Hi,
it appears the little feature might require a little bit more work to find its
value of the patch.
Using the following debug module , some debugging shows the
TIF_USER_RETURN_NOTIFY
bit is propagated in __switch_to among tasks , but USER_RETURN_NOTIFY call back
seems to
be dropped somewhere on somone who carries the bit return to user space.
side notes:
there is an issue that the module symbols is not appended to
/sys/kernel/debug/tracing/available_filter_functions
which should be sovled first to make it easier for further debuggig.
[root@localhost linux]# cat lib/user-return-test.c
#include <linux/module.h>
#include <linux/kernel.h>
#include <linux/init.h>
#include <linux/container_of.h>
#include <linux/user-return-notifier.h>
#include <linux/delay.h>
#include <linux/kthread.h>
#include <linux/sched.h>
MODULE_LICENSE("GPL");
struct test_user_return {
struct user_return_notifier urn;
bool registered;
int urn_value_changed;
struct task_struct *worker;
};
static struct test_user_return __percpu *user_return_test;
static void test_user_return_cb(struct user_return_notifier *urn)
{
struct test_user_return *tur =
container_of(urn, struct test_user_return, urn);
unsigned long flags;
local_irq_save(flags);
tur->urn_value_changed++;
local_irq_restore(flags);
return;
}
static int test_user_return_worker(void *tur)
{
struct test_user_return *t;
t = (struct test_user_return *) tur;
preempt_disable();
user_return_notifier_register(&t->urn);
preempt_enable();
t->registered = true;
while (!kthread_should_stop()) {
static int err_rate = 0;
msleep (1000);
if (!test_thread_flag(TIF_USER_RETURN_NOTIFY) && (err_rate ==
0)) {
pr_err("TIF_USER_RETURN_NOTIFY is lost");
err_rate++;
}
}
return 0;
}
static int init_test_user_return(void)
{
int r = 0;
user_return_test = alloc_percpu(struct test_user_return);
if (!user_return_test) {
pr_err("failed to allocate percpu test_user_return\n");
r = -ENOMEM;
goto exit;
}
{
unsigned int cpu;
struct task_struct *task;
struct test_user_return *tur;
for_each_online_cpu(cpu) {
tur = per_cpu_ptr(user_return_test, cpu);
if (!tur->registered) {
tur->urn.on_user_return = test_user_return_cb;
task = kthread_create(test_user_return_worker,
tur, "test_user_return");
if (IS_ERR(task))
pr_err("no test_user_return kthread
created for cpu %d",cpu);
else {
tur->worker = task;
wake_up_process(task);
}
}
}
}
exit:
return r;
}
static void exit_test_user_return(void)
{
struct test_user_return *tur;
int i,ret=0;
for_each_online_cpu(i) {
tur = per_cpu_ptr(user_return_test, i);
if (tur->registered) {
pr_info("[cpu=%d, %d] ", i, tur->urn_value_changed);
user_return_notifier_unregister(&tur->urn);
tur->registered = false;
}
if (tur->worker) {
ret = kthread_stop(tur->worker);
if (ret)
pr_err("can't stop test_user_return kthread for
cpu %d", i);
}
}
free_percpu(user_return_test);
return;
}
module_init(init_test_user_return);
module_exit(exit_test_user_return);
------------------ Original ------------------
From: "Christophe Leroy"<christophe.le...@csgroup.eu>;
Date: Tue, Feb 20, 2024 05:02 PM
To: "mpe"<m...@ellerman.id.au>; "Aneesh Kumar K.V"<aneesh.ku...@kernel.org>;
"虞陆铭"<luming...@shingroup.cn>; "linuxppc-dev"<linuxppc-dev@lists.ozlabs.org>;
"linux-kernel"<linux-ker...@vger.kernel.org>; "npiggin"<npig...@gmail.com>;
Cc: "shenghui...@shingroup.cn"<shenghui...@shingroup.cn>;
"dawei...@shingroup.cn"<dawei...@shingroup.cn>;
"ke.z...@shingroup.cn"<ke.z...@shingroup.cn>; "luming.yu"<luming...@gmail.com>;
Subject: Re: [PATCH v1 2/2] powerpc/debug: hook to user return notifier
infrastructure
Le 20/02/2024 à 09:51, Christophe Leroy a écrit :
>
>
> Le 19/12/2023 à 07:33, Michael Ellerman a écrit :
>> Aneesh Kumar K.V <aneesh.ku...@kernel.org> writes:
>>> Luming Yu <luming...@shingroup.cn> writes:
>>>
>>>> Before we have powerpc to use the generic entry infrastructure,
>>>> the call to fire user return notifier is made temporarily in powerpc
>>>> entry code.
>>>>
>>>
>>> It is still not clear what will be registered as user return notifier.
>>> Can you summarize that here?
>>
>> fire_user_return_notifiers() is defined in kernel/user-return-notifier.c
>>
>> That's built when CONFIG_USER_RETURN_NOTIFIER=y.
>>
>> That is not user selectable, it's only enabled by:
>>
>> arch/x86/kvm/Kconfig: select USER_RETURN_NOTIFIER
>>
>> So it looks to me like (currently) it's always a nop and does nothing.
>>
>> Which makes me wonder what the point of wiring this feature up is :)
>> Maybe it's needed for some other feature I don't know about?
>>
>> Arguably we could just enable it because we can, and it currently does
>> nothing so it's unlikely to break anything. But that also makes it
>> impossible to test the implementation is correct, and runs the risk that
>> one day in the future when it does get enabled only then do we discover
>> it doesn't work.
>
> Opened an "issue" for the day we need it:
> https://github.com/KSPP/linux/issues/348
Correct one is https://github.com/linuxppc/issues/issues/477
