* [PATCH] PM: sleep: Replace read_lock/unlock(tasklist_lock) with rcu_read_lock/unlock() @ 2021-06-07 6:57 qiang.zhang 2021-06-07 11:45 ` Rafael J. Wysocki 0 siblings, 1 reply; 7+ messages in thread From: qiang.zhang @ 2021-06-07 6:57 UTC (permalink / raw) To: rafael.j.wysocki, len.brown, pavel, paulmck; +Cc: linux-pm, linux-kernel From: Zqiang <qiang.zhang@windriver.com> Using rcu_read_lock/unlock() instead of read_lock/unlock(tasklist_lock), the task list can be traversed in parallel to any list additions or removals, improve concurrency. Signed-off-by: Zqiang <qiang.zhang@windriver.com> --- kernel/power/process.c | 16 ++++++++-------- 1 file changed, 8 insertions(+), 8 deletions(-) diff --git a/kernel/power/process.c b/kernel/power/process.c index 50cc63534486..0f8dee9ee097 100644 --- a/kernel/power/process.c +++ b/kernel/power/process.c @@ -48,7 +48,7 @@ static int try_to_freeze_tasks(bool user_only) while (true) { todo = 0; - read_lock(&tasklist_lock); + rcu_read_lock(); for_each_process_thread(g, p) { if (p == current || !freeze_task(p)) continue; @@ -56,7 +56,7 @@ static int try_to_freeze_tasks(bool user_only) if (!freezer_should_skip(p)) todo++; } - read_unlock(&tasklist_lock); + rcu_read_unlock(); if (!user_only) { wq_busy = freeze_workqueues_busy(); @@ -97,13 +97,13 @@ static int try_to_freeze_tasks(bool user_only) show_workqueue_state(); if (!wakeup || pm_debug_messages_on) { - read_lock(&tasklist_lock); + rcu_read_lock(); for_each_process_thread(g, p) { if (p != current && !freezer_should_skip(p) && freezing(p) && !frozen(p)) sched_show_task(p); } - read_unlock(&tasklist_lock); + rcu_read_unlock(); } } else { pr_cont("(elapsed %d.%03d seconds) ", elapsed_msecs / 1000, @@ -206,13 +206,13 @@ void thaw_processes(void) cpuset_wait_for_hotplug(); - read_lock(&tasklist_lock); + rcu_read_lock(); for_each_process_thread(g, p) { /* No other threads should have PF_SUSPEND_TASK set */ WARN_ON((p != curr) && (p->flags & PF_SUSPEND_TASK)); __thaw_task(p); } - read_unlock(&tasklist_lock); + rcu_read_unlock(); WARN_ON(!(curr->flags & PF_SUSPEND_TASK)); curr->flags &= ~PF_SUSPEND_TASK; @@ -233,12 +233,12 @@ void thaw_kernel_threads(void) thaw_workqueues(); - read_lock(&tasklist_lock); + rcu_read_lock(); for_each_process_thread(g, p) { if (p->flags & PF_KTHREAD) __thaw_task(p); } - read_unlock(&tasklist_lock); + rcu_read_unlock(); schedule(); pr_cont("done.\n"); -- 2.17.1 ^ permalink raw reply related [flat|nested] 7+ messages in thread
* Re: [PATCH] PM: sleep: Replace read_lock/unlock(tasklist_lock) with rcu_read_lock/unlock() 2021-06-07 6:57 [PATCH] PM: sleep: Replace read_lock/unlock(tasklist_lock) with rcu_read_lock/unlock() qiang.zhang @ 2021-06-07 11:45 ` Rafael J. Wysocki 2021-06-07 14:11 ` Zhang, Qiang 0 siblings, 1 reply; 7+ messages in thread From: Rafael J. Wysocki @ 2021-06-07 11:45 UTC (permalink / raw) To: qiang.zhang Cc: Rafael Wysocki, Len Brown, Pavel Machek, Paul E. McKenney, Linux PM, Linux Kernel Mailing List On Mon, Jun 7, 2021 at 8:57 AM <qiang.zhang@windriver.com> wrote: > > From: Zqiang <qiang.zhang@windriver.com> > > Using rcu_read_lock/unlock() instead of read_lock/unlock(tasklist_lock), > the task list can be traversed in parallel to any list additions or > removals, improve concurrency. > > Signed-off-by: Zqiang <qiang.zhang@windriver.com> This changes the reader side only AFAICS, but what about the writer side? What exactly is there to ensure that the updates of the list will remain safe after this change? > --- > kernel/power/process.c | 16 ++++++++-------- > 1 file changed, 8 insertions(+), 8 deletions(-) > > diff --git a/kernel/power/process.c b/kernel/power/process.c > index 50cc63534486..0f8dee9ee097 100644 > --- a/kernel/power/process.c > +++ b/kernel/power/process.c > @@ -48,7 +48,7 @@ static int try_to_freeze_tasks(bool user_only) > > while (true) { > todo = 0; > - read_lock(&tasklist_lock); > + rcu_read_lock(); > for_each_process_thread(g, p) { > if (p == current || !freeze_task(p)) > continue; > @@ -56,7 +56,7 @@ static int try_to_freeze_tasks(bool user_only) > if (!freezer_should_skip(p)) > todo++; > } > - read_unlock(&tasklist_lock); > + rcu_read_unlock(); > > if (!user_only) { > wq_busy = freeze_workqueues_busy(); > @@ -97,13 +97,13 @@ static int try_to_freeze_tasks(bool user_only) > show_workqueue_state(); > > if (!wakeup || pm_debug_messages_on) { > - read_lock(&tasklist_lock); > + rcu_read_lock(); > for_each_process_thread(g, p) { > if (p != current && !freezer_should_skip(p) > && freezing(p) && !frozen(p)) > sched_show_task(p); > } > - read_unlock(&tasklist_lock); > + rcu_read_unlock(); > } > } else { > pr_cont("(elapsed %d.%03d seconds) ", elapsed_msecs / 1000, > @@ -206,13 +206,13 @@ void thaw_processes(void) > > cpuset_wait_for_hotplug(); > > - read_lock(&tasklist_lock); > + rcu_read_lock(); > for_each_process_thread(g, p) { > /* No other threads should have PF_SUSPEND_TASK set */ > WARN_ON((p != curr) && (p->flags & PF_SUSPEND_TASK)); > __thaw_task(p); > } > - read_unlock(&tasklist_lock); > + rcu_read_unlock(); > > WARN_ON(!(curr->flags & PF_SUSPEND_TASK)); > curr->flags &= ~PF_SUSPEND_TASK; > @@ -233,12 +233,12 @@ void thaw_kernel_threads(void) > > thaw_workqueues(); > > - read_lock(&tasklist_lock); > + rcu_read_lock(); > for_each_process_thread(g, p) { > if (p->flags & PF_KTHREAD) > __thaw_task(p); > } > - read_unlock(&tasklist_lock); > + rcu_read_unlock(); > > schedule(); > pr_cont("done.\n"); > -- > 2.17.1 > ^ permalink raw reply [flat|nested] 7+ messages in thread
* Re: [PATCH] PM: sleep: Replace read_lock/unlock(tasklist_lock) with rcu_read_lock/unlock() 2021-06-07 11:45 ` Rafael J. Wysocki @ 2021-06-07 14:11 ` Zhang, Qiang 2021-06-07 15:41 ` Paul E. McKenney 0 siblings, 1 reply; 7+ messages in thread From: Zhang, Qiang @ 2021-06-07 14:11 UTC (permalink / raw) To: Rafael J. Wysocki, paulmck Cc: Rafael Wysocki, Len Brown, Pavel Machek, Linux PM, Linux Kernel Mailing List ________________________________________ From: Rafael J. Wysocki <rafael@kernel.org> Sent: Monday, 7 June 2021 19:45 To: Zhang, Qiang Cc: Rafael Wysocki; Len Brown; Pavel Machek; Paul E. McKenney; Linux PM; Linux Kernel Mailing List Subject: Re: [PATCH] PM: sleep: Replace read_lock/unlock(tasklist_lock) with rcu_read_lock/unlock() [Please note: This e-mail is from an EXTERNAL e-mail address] On Mon, Jun 7, 2021 at 8:57 AM <qiang.zhang@windriver.com> wrote: > > From: Zqiang <qiang.zhang@windriver.com> > > Using rcu_read_lock/unlock() instead of read_lock/unlock(tasklist_lock), > the task list can be traversed in parallel to any list additions or > removals, improve concurrency. > > Signed-off-by: Zqiang <qiang.zhang@windriver.com> >This changes the reader side only AFAICS, but what about the >writer side? Hello rafael In the case of holding read_lock(tasklist_lock) if there are a lot of tasks in the system that need to be frozen, the read_lock(tasklist_lock) has not been released for a short time, when clone task, we will acquire write_lock(tasklist_lock) and insert task to list, if Ifcan't get the write lock all the time. the hung task may be trigger and report warning. for example: copy_process() {........ write_lock_irq(&tasklist_lock); ........ list_add_tail_rcu(&p->tasks, &init_task.tasks); } > >What exactly is there to ensure that the updates of the list will >remain safe after this change? The RCU can guarantee that the list can be traversed and added at the same time, and in the read critical area, it is guaranteed that the task structure will not be released. In ./Documentation/RCU/listRCU.rst , it is also explained. Maybe we can ask Paul E, Paul E what's your opinion ? Thanks Qiang > --- > kernel/power/process.c | 16 ++++++++-------- > 1 file changed, 8 insertions(+), 8 deletions(-) > > diff --git a/kernel/power/process.c b/kernel/power/process.c > index 50cc63534486..0f8dee9ee097 100644 > --- a/kernel/power/process.c > +++ b/kernel/power/process.c > @@ -48,7 +48,7 @@ static int try_to_freeze_tasks(bool user_only) > > while (true) { > todo = 0; > - read_lock(&tasklist_lock); > + rcu_read_lock(); > for_each_process_thread(g, p) { > if (p == current || !freeze_task(p)) > continue; > @@ -56,7 +56,7 @@ static int try_to_freeze_tasks(bool user_only) > if (!freezer_should_skip(p)) > todo++; > } > - read_unlock(&tasklist_lock); > + rcu_read_unlock(); > > if (!user_only) { > wq_busy = freeze_workqueues_busy(); > @@ -97,13 +97,13 @@ static int try_to_freeze_tasks(bool user_only) > show_workqueue_state(); > > if (!wakeup || pm_debug_messages_on) { > - read_lock(&tasklist_lock); > + rcu_read_lock(); > for_each_process_thread(g, p) { > if (p != current && !freezer_should_skip(p) > && freezing(p) && !frozen(p)) > sched_show_task(p); > } > - read_unlock(&tasklist_lock); > + rcu_read_unlock(); > } > } else { > pr_cont("(elapsed %d.%03d seconds) ", elapsed_msecs / 1000, > @@ -206,13 +206,13 @@ void thaw_processes(void) > > cpuset_wait_for_hotplug(); > > - read_lock(&tasklist_lock); > + rcu_read_lock(); > for_each_process_thread(g, p) { > /* No other threads should have PF_SUSPEND_TASK set */ > WARN_ON((p != curr) && (p->flags & PF_SUSPEND_TASK)); > __thaw_task(p); > } > - read_unlock(&tasklist_lock); > + rcu_read_unlock(); > > WARN_ON(!(curr->flags & PF_SUSPEND_TASK)); > curr->flags &= ~PF_SUSPEND_TASK; > @@ -233,12 +233,12 @@ void thaw_kernel_threads(void) > > thaw_workqueues(); > > - read_lock(&tasklist_lock); > + rcu_read_lock(); > for_each_process_thread(g, p) { > if (p->flags & PF_KTHREAD) > __thaw_task(p); > } > - read_unlock(&tasklist_lock); > + rcu_read_unlock(); > > schedule(); > pr_cont("done.\n"); > -- > 2.17.1 > ^ permalink raw reply [flat|nested] 7+ messages in thread
* Re: [PATCH] PM: sleep: Replace read_lock/unlock(tasklist_lock) with rcu_read_lock/unlock() 2021-06-07 14:11 ` Zhang, Qiang @ 2021-06-07 15:41 ` Paul E. McKenney 2021-06-08 2:50 ` Zhang, Qiang 0 siblings, 1 reply; 7+ messages in thread From: Paul E. McKenney @ 2021-06-07 15:41 UTC (permalink / raw) To: Zhang, Qiang Cc: Rafael J. Wysocki, Rafael Wysocki, Len Brown, Pavel Machek, Linux PM, Linux Kernel Mailing List On Mon, Jun 07, 2021 at 02:11:37PM +0000, Zhang, Qiang wrote: > ________________________________________ > From: Rafael J. Wysocki <rafael@kernel.org> > Sent: Monday, 7 June 2021 19:45 > To: Zhang, Qiang > Cc: Rafael Wysocki; Len Brown; Pavel Machek; Paul E. McKenney; Linux PM; Linux Kernel Mailing List > Subject: Re: [PATCH] PM: sleep: Replace read_lock/unlock(tasklist_lock) with rcu_read_lock/unlock() > > [Please note: This e-mail is from an EXTERNAL e-mail address] > > On Mon, Jun 7, 2021 at 8:57 AM <qiang.zhang@windriver.com> wrote: > > > > From: Zqiang <qiang.zhang@windriver.com> > > > > Using rcu_read_lock/unlock() instead of read_lock/unlock(tasklist_lock), > > the task list can be traversed in parallel to any list additions or > > removals, improve concurrency. > > > > Signed-off-by: Zqiang <qiang.zhang@windriver.com> > > >This changes the reader side only AFAICS, but what about the >writer side? > > Hello rafael > > In the case of holding read_lock(tasklist_lock) > if there are a lot of tasks in the system that need to be frozen, > the read_lock(tasklist_lock) has not been released for a short time, > when clone task, we will acquire write_lock(tasklist_lock) and insert task to list, if Ifcan't get the write lock all the time. the hung task > may be trigger and report warning. > > for example: > > copy_process() > {........ > write_lock_irq(&tasklist_lock); > ........ > list_add_tail_rcu(&p->tasks, &init_task.tasks); > } > > > > >What exactly is there to ensure that the updates of the list will > >remain safe after this change? > > The RCU can guarantee that the list can be traversed and added at the same time, and in the read critical area, it is guaranteed that the task structure will not be released. > > In ./Documentation/RCU/listRCU.rst , it is also explained. > > Maybe we can ask Paul E, Paul E what's your opinion ? Although RCU does guarantee that the task structures are not freed while an RCU reader references them, that is only one of the questions that must be answered. Changing from read_lock(&tasklist_lock) to rcu_read_lock() also allows all of the code currently under write_lock(&tasklist_lock) to execute concurrently with these newly created RCU readers. Is that safe? If it is safe, why? For example, the first pair of hunks in the diff -might- be safe because "todo" is updated while read-holding tasklist_lock, and then that value is used after releasing that lock. But I do not know this code well, and it is quite possible that something else is going on. Thanx, Paul > Thanks > Qiang > > > --- > > kernel/power/process.c | 16 ++++++++-------- > > 1 file changed, 8 insertions(+), 8 deletions(-) > > > > diff --git a/kernel/power/process.c b/kernel/power/process.c > > index 50cc63534486..0f8dee9ee097 100644 > > --- a/kernel/power/process.c > > +++ b/kernel/power/process.c > > @@ -48,7 +48,7 @@ static int try_to_freeze_tasks(bool user_only) > > > > while (true) { > > todo = 0; > > - read_lock(&tasklist_lock); > > + rcu_read_lock(); > > for_each_process_thread(g, p) { > > if (p == current || !freeze_task(p)) > > continue; > > @@ -56,7 +56,7 @@ static int try_to_freeze_tasks(bool user_only) > > if (!freezer_should_skip(p)) > > todo++; > > } > > - read_unlock(&tasklist_lock); > > + rcu_read_unlock(); > > > > if (!user_only) { > > wq_busy = freeze_workqueues_busy(); > > @@ -97,13 +97,13 @@ static int try_to_freeze_tasks(bool user_only) > > show_workqueue_state(); > > > > if (!wakeup || pm_debug_messages_on) { > > - read_lock(&tasklist_lock); > > + rcu_read_lock(); > > for_each_process_thread(g, p) { > > if (p != current && !freezer_should_skip(p) > > && freezing(p) && !frozen(p)) > > sched_show_task(p); > > } > > - read_unlock(&tasklist_lock); > > + rcu_read_unlock(); > > } > > } else { > > pr_cont("(elapsed %d.%03d seconds) ", elapsed_msecs / 1000, > > @@ -206,13 +206,13 @@ void thaw_processes(void) > > > > cpuset_wait_for_hotplug(); > > > > - read_lock(&tasklist_lock); > > + rcu_read_lock(); > > for_each_process_thread(g, p) { > > /* No other threads should have PF_SUSPEND_TASK set */ > > WARN_ON((p != curr) && (p->flags & PF_SUSPEND_TASK)); > > __thaw_task(p); > > } > > - read_unlock(&tasklist_lock); > > + rcu_read_unlock(); > > > > WARN_ON(!(curr->flags & PF_SUSPEND_TASK)); > > curr->flags &= ~PF_SUSPEND_TASK; > > @@ -233,12 +233,12 @@ void thaw_kernel_threads(void) > > > > thaw_workqueues(); > > > > - read_lock(&tasklist_lock); > > + rcu_read_lock(); > > for_each_process_thread(g, p) { > > if (p->flags & PF_KTHREAD) > > __thaw_task(p); > > } > > - read_unlock(&tasklist_lock); > > + rcu_read_unlock(); > > > > schedule(); > > pr_cont("done.\n"); > > -- > > 2.17.1 > > ^ permalink raw reply [flat|nested] 7+ messages in thread
* Re: [PATCH] PM: sleep: Replace read_lock/unlock(tasklist_lock) with rcu_read_lock/unlock() 2021-06-07 15:41 ` Paul E. McKenney @ 2021-06-08 2:50 ` Zhang, Qiang 2021-06-08 4:56 ` Paul E. McKenney 0 siblings, 1 reply; 7+ messages in thread From: Zhang, Qiang @ 2021-06-08 2:50 UTC (permalink / raw) To: paulmck Cc: Rafael J. Wysocki, Rafael Wysocki, Len Brown, Pavel Machek, Linux PM, Linux Kernel Mailing List ________________________________________ From: Paul E. McKenney <paulmck@kernel.org> Sent: Monday, 7 June 2021 23:41 To: Zhang, Qiang Cc: Rafael J. Wysocki; Rafael Wysocki; Len Brown; Pavel Machek; Linux PM; Linux Kernel Mailing List Subject: Re: [PATCH] PM: sleep: Replace read_lock/unlock(tasklist_lock) with rcu_read_lock/unlock() [Please note: This e-mail is from an EXTERNAL e-mail address] On Mon, Jun 07, 2021 at 02:11:37PM +0000, Zhang, Qiang wrote: > ________________________________________ > From: Rafael J. Wysocki <rafael@kernel.org> > Sent: Monday, 7 June 2021 19:45 > To: Zhang, Qiang > Cc: Rafael Wysocki; Len Brown; Pavel Machek; Paul E. McKenney; Linux PM; Linux Kernel Mailing List > Subject: Re: [PATCH] PM: sleep: Replace read_lock/unlock(tasklist_lock) with rcu_read_lock/unlock() > > [Please note: This e-mail is from an EXTERNAL e-mail address] > > On Mon, Jun 7, 2021 at 8:57 AM <qiang.zhang@windriver.com> wrote: > > > > From: Zqiang <qiang.zhang@windriver.com> > > > > Using rcu_read_lock/unlock() instead of read_lock/unlock(tasklist_lock), > > the task list can be traversed in parallel to any list additions or > > removals, improve concurrency. > > > > Signed-off-by: Zqiang <qiang.zhang@windriver.com> > > >This changes the reader side only AFAICS, but what about the >writer side? > > Hello rafael > > In the case of holding read_lock(tasklist_lock) > if there are a lot of tasks in the system that need to be frozen, > the read_lock(tasklist_lock) has not been released for a short time, > when clone task, we will acquire write_lock(tasklist_lock) and insert task to list, if Ifcan't get the write lock all the time. the hung task > may be trigger and report warning. > > for example: > > copy_process() > {........ > write_lock_irq(&tasklist_lock); > ........ > list_add_tail_rcu(&p->tasks, &init_task.tasks); > } > > > > >What exactly is there to ensure that the updates of the list will > >remain safe after this change? > > The RCU can guarantee that the list can be traversed and added at the same time, and in the read critical area, it is guaranteed that the task structure will not be released. > > In ./Documentation/RCU/listRCU.rst , it is also explained. > > Maybe we can ask Paul E, Paul E what's your opinion ? >Although RCU does guarantee that the task structures are not freed while >an RCU reader references them, that is only one of the questions that >must be answered. > >Changing from read_lock(&tasklist_lock) to rcu_read_lock() also allows >all of the code currently under write_lock(&tasklist_lock) to execute >concurrently with these newly created RCU readers. Is that safe? >If it is safe, why? > >For example, the first pair of hunks in the diff -might- be safe because >"todo" is updated while read-holding tasklist_lock, and then that value >is used after releasing that lock. But I do not know this code well, >and it is quite possible that something else is going on. Thanks Pual E you mean is the rcu ensures that the task_struct is not released but it does not guarantee that some members of the task_struct object are valid. There may be the following scenarios: In rcu read critical area, when accessing some members of task_struct object , it may have been released, when the task exiting. Is my understanding correct? Qiang > > Thanx, Paul > Thanks > Qiang > > > --- > > kernel/power/process.c | 16 ++++++++-------- > > 1 file changed, 8 insertions(+), 8 deletions(-) > > > > diff --git a/kernel/power/process.c b/kernel/power/process.c > > index 50cc63534486..0f8dee9ee097 100644 > > --- a/kernel/power/process.c > > +++ b/kernel/power/process.c > > @@ -48,7 +48,7 @@ static int try_to_freeze_tasks(bool user_only) > > > > while (true) { > > todo = 0; > > - read_lock(&tasklist_lock); > > + rcu_read_lock(); > > for_each_process_thread(g, p) { > > if (p == current || !freeze_task(p)) > > continue; > > @@ -56,7 +56,7 @@ static int try_to_freeze_tasks(bool user_only) > > if (!freezer_should_skip(p)) > > todo++; > > } > > - read_unlock(&tasklist_lock); > > + rcu_read_unlock(); > > > > if (!user_only) { > > wq_busy = freeze_workqueues_busy(); > > @@ -97,13 +97,13 @@ static int try_to_freeze_tasks(bool user_only) > > show_workqueue_state(); > > > > if (!wakeup || pm_debug_messages_on) { > > - read_lock(&tasklist_lock); > > + rcu_read_lock(); > > for_each_process_thread(g, p) { > > if (p != current && !freezer_should_skip(p) > > && freezing(p) && !frozen(p)) > > sched_show_task(p); > > } > > - read_unlock(&tasklist_lock); > > + rcu_read_unlock(); > > } > > } else { > > pr_cont("(elapsed %d.%03d seconds) ", elapsed_msecs / 1000, > > @@ -206,13 +206,13 @@ void thaw_processes(void) > > > > cpuset_wait_for_hotplug(); > > > > - read_lock(&tasklist_lock); > > + rcu_read_lock(); > > for_each_process_thread(g, p) { > > /* No other threads should have PF_SUSPEND_TASK set */ > > WARN_ON((p != curr) && (p->flags & PF_SUSPEND_TASK)); > > __thaw_task(p); > > } > > - read_unlock(&tasklist_lock); > > + rcu_read_unlock(); > > > > WARN_ON(!(curr->flags & PF_SUSPEND_TASK)); > > curr->flags &= ~PF_SUSPEND_TASK; > > @@ -233,12 +233,12 @@ void thaw_kernel_threads(void) > > > > thaw_workqueues(); > > > > - read_lock(&tasklist_lock); > > + rcu_read_lock(); > > for_each_process_thread(g, p) { > > if (p->flags & PF_KTHREAD) > > __thaw_task(p); > > } > > - read_unlock(&tasklist_lock); > > + rcu_read_unlock(); > > > > schedule(); > > pr_cont("done.\n"); > > -- > > 2.17.1 > > ^ permalink raw reply [flat|nested] 7+ messages in thread
* Re: [PATCH] PM: sleep: Replace read_lock/unlock(tasklist_lock) with rcu_read_lock/unlock() 2021-06-08 2:50 ` Zhang, Qiang @ 2021-06-08 4:56 ` Paul E. McKenney 2021-06-09 7:04 ` Zhang, Qiang 0 siblings, 1 reply; 7+ messages in thread From: Paul E. McKenney @ 2021-06-08 4:56 UTC (permalink / raw) To: Zhang, Qiang Cc: Rafael J. Wysocki, Rafael Wysocki, Len Brown, Pavel Machek, Linux PM, Linux Kernel Mailing List On Tue, Jun 08, 2021 at 02:50:26AM +0000, Zhang, Qiang wrote: > > > ________________________________________ > From: Paul E. McKenney <paulmck@kernel.org> > Sent: Monday, 7 June 2021 23:41 > To: Zhang, Qiang > Cc: Rafael J. Wysocki; Rafael Wysocki; Len Brown; Pavel Machek; Linux PM; Linux Kernel Mailing List > Subject: Re: [PATCH] PM: sleep: Replace read_lock/unlock(tasklist_lock) with rcu_read_lock/unlock() > > [Please note: This e-mail is from an EXTERNAL e-mail address] > > On Mon, Jun 07, 2021 at 02:11:37PM +0000, Zhang, Qiang wrote: > > ________________________________________ > > From: Rafael J. Wysocki <rafael@kernel.org> > > Sent: Monday, 7 June 2021 19:45 > > To: Zhang, Qiang > > Cc: Rafael Wysocki; Len Brown; Pavel Machek; Paul E. McKenney; Linux PM; Linux Kernel Mailing List > > Subject: Re: [PATCH] PM: sleep: Replace read_lock/unlock(tasklist_lock) with rcu_read_lock/unlock() > > > > [Please note: This e-mail is from an EXTERNAL e-mail address] > > > > On Mon, Jun 7, 2021 at 8:57 AM <qiang.zhang@windriver.com> wrote: > > > > > > From: Zqiang <qiang.zhang@windriver.com> > > > > > > Using rcu_read_lock/unlock() instead of read_lock/unlock(tasklist_lock), > > > the task list can be traversed in parallel to any list additions or > > > removals, improve concurrency. > > > > > > Signed-off-by: Zqiang <qiang.zhang@windriver.com> > > > > >This changes the reader side only AFAICS, but what about the >writer side? > > > > Hello rafael > > > > In the case of holding read_lock(tasklist_lock) > > if there are a lot of tasks in the system that need to be frozen, > > the read_lock(tasklist_lock) has not been released for a short time, > > when clone task, we will acquire write_lock(tasklist_lock) and insert task to list, if Ifcan't get the write lock all the time. the hung task > > may be trigger and report warning. > > > > for example: > > > > copy_process() > > {........ > > write_lock_irq(&tasklist_lock); > > ........ > > list_add_tail_rcu(&p->tasks, &init_task.tasks); > > } > > > > > > > >What exactly is there to ensure that the updates of the list will > > >remain safe after this change? > > > > The RCU can guarantee that the list can be traversed and added at the same time, and in the read critical area, it is guaranteed that the task structure will not be released. > > > > In ./Documentation/RCU/listRCU.rst , it is also explained. > > > > Maybe we can ask Paul E, Paul E what's your opinion ? > > >Although RCU does guarantee that the task structures are not freed while > >an RCU reader references them, that is only one of the questions that > >must be answered. > > > >Changing from read_lock(&tasklist_lock) to rcu_read_lock() also allows > >all of the code currently under write_lock(&tasklist_lock) to execute > >concurrently with these newly created RCU readers. Is that safe? > >If it is safe, why? > > > >For example, the first pair of hunks in the diff -might- be safe because > >"todo" is updated while read-holding tasklist_lock, and then that value > >is used after releasing that lock. But I do not know this code well, > >and it is quite possible that something else is going on. > > Thanks Pual E > > you mean is the rcu ensures that the task_struct is not released but it does not guarantee that some members of the task_struct object are valid. > There may be the following scenarios: > In rcu read critical area, when accessing some members of task_struct object , it may have been released, when the task exiting. > > Is my understanding correct? Yes, that is a concern. The freezer code would need to be checked carefully to see whether or not the transformation from read_lock() to rcu_read_lock() was safe. Thanx, Paul > Qiang > > > > > Thanx, Paul > > > Thanks > > Qiang > > > > > --- > > > kernel/power/process.c | 16 ++++++++-------- > > > 1 file changed, 8 insertions(+), 8 deletions(-) > > > > > > diff --git a/kernel/power/process.c b/kernel/power/process.c > > > index 50cc63534486..0f8dee9ee097 100644 > > > --- a/kernel/power/process.c > > > +++ b/kernel/power/process.c > > > @@ -48,7 +48,7 @@ static int try_to_freeze_tasks(bool user_only) > > > > > > while (true) { > > > todo = 0; > > > - read_lock(&tasklist_lock); > > > + rcu_read_lock(); > > > for_each_process_thread(g, p) { > > > if (p == current || !freeze_task(p)) > > > continue; > > > @@ -56,7 +56,7 @@ static int try_to_freeze_tasks(bool user_only) > > > if (!freezer_should_skip(p)) > > > todo++; > > > } > > > - read_unlock(&tasklist_lock); > > > + rcu_read_unlock(); > > > > > > if (!user_only) { > > > wq_busy = freeze_workqueues_busy(); > > > @@ -97,13 +97,13 @@ static int try_to_freeze_tasks(bool user_only) > > > show_workqueue_state(); > > > > > > if (!wakeup || pm_debug_messages_on) { > > > - read_lock(&tasklist_lock); > > > + rcu_read_lock(); > > > for_each_process_thread(g, p) { > > > if (p != current && !freezer_should_skip(p) > > > && freezing(p) && !frozen(p)) > > > sched_show_task(p); > > > } > > > - read_unlock(&tasklist_lock); > > > + rcu_read_unlock(); > > > } > > > } else { > > > pr_cont("(elapsed %d.%03d seconds) ", elapsed_msecs / 1000, > > > @@ -206,13 +206,13 @@ void thaw_processes(void) > > > > > > cpuset_wait_for_hotplug(); > > > > > > - read_lock(&tasklist_lock); > > > + rcu_read_lock(); > > > for_each_process_thread(g, p) { > > > /* No other threads should have PF_SUSPEND_TASK set */ > > > WARN_ON((p != curr) && (p->flags & PF_SUSPEND_TASK)); > > > __thaw_task(p); > > > } > > > - read_unlock(&tasklist_lock); > > > + rcu_read_unlock(); > > > > > > WARN_ON(!(curr->flags & PF_SUSPEND_TASK)); > > > curr->flags &= ~PF_SUSPEND_TASK; > > > @@ -233,12 +233,12 @@ void thaw_kernel_threads(void) > > > > > > thaw_workqueues(); > > > > > > - read_lock(&tasklist_lock); > > > + rcu_read_lock(); > > > for_each_process_thread(g, p) { > > > if (p->flags & PF_KTHREAD) > > > __thaw_task(p); > > > } > > > - read_unlock(&tasklist_lock); > > > + rcu_read_unlock(); > > > > > > schedule(); > > > pr_cont("done.\n"); > > > -- > > > 2.17.1 > > > ^ permalink raw reply [flat|nested] 7+ messages in thread
* Re: [PATCH] PM: sleep: Replace read_lock/unlock(tasklist_lock) with rcu_read_lock/unlock() 2021-06-08 4:56 ` Paul E. McKenney @ 2021-06-09 7:04 ` Zhang, Qiang 0 siblings, 0 replies; 7+ messages in thread From: Zhang, Qiang @ 2021-06-09 7:04 UTC (permalink / raw) To: Rafael J. Wysocki Cc: Len Brown, Pavel Machek, Linux PM, Linux Kernel Mailing List, paulmck ________________________________________ From: Paul E. McKenney <paulmck@kernel.org> Sent: Tuesday, 8 June 2021 12:56 To: Zhang, Qiang Cc: Rafael J. Wysocki; Rafael Wysocki; Len Brown; Pavel Machek; Linux PM; Linux Kernel Mailing List Subject: Re: [PATCH] PM: sleep: Replace read_lock/unlock(tasklist_lock) with rcu_read_lock/unlock() [Please note: This e-mail is from an EXTERNAL e-mail address] On Tue, Jun 08, 2021 at 02:50:26AM +0000, Zhang, Qiang wrote: > > > ________________________________________ > From: Paul E. McKenney <paulmck@kernel.org> > Sent: Monday, 7 June 2021 23:41 > To: Zhang, Qiang > Cc: Rafael J. Wysocki; Rafael Wysocki; Len Brown; Pavel Machek; Linux PM; Linux Kernel Mailing List > Subject: Re: [PATCH] PM: sleep: Replace read_lock/unlock(tasklist_lock) with rcu_read_lock/unlock() > > [Please note: This e-mail is from an EXTERNAL e-mail address] > > On Mon, Jun 07, 2021 at 02:11:37PM +0000, Zhang, Qiang wrote: > > ________________________________________ > > From: Rafael J. Wysocki <rafael@kernel.org> > > Sent: Monday, 7 June 2021 19:45 > > To: Zhang, Qiang > > Cc: Rafael Wysocki; Len Brown; Pavel Machek; Paul E. McKenney; Linux PM; Linux Kernel Mailing List > > Subject: Re: [PATCH] PM: sleep: Replace read_lock/unlock(tasklist_lock) with rcu_read_lock/unlock() > > > > [Please note: This e-mail is from an EXTERNAL e-mail address] > > > > On Mon, Jun 7, 2021 at 8:57 AM <qiang.zhang@windriver.com> wrote: > > > > > > From: Zqiang <qiang.zhang@windriver.com> > > > > > > Using rcu_read_lock/unlock() instead of read_lock/unlock(tasklist_lock), > > > the task list can be traversed in parallel to any list additions or > > > removals, improve concurrency. > > > > > > Signed-off-by: Zqiang <qiang.zhang@windriver.com> > > > > >This changes the reader side only AFAICS, but what about the >writer side? > > > > Hello rafael > > > > In the case of holding read_lock(tasklist_lock) > > if there are a lot of tasks in the system that need to be frozen, > > the read_lock(tasklist_lock) has not been released for a short time, > > when clone task, we will acquire write_lock(tasklist_lock) and insert task to list, if Ifcan't get the write lock all the time. the hung task > > may be trigger and report warning. > > > > for example: > > > > copy_process() > > {........ > > write_lock_irq(&tasklist_lock); > > ........ > > list_add_tail_rcu(&p->tasks, &init_task.tasks); > > } > > > > > > > >What exactly is there to ensure that the updates of the list will > > >remain safe after this change? > > > > The RCU can guarantee that the list can be traversed and added at the same time, and in the read critical area, it is guaranteed that the task structure will not be released. > > > > In ./Documentation/RCU/listRCU.rst , it is also explained. > > > > Maybe we can ask Paul E, Paul E what's your opinion ? > > >Although RCU does guarantee that the task structures are not freed while > >an RCU reader references them, that is only one of the questions that > >must be answered. > > > >Changing from read_lock(&tasklist_lock) to rcu_read_lock() also allows > >all of the code currently under write_lock(&tasklist_lock) to execute > >concurrently with these newly created RCU readers. Is that safe? > >If it is safe, why? > > > >For example, the first pair of hunks in the diff -might- be safe because > >"todo" is updated while read-holding tasklist_lock, and then that value > >is used after releasing that lock. But I do not know this code well, > >and it is quite possible that something else is going on. > > Thanks Pual E > > you mean is the rcu ensures that the task_struct is not released but it does not guarantee that some members of the task_struct object are valid. > There may be the following scenarios: > In rcu read critical area, when accessing some members of task_struct object , it may have been released, when the task exiting. > > Is my understanding correct? >Yes, that is a concern. The freezer code would need to be checked >carefully to see whether or not the transformation from read_lock() >to rcu_read_lock() was safe. Sorry for the noise I made, there is wrong with this patch I read the code, when accessing members of task_struct in the rcu read critical area cannot guarantee its effectiveness. Qiang > > Thanx, Paul > Qiang > > > > > Thanx, Paul > > > Thanks > > Qiang > > > > > --- > > > kernel/power/process.c | 16 ++++++++-------- > > > 1 file changed, 8 insertions(+), 8 deletions(-) > > > > > > diff --git a/kernel/power/process.c b/kernel/power/process.c > > > index 50cc63534486..0f8dee9ee097 100644 > > > --- a/kernel/power/process.c > > > +++ b/kernel/power/process.c > > > @@ -48,7 +48,7 @@ static int try_to_freeze_tasks(bool user_only) > > > > > > while (true) { > > > todo = 0; > > > - read_lock(&tasklist_lock); > > > + rcu_read_lock(); > > > for_each_process_thread(g, p) { > > > if (p == current || !freeze_task(p)) > > > continue; > > > @@ -56,7 +56,7 @@ static int try_to_freeze_tasks(bool user_only) > > > if (!freezer_should_skip(p)) > > > todo++; > > > } > > > - read_unlock(&tasklist_lock); > > > + rcu_read_unlock(); > > > > > > if (!user_only) { > > > wq_busy = freeze_workqueues_busy(); > > > @@ -97,13 +97,13 @@ static int try_to_freeze_tasks(bool user_only) > > > show_workqueue_state(); > > > > > > if (!wakeup || pm_debug_messages_on) { > > > - read_lock(&tasklist_lock); > > > + rcu_read_lock(); > > > for_each_process_thread(g, p) { > > > if (p != current && !freezer_should_skip(p) > > > && freezing(p) && !frozen(p)) > > > sched_show_task(p); > > > } > > > - read_unlock(&tasklist_lock); > > > + rcu_read_unlock(); > > > } > > > } else { > > > pr_cont("(elapsed %d.%03d seconds) ", elapsed_msecs / 1000, > > > @@ -206,13 +206,13 @@ void thaw_processes(void) > > > > > > cpuset_wait_for_hotplug(); > > > > > > - read_lock(&tasklist_lock); > > > + rcu_read_lock(); > > > for_each_process_thread(g, p) { > > > /* No other threads should have PF_SUSPEND_TASK set */ > > > WARN_ON((p != curr) && (p->flags & PF_SUSPEND_TASK)); > > > __thaw_task(p); > > > } > > > - read_unlock(&tasklist_lock); > > > + rcu_read_unlock(); > > > > > > WARN_ON(!(curr->flags & PF_SUSPEND_TASK)); > > > curr->flags &= ~PF_SUSPEND_TASK; > > > @@ -233,12 +233,12 @@ void thaw_kernel_threads(void) > > > > > > thaw_workqueues(); > > > > > > - read_lock(&tasklist_lock); > > > + rcu_read_lock(); > > > for_each_process_thread(g, p) { > > > if (p->flags & PF_KTHREAD) > > > __thaw_task(p); > > > } > > > - read_unlock(&tasklist_lock); > > > + rcu_read_unlock(); > > > > > > schedule(); > > > pr_cont("done.\n"); > > > -- > > > 2.17.1 > > > ^ permalink raw reply [flat|nested] 7+ messages in thread
end of thread, other threads:[~2021-06-09 7:05 UTC | newest] Thread overview: 7+ messages (download: mbox.gz / follow: Atom feed) -- links below jump to the message on this page -- 2021-06-07 6:57 [PATCH] PM: sleep: Replace read_lock/unlock(tasklist_lock) with rcu_read_lock/unlock() qiang.zhang 2021-06-07 11:45 ` Rafael J. Wysocki 2021-06-07 14:11 ` Zhang, Qiang 2021-06-07 15:41 ` Paul E. McKenney 2021-06-08 2:50 ` Zhang, Qiang 2021-06-08 4:56 ` Paul E. McKenney 2021-06-09 7:04 ` Zhang, Qiang
This is a public inbox, see mirroring instructions for how to clone and mirror all data and code used for this inbox; as well as URLs for NNTP newsgroup(s).