Re: [PATCH v3 1/1] mm: introduce put_user_page*(), placeholder versions

[John Hubbard]

On 3/7/19 6:58 PM, Christopher Lameter wrote:
> On Wed, 6 Mar 2019, john.hubbard@xxxxxxxxx wrote:
> 
>> Dave Chinner's description of this is very clear:
>>
>>     "The fundamental issue is that ->page_mkwrite must be called on every
>>     write access to a clean file backed page, not just the first one.
>>     How long the GUP reference lasts is irrelevant, if the page is clean
>>     and you need to dirty it, you must call ->page_mkwrite before it is
>>     marked writeable and dirtied. Every. Time."
>>
>> This is just one symptom of the larger design problem: filesystems do not
>> actually support get_user_pages() being called on their pages, and letting
>> hardware write directly to those pages--even though that patter has been
>> going on since about 2005 or so.
> 
> Can we distinguish between real filesystems that actually write to a
> backing device and the special filesystems (like hugetlbfs, shm and
> friends) that are like anonymous memory and do not require
> ->page_mkwrite() in the same way as regular filesystems?

Yes. I'll change the wording in the commit message to say "real filesystems
that actually write to a backing device", instead of "filesystems". That
does help, thanks.

> 
> The use that I have seen in my section of the world has been restricted to
> RDMA and get_user_pages being limited to anonymous memory and those
> special filesystems. And if the RDMA memory is of such type then the use
> in the past and present is safe.

Agreed.

> 
> So a logical other approach would be to simply not allow the use of
> long term get_user_page() on real filesystem pages. I hope this patch
> supports that?

This patch neither prevents nor provides that. What this patch does is
provide a prerequisite to clear identification of pages that have had
get_user_pages() called on them.


> 
> It is customary after all that a file read or write operation involve one
> single file(!) and that what is written either comes from or goes to
> memory (anonymous or special memory filesystem).
> 
> If you have an mmapped memory segment with a regular device backed file
> then you already have one file associated with a memory segment and a
> filesystem that does take care of synchronizing the contents of the memory
> segment to a backing device.
> 
> If you now perform RDMA or device I/O on such a memory segment then you
> will have *two* different devices interacting with that memory segment. I
> think that ought not to happen and not be supported out of the box. It
> will be difficult to handle and the semantics will be hard for users to
> understand.
> 
> What could happen is that the filesystem could agree on request to allow
> third party I/O to go to such a memory segment. But that needs to be well
> defined and clearly and explicitly handled by some mechanism in user space
> that has well defined semantics for data integrity for the filesystem as
> well as the RDMA or device I/O.
> 

Those discussions are underway. Dave Chinner and others have been talking
about filesystem leases, for example. The key point here is that we'll still
need, in any of these approaches, to be able to identify the gup-pinned
pages. And there are lots (100+) of call sites to change. So I figure we'd
better get that started.

thanks,
-- 
John Hubbard
NVIDIA