commit 15b2219facad ("kernel: freezer should treat PF_IO_WORKER like
PF_KTHREAD for freezing") is to fix the freezeing issue of IO threads
by making the freezer not send them fake signals.
Here live patching consistency model call klp_send_signals to wake up
all tasks by send fake signal to all non-kthread which only check the
PF_KTHREAD flag, so it still send signal to io threads which may lead to
freezeing issue of io threads.
Here we take the same fix action by treating PF_IO_WORKERS as PF_KTHREAD
within klp_send_signal function.
Signed-off-by: Dong Kai <[email protected]>
---
note:
the io threads freeze issue links:
[1] https://lore.kernel.org/io-uring/YEgnIp43%[email protected]/
[2] https://lore.kernel.org/io-uring/[email protected]/
kernel/livepatch/transition.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/kernel/livepatch/transition.c b/kernel/livepatch/transition.c
index f6310f848f34..0e1c35c8f4b4 100644
--- a/kernel/livepatch/transition.c
+++ b/kernel/livepatch/transition.c
@@ -358,7 +358,7 @@ static void klp_send_signals(void)
* Meanwhile the task could migrate itself and the action
* would be meaningless. It is not serious though.
*/
- if (task->flags & PF_KTHREAD) {
+ if (task->flags & (PF_KTHREAD | PF_IO_WORKER)) {
/*
* Wake up a kthread which sleeps interruptedly and
* still has not been migrated.
--
2.17.1
On 3/24/21 9:48 PM, Dong Kai wrote:
> commit 15b2219facad ("kernel: freezer should treat PF_IO_WORKER like
> PF_KTHREAD for freezing") is to fix the freezeing issue of IO threads
nit: s/freezeing/freezing
> by making the freezer not send them fake signals.
>
> Here live patching consistency model call klp_send_signals to wake up
> all tasks by send fake signal to all non-kthread which only check the
> PF_KTHREAD flag, so it still send signal to io threads which may lead to
> freezeing issue of io threads.
>
> Here we take the same fix action by treating PF_IO_WORKERS as PF_KTHREAD
> within klp_send_signal function.
>
> Signed-off-by: Dong Kai <[email protected]>
> ---
> note:
> the io threads freeze issue links:
> [1] https://lore.kernel.org/io-uring/YEgnIp43%[email protected]/
> [2] https://lore.kernel.org/io-uring/[email protected]/
>
> kernel/livepatch/transition.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/kernel/livepatch/transition.c b/kernel/livepatch/transition.c
> index f6310f848f34..0e1c35c8f4b4 100644
> --- a/kernel/livepatch/transition.c
> +++ b/kernel/livepatch/transition.c
> @@ -358,7 +358,7 @@ static void klp_send_signals(void)
> * Meanwhile the task could migrate itself and the action
> * would be meaningless. It is not serious though.
> */
> - if (task->flags & PF_KTHREAD) {
> + if (task->flags & (PF_KTHREAD | PF_IO_WORKER)) {
> /*
> * Wake up a kthread which sleeps interruptedly and
> * still has not been migrated.
>
(PF_KTHREAD | PF_IO_WORKER) is open coded in soo many places maybe this
is a silly question, but...
If the livepatch code could use fake_signal_wake_up(), we could
consolidate the pattern in klp_send_signals() with the one in
freeze_task(). Then there would only one place for wake up / fake
signal logic.
I don't fully understand the differences in the freeze_task() version,
so I only pose this as a question and not v2 request.
As it is, this change seems logical to me, so:
Acked-by: Joe Lawrence <[email protected]>
Thanks,
-- Joe
On Thu, 25 Mar 2021, Dong Kai wrote:
> commit 15b2219facad ("kernel: freezer should treat PF_IO_WORKER like
> PF_KTHREAD for freezing") is to fix the freezeing issue of IO threads
> by making the freezer not send them fake signals.
>
> Here live patching consistency model call klp_send_signals to wake up
> all tasks by send fake signal to all non-kthread which only check the
> PF_KTHREAD flag, so it still send signal to io threads which may lead to
> freezeing issue of io threads.
I suppose this could happen, but it will also affect the live patching
transition if the io threads do not react to signals.
Are you able to reproduce it easily? I mean, is there a testcase I could
use to take a closer look?
> Here we take the same fix action by treating PF_IO_WORKERS as PF_KTHREAD
> within klp_send_signal function.
Yes, this sounds reasonable.
Miroslav
> Signed-off-by: Dong Kai <[email protected]>
> ---
> note:
> the io threads freeze issue links:
> [1] https://lore.kernel.org/io-uring/YEgnIp43%[email protected]/
> [2] https://lore.kernel.org/io-uring/[email protected]/
>
> kernel/livepatch/transition.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/kernel/livepatch/transition.c b/kernel/livepatch/transition.c
> index f6310f848f34..0e1c35c8f4b4 100644
> --- a/kernel/livepatch/transition.c
> +++ b/kernel/livepatch/transition.c
> @@ -358,7 +358,7 @@ static void klp_send_signals(void)
> * Meanwhile the task could migrate itself and the action
> * would be meaningless. It is not serious though.
> */
> - if (task->flags & PF_KTHREAD) {
> + if (task->flags & (PF_KTHREAD | PF_IO_WORKER)) {
> /*
> * Wake up a kthread which sleeps interruptedly and
> * still has not been migrated.
> (PF_KTHREAD | PF_IO_WORKER) is open coded in soo many places maybe this is a
> silly question, but...
>
> If the livepatch code could use fake_signal_wake_up(), we could consolidate
> the pattern in klp_send_signals() with the one in freeze_task(). Then there
> would only one place for wake up / fake signal logic.
>
> I don't fully understand the differences in the freeze_task() version, so I
> only pose this as a question and not v2 request.
The plan was to remove our live patching fake signal completely and use
the new infrastructure Jens proposed in the past.
Something like
diff --git a/kernel/livepatch/transition.c b/kernel/livepatch/transition.c
index f6310f848f34..3a4beb9395c4 100644
--- a/kernel/livepatch/transition.c
+++ b/kernel/livepatch/transition.c
@@ -9,6 +9,7 @@
#include <linux/cpu.h>
#include <linux/stacktrace.h>
+#include <linux/tracehook.h>
#include "core.h"
#include "patch.h"
#include "transition.h"
@@ -369,9 +370,7 @@ static void klp_send_signals(void)
* Send fake signal to all non-kthread tasks which are
* still not migrated.
*/
- spin_lock_irq(&task->sighand->siglock);
- signal_wake_up(task, 0);
- spin_unlock_irq(&task->sighand->siglock);
+ set_notify_signal(task);
}
}
read_unlock(&tasklist_lock);
diff --git a/kernel/signal.c b/kernel/signal.c
index a15c584a0455..b7cf4eda8611 100644
--- a/kernel/signal.c
+++ b/kernel/signal.c
@@ -181,8 +181,7 @@ void recalc_sigpending_and_wake(struct task_struct *t)
void recalc_sigpending(void)
{
- if (!recalc_sigpending_tsk(current) && !freezing(current) &&
- !klp_patch_pending(current))
+ if (!recalc_sigpending_tsk(current) && !freezing(current))
clear_thread_flag(TIF_SIGPENDING);
}
Let me verify it still works and there are all the needed pieces merged
for all the architectures we support (x86_64, ppc64le and s390x). I'll
send a proper patch then.
Miroslav
On 3/24/21 7:48 PM, Dong Kai wrote:
> commit 15b2219facad ("kernel: freezer should treat PF_IO_WORKER like
> PF_KTHREAD for freezing") is to fix the freezeing issue of IO threads
> by making the freezer not send them fake signals.
>
> Here live patching consistency model call klp_send_signals to wake up
> all tasks by send fake signal to all non-kthread which only check the
> PF_KTHREAD flag, so it still send signal to io threads which may lead to
> freezeing issue of io threads.
>
> Here we take the same fix action by treating PF_IO_WORKERS as PF_KTHREAD
> within klp_send_signal function.
Reviewed-by: Jens Axboe <[email protected]>
--
Jens Axboe
On 3/25/21 3:30 AM, Miroslav Benes wrote:
>> (PF_KTHREAD | PF_IO_WORKER) is open coded in soo many places maybe this is a
>> silly question, but...
>>
>> If the livepatch code could use fake_signal_wake_up(), we could consolidate
>> the pattern in klp_send_signals() with the one in freeze_task(). Then there
>> would only one place for wake up / fake signal logic.
>>
>> I don't fully understand the differences in the freeze_task() version, so I
>> only pose this as a question and not v2 request.
>
> The plan was to remove our live patching fake signal completely and use
> the new infrastructure Jens proposed in the past.
That would be great, I've actually been waiting for that to show up!
I would greatly prefer this approach if you deem it suitable for 5.12,
if not we'll still need the temporary work-around for live patching.
--
Jens Axboe
On 3/25/21 5:26 AM, Miroslav Benes wrote:
> On Thu, 25 Mar 2021, Dong Kai wrote:
>
>> commit 15b2219facad ("kernel: freezer should treat PF_IO_WORKER like
>> PF_KTHREAD for freezing") is to fix the freezeing issue of IO threads
>> by making the freezer not send them fake signals.
>>
>> Here live patching consistency model call klp_send_signals to wake up
>> all tasks by send fake signal to all non-kthread which only check the
>> PF_KTHREAD flag, so it still send signal to io threads which may lead to
>> freezeing issue of io threads.
>
> I suppose this could happen, but it will also affect the live patching
> transition if the io threads do not react to signals.
>
> Are you able to reproduce it easily? I mean, is there a testcase I could
> use to take a closer look?
>
If repro is only hypothetical at this point, perhaps we can artificially
create it in selftests? And useful to verify the future change you
mentioned in your other reply?
-- Joe
On 2021/3/25 17:26, Miroslav Benes wrote:
> On Thu, 25 Mar 2021, Dong Kai wrote:
>
>> commit 15b2219facad ("kernel: freezer should treat PF_IO_WORKER like
>> PF_KTHREAD for freezing") is to fix the freezeing issue of IO threads
>> by making the freezer not send them fake signals.
>>
>> Here live patching consistency model call klp_send_signals to wake up
>> all tasks by send fake signal to all non-kthread which only check the
>> PF_KTHREAD flag, so it still send signal to io threads which may lead to
>> freezeing issue of io threads.
>
> I suppose this could happen, but it will also affect the live patching
> transition if the io threads do not react to signals.
>
> Are you able to reproduce it easily? I mean, is there a testcase I could
> use to take a closer look?
>
Um... I tried but failed to reproduce this on real environment as i'm
not familiar with the io uring usage.
So i use a tricky way to verify this possibility by the following patch
which create a fake io thread in module and patch the func which is
always within thread running stack. Then the stack check will failed
when transition and trigger the klp_send_signal flow.
This example may not suitable, but you can get my point
Kai
Note: this patch export some symbols just for test via module because if
i create io thread via sysinit, it will receive SIGKILL signal[set by
zap_other_threads] when run init process and exit the loop, weird...
diff --git a/fs/exec.c b/fs/exec.c
index 18594f11c31f..a64af6cac43b 100644
--- a/fs/exec.c
+++ b/fs/exec.c
@@ -1229,6 +1229,7 @@ void __set_task_comm(struct task_struct *tsk,
const char *buf, bool exec)
task_unlock(tsk);
perf_event_comm(tsk, exec);
}
+EXPORT_SYMBOL_GPL(__set_task_comm);
/*
* Calling this is the point of no return. None of the failures will be
diff --git a/kernel/fork.c b/kernel/fork.c
index 54cc905e5fe0..03064fef7bb1 100644
--- a/kernel/fork.c
+++ b/kernel/fork.c
@@ -2447,6 +2447,7 @@ struct task_struct *create_io_thread(int
(*fn)(void *), void *arg, int node)
}
return tsk;
}
+EXPORT_SYMBOL(create_io_thread);
/*
* Ok, this is the main fork-routine.
index 98191218d891..8151d17149a0 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -3856,6 +3856,7 @@ void wake_up_new_task(struct task_struct *p)
#endif
task_rq_unlock(rq, p, &rf);
}
+EXPORT_SYMBOL_GPL(wake_up_new_task);
#ifdef CONFIG_PREEMPT_NOTIFIERS
diff --git a/samples/test/Makefile b/samples/test/Makefile
new file mode 100644
index 000000000000..efbf01c6477e
--- /dev/null
+++ b/samples/test/Makefile
@@ -0,0 +1 @@
+obj-m += io_thread.o livepatch-sample.o
diff --git a/samples/test/io_thread.c b/samples/test/io_thread.c
new file mode 100644
index 000000000000..e7bdc51a4582
--- /dev/null
+++ b/samples/test/io_thread.c
@@ -0,0 +1,49 @@
+#include <linux/module.h>
+#include <linux/sched.h>
+#include <linux/sched/task.h>
+#include <linux/sched/signal.h>
+
+static __used noinline void func(void)
+{
+ printk("func\n");
+ schedule_timeout(HZ * 5);
+}
+
+static int io_worker(void *data)
+{
+ set_task_comm(current, "io_worker");
+ while (1) {
+ set_current_state(TASK_INTERRUPTIBLE);
+ func();
+
+ if (fatal_signal_pending(current))
+ break;
+ }
+
+ return 0;
+}
+
+static int __init io_thread_init(void)
+{
+ struct task_struct *task = NULL;
+
+ task = create_io_thread(io_worker, NULL, 0);
+ if (task == NULL)
+ return -EINVAL;
+ wake_up_new_task(task);
+
+ /* when insmod exit, io thread got SIGKILL and exit, so... */
+ while (1)
+ schedule_timeout(HZ);
+ return 0;
+}
+
+static void __exit io_thread_exit(void)
+{
+ return;
+}
+
+module_init(io_thread_init);
+module_exit(io_thread_exit);
+
+MODULE_LICENSE("GPL");
diff --git a/samples/test/livepatch-sample.c
b/samples/test/livepatch-sample.c
new file mode 100644
index 000000000000..c35b494f5c5a
--- /dev/null
+++ b/samples/test/livepatch-sample.c
@@ -0,0 +1,43 @@
+#include <linux/module.h>
+#include <linux/kernel.h>
+#include <linux/livepatch.h>
+
+static void new_func(void)
+{
+ schedule_timeout(HZ * 5);
+ printk("new_func\n");
+}
+
+static struct klp_func funcs[] = {
+ {
+ .old_name = "func",
+ .new_func = new_func,
+ }, { }
+};
+
+static struct klp_object objs[] = {
+ {
+ .name = "io_thread",
+ .funcs = funcs,
+ }, { }
+};
+
+static struct klp_patch patch = {
+ .mod = THIS_MODULE,
+ .objs = objs,
+};
+
+static int livepatch_init(void)
+{
+ return klp_enable_patch(&patch);
+}
+
+static void livepatch_exit(void)
+{
+}
+
+module_init(livepatch_init);
+module_exit(livepatch_exit);
+MODULE_INFO(livepatch, "Y");
+
+MODULE_LICENSE("GPL");
--
>> Here we take the same fix action by treating PF_IO_WORKERS as PF_KTHREAD
>> within klp_send_signal function.
>
> Yes, this sounds reasonable.
>
> Miroslav
>
>> Signed-off-by: Dong Kai <[email protected]>
>> ---
>> note:
>> the io threads freeze issue links:
>> [1] https://lore.kernel.org/io-uring/YEgnIp43%[email protected]/
>> [2] https://lore.kernel.org/io-uring/[email protected]/
>>
>> kernel/livepatch/transition.c | 2 +-
>> 1 file changed, 1 insertion(+), 1 deletion(-)
>>
>> diff --git a/kernel/livepatch/transition.c b/kernel/livepatch/transition.c
>> index f6310f848f34..0e1c35c8f4b4 100644
>> --- a/kernel/livepatch/transition.c
>> +++ b/kernel/livepatch/transition.c
>> @@ -358,7 +358,7 @@ static void klp_send_signals(void)
>> * Meanwhile the task could migrate itself and the action
>> * would be meaningless. It is not serious though.
>> */
>> - if (task->flags & PF_KTHREAD) {
>> + if (task->flags & (PF_KTHREAD | PF_IO_WORKER)) {
>> /*
>> * Wake up a kthread which sleeps interruptedly and
>> * still has not been migrated.
>
On Thu, 25 Mar 2021, Jens Axboe wrote:
> On 3/25/21 3:30 AM, Miroslav Benes wrote:
> >> (PF_KTHREAD | PF_IO_WORKER) is open coded in soo many places maybe this is a
> >> silly question, but...
> >>
> >> If the livepatch code could use fake_signal_wake_up(), we could consolidate
> >> the pattern in klp_send_signals() with the one in freeze_task(). Then there
> >> would only one place for wake up / fake signal logic.
> >>
> >> I don't fully understand the differences in the freeze_task() version, so I
> >> only pose this as a question and not v2 request.
> >
> > The plan was to remove our live patching fake signal completely and use
> > the new infrastructure Jens proposed in the past.
>
> That would be great, I've actually been waiting for that to show up!
Sorry about that. I failed to notice that the infrastructure was merged
already. I'll send it soonish.
> I would greatly prefer this approach if you deem it suitable for 5.12,
> if not we'll still need the temporary work-around for live patching.
I noticed there is [email protected] now, so I
suppose we should wait for that to land in mainline and simply do nothing
about PF_IO_WORKER for live patching.
Miroslav