On 2018-01-13 02:32, Lukas Wunner wrote:
On Fri, Jan 12, 2018 at 10:31:12AM -0500, Sinan Kaya wrote:
On 1/12/2018 10:12 AM, Lukas Wunner wrote:
> On Fri, Jan 12, 2018 at 09:26:48AM -0500, Sinan Kaya wrote:
>> On 1/12/2018 5:49 AM, Lukas Wunner wrote:
>> I wonder if we can separate remove from shutdown and just disable the IRQs
>> in shutdown case rather than turning off the slot power etc.
>
> But don't we risk "IRQ xx: nobody cared" splats if we do that?
I assumed code was turning off the slot power etc. aggressively.
After looking at the code some more time, it seems to be doing the
right thing and telling pcie controller not to generate interrupts for
hotplug.
I think this is what is failing for you probably because by the time
you are
shutting down there is nobody to issue the command completion. This
would
repeat for each hotplug capable pcie slot.
You mean we disable Command Completed interrupts and thus the port
can't notify that the command was completed? It seems the code
accommodates to that by polling the PCI_EXP_SLTSTA_CC bit:
if (ctrl->slot_ctrl & PCI_EXP_SLTCTL_HPIE &&
ctrl->slot_ctrl & PCI_EXP_SLTCTL_CCIE)
rc = wait_event_timeout(ctrl->queue, !ctrl->cmd_busy, timeout);
else
rc = pcie_poll_cmd(ctrl, jiffies_to_msecs(timeout));
The problem is that these Thunderbolt controllers never seem to set
the PCI_EXP_SLTSTA_CC bit, resulting in "Timeout on hotplug command"
messages.
I waa thinking of using nowait variant of write function in notification
disable function in order to not introduce new behavior for existing
silicon.
The pciehp code is okay, we just need a workaround for the broken
Thunderbolt 1 chips. This has been a pain point all along, but
your patch made the brokenness visible enough that investigating
and fixing it became unavoidable. So don't worry about your patch,
it's all fine. ;-)
Thanks,
Lukas
