Re: [EXTERNAL] [RFC PATCH] powernv/eeh: Fix oops when probing cxl devices

[Sam Bobroff <sbobroff@linux.ibm.com>]

[-- Attachment #1: Type: text/plain, Size: 3639 bytes --]

On Fri, Sep 27, 2019 at 02:45:10PM +0200, Frederic Barrat wrote:
> Recent cleanup in the way EEH support is added to a device causes a
> kernel oops when the cxl driver probes a device and creates virtual
> devices discovered on the FPGA:
> 
>     BUG: Kernel NULL pointer dereference at 0x000000a0
>     Faulting instruction address: 0xc000000000048070
>     Oops: Kernel access of bad area, sig: 7 [#1]
>     ...
>     NIP [c000000000048070] eeh_add_device_late.part.9+0x50/0x1e0
>     LR [c00000000004805c] eeh_add_device_late.part.9+0x3c/0x1e0
>     Call Trace:
>     [c000200e43983900] [c00000000079e250] _dev_info+0x5c/0x6c (unreliable)
>     [c000200e43983980] [c0000000000d1ad0] pnv_pcibios_bus_add_device+0x60/0xb0
>     [c000200e439839f0] [c0000000000606d0] pcibios_bus_add_device+0x40/0x60
>     [c000200e43983a10] [c0000000006aa3a0] pci_bus_add_device+0x30/0x100
>     [c000200e43983a80] [c0000000006aa4d4] pci_bus_add_devices+0x64/0xd0
>     [c000200e43983ac0] [c00800001c429118] cxl_pci_vphb_add+0xe0/0x130 [cxl]
>     [c000200e43983b00] [c00800001c4242ac] cxl_probe+0x504/0x5b0 [cxl]
>     [c000200e43983bb0] [c0000000006bba1c] local_pci_probe+0x6c/0x110
>     [c000200e43983c30] [c000000000159278] work_for_cpu_fn+0x38/0x60
> 
> The root cause is that those cxl virtual devices don't have a
> representation in the device tree and therefore no associated pci_dn
> structure. In eeh_add_device_late(), pdn is NULL, so edev is NULL and
> we oops.
> 
> We never had explicit support for EEH for those virtual
> devices. Instead, EEH events are reported to the (real) pci device and
> handled by the cxl driver. Which can then forward to the virtual
> devices and handle dependencies. The fact that we try adding EEH
> support for the virtual devices is new and a side-effect of the recent
> cleanup.
> 
> This patch fixes it by skipping adding EEH support on powernv for
> devices which don't have a pci_dn structure.
> 
> Fixes: b905f8cdca77 ("powerpc/eeh: EEH for pSeries hot plug")
> Signed-off-by: Frederic Barrat <fbarrat@linux.ibm.com>
> ---
> 
> Sending as an RFC, as I'm afraid of hiding potential issues and would
> be interested in comments. The powernv eeh code expects a struct
> pci_dn, so the fix seems safe. I'm wondering if there could be cases
> (other than capi virtual devices) where we'd want to blow up and fix
> instead of going undetected with this patch.

Looks good to me.

I do think it would be good to detect a missing pci_dn (WARN_ONCE()
might be appropriate).  However to implement it,
pnv_pcibios_bus_add_device() would need a way to detect that a struct
pci_dev is a cxl virtual device. I don't see an easy way to do that; do
you know if it's possible?

One last thing: pseries_pcibios_bus_add_device() also requires a pci_dn.
Do you know if it's possible to trigger a similar issue there, or is it
not possible for some reason?

Cheers,
Sam.

>  arch/powerpc/platforms/powernv/eeh-powernv.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/arch/powerpc/platforms/powernv/eeh-powernv.c b/arch/powerpc/platforms/powernv/eeh-powernv.c
> index 6bc24a47e9ef..6f300ab7f0e9 100644
> --- a/arch/powerpc/platforms/powernv/eeh-powernv.c
> +++ b/arch/powerpc/platforms/powernv/eeh-powernv.c
> @@ -42,7 +42,7 @@ void pnv_pcibios_bus_add_device(struct pci_dev *pdev)
>  {
>  	struct pci_dn *pdn = pci_get_pdn(pdev);
>  
> -	if (eeh_has_flag(EEH_FORCE_DISABLED))
> +	if (!pdn || eeh_has_flag(EEH_FORCE_DISABLED))
>  		return;
>  
>  	dev_dbg(&pdev->dev, "EEH: Setting up device\n");
> -- 
> 2.21.0
> 

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 488 bytes --]