On Tue, Sep 08, 2020 at 11:09:19AM +0200, David Hildenbrand wrote:
On 20.08.20 17:52, Mike Rapoport wrote:
> On Wed, Aug 19, 2020 at 07:45:29PM +0200, David Hildenbrand wrote:
>> On 19.08.20 19:33, Mike Rapoport wrote:
>>> On Wed, Aug 19, 2020 at 02:10:43PM +0200, David Hildenbrand wrote:
>>>> On 19.08.20 13:53, Mike Rapoport wrote:
>>>>> On Wed, Aug 19, 2020 at 12:49:05PM +0200, David Hildenbrand wrote:
>>>>>> On 18.08.20 16:15, Mike Rapoport wrote:
>>>>>>> From: Mike Rapoport <rppt(a)linux.ibm.com>
>>>>>>> Taking pages out from the direct map and bringing them back
>>>>>>> undesired fragmentation and usage of the smaller pages in
>>>>>>> mapping of the physical memory.
>>>>>>> This can be avoided if a significantly large area of the
>>>>>>> would be reserved for secretmem purposes at boot time.
>>>>>>> Add ability to reserve physical memory for secretmem at boot
>>>>>>> "secretmem" kernel parameter and then use that
reserved memory as a global
>>>>>>> pool for secret memory needs.
>>>>>> Wouldn't something like CMA be the better fit? Just
wondering. Then, the
>>>>>> memory can actually be reused for something else while not
>>>>> The memory allocated as secret is removed from the direct map and
>>>>> boot time reservation is intended to reduce direct map
>>>>> and to avoid splitting 1G pages there. So with CMA I'd still
>>>>> allocate 1G chunks for this and once 1G page is dropped from the
>>>>> map it still cannot be reused for anything else until it is freed.
>>>>> I could use CMA to do the boot time reservation, but doing the
>>>>> reservesion directly seemed simpler and more explicit to me.
>>>> Well, using CMA would give you the possibility to let the memory be
>>>> for other purposes until you decide it's the right time to take it
>>>> remove the direct mapping etc.
>>> I still can't say I follow you here. If I reseve a CMA area as a pool
>>> for secret memory 1G pages, it is still reserved and it still cannot be
>>> used for other purposes, right?
>> So, AFAIK, if you create a CMA pool it can be used for any MOVABLE
>> allocations (similar to ZONE_MOVABLE) until you actually allocate CMA
>> memory from that region. Other allocations on that are will then be
>> migrated away (using alloc_contig_range()).
>> For example, if you have a 1~GiB CMA area, you could allocate 4~MB pages
>> from that CMA area on demand (removing the direct mapping, etc ..), and
>> free when no longer needed (instantiating the direct mapping). The free
>> memory in that area could used for MOVABLE allocations.
> The boot time resrvation is intended to avoid splitting 1G pages in the
> direct map. Without the boot time reservation, we maintain a pool of 2M
> pages so the 1G pages are split and 2M pages remain unsplit.
> If I scale your example to match the requirement to avoid splitting 1G
> pages in the direct map, that would mean creating a CMA area of several
> tens of gigabytes and then doing cma_alloc() of 1G each time we need to
> refill the secretmem pool.
> It is quite probable that we won't be able to get 1G from CMA after the
> system worked for some time.
Why? It should only contain movable pages, and if that is not the case,
it's a bug we have to fix. It should behave just as ZONE_MOVABLE.
(although I agree that in corner cases, alloc_contig_pages() might
temporarily fail on some chunks - e.g., with long/short-term page
pinnings - in contrast to memory offlining, it won't retry forever)
The use-case I had in mind for the boot time reservation in secretmem is
a machine that runs VMs and there is a desire to have the VM memory
protected from the host. In a way this should be similar to booting a
host with mem=X where most of the machine memory never gets to be used
by the host kernel.
For such use case, boot time reservation controlled by the command
line parameter seems to me simpler than using CMA. I agree that there is
no way to use the reserved memory for other purpose, but then we won't
need to create physically contiguous chunk of several gigs every time a
VM is created.
> With boot time reservation we won't need physcally
contiguous 1G to
> satisfy smaller allocation requests for secretmem because we don't need
> to maintain 1G mappings in the secretmem pool.
You can allocate within your CMA area however you want - doesn't need to
be whole gigabytes in case there is no need for it.
The whole point of boot time reservation is to prevent splitting 1G
pages in the direct map. Allocating smaller chunks will still cause
fragmentation of the direct map.
Again, the big benefit of CMA is that the reserved memory can be
for other purpose while nobody is actually making use of it.
Right, but I think if a user explicitly asked to use X gigabytes for the
secretmem we can allow that.
> That said, I believe the addition of the boot time reservation, either
> direct or with CMA, can be added as an incrememntal patch after the
> "core" functionality is merged.
I am not convinced that we want to let random processes to do
alloc_pages() in the range of tens of gigabytes. It's not just mlocked
memory. I prefer either using CMA or relying on the boot time
reservations. But let's see if there are other opinions and people just
Having that said, I have no further comments.
David / dhildenb