Hi Mattias,
On 2022/10/12 4:26, Mattias Rönnblom wrote:
> On 2022-10-11 17:58, Dmitry Kozlyuk wrote:
>> 2022-10-11 12:17 (UTC+0000), Chengwen Feng:
>>> This patch adds a memarea backup mechanism, where an allocation request
>>> which cannot be met by the current memarea is deferred to its backup
>>> memarea.
>>
>> This is a controversial feature.
>>
>> 1. It violates memarea property of freeing all allocated objects
>> at once when the memarea itself is destroyed. Objects allocated
>> in the backup memarea through the destroyed one will remain.
>>
>> 2. If there was an API to check that the object belongs to a memarea
>> (the check from rte_memarea_update_refcnt() in this patch),
>> it would be trivial to implement this feature on top of memarea API.
>>
>> Nit: "Deferred" is about time -> "forwarded", "delegated", or "handled over".
>>
>> A general note about this series.
>> IMO, libraries should have limited scope and allow composition
>> rather than accumulate features and control their use via options.
>> The core idea of memarea is an allocator within a memory region,
>> a fast one and with low overhead, usable to free all objects at once.
>>
>
> What's a typical use case for a memory region? In a packet processing context.
we used it in video system:
udp-packets -> splitter -> stream-1-reorder -> stream-1-decoder --
| |
| |-->
compose-picture -> picture-encoder -> udp-packets
| |
-> stream-2-reorder -> stream-2-decoder --
each stream-decoder use a dedicated memarea which have following advantage:
1. there will no global lock which like rte_malloc
2. memory objects leakage will only impact the decoder, it will not impact
global system.
3. simple the programmer, only destroy the memarea after the session finished.
As you see, this is a different idea of memory optimization, in pktmbuf_pool,
it use
prealloc + per-lcore cache, it readuces lock conficts by lcore cache, but
didn't fix
memory-leakage's impact.
>
> The ability to instantiate a variable number of heaps/regions seems useful,
> although it's not clear to me if the application should order that to happen
> on a per-lcore basis, on a per-NUMA node basis, a per-<some domain
> object>-basis, or something else entirely.
>
> It seems to me that DPDK is lacking a variable-size memory allocator which is
> efficient and safe to use from lcore threads. My impression is that glibc
> malloc() and rte_malloc() are too slow for the packet processing threads, and
> involves code paths taking locks shared with non-EAL threads.
>
>> This is orthogonal to the question from where the memory comes from.
>> HEAP and LIBC sources could be built on top of USER source,
>> which means that the concept of source is less relevant.
>> Backup mechanism could instead be a way to add memory to the area,
>> in which case HEAP and LIBC memarea would also be expandable.
>> Memarea API could be defined as a structure with callbacks,
>> and different types of memarea could be combined,
>> for example, interlocked memarea on top of expandable memarea on top of
>> memarea with a particular memory management algorithm.
>>
>> I'm not saying we should immediately build all this complexity.
>
> The part with implementing runtime polymorphism using a struct with function
> pointers, instead of the enum+switch-based-type-test approach, doesn't sound
> like something that would add complexity. Rather the opposite.
>
> Also, having a clear-cut separation of concern between
> the-thing-that-allocates-and-frees-the-region and the region-internal memory
> manager (what's called an algorithm in this patchset) also seems like
> something that would simplify the code.
>
>> On the contrary, I would merge the basic things first,
>> then try to _stack_ new features on top,
>> then look if interfaces emerge that can be used for composition.
>
> .
