xref: /titanic_41/usr/src/uts/common/vm/anon.h (revision 75d01c9ab5ef6f1bbac9f9d4eb379d5c38583d82)
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, Version 1.0 only
6  * (the "License").  You may not use this file except in compliance
7  * with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or http://www.opensolaris.org/os/licensing.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 /*
23  * Copyright 2005 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 /*	Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T */
28 /*	 All Rights Reserved   */
29 
30 /*
31  * University Copyright- Copyright (c) 1982, 1986, 1988
32  * The Regents of the University of California
33  * All Rights Reserved
34  *
35  * University Acknowledgment- Portions of this document are derived from
36  * software developed by the University of California, Berkeley, and its
37  * contributors.
38  */
39 
40 #ifndef	_VM_ANON_H
41 #define	_VM_ANON_H
42 
43 #pragma ident	"%Z%%M%	%I%	%E% SMI"
44 
45 #include <sys/cred.h>
46 #include <vm/seg.h>
47 #include <vm/vpage.h>
48 
49 #ifdef	__cplusplus
50 extern "C" {
51 #endif
52 
53 /*
54  * VM - Anonymous pages.
55  */
56 
57 typedef	unsigned long anoff_t;		/* anon offsets */
58 
59 /*
60  *	Each anonymous page, either in memory or in swap, has an anon structure.
61  * The structure (slot) provides a level of indirection between anonymous pages
62  * and their backing store.
63  *
64  *	(an_vp, an_off) names the vnode of the anonymous page for this slot.
65  *
66  * 	(an_pvp, an_poff) names the location of the physical backing store
67  * 	for the page this slot represents. If the name is null there is no
68  * 	associated physical store. The physical backing store location can
69  *	change while the slot is in use.
70  *
71  *	an_hash is a hash list of anon slots. The list is hashed by
72  * 	(an_vp, an_off) of the associated anonymous page and provides a
73  *	method of going from the name of an anonymous page to its
74  * 	associated anon slot.
75  *
76  *	an_refcnt holds a reference count which is the number of separate
77  * 	copies that will need to be created in case of copy-on-write.
78  *	A refcnt > 0 protects the existence of the slot. The refcnt is
79  * 	initialized to 1 when the anon slot is created in anon_alloc().
80  *	If a client obtains an anon slot and allows multiple threads to
81  * 	share it, then it is the client's responsibility to insure that
82  *	it does not allow one thread to try to reference the slot at the
83  *	same time as another is trying to decrement the last count and
84  *	destroy the anon slot. E.g., the seg_vn segment type protects
85  *	against this with higher level locks.
86  */
87 
88 struct anon {
89 	struct vnode *an_vp;	/* vnode of anon page */
90 	struct vnode *an_pvp;	/* vnode of physical backing store */
91 	anoff_t an_off;		/* offset of anon page */
92 	anoff_t an_poff;	/* offset in vnode */
93 	struct anon *an_hash;	/* hash table of anon slots */
94 	int an_refcnt;		/* # of people sharing slot */
95 };
96 
97 #ifdef _KERNEL
98 /*
99  * The swapinfo_lock protects:
100  *		swapinfo list
101  *		individual swapinfo structures
102  *
103  * The anoninfo_lock protects:
104  *		anoninfo counters
105  *
106  * The anonhash_lock protects:
107  *		anon hash lists
108  *		anon slot fields
109  *
110  * Fields in the anon slot which are read-only for the life of the slot
111  * (an_vp, an_off) do not require the anonhash_lock be held to access them.
112  * If you access a field without the anonhash_lock held you must be holding
113  * the slot with an_refcnt to make sure it isn't destroyed.
114  * To write (an_pvp, an_poff) in a given slot you must also hold the
115  * p_iolock of the anonymous page for slot.
116  */
117 extern kmutex_t anoninfo_lock;
118 extern kmutex_t swapinfo_lock;
119 extern kmutex_t anonhash_lock[];
120 extern pad_mutex_t anon_array_lock[];
121 extern kcondvar_t anon_array_cv[];
122 
123 /*
124  * Global hash table to provide a function from (vp, off) -> ap
125  */
126 extern size_t anon_hash_size;
127 extern struct anon **anon_hash;
128 #define	ANON_HASH_SIZE	anon_hash_size
129 #define	ANON_HASHAVELEN	4
130 #define	ANON_HASH(VP, OFF)	\
131 ((((uintptr_t)(VP) >> 7)  ^ ((OFF) >> PAGESHIFT)) & (ANON_HASH_SIZE - 1))
132 
133 #define	AH_LOCK_SIZE	64
134 #define	AH_LOCK(vp, off) (ANON_HASH((vp), (off)) & (AH_LOCK_SIZE -1))
135 
136 #endif	/* _KERNEL */
137 
138 /*
139  * Declaration for the Global counters to accurately
140  * track the kernel foot print in memory.
141  */
142 extern  pgcnt_t segvn_pages_locked;
143 extern  pgcnt_t pages_locked;
144 extern  pgcnt_t pages_claimed;
145 extern  pgcnt_t pages_useclaim;
146 extern  pgcnt_t obp_pages;
147 
148 /*
149  * Anonymous backing store accounting structure for swapctl.
150  *
151  * ani_max = maximum amount of swap space
152  *	(including potentially available physical memory)
153  * ani_free = amount of unallocated anonymous memory
154  *	(some of which might be reserved and including
155  *	potentially available physical memory)
156  * ani_resv = amount of claimed (reserved) anonymous memory
157  *
158  * The swap data can be aquired more efficiently through the
159  * kstats interface.
160  * Total slots currently available for reservation =
161  *	MAX(ani_max - ani_resv, 0) + (availrmem - swapfs_minfree)
162  */
163 struct anoninfo {
164 	pgcnt_t	ani_max;
165 	pgcnt_t	ani_free;
166 	pgcnt_t	ani_resv;
167 };
168 
169 #ifdef _SYSCALL32
170 struct anoninfo32 {
171 	size32_t ani_max;
172 	size32_t ani_free;
173 	size32_t ani_resv;
174 };
175 #endif /* _SYSCALL32 */
176 
177 /*
178  * Define the NCPU pool of the ani_free counters. Update the counter
179  * of the cpu on which the thread is running and in every clock intr
180  * sync anoninfo.ani_free with the current total off all the NCPU entries.
181  */
182 
183 typedef	struct	ani_free {
184 	kmutex_t	ani_lock;
185 	pgcnt_t		ani_count;
186 	uchar_t		pad[64 - sizeof (kmutex_t) - sizeof (pgcnt_t)];
187 			/* XXX 64 = cacheline size */
188 } ani_free_t;
189 
190 #define	ANI_MAX_POOL	128
191 extern	ani_free_t	ani_free_pool[];
192 
193 #define	ANI_ADD(inc)	{ \
194 	ani_free_t	*anifp; \
195 	int		index; \
196 	index = (CPU->cpu_id & (ANI_MAX_POOL - 1)); \
197 	anifp = &ani_free_pool[index]; \
198 	mutex_enter(&anifp->ani_lock); \
199 	anifp->ani_count += inc; \
200 	mutex_exit(&anifp->ani_lock); \
201 }
202 
203 /*
204  * Anon array pointers are allocated in chunks. Each chunk
205  * has PAGESIZE/sizeof(u_long *) of anon pointers.
206  * There are two levels of arrays for anon array pointers larger
207  * than a chunk. The first level points to anon array chunks.
208  * The second level consists of chunks of anon pointers.
209  *
210  * If anon array is smaller than a chunk then the whole anon array
211  * is created (memory is allocated for whole anon array).
212  * If anon array is larger than a chunk only first level array is
213  * allocated. Then other arrays (chunks) are allocated only when
214  * they are initialized with anon pointers.
215  */
216 struct anon_hdr {
217 	kmutex_t serial_lock;	/* serialize array chunk allocation */
218 	pgcnt_t	size;		/* number of pointers to (anon) pages */
219 	void	**array_chunk;	/* pointers to anon pointers or chunks of */
220 				/* anon pointers */
221 	int	flags;		/* ANON_ALLOC_FORCE force preallocation of */
222 				/* whole anon array	*/
223 };
224 
225 #ifdef	_LP64
226 #define	ANON_PTRSHIFT	3
227 #define	ANON_PTRMASK	~7
228 #else
229 #define	ANON_PTRSHIFT	2
230 #define	ANON_PTRMASK	~3
231 #endif
232 
233 #define	ANON_CHUNK_SIZE		(PAGESIZE >> ANON_PTRSHIFT)
234 #define	ANON_CHUNK_SHIFT	(PAGESHIFT - ANON_PTRSHIFT)
235 #define	ANON_CHUNK_OFF		(ANON_CHUNK_SIZE - 1)
236 
237 /*
238  * Anon flags.
239  */
240 #define	ANON_SLEEP		0x0	/* ok to block */
241 #define	ANON_NOSLEEP		0x1	/* non-blocking call */
242 #define	ANON_ALLOC_FORCE	0x2	/* force single level anon array */
243 #define	ANON_GROWDOWN		0x4	/* anon array should grow downward */
244 
245 /*
246  * The anon_map structure is used by various clients of the anon layer to
247  * manage anonymous memory.   When anonymous memory is shared,
248  * then the different clients sharing it will point to the
249  * same anon_map structure.  Also, if a segment is unmapped
250  * in the middle where an anon_map structure exists, the
251  * newly created segment will also share the anon_map structure,
252  * although the two segments will use different ranges of the
253  * anon array.  When mappings are private (or shared with
254  * a reference count of 1), an unmap operation will free up
255  * a range of anon slots in the array given by the anon_map
256  * structure.  Because of fragmentation due to this unmapping,
257  * we have to store the size of the anon array in the anon_map
258  * structure so that we can free everything when the referernce
259  * count goes to zero.
260  *
261  * A new rangelock scheme is introduced to make the anon layer scale.
262  * A reader/writer lock per anon_amp and an array of system-wide hash
263  * locks, anon_array_lock[] are introduced to replace serial_lock and
264  * anonmap lock.  The writer lock is held when we want to singlethreaD
265  * the reference to the anon array pointers or when references to
266  * anon_map's members, whereas reader lock and anon_array_lock are
267  * held to allows multiple threads to reference different part of
268  * anon array.  A global set of condition variables, anon_array_cv,
269  * are used with anon_array_lock[] to make the hold time of the locks
270  * short.
271  *
272  * szc is used to calculate the index of hash locks and cv's.  We
273  * could've just used seg->s_szc if not for the possible sharing of
274  * anon_amp between SYSV shared memory and ISM, so now we introduce
275  * szc in the anon_map structure.  For MAP_SHARED, the amp->szc is either
276  * 0 (base page size) or page_num_pagesizes() - 1, while MAP_PRIVATE
277  * the amp->szc could be anything in [0, page_num_pagesizes() - 1].
278  */
279 struct anon_map {
280 	krwlock_t a_rwlock;	/* protect anon_map and anon array */
281 	size_t	size;		/* size in bytes mapped by the anon array */
282 	struct	anon_hdr *ahp; 	/* anon array header pointer, containing */
283 				/* anon pointer array(s) */
284 	size_t	swresv;		/* swap space reserved for this anon_map */
285 	uint_t	refcnt;		/* reference count on this structure */
286 	ushort_t a_szc;		/* max szc among shared processes */
287 	void	*locality;	/* lgroup locality info */
288 };
289 
290 #ifdef _KERNEL
291 
292 #define	ANON_BUSY		0x1
293 #define	ANON_ISBUSY(slot)	(*(slot) & ANON_BUSY)
294 #define	ANON_SETBUSY(slot)	(*(slot) |= ANON_BUSY)
295 #define	ANON_CLRBUSY(slot)	(*(slot) &= ~ANON_BUSY)
296 
297 #define	ANON_MAP_SHIFT		6	/* log2(sizeof (struct anon_map)) */
298 #define	ANON_ARRAY_SHIFT	7	/* log2(ANON_LOCKSIZE) */
299 #define	ANON_LOCKSIZE		128
300 
301 #define	ANON_LOCK_ENTER(lock, type)	rw_enter((lock), (type))
302 #define	ANON_LOCK_EXIT(lock)		rw_exit((lock))
303 
304 #define	ANON_ARRAY_HASH(amp, idx)\
305 	((((idx) + ((idx) >> ANON_ARRAY_SHIFT) +\
306 	((idx) >> (ANON_ARRAY_SHIFT << 1)) +\
307 	((idx) >> (ANON_ARRAY_SHIFT + (ANON_ARRAY_SHIFT << 1)))) ^\
308 	((uintptr_t)(amp) >> ANON_MAP_SHIFT)) & (ANON_LOCKSIZE - 1))
309 
310 typedef struct anon_sync_obj {
311 	kmutex_t	*sync_mutex;
312 	kcondvar_t	*sync_cv;
313 	ulong_t		*sync_data;
314 } anon_sync_obj_t;
315 
316 /*
317  * Anonymous backing store accounting structure for kernel.
318  * ani_max = total reservable slots on physical (disk-backed) swap
319  * ani_phys_resv = total phys slots reserved for use by clients
320  * ani_mem_resv = total mem slots reserved for use by clients
321  * ani_free = # unallocated physical slots + # of reserved unallocated
322  * memory slots
323  */
324 
325 /*
326  * Initial total swap slots available for reservation
327  */
328 #define	TOTAL_AVAILABLE_SWAP \
329 	(k_anoninfo.ani_max + MAX((spgcnt_t)(availrmem - swapfs_minfree), 0))
330 
331 /*
332  * Swap slots currently available for reservation
333  */
334 #define	CURRENT_TOTAL_AVAILABLE_SWAP \
335 	((k_anoninfo.ani_max - k_anoninfo.ani_phys_resv) +	\
336 			MAX((spgcnt_t)(availrmem - swapfs_minfree), 0))
337 
338 struct k_anoninfo {
339 	pgcnt_t	ani_max;	/* total reservable slots on phys */
340 					/* (disk) swap */
341 	pgcnt_t	ani_free;	/* # of unallocated phys and mem slots */
342 	pgcnt_t	ani_phys_resv;	/* # of reserved phys (disk) slots */
343 	pgcnt_t	ani_mem_resv;	/* # of reserved mem slots */
344 	pgcnt_t	ani_locked_swap; /* # of swap slots locked in reserved */
345 				/* mem swap */
346 };
347 
348 extern	struct k_anoninfo k_anoninfo;
349 
350 extern void	anon_init(void);
351 extern struct	anon *anon_alloc(struct vnode *, anoff_t);
352 extern void	anon_dup(struct anon_hdr *, ulong_t,
353 		    struct anon_hdr *, ulong_t, size_t);
354 extern void	anon_dup_fill_holes(struct anon_hdr *, ulong_t,
355 		    struct anon_hdr *, ulong_t, size_t, uint_t, int);
356 extern int	anon_fill_cow_holes(struct seg *, caddr_t, struct anon_hdr *,
357 		    ulong_t, struct vnode *, u_offset_t, size_t, uint_t,
358 		    uint_t, struct vpage [], struct cred *);
359 extern void	anon_free(struct anon_hdr *, ulong_t, size_t);
360 extern void	anon_free_pages(struct anon_hdr *, ulong_t, size_t, uint_t);
361 extern void	anon_disclaim(struct anon_map *, ulong_t, size_t, int);
362 extern int	anon_getpage(struct anon **, uint_t *, struct page **,
363 		    size_t, struct seg *, caddr_t, enum seg_rw, struct cred *);
364 extern int	swap_getconpage(struct vnode *, u_offset_t, size_t,
365 		    uint_t *, page_t *[], size_t, page_t *,
366 		    spgcnt_t *, struct seg *, caddr_t,
367 		    enum seg_rw, struct cred *);
368 extern int	anon_map_getpages(struct anon_map *, ulong_t,
369 		    uint_t, struct seg *, caddr_t, uint_t,
370 		    uint_t *, page_t *[], uint_t *,
371 		    struct vpage [], enum seg_rw, int, int, struct cred *);
372 extern int	anon_map_privatepages(struct anon_map *, ulong_t,
373 		    uint_t, struct seg *, caddr_t, uint_t,
374 		    page_t *[], struct vpage [], int, struct cred *);
375 extern struct	page *anon_private(struct anon **, struct seg *,
376 		    caddr_t, uint_t, struct page *,
377 		    int, struct cred *);
378 extern struct	page *anon_zero(struct seg *, caddr_t,
379 		    struct anon **, struct cred *);
380 extern int	anon_map_createpages(struct anon_map *, ulong_t,
381 		    size_t, struct page **,
382 		    struct seg *, caddr_t,
383 		    enum seg_rw, struct cred *);
384 extern int	anon_map_demotepages(struct anon_map *, ulong_t,
385 		    struct seg *, caddr_t, uint_t,
386 		    struct vpage [], struct cred *);
387 extern int	anon_resvmem(size_t, uint_t);
388 extern void	anon_unresv(size_t);
389 extern struct	anon_map *anonmap_alloc(size_t, size_t);
390 extern void	anonmap_free(struct anon_map *);
391 extern void	anon_decref(struct anon *);
392 extern int	non_anon(struct anon_hdr *, ulong_t, u_offset_t *, size_t *);
393 extern pgcnt_t	anon_pages(struct anon_hdr *, ulong_t, pgcnt_t);
394 extern int	anon_swap_adjust(pgcnt_t);
395 extern void	anon_swap_restore(pgcnt_t);
396 extern struct	anon_hdr *anon_create(pgcnt_t, int);
397 extern void	anon_release(struct anon_hdr *, pgcnt_t);
398 extern struct	anon *anon_get_ptr(struct anon_hdr *, ulong_t);
399 extern ulong_t	*anon_get_slot(struct anon_hdr *, ulong_t);
400 extern struct	anon *anon_get_next_ptr(struct anon_hdr *, ulong_t *);
401 extern int	anon_set_ptr(struct anon_hdr *, ulong_t, struct anon *, int);
402 extern int 	anon_copy_ptr(struct anon_hdr *, ulong_t,
403 		    struct anon_hdr *, ulong_t, pgcnt_t, int);
404 extern pgcnt_t	anon_grow(struct anon_hdr *, ulong_t *, pgcnt_t, pgcnt_t, int);
405 extern void	anon_array_enter(struct anon_map *, ulong_t,
406 			anon_sync_obj_t *);
407 extern int	anon_array_try_enter(struct anon_map *, ulong_t,
408 			anon_sync_obj_t *);
409 extern void	anon_array_exit(anon_sync_obj_t *);
410 
411 /*
412  * anon_resv checks to see if there is enough swap space to fulfill a
413  * request and if so, reserves the appropriate anonymous memory resources.
414  * anon_checkspace just checks to see if there is space to fulfill the request,
415  * without taking any resources.  Both return 1 if successful and 0 if not.
416  */
417 #define	anon_resv(size)		anon_resvmem((size), 1)
418 #define	anon_checkspace(size)	anon_resvmem((size), 0)
419 
420 /*
421  * Flags to anon_private
422  */
423 #define	STEAL_PAGE	0x1	/* page can be stolen */
424 #define	LOCK_PAGE	0x2	/* page must be ``logically'' locked */
425 
426 /*
427  * Flags to anon_disclaim
428  */
429 #define	ANON_PGLOOKUP_BLK	0x1	/* block on locked pages */
430 
431 /*
432  * SEGKP ANON pages that are locked are assumed to be LWP stack pages
433  * and thus count towards the user pages locked count.
434  * This value is protected by the same lock as availrmem.
435  */
436 extern pgcnt_t anon_segkp_pages_locked;
437 
438 extern int anon_debug;
439 
440 #ifdef ANON_DEBUG
441 
442 #define	A_ANON	0x01
443 #define	A_RESV	0x02
444 #define	A_MRESV	0x04
445 
446 /* vararg-like debugging macro. */
447 #define	ANON_PRINT(f, printf_args) \
448 		if (anon_debug & f) \
449 			printf printf_args
450 
451 #else	/* ANON_DEBUG */
452 
453 #define	ANON_PRINT(f, printf_args)
454 
455 #endif	/* ANON_DEBUG */
456 
457 #endif	/* _KERNEL */
458 
459 #ifdef	__cplusplus
460 }
461 #endif
462 
463 #endif	/* _VM_ANON_H */
464