Re: [PATCH v2 00/19] scsi: libsas: Remove in_interrupt() check

["Ahmed S. Darwish" <a.darwish@linutronix.de>]

On Tue, Jan 12, 2021 at 04:00:57PM +0000, John Garry wrote:
...
>
> I boot-tested on my machines which have hisi_sas v2 and v3 hw, and it's ok.
> I will ask some guys to test a bit more.
>

Thanks a lot!

> And generally the changes look ok. But I just have a slight concern that we
> don't pass the gfp_flags all the way from the origin caller.
>
> So we have some really long callchains, for example:
>
> host.c: sci_controller_error_handler(): atomic, irq handler     (*)
> OR host.c: sci_controller_completion_handler(), atomic, tasklet (*)
>   -> sci_controller_process_completions()
>     -> sci_controller_unsolicited_frame()
>       -> phy.c: sci_phy_frame_handler()
>         -> sci_change_state(SCI_PHY_SUB_AWAIT_SAS_POWER)
>           -> sci_phy_starting_await_sas_power_substate_enter()
>             -> host.c: sci_controller_power_control_queue_insert()
>               -> phy.c: sci_phy_consume_power_handler()
>                 -> sci_change_state(SCI_PHY_SUB_FINAL)
>         -> sci_change_state(SCI_PHY_SUB_FINAL)
>     -> sci_controller_event_completion()
>       -> phy.c: sci_phy_event_handler()
>         -> sci_phy_start_sata_link_training()
>           -> sci_change_state(SCI_PHY_SUB_AWAIT_SATA_POWER)
>             -> sci_phy_starting_await_sata_power_substate_enter
>               -> host.c: sci_controller_power_control_queue_insert()
>                 -> phy.c: sci_phy_consume_power_handler()
>                   -> sci_change_state(SCI_PHY_SUB_FINAL)
>
> So if someone rearranges the code later, adds new callchains, etc., it could
> be missed that the context may have changed than what we assume at the
> bottom. But then passing the flags everywhere is cumbersome, and all the
> libsas users see little or no significant changes anyway, apart from a
> couple.
>

The deep call chains like the one you've quoted are all within the isci
Intel driver (patches #5 => #7), due to the *massive* state transitions
that driver has. But as the commit logs of these three patches show,
almost all of such transitions happened under atomic context anyway and
GFP_ATOMIC was thus used.

The GFP_KERNEL call-chains were all very simple: a workqueue, functions
already calling msleep() or wait_event_timeout() two or three lines
nearby, and so on.

All the other libsas clients (that is, except isci) also had normal call
chains that were IMHO easy to follow.

Thanks,

--
Ahmed S. Darwish
Linutronix GmbH