Convert using fixup mechanism to suppressing MCK for the duration of config
read/write transaction: while fixups work fine with the case of a precise
exception, we identified a major drawback with this approach when there's
an imprecise case. In this scenario there is the following race condition:
the fixup is (by design) set to catch the instruction following the one
actually causing the exception; if an interrupt (e.g. decrementer) happens
between those two instructions, the ISR code is executed before the fixup
handler the machine check is no longer protected by the fixup handler as it
appears as within the ISR code. In consequence the fixup approach is being
phased out and replaced with explicit suppressing of MCK during a PCIe
config read/write cycle.
Signed-off-by: Grzegorz Bernacki <gjb@semihalf.com>
During config transactions on the PCIe bus an attempt to scan for a
non-existent device can lead to a machine check exception with certain
peripheral devices. In order to avoid crashing in such scenarios the
instrumented versions of the config cycle read routines are introduced, so
the exceptions fixups framework can gracefully recover.
Signed-off-by: Grzegorz Bernacki <gjb@semihalf.com>
Acked-by: Rafal Jaworowski <raj@semihalf.com>
This brings back separate settings for PCIe bus numbers depending on chip
revision, which got eliminated in 2b393b0f0a
commit. 440SPe rev. A does NOT work properly with the same settings as for
the rev. B (no devices are seen on the bus during enumeration).
Signed-off-by: Rafal Jaworowski <raj@semihalf.com>
