| Message ID | 1525408246-14768-1-git-send-email-iamjoonsoo.kim@lge.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On 05/04/2018 06:30 AM, js1304@gmail.com wrote: > From: Joonsoo Kim <iamjoonsoo.kim@lge.com> > > Currently, we use the zone index of preferred_zone which represents > the best matching zone for allocation, as classzone_idx. It has a problem > on NUMA system with ZONE_MOVABLE. > > In NUMA system, it can be possible that each node has different populated > zones. For example, node 0 could have DMA/DMA32/NORMAL/MOVABLE zone and > node 1 could have only NORMAL zone. In this setup, allocation request > initiated on node 0 and the one on node 1 would have different > classzone_idx, 3 and 2, respectively, since their preferred_zones are > different. If they are handled by only their own node, there is no problem. > However, if they are somtimes handled by the remote node, the problem > would happen. > > In the following setup, allocation initiated on node 1 will have some > precedence than allocation initiated on node 0 when former allocation is > processed on node 0 due to not enough memory on node 1. They will have > different lowmem reserve due to their different classzone_idx thus > an watermark bars are also different. > ... > > min watermark for NORMAL zone on node 0 > allocation initiated on node 0: 750 + 4096 = 4846 > allocation initiated on node 1: 750 + 0 = 750 > > This watermark difference could cause too many numa_miss allocation > in some situation and then performance could be downgraded. > > Recently, there was a regression report about this problem on CMA patches > since CMA memory are placed in ZONE_MOVABLE by those patches. I checked > that problem is disappeared with this fix that uses high_zoneidx > for classzone_idx. > > http://lkml.kernel.org/r/20180102063528.GG30397@yexl-desktop > > Using high_zoneidx for classzone_idx is more consistent way than previous > approach because system's memory layout doesn't affect anything to it. So to summarize; - ac->high_zoneidx is computed via the arcane gfp_zone(gfp_mask) and represents the highest zone the allocation can use - classzone_idx was supposed to be the highest zone that the allocation can use, that is actually available in the system. Somehow that became the highest zone that is available on the preferred node (in the default node-order zonelist), which causes the watermark inconsistencies you mention. I don't see a problem with your change. I would be worried about inflated reserves when e.g. ZONE_MOVABLE doesn't exist, but that doesn't seem to be the case. My laptop has empty ZONE_MOVABLE and the ZONE_NORMAL protection for movable is 0. But there had to be some reason for classzone_idx to be like this and not simple high_zoneidx. Maybe Mel remembers? Maybe it was important then, but is not anymore? Sigh, it seems to be pre-git. > With this patch, both classzone_idx on above example will be 3 so will > have the same min watermark. > > allocation initiated on node 0: 750 + 4096 = 4846 > allocation initiated on node 1: 750 + 4096 = 4846 > > Reported-by: Ye Xiaolong <xiaolong.ye@intel.com> > Tested-by: Ye Xiaolong <xiaolong.ye@intel.com> > Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com> > --- > mm/internal.h | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > > diff --git a/mm/internal.h b/mm/internal.h > index 228dd66..e1d7376 100644 > --- a/mm/internal.h > +++ b/mm/internal.h > @@ -123,7 +123,7 @@ struct alloc_context { > bool spread_dirty_pages; > }; > > -#define ac_classzone_idx(ac) zonelist_zone_idx(ac->preferred_zoneref) > +#define ac_classzone_idx(ac) (ac->high_zoneidx) > > /* > * Locate the struct page for both the matching buddy in our >
2018-05-04 16:03 GMT+09:00 Vlastimil Babka <vbabka@suse.cz>: > On 05/04/2018 06:30 AM, js1304@gmail.com wrote: >> From: Joonsoo Kim <iamjoonsoo.kim@lge.com> >> >> Currently, we use the zone index of preferred_zone which represents >> the best matching zone for allocation, as classzone_idx. It has a problem >> on NUMA system with ZONE_MOVABLE. >> >> In NUMA system, it can be possible that each node has different populated >> zones. For example, node 0 could have DMA/DMA32/NORMAL/MOVABLE zone and >> node 1 could have only NORMAL zone. In this setup, allocation request >> initiated on node 0 and the one on node 1 would have different >> classzone_idx, 3 and 2, respectively, since their preferred_zones are >> different. If they are handled by only their own node, there is no problem. >> However, if they are somtimes handled by the remote node, the problem >> would happen. >> >> In the following setup, allocation initiated on node 1 will have some >> precedence than allocation initiated on node 0 when former allocation is >> processed on node 0 due to not enough memory on node 1. They will have >> different lowmem reserve due to their different classzone_idx thus >> an watermark bars are also different. >> > ... > >> >> min watermark for NORMAL zone on node 0 >> allocation initiated on node 0: 750 + 4096 = 4846 >> allocation initiated on node 1: 750 + 0 = 750 >> >> This watermark difference could cause too many numa_miss allocation >> in some situation and then performance could be downgraded. >> >> Recently, there was a regression report about this problem on CMA patches >> since CMA memory are placed in ZONE_MOVABLE by those patches. I checked >> that problem is disappeared with this fix that uses high_zoneidx >> for classzone_idx. >> >> http://lkml.kernel.org/r/20180102063528.GG30397@yexl-desktop >> >> Using high_zoneidx for classzone_idx is more consistent way than previous >> approach because system's memory layout doesn't affect anything to it. > > So to summarize; > - ac->high_zoneidx is computed via the arcane gfp_zone(gfp_mask) and > represents the highest zone the allocation can use > - classzone_idx was supposed to be the highest zone that the allocation > can use, that is actually available in the system. Somehow that became > the highest zone that is available on the preferred node (in the default > node-order zonelist), which causes the watermark inconsistencies you > mention. Yes! Thanks for summarize! > I don't see a problem with your change. I would be worried about > inflated reserves when e.g. ZONE_MOVABLE doesn't exist, but that doesn't > seem to be the case. My laptop has empty ZONE_MOVABLE and the > ZONE_NORMAL protection for movable is 0. Yes! Protection number is calculated by using the number of managed page in upper zone. If there is no memory on the upper zone, protection will be 0. > But there had to be some reason for classzone_idx to be like this and > not simple high_zoneidx. Maybe Mel remembers? Maybe it was important > then, but is not anymore? Sigh, it seems to be pre-git. Based on my code inspection, this patch changing classzone_idx implementation would not cause the problem. I also have tried to find the reason for classzone_idx implementation by searching git history but I can't. As you said, it seems to be pre-git. It would be really helpful that someone who remembers the reason for current classzone_idx implementation teaches me the reason. Thanks.
On Fri, May 04, 2018 at 09:03:02AM +0200, Vlastimil Babka wrote: > > min watermark for NORMAL zone on node 0 > > allocation initiated on node 0: 750 + 4096 = 4846 > > allocation initiated on node 1: 750 + 0 = 750 > > > > This watermark difference could cause too many numa_miss allocation > > in some situation and then performance could be downgraded. > > > > Recently, there was a regression report about this problem on CMA patches > > since CMA memory are placed in ZONE_MOVABLE by those patches. I checked > > that problem is disappeared with this fix that uses high_zoneidx > > for classzone_idx. > > > > http://lkml.kernel.org/r/20180102063528.GG30397@yexl-desktop > > > > Using high_zoneidx for classzone_idx is more consistent way than previous > > approach because system's memory layout doesn't affect anything to it. > > So to summarize; > - ac->high_zoneidx is computed via the arcane gfp_zone(gfp_mask) and > represents the highest zone the allocation can use It's arcane but it was simply a fast-path calculation. A much older definition would be easier to understand but it was slower. > - classzone_idx was supposed to be the highest zone that the allocation > can use, that is actually available in the system. Somehow that became > the highest zone that is available on the preferred node (in the default > node-order zonelist), which causes the watermark inconsistencies you > mention. > I think it *always* was the index of the first preferred zone of a zonelist. The treatment of classzone has changed a lot over the years and I didn't do a historical check but the general intent was always "protect some pages in lower zones". This was particularly important for 32-bit and highmem albeit that is less of a concern today. When it transferred to NUMA, I don't think it ever was seriously considered if it should change as the critical node was likely to be node 0 with all the zones and the remote nodes all used the highest zone. CMA/MOVABLE changed that slightly by allowing the possibility of node0 having a "higher" zone than every other node. When MOVABLE was introduced, it wasn't much of a problem as the purpose of MOVABLE was for systems that dynamically needed to allocate hugetlbfs later in the runtime but for CMA, it was a lot more critical for ordinary usage so this is primarily a CMA thing. > I don't see a problem with your change. I would be worried about > inflated reserves when e.g. ZONE_MOVABLE doesn't exist, but that doesn't > seem to be the case. My laptop has empty ZONE_MOVABLE and the > ZONE_NORMAL protection for movable is 0. > > But there had to be some reason for classzone_idx to be like this and > not simple high_zoneidx. Maybe Mel remembers? Maybe it was important > then, but is not anymore? Sigh, it seems to be pre-git. > classzone predates my involvement with Linux but I would be less concerneed about what the original intent was and instead ensure that classzone index is consistent, sane and potentially renamed while preserving the intent of "reserve pages in lower zones when an allocation request can use higher zones". While historically the critical intent was to preserve Normal and to a lesser extent DMA on 32-bit systems, there still should be some care of DMA32 so we should not lose that. With the patch, the allocator looks like it would be fine as just reservations change. I think it's unlikely that CMA usage will result in lowmem starvation. Compaction becomes a bit weird as classzone index has no special meaning versis highmem and I think it'll be very easy to forget. Similarly, vmscan can reclaim pages from remote nodes and zones that are higher than the original request. That is not likely to be a problem but it's a change in behaviour and easy to miss. Fundamentally, I find it extremely weird we now have two variables that are essentially the same thing. They should be collapsed into one variable, renamed and documented on what the index means for page allocator, compaction, vmscan and the special casing around CMA.
Hello, Mel. Thanks for precious input! 2018-05-04 19:33 GMT+09:00 Mel Gorman <mgorman@suse.de>: > On Fri, May 04, 2018 at 09:03:02AM +0200, Vlastimil Babka wrote: >> > min watermark for NORMAL zone on node 0 >> > allocation initiated on node 0: 750 + 4096 = 4846 >> > allocation initiated on node 1: 750 + 0 = 750 >> > >> > This watermark difference could cause too many numa_miss allocation >> > in some situation and then performance could be downgraded. >> > >> > Recently, there was a regression report about this problem on CMA patches >> > since CMA memory are placed in ZONE_MOVABLE by those patches. I checked >> > that problem is disappeared with this fix that uses high_zoneidx >> > for classzone_idx. >> > >> > http://lkml.kernel.org/r/20180102063528.GG30397@yexl-desktop >> > >> > Using high_zoneidx for classzone_idx is more consistent way than previous >> > approach because system's memory layout doesn't affect anything to it. >> >> So to summarize; >> - ac->high_zoneidx is computed via the arcane gfp_zone(gfp_mask) and >> represents the highest zone the allocation can use > > It's arcane but it was simply a fast-path calculation. A much older > definition would be easier to understand but it was slower. > >> - classzone_idx was supposed to be the highest zone that the allocation >> can use, that is actually available in the system. Somehow that became >> the highest zone that is available on the preferred node (in the default >> node-order zonelist), which causes the watermark inconsistencies you >> mention. >> > > I think it *always* was the index of the first preferred zone of a > zonelist. The treatment of classzone has changed a lot over the years and > I didn't do a historical check but the general intent was always "protect > some pages in lower zones". This was particularly important for 32-bit > and highmem albeit that is less of a concern today. When it transferred to > NUMA, I don't think it ever was seriously considered if it should change > as the critical node was likely to be node 0 with all the zones and the > remote nodes all used the highest zone. CMA/MOVABLE changed that slightly > by allowing the possibility of node0 having a "higher" zone than every I think that this problem is related to not only protection of the lowmem (that is lower than normal) but also node balance. In fact, problem reported by zeroday-bot is caused by node1 having a "higher" zone. In this case, node0's lowmem is protected well but node balance of the allocation is broken since node1's normal memory cannot be protected from allocation that is initiated on remote node. > other node. When MOVABLE was introduced, it wasn't much of a problem as > the purpose of MOVABLE was for systems that dynamically needed to allocate > hugetlbfs later in the runtime but for CMA, it was a lot more critical > for ordinary usage so this is primarily a CMA thing. I'm not sure that it's primarily a CMA thing. There is an another critical setup for this problem, that is, memory hotplug. If someone plug-in a new memory to the MOVABLE zone, "higher" zone will be created in a specific node and this problem happens. I have checked this with QEMU. >> I don't see a problem with your change. I would be worried about >> inflated reserves when e.g. ZONE_MOVABLE doesn't exist, but that doesn't >> seem to be the case. My laptop has empty ZONE_MOVABLE and the >> ZONE_NORMAL protection for movable is 0. >> >> But there had to be some reason for classzone_idx to be like this and >> not simple high_zoneidx. Maybe Mel remembers? Maybe it was important >> then, but is not anymore? Sigh, it seems to be pre-git. >> > > classzone predates my involvement with Linux but I would be less concerneed > about what the original intent was and instead ensure that classzone index > is consistent, sane and potentially renamed while preserving the intent of > "reserve pages in lower zones when an allocation request can use higher > zones". While historically the critical intent was to preserve Normal and > to a lesser extent DMA on 32-bit systems, there still should be some care > of DMA32 so we should not lose that. Agreed! > With the patch, the allocator looks like it would be fine as just > reservations change. I think it's unlikely that CMA usage will result > in lowmem starvation. Compaction becomes a bit weird as classzone index > has no special meaning versis highmem and I think it'll be very easy to > forget. Similarly, vmscan can reclaim pages from remote nodes and zones > that are higher than the original request. That is not likely to be a > problem but it's a change in behaviour and easy to miss. > > Fundamentally, I find it extremely weird we now have two variables that are > essentially the same thing. They should be collapsed into one variable, > renamed and documented on what the index means for page allocator, > compaction, vmscan and the special casing around CMA. Agreed! I will update this patch to reflect your comment. If someone have an idea on renaming this variable, please let me know. Thanks.
On Fri, 4 May 2018 11:33:22 +0100 Mel Gorman <mgorman@suse.de> wrote: > On Fri, May 04, 2018 at 09:03:02AM +0200, Vlastimil Babka wrote: > > > min watermark for NORMAL zone on node 0 > > > allocation initiated on node 0: 750 + 4096 = 4846 > > > allocation initiated on node 1: 750 + 0 = 750 > > > > > > This watermark difference could cause too many numa_miss allocation > > > in some situation and then performance could be downgraded. > > > > > > Recently, there was a regression report about this problem on CMA patches > > > since CMA memory are placed in ZONE_MOVABLE by those patches. I checked > > > that problem is disappeared with this fix that uses high_zoneidx > > > for classzone_idx. > > > > > > http://lkml.kernel.org/r/20180102063528.GG30397@yexl-desktop > > > > > > Using high_zoneidx for classzone_idx is more consistent way than previous > > > approach because system's memory layout doesn't affect anything to it. > > > > So to summarize; > > - ac->high_zoneidx is computed via the arcane gfp_zone(gfp_mask) and > > represents the highest zone the allocation can use > > It's arcane but it was simply a fast-path calculation. A much older > definition would be easier to understand but it was slower. > > > - classzone_idx was supposed to be the highest zone that the allocation > > can use, that is actually available in the system. Somehow that became > > the highest zone that is available on the preferred node (in the default > > node-order zonelist), which causes the watermark inconsistencies you > > mention. > > > > I think it *always* was the index of the first preferred zone of a > zonelist. The treatment of classzone has changed a lot over the years and > I didn't do a historical check but the general intent was always "protect > some pages in lower zones". This was particularly important for 32-bit > and highmem albeit that is less of a concern today. When it transferred to > NUMA, I don't think it ever was seriously considered if it should change > as the critical node was likely to be node 0 with all the zones and the > remote nodes all used the highest zone. CMA/MOVABLE changed that slightly > by allowing the possibility of node0 having a "higher" zone than every > other node. When MOVABLE was introduced, it wasn't much of a problem as > the purpose of MOVABLE was for systems that dynamically needed to allocate > hugetlbfs later in the runtime but for CMA, it was a lot more critical > for ordinary usage so this is primarily a CMA thing. > > > I don't see a problem with your change. I would be worried about > > inflated reserves when e.g. ZONE_MOVABLE doesn't exist, but that doesn't > > seem to be the case. My laptop has empty ZONE_MOVABLE and the > > ZONE_NORMAL protection for movable is 0. > > > > But there had to be some reason for classzone_idx to be like this and > > not simple high_zoneidx. Maybe Mel remembers? Maybe it was important > > then, but is not anymore? Sigh, it seems to be pre-git. > > > > classzone predates my involvement with Linux but I would be less concerneed > about what the original intent was and instead ensure that classzone index > is consistent, sane and potentially renamed while preserving the intent of > "reserve pages in lower zones when an allocation request can use higher > zones". While historically the critical intent was to preserve Normal and > to a lesser extent DMA on 32-bit systems, there still should be some care > of DMA32 so we should not lose that. > > With the patch, the allocator looks like it would be fine as just > reservations change. I think it's unlikely that CMA usage will result > in lowmem starvation. Compaction becomes a bit weird as classzone index > has no special meaning versis highmem and I think it'll be very easy to > forget. Similarly, vmscan can reclaim pages from remote nodes and zones > that are higher than the original request. That is not likely to be a > problem but it's a change in behaviour and easy to miss. > > Fundamentally, I find it extremely weird we now have two variables that are > essentially the same thing. They should be collapsed into one variable, > renamed and documented on what the index means for page allocator, > compaction, vmscan and the special casing around CMA. You're all so young ;) classzone was Andrea. Perhaps he can shed some light upon the questions which have been raised?
On 05/08/2018 03:00 AM, Joonsoo Kim wrote: >> classzone predates my involvement with Linux but I would be less concerneed >> about what the original intent was and instead ensure that classzone index >> is consistent, sane and potentially renamed while preserving the intent of >> "reserve pages in lower zones when an allocation request can use higher >> zones". While historically the critical intent was to preserve Normal and >> to a lesser extent DMA on 32-bit systems, there still should be some care >> of DMA32 so we should not lose that. > > Agreed! > >> With the patch, the allocator looks like it would be fine as just >> reservations change. I think it's unlikely that CMA usage will result >> in lowmem starvation. Compaction becomes a bit weird as classzone index >> has no special meaning versis highmem and I think it'll be very easy to >> forget. I don't understand this point, what do you mean about highmem here? I've checked and compaction seems to use classzone_idx 1) to pass it to watermark checks as part of compaction suitability checks, i.e. the usual lowmem protection, and 2) to limit compaction of higher zones in kcompactd if the direct compactor can't use them anyway - seems this part has currently the same zone imbalance problem as reclaim. >> Similarly, vmscan can reclaim pages from remote nodes and zones >> that are higher than the original request. That is not likely to be a >> problem but it's a change in behaviour and easy to miss. >> >> Fundamentally, I find it extremely weird we now have two variables that are >> essentially the same thing. They should be collapsed into one variable, >> renamed and documented on what the index means for page allocator, >> compaction, vmscan and the special casing around CMA. > > Agreed! > I will update this patch to reflect your comment. If someone have an idea > on renaming this variable, please let me know. Pehaps max_zone_idx? Seems a bit more clear than "high_zoneidx". And I have no idea what was actually meant by "class". > Thanks. >
On Wed, May 16, 2018 at 11:35:55AM +0200, Vlastimil Babka wrote: > On 05/08/2018 03:00 AM, Joonsoo Kim wrote: > >> classzone predates my involvement with Linux but I would be less concerneed > >> about what the original intent was and instead ensure that classzone index > >> is consistent, sane and potentially renamed while preserving the intent of > >> "reserve pages in lower zones when an allocation request can use higher > >> zones". While historically the critical intent was to preserve Normal and > >> to a lesser extent DMA on 32-bit systems, there still should be some care > >> of DMA32 so we should not lose that. > > > > Agreed! > > > >> With the patch, the allocator looks like it would be fine as just > >> reservations change. I think it's unlikely that CMA usage will result > >> in lowmem starvation. Compaction becomes a bit weird as classzone index > >> has no special meaning versis highmem and I think it'll be very easy to > >> forget. > > I don't understand this point, what do you mean about highmem here? I mean it has no special meaning as compaction is not primarily concerned with lowmem protections as it compacts within a zone. It preserves watermarks but it does not have the same degree of criticality as the page allocator and reclaim is concerned with. > I've > checked and compaction seems to use classzone_idx 1) to pass it to > watermark checks as part of compaction suitability checks, i.e. the > usual lowmem protection, and 2) to limit compaction of higher zones in > kcompactd if the direct compactor can't use them anyway - seems this > part has currently the same zone imbalance problem as reclaim. > Originally the watermark check for compaction was primarily about not depleting a single zone but the checks were duplicated anyway. It's not actually super critical for it to preserve lowmem zones as any memory usage by compaction is transient. > > Agreed! > > I will update this patch to reflect your comment. If someone have an idea > > on renaming this variable, please let me know. > > Pehaps max_zone_idx? Seems a bit more clear than "high_zoneidx". And I > have no idea what was actually meant by "class". > I don't have a better suggestion.
diff --git a/mm/internal.h b/mm/internal.h index 228dd66..e1d7376 100644 --- a/mm/internal.h +++ b/mm/internal.h @@ -123,7 +123,7 @@ struct alloc_context { bool spread_dirty_pages; }; -#define ac_classzone_idx(ac) zonelist_zone_idx(ac->preferred_zoneref) +#define ac_classzone_idx(ac) (ac->high_zoneidx) /* * Locate the struct page for both the matching buddy in our