Subversion Repositories HelenOS

Rev

Rev 3343 | 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. #include <arch/asm.h>
  116.  
  117. SPINLOCK_INITIALIZE(slab_cache_lock);
  118. static LIST_INITIALIZE(slab_cache_list);
  119.  
  120. /** Magazine cache */
  121. static slab_cache_t mag_cache;
  122. /** Cache for cache descriptors */
  123. static slab_cache_t slab_cache_cache;
  124. /** Cache for external slab descriptors
  125.  * This time we want per-cpu cache, so do not make it static
  126.  * - using slab for internal slab structures will not deadlock,
  127.  *   as all slab structures are 'small' - control structures of
  128.  *   their caches do not require further allocation
  129.  */
  130. static slab_cache_t *slab_extern_cache;
  131. /** Caches for malloc */
  132. static slab_cache_t *malloc_caches[SLAB_MAX_MALLOC_W - SLAB_MIN_MALLOC_W + 1];
  133. char *malloc_names[] =  {
  134.     "malloc-16",
  135.     "malloc-32",
  136.     "malloc-64",
  137.     "malloc-128",
  138.     "malloc-256",
  139.     "malloc-512",
  140.     "malloc-1K",
  141.     "malloc-2K",
  142.     "malloc-4K",
  143.     "malloc-8K",
  144.     "malloc-16K",
  145.     "malloc-32K",
  146.     "malloc-64K",
  147.     "malloc-128K",
  148.     "malloc-256K"
  149. };
  150.  
  151. /** Slab descriptor */
  152. typedef struct {
  153.     slab_cache_t *cache;    /**< Pointer to parent cache. */
  154.     link_t link;        /**< List of full/partial slabs. */
  155.     void *start;        /**< Start address of first available item. */
  156.     count_t available;  /**< Count of available items in this slab. */
  157.     index_t nextavail;  /**< The index of next available item. */
  158. } slab_t;
  159.  
  160. #ifdef CONFIG_DEBUG
  161. static int _slab_initialized = 0;
  162. #endif
  163.  
  164. /**************************************/
  165. /* Slab allocation functions          */
  166.  
  167. /**
  168.  * Allocate frames for slab space and initialize
  169.  *
  170.  */
  171. static slab_t *slab_space_alloc(slab_cache_t *cache, int flags)
  172. {
  173.     void *data;
  174.     slab_t *slab;
  175.     size_t fsize;
  176.     unsigned int i;
  177.     unsigned int zone = 0;
  178.    
  179.     data = frame_alloc_generic(cache->order, FRAME_KA | flags, &zone);
  180.     if (!data) {
  181.         return NULL;
  182.     }
  183.     if (!(cache->flags & SLAB_CACHE_SLINSIDE)) {
  184.         slab = slab_alloc(slab_extern_cache, flags);
  185.         if (!slab) {
  186.             frame_free(KA2PA(data));
  187.             return NULL;
  188.         }
  189.     } else {
  190.         fsize = (PAGE_SIZE << cache->order);
  191.         slab = data + fsize - sizeof(*slab);
  192.     }
  193.    
  194.     /* Fill in slab structures */
  195.     for (i = 0; i < ((unsigned int) 1 << cache->order); i++)
  196.         frame_set_parent(ADDR2PFN(KA2PA(data)) + i, slab, zone);
  197.  
  198.     slab->start = data;
  199.     slab->available = cache->objects;
  200.     slab->nextavail = 0;
  201.     slab->cache = cache;
  202.  
  203.     for (i = 0; i < cache->objects; i++)
  204.         *((int *) (slab->start + i*cache->size)) = i + 1;
  205.  
  206.     atomic_inc(&cache->allocated_slabs);
  207.     return slab;
  208. }
  209.  
  210. /**
  211.  * Deallocate space associated with slab
  212.  *
  213.  * @return number of freed frames
  214.  */
  215. static count_t slab_space_free(slab_cache_t *cache, slab_t *slab)
  216. {
  217.     frame_free(KA2PA(slab->start));
  218.     if (! (cache->flags & SLAB_CACHE_SLINSIDE))
  219.         slab_free(slab_extern_cache, slab);
  220.  
  221.     atomic_dec(&cache->allocated_slabs);
  222.    
  223.     return 1 << cache->order;
  224. }
  225.  
  226. /** Map object to slab structure */
  227. static slab_t * obj2slab(void *obj)
  228. {
  229.     return (slab_t *) frame_get_parent(ADDR2PFN(KA2PA(obj)), 0);
  230. }
  231.  
  232. /**************************************/
  233. /* Slab functions */
  234.  
  235.  
  236. /**
  237.  * Return object to slab and call a destructor
  238.  *
  239.  * @param slab If the caller knows directly slab of the object, otherwise NULL
  240.  *
  241.  * @return Number of freed pages
  242.  */
  243. static count_t slab_obj_destroy(slab_cache_t *cache, void *obj, 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,
  305.             link);
  306.         list_remove(&slab->link);
  307.     }
  308.     obj = slab->start + slab->nextavail * cache->size;
  309.     slab->nextavail = *((int *)obj);
  310.     slab->available--;
  311.  
  312.     if (!slab->available)
  313.         list_prepend(&slab->link, &cache->full_slabs);
  314.     else
  315.         list_prepend(&slab->link, &cache->partial_slabs);
  316.  
  317.     spinlock_unlock(&cache->slablock);
  318.  
  319.     if (cache->constructor && cache->constructor(obj, flags)) {
  320.         /* Bad, bad, construction failed */
  321.         slab_obj_destroy(cache, obj, slab);
  322.         return NULL;
  323.     }
  324.     return obj;
  325. }
  326.  
  327. /**************************************/
  328. /* CPU-Cache slab functions */
  329.  
  330. /**
  331.  * Finds a full magazine in cache, takes it from list
  332.  * and returns it
  333.  *
  334.  * @param first If true, return first, else last mag
  335.  */
  336. static slab_magazine_t *get_mag_from_cache(slab_cache_t *cache, 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, slab_magazine_t *mag)
  372. {
  373.     unsigned int i;
  374.     count_t frames = 0;
  375.  
  376.     for (i = 0; i < mag->busy; i++) {
  377.         frames += slab_obj_destroy(cache, mag->objs[i], NULL);
  378.         atomic_dec(&cache->cached_objs);
  379.     }
  380.    
  381.     slab_free(&mag_cache, mag);
  382.  
  383.     return frames;
  384. }
  385.  
  386. /**
  387.  * Find full magazine, set it as current and return it
  388.  *
  389.  * Assume cpu_magazine lock is held
  390.  */
  391. static slab_magazine_t *get_full_current_mag(slab_cache_t *cache)
  392. {
  393.     slab_magazine_t *cmag, *lastmag, *newmag;
  394.  
  395.     cmag = cache->mag_cache[CPU->id].current;
  396.     lastmag = cache->mag_cache[CPU->id].last;
  397.     if (cmag) { /* First try local CPU magazines */
  398.         if (cmag->busy)
  399.             return cmag;
  400.  
  401.         if (lastmag && lastmag->busy) {
  402.             cache->mag_cache[CPU->id].current = lastmag;
  403.             cache->mag_cache[CPU->id].last = cmag;
  404.             return lastmag;
  405.         }
  406.     }
  407.     /* Local magazines are empty, import one from magazine list */
  408.     newmag = get_mag_from_cache(cache, 1);
  409.     if (!newmag)
  410.         return NULL;
  411.  
  412.     if (lastmag)
  413.         magazine_destroy(cache, lastmag);
  414.  
  415.     cache->mag_cache[CPU->id].last = cmag;
  416.     cache->mag_cache[CPU->id].current = newmag;
  417.     return newmag;
  418. }
  419.  
  420. /**
  421.  * Try to find object in CPU-cache magazines
  422.  *
  423.  * @return Pointer to object or NULL if not available
  424.  */
  425. static void *magazine_obj_get(slab_cache_t *cache)
  426. {
  427.     slab_magazine_t *mag;
  428.     void *obj;
  429.  
  430.     if (!CPU)
  431.         return NULL;
  432.  
  433.     spinlock_lock(&cache->mag_cache[CPU->id].lock);
  434.  
  435.     mag = get_full_current_mag(cache);
  436.     if (!mag) {
  437.         spinlock_unlock(&cache->mag_cache[CPU->id].lock);
  438.         return NULL;
  439.     }
  440.     obj = mag->objs[--mag->busy];
  441.     spinlock_unlock(&cache->mag_cache[CPU->id].lock);
  442.     atomic_dec(&cache->cached_objs);
  443.    
  444.     return obj;
  445. }
  446.  
  447. /**
  448.  * Assure that the current magazine is empty, return pointer to it, or NULL if
  449.  * no empty magazine is available and cannot be allocated
  450.  *
  451.  * Assume mag_cache[CPU->id].lock is held
  452.  *
  453.  * We have 2 magazines bound to processor.
  454.  * First try the current.
  455.  *  If full, try the last.
  456.  *   If full, put to magazines list.
  457.  *   allocate new, exchange last & current
  458.  *
  459.  */
  460. static slab_magazine_t *make_empty_current_mag(slab_cache_t *cache)
  461. {
  462.     slab_magazine_t *cmag,*lastmag,*newmag;
  463.  
  464.     cmag = cache->mag_cache[CPU->id].current;
  465.     lastmag = cache->mag_cache[CPU->id].last;
  466.  
  467.     if (cmag) {
  468.         if (cmag->busy < cmag->size)
  469.             return cmag;
  470.         if (lastmag && lastmag->busy < lastmag->size) {
  471.             cache->mag_cache[CPU->id].last = cmag;
  472.             cache->mag_cache[CPU->id].current = lastmag;
  473.             return lastmag;
  474.         }
  475.     }
  476.     /* current | last are full | nonexistent, allocate new */
  477.     /* We do not want to sleep just because of caching */
  478.     /* Especially we do not want reclaiming to start, as
  479.      * this would deadlock */
  480.     newmag = slab_alloc(&mag_cache, FRAME_ATOMIC | FRAME_NO_RECLAIM);
  481.     if (!newmag)
  482.         return NULL;
  483.     newmag->size = SLAB_MAG_SIZE;
  484.     newmag->busy = 0;
  485.  
  486.     /* Flush last to magazine list */
  487.     if (lastmag)
  488.         put_mag_to_cache(cache, lastmag);
  489.  
  490.     /* Move current as last, save new as current */
  491.     cache->mag_cache[CPU->id].last = cmag; 
  492.     cache->mag_cache[CPU->id].current = newmag;
  493.  
  494.     return newmag;
  495. }
  496.  
  497. /**
  498.  * Put object into CPU-cache magazine
  499.  *
  500.  * @return 0 - success, -1 - could not get memory
  501.  */
  502. static int magazine_obj_put(slab_cache_t *cache, void *obj)
  503. {
  504.     slab_magazine_t *mag;
  505.  
  506.     if (!CPU)
  507.         return -1;
  508.  
  509.     spinlock_lock(&cache->mag_cache[CPU->id].lock);
  510.  
  511.     mag = make_empty_current_mag(cache);
  512.     if (!mag) {
  513.         spinlock_unlock(&cache->mag_cache[CPU->id].lock);
  514.         return -1;
  515.     }
  516.    
  517.     mag->objs[mag->busy++] = obj;
  518.  
  519.     spinlock_unlock(&cache->mag_cache[CPU->id].lock);
  520.     atomic_inc(&cache->cached_objs);
  521.     return 0;
  522. }
  523.  
  524.  
  525. /**************************************/
  526. /* Slab cache functions */
  527.  
  528. /** Return number of objects that fit in certain cache size */
  529. static unsigned int comp_objects(slab_cache_t *cache)
  530. {
  531.     if (cache->flags & SLAB_CACHE_SLINSIDE)
  532.         return ((PAGE_SIZE << cache->order) - sizeof(slab_t)) /
  533.             cache->size;
  534.     else
  535.         return (PAGE_SIZE << cache->order) / cache->size;
  536. }
  537.  
  538. /** Return wasted space in slab */
  539. static unsigned int badness(slab_cache_t *cache)
  540. {
  541.     unsigned int objects;
  542.     unsigned 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.     unsigned int i;
  557.    
  558.     ASSERT(_slab_initialized >= 2);
  559.  
  560.     cache->mag_cache = malloc(sizeof(slab_mag_cache_t) * config.cpu_count,
  561.         0);
  562.     for (i = 0; i < config.cpu_count; i++) {
  563.         memsetb(&cache->mag_cache[i], 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, char *name, size_t size, size_t align,
  572.     int (*constructor)(void *obj, int kmflag), int (*destructor)(void *obj),
  573.     int flags)
  574. {
  575.     int pages;
  576.     ipl_t ipl;
  577.  
  578.     memsetb(cache, sizeof(*cache), 0);
  579.     cache->name = name;
  580.  
  581.     if (align < sizeof(unative_t))
  582.         align = sizeof(unative_t);
  583.     size = ALIGN_UP(size, align);
  584.        
  585.     cache->size = size;
  586.  
  587.     cache->constructor = constructor;
  588.     cache->destructor = destructor;
  589.     cache->flags = flags;
  590.  
  591.     list_initialize(&cache->full_slabs);
  592.     list_initialize(&cache->partial_slabs);
  593.     list_initialize(&cache->magazines);
  594.     spinlock_initialize(&cache->slablock, "slab_lock");
  595.     spinlock_initialize(&cache->maglock, "slab_maglock");
  596.     if (!(cache->flags & SLAB_CACHE_NOMAGAZINE))
  597.         make_magcache(cache);
  598.  
  599.     /* Compute slab sizes, object counts in slabs etc. */
  600.     if (cache->size < SLAB_INSIDE_SIZE)
  601.         cache->flags |= SLAB_CACHE_SLINSIDE;
  602.  
  603.     /* Minimum slab order */
  604.     pages = SIZE2FRAMES(cache->size);
  605.     /* We need the 2^order >= pages */
  606.     if (pages == 1)
  607.         cache->order = 0;
  608.     else
  609.         cache->order = fnzb(pages - 1) + 1;
  610.  
  611.     while (badness(cache) > SLAB_MAX_BADNESS(cache)) {
  612.         cache->order += 1;
  613.     }
  614.     cache->objects = comp_objects(cache);
  615.     /* If info fits in, put it inside */
  616.     if (badness(cache) > sizeof(slab_t))
  617.         cache->flags |= SLAB_CACHE_SLINSIDE;
  618.  
  619.     /* Add cache to cache list */
  620.     ipl = interrupts_disable();
  621.     spinlock_lock(&slab_cache_lock);
  622.  
  623.     list_append(&cache->link, &slab_cache_list);
  624.  
  625.     spinlock_unlock(&slab_cache_lock);
  626.     interrupts_restore(ipl);
  627. }
  628.  
  629. /** Create slab cache  */
  630. slab_cache_t *
  631. slab_cache_create(char *name, size_t size, size_t align,
  632.     int (*constructor)(void *obj, int kmflag), int (*destructor)(void *obj),
  633.     int flags)
  634. {
  635.     slab_cache_t *cache;
  636.  
  637.     cache = slab_alloc(&slab_cache_cache, 0);
  638.     _slab_cache_create(cache, name, size, align, constructor, destructor,
  639.         flags);
  640.     return cache;
  641. }
  642.  
  643. /**
  644.  * Reclaim space occupied by objects that are already free
  645.  *
  646.  * @param flags If contains SLAB_RECLAIM_ALL, do aggressive freeing
  647.  * @return Number of freed pages
  648.  */
  649. static count_t _slab_reclaim(slab_cache_t *cache, int flags)
  650. {
  651.     unsigned int i;
  652.     slab_magazine_t *mag;
  653.     count_t frames = 0;
  654.     int magcount;
  655.    
  656.     if (cache->flags & SLAB_CACHE_NOMAGAZINE)
  657.         return 0; /* Nothing to do */
  658.  
  659.     /* We count up to original magazine count to avoid
  660.      * endless loop
  661.      */
  662.     magcount = atomic_get(&cache->magazine_counter);
  663.     while (magcount-- && (mag=get_mag_from_cache(cache, 0))) {
  664.         frames += magazine_destroy(cache,mag);
  665.         if (!(flags & SLAB_RECLAIM_ALL) && frames)
  666.             break;
  667.     }
  668.    
  669.     if (flags & SLAB_RECLAIM_ALL) {
  670.         /* Free cpu-bound magazines */
  671.         /* Destroy CPU magazines */
  672.         for (i = 0; i < config.cpu_count; i++) {
  673.             spinlock_lock(&cache->mag_cache[i].lock);
  674.  
  675.             mag = cache->mag_cache[i].current;
  676.             if (mag)
  677.                 frames += magazine_destroy(cache, mag);
  678.             cache->mag_cache[i].current = NULL;
  679.            
  680.             mag = cache->mag_cache[i].last;
  681.             if (mag)
  682.                 frames += magazine_destroy(cache, mag);
  683.             cache->mag_cache[i].last = NULL;
  684.  
  685.             spinlock_unlock(&cache->mag_cache[i].lock);
  686.         }
  687.     }
  688.  
  689.     return frames;
  690. }
  691.  
  692. /** Check that there are no slabs and remove cache from system  */
  693. void slab_cache_destroy(slab_cache_t *cache)
  694. {
  695.     ipl_t ipl;
  696.  
  697.     /* First remove cache from link, so that we don't need
  698.      * to disable interrupts later
  699.      */
  700.  
  701.     ipl = interrupts_disable();
  702.     spinlock_lock(&slab_cache_lock);
  703.  
  704.     list_remove(&cache->link);
  705.  
  706.     spinlock_unlock(&slab_cache_lock);
  707.     interrupts_restore(ipl);
  708.  
  709.     /* Do not lock anything, we assume the software is correct and
  710.      * does not touch the cache when it decides to destroy it */
  711.    
  712.     /* Destroy all magazines */
  713.     _slab_reclaim(cache, SLAB_RECLAIM_ALL);
  714.  
  715.     /* All slabs must be empty */
  716.     if (!list_empty(&cache->full_slabs) ||
  717.         !list_empty(&cache->partial_slabs))
  718.         panic("Destroying cache that is not empty.");
  719.  
  720.     if (!(cache->flags & SLAB_CACHE_NOMAGAZINE))
  721.         free(cache->mag_cache);
  722.     slab_free(&slab_cache_cache, cache);
  723. }
  724.  
  725. /** Allocate new object from cache - if no flags given, always returns memory */
  726. void *slab_alloc(slab_cache_t *cache, int flags)
  727. {
  728.     ipl_t ipl;
  729.     void *result = NULL;
  730.    
  731.     /* Disable interrupts to avoid deadlocks with interrupt handlers */
  732.     ipl = interrupts_disable();
  733.  
  734.     if (!(cache->flags & SLAB_CACHE_NOMAGAZINE)) {
  735.         result = magazine_obj_get(cache);
  736.     }
  737.     if (!result)
  738.         result = slab_obj_create(cache, flags);
  739.  
  740.     interrupts_restore(ipl);
  741.  
  742.     if (result)
  743.         atomic_inc(&cache->allocated_objs);
  744.  
  745.     return result;
  746. }
  747.  
  748. /** Return object to cache, use slab if known  */
  749. static void _slab_free(slab_cache_t *cache, void *obj, slab_t *slab)
  750. {
  751.     ipl_t ipl;
  752.  
  753.     ipl = interrupts_disable();
  754.  
  755.     if ((cache->flags & SLAB_CACHE_NOMAGAZINE) ||
  756.         magazine_obj_put(cache, obj)) {
  757.         slab_obj_destroy(cache, obj, slab);
  758.  
  759.     }
  760.     interrupts_restore(ipl);
  761.     atomic_dec(&cache->allocated_objs);
  762. }
  763.  
  764. /** Return slab object to cache */
  765. void slab_free(slab_cache_t *cache, void *obj)
  766. {
  767.     _slab_free(cache, obj, NULL);
  768. }
  769.  
  770. /* Go through all caches and reclaim what is possible */
  771. count_t slab_reclaim(int flags)
  772. {
  773.     slab_cache_t *cache;
  774.     link_t *cur;
  775.     count_t frames = 0;
  776.  
  777.     spinlock_lock(&slab_cache_lock);
  778.  
  779.     /* TODO: Add assert, that interrupts are disabled, otherwise
  780.      * memory allocation from interrupts can deadlock.
  781.      */
  782.  
  783.     for (cur = slab_cache_list.next; cur != &slab_cache_list;
  784.         cur = cur->next) {
  785.         cache = list_get_instance(cur, slab_cache_t, link);
  786.         frames += _slab_reclaim(cache, flags);
  787.     }
  788.  
  789.     spinlock_unlock(&slab_cache_lock);
  790.  
  791.     return frames;
  792. }
  793.  
  794.  
  795. /* Print list of slabs */
  796. void slab_print_list(void)
  797. {
  798.     int skip = 0;
  799.  
  800.     printf("slab name        size     pages  obj/pg slabs  cached allocated"
  801.         " ctl\n");
  802.     printf("---------------- -------- ------ ------ ------ ------ ---------"
  803.         " ---\n");
  804.  
  805.     while (true) {
  806.         slab_cache_t *cache;
  807.         link_t *cur;
  808.         ipl_t ipl;
  809.         int i;
  810.  
  811.         /*
  812.          * We must not hold the slab_cache_lock spinlock when printing
  813.          * the statistics. Otherwise we can easily deadlock if the print
  814.          * needs to allocate memory.
  815.          *
  816.          * Therefore, we walk through the slab cache list, skipping some
  817.          * amount of already processed caches during each iteration and
  818.          * gathering statistics about the first unprocessed cache. For
  819.          * the sake of printing the statistics, we realese the
  820.          * slab_cache_lock and reacquire it afterwards. Then the walk
  821.          * starts again.
  822.          *
  823.          * This limits both the efficiency and also accuracy of the
  824.          * obtained statistics. The efficiency is decreased because the
  825.          * time complexity of the algorithm is quadratic instead of
  826.          * linear. The accuracy is impacted because we drop the lock
  827.          * after processing one cache. If there is someone else
  828.          * manipulating the cache list, we might omit an arbitrary
  829.          * number of caches or process one cache multiple times.
  830.          * However, we don't bleed for this algorithm for it is only
  831.          * statistics.
  832.          */
  833.  
  834.         ipl = interrupts_disable();
  835.         spinlock_lock(&slab_cache_lock);
  836.  
  837.         for (i = 0, cur = slab_cache_list.next;
  838.             i < skip && cur != &slab_cache_list;
  839.             i++, cur = cur->next)
  840.             ;
  841.  
  842.         if (cur == &slab_cache_list) {
  843.             spinlock_unlock(&slab_cache_lock);
  844.             interrupts_restore(ipl);
  845.             break;
  846.         }
  847.  
  848.         skip++;
  849.  
  850.         cache = list_get_instance(cur, slab_cache_t, link);
  851.  
  852.         char *name = cache->name;
  853.         uint8_t order = cache->order;
  854.         size_t size = cache->size;
  855.         unsigned int objects = cache->objects;
  856.         long allocated_slabs = atomic_get(&cache->allocated_slabs);
  857.         long cached_objs = atomic_get(&cache->cached_objs);
  858.         long allocated_objs = atomic_get(&cache->allocated_objs);
  859.         int flags = cache->flags;
  860.        
  861.         spinlock_unlock(&slab_cache_lock);
  862.         interrupts_restore(ipl);
  863.        
  864.         printf("%-16s %8" PRIs " %6d %6u %6ld %6ld %9ld %-3s\n",
  865.             name, size, (1 << order), objects, allocated_slabs,
  866.             cached_objs, allocated_objs,
  867.             flags & SLAB_CACHE_SLINSIDE ? "in" : "out");
  868.     }
  869. }
  870.  
  871. void slab_cache_init(void)
  872. {
  873.     int i, size;
  874.  
  875.     /* Initialize magazine cache */
  876.     _slab_cache_create(&mag_cache, "slab_magazine",
  877.         sizeof(slab_magazine_t) + SLAB_MAG_SIZE * sizeof(void*),
  878.         sizeof(uintptr_t), NULL, NULL, SLAB_CACHE_NOMAGAZINE |
  879.         SLAB_CACHE_SLINSIDE);
  880.     /* Initialize slab_cache cache */
  881.     _slab_cache_create(&slab_cache_cache, "slab_cache",
  882.         sizeof(slab_cache_cache), sizeof(uintptr_t), NULL, NULL,
  883.         SLAB_CACHE_NOMAGAZINE | SLAB_CACHE_SLINSIDE);
  884.     /* Initialize external slab cache */
  885.     slab_extern_cache = slab_cache_create("slab_extern", sizeof(slab_t), 0,
  886.         NULL, NULL, SLAB_CACHE_SLINSIDE | SLAB_CACHE_MAGDEFERRED);
  887.  
  888.     /* Initialize structures for malloc */
  889.     for (i = 0, size = (1 << SLAB_MIN_MALLOC_W);
  890.         i < (SLAB_MAX_MALLOC_W - SLAB_MIN_MALLOC_W + 1);
  891.         i++, size <<= 1) {
  892.         malloc_caches[i] = slab_cache_create(malloc_names[i], size, 0,
  893.             NULL, NULL, SLAB_CACHE_MAGDEFERRED);
  894.     }
  895. #ifdef CONFIG_DEBUG      
  896.     _slab_initialized = 1;
  897. #endif
  898. }
  899.  
  900. /** Enable cpu_cache
  901.  *
  902.  * Kernel calls this function, when it knows the real number of
  903.  * processors.
  904.  * Allocate slab for cpucache and enable it on all existing
  905.  * slabs that are SLAB_CACHE_MAGDEFERRED
  906.  */
  907. void slab_enable_cpucache(void)
  908. {
  909.     link_t *cur;
  910.     slab_cache_t *s;
  911.  
  912. #ifdef CONFIG_DEBUG
  913.     _slab_initialized = 2;
  914. #endif
  915.  
  916.     spinlock_lock(&slab_cache_lock);
  917.    
  918.     for (cur = slab_cache_list.next; cur != &slab_cache_list;
  919.         cur = cur->next){
  920.         s = list_get_instance(cur, slab_cache_t, link);
  921.         if ((s->flags & SLAB_CACHE_MAGDEFERRED) !=
  922.             SLAB_CACHE_MAGDEFERRED)
  923.             continue;
  924.         make_magcache(s);
  925.         s->flags &= ~SLAB_CACHE_MAGDEFERRED;
  926.     }
  927.  
  928.     spinlock_unlock(&slab_cache_lock);
  929. }
  930.  
  931. /**************************************/
  932. /* kalloc/kfree functions             */
  933. void *malloc(unsigned int size, int flags)
  934. {
  935.     ASSERT(_slab_initialized);
  936.     ASSERT(size && size <= (1 << SLAB_MAX_MALLOC_W));
  937.    
  938.     if (size < (1 << SLAB_MIN_MALLOC_W))
  939.         size = (1 << SLAB_MIN_MALLOC_W);
  940.  
  941.     int idx = fnzb(size - 1) - SLAB_MIN_MALLOC_W + 1;
  942.  
  943.     return slab_alloc(malloc_caches[idx], flags);
  944. }
  945.  
  946. void *realloc(void *ptr, unsigned int size, int flags)
  947. {
  948.     ASSERT(_slab_initialized);
  949.     ASSERT(size <= (1 << SLAB_MAX_MALLOC_W));
  950.    
  951.     void *new_ptr;
  952.    
  953.     if (size > 0) {
  954.         if (size < (1 << SLAB_MIN_MALLOC_W))
  955.             size = (1 << SLAB_MIN_MALLOC_W);
  956.         int idx = fnzb(size - 1) - SLAB_MIN_MALLOC_W + 1;
  957.        
  958.         new_ptr = slab_alloc(malloc_caches[idx], flags);
  959.     } else
  960.         new_ptr = NULL;
  961.    
  962.     if ((new_ptr != NULL) && (ptr != NULL)) {
  963.         slab_t *slab = obj2slab(ptr);
  964.         memcpy(new_ptr, ptr, min(size, slab->cache->size));
  965.     }
  966.    
  967.     if (ptr != NULL)
  968.         free(ptr);
  969.    
  970.     return new_ptr;
  971. }
  972.  
  973. void free(void *ptr)
  974. {
  975.     if (!ptr)
  976.         return;
  977.  
  978.     slab_t *slab = obj2slab(ptr);
  979.     _slab_free(slab->cache, ptr, slab);
  980. }
  981.  
  982. /** @}
  983.  */
  984.