| Message ID | 158893941613.200862.4094521350329937435.stgit@buzz (mailing list archive) |
|---|---|
| Headers | show |
| Series | dcache: increase poison resistance | expand |
Hi Konstantin, We tested this patch set recently and found it limiting negative dentry to a small part of total memory. And also we don't see any performance regression on it. Do you have any plan to integrate it into mainline? It will help a lot on memory fragmentation issue causing by dentry slab, there were a lot of customer cases where sys% was very high since most cpu were doing memory compaction, dentry slab was taking too much memory and nearly all dentry there were negative. The following is test result we run on two types of servers, one is 256G memory with 24 CPUS and another is 3T memory with 384 CPUS. The test case is using a lot of processes to generate negative dentry in parallel, the following is the test result after 72 hours, the negative dentry number is stable around that number even running longer time. If without the patch set, in less than half an hour 197G was took by negative dentry on 256G system, in 1 day 2.4T was took on 3T system. neg-dentry-number neg-dentry-mem-usage 256G 55259084 10.6G 3T 202306756 38.8G For perf test, we run the following, and no regression found. - create 1M negative dentry and then touch them to convert them to positive dentry - create 10K/100K/1M files - remove 10K/100K/1M files - kernel compile To verify the fsnotify fix, we used inotifywait to watch file create/open in some directory where there is a lot of negative dentry, without the patch set, the system will run into soft lockup, with it, no soft lockup. We also try to defeat the limitation by making different processes generating negative dentry with the same naming way, that will make one negative dentry being accessed couple times around same time, DCACHE_REFERENCED will be set on it and then it can't be trimmed easily. We do see negative dentry will take all the memory slowly from one of our system with 120G memory, for above two system, we see the memory usage were increased, but still a small part of total memory. This looks ok, since the common negative dentry user case will be create some temp files and then remove it, it will be rare to access same negative dentry around same time. Thanks, Junxiao. On 5/8/20 5:23 AM, Konstantin Khlebnikov wrote: > For most filesystems result of every negative lookup is cached, content of > directories is usually cached too. Production of negative dentries isn't > limited with disk speed. It's really easy to generate millions of them if > system has enough memory. > > Getting this memory back ins't that easy because slab frees pages only when > all related objects are gone. While dcache shrinker works in LRU order. > > Typical scenario is an idle system where some process periodically creates > temporary files and removes them. After some time, memory will be filled > with negative dentries for these random file names. > > Simple lookup of random names also generates negative dentries very fast. > Constant flow of such negative denries drains all other inactive caches. > > Negative dentries are linked into siblings list along with normal positive > dentries. Some operations walks dcache tree but looks only for positive > dentries: most important is fsnotify/inotify. Hordes of negative dentries > slow down these operations significantly. > > Time of dentry lookup is usually unaffected because hash table grows along > with size of memory. Unless somebody especially crafts hash collisions. > > This patch set solves all of these problems: > > Move negative denries to the end of sliblings list, thus walkers could > skip them at first sight (patches 3-6). > > Keep in dcache at most three unreferenced negative denties in row in each > hash bucket (patches 7-8). > > --- > > Konstantin Khlebnikov (8): > dcache: show count of hash buckets in sysctl fs.dentry-state > selftests: add stress testing tool for dcache > dcache: sweep cached negative dentries to the end of list of siblings > fsnotify: stop walking child dentries if remaining tail is negative > dcache: add action D_WALK_SKIP_SIBLINGS to d_walk() > dcache: stop walking siblings if remaining dentries all negative > dcache: push releasing dentry lock into sweep_negative > dcache: prevent flooding with negative dentries > > > fs/dcache.c | 144 +++++++++++- > fs/libfs.c | 10 +- > fs/notify/fsnotify.c | 6 +- > include/linux/dcache.h | 6 + > tools/testing/selftests/filesystems/Makefile | 1 + > .../selftests/filesystems/dcache_stress.c | 210 ++++++++++++++++++ > 6 files changed, 370 insertions(+), 7 deletions(-) > create mode 100644 tools/testing/selftests/filesystems/dcache_stress.c > > -- > Signature
On 12/11/20 11:32 PM, Konstantin Khlebnikov wrote: > On Thu, Dec 10, 2020 at 2:01 AM Junxiao Bi <junxiao.bi@oracle.com > <mailto:junxiao.bi@oracle.com>> wrote: > > Hi Konstantin, > > We tested this patch set recently and found it limiting negative > dentry > to a small part of total memory. And also we don't see any > performance > regression on it. Do you have any plan to integrate it into > mainline? It > will help a lot on memory fragmentation issue causing by dentry slab, > there were a lot of customer cases where sys% was very high since > most > cpu were doing memory compaction, dentry slab was taking too much > memory > and nearly all dentry there were negative. > > > Right now I don't have any plans for this. I suspect such problems will > appear much more often since machines are getting bigger. > So, somebody will take care of it. We already had a lot of customer cases. It made no sense to leave so many negative dentry in the system, it caused memory fragmentation and not much benefit. > > First part which collects negative dentries at the end list of > siblings could be > done in a more obvious way by splitting the list in two. > But this touches much more code. That would add new field to dentry? > > Last patch isn't very rigid but does non-trivial changes. > Probably it's better to call some garbage collector thingy periodically. > Lru list needs pressure to age and reorder entries properly. Swap the negative dentry to the head of hash list when it get accessed? Extra ones can be easily trimmed when swapping, using GC is to reduce perf impact? Thanks, Junxioao. > > Gc could be off by default or thresholds set very high (50% of ram for > example). > Final setup could be left up to owners of large systems, which needs > fine tuning.
On 12/13/20 11:43 PM, Konstantin Khlebnikov wrote: > > > On Sun, Dec 13, 2020 at 9:52 PM Junxiao Bi <junxiao.bi@oracle.com > <mailto:junxiao.bi@oracle.com>> wrote: > > On 12/11/20 11:32 PM, Konstantin Khlebnikov wrote: > > > On Thu, Dec 10, 2020 at 2:01 AM Junxiao Bi > <junxiao.bi@oracle.com <mailto:junxiao.bi@oracle.com> > > <mailto:junxiao.bi@oracle.com <mailto:junxiao.bi@oracle.com>>> > wrote: > > > > Hi Konstantin, > > > > We tested this patch set recently and found it limiting negative > > dentry > > to a small part of total memory. And also we don't see any > > performance > > regression on it. Do you have any plan to integrate it into > > mainline? It > > will help a lot on memory fragmentation issue causing by > dentry slab, > > there were a lot of customer cases where sys% was very high > since > > most > > cpu were doing memory compaction, dentry slab was taking too > much > > memory > > and nearly all dentry there were negative. > > > > > > Right now I don't have any plans for this. I suspect such > problems will > > appear much more often since machines are getting bigger. > > So, somebody will take care of it. > We already had a lot of customer cases. It made no sense to leave so > many negative dentry in the system, it caused memory fragmentation > and > not much benefit. > > > Dcache could grow so big only if the system lacks of memory pressure. > > Simplest solution is a cronjob which provinces such pressure by > creating sparse file on disk-based fs and then reading it. > This should wash away all inactive caches with no IO and zero chance > of oom. Sound good, will try. > > > > > First part which collects negative dentries at the end list of > > siblings could be > > done in a more obvious way by splitting the list in two. > > But this touches much more code. > That would add new field to dentry? > > > Yep. Decision is up to maintainers. > > > > > Last patch isn't very rigid but does non-trivial changes. > > Probably it's better to call some garbage collector thingy > periodically. > > Lru list needs pressure to age and reorder entries properly. > > Swap the negative dentry to the head of hash list when it get > accessed? > Extra ones can be easily trimmed when swapping, using GC is to reduce > perf impact? > > > Reclaimer/shrinker scans denties in LRU lists, it's an another list. Ah, you mean GC to reclaim from LRU list. I am not sure it could catch up the speed of negative dentry generating. Thanks, Junxiao. > My patch used order in hash lists is a very unusual way. Don't be > confused. > > There are four lists > parent - siblings > hashtable - hashchain > LRU > inode - alias > > > Thanks, > > Junxioao. > > > > > Gc could be off by default or thresholds set very high (50% of > ram for > > example). > > Final setup could be left up to owners of large systems, which > needs > > fine tuning. >
Hi Konstantin, How would you like to proceed with this patch set? This patchset as it is already fixed the customer issue we faced, it will stop memory fragmentation causing by negative dentry and no performance regression through our test. In production workload, it is common that some app kept creating and removing tmp files, this will leave a lot of negative dentry over time, some time later, it will cause memory fragmentation and system run into memory compaction and not responsible. It will be good to push it to upstream merge. If you are busy, we can try push it again. Thanks, Junxiao. On 12/14/20 3:10 PM, Junxiao Bi wrote: > On 12/13/20 11:43 PM, Konstantin Khlebnikov wrote: > >> >> >> On Sun, Dec 13, 2020 at 9:52 PM Junxiao Bi <junxiao.bi@oracle.com >> <mailto:junxiao.bi@oracle.com>> wrote: >> >> On 12/11/20 11:32 PM, Konstantin Khlebnikov wrote: >> >> > On Thu, Dec 10, 2020 at 2:01 AM Junxiao Bi >> <junxiao.bi@oracle.com <mailto:junxiao.bi@oracle.com> >> > <mailto:junxiao.bi@oracle.com <mailto:junxiao.bi@oracle.com>>> >> wrote: >> > >> > Hi Konstantin, >> > >> > We tested this patch set recently and found it limiting >> negative >> > dentry >> > to a small part of total memory. And also we don't see any >> > performance >> > regression on it. Do you have any plan to integrate it into >> > mainline? It >> > will help a lot on memory fragmentation issue causing by >> dentry slab, >> > there were a lot of customer cases where sys% was very high >> since >> > most >> > cpu were doing memory compaction, dentry slab was taking too >> much >> > memory >> > and nearly all dentry there were negative. >> > >> > >> > Right now I don't have any plans for this. I suspect such >> problems will >> > appear much more often since machines are getting bigger. >> > So, somebody will take care of it. >> We already had a lot of customer cases. It made no sense to leave so >> many negative dentry in the system, it caused memory fragmentation >> and >> not much benefit. >> >> >> Dcache could grow so big only if the system lacks of memory pressure. >> >> Simplest solution is a cronjob which provinces such pressure by >> creating sparse file on disk-based fs and then reading it. >> This should wash away all inactive caches with no IO and zero chance >> of oom. > Sound good, will try. >> >> > >> > First part which collects negative dentries at the end list of >> > siblings could be >> > done in a more obvious way by splitting the list in two. >> > But this touches much more code. >> That would add new field to dentry? >> >> >> Yep. Decision is up to maintainers. >> >> > >> > Last patch isn't very rigid but does non-trivial changes. >> > Probably it's better to call some garbage collector thingy >> periodically. >> > Lru list needs pressure to age and reorder entries properly. >> >> Swap the negative dentry to the head of hash list when it get >> accessed? >> Extra ones can be easily trimmed when swapping, using GC is to >> reduce >> perf impact? >> >> >> Reclaimer/shrinker scans denties in LRU lists, it's an another list. > > Ah, you mean GC to reclaim from LRU list. I am not sure it could catch > up the speed of negative dentry generating. > > Thanks, > > Junxiao. > >> My patch used order in hash lists is a very unusual way. Don't be >> confused. >> >> There are four lists >> parent - siblings >> hashtable - hashchain >> LRU >> inode - alias >> >> >> Thanks, >> >> Junxioao. >> >> > >> > Gc could be off by default or thresholds set very high (50% of >> ram for >> > example). >> > Final setup could be left up to owners of large systems, which >> needs >> > fine tuning. >>
For most filesystems result of every negative lookup is cached, content of directories is usually cached too. Production of negative dentries isn't limited with disk speed. It's really easy to generate millions of them if system has enough memory. Getting this memory back ins't that easy because slab frees pages only when all related objects are gone. While dcache shrinker works in LRU order. Typical scenario is an idle system where some process periodically creates temporary files and removes them. After some time, memory will be filled with negative dentries for these random file names. Simple lookup of random names also generates negative dentries very fast. Constant flow of such negative denries drains all other inactive caches. Negative dentries are linked into siblings list along with normal positive dentries. Some operations walks dcache tree but looks only for positive dentries: most important is fsnotify/inotify. Hordes of negative dentries slow down these operations significantly. Time of dentry lookup is usually unaffected because hash table grows along with size of memory. Unless somebody especially crafts hash collisions. This patch set solves all of these problems: Move negative denries to the end of sliblings list, thus walkers could skip them at first sight (patches 3-6). Keep in dcache at most three unreferenced negative denties in row in each hash bucket (patches 7-8). --- Konstantin Khlebnikov (8): dcache: show count of hash buckets in sysctl fs.dentry-state selftests: add stress testing tool for dcache dcache: sweep cached negative dentries to the end of list of siblings fsnotify: stop walking child dentries if remaining tail is negative dcache: add action D_WALK_SKIP_SIBLINGS to d_walk() dcache: stop walking siblings if remaining dentries all negative dcache: push releasing dentry lock into sweep_negative dcache: prevent flooding with negative dentries fs/dcache.c | 144 +++++++++++- fs/libfs.c | 10 +- fs/notify/fsnotify.c | 6 +- include/linux/dcache.h | 6 + tools/testing/selftests/filesystems/Makefile | 1 + .../selftests/filesystems/dcache_stress.c | 210 ++++++++++++++++++ 6 files changed, 370 insertions(+), 7 deletions(-) create mode 100644 tools/testing/selftests/filesystems/dcache_stress.c -- Signature