Re: [PATCH] PCI: Allow drivers to opt in to async probing

[Bjorn Helgaas <helgaas@kernel.org>]

On Tue, Jul 15, 2025 at 10:42:47AM +0200, Lukas Wunner wrote:
> On Mon, Jul 14, 2025 at 11:35:35PM -0700, dan.j.williams@intel.com wrote:
> > I too could have swore I see async behavior with cxl_pci. I believe this
> > patch is only affecting async behavior when the driver is loaded before
> > initial arrival of the PCI device.
> > 
> > For the typical modular driver case the late arriving driver also
> > arranges async probing. Lo and behold on current upstream:
> > 
> > [   13.002750] __driver_attach: pci 0000:35:00.0: probing driver cxl_pci asynchronously
> > 
> > ...so this patch is only a change in behavior for built-in drivers
> > loaded before PCI initial scan afaics.
> 
> Right, thank you!
> 
> FWIW, below is a rephrased commit message which seeks to make more
> clear that only initial probing of built-in drivers is affected.
> 
> In case Bjorn wants to replace the commit on pci/enumeration.
> I can also submit this as v2 if desired.  No code changes,
> just a rephrased commit message.

Updated the commit log as below, thanks!

> -- >8 --
> 
> Subject: [PATCH] PCI: Allow built-in drivers to use async initial probing
> 
> The PCI core has historically not allowed built-in drivers to opt in to
> async initial probing:  Drivers may set "PROBE_PREFER_ASYNCHRONOUS", but
> initial probing always happens synchronously.  That's because the PCI core
> uses device_attach() instead of device_initial_probe().
> 
> Should a driver return -EPROBE_DEFER on initial probe, reprobing later on
> does honor the PROBE_PREFER_ASYNCHRONOUS setting.  Modular drivers are
> also allowed to probe asynchronously, which is inconsistent.
> 
> The choice of device_attach() is likely not deliberate:  It was introduced
> in 2013 with commit 58d9a38f6fac ("PCI: Skip attaching driver in
> device_add()"), but asynchronous probing was added two years later with
> commit 765230b5f084 ("driver-core: add asynchronous probing support for
> drivers").
> 
> According to the kernel-doc of "enum probe_type", "the end goal is to
> switch the kernel to use asynchronous probing by default".  To this end,
> use device_initial_probe() to allow asynchronous initial probing.  The
> function returns void, making the return value check unnecessary.
> 
> Initial PCI probing often takes on the order of seconds even on laptops,
> so this may speed up booting significantly.
> 
> A small number of PCI drivers already opt in to asynchronous probing.
> Their maintainers (who are all cc'ed) should watch out for issues, now
> that asynchronous probing is not just allowed for deferred and modular
> probing, but also initial probing:
> 
>   hl_pci_driver        drivers/accel/habanalabs/common/habanalabs_drv.c
>   cxl_pci_driver       drivers/cxl/pci.c
>   quicki2c_driver      drivers/hid/intel-thc-hid/intel-quicki2c/pci-quicki2c.c
>   quickspi_driver      drivers/hid/intel-thc-hid/intel-quickspi/pci-quickspi.c
>   i801_driver          drivers/i2c/busses/i2c-i801.c
>   mei_me_driver        drivers/misc/mei/pci-me.c
>   mei_vsc_drv          drivers/misc/mei/platform-vsc.c
>   sdhci_driver         drivers/mmc/host/sdhci-pci-core.c
>   nvme_driver          drivers/nvme/host/pci.c
>   ehci_pci_driver      drivers/usb/host/ehci-pci.c
>   hvfb_pci_stub_driver drivers/video/fbdev/hyperv_fb.c
> 
> All other driver maintainers may test asynchronous probing by specifying
> the command line parameter "driver_async_probe=drv_name1,drv_name2,...",
> and on success setting "probe_type = PROBE_PREFER_ASYNCHRONOUS" in the
> pci_driver struct.
> 
> Signed-off-by: Lukas Wunner <lukas@wunner.de>
> Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
> Link: https://patch.msgid.link/53abe6f5ac7c631f95f5d061aa748b192eda0379.1751614426.git.lukas@wunner.de
> ---
>  drivers/pci/bus.c | 5 +----
>  1 file changed, 1 insertion(+), 4 deletions(-)
> 
> diff --git a/drivers/pci/bus.c b/drivers/pci/bus.c
> index 6904886..b77fd30b 100644
> --- a/drivers/pci/bus.c
> +++ b/drivers/pci/bus.c
> @@ -341,7 +341,6 @@ void pci_bus_add_device(struct pci_dev *dev)
>  {
>  	struct device_node *dn = dev->dev.of_node;
>  	struct platform_device *pdev;
> -	int retval;
>  
>  	/*
>  	 * Can not put in pci_device_add yet because resources
> @@ -372,9 +371,7 @@ void pci_bus_add_device(struct pci_dev *dev)
>  	if (!dn || of_device_is_available(dn))
>  		pci_dev_allow_binding(dev);
>  
> -	retval = device_attach(&dev->dev);
> -	if (retval < 0 && retval != -EPROBE_DEFER)
> -		pci_warn(dev, "device attach failed (%d)\n", retval);
> +	device_initial_probe(&dev->dev);
>  
>  	pci_dev_assign_added(dev);
>  }
> -- 
> 2.47.2
>