Hi,
This is abnormal test case, however, when this happened it will have big impact on the apps using those files. And this can not be restored automatically unless disable some xlator, I think it is unacceptable for the user apps.
cynthia
From: Kotresh Hiremath Ravishankar <khiremat@xxxxxxxxxx>
Sent: 2020年3月12日 14:37
To: Zhou, Cynthia (NSB - CN/Hangzhou) <cynthia.zhou@xxxxxxxxxxxxxxx>
Cc: Gluster Devel <gluster-devel@xxxxxxxxxxx>
Subject: Re: could you help to check about a glusterfs issue seems to be related to ctime
All the perf xlators depend on time (mostly mtime I guess). In my setup, only quick read was enabled and hence disabling it worked for me.
All perf xlators needs to be disabled to make it work correctly. But I still failed to understand how normal this kind of workload ?
When disable both quick-read and performance.io-cache off everything is back to normal
I attached the log when only enable quick-read and performance.io-cache is still on glusterfs trace log
When execute command “cat /mnt/export/testfile”
Can you help to find why this still to fail to show correct content?
The file size showed is 141, but actually in brick it is longer than that.
cynthia
From: Zhou, Cynthia (NSB - CN/Hangzhou)
Sent: 2020年3月12日 12:53
To: 'Kotresh Hiremath Ravishankar' <khiremat@xxxxxxxxxx>
Cc: 'Gluster Devel' <gluster-devel@xxxxxxxxxxx>
Subject: RE: could you help to check about a glusterfs issue seems to be related to ctime
From my local test only when disable both features.ctime and ctime.noatime this issue is gone.
Or
Do echo 3 >/proc/sys/vm/drop_caches after each time when some client change the file , can cat command show correct data(same as brick )
cynthia
From: Zhou, Cynthia (NSB - CN/Hangzhou)
Sent: 2020年3月12日 9:53
To: 'Kotresh Hiremath Ravishankar' <khiremat@xxxxxxxxxx>
Cc: Gluster Devel <gluster-devel@xxxxxxxxxxx>
Subject: RE: could you help to check about a glusterfs issue seems to be related to ctime
Hi,
Thanks for your responding!
I’ve tried to disable quick-read:
[root@mn-0:/home/robot]
# gluster v get export all| grep quick
performance.quick-read off
performance.nfs.quick-read off
however, this issue still exists.
Two clients see different contents.
it seems only after I disable utime this issue is completely gone.
features.ctime off
ctime.noatime off
Do you know why is this?
Cynthia
Nokia storage team
From: Kotresh Hiremath Ravishankar <khiremat@xxxxxxxxxx>
Sent: 2020年3月11日 22:05
To: Zhou, Cynthia (NSB - CN/Hangzhou) <cynthia.zhou@xxxxxxxxxxxxxxx>
Cc: Gluster Devel <gluster-devel@xxxxxxxxxxx>
Subject: Re: could you help to check about a glusterfs issue seems to be related to ctime
Hi,
I figured out what's happening. The issue is that the file has 'c|a|m' time set to future (The file is created after the date is set to +30 days). This
is done from client-1. On client-2 with correct date, when data is appended, it doesn't update the mtime and ctime because of both mtime and ctime is less than
already set time on the file. This protection is required to keep the latest time when two clients are writing to the same file. We update c|m|a time only if it's greater than
existing time. As a result, the perf xlators on client1 which relies on mtime doesn't send read to server as it thinks nothing is changed as in this case the times haven't
changed.
Workarounds:
1. Disabling quick-read solved the issue for me.
I don't know how real this kind of workload is? Is this a normal scenario ?
The other thing to do is to remove that protection of updating time only if it's greater but that would open up the race when two clients are updating the same file.
This would result in keeping the older time than the latest. This requires code change and I don't think that should be done.
Exactly, I am also curious about this. I will debug and update about what's exactly happening.
I used to think the file is cached in some client side buffer, because I’ve checked from different sn brick, the file content are all correct. But when I open client side trace
level log, and cat the file, I only find lookup/open/flush fop from fuse-bridge side, I am just wondering how is file content served to client side? Should not there be readv fop seen from trace log?
cynthia
From: Zhou, Cynthia (NSB - CN/Hangzhou)
Sent: 2020年3月11日 15:54
To: 'Kotresh Hiremath Ravishankar' <khiremat@xxxxxxxxxx>
Subject: RE: could you help to check about a glusterfs issue seems to be related to ctime
Does that require, that for all the time client should be time synched? What if the client time is not synched for a while? And then restored?
I make a test when time has been restored and then client change the file, the file’s modify time, access times remains to be wrong, is that correct?
root@mn-0:/home/robot]
# echo "fromm mn-0">>/mnt/export/testfile
[root@mn-0:/home/robot]
# stat /mnt/export/testfile
File: /mnt/export/testfile
Size: 30 Blocks: 1 IO Block: 131072 regular file
Device: 28h/40d Inode: 9855109080001305442 Links: 1
Access: (0644/-rw-r--r--) Uid: ( 0/ root) Gid: ( 615/_nokfsuifileshare)
Access: 2020-05-10 09:33:59.713840197 +0300
Modify: 2020-05-10 09:33:59.713840197 +0300
Change: 2020-05-10 09:33:59.714413772 +0300 //remains to be future time
Birth: -
[root@mn-0:/home/robot]
# cat /mnt/export/testfil
cat: /mnt/export/testfil: No such file or directory
[root@mn-0:/home/robot]
# cat /mnt/export/testfile
from mn0
from mn-1
fromm mn-0
[root@mn-0:/home/robot]
# date
Wed 11 Mar 2020 09:05:58 AM EET
[root@mn-0:/home/robot]
cynthia
From: Kotresh Hiremath Ravishankar <khiremat@xxxxxxxxxx>
Sent: 2020年3月11日 15:41
To: Zhou, Cynthia (NSB - CN/Hangzhou) <cynthia.zhou@xxxxxxxxxxxxxxx>
Subject: Re: could you help to check about a glusterfs issue seems to be related to ctime
But there are times that ntp service went wrong, and time on two storage nodes may be not synced.
Or do you mean when can not guarantee that the time on two clients is synched, we should not enable this ctime feature?
Yes, that's correct. The ctime feature relies on the time generated at the client (that's the utime xlator loaded in client) and hence
expects all clients to be ntp synced.
Without ctime feature, is there some way to avoid this “file changed as we read it” issue?
Unfortunately no. That's the only way as of now.
cynthia
From: Kotresh Hiremath Ravishankar <khiremat@xxxxxxxxxx>
Sent: 2020年3月11日 15:12
To: Zhou, Cynthia (NSB - CN/Hangzhou) <cynthia.zhou@xxxxxxxxxxxxxxx>
Subject: Re: could you help to check about a glusterfs issue seems to be related to ctime
Hi,
I have not looked at it. I will take a look and update you. But one of the pre-requisite for ctime feature is that the clients should be time synced (ntp or other means).
Could you try your reproducer by syncing the time of all clients and update me back ?
Thanks,
I make some test
After change date to future, and touch a file write sth, then restore time to normal.
Then append sth to the file, the file modify time access time is still future, it is not the same with ext4,
I think this is wrong.
[root@mn-0:/home/robot]
# echo "fromm mn-0">>/mnt/export/testfile
[root@mn-0:/home/robot]
# stat /mnt/export/testfile
File: /mnt/export/testfile
Size: 30 Blocks: 1 IO Block: 131072 regular file
Device: 28h/40d Inode: 9855109080001305442 Links: 1
Access: (0644/-rw-r--r--) Uid: ( 0/ root) Gid: ( 615/_nokfsuifileshare)
Access: 2020-05-10 09:33:59.713840197 +0300
Modify: 2020-05-10 09:33:59.713840197 +0300
Change: 2020-05-10 09:33:59.714413772 +0300
Birth: -
[root@mn-0:/home/robot]
# cat /mnt/export/testfil
cat: /mnt/export/testfil: No such file or directory
[root@mn-0:/home/robot]
# cat /mnt/export/testfile
from mn0
from mn-1
fromm mn-0
[root@mn-0:/home/robot]
# date
Wed 11 Mar 2020 09:05:58 AM EET
[root@mn-0:/home/robot]
cynthia
From: Zhou, Cynthia (NSB - CN/Hangzhou)
Sent: 2020年3月11日 14:41
To: 'khiremat@xxxxxxxxxx' <khiremat@xxxxxxxxxx>
Subject: could you help to check about a glusterfs issue seems to be related to ctime
Hi glusterfs expert,
Good day!
Do you have a time to check the ticket :
https://bugzilla.redhat.com/show_bug.cgi?id=1811907 ?
After I disable feature ctime this issue is gone, however, I meet erro “file changed as we read it” when using tar command.
Have you any idea?
Thanks!
Cynthia
Nokia storage team
--
--
--
--
--
|
