alex chen
2017-Dec-28 02:02 UTC
[Ocfs2-devel] [PATCH] ocfs2: try a blocking lock before return AOP_TRUNCATED_PAGE
Hi Gang, On 2017/12/27 18:37, Gang He wrote:> Hi Jun, > > >>>> >> Hi Gang, >> >> Do you mean that too many retrys in loop cast losts of CPU-time and >> block page-fault interrupt? We should not add any delay in >> ocfs2_fault(), right? And I still feel a little confused why your >> method can solve this problem. > You can see the related code in function filemap_fault(), if ocfs2 fails to read a page since > it can not get a inode lock with non-block mode, the VFS layer code will invoke ocfs2 > read page call back function circularly, this will lead to a softlockup problem (like the below back trace). > So, we should get a blocking lock to let the dlm lock to this node and also can avoid CPU loop,Can we use 'cond_resched()' to allow the thread to release the CPU temperately for solving this softlockup?> second, base on my testing, the patch also can improve the efficiency in case modifying the same > file frequently from multiple nodes, since the lock acquisition chance is more fair. > In fact, the code was modified by a patch 1cce4df04f37 ("ocfs2: do not lock/unlock() inode DLM lock"), > before that patch, the code is the same, this patch can be considered to revert that patch, except adding more > clear comments.In patch 1cce4df04f37 ("ocfs2: do not lock/unlock() inode DLM lock"), Goldwyn says blocking lock and unlock will only make the performance worse where contention over the locks is high, which is the opposite of your described above. IMO, blocking lock and unlock here is indeed unnecessary. Thanks, Alex> > Thanks > Gang > > >> >> thanks, >> Jun >> >> On 2017/12/27 17:29, Gang He wrote: >>> If we can't get inode lock immediately in the function >>> ocfs2_inode_lock_with_page() when reading a page, we should not >>> return directly here, since this will lead to a softlockup problem. >>> The method is to get a blocking lock and immediately unlock before >>> returning, this can avoid CPU resource waste due to lots of retries, >>> and benefits fairness in getting lock among multiple nodes, increase >>> efficiency in case modifying the same file frequently from multiple >>> nodes. >>> The softlockup problem looks like, >>> Kernel panic - not syncing: softlockup: hung tasks >>> CPU: 0 PID: 885 Comm: multi_mmap Tainted: G L 4.12.14-6.1-default #1 >>> Hardware name: Bochs Bochs, BIOS Bochs 01/01/2011 >>> Call Trace: >>> <IRQ> >>> dump_stack+0x5c/0x82 >>> panic+0xd5/0x21e >>> watchdog_timer_fn+0x208/0x210 >>> ? watchdog_park_threads+0x70/0x70 >>> __hrtimer_run_queues+0xcc/0x200 >>> hrtimer_interrupt+0xa6/0x1f0 >>> smp_apic_timer_interrupt+0x34/0x50 >>> apic_timer_interrupt+0x96/0xa0 >>> </IRQ> >>> RIP: 0010:unlock_page+0x17/0x30 >>> RSP: 0000:ffffaf154080bc88 EFLAGS: 00000246 ORIG_RAX: ffffffffffffff10 >>> RAX: dead000000000100 RBX: fffff21e009f5300 RCX: 0000000000000004 >>> RDX: dead0000000000ff RSI: 0000000000000202 RDI: fffff21e009f5300 >>> RBP: 0000000000000000 R08: 0000000000000000 R09: ffffaf154080bb00 >>> R10: ffffaf154080bc30 R11: 0000000000000040 R12: ffff993749a39518 >>> R13: 0000000000000000 R14: fffff21e009f5300 R15: fffff21e009f5300 >>> ocfs2_inode_lock_with_page+0x25/0x30 [ocfs2] >>> ocfs2_readpage+0x41/0x2d0 [ocfs2] >>> ? pagecache_get_page+0x30/0x200 >>> filemap_fault+0x12b/0x5c0 >>> ? recalc_sigpending+0x17/0x50 >>> ? __set_task_blocked+0x28/0x70 >>> ? __set_current_blocked+0x3d/0x60 >>> ocfs2_fault+0x29/0xb0 [ocfs2] >>> __do_fault+0x1a/0xa0 >>> __handle_mm_fault+0xbe8/0x1090 >>> handle_mm_fault+0xaa/0x1f0 >>> __do_page_fault+0x235/0x4b0 >>> trace_do_page_fault+0x3c/0x110 >>> async_page_fault+0x28/0x30 >>> RIP: 0033:0x7fa75ded638e >>> RSP: 002b:00007ffd6657db18 EFLAGS: 00010287 >>> RAX: 000055c7662fb700 RBX: 0000000000000001 RCX: 000055c7662fb700 >>> RDX: 0000000000001770 RSI: 00007fa75e909000 RDI: 000055c7662fb700 >>> RBP: 0000000000000003 R08: 000000000000000e R09: 0000000000000000 >>> R10: 0000000000000483 R11: 00007fa75ded61b0 R12: 00007fa75e90a770 >>> R13: 000000000000000e R14: 0000000000001770 R15: 0000000000000000 >>> >>> Fixes: 1cce4df04f37 ("ocfs2: do not lock/unlock() inode DLM lock") >>> Signed-off-by: Gang He <ghe at suse.com> >>> --- >>> fs/ocfs2/dlmglue.c | 9 +++++++++ >>> 1 file changed, 9 insertions(+) >>> >>> diff --git a/fs/ocfs2/dlmglue.c b/fs/ocfs2/dlmglue.c >>> index 4689940..5193218 100644 >>> --- a/fs/ocfs2/dlmglue.c >>> +++ b/fs/ocfs2/dlmglue.c >>> @@ -2486,6 +2486,15 @@ int ocfs2_inode_lock_with_page(struct inode *inode, >>> ret = ocfs2_inode_lock_full(inode, ret_bh, ex, OCFS2_LOCK_NONBLOCK); >>> if (ret == -EAGAIN) { >>> unlock_page(page); >>> + /* >>> + * If we can't get inode lock immediately, we should not return >>> + * directly here, since this will lead to a softlockup problem. >>> + * The method is to get a blocking lock and immediately unlock >>> + * before returning, this can avoid CPU resource waste due to >>> + * lots of retries, and benefits fairness in getting lock. >>> + */ >>> + if (ocfs2_inode_lock(inode, ret_bh, ex) == 0) >>> + ocfs2_inode_unlock(inode, ex); >>> ret = AOP_TRUNCATED_PAGE; >>> } >>> >>> > > _______________________________________________ > Ocfs2-devel mailing list > Ocfs2-devel at oss.oracle.com > https://oss.oracle.com/mailman/listinfo/ocfs2-devel > > . >
Gang He
2017-Dec-28 02:48 UTC
[Ocfs2-devel] [PATCH] ocfs2: try a blocking lock before return AOP_TRUNCATED_PAGE
Hi Alex,>>> > Hi Gang, > > On 2017/12/27 18:37, Gang He wrote: >> Hi Jun, >> >> >>>>> >>> Hi Gang, >>> >>> Do you mean that too many retrys in loop cast losts of CPU-time and >>> block page-fault interrupt? We should not add any delay in >>> ocfs2_fault(), right? And I still feel a little confused why your >>> method can solve this problem. >> You can see the related code in function filemap_fault(), if ocfs2 fails to > read a page since >> it can not get a inode lock with non-block mode, the VFS layer code will > invoke ocfs2 >> read page call back function circularly, this will lead to a softlockup > problem (like the below back trace). >> So, we should get a blocking lock to let the dlm lock to this node and also > can avoid CPU loop, > Can we use 'cond_resched()' to allow the thread to release the CPU > temperately for solving this softlockup?Yes, we can use cond_resched() function to avoid this softlockup. In fact, if the kernel is configured with CONFIG_PREEMPT=y, this softlockup does not happen since the kernel can help. But, this way still leads to CPU resource waste, CPU usage can reach about 80% - 100% when multiple nodes read/write/mmap-access the same file concurrently, and more, the read/write/mmap-access speed is more lower (50% decrease). Why? Because we need to get DLM lock for each node, before one node gets DLM lock, another node has to down-convert this DLM lock, that means flushing the memory data to the disk before DLM lock down-conversion. this disk IO operation is very slow compared with CPU cycle, that means the node which want to get DLM lock, will do lots of reties before another node complete down-converting this DLM lock, actual, these retries do not make sense, just waste CPU cycle. So, if we add a blocking lock/unlock here, we will avoid these unnecessary reties, especially in case slow-speed disk and more ocfs2 nodes(>=3). I did the ocfs2 test case (multi_mmap in multiple_run.sh), after applied this patch, the CPU rate on each node was about 40%-50%, and the test case execution time reduced by half. the full command is as below, multiple_run.sh -i eth0 -k ~/linux-4.4.21-69.tar.gz -o ~/ocfs2mullog -C hacluster -s pcmk -n nd1,nd2,nd3 -d /dev/sda1 -b 4096 -c 32768 -t multi_mmap /mnt/shared the shared storage is a iscsi disk. Thanks Gang> >> second, base on my testing, the patch also can improve the efficiency in > case modifying the same >> file frequently from multiple nodes, since the lock acquisition chance is > more fair. >> In fact, the code was modified by a patch 1cce4df04f37 ("ocfs2: do not > lock/unlock() inode DLM lock"), >> before that patch, the code is the same, this patch can be considered to > revert that patch, except adding more >> clear comments. > In patch 1cce4df04f37 ("ocfs2: do not lock/unlock() inode DLM lock"), > Goldwyn says blocking lock and unlock will only make > the performance worse where contention over the locks is high, which is the > opposite of your described above. > IMO, blocking lock and unlock here is indeed unnecessary. > > Thanks, > Alex >> >> Thanks >> Gang >> >> >>> >>> thanks, >>> Jun >>> >>> On 2017/12/27 17:29, Gang He wrote: >>>> If we can't get inode lock immediately in the function >>>> ocfs2_inode_lock_with_page() when reading a page, we should not >>>> return directly here, since this will lead to a softlockup problem. >>>> The method is to get a blocking lock and immediately unlock before >>>> returning, this can avoid CPU resource waste due to lots of retries, >>>> and benefits fairness in getting lock among multiple nodes, increase >>>> efficiency in case modifying the same file frequently from multiple >>>> nodes. >>>> The softlockup problem looks like, >>>> Kernel panic - not syncing: softlockup: hung tasks >>>> CPU: 0 PID: 885 Comm: multi_mmap Tainted: G L 4.12.14-6.1-default #1 >>>> Hardware name: Bochs Bochs, BIOS Bochs 01/01/2011 >>>> Call Trace: >>>> <IRQ> >>>> dump_stack+0x5c/0x82 >>>> panic+0xd5/0x21e >>>> watchdog_timer_fn+0x208/0x210 >>>> ? watchdog_park_threads+0x70/0x70 >>>> __hrtimer_run_queues+0xcc/0x200 >>>> hrtimer_interrupt+0xa6/0x1f0 >>>> smp_apic_timer_interrupt+0x34/0x50 >>>> apic_timer_interrupt+0x96/0xa0 >>>> </IRQ> >>>> RIP: 0010:unlock_page+0x17/0x30 >>>> RSP: 0000:ffffaf154080bc88 EFLAGS: 00000246 ORIG_RAX: ffffffffffffff10 >>>> RAX: dead000000000100 RBX: fffff21e009f5300 RCX: 0000000000000004 >>>> RDX: dead0000000000ff RSI: 0000000000000202 RDI: fffff21e009f5300 >>>> RBP: 0000000000000000 R08: 0000000000000000 R09: ffffaf154080bb00 >>>> R10: ffffaf154080bc30 R11: 0000000000000040 R12: ffff993749a39518 >>>> R13: 0000000000000000 R14: fffff21e009f5300 R15: fffff21e009f5300 >>>> ocfs2_inode_lock_with_page+0x25/0x30 [ocfs2] >>>> ocfs2_readpage+0x41/0x2d0 [ocfs2] >>>> ? pagecache_get_page+0x30/0x200 >>>> filemap_fault+0x12b/0x5c0 >>>> ? recalc_sigpending+0x17/0x50 >>>> ? __set_task_blocked+0x28/0x70 >>>> ? __set_current_blocked+0x3d/0x60 >>>> ocfs2_fault+0x29/0xb0 [ocfs2] >>>> __do_fault+0x1a/0xa0 >>>> __handle_mm_fault+0xbe8/0x1090 >>>> handle_mm_fault+0xaa/0x1f0 >>>> __do_page_fault+0x235/0x4b0 >>>> trace_do_page_fault+0x3c/0x110 >>>> async_page_fault+0x28/0x30 >>>> RIP: 0033:0x7fa75ded638e >>>> RSP: 002b:00007ffd6657db18 EFLAGS: 00010287 >>>> RAX: 000055c7662fb700 RBX: 0000000000000001 RCX: 000055c7662fb700 >>>> RDX: 0000000000001770 RSI: 00007fa75e909000 RDI: 000055c7662fb700 >>>> RBP: 0000000000000003 R08: 000000000000000e R09: 0000000000000000 >>>> R10: 0000000000000483 R11: 00007fa75ded61b0 R12: 00007fa75e90a770 >>>> R13: 000000000000000e R14: 0000000000001770 R15: 0000000000000000 >>>> >>>> Fixes: 1cce4df04f37 ("ocfs2: do not lock/unlock() inode DLM lock") >>>> Signed-off-by: Gang He <ghe at suse.com> >>>> --- >>>> fs/ocfs2/dlmglue.c | 9 +++++++++ >>>> 1 file changed, 9 insertions(+) >>>> >>>> diff --git a/fs/ocfs2/dlmglue.c b/fs/ocfs2/dlmglue.c >>>> index 4689940..5193218 100644 >>>> --- a/fs/ocfs2/dlmglue.c >>>> +++ b/fs/ocfs2/dlmglue.c >>>> @@ -2486,6 +2486,15 @@ int ocfs2_inode_lock_with_page(struct inode *inode, >>>> ret = ocfs2_inode_lock_full(inode, ret_bh, ex, OCFS2_LOCK_NONBLOCK); >>>> if (ret == -EAGAIN) { >>>> unlock_page(page); >>>> + /* >>>> + * If we can't get inode lock immediately, we should not return >>>> + * directly here, since this will lead to a softlockup problem. >>>> + * The method is to get a blocking lock and immediately unlock >>>> + * before returning, this can avoid CPU resource waste due to >>>> + * lots of retries, and benefits fairness in getting lock. >>>> + */ >>>> + if (ocfs2_inode_lock(inode, ret_bh, ex) == 0) >>>> + ocfs2_inode_unlock(inode, ex); >>>> ret = AOP_TRUNCATED_PAGE; >>>> } >>>> >>>> >> >> _______________________________________________ >> Ocfs2-devel mailing list >> Ocfs2-devel at oss.oracle.com >> https://oss.oracle.com/mailman/listinfo/ocfs2-devel >> >> . >>