--- /dev/null
+// SPDX-License-Identifier: GPL-2.0
+#include <unistd.h>
+#include <pthread.h>
+#include <sys/mman.h>
+#include <stdatomic.h>
+#include <test_progs.h>
+#include <sys/syscall.h>
+#include <linux/module.h>
+#include <linux/userfaultfd.h>
+
+#include "ksym_race.skel.h"
+#include "bpf_mod_race.skel.h"
+#include "kfunc_call_race.skel.h"
+
+/* This test crafts a race between btf_try_get_module and do_init_module, and
+ * checks whether btf_try_get_module handles the invocation for a well-formed
+ * but uninitialized module correctly. Unless the module has completed its
+ * initcalls, the verifier should fail the program load and return ENXIO.
+ *
+ * userfaultfd is used to trigger a fault in an fmod_ret program, and make it
+ * sleep, then the BPF program is loaded and the return value from verifier is
+ * inspected. After this, the userfaultfd is closed so that the module loading
+ * thread makes forward progress, and fmod_ret injects an error so that the
+ * module load fails and it is freed.
+ *
+ * If the verifier succeeded in loading the supplied program, it will end up
+ * taking reference to freed module, and trigger a crash when the program fd
+ * is closed later. This is true for both kfuncs and ksyms. In both cases,
+ * the crash is triggered inside bpf_prog_free_deferred, when module reference
+ * is finally released.
+ */
+
+struct test_config {
+ const char *str_open;
+ void *(*bpf_open_and_load)();
+ void (*bpf_destroy)(void *);
+};
+
+enum test_state {
+ _TS_INVALID,
+ TS_MODULE_LOAD,
+ TS_MODULE_LOAD_FAIL,
+};
+
+static _Atomic enum test_state state = _TS_INVALID;
+
+static int sys_finit_module(int fd, const char *param_values, int flags)
+{
+ return syscall(__NR_finit_module, fd, param_values, flags);
+}
+
+static int sys_delete_module(const char *name, unsigned int flags)
+{
+ return syscall(__NR_delete_module, name, flags);
+}
+
+static int load_module(const char *mod)
+{
+ int ret, fd;
+
+ fd = open("bpf_testmod.ko", O_RDONLY);
+ if (fd < 0)
+ return fd;
+
+ ret = sys_finit_module(fd, "", 0);
+ close(fd);
+ if (ret < 0)
+ return ret;
+ return 0;
+}
+
+static void *load_module_thread(void *p)
+{
+
+ if (!ASSERT_NEQ(load_module("bpf_testmod.ko"), 0, "load_module_thread must fail"))
+ atomic_store(&state, TS_MODULE_LOAD);
+ else
+ atomic_store(&state, TS_MODULE_LOAD_FAIL);
+ return p;
+}
+
+static int sys_userfaultfd(int flags)
+{
+ return syscall(__NR_userfaultfd, flags);
+}
+
+static int test_setup_uffd(void *fault_addr)
+{
+ struct uffdio_register uffd_register = {};
+ struct uffdio_api uffd_api = {};
+ int uffd;
+
+ uffd = sys_userfaultfd(O_CLOEXEC);
+ if (uffd < 0)
+ return -errno;
+
+ uffd_api.api = UFFD_API;
+ uffd_api.features = 0;
+ if (ioctl(uffd, UFFDIO_API, &uffd_api)) {
+ close(uffd);
+ return -1;
+ }
+
+ uffd_register.range.start = (unsigned long)fault_addr;
+ uffd_register.range.len = 4096;
+ uffd_register.mode = UFFDIO_REGISTER_MODE_MISSING;
+ if (ioctl(uffd, UFFDIO_REGISTER, &uffd_register)) {
+ close(uffd);
+ return -1;
+ }
+ return uffd;
+}
+
+static void test_bpf_mod_race_config(const struct test_config *config)
+{
+ void *fault_addr, *skel_fail;
+ struct bpf_mod_race *skel;
+ struct uffd_msg uffd_msg;
+ pthread_t load_mod_thrd;
+ _Atomic int *blockingp;
+ int uffd, ret;
+
+ fault_addr = mmap(0, 4096, PROT_READ, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
+ if (!ASSERT_NEQ(fault_addr, MAP_FAILED, "mmap for uffd registration"))
+ return;
+
+ if (!ASSERT_OK(sys_delete_module("bpf_testmod", 0), "unload bpf_testmod"))
+ goto end_mmap;
+
+ skel = bpf_mod_race__open();
+ if (!ASSERT_OK_PTR(skel, "bpf_mod_kfunc_race__open"))
+ goto end_module;
+
+ skel->rodata->bpf_mod_race_config.tgid = getpid();
+ skel->rodata->bpf_mod_race_config.inject_error = -4242;
+ skel->rodata->bpf_mod_race_config.fault_addr = fault_addr;
+ if (!ASSERT_OK(bpf_mod_race__load(skel), "bpf_mod___load"))
+ goto end_destroy;
+ blockingp = (_Atomic int *)&skel->bss->bpf_blocking;
+
+ if (!ASSERT_OK(bpf_mod_race__attach(skel), "bpf_mod_kfunc_race__attach"))
+ goto end_destroy;
+
+ uffd = test_setup_uffd(fault_addr);
+ if (!ASSERT_GE(uffd, 0, "userfaultfd open + register address"))
+ goto end_destroy;
+
+ if (!ASSERT_OK(pthread_create(&load_mod_thrd, NULL, load_module_thread, NULL),
+ "load module thread"))
+ goto end_uffd;
+
+ /* Now, we either fail loading module, or block in bpf prog, spin to find out */
+ while (!atomic_load(&state) && !atomic_load(blockingp))
+ ;
+ if (!ASSERT_EQ(state, _TS_INVALID, "module load should block"))
+ goto end_join;
+ if (!ASSERT_EQ(*blockingp, 1, "module load blocked")) {
+ pthread_kill(load_mod_thrd, SIGKILL);
+ goto end_uffd;
+ }
+
+ /* We might have set bpf_blocking to 1, but may have not blocked in
+ * bpf_copy_from_user. Read userfaultfd descriptor to verify that.
+ */
+ if (!ASSERT_EQ(read(uffd, &uffd_msg, sizeof(uffd_msg)), sizeof(uffd_msg),
+ "read uffd block event"))
+ goto end_join;
+ if (!ASSERT_EQ(uffd_msg.event, UFFD_EVENT_PAGEFAULT, "read uffd event is pagefault"))
+ goto end_join;
+
+ /* We know that load_mod_thrd is blocked in the fmod_ret program, the
+ * module state is still MODULE_STATE_COMING because mod->init hasn't
+ * returned. This is the time we try to load a program calling kfunc and
+ * check if we get ENXIO from verifier.
+ */
+ skel_fail = config->bpf_open_and_load();
+ ret = errno;
+ if (!ASSERT_EQ(skel_fail, NULL, config->str_open)) {
+ /* Close uffd to unblock load_mod_thrd */
+ close(uffd);
+ uffd = -1;
+ while (atomic_load(blockingp) != 2)
+ ;
+ ASSERT_OK(kern_sync_rcu(), "kern_sync_rcu");
+ config->bpf_destroy(skel_fail);
+ goto end_join;
+
+ }
+ ASSERT_EQ(ret, ENXIO, "verifier returns ENXIO");
+ ASSERT_EQ(skel->data->res_try_get_module, false, "btf_try_get_module == false");
+
+ close(uffd);
+ uffd = -1;
+end_join:
+ pthread_join(load_mod_thrd, NULL);
+ if (uffd < 0)
+ ASSERT_EQ(atomic_load(&state), TS_MODULE_LOAD_FAIL, "load_mod_thrd success");
+end_uffd:
+ if (uffd >= 0)
+ close(uffd);
+end_destroy:
+ bpf_mod_race__destroy(skel);
+ ASSERT_OK(kern_sync_rcu(), "kern_sync_rcu");
+end_module:
+ sys_delete_module("bpf_testmod", 0);
+ ASSERT_OK(load_module("bpf_testmod.ko"), "restore bpf_testmod");
+end_mmap:
+ munmap(fault_addr, 4096);
+ atomic_store(&state, _TS_INVALID);
+}
+
+static const struct test_config ksym_config = {
+ .str_open = "ksym_race__open_and_load",
+ .bpf_open_and_load = (void *)ksym_race__open_and_load,
+ .bpf_destroy = (void *)ksym_race__destroy,
+};
+
+static const struct test_config kfunc_config = {
+ .str_open = "kfunc_call_race__open_and_load",
+ .bpf_open_and_load = (void *)kfunc_call_race__open_and_load,
+ .bpf_destroy = (void *)kfunc_call_race__destroy,
+};
+
+void serial_test_bpf_mod_race(void)
+{
+ if (test__start_subtest("ksym (used_btfs UAF)"))
+ test_bpf_mod_race_config(&ksym_config);
+ if (test__start_subtest("kfunc (kfunc_btf_tab UAF)"))
+ test_bpf_mod_race_config(&kfunc_config);
+}
--- /dev/null
+// SPDX-License-Identifier: GPL-2.0
+#include <vmlinux.h>
+#include <bpf/bpf_helpers.h>
+#include <bpf/bpf_tracing.h>
+
+const volatile struct {
+ /* thread to activate trace programs for */
+ pid_t tgid;
+ /* return error from __init function */
+ int inject_error;
+ /* uffd monitored range start address */
+ void *fault_addr;
+} bpf_mod_race_config = { -1 };
+
+int bpf_blocking = 0;
+int res_try_get_module = -1;
+
+static __always_inline bool check_thread_id(void)
+{
+ struct task_struct *task = bpf_get_current_task_btf();
+
+ return task->tgid == bpf_mod_race_config.tgid;
+}
+
+/* The trace of execution is something like this:
+ *
+ * finit_module()
+ * load_module()
+ * prepare_coming_module()
+ * notifier_call(MODULE_STATE_COMING)
+ * btf_parse_module()
+ * btf_alloc_id() // Visible to userspace at this point
+ * list_add(btf_mod->list, &btf_modules)
+ * do_init_module()
+ * freeinit = kmalloc()
+ * ret = mod->init()
+ * bpf_prog_widen_race()
+ * bpf_copy_from_user()
+ * ...<sleep>...
+ * if (ret < 0)
+ * ...
+ * free_module()
+ * return ret
+ *
+ * At this point, module loading thread is blocked, we now load the program:
+ *
+ * bpf_check
+ * add_kfunc_call/check_pseudo_btf_id
+ * btf_try_get_module
+ * try_get_module_live == false
+ * return -ENXIO
+ *
+ * Without the fix (try_get_module_live in btf_try_get_module):
+ *
+ * bpf_check
+ * add_kfunc_call/check_pseudo_btf_id
+ * btf_try_get_module
+ * try_get_module == true
+ * <store module reference in btf_kfunc_tab or used_btf array>
+ * ...
+ * return fd
+ *
+ * Now, if we inject an error in the blocked program, our module will be freed
+ * (going straight from MODULE_STATE_COMING to MODULE_STATE_GOING).
+ * Later, when bpf program is freed, it will try to module_put already freed
+ * module. This is why try_get_module_live returns false if mod->state is not
+ * MODULE_STATE_LIVE.
+ */
+
+SEC("fmod_ret.s/bpf_fentry_test1")
+int BPF_PROG(widen_race, int a, int ret)
+{
+ char dst;
+
+ if (!check_thread_id())
+ return 0;
+ /* Indicate that we will attempt to block */
+ bpf_blocking = 1;
+ bpf_copy_from_user(&dst, 1, bpf_mod_race_config.fault_addr);
+ return bpf_mod_race_config.inject_error;
+}
+
+SEC("fexit/do_init_module")
+int BPF_PROG(fexit_init_module, struct module *mod, int ret)
+{
+ if (!check_thread_id())
+ return 0;
+ /* Indicate that we finished blocking */
+ bpf_blocking = 2;
+ return 0;
+}
+
+SEC("fexit/btf_try_get_module")
+int BPF_PROG(fexit_module_get, const struct btf *btf, struct module *mod)
+{
+ res_try_get_module = !!mod;
+ return 0;
+}
+
+char _license[] SEC("license") = "GPL";