[PATCH v3] PCI: Data corruption happening due to race condition

[Benjamin Herrenschmidt]

On Thu, 2018-07-19 at 20:55 +0200, Lukas Wunner wrote:
> On Thu, Jul 19, 2018 at 9:48 AM, Benjamin Herrenschmidt <benh at kernel.crashing.org wrote:
> > Indeed. However I'm not fan of the solution. Shouldn't we instead have
> > some locking for the content of pci_dev? I've always been wary of us
> > having other similar races in there.
> 
> The solution presented is perfectly fine as it uses atomic bitops which
> obviate the need for locking.  Why do you want to add unnecessary locking
> on top?

Atomic bitops tend to be *more* expensive than a lock.

My concern is that the PCIe code historically had no locking and I
worry we may have other fields in there with similar issues. But maybe
I'm wrong.

> Certain other parts of struct pci_dev use their own locking, e.g.
> pci_bus_sem to protect bus_list.  Most elements can and should
> be accessed lockless for performance.
> 
> 
> > > The powerpc PCI code contains a lot of cruft coming from the depth of
> > > history, including rather nasty assumptions. We want to progressively
> > > clean it up, starting with EEH, but it will take time.
> 
> Then I suggest using the #include "../../../drivers/pci/pci.h" for now
> until the powerpc arch code has been consolidated.

There's also the need both in powerpc and sparc to access the guts of
pci_dev because those archs will "fabricate" as pci_dev from the
device-tree rather than probing it under some circumstances.

Cheers,
Ben.