On Tue, Aug 12, 2025 at 07:44:09AM +0930, Qu Wenruo wrote: > > > 在 2025/8/12 01:19, Darrick J. Wong 写道: > > On Sun, Aug 10, 2025 at 07:36:48AM +0930, Qu Wenruo wrote: > > > > > > > > > 在 2025/8/9 18:39, Zhang Yi 写道: > > > > On 2025/8/9 6:11, Qu Wenruo wrote: > > > > > 在 2025/8/8 21:46, Theodore Ts'o 写道: > > > > > > On Fri, Aug 08, 2025 at 06:20:56PM +0930, Qu Wenruo wrote: > > > > > > > > > > > > > > 在 2025/8/8 17:22, Qu Wenruo 写道: > > > > > > > > Hi, > > > > > > > > > > > > > > > > [BACKGROUND] > > > > > > > > Recently I'm testing btrfs with 16KiB block size. > > > > > > > > > > > > > > > > Currently btrfs is artificially limiting subpage block size to 4K. > > > > > > > > But there is a simple patch to change it to support all block sizes <= > > > > > > > > page size in my branch: > > > > > > > > > > > > > > > > https://github.com/adam900710/linux/tree/larger_bs_support > > > > > > > > > > > > > > > > [IOMAP WARNING] > > > > > > > > And I'm running into a very weird kernel warning at btrfs/136, with 16K > > > > > > > > block size and 64K page size. > > > > > > > > > > > > > > > > The problem is, the problem happens with ext3 (using ext4 modeule) with > > > > > > > > 16K block size, and no btrfs is involved yet. > > > > > > > > > > > > > > > > > > Thanks for the bug report! This looks like it's an issue with using > > > > > > indirect block-mapped file with a 16k block size. I tried your > > > > > > reproducer using a 1k block size on an x86_64 system, which is how I > > > > > > test problem caused by the block size < page size. It didn't > > > > > > reproduce there, so it looks like it really needs a 16k block size. > > > > > > > > > > > > Can you say something about what system were you running your testing > > > > > > on --- was it an arm64 system, or a powerpc 64 system (the two most > > > > > > common systems with page size > 4k)? (I assume you're not trying to > > > > > > do this on an Itanic. :-) And was the page size 16k or 64k? > > > > > > > > > > The architecture is aarch64, the host board is Rock5B (cheap and fast enough), the test machine is a VM on that board, with ovmf as the UEFI firmware. > > > > > > > > > > The kernel is configured to use 64K page size, the *ext3* system is using 16K block size. > > > > > > > > > > Currently I tried the following combination with 64K page size and ext3, the result looks like the following > > > > > > > > > > - 2K block size > > > > > - 4K block size > > > > > All fine > > > > > > > > > > - 8K block size > > > > > - 16K block size > > > > > All the same kernel warning and never ending fsstress > > > > > > > > > > - 32K block size > > > > > - 64K block size > > > > > All fine > > > > > > > > > > I am surprised as you that, not all subpage block size are having problems, just 2 of the less common combinations failed. > > > > > > > > > > And the most common ones (4K, page size) are all fine. > > > > > > > > > > Finally, if using ext4 not ext3, all combinations above are fine again. > > > > > > > > > > So I ran out of ideas why only 2 block sizes fail here... > > > > > > > > > > > > > This issue is caused by an overflow in the calculation of the hole's > > > > length on the forth-level depth for non-extent inodes. For a file system > > > > with a 4KB block size, the calculation will not overflow. For a 64KB > > > > block size, the queried position will not reach the fourth level, so this > > > > issue only occur on the filesystem with a 8KB and 16KB block size. > > > > > > > > Hi, Wenruo, could you try the following fix? > > > > > > > > diff --git a/fs/ext4/indirect.c b/fs/ext4/indirect.c > > > > index 7de327fa7b1c..d45124318200 100644 > > > > --- a/fs/ext4/indirect.c > > > > +++ b/fs/ext4/indirect.c > > > > @@ -539,7 +539,7 @@ int ext4_ind_map_blocks(handle_t *handle, struct inode *inode, > > > > int indirect_blks; > > > > int blocks_to_boundary = 0; > > > > int depth; > > > > - int count = 0; > > > > + u64 count = 0; > > > > ext4_fsblk_t first_block = 0; > > > > > > > > trace_ext4_ind_map_blocks_enter(inode, map->m_lblk, map->m_len, flags); > > > > @@ -588,7 +588,7 @@ int ext4_ind_map_blocks(handle_t *handle, struct inode *inode, > > > > count++; > > > > /* Fill in size of a hole we found */ > > > > map->m_pblk = 0; > > > > - map->m_len = min_t(unsigned int, map->m_len, count); > > > > + map->m_len = umin(map->m_len, count); > > > > goto cleanup; > > > > } > > > > > > It indeed solves the problem. > > > > > > Tested-by: Qu Wenruo <wqu@xxxxxxxx> > > > > Can we get the relevant chunks of this test turned into a tests/ext4/ > > fstest so that the ext4 developers have a regression test that doesn't > > require setting up btrfs, please? > > Sure, although I can send out a ext4 specific test case for it, I'm > definitely not the best one to explain why the problem happens. > > Thus I believe Zhang Yi would be the best one to send the test case. > > > > Another thing is, any ext3 run with 16K block size (that's if the system > supports it) should trigger it with the existing test cases. > > The biggest challenge is to get a system supporting 16k bs (aka page size >= > 16K), so it has a high chance that for most people the new test case will > mostly be NOTRUN. I'm curious to try out fuse2fs against whatever test gets written, since it supports large fsblock sizes. --D > Thanks, > Qu > > > > > --D > > > > > Thanks, > > > Qu > > > > > > > Thanks, > > > > Yi. > > > > > > > > > > > >
