* Mathieu Desnoyers:
> Adding the NUMA node id to struct rseq is a straightforward thing to do,
> and a good way to figure out if anything in the user-space ecosystem
> prevents extending struct rseq.
>
> This NUMA node id field allows memory allocators such as tcmalloc to
> take advantage of fast access to the current NUMA node id to perform
> NUMA-aware memory allocation.
>
> It is also useful for implementing NUMA-aware user-space mutexes.
It can be used to implement getcpu purely in userspace, too. I had
plan to hack this together with a node ID cache in TLS, which should
offer pretty much the same functionality (except for weird CPU
topology changes which alter the node ID of a previously used CPU).
However, I do not understand the need for two fields here. Why isn't
one enough?
One field would also avoid the need to mess with rseq_cpu_id_state,
maintaining API compatibility.
----- On Jan 31, 2022, at 4:19 PM, Florian Weimer [email protected] wrote:
> * Mathieu Desnoyers:
>
>> Adding the NUMA node id to struct rseq is a straightforward thing to do,
>> and a good way to figure out if anything in the user-space ecosystem
>> prevents extending struct rseq.
>>
>> This NUMA node id field allows memory allocators such as tcmalloc to
>> take advantage of fast access to the current NUMA node id to perform
>> NUMA-aware memory allocation.
>>
>> It is also useful for implementing NUMA-aware user-space mutexes.
>
> It can be used to implement getcpu purely in userspace, too. I had
> plan to hack this together with a node ID cache in TLS, which should
> offer pretty much the same functionality (except for weird CPU
> topology changes which alter the node ID of a previously used CPU).
I suspect that any approach based on a user-space cache will break with
respect to CRIU. That is one big advantage of using the rseq thread area
for this.
>
> However, I do not understand the need for two fields here. Why isn't
> one enough?
As stated in my self-reply, I don't think those two fields are needed
after all.
>
> One field would also avoid the need to mess with rseq_cpu_id_state,
> maintaining API compatibility.
True. However considering that we plan to remove the buggy "rseq_cs.ptr"
fields from the API, that rseq.h UAPI compatibility does not seem to be
very much relevant.
But still, it's better if we can avoid breaking API, agreed. And the
"node_id_start" does not appear to be needed after all.
Thanks,
Mathieu
--
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com
On Mon, Jan 31, 2022 at 10:19:09PM +0100, Florian Weimer wrote:
> It can be used to implement getcpu purely in userspace, too. I had
> plan to hack this together with a node ID cache in TLS, which should
> offer pretty much the same functionality (except for weird CPU
> topology changes which alter the node ID of a previously used CPU).
PowerPC does that quite a lot..
* Peter Zijlstra:
> On Mon, Jan 31, 2022 at 10:19:09PM +0100, Florian Weimer wrote:
>> It can be used to implement getcpu purely in userspace, too. I had
>> plan to hack this together with a node ID cache in TLS, which should
>> offer pretty much the same functionality (except for weird CPU
>> topology changes which alter the node ID of a previously used CPU).
>
> PowerPC does that quite a lot..
How so? Why make it especially hard for userspace?