Messages in this thread | | | Date | Mon, 9 Feb 2009 12:34:10 +0100 | From | Ingo Molnar <> | Subject | Re: [PATCH] irq: optimize init_kstat_irqs/init_copy_kstat_irqs |
| |
* Andrew Morton <akpm@linux-foundation.org> wrote:
> On Mon, 9 Feb 2009 09:37:39 +0100 Ingo Molnar <mingo@elte.hu> wrote: > > > > > * Andrew Morton <akpm@linux-foundation.org> wrote: > > > > > On Mon, 9 Feb 2009 09:11:24 +0100 Ingo Molnar <mingo@elte.hu> wrote: > > > > > > > > > > > * Andrew Morton <akpm@linux-foundation.org> wrote: > > > > > > > > > On Sat, 7 Feb 2009 01:01:03 -0800 Yinghai Lu <yinghai@kernel.org> wrote: > > > > > > > > > > > > > > > > > add kzalloc_node_safe()? > > > > > > > > > > I cannot find that function. > > > > > > > > His suggestion is to provide that allocator variant. > > > > > > > > > > Oh. > > > > > > It isn't possible to write a kzalloc_node_safe(GFP_ATOMIC). Or at > > > least, we've never worked out a way. > > > > > > Maybe I'm confused again. > > > > Indeed - duh - more morning tea needed. > > > > Yinghai, why are those allocations GFP_ATOMIC to begin with? These: > > > > earth4:~/tip> grep GFP_ATOMIC kernel/irq/*.c > > kernel/irq/handle.c: ptr = kzalloc_node(nr * sizeof(*desc->kstat_irqs), GFP_ATOMIC, node); > > kernel/irq/handle.c: desc = kzalloc_node(sizeof(*desc), GFP_ATOMIC, node); > > kernel/irq/manage.c: action = kmalloc(sizeof(struct irqaction), GFP_ATOMIC); > > > > Should all be GFP_KERNEL. Wherever they are within a spinlocked section the code > > should be restructured. All descriptor data structures should be preallocated at > > __setup_irq() time. If we ever need to allocate dynamically later on, in the middle > > of some difficult codepath that's a structure bug in the code. > > yup, something along those lines. > > > and this one: > > > > kernel/irq/numa_migrate.c: desc = kzalloc_node(sizeof(*desc), GFP_ATOMIC, node); > > > > should fail the migration silently if GFP_ATOMIC returns NULL. > > Silent failure sounds bad? > > The allocation attempt will spew a page-allocation-failure backtrace > anyway, so people will still get alarmed. > > You might instead choose to suppress that warning with __GFP_NOWARN and > instead add a more meaningful warning at the calling codesite.
Well this is a NUMA performance optimization that gets re-tried anyway, so spewing anything because we are so low on RAM that we cannot fulfill GFP_ATOMIC wont help things. Silence is better there i think. At most a WARN_ONCE().
Ingo
| |