Michal Simek wrote:
The flushing is only necessary when a certain memory region is written to, not flushed, and then allocated for DMA from device. This is not what drivers usually do, but it's nevertheless legal to do so. Maybe it can also happen as a result of memory being freed but not flushed, and then allocated as a DMA buffer. So this flushing prevents, at most, an extremely rare problem. I don't expect to see something go wrong right away in the lack of this flush.
DMA_TO_DEVICE is fine - data has to be flushed.
For DMA_FROM_DEVICE you expect that allocated space will contains data from device.
Which means if you flush them, they will be rewritten by DMA in the next step.
Which means that IMHO you can invalidate them which is faster than flushing.
In light of this, I can't see why flushing would be slower than invalidation, if the cache lines aren't expected to be dirty except for very rare conditions.
As for invalidation of a dirty cache line: Looking at the Microblaze reference manual's description of the wdc instruction, I'm under the impression that each cache line has two flags: Valid and Dirty, as they appear in the pseudocode. I'm not on the clear about what happens if a dirty cache line is invalidated. It would make sense to clear both flags. It also makes sense to write the data back to the RAM in this case. But I really don't know what's actually implemented.