on 12/20/2024 9:10 AM, Andreas Dilger wrote: > On Dec 19, 2024, at 4:00 AM, Kemeng Shi <shikemeng@xxxxxxxxxxxxxxx> wrote: >> >> As high four bits of block in dx_entry is not used by any feature for now, we can remove unneeded bits mask in dx_get_block() and add it back >> when it's really needed. > > Actually, the opposite is true. This mask protects the *CURRENT* code > from any future use for these bits, so removing it now means that they > could never be used in the future, since the block number would be > taken as all 32 bits instead of only the bottom 28 bits. I don't think > we are in any danger of having a 16TB single directory any time soon. > > However, the top bits were intended to store a "fullness" for the index > blocks, to optimize online directory shrinking without having to scan > each of the blocks for how many entries are currently in the block. > This would allow the dirent removal to easily see "this block and the > previous/next block are only 1/3 full and could be merged". > > See the following thread for a prototype patch and discussion on this: > https://patchwork.ozlabs.org/project/linux-ext4/patch/20190821182740.97127-1-harshadshirwadkar@xxxxxxxxx/ > > I think removing this mask has a negative effect on future usefulness, > and virtually no benefit to the code today, so I would object to landing > it. Sure, it makes sense to reserve bit for future use if it will likely be used. But I wonder would it be better to catch using high four bits in in ext4_append() in which case we could forbit using high four bits in time rather than lost dir when reserved bits are really used in future. This also reduce cpu cost as dx_get_block() is used likely more frequent than ext4_append(). Just a thought. Thanks, Kemeng
