Eric Ren
2017-Jan-16 06:42 UTC
[Ocfs2-devel] [PATCH v2 0/2] fix deadlock caused by recursive cluster locking
This is a formal patch set v2 to solve the deadlock issue on which I previously started a RFC (draft patch), and the discussion happened here: [https://oss.oracle.com/pipermail/ocfs2-devel/2016-October/012455.html] Compared to the previous draft patch, this one is much simple and neat. It neither messes up the dlmglue core, nor has a performance penalty on the whole cluster locking system. Instead, it is only used in places where such recursive cluster locking may happen. Changes since v1: 1. Let ocfs2_is_locked_by_me() just return true/false to indicate if the process gets the cluster lock - suggested by: Joseph Qi <jiangqi903 at gmail.com> and Junxiao Bi <junxiao.bi at oracle.com>. 2. Change "struct ocfs2_holder" to a more meaningful name "ocfs2_lock_holder", suggested by: Junxiao Bi <junxiao.bi at oracle.com>. 3. Add debugging output at ocfs2_setattr() and ocfs2_permission() to catch exceptional cases, suggested by: Junxiao Bi <junxiao.bi at oracle.com>. 4. Do not inline functions whose bodies are not in scope, changed by: Stephen Rothwell <sfr at canb.auug.org.au>. Your comments and feedbacks are always welcomed. Eric Ren (2): ocfs2/dlmglue: prepare tracking logic to avoid recursive cluster lock ocfs2: fix deadlock issue when taking inode lock at vfs entry points fs/ocfs2/acl.c | 39 ++++++++++++++++++++++++---- fs/ocfs2/dlmglue.c | 48 +++++++++++++++++++++++++++++++--- fs/ocfs2/dlmglue.h | 18 +++++++++++++ fs/ocfs2/file.c | 76 +++++++++++++++++++++++++++++++++++++++++++++++------- fs/ocfs2/ocfs2.h | 1 + 5 files changed, 164 insertions(+), 18 deletions(-) -- 2.10.2
Eric Ren
2017-Jan-16 06:42 UTC
[Ocfs2-devel] [PATCH v2 1/2] ocfs2/dlmglue: prepare tracking logic to avoid recursive cluster lock
We are in the situation that we have to avoid recursive cluster locking,
but there is no way to check if a cluster lock has been taken by a
precess already.
Mostly, we can avoid recursive locking by writing code carefully.
However, we found that it's very hard to handle the routines that
are invoked directly by vfs code. For instance:
const struct inode_operations ocfs2_file_iops = {
.permission = ocfs2_permission,
.get_acl = ocfs2_iop_get_acl,
.set_acl = ocfs2_iop_set_acl,
};
Both ocfs2_permission() and ocfs2_iop_get_acl() call ocfs2_inode_lock(PR):
do_sys_open
may_open
inode_permission
ocfs2_permission
ocfs2_inode_lock() <=== first time
generic_permission
get_acl
ocfs2_iop_get_acl
ocfs2_inode_lock() <=== recursive one
A deadlock will occur if a remote EX request comes in between two
of ocfs2_inode_lock(). Briefly describe how the deadlock is formed:
On one hand, OCFS2_LOCK_BLOCKED flag of this lockres is set in
BAST(ocfs2_generic_handle_bast) when downconvert is started
on behalf of the remote EX lock request. Another hand, the recursive
cluster lock (the second one) will be blocked in in __ocfs2_cluster_lock()
because of OCFS2_LOCK_BLOCKED. But, the downconvert never complete, why?
because there is no chance for the first cluster lock on this node to be
unlocked - we block ourselves in the code path.
The idea to fix this issue is mostly taken from gfs2 code.
1. introduce a new field: struct ocfs2_lock_res.l_holders, to
keep track of the processes' pid who has taken the cluster lock
of this lock resource;
2. introduce a new flag for ocfs2_inode_lock_full: OCFS2_META_LOCK_GETBH;
it means just getting back disk inode bh for us if we've got cluster lock.
3. export a helper: ocfs2_is_locked_by_me() is used to check if we
have got the cluster lock in the upper code path.
The tracking logic should be used by some of the ocfs2 vfs's callbacks,
to solve the recursive locking issue cuased by the fact that vfs routines
can call into each other.
The performance penalty of processing the holder list should only be seen
at a few cases where the tracking logic is used, such as get/set acl.
You may ask what if the first time we got a PR lock, and the second time
we want a EX lock? fortunately, this case never happens in the real world,
as far as I can see, including permission check, (get|set)_(acl|attr), and
the gfs2 code also do so.
Changes since v1:
1. Let ocfs2_is_locked_by_me() just return true/false to indicate if the
process gets the cluster lock - suggested by: Joseph Qi <jiangqi903 at
gmail.com>
and Junxiao Bi <junxiao.bi at oracle.com>.
2. Change "struct ocfs2_holder" to a more meaningful name
"ocfs2_lock_holder",
suggested by: Junxiao Bi <junxiao.bi at oracle.com>.
3. Do not inline functions whose bodies are not in scope, changed by:
Stephen Rothwell <sfr at canb.auug.org.au>.
[sfr at canb.auug.org.au remove some inlines]
Signed-off-by: Eric Ren <zren at suse.com>
---
fs/ocfs2/dlmglue.c | 48 +++++++++++++++++++++++++++++++++++++++++++++---
fs/ocfs2/dlmglue.h | 18 ++++++++++++++++++
fs/ocfs2/ocfs2.h | 1 +
3 files changed, 64 insertions(+), 3 deletions(-)
diff --git a/fs/ocfs2/dlmglue.c b/fs/ocfs2/dlmglue.c
index 77d1632..b045f02 100644
--- a/fs/ocfs2/dlmglue.c
+++ b/fs/ocfs2/dlmglue.c
@@ -532,6 +532,7 @@ void ocfs2_lock_res_init_once(struct ocfs2_lock_res *res)
init_waitqueue_head(&res->l_event);
INIT_LIST_HEAD(&res->l_blocked_list);
INIT_LIST_HEAD(&res->l_mask_waiters);
+ INIT_LIST_HEAD(&res->l_holders);
}
void ocfs2_inode_lock_res_init(struct ocfs2_lock_res *res,
@@ -749,6 +750,46 @@ void ocfs2_lock_res_free(struct ocfs2_lock_res *res)
res->l_flags = 0UL;
}
+void ocfs2_add_holder(struct ocfs2_lock_res *lockres,
+ struct ocfs2_lock_holder *oh)
+{
+ INIT_LIST_HEAD(&oh->oh_list);
+ oh->oh_owner_pid = get_pid(task_pid(current));
+
+ spin_lock(&lockres->l_lock);
+ list_add_tail(&oh->oh_list, &lockres->l_holders);
+ spin_unlock(&lockres->l_lock);
+}
+
+void ocfs2_remove_holder(struct ocfs2_lock_res *lockres,
+ struct ocfs2_lock_holder *oh)
+{
+ spin_lock(&lockres->l_lock);
+ list_del(&oh->oh_list);
+ spin_unlock(&lockres->l_lock);
+
+ put_pid(oh->oh_owner_pid);
+}
+
+int ocfs2_is_locked_by_me(struct ocfs2_lock_res *lockres)
+{
+ struct ocfs2_lock_holder *oh;
+ struct pid *pid;
+
+ /* look in the list of holders for one with the current task as owner */
+ spin_lock(&lockres->l_lock);
+ pid = task_pid(current);
+ list_for_each_entry(oh, &lockres->l_holders, oh_list) {
+ if (oh->oh_owner_pid == pid) {
+ spin_unlock(&lockres->l_lock);
+ return 1;
+ }
+ }
+ spin_unlock(&lockres->l_lock);
+
+ return 0;
+}
+
static inline void ocfs2_inc_holders(struct ocfs2_lock_res *lockres,
int level)
{
@@ -2333,8 +2374,9 @@ int ocfs2_inode_lock_full_nested(struct inode *inode,
goto getbh;
}
- if (ocfs2_mount_local(osb))
- goto local;
+ if ((arg_flags & OCFS2_META_LOCK_GETBH) ||
+ ocfs2_mount_local(osb))
+ goto update;
if (!(arg_flags & OCFS2_META_LOCK_RECOVERY))
ocfs2_wait_for_recovery(osb);
@@ -2363,7 +2405,7 @@ int ocfs2_inode_lock_full_nested(struct inode *inode,
if (!(arg_flags & OCFS2_META_LOCK_RECOVERY))
ocfs2_wait_for_recovery(osb);
-local:
+update:
/*
* We only see this flag if we're being called from
* ocfs2_read_locked_inode(). It means we're locking an inode
diff --git a/fs/ocfs2/dlmglue.h b/fs/ocfs2/dlmglue.h
index d293a22..19d709d 100644
--- a/fs/ocfs2/dlmglue.h
+++ b/fs/ocfs2/dlmglue.h
@@ -70,6 +70,11 @@ struct ocfs2_orphan_scan_lvb {
__be32 lvb_os_seqno;
};
+struct ocfs2_lock_holder {
+ struct list_head oh_list;
+ struct pid *oh_owner_pid;
+};
+
/* ocfs2_inode_lock_full() 'arg_flags' flags */
/* don't wait on recovery. */
#define OCFS2_META_LOCK_RECOVERY (0x01)
@@ -77,6 +82,8 @@ struct ocfs2_orphan_scan_lvb {
#define OCFS2_META_LOCK_NOQUEUE (0x02)
/* don't block waiting for the downconvert thread, instead return -EAGAIN
*/
#define OCFS2_LOCK_NONBLOCK (0x04)
+/* just get back disk inode bh if we've got cluster lock. */
+#define OCFS2_META_LOCK_GETBH (0x08)
/* Locking subclasses of inode cluster lock */
enum {
@@ -170,4 +177,15 @@ void ocfs2_put_dlm_debug(struct ocfs2_dlm_debug
*dlm_debug);
/* To set the locking protocol on module initialization */
void ocfs2_set_locking_protocol(void);
+
+/*
+ * Keep a list of processes who have interest in a lockres.
+ * Note: this is now only uesed for check recursive cluster lock.
+ */
+void ocfs2_add_holder(struct ocfs2_lock_res *lockres,
+ struct ocfs2_lock_holder *oh);
+void ocfs2_remove_holder(struct ocfs2_lock_res *lockres,
+ struct ocfs2_lock_holder *oh);
+int ocfs2_is_locked_by_me(struct ocfs2_lock_res *lockres);
+
#endif /* DLMGLUE_H */
diff --git a/fs/ocfs2/ocfs2.h b/fs/ocfs2/ocfs2.h
index 7e5958b..0c39d71 100644
--- a/fs/ocfs2/ocfs2.h
+++ b/fs/ocfs2/ocfs2.h
@@ -172,6 +172,7 @@ struct ocfs2_lock_res {
struct list_head l_blocked_list;
struct list_head l_mask_waiters;
+ struct list_head l_holders;
unsigned long l_flags;
char l_name[OCFS2_LOCK_ID_MAX_LEN];
--
2.10.2
Eric Ren
2017-Jan-16 06:42 UTC
[Ocfs2-devel] [PATCH v2 2/2] ocfs2: fix deadlock issue when taking inode lock at vfs entry points
Commit 743b5f1434f5 ("ocfs2: take inode lock in
ocfs2_iop_set/get_acl()")
results in a deadlock, as the author "Tariq Saeed" realized shortly
after the patch was merged. The discussion happened here
(https://oss.oracle.com/pipermail/ocfs2-devel/2015-September/011085.html).
The reason why taking cluster inode lock at vfs entry points opens up
a self deadlock window, is explained in the previous patch of this
series.
So far, we have seen two different code paths that have this issue.
1. do_sys_open
may_open
inode_permission
ocfs2_permission
ocfs2_inode_lock() <=== take PR
generic_permission
get_acl
ocfs2_iop_get_acl
ocfs2_inode_lock() <=== take PR
2. fchmod|fchmodat
chmod_common
notify_change
ocfs2_setattr <=== take EX
posix_acl_chmod
get_acl
ocfs2_iop_get_acl <=== take PR
ocfs2_iop_set_acl <=== take EX
Fixes them by adding the tracking logic (in the previous patch) for
these funcs above, ocfs2_permission(), ocfs2_iop_[set|get]_acl(),
ocfs2_setattr().
Changes since v1:
1. Let ocfs2_is_locked_by_me() just return true/false to indicate if the
process gets the cluster lock - suggested by: Joseph Qi <jiangqi903 at
gmail.com>
and Junxiao Bi <junxiao.bi at oracle.com>.
2. Change "struct ocfs2_holder" to a more meaningful name
"ocfs2_lock_holder",
suggested by: Junxiao Bi <junxiao.bi at oracle.com>.
3. Add debugging output at ocfs2_setattr() and ocfs2_permission() to
catch exceptional cases, suggested by: Junxiao Bi <junxiao.bi at
oracle.com>.
Signed-off-by: Eric Ren <zren at suse.com>
---
fs/ocfs2/acl.c | 39 +++++++++++++++++++++++++----
fs/ocfs2/file.c | 76 +++++++++++++++++++++++++++++++++++++++++++++++++--------
2 files changed, 100 insertions(+), 15 deletions(-)
diff --git a/fs/ocfs2/acl.c b/fs/ocfs2/acl.c
index bed1fcb..3e47262 100644
--- a/fs/ocfs2/acl.c
+++ b/fs/ocfs2/acl.c
@@ -284,16 +284,31 @@ int ocfs2_iop_set_acl(struct inode *inode, struct
posix_acl *acl, int type)
{
struct buffer_head *bh = NULL;
int status = 0;
-
- status = ocfs2_inode_lock(inode, &bh, 1);
+ int arg_flags = 0, has_locked;
+ struct ocfs2_lock_holder oh;
+ struct ocfs2_lock_res *lockres;
+
+ lockres = &OCFS2_I(inode)->ip_inode_lockres;
+ has_locked = ocfs2_is_locked_by_me(lockres);
+ if (has_locked)
+ arg_flags = OCFS2_META_LOCK_GETBH;
+ status = ocfs2_inode_lock_full(inode, &bh, 1, arg_flags);
if (status < 0) {
if (status != -ENOENT)
mlog_errno(status);
return status;
}
+ if (!has_locked)
+ ocfs2_add_holder(lockres, &oh);
+
status = ocfs2_set_acl(NULL, inode, bh, type, acl, NULL, NULL);
- ocfs2_inode_unlock(inode, 1);
+
+ if (!has_locked) {
+ ocfs2_remove_holder(lockres, &oh);
+ ocfs2_inode_unlock(inode, 1);
+ }
brelse(bh);
+
return status;
}
@@ -303,21 +318,35 @@ struct posix_acl *ocfs2_iop_get_acl(struct inode *inode,
int type)
struct buffer_head *di_bh = NULL;
struct posix_acl *acl;
int ret;
+ int arg_flags = 0, has_locked;
+ struct ocfs2_lock_holder oh;
+ struct ocfs2_lock_res *lockres;
osb = OCFS2_SB(inode->i_sb);
if (!(osb->s_mount_opt & OCFS2_MOUNT_POSIX_ACL))
return NULL;
- ret = ocfs2_inode_lock(inode, &di_bh, 0);
+
+ lockres = &OCFS2_I(inode)->ip_inode_lockres;
+ has_locked = ocfs2_is_locked_by_me(lockres);
+ if (has_locked)
+ arg_flags = OCFS2_META_LOCK_GETBH;
+ ret = ocfs2_inode_lock_full(inode, &di_bh, 0, arg_flags);
if (ret < 0) {
if (ret != -ENOENT)
mlog_errno(ret);
return ERR_PTR(ret);
}
+ if (!has_locked)
+ ocfs2_add_holder(lockres, &oh);
acl = ocfs2_get_acl_nolock(inode, type, di_bh);
- ocfs2_inode_unlock(inode, 0);
+ if (!has_locked) {
+ ocfs2_remove_holder(lockres, &oh);
+ ocfs2_inode_unlock(inode, 0);
+ }
brelse(di_bh);
+
return acl;
}
diff --git a/fs/ocfs2/file.c b/fs/ocfs2/file.c
index c488965..b620c25 100644
--- a/fs/ocfs2/file.c
+++ b/fs/ocfs2/file.c
@@ -1138,6 +1138,9 @@ int ocfs2_setattr(struct dentry *dentry, struct iattr
*attr)
handle_t *handle = NULL;
struct dquot *transfer_to[MAXQUOTAS] = { };
int qtype;
+ int arg_flags = 0, had_lock;
+ struct ocfs2_lock_holder oh;
+ struct ocfs2_lock_res *lockres;
trace_ocfs2_setattr(inode, dentry,
(unsigned long long)OCFS2_I(inode)->ip_blkno,
@@ -1173,13 +1176,41 @@ int ocfs2_setattr(struct dentry *dentry, struct iattr
*attr)
}
}
- status = ocfs2_inode_lock(inode, &bh, 1);
+ lockres = &OCFS2_I(inode)->ip_inode_lockres;
+ had_lock = ocfs2_is_locked_by_me(lockres);
+ if (had_lock) {
+ arg_flags = OCFS2_META_LOCK_GETBH;
+
+ /*
+ * As far as we know, ocfs2_setattr() could only be the first
+ * VFS entry point in the call chain of recursive cluster
+ * locking issue.
+ *
+ * For instance:
+ * chmod_common()
+ * notify_change()
+ * ocfs2_setattr()
+ * posix_acl_chmod()
+ * ocfs2_iop_get_acl()
+ *
+ * But, we're not 100% sure if it's always true, because the
+ * ordering of the VFS entry points in the call chain is out
+ * of our control. So, we'd better dump the stack here to
+ * catch the other cases of recursive locking.
+ */
+ mlog(ML_ERROR, "Another case of recursive locking:\n");
+ dump_stack();
+ }
+ status = ocfs2_inode_lock_full(inode, &bh, 1, arg_flags);
if (status < 0) {
if (status != -ENOENT)
mlog_errno(status);
goto bail_unlock_rw;
}
- inode_locked = 1;
+ if (!had_lock) {
+ ocfs2_add_holder(lockres, &oh);
+ inode_locked = 1;
+ }
if (size_change) {
status = inode_newsize_ok(inode, attr->ia_size);
@@ -1260,7 +1291,8 @@ int ocfs2_setattr(struct dentry *dentry, struct iattr
*attr)
bail_commit:
ocfs2_commit_trans(osb, handle);
bail_unlock:
- if (status) {
+ if (status && inode_locked) {
+ ocfs2_remove_holder(lockres, &oh);
ocfs2_inode_unlock(inode, 1);
inode_locked = 0;
}
@@ -1278,8 +1310,10 @@ int ocfs2_setattr(struct dentry *dentry, struct iattr
*attr)
if (status < 0)
mlog_errno(status);
}
- if (inode_locked)
+ if (inode_locked) {
+ ocfs2_remove_holder(lockres, &oh);
ocfs2_inode_unlock(inode, 1);
+ }
brelse(bh);
return status;
@@ -1321,20 +1355,42 @@ int ocfs2_getattr(struct vfsmount *mnt,
int ocfs2_permission(struct inode *inode, int mask)
{
int ret;
+ int has_locked;
+ struct ocfs2_lock_holder oh;
+ struct ocfs2_lock_res *lockres;
if (mask & MAY_NOT_BLOCK)
return -ECHILD;
- ret = ocfs2_inode_lock(inode, NULL, 0);
- if (ret) {
- if (ret != -ENOENT)
- mlog_errno(ret);
- goto out;
+ lockres = &OCFS2_I(inode)->ip_inode_lockres;
+ has_locked = ocfs2_is_locked_by_me(lockres);
+ if (!has_locked) {
+ ret = ocfs2_inode_lock(inode, NULL, 0);
+ if (ret) {
+ if (ret != -ENOENT)
+ mlog_errno(ret);
+ goto out;
+ }
+ ocfs2_add_holder(lockres, &oh);
+ } else {
+ /* See comments in ocfs2_setattr() for details.
+ * The call chain of this case could be:
+ * do_sys_open()
+ * may_open()
+ * inode_permission()
+ * ocfs2_permission()
+ * ocfs2_iop_get_acl()
+ */
+ mlog(ML_ERROR, "Another case of recursive locking:\n");
+ dump_stack();
}
ret = generic_permission(inode, mask);
- ocfs2_inode_unlock(inode, 0);
+ if (!has_locked) {
+ ocfs2_remove_holder(lockres, &oh);
+ ocfs2_inode_unlock(inode, 0);
+ }
out:
return ret;
}
--
2.10.2