- 10 Oct, 2014 40 commits
-
-
Minchan Kim authored
zs_get_total_size_bytes returns a amount of memory zsmalloc consumed with *byte unit* but zsmalloc operates *page unit* rather than byte unit so let's change the API so benefit we could get is that reduce unnecessary overhead (ie, change page unit with byte unit) in zsmalloc. Since return type is pages, "zs_get_total_pages" is better than "zs_get_total_size_bytes". Signed-off-by:
Minchan Kim <minchan@kernel.org> Reviewed-by:
Dan Streetman <ddstreet@ieee.org> Cc: Sergey Senozhatsky <sergey.senozhatsky@gmail.com> Cc: Jerome Marchand <jmarchan@redhat.com> Cc: <juno.choi@lge.com> Cc: <seungho1.park@lge.com> Cc: Luigi Semenzato <semenzato@google.com> Cc: Nitin Gupta <ngupta@vflare.org> Cc: Seth Jennings <sjennings@variantweb.net> Cc: David Horner <ds2horner@gmail.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by:
Andrew Morton <akpm@linux-foundation.org> Signed-off-by:
Linus Torvalds <torvalds@linux-foundation.org>
-
Minchan Kim authored
Currently, zram has no feature to limit memory so theoretically zram can deplete system memory. Users have asked for a limit several times as even without exhaustion zram makes it hard to control memory usage of the platform. This patchset adds the feature. Patch 1 makes zs_get_total_size_bytes faster because it would be used frequently in later patches for the new feature. Patch 2 changes zs_get_total_size_bytes's return unit from bytes to page so that zsmalloc doesn't need unnecessary operation(ie, << PAGE_SHIFT). Patch 3 adds new feature. I added the feature into zram layer, not zsmalloc because limiation is zram's requirement, not zsmalloc so any other user using zsmalloc(ie, zpool) shouldn't affected by unnecessary branch of zsmalloc. In future, if every users of zsmalloc want the feature, then, we could move the feature from client side to zsmalloc easily but vice versa would be painful. Patch 4 adds news facility to report maximum memory usage of zram so that this avoids user polling frequently via /sys/block/zram0/ mem_used_total and ensures transient max are not missed. This patch (of 4): pages_allocated has counted in size_class structure and when user of zsmalloc want to see total_size_bytes, it should gather all of count from each size_class to report the sum. It's not bad if user don't see the value often but if user start to see the value frequently, it would be not a good deal for performance pov. This patch moves the count from size_class to zs_pool so it could reduce memory footprint (from [255 * 8byte] to [sizeof(atomic_long_t)]). Signed-off-by:
David Horner <ds2horner@gmail.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by:
-
Davidlohr Bueso authored
Performing vma lookups without taking the mm->mmap_sem is asking for trouble. While doing the search, the vma in question can be modified or even removed before returning to the caller. Take the lock (shared) in order to avoid races while iterating through the vmacache and/or rbtree. In addition, this guarantees that the address space will remain intact during the CPU flushing. Signed-off-by:
Davidlohr Bueso <davidlohr@hp.com> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Signed-off-by:
-
Christoph Lameter authored
vmstat workers are used for folding counter differentials into the zone, per node and global counters at certain time intervals. They currently run at defined intervals on all processors which will cause some holdoff for processors that need minimal intrusion by the OS. The current vmstat_update mechanism depends on a deferrable timer firing every other second by default which registers a work queue item that runs on the local CPU, with the result that we have 1 interrupt and one additional schedulable task on each CPU every 2 seconds If a workload indeed causes VM activity or multiple tasks are running on a CPU, then there are probably bigger issues to deal with. However, some workloads dedicate a CPU for a single CPU bound task. This is done in high performance computing, in high frequency financial applications, in networking (Intel DPDK, EZchip NPS) and with the advent of systems with more and more CPUs over time, this may become more and more common to do since when one has enough CPUs one cares less about efficiently sharing a CPU with other tasks and more about efficiently monopolizing a CPU per task. The difference of having this timer firing and workqueue kernel thread scheduled per second can be enormous. An artificial test measuring the worst case time to do a simple "i++" in an endless loop on a bare metal system and under Linux on an isolated CPU with dynticks and with and without this patch, have Linux match the bare metal performance (~700 cycles) with this patch and loose by couple of orders of magnitude (~200k cycles) without it[*]. The loss occurs for something that just calculates statistics. For networking applications, for example, this could be the difference between dropping packets or sustaining line rate. Statistics are important and useful, but it would be great if there would be a way to not cause statistics gathering produce a huge performance difference. This patche does just that. This patch creates a vmstat shepherd worker that monitors the per cpu differentials on all processors. If there are differentials on a processor then a vmstat worker local to the processors with the differentials is created. That worker will then start folding the diffs in regular intervals. Should the worker find that there is no work to be done then it will make the shepherd worker monitor the differentials again. With this patch it is possible then to have periods longer than 2 seconds without any OS event on a "cpu" (hardware thread). The patch shows a very minor increased in system performance. hackbench -s 512 -l 2000 -g 15 -f 25 -P Results before the patch: Running in process mode with 15 groups using 50 file descriptors each (== 750 tasks) Each sender will pass 2000 messages of 512 bytes Time: 4.992 Running in process mode with 15 groups using 50 file descriptors each (== 750 tasks) Each sender will pass 2000 messages of 512 bytes Time: 4.971 Running in process mode with 15 groups using 50 file descriptors each (== 750 tasks) Each sender will pass 2000 messages of 512 bytes Time: 5.063 Hackbench after the patch: Running in process mode with 15 groups using 50 file descriptors each (== 750 tasks) Each sender will pass 2000 messages of 512 bytes Time: 4.973 Running in process mode with 15 groups using 50 file descriptors each (== 750 tasks) Each sender will pass 2000 messages of 512 bytes Time: 4.990 Running in process mode with 15 groups using 50 file descriptors each (== 750 tasks) Each sender will pass 2000 messages of 512 bytes Time: 4.993 [fengguang.wu@intel.com: cpu_stat_off can be static] Signed-off-by:
Christoph Lameter <cl@linux.com> Reviewed-by:
Gilad Ben-Yossef <gilad@benyossef.com> Cc: Frederic Weisbecker <fweisbec@gmail.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tejun Heo <tj@kernel.org> Cc: John Stultz <john.stultz@linaro.org> Cc: Mike Frysinger <vapier@gentoo.org> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Hakan Akkan <hakanakkan@gmail.com> Cc: Max Krasnyansky <maxk@qti.qualcomm.com> Cc: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com> Cc: Hugh Dickins <hughd@google.com> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Signed-off-by:
Fengguang Wu <fengguang.wu@intel.com> Signed-off-by:
-
Jean Delvare authored
It isn't obvious that CMA can be disabled on the kernel's command line, so document it. Signed-off-by:
Jean Delvare <jdelvare@suse.de> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Akinobu Mita <akinobu.mita@gmail.com> Cc: Chuck Ebbert <cebbert.lkml@gmail.com> Cc: Marek Szyprowski <m.szyprowski@samsung.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Signed-off-by:
-
Sebastien Buisson authored
Increase the buffer-head per-CPU LRU size to allow efficient filesystem operations that access many blocks for each transaction. For example, creating a file in a large ext4 directory with quota enabled will access multiple buffer heads and will overflow the LRU at the default 8-block LRU size: * parent directory inode table block (ctime, nlinks for subdirs) * new inode bitmap * inode table block * 2 quota blocks * directory leaf block (not reused, but pollutes one cache entry) * 2 levels htree blocks (only one is reused, other pollutes cache) * 2 levels indirect/index blocks (only one is reused) The buffer-head per-CPU LRU size is raised to 16, as it shows in metadata performance benchmarks up to 10% gain for create, 4% for lookup and 7% for destroy. Signed-off-by:
Liang Zhen <liang.zhen@intel.com> Signed-off-by:
Andreas Dilger <andreas.dilger@intel.com> Signed-off-by:
Sebastien Buisson <sebastien.buisson@bull.net> Signed-off-by:
-
Mel Gorman authored
PROT_NUMA VMAs are skipped to avoid problems distinguishing between present, prot_none and special entries. MPOL_MF_LAZY is not visible from userspace since commit a720094d ("mm: mempolicy: Hide MPOL_NOOP and MPOL_MF_LAZY from userspace for now") but it should still skip VMAs the same way task_numa_work does. Signed-off-by:
Mel Gorman <mgorman@suse.de> Acked-by:
Rik van Riel <riel@redhat.com> Acked-by:
Hugh Dickins <hughd@google.com> Acked-by:
Peter Zijlstra <peterz@infradead.org> Signed-off-by:
-
Konstantin Khlebnikov authored
This tool induces memory fragmentation via sequential allocation of transparent huge pages and splitting off everything except their last sub-pages. It easily generates pressure to the memory compaction code. $ perf stat -e 'compaction:*' -e 'migrate:*' ./transhuge-stress transhuge-stress: allocate 7858 transhuge pages, using 15716 MiB virtual memory and 61 MiB of ram transhuge-stress: 1.653 s/loop, 0.210 ms/page, 9504.828 MiB/s 7858 succeed, 0 failed, 2439 different pages transhuge-stress: 1.537 s/loop, 0.196 ms/page, 10226.227 MiB/s 7858 succeed, 0 failed, 2364 different pages transhuge-stress: 1.658 s/loop, 0.211 ms/page, 9479.215 MiB/s 7858 succeed, 0 failed, 2179 different pages transhuge-stress: 1.617 s/loop, 0.206 ms/page, 9716.992 MiB/s 7858 succeed, 0 failed, 2421 different pages ^C./transhuge-stress: Interrupt Performance counter stats for './transhuge-stress': 1.744.051 compaction:mm_compaction_isolate_migratepages 1.014 compaction:mm_compaction_isolate_freepages 1.744.051 compaction:mm_compaction_migratepages 1.647 compaction:mm_compaction_begin 1.647 compaction:mm_compaction_end 1.744.051 migrate:mm_migrate_pages 0 migrate:mm_numa_migrate_ratelimit 7,964696835 seconds time elapsed Signed-off-by:Konstantin Khlebnikov <koct9i@gmail.com> Cc: Rafael Aquini <aquini@redhat.com> Cc: Andrey Ryabinin <ryabinin.a.a@gmail.com> Cc: Shuah Khan <shuahkh@osg.samsung.com> Signed-off-by:
-
Konstantin Khlebnikov authored
Always mark pages with PageBalloon even if balloon compaction is disabled and expose this mark in /proc/kpageflags as KPF_BALLOON. Also this patch adds three counters into /proc/vmstat: "balloon_inflate", "balloon_deflate" and "balloon_migrate". They accumulate balloon activity. Current size of balloon is (balloon_inflate - balloon_deflate) pages. All generic balloon code now gathered under option CONFIG_MEMORY_BALLOON. It should be selected by ballooning driver which wants use this feature. Currently virtio-balloon is the only user. Signed-off-by:
Konstantin Khlebnikov <k.khlebnikov@samsung.com> Cc: Rafael Aquini <aquini@redhat.com> Cc: Andrey Ryabinin <ryabinin.a.a@gmail.com> Signed-off-by:
-
Konstantin Khlebnikov authored
Now ballooned pages are detected using PageBalloon(). Fake mapping is no longer required. This patch links ballooned pages to balloon device using field page->private instead of page->mapping. Also this patch embeds balloon_dev_info directly into struct virtio_balloon. Signed-off-by:
-
Konstantin Khlebnikov authored
Sasha Levin reported KASAN splash inside isolate_migratepages_range(). Problem is in the function __is_movable_balloon_page() which tests AS_BALLOON_MAP in page->mapping->flags. This function has no protection against anonymous pages. As result it tried to check address space flags inside struct anon_vma. Further investigation shows more problems in current implementation: * Special branch in __unmap_and_move() never works: balloon_page_movable() checks page flags and page_count. In __unmap_and_move() page is locked, reference counter is elevated, thus balloon_page_movable() always fails. As a result execution goes to the normal migration path. virtballoon_migratepage() returns MIGRATEPAGE_BALLOON_SUCCESS instead of MIGRATEPAGE_SUCCESS, move_to_new_page() thinks this is an error code and assigns newpage->mapping to NULL. Newly migrated page lose connectivity with balloon an all ability for further migration. * lru_lock erroneously required in isolate_migratepages_range() for isolation ballooned page. This function releases lru_lock periodically, this makes migration mostly impossible for some pages. * balloon_page_dequeue have a tight race with balloon_page_isolate: balloon_page_isolate could be executed in parallel with dequeue between picking page from list and locking page_lock. Race is rare because they use trylock_page() for locking. This patch fixes all of them. Instead of fake mapping with special flag this patch uses special state of page->_mapcount: PAGE_BALLOON_MAPCOUNT_VALUE = -256. Buddy allocator uses PAGE_BUDDY_MAPCOUNT_VALUE = -128 for similar purpose. Storing mark directly in struct page makes everything safer and easier. PagePrivate is used to mark pages present in page list (i.e. not isolated, like PageLRU for normal pages). It replaces special rules for reference counter and makes balloon migration similar to migration of normal pages. This flag is protected by page_lock together with link to the balloon device. Signed-off-by:
Sasha Levin <sasha.levin@oracle.com> Link: http://lkml.kernel.org/p/53E6CEAA.9020105@oracle.com Cc: Rafael Aquini <aquini@redhat.com> Cc: Andrey Ryabinin <ryabinin.a.a@gmail.com> Cc: <stable@vger.kernel.org> [3.8+] Signed-off-by:
-
Steve Capper authored
Activate the RCU fast_gup for ARM64. We also need to force THP splits to broadcast an IPI s.t. we block in the fast_gup page walker. As THP splits are comparatively rare, this should not lead to a noticeable performance degradation. Some pre-requisite functions pud_write and pud_page are also added. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by:
Steve Capper <steve.capper@linaro.org> Tested-by:
Dann Frazier <dann.frazier@canonical.com> Acked-by:
Catalin Marinas <catalin.marinas@arm.com> Cc: Hugh Dickins <hughd@google.com> Cc: Russell King <rmk@arm.linux.org.uk> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Will Deacon <will.deacon@arm.com> Cc: Christoffer Dall <christoffer.dall@linaro.org> Cc: Andrea Arcangeli <aarcange@redhat.com> Signed-off-by:
-
Steve Capper authored
In order to implement fast_get_user_pages we need to ensure that the page table walker is protected from page table pages being freed from under it. This patch enables HAVE_RCU_TABLE_FREE, any page table pages belonging to address spaces with multiple users will be call_rcu_sched freed. Meaning that disabling interrupts will block the free and protect the fast gup page walker. Signed-off-by:
-
Steve Capper authored
Activate the RCU fast_gup for ARM. We also need to force THP splits to broadcast an IPI s.t. we block in the fast_gup page walker. As THP splits are comparatively rare, this should not lead to a noticeable performance degradation. Some pre-requisite functions pud_write and pud_page are also added. Signed-off-by:
-
Steve Capper authored
In order to implement fast_get_user_pages we need to ensure that the page table walker is protected from page table pages being freed from under it. This patch enables HAVE_RCU_TABLE_FREE, any page table pages belonging to address spaces with multiple users will be call_rcu_sched freed. Meaning that disabling interrupts will block the free and protect the fast gup page walker. Signed-off-by:
-
Steve Capper authored
We need a mechanism to tag ptes as being special, this indicates that no attempt should be made to access the underlying struct page * associated with the pte. This is used by the fast_gup when operating on ptes as it has no means to access VMAs (that also contain this information) locklessly. The L_PTE_SPECIAL bit is already allocated for LPAE, this patch modifies pte_special and pte_mkspecial to make use of it, and defines __HAVE_ARCH_PTE_SPECIAL. This patch also excludes special ptes from the icache/dcache sync logic. Signed-off-by:
-
Steve Capper authored
This series implements general forms of get_user_pages_fast and __get_user_pages_fast in core code and activates them for arm and arm64. These are required for Transparent HugePages to function correctly, as a futex on a THP tail will otherwise result in an infinite loop (due to the core implementation of __get_user_pages_fast always returning 0). Unfortunately, a futex on THP tail can be quite common for certain workloads; thus THP is unreliable without a __get_user_pages_fast implementation. This series may also be beneficial for direct-IO heavy workloads and certain KVM workloads. This patch (of 6): get_user_pages_fast() attempts to pin user pages by walking the page tables directly and avoids taking locks. Thus the walker needs to be protected from page table pages being freed from under it, and needs to block any THP splits. One way to achieve this is to have the walker disable interrupts, and rely on IPIs from the TLB flushing code blocking before the page table pages are freed. On some platforms we have hardware broadcast of TLB invalidations, thus the TLB flushing code doesn't necessarily need to broadcast IPIs; and spuriously broadcasting IPIs can hurt system performance if done too often. This problem has been solved on PowerPC and Sparc by batching up page table pages belonging to more than one mm_user, then scheduling an rcu_sched callback to free the pages. This RCU page table free logic has been promoted to core code and is activated when one enables HAVE_RCU_TABLE_FREE. Unfortunately, these architectures implement their own get_user_pages_fast routines. The RCU page table free logic coupled with an IPI broadcast on THP split (which is a rare event), allows one to protect a page table walker by merely disabling the interrupts during the walk. This patch provides a general RCU implementation of get_user_pages_fast that can be used by architectures that perform hardware broadcast of TLB invalidations. It is based heavily on the PowerPC implementation by Nick Piggin. [akpm@linux-foundation.org: various comment fixes] Signed-off-by:
-
Paul McQuade authored
Remove 3 brace coding style for any arm of this statement Signed-off-by:
Paul McQuade <paulmcquad@gmail.com> Signed-off-by:
-
Paul McQuade authored
WARNING: Prefer: pr_err(... to printk(KERN_ERR ... [akpm@linux-foundation.org: remove KERN_ERR] Signed-off-by:
-
Yasuaki Ishimatsu authored
By the following commits, we prevented from allocating firmware_map_entry of same memory range: f0093ede: drivers/firmware/memmap.c: don't allocate firmware_map_entry of same memory range 49c8b24d: drivers/firmware/memmap.c: pass the correct argument to firmware_map_find_entry_bootmem() But it's not enough. When PNP0C80 device is added by acpi_scan_init(), memmap sysfses of same firmware_map_entry are created twice as follows: # cat /sys/firmware/memmap/*/start 0x40000000000 0x60000000000 0x4a837000 0x4a83a000 0x4a8b5000 ... 0x40000000000 0x60000000000 ... The flows of the issues are as follows: 1. e820_reserve_resources() allocates firmware_map_entrys of all memory ranges defined in e820. And, these firmware_map_entrys are linked with map_entries list. map_entries -> entry 1 -> ... -> entry N 2. When PNP0C80 device is limited by mem= boot option, acpi_scan_init() added the memory device. In this case, firmware_map_add_hotplug() allocates firmware_map_entry and creates memmap sysfs. map_entries -> entry 1 -> ... -> entry N -> entry N+1 | memmap 1 3. firmware_memmap_init() creates memmap sysfses of firmware_map_entrys linked with map_entries. map_entries -> entry 1 -> ... -> entry N -> entry N+1 | | | memmap 2 memmap N+1 memmap 1 memmap N+2 So while hot removing the PNP0C80 device, kernel panic occurs as follows: BUG: unable to handle kernel paging request at 00000001003e000b IP: sysfs_open_file+0x46/0x2b0 PGD 203a89fe067 PUD 0 Oops: 0000 [#1] SMP ... Call Trace: do_dentry_open+0x1ef/0x2a0 finish_open+0x31/0x40 do_last+0x57c/0x1220 path_openat+0xc2/0x4c0 do_filp_open+0x4b/0xb0 do_sys_open+0xf3/0x1f0 SyS_open+0x1e/0x20 system_call_fastpath+0x16/0x1b The patch adds a check of confirming whether memmap sysfs of firmware_map_entry has been created, and does not create memmap sysfs of same firmware_map_entry. Signed-off-by:
Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com> Cc: Santosh Shilimkar <santosh.shilimkar@ti.com> Cc: Toshi Kani <toshi.kani@hp.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Signed-off-by:
-
Paul McQuade authored
Replace asm. headers with linux/headers: <linux/bug.h> <linux/io.h> Signed-off-by:
-
Paul McQuade authored
Signed-off-by:
-
Paul McQuade authored
"WARNING: Use #include <linux/uaccess.h> instead of <asm/uaccess.h>" Signed-off-by:
-
Vladimir Davydov authored
memcg_can_account_kmem() returns true iff !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) && memcg_kmem_is_active(memcg); To begin with the !mem_cgroup_is_root(memcg) check is useless, because one can't enable kmem accounting for the root cgroup (mem_cgroup_write() returns EINVAL on an attempt to set the limit on the root cgroup). Furthermore, the !mem_cgroup_disabled() check also seems to be redundant. The point is memcg_can_account_kmem() is called from three places: mem_cgroup_salbinfo_read(), __memcg_kmem_get_cache(), and __memcg_kmem_newpage_charge(). The latter two functions are only invoked if memcg_kmem_enabled() returns true, which implies that the memory cgroup subsystem is enabled. And mem_cgroup_slabinfo_read() shows the output of memory.kmem.slabinfo, which won't exist if the memory cgroup is completely disabled. So let's substitute all the calls to memcg_can_account_kmem() with plain memcg_kmem_is_active(), and kill the former. Signed-off-by:Vladimir Davydov <vdavydov@parallels.com> Acked-by:
Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@suse.cz> Signed-off-by:
-
Johannes Weiner authored
In a memcg with even just moderate cache pressure, success rates for transparent huge page allocations drop to zero, wasting a lot of effort that the allocator puts into assembling these pages. The reason for this is that the memcg reclaim code was never designed for higher-order charges. It reclaims in small batches until there is room for at least one page. Huge page charges only succeed when these batches add up over a series of huge faults, which is unlikely under any significant load involving order-0 allocations in the group. Remove that loop on the memcg side in favor of passing the actual reclaim goal to direct reclaim, which is already set up and optimized to meet higher-order goals efficiently. This brings memcg's THP policy in line with the system policy: if the allocator painstakingly assembles a hugepage, memcg will at least make an honest effort to charge it. As a result, transparent hugepage allocation rates amid cache activity are drastically improved: vanilla patched pgalloc 4717530.80 ( +0.00%) 4451376.40 ( -5.64%) pgfault 491370.60 ( +0.00%) 225477.40 ( -54.11%) pgmajfault 2.00 ( +0.00%) 1.80 ( -6.67%) thp_fault_alloc 0.00 ( +0.00%) 531.60 (+100.00%) thp_fault_fallback 749.00 ( +0.00%) 217.40 ( -70.88%) [ Note: this may in turn increase memory consumption from internal fragmentation, which is an inherent risk of transparent hugepages. Some setups may have to adjust the memcg limits accordingly to accomodate this - or, if the machine is already packed to capacity, disable the transparent huge page feature. ] Signed-off-by: -
Johannes Weiner authored
When attempting to charge pages, we first charge the memory counter and then the memory+swap counter. If one of the counters is at its limit, we enter reclaim, but if it's the memory+swap counter, reclaim shouldn't swap because that wouldn't change the situation. However, if the counters have the same limits, we never get to the memory+swap limit. To know whether reclaim should swap or not, there is a state flag that indicates whether the limits are equal and whether hitting the memory limit implies hitting the memory+swap limit. Just try the memory+swap counter first. Signed-off-by:
Michal Hocko <mhocko@suse.cz> Cc: Dave Hansen <dave@sr71.net> Cc: Greg Thelen <gthelen@google.com> Signed-off-by:
-
Michal Hocko authored
free_pages_and_swap_cache limits release_pages to PAGEVEC_SIZE chunks. This is not a big deal for the normal release path but it completely kills memcg uncharge batching which reduces res_counter spin_lock contention. Dave has noticed this with his page fault scalability test case on a large machine when the lock was basically dominating on all CPUs: 80.18% 80.18% [kernel] [k] _raw_spin_lock | --- _raw_spin_lock | |--66.59%-- res_counter_uncharge_until | res_counter_uncharge | uncharge_batch | uncharge_list | mem_cgroup_uncharge_list | release_pages | free_pages_and_swap_cache | tlb_flush_mmu_free | | | |--90.12%-- unmap_single_vma | | unmap_vmas | | unmap_region | | do_munmap | | vm_munmap | | sys_munmap | | system_call_fastpath | | __GI___munmap | | | --9.88%-- tlb_flush_mmu | tlb_finish_mmu | unmap_region | do_munmap | vm_munmap | sys_munmap | system_call_fastpath | __GI___munmap In his case the load was running in the root memcg and that part has been handled by reverting 05b84301 ("mm: memcontrol: use root_mem_cgroup res_counter") because this is a clear regression, but the problem remains inside dedicated memcgs. There is no reason to limit release_pages to PAGEVEC_SIZE batches other than lru_lock held times. This logic, however, can be moved inside the function. mem_cgroup_uncharge_list and free_hot_cold_page_list do not hold any lock for the whole pages_to_free list so it is safe to call them in a single run. The release_pages() code was previously breaking the lru_lock each PAGEVEC_SIZE pages (ie, 14 pages). However this code has no usage of pagevecs so switch to breaking the lock at least every SWAP_CLUSTER_MAX (32) pages. This means that the lock acquisition frequency is approximately halved and the max hold times are approximately doubled. The now unneeded batching is removed from free_pages_and_swap_cache(). Also update the grossly out-of-date release_pages documentation. Signed-off-by:Dave Hansen <dave@sr71.net> Cc: Vladimir Davydov <vdavydov@parallels.com> Cc: Greg Thelen <gthelen@google.com> Signed-off-by:
-
Sebastian Andrzej Siewior authored
cat /sys/.../pools followed by removal the device leads to: |====================================================== |[ INFO: possible circular locking dependency detected ] |3.17.0-rc4+ #1498 Not tainted |------------------------------------------------------- |rmmod/2505 is trying to acquire lock: | (s_active#28){++++.+}, at: [<c017f754>] kernfs_remove_by_name_ns+0x3c/0x88 | |but task is already holding lock: | (pools_lock){+.+.+.}, at: [<c011494c>] dma_pool_destroy+0x18/0x17c | |which lock already depends on the new lock. |the existing dependency chain (in reverse order) is: | |-> #1 (pools_lock){+.+.+.}: | [<c0114ae8>] show_pools+0x30/0xf8 | [<c0313210>] dev_attr_show+0x1c/0x48 | [<c0180e84>] sysfs_kf_seq_show+0x88/0x10c | [<c017f960>] kernfs_seq_show+0x24/0x28 | [<c013efc4>] seq_read+0x1b8/0x480 | [<c011e820>] vfs_read+0x8c/0x148 | [<c011ea10>] SyS_read+0x40/0x8c | [<c000e960>] ret_fast_syscall+0x0/0x48 | |-> #0 (s_active#28){++++.+}: | [<c017e9ac>] __kernfs_remove+0x258/0x2ec | [<c017f754>] kernfs_remove_by_name_ns+0x3c/0x88 | [<c0114a7c>] dma_pool_destroy+0x148/0x17c | [<c03ad288>] hcd_buffer_destroy+0x20/0x34 | [<c03a4780>] usb_remove_hcd+0x110/0x1a4 The problem is the lock order of pools_lock and kernfs_mutex in dma_pool_destroy() vs show_pools() call path. This patch breaks out the creation of the sysfs file outside of the pools_lock mutex. The newly added pools_reg_lock ensures that there is no race of create vs destroy code path in terms whether or not the sysfs file has to be deleted (and was it deleted before we try to create a new one) and what to do if device_create_file() failed. Signed-off-by:Sebastian Andrzej Siewior <bigeasy@linutronix.de> Signed-off-by:
-
Vladimir Davydov authored
`While growing per memcg caches arrays, we jump between memcontrol.c and slab_common.c in a weird way: memcg_alloc_cache_id - memcontrol.c memcg_update_all_caches - slab_common.c memcg_update_cache_size - memcontrol.c There's absolutely no reason why memcg_update_cache_size can't live on the slab's side though. So let's move it there and settle it comfortably amid per-memcg cache allocation functions. Besides, this patch cleans this function up a bit, removing all the useless comments from it, and renames it to memcg_update_cache_params to conform to memcg_alloc/free_cache_params, which we already have in slab_common.c. Signed-off-by: -
Vladimir Davydov authored
memcg_update_all_caches grows arrays of per-memcg caches, so we only need to call it when memcg_limited_groups_array_size is increased. However, currently we invoke it each time a new kmem-active memory cgroup is created. Then it just iterates over all slab_caches and does nothing (memcg_update_cache_size returns immediately). This patch fixes this insanity. In the meantime it moves the code dealing with id allocations to separate functions, memcg_alloc_cache_id and memcg_free_cache_id. Signed-off-by:
-
Vladimir Davydov authored
The only reason why they live in memcontrol.c is that we get/put css reference to the owner memory cgroup in them. However, we can do that in memcg_{un,}register_cache. OTOH, there are several reasons to move them to slab_common.c. First, I think that the less public interface functions we have in memcontrol.h the better. Since the functions I move don't depend on memcontrol, I think it's worth making them private to slab, especially taking into account that the arrays are defined on the slab's side too. Second, the way how per-memcg arrays are updated looks rather awkward: it proceeds from memcontrol.c (__memcg_activate_kmem) to slab_common.c (memcg_update_all_caches) and back to memcontrol.c again (memcg_update_array_size). In the following patches I move the function relocating the arrays (memcg_update_array_size) to slab_common.c and therefore get rid this circular call path. I think we should have the cache allocation stuff in the same place where we have relocation, because it's easier to follow the code then. So I move arrays alloc/free functions to slab_common.c too. The third point isn't obvious. I'm going to make the list_lru structure per-memcg to allow targeted kmem reclaim. That means we will have per-memcg arrays in list_lrus too. It turns out that it's much easier to update these arrays in list_lru.c rather than in memcontrol.c, because all the stuff we need is defined there. This patch makes memcg caches arrays allocation path conform that of the upcoming list_lru. So let's move these functions to slab_common.c and make them static. Signed-off-by: -
Andrew Morton authored
- s/KERN_ALERT/pr_emerg/: we're going BUG so let's maximize the changes of getting the message out. - convert debug.c to pr_foo() Cc: Sasha Levin <sasha.levin@oracle.com> Signed-off-by:
-
Sasha Levin authored
Dump the contents of the relevant struct_mm when we hit the bug condition. Signed-off-by:
-
Sasha Levin authored
Very similar to VM_BUG_ON_PAGE and VM_BUG_ON_VMA, dump struct_mm when the bug is hit. [akpm@linux-foundation.org: coding-style fixes] [mhocko@suse.cz: fix build] [mhocko@suse.cz: fix build some more] [akpm@linux-foundation.org: do strange things to avoid doing strange things for the comma separators] Signed-off-by:
-
Sasha Levin authored
dump_page() and dump_vma() are not specific to page_alloc.c, move them out so page_alloc.c won't turn into the unofficial debug repository. Signed-off-by:
-
Peter Feiner authored
If a /proc/pid/pagemap read spans a [VMA, an unmapped region, then a VM_SOFTDIRTY VMA], the virtual pages in the unmapped region are reported as softdirty. Here's a program to demonstrate the bug: int main() { const uint64_t PAGEMAP_SOFTDIRTY = 1ul << 55; uint64_t pme[3]; int fd = open("/proc/self/pagemap", O_RDONLY);; char *m = mmap(NULL, 3 * getpagesize(), PROT_READ, MAP_ANONYMOUS | MAP_SHARED, -1, 0); munmap(m + getpagesize(), getpagesize()); pread(fd, pme, 24, (unsigned long) m / getpagesize() * 8); assert(pme[0] & PAGEMAP_SOFTDIRTY); /* passes */ assert(!(pme[1] & PAGEMAP_SOFTDIRTY)); /* fails */ assert(pme[2] & PAGEMAP_SOFTDIRTY); /* passes */ return 0; } (Note that all pages in new VMAs are softdirty until cleared). Tested: Used the program given above. I'm going to include this code in a selftest in the future. [n-horiguchi@ah.jp.nec.com: prevent pagemap_pte_range() from overrunning] Signed-off-by:Peter Feiner <pfeiner@google.com> Cc: "Kirill A. Shutemov" <kirill@shutemov.name> Cc: Cyrill Gorcunov <gorcunov@openvz.org> Cc: Pavel Emelyanov <xemul@parallels.com> Cc: Jamie Liu <jamieliu@google.com> Cc: Hugh Dickins <hughd@google.com> Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com> Signed-off-by:
Naoya Horiguchi <n-horiguchi@ah.jp.nec.com> Signed-off-by:
-
Mel Gorman authored
Zones are allocated by the page allocator in either node or zone order. Node ordering is preferred in terms of locality and is applied automatically in one of three cases: 1. If a node has only low memory 2. If DMA/DMA32 is a high percentage of memory 3. If low memory on a single node is greater than 70% of the node size Otherwise zone ordering is used to preserve low memory for devices that require it. Unfortunately a consequence of this is that applications running on a machine with balanced NUMA nodes will experience different performance characteristics depending on which node they happen to start from. The point of zone ordering is to protect lower zones for devices that require DMA/DMA32 memory. When NUMA was first introduced, this was critical as 32-bit NUMA machines existed and exhausting low memory triggered OOMs easily as so many allocations required low memory. On 64-bit machines the primary concern is devices that are 32-bit only which is less severe than the low memory exhaustion problem on 32-bit NUMA. It seems there are really few devices that depends on it. AGP -- I assume this is getting more rare but even then I think the allocations happen early in boot time where lowmem pressure is less of a problem DRM -- If the device is 32-bit only then there may be low pressure. I didn't evaluate these in detail but it looks like some of these are mobile graphics card. Not many NUMA laptops out there. DRM folk should know better though. Some TV cards -- Much demand for 32-bit capable TV cards on NUMA machines? B43 wireless card -- again not really a NUMA thing. I cannot find a good reason to incur a performance penalty on all 64-bit NUMA machines in case someone throws a brain damanged TV or graphics card in there. This patch defaults to node-ordering on 64-bit NUMA machines. I was tempted to make it default everywhere but I understand that some embedded arches may be using 32-bit NUMA where I cannot predict the consequences. The performance impact depends on the workload and the characteristics of the machine and the machine I tested on had a large Normal zone on node 0 so the impact is within the noise for the majority of tests. The allocation stats show more allocation requests were from DMA32 and local node. Running SpecJBB with multiple JVMs and automatic NUMA balancing disabled the results were specjbb 3.17.0-rc2 3.17.0-rc2 vanilla nodeorder-v1r1 Min 1 29534.00 ( 0.00%) 30020.00 ( 1.65%) Min 10 115717.00 ( 0.00%) 134038.00 ( 15.83%) Min 19 109718.00 ( 0.00%) 114186.00 ( 4.07%) Min 28 104459.00 ( 0.00%) 103639.00 ( -0.78%) Min 37 98245.00 ( 0.00%) 103756.00 ( 5.61%) Min 46 97198.00 ( 0.00%) 96197.00 ( -1.03%) Mean 1 30953.25 ( 0.00%) 31917.75 ( 3.12%) Mean 10 124432.50 ( 0.00%) 140904.00 ( 13.24%) Mean 19 116033.50 ( 0.00%) 119294.75 ( 2.81%) Mean 28 108365.25 ( 0.00%) 106879.50 ( -1.37%) Mean 37 102984.75 ( 0.00%) 106924.25 ( 3.83%) Mean 46 100783.25 ( 0.00%) 105368.50 ( 4.55%) Stddev 1 1260.38 ( 0.00%) 1109.66 ( 11.96%) Stddev 10 7434.03 ( 0.00%) 5171.91 ( 30.43%) Stddev 19 8453.84 ( 0.00%) 5309.59 ( 37.19%) Stddev 28 4184.55 ( 0.00%) 2906.63 ( 30.54%) Stddev 37 5409.49 ( 0.00%) 3192.12 ( 40.99%) Stddev 46 4521.95 ( 0.00%) 7392.52 (-63.48%) Max 1 32738.00 ( 0.00%) 32719.00 ( -0.06%) Max 10 136039.00 ( 0.00%) 148614.00 ( 9.24%) Max 19 130566.00 ( 0.00%) 127418.00 ( -2.41%) Max 28 115404.00 ( 0.00%) 111254.00 ( -3.60%) Max 37 112118.00 ( 0.00%) 111732.00 ( -0.34%) Max 46 108541.00 ( 0.00%) 116849.00 ( 7.65%) TPut 1 123813.00 ( 0.00%) 127671.00 ( 3.12%) TPut 10 497730.00 ( 0.00%) 563616.00 ( 13.24%) TPut 19 464134.00 ( 0.00%) 477179.00 ( 2.81%) TPut 28 433461.00 ( 0.00%) 427518.00 ( -1.37%) TPut 37 411939.00 ( 0.00%) 427697.00 ( 3.83%) TPut 46 403133.00 ( 0.00%) 421474.00 ( 4.55%) 3.17.0-rc2 3.17.0-rc2 vanillanodeorder-v1r1 DMA allocs 0 0 DMA32 allocs 57 1491992 Normal allocs 32543566 30026383 Movable allocs 0 0 Direct pages scanned 0 0 Kswapd pages scanned 0 0 Kswapd pages reclaimed 0 0 Direct pages reclaimed 0 0 Kswapd efficiency 100% 100% Kswapd velocity 0.000 0.000 Direct efficiency 100% 100% Direct velocity 0.000 0.000 Percentage direct scans 0% 0% Zone normal velocity 0.000 0.000 Zone dma32 velocity 0.000 0.000 Zone dma velocity 0.000 0.000 THP fault alloc 55164 52987 THP collapse alloc 139 147 THP splits 26 21 NUMA alloc hit 4169066 4250692 NUMA alloc miss 0 0 Note that there were more DMA32 allocations with the patch applied. In this particular case there was no difference in numa_hit and numa_miss. The expectation is that DMA32 was being used at the low watermark instead of falling into the slow path. kswapd was not woken but it's not worken for THP allocations. On 32-bit, this patch defaults to zone-ordering as low memory depletion can be a serious problem on 32-bit large memory machines. If the default ordering was node then processes on node 0 will deplete the Normal zone due to normal activity. The problem is worse if CONFIG_HIGHPTE is not set. If combined with large amounts of dirty/writeback pages in Normal zone then there is also a high risk of OOM. The heuristics are removed as it's not clear they were ever important on 32-bit. They were only relevant for setting node-ordering on 64-bit. Signed-off-by: -
Mel Gorman authored
Since 2.6.24 there has been a paranoid check in move_freepages that looks up the zone of two pages. This is a very slow path and the only time I've seen this bug trigger recently is when memory initialisation was broken during patch development. Despite the fact it's a slow path, this patch converts the check to a VM_BUG_ON anyway as it has served its purpose by now. Signed-off-by:
David Rientjes <rientjes@google.com> Acked-by:
-
Xue jiufei authored
Fix a deadlock problem caused by direct memory reclaim in o2net_wq. The situation is as follows: 1) Receive a connect message from another node, node queues a work_struct o2net_listen_work. 2) o2net_wq processes this work and call the following functions: o2net_wq -> o2net_accept_one -> sock_create_lite -> sock_alloc() -> kmem_cache_alloc with GFP_KERNEL -> ____cache_alloc_node ->__alloc_pages_nodemask -> do_try_to_free_pages -> shrink_slab -> evict -> ocfs2_evict_inode -> ocfs2_drop_lock -> dlmunlock -> o2net_send_message_vec then o2net_wq wait for the unlock reply from master. 3) tcp layer received the reply, call o2net_data_ready() and queue sc_rx_work, waiting o2net_wq to process this work. 4) o2net_wq is a single thread workqueue, it process the work one by one. Right now it is still doing o2net_listen_work and cannot handle sc_rx_work. so we deadlock. Junxiao Bi's patch "mm: clear __GFP_FS when PF_MEMALLOC_NOIO is set" (http://ozlabs.org/~akpm/mmots/broken-out/mm-clear-__gfp_fs-when-pf_memalloc_noio-is-set.patch) clears __GFP_FS in memalloc_noio_flags() besides __GFP_IO. We use memalloc_noio_save() to set process flag PF_MEMALLOC_NOIO so that all allocations done by this process are done as if GFP_NOIO was specified. We are not reentering filesystem while doing memory reclaim. Signed-off-by:joyce.xue <xuejiufei@huawei.com> Cc: Junxiao Bi <junxiao.bi@oracle.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Mark Fasheh <mfasheh@suse.com> Signed-off-by:
-
Junxiao Bi authored
commit 21caf2fc ("mm: teach mm by current context info to not do I/O during memory allocation") introduces PF_MEMALLOC_NOIO flag to avoid doing I/O inside memory allocation, __GFP_IO is cleared when this flag is set, but __GFP_FS implies __GFP_IO, it should also be cleared. Or it may still run into I/O, like in superblock shrinker. And this will make the kernel run into the deadlock case described in that commit. See Dave Chinner's comment about io in superblock shrinker: Filesystem shrinkers do indeed perform IO from the superblock shrinker and have for years. Even clean inodes can require IO before they can be freed - e.g. on an orphan list, need truncation of post-eof blocks, need to wait for ordered operations to complete before it can be freed, etc. IOWs, Ext4, btrfs and XFS all can issue and/or block on arbitrary amounts of IO in the superblock shrinker context. XFS, in particular, has been doing transactions and IO from the VFS inode cache shrinker since it was first introduced.... Fix this by clearing __GFP_FS in memalloc_noio_flags(), this function has masked all the gfp_mask that will be passed into fs for the processes setting PF_MEMALLOC_NOIO in the direct reclaim path. v1 thread at: https://lkml.org/lkml/2014/9/3/32Signed-off-by:
Junxiao Bi <junxiao.bi@oracle.com> Cc: Dave Chinner <david@fromorbit.com> Cc: joyce.xue <xuejiufei@huawei.com> Cc: Ming Lei <ming.lei@canonical.com> Cc: Trond Myklebust <trond.myklebust@primarydata.com> Cc: <stable@vger.kernel.org> Signed-off-by:
-
