Skip to contentSkip to conversation

Query:Regarding percpu_counter debug object destroy

DISCUSSION Kohli, Gaurav 2 participants 3 messages Apr 13, 2018 Arm Qualcomm mailing list

lore
Share & tools
lore
#0 lore raw

Hi ,

I have checked below code and it seems we are calling debug_object_free twice, ideally we should deactivate and later we have to destroy.

1st call -> percpu_counter_destroy->debug_percpu_counter_deactivate -> debug_object_free 2nd call ->
debug_object_free

static bool percpu_counter_fixup_free(void *addr, enum debug_obj_state state)

{
        struct percpu_counter *fbc = addr;

        switch (state) {
        case ODEBUG_STATE_ACTIVE:
                percpu_counter_destroy(fbc);  -> first call
                debug_object_free(fbc, &percpu_counter_debug_descr); 2nd call
                return true;
        default:
                return false;
        }
}

We are seeing one issue, where one list contain garbage data in obj_hash, just before element of that is percpu_counter but still not sure as it is very difficult to reproduce.

Can some one please review above code or can we remove one instance of debug_object_free from above code.

Regards
Gaurav

Signature · 2 lines
Qualcomm India Private Limited, on behalf of Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
a Linux Foundation Collaborative Project.
#1 lore raw
Quoted 18 lines · Gaurav

Hi ,

I have checked below code and it seems we are calling debug_object_free twice, ideally we should deactivate and later we
have to destroy.

1st call -> percpu_counter_destroy->debug_percpu_counter_deactivate -> debug_object_free
2nd call ->
debug_object_free

static bool percpu_counter_fixup_free(void *addr, enum debug_obj_state state)
{
        struct percpu_counter *fbc = addr;

Quoted 4 lines

        switch (state) {
        case ODEBUG_STATE_ACTIVE:
                percpu_counter_destroy(fbc);  -> first call debug_object_free(fbc, &percpu_counter_debug_descr); 2nd

Having looked at the code I'd say this is indeed buggy. I'd say it stemmed from same cargo culting since timer_fixup_free follows the same structure of code, except that in del_timer_sync there is no code which does debug_object_free. The situation is similar in work_fixup_free.

So at this point I guess the question is whether we want to leave the debug_object_free call in percpu_counter_fixup_free and just remove debug_percpu_counter_deactivate and open-code the call to debug_object_deactivate in percpu_counter_destroy. Or remove the explicit call in percpu_counter_fixup_free and leave debug_percpu_counter_deactivate.

In the end it's a matter of style, so perhaps Tejun, as the maintainer, has the final say what style he prefers. Personally, I'd go for the former solution so that the percpu follows the style of the rest of the kernel.

Quoted 18 lines

call
                return true;
        default:
                return false;
        }
}

We are seeing one issue, where one list contain garbage data in obj_hash, just before element of that is percpu_counter but still not sure as it is very difficult to reproduce.

Can some one please review above code or can we remove one instance of debug_object_free from above code.

Regards
Gaurav

#2 lore raw

Hi Nikolay,

Thanks for the comment.

I agree ,like timer , hrtimer we have to mark inactive in destroy function and finally freeing the debug object

after destruction of percpu_counter.

But i am still not sure that this double freeing with same address may create race or not in debug_object list.

Regards

Gaurav

Quoted 62 lines · Nikolay Borisov
Quoted 18 lines · Gaurav

Hi ,

I have checked below code and it seems we are calling debug_object_free twice, ideally we should deactivate and later we
have to destroy.

1st call -> percpu_counter_destroy->debug_percpu_counter_deactivate -> debug_object_free
2nd call ->
debug_object_free

static bool percpu_counter_fixup_free(void *addr, enum debug_obj_state state)
{
        struct percpu_counter *fbc = addr;

Quoted 4 lines

        switch (state) {
        case ODEBUG_STATE_ACTIVE:
                percpu_counter_destroy(fbc);  -> first call debug_object_free(fbc, &percpu_counter_debug_descr); 2nd

Having looked at the code I'd say this is indeed buggy. I'd say it stemmed from same cargo culting since timer_fixup_free follows the same structure of code, except that in del_timer_sync there is no code which does debug_object_free. The situation is similar in work_fixup_free.

So at this point I guess the question is whether we want to leave the debug_object_free call in percpu_counter_fixup_free and just remove debug_percpu_counter_deactivate and open-code the call to debug_object_deactivate in percpu_counter_destroy. Or remove the explicit call in percpu_counter_fixup_free and leave debug_percpu_counter_deactivate.

In the end it's a matter of style, so perhaps Tejun, as the maintainer, has the final say what style he prefers. Personally, I'd go for the former solution so that the percpu follows the style of the rest of the kernel.

Quoted 19 lines

call
                return true;
        default:
                return false;
        }
}

We are seeing one issue, where one list contain garbage data in obj_hash, just before element of that is percpu_counter but still not sure as it is very difficult to reproduce.

Can some one please review above code or can we remove one instance of debug_object_free from above code.

Regards
Gaurav

Signature · 2 lines
Qualcomm India Private Limited, on behalf of Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
a Linux Foundation Collaborative Project.