On Sat, 31 May 2014, Namjae Jeon wrote: > Date: Sat, 31 May 2014 16:40:29 +0900 > From: Namjae Jeon <namjae.jeon@xxxxxxxxxxx> > To: 'Lukáš Czerner' <lczerner@xxxxxxxxxx> > Cc: 'Dave Chinner' <david@xxxxxxxxxxxxx>, 'Theodore Ts'o' <tytso@xxxxxxx>, > 'linux-ext4' <linux-ext4@xxxxxxxxxxxxxxx>, xfs@xxxxxxxxxxx, > linux-fsdevel@xxxxxxxxxxxxxxx, linux-kernel@xxxxxxxxxxxxxxx, > 'Ashish Sangwan' <a.sangwan@xxxxxxxxxxx> > Subject: RE: [PATCH v2 0/10] fs: Introduce FALLOC_FL_INSERT_RANGE for > fallocate > > > > > Date: Thu, 08 May 2014 19:23:19 +0900 > > > From: Namjae Jeon <namjae.jeon@xxxxxxxxxxx> > > > To: Dave Chinner <david@xxxxxxxxxxxxx>, Theodore Ts'o <tytso@xxxxxxx> > > > Cc: linux-ext4 <linux-ext4@xxxxxxxxxxxxxxx>, xfs@xxxxxxxxxxx, > > > linux-fsdevel@xxxxxxxxxxxxxxx, linux-kernel@xxxxxxxxxxxxxxx, > > > Ashish Sangwan <a.sangwan@xxxxxxxxxxx> > > > Subject: [PATCH v2 0/10] fs: Introduce FALLOC_FL_INSERT_RANGE for fallocate > > > > > > In continuation of the work of making the process of non linear editing of > > > media files faster, we introduce here the new flag FALLOC_FL_INSERT_RANGE > > > for fallocate. > > > > > > This flag will work opposite to the newly added FALLOC_FL_COLLAPSE_RANGE flag. > > > As such, specifying FALLOC_FL_INSERT_RANGE flag will insert zeroed-out space > > > in between the file within the range specified by offset and len. User can > > > write new data in this space. e.g. ads. > > > Like collapse range, currently we have the limitation that offset and len should > > > be block size aligned for both XFS and Ext4. > > > > > > The semantics of the flag are : > > > 1) It allocates new zeroed out on disk space of len bytes starting > > > at offset byte without overwriting any existing data. All the data blocks > > > from offset to EOF are shifted towards right to make space for inserting > > > new blocks > > > > Hi, > > > > this sounds a little bit weird to me. I understand the reason for > > this, but this is effectively two operations masking as one. We > > shift the existing data and then we allocate unwritten extents for > > the hole we've created. > > > > What would make more sense to me is to implement only the first > > operation - the shift. And then let the user to allocate unwritten > > extents for the hole using simple fallocate. > > > > The reason is that if you succeed the first part and then fail the > > second due to ENOSPC or any other reason the file will end up in > > undefined state unnecessarily. Yes in your current implementation > > it seems that you'll always end up with the hole in the file and the > > rest is properly shifted, but that may vary from file system to file > > system. Some might choose to roll back the shift, some might not. > > > > If FALLOC_FL_INSERT_RANGE (or any name you wish to choose) would > > just simply shift the extents then you'll get rid of this and the > > only thing that user needs to do (if he chooses to) is to use > > fallocate for the hole created by the shift. If it fails, then > > well, he can try again without any consequences. As a bonus you get > > the possibility to leave the hole in the file which might be useful > > as well. > > > > With current behaviour this might get very confusing very quickly. > > > > What do you and others think ? > Hi Lukas. > Insert range inherently means inserting a real range of space into > the file to provide guaranteed space with user in the inserted area > so that further writes should not fail with an -ENOSPC at least. > If insert range cannot gurantees the above semantics, It should > return error to user space. So what will happen when there is not enough space when "inserting a range" ? And how should user proceed from there ? > > If insert range has been performed on a file, It will reserve space > that write never fail in the inserted area, > In case of full partition or small available size than the range > user want, It seems odd just only left inserting a hole in the middle > of file and return success to user when no one can really write to > this hole. There is a fallocate for allocation, so as I already said you can shift the extents to make a hole in the file and then use fallocate to allocate space for it and you'll get the same result. You are basically doing that now as well, but when the allocation fails the whole "insert range" ioct fails, however the extents are already shifter and there is already a holi in the file so freeing some space and running this ioctl again will not help you at all. While if you fail a fallocate, you can free some space and run it again without any problems. The result will be as expected. What I am arguing about is basically that your insert range ioct is masking two operations as one. Why not to make it transparent and split it into "shift extents" and fallocate ? Then there is a question about the name because it's no longer "insert range" but rather "insert hole" which I think is better and arguably more useful semantic. Thanks! -Lukas > > Thanks! > > > > Thanks! > > -LUkas > > > > > > > 2) It should be used exclusively. No other fallocate flag in combination. > > > 3) Offset and length supplied to fallocate should be fs block size aligned > > > in case of xfs and ext4. > > > 4) Insert range does not work for the case when offset is overlapping/beyond > > > i_size. If the user wants to allocate space at the end of file they are > > > advised to use either ftruncate(2) or fallocate(2) with mode 0. > > > 5) It increses the size of file by len bytes. > > > > > > > > > Namjae Jeon (10): > > > fs: Add support FALLOC_FL_INSERT_RANGE for fallocate > > > xfs: Add support FALLOC_FL_INSERT_RANGE for fallocate > > > ext4: Add support FALLOC_FL_INSERT_RANGE for fallocate > > > xfsprogs: xfs_io: add finsert command for insert range via fallocate > > > xfstests: generic/027: Standard insert range tests > > > xfstests: generic/028: Delayed allocation insert range > > > xfstests: generic/029: Multi insert range tests > > > xfstests: generic/030: Delayed allocation multi insert > > > xfstests: fsstress: Add fallocate insert range operation > > > xfstests: fsx: Add fallocate insert range operation > > > > > > > >
