Hi,
On Wed, Mar 17, 2021 at 10:45 AM Dan Williams <dan.j.williams@xxxxxxxxx> wrote:
>
> On Wed, Mar 17, 2021 at 10:20 AM Sathyanarayanan Kuppuswamy Natarajan
> <sathyanarayanan.nkuppuswamy@xxxxxxxxx> wrote:
> >
> > Hi,
> >
> > On Wed, Mar 17, 2021 at 9:31 AM Dan Williams <dan.j.williams@xxxxxxxxx> wrote:
> > >
> > > On Tue, Mar 16, 2021 at 10:31 PM Lukas Wunner <lukas@xxxxxxxxx> wrote:
> > > >
> > > > On Tue, Mar 16, 2021 at 10:08:31PM -0700, Dan Williams wrote:
> > > > > On Tue, Mar 16, 2021 at 9:14 PM Lukas Wunner <lukas@xxxxxxxxx> wrote:
> > > > > >
> > > > > > On Fri, Mar 12, 2021 at 07:32:08PM -0800, sathyanarayanan.kuppuswamy@xxxxxxxxxxxxxxx wrote:
> > > > > > > + if ((events == PCI_EXP_SLTSTA_DLLSC) && is_dpc_reset_active(pdev)) {
> > > > > > > + ctrl_info(ctrl, "Slot(%s): DLLSC event(DPC), skipped\n",
> > > > > > > + slot_name(ctrl));
> > > > > > > + ret = IRQ_HANDLED;
> > > > > > > + goto out;
> > > > > > > + }
> > > > > >
> > > > > > Two problems here:
> > > > > >
> > > > > > (1) If recovery fails, the link will *remain* down, so there'll be
> > > > > > no Link Up event. You've filtered the Link Down event, thus the
> > > > > > slot will remain in ON_STATE even though the device in the slot is
> > > > > > no longer accessible. That's not good, the slot should be brought
> > > > > > down in this case.
> > > > >
> > > > > Can you elaborate on why that is "not good" from the end user
> > > > > perspective? From a driver perspective the device driver context is
> > > > > lost and the card needs servicing. The service event starts a new
> > > > > cycle of slot-attention being triggered and that syncs the slot-down
> > > > > state at that time.
> > > >
> > > > All of pciehp's code assumes that if the link is down, the slot must be
> > > > off. A slot which is in ON_STATE for a prolonged period of time even
> > > > though the link is down is an oddity the code doesn't account for.
> > > >
> > > > If the link goes down, the slot should be brought into OFF_STATE.
> > > > (It's okay though to delay bringdown until DPC recovery has completed
> > > > unsuccessfully, which is what the patch I'm proposing does.)
> > > >
> > > > I don't understand what you mean by "service event". Someone unplugging
> > > > and replugging the NVMe drive?
> > >
> > > Yes, service meaning a technician physically removes the card.
> > >
> > > >
> > > >
> > > > > > (2) If recovery succeeds, there's a race where pciehp may call
> > > > > > is_dpc_reset_active() *after* dpc_reset_link() has finished.
> > > > > > So both the DPC Trigger Status bit as well as pdev->dpc_reset_active
> > > > > > will be cleared. Thus, the Link Up event is not filtered by pciehp
> > > > > > and the slot is brought down and back up even though DPC recovery
> > > > > > was succesful, which seems undesirable.
> > > > >
> > > > > The hotplug driver never saw the Link Down, so what does it do when
> > > > > the slot transitions from Link Up to Link Up? Do you mean the Link
> > > > > Down might fire after the dpc recovery has completed if the hotplug
> > > > > notification was delayed?
> > > >
> > > > If the Link Down is filtered and the Link Up is not, pciehp will
> > > > bring down the slot and then bring it back up. That's because pciehp
> > > > can't really tell whether a DLLSC event is Link Up or Link Down.
> > > >
> > > > It just knows that the link was previously up, is now up again,
> > > > but must have been down intermittently, so transactions to the
> > > > device in the slot may have been lost and the slot is therefore
> > > > brought down for safety. Because the link is up, it is then
> > > > brought back up.
> > >
> > > I wonder why we're not seeing that effect in testing?
> >
> > In our test case, there is a good chance that the LINK UP event is also
> > filtered. We change the dpc_reset_active status only after we verify
> > the link is up. So if hotplug handler handles the LINK UP event before
> > we change the status of dpc_reset_active, then it will not lead to the
> > issue mentioned by Lukas.
> >
>
> Ah, ok, we're missing a flush of the hotplug event handler after the
> link is up to make sure the hotplug handler does not see the Link Up.
Flush of hotplug event after successful recovery, and a simulated hotplug link
down event after link recovery fails should solve the problems raised
by Lukas. I assume Lukas' proposal adds this support. I will check his patch
shortly.
> I'm not immediately seeing how the new proposal ensures that there is
> no Link Up event still in flight after DPC completes its work.
> Wouldn't it be required to throw away Link Up to Link Up transitions?
--
Sathyanarayanan Kuppuswamy
Linux Kernel Developer
