Nicholas Piggin <npig...@gmail.com> writes:
> On Mon, 30 Apr 2018 18:34:32 +1000
> Balbir Singh <bsinghar...@gmail.com> wrote:
>
>> This reverts commit 53ecde0b9126ff140abe3aefd7f0ec64d6fa36b0.
>>
>> The commit above changed the memblock size to 1GiB, which did some
>> nice things like create fewer TLB entries for mapping memory at
>> the time of hotplug. The downside is that it changes the granularity
>> at which memory can be hot-plugged and hot-unplugged. The implication
>> is that if we had less than a 1GiB to hot-plug/hot-unplug that
>> would not be possible.
>>
>> The reason we had this fix was to resolve an issue where we did not
>> split mappings on hot-unplug, leaving a TLB entry that spanned the
>> region that was unplugged. This is now fixed by 4dd5f8a99e79 which
>> splits the page table, removing the MMU mappings for the hot-unplugged
>> region correctly.
>>
>> This trade-offs performance as we increase the number of TLB's to
>> map 1GiB of memory, mapped in via hot-plug, but gives us the ability
>> to handle smaller memory sizes.
>
> Why does powernv have any memory block size override at all now?
I dunno, history says:
powerpc/powernv: Set memory_block_size_bytes to 256MB
powerpc sets a low SECTION_SIZE_BITS to accomodate small pseries
boxes. We default to 16MB memory blocks, and boxes with a lot
of memory end up with enormous numbers of sysfs memory nodes.
Set a more reasonable default for powernv of 256MB.
Signed-off-by: Anton Blanchard <an...@samba.org>
It might make more sense to go back to a single definition for powerpc
and have it return a sensible size based on the total available memory -
though I guess memory hotplugged after boot might make that difficult.
Because we definitely have pseries machines with a lot of memory these
days too.
cheers
