On Tue, Jul 12, 2022 at 10:56:11AM -0400, Joe Lawrence wrote:
On Thu, Jun 30, 2022 at 11:12:26AM -0300, Marcos Paulo de Souza wrote:
Syscalls are called a tricky way. Some architectures add a prefix to the syscall name (SYSCALL_WRAPPER).
This new test creates one userspace process per online cpu calling getpid continuously and tries to livepatch the getpid function. Add the correct function prefix for all archs that select HAS_SYSCALL_WRAPPER.
Signed-off-by: Marcos Paulo de Souza mpdesouza@suse.com
Hi Marcos,
Thanks for working on this and sorry for the delayed reply. I gave it a spin yesterday and have a few comments below...
It has been sometime since the last version, thanks for the comments. I have some questions bellow.
tools/testing/selftests/livepatch/Makefile | 12 +- .../selftests/livepatch/test-syscall.sh | 52 ++++++ .../test_binaries/test_klp-call_getpid.c | 48 ++++++ .../selftests/livepatch/test_modules/Makefile | 3 +- .../livepatch/test_modules/test_klp_syscall.c | 150 ++++++++++++++++++ 5 files changed, 261 insertions(+), 4 deletions(-) create mode 100755 tools/testing/selftests/livepatch/test-syscall.sh create mode 100644 tools/testing/selftests/livepatch/test_binaries/test_klp-call_getpid.c create mode 100644 tools/testing/selftests/livepatch/test_modules/test_klp_syscall.c
diff --git a/tools/testing/selftests/livepatch/Makefile b/tools/testing/selftests/livepatch/Makefile index 5ef492b87bb1..35014197184e 100644 --- a/tools/testing/selftests/livepatch/Makefile +++ b/tools/testing/selftests/livepatch/Makefile @@ -1,10 +1,14 @@ # SPDX-License-Identifier: GPL-2.0 +include ../../../build/Build.include +include ../../../scripts/Makefile.arch +include ../../../scripts/Makefile.include TEST_FILES := settings \ test_modules # We need the test_modules dir in order to make gen_tar and install to work -TEST_GEN_PROGS_EXTENDED := test_modules/test_klp_atomic_replace.ko \ +TEST_GEN_PROGS_EXTENDED := test_binaries/test_klp-call_getpid \
Before I commit this to muscle memory, would "test_programs" better describe this directory? When I see "binaries", I think of source vs. build and not kernel vs. userspace.
Done locally.
test_modules/test_klp_atomic_replace.ko \ test_modules/test_klp_callbacks_busy.ko \ test_modules/test_klp_callbacks_demo.ko \ test_modules/test_klp_callbacks_demo2.ko \@@ -13,7 +17,8 @@ TEST_GEN_PROGS_EXTENDED := test_modules/test_klp_atomic_replace.ko \ test_modules/test_klp_state.ko \ test_modules/test_klp_state2.ko \ test_modules/test_klp_state3.ko \
test_modules/test_klp_shadow_vars.ko
test_modules/test_klp_shadow_vars.ko \test_modules/test_klp_syscall.koTEST_PROGS_EXTENDED := functions.sh TEST_PROGS := \ @@ -21,7 +26,8 @@ TEST_PROGS := \ test-callbacks.sh \ test-shadow-vars.sh \ test-state.sh \
- test-ftrace.sh
- test-ftrace.sh \
- test-syscall.sh
# override lib.mk's default rules OVERRIDE_TARGETS := 1 diff --git a/tools/testing/selftests/livepatch/test-syscall.sh b/tools/testing/selftests/livepatch/test-syscall.sh new file mode 100755 index 000000000000..fb3270de7f1f --- /dev/null +++ b/tools/testing/selftests/livepatch/test-syscall.sh @@ -0,0 +1,52 @@ +#!/bin/bash +# SPDX-License-Identifier: GPL-2.0 +# Copyright (C) 2022 SUSE +# Author: Marcos Paulo de Souza mpdesouza@suse.com
+. $(dirname $0)/functions.sh
+MOD_SYSCALL=test_klp_syscall
+setup_config
+# - Start _NRPROC processes calling getpid and load a livepatch to patch the +# getpid syscall
+start_test "patch getpid syscall while being heavily hammered"
+for i in $(seq 0 $(( $(getconf _NPROCESSORS_ONLN) -1)) ); do
- ./test_binaries/test_klp-call_getpid &
- pids[$i]="$!"
+done
On second thought, bash should only iterate through the elements that are defined, so your earlier version is probably better:
for i in $(seq 1 $(getconf _NPROCESSORS_ONLN)); do
+pid_list=$(echo ${pids[@]} | tr ' ' ',') +load_lp $MOD_SYSCALL klp_pids=$pid_list
We could wait for all test_klp-call_getpid instances to execute the livepatch code:
loop_until 'grep -q '^0$' /sys/kernel/test_klp_syscall/npids'
This has a built in timeout, so it's not infinitely looping. You could add some die "reason" logic here, or just let the existing code cleanup after the failed test.
Makes sense. Done locally.
+pending_pids=$(cat /sys/kernel/test_klp_syscall/npids)
+for pid in ${pids[@]}; do
- kill $pid || true
+done
+disable_lp $MOD_SYSCALL +unload_lp $MOD_SYSCALL
+if [[ "$pending_pids" != "0" ]]; then
- echo -e "FAIL\n\n"
- die "processes not livepatched: $pending_pids"
+fi
+check_result "% insmod test_modules/$MOD_SYSCALL.ko klp_pids=$pid_list +livepatch: enabling patch '$MOD_SYSCALL' +livepatch: '$MOD_SYSCALL': initializing patching transition +livepatch: '$MOD_SYSCALL': starting patching transition +livepatch: '$MOD_SYSCALL': completing patching transition +livepatch: '$MOD_SYSCALL': patching complete +% echo 0 > /sys/kernel/livepatch/$MOD_SYSCALL/enabled +livepatch: '$MOD_SYSCALL': initializing unpatching transition +livepatch: '$MOD_SYSCALL': starting unpatching transition +livepatch: '$MOD_SYSCALL': completing unpatching transition +livepatch: '$MOD_SYSCALL': unpatching complete +% rmmod $MOD_SYSCALL"
+exit 0 diff --git a/tools/testing/selftests/livepatch/test_binaries/test_klp-call_getpid.c b/tools/testing/selftests/livepatch/test_binaries/test_klp-call_getpid.c new file mode 100644 index 000000000000..ec470660dee6 --- /dev/null +++ b/tools/testing/selftests/livepatch/test_binaries/test_klp-call_getpid.c @@ -0,0 +1,48 @@ +// SPDX-License-Identifier: GPL-2.0 +/*
- Copyright (C) 2017 SUSE
- Author: Libor Pechacek lpechacek@suse.cz
- */
+#include <stdio.h> +#include <unistd.h> +#include <sys/syscall.h> +#include <sys/types.h> +#include <signal.h>
+static int stop = 0;
nit: no need to init global to 0
Indeed. Fixed.
+static int sig_int;
+void hup_handler(int signum) +{
- stop = 1;
+}
+void int_handler(int signum) +{
- stop = 1;
- sig_int = 1;
+}
+int main(int argc, char *argv[]) +{
- pid_t orig_pid, pid;
- long count = 0;
- signal(SIGHUP, &hup_handler);
- signal(SIGINT, &int_handler);
- orig_pid = syscall(SYS_getpid);
- while (!stop) {
pid = syscall(SYS_getpid);if (pid != orig_pid)return 1;This test doesn't care about the user program return code, but I wonder if the status should be flipped -- this is the desired code path, not the one at the end of main(), right?
I got the code from our tests. IIUC, syscall with SYS_getpid cannot fail. I'll double check before sending v3.
count++;- }
- if (sig_int)
printf("%ld iterations done\n", count);- return 0;
+} diff --git a/tools/testing/selftests/livepatch/test_modules/Makefile b/tools/testing/selftests/livepatch/test_modules/Makefile index 1eab43b741cd..ebb754accf46 100644 --- a/tools/testing/selftests/livepatch/test_modules/Makefile +++ b/tools/testing/selftests/livepatch/test_modules/Makefile @@ -10,7 +10,8 @@ obj-m += test_klp_atomic_replace.o \ test_klp_state.o \ test_klp_state2.o \ test_klp_state3.o \
- test_klp_shadow_vars.o
- test_klp_shadow_vars.o \
- test_klp_syscall.o
%.ko: make -C $(KDIR) M=$(TESTMODS_DIR) $@ diff --git a/tools/testing/selftests/livepatch/test_modules/test_klp_syscall.c b/tools/testing/selftests/livepatch/test_modules/test_klp_syscall.c new file mode 100644 index 000000000000..e90f4ac8e7a4 --- /dev/null +++ b/tools/testing/selftests/livepatch/test_modules/test_klp_syscall.c @@ -0,0 +1,150 @@ +// SPDX-License-Identifier: GPL-2.0 +/*
- Copyright (C) 2017-2022 SUSE
- Authors: Libor Pechacek lpechacek@suse.cz
Nicolai Stange <nstange@suse.de>
Marcos Paulo de Souza <mpdesouza@suse.com>- */
+#include <linux/module.h> +#include <linux/kernel.h> +#include <linux/sched.h> +#include <linux/slab.h> +#include <linux/livepatch.h>
+#if defined(__x86_64__) +#define FN_PREFIX __x64_ +#elif defined(__s390x__) +#define FN_PREFIX __s390x_ +#elif defined(__aarch64__) +#define FN_PREFIX __arm64_ +#else +/* powerpc does not select ARCH_HAS_SYSCALL_WRAPPER */ +#define FN_PREFIX +#endif
+struct klp_pid_t {
- pid_t pid;
- struct list_head list;
+}; +static LIST_HEAD(klp_pid_list);
+/* Protects klp_pid_list */ +static DEFINE_MUTEX(kpid_mutex);
+static int klp_pids[NR_CPUS]; +static unsigned int npids; +module_param_array(klp_pids, int, &npids, 0);
+static ssize_t npids_show(struct kobject *kobj, struct kobj_attribute *attr,
char *buf)+{
- return sprintf(buf, "%u\n", npids);
+}
+static struct kobj_attribute klp_attr = __ATTR_RO(npids); +static struct kobject *klp_kobj;
+static void free_klp_pid_list(void) +{
- struct klp_pid_t *kpid, *temp;
- mutex_lock(&kpid_mutex);
- list_for_each_entry_safe(kpid, temp, &klp_pid_list, list) {
list_del(&kpid->list);kfree(kpid);- }
- mutex_unlock(&kpid_mutex);
+}
+asmlinkage long lp_sys_getpid(void) +{
- struct klp_pid_t *kpid, *temp;
- /*
* For each thread calling getpid, check if the pid exists in* klp_pid_list. If yes, decrement the npids variable and remove the pid* from the list. npids will be later used to ensure that all pids* transitioned to the liveaptched state.I go back and forth with this comment: 1) the code is pretty self-explanatory and 2) it explains what test-syscall.sh is doing with npids
Describing the behavioral difference I suggest below would be helpful (if you agree w/the change). Something about removing the current pid from the list, okay, too.
Makes sense. I'll move the comment to the test.
Then as far as describing npids' purpose, I'd put that up at the top of the file where npids is defined. It can be as generic as describing what it's counting.
Thanks for the suggestion. A short description will be added when defining the module argument.
*/- mutex_lock(&kpid_mutex);
- list_for_each_entry_safe(kpid, temp, &klp_pid_list, list) {
if (current->pid == kpid->pid) {list_del(&kpid->list);kfree(kpid);npids--;break;I think it would be safer to return task_tgid_vnr() here, but ...
}- }
- mutex_unlock(&kpid_mutex);
- return task_tgid_vnr(current);
task_pid_vnr() here. That way we're only changing behavior for the processes in the list and not all programs across the system.
So, looking at getpid syscall definition, it calls task_tgid_vnr:
SYSCALL_DEFINE0(getpid) { return task_tgid_vnr(current); }
The function task_pid_vnr is called by gettid. So, IIUC, the test should call task_tgid_vnr, since getpid is being livepatched.
Am I missing something?
+}
+static struct klp_func vmlinux_funcs[] = {
- {
.old_name = __stringify(FN_PREFIX) "sys_getpid",.new_func = lp_sys_getpid,- }, {}
+};
+static struct klp_object objs[] = {
- {
/* name being NULL means vmlinux */.funcs = vmlinux_funcs,- }, {}
+};
+static struct klp_patch patch = {
- .mod = THIS_MODULE,
- .objs = objs,
+};
+static int livepatch_init(void) +{
- int ret;
- struct klp_pid_t *kpid;
- if (npids > 0) {
int i;for (i = 0; i < npids; i++) {kpid = kmalloc(sizeof(struct klp_pid_t), GFP_KERNEL);if (!kpid)goto err_mem;kpid->pid = klp_pids[i];list_add(&kpid->list, &klp_pid_list);}- }
- klp_kobj = kobject_create_and_add("test_klp_syscall", kernel_kobj);
- if (!klp_kobj)
goto err_mem;- ret = sysfs_create_file(klp_kobj, &klp_attr.attr);
- if (ret) {
kobject_put(klp_kobj);goto out_klp_pid_list;- }
- return klp_enable_patch(&patch);
+err_mem:
- ret = -ENOMEM;
+out_klp_pid_list:
- free_klp_pid_list();
- return ret;
+}
+static void livepatch_exit(void) +{
- free_klp_pid_list();
- kobject_put(klp_kobj);
+}
+module_init(livepatch_init); +module_exit(livepatch_exit); +MODULE_LICENSE("GPL"); +MODULE_INFO(livepatch, "Y");
The linked list code looks like it should work and cleans up properly, but if the current test is the extent of what we need, it would be simpler to directly use klp_pids[] to track progress: since we're never going to livepatch pid 0, we could just search klp_pids[] for our pid and then zero it out once we've executed the livepatch code.
This would let us drop out the kmalloc and list management code (looks to be nearly 25% of the file at this point). I think the kpid_mutex still needs to stay though. Iterating through the fixed array shouldn't be too much more expensive than through the (shrinking) klp_pid_list.
On the other hand, if you think that we'll need to adapt this module to handle a dynamically changing list of klp_pids[], then perhaps a separate list is the way to go.
My first idea was to use klp_pids indexed by NR_CPUS, but I thought about bigger machines with more CPUs, making the array too big for just a couple of pids to be tracked...
Thinking again, I don't think it's a problem. Also, I don't believe that we would need to dynamically add pids, so this code can go. In the next version I'll just access klp_pids directly. Thanks for the suggestion!
Final note, kbuild is probably my weakest area, so I'm of limited help review-wise there. In trying out the code, I wasn't sure exactly how to run things. I did:
$ make -C tools/testing/selftests/livepatch/ run_tests
and that built everything and ran the tests (good). I changed a few .c files and tried again, but they didn't rebuild. Then from tools/testing/selftests/livepatch/test_modules/ I ran `make` and it only ran the clean target (not good, I think).
I need to test it and make sure the code is rebuilt. Thanks for noticing this problem.
What is the typical workflow supposed to be in this combined setup?
The plan was to support all forms of running the selftests: by running it in place and by running the "installed" kselftests. I'll test it again to make sure that the code is rebuilt when it should.
Thanks,
-- Joe