purpose of modular mca recovery

purpose of modular mca recovery

[Christoph Hellwig]

is there any reason to built this code modular?  After all it would
usually not be loaded when it was needed.  Not allowing modular mca
recovery would help greatly because it's using some symbols that we'd
prefer to unexport before chaning some implementation details
(tasklist_lock and maybe force_sig)

Re: purpose of modular mca recovery

[Russ Anderson]

Christoph Hellwig wrote:
> 
> is there any reason to built this code modular?  After all it would
> usually not be loaded when it was needed.  Not allowing modular mca
> recovery would help greatly because it's using some symbols that we'd
> prefer to unexport before chaning some implementation details
> (tasklist_lock and maybe force_sig)

It is extremely handy for testing to be able to insmod/rmmod 
mca_recovery to test changes, rather than having to reboot the 
kernel.  It is also a handly way to turn it on/off the recovery
code or upgrade without having to reboot.

-- 
Russ Anderson, OS RAS/Partitioning Project Lead  
SGI - Silicon Graphics Inc          rja@sgi.com

RE: purpose of modular mca recovery

[Luck, Tony]

>is there any reason to built this code modular?  After all it would
>usually not be loaded when it was needed.  Not allowing modular mca
>recovery would help greatly because it's using some symbols that we'd
>prefer to unexport before chaning some implementation details
>(tasklist_lock and maybe force_sig)

At the moment there are two reasons:
1) OS distributors don't trust this code very much and they seem to
prefer to ship a kernel that doesn't have mca memory recovery enabled.
But some do seem ok with the idea of supplying a module that a
suitably knowlegable customer would load.  This doesn't seem too
unreasonable a position, after all memory error recovery does involve
killing user applications.

2) Loading/unloading the module provides a mechanism to enable/disable
memory error recovery ... so we aren't tempted to add some ugly file in
/proc to do this.

Neither of these are really strong long-term arguments, but I'd prefer
to keep things this way for a while longer until we can convice a
few more people that this is good code to have built in as the default.

-Tony