ashmem: Anonymous shared memory subsystem
authorRobert Love <rlove@google.com>
Wed, 21 Dec 2011 00:49:48 +0000 (16:49 -0800)
committerGreg Kroah-Hartman <gregkh@suse.de>
Wed, 21 Dec 2011 21:38:28 +0000 (13:38 -0800)
The anonymous shared memory (ashmem) subsystem provides a
Unix-y,file-based shared memory interface to user-space. It
works like anonymous memory (e.g. mmapping fd=0) except if
you share the file descriptor via the usual means, you will
share the mapping. The shared memory can be accessed via both
mmap or file I/O. The backing store is a simple shmem file.

Additionally, ashmem introduces the concept of page pinning.
Pinned pages (the default) behave like any anonymous memory.
Unpinned pages are available to the kernel for eviction during
VM pressure. When repinning the pages, the return value
instructs user-space as to any eviction. In this manner,
user-space processes may implement caching and similar
resource management that efficiently integrates with kernel
memory management.

Signed-off-by: Robert Love <rlove@google.com>
ashmem: Don't install fault handler for private mmaps.

Ashmem is used to create named private heaps. If this heap is backed
by a tmpfs file it will allocate two pages for every page touched.
In 2.6.27, the extra page would later be freed, but 2.6.29 does not
scan anonymous pages when running without swap so the memory is not
freed while the file is referenced. This change changes the behavior
of private ashmem mmaps to match /dev/zero instead tmpfs.

Signed-off-by: Arve Hjønnevåg <arve@android.com>
ashmem: Add common prefix to name reported in /proc/pid/maps

Signed-off-by: Arve Hjønnevåg <arve@android.com>
ashmem: don't require a page aligned size

This makes ashmem more similar to shmem and mmap, by
not requiring the specified size to be page aligned,
instead rounding it internally as needed.

Signed-off-by: Marco Nelissen <marcone@android.com>
[jstultz: Improved commit subject and included patch description
from rlove. Also moved ashmem files to staging dir, and reworked
code to avoid touching mm/shmem.c while we're in staging.]
CC: Brian Swetland <swetland@google.com>
CC: Colin Cross <ccross@android.com>
CC: Arve Hjønnevåg <arve@android.com>
CC: Dima Zavin <dima@android.com>
CC: Robert Love <rlove@google.com>
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
drivers/staging/android/Kconfig
drivers/staging/android/Makefile
drivers/staging/android/ashmem.c [new file with mode: 0644]
drivers/staging/android/ashmem.h [new file with mode: 0644]

index 6094fd6..becf711 100644 (file)
@@ -12,6 +12,15 @@ config ANDROID_BINDER_IPC
        bool "Android Binder IPC Driver"
        default n
 
+config ASHMEM
+       bool "Enable the Anonymous Shared Memory Subsystem"
+       default n
+       depends on SHMEM || TINY_SHMEM
+       help
+         The ashmem subsystem is a new shared memory allocator, similar to
+         POSIX SHM but with different behavior and sporting a simpler
+         file-based API.
+
 config ANDROID_LOGGER
        tristate "Android log driver"
        default n
index 8fd7391..eaed1ff 100644 (file)
@@ -1,4 +1,5 @@
 obj-$(CONFIG_ANDROID_BINDER_IPC)       += binder.o
+obj-$(CONFIG_ASHMEM)                   += ashmem.o
 obj-$(CONFIG_ANDROID_LOGGER)           += logger.o
 obj-$(CONFIG_ANDROID_RAM_CONSOLE)      += ram_console.o
 obj-$(CONFIG_ANDROID_TIMED_OUTPUT)     += timed_output.o
diff --git a/drivers/staging/android/ashmem.c b/drivers/staging/android/ashmem.c
new file mode 100644 (file)
index 0000000..5775c6c
--- /dev/null
@@ -0,0 +1,678 @@
+/* mm/ashmem.c
+**
+** Anonymous Shared Memory Subsystem, ashmem
+**
+** Copyright (C) 2008 Google, Inc.
+**
+** Robert Love <rlove@google.com>
+**
+** This software is licensed under the terms of the GNU General Public
+** License version 2, as published by the Free Software Foundation, and
+** may be copied, distributed, and modified under those terms.
+**
+** This program is distributed in the hope that it will be useful,
+** but WITHOUT ANY WARRANTY; without even the implied warranty of
+** MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+** GNU General Public License for more details.
+*/
+
+#include <linux/module.h>
+#include <linux/file.h>
+#include <linux/fs.h>
+#include <linux/miscdevice.h>
+#include <linux/security.h>
+#include <linux/mm.h>
+#include <linux/mman.h>
+#include <linux/uaccess.h>
+#include <linux/personality.h>
+#include <linux/bitops.h>
+#include <linux/mutex.h>
+#include <linux/shmem_fs.h>
+#include "ashmem.h"
+
+#define ASHMEM_NAME_PREFIX "dev/ashmem/"
+#define ASHMEM_NAME_PREFIX_LEN (sizeof(ASHMEM_NAME_PREFIX) - 1)
+#define ASHMEM_FULL_NAME_LEN (ASHMEM_NAME_LEN + ASHMEM_NAME_PREFIX_LEN)
+
+/*
+ * ashmem_area - anonymous shared memory area
+ * Lifecycle: From our parent file's open() until its release()
+ * Locking: Protected by `ashmem_mutex'
+ * Big Note: Mappings do NOT pin this structure; it dies on close()
+ */
+struct ashmem_area {
+       char name[ASHMEM_FULL_NAME_LEN];/* optional name for /proc/pid/maps */
+       struct list_head unpinned_list; /* list of all ashmem areas */
+       struct file *file;              /* the shmem-based backing file */
+       size_t size;                    /* size of the mapping, in bytes */
+       unsigned long prot_mask;        /* allowed prot bits, as vm_flags */
+};
+
+/*
+ * ashmem_range - represents an interval of unpinned (evictable) pages
+ * Lifecycle: From unpin to pin
+ * Locking: Protected by `ashmem_mutex'
+ */
+struct ashmem_range {
+       struct list_head lru;           /* entry in LRU list */
+       struct list_head unpinned;      /* entry in its area's unpinned list */
+       struct ashmem_area *asma;       /* associated area */
+       size_t pgstart;                 /* starting page, inclusive */
+       size_t pgend;                   /* ending page, inclusive */
+       unsigned int purged;            /* ASHMEM_NOT or ASHMEM_WAS_PURGED */
+};
+
+/* LRU list of unpinned pages, protected by ashmem_mutex */
+static LIST_HEAD(ashmem_lru_list);
+
+/* Count of pages on our LRU list, protected by ashmem_mutex */
+static unsigned long lru_count;
+
+/*
+ * ashmem_mutex - protects the list of and each individual ashmem_area
+ *
+ * Lock Ordering: ashmex_mutex -> i_mutex -> i_alloc_sem
+ */
+static DEFINE_MUTEX(ashmem_mutex);
+
+static struct kmem_cache *ashmem_area_cachep __read_mostly;
+static struct kmem_cache *ashmem_range_cachep __read_mostly;
+
+#define range_size(range) \
+  ((range)->pgend - (range)->pgstart + 1)
+
+#define range_on_lru(range) \
+  ((range)->purged == ASHMEM_NOT_PURGED)
+
+#define page_range_subsumes_range(range, start, end) \
+  (((range)->pgstart >= (start)) && ((range)->pgend <= (end)))
+
+#define page_range_subsumed_by_range(range, start, end) \
+  (((range)->pgstart <= (start)) && ((range)->pgend >= (end)))
+
+#define page_in_range(range, page) \
+ (((range)->pgstart <= (page)) && ((range)->pgend >= (page)))
+
+#define page_range_in_range(range, start, end) \
+  (page_in_range(range, start) || page_in_range(range, end) || \
+   page_range_subsumes_range(range, start, end))
+
+#define range_before_page(range, page) \
+  ((range)->pgend < (page))
+
+#define PROT_MASK              (PROT_EXEC | PROT_READ | PROT_WRITE)
+
+static inline void lru_add(struct ashmem_range *range)
+{
+       list_add_tail(&range->lru, &ashmem_lru_list);
+       lru_count += range_size(range);
+}
+
+static inline void lru_del(struct ashmem_range *range)
+{
+       list_del(&range->lru);
+       lru_count -= range_size(range);
+}
+
+/*
+ * range_alloc - allocate and initialize a new ashmem_range structure
+ *
+ * 'asma' - associated ashmem_area
+ * 'prev_range' - the previous ashmem_range in the sorted asma->unpinned list
+ * 'purged' - initial purge value (ASMEM_NOT_PURGED or ASHMEM_WAS_PURGED)
+ * 'start' - starting page, inclusive
+ * 'end' - ending page, inclusive
+ *
+ * Caller must hold ashmem_mutex.
+ */
+static int range_alloc(struct ashmem_area *asma,
+                      struct ashmem_range *prev_range, unsigned int purged,
+                      size_t start, size_t end)
+{
+       struct ashmem_range *range;
+
+       range = kmem_cache_zalloc(ashmem_range_cachep, GFP_KERNEL);
+       if (unlikely(!range))
+               return -ENOMEM;
+
+       range->asma = asma;
+       range->pgstart = start;
+       range->pgend = end;
+       range->purged = purged;
+
+       list_add_tail(&range->unpinned, &prev_range->unpinned);
+
+       if (range_on_lru(range))
+               lru_add(range);
+
+       return 0;
+}
+
+static void range_del(struct ashmem_range *range)
+{
+       list_del(&range->unpinned);
+       if (range_on_lru(range))
+               lru_del(range);
+       kmem_cache_free(ashmem_range_cachep, range);
+}
+
+/*
+ * range_shrink - shrinks a range
+ *
+ * Caller must hold ashmem_mutex.
+ */
+static inline void range_shrink(struct ashmem_range *range,
+                               size_t start, size_t end)
+{
+       size_t pre = range_size(range);
+
+       range->pgstart = start;
+       range->pgend = end;
+
+       if (range_on_lru(range))
+               lru_count -= pre - range_size(range);
+}
+
+static int ashmem_open(struct inode *inode, struct file *file)
+{
+       struct ashmem_area *asma;
+       int ret;
+
+       ret = nonseekable_open(inode, file);
+       if (unlikely(ret))
+               return ret;
+
+       asma = kmem_cache_zalloc(ashmem_area_cachep, GFP_KERNEL);
+       if (unlikely(!asma))
+               return -ENOMEM;
+
+       INIT_LIST_HEAD(&asma->unpinned_list);
+       memcpy(asma->name, ASHMEM_NAME_PREFIX, ASHMEM_NAME_PREFIX_LEN);
+       asma->prot_mask = PROT_MASK;
+       file->private_data = asma;
+
+       return 0;
+}
+
+static int ashmem_release(struct inode *ignored, struct file *file)
+{
+       struct ashmem_area *asma = file->private_data;
+       struct ashmem_range *range, *next;
+
+       mutex_lock(&ashmem_mutex);
+       list_for_each_entry_safe(range, next, &asma->unpinned_list, unpinned)
+               range_del(range);
+       mutex_unlock(&ashmem_mutex);
+
+       if (asma->file)
+               fput(asma->file);
+       kmem_cache_free(ashmem_area_cachep, asma);
+
+       return 0;
+}
+
+static int ashmem_mmap(struct file *file, struct vm_area_struct *vma)
+{
+       struct ashmem_area *asma = file->private_data;
+       int ret = 0;
+
+       mutex_lock(&ashmem_mutex);
+
+       /* user needs to SET_SIZE before mapping */
+       if (unlikely(!asma->size)) {
+               ret = -EINVAL;
+               goto out;
+       }
+
+       /* requested protection bits must match our allowed protection mask */
+       if (unlikely((vma->vm_flags & ~asma->prot_mask) & PROT_MASK)) {
+               ret = -EPERM;
+               goto out;
+       }
+
+       if (!asma->file) {
+               char *name = ASHMEM_NAME_DEF;
+               struct file *vmfile;
+
+               if (asma->name[ASHMEM_NAME_PREFIX_LEN] != '\0')
+                       name = asma->name;
+
+               /* ... and allocate the backing shmem file */
+               vmfile = shmem_file_setup(name, asma->size, vma->vm_flags);
+               if (unlikely(IS_ERR(vmfile))) {
+                       ret = PTR_ERR(vmfile);
+                       goto out;
+               }
+               asma->file = vmfile;
+       }
+       get_file(asma->file);
+
+       /*
+        * XXX - Reworked to use shmem_zero_setup() instead of 
+        * shmem_set_file while we're in staging. -jstultz
+        */
+       if (vma->vm_flags & VM_SHARED) {
+               ret = shmem_zero_setup(vma);
+               if (ret) {
+                       fput(asma->file);
+                       goto out;
+               }
+       }
+
+       if (vma->vm_file)
+               fput(vma->vm_file);
+       vma->vm_file = asma->file;
+       vma->vm_flags |= VM_CAN_NONLINEAR;
+
+out:
+       mutex_unlock(&ashmem_mutex);
+       return ret;
+}
+
+/*
+ * ashmem_shrink - our cache shrinker, called from mm/vmscan.c :: shrink_slab
+ *
+ * 'nr_to_scan' is the number of objects (pages) to prune, or 0 to query how
+ * many objects (pages) we have in total.
+ *
+ * 'gfp_mask' is the mask of the allocation that got us into this mess.
+ *
+ * Return value is the number of objects (pages) remaining, or -1 if we cannot
+ * proceed without risk of deadlock (due to gfp_mask).
+ *
+ * We approximate LRU via least-recently-unpinned, jettisoning unpinned partial
+ * chunks of ashmem regions LRU-wise one-at-a-time until we hit 'nr_to_scan'
+ * pages freed.
+ */
+static int ashmem_shrink(int nr_to_scan, gfp_t gfp_mask)
+{
+       struct ashmem_range *range, *next;
+
+       /* We might recurse into filesystem code, so bail out if necessary */
+       if (nr_to_scan && !(gfp_mask & __GFP_FS))
+               return -1;
+       if (!nr_to_scan)
+               return lru_count;
+
+       mutex_lock(&ashmem_mutex);
+       list_for_each_entry_safe(range, next, &ashmem_lru_list, lru) {
+               struct inode *inode = range->asma->file->f_dentry->d_inode;
+               loff_t start = range->pgstart * PAGE_SIZE;
+               loff_t end = (range->pgend + 1) * PAGE_SIZE - 1;
+
+               vmtruncate_range(inode, start, end);
+               range->purged = ASHMEM_WAS_PURGED;
+               lru_del(range);
+
+               nr_to_scan -= range_size(range);
+               if (nr_to_scan <= 0)
+                       break;
+       }
+       mutex_unlock(&ashmem_mutex);
+
+       return lru_count;
+}
+
+static struct shrinker ashmem_shrinker = {
+       .shrink = ashmem_shrink,
+       .seeks = DEFAULT_SEEKS * 4,
+};
+
+static int set_prot_mask(struct ashmem_area *asma, unsigned long prot)
+{
+       int ret = 0;
+
+       mutex_lock(&ashmem_mutex);
+
+       /* the user can only remove, not add, protection bits */
+       if (unlikely((asma->prot_mask & prot) != prot)) {
+               ret = -EINVAL;
+               goto out;
+       }
+
+       /* does the application expect PROT_READ to imply PROT_EXEC? */
+       if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
+               prot |= PROT_EXEC;
+
+       asma->prot_mask = prot;
+
+out:
+       mutex_unlock(&ashmem_mutex);
+       return ret;
+}
+
+static int set_name(struct ashmem_area *asma, void __user *name)
+{
+       int ret = 0;
+
+       mutex_lock(&ashmem_mutex);
+
+       /* cannot change an existing mapping's name */
+       if (unlikely(asma->file)) {
+               ret = -EINVAL;
+               goto out;
+       }
+
+       if (unlikely(copy_from_user(asma->name + ASHMEM_NAME_PREFIX_LEN,
+                                   name, ASHMEM_NAME_LEN)))
+               ret = -EFAULT;
+       asma->name[ASHMEM_FULL_NAME_LEN-1] = '\0';
+
+out:
+       mutex_unlock(&ashmem_mutex);
+
+       return ret;
+}
+
+static int get_name(struct ashmem_area *asma, void __user *name)
+{
+       int ret = 0;
+
+       mutex_lock(&ashmem_mutex);
+       if (asma->name[ASHMEM_NAME_PREFIX_LEN] != '\0') {
+               size_t len;
+
+               /*
+                * Copying only `len', instead of ASHMEM_NAME_LEN, bytes
+                * prevents us from revealing one user's stack to another.
+                */
+               len = strlen(asma->name + ASHMEM_NAME_PREFIX_LEN) + 1;
+               if (unlikely(copy_to_user(name,
+                               asma->name + ASHMEM_NAME_PREFIX_LEN, len)))
+                       ret = -EFAULT;
+       } else {
+               if (unlikely(copy_to_user(name, ASHMEM_NAME_DEF,
+                                         sizeof(ASHMEM_NAME_DEF))))
+                       ret = -EFAULT;
+       }
+       mutex_unlock(&ashmem_mutex);
+
+       return ret;
+}
+
+/*
+ * ashmem_pin - pin the given ashmem region, returning whether it was
+ * previously purged (ASHMEM_WAS_PURGED) or not (ASHMEM_NOT_PURGED).
+ *
+ * Caller must hold ashmem_mutex.
+ */
+static int ashmem_pin(struct ashmem_area *asma, size_t pgstart, size_t pgend)
+{
+       struct ashmem_range *range, *next;
+       int ret = ASHMEM_NOT_PURGED;
+
+       list_for_each_entry_safe(range, next, &asma->unpinned_list, unpinned) {
+               /* moved past last applicable page; we can short circuit */
+               if (range_before_page(range, pgstart))
+                       break;
+
+               /*
+                * The user can ask us to pin pages that span multiple ranges,
+                * or to pin pages that aren't even unpinned, so this is messy.
+                *
+                * Four cases:
+                * 1. The requested range subsumes an existing range, so we
+                *    just remove the entire matching range.
+                * 2. The requested range overlaps the start of an existing
+                *    range, so we just update that range.
+                * 3. The requested range overlaps the end of an existing
+                *    range, so we just update that range.
+                * 4. The requested range punches a hole in an existing range,
+                *    so we have to update one side of the range and then
+                *    create a new range for the other side.
+                */
+               if (page_range_in_range(range, pgstart, pgend)) {
+                       ret |= range->purged;
+
+                       /* Case #1: Easy. Just nuke the whole thing. */
+                       if (page_range_subsumes_range(range, pgstart, pgend)) {
+                               range_del(range);
+                               continue;
+                       }
+
+                       /* Case #2: We overlap from the start, so adjust it */
+                       if (range->pgstart >= pgstart) {
+                               range_shrink(range, pgend + 1, range->pgend);
+                               continue;
+                       }
+
+                       /* Case #3: We overlap from the rear, so adjust it */
+                       if (range->pgend <= pgend) {
+                               range_shrink(range, range->pgstart, pgstart-1);
+                               continue;
+                       }
+
+                       /*
+                        * Case #4: We eat a chunk out of the middle. A bit
+                        * more complicated, we allocate a new range for the
+                        * second half and adjust the first chunk's endpoint.
+                        */
+                       range_alloc(asma, range, range->purged,
+                                   pgend + 1, range->pgend);
+                       range_shrink(range, range->pgstart, pgstart - 1);
+                       break;
+               }
+       }
+
+       return ret;
+}
+
+/*
+ * ashmem_unpin - unpin the given range of pages. Returns zero on success.
+ *
+ * Caller must hold ashmem_mutex.
+ */
+static int ashmem_unpin(struct ashmem_area *asma, size_t pgstart, size_t pgend)
+{
+       struct ashmem_range *range, *next;
+       unsigned int purged = ASHMEM_NOT_PURGED;
+
+restart:
+       list_for_each_entry_safe(range, next, &asma->unpinned_list, unpinned) {
+               /* short circuit: this is our insertion point */
+               if (range_before_page(range, pgstart))
+                       break;
+
+               /*
+                * The user can ask us to unpin pages that are already entirely
+                * or partially pinned. We handle those two cases here.
+                */
+               if (page_range_subsumed_by_range(range, pgstart, pgend))
+                       return 0;
+               if (page_range_in_range(range, pgstart, pgend)) {
+                       pgstart = min_t(size_t, range->pgstart, pgstart),
+                       pgend = max_t(size_t, range->pgend, pgend);
+                       purged |= range->purged;
+                       range_del(range);
+                       goto restart;
+               }
+       }
+
+       return range_alloc(asma, range, purged, pgstart, pgend);
+}
+
+/*
+ * ashmem_get_pin_status - Returns ASHMEM_IS_UNPINNED if _any_ pages in the
+ * given interval are unpinned and ASHMEM_IS_PINNED otherwise.
+ *
+ * Caller must hold ashmem_mutex.
+ */
+static int ashmem_get_pin_status(struct ashmem_area *asma, size_t pgstart,
+                                size_t pgend)
+{
+       struct ashmem_range *range;
+       int ret = ASHMEM_IS_PINNED;
+
+       list_for_each_entry(range, &asma->unpinned_list, unpinned) {
+               if (range_before_page(range, pgstart))
+                       break;
+               if (page_range_in_range(range, pgstart, pgend)) {
+                       ret = ASHMEM_IS_UNPINNED;
+                       break;
+               }
+       }
+
+       return ret;
+}
+
+static int ashmem_pin_unpin(struct ashmem_area *asma, unsigned long cmd,
+                           void __user *p)
+{
+       struct ashmem_pin pin;
+       size_t pgstart, pgend;
+       int ret = -EINVAL;
+
+       if (unlikely(!asma->file))
+               return -EINVAL;
+
+       if (unlikely(copy_from_user(&pin, p, sizeof(pin))))
+               return -EFAULT;
+
+       /* per custom, you can pass zero for len to mean "everything onward" */
+       if (!pin.len)
+               pin.len = PAGE_ALIGN(asma->size) - pin.offset;
+
+       if (unlikely((pin.offset | pin.len) & ~PAGE_MASK))
+               return -EINVAL;
+
+       if (unlikely(((__u32) -1) - pin.offset < pin.len))
+               return -EINVAL;
+
+       if (unlikely(PAGE_ALIGN(asma->size) < pin.offset + pin.len))
+               return -EINVAL;
+
+       pgstart = pin.offset / PAGE_SIZE;
+       pgend = pgstart + (pin.len / PAGE_SIZE) - 1;
+
+       mutex_lock(&ashmem_mutex);
+
+       switch (cmd) {
+       case ASHMEM_PIN:
+               ret = ashmem_pin(asma, pgstart, pgend);
+               break;
+       case ASHMEM_UNPIN:
+               ret = ashmem_unpin(asma, pgstart, pgend);
+               break;
+       case ASHMEM_GET_PIN_STATUS:
+               ret = ashmem_get_pin_status(asma, pgstart, pgend);
+               break;
+       }
+
+       mutex_unlock(&ashmem_mutex);
+
+       return ret;
+}
+
+static long ashmem_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
+{
+       struct ashmem_area *asma = file->private_data;
+       long ret = -ENOTTY;
+
+       switch (cmd) {
+       case ASHMEM_SET_NAME:
+               ret = set_name(asma, (void __user *) arg);
+               break;
+       case ASHMEM_GET_NAME:
+               ret = get_name(asma, (void __user *) arg);
+               break;
+       case ASHMEM_SET_SIZE:
+               ret = -EINVAL;
+               if (!asma->file) {
+                       ret = 0;
+                       asma->size = (size_t) arg;
+               }
+               break;
+       case ASHMEM_GET_SIZE:
+               ret = asma->size;
+               break;
+       case ASHMEM_SET_PROT_MASK:
+               ret = set_prot_mask(asma, arg);
+               break;
+       case ASHMEM_GET_PROT_MASK:
+               ret = asma->prot_mask;
+               break;
+       case ASHMEM_PIN:
+       case ASHMEM_UNPIN:
+       case ASHMEM_GET_PIN_STATUS:
+               ret = ashmem_pin_unpin(asma, cmd, (void __user *) arg);
+               break;
+       case ASHMEM_PURGE_ALL_CACHES:
+               ret = -EPERM;
+               if (capable(CAP_SYS_ADMIN)) {
+                       ret = ashmem_shrink(0, GFP_KERNEL);
+                       ashmem_shrink(ret, GFP_KERNEL);
+               }
+               break;
+       }
+
+       return ret;
+}
+
+static struct file_operations ashmem_fops = {
+       .owner = THIS_MODULE,
+       .open = ashmem_open,
+       .release = ashmem_release,
+       .mmap = ashmem_mmap,
+       .unlocked_ioctl = ashmem_ioctl,
+       .compat_ioctl = ashmem_ioctl,
+};
+
+static struct miscdevice ashmem_misc = {
+       .minor = MISC_DYNAMIC_MINOR,
+       .name = "ashmem",
+       .fops = &ashmem_fops,
+};
+
+static int __init ashmem_init(void)
+{
+       int ret;
+
+       ashmem_area_cachep = kmem_cache_create("ashmem_area_cache",
+                                         sizeof(struct ashmem_area),
+                                         0, 0, NULL);
+       if (unlikely(!ashmem_area_cachep)) {
+               printk(KERN_ERR "ashmem: failed to create slab cache\n");
+               return -ENOMEM;
+       }
+
+       ashmem_range_cachep = kmem_cache_create("ashmem_range_cache",
+                                         sizeof(struct ashmem_range),
+                                         0, 0, NULL);
+       if (unlikely(!ashmem_range_cachep)) {
+               printk(KERN_ERR "ashmem: failed to create slab cache\n");
+               return -ENOMEM;
+       }
+
+       ret = misc_register(&ashmem_misc);
+       if (unlikely(ret)) {
+               printk(KERN_ERR "ashmem: failed to register misc device!\n");
+               return ret;
+       }
+
+       register_shrinker(&ashmem_shrinker);
+
+       printk(KERN_INFO "ashmem: initialized\n");
+
+       return 0;
+}
+
+static void __exit ashmem_exit(void)
+{
+       int ret;
+
+       unregister_shrinker(&ashmem_shrinker);
+
+       ret = misc_deregister(&ashmem_misc);
+       if (unlikely(ret))
+               printk(KERN_ERR "ashmem: failed to unregister misc device!\n");
+
+       kmem_cache_destroy(ashmem_range_cachep);
+       kmem_cache_destroy(ashmem_area_cachep);
+
+       printk(KERN_INFO "ashmem: unloaded\n");
+}
+
+module_init(ashmem_init);
+module_exit(ashmem_exit);
+
+MODULE_LICENSE("GPL");
diff --git a/drivers/staging/android/ashmem.h b/drivers/staging/android/ashmem.h
new file mode 100644 (file)
index 0000000..1976b10
--- /dev/null
@@ -0,0 +1,48 @@
+/*
+ * include/linux/ashmem.h
+ *
+ * Copyright 2008 Google Inc.
+ * Author: Robert Love
+ *
+ * This file is dual licensed.  It may be redistributed and/or modified
+ * under the terms of the Apache 2.0 License OR version 2 of the GNU
+ * General Public License.
+ */
+
+#ifndef _LINUX_ASHMEM_H
+#define _LINUX_ASHMEM_H
+
+#include <linux/limits.h>
+#include <linux/ioctl.h>
+
+#define ASHMEM_NAME_LEN                256
+
+#define ASHMEM_NAME_DEF                "dev/ashmem"
+
+/* Return values from ASHMEM_PIN: Was the mapping purged while unpinned? */
+#define ASHMEM_NOT_PURGED      0
+#define ASHMEM_WAS_PURGED      1
+
+/* Return values from ASHMEM_GET_PIN_STATUS: Is the mapping pinned? */
+#define ASHMEM_IS_UNPINNED     0
+#define ASHMEM_IS_PINNED       1
+
+struct ashmem_pin {
+       __u32 offset;   /* offset into region, in bytes, page-aligned */
+       __u32 len;      /* length forward from offset, in bytes, page-aligned */
+};
+
+#define __ASHMEMIOC            0x77
+
+#define ASHMEM_SET_NAME                _IOW(__ASHMEMIOC, 1, char[ASHMEM_NAME_LEN])
+#define ASHMEM_GET_NAME                _IOR(__ASHMEMIOC, 2, char[ASHMEM_NAME_LEN])
+#define ASHMEM_SET_SIZE                _IOW(__ASHMEMIOC, 3, size_t)
+#define ASHMEM_GET_SIZE                _IO(__ASHMEMIOC, 4)
+#define ASHMEM_SET_PROT_MASK   _IOW(__ASHMEMIOC, 5, unsigned long)
+#define ASHMEM_GET_PROT_MASK   _IO(__ASHMEMIOC, 6)
+#define ASHMEM_PIN             _IOW(__ASHMEMIOC, 7, struct ashmem_pin)
+#define ASHMEM_UNPIN           _IOW(__ASHMEMIOC, 8, struct ashmem_pin)
+#define ASHMEM_GET_PIN_STATUS  _IO(__ASHMEMIOC, 9)
+#define ASHMEM_PURGE_ALL_CACHES        _IO(__ASHMEMIOC, 10)
+
+#endif /* _LINUX_ASHMEM_H */