Re: [PATCH 09/12] thp: introduce khugepaged_prealloc_page andkhugepaged_alloc_page

From: Hugh Dickins
Date: Tue Sep 11 2012 - 22:05:09 EST

On Mon, 13 Aug 2012, Xiao Guangrong wrote:

> They are used to abstract the difference between NUMA enabled and NUMA disabled
> to make the code more readable
> Signed-off-by: Xiao Guangrong <xiaoguangrong@xxxxxxxxxxxxxxxxxx>
> ---
> mm/huge_memory.c | 166 ++++++++++++++++++++++++++++++++----------------------
> 1 files changed, 98 insertions(+), 68 deletions(-)

Hmm, that in itself is not necessarily an improvement.

I'm a bit sceptical about this patch,
in last Thursday's mmotm 2012-09-06-16-46.

What brought me to look at it was hitting "BUG at mm/huge_memory.c:1842!"
running tmpfs kbuild swapping load (with memcg's memory.limit_in_bytes
forcing out to swap), while I happened to have CONFIG_NUMA=y.

That's the VM_BUG_ON(*hpage) on entry to khugepaged_alloc_page().

(If I'm honest, I'll admit I have Michel's "interval trees for anon rmap"
patches in on top, and so the line number was actually shifted to 1839:
but I don't believe his patches were in any way involved here, and
indeed I've not yet found a problem with them: they look very good.)

I expect the BUG could quite easily be fixed up by making another call
to khugepaged_prealloc_page() from somewhere to free up the hpage;
but forgive me if I dislike using "prealloc" to free.

I do agree with you that the several CONFIG_NUMA ifdefs dotted around
mm/huge_memory.c are regrettable, but I'm not at all sure that you're
improving the situation with this patch, which gives misleading names
to functions and moves the mmap_sem upping out of line.

I think you need to revisit it: maybe not go so far (leaving a few
CONFIG_NUMAs behind, if they're not too bad), or maybe go further
(add a separate function for freeing in the NUMA case, instead of
using "prealloc"). I don't know what's best: have a play and see.

That's what I was intending to write yesterday. But overnight I
was running with this 9/12 backed out (I think 10,11,12 should be
independent), and found "BUG at mm/huge_memory.c:1835!" this morning.

That's the VM_BUG_ON(*hpage) below #else in collapse_huge_page()
when 9/12 is reverted.

So maybe 9/12 is just obscuring what was already a BUG, either earlier
in your series or elsewhere in mmotm (I've never seen it on 3.6-rc or
earlier releases, nor without CONFIG_NUMA). I've not spent any time
looking for it, maybe it's obvious - can you spot and fix it?

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at
Please read the FAQ at