On 2018/8/24 下午4:09, Misono Tomohiro wrote: [snip] >> >> BTW, what's the possibility of such problem in your test environment? > > It's like one in several times. > It may depend on hardware performance? (the machine is not so fast), > > I also noticed following warning happens too (not always): >
After digging into the case, it's more complex than just my patch. Firstly, we lacks a lot of underflow check when modifying bytes_may_use. So we need to do all the underflow detection for every modifier of bytes_may_use. Secondly, btrfs_cross_ref_exist() check makes NODATACOW check in __btrfs_buffered_write() unreliable. For the following case, at __btrfs_buffered_write() time we're pretty sure we could do NODATACOW, but at sync time, due to cloned range, btrfs_cross_ref_exist() would detect reflinked prealloc extent, then falls back to CoW, and finally cause bytes_may_use underflow: --- mkfs.btrfs -f $dev > $full_log mount $dev $mnt -o nospace_cache btrfs quota enable $mnt btrfs quota rescan -w $mnt xfs_io -f -c "falloc 0 2M" $mnt/file1 > /dev/null xfs_io -c "pwrite -b 1M 0 1M" $mnt/file1 > /dev/null xfs_io -c "reflink $mnt/file1 1M 4M 1M" $mnt/file1 > /dev/null sync ---- Even without my patch, the "pwrite" command is still CoWed, which could be avoided. And that's the reason my patch is causing the underflow. To fix this, we need more accurate btrfs_cross_ref_exist() check, not only for @disk_bytenr but also check @len. Or we could try to flush the whole inode in clone_range() so we could go through NOCOW routine before clone really happens. Thanks, Qu
signature.asc
Description: OpenPGP digital signature
