On Fri, Dec 15, 2017 at 8:42 AM, zhangyi (F) <yi.zhang@xxxxxxxxxx> wrote: > On 2017/12/14 21:44, Amir Goldstein Wrote: >> On Thu, Dec 14, 2017 at 8:48 AM, zhangyi (F) <yi.zhang@xxxxxxxxxx> wrote: [...] >> Please test the Valid "rename exchange redirect" case: >> set_redirect $upperdir/testdir1 "testdir2" >> set_redirect $upperdir/testdir2 "testdir1" >> fsck should not remove those redirects. >> > I can add this case. I was wondering why add this special one? If I understand right, > these two xattrs will not remove after fsck if $lowerdir/testdir1 and $lowerdir/testdir2 > exists, just like the second test case "Test valid redirect xattr point to a directory > origin in the same directory". > IIUC, the first version of fsck you posted wanted to make sure that origin is covered with either a whiteout or opaque dir and I commented that it is also legal to be covered by a redirected dir. So the test of exchanged redirect dirs checks that this condition is handled correctly by fsck. >>> +# Test missing opaque xattr with valid redirect directory, should fix opaque >> >> Why does fsck prefer redirect over non-redirect? >> Why is this not a private case of duplicate redirect? >> > Sorry, I not follow. > This test, I simulate the case of the following: > 1) mkdir -p lower/testdir upper/testdir > 2) mount overlay > 3) mv merge/testdir merge/dir2 (will create an whiteout 'testdir' in upper) > 4) mkdir merge/testdir (create opaque dir) > 5) umount overlay > 6) remove opaque xattr of upper/testdir (missing opaque xattr lead to lower/testdir/* expose) > > Am I miss something? That is one way of getting to that inconsistency which suggests that the correct and automatic fix is to set opaque on uper/testdir, but what about: 1) extract pre-defined image to lower/testdir upper/testdir 2) mount overlay 3) mv merge/testdir merge/dir2 (will create an whiteout 'testdir' in upper) 4) mkdir merge/testdir (create opaque dir) 5) umount overlay 6) re-extract pre-defined image to lower/testdir upper/testdir (*) (*) Anything that *can* happen *will* happen... Note that prior to redirect_dir feature, use could even get a way with this partial restore/undo of testdir changes. My point is that if you have 2 dirs pointing at the same lower, either because they are both redirect or because one is not opaque, it is not possible to automatically determine which should be the merge dir and need to ask user to resolve the conflict. > >>> +echo "+ Missing opaque" >>> +mkdir $lowerdir/origin >>> +mkdir $upperdir/origin >>> +mkdir $upperdir/testdir >>> +set_redirect $upperdir/testdir "origin" >>> +_overlay_fsck_dirs -a $lowerdir $upperdir $workdir >> $seqres.full 2>&1 || \ >>> + _fail "fsck should not fail" >>> +check_xattr $OVL_REDIRECT_XATTR "origin" $upperdir/testdir >>> +check_xattr $OVL_OPAQUE_XATTR $OVL_OPAQUE_XATTR_VAL $upperdir/origin >>> +rm -rf $lowerdir/* $upperdir/* >>> + >>> +# Test duplicate redirect directories point to one origin, should fail in >>> +# -a 'auto' mode, should remove the duplicate one in -y 'yes' mode >>> +echo "+ Duplicate redirect" >>> +mkdir $lowerdir/origin >>> +mknod $upperdir/origin c 0 0 >>> +mkdir $upperdir/testdir1 >>> +mkdir $upperdir/testdir2 >>> +set_redirect $upperdir/testdir1 "origin" >>> +set_redirect $upperdir/testdir2 "origin" >>> +_overlay_fsck_dirs -a $lowerdir $upperdir $workdir >> $seqres.full 2>&1 && \ >>> + _fail "fsck should fail" >>> +_overlay_fsck_dirs -y $lowerdir $upperdir $workdir >> $seqres.full 2>&1 || \ >>> + _fail "fsck should not fail" >>> +check_xattr $OVL_REDIRECT_XATTR "origin" $upperdir/testdir2 >>> + >> >> So what guaranty you have that fsck clears testdir1 and not testdir2? >> Maybe check there is exactly one redirect left after fsck? >> What does fsck do with -y to the duplicate dir anyway? >> If copying from e2fsck behavior for duplicate referenced block, should >> copy the content of lower dirs to upper dir and make it opaque, but that can >> get nasty for a nested duplicate tree... OTOH, if user got to duplicate >> redirect dir by cp -a of a redirected upper dir, then a full copy of the tree >> is what user intended for. >> > Let me see, now, fsck will remove the *later redirect xattr in 'yes' mode, > which means that we guess user want to copy the tree in upperdir if they > call cp -a when overlayfs is offline, not the whole tree expose in merge > layer. I think if we need to copy whole tree include lower dirs maybe too > complicated. > > *) fsck just remove the later one when scaning, I understand it maybe cannot guaranty > fsck will clears testdir1 and not testdir2 for some base filesystem, so need to find > better way, thoughts? > quite simple I think. Instead of verifying that redirect was removed from testdir1 and remained on testdir2, verify that redirect from removed from either and left on exactly one of them. Less trivial scripting, but not impossible. Amir. -- To unsubscribe from this list: send the line "unsubscribe linux-unionfs" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
