Subversion Repositories HelenOS-historic

Rev

Rev 1757 | 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.  
  115. SPINLOCK_INITIALIZE(slab_cache_lock);
  116. static LIST_INITIALIZE(slab_cache_list);
  117.  
  118. /** Magazine cache */
  119. static slab_cache_t mag_cache;
  120. /** Cache for cache descriptors */
  121. static slab_cache_t slab_cache_cache;
  122. /** Cache for external slab descriptors
  123.  * This time we want per-cpu cache, so do not make it static
  124.  * - using slab for internal slab structures will not deadlock,
  125.  *   as all slab structures are 'small' - control structures of
  126.  *   their caches do not require further allocation
  127.  */
  128. static slab_cache_t *slab_extern_cache;
  129. /** Caches for malloc */
  130. static slab_cache_t *malloc_caches[SLAB_MAX_MALLOC_W-SLAB_MIN_MALLOC_W+1];
  131. char *malloc_names[] =  {
  132.     "malloc-16","malloc-32","malloc-64","malloc-128",
  133.     "malloc-256","malloc-512","malloc-1K","malloc-2K",
  134.     "malloc-4K","malloc-8K","malloc-16K","malloc-32K",
  135.     "malloc-64K","malloc-128K","malloc-256K"
  136. };
  137.  
  138. /** Slab descriptor */
  139. typedef struct {
  140.     slab_cache_t *cache; /**< Pointer to parent cache */
  141.     link_t link;       /* List of full/partial slabs */
  142.     void *start;       /**< Start address of first available item */
  143.     count_t available; /**< Count of available items in this slab */
  144.     index_t nextavail; /**< The index of next available item */
  145. }slab_t;
  146.  
  147. #ifdef CONFIG_DEBUG
  148. static int _slab_initialized = 0;
  149. #endif
  150.  
  151. /**************************************/
  152. /* Slab allocation functions          */
  153.  
  154. /**
  155.  * Allocate frames for slab space and initialize
  156.  *
  157.  */
  158. static slab_t * slab_space_alloc(slab_cache_t *cache, int flags)
  159. {
  160.     void *data;
  161.     slab_t *slab;
  162.     size_t fsize;
  163.     int i;
  164.     int status;
  165.     int zone=0;
  166.    
  167.     data = frame_alloc_rc_zone(cache->order, FRAME_KA | flags, &status, &zone);
  168.     if (status != FRAME_OK) {
  169.         return NULL;
  170.     }
  171.     if (! (cache->flags & SLAB_CACHE_SLINSIDE)) {
  172.         slab = slab_alloc(slab_extern_cache, flags);
  173.         if (!slab) {
  174.             frame_free(KA2PA(data));
  175.             return NULL;
  176.         }
  177.     } else {
  178.         fsize = (PAGE_SIZE << cache->order);
  179.         slab = data + fsize - sizeof(*slab);
  180.     }
  181.    
  182.     /* Fill in slab structures */
  183.     for (i=0; i < (1 << cache->order); i++)
  184.         frame_set_parent(ADDR2PFN(KA2PA(data))+i, slab, zone);
  185.  
  186.     slab->start = data;
  187.     slab->available = cache->objects;
  188.     slab->nextavail = 0;
  189.     slab->cache = cache;
  190.  
  191.     for (i=0; i<cache->objects;i++)
  192.         *((int *) (slab->start + i*cache->size)) = i+1;
  193.  
  194.     atomic_inc(&cache->allocated_slabs);
  195.     return slab;
  196. }
  197.  
  198. /**
  199.  * Deallocate space associated with slab
  200.  *
  201.  * @return number of freed frames
  202.  */
  203. static count_t slab_space_free(slab_cache_t *cache, slab_t *slab)
  204. {
  205.     frame_free(KA2PA(slab->start));
  206.     if (! (cache->flags & SLAB_CACHE_SLINSIDE))
  207.         slab_free(slab_extern_cache, slab);
  208.  
  209.     atomic_dec(&cache->allocated_slabs);
  210.    
  211.     return 1 << cache->order;
  212. }
  213.  
  214. /** Map object to slab structure */
  215. static slab_t * obj2slab(void *obj)
  216. {
  217.     return (slab_t *)frame_get_parent(ADDR2PFN(KA2PA(obj)), 0);
  218. }
  219.  
  220. /**************************************/
  221. /* Slab functions */
  222.  
  223.  
  224. /**
  225.  * Return object to slab and call a destructor
  226.  *
  227.  * @param slab If the caller knows directly slab of the object, otherwise NULL
  228.  *
  229.  * @return Number of freed pages
  230.  */
  231. static count_t slab_obj_destroy(slab_cache_t *cache, void *obj,
  232.                 slab_t *slab)
  233. {
  234.     int freed = 0;
  235.  
  236.     if (!slab)
  237.         slab = obj2slab(obj);
  238.  
  239.     ASSERT(slab->cache == cache);
  240.  
  241.     if (cache->destructor)
  242.         freed = cache->destructor(obj);
  243.    
  244.     spinlock_lock(&cache->slablock);
  245.     ASSERT(slab->available < cache->objects);
  246.  
  247.     *((int *)obj) = slab->nextavail;
  248.     slab->nextavail = (obj - slab->start)/cache->size;
  249.     slab->available++;
  250.  
  251.     /* Move it to correct list */
  252.     if (slab->available == cache->objects) {
  253.         /* Free associated memory */
  254.         list_remove(&slab->link);
  255.         spinlock_unlock(&cache->slablock);
  256.  
  257.         return freed + slab_space_free(cache, slab);
  258.  
  259.     } else if (slab->available == 1) {
  260.         /* It was in full, move to partial */
  261.         list_remove(&slab->link);
  262.         list_prepend(&slab->link, &cache->partial_slabs);
  263.     }
  264.     spinlock_unlock(&cache->slablock);
  265.     return freed;
  266. }
  267.  
  268. /**
  269.  * Take new object from slab or create new if needed
  270.  *
  271.  * @return Object address or null
  272.  */
  273. static void * slab_obj_create(slab_cache_t *cache, int flags)
  274. {
  275.     slab_t *slab;
  276.     void *obj;
  277.  
  278.     spinlock_lock(&cache->slablock);
  279.  
  280.     if (list_empty(&cache->partial_slabs)) {
  281.         /* Allow recursion and reclaiming
  282.          * - this should work, as the slab control structures
  283.          *   are small and do not need to allocate with anything
  284.          *   other than frame_alloc when they are allocating,
  285.          *   that's why we should get recursion at most 1-level deep
  286.          */
  287.         spinlock_unlock(&cache->slablock);
  288.         slab = slab_space_alloc(cache, flags);
  289.         if (!slab)
  290.             return NULL;
  291.         spinlock_lock(&cache->slablock);
  292.     } else {
  293.         slab = list_get_instance(cache->partial_slabs.next,
  294.                      slab_t,
  295.                      link);
  296.         list_remove(&slab->link);
  297.     }
  298.     obj = slab->start + slab->nextavail * cache->size;
  299.     slab->nextavail = *((int *)obj);
  300.     slab->available--;
  301.  
  302.     if (! slab->available)
  303.         list_prepend(&slab->link, &cache->full_slabs);
  304.     else
  305.         list_prepend(&slab->link, &cache->partial_slabs);
  306.  
  307.     spinlock_unlock(&cache->slablock);
  308.  
  309.     if (cache->constructor && cache->constructor(obj, flags)) {
  310.         /* Bad, bad, construction failed */
  311.         slab_obj_destroy(cache, obj, slab);
  312.         return NULL;
  313.     }
  314.     return obj;
  315. }
  316.  
  317. /**************************************/
  318. /* CPU-Cache slab functions */
  319.  
  320. /**
  321.  * Finds a full magazine in cache, takes it from list
  322.  * and returns it
  323.  *
  324.  * @param first If true, return first, else last mag
  325.  */
  326. static slab_magazine_t * get_mag_from_cache(slab_cache_t *cache,
  327.                         int first)
  328. {
  329.     slab_magazine_t *mag = NULL;
  330.     link_t *cur;
  331.  
  332.     spinlock_lock(&cache->maglock);
  333.     if (!list_empty(&cache->magazines)) {
  334.         if (first)
  335.             cur = cache->magazines.next;
  336.         else
  337.             cur = cache->magazines.prev;
  338.         mag = list_get_instance(cur, slab_magazine_t, link);
  339.         list_remove(&mag->link);
  340.         atomic_dec(&cache->magazine_counter);
  341.     }
  342.     spinlock_unlock(&cache->maglock);
  343.     return mag;
  344. }
  345.  
  346. /** Prepend magazine to magazine list in cache */
  347. static void put_mag_to_cache(slab_cache_t *cache, slab_magazine_t *mag)
  348. {
  349.     spinlock_lock(&cache->maglock);
  350.  
  351.     list_prepend(&mag->link, &cache->magazines);
  352.     atomic_inc(&cache->magazine_counter);
  353.    
  354.     spinlock_unlock(&cache->maglock);
  355. }
  356.  
  357. /**
  358.  * Free all objects in magazine and free memory associated with magazine
  359.  *
  360.  * @return Number of freed pages
  361.  */
  362. static count_t magazine_destroy(slab_cache_t *cache,
  363.                 slab_magazine_t *mag)
  364. {
  365.     int i;
  366.     count_t frames = 0;
  367.  
  368.     for (i=0;i < mag->busy; i++) {
  369.         frames += slab_obj_destroy(cache, mag->objs[i], NULL);
  370.         atomic_dec(&cache->cached_objs);
  371.     }
  372.    
  373.     slab_free(&mag_cache, mag);
  374.  
  375.     return frames;
  376. }
  377.  
  378. /**
  379.  * Find full magazine, set it as current and return it
  380.  *
  381.  * Assume cpu_magazine lock is held
  382.  */
  383. static slab_magazine_t * get_full_current_mag(slab_cache_t *cache)
  384. {
  385.     slab_magazine_t *cmag, *lastmag, *newmag;
  386.  
  387.     cmag = cache->mag_cache[CPU->id].current;
  388.     lastmag = cache->mag_cache[CPU->id].last;
  389.     if (cmag) { /* First try local CPU magazines */
  390.         if (cmag->busy)
  391.             return cmag;
  392.  
  393.         if (lastmag && lastmag->busy) {
  394.             cache->mag_cache[CPU->id].current = lastmag;
  395.             cache->mag_cache[CPU->id].last = cmag;
  396.             return lastmag;
  397.         }
  398.     }
  399.     /* Local magazines are empty, import one from magazine list */
  400.     newmag = get_mag_from_cache(cache, 1);
  401.     if (!newmag)
  402.         return NULL;
  403.  
  404.     if (lastmag)
  405.         magazine_destroy(cache, lastmag);
  406.  
  407.     cache->mag_cache[CPU->id].last = cmag;
  408.     cache->mag_cache[CPU->id].current = newmag;
  409.     return newmag;
  410. }
  411.  
  412. /**
  413.  * Try to find object in CPU-cache magazines
  414.  *
  415.  * @return Pointer to object or NULL if not available
  416.  */
  417. static void * magazine_obj_get(slab_cache_t *cache)
  418. {
  419.     slab_magazine_t *mag;
  420.     void *obj;
  421.  
  422.     if (!CPU)
  423.         return NULL;
  424.  
  425.     spinlock_lock(&cache->mag_cache[CPU->id].lock);
  426.  
  427.     mag = get_full_current_mag(cache);
  428.     if (!mag) {
  429.         spinlock_unlock(&cache->mag_cache[CPU->id].lock);
  430.         return NULL;
  431.     }
  432.     obj = mag->objs[--mag->busy];
  433.     spinlock_unlock(&cache->mag_cache[CPU->id].lock);
  434.     atomic_dec(&cache->cached_objs);
  435.    
  436.     return obj;
  437. }
  438.  
  439. /**
  440.  * Assure that the current magazine is empty, return pointer to it, or NULL if
  441.  * no empty magazine is available and cannot be allocated
  442.  *
  443.  * Assume mag_cache[CPU->id].lock is held
  444.  *
  445.  * We have 2 magazines bound to processor.
  446.  * First try the current.
  447.  *  If full, try the last.
  448.  *   If full, put to magazines list.
  449.  *   allocate new, exchange last & current
  450.  *
  451.  */
  452. static slab_magazine_t * make_empty_current_mag(slab_cache_t *cache)
  453. {
  454.     slab_magazine_t *cmag,*lastmag,*newmag;
  455.  
  456.     cmag = cache->mag_cache[CPU->id].current;
  457.     lastmag = cache->mag_cache[CPU->id].last;
  458.  
  459.     if (cmag) {
  460.         if (cmag->busy < cmag->size)
  461.             return cmag;
  462.         if (lastmag && lastmag->busy < lastmag->size) {
  463.             cache->mag_cache[CPU->id].last = cmag;
  464.             cache->mag_cache[CPU->id].current = lastmag;
  465.             return lastmag;
  466.         }
  467.     }
  468.     /* current | last are full | nonexistent, allocate new */
  469.     /* We do not want to sleep just because of caching */
  470.     /* Especially we do not want reclaiming to start, as
  471.      * this would deadlock */
  472.     newmag = slab_alloc(&mag_cache, FRAME_ATOMIC | FRAME_NO_RECLAIM);
  473.     if (!newmag)
  474.         return NULL;
  475.     newmag->size = SLAB_MAG_SIZE;
  476.     newmag->busy = 0;
  477.  
  478.     /* Flush last to magazine list */
  479.     if (lastmag)
  480.         put_mag_to_cache(cache, lastmag);
  481.  
  482.     /* Move current as last, save new as current */
  483.     cache->mag_cache[CPU->id].last = cmag; 
  484.     cache->mag_cache[CPU->id].current = newmag;
  485.  
  486.     return newmag;
  487. }
  488.  
  489. /**
  490.  * Put object into CPU-cache magazine
  491.  *
  492.  * @return 0 - success, -1 - could not get memory
  493.  */
  494. static int magazine_obj_put(slab_cache_t *cache, void *obj)
  495. {
  496.     slab_magazine_t *mag;
  497.  
  498.     if (!CPU)
  499.         return -1;
  500.  
  501.     spinlock_lock(&cache->mag_cache[CPU->id].lock);
  502.  
  503.     mag = make_empty_current_mag(cache);
  504.     if (!mag) {
  505.         spinlock_unlock(&cache->mag_cache[CPU->id].lock);
  506.         return -1;
  507.     }
  508.    
  509.     mag->objs[mag->busy++] = obj;
  510.  
  511.     spinlock_unlock(&cache->mag_cache[CPU->id].lock);
  512.     atomic_inc(&cache->cached_objs);
  513.     return 0;
  514. }
  515.  
  516.  
  517. /**************************************/
  518. /* Slab cache functions */
  519.  
  520. /** Return number of objects that fit in certain cache size */
  521. static int comp_objects(slab_cache_t *cache)
  522. {
  523.     if (cache->flags & SLAB_CACHE_SLINSIDE)
  524.         return ((PAGE_SIZE << cache->order) - sizeof(slab_t)) / cache->size;
  525.     else
  526.         return (PAGE_SIZE << cache->order) / cache->size;
  527. }
  528.  
  529. /** Return wasted space in slab */
  530. static int badness(slab_cache_t *cache)
  531. {
  532.     int objects;
  533.     int ssize;
  534.  
  535.     objects = comp_objects(cache);
  536.     ssize = PAGE_SIZE << cache->order;
  537.     if (cache->flags & SLAB_CACHE_SLINSIDE)
  538.         ssize -= sizeof(slab_t);
  539.     return ssize - objects*cache->size;
  540. }
  541.  
  542. /**
  543.  * Initialize mag_cache structure in slab cache
  544.  */
  545. static void make_magcache(slab_cache_t *cache)
  546. {
  547.     int i;
  548.    
  549.     ASSERT(_slab_initialized >= 2);
  550.  
  551.     cache->mag_cache = malloc(sizeof(slab_mag_cache_t)*config.cpu_count,0);
  552.     for (i=0; i < config.cpu_count; i++) {
  553.         memsetb((__address)&cache->mag_cache[i],
  554.             sizeof(cache->mag_cache[i]), 0);
  555.         spinlock_initialize(&cache->mag_cache[i].lock,
  556.                     "slab_maglock_cpu");
  557.     }
  558. }
  559.  
  560. /** Initialize allocated memory as a slab cache */
  561. static void
  562. _slab_cache_create(slab_cache_t *cache,
  563.            char *name,
  564.            size_t size,
  565.            size_t align,
  566.            int (*constructor)(void *obj, int kmflag),
  567.            int (*destructor)(void *obj),
  568.            int flags)
  569. {
  570.     int pages;
  571.     ipl_t ipl;
  572.  
  573.     memsetb((__address)cache, sizeof(*cache), 0);
  574.     cache->name = name;
  575.  
  576.     if (align < sizeof(__native))
  577.         align = sizeof(__native);
  578.     size = ALIGN_UP(size, align);
  579.        
  580.     cache->size = size;
  581.  
  582.     cache->constructor = constructor;
  583.     cache->destructor = destructor;
  584.     cache->flags = flags;
  585.  
  586.     list_initialize(&cache->full_slabs);
  587.     list_initialize(&cache->partial_slabs);
  588.     list_initialize(&cache->magazines);
  589.     spinlock_initialize(&cache->slablock, "slab_lock");
  590.     spinlock_initialize(&cache->maglock, "slab_maglock");
  591.     if (! (cache->flags & SLAB_CACHE_NOMAGAZINE))
  592.         make_magcache(cache);
  593.  
  594.     /* Compute slab sizes, object counts in slabs etc. */
  595.     if (cache->size < SLAB_INSIDE_SIZE)
  596.         cache->flags |= SLAB_CACHE_SLINSIDE;
  597.  
  598.     /* Minimum slab order */
  599.     pages = SIZE2FRAMES(cache->size);
  600.     /* We need the 2^order >= pages */
  601.     if (pages == 1)
  602.         cache->order = 0;
  603.     else
  604.         cache->order = fnzb(pages-1)+1;
  605.  
  606.     while (badness(cache) > SLAB_MAX_BADNESS(cache)) {
  607.         cache->order += 1;
  608.     }
  609.     cache->objects = comp_objects(cache);
  610.     /* If info fits in, put it inside */
  611.     if (badness(cache) > sizeof(slab_t))
  612.         cache->flags |= SLAB_CACHE_SLINSIDE;
  613.  
  614.     /* Add cache to cache list */
  615.     ipl = interrupts_disable();
  616.     spinlock_lock(&slab_cache_lock);
  617.  
  618.     list_append(&cache->link, &slab_cache_list);
  619.  
  620.     spinlock_unlock(&slab_cache_lock);
  621.     interrupts_restore(ipl);
  622. }
  623.  
  624. /** Create slab cache  */
  625. slab_cache_t * slab_cache_create(char *name,
  626.                  size_t size,
  627.                  size_t align,
  628.                  int (*constructor)(void *obj, int kmflag),
  629.                  int (*destructor)(void *obj),
  630.                  int flags)
  631. {
  632.     slab_cache_t *cache;
  633.  
  634.     cache = slab_alloc(&slab_cache_cache, 0);
  635.     _slab_cache_create(cache, name, size, align, constructor, destructor,
  636.                flags);
  637.     return cache;
  638. }
  639.  
  640. /**
  641.  * Reclaim space occupied by objects that are already free
  642.  *
  643.  * @param flags If contains SLAB_RECLAIM_ALL, do aggressive freeing
  644.  * @return Number of freed pages
  645.  */
  646. static count_t _slab_reclaim(slab_cache_t *cache, int flags)
  647. {
  648.     int i;
  649.     slab_magazine_t *mag;
  650.     count_t frames = 0;
  651.     int magcount;
  652.    
  653.     if (cache->flags & SLAB_CACHE_NOMAGAZINE)
  654.         return 0; /* Nothing to do */
  655.  
  656.     /* We count up to original magazine count to avoid
  657.      * endless loop
  658.      */
  659.     magcount = atomic_get(&cache->magazine_counter);
  660.     while (magcount-- && (mag=get_mag_from_cache(cache,0))) {
  661.         frames += magazine_destroy(cache,mag);
  662.         if (!(flags & SLAB_RECLAIM_ALL) && frames)
  663.             break;
  664.     }
  665.    
  666.     if (flags & SLAB_RECLAIM_ALL) {
  667.         /* Free cpu-bound magazines */
  668.         /* Destroy CPU magazines */
  669.         for (i=0; i<config.cpu_count; i++) {
  670.             spinlock_lock(&cache->mag_cache[i].lock);
  671.  
  672.             mag = cache->mag_cache[i].current;
  673.             if (mag)
  674.                 frames += magazine_destroy(cache, mag);
  675.             cache->mag_cache[i].current = NULL;
  676.            
  677.             mag = cache->mag_cache[i].last;
  678.             if (mag)
  679.                 frames += magazine_destroy(cache, mag);
  680.             cache->mag_cache[i].last = NULL;
  681.  
  682.             spinlock_unlock(&cache->mag_cache[i].lock);
  683.         }
  684.     }
  685.  
  686.     return frames;
  687. }
  688.  
  689. /** Check that there are no slabs and remove cache from system  */
  690. void slab_cache_destroy(slab_cache_t *cache)
  691. {
  692.     ipl_t ipl;
  693.  
  694.     /* First remove cache from link, so that we don't need
  695.      * to disable interrupts later
  696.      */
  697.  
  698.     ipl = interrupts_disable();
  699.     spinlock_lock(&slab_cache_lock);
  700.  
  701.     list_remove(&cache->link);
  702.  
  703.     spinlock_unlock(&slab_cache_lock);
  704.     interrupts_restore(ipl);
  705.  
  706.     /* Do not lock anything, we assume the software is correct and
  707.      * does not touch the cache when it decides to destroy it */
  708.    
  709.     /* Destroy all magazines */
  710.     _slab_reclaim(cache, SLAB_RECLAIM_ALL);
  711.  
  712.     /* All slabs must be empty */
  713.     if (!list_empty(&cache->full_slabs) \
  714.         || !list_empty(&cache->partial_slabs))
  715.         panic("Destroying cache that is not empty.");
  716.  
  717.     if (!(cache->flags & SLAB_CACHE_NOMAGAZINE))
  718.         free(cache->mag_cache);
  719.     slab_free(&slab_cache_cache, cache);
  720. }
  721.  
  722. /** Allocate new object from cache - if no flags given, always returns
  723.     memory */
  724. void * slab_alloc(slab_cache_t *cache, int flags)
  725. {
  726.     ipl_t ipl;
  727.     void *result = NULL;
  728.    
  729.     /* Disable interrupts to avoid deadlocks with interrupt handlers */
  730.     ipl = interrupts_disable();
  731.  
  732.     if (!(cache->flags & SLAB_CACHE_NOMAGAZINE)) {
  733.         result = magazine_obj_get(cache);
  734.     }
  735.     if (!result)
  736.         result = slab_obj_create(cache, flags);
  737.  
  738.     interrupts_restore(ipl);
  739.  
  740.     if (result)
  741.         atomic_inc(&cache->allocated_objs);
  742.  
  743.     return result;
  744. }
  745.  
  746. /** Return object to cache, use slab if known  */
  747. static void _slab_free(slab_cache_t *cache, void *obj, slab_t *slab)
  748. {
  749.     ipl_t ipl;
  750.  
  751.     ipl = interrupts_disable();
  752.  
  753.     if ((cache->flags & SLAB_CACHE_NOMAGAZINE) \
  754.         || magazine_obj_put(cache, obj)) {
  755.  
  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; cur=cur->next) {
  783.         cache = list_get_instance(cur, slab_cache_t, link);
  784.         frames += _slab_reclaim(cache, flags);
  785.     }
  786.  
  787.     spinlock_unlock(&slab_cache_lock);
  788.  
  789.     return frames;
  790. }
  791.  
  792.  
  793. /* Print list of slabs */
  794. void slab_print_list(void)
  795. {
  796.     slab_cache_t *cache;
  797.     link_t *cur;
  798.     ipl_t ipl;
  799.    
  800.     ipl = interrupts_disable();
  801.     spinlock_lock(&slab_cache_lock);
  802.     printf("slab name\t  Osize\t  Pages\t Obj/pg\t  Slabs\t Cached\tAllocobjs\tCtl\n");
  803.     for (cur = slab_cache_list.next;cur!=&slab_cache_list; cur=cur->next) {
  804.         cache = list_get_instance(cur, slab_cache_t, link);
  805.         printf("%s\t%7zd\t%7zd\t%7zd\t%7zd\t%7zd\t%7zd\t\t%s\n", cache->name, cache->size,
  806.                (1 << cache->order), cache->objects,
  807.                atomic_get(&cache->allocated_slabs),
  808.                atomic_get(&cache->cached_objs),
  809.                atomic_get(&cache->allocated_objs),
  810.                cache->flags & SLAB_CACHE_SLINSIDE ? "In" : "Out");
  811.     }
  812.     spinlock_unlock(&slab_cache_lock);
  813.     interrupts_restore(ipl);
  814. }
  815.  
  816. void slab_cache_init(void)
  817. {
  818.     int i, size;
  819.  
  820.     /* Initialize magazine cache */
  821.     _slab_cache_create(&mag_cache,
  822.                "slab_magazine",
  823.                sizeof(slab_magazine_t)+SLAB_MAG_SIZE*sizeof(void*),
  824.                sizeof(__address),
  825.                NULL, NULL,
  826.                SLAB_CACHE_NOMAGAZINE | SLAB_CACHE_SLINSIDE);
  827.     /* Initialize slab_cache cache */
  828.     _slab_cache_create(&slab_cache_cache,
  829.                "slab_cache",
  830.                sizeof(slab_cache_cache),
  831.                sizeof(__address),
  832.                NULL, NULL,
  833.                SLAB_CACHE_NOMAGAZINE | SLAB_CACHE_SLINSIDE);
  834.     /* Initialize external slab cache */
  835.     slab_extern_cache = slab_cache_create("slab_extern",
  836.                           sizeof(slab_t),
  837.                           0, NULL, NULL,
  838.                           SLAB_CACHE_SLINSIDE | SLAB_CACHE_MAGDEFERRED);
  839.  
  840.     /* Initialize structures for malloc */
  841.     for (i=0, size=(1<<SLAB_MIN_MALLOC_W);
  842.          i < (SLAB_MAX_MALLOC_W-SLAB_MIN_MALLOC_W+1);
  843.          i++, size <<= 1) {
  844.         malloc_caches[i] = slab_cache_create(malloc_names[i],
  845.                              size, 0,
  846.                              NULL,NULL, SLAB_CACHE_MAGDEFERRED);
  847.     }
  848. #ifdef CONFIG_DEBUG      
  849.     _slab_initialized = 1;
  850. #endif
  851. }
  852.  
  853. /** Enable cpu_cache
  854.  *
  855.  * Kernel calls this function, when it knows the real number of
  856.  * processors.
  857.  * Allocate slab for cpucache and enable it on all existing
  858.  * slabs that are SLAB_CACHE_MAGDEFERRED
  859.  */
  860. void slab_enable_cpucache(void)
  861. {
  862.     link_t *cur;
  863.     slab_cache_t *s;
  864.  
  865. #ifdef CONFIG_DEBUG
  866.     _slab_initialized = 2;
  867. #endif
  868.  
  869.     spinlock_lock(&slab_cache_lock);
  870.    
  871.     for (cur=slab_cache_list.next; cur != &slab_cache_list;cur=cur->next){
  872.         s = list_get_instance(cur, slab_cache_t, link);
  873.         if ((s->flags & SLAB_CACHE_MAGDEFERRED) != SLAB_CACHE_MAGDEFERRED)
  874.             continue;
  875.         make_magcache(s);
  876.         s->flags &= ~SLAB_CACHE_MAGDEFERRED;
  877.     }
  878.  
  879.     spinlock_unlock(&slab_cache_lock);
  880. }
  881.  
  882. /**************************************/
  883. /* kalloc/kfree functions             */
  884. void * malloc(unsigned int size, int flags)
  885. {
  886.     int idx;
  887.  
  888.     ASSERT(_slab_initialized);
  889.     ASSERT(size && size <= (1 << SLAB_MAX_MALLOC_W));
  890.    
  891.     if (size < (1 << SLAB_MIN_MALLOC_W))
  892.         size = (1 << SLAB_MIN_MALLOC_W);
  893.  
  894.     idx = fnzb(size-1) - SLAB_MIN_MALLOC_W + 1;
  895.  
  896.     return slab_alloc(malloc_caches[idx], flags);
  897. }
  898.  
  899. void free(void *obj)
  900. {
  901.     slab_t *slab;
  902.  
  903.     if (!obj) return;
  904.  
  905.     slab = obj2slab(obj);
  906.     _slab_free(slab->cache, obj, slab);
  907. }
  908.  
  909. /** @}
  910.  */
  911.