Dan Williams <dan.j.williams(a)intel.com> writes:
On Thu, Apr 27, 2017 at 11:41 AM, Jeff Moyer
> Dan Williams <dan.j.williams(a)intel.com> writes:
>>> The sentiment is that programs shouldn't have to grovel around in sysfs
>>> to do stuff related to an open file descriptor or mapping. I don't take
>>> issue with the name. I do worry that something like 'wpq_drain' may
>>> too platform specific, though. The NVM Programming Model specification
>>> is going to call this "deep flush", so maybe that will give you
>>> some inspiration if you do want to change the name.
>> I'll change to "deep_flush", and I quibble that this is related to
>> single open file descriptor or mapping. It really is a "region flush"
>> for giving extra protection for global metadata, but the persistence
>> of individual fds or mappings is handled by ADR. I think an ioctl
>> might give the false impression that every time you flush a cacheline
>> to persistence you need to call the ioctl.
> fsync, for example, may affect more than one fd--all data in the drive
> write cache will be flushed. I don't see how this is so different. I
> think a sysfs file is awkward because it requires an application to
> chase down the correct file in the sysfs hierarchy. If the application
> already has an open fd or a mapping, it should be able to operate on
I'm teetering, but still leaning towards sysfs. The use case that
needs this is device-dax because we otherwise silently do this behind
the application's back on filesystem-dax for fsync / msync.
We may yet get file system support for flush from userspace (NOVA, for
example). So I don't think we should restrict ourselves to only
thinking about the device dax use case.
A device-dax ioctl would be straightforward, but 'deep flush'
that the device-dax instance is fronting persistent memory. There's
nothing persistent memory specific about device-dax except that today
only the nvdimm sub-system knows how to create them, but there's
nothing that prevents other memory regions from being mapped this way.
You're concerned that applications operating on device dax instances
that are not backed by pmem will try to issue a deep flush? Why would
they do that, and why can't you just return failure from the ioctl?
So I'd rather this persistent memory specific mechanism stay with
persistent memory specific portion of the interface rather than plumb
persistent memory details out through the generic device-dax interface
since we have no other intercept point like we do in the
filesystem-dax case to hide this flush.
Look at the block layer. You can issue an ioctl on a block device, and
if the generic block layer can handle it, it does. If not, it gets
passed down to lower layers until either it gets handled, or it bubbles
back up because nobody knew what to do with it. I think you can do the
same thing here, and that solves your layering violation.