On Thu, 23 Mar 2023 20:21:11 +0100 "Uladzislau Rezki (Sony)" <urezki@xxxxxxxxx> wrote:
> A global vmap_blocks-xarray array can be contented under
> heavy usage of the vm_map_ram()/vm_unmap_ram() APIs. The
> lock_stat shows that a "vmap_blocks.xa_lock" lock is a
> second in a top-list when it comes to contentions:
>
> ...
>
> This patch does not fix vmap_area_lock/free_vmap_area_lock and
> purge_vmap_area_lock bottle-necks, it is rather a separate rework.
>
> ...
>
> static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
>
> ...
>
> +static struct vmap_block_queue *
> +addr_to_vbq(unsigned long addr)
> +{
> + int cpu = (addr / VMAP_BLOCK_SIZE) % num_possible_cpus();
> + return &per_cpu(vmap_block_queue, cpu);
> +}
Seems strange. vmap_block_queue is not a per-cpu thing in this usage.
Instead it's a hash table, indexed off the (hashed) address, not off
smp_processor_id().
Yet in other places, vmap_block_queue *is* used in the conventional
cpu-local fashion.
So we can have CPU A using the cpu-local entry in vmap_block_queue
while CPU B is simultaneously using it, having looked it up via `addr'.
AFAICT this all works OK, no races.
But still, what it's doing is mixing an addr-indexed hashtable with the
CPU-indexed array in surprising ways. It would be clearer to make the
vmap_blocks array a separate thing from the per-cpu array, although it
would presumably use a bit more memory.
Can we please at least get a big fat comment in an appropriate place
which explains all this to the reader?
