- 02 Jun, 2020 40 commits
-
-
Christoph Hellwig authored
The non-cached vmalloc mapping was initially added as a hack for the first-gen amigaone platform (6xx/book32s), isn't fully supported upstream, and which used the legacy radeon driver together with non-coherent DMA. However this only ever worked reliably for DRI . Remove the hack as it is the last user of __vmalloc passing a page protection flag other than PAGE_KERNEL and didn't do anything for other platforms with non-coherent DMA. Signed-off-by:
Christoph Hellwig <hch@lst.de> Signed-off-by:
Andrew Morton <akpm@linux-foundation.org> Acked-by:
Daniel Vetter <daniel.vetter@ffwll.ch> Acked-by:
Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Christophe Leroy <christophe.leroy@c-s.fr> Cc: David Airlie <airlied@linux.ie> Cc: Gao Xiang <xiang@kernel.org> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Haiyang Zhang <haiyangz@microsoft.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: "K. Y. Srinivasan" <kys@microsoft.com> Cc: Laura Abbott <labbott@redhat.com> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Michael Kelley <mikelley@microsoft.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Nitin Gupta <ngupta@vflare.org> Cc: Robin Murphy <robin.murphy@arm.com> Cc: Sakari Ailus <sakari.ailus@linux.intel.com> Cc: Stephen Hemminger <sthemmin@microsoft.com> Cc: Sumit Semwal <sumit.semwal@linaro.org> Cc: Wei Liu <wei.liu@kernel.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: Paul Mackerras <paulus@ozlabs.org> Cc: Vasily Gorbik <gor@linux.ibm.com> Cc: Will Deacon <will@kernel.org> Link: http://lkml.kernel.org/r/20200414131348.444715-21-hch@lst.deSigned-off-by:
Linus Torvalds <torvalds@linux-foundation.org>
-
Christoph Hellwig authored
To help enforcing the W^X protection don't allow remapping existing pages as executable. x86 bits from Peter Zijlstra, arm64 bits from Mark Rutland. Signed-off-by:
-
Christoph Hellwig authored
This is always PAGE_KERNEL - for long term mappings with other properties vmap should be used. Signed-off-by:
-
Christoph Hellwig authored
This function just has a single caller, open code it there. Signed-off-by:
-
Christoph Hellwig authored
Switch all callers to map_kernel_range, which symmetric to the unmap side (as well as the _noflush versions). Signed-off-by:
-
Christoph Hellwig authored
None of the callers needs the number of pages, and a 0 / -errno return value is a lot more intuitive. Signed-off-by:
-
Christoph Hellwig authored
This matches the map_kernel_range_noflush API. Also change to pass a size instead of the end, similar to the noflush version. Signed-off-by:
-
Christoph Hellwig authored
These have non-static aliases called map_kernel_range_noflush and unmap_kernel_range_noflush that just differ slightly in the calling conventions that pass addr + size instead of an end. Signed-off-by:
-
Christoph Hellwig authored
Ever use of addr in vb_free casts to unsigned long first, and the caller has an unsigned long version of the address available anyway. Just pass that and avoid all the casts. Signed-off-by:
-
Christoph Hellwig authored
This allows to unexport map_vm_area and unmap_kernel_range, which are rather deep internal and should not be available to modules, as they for example allow fine grained control of mapping permissions, and also allow splitting the setup of a vmalloc area and the actual mapping and thus expose vmalloc internals. zsmalloc is typically built-in and continues to work (just like the percpu-vm code using a similar patter), while modular zsmalloc also continues to work, but must use copies. Signed-off-by:
-
Christoph Hellwig authored
Rename the Kconfig variable to clarify the scope. Signed-off-by:
Minchan Kim <minchan@kernel.org> Acked-by:
-
Christoph Hellwig authored
There are no modular users of this function. Signed-off-by:
-
Christoph Hellwig authored
Switch the two remaining callers to use __get_vm_area_caller instead. Signed-off-by:
-
Christoph Hellwig authored
These helpers are only used for remapping the ISA I/O base. Replace the mapping side with a remap_isa_range helper in isa-bridge.c that hard codes all the known arguments, and just remove __iounmap_at in favour of open coding it in the only caller. Signed-off-by:
-
Christoph Hellwig authored
Factor code shared between pci_64 and electra_cf into a ioremap_pbh helper that follows the normal ioremap semantics, and returns a useful __iomem pointer. Note that it opencodes __ioremap_at as we know from the callers the slab is available. Switch pci_64 to also store the result as __iomem pointer, and unmap the result using iounmap instead of force casting and using vmalloc APIs. Signed-off-by:
-
Christoph Hellwig authored
Replace the open coded instance of vmap with the actual function. In the non-contiguous (IOMMU) case this requires an extra find_vm_area, but given that this isn't a fast path function that is a small price to pay. Signed-off-by:
-
Christoph Hellwig authored
Just use vmap instead of messing with vmalloc internals. Signed-off-by:
-
Christoph Hellwig authored
vm_map_ram can keep mappings around after the vm_unmap_ram. Using that with non-PAGE_KERNEL mappings can lead to all kinds of aliasing issues. Signed-off-by:
Greg Kroah-Hartman <gregkh@linuxfoundation.org> Acked-by:
-
Christoph Hellwig authored
vmap does not take a gfp_t, the flags argument is for VM_* flags. Signed-off-by:
-
Christoph Hellwig authored
Patch series "decruft the vmalloc API", v2. Peter noticed that with some dumb luck you can toast the kernel address space with exported vmalloc symbols. I used this as an opportunity to decruft the vmalloc.c API and make it much more systematic. This also removes any chance to create vmalloc mappings outside the designated areas or using executable permissions from modules. Besides that it removes more than 300 lines of code. This patch (of 29): Use the designated helper for allocating executable kernel memory, and remove the now unused PAGE_KERNEL_RX define. Signed-off-by:
Michael Kelley <mikelley@microsoft.com> Acked-by:
Wei Liu <wei.liu@kernel.org> Acked-by:
-
Wetp Zhang authored
Some processes dont't want to be killed early, but in "Action Required" case, those also may be killed by BUS_MCEERR_AO when sharing memory with other which is accessing the fail memory. And sending SIGBUS with BUS_MCEERR_AO for action required error is strange, so ignore the non-current processes here. Suggested-by:
Naoya Horiguchi <naoya.horiguchi@nec.com> Signed-off-by:
Wetp Zhang <wetp.zy@linux.alibaba.com> Signed-off-by:
Pankaj Gupta <pankaj.gupta.linux@gmail.com> Link: http://lkml.kernel.org/r/1590817116-21281-1-git-send-email-wetp.zy@linux.alibaba.comSigned-off-by:
-
chenqiwu authored
Since commit 25b2995a ("mm: remove MEMORY_DEVICE_PUBLIC support"), the assignment to 'page' for pte_devmap case has been unnecessary. Let's remove it. [willy@infradead.org: changelog] Signed-off-by:
chenqiwu <chenqiwu@xiaomi.com> Signed-off-by:
Matthew Wilcox <willy@infradead.org> Acked-by:
Michal Hocko <mhocko@suse.com> Link: http://lkml.kernel.org/r/1587349685-31712-1-git-send-email-qiwuchen55@gmail.comSigned-off-by:
-
Huang Ying authored
Now, when reading /proc/PID/smaps, the PMD migration entry in page table is simply ignored. To improve the accuracy of /proc/PID/smaps, its parsing and processing is added. To test the patch, we run pmbench to eat 400 MB memory in background, then run /usr/bin/migratepages and `cat /proc/PID/smaps` every second. The issue as follows can be reproduced within 60 seconds. Before the patch, for the fully populated 400 MB anonymous VMA, some THP pages under migration may be lost as below. 7f3f6a7e5000-7f3f837e5000 rw-p 00000000 00:00 0 Size: 409600 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Rss: 407552 kB Pss: 407552 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 407552 kB Referenced: 301056 kB Anonymous: 407552 kB LazyFree: 0 kB AnonHugePages: 405504 kB ShmemPmdMapped: 0 kB FilePmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB Locked: 0 kB THPeligible: 1 VmFlags: rd wr mr mw me ac After the patch, it will be always, 7f3f6a7e5000-7f3f837e5000 rw-p 00000000 00:00 0 Size: 409600 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Rss: 409600 kB Pss: 409600 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 409600 kB Referenced: 294912 kB Anonymous: 409600 kB LazyFree: 0 kB AnonHugePages: 407552 kB ShmemPmdMapped: 0 kB FilePmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB Locked: 0 kB THPeligible: 1 VmFlags: rd wr mr mw me ac Signed-off-by:
"Huang, Ying" <ying.huang@intel.com> Signed-off-by:
Zi Yan <ziy@nvidia.com> Acked-by:
Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by:
Vlastimil Babka <vbabka@suse.cz> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Alexey Dobriyan <adobriyan@gmail.com> Cc: Konstantin Khlebnikov <khlebnikov@yandex-team.ru> Cc: "Jérôme Glisse" <jglisse@redhat.com> Cc: Yang Shi <yang.shi@linux.alibaba.com> Link: http://lkml.kernel.org/r/20200403123059.1846960-1-ying.huang@intel.comSigned-off-by:
-
Steven Price authored
The page table entry is passed in the 'val' argument to note_page(), however this was previously an "unsigned long" which is fine on 64-bit platforms. But for 32 bit x86 it is not always big enough to contain a page table entry which may be 64 bits. Change the type to u64 to ensure that it is always big enough. [akpm@linux-foundation.org: fix riscv] Reported-by:
Jan Beulich <jbeulich@suse.com> Signed-off-by:
Steven Price <steven.price@arm.com> Signed-off-by:
-
Steven Price authored
Patch series "Fix W+X debug feature on x86" Jan alerted me[1] that the W+X detection debug feature was broken in x86 by my change[2] to switch x86 to use the generic ptdump infrastructure. Fundamentally the approach of trying to move the calculation of effective permissions into note_page() was broken because note_page() is only called for 'leaf' entries and the effective permissions are passed down via the internal nodes of the page tree. The solution I've taken here is to create a new (optional) callback which is called for all nodes of the page tree and therefore can calculate the effective permissions. Secondly on some configurations (32 bit with PAE) "unsigned long" is not large enough to store the table entries. The fix here is simple - let's just use a u64. [1] https://lore.kernel.org/lkml/d573dc7e-e742-84de-473d-f971142fa319@suse.com/ [2] 2ae27137 ("x86: mm: convert dump_pagetables to use walk_page_range") This patch (of 2): By switching the x86 page table dump code to use the generic code the effective permissions are no longer calculated correctly because the note_page() function is only called for *leaf* entries. To calculate the actual effective permissions it is necessary to observe the full hierarchy of the page tree. Introduce a new callback for ptdump which is called for every entry and can therefore update the prot_levels array correctly. note_page() can then simply access the appropriate element in the array. [steven.price@arm.com: make the assignment conditional on val != 0] Link: http://lkml.kernel.org/r/430c8ab4-e7cd-6933-dde6-087fac6db872@arm.com Fixes: 2ae27137 ("x86: mm: convert dump_pagetables to use walk_page_range") Reported-by:
-
Zefan Li authored
While trying to use remote memcg charging in an out-of-tree kernel module I found it's not working, because the current thread is a workqueue thread. As we will probably encounter this issue in the future as the users of memalloc_use_memcg() grow, and it's nothing wrong for this usage, it's better we fix it now. Signed-off-by:
Zefan Li <lizefan@huawei.com> Signed-off-by:
Roman Gushchin <guro@fb.com> Reviewed-by:
Shakeel Butt <shakeelb@google.com> Acked-by:
Johannes Weiner <hannes@cmpxchg.org> Acked-by:
-
Jakub Kicinski authored
Add a memory.swap.high knob, which can be used to protect the system from SWAP exhaustion. The mechanism used for penalizing is similar to memory.high penalty (sleep on return to user space). That is not to say that the knob itself is equivalent to memory.high. The objective is more to protect the system from potentially buggy tasks consuming a lot of swap and impacting other tasks, or even bringing the whole system to stand still with complete SWAP exhaustion. Hopefully without the need to find per-task hard limits. Slowing misbehaving tasks down gradually allows user space oom killers or other protection mechanisms to react. oomd and earlyoom already do killing based on swap exhaustion, and memory.swap.high protection will help implement such userspace oom policies more reliably. We can use one counter for number of pages allocated under pressure to save struct task space and avoid two separate hierarchy walks on the hot path. The exact overage is calculated on return to user space, anyway. Take the new high limit into account when determining if swap is "full". Borrowing the explanation from Johannes: The idea behind "swap full" is that as long as the workload has plenty of swap space available and it's not changing its memory contents, it makes sense to generously hold on to copies of data in the swap device, even after the swapin. A later reclaim cycle can drop the page without any IO. Trading disk space for IO. But the only two ways to reclaim a swap slot is when they're faulted in and the references go away, or by scanning the virtual address space like swapoff does - which is very expensive (one could argue it's too expensive even for swapoff, it's often more practical to just reboot). So at some point in the fill level, we have to start freeing up swap slots on fault/swapin. Otherwise we could eventually run out of swap slots while they're filled with copies of data that is also in RAM. We don't want to OOM a workload because its available swap space is filled with redundant cache. Signed-off-by:
Jakub Kicinski <kuba@kernel.org> Signed-off-by:
-
Jakub Kicinski authored
High memory limit is currently recorded directly in struct mem_cgroup. We are about to add a high limit for swap, move the field to struct page_counter and add some helpers. Signed-off-by:
-
Jakub Kicinski authored
We will want to call calculate_high_delay() twice - once for memory and once for swap, and we should apply the clamp value to sum of the penalties. Clamping has to be applied outside of calculate_high_delay(). Signed-off-by:
-
Jakub Kicinski authored
Patch series "memcg: Slow down swap allocation as the available space gets depleted", v6. Tejun describes the problem as follows: When swap runs out, there's an abrupt change in system behavior - the anonymous memory suddenly becomes unmanageable which readily breaks any sort of memory isolation and can bring down the whole system. To avoid that, oomd [1] monitors free swap space and triggers kills when it drops below the specific threshold (e.g. 15%). While this works, it's far from ideal: - Depending on IO performance and total swap size, a given headroom might not be enough or too much. - oomd has to monitor swap depletion in addition to the usual pressure metrics and it currently doesn't consider memory.swap.max. Solve this by adapting parts of the approach that memory.high uses - slow down allocation as the resource gets depleted turning the depletion behavior from abrupt cliff one to gradual degradation observable through memory pressure metric. [1] https://github.com/facebookincubator/oomd This patch (of 4): Slice the memory overage calculation logic a little bit so we can reuse it to apply a similar penalty to the swap. The logic which accesses the memory-specific fields (use and high values) has to be taken out of calculate_high_delay(). Signed-off-by:
-
Shakeel Butt authored
One way to measure the efficiency of memory reclaim is to look at the ratio (pgscan+pfrefill)/pgsteal. However at the moment these stats are not updated consistently at the system level and the ratio of these are not very meaningful. The pgsteal and pgscan are updated for only global reclaim while pgrefill gets updated for global as well as cgroup reclaim. Please note that this difference is only for system level vmstats. The cgroup stats returned by memory.stat are actually consistent. The cgroup's pgsteal contains number of reclaimed pages for global as well as cgroup reclaim. So, one way to get the system level stats is to get these stats from root's memory.stat, so, expose memory.stat for the root cgroup. From Johannes Weiner: There are subtle differences between /proc/vmstat and memory.stat, and cgroup-aware code that wants to watch the full hierarchy currently has to know about these intricacies and translate semantics back and forth. Generally having the fully recursive memory.stat at the root level could help a broader range of usecases. Why not fix the stats by including both the global and cgroup reclaim activity instead of exposing root cgroup's memory.stat? The reason is the benefit of having metrics exposing the activity that happens purely due to machine capacity rather than localized activity that happens due to the limits throughout the cgroup tree. Additionally there are userspace tools like sysstat(sar) which reads these stats to inform about the system level reclaim activity. So, we should not break such use-cases. Suggested-by:
Yafang Shao <laoar.shao@gmail.com> Acked-by:
Chris Down <chris@chrisdown.name> Cc: Mel Gorman <mgorman@suse.de> Cc: Roman Gushchin <guro@fb.com> Cc: Michal Hocko <mhocko@kernel.org> Link: http://lkml.kernel.org/r/20200508170630.94406-1-shakeelb@google.comSigned-off-by:
-
Kaixu Xia authored
When the variables count and limit have the same value(count == limit), the result of min(margin, limit - count) statement should be 0 and the variable margin is set to 0. So in this case, the min() statement is not necessary and we can directly set the variable margin to 0. Signed-off-by:
Kaixu Xia <kaixuxia@tencent.com> Signed-off-by:
-
Yafang Shao authored
There's a new workingset counter introduced in commit 1899ad18 ("mm: workingset: tell cache transitions from workingset thrashing"). With the help of this counter we can know the workingset is transitioning or thrashing. To leverage the benifit of this counter to memcg, we should introduce it into memory.stat. Then we could know the workingset of the workload inside a memcg better. Bellow is the verification of this new counter in memory.stat. Read a file into the memory and then read it again to make these pages be active. The size of this file is 1G. (memory.max is greater than file size) The counters in memory.stat will be inactive_file 0 active_file 1073639424 workingset_refault 0 workingset_activate 0 workingset_restore 0 workingset_nodereclaim 0 Trigger the memcg reclaim by setting a lower value to memory.high, and then some pages will be demoted into inactive list, and then some pages in the inactive list will be evicted into the storage. inactive_file 498094080 active_file 310063104 workingset_refault 0 workingset_activate 0 workingset_restore 0 workingset_nodereclaim 0 Then recover the memory.high and read the file into memory again. As a result of it, the transitioning will occur. Bellow is the result of this transitioning, inactive_file 498094080 active_file 575397888 workingset_refault 64746 workingset_activate 64746 workingset_restore 64746 workingset_nodereclaim 0 Signed-off-by:
-
Miaohe Lin authored
Since commit 8d93b41c ("mm: Convert add_to_swap_cache to XArray"), __add_to_swap_cache and add_to_swap_cache are combined into one function. There is no __add_to_swap_cache() anymore. Signed-off-by:
Miaohe Lin <linmiaohe@huawei.com> Signed-off-by:
-
Randy Dunlap authored
Fix the heading and Size/Used/Priority field alignments in /proc/swaps. If the Size and/or Used value is >= 10000000 (8 bytes), then the alignment by using tab characters is broken. This patch maintains the use of tabs for alignment. If spaces are preferred, we can just use a Field Width specifier for the bytes and inuse fields. That way those fields don't have to be a multiple of 8 bytes in width. E.g., with a field width of 12, both Size and Used would always fit on the first line of an 80-column wide terminal (only Priority would be on the second line). There are actually 2 problems: heading alignment and field width. On an xterm, if Used is 7 bytes in length, the tab does nothing, and the display is like this, with no space/tab between the Used and Priority fields. (ugh) Filename Type Size Used Priority /dev/sda8 partition 16779260 2023012-1 To be clear, if one does 'cat /proc/swaps >/tmp/proc.swaps', it does look different, like so: Filename Type Size Used Priority /dev/sda8 partition 16779260 2086988 -1 Signed-off-by:
Randy Dunlap <rdunlap@infradead.org> Signed-off-by:
-
Huang Ying authored
In some swap scalability test, it is found that there are heavy lock contention on swap cache even if we have split one swap cache radix tree per swap device to one swap cache radix tree every 64 MB trunk in commit 4b3ef9da ("mm/swap: split swap cache into 64MB trunks"). The reason is as follow. After the swap device becomes fragmented so that there's no free swap cluster, the swap device will be scanned linearly to find the free swap slots. swap_info_struct->cluster_next is the next scanning base that is shared by all CPUs. So nearby free swap slots will be allocated for different CPUs. The probability for multiple CPUs to operate on the same 64 MB trunk is high. This causes the lock contention on the swap cache. To solve the issue, in this patch, for SSD swap device, a percpu version next scanning base (cluster_next_cpu) is added. Every CPU will use its own per-cpu next scanning base. And after finishing scanning a 64MB trunk, the per-cpu scanning base will be changed to the beginning of another randomly selected 64MB trunk. In this way, the probability for multiple CPUs to operate on the same 64 MB trunk is reduced greatly. Thus the lock contention is reduced too. For HDD, because sequential access is more important for IO performance, the original shared next scanning base is used. To test the patch, we have run 16-process pmbench memory benchmark on a 2-socket server machine with 48 cores. One ram disk is configured as the swap device per socket. The pmbench working-set size is much larger than the available memory so that swapping is triggered. The memory read/write ratio is 80/20 and the accessing pattern is random. In the original implementation, the lock contention on the swap cache is heavy. The perf profiling data of the lock contention code path is as following, _raw_spin_lock_irq.add_to_swap_cache.add_to_swap.shrink_page_list: 7.91 _raw_spin_lock_irqsave.__remove_mapping.shrink_page_list: 7.11 _raw_spin_lock.swapcache_free_entries.free_swap_slot.__swap_entry_free: 2.51 _raw_spin_lock_irqsave.swap_cgroup_record.mem_cgroup_uncharge_swap: 1.66 _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 1.29 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 1.03 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 0.93 After applying this patch, it becomes, _raw_spin_lock.swapcache_free_entries.free_swap_slot.__swap_entry_free: 3.58 _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 2.3 _raw_spin_lock_irqsave.swap_cgroup_record.mem_cgroup_uncharge_swap: 2.26 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 1.8 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 1.19 The lock contention on the swap cache is almost eliminated. And the pmbench score increases 18.5%. The swapin throughput increases 18.7% from 2.96 GB/s to 3.51 GB/s. While the swapout throughput increases 18.5% from 2.99 GB/s to 3.54 GB/s. We need really fast disk to show the benefit. I have tried this on 2 Intel P3600 NVMe disks. The performance improvement is only about 1%. The improvement should be better on the faster disks, such as Intel Optane disk. [ying.huang@intel.com: fix cluster_next_cpu allocation and freeing, per Daniel] Link: http://lkml.kernel.org/r/20200525002648.336325-1-ying.huang@intel.com [ying.huang@intel.com: v4] Link: http://lkml.kernel.org/r/20200529010840.928819-1-ying.huang@intel.comSigned-off-by:
Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Hugh Dickins <hughd@google.com> Link: http://lkml.kernel.org/r/20200520031502.175659-1-ying.huang@intel.comSigned-off-by:
-
Huang Ying authored
To improve the code readability and take advantage of the common implementation. Signed-off-by:
-
Wei Yang authored
__swap_entry_free() always frees 1 entry. Let's remove the usage. Signed-off-by:
Wei Yang <richard.weiyang@gmail.com> Signed-off-by:
-
Wei Yang authored
swap_info_struct->swap_map[] encodes some flag and count. And to do some condition check, it also introduces some special values. Currently those macros are defined with some magic order, which makes audience hard to understand the exact meaning. This patch split those macros into three categories: flag special value for first swap_map special value for continued swap_map May this help audiences a little. [akpm@linux-foundation.org: tweak capitalization in comments] Signed-off-by: -
Huang Ying authored
Now, the scalability of swap code will drop much when the swap device becomes fragmented, because the swap slots allocation batching stops working. To solve the problem, in this patch, we will try to scan a little more swap slots with restricted effort to batch the swap slots allocation even if the swap device is fragmented. Test shows that the benchmark score can increase up to 37.1% with the patch. Details are as follows. The swap code has a per-cpu cache of swap slots. These batch swap space allocations to improve swap subsystem scaling. In the following code path, add_to_swap() get_swap_page() refill_swap_slots_cache() get_swap_pages() scan_swap_map_slots() scan_swap_map_slots() and get_swap_pages() can return multiple swap slots for each call. These slots will be cached in the per-CPU swap slots cache, so that several following swap slot requests will be fulfilled there to avoid the lock contention in the lower level swap space allocation/freeing code path. But this only works when there are free swap clusters. If a swap device becomes so fragmented that there's no free swap clusters, scan_swap_map_slots() and get_swap_pages() will return only one swap slot for each call in the above code path. Effectively, this falls back to the situation before the swap slots cache was introduced, the heavy lock contention on the swap related locks kills the scalability. Why does it work in this way? Because the swap device could be large, and the free swap slot scanning could be quite time consuming, to avoid taking too much time to scanning free swap slots, the conservative method was used. In fact, this can be improved via scanning a little more free slots with strictly restricted effort. Which is implemented in this patch. In scan_swap_map_slots(), after the first free swap slot is gotten, we will try to scan a little more, but only if we haven't scanned too many slots (< LATENCY_LIMIT). That is, the added scanning latency is strictly restricted. To test the patch, we have run 16-process pmbench memory benchmark on a 2-socket server machine with 48 cores. Multiple ram disks are configured as the swap devices. The pmbench working-set size is much larger than the available memory so that swapping is triggered. The memory read/write ratio is 80/20 and the accessing pattern is random, so the swap space becomes highly fragmented during the test. In the original implementation, the lock contention on swap related locks is very heavy. The perf profiling data of the lock contention code path is as following, _raw_spin_lock.get_swap_pages.get_swap_page.add_to_swap: 21.03 _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 1.92 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 1.72 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 0.69 While after applying this patch, it becomes, _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 4.89 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 3.85 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 1.1 _raw_spin_lock_irqsave.pagevec_lru_move_fn.__lru_cache_add.do_swap_page: 0.88 That is, the lock contention on the swap locks is eliminated. And the pmbench score increases 37.1%. The swapin throughput increases 45.7% from 2.02 GB/s to 2.94 GB/s. While the swapout throughput increases 45.3% from 2.04 GB/s to 2.97 GB/s. Signed-off-by:Tim Chen <tim.c.chen@linux.intel.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Link: http://lkml.kernel.org/r/20200427030023.264780-1-ying.huang@intel.comSigned-off-by:
-
