[RFC v4 0/3] lib/fastmem: fast small-object allocator

[Bruce Richardson]

On Mon, Aug 10, 2026 at 06:09:59PM +0200, Morten Brørup wrote:
> > 3. Per-lcore caches: bounded LIFO stacks (no locks on the hot
> >    path). Cache misses trigger bulk transfers to/from the shared
> >    bin under a spinlock.
> 
> I haven't looked at the v4 code in detail, so forgive me if my concern is already addressed.
> 
> I am worried that a non-EAL thread can be preempted while holding a fastmem lock.
> Could this cause an EAL thread to stall in the fastmem lib under the wrong circumstances?
> If so, can this be avoided in any way?
> 
> The Lock-Free Stack has a solution using single-linked lists and a 128-bit CAS operation.
> The solution is quite slow (due to linked list traversal with retry on collision), but progress is practically guaranteed.
> 
This is an issue with many DPDK libraries, even rings and most mempool
implementations. I think it's up to the app developer to ensure that
preemption of non-EAL threads does not cause data-plane issues.

/Bruce