staging: zsmalloc: zsmalloc memory allocation library
authorNitin Gupta <ngupta@vflare.org>
Mon, 9 Jan 2012 22:51:56 +0000 (16:51 -0600)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Thu, 9 Feb 2012 01:12:53 +0000 (17:12 -0800)
This patch creates a new memory allocation library named
zsmalloc.

NOTE: zsmalloc currently depends on SPARSEMEM for the MAX_PHYSMEM_BITS
value needed to determine the format of the object handle. There may
be a better way to do this.  Feedback is welcome.

Signed-off-by: Nitin Gupta <ngupta@vflare.org>
Signed-off-by: Seth Jennings <sjenning@linux.vnet.ibm.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
drivers/staging/zsmalloc/Kconfig [new file with mode: 0644]
drivers/staging/zsmalloc/Makefile [new file with mode: 0644]
drivers/staging/zsmalloc/zsmalloc-main.c [new file with mode: 0644]
drivers/staging/zsmalloc/zsmalloc.h [new file with mode: 0644]
drivers/staging/zsmalloc/zsmalloc_int.h [new file with mode: 0644]

diff --git a/drivers/staging/zsmalloc/Kconfig b/drivers/staging/zsmalloc/Kconfig
new file mode 100644 (file)
index 0000000..3e7a8d4
--- /dev/null
@@ -0,0 +1,11 @@
+config ZSMALLOC
+       tristate "Memory allocator for compressed pages"
+       depends on SPARSEMEM
+       default n
+       help
+         zsmalloc is a slab-based memory allocator designed to store
+         compressed RAM pages.  zsmalloc uses virtual memory mapping
+         in order to reduce fragmentation.  However, this results in a
+         non-standard allocator interface where a handle, not a pointer, is
+         returned by an alloc().  This handle must be mapped in order to
+         access the allocated space.
diff --git a/drivers/staging/zsmalloc/Makefile b/drivers/staging/zsmalloc/Makefile
new file mode 100644 (file)
index 0000000..b134848
--- /dev/null
@@ -0,0 +1,3 @@
+zsmalloc-y             := zsmalloc-main.o
+
+obj-$(CONFIG_ZSMALLOC) += zsmalloc.o
diff --git a/drivers/staging/zsmalloc/zsmalloc-main.c b/drivers/staging/zsmalloc/zsmalloc-main.c
new file mode 100644 (file)
index 0000000..189fb42
--- /dev/null
@@ -0,0 +1,756 @@
+/*
+ * zsmalloc memory allocator
+ *
+ * Copyright (C) 2011  Nitin Gupta
+ *
+ * This code is released using a dual license strategy: BSD/GPL
+ * You can choose the license that better fits your requirements.
+ *
+ * Released under the terms of 3-clause BSD License
+ * Released under the terms of GNU General Public License Version 2.0
+ */
+
+#ifdef CONFIG_ZSMALLOC_DEBUG
+#define DEBUG
+#endif
+
+#include <linux/module.h>
+#include <linux/kernel.h>
+#include <linux/bitops.h>
+#include <linux/errno.h>
+#include <linux/highmem.h>
+#include <linux/init.h>
+#include <linux/string.h>
+#include <linux/slab.h>
+#include <asm/tlbflush.h>
+#include <asm/pgtable.h>
+#include <linux/cpumask.h>
+#include <linux/cpu.h>
+
+#include "zsmalloc.h"
+#include "zsmalloc_int.h"
+
+/*
+ * A zspage's class index and fullness group
+ * are encoded in its (first)page->mapping
+ */
+#define CLASS_IDX_BITS 28
+#define FULLNESS_BITS  4
+#define CLASS_IDX_MASK ((1 << CLASS_IDX_BITS) - 1)
+#define FULLNESS_MASK  ((1 << FULLNESS_BITS) - 1)
+
+/*
+ * Object location (<PFN>, <obj_idx>) is encoded as
+ * as single (void *) handle value.
+ *
+ * Note that object index <obj_idx> is relative to system
+ * page <PFN> it is stored in, so for each sub-page belonging
+ * to a zspage, obj_idx starts with 0.
+ */
+#define _PFN_BITS              (MAX_PHYSMEM_BITS - PAGE_SHIFT)
+#define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS)
+#define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
+
+/* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
+static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
+
+static int is_first_page(struct page *page)
+{
+       return test_bit(PG_private, &page->flags);
+}
+
+static int is_last_page(struct page *page)
+{
+       return test_bit(PG_private_2, &page->flags);
+}
+
+static void get_zspage_mapping(struct page *page, unsigned int *class_idx,
+                               enum fullness_group *fullness)
+{
+       unsigned long m;
+       BUG_ON(!is_first_page(page));
+
+       m = (unsigned long)page->mapping;
+       *fullness = m & FULLNESS_MASK;
+       *class_idx = (m >> FULLNESS_BITS) & CLASS_IDX_MASK;
+}
+
+static void set_zspage_mapping(struct page *page, unsigned int class_idx,
+                               enum fullness_group fullness)
+{
+       unsigned long m;
+       BUG_ON(!is_first_page(page));
+
+       m = ((class_idx & CLASS_IDX_MASK) << FULLNESS_BITS) |
+                       (fullness & FULLNESS_MASK);
+       page->mapping = (struct address_space *)m;
+}
+
+static int get_size_class_index(int size)
+{
+       int idx = 0;
+
+       if (likely(size > ZS_MIN_ALLOC_SIZE))
+               idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
+                               ZS_SIZE_CLASS_DELTA);
+
+       return idx;
+}
+
+static enum fullness_group get_fullness_group(struct page *page)
+{
+       int inuse, max_objects;
+       enum fullness_group fg;
+       BUG_ON(!is_first_page(page));
+
+       inuse = page->inuse;
+       max_objects = page->objects;
+
+       if (inuse == 0)
+               fg = ZS_EMPTY;
+       else if (inuse == max_objects)
+               fg = ZS_FULL;
+       else if (inuse <= max_objects / fullness_threshold_frac)
+               fg = ZS_ALMOST_EMPTY;
+       else
+               fg = ZS_ALMOST_FULL;
+
+       return fg;
+}
+
+static void insert_zspage(struct page *page, struct size_class *class,
+                               enum fullness_group fullness)
+{
+       struct page **head;
+
+       BUG_ON(!is_first_page(page));
+
+       if (fullness >= _ZS_NR_FULLNESS_GROUPS)
+               return;
+
+       head = &class->fullness_list[fullness];
+       if (*head)
+               list_add_tail(&page->lru, &(*head)->lru);
+
+       *head = page;
+}
+
+static void remove_zspage(struct page *page, struct size_class *class,
+                               enum fullness_group fullness)
+{
+       struct page **head;
+
+       BUG_ON(!is_first_page(page));
+
+       if (fullness >= _ZS_NR_FULLNESS_GROUPS)
+               return;
+
+       head = &class->fullness_list[fullness];
+       BUG_ON(!*head);
+       if (list_empty(&(*head)->lru))
+               *head = NULL;
+       else if (*head == page)
+               *head = (struct page *)list_entry((*head)->lru.next,
+                                       struct page, lru);
+
+       list_del_init(&page->lru);
+}
+
+static enum fullness_group fix_fullness_group(struct zs_pool *pool,
+                                               struct page *page)
+{
+       int class_idx;
+       struct size_class *class;
+       enum fullness_group currfg, newfg;
+
+       BUG_ON(!is_first_page(page));
+
+       get_zspage_mapping(page, &class_idx, &currfg);
+       newfg = get_fullness_group(page);
+       if (newfg == currfg)
+               goto out;
+
+       class = &pool->size_class[class_idx];
+       remove_zspage(page, class, currfg);
+       insert_zspage(page, class, newfg);
+       set_zspage_mapping(page, class_idx, newfg);
+
+out:
+       return newfg;
+}
+
+/*
+ * We have to decide on how many pages to link together
+ * to form a zspage for each size class. This is important
+ * to reduce wastage due to unusable space left at end of
+ * each zspage which is given as:
+ *     wastage = Zp - Zp % size_class
+ * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
+ *
+ * For example, for size class of 3/8 * PAGE_SIZE, we should
+ * link together 3 PAGE_SIZE sized pages to form a zspage
+ * since then we can perfectly fit in 8 such objects.
+ */
+static int get_zspage_order(int class_size)
+{
+       int i, max_usedpc = 0;
+       /* zspage order which gives maximum used size per KB */
+       int max_usedpc_order = 1;
+
+       for (i = 1; i <= max_zspage_order; i++) {
+               int zspage_size;
+               int waste, usedpc;
+
+               zspage_size = i * PAGE_SIZE;
+               waste = zspage_size % class_size;
+               usedpc = (zspage_size - waste) * 100 / zspage_size;
+
+               if (usedpc > max_usedpc) {
+                       max_usedpc = usedpc;
+                       max_usedpc_order = i;
+               }
+       }
+
+       return max_usedpc_order;
+}
+
+/*
+ * A single 'zspage' is composed of many system pages which are
+ * linked together using fields in struct page. This function finds
+ * the first/head page, given any component page of a zspage.
+ */
+static struct page *get_first_page(struct page *page)
+{
+       if (is_first_page(page))
+               return page;
+       else
+               return page->first_page;
+}
+
+static struct page *get_next_page(struct page *page)
+{
+       struct page *next;
+
+       if (is_last_page(page))
+               next = NULL;
+       else if (is_first_page(page))
+               next = (struct page *)page->private;
+       else
+               next = list_entry(page->lru.next, struct page, lru);
+
+       return next;
+}
+
+/* Encode <page, obj_idx> as a single handle value */
+static void *obj_location_to_handle(struct page *page, unsigned long obj_idx)
+{
+       unsigned long handle;
+
+       if (!page) {
+               BUG_ON(obj_idx);
+               return NULL;
+       }
+
+       handle = page_to_pfn(page) << OBJ_INDEX_BITS;
+       handle |= (obj_idx & OBJ_INDEX_MASK);
+
+       return (void *)handle;
+}
+
+/* Decode <page, obj_idx> pair from the given object handle */
+static void obj_handle_to_location(void *handle, struct page **page,
+                               unsigned long *obj_idx)
+{
+       unsigned long hval = (unsigned long)handle;
+
+       *page = pfn_to_page(hval >> OBJ_INDEX_BITS);
+       *obj_idx = hval & OBJ_INDEX_MASK;
+}
+
+static unsigned long obj_idx_to_offset(struct page *page,
+                               unsigned long obj_idx, int class_size)
+{
+       unsigned long off = 0;
+
+       if (!is_first_page(page))
+               off = page->index;
+
+       return off + obj_idx * class_size;
+}
+
+static void free_zspage(struct page *first_page)
+{
+       struct page *nextp, *tmp;
+
+       BUG_ON(!is_first_page(first_page));
+       BUG_ON(first_page->inuse);
+
+       nextp = (struct page *)page_private(first_page);
+
+       clear_bit(PG_private, &first_page->flags);
+       clear_bit(PG_private_2, &first_page->flags);
+       set_page_private(first_page, 0);
+       first_page->mapping = NULL;
+       first_page->freelist = NULL;
+       reset_page_mapcount(first_page);
+       __free_page(first_page);
+
+       /* zspage with only 1 system page */
+       if (!nextp)
+               return;
+
+       list_for_each_entry_safe(nextp, tmp, &nextp->lru, lru) {
+               list_del(&nextp->lru);
+               clear_bit(PG_private_2, &nextp->flags);
+               nextp->index = 0;
+               __free_page(nextp);
+       }
+}
+
+/* Initialize a newly allocated zspage */
+static void init_zspage(struct page *first_page, struct size_class *class)
+{
+       unsigned long off = 0;
+       struct page *page = first_page;
+
+       BUG_ON(!is_first_page(first_page));
+       while (page) {
+               struct page *next_page;
+               struct link_free *link;
+               unsigned int i, objs_on_page;
+
+               /*
+                * page->index stores offset of first object starting
+                * in the page. For the first page, this is always 0,
+                * so we use first_page->index (aka ->freelist) to store
+                * head of corresponding zspage's freelist.
+                */
+               if (page != first_page)
+                       page->index = off;
+
+               link = (struct link_free *)kmap_atomic(page) +
+                                               off / sizeof(*link);
+               objs_on_page = (PAGE_SIZE - off) / class->size;
+
+               for (i = 1; i <= objs_on_page; i++) {
+                       off += class->size;
+                       if (off < PAGE_SIZE) {
+                               link->next = obj_location_to_handle(page, i);
+                               link += class->size / sizeof(*link);
+                       }
+               }
+
+               /*
+                * We now come to the last (full or partial) object on this
+                * page, which must point to the first object on the next
+                * page (if present)
+                */
+               next_page = get_next_page(page);
+               link->next = obj_location_to_handle(next_page, 0);
+               kunmap_atomic(link);
+               page = next_page;
+               off = (off + class->size) % PAGE_SIZE;
+       }
+}
+
+/*
+ * Allocate a zspage for the given size class
+ */
+static struct page *alloc_zspage(struct size_class *class, gfp_t flags)
+{
+       int i, error;
+       struct page *first_page = NULL;
+
+       /*
+        * Allocate individual pages and link them together as:
+        * 1. first page->private = first sub-page
+        * 2. all sub-pages are linked together using page->lru
+        * 3. each sub-page is linked to the first page using page->first_page
+        *
+        * For each size class, First/Head pages are linked together using
+        * page->lru. Also, we set PG_private to identify the first page
+        * (i.e. no other sub-page has this flag set) and PG_private_2 to
+        * identify the last page.
+        */
+       error = -ENOMEM;
+       for (i = 0; i < class->zspage_order; i++) {
+               struct page *page, *prev_page;
+
+               page = alloc_page(flags);
+               if (!page)
+                       goto cleanup;
+
+               INIT_LIST_HEAD(&page->lru);
+               if (i == 0) {   /* first page */
+                       set_bit(PG_private, &page->flags);
+                       set_page_private(page, 0);
+                       first_page = page;
+                       first_page->inuse = 0;
+               }
+               if (i == 1)
+                       first_page->private = (unsigned long)page;
+               if (i >= 1)
+                       page->first_page = first_page;
+               if (i >= 2)
+                       list_add(&page->lru, &prev_page->lru);
+               if (i == class->zspage_order - 1)       /* last page */
+                       set_bit(PG_private_2, &page->flags);
+
+               prev_page = page;
+       }
+
+       init_zspage(first_page, class);
+
+       first_page->freelist = obj_location_to_handle(first_page, 0);
+       /* Maximum number of objects we can store in this zspage */
+       first_page->objects = class->zspage_order * PAGE_SIZE / class->size;
+
+       error = 0; /* Success */
+
+cleanup:
+       if (unlikely(error) && first_page) {
+               free_zspage(first_page);
+               first_page = NULL;
+       }
+
+       return first_page;
+}
+
+static struct page *find_get_zspage(struct size_class *class)
+{
+       int i;
+       struct page *page;
+
+       for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) {
+               page = class->fullness_list[i];
+               if (page)
+                       break;
+       }
+
+       return page;
+}
+
+
+/*
+ * If this becomes a separate module, register zs_init() with
+ * module_init(), zs_exit with module_exit(), and remove zs_initialized
+*/
+static int zs_initialized;
+
+static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action,
+                               void *pcpu)
+{
+       int cpu = (long)pcpu;
+       struct mapping_area *area;
+
+       switch (action) {
+       case CPU_UP_PREPARE:
+               area = &per_cpu(zs_map_area, cpu);
+               if (area->vm)
+                       break;
+               area->vm = alloc_vm_area(2 * PAGE_SIZE, area->vm_ptes);
+               if (!area->vm)
+                       return notifier_from_errno(-ENOMEM);
+               break;
+       case CPU_DEAD:
+       case CPU_UP_CANCELED:
+               area = &per_cpu(zs_map_area, cpu);
+               if (area->vm)
+                       free_vm_area(area->vm);
+               area->vm = NULL;
+               break;
+       }
+
+       return NOTIFY_OK;
+}
+
+static struct notifier_block zs_cpu_nb = {
+       .notifier_call = zs_cpu_notifier
+};
+
+static void zs_exit(void)
+{
+       int cpu;
+
+       for_each_online_cpu(cpu)
+               zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu);
+       unregister_cpu_notifier(&zs_cpu_nb);
+}
+
+static int zs_init(void)
+{
+       int cpu, ret;
+
+       register_cpu_notifier(&zs_cpu_nb);
+       for_each_online_cpu(cpu) {
+               ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
+               if (notifier_to_errno(ret))
+                       goto fail;
+       }
+       return 0;
+fail:
+       zs_exit();
+       return notifier_to_errno(ret);
+}
+
+struct zs_pool *zs_create_pool(const char *name, gfp_t flags)
+{
+       int i, error, ovhd_size;
+       struct zs_pool *pool;
+
+       if (!name)
+               return NULL;
+
+       ovhd_size = roundup(sizeof(*pool), PAGE_SIZE);
+       pool = kzalloc(ovhd_size, GFP_KERNEL);
+       if (!pool)
+               return NULL;
+
+       for (i = 0; i < ZS_SIZE_CLASSES; i++) {
+               int size;
+               struct size_class *class;
+
+               size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
+               if (size > ZS_MAX_ALLOC_SIZE)
+                       size = ZS_MAX_ALLOC_SIZE;
+
+               class = &pool->size_class[i];
+               class->size = size;
+               class->index = i;
+               spin_lock_init(&class->lock);
+               class->zspage_order = get_zspage_order(size);
+
+       }
+
+       /*
+        * If this becomes a separate module, register zs_init with
+        * module_init, and remove this block
+       */
+       if (!zs_initialized) {
+               error = zs_init();
+               if (error)
+                       goto cleanup;
+               zs_initialized = 1;
+       }
+
+       pool->flags = flags;
+       pool->name = name;
+
+       error = 0; /* Success */
+
+cleanup:
+       if (error) {
+               zs_destroy_pool(pool);
+               pool = NULL;
+       }
+
+       return pool;
+}
+EXPORT_SYMBOL_GPL(zs_create_pool);
+
+void zs_destroy_pool(struct zs_pool *pool)
+{
+       int i;
+
+       for (i = 0; i < ZS_SIZE_CLASSES; i++) {
+               int fg;
+               struct size_class *class = &pool->size_class[i];
+
+               for (fg = 0; fg < _ZS_NR_FULLNESS_GROUPS; fg++) {
+                       if (class->fullness_list[fg]) {
+                               pr_info("Freeing non-empty class with size "
+                                       "%db, fullness group %d\n",
+                                       class->size, fg);
+                       }
+               }
+       }
+       kfree(pool);
+}
+EXPORT_SYMBOL_GPL(zs_destroy_pool);
+
+/**
+ * zs_malloc - Allocate block of given size from pool.
+ * @pool: pool to allocate from
+ * @size: size of block to allocate
+ * @page: page no. that holds the object
+ * @offset: location of object within page
+ *
+ * On success, <page, offset> identifies block allocated
+ * and 0 is returned. On failure, <page, offset> is set to
+ * 0 and -ENOMEM is returned.
+ *
+ * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
+ */
+void *zs_malloc(struct zs_pool *pool, size_t size)
+{
+       void *obj;
+       struct link_free *link;
+       int class_idx;
+       struct size_class *class;
+
+       struct page *first_page, *m_page;
+       unsigned long m_objidx, m_offset;
+
+       if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
+               return NULL;
+
+       class_idx = get_size_class_index(size);
+       class = &pool->size_class[class_idx];
+       BUG_ON(class_idx != class->index);
+
+       spin_lock(&class->lock);
+       first_page = find_get_zspage(class);
+
+       if (!first_page) {
+               spin_unlock(&class->lock);
+               first_page = alloc_zspage(class, pool->flags);
+               if (unlikely(!first_page))
+                       return NULL;
+
+               set_zspage_mapping(first_page, class->index, ZS_EMPTY);
+               spin_lock(&class->lock);
+               class->pages_allocated += class->zspage_order;
+       }
+
+       obj = first_page->freelist;
+       obj_handle_to_location(obj, &m_page, &m_objidx);
+       m_offset = obj_idx_to_offset(m_page, m_objidx, class->size);
+
+       link = (struct link_free *)kmap_atomic(m_page) +
+                                       m_offset / sizeof(*link);
+       first_page->freelist = link->next;
+       memset(link, POISON_INUSE, sizeof(*link));
+       kunmap_atomic(link);
+
+       first_page->inuse++;
+       /* Now move the zspage to another fullness group, if required */
+       fix_fullness_group(pool, first_page);
+       spin_unlock(&class->lock);
+
+       return obj;
+}
+EXPORT_SYMBOL_GPL(zs_malloc);
+
+void zs_free(struct zs_pool *pool, void *obj)
+{
+       struct link_free *link;
+       struct page *first_page, *f_page;
+       unsigned long f_objidx, f_offset;
+
+       int class_idx;
+       struct size_class *class;
+       enum fullness_group fullness;
+
+       if (unlikely(!obj))
+               return;
+
+       obj_handle_to_location(obj, &f_page, &f_objidx);
+       first_page = get_first_page(f_page);
+
+       get_zspage_mapping(first_page, &class_idx, &fullness);
+       class = &pool->size_class[class_idx];
+       f_offset = obj_idx_to_offset(f_page, f_objidx, class->size);
+
+       spin_lock(&class->lock);
+
+       /* Insert this object in containing zspage's freelist */
+       link = (struct link_free *)((unsigned char *)kmap_atomic(f_page)
+                                                       + f_offset);
+       link->next = first_page->freelist;
+       kunmap_atomic(link);
+       first_page->freelist = obj;
+
+       first_page->inuse--;
+       fullness = fix_fullness_group(pool, first_page);
+
+       if (fullness == ZS_EMPTY)
+               class->pages_allocated -= class->zspage_order;
+
+       spin_unlock(&class->lock);
+
+       if (fullness == ZS_EMPTY)
+               free_zspage(first_page);
+}
+EXPORT_SYMBOL_GPL(zs_free);
+
+void *zs_map_object(struct zs_pool *pool, void *handle)
+{
+       struct page *page;
+       unsigned long obj_idx, off;
+
+       unsigned int class_idx;
+       enum fullness_group fg;
+       struct size_class *class;
+       struct mapping_area *area;
+
+       BUG_ON(!handle);
+
+       obj_handle_to_location(handle, &page, &obj_idx);
+       get_zspage_mapping(get_first_page(page), &class_idx, &fg);
+       class = &pool->size_class[class_idx];
+       off = obj_idx_to_offset(page, obj_idx, class->size);
+
+       area = &get_cpu_var(zs_map_area);
+       if (off + class->size <= PAGE_SIZE) {
+               /* this object is contained entirely within a page */
+               area->vm_addr = kmap_atomic(page);
+       } else {
+               /* this object spans two pages */
+               struct page *nextp;
+
+               nextp = get_next_page(page);
+               BUG_ON(!nextp);
+
+
+               set_pte(area->vm_ptes[0], mk_pte(page, PAGE_KERNEL));
+               set_pte(area->vm_ptes[1], mk_pte(nextp, PAGE_KERNEL));
+
+               /* We pre-allocated VM area so mapping can never fail */
+               area->vm_addr = area->vm->addr;
+       }
+
+       return area->vm_addr + off;
+}
+EXPORT_SYMBOL_GPL(zs_map_object);
+
+void zs_unmap_object(struct zs_pool *pool, void *handle)
+{
+       struct page *page;
+       unsigned long obj_idx, off;
+
+       unsigned int class_idx;
+       enum fullness_group fg;
+       struct size_class *class;
+       struct mapping_area *area;
+
+       BUG_ON(!handle);
+
+       obj_handle_to_location(handle, &page, &obj_idx);
+       get_zspage_mapping(get_first_page(page), &class_idx, &fg);
+       class = &pool->size_class[class_idx];
+       off = obj_idx_to_offset(page, obj_idx, class->size);
+
+       area = &__get_cpu_var(zs_map_area);
+       if (off + class->size <= PAGE_SIZE) {
+               kunmap_atomic(area->vm_addr);
+       } else {
+               set_pte(area->vm_ptes[0], __pte(0));
+               set_pte(area->vm_ptes[1], __pte(0));
+               __flush_tlb_one((unsigned long)area->vm_addr);
+               __flush_tlb_one((unsigned long)area->vm_addr + PAGE_SIZE);
+       }
+       put_cpu_var(zs_map_area);
+}
+EXPORT_SYMBOL_GPL(zs_unmap_object);
+
+u64 zs_get_total_size_bytes(struct zs_pool *pool)
+{
+       int i;
+       u64 npages = 0;
+
+       for (i = 0; i < ZS_SIZE_CLASSES; i++)
+               npages += pool->size_class[i].pages_allocated;
+
+       return npages << PAGE_SHIFT;
+}
+EXPORT_SYMBOL_GPL(zs_get_total_size_bytes);
diff --git a/drivers/staging/zsmalloc/zsmalloc.h b/drivers/staging/zsmalloc/zsmalloc.h
new file mode 100644 (file)
index 0000000..949384e
--- /dev/null
@@ -0,0 +1,31 @@
+/*
+ * zsmalloc memory allocator
+ *
+ * Copyright (C) 2011  Nitin Gupta
+ *
+ * This code is released using a dual license strategy: BSD/GPL
+ * You can choose the license that better fits your requirements.
+ *
+ * Released under the terms of 3-clause BSD License
+ * Released under the terms of GNU General Public License Version 2.0
+ */
+
+#ifndef _ZS_MALLOC_H_
+#define _ZS_MALLOC_H_
+
+#include <linux/types.h>
+
+struct zs_pool;
+
+struct zs_pool *zs_create_pool(const char *name, gfp_t flags);
+void zs_destroy_pool(struct zs_pool *pool);
+
+void *zs_malloc(struct zs_pool *pool, size_t size);
+void zs_free(struct zs_pool *pool, void *obj);
+
+void *zs_map_object(struct zs_pool *pool, void *handle);
+void zs_unmap_object(struct zs_pool *pool, void *handle);
+
+u64 zs_get_total_size_bytes(struct zs_pool *pool);
+
+#endif
diff --git a/drivers/staging/zsmalloc/zsmalloc_int.h b/drivers/staging/zsmalloc/zsmalloc_int.h
new file mode 100644 (file)
index 0000000..354a020
--- /dev/null
@@ -0,0 +1,126 @@
+/*
+ * zsmalloc memory allocator
+ *
+ * Copyright (C) 2011  Nitin Gupta
+ *
+ * This code is released using a dual license strategy: BSD/GPL
+ * You can choose the license that better fits your requirements.
+ *
+ * Released under the terms of 3-clause BSD License
+ * Released under the terms of GNU General Public License Version 2.0
+ */
+
+#ifndef _ZS_MALLOC_INT_H_
+#define _ZS_MALLOC_INT_H_
+
+#include <linux/kernel.h>
+#include <linux/spinlock.h>
+#include <linux/types.h>
+
+/*
+ * This must be power of 2 and greater than of equal to sizeof(link_free).
+ * These two conditions ensure that any 'struct link_free' itself doesn't
+ * span more than 1 page which avoids complex case of mapping 2 pages simply
+ * to restore link_free pointer values.
+ */
+#define ZS_ALIGN               8
+
+/* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
+#define ZS_MIN_ALLOC_SIZE      32
+#define ZS_MAX_ALLOC_SIZE      PAGE_SIZE
+
+/*
+ * On systems with 4K page size, this gives 254 size classes! There is a
+ * trader-off here:
+ *  - Large number of size classes is potentially wasteful as free page are
+ *    spread across these classes
+ *  - Small number of size classes causes large internal fragmentation
+ *  - Probably its better to use specific size classes (empirically
+ *    determined). NOTE: all those class sizes must be set as multiple of
+ *    ZS_ALIGN to make sure link_free itself never has to span 2 pages.
+ *
+ *  ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
+ *  (reason above)
+ */
+#define ZS_SIZE_CLASS_DELTA    16
+#define ZS_SIZE_CLASSES                ((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / \
+                                       ZS_SIZE_CLASS_DELTA + 1)
+
+/*
+ * A single 'zspage' is composed of N discontiguous 0-order (single) pages.
+ * This defines upper limit on N.
+ */
+static const int max_zspage_order = 4;
+
+/*
+ * We do not maintain any list for completely empty or full pages
+ */
+enum fullness_group {
+       ZS_ALMOST_FULL,
+       ZS_ALMOST_EMPTY,
+       _ZS_NR_FULLNESS_GROUPS,
+
+       ZS_EMPTY,
+       ZS_FULL
+};
+
+/*
+ * We assign a page to ZS_ALMOST_EMPTY fullness group when:
+ *     n <= N / f, where
+ * n = number of allocated objects
+ * N = total number of objects zspage can store
+ * f = 1/fullness_threshold_frac
+ *
+ * Similarly, we assign zspage to:
+ *     ZS_ALMOST_FULL  when n > N / f
+ *     ZS_EMPTY        when n == 0
+ *     ZS_FULL         when n == N
+ *
+ * (see: fix_fullness_group())
+ */
+static const int fullness_threshold_frac = 4;
+
+struct mapping_area {
+       struct vm_struct *vm;
+       pte_t *vm_ptes[2];
+       char *vm_addr;
+};
+
+struct size_class {
+       /*
+        * Size of objects stored in this class. Must be multiple
+        * of ZS_ALIGN.
+        */
+       int size;
+       unsigned int index;
+
+       /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
+       int zspage_order;
+
+       spinlock_t lock;
+
+       /* stats */
+       u64 pages_allocated;
+
+       struct page *fullness_list[_ZS_NR_FULLNESS_GROUPS];
+};
+
+/*
+ * Placed within free objects to form a singly linked list.
+ * For every zspage, first_page->freelist gives head of this list.
+ *
+ * This must be power of 2 and less than or equal to ZS_ALIGN
+ */
+struct link_free {
+       /* Handle of next free chunk (encodes <PFN, obj_idx>) */
+       void *next;
+};
+
+struct zs_pool {
+       struct size_class size_class[ZS_SIZE_CLASSES];
+
+       gfp_t flags;    /* allocation flags used when growing pool */
+       const char *name;
+};
+
+#endif