Re: [RFC PATCH 0/2] Remove shrinker's nr_deferred
On Wed, Sep 30, 2020 at 12:31 AM Dave Chinner wrote:
>
> On Sat, Sep 26, 2020 at 01:31:36PM -0700, Yang Shi wrote:
> > Hi Dave,
> >
> > I was exploring to make the "nr_deferred" per memcg. I looked into and
> > had some prototypes for two approaches so far:
> > 1. Have per memcg data structure for each memcg aware shrinker, just
> > like what shrinker_map does.
> > 2. Have "nr_deferred" on list_lru_one for memcg aware lists.
> >
> > Both seem feasible, however the latter one looks much cleaner, I just
> > need to add two new APIs for shrinker which gets and sets
> > "nr_deferred" respectively. And, just memcg aware shrinkers need
> > define those two callouts. We just need to care about memcg aware
> > shrinkers, and the most memcg aware shrinkers (inode/dentry, nfs and
> > workingset) use list_lru, so I'd prefer the latter one.
>
> The list_lru is completely separate from the shrinker context. The
> structure that tracks objects in a subsystem is not visible or aware
> of how the high level shrinker scanning algorithms work. Not to
> mention that subsystem shrinkers can be memcg aware without using
> list_lru structures to index objects owned by a given memcg. Hence I
> really don't like the idea of tying the shrinker control data deeply
> into subsystem cache indexing
I see your points. Yes, makes sense to me. The list_lru is a common
data structure and could be used by other subsystems, not only memcg
aware shrinkers.
Putting shrinker control data in list_lru seems break the layer. So,
option #1 might be more appropriate. The change looks like:
struct mem_cgroup_per_node {
...
struct memcg_shrinker_map __rcu *shrinker_map;
+ struct memcg_shrinker_deferred __rcu*shrinker_deferred;
...
}
>
>
> > But there are two memcg aware shrinkers are not that straightforward
> > to deal with:
> > 1. The deferred split THP. It doesn't use list_lru, but actually I
> > don't worry about this one, since it is not cache just some partial
> > unmapped THPs. I could try to convert it to use list_lru later on or
> > just kill deferred split by making vmscan split partial unmapped THPs.
> > So TBH I don't think it is a blocker.
>
> What a fantastic abuse of the reclaim infrastructure. :/
>
> First it was just defered work. Then it became NUMA_AWARE. THen it
> became MEMCG_AWARE and
>
> Oh, man what a nasty hack that SHRINKER_NONSLAB flag is so that it
> runs through shrink_slab_memcg() even when memcgs are configured in
> but kmem tracking disabled. We have heaps of shrinkers that reclaim
> from things that aren't slab caches, but this one is just nuts.
>
> > 2. The fs_objects. This one looks weird. It shares the same shrinker
> > with inode/dentry. The only user is XFS currently. But it looks it is
> > not really memcg aware at all, right?
>
> It most definitely is.
>
> The VFS dentry and inode cache reclaim are memcg aware. The
> fs_objects callout is for filesystem level object garbage collection
> that can be done as a result of the dentry and inode caches being
> reclaimed.
>
> i.e. once the VFS has reclaimed the inode attached to the memcg, it
> is no longer attached and accounted to the memcg anymore. It is
> owned by the filesystem at this point, and it is entirely up to the
> filesytem to when it can then be freed. Most filesystems do it in
> the inode cache reclaim via the ->destroy method. XFS, OTOH, tags
> freeable inodes in it's internal radix trees rather than freeing
> them immediately because it still may have to clean the inode before
> it can be freed. Hence we defer freeing of inodes until the
> ->fs_objects pass
Aha, thanks for elaborating. Now I see what it is doing for...
>
> > They are managed by radix tree
> > which is not per memcg by looking into xfs code, so the "per list_lru
> > nr_deferred" can't work for it. I thought of a couple of ways to
> > tackle it off the top of my head:
> > A. Just ignore it. If the amount of fs_objects are negligible
> > comparing to inode/dentry, then I think it can be just ignored and
> > kept it as is.
>
> Ah, no, they are not negliable. Under memory pressure, the number of
> objects is typically 1/3rd dentries, 1/3rd VFS inodes, 1/3rd fs
> objects to be reclaimed. The dentries and VFS inodes are owned by
> VFS level caches and associated with memcgs, the fs_objects are only
> visible to the filesystem.
>
> > B. Move it out of inode/dentry shrinker. Add a dedicated shrinker
> > for it, for example, sb->s_fs_obj_shrink.
>
> No, they are there because the reclaim has to be kept in exact
> proportion to the dentry and inode reclaim quantities. That's the
> reason I put that code there in the first place: a separate inode
> filesystem cache shrinker just didn't work well at all.
>
> > C. Make it really memcg aware and use list_lru.
>
> Two things. Firstly, objects are owned by the filesystem at this
> point, not memcgs. Memcgs were detatched at the VFS inode reclaim
> layer.
>
> Secondly, list-lru does not scale well
Re: [RFC PATCH 0/2] Remove shrinker's nr_deferred
On Sat, Sep 26, 2020 at 01:31:36PM -0700, Yang Shi wrote: > Hi Dave, > > I was exploring to make the "nr_deferred" per memcg. I looked into and > had some prototypes for two approaches so far: > 1. Have per memcg data structure for each memcg aware shrinker, just > like what shrinker_map does. > 2. Have "nr_deferred" on list_lru_one for memcg aware lists. > > Both seem feasible, however the latter one looks much cleaner, I just > need to add two new APIs for shrinker which gets and sets > "nr_deferred" respectively. And, just memcg aware shrinkers need > define those two callouts. We just need to care about memcg aware > shrinkers, and the most memcg aware shrinkers (inode/dentry, nfs and > workingset) use list_lru, so I'd prefer the latter one. The list_lru is completely separate from the shrinker context. The structure that tracks objects in a subsystem is not visible or aware of how the high level shrinker scanning algorithms work. Not to mention that subsystem shrinkers can be memcg aware without using list_lru structures to index objects owned by a given memcg. Hence I really don't like the idea of tying the shrinker control data deeply into subsystem cache indexing > But there are two memcg aware shrinkers are not that straightforward > to deal with: > 1. The deferred split THP. It doesn't use list_lru, but actually I > don't worry about this one, since it is not cache just some partial > unmapped THPs. I could try to convert it to use list_lru later on or > just kill deferred split by making vmscan split partial unmapped THPs. > So TBH I don't think it is a blocker. What a fantastic abuse of the reclaim infrastructure. :/ First it was just defered work. Then it became NUMA_AWARE. THen it became MEMCG_AWARE and Oh, man what a nasty hack that SHRINKER_NONSLAB flag is so that it runs through shrink_slab_memcg() even when memcgs are configured in but kmem tracking disabled. We have heaps of shrinkers that reclaim from things that aren't slab caches, but this one is just nuts. > 2. The fs_objects. This one looks weird. It shares the same shrinker > with inode/dentry. The only user is XFS currently. But it looks it is > not really memcg aware at all, right? It most definitely is. The VFS dentry and inode cache reclaim are memcg aware. The fs_objects callout is for filesystem level object garbage collection that can be done as a result of the dentry and inode caches being reclaimed. i.e. once the VFS has reclaimed the inode attached to the memcg, it is no longer attached and accounted to the memcg anymore. It is owned by the filesystem at this point, and it is entirely up to the filesytem to when it can then be freed. Most filesystems do it in the inode cache reclaim via the ->destroy method. XFS, OTOH, tags freeable inodes in it's internal radix trees rather than freeing them immediately because it still may have to clean the inode before it can be freed. Hence we defer freeing of inodes until the ->fs_objects pass > They are managed by radix tree > which is not per memcg by looking into xfs code, so the "per list_lru > nr_deferred" can't work for it. I thought of a couple of ways to > tackle it off the top of my head: > A. Just ignore it. If the amount of fs_objects are negligible > comparing to inode/dentry, then I think it can be just ignored and > kept it as is. Ah, no, they are not negliable. Under memory pressure, the number of objects is typically 1/3rd dentries, 1/3rd VFS inodes, 1/3rd fs objects to be reclaimed. The dentries and VFS inodes are owned by VFS level caches and associated with memcgs, the fs_objects are only visible to the filesystem. > B. Move it out of inode/dentry shrinker. Add a dedicated shrinker > for it, for example, sb->s_fs_obj_shrink. No, they are there because the reclaim has to be kept in exact proportion to the dentry and inode reclaim quantities. That's the reason I put that code there in the first place: a separate inode filesystem cache shrinker just didn't work well at all. > C. Make it really memcg aware and use list_lru. Two things. Firstly, objects are owned by the filesystem at this point, not memcgs. Memcgs were detatched at the VFS inode reclaim layer. Secondly, list-lru does not scale well enough for the use needed by XFS. We use radix trees so we can do lockless batch lookups and IO-efficient inode-order reclaim passes. We also have concurrent reclaim capabilities because of the lockless tag lookup walks. Using a list_lru for this substantially reduces reclaim performance and greatly increases CPU usage of reclaim because of contention on the internal list lru locks. Been there, measured that > I don't have any experience on XFS code, #C seems the most optimal, > but should be the most time consuming, I'm not sure if it is worth it > or not. So, #B sounds more preferred IMHO. I think you're going completely in the wrong direction. The problem that needs solving is integrating shrinker scanning control state
Re: [RFC PATCH 0/2] Remove shrinker's nr_deferred
Hi Dave, I was exploring to make the "nr_deferred" per memcg. I looked into and had some prototypes for two approaches so far: 1. Have per memcg data structure for each memcg aware shrinker, just like what shrinker_map does. 2. Have "nr_deferred" on list_lru_one for memcg aware lists. Both seem feasible, however the latter one looks much cleaner, I just need to add two new APIs for shrinker which gets and sets "nr_deferred" respectively. And, just memcg aware shrinkers need define those two callouts. We just need to care about memcg aware shrinkers, and the most memcg aware shrinkers (inode/dentry, nfs and workingset) use list_lru, so I'd prefer the latter one. But there are two memcg aware shrinkers are not that straightforward to deal with: 1. The deferred split THP. It doesn't use list_lru, but actually I don't worry about this one, since it is not cache just some partial unmapped THPs. I could try to convert it to use list_lru later on or just kill deferred split by making vmscan split partial unmapped THPs. So TBH I don't think it is a blocker. 2. The fs_objects. This one looks weird. It shares the same shrinker with inode/dentry. The only user is XFS currently. But it looks it is not really memcg aware at all, right? They are managed by radix tree which is not per memcg by looking into xfs code, so the "per list_lru nr_deferred" can't work for it. I thought of a couple of ways to tackle it off the top of my head: A. Just ignore it. If the amount of fs_objects are negligible comparing to inode/dentry, then I think it can be just ignored and kept it as is. B. Move it out of inode/dentry shrinker. Add a dedicated shrinker for it, for example, sb->s_fs_obj_shrink. C. Make it really memcg aware and use list_lru. I don't have any experience on XFS code, #C seems the most optimal, but should be the most time consuming, I'm not sure if it is worth it or not. So, #B sounds more preferred IMHO. Advice is much appreciated. Thanks. On Tue, Sep 22, 2020 at 4:45 PM Yang Shi wrote: > > On Sun, Sep 20, 2020 at 5:32 PM Dave Chinner wrote: > > > > On Thu, Sep 17, 2020 at 05:12:08PM -0700, Yang Shi wrote: > > > On Wed, Sep 16, 2020 at 7:37 PM Dave Chinner wrote: > > > > On Wed, Sep 16, 2020 at 11:58:21AM -0700, Yang Shi wrote: > > > > It clamps the worst case freeing to half the cache, and that is > > > > exactly what you are seeing. This, unfortunately, won't be enough to > > > > fix the windup problem once it's spiralled out of control. It's > > > > fairly rare for this to happen - it takes effort to find an adverse > > > > workload that will cause windup like this. > > > > > > I'm not sure if it is very rare, but my reproducer definitely could > > > generate huge amount of deferred objects easily. In addition it might > > > be easier to run into this case with hundreds of memcgs. Just imaging > > > hundreds memcgs run limit reclaims with __GFP_NOFS, the amount of > > > deferred objects can be built up easily. > > > > This is the first time I've seen a report that indicates excessive > > wind-up is occurring in years. That definitely makes it a rare > > problem in the real world. > > > > > On our production machine, I saw much more absurd deferred objects, > > > check the below tracing result out: > > > > > > <...>-48776 [032] 27970562.458916: mm_shrink_slab_start: > > > super_cache_scan+0x0/0x1a0 9a83046f3458: nid: 0 objects to shrink > > > 2531805877005 gfp_flags GFP_HIGHUSER_MOVABLE pgs_scanned 32 lru_pgs > > > 9300 cache items 1667 delta 11 total_scan 833 > > > > > > There are 2.5 trillion deferred objects on one node! So total > 5 > > > trillion! > > > > Sure, I'm not saying it's impossible to trigger, just that there are > > not many common workloads that actually cause it to occur. And, > > really, if it's wound up that far before you've noticed a problem, > > then wind-up itself isn't typically a serious problem for > > systems > > Actually the problem was observed some time ago, I just got some time > to look into the root cause. > > This kind of problem may be more common with memcg environment. For > example, a misconfigured memcg may incur excessive __GFP_NOFS limit > reclaims. > > > > > > > So, with all that said, a year ago I actually fixed this problem > > > > as part of some work I did to provide non-blocking inode reclaim > > > > infrastructure in the shrinker for XFS inode reclaim. > > > > See this patch: > > > > > > > > https://lore.kernel.org/linux-xfs/20191031234618.15403-13-da...@fromorbit.com/ > > > > > > Thanks for this. I remembered the patches, but I admitted I was not > > > aware deferred objects could go wild like that. > > > > Not many people are > > > > > > It did two things. First it ensured all the deferred work was done > > > > by kswapd so that some poor direct reclaim victim didn't hit a > > > > massive reclaim latency spike because of windup. Secondly, it > > > > clamped the maximum windup to the maximum single pass reclaim scan > > > >
Re: [RFC PATCH 0/2] Remove shrinker's nr_deferred
On Sun, Sep 20, 2020 at 5:32 PM Dave Chinner wrote: > > On Thu, Sep 17, 2020 at 05:12:08PM -0700, Yang Shi wrote: > > On Wed, Sep 16, 2020 at 7:37 PM Dave Chinner wrote: > > > On Wed, Sep 16, 2020 at 11:58:21AM -0700, Yang Shi wrote: > > > It clamps the worst case freeing to half the cache, and that is > > > exactly what you are seeing. This, unfortunately, won't be enough to > > > fix the windup problem once it's spiralled out of control. It's > > > fairly rare for this to happen - it takes effort to find an adverse > > > workload that will cause windup like this. > > > > I'm not sure if it is very rare, but my reproducer definitely could > > generate huge amount of deferred objects easily. In addition it might > > be easier to run into this case with hundreds of memcgs. Just imaging > > hundreds memcgs run limit reclaims with __GFP_NOFS, the amount of > > deferred objects can be built up easily. > > This is the first time I've seen a report that indicates excessive > wind-up is occurring in years. That definitely makes it a rare > problem in the real world. > > > On our production machine, I saw much more absurd deferred objects, > > check the below tracing result out: > > > > <...>-48776 [032] 27970562.458916: mm_shrink_slab_start: > > super_cache_scan+0x0/0x1a0 9a83046f3458: nid: 0 objects to shrink > > 2531805877005 gfp_flags GFP_HIGHUSER_MOVABLE pgs_scanned 32 lru_pgs > > 9300 cache items 1667 delta 11 total_scan 833 > > > > There are 2.5 trillion deferred objects on one node! So total > 5 trillion! > > Sure, I'm not saying it's impossible to trigger, just that there are > not many common workloads that actually cause it to occur. And, > really, if it's wound up that far before you've noticed a problem, > then wind-up itself isn't typically a serious problem for > systems Actually the problem was observed some time ago, I just got some time to look into the root cause. This kind of problem may be more common with memcg environment. For example, a misconfigured memcg may incur excessive __GFP_NOFS limit reclaims. > > > > So, with all that said, a year ago I actually fixed this problem > > > as part of some work I did to provide non-blocking inode reclaim > > > infrastructure in the shrinker for XFS inode reclaim. > > > See this patch: > > > > > > https://lore.kernel.org/linux-xfs/20191031234618.15403-13-da...@fromorbit.com/ > > > > Thanks for this. I remembered the patches, but I admitted I was not > > aware deferred objects could go wild like that. > > Not many people are > > > > It did two things. First it ensured all the deferred work was done > > > by kswapd so that some poor direct reclaim victim didn't hit a > > > massive reclaim latency spike because of windup. Secondly, it > > > clamped the maximum windup to the maximum single pass reclaim scan > > > limit, which is (freeable * 2) objects. > > > > > > Finally it also changed the amount of deferred work a single kswapd > > > pass did to be directly proportional to the reclaim priority. Hence > > > as we get closer to OOM, kswapd tries much harder to get the > > > deferred work backlog down to zero. This means that a single, low > > > priority reclaim pass will never reclaim half the cache - only > > > sustained memory pressure and _reclaim priority windup_ will do > > > that. > > > > Other than these, there are more problems: > > > > - The amount of deferred objects seem get significantly overestimated > > and unbounded. For example, if one lru has 1000 objects, the amount of > > reclaimables is bounded to 1000, but the amount of deferred is not. It > > may go much bigger than 1000, right? As the above tracing result > > shows, 2.5 trillion deferred objects on one node, assuming all of them > > are dentry (192 bytes per object), so the total size of deferred on > > one node is ~480TB! Or this is a bug? > > As the above patchset points out: it can get out of control because > it is unbounded. The above patchset bounds the deferred work to (2 * > current cache item count) and so it cannot ever spiral out of > control like this. I was thinking about cap it to (2 * freeable) too before I looked into your patches :-) > > > - The deferred will be reset by the reclaimer who gets there first, > > then other concurrent reclaimers just see 0 or very few deferred > > objects. > > No, not exactly. > > The current behaviour is that the deferred count is drained by the > current shrinker context, then it does whatever work it can, then it > puts the remainder of the work that was not done back on the > deferred count. This was done so that only a single reclaim context > tried to execute the deferred work (i.e. to prevent the deferred > work being run multiple times by concurrent reclaim contexts), but > if the work didn't get done it was still accounted and would get > done later. Yes, definitely. I should articulated it at the first place. > > A side effect of this was that nothing ever zeros the deferred > count, however,
Re: [RFC PATCH 0/2] Remove shrinker's nr_deferred
On Thu, Sep 17, 2020 at 05:12:08PM -0700, Yang Shi wrote: > On Wed, Sep 16, 2020 at 7:37 PM Dave Chinner wrote: > > On Wed, Sep 16, 2020 at 11:58:21AM -0700, Yang Shi wrote: > > It clamps the worst case freeing to half the cache, and that is > > exactly what you are seeing. This, unfortunately, won't be enough to > > fix the windup problem once it's spiralled out of control. It's > > fairly rare for this to happen - it takes effort to find an adverse > > workload that will cause windup like this. > > I'm not sure if it is very rare, but my reproducer definitely could > generate huge amount of deferred objects easily. In addition it might > be easier to run into this case with hundreds of memcgs. Just imaging > hundreds memcgs run limit reclaims with __GFP_NOFS, the amount of > deferred objects can be built up easily. This is the first time I've seen a report that indicates excessive wind-up is occurring in years. That definitely makes it a rare problem in the real world. > On our production machine, I saw much more absurd deferred objects, > check the below tracing result out: > > <...>-48776 [032] 27970562.458916: mm_shrink_slab_start: > super_cache_scan+0x0/0x1a0 9a83046f3458: nid: 0 objects to shrink > 2531805877005 gfp_flags GFP_HIGHUSER_MOVABLE pgs_scanned 32 lru_pgs > 9300 cache items 1667 delta 11 total_scan 833 > > There are 2.5 trillion deferred objects on one node! So total > 5 trillion! Sure, I'm not saying it's impossible to trigger, just that there are not many common workloads that actually cause it to occur. And, really, if it's wound up that far before you've noticed a problem, then wind-up itself isn't typically a serious problem for systems > > So, with all that said, a year ago I actually fixed this problem > > as part of some work I did to provide non-blocking inode reclaim > > infrastructure in the shrinker for XFS inode reclaim. > > See this patch: > > > > https://lore.kernel.org/linux-xfs/20191031234618.15403-13-da...@fromorbit.com/ > > Thanks for this. I remembered the patches, but I admitted I was not > aware deferred objects could go wild like that. Not many people are > > It did two things. First it ensured all the deferred work was done > > by kswapd so that some poor direct reclaim victim didn't hit a > > massive reclaim latency spike because of windup. Secondly, it > > clamped the maximum windup to the maximum single pass reclaim scan > > limit, which is (freeable * 2) objects. > > > > Finally it also changed the amount of deferred work a single kswapd > > pass did to be directly proportional to the reclaim priority. Hence > > as we get closer to OOM, kswapd tries much harder to get the > > deferred work backlog down to zero. This means that a single, low > > priority reclaim pass will never reclaim half the cache - only > > sustained memory pressure and _reclaim priority windup_ will do > > that. > > Other than these, there are more problems: > > - The amount of deferred objects seem get significantly overestimated > and unbounded. For example, if one lru has 1000 objects, the amount of > reclaimables is bounded to 1000, but the amount of deferred is not. It > may go much bigger than 1000, right? As the above tracing result > shows, 2.5 trillion deferred objects on one node, assuming all of them > are dentry (192 bytes per object), so the total size of deferred on > one node is ~480TB! Or this is a bug? As the above patchset points out: it can get out of control because it is unbounded. The above patchset bounds the deferred work to (2 * current cache item count) and so it cannot ever spiral out of control like this. > - The deferred will be reset by the reclaimer who gets there first, > then other concurrent reclaimers just see 0 or very few deferred > objects. No, not exactly. The current behaviour is that the deferred count is drained by the current shrinker context, then it does whatever work it can, then it puts the remainder of the work that was not done back on the deferred count. This was done so that only a single reclaim context tried to execute the deferred work (i.e. to prevent the deferred work being run multiple times by concurrent reclaim contexts), but if the work didn't get done it was still accounted and would get done later. A side effect of this was that nothing ever zeros the deferred count, however, because there is no serialisation between concurrent shrinker contexts. That's why it can wind up if the number of GFP_NOFS reclaim contexts greatly exceeds the number of GFP_KERNEL reclaim contexts. This is what the above patchset fixes - deferred work is only ever done by kswapd(), which means it doesn't have to care about multiple reclaim contexts doing deferred work. This simplifies it right down, and it allows us to bound the quantity of deferred work as a single reclaimer will be doing it all... > So the clamp may not happen on the lrus which have most > objects. For example, memcg A's dentry lru has 1000 objects,
Re: [RFC PATCH 0/2] Remove shrinker's nr_deferred
On Wed, Sep 16, 2020 at 7:37 PM Dave Chinner wrote: > > On Wed, Sep 16, 2020 at 11:58:21AM -0700, Yang Shi wrote: > > > > Recently huge amount one-off slab drop was seen on some vfs metadata heavy > > workloads, > > it turned out there were huge amount accumulated nr_deferred objects seen > > by the > > shrinker. > > > > I managed to reproduce this problem with kernel build workload plus > > negative dentry > > generator. > > > > First step, run the below kernel build test script: > > > > NR_CPUS=`cat /proc/cpuinfo | grep -e processor | wc -l` > > > > cd /root/Buildarea/linux-stable > > > > for i in `seq 1500`; do > > cgcreate -g memory:kern_build > > echo 4G > /sys/fs/cgroup/memory/kern_build/memory.limit_in_bytes > > > > echo 3 > /proc/sys/vm/drop_caches > > cgexec -g memory:kern_build make clean > /dev/null 2>&1 > > cgexec -g memory:kern_build make -j$NR_CPUS > /dev/null 2>&1 > > > > cgdelete -g memory:kern_build > > done > > > > That would generate huge amount deferred objects due to __GFP_NOFS > > allocations. > > > > Then run the below negative dentry generator script: > > > > NR_CPUS=`cat /proc/cpuinfo | grep -e processor | wc -l` > > > > mkdir /sys/fs/cgroup/memory/test > > echo $$ > /sys/fs/cgroup/memory/test/tasks > > > > for i in `seq $NR_CPUS`; do > > while true; do > > FILE=`head /dev/urandom | tr -dc A-Za-z0-9 | head -c 64` > > cat $FILE 2>/dev/null > > done & > > done > > > > Then kswapd will shrink half of dentry cache in just one loop as the below > > tracing result > > showed: > > > > kswapd0-475 [028] 305968.252561: mm_shrink_slab_start: > > super_cache_scan+0x0/0x190 24acf00c: nid: 0 > > objects to shrink 4994376020 gfp_flags GFP_KERNEL cache items 93689873 > > delta 45746 total_scan 46844936 priority 12 > > kswapd0-475 [021] 306013.099399: mm_shrink_slab_end: > > super_cache_scan+0x0/0x190 24acf00c: nid: 0 unused > > scan count 4994376020 new scan count 4947576838 total_scan 8 last shrinker > > return val 46844928 > > > You have 93M dentries and inodes in the cache, and the reclaim delta is 45746, > which is totally sane for a priority 12 reclaim priority. SO you've > basically had to do a couple of million GFP_NOFS direct reclaim > passes that were unable to reclaim anything to get to a point > where the deferred reclaim would up to 4.9 -billion- objects. > > Basically, you would up the deferred work so far that it got out of > control before a GFP_KERNEL reclaim context could do anything to > bring it under control. > > However, removing defered work is not the solution. If we don't > defer some of this reclaim work, then filesystem intensive workloads > -cannot reclaim memory from their own caches- when they need memory. > And when those caches largely dominate the used memory in the > machine, this will grind the filesystem workload to a halt.. Hence > this deferral mechanism is actually critical to keeping the > filesystem caches balanced with the rest of the system. Yes, I agree there might be imbalance if vfs caches shrinkers can't keep up due to excessive __GFP_NOFS allocations. > > The behaviour you see is the windup clamping code triggering: > > /* > * We need to avoid excessive windup on filesystem shrinkers > * due to large numbers of GFP_NOFS allocations causing the > * shrinkers to return -1 all the time. This results in a large > * nr being built up so when a shrink that can do some work > * comes along it empties the entire cache due to nr >>> > * freeable. This is bad for sustaining a working set in > * memory. > * > * Hence only allow the shrinker to scan the entire cache when > * a large delta change is calculated directly. > */ > if (delta < freeable / 4) > total_scan = min(total_scan, freeable / 2); > > It clamps the worst case freeing to half the cache, and that is > exactly what you are seeing. This, unfortunately, won't be enough to > fix the windup problem once it's spiralled out of control. It's > fairly rare for this to happen - it takes effort to find an adverse > workload that will cause windup like this. I'm not sure if it is very rare, but my reproducer definitely could generate huge amount of deferred objects easily. In addition it might be easier to run into this case with hundreds of memcgs. Just imaging hundreds memcgs run limit reclaims with __GFP_NOFS, the amount of deferred objects can be built up easily. On our production machine, I saw much more absurd deferred objects, check the below tracing result out: <...>-48776 [032] 27970562.458916: mm_shrink_slab_start: super_cache_scan+0x0/0x1a0 9a83046f3458: nid: 0 objects to shrink 2531805877005 gfp_flags GFP_HIGHUSER_MOVABLE pgs_scanned 32 lru_pgs 9300 cache items 1667 delta 11 total_scan 833 There are 2.5
Re: [RFC PATCH 0/2] Remove shrinker's nr_deferred
On Wed, Sep 16, 2020 at 11:58:21AM -0700, Yang Shi wrote: > > Recently huge amount one-off slab drop was seen on some vfs metadata heavy > workloads, > it turned out there were huge amount accumulated nr_deferred objects seen by > the > shrinker. > > I managed to reproduce this problem with kernel build workload plus negative > dentry > generator. > > First step, run the below kernel build test script: > > NR_CPUS=`cat /proc/cpuinfo | grep -e processor | wc -l` > > cd /root/Buildarea/linux-stable > > for i in `seq 1500`; do > cgcreate -g memory:kern_build > echo 4G > /sys/fs/cgroup/memory/kern_build/memory.limit_in_bytes > > echo 3 > /proc/sys/vm/drop_caches > cgexec -g memory:kern_build make clean > /dev/null 2>&1 > cgexec -g memory:kern_build make -j$NR_CPUS > /dev/null 2>&1 > > cgdelete -g memory:kern_build > done > > That would generate huge amount deferred objects due to __GFP_NOFS > allocations. > > Then run the below negative dentry generator script: > > NR_CPUS=`cat /proc/cpuinfo | grep -e processor | wc -l` > > mkdir /sys/fs/cgroup/memory/test > echo $$ > /sys/fs/cgroup/memory/test/tasks > > for i in `seq $NR_CPUS`; do > while true; do > FILE=`head /dev/urandom | tr -dc A-Za-z0-9 | head -c 64` > cat $FILE 2>/dev/null > done & > done > > Then kswapd will shrink half of dentry cache in just one loop as the below > tracing result > showed: > > kswapd0-475 [028] 305968.252561: mm_shrink_slab_start: > super_cache_scan+0x0/0x190 24acf00c: nid: 0 > objects to shrink 4994376020 gfp_flags GFP_KERNEL cache items 93689873 delta > 45746 total_scan 46844936 priority 12 > kswapd0-475 [021] 306013.099399: mm_shrink_slab_end: > super_cache_scan+0x0/0x190 24acf00c: nid: 0 unused > scan count 4994376020 new scan count 4947576838 total_scan 8 last shrinker > return val 46844928 You have 93M dentries and inodes in the cache, and the reclaim delta is 45746, which is totally sane for a priority 12 reclaim priority. SO you've basically had to do a couple of million GFP_NOFS direct reclaim passes that were unable to reclaim anything to get to a point where the deferred reclaim would up to 4.9 -billion- objects. Basically, you would up the deferred work so far that it got out of control before a GFP_KERNEL reclaim context could do anything to bring it under control. However, removing defered work is not the solution. If we don't defer some of this reclaim work, then filesystem intensive workloads -cannot reclaim memory from their own caches- when they need memory. And when those caches largely dominate the used memory in the machine, this will grind the filesystem workload to a halt.. Hence this deferral mechanism is actually critical to keeping the filesystem caches balanced with the rest of the system. The behaviour you see is the windup clamping code triggering: /* * We need to avoid excessive windup on filesystem shrinkers * due to large numbers of GFP_NOFS allocations causing the * shrinkers to return -1 all the time. This results in a large * nr being built up so when a shrink that can do some work * comes along it empties the entire cache due to nr >>> * freeable. This is bad for sustaining a working set in * memory. * * Hence only allow the shrinker to scan the entire cache when * a large delta change is calculated directly. */ if (delta < freeable / 4) total_scan = min(total_scan, freeable / 2); It clamps the worst case freeing to half the cache, and that is exactly what you are seeing. This, unfortunately, won't be enough to fix the windup problem once it's spiralled out of control. It's fairly rare for this to happen - it takes effort to find an adverse workload that will cause windup like this. So, with all that said, a year ago I actually fixed this problem as part of some work I did to provide non-blocking inode reclaim infrastructure in the shrinker for XFS inode reclaim. See this patch: https://lore.kernel.org/linux-xfs/20191031234618.15403-13-da...@fromorbit.com/ It did two things. First it ensured all the deferred work was done by kswapd so that some poor direct reclaim victim didn't hit a massive reclaim latency spike because of windup. Secondly, it clamped the maximum windup to the maximum single pass reclaim scan limit, which is (freeable * 2) objects. Finally it also changed the amount of deferred work a single kswapd pass did to be directly proportional to the reclaim priority. Hence as we get closer to OOM, kswapd tries much harder to get the deferred work backlog down to zero. This means that a single, low priority reclaim pass will never reclaim half the cache - only sustained memory pressure and _reclaim priority windup_ will do that. You probably want to look at all the shrinker infrastructure
[RFC PATCH 0/2] Remove shrinker's nr_deferred
Recently huge amount one-off slab drop was seen on some vfs metadata heavy
workloads,
it turned out there were huge amount accumulated nr_deferred objects seen by the
shrinker.
I managed to reproduce this problem with kernel build workload plus negative
dentry
generator.
First step, run the below kernel build test script:
NR_CPUS=`cat /proc/cpuinfo | grep -e processor | wc -l`
cd /root/Buildarea/linux-stable
for i in `seq 1500`; do
cgcreate -g memory:kern_build
echo 4G > /sys/fs/cgroup/memory/kern_build/memory.limit_in_bytes
echo 3 > /proc/sys/vm/drop_caches
cgexec -g memory:kern_build make clean > /dev/null 2>&1
cgexec -g memory:kern_build make -j$NR_CPUS > /dev/null 2>&1
cgdelete -g memory:kern_build
done
That would generate huge amount deferred objects due to __GFP_NOFS allocations.
Then run the below negative dentry generator script:
NR_CPUS=`cat /proc/cpuinfo | grep -e processor | wc -l`
mkdir /sys/fs/cgroup/memory/test
echo $$ > /sys/fs/cgroup/memory/test/tasks
for i in `seq $NR_CPUS`; do
while true; do
FILE=`head /dev/urandom | tr -dc A-Za-z0-9 | head -c 64`
cat $FILE 2>/dev/null
done &
done
Then kswapd will shrink half of dentry cache in just one loop as the below
tracing result
showed:
kswapd0-475 [028] 305968.252561: mm_shrink_slab_start:
super_cache_scan+0x0/0x190 24acf00c: nid: 0
objects to shrink 4994376020 gfp_flags GFP_KERNEL cache items 93689873 delta
45746 total_scan 46844936 priority 12
kswapd0-475 [021] 306013.099399: mm_shrink_slab_end:
super_cache_scan+0x0/0x190 24acf00c: nid: 0 unused
scan count 4994376020 new scan count 4947576838 total_scan 8 last shrinker
return val 46844928
There were huge deferred objects before the shrinker was called, the behavior
does match the code
but it might be not desirable from the user's stand of point.
IIUC the deferred objects were used to make balance between slab and page
cache, but since commit
9092c71bb724dba2ecba849eae69e5c9d39bd3d2 ("mm: use sc->priority for slab shrink
targets") they
were decoupled. And as that commit stated "these two things have nothing to do
with each other".
So why do we have to still keep it around? I can think of there might be huge
slab accumulated
without taking into account deferred objects, but nowadays the most workloads
are constrained by
memcg which could limit the usage of kmem (by default now), so it seems
maintaining deferred
objects is not that useful anymore. It seems we could remove it to simplify
the shrinker logic
a lot.
I may overlook some other important usecases of nr_deferred, comments are much
appreciated.
