Subversion Repositories HelenOS-historic

Rev

Rev 1760 | 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 zone=0;
  165.    
  166.     data = frame_alloc_generic(cache->order, FRAME_KA | flags, &zone);
  167.     if (!data) {
  168.         return NULL;
  169.     }
  170.     if (! (cache->flags & SLAB_CACHE_SLINSIDE)) {
  171.         slab = slab_alloc(slab_extern_cache, flags);
  172.         if (!slab) {
  173.             frame_free(KA2PA(data));
  174.             return NULL;
  175.         }
  176.     } else {
  177.         fsize = (PAGE_SIZE << cache->order);
  178.         slab = data + fsize - sizeof(*slab);
  179.     }
  180.    
  181.     /* Fill in slab structures */
  182.     for (i=0; i < (1 << cache->order); i++)
  183.         frame_set_parent(ADDR2PFN(KA2PA(data))+i, slab, zone);
  184.  
  185.     slab->start = data;
  186.     slab->available = cache->objects;
  187.     slab->nextavail = 0;
  188.     slab->cache = cache;
  189.  
  190.     for (i=0; i<cache->objects;i++)
  191.         *((int *) (slab->start + i*cache->size)) = i+1;
  192.  
  193.     atomic_inc(&cache->allocated_slabs);
  194.     return slab;
  195. }
  196.  
  197. /**
  198.  * Deallocate space associated with slab
  199.  *
  200.  * @return number of freed frames
  201.  */
  202. static count_t slab_space_free(slab_cache_t *cache, slab_t *slab)
  203. {
  204.     frame_free(KA2PA(slab->start));
  205.     if (! (cache->flags & SLAB_CACHE_SLINSIDE))
  206.         slab_free(slab_extern_cache, slab);
  207.  
  208.     atomic_dec(&cache->allocated_slabs);
  209.    
  210.     return 1 << cache->order;
  211. }
  212.  
  213. /** Map object to slab structure */
  214. static slab_t * obj2slab(void *obj)
  215. {
  216.     return (slab_t *)frame_get_parent(ADDR2PFN(KA2PA(obj)), 0);
  217. }
  218.  
  219. /**************************************/
  220. /* Slab functions */
  221.  
  222.  
  223. /**
  224.  * Return object to slab and call a destructor
  225.  *
  226.  * @param slab If the caller knows directly slab of the object, otherwise NULL
  227.  *
  228.  * @return Number of freed pages
  229.  */
  230. static count_t slab_obj_destroy(slab_cache_t *cache, void *obj,
  231.                 slab_t *slab)
  232. {
  233.     int freed = 0;
  234.  
  235.     if (!slab)
  236.         slab = obj2slab(obj);
  237.  
  238.     ASSERT(slab->cache == cache);
  239.  
  240.     if (cache->destructor)
  241.         freed = cache->destructor(obj);
  242.    
  243.     spinlock_lock(&cache->slablock);
  244.     ASSERT(slab->available < cache->objects);
  245.  
  246.     *((int *)obj) = slab->nextavail;
  247.     slab->nextavail = (obj - slab->start)/cache->size;
  248.     slab->available++;
  249.  
  250.     /* Move it to correct list */
  251.     if (slab->available == cache->objects) {
  252.         /* Free associated memory */
  253.         list_remove(&slab->link);
  254.         spinlock_unlock(&cache->slablock);
  255.  
  256.         return freed + slab_space_free(cache, slab);
  257.  
  258.     } else if (slab->available == 1) {
  259.         /* It was in full, move to partial */
  260.         list_remove(&slab->link);
  261.         list_prepend(&slab->link, &cache->partial_slabs);
  262.     }
  263.     spinlock_unlock(&cache->slablock);
  264.     return freed;
  265. }
  266.  
  267. /**
  268.  * Take new object from slab or create new if needed
  269.  *
  270.  * @return Object address or null
  271.  */
  272. static void * slab_obj_create(slab_cache_t *cache, int flags)
  273. {
  274.     slab_t *slab;
  275.     void *obj;
  276.  
  277.     spinlock_lock(&cache->slablock);
  278.  
  279.     if (list_empty(&cache->partial_slabs)) {
  280.         /* Allow recursion and reclaiming
  281.          * - this should work, as the slab control structures
  282.          *   are small and do not need to allocate with anything
  283.          *   other than frame_alloc when they are allocating,
  284.          *   that's why we should get recursion at most 1-level deep
  285.          */
  286.         spinlock_unlock(&cache->slablock);
  287.         slab = slab_space_alloc(cache, flags);
  288.         if (!slab)
  289.             return NULL;
  290.         spinlock_lock(&cache->slablock);
  291.     } else {
  292.         slab = list_get_instance(cache->partial_slabs.next,
  293.                      slab_t,
  294.                      link);
  295.         list_remove(&slab->link);
  296.     }
  297.     obj = slab->start + slab->nextavail * cache->size;
  298.     slab->nextavail = *((int *)obj);
  299.     slab->available--;
  300.  
  301.     if (! slab->available)
  302.         list_prepend(&slab->link, &cache->full_slabs);
  303.     else
  304.         list_prepend(&slab->link, &cache->partial_slabs);
  305.  
  306.     spinlock_unlock(&cache->slablock);
  307.  
  308.     if (cache->constructor && cache->constructor(obj, flags)) {
  309.         /* Bad, bad, construction failed */
  310.         slab_obj_destroy(cache, obj, slab);
  311.         return NULL;
  312.     }
  313.     return obj;
  314. }
  315.  
  316. /**************************************/
  317. /* CPU-Cache slab functions */
  318.  
  319. /**
  320.  * Finds a full magazine in cache, takes it from list
  321.  * and returns it
  322.  *
  323.  * @param first If true, return first, else last mag
  324.  */
  325. static slab_magazine_t * get_mag_from_cache(slab_cache_t *cache,
  326.                         int first)
  327. {
  328.     slab_magazine_t *mag = NULL;
  329.     link_t *cur;
  330.  
  331.     spinlock_lock(&cache->maglock);
  332.     if (!list_empty(&cache->magazines)) {
  333.         if (first)
  334.             cur = cache->magazines.next;
  335.         else
  336.             cur = cache->magazines.prev;
  337.         mag = list_get_instance(cur, slab_magazine_t, link);
  338.         list_remove(&mag->link);
  339.         atomic_dec(&cache->magazine_counter);
  340.     }
  341.     spinlock_unlock(&cache->maglock);
  342.     return mag;
  343. }
  344.  
  345. /** Prepend magazine to magazine list in cache */
  346. static void put_mag_to_cache(slab_cache_t *cache, slab_magazine_t *mag)
  347. {
  348.     spinlock_lock(&cache->maglock);
  349.  
  350.     list_prepend(&mag->link, &cache->magazines);
  351.     atomic_inc(&cache->magazine_counter);
  352.    
  353.     spinlock_unlock(&cache->maglock);
  354. }
  355.  
  356. /**
  357.  * Free all objects in magazine and free memory associated with magazine
  358.  *
  359.  * @return Number of freed pages
  360.  */
  361. static count_t magazine_destroy(slab_cache_t *cache,
  362.                 slab_magazine_t *mag)
  363. {
  364.     int i;
  365.     count_t frames = 0;
  366.  
  367.     for (i=0;i < mag->busy; i++) {
  368.         frames += slab_obj_destroy(cache, mag->objs[i], NULL);
  369.         atomic_dec(&cache->cached_objs);
  370.     }
  371.    
  372.     slab_free(&mag_cache, mag);
  373.  
  374.     return frames;
  375. }
  376.  
  377. /**
  378.  * Find full magazine, set it as current and return it
  379.  *
  380.  * Assume cpu_magazine lock is held
  381.  */
  382. static slab_magazine_t * get_full_current_mag(slab_cache_t *cache)
  383. {
  384.     slab_magazine_t *cmag, *lastmag, *newmag;
  385.  
  386.     cmag = cache->mag_cache[CPU->id].current;
  387.     lastmag = cache->mag_cache[CPU->id].last;
  388.     if (cmag) { /* First try local CPU magazines */
  389.         if (cmag->busy)
  390.             return cmag;
  391.  
  392.         if (lastmag && lastmag->busy) {
  393.             cache->mag_cache[CPU->id].current = lastmag;
  394.             cache->mag_cache[CPU->id].last = cmag;
  395.             return lastmag;
  396.         }
  397.     }
  398.     /* Local magazines are empty, import one from magazine list */
  399.     newmag = get_mag_from_cache(cache, 1);
  400.     if (!newmag)
  401.         return NULL;
  402.  
  403.     if (lastmag)
  404.         magazine_destroy(cache, lastmag);
  405.  
  406.     cache->mag_cache[CPU->id].last = cmag;
  407.     cache->mag_cache[CPU->id].current = newmag;
  408.     return newmag;
  409. }
  410.  
  411. /**
  412.  * Try to find object in CPU-cache magazines
  413.  *
  414.  * @return Pointer to object or NULL if not available
  415.  */
  416. static void * magazine_obj_get(slab_cache_t *cache)
  417. {
  418.     slab_magazine_t *mag;
  419.     void *obj;
  420.  
  421.     if (!CPU)
  422.         return NULL;
  423.  
  424.     spinlock_lock(&cache->mag_cache[CPU->id].lock);
  425.  
  426.     mag = get_full_current_mag(cache);
  427.     if (!mag) {
  428.         spinlock_unlock(&cache->mag_cache[CPU->id].lock);
  429.         return NULL;
  430.     }
  431.     obj = mag->objs[--mag->busy];
  432.     spinlock_unlock(&cache->mag_cache[CPU->id].lock);
  433.     atomic_dec(&cache->cached_objs);
  434.    
  435.     return obj;
  436. }
  437.  
  438. /**
  439.  * Assure that the current magazine is empty, return pointer to it, or NULL if
  440.  * no empty magazine is available and cannot be allocated
  441.  *
  442.  * Assume mag_cache[CPU->id].lock is held
  443.  *
  444.  * We have 2 magazines bound to processor.
  445.  * First try the current.
  446.  *  If full, try the last.
  447.  *   If full, put to magazines list.
  448.  *   allocate new, exchange last & current
  449.  *
  450.  */
  451. static slab_magazine_t * make_empty_current_mag(slab_cache_t *cache)
  452. {
  453.     slab_magazine_t *cmag,*lastmag,*newmag;
  454.  
  455.     cmag = cache->mag_cache[CPU->id].current;
  456.     lastmag = cache->mag_cache[CPU->id].last;
  457.  
  458.     if (cmag) {
  459.         if (cmag->busy < cmag->size)
  460.             return cmag;
  461.         if (lastmag && lastmag->busy < lastmag->size) {
  462.             cache->mag_cache[CPU->id].last = cmag;
  463.             cache->mag_cache[CPU->id].current = lastmag;
  464.             return lastmag;
  465.         }
  466.     }
  467.     /* current | last are full | nonexistent, allocate new */
  468.     /* We do not want to sleep just because of caching */
  469.     /* Especially we do not want reclaiming to start, as
  470.      * this would deadlock */
  471.     newmag = slab_alloc(&mag_cache, FRAME_ATOMIC | FRAME_NO_RECLAIM);
  472.     if (!newmag)
  473.         return NULL;
  474.     newmag->size = SLAB_MAG_SIZE;
  475.     newmag->busy = 0;
  476.  
  477.     /* Flush last to magazine list */
  478.     if (lastmag)
  479.         put_mag_to_cache(cache, lastmag);
  480.  
  481.     /* Move current as last, save new as current */
  482.     cache->mag_cache[CPU->id].last = cmag; 
  483.     cache->mag_cache[CPU->id].current = newmag;
  484.  
  485.     return newmag;
  486. }
  487.  
  488. /**
  489.  * Put object into CPU-cache magazine
  490.  *
  491.  * @return 0 - success, -1 - could not get memory
  492.  */
  493. static int magazine_obj_put(slab_cache_t *cache, void *obj)
  494. {
  495.     slab_magazine_t *mag;
  496.  
  497.     if (!CPU)
  498.         return -1;
  499.  
  500.     spinlock_lock(&cache->mag_cache[CPU->id].lock);
  501.  
  502.     mag = make_empty_current_mag(cache);
  503.     if (!mag) {
  504.         spinlock_unlock(&cache->mag_cache[CPU->id].lock);
  505.         return -1;
  506.     }
  507.    
  508.     mag->objs[mag->busy++] = obj;
  509.  
  510.     spinlock_unlock(&cache->mag_cache[CPU->id].lock);
  511.     atomic_inc(&cache->cached_objs);
  512.     return 0;
  513. }
  514.  
  515.  
  516. /**************************************/
  517. /* Slab cache functions */
  518.  
  519. /** Return number of objects that fit in certain cache size */
  520. static int comp_objects(slab_cache_t *cache)
  521. {
  522.     if (cache->flags & SLAB_CACHE_SLINSIDE)
  523.         return ((PAGE_SIZE << cache->order) - sizeof(slab_t)) / cache->size;
  524.     else
  525.         return (PAGE_SIZE << cache->order) / cache->size;
  526. }
  527.  
  528. /** Return wasted space in slab */
  529. static int badness(slab_cache_t *cache)
  530. {
  531.     int objects;
  532.     int ssize;
  533.  
  534.     objects = comp_objects(cache);
  535.     ssize = PAGE_SIZE << cache->order;
  536.     if (cache->flags & SLAB_CACHE_SLINSIDE)
  537.         ssize -= sizeof(slab_t);
  538.     return ssize - objects*cache->size;
  539. }
  540.  
  541. /**
  542.  * Initialize mag_cache structure in slab cache
  543.  */
  544. static void make_magcache(slab_cache_t *cache)
  545. {
  546.     int i;
  547.    
  548.     ASSERT(_slab_initialized >= 2);
  549.  
  550.     cache->mag_cache = malloc(sizeof(slab_mag_cache_t)*config.cpu_count,0);
  551.     for (i=0; i < config.cpu_count; i++) {
  552.         memsetb((__address)&cache->mag_cache[i],
  553.             sizeof(cache->mag_cache[i]), 0);
  554.         spinlock_initialize(&cache->mag_cache[i].lock,
  555.                     "slab_maglock_cpu");
  556.     }
  557. }
  558.  
  559. /** Initialize allocated memory as a slab cache */
  560. static void
  561. _slab_cache_create(slab_cache_t *cache,
  562.            char *name,
  563.            size_t size,
  564.            size_t align,
  565.            int (*constructor)(void *obj, int kmflag),
  566.            int (*destructor)(void *obj),
  567.            int flags)
  568. {
  569.     int pages;
  570.     ipl_t ipl;
  571.  
  572.     memsetb((__address)cache, sizeof(*cache), 0);
  573.     cache->name = name;
  574.  
  575.     if (align < sizeof(__native))
  576.         align = sizeof(__native);
  577.     size = ALIGN_UP(size, align);
  578.        
  579.     cache->size = size;
  580.  
  581.     cache->constructor = constructor;
  582.     cache->destructor = destructor;
  583.     cache->flags = flags;
  584.  
  585.     list_initialize(&cache->full_slabs);
  586.     list_initialize(&cache->partial_slabs);
  587.     list_initialize(&cache->magazines);
  588.     spinlock_initialize(&cache->slablock, "slab_lock");
  589.     spinlock_initialize(&cache->maglock, "slab_maglock");
  590.     if (! (cache->flags & SLAB_CACHE_NOMAGAZINE))
  591.         make_magcache(cache);
  592.  
  593.     /* Compute slab sizes, object counts in slabs etc. */
  594.     if (cache->size < SLAB_INSIDE_SIZE)
  595.         cache->flags |= SLAB_CACHE_SLINSIDE;
  596.  
  597.     /* Minimum slab order */
  598.     pages = SIZE2FRAMES(cache->size);
  599.     /* We need the 2^order >= pages */
  600.     if (pages == 1)
  601.         cache->order = 0;
  602.     else
  603.         cache->order = fnzb(pages-1)+1;
  604.  
  605.     while (badness(cache) > SLAB_MAX_BADNESS(cache)) {
  606.         cache->order += 1;
  607.     }
  608.     cache->objects = comp_objects(cache);
  609.     /* If info fits in, put it inside */
  610.     if (badness(cache) > sizeof(slab_t))
  611.         cache->flags |= SLAB_CACHE_SLINSIDE;
  612.  
  613.     /* Add cache to cache list */
  614.     ipl = interrupts_disable();
  615.     spinlock_lock(&slab_cache_lock);
  616.  
  617.     list_append(&cache->link, &slab_cache_list);
  618.  
  619.     spinlock_unlock(&slab_cache_lock);
  620.     interrupts_restore(ipl);
  621. }
  622.  
  623. /** Create slab cache  */
  624. slab_cache_t * slab_cache_create(char *name,
  625.                  size_t size,
  626.                  size_t align,
  627.                  int (*constructor)(void *obj, int kmflag),
  628.                  int (*destructor)(void *obj),
  629.                  int flags)
  630. {
  631.     slab_cache_t *cache;
  632.  
  633.     cache = slab_alloc(&slab_cache_cache, 0);
  634.     _slab_cache_create(cache, name, size, align, constructor, destructor,
  635.                flags);
  636.     return cache;
  637. }
  638.  
  639. /**
  640.  * Reclaim space occupied by objects that are already free
  641.  *
  642.  * @param flags If contains SLAB_RECLAIM_ALL, do aggressive freeing
  643.  * @return Number of freed pages
  644.  */
  645. static count_t _slab_reclaim(slab_cache_t *cache, int flags)
  646. {
  647.     int i;
  648.     slab_magazine_t *mag;
  649.     count_t frames = 0;
  650.     int magcount;
  651.    
  652.     if (cache->flags & SLAB_CACHE_NOMAGAZINE)
  653.         return 0; /* Nothing to do */
  654.  
  655.     /* We count up to original magazine count to avoid
  656.      * endless loop
  657.      */
  658.     magcount = atomic_get(&cache->magazine_counter);
  659.     while (magcount-- && (mag=get_mag_from_cache(cache,0))) {
  660.         frames += magazine_destroy(cache,mag);
  661.         if (!(flags & SLAB_RECLAIM_ALL) && frames)
  662.             break;
  663.     }
  664.    
  665.     if (flags & SLAB_RECLAIM_ALL) {
  666.         /* Free cpu-bound magazines */
  667.         /* Destroy CPU magazines */
  668.         for (i=0; i<config.cpu_count; i++) {
  669.             spinlock_lock(&cache->mag_cache[i].lock);
  670.  
  671.             mag = cache->mag_cache[i].current;
  672.             if (mag)
  673.                 frames += magazine_destroy(cache, mag);
  674.             cache->mag_cache[i].current = NULL;
  675.            
  676.             mag = cache->mag_cache[i].last;
  677.             if (mag)
  678.                 frames += magazine_destroy(cache, mag);
  679.             cache->mag_cache[i].last = NULL;
  680.  
  681.             spinlock_unlock(&cache->mag_cache[i].lock);
  682.         }
  683.     }
  684.  
  685.     return frames;
  686. }
  687.  
  688. /** Check that there are no slabs and remove cache from system  */
  689. void slab_cache_destroy(slab_cache_t *cache)
  690. {
  691.     ipl_t ipl;
  692.  
  693.     /* First remove cache from link, so that we don't need
  694.      * to disable interrupts later
  695.      */
  696.  
  697.     ipl = interrupts_disable();
  698.     spinlock_lock(&slab_cache_lock);
  699.  
  700.     list_remove(&cache->link);
  701.  
  702.     spinlock_unlock(&slab_cache_lock);
  703.     interrupts_restore(ipl);
  704.  
  705.     /* Do not lock anything, we assume the software is correct and
  706.      * does not touch the cache when it decides to destroy it */
  707.    
  708.     /* Destroy all magazines */
  709.     _slab_reclaim(cache, SLAB_RECLAIM_ALL);
  710.  
  711.     /* All slabs must be empty */
  712.     if (!list_empty(&cache->full_slabs) \
  713.         || !list_empty(&cache->partial_slabs))
  714.         panic("Destroying cache that is not empty.");
  715.  
  716.     if (!(cache->flags & SLAB_CACHE_NOMAGAZINE))
  717.         free(cache->mag_cache);
  718.     slab_free(&slab_cache_cache, cache);
  719. }
  720.  
  721. /** Allocate new object from cache - if no flags given, always returns
  722.     memory */
  723. void * slab_alloc(slab_cache_t *cache, int flags)
  724. {
  725.     ipl_t ipl;
  726.     void *result = NULL;
  727.    
  728.     /* Disable interrupts to avoid deadlocks with interrupt handlers */
  729.     ipl = interrupts_disable();
  730.  
  731.     if (!(cache->flags & SLAB_CACHE_NOMAGAZINE)) {
  732.         result = magazine_obj_get(cache);
  733.     }
  734.     if (!result)
  735.         result = slab_obj_create(cache, flags);
  736.  
  737.     interrupts_restore(ipl);
  738.  
  739.     if (result)
  740.         atomic_inc(&cache->allocated_objs);
  741.  
  742.     return result;
  743. }
  744.  
  745. /** Return object to cache, use slab if known  */
  746. static void _slab_free(slab_cache_t *cache, void *obj, slab_t *slab)
  747. {
  748.     ipl_t ipl;
  749.  
  750.     ipl = interrupts_disable();
  751.  
  752.     if ((cache->flags & SLAB_CACHE_NOMAGAZINE) \
  753.         || magazine_obj_put(cache, obj)) {
  754.  
  755.         slab_obj_destroy(cache, obj, slab);
  756.  
  757.     }
  758.     interrupts_restore(ipl);
  759.     atomic_dec(&cache->allocated_objs);
  760. }
  761.  
  762. /** Return slab object to cache */
  763. void slab_free(slab_cache_t *cache, void *obj)
  764. {
  765.     _slab_free(cache,obj,NULL);
  766. }
  767.  
  768. /* Go through all caches and reclaim what is possible */
  769. count_t slab_reclaim(int flags)
  770. {
  771.     slab_cache_t *cache;
  772.     link_t *cur;
  773.     count_t frames = 0;
  774.  
  775.     spinlock_lock(&slab_cache_lock);
  776.  
  777.     /* TODO: Add assert, that interrupts are disabled, otherwise
  778.      * memory allocation from interrupts can deadlock.
  779.      */
  780.  
  781.     for (cur = slab_cache_list.next;cur!=&slab_cache_list; cur=cur->next) {
  782.         cache = list_get_instance(cur, slab_cache_t, link);
  783.         frames += _slab_reclaim(cache, flags);
  784.     }
  785.  
  786.     spinlock_unlock(&slab_cache_lock);
  787.  
  788.     return frames;
  789. }
  790.  
  791.  
  792. /* Print list of slabs */
  793. void slab_print_list(void)
  794. {
  795.     slab_cache_t *cache;
  796.     link_t *cur;
  797.     ipl_t ipl;
  798.    
  799.     ipl = interrupts_disable();
  800.     spinlock_lock(&slab_cache_lock);
  801.     printf("slab name\t  Osize\t  Pages\t Obj/pg\t  Slabs\t Cached\tAllocobjs\tCtl\n");
  802.     for (cur = slab_cache_list.next;cur!=&slab_cache_list; cur=cur->next) {
  803.         cache = list_get_instance(cur, slab_cache_t, link);
  804.         printf("%s\t%7zd\t%7zd\t%7zd\t%7zd\t%7zd\t%7zd\t\t%s\n", cache->name, cache->size,
  805.                (1 << cache->order), cache->objects,
  806.                atomic_get(&cache->allocated_slabs),
  807.                atomic_get(&cache->cached_objs),
  808.                atomic_get(&cache->allocated_objs),
  809.                cache->flags & SLAB_CACHE_SLINSIDE ? "In" : "Out");
  810.     }
  811.     spinlock_unlock(&slab_cache_lock);
  812.     interrupts_restore(ipl);
  813. }
  814.  
  815. void slab_cache_init(void)
  816. {
  817.     int i, size;
  818.  
  819.     /* Initialize magazine cache */
  820.     _slab_cache_create(&mag_cache,
  821.                "slab_magazine",
  822.                sizeof(slab_magazine_t)+SLAB_MAG_SIZE*sizeof(void*),
  823.                sizeof(__address),
  824.                NULL, NULL,
  825.                SLAB_CACHE_NOMAGAZINE | SLAB_CACHE_SLINSIDE);
  826.     /* Initialize slab_cache cache */
  827.     _slab_cache_create(&slab_cache_cache,
  828.                "slab_cache",
  829.                sizeof(slab_cache_cache),
  830.                sizeof(__address),
  831.                NULL, NULL,
  832.                SLAB_CACHE_NOMAGAZINE | SLAB_CACHE_SLINSIDE);
  833.     /* Initialize external slab cache */
  834.     slab_extern_cache = slab_cache_create("slab_extern",
  835.                           sizeof(slab_t),
  836.                           0, NULL, NULL,
  837.                           SLAB_CACHE_SLINSIDE | SLAB_CACHE_MAGDEFERRED);
  838.  
  839.     /* Initialize structures for malloc */
  840.     for (i=0, size=(1<<SLAB_MIN_MALLOC_W);
  841.          i < (SLAB_MAX_MALLOC_W-SLAB_MIN_MALLOC_W+1);
  842.          i++, size <<= 1) {
  843.         malloc_caches[i] = slab_cache_create(malloc_names[i],
  844.                              size, 0,
  845.                              NULL,NULL, SLAB_CACHE_MAGDEFERRED);
  846.     }
  847. #ifdef CONFIG_DEBUG      
  848.     _slab_initialized = 1;
  849. #endif
  850. }
  851.  
  852. /** Enable cpu_cache
  853.  *
  854.  * Kernel calls this function, when it knows the real number of
  855.  * processors.
  856.  * Allocate slab for cpucache and enable it on all existing
  857.  * slabs that are SLAB_CACHE_MAGDEFERRED
  858.  */
  859. void slab_enable_cpucache(void)
  860. {
  861.     link_t *cur;
  862.     slab_cache_t *s;
  863.  
  864. #ifdef CONFIG_DEBUG
  865.     _slab_initialized = 2;
  866. #endif
  867.  
  868.     spinlock_lock(&slab_cache_lock);
  869.    
  870.     for (cur=slab_cache_list.next; cur != &slab_cache_list;cur=cur->next){
  871.         s = list_get_instance(cur, slab_cache_t, link);
  872.         if ((s->flags & SLAB_CACHE_MAGDEFERRED) != SLAB_CACHE_MAGDEFERRED)
  873.             continue;
  874.         make_magcache(s);
  875.         s->flags &= ~SLAB_CACHE_MAGDEFERRED;
  876.     }
  877.  
  878.     spinlock_unlock(&slab_cache_lock);
  879. }
  880.  
  881. /**************************************/
  882. /* kalloc/kfree functions             */
  883. void * malloc(unsigned int size, int flags)
  884. {
  885.     int idx;
  886.  
  887.     ASSERT(_slab_initialized);
  888.     ASSERT(size && size <= (1 << SLAB_MAX_MALLOC_W));
  889.    
  890.     if (size < (1 << SLAB_MIN_MALLOC_W))
  891.         size = (1 << SLAB_MIN_MALLOC_W);
  892.  
  893.     idx = fnzb(size-1) - SLAB_MIN_MALLOC_W + 1;
  894.  
  895.     return slab_alloc(malloc_caches[idx], flags);
  896. }
  897.  
  898. void free(void *obj)
  899. {
  900.     slab_t *slab;
  901.  
  902.     if (!obj) return;
  903.  
  904.     slab = obj2slab(obj);
  905.     _slab_free(slab->cache, obj, slab);
  906. }
  907.  
  908. /** @}
  909.  */
  910.