Subversion Repositories HelenOS

Rev

Rev 2124 | Blame | Compare with Previous | Last modification | View Log | Download | RSS feed

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