Re: [PATCH 23/51] e2fsck: Verify htree root/node checksums

On Mon, Dec 19, 2011 at 07:57:21AM +0100, Andreas Dilger wrote:

On 2011-12-14, at 2:15, "Darrick J. Wong" [off-list ref] wrote:

quoted

Check htree internal node checksums.  If broken, ask user to clear the htree
index and recreate it later.

@@ -560,8 +572,12 @@ static void parse_int_node(ext2_filsys fs,

#endif

   count = ext2fs_le16_to_cpu(limit->count);
-    expect_limit = (fs->blocksize - ((char *) ent - block_buf)) /
-        sizeof(struct ext2_dx_entry);
+    if (EXT2_HAS_RO_COMPAT_FEATURE(fs->super,
+                       EXT4_FEATURE_RO_COMPAT_METADATA_CSUM))
+        csum_size = sizeof(struct ext2_dx_tail);
+    expect_limit = (fs->blocksize -
+            (csum_size + ((char *) ent - block_buf))) /
+               sizeof(struct ext2_dx_entry);

How does this handle the case where the METADATA_CSUM feature is newly
enabled but there is a dirent usinf the end of the leaf block?  It definitely
shouldn't cause the last entry to be considered invalid. 

For htree nodes, tune2fs decrements limit if count < limit.  If count == limit,
then fsck is run to rebuild the htree index.  For its part, e2fsck will notice
that count == limit, zero the htree node, and add the directory inode to the
pass3 rebuild list.

quoted

@@ -829,7 +850,7 @@ static int check_dir_block(ext2_filsys fs,
              (rec_len == fs->blocksize) &&
              (dirent->name_len == 0) &&
              (ext2fs_le16_to_cpu(limit->limit) ==
-                ((fs->blocksize-8) /
+                ((fs->blocksize - (8 + csum_size)) /
                sizeof(struct ext2_dx_entry))))
           dx_db->type = DX_DIRBLOCK_NODE;

Same here.  

quoted

diff --git a/e2fsck/problem.c b/e2fsck/problem.c
index 96b0de5..2e9ab7f 100644
--- a/e2fsck/problem.c
+++ b/e2fsck/problem.c

@@ -1374,6 +1374,16 @@ static struct e2fsck_problem problem_table[] = {
     N_("i_file_acl_hi @F %N, @s zero.\n"),
     PROMPT_CLEAR, PR_PREEN_OK },

+    /* htree root node fails checksum */
+    { PR_2_HTREE_ROOT_CSUM_INVALID,
+      N_("@p @h %d: root node fails checksum\n"),
+      PROMPT_CLEAR_HTREE, PR_PREEN_OK },
+
+    /* htree internal node fails checksum */
+    { PR_2_HTREE_NODE_CSUM_INVALID,
+      N_("@p @h %d: node fails checksum\n"),

This error message should include "internal" in the printed message. 

How about "htree node"?  I wondered why the other PR_*_HTREE messages don't say
"htree" -- is there a particular reason for this?

--D