On 2021/12/16 19:07, Borislav Petkov wrote:
> On Thu, Dec 16, 2021 at 10:46:12AM +0800, Leizhen (ThunderTown) wrote:
>> The original value (1ULL << 32) is inaccurate
>
> I keep asking *why*?
>
>> and it enlarged the CRASH_ADDR_LOW upper limit.
>
> $ git grep -E "CRASH_ADDR_LOW\W"
> $
>
> I have no clue what you mean here.
#ifdef CONFIG_X86_32
# define CRASH_ADDR_LOW_MAX SZ_512M
# define CRASH_ADDR_HIGH_MAX SZ_512M
#endif
if (!high)
(1) crash_base = memblock_phys_alloc_range(crash_size,
CRASH_ALIGN, CRASH_ALIGN,
CRASH_ADDR_LOW_MAX);
if (!crash_base)
(2) crash_base = memblock_phys_alloc_range(crash_size,
CRASH_ALIGN, CRASH_ALIGN,
CRASH_ADDR_HIGH_MAX);
- if (crash_base >= (1ULL << 32) && reserve_crashkernel_low())
+(3) if (crash_base >= CRASH_ADDR_LOW_MAX && reserve_crashkernel_low())
If the memory of 'crash_base' is successfully allocated at (1), because the last
parameter CRASH_ADDR_LOW_MAX is the upper bound, so we can sure that
"crash_base < CRASH_ADDR_LOW_MAX". So that, reserve_crashkernel_low() will not be
invoked at (3). That's why I said (1ULL << 32) is inaccurate and enlarge the CRASH_ADDR_LOW
upper limit.
If the memory of 'crash_base' is successfully allocated at (2), you see,
CRASH_ADDR_HIGH_MAX = CRASH_ADDR_LOW_MAX = SZ_512M, the same as (1). In fact,
"crashkernel=high," may not be recommended on X86_32.
Is it possible that (CRASH_ADDR_HIGH_MAX >= 4G) and (CRASH_ADDR_LOW_MAX < 4G)?
In this case, the memory allocated at (2) maybe over 4G. But why shouldn't
CRASH_ADDR_LOW_MAX be equal to 4G at this point?
>
>> This is because when the memory is allocated from the low end, the
>> address cannot exceed CRASH_ADDR_LOW_MAX, see "if (!high)" branch.
>
>> If
>> the memory is allocated from the high end, 'crash_base' is greater than or
>> equal to (1ULL << 32), and naturally, it is greater than CRASH_ADDR_LOW_MAX.
>>
>> I think I should update the description, thanks.
>
> I think you should explain why is (1ULL << 32) wrong.
>
> It came from:
>
> eb6db83d1059 ("x86/setup: Do not reserve crashkernel high memory if low reservation failed")
>
> which simply frees the high memory portion when the low reservation
> fails. And the test for that is, is crash base > 4G. So that makes
> perfect sense to me.
>
> So your change is a NOP on 64-bit and it is a NOP on 32-bit by virtue of
> the _low() variant always returning 0 on 32-bit.
>
