Re: [PATCH] fix to serialize unload and discovery

[malahal@us.ibm.com]

Lubin, I am not sure if you are referring to your version of the driver
or the mainline version! If I read the mainline version correctly,
sas_deform_port() actually removes a phy from a port and also removes
all attached devices if it is the last phy in the port.

When a phy/port is gone, an event is posted to the work queue to
eventually call sas_deform_port(). Let me know if I misunderstood the
mainline version.

Thanks, Malahal.

Luben Tuikov [ltuikov@yahoo.com] wrote:
> --- malahal@us.ibm.com wrote:
> > sas_discover_domain() and sas_deform_port() assume that they are single
> > threaded and never run in parallel.
> 
> This is indeed not true.
>
> > That is mostly true as we have a
> > single threaded work queue to handle discovery and other events.
> 
> Hint: when the port is gone, it is gone...
> 
> > The
> > only race I find is unloading a module will call sas_deform_port()
> > without watching if there is a discovery event is in progress. This
> > patch will stop queuing further events, flush the queue, and then call
> > sas_deform_port() to avoid the race.
> 
> You need a much more elaborate (read: general solution, aka algorithm)
> framework to solve this.
> 
> Needless to say, I don't observe these bugs and errors in my version
> of the SAS Stack.
> 
>     Luben
>