xref: /freebsd/sys/vm/vm_page.h (revision 0c54932d50a0cbffdd083bf6b2e8d587902f90c9)
1 /*-
2  * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
3  *
4  * Copyright (c) 1991, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * This code is derived from software contributed to Berkeley by
8  * The Mach Operating System project at Carnegie-Mellon University.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *	from: @(#)vm_page.h	8.2 (Berkeley) 12/13/93
35  *
36  *
37  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38  * All rights reserved.
39  *
40  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
41  *
42  * Permission to use, copy, modify and distribute this software and
43  * its documentation is hereby granted, provided that both the copyright
44  * notice and this permission notice appear in all copies of the
45  * software, derivative works or modified versions, and any portions
46  * thereof, and that both notices appear in supporting documentation.
47  *
48  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
51  *
52  * Carnegie Mellon requests users of this software to return to
53  *
54  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
55  *  School of Computer Science
56  *  Carnegie Mellon University
57  *  Pittsburgh PA 15213-3890
58  *
59  * any improvements or extensions that they make and grant Carnegie the
60  * rights to redistribute these changes.
61  *
62  * $FreeBSD$
63  */
64 
65 /*
66  *	Resident memory system definitions.
67  */
68 
69 #ifndef	_VM_PAGE_
70 #define	_VM_PAGE_
71 
72 #include <vm/pmap.h>
73 
74 /*
75  *	Management of resident (logical) pages.
76  *
77  *	A small structure is kept for each resident
78  *	page, indexed by page number.  Each structure
79  *	is an element of several collections:
80  *
81  *		A radix tree used to quickly
82  *		perform object/offset lookups
83  *
84  *		A list of all pages for a given object,
85  *		so they can be quickly deactivated at
86  *		time of deallocation.
87  *
88  *		An ordered list of pages due for pageout.
89  *
90  *	In addition, the structure contains the object
91  *	and offset to which this page belongs (for pageout),
92  *	and sundry status bits.
93  *
94  *	In general, operations on this structure's mutable fields are
95  *	synchronized using either one of or a combination of locks.  If a
96  *	field is annotated with two of these locks then holding either is
97  *	sufficient for read access but both are required for write access.
98  *	The queue lock for a page depends on the value of its queue field and is
99  *	described in detail below.
100  *
101  *	The following annotations are possible:
102  *	(A) the field must be accessed using atomic(9) and may require
103  *	    additional synchronization.
104  *	(B) the page busy lock.
105  *	(C) the field is immutable.
106  *	(F) the per-domain lock for the free queues.
107  *	(M) Machine dependent, defined by pmap layer.
108  *	(O) the object that the page belongs to.
109  *	(Q) the page's queue lock.
110  *
111  *	The busy lock is an embedded reader-writer lock that protects the
112  *	page's contents and identity (i.e., its <object, pindex> tuple) as
113  *	well as certain valid/dirty modifications.  To avoid bloating the
114  *	the page structure, the busy lock lacks some of the features available
115  *	the kernel's general-purpose synchronization primitives.  As a result,
116  *	busy lock ordering rules are not verified, lock recursion is not
117  *	detected, and an attempt to xbusy a busy page or sbusy an xbusy page
118  *	results will trigger a panic rather than causing the thread to block.
119  *	vm_page_sleep_if_busy() can be used to sleep until the page's busy
120  *	state changes, after which the caller must re-lookup the page and
121  *	re-evaluate its state.  vm_page_busy_acquire() will block until
122  *	the lock is acquired.
123  *
124  *	The valid field is protected by the page busy lock (B) and object
125  *	lock (O).  Transitions from invalid to valid are generally done
126  *	via I/O or zero filling and do not require the object lock.
127  *	These must be protected with the busy lock to prevent page-in or
128  *	creation races.  Page invalidation generally happens as a result
129  *	of truncate or msync.  When invalidated, pages must not be present
130  *	in pmap and must hold the object lock to prevent concurrent
131  *	speculative read-only mappings that do not require busy.  I/O
132  *	routines may check for validity without a lock if they are prepared
133  *	to handle invalidation races with higher level locks (vnode) or are
134  *	unconcerned with races so long as they hold a reference to prevent
135  *	recycling.  When a valid bit is set while holding a shared busy
136  *	lock (A) atomic operations are used to protect against concurrent
137  *	modification.
138  *
139  *	In contrast, the synchronization of accesses to the page's
140  *	dirty field is a mix of machine dependent (M) and busy (B).  In
141  *	the machine-independent layer, the page busy must be held to
142  *	operate on the field.  However, the pmap layer is permitted to
143  *	set all bits within the field without holding that lock.  If the
144  *	underlying architecture does not support atomic read-modify-write
145  *	operations on the field's type, then the machine-independent
146  *	layer uses a 32-bit atomic on the aligned 32-bit word that
147  *	contains the dirty field.  In the machine-independent layer,
148  *	the implementation of read-modify-write operations on the
149  *	field is encapsulated in vm_page_clear_dirty_mask().  An
150  *	exclusive busy lock combined with pmap_remove_{write/all}() is the
151  *	only way to ensure a page can not become dirty.  I/O generally
152  *	removes the page from pmap to ensure exclusive access and atomic
153  *	writes.
154  *
155  *	The ref_count field tracks references to the page.  References that
156  *	prevent the page from being reclaimable are called wirings and are
157  *	counted in the low bits of ref_count.  The containing object's
158  *	reference, if one exists, is counted using the VPRC_OBJREF bit in the
159  *	ref_count field.  Additionally, the VPRC_BLOCKED bit is used to
160  *	atomically check for wirings and prevent new wirings via
161  *	pmap_extract_and_hold().  When a page belongs to an object, it may be
162  *	wired only when the object is locked, or the page is busy, or by
163  *	pmap_extract_and_hold().  As a result, if the object is locked and the
164  *	page is not busy (or is exclusively busied by the current thread), and
165  *	the page is unmapped, its wire count will not increase.  The ref_count
166  *	field is updated using atomic operations in most cases, except when it
167  *	is known that no other references to the page exist, such as in the page
168  *	allocator.  A page may be present in the page queues, or even actively
169  *	scanned by the page daemon, without an explicitly counted referenced.
170  *	The page daemon must therefore handle the possibility of a concurrent
171  *	free of the page.
172  *
173  *	The queue state of a page consists of the queue and act_count fields of
174  *	its atomically updated state, and the subset of atomic flags specified
175  *	by PGA_QUEUE_STATE_MASK.  The queue field contains the page's page queue
176  *	index, or PQ_NONE if it does not belong to a page queue.  To modify the
177  *	queue field, the page queue lock corresponding to the old value must be
178  *	held, unless that value is PQ_NONE, in which case the queue index must
179  *	be updated using an atomic RMW operation.  There is one exception to
180  *	this rule: the page daemon may transition the queue field from
181  *	PQ_INACTIVE to PQ_NONE immediately prior to freeing the page during an
182  *	inactive queue scan.  At that point the page is already dequeued and no
183  *	other references to that vm_page structure can exist.  The PGA_ENQUEUED
184  *	flag, when set, indicates that the page structure is physically inserted
185  *	into the queue corresponding to the page's queue index, and may only be
186  *	set or cleared with the corresponding page queue lock held.
187  *
188  *	To avoid contention on page queue locks, page queue operations (enqueue,
189  *	dequeue, requeue) are batched using fixed-size per-CPU queues.  A
190  *	deferred operation is requested by setting one of the flags in
191  *	PGA_QUEUE_OP_MASK and inserting an entry into a batch queue.  When a
192  *	queue is full, an attempt to insert a new entry will lock the page
193  *	queues and trigger processing of the pending entries.  The
194  *	type-stability of vm_page structures is crucial to this scheme since the
195  *	processing of entries in a given batch queue may be deferred
196  *	indefinitely.  In particular, a page may be freed with pending batch
197  *	queue entries.  The page queue operation flags must be set using atomic
198  *	RWM operations.
199  */
200 
201 #if PAGE_SIZE == 4096
202 #define VM_PAGE_BITS_ALL 0xffu
203 typedef uint8_t vm_page_bits_t;
204 #elif PAGE_SIZE == 8192
205 #define VM_PAGE_BITS_ALL 0xffffu
206 typedef uint16_t vm_page_bits_t;
207 #elif PAGE_SIZE == 16384
208 #define VM_PAGE_BITS_ALL 0xffffffffu
209 typedef uint32_t vm_page_bits_t;
210 #elif PAGE_SIZE == 32768
211 #define VM_PAGE_BITS_ALL 0xfffffffffffffffflu
212 typedef uint64_t vm_page_bits_t;
213 #endif
214 
215 typedef union vm_page_astate {
216 	struct {
217 		uint16_t flags;
218 		uint8_t	queue;
219 		uint8_t act_count;
220 	};
221 	uint32_t _bits;
222 } vm_page_astate_t;
223 
224 struct vm_page {
225 	union {
226 		TAILQ_ENTRY(vm_page) q; /* page queue or free list (Q) */
227 		struct {
228 			SLIST_ENTRY(vm_page) ss; /* private slists */
229 		} s;
230 		struct {
231 			u_long p;
232 			u_long v;
233 		} memguard;
234 		struct {
235 			void *slab;
236 			void *zone;
237 		} uma;
238 	} plinks;
239 	TAILQ_ENTRY(vm_page) listq;	/* pages in same object (O) */
240 	vm_object_t object;		/* which object am I in (O) */
241 	vm_pindex_t pindex;		/* offset into object (O,P) */
242 	vm_paddr_t phys_addr;		/* physical address of page (C) */
243 	struct md_page md;		/* machine dependent stuff */
244 	u_int ref_count;		/* page references (A) */
245 	u_int busy_lock;		/* busy owners lock (A) */
246 	union vm_page_astate a;		/* state accessed atomically (A) */
247 	uint8_t order;			/* index of the buddy queue (F) */
248 	uint8_t pool;			/* vm_phys freepool index (F) */
249 	uint8_t flags;			/* page PG_* flags (P) */
250 	uint8_t oflags;			/* page VPO_* flags (O) */
251 	int8_t psind;			/* pagesizes[] index (O) */
252 	int8_t segind;			/* vm_phys segment index (C) */
253 	/* NOTE that these must support one bit per DEV_BSIZE in a page */
254 	/* so, on normal X86 kernels, they must be at least 8 bits wide */
255 	vm_page_bits_t valid;		/* valid DEV_BSIZE chunk map (O,B) */
256 	vm_page_bits_t dirty;		/* dirty DEV_BSIZE chunk map (M,B) */
257 };
258 
259 /*
260  * Special bits used in the ref_count field.
261  *
262  * ref_count is normally used to count wirings that prevent the page from being
263  * reclaimed, but also supports several special types of references that do not
264  * prevent reclamation.  Accesses to the ref_count field must be atomic unless
265  * the page is unallocated.
266  *
267  * VPRC_OBJREF is the reference held by the containing object.  It can set or
268  * cleared only when the corresponding object's write lock is held.
269  *
270  * VPRC_BLOCKED is used to atomically block wirings via pmap lookups while
271  * attempting to tear down all mappings of a given page.  The page busy lock and
272  * object write lock must both be held in order to set or clear this bit.
273  */
274 #define	VPRC_BLOCKED	0x40000000u	/* mappings are being removed */
275 #define	VPRC_OBJREF	0x80000000u	/* object reference, cleared with (O) */
276 #define	VPRC_WIRE_COUNT(c)	((c) & ~(VPRC_BLOCKED | VPRC_OBJREF))
277 #define	VPRC_WIRE_COUNT_MAX	(~(VPRC_BLOCKED | VPRC_OBJREF))
278 
279 /*
280  * Page flags stored in oflags:
281  *
282  * Access to these page flags is synchronized by the lock on the object
283  * containing the page (O).
284  *
285  * Note: VPO_UNMANAGED (used by OBJT_DEVICE, OBJT_PHYS and OBJT_SG)
286  * 	 indicates that the page is not under PV management but
287  * 	 otherwise should be treated as a normal page.  Pages not
288  * 	 under PV management cannot be paged out via the
289  * 	 object/vm_page_t because there is no knowledge of their pte
290  * 	 mappings, and such pages are also not on any PQ queue.
291  *
292  */
293 #define	VPO_KMEM_EXEC	0x01		/* kmem mapping allows execution */
294 #define	VPO_SWAPSLEEP	0x02		/* waiting for swap to finish */
295 #define	VPO_UNMANAGED	0x04		/* no PV management for page */
296 #define	VPO_SWAPINPROG	0x08		/* swap I/O in progress on page */
297 
298 /*
299  * Busy page implementation details.
300  * The algorithm is taken mostly by rwlock(9) and sx(9) locks implementation,
301  * even if the support for owner identity is removed because of size
302  * constraints.  Checks on lock recursion are then not possible, while the
303  * lock assertions effectiveness is someway reduced.
304  */
305 #define	VPB_BIT_SHARED		0x01
306 #define	VPB_BIT_EXCLUSIVE	0x02
307 #define	VPB_BIT_WAITERS		0x04
308 #define	VPB_BIT_FLAGMASK						\
309 	(VPB_BIT_SHARED | VPB_BIT_EXCLUSIVE | VPB_BIT_WAITERS)
310 
311 #define	VPB_SHARERS_SHIFT	3
312 #define	VPB_SHARERS(x)							\
313 	(((x) & ~VPB_BIT_FLAGMASK) >> VPB_SHARERS_SHIFT)
314 #define	VPB_SHARERS_WORD(x)	((x) << VPB_SHARERS_SHIFT | VPB_BIT_SHARED)
315 #define	VPB_ONE_SHARER		(1 << VPB_SHARERS_SHIFT)
316 
317 #define	VPB_SINGLE_EXCLUSIVE	VPB_BIT_EXCLUSIVE
318 #ifdef INVARIANTS
319 #define	VPB_CURTHREAD_EXCLUSIVE						\
320 	(VPB_BIT_EXCLUSIVE | ((u_int)(uintptr_t)curthread & ~VPB_BIT_FLAGMASK))
321 #else
322 #define	VPB_CURTHREAD_EXCLUSIVE	VPB_SINGLE_EXCLUSIVE
323 #endif
324 
325 #define	VPB_UNBUSIED		VPB_SHARERS_WORD(0)
326 
327 /* Freed lock blocks both shared and exclusive. */
328 #define	VPB_FREED		(0xffffffff - VPB_BIT_SHARED)
329 
330 #define	PQ_NONE		255
331 #define	PQ_INACTIVE	0
332 #define	PQ_ACTIVE	1
333 #define	PQ_LAUNDRY	2
334 #define	PQ_UNSWAPPABLE	3
335 #define	PQ_COUNT	4
336 
337 #ifndef VM_PAGE_HAVE_PGLIST
338 TAILQ_HEAD(pglist, vm_page);
339 #define VM_PAGE_HAVE_PGLIST
340 #endif
341 SLIST_HEAD(spglist, vm_page);
342 
343 #ifdef _KERNEL
344 extern vm_page_t bogus_page;
345 #endif	/* _KERNEL */
346 
347 extern struct mtx_padalign pa_lock[];
348 
349 #if defined(__arm__)
350 #define	PDRSHIFT	PDR_SHIFT
351 #elif !defined(PDRSHIFT)
352 #define PDRSHIFT	21
353 #endif
354 
355 #define	pa_index(pa)	((pa) >> PDRSHIFT)
356 #define	PA_LOCKPTR(pa)	((struct mtx *)(&pa_lock[pa_index(pa) % PA_LOCK_COUNT]))
357 #define	PA_LOCKOBJPTR(pa)	((struct lock_object *)PA_LOCKPTR((pa)))
358 #define	PA_LOCK(pa)	mtx_lock(PA_LOCKPTR(pa))
359 #define	PA_TRYLOCK(pa)	mtx_trylock(PA_LOCKPTR(pa))
360 #define	PA_UNLOCK(pa)	mtx_unlock(PA_LOCKPTR(pa))
361 #define	PA_UNLOCK_COND(pa) 			\
362 	do {		   			\
363 		if ((pa) != 0) {		\
364 			PA_UNLOCK((pa));	\
365 			(pa) = 0;		\
366 		}				\
367 	} while (0)
368 
369 #define	PA_LOCK_ASSERT(pa, a)	mtx_assert(PA_LOCKPTR(pa), (a))
370 
371 #if defined(KLD_MODULE) && !defined(KLD_TIED)
372 #define	vm_page_lock(m)		vm_page_lock_KBI((m), LOCK_FILE, LOCK_LINE)
373 #define	vm_page_unlock(m)	vm_page_unlock_KBI((m), LOCK_FILE, LOCK_LINE)
374 #define	vm_page_trylock(m)	vm_page_trylock_KBI((m), LOCK_FILE, LOCK_LINE)
375 #else	/* !KLD_MODULE */
376 #define	vm_page_lockptr(m)	(PA_LOCKPTR(VM_PAGE_TO_PHYS((m))))
377 #define	vm_page_lock(m)		mtx_lock(vm_page_lockptr((m)))
378 #define	vm_page_unlock(m)	mtx_unlock(vm_page_lockptr((m)))
379 #define	vm_page_trylock(m)	mtx_trylock(vm_page_lockptr((m)))
380 #endif
381 #if defined(INVARIANTS)
382 #define	vm_page_assert_locked(m)		\
383     vm_page_assert_locked_KBI((m), __FILE__, __LINE__)
384 #define	vm_page_lock_assert(m, a)		\
385     vm_page_lock_assert_KBI((m), (a), __FILE__, __LINE__)
386 #else
387 #define	vm_page_assert_locked(m)
388 #define	vm_page_lock_assert(m, a)
389 #endif
390 
391 /*
392  * The vm_page's aflags are updated using atomic operations.  To set or clear
393  * these flags, the functions vm_page_aflag_set() and vm_page_aflag_clear()
394  * must be used.  Neither these flags nor these functions are part of the KBI.
395  *
396  * PGA_REFERENCED may be cleared only if the page is locked.  It is set by
397  * both the MI and MD VM layers.  However, kernel loadable modules should not
398  * directly set this flag.  They should call vm_page_reference() instead.
399  *
400  * PGA_WRITEABLE is set exclusively on managed pages by pmap_enter().
401  * When it does so, the object must be locked, or the page must be
402  * exclusive busied.  The MI VM layer must never access this flag
403  * directly.  Instead, it should call pmap_page_is_write_mapped().
404  *
405  * PGA_EXECUTABLE may be set by pmap routines, and indicates that a page has
406  * at least one executable mapping.  It is not consumed by the MI VM layer.
407  *
408  * PGA_NOSYNC must be set and cleared with the page busy lock held.
409  *
410  * PGA_ENQUEUED is set and cleared when a page is inserted into or removed
411  * from a page queue, respectively.  It determines whether the plinks.q field
412  * of the page is valid.  To set or clear this flag, page's "queue" field must
413  * be a valid queue index, and the corresponding page queue lock must be held.
414  *
415  * PGA_DEQUEUE is set when the page is scheduled to be dequeued from a page
416  * queue, and cleared when the dequeue request is processed.  A page may
417  * have PGA_DEQUEUE set and PGA_ENQUEUED cleared, for instance if a dequeue
418  * is requested after the page is scheduled to be enqueued but before it is
419  * actually inserted into the page queue.
420  *
421  * PGA_REQUEUE is set when the page is scheduled to be enqueued or requeued
422  * in its page queue.
423  *
424  * PGA_REQUEUE_HEAD is a special flag for enqueuing pages near the head of
425  * the inactive queue, thus bypassing LRU.
426  *
427  * The PGA_DEQUEUE, PGA_REQUEUE and PGA_REQUEUE_HEAD flags must be set using an
428  * atomic RMW operation to ensure that the "queue" field is a valid queue index,
429  * and the corresponding page queue lock must be held when clearing any of the
430  * flags.
431  *
432  * PGA_SWAP_FREE is used to defer freeing swap space to the pageout daemon
433  * when the context that dirties the page does not have the object write lock
434  * held.
435  */
436 #define	PGA_WRITEABLE	0x0001		/* page may be mapped writeable */
437 #define	PGA_REFERENCED	0x0002		/* page has been referenced */
438 #define	PGA_EXECUTABLE	0x0004		/* page may be mapped executable */
439 #define	PGA_ENQUEUED	0x0008		/* page is enqueued in a page queue */
440 #define	PGA_DEQUEUE	0x0010		/* page is due to be dequeued */
441 #define	PGA_REQUEUE	0x0020		/* page is due to be requeued */
442 #define	PGA_REQUEUE_HEAD 0x0040		/* page requeue should bypass LRU */
443 #define	PGA_NOSYNC	0x0080		/* do not collect for syncer */
444 #define	PGA_SWAP_FREE	0x0100		/* page with swap space was dirtied */
445 #define	PGA_SWAP_SPACE	0x0200		/* page has allocated swap space */
446 
447 #define	PGA_QUEUE_OP_MASK	(PGA_DEQUEUE | PGA_REQUEUE | PGA_REQUEUE_HEAD)
448 #define	PGA_QUEUE_STATE_MASK	(PGA_ENQUEUED | PGA_QUEUE_OP_MASK)
449 
450 /*
451  * Page flags.  Updates to these flags are not synchronized, and thus they must
452  * be set during page allocation or free to avoid races.
453  *
454  * The PG_PCPU_CACHE flag is set at allocation time if the page was
455  * allocated from a per-CPU cache.  It is cleared the next time that the
456  * page is allocated from the physical memory allocator.
457  */
458 #define	PG_PCPU_CACHE	0x01		/* was allocated from per-CPU caches */
459 #define	PG_FICTITIOUS	0x02		/* physical page doesn't exist */
460 #define	PG_ZERO		0x04		/* page is zeroed */
461 #define	PG_MARKER	0x08		/* special queue marker page */
462 #define	PG_NODUMP	0x10		/* don't include this page in a dump */
463 
464 /*
465  * Misc constants.
466  */
467 #define ACT_DECLINE		1
468 #define ACT_ADVANCE		3
469 #define ACT_INIT		5
470 #define ACT_MAX			64
471 
472 #ifdef _KERNEL
473 
474 #include <sys/systm.h>
475 
476 #include <machine/atomic.h>
477 
478 /*
479  * Each pageable resident page falls into one of five lists:
480  *
481  *	free
482  *		Available for allocation now.
483  *
484  *	inactive
485  *		Low activity, candidates for reclamation.
486  *		This list is approximately LRU ordered.
487  *
488  *	laundry
489  *		This is the list of pages that should be
490  *		paged out next.
491  *
492  *	unswappable
493  *		Dirty anonymous pages that cannot be paged
494  *		out because no swap device is configured.
495  *
496  *	active
497  *		Pages that are "active", i.e., they have been
498  *		recently referenced.
499  *
500  */
501 
502 extern vm_page_t vm_page_array;		/* First resident page in table */
503 extern long vm_page_array_size;		/* number of vm_page_t's */
504 extern long first_page;			/* first physical page number */
505 
506 #define VM_PAGE_TO_PHYS(entry)	((entry)->phys_addr)
507 
508 /*
509  * PHYS_TO_VM_PAGE() returns the vm_page_t object that represents a memory
510  * page to which the given physical address belongs. The correct vm_page_t
511  * object is returned for addresses that are not page-aligned.
512  */
513 vm_page_t PHYS_TO_VM_PAGE(vm_paddr_t pa);
514 
515 /*
516  * Page allocation parameters for vm_page for the functions
517  * vm_page_alloc(), vm_page_grab(), vm_page_alloc_contig() and
518  * vm_page_alloc_freelist().  Some functions support only a subset
519  * of the flags, and ignore others, see the flags legend.
520  *
521  * The meaning of VM_ALLOC_ZERO differs slightly between the vm_page_alloc*()
522  * and the vm_page_grab*() functions.  See these functions for details.
523  *
524  * Bits 0 - 1 define class.
525  * Bits 2 - 15 dedicated for flags.
526  * Legend:
527  * (a) - vm_page_alloc() supports the flag.
528  * (c) - vm_page_alloc_contig() supports the flag.
529  * (f) - vm_page_alloc_freelist() supports the flag.
530  * (g) - vm_page_grab() supports the flag.
531  * (p) - vm_page_grab_pages() supports the flag.
532  * Bits above 15 define the count of additional pages that the caller
533  * intends to allocate.
534  */
535 #define VM_ALLOC_NORMAL		0
536 #define VM_ALLOC_INTERRUPT	1
537 #define VM_ALLOC_SYSTEM		2
538 #define	VM_ALLOC_CLASS_MASK	3
539 #define	VM_ALLOC_WAITOK		0x0008	/* (acf) Sleep and retry */
540 #define	VM_ALLOC_WAITFAIL	0x0010	/* (acf) Sleep and return error */
541 #define	VM_ALLOC_WIRED		0x0020	/* (acfgp) Allocate a wired page */
542 #define	VM_ALLOC_ZERO		0x0040	/* (acfgp) Allocate a prezeroed page */
543 #define	VM_ALLOC_NOOBJ		0x0100	/* (acg) No associated object */
544 #define	VM_ALLOC_NOBUSY		0x0200	/* (acgp) Do not excl busy the page */
545 #define	VM_ALLOC_NOCREAT	0x0400	/* (gp) Don't create a page */
546 #define	VM_ALLOC_IGN_SBUSY	0x1000	/* (gp) Ignore shared busy flag */
547 #define	VM_ALLOC_NODUMP		0x2000	/* (ag) don't include in dump */
548 #define	VM_ALLOC_SBUSY		0x4000	/* (acgp) Shared busy the page */
549 #define	VM_ALLOC_NOWAIT		0x8000	/* (acfgp) Do not sleep */
550 #define	VM_ALLOC_COUNT_SHIFT	16
551 #define	VM_ALLOC_COUNT(count)	((count) << VM_ALLOC_COUNT_SHIFT)
552 
553 #ifdef M_NOWAIT
554 static inline int
555 malloc2vm_flags(int malloc_flags)
556 {
557 	int pflags;
558 
559 	KASSERT((malloc_flags & M_USE_RESERVE) == 0 ||
560 	    (malloc_flags & M_NOWAIT) != 0,
561 	    ("M_USE_RESERVE requires M_NOWAIT"));
562 	pflags = (malloc_flags & M_USE_RESERVE) != 0 ? VM_ALLOC_INTERRUPT :
563 	    VM_ALLOC_SYSTEM;
564 	if ((malloc_flags & M_ZERO) != 0)
565 		pflags |= VM_ALLOC_ZERO;
566 	if ((malloc_flags & M_NODUMP) != 0)
567 		pflags |= VM_ALLOC_NODUMP;
568 	if ((malloc_flags & M_NOWAIT))
569 		pflags |= VM_ALLOC_NOWAIT;
570 	if ((malloc_flags & M_WAITOK))
571 		pflags |= VM_ALLOC_WAITOK;
572 	return (pflags);
573 }
574 #endif
575 
576 /*
577  * Predicates supported by vm_page_ps_test():
578  *
579  *	PS_ALL_DIRTY is true only if the entire (super)page is dirty.
580  *	However, it can be spuriously false when the (super)page has become
581  *	dirty in the pmap but that information has not been propagated to the
582  *	machine-independent layer.
583  */
584 #define	PS_ALL_DIRTY	0x1
585 #define	PS_ALL_VALID	0x2
586 #define	PS_NONE_BUSY	0x4
587 
588 bool vm_page_busy_acquire(vm_page_t m, int allocflags);
589 void vm_page_busy_downgrade(vm_page_t m);
590 int vm_page_busy_tryupgrade(vm_page_t m);
591 void vm_page_busy_sleep(vm_page_t m, const char *msg, bool nonshared);
592 void vm_page_busy_sleep_unlocked(vm_object_t obj, vm_page_t m,
593     vm_pindex_t pindex, const char *wmesg, bool nonshared);
594 void vm_page_free(vm_page_t m);
595 void vm_page_free_zero(vm_page_t m);
596 
597 void vm_page_activate (vm_page_t);
598 void vm_page_advise(vm_page_t m, int advice);
599 vm_page_t vm_page_alloc(vm_object_t, vm_pindex_t, int);
600 vm_page_t vm_page_alloc_domain(vm_object_t, vm_pindex_t, int, int);
601 vm_page_t vm_page_alloc_after(vm_object_t, vm_pindex_t, int, vm_page_t);
602 vm_page_t vm_page_alloc_domain_after(vm_object_t, vm_pindex_t, int, int,
603     vm_page_t);
604 vm_page_t vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
605     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
606     vm_paddr_t boundary, vm_memattr_t memattr);
607 vm_page_t vm_page_alloc_contig_domain(vm_object_t object,
608     vm_pindex_t pindex, int domain, int req, u_long npages, vm_paddr_t low,
609     vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
610     vm_memattr_t memattr);
611 vm_page_t vm_page_alloc_freelist(int, int);
612 vm_page_t vm_page_alloc_freelist_domain(int, int, int);
613 void vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set);
614 bool vm_page_blacklist_add(vm_paddr_t pa, bool verbose);
615 vm_page_t vm_page_grab(vm_object_t, vm_pindex_t, int);
616 vm_page_t vm_page_grab_unlocked(vm_object_t, vm_pindex_t, int);
617 int vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
618     vm_page_t *ma, int count);
619 int vm_page_grab_pages_unlocked(vm_object_t object, vm_pindex_t pindex,
620     int allocflags, vm_page_t *ma, int count);
621 int vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex,
622     int allocflags);
623 int vm_page_grab_valid_unlocked(vm_page_t *mp, vm_object_t object,
624     vm_pindex_t pindex, int allocflags);
625 void vm_page_deactivate(vm_page_t);
626 void vm_page_deactivate_noreuse(vm_page_t);
627 void vm_page_dequeue(vm_page_t m);
628 void vm_page_dequeue_deferred(vm_page_t m);
629 vm_page_t vm_page_find_least(vm_object_t, vm_pindex_t);
630 void vm_page_free_invalid(vm_page_t);
631 vm_page_t vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr);
632 void vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr);
633 void vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags);
634 int vm_page_insert (vm_page_t, vm_object_t, vm_pindex_t);
635 void vm_page_invalid(vm_page_t m);
636 void vm_page_launder(vm_page_t m);
637 vm_page_t vm_page_lookup(vm_object_t, vm_pindex_t);
638 vm_page_t vm_page_next(vm_page_t m);
639 void vm_page_pqbatch_drain(void);
640 void vm_page_pqbatch_submit(vm_page_t m, uint8_t queue);
641 bool vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old,
642     vm_page_astate_t new);
643 vm_page_t vm_page_prev(vm_page_t m);
644 bool vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m);
645 void vm_page_putfake(vm_page_t m);
646 void vm_page_readahead_finish(vm_page_t m);
647 bool vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low,
648     vm_paddr_t high, u_long alignment, vm_paddr_t boundary);
649 bool vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
650     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary);
651 void vm_page_reference(vm_page_t m);
652 #define	VPR_TRYFREE	0x01
653 #define	VPR_NOREUSE	0x02
654 void vm_page_release(vm_page_t m, int flags);
655 void vm_page_release_locked(vm_page_t m, int flags);
656 vm_page_t vm_page_relookup(vm_object_t, vm_pindex_t);
657 bool vm_page_remove(vm_page_t);
658 bool vm_page_remove_xbusy(vm_page_t);
659 int vm_page_rename(vm_page_t, vm_object_t, vm_pindex_t);
660 void vm_page_replace(vm_page_t mnew, vm_object_t object,
661     vm_pindex_t pindex, vm_page_t mold);
662 int vm_page_sbusied(vm_page_t m);
663 vm_page_t vm_page_scan_contig(u_long npages, vm_page_t m_start,
664     vm_page_t m_end, u_long alignment, vm_paddr_t boundary, int options);
665 vm_page_bits_t vm_page_set_dirty(vm_page_t m);
666 void vm_page_set_valid_range(vm_page_t m, int base, int size);
667 int vm_page_sleep_if_busy(vm_page_t m, const char *msg);
668 int vm_page_sleep_if_xbusy(vm_page_t m, const char *msg);
669 vm_offset_t vm_page_startup(vm_offset_t vaddr);
670 void vm_page_sunbusy(vm_page_t m);
671 bool vm_page_try_remove_all(vm_page_t m);
672 bool vm_page_try_remove_write(vm_page_t m);
673 int vm_page_trysbusy(vm_page_t m);
674 int vm_page_tryxbusy(vm_page_t m);
675 void vm_page_unhold_pages(vm_page_t *ma, int count);
676 void vm_page_unswappable(vm_page_t m);
677 void vm_page_unwire(vm_page_t m, uint8_t queue);
678 bool vm_page_unwire_noq(vm_page_t m);
679 void vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr);
680 void vm_page_wire(vm_page_t);
681 bool vm_page_wire_mapped(vm_page_t m);
682 void vm_page_xunbusy_hard(vm_page_t m);
683 void vm_page_xunbusy_hard_unchecked(vm_page_t m);
684 void vm_page_set_validclean (vm_page_t, int, int);
685 void vm_page_clear_dirty(vm_page_t, int, int);
686 void vm_page_set_invalid(vm_page_t, int, int);
687 void vm_page_valid(vm_page_t m);
688 int vm_page_is_valid(vm_page_t, int, int);
689 void vm_page_test_dirty(vm_page_t);
690 vm_page_bits_t vm_page_bits(int base, int size);
691 void vm_page_zero_invalid(vm_page_t m, boolean_t setvalid);
692 void vm_page_free_pages_toq(struct spglist *free, bool update_wire_count);
693 
694 void vm_page_dirty_KBI(vm_page_t m);
695 void vm_page_lock_KBI(vm_page_t m, const char *file, int line);
696 void vm_page_unlock_KBI(vm_page_t m, const char *file, int line);
697 int vm_page_trylock_KBI(vm_page_t m, const char *file, int line);
698 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
699 void vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line);
700 void vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line);
701 #endif
702 
703 #define	vm_page_busy_fetch(m)	atomic_load_int(&(m)->busy_lock)
704 
705 #define	vm_page_assert_busied(m)					\
706 	KASSERT(vm_page_busied(m),					\
707 	    ("vm_page_assert_busied: page %p not busy @ %s:%d", \
708 	    (m), __FILE__, __LINE__))
709 
710 #define	vm_page_assert_sbusied(m)					\
711 	KASSERT(vm_page_sbusied(m),					\
712 	    ("vm_page_assert_sbusied: page %p not shared busy @ %s:%d", \
713 	    (m), __FILE__, __LINE__))
714 
715 #define	vm_page_assert_unbusied(m)					\
716 	KASSERT((vm_page_busy_fetch(m) & ~VPB_BIT_WAITERS) !=		\
717 	    VPB_CURTHREAD_EXCLUSIVE,					\
718 	    ("vm_page_assert_xbusied: page %p busy_lock %#x owned"	\
719             " by me @ %s:%d",						\
720 	    (m), (m)->busy_lock, __FILE__, __LINE__));			\
721 
722 #define	vm_page_assert_xbusied_unchecked(m) do {			\
723 	KASSERT(vm_page_xbusied(m),					\
724 	    ("vm_page_assert_xbusied: page %p not exclusive busy @ %s:%d", \
725 	    (m), __FILE__, __LINE__));					\
726 } while (0)
727 #define	vm_page_assert_xbusied(m) do {					\
728 	vm_page_assert_xbusied_unchecked(m);				\
729 	KASSERT((vm_page_busy_fetch(m) & ~VPB_BIT_WAITERS) ==		\
730 	    VPB_CURTHREAD_EXCLUSIVE,					\
731 	    ("vm_page_assert_xbusied: page %p busy_lock %#x not owned"	\
732             " by me @ %s:%d",						\
733 	    (m), (m)->busy_lock, __FILE__, __LINE__));			\
734 } while (0)
735 
736 #define	vm_page_busied(m)						\
737 	(vm_page_busy_fetch(m) != VPB_UNBUSIED)
738 
739 #define	vm_page_sbusy(m) do {						\
740 	if (!vm_page_trysbusy(m))					\
741 		panic("%s: page %p failed shared busying", __func__,	\
742 		    (m));						\
743 } while (0)
744 
745 #define	vm_page_xbusied(m)						\
746 	((vm_page_busy_fetch(m) & VPB_SINGLE_EXCLUSIVE) != 0)
747 
748 #define	vm_page_busy_freed(m)						\
749 	(vm_page_busy_fetch(m) == VPB_FREED)
750 
751 #define	vm_page_xbusy(m) do {						\
752 	if (!vm_page_tryxbusy(m))					\
753 		panic("%s: page %p failed exclusive busying", __func__,	\
754 		    (m));						\
755 } while (0)
756 
757 /* Note: page m's lock must not be owned by the caller. */
758 #define	vm_page_xunbusy(m) do {						\
759 	if (!atomic_cmpset_rel_int(&(m)->busy_lock,			\
760 	    VPB_CURTHREAD_EXCLUSIVE, VPB_UNBUSIED))			\
761 		vm_page_xunbusy_hard(m);				\
762 } while (0)
763 #define	vm_page_xunbusy_unchecked(m) do {				\
764 	if (!atomic_cmpset_rel_int(&(m)->busy_lock,			\
765 	    VPB_CURTHREAD_EXCLUSIVE, VPB_UNBUSIED))			\
766 		vm_page_xunbusy_hard_unchecked(m);			\
767 } while (0)
768 
769 #ifdef INVARIANTS
770 void vm_page_object_busy_assert(vm_page_t m);
771 #define	VM_PAGE_OBJECT_BUSY_ASSERT(m)	vm_page_object_busy_assert(m)
772 void vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits);
773 #define	VM_PAGE_ASSERT_PGA_WRITEABLE(m, bits)				\
774 	vm_page_assert_pga_writeable(m, bits)
775 /*
776  * Claim ownership of a page's xbusy state.  In non-INVARIANTS kernels this
777  * operation is a no-op since ownership is not tracked.  In particular
778  * this macro does not provide any synchronization with the previous owner.
779  */
780 #define	vm_page_xbusy_claim(m) do {					\
781 	u_int _busy_lock;						\
782 									\
783 	vm_page_assert_xbusied_unchecked((m));				\
784 	do {								\
785 		_busy_lock = vm_page_busy_fetch(m);			\
786 	} while (!atomic_cmpset_int(&(m)->busy_lock, _busy_lock,	\
787 	    (_busy_lock & VPB_BIT_FLAGMASK) | VPB_CURTHREAD_EXCLUSIVE)); \
788 } while (0)
789 #else
790 #define	VM_PAGE_OBJECT_BUSY_ASSERT(m)	(void)0
791 #define	VM_PAGE_ASSERT_PGA_WRITEABLE(m, bits)	(void)0
792 #define	vm_page_xbusy_claim(m)
793 #endif
794 
795 #if BYTE_ORDER == BIG_ENDIAN
796 #define	VM_PAGE_AFLAG_SHIFT	16
797 #else
798 #define	VM_PAGE_AFLAG_SHIFT	0
799 #endif
800 
801 /*
802  *	Load a snapshot of a page's 32-bit atomic state.
803  */
804 static inline vm_page_astate_t
805 vm_page_astate_load(vm_page_t m)
806 {
807 	vm_page_astate_t a;
808 
809 	a._bits = atomic_load_32(&m->a._bits);
810 	return (a);
811 }
812 
813 /*
814  *	Atomically compare and set a page's atomic state.
815  */
816 static inline bool
817 vm_page_astate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
818 {
819 
820 	KASSERT(new.queue == PQ_INACTIVE || (new.flags & PGA_REQUEUE_HEAD) == 0,
821 	    ("%s: invalid head requeue request for page %p", __func__, m));
822 	KASSERT((new.flags & PGA_ENQUEUED) == 0 || new.queue != PQ_NONE,
823 	    ("%s: setting PGA_ENQUEUED with PQ_NONE in page %p", __func__, m));
824 	KASSERT(new._bits != old->_bits,
825 	    ("%s: bits are unchanged", __func__));
826 
827 	return (atomic_fcmpset_32(&m->a._bits, &old->_bits, new._bits) != 0);
828 }
829 
830 /*
831  *	Clear the given bits in the specified page.
832  */
833 static inline void
834 vm_page_aflag_clear(vm_page_t m, uint16_t bits)
835 {
836 	uint32_t *addr, val;
837 
838 	/*
839 	 * Access the whole 32-bit word containing the aflags field with an
840 	 * atomic update.  Parallel non-atomic updates to the other fields
841 	 * within this word are handled properly by the atomic update.
842 	 */
843 	addr = (void *)&m->a;
844 	val = bits << VM_PAGE_AFLAG_SHIFT;
845 	atomic_clear_32(addr, val);
846 }
847 
848 /*
849  *	Set the given bits in the specified page.
850  */
851 static inline void
852 vm_page_aflag_set(vm_page_t m, uint16_t bits)
853 {
854 	uint32_t *addr, val;
855 
856 	VM_PAGE_ASSERT_PGA_WRITEABLE(m, bits);
857 
858 	/*
859 	 * Access the whole 32-bit word containing the aflags field with an
860 	 * atomic update.  Parallel non-atomic updates to the other fields
861 	 * within this word are handled properly by the atomic update.
862 	 */
863 	addr = (void *)&m->a;
864 	val = bits << VM_PAGE_AFLAG_SHIFT;
865 	atomic_set_32(addr, val);
866 }
867 
868 /*
869  *	vm_page_dirty:
870  *
871  *	Set all bits in the page's dirty field.
872  *
873  *	The object containing the specified page must be locked if the
874  *	call is made from the machine-independent layer.
875  *
876  *	See vm_page_clear_dirty_mask().
877  */
878 static __inline void
879 vm_page_dirty(vm_page_t m)
880 {
881 
882 	/* Use vm_page_dirty_KBI() under INVARIANTS to save memory. */
883 #if (defined(KLD_MODULE) && !defined(KLD_TIED)) || defined(INVARIANTS)
884 	vm_page_dirty_KBI(m);
885 #else
886 	m->dirty = VM_PAGE_BITS_ALL;
887 #endif
888 }
889 
890 /*
891  *	vm_page_undirty:
892  *
893  *	Set page to not be dirty.  Note: does not clear pmap modify bits
894  */
895 static __inline void
896 vm_page_undirty(vm_page_t m)
897 {
898 
899 	VM_PAGE_OBJECT_BUSY_ASSERT(m);
900 	m->dirty = 0;
901 }
902 
903 static inline uint8_t
904 _vm_page_queue(vm_page_astate_t as)
905 {
906 
907 	if ((as.flags & PGA_DEQUEUE) != 0)
908 		return (PQ_NONE);
909 	return (as.queue);
910 }
911 
912 /*
913  *	vm_page_queue:
914  *
915  *	Return the index of the queue containing m.
916  */
917 static inline uint8_t
918 vm_page_queue(vm_page_t m)
919 {
920 
921 	return (_vm_page_queue(vm_page_astate_load(m)));
922 }
923 
924 static inline bool
925 vm_page_active(vm_page_t m)
926 {
927 
928 	return (vm_page_queue(m) == PQ_ACTIVE);
929 }
930 
931 static inline bool
932 vm_page_inactive(vm_page_t m)
933 {
934 
935 	return (vm_page_queue(m) == PQ_INACTIVE);
936 }
937 
938 static inline bool
939 vm_page_in_laundry(vm_page_t m)
940 {
941 	uint8_t queue;
942 
943 	queue = vm_page_queue(m);
944 	return (queue == PQ_LAUNDRY || queue == PQ_UNSWAPPABLE);
945 }
946 
947 /*
948  *	vm_page_drop:
949  *
950  *	Release a reference to a page and return the old reference count.
951  */
952 static inline u_int
953 vm_page_drop(vm_page_t m, u_int val)
954 {
955 	u_int old;
956 
957 	/*
958 	 * Synchronize with vm_page_free_prep(): ensure that all updates to the
959 	 * page structure are visible before it is freed.
960 	 */
961 	atomic_thread_fence_rel();
962 	old = atomic_fetchadd_int(&m->ref_count, -val);
963 	KASSERT(old != VPRC_BLOCKED,
964 	    ("vm_page_drop: page %p has an invalid refcount value", m));
965 	return (old);
966 }
967 
968 /*
969  *	vm_page_wired:
970  *
971  *	Perform a racy check to determine whether a reference prevents the page
972  *	from being reclaimable.  If the page's object is locked, and the page is
973  *	unmapped and exclusively busied by the current thread, no new wirings
974  *	may be created.
975  */
976 static inline bool
977 vm_page_wired(vm_page_t m)
978 {
979 
980 	return (VPRC_WIRE_COUNT(m->ref_count) > 0);
981 }
982 
983 static inline bool
984 vm_page_all_valid(vm_page_t m)
985 {
986 
987 	return (m->valid == VM_PAGE_BITS_ALL);
988 }
989 
990 static inline bool
991 vm_page_none_valid(vm_page_t m)
992 {
993 
994 	return (m->valid == 0);
995 }
996 
997 #endif				/* _KERNEL */
998 #endif				/* !_VM_PAGE_ */
999