Re: [PATCH] PCI: controller: Restore PCI_REASSIGN_ALL_BUS when PCI_PROBE_ONLY is enabled

[Manivannan Sadhasivam <manivannan.sadhasivam@linaro.org>]

On Fri, Jan 17, 2025 at 04:24:14PM +0800, Bo Sun wrote:
> On our Marvell OCTEON CN96XX board, we observed the following panic on
> the latest kernel:
> Unable to handle kernel NULL pointer dereference at virtual address 0000000000000080
> Mem abort info:
>   ESR = 0x0000000096000005
>   EC = 0x25: DABT (current EL), IL = 32 bits
>   SET = 0, FnV = 0
>   EA = 0, S1PTW = 0
>   FSC = 0x05: level 1 translation fault
> Data abort info:
>   ISV = 0, ISS = 0x00000005, ISS2 = 0x00000000
>   CM = 0, WnR = 0, TnD = 0, TagAccess = 0
>   GCS = 0, Overlay = 0, DirtyBit = 0, Xs = 0
> [0000000000000080] user address but active_mm is swapper
> Internal error: Oops: 0000000096000005 [#1] PREEMPT SMP
> Modules linked in:
> CPU: 9 UID: 0 PID: 1 Comm: swapper/0 Not tainted 6.13.0-rc7-00149-g9bffa1ad25b8 #1
> Hardware name: Marvell OcteonTX CN96XX board (DT)
> pstate: 60400009 (nZCv daif +PAN -UAO -TCO -DIT -SSBS BTYPE=--)
> pc : of_pci_add_properties+0x278/0x4c8
> lr : of_pci_add_properties+0x258/0x4c8
> sp : ffff8000822ef9b0
> x29: ffff8000822ef9b0 x28: ffff000106dd8000 x27: ffff800081bc3b30
> x26: ffff800081540118 x25: ffff8000813d2be0 x24: 0000000000000000
> x23: ffff00010528a800 x22: ffff000107c50000 x21: ffff0001039c2630
> x20: ffff0001039c2630 x19: 0000000000000000 x18: ffffffffffffffff
> x17: 00000000a49c1b85 x16: 0000000084c07b58 x15: ffff000103a10f98
> x14: ffffffffffffffff x13: ffff000103a10f96 x12: 0000000000000003
> x11: 0101010101010101 x10: 000000000000002c x9 : ffff800080ca7acc
> x8 : ffff0001038fd900 x7 : 0000000000000000 x6 : 0000000000696370
> x5 : 0000000000000000 x4 : 0000000000000002 x3 : ffff8000822efa40
> x2 : ffff800081341000 x1 : ffff000107c50000 x0 : 0000000000000000
> Call trace:
>  of_pci_add_properties+0x278/0x4c8 (P)
>  of_pci_make_dev_node+0xe0/0x158
>  pci_bus_add_device+0x158/0x210
>  pci_bus_add_devices+0x40/0x98
>  pci_host_probe+0x94/0x118
>  pci_host_common_probe+0x120/0x1a0
>  platform_probe+0x70/0xf0
>  really_probe+0xb4/0x2a8
>  __driver_probe_device+0x80/0x140
>  driver_probe_device+0x48/0x170
>  __driver_attach+0x9c/0x1b0
>  bus_for_each_dev+0x7c/0xe8
>  driver_attach+0x2c/0x40
>  bus_add_driver+0xec/0x218
>  driver_register+0x68/0x138
>  __platform_driver_register+0x2c/0x40
>  gen_pci_driver_init+0x24/0x38
>  do_one_initcall+0x4c/0x278
>  kernel_init_freeable+0x1f4/0x3d0
>  kernel_init+0x28/0x1f0
>  ret_from_fork+0x10/0x20
> Code: aa1603e1 f0005522 d2800044 91000042 (f94040a0)
> 
> This regression was introduced by commit 7246a4520b4b ("PCI: Use
> preserve_config in place of pci_flags"). On our board, the 002:00:07.0
> bridge is misconfigured by the bootloader. Both its secondary and
> subordinate bus numbers are initialized to 0, while its fixed secondary
> bus number is set to 8.

What do you mean by 'fixed secondary bus number'?

> However, bus number 8 is also assigned to another
> bridge (0002:00:0f.0). Although this is a bootloader issue, before the
> change in commit 7246a4520b4b, the PCI_REASSIGN_ALL_BUS flag was
> set by default when PCI_PROBE_ONLY was enabled, ensuing that all the
> bus number for these bridges were reassigned, avoiding any conflicts.
> 

Isn't the opposite? PCI_REASSIGN_ALL_BUS was only added if the PCI_PROBE_ONLY
flag was not set:

	/* Do not reassign resources if probe only */
	if (!pci_has_flag(PCI_PROBE_ONLY))
		pci_add_flags(PCI_REASSIGN_ALL_BUS);


> After the change introduced in commit 7246a4520b4b, the bus numbers
> assigned by the bootloader are reused by all other bridges, except
> the misconfigured 002:00:07.0 bridge. The kernel attempt to reconfigure
> 002:00:07.0 by reusing the fixed secondary bus number 8 assigned by
> bootloader. However, since a pci_bus has already been allocated for
> bus 8 due to the probe of 0002:00:0f.0, no new pci_bus allocated for
> 002:00:07.0.

How come 0002:00:0f.0 is enumerated before 0002:00:07.0 in a depth first manner?

> This results in a pci bridge device without a pci_bus
> attached (pdev->subordinate == NULL). Consequently, accessing
> pdev->subordinate in of_pci_prop_bus_range() leads to a NULL pointer
> dereference.
> 

Looks like it is a bug to let a bridge proceed without 'pdev->subordinate'
assigned.

- Mani

-- 
மணிவண்ணன் சதாசிவம்