Re: [RFC PATCH 0/4] KVM: ioctl for populating guest_memfd

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 





On 20/11/2024 15:13, David Hildenbrand wrote:
> Hi!

Hi! :)

>> Results:
>>    - MAP_PRIVATE: 968 ms
>>    - MAP_SHARED: 1646 ms
>
> At least here it is expected to some degree: as soon as the page cache
> is involved map/unmap gets slower, because we are effectively
> maintaining two datastructures (page tables + page cache) instead of
> only a single one (page cache)
>
> Can you make sure that THP/large folios don't interfere in your
> experiments (e.g., madvise(MADV_NOHUGEPAGE))?

I was using transparent_hugepage=never command line argument in my testing.

$ cat /sys/kernel/mm/transparent_hugepage/enabled
always madvise [never]

Is that sufficient to exclude the THP/large folio factor?

>> While this logic is intuitive, its performance effect is more
>> significant that I would expect.
>
> Yes. How much of the performance difference would remain if you hack out
> the atomic op just to play with it? I suspect there will still be some
> difference.

I have tried that, but could not see any noticeable difference in the overall results.

It looks like a big portion of the bottleneck has moved from shmem_get_folio_gfp/folio_mark_uptodate to finish_fault/__pte_offset_map_lock somehow. I have no good explanation for why:

Orig:
                  - 69.62% do_fault
                     + 44.61% __do_fault
                     + 20.26% filemap_map_pages
                     + 3.48% finish_fault
Hacked:
                  - 67.39% do_fault
                     + 32.45% __do_fault
                     + 21.87% filemap_map_pages
                     + 11.97% finish_fault

Orig:
                     - 3.48% finish_fault
                        - 1.28% set_pte_range
                             0.96% folio_add_file_rmap_ptes
                        - 0.91% __pte_offset_map_lock
                             0.54% _raw_spin_lock
Hacked:
                     - 11.97% finish_fault
                        - 8.59% __pte_offset_map_lock
                           - 6.27% _raw_spin_lock
                                preempt_count_add
                             1.00% __pte_offset_map
                        - 1.28% set_pte_range
                           - folio_add_file_rmap_ptes
                                __mod_node_page_state

> Note that we might improve allocation times with guest_memfd when
> allocating larger folios.

I suppose it may not always be an option depending on requirements to consistency of the allocation latency. Eg if a large folio isn't available at the time, the performance would degrade to the base case (please correct me if I'm missing something).

Heh, now I spot that your comment was as reply to a series.

Yeah, sorry if it wasn't obvious.

If your ioctl is supposed to to more than "allocating memory" like
MAP_POPULATE/MADV_POPULATE+* ... then POPULATE is a suboptimal choice.
Because for allocating memory, we would want to use fallocate() instead.
I assume you want to "allocate+copy"?

Yes, the ultimate use case is "allocate+copy".

I'll note that, as we're moving into the direction of moving
guest_memfd.c into mm/guestmem.c, we'll likely want to avoid "KVM_*"
ioctls, and think about something generic.

Good point, thanks. Are we at the stage where some concrete API has been proposed yet? I might have missed that.

Any clue how your new ioctl will interact with the WIP to have shared
memory as part of guest_memfd? For example, could it be reasonable to
"populate" the shared memory first (via VMA) and then convert that
"allocated+filled" memory to private?

No, I can't immediately see why it shouldn't work. My main concern would probably still be about the latency of the population stage as I can't see why it would improve compared to what we have now, because my feeling is this is linked with the sharedness property of guest_memfd.

Cheers,

David / dhildenb






[Index of Archives]     [Kernel Newbies]     [Security]     [Netfilter]     [Bugtraq]     [Linux FS]     [Yosemite Forum]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Samba]     [Video 4 Linux]     [Device Mapper]     [Linux Resources]

  Powered by Linux