Re: [PATCH v4 12/18] PCI: rcar-gen4: Recover the Root Port on link down

[Marek Vasut]

On 10/15/26 5:30 AM, Koichiro Den wrote:
> On R-Car, the controller's reset requests (smlh_req_rst_not,
> link_req_rst_not) arrive on intreq_pcim_sub, the interrupt the MSIs are
> demultiplexed from, so DBI is read as soon as the link goes down. On
> R-Car S4 that is a hazard: DBI accesses issued within a few hundred
> microseconds of an unexpected link down do not complete and hang the
> host. In testing, the first Root Port config read after powering off
> the link partner hung unless delayed by ~300 us.
>
> Check the APP reset status in the interrupt handler before DBI is
> touched. When a reset request is latched, flag the controller as
> needing reinitialization, which masks the sources, ack the request and
> schedule recovery work. The work calls pci_host_handle_link_down() on
> the Root Port, which runs the AER-style recovery and resets the
> controller through reset_root_port(), holding a reference on the port.
> If the Root Port is absent when the work runs, reset the controller
> directly so that successful reinitialization can restore interrupt
> delivery. Hold the rescan lock across the absence check and reset.
> The APP reset-status check also reports a recovery already pending,
> whether started by another interrupt or left over from a failed reset,
> so the handler issues no further DBI accesses until a reset succeeds.
>
> Arm detection after initial PCI enumeration, so recovery does not
> interrupt device discovery or resource assignment. Only unmasked status
> bits are handled and pending latches are cleared when the sources are
> re-enabled, so requests recorded during probe or the reset itself do
> not trigger another recovery. Teardown only disarms the detection: MSI
> delivery has to keep working while devices are removed.
>
> Run recovery on a freezable workqueue to keep it outside device
> suspend/resume. Reject suspend while a reset request is pending,
> including one latched after the parent IRQs were suspended.
> dw_pcie_suspend_noirq() would otherwise access DBI, and the queued
> recovery can run once the system thaws. Disarm Root Port events during
> suspend and re-arm them in .post_init() after resume has set up the
> link, so taking the link down for suspend does not start recovery.
>
> Signed-off-by: Koichiro Den <den@xxxxxxxxxxxxx>
Reviewed-by: Marek Vasut <marek.vasut+renesas@xxxxxxxxxxx>