[PATCH v3 0/7] Allow setting caching mode in arch_add_memory() for P2PDMA

[Dan Williams]

On Thu, Feb 27, 2020 at 10:03 AM Jason Gunthorpe <jgg at ziepe.ca> wrote:
>
> On Thu, Feb 27, 2020 at 09:55:04AM -0800, Dan Williams wrote:
> > On Thu, Feb 27, 2020 at 9:43 AM Jason Gunthorpe <jgg at ziepe.ca> wrote:
> > >
> > > On Thu, Feb 27, 2020 at 10:21:50AM -0700, Logan Gunthorpe wrote:
> > > >
> > > >
> > > > On 2020-02-27 10:17 a.m., Jason Gunthorpe wrote:
> > > > >> Instead of this, this series proposes a change to arch_add_memory()
> > > > >> to take the pgprot required by the mapping which allows us to
> > > > >> explicitly set pagetable entries for P2PDMA memory to WC.
> > > > >
> > > > > Is there a particular reason why WC was selected here? I thought for
> > > > > the p2pdma cases there was no kernel user that touched the memory?
> > > >
> > > > Yes, that's correct. I choose WC here because the existing users are
> > > > registering memory blocks without side effects which fit the WC
> > > > semantics well.
> > >
> > > Hm, AFAIK WC memory is not compatible with the spinlocks/mutexs/etc in
> > > Linux, so while it is true the memory has no side effects, there would
> > > be surprising concurrency risks if anything in the kernel tried to
> > > write to it.
> > >
> > > Not compatible means the locks don't contain stores to WC memory the
> > > way you would expect. AFAIK on many CPUs extra barriers are required
> > > to keep WC stores ordered, the same way ARM already has extra barriers
> > > to keep UC stores ordered with locking..
> > >
> > > The spinlocks are defined to contain UC stores though.
> >
> > How are spinlocks and mutexes getting into p2pdma ranges in the first
> > instance? Even with UC, the system has bigger problems if it's trying
> > to send bus locks targeting PCI, see the flurry of activity of trying
> > to trigger faults on split locks [1].
>
> This is not what I was trying to explain.
>
> Consider
>
>  static spinlock lock; // CPU DRAM
>  static idx = 0;
>  u64 *wc_memory = [..];
>
>  spin_lock(&lock);
>  wc_memory[0] = idx++;
>  spin_unlock(&lock);
>
> You'd expect that the PCI device will observe stores where idx is
> strictly increasing, but this is not guarenteed. idx may decrease, idx
> may skip. It just won't duplicate.
>
> Or perhaps
>
>  wc_memory[0] = foo;
>  writel(doorbell)
>
> foo is not guarenteed observable by the device before doorbell reaches
> the device.
>
> All of these are things that do not happen with UC or NC memory, and
> are surprising violations of our programming model.
>
> Generic kernel code should never touch WC memory unless the code is
> specifically designed to handle it.

Ah, yes, agree.