On 2023/3/31 16:31, Petr Mladek wrote:
> On Thu 2023-03-30 22:59:12, Jiri Olsa wrote:
>> On Thu, Mar 30, 2023 at 08:26:41PM +0800, Leizhen (ThunderTown) wrote:
>>>
>>>
>>> On 2023/3/30 15:29, Jiri Olsa wrote:
>>>> ping,
>>>>
>>>> Petr, Zhen, any comment on discussion below?
>>>>
>>>> thanks,
>>>> jirka
>>>>
>>>> On Thu, Mar 23, 2023 at 03:00:25PM +0100, Jiri Olsa wrote:
>>>>> On Wed, Mar 22, 2023 at 09:03:46AM -0700, Alexei Starovoitov wrote:
>>>>>> On Wed, Mar 22, 2023 at 5:14 AM Jiri Olsa <olsajiri@xxxxxxxxx> wrote:
>>>>>>>
>>>>>>> On Wed, Mar 22, 2023 at 10:49:38AM +0100, Artem Savkov wrote:
>>>>>>>
>>>>>>> SNIP
>>>>>>>
>>>>>>>>>> Hm, do we even need to preempt_disable? IIUC, preempt_disable is used
>>>>>>>>>> in module kallsyms to prevent taking the module lock b/c kallsyms are
>>>>>>>>>> used in the oops path. That shouldn't be an issue here, is that correct?
>>>>>>>>>
>>>>>>>>> btf_try_get_module calls try_module_get which disables the preemption,
>>>>>>>>> so no need to call it in here
>>>>>>>>
>>>>>>>> It does, but it reenables preemption right away so it is enabled by the
>>>>>>>> time we call find_kallsyms_symbol_value(). I am getting the following
>>>>>>>> lockdep splat while running module_fentry_shadow test from test_progs.
>>>>>>>>
>>>>>>>> [ 12.017973][ T488] =============================
>>>>>>>> [ 12.018529][ T488] WARNING: suspicious RCU usage
>>>>>>>> [ 12.018987][ T488] 6.2.0.bpf-test-13063-g6a9f5cdba3c5 #804 Tainted: G OE
>>>>>>>> [ 12.019898][ T488] -----------------------------
>>>>>>>> [ 12.020391][ T488] kernel/module/kallsyms.c:448 suspicious rcu_dereference_check() usage!
>>>>>>>> [ 12.021335][ T488]
>>>>>>>> [ 12.021335][ T488] other info that might help us debug this:
>>>>>>>> [ 12.021335][ T488]
>>>>>>>> [ 12.022416][ T488]
>>>>>>>> [ 12.022416][ T488] rcu_scheduler_active = 2, debug_locks = 1
>>>>>>>> [ 12.023297][ T488] no locks held by test_progs/488.
>>>>>>>> [ 12.023854][ T488]
>>>>>>>> [ 12.023854][ T488] stack backtrace:
>>>>>>>> [ 12.024336][ T488] CPU: 0 PID: 488 Comm: test_progs Tainted: G OE 6.2.0.bpf-test-13063-g6a9f5cdba3c5 #804
>>>>>>>> [ 12.025290][ T488] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.16.1-2.fc37 04/01/2014
>>>>>>>> [ 12.026108][ T488] Call Trace:
>>>>>>>> [ 12.026381][ T488] <TASK>
>>>>>>>> [ 12.026649][ T488] dump_stack_lvl+0xb4/0x110
>>>>>>>> [ 12.027060][ T488] lockdep_rcu_suspicious+0x158/0x1f0
>>>>>>>> [ 12.027541][ T488] find_kallsyms_symbol_value+0xe8/0x110
>>>>>>>> [ 12.028028][ T488] bpf_check_attach_target+0x838/0xa20
>>>>>>>> [ 12.028511][ T488] check_attach_btf_id+0x144/0x3f0
>>>>>>>> [ 12.028957][ T488] ? __pfx_cmp_subprogs+0x10/0x10
>>>>>>>> [ 12.029408][ T488] bpf_check+0xeec/0x1850
>>>>>>>> [ 12.029799][ T488] ? ktime_get_with_offset+0x124/0x1d0
>>>>>>>> [ 12.030247][ T488] bpf_prog_load+0x87a/0xed0
>>>>>>>> [ 12.030627][ T488] ? __lock_release+0x5f/0x160
>>>>>>>> [ 12.031010][ T488] ? __might_fault+0x53/0xb0
>>>>>>>> [ 12.031394][ T488] ? selinux_bpf+0x6c/0xa0
>>>>>>>> [ 12.031756][ T488] __sys_bpf+0x53c/0x1240
>>>>>>>> [ 12.032115][ T488] __x64_sys_bpf+0x27/0x40
>>>>>>>> [ 12.032476][ T488] do_syscall_64+0x3e/0x90
>>>>>>>> [ 12.032835][ T488] entry_SYSCALL_64_after_hwframe+0x72/0xdc
>>>>>>>
>>>>>>> --- a/kernel/module/kallsyms.c
>>>>>>> +++ b/kernel/module/kallsyms.c
>>> Commit 91fb02f31505 ("module: Move kallsyms support into a separate file") hides
>>> the answer. find_kallsyms_symbol_value() was originally a static function, and it
>>> is only called by module_kallsyms_lookup_name() and is preemptive-protected.
>>>
>>> Now that we've added a call to function find_kallsyms_symbol_value(), it seems like
>>> we should do the same thing as function module_kallsyms_lookup_name().
>>>
>>> Like this?
>>> + mod = btf_try_get_module(btf);
>>> + if (mod) {
>>> + preempt_disable();
>>> + addr = find_kallsyms_symbol_value(mod, tname);
>>> + preempt_enable();
>>> + } else
>>> + addr = 0;
>>
>> yes, that's what I did above, but I was just curious about the strange
>> RCU usage Alexei commented on earlier:
>>
>> >>> +unsigned long find_kallsyms_symbol_value(struct module *mod, const char *name)
>> >>> +{
>> >>> + unsigned long ret;
>> >>> +
>> >>> + preempt_disable();
>> >>> + ret = __find_kallsyms_symbol_value(mod, name);
>> >>> + preempt_enable();
>> >>> + return ret;
>> >>> +}
>> >>
>> >> That doesn't look right.
>> >> I think the issue is misuse of rcu_dereference_sched in
>> >> find_kallsyms_symbol_value.
>> >
>> > it seems to be using rcu pointer to keep symbols for module init time and
>> > then core symbols for after init.. and switch between them when module is
>> > loaded, hence the strange rcu usage I think
load_module
post_relocation
add_kallsyms
mod->kallsyms = (void __rcu *)mod->init_layout.base + info->mod_kallsyms_init_off; (1)
do_init_module
freeinit->module_init = mod->init_layout.base;
rcu_assign_pointer(mod->kallsyms, &mod->core_kallsyms); (2)
if (llist_add(&freeinit->node, &init_free_list))
schedule_work(&init_free_wq);
do_free_init
synchronize_rcu();
module_memfree(initfree->module_init);
IIUC, the RCU can help synchronize_rcu() in do_free_init() to make sure that no one
is still using the first mod->kallsyms (1). If find_kallsyms_symbol_value() is executed
between (1) and (2).
>
> My understanding is that rcu is needed to prevent module from being freed.
> It should be related to:
>
> static void free_module(struct module *mod)
> {
> [...]
> /* Now we can delete it from the lists */
> mutex_lock(&module_mutex);
> /* Unlink carefully: kallsyms could be walking list. */
> list_del_rcu(&mod->list);
> [...]
> }
>
> I am sorry for the late reply. I was busy and I thought that it was
> related to the refactoring. I hoped that peopled doing the refactoring
> would answer.
>
> Best Regards,
> Petr
> .
>
--
Regards,
Zhen Lei
