On Tue, Apr 2, 2024 at 9:39 PM Joe Lawrence <joe.lawrence@xxxxxxxxxx> wrote: > > On 4/1/24 22:45, Yafang Shao wrote: > > On Mon, Apr 1, 2024 at 11:02 PM Joe Lawrence <joe.lawrence@xxxxxxxxxx> wrote: > >> > >> On Sun, Mar 31, 2024 at 09:38:39PM +0800, Yafang Shao wrote: > >>> Enhance the functionality of kpatch to automatically remove the associated > >>> module when replacing an old livepatch with a new one. This ensures that no > >>> leftover modules remain in the system. For instance: > >>> > >>> - Load the first livepatch > >>> $ kpatch load 6.9.0-rc1+/livepatch-test_0.ko > >>> loading patch module: 6.9.0-rc1+/livepatch-test_0.ko > >>> waiting (up to 15 seconds) for patch transition to complete... > >>> transition complete (2 seconds) > >>> > >>> $ kpatch list > >>> Loaded patch modules: > >>> livepatch_test_0 [enabled] > >>> > >>> $ lsmod |grep livepatch > >>> livepatch_test_0 16384 1 > >>> > >>> - Load a new livepatch > >>> $ kpatch load 6.9.0-rc1+/livepatch-test_1.ko > >>> loading patch module: 6.9.0-rc1+/livepatch-test_1.ko > >>> waiting (up to 15 seconds) for patch transition to complete... > >>> transition complete (2 seconds) > >>> > >>> $ kpatch list > >>> Loaded patch modules: > >>> livepatch_test_1 [enabled] > >>> > >>> $ lsmod |grep livepatch > >>> livepatch_test_1 16384 1 > >>> livepatch_test_0 16384 0 <<<< leftover > >>> > >>> With this improvement, executing > >>> `kpatch load 6.9.0-rc1+/livepatch-test_1.ko` will automatically remove the > >>> livepatch-test_0.ko module. > >>> > >> > >> Hi Yafang, > >> > >> I think it would be better if the commit message reasoning used > >> insmod/modprobe directly rather than the kpatch user utility wrapper. > >> That would be more generic and remove any potential kpatch utility > >> variants from the picture. (For example, it is possible to add `rmmod` > >> in the wrapper and then this patch would be redundant.) > > > > Hi Joe, > > > > I attempted to incorporate an `rmmod` operation within the kpatch > > replacement process, but encountered challenges in devising a safe and > > effective solution. The difficulty arises from the uncertainty > > regarding which livepatch will be replaced in userspace, necessitating > > the operation to be conducted within the kernel itself. > > > > I wasn't suggesting that the kpatch user utility should or could solve > this problem, just that this scenario is not specific to kpatch. And > since this is a kernel patch, it would be consistent to speak in terms > of livepatches: the repro can be phrased in terms of modprobe/insmod, > /sys/kernel/livepatch/ sysfs, rmmod, etc. for which those not using the > kpatch utility are more familiar with. Understood. Thanks for your explanation. I will try it. > > >> > >>> Signed-off-by: Yafang Shao <laoar.shao@xxxxxxxxx> > >>> --- > >>> include/linux/module.h | 1 + > >>> kernel/livepatch/core.c | 11 +++++++++-- > >>> kernel/module/main.c | 43 ++++++++++++++++++++++++----------------- > >>> 3 files changed, 35 insertions(+), 20 deletions(-) > >>> > >>> diff --git a/include/linux/module.h b/include/linux/module.h > >>> index 1153b0d99a80..9a95174a919b 100644 > >>> --- a/include/linux/module.h > >>> +++ b/include/linux/module.h > >>> @@ -75,6 +75,7 @@ extern struct module_attribute module_uevent; > >>> /* These are either module local, or the kernel's dummy ones. */ > >>> extern int init_module(void); > >>> extern void cleanup_module(void); > >>> +extern void delete_module(struct module *mod); > >>> > >>> #ifndef MODULE > >>> /** > >>> diff --git a/kernel/livepatch/core.c b/kernel/livepatch/core.c > >>> index ecbc9b6aba3a..f1edc999f3ef 100644 > >>> --- a/kernel/livepatch/core.c > >>> +++ b/kernel/livepatch/core.c > >>> @@ -711,6 +711,8 @@ static void klp_free_patch_start(struct klp_patch *patch) > >>> */ > >>> static void klp_free_patch_finish(struct klp_patch *patch) > >>> { > >>> + struct module *mod = patch->mod; > >>> + > >>> /* > >>> * Avoid deadlock with enabled_store() sysfs callback by > >>> * calling this outside klp_mutex. It is safe because > >>> @@ -721,8 +723,13 @@ static void klp_free_patch_finish(struct klp_patch *patch) > >>> wait_for_completion(&patch->finish); > >>> > >>> /* Put the module after the last access to struct klp_patch. */ > >>> - if (!patch->forced) > >>> - module_put(patch->mod); > >>> + if (!patch->forced) { > >>> + module_put(mod); > >>> + if (module_refcount(mod)) > >>> + return; > >>> + mod->state = MODULE_STATE_GOING; > >>> + delete_module(mod); > >>> + } > > I'm gonna have to read study code in kernel/module/ to be confident that > this is completely safe. What happens if this code races a concurrent > `rmmod` from the user (perhaps that pesky kpatch utility)? Can a stray > module reference sneak between the code here. Etc. The existing > delete_module syscall does some additional safety checks under the > module_mutex, which may or may not make sense for this use case... Petr, > Miroslav any thoughts? A race condition may occur. It appears necessary to modify the mod->state under the protection of module_mutex. If the state is not MODULE_STATE_LIVE, it must be skipped. > > Also, code-wise, it would be nice if the mod->state were only assigned > inside the kernel/module/main.c code... maybe this little sequence can > be pushed into that file so it's all in one place? good suggestion. will do it. > > >>> } > >>> > >>> /* > >>> diff --git a/kernel/module/main.c b/kernel/module/main.c > >>> index e1e8a7a9d6c1..e863e1f87dfd 100644 > >>> --- a/kernel/module/main.c > >>> +++ b/kernel/module/main.c > >>> @@ -695,12 +695,35 @@ EXPORT_SYMBOL(module_refcount); > >>> /* This exists whether we can unload or not */ > >>> static void free_module(struct module *mod); > >>> > >>> +void delete_module(struct module *mod) > >>> +{ > >>> + char buf[MODULE_FLAGS_BUF_SIZE]; > >>> + > >>> + /* Final destruction now no one is using it. */ > >>> + if (mod->exit != NULL) > >>> + mod->exit(); > >>> + blocking_notifier_call_chain(&module_notify_list, > >>> + MODULE_STATE_GOING, mod); > >>> + klp_module_going(mod); > >>> + ftrace_release_mod(mod); > >>> + > >>> + async_synchronize_full(); > >>> + > >>> + /* Store the name and taints of the last unloaded module for diagnostic purposes */ > >>> + strscpy(last_unloaded_module.name, mod->name, sizeof(last_unloaded_module.name)); > >>> + strscpy(last_unloaded_module.taints, module_flags(mod, buf, false), > >>> + sizeof(last_unloaded_module.taints)); > >>> + > >>> + free_module(mod); > >>> + /* someone could wait for the module in add_unformed_module() */ > >>> + wake_up_all(&module_wq); > >>> +} > >>> + > >>> SYSCALL_DEFINE2(delete_module, const char __user *, name_user, > >>> unsigned int, flags) > >>> { > >>> struct module *mod; > >>> char name[MODULE_NAME_LEN]; > >>> - char buf[MODULE_FLAGS_BUF_SIZE]; > >>> int ret, forced = 0; > >>> > >>> if (!capable(CAP_SYS_MODULE) || modules_disabled) > >>> @@ -750,23 +773,7 @@ SYSCALL_DEFINE2(delete_module, const char __user *, name_user, > >>> goto out; > >>> > >>> mutex_unlock(&module_mutex); > >>> - /* Final destruction now no one is using it. */ > >>> - if (mod->exit != NULL) > >>> - mod->exit(); > >>> - blocking_notifier_call_chain(&module_notify_list, > >>> - MODULE_STATE_GOING, mod); > >>> - klp_module_going(mod); > >>> - ftrace_release_mod(mod); > >>> - > >>> - async_synchronize_full(); > >>> - > >>> - /* Store the name and taints of the last unloaded module for diagnostic purposes */ > >>> - strscpy(last_unloaded_module.name, mod->name, sizeof(last_unloaded_module.name)); > >>> - strscpy(last_unloaded_module.taints, module_flags(mod, buf, false), sizeof(last_unloaded_module.taints)); > >>> - > >>> - free_module(mod); > >>> - /* someone could wait for the module in add_unformed_module() */ > >>> - wake_up_all(&module_wq); > >>> + delete_module(mod); > >>> return 0; > >>> out: > >>> mutex_unlock(&module_mutex); > >>> -- > >>> 2.39.1 > >>> > >> > >> It's been a while since atomic replace was added and so I forget why the > >> implementation doesn't try this -- is it possible for the livepatch > >> module to have additional references that this patch would force its way > >> through? > > > > In the klp_free_patch_finish() function, a check is performed on the > > reference count of the livepatch module. If the reference count is not > > zero, the function will skip further processing. > > > >> > >> Also, this patch will break the "atomic replace livepatch" kselftest in > >> test-livepatch.sh [1]. I think it would need to drop the `unload_lp > >> $MOD_LIVEPATCH` command, the following 'live patched' greps and their > >> corresponding dmesg output in the test's final check_result() call. > > > > Thanks for your information. I will check it. > > > > Let me know if you have any questions, I'm more familiar with that code > than the atomic replace / module interactions :) > You're correct in noting that we need to discard certain unload_lps and rmmods. This is because, after implementing the change, executing `echo 0 > /sys/kernel/livepatch/${livepatch}/enabled` will remove the associated kernel module. The question then arises: is this change in behavior acceptable, or should we avoid it? -- Regards Yafang