[PATCH v2 1/2] perf/dwc_pcie: Avoid nested bus notifier locking

[Yicong Yang]

On 8/12/26 12:06 AM, Zhanpeng Zhang wrote:
> dwc_pcie_pmu_notifier() registers a synthetic platform device directly from
> the BUS_NOTIFY_ADD_DEVICE callback. Bus notifier callbacks run with the
> bus_notifier rwsem held, so platform_device_register_data() enters the
> platform bus notifier chain before the PCI bus notifier returns. When PCI
> host bridges are registered during late init, lockdep reports this as
> possible recursive locking.
> 
> Move platform device reconciliation to a work item that runs after the PCI
> notifier returns. Serialize it with PCI rescan and removal, and parent each
> synthetic device to its Root Port so their lifetimes remain coupled.
> 
> During removal, unregister the perf PMU and its CPU hotplug instance before
> PCI resources disappear, then defer only platform device destruction.
> Disable manual bind controls to preserve that ordering.
> 
> Fixes: af9597adc2f1 ("drivers/perf: add DesignWare PCIe PMU driver")
> Signed-off-by: Zhanpeng Zhang <zhangzhanpeng.jasper at bytedance.com>

please attach the lockdep warning/trace in the commit for better
understanding the issue..

> ---
> Tested on a RISC-V system with CONFIG_KASAN, CONFIG_PROVE_LOCKING,
> CONFIG_DEBUG_LOCK_ALLOC and CONFIG_LOCKDEP enabled. Late-init PCI discovery
> no longer triggers the nested bus notifier warning.
> 
>  drivers/perf/dwc_pcie_pmu.c | 149 +++++++++++++++++++++++++++---------
>  1 file changed, 113 insertions(+), 36 deletions(-)
> 

[...]

>  
> @@ -815,7 +891,11 @@ static int dwc_pcie_pmu_offline_cpu(unsigned int cpu, struct hlist_node *cpuhp_n
>  
>  static struct platform_driver dwc_pcie_pmu_driver = {
>  	.probe = dwc_pcie_pmu_probe,
> -	.driver = {.name = "dwc_pcie_pmu",},
> +	.driver = {
> +		.name = "dwc_pcie_pmu",
> +		.probe_type = PROBE_FORCE_SYNCHRONOUS,

this looks irrelevant to the issue to resolve. need description here.

I suppose this issue is because both the PMU and the PCIe controller
are on the platform bus, right? If we switch the PMU to the faux device
interface [1] this could be solved?

thanks.

[1] https://lore.kernel.org/all/20260708084507.98242-4-yang.yicong@picoheart.com/