On Tue, Apr 08, 2025 at 07:51:03PM +0100, Matthew Wilcox wrote: > On Tue, Apr 08, 2025 at 11:02:40AM -0700, Darrick J. Wong wrote: > > On Tue, Apr 08, 2025 at 06:51:14PM +0100, Matthew Wilcox wrote: > > > On Tue, Apr 08, 2025 at 10:48:55AM -0700, Darrick J. Wong wrote: > > > > On Tue, Apr 08, 2025 at 10:24:40AM -0700, Luis Chamberlain wrote: > > > > > On Tue, Apr 8, 2025 at 10:06 AM Luis Chamberlain <mcgrof@xxxxxxxxxx> wrote: > > > > > > Fun > > > > > > puzzle for the community is figuring out *why* oh why did a large folio > > > > > > end up being used on buffer-heads for your use case *without* an LBS > > > > > > device (logical block size) being present, as I assume you didn't have > > > > > > one, ie say a nvme or virtio block device with logical block size > > > > > > > PAGE_SIZE. The area in question would trigger on folio migration *only* > > > > > > if you are migrating large buffer-head folios. We only create those > > > > > > > > > > To be clear, large folios for buffer-heads. > > > > > > if > > > > > > you have an LBS device and are leveraging the block device cache or a > > > > > > filesystem with buffer-heads with LBS (they don't exist yet other than > > > > > > the block device cache). > > > > > > > > My guess is that udev or something tries to read the disk label in > > > > response to some uevent (mkfs, mount, unmount, etc), which creates a > > > > large folio because min_order > 0, and attaches a buffer head. There's > > > > a separate crash report that I'll cc you on. > > > > > > But you said: > > > > > > > the machine is arm64 with 64k basepages and 4k fsblock size: > > > > > > so that shouldn't be using large folios because you should have set the > > > order to 0. Right? Or did you mis-speak and use a 4K PAGE_SIZE kernel > > > with a 64k fsblocksize? > > > > This particular kernel warning is arm64 with 64k base pages and a 4k > > fsblock size, and my suspicion is that udev/libblkid are creating the > > buffer heads or something weird like that. > > > > On x64 with 4k base pages, xfs/032 creates a filesystem with 64k sector > > size and there's an actual kernel crash resulting from a udev worker: > > https://lore.kernel.org/linux-fsdevel/20250408175125.GL6266@frogsfrogsfrogs/T/#u > > > > So I didn't misspeak, I just have two problems. I actually have four > > problems, but the others are loop device behavior changes. > > Right, but this warning only triggers for large folios. So somehow > we've got a multi-page folio in the bdev's page cache. > > Ah. I see. > > block/bdev.c: mapping_set_folio_min_order(BD_INODE(bdev)->i_mapping, > > so we're telling the bdev that it can go up to MAX_PAGECACHE_ORDER. Ah yes silly me that would explain the large folios without LBS devices. > And then we call readahead, which will happily put order-2 folios > in the pagecache because of my bug that we've never bothered fixing. > > We should probably fix that now, but as a temporary measure if > you'd like to put: > > mapping_set_folio_order_range(BD_INODE(bdev)->i_mapping, min, min) > > instead of the mapping_set_folio_min_order(), that would make the bug > no longer appear for you. Agreed. Luis
