page and the caller 1846 * might be holding the lock shared, we don't want to 1847 * deadlock. 1848 */ 1849 if (m != NULL) { 1850 KASSERT(m->dirty == 0, 1851 ("Found dirty cache page %p", m)); 1852 if ((obj = m->object) != NULL) { 1853 if (vm_object_hold_try(obj)) { 1854 vm_page_protect(m, VM_PROT_NONE); 1855 vm_page_free(m); 1856 /* m->object NULL here */ 1857 vm_object_drop(obj); 1858 } else { 1859 vm_page_deactivate(m); 1860 vm_page_wakeup(m); 1861 } 1862 } else { 1863 vm_page_protect(m, VM_PROT_NONE); 1864 vm_page_free(m); 1865 } 1866 goto loop; 1867 } 1868 1869 /* 1870 * On failure return NULL 1871 */ 1872 atomic_add_int(&vm_pageout_deficit, 1); 1873 pagedaemon_wakeup(); 1874 return (NULL); 1875 } else { 1876 /* 1877 * No pages available, wakeup the pageout daemon and give up. 1878 */ 1879 atomic_add_int(&vm_pageout_deficit, 1); 1880 pagedaemon_wakeup(); 1881 return (NULL); 1882 } 1883 1884 /* 1885 * v_free_count can race so loop if we don't find the expected 1886 * page. 1887 */ 1888 if (m == NULL) { 1889 vmstats_rollup(); 1890 goto loop; 1891 } 1892 1893 /* 1894 * Good page found. The page has already been busied for us and 1895 * removed from its queues. 1896 */ 1897 KASSERT(m->dirty == 0, 1898 ("vm_page_alloc: free/cache page %p was dirty", m)); 1899 KKASSERT(m->queue == PQ_NONE); 1900 1901 #if 0 1902 done: 1903 #endif 1904 /* 1905 * Initialize the structure, inheriting some flags but clearing 1906 * all the rest. The page has already been busied for us. 1907 */ 1908 vm_page_flag_clear(m, ~PG_KEEP_NEWPAGE_MASK); 1909 1910 KKASSERT(m->wire_count == 0); 1911 KKASSERT((m->busy_count & PBUSY_MASK) == 0); 1912 m->act_count = 0; 1913 m->valid = 0; 1914 1915 /* 1916 * Caller must be holding the object lock (asserted by 1917 * vm_page_insert()). 1918 * 1919 * NOTE: Inserting a page here does not insert it into any pmaps 1920 * (which could cause us to block allocating memory). 1921 * 1922 * NOTE: If no object an unassociated page is allocated, m->pindex 1923 * can be used by the caller for any purpose. 1924 */ 1925 if (object) { 1926 if (vm_page_insert(m, object, pindex) == FALSE) { 1927 vm_page_free(m); 1928 if ((page_req & VM_ALLOC_NULL_OK) == 0) 1929 panic("PAGE RACE %p[%ld]/%p", 1930 object, (long)pindex, m); 1931 m = NULL; 1932 } 1933 } else { 1934 m->pindex = pindex; 1935 } 1936 1937 /* 1938 * Don't wakeup too often - wakeup the pageout daemon when 1939 * we would be nearly out of memory. 1940 */ 1941 pagedaemon_wakeup(); 1942 1943 /* 1944 * A BUSY page is returned. 1945 */ 1946 return (m); 1947 } 1948 1949 /* 1950 * Returns number of pages available in our DMA memory reserve 1951 * (adjusted with vm.dma_reserved=<value>m in /boot/loader.conf) 1952 */ 1953 vm_size_t 1954 vm_contig_avail_pages(void) 1955 { 1956 alist_blk_t blk; 1957 alist_blk_t count; 1958 alist_blk_t bfree; 1959 spin_lock(&vm_contig_spin); 1960 bfree = alist_free_info(&vm_contig_alist, &blk, &count); 1961 spin_unlock(&vm_contig_spin); 1962 1963 return bfree; 1964 } 1965 1966 /* 1967 * Attempt to allocate contiguous physical memory with the specified 1968 * requirements. 1969 */ 1970 vm_page_t 1971 vm_page_alloc_contig(vm_paddr_t low, vm_paddr_t high, 1972 unsigned long alignment, unsigned long boundary, 1973 unsigned long size, vm_memattr_t memattr) 1974 { 1975 alist_blk_t blk; 1976 vm_page_t m; 1977 vm_pindex_t i; 1978 #if 0 1979 static vm_pindex_t contig_rover; 1980 #endif 1981 1982 alignment >>= PAGE_SHIFT; 1983 if (alignment == 0) 1984 alignment = 1; 1985 boundary >>= PAGE_SHIFT; 1986 if (boundary == 0) 1987 boundary = 1; 1988 size = (size + PAGE_MASK) >> PAGE_SHIFT; 1989 1990 #if 0 1991 /* 1992 * Disabled temporarily until we find a solution for DRM (a flag 1993 * to always use the free space reserve, for performance). 1994 */ 1995 if (high == BUS_SPACE_MAXADDR && alignment <= PAGE_SIZE && 1996 boundary <= PAGE_SIZE && size == 1 && 1997 memattr == VM_MEMATTR_DEFAULT) { 1998 /* 1999 * Any page will work, use vm_page_alloc() 2000 * (e.g. when used from kmem_alloc_attr()) 2001 */ 2002 m = vm_page_alloc(NULL, (contig_rover++) & 0x7FFFFFFF, 2003 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM | 2004 VM_ALLOC_INTERRUPT); 2005 m->valid = VM_PAGE_BITS_ALL; 2006 vm_page_wire(m); 2007 vm_page_wakeup(m); 2008 } else 2009 #endif 2010 { 2011 /* 2012 * Use the low-memory dma reserve 2013 */ 2014 spin_lock(&vm_contig_spin); 2015 blk = alist_alloc(&vm_contig_alist, 0, size); 2016 if (blk == ALIST_BLOCK_NONE) { 2017 spin_unlock(&vm_contig_spin); 2018 if (bootverbose) { 2019 kprintf("vm_page_alloc_contig: %ldk nospace\n", 2020 (size << PAGE_SHIFT) / 1024); 2021 print_backtrace(5); 2022 } 2023 return(NULL); 2024 } 2025 if (high && ((vm_paddr_t)(blk + size) << PAGE_SHIFT) > high) { 2026 alist_free(&vm_contig_alist, blk, size); 2027 spin_unlock(&vm_contig_spin); 2028 if (bootverbose) { 2029 kprintf("vm_page_alloc_contig: %ldk high " 2030 "%016jx failed\n", 2031 (size << PAGE_SHIFT) / 1024, 2032 (intmax_t)high); 2033 } 2034 return(NULL); 2035 } 2036 spin_unlock(&vm_contig_spin); 2037 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT); 2038 } 2039 if (vm_contig_verbose) { 2040 kprintf("vm_page_alloc_contig: %016jx/%ldk " 2041 "(%016jx-%016jx al=%lu bo=%lu pgs=%lu attr=%d\n", 2042 (intmax_t)m->phys_addr, 2043 (size << PAGE_SHIFT) / 1024, 2044 low, high, alignment, boundary, size, memattr); 2045 } 2046 if (memattr != VM_MEMATTR_DEFAULT) { 2047 for (i = 0;i < size; i++) 2048 pmap_page_set_memattr(&m[i], memattr); 2049 } 2050 return m; 2051 } 2052 2053 /* 2054 * Free contiguously allocated pages. The pages will be wired but not busy. 2055 * When freeing to the alist we leave them wired and not busy. 2056 */ 2057 void 2058 vm_page_free_contig(vm_page_t m, unsigned long size) 2059 { 2060 vm_paddr_t pa = VM_PAGE_TO_PHYS(m); 2061 vm_pindex_t start = pa >> PAGE_SHIFT; 2062 vm_pindex_t pages = (size + PAGE_MASK) >> PAGE_SHIFT; 2063 2064 if (vm_contig_verbose) { 2065 kprintf("vm_page_free_contig: %016jx/%ldk\n", 2066 (intmax_t)pa, size / 1024); 2067 } 2068 if (pa < vm_low_phys_reserved) { 2069 KKASSERT(pa + size <= vm_low_phys_reserved); 2070 spin_lock(&vm_contig_spin); 2071 alist_free(&vm_contig_alist, start, pages); 2072 spin_unlock(&vm_contig_spin); 2073 } else { 2074 while (pages) { 2075 vm_page_busy_wait(m, FALSE, "cpgfr"); 2076 vm_page_unwire(m, 0); 2077 vm_page_free(m); 2078 --pages; 2079 ++m; 2080 } 2081 2082 } 2083 } 2084 2085 2086 /* 2087 * Wait for sufficient free memory for nominal heavy memory use kernel 2088 * operations. 2089 * 2090 * WARNING! Be sure never to call this in any vm_pageout code path, which 2091 * will trivially deadlock the system. 2092 */ 2093 void 2094 vm_wait_nominal(void) 2095 { 2096 while (vm_page_count_min(0)) 2097 vm_wait(0); 2098 } 2099 2100 /* 2101 * Test if vm_wait_nominal() would block. 2102 */ 2103 int 2104 vm_test_nominal(void) 2105 { 2106 if (vm_page_count_min(0)) 2107 return(1); 2108 return(0); 2109 } 2110 2111 /* 2112 * Block until free pages are available for allocation, called in various 2113 * places before memory allocations. 2114 * 2115 * The caller may loop if vm_page_count_min() == FALSE so we cannot be 2116 * more generous then that. 2117 */ 2118 void 2119 vm_wait(int timo) 2120 { 2121 /* 2122 * never wait forever 2123 */ 2124 if (timo == 0) 2125 timo = hz; 2126 lwkt_gettoken(&vm_token); 2127 2128 if (curthread == pagethread || 2129 curthread == emergpager) { 2130 /* 2131 * The pageout daemon itself needs pages, this is bad. 2132 */ 2133 if (vm_page_count_min(0)) { 2134 vm_pageout_pages_needed = 1; 2135 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo); 2136 } 2137 } else { 2138 /* 2139 * Wakeup the pageout daemon if necessary and wait. 2140 * 2141 * Do not wait indefinitely for the target to be reached, 2142 * as load might prevent it from being reached any time soon. 2143 * But wait a little to try to slow down page allocations 2144 * and to give more important threads (the pagedaemon) 2145 * allocation priority. 2146 */ 2147 if (vm_page_count_target()) { 2148 if (vm_pages_needed == 0) { 2149 vm_pages_needed = 1; 2150 wakeup(&vm_pages_needed); 2151 } 2152 ++vm_pages_waiting; /* SMP race ok */ 2153 tsleep(&vmstats.v_free_count, 0, "vmwait", timo); 2154 } 2155 } 2156 lwkt_reltoken(&vm_token); 2157 } 2158 2159 /* 2160 * Block until free pages are available for allocation 2161 * 2162 * Called only from vm_fault so that processes page faulting can be 2163 * easily tracked. 2164 */ 2165 void 2166 vm_wait_pfault(void) 2167 { 2168 /* 2169 * Wakeup the pageout daemon if necessary and wait. 2170 * 2171 * Do not wait indefinitely for the target to be reached, 2172 * as load might prevent it from being reached any time soon. 2173 * But wait a little to try to slow down page allocations 2174 * and to give more important threads (the pagedaemon) 2175 * allocation priority. 2176 */ 2177 if (vm_page_count_min(0)) { 2178 lwkt_gettoken(&vm_token); 2179 while (vm_page_count_severe()) { 2180 if (vm_page_count_target()) { 2181 thread_t td; 2182 2183 if (vm_pages_needed == 0) { 2184 vm_pages_needed = 1; 2185 wakeup(&vm_pages_needed); 2186 } 2187 ++vm_pages_waiting; /* SMP race ok */ 2188 tsleep(&vmstats.v_free_count, 0, "pfault", hz); 2189 2190 /* 2191 * Do not stay stuck in the loop if the system is trying 2192 * to kill the process. 2193 */ 2194 td = curthread; 2195 if (td->td_proc && (td->td_proc->p_flags & P_LOWMEMKILL)) 2196 break; 2197 } 2198 } 2199 lwkt_reltoken(&vm_token); 2200 } 2201 } 2202 2203 /* 2204 * Put the specified page on the active list (if appropriate). Ensure 2205 * that act_count is at least ACT_INIT but do not otherwise mess with it. 2206 * 2207 * The caller should be holding the page busied ? XXX 2208 * This routine may not block. 2209 */ 2210 void 2211 vm_page_activate(vm_page_t m) 2212 { 2213 u_short oqueue; 2214 2215 vm_page_spin_lock(m); 2216 if (m->queue - m->pc != PQ_ACTIVE) { 2217 _vm_page_queue_spin_lock(m); 2218 oqueue = _vm_page_rem_queue_spinlocked(m); 2219 /* page is left spinlocked, queue is unlocked */ 2220 2221 if (oqueue == PQ_CACHE) 2222 mycpu->gd_cnt.v_reactivated++; 2223 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 2224 if (m->act_count < ACT_INIT) 2225 m->act_count = ACT_INIT; 2226 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE + m->pc, 0); 2227 } 2228 _vm_page_and_queue_spin_unlock(m); 2229 if (oqueue == PQ_CACHE || oqueue == PQ_FREE) 2230 pagedaemon_wakeup(); 2231 } else { 2232 if (m->act_count < ACT_INIT) 2233 m->act_count = ACT_INIT; 2234 vm_page_spin_unlock(m); 2235 } 2236 } 2237 2238 /* 2239 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 2240 * routine is called when a page has been added to the cache or free 2241 * queues. 2242 * 2243 * This routine may not block. 2244 */ 2245 static __inline void 2246 vm_page_free_wakeup(void) 2247 { 2248 globaldata_t gd = mycpu; 2249 2250 /* 2251 * If the pageout daemon itself needs pages, then tell it that 2252 * there are some free. 2253 */ 2254 if (vm_pageout_pages_needed && 2255 gd->gd_vmstats.v_cache_count + gd->gd_vmstats.v_free_count >= 2256 gd->gd_vmstats.v_pageout_free_min 2257 ) { 2258 vm_pageout_pages_needed = 0; 2259 wakeup(&vm_pageout_pages_needed); 2260 } 2261 2262 /* 2263 * Wakeup processes that are waiting on memory. 2264 * 2265 * Generally speaking we want to wakeup stuck processes as soon as 2266 * possible. !vm_page_count_min(0) is the absolute minimum point 2267 * where we can do this. Wait a bit longer to reduce degenerate 2268 * re-blocking (vm_page_free_hysteresis). The target check is just 2269 * to make sure the min-check w/hysteresis does not exceed the 2270 * normal target. 2271 */ 2272 if (vm_pages_waiting) { 2273 if (!vm_page_count_min(vm_page_free_hysteresis) || 2274 !vm_page_count_target()) { 2275 vm_pages_waiting = 0; 2276 wakeup(&vmstats.v_free_count); 2277 ++mycpu->gd_cnt.v_ppwakeups; 2278 } 2279 #if 0 2280 if (!vm_page_count_target()) { 2281 /* 2282 * Plenty of pages are free, wakeup everyone. 2283 */ 2284 vm_pages_waiting = 0; 2285 wakeup(&vmstats.v_free_count); 2286 ++mycpu->gd_cnt.v_ppwakeups; 2287 } else if (!vm_page_count_min(0)) { 2288 /* 2289 * Some pages are free, wakeup someone. 2290 */ 2291 int wcount = vm_pages_waiting; 2292 if (wcount > 0) 2293 --wcount; 2294 vm_pages_waiting = wcount; 2295 wakeup_one(&vmstats.v_free_count); 2296 ++mycpu->gd_cnt.v_ppwakeups; 2297 } 2298 #endif 2299 } 2300 } 2301 2302 /* 2303 * Returns the given page to the PQ_FREE or PQ_HOLD list and disassociates 2304 * it from its VM object. 2305 * 2306 * The vm_page must be BUSY on entry. BUSY will be released on 2307 * return (the page will have been freed). 2308 */ 2309 void 2310 vm_page_free_toq(vm_page_t m) 2311 { 2312 mycpu->gd_cnt.v_tfree++; 2313 KKASSERT((m->flags & PG_MAPPED) == 0); 2314 KKASSERT(m->busy_count & PBUSY_LOCKED); 2315 2316 if ((m->busy_count & PBUSY_MASK) || ((m->queue - m->pc) == PQ_FREE)) { 2317 kprintf("vm_page_free: pindex(%lu), busy %08x, " 2318 "hold(%d)\n", 2319 (u_long)m->pindex, m->busy_count, m->hold_count); 2320 if ((m->queue - m->pc) == PQ_FREE) 2321 panic("vm_page_free: freeing free page"); 2322 else 2323 panic("vm_page_free: freeing busy page"); 2324 } 2325 2326 /* 2327 * Remove from object, spinlock the page and its queues and 2328 * remove from any queue. No queue spinlock will be held 2329 * after this section (because the page was removed from any 2330 * queue). 2331 */ 2332 vm_page_remove(m); 2333 vm_page_and_queue_spin_lock(m); 2334 _vm_page_rem_queue_spinlocked(m); 2335 2336 /* 2337 * No further management of fictitious pages occurs beyond object 2338 * and queue removal. 2339 */ 2340 if ((m->flags & PG_FICTITIOUS) != 0) { 2341 vm_page_spin_unlock(m); 2342 vm_page_wakeup(m); 2343 return; 2344 } 2345 2346 m->valid = 0; 2347 vm_page_undirty(m); 2348 2349 if (m->wire_count != 0) { 2350 if (m->wire_count > 1) { 2351 panic( 2352 "vm_page_free: invalid wire count (%d), pindex: 0x%lx", 2353 m->wire_count, (long)m->pindex); 2354 } 2355 panic("vm_page_free: freeing wired page"); 2356 } 2357 2358 /* 2359 * Clear the UNMANAGED flag when freeing an unmanaged page. 2360 * Clear the NEED_COMMIT flag 2361 */ 2362 if (m->flags & PG_UNMANAGED) 2363 vm_page_flag_clear(m, PG_UNMANAGED); 2364 if (m->flags & PG_NEED_COMMIT) 2365 vm_page_flag_clear(m, PG_NEED_COMMIT); 2366 2367 if (m->hold_count != 0) { 2368 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0); 2369 } else { 2370 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 1); 2371 } 2372 2373 /* 2374 * This sequence allows us to clear BUSY while still holding 2375 * its spin lock, which reduces contention vs allocators. We 2376 * must not leave the queue locked or _vm_page_wakeup() may 2377 * deadlock. 2378 */ 2379 _vm_page_queue_spin_unlock(m); 2380 if (_vm_page_wakeup(m)) { 2381 vm_page_spin_unlock(m); 2382 wakeup(m); 2383 } else { 2384 vm_page_spin_unlock(m); 2385 } 2386 vm_page_free_wakeup(); 2387 } 2388 2389 /* 2390 * vm_page_unmanage() 2391 * 2392 * Prevent PV management from being done on the page. The page is 2393 * removed from the paging queues as if it were wired, and as a 2394 * consequence of no longer being managed the pageout daemon will not 2395 * touch it (since there is no way to locate the pte mappings for the 2396 * page). madvise() calls that mess with the pmap will also no longer 2397 * operate on the page. 2398 * 2399 * Beyond that the page is still reasonably 'normal'. Freeing the page 2400 * will clear the flag. 2401 * 2402 * This routine is used by OBJT_PHYS objects - objects using unswappable 2403 * physical memory as backing store rather then swap-backed memory and 2404 * will eventually be extended to support 4MB unmanaged physical 2405 * mappings. 2406 * 2407 * Caller must be holding the page busy. 2408 */ 2409 void 2410 vm_page_unmanage(vm_page_t m) 2411 { 2412 KKASSERT(m->busy_count & PBUSY_LOCKED); 2413 if ((m->flags & PG_UNMANAGED) == 0) { 2414 if (m->wire_count == 0) 2415 vm_page_unqueue(m); 2416 } 2417 vm_page_flag_set(m, PG_UNMANAGED); 2418 } 2419 2420 /* 2421 * Mark this page as wired down by yet another map, removing it from 2422 * paging queues as necessary. 2423 * 2424 * Caller must be holding the page busy. 2425 */ 2426 void 2427 vm_page_wire(vm_page_t m) 2428 { 2429 /* 2430 * Only bump the wire statistics if the page is not already wired, 2431 * and only unqueue the page if it is on some queue (if it is unmanaged 2432 * it is already off the queues). Don't do anything with fictitious 2433 * pages because they are always wired. 2434 */ 2435 KKASSERT(m->busy_count & PBUSY_LOCKED); 2436 if ((m->flags & PG_FICTITIOUS) == 0) { 2437 if (atomic_fetchadd_int(&m->wire_count, 1) == 0) { 2438 if ((m->flags & PG_UNMANAGED) == 0) 2439 vm_page_unqueue(m); 2440 atomic_add_long(&mycpu->gd_vmstats_adj.v_wire_count, 1); 2441 } 2442 KASSERT(m->wire_count != 0, 2443 ("vm_page_wire: wire_count overflow m=%p", m)); 2444 } 2445 } 2446 2447 /* 2448 * Release one wiring of this page, potentially enabling it to be paged again. 2449 * 2450 * Many pages placed on the inactive queue should actually go 2451 * into the cache, but it is difficult to figure out which. What 2452 * we do instead, if the inactive target is well met, is to put 2453 * clean pages at the head of the inactive queue instead of the tail. 2454 * This will cause them to be moved to the cache more quickly and 2455 * if not actively re-referenced, freed more quickly. If we just 2456 * stick these pages at the end of the inactive queue, heavy filesystem 2457 * meta-data accesses can cause an unnecessary paging load on memory bound 2458 * processes. This optimization causes one-time-use metadata to be 2459 * reused more quickly. 2460 * 2461 * Pages marked PG_NEED_COMMIT are always activated and never placed on 2462 * the inactive queue. This helps the pageout daemon determine memory 2463 * pressure and act on out-of-memory situations more quickly. 2464 * 2465 * BUT, if we are in a low-memory situation we have no choice but to 2466 * put clean pages on the cache queue. 2467 * 2468 * A number of routines use vm_page_unwire() to guarantee that the page 2469 * will go into either the inactive or active queues, and will NEVER 2470 * be placed in the cache - for example, just after dirtying a page. 2471 * dirty pages in the cache are not allowed. 2472 * 2473 * This routine may not block. 2474 */ 2475 void 2476 vm_page_unwire(vm_page_t m, int activate) 2477 { 2478 KKASSERT(m->busy_count & PBUSY_LOCKED); 2479 if (m->flags & PG_FICTITIOUS) { 2480 /* do nothing */ 2481 } else if (m->wire_count <= 0) { 2482 panic("vm_page_unwire: invalid wire count: %d", m->wire_count); 2483 } else { 2484 if (atomic_fetchadd_int(&m->wire_count, -1) == 1) { 2485 atomic_add_long(&mycpu->gd_vmstats_adj.v_wire_count,-1); 2486 if (m->flags & PG_UNMANAGED) { 2487 ; 2488 } else if (activate || (m->flags & PG_NEED_COMMIT)) { 2489 vm_page_spin_lock(m); 2490 _vm_page_add_queue_spinlocked(m, 2491 PQ_ACTIVE + m->pc, 0); 2492 _vm_page_and_queue_spin_unlock(m); 2493 } else { 2494 vm_page_spin_lock(m); 2495 vm_page_flag_clear(m, PG_WINATCFLS); 2496 _vm_page_add_queue_spinlocked(m, 2497 PQ_INACTIVE + m->pc, 0); 2498 ++vm_swapcache_inactive_heuristic; 2499 _vm_page_and_queue_spin_unlock(m); 2500 } 2501 } 2502 } 2503 } 2504 2505 /* 2506 * Move the specified page to the inactive queue. If the page has 2507 * any associated swap, the swap is deallocated. 2508 * 2509 * Normally athead is 0 resulting in LRU operation. athead is set 2510 * to 1 if we want this page to be 'as if it were placed in the cache', 2511 * except without unmapping it from the process address space. 2512 * 2513 * vm_page's spinlock must be held on entry and will remain held on return. 2514 * This routine may not block. 2515 */ 2516 static void 2517 _vm_page_deactivate_locked(vm_page_t m, int athead) 2518 { 2519 u_short oqueue; 2520 2521 /* 2522 * Ignore if already inactive. 2523 */ 2524 if (m->queue - m->pc == PQ_INACTIVE) 2525 return; 2526 _vm_page_queue_spin_lock(m); 2527 oqueue = _vm_page_rem_queue_spinlocked(m); 2528 2529 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 2530 if (oqueue == PQ_CACHE) 2531 mycpu->gd_cnt.v_reactivated++; 2532 vm_page_flag_clear(m, PG_WINATCFLS); 2533 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE + m->pc, athead); 2534 if (athead == 0) 2535 ++vm_swapcache_inactive_heuristic; 2536 } 2537 /* NOTE: PQ_NONE if condition not taken */ 2538 _vm_page_queue_spin_unlock(m); 2539 /* leaves vm_page spinlocked */ 2540 } 2541 2542 /* 2543 * Attempt to deactivate a page. 2544 * 2545 * No requirements. 2546 */ 2547 void 2548 vm_page_deactivate(vm_page_t m) 2549 { 2550 vm_page_spin_lock(m); 2551 _vm_page_deactivate_locked(m, 0); 2552 vm_page_spin_unlock(m); 2553 } 2554 2555 void 2556 vm_page_deactivate_locked(vm_page_t m) 2557 { 2558 _vm_page_deactivate_locked(m, 0); 2559 } 2560 2561 /* 2562 * Attempt to move a busied page to PQ_CACHE, then unconditionally unbusy it. 2563 * 2564 * This function returns non-zero if it successfully moved the page to 2565 * PQ_CACHE. 2566 * 2567 * This function unconditionally unbusies the page on return. 2568 */ 2569 int 2570 vm_page_try_to_cache(vm_page_t m) 2571 { 2572 vm_page_spin_lock(m); 2573 if (m->dirty || m->hold_count || m->wire_count || 2574 (m->flags & (PG_UNMANAGED | PG_NEED_COMMIT))) { 2575 if (_vm_page_wakeup(m)) { 2576 vm_page_spin_unlock(m); 2577 wakeup(m); 2578 } else { 2579 vm_page_spin_unlock(m); 2580 } 2581 return(0); 2582 } 2583 vm_page_spin_unlock(m); 2584 2585 /* 2586 * Page busied by us and no longer spinlocked. Dirty pages cannot 2587 * be moved to the cache. 2588 */ 2589 vm_page_test_dirty(m); 2590 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2591 vm_page_wakeup(m); 2592 return(0); 2593 } 2594 vm_page_cache(m); 2595 return(1); 2596 } 2597 2598 /* 2599 * Attempt to free the page. If we cannot free it, we do nothing. 2600 * 1 is returned on success, 0 on failure. 2601 * 2602 * No requirements. 2603 */ 2604 int 2605 vm_page_try_to_free(vm_page_t m) 2606 { 2607 vm_page_spin_lock(m); 2608 if (vm_page_busy_try(m, TRUE)) { 2609 vm_page_spin_unlock(m); 2610 return(0); 2611 } 2612 2613 /* 2614 * The page can be in any state, including already being on the free 2615 * queue. Check to see if it really can be freed. 2616 */ 2617 if (m->dirty || /* can't free if it is dirty */ 2618 m->hold_count || /* or held (XXX may be wrong) */ 2619 m->wire_count || /* or wired */ 2620 (m->flags & (PG_UNMANAGED | /* or unmanaged */ 2621 PG_NEED_COMMIT)) || /* or needs a commit */ 2622 m->queue - m->pc == PQ_FREE || /* already on PQ_FREE */ 2623 m->queue - m->pc == PQ_HOLD) { /* already on PQ_HOLD */ 2624 if (_vm_page_wakeup(m)) { 2625 vm_page_spin_unlock(m); 2626 wakeup(m); 2627 } else { 2628 vm_page_spin_unlock(m); 2629 } 2630 return(0); 2631 } 2632 vm_page_spin_unlock(m); 2633 2634 /* 2635 * We can probably free the page. 2636 * 2637 * Page busied by us and no longer spinlocked. Dirty pages will 2638 * not be freed by this function. We have to re-test the 2639 * dirty bit after cleaning out the pmaps. 2640 */ 2641 vm_page_test_dirty(m); 2642 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2643 vm_page_wakeup(m); 2644 return(0); 2645 } 2646 vm_page_protect(m, VM_PROT_NONE); 2647 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2648 vm_page_wakeup(m); 2649 return(0); 2650 } 2651 vm_page_free(m); 2652 return(1); 2653 } 2654 2655 /* 2656 * vm_page_cache 2657 * 2658 * Put the specified page onto the page cache queue (if appropriate). 2659 * 2660 * The page must be busy, and this routine will release the busy and 2661 * possibly even free the page. 2662 */ 2663 void 2664 vm_page_cache(vm_page_t m) 2665 { 2666 /* 2667 * Not suitable for the cache 2668 */ 2669 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) || 2670 (m->busy_count & PBUSY_MASK) || 2671 m->wire_count || m->hold_count) { 2672 vm_page_wakeup(m); 2673 return; 2674 } 2675 2676 /* 2677 * Already in the cache (and thus not mapped) 2678 */ 2679 if ((m->queue - m->pc) == PQ_CACHE) { 2680 KKASSERT((m->flags & PG_MAPPED) == 0); 2681 vm_page_wakeup(m); 2682 return; 2683 } 2684 2685 /* 2686 * Caller is required to test m->dirty, but note that the act of 2687 * removing the page from its maps can cause it to become dirty 2688 * on an SMP system due to another cpu running in usermode. 2689 */ 2690 if (m->dirty) { 2691 panic("vm_page_cache: caching a dirty page, pindex: %ld", 2692 (long)m->pindex); 2693 } 2694 2695 /* 2696 * Remove all pmaps and indicate that the page is not 2697 * writeable or mapped. Our vm_page_protect() call may 2698 * have blocked (especially w/ VM_PROT_NONE), so recheck 2699 * everything. 2700 */ 2701 vm_page_protect(m, VM_PROT_NONE); 2702 if ((m->flags & (PG_UNMANAGED | PG_MAPPED)) || 2703 (m->busy_count & PBUSY_MASK) || 2704 m->wire_count || m->hold_count) { 2705 vm_page_wakeup(m); 2706 } else if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2707 vm_page_deactivate(m); 2708 vm_page_wakeup(m); 2709 } else { 2710 _vm_page_and_queue_spin_lock(m); 2711 _vm_page_rem_queue_spinlocked(m); 2712 _vm_page_add_queue_spinlocked(m, PQ_CACHE + m->pc, 0); 2713 _vm_page_queue_spin_unlock(m); 2714 if (_vm_page_wakeup(m)) { 2715 vm_page_spin_unlock(m); 2716 wakeup(m); 2717 } else { 2718 vm_page_spin_unlock(m); 2719 } 2720 vm_page_free_wakeup(); 2721 } 2722 } 2723 2724 /* 2725 * vm_page_dontneed() 2726 * 2727 * Cache, deactivate, or do nothing as appropriate. This routine 2728 * is typically used by madvise() MADV_DONTNEED. 2729 * 2730 * Generally speaking we want to move the page into the cache so 2731 * it gets reused quickly. However, this can result in a silly syndrome 2732 * due to the page recycling too quickly. Small objects will not be 2733 * fully cached. On the otherhand, if we move the page to the inactive 2734 * queue we wind up with a problem whereby very large objects 2735 * unnecessarily blow away our inactive and cache queues. 2736 * 2737 * The solution is to move the pages based on a fixed weighting. We 2738 * either leave them alone, deactivate them, or move them to the cache, 2739 * where moving them to the cache has the highest weighting. 2740 * By forcing some pages into other queues we eventually force the 2741 * system to balance the queues, potentially recovering other unrelated 2742 * space from active. The idea is to not force this to happen too 2743 * often. 2744 * 2745 * The page must be busied. 2746 */ 2747 void 2748 vm_page_dontneed(vm_page_t m) 2749 { 2750 static int dnweight; 2751 int dnw; 2752 int head; 2753 2754 dnw = ++dnweight; 2755 2756 /* 2757 * occassionally leave the page alone 2758 */ 2759 if ((dnw & 0x01F0) == 0 || 2760 m->queue - m->pc == PQ_INACTIVE || 2761 m->queue - m->pc == PQ_CACHE 2762 ) { 2763 if (m->act_count >= ACT_INIT) 2764 --m->act_count; 2765 return; 2766 } 2767 2768 /* 2769 * If vm_page_dontneed() is inactivating a page, it must clear 2770 * the referenced flag; otherwise the pagedaemon will see references 2771 * on the page in the inactive queue and reactivate it. Until the 2772 * page can move to the cache queue, madvise's job is not done. 2773 */ 2774 vm_page_flag_clear(m, PG_REFERENCED); 2775 pmap_clear_reference(m); 2776 2777 if (m->dirty == 0) 2778 vm_page_test_dirty(m); 2779 2780 if (m->dirty || (dnw & 0x0070) == 0) { 2781 /* 2782 * Deactivate the page 3 times out of 32. 2783 */ 2784 head = 0; 2785 } else { 2786 /* 2787 * Cache the page 28 times out of every 32. Note that 2788 * the page is deactivated instead of cached, but placed 2789 * at the head of the queue instead of the tail. 2790 */ 2791 head = 1; 2792 } 2793 vm_page_spin_lock(m); 2794 _vm_page_deactivate_locked(m, head); 2795 vm_page_spin_unlock(m); 2796 } 2797 2798 /* 2799 * These routines manipulate the 'soft busy' count for a page. A soft busy 2800 * is almost like a hard BUSY except that it allows certain compatible 2801 * operations to occur on the page while it is busy. For example, a page 2802 * undergoing a write can still be mapped read-only. 2803 * 2804 * We also use soft-busy to quickly pmap_enter shared read-only pages 2805 * without having to hold the page locked. 2806 * 2807 * The soft-busy count can be > 1 in situations where multiple threads 2808 * are pmap_enter()ing the same page simultaneously, or when two buffer 2809 * cache buffers overlap the same page. 2810 * 2811 * The caller must hold the page BUSY when making these two calls. 2812 */ 2813 void 2814 vm_page_io_start(vm_page_t m) 2815 { 2816 uint32_t ocount; 2817 2818 ocount = atomic_fetchadd_int(&m->busy_count, 1); 2819 KKASSERT(ocount & PBUSY_LOCKED); 2820 } 2821 2822 void 2823 vm_page_io_finish(vm_page_t m) 2824 { 2825 uint32_t ocount; 2826 2827 ocount = atomic_fetchadd_int(&m->busy_count, -1); 2828 KKASSERT(ocount & PBUSY_MASK); 2829 #if 0 2830 if (((ocount - 1) & (PBUSY_LOCKED | PBUSY_MASK)) == 0) 2831 wakeup(m); 2832 #endif 2833 } 2834 2835 /* 2836 * Attempt to soft-busy a page. The page must not be PBUSY_LOCKED. 2837 * 2838 * We can't use fetchadd here because we might race a hard-busy and the 2839 * page freeing code asserts on a non-zero soft-busy count (even if only 2840 * temporary). 2841 * 2842 * Returns 0 on success, non-zero on failure. 2843 */ 2844 int 2845 vm_page_sbusy_try(vm_page_t m) 2846 { 2847 uint32_t ocount; 2848 2849 for (;;) { 2850 ocount = m->busy_count; 2851 cpu_ccfence(); 2852 if (ocount & PBUSY_LOCKED) 2853 return 1; 2854 if (atomic_cmpset_int(&m->busy_count, ocount, ocount + 1)) 2855 break; 2856 } 2857 return 0; 2858 #if 0 2859 if (m->busy_count & PBUSY_LOCKED) 2860 return 1; 2861 ocount = atomic_fetchadd_int(&m->busy_count, 1); 2862 if (ocount & PBUSY_LOCKED) { 2863 vm_page_sbusy_drop(m); 2864 return 1; 2865 } 2866 return 0; 2867 #endif 2868 } 2869 2870 /* 2871 * Indicate that a clean VM page requires a filesystem commit and cannot 2872 * be reused. Used by tmpfs. 2873 */ 2874 void 2875 vm_page_need_commit(vm_page_t m) 2876 { 2877 vm_page_flag_set(m, PG_NEED_COMMIT); 2878 vm_object_set_writeable_dirty(m->object); 2879 } 2880 2881 void 2882 vm_page_clear_commit(vm_page_t m) 2883 { 2884 vm_page_flag_clear(m, PG_NEED_COMMIT); 2885 } 2886 2887 /* 2888 * Grab a page, blocking if it is busy and allocating a page if necessary. 2889 * A busy page is returned or NULL. The page may or may not be valid and 2890 * might not be on a queue (the caller is responsible for the disposition of 2891 * the page). 2892 * 2893 * If VM_ALLOC_ZERO is specified and the grab must allocate a new page, the 2894 * page will be zero'd and marked valid. 2895 * 2896 * If VM_ALLOC_FORCE_ZERO is specified the page will be zero'd and marked 2897 * valid even if it already exists. 2898 * 2899 * If VM_ALLOC_RETRY is specified this routine will never return NULL. Also 2900 * note that VM_ALLOC_NORMAL must be specified if VM_ALLOC_RETRY is specified. 2901 * VM_ALLOC_NULL_OK is implied when VM_ALLOC_RETRY is specified. 2902 * 2903 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is 2904 * always returned if we had blocked. 2905 * 2906 * This routine may not be called from an interrupt. 2907 * 2908 * No other requirements. 2909 */ 2910 vm_page_t 2911 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 2912 { 2913 vm_page_t m; 2914 int error; 2915 int shared = 1; 2916 2917 KKASSERT(allocflags & 2918 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 2919 vm_object_hold_shared(object); 2920 for (;;) { 2921 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error); 2922 if (error) { 2923 vm_page_sleep_busy(m, TRUE, "pgrbwt"); 2924 if ((allocflags & VM_ALLOC_RETRY) == 0) { 2925 m = NULL; 2926 break; 2927 } 2928 /* retry */ 2929 } else if (m == NULL) { 2930 if (shared) { 2931 vm_object_upgrade(object); 2932 shared = 0; 2933 } 2934 if (allocflags & VM_ALLOC_RETRY) 2935 allocflags |= VM_ALLOC_NULL_OK; 2936 m = vm_page_alloc(object, pindex, 2937 allocflags & ~VM_ALLOC_RETRY); 2938 if (m) 2939 break; 2940 vm_wait(0); 2941 if ((allocflags & VM_ALLOC_RETRY) == 0) 2942 goto failed; 2943 } else { 2944 /* m found */ 2945 break; 2946 } 2947 } 2948 2949 /* 2950 * If VM_ALLOC_ZERO an invalid page will be zero'd and set valid. 2951 * 2952 * If VM_ALLOC_FORCE_ZERO the page is unconditionally zero'd and set 2953 * valid even if already valid. 2954 * 2955 * NOTE! We have removed all of the PG_ZERO optimizations and also 2956 * removed the idle zeroing code. These optimizations actually 2957 * slow things down on modern cpus because the zerod area is 2958 * likely uncached, placing a memory-access burden on the 2959 * accesors taking the fault. 2960 * 2961 * By always zeroing the page in-line with the fault, no 2962 * dynamic ram reads are needed and the caches are hot, ready 2963 * for userland to access the memory. 2964 */ 2965 if (m->valid == 0) { 2966 if (allocflags & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) { 2967 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 2968 m->valid = VM_PAGE_BITS_ALL; 2969 } 2970 } else if (allocflags & VM_ALLOC_FORCE_ZERO) { 2971 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 2972 m->valid = VM_PAGE_BITS_ALL; 2973 } 2974 failed: 2975 vm_object_drop(object); 2976 return(m); 2977 } 2978 2979 /* 2980 * Mapping function for valid bits or for dirty bits in 2981 * a page. May not block. 2982 * 2983 * Inputs are required to range within a page. 2984 * 2985 * No requirements. 2986 * Non blocking. 2987 */ 2988 int 2989 vm_page_bits(int base, int size) 2990 { 2991 int first_bit; 2992 int last_bit; 2993 2994 KASSERT( 2995 base + size <= PAGE_SIZE, 2996 ("vm_page_bits: illegal base/size %d/%d", base, size) 2997 ); 2998 2999 if (size == 0) /* handle degenerate case */ 3000 return(0); 3001 3002 first_bit = base >> DEV_BSHIFT; 3003 last_bit = (base + size - 1) >> DEV_BSHIFT; 3004 3005 return ((2 << last_bit) - (1 << first_bit)); 3006 } 3007 3008 /* 3009 * Sets portions of a page valid and clean. The arguments are expected 3010 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 3011 * of any partial chunks touched by the range. The invalid portion of 3012 * such chunks will be zero'd. 3013 * 3014 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically 3015 * align base to DEV_BSIZE so as not to mark clean a partially 3016 * truncated device block. Otherwise the dirty page status might be 3017 * lost. 3018 * 3019 * This routine may not block. 3020 * 3021 * (base + size) must be less then or equal to PAGE_SIZE. 3022 */ 3023 static void 3024 _vm_page_zero_valid(vm_page_t m, int base, int size) 3025 { 3026 int frag; 3027 int endoff; 3028 3029 if (size == 0) /* handle degenerate case */ 3030 return; 3031 3032 /* 3033 * If the base is not DEV_BSIZE aligned and the valid 3034 * bit is clear, we have to zero out a portion of the 3035 * first block. 3036 */ 3037 3038 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 3039 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0 3040 ) { 3041 pmap_zero_page_area( 3042 VM_PAGE_TO_PHYS(m), 3043 frag, 3044 base - frag 3045 ); 3046 } 3047 3048 /* 3049 * If the ending offset is not DEV_BSIZE aligned and the 3050 * valid bit is clear, we have to zero out a portion of 3051 * the last block. 3052 */ 3053 3054 endoff = base + size; 3055 3056 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 3057 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0 3058 ) { 3059 pmap_zero_page_area( 3060 VM_PAGE_TO_PHYS(m), 3061 endoff, 3062 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)) 3063 ); 3064 } 3065 } 3066 3067 /* 3068 * Set valid, clear dirty bits. If validating the entire 3069 * page we can safely clear the pmap modify bit. We also 3070 * use this opportunity to clear the PG_NOSYNC flag. If a process 3071 * takes a write fault on a MAP_NOSYNC memory area the flag will 3072 * be set again. 3073 * 3074 * We set valid bits inclusive of any overlap, but we can only 3075 * clear dirty bits for DEV_BSIZE chunks that are fully within 3076 * the range. 3077 * 3078 * Page must be busied? 3079 * No other requirements. 3080 */ 3081 void 3082 vm_page_set_valid(vm_page_t m, int base, int size) 3083 { 3084 _vm_page_zero_valid(m, base, size); 3085 m->valid |= vm_page_bits(base, size); 3086 } 3087 3088 3089 /* 3090 * Set valid bits and clear dirty bits. 3091 * 3092 * Page must be busied by caller. 3093 * 3094 * NOTE: This function does not clear the pmap modified bit. 3095 * Also note that e.g. NFS may use a byte-granular base 3096 * and size. 3097 * 3098 * No other requirements. 3099 */ 3100 void 3101 vm_page_set_validclean(vm_page_t m, int base, int size) 3102 { 3103 int pagebits; 3104 3105 _vm_page_zero_valid(m, base, size); 3106 pagebits = vm_page_bits(base, size); 3107 m->valid |= pagebits; 3108 m->dirty &= ~pagebits; 3109 if (base == 0 && size == PAGE_SIZE) { 3110 /*pmap_clear_modify(m);*/ 3111 vm_page_flag_clear(m, PG_NOSYNC); 3112 } 3113 } 3114 3115 /* 3116 * Set valid & dirty. Used by buwrite() 3117 * 3118 * Page must be busied by caller. 3119 */ 3120 void 3121 vm_page_set_validdirty(vm_page_t m, int base, int size) 3122 { 3123 int pagebits; 3124 3125 pagebits = vm_page_bits(base, size); 3126 m->valid |= pagebits; 3127 m->dirty |= pagebits; 3128 if (m->object) 3129 vm_object_set_writeable_dirty(m->object); 3130 } 3131 3132 /* 3133 * Clear dirty bits. 3134 * 3135 * NOTE: This function does not clear the pmap modified bit. 3136 * Also note that e.g. NFS may use a byte-granular base 3137 * and size. 3138 * 3139 * Page must be busied? 3140 * No other requirements. 3141 */ 3142 void 3143 vm_page_clear_dirty(vm_page_t m, int base, int size) 3144 { 3145 m->dirty &= ~vm_page_bits(base, size); 3146 if (base == 0 && size == PAGE_SIZE) { 3147 /*pmap_clear_modify(m);*/ 3148 vm_page_flag_clear(m, PG_NOSYNC); 3149 } 3150 } 3151 3152 /* 3153 * Make the page all-dirty. 3154 * 3155 * Also make sure the related object and vnode reflect the fact that the 3156 * object may now contain a dirty page. 3157 * 3158 * Page must be busied? 3159 * No other requirements. 3160 */ 3161 void 3162 vm_page_dirty(vm_page_t m) 3163 { 3164 #ifdef INVARIANTS 3165 int pqtype = m->queue - m->pc; 3166 #endif 3167 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE, 3168 ("vm_page_dirty: page in free/cache queue!")); 3169 if (m->dirty != VM_PAGE_BITS_ALL) { 3170 m->dirty = VM_PAGE_BITS_ALL; 3171 if (m->object) 3172 vm_object_set_writeable_dirty(m->object); 3173 } 3174 } 3175 3176 /* 3177 * Invalidates DEV_BSIZE'd chunks within a page. Both the 3178 * valid and dirty bits for the effected areas are cleared. 3179 * 3180 * Page must be busied? 3181 * Does not block. 3182 * No other requirements. 3183 */ 3184 void 3185 vm_page_set_invalid(vm_page_t m, int base, int size) 3186 { 3187 int bits; 3188 3189 bits = vm_page_bits(base, size); 3190 m->valid &= ~bits; 3191 m->dirty &= ~bits; 3192 atomic_add_int(&m->object->generation, 1); 3193 } 3194 3195 /* 3196 * The kernel assumes that the invalid portions of a page contain 3197 * garbage, but such pages can be mapped into memory by user code. 3198 * When this occurs, we must zero out the non-valid portions of the 3199 * page so user code sees what it expects. 3200 * 3201 * Pages are most often semi-valid when the end of a file is mapped 3202 * into memory and the file's size is not page aligned. 3203 * 3204 * Page must be busied? 3205 * No other requirements. 3206 */ 3207 void 3208 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 3209 { 3210 int b; 3211 int i; 3212 3213 /* 3214 * Scan the valid bits looking for invalid sections that 3215 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 3216 * valid bit may be set ) have already been zerod by 3217 * vm_page_set_validclean(). 3218 */ 3219 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 3220 if (i == (PAGE_SIZE / DEV_BSIZE) || 3221 (m->valid & (1 << i)) 3222 ) { 3223 if (i > b) { 3224 pmap_zero_page_area( 3225 VM_PAGE_TO_PHYS(m), 3226 b << DEV_BSHIFT, 3227 (i - b) << DEV_BSHIFT 3228 ); 3229 } 3230 b = i + 1; 3231 } 3232 } 3233 3234 /* 3235 * setvalid is TRUE when we can safely set the zero'd areas 3236 * as being valid. We can do this if there are no cache consistency 3237 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 3238 */ 3239 if (setvalid) 3240 m->valid = VM_PAGE_BITS_ALL; 3241 } 3242 3243 /* 3244 * Is a (partial) page valid? Note that the case where size == 0 3245 * will return FALSE in the degenerate case where the page is entirely 3246 * invalid, and TRUE otherwise. 3247 * 3248 * Does not block. 3249 * No other requirements. 3250 */ 3251 int 3252 vm_page_is_valid(vm_page_t m, int base, int size) 3253 { 3254 int bits = vm_page_bits(base, size); 3255 3256 if (m->valid && ((m->valid & bits) == bits)) 3257 return 1; 3258 else 3259 return 0; 3260 } 3261 3262 /* 3263 * update dirty bits from pmap/mmu. May not block. 3264 * 3265 * Caller must hold the page busy 3266 */ 3267 void 3268 vm_page_test_dirty(vm_page_t m) 3269 { 3270 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 3271 vm_page_dirty(m); 3272 } 3273 } 3274 3275 #include "opt_ddb.h" 3276 #ifdef DDB 3277 #include <ddb/ddb.h> 3278 3279 DB_SHOW_COMMAND(page, vm_page_print_page_info) 3280 { 3281 db_printf("vmstats.v_free_count: %ld\n", vmstats.v_free_count); 3282 db_printf("vmstats.v_cache_count: %ld\n", vmstats.v_cache_count); 3283 db_printf("vmstats.v_inactive_count: %ld\n", vmstats.v_inactive_count); 3284 db_printf("vmstats.v_active_count: %ld\n", vmstats.v_active_count); 3285 db_printf("vmstats.v_wire_count: %ld\n", vmstats.v_wire_count); 3286 db_printf("vmstats.v_free_reserved: %ld\n", vmstats.v_free_reserved); 3287 db_printf("vmstats.v_free_min: %ld\n", vmstats.v_free_min); 3288 db_printf("vmstats.v_free_target: %ld\n", vmstats.v_free_target); 3289 db_printf("vmstats.v_cache_min: %ld\n", vmstats.v_cache_min); 3290 db_printf("vmstats.v_inactive_target: %ld\n", 3291 vmstats.v_inactive_target); 3292 } 3293 3294 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 3295 { 3296 int i; 3297 db_printf("PQ_FREE:"); 3298 for (i = 0; i < PQ_L2_SIZE; i++) { 3299 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 3300 } 3301 db_printf("\n"); 3302 3303 db_printf("PQ_CACHE:"); 3304 for(i = 0; i < PQ_L2_SIZE; i++) { 3305 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 3306 } 3307 db_printf("\n"); 3308 3309 db_printf("PQ_ACTIVE:"); 3310 for(i = 0; i < PQ_L2_SIZE; i++) { 3311 db_printf(" %d", vm_page_queues[PQ_ACTIVE + i].lcnt); 3312 } 3313 db_printf("\n"); 3314 3315 db_printf("PQ_INACTIVE:"); 3316 for(i = 0; i < PQ_L2_SIZE; i++) { 3317 db_printf(" %d", vm_page_queues[PQ_INACTIVE + i].lcnt); 3318 } 3319 db_printf("\n"); 3320 } 3321 #endif /* DDB */ 3322