Subversion Repositories HelenOS-historic

Rev

Rev 788 | Rev 791 | 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.  * The SLAB allocator is closely modelled after OpenSolaris SLAB allocator
  31.  * http://www.usenix.org/events/usenix01/full_papers/bonwick/bonwick_html/
  32.  *
  33.  * with the following exceptions:
  34.  *   - empty SLABS are deallocated immediately
  35.  *     (in Linux they are kept in linked list, in Solaris ???)
  36.  *   - empty magazines are deallocated when not needed
  37.  *     (in Solaris they are held in linked list in slab cache)
  38.  *
  39.  *   Following features are not currently supported but would be easy to do:
  40.  *   - cache coloring
  41.  *   - dynamic magazine growing (different magazine sizes are already
  42.  *     supported, but we would need to adjust allocating strategy)
  43.  *
  44.  * The SLAB allocator supports per-CPU caches ('magazines') to facilitate
  45.  * good SMP scaling.
  46.  *
  47.  * When a new object is being allocated, it is first checked, if it is
  48.  * available in CPU-bound magazine. If it is not found there, it is
  49.  * allocated from CPU-shared SLAB - if partial full is found, it is used,
  50.  * otherwise a new one is allocated.
  51.  *
  52.  * When an object is being deallocated, it is put to CPU-bound magazine.
  53.  * If there is no such magazine, new one is allocated (if it fails,
  54.  * the object is deallocated into SLAB). If the magazine is full, it is
  55.  * put into cpu-shared list of magazines and new one is allocated.
  56.  *
  57.  * The CPU-bound magazine is actually a pair of magazine to avoid
  58.  * thrashing when somebody is allocating/deallocating 1 item at the magazine
  59.  * size boundary. LIFO order is enforced, which should avoid fragmentation
  60.  * as much as possible.
  61.  *  
  62.  * Every cache contains list of full slabs and list of partialy full slabs.
  63.  * Empty SLABS are immediately freed (thrashing will be avoided because
  64.  * of magazines).
  65.  *
  66.  * The SLAB information structure is kept inside the data area, if possible.
  67.  * The cache can be marked that it should not use magazines. This is used
  68.  * only for SLAB related caches to avoid deadlocks and infinite recursion
  69.  * (the SLAB allocator uses itself for allocating all it's control structures).
  70.  *
  71.  * The SLAB allocator allocates lot of space and does not free it. When
  72.  * frame allocator fails to allocate the frame, it calls slab_reclaim().
  73.  * It tries 'light reclaim' first, then brutal reclaim. The light reclaim
  74.  * releases slabs from cpu-shared magazine-list, until at least 1 slab
  75.  * is deallocated in each cache (this algorithm should probably change).
  76.  * The brutal reclaim removes all cached objects, even from CPU-bound
  77.  * magazines.
  78.  *
  79.  * TODO: For better CPU-scaling the magazine allocation strategy should
  80.  * be extended. Currently, if the cache does not have magazine, it asks
  81.  * for non-cpu cached magazine cache to provide one. It might be feasible
  82.  * to add cpu-cached magazine cache (which would allocate it's magazines
  83.  * from non-cpu-cached mag. cache). This would provide a nice per-cpu
  84.  * buffer. The other possibility is to use the per-cache
  85.  * 'empty-magazine-list', which decreases competing for 1 per-system
  86.  * magazine cache.
  87.  *
  88.  * - it might be good to add granularity of locks even to slab level,
  89.  *   we could then try_spinlock over all partial slabs and thus improve
  90.  *   scalability even on slab level
  91.  */
  92.  
  93.  
  94. #include <synch/spinlock.h>
  95. #include <mm/slab.h>
  96. #include <adt/list.h>
  97. #include <memstr.h>
  98. #include <align.h>
  99. #include <mm/heap.h>
  100. #include <mm/frame.h>
  101. #include <config.h>
  102. #include <print.h>
  103. #include <arch.h>
  104. #include <panic.h>
  105. #include <debug.h>
  106. #include <bitops.h>
  107.  
  108. SPINLOCK_INITIALIZE(slab_cache_lock);
  109. static LIST_INITIALIZE(slab_cache_list);
  110.  
  111. /** Magazine cache */
  112. static slab_cache_t mag_cache;
  113. /** Cache for cache descriptors */
  114. static slab_cache_t slab_cache_cache;
  115. /** Cache for magcache structure from cache_t */
  116. static slab_cache_t *cpu_cache = NULL;
  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-8","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"
  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. /**************************************/
  143. /* SLAB allocation functions          */
  144.  
  145. /**
  146.  * Allocate frames for slab space and initialize
  147.  *
  148.  */
  149. static slab_t * slab_space_alloc(slab_cache_t *cache, int flags)
  150. {
  151.     void *data;
  152.     slab_t *slab;
  153.     size_t fsize;
  154.     int i;
  155.     zone_t *zone = NULL;
  156.     int status;
  157.     frame_t *frame;
  158.  
  159.     data = (void *)frame_alloc_rc_zone(cache->order, FRAME_KA | flags, &status, &zone);
  160.     if (status != FRAME_OK) {
  161.         return NULL;
  162.     }
  163.     if (! (cache->flags & SLAB_CACHE_SLINSIDE)) {
  164.         slab = slab_alloc(slab_extern_cache, flags);
  165.         if (!slab) {
  166.             frame_free((__address)data);
  167.             return NULL;
  168.         }
  169.     } else {
  170.         fsize = (PAGE_SIZE << cache->order);
  171.         slab = data + fsize - sizeof(*slab);
  172.     }
  173.        
  174.     /* Fill in slab structures */
  175.     /* TODO: some better way of accessing the frame */
  176.     for (i=0; i < (1 << cache->order); i++) {
  177.         frame = ADDR2FRAME(zone, KA2PA((__address)(data+i*PAGE_SIZE)));
  178.         frame->parent = slab;
  179.     }
  180.  
  181.     slab->start = data;
  182.     slab->available = cache->objects;
  183.     slab->nextavail = 0;
  184.     slab->cache = cache;
  185.  
  186.     for (i=0; i<cache->objects;i++)
  187.         *((int *) (slab->start + i*cache->size)) = i+1;
  188.  
  189.     atomic_inc(&cache->allocated_slabs);
  190.     return slab;
  191. }
  192.  
  193. /**
  194.  * Deallocate space associated with SLAB
  195.  *
  196.  * @return number of freed frames
  197.  */
  198. static count_t slab_space_free(slab_cache_t *cache, slab_t *slab)
  199. {
  200.     frame_free((__address)slab->start);
  201.     if (! (cache->flags & SLAB_CACHE_SLINSIDE))
  202.         slab_free(slab_extern_cache, slab);
  203.  
  204.     atomic_dec(&cache->allocated_slabs);
  205.    
  206.     return 1 << cache->order;
  207. }
  208.  
  209. /** Map object to slab structure */
  210. static slab_t * obj2slab(void *obj)
  211. {
  212.     frame_t *frame;
  213.  
  214.     frame = frame_addr2frame((__address)obj);
  215.     return (slab_t *)frame->parent;
  216. }
  217.  
  218. /**************************************/
  219. /* SLAB functions */
  220.  
  221.  
  222. /**
  223.  * Return object to slab and call a destructor
  224.  *
  225.  * @param slab If the caller knows directly slab of the object, otherwise NULL
  226.  *
  227.  * @return Number of freed pages
  228.  */
  229. static count_t slab_obj_destroy(slab_cache_t *cache, void *obj,
  230.                 slab_t *slab)
  231. {
  232.     int freed = 0;
  233.  
  234.     if (!slab)
  235.         slab = obj2slab(obj);
  236.  
  237.     ASSERT(slab->cache == cache);
  238.  
  239.     if (cache->destructor)
  240.         freed = cache->destructor(obj);
  241.    
  242.     spinlock_lock(&cache->slablock);
  243.     ASSERT(slab->available < cache->objects);
  244.  
  245.     *((int *)obj) = slab->nextavail;
  246.     slab->nextavail = (obj - slab->start)/cache->size;
  247.     slab->available++;
  248.  
  249.     /* Move it to correct list */
  250.     if (slab->available == cache->objects) {
  251.         /* Free associated memory */
  252.         list_remove(&slab->link);
  253.         spinlock_unlock(&cache->slablock);
  254.  
  255.         return freed + slab_space_free(cache, slab);
  256.  
  257.     } else if (slab->available == 1) {
  258.         /* It was in full, move to partial */
  259.         list_remove(&slab->link);
  260.         list_prepend(&slab->link, &cache->partial_slabs);
  261.     }
  262.     spinlock_unlock(&cache->slablock);
  263.     return freed;
  264. }
  265.  
  266. /**
  267.  * Take new object from slab or create new if needed
  268.  *
  269.  * @return Object address or null
  270.  */
  271. static void * slab_obj_create(slab_cache_t *cache, int flags)
  272. {
  273.     slab_t *slab;
  274.     void *obj;
  275.  
  276.     spinlock_lock(&cache->slablock);
  277.  
  278.     if (list_empty(&cache->partial_slabs)) {
  279.         /* Allow recursion and reclaiming
  280.          * - this should work, as the SLAB control structures
  281.          *   are small and do not need to allocte with anything
  282.          *   other ten frame_alloc when they are allocating,
  283.          *   that's why we should get recursion at most 1-level deep
  284.          */
  285.         spinlock_unlock(&cache->slablock);
  286.         slab = slab_space_alloc(cache, flags);
  287.         if (!slab)
  288.             return NULL;
  289.         spinlock_lock(&cache->slablock);
  290.     } else {
  291.         slab = list_get_instance(cache->partial_slabs.next,
  292.                      slab_t,
  293.                      link);
  294.         list_remove(&slab->link);
  295.     }
  296.     obj = slab->start + slab->nextavail * cache->size;
  297.     slab->nextavail = *((int *)obj);
  298.     slab->available--;
  299.  
  300.     if (! slab->available)
  301.         list_prepend(&slab->link, &cache->full_slabs);
  302.     else
  303.         list_prepend(&slab->link, &cache->partial_slabs);
  304.  
  305.     spinlock_unlock(&cache->slablock);
  306.  
  307.     if (cache->constructor && cache->constructor(obj, flags)) {
  308.         /* Bad, bad, construction failed */
  309.         slab_obj_destroy(cache, obj, slab);
  310.         return NULL;
  311.     }
  312.     return obj;
  313. }
  314.  
  315. /**************************************/
  316. /* CPU-Cache slab functions */
  317.  
  318. /**
  319.  * Finds a full magazine in cache, takes it from list
  320.  * and returns it
  321.  *
  322.  * @param first If true, return first, else last mag
  323.  */
  324. static slab_magazine_t * get_mag_from_cache(slab_cache_t *cache,
  325.                         int first)
  326. {
  327.     slab_magazine_t *mag = NULL;
  328.     link_t *cur;
  329.  
  330.     spinlock_lock(&cache->maglock);
  331.     if (!list_empty(&cache->magazines)) {
  332.         if (first)
  333.             cur = cache->magazines.next;
  334.         else
  335.             cur = cache->magazines.prev;
  336.         mag = list_get_instance(cur, slab_magazine_t, link);
  337.         list_remove(&mag->link);
  338.         atomic_dec(&cache->magazine_counter);
  339.     }
  340.     spinlock_unlock(&cache->maglock);
  341.     return mag;
  342. }
  343.  
  344. /** Prepend magazine to magazine list in cache */
  345. static void put_mag_to_cache(slab_cache_t *cache, slab_magazine_t *mag)
  346. {
  347.     spinlock_lock(&cache->maglock);
  348.  
  349.     list_prepend(&mag->link, &cache->magazines);
  350.     atomic_inc(&cache->magazine_counter);
  351.    
  352.     spinlock_unlock(&cache->maglock);
  353. }
  354.  
  355. /**
  356.  * Free all objects in magazine and free memory associated with magazine
  357.  *
  358.  * @return Number of freed pages
  359.  */
  360. static count_t magazine_destroy(slab_cache_t *cache,
  361.                 slab_magazine_t *mag)
  362. {
  363.     int i;
  364.     count_t frames = 0;
  365.  
  366.     for (i=0;i < mag->busy; i++) {
  367.         frames += slab_obj_destroy(cache, mag->objs[i], NULL);
  368.         atomic_dec(&cache->cached_objs);
  369.     }
  370.    
  371.     slab_free(&mag_cache, mag);
  372.  
  373.     return frames;
  374. }
  375.  
  376. /**
  377.  * Find full magazine, set it as current and return it
  378.  *
  379.  * Assume cpu_magazine lock is held
  380.  */
  381. static slab_magazine_t * get_full_current_mag(slab_cache_t *cache)
  382. {
  383.     slab_magazine_t *cmag, *lastmag, *newmag;
  384.  
  385.     cmag = cache->mag_cache[CPU->id].current;
  386.     lastmag = cache->mag_cache[CPU->id].last;
  387.     if (cmag) { /* First try local CPU magazines */
  388.         if (cmag->busy)
  389.             return cmag;
  390.  
  391.         if (lastmag && lastmag->busy) {
  392.             cache->mag_cache[CPU->id].current = lastmag;
  393.             cache->mag_cache[CPU->id].last = cmag;
  394.             return lastmag;
  395.         }
  396.     }
  397.     /* Local magazines are empty, import one from magazine list */
  398.     newmag = get_mag_from_cache(cache, 1);
  399.     if (!newmag)
  400.         return NULL;
  401.  
  402.     if (lastmag)
  403.         magazine_destroy(cache, lastmag);
  404.  
  405.     cache->mag_cache[CPU->id].last = cmag;
  406.     cache->mag_cache[CPU->id].current = newmag;
  407.     return newmag;
  408. }
  409.  
  410. /**
  411.  * Try to find object in CPU-cache magazines
  412.  *
  413.  * @return Pointer to object or NULL if not available
  414.  */
  415. static void * magazine_obj_get(slab_cache_t *cache)
  416. {
  417.     slab_magazine_t *mag;
  418.     void *obj;
  419.  
  420.     if (!CPU)
  421.         return NULL;
  422.  
  423.     spinlock_lock(&cache->mag_cache[CPU->id].lock);
  424.  
  425.     mag = get_full_current_mag(cache);
  426.     if (!mag) {
  427.         spinlock_unlock(&cache->mag_cache[CPU->id].lock);
  428.         return NULL;
  429.     }
  430.     obj = mag->objs[--mag->busy];
  431.     spinlock_unlock(&cache->mag_cache[CPU->id].lock);
  432.     atomic_dec(&cache->cached_objs);
  433.    
  434.     return obj;
  435. }
  436.  
  437. /**
  438.  * Assure that the current magazine is empty, return pointer to it, or NULL if
  439.  * no empty magazine is available and cannot be allocated
  440.  *
  441.  * Assume mag_cache[CPU->id].lock is held
  442.  *
  443.  * We have 2 magazines bound to processor.
  444.  * First try the current.
  445.  *  If full, try the last.
  446.  *   If full, put to magazines list.
  447.  *   allocate new, exchange last & current
  448.  *
  449.  */
  450. static slab_magazine_t * make_empty_current_mag(slab_cache_t *cache)
  451. {
  452.     slab_magazine_t *cmag,*lastmag,*newmag;
  453.  
  454.     cmag = cache->mag_cache[CPU->id].current;
  455.     lastmag = cache->mag_cache[CPU->id].last;
  456.  
  457.     if (cmag) {
  458.         if (cmag->busy < cmag->size)
  459.             return cmag;
  460.         if (lastmag && lastmag->busy < lastmag->size) {
  461.             cache->mag_cache[CPU->id].last = cmag;
  462.             cache->mag_cache[CPU->id].current = lastmag;
  463.             return lastmag;
  464.         }
  465.     }
  466.     /* current | last are full | nonexistent, allocate new */
  467.     /* We do not want to sleep just because of caching */
  468.     /* Especially we do not want reclaiming to start, as
  469.      * this would deadlock */
  470.     newmag = slab_alloc(&mag_cache, FRAME_ATOMIC | FRAME_NO_RECLAIM);
  471.     if (!newmag)
  472.         return NULL;
  473.     newmag->size = SLAB_MAG_SIZE;
  474.     newmag->busy = 0;
  475.  
  476.     /* Flush last to magazine list */
  477.     if (lastmag)
  478.         put_mag_to_cache(cache, lastmag);
  479.  
  480.     /* Move current as last, save new as current */
  481.     cache->mag_cache[CPU->id].last = cmag; 
  482.     cache->mag_cache[CPU->id].current = newmag;
  483.  
  484.     return newmag;
  485. }
  486.  
  487. /**
  488.  * Put object into CPU-cache magazine
  489.  *
  490.  * @return 0 - success, -1 - could not get memory
  491.  */
  492. static int magazine_obj_put(slab_cache_t *cache, void *obj)
  493. {
  494.     slab_magazine_t *mag;
  495.  
  496.     if (!CPU)
  497.         return -1;
  498.  
  499.     spinlock_lock(&cache->mag_cache[CPU->id].lock);
  500.  
  501.     mag = make_empty_current_mag(cache);
  502.     if (!mag) {
  503.         spinlock_unlock(&cache->mag_cache[CPU->id].lock);
  504.         return -1;
  505.     }
  506.    
  507.     mag->objs[mag->busy++] = obj;
  508.  
  509.     spinlock_unlock(&cache->mag_cache[CPU->id].lock);
  510.     atomic_inc(&cache->cached_objs);
  511.     return 0;
  512. }
  513.  
  514.  
  515. /**************************************/
  516. /* SLAB CACHE functions */
  517.  
  518. /** Return number of objects that fit in certain cache size */
  519. static int comp_objects(slab_cache_t *cache)
  520. {
  521.     if (cache->flags & SLAB_CACHE_SLINSIDE)
  522.         return ((PAGE_SIZE << cache->order) - sizeof(slab_t)) / cache->size;
  523.     else
  524.         return (PAGE_SIZE << cache->order) / cache->size;
  525. }
  526.  
  527. /** Return wasted space in slab */
  528. static int badness(slab_cache_t *cache)
  529. {
  530.     int objects;
  531.     int ssize;
  532.  
  533.     objects = comp_objects(cache);
  534.     ssize = PAGE_SIZE << cache->order;
  535.     if (cache->flags & SLAB_CACHE_SLINSIDE)
  536.         ssize -= sizeof(slab_t);
  537.     return ssize - objects*cache->size;
  538. }
  539.  
  540. /**
  541.  * Initialize mag_cache structure in slab cache
  542.  */
  543. static void make_magcache(slab_cache_t *cache)
  544. {
  545.     int i;
  546.  
  547.     ASSERT(cpu_cache);
  548.     cache->mag_cache = slab_alloc(cpu_cache, 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.     cache->order = fnzb(pages);
  598.  
  599.     while (badness(cache) > SLAB_MAX_BADNESS(cache)) {
  600.         cache->order += 1;
  601.     }
  602.     cache->objects = comp_objects(cache);
  603.     /* If info fits in, put it inside */
  604.     if (badness(cache) > sizeof(slab_t))
  605.         cache->flags |= SLAB_CACHE_SLINSIDE;
  606.  
  607.     /* Add cache to cache list */
  608.     ipl = interrupts_disable();
  609.     spinlock_lock(&slab_cache_lock);
  610.  
  611.     list_append(&cache->link, &slab_cache_list);
  612.  
  613.     spinlock_unlock(&slab_cache_lock);
  614.     interrupts_restore(ipl);
  615. }
  616.  
  617. /** Create slab cache  */
  618. slab_cache_t * slab_cache_create(char *name,
  619.                  size_t size,
  620.                  size_t align,
  621.                  int (*constructor)(void *obj, int kmflag),
  622.                  int (*destructor)(void *obj),
  623.                  int flags)
  624. {
  625.     slab_cache_t *cache;
  626.  
  627.     cache = slab_alloc(&slab_cache_cache, 0);
  628.     _slab_cache_create(cache, name, size, align, constructor, destructor,
  629.                flags);
  630.     return cache;
  631. }
  632.  
  633. /**
  634.  * Reclaim space occupied by objects that are already free
  635.  *
  636.  * @param flags If contains SLAB_RECLAIM_ALL, do aggressive freeing
  637.  * @return Number of freed pages
  638.  */
  639. static count_t _slab_reclaim(slab_cache_t *cache, int flags)
  640. {
  641.     int i;
  642.     slab_magazine_t *mag;
  643.     count_t frames = 0;
  644.     int magcount;
  645.    
  646.     if (cache->flags & SLAB_CACHE_NOMAGAZINE)
  647.         return 0; /* Nothing to do */
  648.  
  649.     /* We count up to original magazine count to avoid
  650.      * endless loop
  651.      */
  652.     magcount = atomic_get(&cache->magazine_counter);
  653.     while (magcount-- && (mag=get_mag_from_cache(cache,0))) {
  654.         frames += magazine_destroy(cache,mag);
  655.         if (!(flags & SLAB_RECLAIM_ALL) && frames)
  656.             break;
  657.     }
  658.    
  659.     if (flags & SLAB_RECLAIM_ALL) {
  660.         /* Free cpu-bound magazines */
  661.         /* Destroy CPU magazines */
  662.         for (i=0; i<config.cpu_count; i++) {
  663.             spinlock_lock(&cache->mag_cache[i].lock);
  664.  
  665.             mag = cache->mag_cache[i].current;
  666.             if (mag)
  667.                 frames += magazine_destroy(cache, mag);
  668.             cache->mag_cache[i].current = NULL;
  669.            
  670.             mag = cache->mag_cache[i].last;
  671.             if (mag)
  672.                 frames += magazine_destroy(cache, mag);
  673.             cache->mag_cache[i].last = NULL;
  674.  
  675.             spinlock_unlock(&cache->mag_cache[i].lock);
  676.         }
  677.     }
  678.  
  679.     return frames;
  680. }
  681.  
  682. /** Check that there are no slabs and remove cache from system  */
  683. void slab_cache_destroy(slab_cache_t *cache)
  684. {
  685.     ipl_t ipl;
  686.  
  687.     /* First remove cache from link, so that we don't need
  688.      * to disable interrupts later
  689.      */
  690.  
  691.     ipl = interrupts_disable();
  692.     spinlock_lock(&slab_cache_lock);
  693.  
  694.     list_remove(&cache->link);
  695.  
  696.     spinlock_unlock(&slab_cache_lock);
  697.     interrupts_restore(ipl);
  698.  
  699.     /* Do not lock anything, we assume the software is correct and
  700.      * does not touch the cache when it decides to destroy it */
  701.    
  702.     /* Destroy all magazines */
  703.     _slab_reclaim(cache, SLAB_RECLAIM_ALL);
  704.  
  705.     /* All slabs must be empty */
  706.     if (!list_empty(&cache->full_slabs) \
  707.         || !list_empty(&cache->partial_slabs))
  708.         panic("Destroying cache that is not empty.");
  709.  
  710.     if (!(cache->flags & SLAB_CACHE_NOMAGAZINE))
  711.         slab_free(cpu_cache, cache->mag_cache);
  712.     slab_free(&slab_cache_cache, cache);
  713. }
  714.  
  715. /** Allocate new object from cache - if no flags given, always returns
  716.     memory */
  717. void * slab_alloc(slab_cache_t *cache, int flags)
  718. {
  719.     ipl_t ipl;
  720.     void *result = NULL;
  721.    
  722.     /* Disable interrupts to avoid deadlocks with interrupt handlers */
  723.     ipl = interrupts_disable();
  724.  
  725.     if (!(cache->flags & SLAB_CACHE_NOMAGAZINE))
  726.         result = magazine_obj_get(cache);
  727.     if (!result)
  728.         result = slab_obj_create(cache, flags);
  729.  
  730.     interrupts_restore(ipl);
  731.  
  732.     if (result)
  733.         atomic_inc(&cache->allocated_objs);
  734.  
  735.     return result;
  736. }
  737.  
  738. /** Return object to cache, use slab if known  */
  739. static void _slab_free(slab_cache_t *cache, void *obj, slab_t *slab)
  740. {
  741.     ipl_t ipl;
  742.  
  743.     ipl = interrupts_disable();
  744.  
  745.     if ((cache->flags & SLAB_CACHE_NOMAGAZINE) \
  746.         || magazine_obj_put(cache, obj)) {
  747.  
  748.         slab_obj_destroy(cache, obj, slab);
  749.  
  750.     }
  751.     interrupts_restore(ipl);
  752.     atomic_dec(&cache->allocated_objs);
  753. }
  754.  
  755. /** Return slab object to cache */
  756. void slab_free(slab_cache_t *cache, void *obj)
  757. {
  758.     _slab_free(cache,obj,NULL);
  759. }
  760.  
  761. /* Go through all caches and reclaim what is possible */
  762. count_t slab_reclaim(int flags)
  763. {
  764.     slab_cache_t *cache;
  765.     link_t *cur;
  766.     count_t frames = 0;
  767.  
  768.     spinlock_lock(&slab_cache_lock);
  769.  
  770.     /* TODO: Add assert, that interrupts are disabled, otherwise
  771.      * memory allocation from interrupts can deadlock.
  772.      */
  773.  
  774.     for (cur = slab_cache_list.next;cur!=&slab_cache_list; cur=cur->next) {
  775.         cache = list_get_instance(cur, slab_cache_t, link);
  776.         frames += _slab_reclaim(cache, flags);
  777.     }
  778.  
  779.     spinlock_unlock(&slab_cache_lock);
  780.  
  781.     return frames;
  782. }
  783.  
  784.  
  785. /* Print list of slabs */
  786. void slab_print_list(void)
  787. {
  788.     slab_cache_t *cache;
  789.     link_t *cur;
  790.     ipl_t ipl;
  791.    
  792.     ipl = interrupts_disable();
  793.     spinlock_lock(&slab_cache_lock);
  794.     printf("SLAB name\tOsize\tPages\tObj/pg\tSlabs\tCached\tAllocobjs\tCtl\n");
  795.     for (cur = slab_cache_list.next;cur!=&slab_cache_list; cur=cur->next) {
  796.         cache = list_get_instance(cur, slab_cache_t, link);
  797.         printf("%s\t%d\t%d\t%d\t%d\t%d\t%d\t\t%s\n", cache->name, cache->size,
  798.                (1 << cache->order), cache->objects,
  799.                atomic_get(&cache->allocated_slabs),
  800.                atomic_get(&cache->cached_objs),
  801.                atomic_get(&cache->allocated_objs),
  802.                cache->flags & SLAB_CACHE_SLINSIDE ? "In" : "Out");
  803.     }
  804.     spinlock_unlock(&slab_cache_lock);
  805.     interrupts_restore(ipl);
  806. }
  807.  
  808. #ifdef CONFIG_DEBUG
  809. static int _slab_initialized = 0;
  810. #endif
  811.  
  812. void slab_cache_init(void)
  813. {
  814.     int i, size;
  815.  
  816.     /* Initialize magazine cache */
  817.     _slab_cache_create(&mag_cache,
  818.                "slab_magazine",
  819.                sizeof(slab_magazine_t)+SLAB_MAG_SIZE*sizeof(void*),
  820.                sizeof(__address),
  821.                NULL, NULL,
  822.                SLAB_CACHE_NOMAGAZINE | SLAB_CACHE_SLINSIDE);
  823.     /* Initialize slab_cache cache */
  824.     _slab_cache_create(&slab_cache_cache,
  825.                "slab_cache",
  826.                sizeof(slab_cache_cache),
  827.                sizeof(__address),
  828.                NULL, NULL,
  829.                SLAB_CACHE_NOMAGAZINE | SLAB_CACHE_SLINSIDE);
  830.     /* Initialize external slab cache */
  831.     slab_extern_cache = slab_cache_create("slab_extern",
  832.                           sizeof(slab_t),
  833.                           0, NULL, NULL,
  834.                           SLAB_CACHE_SLINSIDE | SLAB_CACHE_MAGDEFERRED);
  835.  
  836.     /* Initialize structures for malloc */
  837.     for (i=0, size=(1<<SLAB_MIN_MALLOC_W);
  838.          i < (SLAB_MAX_MALLOC_W-SLAB_MIN_MALLOC_W+1);
  839.          i++, size <<= 1) {
  840.         malloc_caches[i] = slab_cache_create(malloc_names[i],
  841.                              size, 0,
  842.                              NULL,NULL, SLAB_CACHE_MAGDEFERRED);
  843.     }
  844. #ifdef CONFIG_DEBUG      
  845.     _slab_initialized = 1;
  846. #endif
  847. }
  848.  
  849. /** Enable cpu_cache
  850.  *
  851.  * Kernel calls this function, when it knows the real number of
  852.  * processors.
  853.  * Allocate slab for cpucache and enable it on all existing
  854.  * slabs that are SLAB_CACHE_MAGDEFERRED
  855.  */
  856. void slab_enable_cpucache(void)
  857. {
  858.     link_t *cur;
  859.     slab_cache_t *s;
  860.  
  861.     cpu_cache = slab_cache_create("magcpucache",
  862.                       sizeof(slab_mag_cache_t) * config.cpu_count,
  863.                       0, NULL, NULL,
  864.                       SLAB_CACHE_NOMAGAZINE);
  865.     spinlock_lock(&slab_cache_lock);
  866.    
  867.     for (cur=slab_cache_list.next; cur != &slab_cache_list;cur=cur->next){
  868.         s = list_get_instance(cur, slab_cache_t, link);
  869.         if ((s->flags & SLAB_CACHE_MAGDEFERRED) != SLAB_CACHE_MAGDEFERRED)
  870.             continue;
  871.         make_magcache(s);
  872.         s->flags &= ~SLAB_CACHE_MAGDEFERRED;
  873.     }
  874.  
  875.     spinlock_unlock(&slab_cache_lock);
  876. }
  877.  
  878. /**************************************/
  879. /* kalloc/kfree functions             */
  880. void * kalloc(unsigned int size, int flags)
  881. {
  882.     int idx;
  883.  
  884.     ASSERT(_slab_initialized);
  885.     ASSERT( size && size <= (1 << SLAB_MAX_MALLOC_W));
  886.    
  887.     if (size < (1 << SLAB_MIN_MALLOC_W))
  888.         size = (1 << SLAB_MIN_MALLOC_W);
  889.  
  890.     idx = fnzb(size-1) - SLAB_MIN_MALLOC_W + 1;
  891.  
  892.     return slab_alloc(malloc_caches[idx], flags);
  893. }
  894.  
  895.  
  896. void kfree(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.