On Mon, Jul 01, 2019 at 08:15:50PM -0700, Mike Kravetz wrote:
> On 7/1/19 1:59 AM, Mel Gorman wrote:
> > On Fri, Jun 28, 2019 at 11:20:42AM -0700, Mike Kravetz wrote:
> >> On 4/24/19 7:35 AM, Vlastimil Babka wrote:
> >>> On 4/23/19 6:39 PM, Mike Kravetz wrote:
> >>>>> That being said, I do not think __GFP_RETRY_MAYFAIL is wrong here. It
> >>>>> looks like there is something wrong in the reclaim going on.
> >>>>
> >>>> Ok, I will start digging into that. Just wanted to make sure before I
> >>>> got
> >>>> into it too deep.
> >>>>
> >>>> BTW - This is very easy to reproduce. Just try to allocate more huge
> >>>> pages
> >>>> than will fit into memory. I see this 'reclaim taking forever' behavior
> >>>> on
> >>>> v5.1-rc5-mmotm-2019-04-19-14-53. Looks like it was there in v5.0 as
> >>>> well.
> >>>
> >>> I'd suspect this in should_continue_reclaim():
> >>>
> >>> /* Consider stopping depending on scan and reclaim activity */
> >>> if (sc->gfp_mask & __GFP_RETRY_MAYFAIL) {
> >>> /*
> >>> * For __GFP_RETRY_MAYFAIL allocations, stop reclaiming
> >>> if the
> >>> * full LRU list has been scanned and we are still failing
> >>> * to reclaim pages. This full LRU scan is potentially
> >>> * expensive but a __GFP_RETRY_MAYFAIL caller really
> >>> wants to succeed
> >>> */
> >>> if (!nr_reclaimed && !nr_scanned)
> >>> return false;
> >>>
> >>> And that for some reason, nr_scanned never becomes zero. But it's hard
> >>> to figure out through all the layers of functions :/
> >>
> >> I got back to looking into the direct reclaim/compaction stalls when
> >> trying to allocate huge pages. As previously mentioned, the code is
> >> looping for a long time in shrink_node(). The routine
> >> should_continue_reclaim() returns true perhaps more often than it should.
> >>
> >> As Vlastmil guessed, my debug code output below shows nr_scanned is
> >> remaining
> >> non-zero for quite a while. This was on v5.2-rc6.
> >>
> >
> > I think it would be reasonable to have should_continue_reclaim allow an
> > exit if scanning at higher priority than DEF_PRIORITY - 2, nr_scanned is
> > less than SWAP_CLUSTER_MAX and no pages are being reclaimed.
>
> Thanks Mel,
>
> I added such a check to should_continue_reclaim. However, it does not
> address the issue I am seeing. In that do-while loop in shrink_node,
> the scan priority is not raised (priority--). We can enter the loop
> with priority == DEF_PRIORITY and continue to loop for minutes as seen
> in my previous debug output.
>
Indeed. I'm getting knocked offline shortly so I didn't give this the
time it deserves but it appears that part of this problem is
hugetlb-specific when one node is full and can enter into this continual
loop due to __GFP_RETRY_MAYFAIL requiring both nr_reclaimed and
nr_scanned to be zero.
Have you considered one of the following as an option?
1. Always use the on-stack nodes_allowed in __nr_hugepages_store_common
and copy nodes_states if necessary. Add a bool parameter to
alloc_pool_huge_page that is true when called from set_max_huge_pages.
If an allocation from alloc_fresh_huge_page, clear the failing node
from the mask so it's not retried, bail if the mask is empty. The
consequences are that round-robin allocation of huge pages will be
different if a node failed to allocate for transient reasons.
2. Alter the condition in should_continue_reclaim for
__GFP_RETRY_MAYFAIL to consider if nr_scanned < SWAP_CLUSTER_MAX.
Either raise priority (will interfere with kswapd though) or
bail entirely. Consequences may be that other __GFP_RETRY_MAYFAIL
allocations do not want this behaviour. There are a lot of users.
3. Move where __GFP_RETRY_MAYFAIL is set in a gfp_mask in mm/hugetlb.c.
Strip the flag if an allocation fails on a node. Consequences are
that setting the required number of huge pages is more likely to
return without all the huge pages set.
--
Mel Gorman
SUSE Labs
