xref: /illumos-gate/usr/src/uts/common/vm/anon.h (revision 8d0c3d29bb99f6521f2dc5058a7e4debebad7899)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 1986, 2010, Oracle and/or its affiliates. All rights reserved.
23  */
24 
25 /*	Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T */
26 /*	 All Rights Reserved   */
27 
28 /*
29  * University Copyright- Copyright (c) 1982, 1986, 1988
30  * The Regents of the University of California
31  * All Rights Reserved
32  *
33  * University Acknowledgment- Portions of this document are derived from
34  * software developed by the University of California, Berkeley, and its
35  * contributors.
36  */
37 
38 #ifndef	_VM_ANON_H
39 #define	_VM_ANON_H
40 
41 #include <sys/cred.h>
42 #include <sys/zone.h>
43 #include <vm/seg.h>
44 #include <vm/vpage.h>
45 
46 #ifdef	__cplusplus
47 extern "C" {
48 #endif
49 
50 /*
51  * VM - Anonymous pages.
52  */
53 
54 typedef	unsigned long anoff_t;		/* anon offsets */
55 
56 /*
57  *	Each anonymous page, either in memory or in swap, has an anon structure.
58  * The structure (slot) provides a level of indirection between anonymous pages
59  * and their backing store.
60  *
61  *	(an_vp, an_off) names the vnode of the anonymous page for this slot.
62  *
63  * 	(an_pvp, an_poff) names the location of the physical backing store
64  * 	for the page this slot represents. If the name is null there is no
65  * 	associated physical store. The physical backing store location can
66  *	change while the slot is in use.
67  *
68  *	an_hash is a hash list of anon slots. The list is hashed by
69  * 	(an_vp, an_off) of the associated anonymous page and provides a
70  *	method of going from the name of an anonymous page to its
71  * 	associated anon slot.
72  *
73  *	an_refcnt holds a reference count which is the number of separate
74  * 	copies that will need to be created in case of copy-on-write.
75  *	A refcnt > 0 protects the existence of the slot. The refcnt is
76  * 	initialized to 1 when the anon slot is created in anon_alloc().
77  *	If a client obtains an anon slot and allows multiple threads to
78  * 	share it, then it is the client's responsibility to insure that
79  *	it does not allow one thread to try to reference the slot at the
80  *	same time as another is trying to decrement the last count and
81  *	destroy the anon slot. E.g., the seg_vn segment type protects
82  *	against this with higher level locks.
83  */
84 
85 struct anon {
86 	struct vnode *an_vp;	/* vnode of anon page */
87 	struct vnode *an_pvp;	/* vnode of physical backing store */
88 	anoff_t an_off;		/* offset of anon page */
89 	anoff_t an_poff;	/* offset in vnode */
90 	struct anon *an_hash;	/* hash table of anon slots */
91 	int an_refcnt;		/* # of people sharing slot */
92 };
93 
94 #define	AN_CACHE_ALIGN_LOG2	4	/* log2(AN_CACHE_ALIGN) */
95 #define	AN_CACHE_ALIGN	(1U << AN_CACHE_ALIGN_LOG2) /* anon address aligned */
96 						/* 16 bytes */
97 
98 
99 #ifdef _KERNEL
100 /*
101  * The swapinfo_lock protects:
102  *		swapinfo list
103  *		individual swapinfo structures
104  *
105  * The anoninfo_lock protects:
106  *		anoninfo counters
107  *
108  * The anonhash_lock protects:
109  *		anon hash lists
110  *		anon slot fields
111  *
112  * Fields in the anon slot which are read-only for the life of the slot
113  * (an_vp, an_off) do not require the anonhash_lock be held to access them.
114  * If you access a field without the anonhash_lock held you must be holding
115  * the slot with an_refcnt to make sure it isn't destroyed.
116  * To write (an_pvp, an_poff) in a given slot you must also hold the
117  * p_iolock of the anonymous page for slot.
118  */
119 extern kmutex_t anoninfo_lock;
120 extern kmutex_t swapinfo_lock;
121 extern pad_mutex_t *anonhash_lock;
122 extern pad_mutex_t anon_array_lock[];
123 extern kcondvar_t anon_array_cv[];
124 
125 /*
126  * Global hash table to provide a function from (vp, off) -> ap
127  */
128 extern size_t anon_hash_size;
129 extern unsigned int anon_hash_shift;
130 extern struct anon **anon_hash;
131 #define	ANON_HASH_SIZE	anon_hash_size
132 #define	ANON_HASHAVELEN	4
133 /*
134  * Try to use as many bits of randomness from both vp and off as we can.
135  * This should help spreading evenly for a variety of workloads.  See comments
136  * for PAGE_HASH_FUNC for more explanation.
137  */
138 #define	ANON_HASH(vp, off)	\
139 	(((((uintptr_t)(off) >> PAGESHIFT) ^ \
140 		((uintptr_t)(off) >> (PAGESHIFT + anon_hash_shift))) ^ \
141 		(((uintptr_t)(vp) >> 3) ^ \
142 		((uintptr_t)(vp) >> (3 + anon_hash_shift)) ^ \
143 		((uintptr_t)(vp) >> (3 + 2 * anon_hash_shift)) ^ \
144 		((uintptr_t)(vp) << \
145 		    (anon_hash_shift - AN_VPSHIFT - VNODE_ALIGN_LOG2)))) & \
146 		(anon_hash_size - 1))
147 
148 #define	AH_LOCK_SIZE	(2 << NCPU_LOG2)
149 
150 #define	AH_MUTEX(vp, off)				\
151 	(&anonhash_lock[(ANON_HASH((vp), (off)) &	\
152 	    (AH_LOCK_SIZE - 1))].pad_mutex)
153 
154 #endif	/* _KERNEL */
155 
156 /*
157  * Declaration for the Global counters to accurately
158  * track the kernel foot print in memory.
159  */
160 extern  pgcnt_t pages_locked;
161 extern  pgcnt_t pages_claimed;
162 extern  pgcnt_t pages_useclaim;
163 extern  pgcnt_t obp_pages;
164 
165 /*
166  * Anonymous backing store accounting structure for swapctl.
167  *
168  * ani_max = maximum amount of swap space
169  *	(including potentially available physical memory)
170  * ani_free = amount of unallocated anonymous memory
171  *	(some of which might be reserved and including
172  *	potentially available physical memory)
173  * ani_resv = amount of claimed (reserved) anonymous memory
174  *
175  * The swap data can be aquired more efficiently through the
176  * kstats interface.
177  * Total slots currently available for reservation =
178  *	MAX(ani_max - ani_resv, 0) + (availrmem - swapfs_minfree)
179  */
180 struct anoninfo {
181 	pgcnt_t	ani_max;
182 	pgcnt_t	ani_free;
183 	pgcnt_t	ani_resv;
184 };
185 
186 #ifdef _SYSCALL32
187 struct anoninfo32 {
188 	size32_t ani_max;
189 	size32_t ani_free;
190 	size32_t ani_resv;
191 };
192 #endif /* _SYSCALL32 */
193 
194 /*
195  * Define the NCPU pool of the ani_free counters. Update the counter
196  * of the cpu on which the thread is running and in every clock intr
197  * sync anoninfo.ani_free with the current total off all the NCPU entries.
198  */
199 
200 typedef	struct	ani_free {
201 	kmutex_t	ani_lock;
202 	pgcnt_t		ani_count;
203 	uchar_t		pad[64 - sizeof (kmutex_t) - sizeof (pgcnt_t)];
204 			/* XXX 64 = cacheline size */
205 } ani_free_t;
206 
207 #define	ANI_MAX_POOL	128
208 extern	ani_free_t	ani_free_pool[];
209 
210 #define	ANI_ADD(inc)	{ \
211 	ani_free_t	*anifp; \
212 	int		index; \
213 	index = (CPU->cpu_id & (ANI_MAX_POOL - 1)); \
214 	anifp = &ani_free_pool[index]; \
215 	mutex_enter(&anifp->ani_lock); \
216 	anifp->ani_count += inc; \
217 	mutex_exit(&anifp->ani_lock); \
218 }
219 
220 /*
221  * Anon array pointers are allocated in chunks. Each chunk
222  * has PAGESIZE/sizeof(u_long *) of anon pointers.
223  * There are two levels of arrays for anon array pointers larger
224  * than a chunk. The first level points to anon array chunks.
225  * The second level consists of chunks of anon pointers.
226  *
227  * If anon array is smaller than a chunk then the whole anon array
228  * is created (memory is allocated for whole anon array).
229  * If anon array is larger than a chunk only first level array is
230  * allocated. Then other arrays (chunks) are allocated only when
231  * they are initialized with anon pointers.
232  */
233 struct anon_hdr {
234 	kmutex_t serial_lock;	/* serialize array chunk allocation */
235 	pgcnt_t	size;		/* number of pointers to (anon) pages */
236 	void	**array_chunk;	/* pointers to anon pointers or chunks of */
237 				/* anon pointers */
238 	int	flags;		/* ANON_ALLOC_FORCE force preallocation of */
239 				/* whole anon array	*/
240 };
241 
242 #ifdef	_LP64
243 #define	ANON_PTRSHIFT	3
244 #define	ANON_PTRMASK	~7
245 #else
246 #define	ANON_PTRSHIFT	2
247 #define	ANON_PTRMASK	~3
248 #endif
249 
250 #define	ANON_CHUNK_SIZE		(PAGESIZE >> ANON_PTRSHIFT)
251 #define	ANON_CHUNK_SHIFT	(PAGESHIFT - ANON_PTRSHIFT)
252 #define	ANON_CHUNK_OFF		(ANON_CHUNK_SIZE - 1)
253 
254 /*
255  * Anon flags.
256  */
257 #define	ANON_SLEEP		0x0	/* ok to block */
258 #define	ANON_NOSLEEP		0x1	/* non-blocking call */
259 #define	ANON_ALLOC_FORCE	0x2	/* force single level anon array */
260 #define	ANON_GROWDOWN		0x4	/* anon array should grow downward */
261 
262 struct kshmid;
263 
264 /*
265  * The anon_map structure is used by various clients of the anon layer to
266  * manage anonymous memory.   When anonymous memory is shared,
267  * then the different clients sharing it will point to the
268  * same anon_map structure.  Also, if a segment is unmapped
269  * in the middle where an anon_map structure exists, the
270  * newly created segment will also share the anon_map structure,
271  * although the two segments will use different ranges of the
272  * anon array.  When mappings are private (or shared with
273  * a reference count of 1), an unmap operation will free up
274  * a range of anon slots in the array given by the anon_map
275  * structure.  Because of fragmentation due to this unmapping,
276  * we have to store the size of the anon array in the anon_map
277  * structure so that we can free everything when the referernce
278  * count goes to zero.
279  *
280  * A new rangelock scheme is introduced to make the anon layer scale.
281  * A reader/writer lock per anon_amp and an array of system-wide hash
282  * locks, anon_array_lock[] are introduced to replace serial_lock and
283  * anonmap lock.  The writer lock is held when we want to singlethreaD
284  * the reference to the anon array pointers or when references to
285  * anon_map's members, whereas reader lock and anon_array_lock are
286  * held to allows multiple threads to reference different part of
287  * anon array.  A global set of condition variables, anon_array_cv,
288  * are used with anon_array_lock[] to make the hold time of the locks
289  * short.
290  *
291  * szc is used to calculate the index of hash locks and cv's.  We
292  * could've just used seg->s_szc if not for the possible sharing of
293  * anon_amp between SYSV shared memory and ISM, so now we introduce
294  * szc in the anon_map structure.  For MAP_SHARED, the amp->szc is either
295  * 0 (base page size) or page_num_pagesizes() - 1, while MAP_PRIVATE
296  * the amp->szc could be anything in [0, page_num_pagesizes() - 1].
297  */
298 typedef struct anon_map {
299 	krwlock_t a_rwlock;	/* protect anon_map and anon array */
300 	size_t	size;		/* size in bytes mapped by the anon array */
301 	struct	anon_hdr *ahp; 	/* anon array header pointer, containing */
302 				/* anon pointer array(s) */
303 	size_t	swresv;		/* swap space reserved for this anon_map */
304 	ulong_t	refcnt;		/* reference count on this structure */
305 	ushort_t a_szc;		/* max szc among shared processes */
306 	void	*locality;	/* lgroup locality info */
307 	struct kshmid *a_sp;	/* kshmid if amp backs sysV, or NULL */
308 	int	a_purgewait;	/* somebody waits for slocks to go away */
309 	kcondvar_t a_purgecv;	/* cv for waiting for slocks to go away */
310 	kmutex_t a_purgemtx;	/* mutex for anonmap_purge() */
311 	spgcnt_t a_softlockcnt; /* number of pages locked in pcache */
312 	kmutex_t a_pmtx;	/* protects amp's pcache list */
313 	pcache_link_t a_phead;	/* head of amp's pcache list */
314 } amp_t;
315 
316 #ifdef _KERNEL
317 
318 #define	ANON_BUSY		0x1
319 #define	ANON_ISBUSY(slot)	(*(slot) & ANON_BUSY)
320 #define	ANON_SETBUSY(slot)	(*(slot) |= ANON_BUSY)
321 #define	ANON_CLRBUSY(slot)	(*(slot) &= ~ANON_BUSY)
322 
323 #define	ANON_MAP_SHIFT		6	/* log2(sizeof (struct anon_map)) */
324 #define	ANON_ARRAY_SHIFT	7	/* log2(ANON_LOCKSIZE) */
325 #define	ANON_LOCKSIZE		128
326 
327 #define	ANON_LOCK_ENTER(lock, type)	rw_enter((lock), (type))
328 #define	ANON_LOCK_EXIT(lock)		rw_exit((lock))
329 #define	ANON_LOCK_HELD(lock)		RW_LOCK_HELD((lock))
330 #define	ANON_READ_HELD(lock)		RW_READ_HELD((lock))
331 #define	ANON_WRITE_HELD(lock)		RW_WRITE_HELD((lock))
332 
333 #define	ANON_ARRAY_HASH(amp, idx)\
334 	((((idx) + ((idx) >> ANON_ARRAY_SHIFT) +\
335 	((idx) >> (ANON_ARRAY_SHIFT << 1)) +\
336 	((idx) >> (ANON_ARRAY_SHIFT + (ANON_ARRAY_SHIFT << 1)))) ^\
337 	((uintptr_t)(amp) >> ANON_MAP_SHIFT)) & (ANON_LOCKSIZE - 1))
338 
339 typedef struct anon_sync_obj {
340 	kmutex_t	*sync_mutex;
341 	kcondvar_t	*sync_cv;
342 	ulong_t		*sync_data;
343 } anon_sync_obj_t;
344 
345 /*
346  * Anonymous backing store accounting structure for kernel.
347  * ani_max = total reservable slots on physical (disk-backed) swap
348  * ani_phys_resv = total phys slots reserved for use by clients
349  * ani_mem_resv = total mem slots reserved for use by clients
350  * ani_free = # unallocated physical slots + # of reserved unallocated
351  * memory slots
352  */
353 
354 /*
355  * Initial total swap slots available for reservation
356  */
357 #define	TOTAL_AVAILABLE_SWAP \
358 	(k_anoninfo.ani_max + MAX((spgcnt_t)(availrmem - swapfs_minfree), 0))
359 
360 /*
361  * Swap slots currently available for reservation
362  */
363 #define	CURRENT_TOTAL_AVAILABLE_SWAP				\
364 	((k_anoninfo.ani_max - k_anoninfo.ani_phys_resv) +	\
365 	    MAX((spgcnt_t)(availrmem - swapfs_minfree), 0))
366 
367 struct k_anoninfo {
368 	pgcnt_t	ani_max;	/* total reservable slots on phys */
369 					/* (disk) swap */
370 	pgcnt_t	ani_free;	/* # of unallocated phys and mem slots */
371 	pgcnt_t	ani_phys_resv;	/* # of reserved phys (disk) slots */
372 	pgcnt_t	ani_mem_resv;	/* # of reserved mem slots */
373 	pgcnt_t	ani_locked_swap; /* # of swap slots locked in reserved */
374 				/* mem swap */
375 };
376 
377 extern	struct k_anoninfo k_anoninfo;
378 
379 extern void	anon_init(void);
380 extern struct	anon *anon_alloc(struct vnode *, anoff_t);
381 extern void	anon_dup(struct anon_hdr *, ulong_t,
382 		    struct anon_hdr *, ulong_t, size_t);
383 extern void	anon_dup_fill_holes(struct anon_hdr *, ulong_t,
384 		    struct anon_hdr *, ulong_t, size_t, uint_t, int);
385 extern int	anon_fill_cow_holes(struct seg *, caddr_t, struct anon_hdr *,
386 		    ulong_t, struct vnode *, u_offset_t, size_t, uint_t,
387 		    uint_t, struct vpage [], struct cred *);
388 extern void	anon_free(struct anon_hdr *, ulong_t, size_t);
389 extern void	anon_free_pages(struct anon_hdr *, ulong_t, size_t, uint_t);
390 extern void	anon_disclaim(struct anon_map *, ulong_t, size_t);
391 extern int	anon_getpage(struct anon **, uint_t *, struct page **,
392 		    size_t, struct seg *, caddr_t, enum seg_rw, struct cred *);
393 extern int	swap_getconpage(struct vnode *, u_offset_t, size_t,
394 		    uint_t *, page_t *[], size_t, page_t *, uint_t *,
395 		    spgcnt_t *, struct seg *, caddr_t,
396 		    enum seg_rw, struct cred *);
397 extern int	anon_map_getpages(struct anon_map *, ulong_t,
398 		    uint_t, struct seg *, caddr_t, uint_t,
399 		    uint_t *, page_t *[], uint_t *,
400 		    struct vpage [], enum seg_rw, int, int, int, struct cred *);
401 extern int	anon_map_privatepages(struct anon_map *, ulong_t,
402 		    uint_t, struct seg *, caddr_t, uint_t,
403 		    page_t *[], struct vpage [], int, int, struct cred *);
404 extern struct	page *anon_private(struct anon **, struct seg *,
405 		    caddr_t, uint_t, struct page *,
406 		    int, struct cred *);
407 extern struct	page *anon_zero(struct seg *, caddr_t,
408 		    struct anon **, struct cred *);
409 extern int	anon_map_createpages(struct anon_map *, ulong_t,
410 		    size_t, struct page **,
411 		    struct seg *, caddr_t,
412 		    enum seg_rw, struct cred *);
413 extern int	anon_map_demotepages(struct anon_map *, ulong_t,
414 		    struct seg *, caddr_t, uint_t,
415 		    struct vpage [], struct cred *);
416 extern void	anon_shmap_free_pages(struct anon_map *, ulong_t, size_t);
417 extern int	anon_resvmem(size_t, boolean_t, zone_t *, int);
418 extern void	anon_unresvmem(size_t, zone_t *);
419 extern struct	anon_map *anonmap_alloc(size_t, size_t, int);
420 extern void	anonmap_free(struct anon_map *);
421 extern void	anonmap_purge(struct anon_map *);
422 extern void	anon_swap_free(struct anon *, struct page *);
423 extern void	anon_decref(struct anon *);
424 extern int	non_anon(struct anon_hdr *, ulong_t, u_offset_t *, size_t *);
425 extern pgcnt_t	anon_pages(struct anon_hdr *, ulong_t, pgcnt_t);
426 extern int	anon_swap_adjust(pgcnt_t);
427 extern void	anon_swap_restore(pgcnt_t);
428 extern struct	anon_hdr *anon_create(pgcnt_t, int);
429 extern void	anon_release(struct anon_hdr *, pgcnt_t);
430 extern struct	anon *anon_get_ptr(struct anon_hdr *, ulong_t);
431 extern ulong_t	*anon_get_slot(struct anon_hdr *, ulong_t);
432 extern struct	anon *anon_get_next_ptr(struct anon_hdr *, ulong_t *);
433 extern int	anon_set_ptr(struct anon_hdr *, ulong_t, struct anon *, int);
434 extern int 	anon_copy_ptr(struct anon_hdr *, ulong_t,
435 		    struct anon_hdr *, ulong_t, pgcnt_t, int);
436 extern pgcnt_t	anon_grow(struct anon_hdr *, ulong_t *, pgcnt_t, pgcnt_t, int);
437 extern void	anon_array_enter(struct anon_map *, ulong_t,
438 			anon_sync_obj_t *);
439 extern int	anon_array_try_enter(struct anon_map *, ulong_t,
440 			anon_sync_obj_t *);
441 extern void	anon_array_exit(anon_sync_obj_t *);
442 
443 /*
444  * anon_resv checks to see if there is enough swap space to fulfill a
445  * request and if so, reserves the appropriate anonymous memory resources.
446  * anon_checkspace just checks to see if there is space to fulfill the request,
447  * without taking any resources.  Both return 1 if successful and 0 if not.
448  *
449  * Macros are provided as anon reservation is usually charged to the zone of
450  * the current process.  In some cases (such as anon reserved by tmpfs), a
451  * zone pointer is needed to charge the appropriate zone.
452  */
453 #define	anon_unresv(size)		anon_unresvmem(size, curproc->p_zone)
454 #define	anon_unresv_zone(size, zone)	anon_unresvmem(size, zone)
455 #define	anon_resv(size)			\
456 	anon_resvmem((size), 1, curproc->p_zone, 1)
457 #define	anon_resv_zone(size, zone)	anon_resvmem((size), 1, zone, 1)
458 #define	anon_checkspace(size, zone)	anon_resvmem((size), 0, zone, 0)
459 #define	anon_try_resv_zone(size, zone)	anon_resvmem((size), 1, zone, 0)
460 
461 /*
462  * Flags to anon_private
463  */
464 #define	STEAL_PAGE	0x1	/* page can be stolen */
465 #define	LOCK_PAGE	0x2	/* page must be ``logically'' locked */
466 
467 /*
468  * SEGKP ANON pages that are locked are assumed to be LWP stack pages
469  * and thus count towards the user pages locked count.
470  * This value is protected by the same lock as availrmem.
471  */
472 extern pgcnt_t anon_segkp_pages_locked;
473 
474 extern int anon_debug;
475 
476 #ifdef ANON_DEBUG
477 
478 #define	A_ANON	0x01
479 #define	A_RESV	0x02
480 #define	A_MRESV	0x04
481 
482 /* vararg-like debugging macro. */
483 #define	ANON_PRINT(f, printf_args) \
484 		if (anon_debug & f) \
485 			printf printf_args
486 
487 #else	/* ANON_DEBUG */
488 
489 #define	ANON_PRINT(f, printf_args)
490 
491 #endif	/* ANON_DEBUG */
492 
493 #endif	/* _KERNEL */
494 
495 #ifdef	__cplusplus
496 }
497 #endif
498 
499 #endif	/* _VM_ANON_H */
500