- 07 Apr, 2014 40 commits
-
-
Sergey Senozhatsky authored
This is a preparation patch for stats code duplication removal. 1) use atomic64_t for `pages_zero' and `pages_stored' zram stats. 2) `compr_size' and `pages_zero' struct zram_stats members did not follow the existing device attr naming scheme: zram_stats.ATTR has ATTR_show() function. rename them: -- compr_size -> compr_data_size -- pages_zero -> zero_pages Minchan Kim's note: If we really have trouble with atomic stat operation, we could change it with percpu_counter so that it could solve atomic overhead and unnecessary memory space by introducing unsigned long instead of 64bit atomic_t. Signed-off-by:
Sergey Senozhatsky <sergey.senozhatsky@gmail.com> Acked-by:
Minchan Kim <minchan@kernel.org> Acked-by:
Jerome Marchand <jmarchan@redhat.com> Cc: Nitin Gupta <ngupta@vflare.org> Signed-off-by:
Andrew Morton <akpm@linux-foundation.org> Signed-off-by:
Linus Torvalds <torvalds@linux-foundation.org>
-
Sergey Senozhatsky authored
Remove `good' and `bad' compressed sub-requests stats. RW request may cause a number of RW sub-requests. zram used to account `good' compressed sub-queries (with compressed size less than 50% of original size), `bad' compressed sub-queries (with compressed size greater that 75% of original size), leaving sub-requests with compression size between 50% and 75% of original size not accounted and not reported. zram already accounts each sub-request's compression size so we can calculate real device compression ratio. Signed-off-by:
-
Sergey Senozhatsky authored
Do not pass rw argument down the __zram_make_request() -> zram_bvec_rw() chain, decode it in zram_bvec_rw() instead. Besides, this is the place where we distinguish READ and WRITE bio data directions, so account zram RW stats here, instead of __zram_make_request(). This also allows to account a real number of zram READ/WRITE operations, not just requests (single RW request may cause a number of zram RW ops with separate locking, compression/decompression, etc). Signed-off-by:
-
Sergey Senozhatsky authored
Introduce init_done() helper function which allows us to drop `init_done' struct zram member. init_done() uses the fact that ->init_done == 1 equals to ->meta != NULL. Signed-off-by:
-
John Hubbard authored
A new dump_page() routine was recently added, and marked EXPORT_SYMBOL_GPL. dump_page() was also added to the VM_BUG_ON_PAGE() macro, and so the end result is that non-GPL code can no longer call get_page() and a few other routines. This only happens if the kernel was compiled with CONFIG_DEBUG_VM. Change dump_page() to be EXPORT_SYMBOL. Longer explanation: Prior to commit 309381fe ("mm: dump page when hitting a VM_BUG_ON using VM_BUG_ON_PAGE") , it was possible to build MIT-licensed (non-GPL) drivers on Fedora. Fedora is semi-unique, in that it sets CONFIG_VM_DEBUG. Because Fedora sets CONFIG_VM_DEBUG, they end up pulling in dump_page(), via VM_BUG_ON_PAGE, via get_page(). As one of the authors of NVIDIA's new, open source, "UVM-Lite" kernel module, I originally choose to use the kernel's get_page() routine from within nvidia_uvm_page_cache.c, because get_page() has always seemed to be very clearly intended for use by non-GPL, driver code. So I'm hoping that making get_page() widely accessible again will not be too controversial. We did check with Fedora first, and they responded (https://bugzilla.redhat.com/show_bug.cgi?id=1074710#c3) that we should try to get upstream changed, before asking Fedora to change. Their reasoning seems beneficial to Linux: leaving CONFIG_DEBUG_VM set allows Fedora to help catch mm bugs. Signed-off-by:
John Hubbard <jhubbard@nvidia.com> Cc: Sasha Levin <sasha.levin@oracle.com> Cc: Josh Boyer <jwboyer@redhat.com> Signed-off-by:
-
Srikar Dronamraju authored
LAST_CPUPID_MASK is calculated using LAST_CPUPID_WIDTH. However LAST_CPUPID_WIDTH itself can be 0. (when LAST_CPUPID_NOT_IN_PAGE_FLAGS is set). In such a case LAST_CPUPID_MASK turns out to be 0. But with recent commit 1ae71d03: (mm: numa: bugfix for LAST_CPUPID_NOT_IN_PAGE_FLAGS) if LAST_CPUPID_MASK is 0, page_cpupid_xchg_last() and page_cpupid_reset_last() causes page->_last_cpupid to be set to 0. This causes performance regression. Its almost as if numa_balancing is off. Fix LAST_CPUPID_MASK by using LAST_CPUPID_SHIFT instead of LAST_CPUPID_WIDTH. Some performance numbers and perf stats with and without the fix. (3.14-rc6) ---------- numa01 Performance counter stats for '/usr/bin/time -f %e %S %U %c %w -o start_bench.out -a ./numa01': 12,27,462 cs [100.00%] 2,41,957 migrations [100.00%] 1,68,01,713 faults [100.00%] 7,99,35,29,041 cache-misses 98,808 migrate:mm_migrate_pages [100.00%] 1407.690148814 seconds time elapsed numa02 Performance counter stats for '/usr/bin/time -f %e %S %U %c %w -o start_bench.out -a ./numa02': 63,065 cs [100.00%] 14,364 migrations [100.00%] 2,08,118 faults [100.00%] 25,32,59,404 cache-misses 12 migrate:mm_migrate_pages [100.00%] 63.840827219 seconds time elapsed (3.14-rc6 with fix) ------------------- numa01 Performance counter stats for '/usr/bin/time -f %e %S %U %c %w -o start_bench.out -a ./numa01': 9,68,911 cs [100.00%] 1,01,414 migrations [100.00%] 88,38,697 faults [100.00%] 4,42,92,51,042 cache-misses 4,25,060 migrate:mm_migrate_pages [100.00%] 685.965331189 seconds time elapsed numa02 Performance counter stats for '/usr/bin/time -f %e %S %U %c %w -o start_bench.out -a ./numa02': 17,543 cs [100.00%] 2,962 migrations [100.00%] 1,17,843 faults [100.00%] 11,80,61,644 cache-misses 12,358 migrate:mm_migrate_pages [100.00%] 20.380132343 seconds time elapsed Signed-off-by:
Srikar Dronamraju <srikar@linux.vnet.ibm.com> Cc: Liu Ping Fan <pingfank@linux.vnet.ibm.com> Reviewed-by:
Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: Mel Gorman <mel@csn.ul.ie> Signed-off-by:
-
Zhang Yanfei authored
s/MADV_NODUMP/MADV_DONTDUMP/ Signed-off-by:
Zhang Yanfei <zhangyanfei@cn.fujitsu.com> Signed-off-by:
-
Fabian Frederick authored
Commit f9acc8c7 ("readahead: sanify file_ra_state names") left ra_submit with a single function call. Move ra_submit to internal.h and inline it to save some stack. Thanks to Andrew Morton for commenting different versions. Signed-off-by:
Fabian Frederick <fabf@skynet.be> Suggested-by:
-
Mizuma, Masayoshi authored
When I decrease the value of nr_hugepage in procfs a lot, softlockup happens. It is because there is no chance of context switch during this process. On the other hand, when I allocate a large number of hugepages, there is some chance of context switch. Hence softlockup doesn't happen during this process. So it's necessary to add the context switch in the freeing process as same as allocating process to avoid softlockup. When I freed 12 TB hugapages with kernel-2.6.32-358.el6, the freeing process occupied a CPU over 150 seconds and following softlockup message appeared twice or more. $ echo 6000000 > /proc/sys/vm/nr_hugepages $ cat /proc/sys/vm/nr_hugepages 6000000 $ grep ^Huge /proc/meminfo HugePages_Total: 6000000 HugePages_Free: 6000000 HugePages_Rsvd: 0 HugePages_Surp: 0 Hugepagesize: 2048 kB $ echo 0 > /proc/sys/vm/nr_hugepages BUG: soft lockup - CPU#16 stuck for 67s! [sh:12883] ... Pid: 12883, comm: sh Not tainted 2.6.32-358.el6.x86_64 #1 Call Trace: free_pool_huge_page+0xb8/0xd0 set_max_huge_pages+0x128/0x190 hugetlb_sysctl_handler_common+0x113/0x140 hugetlb_sysctl_handler+0x1e/0x20 proc_sys_call_handler+0x97/0xd0 proc_sys_write+0x14/0x20 vfs_write+0xb8/0x1a0 sys_write+0x51/0x90 __audit_syscall_exit+0x265/0x290 system_call_fastpath+0x16/0x1b I have not confirmed this problem with upstream kernels because I am not able to prepare the machine equipped with 12TB memory now. However I confirmed that the amount of decreasing hugepages was directly proportional to the amount of required time. I measured required times on a smaller machine. It showed 130-145 hugepages decreased in a millisecond. Amount of decreasing Required time Decreasing rate hugepages (msec) (pages/msec) ------------------------------------------------------------ 10,000 pages == 20GB 70 - 74 135-142 30,000 pages == 60GB 208 - 229 131-144 It means decrement of 6TB hugepages will trigger softlockup with the default threshold 20sec, in this decreasing rate. Signed-off-by:
Masayoshi Mizuma <m.mizuma@jp.fujitsu.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Michal Hocko <mhocko@suse.cz> Cc: Wanpeng Li <liwanp@linux.vnet.ibm.com> Cc: Aneesh Kumar <aneesh.kumar@linux.vnet.ibm.com> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com> Cc: <stable@vger.kernel.org> Signed-off-by:
-
Fabian Frederick authored
Replace ((phys_addr_t)(x) << PAGE_SHIFT) by pfn macro. Signed-off-by:
-
Emil Medve authored
This is a small cleanup. Signed-off-by:
Emil Medve <Emilian.Medve@Freescale.com> Signed-off-by:
-
Miklos Szeredi authored
There's only one caller of set_page_dirty_balance() and that will call it with page_mkwrite == 0. The page_mkwrite argument was unused since commit b827e496 "mm: close page_mkwrite races". Signed-off-by:
Miklos Szeredi <mszeredi@suse.cz> Signed-off-by:
-
Vlastimil Babka authored
A BUG_ON(!PageLocked) was triggered in mlock_vma_page() by Sasha Levin fuzzing with trinity. The call site try_to_unmap_cluster() does not lock the pages other than its check_page parameter (which is already locked). The BUG_ON in mlock_vma_page() is not documented and its purpose is somewhat unclear, but apparently it serializes against page migration, which could otherwise fail to transfer the PG_mlocked flag. This would not be fatal, as the page would be eventually encountered again, but NR_MLOCK accounting would become distorted nevertheless. This patch adds a comment to the BUG_ON in mlock_vma_page() and munlock_vma_page() to that effect. The call site try_to_unmap_cluster() is fixed so that for page != check_page, trylock_page() is attempted (to avoid possible deadlocks as we already have check_page locked) and mlock_vma_page() is performed only upon success. If the page lock cannot be obtained, the page is left without PG_mlocked, which is again not a problem in the whole unevictable memory design. Signed-off-by:
Vlastimil Babka <vbabka@suse.cz> Signed-off-by:
Bob Liu <bob.liu@oracle.com> Reported-by:
Sasha Levin <sasha.levin@oracle.com> Cc: Wanpeng Li <liwanp@linux.vnet.ibm.com> Cc: Michel Lespinasse <walken@google.com> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Acked-by:
Rik van Riel <riel@redhat.com> Cc: David Rientjes <rientjes@google.com> Cc: Mel Gorman <mgorman@suse.de> Cc: Hugh Dickins <hughd@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: <stable@vger.kernel.org> Signed-off-by:
-
Johannes Weiner authored
On NUMA systems, a node may start thrashing cache or even swap anonymous pages while there are still free pages on remote nodes. This is a result of commits 81c0a2bb ("mm: page_alloc: fair zone allocator policy") and fff4068c ("mm: page_alloc: revert NUMA aspect of fair allocation policy"). Before those changes, the allocator would first try all allowed zones, including those on remote nodes, before waking any kswapds. But now, the allocator fastpath doubles as the fairness pass, which in turn can only consider the local node to prevent remote spilling based on exhausted fairness batches alone. Remote nodes are only considered in the slowpath, after the kswapds are woken up. But if remote nodes still have free memory, kswapd should not be woken to rebalance the local node or it may thrash cash or swap prematurely. Fix this by adding one more unfair pass over the zonelist that is allowed to spill to remote nodes after the local fairness pass fails but before entering the slowpath and waking the kswapds. This also gets rid of the GFP_THISNODE exemption from the fairness protocol because the unfair pass is no longer tied to kswapd, which GFP_THISNODE is not allowed to wake up. However, because remote spills can be more frequent now - we prefer them over local kswapd reclaim - the allocation batches on remote nodes could underflow more heavily. When resetting the batches, use atomic_long_read() directly instead of zone_page_state() to calculate the delta as the latter filters negative counter values. Signed-off-by:
Johannes Weiner <hannes@cmpxchg.org> Acked-by:
Mel Gorman <mgorman@suse.de> Cc: <stable@kernel.org> [3.12+] Signed-off-by:
-
Michal Hocko authored
mem_cgroup_newpage_charge is used only for charging anonymous memory so it is better to rename it to mem_cgroup_charge_anon. mem_cgroup_cache_charge is used for file backed memory so rename it to mem_cgroup_charge_file. Signed-off-by:
Michal Hocko <mhocko@suse.cz> Cc: Johannes Weiner <hannes@cmpxchg.org> Signed-off-by:
-
Johannes Weiner authored
Some callsites pass a memcg directly, some callsites pass an mm that then has to be translated to a memcg. This makes for a terrible function interface. Just push the mm-to-memcg translation into the respective callsites and always pass a memcg to mem_cgroup_try_charge(). [mhocko@suse.cz: add charge mm helper] Signed-off-by:
-
Michal Hocko authored
__mem_cgroup_try_charge duplicates get_mem_cgroup_from_mm for charges which came without a memcg. The only reason seems to be a tiny optimization when css_tryget is not called if the charge can be consumed from the stock. Nevertheless css_tryget is very cheap since it has been reworked to use per-cpu counting so this optimization doesn't give us anything these days. So let's drop the code duplication so that the code is more readable. Signed-off-by:
-
Johannes Weiner authored
Instead of returning NULL from try_get_mem_cgroup_from_mm() when the mm owner is exiting, just return root_mem_cgroup. This makes sense for all callsites and gets rid of some of them having to fallback manually. [fengguang.wu@intel.com: fix warnings] Signed-off-by:
Fengguang Wu <fengguang.wu@intel.com> Acked-by:
-
Johannes Weiner authored
Users pass either a mm that has been established under task lock, or use a verified current->mm, which means the task can't be exiting. Signed-off-by:
-
Johannes Weiner authored
Only page cache charges can happen without an mm context, so push this special case out of the inner core and into the cache charge function. An ancient comment explains that the mm can also be NULL in case the task is currently being migrated, but that is not actually true with the current case, so just remove it. Signed-off-by:
-
Johannes Weiner authored
mem_cgroup_charge_common() is used by both cache and anon pages, but most of its body only applies to anon pages and the remainder is not worth having in a separate function. Signed-off-by:
-
Johannes Weiner authored
It used to disable preemption and run sanity checks but now it's only taking a number out of one percpu counter and putting it into another. Do this directly in the callsite and save the indirection. Signed-off-by:
-
Johannes Weiner authored
lock_page_cgroup() disables preemption, remove explicit preemption disabling for code paths holding this lock. Signed-off-by:
-
Kirill A. Shutemov authored
I tried to use 'dump_page(page, __func__)' for debugging, but it triggers warning: warning: passing argument 2 of `dump_page' discards `const' qualifier from pointer target type [enabled by default] Let's convert 'reason' to 'const char *' in dump_page() and friends: we shouldn't modify it anyway. Signed-off-by:
Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Signed-off-by:
-
Gioh Kim authored
vm_map_ram() has a fragmentation problem when it cannot purge a chunk(ie, 4M address space) if there is a pinning object in that addresss space. So it could consume all VMALLOC address space easily. We can fix the fragmentation problem by using vmap instead of vm_map_ram() but vmap() is known to be slow compared to vm_map_ram(). Minchan said vm_map_ram is 5 times faster than vmap in his tests. So I thought we should fix fragment problem of vm_map_ram because our proprietary GPU driver has used it heavily. On second thought, it's not an easy because we should reuse freed space for solving the problem and it could make more IPI and bitmap operation for searching hole. It could mitigate API's goal which is very fast mapping. And even fragmentation problem wouldn't show in 64 bit machine. Another option is that the user should separate long-life and short-life object and use vmap for long-life but vm_map_ram for short-life. If we inform the user about the characteristic of vm_map_ram the user can choose one according to the page lifetime. Let's add some notice messages to user. [akpm@linux-foundation.org: tweak comment text] Signed-off-by:
Gioh Kim <gioh.kim@lge.com> Reviewed-by:
-
Choi Gi-yong authored
Signed-off-by:
Choi Gi-yong <yong@gnoy.org> Signed-off-by:
-
Mikulas Patocka authored
Add unlikely and likely hints to the function mempool_free. It lays out the code in such a way that the common path is executed straighforward and saves a cache line. Signed-off-by:
Mikulas Patocka <mpatocka@redhat.com> Cc: Ingo Molnar <mingo@elte.hu> Signed-off-by:
-
David Rientjes authored
The conditions that control the isolation mode in isolate_migratepages_range() do not change during the iteration, so extract them out and only define the value once. This actually does have an effect, gcc doesn't optimize it itself because of cc->sync. Signed-off-by:
David Rientjes <rientjes@google.com> Cc: Mel Gorman <mgorman@suse.de> Acked-by:
-
David Rientjes authored
The res_counter_{charge,uncharge}_locked() variants are not used in the kernel outside of the resource counter code itself, so remove the interface. Signed-off-by: -
David Rientjes authored
PF_MEMPOLICY is an unnecessary optimization for CONFIG_SLAB users. There's no significant performance degradation to checking current->mempolicy rather than current->flags & PF_MEMPOLICY in the allocation path, especially since this is considered unlikely(). Running TCP_RR with netperf-2.4.5 through localhost on 16 cpu machine with 64GB of memory and without a mempolicy: threads before after 16 1249409 1244487 32 1281786 1246783 48 1239175 1239138 64 1244642 1241841 80 1244346 1248918 96 1266436 1254316 112 1307398 1312135 128 1327607 1326502 Per-process flags are a scarce resource so we should free them up whenever possible and make them available. We'll be using it shortly for memcg oom reserves. Signed-off-by:
-
David Rientjes authored
slab_node() is actually a mempolicy function, so rename it to mempolicy_slab_node() to make it clearer that it used for processes with mempolicies. At the same time, cleanup its code by saving numa_mem_id() in a local variable (since we require a node with memory, not just any node) and remove an obsolete comment that assumes the mempolicy is actually passed into the function. Signed-off-by:
Christoph Lameter <cl@linux.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Rik van Riel <riel@redhat.com> Cc: Jianguo Wu <wujianguo@huawei.com> Cc: Tim Hockin <thockin@google.com> Signed-off-by:
-
David Rientjes authored
copy_flags() does not use the clone_flags formal and can be collapsed into copy_process() for cleaner code. Signed-off-by:
-
Gideon Israel Dsouza authored
To increase compiler portability there is <linux/compiler.h> which provides convenience macros for various gcc constructs. Eg: __weak for __attribute__((weak)). I've replaced all instances of gcc attributes with the right macro in the memory management (/mm) subsystem. [akpm@linux-foundation.org: while-we're-there consistency tweaks] Signed-off-by:
Gideon Israel Dsouza <gidisrael@gmail.com> Signed-off-by:
-
Davidlohr Bueso authored
This patch is a continuation of efforts trying to optimize find_vma(), avoiding potentially expensive rbtree walks to locate a vma upon faults. The original approach (https://lkml.org/lkml/2013/11/1/410), where the largest vma was also cached, ended up being too specific and random, thus further comparison with other approaches were needed. There are two things to consider when dealing with this, the cache hit rate and the latency of find_vma(). Improving the hit-rate does not necessarily translate in finding the vma any faster, as the overhead of any fancy caching schemes can be too high to consider. We currently cache the last used vma for the whole address space, which provides a nice optimization, reducing the total cycles in find_vma() by up to 250%, for workloads with good locality. On the other hand, this simple scheme is pretty much useless for workloads with poor locality. Analyzing ebizzy runs shows that, no matter how many threads are running, the mmap_cache hit rate is less than 2%, and in many situations below 1%. The proposed approach is to replace this scheme with a small per-thread cache, maximizing hit rates at a very low maintenance cost. Invalidations are performed by simply bumping up a 32-bit sequence number. The only expensive operation is in the rare case of a seq number overflow, where all caches that share the same address space are flushed. Upon a miss, the proposed replacement policy is based on the page number that contains the virtual address in question. Concretely, the following results are seen on an 80 core, 8 socket x86-64 box: 1) System bootup: Most programs are single threaded, so the per-thread scheme does improve ~50% hit rate by just adding a few more slots to the cache. +----------------+----------+------------------+ | caching scheme | hit-rate | cycles (billion) | +----------------+----------+------------------+ | baseline | 50.61% | 19.90 | | patched | 73.45% | 13.58 | +----------------+----------+------------------+ 2) Kernel build: This one is already pretty good with the current approach as we're dealing with good locality. +----------------+----------+------------------+ | caching scheme | hit-rate | cycles (billion) | +----------------+----------+------------------+ | baseline | 75.28% | 11.03 | | patched | 88.09% | 9.31 | +----------------+----------+------------------+ 3) Oracle 11g Data Mining (4k pages): Similar to the kernel build workload. +----------------+----------+------------------+ | caching scheme | hit-rate | cycles (billion) | +----------------+----------+------------------+ | baseline | 70.66% | 17.14 | | patched | 91.15% | 12.57 | +----------------+----------+------------------+ 4) Ebizzy: There's a fair amount of variation from run to run, but this approach always shows nearly perfect hit rates, while baseline is just about non-existent. The amounts of cycles can fluctuate between anywhere from ~60 to ~116 for the baseline scheme, but this approach reduces it considerably. For instance, with 80 threads: +----------------+----------+------------------+ | caching scheme | hit-rate | cycles (billion) | +----------------+----------+------------------+ | baseline | 1.06% | 91.54 | | patched | 99.97% | 14.18 | +----------------+----------+------------------+ [akpm@linux-foundation.org: fix nommu build, per Davidlohr] [akpm@linux-foundation.org: document vmacache_valid() logic] [akpm@linux-foundation.org: attempt to untangle header files] [akpm@linux-foundation.org: add vmacache_find() BUG_ON] [hughd@google.com: add vmacache_valid_mm() (from Oleg)] [akpm@linux-foundation.org: coding-style fixes] [akpm@linux-foundation.org: adjust and enhance comments] Signed-off-by:
Davidlohr Bueso <davidlohr@hp.com> Reviewed-by:
Michel Lespinasse <walken@google.com> Cc: Oleg Nesterov <oleg@redhat.com> Tested-by:
Hugh Dickins <hughd@google.com> Signed-off-by:
-
Ning Qu authored
In shmem/tmpfs, we also use the generic filemap_map_pages, seems the additional checking is not worth a separate version of map_pages for it. Signed-off-by:
Ning Qu <quning@google.com> Acked-by:
-
Kirill A. Shutemov authored
Let's allow people to tweak faultaround at runtime. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by:
-
Kirill A. Shutemov authored
Minor cleanups: - 'size' variable is now in bytes, not pages; - use round_up(): it should be easier to read. Signed-off-by:
-
Kirill A. Shutemov authored
filemap_map_pages() is generic implementation of ->map_pages() for filesystems who uses page cache. It should be safe to use filemap_map_pages() for ->map_pages() if filesystem use filemap_fault() for ->fault(). Signed-off-by:
-
Kirill A. Shutemov authored
Here's new version of faultaround patchset. It took a while to tune it and collect performance data. First patch adds new callback ->map_pages to vm_operations_struct. ->map_pages() is called when VM asks to map easy accessible pages. Filesystem should find and map pages associated with offsets from "pgoff" till "max_pgoff". ->map_pages() is called with page table locked and must not block. If it's not possible to reach a page without blocking, filesystem should skip it. Filesystem should use do_set_pte() to setup page table entry. Pointer to entry associated with offset "pgoff" is passed in "pte" field in vm_fault structure. Pointers to entries for other offsets should be calculated relative to "pte". Currently VM use ->map_pages only on read page fault path. We try to map FAULT_AROUND_PAGES a time. FAULT_AROUND_PAGES is 16 for now. Performance data for different FAULT_AROUND_ORDER is below. TODO: - implement ->map_pages() for shmem/tmpfs; - modify get_user_pages() to be able to use ->map_pages() and implement mmap(MAP_POPULATE|MAP_NONBLOCK) on top. ========================================================================= Tested on 4-socket machine (120 threads) with 128GiB of RAM. Few real-world workloads. The sweet spot for FAULT_AROUND_ORDER here is somewhere between 3 and 5. Let's say 4 :) Linux build (make -j60) FAULT_AROUND_ORDER Baseline 1 3 4 5 7 9 minor-faults 283,301,572 247,151,987 212,215,789 204,772,882 199,568,944 194,703,779 193,381,485 time, seconds 151.227629483 153.920996480 151.356125472 150.863792049 150.879207877 151.150764954 151.450962358 Linux rebuild (make -j60) FAULT_AROUND_ORDER Baseline 1 3 4 5 7 9 minor-faults 5,396,854 4,148,444 2,855,286 2,577,282 2,361,957 2,169,573 2,112,643 time, seconds 27.404543757 27.559725591 27.030057426 26.855045126 26.678618635 26.974523490 26.761320095 Git test suite (make -j60 test) FAULT_AROUND_ORDER Baseline 1 3 4 5 7 9 minor-faults 129,591,823 99,200,751 66,106,718 57,606,410 51,510,808 45,776,813 44,085,515 time, seconds 66.087215026 64.784546905 64.401156567 65.282708668 66.034016829 66.793780811 67.237810413 Two synthetic tests: access every word in file in sequential/random order. It doesn't improve much after FAULT_AROUND_ORDER == 4. Sequential access 16GiB file FAULT_AROUND_ORDER Baseline 1 3 4 5 7 9 1 thread minor-faults 4,195,437 2,098,275 525,068 262,251 131,170 32,856 8,282 time, seconds 7.250461742 6.461711074 5.493859139 5.488488147 5.707213983 5.898510832 5.109232856 8 threads minor-faults 33,557,540 16,892,728 4,515,848 2,366,999 1,423,382 442,732 142,339 time, seconds 16.649304881 9.312555263 6.612490639 6.394316732 6.669827501 6.75078944 6.371900528 32 threads minor-faults 134,228,222 67,526,810 17,725,386 9,716,537 4,763,731 1,668,921 537,200 time, seconds 49.164430543 29.712060103 12.938649729 10.175151004 11.840094583 9.594081325 9.928461797 60 threads minor-faults 251,687,988 126,146,952 32,919,406 18,208,804 10,458,947 2,733,907 928,217 time, seconds 86.260656897 49.626551828 22.335007632 17.608243696 16.523119035 16.339489186 16.326390902 120 threads minor-faults 503,352,863 252,939,677 67,039,168 35,191,827 19,170,091 4,688,357 1,471,862 time, seconds 124.589206333 79.757867787 39.508707872 32.167281632 29.972989292 28.729834575 28.042251622 Random access 1GiB file 1 thread minor-faults 262,636 132,743 34,369 17,299 8,527 3,451 1,222 time, seconds 15.351890914 16.613802482 16.569227308 15.179220992 16.557356122 16.578247824 15.365266994 8 threads minor-faults 2,098,948 1,061,871 273,690 154,501 87,110 25,663 7,384 time, seconds 15.040026343 15.096933500 14.474757288 14.289129964 14.411537468 14.296316837 14.395635804 32 threads minor-faults 8,390,734 4,231,023 1,054,432 528,847 269,242 97,746 26,881 time, seconds 20.430433109 21.585235358 22.115062928 14.872878951 14.880856305 14.883370649 14.821261690 60 threads minor-faults 15,733,258 7,892,809 1,973,393 988,266 594,789 164,994 51,691 time, seconds 26.577302548 25.692397770 18.728863715 20.153026398 21.619101933 17.745086260 17.613215273 120 threads minor-faults 31,471,111 15,816,616 3,959,209 1,978,685 1,008,299 264,635 96,010 time, seconds 41.835322703 40.459786095 36.085306105 35.313894834 35.814445675 36.552633793 34.289210594 Touch only one page in page table in 16GiB file FAULT_AROUND_ORDER Baseline 1 3 4 5 7 9 1 thread minor-faults 8,372 8,324 8,270 8,260 8,249 8,239 8,237 time, seconds 0.039892712 0.045369149 0.051846126 0.063681685 0.079095975 0.17652406 0.541213386 8 threads minor-faults 65,731 65,681 65,628 65,620 65,608 65,599 65,596 time, seconds 0.124159196 0.488600638 0.156854426 0.191901957 0.242631486 0.543569456 1.677303984 32 threads minor-faults 262,388 262,341 262,285 262,276 262,266 262,257 263,183 time, seconds 0.452421421 0.488600638 0.565020946 0.648229739 0.789850823 1.651584361 5.000361559 60 threads minor-faults 491,822 491,792 491,723 491,711 491,701 491,691 491,825 time, seconds 0.763288616 0.869620515 0.980727360 1.161732354 1.466915814 3.04041448 9.308612938 120 threads minor-faults 983,466 983,655 983,366 983,372 983,363 984,083 984,164 time, seconds 1.595846553 1.667902182 2.008959376 2.425380942 2.941368804 5.977807890 18.401846125 This patch (of 2): Introduce new vm_ops callback ->map_pages() and uses it for mapping easy accessible pages around fault address. On read page fault, if filesystem provides ->map_pages(), we try to map up to FAULT_AROUND_PAGES pages around page fault address in hope to reduce number of minor page faults. We call ->map_pages first and use ->fault() as fallback if page by the offset is not ready to be mapped (cold page cache or something). Signed-off-by:
-
Andrew Morton authored
"mm: introduce vm_ops->map_pages()" wants to export a do_set_pte() from core kernel. Rename lguest's do_set_pte() to something more lguest-specific. Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com> Cc: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by:
-
