peterz-raw_pagefault_disable.patch

Subject: mm: raw_pagefault_disable
From: Peter Zijlstra <a.p.zijlstra@chello.nl>
Date: Fri Aug 05 17:16:58 CEST 2011

Adding migrate_disable() to pagefault_disable() to preserve the
per-cpu thing for kmap_atomic might not have been the best of choices.
But short of adding preempt_disable/migrate_disable foo all over the
kmap code it still seems the best way.

It does however yield the below borkage as well as wreck !-rt builds
since !-rt does rely on pagefault_disable() not preempting. So fix all
that up by adding raw_pagefault_disable().

 <NMI>  [<ffffffff81076d5c>] warn_slowpath_common+0x85/0x9d
 [<ffffffff81076e17>] warn_slowpath_fmt+0x46/0x48
 [<ffffffff814f7fca>] ? _raw_spin_lock+0x6c/0x73
 [<ffffffff810cac87>] ? watchdog_overflow_callback+0x9b/0xd0
 [<ffffffff810caca3>] watchdog_overflow_callback+0xb7/0xd0
 [<ffffffff810f51bb>] __perf_event_overflow+0x11c/0x1fe
 [<ffffffff810f298f>] ? perf_event_update_userpage+0x149/0x151
 [<ffffffff810f2846>] ? perf_event_task_disable+0x7c/0x7c
 [<ffffffff810f5b7c>] perf_event_overflow+0x14/0x16
 [<ffffffff81046e02>] x86_pmu_handle_irq+0xcb/0x108
 [<ffffffff814f9a6b>] perf_event_nmi_handler+0x46/0x91
 [<ffffffff814fb2ba>] notifier_call_chain+0x79/0xa6
 [<ffffffff814fb34d>] __atomic_notifier_call_chain+0x66/0x98
 [<ffffffff814fb2e7>] ? notifier_call_chain+0xa6/0xa6
 [<ffffffff814fb393>] atomic_notifier_call_chain+0x14/0x16
 [<ffffffff814fb3c3>] notify_die+0x2e/0x30
 [<ffffffff814f8f75>] do_nmi+0x7e/0x22b
 [<ffffffff814f8bca>] nmi+0x1a/0x2c
 [<ffffffff814fb130>] ? sub_preempt_count+0x4b/0xaa
 <<EOE>>  <IRQ>  [<ffffffff812d44cc>] delay_tsc+0xac/0xd1
 [<ffffffff812d4399>] __delay+0xf/0x11
 [<ffffffff812d95d9>] do_raw_spin_lock+0xd2/0x13c
 [<ffffffff814f813e>] _raw_spin_lock_irqsave+0x6b/0x85
 [<ffffffff8106772a>] ? task_rq_lock+0x35/0x8d
 [<ffffffff8106772a>] task_rq_lock+0x35/0x8d
 [<ffffffff8106fe2f>] migrate_disable+0x65/0x12c
 [<ffffffff81114e69>] pagefault_disable+0xe/0x1f
 [<ffffffff81039c73>] dump_trace+0x21f/0x2e2
 [<ffffffff8103ad79>] show_trace_log_lvl+0x54/0x5d
 [<ffffffff8103ad97>] show_trace+0x15/0x17
 [<ffffffff814f4f5f>] dump_stack+0x77/0x80
 [<ffffffff812d94b0>] spin_bug+0x9c/0xa3
 [<ffffffff81067745>] ? task_rq_lock+0x50/0x8d
 [<ffffffff812d954e>] do_raw_spin_lock+0x47/0x13c
 [<ffffffff814f7fbe>] _raw_spin_lock+0x60/0x73
 [<ffffffff81067745>] ? task_rq_lock+0x50/0x8d
 [<ffffffff81067745>] task_rq_lock+0x50/0x8d
 [<ffffffff8106fe2f>] migrate_disable+0x65/0x12c
 [<ffffffff81114e69>] pagefault_disable+0xe/0x1f
 [<ffffffff81039c73>] dump_trace+0x21f/0x2e2
 [<ffffffff8104369b>] save_stack_trace+0x2f/0x4c
 [<ffffffff810a7848>] save_trace+0x3f/0xaf
 [<ffffffff810aa2bd>] mark_lock+0x228/0x530
 [<ffffffff810aac27>] __lock_acquire+0x662/0x1812
 [<ffffffff8103dad4>] ? native_sched_clock+0x37/0x6d
 [<ffffffff810a790e>] ? trace_hardirqs_off_caller+0x1f/0x99
 [<ffffffff810693f6>] ? sched_rt_period_timer+0xbd/0x218
 [<ffffffff810ac403>] lock_acquire+0x145/0x18a
 [<ffffffff810693f6>] ? sched_rt_period_timer+0xbd/0x218
 [<ffffffff814f7f9e>] _raw_spin_lock+0x40/0x73
 [<ffffffff810693f6>] ? sched_rt_period_timer+0xbd/0x218
 [<ffffffff810693f6>] sched_rt_period_timer+0xbd/0x218
 [<ffffffff8109aa39>] __run_hrtimer+0x1e4/0x347
 [<ffffffff81069339>] ? can_migrate_task.clone.82+0x14a/0x14a
 [<ffffffff8109b97c>] hrtimer_interrupt+0xee/0x1d6
 [<ffffffff814fb23d>] ? add_preempt_count+0xae/0xb2
 [<ffffffff814ffb38>] smp_apic_timer_interrupt+0x85/0x98
 [<ffffffff814fef13>] apic_timer_interrupt+0x13/0x20


Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/n/tip-31keae8mkjiv8esq4rl76cib@git.kernel.org
---
 include/linux/uaccess.h |   30 ++++++++++++++++++++++++++++--
 mm/memory.c             |    2 ++
 2 files changed, 30 insertions(+), 2 deletions(-)

Index: linux-stable/include/linux/uaccess.h
===================================================================
--- linux-stable.orig/include/linux/uaccess.h
+++ linux-stable/include/linux/uaccess.h
@@ -8,8 +8,34 @@
  * These routines enable/disable the pagefault handler in that
  * it will not take any MM locks and go straight to the fixup table.
  */
+static inline void raw_pagefault_disable(void)
+{
+	inc_preempt_count();
+	barrier();
+}
+
+static inline void raw_pagefault_enable(void)
+{
+	barrier();
+	dec_preempt_count();
+	barrier();
+	preempt_check_resched();
+}
+
+#ifndef CONFIG_PREEMPT_RT_FULL
+static inline void pagefault_disable(void)
+{
+	raw_pagefault_disable();
+}
+
+static inline void pagefault_enable(void)
+{
+	raw_pagefault_enable();
+}
+#else
 extern void pagefault_disable(void);
 extern void pagefault_enable(void);
+#endif
 
 #ifndef ARCH_HAS_NOCACHE_UACCESS
 
@@ -50,9 +76,9 @@ static inline unsigned long __copy_from_
 		mm_segment_t old_fs = get_fs();		\
 							\
 		set_fs(KERNEL_DS);			\
-		pagefault_disable();			\
+		raw_pagefault_disable();		\
 		ret = __copy_from_user_inatomic(&(retval), (__force typeof(retval) __user *)(addr), sizeof(retval));		\
-		pagefault_enable();			\
+		raw_pagefault_enable();			\
 		set_fs(old_fs);				\
 		ret;					\
 	})
Index: linux-stable/mm/memory.c
===================================================================
--- linux-stable.orig/mm/memory.c
+++ linux-stable/mm/memory.c
@@ -3484,6 +3484,7 @@ unlock:
 	return 0;
 }
 
+#ifdef CONFIG_PREEMPT_RT_FULL
 void pagefault_disable(void)
 {
 	inc_preempt_count();
@@ -3512,6 +3513,7 @@ void pagefault_enable(void)
 	preempt_check_resched();
 }
 EXPORT_SYMBOL_GPL(pagefault_enable);
+#endif
 
 /*
  * By the time we get here, we already hold the mm semaphore