xref: /titanic_41/usr/src/uts/common/sys/kmem_impl.h (revision e11c3f44f531fdff80941ce57c065d2ae861cefc)
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 /*
23  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #ifndef _SYS_KMEM_IMPL_H
28 #define	_SYS_KMEM_IMPL_H
29 
30 #pragma ident	"%Z%%M%	%I%	%E% SMI"
31 
32 #include <sys/kmem.h>
33 #include <sys/vmem.h>
34 #include <sys/thread.h>
35 #include <sys/t_lock.h>
36 #include <sys/time.h>
37 #include <sys/kstat.h>
38 #include <sys/cpuvar.h>
39 #include <sys/systm.h>
40 #include <vm/page.h>
41 #include <sys/avl.h>
42 #include <sys/list.h>
43 
44 #ifdef	__cplusplus
45 extern "C" {
46 #endif
47 
48 /*
49  * kernel memory allocator: implementation-private data structures
50  *
51  * Lock order:
52  * 1. cache_lock
53  * 2. cc_lock in order by CPU ID
54  * 3. cache_depot_lock
55  *
56  * Do not call kmem_cache_alloc() or taskq_dispatch() while holding any of the
57  * above locks.
58  */
59 
60 #define	KMF_AUDIT	0x00000001	/* transaction auditing */
61 #define	KMF_DEADBEEF	0x00000002	/* deadbeef checking */
62 #define	KMF_REDZONE	0x00000004	/* redzone checking */
63 #define	KMF_CONTENTS	0x00000008	/* freed-buffer content logging */
64 #define	KMF_STICKY	0x00000010	/* if set, override /etc/system */
65 #define	KMF_NOMAGAZINE	0x00000020	/* disable per-cpu magazines */
66 #define	KMF_FIREWALL	0x00000040	/* put all bufs before unmapped pages */
67 #define	KMF_LITE	0x00000100	/* lightweight debugging */
68 
69 #define	KMF_HASH	0x00000200	/* cache has hash table */
70 #define	KMF_RANDOMIZE	0x00000400	/* randomize other kmem_flags */
71 
72 #define	KMF_BUFTAG	(KMF_DEADBEEF | KMF_REDZONE)
73 #define	KMF_TOUCH	(KMF_BUFTAG | KMF_LITE | KMF_CONTENTS)
74 #define	KMF_RANDOM	(KMF_TOUCH | KMF_AUDIT | KMF_NOMAGAZINE)
75 #define	KMF_DEBUG	(KMF_RANDOM | KMF_FIREWALL)
76 
77 #define	KMEM_STACK_DEPTH	15
78 
79 #define	KMEM_FREE_PATTERN		0xdeadbeefdeadbeefULL
80 #define	KMEM_UNINITIALIZED_PATTERN	0xbaddcafebaddcafeULL
81 #define	KMEM_REDZONE_PATTERN		0xfeedfacefeedfaceULL
82 #define	KMEM_REDZONE_BYTE		0xbb
83 
84 /*
85  * Redzone size encodings for kmem_alloc() / kmem_free().  We encode the
86  * allocation size, rather than storing it directly, so that kmem_free()
87  * can distinguish frees of the wrong size from redzone violations.
88  *
89  * A size of zero is never valid.
90  */
91 #define	KMEM_SIZE_ENCODE(x)	(251 * (x) + 1)
92 #define	KMEM_SIZE_DECODE(x)	((x) / 251)
93 #define	KMEM_SIZE_VALID(x)	((x) % 251 == 1 && (x) != 1)
94 
95 
96 #define	KMEM_ALIGN		8	/* min guaranteed alignment */
97 #define	KMEM_ALIGN_SHIFT	3	/* log2(KMEM_ALIGN) */
98 #define	KMEM_VOID_FRACTION	8	/* never waste more than 1/8 of slab */
99 
100 #define	KMEM_SLAB_IS_PARTIAL(sp)		\
101 	((sp)->slab_refcnt > 0 && (sp)->slab_refcnt < (sp)->slab_chunks)
102 #define	KMEM_SLAB_IS_ALL_USED(sp)		\
103 	((sp)->slab_refcnt == (sp)->slab_chunks)
104 
105 /*
106  * The bufctl (buffer control) structure keeps some minimal information
107  * about each buffer: its address, its slab, and its current linkage,
108  * which is either on the slab's freelist (if the buffer is free), or
109  * on the cache's buf-to-bufctl hash table (if the buffer is allocated).
110  * In the case of non-hashed, or "raw", caches (the common case), only
111  * the freelist linkage is necessary: the buffer address is at a fixed
112  * offset from the bufctl address, and the slab is at the end of the page.
113  *
114  * NOTE: bc_next must be the first field; raw buffers have linkage only.
115  */
116 typedef struct kmem_bufctl {
117 	struct kmem_bufctl	*bc_next;	/* next bufctl struct */
118 	void			*bc_addr;	/* address of buffer */
119 	struct kmem_slab	*bc_slab;	/* controlling slab */
120 } kmem_bufctl_t;
121 
122 /*
123  * The KMF_AUDIT version of the bufctl structure.  The beginning of this
124  * structure must be identical to the normal bufctl structure so that
125  * pointers are interchangeable.
126  */
127 typedef struct kmem_bufctl_audit {
128 	struct kmem_bufctl	*bc_next;	/* next bufctl struct */
129 	void			*bc_addr;	/* address of buffer */
130 	struct kmem_slab	*bc_slab;	/* controlling slab */
131 	kmem_cache_t		*bc_cache;	/* controlling cache */
132 	hrtime_t		bc_timestamp;	/* transaction time */
133 	kthread_t		*bc_thread;	/* thread doing transaction */
134 	struct kmem_bufctl	*bc_lastlog;	/* last log entry */
135 	void			*bc_contents;	/* contents at last free */
136 	int			bc_depth;	/* stack depth */
137 	pc_t			bc_stack[KMEM_STACK_DEPTH];	/* pc stack */
138 } kmem_bufctl_audit_t;
139 
140 /*
141  * A kmem_buftag structure is appended to each buffer whenever any of the
142  * KMF_BUFTAG flags (KMF_DEADBEEF, KMF_REDZONE, KMF_VERIFY) are set.
143  */
144 typedef struct kmem_buftag {
145 	uint64_t		bt_redzone;	/* 64-bit redzone pattern */
146 	kmem_bufctl_t		*bt_bufctl;	/* bufctl */
147 	intptr_t		bt_bxstat;	/* bufctl ^ (alloc/free) */
148 } kmem_buftag_t;
149 
150 /*
151  * A variant of the kmem_buftag structure used for KMF_LITE caches.
152  * Previous callers are stored in reverse chronological order. (i.e. most
153  * recent first)
154  */
155 typedef struct kmem_buftag_lite {
156 	kmem_buftag_t		bt_buftag;	/* a normal buftag */
157 	pc_t			bt_history[1];	/* zero or more callers */
158 } kmem_buftag_lite_t;
159 
160 #define	KMEM_BUFTAG_LITE_SIZE(f)	\
161 	(offsetof(kmem_buftag_lite_t, bt_history[f]))
162 
163 #define	KMEM_BUFTAG(cp, buf)		\
164 	((kmem_buftag_t *)((char *)(buf) + (cp)->cache_buftag))
165 
166 #define	KMEM_BUFCTL(cp, buf)		\
167 	((kmem_bufctl_t *)((char *)(buf) + (cp)->cache_bufctl))
168 
169 #define	KMEM_BUF(cp, bcp)		\
170 	((void *)((char *)(bcp) - (cp)->cache_bufctl))
171 
172 #define	KMEM_SLAB(cp, buf)		\
173 	((kmem_slab_t *)P2END((uintptr_t)(buf), (cp)->cache_slabsize) - 1)
174 
175 #define	KMEM_CPU_CACHE(cp)		\
176 	(kmem_cpu_cache_t *)((char *)cp + CPU->cpu_cache_offset)
177 
178 #define	KMEM_MAGAZINE_VALID(cp, mp)	\
179 	(((kmem_slab_t *)P2END((uintptr_t)(mp), PAGESIZE) - 1)->slab_cache == \
180 	    (cp)->cache_magtype->mt_cache)
181 
182 #define	KMEM_SLAB_OFFSET(sp, buf)	\
183 	((size_t)((uintptr_t)(buf) - (uintptr_t)((sp)->slab_base)))
184 
185 #define	KMEM_SLAB_MEMBER(sp, buf)	\
186 	(KMEM_SLAB_OFFSET(sp, buf) < (sp)->slab_cache->cache_slabsize)
187 
188 #define	KMEM_BUFTAG_ALLOC	0xa110c8edUL
189 #define	KMEM_BUFTAG_FREE	0xf4eef4eeUL
190 
191 /* slab_later_count thresholds */
192 #define	KMEM_DISBELIEF		3
193 
194 /* slab_flags */
195 #define	KMEM_SLAB_NOMOVE	0x1
196 #define	KMEM_SLAB_MOVE_PENDING	0x2
197 
198 typedef struct kmem_slab {
199 	struct kmem_cache	*slab_cache;	/* controlling cache */
200 	void			*slab_base;	/* base of allocated memory */
201 	avl_node_t		slab_link;	/* slab linkage */
202 	struct kmem_bufctl	*slab_head;	/* first free buffer */
203 	long			slab_refcnt;	/* outstanding allocations */
204 	long			slab_chunks;	/* chunks (bufs) in this slab */
205 	uint32_t		slab_stuck_offset; /* unmoved buffer offset */
206 	uint16_t		slab_later_count; /* cf KMEM_CBRC_LATER */
207 	uint16_t		slab_flags;	/* bits to mark the slab */
208 } kmem_slab_t;
209 
210 #define	KMEM_HASH_INITIAL	64
211 
212 #define	KMEM_HASH(cp, buf)	\
213 	((cp)->cache_hash_table +	\
214 	(((uintptr_t)(buf) >> (cp)->cache_hash_shift) & (cp)->cache_hash_mask))
215 
216 typedef struct kmem_magazine {
217 	void	*mag_next;
218 	void	*mag_round[1];		/* one or more rounds */
219 } kmem_magazine_t;
220 
221 /*
222  * The magazine types for fast per-cpu allocation
223  */
224 typedef struct kmem_magtype {
225 	int		mt_magsize;	/* magazine size (number of rounds) */
226 	int		mt_align;	/* magazine alignment */
227 	size_t		mt_minbuf;	/* all smaller buffers qualify */
228 	size_t		mt_maxbuf;	/* no larger buffers qualify */
229 	kmem_cache_t	*mt_cache;	/* magazine cache */
230 } kmem_magtype_t;
231 
232 #define	KMEM_CPU_CACHE_SIZE	64	/* must be power of 2 */
233 #define	KMEM_CPU_PAD		(KMEM_CPU_CACHE_SIZE - sizeof (kmutex_t) - \
234 	2 * sizeof (uint64_t) - 2 * sizeof (void *) - 4 * sizeof (int))
235 #define	KMEM_CACHE_SIZE(ncpus)	\
236 	((size_t)(&((kmem_cache_t *)0)->cache_cpu[ncpus]))
237 
238 typedef struct kmem_cpu_cache {
239 	kmutex_t	cc_lock;	/* protects this cpu's local cache */
240 	uint64_t	cc_alloc;	/* allocations from this cpu */
241 	uint64_t	cc_free;	/* frees to this cpu */
242 	kmem_magazine_t	*cc_loaded;	/* the currently loaded magazine */
243 	kmem_magazine_t	*cc_ploaded;	/* the previously loaded magazine */
244 	int		cc_rounds;	/* number of objects in loaded mag */
245 	int		cc_prounds;	/* number of objects in previous mag */
246 	int		cc_magsize;	/* number of rounds in a full mag */
247 	int		cc_flags;	/* CPU-local copy of cache_flags */
248 	char		cc_pad[KMEM_CPU_PAD]; /* for nice alignment */
249 } kmem_cpu_cache_t;
250 
251 /*
252  * The magazine lists used in the depot.
253  */
254 typedef struct kmem_maglist {
255 	kmem_magazine_t	*ml_list;	/* magazine list */
256 	long		ml_total;	/* number of magazines */
257 	long		ml_min;		/* min since last update */
258 	long		ml_reaplimit;	/* max reapable magazines */
259 	uint64_t	ml_alloc;	/* allocations from this list */
260 } kmem_maglist_t;
261 
262 typedef struct kmem_defrag {
263 	/*
264 	 * Statistics
265 	 */
266 	uint64_t	kmd_callbacks;		/* move callbacks */
267 	uint64_t	kmd_yes;		/* KMEM_CBRC_YES responses */
268 	uint64_t	kmd_no;			/* NO responses */
269 	uint64_t	kmd_later;		/* LATER responses */
270 	uint64_t	kmd_dont_need;		/* DONT_NEED responses */
271 	uint64_t	kmd_dont_know;		/* DONT_KNOW responses */
272 	uint64_t	kmd_hunt_found;		/* DONT_KNOW: # found in mag */
273 
274 	/*
275 	 * Consolidator fields
276 	 */
277 	avl_tree_t	kmd_moves_pending;	/* buffer moves pending */
278 	list_t		kmd_deadlist;		/* deferred slab frees */
279 	size_t		kmd_deadcount;		/* # of slabs in kmd_deadlist */
280 	uint8_t		kmd_reclaim_numer;	/* slab usage threshold */
281 	uint8_t		kmd_pad1;		/* compiler padding */
282 	size_t		kmd_slabs_sought;	/* reclaimable slabs sought */
283 	size_t		kmd_slabs_found;	/* reclaimable slabs found */
284 	size_t		kmd_scans;		/* nth scan interval counter */
285 	/*
286 	 * Fields used to ASSERT that the client does not kmem_cache_free()
287 	 * objects passed to the move callback.
288 	 */
289 	void		*kmd_from_buf;		/* object to move */
290 	void		*kmd_to_buf;		/* move destination */
291 	kthread_t	*kmd_thread;		/* thread calling move */
292 } kmem_defrag_t;
293 
294 #define	KMEM_CACHE_NAMELEN	31
295 
296 struct kmem_cache {
297 	/*
298 	 * Statistics
299 	 */
300 	uint64_t	cache_slab_create;	/* slab creates */
301 	uint64_t	cache_slab_destroy;	/* slab destroys */
302 	uint64_t	cache_slab_alloc;	/* slab layer allocations */
303 	uint64_t	cache_slab_free;	/* slab layer frees */
304 	uint64_t	cache_alloc_fail;	/* total failed allocations */
305 	uint64_t	cache_buftotal;		/* total buffers */
306 	uint64_t	cache_bufmax;		/* max buffers ever */
307 	uint64_t	cache_bufslab;		/* buffers free in slab layer */
308 	uint64_t	cache_rescale;		/* # of hash table rescales */
309 	uint64_t	cache_lookup_depth;	/* hash lookup depth */
310 	uint64_t	cache_depot_contention;	/* mutex contention count */
311 	uint64_t	cache_depot_contention_prev; /* previous snapshot */
312 
313 	/*
314 	 * Cache properties
315 	 */
316 	char		cache_name[KMEM_CACHE_NAMELEN + 1];
317 	size_t		cache_bufsize;		/* object size */
318 	size_t		cache_align;		/* object alignment */
319 	int		(*cache_constructor)(void *, void *, int);
320 	void		(*cache_destructor)(void *, void *);
321 	void		(*cache_reclaim)(void *);
322 	kmem_cbrc_t	(*cache_move)(void *, void *, size_t, void *);
323 	void		*cache_private;		/* opaque arg to callbacks */
324 	vmem_t		*cache_arena;		/* vmem source for slabs */
325 	int		cache_cflags;		/* cache creation flags */
326 	int		cache_flags;		/* various cache state info */
327 	uint32_t	cache_mtbf;		/* induced alloc failure rate */
328 	uint32_t	cache_pad1;		/* compiler padding */
329 	kstat_t		*cache_kstat;		/* exported statistics */
330 	list_node_t	cache_link;		/* cache linkage */
331 
332 	/*
333 	 * Slab layer
334 	 */
335 	kmutex_t	cache_lock;		/* protects slab layer */
336 	size_t		cache_chunksize;	/* buf + alignment [+ debug] */
337 	size_t		cache_slabsize;		/* size of a slab */
338 	size_t		cache_maxchunks;	/* max buffers per slab */
339 	size_t		cache_bufctl;		/* buf-to-bufctl distance */
340 	size_t		cache_buftag;		/* buf-to-buftag distance */
341 	size_t		cache_verify;		/* bytes to verify */
342 	size_t		cache_contents;		/* bytes of saved content */
343 	size_t		cache_color;		/* next slab color */
344 	size_t		cache_mincolor;		/* maximum slab color */
345 	size_t		cache_maxcolor;		/* maximum slab color */
346 	size_t		cache_hash_shift;	/* get to interesting bits */
347 	size_t		cache_hash_mask;	/* hash table mask */
348 	list_t		cache_complete_slabs;	/* completely allocated slabs */
349 	size_t		cache_complete_slab_count;
350 	avl_tree_t	cache_partial_slabs;	/* partial slab freelist */
351 	size_t		cache_partial_binshift;	/* for AVL sort bins */
352 	kmem_cache_t	*cache_bufctl_cache;	/* source of bufctls */
353 	kmem_bufctl_t	**cache_hash_table;	/* hash table base */
354 	kmem_defrag_t	*cache_defrag;		/* slab consolidator fields */
355 
356 	/*
357 	 * Depot layer
358 	 */
359 	kmutex_t	cache_depot_lock;	/* protects depot */
360 	kmem_magtype_t	*cache_magtype;		/* magazine type */
361 	kmem_maglist_t	cache_full;		/* full magazines */
362 	kmem_maglist_t	cache_empty;		/* empty magazines */
363 
364 	/*
365 	 * Per-CPU layer
366 	 */
367 	kmem_cpu_cache_t cache_cpu[1];		/* max_ncpus actual elements */
368 };
369 
370 typedef struct kmem_cpu_log_header {
371 	kmutex_t	clh_lock;
372 	char		*clh_current;
373 	size_t		clh_avail;
374 	int		clh_chunk;
375 	int		clh_hits;
376 	char		clh_pad[64 - sizeof (kmutex_t) - sizeof (char *) -
377 				sizeof (size_t) - 2 * sizeof (int)];
378 } kmem_cpu_log_header_t;
379 
380 typedef struct kmem_log_header {
381 	kmutex_t	lh_lock;
382 	char		*lh_base;
383 	int		*lh_free;
384 	size_t		lh_chunksize;
385 	int		lh_nchunks;
386 	int		lh_head;
387 	int		lh_tail;
388 	int		lh_hits;
389 	kmem_cpu_log_header_t lh_cpu[1];	/* ncpus actually allocated */
390 } kmem_log_header_t;
391 
392 /* kmem_move kmm_flags */
393 #define	KMM_DESPERATE		0x1
394 #define	KMM_NOTIFY		0x2
395 
396 typedef struct kmem_move {
397 	kmem_slab_t	*kmm_from_slab;
398 	void		*kmm_from_buf;
399 	void		*kmm_to_buf;
400 	avl_node_t	kmm_entry;
401 	int		kmm_flags;
402 } kmem_move_t;
403 
404 /*
405  * In order to consolidate partial slabs, it must be possible for the cache to
406  * have partial slabs.
407  */
408 #define	KMEM_IS_MOVABLE(cp)						\
409 	(((cp)->cache_chunksize * 2) <= (cp)->cache_slabsize)
410 
411 #ifdef	__cplusplus
412 }
413 #endif
414 
415 #endif	/* _SYS_KMEM_IMPL_H */
416