Fix memcg/memory.high in case kmem accounting is enabled

Hi,

Tejun reported that sometimes memcg/memory.high threshold seems to be silently ignored if kmem accounting is enabled:

http://www.spinics.net/lists/linux-mm/msg93613.html

It turned out that both SLAB and SLUB try to allocate without __GFP_WAIT first. As a result, if there is enough free pages, memcg reclaim will not get invoked on kmem allocations, which will lead to uncontrollable growth of memory usage no matter what memory.high is set to.

This patch set attempts to fix this issue. For more details please see comments to individual patches.

Thanks,

Vladimir Davydov (2):
mm/slab: skip memcg reclaim only if in atomic context mm/slub: do not bypass memcg reclaim for high-order page allocation

 mm/slab.c | 32 +++++++++++---------------------
 mm/slub.c | 24 +++++++++++-------------
 2 files changed, 22 insertions(+), 34 deletions(-)

Signature · 1 line

2.1.4

Right but isn't that what the caller explicitly asked for? Why should we ignore that for kmem accounting? It seems like a fix at a wrong layer to me. Either we should start failing GFP_NOWAIT charges when we are above high wmark or deploy an additional catchup mechanism as suggested by Tejun. I like the later more because it allows to better handle GFP_NOFS requests as well and there are many sources of these from kmem paths.

Quoted 6 lines · Michal Hocko

Right but isn't that what the caller explicitly asked for? Why should we ignore that for kmem accounting? It seems like a fix at a wrong layer to me. Either we should start failing GFP_NOWAIT charges when we are above high wmark or deploy an additional catchup mechanism as suggested by Tejun. I like the later more because it allows to better handle GFP_NOFS requests as well and there are many sources of these from kmem paths.

Yeah, this is beginning to look like we're trying to solve the problem at the wrong layer. slab/slub or whatever else should be able to use GFP_NOWAIT in whatever frequency they want for speculative allocations.

slab/slub can issue alloc_pages() any time with any flags they want and it won't be accounted to memcg, because kmem is accounted at slab/slub layer, not in buddy.

Quoted 3 lines · Vladimir Davydov

slab/slub can issue alloc_pages() any time with any flags they want and it won't be accounted to memcg, because kmem is accounted at slab/slub layer, not in buddy.

Hmmm? I meant the eventual calling into try_charge w/ GFP_NOWAIT. Speculative usage of GFP_NOWAIT is bound to increase and we don't want to put on extra restrictions from memcg side.

For memory.high,
punting to the return path is a pretty stright-forward solution which should make the problem go away almost entirely.

We have to be cautious about placing memcg_charge in slab/slub. To understand why, consider SLAB case, which first tries to allocate from all nodes in the order of preference w/o __GFP_WAIT and only if it fails falls back on an allocation from any node w/ __GFP_WAIT. This is its internal algorithm. If we blindly put memcg_charge to alloc_slab method, then, when we are near the memcg limit, we will go over all NUMA nodes in vain, then finally fall back to __GFP_WAIT allocation, which will get a slab from a random node. Not only we do more work than necessary due to walking over all NUMA nodes for nothing, but we also break SLAB internal logic! And you just can't fix it in memcg, because memcg knows nothing about the internal logic of SLAB, how it handles NUMA nodes.

SLUB has a different problem. It tries to avoid high-order allocations if there is a risk of invoking costly memory compactor. It has nothing to do with memcg, because memcg does not care if the charge is for a high order page or not.

Quoted 12 lines · Vladimir Davydov

We have to be cautious about placing memcg_charge in slab/slub. To understand why, consider SLAB case, which first tries to allocate from all nodes in the order of preference w/o __GFP_WAIT and only if it fails falls back on an allocation from any node w/ __GFP_WAIT. This is its internal algorithm. If we blindly put memcg_charge to alloc_slab method, then, when we are near the memcg limit, we will go over all NUMA nodes in vain, then finally fall back to __GFP_WAIT allocation, which will get a slab from a random node. Not only we do more work than necessary due to walking over all NUMA nodes for nothing, but we also break SLAB internal logic! And you just can't fix it in memcg, because memcg knows nothing about the internal logic of SLAB, how it handles NUMA nodes.

Quoted 4 lines

SLUB has a different problem. It tries to avoid high-order allocations if there is a risk of invoking costly memory compactor. It has nothing to do with memcg, because memcg does not care if the charge is for a high order page or not.

Maybe I'm missing something but aren't both issues caused by memcg failing to provide headroom for NOWAIT allocations when the consumption gets close to the max limit?

Regardless of the specific usage, !__GFP_WAIT means "give me memory if it can be spared w/o inducing direct time-consuming maintenance work" and the contract around it is that such requests will mostly succeed under nominal conditions. Also, slab/slub might not stay as the only user of try_charge().

I still think solving this from memcg side is the right direction.

If we want to allow slab/slub implementation to invoke try_charge wherever it wants, we need to introduce an asynchronous thread doing reclaim when a memcg is approaching its limit (or teach kswapd do that).

That's a way to go, but what's the point to complicate things prematurely while it seems we can fix the problem by using the technique similar to the one behind memory.high?

Nevertheless, even if we introduced such a thread, it'd be just insane to allow slab/slub blindly insert try_charge. Let me repeat the examples of SLAB/SLUB sub-optimal behavior caused by thoughtless usage of try_charge I gave above:

- memcg knows nothing about NUMA nodes, so what's the point in failing !__GFP_WAIT allocations used by SLAB while inspecting NUMA nodes?
- memcg knows nothing about high order pages, so what's the point in failing !__GFP_WAIT allocations used by SLUB to try to allocate a high order page?

On Mon, Aug 31, 2015 at 07:51:32PM +0300, Vladimir Davydov wrote:
...

Quoted 3 lines

If we want to allow slab/slub implementation to invoke try_charge wherever it wants, we need to introduce an asynchronous thread doing reclaim when a memcg is approaching its limit (or teach kswapd do that).

In the long term, I think this is the way to go.

Quoted 4 lines

That's a way to go, but what's the point to complicate things prematurely while it seems we can fix the problem by using the technique similar to the one behind memory.high?

Cuz we're now scattering workarounds to multiple places and I'm sure we'll add more try_charge() users (e.g. we want to fold in tcp memcg under the same knobs) and we'll have to worry about the same problem all over again and will inevitably miss some cases leading to subtle failures.

Quoted 11 lines

Nevertheless, even if we introduced such a thread, it'd be just insane to allow slab/slub blindly insert try_charge. Let me repeat the examples of SLAB/SLUB sub-optimal behavior caused by thoughtless usage of try_charge I gave above:

- memcg knows nothing about NUMA nodes, so what's the point in failing !__GFP_WAIT allocations used by SLAB while inspecting NUMA nodes?
- memcg knows nothing about high order pages, so what's the point in failing !__GFP_WAIT allocations used by SLUB to try to allocate a high order page?

Both are optimistic speculative actions and as long as memcg can guarantee that those requests will succeed under normal circumstances, as does the system-wide mm does, it isn't a problem.

In general, we want to make sure inside-cgroup behaviors as close to system-wide behaviors as possible, scoped but equivalent in kind. Doing things differently, while inevitable in certain cases, is likely to get messy in the long term.

I totally agree that we should strive to make a kmem user feel roughly the same in memcg as if it were running on a host with equal amount of RAM. There are two ways to achieve that:

1. Make the API functions, i.e. kmalloc and friends, behave inside memcg roughly the same way as they do in the root cgroup.
2. Make the internal memcg functions, i.e. try_charge and friends, behave roughly the same way as alloc_pages.

I find way 1 more flexible, because we don't have to blindly follow heuristics used on global memory reclaim and therefore have more opportunities to achieve the same goal.

Quoted 9 lines · Vladimir Davydov

I totally agree that we should strive to make a kmem user feel roughly the same in memcg as if it were running on a host with equal amount of RAM. There are two ways to achieve that:

1. Make the API functions, i.e. kmalloc and friends, behave inside memcg roughly the same way as they do in the root cgroup.
2. Make the internal memcg functions, i.e. try_charge and friends, behave roughly the same way as alloc_pages.

Quoted 3 lines

I find way 1 more flexible, because we don't have to blindly follow heuristics used on global memory reclaim and therefore have more opportunities to achieve the same goal.

The heuristics need to integrate well if its in a cgroup or not. In general make use of cgroups as transparent as possible to the rest of the code.

Quoted 3 lines · Vladimir Davydov

Tejun reported that sometimes memcg/memory.high threshold seems to be silently ignored if kmem accounting is enabled:

Quoted 6 lines

http://www.spinics.net/lists/linux-mm/msg93613.html

It turned out that both SLAB and SLUB try to allocate without __GFP_WAIT first. As a result, if there is enough free pages, memcg reclaim will not get invoked on kmem allocations, which will lead to uncontrollable growth of memory usage no matter what memory.high is set to.

Right but isn't that what the caller explicitly asked for?

Why should we ignore that for kmem accounting? It seems like a fix at a wrong layer to me.

Either we should start failing GFP_NOWAIT charges when we are above high wmark or deploy an additional catchup mechanism as suggested by Tejun.

I like the later more because it allows to better handle GFP_NOFS requests as well and there are many sources of these from kmem paths.

That being said, this is the fix at the right layer.

Hello, Vladimir.

On Mon, Aug 31, 2015 at 05:20:49PM +0300, Vladimir Davydov wrote:
...

Quoted 1 line

That being said, this is the fix at the right layer.

While this *might* be a necessary workaround for the hard limit case right now, this is by no means the fix at the right layer. The expectation is that mm keeps a reasonable amount of memory available for allocations which can't block. These allocations may fail from time to time depending on luck and under extreme memory pressure but the caller should be able to depend on it as a speculative allocation mechanism which doesn't fail willy-nilly.

Hardlimit breaking GFP_NOWAIT behavior is a bug on memcg side, not slab or slub.

Quoted 12 lines · Michal Hocko

Quoted 9 lines · Vladimir Davydov

Tejun reported that sometimes memcg/memory.high threshold seems to be silently ignored if kmem accounting is enabled:

http://www.spinics.net/lists/linux-mm/msg93613.html

It turned out that both SLAB and SLUB try to allocate without __GFP_WAIT first. As a result, if there is enough free pages, memcg reclaim will not get invoked on kmem allocations, which will lead to uncontrollable growth of memory usage no matter what memory.high is set to.

Quoted 1 line

Right but isn't that what the caller explicitly asked for?

No. If the caller of kmalloc() asked for a __GFP_WAIT allocation, we might ignore that and charge memcg w/o __GFP_WAIT.

Quoted 2 lines

Why should we ignore that for kmem accounting? It seems like a fix at a wrong layer to me.

Let's forget about memory.high for a minute.

1. SLAB. Suppose someone calls kmalloc_node and there is enough free memory on the preferred node. W/o memcg limit set, the allocation will happen from the preferred node, which is OK. If there is memcg limit, we can currently fail to allocate from the preferred node if we are near the limit. We issue memcg reclaim and go to fallback alloc then, which will most probably allocate from a different node, although there is no reason for that. This is a bug.

2. SLUB. Someone calls kmalloc and there is enough free high order pages. If there is no memcg limit, we will allocate a high order slab page, which is in accordance with SLUB internal logic. With memcg limit set, we are likely to fail to charge high order page (because we currently try to charge high order pages w/o __GFP_WAIT) and fallback on a low order page. The latter is unexpected and unjustified.

That being said, this is the fix at the right layer.

Quoted 3 lines

Either we should start failing GFP_NOWAIT charges when we are above high wmark or deploy an additional catchup mechanism as suggested by Tejun.

The mechanism proposed by Tejun won't help us to avoid allocation failures if we are hitting memory.max w/o __GFP_WAIT or __GFP_FS.

To fix GFP_NOFS/GFP_NOWAIT failures we just need to start reclaim when the gap between limit and usage is getting too small. It may be done from a workqueue or from task_work, but currently I don't see any reason why complicate and not just start reclaim directly, just like memory.high does.

I mean, currently you can protect against GFP_NOWAIT failures by setting memory.high to be 1-2 MB lower than memory.high and this *will* work, because GFP_NOWAIT/GFP_NOFS allocations can't go on infinitely - they will alternate with normal GFP_KERNEL allocations sooner or later. It does not mean we should encourage users to set memory.high to protect against such failures, because, as pointed out by Tejun, logic behind memory.high is currently opaque and can change, but we can introduce memcg-internal watermarks that would work exactly as memory.high and hence help us against GFP_NOWAIT/GFP_NOFS failures.

Quoted 13 lines · Vladimir Davydov

Quoted 12 lines · Michal Hocko

Quoted 9 lines · Vladimir Davydov

Tejun reported that sometimes memcg/memory.high threshold seems to be silently ignored if kmem accounting is enabled:

http://www.spinics.net/lists/linux-mm/msg93613.html

It turned out that both SLAB and SLUB try to allocate without __GFP_WAIT first. As a result, if there is enough free pages, memcg reclaim will not get invoked on kmem allocations, which will lead to uncontrollable growth of memory usage no matter what memory.high is set to.

Quoted 4 lines

Quoted 1 line

Right but isn't that what the caller explicitly asked for?

No. If the caller of kmalloc() asked for a __GFP_WAIT allocation, we might ignore that and charge memcg w/o __GFP_WAIT.

I was referring to the slab allocator as the caller. Sorry for not being clear about that.

Quoted 5 lines

Quoted 2 lines

Why should we ignore that for kmem accounting? It seems like a fix at a wrong layer to me.

Let's forget about memory.high for a minute.

Quoted 7 lines

1. SLAB. Suppose someone calls kmalloc_node and there is enough free memory on the preferred node. W/o memcg limit set, the allocation will happen from the preferred node, which is OK. If there is memcg limit, we can currently fail to allocate from the preferred node if we are near the limit. We issue memcg reclaim and go to fallback alloc then, which will most probably allocate from a different node, although there is no reason for that. This is a bug.

I am not familiar with the SLAB internals much but how is it different from the global case. If the preferred node is full then __GFP_THISNODE request will make it fail early even without giving GFP_NOWAIT additional access to atomic memory reserves. The fact that memcg case fails earlier is perfectly expected because the restriction is tighter than the global case.

How the fallback is implemented and whether trying other node before reclaiming from the preferred one is reasonable I dunno. This is for SLAB to decide. But ignoring GFP_NOWAIT for this path makes the behavior for memcg enabled setups subtly different. And that is bad.

Quoted 8 lines

2. SLUB. Someone calls kmalloc and there is enough free high order pages. If there is no memcg limit, we will allocate a high order slab page, which is in accordance with SLUB internal logic. With memcg limit set, we are likely to fail to charge high order page (because we currently try to charge high order pages w/o __GFP_WAIT) and fallback on a low order page. The latter is unexpected and unjustified.

And this case very similar and I even argue that it shows more brokenness with your patch. The SLUB allocator has _explicitly_ asked for an allocation _without_ reclaim because that would be unnecessarily too costly and there is other less expensive fallback. But memcg would

be ignoring this with your patch AFAIU and break the optimization. There are other cases like that. E.g. THP pages are allocated without GFP_WAIT when defrag is disabled.

Quoted 7 lines

That being said, this is the fix at the right layer.

Quoted 3 lines

Either we should start failing GFP_NOWAIT charges when we are above high wmark or deploy an additional catchup mechanism as suggested by Tejun.

Quoted 2 lines

The mechanism proposed by Tejun won't help us to avoid allocation failures if we are hitting memory.max w/o __GFP_WAIT or __GFP_FS.

Why would be that a problem. The _hard_ limit is reached and reclaim cannot make any progress. An allocation failure is to be expected. GFP_NOWAIT will fail normally and GFP_NOFS will attempt to reclaim before failing.

Quoted 6 lines

To fix GFP_NOFS/GFP_NOWAIT failures we just need to start reclaim when the gap between limit and usage is getting too small. It may be done from a workqueue or from task_work, but currently I don't see any reason why complicate and not just start reclaim directly, just like memory.high does.

Yes we can do better than we do right now. But that doesn't mean we should put hacks all over the place and lie about the allocation context.

Quoted 10 lines

I mean, currently you can protect against GFP_NOWAIT failures by setting memory.high to be 1-2 MB lower than memory.high and this *will* work, because GFP_NOWAIT/GFP_NOFS allocations can't go on infinitely - they will alternate with normal GFP_KERNEL allocations sooner or later. It does not mean we should encourage users to set memory.high to protect against such failures, because, as pointed out by Tejun, logic behind memory.high is currently opaque and can change, but we can introduce memcg-internal watermarks that would work exactly as memory.high and hence help us against GFP_NOWAIT/GFP_NOFS failures.

I am not against something like watermarks and doing more pro-active reclaim but this is far from easy to do - which is one of the reason we do not have it yet. The idea from Tejun about the return to userspace reclaim is nice in that regards that it happens from a well defined context and helps to keep memory.high behavior much saner.

Quoted 1 line · Michal Hocko

On Mon 31-08-15 17:20:49, Vladimir Davydov wrote:

Quoted 8 lines

Quoted 7 lines

1. SLAB. Suppose someone calls kmalloc_node and there is enough free memory on the preferred node. W/o memcg limit set, the allocation will happen from the preferred node, which is OK. If there is memcg limit, we can currently fail to allocate from the preferred node if we are near the limit. We issue memcg reclaim and go to fallback alloc then, which will most probably allocate from a different node, although there is no reason for that. This is a bug.

Quoted 6 lines

I am not familiar with the SLAB internals much but how is it different from the global case. If the preferred node is full then __GFP_THISNODE request will make it fail early even without giving GFP_NOWAIT additional access to atomic memory reserves. The fact that memcg case fails earlier is perfectly expected because the restriction is tighter than the global case.

memcg restrictions are orthogonal to NUMA: failing an allocation from a particular node does not mean failing memcg charge and vice versa.

Quoted 4 lines

How the fallback is implemented and whether trying other node before reclaiming from the preferred one is reasonable I dunno. This is for SLAB to decide. But ignoring GFP_NOWAIT for this path makes the behavior for memcg enabled setups subtly different. And that is bad.

Quite the contrary. Trying to charge memcg w/o __GFP_WAIT while inspecting if a NUMA node has free pages makes SLAB behaviour subtly
differently: SLAB will walk over all NUMA nodes for nothing instead of invoking memcg reclaim once a free page is found.

You are talking about memcg/kmem accounting as if it were done in the buddy allocator on top of which the slab layer is built knowing nothing about memcg accounting on the lower layer. That's not true and that simply can't be true. Kmem accounting is implemented at the slab layer. Memcg provides its memcg_charge_slab/uncharge methods solely for slab core, so it's OK to have some calling conventions between them. What we are really obliged to do is to preserve behavior of slab's external API, i.e. kmalloc and friends.

Quoted 8 lines

Quoted 7 lines

2. SLUB. Someone calls kmalloc and there is enough free high order pages. If there is no memcg limit, we will allocate a high order slab page, which is in accordance with SLUB internal logic. With memcg limit set, we are likely to fail to charge high order page (because we currently try to charge high order pages w/o __GFP_WAIT) and fallback on a low order page. The latter is unexpected and unjustified.

Quoted 4 lines

And this case very similar and I even argue that it shows more brokenness with your patch. The SLUB allocator has _explicitly_ asked for an allocation _without_ reclaim because that would be unnecessarily too costly and there is other less expensive fallback. But memcg would

You are ignoring the fact that, in contrast to alloc_pages, for memcg there is practically no difference between charging a 4-order page or a 1-order page.

OTOH, using 1-order pages where we could go with 4-order pages increases page fragmentation at the global level. This subtly breaks internal SLUB optimization. Once again, kmem accounting is not something staying aside from slab core, it's a part of slab core.

Quoted 3 lines

be ignoring this with your patch AFAIU and break the optimization. There are other cases like that. E.g. THP pages are allocated without GFP_WAIT when defrag is disabled.

It might be wrong. If we can't find a continuous 2Mb page, we should probably give up instead of calling compactor. For memcg it might be better to reclaim some space for 2Mb page right now and map a 2Mb page instead of reclaiming space for 512 4Kb pages a moment later, because in memcg case there is absolutely no difference between reclaiming 2Mb for a huge page and 2Mb for 512 4Kb pages.

Quoted 6 lines

Quoted 6 lines

That being said, this is the fix at the right layer.

Quoted 3 lines

Either we should start failing GFP_NOWAIT charges when we are above high wmark or deploy an additional catchup mechanism as suggested by Tejun.

Quoted 7 lines

Quoted 2 lines

The mechanism proposed by Tejun won't help us to avoid allocation failures if we are hitting memory.max w/o __GFP_WAIT or __GFP_FS.

Why would be that a problem. The _hard_ limit is reached and reclaim cannot make any progress. An allocation failure is to be expected. GFP_NOWAIT will fail normally and GFP_NOFS will attempt to reclaim before failing.

Quoting my e-mail to Tejun explaining why using task_work won't help if we don't fix SLAB/SLUB:

: Generally speaking, handing over reclaim responsibility to task_work : won't help, because there might be cases when a process spends quite a : lot of time in kernel invoking lots of GFP_KERNEL allocations before : returning to userspace. Without fixing slab/slub, such a process will : charge w/o __GFP_WAIT and therefore can exceed memory.high and reach : memory.max. If there are no other active processes in the cgroup, the : cgroup can stay with memory.high excess for a relatively long time : (suppose the process was throttled in kernel), possibly hurting the rest : of the system. What is worse, if the process happens to invoke a real : GFP_NOWAIT allocation when it's about to hit the limit, it will fail.

For a kmalloc user that's completely unexpected.

Quoted 6 lines

Quoted 5 lines

To fix GFP_NOFS/GFP_NOWAIT failures we just need to start reclaim when the gap between limit and usage is getting too small. It may be done from a workqueue or from task_work, but currently I don't see any reason why complicate and not just start reclaim directly, just like memory.high does.

Quoted 3 lines

Yes we can do better than we do right now. But that doesn't mean we should put hacks all over the place and lie about the allocation context.

What do you mean by saying "all over the place"? It's a fix for kmem implementation, to be more exact for the part of it residing in the slab core.

Everyone else, except a couple of kmem users issuing alloc_page directly like threadinfo, will use kmalloc and know nothing what's going on there and how all this accounting stuff is handled - they will just use plain old convenient kmalloc, which works exactly as it does in the root cgroup.

Quoted 2 lines · Vladimir Davydov

Quoted 1 line · Michal Hocko

On Mon 31-08-15 17:20:49, Vladimir Davydov wrote:

{...}

Quoted 15 lines

Quoted 8 lines

Quoted 7 lines

1. SLAB. Suppose someone calls kmalloc_node and there is enough free memory on the preferred node. W/o memcg limit set, the allocation will happen from the preferred node, which is OK. If there is memcg limit, we can currently fail to allocate from the preferred node if we are near the limit. We issue memcg reclaim and go to fallback alloc then, which will most probably allocate from a different node, although there is no reason for that. This is a bug.

Quoted 6 lines

I am not familiar with the SLAB internals much but how is it different from the global case. If the preferred node is full then __GFP_THISNODE request will make it fail early even without giving GFP_NOWAIT additional access to atomic memory reserves. The fact that memcg case fails earlier is perfectly expected because the restriction is tighter than the global case.

Quoted 2 lines

memcg restrictions are orthogonal to NUMA: failing an allocation from a particular node does not mean failing memcg charge and vice versa.

Sure memcg doesn't care about NUMA it just puts an additional constrain on top of all existing ones. The point I've tried to make is that the logic is currently same whether it is page allocator (with the node restriction) or memcg (cumulative amount restriction) are behaving consistently. Neither of them try to reclaim in order to achieve its goals. How conservative is memcg about allowing GFP_NOWAIT allocation is a separate issue and all those details belong to memcg proper same as the allocation strategy for these allocations belongs to the page allocator.

Quoted 5 lines

Quoted 4 lines

How the fallback is implemented and whether trying other node before reclaiming from the preferred one is reasonable I dunno. This is for SLAB to decide. But ignoring GFP_NOWAIT for this path makes the behavior for memcg enabled setups subtly different. And that is bad.

Quoted 4 lines

Quite the contrary. Trying to charge memcg w/o __GFP_WAIT while inspecting if a NUMA node has free pages makes SLAB behaviour subtly
differently: SLAB will walk over all NUMA nodes for nothing instead of invoking memcg reclaim once a free page is found.

So you are saying that the SLAB kmem accounting in this particular path is suboptimal because the fallback mode doesn't retry local node with the reclaim enabled before falling back to other nodes?

I would consider it quite surprising as well even for the global case because __GFP_THISNODE doesn't wake up kswapd to make room on that node.

Quoted 8 lines

You are talking about memcg/kmem accounting as if it were done in the buddy allocator on top of which the slab layer is built knowing nothing about memcg accounting on the lower layer. That's not true and that simply can't be true. Kmem accounting is implemented at the slab layer. Memcg provides its memcg_charge_slab/uncharge methods solely for slab core, so it's OK to have some calling conventions between them. What we are really obliged to do is to preserve behavior of slab's external API, i.e. kmalloc and friends.

I guess I understand what you are saying here but it sounds like special casing which tries to be clever because the current code understands both the lower level allocator and kmem charge paths to decide how to

juggle with them. This is imho bad and hard to maintain long term.

Quoted 14 lines

Quoted 13 lines

Quoted 8 lines

2. SLUB. Someone calls kmalloc and there is enough free high order pages. If there is no memcg limit, we will allocate a high order slab page, which is in accordance with SLUB internal logic. With memcg limit set, we are likely to fail to charge high order page (because we currently try to charge high order pages w/o __GFP_WAIT) and fallback on a low order page. The latter is unexpected and unjustified.

And this case very similar and I even argue that it shows more brokenness with your patch. The SLUB allocator has _explicitly_ asked for an allocation _without_ reclaim because that would be unnecessarily too costly and there is other less expensive fallback. But memcg would

Quoted 3 lines

You are ignoring the fact that, in contrast to alloc_pages, for memcg there is practically no difference between charging a 4-order page or a 1-order page.

But this is an implementation details which might change anytime in future.

Quoted 5 lines

OTOH, using 1-order pages where we could go with 4-order pages increases page fragmentation at the global level. This subtly breaks internal SLUB optimization. Once again, kmem accounting is not something staying aside from slab core, it's a part of slab core.

This is certainly true and it is what you get when you put an additional constrain on top of an existing one. You simply cannot get both the great performance _and_ a local memory restriction.

Quoted 11 lines

Quoted 4 lines

be ignoring this with your patch AFAIU and break the optimization. There are other cases like that. E.g. THP pages are allocated without GFP_WAIT when defrag is disabled.

It might be wrong. If we can't find a continuous 2Mb page, we should probably give up instead of calling compactor. For memcg it might be better to reclaim some space for 2Mb page right now and map a 2Mb page instead of reclaiming space for 512 4Kb pages a moment later, because in memcg case there is absolutely no difference between reclaiming 2Mb for a huge page and 2Mb for 512 4Kb pages.

Or maybe the whole reclaim just doesn't pay off because the TLB savings will never compensate for the reclaim. The defrag knob basically says that we shouldn't try to opportunistically prepare a room for the THP page.

Quoted 29 lines

Quoted 7 lines

Quoted 7 lines

That being said, this is the fix at the right layer.

Quoted 3 lines

Either we should start failing GFP_NOWAIT charges when we are above high wmark or deploy an additional catchup mechanism as suggested by Tejun.

Quoted 7 lines

Quoted 2 lines

The mechanism proposed by Tejun won't help us to avoid allocation failures if we are hitting memory.max w/o __GFP_WAIT or __GFP_FS.

Why would be that a problem. The _hard_ limit is reached and reclaim cannot make any progress. An allocation failure is to be expected. GFP_NOWAIT will fail normally and GFP_NOFS will attempt to reclaim before failing.

Quoting my e-mail to Tejun explaining why using task_work won't help if we don't fix SLAB/SLUB:

: Generally speaking, handing over reclaim responsibility to task_work : won't help, because there might be cases when a process spends quite a : lot of time in kernel invoking lots of GFP_KERNEL allocations before : returning to userspace. Without fixing slab/slub, such a process will : charge w/o __GFP_WAIT and therefore can exceed memory.high and reach : memory.max. If there are no other active processes in the cgroup, the : cgroup can stay with memory.high excess for a relatively long time : (suppose the process was throttled in kernel), possibly hurting the rest : of the system. What is worse, if the process happens to invoke a real : GFP_NOWAIT allocation when it's about to hit the limit, it will fail.

Quoted 1 line

For a kmalloc user that's completely unexpected.

We have the global reclaim which handles the global memory pressure. And until the hard limit is enforced I do not see what is the huge problem here. Sure we can have high limit in excess but that is to be expected.

Same as failing allocations for the hard limit enforcement.

Maybe moving whole high limit reclaim to the delayed context is not what we will end up with and reduce this only for GFP_NOWAIT or other weak reclaim contexts. This is to be discussed of course.

Quoted 11 lines

Quoted 10 lines

Quoted 6 lines

To fix GFP_NOFS/GFP_NOWAIT failures we just need to start reclaim when the gap between limit and usage is getting too small. It may be done from a workqueue or from task_work, but currently I don't see any reason why complicate and not just start reclaim directly, just like memory.high does.

Yes we can do better than we do right now. But that doesn't mean we should put hacks all over the place and lie about the allocation context.

Quoted 3 lines

What do you mean by saying "all over the place"? It's a fix for kmem implementation, to be more exact for the part of it residing in the slab core.

I meant into two slab allocators currently because of the implementation details which are spread into three different places - page allocator, memcg charging code and the respective slab allocator specific details.

Quoted 6 lines

Everyone else, except a couple of kmem users issuing alloc_page directly like threadinfo, will use kmalloc and know nothing what's going on there and how all this accounting stuff is handled - they will just use plain old convenient kmalloc, which works exactly as it does in the root cgroup.

If we ever grow more users and charge more kernel memory then they might be doing similar assumptions and tweak allocation/charge context and we would end up in a bigger mess. It makes much more sense to have allocation and charge context consistent.

Quoted 2 lines · Michal Hocko

Quoted 1 line · Vladimir Davydov

On Tue, Sep 01, 2015 at 02:36:12PM +0200, Michal Hocko wrote:

Quoted 10 lines

Quoted 9 lines

Quoted 4 lines

How the fallback is implemented and whether trying other node before reclaiming from the preferred one is reasonable I dunno. This is for SLAB to decide. But ignoring GFP_NOWAIT for this path makes the behavior for memcg enabled setups subtly different. And that is bad.

Quite the contrary. Trying to charge memcg w/o __GFP_WAIT while inspecting if a NUMA node has free pages makes SLAB behaviour subtly
differently: SLAB will walk over all NUMA nodes for nothing instead of invoking memcg reclaim once a free page is found.

Quoted 3 lines

So you are saying that the SLAB kmem accounting in this particular path is suboptimal because the fallback mode doesn't retry local node with the reclaim enabled before falling back to other nodes?

I'm just pointing out some subtle behavior changes in slab you were opposed to.

Quoted 12 lines

I would consider it quite surprising as well even for the global case because __GFP_THISNODE doesn't wake up kswapd to make room on that node.

Quoted 8 lines

You are talking about memcg/kmem accounting as if it were done in the buddy allocator on top of which the slab layer is built knowing nothing about memcg accounting on the lower layer. That's not true and that simply can't be true. Kmem accounting is implemented at the slab layer. Memcg provides its memcg_charge_slab/uncharge methods solely for slab core, so it's OK to have some calling conventions between them. What we are really obliged to do is to preserve behavior of slab's external API, i.e. kmalloc and friends.

Quoted 3 lines

I guess I understand what you are saying here but it sounds like special casing which tries to be clever because the current code understands both the lower level allocator and kmem charge paths to decide how to

What do you mean by saying "it understands the lower level allocator"?

AFAIK we have memcg callbacks only in special places, like page fault handler or kmalloc.

Quoted 1 line

juggle with them. This is imho bad and hard to maintain long term.

We already juggle. Just grep where and how we insert mem_cgroup_try_charge.

Quoted 17 lines

Quoted 16 lines

Quoted 8 lines

Quoted 7 lines

2. SLUB. Someone calls kmalloc and there is enough free high order pages. If there is no memcg limit, we will allocate a high order slab page, which is in accordance with SLUB internal logic. With memcg limit set, we are likely to fail to charge high order page (because we currently try to charge high order pages w/o __GFP_WAIT) and fallback on a low order page. The latter is unexpected and unjustified.

Quoted 4 lines

And this case very similar and I even argue that it shows more brokenness with your patch. The SLUB allocator has _explicitly_ asked for an allocation _without_ reclaim because that would be unnecessarily too costly and there is other less expensive fallback. But memcg would

You are ignoring the fact that, in contrast to alloc_pages, for memcg there is practically no difference between charging a 4-order page or a 1-order page.

Quoted 2 lines

But this is an implementation details which might change anytime in future.

The fact that memcg reclaim does not invoke compactor is indeed an implementation detail, but how can it change?

Quoted 8 lines

Quoted 4 lines

OTOH, using 1-order pages where we could go with 4-order pages increases page fragmentation at the global level. This subtly breaks internal SLUB optimization. Once again, kmem accounting is not something staying aside from slab core, it's a part of slab core.

This is certainly true and it is what you get when you put an additional constrain on top of an existing one. You simply cannot get both the great performance _and_ a local memory restriction.

So what? We shouldn't even try?

Quoted 11 lines

Quoted 10 lines

Quoted 3 lines

be ignoring this with your patch AFAIU and break the optimization. There are other cases like that. E.g. THP pages are allocated without GFP_WAIT when defrag is disabled.

It might be wrong. If we can't find a continuous 2Mb page, we should probably give up instead of calling compactor. For memcg it might be better to reclaim some space for 2Mb page right now and map a 2Mb page instead of reclaiming space for 512 4Kb pages a moment later, because in memcg case there is absolutely no difference between reclaiming 2Mb for a huge page and 2Mb for 512 4Kb pages.

Quoted 4 lines

Or maybe the whole reclaim just doesn't pay off because the TLB savings will never compensate for the reclaim. The defrag knob basically says that we shouldn't try to opportunistically prepare a room for the THP page.

And why is it called "defrag" then?

Quoted 14 lines

Quoted 14 lines

Quoting my e-mail to Tejun explaining why using task_work won't help if we don't fix SLAB/SLUB:

: Generally speaking, handing over reclaim responsibility to task_work : won't help, because there might be cases when a process spends quite a : lot of time in kernel invoking lots of GFP_KERNEL allocations before : returning to userspace. Without fixing slab/slub, such a process will : charge w/o __GFP_WAIT and therefore can exceed memory.high and reach : memory.max. If there are no other active processes in the cgroup, the : cgroup can stay with memory.high excess for a relatively long time : (suppose the process was throttled in kernel), possibly hurting the rest : of the system. What is worse, if the process happens to invoke a real : GFP_NOWAIT allocation when it's about to hit the limit, it will fail.

Quoted 5 lines

Quoted 1 line

For a kmalloc user that's completely unexpected.

We have the global reclaim which handles the global memory pressure. And until the hard limit is enforced I do not see what is the huge problem here. Sure we can have high limit in excess but that is to be expected.

What exactly is to be expected? Is it OK if memory.high is just ignored?

Quoted 1 line

Same as failing allocations for the hard limit enforcement.

If a kmem allocation fails, your app is likely to fail too. Nobody expects write/read fail with ENOMEM if there seems to be enough reclaimable memory. If we try to fix the GFP_NOWAIT problem only by using task_work reclaim, it won't be a complete fix, because a failure may still occur as I described above.

Quoted 3 lines

Maybe moving whole high limit reclaim to the delayed context is not what we will end up with and reduce this only for GFP_NOWAIT or other weak reclaim contexts. This is to be discussed of course.

Yeah, but w/o fixing kmalloc it may happen that *every* allocation will be GFP_NOWAIT. It'd complicate the implementation.

Quoted 4 lines

Quoted 3 lines

What do you mean by saying "all over the place"? It's a fix for kmem implementation, to be more exact for the part of it residing in the slab core.

Quoted 3 lines

I meant into two slab allocators currently because of the implementation details which are spread into three different places - page allocator, memcg charging code and the respective slab allocator specific details.

If we remove kmem accounting, we will still have implementation details spread over page allocator, reclaimer, rmap, memcg. Slab is not the worst part of it IMO. Anyway, kmem accounting can't be implemented solely in memcg.

Quoted 6 lines

Quoted 5 lines

Everyone else, except a couple of kmem users issuing alloc_page directly like threadinfo, will use kmalloc and know nothing what's going on there and how all this accounting stuff is handled - they will just use plain old convenient kmalloc, which works exactly as it does in the root cgroup.

Quoted 4 lines

If we ever grow more users and charge more kernel memory then they might be doing similar assumptions and tweak allocation/charge context and we would end up in a bigger mess. It makes much more sense to have allocation and charge context consistent.

What new users? Why can't they just call kmalloc?

Quoted 3 lines · Vladimir Davydov

Quoted 2 lines · Michal Hocko

Quoted 1 line · Vladimir Davydov

On Tue, Sep 01, 2015 at 02:36:12PM +0200, Michal Hocko wrote:

[...]

Quoted 10 lines

Quoted 10 lines

Quoted 9 lines

Quoted 4 lines

How the fallback is implemented and whether trying other node before reclaiming from the preferred one is reasonable I dunno. This is for SLAB to decide. But ignoring GFP_NOWAIT for this path makes the behavior for memcg enabled setups subtly different. And that is bad.

Quite the contrary. Trying to charge memcg w/o __GFP_WAIT while inspecting if a NUMA node has free pages makes SLAB behaviour subtly
differently: SLAB will walk over all NUMA nodes for nothing instead of invoking memcg reclaim once a free page is found.

Quoted 6 lines

Quoted 3 lines

So you are saying that the SLAB kmem accounting in this particular path is suboptimal because the fallback mode doesn't retry local node with the reclaim enabled before falling back to other nodes?

I'm just pointing out some subtle behavior changes in slab you were opposed to.

I guess we are still not at the same page here. If the slab has a subtle behavior (and from what you are saying it seems it has the same behavior at the global scope) then we should strive to fix it rather than making it more obscure just to not expose GFP_NOWAIT to memcg which is not handled properly currently wrt. high limit (more on that below) which was the primary motivation for the patch AFAIU.

Quoted 12 lines

Quoted 12 lines

I would consider it quite surprising as well even for the global case because __GFP_THISNODE doesn't wake up kswapd to make room on that node.

Quoted 7 lines

You are talking about memcg/kmem accounting as if it were done in the > > > > buddy allocator on top of which the slab layer is built knowing nothing about memcg accounting on the lower layer. That's not true and that
simply can't be true. Kmem accounting is implemented at the slab layer.
Memcg provides its memcg_charge_slab/uncharge methods solely for
slab core, so it's OK to have some calling conventions between them.
What we are really obliged to do is to preserve behavior of slab's
external API, i.e. kmalloc and friends.

Quoted 5 lines

Quoted 3 lines

I guess I understand what you are saying here but it sounds like special casing which tries to be clever because the current code understands both the lower level allocator and kmem charge paths to decide how to

What do you mean by saying "it understands the lower level allocator"?

I mean it requires/abuses special behavior from the page allocator like __GFP_THISNODE && !wait for the hot path.

Quoted 2 lines

AFAIK we have memcg callbacks only in special places, like page fault handler or kmalloc.

But anybody might opt-in to be charged. I can see some other buffers which are even not accounted for right now will be charged in future.

Quoted 4 lines

Quoted 1 line

juggle with them. This is imho bad and hard to maintain long term.

We already juggle. Just grep where and how we insert mem_cgroup_try_charge.

We should always preserve the gfp context (at least its reclaim part). If we are not then it is a bug.

Quoted 17 lines

Quoted 17 lines

Quoted 16 lines

Quoted 8 lines

Quoted 7 lines

2. SLUB. Someone calls kmalloc and there is enough free high order pages. If there is no memcg limit, we will allocate a high order slab page, which is in accordance with SLUB internal logic. With memcg limit set, we are likely to fail to charge high order page (because we currently try to charge high order pages w/o __GFP_WAIT) and fallback on a low order page. The latter is unexpected and unjustified.

Quoted 4 lines

And this case very similar and I even argue that it shows more brokenness with your patch. The SLUB allocator has _explicitly_ asked for an allocation _without_ reclaim because that would be unnecessarily too costly and there is other less expensive fallback. But memcg would

You are ignoring the fact that, in contrast to alloc_pages, for memcg there is practically no difference between charging a 4-order page or a 1-order page.

Quoted 5 lines

Quoted 2 lines

But this is an implementation details which might change anytime in future.

The fact that memcg reclaim does not invoke compactor is indeed an implementation detail, but how can it change?

Compaction is indeed not something memcg reclaim cares about right now or will care in foreseeable future. I meant something else. order-1 vs. ordern-N differ in the reclaim target which then controls the potential latency of the reclaim. The fact that order-1 and order-4 do not really make any difference _right now_ because of the large SWAP_CLUSTER_MAX is the implementation detail I was referring to.

Quoted 10 lines

Quoted 8 lines

Quoted 4 lines

OTOH, using 1-order pages where we could go with 4-order pages increases page fragmentation at the global level. This subtly breaks internal SLUB optimization. Once again, kmem accounting is not something staying aside from slab core, it's a part of slab core.

This is certainly true and it is what you get when you put an additional constrain on top of an existing one. You simply cannot get both the great performance _and_ a local memory restriction.

So what? We shouldn't even try?

Of course you can try. Then the question is what are costs/benefits (both performance and maintainability). I didn't say those two patches are incorrect (the original kmalloc gfp mask is obeyed). They just seem targeting a wrong layer IMO. Alternative solutions were not attempted and measured for typical workloads. If we find out that addressing GFP_NOWAIT at memcg level will be viable for most reasonable loads and corner cases are at least not causing runaways which would be hard to address then let's put workarounds where they are necessary.

Quoted 11 lines

Quoted 11 lines

Quoted 10 lines

Quoted 3 lines

be ignoring this with your patch AFAIU and break the optimization. There are other cases like that. E.g. THP pages are allocated without GFP_WAIT when defrag is disabled.

It might be wrong. If we can't find a continuous 2Mb page, we should probably give up instead of calling compactor. For memcg it might be better to reclaim some space for 2Mb page right now and map a 2Mb page instead of reclaiming space for 512 4Kb pages a moment later, because in memcg case there is absolutely no difference between reclaiming 2Mb for a huge page and 2Mb for 512 4Kb pages.

Quoted 6 lines

Quoted 4 lines

Or maybe the whole reclaim just doesn't pay off because the TLB savings will never compensate for the reclaim. The defrag knob basically says that we shouldn't try to opportunistically prepare a room for the THP page.

And why is it called "defrag" then?

Do not ask me about the naming. If this was only about compaction then the allocator might be told about that by a special GFP flag. The memcg could be in line with that. But the point remains. If the defrag is a knob to make the page fault THP path lighter then no memcg reclaim is a reasonable to do.

[...]

Quoted 14 lines

Quoted 14 lines

Quoted 14 lines

Quoting my e-mail to Tejun explaining why using task_work won't help if we don't fix SLAB/SLUB:

: Generally speaking, handing over reclaim responsibility to task_work : won't help, because there might be cases when a process spends quite a : lot of time in kernel invoking lots of GFP_KERNEL allocations before : returning to userspace. Without fixing slab/slub, such a process will : charge w/o __GFP_WAIT and therefore can exceed memory.high and reach : memory.max. If there are no other active processes in the cgroup, the : cgroup can stay with memory.high excess for a relatively long time : (suppose the process was throttled in kernel), possibly hurting the rest : of the system. What is worse, if the process happens to invoke a real : GFP_NOWAIT allocation when it's about to hit the limit, it will fail.

Quoted 7 lines

Quoted 5 lines

Quoted 1 line

For a kmalloc user that's completely unexpected.

We have the global reclaim which handles the global memory pressure. And until the hard limit is enforced I do not see what is the huge problem here. Sure we can have high limit in excess but that is to be expected.

What exactly is to be expected? Is it OK if memory.high is just ignored?

It is not OK to be ignored altogether. The high limit is where the throttling should start. And we currently do not handle GFP_NOWAIT which is something to be solved. We shouldn't remove GFP_NOWAIT callers as a workaround.

There are more things to do here. We can perform the reclaim from the delayed context where the direct reclaim is not allowed/requested. And we can start failing GFP_NOWAIT on an excessive high limit breach when the delayed reclaim doesn't catch up with the demand. This is basically what we do on the global level.
If even this is not sufficient and the kernel allows for a lot of allocations in the single run, which would be something to look at in the first place, then we have global mechanisms to mitigate that.

memory.high is an opportunistic memory isolation. It doesn't guarantee a complete isolation. The hard limit is for that purpose.

Quoted 7 lines

Quoted 1 line

Same as failing allocations for the hard limit enforcement.

If a kmem allocation fails, your app is likely to fail too. Nobody expects write/read fail with ENOMEM if there seems to be enough reclaimable memory. If we try to fix the GFP_NOWAIT problem only by using task_work reclaim, it won't be a complete fix, because a failure may still occur as I described above.

You cannot have a system which cannot tolerate failures and require memory restrictions. These two requirements simply go against each other. Moreover GPF_NOWAIT context is really light and should always have a fallback mode otherwise you get what you are saying - failures with reclaimable memory. And this is very much the case for the global case as well.

Quoted 6 lines

Quoted 3 lines

Maybe moving whole high limit reclaim to the delayed context is not what we will end up with and reduce this only for GFP_NOWAIT or other weak reclaim contexts. This is to be discussed of course.

Yeah, but w/o fixing kmalloc it may happen that *every* allocation will be GFP_NOWAIT. It'd complicate the implementation.

OK, but that is the case for the global case already. MM resp. memcg has to say at when to stop it. The global case handles that at the page allocator layer and memcg should do something similar at the charge level.

[...]

Quoted 4 lines

Quoted 4 lines

Quoted 3 lines

What do you mean by saying "all over the place"? It's a fix for kmem implementation, to be more exact for the part of it residing in the slab core.

Quoted 8 lines

Quoted 3 lines

I meant into two slab allocators currently because of the implementation details which are spread into three different places - page allocator, memcg charging code and the respective slab allocator specific details.

If we remove kmem accounting, we will still have implementation details spread over page allocator, reclaimer, rmap, memcg. Slab is not the worst part of it IMO. Anyway, kmem accounting can't be implemented solely in memcg.

The current state is quite complex already and making it even more complex by making allocation and charge context inconsistent is not really desirable.

Quoted 11 lines

Quoted 10 lines

Quoted 5 lines

Everyone else, except a couple of kmem users issuing alloc_page directly like threadinfo, will use kmalloc and know nothing what's going on there and how all this accounting stuff is handled - they will just use plain old convenient kmalloc, which works exactly as it does in the root cgroup.

If we ever grow more users and charge more kernel memory then they might be doing similar assumptions and tweak allocation/charge context and we would end up in a bigger mess. It makes much more sense to have allocation and charge context consistent.

Quoted 1 line

What new users? Why can't they just call kmalloc?

What about direct users of the page allocator. Why should they pay cost for more complex/expensive code paths when they do not need sub-page sizes.

Slab is a kind of abnormal alloc_pages user. By calling alloc_pages_node with __GFP_THISNODE and w/o __GFP_WAIT before falling back to alloc_pages with the caller's context, it does the job normally done by alloc_pages itself. It's not what is done massively.

Leaving slab charge path as is looks really ugly to me. Look, slab iterates over all nodes, inspecting if they have free pages and fails even if they do due to the memcg constraint...

To sum it up. Basically, there are two ways of handling kmemcg charges:

1. Make the memcg try_charge mimic alloc_pages behavior.
2. Make API functions (kmalloc, etc) work in memcg as if they were called from the root cgroup, while keeping interactions between the low level subsys (slab) and memcg private.

Way 1 might look appealing at the first glance, but at the same time it is much more complex, because alloc_pages has grown over the years to handle a lot of subtle situations that may arise on global memory pressure, but impossible in memcg. What does way 1 give us then? We can't insert try_charge directly to alloc_pages and have to spread its calls all over the code anyway, so why is it better? Easier to use it in places where users depend on buddy allocator peculiarities? There are not many such users.

exists solely for memcg-vs-list_lru and memcg-vs-slab interactions. We even handle kmem_cache destruction on memcg offline differently for SLAB and SLUB for performance reasons.

Way 2 gives us more space to maneuver IMO. SLAB/SLUB may do weird tricks for optimization, but their API is well defined, so we just make kmalloc work as expected while providing inter-subsys calls, like memcg_charge_slab, for SLAB/SLUB that have their own conventions. You mentioned kmem users that allocate memory using alloc_pages. There is an API function for them too, alloc_kmem_pages. Everything behind the API is hidden and may be done in such a way to achieve optimal performance.

Quoted 8 lines · Vladimir Davydov

Slab is a kind of abnormal alloc_pages user. By calling alloc_pages_node with __GFP_THISNODE and w/o __GFP_WAIT before falling back to alloc_pages with the caller's context, it does the job normally done by alloc_pages itself. It's not what is done massively.

Leaving slab charge path as is looks really ugly to me. Look, slab iterates over all nodes, inspecting if they have free pages and fails even if they do due to the memcg constraint...

Well yes it needs to do that due to the way NUMA support was designed in. SLAB needs to check the per node caches if objects are present before going to more remote nodes. Sorry about this. I realized the design issue in 2006 and SLUB was the result in 2007 of an alternate design to let the page allocator do its proper job.

Quoted 8 lines

To sum it up. Basically, there are two ways of handling kmemcg charges:

1. Make the memcg try_charge mimic alloc_pages behavior.
2. Make API functions (kmalloc, etc) work in memcg as if they were called from the root cgroup, while keeping interactions between the low level subsys (slab) and memcg private.

Quoted 8 lines

Way 1 might look appealing at the first glance, but at the same time it is much more complex, because alloc_pages has grown over the years to handle a lot of subtle situations that may arise on global memory pressure, but impossible in memcg. What does way 1 give us then? We can't insert try_charge directly to alloc_pages and have to spread its calls all over the code anyway, so why is it better? Easier to use it in places where users depend on buddy allocator peculiarities? There are not many such users.

Would it be possible to have a special alloc_pages_memcg with different semantics?

On the other hand alloc_pages() has grown to handle all the special cases. Why cant it also handle the special memcg case? There are numerous other

allocators that cache memory in the kernel from networking to the bizarre compressed swap approaches. How does memcg handle that? Isnt

that situation similar to what the slab allocators do?

Quoted 4 lines

exists solely for memcg-vs-list_lru and memcg-vs-slab interactions. We even handle kmem_cache destruction on memcg offline differently for SLAB and SLUB for performance reasons.

Ugly. Internal allocator design impacts container handling.

Quoted 8 lines

Way 2 gives us more space to maneuver IMO. SLAB/SLUB may do weird tricks for optimization, but their API is well defined, so we just make kmalloc work as expected while providing inter-subsys calls, like memcg_charge_slab, for SLAB/SLUB that have their own conventions. You mentioned kmem users that allocate memory using alloc_pages. There is an API function for them too, alloc_kmem_pages. Everything behind the API is hidden and may be done in such a way to achieve optimal performance.

Can we also hide cgroups memory handling behind the page based schemes without having extra handling for the slab allocators?

To sum it up. Basically, there are two ways of handling kmemcg charges:

1. Make the memcg try_charge mimic alloc_pages behavior.
2. Make API functions (kmalloc, etc) work in memcg as if they were called from the root cgroup, while keeping interactions between the low level subsys (slab) and memcg private.

Way 1 might look appealing at the first glance, but at the same time it is much more complex, because alloc_pages has grown over the years to handle a lot of subtle situations that may arise on global memory pressure, but impossible in memcg. What does way 1 give us then? We can't insert try_charge directly to alloc_pages and have to spread its calls all over the code anyway, so why is it better? Easier to use it in places where users depend on buddy allocator peculiarities? There are not many such users.

I understand that the idea of way 1 is to provide a well-defined memcg API independent of the rest of the code, but that's just impossible. You need special casing anyway. E.g. you need those get/put_kmem_cache helpers, which exist solely for SLAB/SLUB. You need all this special stuff for growing per-memcg array in list_lru and kmem_cache, which exists solely for memcg-vs-list_lru and memcg-vs-slab interactions. We even handle kmem_cache destruction on memcg offline differently for SLAB and SLUB for performance reasons.

Way 2 gives us more space to maneuver IMO. SLAB/SLUB may do weird tricks for optimization, but their API is well defined, so we just make kmalloc work as expected while providing inter-subsys calls, like memcg_charge_slab, for SLAB/SLUB that have their own conventions. You mentioned kmem users that allocate memory using alloc_pages. There is an API function for them too, alloc_kmem_pages. Everything behind the API is hidden and may be done in such a way to achieve optimal performance.

On Wed, Sep 02, 2015 at 12:30:39PM +0300, Vladimir Davydov wrote:
...

Quoted 7 lines

To sum it up. Basically, there are two ways of handling kmemcg charges:

1. Make the memcg try_charge mimic alloc_pages behavior.
2. Make API functions (kmalloc, etc) work in memcg as if they were called from the root cgroup, while keeping interactions between the low level subsys (slab) and memcg private.

Quoted 8 lines

Way 1 might look appealing at the first glance, but at the same time it is much more complex, because alloc_pages has grown over the years to handle a lot of subtle situations that may arise on global memory pressure, but impossible in memcg. What does way 1 give us then? We can't insert try_charge directly to alloc_pages and have to spread its calls all over the code anyway, so why is it better? Easier to use it in places where users depend on buddy allocator peculiarities? There are not many such users.

Maybe this is from inexperience but wouldn't 1 also be simpler than the global case for the same reasons that doing 2 is simpler? It's not like the fact that memory shortage inside memcg usually doesn't mean global shortage goes away depending on whether we take 1 or 2.

That said, it is true that slab is an integral part of kmemcg and I can't see how it can be made oblivious of memcg operations, so yeah one way or the other slab has to know the details and we may have to do some unusual things at that layer.

Quoted 8 lines

I understand that the idea of way 1 is to provide a well-defined memcg API independent of the rest of the code, but that's just impossible. You need special casing anyway. E.g. you need those get/put_kmem_cache helpers, which exist solely for SLAB/SLUB. You need all this special stuff for growing per-memcg array in list_lru and kmem_cache, which exists solely for memcg-vs-list_lru and memcg-vs-slab interactions. We even handle kmem_cache destruction on memcg offline differently for SLAB and SLUB for performance reasons.

It isn't a black or white thing. Sure, slab should be involved in kmemcg but at the same time if we can keep the amount of exposure in check, that's the better way to go.

Quoted 7 lines

Way 2 gives us more space to maneuver IMO. SLAB/SLUB may do weird tricks for optimization, but their API is well defined, so we just make kmalloc work as expected while providing inter-subsys calls, like memcg_charge_slab, for SLAB/SLUB that have their own conventions. You mentioned kmem users that allocate memory using alloc_pages. There is an API function for them too, alloc_kmem_pages. Everything behind the API is hidden and may be done in such a way to achieve optimal performance.

Ditto. Nobody is arguing that we can get it out completely but at the same time handling of GFP_NOWAIT seems like a pretty fundamental proprety that we'd wanna maintain at memcg boundary.

You said elsewhere that GFP_NOWAIT happening back-to-back is unlikely. I'm not sure how much we can commit to that statement. GFP_KERNEL allocating huge amount of memory in a single go is a kernel bug. GFP_NOWAIT optimization in a hot path which is accessible to userland isn't and we'll be growing more and more of them. We need to be protected against back-to-back GFP_NOWAIT allocations.

Trying a high-order page before falling back on lower order is not something really common. It implicitly relies on the fact that reclaiming memory for a new continuous high-order page is much more expensive than getting the same amount of order-1 pages. This is true for buddy alloc, but not for memcg. That's why playing such a trick with try_charge is wrong IMO. If such a trick becomes common, I think we will have to introduce a helper for it, because otherwise a change in buddy alloc internal logic (e.g. a defrag optimization making high order pages cheaper) may affect its users.

That said, I totally agree that memcg should handle GFP_NOWAIT, but I'm opposed to the idea that it should handle the tricks that rely on internal buddy alloc logic similar to those used by SLAB and SLUB. We'd better strive to hide these tricks in buddy alloc helpers and never use them directly.

That's why I think we need these patches and they aren't workarounds that can be reverted once try_charge has been taught to handle GFP_NOWAIT properly.

Quoted 6 lines

You said elsewhere that GFP_NOWAIT happening back-to-back is unlikely. I'm not sure how much we can commit to that statement. GFP_KERNEL allocating huge amount of memory in a single go is a kernel bug. GFP_NOWAIT optimization in a hot path which is accessible to userland isn't and we'll be growing more and more of them. We need to be protected against back-to-back GFP_NOWAIT allocations.

AFAIU if someone tries to allocate with GFP_NOWAIT (i.e. w/o __GFP_NOFAIL or __GFP_HIGH), he/she must be prepared to allocation failures, so there should be a safe fall back path, which fixes things in normal context. It doesn't mean we shouldn't do anything to satisfy such optimistic requests from memcg, but we may occasionally fail them.

OTOH if someone allocates with GFP_KERNEL, he/she should be prepared to get NULL, but in this case the whole operation will usually be aborted. Therefore with the possibility of all GFP_KERNEL being transformed to GFP_NOWAIT inside slab, memcg has to be extra cautious, because failing a usual GFP_NOWAIT in such a case may result not in falling back on slow path, but in user-visible effects like failing to open a file with ENOMEM. This is really difficult to achieve and I doubt it's worth complicating memcg code, because we can just fix SLAB/SLUB.

Regarding __GFP_NOFAIL and __GFP_HIGH, IMO we can let them go uncharged or charge them forcefully even if they breach the limit, because there shouldn't be many of them (if there were really a lot of them, they could deplete memory reserves and hang the system).

If all these assumptions are true, we don't need to do anything (apart from forcefully charging high prio allocations may be) for kmemcg to work satisfactory. For optimizing optimistic GFP_NOWAIT callers one can use memory.high instead or along with memory.max. Reclaiming memory.high in kernel while holding various locks can result in prio inversions though, but that's a different story, which could be fixed by task_work reclaim.

I admit I may be mistaken, but if I'm right, we may end up with really complex memcg reclaim logic trying to closely mimic behavior of buddy alloc with all its historic peculiarities. That's why I don't want to rush ahead "fixing" memcg reclaim before an agreement among all interested people is reached...

Quoted 5 lines

I admit I may be mistaken, but if I'm right, we may end up with really complex memcg reclaim logic trying to closely mimic behavior of buddy alloc with all its historic peculiarities. That's why I don't want to rush ahead "fixing" memcg reclaim before an agreement among all interested people is reached...

I think that's a bit out of proportion. I'm not suggesting bringing in all complexities of global reclaim. There's no reason to and what memcg deals with is inherently way simpler than actual memory allocation. The original patch was about fixing systematic failure around GFP_NOWAIT close to the high limit. We might want to do background reclaim close to max but as long as high limit functions correctly, that's much less of a problem at least on the v2 interface.

Now I think task_work reclaim initially proposed by Tejun would be a much better fix.

I'm terribly sorry for being so annoying and stubborn and want to thank you for all your feedback!

[
I'll try to summarize my point in one hunk instead of spreading it all over the e-mail, because IMO it's becoming a kind of difficult to follow. If you think that there's a question I dodge, please let me now and I'll try to address it separately.

Also, adding Johannes to Cc (I noticed that I accidentally left him out), because this discussion seems to be fundamental and may affect our further steps dramatically.
]

On Tue, Sep 01, 2015 at 08:38:50PM +0200, Michal Hocko wrote:

Quoted 6 lines

I guess we are still not at the same page here. If the slab has a subtle behavior (and from what you are saying it seems it has the same behavior at the global scope) then we should strive to fix it rather than making it more obscure just to not expose GFP_NOWAIT to memcg which is not handled properly currently wrt. high limit (more on that below) which was the primary motivation for the patch AFAIU.

Slab is a kind of abnormal alloc_pages user. By calling alloc_pages_node with __GFP_THISNODE and w/o __GFP_WAIT before falling back to alloc_pages with the caller's context, it does the job normally done by alloc_pages itself. It's not what is done massively.

Leaving slab charge path as is looks really ugly to me. Look, slab iterates over all nodes, inspecting if they have free pages and fails even if they do due to the memcg constraint...

My point is that what slab does is a pretty low level thing, normal users call alloc_pages or kmalloc with flags corresponding to their context. Of course, there may be special users trying optimistically GFP_NOWAIT, but they aren't massive, and that simplifies things for memcg a lot.

I mean if we can rely on the fact that the number of GFP_NOWAIT allocations that can occur in a row is limited we can use direct reclaim (like memory.high) and/or task_work reclaim to fix GFP_NOWAIT failures. Otherwise, we have to mimic the global alloc with most its heuristics. I don't think that copying those heuristics is the right thing to do, because in memcg case the same problems may be resolved much easier, because we don't actually experience real memory shortage when hitting the limit.

Moreover, we already treat some flags not in the same way as in case of slab for simplicity. E.g. we let __GFP_NOFAIL allocations go uncharged instead of retrying infinitely.

We ignore __GFP_THISNODE thing and we just cannot take it into account.

We ignore allocation order, because that makes no sense for memcg.

To sum it up. Basically, there are two ways of handling kmemcg charges:

1. Make the memcg try_charge mimic alloc_pages behavior.
2. Make API functions (kmalloc, etc) work in memcg as if they were called from the root cgroup, while keeping interactions between the low level subsys (slab) and memcg private.

Way 1 might look appealing at the first glance, but at the same time it is much more complex, because alloc_pages has grown over the years to handle a lot of subtle situations that may arise on global memory pressure, but impossible in memcg. What does way 1 give us then? We can't insert try_charge directly to alloc_pages and have to spread its calls all over the code anyway, so why is it better? Easier to use it in places where users depend on buddy allocator peculiarities? There are not many such users.

I understand that the idea of way 1 is to provide a well-defined memcg API independent of the rest of the code, but that's just impossible. You need special casing anyway. E.g. you need those get/put_kmem_cache helpers, which exist solely for SLAB/SLUB. You need all this special stuff for growing per-memcg array in list_lru and kmem_cache, which exists solely for memcg-vs-list_lru and memcg-vs-slab interactions. We even handle kmem_cache destruction on memcg offline differently for SLAB and SLUB for performance reasons.

Way 2 gives us more space to maneuver IMO. SLAB/SLUB may do weird tricks for optimization, but their API is well defined, so we just make kmalloc work as expected while providing inter-subsys calls, like memcg_charge_slab, for SLAB/SLUB that have their own conventions.

You mentioned kmem users that allocate memory using alloc_pages. There is an API function for them too, alloc_kmem_pages. Everything behind the API is hidden and may be done in such a way to achieve optimal performance.