On Thu, Aug 8, 2024 at 1:35 PM Andrii Nakryiko
[off-list ref] wrote:

On Thu, Aug 8, 2024 at 9:51 AM Alexei Starovoitov
[off-list ref] wrote:

quoted

On Wed, Aug 7, 2024 at 8:31 AM Andrii Nakryiko
[off-list ref] wrote:

quoted

On Wed, Aug 7, 2024 at 3:30 AM Christian Brauner [off-list ref] wrote:

quoted

On Tue, Aug 06, 2024 at 03:32:20PM GMT, Andrii Nakryiko wrote:

quoted

On Mon, Jul 29, 2024 at 10:20 PM [off-list ref] wrote:

quoted

From: Al Viro <viro@zeniv.linux.org.uk>

Equivalent transformation.  For one thing, it's easier to follow that way.
For another, that simplifies the control flow in the vicinity of struct fd
handling in there, which will allow a switch to CLASS(fd) and make the
thing much easier to verify wrt leaks.

Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
---
 kernel/bpf/verifier.c | 342 +++++++++++++++++++++---------------------
 1 file changed, 172 insertions(+), 170 deletions(-)

This looks unnecessarily intrusive. I think it's best to extract the
logic of fetching and adding bpf_map by fd into a helper and that way
contain fdget + fdput logic nicely. Something like below, which I can
send to bpf-next.

commit b5eec08241cc0263e560551de91eda73ccc5987d
Author: Andrii Nakryiko [off-list ref]
Date:   Tue Aug 6 14:31:34 2024 -0700

    bpf: factor out fetching bpf_map from FD and adding it to used_maps list

    Factor out the logic to extract bpf_map instances from FD embedded in
    bpf_insns, adding it to the list of used_maps (unless it's already
    there, in which case we just reuse map's index). This simplifies the
    logic in resolve_pseudo_ldimm64(), especially around `struct fd`
    handling, as all that is now neatly contained in the helper and doesn't
    leak into a dozen error handling paths.

    Signed-off-by: Andrii Nakryiko [off-list ref]

diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
index df3be12096cf..14e4ef687a59 100644
--- a/kernel/bpf/verifier.c
+++ b/kernel/bpf/verifier.c

@@ -18865,6 +18865,58 @@ static bool bpf_map_is_cgroup_storage(struct

bpf_map *map)
         map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
 }

+/* Add map behind fd to used maps list, if it's not already there, and return
+ * its index. Also set *reused to true if this map was already in the list of
+ * used maps.
+ * Returns <0 on error, or >= 0 index, on success.
+ */
+static int add_used_map_from_fd(struct bpf_verifier_env *env, int fd,
bool *reused)
+{
+    struct fd f = fdget(fd);

Use CLASS(fd, f)(fd) and you can avoid all that fdput() stuff.

That was the point of Al's next patch in the series, so I didn't want
to do it in this one that just refactored the logic of adding maps.
But I can fold that in and send it to bpf-next.

+1.

The bpf changes look ok and Andrii's approach is easier to grasp.
It's better to route bpf conversion to CLASS(fd,..) via bpf-next,
so it goes through bpf CI and our other testing.

bpf patches don't seem to depend on newly added CLASS(fd_pos, ...
and fderr, so pretty much independent from other patches.

Ok, so CLASS(fd, f) won't work just yet because of peculiar
__bpf_map_get() contract: if it gets valid struct fd but it doesn't
contain a valid struct bpf_map, then __bpf_map_get() does fdput()
internally. In all other cases the caller has to do fdput() and
returned struct bpf_map's refcount has to be bumped by the caller
(__bpf_map_get() doesn't do that, I guess that's why it's
double-underscored).

I think the reason it was done was just a convenience to not have to
get/put bpf_map for temporary uses (and instead rely on file's
reference keeping bpf_map alive), plus we have bpf_map_inc() and
bpf_map_inc_uref() variants, so in some cases we need to bump just
refcount, and in some both user and normal refcounts.

So can't use CLASS(fd, ...) without some more clean up.

Alexei, how about changing __bpf_map_get(struct fd f) to
__bpf_map_get_from_fd(int ufd), doing fdget/fdput internally, and
always returning bpf_map with (normal) refcount bumped (if successful,
of course). We can then split bpf_map_inc_with_uref() into just
bpf_map_inc() and bpf_map_inc_uref(), and callers will be able to do
extra uref-only increment, if necessary.

I can do that as a pre-patch, there are about 15 callers, so not too
much work to clean this up. Let me know.

Yeah. Let's kill __bpf_map_get(struct fd ..) altogether.
This logic was added in 2014.
fdget() had to be first and fdput() last to make sure
the map won't disappear while sys_bpf command is running.
All of the places can use bpf_map_get(), bpf_map_put() pair
and rely on map->refcnt, but...

- it's atomic64_inc(&map->refcnt); The cost is probably
in the noise compared to all the work that map sys_bpf commands do.

- It also opens new fuzzing opportunity to do some map operation
in one thread and close(map_fd) in the other, so map->usercnt can
drop to zero and map_release_uref() cleanup can start while
the other thread is still busy doing something like map_update_elem().
It can be mitigated by doing bpf_map_get_with_uref(), but two
atomic64_inc() is kinda too much.

So let's remove __bpf_map_get() and replace all users with bpf_map_get(),
but we may need to revisit that later.