When reviewing my LAN9118 driver PCD patch [1], Ard Biesheuvel noted
that most calls to gBS->Stall() in this driver seem to be used to
prevent timing issues between the device updating data and the host
reading the values. And that replacing most of these calls with a
MemoryFence() would be more robust.
The only exceptions are the stalls that are enclosed inside retry loops:
- in the AutoNegotiate() function.
This stall is waiting for the link to negotiate, which may require
stalling until it is ready.
- in the Lan9118Initialize() function.
These two stalls are waiting for devices and time out after a number
of retries.
- in the SoftReset() function.
This stall is inside a loop where the comment states:
"If time taken exceeds 100us, then there was an error condition"
In these instances, I kept the stall, but also added a MemoryFence().