Re: [PATCH] bpf: don't fail kmalloc while releasing raw_tp

[Mathieu Desnoyers]

----- On Nov 16, 2020, at 3:44 PM, rostedt rostedt@xxxxxxxxxxx wrote:

> On Mon, 16 Nov 2020 15:37:27 -0500 (EST)
> Mathieu Desnoyers <mathieu.desnoyers@xxxxxxxxxxxx> wrote:
>> > 
>> > Mathieu,
>> > 
>> > Can't we do something that would still allow to unregister a probe even if
>> > a new probe array fails to allocate? We could kick off a irq work to try to
>> > clean up the probe at a later time, but still, the unregister itself should
>> > not fail due to memory failure.
>> 
>> Currently, the fast path iteration looks like:
>> 
>>                 struct tracepoint_func *it_func_ptr;
>>                 void *it_func;
>> 
>>                 it_func_ptr =                                           \
>>                         rcu_dereference_raw((&__tracepoint_##_name)->funcs); \
>>                 do {                                                    \
>>                         it_func = (it_func_ptr)->func;                  \
>>                         __data = (it_func_ptr)->data;                   \
>>                         ((void(*)(void *, proto))(it_func))(__data, args); \
>>                 } while ((++it_func_ptr)->func);
>> 
>> So we RCU dereference the array, and iterate on the array until we find a NULL
>> func. So you could not use NULL to skip items, but you could perhaps reserve
>> a (void *)0x1UL tombstone for this.
> 
> Actually, you could just set it to a stub callback that does nothing. then
> you don't even need to touch the above macro. Not sure why I didn't
> recommend this to begin with, because that's exactly what the function
> tracer does with ftrace_stub.

I like the stub idea.

What prototype should the stub function have ? Is it acceptable to pass
arguments to a function expecting (void) ? If not, then we may need to
create stub functions for each tracepoint.

> 
> 
>> 
>> It should ideally be an unlikely branch, and it would be good to benchmark the
>> change when multiple tracing probes are attached to figure out whether the
>> overhead is significant when tracing is enabled.
> 
> If you use a stub function, it shouldn't affect anything. And the worse
> that would happen is that you have a slight overhead of calling the stub
> until you can properly remove the callback.
> 
>> 
>> I wonder whether we really mind that much about using slightly more memory
>> than required after a failed reallocation due to ENOMEM. Perhaps the irq work
>> is not even needed. Chances are that the irq work would fail again and again if
>> it's in low memory conditions. So maybe it's better to just keep the tombstone
>> in place until the next successful callback array reallocation.
>> 
> 
> True. If we just replace the function with a stub on memory failure (always
> using __GFP_NOFAIL, and if it fails to reallocate a new array, just replace
> the callback with the stub and be done with it. It may require some more
> accounting to make sure the tracepoint.c code can handle these stubs, and
> remove them on new additions to the code.

Yes.

> Heck, if a stub exists, you could just swap it with a new item.

Not without proper synchronization, otherwise you could end up with
mismatch between function callback and private data. The only transition
valid without waiting for an rcu grace period is to change the function
pointer to a stub function. Anything else (e.g. replacing the stub by
some other callback function) would require to wait for a grace period
beforehand.

> But on any new changes to the list, the stubs should be purged.

Yes,

Thanks,

Mathieu

> 
> 
> -- Steve

-- 
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com