ext4: Don't allow lg prealloc list to be grow large.
authorAneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Wed, 23 Jul 2008 18:14:05 +0000 (14:14 -0400)
committerTheodore Ts'o <tytso@mit.edu>
Wed, 23 Jul 2008 18:14:05 +0000 (14:14 -0400)
Currently, the locality group prealloc list is freed only when there
is a block allocation failure. This can result in large number of
entries in the preallocation list making ext4_mb_use_preallocated()
expensive.

To fix this, we convert the locality group prealloc list to a hash
list. The hash index is the order of number of blocks in the prealloc
space with a max order of 9. When adding prealloc space to the list we
make sure total entries for each order does not exceed 8. If it is
more than 8 we discard few entries and make sure the we have only <= 5
entries.

Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
fs/ext4/mballoc.c
fs/ext4/mballoc.h

index 49bec8404c5fe122974ee68363c6faa588b2d9f4..865e9ddb44d406d298da48b0137a4f426f8d45dd 100644 (file)
@@ -2480,7 +2480,7 @@ err_freesgi:
 int ext4_mb_init(struct super_block *sb, int needs_recovery)
 {
        struct ext4_sb_info *sbi = EXT4_SB(sb);
-       unsigned i;
+       unsigned i, j;
        unsigned offset;
        unsigned max;
        int ret;
@@ -2552,7 +2552,8 @@ int ext4_mb_init(struct super_block *sb, int needs_recovery)
                struct ext4_locality_group *lg;
                lg = &sbi->s_locality_groups[i];
                mutex_init(&lg->lg_mutex);
-               INIT_LIST_HEAD(&lg->lg_prealloc_list);
+               for (j = 0; j < PREALLOC_TB_SIZE; j++)
+                       INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
                spin_lock_init(&lg->lg_prealloc_lock);
        }
 
@@ -3263,6 +3264,7 @@ static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
                                struct ext4_prealloc_space *pa)
 {
        unsigned int len = ac->ac_o_ex.fe_len;
+
        ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
                                        &ac->ac_b_ex.fe_group,
                                        &ac->ac_b_ex.fe_start);
@@ -3285,6 +3287,7 @@ static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
 static noinline_for_stack int
 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
 {
+       int order, i;
        struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
        struct ext4_locality_group *lg;
        struct ext4_prealloc_space *pa;
@@ -3325,22 +3328,29 @@ ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
        lg = ac->ac_lg;
        if (lg == NULL)
                return 0;
-
-       rcu_read_lock();
-       list_for_each_entry_rcu(pa, &lg->lg_prealloc_list, pa_inode_list) {
-               spin_lock(&pa->pa_lock);
-               if (pa->pa_deleted == 0 && pa->pa_free >= ac->ac_o_ex.fe_len) {
-                       atomic_inc(&pa->pa_count);
-                       ext4_mb_use_group_pa(ac, pa);
+       order  = fls(ac->ac_o_ex.fe_len) - 1;
+       if (order > PREALLOC_TB_SIZE - 1)
+               /* The max size of hash table is PREALLOC_TB_SIZE */
+               order = PREALLOC_TB_SIZE - 1;
+
+       for (i = order; i < PREALLOC_TB_SIZE; i++) {
+               rcu_read_lock();
+               list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
+                                       pa_inode_list) {
+                       spin_lock(&pa->pa_lock);
+                       if (pa->pa_deleted == 0 &&
+                                       pa->pa_free >= ac->ac_o_ex.fe_len) {
+                               atomic_inc(&pa->pa_count);
+                               ext4_mb_use_group_pa(ac, pa);
+                               spin_unlock(&pa->pa_lock);
+                               ac->ac_criteria = 20;
+                               rcu_read_unlock();
+                               return 1;
+                       }
                        spin_unlock(&pa->pa_lock);
-                       ac->ac_criteria = 20;
-                       rcu_read_unlock();
-                       return 1;
                }
-               spin_unlock(&pa->pa_lock);
+               rcu_read_unlock();
        }
-       rcu_read_unlock();
-
        return 0;
 }
 
@@ -3563,6 +3573,7 @@ ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
        pa->pa_free = pa->pa_len;
        atomic_set(&pa->pa_count, 1);
        spin_lock_init(&pa->pa_lock);
+       INIT_LIST_HEAD(&pa->pa_inode_list);
        pa->pa_deleted = 0;
        pa->pa_linear = 1;
 
@@ -3583,10 +3594,10 @@ ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
        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_tail_rcu(&pa->pa_inode_list, &lg->lg_prealloc_list);
-       spin_unlock(pa->pa_obj_lock);
-
+       /*
+        * We will later add the new pa to the right bucket
+        * after updating the pa_free in ext4_mb_release_context
+        */
        return 0;
 }
 
@@ -4123,22 +4134,168 @@ ext4_mb_initialize_context(struct ext4_allocation_context *ac,
 
 }
 
+static noinline_for_stack void
+ext4_mb_discard_lg_preallocations(struct super_block *sb,
+                                       struct ext4_locality_group *lg,
+                                       int order, int total_entries)
+{
+       ext4_group_t group = 0;
+       struct ext4_buddy e4b;
+       struct list_head discard_list;
+       struct ext4_prealloc_space *pa, *tmp;
+       struct ext4_allocation_context *ac;
+
+       mb_debug("discard locality group preallocation\n");
+
+       INIT_LIST_HEAD(&discard_list);
+       ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
+
+       spin_lock(&lg->lg_prealloc_lock);
+       list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
+                                               pa_inode_list) {
+               spin_lock(&pa->pa_lock);
+               if (atomic_read(&pa->pa_count)) {
+                       /*
+                        * This is the pa that we just used
+                        * for block allocation. So don't
+                        * free that
+                        */
+                       spin_unlock(&pa->pa_lock);
+                       continue;
+               }
+               if (pa->pa_deleted) {
+                       spin_unlock(&pa->pa_lock);
+                       continue;
+               }
+               /* only lg prealloc space */
+               BUG_ON(!pa->pa_linear);
+
+               /* seems this one can be freed ... */
+               pa->pa_deleted = 1;
+               spin_unlock(&pa->pa_lock);
+
+               list_del_rcu(&pa->pa_inode_list);
+               list_add(&pa->u.pa_tmp_list, &discard_list);
+
+               total_entries--;
+               if (total_entries <= 5) {
+                       /*
+                        * we want to keep only 5 entries
+                        * allowing it to grow to 8. This
+                        * mak sure we don't call discard
+                        * soon for this list.
+                        */
+                       break;
+               }
+       }
+       spin_unlock(&lg->lg_prealloc_lock);
+
+       list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
+
+               ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
+               if (ext4_mb_load_buddy(sb, group, &e4b)) {
+                       ext4_error(sb, __func__, "Error in loading buddy "
+                                       "information for %lu\n", group);
+                       continue;
+               }
+               ext4_lock_group(sb, group);
+               list_del(&pa->pa_group_list);
+               ext4_mb_release_group_pa(&e4b, pa, ac);
+               ext4_unlock_group(sb, group);
+
+               ext4_mb_release_desc(&e4b);
+               list_del(&pa->u.pa_tmp_list);
+               call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
+       }
+       if (ac)
+               kmem_cache_free(ext4_ac_cachep, ac);
+}
+
+/*
+ * We have incremented pa_count. So it cannot be freed at this
+ * point. Also we hold lg_mutex. So no parallel allocation is
+ * possible from this lg. That means pa_free cannot be updated.
+ *
+ * A parallel ext4_mb_discard_group_preallocations is possible.
+ * which can cause the lg_prealloc_list to be updated.
+ */
+
+static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
+{
+       int order, added = 0, lg_prealloc_count = 1;
+       struct super_block *sb = ac->ac_sb;
+       struct ext4_locality_group *lg = ac->ac_lg;
+       struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
+
+       order = fls(pa->pa_free) - 1;
+       if (order > PREALLOC_TB_SIZE - 1)
+               /* The max size of hash table is PREALLOC_TB_SIZE */
+               order = PREALLOC_TB_SIZE - 1;
+       /* Add the prealloc space to lg */
+       rcu_read_lock();
+       list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
+                                               pa_inode_list) {
+               spin_lock(&tmp_pa->pa_lock);
+               if (tmp_pa->pa_deleted) {
+                       spin_unlock(&pa->pa_lock);
+                       continue;
+               }
+               if (!added && pa->pa_free < tmp_pa->pa_free) {
+                       /* Add to the tail of the previous entry */
+                       list_add_tail_rcu(&pa->pa_inode_list,
+                                               &tmp_pa->pa_inode_list);
+                       added = 1;
+                       /*
+                        * we want to count the total
+                        * number of entries in the list
+                        */
+               }
+               spin_unlock(&tmp_pa->pa_lock);
+               lg_prealloc_count++;
+       }
+       if (!added)
+               list_add_tail_rcu(&pa->pa_inode_list,
+                                       &lg->lg_prealloc_list[order]);
+       rcu_read_unlock();
+
+       /* Now trim the list to be not more than 8 elements */
+       if (lg_prealloc_count > 8) {
+               ext4_mb_discard_lg_preallocations(sb, lg,
+                                               order, lg_prealloc_count);
+               return;
+       }
+       return ;
+}
+
 /*
  * release all resource we used in allocation
  */
 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
 {
-       if (ac->ac_pa) {
-               if (ac->ac_pa->pa_linear) {
+       struct ext4_prealloc_space *pa = ac->ac_pa;
+       if (pa) {
+               if (pa->pa_linear) {
                        /* 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;
-                       ac->ac_pa->pa_lstart += ac->ac_b_ex.fe_len;
-                       ac->ac_pa->pa_free -= ac->ac_b_ex.fe_len;
-                       ac->ac_pa->pa_len -= ac->ac_b_ex.fe_len;
-                       spin_unlock(&ac->ac_pa->pa_lock);
+                       spin_lock(&pa->pa_lock);
+                       pa->pa_pstart += ac->ac_b_ex.fe_len;
+                       pa->pa_lstart += ac->ac_b_ex.fe_len;
+                       pa->pa_free -= ac->ac_b_ex.fe_len;
+                       pa->pa_len -= ac->ac_b_ex.fe_len;
+                       spin_unlock(&pa->pa_lock);
+                       /*
+                        * We want to add the pa to the right bucket.
+                        * Remove it from the list and while adding
+                        * make sure the list to which we are adding
+                        * doesn't grow big.
+                        */
+                       if (likely(pa->pa_free)) {
+                               spin_lock(pa->pa_obj_lock);
+                               list_del_rcu(&pa->pa_inode_list);
+                               spin_unlock(pa->pa_obj_lock);
+                               ext4_mb_add_n_trim(ac);
+                       }
                }
-               ext4_mb_put_pa(ac, ac->ac_sb, ac->ac_pa);
+               ext4_mb_put_pa(ac, ac->ac_sb, pa);
        }
        if (ac->ac_bitmap_page)
                page_cache_release(ac->ac_bitmap_page);
index bfe6add46bcfdaa2988a449b2c6cd316c4802419..c7c9906c2a754dec0953412b1c0843873ef7f6b9 100644 (file)
@@ -164,11 +164,17 @@ struct ext4_free_extent {
  * Locality group:
  *   we try to group all related changes together
  *   so that writeback can flush/allocate them together as well
+ *   Size of lg_prealloc_list hash is determined by MB_DEFAULT_GROUP_PREALLOC
+ *   (512). We store prealloc space into the hash based on the pa_free blocks
+ *   order value.ie, fls(pa_free)-1;
  */
+#define PREALLOC_TB_SIZE 10
 struct ext4_locality_group {
        /* for allocator */
-       struct mutex            lg_mutex;       /* to serialize allocates */
-       struct list_head        lg_prealloc_list;/* list of preallocations */
+       /* to serialize allocates */
+       struct mutex            lg_mutex;
+       /* list of preallocations */
+       struct list_head        lg_prealloc_list[PREALLOC_TB_SIZE];
        spinlock_t              lg_prealloc_lock;
 };