Re: [PATCH 8/9] mm: vmalloc: Offload free_vmap_area_lock global lock

[Uladzislau Rezki]

On Tue, Jun 06, 2023 at 08:11:04PM +0800, Baoquan He wrote:
> On 06/06/23 at 11:01am, Uladzislau Rezki wrote:
> > On Mon, Jun 05, 2023 at 08:43:39AM +0800, Baoquan He wrote:
> > > On 05/22/23 at 01:08pm, Uladzislau Rezki (Sony) wrote:
> > > ......  
> > > > +static unsigned long
> > > > +this_cpu_zone_alloc_fill(struct cpu_vmap_zone *z,
> > > > +	unsigned long size, unsigned long align,
> > > > +	gfp_t gfp_mask, int node)
> > > > +{
> > > > +	unsigned long addr = VMALLOC_END;
> > > > +	struct vmap_area *va;
> > > > +
> > > > +	/*
> > > > +	 * It still can race. One task sets a progress to
> > > > +	 * 1 a second one gets preempted on entry, the first
> > > > +	 * zeroed the progress flag and second proceed with
> > > > +	 * an extra prefetch.
> > > > +	 */
> > > > +	if (atomic_xchg(&z->fill_in_progress, 1))
> > > > +		return addr;
> > > > +
> > > > +	va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
> > > > +	if (unlikely(!va))
> > > > +		goto out;
> > > > +
> > > > +	spin_lock(&free_vmap_area_lock);
> > > > +	addr = __alloc_vmap_area(&free_vmap_area_root, &free_vmap_area_list,
> > > > +		cvz_size, 1, VMALLOC_START, VMALLOC_END);
> > > > +	spin_unlock(&free_vmap_area_lock);
> > > 
> > > The 'z' is passed in from this_cpu_zone_alloc(), and it's got with
> > > raw_cpu_ptr(&cpu_vmap_zone). Here when we try to get chunk of cvz_size
> > > from free_vmap_area_root/free_vmap_area_list, how can we guarantee it
> > > must belong to the 'z' zone? With my understanding, __alloc_vmap_area()
> > > will get efficient address range sequentially bottom up from
> > > free_vmap_area_root. Please correct me if I am wrong.
> > > 
> > We do not guarantee that and it does not worth it. The most important is:
> > 
> > If we search a zone that exactly match a CPU-id the usage of a global
> > vmap space becomes more wider, i.e. toward a high address space. This is
> > not good because we can affect other users which allocate within a specific
> > range. On a big system it might be a problem. Therefore a pre-fetch is done 
> > sequentially on demand.
> > 
> > Secondly, i do not see much difference in performance if we follow
> > exactly CPU-zone-id.
> 
> Ah, I see, the allocated range will be put into appropriate zone's
> busy tree by calculating its zone via addr_to_cvz(va->va_start). The
> cvz->free tree is only a percpu pre-fetch cache. This is smart, thanks a
> lot for explanation. 
> 
Yes. The busy/lazy are placed per-cpu zone(using addr_to_cvz(addr)) whereas
the allocated chunk on a current CPU.

Thanks!

--
Uladzislau Rezki