[PATCH 4/4] dma-mapping: clear dev->dma_ops in arch_teardown_dma_ops

Guenter Roeck [off-list ref] writes:

Hi,

On Mon, Aug 27, 2018 at 10:47:11AM +0200, Christoph Hellwig wrote:

quoted

There is no reason to leave the per-device dma_ops around when
deconfiguring a device, so move this code from arm64 into the
common code.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Robin Murphy <robin.murphy@arm.com>

This patch causes various ppc images to crash in -next due to NULL
DMA ops in dma_alloc_attrs().

I finally remembered where you autobuilder is :)

https://kerneltests.org/builders/qemu-ppc-next/builds/970/steps/qemubuildcommand_1/logs/stdio

Looks like mac99 at least is failing.

Just reproduced it on my setup.

Looking into the code, the macio driver tries to instantiate on
the 0000:f0:0d.0 PCI address (the driver maps to all Apple PCI IDs)
and fails. This results in a call to arch_teardown_dma_ops(). Later,
the same device pointer is used to instantiate ohci-pci, which
crashes in probe because the dma_ops pointer has been cleared.

I don't claim to fully understand the code, but to me it looks like
the pci device is allocated and waiting for a driver to attach to.
See arch/powerpc/kernel/pci_of_scan.c:of_create_pci_dev().
If attaching a driver (here: macio) fails, the same device pointer
is then reused for the next matching driver until a match is found
and the device is successfully instantiated. Of course this fails
quite badly if the device pointer has been scrubbed after the first
failure.

I don't know if this is generic PCI behavior or ppc specific.
I am copying pci and ppc maintainers for additional input.

I would guess this is some PPC special sauce  O_o

Will have a look.

cheers