cluster: RHEL6 - fsck.gfs2: Don't allocate leaf blocks in pass1

Bob Peterson rpeterso at fedoraproject.org
Tue May 21 12:24:24 UTC 2013


Gitweb:        http://git.fedorahosted.org/git/?p=cluster.git;a=commitdiff;h=bf9a851204b747f850a3a95a907cd8101bc2b726
Commit:        bf9a851204b747f850a3a95a907cd8101bc2b726
Parent:        100e0e1961e7f26c72ae59b57f5c86873053eb2f
Author:        Bob Peterson <rpeterso at redhat.com>
AuthorDate:    Wed Apr 17 14:09:30 2013 -0700
Committer:     Bob Peterson <rpeterso at redhat.com>
CommitterDate: Fri May 17 15:33:01 2013 -0500

fsck.gfs2: Don't allocate leaf blocks in pass1

Before this patch, if leaf blocks were found to be corrupt, pass1
tried to fix them by allocating new leaf blocks in place of the bad
ones. That's a bad idea, because pass1 populates the blockmap and
sets the bitmap accordingly. In other words, it's dynamically changing.
Say, for example, that you're checking a directory a dinode 0x1234, and
it has a corrupt hash table, and needs new leaf blocks inserted.
Now suppose you have a second directory that occurs later in the bitmap,
say at block 0x2345, and it references leaf block 0x2346, but for some
reason that block (0x2346) is improperly set to "free" in the bitmap.
If pass1 goes out looking for a free block in order to allocate a new
leaf for 0x1234, it will naturally find block 0x2346, because it's
marked free. It writes a new leaf at that block and adds a new
reference in the hash table of 0x1234. Later, when pass1 processes
directory 0x2345, it discovers the reference to 0x2346. Not only has
it wiped out the perfectly good leaf block, it has also created a
duplicate block reference that it needs to sort out in pass1b, which
will likely keep the replaced reference and throw the good one we
had. Thus, we introduced corruption into the file system when we
should have kept the only good reference to 0x2346 and fixed the
bitmap.

The solution provided by this patch is to simply zero out the bad
hash table entries when pass1 comes across them. Later, when pass2
discovers the zero leaf blocks, it can safely allocate new blocks
(since pass1 synced the bitmap according to the blockmap) for the new
leaf blocks and replace the zeros with valid block references.

rhbz#902920
---
 gfs2/fsck/metawalk.c |   31 ++++++++++++++++++++++++++++++-
 gfs2/fsck/metawalk.h |    2 +-
 gfs2/fsck/pass1.c    |    9 ++-------
 gfs2/fsck/pass2.c    |    2 +-
 4 files changed, 34 insertions(+), 10 deletions(-)

diff --git a/gfs2/fsck/metawalk.c b/gfs2/fsck/metawalk.c
index 37b2d4b..78047ec 100644
--- a/gfs2/fsck/metawalk.c
+++ b/gfs2/fsck/metawalk.c
@@ -1954,7 +1954,7 @@ int write_new_leaf(struct gfs2_inode *dip, int start_lindex, int num_copies,
  * leaf a bit, but it's better than deleting the whole directory,
  * which is what used to happen before. */
 int repair_leaf(struct gfs2_inode *ip, uint64_t *leaf_no, int lindex,
-		int ref_count, const char *msg)
+		int ref_count, const char *msg, int allow_alloc)
 {
 	int new_leaf_blks = 0, error, refs;
 	uint64_t bn = 0;
@@ -1969,6 +1969,35 @@ int repair_leaf(struct gfs2_inode *ip, uint64_t *leaf_no, int lindex,
 		log_err( _("Bad leaf left in place.\n"));
 		goto out;
 	}
+	if (!allow_alloc) {
+		uint64_t *cpyptr;
+		char *padbuf;
+		int pad_size, i;
+
+		padbuf = malloc(ref_count * sizeof(uint64_t));
+		cpyptr = (uint64_t *)padbuf;
+		for (i = 0; i < ref_count; i++) {
+			*cpyptr = 0;
+			cpyptr++;
+		}
+		pad_size = ref_count * sizeof(uint64_t);
+		log_err(_("Writing zeros to the hash table of directory %lld "
+			  "(0x%llx) at index: 0x%x for 0x%x pointers.\n"),
+			(unsigned long long)ip->i_di.di_num.no_addr,
+			(unsigned long long)ip->i_di.di_num.no_addr,
+			lindex, ref_count);
+		if (ip->i_sbd->gfs1)
+			gfs1_writei(ip, padbuf, lindex * sizeof(uint64_t),
+				    pad_size);
+		else
+			gfs2_writei(ip, padbuf, lindex * sizeof(uint64_t),
+				    pad_size);
+		free(padbuf);
+		log_err( _("Directory Inode %llu (0x%llx) patched.\n"),
+			 (unsigned long long)ip->i_di.di_num.no_addr,
+			 (unsigned long long)ip->i_di.di_num.no_addr);
+		goto out;
+	}
 	/* We can only write leafs in quantities that are factors of
 	   two, since leaves are doubled, not added sequentially.
 	   So if we have a hole that's not a factor of 2, we have to
diff --git a/gfs2/fsck/metawalk.h b/gfs2/fsck/metawalk.h
index aacb962..a5a51c2 100644
--- a/gfs2/fsck/metawalk.h
+++ b/gfs2/fsck/metawalk.h
@@ -61,7 +61,7 @@ extern int write_new_leaf(struct gfs2_inode *dip, int start_lindex,
 			  int num_copies, const char *before_or_after,
 			  uint64_t *bn);
 extern int repair_leaf(struct gfs2_inode *ip, uint64_t *leaf_no, int lindex,
-		       int ref_count, const char *msg);
+		       int ref_count, const char *msg, int allow_alloc);
 
 #define is_duplicate(dblock) ((dupfind(dblock)) ? 1 : 0)
 
diff --git a/gfs2/fsck/pass1.c b/gfs2/fsck/pass1.c
index 2355de2..7208335 100644
--- a/gfs2/fsck/pass1.c
+++ b/gfs2/fsck/pass1.c
@@ -83,13 +83,8 @@ static int pass1_repair_leaf(struct gfs2_inode *ip, uint64_t *leaf_no,
 			     int lindex, int ref_count, const char *msg,
 			     void *private)
 {
-	struct block_count *bc = (struct block_count *)private;
-	int new_leaf_blks;
-
-	new_leaf_blks = repair_leaf(ip, leaf_no, lindex, ref_count, msg);
-	bc->indir_count += new_leaf_blks;
-
-	return new_leaf_blks;
+	repair_leaf(ip, leaf_no, lindex, ref_count, msg, 0);
+	return 0;
 }
 
 struct metawalk_fxns pass1_fxns = {
diff --git a/gfs2/fsck/pass2.c b/gfs2/fsck/pass2.c
index 02fa577..4b3a5c0 100644
--- a/gfs2/fsck/pass2.c
+++ b/gfs2/fsck/pass2.c
@@ -1038,7 +1038,7 @@ static int pass2_repair_leaf(struct gfs2_inode *ip, uint64_t *leaf_no,
 			     int lindex, int ref_count, const char *msg,
 			     void *private)
 {
-	return repair_leaf(ip, leaf_no, lindex, ref_count, msg);
+	return repair_leaf(ip, leaf_no, lindex, ref_count, msg, 1);
 }
 
 /* The purpose of leafck_fxns is to provide a means for function fix_hashtable


More information about the cluster-commits mailing list