ext4: introduce reserved space
authorLukas Czerner <lczerner@redhat.com>
Wed, 10 Apr 2013 02:11:22 +0000 (22:11 -0400)
committerTheodore Ts'o <tytso@mit.edu>
Wed, 10 Apr 2013 02:11:22 +0000 (22:11 -0400)
Currently in ENOSPC condition when writing into unwritten space, or
punching a hole, we might need to split the extent and grow extent tree.
However since we can not allocate any new metadata blocks we'll have to
zero out unwritten part of extent or punched out part of extent, or in
the worst case return ENOSPC even though use actually does not allocate
any space.

Also in delalloc path we do reserve metadata and data blocks for the
time we're going to write out, however metadata block reservation is
very tricky especially since we expect that logical connectivity implies
physical connectivity, however that might not be the case and hence we
might end up allocating more metadata blocks than previously reserved.
So in future, metadata reservation checks should be removed since we can
not assure that we do not under reserve.

And this is where reserved space comes into the picture. When mounting
the file system we slice off a little bit of the file system space (2%
or 4096 clusters, whichever is smaller) which can be then used for the
cases mentioned above to prevent costly zeroout, or unexpected ENOSPC.

The number of reserved clusters can be set via sysfs, however it can
never be bigger than number of free clusters in the file system.

Note that this patch fixes the failure of xfstest 274 as expected.

Signed-off-by: Lukas Czerner <lczerner@redhat.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Reviewed-by: Carlos Maiolino <cmaiolino@redhat.com>
Documentation/filesystems/ext4.txt
fs/ext4/balloc.c
fs/ext4/ext4.h
fs/ext4/extents.c
fs/ext4/inode.c
fs/ext4/super.c

index 5dd957d8b25b426faf59776f50fffd27e0fb0119..f7cbf574a875271296d93bee53b290ca36c44a66 100644 (file)
@@ -494,6 +494,17 @@ Files in /sys/fs/ext4/<devname>
  session_write_kbytes         This file is read-only and shows the number of
                               kilobytes of data that have been written to this
                               filesystem since it was mounted.
+
+ reserved_clusters            This is RW file and contains number of reserved
+                              clusters in the file system which will be used
+                              in the specific situations to avoid costly
+                              zeroout, unexpected ENOSPC, or possible data
+                              loss. The default is 2% or 4096 clusters,
+                              whichever is smaller and this can be changed
+                              however it can never exceed number of clusters
+                              in the file system. If there is not enough space
+                              for the reserved space when mounting the file
+                              mount will _not_ fail.
 ..............................................................................
 
 Ioctls
index 9e8d8ffb063f8d7b7346391bcc9fe7c552b44a35..8dcaea69e37fdbf0de0acb9f92c251166851afbb 100644 (file)
@@ -499,20 +499,22 @@ ext4_read_block_bitmap(struct super_block *sb, ext4_group_t block_group)
 static int ext4_has_free_clusters(struct ext4_sb_info *sbi,
                                  s64 nclusters, unsigned int flags)
 {
-       s64 free_clusters, dirty_clusters, root_clusters;
+       s64 free_clusters, dirty_clusters, rsv, resv_clusters;
        struct percpu_counter *fcc = &sbi->s_freeclusters_counter;
        struct percpu_counter *dcc = &sbi->s_dirtyclusters_counter;
 
        free_clusters  = percpu_counter_read_positive(fcc);
        dirty_clusters = percpu_counter_read_positive(dcc);
+       resv_clusters = atomic64_read(&sbi->s_resv_clusters);
 
        /*
         * r_blocks_count should always be multiple of the cluster ratio so
         * we are safe to do a plane bit shift only.
         */
-       root_clusters = ext4_r_blocks_count(sbi->s_es) >> sbi->s_cluster_bits;
+       rsv = (ext4_r_blocks_count(sbi->s_es) >> sbi->s_cluster_bits) +
+             resv_clusters;
 
-       if (free_clusters - (nclusters + root_clusters + dirty_clusters) <
+       if (free_clusters - (nclusters + rsv + dirty_clusters) <
                                        EXT4_FREECLUSTERS_WATERMARK) {
                free_clusters  = percpu_counter_sum_positive(fcc);
                dirty_clusters = percpu_counter_sum_positive(dcc);
@@ -520,15 +522,21 @@ static int ext4_has_free_clusters(struct ext4_sb_info *sbi,
        /* Check whether we have space after accounting for current
         * dirty clusters & root reserved clusters.
         */
-       if (free_clusters >= ((root_clusters + nclusters) + dirty_clusters))
+       if (free_clusters >= (rsv + nclusters + dirty_clusters))
                return 1;
 
        /* Hm, nope.  Are (enough) root reserved clusters available? */
        if (uid_eq(sbi->s_resuid, current_fsuid()) ||
            (!gid_eq(sbi->s_resgid, GLOBAL_ROOT_GID) && in_group_p(sbi->s_resgid)) ||
            capable(CAP_SYS_RESOURCE) ||
-               (flags & EXT4_MB_USE_ROOT_BLOCKS)) {
+           (flags & EXT4_MB_USE_ROOT_BLOCKS)) {
 
+               if (free_clusters >= (nclusters + dirty_clusters +
+                                     resv_clusters))
+                       return 1;
+       }
+       /* No free blocks. Let's see if we can dip into reserved pool */
+       if (flags & EXT4_MB_USE_RESERVED) {
                if (free_clusters >= (nclusters + dirty_clusters))
                        return 1;
        }
index d91871570982bc1f9659a63c29c7b085c960fe6c..12b560435aba243d5a72a180298d41c33159e8e3 100644 (file)
@@ -121,6 +121,8 @@ typedef unsigned int ext4_group_t;
 #define EXT4_MB_STREAM_ALLOC           0x0800
 /* Use reserved root blocks if needed */
 #define EXT4_MB_USE_ROOT_BLOCKS                0x1000
+/* Use blocks from reserved pool */
+#define EXT4_MB_USE_RESERVED           0x2000
 
 struct ext4_allocation_request {
        /* target inode for block we're allocating */
@@ -557,9 +559,8 @@ enum {
 #define EXT4_GET_BLOCKS_UNINIT_EXT             0x0002
 #define EXT4_GET_BLOCKS_CREATE_UNINIT_EXT      (EXT4_GET_BLOCKS_UNINIT_EXT|\
                                                 EXT4_GET_BLOCKS_CREATE)
-       /* Caller is from the delayed allocation writeout path,
-          so set the magic i_delalloc_reserve_flag after taking the
-          inode allocation semaphore for */
+       /* Caller is from the delayed allocation writeout path
+        * finally doing the actual allocation of delayed blocks */
 #define EXT4_GET_BLOCKS_DELALLOC_RESERVE       0x0004
        /* caller is from the direct IO path, request to creation of an
        unitialized extents if not allocated, split the uninitialized
@@ -571,8 +572,9 @@ enum {
        /* Convert extent to initialized after IO complete */
 #define EXT4_GET_BLOCKS_IO_CONVERT_EXT         (EXT4_GET_BLOCKS_CONVERT|\
                                         EXT4_GET_BLOCKS_CREATE_UNINIT_EXT)
-       /* Punch out blocks of an extent */
-#define EXT4_GET_BLOCKS_PUNCH_OUT_EXT          0x0020
+       /* Eventual metadata allocation (due to growing extent tree)
+        * should not fail, so try to use reserved blocks for that.*/
+#define EXT4_GET_BLOCKS_METADATA_NOFAIL                0x0020
        /* Don't normalize allocation size (used for fallocate) */
 #define EXT4_GET_BLOCKS_NO_NORMALIZE           0x0040
        /* Request will not result in inode size update (user for fallocate) */
@@ -1188,6 +1190,7 @@ struct ext4_sb_info {
        unsigned int s_mount_flags;
        unsigned int s_def_mount_opt;
        ext4_fsblk_t s_sb_block;
+       atomic64_t s_resv_clusters;
        kuid_t s_resuid;
        kgid_t s_resgid;
        unsigned short s_mount_state;
index ea607f9072323e486df70ccb4c2c6597bff09aed..8b158ae2443bcd95509cc0b717d9f526e28aac87 100644 (file)
@@ -1942,8 +1942,8 @@ prepend:
         * There is no free space in the found leaf.
         * We're gonna add a new leaf in the tree.
         */
-       if (flag & EXT4_GET_BLOCKS_PUNCH_OUT_EXT)
-               flags = EXT4_MB_USE_ROOT_BLOCKS;
+       if (flag & EXT4_GET_BLOCKS_METADATA_NOFAIL)
+               flags = EXT4_MB_USE_RESERVED;
        err = ext4_ext_create_new_leaf(handle, inode, flags, path, newext);
        if (err)
                goto cleanup;
@@ -2729,12 +2729,14 @@ again:
 
                        /*
                         * Split the extent in two so that 'end' is the last
-                        * block in the first new extent
+                        * block in the first new extent. Also we should not
+                        * fail removing space due to ENOSPC so try to use
+                        * reserved block if that happens.
                         */
                        err = ext4_split_extent_at(handle, inode, path,
-                                               end + 1, split_flag,
-                                               EXT4_GET_BLOCKS_PRE_IO |
-                                               EXT4_GET_BLOCKS_PUNCH_OUT_EXT);
+                                       end + 1, split_flag,
+                                       EXT4_GET_BLOCKS_PRE_IO |
+                                       EXT4_GET_BLOCKS_METADATA_NOFAIL);
 
                        if (err < 0)
                                goto out;
@@ -3209,7 +3211,8 @@ out:
 static int ext4_ext_convert_to_initialized(handle_t *handle,
                                           struct inode *inode,
                                           struct ext4_map_blocks *map,
-                                          struct ext4_ext_path *path)
+                                          struct ext4_ext_path *path,
+                                          int flags)
 {
        struct ext4_sb_info *sbi;
        struct ext4_extent_header *eh;
@@ -3435,7 +3438,7 @@ static int ext4_ext_convert_to_initialized(handle_t *handle,
        }
 
        allocated = ext4_split_extent(handle, inode, path,
-                                     &split_map, split_flag, 0);
+                                     &split_map, split_flag, flags);
        if (allocated < 0)
                err = allocated;
 
@@ -3755,6 +3758,12 @@ ext4_ext_handle_uninitialized_extents(handle_t *handle, struct inode *inode,
                  flags, allocated);
        ext4_ext_show_leaf(inode, path);
 
+       /*
+        * When writing into uninitialized space, we should not fail to
+        * allocate metadata blocks for the new extent block if needed.
+        */
+       flags |= EXT4_GET_BLOCKS_METADATA_NOFAIL;
+
        trace_ext4_ext_handle_uninitialized_extents(inode, map, flags,
                                                    allocated, newblock);
 
@@ -3818,7 +3827,7 @@ ext4_ext_handle_uninitialized_extents(handle_t *handle, struct inode *inode,
        }
 
        /* buffered write, writepage time, convert*/
-       ret = ext4_ext_convert_to_initialized(handle, inode, map, path);
+       ret = ext4_ext_convert_to_initialized(handle, inode, map, path, flags);
        if (ret >= 0)
                ext4_update_inode_fsync_trans(handle, inode, 1);
 out:
index f9b0b479ff4cce8024c98ef78ac13550c001b382..629d67b62dfba56cc9009c6a54c56866f6149581 100644 (file)
@@ -1688,12 +1688,21 @@ static void mpage_da_map_and_submit(struct mpage_da_data *mpd)
         */
        map.m_lblk = next;
        map.m_len = max_blocks;
-       get_blocks_flags = EXT4_GET_BLOCKS_CREATE;
+       /*
+        * We're in delalloc path and it is possible that we're going to
+        * need more metadata blocks than previously reserved. However
+        * we must not fail because we're in writeback and there is
+        * nothing we can do about it so it might result in data loss.
+        * So use reserved blocks to allocate metadata if possible.
+        */
+       get_blocks_flags = EXT4_GET_BLOCKS_CREATE |
+                          EXT4_GET_BLOCKS_METADATA_NOFAIL;
        if (ext4_should_dioread_nolock(mpd->inode))
                get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
        if (mpd->b_state & (1 << BH_Delay))
                get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
 
+
        blks = ext4_map_blocks(handle, mpd->inode, &map, get_blocks_flags);
        if (blks < 0) {
                struct super_block *sb = mpd->inode->i_sb;
index 968ca93691751781931e4a1959eb51db06fb0c02..6fea87db7daa4696e1781a2817fbd6ff61d14673 100644 (file)
@@ -81,6 +81,7 @@ static int ext4_feature_set_ok(struct super_block *sb, int readonly);
 static void ext4_destroy_lazyinit_thread(void);
 static void ext4_unregister_li_request(struct super_block *sb);
 static void ext4_clear_request_list(void);
+static int ext4_reserve_clusters(struct ext4_sb_info *, ext4_fsblk_t);
 
 #if !defined(CONFIG_EXT2_FS) && !defined(CONFIG_EXT2_FS_MODULE) && defined(CONFIG_EXT4_USE_FOR_EXT23)
 static struct file_system_type ext2_fs_type = {
@@ -2382,6 +2383,17 @@ struct ext4_attr {
        int offset;
 };
 
+static int parse_strtoull(const char *buf,
+               unsigned long long max, unsigned long long *value)
+{
+       int ret;
+
+       ret = kstrtoull(skip_spaces(buf), 0, value);
+       if (!ret && *value > max)
+               ret = -EINVAL;
+       return ret;
+}
+
 static int parse_strtoul(const char *buf,
                unsigned long max, unsigned long *value)
 {
@@ -2466,6 +2478,27 @@ static ssize_t sbi_ui_store(struct ext4_attr *a,
        return count;
 }
 
+static ssize_t reserved_clusters_show(struct ext4_attr *a,
+                                 struct ext4_sb_info *sbi, char *buf)
+{
+       return snprintf(buf, PAGE_SIZE, "%llu\n",
+               (unsigned long long) atomic64_read(&sbi->s_resv_clusters));
+}
+
+static ssize_t reserved_clusters_store(struct ext4_attr *a,
+                                  struct ext4_sb_info *sbi,
+                                  const char *buf, size_t count)
+{
+       unsigned long long val;
+       int ret;
+
+       if (parse_strtoull(buf, -1ULL, &val))
+               return -EINVAL;
+       ret = ext4_reserve_clusters(sbi, val);
+
+       return ret ? ret : count;
+}
+
 static ssize_t trigger_test_error(struct ext4_attr *a,
                                  struct ext4_sb_info *sbi,
                                  const char *buf, size_t count)
@@ -2503,6 +2536,7 @@ static struct ext4_attr ext4_attr_##name = __ATTR(name, mode, show, store)
 EXT4_RO_ATTR(delayed_allocation_blocks);
 EXT4_RO_ATTR(session_write_kbytes);
 EXT4_RO_ATTR(lifetime_write_kbytes);
+EXT4_RW_ATTR(reserved_clusters);
 EXT4_ATTR_OFFSET(inode_readahead_blks, 0644, sbi_ui_show,
                 inode_readahead_blks_store, s_inode_readahead_blks);
 EXT4_RW_ATTR_SBI_UI(inode_goal, s_inode_goal);
@@ -2520,6 +2554,7 @@ static struct attribute *ext4_attrs[] = {
        ATTR_LIST(delayed_allocation_blocks),
        ATTR_LIST(session_write_kbytes),
        ATTR_LIST(lifetime_write_kbytes),
+       ATTR_LIST(reserved_clusters),
        ATTR_LIST(inode_readahead_blks),
        ATTR_LIST(inode_goal),
        ATTR_LIST(mb_stats),
@@ -3195,6 +3230,40 @@ int ext4_calculate_overhead(struct super_block *sb)
        return 0;
 }
 
+
+static ext4_fsblk_t ext4_calculate_resv_clusters(struct ext4_sb_info *sbi)
+{
+       ext4_fsblk_t resv_clusters;
+
+       /*
+        * By default we reserve 2% or 4096 clusters, whichever is smaller.
+        * This should cover the situations where we can not afford to run
+        * out of space like for example punch hole, or converting
+        * uninitialized extents in delalloc path. In most cases such
+        * allocation would require 1, or 2 blocks, higher numbers are
+        * very rare.
+        */
+       resv_clusters = ext4_blocks_count(sbi->s_es) >> sbi->s_cluster_bits;
+
+       do_div(resv_clusters, 50);
+       resv_clusters = min_t(ext4_fsblk_t, resv_clusters, 4096);
+
+       return resv_clusters;
+}
+
+
+static int ext4_reserve_clusters(struct ext4_sb_info *sbi, ext4_fsblk_t count)
+{
+       ext4_fsblk_t clusters = ext4_blocks_count(sbi->s_es) >>
+                               sbi->s_cluster_bits;
+
+       if (count >= clusters)
+               return -EINVAL;
+
+       atomic64_set(&sbi->s_resv_clusters, count);
+       return 0;
+}
+
 static int ext4_fill_super(struct super_block *sb, void *data, int silent)
 {
        char *orig_data = kstrdup(data, GFP_KERNEL);
@@ -3918,6 +3987,13 @@ no_journal:
                         "available");
        }
 
+       err = ext4_reserve_clusters(sbi, ext4_calculate_resv_clusters(sbi));
+       if (err) {
+               ext4_msg(sb, KERN_ERR, "failed to reserve %llu clusters for "
+                        "reserved pool", ext4_calculate_resv_clusters(sbi));
+               goto failed_mount4a;
+       }
+
        err = ext4_setup_system_zone(sb);
        if (err) {
                ext4_msg(sb, KERN_ERR, "failed to initialize system "
@@ -4750,9 +4826,10 @@ static int ext4_statfs(struct dentry *dentry, struct kstatfs *buf)
        struct super_block *sb = dentry->d_sb;
        struct ext4_sb_info *sbi = EXT4_SB(sb);
        struct ext4_super_block *es = sbi->s_es;
-       ext4_fsblk_t overhead = 0;
+       ext4_fsblk_t overhead = 0, resv_blocks;
        u64 fsid;
        s64 bfree;
+       resv_blocks = EXT4_C2B(sbi, atomic64_read(&sbi->s_resv_clusters));
 
        if (!test_opt(sb, MINIX_DF))
                overhead = sbi->s_overhead;
@@ -4764,8 +4841,9 @@ static int ext4_statfs(struct dentry *dentry, struct kstatfs *buf)
                percpu_counter_sum_positive(&sbi->s_dirtyclusters_counter);
        /* prevent underflow in case that few free space is available */
        buf->f_bfree = EXT4_C2B(sbi, max_t(s64, bfree, 0));
-       buf->f_bavail = buf->f_bfree - ext4_r_blocks_count(es);
-       if (buf->f_bfree < ext4_r_blocks_count(es))
+       buf->f_bavail = buf->f_bfree -
+                       (ext4_r_blocks_count(es) + resv_blocks);
+       if (buf->f_bfree < (ext4_r_blocks_count(es) + resv_blocks))
                buf->f_bavail = 0;
        buf->f_files = le32_to_cpu(es->s_inodes_count);
        buf->f_ffree = percpu_counter_sum_positive(&sbi->s_freeinodes_counter);