Btrfs: Avoid memcpy where possible in extent_buffers
authorChris Mason <chris.mason@oracle.com>
Mon, 15 Oct 2007 20:14:37 +0000 (16:14 -0400)
committerChris Mason <chris.mason@oracle.com>
Thu, 25 Sep 2008 15:03:56 +0000 (11:03 -0400)
Signed-off-by: Chris Mason <chris.mason@oracle.com>
fs/btrfs/ctree.h
fs/btrfs/disk-io.c
fs/btrfs/extent_map.c
fs/btrfs/extent_map.h

index 30fbbd7221a96086eea849efcb63c6b2efab7a09..aed08618aca67167de95bf1d50e422efe823d261 100644 (file)
@@ -483,15 +483,50 @@ static inline void btrfs_set_##name(struct extent_buffer *eb,             \
 #define BTRFS_SETGET_HEADER_FUNCS(name, type, member, bits)            \
 static inline u##bits btrfs_##name(struct extent_buffer *eb)           \
 {                                                                      \
-       __le##bits res;                                                 \
-       read_eb_member(eb, NULL, type, member, &res);                   \
-       return le##bits##_to_cpu(res);                                  \
+       int err;                                                        \
+       char *map_token;                                                \
+       char *kaddr;                                                    \
+       unsigned long map_start;                                        \
+       unsigned long map_len;                                          \
+       unsigned long offset = offsetof(type, member);                  \
+       err = map_extent_buffer(eb, offset,                             \
+                               sizeof(((type *)0)->member),            \
+                               &map_token, &kaddr,                     \
+                               &map_start, &map_len, KM_USER0);        \
+       if (!err) {                                                     \
+               __le##bits *tmp = (__le##bits *)(kaddr + offset -       \
+                                              map_start);              \
+               u##bits res = le##bits##_to_cpu(*tmp);                  \
+               unmap_extent_buffer(eb, map_token, KM_USER0);           \
+               return res;                                             \
+       } else {                                                        \
+               __le##bits res;                                         \
+               read_eb_member(eb, NULL, type, member, &res);           \
+               return le##bits##_to_cpu(res);                          \
+       }                                                               \
 }                                                                      \
 static inline void btrfs_set_##name(struct extent_buffer *eb,          \
                                    u##bits val)                        \
 {                                                                      \
-       val = cpu_to_le##bits(val);                                     \
-       write_eb_member(eb, NULL, type, member, &val);                  \
+       int err;                                                        \
+       char *map_token;                                                \
+       char *kaddr;                                                    \
+       unsigned long map_start;                                        \
+       unsigned long map_len;                                          \
+       unsigned long offset = offsetof(type, member);                  \
+       err = map_extent_buffer(eb, offset,                             \
+                               sizeof(((type *)0)->member),            \
+                               &map_token, &kaddr,                     \
+                               &map_start, &map_len, KM_USER0);        \
+       if (!err) {                                                     \
+               __le##bits *tmp = (__le##bits *)(kaddr + offset -       \
+                                              map_start);              \
+               *tmp = cpu_to_le##bits(val);                            \
+               unmap_extent_buffer(eb, map_token, KM_USER0);           \
+       } else {                                                        \
+               val = cpu_to_le##bits(val);                             \
+               write_eb_member(eb, NULL, type, member, &val);          \
+       }                                                               \
 }
 
 #define BTRFS_SETGET_STACK_FUNCS(name, type, member, bits)             \
index 0c1f90cbedb077b72ddc84094297523970478cbf..8242933a1d8b411bc2ed7e9dec030d04ec9ac263 100644 (file)
@@ -609,20 +609,20 @@ int close_ctree(struct btrfs_root *root)
 
 int btrfs_buffer_uptodate(struct extent_buffer *buf)
 {
-       struct inode *btree_inode = buf->pages[0]->mapping->host;
+       struct inode *btree_inode = buf->first_page->mapping->host;
        return extent_buffer_uptodate(&BTRFS_I(btree_inode)->extent_tree, buf);
 }
 
 int btrfs_set_buffer_uptodate(struct extent_buffer *buf)
 {
-       struct inode *btree_inode = buf->pages[0]->mapping->host;
+       struct inode *btree_inode = buf->first_page->mapping->host;
        return set_extent_buffer_uptodate(&BTRFS_I(btree_inode)->extent_tree,
                                          buf);
 }
 
 void btrfs_mark_buffer_dirty(struct extent_buffer *buf)
 {
-       struct btrfs_root *root = BTRFS_I(buf->pages[0]->mapping->host)->root;
+       struct btrfs_root *root = BTRFS_I(buf->first_page->mapping->host)->root;
        u64 transid = btrfs_header_generation(buf);
        struct inode *btree_inode = root->fs_info->btree_inode;
 
index 8bef309e1b3737dce85939e420b2e6bf1852226b..d2c733c68b4c2cbe2cfac86a0e1ce260027a4a5a 100644 (file)
@@ -18,6 +18,11 @@ struct kmem_cache *btrfs_cache_create(const char *name, size_t size,
 
 static struct kmem_cache *extent_map_cache;
 static struct kmem_cache *extent_state_cache;
+static struct kmem_cache *extent_buffer_cache;
+static LIST_HEAD(extent_buffers);
+static spinlock_t extent_buffers_lock;
+static int nr_extent_buffers;
+#define MAX_EXTENT_BUFFER_CACHE 128
 
 struct tree_entry {
        u64 start;
@@ -29,21 +34,33 @@ struct tree_entry {
 void __init extent_map_init(void)
 {
        extent_map_cache = btrfs_cache_create("extent_map",
-                                           sizeof(struct extent_map),
-                                           SLAB_DESTROY_BY_RCU,
+                                           sizeof(struct extent_map), 0,
                                            NULL);
        extent_state_cache = btrfs_cache_create("extent_state",
-                                           sizeof(struct extent_state),
-                                           SLAB_DESTROY_BY_RCU,
+                                           sizeof(struct extent_state), 0,
                                            NULL);
+       extent_buffer_cache = btrfs_cache_create("extent_buffers",
+                                           sizeof(struct extent_buffer), 0,
+                                           NULL);
+       spin_lock_init(&extent_buffers_lock);
 }
 
 void __exit extent_map_exit(void)
 {
+       struct extent_buffer *eb;
+
+       while (!list_empty(&extent_buffers)) {
+               eb = list_entry(extent_buffers.next,
+                               struct extent_buffer, list);
+               list_del(&eb->list);
+               kmem_cache_free(extent_buffer_cache, eb);
+       }
        if (extent_map_cache)
                kmem_cache_destroy(extent_map_cache);
        if (extent_state_cache)
                kmem_cache_destroy(extent_state_cache);
+       if (extent_buffer_cache)
+               kmem_cache_destroy(extent_buffer_cache);
 }
 
 void extent_map_tree_init(struct extent_map_tree *tree,
@@ -1858,6 +1875,48 @@ sector_t extent_bmap(struct address_space *mapping, sector_t iblock,
        return (em->block_start + start - em->start) >> inode->i_blkbits;
 }
 
+static struct extent_buffer *__alloc_extent_buffer(gfp_t mask)
+{
+       struct extent_buffer *eb = NULL;
+       spin_lock(&extent_buffers_lock);
+       if (!list_empty(&extent_buffers)) {
+               eb = list_entry(extent_buffers.next, struct extent_buffer,
+                               list);
+               list_del(&eb->list);
+               WARN_ON(nr_extent_buffers == 0);
+               nr_extent_buffers--;
+       }
+       spin_unlock(&extent_buffers_lock);
+       if (eb) {
+               memset(eb, 0, sizeof(*eb));
+               return eb;
+       }
+       return kmem_cache_zalloc(extent_buffer_cache, mask);
+}
+
+static void __free_extent_buffer(struct extent_buffer *eb)
+{
+       if (nr_extent_buffers >= MAX_EXTENT_BUFFER_CACHE) {
+               kmem_cache_free(extent_buffer_cache, eb);
+       } else {
+               spin_lock(&extent_buffers_lock);
+               list_add(&eb->list, &extent_buffers);
+               nr_extent_buffers++;
+               spin_unlock(&extent_buffers_lock);
+       }
+}
+
+static inline struct page *extent_buffer_page(struct extent_buffer *eb, int i)
+{
+       struct page *p;
+       if (i == 0)
+               return eb->first_page;
+       i += eb->start >> PAGE_CACHE_SHIFT;
+       p = find_get_page(eb->first_page->mapping, i);
+       page_cache_release(p);
+       return p;
+}
+
 struct extent_buffer *alloc_extent_buffer(struct extent_map_tree *tree,
                                          u64 start, unsigned long len,
                                          gfp_t mask)
@@ -1871,7 +1930,7 @@ struct extent_buffer *alloc_extent_buffer(struct extent_map_tree *tree,
        struct address_space *mapping = tree->mapping;
        int uptodate = 0;
 
-       eb = kzalloc(EXTENT_BUFFER_SIZE(num_pages), mask);
+       eb = __alloc_extent_buffer(mask);
        if (!eb || IS_ERR(eb))
                return NULL;
 
@@ -1881,9 +1940,16 @@ struct extent_buffer *alloc_extent_buffer(struct extent_map_tree *tree,
 
        for (i = 0; i < num_pages; i++, index++) {
                p = find_or_create_page(mapping, index, mask | __GFP_HIGHMEM);
-               if (!p)
+               if (!p) {
+                       /* make sure the free only frees the pages we've
+                        * grabbed a reference on
+                        */
+                       eb->len = i << PAGE_CACHE_SHIFT;
+                       eb->start &= ~((u64)PAGE_CACHE_SIZE - 1);
                        goto fail;
-               eb->pages[i] = p;
+               }
+               if (i == 0)
+                       eb->first_page = p;
                if (!PageUptodate(p))
                        uptodate = 0;
                unlock_page(p);
@@ -1909,7 +1975,7 @@ struct extent_buffer *find_extent_buffer(struct extent_map_tree *tree,
        struct page *p;
        struct address_space *mapping = tree->mapping;
 
-       eb = kzalloc(EXTENT_BUFFER_SIZE(num_pages), mask);
+       eb = __alloc_extent_buffer(mask);
        if (!eb || IS_ERR(eb))
                return NULL;
 
@@ -1919,9 +1985,16 @@ struct extent_buffer *find_extent_buffer(struct extent_map_tree *tree,
 
        for (i = 0; i < num_pages; i++, index++) {
                p = find_get_page(mapping, index);
-               if (!p)
+               if (!p) {
+                       /* make sure the free only frees the pages we've
+                        * grabbed a reference on
+                        */
+                       eb->len = i << PAGE_CACHE_SHIFT;
+                       eb->start &= ~((u64)PAGE_CACHE_SIZE - 1);
                        goto fail;
-               eb->pages[i] = p;
+               }
+               if (i == 0)
+                       eb->first_page = p;
        }
        return eb;
 fail:
@@ -1944,11 +2017,12 @@ void free_extent_buffer(struct extent_buffer *eb)
        num_pages = ((eb->start + eb->len - 1) >> PAGE_CACHE_SHIFT) -
                (eb->start >> PAGE_CACHE_SHIFT) + 1;
 
-       for (i = 0; i < num_pages; i++) {
-               if (eb->pages[i])
-                       page_cache_release(eb->pages[i]);
+       if (eb->first_page)
+               page_cache_release(eb->first_page);
+       for (i = 1; i < num_pages; i++) {
+               page_cache_release(extent_buffer_page(eb, i));
        }
-       kfree(eb);
+       __free_extent_buffer(eb);
 }
 EXPORT_SYMBOL(free_extent_buffer);
 
@@ -1968,7 +2042,7 @@ int clear_extent_buffer_dirty(struct extent_map_tree *tree,
                (eb->start >> PAGE_CACHE_SHIFT) + 1;
 
        for (i = 0; i < num_pages; i++) {
-               page = eb->pages[i];
+               page = extent_buffer_page(eb, i);
                lock_page(page);
                /*
                 * if we're on the last page or the first page and the
@@ -2021,7 +2095,7 @@ int set_extent_buffer_uptodate(struct extent_map_tree *tree,
        set_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
                            GFP_NOFS);
        for (i = 0; i < num_pages; i++) {
-               page = eb->pages[i];
+               page = extent_buffer_page(eb, i);
                if ((i == 0 && (eb->start & (PAGE_CACHE_SIZE - 1))) ||
                    ((i == num_pages - 1) &&
                     ((eb->start + eb->len - 1) & (PAGE_CACHE_SIZE - 1)))) {
@@ -2064,7 +2138,7 @@ int read_extent_buffer_pages(struct extent_map_tree *tree,
        num_pages = ((eb->start + eb->len - 1) >> PAGE_CACHE_SHIFT) -
                (eb->start >> PAGE_CACHE_SHIFT) + 1;
        for (i = 0; i < num_pages; i++) {
-               page = eb->pages[i];
+               page = extent_buffer_page(eb, i);
                if (PageUptodate(page)) {
                        continue;
                }
@@ -2090,7 +2164,7 @@ int read_extent_buffer_pages(struct extent_map_tree *tree,
        }
 
        for (i = 0; i < num_pages; i++) {
-               page = eb->pages[i];
+               page = extent_buffer_page(eb, i);
                wait_on_page_locked(page);
                if (!PageUptodate(page)) {
                        ret = -EIO;
@@ -2116,12 +2190,12 @@ void read_extent_buffer(struct extent_buffer *eb, void *dstv,
        WARN_ON(start > eb->len);
        WARN_ON(start + len > eb->start + eb->len);
 
-       page = eb->pages[i];
        offset = start & ((unsigned long)PAGE_CACHE_SIZE - 1);
        if (i == 0)
                offset += start_offset;
 
        while(len > 0) {
+               page = extent_buffer_page(eb, i);
                WARN_ON(!PageUptodate(page));
 
                cur = min(len, (PAGE_CACHE_SIZE - offset));
@@ -2134,7 +2208,6 @@ void read_extent_buffer(struct extent_buffer *eb, void *dstv,
                len -= cur;
                offset = 0;
                i++;
-               page = eb->pages[i];
        }
 }
 EXPORT_SYMBOL(read_extent_buffer);
@@ -2165,7 +2238,7 @@ int map_extent_buffer(struct extent_buffer *eb, unsigned long start,
        }
 
        // kaddr = kmap_atomic(eb->pages[i], km);
-       kaddr = page_address(eb->pages[i]);
+       kaddr = page_address(extent_buffer_page(eb, i));
        *token = kaddr;
        *map = kaddr + offset;
        *map_len = PAGE_CACHE_SIZE - offset;
@@ -2195,12 +2268,12 @@ int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
        WARN_ON(start > eb->len);
        WARN_ON(start + len > eb->start + eb->len);
 
-       page = eb->pages[i];
        offset = start & ((unsigned long)PAGE_CACHE_SIZE - 1);
        if (i == 0)
                offset += start_offset;
 
        while(len > 0) {
+               page = extent_buffer_page(eb, i);
                WARN_ON(!PageUptodate(page));
 
                cur = min(len, (PAGE_CACHE_SIZE - offset));
@@ -2216,7 +2289,6 @@ int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
                len -= cur;
                offset = 0;
                i++;
-               page = eb->pages[i];
        }
        return ret;
 }
@@ -2236,12 +2308,12 @@ void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
        WARN_ON(start > eb->len);
        WARN_ON(start + len > eb->start + eb->len);
 
-       page = eb->pages[i];
        offset = start & ((unsigned long)PAGE_CACHE_SIZE - 1);
        if (i == 0)
                offset += start_offset;
 
        while(len > 0) {
+               page = extent_buffer_page(eb, i);
                WARN_ON(!PageUptodate(page));
 
                cur = min(len, PAGE_CACHE_SIZE - offset);
@@ -2254,7 +2326,6 @@ void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
                len -= cur;
                offset = 0;
                i++;
-               page = eb->pages[i];
        }
 }
 EXPORT_SYMBOL(write_extent_buffer);
@@ -2272,12 +2343,12 @@ void memset_extent_buffer(struct extent_buffer *eb, char c,
        WARN_ON(start > eb->len);
        WARN_ON(start + len > eb->start + eb->len);
 
-       page = eb->pages[i];
        offset = start & ((unsigned long)PAGE_CACHE_SIZE - 1);
        if (i == 0)
                offset += start_offset;
 
        while(len > 0) {
+               page = extent_buffer_page(eb, i);
                WARN_ON(!PageUptodate(page));
 
                cur = min(len, PAGE_CACHE_SIZE - offset);
@@ -2289,7 +2360,6 @@ void memset_extent_buffer(struct extent_buffer *eb, char c,
                len -= cur;
                offset = 0;
                i++;
-               page = eb->pages[i];
        }
 }
 EXPORT_SYMBOL(memset_extent_buffer);
@@ -2313,7 +2383,7 @@ void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
                offset += start_offset;
 
        while(len > 0) {
-               page = dst->pages[i];
+               page = extent_buffer_page(dst, i);
                WARN_ON(!PageUptodate(page));
 
                cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
@@ -2414,7 +2484,8 @@ void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
                cur = min(cur, (unsigned long)(PAGE_CACHE_SIZE -
                                               dst_off_in_page));
 
-               copy_pages(dst->pages[dst_i], dst->pages[src_i],
+               copy_pages(extent_buffer_page(dst, dst_i),
+                          extent_buffer_page(dst, src_i),
                           dst_off_in_page, src_off_in_page, cur);
 
                src_offset += cur;
@@ -2467,7 +2538,8 @@ void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
                cur = min(len, src_off_in_page + 1);
                cur = min(cur, dst_off_in_page + 1);
 // printk("move pages orig dst %lu src %lu len %lu, this %lu %lu %lu\n", dst_offset, src_offset, len, dst_off_in_page - cur + 1, src_off_in_page - cur + 1, cur);
-               move_pages(dst->pages[dst_i], dst->pages[src_i],
+               move_pages(extent_buffer_page(dst, dst_i),
+                          extent_buffer_page(dst, src_i),
                           dst_off_in_page - cur + 1,
                           src_off_in_page - cur + 1, cur);
 
index 895789039ddd9faf9c36b15285292bb5e8805343..3b3abf34c4038a4aedd06003110b2870e62ebe5f 100644 (file)
@@ -62,14 +62,13 @@ struct extent_state {
        struct list_head list;
 };
 
-#define EXTENT_BUFFER_SIZE(nr) (sizeof(struct extent_buffer) + \
-                              (nr - 1) * sizeof(struct page *))
 struct extent_buffer {
        u64 start;
        unsigned long len;
        atomic_t refs;
        int flags;
-       struct page *pages[];
+       struct list_head list;
+       struct page *first_page;
 };
 
 typedef struct extent_map *(get_extent_t)(struct inode *inode,