Now available: The Design and Implementation of the FreeBSD Operating System (Second Edition)
[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]
FreeBSD/Linux Kernel Cross Reference
sys/uvm/uvm_km.c
Version:
- FREEBSD - FREEBSD-13-STABLE - FREEBSD-13-0 - FREEBSD-12-STABLE - FREEBSD-12-0 - FREEBSD-11-STABLE - FREEBSD-11-0 - FREEBSD-10-STABLE - FREEBSD-10-0 - FREEBSD-9-STABLE - FREEBSD-9-0 - FREEBSD-8-STABLE - FREEBSD-8-0 - FREEBSD-7-STABLE - FREEBSD-7-0 - FREEBSD-6-STABLE - FREEBSD-6-0 - FREEBSD-5-STABLE - FREEBSD-5-0 - FREEBSD-4-STABLE - FREEBSD-3-STABLE - FREEBSD22 - l41 - OPENBSD - linux-2.6 - MK84 - PLAN9 - xnu-8792
SearchContext: - none - 3 - 10
SearchContext: - none - 3 - 10
1 /* $OpenBSD: uvm_km.c,v 1.151 2022/08/01 14:15:46 mpi Exp $ */ 2 /* $NetBSD: uvm_km.c,v 1.42 2001/01/14 02:10:01 thorpej Exp $ */ 3 4 /* 5 * Copyright (c) 1997 Charles D. Cranor and Washington University. 6 * Copyright (c) 1991, 1993, The Regents of the University of California. 7 * 8 * All rights reserved. 9 * 10 * This code is derived from software contributed to Berkeley by 11 * The Mach Operating System project at Carnegie-Mellon University. 12 * 13 * Redistribution and use in source and binary forms, with or without 14 * modification, are permitted provided that the following conditions 15 * are met: 16 * 1. Redistributions of source code must retain the above copyright 17 * notice, this list of conditions and the following disclaimer. 18 * 2. Redistributions in binary form must reproduce the above copyright 19 * notice, this list of conditions and the following disclaimer in the 20 * documentation and/or other materials provided with the distribution. 21 * 3. Neither the name of the University nor the names of its contributors 22 * may be used to endorse or promote products derived from this software 23 * without specific prior written permission. 24 * 25 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 26 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 27 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 28 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 29 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 30 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 31 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 32 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 33 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 34 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 35 * SUCH DAMAGE. 36 * 37 * @(#)vm_kern.c 8.3 (Berkeley) 1/12/94 38 * from: Id: uvm_km.c,v 1.1.2.14 1998/02/06 05:19:27 chs Exp 39 * 40 * 41 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 42 * All rights reserved. 43 * 44 * Permission to use, copy, modify and distribute this software and 45 * its documentation is hereby granted, provided that both the copyright 46 * notice and this permission notice appear in all copies of the 47 * software, derivative works or modified versions, and any portions 48 * thereof, and that both notices appear in supporting documentation. 49 * 50 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 51 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 52 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 53 * 54 * Carnegie Mellon requests users of this software to return to 55 * 56 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 57 * School of Computer Science 58 * Carnegie Mellon University 59 * Pittsburgh PA 15213-3890 60 * 61 * any improvements or extensions that they make and grant Carnegie the 62 * rights to redistribute these changes. 63 */ 64 65 /* 66 * uvm_km.c: handle kernel memory allocation and management 67 */ 68 69 /* 70 * overview of kernel memory management: 71 * 72 * the kernel virtual address space is mapped by "kernel_map." kernel_map 73 * starts at a machine-dependent address and is VM_KERNEL_SPACE_SIZE bytes 74 * large. 75 * 76 * the kernel_map has several "submaps." submaps can only appear in 77 * the kernel_map (user processes can't use them). submaps "take over" 78 * the management of a sub-range of the kernel's address space. submaps 79 * are typically allocated at boot time and are never released. kernel 80 * virtual address space that is mapped by a submap is locked by the 81 * submap's lock -- not the kernel_map's lock. 82 * 83 * thus, the useful feature of submaps is that they allow us to break 84 * up the locking and protection of the kernel address space into smaller 85 * chunks. 86 * 87 * The VM system has several standard kernel submaps: 88 * kmem_map: Contains only wired kernel memory for malloc(9). 89 * Note: All access to this map must be protected by splvm as 90 * calls to malloc(9) are allowed in interrupt handlers. 91 * exec_map: Memory to hold arguments to system calls are allocated from 92 * this map. 93 * XXX: This is primeraly used to artificially limit the number 94 * of concurrent processes doing an exec. 95 * phys_map: Buffers for vmapbuf (physio) are allocated from this map. 96 * 97 * the kernel allocates its private memory out of special uvm_objects whose 98 * reference count is set to UVM_OBJ_KERN (thus indicating that the objects 99 * are "special" and never die). all kernel objects should be thought of 100 * as large, fixed-sized, sparsely populated uvm_objects. each kernel 101 * object is equal to the size of kernel virtual address space (i.e. 102 * VM_KERNEL_SPACE_SIZE). 103 * 104 * most kernel private memory lives in kernel_object. the only exception 105 * to this is for memory that belongs to submaps that must be protected 106 * by splvm(). each of these submaps manages their own pages. 107 * 108 * note that just because a kernel object spans the entire kernel virtual 109 * address space doesn't mean that it has to be mapped into the entire space. 110 * large chunks of a kernel object's space go unused either because 111 * that area of kernel VM is unmapped, or there is some other type of 112 * object mapped into that range (e.g. a vnode). for submap's kernel 113 * objects, the only part of the object that can ever be populated is the 114 * offsets that are managed by the submap. 115 * 116 * note that the "offset" in a kernel object is always the kernel virtual 117 * address minus the vm_map_min(kernel_map). 118 * example: 119 * suppose kernel_map starts at 0xf8000000 and the kernel does a 120 * uvm_km_alloc(kernel_map, PAGE_SIZE) [allocate 1 wired down page in the 121 * kernel map]. if uvm_km_alloc returns virtual address 0xf8235000, 122 * then that means that the page at offset 0x235000 in kernel_object is 123 * mapped at 0xf8235000. 124 * 125 * kernel objects have one other special property: when the kernel virtual 126 * memory mapping them is unmapped, the backing memory in the object is 127 * freed right away. this is done with the uvm_km_pgremove() function. 128 * this has to be done because there is no backing store for kernel pages 129 * and no need to save them after they are no longer referenced. 130 */ 131 132 #include <sys/param.h> 133 #include <sys/systm.h> 134 #include <sys/proc.h> 135 #include <sys/kthread.h> 136 #include <uvm/uvm.h> 137 138 /* 139 * global data structures 140 */ 141 142 struct vm_map *kernel_map = NULL; 143 144 /* Unconstraint range. */ 145 struct uvm_constraint_range no_constraint = { 0x0, (paddr_t)-1 }; 146 147 /* 148 * local data structures 149 */ 150 static struct vm_map kernel_map_store; 151 152 /* 153 * uvm_km_init: init kernel maps and objects to reflect reality (i.e. 154 * KVM already allocated for text, data, bss, and static data structures). 155 * 156 * => KVM is defined by [base.. base + VM_KERNEL_SPACE_SIZE]. 157 * we assume that [base -> start] has already been allocated and that 158 * "end" is the end of the kernel image span. 159 */ 160 void 161 uvm_km_init(vaddr_t base, vaddr_t start, vaddr_t end) 162 { 163 /* kernel_object: for pageable anonymous kernel memory */ 164 uao_init(); 165 uvm.kernel_object = uao_create(VM_KERNEL_SPACE_SIZE, UAO_FLAG_KERNOBJ); 166 167 /* 168 * init the map and reserve already allocated kernel space 169 * before installing. 170 */ 171 172 uvm_map_setup(&kernel_map_store, pmap_kernel(), base, end, 173 #ifdef KVA_GUARDPAGES 174 VM_MAP_PAGEABLE | VM_MAP_GUARDPAGES 175 #else 176 VM_MAP_PAGEABLE 177 #endif 178 ); 179 if (base != start && uvm_map(&kernel_map_store, &base, start - base, 180 NULL, UVM_UNKNOWN_OFFSET, 0, 181 UVM_MAPFLAG(PROT_READ | PROT_WRITE, PROT_READ | PROT_WRITE, 182 MAP_INHERIT_NONE, MADV_RANDOM, UVM_FLAG_FIXED)) != 0) 183 panic("uvm_km_init: could not reserve space for kernel"); 184 185 kernel_map = &kernel_map_store; 186 } 187 188 /* 189 * uvm_km_suballoc: allocate a submap in the kernel map. once a submap 190 * is allocated all references to that area of VM must go through it. this 191 * allows the locking of VAs in kernel_map to be broken up into regions. 192 * 193 * => if `fixed' is true, *min specifies where the region described 194 * by the submap must start 195 * => if submap is non NULL we use that as the submap, otherwise we 196 * alloc a new map 197 */ 198 struct vm_map * 199 uvm_km_suballoc(struct vm_map *map, vaddr_t *min, vaddr_t *max, vsize_t size, 200 int flags, boolean_t fixed, struct vm_map *submap) 201 { 202 int mapflags = UVM_FLAG_NOMERGE | (fixed ? UVM_FLAG_FIXED : 0); 203 204 size = round_page(size); /* round up to pagesize */ 205 206 /* first allocate a blank spot in the parent map */ 207 if (uvm_map(map, min, size, NULL, UVM_UNKNOWN_OFFSET, 0, 208 UVM_MAPFLAG(PROT_READ | PROT_WRITE, PROT_READ | PROT_WRITE, 209 MAP_INHERIT_NONE, MADV_RANDOM, mapflags)) != 0) { 210 panic("uvm_km_suballoc: unable to allocate space in parent map"); 211 } 212 213 /* set VM bounds (min is filled in by uvm_map) */ 214 *max = *min + size; 215 216 /* add references to pmap and create or init the submap */ 217 pmap_reference(vm_map_pmap(map)); 218 if (submap == NULL) { 219 submap = uvm_map_create(vm_map_pmap(map), *min, *max, flags); 220 if (submap == NULL) 221 panic("uvm_km_suballoc: unable to create submap"); 222 } else { 223 uvm_map_setup(submap, vm_map_pmap(map), *min, *max, flags); 224 } 225 226 /* 227 * now let uvm_map_submap plug in it... 228 */ 229 if (uvm_map_submap(map, *min, *max, submap) != 0) 230 panic("uvm_km_suballoc: submap allocation failed"); 231 232 return(submap); 233 } 234 235 /* 236 * uvm_km_pgremove: remove pages from a kernel uvm_object. 237 * 238 * => when you unmap a part of anonymous kernel memory you want to toss 239 * the pages right away. (this gets called from uvm_unmap_...). 240 */ 241 void 242 uvm_km_pgremove(struct uvm_object *uobj, vaddr_t startva, vaddr_t endva) 243 { 244 const voff_t start = startva - vm_map_min(kernel_map); 245 const voff_t end = endva - vm_map_min(kernel_map); 246 struct vm_page *pp; 247 voff_t curoff; 248 int slot; 249 int swpgonlydelta = 0; 250 251 KASSERT(UVM_OBJ_IS_AOBJ(uobj)); 252 KASSERT(rw_write_held(uobj->vmobjlock)); 253 254 pmap_remove(pmap_kernel(), startva, endva); 255 for (curoff = start ; curoff < end ; curoff += PAGE_SIZE) { 256 pp = uvm_pagelookup(uobj, curoff); 257 if (pp && pp->pg_flags & PG_BUSY) { 258 uvm_pagewait(pp, uobj->vmobjlock, "km_pgrm"); 259 rw_enter(uobj->vmobjlock, RW_WRITE); 260 curoff -= PAGE_SIZE; /* loop back to us */ 261 continue; 262 } 263 264 /* free the swap slot, then the page */ 265 slot = uao_dropswap(uobj, curoff >> PAGE_SHIFT); 266 267 if (pp != NULL) { 268 uvm_lock_pageq(); 269 uvm_pagefree(pp); 270 uvm_unlock_pageq(); 271 } else if (slot != 0) { 272 swpgonlydelta++; 273 } 274 } 275 276 if (swpgonlydelta > 0) { 277 KASSERT(uvmexp.swpgonly >= swpgonlydelta); 278 atomic_add_int(&uvmexp.swpgonly, -swpgonlydelta); 279 } 280 } 281 282 283 /* 284 * uvm_km_pgremove_intrsafe: like uvm_km_pgremove(), but for "intrsafe" 285 * objects 286 * 287 * => when you unmap a part of anonymous kernel memory you want to toss 288 * the pages right away. (this gets called from uvm_unmap_...). 289 * => none of the pages will ever be busy, and none of them will ever 290 * be on the active or inactive queues (because these objects are 291 * never allowed to "page"). 292 */ 293 void 294 uvm_km_pgremove_intrsafe(vaddr_t start, vaddr_t end) 295 { 296 struct vm_page *pg; 297 vaddr_t va; 298 paddr_t pa; 299 300 for (va = start; va < end; va += PAGE_SIZE) { 301 if (!pmap_extract(pmap_kernel(), va, &pa)) 302 continue; 303 pg = PHYS_TO_VM_PAGE(pa); 304 if (pg == NULL) 305 panic("uvm_km_pgremove_intrsafe: no page"); 306 uvm_pagefree(pg); 307 } 308 pmap_kremove(start, end - start); 309 } 310 311 /* 312 * uvm_km_kmemalloc: lower level kernel memory allocator for malloc() 313 * 314 * => we map wired memory into the specified map using the obj passed in 315 * => NOTE: we can return NULL even if we can wait if there is not enough 316 * free VM space in the map... caller should be prepared to handle 317 * this case. 318 * => we return KVA of memory allocated 319 * => flags: NOWAIT, VALLOC - just allocate VA, TRYLOCK - fail if we can't 320 * lock the map 321 * => low, high, alignment, boundary, nsegs are the corresponding parameters 322 * to uvm_pglistalloc 323 * => flags: ZERO - correspond to uvm_pglistalloc flags 324 */ 325 vaddr_t 326 uvm_km_kmemalloc_pla(struct vm_map *map, struct uvm_object *obj, vsize_t size, 327 vsize_t valign, int flags, paddr_t low, paddr_t high, paddr_t alignment, 328 paddr_t boundary, int nsegs) 329 { 330 vaddr_t kva, loopva; 331 voff_t offset; 332 struct vm_page *pg; 333 struct pglist pgl; 334 int pla_flags; 335 336 KASSERT(vm_map_pmap(map) == pmap_kernel()); 337 /* UVM_KMF_VALLOC => !UVM_KMF_ZERO */ 338 KASSERT(!(flags & UVM_KMF_VALLOC) || 339 !(flags & UVM_KMF_ZERO)); 340 341 /* setup for call */ 342 size = round_page(size); 343 kva = vm_map_min(map); /* hint */ 344 if (nsegs == 0) 345 nsegs = atop(size); 346 347 /* allocate some virtual space */ 348 if (__predict_false(uvm_map(map, &kva, size, obj, UVM_UNKNOWN_OFFSET, 349 valign, UVM_MAPFLAG(PROT_READ | PROT_WRITE, PROT_READ | PROT_WRITE, 350 MAP_INHERIT_NONE, MADV_RANDOM, (flags & UVM_KMF_TRYLOCK))) != 0)) { 351 return 0; 352 } 353 354 /* if all we wanted was VA, return now */ 355 if (flags & UVM_KMF_VALLOC) { 356 return kva; 357 } 358 359 /* recover object offset from virtual address */ 360 if (obj != NULL) 361 offset = kva - vm_map_min(kernel_map); 362 else 363 offset = 0; 364 365 /* 366 * now allocate and map in the memory... note that we are the only ones 367 * whom should ever get a handle on this area of VM. 368 */ 369 TAILQ_INIT(&pgl); 370 pla_flags = 0; 371 KASSERT(uvmexp.swpgonly <= uvmexp.swpages); 372 if ((flags & UVM_KMF_NOWAIT) || 373 ((flags & UVM_KMF_CANFAIL) && 374 uvmexp.swpages - uvmexp.swpgonly <= atop(size))) 375 pla_flags |= UVM_PLA_NOWAIT; 376 else 377 pla_flags |= UVM_PLA_WAITOK; 378 if (flags & UVM_KMF_ZERO) 379 pla_flags |= UVM_PLA_ZERO; 380 if (uvm_pglistalloc(size, low, high, alignment, boundary, &pgl, nsegs, 381 pla_flags) != 0) { 382 /* Failed. */ 383 uvm_unmap(map, kva, kva + size); 384 return (0); 385 } 386 387 if (obj != NULL) 388 rw_enter(obj->vmobjlock, RW_WRITE); 389 390 loopva = kva; 391 while (loopva != kva + size) { 392 pg = TAILQ_FIRST(&pgl); 393 TAILQ_REMOVE(&pgl, pg, pageq); 394 uvm_pagealloc_pg(pg, obj, offset, NULL); 395 atomic_clearbits_int(&pg->pg_flags, PG_BUSY); 396 UVM_PAGE_OWN(pg, NULL); 397 398 /* 399 * map it in: note that we call pmap_enter with the map and 400 * object unlocked in case we are kmem_map. 401 */ 402 if (obj == NULL) { 403 pmap_kenter_pa(loopva, VM_PAGE_TO_PHYS(pg), 404 PROT_READ | PROT_WRITE); 405 } else { 406 pmap_enter(map->pmap, loopva, VM_PAGE_TO_PHYS(pg), 407 PROT_READ | PROT_WRITE, 408 PROT_READ | PROT_WRITE | PMAP_WIRED); 409 } 410 loopva += PAGE_SIZE; 411 offset += PAGE_SIZE; 412 } 413 KASSERT(TAILQ_EMPTY(&pgl)); 414 pmap_update(pmap_kernel()); 415 416 if (obj != NULL) 417 rw_exit(obj->vmobjlock); 418 419 return kva; 420 } 421 422 /* 423 * uvm_km_free: free an area of kernel memory 424 */ 425 void 426 uvm_km_free(struct vm_map *map, vaddr_t addr, vsize_t size) 427 { 428 uvm_unmap(map, trunc_page(addr), round_page(addr+size)); 429 } 430 431 /* 432 * uvm_km_alloc1: allocate wired down memory in the kernel map. 433 * 434 * => we can sleep if needed 435 */ 436 vaddr_t 437 uvm_km_alloc1(struct vm_map *map, vsize_t size, vsize_t align, boolean_t zeroit) 438 { 439 vaddr_t kva, loopva; 440 voff_t offset; 441 struct vm_page *pg; 442 443 KASSERT(vm_map_pmap(map) == pmap_kernel()); 444 445 size = round_page(size); 446 kva = vm_map_min(map); /* hint */ 447 448 /* allocate some virtual space */ 449 if (__predict_false(uvm_map(map, &kva, size, uvm.kernel_object, 450 UVM_UNKNOWN_OFFSET, align, 451 UVM_MAPFLAG(PROT_READ | PROT_WRITE, 452 PROT_READ | PROT_WRITE | PROT_EXEC, 453 MAP_INHERIT_NONE, MADV_RANDOM, 0)) != 0)) { 454 return 0; 455 } 456 457 /* recover object offset from virtual address */ 458 offset = kva - vm_map_min(kernel_map); 459 460 /* now allocate the memory. we must be careful about released pages. */ 461 loopva = kva; 462 while (size) { 463 rw_enter(uvm.kernel_object->vmobjlock, RW_WRITE); 464 /* allocate ram */ 465 pg = uvm_pagealloc(uvm.kernel_object, offset, NULL, 0); 466 if (pg) { 467 atomic_clearbits_int(&pg->pg_flags, PG_BUSY); 468 UVM_PAGE_OWN(pg, NULL); 469 } 470 rw_exit(uvm.kernel_object->vmobjlock); 471 if (__predict_false(pg == NULL)) { 472 if (curproc == uvm.pagedaemon_proc) { 473 /* 474 * It is unfeasible for the page daemon to 475 * sleep for memory, so free what we have 476 * allocated and fail. 477 */ 478 uvm_unmap(map, kva, loopva - kva); 479 return (0); 480 } else { 481 uvm_wait("km_alloc1w"); /* wait for memory */ 482 continue; 483 } 484 } 485 486 /* 487 * map it in; note we're never called with an intrsafe 488 * object, so we always use regular old pmap_enter(). 489 */ 490 pmap_enter(map->pmap, loopva, VM_PAGE_TO_PHYS(pg), 491 PROT_READ | PROT_WRITE, 492 PROT_READ | PROT_WRITE | PMAP_WIRED); 493 494 loopva += PAGE_SIZE; 495 offset += PAGE_SIZE; 496 size -= PAGE_SIZE; 497 } 498 pmap_update(map->pmap); 499 500 /* 501 * zero on request (note that "size" is now zero due to the above loop 502 * so we need to subtract kva from loopva to reconstruct the size). 503 */ 504 if (zeroit) 505 memset((caddr_t)kva, 0, loopva - kva); 506 507 return kva; 508 } 509 510 #if defined(__HAVE_PMAP_DIRECT) 511 /* 512 * uvm_km_page allocator, __HAVE_PMAP_DIRECT arch 513 * On architectures with machine memory direct mapped into a portion 514 * of KVM, we have very little work to do. Just get a physical page, 515 * and find and return its VA. 516 */ 517 void 518 uvm_km_page_init(void) 519 { 520 /* nothing */ 521 } 522 523 void 524 uvm_km_page_lateinit(void) 525 { 526 /* nothing */ 527 } 528 529 #else 530 /* 531 * uvm_km_page allocator, non __HAVE_PMAP_DIRECT archs 532 * This is a special allocator that uses a reserve of free pages 533 * to fulfill requests. It is fast and interrupt safe, but can only 534 * return page sized regions. Its primary use is as a backend for pool. 535 * 536 * The memory returned is allocated from the larger kernel_map, sparing 537 * pressure on the small interrupt-safe kmem_map. It is wired, but 538 * not zero filled. 539 */ 540 541 struct uvm_km_pages uvm_km_pages; 542 543 void uvm_km_createthread(void *); 544 void uvm_km_thread(void *); 545 struct uvm_km_free_page *uvm_km_doputpage(struct uvm_km_free_page *); 546 547 /* 548 * Allocate the initial reserve, and create the thread which will 549 * keep the reserve full. For bootstrapping, we allocate more than 550 * the lowat amount, because it may be a while before the thread is 551 * running. 552 */ 553 void 554 uvm_km_page_init(void) 555 { 556 int lowat_min; 557 int i; 558 int len, bulk; 559 vaddr_t addr; 560 561 mtx_init(&uvm_km_pages.mtx, IPL_VM); 562 if (!uvm_km_pages.lowat) { 563 /* based on physmem, calculate a good value here */ 564 uvm_km_pages.lowat = physmem / 256; 565 lowat_min = physmem < atop(16 * 1024 * 1024) ? 32 : 128; 566 if (uvm_km_pages.lowat < lowat_min) 567 uvm_km_pages.lowat = lowat_min; 568 } 569 if (uvm_km_pages.lowat > UVM_KM_PAGES_LOWAT_MAX) 570 uvm_km_pages.lowat = UVM_KM_PAGES_LOWAT_MAX; 571 uvm_km_pages.hiwat = 4 * uvm_km_pages.lowat; 572 if (uvm_km_pages.hiwat > UVM_KM_PAGES_HIWAT_MAX) 573 uvm_km_pages.hiwat = UVM_KM_PAGES_HIWAT_MAX; 574 575 /* Allocate all pages in as few allocations as possible. */ 576 len = 0; 577 bulk = uvm_km_pages.hiwat; 578 while (len < uvm_km_pages.hiwat && bulk > 0) { 579 bulk = MIN(bulk, uvm_km_pages.hiwat - len); 580 addr = vm_map_min(kernel_map); 581 if (uvm_map(kernel_map, &addr, (vsize_t)bulk << PAGE_SHIFT, 582 NULL, UVM_UNKNOWN_OFFSET, 0, 583 UVM_MAPFLAG(PROT_READ | PROT_WRITE, 584 PROT_READ | PROT_WRITE, MAP_INHERIT_NONE, 585 MADV_RANDOM, UVM_KMF_TRYLOCK)) != 0) { 586 bulk /= 2; 587 continue; 588 } 589 590 for (i = len; i < len + bulk; i++, addr += PAGE_SIZE) 591 uvm_km_pages.page[i] = addr; 592 len += bulk; 593 } 594 595 uvm_km_pages.free = len; 596 for (i = len; i < UVM_KM_PAGES_HIWAT_MAX; i++) 597 uvm_km_pages.page[i] = 0; 598 599 /* tone down if really high */ 600 if (uvm_km_pages.lowat > 512) 601 uvm_km_pages.lowat = 512; 602 } 603 604 void 605 uvm_km_page_lateinit(void) 606 { 607 kthread_create_deferred(uvm_km_createthread, NULL); 608 } 609 610 void 611 uvm_km_createthread(void *arg) 612 { 613 kthread_create(uvm_km_thread, NULL, &uvm_km_pages.km_proc, "kmthread"); 614 } 615 616 /* 617 * Endless loop. We grab pages in increments of 16 pages, then 618 * quickly swap them into the list. 619 */ 620 void 621 uvm_km_thread(void *arg) 622 { 623 vaddr_t pg[16]; 624 int i; 625 int allocmore = 0; 626 int flags; 627 struct uvm_km_free_page *fp = NULL; 628 629 KERNEL_UNLOCK(); 630 631 for (;;) { 632 mtx_enter(&uvm_km_pages.mtx); 633 if (uvm_km_pages.free >= uvm_km_pages.lowat && 634 uvm_km_pages.freelist == NULL) { 635 msleep_nsec(&uvm_km_pages.km_proc, &uvm_km_pages.mtx, 636 PVM, "kmalloc", INFSLP); 637 } 638 allocmore = uvm_km_pages.free < uvm_km_pages.lowat; 639 fp = uvm_km_pages.freelist; 640 uvm_km_pages.freelist = NULL; 641 uvm_km_pages.freelistlen = 0; 642 mtx_leave(&uvm_km_pages.mtx); 643 644 if (allocmore) { 645 /* 646 * If there was nothing on the freelist, then we 647 * must obtain at least one page to make progress. 648 * So, only use UVM_KMF_TRYLOCK for the first page 649 * if fp != NULL 650 */ 651 flags = UVM_MAPFLAG(PROT_READ | PROT_WRITE, 652 PROT_READ | PROT_WRITE, MAP_INHERIT_NONE, 653 MADV_RANDOM, fp != NULL ? UVM_KMF_TRYLOCK : 0); 654 memset(pg, 0, sizeof(pg)); 655 for (i = 0; i < nitems(pg); i++) { 656 pg[i] = vm_map_min(kernel_map); 657 if (uvm_map(kernel_map, &pg[i], PAGE_SIZE, 658 NULL, UVM_UNKNOWN_OFFSET, 0, flags) != 0) { 659 pg[i] = 0; 660 break; 661 } 662 663 /* made progress, so don't sleep for more */ 664 flags = UVM_MAPFLAG(PROT_READ | PROT_WRITE, 665 PROT_READ | PROT_WRITE, MAP_INHERIT_NONE, 666 MADV_RANDOM, UVM_KMF_TRYLOCK); 667 } 668 669 mtx_enter(&uvm_km_pages.mtx); 670 for (i = 0; i < nitems(pg); i++) { 671 if (uvm_km_pages.free == 672 nitems(uvm_km_pages.page)) 673 break; 674 else if (pg[i] != 0) 675 uvm_km_pages.page[uvm_km_pages.free++] 676 = pg[i]; 677 } 678 wakeup(&uvm_km_pages.free); 679 mtx_leave(&uvm_km_pages.mtx); 680 681 /* Cleanup left-over pages (if any). */ 682 for (; i < nitems(pg); i++) { 683 if (pg[i] != 0) { 684 uvm_unmap(kernel_map, 685 pg[i], pg[i] + PAGE_SIZE); 686 } 687 } 688 } 689 while (fp) { 690 fp = uvm_km_doputpage(fp); 691 } 692 } 693 } 694 695 struct uvm_km_free_page * 696 uvm_km_doputpage(struct uvm_km_free_page *fp) 697 { 698 vaddr_t va = (vaddr_t)fp; 699 struct vm_page *pg; 700 int freeva = 1; 701 struct uvm_km_free_page *nextfp = fp->next; 702 703 pg = uvm_atopg(va); 704 705 pmap_kremove(va, PAGE_SIZE); 706 pmap_update(kernel_map->pmap); 707 708 mtx_enter(&uvm_km_pages.mtx); 709 if (uvm_km_pages.free < uvm_km_pages.hiwat) { 710 uvm_km_pages.page[uvm_km_pages.free++] = va; 711 freeva = 0; 712 } 713 mtx_leave(&uvm_km_pages.mtx); 714 715 if (freeva) 716 uvm_unmap(kernel_map, va, va + PAGE_SIZE); 717 718 uvm_pagefree(pg); 719 return (nextfp); 720 } 721 #endif /* !__HAVE_PMAP_DIRECT */ 722 723 void * 724 km_alloc(size_t sz, const struct kmem_va_mode *kv, 725 const struct kmem_pa_mode *kp, const struct kmem_dyn_mode *kd) 726 { 727 struct vm_map *map; 728 struct vm_page *pg; 729 struct pglist pgl; 730 int mapflags = 0; 731 vm_prot_t prot; 732 paddr_t pla_align; 733 int pla_flags; 734 int pla_maxseg; 735 vaddr_t va, sva = 0; 736 737 KASSERT(sz == round_page(sz)); 738 739 TAILQ_INIT(&pgl); 740 741 if (kp->kp_nomem || kp->kp_pageable) 742 goto alloc_va; 743 744 pla_flags = kd->kd_waitok ? UVM_PLA_WAITOK : UVM_PLA_NOWAIT; 745 pla_flags |= UVM_PLA_TRYCONTIG; 746 if (kp->kp_zero) 747 pla_flags |= UVM_PLA_ZERO; 748 749 pla_align = kp->kp_align; 750 #ifdef __HAVE_PMAP_DIRECT 751 if (pla_align < kv->kv_align) 752 pla_align = kv->kv_align; 753 #endif 754 pla_maxseg = kp->kp_maxseg; 755 if (pla_maxseg == 0) 756 pla_maxseg = sz / PAGE_SIZE; 757 758 if (uvm_pglistalloc(sz, kp->kp_constraint->ucr_low, 759 kp->kp_constraint->ucr_high, pla_align, kp->kp_boundary, 760 &pgl, pla_maxseg, pla_flags)) { 761 return (NULL); 762 } 763 764 #ifdef __HAVE_PMAP_DIRECT 765 /* 766 * Only use direct mappings for single page or single segment 767 * allocations. 768 */ 769 if (kv->kv_singlepage || kp->kp_maxseg == 1) { 770 TAILQ_FOREACH(pg, &pgl, pageq) { 771 va = pmap_map_direct(pg); 772 if (pg == TAILQ_FIRST(&pgl)) 773 sva = va; 774 } 775 return ((void *)sva); 776 } 777 #endif 778 alloc_va: 779 prot = PROT_READ | PROT_WRITE; 780 781 if (kp->kp_pageable) { 782 KASSERT(kp->kp_object); 783 KASSERT(!kv->kv_singlepage); 784 } else { 785 KASSERT(kp->kp_object == NULL); 786 } 787 788 if (kv->kv_singlepage) { 789 KASSERT(sz == PAGE_SIZE); 790 #ifdef __HAVE_PMAP_DIRECT 791 panic("km_alloc: DIRECT single page"); 792 #else 793 mtx_enter(&uvm_km_pages.mtx); 794 while (uvm_km_pages.free == 0) { 795 if (kd->kd_waitok == 0) { 796 mtx_leave(&uvm_km_pages.mtx); 797 uvm_pglistfree(&pgl); 798 return NULL; 799 } 800 msleep_nsec(&uvm_km_pages.free, &uvm_km_pages.mtx, 801 PVM, "getpage", INFSLP); 802 } 803 va = uvm_km_pages.page[--uvm_km_pages.free]; 804 if (uvm_km_pages.free < uvm_km_pages.lowat && 805 curproc != uvm_km_pages.km_proc) { 806 if (kd->kd_slowdown) 807 *kd->kd_slowdown = 1; 808 wakeup(&uvm_km_pages.km_proc); 809 } 810 mtx_leave(&uvm_km_pages.mtx); 811 #endif 812 } else { 813 struct uvm_object *uobj = NULL; 814 815 if (kd->kd_trylock) 816 mapflags |= UVM_KMF_TRYLOCK; 817 818 if (kp->kp_object) 819 uobj = *kp->kp_object; 820 try_map: 821 map = *kv->kv_map; 822 va = vm_map_min(map); 823 if (uvm_map(map, &va, sz, uobj, kd->kd_prefer, 824 kv->kv_align, UVM_MAPFLAG(prot, prot, MAP_INHERIT_NONE, 825 MADV_RANDOM, mapflags))) { 826 if (kv->kv_wait && kd->kd_waitok) { 827 tsleep_nsec(map, PVM, "km_allocva", INFSLP); 828 goto try_map; 829 } 830 uvm_pglistfree(&pgl); 831 return (NULL); 832 } 833 } 834 sva = va; 835 TAILQ_FOREACH(pg, &pgl, pageq) { 836 if (kp->kp_pageable) 837 pmap_enter(pmap_kernel(), va, VM_PAGE_TO_PHYS(pg), 838 prot, prot | PMAP_WIRED); 839 else 840 pmap_kenter_pa(va, VM_PAGE_TO_PHYS(pg), prot); 841 va += PAGE_SIZE; 842 } 843 pmap_update(pmap_kernel()); 844 return ((void *)sva); 845 } 846 847 void 848 km_free(void *v, size_t sz, const struct kmem_va_mode *kv, 849 const struct kmem_pa_mode *kp) 850 { 851 vaddr_t sva, eva, va; 852 struct vm_page *pg; 853 struct pglist pgl; 854 855 sva = (vaddr_t)v; 856 eva = sva + sz; 857 858 if (kp->kp_nomem) 859 goto free_va; 860 861 #ifdef __HAVE_PMAP_DIRECT 862 if (kv->kv_singlepage || kp->kp_maxseg == 1) { 863 TAILQ_INIT(&pgl); 864 for (va = sva; va < eva; va += PAGE_SIZE) { 865 pg = pmap_unmap_direct(va); 866 TAILQ_INSERT_TAIL(&pgl, pg, pageq); 867 } 868 uvm_pglistfree(&pgl); 869 return; 870 } 871 #else 872 if (kv->kv_singlepage) { 873 struct uvm_km_free_page *fp = v; 874 875 mtx_enter(&uvm_km_pages.mtx); 876 fp->next = uvm_km_pages.freelist; 877 uvm_km_pages.freelist = fp; 878 if (uvm_km_pages.freelistlen++ > 16) 879 wakeup(&uvm_km_pages.km_proc); 880 mtx_leave(&uvm_km_pages.mtx); 881 return; 882 } 883 #endif 884 885 if (kp->kp_pageable) { 886 pmap_remove(pmap_kernel(), sva, eva); 887 pmap_update(pmap_kernel()); 888 } else { 889 TAILQ_INIT(&pgl); 890 for (va = sva; va < eva; va += PAGE_SIZE) { 891 paddr_t pa; 892 893 if (!pmap_extract(pmap_kernel(), va, &pa)) 894 continue; 895 896 pg = PHYS_TO_VM_PAGE(pa); 897 if (pg == NULL) { 898 panic("km_free: unmanaged page 0x%lx", pa); 899 } 900 TAILQ_INSERT_TAIL(&pgl, pg, pageq); 901 } 902 pmap_kremove(sva, sz); 903 pmap_update(pmap_kernel()); 904 uvm_pglistfree(&pgl); 905 } 906 free_va: 907 uvm_unmap(*kv->kv_map, sva, eva); 908 if (kv->kv_wait) 909 wakeup(*kv->kv_map); 910 } 911 912 const struct kmem_va_mode kv_any = { 913 .kv_map = &kernel_map, 914 }; 915 916 const struct kmem_va_mode kv_intrsafe = { 917 .kv_map = &kmem_map, 918 }; 919 920 const struct kmem_va_mode kv_page = { 921 .kv_singlepage = 1 922 }; 923 924 const struct kmem_pa_mode kp_dirty = { 925 .kp_constraint = &no_constraint 926 }; 927 928 const struct kmem_pa_mode kp_dma = { 929 .kp_constraint = &dma_constraint 930 }; 931 932 const struct kmem_pa_mode kp_dma_contig = { 933 .kp_constraint = &dma_constraint, 934 .kp_maxseg = 1 935 }; 936 937 const struct kmem_pa_mode kp_dma_zero = { 938 .kp_constraint = &dma_constraint, 939 .kp_zero = 1 940 }; 941 942 const struct kmem_pa_mode kp_zero = { 943 .kp_constraint = &no_constraint, 944 .kp_zero = 1 945 }; 946 947 const struct kmem_pa_mode kp_pageable = { 948 .kp_object = &uvm.kernel_object, 949 .kp_pageable = 1 950 /* XXX - kp_nomem, maybe, but we'll need to fix km_free. */ 951 }; 952 953 const struct kmem_pa_mode kp_none = { 954 .kp_nomem = 1 955 }; 956 957 const struct kmem_dyn_mode kd_waitok = { 958 .kd_waitok = 1, 959 .kd_prefer = UVM_UNKNOWN_OFFSET 960 }; 961 962 const struct kmem_dyn_mode kd_nowait = { 963 .kd_prefer = UVM_UNKNOWN_OFFSET 964 }; 965 966 const struct kmem_dyn_mode kd_trylock = { 967 .kd_trylock = 1, 968 .kd_prefer = UVM_UNKNOWN_OFFSET 969 };
Cache object: a9dc06628664e8bea4b9f3bdcb3ce1b0
[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]