On Mon, Aug 23, 2021 at 09:44:55AM -0500, Rob Herring wrote:

On Mon, Aug 23, 2021 at 8:10 AM Mike Rapoport [off-list ref] wrote:

quoted

On Mon, Aug 23, 2021 at 12:17:50PM +0200, Geert Uytterhoeven wrote:

quoted

Hi Mike,

On Mon, Aug 16, 2021 at 7:52 AM Mike Rapoport [off-list ref] wrote:

quoted

On Wed, Aug 11, 2021 at 10:50:59AM +0200, Geert Uytterhoeven wrote:

quoted

Prepare for early_init_fdt_scan_reserved_mem() reserving the memory
occupied by an elf core header described in the device tree.
As arch_mem_init() calls early_init_fdt_scan_reserved_mem() before
mips_reserve_vmcore(), the latter needs to check if the memory has
already been reserved before.

Doing memblock_reserve() for the same region is usually fine, did you
encounter any issues without this patch?

Does it also work if the same region is part of an earlier larger
reservation?  I am no memblock expert, so I don't know.
I didn't run into any issues, as my MIPS platform is non-DT, but I
assume arch/arm64/mm/init.c:reserve_elfcorehdr() had the check for
a reason.

The memory will be reserved regardless of the earlier reservation, the
issue may appear when the reservations are made for different purpose. E.g.
if there was crash kernel allocation before the reservation of elfcorehdr.

The check in such case will prevent the second reservation, but, at least
in arch/arm64/mm/init.c:reserve_elfcorehdr() it does not seem to prevent
different users of the overlapping regions to step on each others toes.

If the kernel has been passed in overlapping regions, is there
anything you can do other than hope to get a message out?

Nothing really. I've been thinking about adding flags to memblock.reserved
to at least distinguish firmware regions from the kernel allocations, but I
never got to that.
 

quoted

Moreover, arm64::reserve_elfcorehdr() seems buggy to me, because of there
is only a partial overlap of the elfcorehdr with the previous reservation,
the non-overlapping part of elfcorehdr won't get reserved at all.

What do you suggest as the arm64 version is not the common version?

I'm not really familiar with crash dump internals, so I don't know if
resetting elfcorehdr_addr to ELFCORE_ADDR_ERR is a good idea. I think at
least arm64::reserve_elfcorehdr() should reserve the entire elfcorehdr area
regardless of the overlap. Otherwise it might get overwritten by a random
memblock_alloc().

Rob

-- 
Sincerely yours,
Mike.