Re: [PATCH] PCI: aspeed: Drop PROBE_PREFER_ASYNCHRONOUS to fix enumeration race

[sashiko-bot@kernel.org]

Thank you for your contribution! Sashiko AI review found 1 potential
issue to consider:
- [Medium] PCI: aspeed: dropping PROBE_PREFER_ASYNCHRONOUS does not fix
  EP probe order
--

commit 0a5273357bc4c6f662270192044fe2acfee34187
Author: Jacky Chou <jacky_chou@aspeedtech.com>

PCI: aspeed: Drop PROBE_PREFER_ASYNCHRONOUS to fix enumeration race

This commit removes PROBE_PREFER_ASYNCHRONOUS from aspeed_pcie_driver so
that the ASPEED PCIe Root Complex driver probes synchronously.

> diff --git a/drivers/pci/controller/pcie-aspeed.c b/drivers/pci/controller/pcie-aspeed.c
> index ecb7562f84c01..def5591010cac 100644
> --- a/drivers/pci/controller/pcie-aspeed.c
> +++ b/drivers/pci/controller/pcie-aspeed.c
> @@ -1104,7 +1104,6 @@ static struct platform_driver aspeed_pcie_driver = {
>  		.name = "aspeed-pcie",
>  		.of_match_table = aspeed_pcie_of_match,
>  		.suppress_bind_attrs = true,
> -		.probe_type = PROBE_PREFER_ASYNCHRONOUS,
>  	},
>  	.probe = aspeed_pcie_probe,
>  };

[Severity: Medium]
Does dropping PROBE_PREFER_ASYNCHRONOUS from aspeed_pcie_driver actually
solve the endpoint probe ordering problem described in the commit message?

Registering a struct pci_driver before the bus is enumerated succeeds and
adds the driver to pci_bus_type; its probe() callback does not run until
aspeed_pcie_probe() finishes enumeration and pci_bus_add_device() calls
device_initial_probe(&dev->dev), which binds any already-registered
pci_driver even when aspeed_pcie_probe() runs on an async worker.

If a consumer instead looks up a pci_dev directly from an initcall or an
on-board endpoint requires another driver to run before bus enumeration,
removing PROBE_PREFER_ASYNCHRONOUS still does not guarantee that
aspeed_pcie_probe() runs before late_initcall: aspeed_pcie_driver is
registered at device_initcall via builtin_platform_driver(), and
aspeed_pcie_probe() acquires resets, clocks, and PHYs (h2xrst, port->clk,
port->phy, port->perst) that can defer probing via -EPROBE_DEFER or
fw_devlink until deferred_probe_initcall() at late_initcall.

Meanwhile, forcing synchronous probing serializes the 10 ms mdelay() in
aspeed_host_reset() and the 100 ms msleep() per port in
aspeed_pcie_port_init() on the main boot thread across all RC instances.

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20261109-aspeed-pcie-sync-probe-v1-1-71cc9e718623@aspeedtech.com?part=1