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