On 06/25/24 at 10:05pm, Uladzislau Rezki wrote:
> > > > > > /**
> > > > > > * cpumask_next - get the next cpu in a cpumask
> > > > > > * @n: the cpu prior to the place to search (i.e. return will be > @n)
> > > > > > * @srcp: the cpumask pointer
> > > > > > *
> > > > > > * Return: >= nr_cpu_ids if no further cpus set.
> > > > >
> > > > > Ah, I got what you mean. In the vbq case, it may not have chance to get
> > > > > a return number as nr_cpu_ids. Becuase the hashed index limits the
> > > > > range to [0, nr_cpu_ids-1], and cpu_possible(index) will guarantee it
> > > > > won't be the highest cpu number [nr_cpu_ids-1] since CPU[nr_cpu_ids-1] must
> > > > > be possible CPU.
> > > > >
> > > > > Do I miss some corner cases?
> > > > >
> > > > Right. We guarantee that a highest CPU is available by doing: % nr_cpu_ids.
> > > > So we do not need to use *next_wrap() variant. You do not miss anything :)
> > > >
> > > > Hailong Liu has proposed more simpler version:
> > > >
> > > > <snip>
> > > > diff --git a/mm/vmalloc.c b/mm/vmalloc.c
> > > > index 11fe5ea208aa..e1e63ffb9c57 100644
> > > > --- a/mm/vmalloc.c
> > > > +++ b/mm/vmalloc.c
> > > > @@ -1994,8 +1994,9 @@ static struct xarray *
> > > > addr_to_vb_xa(unsigned long addr)
> > > > {
> > > > int index = (addr / VMAP_BLOCK_SIZE) % num_possible_cpus();
> > > > + int cpu = cpumask_nth(index, cpu_possible_mask);
> > > >
> > > > - return &per_cpu(vmap_block_queue, index).vmap_blocks;
> > > > + return &per_cpu(vmap_block_queue, cpu).vmap_blocks;
> > > > <snip>
> > > >
> > > > which just takes a next CPU if an index is not set in the cpu_possible_mask.
> > > >
> > > > The only thing that can be updated in the patch is to replace num_possible_cpu()
> > > > by the nr_cpu_ids.
> > > >
> > > > Any thoughts? I think we need to fix it by a minor change so it is
> > > > easier to back-port on stable kernels.
> > >
> > > Yeah, sounds good since the regresson commit is merged in v6.3.
> > > Please feel free to post this and the hash array patch separately for
> > > formal reviewing.
> > >
> > Agreed! The patch about hash array i will post later.
> >
> > > By the way, when I am replying this mail, I check the cpumask_nth()
> > > again. I doubt it may take more checking then cpu_possible(), given most
> > > of systems don't have gaps in cpu_possible_mask. I could be dizzy at
> > > this moment.
> > >
> > > static inline unsigned int cpumask_nth(unsigned int cpu, const struct cpumask *srcp)
> > > {
> > > return find_nth_bit(cpumask_bits(srcp), small_cpumask_bits, cpumask_check(cpu));
> > > }
> > >
> > Yep, i do not think it is a big problem based on your noted fact.
> >
> Checked. There is a difference:
>
> 1. Default
>
> <snip>
> ...
> + 15.95% 6.05% [kernel] [k] __vmap_pages_range_noflush
> + 15.91% 1.74% [kernel] [k] addr_to_vb_xa <---------------
> + 15.13% 12.05% [kernel] [k] vunmap_p4d_range
> + 14.17% 13.38% [kernel] [k] __find_nth_bit <--------------
> + 10.62% 0.00% [kernel] [k] ret_from_fork_asm
> + 10.62% 0.00% [kernel] [k] ret_from_fork
> + 10.62% 0.00% [kernel] [k] kthread
> ...
> <snip>
>
> 2. Check if cpu_possible() and then fallback to cpumask_nth() if not
>
> <snip>
> ...
> + 6.84% 0.29% [kernel] [k] alloc_vmap_area
> + 6.80% 6.70% [kernel] [k] native_queued_spin_lock_slowpath
> + 4.24% 0.09% [kernel] [k] free_vmap_block
> + 2.41% 2.38% [kernel] [k] addr_to_vb_xa <-----------
> + 1.94% 1.91% [kernel] [k] xas_start
> ...
> <snip>
>
> It is _worth_ to check if an index is in possible mask:
>
> diff --git a/mm/vmalloc.c b/mm/vmalloc.c
> index 45e1506d58c3..af20f78c2cbf 100644
> --- a/mm/vmalloc.c
> +++ b/mm/vmalloc.c
> @@ -2542,7 +2542,10 @@ static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
> static struct xarray *
> addr_to_vb_xa(unsigned long addr)
> {
> - int index = (addr / VMAP_BLOCK_SIZE) % num_possible_cpus();
> + int index = (addr / VMAP_BLOCK_SIZE) % nr_cpu_ids;
> +
> + if (!cpu_possible(index))
> + index = cpumask_nth(index, cpu_possible_mask);
>
> return &per_cpu(vmap_block_queue, index).vmap_blocks;
> }
>
> cpumask_nth() is not cheap. My measurements are based on a synthetic
> tight test and it detects a difference. In a real workloads it should
> not be visible. Having gaps is not a common case plus a "slow path"
> will be mitigated by the hit against possible mask.
Ah, this is consistent with my understanding from the code, thanks
for confirming by testing.
