- Oct 16, 2020
-
-
David Howells authored
The file_ra_state being passed into page_cache_sync_readahead() was being ignored in favour of using the one embedded in the struct file. The only caller for which this makes a difference is the fsverity code if the file has been marked as POSIX_FADV_RANDOM, but it's confusing and worth fixing. Signed-off-by:
David Howells <dhowells@redhat.com> Signed-off-by:
Matthew Wilcox (Oracle) <willy@infradead.org> Signed-off-by:
Andrew Morton <akpm@linux-foundation.org> Cc: Eric Biggers <ebiggers@google.com> Link: https://lkml.kernel.org/r/20200903140844.14194-10-willy@infradead.org Signed-off-by:
Linus Torvalds <torvalds@linux-foundation.org>
-
David Howells authored
Fold ra_submit() into its last remaining user and pass the readahead_control struct to both do_page_cache_ra() and page_cache_sync_ra(). Signed-off-by:
-
David Howells authored
Reimplement force_page_cache_readahead() as a wrapper around force_page_cache_ra(). Pass the existing readahead_control from page_cache_sync_readahead(). Signed-off-by:
-
Matthew Wilcox (Oracle) authored
Rename __do_page_cache_readahead() to do_page_cache_ra() and call it directly from ondemand_readahead() instead of indirecting via ra_submit(). Signed-off-by:
-
- Oct 14, 2020
-
-
Matthew Wilcox (Oracle) authored
i915 does not want to see value entries. Switch it to use find_lock_page() instead, and remove the export of find_lock_entry(). Move find_lock_entry() and find_get_entry() to mm/internal.h to discourage any future use. Signed-off-by:
Johannes Weiner <hannes@cmpxchg.org> Cc: Alexey Dobriyan <adobriyan@gmail.com> Cc: Chris Wilson <chris@chris-wilson.co.uk> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Jani Nikula <jani.nikula@linux.intel.com> Cc: Matthew Auld <matthew.auld@intel.com> Cc: William Kucharski <william.kucharski@oracle.com> Link: https://lkml.kernel.org/r/20200910183318.20139-6-willy@infradead.org Signed-off-by:
-
- Aug 15, 2020
-
-
Matthew Wilcox (Oracle) authored
The thp prefix is more frequently used than hpage and we should be consistent between the various functions. [akpm@linux-foundation.org: fix mm/migrate.c] Signed-off-by:
William Kucharski <william.kucharski@oracle.com> Reviewed-by:
Zi Yan <ziy@nvidia.com> Cc: Mike Kravetz <mike.kravetz@oracle.com> Cc: David Hildenbrand <david@redhat.com> Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com> Link: http://lkml.kernel.org/r/20200629151959.15779-6-willy@infradead.org Signed-off-by:
-
Matthew Wilcox (Oracle) authored
This function returns the number of bytes in a THP. It is like page_size(), but compiles to just PAGE_SIZE if CONFIG_TRANSPARENT_HUGEPAGE is disabled. Signed-off-by:
-
- Aug 12, 2020
-
-
Joonsoo Kim authored
There is a well-defined migration target allocation callback. Use it. Signed-off-by:
Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by:
Michal Hocko <mhocko@suse.com> Acked-by:
Vlastimil Babka <vbabka@suse.cz> Cc: Christoph Hellwig <hch@infradead.org> Cc: Mike Kravetz <mike.kravetz@oracle.com> Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com> Cc: Roman Gushchin <guro@fb.com> Link: http://lkml.kernel.org/r/1594622517-20681-7-git-send-email-iamjoonsoo.kim@lge.com Signed-off-by:
-
Joonsoo Kim authored
There are some similar functions for migration target allocation. Since there is no fundamental difference, it's better to keep just one rather than keeping all variants. This patch implements base migration target allocation function. In the following patches, variants will be converted to use this function. Changes should be mechanical, but, unfortunately, there are some differences. First, some callers' nodemask is assgined to NULL since NULL nodemask will be considered as all available nodes, that is, &node_states[N_MEMORY]. Second, for hugetlb page allocation, gfp_mask is redefined as regular hugetlb allocation gfp_mask plus __GFP_THISNODE if user provided gfp_mask has it. This is because future caller of this function requires to set this node constaint. Lastly, if provided nodeid is NUMA_NO_NODE, nodeid is set up to the node where migration source lives. It helps to remove simple wrappers for setting up the nodeid. Note that PageHighmem() call in previous function is changed to open-code "is_highmem_idx()" since it provides more readability. [akpm@linux-foundation.org: tweak patch title, per Vlastimil] [akpm@linux-foundation.org: fix typo in comment] Signed-off-by:
-
Nitin Gupta authored
For some applications, we need to allocate almost all memory as hugepages. However, on a running system, higher-order allocations can fail if the memory is fragmented. Linux kernel currently does on-demand compaction as we request more hugepages, but this style of compaction incurs very high latency. Experiments with one-time full memory compaction (followed by hugepage allocations) show that kernel is able to restore a highly fragmented memory state to a fairly compacted memory state within <1 sec for a 32G system. Such data suggests that a more proactive compaction can help us allocate a large fraction of memory as hugepages keeping allocation latencies low. For a more proactive compaction, the approach taken here is to define a new sysctl called 'vm.compaction_proactiveness' which dictates bounds for external fragmentation which kcompactd tries to maintain. The tunable takes a value in range [0, 100], with a default of 20. Note that a previous version of this patch [1] was found to introduce too many tunables (per-order extfrag{low, high}), but this one reduces them to just one sysctl. Also, the new tunable is an opaque value instead of asking for specific bounds of "external fragmentation", which would have been difficult to estimate. The internal interpretation of this opaque value allows for future fine-tuning. Currently, we use a simple translation from this tunable to [low, high] "fragmentation score" thresholds (low=100-proactiveness, high=low+10%). The score for a node is defined as weighted mean of per-zone external fragmentation. A zone's present_pages determines its weight. To periodically check per-node score, we reuse per-node kcompactd threads, which are woken up every 500 milliseconds to check the same. If a node's score exceeds its high threshold (as derived from user-provided proactiveness value), proactive compaction is started until its score reaches its low threshold value. By default, proactiveness is set to 20, which implies threshold values of low=80 and high=90. This patch is largely based on ideas from Michal Hocko [2]. See also the LWN article [3]. Performance data ================ System: x64_64, 1T RAM, 80 CPU threads. Kernel: 5.6.0-rc3 + this patch echo madvise | sudo tee /sys/kernel/mm/transparent_hugepage/enabled echo madvise | sudo tee /sys/kernel/mm/transparent_hugepage/defrag Before starting the driver, the system was fragmented from a userspace program that allocates all memory and then for each 2M aligned section, frees 3/4 of base pages using munmap. The workload is mainly anonymous userspace pages, which are easy to move around. I intentionally avoided unmovable pages in this test to see how much latency we incur when hugepage allocations hit direct compaction. 1. Kernel hugepage allocation latencies With the system in such a fragmented state, a kernel driver then allocates as many hugepages as possible and measures allocation latency: (all latency values are in microseconds) - With vanilla 5.6.0-rc3 percentile latency –––––––––– ––––––– 5 7894 10 9496 25 12561 30 15295 40 18244 50 21229 60 27556 75 30147 80 31047 90 32859 95 33799 Total 2M hugepages allocated = 383859 (749G worth of hugepages out of 762G total free => 98% of free memory could be allocated as hugepages) - With 5.6.0-rc3 + this patch, with proactiveness=20 sysctl -w vm.compaction_proactiveness=20 percentile latency –––––––––– ––––––– 5 2 10 2 25 3 30 3 40 3 50 4 60 4 75 4 80 4 90 5 95 429 Total 2M hugepages allocated = 384105 (750G worth of hugepages out of 762G total free => 98% of free memory could be allocated as hugepages) 2. JAVA heap allocation In this test, we first fragment memory using the same method as for (1). Then, we start a Java process with a heap size set to 700G and request the heap to be allocated with THP hugepages. We also set THP to madvise to allow hugepage backing of this heap. /usr/bin/time java -Xms700G -Xmx700G -XX:+UseTransparentHugePages -XX:+AlwaysPreTouch The above command allocates 700G of Java heap using hugepages. - With vanilla 5.6.0-rc3 17.39user 1666.48system 27:37.89elapsed - With 5.6.0-rc3 + this patch, with proactiveness=20 8.35user 194.58system 3:19.62elapsed Elapsed time remains around 3:15, as proactiveness is further increased. Note that proactive compaction happens throughout the runtime of these workloads. The situation of one-time compaction, sufficient to supply hugepages for following allocation stream, can probably happen for more extreme proactiveness values, like 80 or 90. In the above Java workload, proactiveness is set to 20. The test starts with a node's score of 80 or higher, depending on the delay between the fragmentation step and starting the benchmark, which gives more-or-less time for the initial round of compaction. As t he benchmark consumes hugepages, node's score quickly rises above the high threshold (90) and proactive compaction starts again, which brings down the score to the low threshold level (80). Repeat. bpftrace also confirms proactive compaction running 20+ times during the runtime of this Java benchmark. kcompactd threads consume 100% of one of the CPUs while it tries to bring a node's score within thresholds. Backoff behavior ================ Above workloads produce a memory state which is easy to compact. However, if memory is filled with unmovable pages, proactive compaction should essentially back off. To test this aspect: - Created a kernel driver that allocates almost all memory as hugepages followed by freeing first 3/4 of each hugepage. - Set proactiveness=40 - Note that proactive_compact_node() is deferred maximum number of times with HPAGE_FRAG_CHECK_INTERVAL_MSEC of wait between each check (=> ~30 seconds between retries). [1] https://patchwork.kernel.org/patch/11098289/ [2] https://lore.kernel.org/linux-mm/20161230131412.GI13301@dhcp22.suse.cz/ [3] https://lwn.net/Articles/817905/ Signed-off-by:Nitin Gupta <nigupta@nvidia.com> Signed-off-by:
Oleksandr Natalenko <oleksandr@redhat.com> Reviewed-by:
Khalid Aziz <khalid.aziz@oracle.com> Reviewed-by:
-
- Jun 09, 2020
-
-
Michel Lespinasse authored
Convert comments that reference mmap_sem to reference mmap_lock instead. [akpm@linux-foundation.org: fix up linux-next leftovers] [akpm@linux-foundation.org: s/lockaphore/lock/, per Vlastimil] [akpm@linux-foundation.org: more linux-next fixups, per Michel] Signed-off-by:
Michel Lespinasse <walken@google.com> Signed-off-by:
Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Davidlohr Bueso <dbueso@suse.de> Cc: David Rientjes <rientjes@google.com> Cc: Hugh Dickins <hughd@google.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Jerome Glisse <jglisse@redhat.com> Cc: John Hubbard <jhubbard@nvidia.com> Cc: Laurent Dufour <ldufour@linux.ibm.com> Cc: Liam Howlett <Liam.Howlett@oracle.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Ying Han <yinghan@google.com> Link: http://lkml.kernel.org/r/20200520052908.204642-13-walken@google.com Signed-off-by:
-
Michel Lespinasse authored
This change converts the existing mmap_sem rwsem calls to use the new mmap locking API instead. The change is generated using coccinelle with the following rule: // spatch --sp-file mmap_lock_api.cocci --in-place --include-headers --dir . @@ expression mm; @@ ( -init_rwsem +mmap_init_lock | -down_write +mmap_write_lock | -down_write_killable +mmap_write_lock_killable | -down_write_trylock +mmap_write_trylock | -up_write +mmap_write_unlock | -downgrade_write +mmap_write_downgrade | -down_read +mmap_read_lock | -down_read_killable +mmap_read_lock_killable | -down_read_trylock +mmap_read_trylock | -up_read +mmap_read_unlock ) -(&mm->mmap_sem) +(mm) Signed-off-by:
Laurent Dufour <ldufour@linux.ibm.com> Reviewed-by:
-
- Jun 05, 2020
-
-
Ethon Paul authored
There is a typo in comment, fix it. Signed-off-by:
Ethon Paul <ethp@qq.com> Signed-off-by:
Ralph Campbell <rcampbell@nvidia.com> Link: http://lkml.kernel.org/r/20200411064723.15855-1-ethp@qq.com Signed-off-by:
-
- Jun 04, 2020
-
-
Maninder Singh authored
commit 3c710c1a ("mm, vmscan extract shrink_page_list reclaim counters into a struct") changed data type for the function, so changing return type for funciton and its caller. Signed-off-by:
Vaneet Narang <v.narang@samsung.com> Signed-off-by:
Maninder Singh <maninder1.s@samsung.com> Signed-off-by:
-
Joonsoo Kim authored
classzone_idx is just different name for high_zoneidx now. So, integrate them and add some comment to struct alloc_context in order to reduce future confusion about the meaning of this variable. The accessor, ac_classzone_idx() is also removed since it isn't needed after integration. In addition to integration, this patch also renames high_zoneidx to highest_zoneidx since it represents more precise meaning. Signed-off-by:
Baoquan He <bhe@redhat.com> Acked-by:
David Rientjes <rientjes@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Ye Xiaolong <xiaolong.ye@intel.com> Link: http://lkml.kernel.org/r/1587095923-7515-3-git-send-email-iamjoonsoo.kim@lge.com Signed-off-by:
-
Joonsoo Kim authored
Patch series "integrate classzone_idx and high_zoneidx", v5. This patchset is followup of the problem reported and discussed two years ago [1, 2]. The problem this patchset solves is related to the classzone_idx on the NUMA system. It causes a problem when the lowmem reserve protection exists for some zones on a node that do not exist on other nodes. This problem was reported two years ago, and, at that time, the solution got general agreements [2]. But it was not upstreamed. [1]: http://lkml.kernel.org/r/20180102063528.GG30397@yexl-desktop [2]: http://lkml.kernel.org/r/1525408246-14768-1-git-send-email-iamjoonsoo.kim@lge.com This patch (of 2): Currently, we use classzone_idx to calculate lowmem reserve proetection for an allocation request. This classzone_idx causes a problem on NUMA systems when the lowmem reserve protection exists for some zones on a node that do not exist on other nodes. Before further explanation, I should first clarify how to compute the classzone_idx and the high_zoneidx. - ac->high_zoneidx is computed via the arcane gfp_zone(gfp_mask) and represents the index of the highest zone the allocation can use - classzone_idx was supposed to be the index of the highest zone on the local node that the allocation can use, that is actually available in the system Think about following example. Node 0 has 4 populated zone, DMA/DMA32/NORMAL/MOVABLE. Node 1 has 1 populated zone, NORMAL. Some zones, such as MOVABLE, doesn't exist on node 1 and this makes following difference. Assume that there is an allocation request whose gfp_zone(gfp_mask) is the zone, MOVABLE. Then, it's high_zoneidx is 3. If this allocation is initiated on node 0, it's classzone_idx is 3 since actually available/usable zone on local (node 0) is MOVABLE. If this allocation is initiated on node 1, it's classzone_idx is 2 since actually available/usable zone on local (node 1) is NORMAL. You can see that classzone_idx of the allocation request are different according to their starting node, even if their high_zoneidx is the same. Think more about these two allocation requests. If they are processed on local, there is no problem. However, if allocation is initiated on node 1 are processed on remote, in this example, at the NORMAL zone on node 0, due to memory shortage, problem occurs. Their different classzone_idx leads to different lowmem reserve and then different min watermark. See the following example. root@ubuntu:/sys/devices/system/memory# cat /proc/zoneinfo Node 0, zone DMA per-node stats ... pages free 3965 min 5 low 8 high 11 spanned 4095 present 3998 managed 3977 protection: (0, 2961, 4928, 5440) ... Node 0, zone DMA32 pages free 757955 min 1129 low 1887 high 2645 spanned 1044480 present 782303 managed 758116 protection: (0, 0, 1967, 2479) ... Node 0, zone Normal pages free 459806 min 750 low 1253 high 1756 spanned 524288 present 524288 managed 503620 protection: (0, 0, 0, 4096) ... Node 0, zone Movable pages free 130759 min 195 low 326 high 457 spanned 1966079 present 131072 managed 131072 protection: (0, 0, 0, 0) ... Node 1, zone DMA pages free 0 min 0 low 0 high 0 spanned 0 present 0 managed 0 protection: (0, 0, 1006, 1006) Node 1, zone DMA32 pages free 0 min 0 low 0 high 0 spanned 0 present 0 managed 0 protection: (0, 0, 1006, 1006) Node 1, zone Normal per-node stats ... pages free 233277 min 383 low 640 high 897 spanned 262144 present 262144 managed 257744 protection: (0, 0, 0, 0) ... Node 1, zone Movable pages free 0 min 0 low 0 high 0 spanned 262144 present 0 managed 0 protection: (0, 0, 0, 0) - static min watermark for the NORMAL zone on node 0 is 750. - lowmem reserve for the request with classzone idx 3 at the NORMAL on node 0 is 4096. - lowmem reserve for the request with classzone idx 2 at the NORMAL on node 0 is 0. So, overall min watermark is: allocation initiated on node 0 (classzone_idx 3): 750 + 4096 = 4846 allocation initiated on node 1 (classzone_idx 2): 750 + 0 = 750 Allocation initiated on node 1 will have some precedence than allocation initiated on node 0 because min watermark of the former allocation is lower than the other. So, allocation initiated on node 1 could succeed on node 0 when allocation initiated on node 0 could not, and, this could cause too many numa_miss allocation. Then, performance could be downgraded. Recently, there was a regression report about this problem on CMA patches since CMA memory are placed in ZONE_MOVABLE by those patches. I checked that problem is disappeared with this fix that uses high_zoneidx for classzone_idx. http://lkml.kernel.org/r/20180102063528.GG30397@yexl-desktop Using high_zoneidx for classzone_idx is more consistent way than previous approach because system's memory layout doesn't affect anything to it. With this patch, both classzone_idx on above example will be 3 so will have the same min watermark. allocation initiated on node 0: 750 + 4096 = 4846 allocation initiated on node 1: 750 + 4096 = 4846 One could wonder if there is a side effect that allocation initiated on node 1 will use higher bar when allocation is handled on local since classzone_idx could be higher than before. It will not happen because the zone without managed page doesn't contributes lowmem_reserve at all. Reported-by:
Ye Xiaolong <xiaolong.ye@intel.com> Signed-off-by:
-
- Jun 02, 2020
-
-
Matthew Wilcox (Oracle) authored
ondemand_readahead has two callers, neither of which use the return value. That means that both ra_submit and __do_page_cache_readahead() can return void, and we don't need to worry that a present page in the readahead window causes us to return a smaller nr_pages than we ought to have. Similarly, no caller uses the return value from force_page_cache_readahead(). Signed-off-by:
Dave Chinner <dchinner@redhat.com> Reviewed-by:
John Hubbard <jhubbard@nvidia.com> Reviewed-by:
Christoph Hellwig <hch@lst.de> Reviewed-by:
-
Matthew Wilcox (Oracle) authored
Patch series "Change readahead API", v11. This series adds a readahead address_space operation to replace the readpages operation. The key difference is that pages are added to the page cache as they are allocated (and then looked up by the filesystem) instead of passing them on a list to the readpages operation and having the filesystem add them to the page cache. It's a net reduction in code for each implementation, more efficient than walking a list, and solves the direct-write vs buffered-read problem reported by yu kuai at http://lkml.kernel.org/r/20200116063601.39201-1-yukuai3@huawei.com The only unconverted filesystems are those which use fscache. Their conversion is pending Dave Howells' rewrite which will make the conversion substantially easier. This should be completed by the end of the year. I want to thank the reviewers/testers; Dave Chinner, John Hubbard, Eric Biggers, Johannes Thumshirn, Dave Sterba, Zi Yan, Christoph Hellwig and Miklos Szeredi have done a marvellous job of providing constructive criticism. These patches pass an xfstests run on ext4, xfs & btrfs with no regressions that I can tell (some of the tests seem a little flaky before and remain flaky afterwards). This patch (of 25): The readahead code is part of the page cache so should be found in the pagemap.h file. force_page_cache_readahead is only used within mm, so move it to mm/internal.h instead. Remove the parameter names where they add no value, and rename the ones which were actively misleading. Signed-off-by:
Johannes Thumshirn <johannes.thumshirn@wdc.com> Cc: Chao Yu <yuchao0@huawei.com> Cc: Cong Wang <xiyou.wangcong@gmail.com> Cc: Darrick J. Wong <darrick.wong@oracle.com> Cc: Dave Chinner <dchinner@redhat.com> Cc: Eric Biggers <ebiggers@google.com> Cc: Gao Xiang <gaoxiang25@huawei.com> Cc: Jaegeuk Kim <jaegeuk@kernel.org> Cc: Joseph Qi <joseph.qi@linux.alibaba.com> Cc: Junxiao Bi <junxiao.bi@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Zi Yan <ziy@nvidia.com> Cc: Miklos Szeredi <mszeredi@redhat.com> Link: http://lkml.kernel.org/r/20200414150233.24495-1-willy@infradead.org Link: http://lkml.kernel.org/r/20200414150233.24495-2-willy@infradead.org Signed-off-by:
-
- Apr 07, 2020
-
-
Alexander Duyck authored
There are cases where we would benefit from avoiding having to go through the allocation and free cycle to return an isolated page. Examples for this might include page poisoning in which we isolate a page and then put it back in the free list without ever having actually allocated it. This will enable us to also avoid notifiers for the future free page reporting which will need to avoid retriggering page reporting when returning pages that have been reported on. Signed-off-by:
Alexander Duyck <alexander.h.duyck@linux.intel.com> Signed-off-by:
David Hildenbrand <david@redhat.com> Acked-by:
Mel Gorman <mgorman@techsingularity.net> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Luiz Capitulino <lcapitulino@redhat.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: Michael S. Tsirkin <mst@redhat.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Nitesh Narayan Lal <nitesh@redhat.com> Cc: Oscar Salvador <osalvador@suse.de> Cc: Pankaj Gupta <pagupta@redhat.com> Cc: Paolo Bonzini <pbonzini@redhat.com> Cc: Rik van Riel <riel@surriel.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Wei Wang <wei.w.wang@intel.com> Cc: Yang Zhang <yang.zhang.wz@gmail.com> Cc: wei qi <weiqi4@huawei.com> Link: http://lkml.kernel.org/r/20200211224624.29318.89287.stgit@localhost.localdomain Signed-off-by:
-
- Apr 02, 2020
-
-
Rik van Riel authored
Patch series "fix THP migration for CMA allocations", v2. Transparent huge pages are allocated with __GFP_MOVABLE, and can end up in CMA memory blocks. Transparent huge pages also have most of the infrastructure in place to allow migration. However, a few pieces were missing, causing THP migration to fail when attempting to use CMA to allocate 1GB hugepages. With these patches in place, THP migration from CMA blocks seems to work, both for anonymous THPs and for tmpfs/shmem THPs. This patch (of 2): Add information to struct compact_control to indicate that the allocator would really like to clear out this specific part of memory, used by for example CMA. Signed-off-by:
Rik van Riel <riel@surriel.com> Signed-off-by:
-
Mateusz Nosek authored
This patch makes ALLOC_KSWAPD equal to __GFP_KSWAPD_RECLAIM (cast to int). Thanks to that code like: if (gfp_mask & __GFP_KSWAPD_RECLAIM) alloc_flags |= ALLOC_KSWAPD; can be changed to: alloc_flags |= (__force int) (gfp_mask &__GFP_KSWAPD_RECLAIM); Thanks to this one branch less is generated in the assembly. In case of ALLOC_KSWAPD flag two branches are saved, first one in code that always executes in the beginning of page allocation and the second one in loop in page allocator slowpath. Signed-off-by:Mateusz Nosek <mateusznosek0@gmail.com> Signed-off-by:
-
Peter Xu authored
The idea comes from a discussion between Linus and Andrea [1]. Before this patch we only allow a page fault to retry once. We achieved this by clearing the FAULT_FLAG_ALLOW_RETRY flag when doing handle_mm_fault() the second time. This was majorly used to avoid unexpected starvation of the system by looping over forever to handle the page fault on a single page. However that should hardly happen, and after all for each code path to return a VM_FAULT_RETRY we'll first wait for a condition (during which time we should possibly yield the cpu) to happen before VM_FAULT_RETRY is really returned. This patch removes the restriction by keeping the FAULT_FLAG_ALLOW_RETRY flag when we receive VM_FAULT_RETRY. It means that the page fault handler now can retry the page fault for multiple times if necessary without the need to generate another page fault event. Meanwhile we still keep the FAULT_FLAG_TRIED flag so page fault handler can still identify whether a page fault is the first attempt or not. Then we'll have these combinations of fault flags (only considering ALLOW_RETRY flag and TRIED flag): - ALLOW_RETRY and !TRIED: this means the page fault allows to retry, and this is the first try - ALLOW_RETRY and TRIED: this means the page fault allows to retry, and this is not the first try - !ALLOW_RETRY and !TRIED: this means the page fault does not allow to retry at all - !ALLOW_RETRY and TRIED: this is forbidden and should never be used In existing code we have multiple places that has taken special care of the first condition above by checking against (fault_flags & FAULT_FLAG_ALLOW_RETRY). This patch introduces a simple helper to detect the first retry of a page fault by checking against both (fault_flags & FAULT_FLAG_ALLOW_RETRY) and !(fault_flag & FAULT_FLAG_TRIED) because now even the 2nd try will have the ALLOW_RETRY set, then use that helper in all existing special paths. One example is in __lock_page_or_retry(), now we'll drop the mmap_sem only in the first attempt of page fault and we'll keep it in follow up retries, so old locking behavior will be retained. This will be a nice enhancement for current code [2] at the same time a supporting material for the future userfaultfd-writeprotect work, since in that work there will always be an explicit userfault writeprotect retry for protected pages, and if that cannot resolve the page fault (e.g., when userfaultfd-writeprotect is used in conjunction with swapped pages) then we'll possibly need a 3rd retry of the page fault. It might also benefit other potential users who will have similar requirement like userfault write-protection. GUP code is not touched yet and will be covered in follow up patch. Please read the thread below for more information. [1] https://lore.kernel.org/lkml/20171102193644.GB22686@redhat.com/ [2] https://lore.kernel.org/lkml/20181230154648.GB9832@redhat.com/ Suggested-by:Andrea Arcangeli <aarcange@redhat.com> Signed-off-by:
Peter Xu <peterx@redhat.com> Signed-off-by:
Brian Geffon <bgeffon@google.com> Cc: Bobby Powers <bobbypowers@gmail.com> Cc: David Hildenbrand <david@redhat.com> Cc: Denis Plotnikov <dplotnikov@virtuozzo.com> Cc: "Dr . David Alan Gilbert" <dgilbert@redhat.com> Cc: Hugh Dickins <hughd@google.com> Cc: Jerome Glisse <jglisse@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: "Kirill A . Shutemov" <kirill@shutemov.name> Cc: Martin Cracauer <cracauer@cons.org> Cc: Marty McFadden <mcfadden8@llnl.gov> Cc: Matthew Wilcox <willy@infradead.org> Cc: Maya Gokhale <gokhale2@llnl.gov> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Kravetz <mike.kravetz@oracle.com> Cc: Mike Rapoport <rppt@linux.vnet.ibm.com> Cc: Pavel Emelyanov <xemul@openvz.org> Link: http://lkml.kernel.org/r/20200220160246.9790-1-peterx@redhat.com Signed-off-by:
-
Yang Shi authored
When backporting commit 9c4e6b1a ("mm, mlock, vmscan: no more skipping pagevecs") to our 4.9 kernel, our test bench noticed around 10% down with a couple of vm-scalability's test cases (lru-file-readonce, lru-file-readtwice and lru-file-mmap-read). I didn't see that much down on my VM (32c-64g-2nodes). It might be caused by the test configuration, which is 32c-256g with NUMA disabled and the tests were run in root memcg, so the tests actually stress only one inactive and active lru. It sounds not very usual in mordern production environment. That commit did two major changes: 1. Call page_evictable() 2. Use smp_mb to force the PG_lru set visible It looks they contribute the most overhead. The page_evictable() is a function which does function prologue and epilogue, and that was used by page reclaim path only. However, lru add is a very hot path, so it sounds better to make it inline. However, it calls page_mapping() which is not inlined either, but the disassemble shows it doesn't do push and pop operations and it sounds not very straightforward to inline it. Other than this, it sounds smp_mb() is not necessary for x86 since SetPageLRU is atomic which enforces memory barrier already, replace it with smp_mb__after_atomic() in the following patch. With the two fixes applied, the tests can get back around 5% on that test bench and get back normal on my VM. Since the test bench configuration is not that usual and I also saw around 6% up on the latest upstream, so it sounds good enough IMHO. The below is test data (lru-file-readtwice throughput) against the v5.6-rc4: mainline w/ inline fix 150MB 154MB With this patch the throughput gets 2.67% up. The data with using smp_mb__after_atomic() is showed in the following patch. Shakeel Butt did the below test: On a real machine with limiting the 'dd' on a single node and reading 100 GiB sparse file (less than a single node). Just ran a single instance to not cause the lru lock contention. The cmdline used is "dd if=file-100GiB of=/dev/null bs=4k". Ran the cmd 10 times with drop_caches in between and measured the time it took. Without patch: 56.64143 +- 0.672 sec With patches: 56.10 +- 0.21 sec [akpm@linux-foundation.org: move page_evictable() to internal.h] Fixes: 9c4e6b1a ("mm, mlock, vmscan: no more skipping pagevecs") Signed-off-by:
Yang Shi <yang.shi@linux.alibaba.com> Signed-off-by:
Shakeel Butt <shakeelb@google.com> Reviewed-by:
-
- Dec 01, 2019
-
-
Mel Gorman authored
Memory hotplug needs to be able to reset and reinit the pcpu allocator batch and high limits but this action is internal to the VM. Move the declaration to internal.h Link: http://lkml.kernel.org/r/20191021094808.28824-4-mgorman@techsingularity.net Signed-off-by:
-
Wei Yang authored
Now we use rb_parent to get next, while this is not necessary. When prev is NULL, this means vma should be the first element in the list. Then next should be current first one (mm->mmap), no matter whether we have parent or not. After removing it, the code shows the beauty of symmetry. Link: http://lkml.kernel.org/r/20190813032656.16625-1-richardw.yang@linux.intel.com Signed-off-by:
Wei Yang <richardw.yang@linux.intel.com> Acked-by:
-
Wei Yang authored
Just make the code a little easier to read. Link: http://lkml.kernel.org/r/20191006012636.31521-3-richardw.yang@linux.intel.com Signed-off-by:
-
Johannes Weiner authored
One of our services is observing hanging ps/top/etc under heavy write IO, and the task states show this is an mmap_sem priority inversion: A write fault is holding the mmap_sem in read-mode and waiting for (heavily cgroup-limited) IO in balance_dirty_pages(): balance_dirty_pages+0x724/0x905 balance_dirty_pages_ratelimited+0x254/0x390 fault_dirty_shared_page.isra.96+0x4a/0x90 do_wp_page+0x33e/0x400 __handle_mm_fault+0x6f0/0xfa0 handle_mm_fault+0xe4/0x200 __do_page_fault+0x22b/0x4a0 page_fault+0x45/0x50 Somebody tries to change the address space, contending for the mmap_sem in write-mode: call_rwsem_down_write_failed_killable+0x13/0x20 do_mprotect_pkey+0xa8/0x330 SyS_mprotect+0xf/0x20 do_syscall_64+0x5b/0x100 entry_SYSCALL_64_after_hwframe+0x3d/0xa2 The waiting writer locks out all subsequent readers to avoid lock starvation, and several threads can be seen hanging like this: call_rwsem_down_read_failed+0x14/0x30 proc_pid_cmdline_read+0xa0/0x480 __vfs_read+0x23/0x140 vfs_read+0x87/0x130 SyS_read+0x42/0x90 do_syscall_64+0x5b/0x100 entry_SYSCALL_64_after_hwframe+0x3d/0xa2 To fix this, do what we do for cache read faults already: drop the mmap_sem before calling into anything IO bound, in this case the balance_dirty_pages() function, and return VM_FAULT_RETRY. Link: http://lkml.kernel.org/r/20190924194238.GA29030@cmpxchg.org Signed-off-by:Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Hillf Danton <hdanton@sina.com> Cc: Hugh Dickins <hughd@google.com> Signed-off-by:
-
- Sep 26, 2019
-
-
Minchan Kim authored
Patch series "Introduce MADV_COLD and MADV_PAGEOUT", v7. - Background The Android terminology used for forking a new process and starting an app from scratch is a cold start, while resuming an existing app is a hot start. While we continually try to improve the performance of cold starts, hot starts will always be significantly less power hungry as well as faster so we are trying to make hot start more likely than cold start. To increase hot start, Android userspace manages the order that apps should be killed in a process called ActivityManagerService. ActivityManagerService tracks every Android app or service that the user could be interacting with at any time and translates that into a ranked list for lmkd(low memory killer daemon). They are likely to be killed by lmkd if the system has to reclaim memory. In that sense they are similar to entries in any other cache. Those apps are kept alive for opportunistic performance improvements but those performance improvements will vary based on the memory requirements of individual workloads. - Problem Naturally, cached apps were dominant consumers of memory on the system. However, they were not significant consumers of swap even though they are good candidate for swap. Under investigation, swapping out only begins once the low zone watermark is hit and kswapd wakes up, but the overall allocation rate in the system might trip lmkd thresholds and cause a cached process to be killed(we measured performance swapping out vs. zapping the memory by killing a process. Unsurprisingly, zapping is 10x times faster even though we use zram which is much faster than real storage) so kill from lmkd will often satisfy the high zone watermark, resulting in very few pages actually being moved to swap. - Approach The approach we chose was to use a new interface to allow userspace to proactively reclaim entire processes by leveraging platform information. This allowed us to bypass the inaccuracy of the kernel’s LRUs for pages that are known to be cold from userspace and to avoid races with lmkd by reclaiming apps as soon as they entered the cached state. Additionally, it could provide many chances for platform to use much information to optimize memory efficiency. To achieve the goal, the patchset introduce two new options for madvise. One is MADV_COLD which will deactivate activated pages and the other is MADV_PAGEOUT which will reclaim private pages instantly. These new options complement MADV_DONTNEED and MADV_FREE by adding non-destructive ways to gain some free memory space. MADV_PAGEOUT is similar to MADV_DONTNEED in a way that it hints the kernel that memory region is not currently needed and should be reclaimed immediately; MADV_COLD is similar to MADV_FREE in a way that it hints the kernel that memory region is not currently needed and should be reclaimed when memory pressure rises. This patch (of 5): When a process expects no accesses to a certain memory range, it could give a hint to kernel that the pages can be reclaimed when memory pressure happens but data should be preserved for future use. This could reduce workingset eviction so it ends up increasing performance. This patch introduces the new MADV_COLD hint to madvise(2) syscall. MADV_COLD can be used by a process to mark a memory range as not expected to be used in the near future. The hint can help kernel in deciding which pages to evict early during memory pressure. It works for every LRU pages like MADV_[DONTNEED|FREE]. IOW, It moves active file page -> inactive file LRU active anon page -> inacdtive anon LRU Unlike MADV_FREE, it doesn't move active anonymous pages to inactive file LRU's head because MADV_COLD is a little bit different symantic. MADV_FREE means it's okay to discard when the memory pressure because the content of the page is *garbage* so freeing such pages is almost zero overhead since we don't need to swap out and access afterward causes just minor fault. Thus, it would make sense to put those freeable pages in inactive file LRU to compete other used-once pages. It makes sense for implmentaion point of view, too because it's not swapbacked memory any longer until it would be re-dirtied. Even, it could give a bonus to make them be reclaimed on swapless system. However, MADV_COLD doesn't mean garbage so reclaiming them requires swap-out/in in the end so it's bigger cost. Since we have designed VM LRU aging based on cost-model, anonymous cold pages would be better to position inactive anon's LRU list, not file LRU. Furthermore, it would help to avoid unnecessary scanning if system doesn't have a swap device. Let's start simpler way without adding complexity at this moment. However, keep in mind, too that it's a caveat that workloads with a lot of pages cache are likely to ignore MADV_COLD on anonymous memory because we rarely age anonymous LRU lists. * man-page material MADV_COLD (since Linux x.x) Pages in the specified regions will be treated as less-recently-accessed compared to pages in the system with similar access frequencies. In contrast to MADV_FREE, the contents of the region are preserved regardless of subsequent writes to pages. MADV_COLD cannot be applied to locked pages, Huge TLB pages, or VM_PFNMAP pages. [akpm@linux-foundation.org: resolve conflicts with hmm.git] Link: http://lkml.kernel.org/r/20190726023435.214162-2-minchan@kernel.org Signed-off-by:
Minchan Kim <minchan@kernel.org> Reported-by:
kbuild test robot <lkp@intel.com> Acked-by:
-
- May 30, 2019
-
-
Thomas Gleixner authored
Based on 1 normalized pattern(s): this program is free software you can redistribute it and or modify it under the terms of the gnu general public license as published by the free software foundation either version 2 of the license or at your option any later version extracted by the scancode license scanner the SPDX license identifier GPL-2.0-or-later has been chosen to replace the boilerplate/reference in 3029 file(s). Signed-off-by:
Thomas Gleixner <tglx@linutronix.de> Reviewed-by:
Allison Randal <allison@lohutok.net> Cc: linux-spdx@vger.kernel.org Link: https://lkml.kernel.org/r/20190527070032.746973796@linutronix.de Signed-off-by:
Greg Kroah-Hartman <gregkh@linuxfoundation.org>
-
- Mar 06, 2019
-
-
Mel Gorman authored
Compaction is inherently race-prone as a suitable page freed during compaction can be allocated by any parallel task. This patch uses a capture_control structure to isolate a page immediately when it is freed by a direct compactor in the slow path of the page allocator. The intent is to avoid redundant scanning. 5.0.0-rc1 5.0.0-rc1 selective-v3r17 capture-v3r19 Amean fault-both-1 0.00 ( 0.00%) 0.00 * 0.00%* Amean fault-both-3 2582.11 ( 0.00%) 2563.68 ( 0.71%) Amean fault-both-5 4500.26 ( 0.00%) 4233.52 ( 5.93%) Amean fault-both-7 5819.53 ( 0.00%) 6333.65 ( -8.83%) Amean fault-both-12 9321.18 ( 0.00%) 9759.38 ( -4.70%) Amean fault-both-18 9782.76 ( 0.00%) 10338.76 ( -5.68%) Amean fault-both-24 15272.81 ( 0.00%) 13379.55 * 12.40%* Amean fault-both-30 15121.34 ( 0.00%) 16158.25 ( -6.86%) Amean fault-both-32 18466.67 ( 0.00%) 18971.21 ( -2.73%) Latency is only moderately affected but the devil is in the details. A closer examination indicates that base page fault latency is reduced but latency of huge pages is increased as it takes creater care to succeed. Part of the "problem" is that allocation success rates are close to 100% even when under pressure and compaction gets harder 5.0.0-rc1 5.0.0-rc1 selective-v3r17 capture-v3r19 Percentage huge-3 96.70 ( 0.00%) 98.23 ( 1.58%) Percentage huge-5 96.99 ( 0.00%) 95.30 ( -1.75%) Percentage huge-7 94.19 ( 0.00%) 97.24 ( 3.24%) Percentage huge-12 94.95 ( 0.00%) 97.35 ( 2.53%) Percentage huge-18 96.74 ( 0.00%) 97.30 ( 0.58%) Percentage huge-24 97.07 ( 0.00%) 97.55 ( 0.50%) Percentage huge-30 95.69 ( 0.00%) 98.50 ( 2.95%) Percentage huge-32 96.70 ( 0.00%) 99.27 ( 2.65%) And scan rates are reduced as expected by 6% for the migration scanner and 29% for the free scanner indicating that there is less redundant work. Compaction migrate scanned 20815362 19573286 Compaction free scanned 16352612 11510663 [mgorman@techsingularity.net: remove redundant check] Link: http://lkml.kernel.org/r/20190201143853.GH9565@techsingularity.net Link: http://lkml.kernel.org/r/20190118175136.31341-23-mgorman@techsingularity.net Signed-off-by: -
Mel Gorman authored
As compaction proceeds and creates high-order blocks, the free list search gets less efficient as the larger blocks are used as compaction targets. Eventually, the larger blocks will be behind the migration scanner for partially migrated pageblocks and the search fails. This patch round-robins what orders are searched so that larger blocks can be ignored and find smaller blocks that can be used as migration targets. The overall impact was small on 1-socket but it avoids corner cases where the migration/free scanners meet prematurely or situations where many of the pageblocks encountered by the free scanner are almost full instead of being properly packed. Previous testing had indicated that without this patch there were occasional large spikes in the free scanner without this patch. [dan.carpenter@oracle.com: fix static checker warning] Link: http://lkml.kernel.org/r/20190118175136.31341-20-mgorman@techsingularity.net Signed-off-by:
-
Mel Gorman authored
Pageblocks are marked for skip when no pages are isolated after a scan. However, it's possible to hit corner cases where the migration scanner gets stuck near the boundary between the source and target scanner. Due to pages being migrated in blocks of COMPACT_CLUSTER_MAX, pages that are migrated can be reallocated before the pageblock is complete. The pageblock is not necessarily skipped so it can be rescanned multiple times. Similarly, a pageblock with some dirty/writeback pages may fail to migrate and be rescanned until writeback completes which is wasteful. This patch tracks if a pageblock is being rescanned. If so, then the entire pageblock will be migrated as one operation. This narrows the race window during which pages can be reallocated during migration. Secondly, if there are pages that cannot be isolated then the pageblock will still be fully scanned and marked for skipping. On the second rescan, the pageblock skip is set and the migration scanner makes progress. 5.0.0-rc1 5.0.0-rc1 findfree-v3r16 norescan-v3r16 Amean fault-both-1 0.00 ( 0.00%) 0.00 * 0.00%* Amean fault-both-3 3200.68 ( 0.00%) 3002.07 ( 6.21%) Amean fault-both-5 4847.75 ( 0.00%) 4684.47 ( 3.37%) Amean fault-both-7 6658.92 ( 0.00%) 6815.54 ( -2.35%) Amean fault-both-12 11077.62 ( 0.00%) 10864.02 ( 1.93%) Amean fault-both-18 12403.97 ( 0.00%) 12247.52 ( 1.26%) Amean fault-both-24 15607.10 ( 0.00%) 15683.99 ( -0.49%) Amean fault-both-30 18752.27 ( 0.00%) 18620.02 ( 0.71%) Amean fault-both-32 21207.54 ( 0.00%) 19250.28 * 9.23%* 5.0.0-rc1 5.0.0-rc1 findfree-v3r16 norescan-v3r16 Percentage huge-3 96.86 ( 0.00%) 95.00 ( -1.91%) Percentage huge-5 93.72 ( 0.00%) 94.22 ( 0.53%) Percentage huge-7 94.31 ( 0.00%) 92.35 ( -2.08%) Percentage huge-12 92.66 ( 0.00%) 91.90 ( -0.82%) Percentage huge-18 91.51 ( 0.00%) 89.58 ( -2.11%) Percentage huge-24 90.50 ( 0.00%) 90.03 ( -0.52%) Percentage huge-30 91.57 ( 0.00%) 89.14 ( -2.65%) Percentage huge-32 91.00 ( 0.00%) 90.58 ( -0.46%) Negligible difference but this was likely a case when the specific corner case was not hit. A previous run of the same patch based on an earlier iteration of the series showed large differences where migration rates could be halved when the corner case was hit. The specific corner case where migration scan rates go through the roof was due to a dirty/writeback pageblock located at the boundary of the migration/free scanner did not happen in this case. When it does happen, the scan rates multipled by massive margins. Link: http://lkml.kernel.org/r/20190118175136.31341-13-mgorman@techsingularity.net Signed-off-by: -
Mel Gorman authored
The migration scanner is a linear scan of a zone with a potentiall large search space. Furthermore, many pageblocks are unusable such as those filled with reserved pages or partially filled with pages that cannot migrate. These still get scanned in the common case of allocating a THP and the cost accumulates. The patch uses a partial search of the free lists to locate a migration source candidate that is marked as MOVABLE when allocating a THP. It prefers picking a block with a larger number of free pages already on the basis that there are fewer pages to migrate to free the entire block. The lowest PFN found during searches is tracked as the basis of the start for the linear search after the first search of the free list fails. After the search, the free list is shuffled so that the next search will not encounter the same page. If the search fails then the subsequent searches will be shorter and the linear scanner is used. If this search fails, or if the request is for a small or unmovable/reclaimable allocation then the linear scanner is still used. It is somewhat pointless to use the list search in those cases. Small free pages must be used for the search and there is no guarantee that movable pages are located within that block that are contiguous. 5.0.0-rc1 5.0.0-rc1 noboost-v3r10 findmig-v3r15 Amean fault-both-3 3771.41 ( 0.00%) 3390.40 ( 10.10%) Amean fault-both-5 5409.05 ( 0.00%) 5082.28 ( 6.04%) Amean fault-both-7 7040.74 ( 0.00%) 7012.51 ( 0.40%) Amean fault-both-12 11887.35 ( 0.00%) 11346.63 ( 4.55%) Amean fault-both-18 16718.19 ( 0.00%) 15324.19 ( 8.34%) Amean fault-both-24 21157.19 ( 0.00%) 16088.50 * 23.96%* Amean fault-both-30 21175.92 ( 0.00%) 18723.42 * 11.58%* Amean fault-both-32 21339.03 ( 0.00%) 18612.01 * 12.78%* 5.0.0-rc1 5.0.0-rc1 noboost-v3r10 findmig-v3r15 Percentage huge-3 86.50 ( 0.00%) 89.83 ( 3.85%) Percentage huge-5 92.52 ( 0.00%) 91.96 ( -0.61%) Percentage huge-7 92.44 ( 0.00%) 92.85 ( 0.44%) Percentage huge-12 92.98 ( 0.00%) 92.74 ( -0.25%) Percentage huge-18 91.70 ( 0.00%) 91.71 ( 0.02%) Percentage huge-24 91.59 ( 0.00%) 92.13 ( 0.60%) Percentage huge-30 90.14 ( 0.00%) 93.79 ( 4.04%) Percentage huge-32 90.03 ( 0.00%) 91.27 ( 1.37%) This shows an improvement in allocation latencies with similar allocation success rates. While not presented, there was a 31% reduction in migration scanning and a 8% reduction on system CPU usage. A 2-socket machine showed similar benefits. [mgorman@techsingularity.net: several fixes] Link: http://lkml.kernel.org/r/20190204120111.GL9565@techsingularity.net [vbabka@suse.cz: migrate block that was found-fast, some optimisations] Link: http://lkml.kernel.org/r/20190118175136.31341-10-mgorman@techsingularity.net Signed-off-by: -
Mel Gorman authored
When compaction is finishing, it uses a flag to ensure the pageblock is complete but it makes sense to always complete migration of a pageblock. Minimally, skip information is based on a pageblock and partially scanned pageblocks may incur more scanning in the future. The pageblock skip handling also becomes more strict later in the series and the hint is more useful if a complete pageblock was always scanned. The potentially impacts latency as more scanning is done but it's not a consistent win or loss as the scanning is not always a high percentage of the pageblock and sometimes it is offset by future reductions in scanning. Hence, the results are not presented this time due to a misleading mix of gains/losses without any clear pattern. However, full scanning of the pageblock is important for later patches. Link: http://lkml.kernel.org/r/20190118175136.31341-8-mgorman@techsingularity.net Signed-off-by:
-
Mel Gorman authored
The last_migrated_pfn field is a bit dubious as to whether it really helps but either way, the information from it can be inferred without increasing the size of compact_control so remove the field. Link: http://lkml.kernel.org/r/20190118175136.31341-4-mgorman@techsingularity.net Signed-off-by:
-
Mel Gorman authored
compact_control spans two cache lines with write-intensive lines on both. Rearrange so the most write-intensive fields are in the same cache line. This has a negligible impact on the overall performance of compaction and is more a tidying exercise than anything. Link: http://lkml.kernel.org/r/20190118175136.31341-3-mgorman@techsingularity.net Signed-off-by:
-
Mel Gorman authored
Patch series "Increase success rates and reduce latency of compaction", v3. This series reduces scan rates and success rates of compaction, primarily by using the free lists to shorten scans, better controlling of skip information and whether multiple scanners can target the same block and capturing pageblocks before being stolen by parallel requests. The series is based on mmotm from January 9th, 2019 with the previous compaction series reverted. I'm mostly using thpscale to measure the impact of the series. The benchmark creates a large file, maps it, faults it, punches holes in the mapping so that the virtual address space is fragmented and then tries to allocate THP. It re-executes for different numbers of threads. From a fragmentation perspective, the workload is relatively benign but it does stress compaction. The overall impact on latencies for a 1-socket machine is baseline patches Amean fault-both-3 3832.09 ( 0.00%) 2748.56 * 28.28%* Amean fault-both-5 4933.06 ( 0.00%) 4255.52 ( 13.73%) Amean fault-both-7 7017.75 ( 0.00%) 6586.93 ( 6.14%) Amean fault-both-12 11610.51 ( 0.00%) 9162.34 * 21.09%* Amean fault-both-18 17055.85 ( 0.00%) 11530.06 * 32.40%* Amean fault-both-24 19306.27 ( 0.00%) 17956.13 ( 6.99%) Amean fault-both-30 22516.49 ( 0.00%) 15686.47 * 30.33%* Amean fault-both-32 23442.93 ( 0.00%) 16564.83 * 29.34%* The allocation success rates are much improved baseline patches Percentage huge-3 85.99 ( 0.00%) 97.96 ( 13.92%) Percentage huge-5 88.27 ( 0.00%) 96.87 ( 9.74%) Percentage huge-7 85.87 ( 0.00%) 94.53 ( 10.09%) Percentage huge-12 82.38 ( 0.00%) 98.44 ( 19.49%) Percentage huge-18 83.29 ( 0.00%) 99.14 ( 19.04%) Percentage huge-24 81.41 ( 0.00%) 97.35 ( 19.57%) Percentage huge-30 80.98 ( 0.00%) 98.05 ( 21.08%) Percentage huge-32 80.53 ( 0.00%) 97.06 ( 20.53%) That's a nearly perfect allocation success rate. The biggest impact is on the scan rates Compaction migrate scanned 55893379 19341254 Compaction free scanned 474739990 11903963 The number of pages scanned for migration was reduced by 65% and the free scanner was reduced by 97.5%. So much less work in exchange for lower latency and better success rates. The series was also evaluated using a workload that heavily fragments memory but the benefits there are also significant, albeit not presented. It was commented that we should be rethinking scanning entirely and to a large extent I agree. However, to achieve that you need a lot of this series in place first so it's best to make the linear scanners as best as possible before ripping them out. This patch (of 22): The isolate and migrate scanners should never isolate more than a pageblock of pages so unsigned int is sufficient saving 8 bytes on a 64-bit build. Link: http://lkml.kernel.org/r/20190118175136.31341-2-mgorman@techsingularity.net Signed-off-by:
-
Arun KS authored
When freeing pages are done with higher order, time spent on coalescing pages by buddy allocator can be reduced. With section size of 256MB, hot add latency of a single section shows improvement from 50-60 ms to less than 1 ms, hence improving the hot add latency by 60 times. Modify external providers of online callback to align with the change. [arunks@codeaurora.org: v11] Link: http://lkml.kernel.org/r/1547792588-18032-1-git-send-email-arunks@codeaurora.org [akpm@linux-foundation.org: remove unused local, per Arun] [akpm@linux-foundation.org: avoid return of void-returning __free_pages_core(), per Oscar] [akpm@linux-foundation.org: fix it for mm-convert-totalram_pages-and-totalhigh_pages-variables-to-atomic.patch] [arunks@codeaurora.org: v8] Link: http://lkml.kernel.org/r/1547032395-24582-1-git-send-email-arunks@codeaurora.org [arunks@codeaurora.org: v9] Link: http://lkml.kernel.org/r/1547098543-26452-1-git-send-email-arunks@codeaurora.org Link: http://lkml.kernel.org/r/1538727006-5727-1-git-send-email-arunks@codeaurora.org Signed-off-by:
Arun KS <arunks@codeaurora.org> Reviewed-by:
Oscar Salvador <osalvador@suse.de> Reviewed-by:
-
- Dec 28, 2018
-
-
Mel Gorman authored
This is a preparation patch that copies the GFP flag __GFP_KSWAPD_RECLAIM into alloc_flags. This is a preparation patch only that avoids having to pass gfp_mask through a long callchain in a future patch. Note that the setting in the fast path happens in alloc_flags_nofragment() and it may be claimed that this has nothing to do with ALLOC_NO_FRAGMENT. That's true in this patch but is not true later so it's done now for easier review to show where the flag needs to be recorded. No functional change. [mgorman@techsingularity.net: ALLOC_KSWAPD flag needs to be applied in the !CONFIG_ZONE_DMA32 case] Link: http://lkml.kernel.org/r/20181126143503.GO23260@techsingularity.net Link: http://lkml.kernel.org/r/20181123114528.28802-4-mgorman@techsingularity.net Signed-off-by:
-
Mel Gorman authored
Patch series "Fragmentation avoidance improvements", v5. It has been noted before that fragmentation avoidance (aka anti-fragmentation) is not perfect. Given sufficient time or an adverse workload, memory gets fragmented and the long-term success of high-order allocations degrades. This series defines an adverse workload, a definition of external fragmentation events (including serious) ones and a series that reduces the level of those fragmentation events. The details of the workload and the consequences are described in more detail in the changelogs. However, from patch 1, this is a high-level summary of the adverse workload. The exact details are found in the mmtests implementation. The broad details of the workload are as follows; 1. Create an XFS filesystem (not specified in the configuration but done as part of the testing for this patch) 2. Start 4 fio threads that write a number of 64K files inefficiently. Inefficiently means that files are created on first access and not created in advance (fio parameterr create_on_open=1) and fallocate is not used (fallocate=none). With multiple IO issuers this creates a mix of slab and page cache allocations over time. The total size of the files is 150% physical memory so that the slabs and page cache pages get mixed 3. Warm up a number of fio read-only threads accessing the same files created in step 2. This part runs for the same length of time it took to create the files. It'll fault back in old data and further interleave slab and page cache allocations. As it's now low on memory due to step 2, fragmentation occurs as pageblocks get stolen. 4. While step 3 is still running, start a process that tries to allocate 75% of memory as huge pages with a number of threads. The number of threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP threads contending with fio, any other threads or forcing cross-NUMA scheduling. Note that the test has not been used on a machine with less than 8 cores. The benchmark records whether huge pages were allocated and what the fault latency was in microseconds 5. Measure the number of events potentially causing external fragmentation, the fault latency and the huge page allocation success rate. 6. Cleanup Overall the series reduces external fragmentation causing events by over 94% on 1 and 2 socket machines, which in turn impacts high-order allocation success rates over the long term. There are differences in latencies and high-order allocation success rates. Latencies are a mixed bag as they are vulnerable to exact system state and whether allocations succeeded so they are treated as a secondary metric. Patch 1 uses lower zones if they are populated and have free memory instead of fragmenting a higher zone. It's special cased to handle a Normal->DMA32 fallback with the reasons explained in the changelog. Patch 2-4 boosts watermarks temporarily when an external fragmentation event occurs. kswapd wakes to reclaim a small amount of old memory and then wakes kcompactd on completion to recover the system slightly. This introduces some overhead in the slowpath. The level of boosting can be tuned or disabled depending on the tolerance for fragmentation vs allocation latency. Patch 5 stalls some movable allocation requests to let kswapd from patch 4 make some progress. The duration of the stalls is very low but it is possible to tune the system to avoid fragmentation events if larger stalls can be tolerated. The bulk of the improvement in fragmentation avoidance is from patches 1-4 but patch 5 can deal with a rare corner case and provides the option of tuning a system for THP allocation success rates in exchange for some stalls to control fragmentation. This patch (of 5): The page allocator zone lists are iterated based on the watermarks of each zone which does not take anti-fragmentation into account. On x86, node 0 may have multiple zones while other nodes have one zone. A consequence is that tasks running on node 0 may fragment ZONE_NORMAL even though ZONE_DMA32 has plenty of free memory. This patch special cases the allocator fast path such that it'll try an allocation from a lower local zone before fragmenting a higher zone. In this case, stealing of pageblocks or orders larger than a pageblock are still allowed in the fast path as they are uninteresting from a fragmentation point of view. This was evaluated using a benchmark designed to fragment memory before attempting THP allocations. It's implemented in mmtests as the following configurations configs/config-global-dhp__workload_thpfioscale configs/config-global-dhp__workload_thpfioscale-defrag configs/config-global-dhp__workload_thpfioscale-madvhugepage e.g. from mmtests ./run-mmtests.sh --run-monitor --config configs/config-global-dhp__workload_thpfioscale test-run-1 The broad details of the workload are as follows; 1. Create an XFS filesystem (not specified in the configuration but done as part of the testing for this patch). 2. Start 4 fio threads that write a number of 64K files inefficiently. Inefficiently means that files are created on first access and not created in advance (fio parameter create_on_open=1) and fallocate is not used (fallocate=none). With multiple IO issuers this creates a mix of slab and page cache allocations over time. The total size of the files is 150% physical memory so that the slabs and page cache pages get mixed. 3. Warm up a number of fio read-only processes accessing the same files created in step 2. This part runs for the same length of time it took to create the files. It'll refault old data and further interleave slab and page cache allocations. As it's now low on memory due to step 2, fragmentation occurs as pageblocks get stolen. 4. While step 3 is still running, start a process that tries to allocate 75% of memory as huge pages with a number of threads. The number of threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP threads contending with fio, any other threads or forcing cross-NUMA scheduling. Note that the test has not been used on a machine with less than 8 cores. The benchmark records whether huge pages were allocated and what the fault latency was in microseconds. 5. Measure the number of events potentially causing external fragmentation, the fault latency and the huge page allocation success rate. 6. Cleanup the test files. Note that due to the use of IO and page cache that this benchmark is not suitable for running on large machines where the time to fragment memory may be excessive. Also note that while this is one mix that generates fragmentation that it's not the only mix that generates fragmentation. Differences in workload that are more slab-intensive or whether SLUB is used with high-order pages may yield different results. When the page allocator fragments memory, it records the event using the mm_page_alloc_extfrag ftrace event. If the fallback_order is smaller than a pageblock order (order-9 on 64-bit x86) then it's considered to be an "external fragmentation event" that may cause issues in the future. Hence, the primary metric here is the number of external fragmentation events that occur with order < 9. The secondary metric is allocation latency and huge page allocation success rates but note that differences in latencies and what the success rate also can affect the number of external fragmentation event which is why it's a secondary metric. 1-socket Skylake machine config-global-dhp__workload_thpfioscale XFS (no special madvise) 4 fio threads, 1 THP allocating thread -------------------------------------- 4.20-rc3 extfrag events < order 9: 804694 4.20-rc3+patch: 408912 (49% reduction) thpfioscale Fault Latencies 4.20.0-rc3 4.20.0-rc3 vanilla lowzone-v5r8 Amean fault-base-1 662.92 ( 0.00%) 653.58 * 1.41%* Amean fault-huge-1 0.00 ( 0.00%) 0.00 ( 0.00%) 4.20.0-rc3 4.20.0-rc3 vanilla lowzone-v5r8 Percentage huge-1 0.00 ( 0.00%) 0.00 ( 0.00%) Fault latencies are slightly reduced while allocation success rates remain at zero as this configuration does not make any special effort to allocate THP and fio is heavily active at the time and either filling memory or keeping pages resident. However, a 49% reduction of serious fragmentation events reduces the changes of external fragmentation being a problem in the future. Vlastimil asked during review for a breakdown of the allocation types that are falling back. vanilla 3816 MIGRATE_UNMOVABLE 800845 MIGRATE_MOVABLE 33 MIGRATE_UNRECLAIMABLE patch 735 MIGRATE_UNMOVABLE 408135 MIGRATE_MOVABLE 42 MIGRATE_UNRECLAIMABLE The majority of the fallbacks are due to movable allocations and this is consistent for the workload throughout the series so will not be presented again as the primary source of fallbacks are movable allocations. Movable fallbacks are sometimes considered "ok" to fallback because they can be migrated. The problem is that they can fill an unmovable/reclaimable pageblock causing those allocations to fallback later and polluting pageblocks with pages that cannot move. If there is a movable fallback, it is pretty much guaranteed to affect an unmovable/reclaimable pageblock and while it might not be enough to actually cause a unmovable/reclaimable fallback in the future, we cannot know that in advance so the patch takes the only option available to it. Hence, it's important to control them. This point is also consistent throughout the series and will not be repeated. 1-socket Skylake machine global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE) ----------------------------------------------------------------- 4.20-rc3 extfrag events < order 9: 291392 4.20-rc3+patch: 191187 (34% reduction) thpfioscale Fault Latencies 4.20.0-rc3 4.20.0-rc3 vanilla lowzone-v5r8 Amean fault-base-1 1495.14 ( 0.00%) 1467.55 ( 1.85%) Amean fault-huge-1 1098.48 ( 0.00%) 1127.11 ( -2.61%) thpfioscale Percentage Faults Huge 4.20.0-rc3 4.20.0-rc3 vanilla lowzone-v5r8 Percentage huge-1 78.57 ( 0.00%) 77.64 ( -1.18%) Fragmentation events were reduced quite a bit although this is known to be a little variable. The latencies and allocation success rates are similar but they were already quite high. 2-socket Haswell machine config-global-dhp__workload_thpfioscale XFS (no special madvise) 4 fio threads, 5 THP allocating threads ---------------------------------------------------------------- 4.20-rc3 extfrag events < order 9: 215698 4.20-rc3+patch: 200210 (7% reduction) thpfioscale Fault Latencies 4.20.0-rc3 4.20.0-rc3 vanilla lowzone-v5r8 Amean fault-base-5 1350.05 ( 0.00%) 1346.45 ( 0.27%) Amean fault-huge-5 4181.01 ( 0.00%) 3418.60 ( 18.24%) 4.20.0-rc3 4.20.0-rc3 vanilla lowzone-v5r8 Percentage huge-5 1.15 ( 0.00%) 0.78 ( -31.88%) The reduction of external fragmentation events is slight and this is partially due to the removal of __GFP_THISNODE in commit ac5b2c18 ("mm: thp: relax __GFP_THISNODE for MADV_HUGEPAGE mappings") as THP allocations can now spill over to remote nodes instead of fragmenting local memory. 2-socket Haswell machine global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE) ----------------------------------------------------------------- 4.20-rc3 extfrag events < order 9: 166352 4.20-rc3+patch: 147463 (11% reduction) thpfioscale Fault Latencies 4.20.0-rc3 4.20.0-rc3 vanilla lowzone-v5r8 Amean fault-base-5 6138.97 ( 0.00%) 6217.43 ( -1.28%) Amean fault-huge-5 2294.28 ( 0.00%) 3163.33 * -37.88%* thpfioscale Percentage Faults Huge 4.20.0-rc3 4.20.0-rc3 vanilla lowzone-v5r8 Percentage huge-5 96.82 ( 0.00%) 95.14 ( -1.74%) There was a slight reduction in external fragmentation events although the latencies were higher. The allocation success rate is high enough that the system is struggling and there is quite a lot of parallel reclaim and compaction activity. There is also a certain degree of luck on whether processes start on node 0 or not for this patch but the relevance is reduced later in the series. Overall, the patch reduces the number of external fragmentation causing events so the success of THP over long periods of time would be improved for this adverse workload. Link: http://lkml.kernel.org/r/20181123114528.28802-2-mgorman@techsingularity.net Signed-off-by:
-
