On Fri, Dec 18, 2015 at 01:16:14AM +0100, Christoph Anton Mitterer wrote: > > It would be nice if filefrag supports btrfs compression. > Right now, it seems to assume that each 128KiB compression "block" > is one extent, though, AFAIU discussion on linux-btrfs, > it's in realliy one extent.
How, exactly, do you propose that filefrag understand this? It is getting the information from the fiemap ioctl: https://www.kernel.org/doc/Documentation/filesystems/fiemap.txt and the problem is that fiemap returns the logical length of the extent, but not the *physical* length of the extent. So this is why the filefrag is show what appears to be overlapping extents: > ext: logical_offset: physical_offset: length: expected: flags: > 0: 0.. 31: 3072.. 3103: 32: encoded > 1: 32.. 63: 3078.. 3109: 32: 3104: encoded Presumably, the first extent for logical blocks 0 -- 31 is taking only 6 blocks --- but there's no way filefrag can know that, not having any magical or psychic abilities. It could be that that the blocks only take 5 blocks, and there is a one block "hole" meaning there was true fragmentation --- but there's no way to tell that, and having filefrag send e-mail to start a discussion on linux-btrfs is obviously not something that is going to work given the current state of Artificial Intelligence. :-) We could display: ext: logical_offset: physical_offset: length: expected: flags: 0: 0.. 31: 3072.. ????: 32: encoded 1: 32.. 63: 3078.. ????: 32: 3104: encoded ... ... since the encoded flag means that it is possible that physical length is^H^H^H might be different from the logical length. But even that is not guaranteed; it could be that the blocks are encrypted, and the physical length and logical lengths of the extent are the same --- all FM_ENCODED means is that it is not safe to try to read the data blocks directly if you expect to get the actual data. (One of the uses of fiemap is for bootloaders who want to store the block list so they can boot the system without understanding the low-level file system.) So I'm afraid there's not much I can do here. My suggestion is that you kick off a discussion on linux-btrfs about adding a new, expanded ioctl to the kernel that provides a superset of the FIEMAP ioctl, which also returns the physical length. Best regards, - Ted
