- 03 Jul, 2013 40 commits
-
-
Mel Gorman authored
Andrew Perepechko reported a problem whereby pages are being prematurely evicted as the mark_page_accessed() hint is ignored for pages that are currently on a pagevec -- http://www.spinics.net/lists/linux-ext4/msg37340.html . Alexey Lyahkov and Robin Dong have also reported problems recently that could be due to hot pages reaching the end of the inactive list too quickly and be reclaimed. Rather than addressing this on a per-filesystem basis, this series aims to fix the mark_page_accessed() interface by deferring what LRU a page is added to pagevec drain time and allowing mark_page_accessed() to call SetPageActive on a pagevec page. Patch 1 adds two tracepoints for LRU page activation and insertion. Using these processes it's possible to build a model of pages in the LRU that can be processed offline. Patch 2 defers making the decision on what LRU to add a page to until when the pagevec is drained. Patch 3 searches the local pagevec for pages to mark PageActive on mark_page_accessed. The changelog explains why only the local pagevec is examined. Patches 4 and 5 tidy up the API. postmark, a dd-based test and fs-mark both single and threaded mode were run but none of them showed any performance degradation or gain as a result of the patch. Using patch 1, I built a *very* basic model of the LRU to examine offline what the average age of different page types on the LRU were in milliseconds. Of course, capturing the trace distorts the test as it's written to local disk but it does not matter for the purposes of this test. The average age of pages in milliseconds were vanilla deferdrain Average age mapped anon: 1454 1250 Average age mapped file: 127841 155552 Average age unmapped anon: 85 235 Average age unmapped file: 73633 38884 Average age unmapped buffers: 74054 116155 The LRU activity was mostly files which you'd expect for a dd-based workload. Note that the average age of buffer pages is increased by the series and it is expected this is due to the fact that the buffer pages are now getting added to the active list when drained from the pagevecs. Note that the average age of the unmapped file data is decreased as they are still added to the inactive list and are reclaimed before the buffers. There is no guarantee this is a universal win for all workloads and it would be nice if the filesystem people gave some thought as to whether this decision is generally a win or a loss. This patch: Using these tracepoints it is possible to model LRU activity and the average residency of pages of different types. This can be used to debug problems related to premature reclaim of pages of particular types. Signed-off-by:
Mel Gorman <mgorman@suse.de> Reviewed-by:
Rik van Riel <riel@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Alexey Lyahkov <alexey.lyashkov@gmail.com> Cc: Andrew Perepechko <anserper@ya.ru> Cc: Robin Dong <sanbai@taobao.com> Cc: Theodore Tso <tytso@mit.edu> Cc: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Bernd Schubert <bernd.schubert@fastmail.fm> Cc: David Howells <dhowells@redhat.com> Cc: Trond Myklebust <Trond.Myklebust@netapp.com> Signed-off-by:
Andrew Morton <akpm@linux-foundation.org> Signed-off-by:
Linus Torvalds <torvalds@linux-foundation.org>
-
Li Zefan authored
hugetlb cgroup has already been implemented. Signed-off-by:
Li Zefan <lizefan@huawei.com> Acked-by:
KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Acked-by:
Rob Landley <rob@landley.net> Cc: Michal Hocko <mhocko@suse.cz> Cc: Johannes Weiner <hannes@cmpxchg.org> Signed-off-by:
-
HATAYAMA Daisuke authored
This patch introduces mmap_vmcore(). Don't permit writable nor executable mapping even with mprotect() because this mmap() is aimed at reading crash dump memory. Non-writable mapping is also requirement of remap_pfn_range() when mapping linear pages on non-consecutive physical pages; see is_cow_mapping(). Set VM_MIXEDMAP flag to remap memory by remap_pfn_range and by remap_vmalloc_range_pertial at the same time for a single vma. do_munmap() can correctly clean partially remapped vma with two functions in abnormal case. See zap_pte_range(), vm_normal_page() and their comments for details. On x86-32 PAE kernels, mmap() supports at most 16TB memory only. This limitation comes from the fact that the third argument of remap_pfn_range(), pfn, is of 32-bit length on x86-32: unsigned long. [akpm@linux-foundation.org: use min(), switch to conventional error-unwinding approach] Signed-off-by:
HATAYAMA Daisuke <d.hatayama@jp.fujitsu.com> Acked-by:
Vivek Goyal <vgoyal@redhat.com> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Atsushi Kumagai <kumagai-atsushi@mxc.nes.nec.co.jp> Cc: Lisa Mitchell <lisa.mitchell@hp.com> Cc: Zhang Yanfei <zhangyanfei@cn.fujitsu.com> Tested-by:
Maxim Uvarov <muvarov@gmail.com> Cc: Arnd Bergmann <arnd@arndb.de> Signed-off-by:
-
HATAYAMA Daisuke authored
The previous patches newly added holes before each chunk of memory and the holes need to be count in vmcore file size. There are two ways to count file size in such a way: 1) suppose m is a poitner to the last vmcore object in vmcore_list. Then file size is (m->offset + m->size), or 2) calculate sum of size of buffers for ELF header, program headers, ELF note segments and objects in vmcore_list. Although 1) is more direct and simpler than 2), 2) seems better in that it reflects internal object structure of /proc/vmcore. Thus, this patch changes get_vmcore_size_elf{64, 32} so that it calculates size in the way of 2). As a result, both get_vmcore_size_elf{64, 32} have the same definition. Merge them as get_vmcore_size. Signed-off-by: -
HATAYAMA Daisuke authored
Now ELF note segment has been copied in the buffer on vmalloc memory. To allow user process to remap the ELF note segment buffer with remap_vmalloc_page, the corresponding VM area object has to have VM_USERMAP flag set. [akpm@linux-foundation.org: use the conventional comment layout] Signed-off-by:
-
HATAYAMA Daisuke authored
The reasons why we don't allocate ELF note segment in the 1st kernel (old memory) on page boundary is to keep backward compatibility for old kernels, and that if doing so, we waste not a little memory due to round-up operation to fit the memory to page boundary since most of the buffers are in per-cpu area. ELF notes are per-cpu, so total size of ELF note segments depends on number of CPUs. The current maximum number of CPUs on x86_64 is 5192, and there's already system with 4192 CPUs in SGI, where total size amounts to 1MB. This can be larger in the near future or possibly even now on another architecture that has larger size of note per a single cpu. Thus, to avoid the case where memory allocation for large block fails, we allocate vmcore objects on vmalloc memory. This patch adds elfnotes_buf and elfnotes_sz variables to keep pointer to the ELF note segment buffer and its size. There's no longer the vmcore object that corresponds to the ELF note segment in vmcore_list. Accordingly, read_vmcore() has new case for ELF note segment and set_vmcore_list_offsets_elf{64,32}() and other helper functions starts calculating offset from sum of size of ELF headers and size of ELF note segment. [akpm@linux-foundation.org: use min(), fix error-path vzalloc() leaks] Signed-off-by: -
HATAYAMA Daisuke authored
We want to allocate ELF note segment buffer on the 2nd kernel in vmalloc space and remap it to user-space in order to reduce the risk that memory allocation fails on system with huge number of CPUs and so with huge ELF note segment that exceeds 11-order block size. Although there's already remap_vmalloc_range for the purpose of remapping vmalloc memory to user-space, we need to specify user-space range via vma. Mmap on /proc/vmcore needs to remap range across multiple objects, so the interface that requires vma to cover full range is problematic. This patch introduces remap_vmalloc_range_partial that receives user-space range as a pair of base address and size and can be used for mmap on /proc/vmcore case. remap_vmalloc_range is rewritten using remap_vmalloc_range_partial. [akpm@linux-foundation.org: use PAGE_ALIGNED()] Signed-off-by:
-
HATAYAMA Daisuke authored
Currently, __find_vmap_area searches for the kernel VM area starting at a given address. This patch changes this behavior so that it searches for the kernel VM area to which the address belongs. This change is needed by remap_vmalloc_range_partial to be introduced in later patch that receives any position of kernel VM area as target address. This patch changes the condition (addr > va->va_start) to the equivalent (addr >= va->va_end) by taking advantage of the fact that each kernel VM area is non-overlapping. Signed-off-by:
KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Vivek Goyal <vgoyal@redhat.com> Cc: Atsushi Kumagai <kumagai-atsushi@mxc.nes.nec.co.jp> Cc: Lisa Mitchell <lisa.mitchell@hp.com> Cc: Zhang Yanfei <zhangyanfei@cn.fujitsu.com> Signed-off-by:
-
HATAYAMA Daisuke authored
vmcore: treat memory chunks referenced by PT_LOAD program header entries in page-size boundary in vmcore_list Treat memory chunks referenced by PT_LOAD program header entries in page-size boundary in vmcore_list. Formally, for each range [start, end], we set up the corresponding vmcore object in vmcore_list to [rounddown(start, PAGE_SIZE), roundup(end, PAGE_SIZE)]. This change affects layout of /proc/vmcore. The gaps generated by the rearrangement are newly made visible to applications as holes. Concretely, they are two ranges [rounddown(start, PAGE_SIZE), start] and [end, roundup(end, PAGE_SIZE)]. Suppose variable m points at a vmcore object in vmcore_list, and variable phdr points at the program header of PT_LOAD type the variable m corresponds to. Then, pictorially: m->offset +---------------+ | hole | phdr->p_offset = +---------------+ m->offset + (paddr - start) | |\ | kernel memory | phdr->p_memsz | |/ +---------------+ | hole | m->offset + m->size +---------------+ where m->offset and m->offset + m->size are always page-size aligned. Signed-off-by: -
HATAYAMA Daisuke authored
Allocate ELF headers on page-size boundary using __get_free_pages() instead of kmalloc(). Later patch will merge PT_NOTE entries into a single unique one and decrease the buffer size actually used. Keep original buffer size in variable elfcorebuf_sz_orig to kfree the buffer later and actually used buffer size with rounded up to page-size boundary in variable elfcorebuf_sz separately. The size of part of the ELF buffer exported from /proc/vmcore is elfcorebuf_sz. The merged, removed PT_NOTE entries, i.e. the range [elfcorebuf_sz, elfcorebuf_sz_orig], is filled with 0. Use size of the ELF headers as an initial offset value in set_vmcore_list_offsets_elf{64,32} and process_ptload_program_headers_elf{64,32} in order to indicate that the offset includes the holes towards the page boundary. As a result, both set_vmcore_list_offsets_elf{64,32} have the same definition. Merge them as set_vmcore_list_offsets. [akpm@linux-foundation.org: add free_elfcorebuf(), cleanups] Signed-off-by: -
HATAYAMA Daisuke authored
Rewrite part of read_vmcore() that reads objects in vmcore_list in the same way as part reading ELF headers, by which some duplicated and redundant codes are removed. Signed-off-by:
-
Andrew Morton authored
To test whether an address is aligned to PAGE_SIZE. Cc: HATAYAMA Daisuke <d.hatayama@jp.fujitsu.com> Cc: "Eric W. Biederman" <ebiederm@xmission.com>, Signed-off-by:
-
Cody P Schafer authored
mmzone.h documents node_size_lock (which pgdat_resize_lock() locks) as follows: * Must be held any time you expect node_start_pfn, node_present_pages * or node_spanned_pages stay constant. [...] So actually hold it when we update node_present_pages in __offline_pages(). [akpm@linux-foundation.org: fix build] Signed-off-by:Cody P Schafer <cody@linux.vnet.ibm.com> Cc: David Rientjes <rientjes@google.com> Signed-off-by:
-
Cody P Schafer authored
mmzone.h documents node_size_lock (which pgdat_resize_lock() locks) as follows: * Must be held any time you expect node_start_pfn, node_present_pages * or node_spanned_pages stay constant. [...] So actually hold it when we update node_present_pages in online_pages(). Signed-off-by: -
Cody P Schafer authored
Signed-off-by:
-
Cody P Schafer authored
Signed-off-by:
David Rientjes <rientjes@google.com> Signed-off-by:
-
Mel Gorman authored
VM page reclaim uses dirty and writeback page states to determine if flushers are cleaning pages too slowly and that page reclaim should stall waiting on flushers to catch up. Page state in NFS is a bit more complex and a clean page can be unreclaimable due to being unstable which is effectively "dirty" from the perspective of the VM from reclaim context. Similarly, if the inode is currently being committed then it's similar to being under writeback. This patch adds a is_dirty_writeback() handled for NFS that checks if a pages backing inode is being committed and should be accounted as writeback and if a page has private state indicating that it is effectively dirty. Signed-off-by:
-
Mel Gorman authored
Page reclaim keeps track of dirty and under writeback pages and uses it to determine if wait_iff_congested() should stall or if kswapd should begin writing back pages. This fails to account for buffer pages that can be under writeback but not PageWriteback which is the case for filesystems like ext3 ordered mode. Furthermore, PageDirty buffer pages can have all the buffers clean and writepage does no IO so it should not be accounted as congested. This patch adds an address_space operation that filesystems may optionally use to check if a page is really dirty or really under writeback. An implementation is provided for for buffer_heads is added and used for block operations and ext3 in ordered mode. By default the page flags are obeyed. Credit goes to Jan Kara for identifying that the page flags alone are not sufficient for ext3 and sanity checking a number of ideas on how the problem could be addressed. Signed-off-by:
-
Mel Gorman authored
Currently a zone will only be marked congested if the underlying BDI is congested but if dirty pages are spread across zones it is possible that an individual zone is full of dirty pages without being congested. The impact is that zone gets scanned very quickly potentially reclaiming really clean pages. This patch treats pages marked for immediate reclaim as congested for the purposes of marking a zone ZONE_CONGESTED and stalling in wait_iff_congested. Signed-off-by:
-
Mel Gorman authored
shrink_inactive_list makes decisions on whether to stall based on the number of dirty pages encountered. The wait_iff_congested() call in shrink_page_list does no such thing and it's arbitrary. This patch moves the decision on whether to set ZONE_CONGESTED and the wait_iff_congested call into shrink_page_list. This keeps all the decisions on whether to stall or not in the one place. Signed-off-by:
-
Mel Gorman authored
In shrink_page_list a decision may be made to stall and flag a zone as ZONE_WRITEBACK so that if a large number of unqueued dirty pages are encountered later then the reclaimer will stall. Set ZONE_WRITEBACK before potentially going to sleep so it is noticed sooner. Signed-off-by:
-
Mel Gorman authored
Commit "mm: vmscan: Block kswapd if it is encountering pages under writeback" blocks page reclaim if it encounters pages under writeback marked for immediate reclaim. It blocks while pages are still isolated from the LRU which is unnecessary. This patch defers the blocking until after the isolated pages have been processed and tidies up some of the comments. Signed-off-by:
-
Mel Gorman authored
Further testing of the "Reduce system disruption due to kswapd" discovered a few problems. First and foremost, it's possible for pages under writeback to be freed which will lead to badness. Second, as pages were not being swapped the file LRU was being scanned faster and clean file pages were being reclaimed. In some cases this results in increased read IO to re-read data from disk. Third, more pages were being written from kswapd context which can adversly affect IO performance. Lastly, it was observed that PageDirty pages are not necessarily dirty on all filesystems (buffers can be clean while PageDirty is set and ->writepage generates no IO) and not all filesystems set PageWriteback when the page is being written (e.g. ext3). This disconnect confuses the reclaim stalling logic. This follow-up series is aimed at these problems. The tests were based on three kernels vanilla: kernel 3.9 as that is what the current mmotm uses as a baseline mmotm-20130522 is mmotm as of 22nd May with "Reduce system disruption due to kswapd" applied on top as per what should be in Andrew's tree right now lessdisrupt-v7r10 is this follow-up series on top of the mmotm kernel The first test used memcached+memcachetest while some background IO was in progress as implemented by the parallel IO tests implement in MM Tests. memcachetest benchmarks how many operations/second memcached can service. It starts with no background IO on a freshly created ext4 filesystem and then re-runs the test with larger amounts of IO in the background to roughly simulate a large copy in progress. The expectation is that the IO should have little or no impact on memcachetest which is running entirely in memory. parallelio 3.9.0 3.9.0 3.9.0 vanilla mm1-mmotm-20130522 mm1-lessdisrupt-v7r10 Ops memcachetest-0M 23117.00 ( 0.00%) 22780.00 ( -1.46%) 22763.00 ( -1.53%) Ops memcachetest-715M 23774.00 ( 0.00%) 23299.00 ( -2.00%) 22934.00 ( -3.53%) Ops memcachetest-2385M 4208.00 ( 0.00%) 24154.00 (474.00%) 23765.00 (464.76%) Ops memcachetest-4055M 4104.00 ( 0.00%) 25130.00 (512.33%) 24614.00 (499.76%) Ops io-duration-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) Ops io-duration-715M 12.00 ( 0.00%) 7.00 ( 41.67%) 6.00 ( 50.00%) Ops io-duration-2385M 116.00 ( 0.00%) 21.00 ( 81.90%) 21.00 ( 81.90%) Ops io-duration-4055M 160.00 ( 0.00%) 36.00 ( 77.50%) 35.00 ( 78.12%) Ops swaptotal-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) Ops swaptotal-715M 140138.00 ( 0.00%) 18.00 ( 99.99%) 18.00 ( 99.99%) Ops swaptotal-2385M 385682.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) Ops swaptotal-4055M 418029.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) Ops swapin-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) Ops swapin-715M 144.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) Ops swapin-2385M 134227.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) Ops swapin-4055M 125618.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) Ops minorfaults-0M 1536429.00 ( 0.00%) 1531632.00 ( 0.31%) 1533541.00 ( 0.19%) Ops minorfaults-715M 1786996.00 ( 0.00%) 1612148.00 ( 9.78%) 1608832.00 ( 9.97%) Ops minorfaults-2385M 1757952.00 ( 0.00%) 1614874.00 ( 8.14%) 1613541.00 ( 8.21%) Ops minorfaults-4055M 1774460.00 ( 0.00%) 1633400.00 ( 7.95%) 1630881.00 ( 8.09%) Ops majorfaults-0M 1.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) Ops majorfaults-715M 184.00 ( 0.00%) 167.00 ( 9.24%) 166.00 ( 9.78%) Ops majorfaults-2385M 24444.00 ( 0.00%) 155.00 ( 99.37%) 93.00 ( 99.62%) Ops majorfaults-4055M 21357.00 ( 0.00%) 147.00 ( 99.31%) 134.00 ( 99.37%) memcachetest is the transactions/second reported by memcachetest. In the vanilla kernel note that performance drops from around 23K/sec to just over 4K/second when there is 2385M of IO going on in the background. With current mmotm, there is no collapse in performance and with this follow-up series there is little change. swaptotal is the total amount of swap traffic. With mmotm and the follow-up series, the total amount of swapping is much reduced. 3.9.0 3.9.0 3.9.0 vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10 Minor Faults 11160152 10706748 10622316 Major Faults 46305 755 678 Swap Ins 260249 0 0 Swap Outs 683860 18 18 Direct pages scanned 0 678 2520 Kswapd pages scanned 6046108 8814900 1639279 Kswapd pages reclaimed 1081954 1172267 1094635 Direct pages reclaimed 0 566 2304 Kswapd efficiency 17% 13% 66% Kswapd velocity 5217.560 7618.953 1414.879 Direct efficiency 100% 83% 91% Direct velocity 0.000 0.586 2.175 Percentage direct scans 0% 0% 0% Zone normal velocity 5105.086 6824.681 671.158 Zone dma32 velocity 112.473 794.858 745.896 Zone dma velocity 0.000 0.000 0.000 Page writes by reclaim 1929612.000 6861768.000 32821.000 Page writes file 1245752 6861750 32803 Page writes anon 683860 18 18 Page reclaim immediate 7484 40 239 Sector Reads 1130320 93996 86900 Sector Writes 13508052 10823500 11804436 Page rescued immediate 0 0 0 Slabs scanned 33536 27136 18560 Direct inode steals 0 0 0 Kswapd inode steals 8641 1035 0 Kswapd skipped wait 0 0 0 THP fault alloc 8 37 33 THP collapse alloc 508 552 515 THP splits 24 1 1 THP fault fallback 0 0 0 THP collapse fail 0 0 0 There are a number of observations to make here 1. Swap outs are almost eliminated. Swap ins are 0 indicating that the pages swapped were really unused anonymous pages. Related to that, major faults are much reduced. 2. kswapd efficiency was impacted by the initial series but with these follow-up patches, the efficiency is now at 66% indicating that far fewer pages were skipped during scanning due to dirty or writeback pages. 3. kswapd velocity is reduced indicating that fewer pages are being scanned with the follow-up series as kswapd now stalls when the tail of the LRU queue is full of unqueued dirty pages. The stall gives flushers a chance to catch-up so kswapd can reclaim clean pages when it wakes 4. In light of Zlatko's recent reports about zone scanning imbalances, mmtests now reports scanning velocity on a per-zone basis. With mainline, you can see that the scanning activity is dominated by the Normal zone with over 45 times more scanning in Normal than the DMA32 zone. With the series currently in mmotm, the ratio is slightly better but it is still the case that the bulk of scanning is in the highest zone. With this follow-up series, the ratio of scanning between the Normal and DMA32 zone is roughly equal. 5. As Dave Chinner observed, the current patches in mmotm increased the number of pages written from kswapd context which is expected to adversly impact IO performance. With the follow-up patches, far fewer pages are written from kswapd context than the mainline kernel 6. With the series in mmotm, fewer inodes were reclaimed by kswapd. With the follow-up series, there is less slab shrinking activity and no inodes were reclaimed. 7. Note that "Sectors Read" is drastically reduced implying that the source data being used for the IO is not being aggressively discarded due to page reclaim skipping over dirty pages and reclaiming clean pages. Note that the reducion in reads could also be due to inode data not being re-read from disk after a slab shrink. 3.9.0 3.9.0 3.9.0 vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10 Mean sda-avgqz 166.99 32.09 33.44 Mean sda-await 853.64 192.76 185.43 Mean sda-r_await 6.31 9.24 5.97 Mean sda-w_await 2992.81 202.65 192.43 Max sda-avgqz 1409.91 718.75 698.98 Max sda-await 6665.74 3538.00 3124.23 Max sda-r_await 58.96 111.95 58.00 Max sda-w_await 28458.94 3977.29 3148.61 In light of the changes in writes from reclaim context, the number of reads and Dave Chinner's concerns about IO performance I took a closer look at the IO stats for the test disk. Few observations 1. The average queue size is reduced by the initial series and roughly the same with this follow up. 2. Average wait times for writes are reduced and as the IO is completing faster it at least implies that the gain is because flushers are writing the files efficiently instead of page reclaim getting in the way. 3. The reduction in maximum write latency is staggering. 28 seconds down to 3 seconds. Jan Kara asked how NFS is affected by all of this. Unstable pages can be taken into account as one of the patches in the series shows but it is still the case that filesystems with unusual handling of dirty or writeback could still be treated better. Tests like postmark, fsmark and largedd showed up nothing useful. On my test setup, pages are simply not being written back from reclaim context with or without the patches and there are no changes in performance. My test setup probably is just not strong enough network-wise to be really interesting. I ran a longer-lived memcached test with IO going to NFS instead of a local disk parallelio 3.9.0 3.9.0 3.9.0 vanilla mm1-mmotm-20130522 mm1-lessdisrupt-v7r10 Ops memcachetest-0M 23323.00 ( 0.00%) 23241.00 ( -0.35%) 23321.00 ( -0.01%) Ops memcachetest-715M 25526.00 ( 0.00%) 24763.00 ( -2.99%) 23242.00 ( -8.95%) Ops memcachetest-2385M 8814.00 ( 0.00%) 26924.00 (205.47%) 23521.00 (166.86%) Ops memcachetest-4055M 5835.00 ( 0.00%) 26827.00 (359.76%) 25560.00 (338.05%) Ops io-duration-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) Ops io-duration-715M 65.00 ( 0.00%) 71.00 ( -9.23%) 11.00 ( 83.08%) Ops io-duration-2385M 129.00 ( 0.00%) 94.00 ( 27.13%) 53.00 ( 58.91%) Ops io-duration-4055M 301.00 ( 0.00%) 100.00 ( 66.78%) 108.00 ( 64.12%) Ops swaptotal-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) Ops swaptotal-715M 14394.00 ( 0.00%) 949.00 ( 93.41%) 63.00 ( 99.56%) Ops swaptotal-2385M 401483.00 ( 0.00%) 24437.00 ( 93.91%) 30118.00 ( 92.50%) Ops swaptotal-4055M 554123.00 ( 0.00%) 35688.00 ( 93.56%) 63082.00 ( 88.62%) Ops swapin-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) Ops swapin-715M 4522.00 ( 0.00%) 560.00 ( 87.62%) 63.00 ( 98.61%) Ops swapin-2385M 169861.00 ( 0.00%) 5026.00 ( 97.04%) 13917.00 ( 91.81%) Ops swapin-4055M 192374.00 ( 0.00%) 10056.00 ( 94.77%) 25729.00 ( 86.63%) Ops minorfaults-0M 1445969.00 ( 0.00%) 1520878.00 ( -5.18%) 1454024.00 ( -0.56%) Ops minorfaults-715M 1557288.00 ( 0.00%) 1528482.00 ( 1.85%) 1535776.00 ( 1.38%) Ops minorfaults-2385M 1692896.00 ( 0.00%) 1570523.00 ( 7.23%) 1559622.00 ( 7.87%) Ops minorfaults-4055M 1654985.00 ( 0.00%) 1581456.00 ( 4.44%) 1596713.00 ( 3.52%) Ops majorfaults-0M 0.00 ( 0.00%) 1.00 (-99.00%) 0.00 ( 0.00%) Ops majorfaults-715M 763.00 ( 0.00%) 265.00 ( 65.27%) 75.00 ( 90.17%) Ops majorfaults-2385M 23861.00 ( 0.00%) 894.00 ( 96.25%) 2189.00 ( 90.83%) Ops majorfaults-4055M 27210.00 ( 0.00%) 1569.00 ( 94.23%) 4088.00 ( 84.98%) 1. Performance does not collapse due to IO which is good. IO is also completing faster. Note with mmotm, IO completes in a third of the time and faster again with this series applied 2. Swapping is reduced, although not eliminated. The figures for the follow-up look bad but it does vary a bit as the stalling is not perfect for nfs or filesystems like ext3 with unusual handling of dirty and writeback pages 3. There are swapins, particularly with larger amounts of IO indicating that active pages are being reclaimed. However, the number of much reduced. 3.9.0 3.9.0 3.9.0 vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10 Minor Faults 36339175 35025445 35219699 Major Faults 310964 27108 51887 Swap Ins 2176399 173069 333316 Swap Outs 3344050 357228 504824 Direct pages scanned 8972 77283 43242 Kswapd pages scanned 20899983 8939566 14772851 Kswapd pages reclaimed 6193156 5172605 5231026 Direct pages reclaimed 8450 73802 39514 Kswapd efficiency 29% 57% 35% Kswapd velocity 3929.743 1847.499 3058.840 Direct efficiency 94% 95% 91% Direct velocity 1.687 15.972 8.954 Percentage direct scans 0% 0% 0% Zone normal velocity 3721.907 939.103 2185.142 Zone dma32 velocity 209.522 924.368 882.651 Zone dma velocity 0.000 0.000 0.000 Page writes by reclaim 4082185.000 526319.000 537114.000 Page writes file 738135 169091 32290 Page writes anon 3344050 357228 504824 Page reclaim immediate 9524 170 5595843 Sector Reads 8909900 861192 1483680 Sector Writes 13428980 1488744 2076800 Page rescued immediate 0 0 0 Slabs scanned 38016 31744 28672 Direct inode steals 0 0 0 Kswapd inode steals 424 0 0 Kswapd skipped wait 0 0 0 THP fault alloc 14 15 119 THP collapse alloc 1767 1569 1618 THP splits 30 29 25 THP fault fallback 0 0 0 THP collapse fail 8 5 0 Compaction stalls 17 41 100 Compaction success 7 31 95 Compaction failures 10 10 5 Page migrate success 7083 22157 62217 Page migrate failure 0 0 0 Compaction pages isolated 14847 48758 135830 Compaction migrate scanned 18328 48398 138929 Compaction free scanned 2000255 355827 1720269 Compaction cost 7 24 68 I guess the main takeaway again is the much reduced page writes from reclaim context and reduced reads. 3.9.0 3.9.0 3.9.0 vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10 Mean sda-avgqz 23.58 0.35 0.44 Mean sda-await 133.47 15.72 15.46 Mean sda-r_await 4.72 4.69 3.95 Mean sda-w_await 507.69 28.40 33.68 Max sda-avgqz 680.60 12.25 23.14 Max sda-await 3958.89 221.83 286.22 Max sda-r_await 63.86 61.23 67.29 Max sda-w_await 11710.38 883.57 1767.28 And as before, write wait times are much reduced. This patch: The patch "mm: vmscan: Have kswapd writeback pages based on dirty pages encountered, not priority" decides whether to writeback pages from reclaim context based on the number of dirty pages encountered. This situation is flagged too easily and flushers are not given the chance to catch up resulting in more pages being written from reclaim context and potentially impacting IO performance. The check for PageWriteback is also misplaced as it happens within a PageDirty check which is nonsense as the dirty may have been cleared for IO. The accounting is updated very late and pages that are already under writeback, were reactivated, could not unmapped or could not be released are all missed. Similarly, a page is considered congested for reasons other than being congested and pages that cannot be written out in the correct context are skipped. Finally, it considers stalling and writing back filesystem pages due to encountering dirty anonymous pages at the tail of the LRU which is dumb. This patch causes kswapd to begin writing filesystem pages from reclaim context only if page reclaim found that all filesystem pages at the tail of the LRU were unqueued dirty pages. Before it starts writing filesystem pages, it will stall to give flushers a chance to catch up. The decision on whether wait_iff_congested is also now determined by dirty filesystem pages only. Congested pages are based on whether the underlying BDI is congested regardless of the context of the reclaiming process. Signed-off-by: -
Mel Gorman authored
balance_pgdat() is very long and some of the logic can and should be internal to kswapd_shrink_zone(). Move it so the flow of balance_pgdat() is marginally easier to follow. Signed-off-by:
Johannes Weiner <hannes@cmpxchg.org> Reviewed-by:
Michal Hocko <mhocko@suse.cz> Acked-by:
Zlatko Calusic <zcalusic@bitsync.net> Cc: dormando <dormando@rydia.net> Signed-off-by:
-
Mel Gorman authored
Currently kswapd checks if it should start writepage as it shrinks each zone without taking into consideration if the zone is balanced or not. This is not wrong as such but it does not make much sense either. This patch checks once per pgdat scan if kswapd should be writing pages. Signed-off-by:
-
Mel Gorman authored
Historically, kswapd used to congestion_wait() at higher priorities if it was not making forward progress. This made no sense as the failure to make progress could be completely independent of IO. It was later replaced by wait_iff_congested() and removed entirely by commit 258401a6 (mm: don't wait on congested zones in balance_pgdat()) as it was duplicating logic in shrink_inactive_list(). This is problematic. If kswapd encounters many pages under writeback and it continues to scan until it reaches the high watermark then it will quickly skip over the pages under writeback and reclaim clean young pages or push applications out to swap. The use of wait_iff_congested() is not suited to kswapd as it will only stall if the underlying BDI is really congested or a direct reclaimer was unable to write to the underlying BDI. kswapd bypasses the BDI congestion as it sets PF_SWAPWRITE but even if this was taken into account then it would cause direct reclaimers to stall on writeback which is not desirable. This patch sets a ZONE_WRITEBACK flag if direct reclaim or kswapd is encountering too many pages under writeback. If this flag is set and kswapd encounters a PageReclaim page under writeback then it'll assume that the LRU lists are being recycled too quickly before IO can complete and block waiting for some IO to complete. Signed-off-by:
-
Mel Gorman authored
Currently kswapd queues dirty pages for writeback if scanning at an elevated priority but the priority kswapd scans at is not related to the number of unqueued dirty encountered. Since commit "mm: vmscan: Flatten kswapd priority loop", the priority is related to the size of the LRU and the zone watermark which is no indication as to whether kswapd should write pages or not. This patch tracks if an excessive number of unqueued dirty pages are being encountered at the end of the LRU. If so, it indicates that dirty pages are being recycled before flusher threads can clean them and flags the zone so that kswapd will start writing pages until the zone is balanced. Signed-off-by:
-
Mel Gorman authored
Page reclaim at priority 0 will scan the entire LRU as priority 0 is considered to be a near OOM condition. Kswapd can reach priority 0 quite easily if it is encountering a large number of pages it cannot reclaim such as pages under writeback. When this happens, kswapd reclaims very aggressively even though there may be no real risk of allocation failure or OOM. This patch prevents kswapd reaching priority 0 and trying to reclaim the world. Direct reclaimers will still reach priority 0 in the event of an OOM situation. Signed-off-by:
-
Mel Gorman authored
In the past, kswapd makes a decision on whether to compact memory after the pgdat was considered balanced. This more or less worked but it is late to make such a decision and does not fit well now that kswapd makes a decision whether to exit the zone scanning loop depending on reclaim progress. This patch will compact a pgdat if at least the requested number of pages were reclaimed from unbalanced zones for a given priority. If any zone is currently balanced, kswapd will not call compaction as it is expected the necessary pages are already available. Signed-off-by:
-
Mel Gorman authored
kswapd stops raising the scanning priority when at least SWAP_CLUSTER_MAX pages have been reclaimed or the pgdat is considered balanced. It then rechecks if it needs to restart at DEF_PRIORITY and whether high-order reclaim needs to be reset. This is not wrong per-se but it is confusing to follow and forcing kswapd to stay at DEF_PRIORITY may require several restarts before it has scanned enough pages to meet the high watermark even at 100% efficiency. This patch irons out the logic a bit by controlling when priority is raised and removing the "goto loop_again". This patch has kswapd raise the scanning priority until it is scanning enough pages that it could meet the high watermark in one shrink of the LRU lists if it is able to reclaim at 100% efficiency. It will not raise the scanning prioirty higher unless it is failing to reclaim any pages. To avoid infinite looping for high-order allocation requests kswapd will not reclaim for high-order allocations when it has reclaimed at least twice the number of pages as the allocation request. Signed-off-by:
-
Mel Gorman authored
Simplistically, the anon and file LRU lists are scanned proportionally depending on the value of vm.swappiness although there are other factors taken into account by get_scan_count(). The patch "mm: vmscan: Limit the number of pages kswapd reclaims" limits the number of pages kswapd reclaims but it breaks this proportional scanning and may evenly shrink anon/file LRUs regardless of vm.swappiness. This patch preserves the proportional scanning and reclaim. It does mean that kswapd will reclaim more than requested but the number of pages will be related to the high watermark. [mhocko@suse.cz: Correct proportional reclaim for memcg and simplify] [kamezawa.hiroyu@jp.fujitsu.com: Recalculate scan based on target] [hannes@cmpxchg.org: Account for already scanned pages properly] Signed-off-by:
-
Mel Gorman authored
This series does not fix all the current known problems with reclaim but it addresses one important swapping bug when there is background IO. Changelog since V3 - Drop the slab shrink changes in light of Glaubers series and discussions highlighted that there were a number of potential problems with the patch. (mel) - Rebased to 3.10-rc1 Changelog since V2 - Preserve ratio properly for proportional scanning (kamezawa) Changelog since V1 - Rename ZONE_DIRTY to ZONE_TAIL_LRU_DIRTY (andi) - Reformat comment in shrink_page_list (andi) - Clarify some comments (dhillf) - Rework how the proportional scanning is preserved - Add PageReclaim check before kswapd starts writeback - Reset sc.nr_reclaimed on every full zone scan Kswapd and page reclaim behaviour has been screwy in one way or the other for a long time. Very broadly speaking it worked in the far past because machines were limited in memory so it did not have that many pages to scan and it stalled congestion_wait() frequently to prevent it going completely nuts. In recent times it has behaved very unsatisfactorily with some of the problems compounded by the removal of stall logic and the introduction of transparent hugepage support with high-order reclaims. There are many variations of bugs that are rooted in this area. One example is reports of a large copy operations or backup causing the machine to grind to a halt or applications pushed to swap. Sometimes in low memory situations a large percentage of memory suddenly gets reclaimed. In other cases an application starts and kswapd hits 100% CPU usage for prolonged periods of time and so on. There is now talk of introducing features like an extra free kbytes tunable to work around aspects of the problem instead of trying to deal with it. It's compounded by the problem that it can be very workload and machine specific. This series aims at addressing some of the worst of these problems without attempting to fundmentally alter how page reclaim works. Patches 1-2 limits the number of pages kswapd reclaims while still obeying the anon/file proportion of the LRUs it should be scanning. Patches 3-4 control how and when kswapd raises its scanning priority and deletes the scanning restart logic which is tricky to follow. Patch 5 notes that it is too easy for kswapd to reach priority 0 when scanning and then reclaim the world. Down with that sort of thing. Patch 6 notes that kswapd starts writeback based on scanning priority which is not necessarily related to dirty pages. It will have kswapd writeback pages if a number of unqueued dirty pages have been recently encountered at the tail of the LRU. Patch 7 notes that sometimes kswapd should stall waiting on IO to complete to reduce LRU churn and the likelihood that it'll reclaim young clean pages or push applications to swap. It will cause kswapd to block on IO if it detects that pages being reclaimed under writeback are recycling through the LRU before the IO completes. Patchies 8-9 are cosmetic but balance_pgdat() is easier to follow after they are applied. This was tested using memcached+memcachetest while some background IO was in progress as implemented by the parallel IO tests implement in MM Tests. memcachetest benchmarks how many operations/second memcached can service and it is run multiple times. It starts with no background IO and then re-runs the test with larger amounts of IO in the background to roughly simulate a large copy in progress. The expectation is that the IO should have little or no impact on memcachetest which is running entirely in memory. 3.10.0-rc1 3.10.0-rc1 vanilla lessdisrupt-v4 Ops memcachetest-0M 22155.00 ( 0.00%) 22180.00 ( 0.11%) Ops memcachetest-715M 22720.00 ( 0.00%) 22355.00 ( -1.61%) Ops memcachetest-2385M 3939.00 ( 0.00%) 23450.00 (495.33%) Ops memcachetest-4055M 3628.00 ( 0.00%) 24341.00 (570.92%) Ops io-duration-0M 0.00 ( 0.00%) 0.00 ( 0.00%) Ops io-duration-715M 12.00 ( 0.00%) 7.00 ( 41.67%) Ops io-duration-2385M 118.00 ( 0.00%) 21.00 ( 82.20%) Ops io-duration-4055M 162.00 ( 0.00%) 36.00 ( 77.78%) Ops swaptotal-0M 0.00 ( 0.00%) 0.00 ( 0.00%) Ops swaptotal-715M 140134.00 ( 0.00%) 18.00 ( 99.99%) Ops swaptotal-2385M 392438.00 ( 0.00%) 0.00 ( 0.00%) Ops swaptotal-4055M 449037.00 ( 0.00%) 27864.00 ( 93.79%) Ops swapin-0M 0.00 ( 0.00%) 0.00 ( 0.00%) Ops swapin-715M 0.00 ( 0.00%) 0.00 ( 0.00%) Ops swapin-2385M 148031.00 ( 0.00%) 0.00 ( 0.00%) Ops swapin-4055M 135109.00 ( 0.00%) 0.00 ( 0.00%) Ops minorfaults-0M 1529984.00 ( 0.00%) 1530235.00 ( -0.02%) Ops minorfaults-715M 1794168.00 ( 0.00%) 1613750.00 ( 10.06%) Ops minorfaults-2385M 1739813.00 ( 0.00%) 1609396.00 ( 7.50%) Ops minorfaults-4055M 1754460.00 ( 0.00%) 1614810.00 ( 7.96%) Ops majorfaults-0M 0.00 ( 0.00%) 0.00 ( 0.00%) Ops majorfaults-715M 185.00 ( 0.00%) 180.00 ( 2.70%) Ops majorfaults-2385M 24472.00 ( 0.00%) 101.00 ( 99.59%) Ops majorfaults-4055M 22302.00 ( 0.00%) 229.00 ( 98.97%) Note how the vanilla kernels performance collapses when there is enough IO taking place in the background. This drop in performance is part of what users complain of when they start backups. Note how the swapin and major fault figures indicate that processes were being pushed to swap prematurely. With the series applied, there is no noticable performance drop and while there is still some swap activity, it's tiny. 20 iterations of this test were run in total and averaged. Every 5 iterations, additional IO was generated in the background using dd to measure how the workload was impacted. The 0M, 715M, 2385M and 4055M subblock refer to the amount of IO going on in the background at each iteration. So memcachetest-2385M is reporting how many transactions/second memcachetest recorded on average over 5 iterations while there was 2385M of IO going on in the ground. There are six blocks of information reported here memcachetest is the transactions/second reported by memcachetest. In the vanilla kernel note that performance drops from around 22K/sec to just under 4K/second when there is 2385M of IO going on in the background. This is one type of performance collapse users complain about if a large cp or backup starts in the background io-duration refers to how long it takes for the background IO to complete. It's showing that with the patched kernel that the IO completes faster while not interfering with the memcache workload swaptotal is the total amount of swap traffic. With the patched kernel, the total amount of swapping is much reduced although it is still not zero. swapin in this case is an indication as to whether we are swap trashing. The closer the swapin/swapout ratio is to 1, the worse the trashing is. Note with the patched kernel that there is no swapin activity indicating that all the pages swapped were really inactive unused pages. minorfaults are just minor faults. An increased number of minor faults can indicate that page reclaim is unmapping the pages but not swapping them out before they are faulted back in. With the patched kernel, there is only a small change in minor faults majorfaults are just major faults in the target workload and a high number can indicate that a workload is being prematurely swapped. With the patched kernel, major faults are much reduced. As there are no swapin's recorded so it's not being swapped. The likely explanation is that that libraries or configuration files used by the workload during startup get paged out by the background IO. Overall with the series applied, there is no noticable performance drop due to background IO and while there is still some swap activity, it's tiny and the lack of swapins imply that the swapped pages were inactive and unused. 3.10.0-rc1 3.10.0-rc1 vanilla lessdisrupt-v4 Page Ins 1234608 101892 Page Outs 12446272 11810468 Swap Ins 283406 0 Swap Outs 698469 27882 Direct pages scanned 0 136480 Kswapd pages scanned 6266537 5369364 Kswapd pages reclaimed 1088989 930832 Direct pages reclaimed 0 120901 Kswapd efficiency 17% 17% Kswapd velocity 5398.371 4635.115 Direct efficiency 100% 88% Direct velocity 0.000 117.817 Percentage direct scans 0% 2% Page writes by reclaim 1655843 4009929 Page writes file 957374 3982047 Page writes anon 698469 27882 Page reclaim immediate 5245 1745 Page rescued immediate 0 0 Slabs scanned 33664 25216 Direct inode steals 0 0 Kswapd inode steals 19409 778 Kswapd skipped wait 0 0 THP fault alloc 35 30 THP collapse alloc 472 401 THP splits 27 22 THP fault fallback 0 0 THP collapse fail 0 1 Compaction stalls 0 4 Compaction success 0 0 Compaction failures 0 4 Page migrate success 0 0 Page migrate failure 0 0 Compaction pages isolated 0 0 Compaction migrate scanned 0 0 Compaction free scanned 0 0 Compaction cost 0 0 NUMA PTE updates 0 0 NUMA hint faults 0 0 NUMA hint local faults 0 0 NUMA pages migrated 0 0 AutoNUMA cost 0 0 Unfortunately, note that there is a small amount of direct reclaim due to kswapd no longer reclaiming the world. ftrace indicates that the direct reclaim stalls are mostly harmless with the vast bulk of the stalls incurred by dd 23 tclsh-3367 38 memcachetest-13733 49 memcachetest-12443 57 tee-3368 1541 dd-13826 1981 dd-12539 A consequence of the direct reclaim for dd is that the processes for the IO workload may show a higher system CPU usage. There is also a risk that kswapd not reclaiming the world may mean that it stays awake balancing zones, does not stall on the appropriate events and continually scans pages it cannot reclaim consuming CPU. This will be visible as continued high CPU usage but in my own tests I only saw a single spike lasting less than a second and I did not observe any problems related to reclaim while running the series on my desktop. This patch: The number of pages kswapd can reclaim is bound by the number of pages it scans which is related to the size of the zone and the scanning priority. In many cases the priority remains low because it's reset every SWAP_CLUSTER_MAX reclaimed pages but in the event kswapd scans a large number of pages it cannot reclaim, it will raise the priority and potentially discard a large percentage of the zone as sc->nr_to_reclaim is ULONG_MAX. The user-visible effect is a reclaim "spike" where a large percentage of memory is suddenly freed. It would be bad enough if this was just unused memory but because of how anon/file pages are balanced it is possible that applications get pushed to swap unnecessarily. This patch limits the number of pages kswapd will reclaim to the high watermark. Reclaim will still overshoot due to it not being a hard limit as shrink_lruvec() will ignore the sc.nr_to_reclaim at DEF_PRIORITY but it prevents kswapd reclaiming the world at higher priorities. The number of pages it reclaims is not adjusted for high-order allocations as kswapd will reclaim excessively if it is to balance zones for high-order allocations. Signed-off-by: -
Cody P Schafer authored
When memory hotplug is triggered, we call pageset_init() on per-cpu-pagesets which both contain pages and are in use, causing both the leakage of those pages and (potentially) bad behaviour if a page is allocated from a pageset while it is being cleared. Avoid this by factoring out pageset_set_high_and_batch() (which contains all needed logic too set a pageset's ->high and ->batch inrespective of system state) from zone_pageset_init() and using the new pageset_set_high_and_batch() instead of zone_pageset_init() in zone_pcp_update(). Signed-off-by:
-
Cody P Schafer authored
Signed-off-by:
-
Cody P Schafer authored
Previously, zone_pcp_update() called pageset_set_batch() directly, essentially assuming that percpu_pagelist_fraction == 0. Correct this by calling zone_pageset_init(), which chooses the appropriate ->batch and ->high calculations. Signed-off-by:
-
Cody P Schafer authored
Signed-off-by:
-
Cody P Schafer authored
Signed-off-by:
-
Cody P Schafer authored
Signed-off-by:
-
Cody P Schafer authored
Simply moves calculation of the new 'high' value outside the for_each_possible_cpu() loop, as it does not depend on the cpu. Signed-off-by:
-
Cody P Schafer authored
zone_pcp_update()'s goal is to adjust the ->high and ->mark members of a percpu pageset based on a zone's ->managed_pages. We don't need to drain the entire percpu pageset just to modify these fields. This lets us avoid calling setup_pageset() (and the draining required to call it) and instead allows simply setting the fields' values (with some attention paid to memory barriers to prevent the relationship between ->batch and ->high from being thrown off). This does change the behavior of zone_pcp_update() as the percpu pagesets will not be drained when zone_pcp_update() is called (they will end up being shrunk, not completely drained, later when a 0-order page is freed in free_hot_cold_page()). Signed-off-by:
-
