ext4: fix potential race between s_flex_groups online resizing and access
authorSuraj Jitindar Singh <surajjs@amazon.com>
Wed, 19 Feb 2020 03:08:51 +0000 (19:08 -0800)
committerTheodore Ts'o <tytso@mit.edu>
Sat, 22 Feb 2020 00:31:46 +0000 (19:31 -0500)
During an online resize an array of s_flex_groups structures gets replaced
so it can get enlarged. If there is a concurrent access to the array and
this memory has been reused then this can lead to an invalid memory access.

The s_flex_group array has been converted into an array of pointers rather
than an array of structures. This is to ensure that the information
contained in the structures cannot get out of sync during a resize due to
an accessor updating the value in the old structure after it has been
copied but before the array pointer is updated. Since the structures them-
selves are no longer copied but only the pointers to them this case is
mitigated.

Link: https://bugzilla.kernel.org/show_bug.cgi?id=206443
Link: https://lore.kernel.org/r/20200221053458.730016-4-tytso@mit.edu
Signed-off-by: Suraj Jitindar Singh <surajjs@amazon.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Cc: stable@kernel.org
fs/ext4/ext4.h
fs/ext4/ialloc.c
fs/ext4/mballoc.c
fs/ext4/resize.c
fs/ext4/super.c

index b1ece5329738a413a901ff946260a80b309e55e4..614fefa7dc7ab22a980b9d34d72dad3d6500c2c5 100644 (file)
@@ -1512,7 +1512,7 @@ struct ext4_sb_info {
        unsigned int s_extent_max_zeroout_kb;
 
        unsigned int s_log_groups_per_flex;
-       struct flex_groups *s_flex_groups;
+       struct flex_groups * __rcu *s_flex_groups;
        ext4_group_t s_flex_groups_allocated;
 
        /* workqueue for reserved extent conversions (buffered io) */
index c66e8f9451a266669bc70a40b40b3ac1af5849d6..f95ee99091e4c55a827d4cdad01bcc89ef8b3e89 100644 (file)
@@ -328,11 +328,13 @@ void ext4_free_inode(handle_t *handle, struct inode *inode)
 
        percpu_counter_inc(&sbi->s_freeinodes_counter);
        if (sbi->s_log_groups_per_flex) {
-               ext4_group_t f = ext4_flex_group(sbi, block_group);
+               struct flex_groups *fg;
 
-               atomic_inc(&sbi->s_flex_groups[f].free_inodes);
+               fg = sbi_array_rcu_deref(sbi, s_flex_groups,
+                                        ext4_flex_group(sbi, block_group));
+               atomic_inc(&fg->free_inodes);
                if (is_directory)
-                       atomic_dec(&sbi->s_flex_groups[f].used_dirs);
+                       atomic_dec(&fg->used_dirs);
        }
        BUFFER_TRACE(bh2, "call ext4_handle_dirty_metadata");
        fatal = ext4_handle_dirty_metadata(handle, NULL, bh2);
@@ -368,12 +370,13 @@ static void get_orlov_stats(struct super_block *sb, ext4_group_t g,
                            int flex_size, struct orlov_stats *stats)
 {
        struct ext4_group_desc *desc;
-       struct flex_groups *flex_group = EXT4_SB(sb)->s_flex_groups;
 
        if (flex_size > 1) {
-               stats->free_inodes = atomic_read(&flex_group[g].free_inodes);
-               stats->free_clusters = atomic64_read(&flex_group[g].free_clusters);
-               stats->used_dirs = atomic_read(&flex_group[g].used_dirs);
+               struct flex_groups *fg = sbi_array_rcu_deref(EXT4_SB(sb),
+                                                            s_flex_groups, g);
+               stats->free_inodes = atomic_read(&fg->free_inodes);
+               stats->free_clusters = atomic64_read(&fg->free_clusters);
+               stats->used_dirs = atomic_read(&fg->used_dirs);
                return;
        }
 
@@ -1054,7 +1057,8 @@ got:
                if (sbi->s_log_groups_per_flex) {
                        ext4_group_t f = ext4_flex_group(sbi, group);
 
-                       atomic_inc(&sbi->s_flex_groups[f].used_dirs);
+                       atomic_inc(&sbi_array_rcu_deref(sbi, s_flex_groups,
+                                                       f)->used_dirs);
                }
        }
        if (ext4_has_group_desc_csum(sb)) {
@@ -1077,7 +1081,8 @@ got:
 
        if (sbi->s_log_groups_per_flex) {
                flex_group = ext4_flex_group(sbi, group);
-               atomic_dec(&sbi->s_flex_groups[flex_group].free_inodes);
+               atomic_dec(&sbi_array_rcu_deref(sbi, s_flex_groups,
+                                               flex_group)->free_inodes);
        }
 
        inode->i_ino = ino + group * EXT4_INODES_PER_GROUP(sb);
index 1b46fb63692abc8351ce2f5eb2c698f1a5f55752..51a78eb65f3cf64744dd82a7ed0735f350a4459f 100644 (file)
@@ -3038,7 +3038,8 @@ ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
                ext4_group_t flex_group = ext4_flex_group(sbi,
                                                          ac->ac_b_ex.fe_group);
                atomic64_sub(ac->ac_b_ex.fe_len,
-                            &sbi->s_flex_groups[flex_group].free_clusters);
+                            &sbi_array_rcu_deref(sbi, s_flex_groups,
+                                                 flex_group)->free_clusters);
        }
 
        err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
@@ -4936,7 +4937,8 @@ do_more:
        if (sbi->s_log_groups_per_flex) {
                ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
                atomic64_add(count_clusters,
-                            &sbi->s_flex_groups[flex_group].free_clusters);
+                            &sbi_array_rcu_deref(sbi, s_flex_groups,
+                                                 flex_group)->free_clusters);
        }
 
        /*
@@ -5093,7 +5095,8 @@ int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
        if (sbi->s_log_groups_per_flex) {
                ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
                atomic64_add(clusters_freed,
-                            &sbi->s_flex_groups[flex_group].free_clusters);
+                            &sbi_array_rcu_deref(sbi, s_flex_groups,
+                                                 flex_group)->free_clusters);
        }
 
        ext4_mb_unload_buddy(&e4b);
index 536cc9f380914316fe7ee0f4a46002190bf58979..a50b51270ea9ad976c2f9da7c2cc97b8d2254eb3 100644 (file)
@@ -1430,11 +1430,14 @@ static void ext4_update_super(struct super_block *sb,
                   percpu_counter_read(&sbi->s_freeclusters_counter));
        if (ext4_has_feature_flex_bg(sb) && sbi->s_log_groups_per_flex) {
                ext4_group_t flex_group;
+               struct flex_groups *fg;
+
                flex_group = ext4_flex_group(sbi, group_data[0].group);
+               fg = sbi_array_rcu_deref(sbi, s_flex_groups, flex_group);
                atomic64_add(EXT4_NUM_B2C(sbi, free_blocks),
-                            &sbi->s_flex_groups[flex_group].free_clusters);
+                            &fg->free_clusters);
                atomic_add(EXT4_INODES_PER_GROUP(sb) * flex_gd->count,
-                          &sbi->s_flex_groups[flex_group].free_inodes);
+                          &fg->free_inodes);
        }
 
        /*
index e00bcc19099fc81822164722bee9da593218e913..6b7e628b7903a4206bb4bf75e8446a6b4ea7510d 100644 (file)
@@ -1015,6 +1015,7 @@ static void ext4_put_super(struct super_block *sb)
        struct ext4_sb_info *sbi = EXT4_SB(sb);
        struct ext4_super_block *es = sbi->s_es;
        struct buffer_head **group_desc;
+       struct flex_groups **flex_groups;
        int aborted = 0;
        int i, err;
 
@@ -1052,8 +1053,13 @@ static void ext4_put_super(struct super_block *sb)
        for (i = 0; i < sbi->s_gdb_count; i++)
                brelse(group_desc[i]);
        kvfree(group_desc);
+       flex_groups = rcu_dereference(sbi->s_flex_groups);
+       if (flex_groups) {
+               for (i = 0; i < sbi->s_flex_groups_allocated; i++)
+                       kvfree(flex_groups[i]);
+               kvfree(flex_groups);
+       }
        rcu_read_unlock();
-       kvfree(sbi->s_flex_groups);
        percpu_counter_destroy(&sbi->s_freeclusters_counter);
        percpu_counter_destroy(&sbi->s_freeinodes_counter);
        percpu_counter_destroy(&sbi->s_dirs_counter);
@@ -2384,8 +2390,8 @@ done:
 int ext4_alloc_flex_bg_array(struct super_block *sb, ext4_group_t ngroup)
 {
        struct ext4_sb_info *sbi = EXT4_SB(sb);
-       struct flex_groups *new_groups;
-       int size;
+       struct flex_groups **old_groups, **new_groups;
+       int size, i;
 
        if (!sbi->s_log_groups_per_flex)
                return 0;
@@ -2394,22 +2400,37 @@ int ext4_alloc_flex_bg_array(struct super_block *sb, ext4_group_t ngroup)
        if (size <= sbi->s_flex_groups_allocated)
                return 0;
 
-       size = roundup_pow_of_two(size * sizeof(struct flex_groups));
-       new_groups = kvzalloc(size, GFP_KERNEL);
+       new_groups = kvzalloc(roundup_pow_of_two(size *
+                             sizeof(*sbi->s_flex_groups)), GFP_KERNEL);
        if (!new_groups) {
-               ext4_msg(sb, KERN_ERR, "not enough memory for %d flex groups",
-                        size / (int) sizeof(struct flex_groups));
+               ext4_msg(sb, KERN_ERR,
+                        "not enough memory for %d flex group pointers", size);
                return -ENOMEM;
        }
-
-       if (sbi->s_flex_groups) {
-               memcpy(new_groups, sbi->s_flex_groups,
-                      (sbi->s_flex_groups_allocated *
-                       sizeof(struct flex_groups)));
-               kvfree(sbi->s_flex_groups);
+       for (i = sbi->s_flex_groups_allocated; i < size; i++) {
+               new_groups[i] = kvzalloc(roundup_pow_of_two(
+                                        sizeof(struct flex_groups)),
+                                        GFP_KERNEL);
+               if (!new_groups[i]) {
+                       for (i--; i >= sbi->s_flex_groups_allocated; i--)
+                               kvfree(new_groups[i]);
+                       kvfree(new_groups);
+                       ext4_msg(sb, KERN_ERR,
+                                "not enough memory for %d flex groups", size);
+                       return -ENOMEM;
+               }
        }
-       sbi->s_flex_groups = new_groups;
-       sbi->s_flex_groups_allocated = size / sizeof(struct flex_groups);
+       rcu_read_lock();
+       old_groups = rcu_dereference(sbi->s_flex_groups);
+       if (old_groups)
+               memcpy(new_groups, old_groups,
+                      (sbi->s_flex_groups_allocated *
+                       sizeof(struct flex_groups *)));
+       rcu_read_unlock();
+       rcu_assign_pointer(sbi->s_flex_groups, new_groups);
+       sbi->s_flex_groups_allocated = size;
+       if (old_groups)
+               ext4_kvfree_array_rcu(old_groups);
        return 0;
 }
 
@@ -2417,6 +2438,7 @@ static int ext4_fill_flex_info(struct super_block *sb)
 {
        struct ext4_sb_info *sbi = EXT4_SB(sb);
        struct ext4_group_desc *gdp = NULL;
+       struct flex_groups *fg;
        ext4_group_t flex_group;
        int i, err;
 
@@ -2434,12 +2456,11 @@ static int ext4_fill_flex_info(struct super_block *sb)
                gdp = ext4_get_group_desc(sb, i, NULL);
 
                flex_group = ext4_flex_group(sbi, i);
-               atomic_add(ext4_free_inodes_count(sb, gdp),
-                          &sbi->s_flex_groups[flex_group].free_inodes);
+               fg = sbi_array_rcu_deref(sbi, s_flex_groups, flex_group);
+               atomic_add(ext4_free_inodes_count(sb, gdp), &fg->free_inodes);
                atomic64_add(ext4_free_group_clusters(sb, gdp),
-                            &sbi->s_flex_groups[flex_group].free_clusters);
-               atomic_add(ext4_used_dirs_count(sb, gdp),
-                          &sbi->s_flex_groups[flex_group].used_dirs);
+                            &fg->free_clusters);
+               atomic_add(ext4_used_dirs_count(sb, gdp), &fg->used_dirs);
        }
 
        return 1;
@@ -3641,6 +3662,7 @@ static int ext4_fill_super(struct super_block *sb, void *data, int silent)
        struct buffer_head *bh, **group_desc;
        struct ext4_super_block *es = NULL;
        struct ext4_sb_info *sbi = kzalloc(sizeof(*sbi), GFP_KERNEL);
+       struct flex_groups **flex_groups;
        ext4_fsblk_t block;
        ext4_fsblk_t sb_block = get_sb_block(&data);
        ext4_fsblk_t logical_sb_block;
@@ -4692,8 +4714,14 @@ failed_mount7:
        ext4_unregister_li_request(sb);
 failed_mount6:
        ext4_mb_release(sb);
-       if (sbi->s_flex_groups)
-               kvfree(sbi->s_flex_groups);
+       rcu_read_lock();
+       flex_groups = rcu_dereference(sbi->s_flex_groups);
+       if (flex_groups) {
+               for (i = 0; i < sbi->s_flex_groups_allocated; i++)
+                       kvfree(flex_groups[i]);
+               kvfree(flex_groups);
+       }
+       rcu_read_unlock();
        percpu_counter_destroy(&sbi->s_freeclusters_counter);
        percpu_counter_destroy(&sbi->s_freeinodes_counter);
        percpu_counter_destroy(&sbi->s_dirs_counter);