Re: [PATCH v2] PCI: dw-rockchip: Enable async probe by default

[Niklas Cassel <cassel@kernel.org>]

Hello Grimmauld,

On Thu, Jan 29, 2026 at 03:06:59PM +0100, Grimmauld wrote:
> Hello Anand,
> 
> I have tested this patch.
> Hardware/Kernel information:
> - radxa rock 5c lite
> - rk3588s CPU, arm64
> - defconfig NixOS kernel
> - picked onto 6.18.7
> - DT: rockchip/rk3588s-rock-5c.dtb
> - tested both uboot (mainline) and edk2 (vendor)
> 
> On Fri, Aug 09, 2024 at 01:06:09PM +0530, Anand Moon wrote:
> > Rockchip DWC PCIe driver currently waits for the combo PHY link
> > (PCIe 3.0, PCIe 2.0, and SATA 3.0) to be established link training
> > during boot, it also waits for the link to be up, which could consume
> > several milliseconds during boot.
> 
> I found that without this patch, USB 3 ports as well as the PCIe connector seemingly stay uninitialized during boot on my hardware. 
> This manifests in a bootable USB flash drive loading initrd from bootloader (both uboot and edk2) perfectly, but then fails to mount the rootfs from the drive.
> In effect, boot is not just slower than it should be, it just does not boot all the way at all.
> In that scenario, the devfs entries corresponding to the flash drive are simply missing, same for the sysfs where i would expect the USB device listed.
> Replugging the USB flash drive during initrd does seem to fix that, but is tedious and not viable for a server.
> Similarly, booting from m.2 SSD attached via PCIe fails the same way, with rootfs timing out despite the bootloader correctly reading initrd on the same drive.
> Fwiw, replugging the SSD does not work like it does for USB flash drives, and is even worse of an idea.
> USB 2 ports as well as SD card boots correctly, even without your patch.
> Without your patch, i am seeing "deferred probe pending" in dmesg before boot gets stuck, which was the hint which made me find your patch.
> I am not sure whether that is the actual cause or just a symptom for why drives are not recognized during boot, and am not quite sure how to debug this further.
> 
> > To optimize boot time, this commit allows asynchronous probing.
> > This change enables the PCIe link establishment to occur in the
> > background while other devices are being probed.
> 
> With this patch, booting from SSD or USB 3 port works flawlessly, and i have not seen any regressions with SD card or USB 2 boot, nor any other hardware component.
> This setup has worked for multiple boots without fail, both with traditional ext4 and zfs rootfs being loaded from USB 3 and PCIe.
> Because i require this patch to run my rock 5c from SSD, i am currently running a custom patched kernel, but would highly appreciate this patch making its way to mainline eventually.
> There might well be something else going on here. The proposed patch may not be the "proper" fix to the issues i am seeing, but it does at least work.
> 
> I have NOT tested boot from eMMC (either with or without this patch), though i have no reason to believe it would be impacted.
> 
> I am happy to provide more info as needed. First time posting to the LKML, i hope i am doing this right...
> 
> Tested-by: Grimmauld <grimmauld@grimmauld.de>

I tested this patch again on the latest kernel, and it still results in
the "requesting loading a module with wait allowed while being called from
async context can result in a deadlock" warning from the modules code.
(With the calling code being phylib.)
See the phylib splat that I previously reported in this thread.

Note that I've built the network PHY driver that phylib wants to load
(CONFIG_REALTEK_PHY=y) as built-in. As long as the PHY driver is built
as built-in, I don't think that the problem the modules code is warning
about can happen. (But I also don't understand why it is trying to load
a module when the driver is built as built-in in the first place...)

Anyway, my networking is working perfectly fine even with the splat.


Having async probing for the Rockchip PCIe controller driver, which is
used a LOT of rockchip based SoCs, is a good thing, so I don't think it
is right to avoid enabling async probing just because it results in a
splat on a single rockchip based board (rock5b).


Kind regards,
Niklas