selftests/bpf: Add test for race in btf_try_get_module

This adds a complete test case to ensure we never take references to
modules not in MODULE_STATE_LIVE, which can lead to UAF, and it also
ensures we never access btf->kfunc_set_tab in an inconsistent state.

The test uses userfaultfd to artificially widen the race.

When run on an unpatched kernel, it leads to the following splat:

[root@(none) bpf]# ./test_progs -t bpf_mod_race/ksym
[ 55.498171] BUG: unable to handle page fault for address: fffffbfff802548b
[ 55.499206] #PF: supervisor read access in kernel mode
[ 55.499855] #PF: error_code(0x0000) - not-present page
[ 55.500555] PGD a4fa9067 P4D a4fa9067 PUD a4fa5067 PMD 1b44067 PTE 0
[ 55.501499] Oops: 0000 [#1] PREEMPT SMP KASAN NOPTI
[ 55.502195] CPU: 0 PID: 83 Comm: kworker/0:2 Tainted: G OE 5.16.0-rc4+ #151
[ 55.503388] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS ArchLinux 1.15.0-1 04/01/2014
[ 55.504777] Workqueue: events bpf_prog_free_deferred
[ 55.505563] RIP: 0010:kasan_check_range+0x184/0x1d0
[ 55.509140] RSP: 0018:ffff88800560fcf0 EFLAGS: 00010282
[ 55.509977] RAX: fffffbfff802548b RBX: fffffbfff802548c RCX: ffffffff9337b6ba
[ 55.511096] RDX: fffffbfff802548c RSI: 0000000000000004 RDI: ffffffffc012a458
[ 55.512143] RBP: fffffbfff802548b R08: 0000000000000001 R09: ffffffffc012a45b
[ 55.513228] R10: fffffbfff802548b R11: 0000000000000001 R12: ffff888001b5f598
[ 55.514332] R13: ffff888004f49ac8 R14: 0000000000000000 R15: ffff888092449400
[ 55.515418] FS: 0000000000000000(0000) GS:ffff888092400000(0000) knlGS:0000000000000000
[ 55.516705] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[ 55.517560] CR2: fffffbfff802548b CR3: 0000000007c10006 CR4: 0000000000770ef0
[ 55.518672] PKRU: 55555554
[ 55.519022] Call Trace:
[ 55.519483] <TASK>
[ 55.519884] module_put.part.0+0x2a/0x180
[ 55.520642] bpf_prog_free_deferred+0x129/0x2e0
[ 55.521478] process_one_work+0x4fa/0x9e0
[ 55.522122] ? pwq_dec_nr_in_flight+0x100/0x100
[ 55.522878] ? rwlock_bug.part.0+0x60/0x60
[ 55.523551] worker_thread+0x2eb/0x700
[ 55.524176] ? __kthread_parkme+0xd8/0xf0
[ 55.524853] ? process_one_work+0x9e0/0x9e0
[ 55.525544] kthread+0x23a/0x270
[ 55.526088] ? set_kthread_struct+0x80/0x80
[ 55.526798] ret_from_fork+0x1f/0x30
[ 55.527413] </TASK>
[ 55.527813] Modules linked in: bpf_testmod(OE) [last unloaded: bpf_testmod]
[ 55.530846] CR2: fffffbfff802548b
[ 55.531341] ---[ end trace 1af41803c054ad6d ]---
[ 55.532136] RIP: 0010:kasan_check_range+0x184/0x1d0
[ 55.535887] RSP: 0018:ffff88800560fcf0 EFLAGS: 00010282
[ 55.536711] RAX: fffffbfff802548b RBX: fffffbfff802548c RCX: ffffffff9337b6ba
[ 55.537821] RDX: fffffbfff802548c RSI: 0000000000000004 RDI: ffffffffc012a458
[ 55.538899] RBP: fffffbfff802548b R08: 0000000000000001 R09: ffffffffc012a45b
[ 55.539928] R10: fffffbfff802548b R11: 0000000000000001 R12: ffff888001b5f598
[ 55.541021] R13: ffff888004f49ac8 R14: 0000000000000000 R15: ffff888092449400
[ 55.542108] FS: 0000000000000000(0000) GS:ffff888092400000(0000) knlGS:0000000000000000
[ 55.543260]CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[ 55.544136] CR2: fffffbfff802548b CR3: 0000000007c10006 CR4: 0000000000770ef0
[ 55.545317] PKRU: 55555554
[ 55.545671] note: kworker/0:2[83] exited with preempt_count 1

Signed-off-by: Kumar Kartikeya Dwivedi <[e-mail]>
Link: https://lore.kernel.org/r/[e-mail]
Signed-off-by: Alexei Starovoitov <[e-mail]>

net/bpf/test_run.c

···
 {
 	return a + 1;
 }
+EXPORT_SYMBOL_GPL(bpf_fentry_test1);
+ALLOW_ERROR_INJECTION(bpf_fentry_test1, ERRNO);
 
 int noinline bpf_fentry_test2(int a, u64 b)
 {

tools/testing/selftests/bpf/bpf_testmod/bpf_testmod.c

···
 	.check_set = &bpf_testmod_check_kfunc_ids,
 };
 
+extern int bpf_fentry_test1(int a);
+
 static int bpf_testmod_init(void)
 {
 	int ret;
···
 	ret = register_btf_kfunc_id_set(BPF_PROG_TYPE_SCHED_CLS, &bpf_testmod_kfunc_set);
 	if (ret < 0)
 		return ret;
+	if (bpf_fentry_test1(0) < 0)
+		return -EINVAL;
 	return sysfs_create_bin_file(kernel_kobj, &bin_attr_bpf_testmod_file);
 }
 

tools/testing/selftests/bpf/config

···
 CONFIG_NF_DEFRAG_IPV4=y
 CONFIG_NF_DEFRAG_IPV6=y
 CONFIG_NF_CONNTRACK=y
+CONFIG_USERFAULTFD=y

tools/testing/selftests/bpf/prog_tests/bpf_mod_race.c

···
+// 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);
+}

tools/testing/selftests/bpf/progs/bpf_mod_race.c

···
+// 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";

tools/testing/selftests/bpf/progs/kfunc_call_race.c

···
+// SPDX-License-Identifier: GPL-2.0
+#include <vmlinux.h>
+#include <bpf/bpf_helpers.h>
+
+extern void bpf_testmod_test_mod_kfunc(int i) __ksym;
+
+SEC("tc")
+int kfunc_call_fail(struct __sk_buff *ctx)
+{
+	bpf_testmod_test_mod_kfunc(0);
+	return 0;
+}
+
+char _license[] SEC("license") = "GPL";

tools/testing/selftests/bpf/progs/ksym_race.c

···
+// SPDX-License-Identifier: GPL-2.0
+#include <vmlinux.h>
+#include <bpf/bpf_helpers.h>
+
+extern int bpf_testmod_ksym_percpu __ksym;
+
+SEC("tc")
+int ksym_fail(struct __sk_buff *ctx)
+{
+	return *(int *)bpf_this_cpu_ptr(&bpf_testmod_ksym_percpu);
+}
+
+char _license[] SEC("license") = "GPL";