Subversion Repositories HelenOS-historic

Rev

Rev 1554 | Rev 1682 | 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. /**
  30.  * @file    slab.c
  31.  * @brief   Slab allocator.
  32.  *
  33.  * The slab allocator is closely modelled after OpenSolaris slab allocator.
  34.  * @see http://www.usenix.org/events/usenix01/full_papers/bonwick/bonwick_html/
  35.  *
  36.  * with the following exceptions:
  37.  * @li empty slabs are deallocated immediately
  38.  *     (in Linux they are kept in linked list, in Solaris ???)
  39.  * @li empty magazines are deallocated when not needed
  40.  *     (in Solaris they are held in linked list in slab cache)
  41.  *
  42.  * Following features are not currently supported but would be easy to do:
  43.  * @li cache coloring
  44.  * @li dynamic magazine growing (different magazine sizes are already
  45.  *     supported, but we would need to adjust allocation strategy)
  46.  *
  47.  * The slab allocator supports per-CPU caches ('magazines') to facilitate
  48.  * good SMP scaling.
  49.  *
  50.  * When a new object is being allocated, it is first checked, if it is
  51.  * available in a CPU-bound magazine. If it is not found there, it is
  52.  * allocated from a CPU-shared slab - if a partially full one is found,
  53.  * it is used, otherwise a new one is allocated.
  54.  *
  55.  * When an object is being deallocated, it is put to a CPU-bound magazine.
  56.  * If there is no such magazine, a new one is allocated (if this fails,
  57.  * the object is deallocated into slab). If the magazine is full, it is
  58.  * put into cpu-shared list of magazines and a new one is allocated.
  59.  *
  60.  * The CPU-bound magazine is actually a pair of magazines in order to avoid
  61.  * thrashing when somebody is allocating/deallocating 1 item at the magazine
  62.  * size boundary. LIFO order is enforced, which should avoid fragmentation
  63.  * as much as possible.
  64.  *  
  65.  * Every cache contains list of full slabs and list of partially full slabs.
  66.  * Empty slabs are immediately freed (thrashing will be avoided because
  67.  * of magazines).
  68.  *
  69.  * The slab information structure is kept inside the data area, if possible.
  70.  * The cache can be marked that it should not use magazines. This is used
  71.  * only for slab related caches to avoid deadlocks and infinite recursion
  72.  * (the slab allocator uses itself for allocating all it's control structures).
  73.  *
  74.  * The slab allocator allocates a lot of space and does not free it. When
  75.  * the frame allocator fails to allocate a frame, it calls slab_reclaim().
  76.  * It tries 'light reclaim' first, then brutal reclaim. The light reclaim
  77.  * releases slabs from cpu-shared magazine-list, until at least 1 slab
  78.  * is deallocated in each cache (this algorithm should probably change).
  79.  * The brutal reclaim removes all cached objects, even from CPU-bound
  80.  * magazines.
  81.  *
  82.  * TODO:@n
  83.  * For better CPU-scaling the magazine allocation strategy should
  84.  * be extended. Currently, if the cache does not have magazine, it asks
  85.  * for non-cpu cached magazine cache to provide one. It might be feasible
  86.  * to add cpu-cached magazine cache (which would allocate it's magazines
  87.  * from non-cpu-cached mag. cache). This would provide a nice per-cpu
  88.  * buffer. The other possibility is to use the per-cache
  89.  * 'empty-magazine-list', which decreases competing for 1 per-system
  90.  * magazine cache.
  91.  *
  92.  * @li it might be good to add granularity of locks even to slab level,
  93.  *     we could then try_spinlock over all partial slabs and thus improve
  94.  *     scalability even on slab level
  95.  */
  96.  
  97. #include <synch/spinlock.h>
  98. #include <mm/slab.h>
  99. #include <adt/list.h>
  100. #include <memstr.h>
  101. #include <align.h>
  102. #include <mm/frame.h>
  103. #include <config.h>
  104. #include <print.h>
  105. #include <arch.h>
  106. #include <panic.h>
  107. #include <debug.h>
  108. #include <bitops.h>
  109.  
  110. SPINLOCK_INITIALIZE(slab_cache_lock);
  111. static LIST_INITIALIZE(slab_cache_list);
  112.  
  113. /** Magazine cache */
  114. static slab_cache_t mag_cache;
  115. /** Cache for cache descriptors */
  116. static slab_cache_t slab_cache_cache;
  117. /** Cache for external slab descriptors
  118.  * This time we want per-cpu cache, so do not make it static
  119.  * - using slab for internal slab structures will not deadlock,
  120.  *   as all slab structures are 'small' - control structures of
  121.  *   their caches do not require further allocation
  122.  */
  123. static slab_cache_t *slab_extern_cache;
  124. /** Caches for malloc */
  125. static slab_cache_t *malloc_caches[SLAB_MAX_MALLOC_W-SLAB_MIN_MALLOC_W+1];
  126. char *malloc_names[] =  {
  127.     "malloc-16","malloc-32","malloc-64","malloc-128",
  128.     "malloc-256","malloc-512","malloc-1K","malloc-2K",
  129.     "malloc-4K","malloc-8K","malloc-16K","malloc-32K",
  130.     "malloc-64K","malloc-128K","malloc-256K"
  131. };
  132.  
  133. /** Slab descriptor */
  134. typedef struct {
  135.     slab_cache_t *cache; /**< Pointer to parent cache */
  136.     link_t link;       /* List of full/partial slabs */
  137.     void *start;       /**< Start address of first available item */
  138.     count_t available; /**< Count of available items in this slab */
  139.     index_t nextavail; /**< The index of next available item */
  140. }slab_t;
  141.  
  142. #ifdef CONFIG_DEBUG
  143. static int _slab_initialized = 0;
  144. #endif
  145.  
  146. /**************************************/
  147. /* Slab allocation functions          */
  148.  
  149. /**
  150.  * Allocate frames for slab space and initialize
  151.  *
  152.  */
  153. static slab_t * slab_space_alloc(slab_cache_t *cache, int flags)
  154. {
  155.     void *data;
  156.     slab_t *slab;
  157.     size_t fsize;
  158.     int i;
  159.     int status;
  160.     pfn_t pfn;
  161.     int zone=0;
  162.    
  163.     pfn = frame_alloc_rc_zone(cache->order, FRAME_KA | flags, &status, &zone);
  164.     data = (void *) PA2KA(PFN2ADDR(pfn));
  165.     if (status != FRAME_OK) {
  166.         return NULL;
  167.     }
  168.     if (! (cache->flags & SLAB_CACHE_SLINSIDE)) {
  169.         slab = slab_alloc(slab_extern_cache, flags);
  170.         if (!slab) {
  171.             frame_free(ADDR2PFN(KA2PA(data)));
  172.             return NULL;
  173.         }
  174.     } else {
  175.         fsize = (PAGE_SIZE << cache->order);
  176.         slab = data + fsize - sizeof(*slab);
  177.     }
  178.    
  179.     /* Fill in slab structures */
  180.     for (i=0; i < (1 << cache->order); i++)
  181.         frame_set_parent(pfn+i, slab, zone);
  182.  
  183.     slab->start = data;
  184.     slab->available = cache->objects;
  185.     slab->nextavail = 0;
  186.     slab->cache = cache;
  187.  
  188.     for (i=0; i<cache->objects;i++)
  189.         *((int *) (slab->start + i*cache->size)) = i+1;
  190.  
  191.     atomic_inc(&cache->allocated_slabs);
  192.     return slab;
  193. }
  194.  
  195. /**
  196.  * Deallocate space associated with slab
  197.  *
  198.  * @return number of freed frames
  199.  */
  200. static count_t slab_space_free(slab_cache_t *cache, slab_t *slab)
  201. {
  202.     frame_free(ADDR2PFN(KA2PA(slab->start)));
  203.     if (! (cache->flags & SLAB_CACHE_SLINSIDE))
  204.         slab_free(slab_extern_cache, slab);
  205.  
  206.     atomic_dec(&cache->allocated_slabs);
  207.    
  208.     return 1 << cache->order;
  209. }
  210.  
  211. /** Map object to slab structure */
  212. static slab_t * obj2slab(void *obj)
  213. {
  214.     return (slab_t *)frame_get_parent(ADDR2PFN(KA2PA(obj)), 0);
  215. }
  216.  
  217. /**************************************/
  218. /* Slab functions */
  219.  
  220.  
  221. /**
  222.  * Return object to slab and call a destructor
  223.  *
  224.  * @param slab If the caller knows directly slab of the object, otherwise NULL
  225.  *
  226.  * @return Number of freed pages
  227.  */
  228. static count_t slab_obj_destroy(slab_cache_t *cache, void *obj,
  229.                 slab_t *slab)
  230. {
  231.     int freed = 0;
  232.  
  233.     if (!slab)
  234.         slab = obj2slab(obj);
  235.  
  236.     ASSERT(slab->cache == cache);
  237.  
  238.     if (cache->destructor)
  239.         freed = cache->destructor(obj);
  240.    
  241.     spinlock_lock(&cache->slablock);
  242.     ASSERT(slab->available < cache->objects);
  243.  
  244.     *((int *)obj) = slab->nextavail;
  245.     slab->nextavail = (obj - slab->start)/cache->size;
  246.     slab->available++;
  247.  
  248.     /* Move it to correct list */
  249.     if (slab->available == cache->objects) {
  250.         /* Free associated memory */
  251.         list_remove(&slab->link);
  252.         spinlock_unlock(&cache->slablock);
  253.  
  254.         return freed + slab_space_free(cache, slab);
  255.  
  256.     } else if (slab->available == 1) {
  257.         /* It was in full, move to partial */
  258.         list_remove(&slab->link);
  259.         list_prepend(&slab->link, &cache->partial_slabs);
  260.     }
  261.     spinlock_unlock(&cache->slablock);
  262.     return freed;
  263. }
  264.  
  265. /**
  266.  * Take new object from slab or create new if needed
  267.  *
  268.  * @return Object address or null
  269.  */
  270. static void * slab_obj_create(slab_cache_t *cache, int flags)
  271. {
  272.     slab_t *slab;
  273.     void *obj;
  274.  
  275.     spinlock_lock(&cache->slablock);
  276.  
  277.     if (list_empty(&cache->partial_slabs)) {
  278.         /* Allow recursion and reclaiming
  279.          * - this should work, as the slab control structures
  280.          *   are small and do not need to allocate with anything
  281.          *   other than frame_alloc when they are allocating,
  282.          *   that's why we should get recursion at most 1-level deep
  283.          */
  284.         spinlock_unlock(&cache->slablock);
  285.         slab = slab_space_alloc(cache, flags);
  286.         if (!slab)
  287.             return NULL;
  288.         spinlock_lock(&cache->slablock);
  289.     } else {
  290.         slab = list_get_instance(cache->partial_slabs.next,
  291.                      slab_t,
  292.                      link);
  293.         list_remove(&slab->link);
  294.     }
  295.     obj = slab->start + slab->nextavail * cache->size;
  296.     slab->nextavail = *((int *)obj);
  297.     slab->available--;
  298.  
  299.     if (! slab->available)
  300.         list_prepend(&slab->link, &cache->full_slabs);
  301.     else
  302.         list_prepend(&slab->link, &cache->partial_slabs);
  303.  
  304.     spinlock_unlock(&cache->slablock);
  305.  
  306.     if (cache->constructor && cache->constructor(obj, flags)) {
  307.         /* Bad, bad, construction failed */
  308.         slab_obj_destroy(cache, obj, slab);
  309.         return NULL;
  310.     }
  311.     return obj;
  312. }
  313.  
  314. /**************************************/
  315. /* CPU-Cache slab functions */
  316.  
  317. /**
  318.  * Finds a full magazine in cache, takes it from list
  319.  * and returns it
  320.  *
  321.  * @param first If true, return first, else last mag
  322.  */
  323. static slab_magazine_t * get_mag_from_cache(slab_cache_t *cache,
  324.                         int first)
  325. {
  326.     slab_magazine_t *mag = NULL;
  327.     link_t *cur;
  328.  
  329.     spinlock_lock(&cache->maglock);
  330.     if (!list_empty(&cache->magazines)) {
  331.         if (first)
  332.             cur = cache->magazines.next;
  333.         else
  334.             cur = cache->magazines.prev;
  335.         mag = list_get_instance(cur, slab_magazine_t, link);
  336.         list_remove(&mag->link);
  337.         atomic_dec(&cache->magazine_counter);
  338.     }
  339.     spinlock_unlock(&cache->maglock);
  340.     return mag;
  341. }
  342.  
  343. /** Prepend magazine to magazine list in cache */
  344. static void put_mag_to_cache(slab_cache_t *cache, slab_magazine_t *mag)
  345. {
  346.     spinlock_lock(&cache->maglock);
  347.  
  348.     list_prepend(&mag->link, &cache->magazines);
  349.     atomic_inc(&cache->magazine_counter);
  350.    
  351.     spinlock_unlock(&cache->maglock);
  352. }
  353.  
  354. /**
  355.  * Free all objects in magazine and free memory associated with magazine
  356.  *
  357.  * @return Number of freed pages
  358.  */
  359. static count_t magazine_destroy(slab_cache_t *cache,
  360.                 slab_magazine_t *mag)
  361. {
  362.     int i;
  363.     count_t frames = 0;
  364.  
  365.     for (i=0;i < mag->busy; i++) {
  366.         frames += slab_obj_destroy(cache, mag->objs[i], NULL);
  367.         atomic_dec(&cache->cached_objs);
  368.     }
  369.    
  370.     slab_free(&mag_cache, mag);
  371.  
  372.     return frames;
  373. }
  374.  
  375. /**
  376.  * Find full magazine, set it as current and return it
  377.  *
  378.  * Assume cpu_magazine lock is held
  379.  */
  380. static slab_magazine_t * get_full_current_mag(slab_cache_t *cache)
  381. {
  382.     slab_magazine_t *cmag, *lastmag, *newmag;
  383.  
  384.     cmag = cache->mag_cache[CPU->id].current;
  385.     lastmag = cache->mag_cache[CPU->id].last;
  386.     if (cmag) { /* First try local CPU magazines */
  387.         if (cmag->busy)
  388.             return cmag;
  389.  
  390.         if (lastmag && lastmag->busy) {
  391.             cache->mag_cache[CPU->id].current = lastmag;
  392.             cache->mag_cache[CPU->id].last = cmag;
  393.             return lastmag;
  394.         }
  395.     }
  396.     /* Local magazines are empty, import one from magazine list */
  397.     newmag = get_mag_from_cache(cache, 1);
  398.     if (!newmag)
  399.         return NULL;
  400.  
  401.     if (lastmag)
  402.         magazine_destroy(cache, lastmag);
  403.  
  404.     cache->mag_cache[CPU->id].last = cmag;
  405.     cache->mag_cache[CPU->id].current = newmag;
  406.     return newmag;
  407. }
  408.  
  409. /**
  410.  * Try to find object in CPU-cache magazines
  411.  *
  412.  * @return Pointer to object or NULL if not available
  413.  */
  414. static void * magazine_obj_get(slab_cache_t *cache)
  415. {
  416.     slab_magazine_t *mag;
  417.     void *obj;
  418.  
  419.     if (!CPU)
  420.         return NULL;
  421.  
  422.     spinlock_lock(&cache->mag_cache[CPU->id].lock);
  423.  
  424.     mag = get_full_current_mag(cache);
  425.     if (!mag) {
  426.         spinlock_unlock(&cache->mag_cache[CPU->id].lock);
  427.         return NULL;
  428.     }
  429.     obj = mag->objs[--mag->busy];
  430.     spinlock_unlock(&cache->mag_cache[CPU->id].lock);
  431.     atomic_dec(&cache->cached_objs);
  432.    
  433.     return obj;
  434. }
  435.  
  436. /**
  437.  * Assure that the current magazine is empty, return pointer to it, or NULL if
  438.  * no empty magazine is available and cannot be allocated
  439.  *
  440.  * Assume mag_cache[CPU->id].lock is held
  441.  *
  442.  * We have 2 magazines bound to processor.
  443.  * First try the current.
  444.  *  If full, try the last.
  445.  *   If full, put to magazines list.
  446.  *   allocate new, exchange last & current
  447.  *
  448.  */
  449. static slab_magazine_t * make_empty_current_mag(slab_cache_t *cache)
  450. {
  451.     slab_magazine_t *cmag,*lastmag,*newmag;
  452.  
  453.     cmag = cache->mag_cache[CPU->id].current;
  454.     lastmag = cache->mag_cache[CPU->id].last;
  455.  
  456.     if (cmag) {
  457.         if (cmag->busy < cmag->size)
  458.             return cmag;
  459.         if (lastmag && lastmag->busy < lastmag->size) {
  460.             cache->mag_cache[CPU->id].last = cmag;
  461.             cache->mag_cache[CPU->id].current = lastmag;
  462.             return lastmag;
  463.         }
  464.     }
  465.     /* current | last are full | nonexistent, allocate new */
  466.     /* We do not want to sleep just because of caching */
  467.     /* Especially we do not want reclaiming to start, as
  468.      * this would deadlock */
  469.     newmag = slab_alloc(&mag_cache, FRAME_ATOMIC | FRAME_NO_RECLAIM);
  470.     if (!newmag)
  471.         return NULL;
  472.     newmag->size = SLAB_MAG_SIZE;
  473.     newmag->busy = 0;
  474.  
  475.     /* Flush last to magazine list */
  476.     if (lastmag)
  477.         put_mag_to_cache(cache, lastmag);
  478.  
  479.     /* Move current as last, save new as current */
  480.     cache->mag_cache[CPU->id].last = cmag; 
  481.     cache->mag_cache[CPU->id].current = newmag;
  482.  
  483.     return newmag;
  484. }
  485.  
  486. /**
  487.  * Put object into CPU-cache magazine
  488.  *
  489.  * @return 0 - success, -1 - could not get memory
  490.  */
  491. static int magazine_obj_put(slab_cache_t *cache, void *obj)
  492. {
  493.     slab_magazine_t *mag;
  494.  
  495.     if (!CPU)
  496.         return -1;
  497.  
  498.     spinlock_lock(&cache->mag_cache[CPU->id].lock);
  499.  
  500.     mag = make_empty_current_mag(cache);
  501.     if (!mag) {
  502.         spinlock_unlock(&cache->mag_cache[CPU->id].lock);
  503.         return -1;
  504.     }
  505.    
  506.     mag->objs[mag->busy++] = obj;
  507.  
  508.     spinlock_unlock(&cache->mag_cache[CPU->id].lock);
  509.     atomic_inc(&cache->cached_objs);
  510.     return 0;
  511. }
  512.  
  513.  
  514. /**************************************/
  515. /* Slab cache functions */
  516.  
  517. /** Return number of objects that fit in certain cache size */
  518. static int comp_objects(slab_cache_t *cache)
  519. {
  520.     if (cache->flags & SLAB_CACHE_SLINSIDE)
  521.         return ((PAGE_SIZE << cache->order) - sizeof(slab_t)) / cache->size;
  522.     else
  523.         return (PAGE_SIZE << cache->order) / cache->size;
  524. }
  525.  
  526. /** Return wasted space in slab */
  527. static int badness(slab_cache_t *cache)
  528. {
  529.     int objects;
  530.     int ssize;
  531.  
  532.     objects = comp_objects(cache);
  533.     ssize = PAGE_SIZE << cache->order;
  534.     if (cache->flags & SLAB_CACHE_SLINSIDE)
  535.         ssize -= sizeof(slab_t);
  536.     return ssize - objects*cache->size;
  537. }
  538.  
  539. /**
  540.  * Initialize mag_cache structure in slab cache
  541.  */
  542. static void make_magcache(slab_cache_t *cache)
  543. {
  544.     int i;
  545.    
  546.     ASSERT(_slab_initialized >= 2);
  547.  
  548.     cache->mag_cache = malloc(sizeof(slab_mag_cache_t)*config.cpu_count,0);
  549.     for (i=0; i < config.cpu_count; i++) {
  550.         memsetb((__address)&cache->mag_cache[i],
  551.             sizeof(cache->mag_cache[i]), 0);
  552.         spinlock_initialize(&cache->mag_cache[i].lock,
  553.                     "slab_maglock_cpu");
  554.     }
  555. }
  556.  
  557. /** Initialize allocated memory as a slab cache */
  558. static void
  559. _slab_cache_create(slab_cache_t *cache,
  560.            char *name,
  561.            size_t size,
  562.            size_t align,
  563.            int (*constructor)(void *obj, int kmflag),
  564.            int (*destructor)(void *obj),
  565.            int flags)
  566. {
  567.     int pages;
  568.     ipl_t ipl;
  569.  
  570.     memsetb((__address)cache, sizeof(*cache), 0);
  571.     cache->name = name;
  572.  
  573.     if (align < sizeof(__native))
  574.         align = sizeof(__native);
  575.     size = ALIGN_UP(size, align);
  576.        
  577.     cache->size = size;
  578.  
  579.     cache->constructor = constructor;
  580.     cache->destructor = destructor;
  581.     cache->flags = flags;
  582.  
  583.     list_initialize(&cache->full_slabs);
  584.     list_initialize(&cache->partial_slabs);
  585.     list_initialize(&cache->magazines);
  586.     spinlock_initialize(&cache->slablock, "slab_lock");
  587.     spinlock_initialize(&cache->maglock, "slab_maglock");
  588.     if (! (cache->flags & SLAB_CACHE_NOMAGAZINE))
  589.         make_magcache(cache);
  590.  
  591.     /* Compute slab sizes, object counts in slabs etc. */
  592.     if (cache->size < SLAB_INSIDE_SIZE)
  593.         cache->flags |= SLAB_CACHE_SLINSIDE;
  594.  
  595.     /* Minimum slab order */
  596.     pages = ((cache->size-1) >> PAGE_WIDTH) + 1;
  597.     /* We need the 2^order >= pages */
  598.     if (pages == 1)
  599.         cache->order = 0;
  600.     else
  601.         cache->order = fnzb(pages-1)+1;
  602.  
  603.     while (badness(cache) > SLAB_MAX_BADNESS(cache)) {
  604.         cache->order += 1;
  605.     }
  606.     cache->objects = comp_objects(cache);
  607.     /* If info fits in, put it inside */
  608.     if (badness(cache) > sizeof(slab_t))
  609.         cache->flags |= SLAB_CACHE_SLINSIDE;
  610.  
  611.     /* Add cache to cache list */
  612.     ipl = interrupts_disable();
  613.     spinlock_lock(&slab_cache_lock);
  614.  
  615.     list_append(&cache->link, &slab_cache_list);
  616.  
  617.     spinlock_unlock(&slab_cache_lock);
  618.     interrupts_restore(ipl);
  619. }
  620.  
  621. /** Create slab cache  */
  622. slab_cache_t * slab_cache_create(char *name,
  623.                  size_t size,
  624.                  size_t align,
  625.                  int (*constructor)(void *obj, int kmflag),
  626.                  int (*destructor)(void *obj),
  627.                  int flags)
  628. {
  629.     slab_cache_t *cache;
  630.  
  631.     cache = slab_alloc(&slab_cache_cache, 0);
  632.     _slab_cache_create(cache, name, size, align, constructor, destructor,
  633.                flags);
  634.     return cache;
  635. }
  636.  
  637. /**
  638.  * Reclaim space occupied by objects that are already free
  639.  *
  640.  * @param flags If contains SLAB_RECLAIM_ALL, do aggressive freeing
  641.  * @return Number of freed pages
  642.  */
  643. static count_t _slab_reclaim(slab_cache_t *cache, int flags)
  644. {
  645.     int i;
  646.     slab_magazine_t *mag;
  647.     count_t frames = 0;
  648.     int magcount;
  649.    
  650.     if (cache->flags & SLAB_CACHE_NOMAGAZINE)
  651.         return 0; /* Nothing to do */
  652.  
  653.     /* We count up to original magazine count to avoid
  654.      * endless loop
  655.      */
  656.     magcount = atomic_get(&cache->magazine_counter);
  657.     while (magcount-- && (mag=get_mag_from_cache(cache,0))) {
  658.         frames += magazine_destroy(cache,mag);
  659.         if (!(flags & SLAB_RECLAIM_ALL) && frames)
  660.             break;
  661.     }
  662.    
  663.     if (flags & SLAB_RECLAIM_ALL) {
  664.         /* Free cpu-bound magazines */
  665.         /* Destroy CPU magazines */
  666.         for (i=0; i<config.cpu_count; i++) {
  667.             spinlock_lock(&cache->mag_cache[i].lock);
  668.  
  669.             mag = cache->mag_cache[i].current;
  670.             if (mag)
  671.                 frames += magazine_destroy(cache, mag);
  672.             cache->mag_cache[i].current = NULL;
  673.            
  674.             mag = cache->mag_cache[i].last;
  675.             if (mag)
  676.                 frames += magazine_destroy(cache, mag);
  677.             cache->mag_cache[i].last = NULL;
  678.  
  679.             spinlock_unlock(&cache->mag_cache[i].lock);
  680.         }
  681.     }
  682.  
  683.     return frames;
  684. }
  685.  
  686. /** Check that there are no slabs and remove cache from system  */
  687. void slab_cache_destroy(slab_cache_t *cache)
  688. {
  689.     ipl_t ipl;
  690.  
  691.     /* First remove cache from link, so that we don't need
  692.      * to disable interrupts later
  693.      */
  694.  
  695.     ipl = interrupts_disable();
  696.     spinlock_lock(&slab_cache_lock);
  697.  
  698.     list_remove(&cache->link);
  699.  
  700.     spinlock_unlock(&slab_cache_lock);
  701.     interrupts_restore(ipl);
  702.  
  703.     /* Do not lock anything, we assume the software is correct and
  704.      * does not touch the cache when it decides to destroy it */
  705.    
  706.     /* Destroy all magazines */
  707.     _slab_reclaim(cache, SLAB_RECLAIM_ALL);
  708.  
  709.     /* All slabs must be empty */
  710.     if (!list_empty(&cache->full_slabs) \
  711.         || !list_empty(&cache->partial_slabs))
  712.         panic("Destroying cache that is not empty.");
  713.  
  714.     if (!(cache->flags & SLAB_CACHE_NOMAGAZINE))
  715.         free(cache->mag_cache);
  716.     slab_free(&slab_cache_cache, cache);
  717. }
  718.  
  719. /** Allocate new object from cache - if no flags given, always returns
  720.     memory */
  721. void * slab_alloc(slab_cache_t *cache, int flags)
  722. {
  723.     ipl_t ipl;
  724.     void *result = NULL;
  725.    
  726.     /* Disable interrupts to avoid deadlocks with interrupt handlers */
  727.     ipl = interrupts_disable();
  728.  
  729.     if (!(cache->flags & SLAB_CACHE_NOMAGAZINE)) {
  730.         result = magazine_obj_get(cache);
  731.     }
  732.     if (!result)
  733.         result = slab_obj_create(cache, flags);
  734.  
  735.     interrupts_restore(ipl);
  736.  
  737.     if (result)
  738.         atomic_inc(&cache->allocated_objs);
  739.  
  740.     return result;
  741. }
  742.  
  743. /** Return object to cache, use slab if known  */
  744. static void _slab_free(slab_cache_t *cache, void *obj, slab_t *slab)
  745. {
  746.     ipl_t ipl;
  747.  
  748.     ipl = interrupts_disable();
  749.  
  750.     if ((cache->flags & SLAB_CACHE_NOMAGAZINE) \
  751.         || magazine_obj_put(cache, obj)) {
  752.  
  753.         slab_obj_destroy(cache, obj, slab);
  754.  
  755.     }
  756.     interrupts_restore(ipl);
  757.     atomic_dec(&cache->allocated_objs);
  758. }
  759.  
  760. /** Return slab object to cache */
  761. void slab_free(slab_cache_t *cache, void *obj)
  762. {
  763.     _slab_free(cache,obj,NULL);
  764. }
  765.  
  766. /* Go through all caches and reclaim what is possible */
  767. count_t slab_reclaim(int flags)
  768. {
  769.     slab_cache_t *cache;
  770.     link_t *cur;
  771.     count_t frames = 0;
  772.  
  773.     spinlock_lock(&slab_cache_lock);
  774.  
  775.     /* TODO: Add assert, that interrupts are disabled, otherwise
  776.      * memory allocation from interrupts can deadlock.
  777.      */
  778.  
  779.     for (cur = slab_cache_list.next;cur!=&slab_cache_list; cur=cur->next) {
  780.         cache = list_get_instance(cur, slab_cache_t, link);
  781.         frames += _slab_reclaim(cache, flags);
  782.     }
  783.  
  784.     spinlock_unlock(&slab_cache_lock);
  785.  
  786.     return frames;
  787. }
  788.  
  789.  
  790. /* Print list of slabs */
  791. void slab_print_list(void)
  792. {
  793.     slab_cache_t *cache;
  794.     link_t *cur;
  795.     ipl_t ipl;
  796.    
  797.     ipl = interrupts_disable();
  798.     spinlock_lock(&slab_cache_lock);
  799.     printf("slab name\t  Osize\t  Pages\t Obj/pg\t  Slabs\t Cached\tAllocobjs\tCtl\n");
  800.     for (cur = slab_cache_list.next;cur!=&slab_cache_list; cur=cur->next) {
  801.         cache = list_get_instance(cur, slab_cache_t, link);
  802.         printf("%s\t%7zd\t%7zd\t%7zd\t%7zd\t%7zd\t%7zd\t\t%s\n", cache->name, cache->size,
  803.                (1 << cache->order), cache->objects,
  804.                atomic_get(&cache->allocated_slabs),
  805.                atomic_get(&cache->cached_objs),
  806.                atomic_get(&cache->allocated_objs),
  807.                cache->flags & SLAB_CACHE_SLINSIDE ? "In" : "Out");
  808.     }
  809.     spinlock_unlock(&slab_cache_lock);
  810.     interrupts_restore(ipl);
  811. }
  812.  
  813. void slab_cache_init(void)
  814. {
  815.     int i, size;
  816.  
  817.     /* Initialize magazine cache */
  818.     _slab_cache_create(&mag_cache,
  819.                "slab_magazine",
  820.                sizeof(slab_magazine_t)+SLAB_MAG_SIZE*sizeof(void*),
  821.                sizeof(__address),
  822.                NULL, NULL,
  823.                SLAB_CACHE_NOMAGAZINE | SLAB_CACHE_SLINSIDE);
  824.     /* Initialize slab_cache cache */
  825.     _slab_cache_create(&slab_cache_cache,
  826.                "slab_cache",
  827.                sizeof(slab_cache_cache),
  828.                sizeof(__address),
  829.                NULL, NULL,
  830.                SLAB_CACHE_NOMAGAZINE | SLAB_CACHE_SLINSIDE);
  831.     /* Initialize external slab cache */
  832.     slab_extern_cache = slab_cache_create("slab_extern",
  833.                           sizeof(slab_t),
  834.                           0, NULL, NULL,
  835.                           SLAB_CACHE_SLINSIDE | SLAB_CACHE_MAGDEFERRED);
  836.  
  837.     /* Initialize structures for malloc */
  838.     for (i=0, size=(1<<SLAB_MIN_MALLOC_W);
  839.          i < (SLAB_MAX_MALLOC_W-SLAB_MIN_MALLOC_W+1);
  840.          i++, size <<= 1) {
  841.         malloc_caches[i] = slab_cache_create(malloc_names[i],
  842.                              size, 0,
  843.                              NULL,NULL, SLAB_CACHE_MAGDEFERRED);
  844.     }
  845. #ifdef CONFIG_DEBUG      
  846.     _slab_initialized = 1;
  847. #endif
  848. }
  849.  
  850. /** Enable cpu_cache
  851.  *
  852.  * Kernel calls this function, when it knows the real number of
  853.  * processors.
  854.  * Allocate slab for cpucache and enable it on all existing
  855.  * slabs that are SLAB_CACHE_MAGDEFERRED
  856.  */
  857. void slab_enable_cpucache(void)
  858. {
  859.     link_t *cur;
  860.     slab_cache_t *s;
  861.  
  862. #ifdef CONFIG_DEBUG
  863.     _slab_initialized = 2;
  864. #endif
  865.  
  866.     spinlock_lock(&slab_cache_lock);
  867.    
  868.     for (cur=slab_cache_list.next; cur != &slab_cache_list;cur=cur->next){
  869.         s = list_get_instance(cur, slab_cache_t, link);
  870.         if ((s->flags & SLAB_CACHE_MAGDEFERRED) != SLAB_CACHE_MAGDEFERRED)
  871.             continue;
  872.         make_magcache(s);
  873.         s->flags &= ~SLAB_CACHE_MAGDEFERRED;
  874.     }
  875.  
  876.     spinlock_unlock(&slab_cache_lock);
  877. }
  878.  
  879. /**************************************/
  880. /* kalloc/kfree functions             */
  881. void * malloc(unsigned int size, int flags)
  882. {
  883.     int idx;
  884.  
  885.     ASSERT(_slab_initialized);
  886.     ASSERT(size && size <= (1 << SLAB_MAX_MALLOC_W));
  887.    
  888.     if (size < (1 << SLAB_MIN_MALLOC_W))
  889.         size = (1 << SLAB_MIN_MALLOC_W);
  890.  
  891.     idx = fnzb(size-1) - SLAB_MIN_MALLOC_W + 1;
  892.  
  893.     return slab_alloc(malloc_caches[idx], flags);
  894. }
  895.  
  896. void free(void *obj)
  897. {
  898.     slab_t *slab;
  899.  
  900.     if (!obj) return;
  901.  
  902.     slab = obj2slab(obj);
  903.     _slab_free(slab->cache, obj, slab);
  904. }
  905.