Re: [RFC PATCH v3 1/2] mempinfd: Add new syscall to provide memory pin

[David Hildenbrand]

On 08.02.21 03:27, Song Bao Hua (Barry Song) wrote:
> > -----Original Message-----
> > From: owner-linux-mm@xxxxxxxxx [mailto:owner-linux-mm@xxxxxxxxx] On Behalf Of
> > Matthew Wilcox
> > Sent: Monday, February 8, 2021 2:31 PM
> > To: Song Bao Hua (Barry Song) <song.bao.hua@xxxxxxxxxxxxx>
> > Cc: Wangzhou (B) <wangzhou1@xxxxxxxxxxxxx>; linux-kernel@xxxxxxxxxxxxxxx;
> > iommu@xxxxxxxxxxxxxxxxxxxxxxxxxx; linux-mm@xxxxxxxxx;
> > linux-arm-kernel@xxxxxxxxxxxxxxxxxxx; linux-api@xxxxxxxxxxxxxxx; Andrew
> > Morton <akpm@xxxxxxxxxxxxxxxxxxxx>; Alexander Viro <viro@xxxxxxxxxxxxxxxxxx>;
> > gregkh@xxxxxxxxxxxxxxxxxxx; jgg@xxxxxxxx; kevin.tian@xxxxxxxxx;
> > jean-philippe@xxxxxxxxxx; eric.auger@xxxxxxxxxx; Liguozhu (Kenneth)
> > <liguozhu@xxxxxxxxxxxxx>; zhangfei.gao@xxxxxxxxxx; chensihang (A)
> > <chensihang1@xxxxxxxxxxxxx>
> > Subject: Re: [RFC PATCH v3 1/2] mempinfd: Add new syscall to provide memory
> > pin
> >
> > On Sun, Feb 07, 2021 at 10:24:28PM +0000, Song Bao Hua (Barry Song) wrote:
> > > > > In high-performance I/O cases, accelerators might want to perform
> > > > > I/O on a memory without IO page faults which can result in dramatically
> > > > > increased latency. Current memory related APIs could not achieve this
> > > > > requirement, e.g. mlock can only avoid memory to swap to backup device,
> > > > > page migration can still trigger IO page fault.
> > > >
> > > > Well ... we have two requirements.  The application wants to not take
> > > > page faults.  The system wants to move the application to a different
> > > > NUMA node in order to optimise overall performance.  Why should the
> > > > application's desires take precedence over the kernel's desires?  And why
> > > > should it be done this way rather than by the sysadmin using numactl to
> > > > lock the application to a particular node?
> > >
> > > NUMA balancer is just one of many reasons for page migration. Even one
> > > simple alloc_pages() can cause memory migration in just single NUMA
> > > node or UMA system.
> > >
> > > The other reasons for page migration include but are not limited to:
> > > * memory move due to CMA
> > > * memory move due to huge pages creation
> > >
> > > Hardly we can ask users to disable the COMPACTION, CMA and Huge Page
> > > in the whole system.
> >
> > You're dodging the question.  Should the CMA allocation fail because
> > another application is using SVA?
> >
> > I would say no.
>
> I would say no as well.
>
> While IOMMU is enabled, CMA almost has one user only: IOMMU driver
> as other drivers will depend on iommu to use non-contiguous memory
> though they are still calling dma_alloc_coherent().
>
> In iommu driver, dma_alloc_coherent is called during initialization
> and there is no new allocation afterwards. So it wouldn't cause
> runtime impact on SVA performance. Even there is new allocations,
> CMA will fall back to general alloc_pages() and iommu drivers are
> almost allocating small memory for command queues.
>
> So I would say general compound pages, huge pages, especially
> transparent huge pages, would be bigger concerns than CMA for
> internal page migration within one NUMA.
>
> Not like CMA, general alloc_pages() can get memory by moving
> pages other than those pinned.
>
> And there is no guarantee we can always bind the memory of
> SVA applications to single one NUMA, so NUMA balancing is
> still a concern.
>
> But I agree we need a way to make CMA success while the userspace
> pages are pinned. Since pin has been viral in many drivers, I
> assume there is a way to handle this. Otherwise, APIs like
> V4L2_MEMORY_USERPTR[1] will possibly make CMA fail as there
> is no guarantee that usersspace will allocate unmovable memory
> and there is no guarantee the fallback path- alloc_pages() can
> succeed while allocating big memory.

Long term pinnings cannot go onto CMA-reserved memory, and there is 
similar work to also fix ZONE_MOVABLE in that regard.

https://lkml.kernel.org/r/20210125194751.1275316-1-pasha.tatashin@xxxxxxxxxx

One of the reasons I detest using long term pinning of pages where it 
could be avoided. Take VFIO and RDMA as an example: these things 
currently can't work without them.

What I read here: "DMA performance will be affected severely". That does 
not sound like a compelling argument to me for long term pinnings. 
Please find another way to achieve the same goal without long term 
pinnings controlled by user space - e.g., controlling when migration 
actually happens.

For example, CMA/alloc_contig_range()/memory unplug are corner cases 
that happen rarely, you shouldn't have to worry about them messing with 
your DMA performance.

--
Thanks,

David / dhildenb