Ashish Samant
2018-Apr-11 19:31 UTC
[Ocfs2-devel] [PATCH V2] ocfs2: Take inode cluster lock before moving reflinked inode from orphan dir
While reflinking an inode, we create a new inode in orphan directory, then take EX lock on it, reflink the original inode to orphan inode and release EX lock. Once the lock is released another node could request it in EX mode from ocfs2_recover_orphans() which causes downconvert of the lock, on this node, to NL mode. Later we attempt to initialize security acl for the orphan inode and move it to the reflink destination. However, while doing this we dont take EX lock on the inode. This could potentially cause problems because we could be starting transaction, accessing journal and modifying metadata of the inode while holding NL lock and with another node holding EX lock on the inode. Fix this by taking orphan inode cluster lock in EX mode before initializing security and moving orphan inode to reflink destination. Use the __tracker variant while taking inode lock to avoid recursive locking in the ocfs2_init_security_and_acl() call chain. Signed-off-by: Ashish Samant <ashish.samant at oracle.com> V1->V2: Modify commit message to better reflect the problem in upstream kernel. --- fs/ocfs2/refcounttree.c | 14 ++++++++++++-- 1 file changed, 12 insertions(+), 2 deletions(-) diff --git a/fs/ocfs2/refcounttree.c b/fs/ocfs2/refcounttree.c index ab156e3..1b1283f 100644 --- a/fs/ocfs2/refcounttree.c +++ b/fs/ocfs2/refcounttree.c @@ -4250,10 +4250,11 @@ static int __ocfs2_reflink(struct dentry *old_dentry, static int ocfs2_reflink(struct dentry *old_dentry, struct inode *dir, struct dentry *new_dentry, bool preserve) { - int error; + int error, had_lock; struct inode *inode = d_inode(old_dentry); struct buffer_head *old_bh = NULL; struct inode *new_orphan_inode = NULL; + struct ocfs2_lock_holder oh; if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) return -EOPNOTSUPP; @@ -4295,6 +4296,14 @@ static int ocfs2_reflink(struct dentry *old_dentry, struct inode *dir, goto out; } + had_lock = ocfs2_inode_lock_tracker(new_orphan_inode, NULL, 1, + &oh); + if (had_lock < 0) { + error = had_lock; + mlog_errno(error); + goto out; + } + /* If the security isn't preserved, we need to re-initialize them. */ if (!preserve) { error = ocfs2_init_security_and_acl(dir, new_orphan_inode, @@ -4302,14 +4311,15 @@ static int ocfs2_reflink(struct dentry *old_dentry, struct inode *dir, if (error) mlog_errno(error); } -out: if (!error) { error = ocfs2_mv_orphaned_inode_to_new(dir, new_orphan_inode, new_dentry); if (error) mlog_errno(error); } + ocfs2_inode_unlock_tracker(new_orphan_inode, 1, &oh, had_lock); +out: if (new_orphan_inode) { /* * We need to open_unlock the inode no matter whether we -- 1.9.1
Junxiao Bi
2018-Apr-12 01:17 UTC
[Ocfs2-devel] [PATCH V2] ocfs2: Take inode cluster lock before moving reflinked inode from orphan dir
On 04/12/2018 03:31 AM, Ashish Samant wrote:> While reflinking an inode, we create a new inode in orphan directory, then > take EX lock on it, reflink the original inode to orphan inode and release > EX lock. Once the lock is released another node could request it in EX mode > from ocfs2_recover_orphans() which causes downconvert of the lock, on this > node, to NL mode. > > Later we attempt to initialize security acl for the orphan inode and move > it to the reflink destination. However, while doing this we dont take EX > lock on the inode. This could potentially cause problems because we could > be starting transaction, accessing journal and modifying metadata of the > inode while holding NL lock and with another node holding EX lock on the > inode. > > Fix this by taking orphan inode cluster lock in EX mode before > initializing security and moving orphan inode to reflink destination. > Use the __tracker variant while taking inode lock to avoid recursive > locking in the ocfs2_init_security_and_acl() call chain. > > Signed-off-by: Ashish Samant <ashish.samant at oracle.com>Reviewed-by: Junxiao Bi <junxiao.bi at oracle.com>> > V1->V2: > Modify commit message to better reflect the problem in upstream kernel. > --- > fs/ocfs2/refcounttree.c | 14 ++++++++++++-- > 1 file changed, 12 insertions(+), 2 deletions(-) > > diff --git a/fs/ocfs2/refcounttree.c b/fs/ocfs2/refcounttree.c > index ab156e3..1b1283f 100644 > --- a/fs/ocfs2/refcounttree.c > +++ b/fs/ocfs2/refcounttree.c > @@ -4250,10 +4250,11 @@ static int __ocfs2_reflink(struct dentry *old_dentry, > static int ocfs2_reflink(struct dentry *old_dentry, struct inode *dir, > struct dentry *new_dentry, bool preserve) > { > - int error; > + int error, had_lock; > struct inode *inode = d_inode(old_dentry); > struct buffer_head *old_bh = NULL; > struct inode *new_orphan_inode = NULL; > + struct ocfs2_lock_holder oh; > > if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) > return -EOPNOTSUPP; > @@ -4295,6 +4296,14 @@ static int ocfs2_reflink(struct dentry *old_dentry, struct inode *dir, > goto out; > } > > + had_lock = ocfs2_inode_lock_tracker(new_orphan_inode, NULL, 1, > + &oh); > + if (had_lock < 0) { > + error = had_lock; > + mlog_errno(error); > + goto out; > + } > + > /* If the security isn't preserved, we need to re-initialize them. */ > if (!preserve) { > error = ocfs2_init_security_and_acl(dir, new_orphan_inode, > @@ -4302,14 +4311,15 @@ static int ocfs2_reflink(struct dentry *old_dentry, struct inode *dir, > if (error) > mlog_errno(error); > } > -out: > if (!error) { > error = ocfs2_mv_orphaned_inode_to_new(dir, new_orphan_inode, > new_dentry); > if (error) > mlog_errno(error); > } > + ocfs2_inode_unlock_tracker(new_orphan_inode, 1, &oh, had_lock); > > +out: > if (new_orphan_inode) { > /* > * We need to open_unlock the inode no matter whether we
piaojun
2018-Apr-12 01:37 UTC
[Ocfs2-devel] [PATCH V2] ocfs2: Take inode cluster lock before moving reflinked inode from orphan dir
On 2018/4/12 3:31, Ashish Samant wrote:> While reflinking an inode, we create a new inode in orphan directory, then > take EX lock on it, reflink the original inode to orphan inode and release > EX lock. Once the lock is released another node could request it in EX mode > from ocfs2_recover_orphans() which causes downconvert of the lock, on this > node, to NL mode. > > Later we attempt to initialize security acl for the orphan inode and move > it to the reflink destination. However, while doing this we dont take EX > lock on the inode. This could potentially cause problems because we could > be starting transaction, accessing journal and modifying metadata of the > inode while holding NL lock and with another node holding EX lock on the > inode. > > Fix this by taking orphan inode cluster lock in EX mode before > initializing security and moving orphan inode to reflink destination. > Use the __tracker variant while taking inode lock to avoid recursive > locking in the ocfs2_init_security_and_acl() call chain. > > Signed-off-by: Ashish Samant <ashish.samant at oracle.com>Acked-by: Jun Piao <piaojun at huawei.com>> > V1->V2: > Modify commit message to better reflect the problem in upstream kernel. > --- > fs/ocfs2/refcounttree.c | 14 ++++++++++++-- > 1 file changed, 12 insertions(+), 2 deletions(-) > > diff --git a/fs/ocfs2/refcounttree.c b/fs/ocfs2/refcounttree.c > index ab156e3..1b1283f 100644 > --- a/fs/ocfs2/refcounttree.c > +++ b/fs/ocfs2/refcounttree.c > @@ -4250,10 +4250,11 @@ static int __ocfs2_reflink(struct dentry *old_dentry, > static int ocfs2_reflink(struct dentry *old_dentry, struct inode *dir, > struct dentry *new_dentry, bool preserve) > { > - int error; > + int error, had_lock; > struct inode *inode = d_inode(old_dentry); > struct buffer_head *old_bh = NULL; > struct inode *new_orphan_inode = NULL; > + struct ocfs2_lock_holder oh; > > if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) > return -EOPNOTSUPP; > @@ -4295,6 +4296,14 @@ static int ocfs2_reflink(struct dentry *old_dentry, struct inode *dir, > goto out; > } > > + had_lock = ocfs2_inode_lock_tracker(new_orphan_inode, NULL, 1, > + &oh); > + if (had_lock < 0) { > + error = had_lock; > + mlog_errno(error); > + goto out; > + } > + > /* If the security isn't preserved, we need to re-initialize them. */ > if (!preserve) { > error = ocfs2_init_security_and_acl(dir, new_orphan_inode, > @@ -4302,14 +4311,15 @@ static int ocfs2_reflink(struct dentry *old_dentry, struct inode *dir, > if (error) > mlog_errno(error); > } > -out: > if (!error) { > error = ocfs2_mv_orphaned_inode_to_new(dir, new_orphan_inode, > new_dentry); > if (error) > mlog_errno(error); > } > + ocfs2_inode_unlock_tracker(new_orphan_inode, 1, &oh, had_lock); > > +out: > if (new_orphan_inode) { > /* > * We need to open_unlock the inode no matter whether we >
Joseph Qi
2018-Apr-12 05:38 UTC
[Ocfs2-devel] [PATCH V2] ocfs2: Take inode cluster lock before moving reflinked inode from orphan dir
On 18/4/12 03:31, Ashish Samant wrote:> While reflinking an inode, we create a new inode in orphan directory, then > take EX lock on it, reflink the original inode to orphan inode and release > EX lock. Once the lock is released another node could request it in EX mode > from ocfs2_recover_orphans() which causes downconvert of the lock, on this > node, to NL mode. > > Later we attempt to initialize security acl for the orphan inode and move > it to the reflink destination. However, while doing this we dont take EX > lock on the inode. This could potentially cause problems because we could > be starting transaction, accessing journal and modifying metadata of the > inode while holding NL lock and with another node holding EX lock on the > inode. > > Fix this by taking orphan inode cluster lock in EX mode before > initializing security and moving orphan inode to reflink destination. > Use the __tracker variant while taking inode lock to avoid recursive > locking in the ocfs2_init_security_and_acl() call chain. > > Signed-off-by: Ashish Samant <ashish.samant at oracle.com> >Reviewed-by: Joseph Qi <jiangqi903 at gmail.com>> V1->V2: > Modify commit message to better reflect the problem in upstream kernel. > --- > fs/ocfs2/refcounttree.c | 14 ++++++++++++-- > 1 file changed, 12 insertions(+), 2 deletions(-) > > diff --git a/fs/ocfs2/refcounttree.c b/fs/ocfs2/refcounttree.c > index ab156e3..1b1283f 100644 > --- a/fs/ocfs2/refcounttree.c > +++ b/fs/ocfs2/refcounttree.c > @@ -4250,10 +4250,11 @@ static int __ocfs2_reflink(struct dentry *old_dentry, > static int ocfs2_reflink(struct dentry *old_dentry, struct inode *dir, > struct dentry *new_dentry, bool preserve) > { > - int error; > + int error, had_lock; > struct inode *inode = d_inode(old_dentry); > struct buffer_head *old_bh = NULL; > struct inode *new_orphan_inode = NULL; > + struct ocfs2_lock_holder oh; > > if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) > return -EOPNOTSUPP; > @@ -4295,6 +4296,14 @@ static int ocfs2_reflink(struct dentry *old_dentry, struct inode *dir, > goto out; > } > > + had_lock = ocfs2_inode_lock_tracker(new_orphan_inode, NULL, 1, > + &oh); > + if (had_lock < 0) { > + error = had_lock; > + mlog_errno(error); > + goto out; > + } > + > /* If the security isn't preserved, we need to re-initialize them. */ > if (!preserve) { > error = ocfs2_init_security_and_acl(dir, new_orphan_inode, > @@ -4302,14 +4311,15 @@ static int ocfs2_reflink(struct dentry *old_dentry, struct inode *dir, > if (error) > mlog_errno(error); > } > -out: > if (!error) { > error = ocfs2_mv_orphaned_inode_to_new(dir, new_orphan_inode, > new_dentry); > if (error) > mlog_errno(error); > } > + ocfs2_inode_unlock_tracker(new_orphan_inode, 1, &oh, had_lock); > > +out: > if (new_orphan_inode) { > /* > * We need to open_unlock the inode no matter whether we >
Andrew Morton
2018-May-02 00:08 UTC
[Ocfs2-devel] [PATCH V2] ocfs2: Take inode cluster lock before moving reflinked inode from orphan dir
On Wed, 11 Apr 2018 12:31:47 -0700 Ashish Samant <ashish.samant at oracle.com> wrote:> While reflinking an inode, we create a new inode in orphan directory, then > take EX lock on it, reflink the original inode to orphan inode and release > EX lock. Once the lock is released another node could request it in EX mode > from ocfs2_recover_orphans() which causes downconvert of the lock, on this > node, to NL mode. > > Later we attempt to initialize security acl for the orphan inode and move > it to the reflink destination. However, while doing this we dont take EX > lock on the inode. This could potentially cause problems because we could > be starting transaction, accessing journal and modifying metadata of the > inode while holding NL lock and with another node holding EX lock on the > inode.Someone (eg, me) needs to decide which kernel version(s) need the patch. And all I have is "this could potentially cause problems". Can you please be more specific about the end-user impact? *Does* it cause problems? Is the problem sufficiently urgent to warrant a merge into 4.17? Into -stable kernels? Please always include all the end-user-impact info when fixing bugs. Thanks.> Fix this by taking orphan inode cluster lock in EX mode before > initializing security and moving orphan inode to reflink destination. > Use the __tracker variant while taking inode lock to avoid recursive > locking in the ocfs2_init_security_and_acl() call chain.