On 11/12/18 4:52 PM, Florian Fainelli wrote:

On 11/12/18 4:40 PM, Vineet Gupta wrote:

quoted

On 11/12/18 4:38 PM, Florian Fainelli wrote:

quoted

quoted

quoted

 #ifdef CONFIG_BLK_DEV_INITRD
-	if (initrd_start)
-		memblock_reserve(__pa(initrd_start), initrd_end - initrd_start);
+	if (phys_initrd_size) {
+		memblock_reserve(phys_initrd_start, phys_initrd_size);
+		initrd_start = (unsigned long)__va(phys_initrd_start);
+		initrd_end = initrd_start + phys_initrd_size;
+	}
 #endif

The common code now uses phys_initrd*, and you also use the same in ARC code, do
we still need the initrd_* setting here ?
ARC semantics was using them as PA anyways.

Yes, the generic initrd code expects initrd_start/end to be virtual
addresses, which we now directly derive from phys_initrd_start, that
should really be equivalent.

So we can skip this explicit setting above - ARC arch code doesn't access the virt
initrd_start

OK, you are saying we could just have the generic initrd code do this
assignment instead of having each architecture do it, is that a correct
understanding? 

Correct !

If so, I suppose it could be done, whether as of this
patch series or as a follow-up, either way is fine with me.

If it is not too much trouble, I'd prefer this now. I should have chimed earlier.

One possible caveat is if __va() and __phys_to_virt() behave differently
(e.g: because of CONFIG_DEBUG_VIRTUAL or other things).

Thing is, after your patches, we don't use the vanilla initrd_xxx in arch code any
longer. So this becomes just an implementation detail, which core code may or
maynot need and if it does, this needs to work already w/o having to set anything
in arch code. Agree ?

Thx,
-Vineet