On Tue, 2018-02-13 at 11:34 -0600, Dennis Zhou wrote: > Hi Eric, > > On Tue, Feb 13, 2018 at 05:35:26AM -0800, Eric Dumazet wrote: > > > > Also I would consider using this fix as I had warnings of cpus being > > stuck there for more than 50 ms : > > > > > > diff --git a/mm/percpu-vm.c b/mm/percpu-vm.c > > index > > 9158e5a81391ced4e268e3d5dd9879c2bc7280ce..6309b01ceb357be01e857e5f899429403836f41f > > 100644 > > --- a/mm/percpu-vm.c > > +++ b/mm/percpu-vm.c > > @@ -92,6 +92,7 @@ static int pcpu_alloc_pages(struct pcpu_chunk *chunk, > > *pagep = alloc_pages_node(cpu_to_node(cpu), gfp, 0); > > if (!*pagep) > > goto err; > > + cond_resched(); > > } > > } > > return 0; > > > > > > This function gets called from pcpu_populate_chunk while holding the > pcpu_alloc_mutex and is called from two scenarios. First, when an > allocation occurs to a place without backing pages, and second when the > workqueue item is scheduled to replenish the number of empty pages. So, > I don't think this is a good idea. >
That _is_ a good idea, we do this already in vmalloc(), and vmalloc() can absolutely be called while some mutex(es) are held. > My understanding is if we're seeing warnings here, that means we're > struggling to find backing pages. I believe adding __GFP_NORETRY on the > workqueue path as Tejun mentioned above would help with warnings as > well, but not if they are caused by the allocation path. > That is a separate concern. My patch simply avoids latency spikes when huge percpu allocations are happening, on systems with say 1024 cpus.