On 08/05/2015 10:08 AM, Alexey Kardashevskiy wrote:
This is about VFIO aka PCI passthrough used from QEMU.
KVM is irrelevant here.
QEMU is a machine emulator. It allocates guest RAM from anonymous memory
and these pages are movable which is ok. They may happen to be allocated
from the contiguous memory allocation zone (CMA). Which is also ok as
long they are movable.
However if the guest starts using VFIO (which can be hotplugged into
the guest), in most cases it involves DMA which requires guest RAM pages
to be pinned and not move once their addresses are programmed to
the hardware for DMA.
So we end up in a situation when quite many pages in CMA are not movable
anymore. And we get bunch of these:
[77306.513966] alloc_contig_range: [1f3800, 1f78c4) PFNs busy
[77306.514448] alloc_contig_range: [1f3800, 1f78c8) PFNs busy
[77306.514927] alloc_contig_range: [1f3800, 1f78cc) PFNs busy
IIRC CMA was for mobile devices and their camera/codec drivers and you
don't use QEMU on those? What do you need CMA for in your case?
This is a very rough patch to start the conversation about how to move
pages properly. mm/page_alloc.c does this and
arch/powerpc/mm/mmu_context_iommu.c exploits it.
OK such conversation should probably start by mentioning the VM_PINNED
effort by Peter Zijlstra: https://lkml.org/lkml/2014/5/26/345
It's more general approach to dealing with pinned pages, and moving them
out of CMA area (and compacting them in general) prior pinning is one of
the things that should be done within that framework.
Then there's the effort to enable migrating pages other than LRU during
compaction (and thus CMA allocation): https://lwn.net/Articles/650864/
I don't know if that would be applicable in your use case, i.e. are the
pins for DMA short-lived and can the isolation/migration code wait a bit
for the transfer to finish so it can grab them, or something?
Please do not comment on the style and code placement,
this is just to give some context :)
Obviously, this does not work well - it manages to migrate only few pages
and crashes as it is missing locks/disabling interrupts and I probably
should not just remove pages from LRU list (normally, I guess, only these
can migrate) and a million of other things.
The questions are:
- what is the correct way of telling if the page is in CMA?
is (get_pageblock_migratetype(page) == MIGRATE_CMA) good enough?
Should be.
- how to tell MM to move page away? I am calling migrate_pages() with
an get_new_page callback which allocates a page with GFP_USER but without
GFP_MOVABLE which should allocate new page out of CMA which seems ok but
there is a little convern that we might want to add MOVABLE back when
VFIO device is unplugged from the guest.
Hmm, once the page is allocated, then the migratetype is not tracked
anywhere (except in page_owner debug data). But the unmovable allocations
might exhaust available unmovable pageblocks and lead to fragmentation. So
"add MOVABLE back" would be too late. Instead we would need to tell the
allocator somehow to give us movable page but outside of CMA.
CMA's own
__alloc_contig_migrate_range() avoids this problem by allocating movable
pages, but the range has been already page-isolated and thus the allocator
won't see the pages there.You obviously can't take this approach and
isolate all CMA pageblocks like that. That smells like a new __GFP_FLAG, meh.
- do I need to isolate pages by using isolate_migratepages_range,
reclaim_clean_pages_from_list like __alloc_contig_migrate_range does?
I dropped them for now and the patch uses only @migratepages from
the compact_control struct.
You don't have to do reclaim_clean_pages_from_list(), but the isolation has
to be careful, yeah.
- are there any flags in madvise() to address this (could not
locate any relevant)?
AFAIK there's no madvise(I_WILL_BE_PINNING_THIS_RANGE)
- what else is missing? disabled interrupts? locks?
See what isolate_migratepages_block() does.