[OpenAFS] Open files disappear (revisited)

[Simon Wilkinson]

Back in February of this year, there was a thread on this list  
entitled "Open files disappear", which detailed I/O errors reported by  
the 'blogbench' benchmark when running on AFS. The archive is at https://lists.openafs.org/pipermail/openafs-info/2009-February/030928.html

I discovered a similar problem yesterday, whilst testing some  
modifications to the cache manager. After a fair amount of sleuthing  
(blow by blow details in the openafs@conference.openafs.org chatroom),  
a group of us managed to get to the bottom of this. What is happening  
is that blogbench is renaming a file on top of a file which it already  
has open, which causes the fileserver to remove the open file. When we  
do a normal remove of a file which is still open, AFS's silly-rename  
support moves it to one side until all of the processes on the local  
machine have closed the file. Silly rename hasn't been implemented for  
files which get clobbered by a rename, hence the errors.

Whether the rename occurs whilst files are still open seems to be down  
to a timing issue in the benchmark itself - the faster the machine, or  
the larger the number of available processors, the more likely this is  
to occur.

RT #124973 is tracking this issue - my intention is to implement the  
missing silly rename support.

Cheers,

Simon.