[OpenAFS-devel] Linux cache manager throwing away callbacks

[Chas Williams (CONTRACTOR)]

In message <1B1DC52E-3763-44B8-B050-0729B6916E94@inf.ed.ac.uk>,Simon Wilkinson 
writes:
>> checking CUnlinked doesnt prevent that.  this code in .dentry_iput has
>> been around quite a while i think.
>
>So, this is the behaviour I'm seeing... dentry_iput gets called when  
>the dentry count reaches 0, regardless of what the inode count is  
>doing. So, the kernel can end up dereferencing the dentry by calling  
>dentry_iput whilst the cache manager is still holding references on  
>the inode.

the problem would be that the cache manager takes a reference to the
inode to keep it from returning to the kernel.  this keeps you from
using any of the iput() hooks, like .delete_inode or .drop_inode.
.put_inode (which was called for each iput() disappeared at some
point).

the right fix seems like it wont be simple.

>This is hurting me in disconnected mode, but I also wonder if there  
>are potential locations where invalidating the segments may break  
>things when running normally, in particular where we are using  
>background stores against a slow file server.

not sure.  disconnected is new and i never tested background writes.