Re: [RFC PATCH V2 0/5] vhost: accelerate metadata access through vmap()

["Michael S. Tsirkin" <mst@redhat.com>]

On Wed, Mar 13, 2019 at 09:37:08AM -0700, James Bottomley wrote:
> On Wed, 2019-03-13 at 09:05 -0700, Christoph Hellwig wrote:
> > On Tue, Mar 12, 2019 at 01:53:37PM -0700, James Bottomley wrote:
> > > I've got to say: optimize what?  What code do we ever have in the
> > > kernel that kmap's a page and then doesn't do anything with it? You
> > > can
> > > guarantee that on kunmap the page is either referenced (needs
> > > invalidating) or updated (needs flushing). The in-kernel use of
> > > kmap is
> > > always
> > > 
> > > kmap
> > > do something with the mapped page
> > > kunmap
> > > 
> > > In a very short interval.  It seems just a simplification to make
> > > kunmap do the flush if needed rather than try to have the users
> > > remember.  The thing which makes this really simple is that on most
> > > architectures flush and invalidate is the same operation.  If you
> > > really want to optimize you can use the referenced and dirty bits
> > > on the kmapped pte to tell you what operation to do, but if your
> > > flush is your invalidate, you simply assume the data needs flushing
> > > on kunmap without checking anything.
> > 
> > I agree that this would be a good way to simplify the API.   Now
> > we'd just need volunteers to implement this for all architectures
> > that need cache flushing and then remove the explicit flushing in
> > the callers..
> 
> Well, it's already done on parisc ...  I can help with this if we agree
> it's the best way forward.  It's really only architectures that
> implement flush_dcache_page that would need modifying.
> 
> It may also improve performance because some kmap/use/flush/kunmap
> sequences have flush_dcache_page() instead of
> flush_kernel_dcache_page() and the former is hugely expensive and
> usually unnecessary because GUP already flushed all the user aliases.
> 
> In the interests of full disclosure the reason we do it for parisc is
> because our later machines have problems even with clean aliases.  So
> on most VIPT systems, doing kmap/read/kunmap creates a fairly harmless
> clean alias.  Technically it should be invalidated, because if you
> remap the same page to the same colour you get cached stale data, but
> in practice the data is expired from the cache long before that
> happens, so the problem is almost never seen if the flush is forgotten.
>  Our problem is on the P9xxx processor: they have a L1/L2 VIPT L3 PIPT
> cache.  As the L1/L2 caches expire clean data, they place the expiring
> contents into L3, but because L3 is PIPT, the stale alias suddenly
> becomes the default for any read of they physical page because any
> update which dirtied the cache line often gets written to main memory
> and placed into the L3 as clean *before* the clean alias in L1/L2 gets
> expired, so the older clean alias replaces it.
> 
> Our only recourse is to kill all aliases with prejudice before the
> kernel loses ownership.
> 
> > > > Which means after we fix vhost to add the flush_dcache_page after
> > > > kunmap, Parisc will get a double hit (but it also means Parisc
> > > > was the only one of those archs needed explicit cache flushes,
> > > > where vhost worked correctly so far.. so it kinds of proofs your
> > > > point of giving up being the safe choice).
> > > 
> > > What double hit?  If there's no cache to flush then cache flush is
> > > a no-op.  It's also a highly piplineable no-op because the CPU has
> > > the L1 cache within easy reach.  The only event when flush takes a
> > > large amount time is if we actually have dirty data to write back
> > > to main memory.
> > 
> > I've heard people complaining that on some microarchitectures even
> > no-op cache flushes are relatively expensive.  Don't ask me why,
> > but if we can easily avoid double flushes we should do that.
> 
> It's still not entirely free for us.  Our internal cache line is around
> 32 bytes (some have 16 and some have 64) but that means we need 128
> flushes for a page ... we definitely can't pipeline them all.  So I
> agree duplicate flush elimination would be a small improvement.
> 
> James

I suspect we'll keep the copyXuser path around for 32 bit anyway -
right Jason?
So we can also keep using that on parisc...

-- 
MST