Subversion Repositories HelenOS

Rev

Rev 3022 | Rev 4201 | Go to most recent revision | Only display areas with differences | Ignore whitespace | Details | Blame | Last modification | View Log | RSS feed

Rev 3022 Rev 4055
1
/*
1
/*
2
 * Copyright (c) 2006 Ondrej Palkovsky
2
 * Copyright (c) 2006 Ondrej Palkovsky
3
 * All rights reserved.
3
 * All rights reserved.
4
 *
4
 *
5
 * Redistribution and use in source and binary forms, with or without
5
 * Redistribution and use in source and binary forms, with or without
6
 * modification, are permitted provided that the following conditions
6
 * modification, are permitted provided that the following conditions
7
 * are met:
7
 * are met:
8
 *
8
 *
9
 * - Redistributions of source code must retain the above copyright
9
 * - Redistributions of source code must retain the above copyright
10
 *   notice, this list of conditions and the following disclaimer.
10
 *   notice, this list of conditions and the following disclaimer.
11
 * - Redistributions in binary form must reproduce the above copyright
11
 * - Redistributions in binary form must reproduce the above copyright
12
 *   notice, this list of conditions and the following disclaimer in the
12
 *   notice, this list of conditions and the following disclaimer in the
13
 *   documentation and/or other materials provided with the distribution.
13
 *   documentation and/or other materials provided with the distribution.
14
 * - The name of the author may not be used to endorse or promote products
14
 * - The name of the author may not be used to endorse or promote products
15
 *   derived from this software without specific prior written permission.
15
 *   derived from this software without specific prior written permission.
16
 *
16
 *
17
 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
17
 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18
 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
18
 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19
 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
19
 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20
 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
20
 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21
 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
21
 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22
 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
22
 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23
 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
23
 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24
 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
24
 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25
 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
25
 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26
 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
26
 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27
 */
27
 */
28
 
28
 
29
/** @addtogroup genericmm
29
/** @addtogroup genericmm
30
 * @{
30
 * @{
31
 */
31
 */
32
 
32
 
33
/**
33
/**
34
 * @file
34
 * @file
35
 * @brief   Slab allocator.
35
 * @brief   Slab allocator.
36
 *
36
 *
37
 * The slab allocator is closely modelled after OpenSolaris slab allocator.
37
 * The slab allocator is closely modelled after OpenSolaris slab allocator.
38
 * @see http://www.usenix.org/events/usenix01/full_papers/bonwick/bonwick_html/
38
 * @see http://www.usenix.org/events/usenix01/full_papers/bonwick/bonwick_html/
39
 *
39
 *
40
 * with the following exceptions:
40
 * with the following exceptions:
41
 * @li empty slabs are deallocated immediately
41
 * @li empty slabs are deallocated immediately
42
 *     (in Linux they are kept in linked list, in Solaris ???)
42
 *     (in Linux they are kept in linked list, in Solaris ???)
43
 * @li empty magazines are deallocated when not needed
43
 * @li empty magazines are deallocated when not needed
44
 *     (in Solaris they are held in linked list in slab cache)
44
 *     (in Solaris they are held in linked list in slab cache)
45
 *
45
 *
46
 * Following features are not currently supported but would be easy to do:
46
 * Following features are not currently supported but would be easy to do:
47
 * @li cache coloring
47
 * @li cache coloring
48
 * @li dynamic magazine growing (different magazine sizes are already
48
 * @li dynamic magazine growing (different magazine sizes are already
49
 *     supported, but we would need to adjust allocation strategy)
49
 *     supported, but we would need to adjust allocation strategy)
50
 *
50
 *
51
 * The slab allocator supports per-CPU caches ('magazines') to facilitate
51
 * The slab allocator supports per-CPU caches ('magazines') to facilitate
52
 * good SMP scaling.
52
 * good SMP scaling.
53
 *
53
 *
54
 * When a new object is being allocated, it is first checked, if it is
54
 * When a new object is being allocated, it is first checked, if it is
55
 * available in a CPU-bound magazine. If it is not found there, it is
55
 * available in a CPU-bound magazine. If it is not found there, it is
56
 * allocated from a CPU-shared slab - if a partially full one is found,
56
 * allocated from a CPU-shared slab - if a partially full one is found,
57
 * it is used, otherwise a new one is allocated.
57
 * it is used, otherwise a new one is allocated.
58
 *
58
 *
59
 * When an object is being deallocated, it is put to a CPU-bound magazine.
59
 * When an object is being deallocated, it is put to a CPU-bound magazine.
60
 * If there is no such magazine, a new one is allocated (if this fails,
60
 * If there is no such magazine, a new one is allocated (if this fails,
61
 * the object is deallocated into slab). If the magazine is full, it is
61
 * the object is deallocated into slab). If the magazine is full, it is
62
 * put into cpu-shared list of magazines and a new one is allocated.
62
 * put into cpu-shared list of magazines and a new one is allocated.
63
 *
63
 *
64
 * The CPU-bound magazine is actually a pair of magazines in order to avoid
64
 * The CPU-bound magazine is actually a pair of magazines in order to avoid
65
 * thrashing when somebody is allocating/deallocating 1 item at the magazine
65
 * thrashing when somebody is allocating/deallocating 1 item at the magazine
66
 * size boundary. LIFO order is enforced, which should avoid fragmentation
66
 * size boundary. LIFO order is enforced, which should avoid fragmentation
67
 * as much as possible.
67
 * as much as possible.
68
 *  
68
 *  
69
 * Every cache contains list of full slabs and list of partially full slabs.
69
 * Every cache contains list of full slabs and list of partially full slabs.
70
 * Empty slabs are immediately freed (thrashing will be avoided because
70
 * Empty slabs are immediately freed (thrashing will be avoided because
71
 * of magazines).
71
 * of magazines).
72
 *
72
 *
73
 * The slab information structure is kept inside the data area, if possible.
73
 * The slab information structure is kept inside the data area, if possible.
74
 * The cache can be marked that it should not use magazines. This is used
74
 * The cache can be marked that it should not use magazines. This is used
75
 * only for slab related caches to avoid deadlocks and infinite recursion
75
 * only for slab related caches to avoid deadlocks and infinite recursion
76
 * (the slab allocator uses itself for allocating all it's control structures).
76
 * (the slab allocator uses itself for allocating all it's control structures).
77
 *
77
 *
78
 * The slab allocator allocates a lot of space and does not free it. When
78
 * The slab allocator allocates a lot of space and does not free it. When
79
 * the frame allocator fails to allocate a frame, it calls slab_reclaim().
79
 * the frame allocator fails to allocate a frame, it calls slab_reclaim().
80
 * It tries 'light reclaim' first, then brutal reclaim. The light reclaim
80
 * It tries 'light reclaim' first, then brutal reclaim. The light reclaim
81
 * releases slabs from cpu-shared magazine-list, until at least 1 slab
81
 * releases slabs from cpu-shared magazine-list, until at least 1 slab
82
 * is deallocated in each cache (this algorithm should probably change).
82
 * is deallocated in each cache (this algorithm should probably change).
83
 * The brutal reclaim removes all cached objects, even from CPU-bound
83
 * The brutal reclaim removes all cached objects, even from CPU-bound
84
 * magazines.
84
 * magazines.
85
 *
85
 *
86
 * @todo
86
 * @todo
87
 * For better CPU-scaling the magazine allocation strategy should
87
 * For better CPU-scaling the magazine allocation strategy should
88
 * be extended. Currently, if the cache does not have magazine, it asks
88
 * be extended. Currently, if the cache does not have magazine, it asks
89
 * for non-cpu cached magazine cache to provide one. It might be feasible
89
 * for non-cpu cached magazine cache to provide one. It might be feasible
90
 * to add cpu-cached magazine cache (which would allocate it's magazines
90
 * to add cpu-cached magazine cache (which would allocate it's magazines
91
 * from non-cpu-cached mag. cache). This would provide a nice per-cpu
91
 * from non-cpu-cached mag. cache). This would provide a nice per-cpu
92
 * buffer. The other possibility is to use the per-cache
92
 * buffer. The other possibility is to use the per-cache
93
 * 'empty-magazine-list', which decreases competing for 1 per-system
93
 * 'empty-magazine-list', which decreases competing for 1 per-system
94
 * magazine cache.
94
 * magazine cache.
95
 *
95
 *
96
 * @todo
96
 * @todo
97
 * it might be good to add granularity of locks even to slab level,
97
 * it might be good to add granularity of locks even to slab level,
98
 * we could then try_spinlock over all partial slabs and thus improve
98
 * we could then try_spinlock over all partial slabs and thus improve
99
 * scalability even on slab level
99
 * scalability even on slab level
100
 */
100
 */
101
 
101
 
102
#include <synch/spinlock.h>
102
#include <synch/spinlock.h>
103
#include <mm/slab.h>
103
#include <mm/slab.h>
104
#include <adt/list.h>
104
#include <adt/list.h>
105
#include <memstr.h>
105
#include <memstr.h>
106
#include <align.h>
106
#include <align.h>
107
#include <mm/frame.h>
107
#include <mm/frame.h>
108
#include <config.h>
108
#include <config.h>
109
#include <print.h>
109
#include <print.h>
110
#include <arch.h>
110
#include <arch.h>
111
#include <panic.h>
111
#include <panic.h>
112
#include <debug.h>
112
#include <debug.h>
113
#include <bitops.h>
113
#include <bitops.h>
114
#include <macros.h>
114
#include <macros.h>
115
 
115
 
116
SPINLOCK_INITIALIZE(slab_cache_lock);
116
SPINLOCK_INITIALIZE(slab_cache_lock);
117
static LIST_INITIALIZE(slab_cache_list);
117
static LIST_INITIALIZE(slab_cache_list);
118
 
118
 
119
/** Magazine cache */
119
/** Magazine cache */
120
static slab_cache_t mag_cache;
120
static slab_cache_t mag_cache;
121
/** Cache for cache descriptors */
121
/** Cache for cache descriptors */
122
static slab_cache_t slab_cache_cache;
122
static slab_cache_t slab_cache_cache;
123
/** Cache for external slab descriptors
123
/** Cache for external slab descriptors
124
 * This time we want per-cpu cache, so do not make it static
124
 * This time we want per-cpu cache, so do not make it static
125
 * - using slab for internal slab structures will not deadlock,
125
 * - using slab for internal slab structures will not deadlock,
126
 *   as all slab structures are 'small' - control structures of
126
 *   as all slab structures are 'small' - control structures of
127
 *   their caches do not require further allocation
127
 *   their caches do not require further allocation
128
 */
128
 */
129
static slab_cache_t *slab_extern_cache;
129
static slab_cache_t *slab_extern_cache;
130
/** Caches for malloc */
130
/** Caches for malloc */
131
static slab_cache_t *malloc_caches[SLAB_MAX_MALLOC_W - SLAB_MIN_MALLOC_W + 1];
131
static slab_cache_t *malloc_caches[SLAB_MAX_MALLOC_W - SLAB_MIN_MALLOC_W + 1];
132
char *malloc_names[] =  {
132
char *malloc_names[] =  {
133
    "malloc-16",
133
    "malloc-16",
134
    "malloc-32",
134
    "malloc-32",
135
    "malloc-64",
135
    "malloc-64",
136
    "malloc-128",
136
    "malloc-128",
137
    "malloc-256",
137
    "malloc-256",
138
    "malloc-512",
138
    "malloc-512",
139
    "malloc-1K",
139
    "malloc-1K",
140
    "malloc-2K",
140
    "malloc-2K",
141
    "malloc-4K",
141
    "malloc-4K",
142
    "malloc-8K",
142
    "malloc-8K",
143
    "malloc-16K",
143
    "malloc-16K",
144
    "malloc-32K",
144
    "malloc-32K",
145
    "malloc-64K",
145
    "malloc-64K",
146
    "malloc-128K",
146
    "malloc-128K",
147
    "malloc-256K"
147
    "malloc-256K"
148
};
148
};
149
 
149
 
150
/** Slab descriptor */
150
/** Slab descriptor */
151
typedef struct {
151
typedef struct {
152
    slab_cache_t *cache;    /**< Pointer to parent cache. */
152
    slab_cache_t *cache;    /**< Pointer to parent cache. */
153
    link_t link;        /**< List of full/partial slabs. */
153
    link_t link;        /**< List of full/partial slabs. */
154
    void *start;        /**< Start address of first available item. */
154
    void *start;        /**< Start address of first available item. */
155
    count_t available;  /**< Count of available items in this slab. */
155
    count_t available;  /**< Count of available items in this slab. */
156
    index_t nextavail;  /**< The index of next available item. */
156
    index_t nextavail;  /**< The index of next available item. */
157
} slab_t;
157
} slab_t;
158
 
158
 
159
#ifdef CONFIG_DEBUG
159
#ifdef CONFIG_DEBUG
160
static int _slab_initialized = 0;
160
static int _slab_initialized = 0;
161
#endif
161
#endif
162
 
162
 
163
/**************************************/
163
/**************************************/
164
/* Slab allocation functions          */
164
/* Slab allocation functions          */
165
 
165
 
166
/**
166
/**
167
 * Allocate frames for slab space and initialize
167
 * Allocate frames for slab space and initialize
168
 *
168
 *
169
 */
169
 */
170
static slab_t * slab_space_alloc(slab_cache_t *cache, int flags)
170
static slab_t *slab_space_alloc(slab_cache_t *cache, int flags)
171
{
171
{
172
    void *data;
172
    void *data;
173
    slab_t *slab;
173
    slab_t *slab;
174
    size_t fsize;
174
    size_t fsize;
175
    unsigned int i;
175
    unsigned int i;
176
    unsigned int zone = 0;
176
    count_t zone = 0;
177
   
177
   
178
    data = frame_alloc_generic(cache->order, FRAME_KA | flags, &zone);
178
    data = frame_alloc_generic(cache->order, FRAME_KA | flags, &zone);
179
    if (!data) {
179
    if (!data) {
180
        return NULL;
180
        return NULL;
181
    }
181
    }
182
    if (! (cache->flags & SLAB_CACHE_SLINSIDE)) {
182
    if (!(cache->flags & SLAB_CACHE_SLINSIDE)) {
183
        slab = slab_alloc(slab_extern_cache, flags);
183
        slab = slab_alloc(slab_extern_cache, flags);
184
        if (!slab) {
184
        if (!slab) {
185
            frame_free(KA2PA(data));
185
            frame_free(KA2PA(data));
186
            return NULL;
186
            return NULL;
187
        }
187
        }
188
    } else {
188
    } else {
189
        fsize = (PAGE_SIZE << cache->order);
189
        fsize = (PAGE_SIZE << cache->order);
190
        slab = data + fsize - sizeof(*slab);
190
        slab = data + fsize - sizeof(*slab);
191
    }
191
    }
192
   
192
   
193
    /* Fill in slab structures */
193
    /* Fill in slab structures */
194
    for (i = 0; i < ((unsigned int) 1 << cache->order); i++)
194
    for (i = 0; i < ((unsigned int) 1 << cache->order); i++)
195
        frame_set_parent(ADDR2PFN(KA2PA(data)) + i, slab, zone);
195
        frame_set_parent(ADDR2PFN(KA2PA(data)) + i, slab, zone);
196
 
196
 
197
    slab->start = data;
197
    slab->start = data;
198
    slab->available = cache->objects;
198
    slab->available = cache->objects;
199
    slab->nextavail = 0;
199
    slab->nextavail = 0;
200
    slab->cache = cache;
200
    slab->cache = cache;
201
 
201
 
202
    for (i = 0; i < cache->objects; i++)
202
    for (i = 0; i < cache->objects; i++)
203
        *((int *) (slab->start + i*cache->size)) = i+1;
203
        *((int *) (slab->start + i*cache->size)) = i + 1;
204
 
204
 
205
    atomic_inc(&cache->allocated_slabs);
205
    atomic_inc(&cache->allocated_slabs);
206
    return slab;
206
    return slab;
207
}
207
}
208
 
208
 
209
/**
209
/**
210
 * Deallocate space associated with slab
210
 * Deallocate space associated with slab
211
 *
211
 *
212
 * @return number of freed frames
212
 * @return number of freed frames
213
 */
213
 */
214
static count_t slab_space_free(slab_cache_t *cache, slab_t *slab)
214
static count_t slab_space_free(slab_cache_t *cache, slab_t *slab)
215
{
215
{
216
    frame_free(KA2PA(slab->start));
216
    frame_free(KA2PA(slab->start));
217
    if (! (cache->flags & SLAB_CACHE_SLINSIDE))
217
    if (! (cache->flags & SLAB_CACHE_SLINSIDE))
218
        slab_free(slab_extern_cache, slab);
218
        slab_free(slab_extern_cache, slab);
219
 
219
 
220
    atomic_dec(&cache->allocated_slabs);
220
    atomic_dec(&cache->allocated_slabs);
221
   
221
   
222
    return 1 << cache->order;
222
    return 1 << cache->order;
223
}
223
}
224
 
224
 
225
/** Map object to slab structure */
225
/** Map object to slab structure */
226
static slab_t * obj2slab(void *obj)
226
static slab_t * obj2slab(void *obj)
227
{
227
{
228
    return (slab_t *) frame_get_parent(ADDR2PFN(KA2PA(obj)), 0);
228
    return (slab_t *) frame_get_parent(ADDR2PFN(KA2PA(obj)), 0);
229
}
229
}
230
 
230
 
231
/**************************************/
231
/**************************************/
232
/* Slab functions */
232
/* Slab functions */
233
 
233
 
234
 
234
 
235
/**
235
/**
236
 * Return object to slab and call a destructor
236
 * Return object to slab and call a destructor
237
 *
237
 *
238
 * @param slab If the caller knows directly slab of the object, otherwise NULL
238
 * @param slab If the caller knows directly slab of the object, otherwise NULL
239
 *
239
 *
240
 * @return Number of freed pages
240
 * @return Number of freed pages
241
 */
241
 */
242
static count_t slab_obj_destroy(slab_cache_t *cache, void *obj,
242
static count_t slab_obj_destroy(slab_cache_t *cache, void *obj, slab_t *slab)
243
                slab_t *slab)
-
 
244
{
243
{
245
    int freed = 0;
244
    int freed = 0;
246
 
245
 
247
    if (!slab)
246
    if (!slab)
248
        slab = obj2slab(obj);
247
        slab = obj2slab(obj);
249
 
248
 
250
    ASSERT(slab->cache == cache);
249
    ASSERT(slab->cache == cache);
251
 
250
 
252
    if (cache->destructor)
251
    if (cache->destructor)
253
        freed = cache->destructor(obj);
252
        freed = cache->destructor(obj);
254
   
253
   
255
    spinlock_lock(&cache->slablock);
254
    spinlock_lock(&cache->slablock);
256
    ASSERT(slab->available < cache->objects);
255
    ASSERT(slab->available < cache->objects);
257
 
256
 
258
    *((int *)obj) = slab->nextavail;
257
    *((int *)obj) = slab->nextavail;
259
    slab->nextavail = (obj - slab->start)/cache->size;
258
    slab->nextavail = (obj - slab->start) / cache->size;
260
    slab->available++;
259
    slab->available++;
261
 
260
 
262
    /* Move it to correct list */
261
    /* Move it to correct list */
263
    if (slab->available == cache->objects) {
262
    if (slab->available == cache->objects) {
264
        /* Free associated memory */
263
        /* Free associated memory */
265
        list_remove(&slab->link);
264
        list_remove(&slab->link);
266
        spinlock_unlock(&cache->slablock);
265
        spinlock_unlock(&cache->slablock);
267
 
266
 
268
        return freed + slab_space_free(cache, slab);
267
        return freed + slab_space_free(cache, slab);
269
 
268
 
270
    } else if (slab->available == 1) {
269
    } else if (slab->available == 1) {
271
        /* It was in full, move to partial */
270
        /* It was in full, move to partial */
272
        list_remove(&slab->link);
271
        list_remove(&slab->link);
273
        list_prepend(&slab->link, &cache->partial_slabs);
272
        list_prepend(&slab->link, &cache->partial_slabs);
274
    }
273
    }
275
    spinlock_unlock(&cache->slablock);
274
    spinlock_unlock(&cache->slablock);
276
    return freed;
275
    return freed;
277
}
276
}
278
 
277
 
279
/**
278
/**
280
 * Take new object from slab or create new if needed
279
 * Take new object from slab or create new if needed
281
 *
280
 *
282
 * @return Object address or null
281
 * @return Object address or null
283
 */
282
 */
284
static void * slab_obj_create(slab_cache_t *cache, int flags)
283
static void *slab_obj_create(slab_cache_t *cache, int flags)
285
{
284
{
286
    slab_t *slab;
285
    slab_t *slab;
287
    void *obj;
286
    void *obj;
288
 
287
 
289
    spinlock_lock(&cache->slablock);
288
    spinlock_lock(&cache->slablock);
290
 
289
 
291
    if (list_empty(&cache->partial_slabs)) {
290
    if (list_empty(&cache->partial_slabs)) {
292
        /* Allow recursion and reclaiming
291
        /* Allow recursion and reclaiming
293
         * - this should work, as the slab control structures
292
         * - this should work, as the slab control structures
294
         *   are small and do not need to allocate with anything
293
         *   are small and do not need to allocate with anything
295
         *   other than frame_alloc when they are allocating,
294
         *   other than frame_alloc when they are allocating,
296
         *   that's why we should get recursion at most 1-level deep
295
         *   that's why we should get recursion at most 1-level deep
297
         */
296
         */
298
        spinlock_unlock(&cache->slablock);
297
        spinlock_unlock(&cache->slablock);
299
        slab = slab_space_alloc(cache, flags);
298
        slab = slab_space_alloc(cache, flags);
300
        if (!slab)
299
        if (!slab)
301
            return NULL;
300
            return NULL;
302
        spinlock_lock(&cache->slablock);
301
        spinlock_lock(&cache->slablock);
303
    } else {
302
    } else {
304
        slab = list_get_instance(cache->partial_slabs.next, slab_t, link);
303
        slab = list_get_instance(cache->partial_slabs.next, slab_t,
-
 
304
            link);
305
        list_remove(&slab->link);
305
        list_remove(&slab->link);
306
    }
306
    }
307
    obj = slab->start + slab->nextavail * cache->size;
307
    obj = slab->start + slab->nextavail * cache->size;
308
    slab->nextavail = *((int *)obj);
308
    slab->nextavail = *((int *)obj);
309
    slab->available--;
309
    slab->available--;
310
 
310
 
311
    if (!slab->available)
311
    if (!slab->available)
312
        list_prepend(&slab->link, &cache->full_slabs);
312
        list_prepend(&slab->link, &cache->full_slabs);
313
    else
313
    else
314
        list_prepend(&slab->link, &cache->partial_slabs);
314
        list_prepend(&slab->link, &cache->partial_slabs);
315
 
315
 
316
    spinlock_unlock(&cache->slablock);
316
    spinlock_unlock(&cache->slablock);
317
 
317
 
318
    if (cache->constructor && cache->constructor(obj, flags)) {
318
    if (cache->constructor && cache->constructor(obj, flags)) {
319
        /* Bad, bad, construction failed */
319
        /* Bad, bad, construction failed */
320
        slab_obj_destroy(cache, obj, slab);
320
        slab_obj_destroy(cache, obj, slab);
321
        return NULL;
321
        return NULL;
322
    }
322
    }
323
    return obj;
323
    return obj;
324
}
324
}
325
 
325
 
326
/**************************************/
326
/**************************************/
327
/* CPU-Cache slab functions */
327
/* CPU-Cache slab functions */
328
 
328
 
329
/**
329
/**
330
 * Finds a full magazine in cache, takes it from list
330
 * Finds a full magazine in cache, takes it from list
331
 * and returns it
331
 * and returns it
332
 *
332
 *
333
 * @param first If true, return first, else last mag
333
 * @param first If true, return first, else last mag
334
 */
334
 */
335
static slab_magazine_t * get_mag_from_cache(slab_cache_t *cache,
335
static slab_magazine_t *get_mag_from_cache(slab_cache_t *cache, int first)
336
                        int first)
-
 
337
{
336
{
338
    slab_magazine_t *mag = NULL;
337
    slab_magazine_t *mag = NULL;
339
    link_t *cur;
338
    link_t *cur;
340
 
339
 
341
    spinlock_lock(&cache->maglock);
340
    spinlock_lock(&cache->maglock);
342
    if (!list_empty(&cache->magazines)) {
341
    if (!list_empty(&cache->magazines)) {
343
        if (first)
342
        if (first)
344
            cur = cache->magazines.next;
343
            cur = cache->magazines.next;
345
        else
344
        else
346
            cur = cache->magazines.prev;
345
            cur = cache->magazines.prev;
347
        mag = list_get_instance(cur, slab_magazine_t, link);
346
        mag = list_get_instance(cur, slab_magazine_t, link);
348
        list_remove(&mag->link);
347
        list_remove(&mag->link);
349
        atomic_dec(&cache->magazine_counter);
348
        atomic_dec(&cache->magazine_counter);
350
    }
349
    }
351
    spinlock_unlock(&cache->maglock);
350
    spinlock_unlock(&cache->maglock);
352
    return mag;
351
    return mag;
353
}
352
}
354
 
353
 
355
/** Prepend magazine to magazine list in cache */
354
/** Prepend magazine to magazine list in cache */
356
static void put_mag_to_cache(slab_cache_t *cache, slab_magazine_t *mag)
355
static void put_mag_to_cache(slab_cache_t *cache, slab_magazine_t *mag)
357
{
356
{
358
    spinlock_lock(&cache->maglock);
357
    spinlock_lock(&cache->maglock);
359
 
358
 
360
    list_prepend(&mag->link, &cache->magazines);
359
    list_prepend(&mag->link, &cache->magazines);
361
    atomic_inc(&cache->magazine_counter);
360
    atomic_inc(&cache->magazine_counter);
362
   
361
   
363
    spinlock_unlock(&cache->maglock);
362
    spinlock_unlock(&cache->maglock);
364
}
363
}
365
 
364
 
366
/**
365
/**
367
 * Free all objects in magazine and free memory associated with magazine
366
 * Free all objects in magazine and free memory associated with magazine
368
 *
367
 *
369
 * @return Number of freed pages
368
 * @return Number of freed pages
370
 */
369
 */
371
static count_t magazine_destroy(slab_cache_t *cache,
370
static count_t magazine_destroy(slab_cache_t *cache, slab_magazine_t *mag)
372
                slab_magazine_t *mag)
-
 
373
{
371
{
374
    unsigned int i;
372
    unsigned int i;
375
    count_t frames = 0;
373
    count_t frames = 0;
376
 
374
 
377
    for (i = 0; i < mag->busy; i++) {
375
    for (i = 0; i < mag->busy; i++) {
378
        frames += slab_obj_destroy(cache, mag->objs[i], NULL);
376
        frames += slab_obj_destroy(cache, mag->objs[i], NULL);
379
        atomic_dec(&cache->cached_objs);
377
        atomic_dec(&cache->cached_objs);
380
    }
378
    }
381
   
379
   
382
    slab_free(&mag_cache, mag);
380
    slab_free(&mag_cache, mag);
383
 
381
 
384
    return frames;
382
    return frames;
385
}
383
}
386
 
384
 
387
/**
385
/**
388
 * Find full magazine, set it as current and return it
386
 * Find full magazine, set it as current and return it
389
 *
387
 *
390
 * Assume cpu_magazine lock is held
388
 * Assume cpu_magazine lock is held
391
 */
389
 */
392
static slab_magazine_t * get_full_current_mag(slab_cache_t *cache)
390
static slab_magazine_t *get_full_current_mag(slab_cache_t *cache)
393
{
391
{
394
    slab_magazine_t *cmag, *lastmag, *newmag;
392
    slab_magazine_t *cmag, *lastmag, *newmag;
395
 
393
 
396
    cmag = cache->mag_cache[CPU->id].current;
394
    cmag = cache->mag_cache[CPU->id].current;
397
    lastmag = cache->mag_cache[CPU->id].last;
395
    lastmag = cache->mag_cache[CPU->id].last;
398
    if (cmag) { /* First try local CPU magazines */
396
    if (cmag) { /* First try local CPU magazines */
399
        if (cmag->busy)
397
        if (cmag->busy)
400
            return cmag;
398
            return cmag;
401
 
399
 
402
        if (lastmag && lastmag->busy) {
400
        if (lastmag && lastmag->busy) {
403
            cache->mag_cache[CPU->id].current = lastmag;
401
            cache->mag_cache[CPU->id].current = lastmag;
404
            cache->mag_cache[CPU->id].last = cmag;
402
            cache->mag_cache[CPU->id].last = cmag;
405
            return lastmag;
403
            return lastmag;
406
        }
404
        }
407
    }
405
    }
408
    /* Local magazines are empty, import one from magazine list */
406
    /* Local magazines are empty, import one from magazine list */
409
    newmag = get_mag_from_cache(cache, 1);
407
    newmag = get_mag_from_cache(cache, 1);
410
    if (!newmag)
408
    if (!newmag)
411
        return NULL;
409
        return NULL;
412
 
410
 
413
    if (lastmag)
411
    if (lastmag)
414
        magazine_destroy(cache, lastmag);
412
        magazine_destroy(cache, lastmag);
415
 
413
 
416
    cache->mag_cache[CPU->id].last = cmag;
414
    cache->mag_cache[CPU->id].last = cmag;
417
    cache->mag_cache[CPU->id].current = newmag;
415
    cache->mag_cache[CPU->id].current = newmag;
418
    return newmag;
416
    return newmag;
419
}
417
}
420
 
418
 
421
/**
419
/**
422
 * Try to find object in CPU-cache magazines
420
 * Try to find object in CPU-cache magazines
423
 *
421
 *
424
 * @return Pointer to object or NULL if not available
422
 * @return Pointer to object or NULL if not available
425
 */
423
 */
426
static void * magazine_obj_get(slab_cache_t *cache)
424
static void *magazine_obj_get(slab_cache_t *cache)
427
{
425
{
428
    slab_magazine_t *mag;
426
    slab_magazine_t *mag;
429
    void *obj;
427
    void *obj;
430
 
428
 
431
    if (!CPU)
429
    if (!CPU)
432
        return NULL;
430
        return NULL;
433
 
431
 
434
    spinlock_lock(&cache->mag_cache[CPU->id].lock);
432
    spinlock_lock(&cache->mag_cache[CPU->id].lock);
435
 
433
 
436
    mag = get_full_current_mag(cache);
434
    mag = get_full_current_mag(cache);
437
    if (!mag) {
435
    if (!mag) {
438
        spinlock_unlock(&cache->mag_cache[CPU->id].lock);
436
        spinlock_unlock(&cache->mag_cache[CPU->id].lock);
439
        return NULL;
437
        return NULL;
440
    }
438
    }
441
    obj = mag->objs[--mag->busy];
439
    obj = mag->objs[--mag->busy];
442
    spinlock_unlock(&cache->mag_cache[CPU->id].lock);
440
    spinlock_unlock(&cache->mag_cache[CPU->id].lock);
443
    atomic_dec(&cache->cached_objs);
441
    atomic_dec(&cache->cached_objs);
444
   
442
   
445
    return obj;
443
    return obj;
446
}
444
}
447
 
445
 
448
/**
446
/**
449
 * Assure that the current magazine is empty, return pointer to it, or NULL if
447
 * Assure that the current magazine is empty, return pointer to it, or NULL if
450
 * no empty magazine is available and cannot be allocated
448
 * no empty magazine is available and cannot be allocated
451
 *
449
 *
452
 * Assume mag_cache[CPU->id].lock is held
450
 * Assume mag_cache[CPU->id].lock is held
453
 *
451
 *
454
 * We have 2 magazines bound to processor.
452
 * We have 2 magazines bound to processor.
455
 * First try the current.
453
 * First try the current.
456
 *  If full, try the last.
454
 *  If full, try the last.
457
 *   If full, put to magazines list.
455
 *   If full, put to magazines list.
458
 *   allocate new, exchange last & current
456
 *   allocate new, exchange last & current
459
 *
457
 *
460
 */
458
 */
461
static slab_magazine_t * make_empty_current_mag(slab_cache_t *cache)
459
static slab_magazine_t *make_empty_current_mag(slab_cache_t *cache)
462
{
460
{
463
    slab_magazine_t *cmag,*lastmag,*newmag;
461
    slab_magazine_t *cmag,*lastmag,*newmag;
464
 
462
 
465
    cmag = cache->mag_cache[CPU->id].current;
463
    cmag = cache->mag_cache[CPU->id].current;
466
    lastmag = cache->mag_cache[CPU->id].last;
464
    lastmag = cache->mag_cache[CPU->id].last;
467
 
465
 
468
    if (cmag) {
466
    if (cmag) {
469
        if (cmag->busy < cmag->size)
467
        if (cmag->busy < cmag->size)
470
            return cmag;
468
            return cmag;
471
        if (lastmag && lastmag->busy < lastmag->size) {
469
        if (lastmag && lastmag->busy < lastmag->size) {
472
            cache->mag_cache[CPU->id].last = cmag;
470
            cache->mag_cache[CPU->id].last = cmag;
473
            cache->mag_cache[CPU->id].current = lastmag;
471
            cache->mag_cache[CPU->id].current = lastmag;
474
            return lastmag;
472
            return lastmag;
475
        }
473
        }
476
    }
474
    }
477
    /* current | last are full | nonexistent, allocate new */
475
    /* current | last are full | nonexistent, allocate new */
478
    /* We do not want to sleep just because of caching */
476
    /* We do not want to sleep just because of caching */
479
    /* Especially we do not want reclaiming to start, as
477
    /* Especially we do not want reclaiming to start, as
480
     * this would deadlock */
478
     * this would deadlock */
481
    newmag = slab_alloc(&mag_cache, FRAME_ATOMIC | FRAME_NO_RECLAIM);
479
    newmag = slab_alloc(&mag_cache, FRAME_ATOMIC | FRAME_NO_RECLAIM);
482
    if (!newmag)
480
    if (!newmag)
483
        return NULL;
481
        return NULL;
484
    newmag->size = SLAB_MAG_SIZE;
482
    newmag->size = SLAB_MAG_SIZE;
485
    newmag->busy = 0;
483
    newmag->busy = 0;
486
 
484
 
487
    /* Flush last to magazine list */
485
    /* Flush last to magazine list */
488
    if (lastmag)
486
    if (lastmag)
489
        put_mag_to_cache(cache, lastmag);
487
        put_mag_to_cache(cache, lastmag);
490
 
488
 
491
    /* Move current as last, save new as current */
489
    /* Move current as last, save new as current */
492
    cache->mag_cache[CPU->id].last = cmag; 
490
    cache->mag_cache[CPU->id].last = cmag; 
493
    cache->mag_cache[CPU->id].current = newmag;
491
    cache->mag_cache[CPU->id].current = newmag;
494
 
492
 
495
    return newmag;
493
    return newmag;
496
}
494
}
497
 
495
 
498
/**
496
/**
499
 * Put object into CPU-cache magazine
497
 * Put object into CPU-cache magazine
500
 *
498
 *
501
 * @return 0 - success, -1 - could not get memory
499
 * @return 0 - success, -1 - could not get memory
502
 */
500
 */
503
static int magazine_obj_put(slab_cache_t *cache, void *obj)
501
static int magazine_obj_put(slab_cache_t *cache, void *obj)
504
{
502
{
505
    slab_magazine_t *mag;
503
    slab_magazine_t *mag;
506
 
504
 
507
    if (!CPU)
505
    if (!CPU)
508
        return -1;
506
        return -1;
509
 
507
 
510
    spinlock_lock(&cache->mag_cache[CPU->id].lock);
508
    spinlock_lock(&cache->mag_cache[CPU->id].lock);
511
 
509
 
512
    mag = make_empty_current_mag(cache);
510
    mag = make_empty_current_mag(cache);
513
    if (!mag) {
511
    if (!mag) {
514
        spinlock_unlock(&cache->mag_cache[CPU->id].lock);
512
        spinlock_unlock(&cache->mag_cache[CPU->id].lock);
515
        return -1;
513
        return -1;
516
    }
514
    }
517
   
515
   
518
    mag->objs[mag->busy++] = obj;
516
    mag->objs[mag->busy++] = obj;
519
 
517
 
520
    spinlock_unlock(&cache->mag_cache[CPU->id].lock);
518
    spinlock_unlock(&cache->mag_cache[CPU->id].lock);
521
    atomic_inc(&cache->cached_objs);
519
    atomic_inc(&cache->cached_objs);
522
    return 0;
520
    return 0;
523
}
521
}
524
 
522
 
525
 
523
 
526
/**************************************/
524
/**************************************/
527
/* Slab cache functions */
525
/* Slab cache functions */
528
 
526
 
529
/** Return number of objects that fit in certain cache size */
527
/** Return number of objects that fit in certain cache size */
530
static unsigned int comp_objects(slab_cache_t *cache)
528
static unsigned int comp_objects(slab_cache_t *cache)
531
{
529
{
532
    if (cache->flags & SLAB_CACHE_SLINSIDE)
530
    if (cache->flags & SLAB_CACHE_SLINSIDE)
533
        return ((PAGE_SIZE << cache->order) - sizeof(slab_t)) / cache->size;
531
        return ((PAGE_SIZE << cache->order) - sizeof(slab_t)) /
-
 
532
            cache->size;
534
    else
533
    else
535
        return (PAGE_SIZE << cache->order) / cache->size;
534
        return (PAGE_SIZE << cache->order) / cache->size;
536
}
535
}
537
 
536
 
538
/** Return wasted space in slab */
537
/** Return wasted space in slab */
539
static unsigned int badness(slab_cache_t *cache)
538
static unsigned int badness(slab_cache_t *cache)
540
{
539
{
541
    unsigned int objects;
540
    unsigned int objects;
542
    unsigned int ssize;
541
    unsigned int ssize;
543
 
542
 
544
    objects = comp_objects(cache);
543
    objects = comp_objects(cache);
545
    ssize = PAGE_SIZE << cache->order;
544
    ssize = PAGE_SIZE << cache->order;
546
    if (cache->flags & SLAB_CACHE_SLINSIDE)
545
    if (cache->flags & SLAB_CACHE_SLINSIDE)
547
        ssize -= sizeof(slab_t);
546
        ssize -= sizeof(slab_t);
548
    return ssize - objects * cache->size;
547
    return ssize - objects * cache->size;
549
}
548
}
550
 
549
 
551
/**
550
/**
552
 * Initialize mag_cache structure in slab cache
551
 * Initialize mag_cache structure in slab cache
553
 */
552
 */
554
static void make_magcache(slab_cache_t *cache)
553
static void make_magcache(slab_cache_t *cache)
555
{
554
{
556
    unsigned int i;
555
    unsigned int i;
557
   
556
   
558
    ASSERT(_slab_initialized >= 2);
557
    ASSERT(_slab_initialized >= 2);
559
 
558
 
560
    cache->mag_cache = malloc(sizeof(slab_mag_cache_t)*config.cpu_count,0);
559
    cache->mag_cache = malloc(sizeof(slab_mag_cache_t) * config.cpu_count,
-
 
560
        0);
561
    for (i = 0; i < config.cpu_count; i++) {
561
    for (i = 0; i < config.cpu_count; i++) {
562
        memsetb((uintptr_t)&cache->mag_cache[i],
562
        memsetb(&cache->mag_cache[i], sizeof(cache->mag_cache[i]), 0);
563
            sizeof(cache->mag_cache[i]), 0);
563
        spinlock_initialize(&cache->mag_cache[i].lock,
564
        spinlock_initialize(&cache->mag_cache[i].lock, "slab_maglock_cpu");
564
            "slab_maglock_cpu");
565
    }
565
    }
566
}
566
}
567
 
567
 
568
/** Initialize allocated memory as a slab cache */
568
/** Initialize allocated memory as a slab cache */
569
static void
569
static void
570
_slab_cache_create(slab_cache_t *cache,
570
_slab_cache_create(slab_cache_t *cache, char *name, size_t size, size_t align,
571
           char *name,
-
 
572
           size_t size,
-
 
573
           size_t align,
-
 
574
           int (*constructor)(void *obj, int kmflag),
571
    int (*constructor)(void *obj, int kmflag), int (*destructor)(void *obj),
575
           int (*destructor)(void *obj),
-
 
576
           int flags)
572
    int flags)
577
{
573
{
578
    int pages;
574
    int pages;
579
    ipl_t ipl;
575
    ipl_t ipl;
580
 
576
 
581
    memsetb((uintptr_t)cache, sizeof(*cache), 0);
577
    memsetb(cache, sizeof(*cache), 0);
582
    cache->name = name;
578
    cache->name = name;
583
 
579
 
584
    if (align < sizeof(unative_t))
580
    if (align < sizeof(unative_t))
585
        align = sizeof(unative_t);
581
        align = sizeof(unative_t);
586
    size = ALIGN_UP(size, align);
582
    size = ALIGN_UP(size, align);
587
       
583
       
588
    cache->size = size;
584
    cache->size = size;
589
 
585
 
590
    cache->constructor = constructor;
586
    cache->constructor = constructor;
591
    cache->destructor = destructor;
587
    cache->destructor = destructor;
592
    cache->flags = flags;
588
    cache->flags = flags;
593
 
589
 
594
    list_initialize(&cache->full_slabs);
590
    list_initialize(&cache->full_slabs);
595
    list_initialize(&cache->partial_slabs);
591
    list_initialize(&cache->partial_slabs);
596
    list_initialize(&cache->magazines);
592
    list_initialize(&cache->magazines);
597
    spinlock_initialize(&cache->slablock, "slab_lock");
593
    spinlock_initialize(&cache->slablock, "slab_lock");
598
    spinlock_initialize(&cache->maglock, "slab_maglock");
594
    spinlock_initialize(&cache->maglock, "slab_maglock");
599
    if (! (cache->flags & SLAB_CACHE_NOMAGAZINE))
595
    if (!(cache->flags & SLAB_CACHE_NOMAGAZINE))
600
        make_magcache(cache);
596
        make_magcache(cache);
601
 
597
 
602
    /* Compute slab sizes, object counts in slabs etc. */
598
    /* Compute slab sizes, object counts in slabs etc. */
603
    if (cache->size < SLAB_INSIDE_SIZE)
599
    if (cache->size < SLAB_INSIDE_SIZE)
604
        cache->flags |= SLAB_CACHE_SLINSIDE;
600
        cache->flags |= SLAB_CACHE_SLINSIDE;
605
 
601
 
606
    /* Minimum slab order */
602
    /* Minimum slab order */
607
    pages = SIZE2FRAMES(cache->size);
603
    pages = SIZE2FRAMES(cache->size);
608
    /* We need the 2^order >= pages */
604
    /* We need the 2^order >= pages */
609
    if (pages == 1)
605
    if (pages == 1)
610
        cache->order = 0;
606
        cache->order = 0;
611
    else
607
    else
612
        cache->order = fnzb(pages-1)+1;
608
        cache->order = fnzb(pages - 1) + 1;
613
 
609
 
614
    while (badness(cache) > SLAB_MAX_BADNESS(cache)) {
610
    while (badness(cache) > SLAB_MAX_BADNESS(cache)) {
615
        cache->order += 1;
611
        cache->order += 1;
616
    }
612
    }
617
    cache->objects = comp_objects(cache);
613
    cache->objects = comp_objects(cache);
618
    /* If info fits in, put it inside */
614
    /* If info fits in, put it inside */
619
    if (badness(cache) > sizeof(slab_t))
615
    if (badness(cache) > sizeof(slab_t))
620
        cache->flags |= SLAB_CACHE_SLINSIDE;
616
        cache->flags |= SLAB_CACHE_SLINSIDE;
621
 
617
 
622
    /* Add cache to cache list */
618
    /* Add cache to cache list */
623
    ipl = interrupts_disable();
619
    ipl = interrupts_disable();
624
    spinlock_lock(&slab_cache_lock);
620
    spinlock_lock(&slab_cache_lock);
625
 
621
 
626
    list_append(&cache->link, &slab_cache_list);
622
    list_append(&cache->link, &slab_cache_list);
627
 
623
 
628
    spinlock_unlock(&slab_cache_lock);
624
    spinlock_unlock(&slab_cache_lock);
629
    interrupts_restore(ipl);
625
    interrupts_restore(ipl);
630
}
626
}
631
 
627
 
632
/** Create slab cache  */
628
/** Create slab cache  */
633
slab_cache_t * slab_cache_create(char *name,
-
 
634
                 size_t size,
629
slab_cache_t *
635
                 size_t align,
630
slab_cache_create(char *name, size_t size, size_t align,
636
                 int (*constructor)(void *obj, int kmflag),
631
    int (*constructor)(void *obj, int kmflag), int (*destructor)(void *obj),
637
                 int (*destructor)(void *obj),
-
 
638
                 int flags)
632
    int flags)
639
{
633
{
640
    slab_cache_t *cache;
634
    slab_cache_t *cache;
641
 
635
 
642
    cache = slab_alloc(&slab_cache_cache, 0);
636
    cache = slab_alloc(&slab_cache_cache, 0);
643
    _slab_cache_create(cache, name, size, align, constructor, destructor,
637
    _slab_cache_create(cache, name, size, align, constructor, destructor,
644
               flags);
638
        flags);
645
    return cache;
639
    return cache;
646
}
640
}
647
 
641
 
648
/**
642
/**
649
 * Reclaim space occupied by objects that are already free
643
 * Reclaim space occupied by objects that are already free
650
 *
644
 *
651
 * @param flags If contains SLAB_RECLAIM_ALL, do aggressive freeing
645
 * @param flags If contains SLAB_RECLAIM_ALL, do aggressive freeing
652
 * @return Number of freed pages
646
 * @return Number of freed pages
653
 */
647
 */
654
static count_t _slab_reclaim(slab_cache_t *cache, int flags)
648
static count_t _slab_reclaim(slab_cache_t *cache, int flags)
655
{
649
{
656
    unsigned int i;
650
    unsigned int i;
657
    slab_magazine_t *mag;
651
    slab_magazine_t *mag;
658
    count_t frames = 0;
652
    count_t frames = 0;
659
    int magcount;
653
    int magcount;
660
   
654
   
661
    if (cache->flags & SLAB_CACHE_NOMAGAZINE)
655
    if (cache->flags & SLAB_CACHE_NOMAGAZINE)
662
        return 0; /* Nothing to do */
656
        return 0; /* Nothing to do */
663
 
657
 
664
    /* We count up to original magazine count to avoid
658
    /* We count up to original magazine count to avoid
665
     * endless loop
659
     * endless loop
666
     */
660
     */
667
    magcount = atomic_get(&cache->magazine_counter);
661
    magcount = atomic_get(&cache->magazine_counter);
668
    while (magcount-- && (mag=get_mag_from_cache(cache,0))) {
662
    while (magcount-- && (mag=get_mag_from_cache(cache, 0))) {
669
        frames += magazine_destroy(cache,mag);
663
        frames += magazine_destroy(cache,mag);
670
        if (!(flags & SLAB_RECLAIM_ALL) && frames)
664
        if (!(flags & SLAB_RECLAIM_ALL) && frames)
671
            break;
665
            break;
672
    }
666
    }
673
   
667
   
674
    if (flags & SLAB_RECLAIM_ALL) {
668
    if (flags & SLAB_RECLAIM_ALL) {
675
        /* Free cpu-bound magazines */
669
        /* Free cpu-bound magazines */
676
        /* Destroy CPU magazines */
670
        /* Destroy CPU magazines */
677
        for (i = 0; i < config.cpu_count; i++) {
671
        for (i = 0; i < config.cpu_count; i++) {
678
            spinlock_lock(&cache->mag_cache[i].lock);
672
            spinlock_lock(&cache->mag_cache[i].lock);
679
 
673
 
680
            mag = cache->mag_cache[i].current;
674
            mag = cache->mag_cache[i].current;
681
            if (mag)
675
            if (mag)
682
                frames += magazine_destroy(cache, mag);
676
                frames += magazine_destroy(cache, mag);
683
            cache->mag_cache[i].current = NULL;
677
            cache->mag_cache[i].current = NULL;
684
           
678
           
685
            mag = cache->mag_cache[i].last;
679
            mag = cache->mag_cache[i].last;
686
            if (mag)
680
            if (mag)
687
                frames += magazine_destroy(cache, mag);
681
                frames += magazine_destroy(cache, mag);
688
            cache->mag_cache[i].last = NULL;
682
            cache->mag_cache[i].last = NULL;
689
 
683
 
690
            spinlock_unlock(&cache->mag_cache[i].lock);
684
            spinlock_unlock(&cache->mag_cache[i].lock);
691
        }
685
        }
692
    }
686
    }
693
 
687
 
694
    return frames;
688
    return frames;
695
}
689
}
696
 
690
 
697
/** Check that there are no slabs and remove cache from system  */
691
/** Check that there are no slabs and remove cache from system  */
698
void slab_cache_destroy(slab_cache_t *cache)
692
void slab_cache_destroy(slab_cache_t *cache)
699
{
693
{
700
    ipl_t ipl;
694
    ipl_t ipl;
701
 
695
 
702
    /* First remove cache from link, so that we don't need
696
    /* First remove cache from link, so that we don't need
703
     * to disable interrupts later
697
     * to disable interrupts later
704
     */
698
     */
705
 
699
 
706
    ipl = interrupts_disable();
700
    ipl = interrupts_disable();
707
    spinlock_lock(&slab_cache_lock);
701
    spinlock_lock(&slab_cache_lock);
708
 
702
 
709
    list_remove(&cache->link);
703
    list_remove(&cache->link);
710
 
704
 
711
    spinlock_unlock(&slab_cache_lock);
705
    spinlock_unlock(&slab_cache_lock);
712
    interrupts_restore(ipl);
706
    interrupts_restore(ipl);
713
 
707
 
714
    /* Do not lock anything, we assume the software is correct and
708
    /* Do not lock anything, we assume the software is correct and
715
     * does not touch the cache when it decides to destroy it */
709
     * does not touch the cache when it decides to destroy it */
716
   
710
   
717
    /* Destroy all magazines */
711
    /* Destroy all magazines */
718
    _slab_reclaim(cache, SLAB_RECLAIM_ALL);
712
    _slab_reclaim(cache, SLAB_RECLAIM_ALL);
719
 
713
 
720
    /* All slabs must be empty */
714
    /* All slabs must be empty */
721
    if (!list_empty(&cache->full_slabs) \
715
    if (!list_empty(&cache->full_slabs) ||
722
        || !list_empty(&cache->partial_slabs))
716
        !list_empty(&cache->partial_slabs))
723
        panic("Destroying cache that is not empty.");
717
        panic("Destroying cache that is not empty.");
724
 
718
 
725
    if (!(cache->flags & SLAB_CACHE_NOMAGAZINE))
719
    if (!(cache->flags & SLAB_CACHE_NOMAGAZINE))
726
        free(cache->mag_cache);
720
        free(cache->mag_cache);
727
    slab_free(&slab_cache_cache, cache);
721
    slab_free(&slab_cache_cache, cache);
728
}
722
}
729
 
723
 
730
/** Allocate new object from cache - if no flags given, always returns
724
/** Allocate new object from cache - if no flags given, always returns memory */
731
    memory */
-
 
732
void * slab_alloc(slab_cache_t *cache, int flags)
725
void *slab_alloc(slab_cache_t *cache, int flags)
733
{
726
{
734
    ipl_t ipl;
727
    ipl_t ipl;
735
    void *result = NULL;
728
    void *result = NULL;
736
   
729
   
737
    /* Disable interrupts to avoid deadlocks with interrupt handlers */
730
    /* Disable interrupts to avoid deadlocks with interrupt handlers */
738
    ipl = interrupts_disable();
731
    ipl = interrupts_disable();
739
 
732
 
740
    if (!(cache->flags & SLAB_CACHE_NOMAGAZINE)) {
733
    if (!(cache->flags & SLAB_CACHE_NOMAGAZINE)) {
741
        result = magazine_obj_get(cache);
734
        result = magazine_obj_get(cache);
742
    }
735
    }
743
    if (!result)
736
    if (!result)
744
        result = slab_obj_create(cache, flags);
737
        result = slab_obj_create(cache, flags);
745
 
738
 
746
    interrupts_restore(ipl);
739
    interrupts_restore(ipl);
747
 
740
 
748
    if (result)
741
    if (result)
749
        atomic_inc(&cache->allocated_objs);
742
        atomic_inc(&cache->allocated_objs);
750
 
743
 
751
    return result;
744
    return result;
752
}
745
}
753
 
746
 
754
/** Return object to cache, use slab if known  */
747
/** Return object to cache, use slab if known  */
755
static void _slab_free(slab_cache_t *cache, void *obj, slab_t *slab)
748
static void _slab_free(slab_cache_t *cache, void *obj, slab_t *slab)
756
{
749
{
757
    ipl_t ipl;
750
    ipl_t ipl;
758
 
751
 
759
    ipl = interrupts_disable();
752
    ipl = interrupts_disable();
760
 
753
 
761
    if ((cache->flags & SLAB_CACHE_NOMAGAZINE) \
754
    if ((cache->flags & SLAB_CACHE_NOMAGAZINE) ||
762
        || magazine_obj_put(cache, obj)) {
755
        magazine_obj_put(cache, obj)) {
763
 
-
 
764
        slab_obj_destroy(cache, obj, slab);
756
        slab_obj_destroy(cache, obj, slab);
765
 
757
 
766
    }
758
    }
767
    interrupts_restore(ipl);
759
    interrupts_restore(ipl);
768
    atomic_dec(&cache->allocated_objs);
760
    atomic_dec(&cache->allocated_objs);
769
}
761
}
770
 
762
 
771
/** Return slab object to cache */
763
/** Return slab object to cache */
772
void slab_free(slab_cache_t *cache, void *obj)
764
void slab_free(slab_cache_t *cache, void *obj)
773
{
765
{
774
    _slab_free(cache, obj, NULL);
766
    _slab_free(cache, obj, NULL);
775
}
767
}
776
 
768
 
777
/* Go through all caches and reclaim what is possible */
769
/* Go through all caches and reclaim what is possible */
778
count_t slab_reclaim(int flags)
770
count_t slab_reclaim(int flags)
779
{
771
{
780
    slab_cache_t *cache;
772
    slab_cache_t *cache;
781
    link_t *cur;
773
    link_t *cur;
782
    count_t frames = 0;
774
    count_t frames = 0;
783
 
775
 
784
    spinlock_lock(&slab_cache_lock);
776
    spinlock_lock(&slab_cache_lock);
785
 
777
 
786
    /* TODO: Add assert, that interrupts are disabled, otherwise
778
    /* TODO: Add assert, that interrupts are disabled, otherwise
787
     * memory allocation from interrupts can deadlock.
779
     * memory allocation from interrupts can deadlock.
788
     */
780
     */
789
 
781
 
790
    for (cur = slab_cache_list.next;cur!=&slab_cache_list; cur=cur->next) {
782
    for (cur = slab_cache_list.next; cur != &slab_cache_list;
-
 
783
        cur = cur->next) {
791
        cache = list_get_instance(cur, slab_cache_t, link);
784
        cache = list_get_instance(cur, slab_cache_t, link);
792
        frames += _slab_reclaim(cache, flags);
785
        frames += _slab_reclaim(cache, flags);
793
    }
786
    }
794
 
787
 
795
    spinlock_unlock(&slab_cache_lock);
788
    spinlock_unlock(&slab_cache_lock);
796
 
789
 
797
    return frames;
790
    return frames;
798
}
791
}
799
 
792
 
800
 
793
 
801
/* Print list of slabs */
794
/* Print list of slabs */
802
void slab_print_list(void)
795
void slab_print_list(void)
803
{
796
{
-
 
797
    int skip = 0;
-
 
798
 
-
 
799
    printf("slab name        size     pages  obj/pg slabs  cached allocated"
-
 
800
        " ctl\n");
-
 
801
    printf("---------------- -------- ------ ------ ------ ------ ---------"
-
 
802
        " ---\n");
-
 
803
 
-
 
804
    while (true) {
804
    slab_cache_t *cache;
805
        slab_cache_t *cache;
805
    link_t *cur;
806
        link_t *cur;
806
    ipl_t ipl;
807
        ipl_t ipl;
-
 
808
        int i;
-
 
809
 
-
 
810
        /*
-
 
811
         * We must not hold the slab_cache_lock spinlock when printing
-
 
812
         * the statistics. Otherwise we can easily deadlock if the print
-
 
813
         * needs to allocate memory.
-
 
814
         *
-
 
815
         * Therefore, we walk through the slab cache list, skipping some
-
 
816
         * amount of already processed caches during each iteration and
-
 
817
         * gathering statistics about the first unprocessed cache. For
-
 
818
         * the sake of printing the statistics, we realese the
-
 
819
         * slab_cache_lock and reacquire it afterwards. Then the walk
-
 
820
         * starts again.
-
 
821
         *
-
 
822
         * This limits both the efficiency and also accuracy of the
-
 
823
         * obtained statistics. The efficiency is decreased because the
-
 
824
         * time complexity of the algorithm is quadratic instead of
-
 
825
         * linear. The accuracy is impacted because we drop the lock
-
 
826
         * after processing one cache. If there is someone else
-
 
827
         * manipulating the cache list, we might omit an arbitrary
-
 
828
         * number of caches or process one cache multiple times.
-
 
829
         * However, we don't bleed for this algorithm for it is only
-
 
830
         * statistics.
-
 
831
         */
807
   
832
 
808
    ipl = interrupts_disable();
833
        ipl = interrupts_disable();
809
    spinlock_lock(&slab_cache_lock);
834
        spinlock_lock(&slab_cache_lock);
-
 
835
 
810
    printf("slab name        size     pages  obj/pg slabs  cached allocated ctl\n");
836
        for (i = 0, cur = slab_cache_list.next;
811
    printf("---------------- -------- ------ ------ ------ ------ --------- ---\n");
837
            i < skip && cur != &slab_cache_list;
-
 
838
            i++, cur = cur->next)
-
 
839
            ;
812
   
840
 
813
    for (cur = slab_cache_list.next; cur != &slab_cache_list; cur = cur->next) {
841
        if (cur == &slab_cache_list) {
-
 
842
            spinlock_unlock(&slab_cache_lock);
-
 
843
            interrupts_restore(ipl);
-
 
844
            break;
-
 
845
        }
-
 
846
 
-
 
847
        skip++;
-
 
848
 
814
        cache = list_get_instance(cur, slab_cache_t, link);
849
        cache = list_get_instance(cur, slab_cache_t, link);
-
 
850
 
-
 
851
        char *name = cache->name;
-
 
852
        uint8_t order = cache->order;
-
 
853
        size_t size = cache->size;
-
 
854
        unsigned int objects = cache->objects;
-
 
855
        long allocated_slabs = atomic_get(&cache->allocated_slabs);
-
 
856
        long cached_objs = atomic_get(&cache->cached_objs);
-
 
857
        long allocated_objs = atomic_get(&cache->allocated_objs);
-
 
858
        int flags = cache->flags;
-
 
859
       
-
 
860
        spinlock_unlock(&slab_cache_lock);
-
 
861
        interrupts_restore(ipl);
815
       
862
       
-
 
863
        printf("%-16s %8" PRIs " %6d %6u %6ld %6ld %9ld %-3s\n",
-
 
864
            name, size, (1 << order), objects, allocated_slabs,
-
 
865
            cached_objs, allocated_objs,
816
        printf("%-16s %8zd %6zd %6zd %6zd %6zd %9zd %-3s\n", cache->name, cache->size, (1 << cache->order), cache->objects, atomic_get(&cache->allocated_slabs), atomic_get(&cache->cached_objs), atomic_get(&cache->allocated_objs), cache->flags & SLAB_CACHE_SLINSIDE ? "in" : "out");
866
            flags & SLAB_CACHE_SLINSIDE ? "in" : "out");
817
    }
867
    }
818
    spinlock_unlock(&slab_cache_lock);
-
 
819
    interrupts_restore(ipl);
-
 
820
}
868
}
821
 
869
 
822
void slab_cache_init(void)
870
void slab_cache_init(void)
823
{
871
{
824
    int i, size;
872
    int i, size;
825
 
873
 
826
    /* Initialize magazine cache */
874
    /* Initialize magazine cache */
827
    _slab_cache_create(&mag_cache,
875
    _slab_cache_create(&mag_cache, "slab_magazine",
828
               "slab_magazine",
-
 
829
               sizeof(slab_magazine_t)+SLAB_MAG_SIZE*sizeof(void*),
876
        sizeof(slab_magazine_t) + SLAB_MAG_SIZE * sizeof(void*),
830
               sizeof(uintptr_t),
877
        sizeof(uintptr_t), NULL, NULL, SLAB_CACHE_NOMAGAZINE |
831
               NULL, NULL,
-
 
832
               SLAB_CACHE_NOMAGAZINE | SLAB_CACHE_SLINSIDE);
878
        SLAB_CACHE_SLINSIDE);
833
    /* Initialize slab_cache cache */
879
    /* Initialize slab_cache cache */
834
    _slab_cache_create(&slab_cache_cache,
880
    _slab_cache_create(&slab_cache_cache, "slab_cache",
835
               "slab_cache",
-
 
836
               sizeof(slab_cache_cache),
881
        sizeof(slab_cache_cache), sizeof(uintptr_t), NULL, NULL,
837
               sizeof(uintptr_t),
-
 
838
               NULL, NULL,
-
 
839
               SLAB_CACHE_NOMAGAZINE | SLAB_CACHE_SLINSIDE);
882
        SLAB_CACHE_NOMAGAZINE | SLAB_CACHE_SLINSIDE);
840
    /* Initialize external slab cache */
883
    /* Initialize external slab cache */
841
    slab_extern_cache = slab_cache_create("slab_extern",
884
    slab_extern_cache = slab_cache_create("slab_extern", sizeof(slab_t), 0,
842
                          sizeof(slab_t),
-
 
843
                          0, NULL, NULL,
-
 
844
                          SLAB_CACHE_SLINSIDE | SLAB_CACHE_MAGDEFERRED);
885
        NULL, NULL, SLAB_CACHE_SLINSIDE | SLAB_CACHE_MAGDEFERRED);
845
 
886
 
846
    /* Initialize structures for malloc */
887
    /* Initialize structures for malloc */
847
    for (i=0, size=(1<<SLAB_MIN_MALLOC_W);
888
    for (i = 0, size = (1 << SLAB_MIN_MALLOC_W);
848
         i < (SLAB_MAX_MALLOC_W-SLAB_MIN_MALLOC_W+1);
889
        i < (SLAB_MAX_MALLOC_W - SLAB_MIN_MALLOC_W + 1);
849
         i++, size <<= 1) {
890
        i++, size <<= 1) {
850
        malloc_caches[i] = slab_cache_create(malloc_names[i],
891
        malloc_caches[i] = slab_cache_create(malloc_names[i], size, 0,
851
                             size, 0,
-
 
852
                             NULL,NULL, SLAB_CACHE_MAGDEFERRED);
892
            NULL, NULL, SLAB_CACHE_MAGDEFERRED);
853
    }
893
    }
854
#ifdef CONFIG_DEBUG       
894
#ifdef CONFIG_DEBUG       
855
    _slab_initialized = 1;
895
    _slab_initialized = 1;
856
#endif
896
#endif
857
}
897
}
858
 
898
 
859
/** Enable cpu_cache
899
/** Enable cpu_cache
860
 *
900
 *
861
 * Kernel calls this function, when it knows the real number of
901
 * Kernel calls this function, when it knows the real number of
862
 * processors.
902
 * processors.
863
 * Allocate slab for cpucache and enable it on all existing
903
 * Allocate slab for cpucache and enable it on all existing
864
 * slabs that are SLAB_CACHE_MAGDEFERRED
904
 * slabs that are SLAB_CACHE_MAGDEFERRED
865
 */
905
 */
866
void slab_enable_cpucache(void)
906
void slab_enable_cpucache(void)
867
{
907
{
868
    link_t *cur;
908
    link_t *cur;
869
    slab_cache_t *s;
909
    slab_cache_t *s;
870
 
910
 
871
#ifdef CONFIG_DEBUG
911
#ifdef CONFIG_DEBUG
872
    _slab_initialized = 2;
912
    _slab_initialized = 2;
873
#endif
913
#endif
874
 
914
 
875
    spinlock_lock(&slab_cache_lock);
915
    spinlock_lock(&slab_cache_lock);
876
   
916
   
877
    for (cur=slab_cache_list.next; cur != &slab_cache_list;cur=cur->next){
917
    for (cur = slab_cache_list.next; cur != &slab_cache_list;
-
 
918
        cur = cur->next){
878
        s = list_get_instance(cur, slab_cache_t, link);
919
        s = list_get_instance(cur, slab_cache_t, link);
879
        if ((s->flags & SLAB_CACHE_MAGDEFERRED) != SLAB_CACHE_MAGDEFERRED)
920
        if ((s->flags & SLAB_CACHE_MAGDEFERRED) !=
-
 
921
            SLAB_CACHE_MAGDEFERRED)
880
            continue;
922
            continue;
881
        make_magcache(s);
923
        make_magcache(s);
882
        s->flags &= ~SLAB_CACHE_MAGDEFERRED;
924
        s->flags &= ~SLAB_CACHE_MAGDEFERRED;
883
    }
925
    }
884
 
926
 
885
    spinlock_unlock(&slab_cache_lock);
927
    spinlock_unlock(&slab_cache_lock);
886
}
928
}
887
 
929
 
888
/**************************************/
930
/**************************************/
889
/* kalloc/kfree functions             */
931
/* kalloc/kfree functions             */
890
void * malloc(unsigned int size, int flags)
932
void *malloc(unsigned int size, int flags)
891
{
933
{
892
    ASSERT(_slab_initialized);
934
    ASSERT(_slab_initialized);
893
    ASSERT(size && size <= (1 << SLAB_MAX_MALLOC_W));
935
    ASSERT(size && size <= (1 << SLAB_MAX_MALLOC_W));
894
   
936
   
895
    if (size < (1 << SLAB_MIN_MALLOC_W))
937
    if (size < (1 << SLAB_MIN_MALLOC_W))
896
        size = (1 << SLAB_MIN_MALLOC_W);
938
        size = (1 << SLAB_MIN_MALLOC_W);
897
 
939
 
898
    int idx = fnzb(size - 1) - SLAB_MIN_MALLOC_W + 1;
940
    int idx = fnzb(size - 1) - SLAB_MIN_MALLOC_W + 1;
899
 
941
 
900
    return slab_alloc(malloc_caches[idx], flags);
942
    return slab_alloc(malloc_caches[idx], flags);
901
}
943
}
902
 
944
 
903
void * realloc(void *ptr, unsigned int size, int flags)
945
void *realloc(void *ptr, unsigned int size, int flags)
904
{
946
{
905
    ASSERT(_slab_initialized);
947
    ASSERT(_slab_initialized);
906
    ASSERT(size <= (1 << SLAB_MAX_MALLOC_W));
948
    ASSERT(size <= (1 << SLAB_MAX_MALLOC_W));
907
   
949
   
908
    void *new_ptr;
950
    void *new_ptr;
909
   
951
   
910
    if (size > 0) {
952
    if (size > 0) {
911
        if (size < (1 << SLAB_MIN_MALLOC_W))
953
        if (size < (1 << SLAB_MIN_MALLOC_W))
912
            size = (1 << SLAB_MIN_MALLOC_W);
954
            size = (1 << SLAB_MIN_MALLOC_W);
913
        int idx = fnzb(size - 1) - SLAB_MIN_MALLOC_W + 1;
955
        int idx = fnzb(size - 1) - SLAB_MIN_MALLOC_W + 1;
914
       
956
       
915
        new_ptr = slab_alloc(malloc_caches[idx], flags);
957
        new_ptr = slab_alloc(malloc_caches[idx], flags);
916
    } else
958
    } else
917
        new_ptr = NULL;
959
        new_ptr = NULL;
918
   
960
   
919
    if ((new_ptr != NULL) && (ptr != NULL)) {
961
    if ((new_ptr != NULL) && (ptr != NULL)) {
920
        slab_t *slab = obj2slab(ptr);
962
        slab_t *slab = obj2slab(ptr);
921
        memcpy(new_ptr, ptr, min(size, slab->cache->size));
963
        memcpy(new_ptr, ptr, min(size, slab->cache->size));
922
    }
964
    }
923
   
965
   
924
    if (ptr != NULL)
966
    if (ptr != NULL)
925
        free(ptr);
967
        free(ptr);
926
   
968
   
927
    return new_ptr;
969
    return new_ptr;
928
}
970
}
929
 
971
 
930
void free(void *ptr)
972
void free(void *ptr)
931
{
973
{
932
    if (!ptr)
974
    if (!ptr)
933
        return;
975
        return;
934
 
976
 
935
    slab_t *slab = obj2slab(ptr);
977
    slab_t *slab = obj2slab(ptr);
936
    _slab_free(slab->cache, ptr, slab);
978
    _slab_free(slab->cache, ptr, slab);
937
}
979
}
938
 
980
 
939
/** @}
981
/** @}
940
 */
982
 */
941
 
983