On 07-May-19 11:04 PM, Carrillo, Erik G wrote:
Hi Anatoly,
Thanks for the review. Comments in-line:
<...snipped...>
#define RTE_MAX_DATA_ELS 64
+static const struct rte_memzone *rte_timer_data_mz; static
+rte_atomic16_t *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; @@ -155,6 +157,7 @@
rte_timer_subsystem_init_v1905(void)
struct rte_timer_data *data;
int i, lcore_id;
static const char *mz_name = "rte_timer_mz";
+ size_t data_arr_size = RTE_MAX_DATA_ELS *
+sizeof(*rte_timer_data_arr);
nitpicking, but... const?
No problem - I'll make this change if this line persists into the next version.
<...snipped...>
@@ -205,8 +216,11 @@
BIND_DEFAULT_SYMBOL(rte_timer_subsystem_init, _v1905, 19.05);
void __rte_experimental
rte_timer_subsystem_finalize(void)
{
- if (rte_timer_data_arr)
- rte_free(rte_timer_data_arr);
+ if (!rte_timer_subsystem_initialized)
+ return;
+
+ if (rte_atomic16_dec_and_test(rte_timer_mz_refcnt))
+ rte_memzone_free(rte_timer_data_mz);
I think there's a race here. You may get preempted after test but before
free, where another secondary could initialize. As far as i know, we also
Indeed, thanks for catching this.
support a case when secondary initializes after primary stops running.
Let's even suppose that we allow secondary processes to initialize the timer
subsystem by reserving memzone and checking rte_errno. You would still
have a chance of two init/deinit conflicting, because there's a hole between
memzone allocation and atomic increment.
I don't think this race can be resolved in a safe way, so we might just have to
settle for a memory leak.
I don't see a solution here currently either. I'll look at removing the
memzone_free()
call and possibly the rte_timer_subsystem_finalize() API, since it seems like
there's no reason for it to exist if it can't free the allocations.
I wonder if there are other places in DPDK where this pattern is used.
Technically, this kind of thing /could/ be resolved by having something
in our multiprocess shared memory outside of DPDK heap. I.e. store
something in rte_eal_memconfig like some other things do. This change,
however, would require an ABI break, so while changing this particular
API won't need a deprecation notice, the change itself would.
Regards,
Erik
rte_timer_subsystem_initialized = 0;
}
--
Thanks,
Anatoly
--
Thanks,
Anatoly
