Subversion Repositories HelenOS

Rev

Rev 3022 | Rev 4201 | Go to most recent revision | 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.     unsigned int i;
  176.     count_t 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 < ((unsigned int) 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, slab_t *slab)
  243. {
  244.     int freed = 0;
  245.  
  246.     if (!slab)
  247.         slab = obj2slab(obj);
  248.  
  249.     ASSERT(slab->cache == cache);
  250.  
  251.     if (cache->destructor)
  252.         freed = cache->destructor(obj);
  253.    
  254.     spinlock_lock(&cache->slablock);
  255.     ASSERT(slab->available < cache->objects);
  256.  
  257.     *((int *)obj) = slab->nextavail;
  258.     slab->nextavail = (obj - slab->start) / cache->size;
  259.     slab->available++;
  260.  
  261.     /* Move it to correct list */
  262.     if (slab->available == cache->objects) {
  263.         /* Free associated memory */
  264.         list_remove(&slab->link);
  265.         spinlock_unlock(&cache->slablock);
  266.  
  267.         return freed + slab_space_free(cache, slab);
  268.  
  269.     } else if (slab->available == 1) {
  270.         /* It was in full, move to partial */
  271.         list_remove(&slab->link);
  272.         list_prepend(&slab->link, &cache->partial_slabs);
  273.     }
  274.     spinlock_unlock(&cache->slablock);
  275.     return freed;
  276. }
  277.  
  278. /**
  279.  * Take new object from slab or create new if needed
  280.  *
  281.  * @return Object address or null
  282.  */
  283. static void *slab_obj_create(slab_cache_t *cache, int flags)
  284. {
  285.     slab_t *slab;
  286.     void *obj;
  287.  
  288.     spinlock_lock(&cache->slablock);
  289.  
  290.     if (list_empty(&cache->partial_slabs)) {
  291.         /* Allow recursion and reclaiming
  292.          * - this should work, as the slab control structures
  293.          *   are small and do not need to allocate with anything
  294.          *   other than frame_alloc when they are allocating,
  295.          *   that's why we should get recursion at most 1-level deep
  296.          */
  297.         spinlock_unlock(&cache->slablock);
  298.         slab = slab_space_alloc(cache, flags);
  299.         if (!slab)
  300.             return NULL;
  301.         spinlock_lock(&cache->slablock);
  302.     } else {
  303.         slab = list_get_instance(cache->partial_slabs.next, slab_t,
  304.             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, int first)
  336. {
  337.     slab_magazine_t *mag = NULL;
  338.     link_t *cur;
  339.  
  340.     spinlock_lock(&cache->maglock);
  341.     if (!list_empty(&cache->magazines)) {
  342.         if (first)
  343.             cur = cache->magazines.next;
  344.         else
  345.             cur = cache->magazines.prev;
  346.         mag = list_get_instance(cur, slab_magazine_t, link);
  347.         list_remove(&mag->link);
  348.         atomic_dec(&cache->magazine_counter);
  349.     }
  350.     spinlock_unlock(&cache->maglock);
  351.     return mag;
  352. }
  353.  
  354. /** Prepend magazine to magazine list in cache */
  355. static void put_mag_to_cache(slab_cache_t *cache, slab_magazine_t *mag)
  356. {
  357.     spinlock_lock(&cache->maglock);
  358.  
  359.     list_prepend(&mag->link, &cache->magazines);
  360.     atomic_inc(&cache->magazine_counter);
  361.    
  362.     spinlock_unlock(&cache->maglock);
  363. }
  364.  
  365. /**
  366.  * Free all objects in magazine and free memory associated with magazine
  367.  *
  368.  * @return Number of freed pages
  369.  */
  370. static count_t magazine_destroy(slab_cache_t *cache, slab_magazine_t *mag)
  371. {
  372.     unsigned int i;
  373.     count_t frames = 0;
  374.  
  375.     for (i = 0; i < mag->busy; i++) {
  376.         frames += slab_obj_destroy(cache, mag->objs[i], NULL);
  377.         atomic_dec(&cache->cached_objs);
  378.     }
  379.    
  380.     slab_free(&mag_cache, mag);
  381.  
  382.     return frames;
  383. }
  384.  
  385. /**
  386.  * Find full magazine, set it as current and return it
  387.  *
  388.  * Assume cpu_magazine lock is held
  389.  */
  390. static slab_magazine_t *get_full_current_mag(slab_cache_t *cache)
  391. {
  392.     slab_magazine_t *cmag, *lastmag, *newmag;
  393.  
  394.     cmag = cache->mag_cache[CPU->id].current;
  395.     lastmag = cache->mag_cache[CPU->id].last;
  396.     if (cmag) { /* First try local CPU magazines */
  397.         if (cmag->busy)
  398.             return cmag;
  399.  
  400.         if (lastmag && lastmag->busy) {
  401.             cache->mag_cache[CPU->id].current = lastmag;
  402.             cache->mag_cache[CPU->id].last = cmag;
  403.             return lastmag;
  404.         }
  405.     }
  406.     /* Local magazines are empty, import one from magazine list */
  407.     newmag = get_mag_from_cache(cache, 1);
  408.     if (!newmag)
  409.         return NULL;
  410.  
  411.     if (lastmag)
  412.         magazine_destroy(cache, lastmag);
  413.  
  414.     cache->mag_cache[CPU->id].last = cmag;
  415.     cache->mag_cache[CPU->id].current = newmag;
  416.     return newmag;
  417. }
  418.  
  419. /**
  420.  * Try to find object in CPU-cache magazines
  421.  *
  422.  * @return Pointer to object or NULL if not available
  423.  */
  424. static void *magazine_obj_get(slab_cache_t *cache)
  425. {
  426.     slab_magazine_t *mag;
  427.     void *obj;
  428.  
  429.     if (!CPU)
  430.         return NULL;
  431.  
  432.     spinlock_lock(&cache->mag_cache[CPU->id].lock);
  433.  
  434.     mag = get_full_current_mag(cache);
  435.     if (!mag) {
  436.         spinlock_unlock(&cache->mag_cache[CPU->id].lock);
  437.         return NULL;
  438.     }
  439.     obj = mag->objs[--mag->busy];
  440.     spinlock_unlock(&cache->mag_cache[CPU->id].lock);
  441.     atomic_dec(&cache->cached_objs);
  442.    
  443.     return obj;
  444. }
  445.  
  446. /**
  447.  * Assure that the current magazine is empty, return pointer to it, or NULL if
  448.  * no empty magazine is available and cannot be allocated
  449.  *
  450.  * Assume mag_cache[CPU->id].lock is held
  451.  *
  452.  * We have 2 magazines bound to processor.
  453.  * First try the current.
  454.  *  If full, try the last.
  455.  *   If full, put to magazines list.
  456.  *   allocate new, exchange last & current
  457.  *
  458.  */
  459. static slab_magazine_t *make_empty_current_mag(slab_cache_t *cache)
  460. {
  461.     slab_magazine_t *cmag,*lastmag,*newmag;
  462.  
  463.     cmag = cache->mag_cache[CPU->id].current;
  464.     lastmag = cache->mag_cache[CPU->id].last;
  465.  
  466.     if (cmag) {
  467.         if (cmag->busy < cmag->size)
  468.             return cmag;
  469.         if (lastmag && lastmag->busy < lastmag->size) {
  470.             cache->mag_cache[CPU->id].last = cmag;
  471.             cache->mag_cache[CPU->id].current = lastmag;
  472.             return lastmag;
  473.         }
  474.     }
  475.     /* current | last are full | nonexistent, allocate new */
  476.     /* We do not want to sleep just because of caching */
  477.     /* Especially we do not want reclaiming to start, as
  478.      * this would deadlock */
  479.     newmag = slab_alloc(&mag_cache, FRAME_ATOMIC | FRAME_NO_RECLAIM);
  480.     if (!newmag)
  481.         return NULL;
  482.     newmag->size = SLAB_MAG_SIZE;
  483.     newmag->busy = 0;
  484.  
  485.     /* Flush last to magazine list */
  486.     if (lastmag)
  487.         put_mag_to_cache(cache, lastmag);
  488.  
  489.     /* Move current as last, save new as current */
  490.     cache->mag_cache[CPU->id].last = cmag; 
  491.     cache->mag_cache[CPU->id].current = newmag;
  492.  
  493.     return newmag;
  494. }
  495.  
  496. /**
  497.  * Put object into CPU-cache magazine
  498.  *
  499.  * @return 0 - success, -1 - could not get memory
  500.  */
  501. static int magazine_obj_put(slab_cache_t *cache, void *obj)
  502. {
  503.     slab_magazine_t *mag;
  504.  
  505.     if (!CPU)
  506.         return -1;
  507.  
  508.     spinlock_lock(&cache->mag_cache[CPU->id].lock);
  509.  
  510.     mag = make_empty_current_mag(cache);
  511.     if (!mag) {
  512.         spinlock_unlock(&cache->mag_cache[CPU->id].lock);
  513.         return -1;
  514.     }
  515.    
  516.     mag->objs[mag->busy++] = obj;
  517.  
  518.     spinlock_unlock(&cache->mag_cache[CPU->id].lock);
  519.     atomic_inc(&cache->cached_objs);
  520.     return 0;
  521. }
  522.  
  523.  
  524. /**************************************/
  525. /* Slab cache functions */
  526.  
  527. /** Return number of objects that fit in certain cache size */
  528. static unsigned int comp_objects(slab_cache_t *cache)
  529. {
  530.     if (cache->flags & SLAB_CACHE_SLINSIDE)
  531.         return ((PAGE_SIZE << cache->order) - sizeof(slab_t)) /
  532.             cache->size;
  533.     else
  534.         return (PAGE_SIZE << cache->order) / cache->size;
  535. }
  536.  
  537. /** Return wasted space in slab */
  538. static unsigned int badness(slab_cache_t *cache)
  539. {
  540.     unsigned int objects;
  541.     unsigned int ssize;
  542.  
  543.     objects = comp_objects(cache);
  544.     ssize = PAGE_SIZE << cache->order;
  545.     if (cache->flags & SLAB_CACHE_SLINSIDE)
  546.         ssize -= sizeof(slab_t);
  547.     return ssize - objects * cache->size;
  548. }
  549.  
  550. /**
  551.  * Initialize mag_cache structure in slab cache
  552.  */
  553. static void make_magcache(slab_cache_t *cache)
  554. {
  555.     unsigned int i;
  556.    
  557.     ASSERT(_slab_initialized >= 2);
  558.  
  559.     cache->mag_cache = malloc(sizeof(slab_mag_cache_t) * config.cpu_count,
  560.         0);
  561.     for (i = 0; i < config.cpu_count; i++) {
  562.         memsetb(&cache->mag_cache[i], sizeof(cache->mag_cache[i]), 0);
  563.         spinlock_initialize(&cache->mag_cache[i].lock,
  564.             "slab_maglock_cpu");
  565.     }
  566. }
  567.  
  568. /** Initialize allocated memory as a slab cache */
  569. static void
  570. _slab_cache_create(slab_cache_t *cache, char *name, size_t size, size_t align,
  571.     int (*constructor)(void *obj, int kmflag), int (*destructor)(void *obj),
  572.     int flags)
  573. {
  574.     int pages;
  575.     ipl_t ipl;
  576.  
  577.     memsetb(cache, sizeof(*cache), 0);
  578.     cache->name = name;
  579.  
  580.     if (align < sizeof(unative_t))
  581.         align = sizeof(unative_t);
  582.     size = ALIGN_UP(size, align);
  583.        
  584.     cache->size = size;
  585.  
  586.     cache->constructor = constructor;
  587.     cache->destructor = destructor;
  588.     cache->flags = flags;
  589.  
  590.     list_initialize(&cache->full_slabs);
  591.     list_initialize(&cache->partial_slabs);
  592.     list_initialize(&cache->magazines);
  593.     spinlock_initialize(&cache->slablock, "slab_lock");
  594.     spinlock_initialize(&cache->maglock, "slab_maglock");
  595.     if (!(cache->flags & SLAB_CACHE_NOMAGAZINE))
  596.         make_magcache(cache);
  597.  
  598.     /* Compute slab sizes, object counts in slabs etc. */
  599.     if (cache->size < SLAB_INSIDE_SIZE)
  600.         cache->flags |= SLAB_CACHE_SLINSIDE;
  601.  
  602.     /* Minimum slab order */
  603.     pages = SIZE2FRAMES(cache->size);
  604.     /* We need the 2^order >= pages */
  605.     if (pages == 1)
  606.         cache->order = 0;
  607.     else
  608.         cache->order = fnzb(pages - 1) + 1;
  609.  
  610.     while (badness(cache) > SLAB_MAX_BADNESS(cache)) {
  611.         cache->order += 1;
  612.     }
  613.     cache->objects = comp_objects(cache);
  614.     /* If info fits in, put it inside */
  615.     if (badness(cache) > sizeof(slab_t))
  616.         cache->flags |= SLAB_CACHE_SLINSIDE;
  617.  
  618.     /* Add cache to cache list */
  619.     ipl = interrupts_disable();
  620.     spinlock_lock(&slab_cache_lock);
  621.  
  622.     list_append(&cache->link, &slab_cache_list);
  623.  
  624.     spinlock_unlock(&slab_cache_lock);
  625.     interrupts_restore(ipl);
  626. }
  627.  
  628. /** Create slab cache  */
  629. slab_cache_t *
  630. slab_cache_create(char *name, size_t size, size_t align,
  631.     int (*constructor)(void *obj, int kmflag), int (*destructor)(void *obj),
  632.     int flags)
  633. {
  634.     slab_cache_t *cache;
  635.  
  636.     cache = slab_alloc(&slab_cache_cache, 0);
  637.     _slab_cache_create(cache, name, size, align, constructor, destructor,
  638.         flags);
  639.     return cache;
  640. }
  641.  
  642. /**
  643.  * Reclaim space occupied by objects that are already free
  644.  *
  645.  * @param flags If contains SLAB_RECLAIM_ALL, do aggressive freeing
  646.  * @return Number of freed pages
  647.  */
  648. static count_t _slab_reclaim(slab_cache_t *cache, int flags)
  649. {
  650.     unsigned int i;
  651.     slab_magazine_t *mag;
  652.     count_t frames = 0;
  653.     int magcount;
  654.    
  655.     if (cache->flags & SLAB_CACHE_NOMAGAZINE)
  656.         return 0; /* Nothing to do */
  657.  
  658.     /* We count up to original magazine count to avoid
  659.      * endless loop
  660.      */
  661.     magcount = atomic_get(&cache->magazine_counter);
  662.     while (magcount-- && (mag=get_mag_from_cache(cache, 0))) {
  663.         frames += magazine_destroy(cache,mag);
  664.         if (!(flags & SLAB_RECLAIM_ALL) && frames)
  665.             break;
  666.     }
  667.    
  668.     if (flags & SLAB_RECLAIM_ALL) {
  669.         /* Free cpu-bound magazines */
  670.         /* Destroy CPU magazines */
  671.         for (i = 0; i < config.cpu_count; i++) {
  672.             spinlock_lock(&cache->mag_cache[i].lock);
  673.  
  674.             mag = cache->mag_cache[i].current;
  675.             if (mag)
  676.                 frames += magazine_destroy(cache, mag);
  677.             cache->mag_cache[i].current = NULL;
  678.            
  679.             mag = cache->mag_cache[i].last;
  680.             if (mag)
  681.                 frames += magazine_destroy(cache, mag);
  682.             cache->mag_cache[i].last = NULL;
  683.  
  684.             spinlock_unlock(&cache->mag_cache[i].lock);
  685.         }
  686.     }
  687.  
  688.     return frames;
  689. }
  690.  
  691. /** Check that there are no slabs and remove cache from system  */
  692. void slab_cache_destroy(slab_cache_t *cache)
  693. {
  694.     ipl_t ipl;
  695.  
  696.     /* First remove cache from link, so that we don't need
  697.      * to disable interrupts later
  698.      */
  699.  
  700.     ipl = interrupts_disable();
  701.     spinlock_lock(&slab_cache_lock);
  702.  
  703.     list_remove(&cache->link);
  704.  
  705.     spinlock_unlock(&slab_cache_lock);
  706.     interrupts_restore(ipl);
  707.  
  708.     /* Do not lock anything, we assume the software is correct and
  709.      * does not touch the cache when it decides to destroy it */
  710.    
  711.     /* Destroy all magazines */
  712.     _slab_reclaim(cache, SLAB_RECLAIM_ALL);
  713.  
  714.     /* All slabs must be empty */
  715.     if (!list_empty(&cache->full_slabs) ||
  716.         !list_empty(&cache->partial_slabs))
  717.         panic("Destroying cache that is not empty.");
  718.  
  719.     if (!(cache->flags & SLAB_CACHE_NOMAGAZINE))
  720.         free(cache->mag_cache);
  721.     slab_free(&slab_cache_cache, cache);
  722. }
  723.  
  724. /** Allocate new object from cache - if no flags given, always returns memory */
  725. void *slab_alloc(slab_cache_t *cache, int flags)
  726. {
  727.     ipl_t ipl;
  728.     void *result = NULL;
  729.    
  730.     /* Disable interrupts to avoid deadlocks with interrupt handlers */
  731.     ipl = interrupts_disable();
  732.  
  733.     if (!(cache->flags & SLAB_CACHE_NOMAGAZINE)) {
  734.         result = magazine_obj_get(cache);
  735.     }
  736.     if (!result)
  737.         result = slab_obj_create(cache, flags);
  738.  
  739.     interrupts_restore(ipl);
  740.  
  741.     if (result)
  742.         atomic_inc(&cache->allocated_objs);
  743.  
  744.     return result;
  745. }
  746.  
  747. /** Return object to cache, use slab if known  */
  748. static void _slab_free(slab_cache_t *cache, void *obj, slab_t *slab)
  749. {
  750.     ipl_t ipl;
  751.  
  752.     ipl = interrupts_disable();
  753.  
  754.     if ((cache->flags & SLAB_CACHE_NOMAGAZINE) ||
  755.         magazine_obj_put(cache, obj)) {
  756.         slab_obj_destroy(cache, obj, slab);
  757.  
  758.     }
  759.     interrupts_restore(ipl);
  760.     atomic_dec(&cache->allocated_objs);
  761. }
  762.  
  763. /** Return slab object to cache */
  764. void slab_free(slab_cache_t *cache, void *obj)
  765. {
  766.     _slab_free(cache, obj, NULL);
  767. }
  768.  
  769. /* Go through all caches and reclaim what is possible */
  770. count_t slab_reclaim(int flags)
  771. {
  772.     slab_cache_t *cache;
  773.     link_t *cur;
  774.     count_t frames = 0;
  775.  
  776.     spinlock_lock(&slab_cache_lock);
  777.  
  778.     /* TODO: Add assert, that interrupts are disabled, otherwise
  779.      * memory allocation from interrupts can deadlock.
  780.      */
  781.  
  782.     for (cur = slab_cache_list.next; cur != &slab_cache_list;
  783.         cur = cur->next) {
  784.         cache = list_get_instance(cur, slab_cache_t, link);
  785.         frames += _slab_reclaim(cache, flags);
  786.     }
  787.  
  788.     spinlock_unlock(&slab_cache_lock);
  789.  
  790.     return frames;
  791. }
  792.  
  793.  
  794. /* Print list of slabs */
  795. void slab_print_list(void)
  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) {
  805.         slab_cache_t *cache;
  806.         link_t *cur;
  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.          */
  832.  
  833.         ipl = interrupts_disable();
  834.         spinlock_lock(&slab_cache_lock);
  835.  
  836.         for (i = 0, cur = slab_cache_list.next;
  837.             i < skip && cur != &slab_cache_list;
  838.             i++, cur = cur->next)
  839.             ;
  840.  
  841.         if (cur == &slab_cache_list) {
  842.             spinlock_unlock(&slab_cache_lock);
  843.             interrupts_restore(ipl);
  844.             break;
  845.         }
  846.  
  847.         skip++;
  848.  
  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);
  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,
  866.             flags & SLAB_CACHE_SLINSIDE ? "in" : "out");
  867.     }
  868. }
  869.  
  870. void slab_cache_init(void)
  871. {
  872.     int i, size;
  873.  
  874.     /* Initialize magazine cache */
  875.     _slab_cache_create(&mag_cache, "slab_magazine",
  876.         sizeof(slab_magazine_t) + SLAB_MAG_SIZE * sizeof(void*),
  877.         sizeof(uintptr_t), NULL, NULL, SLAB_CACHE_NOMAGAZINE |
  878.         SLAB_CACHE_SLINSIDE);
  879.     /* Initialize slab_cache cache */
  880.     _slab_cache_create(&slab_cache_cache, "slab_cache",
  881.         sizeof(slab_cache_cache), sizeof(uintptr_t), NULL, NULL,
  882.         SLAB_CACHE_NOMAGAZINE | SLAB_CACHE_SLINSIDE);
  883.     /* Initialize external slab cache */
  884.     slab_extern_cache = slab_cache_create("slab_extern", sizeof(slab_t), 0,
  885.         NULL, NULL, SLAB_CACHE_SLINSIDE | SLAB_CACHE_MAGDEFERRED);
  886.  
  887.     /* Initialize structures for malloc */
  888.     for (i = 0, size = (1 << SLAB_MIN_MALLOC_W);
  889.         i < (SLAB_MAX_MALLOC_W - SLAB_MIN_MALLOC_W + 1);
  890.         i++, size <<= 1) {
  891.         malloc_caches[i] = slab_cache_create(malloc_names[i], size, 0,
  892.             NULL, NULL, SLAB_CACHE_MAGDEFERRED);
  893.     }
  894. #ifdef CONFIG_DEBUG      
  895.     _slab_initialized = 1;
  896. #endif
  897. }
  898.  
  899. /** Enable cpu_cache
  900.  *
  901.  * Kernel calls this function, when it knows the real number of
  902.  * processors.
  903.  * Allocate slab for cpucache and enable it on all existing
  904.  * slabs that are SLAB_CACHE_MAGDEFERRED
  905.  */
  906. void slab_enable_cpucache(void)
  907. {
  908.     link_t *cur;
  909.     slab_cache_t *s;
  910.  
  911. #ifdef CONFIG_DEBUG
  912.     _slab_initialized = 2;
  913. #endif
  914.  
  915.     spinlock_lock(&slab_cache_lock);
  916.    
  917.     for (cur = slab_cache_list.next; cur != &slab_cache_list;
  918.         cur = cur->next){
  919.         s = list_get_instance(cur, slab_cache_t, link);
  920.         if ((s->flags & SLAB_CACHE_MAGDEFERRED) !=
  921.             SLAB_CACHE_MAGDEFERRED)
  922.             continue;
  923.         make_magcache(s);
  924.         s->flags &= ~SLAB_CACHE_MAGDEFERRED;
  925.     }
  926.  
  927.     spinlock_unlock(&slab_cache_lock);
  928. }
  929.  
  930. /**************************************/
  931. /* kalloc/kfree functions             */
  932. void *malloc(unsigned int size, int flags)
  933. {
  934.     ASSERT(_slab_initialized);
  935.     ASSERT(size && size <= (1 << SLAB_MAX_MALLOC_W));
  936.    
  937.     if (size < (1 << SLAB_MIN_MALLOC_W))
  938.         size = (1 << SLAB_MIN_MALLOC_W);
  939.  
  940.     int idx = fnzb(size - 1) - SLAB_MIN_MALLOC_W + 1;
  941.  
  942.     return slab_alloc(malloc_caches[idx], flags);
  943. }
  944.  
  945. void *realloc(void *ptr, unsigned int size, int flags)
  946. {
  947.     ASSERT(_slab_initialized);
  948.     ASSERT(size <= (1 << SLAB_MAX_MALLOC_W));
  949.    
  950.     void *new_ptr;
  951.    
  952.     if (size > 0) {
  953.         if (size < (1 << SLAB_MIN_MALLOC_W))
  954.             size = (1 << SLAB_MIN_MALLOC_W);
  955.         int idx = fnzb(size - 1) - SLAB_MIN_MALLOC_W + 1;
  956.        
  957.         new_ptr = slab_alloc(malloc_caches[idx], flags);
  958.     } else
  959.         new_ptr = NULL;
  960.    
  961.     if ((new_ptr != NULL) && (ptr != NULL)) {
  962.         slab_t *slab = obj2slab(ptr);
  963.         memcpy(new_ptr, ptr, min(size, slab->cache->size));
  964.     }
  965.    
  966.     if (ptr != NULL)
  967.         free(ptr);
  968.    
  969.     return new_ptr;
  970. }
  971.  
  972. void free(void *ptr)
  973. {
  974.     if (!ptr)
  975.         return;
  976.  
  977.     slab_t *slab = obj2slab(ptr);
  978.     _slab_free(slab->cache, ptr, slab);
  979. }
  980.  
  981. /** @}
  982.  */
  983.