Re: [PATCH net-next RFC 00/13] Add devlink reload level option

[Moshe Shemesh]

On 8/5/2020 9:55 AM, Vasundhara Volam wrote:
> On Wed, Aug 5, 2020 at 12:02 PM Moshe Shemesh <moshe@xxxxxxxxxxxx> wrote:
> > On 8/4/2020 1:13 PM, Vasundhara Volam wrote:
> > > On Mon, Aug 3, 2020 at 7:23 PM Moshe Shemesh <moshe@xxxxxxxxxxxx> wrote:
> > > > On 8/3/2020 3:47 PM, Vasundhara Volam wrote:
> > > > > On Mon, Aug 3, 2020 at 5:47 PM Moshe Shemesh <moshe@xxxxxxxxxxxx> wrote:
> > > > > > On 8/3/2020 1:24 PM, Vasundhara Volam wrote:
> > > > > > > On Tue, Jul 28, 2020 at 10:13 PM Jacob Keller <jacob.e.keller@xxxxxxxxx> wrote:
> > > > > > > > On 7/27/2020 10:25 PM, Vasundhara Volam wrote:
> > > > > > > > > On Mon, Jul 27, 2020 at 4:36 PM Moshe Shemesh <moshe@xxxxxxxxxxxx> wrote:
> > > > > > > > > > Introduce new option on devlink reload API to enable the user to select the
> > > > > > > > > > reload level required. Complete support for all levels in mlx5.
> > > > > > > > > > The following reload levels are supported:
> > > > > > > > > >       driver: Driver entities re-instantiation only.
> > > > > > > > > >       fw_reset: Firmware reset and driver entities re-instantiation.
> > > > > > > > > The Name is a little confusing. I think it should be renamed to
> > > > > > > > > fw_live_reset (in which both firmware and driver entities are
> > > > > > > > > re-instantiated).  For only fw_reset, the driver should not undergo
> > > > > > > > > reset (it requires a driver reload for firmware to undergo reset).
> > > > > > > > So, I think the differentiation here is that "live_patch" doesn't reset
> > > > > > > > anything.
> > > > > > > This seems similar to flashing the firmware and does not reset anything.
> > > > > > The live patch is activating fw change without reset.
> > > > > >
> > > > > > It is not suitable for any fw change but fw gaps which don't require reset.
> > > > > >
> > > > > > I can query the fw to check if the pending image change is suitable or
> > > > > > require fw reset.
> > > > > Okay.
> > > > > > > > > >       fw_live_patch: Firmware live patching only.
> > > > > > > > > This level is not clear. Is this similar to flashing??
> > > > > > > > >
> > > > > > > > > Also I have a basic query. The reload command is split into
> > > > > > > > > reload_up/reload_down handlers (Please correct me if this behaviour is
> > > > > > > > > changed with this patchset). What if the vendor specific driver does
> > > > > > > > > not support up/down and needs only a single handler to fire a firmware
> > > > > > > > > reset or firmware live reset command?
> > > > > > > > In the "reload_down" handler, they would trigger the appropriate reset,
> > > > > > > > and quiesce anything that needs to be done. Then on reload up, it would
> > > > > > > > restore and bring up anything quiesced in the first stage.
> > > > > > > Yes, I got the "reload_down" and "reload_up". Similar to the device
> > > > > > > "remove" and "re-probe" respectively.
> > > > > > >
> > > > > > > But our requirement is a similar "ethtool reset" command, where
> > > > > > > ethtool calls a single callback in driver and driver just sends a
> > > > > > > firmware command for doing the reset. Once firmware receives the
> > > > > > > command, it will initiate the reset of driver and firmware entities
> > > > > > > asynchronously.
> > > > > > It is similar to mlx5 case here for fw_reset. The driver triggers the fw
> > > > > > command to reset and all PFs drivers gets events to handle and do
> > > > > > re-initialization.  To fit it to the devlink reload_down and reload_up,
> > > > > > I wait for the event handler to complete and it stops at driver unload
> > > > > > to have the driver up by devlink reload_up. See patch 8 in this patchset.
> > > > > Yes, I see reload_down is triggering the reset. In our driver, after
> > > > > triggering the reset through a firmware command, reset is done in
> > > > > another context as the driver initiates the reset only after receiving
> > > > > an ASYNC event from the firmware.
> > > > Same here.
> > > >
> > > > > Probably, we have to use reload_down() to send firmware command to
> > > > > trigger reset and do nothing in reload_up.
> > > > I had that in previous version, but its wrong to use devlink reload this
> > > > way, so I added wait with timeout for the event handling to complete
> > > > before unload_down function ends. See mlx5_fw_wait_fw_reset_done(). Also
> > > > the event handler stops before load back to have that done by devlink
> > > > reload_up.
> > > But "devlink dev reload" will be invoked by the user only on a single
> > > dev handler and all function drivers will be re-instantiated upon the
> > > ASYNC event. reload_down and reload_up are invoked only the function
> > > which the user invoked.
> > >
> > > Take an example of a 2-port (PF0 and PF1) adapter on a single host and
> > > with some VFs loaded on the device. User invokes "devlink dev reload"
> > > on PF0, ASYNC event is received on 2 PFs and VFs for reset. All the
> > > function drivers will be re-instantiated including PF0.
> > >
> > > If we wait for some time in reload_down() of PF0 and then call load in
> > > reload_up(), this code will be different from other function drivers.
> >
> > I see your point here, but the user run devlink reload command on one
> > PF, in this case of fw-reset it will influence other PFs, but that's a
> > result of the fw-reset, the user if asked for params change or namespace
> > change that was for this PF.
> Right, if any driver is implementing only fw-reset have to leave
> reload_up as an empty function.


No, its not only up the driver. The netns option is implemented by 
devlink and its running between reload_down and reload_up.

> > > > >     And returning from reload
> > > > > does not mean that reset is complete as it is done in another context
> > > > > and the driver notifies the health reporter once the reset is
> > > > > complete. devlink framework may have to allow drivers to implement
> > > > > reload_down only to look more clean or call reload_up only if the
> > > > > driver notifies the devlink once reset is completed from another
> > > > > context. Please suggest.