Re: [PATCH 0/5] ext4: direct IO via iomap infrastructure

["Darrick J. Wong" <darrick.wong@oracle.com>]

On Sat, Aug 24, 2019 at 01:18:31PM +1000, Matthew Bobrowski wrote:
> On Thu, Aug 22, 2019 at 07:41:25PM +0530, Ritesh Harjani wrote:
> > 
> > 
> > On 8/22/19 5:30 PM, Matthew Bobrowski wrote:
> > > On Wed, Aug 21, 2019 at 11:14:07PM +1000, Matthew Bobrowski wrote:
> > > > On Tue, Aug 13, 2019 at 05:57:22PM +0530, RITESH HARJANI wrote:
> > > > > But what I meant was this (I may be wrong here since I haven't
> > > > > really looked into it), but for my understanding I would like to
> > > > > discuss this -
> > > > > 
> > > > > So earlier with this flag(EXT4_STATE_DIO_UNWRITTEN) we were determining on
> > > > > whether a newextent can be merged with ex1 in function
> > > > > ext4_extents_can_be_merged. But now since we have removed that flag we have
> > > > > no way of knowing that whether this inode has any unwritten extents or not
> > > > > from any DIO path.
> > > > > 
> > > > > What I meant is isn't this removal of setting/unsetting of
> > > > > flag(EXT4_STATE_DIO_UNWRITTEN) changing the behavior of this func -
> > > > > ext4_extents_can_be_merged?
> > > > 
> > > > OK, I'm stuck and looking for either clarity, revalidation of my
> > > > thought process, or any input on how to solve this problem for that
> > > > matter.
> > > > 
> > > > In the current ext4 direct IO implementation, the dynamic state flag
> > > > EXT4_STATE_DIO_UNWRITTEN is set/unset for synchronous direct IO
> > > > writes. On the other hand, the flag EXT4_IO_END_UNWRITTEN is set/unset
> > > > for ext4_io_end->flag, and the value of i_unwritten is
> > > > incremented/decremented for asynchronous direct IO writes. All
> > > > mechanisms by which are used to track and determine whether the inode,
> > > > or an IO in flight against a particular inode have any pending
> > > > unwritten extents that need to be converted after the IO has
> > > > completed. In addition to this, we have ext4_can_extents_be_merged()
> > > > performing explicit checks against both EXT4_STATE_DIO_UNWRITTEN and
> > > > i_unwritten and using them to determine whether it can or cannot merge
> > > > a requested extent into an existing extent.
> > > > 
> > > > This is all fine for the current direct IO implementation. However,
> > > > while switching the direct IO code paths over to make use of the iomap
> > > > infrastructure, I believe that we can no longer simply track whether
> > > > an inode has unwritten extents needing to be converted by simply
> > > > setting and checking the EXT4_STATE_DIO_UNWRITTEN flag on the
> > > > inode. The reason being is that there can be multiple direct IO
> > > > operations to unwritten extents running against the inode and we don't
> > > > particularly distinguish synchronous from asynchronous writes within
> > > > ext4_iomap_begin() as there's really no need. So, the only way to
> > > > accurately determine whether extent conversion is deemed necessary for
> > > > an IO operation whether it'd be synchronous or asynchronous is by
> > > > checking the IOMAP_DIO_UNWRITTEN flag within the ->end_io()
> > > > callback. I'm certain that this portion of the logic is correct, but
> > > > we're still left with some issues when it comes to the checks that I
> > > > previously mentioned in ext4_can_extents_be_merged(), which is the
> > > > part I need some input on.
> > > > 
> > > > I was doing some thinking and I don't believe that making use of the
> > > > EXT4_STATE_DIO_UNWRITTEN flag is the solution at all here. This is not
> > > > only for reasons that I've briefly mentioned above, but also because
> > > > of the fact that it'll probably lead to a lot of inaccurate judgements
> > > > while taking particular code paths and some really ugly code that
> > > > creeps close to the definition of insanity. Rather, what if we solve
> > > > this problem by continuing to just use i_unwritten to keep track of
> > > > all the direct IOs to unwritten against running against an inode?
> > > > Within ext4_iomap_begin() post successful creation of unwritten
> > > > extents we'd call atomic_inc(&EXT4_I(inode)->i_unwritten) and
> > > > subsequently within the ->end_io() callback whether we take the
> > > > success or error path we'd call
> > > > atomic_dec(&EXT4_I(inode)->i_unwritten) accordingly? This way we can
> > > > still rely on this value to be used in the check within
> > > > ext4_can_extents_be_merged(). Open for alternate suggestions if anyone
> > > > has any...
> > > 
> > > Actually, no...
> > > 
> > > I've done some more thinking and what I suggested above around the use
> > > of i_unwritten will also not work properly. Using iomap
> > > infrastructure, there is the possibility of calling into the
> > > ->iomap_begin() more than once for a single direct IO operation. This
> > > means that by the time we even get to decrementing i_unwritten in the
> > > ->end_io() callback after converting the unwritten extents we're
> > > already running the possibility of i_unwritten becoming unbalanced
> > > really quickly and staying that way. This also means that the
> > > statement checking i_unwritten in ext4_can_extents_be_merged() will be
> > > affected and potentially result in it being evaluated incorrectly. I
> > > was thinking that we could just decrement i_unwritten in
> > > ->iomap_end(), but that seems to me like it would be racy and also
> > > lead to incorrect results. At this point I'm out of ideas on how to
> > > solve this, so any other ideas would be appreciated!
> > 
> > I will let others also comment, if someone has any other better approach.
> > 
> > 1. One approach is to add the infrastructure in iomap with
> > iomap_dio->private which is filesystem specific pointer. This can be
> > updated by filesystem in ->iomap_begin call into iomap->private.
> > And in case of iomap_dio_rw, this iomap->private will be copied to
> > iomap_dio->private if not already set.
> > 
> > But I think this will eventually become hacky in the sense when you will
> > have to determine on whether the dio->private is already set or not when
> > iomap_apply will be called second time. It will become a problem with AIO
> > DIO in ext4 since we use i_unwritten which tells us whether there is any
> > unwritten extent but it does not tell whether this unwritten extent is due
> > to a DIRECT AIO DIO in progress or a buffered one.
> > 
> > So we can ignore this approach - unless you or someone else have some good
> > design ideas to build on top of above.
> 
> I'm not sure whether _this_ is the solution or not, so let's maybe
> wait for others to comment. One thing that I and probably the iomap
> maintainers would like to avoid is adding any special case code to
> iomap infrastructure, if possible. I mean, from what you suggest it
> seems to be rather generic and not overly intrusive, although I know
> for a fact that iomap infrastructure exists because stuff like
> buffer_heads and the old direct IO code ended up accommodating so many
> different edge cases making it almost unmodifiable and unmaintainable.

I'm probably misunderstanding the ext4 extent cache horribly, but I keep
wondering why any of this is necessary -- why can't ext4 track the
unwritten status in the extent records directly?  And why is there all
this strange "can merge" logic?  If you need to convert blocks X to Y
to written state because a write to those blocks completed, isn't that
just manipulation of a bunch of incore records?  And can't you just seek
back and forth in the extent cache to look for adjacent records to merge
with? <confuseD>

(I'd really prefer not to go adding private fields all over the
place...)

--D

> > 2. Second approach which I was thinking is to track only those extents which
> > are marked unwritten and are under IO. This can be done in ext4_map_blocks.
> > This way we will not have to track a particular inode has any unwritten
> > extents or not, but it will be extent based.
> > Something similar was also done a while ago. Do you think this approach will
> > work in our case?
> > 
> > So with this extents will be scanned in extent status tree to see if any
> > among those are under IO and are unwritten in func
> > ext4_can_extents_be_merged.
> > 
> > https://patchwork.ozlabs.org/patch/1013837/
> 
> Maybe this would be a better approach and I think that it'd
> work. Based upon what I read within in that thread there weren't
> really any objections to the idea, although I can't see that it made
> it upstream, so I may be missing something?
> 
> --M