RISC-V reserved memory problems

[Conor Dooley]

On Thu, Mar 09, 2023 at 04:12:57PM +0100, Alexandre Ghiti wrote:
> On 3/9/23 13:51, Conor Dooley wrote:
> > On Thu, Mar 09, 2023 at 01:45:05PM +0100, Alexandre Ghiti wrote:
> > > On 3/7/23 12:35, Mike Rapoport wrote:
> > > > Hi Conor,
> > > > 
> > > > Sorry for the delay, somehow this slipped between the cracks.
> > > > 
> > > > On Thu, Feb 02, 2023 at 10:01:26PM +0000, Conor Dooley wrote:
> > > > > Hullo Palmer, Mike & whoever else may read this,
> > > > > 
> > > > > Just reviving this thread from a little while ago as I have been in the
> > > > > area again recently...
> > > > TBH, I didn't really dig deep into the issues, but the thought I had was
> > > > what if DT was mapped via fixmap until the setup_vm_final() and then it
> > > > would be possible to call DT methods early.
> > > > 
> > > > Could be I'm shooting in the dark :)
> > > 
> > > I think I understand the issue now, it's because In riscv, we establish 2
> > > different virtual mappings and we map the device tree at 2 different virtual
> > > addresses, which is the problem.
> > > 
> > > So to me, the solution is:
> > > 
> > > - to revert your previous fix, that is calling
> > > early_init_fdt_scan_reserved_mem() before any call to memblock_alloc()
> > > (which could result in an allocation in the area you want to reserve)
> > > 
> > > - to map the device tree at the same virtual address, because
> > > early_init_fdt_scan_reserved_mem() initializes reserved_mem with the dtb
> > > mapping established in setup_vm() and uses reserved_mem with the new mapping
> > > from setup_vm_final (which is what Mike proposes, we should use the fixmap
> > > region to have the same virtual addresses)
> > > 
> > > Hope that makes sense: I'll come up with something this afternoon for you to
> > > test!
> > Sounds good. Please give me some ELI5 commit messages if you can,
> > explanations for this stuff (which I found took a lot of archaeology to
> > understand) would be very welcome next time we need to go back looking
> > at this stuff.
> 
> 
> Can you give it a try here:
> https://github.com/AlexGhiti/riscv-linux/commits/dev/alex/conor_dtb_fixmap_v1
> ?
> 
> That works for me but I need to carefully explain and check that's correct
> though, not upstreamable as is.

Hey Alex,

So I ended up being pretty sick & had to take a week off. I gave this an
initial spin today & it appears to work.
I'll take it for a longer test-drive when you send a "real" patch for
it, but I tested both the lookup by name & the situation that was
allocating in reserved memory and both were not an issue.

Thanks for working on this,
Conor.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 228 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/linux-riscv/attachments/20230320/85fe78df/attachment.sig>