-
Bug
-
Resolution: Unresolved
-
Minor
-
None
-
None
-
None
-
3
-
9223372036854775807
lfs changelog --user NAME --mask creat returns every record type, not just
CREAT, when NAME was registered without a mask. No error is reported. The
request is a wrong answer rather than a failure, which is what makes it worth
filing.
A zero cf_mask means two different things, and the client composes them as
if it meant one.
Meaning one, on the server: "this user registered no mask of their own".
mdd_chlg_usermask() returns cur_mask for a CHANGELOG_USER_REC2 and
0 for the older record (lustre/mdd/mdd_device.c:233), and the lookup puts
that in the reply (lustre/mdd/mdd_changelog.c:53).
Meaning two, on the client: "do not filter". A zero
crs_user_mask disables the test altogether
(lustre/mdc/mdc_changelog.c:222-224):
/* Check if this record type matches the user's mask */ if (crs->crs_user_mask && !(crs->crs_user_mask & BIT(rec->cr.cr_type))) RETURN(0);
The two meanings meet in chlg_ioctl()
(lustre/mdc/mdc_changelog.c:811-814), where in is the mask the caller
asked for and out is the mask the user registered:
if (in.cf_mask == 0) crs->crs_user_mask = out.cf_mask; else crs->crs_user_mask = in.cf_mask & out.cf_mask;
So when the caller passes a mask and the user has none, the result is
in.cf_mask & 0, which is 0, which the record test then reads as *do not
filter*. The caller's mask is dropped and everything comes back.
Which users have no mask
Registration writes a rec2 when either a mask or a name was given
(lustre/mdd/mdd_device.c:1789), but cur_mask is only assigned inside
if (mask) (lustre/mdd/mdd_device.c:1817-1826). So a registration with a
name and no mask produces a rec2 whose cur_mask is 0, and that user
hits this today:
lctl --device testfs-MDT0000 changelog register --user foo touch /mnt/testfs/a ; mkdir /mnt/testfs/d lfs changelog --user foo --mask creat testfs-MDT0000 -> the MKDIR record is listed too
A user registered with neither a mask nor a name gets the older record type,
whose mask is 0 by definition, so every such user is affected as well. Those
users cannot be looked up at all until LU-20647 is fixed, which is the only
reason this is not visible for them yet.
The fix
Either give "no per-user mask" a value that is not zero, or stop treating the
two zeros as the same thing when composing. The smaller change is a third arm:
if (in.cf_mask == 0) crs->crs_user_mask = out.cf_mask; else if (out.cf_mask == 0) crs->crs_user_mask = in.cf_mask; else crs->crs_user_mask = in.cf_mask & out.cf_mask;
A user with no registered mask is not restricted to nothing; they are
unrestricted, so intersecting with them should leave the caller's mask alone.
How it was found
Reading the reply path while fixing LU-20647, which makes a plainly registered
user reachable by lookup for the first time and so would have exposed this to
every such user.