xref: /titanic_41/usr/src/uts/common/sys/kmem_impl.h (revision 9b214d32697277d03ed2e5d98c4a7bfef16dcf4d)
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 2009 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 #include <sys/kmem.h>
31 #include <sys/vmem.h>
32 #include <sys/thread.h>
33 #include <sys/t_lock.h>
34 #include <sys/time.h>
35 #include <sys/kstat.h>
36 #include <sys/cpuvar.h>
37 #include <sys/systm.h>
38 #include <vm/page.h>
39 #include <sys/avl.h>
40 #include <sys/list.h>
41 
42 #ifdef	__cplusplus
43 extern "C" {
44 #endif
45 
46 /*
47  * kernel memory allocator: implementation-private data structures
48  *
49  * Lock order:
50  * 1. cache_lock
51  * 2. cc_lock in order by CPU ID
52  * 3. cache_depot_lock
53  *
54  * Do not call kmem_cache_alloc() or taskq_dispatch() while holding any of the
55  * above locks.
56  */
57 
58 #define	KMF_AUDIT	0x00000001	/* transaction auditing */
59 #define	KMF_DEADBEEF	0x00000002	/* deadbeef checking */
60 #define	KMF_REDZONE	0x00000004	/* redzone checking */
61 #define	KMF_CONTENTS	0x00000008	/* freed-buffer content logging */
62 #define	KMF_STICKY	0x00000010	/* if set, override /etc/system */
63 #define	KMF_NOMAGAZINE	0x00000020	/* disable per-cpu magazines */
64 #define	KMF_FIREWALL	0x00000040	/* put all bufs before unmapped pages */
65 #define	KMF_LITE	0x00000100	/* lightweight debugging */
66 
67 #define	KMF_HASH	0x00000200	/* cache has hash table */
68 #define	KMF_RANDOMIZE	0x00000400	/* randomize other kmem_flags */
69 
70 #define	KMF_BUFTAG	(KMF_DEADBEEF | KMF_REDZONE)
71 #define	KMF_TOUCH	(KMF_BUFTAG | KMF_LITE | KMF_CONTENTS)
72 #define	KMF_RANDOM	(KMF_TOUCH | KMF_AUDIT | KMF_NOMAGAZINE)
73 #define	KMF_DEBUG	(KMF_RANDOM | KMF_FIREWALL)
74 
75 #define	KMEM_STACK_DEPTH	15
76 
77 #define	KMEM_FREE_PATTERN		0xdeadbeefdeadbeefULL
78 #define	KMEM_UNINITIALIZED_PATTERN	0xbaddcafebaddcafeULL
79 #define	KMEM_REDZONE_PATTERN		0xfeedfacefeedfaceULL
80 #define	KMEM_REDZONE_BYTE		0xbb
81 
82 /*
83  * Redzone size encodings for kmem_alloc() / kmem_free().  We encode the
84  * allocation size, rather than storing it directly, so that kmem_free()
85  * can distinguish frees of the wrong size from redzone violations.
86  *
87  * A size of zero is never valid.
88  */
89 #define	KMEM_SIZE_ENCODE(x)	(251 * (x) + 1)
90 #define	KMEM_SIZE_DECODE(x)	((x) / 251)
91 #define	KMEM_SIZE_VALID(x)	((x) % 251 == 1 && (x) != 1)
92 
93 
94 #define	KMEM_ALIGN		8	/* min guaranteed alignment */
95 #define	KMEM_ALIGN_SHIFT	3	/* log2(KMEM_ALIGN) */
96 #define	KMEM_VOID_FRACTION	8	/* never waste more than 1/8 of slab */
97 
98 #define	KMEM_SLAB_IS_PARTIAL(sp)		\
99 	((sp)->slab_refcnt > 0 && (sp)->slab_refcnt < (sp)->slab_chunks)
100 #define	KMEM_SLAB_IS_ALL_USED(sp)		\
101 	((sp)->slab_refcnt == (sp)->slab_chunks)
102 
103 /*
104  * The bufctl (buffer control) structure keeps some minimal information
105  * about each buffer: its address, its slab, and its current linkage,
106  * which is either on the slab's freelist (if the buffer is free), or
107  * on the cache's buf-to-bufctl hash table (if the buffer is allocated).
108  * In the case of non-hashed, or "raw", caches (the common case), only
109  * the freelist linkage is necessary: the buffer address is at a fixed
110  * offset from the bufctl address, and the slab is at the end of the page.
111  *
112  * NOTE: bc_next must be the first field; raw buffers have linkage only.
113  */
114 typedef struct kmem_bufctl {
115 	struct kmem_bufctl	*bc_next;	/* next bufctl struct */
116 	void			*bc_addr;	/* address of buffer */
117 	struct kmem_slab	*bc_slab;	/* controlling slab */
118 } kmem_bufctl_t;
119 
120 /*
121  * The KMF_AUDIT version of the bufctl structure.  The beginning of this
122  * structure must be identical to the normal bufctl structure so that
123  * pointers are interchangeable.
124  */
125 typedef struct kmem_bufctl_audit {
126 	struct kmem_bufctl	*bc_next;	/* next bufctl struct */
127 	void			*bc_addr;	/* address of buffer */
128 	struct kmem_slab	*bc_slab;	/* controlling slab */
129 	kmem_cache_t		*bc_cache;	/* controlling cache */
130 	hrtime_t		bc_timestamp;	/* transaction time */
131 	kthread_t		*bc_thread;	/* thread doing transaction */
132 	struct kmem_bufctl	*bc_lastlog;	/* last log entry */
133 	void			*bc_contents;	/* contents at last free */
134 	int			bc_depth;	/* stack depth */
135 	pc_t			bc_stack[KMEM_STACK_DEPTH];	/* pc stack */
136 } kmem_bufctl_audit_t;
137 
138 /*
139  * A kmem_buftag structure is appended to each buffer whenever any of the
140  * KMF_BUFTAG flags (KMF_DEADBEEF, KMF_REDZONE, KMF_VERIFY) are set.
141  */
142 typedef struct kmem_buftag {
143 	uint64_t		bt_redzone;	/* 64-bit redzone pattern */
144 	kmem_bufctl_t		*bt_bufctl;	/* bufctl */
145 	intptr_t		bt_bxstat;	/* bufctl ^ (alloc/free) */
146 } kmem_buftag_t;
147 
148 /*
149  * A variant of the kmem_buftag structure used for KMF_LITE caches.
150  * Previous callers are stored in reverse chronological order. (i.e. most
151  * recent first)
152  */
153 typedef struct kmem_buftag_lite {
154 	kmem_buftag_t		bt_buftag;	/* a normal buftag */
155 	pc_t			bt_history[1];	/* zero or more callers */
156 } kmem_buftag_lite_t;
157 
158 #define	KMEM_BUFTAG_LITE_SIZE(f)	\
159 	(offsetof(kmem_buftag_lite_t, bt_history[f]))
160 
161 #define	KMEM_BUFTAG(cp, buf)		\
162 	((kmem_buftag_t *)((char *)(buf) + (cp)->cache_buftag))
163 
164 #define	KMEM_BUFCTL(cp, buf)		\
165 	((kmem_bufctl_t *)((char *)(buf) + (cp)->cache_bufctl))
166 
167 #define	KMEM_BUF(cp, bcp)		\
168 	((void *)((char *)(bcp) - (cp)->cache_bufctl))
169 
170 #define	KMEM_SLAB(cp, buf)		\
171 	((kmem_slab_t *)P2END((uintptr_t)(buf), (cp)->cache_slabsize) - 1)
172 
173 /*
174  * The "CPU" macro loads a cpu_t that refers to the cpu that the current
175  * thread is running on at the time the macro is executed.  A context switch
176  * may occur immediately after loading this data structure, leaving this
177  * thread pointing at the cpu_t for the previous cpu.  This is not a problem;
178  * we'd just end up checking the previous cpu's per-cpu cache, and then check
179  * the other layers of the kmem cache if need be.
180  *
181  * It's not even a problem if the old cpu gets DR'ed out during the context
182  * switch.  The cpu-remove DR operation bzero()s the cpu_t, but doesn't free
183  * it.  So the cpu_t's cpu_cache_offset would read as 0, causing us to use
184  * cpu 0's per-cpu cache.
185  *
186  * So, there is no need to disable kernel preemption while using the CPU macro
187  * below since if we have been context switched, there will not be any
188  * correctness problem, just a momentary use of a different per-cpu cache.
189  */
190 
191 #define	KMEM_CPU_CACHE(cp)						\
192 	(kmem_cpu_cache_t *)((char *)(&cp->cache_cpu) + CPU->cpu_cache_offset)
193 
194 #define	KMEM_MAGAZINE_VALID(cp, mp)	\
195 	(((kmem_slab_t *)P2END((uintptr_t)(mp), PAGESIZE) - 1)->slab_cache == \
196 	    (cp)->cache_magtype->mt_cache)
197 
198 #define	KMEM_SLAB_OFFSET(sp, buf)	\
199 	((size_t)((uintptr_t)(buf) - (uintptr_t)((sp)->slab_base)))
200 
201 #define	KMEM_SLAB_MEMBER(sp, buf)	\
202 	(KMEM_SLAB_OFFSET(sp, buf) < (sp)->slab_cache->cache_slabsize)
203 
204 #define	KMEM_BUFTAG_ALLOC	0xa110c8edUL
205 #define	KMEM_BUFTAG_FREE	0xf4eef4eeUL
206 
207 /* slab_later_count thresholds */
208 #define	KMEM_DISBELIEF		3
209 
210 /* slab_flags */
211 #define	KMEM_SLAB_NOMOVE	0x1
212 #define	KMEM_SLAB_MOVE_PENDING	0x2
213 
214 typedef struct kmem_slab {
215 	struct kmem_cache	*slab_cache;	/* controlling cache */
216 	void			*slab_base;	/* base of allocated memory */
217 	avl_node_t		slab_link;	/* slab linkage */
218 	struct kmem_bufctl	*slab_head;	/* first free buffer */
219 	long			slab_refcnt;	/* outstanding allocations */
220 	long			slab_chunks;	/* chunks (bufs) in this slab */
221 	uint32_t		slab_stuck_offset; /* unmoved buffer offset */
222 	uint16_t		slab_later_count; /* cf KMEM_CBRC_LATER */
223 	uint16_t		slab_flags;	/* bits to mark the slab */
224 } kmem_slab_t;
225 
226 #define	KMEM_HASH_INITIAL	64
227 
228 #define	KMEM_HASH(cp, buf)	\
229 	((cp)->cache_hash_table +	\
230 	(((uintptr_t)(buf) >> (cp)->cache_hash_shift) & (cp)->cache_hash_mask))
231 
232 typedef struct kmem_magazine {
233 	void	*mag_next;
234 	void	*mag_round[1];		/* one or more rounds */
235 } kmem_magazine_t;
236 
237 /*
238  * The magazine types for fast per-cpu allocation
239  */
240 typedef struct kmem_magtype {
241 	int		mt_magsize;	/* magazine size (number of rounds) */
242 	int		mt_align;	/* magazine alignment */
243 	size_t		mt_minbuf;	/* all smaller buffers qualify */
244 	size_t		mt_maxbuf;	/* no larger buffers qualify */
245 	kmem_cache_t	*mt_cache;	/* magazine cache */
246 } kmem_magtype_t;
247 
248 #define	KMEM_CPU_CACHE_SIZE	64	/* must be power of 2 */
249 #define	KMEM_CPU_PAD		(KMEM_CPU_CACHE_SIZE - sizeof (kmutex_t) - \
250 	2 * sizeof (uint64_t) - 2 * sizeof (void *) - 4 * sizeof (int))
251 #define	KMEM_CACHE_SIZE(ncpus)	\
252 	((size_t)(&((kmem_cache_t *)0)->cache_cpu[ncpus]))
253 
254 /* Offset from kmem_cache->cache_cpu for per cpu caches */
255 #define	KMEM_CPU_CACHE_OFFSET(cpuid)					\
256 	((size_t)(&((kmem_cache_t *)0)->cache_cpu[cpuid]) -		\
257 	(size_t)(&((kmem_cache_t *)0)->cache_cpu))
258 
259 typedef struct kmem_cpu_cache {
260 	kmutex_t	cc_lock;	/* protects this cpu's local cache */
261 	uint64_t	cc_alloc;	/* allocations from this cpu */
262 	uint64_t	cc_free;	/* frees to this cpu */
263 	kmem_magazine_t	*cc_loaded;	/* the currently loaded magazine */
264 	kmem_magazine_t	*cc_ploaded;	/* the previously loaded magazine */
265 	int		cc_rounds;	/* number of objects in loaded mag */
266 	int		cc_prounds;	/* number of objects in previous mag */
267 	int		cc_magsize;	/* number of rounds in a full mag */
268 	int		cc_flags;	/* CPU-local copy of cache_flags */
269 	char		cc_pad[KMEM_CPU_PAD]; /* for nice alignment */
270 } kmem_cpu_cache_t;
271 
272 /*
273  * The magazine lists used in the depot.
274  */
275 typedef struct kmem_maglist {
276 	kmem_magazine_t	*ml_list;	/* magazine list */
277 	long		ml_total;	/* number of magazines */
278 	long		ml_min;		/* min since last update */
279 	long		ml_reaplimit;	/* max reapable magazines */
280 	uint64_t	ml_alloc;	/* allocations from this list */
281 } kmem_maglist_t;
282 
283 typedef struct kmem_defrag {
284 	/*
285 	 * Statistics
286 	 */
287 	uint64_t	kmd_callbacks;		/* move callbacks */
288 	uint64_t	kmd_yes;		/* KMEM_CBRC_YES responses */
289 	uint64_t	kmd_no;			/* NO responses */
290 	uint64_t	kmd_later;		/* LATER responses */
291 	uint64_t	kmd_dont_need;		/* DONT_NEED responses */
292 	uint64_t	kmd_dont_know;		/* DONT_KNOW responses */
293 	uint64_t	kmd_hunt_found;		/* DONT_KNOW: # found in mag */
294 
295 	/*
296 	 * Consolidator fields
297 	 */
298 	avl_tree_t	kmd_moves_pending;	/* buffer moves pending */
299 	list_t		kmd_deadlist;		/* deferred slab frees */
300 	size_t		kmd_deadcount;		/* # of slabs in kmd_deadlist */
301 	uint8_t		kmd_reclaim_numer;	/* slab usage threshold */
302 	uint8_t		kmd_pad1;		/* compiler padding */
303 	size_t		kmd_slabs_sought;	/* reclaimable slabs sought */
304 	size_t		kmd_slabs_found;	/* reclaimable slabs found */
305 	size_t		kmd_scans;		/* nth scan interval counter */
306 	/*
307 	 * Fields used to ASSERT that the client does not kmem_cache_free()
308 	 * objects passed to the move callback.
309 	 */
310 	void		*kmd_from_buf;		/* object to move */
311 	void		*kmd_to_buf;		/* move destination */
312 	kthread_t	*kmd_thread;		/* thread calling move */
313 } kmem_defrag_t;
314 
315 #define	KMEM_CACHE_NAMELEN	31
316 
317 struct kmem_cache {
318 	/*
319 	 * Statistics
320 	 */
321 	uint64_t	cache_slab_create;	/* slab creates */
322 	uint64_t	cache_slab_destroy;	/* slab destroys */
323 	uint64_t	cache_slab_alloc;	/* slab layer allocations */
324 	uint64_t	cache_slab_free;	/* slab layer frees */
325 	uint64_t	cache_alloc_fail;	/* total failed allocations */
326 	uint64_t	cache_buftotal;		/* total buffers */
327 	uint64_t	cache_bufmax;		/* max buffers ever */
328 	uint64_t	cache_bufslab;		/* buffers free in slab layer */
329 	uint64_t	cache_rescale;		/* # of hash table rescales */
330 	uint64_t	cache_lookup_depth;	/* hash lookup depth */
331 	uint64_t	cache_depot_contention;	/* mutex contention count */
332 	uint64_t	cache_depot_contention_prev; /* previous snapshot */
333 
334 	/*
335 	 * Cache properties
336 	 */
337 	char		cache_name[KMEM_CACHE_NAMELEN + 1];
338 	size_t		cache_bufsize;		/* object size */
339 	size_t		cache_align;		/* object alignment */
340 	int		(*cache_constructor)(void *, void *, int);
341 	void		(*cache_destructor)(void *, void *);
342 	void		(*cache_reclaim)(void *);
343 	kmem_cbrc_t	(*cache_move)(void *, void *, size_t, void *);
344 	void		*cache_private;		/* opaque arg to callbacks */
345 	vmem_t		*cache_arena;		/* vmem source for slabs */
346 	int		cache_cflags;		/* cache creation flags */
347 	int		cache_flags;		/* various cache state info */
348 	uint32_t	cache_mtbf;		/* induced alloc failure rate */
349 	uint32_t	cache_pad1;		/* compiler padding */
350 	kstat_t		*cache_kstat;		/* exported statistics */
351 	list_node_t	cache_link;		/* cache linkage */
352 
353 	/*
354 	 * Slab layer
355 	 */
356 	kmutex_t	cache_lock;		/* protects slab layer */
357 	size_t		cache_chunksize;	/* buf + alignment [+ debug] */
358 	size_t		cache_slabsize;		/* size of a slab */
359 	size_t		cache_maxchunks;	/* max buffers per slab */
360 	size_t		cache_bufctl;		/* buf-to-bufctl distance */
361 	size_t		cache_buftag;		/* buf-to-buftag distance */
362 	size_t		cache_verify;		/* bytes to verify */
363 	size_t		cache_contents;		/* bytes of saved content */
364 	size_t		cache_color;		/* next slab color */
365 	size_t		cache_mincolor;		/* maximum slab color */
366 	size_t		cache_maxcolor;		/* maximum slab color */
367 	size_t		cache_hash_shift;	/* get to interesting bits */
368 	size_t		cache_hash_mask;	/* hash table mask */
369 	list_t		cache_complete_slabs;	/* completely allocated slabs */
370 	size_t		cache_complete_slab_count;
371 	avl_tree_t	cache_partial_slabs;	/* partial slab freelist */
372 	size_t		cache_partial_binshift;	/* for AVL sort bins */
373 	kmem_cache_t	*cache_bufctl_cache;	/* source of bufctls */
374 	kmem_bufctl_t	**cache_hash_table;	/* hash table base */
375 	kmem_defrag_t	*cache_defrag;		/* slab consolidator fields */
376 
377 	/*
378 	 * Depot layer
379 	 */
380 	kmutex_t	cache_depot_lock;	/* protects depot */
381 	kmem_magtype_t	*cache_magtype;		/* magazine type */
382 	kmem_maglist_t	cache_full;		/* full magazines */
383 	kmem_maglist_t	cache_empty;		/* empty magazines */
384 
385 	/*
386 	 * Per-CPU layer
387 	 */
388 	kmem_cpu_cache_t cache_cpu[1];		/* max_ncpus actual elements */
389 };
390 
391 typedef struct kmem_cpu_log_header {
392 	kmutex_t	clh_lock;
393 	char		*clh_current;
394 	size_t		clh_avail;
395 	int		clh_chunk;
396 	int		clh_hits;
397 	char		clh_pad[64 - sizeof (kmutex_t) - sizeof (char *) -
398 				sizeof (size_t) - 2 * sizeof (int)];
399 } kmem_cpu_log_header_t;
400 
401 typedef struct kmem_log_header {
402 	kmutex_t	lh_lock;
403 	char		*lh_base;
404 	int		*lh_free;
405 	size_t		lh_chunksize;
406 	int		lh_nchunks;
407 	int		lh_head;
408 	int		lh_tail;
409 	int		lh_hits;
410 	kmem_cpu_log_header_t lh_cpu[1];	/* ncpus actually allocated */
411 } kmem_log_header_t;
412 
413 /* kmem_move kmm_flags */
414 #define	KMM_DESPERATE		0x1
415 #define	KMM_NOTIFY		0x2
416 
417 typedef struct kmem_move {
418 	kmem_slab_t	*kmm_from_slab;
419 	void		*kmm_from_buf;
420 	void		*kmm_to_buf;
421 	avl_node_t	kmm_entry;
422 	int		kmm_flags;
423 } kmem_move_t;
424 
425 /*
426  * In order to consolidate partial slabs, it must be possible for the cache to
427  * have partial slabs.
428  */
429 #define	KMEM_IS_MOVABLE(cp)						\
430 	(((cp)->cache_chunksize * 2) <= (cp)->cache_slabsize)
431 
432 #ifdef	__cplusplus
433 }
434 #endif
435 
436 #endif	/* _SYS_KMEM_IMPL_H */
437