Prev: [RFC] Save/restore MISC_ENABLE register
Next: [PATCH] Staging: dt3155: fix different address spaces noise in dt3155_drv.c
From: Miles Lane on 7 Jun 2010 14:20 Hi All, I just reproduced a warning I reported quite a while ago.� Is a patch for this in the pipeline? [��� 0.167267] [ INFO: suspicious rcu_dereference_check() usage. ] [��� 0.167396] --------------------------------------------------- [��� 0.167526] include/linux/cgroup.h:534 invoked rcu_dereference_check() without protection! [��� 0.167728] [��� 0.167729] other info that might help us debug this: [��� 0.167731] [��� 0.168092] [��� 0.168093] rcu_scheduler_active = 1, debug_locks = 1 [��� 0.168337] no locks held by watchdog/0/5. [��� 0.168462] [��� 0.168463] stack backtrace: [��� 0.168704] Pid: 5, comm: watchdog/0 Not tainted 2.6.35-rc2-git1 #8 [��� 0.168834] Call Trace: [��� 0.168965]� [<ffffffff81064e9c>] lockdep_rcu_dereference+0x9d/0xa5 [��� 0.169100]� [<ffffffff8102c1ce>] task_subsys_state+0x59/0x70 [��� 0.169232]� [<ffffffff8103189b>] __sched_setscheduler+0x19d/0x2f8 [��� 0.169365]� [<ffffffff8102a5ef>] ? need_resched+0x1e/0x28 [��� 0.169497]� [<ffffffff813c7d01>] ? schedule+0x586/0x619 [��� 0.169628]� [<ffffffff81081c33>] ? watchdog+0x0/0x8c [��� 0.169758]� [<ffffffff81031a11>] sched_setscheduler+0xe/0x10 [��� 0.169889]� [<ffffffff81081c5d>] watchdog+0x2a/0x8c [��� 0.170010]� [<ffffffff81081c33>] ? watchdog+0x0/0x8c [��� 0.170141]� [<ffffffff81054a82>] kthread+0x89/0x91 [��� 0.170274]� [<ffffffff81003054>] kernel_thread_helper+0x4/0x10 [��� 0.170405]� [<ffffffff813ca480>] ? restore_args+0x0/0x30 [��� 0.170536]� [<ffffffff810549f9>] ? kthread+0x0/0x91 [��� 0.170667]� [<ffffffff81003050>] ? kernel_thread_helper+0x0/0x10 [ 0.176751] lockdep: fixing up alternatives. -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo(a)vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
From: Miles Lane on 8 Jun 2010 00:20 On Mon, Jun 7, 2010 at 8:19 PM, Paul E. McKenney <paulmck(a)linux.vnet.ibm.com> wrote: > On Mon, Jun 07, 2010 at 02:14:25PM -0400, Miles Lane wrote: >> Hi All, >> >> I just reproduced a warning I reported quite a while ago.� Is a patch >> for this in the pipeline? > > I proposed a patch, thinking that it was a false positive. �Peter Zijlstra > pointed out that there was a real race, and proposed an alternative patch, > which may be found at http://lkml.org/lkml/2010/4/22/603. > > Could you please test Peter's patch and let us know if it cures the problem? > > � � � � � � � � � � � � � � � � � � � � � � � � � � � �Thanx, Paul > >> [��� 0.167267] [ INFO: suspicious rcu_dereference_check() usage. ] >> [��� 0.167396] --------------------------------------------------- >> [��� 0.167526] include/linux/cgroup.h:534 invoked >> rcu_dereference_check() without protection! >> [��� 0.167728] >> [��� 0.167729] other info that might help us debug this: >> [��� 0.167731] >> [��� 0.168092] >> [��� 0.168093] rcu_scheduler_active = 1, debug_locks = 1 >> [��� 0.168337] no locks held by watchdog/0/5. >> [��� 0.168462] >> [��� 0.168463] stack backtrace: >> [��� 0.168704] Pid: 5, comm: watchdog/0 Not tainted 2.6.35-rc2-git1 #8 >> [��� 0.168834] Call Trace: >> [��� 0.168965]� [<ffffffff81064e9c>] lockdep_rcu_dereference+0x9d/0xa5 >> [��� 0.169100]� [<ffffffff8102c1ce>] task_subsys_state+0x59/0x70 >> [��� 0.169232]� [<ffffffff8103189b>] __sched_setscheduler+0x19d/0x2f8 >> [��� 0.169365]� [<ffffffff8102a5ef>] ? need_resched+0x1e/0x28 >> [��� 0.169497]� [<ffffffff813c7d01>] ? schedule+0x586/0x619 >> [��� 0.169628]� [<ffffffff81081c33>] ? watchdog+0x0/0x8c >> [��� 0.169758]� [<ffffffff81031a11>] sched_setscheduler+0xe/0x10 >> [��� 0.169889]� [<ffffffff81081c5d>] watchdog+0x2a/0x8c >> [��� 0.170010]� [<ffffffff81081c33>] ? watchdog+0x0/0x8c >> [��� 0.170141]� [<ffffffff81054a82>] kthread+0x89/0x91 >> [��� 0.170274]� [<ffffffff81003054>] kernel_thread_helper+0x4/0x10 >> [��� 0.170405]� [<ffffffff813ca480>] ? restore_args+0x0/0x30 >> [��� 0.170536]� [<ffffffff810549f9>] ? kthread+0x0/0x91 >> [��� 0.170667]� [<ffffffff81003050>] ? kernel_thread_helper+0x0/0x10 >> [ � �0.176751] lockdep: fixing up alternatives. > With the patch, I get: [ 0.151274] [ INFO: suspicious rcu_dereference_check() usage. ] [ 0.151390] --------------------------------------------------- [ 0.151520] include/linux/cgroup.h:534 invoked rcu_dereference_check() without protection! [ 0.151723] [ 0.151724] other info that might help us debug this: [ 0.151726] [ 0.151999] [ 0.151999] rcu_scheduler_active = 1, debug_locks = 1 [ 0.151999] 2 locks held by kthreadd/10: [ 0.151999] #0: (key){......}, at: [<ffffffff81036578>] complete+0x1c/0x4e [ 0.151999] #1: (&rq->lock){-.-...}, at: [<ffffffff81037875>] select_task_rq_fair+0x21f/0x791 [ 0.151999] [ 0.151999] stack backtrace: [ 0.151999] Pid: 10, comm: kthreadd Not tainted 2.6.35-rc2-git1 #11 [ 0.151999] Call Trace: [ 0.151999] [<ffffffff81070a45>] lockdep_rcu_dereference+0x9d/0xa5 [ 0.151999] [<ffffffff8103675e>] task_subsys_state+0x59/0x70 [ 0.151999] [<ffffffff8103799a>] select_task_rq_fair+0x344/0x791 [ 0.151999] [<ffffffff81037335>] ? task_rq_lock+0x68/0x9d [ 0.151999] [<ffffffff811d62f3>] ? do_raw_spin_lock+0x79/0x13e [ 0.151999] [<ffffffff81037335>] ? task_rq_lock+0x68/0x9d [ 0.151999] [<ffffffff8103ac1e>] select_task_rq+0x13/0x44 [ 0.151999] [<ffffffff810417c3>] try_to_wake_up+0xf2/0x37d [ 0.151999] [<ffffffff81041a5b>] default_wake_function+0xd/0xf [ 0.151999] [<ffffffff81034272>] __wake_up_common+0x49/0x7f [ 0.151999] [<ffffffff81036596>] complete+0x3a/0x4e [ 0.151999] [<ffffffff8105b598>] ? worker_thread+0x0/0x3a7 [ 0.151999] [<ffffffff8105f7d0>] kthread+0x73/0x91 [ 0.151999] [<ffffffff8100aba4>] kernel_thread_helper+0x4/0x10 [ 0.151999] [<ffffffff813e3694>] ? restore_args+0x0/0x30 [ 0.151999] [<ffffffff8105f75d>] ? kthread+0x0/0x91 [ 0.151999] [<ffffffff8100aba0>] ? kernel_thread_helper+0x0/0x10 -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo(a)vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
From: Miles Lane on 8 Jun 2010 09:20 On Tue, Jun 8, 2010 at 4:40 AM, Peter Zijlstra <peterz(a)infradead.org> wrote: > On Tue, 2010-06-08 at 00:16 -0400, Miles Lane wrote: >> On Mon, Jun 7, 2010 at 8:19 PM, Paul E. McKenney >> <paulmck(a)linux.vnet.ibm.com> wrote: >> > On Mon, Jun 07, 2010 at 02:14:25PM -0400, Miles Lane wrote: >> >> Hi All, >> >> >> >> I just reproduced a warning I reported quite a while ago. �Is a patch >> >> for this in the pipeline? >> > >> > I proposed a patch, thinking that it was a false positive. �Peter Zijlstra >> > pointed out that there was a real race, and proposed an alternative patch, >> > which may be found at http://lkml.org/lkml/2010/4/22/603. >> > >> > Could you please test Peter's patch and let us know if it cures the problem? >> > > > Gah, this task_group() stuff is annoying, how about something like the > below which teaches task_group() about the task_rq()->lock rule? > > --- > �include/linux/cgroup.h | � 20 +++++++++++---- > �kernel/sched.c � � � � | � 61 +++++++++++++++++++++++++---------------------- > �2 files changed, 46 insertions(+), 35 deletions(-) > > diff --git a/include/linux/cgroup.h b/include/linux/cgroup.h > index 0c62160..1efd212 100644 > --- a/include/linux/cgroup.h > +++ b/include/linux/cgroup.h > @@ -525,13 +525,21 @@ static inline struct cgroup_subsys_state *cgroup_subsys_state( > � � � �return cgrp->subsys[subsys_id]; > �} > > -static inline struct cgroup_subsys_state *task_subsys_state( > - � � � struct task_struct *task, int subsys_id) > +/* > + * function to get the cgroup_subsys_state which allows for extra > + * rcu_dereference_check() conditions, such as locks used during the > + * cgroup_subsys::attach() methods. > + */ > +#define task_subsys_state_check(task, subsys_id, __c) � � � � � � � � �\ > + � � � rcu_dereference_check(task->cgroups->subsys[subsys_id], � � � � \ > + � � � � � � � � � � � � � � rcu_read_lock_held() || � � � � � � � � � \ > + � � � � � � � � � � � � � � lockdep_is_held(&task->alloc_lock) || � � \ > + � � � � � � � � � � � � � � cgroup_lock_is_held() || (__c)) > + > +static inline struct cgroup_subsys_state * > +task_subsys_state(struct task_struct *task, int subsys_id) > �{ > - � � � return rcu_dereference_check(task->cgroups->subsys[subsys_id], > - � � � � � � � � � � � � � � � � � �rcu_read_lock_held() || > - � � � � � � � � � � � � � � � � � �lockdep_is_held(&task->alloc_lock) || > - � � � � � � � � � � � � � � � � � �cgroup_lock_is_held()); > + � � � return task_subsys_state_check(task, subsys_id, false); > �} > > �static inline struct cgroup* task_cgroup(struct task_struct *task, > diff --git a/kernel/sched.c b/kernel/sched.c > index f8b8996..e01bb45 100644 > --- a/kernel/sched.c > +++ b/kernel/sched.c > @@ -306,32 +306,26 @@ static int init_task_group_load = INIT_TASK_GROUP_LOAD; > �*/ > �struct task_group init_task_group; > > -/* return group to which a task belongs */ > +/* > + * Return the group to which this tasks belongs. > + * > + * We use task_subsys_state_check() and extend the RCU verification > + * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach() > + * holds that lock for each task it moves into the cgroup. Therefore > + * by holding that lock, we pin the task to the current cgroup. > + */ > �static inline struct task_group *task_group(struct task_struct *p) > �{ > - � � � struct task_group *tg; > + � � � struct cgroup_subsys_state *css; > > -#ifdef CONFIG_CGROUP_SCHED > - � � � tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id), > - � � � � � � � � � � � � � � � struct task_group, css); > -#else > - � � � tg = &init_task_group; > -#endif > - � � � return tg; > + � � � css = task_subsys_state_check(p, cpu_cgroup_subsys_id, > + � � � � � � � � � � � lockdep_is_held(&task_rq(p)->lock)); > + � � � return container_of(css, struct task_group, css); > �} > > �/* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */ > �static inline void set_task_rq(struct task_struct *p, unsigned int cpu) > �{ > - � � � /* > - � � � �* Strictly speaking this rcu_read_lock() is not needed since the > - � � � �* task_group is tied to the cgroup, which in turn can never go away > - � � � �* as long as there are tasks attached to it. > - � � � �* > - � � � �* However since task_group() uses task_subsys_state() which is an > - � � � �* rcu_dereference() user, this quiets CONFIG_PROVE_RCU. > - � � � �*/ > - � � � rcu_read_lock(); > �#ifdef CONFIG_FAIR_GROUP_SCHED > � � � �p->se.cfs_rq = task_group(p)->cfs_rq[cpu]; > � � � �p->se.parent = task_group(p)->se[cpu]; > @@ -341,7 +335,6 @@ static inline void set_task_rq(struct task_struct *p, unsigned int cpu) > � � � �p->rt.rt_rq �= task_group(p)->rt_rq[cpu]; > � � � �p->rt.parent = task_group(p)->rt_se[cpu]; > �#endif > - � � � rcu_read_unlock(); > �} > > �#else > @@ -4465,16 +4458,6 @@ recheck: > � � � �} > > � � � �if (user) { > -#ifdef CONFIG_RT_GROUP_SCHED > - � � � � � � � /* > - � � � � � � � �* Do not allow realtime tasks into groups that have no runtime > - � � � � � � � �* assigned. > - � � � � � � � �*/ > - � � � � � � � if (rt_bandwidth_enabled() && rt_policy(policy) && > - � � � � � � � � � � � � � � � task_group(p)->rt_bandwidth.rt_runtime == 0) > - � � � � � � � � � � � return -EPERM; > -#endif > - > � � � � � � � �retval = security_task_setscheduler(p, policy, param); > � � � � � � � �if (retval) > � � � � � � � � � � � �return retval; > @@ -4490,6 +4473,26 @@ recheck: > � � � � * runqueue lock must be held. > � � � � */ > � � � �rq = __task_rq_lock(p); > + > + � � � retval = 0; > +#ifdef CONFIG_RT_GROUP_SCHED > + � � � if (user) { > + � � � � � � � /* > + � � � � � � � �* Do not allow realtime tasks into groups that have no runtime > + � � � � � � � �* assigned. > + � � � � � � � �*/ > + � � � � � � � if (rt_bandwidth_enabled() && rt_policy(policy) && > + � � � � � � � � � � � � � � � task_group(p)->rt_bandwidth.rt_runtime == 0) > + � � � � � � � � � � � retval = -EPERM; > + > + � � � � � � � if (retval) { > + � � � � � � � � � � � __task_rq_unlock(rq); > + � � � � � � � � � � � raw_spin_unlock_irqrestore(&p->pi_lock, flags); > + � � � � � � � � � � � return retval; > + � � � � � � � } > + � � � } > +#endif > + > � � � �/* recheck policy now with rq lock held */ > � � � �if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { > � � � � � � � �policy = oldpolicy = -1; > > CC kernel/sched.o kernel/sched.c: In function �task_group�: kernel/sched.c:321: error: implicit declaration of function �task_rq� kernel/sched.c:321: error: invalid type argument of �->� (have �int�) make[1]: *** [kernel/sched.o] Error 1 I had to apply with fuzz. Did it mess up? static inline struct task_group *task_group(struct task_struct *p) { struct cgroup_subsys_state *css; css = task_subsys_state_check(p, cpu_cgroup_subsys_id, lockdep_is_held(&task_rq(p)->lock)); return container_of(css, struct task_group, css); } -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo(a)vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
From: Miles Lane on 9 Jun 2010 11:20 On Tue, Jun 8, 2010 at 9:34 AM, Peter Zijlstra <peterz(a)infradead.org> wrote: > On Tue, 2010-06-08 at 09:14 -0400, Miles Lane wrote: > >> � CC � � �kernel/sched.o >> kernel/sched.c: In function �task_group�: >> kernel/sched.c:321: error: implicit declaration of function �task_rq� >> kernel/sched.c:321: error: invalid type argument of �->� (have �int�) >> make[1]: *** [kernel/sched.o] Error 1 >> >> I had to apply with fuzz. �Did it mess up? > > > No, I probably did.. task_rq() is defined on line 636 or thereabouts, > and this function landed around line 320. > > Ahh, and it compiled here because I have CGROUP_SCHED=y, but > PROVE_RCU=n, so that whole check expression disappears and is never > evaluated... > > /me fixes > > --- > Subject: sched: PROVE_RCU vs cpu_cgroup > From: Peter Zijlstra <a.p.zijlstra(a)chello.nl> > Date: Tue Jun 08 11:40:42 CEST 2010 > > PROVE_RCU has a few issues with the cpu_cgroup because the scheduler > typically holds rq->lock around the css rcu derefs but the generic > cgroup code doesn't (and can't) know about that lock. > > Provide means to add extra checks to the css dereference and use that > in the scheduler to annotate its users. > > The addition of rq->lock to these checks is correct because the > cgroup_subsys::attach() method takes the rq->lock for each task it > moves, therefore by holding that lock, we ensure the task is pinned to > the current cgroup and the RCU dereference is valid. > > That leaves one genuine race in __sched_setscheduler() where we used > task_group() without holding any of the required locks and thus raced > with the cgroup code. Solve this by moving the check under the rq->lock. > > Signed-off-by: Peter Zijlstra <a.p.zijlstra(a)chello.nl> > --- > �include/linux/cgroup.h | � 20 +++++--- > �kernel/sched.c � � � � | �115 +++++++++++++++++++++++++------------------------ > �2 files changed, 73 insertions(+), 62 deletions(-) > > Index: linux-2.6/include/linux/cgroup.h > =================================================================== > --- linux-2.6.orig/include/linux/cgroup.h > +++ linux-2.6/include/linux/cgroup.h > @@ -525,13 +525,21 @@ static inline struct cgroup_subsys_state > � � � �return cgrp->subsys[subsys_id]; > �} > > -static inline struct cgroup_subsys_state *task_subsys_state( > - � � � struct task_struct *task, int subsys_id) > +/* > + * function to get the cgroup_subsys_state which allows for extra > + * rcu_dereference_check() conditions, such as locks used during the > + * cgroup_subsys::attach() methods. > + */ > +#define task_subsys_state_check(task, subsys_id, __c) � � � � � � � � �\ > + � � � rcu_dereference_check(task->cgroups->subsys[subsys_id], � � � � \ > + � � � � � � � � � � � � � � rcu_read_lock_held() || � � � � � � � � � \ > + � � � � � � � � � � � � � � lockdep_is_held(&task->alloc_lock) || � � \ > + � � � � � � � � � � � � � � cgroup_lock_is_held() || (__c)) > + > +static inline struct cgroup_subsys_state * > +task_subsys_state(struct task_struct *task, int subsys_id) > �{ > - � � � return rcu_dereference_check(task->cgroups->subsys[subsys_id], > - � � � � � � � � � � � � � � � � � �rcu_read_lock_held() || > - � � � � � � � � � � � � � � � � � �lockdep_is_held(&task->alloc_lock) || > - � � � � � � � � � � � � � � � � � �cgroup_lock_is_held()); > + � � � return task_subsys_state_check(task, subsys_id, false); > �} > > �static inline struct cgroup* task_cgroup(struct task_struct *task, > Index: linux-2.6/kernel/sched.c > =================================================================== > --- linux-2.6.orig/kernel/sched.c > +++ linux-2.6/kernel/sched.c > @@ -306,52 +306,6 @@ static int init_task_group_load = INIT_T > �*/ > �struct task_group init_task_group; > > -/* return group to which a task belongs */ > -static inline struct task_group *task_group(struct task_struct *p) > -{ > - � � � struct task_group *tg; > - > -#ifdef CONFIG_CGROUP_SCHED > - � � � tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id), > - � � � � � � � � � � � � � � � struct task_group, css); > -#else > - � � � tg = &init_task_group; > -#endif > - � � � return tg; > -} > - > -/* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */ > -static inline void set_task_rq(struct task_struct *p, unsigned int cpu) > -{ > - � � � /* > - � � � �* Strictly speaking this rcu_read_lock() is not needed since the > - � � � �* task_group is tied to the cgroup, which in turn can never go away > - � � � �* as long as there are tasks attached to it. > - � � � �* > - � � � �* However since task_group() uses task_subsys_state() which is an > - � � � �* rcu_dereference() user, this quiets CONFIG_PROVE_RCU. > - � � � �*/ > - � � � rcu_read_lock(); > -#ifdef CONFIG_FAIR_GROUP_SCHED > - � � � p->se.cfs_rq = task_group(p)->cfs_rq[cpu]; > - � � � p->se.parent = task_group(p)->se[cpu]; > -#endif > - > -#ifdef CONFIG_RT_GROUP_SCHED > - � � � p->rt.rt_rq �= task_group(p)->rt_rq[cpu]; > - � � � p->rt.parent = task_group(p)->rt_se[cpu]; > -#endif > - � � � rcu_read_unlock(); > -} > - > -#else > - > -static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { } > -static inline struct task_group *task_group(struct task_struct *p) > -{ > - � � � return NULL; > -} > - > �#endif /* CONFIG_CGROUP_SCHED */ > > �/* CFS-related fields in a runqueue */ > @@ -644,6 +598,49 @@ static inline int cpu_of(struct rq *rq) > �#define cpu_curr(cpu) � � � � �(cpu_rq(cpu)->curr) > �#define raw_rq() � � � � � � � (&__raw_get_cpu_var(runqueues)) > > +#ifdef CONFIG_CGROUP_SCHED > + > +/* > + * Return the group to which this tasks belongs. > + * > + * We use task_subsys_state_check() and extend the RCU verification > + * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach() > + * holds that lock for each task it moves into the cgroup. Therefore > + * by holding that lock, we pin the task to the current cgroup. > + */ > +static inline struct task_group *task_group(struct task_struct *p) > +{ > + � � � struct cgroup_subsys_state *css; > + > + � � � css = task_subsys_state_check(p, cpu_cgroup_subsys_id, > + � � � � � � � � � � � lockdep_is_held(&task_rq(p)->lock)); > + � � � return container_of(css, struct task_group, css); > +} > + > +/* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */ > +static inline void set_task_rq(struct task_struct *p, unsigned int cpu) > +{ > +#ifdef CONFIG_FAIR_GROUP_SCHED > + � � � p->se.cfs_rq = task_group(p)->cfs_rq[cpu]; > + � � � p->se.parent = task_group(p)->se[cpu]; > +#endif > + > +#ifdef CONFIG_RT_GROUP_SCHED > + � � � p->rt.rt_rq �= task_group(p)->rt_rq[cpu]; > + � � � p->rt.parent = task_group(p)->rt_se[cpu]; > +#endif > +} > + > +#else /* CONFIG_CGROUP_SCHED */ > + > +static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { } > +static inline struct task_group *task_group(struct task_struct *p) > +{ > + � � � return NULL; > +} > + > +#endif /* CONFIG_CGROUP_SCHED */ > + > �inline void update_rq_clock(struct rq *rq) > �{ > � � � �if (!rq->skip_clock_update) > @@ -4465,16 +4462,6 @@ recheck: > � � � �} > > � � � �if (user) { > -#ifdef CONFIG_RT_GROUP_SCHED > - � � � � � � � /* > - � � � � � � � �* Do not allow realtime tasks into groups that have no runtime > - � � � � � � � �* assigned. > - � � � � � � � �*/ > - � � � � � � � if (rt_bandwidth_enabled() && rt_policy(policy) && > - � � � � � � � � � � � � � � � task_group(p)->rt_bandwidth.rt_runtime == 0) > - � � � � � � � � � � � return -EPERM; > -#endif > - > � � � � � � � �retval = security_task_setscheduler(p, policy, param); > � � � � � � � �if (retval) > � � � � � � � � � � � �return retval; > @@ -4490,6 +4477,22 @@ recheck: > � � � � * runqueue lock must be held. > � � � � */ > � � � �rq = __task_rq_lock(p); > + > +#ifdef CONFIG_RT_GROUP_SCHED > + � � � if (user) { > + � � � � � � � /* > + � � � � � � � �* Do not allow realtime tasks into groups that have no runtime > + � � � � � � � �* assigned. > + � � � � � � � �*/ > + � � � � � � � if (rt_bandwidth_enabled() && rt_policy(policy) && > + � � � � � � � � � � � � � � � task_group(p)->rt_bandwidth.rt_runtime == 0) { > + � � � � � � � � � � � __task_rq_unlock(rq); > + � � � � � � � � � � � raw_spin_unlock_irqrestore(&p->pi_lock, flags); > + � � � � � � � � � � � return -EPERM; > + � � � � � � � } > + � � � } > +#endif > + > � � � �/* recheck policy now with rq lock held */ > � � � �if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { > � � � � � � � �policy = oldpolicy = -1; > > Sorry. I misunderstood this message when I first read it. I didn't realize this message include a new version of the patch. Anyhow, I just tried to apply the patch to 2.6.35-rc2-git3 and got this: # patch -p1 -l -F 20 --dry-run < ../5.patch patching file include/linux/cgroup.h patching file kernel/sched.c Hunk #1 succeeded at 306 with fuzz 1. Hunk #3 FAILED at 4462. Hunk #4 succeeded at 4487 with fuzz 3. 1 out of 4 hunks FAILED -- saving rejects to file kernel/sched.c.rej -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo(a)vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
From: Miles Lane on 9 Jun 2010 12:30
On Wed, Jun 9, 2010 at 11:24 AM, Peter Zijlstra <peterz(a)infradead.org> wrote: > On Wed, 2010-06-09 at 11:11 -0400, Miles Lane wrote: >> >> Sorry. �I misunderstood this message when I first read it. �I didn't >> realize this message include a new version of the patch. >> Anyhow, I just tried to apply the patch to 2.6.35-rc2-git3 and got this: >> >> # patch -p1 -l -F 20 --dry-run < ../5.patch >> patching file include/linux/cgroup.h >> patching file kernel/sched.c >> Hunk #1 succeeded at 306 with fuzz 1. >> Hunk #3 FAILED at 4462. >> Hunk #4 succeeded at 4487 with fuzz 3. >> 1 out of 4 hunks FAILED -- saving rejects to file kernel/sched.c.rej > > Weird.. it seems to apply without trouble to Linus' git tree. > > root(a)twins:/usr/src/linux-2.6# git checkout -f origin/master > HEAD is now at 84f7586... Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/ieee1394/linux1394-2.6 > root(a)twins:/usr/src/linux-2.6# quilt push > Applying patch patches/sched-rcu-validation.patch > patching file include/linux/cgroup.h > patching file kernel/sched.c > > Now at patch patches/sched-rcu-validation.patch > root(a)twins:/usr/src/linux-2.6# git describe > v2.6.35-rc2-54-g84f7586 Oh. Duh. I know what is going on. I had received another patch to sched.c. They must conflict. I will test with just your patch now. -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo(a)vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/ |