[lkml]   [2014]   [Jul]   [1]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: mm: memcontrol: rewrite uncharge API: problems
Hi Hugh,

On Mon, Jun 30, 2014 at 04:55:10PM -0700, Hugh Dickins wrote:
> Hi Hannes,
> Your rewrite of the memcg charge/uncharge API is bold and attractive,
> but I'm having some problems with the way release_pages() now does
> uncharging in I/O completion context.

Yes, I need to make the uncharge path IRQ-safe. This looks doable.

> At the bottom see the lockdep message I get when I start shmem swapping.
> Which I have not begun to attempt to decipher (over to you!), but I do
> see release_pages() mentioned in there (also i915, hope it's irrelevant).

This seems to be about uncharge acquiring the IRQ-unsafe soft limit
tree lock while the outer release_pages() holds the IRQ-safe lru_lock.
A separate issue, AFAICS, that would also be fixed by IRQ-proofing the
uncharge path.

> Which was already worrying me on the PowerPC G5, when moving tasks from
> one memcg to another and removing the old, while swapping and swappingoff
> (I haven't tried much else actually, maybe it's much easier to reproduce).
> I get "unable to handle kernel paging at 0x180" oops in __raw_spinlock <
> res_counter_uncharge_until < mem_cgroup_uncharge_end < release_pages <
> free_pages_and_swap_cache < tlb_flush_mmu_free < tlb_finish_mmu <
> unmap_region < do_munmap (or from exit_mmap < mmput < do_exit).
> I do have CONFIG_MEMCG_SWAP=y, and I think 0x180 corresponds to the
> memsw res_counter spinlock, if memcg is NULL. I don't understand why
> usually the PowerPC: I did see something like it once on this x86 laptop,
> maybe having lockdep in on this slows things down enough not to hit that.
> I've stopped those crashes with patch below: the memcg_batch uncharging
> was never designed for use from interrupts. But I bet it needs more work:
> to disable interrupts, or do something clever with atomics, or... over to
> you again.

I was convinced I had tested these changes with lockdep enabled, but
it must have been at an earlier stage while developing the series.
Otherwise, I should have gotten the same splat as you report.

Thanks for the report, I hope to have something useful ASAP.

 \ /
  Last update: 2014-07-01 20:21    [W:0.069 / U:0.360 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site