bpf: Fix potential race in tail call compatibility check
authorToke Høiland-Jørgensen <toke@redhat.com>
Tue, 26 Oct 2021 11:00:19 +0000 (13:00 +0200)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Tue, 2 Nov 2021 18:48:21 +0000 (19:48 +0100)
commit 54713c85f536048e685258f880bf298a74c3620d upstream.

Lorenzo noticed that the code testing for program type compatibility of
tail call maps is potentially racy in that two threads could encounter a
map with an unset type simultaneously and both return true even though they
are inserting incompatible programs.

The race window is quite small, but artificially enlarging it by adding a
usleep_range() inside the check in bpf_prog_array_compatible() makes it
trivial to trigger from userspace with a program that does, essentially:

        map_fd = bpf_create_map(BPF_MAP_TYPE_PROG_ARRAY, 4, 4, 2, 0);
        pid = fork();
        if (pid) {
                key = 0;
                value = xdp_fd;
        } else {
                key = 1;
                value = tc_fd;
        }
        err = bpf_map_update_elem(map_fd, &key, &value, 0);

While the race window is small, it has potentially serious ramifications in
that triggering it would allow a BPF program to tail call to a program of a
different type. So let's get rid of it by protecting the update with a
spinlock. The commit in the Fixes tag is the last commit that touches the
code in question.

v2:
- Use a spinlock instead of an atomic variable and cmpxchg() (Alexei)
v3:
- Put lock and the members it protects into an embedded 'owner' struct (Daniel)

Fixes: 3324b584b6f6 ("ebpf: misc core cleanup")
Reported-by: Lorenzo Bianconi <lorenzo.bianconi@redhat.com>
Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20211026110019.363464-1-toke@redhat.com
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
include/linux/bpf.h
kernel/bpf/arraymap.c
kernel/bpf/core.c
kernel/bpf/syscall.c

index 0caa448..1f62a4e 100644 (file)
@@ -862,8 +862,11 @@ struct bpf_array_aux {
         * stored in the map to make sure that all callers and callees have
         * the same prog type and JITed flag.
         */
-       enum bpf_prog_type type;
-       bool jited;
+       struct {
+               spinlock_t lock;
+               enum bpf_prog_type type;
+               bool jited;
+       } owner;
        /* Programs with direct jumps into programs part of this array. */
        struct list_head poke_progs;
        struct bpf_map *map;
index c6c81ec..36c68dc 100644 (file)
@@ -1025,6 +1025,7 @@ static struct bpf_map *prog_array_map_alloc(union bpf_attr *attr)
        INIT_WORK(&aux->work, prog_array_map_clear_deferred);
        INIT_LIST_HEAD(&aux->poke_progs);
        mutex_init(&aux->poke_mutex);
+       spin_lock_init(&aux->owner.lock);
 
        map = array_map_alloc(attr);
        if (IS_ERR(map)) {
index 2e4a658..72e4bf0 100644 (file)
@@ -1775,20 +1775,26 @@ static unsigned int __bpf_prog_ret0_warn(const void *ctx,
 bool bpf_prog_array_compatible(struct bpf_array *array,
                               const struct bpf_prog *fp)
 {
+       bool ret;
+
        if (fp->kprobe_override)
                return false;
 
-       if (!array->aux->type) {
+       spin_lock(&array->aux->owner.lock);
+
+       if (!array->aux->owner.type) {
                /* There's no owner yet where we could check for
                 * compatibility.
                 */
-               array->aux->type  = fp->type;
-               array->aux->jited = fp->jited;
-               return true;
+               array->aux->owner.type  = fp->type;
+               array->aux->owner.jited = fp->jited;
+               ret = true;
+       } else {
+               ret = array->aux->owner.type  == fp->type &&
+                     array->aux->owner.jited == fp->jited;
        }
-
-       return array->aux->type  == fp->type &&
-              array->aux->jited == fp->jited;
+       spin_unlock(&array->aux->owner.lock);
+       return ret;
 }
 
 static int bpf_check_tail_call(const struct bpf_prog *fp)
index 9433ab9..9cae74a 100644 (file)
@@ -535,8 +535,10 @@ static void bpf_map_show_fdinfo(struct seq_file *m, struct file *filp)
 
        if (map->map_type == BPF_MAP_TYPE_PROG_ARRAY) {
                array = container_of(map, struct bpf_array, map);
-               type  = array->aux->type;
-               jited = array->aux->jited;
+               spin_lock(&array->aux->owner.lock);
+               type  = array->aux->owner.type;
+               jited = array->aux->owner.jited;
+               spin_unlock(&array->aux->owner.lock);
        }
 
        seq_printf(m,