pci: Don't cache the count of MSI/MSI-X messages before allocation

[jhb]

Summary

A device can in theory change the read-only fields in the MSI/MSI-X
control registers that indicate the maximum number of supported
registers in response to changing other device registers. For
example, certain Intel networking VFs change the number of messages as
a result of changes in the PCI_IOV_ADD_VF callback.

To support this, always read the current value of the relevant control
register in the *_count and *_alloc methods. Once messages have been
allocated, the control register value remains cached.

Reported by: Krzysztof Galazka <[e-mail]>

[kgalazka]

sys/dev/pci/pci.c

1950

Is it possible that e.g. suspend modifies ctrl register between read and write here?

[jhb]

sys/dev/pci/pci.c

1950

It shouldn't be and is probably prevented by Giant atm. Eventually when we have proper locking for new-bus it will ensure that these can't race with suspend. Even now actually, suspend will only try to save and restore this register after a device driver's DEVICE_SUSPEND() method has returned, and a device driver shouldn't be allocating MSI messages while suspended.

[erj]

sys/dev/iavf/iavf_lib.c

1478

Rest in well-deserved peace, my dumb workaround!

[kgalazka]

Ah, that makes sense. Thanks a lot for clarification!