Fix a use-after-free on dentry's d_fsdata fid list when a thread
lookups a fid through dentry while another thread unlinks it:
UAF thread:
refcount_t: addition on 0; use-after-free.
p9_fid_get linux/./include/net/9p/client.h:262
v9fs_fid_find+0x236/0x280 linux/fs/9p/fid.c:129
v9fs_fid_lookup_with_uid linux/fs/9p/fid.c:181
v9fs_fid_lookup+0xbf/0xc20 linux/fs/9p/fid.c:314
v9fs_vfs_getattr_dotl+0xf9/0x360 linux/fs/9p/vfs_inode_dotl.c:400
vfs_statx+0xdd/0x4d0 linux/fs/stat.c:248
Freed by:
p9_client_clunk+0xb0/0xe0 linux/net/9p/client.c:1456
p9_fid_put linux/./include/net/9p/client.h:278
v9fs_dentry_release+0xb5/0x140 linux/fs/9p/vfs_dentry.c:55
v9fs_remove+0x38f/0x620 linux/fs/9p/vfs_inode.c:518
vfs_unlink+0x29a/0x810 linux/fs/namei.c:4335
The problem is that d_fsdata was not accessed under d_lock, because
d_release() normally is only called once the dentry is otherwise no
longer accessible but since we also call it explicitly in v9fs_remove
that lock is required:
move the hlist out of the dentry under lock then unref its fids once
they are no longer accessible.
Fixes: 154372e67d40 ("fs/9p: fix create-unlink-getattr idiom")
Cc: [email protected]
Reported-by: Meysam Firouzi
Reported-by: Amirmohammad Eftekhar
Signed-off-by: Dominique Martinet <[email protected]>
---
fs/9p/vfs_dentry.c | 9 +++++++--
1 file changed, 7 insertions(+), 2 deletions(-)
diff --git a/fs/9p/vfs_dentry.c b/fs/9p/vfs_dentry.c
index f16f73581634..01338d4c2d9e 100644
--- a/fs/9p/vfs_dentry.c
+++ b/fs/9p/vfs_dentry.c
@@ -48,12 +48,17 @@ static int v9fs_cached_dentry_delete(const struct dentry *dentry)
static void v9fs_dentry_release(struct dentry *dentry)
{
struct hlist_node *p, *n;
+ struct hlist_head head;
p9_debug(P9_DEBUG_VFS, " dentry: %pd (%p)\n",
dentry, dentry);
- hlist_for_each_safe(p, n, (struct hlist_head *)&dentry->d_fsdata)
+
+ spin_lock(&dentry->d_lock);
+ hlist_move_list((struct hlist_head *)&dentry->d_fsdata, &head);
+ spin_unlock(&dentry->d_lock);
+
+ hlist_for_each_safe(p, n, &head)
p9_fid_put(hlist_entry(p, struct p9_fid, dlist));
- dentry->d_fsdata = NULL;
}
static int v9fs_lookup_revalidate(struct dentry *dentry, unsigned int flags)
--
2.44.0
On Tuesday, May 21, 2024 2:29:46 PM CEST Dominique Martinet wrote:
> Fix a use-after-free on dentry's d_fsdata fid list when a thread
> lookups a fid through dentry while another thread unlinks it:
I guess that's "looks up". :)
> UAF thread:
> refcount_t: addition on 0; use-after-free.
> p9_fid_get linux/./include/net/9p/client.h:262
> v9fs_fid_find+0x236/0x280 linux/fs/9p/fid.c:129
> v9fs_fid_lookup_with_uid linux/fs/9p/fid.c:181
> v9fs_fid_lookup+0xbf/0xc20 linux/fs/9p/fid.c:314
> v9fs_vfs_getattr_dotl+0xf9/0x360 linux/fs/9p/vfs_inode_dotl.c:400
> vfs_statx+0xdd/0x4d0 linux/fs/stat.c:248
>
> Freed by:
> p9_client_clunk+0xb0/0xe0 linux/net/9p/client.c:1456
That line number looks weird.
> p9_fid_put linux/./include/net/9p/client.h:278
> v9fs_dentry_release+0xb5/0x140 linux/fs/9p/vfs_dentry.c:55
> v9fs_remove+0x38f/0x620 linux/fs/9p/vfs_inode.c:518
> vfs_unlink+0x29a/0x810 linux/fs/namei.c:4335
>
> The problem is that d_fsdata was not accessed under d_lock, because
> d_release() normally is only called once the dentry is otherwise no
> longer accessible but since we also call it explicitly in v9fs_remove
> that lock is required:
> move the hlist out of the dentry under lock then unref its fids once
> they are no longer accessible.
>
> Fixes: 154372e67d40 ("fs/9p: fix create-unlink-getattr idiom")
> Cc: [email protected]
> Reported-by: Meysam Firouzi
> Reported-by: Amirmohammad Eftekhar
> Signed-off-by: Dominique Martinet <[email protected]>
> ---
> fs/9p/vfs_dentry.c | 9 +++++++--
> 1 file changed, 7 insertions(+), 2 deletions(-)
>
> diff --git a/fs/9p/vfs_dentry.c b/fs/9p/vfs_dentry.c
> index f16f73581634..01338d4c2d9e 100644
> --- a/fs/9p/vfs_dentry.c
> +++ b/fs/9p/vfs_dentry.c
> @@ -48,12 +48,17 @@ static int v9fs_cached_dentry_delete(const struct dentry *dentry)
> static void v9fs_dentry_release(struct dentry *dentry)
> {
> struct hlist_node *p, *n;
> + struct hlist_head head;
>
> p9_debug(P9_DEBUG_VFS, " dentry: %pd (%p)\n",
> dentry, dentry);
> - hlist_for_each_safe(p, n, (struct hlist_head *)&dentry->d_fsdata)
> +
> + spin_lock(&dentry->d_lock);
> + hlist_move_list((struct hlist_head *)&dentry->d_fsdata, &head);
> + spin_unlock(&dentry->d_lock);
> +
> + hlist_for_each_safe(p, n, &head)
> p9_fid_put(hlist_entry(p, struct p9_fid, dlist));
> - dentry->d_fsdata = NULL;
> }
I'm not sure if that works out. So you are moving the list from dentry to a
local variable. But if you look at v9fs_fid_find() [fs/9p/fid.c#123] it reads
dentry->d_fsdata (twice) and holds it as local variable before taking a
lock. So the lock in v9fs_fid_find() should happen earlier, no?
>
> static int v9fs_lookup_revalidate(struct dentry *dentry, unsigned int flags)
>
Christian Schoenebeck wrote on Wed, May 22, 2024 at 04:35:19PM +0200:
Thanks for the review!
> On Tuesday, May 21, 2024 2:29:46 PM CEST Dominique Martinet wrote:
> > Fix a use-after-free on dentry's d_fsdata fid list when a thread
> > lookups a fid through dentry while another thread unlinks it:
>
> I guess that's "looks up". :)
Err, I guess.
> > UAF thread:
> > refcount_t: addition on 0; use-after-free.
> > p9_fid_get linux/./include/net/9p/client.h:262
> > v9fs_fid_find+0x236/0x280 linux/fs/9p/fid.c:129
> > v9fs_fid_lookup_with_uid linux/fs/9p/fid.c:181
> > v9fs_fid_lookup+0xbf/0xc20 linux/fs/9p/fid.c:314
> > v9fs_vfs_getattr_dotl+0xf9/0x360 linux/fs/9p/vfs_inode_dotl.c:400
> > vfs_statx+0xdd/0x4d0 linux/fs/stat.c:248
> >
> > Freed by:
> > p9_client_clunk+0xb0/0xe0 linux/net/9p/client.c:1456
>
> That line number looks weird.
I have a p9_fid_destroy there (as of a v6.9-rc5 tree); might have moved
a bit though.
Unfortunately it's inlined so the stack trace only has kfree() next
which is why I cut the trace there; I don't think it really matters?
> > p9_fid_put linux/./include/net/9p/client.h:278
> > v9fs_dentry_release+0xb5/0x140 linux/fs/9p/vfs_dentry.c:55
> > v9fs_remove+0x38f/0x620 linux/fs/9p/vfs_inode.c:518
> > vfs_unlink+0x29a/0x810 linux/fs/namei.c:4335
> >
> > The problem is that d_fsdata was not accessed under d_lock, because
> > d_release() normally is only called once the dentry is otherwise no
> > longer accessible but since we also call it explicitly in v9fs_remove
> > that lock is required:
> > move the hlist out of the dentry under lock then unref its fids once
> > they are no longer accessible.
> >
> > Fixes: 154372e67d40 ("fs/9p: fix create-unlink-getattr idiom")
> > Cc: [email protected]
> > Reported-by: Meysam Firouzi
> > Reported-by: Amirmohammad Eftekhar
> > Signed-off-by: Dominique Martinet <[email protected]>
> > ---
> > fs/9p/vfs_dentry.c | 9 +++++++--
> > 1 file changed, 7 insertions(+), 2 deletions(-)
> >
> > diff --git a/fs/9p/vfs_dentry.c b/fs/9p/vfs_dentry.c
> > index f16f73581634..01338d4c2d9e 100644
> > --- a/fs/9p/vfs_dentry.c
> > +++ b/fs/9p/vfs_dentry.c
> > @@ -48,12 +48,17 @@ static int v9fs_cached_dentry_delete(const struct dentry *dentry)
> > static void v9fs_dentry_release(struct dentry *dentry)
> > {
> > struct hlist_node *p, *n;
> > + struct hlist_head head;
> >
> > p9_debug(P9_DEBUG_VFS, " dentry: %pd (%p)\n",
> > dentry, dentry);
> > - hlist_for_each_safe(p, n, (struct hlist_head *)&dentry->d_fsdata)
> > +
> > + spin_lock(&dentry->d_lock);
> > + hlist_move_list((struct hlist_head *)&dentry->d_fsdata, &head);
> > + spin_unlock(&dentry->d_lock);
> > +
> > + hlist_for_each_safe(p, n, &head)
> > p9_fid_put(hlist_entry(p, struct p9_fid, dlist));
> > - dentry->d_fsdata = NULL;
> > }
>
> I'm not sure if that works out. So you are moving the list from dentry to a
> local variable. But if you look at v9fs_fid_find() [fs/9p/fid.c#123] it reads
> dentry->d_fsdata (twice) and holds it as local variable before taking a
> lock. So the lock in v9fs_fid_find() should happen earlier, no?
The comment still works -- if detry->d_fsdata is NULL then
hlist_for_each_entry will stop short and not iterate over anything (it
won't bug out), so that part is fine in my opinion.
What should be improved though is that if dentry->d_inode we can still
look by inode even if there was a d_fsdata as log as fid wasn't found,
e.g.:
-----
diff --git a/fs/9p/fid.c b/fs/9p/fid.c
index de009a33e0e2..c72825fb0ece 100644
--- a/fs/9p/fid.c
+++ b/fs/9p/fid.c
@@ -131,9 +131,9 @@ static struct p9_fid *v9fs_fid_find(struct dentry *dentry, kuid_t uid, int any)
}
}
spin_unlock(&dentry->d_lock);
- } else {
- if (dentry->d_inode)
- ret = v9fs_fid_find_inode(dentry->d_inode, false, uid, any);
+ }
+ if (!ret && dentry->d_inode)
+ ret = v9fs_fid_find_inode(dentry->d_inode, false, uid, any);
}
return ret;
----
I don't think that has to be part of this commit though, the worst that
can happen here is an extra lookup to server instead of a use after
free; I'll send a separate patch for this.
--
Dominique Martinet | Asmadeus
On Wednesday, May 22, 2024 7:25:06 PM CEST Dominique Martinet wrote:
> Christian Schoenebeck wrote on Wed, May 22, 2024 at 04:35:19PM +0200:
[...]
> > > diff --git a/fs/9p/vfs_dentry.c b/fs/9p/vfs_dentry.c
> > > index f16f73581634..01338d4c2d9e 100644
> > > --- a/fs/9p/vfs_dentry.c
> > > +++ b/fs/9p/vfs_dentry.c
> > > @@ -48,12 +48,17 @@ static int v9fs_cached_dentry_delete(const struct dentry *dentry)
> > > static void v9fs_dentry_release(struct dentry *dentry)
> > > {
> > > struct hlist_node *p, *n;
> > > + struct hlist_head head;
> > >
> > > p9_debug(P9_DEBUG_VFS, " dentry: %pd (%p)\n",
> > > dentry, dentry);
> > > - hlist_for_each_safe(p, n, (struct hlist_head *)&dentry->d_fsdata)
> > > +
> > > + spin_lock(&dentry->d_lock);
> > > + hlist_move_list((struct hlist_head *)&dentry->d_fsdata, &head);
> > > + spin_unlock(&dentry->d_lock);
> > > +
> > > + hlist_for_each_safe(p, n, &head)
> > > p9_fid_put(hlist_entry(p, struct p9_fid, dlist));
> > > - dentry->d_fsdata = NULL;
> > > }
> >
> > I'm not sure if that works out. So you are moving the list from dentry to a
> > local variable. But if you look at v9fs_fid_find() [fs/9p/fid.c#123] it reads
> > dentry->d_fsdata (twice) and holds it as local variable before taking a
> > lock. So the lock in v9fs_fid_find() should happen earlier, no?
>
> The comment still works -- if detry->d_fsdata is NULL then
> hlist_for_each_entry will stop short and not iterate over anything (it
> won't bug out), so that part is fine in my opinion.
I meant the opposite: dentry->d_fsdata not being NULL. In this case
v9fs_fid_find() takes a local copy of the list head pointer as `h` without
taking a lock before.
Then v9fs_fid_find() takes the lock to run hlist_for_each_entry(), but at this
point `h` could already point at garbage.
/Christian
Christian Schoenebeck wrote on Thu, May 23, 2024 at 10:34:14AM +0200:
> > The comment still works -- if detry->d_fsdata is NULL then
> > hlist_for_each_entry will stop short and not iterate over anything (it
> > won't bug out), so that part is fine in my opinion.
>
> I meant the opposite: dentry->d_fsdata not being NULL.
I also meant that in the d_fsdata not being NULL branch, if d_fsdata
turns out to be NULL when it is read under lock later.
> In this case v9fs_fid_find() takes a local copy of the list head
> pointer as `h` without taking a lock before.
It doesn't, it takes &dentry->d_fsdata so the address of d_fsdata before
the lock, but that address cannot change here (another thread cannot
change the address of the dentry) ...(continuing below)
> Then v9fs_fid_find() takes the lock to run hlist_for_each_entry(), but at this
> point `h` could already point at garbage.
.. so *h (in practice, head->first in hlist_for_each_entry()) will
properly contain the first node of the list under lock: either NULL if
we just cleared it (at which point the loop won't iterate anything), or
a new list if other items have been added meanwhile.
I really think it's safe, but I do agree that it's hard to read, happy
to move the `h = &dentry->d_fsdata` inside the lock if you prefer -- it
compiles to the same code for me (x86_64/gcc 13.2.0)
--
Dominique Martinet | Asmadeus
On Thursday, May 23, 2024 11:27:28 AM CEST Dominique Martinet wrote:
> Christian Schoenebeck wrote on Thu, May 23, 2024 at 10:34:14AM +0200:
> > > The comment still works -- if detry->d_fsdata is NULL then
> > > hlist_for_each_entry will stop short and not iterate over anything (it
> > > won't bug out), so that part is fine in my opinion.
> >
> > I meant the opposite: dentry->d_fsdata not being NULL.
>
> I also meant that in the d_fsdata not being NULL branch, if d_fsdata
> turns out to be NULL when it is read under lock later.
>
> > In this case v9fs_fid_find() takes a local copy of the list head
> > pointer as `h` without taking a lock before.
>
> It doesn't, it takes &dentry->d_fsdata so the address of d_fsdata before
> the lock, but that address cannot change here (another thread cannot
> change the address of the dentry) ...(continuing below)
Aaah right, I was missing the `&`, my bad!
> > Then v9fs_fid_find() takes the lock to run hlist_for_each_entry(), but at this
> > point `h` could already point at garbage.
>
> ... so *h (in practice, head->first in hlist_for_each_entry()) will
> properly contain the first node of the list under lock: either NULL if
> we just cleared it (at which point the loop won't iterate anything), or
> a new list if other items have been added meanwhile.
Yeah, looks fine to me.
> I really think it's safe, but I do agree that it's hard to read, happy
> to move the `h = &dentry->d_fsdata` inside the lock if you prefer -- it
> compiles to the same code for me (x86_64/gcc 13.2.0)
No need, you can add my RB. Thanks for the clarification!
Reviewed-by: Christian Schoenebeck <[email protected]>
Christian Schoenebeck wrote on Thu, May 23, 2024 at 12:05:44PM +0200:
> > I really think it's safe, but I do agree that it's hard to read, happy
> > to move the `h = &dentry->d_fsdata` inside the lock if you prefer -- it
> > compiles to the same code for me (x86_64/gcc 13.2.0)
>
> No need, you can add my RB. Thanks for the clarification!
>
> Reviewed-by: Christian Schoenebeck <[email protected]>
Thanks!
I've fixed the typo in the commit message and queued it up in -next,
will send this patch and the trace uninit fix to Linus early next week.
--
Dominique Martinet | Asmadeus