Re: [RFC PATCH v2 00/18] scsi: scsi_error: Introduce new error handle mechanism

[Wenchao Hao]

On 2023/9/27 1:37, Mike Christie wrote:
> On 9/26/23 7:57 AM, Wenchao Hao wrote:
> > On 2023/9/26 1:54, Mike Christie wrote:
> > > On 9/25/23 10:07 AM, Wenchao Hao wrote:
> > > > On 2023/9/25 22:55, Christoph Hellwig wrote:
> > > > > Before we add another new error handling mechanism we need to fix the
> > > > > old one first.  Hannes' work on not passing the scsi_cmnd to the various
> > > > > reset handlers hasn't made a lot of progress in the last five years and
> > > > > we'll need to urgently fix that first before adding even more
> > > > > complexity.
> > > > I observed Hannes's patches posted about one year ago, it has not been
> > > > applied yet. I don't know if he is still working on it.
> > > >
> > > > My patches do not depend much on that work, I think the conflict can be
> > > > solved fast between two changes.
> > >
> > > I think we want to figure out Hannes's patches first.
> > >
> > > For a new EH design we will want to be able to do multiple TMFs in parallel
> > > on the same host/target right?
> >
> > It's not necessary to do multiple TMFs in parallel, it's ok to make sure
> > each TMFs do not affect each other.
> >
> > For example, we have two devices: 0:0:0:0 and 0:0:0:1
> >
> > Both of them request device reset, they do not happened in parallel, but
> > would in serial. If 0:0:0:0 is performing device reset in progress, 0:0:0:1
> > just wait 0:0:0:0 to finish.
>
> I see. I guess we still get the benefit of not having to stop other devices
> when doing TMFs.

Yes, it's better to support multiple TMFs in parallel than just run in serial.
I would wait for Hannes's changes to be applied and send my change again.

> I think we still want a common way to allocate/free and manage resources
> drivers will use during this time. Maybe have a init_device/target/cmd/eh_priv
> and exit_device/target/eh_priv (I'm not sure of the name, but something similar
> to the init_cmd_priv/exit_cmd_priv we have for normal commands.
>
> scsi-ml then calls into the new eh with the priv data. Drivers don't have to
> do the preallocation and worry if it's per device/target/host.
>
> I'm not 100% sure about the low level details. Check out how Hannes's is
> handling tag management for TMFs as well.
>
>
> > > The problem is that we need to be able to make forward progress in the EH
> > > path and not fail just because we can't allocate memory for a TMF related
> > > struct. To accomplish this now, drivers will use mempools, preallocate TMF
> > > related structs/mem/tags with their scsi_cmnd related structs, preallocate
> > > per host/target/device related structs or ignore what I wrote above and just
> > > fail.
> > >
> > > Hannes's patches fix up the eh callouts so they don't pass in a scsi_cmnd
> > > when it's not needed. That seems nice because after that, then for your new
> > > EH we can begin to standardize on how to handle preallocation of drivers
> > > resources needed to perform TMFs for your new EH. It could be a per
> > > device/target/host callout to allow drivers to preallocate, then scsi-ml calls
> > > into the drivers with that data. It doesn't have to be exactly like that or
> > > anything close. It would be nice for drivers to not have to think about this
> > > type of thing and scsi-ml just to handle the resource management for us when
> > > there are multiple TMFs in progress.