[v3] timer: fix resource leak in finalize

Message ID 1557354906-2500-1-git-send-email-erik.g.carrillo@intel.com (mailing list archive)
State Superseded, archived
Delegated to: Thomas Monjalon
Headers
Series [v3] timer: fix resource leak in finalize |

Checks

Context Check Description
ci/checkpatch success coding style OK

Commit Message

Carrillo, Erik G May 8, 2019, 10:35 p.m. UTC
By using a lock added to the rte_mem_config (which lives in shared
memory), we can synchronize multiple processes in init/finalize and
safely free allocations made during init.

Signed-off-by: Erik Gabriel Carrillo <erik.g.carrillo@intel.com>
---
changes in v3:
 - The previous version had race condition.  This version fixes the race
   by adding a lock in shared memory outside of the DPDK heap area 
   that can be used to safely free the memzone reserved in the subsystem
   init call. (Anatoly)

   This patch depends on http://patches.dpdk.org/patch/53333/.
 
changes in v2:
 - Handle scenario where primary process exits before secondaries such
   that memzone is not freed early (Anatoly)

 lib/librte_eal/common/include/rte_eal_memconfig.h |  3 +++
 lib/librte_timer/rte_timer.c                      | 23 ++++++++++++++++++++++-
 2 files changed, 25 insertions(+), 1 deletion(-)
  

Comments

Burakov, Anatoly May 9, 2019, 8:29 a.m. UTC | #1
On 08-May-19 11:35 PM, Erik Gabriel Carrillo wrote:
> By using a lock added to the rte_mem_config (which lives in shared
> memory), we can synchronize multiple processes in init/finalize and
> safely free allocations made during init.
> 
> Signed-off-by: Erik Gabriel Carrillo <erik.g.carrillo@intel.com>
> ---
> changes in v3:
>   - The previous version had race condition.  This version fixes the race
>     by adding a lock in shared memory outside of the DPDK heap area
>     that can be used to safely free the memzone reserved in the subsystem
>     init call. (Anatoly)
> 
>     This patch depends on http://patches.dpdk.org/patch/53333/.
>   
> changes in v2:
>   - Handle scenario where primary process exits before secondaries such
>     that memzone is not freed early (Anatoly)
> 
>   lib/librte_eal/common/include/rte_eal_memconfig.h |  3 +++
>   lib/librte_timer/rte_timer.c                      | 23 ++++++++++++++++++++++-
>   2 files changed, 25 insertions(+), 1 deletion(-)
> 
> diff --git a/lib/librte_eal/common/include/rte_eal_memconfig.h b/lib/librte_eal/common/include/rte_eal_memconfig.h
> index 84aabe3..8cbc09c 100644
> --- a/lib/librte_eal/common/include/rte_eal_memconfig.h
> +++ b/lib/librte_eal/common/include/rte_eal_memconfig.h
> @@ -64,6 +64,9 @@ struct rte_mem_config {
>   	rte_rwlock_t memory_hotplug_lock;
>   	/**< indicates whether memory hotplug request is in progress. */
>   
> +	rte_spinlock_t timer_subsystem_lock;
> +	/**< indicates whether timer subsystem init/finalize is in progress. */
> +

I believe there's an initialize function somewhere which will initialize 
all of these locks. This lock should be in there as well.

Other than that, i'm OK with this change, however i feel like /just/ 
adding this would be a missed opportunity, because next time we want to 
add something here it will be an ABI break again.

We could perhaps use this opportunity to leave some padding for future 
data. I'm not sure how would that look like, just an idea floating in my 
head :)
  
Thomas Monjalon June 5, 2019, 9:33 a.m. UTC | #2
09/05/2019 10:29, Burakov, Anatoly:
> On 08-May-19 11:35 PM, Erik Gabriel Carrillo wrote:
> > By using a lock added to the rte_mem_config (which lives in shared
> > memory), we can synchronize multiple processes in init/finalize and
> > safely free allocations made during init.
> > 
> > Signed-off-by: Erik Gabriel Carrillo <erik.g.carrillo@intel.com>
> > ---
> > changes in v3:
> >   - The previous version had race condition.  This version fixes the race
> >     by adding a lock in shared memory outside of the DPDK heap area
> >     that can be used to safely free the memzone reserved in the subsystem
> >     init call. (Anatoly)
> > 
> >     This patch depends on http://patches.dpdk.org/patch/53333/.
> >   
> > changes in v2:
> >   - Handle scenario where primary process exits before secondaries such
> >     that memzone is not freed early (Anatoly)
> > 
> >   lib/librte_eal/common/include/rte_eal_memconfig.h |  3 +++
> >   lib/librte_timer/rte_timer.c                      | 23 ++++++++++++++++++++++-
> >   2 files changed, 25 insertions(+), 1 deletion(-)
> > 
> > diff --git a/lib/librte_eal/common/include/rte_eal_memconfig.h b/lib/librte_eal/common/include/rte_eal_memconfig.h
> > index 84aabe3..8cbc09c 100644
> > --- a/lib/librte_eal/common/include/rte_eal_memconfig.h
> > +++ b/lib/librte_eal/common/include/rte_eal_memconfig.h
> > @@ -64,6 +64,9 @@ struct rte_mem_config {
> >   	rte_rwlock_t memory_hotplug_lock;
> >   	/**< indicates whether memory hotplug request is in progress. */
> >   
> > +	rte_spinlock_t timer_subsystem_lock;
> > +	/**< indicates whether timer subsystem init/finalize is in progress. */
> > +
> 
> I believe there's an initialize function somewhere which will initialize 
> all of these locks. This lock should be in there as well.
> 
> Other than that, i'm OK with this change, however i feel like /just/ 
> adding this would be a missed opportunity, because next time we want to 
> add something here it will be an ABI break again.
> 
> We could perhaps use this opportunity to leave some padding for future 
> data. I'm not sure how would that look like, just an idea floating in my 
> head :)

Any update or other opinion?
  
Burakov, Anatoly June 5, 2019, 9:47 a.m. UTC | #3
On 05-Jun-19 10:33 AM, Thomas Monjalon wrote:
> 09/05/2019 10:29, Burakov, Anatoly:
>> On 08-May-19 11:35 PM, Erik Gabriel Carrillo wrote:
>>> By using a lock added to the rte_mem_config (which lives in shared
>>> memory), we can synchronize multiple processes in init/finalize and
>>> safely free allocations made during init.
>>>
>>> Signed-off-by: Erik Gabriel Carrillo <erik.g.carrillo@intel.com>
>>> ---
>>> changes in v3:
>>>    - The previous version had race condition.  This version fixes the race
>>>      by adding a lock in shared memory outside of the DPDK heap area
>>>      that can be used to safely free the memzone reserved in the subsystem
>>>      init call. (Anatoly)
>>>
>>>      This patch depends on http://patches.dpdk.org/patch/53333/.
>>>    
>>> changes in v2:
>>>    - Handle scenario where primary process exits before secondaries such
>>>      that memzone is not freed early (Anatoly)
>>>
>>>    lib/librte_eal/common/include/rte_eal_memconfig.h |  3 +++
>>>    lib/librte_timer/rte_timer.c                      | 23 ++++++++++++++++++++++-
>>>    2 files changed, 25 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/lib/librte_eal/common/include/rte_eal_memconfig.h b/lib/librte_eal/common/include/rte_eal_memconfig.h
>>> index 84aabe3..8cbc09c 100644
>>> --- a/lib/librte_eal/common/include/rte_eal_memconfig.h
>>> +++ b/lib/librte_eal/common/include/rte_eal_memconfig.h
>>> @@ -64,6 +64,9 @@ struct rte_mem_config {
>>>    	rte_rwlock_t memory_hotplug_lock;
>>>    	/**< indicates whether memory hotplug request is in progress. */
>>>    
>>> +	rte_spinlock_t timer_subsystem_lock;
>>> +	/**< indicates whether timer subsystem init/finalize is in progress. */
>>> +
>>
>> I believe there's an initialize function somewhere which will initialize
>> all of these locks. This lock should be in there as well.
>>
>> Other than that, i'm OK with this change, however i feel like /just/
>> adding this would be a missed opportunity, because next time we want to
>> add something here it will be an ABI break again.
>>
>> We could perhaps use this opportunity to leave some padding for future
>> data. I'm not sure how would that look like, just an idea floating in my
>> head :)
> 
> Any update or other opinion?
> 

I have a patchset that hides mem config - if we are to add a lock, it 
would be after adding that patch, as a bugfix (since by then it would n 
longer be an ABI breakage or an API change).
  

Patch

diff --git a/lib/librte_eal/common/include/rte_eal_memconfig.h b/lib/librte_eal/common/include/rte_eal_memconfig.h
index 84aabe3..8cbc09c 100644
--- a/lib/librte_eal/common/include/rte_eal_memconfig.h
+++ b/lib/librte_eal/common/include/rte_eal_memconfig.h
@@ -64,6 +64,9 @@  struct rte_mem_config {
 	rte_rwlock_t memory_hotplug_lock;
 	/**< indicates whether memory hotplug request is in progress. */
 
+	rte_spinlock_t timer_subsystem_lock;
+	/**< indicates whether timer subsystem init/finalize is in progress. */
+
 	/* memory segments and zones */
 	struct rte_fbarray memzones; /**< Memzone descriptors. */
 
diff --git a/lib/librte_timer/rte_timer.c b/lib/librte_timer/rte_timer.c
index c0f5b87..a055d6e 100644
--- a/lib/librte_timer/rte_timer.c
+++ b/lib/librte_timer/rte_timer.c
@@ -26,6 +26,7 @@ 
 #include <rte_malloc.h>
 #include <rte_compat.h>
 #include <rte_errno.h>
+#include <rte_eal_memconfig.h>
 
 #include "rte_timer.h"
 
@@ -60,7 +61,12 @@  struct rte_timer_data {
 	uint8_t internal_flags;
 };
 
+#define RTE_TIMER_SUBSYSTEM_LOCK \
+	(&rte_eal_get_configuration()->mem_config->timer_subsystem_lock)
+
 #define RTE_MAX_DATA_ELS 64
+static const struct rte_memzone *rte_timer_data_mz;
+static int *rte_timer_mz_refcnt;
 static struct rte_timer_data *rte_timer_data_arr;
 static const uint32_t default_data_id;
 static uint32_t rte_timer_subsystem_initialized;
@@ -163,22 +169,27 @@  rte_timer_subsystem_init_v1905(void)
 	if (rte_timer_subsystem_initialized)
 		return -EALREADY;
 
+	rte_spinlock_lock(RTE_TIMER_SUBSYSTEM_LOCK);
 lookup:
 	mz = rte_memzone_lookup(mz_name);
 	if (mz == NULL) {
-		mz = rte_memzone_reserve_aligned(mz_name, data_arr_size,
+		mz = rte_memzone_reserve_aligned(mz_name,
+				data_arr_size + sizeof(*rte_timer_mz_refcnt),
 				SOCKET_ID_ANY, 0, RTE_CACHE_LINE_SIZE);
 		if (mz == NULL) {
 			if (rte_errno == EEXIST)
 				goto lookup;
 
+			rte_spinlock_unlock(RTE_TIMER_SUBSYSTEM_LOCK);
 			return -ENOMEM;
 		}
 
 		do_full_init = true;
 	}
 
+	rte_timer_data_mz = mz;
 	rte_timer_data_arr = mz->addr;
+	rte_timer_mz_refcnt = (void *)((char *)mz->addr + data_arr_size);
 
 	if (do_full_init) {
 		for (i = 0; i < RTE_MAX_DATA_ELS; i++) {
@@ -195,6 +206,9 @@  rte_timer_subsystem_init_v1905(void)
 	}
 
 	rte_timer_data_arr[default_data_id].internal_flags |= FL_ALLOCATED;
+	(*rte_timer_mz_refcnt)++;
+
+	rte_spinlock_unlock(RTE_TIMER_SUBSYSTEM_LOCK);
 
 	rte_timer_subsystem_initialized = 1;
 
@@ -210,6 +224,13 @@  rte_timer_subsystem_finalize(void)
 	if (!rte_timer_subsystem_initialized)
 		return;
 
+	rte_spinlock_lock(RTE_TIMER_SUBSYSTEM_LOCK);
+
+	if (--(*rte_timer_mz_refcnt) == 0)
+		rte_memzone_free(rte_timer_data_mz);
+
+	rte_spinlock_unlock(RTE_TIMER_SUBSYSTEM_LOCK);
+
 	rte_timer_subsystem_initialized = 0;
 }