GFS2 fatal: filesystem consistency error on rename
authorBob Peterson <rpeterso@redhat.com>
Thu, 30 Sep 2010 14:34:00 +0000 (10:34 -0400)
committerSteven Whitehouse <swhiteho@redhat.com>
Thu, 30 Sep 2010 16:23:03 +0000 (17:23 +0100)
This patch fixes a GFS2 problem whereby the first rename after a
mount can result in a file system consistency error being flagged
improperly and cause the file system to withdraw.  The problem is
that the rename code tries to run the rgrp list with function
gfs2_blk2rgrpd before the rgrp list is guaranteed to be read in
from disk.  The patch makes the rename function hold the rindex
glock (as the gfs2_unlink code does today) which reads in the rgrp
list if need be.  There were a total of three places in the rename
code that improperly referenced the rgrp list without the rindex
glock and this patch fixes all three.

Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
fs/gfs2/ops_inode.c
fs/gfs2/rgrp.c
fs/gfs2/rgrp.h

index fba00171d915252c9f9433d26a29d24059a241fb..0534510200d5961887d3ee957d799e3b4669e76b 100644 (file)
@@ -736,7 +736,7 @@ static int gfs2_rename(struct inode *odir, struct dentry *odentry,
        struct gfs2_inode *ip = GFS2_I(odentry->d_inode);
        struct gfs2_inode *nip = NULL;
        struct gfs2_sbd *sdp = GFS2_SB(odir);
-       struct gfs2_holder ghs[5], r_gh = { .gh_gl = NULL, };
+       struct gfs2_holder ghs[5], r_gh = { .gh_gl = NULL, }, ri_gh;
        struct gfs2_rgrpd *nrgd;
        unsigned int num_gh;
        int dir_rename = 0;
@@ -750,6 +750,9 @@ static int gfs2_rename(struct inode *odir, struct dentry *odentry,
                        return 0;
        }
 
+       error = gfs2_rindex_hold(sdp, &ri_gh);
+       if (error)
+               return error;
 
        if (odip != ndip) {
                error = gfs2_glock_nq_init(sdp->sd_rename_gl, LM_ST_EXCLUSIVE,
@@ -879,7 +882,7 @@ static int gfs2_rename(struct inode *odir, struct dentry *odentry,
 
                al->al_requested = sdp->sd_max_dirres;
 
-               error = gfs2_inplace_reserve(ndip);
+               error = gfs2_inplace_reserve_ri(ndip);
                if (error)
                        goto out_gunlock_q;
 
@@ -961,6 +964,7 @@ out_gunlock_r:
        if (r_gh.gh_gl)
                gfs2_glock_dq_uninit(&r_gh);
 out:
+       gfs2_glock_dq_uninit(&ri_gh);
        return error;
 }
 
index f9ddcf40175313810b5ee215dde5641f55f3abc2..fb67f593f40856b213f03887642c44a3e4c7ccc5 100644 (file)
@@ -1200,7 +1200,8 @@ out:
  * Returns: errno
  */
 
-int gfs2_inplace_reserve_i(struct gfs2_inode *ip, char *file, unsigned int line)
+int gfs2_inplace_reserve_i(struct gfs2_inode *ip, int hold_rindex,
+                          char *file, unsigned int line)
 {
        struct gfs2_sbd *sdp = GFS2_SB(&ip->i_inode);
        struct gfs2_alloc *al = ip->i_alloc;
@@ -1211,12 +1212,15 @@ int gfs2_inplace_reserve_i(struct gfs2_inode *ip, char *file, unsigned int line)
                return -EINVAL;
 
 try_again:
-       /* We need to hold the rindex unless the inode we're using is
-          the rindex itself, in which case it's already held. */
-       if (ip != GFS2_I(sdp->sd_rindex))
-               error = gfs2_rindex_hold(sdp, &al->al_ri_gh);
-       else if (!sdp->sd_rgrps) /* We may not have the rindex read in, so: */
-               error = gfs2_ri_update_special(ip);
+       if (hold_rindex) {
+               /* We need to hold the rindex unless the inode we're using is
+                  the rindex itself, in which case it's already held. */
+               if (ip != GFS2_I(sdp->sd_rindex))
+                       error = gfs2_rindex_hold(sdp, &al->al_ri_gh);
+               else if (!sdp->sd_rgrps) /* We may not have the rindex read
+                                           in, so: */
+                       error = gfs2_ri_update_special(ip);
+       }
 
        if (error)
                return error;
@@ -1227,7 +1231,7 @@ try_again:
           try to free it, and try the allocation again. */
        error = get_local_rgrp(ip, &unlinked, &last_unlinked);
        if (error) {
-               if (ip != GFS2_I(sdp->sd_rindex))
+               if (hold_rindex && ip != GFS2_I(sdp->sd_rindex))
                        gfs2_glock_dq_uninit(&al->al_ri_gh);
                if (error != -EAGAIN)
                        return error;
@@ -1269,7 +1273,7 @@ void gfs2_inplace_release(struct gfs2_inode *ip)
        al->al_rgd = NULL;
        if (al->al_rgd_gh.gh_gl)
                gfs2_glock_dq_uninit(&al->al_rgd_gh);
-       if (ip != GFS2_I(sdp->sd_rindex))
+       if (ip != GFS2_I(sdp->sd_rindex) && al->al_ri_gh.gh_gl)
                gfs2_glock_dq_uninit(&al->al_ri_gh);
 }
 
index f07119d89557855fc9a8673b32e9119cad921570..0e35c0466f9a6c5979a3fe8c339def323bc37fad 100644 (file)
@@ -39,10 +39,12 @@ static inline void gfs2_alloc_put(struct gfs2_inode *ip)
        ip->i_alloc = NULL;
 }
 
-extern int gfs2_inplace_reserve_i(struct gfs2_inode *ip, char *file,
-                                 unsigned int line);
+extern int gfs2_inplace_reserve_i(struct gfs2_inode *ip, int hold_rindex,
+                                 char *file, unsigned int line);
 #define gfs2_inplace_reserve(ip) \
-gfs2_inplace_reserve_i((ip), __FILE__, __LINE__)
+       gfs2_inplace_reserve_i((ip), 1, __FILE__, __LINE__)
+#define gfs2_inplace_reserve_ri(ip) \
+       gfs2_inplace_reserve_i((ip), 0, __FILE__, __LINE__)
 
 extern void gfs2_inplace_release(struct gfs2_inode *ip);