Re: nfs race condition

[Nicolas Parpandet <npa@1g6.biz>]

Hello,

I did some more investigations, (wireshark)

The unlink of /tmp/crontab.vYPoHR/crontab is tranlated into a rename nfs call ? (.nfs file)

NFS	670	V3 READDIRPLUS Reply (Call In 731) crontab .. .
NFS	210	V3 LOOKUP Call (Reply In 734), DH: 0x038abb3d/.nfs000000000098008a00000019
NFS	194	V3 LOOKUP Reply (Call In 733) Error: NFS3ERR_NOENT
NFS	254	V3 RENAME Call (Reply In 736), From DH: 0x038abb3d/crontab To DH: 0x038abb3d/.nfs000000000098008a00000019
NFS	338	V3 RENAME Reply (Call In 735)

and after, it tries to remove the directory /tmp/crontab.0wuoNx :
NFS	198	V3 RMDIR Call (Reply In 738), DH: 0x954c28ec/crontab.0wuoNx
NFS	222	V3 RMDIR Reply (Call In 737) Error: NFS3ERR_NOTEMPTY

but the file /tmp/crontab.0wuoNx/crontab was renamed instead of being removed (nfs network, I cannot see .nfs000 file in fact),
this prevent the removing of the directory  /tmp/crontab.0wuoNx (notempty)

It seems a race condition, see "nfs silly rename" keyword, but I have the problem with mysql temp files, and crontab temp file (crontab -e),
am I the only one doing mysql and crontabs over NFS !!!?!?, does anybody have a solution ?


Regards

Nicolas


----- Mail original -----
De: "npa" <npa@1g6.biz>
À: linux-kernel@vger.kernel.org
Cc: "pcoustillas" <pcoustillas@1g6.biz>
Envoyé: Dimanche 4 Octobre 2015 09:57:26
Objet: nfs race condition

Hello, 

I'm doing some " crontab -e, exit with :x" on NFS mount point, 
and there is a race condition whent removing temporary files : 


7701 write(2, "No modification made\n", 21) = 21 
7701 open("/tmp/crontab.vYPoHR", O_RDONLY|O_NONBLOCK|O_DIRECTORY|O_CLOEXEC) = 4 
7701 getdents(4, {{d_ino=73336409, d_off=1, d_reclen=24, d_name="."} {d_ino=73335455, d_off=2, d_reclen=24, d_name=".."} {d_ino=73340547, d_off=3, d_reclen=32, d_name="crontab"}}, 32768) = 80 
7701 unlink("/tmp/crontab.vYPoHR/crontab") = 0 
7701 getdents(4, {}, 32768) = 0 
7701 close(4) = 0 
7701 rmdir("/tmp/crontab.vYPoHR") = -1 ENOTEMPTY (Directory not empty) 
7701 write(2, "/tmp/crontab.vYPoHR: Directory n"..., 41) = 41 
7701 socket(PF_FILE, SOCK_DGRAM|SOCK_CLOEXEC, 0) = 4

The unlink is really done after the rmdir,
nfsv4 & nfsv3 over ipv6, so the rmdir fails and the directory isn't removed.
I have same problem with mysql temporary files.

If I mount with "nocto,noac,lookupcache=none" : same problem
kernel version on server : 3.16.0-4-amd64 (debian8)
kernel version on client : 2.6.32-39-pve (proxmox), same thing with 3.13.0-63-generic (ubuntu)

Please answer me in CC, I'm not subscribed to the list.

Regards

Nicolas