- 23 Feb, 2017 40 commits
-
-
Mike Kravetz authored
Test userfaultfd hugetlb functionality by using the existing testing method (in userfaultfd.c). Instead of an anonymous memeory, a hugetlbfs file is mmap'ed private. In this way fallocate hole punch can be used to release pages. This is because madvise(MADV_DONTNEED) is not supported for huge pages. Use the same file, but create wrappers for allocating ranges and releasing pages. Compile userfaultfd.c with HUGETLB_TEST defined to produce an executable to test userfaultfd hugetlb functionality. Link: http://lkml.kernel.org/r/20161216144821.5183-23-aarcange@redhat.comSigned-off-by:
Mike Kravetz <mike.kravetz@oracle.com> Signed-off-by:
Andrea Arcangeli <aarcange@redhat.com> Cc: "Dr. David Alan Gilbert" <dgilbert@redhat.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Michael Rapoport <RAPOPORT@il.ibm.com> Cc: Mike Rapoport <rppt@linux.vnet.ibm.com> Cc: Pavel Emelyanov <xemul@parallels.com> Signed-off-by:
Andrew Morton <akpm@linux-foundation.org> Signed-off-by:
Linus Torvalds <torvalds@linux-foundation.org>
-
Mike Kravetz authored
Expand the userfaultfd_register/unregister routines to allow VM_HUGETLB vmas. huge page alignment checking is performed after a VM_HUGETLB vma is encountered. Also, since there is no UFFDIO_ZEROPAGE support for huge pages do not return that as a valid ioctl method for huge page ranges. Link: http://lkml.kernel.org/r/20161216144821.5183-22-aarcange@redhat.comSigned-off-by:
-
Mike Kravetz authored
When processing a hugetlb fault for no page present, check the vma to determine if faults are to be handled via userfaultfd. If so, drop the hugetlb_fault_mutex and call handle_userfault(). Link: http://lkml.kernel.org/r/20161216144821.5183-21-aarcange@redhat.comSigned-off-by:
Hillf Danton <hillf.zj@alibaba-inc.com> Cc: "Dr. David Alan Gilbert" <dgilbert@redhat.com> Cc: Michael Rapoport <RAPOPORT@il.ibm.com> Cc: Mike Rapoport <rppt@linux.vnet.ibm.com> Cc: Pavel Emelyanov <xemul@parallels.com> Signed-off-by:
-
Mike Kravetz authored
The new routine copy_huge_page_from_user() uses kmap_atomic() to map PAGE_SIZE pages. However, this prevents page faults in the subsequent call to copy_from_user(). This is OK in the case where the routine is copied with mmap_sema held. However, in another case we want to allow page faults. So, add a new argument allow_pagefault to indicate if the routine should allow page faults. [dan.carpenter@oracle.com: unmap the correct pointer] Link: http://lkml.kernel.org/r/20170113082608.GA3548@mwanda [akpm@linux-foundation.org: kunmap() takes a page*, per Hugh] Link: http://lkml.kernel.org/r/20161216144821.5183-20-aarcange@redhat.comSigned-off-by:
Dan Carpenter <dan.carpenter@oracle.com> Cc: "Dr. David Alan Gilbert" <dgilbert@redhat.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Michael Rapoport <RAPOPORT@il.ibm.com> Cc: Mike Rapoport <rppt@linux.vnet.ibm.com> Cc: Pavel Emelyanov <xemul@parallels.com> Cc: Hugh Dickins <hughd@google.com> Cc: Hugh Dickins <hughd@google.com> Signed-off-by:
-
Mike Kravetz authored
__mcopy_atomic_hugetlb performs the UFFDIO_COPY operation for huge pages. It is based on the existing __mcopy_atomic routine for normal pages. Unlike normal pages, there is no huge page support for the UFFDIO_ZEROPAGE operation. Link: http://lkml.kernel.org/r/20161216144821.5183-19-aarcange@redhat.comSigned-off-by:
-
Mike Kravetz authored
hugetlb_mcopy_atomic_pte is the low level routine that implements the userfaultfd UFFDIO_COPY command. It is based on the existing mcopy_atomic_pte routine with modifications for huge pages. Link: http://lkml.kernel.org/r/20161216144821.5183-18-aarcange@redhat.comSigned-off-by:
-
Mike Kravetz authored
userfaultfd UFFDIO_COPY allows user level code to copy data to a page at fault time. The data is copied from user space to a newly allocated huge page. The new routine copy_huge_page_from_user performs this copy. Link: http://lkml.kernel.org/r/20161216144821.5183-17-aarcange@redhat.comSigned-off-by:
-
Andrea Arcangeli authored
Userfaults may still happen after the userfaultfd monitor thread received a UFFD_EVENT_MADVDONTNEED until UFFDIO_UNREGISTER is run. Wake any pending userfault within UFFDIO_UNREGISTER protected by the mmap_sem for writing, so they will not be reported to userland leading to UFFDIO_COPY returning -EINVAL (as the range was already unregistered) and they will not hang permanently either. Link: http://lkml.kernel.org/r/20161216144821.5183-16-aarcange@redhat.comSigned-off-by:
-
Andrea Arcangeli authored
MADV_DONTNEED must be notified to userland before the pages are zapped. This allows userland to immediately stop adding pages to the userfaultfd ranges before the pages are actually zapped or there could be non-zeropage leftovers as result of concurrent UFFDIO_COPY run in between zap_page_range and madvise_userfault_dontneed (both MADV_DONTNEED and UFFDIO_COPY runs under the mmap_sem for reading, so they can run concurrently). Link: http://lkml.kernel.org/r/20161216144821.5183-15-aarcange@redhat.comSigned-off-by:
-
Pavel Emelyanov authored
If the page is punched out of the address space the uffd reader should know this and zeromap the respective area in case of the #PF event. Link: http://lkml.kernel.org/r/20161216144821.5183-14-aarcange@redhat.comSigned-off-by:
Pavel Emelyanov <xemul@parallels.com> Signed-off-by:
Mike Rapoport <rppt@linux.vnet.ibm.com> Signed-off-by:
-
Andrea Arcangeli authored
Optimize the mremap_userfaultfd_complete() interface to pass only the vm_userfaultfd_ctx pointer through the stack as a microoptimization. Link: http://lkml.kernel.org/r/20161216144821.5183-13-aarcange@redhat.comSigned-off-by:
-
Pavel Emelyanov authored
The event denotes that an area [start:end] moves to different location. Length change isn't reported as "new" addresses, if they appear on the uffd reader side they will not contain any data and the latter can just zeromap them. Waiting for the event ACK is also done outside of mmap sem, as for fork event. Link: http://lkml.kernel.org/r/20161216144821.5183-12-aarcange@redhat.comSigned-off-by:
-
Mike Rapoport authored
Since commit d2005e3f ("userfaultfd: don't pin the user memory in userfaultfd_file_create()") userfaultfd uses mm_count rather than mm_users to pin mm_struct. Make dup_userfaultfd consistent with this behaviour Link: http://lkml.kernel.org/r/20161216144821.5183-11-aarcange@redhat.comSigned-off-by:
-
Pavel Emelyanov authored
When the mm with uffd-ed vmas fork()-s the respective vmas notify their uffds with the event which contains a descriptor with new uffd. This new descriptor can then be used to get events from the child and populate its mm with data. Note, that there can be different uffd-s controlling different vmas within one mm, so first we should collect all those uffds (and ctx-s) in a list and then notify them all one by one but only once per fork(). The context is created at fork() time but the descriptor, file struct and anon inode object is created at event read time. So some trickery is added to the userfaultfd_ctx_read() to handle the ctx queues' locking vs file creation. Another thing worth noticing is that the task that fork()-s waits for the uffd event to get processed WITHOUT the mmap sem. [aarcange@redhat.com: build warning fix] Link: http://lkml.kernel.org/r/20161216144821.5183-10-aarcange@redhat.com Link: http://lkml.kernel.org/r/20161216144821.5183-9-aarcange@redhat.comSigned-off-by:
-
Andrea Arcangeli authored
This will allow userland to probe all features available in the kernel. It will however only enable the requested features in the open userfaultfd context. Link: http://lkml.kernel.org/r/20161216144821.5183-8-aarcange@redhat.comSigned-off-by:
-
Pavel Emelyanov authored
The custom events are queued in ctx->event_wqh not to disturb the fast-path-ed PF queue-wait-wakeup functions. The events to be generated (other than PF-s) are requested in UFFD_API ioctl with the uffd_api.features bits. Those, known by the kernel, are then turned on and reported back to the user-space. Link: http://lkml.kernel.org/r/20161216144821.5183-7-aarcange@redhat.comSigned-off-by:
-
Pavel Emelyanov authored
I will need one to lookup for userfaultfd_wait_queue-s in different wait queue Link: http://lkml.kernel.org/r/20161216144821.5183-6-aarcange@redhat.comSigned-off-by:
-
Andrea Arcangeli authored
Cleanup the vma->vm_ops usage. Side note: it would be more robust if vma_is_anonymous() would also check that vm_flags hasn't VM_PFNMAP set. Link: http://lkml.kernel.org/r/20161216144821.5183-5-aarcange@redhat.comSigned-off-by:
-
Andrea Arcangeli authored
Avoid BUG_ON()s and only WARN instead. This is just a cleanup, it can't make any runtime difference. This BUG_ON has never triggered and cannot trigger. Link: http://lkml.kernel.org/r/20161216144821.5183-4-aarcange@redhat.comSigned-off-by:
-
Andrea Arcangeli authored
Minor comment correction. Link: http://lkml.kernel.org/r/20161216144821.5183-3-aarcange@redhat.comSigned-off-by:
-
Andrea Arcangeli authored
Patch series "userfaultfd tmpfs/hugetlbfs/non-cooperative", v2 These userfaultfd features are finished and are ready for larger exposure in -mm and upstream merging. 1) tmpfs non present userfault 2) hugetlbfs non present userfault 3) non cooperative userfault for fork/madvise/mremap qemu development code is already exercising 2) and container postcopy live migration needs 3). 1) is not currently used but there's a self test and we know some qemu user for various reasons uses tmpfs as backing for KVM so it'll need it too to use postcopy live migration with tmpfs memory. All review feedback from the previous submit has been handled and the fixes are included. There's no outstanding issue AFIK. Upstream code just did a s/fe/vmf/ conversion in the page faults and this has been converted as well incrementally. In addition to the previous submits, this also wakes up stuck userfaults during UFFDIO_UNREGISTER. The non cooperative testcase actually reproduced this problem by getting stuck instead of quitting clean in some rare case as it could call UFFDIO_UNREGISTER while some userfault could be still in flight. The other option would have been to keep leaving it up to userland to serialize itself and to patch the testcase instead but the wakeup during unregister I think is preferable. David also asked the UFFD_FEATURE_MISSING_HUGETLBFS and UFFD_FEATURE_MISSING_SHMEM feature flags to be added so QEMU can avoid to probe if the hugetlbfs/shmem missing support is available by calling UFFDIO_REGISTER. QEMU already checks HUGETLBFS_MAGIC with fstatfs so if UFFD_FEATURE_MISSING_HUGETLBFS is also set, it knows UFFDIO_REGISTER will succeed (or if it fails, it's for some other more concerning reason). There's no reason to worry about adding too many feature flags. There are 64 available and worst case we've to bump the API if someday we're really going to run out of them. The round-trip network latency of hugetlbfs userfaults during postcopy live migration is still of the order of dozen milliseconds on 10GBit if at 2MB hugepage granularity so it's working perfectly and it should provide for higher bandwidth or lower CPU usage (which makes it interesting to add an option in the future to support THP granularity too for anonymous memory, UFFDIO_COPY would then have to create THP if alignment/len allows for it). 1GB hugetlbfs granularity will require big changes in hugetlbfs to work so it's deferred for later. This patch (of 42): This adds proper documentation (inline) to avoid the risk of further misunderstandings about the semantics of _IOW/_IOR and it also reminds whoever will bump the UFFDIO_API in the future, to change the two ioctl to _IOW. This was found while implementing strace support for those ioctl, otherwise we could have never found it by just reviewing kernel code and testing it. _IOC_READ or _IOC_WRITE alters nothing but the ioctl number itself, so it's only worth fixing if the UFFDIO_API is bumped someday. Link: http://lkml.kernel.org/r/20161216144821.5183-2-aarcange@redhat.comSigned-off-by:
"Dmitry V. Levin" <ldv@altlinux.org> Cc: Michael Rapoport <RAPOPORT@il.ibm.com> Cc: "Dr. David Alan Gilbert" <dgilbert@redhat.com> Cc: Mike Kravetz <mike.kravetz@oracle.com> Cc: Pavel Emelyanov <xemul@parallels.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Signed-off-by:
-
Michal Hocko authored
Higher order requests oom debugging is currently quite hard. We do have some compaction points which can tell us how the compaction is operating but there is no trace point to tell us about compaction retry logic. This patch adds a one which will have the following format bash-3126 [001] .... 1498.220001: compact_retry: order=9 priority=COMPACT_PRIO_SYNC_LIGHT compaction_result=withdrawn retries=0 max_retries=16 should_retry=0 we can see that the order 9 request is not retried even though we are in the highest compaction priority mode becase the last compaction attempt was withdrawn. This means that compaction_zonelist_suitable must have returned false and there is no suitable zone to compact for this request and so no need to retry further. another example would be <...>-3137 [001] .... 81.501689: compact_retry: order=9 priority=COMPACT_PRIO_SYNC_LIGHT compaction_result=failed retries=0 max_retries=16 should_retry=0 in this case the order-9 compaction failed to find any suitable block. We do not retry anymore because this is a costly request and those do not go below COMPACT_PRIO_SYNC_LIGHT priority. Link: http://lkml.kernel.org/r/20161220130135.15719-4-mhocko@kernel.orgSigned-off-by:Michal Hocko <mhocko@suse.com> Acked-by:
Vlastimil Babka <vbabka@suse.cz> Cc: David Rientjes <rientjes@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Signed-off-by:
-
Michal Hocko authored
should_reclaim_retry is the central decision point for declaring the OOM. It might be really useful to expose data used for this decision making when debugging an unexpected oom situations. Say we have an OOM report: [ 52.264001] mem_eater invoked oom-killer: gfp_mask=0x24280ca(GFP_HIGHUSER_MOVABLE|__GFP_ZERO), nodemask=0, order=0, oom_score_adj=0 [ 52.267549] CPU: 3 PID: 3148 Comm: mem_eater Tainted: G W 4.8.0-oomtrace3-00006-gb21338b386d2 #1024 Now we can check the tracepoint data to see how we have ended up in this situation: mem_eater-3148 [003] .... 52.432801: reclaim_retry_zone: node=0 zone=DMA32 order=0 reclaimable=51 available=11134 min_wmark=11084 no_progress_loops=1 wmark_check=1 mem_eater-3148 [003] .... 52.433269: reclaim_retry_zone: node=0 zone=DMA32 order=0 reclaimable=51 available=11103 min_wmark=11084 no_progress_loops=1 wmark_check=1 mem_eater-3148 [003] .... 52.433712: reclaim_retry_zone: node=0 zone=DMA32 order=0 reclaimable=51 available=11100 min_wmark=11084 no_progress_loops=2 wmark_check=1 mem_eater-3148 [003] .... 52.434067: reclaim_retry_zone: node=0 zone=DMA32 order=0 reclaimable=51 available=11097 min_wmark=11084 no_progress_loops=3 wmark_check=1 mem_eater-3148 [003] .... 52.434414: reclaim_retry_zone: node=0 zone=DMA32 order=0 reclaimable=51 available=11094 min_wmark=11084 no_progress_loops=4 wmark_check=1 mem_eater-3148 [003] .... 52.434761: reclaim_retry_zone: node=0 zone=DMA32 order=0 reclaimable=51 available=11091 min_wmark=11084 no_progress_loops=5 wmark_check=1 mem_eater-3148 [003] .... 52.435108: reclaim_retry_zone: node=0 zone=DMA32 order=0 reclaimable=51 available=11087 min_wmark=11084 no_progress_loops=6 wmark_check=1 mem_eater-3148 [003] .... 52.435478: reclaim_retry_zone: node=0 zone=DMA32 order=0 reclaimable=51 available=11084 min_wmark=11084 no_progress_loops=7 wmark_check=0 mem_eater-3148 [003] .... 52.435478: reclaim_retry_zone: node=0 zone=DMA order=0 reclaimable=0 available=1126 min_wmark=179 no_progress_loops=7 wmark_check=0 The above shows that we can quickly deduce that the reclaim stopped making any progress (see no_progress_loops increased in each round) and while there were still some 51 reclaimable pages they couldn't be dropped for some reason (vmscan trace points would tell us more about that part). available will represent reclaimable + free_pages scaled down per no_progress_loops factor. This is essentially an optimistic estimate of how much memory we would have when reclaiming everything. This can be compared to min_wmark to get a rought idea but the wmark_check tells the result of the watermark check which is more precise (includes lowmem reserves, considers the order etc.). As we can see no zone is eligible in the end and that is why we have triggered the oom in this situation. Please note that higher order requests might fail on the wmark_check even when there is much more memory available than min_wmark - e.g. when the memory is fragmented. A follow up tracepoint will help to debug those situations. Link: http://lkml.kernel.org/r/20161220130135.15719-3-mhocko@kernel.orgSigned-off-by: -
Michal Hocko authored
COMPACTION_STATUS resp. ZONE_TYPE are currently used to translate enum compact_result resp. struct zone index into their symbolic names for an easier post processing. The follow up patch would like to reuse this as well. The code involves some preprocessor black magic which is better not duplicated elsewhere so move it to a common mm tracing relate header. Link: http://lkml.kernel.org/r/20161220130135.15719-2-mhocko@kernel.orgSigned-off-by:
-
Geliang Tang authored
Use rb_entry_safe() instead of open-coding it. Link: http://lkml.kernel.org/r/81bb9820e5b9e4a1c596b3e76f88abf8c4a76cb0.1482221947.git.geliangtang@gmail.comSigned-off-by:
Geliang Tang <geliangtang@gmail.com> Acked-by:
-
Vlastimil Babka authored
On architectures that allow memory holes, page_is_buddy() has to perform page_to_pfn() to check for the memory hole. After the previous patch, we have the pfn already available in __free_one_page(), which is the only caller of page_is_buddy(), so move the check there and avoid page_to_pfn(). Link: http://lkml.kernel.org/r/20161216120009.20064-2-vbabka@suse.czSigned-off-by:
Mel Gorman <mgorman@techsingularity.net> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Signed-off-by:
-
Vlastimil Babka authored
In __free_one_page() we do the buddy merging arithmetics on "page/buddy index", which is just the lower MAX_ORDER bits of pfn. The operations we do that affect the higher bits are bitwise AND and subtraction (in that order), where the final result will be the same with the higher bits left unmasked, as long as these bits are equal for both buddies - which must be true by the definition of a buddy. We can therefore use pfn's directly instead of "index" and skip the zeroing of >MAX_ORDER bits. This can help a bit by itself, although compiler might be smart enough already. It also helps the next patch to avoid page_to_pfn() for memory hole checks. Link: http://lkml.kernel.org/r/20161216120009.20064-1-vbabka@suse.czSigned-off-by:
-
Michal Hocko authored
Tetsuo has been stressing OOM killer path with many parallel allocation requests when he has noticed that it is not all that hard to swamp kernel logs with warn_alloc messages caused by allocation stalls. Even though the allocation stall message is triggered only once in 10s there might be many different tasks hitting it roughly around the same time. A big part of the output is show_mem() which can generate a lot of output even on a small machines. There is no reason to show the state of memory counter for each allocation stall, especially when multiple of them are reported in a short time period. Chances are that not much has changed since the last report. This patch simply rate limits show_mem called from warn_alloc to only dump something once per second. This should be enough to give us a clue why an allocation might be stalling while burst of warnings will not swamp log with too much data. While we are at it, extract all the show_mem related handling (filters) into a separate function warn_alloc_show_mem. This will make the code cleaner and as a bonus point we can distinguish which part of warn_alloc got throttled due to rate limiting as ___ratelimit dumps the caller. [akpm@linux-foundation.org: reduce scope of the ratelimit_states] Link: http://lkml.kernel.org/r/20161215101510.9030-1-mhocko@kernel.orgSigned-off-by:
Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp> Signed-off-by:
-
Hugh Dickins authored
Callers of shmem_mapping() are interested in whether the mapping is swap backed - except for uprobes, which is interested in whether it should use shmem_read_mapping_page(). All these callers are better served by a shmem_mapping() which checks for shmem_aops, than the current version which goes through several indirections to find where the inode lives - and has the surprising effect that a private mmap of /dev/zero satisfies both vma_is_anonymous() and shmem_mapping(), when that device node is on devtmpfs. I don't think anything in the tree suffers from that surprise, but it caught me out, and is better fixed. Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1612052148530.13021@eggly.anvilsSigned-off-by:
Hugh Dickins <hughd@google.com> Acked-by:
Johannes Weiner <hannes@cmpxchg.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com> Signed-off-by:
-
Tejun Heo authored
SLUB creates a per-cache directory under /sys/kernel/slab which hosts a bunch of debug files. Usually, there aren't that many caches on a system and this doesn't really matter; however, if memcg is in use, each cache can have per-cgroup sub-caches. SLUB creates the same directories for these sub-caches under /sys/kernel/slab/$CACHE/cgroup. Unfortunately, because there can be a lot of cgroups, active or draining, the product of the numbers of caches, cgroups and files in each directory can reach a very high number - hundreds of thousands is commonplace. Millions and beyond aren't difficult to reach either. What's under /sys/kernel/slab is primarily for debugging and the information and control on the a root cache already cover its sub-caches. While having a separate directory for each sub-cache can be helpful for development, it doesn't make much sense to pay this amount of overhead by default. This patch introduces a boot parameter slub_memcg_sysfs which determines whether to create sysfs directories for per-memcg sub-caches. It also adds CONFIG_SLUB_MEMCG_SYSFS_ON which determines the boot parameter's default value and defaults to 0. [akpm@linux-foundation.org: kset_unregister(NULL) is legal] Link: http://lkml.kernel.org/r/20170204145203.GB26958@mtj.duckdns.orgSigned-off-by:
Tejun Heo <tj@kernel.org> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Signed-off-by:
-
Tejun Heo authored
If there's contention on slab_mutex, queueing the per-cache destruction work item on the system_wq can unnecessarily create and tie up a lot of kworkers. Rename memcg_kmem_cache_create_wq to memcg_kmem_cache_wq and make it global and use that workqueue for the destruction work items too. While at it, convert the workqueue from an unbound workqueue to a per-cpu one with concurrency limited to 1. It's generally preferable to use per-cpu workqueues and concurrency limit of 1 is safe enough. This is suggested by Joonsoo Kim. Link: http://lkml.kernel.org/r/20170117235411.9408-11-tj@kernel.orgSigned-off-by:
Jay Vana <jsvana@fb.com> Acked-by:
Vladimir Davydov <vdavydov@tarantool.org> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by:
-
Tejun Heo authored
With kmem cgroup support enabled, kmem_caches can be created and destroyed frequently and a great number of near empty kmem_caches can accumulate if there are a lot of transient cgroups and the system is not under memory pressure. When memory reclaim starts under such conditions, it can lead to consecutive deactivation and destruction of many kmem_caches, easily hundreds of thousands on moderately large systems, exposing scalability issues in the current slab management code. This is one of the patches to address the issue. Each cache has a number of sysfs interface files under /sys/kernel/slab. On a system with a lot of memory and transient memcgs, the number of interface files which have to be removed once memory reclaim kicks in can reach millions. Link: http://lkml.kernel.org/r/20170117235411.9408-10-tj@kernel.orgSigned-off-by:
Vladimir Davydov <vdavydov.dev@gmail.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by:
-
Tejun Heo authored
With kmem cgroup support enabled, kmem_caches can be created and destroyed frequently and a great number of near empty kmem_caches can accumulate if there are a lot of transient cgroups and the system is not under memory pressure. When memory reclaim starts under such conditions, it can lead to consecutive deactivation and destruction of many kmem_caches, easily hundreds of thousands on moderately large systems, exposing scalability issues in the current slab management code. This is one of the patches to address the issue. slub uses synchronize_sched() to deactivate a memcg cache. synchronize_sched() is an expensive and slow operation and doesn't scale when a huge number of caches are destroyed back-to-back. While there used to be a simple batching mechanism, the batching was too restricted to be helpful. This patch implements slab_deactivate_memcg_cache_rcu_sched() which slub can use to schedule sched RCU callback instead of performing synchronize_sched() synchronously while holding cgroup_mutex. While this adds online cpus, mems and slab_mutex operations, operating on these locks back-to-back from the same kworker, which is what's gonna happen when there are many to deactivate, isn't expensive at all and this gets rid of the scalability problem completely. Link: http://lkml.kernel.org/r/20170117235411.9408-9-tj@kernel.orgSigned-off-by:
-
Tejun Heo authored
__kmem_cache_shrink() is called with %true @deactivate only for memcg caches. Remove @deactivate from __kmem_cache_shrink() and introduce __kmemcg_cache_deactivate() instead. Each memcg-supporting allocator should implement it and it should deactivate and drain the cache. This is to allow memcg cache deactivation behavior to further deviate from simple shrinking without messing up __kmem_cache_shrink(). This is pure reorganization and doesn't introduce any observable behavior changes. v2: Dropped unnecessary ifdef in mm/slab.h as suggested by Vladimir. Link: http://lkml.kernel.org/r/20170117235411.9408-8-tj@kernel.orgSigned-off-by:
-
Tejun Heo authored
With kmem cgroup support enabled, kmem_caches can be created and destroyed frequently and a great number of near empty kmem_caches can accumulate if there are a lot of transient cgroups and the system is not under memory pressure. When memory reclaim starts under such conditions, it can lead to consecutive deactivation and destruction of many kmem_caches, easily hundreds of thousands on moderately large systems, exposing scalability issues in the current slab management code. This is one of the patches to address the issue. slab_caches currently lists all caches including root and memcg ones. This is the only data structure which lists the root caches and iterating root caches can only be done by walking the list while skipping over memcg caches. As there can be a huge number of memcg caches, this can become very expensive. This also can make /proc/slabinfo behave very badly. seq_file processes reads in 4k chunks and seeks to the previous Nth position on slab_caches list to resume after each chunk. With a lot of memcg cache churns on the list, reading /proc/slabinfo can become very slow and its content often ends up with duplicate and/or missing entries. This patch adds a new list slab_root_caches which lists only the root caches. When memcg is not enabled, it becomes just an alias of slab_caches. memcg specific list operations are collected into memcg_[un]link_cache(). Link: http://lkml.kernel.org/r/20170117235411.9408-7-tj@kernel.orgSigned-off-by:
-
Tejun Heo authored
With kmem cgroup support enabled, kmem_caches can be created and destroyed frequently and a great number of near empty kmem_caches can accumulate if there are a lot of transient cgroups and the system is not under memory pressure. When memory reclaim starts under such conditions, it can lead to consecutive deactivation and destruction of many kmem_caches, easily hundreds of thousands on moderately large systems, exposing scalability issues in the current slab management code. This is one of the patches to address the issue. While a memcg kmem_cache is listed on its root cache's ->children list, there is no direct way to iterate all kmem_caches which are assocaited with a memory cgroup. The only way to iterate them is walking all caches while filtering out caches which don't match, which would be most of them. This makes memcg destruction operations O(N^2) where N is the total number of slab caches which can be huge. This combined with the synchronous RCU operations can tie up a CPU and affect the whole machine for many hours when memory reclaim triggers offlining and destruction of the stale memcgs. This patch adds mem_cgroup->kmem_caches list which goes through memcg_cache_params->kmem_caches_node of all kmem_caches which are associated with the memcg. All memcg specific iterations, including stat file access, are updated to use the new list instead. Link: http://lkml.kernel.org/r/20170117235411.9408-6-tj@kernel.orgSigned-off-by:
-
Tejun Heo authored
We're going to change how memcg caches are iterated. In preparation, clean up and reorganize memcg_cache_params. * The shared ->list is replaced by ->children in root and ->children_node in children. * ->is_root_cache is removed. Instead ->root_cache is moved out of the child union and now used by both root and children. NULL indicates root cache. Non-NULL a memcg one. This patch doesn't cause any observable behavior changes. Link: http://lkml.kernel.org/r/20170117235411.9408-5-tj@kernel.orgSigned-off-by:
-
Tejun Heo authored
With kmem cgroup support enabled, kmem_caches can be created and destroyed frequently and a great number of near empty kmem_caches can accumulate if there are a lot of transient cgroups and the system is not under memory pressure. When memory reclaim starts under such conditions, it can lead to consecutive deactivation and destruction of many kmem_caches, easily hundreds of thousands on moderately large systems, exposing scalability issues in the current slab management code. This is one of the patches to address the issue. SLAB_DESTORY_BY_RCU caches need to flush all RCU operations before destruction because slab pages are freed through RCU and they need to be able to dereference the associated kmem_cache. Currently, it's done synchronously with rcu_barrier(). As rcu_barrier() is expensive time-wise, slab implements a batching mechanism so that rcu_barrier() can be done for multiple caches at the same time. Unfortunately, the rcu_barrier() is in synchronous path which is called while holding cgroup_mutex and the batching is too limited to be actually helpful. This patch updates the cache release path so that the batching is asynchronous and global. All SLAB_DESTORY_BY_RCU caches are queued globally and a work item consumes the list. The work item calls rcu_barrier() only once for all caches that are currently queued. * release_caches() is removed and shutdown_cache() now either directly release the cache or schedules a RCU callback to do that. This makes the cache inaccessible once shutdown_cache() is called and makes it impossible for shutdown_memcg_caches() to do memcg-specific cleanups afterwards. Move memcg-specific part into a helper, unlink_memcg_cache(), and make shutdown_cache() call it directly. Link: http://lkml.kernel.org/r/20170117235411.9408-4-tj@kernel.orgSigned-off-by:
-
Tejun Heo authored
Separate out slub sysfs removal and release, and call the former earlier from __kmem_cache_shutdown(). There's no reason to defer sysfs removal through RCU and this will later allow us to remove sysfs files way earlier during memory cgroup offline instead of release. Link: http://lkml.kernel.org/r/20170117235411.9408-3-tj@kernel.orgSigned-off-by:
-
Tejun Heo authored
Patch series "slab: make memcg slab destruction scalable", v3. With kmem cgroup support enabled, kmem_caches can be created and destroyed frequently and a great number of near empty kmem_caches can accumulate if there are a lot of transient cgroups and the system is not under memory pressure. When memory reclaim starts under such conditions, it can lead to consecutive deactivation and destruction of many kmem_caches, easily hundreds of thousands on moderately large systems, exposing scalability issues in the current slab management code. I've seen machines which end up with hundred thousands of caches and many millions of kernfs_nodes. The current code is O(N^2) on the total number of caches and has synchronous rcu_barrier() and synchronize_sched() in cgroup offline / release path which is executed while holding cgroup_mutex. Combined, this leads to very expensive and slow cache destruction operations which can easily keep running for half a day. This also messes up /proc/slabinfo along with other cache iterating operations. seq_file operates on 4k chunks and on each 4k boundary tries to seek to the last position in the list. With a huge number of caches on the list, this becomes very slow and very prone to the list content changing underneath it leading to a lot of missing and/or duplicate entries. This patchset addresses the scalability problem. * Add root and per-memcg lists. Update each user to use the appropriate list. * Make rcu_barrier() for SLAB_DESTROY_BY_RCU caches globally batched and asynchronous. * For dying empty slub caches, remove the sysfs files after deactivation so that we don't end up with millions of sysfs files without any useful information on them. This patchset contains the following nine patches. 0001-Revert-slub-move-synchronize_sched-out-of-slab_mutex.patch 0002-slub-separate-out-sysfs_slab_release-from-sysfs_slab.patch 0003-slab-remove-synchronous-rcu_barrier-call-in-memcg-ca.patch 0004-slab-reorganize-memcg_cache_params.patch 0005-slab-link-memcg-kmem_caches-on-their-associated-memo.patch 0006-slab-implement-slab_root_caches-list.patch 0007-slab-introduce-__kmemcg_cache_deactivate.patch 0008-slab-remove-synchronous-synchronize_sched-from-memcg.patch 0009-slab-remove-slub-sysfs-interface-files-early-for-emp.patch 0010-slab-use-memcg_kmem_cache_wq-for-slab-destruction-op.patch 0001 reverts an existing optimization to prepare for the following changes. 0002 is a prep patch. 0003 makes rcu_barrier() in release path batched and asynchronous. 0004-0006 separate out the lists. 0007-0008 replace synchronize_sched() in slub destruction path with call_rcu_sched(). 0009 removes sysfs files early for empty dying caches. 0010 makes destruction work items use a workqueue with limited concurrency. This patch (of 10): Revert 89e364db ("slub: move synchronize_sched out of slab_mutex on shrink"). With kmem cgroup support enabled, kmem_caches can be created and destroyed frequently and a great number of near empty kmem_caches can accumulate if there are a lot of transient cgroups and the system is not under memory pressure. When memory reclaim starts under such conditions, it can lead to consecutive deactivation and destruction of many kmem_caches, easily hundreds of thousands on moderately large systems, exposing scalability issues in the current slab management code. This is one of the patches to address the issue. Moving synchronize_sched() out of slab_mutex isn't enough as it's still inside cgroup_mutex. The whole deactivation / release path will be updated to avoid all synchronous RCU operations. Revert this insufficient optimization in preparation to ease future changes. Link: http://lkml.kernel.org/r/20170117235411.9408-2-tj@kernel.orgSigned-off-by:
-
