| Message ID | 20180525130853.13915-1-vbabka@suse.cz (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Fri, 25 May 2018 15:08:53 +0200 Vlastimil Babka <vbabka@suse.cz> wrote: > we might consider this for 4.17 although I don't know if there's anything > currently broken. Stable backports should be more important, but will have to > be reviewed carefully, as the code went through many changes. > BTW I think that also the ac->preferred_zoneref reset is currently useless if > we don't also reset ac->nodemask from a mempolicy to NULL first (which we > probably should for the OOM victims etc?), but I would leave that for a > separate patch. Confused. If nothing is currently broken then why is a backport needed? Presumably because we expect breakage in the future? Can you expand on this?
On 05/25/2018 09:43 PM, Andrew Morton wrote: > On Fri, 25 May 2018 15:08:53 +0200 Vlastimil Babka <vbabka@suse.cz> wrote: > >> we might consider this for 4.17 although I don't know if there's anything >> currently broken. Stable backports should be more important, but will have to >> be reviewed carefully, as the code went through many changes. >> BTW I think that also the ac->preferred_zoneref reset is currently useless if >> we don't also reset ac->nodemask from a mempolicy to NULL first (which we >> probably should for the OOM victims etc?), but I would leave that for a >> separate patch. > > Confused. If nothing is currently broken then why is a backport > needed? Presumably because we expect breakage in the future? Can you > expand on this? I mean that SLAB is currently not affected, but in older kernels than 4.7 that don't yet have 511e3a058812 ("mm/slab: make cache_grow() handle the page allocated on arbitrary node") it is. That's at least 4.4 LTS. Older ones I'll have to check.
On Fri 25-05-18 12:43:00, Andrew Morton wrote: > On Fri, 25 May 2018 15:08:53 +0200 Vlastimil Babka <vbabka@suse.cz> wrote: > > > we might consider this for 4.17 although I don't know if there's anything > > currently broken. Stable backports should be more important, but will have to > > be reviewed carefully, as the code went through many changes. > > BTW I think that also the ac->preferred_zoneref reset is currently useless if > > we don't also reset ac->nodemask from a mempolicy to NULL first (which we > > probably should for the OOM victims etc?), but I would leave that for a > > separate patch. > > Confused. If nothing is currently broken then why is a backport > needed? Presumably because we expect breakage in the future? Can you > expand on this? __GFP_THISNODE is documented to _use_ the given node. Allocating from a different one is a bug. Maybe the current code can cope with that or at least doesn't blow up in an obvious way but the bug is still there. I am still not sure what to do about the zonelist reset. It still seems like an echo from the past but using numa_node_id for __GFP_THISNODE is a clear bug because our task could have been migrated to a cpu on a different than requested node. Acked-by: Michal Hocko <mhocko@suse.com>
On Fri, May 25, 2018 at 03:08:53PM +0200, Vlastimil Babka wrote: > In __alloc_pages_slowpath() we reset zonelist and preferred_zoneref for > allocations that can ignore memory policies. The zonelist is obtained from > current CPU's node. This is a problem for __GFP_THISNODE allocations that want > to allocate on a different node, e.g. because the allocating thread has been > migrated to a different CPU. > > This has been observed to break SLAB in our 4.4-based kernel, because there it > relies on __GFP_THISNODE working as intended. If a slab page is put on wrong > node's list, then further list manipulations may corrupt the list because > page_to_nid() is used to determine which node's list_lock should be locked and > thus we may take a wrong lock and race. > > Current SLAB implementation seems to be immune by luck thanks to commit > 511e3a058812 ("mm/slab: make cache_grow() handle the page allocated on > arbitrary node") but there may be others assuming that __GFP_THISNODE works as > promised. > > We can fix it by simply removing the zonelist reset completely. There is > actually no reason to reset it, because memory policies and cpusets don't > affect the zonelist choice in the first place. This was different when commit > 183f6371aac2 ("mm: ignore mempolicies when using ALLOC_NO_WATERMARK") > introduced the code, as mempolicies provided their own restricted zonelists. > > Signed-off-by: Vlastimil Babka <vbabka@suse.cz> > Fixes: 183f6371aac2 ("mm: ignore mempolicies when using ALLOC_NO_WATERMARK") Acked-by: Mel Gorman <mgorman@techsingularity.net> Thanks.
On 05/25/2018 10:48 PM, Vlastimil Babka wrote: > On 05/25/2018 09:43 PM, Andrew Morton wrote: >> On Fri, 25 May 2018 15:08:53 +0200 Vlastimil Babka <vbabka@suse.cz> wrote: >> >>> we might consider this for 4.17 although I don't know if there's anything >>> currently broken. Stable backports should be more important, but will have to >>> be reviewed carefully, as the code went through many changes. >>> BTW I think that also the ac->preferred_zoneref reset is currently useless if >>> we don't also reset ac->nodemask from a mempolicy to NULL first (which we >>> probably should for the OOM victims etc?), but I would leave that for a >>> separate patch. >> >> Confused. If nothing is currently broken then why is a backport >> needed? Presumably because we expect breakage in the future? Can you >> expand on this? > > I mean that SLAB is currently not affected, but in older kernels than > 4.7 that don't yet have 511e3a058812 ("mm/slab: make cache_grow() handle > the page allocated on arbitrary node") it is. That's at least 4.4 LTS. > Older ones I'll have to check. So I've checked the non-EOL LTS's at kernel.org and: 4.16, 4.14, 4.9 - same as mainline (__GFP_THISNODE broken, but SLAB is OK) 4.4, 4.1, 3.16 - SLAB potentially broken if it makes an ALLOC_NO_WATERMARKS allocation (our 4.4 kernel has backports that extend it to also !ALLOC_CPUSET so it's more likely).
On 05/28/2018 09:21 AM, Michal Hocko wrote: > On Fri 25-05-18 12:43:00, Andrew Morton wrote: >> On Fri, 25 May 2018 15:08:53 +0200 Vlastimil Babka <vbabka@suse.cz> wrote: >> >>> we might consider this for 4.17 although I don't know if there's anything >>> currently broken. Stable backports should be more important, but will have to >>> be reviewed carefully, as the code went through many changes. >>> BTW I think that also the ac->preferred_zoneref reset is currently useless if >>> we don't also reset ac->nodemask from a mempolicy to NULL first (which we >>> probably should for the OOM victims etc?), but I would leave that for a >>> separate patch. >> >> Confused. If nothing is currently broken then why is a backport >> needed? Presumably because we expect breakage in the future? Can you >> expand on this? > > __GFP_THISNODE is documented to _use_ the given node. Allocating from a > different one is a bug. Maybe the current code can cope with that or at > least doesn't blow up in an obvious way but the bug is still there. > > I am still not sure what to do about the zonelist reset. It still seems > like an echo from the past Hmm actually it seems that even at the time of commit 183f6371aac2 introduced the reset, the per-policy zonelists for MPOL_BIND policies were gone for years. Mempolicy only affects which node's zonelist is used, but that always contains all the nodes (unless __GFP_THISNODE) so there's no reason to get another node's zonelist to escape mempolicy restrictions. Mempolicy restrictions are given as nodemask, so if we want to ignore them for OOM victims etc, we have to reset nodemask instead. But again we have to be careful in case the nodemask doesn't come from mempolicy, but from somebody who might be broken if we ignore it. > but using numa_node_id for __GFP_THISNODE is > a clear bug because our task could have been migrated to a cpu on a > different than requested node. > > Acked-by: Michal Hocko <mhocko@suse.com> >
diff --git a/mm/page_alloc.c b/mm/page_alloc.c index 905db9d7962f..be0f0b5d3935 100644 --- a/mm/page_alloc.c +++ b/mm/page_alloc.c @@ -4165,7 +4165,6 @@ __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order, * orientated. */ if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) { - ac->zonelist = node_zonelist(numa_node_id(), gfp_mask); ac->preferred_zoneref = first_zones_zonelist(ac->zonelist, ac->high_zoneidx, ac->nodemask); }
In __alloc_pages_slowpath() we reset zonelist and preferred_zoneref for allocations that can ignore memory policies. The zonelist is obtained from current CPU's node. This is a problem for __GFP_THISNODE allocations that want to allocate on a different node, e.g. because the allocating thread has been migrated to a different CPU. This has been observed to break SLAB in our 4.4-based kernel, because there it relies on __GFP_THISNODE working as intended. If a slab page is put on wrong node's list, then further list manipulations may corrupt the list because page_to_nid() is used to determine which node's list_lock should be locked and thus we may take a wrong lock and race. Current SLAB implementation seems to be immune by luck thanks to commit 511e3a058812 ("mm/slab: make cache_grow() handle the page allocated on arbitrary node") but there may be others assuming that __GFP_THISNODE works as promised. We can fix it by simply removing the zonelist reset completely. There is actually no reason to reset it, because memory policies and cpusets don't affect the zonelist choice in the first place. This was different when commit 183f6371aac2 ("mm: ignore mempolicies when using ALLOC_NO_WATERMARK") introduced the code, as mempolicies provided their own restricted zonelists. Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Fixes: 183f6371aac2 ("mm: ignore mempolicies when using ALLOC_NO_WATERMARK") Cc: <stable@vger.kernel.org> --- Hi, we might consider this for 4.17 although I don't know if there's anything currently broken. Stable backports should be more important, but will have to be reviewed carefully, as the code went through many changes. BTW I think that also the ac->preferred_zoneref reset is currently useless if we don't also reset ac->nodemask from a mempolicy to NULL first (which we probably should for the OOM victims etc?), but I would leave that for a separate patch. Vlastimil mm/page_alloc.c | 1 - 1 file changed, 1 deletion(-)