uvm_page.h revision 1.96
1 /* $NetBSD: uvm_page.h,v 1.96 2020/01/15 17:55:45 ad Exp $ */ 2 3 /* 4 * Copyright (c) 1997 Charles D. Cranor and Washington University. 5 * Copyright (c) 1991, 1993, The Regents of the University of California. 6 * 7 * All rights reserved. 8 * 9 * This code is derived from software contributed to Berkeley by 10 * The Mach Operating System project at Carnegie-Mellon University. 11 * 12 * Redistribution and use in source and binary forms, with or without 13 * modification, are permitted provided that the following conditions 14 * are met: 15 * 1. Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 2. Redistributions in binary form must reproduce the above copyright 18 * notice, this list of conditions and the following disclaimer in the 19 * documentation and/or other materials provided with the distribution. 20 * 3. Neither the name of the University nor the names of its contributors 21 * may be used to endorse or promote products derived from this software 22 * without specific prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 * 36 * @(#)vm_page.h 7.3 (Berkeley) 4/21/91 37 * from: Id: uvm_page.h,v 1.1.2.6 1998/02/04 02:31:42 chuck Exp 38 * 39 * 40 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 41 * All rights reserved. 42 * 43 * Permission to use, copy, modify and distribute this software and 44 * its documentation is hereby granted, provided that both the copyright 45 * notice and this permission notice appear in all copies of the 46 * software, derivative works or modified versions, and any portions 47 * thereof, and that both notices appear in supporting documentation. 48 * 49 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 50 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 51 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 52 * 53 * Carnegie Mellon requests users of this software to return to 54 * 55 * Software Distribution Coordinator or Software.Distribution (at) CS.CMU.EDU 56 * School of Computer Science 57 * Carnegie Mellon University 58 * Pittsburgh PA 15213-3890 59 * 60 * any improvements or extensions that they make and grant Carnegie the 61 * rights to redistribute these changes. 62 */ 63 64 #ifndef _UVM_UVM_PAGE_H_ 65 #define _UVM_UVM_PAGE_H_ 66 67 #include <uvm/uvm_extern.h> 68 #include <uvm/uvm_pglist.h> 69 70 /* 71 * Management of resident (logical) pages. 72 * 73 * Each resident page has a vm_page structure, indexed by page number. 74 * There are several lists in the structure: 75 * 76 * - A red-black tree rooted with the containing object is used to 77 * quickly perform object+offset lookups. 78 * - A list of all pages for a given object, for a quick deactivation 79 * at a time of deallocation. 80 * - An ordered list of pages due for pageout. 81 * 82 * In addition, the structure contains the object and offset to which 83 * this page belongs (for pageout) and sundry status bits. 84 * 85 * Note that the page structure has no lock of its own. The page is 86 * generally protected by its owner's lock (UVM object or amap/anon). 87 * It should be noted that UVM has to serialize pmap(9) operations on 88 * the managed pages, e.g. for pmap_enter() calls. Hence, the lock 89 * order is as follows: 90 * 91 * [vmpage-owner-lock] -> 92 * any pmap locks (e.g. PV hash lock) 93 * 94 * Since the kernel is always self-consistent, no serialization is 95 * required for unmanaged mappings, e.g. for pmap_kenter_pa() calls. 96 * 97 * Field markings and the corresponding locks: 98 * 99 * f: free page queue lock, uvm_fpageqlock 100 * o: page owner (uvm_object::vmobjlock, vm_amap::am_lock, vm_anon::an_lock) 101 * i: vm_page::interlock 102 * => flags set and cleared only with o&i held can 103 * safely be tested for with only o held. 104 * o,i: o|i for read, o&i for write (depends on context - if could be loaned) 105 * => see uvm_loan.c 106 * w: wired page queue or uvm_pglistalloc: 107 * => wired page queue: o&i to change, stable from wire to unwire 108 * XXX What about concurrent or nested wire? 109 * => uvm_pglistalloc: owned by caller 110 * ?: locked by pmap or assumed page owner's lock 111 * p: locked by pagedaemon policy module (pdpolicy) 112 * c: cpu private 113 * s: stable, does not change 114 * 115 * UVM and pmap(9) may use uvm_page_owner_locked_p() to assert whether the 116 * page owner's lock is acquired. 117 * 118 * A page can have one of four identities: 119 * 120 * o free 121 * => pageq.list is entry on global free page queue 122 * => uanon is unused (or (void *)0xdeadbeef for DEBUG) 123 * => uobject is unused (or (void *)0xdeadbeef for DEBUG) 124 * => PG_FREE is set in flags 125 * o owned by a uvm_object 126 * => pageq.queue is entry on wired page queue, if any 127 * => uanon is NULL or the vm_anon to which it has been O->A loaned 128 * => uobject is owner 129 * o owned by a vm_anon 130 * => pageq is unused (XXX correct?) 131 * => uanon is owner 132 * => uobject is NULL 133 * => PG_ANON is set in flags 134 * o allocated by uvm_pglistalloc 135 * => pageq.queue is entry on resulting pglist, owned by caller 136 * => uanon is unused 137 * => uobject is unused 138 * 139 * The following transitions are allowed: 140 * 141 * - uvm_pagealloc: free -> owned by a uvm_object/vm_anon 142 * - uvm_pagefree: owned by a uvm_object/vm_anon -> free 143 * - uvm_pglistalloc: free -> allocated by uvm_pglistalloc 144 * - uvm_pglistfree: allocated by uvm_pglistalloc -> free 145 * 146 * On the ordering of fields: 147 * 148 * The fields most heavily used by the page allocator and uvmpdpol are 149 * clustered together at the start of the structure, so that while under 150 * global lock it's more likely that only one cache line for each page need 151 * be touched. 152 */ 153 154 struct vm_page { 155 union { 156 TAILQ_ENTRY(vm_page) queue; /* w: wired page queue 157 * or uvm_pglistalloc output */ 158 LIST_ENTRY(vm_page) list; /* f: global free page queue */ 159 } pageq; 160 TAILQ_ENTRY(vm_page) pdqueue; /* p: pagedaemon queue */ 161 kmutex_t interlock; /* s: lock on identity */ 162 uint32_t pqflags; /* i: pagedaemon flags */ 163 uint32_t flags; /* o: object flags */ 164 paddr_t phys_addr; /* o: physical address of pg */ 165 uint32_t loan_count; /* o,i: num. active loans */ 166 uint32_t wire_count; /* o,i: wired down map refs */ 167 struct vm_anon *uanon; /* o,i: anon */ 168 struct uvm_object *uobject; /* o,i: object */ 169 voff_t offset; /* o: offset into object */ 170 171 #ifdef __HAVE_VM_PAGE_MD 172 struct vm_page_md mdpage; /* ?: pmap-specific data */ 173 #endif 174 175 #if defined(UVM_PAGE_TRKOWN) 176 /* debugging fields to track page ownership */ 177 pid_t owner; /* proc that set PG_BUSY */ 178 lwpid_t lowner; /* lwp that set PG_BUSY */ 179 const char *owner_tag; /* why it was set busy */ 180 #endif 181 }; 182 183 /* 184 * Overview of UVM page flags. 185 * 186 * Locking notes: 187 * 188 * PG_, struct vm_page::flags => locked by owner 189 * PG_AOBJ => additionally locked by vm_page::interlock 190 * PG_ANON => additionally locked by vm_page::interlock 191 * PG_FREE => additionally locked by uvm_fpageqlock 192 * for uvm_pglistalloc() 193 * 194 * Flag descriptions: 195 * 196 * PG_CLEAN: 197 * Page is known clean. 198 * The contents of the page is consistent with its backing store. 199 * 200 * PG_DIRTY: 201 * Page is known dirty. 202 * To avoid losing data, the contents of the page should be written 203 * back to the backing store before freeing the page. 204 * 205 * PG_BUSY: 206 * Page is long-term locked, usually because of I/O (transfer from the 207 * page memory to the backing store) is in progress. LWP attempting 208 * to access the page shall set PG_WANTED and wait. 209 * 210 * PG_WANTED: 211 * Indicates that the page, which is currently PG_BUSY, is wanted by 212 * some other LWP. The page owner (i.e. LWP which set PG_BUSY) is 213 * responsible to clear both flags and wake up any waiters once it has 214 * released the long-term lock (PG_BUSY). 215 * 216 * PG_PAGEOUT: 217 * Indicates that the page is being paged-out in preparation for 218 * being freed. 219 * 220 * PG_RELEASED: 221 * Indicates that the page, which is currently PG_BUSY, should be freed 222 * after the release of long-term lock. It is responsibility of the 223 * owning LWP (i.e. which set PG_BUSY) to do it. 224 * 225 * PG_FAKE: 226 * Page has been allocated, but not yet initialised. The flag is used 227 * to avoid overwriting of valid data, e.g. to prevent read from the 228 * backing store when in-core data is newer. 229 * 230 * PG_RDONLY: 231 * Indicates that the page must be mapped read-only. 232 * 233 * PG_ZERO: 234 * Indicates that the page has been pre-zeroed. This flag is only 235 * set when the page is not in the queues and is cleared when the 236 * page is placed on the free list. 237 * 238 * PG_MARKER: 239 * Dummy marker page, generally used for list traversal. 240 */ 241 242 /* 243 * if you want to renumber PG_CLEAN and PG_DIRTY, check __CTASSERTs in 244 * uvm_page_status.c first. 245 */ 246 247 #define PG_CLEAN 0x00000001 /* page is known clean */ 248 #define PG_DIRTY 0x00000002 /* page is known dirty */ 249 #define PG_BUSY 0x00000004 /* page is locked */ 250 #define PG_WANTED 0x00000008 /* someone is waiting for page */ 251 #define PG_PAGEOUT 0x00000010 /* page to be freed for pagedaemon */ 252 #define PG_RELEASED 0x00000020 /* page to be freed when unbusied */ 253 #define PG_FAKE 0x00000040 /* page is not yet initialized */ 254 #define PG_RDONLY 0x00000080 /* page must be mapped read-only */ 255 #define PG_ZERO 0x00000100 /* page is pre-zero'd */ 256 #define PG_TABLED 0x00000200 /* page is tabled in object */ 257 #define PG_AOBJ 0x00000400 /* page is part of an anonymous 258 uvm_object */ 259 #define PG_ANON 0x00000800 /* page is part of an anon, rather 260 than an uvm_object */ 261 #define PG_FILE 0x00001000 /* file backed (non-anonymous) */ 262 #define PG_READAHEAD 0x00002000 /* read-ahead but not "hit" yet */ 263 #define PG_FREE 0x00004000 /* page is on free list */ 264 #define PG_MARKER 0x00008000 /* dummy marker page */ 265 #define PG_PAGER1 0x00010000 /* pager-specific flag */ 266 267 #define PG_STAT (PG_ANON|PG_AOBJ|PG_FILE) 268 #define PG_SWAPBACKED (PG_ANON|PG_AOBJ) 269 270 #define UVM_PGFLAGBITS \ 271 "\20\1CLEAN\2DIRTY\3BUSY\4WANTED" \ 272 "\5PAGEOUT\6RELEASED\7FAKE\10RDONLY" \ 273 "\11ZERO\12TABLED\13AOBJ\14ANON" \ 274 "\15FILE\16READAHEAD\17FREE\20MARKER" \ 275 "\21PAGER1" 276 277 /* 278 * uvmpdpol state flags. 279 * 280 * => may only be changed with pg->interlock held. 281 * => changing them is the responsibility of uvmpdpol .. 282 * => .. but uvm_page needs to know about them in order to purge updates. 283 * => PQ_PRIVATE is private to the individual uvmpdpol implementation. 284 */ 285 286 #define PQ_INTENT_A 0x00000000 /* intend activation */ 287 #define PQ_INTENT_I 0x00000001 /* intend deactivation */ 288 #define PQ_INTENT_E 0x00000002 /* intend enqueue */ 289 #define PQ_INTENT_D 0x00000003 /* intend dequeue */ 290 #define PQ_INTENT_MASK 0x00000003 /* mask of intended state */ 291 #define PQ_INTENT_SET 0x00000004 /* not realized yet */ 292 #define PQ_INTENT_QUEUED 0x00000008 /* queued for processing */ 293 #define PQ_PRIVATE 0xfffffff0 294 295 /* 296 * physical memory layout structure 297 * 298 * MD vmparam.h must #define: 299 * VM_PHYSEG_MAX = max number of physical memory segments we support 300 * (if this is "1" then we revert to a "contig" case) 301 * VM_PHYSSEG_STRAT: memory sort/search options (for VM_PHYSEG_MAX > 1) 302 * - VM_PSTRAT_RANDOM: linear search (random order) 303 * - VM_PSTRAT_BSEARCH: binary search (sorted by address) 304 * - VM_PSTRAT_BIGFIRST: linear search (sorted by largest segment first) 305 * - others? 306 * XXXCDC: eventually we should purge all left-over global variables... 307 */ 308 #define VM_PSTRAT_RANDOM 1 309 #define VM_PSTRAT_BSEARCH 2 310 #define VM_PSTRAT_BIGFIRST 3 311 312 #ifdef _KERNEL 313 314 /* 315 * globals 316 */ 317 318 extern bool vm_page_zero_enable; 319 320 /* 321 * prototypes: the following prototypes define the interface to pages 322 */ 323 324 void uvm_page_init(vaddr_t *, vaddr_t *); 325 #if defined(UVM_PAGE_TRKOWN) 326 void uvm_page_own(struct vm_page *, const char *); 327 #endif 328 #if !defined(PMAP_STEAL_MEMORY) 329 bool uvm_page_physget(paddr_t *); 330 #endif 331 void uvm_page_recolor(int); 332 void uvm_page_rebucket(void); 333 void uvm_pageidlezero(void); 334 335 void uvm_pageactivate(struct vm_page *); 336 vaddr_t uvm_pageboot_alloc(vsize_t); 337 void uvm_pagecopy(struct vm_page *, struct vm_page *); 338 void uvm_pagedeactivate(struct vm_page *); 339 void uvm_pagedequeue(struct vm_page *); 340 void uvm_pageenqueue(struct vm_page *); 341 void uvm_pagefree(struct vm_page *); 342 void uvm_pagelock(struct vm_page *); 343 void uvm_pagelock2(struct vm_page *, struct vm_page *); 344 void uvm_pageunlock(struct vm_page *); 345 void uvm_pageunlock2(struct vm_page *, struct vm_page *); 346 void uvm_page_unbusy(struct vm_page **, int); 347 struct vm_page *uvm_pagelookup(struct uvm_object *, voff_t); 348 void uvm_pageunwire(struct vm_page *); 349 void uvm_pagewire(struct vm_page *); 350 void uvm_pagezero(struct vm_page *); 351 bool uvm_pageismanaged(paddr_t); 352 bool uvm_page_owner_locked_p(struct vm_page *); 353 void uvm_pgfl_lock(void); 354 void uvm_pgfl_unlock(void); 355 unsigned int uvm_pagegetdirty(struct vm_page *); 356 void uvm_pagemarkdirty(struct vm_page *, unsigned int); 357 bool uvm_pagecheckdirty(struct vm_page *, bool); 358 bool uvm_pagereadonly_p(struct vm_page *); 359 bool uvm_page_locked_p(struct vm_page *); 360 361 int uvm_page_lookup_freelist(struct vm_page *); 362 363 struct vm_page *uvm_phys_to_vm_page(paddr_t); 364 paddr_t uvm_vm_page_to_phys(const struct vm_page *); 365 366 #if defined(PMAP_DIRECT) 367 extern bool ubc_direct; 368 int uvm_direct_process(struct vm_page **, u_int, voff_t, vsize_t, 369 int (*)(void *, size_t, void *), void *); 370 #endif 371 372 /* 373 * page dirtiness status for uvm_pagegetdirty and uvm_pagemarkdirty 374 * 375 * UNKNOWN means that we need to consult pmap to know if the page is 376 * dirty or not. 377 * basically, UVM_PAGE_STATUS_CLEAN implies that the page has no writable 378 * mapping. 379 * 380 * if you want to renumber these, check __CTASSERTs in 381 * uvm_page_status.c first. 382 */ 383 384 #define UVM_PAGE_STATUS_UNKNOWN 0 385 #define UVM_PAGE_STATUS_CLEAN 1 386 #define UVM_PAGE_STATUS_DIRTY 2 387 #define UVM_PAGE_NUM_STATUS 3 388 389 /* 390 * macros 391 */ 392 393 #define VM_PAGE_TO_PHYS(entry) uvm_vm_page_to_phys(entry) 394 395 #ifdef __HAVE_VM_PAGE_MD 396 #define VM_PAGE_TO_MD(pg) (&(pg)->mdpage) 397 #endif 398 399 /* 400 * Compute the page color for a given page. 401 */ 402 #define VM_PGCOLOR(pg) \ 403 (atop(VM_PAGE_TO_PHYS((pg))) & uvmexp.colormask) 404 #define PHYS_TO_VM_PAGE(pa) uvm_phys_to_vm_page(pa) 405 406 /* 407 * VM_PAGE_IS_FREE() can't tell if the page is on global free list, or a 408 * per-CPU cache. If you need to be certain, pause caching. 409 */ 410 #define VM_PAGE_IS_FREE(entry) ((entry)->flags & PG_FREE) 411 412 /* 413 * Use the lower 10 bits of pg->phys_addr to cache some some locators for 414 * the page. This implies that the smallest possible page size is 1kB, and 415 * that nobody should use pg->phys_addr directly (use VM_PAGE_TO_PHYS()). 416 * 417 * - 5 bits for the freelist index, because uvm_page_lookup_freelist() 418 * traverses an rbtree and therefore features prominently in traces 419 * captured during performance test. It would probably be more useful to 420 * cache physseg index here because freelist can be inferred from physseg, 421 * but it requires changes to allocation for UVM_HOTPLUG, so for now we'll 422 * go with freelist. 423 * 424 * - 5 bits for "bucket", a way for us to categorise pages further as 425 * needed (e.g. NUMA node). 426 * 427 * None of this is set in stone; it can be adjusted as needed. 428 */ 429 430 #define UVM_PHYSADDR_FREELIST __BITS(0,4) 431 #define UVM_PHYSADDR_BUCKET __BITS(5,9) 432 433 static inline unsigned 434 uvm_page_get_freelist(struct vm_page *pg) 435 { 436 unsigned fl = __SHIFTOUT(pg->phys_addr, UVM_PHYSADDR_FREELIST); 437 KASSERT(fl == (unsigned)uvm_page_lookup_freelist(pg)); 438 return fl; 439 } 440 441 static inline unsigned 442 uvm_page_get_bucket(struct vm_page *pg) 443 { 444 return __SHIFTOUT(pg->phys_addr, UVM_PHYSADDR_BUCKET); 445 } 446 447 static inline void 448 uvm_page_set_freelist(struct vm_page *pg, unsigned fl) 449 { 450 KASSERT(fl < 32); 451 pg->phys_addr &= ~UVM_PHYSADDR_FREELIST; 452 pg->phys_addr |= __SHIFTIN(fl, UVM_PHYSADDR_FREELIST); 453 } 454 455 static inline void 456 uvm_page_set_bucket(struct vm_page *pg, unsigned b) 457 { 458 KASSERT(b < 32); 459 pg->phys_addr &= ~UVM_PHYSADDR_BUCKET; 460 pg->phys_addr |= __SHIFTIN(b, UVM_PHYSADDR_BUCKET); 461 } 462 463 #ifdef DEBUG 464 void uvm_pagezerocheck(struct vm_page *); 465 #endif /* DEBUG */ 466 467 #endif /* _KERNEL */ 468 469 #endif /* _UVM_UVM_PAGE_H_ */ 470
Indexes created Fri Sep 26 20:09:58 GMT 2025