On Fri, Jul 31, 2020 at 10:16:49AM +0000, Damien Le Moal wrote:

quoted

Let's keep semantics and implementation separate.  For the case
where we report the actual offset we need a size imitation and no
short writes.

OK. So the name of the flag confused me. The flag name should reflect "Do zone
append and report written offset", right ?

Well, we already have O_APPEND, which is the equivalent to append to
the write pointer.  The only interesting addition is that we also want
to report where we wrote.  So I'd rather have RWF_REPORT_OFFSET or so.

Just to clarify, here was my thinking for zonefs:
1) file open with O_APPEND/aio has RWF_APPEND: then it is OK to assume that the
application did not set the aio offset since APPEND means offset==file size. In
that case, do zone append and report back the written offset.

Yes.

2) file open without O_APPEND/aio does not have RWF_APPEND: the application
specified an aio offset and we must respect it, write it that exact same order,
so use regular writes.

Yes.

For regular file systems, with case (1) condition, the FS use whatever it wants
for the implementation, and report back the written offset, which  will always
be the file size at the time the aio was issued.

Yes. 

Your method with a new flag to switch between (1) and (2) is OK with me, but the
"no short writes" may be difficult to achieve in a regular FS, no ? I do not
think current FSes have such guarantees... Especially in the case of buffered
async writes I think.

I'll have to check what Jens recently changed with io_uring, but
traditionally the only short write case for a normal file system is the
artifical INT_MAX limit for the write size.

quoted

Anything with those semantics can be implemented using Zone Append
trivially in zonefs, and we don't even need the exclusive lock in that
case.  But even without that flag anything that has an exclusive lock can
at least in theory be implemented using Zone Append, it just need
support for submitting another request from the I/O completion handler
of the first.  I just don't think it is worth it - with the exclusive
lock we do have access to the zone serialied so a normal write works
just fine.  Both for the sync and async case.

We did switch to have zonefs do append writes in the sync case, always. Hmmm...
Not sure anymore it was such a good idea.

It might be a good idea as long as we have the heavy handed scheduler
locking.  But if we don't have that there doesn't seem to be much of
a reason for zone append.

OK. Makes sense. That said, taking Naohiro's work on btrfs as an example, zone
append is used for every data write, no matter if it is O_APPEND/RWF_APPEND or
not. The size limitation for zone append writes is not needed at all by
applications. Maximum extent size is aligned to the max append write size
internally, and if the application issued a larger write, it loops over multiple
extents, similarly to any regular write may do (if there is overwrite etc...).

True, we probably don't need the limit for a normal file system.

For the regular FS case, my thinking on the semantic really was: if asked to do
so, return the written offset for a RWF_APPEND aios. And I think that
implementing that does not depend in any way on what the FS does internally.

Exactly.

But I think I am starting to see the picture you are drawing here:
1) Introduce a fcntl() to get "maximum size for atomic append writes"
2) Introduce an aio flag specifying "Do atomic append write and report written
offset"

I think we just need the 'report written offset flag', in fact even for
zonefs with the right locking we can handle unlimited write sizes, just
with lower performance.  E.g.

1) check if the write size is larger than the zone append limit

if no:

 - take the shared lock  and issue a zone append, done

if yes:

 - take the exclusive per-inode (zone) lock and just issue either normal
   writes or zone append at your choice, relying on the lock to
   serialize other writers.  For the async case this means we need a
   lock than can be release in a different context than it was acquired,
   which is a little ugly but can be done.

And the implementation is actually completely free to use zone append writes or
regular writes regardless of the "Do atomic append write and report written
offset" being used or not.

Yes, that was my point about keeping the semantics and implementation
separate.