RPi4's memreserve use in fdt not handled by aarch64 fdt_get_reserved_mem (called via its initarm)

[Mark Millard]

On 2020-Feb-10, at 11:36, Kyle Evans <kevans at freebsd.org> wrote:

> (Removing Jeff from CC- it's becoming clear that this isn't his problem)
> 
> On Mon, Feb 10, 2020 at 12:00 AM Mark Millard <marklmi at yahoo.com> wrote:
>> On 2020-Feb-9, at 21:12, Mark Millard <marklmi at yahoo.com> wrote:
>> 
>>> aarch64 seems to be ignoring the RPi4B's memreserve use.
>>> 
>>> My hypothesis is that head -r356776 and later are
>>> allocating RAM areas that overlap with the
>>> memreserve area.
>>> 
>>> 1st see separate submittals about the live dts text reported
>>> by fdt print in u-boot on the example RPi4B 4 GiByte machine:
>>> 
>>> https://lists.freebsd.org/pipermail/freebsd-arm/2020-February/021207.html
>>> https://lists.freebsd.org/pipermail/freebsd-arm/2020-February/021205.html
>>> 
>>> Then:
>>> 
>>> aarch64's initarm uses fdt_get_reserved_mem and not
>>> fdt_get_reserved_regions. So, looking at what that
>>> implies:
>>> 
>>> int
>>> fdt_get_reserved_mem(struct mem_region *reserved, int *mreserved)
>>> {
>>>       pcell_t reg[FDT_REG_CELLS];
>>>       phandle_t child, root;
>>>       int addr_cells, size_cells;
>>>       int i, rv;
>>> 
>>>       root = OF_finddevice("/reserved-memory");
>>>       if (root == -1) {
>>>               return (ENXIO);
>>>       }
>>> . . . (I'll not list it all to show the lack of
>>> memreserve handling) . . .
>>> 
>>> This does not check for and handle memreserve.
>>> 
>>> By contrast armv7 and armv6 each have an initarm
>>> that uses fdt_get_reserved_regions and that in
>>> turn has:
>>> 
>>> int
>>> fdt_get_reserved_regions(struct mem_region *mr, int *mrcnt)
>>> {
>>>       pcell_t reserve[FDT_REG_CELLS * FDT_MEM_REGIONS];
>>>       pcell_t *reservep;
>>>       phandle_t memory, root;
>>>       int addr_cells, size_cells;
>>>       int i, res_len, rv, tuple_size, tuples;
>>> 
>>>       root = OF_finddevice("/");
>>>       memory = OF_finddevice("/memory");
>>>       if (memory == -1) {
>>>               rv = ENXIO;
>>>               goto out;
>>>       }
>>> 
>>>       if ((rv = fdt_addrsize_cells(OF_parent(memory), &addr_cells,
>>>           &size_cells)) != 0)
>>>               goto out;
>>> 
>>>       if (addr_cells > 2) {
>>>               rv = ERANGE;
>>>               goto out;
>>>       }
>>> 
>>>       tuple_size = sizeof(pcell_t) * (addr_cells + size_cells);
>>> 
>>>       res_len = OF_getproplen(root, "memreserve");
>>>       if (res_len <= 0 || res_len > sizeof(reserve)) {
>>>               rv = ERANGE;
>>>               goto out;
>>>       }
>>> 
>>>       if (OF_getprop(root, "memreserve", reserve, res_len) <= 0) {
>>>               rv = ENXIO;
>>>               goto out;
>>>       }
>>> . . .
>>> 
>>> So this handles memreserve but not /reserved-memory .
>>> 
>>> It appears that for the RPi4B's the 32-bit "normal use"
>>> has lead to aarch64 having memreserve instead of
>>> /reserved-memory .
>>> 
>> 
>> I managed to make a quick test patch for head -r356529
>> but it did not make the RPi4B boot attempt behave
>> differently. So, either I messed up or handling memreserve
>> is not sufficient. (Some alternate information might be
>> covering the address range already?)
>> 
>> I'm not familiar with the subject matter in the code, so
>> I may have messed up the test. I just used:
>> 
>> # svnlite diff /usr/src/sys/dev/fdt/fdt_common.c
>> Index: /usr/src/sys/dev/fdt/fdt_common.c
>> ===================================================================
>> --- /usr/src/sys/dev/fdt/fdt_common.c   (revision 357529)
>> +++ /usr/src/sys/dev/fdt/fdt_common.c   (working copy)
>> @@ -512,6 +512,11 @@
>> 
>>        root = OF_finddevice("/reserved-memory");
>>        if (root == -1) {
>> +               // Fail over to checking for and handling memreserve,
>> +               // such as for a RPi4B.
>> +               if (0 == fdt_get_reserved_regions(reserved,mreserved))
>> +                       return (0);
>> +
>>                return (ENXIO);
>>        }
>> 
> 
> This seems reasonable; specifically CC'ing andrew@ and manu@ to get
> their opinion on the patch, as having authored and reviewed the
> relevant section respectively.

I wrote in a later note, not CC'd to anyone specific at
the time:

Example alternate information was:

       memory at 0 {
               device_type = "memory";
               reg = <0x00000000 0x00000000 0x3b400000 0x00000000 0x40000000 0xbc000000>;
       };

That looks like it avoids indicating anything from the
memreserve region. (I've no clue if such is a requirement
for memory at 0, but it happens to be the case for the test
context.)

I do not know if such is appropriate code or not. For my
specific test context, it makes no difference.

===
Mark Millard
marklmi at yahoo.com
( dsl-only.net went
away in early 2018-Mar)