When mouting ext4 with -o noextents, request for
file data blocks from inode prealloc space.
Signed-off-by: Aneesh Kumar K.V <[email protected]>
---
fs/ext4/balloc.c | 9 ++++++++-
fs/ext4/ext4.h | 5 +++--
fs/ext4/inode.c | 20 ++++++++++++--------
3 files changed, 23 insertions(+), 11 deletions(-)
diff --git a/fs/ext4/balloc.c b/fs/ext4/balloc.c
index da99437..769b2b3 100644
--- a/fs/ext4/balloc.c
+++ b/fs/ext4/balloc.c
@@ -1878,7 +1878,8 @@ ext4_fsblk_t ext4_new_block(handle_t *handle, struct inode *inode,
}
ext4_fsblk_t ext4_new_blocks(handle_t *handle, struct inode *inode,
- ext4_fsblk_t goal, unsigned long *count, int *errp)
+ ext4_lblk_t iblock, ext4_fsblk_t goal,
+ unsigned long *count, int *errp)
{
struct ext4_allocation_request ar;
ext4_fsblk_t ret;
@@ -1892,6 +1893,12 @@ ext4_fsblk_t ext4_new_blocks(handle_t *handle, struct inode *inode,
ar.inode = inode;
ar.goal = goal;
ar.len = *count;
+ ar.logical = iblock;
+ if (S_ISREG(inode->i_mode))
+ ar.flags = EXT4_MB_HINT_DATA;
+ else
+ /* disable in-core preallocation for non-regular files */
+ ar.flags = 0;
ret = ext4_mb_new_blocks(handle, &ar, errp);
*count = ar.len;
return ret;
diff --git a/fs/ext4/ext4.h b/fs/ext4/ext4.h
index 8158083..1bd8e28 100644
--- a/fs/ext4/ext4.h
+++ b/fs/ext4/ext4.h
@@ -960,8 +960,9 @@ extern unsigned long ext4_bg_num_gdb(struct super_block *sb,
ext4_group_t group);
extern ext4_fsblk_t ext4_new_block (handle_t *handle, struct inode *inode,
ext4_fsblk_t goal, int *errp);
-extern ext4_fsblk_t ext4_new_blocks (handle_t *handle, struct inode *inode,
- ext4_fsblk_t goal, unsigned long *count, int *errp);
+extern ext4_fsblk_t ext4_new_blocks(handle_t *handle, struct inode *inode,
+ ext4_lblk_t iblock, ext4_fsblk_t goal,
+ unsigned long *count, int *errp);
extern ext4_fsblk_t ext4_new_blocks_old(handle_t *handle, struct inode *inode,
ext4_fsblk_t goal, unsigned long *count, int *errp);
extern void ext4_free_blocks (handle_t *handle, struct inode *inode,
diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
index 8d97077..0d1923e 100644
--- a/fs/ext4/inode.c
+++ b/fs/ext4/inode.c
@@ -508,8 +508,9 @@ static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned long blks,
* direct blocks
*/
static int ext4_alloc_blocks(handle_t *handle, struct inode *inode,
- ext4_fsblk_t goal, int indirect_blks, int blks,
- ext4_fsblk_t new_blocks[4], int *err)
+ ext4_lblk_t iblock, ext4_fsblk_t goal,
+ int indirect_blks, int blks,
+ ext4_fsblk_t new_blocks[4], int *err)
{
int target, i;
unsigned long count = 0;
@@ -530,7 +531,8 @@ static int ext4_alloc_blocks(handle_t *handle, struct inode *inode,
while (1) {
count = target;
/* allocating blocks for indirect blocks and direct blocks */
- current_block = ext4_new_blocks(handle,inode,goal,&count,err);
+ current_block = ext4_new_blocks(handle, inode, iblock,
+ goal, &count, err);
if (*err)
goto failed_out;
@@ -584,8 +586,9 @@ static int ext4_alloc_blocks(handle_t *handle, struct inode *inode,
* as described above and return 0.
*/
static int ext4_alloc_branch(handle_t *handle, struct inode *inode,
- int indirect_blks, int *blks, ext4_fsblk_t goal,
- ext4_lblk_t *offsets, Indirect *branch)
+ ext4_lblk_t iblock, int indirect_blks,
+ int *blks, ext4_fsblk_t goal,
+ ext4_lblk_t *offsets, Indirect *branch)
{
int blocksize = inode->i_sb->s_blocksize;
int i, n = 0;
@@ -595,7 +598,7 @@ static int ext4_alloc_branch(handle_t *handle, struct inode *inode,
ext4_fsblk_t new_blocks[4];
ext4_fsblk_t current_block;
- num = ext4_alloc_blocks(handle, inode, goal, indirect_blks,
+ num = ext4_alloc_blocks(handle, inode, iblock, goal, indirect_blks,
*blks, new_blocks, &err);
if (err)
return err;
@@ -855,8 +858,9 @@ int ext4_get_blocks_handle(handle_t *handle, struct inode *inode,
/*
* Block out ext4_truncate while we alter the tree
*/
- err = ext4_alloc_branch(handle, inode, indirect_blks, &count, goal,
- offsets + (partial - chain), partial);
+ err = ext4_alloc_branch(handle, inode, iblock, indirect_blks,
+ &count, goal,
+ offsets + (partial - chain), partial);
/*
* The ext4_splice_branch call will free and forget any buffers
--
1.5.5.1.211.g65ea3.dirty
This adds a per inode meta-block prealloc space from which
meta-data block requests are served. This help in making
sure meta-data block are closer. This is needed to speedup
unlink of the file. Any new prealloc space is allocated near
the goal block specified. The goal block is the last block
allocated for the file. So we don't keep the data-block and
meta-data block far apart.
Signed-off-by: Aneesh Kumar K.V <[email protected]>
---
fs/ext4/balloc.c | 27 +++++-
fs/ext4/ext4.h | 26 +++--
fs/ext4/ext4_i.h | 1 +
fs/ext4/extents.c | 6 +-
fs/ext4/inode.c | 54 +++++++++--
fs/ext4/mballoc.c | 266 ++++++++++++++++++++++++++++++++++++++++++++++++-----
fs/ext4/mballoc.h | 7 +-
fs/ext4/super.c | 1 +
fs/ext4/xattr.c | 2 +-
9 files changed, 335 insertions(+), 55 deletions(-)
diff --git a/fs/ext4/balloc.c b/fs/ext4/balloc.c
index 769b2b3..5c80eb5 100644
--- a/fs/ext4/balloc.c
+++ b/fs/ext4/balloc.c
@@ -1857,7 +1857,7 @@ ext4_fsblk_t ext4_new_blocks_old(handle_t *handle, struct inode *inode,
return 0;
}
-ext4_fsblk_t ext4_new_block(handle_t *handle, struct inode *inode,
+ext4_fsblk_t ext4_new_meta_block(handle_t *handle, struct inode *inode,
ext4_fsblk_t goal, int *errp)
{
struct ext4_allocation_request ar;
@@ -1873,9 +1873,30 @@ ext4_fsblk_t ext4_new_block(handle_t *handle, struct inode *inode,
ar.inode = inode;
ar.goal = goal;
ar.len = 1;
+ ar.flags = EXT4_MB_HINT_META_DATA;
ret = ext4_mb_new_blocks(handle, &ar, errp);
return ret;
}
+ext4_fsblk_t ext4_new_meta_blocks(handle_t *handle, struct inode *inode,
+ ext4_fsblk_t goal, unsigned long *count, int *errp)
+{
+ struct ext4_allocation_request ar;
+ ext4_fsblk_t ret;
+
+ if (!test_opt(inode->i_sb, MBALLOC)) {
+ ret = ext4_new_blocks_old(handle, inode, goal, count, errp);
+ return ret;
+ }
+
+ memset(&ar, 0, sizeof(ar));
+ ar.inode = inode;
+ ar.goal = goal;
+ ar.len = *count;
+ ar.flags = EXT4_MB_HINT_META_DATA;
+ ret = ext4_mb_new_blocks(handle, &ar, errp);
+ *count = ar.len;
+ return ret;
+}
ext4_fsblk_t ext4_new_blocks(handle_t *handle, struct inode *inode,
ext4_lblk_t iblock, ext4_fsblk_t goal,
@@ -1895,10 +1916,10 @@ ext4_fsblk_t ext4_new_blocks(handle_t *handle, struct inode *inode,
ar.len = *count;
ar.logical = iblock;
if (S_ISREG(inode->i_mode))
- ar.flags = EXT4_MB_HINT_DATA;
+ ar.flags = EXT4_MB_HINT_FILE_DATA;
else
/* disable in-core preallocation for non-regular files */
- ar.flags = 0;
+ ar.flags = EXT4_MB_HINT_DIR_DATA;
ret = ext4_mb_new_blocks(handle, &ar, errp);
*count = ar.len;
return ret;
diff --git a/fs/ext4/ext4.h b/fs/ext4/ext4.h
index 1bd8e28..b4bd67f 100644
--- a/fs/ext4/ext4.h
+++ b/fs/ext4/ext4.h
@@ -55,25 +55,27 @@
#define EXT4_MULTIBLOCK_ALLOCATOR 1
/* prefer goal again. length */
-#define EXT4_MB_HINT_MERGE 1
+#define EXT4_MB_HINT_MERGE 0x001
/* blocks already reserved */
-#define EXT4_MB_HINT_RESERVED 2
+#define EXT4_MB_HINT_RESERVED 0x002
/* metadata is being allocated */
-#define EXT4_MB_HINT_METADATA 4
+#define EXT4_MB_HINT_METADATA 0x004
/* first blocks in the file */
-#define EXT4_MB_HINT_FIRST 8
+#define EXT4_MB_HINT_FIRST 0x008
/* search for the best chunk */
-#define EXT4_MB_HINT_BEST 16
+#define EXT4_MB_HINT_BEST 0x010
/* data is being allocated */
-#define EXT4_MB_HINT_DATA 32
+#define EXT4_MB_HINT_FILE_DATA 0x020
+#define EXT4_MB_HINT_DIR_DATA 0x040
+#define EXT4_MB_HINT_META_DATA 0x080
/* don't preallocate (for tails) */
-#define EXT4_MB_HINT_NOPREALLOC 64
+#define EXT4_MB_HINT_NOPREALLOC 0x100
/* allocate for locality group */
-#define EXT4_MB_HINT_GROUP_ALLOC 128
+#define EXT4_MB_HINT_GROUP_ALLOC 0x200
/* allocate goal blocks or none */
-#define EXT4_MB_HINT_GOAL_ONLY 256
+#define EXT4_MB_HINT_GOAL_ONLY 0x400
/* goal is meaningful */
-#define EXT4_MB_HINT_TRY_GOAL 512
+#define EXT4_MB_HINT_TRY_GOAL 0x800
struct ext4_allocation_request {
/* target inode for block we're allocating */
@@ -958,8 +960,10 @@ extern ext4_grpblk_t ext4_block_group_offset(struct super_block *sb,
extern int ext4_bg_has_super(struct super_block *sb, ext4_group_t group);
extern unsigned long ext4_bg_num_gdb(struct super_block *sb,
ext4_group_t group);
-extern ext4_fsblk_t ext4_new_block (handle_t *handle, struct inode *inode,
+extern ext4_fsblk_t ext4_new_meta_block(handle_t *handle, struct inode *inode,
ext4_fsblk_t goal, int *errp);
+extern ext4_fsblk_t ext4_new_meta_blocks(handle_t *handle, struct inode *inode,
+ ext4_fsblk_t goal, unsigned long *count, int *errp);
extern ext4_fsblk_t ext4_new_blocks(handle_t *handle, struct inode *inode,
ext4_lblk_t iblock, ext4_fsblk_t goal,
unsigned long *count, int *errp);
diff --git a/fs/ext4/ext4_i.h b/fs/ext4/ext4_i.h
index 26a4ae2..4f11ec4 100644
--- a/fs/ext4/ext4_i.h
+++ b/fs/ext4/ext4_i.h
@@ -161,6 +161,7 @@ struct ext4_inode_info {
/* mballoc */
struct list_head i_prealloc_list;
+ struct list_head i_metaprealloc_list;
spinlock_t i_prealloc_lock;
};
diff --git a/fs/ext4/extents.c b/fs/ext4/extents.c
index 47929c4..c58ebd8 100644
--- a/fs/ext4/extents.c
+++ b/fs/ext4/extents.c
@@ -188,7 +188,7 @@ ext4_ext_new_block(handle_t *handle, struct inode *inode,
ext4_fsblk_t goal, newblock;
goal = ext4_ext_find_goal(inode, path, le32_to_cpu(ex->ee_block));
- newblock = ext4_new_block(handle, inode, goal, err);
+ newblock = ext4_new_meta_block(handle, inode, goal, err);
return newblock;
}
@@ -2690,10 +2690,10 @@ int ext4_ext_get_blocks(handle_t *handle, struct inode *inode,
ar.logical = iblock;
ar.len = allocated;
if (S_ISREG(inode->i_mode))
- ar.flags = EXT4_MB_HINT_DATA;
+ ar.flags = EXT4_MB_HINT_FILE_DATA;
else
/* disable in-core preallocation for non-regular files */
- ar.flags = 0;
+ ar.flags = EXT4_MB_HINT_DIR_DATA;
newblock = ext4_mb_new_blocks(handle, &ar, &err);
if (!newblock)
goto out2;
diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
index 0d1923e..3f4182f 100644
--- a/fs/ext4/inode.c
+++ b/fs/ext4/inode.c
@@ -513,7 +513,7 @@ static int ext4_alloc_blocks(handle_t *handle, struct inode *inode,
ext4_fsblk_t new_blocks[4], int *err)
{
int target, i;
- unsigned long count = 0;
+ long count = 0, blk_allocated = 0;
int index = 0;
ext4_fsblk_t current_block = 0;
int ret = 0;
@@ -526,12 +526,12 @@ static int ext4_alloc_blocks(handle_t *handle, struct inode *inode,
* the first direct block of this branch. That's the
* minimum number of blocks need to allocate(required)
*/
- target = blks + indirect_blks;
-
- while (1) {
+ /* first we try to allocate the indirect blocks */
+ target = indirect_blks;
+ while (target > 0) {
count = target;
/* allocating blocks for indirect blocks and direct blocks */
- current_block = ext4_new_blocks(handle, inode, iblock,
+ current_block = ext4_new_meta_blocks(handle, inode,
goal, &count, err);
if (*err)
goto failed_out;
@@ -542,16 +542,48 @@ static int ext4_alloc_blocks(handle_t *handle, struct inode *inode,
new_blocks[index++] = current_block++;
count--;
}
-
- if (count > 0)
+ if (count > 0) {
+ /*
+ * save the new block number
+ * for the first direct block
+ */
+ new_blocks[index] = current_block;
+ printk(KERN_INFO "%s returned more blocks than "
+ "requested\n", __func__);
+ WARN_ON(1);
break;
+ }
}
- /* save the new block number for the first direct block */
- new_blocks[index] = current_block;
-
+ target = blks - count ;
+ blk_allocated = count;
+ if (!target)
+ goto allocated;
+ /* Now allocate data blocks */
+ count = target;
+ /* allocating blocks for indirect blocks and direct blocks */
+ current_block = ext4_new_blocks(handle, inode, iblock,
+ goal, &count, err);
+ if (*err && (target == blks)) {
+ /*
+ * if the allocation failed and we didn't allocate
+ * any blocks before
+ */
+ goto failed_out;
+ }
+ if (!*err) {
+ if (target == blks) {
+ /*
+ * save the new block number
+ * for the first direct block
+ */
+ new_blocks[index] = current_block;
+ }
+ blk_allocated += count;
+ }
+allocated:
/* total number of blocks allocated for direct blocks */
- ret = count;
+ ret = blk_allocated;
*err = 0;
return ret;
failed_out:
diff --git a/fs/ext4/mballoc.c b/fs/ext4/mballoc.c
index ceee679..7871f46 100644
--- a/fs/ext4/mballoc.c
+++ b/fs/ext4/mballoc.c
@@ -1282,7 +1282,7 @@ static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
get_page(ac->ac_buddy_page);
/* store last allocated for subsequent stream allocation */
- if ((ac->ac_flags & EXT4_MB_HINT_DATA)) {
+ if ((ac->ac_flags & EXT4_MB_HINT_FILE_DATA)) {
spin_lock(&sbi->s_md_lock);
sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
@@ -1723,7 +1723,7 @@ ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
size = isize;
if (size < sbi->s_mb_stream_request &&
- (ac->ac_flags & EXT4_MB_HINT_DATA)) {
+ (ac->ac_flags & EXT4_MB_HINT_FILE_DATA)) {
/* TBD: may be hot point */
spin_lock(&sbi->s_md_lock);
ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
@@ -1744,7 +1744,6 @@ ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
* from the goal value specified
*/
group = ac->ac_g_ex.fe_group;
-
for (i = 0; i < EXT4_SB(sb)->s_groups_count; group++, i++) {
struct ext4_group_info *grp;
struct ext4_group_desc *desc;
@@ -2819,6 +2818,24 @@ ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
return err;
}
+static void
+ext4_mb_normalize_meta_data_request(struct ext4_allocation_context *ac)
+{
+ /*
+ * Need to find what the right nomalized block num should be
+ */
+ struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
+ /* fe_len should be power of 2 */
+ if (i_size_read(ac->ac_inode) >= sbi->s_mb_stream_request) {
+ /* large inode which is using inode prealloc */
+ ac->ac_g_ex.fe_len = 16;
+ } else {
+ ac->ac_g_ex.fe_len = 2;
+ }
+ mb_debug("#%u: goal %u blocks for meta-data group\n",
+ current->pid, ac->ac_g_ex.fe_len);
+}
+
/*
* here we normalize request for locality group
* Group request are normalized to s_strip size if we set the same via mount
@@ -2856,11 +2873,6 @@ ext4_mb_normalize_request(struct ext4_allocation_context *ac,
struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
struct ext4_prealloc_space *pa;
- /* do normalize only data requests, metadata requests
- do not need preallocation */
- if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
- return;
-
/* sometime caller may want exact blocks */
if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
return;
@@ -2870,6 +2882,21 @@ ext4_mb_normalize_request(struct ext4_allocation_context *ac,
if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
return;
+ /*
+ * Normalize only data and meta-data request
+ * Other block requests are not preallocated
+ */
+ if (ac->ac_flags & EXT4_MB_HINT_DIR_DATA)
+ return;
+
+ if (ac->ac_flags & EXT4_MB_HINT_META_DATA) {
+ /* meta-data preallocation space
+ * depends on the file size.
+ */
+ ext4_mb_normalize_meta_data_request(ac);
+ return;
+ }
+
if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
ext4_mb_normalize_group_request(ac);
return ;
@@ -3050,6 +3077,28 @@ static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
ext4_mb_store_history(ac);
}
+/*
+ * use blocks preallocated to meta-data prealloc space
+ */
+static void ext4_mb_use_meta_block_pa(struct ext4_allocation_context *ac,
+ struct ext4_prealloc_space *pa)
+{
+ unsigned int len = ac->ac_o_ex.fe_len;
+ if (len > pa->pa_free)
+ len = pa->pa_free;
+
+ ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
+ &ac->ac_b_ex.fe_group,
+ &ac->ac_b_ex.fe_start);
+ ac->ac_b_ex.fe_len = len;
+ ac->ac_status = AC_STATUS_FOUND;
+ ac->ac_pa = pa;
+
+ mb_debug("use %llu/%u from meta group pa %p\n", pa->pa_pstart, len, pa);
+ pa->pa_pstart += ac->ac_b_ex.fe_len;
+ pa->pa_free -= ac->ac_b_ex.fe_len;
+ pa->pa_len -= ac->ac_b_ex.fe_len;
+}
/*
* use blocks preallocated to inode
@@ -3113,9 +3162,32 @@ ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
struct ext4_locality_group *lg;
struct ext4_prealloc_space *pa;
- /* only data can be preallocated */
- if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
+ /*
+ * non-file and non-metadata always use regular allocator
+ */
+ if (ac->ac_flags & EXT4_MB_HINT_DIR_DATA)
+ return 0;
+
+ if (ac->ac_flags & EXT4_MB_HINT_META_DATA) {
+ /* meta-data allocation request */
+ rcu_read_lock();
+ list_for_each_entry_rcu(pa, &ei->i_metaprealloc_list,
+ pa_inode_list) {
+ /* found preallocated blocks, use them */
+ spin_lock(&pa->pa_lock);
+ if (pa->pa_deleted == 0 && pa->pa_free) {
+ atomic_inc(&pa->pa_count);
+ ext4_mb_use_meta_block_pa(ac, pa);
+ spin_unlock(&pa->pa_lock);
+ ac->ac_criteria = 10;
+ rcu_read_unlock();
+ return 1;
+ }
+ spin_unlock(&pa->pa_lock);
+ }
+ rcu_read_unlock();
return 0;
+ }
/* first, try per-file preallocation */
rcu_read_lock();
@@ -3268,6 +3340,58 @@ static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
}
+static noinline int
+ext4_mb_new_meta_block_pa(struct ext4_allocation_context *ac)
+{
+ struct super_block *sb = ac->ac_sb;
+ struct ext4_prealloc_space *pa;
+ struct ext4_group_info *grp;
+ struct ext4_inode_info *ei;
+
+ /* preallocate only when found space is larger then requested */
+ BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
+ BUG_ON(ac->ac_status != AC_STATUS_FOUND);
+
+ pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
+ if (pa == NULL)
+ return -ENOMEM;
+
+ /* preallocation can change ac_b_ex, thus we store actually
+ * allocated blocks for history */
+ ac->ac_f_ex = ac->ac_b_ex;
+
+ pa->pa_lstart = 0;
+ pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
+ pa->pa_len = ac->ac_b_ex.fe_len;
+ pa->pa_free = pa->pa_len;
+ atomic_set(&pa->pa_count, 1);
+ spin_lock_init(&pa->pa_lock);
+ pa->pa_deleted = 0;
+ pa->pa_type = PA_META_PA;
+
+ mb_debug("new meta pa %p: %llu/%u\n", pa,
+ pa->pa_pstart, pa->pa_len);
+
+ ext4_mb_use_meta_block_pa(ac, pa);
+ atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
+
+ ei = EXT4_I(ac->ac_inode);
+ grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
+
+ pa->pa_obj_lock = &ei->i_prealloc_lock;
+ pa->pa_inode = ac->ac_inode;
+
+ ext4_lock_group(sb, ac->ac_b_ex.fe_group);
+ list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
+ ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
+
+ spin_lock(pa->pa_obj_lock);
+ list_add_rcu(&pa->pa_inode_list, &ei->i_metaprealloc_list);
+ spin_unlock(pa->pa_obj_lock);
+
+ return 0;
+}
+
/*
* creates new preallocated space for given inode
*/
@@ -3331,7 +3455,7 @@ ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
atomic_set(&pa->pa_count, 1);
spin_lock_init(&pa->pa_lock);
pa->pa_deleted = 0;
- pa->pa_linear = 0;
+ pa->pa_type = PA_INODE_PA;
mb_debug("new inode pa %p: %llu/%u for %u\n", pa,
pa->pa_pstart, pa->pa_len, pa->pa_lstart);
@@ -3388,7 +3512,7 @@ ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
atomic_set(&pa->pa_count, 1);
spin_lock_init(&pa->pa_lock);
pa->pa_deleted = 0;
- pa->pa_linear = 1;
+ pa->pa_type = PA_GROUP_PA;
mb_debug("new group pa %p: %llu/%u for %u\n", pa,
pa->pa_pstart, pa->pa_len, pa->pa_lstart);
@@ -3418,7 +3542,9 @@ static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
{
int err;
- if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
+ if (ac->ac_flags & EXT4_MB_HINT_META_DATA)
+ err = ext4_mb_new_meta_block_pa(ac);
+ else if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
err = ext4_mb_new_group_pa(ac);
else
err = ext4_mb_new_inode_pa(ac);
@@ -3500,6 +3626,35 @@ ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
return err;
}
+static noinline int ext4_mb_release_meta_block_pa(struct ext4_buddy *e4b,
+ struct ext4_prealloc_space *pa,
+ struct ext4_allocation_context *ac)
+{
+ struct super_block *sb = e4b->bd_sb;
+ ext4_group_t group;
+ ext4_grpblk_t bit;
+
+ if (ac)
+ ac->ac_op = EXT4_MB_HISTORY_DISCARD;
+
+ BUG_ON(pa->pa_deleted == 0);
+ ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
+ BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
+ mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
+ atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
+
+ if (ac) {
+ ac->ac_sb = sb;
+ ac->ac_inode = NULL;
+ ac->ac_b_ex.fe_group = group;
+ ac->ac_b_ex.fe_start = bit;
+ ac->ac_b_ex.fe_len = pa->pa_len;
+ ac->ac_b_ex.fe_logical = 0;
+ ext4_mb_store_history(ac);
+ }
+
+ return 0;
+}
static noinline_for_stack int
ext4_mb_release_group_pa(struct ext4_buddy *e4b,
@@ -3630,11 +3785,18 @@ ext4_mb_discard_group_preallocations(struct super_block *sb,
list_del_rcu(&pa->pa_inode_list);
spin_unlock(pa->pa_obj_lock);
- if (pa->pa_linear)
+ switch (pa->pa_type) {
+ case PA_META_PA:
+ ext4_mb_release_meta_block_pa(&e4b, pa, ac);
+ break;
+ case PA_GROUP_PA:
ext4_mb_release_group_pa(&e4b, pa, ac);
- else
- ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac);
-
+ break;
+ case PA_INODE_PA:
+ ext4_mb_release_inode_pa(&e4b,
+ bitmap_bh, pa, ac);
+ break;
+ }
list_del(&pa->u.pa_tmp_list);
call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
}
@@ -3669,10 +3831,8 @@ void ext4_mb_discard_inode_preallocations(struct inode *inode)
struct ext4_buddy e4b;
int err;
- if (!test_opt(sb, MBALLOC) || !S_ISREG(inode->i_mode)) {
- /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
+ if (!test_opt(sb, MBALLOC))
return;
- }
mb_debug("discard preallocation for inode %lu\n", inode->i_ino);
@@ -3682,6 +3842,49 @@ void ext4_mb_discard_inode_preallocations(struct inode *inode)
repeat:
/* first, collect all pa's in the inode */
spin_lock(&ei->i_prealloc_lock);
+ while (!list_empty(&ei->i_metaprealloc_list)) {
+ pa = list_entry(ei->i_metaprealloc_list.next,
+ struct ext4_prealloc_space, pa_inode_list);
+ BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
+ spin_lock(&pa->pa_lock);
+ if (atomic_read(&pa->pa_count)) {
+ /* this shouldn't happen often - nobody should
+ * use preallocation while we're discarding it */
+ spin_unlock(&pa->pa_lock);
+ spin_unlock(&ei->i_prealloc_lock);
+ printk(KERN_ERR "uh-oh! used pa while discarding\n");
+ WARN_ON(1);
+ schedule_timeout_uninterruptible(HZ);
+ goto repeat;
+
+ }
+ if (pa->pa_deleted == 0) {
+ pa->pa_deleted = 1;
+ spin_unlock(&pa->pa_lock);
+ list_del_rcu(&pa->pa_inode_list);
+ list_add(&pa->u.pa_tmp_list, &list);
+ continue;
+ }
+
+ /* someone is deleting pa right now */
+ spin_unlock(&pa->pa_lock);
+ spin_unlock(&ei->i_prealloc_lock);
+
+ /* we have to wait here because pa_deleted
+ * doesn't mean pa is already unlinked from
+ * the list. as we might be called from
+ * ->clear_inode() the inode will get freed
+ * and concurrent thread which is unlinking
+ * pa from inode's list may access already
+ * freed memory, bad-bad-bad */
+
+ /* XXX: if this happens too often, we can
+ * add a flag to force wait only in case
+ * of ->clear_inode(), but not in case of
+ * regular truncate */
+ schedule_timeout_uninterruptible(HZ);
+ goto repeat;
+ }
while (!list_empty(&ei->i_prealloc_list)) {
pa = list_entry(ei->i_prealloc_list.next,
struct ext4_prealloc_space, pa_inode_list);
@@ -3728,7 +3931,6 @@ void ext4_mb_discard_inode_preallocations(struct inode *inode)
spin_unlock(&ei->i_prealloc_lock);
list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
- BUG_ON(pa->pa_linear != 0);
ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
err = ext4_mb_load_buddy(sb, group, &e4b);
@@ -3743,7 +3945,18 @@ void ext4_mb_discard_inode_preallocations(struct inode *inode)
ext4_lock_group(sb, group);
list_del(&pa->pa_group_list);
- ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac);
+
+ switch (pa->pa_type) {
+ case PA_META_PA:
+ ext4_mb_release_meta_block_pa(&e4b, pa, ac);
+ break;
+ case PA_INODE_PA:
+ ext4_mb_release_inode_pa(&e4b,
+ bitmap_bh, pa, ac);
+ break;
+ default:
+ BUG();
+ }
ext4_unlock_group(sb, group);
ext4_mb_release_desc(&e4b);
@@ -3842,8 +4055,13 @@ static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
int bsbits = ac->ac_sb->s_blocksize_bits;
loff_t size, isize;
- if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
+ if (!(ac->ac_flags & EXT4_MB_HINT_FILE_DATA)) {
+ /*
+ * group and inode prealloc space is used
+ * only for file data
+ */
return;
+ }
size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
isize = i_size_read(ac->ac_inode) >> bsbits;
@@ -3947,7 +4165,7 @@ ext4_mb_initialize_context(struct ext4_allocation_context *ac,
static int ext4_mb_release_context(struct ext4_allocation_context *ac)
{
if (ac->ac_pa) {
- if (ac->ac_pa->pa_linear) {
+ if (ac->ac_pa->pa_type == PA_GROUP_PA) {
/* see comment in ext4_mb_use_group_pa() */
spin_lock(&ac->ac_pa->pa_lock);
ac->ac_pa->pa_pstart += ac->ac_b_ex.fe_len;
diff --git a/fs/ext4/mballoc.h b/fs/ext4/mballoc.h
index bfe6add..2cc8440 100644
--- a/fs/ext4/mballoc.h
+++ b/fs/ext4/mballoc.h
@@ -131,6 +131,10 @@ struct ext4_group_info {
#define EXT4_MB_GRP_NEED_INIT(grp) \
(test_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &((grp)->bb_state)))
+#define PA_INODE_PA 0
+#define PA_GROUP_PA 1
+#define PA_META_PA 2
+
struct ext4_prealloc_space {
struct list_head pa_inode_list;
@@ -146,8 +150,7 @@ struct ext4_prealloc_space {
ext4_lblk_t pa_lstart; /* log. block */
unsigned short pa_len; /* len of preallocated chunk */
unsigned short pa_free; /* how many blocks are free */
- unsigned short pa_linear; /* consumed in one direction
- * strictly, for grp prealloc */
+ unsigned short pa_type; /* Trype of prealloc space */
spinlock_t *pa_obj_lock;
struct inode *pa_inode; /* hack, for history only */
};
diff --git a/fs/ext4/super.c b/fs/ext4/super.c
index d70165a..cd7cac0 100644
--- a/fs/ext4/super.c
+++ b/fs/ext4/super.c
@@ -570,6 +570,7 @@ static struct inode *ext4_alloc_inode(struct super_block *sb)
ei->vfs_inode.i_version = 1;
memset(&ei->i_cached_extent, 0, sizeof(struct ext4_ext_cache));
INIT_LIST_HEAD(&ei->i_prealloc_list);
+ INIT_LIST_HEAD(&ei->i_metaprealloc_list);
spin_lock_init(&ei->i_prealloc_lock);
return &ei->vfs_inode;
}
diff --git a/fs/ext4/xattr.c b/fs/ext4/xattr.c
index 3fbc2c6..4c8c742 100644
--- a/fs/ext4/xattr.c
+++ b/fs/ext4/xattr.c
@@ -810,7 +810,7 @@ ext4_xattr_block_set(handle_t *handle, struct inode *inode,
/* We need to allocate a new block */
ext4_fsblk_t goal = ext4_group_first_block_no(sb,
EXT4_I(inode)->i_block_group);
- ext4_fsblk_t block = ext4_new_block(handle, inode,
+ ext4_fsblk_t block = ext4_new_meta_block(handle, inode,
goal, &error);
if (error)
goto cleanup;
--
1.5.5.1.211.g65ea3.dirty
On Thu, May 15, 2008 at 09:23:59PM +0530, Aneesh Kumar K.V wrote:
> This adds a per inode meta-block prealloc space from which
> meta-data block requests are served. This help in making
> sure meta-data block are closer. This is needed to speedup
> unlink of the file. Any new prealloc space is allocated near
> the goal block specified. The goal block is the last block
> allocated for the file. So we don't keep the data-block and
> meta-data block far apart.
>
The result can be found at
http://www.radian.org/~kvaneesh/ext4/meta-group/
This patch need to be further improved by grouping the meta-data
at FLEX_BG or block group level rather than per inode. I am sending
it so that others can take a look at the changes
-aneesh
Aneesh Kumar K.V wrote:
> On Thu, May 15, 2008 at 09:23:59PM +0530, Aneesh Kumar K.V wrote:
>> This adds a per inode meta-block prealloc space from which
>> meta-data block requests are served. This help in making
>> sure meta-data block are closer. This is needed to speedup
>> unlink of the file. Any new prealloc space is allocated near
>> the goal block specified. The goal block is the last block
>> allocated for the file. So we don't keep the data-block and
>> meta-data block far apart.
>>
>
> The result can be found at
>
> http://www.radian.org/~kvaneesh/ext4/meta-group/
Out of curiosity, do you have graphs similar to
http://www.radian.org/~kvaneesh/ext4/meta-group/ext4-noextents-metagroup.png
before and after your change?
Thanks,
-Eric
On Thu, May 15, 2008 at 11:14:51AM -0500, Eric Sandeen wrote:
> Aneesh Kumar K.V wrote:
> > On Thu, May 15, 2008 at 09:23:59PM +0530, Aneesh Kumar K.V wrote:
> >> This adds a per inode meta-block prealloc space from which
> >> meta-data block requests are served. This help in making
> >> sure meta-data block are closer. This is needed to speedup
> >> unlink of the file. Any new prealloc space is allocated near
> >> the goal block specified. The goal block is the last block
> >> allocated for the file. So we don't keep the data-block and
> >> meta-data block far apart.
> >>
> >
> > The result can be found at
> >
> > http://www.radian.org/~kvaneesh/ext4/meta-group/
>
> Out of curiosity, do you have graphs similar to
> http://www.radian.org/~kvaneesh/ext4/meta-group/ext4-noextents-metagroup.png
>
> before and after your change?
>
No, But I can do that.
-aneesh
On Thu, May 15, 2008 at 11:14:51AM -0500, Eric Sandeen wrote:
> Aneesh Kumar K.V wrote:
> > On Thu, May 15, 2008 at 09:23:59PM +0530, Aneesh Kumar K.V wrote:
> >> This adds a per inode meta-block prealloc space from which
> >> meta-data block requests are served. This help in making
> >> sure meta-data block are closer. This is needed to speedup
> >> unlink of the file. Any new prealloc space is allocated near
> >> the goal block specified. The goal block is the last block
> >> allocated for the file. So we don't keep the data-block and
> >> meta-data block far apart.
> >>
> >
> > The result can be found at
> >
> > http://www.radian.org/~kvaneesh/ext4/meta-group/
>
> Out of curiosity, do you have graphs similar to
> http://www.radian.org/~kvaneesh/ext4/meta-group/ext4-noextents-metagroup.png
>
> before and after your change?
http://www.radian.org/~kvaneesh/ext4/meta-group/exp1/
-aneesh
Aneesh Kumar K.V wrote:
> On Thu, May 15, 2008 at 11:14:51AM -0500, Eric Sandeen wrote:
>> Aneesh Kumar K.V wrote:
>>> On Thu, May 15, 2008 at 09:23:59PM +0530, Aneesh Kumar K.V wrote:
>>>> This adds a per inode meta-block prealloc space from which
>>>> meta-data block requests are served. This help in making
>>>> sure meta-data block are closer. This is needed to speedup
>>>> unlink of the file. Any new prealloc space is allocated near
>>>> the goal block specified. The goal block is the last block
>>>> allocated for the file. So we don't keep the data-block and
>>>> meta-data block far apart.
>>>>
>>> The result can be found at
>>>
>>> http://www.radian.org/~kvaneesh/ext4/meta-group/
>> Out of curiosity, do you have graphs similar to
>> http://www.radian.org/~kvaneesh/ext4/meta-group/ext4-noextents-metagroup.png
>>
>> before and after your change?
>
> http://www.radian.org/~kvaneesh/ext4/meta-group/exp1/
Thanks!
On:
http://www.radian.org/~kvaneesh/ext4/meta-group/exp1/ext4-noextents-withoutpatch-rm.png
it seems interesting; do you know what's going on between ~40s and 100s?
Also:
between:
http://www.radian.org/~kvaneesh/ext4/meta-group/exp1/ext4-noextents-read.png
and
http://www.radian.org/~kvaneesh/ext4/meta-group/exp1/ext4-noextents-withoutpatch-read.png
it looks like about a 10% slowdown on the read....
-Eric
