On Wed, 2018-10-24 at 14:24 +0800, Yan, Zheng wrote:
On Mon, Oct 15, 2018 at 9:05 PM Jeff Layton jlayton@redhat.com wrote:
On Mon, 2018-10-15 at 20:37 +0800, Yan, Zheng wrote:
On Oct 15, 2018, at 19:29, Jeff Layton jlayton@redhat.com wrote:
On Mon, 2018-10-15 at 09:58 +0800, Yan, Zheng wrote:
On Oct 12, 2018, at 18:56, Jeff Layton jlayton@redhat.com wrote:
On Fri, 2018-10-12 at 10:39 +0800, Yan, Zheng wrote: > > On Oct 11, 2018, at 22:49, Jeff Layton jlayton@redhat.com wrote: > > > > On Fri, 2018-09-28 at 17:43 +0800, Yan, Zheng wrote: > > > This reverts commit 8b8f53af1ed9df88a4c0fbfdf3db58f62060edf3. > > > > > > splice_dentry() is used by three places. For two places, req->r_dentry > > > is passed to splice_dentry(). In the case of error, req->r_dentry does > > > not get updated. So splice_dentry() should not drop reference. > > > > > > Cc: stable@vger.kernel.org #4.18 > > > Signed-off-by: "Yan, Zheng" zyan@redhat.com > > > --- > > > fs/ceph/inode.c | 1 - > > > 1 file changed, 1 deletion(-) > > > > > > diff --git a/fs/ceph/inode.c b/fs/ceph/inode.c > > > index c6bbb7aa99e4..375924b2bc86 100644 > > > --- a/fs/ceph/inode.c > > > +++ b/fs/ceph/inode.c > > > @@ -1140,7 +1140,6 @@ static struct dentry *splice_dentry(struct dentry *dn, struct inode *in) > > > if (IS_ERR(realdn)) { > > > pr_err("splice_dentry error %ld %p inode %p ino %llx.%llx\n", > > > PTR_ERR(realdn), dn, in, ceph_vinop(in)); > > > - dput(dn); > > > dn = realdn; /* note realdn contains the error */ > > > goto out; > > > } else if (realdn) { > > > > This might be ok, buI have some real concerns about splice_dentry and > > its callers -- particularly ceph_fill_trace: > > > > We hold a reference to dn on entry to splice_dentry. We then call > > d_splice_alias and get back an error, and now we don't put the old > > dentry. > > > > Fine -- we have to then expect the caller to do it. Unfortunately, the > > callers in ceph_fill_trace do this: > > > > dn = splice_dentry(dn, in); > > if (IS_ERR(dn)) { > > err = PTR_ERR(dn); > > goto done; > > } > > req->r_dentry = dn; /* may have spliced */ > > > > The old value of dn gets clobbered once that comes back with an ERR_PTR. > > I guess we could claim that r_dentry will still be set to the old value > > at that point and that it would get cleaned up when it gets cleaned up. > > > > But...I see this higher up in ceph_fill_trace at the end of the > > CEPH_MFS_OP_RENAME condition block: > > > > dn = req->r_old_dentry; /* use old_dentry */ > > > > So now I'm worried about the case where the splice succeeds. ISTM that > > "dn" can represent either r_dentry or r_old_dentry at the point where > > splice_dentry gets called, but we only ever reset the value of r_dentry > > there. > > > > If dn == r_old_dentry at the time that splice_dentry is called, and then > > that succeeds, we'll end up leaking the reference to r_dentry and then > > doing an overput on r_old_dentry. > > > > Good catch > > > I think it might help to establish clear "ownership" of the dentry > > references throughout that function. Consider zeroing out r_dentry and > > r_old_dentry at the time that you set the local variables? That might > > make this whole thing less fragile. > > Before the function return, we need to set req->r_dentry again. This will introduce duplicated code because there are several ‘goto’ in the function. How about following change. > > > - dn = req->r_old_dentry; /* use old_dentry */ > + /* swap r_dentry and r_old_dentry */ > + req->r_dentry = req->r_old_dentry; > + req->r_old_dentry = dn; > + dn = req->r_dentry; >
Why do we need to reset r_dentry? I don't see where it gets used after ceph_fill_trace returns, so I gather we're just resetting it to clean up the references? Your proposed fix looks like it would fix that bug. It will mean that the values in the req will change but maybe that's ok.
It’s used by ceph_finish_lookup
In that case, it might be bad idea to overwrite the value of r_dentry with the r_old_dentry pointer?
ceph_finish_lookup() is not used in the rename case ;)
Ok, but still...I worry about this sort of special-casing, as it can lead to people making incorrect assumptions later. I think it would be best to just take a dentry ref for "dn" and just put it before it goes out of scope.
how about below patch. let splice_dentry() to update req->r_dentry.
diff --git a/fs/ceph/inode.c b/fs/ceph/inode.c index e8ce4e66bc47..739737f4e0e1 100644 --- a/fs/ceph/inode.c +++ b/fs/ceph/inode.c @@ -1106,8 +1106,9 @@ static void update_dentry_lease(struct dentry *dentry,
- splice a dentry to an inode.
- caller must hold directory i_mutex for this to be safe.
*/ -static struct dentry *splice_dentry(struct dentry *dn, struct inode *in) +static int splice_dentry(struct dentry **pdn, struct inode *in) {
- struct dentry *dn = *pdn; struct dentry *realdn;
BUG_ON(d_inode(dn)); @@ -1140,28 +1141,23 @@ static struct dentry *splice_dentry(struct dentry *dn, struct inode *in) if (IS_ERR(realdn)) { pr_err("splice_dentry error %ld %p inode %p ino %llx.%llx\n", PTR_ERR(realdn), dn, in, ceph_vinop(in));
dn = realdn;/** Caller should release 'dn' in the case of error.* If 'req->r_dentry' is passed to this function,* caller should leave 'req->r_dentry' untouched.*/goto out;- } else if (realdn) {
return PTR_ERR(realdn);- }
- if (realdn) { dout("dn %p (%d) spliced with %p (%d) " "inode %p ino %llx.%llx\n", dn, d_count(dn), realdn, d_count(realdn), d_inode(realdn), ceph_vinop(d_inode(realdn))); dput(dn);
dn = realdn;
*pdn = realdn;-out:
- return dn;
- return 0;
} /* @@ -1348,7 +1344,10 @@ int ceph_fill_trace(struct super_block *sb, struct ceph_mds_request *req) dout("dn %p gets new offset %lld\n", req->r_old_dentry, ceph_dentry(req->r_old_dentry)->offset);
dn = req->r_old_dentry; /* use old_dentry */
/* swap r_dentry and r_old_dentry */req->r_dentry = req->r_old_dentry;req->r_old_dentry = dn;dn = req->r_dentry;
I think the above deserves a comment explaining why this is safe to do for rename ops.
}/* null dentry? */ @@ -1373,12 +1372,10 @@ int ceph_fill_trace(struct super_block *sb, struct ceph_mds_request *req) if (d_really_is_negative(dn)) { ceph_dir_clear_ordered(dir); ihold(in);
dn = splice_dentry(dn, in);if (IS_ERR(dn)) {err = PTR_ERR(dn);
err = splice_dentry(&req->r_dentry, in);if (err < 0) goto done;
}req->r_dentry = dn; /* may have spliced */
dn = req->r_dentry; /* may have spliced */@@ -1398,22 +1395,18 @@ int ceph_fill_trace(struct super_block *sb, struct ceph_mds_request *req) } else if ((req->r_op == CEPH_MDS_OP_LOOKUPSNAP || req->r_op == CEPH_MDS_OP_MKSNAP) && !test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
struct dentry *dn = req->r_dentry;/* fill out a snapdir LOOKUPSNAP dentry */
BUG_ON(!dn);dout(" linking snapped dir %p to dn %p\n", in, dn);
BUG_ON(!req->r_dentry);dout(" linking snapped dir %p to dn %p\n", in, req->r_dentry);
dn = splice_dentry(dn, in);if (IS_ERR(dn)) {err = PTR_ERR(dn);
err = splice_dentry(&req->r_dentry, in);if (err < 0) goto done;
}req->r_dentry = dn; /* may have spliced */@@ -1677,8 +1670,6 @@ int ceph_readdir_prepopulate(struct ceph_mds_request *req, } if (d_really_is_negative(dn)) {
struct dentry *realdn;if (ceph_security_xattr_deadlock(in)) { dout(" skip splicing dn %p to inode %p" " (security xattr deadlock)\n", dn, in);@@ -1687,13 +1678,9 @@ int ceph_readdir_prepopulate(struct ceph_mds_request *req, goto next_item; }
realdn = splice_dentry(dn, in);if (IS_ERR(realdn)) {err = PTR_ERR(realdn);d_drop(dn);
err = splice_dentry(&dn, in);if (err < 0) goto next_item;
}dn = realdn;ceph_dentry(dn)->offset = rde->offset; @@ -1709,8 +1696,7 @@ int ceph_readdir_prepopulate(struct ceph_mds_request *req, err = ret; } next_item:
if (dn)dput(dn);
dput(dn);out: if (err == 0 && skipped == 0) {
The rest seems to be ok.
At the very least, if you are going to override r_dentry like this, please add a comment explaining why it's safe to do this in the rename case.
Another idea might be to just keep an extra reference to "dn", and not reset r_dentry and r_old_dentry? It's a bit more code, but it might be a cleaner fix, long-term.
I will check, Thanks
Yan, Zheng
-- Jeff Layton jlayton@redhat.com
-- Jeff Layton jlayton@redhat.com
-- Jeff Layton jlayton@redhat.com