ext4: Patch up how we claim metadata blocks for quota purposes
Theodore Ts'o [Wed, 30 Dec 2009 19:20:45 +0000 (14:20 -0500)]
As reported in Kernel Bugzilla #14936, commit d21cd8f triggered a BUG
in the function ext4_da_update_reserve_space() found in
fs/ext4/inode.c.  The root cause of this BUG() was caused by the fact
that ext4_calc_metadata_amount() can severely over-estimate how many
metadata blocks will be needed, especially when using direct
block-mapped files.

In addition, it can also badly *under* estimate how much space is
needed, since ext4_calc_metadata_amount() assumes that the blocks are
contiguous, and this is not always true.  If the application is
writing blocks to a sparse file, the number of metadata blocks
necessary can be severly underestimated by the functions
ext4_da_reserve_space(), ext4_da_update_reserve_space() and
ext4_da_release_space().  This was the cause of the dq_claim_space
reports found on kerneloops.org.

Unfortunately, doing this right means that we need to massively
over-estimate the amount of free space needed.  So in some cases we
may need to force the inode to be written to disk asynchronously in
to avoid spurious quota failures.

http://bugzilla.kernel.org/show_bug.cgi?id=14936

Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>

fs/ext4/inode.c

index 3e3b454..84eeb8f 100644 (file)
@@ -1043,43 +1043,47 @@ static int ext4_calc_metadata_amount(struct inode *inode, int blocks)
        return ext4_indirect_calc_metadata_amount(inode, blocks);
 }
 
+/*
+ * Called with i_data_sem down, which is important since we can call
+ * ext4_discard_preallocations() from here.
+ */
 static void ext4_da_update_reserve_space(struct inode *inode, int used)
 {
        struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
-       int total, mdb, mdb_free, mdb_claim = 0;
-
-       spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
-       /* recalculate the number of metablocks still need to be reserved */
-       total = EXT4_I(inode)->i_reserved_data_blocks - used;
-       mdb = ext4_calc_metadata_amount(inode, total);
-
-       /* figure out how many metablocks to release */
-       BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
-       mdb_free = EXT4_I(inode)->i_reserved_meta_blocks - mdb;
-
-       if (mdb_free) {
-               /* Account for allocated meta_blocks */
-               mdb_claim = EXT4_I(inode)->i_allocated_meta_blocks;
-               BUG_ON(mdb_free < mdb_claim);
-               mdb_free -= mdb_claim;
-
-               /* update fs dirty blocks counter */
+       struct ext4_inode_info *ei = EXT4_I(inode);
+       int mdb_free = 0;
+
+       spin_lock(&ei->i_block_reservation_lock);
+       if (unlikely(used > ei->i_reserved_data_blocks)) {
+               ext4_msg(inode->i_sb, KERN_NOTICE, "%s: ino %lu, used %d "
+                        "with only %d reserved data blocks\n",
+                        __func__, inode->i_ino, used,
+                        ei->i_reserved_data_blocks);
+               WARN_ON(1);
+               used = ei->i_reserved_data_blocks;
+       }
+
+       /* Update per-inode reservations */
+       ei->i_reserved_data_blocks -= used;
+       used += ei->i_allocated_meta_blocks;
+       ei->i_reserved_meta_blocks -= ei->i_allocated_meta_blocks;
+       ei->i_allocated_meta_blocks = 0;
+       percpu_counter_sub(&sbi->s_dirtyblocks_counter, used);
+
+       if (ei->i_reserved_data_blocks == 0) {
+               /*
+                * We can release all of the reserved metadata blocks
+                * only when we have written all of the delayed
+                * allocation blocks.
+                */
+               mdb_free = ei->i_allocated_meta_blocks;
                percpu_counter_sub(&sbi->s_dirtyblocks_counter, mdb_free);
-               EXT4_I(inode)->i_allocated_meta_blocks = 0;
-               EXT4_I(inode)->i_reserved_meta_blocks = mdb;
+               ei->i_allocated_meta_blocks = 0;
        }
-
-       /* update per-inode reservations */
-       BUG_ON(used  > EXT4_I(inode)->i_reserved_data_blocks);
-       EXT4_I(inode)->i_reserved_data_blocks -= used;
-       percpu_counter_sub(&sbi->s_dirtyblocks_counter, used + mdb_claim);
        spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
 
-       vfs_dq_claim_block(inode, used + mdb_claim);
-
-       /*
-        * free those over-booking quota for metadata blocks
-        */
+       /* Update quota subsystem */
+       vfs_dq_claim_block(inode, used);
        if (mdb_free)
                vfs_dq_release_reservation_block(inode, mdb_free);
 
@@ -1088,7 +1092,8 @@ static void ext4_da_update_reserve_space(struct inode *inode, int used)
         * there aren't any writers on the inode, we can discard the
         * inode's preallocations.
         */
-       if (!total && (atomic_read(&inode->i_writecount) == 0))
+       if ((ei->i_reserved_data_blocks == 0) &&
+           (atomic_read(&inode->i_writecount) == 0))
                ext4_discard_preallocations(inode);
 }
 
@@ -1801,7 +1806,8 @@ static int ext4_da_reserve_space(struct inode *inode, int nrblocks)
 {
        int retries = 0;
        struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
-       unsigned long md_needed, mdblocks, total = 0;
+       struct ext4_inode_info *ei = EXT4_I(inode);
+       unsigned long md_needed, md_reserved, total = 0;
 
        /*
         * recalculate the amount of metadata blocks to reserve
@@ -1809,35 +1815,44 @@ static int ext4_da_reserve_space(struct inode *inode, int nrblocks)
         * worse case is one extent per block
         */
 repeat:
-       spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
-       total = EXT4_I(inode)->i_reserved_data_blocks + nrblocks;
-       mdblocks = ext4_calc_metadata_amount(inode, total);
-       BUG_ON(mdblocks < EXT4_I(inode)->i_reserved_meta_blocks);
-
-       md_needed = mdblocks - EXT4_I(inode)->i_reserved_meta_blocks;
+       spin_lock(&ei->i_block_reservation_lock);
+       md_reserved = ei->i_reserved_meta_blocks;
+       md_needed = ext4_calc_metadata_amount(inode, nrblocks);
        total = md_needed + nrblocks;
-       spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
+       spin_unlock(&ei->i_block_reservation_lock);
 
        /*
         * Make quota reservation here to prevent quota overflow
         * later. Real quota accounting is done at pages writeout
         * time.
         */
-       if (vfs_dq_reserve_block(inode, total))
+       if (vfs_dq_reserve_block(inode, total)) {
+               /* 
+                * We tend to badly over-estimate the amount of
+                * metadata blocks which are needed, so if we have
+                * reserved any metadata blocks, try to force out the
+                * inode and see if we have any better luck.
+                */
+               if (md_reserved && retries++ <= 3)
+                       goto retry;
                return -EDQUOT;
+       }
 
        if (ext4_claim_free_blocks(sbi, total)) {
                vfs_dq_release_reservation_block(inode, total);
                if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
+               retry:
+                       if (md_reserved)
+                               write_inode_now(inode, (retries == 3));
                        yield();
                        goto repeat;
                }
                return -ENOSPC;
        }
-       spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
-       EXT4_I(inode)->i_reserved_data_blocks += nrblocks;
-       EXT4_I(inode)->i_reserved_meta_blocks += md_needed;
-       spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
+       spin_lock(&ei->i_block_reservation_lock);
+       ei->i_reserved_data_blocks += nrblocks;
+       ei->i_reserved_meta_blocks += md_needed;
+       spin_unlock(&ei->i_block_reservation_lock);
 
        return 0;       /* success */
 }
@@ -1845,49 +1860,45 @@ repeat:
 static void ext4_da_release_space(struct inode *inode, int to_free)
 {
        struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
-       int total, mdb, mdb_free, release;
+       struct ext4_inode_info *ei = EXT4_I(inode);
 
        if (!to_free)
                return;         /* Nothing to release, exit */
 
        spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
 
-       if (!EXT4_I(inode)->i_reserved_data_blocks) {
+       if (unlikely(to_free > ei->i_reserved_data_blocks)) {
                /*
-                * if there is no reserved blocks, but we try to free some
-                * then the counter is messed up somewhere.
-                * but since this function is called from invalidate
-                * page, it's harmless to return without any action
+                * if there aren't enough reserved blocks, then the
+                * counter is messed up somewhere.  Since this
+                * function is called from invalidate page, it's
+                * harmless to return without any action.
                 */
-               printk(KERN_INFO "ext4 delalloc try to release %d reserved "
-                           "blocks for inode %lu, but there is no reserved "
-                           "data blocks\n", to_free, inode->i_ino);
-               spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
-               return;
+               ext4_msg(inode->i_sb, KERN_NOTICE, "ext4_da_release_space: "
+                        "ino %lu, to_free %d with only %d reserved "
+                        "data blocks\n", inode->i_ino, to_free,
+                        ei->i_reserved_data_blocks);
+               WARN_ON(1);
+               to_free = ei->i_reserved_data_blocks;
        }
+       ei->i_reserved_data_blocks -= to_free;
 
-       /* recalculate the number of metablocks still need to be reserved */
-       total = EXT4_I(inode)->i_reserved_data_blocks - to_free;
-       mdb = ext4_calc_metadata_amount(inode, total);
-
-       /* figure out how many metablocks to release */
-       BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
-       mdb_free = EXT4_I(inode)->i_reserved_meta_blocks - mdb;
-
-       release = to_free + mdb_free;
-
-       /* update fs dirty blocks counter for truncate case */
-       percpu_counter_sub(&sbi->s_dirtyblocks_counter, release);
+       if (ei->i_reserved_data_blocks == 0) {
+               /*
+                * We can release all of the reserved metadata blocks
+                * only when we have written all of the delayed
+                * allocation blocks.
+                */
+               to_free += ei->i_allocated_meta_blocks;
+               ei->i_allocated_meta_blocks = 0;
+       }
 
-       /* update per-inode reservations */
-       BUG_ON(to_free > EXT4_I(inode)->i_reserved_data_blocks);
-       EXT4_I(inode)->i_reserved_data_blocks -= to_free;
+       /* update fs dirty blocks counter */
+       percpu_counter_sub(&sbi->s_dirtyblocks_counter, to_free);
 
-       BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
-       EXT4_I(inode)->i_reserved_meta_blocks = mdb;
        spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
 
-       vfs_dq_release_reservation_block(inode, release);
+       vfs_dq_release_reservation_block(inode, to_free);
 }
 
 static void ext4_da_page_release_reservation(struct page *page,