Hi! On Mon 09-12-19 13:54:24, Kirill Tkhai wrote: > on one of production nodes I observe the situation, when many fragmented files > never become defragmented, becase of they have "a" extended attribute. > The reason is append-only file can't be open for write without O_APPEND attribute: > > $lsattr a.txt > -----a--------e----- a.txt > > $strace e4defrag a.txt > openat(AT_FDCWD, "a.txt", O_RDWR) = -1 EPERM (Operation not permitted) > > Simple O_APPEND passed to open() solves the situation. > > The question is: can't we just do this? > > Let's observe the file restrictions we may have. > > "Append-only" extended attribute restriction is weaker, than RO file permissions (0444). > But RO files are being processed by e4defrag, since e4defrag runs by root, and it easily > ignores RO file permissions, while "append-only" files are always ignored by the util. > Is there a fundamental reason we must skip them? I don't think there's a technical reason why we cannot defragment inodes with 'append-only' or even 'immutable' attribute. Just e4defrag would have to remove these flags so that open can succeed (because neither append-only nor immutable flags are overridden by CAP_DAC_OVERRIDE capability - unlike standard unix permissions or xattrs) and then restore the flags after the fact. Whether we really want to do this is another question. I don't think e4defrag should touch immutable files, for append-only files I don't have a strong opinion... Honza -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR
