Message ID | 548f65d4.jhly8vWOTpWuHPjg%akpm@linux-foundation.org (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
On Mon, Dec 15, 2014 at 02:51:00PM -0800, Andrew Morton wrote: > In such a race case, invalid address access may occurs. So we should > delete list res->tracking before resA->refs decrease to 0. > > diff -puN fs/ocfs2/dlm/dlmmaster.c~ocfs2-avoid-access-invalid-address-when-read-o2dlm-debug-messages fs/ocfs2/dlm/dlmmaster.c > --- a/fs/ocfs2/dlm/dlmmaster.c~ocfs2-avoid-access-invalid-address-when-read-o2dlm-debug-messages > +++ a/fs/ocfs2/dlm/dlmmaster.c > @@ -498,16 +498,6 @@ static void dlm_lockres_release(struct k > mlog(0, "destroying lockres %.*s\n", res->lockname.len, > res->lockname.name); > > - spin_lock(&dlm->track_lock); > - if (!list_empty(&res->tracking)) > - list_del_init(&res->tracking); > - else { > - mlog(ML_ERROR, "Resource %.*s not on the Tracking list\n", > - res->lockname.len, res->lockname.name); > - dlm_print_one_lock_resource(res); > - } > - spin_unlock(&dlm->track_lock); > - > atomic_dec(&dlm->res_cur_count); > > if (!hlist_unhashed(&res->hash_node) || > diff -puN fs/ocfs2/dlm/dlmthread.c~ocfs2-avoid-access-invalid-address-when-read-o2dlm-debug-messages fs/ocfs2/dlm/dlmthread.c > --- a/fs/ocfs2/dlm/dlmthread.c~ocfs2-avoid-access-invalid-address-when-read-o2dlm-debug-messages > +++ a/fs/ocfs2/dlm/dlmthread.c > @@ -211,6 +211,16 @@ static void dlm_purge_lockres(struct dlm > > __dlm_unhash_lockres(dlm, res); > > + spin_lock(&dlm->track_lock); > + if (!list_empty(&res->tracking)) > + list_del_init(&res->tracking); > + else { > + mlog(ML_ERROR, "Resource %.*s not on the Tracking list\n", > + res->lockname.len, res->lockname.name); > + __dlm_print_one_lock_resource(res); > + } > + spin_unlock(&dlm->track_lock); > + > /* lockres is not in the hash now. drop the flag and wake up > * any processes waiting in dlm_get_lock_resource. */ > if (!master) { > _ How do we know that dlm_purge_lockres() is the last caller of dlm_lockres_put()? Don't we now have a problem where if the last ref is dropped by any other function than dlm_purge_lockres() the lockres is freed while on the tracking list? --Mark -- Mark Fasheh
? 2014/12/17 6:26, Mark Fasheh ??: > On Mon, Dec 15, 2014 at 02:51:00PM -0800, Andrew Morton wrote: >> In such a race case, invalid address access may occurs. So we should >> delete list res->tracking before resA->refs decrease to 0. >> > > >> diff -puN fs/ocfs2/dlm/dlmmaster.c~ocfs2-avoid-access-invalid-address-when-read-o2dlm-debug-messages fs/ocfs2/dlm/dlmmaster.c >> --- a/fs/ocfs2/dlm/dlmmaster.c~ocfs2-avoid-access-invalid-address-when-read-o2dlm-debug-messages >> +++ a/fs/ocfs2/dlm/dlmmaster.c >> @@ -498,16 +498,6 @@ static void dlm_lockres_release(struct k >> mlog(0, "destroying lockres %.*s\n", res->lockname.len, >> res->lockname.name); >> >> - spin_lock(&dlm->track_lock); >> - if (!list_empty(&res->tracking)) >> - list_del_init(&res->tracking); >> - else { >> - mlog(ML_ERROR, "Resource %.*s not on the Tracking list\n", >> - res->lockname.len, res->lockname.name); >> - dlm_print_one_lock_resource(res); >> - } >> - spin_unlock(&dlm->track_lock); >> - >> atomic_dec(&dlm->res_cur_count); >> >> if (!hlist_unhashed(&res->hash_node) || >> diff -puN fs/ocfs2/dlm/dlmthread.c~ocfs2-avoid-access-invalid-address-when-read-o2dlm-debug-messages fs/ocfs2/dlm/dlmthread.c >> --- a/fs/ocfs2/dlm/dlmthread.c~ocfs2-avoid-access-invalid-address-when-read-o2dlm-debug-messages >> +++ a/fs/ocfs2/dlm/dlmthread.c >> @@ -211,6 +211,16 @@ static void dlm_purge_lockres(struct dlm >> >> __dlm_unhash_lockres(dlm, res); >> >> + spin_lock(&dlm->track_lock); >> + if (!list_empty(&res->tracking)) >> + list_del_init(&res->tracking); >> + else { >> + mlog(ML_ERROR, "Resource %.*s not on the Tracking list\n", >> + res->lockname.len, res->lockname.name); >> + __dlm_print_one_lock_resource(res); >> + } >> + spin_unlock(&dlm->track_lock); >> + >> /* lockres is not in the hash now. drop the flag and wake up >> * any processes waiting in dlm_get_lock_resource. */ >> if (!master) { >> _ > > How do we know that dlm_purge_lockres() is the last caller of > dlm_lockres_put()? Don't we now have a problem where if the last ref is > dropped by any other function than dlm_purge_lockres() the lockres is freed > while on the tracking list? > --Mark > > -- > Mark Fasheh > > . > dlm_purge_lockres is not necessarily the last caller of dlm_lockres_put(), but it means lockres will be purged if dlm_purge_lockres is called. Besides, lockres is also unhashed in dlm_purge_lockres, so lockres can be removed from tracking list. contents of dlm->tracking_list will be consistent with dlm->lockres_hash. --YiWen Jiang
On Fri, Dec 19, 2014 at 04:34:39PM +0800, jiangyiwen wrote: > ? 2014/12/17 6:26, Mark Fasheh ??: > > On Mon, Dec 15, 2014 at 02:51:00PM -0800, Andrew Morton wrote: > >> In such a race case, invalid address access may occurs. So we should > >> delete list res->tracking before resA->refs decrease to 0. > >> > > > > > >> diff -puN fs/ocfs2/dlm/dlmmaster.c~ocfs2-avoid-access-invalid-address-when-read-o2dlm-debug-messages fs/ocfs2/dlm/dlmmaster.c > >> --- a/fs/ocfs2/dlm/dlmmaster.c~ocfs2-avoid-access-invalid-address-when-read-o2dlm-debug-messages > >> +++ a/fs/ocfs2/dlm/dlmmaster.c > >> @@ -498,16 +498,6 @@ static void dlm_lockres_release(struct k > >> mlog(0, "destroying lockres %.*s\n", res->lockname.len, > >> res->lockname.name); > >> > >> - spin_lock(&dlm->track_lock); > >> - if (!list_empty(&res->tracking)) > >> - list_del_init(&res->tracking); > >> - else { > >> - mlog(ML_ERROR, "Resource %.*s not on the Tracking list\n", > >> - res->lockname.len, res->lockname.name); > >> - dlm_print_one_lock_resource(res); > >> - } > >> - spin_unlock(&dlm->track_lock); > >> - > >> atomic_dec(&dlm->res_cur_count); > >> > >> if (!hlist_unhashed(&res->hash_node) || > >> diff -puN fs/ocfs2/dlm/dlmthread.c~ocfs2-avoid-access-invalid-address-when-read-o2dlm-debug-messages fs/ocfs2/dlm/dlmthread.c > >> --- a/fs/ocfs2/dlm/dlmthread.c~ocfs2-avoid-access-invalid-address-when-read-o2dlm-debug-messages > >> +++ a/fs/ocfs2/dlm/dlmthread.c > >> @@ -211,6 +211,16 @@ static void dlm_purge_lockres(struct dlm > >> > >> __dlm_unhash_lockres(dlm, res); > >> > >> + spin_lock(&dlm->track_lock); > >> + if (!list_empty(&res->tracking)) > >> + list_del_init(&res->tracking); > >> + else { > >> + mlog(ML_ERROR, "Resource %.*s not on the Tracking list\n", > >> + res->lockname.len, res->lockname.name); > >> + __dlm_print_one_lock_resource(res); > >> + } > >> + spin_unlock(&dlm->track_lock); > >> + > >> /* lockres is not in the hash now. drop the flag and wake up > >> * any processes waiting in dlm_get_lock_resource. */ > >> if (!master) { > >> _ > > > > How do we know that dlm_purge_lockres() is the last caller of > > dlm_lockres_put()? Don't we now have a problem where if the last ref is > > dropped by any other function than dlm_purge_lockres() the lockres is freed > > while on the tracking list? > > --Mark > > > > -- > > Mark Fasheh > > > > . > > > dlm_purge_lockres is not necessarily the last caller of > dlm_lockres_put(), but it means lockres will be purged if > dlm_purge_lockres is called. Besides, lockres is also unhashed in > dlm_purge_lockres, so lockres can be removed from tracking list. > contents of dlm->tracking_list will be consistent with > dlm->lockres_hash. I'm still confused. This is what I'm worried about: 1) a procss calls dlm_lockres_put(), NOT from dlm_purge_lockres(). 2) if the count goes to zero, then that process will call dlm_lockres_release() 3) dlm_lockres_release() will free the lockres without removing it from the tracking list. Thus we will have a corrupted list. Does that make sense? Am I wrong here? --Mark -- Mark Fasheh
? 2014/12/20 4:25, Mark Fasheh ??: > On Fri, Dec 19, 2014 at 04:34:39PM +0800, jiangyiwen wrote: >> ? 2014/12/17 6:26, Mark Fasheh ??: >>> On Mon, Dec 15, 2014 at 02:51:00PM -0800, Andrew Morton wrote: >>>> In such a race case, invalid address access may occurs. So we should >>>> delete list res->tracking before resA->refs decrease to 0. >>>> >>> >>> >>>> diff -puN fs/ocfs2/dlm/dlmmaster.c~ocfs2-avoid-access-invalid-address-when-read-o2dlm-debug-messages fs/ocfs2/dlm/dlmmaster.c >>>> --- a/fs/ocfs2/dlm/dlmmaster.c~ocfs2-avoid-access-invalid-address-when-read-o2dlm-debug-messages >>>> +++ a/fs/ocfs2/dlm/dlmmaster.c >>>> @@ -498,16 +498,6 @@ static void dlm_lockres_release(struct k >>>> mlog(0, "destroying lockres %.*s\n", res->lockname.len, >>>> res->lockname.name); >>>> >>>> - spin_lock(&dlm->track_lock); >>>> - if (!list_empty(&res->tracking)) >>>> - list_del_init(&res->tracking); >>>> - else { >>>> - mlog(ML_ERROR, "Resource %.*s not on the Tracking list\n", >>>> - res->lockname.len, res->lockname.name); >>>> - dlm_print_one_lock_resource(res); >>>> - } >>>> - spin_unlock(&dlm->track_lock); >>>> - >>>> atomic_dec(&dlm->res_cur_count); >>>> >>>> if (!hlist_unhashed(&res->hash_node) || >>>> diff -puN fs/ocfs2/dlm/dlmthread.c~ocfs2-avoid-access-invalid-address-when-read-o2dlm-debug-messages fs/ocfs2/dlm/dlmthread.c >>>> --- a/fs/ocfs2/dlm/dlmthread.c~ocfs2-avoid-access-invalid-address-when-read-o2dlm-debug-messages >>>> +++ a/fs/ocfs2/dlm/dlmthread.c >>>> @@ -211,6 +211,16 @@ static void dlm_purge_lockres(struct dlm >>>> >>>> __dlm_unhash_lockres(dlm, res); >>>> >>>> + spin_lock(&dlm->track_lock); >>>> + if (!list_empty(&res->tracking)) >>>> + list_del_init(&res->tracking); >>>> + else { >>>> + mlog(ML_ERROR, "Resource %.*s not on the Tracking list\n", >>>> + res->lockname.len, res->lockname.name); >>>> + __dlm_print_one_lock_resource(res); >>>> + } >>>> + spin_unlock(&dlm->track_lock); >>>> + >>>> /* lockres is not in the hash now. drop the flag and wake up >>>> * any processes waiting in dlm_get_lock_resource. */ >>>> if (!master) { >>>> _ >>> >>> How do we know that dlm_purge_lockres() is the last caller of >>> dlm_lockres_put()? Don't we now have a problem where if the last ref is >>> dropped by any other function than dlm_purge_lockres() the lockres is freed >>> while on the tracking list? >>> --Mark >>> >>> -- >>> Mark Fasheh >>> >>> . >>> >> dlm_purge_lockres is not necessarily the last caller of >> dlm_lockres_put(), but it means lockres will be purged if >> dlm_purge_lockres is called. Besides, lockres is also unhashed in >> dlm_purge_lockres, so lockres can be removed from tracking list. >> contents of dlm->tracking_list will be consistent with >> dlm->lockres_hash. > > I'm still confused. This is what I'm worried about: > > 1) a procss calls dlm_lockres_put(), NOT from dlm_purge_lockres(). > > 2) if the count goes to zero, then that process will call dlm_lockres_release() > But at this time, it has already called dlm_purge_lockres(). The reasons are as follows: 1) lockres is created in dlm_init_lockres(), it call kref_init(), count is 1; 2) Only when lockres is unused, it will call dlm_lockres_put() twice by dlm_run_purge_list(). So dlm_purge_lockres() has been called if the count goes to zero. > 3) dlm_lockres_release() will free the lockres without removing it from the > tracking list. Thus we will have a corrupted list. > Without this scene. dlm_purge_lockres() is called before dlm_lockres_release(). > Does that make sense? Am I wrong here? > --Mark > > -- > Mark Fasheh > > . >
Hi jiangyiwen, On 2014/12/20 17:39, jiangyiwen wrote: > ? 2014/12/20 4:25, Mark Fasheh ??: >> On Fri, Dec 19, 2014 at 04:34:39PM +0800, jiangyiwen wrote: >>> ? 2014/12/17 6:26, Mark Fasheh ??: >>>> On Mon, Dec 15, 2014 at 02:51:00PM -0800, Andrew Morton wrote: >>>>> In such a race case, invalid address access may occurs. So we should >>>>> delete list res->tracking before resA->refs decrease to 0. >>>>> >>>> >>>> >>>>> diff -puN fs/ocfs2/dlm/dlmmaster.c~ocfs2-avoid-access-invalid-address-when-read-o2dlm-debug-messages fs/ocfs2/dlm/dlmmaster.c >>>>> --- a/fs/ocfs2/dlm/dlmmaster.c~ocfs2-avoid-access-invalid-address-when-read-o2dlm-debug-messages >>>>> +++ a/fs/ocfs2/dlm/dlmmaster.c >>>>> @@ -498,16 +498,6 @@ static void dlm_lockres_release(struct k >>>>> mlog(0, "destroying lockres %.*s\n", res->lockname.len, >>>>> res->lockname.name); >>>>> >>>>> - spin_lock(&dlm->track_lock); >>>>> - if (!list_empty(&res->tracking)) >>>>> - list_del_init(&res->tracking); >>>>> - else { >>>>> - mlog(ML_ERROR, "Resource %.*s not on the Tracking list\n", >>>>> - res->lockname.len, res->lockname.name); >>>>> - dlm_print_one_lock_resource(res); >>>>> - } >>>>> - spin_unlock(&dlm->track_lock); >>>>> - >>>>> atomic_dec(&dlm->res_cur_count); >>>>> >>>>> if (!hlist_unhashed(&res->hash_node) || >>>>> diff -puN fs/ocfs2/dlm/dlmthread.c~ocfs2-avoid-access-invalid-address-when-read-o2dlm-debug-messages fs/ocfs2/dlm/dlmthread.c >>>>> --- a/fs/ocfs2/dlm/dlmthread.c~ocfs2-avoid-access-invalid-address-when-read-o2dlm-debug-messages >>>>> +++ a/fs/ocfs2/dlm/dlmthread.c >>>>> @@ -211,6 +211,16 @@ static void dlm_purge_lockres(struct dlm >>>>> >>>>> __dlm_unhash_lockres(dlm, res); >>>>> >>>>> + spin_lock(&dlm->track_lock); >>>>> + if (!list_empty(&res->tracking)) >>>>> + list_del_init(&res->tracking); >>>>> + else { >>>>> + mlog(ML_ERROR, "Resource %.*s not on the Tracking list\n", >>>>> + res->lockname.len, res->lockname.name); >>>>> + __dlm_print_one_lock_resource(res); >>>>> + } >>>>> + spin_unlock(&dlm->track_lock); >>>>> + >>>>> /* lockres is not in the hash now. drop the flag and wake up >>>>> * any processes waiting in dlm_get_lock_resource. */ >>>>> if (!master) { >>>>> _ >>>> >>>> How do we know that dlm_purge_lockres() is the last caller of >>>> dlm_lockres_put()? Don't we now have a problem where if the last ref is >>>> dropped by any other function than dlm_purge_lockres() the lockres is freed >>>> while on the tracking list? >>>> --Mark >>>> >>>> -- >>>> Mark Fasheh >>>> >>>> . >>>> >>> dlm_purge_lockres is not necessarily the last caller of >>> dlm_lockres_put(), but it means lockres will be purged if >>> dlm_purge_lockres is called. Besides, lockres is also unhashed in >>> dlm_purge_lockres, so lockres can be removed from tracking list. >>> contents of dlm->tracking_list will be consistent with >>> dlm->lockres_hash. >> >> I'm still confused. This is what I'm worried about: >> >> 1) a procss calls dlm_lockres_put(), NOT from dlm_purge_lockres(). >> >> 2) if the count goes to zero, then that process will call dlm_lockres_release() >> > But at this time, it has already called dlm_purge_lockres(). The reasons are as follows: > 1) lockres is created in dlm_init_lockres(), it call kref_init(), count is 1; > 2) Only when lockres is unused, it will call dlm_lockres_put() twice by dlm_run_purge_list(). > So dlm_purge_lockres() has been called if the count goes to zero. > >> 3) dlm_lockres_release() will free the lockres without removing it from the >> tracking list. Thus we will have a corrupted list. >> > Without this scene. dlm_purge_lockres() is called before dlm_lockres_release(). Once lock resource is inserted into hash list, it should call dlm_purge_lockres()->__dlm_unhash_lockres() to put the last ref. However, if lock resource is initialized and inserted into tracking list but not inserted into hash list, it can happen that dlm_lockres_release() will free the lockres without removing it from the tracking list. So we should remove lockres from tracking list if we call dlm_lockres_put() after lockres is created but not inserted into hash list yet. Thanks, Xuejiufei >> Does that make sense? Am I wrong here? >> --Mark >> >> -- >> Mark Fasheh >> >> . >> > > > > _______________________________________________ > Ocfs2-devel mailing list > Ocfs2-devel@oss.oracle.com > https://oss.oracle.com/mailman/listinfo/ocfs2-devel >
? 2014/12/23 14:06, Xue jiufei ??: > Hi jiangyiwen, > On 2014/12/20 17:39, jiangyiwen wrote: >> ? 2014/12/20 4:25, Mark Fasheh ??: >>> On Fri, Dec 19, 2014 at 04:34:39PM +0800, jiangyiwen wrote: >>>> ? 2014/12/17 6:26, Mark Fasheh ??: >>>>> On Mon, Dec 15, 2014 at 02:51:00PM -0800, Andrew Morton wrote: >>>>>> In such a race case, invalid address access may occurs. So we should >>>>>> delete list res->tracking before resA->refs decrease to 0. >>>>>> >>>>> >>>>> >>>>>> diff -puN fs/ocfs2/dlm/dlmmaster.c~ocfs2-avoid-access-invalid-address-when-read-o2dlm-debug-messages fs/ocfs2/dlm/dlmmaster.c >>>>>> --- a/fs/ocfs2/dlm/dlmmaster.c~ocfs2-avoid-access-invalid-address-when-read-o2dlm-debug-messages >>>>>> +++ a/fs/ocfs2/dlm/dlmmaster.c >>>>>> @@ -498,16 +498,6 @@ static void dlm_lockres_release(struct k >>>>>> mlog(0, "destroying lockres %.*s\n", res->lockname.len, >>>>>> res->lockname.name); >>>>>> >>>>>> - spin_lock(&dlm->track_lock); >>>>>> - if (!list_empty(&res->tracking)) >>>>>> - list_del_init(&res->tracking); >>>>>> - else { >>>>>> - mlog(ML_ERROR, "Resource %.*s not on the Tracking list\n", >>>>>> - res->lockname.len, res->lockname.name); >>>>>> - dlm_print_one_lock_resource(res); >>>>>> - } >>>>>> - spin_unlock(&dlm->track_lock); >>>>>> - >>>>>> atomic_dec(&dlm->res_cur_count); >>>>>> >>>>>> if (!hlist_unhashed(&res->hash_node) || >>>>>> diff -puN fs/ocfs2/dlm/dlmthread.c~ocfs2-avoid-access-invalid-address-when-read-o2dlm-debug-messages fs/ocfs2/dlm/dlmthread.c >>>>>> --- a/fs/ocfs2/dlm/dlmthread.c~ocfs2-avoid-access-invalid-address-when-read-o2dlm-debug-messages >>>>>> +++ a/fs/ocfs2/dlm/dlmthread.c >>>>>> @@ -211,6 +211,16 @@ static void dlm_purge_lockres(struct dlm >>>>>> >>>>>> __dlm_unhash_lockres(dlm, res); >>>>>> >>>>>> + spin_lock(&dlm->track_lock); >>>>>> + if (!list_empty(&res->tracking)) >>>>>> + list_del_init(&res->tracking); >>>>>> + else { >>>>>> + mlog(ML_ERROR, "Resource %.*s not on the Tracking list\n", >>>>>> + res->lockname.len, res->lockname.name); >>>>>> + __dlm_print_one_lock_resource(res); >>>>>> + } >>>>>> + spin_unlock(&dlm->track_lock); >>>>>> + >>>>>> /* lockres is not in the hash now. drop the flag and wake up >>>>>> * any processes waiting in dlm_get_lock_resource. */ >>>>>> if (!master) { >>>>>> _ >>>>> >>>>> How do we know that dlm_purge_lockres() is the last caller of >>>>> dlm_lockres_put()? Don't we now have a problem where if the last ref is >>>>> dropped by any other function than dlm_purge_lockres() the lockres is freed >>>>> while on the tracking list? >>>>> --Mark >>>>> >>>>> -- >>>>> Mark Fasheh >>>>> >>>>> . >>>>> >>>> dlm_purge_lockres is not necessarily the last caller of >>>> dlm_lockres_put(), but it means lockres will be purged if >>>> dlm_purge_lockres is called. Besides, lockres is also unhashed in >>>> dlm_purge_lockres, so lockres can be removed from tracking list. >>>> contents of dlm->tracking_list will be consistent with >>>> dlm->lockres_hash. >>> >>> I'm still confused. This is what I'm worried about: >>> >>> 1) a procss calls dlm_lockres_put(), NOT from dlm_purge_lockres(). >>> >>> 2) if the count goes to zero, then that process will call dlm_lockres_release() >>> >> But at this time, it has already called dlm_purge_lockres(). The reasons are as follows: >> 1) lockres is created in dlm_init_lockres(), it call kref_init(), count is 1; >> 2) Only when lockres is unused, it will call dlm_lockres_put() twice by dlm_run_purge_list(). >> So dlm_purge_lockres() has been called if the count goes to zero. >> >>> 3) dlm_lockres_release() will free the lockres without removing it from the >>> tracking list. Thus we will have a corrupted list. >>> >> Without this scene. dlm_purge_lockres() is called before dlm_lockres_release(). > Once lock resource is inserted into hash list, it should call > dlm_purge_lockres()->__dlm_unhash_lockres() to put the last ref. > However, if lock resource is initialized and inserted into tracking > list but not inserted into hash list, it can happen that > dlm_lockres_release() will free the lockres without removing it from > the tracking list. > So we should remove lockres from tracking list if we call > dlm_lockres_put() after lockres is created but not inserted into hash > list yet. > > Thanks, > Xuejiufei Thanks, you are right! I will send new version later. --Yiwen Jiang >>> Does that make sense? Am I wrong here? >>> --Mark >>> >>> -- >>> Mark Fasheh >>> >>> . >>> >> >> >> >> _______________________________________________ >> Ocfs2-devel mailing list >> Ocfs2-devel@oss.oracle.com >> https://oss.oracle.com/mailman/listinfo/ocfs2-devel >> > > > > . >
diff -puN fs/ocfs2/dlm/dlmmaster.c~ocfs2-avoid-access-invalid-address-when-read-o2dlm-debug-messages fs/ocfs2/dlm/dlmmaster.c --- a/fs/ocfs2/dlm/dlmmaster.c~ocfs2-avoid-access-invalid-address-when-read-o2dlm-debug-messages +++ a/fs/ocfs2/dlm/dlmmaster.c @@ -498,16 +498,6 @@ static void dlm_lockres_release(struct k mlog(0, "destroying lockres %.*s\n", res->lockname.len, res->lockname.name); - spin_lock(&dlm->track_lock); - if (!list_empty(&res->tracking)) - list_del_init(&res->tracking); - else { - mlog(ML_ERROR, "Resource %.*s not on the Tracking list\n", - res->lockname.len, res->lockname.name); - dlm_print_one_lock_resource(res); - } - spin_unlock(&dlm->track_lock); - atomic_dec(&dlm->res_cur_count); if (!hlist_unhashed(&res->hash_node) || diff -puN fs/ocfs2/dlm/dlmthread.c~ocfs2-avoid-access-invalid-address-when-read-o2dlm-debug-messages fs/ocfs2/dlm/dlmthread.c --- a/fs/ocfs2/dlm/dlmthread.c~ocfs2-avoid-access-invalid-address-when-read-o2dlm-debug-messages +++ a/fs/ocfs2/dlm/dlmthread.c @@ -211,6 +211,16 @@ static void dlm_purge_lockres(struct dlm __dlm_unhash_lockres(dlm, res); + spin_lock(&dlm->track_lock); + if (!list_empty(&res->tracking)) + list_del_init(&res->tracking); + else { + mlog(ML_ERROR, "Resource %.*s not on the Tracking list\n", + res->lockname.len, res->lockname.name); + __dlm_print_one_lock_resource(res); + } + spin_unlock(&dlm->track_lock); + /* lockres is not in the hash now. drop the flag and wake up * any processes waiting in dlm_get_lock_resource. */ if (!master) {