Re: [PATCH] mm: compaction: Abort compaction if too many pages are isolated and caller is asynchronous

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Thu, Jun 2, 2011 at 10:03 AM, Mel Gorman <mgorman@xxxxxxx> wrote:
> On Thu, Jun 02, 2011 at 01:30:36AM +0200, Andrea Arcangeli wrote:
>> Hi Mel,
>>
>> On Wed, Jun 01, 2011 at 10:40:18PM +0100, Mel Gorman wrote:
>> > On Wed, Jun 01, 2011 at 09:15:29PM +0200, Andrea Arcangeli wrote:
>> > > On Wed, Jun 01, 2011 at 06:58:09PM +0100, Mel Gorman wrote:
>> > > > Umm, HIGHMEM4G implies a two-level pagetable layout so where are
>> > > > things like _PAGE_BIT_SPLITTING being set when THP is enabled?
>> > >
>> > > They should be set on the pgd, pud_offset/pgd_offset will just bypass.
>> > > The splitting bit shouldn't be special about it, the present bit
>> > > should work the same.
>> >
>> > This comment is misleading at best then.
>> >
>> > #define _PAGE_BIT_SPLITTING Â Â _PAGE_BIT_UNUSED1 /* only valid on a PSE pmd */
>>
>> From common code point of view it's set in the pmd, the comment can be
>> extended to specify it's actually the pgd in case of 32bit noPAE but I
>> didn't think it was too misleading as we think in common code terms
>> all over the code, the fact it's a bypass is pretty clear across the
>> whole archs.
>>
>
> Fair point.
>
>> > At the PGD level, it can have PSE set obviously but it's not a
>> > PMD. I confess I haven't checked the manual to see if it's safe to
>> > use _PAGE_BIT_UNUSED1 like this so am taking your word for it. I
>>
>> To be sure I re-checked on 253668.pdf page 113/114 noPAE and page 122
>> PAE, on x86 32bit/64 all ptes/pmd/pgd (32bit/64bit PAE/noPAE) have bit
>> 9-11 "Avail" to software. So I think we should be safe here.
>>
>
> Good stuff. I was reasonably sure this was the case but as this was
> already "impossible", it needed to be ruled out.
>
>> > found that the bug is far harder to reproduce with 3 pagetable levels
>> > than with 2 but that is just timing. So far it has proven impossible
>> > on x86-64 at least within 27 hours so that has me looking at how
>> > pagetable management between x86 and x86-64 differ.
>>
>> Weird.
>>
>> However I could see it screwing the nr_inactive/active_* stats, but
>> the nr_isolated should never go below zero, and especially not anon
>> even if split_huge_page does the accounting wrong (and
>> migrate/compaction won't mess with THP), or at least I'd expect things
>> to fall apart in other ways and not with just a fairly innocuous and
>> not-memory corrupting nr_isolated_ counter going off just by one.
>>
>
> Again, agreed. I found it hard to come up with a reason why file would
> get messed up particularly as PageSwapBacked does not get cleared in the
> ordinary case until the page is freed. If we were using pages after
> being freed due to bad refcounting, it would show up in all sorts of bad
> ways.
>
>> The khugepaged nr_isolated_anon increment couldn't affect the file one
>> and we hold mmap_sem write mode there to prevent the pte to change
>> from under us, in addition to the PT and anon_vma lock. Anon_vma lock
>> being wrong sounds unlikely too, and even if it was it should screw
>> the nr_isolated_anon counter, impossible to screw the nr_isolated_file
>> with khugepaged.
>>
>
> After reviewing, I still could not find a problem with the locking that
> might explain this. I thought last night anon_vma might be bust in some
> way but today I couldn't find a problem.
>
>> Where did you put your bugcheck? It looked like you put it in the < 0
>> reader, can you add it to all _inc/dec/mod (even _inc just in case) so
>> we may get a stack trace including the culprit? (not guaranteed but
>> better chance)
>>
>
> Did that, didn't really help other than showing the corruption happens
> early in the process lifetime while huge PMDs are being faulted. This
> made me think the problem might be on or near fork.
>
>> > Barriers are a big different between how 32-bit !SMP and X86-64 but
>> > don't know yet which one is relevant or if this is even the right
>> > direction.
>>
>> The difference is we need xchg on SMP to avoid losing the dirty
>> bit. Otherwise if we do pmd_t pmd = *pmdp; *pmdp = 0; the dirty bit
>> may have been set in between the two by another thread running in
>> userland in a different CPU, while the pmd was still "present". As
>> long as interrupts don't write to read-write userland memory with the
>> pte dirty bit clear, we shouldn't need xchg on !SMP.
>>
>
> Yep.
>
>> On PAE we also need to write 0 into pmd_low before worrying about
>> pmd_high so the present bit is cleared before clearing the high part
>> of the 32bit PAE pte, and we relay on xchg implicit lock to avoid a
>> smp_wmb() in between the two writes.
>>
>
> Yep.
>
>> I'm unsure if any of this could be relevant to our problem, also there
>
> I concluded after a while that it wasn't. Partially from reasoning about
> it and part by testing forcing the use of the SMP versions and finding
> the bug was still reproducible.
>
>> can't be more than one writer at once in the pmd, as nobody can modify
>> it without the page_table_lock held. xchg there is just to be safe for
>> the dirty bit (or we'd corrupt memory with threads running in userland
>> and writing to memory on other cpus while we ptep_clear_flush).
>>
>> I've been wondering about the lack of "lock" on the bus in atomic.h
>> too, but I can't see how it can possibly matter on !SMP, vmstat
>> modifications should execute only 1 asm insn so preempt or irq can't
>> interrupt it.
>
> To be honest, I haven't fully figured out yet why it makes such
> a difference on !SMP. I have a vague notion that it's because
> the page table page and the data is visible before the bit set by
> SetPageSwapBacked on the struct page is visible but haven't reasoned
> it out yet. If this was the case, it might allow an "anon" page to
> be treated as a file by compaction for accounting purposes and push
> the counter negative but you'd think then the anon isolation would
> be positive so it's something else.
>
> As I thought fork() be an issue, I looked closer at what we do
> there. We are calling pmd_alloc at copy_pmd_range which is a no-op
> when PMD is folded and copy_huge_pmd() is calling pte_alloc_one()
> which also has no barrier. I haven't checked this fully (it's very
> late again as I wasn't able to work on this during most of the day)
> but I wonder if it's then possible the PMD setup is not visible before
> insertion into the page table leading to weirdness? Why it matters to
> SMP is unclear unless this is a preemption thing I'm not thinking of.
>
> On a similar vein, during collapse_huge_page(), we use a barrier
> to ensure the data copy is visible before the PMD insertion but in
> __do_huge_pmd_anonymous_page(), we assume the "spinlocking to take the
> lru_lock inside page_add_new_anon_rmap() acts as a full memory". Thing
> is, it's calling lru_cache_add_lru() adding the page to a pagevec
> which is not necessarily taking the LRU lock and !SMP is leaving a
> big enough race before the pagevec gets drained to cause a problem.
> Of course, maybe it *is* happening on SMP but the negative counters
> are being reported as zero :)

Yes. although we have atomic_inc of in get_page, it doesn't imply full
memory barrier.
So we need explicit memory barrier. I think you're right.
But I can't think of that it's related to this problem(UP, preemption).


-- 
Kind regards,
Minchan Kim

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxxx  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href


[Index of Archives]     [Linux ARM Kernel]     [Linux ARM]     [Linux Omap]     [Fedora ARM]     [IETF Annouce]     [Bugtraq]     [Linux]     [Linux OMAP]     [Linux MIPS]     [ECOS]     [Asterisk Internet PBX]     [Linux API]