Re: a link issue maybe introduced in a bug fix " Don't let NFS cache stat after writes"

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 





On Wed, Jan 24, 2018 at 2:24 PM, Pranith Kumar Karampuri <pkarampu@xxxxxxxxxx> wrote:
hi,
       In the same commit you mentioned earlier, there was this code earlier:
-/* Returns 1 if the stat seems to be filled with zeroes. */
-int
-nfs_zero_filled_stat (struct iatt *buf)
-{
-        if (!buf)
-                return 1;
-
-        /* Do not use st_dev because it is transformed to store the xlator id
-         * in place of the device number. Do not use st_ino because by this time
-         * we've already mapped the root ino to 1 so it is not guaranteed to be
-         * 0.
-         */
-        if ((buf->ia_nlink == 0) && (buf->ia_ctime == 0))
-                return 1;
-
-        return 0;
-}
-
-

I moved this to a common library function that can be used in afr as well. Why was it there in NFS? +Niels for answering that question.

If I give you a patch which will assert the error condition, would it be possible for you to figure out the first xlator which is unwinding the iatt with nlink count as zero but ctime as non-zero?

Hey,
      I went through the code, I think gf_fuse_stat2attr() function could be sending both ctime and nlink count as zero to the kernel like you were mentioning. So I guess we should wait for Niels' answer about the need for marking nlink count as zero. We may need to patch fuse code differently and mark entry, attr valid seconds/nseconds to zero, so that we get a lookup on the entry.


On Wed, Jan 24, 2018 at 1:03 PM, Lian, George (NSB - CN/Hangzhou) <george.lian@xxxxxxxxxxxxxxx> wrote:

Hi,  Pranith Kumar,

 

Can you tell me while need set buf->ia_nlink to “0”in function gf_zero_fill_stat(), which API or Application will care it?

If I remove this line and also update corresponding in function gf_is_zero_filled_stat,

The issue seems gone, but I can’t confirm will lead to other issues.

 

So could you please double check it and give your comments?

 

My change is as the below:

 

gf_boolean_t

gf_is_zero_filled_stat (struct iatt *buf)

{

        if (!buf)

                return 1;

 

        /* Do not use st_dev because it is transformed to store the xlator id

         * in place of the device number. Do not use st_ino because by this time

         * we've already mapped the root ino to 1 so it is not guaranteed to be

         * 0.

         */

//        if ((buf->ia_nlink == 0) && (buf->ia_ctime == 0))

        if (buf->ia_ctime == )

                return 1;

 

        return 0;

}

 

void

gf_zero_fill_stat (struct iatt *buf)

{

//       buf->ia_nlink = 0;

        buf->ia_ctime = 0;

}

 

Thanks & Best Regards

George

From: Lian, George (NSB - CN/Hangzhou)
Sent: Friday, January 19, 2018 10:03 AM
To: Pranith Kumar Karampuri <pkarampu@xxxxxxxxxx>; Zhou, Cynthia (NSB - CN/Hangzhou) <cynthia.zhou@xxxxxxxxxxxxxxx>
Cc: Li, Deqian (NSB - CN/Hangzhou) <deqian.li@xxxxxxxxxxxxxxx>; Gluster-devel@xxxxxxxxxxx; Sun, Ping (NSB - CN/Hangzhou) <ping.sun@xxxxxxxxxxxxxxx>


Subject: RE: a link issue maybe introduced in a bug fix " Don't let NFS cache stat after writes"

 

Hi,

>>> Cool, this works for me too. Send me a mail off-list once you are available and we can figure out a way to get into a call and work on this.

 

Have you reproduced the issue per the step I listed in https://bugzilla.redhat.com/show_bug.cgi?id=1531457 and last mail?

 

If not, I would like you could try it yourself , which the difference between yours and mine is just create only 2 bricks instead of 6 bricks.

 

And Cynthia could have a session with you if you needed when I am not available in next Monday and Tuesday.

 

Thanks & Best Regards,

George

 

From: gluster-devel-bounces@gluster.org [mailto:gluster-devel-bounces@gluster.org] On Behalf Of Pranith Kumar Karampuri
Sent: Thursday, January 18, 2018 6:03 PM
To: Lian, George (NSB - CN/Hangzhou) <george.lian@xxxxxxxxxxxxxxx>
Cc: Zhou, Cynthia (NSB - CN/Hangzhou) <cynthia.zhou@xxxxxxxxxxxxxxx>; Li, Deqian (NSB - CN/Hangzhou) <deqian.li@xxxxxxxxxxxxxxx>; Gluster-devel@xxxxxxxxxxx; Sun, Ping (NSB - CN/Hangzhou) <ping.sun@xxxxxxxxxxxxxxx>
Subject: Re: a link issue maybe introduced in a bug fix " Don't let NFS cache stat after writes"

 

 

 

On Thu, Jan 18, 2018 at 12:17 PM, Lian, George (NSB - CN/Hangzhou) <george.lian@xxxxxxxxxxxxxxx> wrote:

Hi,

>>>I actually tried it with replica-2 and replica-3 and then distributed replica-2 before replying to the earlier mail. We can have a debugging session if you are okay with it.

 

It is fine if you cant reproduce the issue in your ENV.

And I has attached the detail reproduce log in the Bugzilla FYI

 

But I am sorry I maybe OOO at Monday and Tuesday next week, so debug session will be fine to me at next Wednesday.

 

Cool, this works for me too. Send me a mail off-list once you are available and we can figure out a way to get into a call and work on this.

 

 

 

Paste the detail reproduce log FYI here:

root@ubuntu:~# gluster peer probe ubuntu

peer probe: success. Probe on localhost not needed

root@ubuntu:~# gluster v create test replica 2 ubuntu:/home/gfs/b1 ubuntu:/home/gfs/b2 force

volume create: test: success: please start the volume to access data

root@ubuntu:~# gluster v start test

volume start: test: success

root@ubuntu:~# gluster v info test

 

Volume Name: test

Type: Replicate

Volume ID: fef5fca3-81d9-46d3-8847-74cde6f701a5

Status: Started

Snapshot Count: 0

Number of Bricks: 1 x 2 = 2

Transport-type: tcp

Bricks:

Brick1: ubuntu:/home/gfs/b1

Brick2: ubuntu:/home/gfs/b2

Options Reconfigured:

transport.address-family: inet

nfs.disable: on

performance.client-io-threads: off

root@ubuntu:~# gluster v status

Status of volume: test

Gluster process                             TCP Port  RDMA Port  Online  Pid

------------------------------------------------------------------------------

Brick ubuntu:/home/gfs/b1                   49152     0          Y       7798

Brick ubuntu:/home/gfs/b2                   49153     0          Y       7818

Self-heal Daemon on localhost               N/A       N/A        Y       7839

 

Task Status of Volume test

------------------------------------------------------------------------------

There are no active volume tasks

 

 

root@ubuntu:~# gluster v set test cluster.consistent-metadata on

volume set: success

 

root@ubuntu:~# ls /mnt/test

ls: cannot access '/mnt/test': No such file or directory

root@ubuntu:~# mkdir -p /mnt/test

root@ubuntu:~# mount -t glusterfs ubuntu:/test /mnt/test

 

root@ubuntu:~# cd /mnt/test

root@ubuntu:/mnt/test# echo "abc">aaa

root@ubuntu:/mnt/test# cp aaa bbb;link bbb ccc

 

root@ubuntu:/mnt/test# kill -9 7818

root@ubuntu:/mnt/test# cp aaa ddd;link ddd eee

link: cannot create link 'eee' to 'ddd': No such file or directory

 

 

Best Regards,

George

 

From: gluster-devel-bounces@gluster.org [mailto:gluster-devel-bounces@gluster.org] On Behalf Of Pranith Kumar Karampuri
Sent: Thursday, January 18, 2018 2:40 PM


To: Lian, George (NSB - CN/Hangzhou) <george.lian@xxxxxxxxxxxxxxx>
Cc: Zhou, Cynthia (NSB - CN/Hangzhou) <cynthia.zhou@xxxxxxxxxxxxxxx>; Gluster-devel@xxxxxxxxxxx; Li, Deqian (NSB - CN/Hangzhou) <deqian.li@xxxxxxxxxxxxxxx>; Sun, Ping (NSB - CN/Hangzhou) <ping.sun@xxxxxxxxxxxxxxx>
Subject: Re: a link issue maybe introduced in a bug fix " Don't let NFS cache stat after writes"

 

 

 

On Thu, Jan 18, 2018 at 6:33 AM, Lian, George (NSB - CN/Hangzhou) <george.lian@xxxxxxxxxxxxxxx> wrote:

Hi,

I suppose the brick numbers in your testing is six, and you just shut down the 3 process.

When I reproduce the issue, I only create a replicate volume with 2 bricks, only let ONE brick working and set cluster.consistent-metadata on,

With this 2 test condition, the issue could 100% reproducible.

 

Hi,

      I actually tried it with replica-2 and replica-3 and then distributed replica-2 before replying to the earlier mail. We can have a debugging session if you are okay with it.

I am in the middle of a customer issue myself(That is the reason for this delay :-( ) and thinking of wrapping it up early next week. Would that be fine with you?

 

 

 

 

16:44:28 :) gluster v status

Status of volume: r2

Gluster process                             TCP Port  RDMA Port  Online  Pid

------------------------------------------------------------------------------

Brick localhost.localdomain:/home/gfs/r2_0  49152     0          Y       5309

Brick localhost.localdomain:/home/gfs/r2_1  49154     0          Y       5330

Brick localhost.localdomain:/home/gfs/r2_2  49156     0          Y       5351

Brick localhost.localdomain:/home/gfs/r2_3  49158     0          Y       5372

Brick localhost.localdomain:/home/gfs/r2_4  49159     0          Y       5393

Brick localhost.localdomain:/home/gfs/r2_5  49160     0          Y       5414

Self-heal Daemon on localhost               N/A       N/A        Y       5436

 

Task Status of Volume r2

------------------------------------------------------------------------------

There are no active volume tasks

 

root@dhcp35-190 - ~

16:44:38 :) kill -9 5309 5351 5393

 

Best Regards,

George

From: gluster-devel-bounces@gluster.org [mailto:gluster-devel-bounces@gluster.org] On Behalf Of Pranith Kumar Karampuri
Sent: Wednesday, January 17, 2018 7:27 PM
To: Lian, George (NSB - CN/Hangzhou) <george.lian@xxxxxxxxxxxxxxx>
Cc: Li, Deqian (NSB - CN/Hangzhou) <deqian.li@xxxxxxxxxxxxxxx>; Gluster-devel@xxxxxxxxxxx; Zhou, Cynthia (NSB - CN/Hangzhou) <cynthia.zhou@xxxxxxxxxxxxxxx>; Sun, Ping (NSB - CN/Hangzhou) <ping.sun@xxxxxxxxxxxxxxx>


Subject: Re: a link issue maybe introduced in a bug fix " Don't let NFS cache stat after writes"

 

 

 

On Mon, Jan 15, 2018 at 1:55 PM, Pranith Kumar Karampuri <pkarampu@xxxxxxxxxx> wrote:

 

 

On Mon, Jan 15, 2018 at 8:46 AM, Lian, George (NSB - CN/Hangzhou) <george.lian@xxxxxxxxxxxxxxx> wrote:

Hi,

 

Have you reproduced this issue? If yes, could you please confirm whether it is an issue or not?

 

Hi,

       I tried recreating this on my laptop and on both master and 3.12 and I am not able to recreate the issue :-(.

Since I was doing this on my laptop, I changed shutting down of the replica to killing the brick process to simulate this test.

Let me know if I missed something.

 

 

Sorry, I am held up with some issue at work, so I think I will get some time day after tomorrow to look at this. In the mean time I am adding more people who know about afr to see if they get a chance to work on this before me.

 

 

And if it is an issue,  do you have any solution for this issue?

 

Thanks & Best Regards,

George

 

From: Lian, George (NSB - CN/Hangzhou)
Sent: Thursday, January 11, 2018 2:01 PM
To: Pranith Kumar Karampuri <pkarampu@xxxxxxxxxx>
Cc: Zhou, Cynthia (NSB - CN/Hangzhou) <cynthia.zhou@xxxxxxxxxxxxxxx>; Gluster-devel@xxxxxxxxxxx; Li, Deqian (NSB - CN/Hangzhou) <deqian.li@xxxxxxxxxxxxxxx>; Sun, Ping (NSB - CN/Hangzhou) <ping.sun@xxxxxxxxxxxxxxx>
Subject: RE: a link issue maybe introduced in a bug fix " Don't let NFS cache stat after writes"

 

Hi,

 

Please see detail test step on https://bugzilla.redhat.com/show_bug.cgi?id=1531457

 

How reproducible:

 

 

Steps to Reproduce:

1.create a volume name "test" with replicated

2.set volume option cluster.consistent-metadata with on: 

  gluster v set test cluster.consistent-metadata on

3. mount volume test on client on /mnt/test

4. create a file aaa size more than 1 byte

   echo "1234567890" >/mnt/test/aaa

5. shutdown a replicat node, let's say sn-1, only let sn-0 worked

6. cp /mnt/test/aaa /mnt/test/bbb; link /mnt/test/bbb /mnt/test/ccc

 

 

BRs

George

 

From: gluster-devel-bounces@gluster.org [mailto:gluster-devel-bounces@gluster.org] On Behalf Of Pranith Kumar Karampuri
Sent: Thursday, January 11, 2018 12:39 PM
To: Lian, George (NSB - CN/Hangzhou) <george.lian@xxxxxxxxxxxxxxx>
Cc: Zhou, Cynthia (NSB - CN/Hangzhou) <cynthia.zhou@xxxxxxxxxxxxxxx>; Gluster-devel@xxxxxxxxxxx; Li, Deqian (NSB - CN/Hangzhou) <deqian.li@xxxxxxxxxxxxxxx>; Sun, Ping (NSB - CN/Hangzhou) <ping.sun@xxxxxxxxxxxxxxx>
Subject: Re: a link issue maybe introduced in a bug fix " Don't let NFS cache stat after writes"

 

 

 

On Thu, Jan 11, 2018 at 6:35 AM, Lian, George (NSB - CN/Hangzhou) <george.lian@xxxxxxxxxxxxxxx> wrote:

Hi,

>>> In which protocol are you seeing this issue? Fuse/NFS/SMB?

It is fuse, within mountpoint by mount -t glusterfs  …“ command.

 

Could you let me know the test you did so that I can try to re-create and see what exactly is going on?

Configuration of the volume and the steps to re-create the issue you are seeing would be helpful in debugging the issue further.

 

 

Thanks & Best Regards,

George

 

From: gluster-devel-bounces@gluster.org [mailto:gluster-devel-bounces@gluster.org] On Behalf Of Pranith Kumar Karampuri
Sent: Wednesday, January 10, 2018 8:08 PM
To: Lian, George (NSB - CN/Hangzhou) <george.lian@xxxxxxxxxxxxxxx>
Cc: Zhou, Cynthia (NSB - CN/Hangzhou) <cynthia.zhou@xxxxxxxxxxxxxxx>; Zhong, Hua (NSB - CN/Hangzhou) <hua.zhong@xxxxxxxxxxxxxxx>; Li, Deqian (NSB - CN/Hangzhou) <deqian.li@xxxxxxxxxxxxxxx>; Gluster-devel@xxxxxxxxxxx; Sun, Ping (NSB - CN/Hangzhou) <ping.sun@xxxxxxxxxxxxxxx>
Subject: Re: a link issue maybe introduced in a bug fix " Don't let NFS cache stat after writes"

 

 

 

On Wed, Jan 10, 2018 at 11:09 AM, Lian, George (NSB - CN/Hangzhou) <george.lian@xxxxxxxxxxxxxxx> wrote:

Hi, Pranith Kumar,

 

I has create a bug on Bugzilla https://bugzilla.redhat.com/show_bug.cgi?id=1531457

After my investigation for this link issue, I suppose your changes on afr-dir-write.c with issue " Don't let NFS cache stat after writes" , your fix is like:

--------------------------------------

       if (afr_txn_nothing_failed (frame, this)) {

                        /*if it did pre-op, it will do post-op changing ctime*/

                        if (priv->consistent_metadata &&

                            afr_needs_changelog_update (local))

                                afr_zero_fill_stat (local);

                        local->transaction.unwind (frame, this);

                }

In the above fix, it set the ia_nlink to 0 if option consistent-metadata is set to on.

And hard link a file with which just created will lead to an error, and the error is caused in kernel function vfs_link:

if (inode->i_nlink == 0 && !(inode->i_state & I_LINKABLE))

             error =  -ENOENT;

 

could you please have a check and give some comments here?

 

When stat is "zero filled", understanding is that the higher layer protocol doesn't send stat value to the kernel and a separate lookup is sent by the kernel to get the latest stat value. In which protocol are you seeing this issue? Fuse/NFS/SMB?

 

 

Thanks & Best Regards,

George




--

Pranith




--

Pranith




--

Pranith




--

Pranith




--

Pranith




--

Pranith




--
Pranith



--
Pranith
_______________________________________________
Gluster-devel mailing list
Gluster-devel@xxxxxxxxxxx
http://lists.gluster.org/mailman/listinfo/gluster-devel

[Index of Archives]     [Gluster Users]     [Ceph Users]     [Linux ARM Kernel]     [Linux ARM]     [Linux Omap]     [Fedora ARM]     [IETF Annouce]     [Security]     [Bugtraq]     [Linux]     [Linux OMAP]     [Linux MIPS]     [eCos]     [Asterisk Internet PBX]     [Linux API]

  Powered by Linux