On Wed, Jul 23, 2025 at 2:20 PM Darrick J. Wong djwong@kernel.org wrote:
On Wed, Jul 23, 2025 at 11:42:42AM -0700, Joanne Koong wrote:
On Wed, Jul 23, 2025 at 7:46 AM Darrick J. Wong djwong@kernel.org wrote:
[cc Joanne]
On Wed, Jul 23, 2025 at 05:14:28PM +0530, Naresh Kamboju wrote:
Regressions found while running LTP msync04 tests on qemu-arm64 running Linux next-20250721, next-20250722 and next-20250723 with 16K and 64K page size enabled builds.
CONFIG_ARM64_64K_PAGES=y ( kernel warning as below ) CONFIG_ARM64_16K_PAGES=y ( kernel warning as below )
No warning noticed with 4K page size. CONFIG_ARM64_4K_PAGES=y works as expected
You might want to cc Joanne since she's been working on large folio support in fuse.
First seen on the tag next-20250721. Good: next-20250718 Bad: next-20250721 to next-20250723
Thanks for the report. Is there a link to the script that mounts the fuse server for these tests? I'm curious whether this was mounted as a fuseblk filesystem.
Regression Analysis:
- New regression? Yes
- Reproducibility? Yes
Test regression: next-20250721 arm64 16K and 64K page size WARNING fs fuse file.c at fuse_iomap_writeback_range
Reported-by: Linux Kernel Functional Testing lkft@linaro.org
## Test log ------------[ cut here ]------------ [ 343.828105] WARNING: fs/fuse/file.c:2146 at fuse_iomap_writeback_range+0x478/0x558 [fuse], CPU#0: msync04/4190
WARN_ON_ONCE(len & (PAGE_SIZE - 1));/me speculates that this might be triggered by an attempt to write back some 4k fsblock within a 16/64k base page?
I think this can happen on 4k base pages as well actually. On the iomap side, the length passed is always block-aligned and in fuse, we set blkbits to be PAGE_SHIFT so theoretically block-aligned is always page-aligned, but I missed that if it's a "fuseblk" filesystem, that isn't true and the blocksize is initialized to a default size of 512 or whatever block size is passed in when it's mounted.
<nod> I think you're correct.
I'll send out a patch to remove this line. It doesn't make any difference for fuse_iomap_writeback_range() logic whether len is page-aligned or not; I had added it as a sanity-check against sketchy ranges.
Also, I just noticed that apparently the blocksize can change dynamically for an inode in fuse through getattr replies from the server (see fuse_change_attributes_common()). This is a problem since the iomap uses inode->i_blkbits for reading/writing to the bitmap. I think we will have to cache the inode blkbits in the iomap_folio_state struct unfortunately :( I'll think about this some more and send out a patch for this.
From my understanding of the iomap code, it's possible to do that if you flush and unmap the entire pagecache (whilst holding i_rwsem and mmap_invalidate_lock) before you change i_blkbits. Nobody *does* this so I have no idea if it actually works, however. Note that even I don't implement the flush and unmap bit; I just scream loudly and do nothing:
lol! i wish I could scream loudly and do nothing too for my case.
AFAICT, I think I just need to flush and unmap that file and can leave the rest of the files/folios in the pagecache as is? But then if the file has active refcounts on it or has been pinned into memory, can I still unmap and remove it from the page cache? I see the invalidate_inode_pages2() function but my understanding is that the page still stays in the cache if it has has active references, and if the page gets mmaped and there's a page fault on it, it'll end up using the preexisting old page in the page cache.
I don't think I really need to have it removed from the page cache so much as just have the ifs state for all the folios in the file freed (after flushing the file) so that it can start over with a new ifs. Ideally we could just flush the file, then iterate through all the folios in the mapping in order of ascending index, and kfree their ->private, but I'm not seeing how we can prevent the case of new writes / a new ifs getting allocated for folios at previous indexes while we're trying to do the iteration/kfreeing.
void fuse_iomap_set_i_blkbits(struct inode *inode, u8 new_blkbits) { trace_fuse_iomap_set_i_blkbits(inode, new_blkbits);
if (inode->i_blkbits == new_blkbits) return; if (!S_ISREG(inode->i_mode)) goto set_it; /* * iomap attaches per-block state to each folio, so we cannot allow * the file block size to change if there's anything in the page cache. * In theory, fuse servers should never be doing this. */ if (inode->i_mapping->nrpages > 0) { WARN_ON(inode->i_blkbits != new_blkbits && inode->i_mapping->nrpages > 0); return; }set_it: inode->i_blkbits = new_blkbits; }
https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/commit/...
--D
Thanks, Joanne