| Message ID | 1389403153-4220-1-git-send-email-srinivas.eeda@oracle.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Fri, Jan 10, 2014 at 05:19:13PM -0800, Srinivas Eeda wrote: > From: Srinivas Eeda <seeda@srini.(none)> > > A tiny race between BAST and unlock message causes the NULL dereference. > > A node sends an unlock request to master and receives a response. Before > processing the response it receives a BAST from the master. Since both requests > are processed by different threads it creates a race. While the BAST is being > processed, lock can get freed by unlock code. > > This patch makes bast to return immediately if lock is found but unlock is > pending. The code should handle this race. We also have to fix master node to > skip sending BAST after receiving unlock message. Did the master send the BAST after the unlock, or does that race too? Does the master know the unlock has succeeded, or does it just think so? > @@ -385,8 +385,13 @@ int dlm_proxy_ast_handler(struct o2net_msg *msg, u32 len, void *data, > head = &res->granted; > > list_for_each_entry(lock, head, list) { > - if (lock->ml.cookie == cookie) > - goto do_ast; > + /* if lock is found but unlock is pending ignore the bast */ > + if (lock->ml.cookie == cookie) { > + if (lock->unlock_pending) > + break; > + else > + goto do_ast; > + } This breaks out for asts as well as basts. Can't that cause problems with the unlock ast expected by the caller? Joel
On 2014/1/11 9:19, Srinivas Eeda wrote: > From: Srinivas Eeda <seeda@srini.(none)> > > A tiny race between BAST and unlock message causes the NULL dereference. > > A node sends an unlock request to master and receives a response. Before > processing the response it receives a BAST from the master. Since both requests > are processed by different threads it creates a race. While the BAST is being > processed, lock can get freed by unlock code. > > This patch makes bast to return immediately if lock is found but unlock is > pending. The code should handle this race. We also have to fix master node to > skip sending BAST after receiving unlock message. > > Below is the crash stack > > BUG: unable to handle kernel NULL pointer dereference at 0000000000000048 > IP: [<ffffffffa015e023>] o2dlm_blocking_ast_wrapper+0xd/0x16 > [<ffffffffa034e3db>] dlm_do_local_bast+0x8e/0x97 [ocfs2_dlm] > [<ffffffffa034f366>] dlm_proxy_ast_handler+0x838/0x87e [ocfs2_dlm] > [<ffffffffa0308abe>] o2net_process_message+0x395/0x5b8 [ocfs2_nodemanager] > [<ffffffffa030aac8>] o2net_rx_until_empty+0x762/0x90d [ocfs2_nodemanager] > [<ffffffff81071802>] worker_thread+0x14d/0x1ed > > Signed-off-by: Srinivas Eeda <srinivas.eeda@oracle.com> > --- > fs/ocfs2/dlm/dlmast.c | 9 +++++++-- > 1 file changed, 7 insertions(+), 2 deletions(-) > > diff --git a/fs/ocfs2/dlm/dlmast.c b/fs/ocfs2/dlm/dlmast.c > index b46278f..dbc6cee 100644 > --- a/fs/ocfs2/dlm/dlmast.c > +++ b/fs/ocfs2/dlm/dlmast.c > @@ -385,8 +385,13 @@ int dlm_proxy_ast_handler(struct o2net_msg *msg, u32 len, void *data, > head = &res->granted; > > list_for_each_entry(lock, head, list) { > - if (lock->ml.cookie == cookie) > - goto do_ast; > + /* if lock is found but unlock is pending ignore the bast */ > + if (lock->ml.cookie == cookie) { > + if (lock->unlock_pending) > + break; > + else > + goto do_ast; > + } > } > > mlog(0, "Got %sast for unknown lock! cookie=%u:%llu, name=%.*s, " > I found you sent a version on Jan 30, 2012. https://oss.oracle.com/pipermail/ocfs2-devel/2012-January/008469.html Compared with the old version, this version only saves a little bit CPU, am I right?
On 01/13/2014 07:37 AM, Joel Becker wrote: > On Fri, Jan 10, 2014 at 05:19:13PM -0800, Srinivas Eeda wrote: >> From: Srinivas Eeda <seeda@srini.(none)> >> >> A tiny race between BAST and unlock message causes the NULL dereference. >> >> A node sends an unlock request to master and receives a response. Before >> processing the response it receives a BAST from the master. Since both requests >> are processed by different threads it creates a race. While the BAST is being >> processed, lock can get freed by unlock code. >> >> This patch makes bast to return immediately if lock is found but unlock is >> pending. The code should handle this race. We also have to fix master node to >> skip sending BAST after receiving unlock message. > Did the master send the BAST after the unlock, or does that race too? > Does the master know the unlock has succeeded, or does it just think so? I think it's due to a race but I haven't debugged the master. My guess is unlock request sneaked in before the dlm_flush_asts was called. However non master node should handle this race as well, so just did that part which fixed a bug we were seeing. > >> @@ -385,8 +385,13 @@ int dlm_proxy_ast_handler(struct o2net_msg *msg, u32 len, void *data, >> head = &res->granted; >> >> list_for_each_entry(lock, head, list) { >> - if (lock->ml.cookie == cookie) >> - goto do_ast; >> + /* if lock is found but unlock is pending ignore the bast */ >> + if (lock->ml.cookie == cookie) { >> + if (lock->unlock_pending) >> + break; >> + else >> + goto do_ast; >> + } > This breaks out for asts as well as basts. Can't that cause problems > with the unlock ast expected by the caller? if unlock_pending is set, then the node is trying to unlock an existing lock and shouldn't receive any asts ? > > Joel > >
On 01/13/2014 08:06 PM, Joseph Qi wrote: > On 2014/1/11 9:19, Srinivas Eeda wrote: >> From: Srinivas Eeda <seeda@srini.(none)> >> >> A tiny race between BAST and unlock message causes the NULL dereference. >> >> A node sends an unlock request to master and receives a response. Before >> processing the response it receives a BAST from the master. Since both requests >> are processed by different threads it creates a race. While the BAST is being >> processed, lock can get freed by unlock code. >> >> This patch makes bast to return immediately if lock is found but unlock is >> pending. The code should handle this race. We also have to fix master node to >> skip sending BAST after receiving unlock message. >> >> Below is the crash stack >> >> BUG: unable to handle kernel NULL pointer dereference at 0000000000000048 >> IP: [<ffffffffa015e023>] o2dlm_blocking_ast_wrapper+0xd/0x16 >> [<ffffffffa034e3db>] dlm_do_local_bast+0x8e/0x97 [ocfs2_dlm] >> [<ffffffffa034f366>] dlm_proxy_ast_handler+0x838/0x87e [ocfs2_dlm] >> [<ffffffffa0308abe>] o2net_process_message+0x395/0x5b8 [ocfs2_nodemanager] >> [<ffffffffa030aac8>] o2net_rx_until_empty+0x762/0x90d [ocfs2_nodemanager] >> [<ffffffff81071802>] worker_thread+0x14d/0x1ed >> >> Signed-off-by: Srinivas Eeda <srinivas.eeda@oracle.com> >> --- >> fs/ocfs2/dlm/dlmast.c | 9 +++++++-- >> 1 file changed, 7 insertions(+), 2 deletions(-) >> >> diff --git a/fs/ocfs2/dlm/dlmast.c b/fs/ocfs2/dlm/dlmast.c >> index b46278f..dbc6cee 100644 >> --- a/fs/ocfs2/dlm/dlmast.c >> +++ b/fs/ocfs2/dlm/dlmast.c >> @@ -385,8 +385,13 @@ int dlm_proxy_ast_handler(struct o2net_msg *msg, u32 len, void *data, >> head = &res->granted; >> >> list_for_each_entry(lock, head, list) { >> - if (lock->ml.cookie == cookie) >> - goto do_ast; >> + /* if lock is found but unlock is pending ignore the bast */ >> + if (lock->ml.cookie == cookie) { >> + if (lock->unlock_pending) >> + break; >> + else >> + goto do_ast; >> + } >> } >> >> mlog(0, "Got %sast for unknown lock! cookie=%u:%llu, name=%.*s, " >> > I found you sent a version on Jan 30, 2012. > https://oss.oracle.com/pipermail/ocfs2-devel/2012-January/008469.html > Compared with the old version, this version only saves a little bit CPU, > am I right? Yes you are right. I made the change as Goldwyn suggested which is a good thing to have :)
diff --git a/fs/ocfs2/dlm/dlmast.c b/fs/ocfs2/dlm/dlmast.c index b46278f..dbc6cee 100644 --- a/fs/ocfs2/dlm/dlmast.c +++ b/fs/ocfs2/dlm/dlmast.c @@ -385,8 +385,13 @@ int dlm_proxy_ast_handler(struct o2net_msg *msg, u32 len, void *data, head = &res->granted; list_for_each_entry(lock, head, list) { - if (lock->ml.cookie == cookie) - goto do_ast; + /* if lock is found but unlock is pending ignore the bast */ + if (lock->ml.cookie == cookie) { + if (lock->unlock_pending) + break; + else + goto do_ast; + } } mlog(0, "Got %sast for unknown lock! cookie=%u:%llu, name=%.*s, "