Merge branch 'tracing/hw-branch-tracing' into tracing/core

Merge reason: this topic is ready for upstream now. It passed
              Oleg's review and Andrew had no further mm/*
              objections/observations either.

Signed-off-by: Ingo Molnar <mingo@elte.hu>

25 files changed, 1616 insertions, 565 deletions

diff --git a/arch/x86/Kconfig.cpu b/arch/x86/Kconfig.cpu
index 8130334..924e156 100644
--- a/arch/x86/Kconfig.cpu
+++ b/arch/x86/Kconfig.cpu
@@ -506,7 +506,6 @@ config X86_PTRACE_BTS
 	bool "Branch Trace Store"
 	default y
 	depends on X86_DEBUGCTLMSR
-	depends on BROKEN
 	---help---
 	  This adds a ptrace interface to the hardware's branch trace store.
 
diff --git a/arch/x86/Kconfig.debug b/arch/x86/Kconfig.debug
index d8359e7..22b752e0 100644
--- a/arch/x86/Kconfig.debug
+++ b/arch/x86/Kconfig.debug
@@ -167,6 +167,15 @@ config IOMMU_LEAK
 	  Add a simple leak tracer to the IOMMU code. This is useful when you
 	  are debugging a buggy device driver that leaks IOMMU mappings.
 
+config X86_DS_SELFTEST
+    bool "DS selftest"
+    default y
+    depends on DEBUG_KERNEL
+    depends on X86_DS
+	---help---
+	  Perform Debug Store selftests at boot time.
+	  If in doubt, say "N".
+
 config HAVE_MMIOTRACE_SUPPORT
 	def_bool y
 
diff --git a/arch/x86/include/asm/ds.h b/arch/x86/include/asm/ds.h
index a8f672b..70dac19 100644
--- a/arch/x86/include/asm/ds.h
+++ b/arch/x86/include/asm/ds.h
@@ -15,8 +15,8 @@
  * - buffer allocation (memory accounting)
  *
  *
- * Copyright (C) 2007-2008 Intel Corporation.
- * Markus Metzger <markus.t.metzger@intel.com>, 2007-2008
+ * Copyright (C) 2007-2009 Intel Corporation.
+ * Markus Metzger <markus.t.metzger@intel.com>, 2007-2009
  */
 
 #ifndef _ASM_X86_DS_H
@@ -83,8 +83,10 @@ enum ds_feature {
  * The interrupt threshold is independent from the overflow callback
  * to allow users to use their own overflow interrupt handling mechanism.
  *
- * task: the task to request recording for;
- *       NULL for per-cpu recording on the current cpu
+ * The function might sleep.
+ *
+ * task: the task to request recording for
+ * cpu:  the cpu to request recording for
  * base: the base pointer for the (non-pageable) buffer;
  * size: the size of the provided buffer in bytes
  * ovfl: pointer to a function to be called on buffer overflow;
@@ -93,19 +95,28 @@ enum ds_feature {
  *     -1 if no interrupt threshold is requested.
  * flags: a bit-mask of the above flags
  */
-extern struct bts_tracer *ds_request_bts(struct task_struct *task,
-					 void *base, size_t size,
-					 bts_ovfl_callback_t ovfl,
-					 size_t th, unsigned int flags);
-extern struct pebs_tracer *ds_request_pebs(struct task_struct *task,
-					   void *base, size_t size,
-					   pebs_ovfl_callback_t ovfl,
-					   size_t th, unsigned int flags);
+extern struct bts_tracer *ds_request_bts_task(struct task_struct *task,
+					      void *base, size_t size,
+					      bts_ovfl_callback_t ovfl,
+					      size_t th, unsigned int flags);
+extern struct bts_tracer *ds_request_bts_cpu(int cpu, void *base, size_t size,
+					     bts_ovfl_callback_t ovfl,
+					     size_t th, unsigned int flags);
+extern struct pebs_tracer *ds_request_pebs_task(struct task_struct *task,
+						void *base, size_t size,
+						pebs_ovfl_callback_t ovfl,
+						size_t th, unsigned int flags);
+extern struct pebs_tracer *ds_request_pebs_cpu(int cpu,
+					       void *base, size_t size,
+					       pebs_ovfl_callback_t ovfl,
+					       size_t th, unsigned int flags);
 
 /*
  * Release BTS or PEBS resources
  * Suspend and resume BTS or PEBS tracing
  *
+ * Must be called with irq's enabled.
+ *
  * tracer: the tracer handle returned from ds_request_~()
  */
 extern void ds_release_bts(struct bts_tracer *tracer);
@@ -115,6 +126,28 @@ extern void ds_release_pebs(struct pebs_tracer *tracer);
 extern void ds_suspend_pebs(struct pebs_tracer *tracer);
 extern void ds_resume_pebs(struct pebs_tracer *tracer);
 
+/*
+ * Release BTS or PEBS resources
+ * Suspend and resume BTS or PEBS tracing
+ *
+ * Cpu tracers must call this on the traced cpu.
+ * Task tracers must call ds_release_~_noirq() for themselves.
+ *
+ * May be called with irq's disabled.
+ *
+ * Returns 0 if successful;
+ * -EPERM if the cpu tracer does not trace the current cpu.
+ * -EPERM if the task tracer does not trace itself.
+ *
+ * tracer: the tracer handle returned from ds_request_~()
+ */
+extern int ds_release_bts_noirq(struct bts_tracer *tracer);
+extern int ds_suspend_bts_noirq(struct bts_tracer *tracer);
+extern int ds_resume_bts_noirq(struct bts_tracer *tracer);
+extern int ds_release_pebs_noirq(struct pebs_tracer *tracer);
+extern int ds_suspend_pebs_noirq(struct pebs_tracer *tracer);
+extern int ds_resume_pebs_noirq(struct pebs_tracer *tracer);
+
 
 /*
  * The raw DS buffer state as it is used for BTS and PEBS recording.
@@ -170,9 +203,9 @@ struct bts_struct {
 		} lbr;
 		/* BTS_TASK_ARRIVES or BTS_TASK_DEPARTS */
 		struct {
-			__u64 jiffies;
+			__u64 clock;
 			pid_t pid;
-		} timestamp;
+		} event;
 	} variant;
 };
 
@@ -201,8 +234,12 @@ struct bts_trace {
 struct pebs_trace {
 	struct ds_trace ds;
 
-	/* the PEBS reset value */
-	unsigned long long reset_value;
+	/* the number of valid counters in the below array */
+	unsigned int counters;
+
+#define MAX_PEBS_COUNTERS 4
+	/* the counter reset value */
+	unsigned long long counter_reset[MAX_PEBS_COUNTERS];
 };
 
 
@@ -237,9 +274,11 @@ extern int ds_reset_pebs(struct pebs_tracer *tracer);
  * Returns 0 on success; -Eerrno on error
  *
  * tracer: the tracer handle returned from ds_request_pebs()
+ * counter: the index of the counter
  * value: the new counter reset value
  */
-extern int ds_set_pebs_reset(struct pebs_tracer *tracer, u64 value);
+extern int ds_set_pebs_reset(struct pebs_tracer *tracer,
+			     unsigned int counter, u64 value);
 
 /*
  * Initialization
@@ -252,21 +291,12 @@ extern void __cpuinit ds_init_intel(struct cpuinfo_x86 *);
  */
 extern void ds_switch_to(struct task_struct *prev, struct task_struct *next);
 
-/*
- * Task clone/init and cleanup work
- */
-extern void ds_copy_thread(struct task_struct *tsk, struct task_struct *father);
-extern void ds_exit_thread(struct task_struct *tsk);
-
 #else /* CONFIG_X86_DS */
 
 struct cpuinfo_x86;
 static inline void __cpuinit ds_init_intel(struct cpuinfo_x86 *ignored) {}
 static inline void ds_switch_to(struct task_struct *prev,
 				struct task_struct *next) {}
-static inline void ds_copy_thread(struct task_struct *tsk,
-				  struct task_struct *father) {}
-static inline void ds_exit_thread(struct task_struct *tsk) {}
 
 #endif /* CONFIG_X86_DS */
 #endif /* _ASM_X86_DS_H */
diff --git a/arch/x86/include/asm/processor.h b/arch/x86/include/asm/processor.h
index c2cceae..0b2fab0 100644
--- a/arch/x86/include/asm/processor.h
+++ b/arch/x86/include/asm/processor.h
@@ -460,14 +460,8 @@ struct thread_struct {
 	unsigned		io_bitmap_max;
 /* MSR_IA32_DEBUGCTLMSR value to switch in if TIF_DEBUGCTLMSR is set.  */
 	unsigned long	debugctlmsr;
-#ifdef CONFIG_X86_DS
-/* Debug Store context; see include/asm-x86/ds.h; goes into MSR_IA32_DS_AREA */
+	/* Debug Store context; see asm/ds.h */
 	struct ds_context	*ds_ctx;
-#endif /* CONFIG_X86_DS */
-#ifdef CONFIG_X86_PTRACE_BTS
-/* the signal to send on a bts buffer overflow */
-	unsigned int	bts_ovfl_signal;
-#endif /* CONFIG_X86_PTRACE_BTS */
 };
 
 static inline unsigned long native_get_debugreg(int regno)
@@ -795,6 +789,21 @@ static inline unsigned long get_debugctlmsr(void)
     return debugctlmsr;
 }
 
+static inline unsigned long get_debugctlmsr_on_cpu(int cpu)
+{
+	u64 debugctlmsr = 0;
+	u32 val1, val2;
+
+#ifndef CONFIG_X86_DEBUGCTLMSR
+	if (boot_cpu_data.x86 < 6)
+		return 0;
+#endif
+	rdmsr_on_cpu(cpu, MSR_IA32_DEBUGCTLMSR, &val1, &val2);
+	debugctlmsr = val1 | ((u64)val2 << 32);
+
+	return debugctlmsr;
+}
+
 static inline void update_debugctlmsr(unsigned long debugctlmsr)
 {
 #ifndef CONFIG_X86_DEBUGCTLMSR
@@ -804,6 +813,18 @@ static inline void update_debugctlmsr(unsigned long debugctlmsr)
 	wrmsrl(MSR_IA32_DEBUGCTLMSR, debugctlmsr);
 }
 
+static inline void update_debugctlmsr_on_cpu(int cpu,
+					     unsigned long debugctlmsr)
+{
+#ifndef CONFIG_X86_DEBUGCTLMSR
+	if (boot_cpu_data.x86 < 6)
+		return;
+#endif
+	wrmsr_on_cpu(cpu, MSR_IA32_DEBUGCTLMSR,
+		     (u32)((u64)debugctlmsr),
+		     (u32)((u64)debugctlmsr >> 32));
+}
+
 /*
  * from system description table in BIOS. Mostly for MCA use, but
  * others may find it useful:
diff --git a/arch/x86/include/asm/ptrace.h b/arch/x86/include/asm/ptrace.h
index e304b66..5cdd19f 100644
--- a/arch/x86/include/asm/ptrace.h
+++ b/arch/x86/include/asm/ptrace.h
@@ -235,12 +235,11 @@ extern int do_get_thread_area(struct task_struct *p, int idx,
 extern int do_set_thread_area(struct task_struct *p, int idx,
 			      struct user_desc __user *info, int can_allocate);
 
-extern void x86_ptrace_untrace(struct task_struct *);
-extern void x86_ptrace_fork(struct task_struct *child,
-			    unsigned long clone_flags);
+#ifdef CONFIG_X86_PTRACE_BTS
+extern void ptrace_bts_untrace(struct task_struct *tsk);
 
-#define arch_ptrace_untrace(tsk) x86_ptrace_untrace(tsk)
-#define arch_ptrace_fork(child, flags) x86_ptrace_fork(child, flags)
+#define arch_ptrace_untrace(tsk)	ptrace_bts_untrace(tsk)
+#endif /* CONFIG_X86_PTRACE_BTS */
 
 #endif /* __KERNEL__ */
 
diff --git a/arch/x86/kernel/Makefile b/arch/x86/kernel/Makefile
index 145cce7..77df4d6 100644
--- a/arch/x86/kernel/Makefile
+++ b/arch/x86/kernel/Makefile
@@ -44,6 +44,7 @@ obj-y				+= process.o
 obj-y				+= i387.o xsave.o
 obj-y				+= ptrace.o
 obj-$(CONFIG_X86_DS)		+= ds.o
+obj-$(CONFIG_X86_DS_SELFTEST)		+= ds_selftest.o
 obj-$(CONFIG_X86_32)		+= tls.o
 obj-$(CONFIG_IA32_EMULATION)	+= tls.o
 obj-y				+= step.o
diff --git a/arch/x86/kernel/ds.c b/arch/x86/kernel/ds.c
index 87b67e3..48bfe13 100644
--- a/arch/x86/kernel/ds.c
+++ b/arch/x86/kernel/ds.c
@@ -19,45 +19,61 @@
  * Markus Metzger <markus.t.metzger@intel.com>, 2007-2009
  */
 
-
-#include <asm/ds.h>
-
-#include <linux/errno.h>
+#include <linux/kernel.h>
 #include <linux/string.h>
-#include <linux/slab.h>
+#include <linux/errno.h>
 #include <linux/sched.h>
+#include <linux/slab.h>
 #include <linux/mm.h>
-#include <linux/kernel.h>
+#include <linux/trace_clock.h>
+
+#include <asm/ds.h>
 
+#include "ds_selftest.h"
 
 /*
- * The configuration for a particular DS hardware implementation.
+ * The configuration for a particular DS hardware implementation:
  */
 struct ds_configuration {
-	/* the name of the configuration */
-	const char *name;
-	/* the size of one pointer-typed field in the DS structure and
-	   in the BTS and PEBS buffers in bytes;
-	   this covers the first 8 DS fields related to buffer management. */
-	unsigned char  sizeof_field;
-	/* the size of a BTS/PEBS record in bytes */
-	unsigned char  sizeof_rec[2];
-	/* a series of bit-masks to control various features indexed
-	 * by enum ds_feature */
-	unsigned long ctl[dsf_ctl_max];
+	/* The name of the configuration: */
+	const char		*name;
+
+	/* The size of pointer-typed fields in DS, BTS, and PEBS: */
+	unsigned char		sizeof_ptr_field;
+
+	/* The size of a BTS/PEBS record in bytes: */
+	unsigned char		sizeof_rec[2];
+
+	/* The number of pebs counter reset values in the DS structure. */
+	unsigned char		nr_counter_reset;
+
+	/* Control bit-masks indexed by enum ds_feature: */
+	unsigned long		ctl[dsf_ctl_max];
 };
-static DEFINE_PER_CPU(struct ds_configuration, ds_cfg_array);
+static struct ds_configuration ds_cfg __read_mostly;
+
+
+/* Maximal size of a DS configuration: */
+#define MAX_SIZEOF_DS		0x80
 
-#define ds_cfg per_cpu(ds_cfg_array, smp_processor_id())
+/* Maximal size of a BTS record: */
+#define MAX_SIZEOF_BTS		(3 * 8)
 
-#define MAX_SIZEOF_DS (12 * 8)	/* maximal size of a DS configuration */
-#define MAX_SIZEOF_BTS (3 * 8)	/* maximal size of a BTS record */
-#define DS_ALIGNMENT (1 << 3)	/* BTS and PEBS buffer alignment */
+/* BTS and PEBS buffer alignment: */
+#define DS_ALIGNMENT		(1 << 3)
 
-#define BTS_CONTROL \
- (ds_cfg.ctl[dsf_bts] | ds_cfg.ctl[dsf_bts_kernel] | ds_cfg.ctl[dsf_bts_user] |\
-  ds_cfg.ctl[dsf_bts_overflow])
+/* Number of buffer pointers in DS: */
+#define NUM_DS_PTR_FIELDS	8
 
+/* Size of a pebs reset value in DS: */
+#define PEBS_RESET_FIELD_SIZE	8
+
+/* Mask of control bits in the DS MSR register: */
+#define BTS_CONTROL				  \
+	( ds_cfg.ctl[dsf_bts]			| \
+	  ds_cfg.ctl[dsf_bts_kernel]		| \
+	  ds_cfg.ctl[dsf_bts_user]		| \
+	  ds_cfg.ctl[dsf_bts_overflow] )
 
 /*
  * A BTS or PEBS tracer.
@@ -66,29 +82,36 @@ static DEFINE_PER_CPU(struct ds_configuration, ds_cfg_array);
  * to identify tracers.
  */
 struct ds_tracer {
-	/* the DS context (partially) owned by this tracer */
-	struct ds_context *context;
-	/* the buffer provided on ds_request() and its size in bytes */
-	void *buffer;
-	size_t size;
+	/* The DS context (partially) owned by this tracer. */
+	struct ds_context	*context;
+	/* The buffer provided on ds_request() and its size in bytes. */
+	void			*buffer;
+	size_t			size;
 };
 
 struct bts_tracer {
-	/* the common DS part */
-	struct ds_tracer ds;
-	/* the trace including the DS configuration */
-	struct bts_trace trace;
-	/* buffer overflow notification function */
-	bts_ovfl_callback_t ovfl;
+	/* The common DS part: */
+	struct ds_tracer	ds;
+
+	/* The trace including the DS configuration: */
+	struct bts_trace	trace;
+
+	/* Buffer overflow notification function: */
+	bts_ovfl_callback_t	ovfl;
+
+	/* Active flags affecting trace collection. */
+	unsigned int		flags;
 };
 
 struct pebs_tracer {
-	/* the common DS part */
-	struct ds_tracer ds;
-	/* the trace including the DS configuration */
-	struct pebs_trace trace;
-	/* buffer overflow notification function */
-	pebs_ovfl_callback_t ovfl;
+	/* The common DS part: */
+	struct ds_tracer	ds;
+
+	/* The trace including the DS configuration: */
+	struct pebs_trace	trace;
+
+	/* Buffer overflow notification function: */
+	pebs_ovfl_callback_t	ovfl;
 };
 
 /*
@@ -97,6 +120,7 @@ struct pebs_tracer {
  *
  * The DS configuration consists of the following fields; different
  * architetures vary in the size of those fields.
+ *
  * - double-word aligned base linear address of the BTS buffer
  * - write pointer into the BTS buffer
  * - end linear address of the BTS buffer (one byte beyond the end of
@@ -135,21 +159,22 @@ enum ds_field {
 };
 
 enum ds_qualifier {
-	ds_bts  = 0,
+	ds_bts = 0,
 	ds_pebs
 };
 
-static inline unsigned long ds_get(const unsigned char *base,
-				   enum ds_qualifier qual, enum ds_field field)
+static inline unsigned long
+ds_get(const unsigned char *base, enum ds_qualifier qual, enum ds_field field)
 {
-	base += (ds_cfg.sizeof_field * (field + (4 * qual)));
+	base += (ds_cfg.sizeof_ptr_field * (field + (4 * qual)));
 	return *(unsigned long *)base;
 }
 
-static inline void ds_set(unsigned char *base, enum ds_qualifier qual,
-			  enum ds_field field, unsigned long value)
+static inline void
+ds_set(unsigned char *base, enum ds_qualifier qual, enum ds_field field,
+       unsigned long value)
 {
-	base += (ds_cfg.sizeof_field * (field + (4 * qual)));
+	base += (ds_cfg.sizeof_ptr_field * (field + (4 * qual)));
 	(*(unsigned long *)base) = value;
 }
 
@@ -159,7 +184,6 @@ static inline void ds_set(unsigned char *base, enum ds_qualifier qual,
  */
 static DEFINE_SPINLOCK(ds_lock);
 
-
 /*
  * We either support (system-wide) per-cpu or per-thread allocation.
  * We distinguish the two based on the task_struct pointer, where a
@@ -178,12 +202,28 @@ static DEFINE_SPINLOCK(ds_lock);
  */
 static atomic_t tracers = ATOMIC_INIT(0);
 
-static inline void get_tracer(struct task_struct *task)
+static inline int get_tracer(struct task_struct *task)
 {
-	if (task)
+	int error;
+
+	spin_lock_irq(&ds_lock);
+
+	if (task) {
+		error = -EPERM;
+		if (atomic_read(&tracers) < 0)
+			goto out;
 		atomic_inc(&tracers);
-	else
+	} else {
+		error = -EPERM;
+		if (atomic_read(&tracers) > 0)
+			goto out;
 		atomic_dec(&tracers);
+	}
+
+	error = 0;
+out:
+	spin_unlock_irq(&ds_lock);
+	return error;
 }
 
 static inline void put_tracer(struct task_struct *task)
@@ -194,14 +234,6 @@ static inline void put_tracer(struct task_struct *task)
 		atomic_inc(&tracers);
 }
 
-static inline int check_tracer(struct task_struct *task)
-{
-	return task ?
-		(atomic_read(&tracers) >= 0) :
-		(atomic_read(&tracers) <= 0);
-}
-
-
 /*
  * The DS context is either attached to a thread or to a cpu:
  * - in the former case, the thread_struct contains a pointer to the
@@ -213,61 +245,58 @@ static inline int check_tracer(struct task_struct *task)
  * deallocated when the last user puts the context.
  */
 struct ds_context {
-	/* pointer to the DS configuration; goes into MSR_IA32_DS_AREA */
-	unsigned char ds[MAX_SIZEOF_DS];
-	/* the owner of the BTS and PEBS configuration, respectively */
-	struct bts_tracer *bts_master;
-	struct pebs_tracer *pebs_master;
-	/* use count */
-	unsigned long count;
-	/* a pointer to the context location inside the thread_struct
-	 * or the per_cpu context array */
-	struct ds_context **this;
-	/* a pointer to the task owning this context, or NULL, if the
-	 * context is owned by a cpu */
-	struct task_struct *task;
-};
+	/* The DS configuration; goes into MSR_IA32_DS_AREA: */
+	unsigned char		ds[MAX_SIZEOF_DS];
+
+	/* The owner of the BTS and PEBS configuration, respectively: */
+	struct bts_tracer	*bts_master;
+	struct pebs_tracer	*pebs_master;
 
-static DEFINE_PER_CPU(struct ds_context *, system_context_array);
+	/* Use count: */
+	unsigned long		count;
 
-#define system_context per_cpu(system_context_array, smp_processor_id())
+	/* Pointer to the context pointer field: */
+	struct ds_context	**this;
+
+	/* The traced task; NULL for cpu tracing: */
+	struct task_struct	*task;
+
+	/* The traced cpu; only valid if task is NULL: */
+	int			cpu;
+};
 
+static DEFINE_PER_CPU(struct ds_context *, cpu_context);
 
-static inline struct ds_context *ds_get_context(struct task_struct *task)
+
+static struct ds_context *ds_get_context(struct task_struct *task, int cpu)
 {
 	struct ds_context **p_context =
-		(task ? &task->thread.ds_ctx : &system_context);
+		(task ? &task->thread.ds_ctx : &per_cpu(cpu_context, cpu));
 	struct ds_context *context = NULL;
 	struct ds_context *new_context = NULL;
-	unsigned long irq;
 
 	/* Chances are small that we already have a context. */
 	new_context = kzalloc(sizeof(*new_context), GFP_KERNEL);
 	if (!new_context)
 		return NULL;
 
-	spin_lock_irqsave(&ds_lock, irq);
+	spin_lock_irq(&ds_lock);
 
 	context = *p_context;
-	if (!context) {
+	if (likely(!context)) {
 		context = new_context;
 
 		context->this = p_context;
 		context->task = task;
+		context->cpu = cpu;
 		context->count = 0;
 
-		if (task)
-			set_tsk_thread_flag(task, TIF_DS_AREA_MSR);
-
-		if (!task || (task == current))
-			wrmsrl(MSR_IA32_DS_AREA, (unsigned long)context->ds);
-
 		*p_context = context;
 	}
 
 	context->count++;
 
-	spin_unlock_irqrestore(&ds_lock, irq);
+	spin_unlock_irq(&ds_lock);
 
 	if (context != new_context)
 		kfree(new_context);
@@ -275,8 +304,9 @@ static inline struct ds_context *ds_get_context(struct task_struct *task)
 	return context;
 }
 
-static inline void ds_put_context(struct ds_context *context)
+static void ds_put_context(struct ds_context *context)
 {
+	struct task_struct *task;
 	unsigned long irq;
 
 	if (!context)
@@ -291,17 +321,55 @@ static inline void ds_put_context(struct ds_context *context)
 
 	*(context->this) = NULL;
 
-	if (context->task)
-		clear_tsk_thread_flag(context->task, TIF_DS_AREA_MSR);
+	task = context->task;
+
+	if (task)
+		clear_tsk_thread_flag(task, TIF_DS_AREA_MSR);
 
-	if (!context->task || (context->task == current))
-		wrmsrl(MSR_IA32_DS_AREA, 0);
+	/*
+	 * We leave the (now dangling) pointer to the DS configuration in
+	 * the DS_AREA msr. This is as good or as bad as replacing it with
+	 * NULL - the hardware would crash if we enabled tracing.
+	 *
+	 * This saves us some problems with having to write an msr on a
+	 * different cpu while preventing others from doing the same for the
+	 * next context for that same cpu.
+	 */
 
 	spin_unlock_irqrestore(&ds_lock, irq);
 
+	/* The context might still be in use for context switching. */
+	if (task && (task != current))
+		wait_task_context_switch(task);
+
 	kfree(context);
 }
 
+static void ds_install_ds_area(struct ds_context *context)
+{
+	unsigned long ds;
+
+	ds = (unsigned long)context->ds;
+
+	/*
+	 * There is a race between the bts master and the pebs master.
+	 *
+	 * The thread/cpu access is synchronized via get/put_cpu() for
+	 * task tracing and via wrmsr_on_cpu for cpu tracing.
+	 *
+	 * If bts and pebs are collected for the same task or same cpu,
+	 * the same confiuration is written twice.
+	 */
+	if (context->task) {
+		get_cpu();
+		if (context->task == current)
+			wrmsrl(MSR_IA32_DS_AREA, ds);
+		set_tsk_thread_flag(context->task, TIF_DS_AREA_MSR);
+		put_cpu();
+	} else
+		wrmsr_on_cpu(context->cpu, MSR_IA32_DS_AREA,
+			     (u32)((u64)ds), (u32)((u64)ds >> 32));
+}
 
 /*
  * Call the tracer's callback on a buffer overflow.
@@ -332,9 +400,9 @@ static void ds_overflow(struct ds_context *context, enum ds_qualifier qual)
  * The remainder of any partially written record is zeroed out.
  *
  * context: the DS context
- * qual: the buffer type
- * record: the data to write
- * size: the size of the data
+ * qual:    the buffer type
+ * record:  the data to write
+ * size:    the size of the data
  */
 static int ds_write(struct ds_context *context, enum ds_qualifier qual,
 		    const void *record, size_t size)
@@ -349,14 +417,14 @@ static int ds_write(struct ds_context *context, enum ds_qualifier qual,
 		unsigned long write_size, adj_write_size;
 
 		/*
-		 * write as much as possible without producing an
+		 * Write as much as possible without producing an
 		 * overflow interrupt.
 		 *
-		 * interrupt_threshold must either be
+		 * Interrupt_threshold must either be
 		 * - bigger than absolute_maximum or
 		 * - point to a record between buffer_base and absolute_maximum
 		 *
-		 * index points to a valid record.
+		 * Index points to a valid record.
 		 */
 		base   = ds_get(context->ds, qual, ds_buffer_base);
 		index  = ds_get(context->ds, qual, ds_index);
@@ -365,8 +433,10 @@ static int ds_write(struct ds_context *context, enum ds_qualifier qual,
 
 		write_end = min(end, int_th);
 
-		/* if we are already beyond the interrupt threshold,
-		 * we fill the entire buffer */
+		/*
+		 * If we are already beyond the interrupt threshold,
+		 * we fill the entire buffer.
+		 */
 		if (write_end <= index)
 			write_end = end;
 
@@ -383,7 +453,7 @@ static int ds_write(struct ds_context *context, enum ds_qualifier qual,
 		adj_write_size = write_size / ds_cfg.sizeof_rec[qual];
 		adj_write_size *= ds_cfg.sizeof_rec[qual];
 
-		/* zero out trailing bytes */
+		/* Zero out trailing bytes. */
 		memset((char *)index + write_size, 0,
 		       adj_write_size - write_size);
 		index += adj_write_size;
@@ -410,7 +480,7 @@ static int ds_write(struct ds_context *context, enum ds_qualifier qual,
  * Later architectures use 64bit pointers throughout, whereas earlier
  * architectures use 32bit pointers in 32bit mode.
  *
- * We compute the base address for the first 8 fields based on:
+ * We compute the base address for the fields based on:
  * - the field size stored in the DS configuration
  * - the relative field position
  *
@@ -431,23 +501,23 @@ enum bts_field {
 	bts_to,
 	bts_flags,
 
-	bts_qual = bts_from,
-	bts_jiffies = bts_to,
-	bts_pid = bts_flags,
+	bts_qual		= bts_from,
+	bts_clock		= bts_to,
+	bts_pid			= bts_flags,
 
-	bts_qual_mask = (bts_qual_max - 1),
-	bts_escape = ((unsigned long)-1 & ~bts_qual_mask)
+	bts_qual_mask		= (bts_qual_max - 1),
+	bts_escape		= ((unsigned long)-1 & ~bts_qual_mask)
 };
 
 static inline unsigned long bts_get(const char *base, enum bts_field field)
 {
-	base += (ds_cfg.sizeof_field * field);
+	base += (ds_cfg.sizeof_ptr_field * field);
 	return *(unsigned long *)base;
 }
 
 static inline void bts_set(char *base, enum bts_field field, unsigned long val)
 {
-	base += (ds_cfg.sizeof_field * field);;
+	base += (ds_cfg.sizeof_ptr_field * field);;
 	(*(unsigned long *)base) = val;
 }
 
@@ -463,8 +533,8 @@ static inline void bts_set(char *base, enum bts_field field, unsigned long val)
  *
  * return: bytes read/written on success; -Eerrno, otherwise
  */
-static int bts_read(struct bts_tracer *tracer, const void *at,
-		    struct bts_struct *out)
+static int
+bts_read(struct bts_tracer *tracer, const void *at, struct bts_struct *out)
 {
 	if (!tracer)
 		return -EINVAL;
@@ -478,8 +548,8 @@ static int bts_read(struct bts_tracer *tracer, const void *at,
 	memset(out, 0, sizeof(*out));
 	if ((bts_get(at, bts_qual) & ~bts_qual_mask) == bts_escape) {
 		out->qualifier = (bts_get(at, bts_qual) & bts_qual_mask);
-		out->variant.timestamp.jiffies = bts_get(at, bts_jiffies);
-		out->variant.timestamp.pid = bts_get(at, bts_pid);
+		out->variant.event.clock = bts_get(at, bts_clock);
+		out->variant.event.pid = bts_get(at, bts_pid);
 	} else {
 		out->qualifier = bts_branch;
 		out->variant.lbr.from = bts_get(at, bts_from);
@@ -516,8 +586,8 @@ static int bts_write(struct bts_tracer *tracer, const struct bts_struct *in)
 	case bts_task_arrives:
 	case bts_task_departs:
 		bts_set(raw, bts_qual, (bts_escape | in->qualifier));
-		bts_set(raw, bts_jiffies, in->variant.timestamp.jiffies);
-		bts_set(raw, bts_pid, in->variant.timestamp.pid);
+		bts_set(raw, bts_clock, in->variant.event.clock);
+		bts_set(raw, bts_pid, in->variant.event.pid);
 		break;
 	default:
 		return -EINVAL;
@@ -555,7 +625,8 @@ static void ds_init_ds_trace(struct ds_trace *trace, enum ds_qualifier qual,
 			     unsigned int flags) {
 	unsigned long buffer, adj;
 
-	/* adjust the buffer address and size to meet alignment
+	/*
+	 * Adjust the buffer address and size to meet alignment
 	 * constraints:
 	 * - buffer is double-word aligned
 	 * - size is multiple of record size
@@ -577,9 +648,11 @@ static void ds_init_ds_trace(struct ds_trace *trace, enum ds_qualifier qual,
 	trace->begin = (void *)buffer;
 	trace->top = trace->begin;
 	trace->end = (void *)(buffer + size);
-	/* The value for 'no threshold' is -1, which will set the
+	/*
+	 * The value for 'no threshold' is -1, which will set the
 	 * threshold outside of the buffer, just like we want it.
 	 */
+	ith *= ds_cfg.sizeof_rec[qual];
 	trace->ith = (void *)(buffer + size - ith);
 
 	trace->flags = flags;
@@ -588,18 +661,27 @@ static void ds_init_ds_trace(struct ds_trace *trace, enum ds_qualifier qual,
 
 static int ds_request(struct ds_tracer *tracer, struct ds_trace *trace,
 		      enum ds_qualifier qual, struct task_struct *task,
-		      void *base, size_t size, size_t th, unsigned int flags)
+		      int cpu, void *base, size_t size, size_t th)
 {
 	struct ds_context *context;
 	int error;
+	size_t req_size;
+
+	error = -EOPNOTSUPP;
+	if (!ds_cfg.sizeof_rec[qual])
+		goto out;
 
 	error = -EINVAL;
 	if (!base)
 		goto out;
 
-	/* we require some space to do alignment adjustments below */
+	req_size = ds_cfg.sizeof_rec[qual];
+	/* We might need space for alignment adjustments. */
+	if (!IS_ALIGNED((unsigned long)base, DS_ALIGNMENT))
+		req_size += DS_ALIGNMENT;
+
 	error = -EINVAL;
-	if (size < (DS_ALIGNMENT + ds_cfg.sizeof_rec[qual]))
+	if (size < req_size)
 		goto out;
 
 	if (th != (size_t)-1) {
@@ -614,182 +696,318 @@ static int ds_request(struct ds_tracer *tracer, struct ds_trace *trace,
 	tracer->size = size;
 
 	error = -ENOMEM;
-	context = ds_get_context(task);
+	context = ds_get_context(task, cpu);
 	if (!context)
 		goto out;
 	tracer->context = context;
 
-	ds_init_ds_trace(trace, qual, base, size, th, flags);
+	/*
+	 * Defer any tracer-specific initialization work for the context until
+	 * context ownership has been clarified.
+	 */
 
 	error = 0;
  out:
 	return error;
 }
 
-struct bts_tracer *ds_request_bts(struct task_struct *task,
-				  void *base, size_t size,
-				  bts_ovfl_callback_t ovfl, size_t th,
-				  unsigned int flags)
+static struct bts_tracer *ds_request_bts(struct task_struct *task, int cpu,
+					 void *base, size_t size,
+					 bts_ovfl_callback_t ovfl, size_t th,
+					 unsigned int flags)
 {
 	struct bts_tracer *tracer;
-	unsigned long irq;
 	int error;
 
+	/* Buffer overflow notification is not yet implemented. */
 	error = -EOPNOTSUPP;
-	if (!ds_cfg.ctl[dsf_bts])
+	if (ovfl)
 		goto out;
 
-	/* buffer overflow notification is not yet implemented */
-	error = -EOPNOTSUPP;
-	if (ovfl)
+	error = get_tracer(task);
+	if (error < 0)
 		goto out;
 
 	error = -ENOMEM;
 	tracer = kzalloc(sizeof(*tracer), GFP_KERNEL);
 	if (!tracer)
-		goto out;
+		goto out_put_tracer;
 	tracer->ovfl = ovfl;
 
+	/* Do some more error checking and acquire a tracing context. */
 	error = ds_request(&tracer->ds, &tracer->trace.ds,
-			   ds_bts, task, base, size, th, flags);
+			   ds_bts, task, cpu, base, size, th);
 	if (error < 0)
 		goto out_tracer;
 
-
-	spin_lock_irqsave(&ds_lock, irq);
-
-	error = -EPERM;
-	if (!check_tracer(task))
-		goto out_unlock;
-	get_tracer(task);
+	/* Claim the bts part of the tracing context we acquired above. */
+	spin_lock_irq(&ds_lock);
 
 	error = -EPERM;
 	if (tracer->ds.context->bts_master)
-		goto out_put_tracer;
+		goto out_unlock;
 	tracer->ds.context->bts_master = tracer;
 
-	spin_unlock_irqrestore(&ds_lock, irq);
+	spin_unlock_irq(&ds_lock);
 
+	/*
+	 * Now that we own the bts part of the context, let's complete the
+	 * initialization for that part.
+	 */
+	ds_init_ds_trace(&tracer->trace.ds, ds_bts, base, size, th, flags);
+	ds_write_config(tracer->ds.context, &tracer->trace.ds, ds_bts);
+	ds_install_ds_area(tracer->ds.context);
 
 	tracer->trace.read  = bts_read;
 	tracer->trace.write = bts_write;
 
-	ds_write_config(tracer->ds.context, &tracer->trace.ds, ds_bts);
+	/* Start tracing. */
 	ds_resume_bts(tracer);
 
 	return tracer;
 
- out_put_tracer:
-	put_tracer(task);
  out_unlock:
-	spin_unlock_irqrestore(&ds_lock, irq);
+	spin_unlock_irq(&ds_lock);
 	ds_put_context(tracer->ds.context);
  out_tracer:
 	kfree(tracer);
+ out_put_tracer:
+	put_tracer(task);
  out:
 	return ERR_PTR(error);
 }
 
-struct pebs_tracer *ds_request_pebs(struct task_struct *task,
-				    void *base, size_t size,
-				    pebs_ovfl_callback_t ovfl, size_t th,
-				    unsigned int flags)
+struct bts_tracer *ds_request_bts_task(struct task_struct *task,
+				       void *base, size_t size,
+				       bts_ovfl_callback_t ovfl,
+				       size_t th, unsigned int flags)
+{
+	return ds_request_bts(task, 0, base, size, ovfl, th, flags);
+}
+
+struct bts_tracer *ds_request_bts_cpu(int cpu, void *base, size_t size,
+				      bts_ovfl_callback_t ovfl,
+				      size_t th, unsigned int flags)
+{
+	return ds_request_bts(NULL, cpu, base, size, ovfl, th, flags);
+}
+
+static struct pebs_tracer *ds_request_pebs(struct task_struct *task, int cpu,
+					   void *base, size_t size,
+					   pebs_ovfl_callback_t ovfl, size_t th,
+					   unsigned int flags)
 {
 	struct pebs_tracer *tracer;
-	unsigned long irq;
 	int error;
 
-	/* buffer overflow notification is not yet implemented */
+	/* Buffer overflow notification is not yet implemented. */
 	error = -EOPNOTSUPP;
 	if (ovfl)
 		goto out;
 
+	error = get_tracer(task);
+	if (error < 0)
+		goto out;
+
 	error = -ENOMEM;
 	tracer = kzalloc(sizeof(*tracer), GFP_KERNEL);
 	if (!tracer)
-		goto out;
+		goto out_put_tracer;
 	tracer->ovfl = ovfl;
 
+	/* Do some more error checking and acquire a tracing context. */
 	error = ds_request(&tracer->ds, &tracer->trace.ds,
-			   ds_pebs, task, base, size, th, flags);
+			   ds_pebs, task, cpu, base, size, th);
 	if (error < 0)
 		goto out_tracer;
 
-	spin_lock_irqsave(&ds_lock, irq);
-
-	error = -EPERM;
-	if (!check_tracer(task))
-		goto out_unlock;
-	get_tracer(task);
+	/* Claim the pebs part of the tracing context we acquired above. */
+	spin_lock_irq(&ds_lock);
 
 	error = -EPERM;
 	if (tracer->ds.context->pebs_master)
-		goto out_put_tracer;
+		goto out_unlock;
 	tracer->ds.context->pebs_master = tracer;
 
-	spin_unlock_irqrestore(&ds_lock, irq);
+	spin_unlock_irq(&ds_lock);
 
+	/*
+	 * Now that we own the pebs part of the context, let's complete the
+	 * initialization for that part.
+	 */
+	ds_init_ds_trace(&tracer->trace.ds, ds_pebs, base, size, th, flags);
 	ds_write_config(tracer->ds.context, &tracer->trace.ds, ds_pebs);
+	ds_install_ds_area(tracer->ds.context);
+
+	/* Start tracing. */
 	ds_resume_pebs(tracer);
 
 	return tracer;
 
- out_put_tracer:
-	put_tracer(task);
  out_unlock:
-	spin_unlock_irqrestore(&ds_lock, irq);
+	spin_unlock_irq(&ds_lock);
 	ds_put_context(tracer->ds.context);
  out_tracer:
 	kfree(tracer);
+ out_put_tracer:
+	put_tracer(task);
  out:
 	return ERR_PTR(error);
 }
 
-void ds_release_bts(struct bts_tracer *tracer)
+struct pebs_tracer *ds_request_pebs_task(struct task_struct *task,
+					 void *base, size_t size,
+					 pebs_ovfl_callback_t ovfl,
+					 size_t th, unsigned int flags)
 {
-	if (!tracer)
-		return;
+	return ds_request_pebs(task, 0, base, size, ovfl, th, flags);
+}
 
-	ds_suspend_bts(tracer);
+struct pebs_tracer *ds_request_pebs_cpu(int cpu, void *base, size_t size,
+					pebs_ovfl_callback_t ovfl,
+					size_t th, unsigned int flags)
+{
+	return ds_request_pebs(NULL, cpu, base, size, ovfl, th, flags);
+}
+
+static void ds_free_bts(struct bts_tracer *tracer)
+{
+	struct task_struct *task;
+
+	task = tracer->ds.context->task;
 
 	WARN_ON_ONCE(tracer->ds.context->bts_master != tracer);
 	tracer->ds.context->bts_master = NULL;
 
-	put_tracer(tracer->ds.context->task);
+	/* Make sure tracing stopped and the tracer is not in use. */
+	if (task && (task != current))
+		wait_task_context_switch(task);
+
 	ds_put_context(tracer->ds.context);
+	put_tracer(task);
 
 	kfree(tracer);
 }
 
+void ds_release_bts(struct bts_tracer *tracer)
+{
+	might_sleep();
+
+	if (!tracer)
+		return;
+
+	ds_suspend_bts(tracer);
+	ds_free_bts(tracer);
+}
+
+int ds_release_bts_noirq(struct bts_tracer *tracer)
+{
+	struct task_struct *task;
+	unsigned long irq;
+	int error;
+
+	if (!tracer)
+		return 0;
+
+	task = tracer->ds.context->task;
+
+	local_irq_save(irq);
+
+	error = -EPERM;
+	if (!task &&
+	    (tracer->ds.context->cpu != smp_processor_id()))
+		goto out;
+
+	error = -EPERM;
+	if (task && (task != current))
+		goto out;
+
+	ds_suspend_bts_noirq(tracer);
+	ds_free_bts(tracer);
+
+	error = 0;
+ out:
+	local_irq_restore(irq);
+	return error;
+}
+
+static void update_task_debugctlmsr(struct task_struct *task,
+				    unsigned long debugctlmsr)
+{
+	task->thread.debugctlmsr = debugctlmsr;
+
+	get_cpu();
+	if (task == current)
+		update_debugctlmsr(debugctlmsr);
+	put_cpu();
+}
+
 void ds_suspend_bts(struct bts_tracer *tracer)
 {
 	struct task_struct *task;
+	unsigned long debugctlmsr;
+	int cpu;
 
 	if (!tracer)
 		return;
 
+	tracer->flags = 0;
+
 	task = tracer->ds.context->task;
+	cpu  = tracer->ds.context->cpu;
 
-	if (!task || (task == current))
-		update_debugctlmsr(get_debugctlmsr() & ~BTS_CONTROL);
+	WARN_ON(!task && irqs_disabled());
 
-	if (task) {
-		task->thread.debugctlmsr &= ~BTS_CONTROL;
+	debugctlmsr = (task ?
+		       task->thread.debugctlmsr :
+		       get_debugctlmsr_on_cpu(cpu));
+	debugctlmsr &= ~BTS_CONTROL;
 
-		if (!task->thread.debugctlmsr)
-			clear_tsk_thread_flag(task, TIF_DEBUGCTLMSR);
-	}
+	if (task)
+		update_task_debugctlmsr(task, debugctlmsr);
+	else
+		update_debugctlmsr_on_cpu(cpu, debugctlmsr);
 }
 
-void ds_resume_bts(struct bts_tracer *tracer)
+int ds_suspend_bts_noirq(struct bts_tracer *tracer)
 {
 	struct task_struct *task;
-	unsigned long control;
+	unsigned long debugctlmsr, irq;
+	int cpu, error = 0;
 
 	if (!tracer)
-		return;
+		return 0;
+
+	tracer->flags = 0;
 
 	task = tracer->ds.context->task;
+	cpu  = tracer->ds.context->cpu;
+
+	local_irq_save(irq);
+
+	error = -EPERM;
+	if (!task && (cpu != smp_processor_id()))
+		goto out;
+
+	debugctlmsr = (task ?
+		       task->thread.debugctlmsr :
+		       get_debugctlmsr());
+	debugctlmsr &= ~BTS_CONTROL;
+
+	if (task)
+		update_task_debugctlmsr(task, debugctlmsr);
+	else
+		update_debugctlmsr(debugctlmsr);
+
+	error = 0;
+ out:
+	local_irq_restore(irq);
+	return error;
+}
+
+static unsigned long ds_bts_control(struct bts_tracer *tracer)
+{
+	unsigned long control;
 
 	control = ds_cfg.ctl[dsf_bts];
 	if (!(tracer->trace.ds.flags & BTS_KERNEL))
@@ -797,41 +1015,149 @@ void ds_resume_bts(struct bts_tracer *tracer)
 	if (!(tracer->trace.ds.flags & BTS_USER))
 		control |= ds_cfg.ctl[dsf_bts_user];
 
-	if (task) {
-		task->thread.debugctlmsr |= control;
-		set_tsk_thread_flag(task, TIF_DEBUGCTLMSR);
-	}
-
-	if (!task || (task == current))
-		update_debugctlmsr(get_debugctlmsr() | control);
+	return control;
 }
 
-void ds_release_pebs(struct pebs_tracer *tracer)
+void ds_resume_bts(struct bts_tracer *tracer)
 {
+	struct task_struct *task;
+	unsigned long debugctlmsr;
+	int cpu;
+
 	if (!tracer)
 		return;
 
-	ds_suspend_pebs(tracer);
+	tracer->flags = tracer->trace.ds.flags;
+
+	task = tracer->ds.context->task;
+	cpu  = tracer->ds.context->cpu;
+
+	WARN_ON(!task && irqs_disabled());
+
+	debugctlmsr = (task ?
+		       task->thread.debugctlmsr :
+		       get_debugctlmsr_on_cpu(cpu));
+	debugctlmsr |= ds_bts_control(tracer);
+
+	if (task)
+		update_task_debugctlmsr(task, debugctlmsr);
+	else
+		update_debugctlmsr_on_cpu(cpu, debugctlmsr);
+}
+
+int ds_resume_bts_noirq(struct bts_tracer *tracer)
+{
+	struct task_struct *task;
+	unsigned long debugctlmsr, irq;
+	int cpu, error = 0;
+
+	if (!tracer)
+		return 0;
+
+	tracer->flags = tracer->trace.ds.flags;
+
+	task = tracer->ds.context->task;
+	cpu  = tracer->ds.context->cpu;
+
+	local_irq_save(irq);
+
+	error = -EPERM;
+	if (!task && (cpu != smp_processor_id()))
+		goto out;
+
+	debugctlmsr = (task ?
+		       task->thread.debugctlmsr :
+		       get_debugctlmsr());
+	debugctlmsr |= ds_bts_control(tracer);
+
+	if (task)
+		update_task_debugctlmsr(task, debugctlmsr);
+	else
+		update_debugctlmsr(debugctlmsr);
+
+	error = 0;
+ out:
+	local_irq_restore(irq);
+	return error;
+}
+
+static void ds_free_pebs(struct pebs_tracer *tracer)
+{
+	struct task_struct *task;
+
+	task = tracer->ds.context->task;
 
 	WARN_ON_ONCE(tracer->ds.context->pebs_master != tracer);
 	tracer->ds.context->pebs_master = NULL;
 
-	put_tracer(tracer->ds.context->task);
 	ds_put_context(tracer->ds.context);
+	put_tracer(task);
 
 	kfree(tracer);
 }
 
+void ds_release_pebs(struct pebs_tracer *tracer)
+{
+	might_sleep();
+
+	if (!tracer)
+		return;
+
+	ds_suspend_pebs(tracer);
+	ds_free_pebs(tracer);
+}
+
+int ds_release_pebs_noirq(struct pebs_tracer *tracer)
+{
+	struct task_struct *task;
+	unsigned long irq;
+	int error;
+
+	if (!tracer)
+		return 0;
+
+	task = tracer->ds.context->task;
+
+	local_irq_save(irq);
+
+	error = -EPERM;
+	if (!task &&
+	    (tracer->ds.context->cpu != smp_processor_id()))
+		goto out;
+
+	error = -EPERM;
+	if (task && (task != current))
+		goto out;
+
+	ds_suspend_pebs_noirq(tracer);
+	ds_free_pebs(tracer);
+
+	error = 0;
+ out:
+	local_irq_restore(irq);
+	return error;
+}
+
 void ds_suspend_pebs(struct pebs_tracer *tracer)
 {
 
 }
 
+int ds_suspend_pebs_noirq(struct pebs_tracer *tracer)
+{
+	return 0;
+}
+
 void ds_resume_pebs(struct pebs_tracer *tracer)
 {
 
 }
 
+int ds_resume_pebs_noirq(struct pebs_tracer *tracer)
+{
+	return 0;
+}
+
 const struct bts_trace *ds_read_bts(struct bts_tracer *tracer)
 {
 	if (!tracer)
@@ -847,8 +1173,12 @@ const struct pebs_trace *ds_read_pebs(struct pebs_tracer *tracer)
 		return NULL;
 
 	ds_read_config(tracer->ds.context, &tracer->trace.ds, ds_pebs);
-	tracer->trace.reset_value =
-		*(u64 *)(tracer->ds.context->ds + (ds_cfg.sizeof_field * 8));
+
+	tracer->trace.counters = ds_cfg.nr_counter_reset;
+	memcpy(tracer->trace.counter_reset,
+	       tracer->ds.context->ds +
+	       (NUM_DS_PTR_FIELDS * ds_cfg.sizeof_ptr_field),
+	       ds_cfg.nr_counter_reset * PEBS_RESET_FIELD_SIZE);
 
 	return &tracer->trace;
 }
@@ -873,18 +1203,24 @@ int ds_reset_pebs(struct pebs_tracer *tracer)
 
 	tracer->trace.ds.top = tracer->trace.ds.begin;
 
-	ds_set(tracer->ds.context->ds, ds_bts, ds_index,
+	ds_set(tracer->ds.context->ds, ds_pebs, ds_index,
 	       (unsigned long)tracer->trace.ds.top);
 
 	return 0;
 }
 
-int ds_set_pebs_reset(struct pebs_tracer *tracer, u64 value)
+int ds_set_pebs_reset(struct pebs_tracer *tracer,
+		      unsigned int counter, u64 value)
 {
 	if (!tracer)
 		return -EINVAL;
 
-	*(u64 *)(tracer->ds.context->ds + (ds_cfg.sizeof_field * 8)) = value;
+	if (ds_cfg.nr_counter_reset < counter)
+		return -EINVAL;
+
+	*(u64 *)(tracer->ds.context->ds +
+		 (NUM_DS_PTR_FIELDS * ds_cfg.sizeof_ptr_field) +
+		 (counter * PEBS_RESET_FIELD_SIZE)) = value;
 
 	return 0;
 }
@@ -894,73 +1230,117 @@ static const struct ds_configuration ds_cfg_netburst = {
 	.ctl[dsf_bts]		= (1 << 2) | (1 << 3),
 	.ctl[dsf_bts_kernel]	= (1 << 5),
 	.ctl[dsf_bts_user]	= (1 << 6),
-
-	.sizeof_field		= sizeof(long),
-	.sizeof_rec[ds_bts]	= sizeof(long) * 3,
-#ifdef __i386__
-	.sizeof_rec[ds_pebs]	= sizeof(long) * 10,
-#else
-	.sizeof_rec[ds_pebs]	= sizeof(long) * 18,
-#endif
+	.nr_counter_reset	= 1,
 };
 static const struct ds_configuration ds_cfg_pentium_m = {
 	.name = "Pentium M",
 	.ctl[dsf_bts]		= (1 << 6) | (1 << 7),
-
-	.sizeof_field		= sizeof(long),
-	.sizeof_rec[ds_bts]	= sizeof(long) * 3,
-#ifdef __i386__
-	.sizeof_rec[ds_pebs]	= sizeof(long) * 10,
-#else
-	.sizeof_rec[ds_pebs]	= sizeof(long) * 18,
-#endif
+	.nr_counter_reset	= 1,
 };
 static const struct ds_configuration ds_cfg_core2_atom = {
 	.name = "Core 2/Atom",
 	.ctl[dsf_bts]		= (1 << 6) | (1 << 7),
 	.ctl[dsf_bts_kernel]	= (1 << 9),
 	.ctl[dsf_bts_user]	= (1 << 10),
-
-	.sizeof_field		= 8,
-	.sizeof_rec[ds_bts]	= 8 * 3,
-	.sizeof_rec[ds_pebs]	= 8 * 18,
+	.nr_counter_reset	= 1,
+};
+static const struct ds_configuration ds_cfg_core_i7 = {
+	.name = "Core i7",
+	.ctl[dsf_bts]		= (1 << 6) | (1 << 7),
+	.ctl[dsf_bts_kernel]	= (1 << 9),
+	.ctl[dsf_bts_user]	= (1 << 10),
+	.nr_counter_reset	= 4,
 };
 
 static void
-ds_configure(const struct ds_configuration *cfg)
+ds_configure(const struct ds_configuration *cfg,
+	     struct cpuinfo_x86 *cpu)
 {
+	unsigned long nr_pebs_fields = 0;
+
+	printk(KERN_INFO "[ds] using %s configuration\n", cfg->name);
+
+#ifdef __i386__
+	nr_pebs_fields = 10;
+#else
+	nr_pebs_fields = 18;
+#endif
+
+	/*
+	 * Starting with version 2, architectural performance
+	 * monitoring supports a format specifier.
+	 */
+	if ((cpuid_eax(0xa) & 0xff) > 1) {
+		unsigned long perf_capabilities, format;
+
+		rdmsrl(MSR_IA32_PERF_CAPABILITIES, perf_capabilities);
+
+		format = (perf_capabilities >> 8) & 0xf;
+
+		switch (format) {
+		case 0:
+			nr_pebs_fields = 18;
+			break;
+		case 1:
+			nr_pebs_fields = 22;
+			break;
+		default:
+			printk(KERN_INFO
+			       "[ds] unknown PEBS format: %lu\n", format);
+			nr_pebs_fields = 0;
+			break;
+		}
+	}
+
 	memset(&ds_cfg, 0, sizeof(ds_cfg));
 	ds_cfg = *cfg;
 
-	printk(KERN_INFO "[ds] using %s configuration\n", ds_cfg.name);
+	ds_cfg.sizeof_ptr_field =
+		(cpu_has(cpu, X86_FEATURE_DTES64) ? 8 : 4);
+
+	ds_cfg.sizeof_rec[ds_bts]  = ds_cfg.sizeof_ptr_field * 3;
+	ds_cfg.sizeof_rec[ds_pebs] = ds_cfg.sizeof_ptr_field * nr_pebs_fields;
 
-	if (!cpu_has_bts) {
-		ds_cfg.ctl[dsf_bts] = 0;
+	if (!cpu_has(cpu, X86_FEATURE_BTS)) {
+		ds_cfg.sizeof_rec[ds_bts] = 0;
 		printk(KERN_INFO "[ds] bts not available\n");
 	}
-	if (!cpu_has_pebs)
+	if (!cpu_has(cpu, X86_FEATURE_PEBS)) {
+		ds_cfg.sizeof_rec[ds_pebs] = 0;
 		printk(KERN_INFO "[ds] pebs not available\n");
+	}
+
+	printk(KERN_INFO "[ds] sizes: address: %u bit, ",
+	       8 * ds_cfg.sizeof_ptr_field);
+	printk("bts/pebs record: %u/%u bytes\n",
+	       ds_cfg.sizeof_rec[ds_bts], ds_cfg.sizeof_rec[ds_pebs]);
 
-	WARN_ON_ONCE(MAX_SIZEOF_DS < (12 * ds_cfg.sizeof_field));
+	WARN_ON_ONCE(MAX_PEBS_COUNTERS < ds_cfg.nr_counter_reset);
 }
 
 void __cpuinit ds_init_intel(struct cpuinfo_x86 *c)
 {
+	/* Only configure the first cpu. Others are identical. */
+	if (ds_cfg.name)
+		return;
+
 	switch (c->x86) {
 	case 0x6:
 		switch (c->x86_model) {
 		case 0x9:
 		case 0xd: /* Pentium M */
-			ds_configure(&ds_cfg_pentium_m);
+			ds_configure(&ds_cfg_pentium_m, c);
 			break;
 		case 0xf:
 		case 0x17: /* Core2 */
 		case 0x1c: /* Atom */
-			ds_configure(&ds_cfg_core2_atom);
+			ds_configure(&ds_cfg_core2_atom, c);
+			break;
+		case 0x1a: /* Core i7 */
+			ds_configure(&ds_cfg_core_i7, c);
 			break;
-		case 0x1a: /* i7 */
 		default:
-			/* sorry, don't know about them */
+			/* Sorry, don't know about them. */
 			break;
 		}
 		break;
@@ -969,64 +1349,89 @@ void __cpuinit ds_init_intel(struct cpuinfo_x86 *c)
 		case 0x0:
 		case 0x1:
 		case 0x2: /* Netburst */
-			ds_configure(&ds_cfg_netburst);
+			ds_configure(&ds_cfg_netburst, c);
 			break;
 		default:
-			/* sorry, don't know about them */
+			/* Sorry, don't know about them. */
 			break;
 		}
 		break;
 	default:
-		/* sorry, don't know about them */
+		/* Sorry, don't know about them. */
 		break;
 	}
 }
 
+static inline void ds_take_timestamp(struct ds_context *context,
+				     enum bts_qualifier qualifier,
+				     struct task_struct *task)
+{
+	struct bts_tracer *tracer = context->bts_master;
+	struct bts_struct ts;
+
+	/* Prevent compilers from reading the tracer pointer twice. */
+	barrier();
+
+	if (!tracer || !(tracer->flags & BTS_TIMESTAMPS))
+		return;
+
+	memset(&ts, 0, sizeof(ts));
+	ts.qualifier		= qualifier;
+	ts.variant.event.clock	= trace_clock_global();
+	ts.variant.event.pid	= task->pid;
+
+	bts_write(tracer, &ts);
+}
+
 /*
  * Change the DS configuration from tracing prev to tracing next.
  */
 void ds_switch_to(struct task_struct *prev, struct task_struct *next)
 {
-	struct ds_context *prev_ctx = prev->thread.ds_ctx;
-	struct ds_context *next_ctx = next->thread.ds_ctx;
+	struct ds_context *prev_ctx	= prev->thread.ds_ctx;
+	struct ds_context *next_ctx	= next->thread.ds_ctx;
+	unsigned long debugctlmsr	= next->thread.debugctlmsr;
+
+	/* Make sure all data is read before we start. */
+	barrier();
 
 	if (prev_ctx) {
 		update_debugctlmsr(0);
 
-		if (prev_ctx->bts_master &&
-		    (prev_ctx->bts_master->trace.ds.flags & BTS_TIMESTAMPS)) {
-			struct bts_struct ts = {
-				.qualifier = bts_task_departs,
-				.variant.timestamp.jiffies = jiffies_64,
-				.variant.timestamp.pid = prev->pid
-			};
-			bts_write(prev_ctx->bts_master, &ts);
-		}
+		ds_take_timestamp(prev_ctx, bts_task_departs, prev);
 	}
 
 	if (next_ctx) {
-		if (next_ctx->bts_master &&
-		    (next_ctx->bts_master->trace.ds.flags & BTS_TIMESTAMPS)) {
-			struct bts_struct ts = {
-				.qualifier = bts_task_arrives,
-				.variant.timestamp.jiffies = jiffies_64,
-				.variant.timestamp.pid = next->pid
-			};
-			bts_write(next_ctx->bts_master, &ts);
-		}
+		ds_take_timestamp(next_ctx, bts_task_arrives, next);
 
 		wrmsrl(MSR_IA32_DS_AREA, (unsigned long)next_ctx->ds);
 	}
 
-	update_debugctlmsr(next->thread.debugctlmsr);
+	update_debugctlmsr(debugctlmsr);
 }
 
-void ds_copy_thread(struct task_struct *tsk, struct task_struct *father)
+static __init int ds_selftest(void)
 {
-	clear_tsk_thread_flag(tsk, TIF_DS_AREA_MSR);
-	tsk->thread.ds_ctx = NULL;
-}
+	if (ds_cfg.sizeof_rec[ds_bts]) {
+		int error;
 
-void ds_exit_thread(struct task_struct *tsk)
-{
+		error = ds_selftest_bts();
+		if (error) {
+			WARN(1, "[ds] selftest failed. disabling bts.\n");
+			ds_cfg.sizeof_rec[ds_bts] = 0;
+		}
+	}
+
+	if (ds_cfg.sizeof_rec[ds_pebs]) {
+		int error;
+
+		error = ds_selftest_pebs();
+		if (error) {
+			WARN(1, "[ds] selftest failed. disabling pebs.\n");
+			ds_cfg.sizeof_rec[ds_pebs] = 0;
+		}
+	}
+
+	return 0;
 }
+device_initcall(ds_selftest);
diff --git a/arch/x86/kernel/ds_selftest.c b/arch/x86/kernel/ds_selftest.c
new file mode 100644
index 0000000..6bc7c19
--- /dev/null
+++ b/arch/x86/kernel/ds_selftest.c
@@ -0,0 +1,408 @@
+/*
+ * Debug Store support - selftest
+ *
+ *
+ * Copyright (C) 2009 Intel Corporation.
+ * Markus Metzger <markus.t.metzger@intel.com>, 2009
+ */
+
+#include "ds_selftest.h"
+
+#include <linux/kernel.h>
+#include <linux/string.h>
+#include <linux/smp.h>
+#include <linux/cpu.h>
+
+#include <asm/ds.h>
+
+
+#define BUFFER_SIZE		521	/* Intentionally chose an odd size. */
+#define SMALL_BUFFER_SIZE	24	/* A single bts entry. */
+
+struct ds_selftest_bts_conf {
+	struct bts_tracer *tracer;
+	int error;
+	int (*suspend)(struct bts_tracer *);
+	int (*resume)(struct bts_tracer *);
+};
+
+static int ds_selftest_bts_consistency(const struct bts_trace *trace)
+{
+	int error = 0;
+
+	if (!trace) {
+		printk(KERN_CONT "failed to access trace...");
+		/* Bail out. Other tests are pointless. */
+		return -1;
+	}
+
+	if (!trace->read) {
+		printk(KERN_CONT "bts read not available...");
+		error = -1;
+	}
+
+	/* Do some sanity checks on the trace configuration. */
+	if (!trace->ds.n) {
+		printk(KERN_CONT "empty bts buffer...");
+		error = -1;
+	}
+	if (!trace->ds.size) {
+		printk(KERN_CONT "bad bts trace setup...");
+		error = -1;
+	}
+	if (trace->ds.end !=
+	    (char *)trace->ds.begin + (trace->ds.n * trace->ds.size)) {
+		printk(KERN_CONT "bad bts buffer setup...");
+		error = -1;
+	}
+	/*
+	 * We allow top in [begin; end], since its not clear when the
+	 * overflow adjustment happens: after the increment or before the
+	 * write.
+	 */
+	if ((trace->ds.top < trace->ds.begin) ||
+	    (trace->ds.end < trace->ds.top)) {
+		printk(KERN_CONT "bts top out of bounds...");
+		error = -1;
+	}
+
+	return error;
+}
+
+static int ds_selftest_bts_read(struct bts_tracer *tracer,
+				const struct bts_trace *trace,
+				const void *from, const void *to)
+{
+	const unsigned char *at;
+
+	/*
+	 * Check a few things which do not belong to this test.
+	 * They should be covered by other tests.
+	 */
+	if (!trace)
+		return -1;
+
+	if (!trace->read)
+		return -1;
+
+	if (to < from)
+		return -1;
+
+	if (from < trace->ds.begin)
+		return -1;
+
+	if (trace->ds.end < to)
+		return -1;
+
+	if (!trace->ds.size)
+		return -1;
+
+	/* Now to the test itself. */
+	for (at = from; (void *)at < to; at += trace->ds.size) {
+		struct bts_struct bts;
+		unsigned long index;
+		int error;
+
+		if (((void *)at - trace->ds.begin) % trace->ds.size) {
+			printk(KERN_CONT
+			       "read from non-integer index...");
+			return -1;
+		}
+		index = ((void *)at - trace->ds.begin) / trace->ds.size;
+
+		memset(&bts, 0, sizeof(bts));
+		error = trace->read(tracer, at, &bts);
+		if (error < 0) {
+			printk(KERN_CONT
+			       "error reading bts trace at [%lu] (0x%p)...",
+			       index, at);
+			return error;
+		}
+
+		switch (bts.qualifier) {
+		case BTS_BRANCH:
+			break;
+		default:
+			printk(KERN_CONT
+			       "unexpected bts entry %llu at [%lu] (0x%p)...",
+			       bts.qualifier, index, at);
+			return -1;
+		}
+	}
+
+	return 0;
+}
+
+static void ds_selftest_bts_cpu(void *arg)
+{
+	struct ds_selftest_bts_conf *conf = arg;
+	const struct bts_trace *trace;
+	void *top;
+
+	if (IS_ERR(conf->tracer)) {
+		conf->error = PTR_ERR(conf->tracer);
+		conf->tracer = NULL;
+
+		printk(KERN_CONT
+		       "initialization failed (err: %d)...", conf->error);
+		return;
+	}
+
+	/* We should meanwhile have enough trace. */
+	conf->error = conf->suspend(conf->tracer);
+	if (conf->error < 0)
+		return;
+
+	/* Let's see if we can access the trace. */
+	trace = ds_read_bts(conf->tracer);
+
+	conf->error = ds_selftest_bts_consistency(trace);
+	if (conf->error < 0)
+		return;
+
+	/* If everything went well, we should have a few trace entries. */
+	if (trace->ds.top == trace->ds.begin) {
+		/*
+		 * It is possible but highly unlikely that we got a
+		 * buffer overflow and end up at exactly the same
+		 * position we started from.
+		 * Let's issue a warning, but continue.
+		 */
+		printk(KERN_CONT "no trace/overflow...");
+	}
+
+	/* Let's try to read the trace we collected. */
+	conf->error =
+		ds_selftest_bts_read(conf->tracer, trace,
+				     trace->ds.begin, trace->ds.top);
+	if (conf->error < 0)
+		return;
+
+	/*
+	 * Let's read the trace again.
+	 * Since we suspended tracing, we should get the same result.
+	 */
+	top = trace->ds.top;
+
+	trace = ds_read_bts(conf->tracer);
+	conf->error = ds_selftest_bts_consistency(trace);
+	if (conf->error < 0)
+		return;
+
+	if (top != trace->ds.top) {
+		printk(KERN_CONT "suspend not working...");
+		conf->error = -1;
+		return;
+	}
+
+	/* Let's collect some more trace - see if resume is working. */
+	conf->error = conf->resume(conf->tracer);
+	if (conf->error < 0)
+		return;
+
+	conf->error = conf->suspend(conf->tracer);
+	if (conf->error < 0)
+		return;
+
+	trace = ds_read_bts(conf->tracer);
+
+	conf->error = ds_selftest_bts_consistency(trace);
+	if (conf->error < 0)
+		return;
+
+	if (trace->ds.top == top) {
+		/*
+		 * It is possible but highly unlikely that we got a
+		 * buffer overflow and end up at exactly the same
+		 * position we started from.
+		 * Let's issue a warning and check the full trace.
+		 */
+		printk(KERN_CONT
+		       "no resume progress/overflow...");
+
+		conf->error =
+			ds_selftest_bts_read(conf->tracer, trace,
+					     trace->ds.begin, trace->ds.end);
+	} else if (trace->ds.top < top) {
+		/*
+		 * We had a buffer overflow - the entire buffer should
+		 * contain trace records.
+		 */
+		conf->error =
+			ds_selftest_bts_read(conf->tracer, trace,
+					     trace->ds.begin, trace->ds.end);
+	} else {
+		/*
+		 * It is quite likely that the buffer did not overflow.
+		 * Let's just check the delta trace.
+		 */
+		conf->error =
+			ds_selftest_bts_read(conf->tracer, trace, top,
+					     trace->ds.top);
+	}
+	if (conf->error < 0)
+		return;
+
+	conf->error = 0;
+}
+
+static int ds_suspend_bts_wrap(struct bts_tracer *tracer)
+{
+	ds_suspend_bts(tracer);
+	return 0;
+}
+
+static int ds_resume_bts_wrap(struct bts_tracer *tracer)
+{
+	ds_resume_bts(tracer);
+	return 0;
+}
+
+static void ds_release_bts_noirq_wrap(void *tracer)
+{
+	(void)ds_release_bts_noirq(tracer);
+}
+
+static int ds_selftest_bts_bad_release_noirq(int cpu,
+					     struct bts_tracer *tracer)
+{
+	int error = -EPERM;
+
+	/* Try to release the tracer on the wrong cpu. */
+	get_cpu();
+	if (cpu != smp_processor_id()) {
+		error = ds_release_bts_noirq(tracer);
+		if (error != -EPERM)
+			printk(KERN_CONT "release on wrong cpu...");
+	}
+	put_cpu();
+
+	return error ? 0 : -1;
+}
+
+static int ds_selftest_bts_bad_request_cpu(int cpu, void *buffer)
+{
+	struct bts_tracer *tracer;
+	int error;
+
+	/* Try to request cpu tracing while task tracing is active. */
+	tracer = ds_request_bts_cpu(cpu, buffer, BUFFER_SIZE, NULL,
+				    (size_t)-1, BTS_KERNEL);
+	error = PTR_ERR(tracer);
+	if (!IS_ERR(tracer)) {
+		ds_release_bts(tracer);
+		error = 0;
+	}
+
+	if (error != -EPERM)
+		printk(KERN_CONT "cpu/task tracing overlap...");
+
+	return error ? 0 : -1;
+}
+
+static int ds_selftest_bts_bad_request_task(void *buffer)
+{
+	struct bts_tracer *tracer;
+	int error;
+
+	/* Try to request cpu tracing while task tracing is active. */
+	tracer = ds_request_bts_task(current, buffer, BUFFER_SIZE, NULL,
+				    (size_t)-1, BTS_KERNEL);
+	error = PTR_ERR(tracer);
+	if (!IS_ERR(tracer)) {
+		error = 0;
+		ds_release_bts(tracer);
+	}
+
+	if (error != -EPERM)
+		printk(KERN_CONT "task/cpu tracing overlap...");
+
+	return error ? 0 : -1;
+}
+
+int ds_selftest_bts(void)
+{
+	struct ds_selftest_bts_conf conf;
+	unsigned char buffer[BUFFER_SIZE], *small_buffer;
+	unsigned long irq;
+	int cpu;
+
+	printk(KERN_INFO "[ds] bts selftest...");
+	conf.error = 0;
+
+	small_buffer = (unsigned char *)ALIGN((unsigned long)buffer, 8) + 8;
+
+	get_online_cpus();
+	for_each_online_cpu(cpu) {
+		conf.suspend = ds_suspend_bts_wrap;
+		conf.resume = ds_resume_bts_wrap;
+		conf.tracer =
+			ds_request_bts_cpu(cpu, buffer, BUFFER_SIZE,
+					   NULL, (size_t)-1, BTS_KERNEL);
+		ds_selftest_bts_cpu(&conf);
+		if (conf.error >= 0)
+			conf.error = ds_selftest_bts_bad_request_task(buffer);
+		ds_release_bts(conf.tracer);
+		if (conf.error < 0)
+			goto out;
+
+		conf.suspend = ds_suspend_bts_noirq;
+		conf.resume = ds_resume_bts_noirq;
+		conf.tracer =
+			ds_request_bts_cpu(cpu, buffer, BUFFER_SIZE,
+					   NULL, (size_t)-1, BTS_KERNEL);
+		smp_call_function_single(cpu, ds_selftest_bts_cpu, &conf, 1);
+		if (conf.error >= 0) {
+			conf.error =
+				ds_selftest_bts_bad_release_noirq(cpu,
+								  conf.tracer);
+			/* We must not release the tracer twice. */
+			if (conf.error < 0)
+				conf.tracer = NULL;
+		}
+		if (conf.error >= 0)
+			conf.error = ds_selftest_bts_bad_request_task(buffer);
+		smp_call_function_single(cpu, ds_release_bts_noirq_wrap,
+					 conf.tracer, 1);
+		if (conf.error < 0)
+			goto out;
+	}
+
+	conf.suspend = ds_suspend_bts_wrap;
+	conf.resume = ds_resume_bts_wrap;
+	conf.tracer =
+		ds_request_bts_task(current, buffer, BUFFER_SIZE,
+				    NULL, (size_t)-1, BTS_KERNEL);
+	ds_selftest_bts_cpu(&conf);
+	if (conf.error >= 0)
+		conf.error = ds_selftest_bts_bad_request_cpu(0, buffer);
+	ds_release_bts(conf.tracer);
+	if (conf.error < 0)
+		goto out;
+
+	conf.suspend = ds_suspend_bts_noirq;
+	conf.resume = ds_resume_bts_noirq;
+	conf.tracer =
+		ds_request_bts_task(current, small_buffer, SMALL_BUFFER_SIZE,
+				   NULL, (size_t)-1, BTS_KERNEL);
+	local_irq_save(irq);
+	ds_selftest_bts_cpu(&conf);
+	if (conf.error >= 0)
+		conf.error = ds_selftest_bts_bad_request_cpu(0, buffer);
+	ds_release_bts_noirq(conf.tracer);
+	local_irq_restore(irq);
+	if (conf.error < 0)
+		goto out;
+
+	conf.error = 0;
+ out:
+	put_online_cpus();
+	printk(KERN_CONT "%s.\n", (conf.error ? "failed" : "passed"));
+
+	return conf.error;
+}
+
+int ds_selftest_pebs(void)
+{
+	return 0;
+}
diff --git a/arch/x86/kernel/ds_selftest.h b/arch/x86/kernel/ds_selftest.h
new file mode 100644
index 0000000..2ba8745
--- /dev/null
+++ b/arch/x86/kernel/ds_selftest.h
@@ -0,0 +1,15 @@
+/*
+ * Debug Store support - selftest
+ *
+ *
+ * Copyright (C) 2009 Intel Corporation.
+ * Markus Metzger <markus.t.metzger@intel.com>, 2009
+ */
+
+#ifdef CONFIG_X86_DS_SELFTEST
+extern int ds_selftest_bts(void);
+extern int ds_selftest_pebs(void);
+#else
+static inline int ds_selftest_bts(void) { return 0; }
+static inline int ds_selftest_pebs(void) { return 0; }
+#endif
diff --git a/arch/x86/kernel/process.c b/arch/x86/kernel/process.c
index ca98915..fb5dfb8 100644
--- a/arch/x86/kernel/process.c
+++ b/arch/x86/kernel/process.c
@@ -14,6 +14,7 @@
 #include <asm/idle.h>
 #include <asm/uaccess.h>
 #include <asm/i387.h>
+#include <asm/ds.h>
 
 unsigned long idle_halt;
 EXPORT_SYMBOL(idle_halt);
@@ -45,6 +46,8 @@ void free_thread_xstate(struct task_struct *tsk)
 		kmem_cache_free(task_xstate_cachep, tsk->thread.xstate);
 		tsk->thread.xstate = NULL;
 	}
+
+	WARN(tsk->thread.ds_ctx, "leaking DS context\n");
 }
 
 void free_thread_info(struct thread_info *ti)
@@ -83,8 +86,6 @@ void exit_thread(void)
 		put_cpu();
 		kfree(bp);
 	}
-
-	ds_exit_thread(current);
 }
 
 void flush_thread(void)
diff --git a/arch/x86/kernel/process_32.c b/arch/x86/kernel/process_32.c
index 76f8f84..b5e4bfe 100644
--- a/arch/x86/kernel/process_32.c
+++ b/arch/x86/kernel/process_32.c
@@ -290,7 +290,8 @@ int copy_thread(unsigned long clone_flags, unsigned long sp,
 		p->thread.io_bitmap_max = 0;
 	}
 
-	ds_copy_thread(p, current);
+	clear_tsk_thread_flag(p, TIF_DS_AREA_MSR);
+	p->thread.ds_ctx = NULL;
 
 	clear_tsk_thread_flag(p, TIF_DEBUGCTLMSR);
 	p->thread.debugctlmsr = 0;
diff --git a/arch/x86/kernel/process_64.c b/arch/x86/kernel/process_64.c
index b751a41..5a1a1de 100644
--- a/arch/x86/kernel/process_64.c
+++ b/arch/x86/kernel/process_64.c
@@ -335,7 +335,8 @@ int copy_thread(unsigned long clone_flags, unsigned long sp,
 			goto out;
 	}
 
-	ds_copy_thread(p, me);
+	clear_tsk_thread_flag(p, TIF_DS_AREA_MSR);
+	p->thread.ds_ctx = NULL;
 
 	clear_tsk_thread_flag(p, TIF_DEBUGCTLMSR);
 	p->thread.debugctlmsr = 0;
diff --git a/arch/x86/kernel/ptrace.c b/arch/x86/kernel/ptrace.c
index 23b7c8f..09ecbde 100644
--- a/arch/x86/kernel/ptrace.c
+++ b/arch/x86/kernel/ptrace.c
@@ -21,6 +21,7 @@
 #include <linux/audit.h>
 #include <linux/seccomp.h>
 #include <linux/signal.h>
+#include <linux/workqueue.h>
 
 #include <asm/uaccess.h>
 #include <asm/pgtable.h>
@@ -578,17 +579,130 @@ static int ioperm_get(struct task_struct *target,
 }
 
 #ifdef CONFIG_X86_PTRACE_BTS
+/*
+ * A branch trace store context.
+ *
+ * Contexts may only be installed by ptrace_bts_config() and only for
+ * ptraced tasks.
+ *
+ * Contexts are destroyed when the tracee is detached from the tracer.
+ * The actual destruction work requires interrupts enabled, so the
+ * work is deferred and will be scheduled during __ptrace_unlink().
+ *
+ * Contexts hold an additional task_struct reference on the traced
+ * task, as well as a reference on the tracer's mm.
+ *
+ * Ptrace already holds a task_struct for the duration of ptrace operations,
+ * but since destruction is deferred, it may be executed after both
+ * tracer and tracee exited.
+ */
+struct bts_context {
+	/* The branch trace handle. */
+	struct bts_tracer	*tracer;
+
+	/* The buffer used to store the branch trace and its size. */
+	void			*buffer;
+	unsigned int		size;
+
+	/* The mm that paid for the above buffer. */
+	struct mm_struct	*mm;
+
+	/* The task this context belongs to. */
+	struct task_struct	*task;
+
+	/* The signal to send on a bts buffer overflow. */
+	unsigned int		bts_ovfl_signal;
+
+	/* The work struct to destroy a context. */
+	struct work_struct	work;
+};
+
+static int alloc_bts_buffer(struct bts_context *context, unsigned int size)
+{
+	void *buffer = NULL;
+	int err = -ENOMEM;
+
+	err = account_locked_memory(current->mm, current->signal->rlim, size);
+	if (err < 0)
+		return err;
+
+	buffer = kzalloc(size, GFP_KERNEL);
+	if (!buffer)
+		goto out_refund;
+
+	context->buffer = buffer;
+	context->size = size;
+	context->mm = get_task_mm(current);
+
+	return 0;
+
+ out_refund:
+	refund_locked_memory(current->mm, size);
+	return err;
+}
+
+static inline void free_bts_buffer(struct bts_context *context)
+{
+	if (!context->buffer)
+		return;
+
+	kfree(context->buffer);
+	context->buffer = NULL;
+
+	refund_locked_memory(context->mm, context->size);
+	context->size = 0;
+
+	mmput(context->mm);
+	context->mm = NULL;
+}
+
+static void free_bts_context_work(struct work_struct *w)
+{
+	struct bts_context *context;
+
+	context = container_of(w, struct bts_context, work);
+
+	ds_release_bts(context->tracer);
+	put_task_struct(context->task);
+	free_bts_buffer(context);
+	kfree(context);
+}
+
+static inline void free_bts_context(struct bts_context *context)
+{
+	INIT_WORK(&context->work, free_bts_context_work);
+	schedule_work(&context->work);
+}
+
+static inline struct bts_context *alloc_bts_context(struct task_struct *task)
+{
+	struct bts_context *context = kzalloc(sizeof(*context), GFP_KERNEL);
+	if (context) {
+		context->task = task;
+		task->bts = context;
+
+		get_task_struct(task);
+	}
+
+	return context;
+}
+
 static int ptrace_bts_read_record(struct task_struct *child, size_t index,
 				  struct bts_struct __user *out)
 {
+	struct bts_context *context;
 	const struct bts_trace *trace;
 	struct bts_struct bts;
 	const unsigned char *at;
 	int error;
 
-	trace = ds_read_bts(child->bts);
+	context = child->bts;
+	if (!context)
+		return -ESRCH;
+
+	trace = ds_read_bts(context->tracer);
 	if (!trace)
-		return -EPERM;
+		return -ESRCH;
 
 	at = trace->ds.top - ((index + 1) * trace->ds.size);
 	if ((void *)at < trace->ds.begin)
@@ -597,7 +711,7 @@ static int ptrace_bts_read_record(struct task_struct *child, size_t index,
 	if (!trace->read)
 		return -EOPNOTSUPP;
 
-	error = trace->read(child->bts, at, &bts);
+	error = trace->read(context->tracer, at, &bts);
 	if (error < 0)
 		return error;
 
@@ -611,13 +725,18 @@ static int ptrace_bts_drain(struct task_struct *child,
 			    long size,
 			    struct bts_struct __user *out)
 {
+	struct bts_context *context;
 	const struct bts_trace *trace;
 	const unsigned char *at;
 	int error, drained = 0;
 
-	trace = ds_read_bts(child->bts);
+	context = child->bts;
+	if (!context)
+		return -ESRCH;
+
+	trace = ds_read_bts(context->tracer);
 	if (!trace)
-		return -EPERM;
+		return -ESRCH;
 
 	if (!trace->read)
 		return -EOPNOTSUPP;
@@ -628,9 +747,8 @@ static int ptrace_bts_drain(struct task_struct *child,
 	for (at = trace->ds.begin; (void *)at < trace->ds.top;
 	     out++, drained++, at += trace->ds.size) {
 		struct bts_struct bts;
-		int error;
 
-		error = trace->read(child->bts, at, &bts);
+		error = trace->read(context->tracer, at, &bts);
 		if (error < 0)
 			return error;
 
@@ -640,35 +758,18 @@ static int ptrace_bts_drain(struct task_struct *child,
 
 	memset(trace->ds.begin, 0, trace->ds.n * trace->ds.size);
 
-	error = ds_reset_bts(child->bts);
+	error = ds_reset_bts(context->tracer);
 	if (error < 0)
 		return error;
 
 	return drained;
 }
 
-static int ptrace_bts_allocate_buffer(struct task_struct *child, size_t size)
-{
-	child->bts_buffer = alloc_locked_buffer(size);
-	if (!child->bts_buffer)
-		return -ENOMEM;
-
-	child->bts_size = size;
-
-	return 0;
-}
-
-static void ptrace_bts_free_buffer(struct task_struct *child)
-{
-	free_locked_buffer(child->bts_buffer, child->bts_size);
-	child->bts_buffer = NULL;
-	child->bts_size = 0;
-}
-
 static int ptrace_bts_config(struct task_struct *child,
 			     long cfg_size,
 			     const struct ptrace_bts_config __user *ucfg)
 {
+	struct bts_context *context;
 	struct ptrace_bts_config cfg;
 	unsigned int flags = 0;
 
@@ -678,28 +779,33 @@ static int ptrace_bts_config(struct task_struct *child,
 	if (copy_from_user(&cfg, ucfg, sizeof(cfg)))
 		return -EFAULT;
 
-	if (child->bts) {
-		ds_release_bts(child->bts);
-		child->bts = NULL;
-	}
+	context = child->bts;
+	if (!context)
+		context = alloc_bts_context(child);
+	if (!context)
+		return -ENOMEM;
 
 	if (cfg.flags & PTRACE_BTS_O_SIGNAL) {
 		if (!cfg.signal)
 			return -EINVAL;
 
-		child->thread.bts_ovfl_signal = cfg.signal;
 		return -EOPNOTSUPP;
+		context->bts_ovfl_signal = cfg.signal;
 	}
 
-	if ((cfg.flags & PTRACE_BTS_O_ALLOC) &&
-	    (cfg.size != child->bts_size)) {
-		int error;
+	ds_release_bts(context->tracer);
+	context->tracer = NULL;
 
-		ptrace_bts_free_buffer(child);
+	if ((cfg.flags & PTRACE_BTS_O_ALLOC) && (cfg.size != context->size)) {
+		int err;
 
-		error = ptrace_bts_allocate_buffer(child, cfg.size);
-		if (error < 0)
-			return error;
+		free_bts_buffer(context);
+		if (!cfg.size)
+			return 0;
+
+		err = alloc_bts_buffer(context, cfg.size);
+		if (err < 0)
+			return err;
 	}
 
 	if (cfg.flags & PTRACE_BTS_O_TRACE)
@@ -708,15 +814,14 @@ static int ptrace_bts_config(struct task_struct *child,
 	if (cfg.flags & PTRACE_BTS_O_SCHED)
 		flags |= BTS_TIMESTAMPS;
 
-	child->bts = ds_request_bts(child, child->bts_buffer, child->bts_size,
-				    /* ovfl = */ NULL, /* th = */ (size_t)-1,
-				    flags);
-	if (IS_ERR(child->bts)) {
-		int error = PTR_ERR(child->bts);
-
-		ptrace_bts_free_buffer(child);
-		child->bts = NULL;
+	context->tracer =
+		ds_request_bts_task(child, context->buffer, context->size,
+				    NULL, (size_t)-1, flags);
+	if (unlikely(IS_ERR(context->tracer))) {
+		int error = PTR_ERR(context->tracer);
 
+		free_bts_buffer(context);
+		context->tracer = NULL;
 		return error;
 	}
 
@@ -727,20 +832,25 @@ static int ptrace_bts_status(struct task_struct *child,
 			     long cfg_size,
 			     struct ptrace_bts_config __user *ucfg)
 {
+	struct bts_context *context;
 	const struct bts_trace *trace;
 	struct ptrace_bts_config cfg;
 
+	context = child->bts;
+	if (!context)
+		return -ESRCH;
+
 	if (cfg_size < sizeof(cfg))
 		return -EIO;
 
-	trace = ds_read_bts(child->bts);
+	trace = ds_read_bts(context->tracer);
 	if (!trace)
-		return -EPERM;
+		return -ESRCH;
 
 	memset(&cfg, 0, sizeof(cfg));
-	cfg.size = trace->ds.end - trace->ds.begin;
-	cfg.signal = child->thread.bts_ovfl_signal;
-	cfg.bts_size = sizeof(struct bts_struct);
+	cfg.size	= trace->ds.end - trace->ds.begin;
+	cfg.signal	= context->bts_ovfl_signal;
+	cfg.bts_size	= sizeof(struct bts_struct);
 
 	if (cfg.signal)
 		cfg.flags |= PTRACE_BTS_O_SIGNAL;
@@ -759,80 +869,51 @@ static int ptrace_bts_status(struct task_struct *child,
 
 static int ptrace_bts_clear(struct task_struct *child)
 {
+	struct bts_context *context;
 	const struct bts_trace *trace;
 
-	trace = ds_read_bts(child->bts);
+	context = child->bts;
+	if (!context)
+		return -ESRCH;
+
+	trace = ds_read_bts(context->tracer);
 	if (!trace)
-		return -EPERM;
+		return -ESRCH;
 
 	memset(trace->ds.begin, 0, trace->ds.n * trace->ds.size);
 
-	return ds_reset_bts(child->bts);
+	return ds_reset_bts(context->tracer);
 }
 
 static int ptrace_bts_size(struct task_struct *child)
 {
+	struct bts_context *context;
 	const struct bts_trace *trace;
 
-	trace = ds_read_bts(child->bts);
+	context = child->bts;
+	if (!context)
+		return -ESRCH;
+
+	trace = ds_read_bts(context->tracer);
 	if (!trace)
-		return -EPERM;
+		return -ESRCH;
 
 	return (trace->ds.top - trace->ds.begin) / trace->ds.size;
 }
 
-static void ptrace_bts_fork(struct task_struct *tsk)
-{
-	tsk->bts = NULL;
-	tsk->bts_buffer = NULL;
-	tsk->bts_size = 0;
-	tsk->thread.bts_ovfl_signal = 0;
-}
-
-static void ptrace_bts_untrace(struct task_struct *child)
+/*
+ * Called from __ptrace_unlink() after the child has been moved back
+ * to its original parent.
+ */
+void ptrace_bts_untrace(struct task_struct *child)
 {
 	if (unlikely(child->bts)) {
-		ds_release_bts(child->bts);
+		free_bts_context(child->bts);
 		child->bts = NULL;
-
-		/* We cannot update total_vm and locked_vm since
-		   child's mm is already gone. But we can reclaim the
-		   memory. */
-		kfree(child->bts_buffer);
-		child->bts_buffer = NULL;
-		child->bts_size = 0;
 	}
 }
-
-static void ptrace_bts_detach(struct task_struct *child)
-{
-	/*
-	 * Ptrace_detach() races with ptrace_untrace() in case
-	 * the child dies and is reaped by another thread.
-	 *
-	 * We only do the memory accounting at this point and
-	 * leave the buffer deallocation and the bts tracer
-	 * release to ptrace_bts_untrace() which will be called
-	 * later on with tasklist_lock held.
-	 */
-	release_locked_buffer(child->bts_buffer, child->bts_size);
-}
-#else
-static inline void ptrace_bts_fork(struct task_struct *tsk) {}
-static inline void ptrace_bts_detach(struct task_struct *child) {}
-static inline void ptrace_bts_untrace(struct task_struct *child) {}
 #endif /* CONFIG_X86_PTRACE_BTS */
 
-void x86_ptrace_fork(struct task_struct *child, unsigned long clone_flags)
-{
-	ptrace_bts_fork(child);
-}
-
-void x86_ptrace_untrace(struct task_struct *child)
-{
-	ptrace_bts_untrace(child);
-}
-
 /*
  * Called by kernel/ptrace.c when detaching..
  *
@@ -844,7 +925,6 @@ void ptrace_disable(struct task_struct *child)
 #ifdef TIF_SYSCALL_EMU
 	clear_tsk_thread_flag(child, TIF_SYSCALL_EMU);
 #endif
-	ptrace_bts_detach(child);
 }
 
 #if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
diff --git a/include/linux/mm.h b/include/linux/mm.h
index bff1f0d..009eabd 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -19,6 +19,7 @@ struct anon_vma;
 struct file_ra_state;
 struct user_struct;
 struct writeback_control;
+struct rlimit;
 
 #ifndef CONFIG_DISCONTIGMEM          /* Don't use mapnrs, do it properly */
 extern unsigned long max_mapnr;
@@ -1319,8 +1320,8 @@ int vmemmap_populate_basepages(struct page *start_page,
 int vmemmap_populate(struct page *start_page, unsigned long pages, int node);
 void vmemmap_populate_print_last(void);
 
-extern void *alloc_locked_buffer(size_t size);
-extern void free_locked_buffer(void *buffer, size_t size);
-extern void release_locked_buffer(void *buffer, size_t size);
+extern int account_locked_memory(struct mm_struct *mm, struct rlimit *rlim,
+				 size_t size);
+extern void refund_locked_memory(struct mm_struct *mm, size_t size);
 #endif /* __KERNEL__ */
 #endif /* _LINUX_MM_H */
diff --git a/include/linux/ptrace.h b/include/linux/ptrace.h
index 67c1565..59e133d 100644
--- a/include/linux/ptrace.h
+++ b/include/linux/ptrace.h
@@ -95,7 +95,6 @@ extern void __ptrace_link(struct task_struct *child,
 			  struct task_struct *new_parent);
 extern void __ptrace_unlink(struct task_struct *child);
 extern void exit_ptrace(struct task_struct *tracer);
-extern void ptrace_fork(struct task_struct *task, unsigned long clone_flags);
 #define PTRACE_MODE_READ   1
 #define PTRACE_MODE_ATTACH 2
 /* Returns 0 on success, -errno on denial. */
@@ -327,15 +326,6 @@ static inline void user_enable_block_step(struct task_struct *task)
 #define arch_ptrace_untrace(task)		do { } while (0)
 #endif
 
-#ifndef arch_ptrace_fork
-/*
- * Do machine-specific work to initialize a new task.
- *
- * This is called from copy_process().
- */
-#define arch_ptrace_fork(child, clone_flags)	do { } while (0)
-#endif
-
 extern int task_current_syscall(struct task_struct *target, long *callno,
 				unsigned long args[6], unsigned int maxargs,
 				unsigned long *sp, unsigned long *pc);
diff --git a/include/linux/sched.h b/include/linux/sched.h
index 7ede5e4..1ed4ef5 100644
--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -96,8 +96,8 @@ struct exec_domain;
 struct futex_pi_state;
 struct robust_list_head;
 struct bio;
-struct bts_tracer;
 struct fs_struct;
+struct bts_context;
 
 /*
  * List of flags we want to share for kernel threads,
@@ -1209,18 +1209,11 @@ struct task_struct {
 	struct list_head ptraced;
 	struct list_head ptrace_entry;
 
-#ifdef CONFIG_X86_PTRACE_BTS
 	/*
 	 * This is the tracer handle for the ptrace BTS extension.
 	 * This field actually belongs to the ptracer task.
 	 */
-	struct bts_tracer *bts;
-	/*
-	 * The buffer to hold the BTS data.
-	 */
-	void *bts_buffer;
-	size_t bts_size;
-#endif /* CONFIG_X86_PTRACE_BTS */
+	struct bts_context *bts;
 
 	/* PID/PID hash table linkage. */
 	struct pid_link pids[PIDTYPE_MAX];
@@ -2003,8 +1996,10 @@ extern void set_task_comm(struct task_struct *tsk, char *from);
 extern char *get_task_comm(char *to, struct task_struct *tsk);
 
 #ifdef CONFIG_SMP
+extern void wait_task_context_switch(struct task_struct *p);
 extern unsigned long wait_task_inactive(struct task_struct *, long match_state);
 #else
+static inline void wait_task_context_switch(struct task_struct *p) {}
 static inline unsigned long wait_task_inactive(struct task_struct *p,
 					       long match_state)
 {
diff --git a/kernel/Makefile b/kernel/Makefile
index 4242366..a35eee3 100644
--- a/kernel/Makefile
+++ b/kernel/Makefile
@@ -93,6 +93,7 @@ obj-$(CONFIG_LATENCYTOP) += latencytop.o
 obj-$(CONFIG_HAVE_GENERIC_DMA_COHERENT) += dma-coherent.o
 obj-$(CONFIG_FUNCTION_TRACER) += trace/
 obj-$(CONFIG_TRACING) += trace/
+obj-$(CONFIG_X86_DS) += trace/
 obj-$(CONFIG_SMP) += sched_cpupri.o
 obj-$(CONFIG_SLOW_WORK) += slow-work.o
 
diff --git a/kernel/fork.c b/kernel/fork.c
index 085f73e..711468f 100644
--- a/kernel/fork.c
+++ b/kernel/fork.c
@@ -1088,8 +1088,8 @@ static struct task_struct *copy_process(unsigned long clone_flags,
 #ifdef CONFIG_DEBUG_MUTEXES
 	p->blocked_on = NULL; /* not blocked yet */
 #endif
-	if (unlikely(current->ptrace))
-		ptrace_fork(p, clone_flags);
+
+	p->bts = NULL;
 
 	/* Perform scheduler related setup. Assign this task to a CPU. */
 	sched_fork(p, clone_flags);
diff --git a/kernel/ptrace.c b/kernel/ptrace.c
index 0692ab5..e950805 100644
--- a/kernel/ptrace.c
+++ b/kernel/ptrace.c
@@ -25,16 +25,6 @@
 
 
 /*
- * Initialize a new task whose father had been ptraced.
- *
- * Called from copy_process().
- */
-void ptrace_fork(struct task_struct *child, unsigned long clone_flags)
-{
-	arch_ptrace_fork(child, clone_flags);
-}
-
-/*
  * ptrace a task: make the debugger its new parent and
  * move it to the ptrace list.
  *
diff --git a/kernel/sched.c b/kernel/sched.c
index 14a19b1..6530a27 100644
--- a/kernel/sched.c
+++ b/kernel/sched.c
@@ -2011,6 +2011,49 @@ migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
 }
 
 /*
+ * wait_task_context_switch -	wait for a thread to complete at least one
+ *				context switch.
+ *
+ * @p must not be current.
+ */
+void wait_task_context_switch(struct task_struct *p)
+{
+	unsigned long nvcsw, nivcsw, flags;
+	int running;
+	struct rq *rq;
+
+	nvcsw	= p->nvcsw;
+	nivcsw	= p->nivcsw;
+	for (;;) {
+		/*
+		 * The runqueue is assigned before the actual context
+		 * switch. We need to take the runqueue lock.
+		 *
+		 * We could check initially without the lock but it is
+		 * very likely that we need to take the lock in every
+		 * iteration.
+		 */
+		rq = task_rq_lock(p, &flags);
+		running = task_running(rq, p);
+		task_rq_unlock(rq, &flags);
+
+		if (likely(!running))
+			break;
+		/*
+		 * The switch count is incremented before the actual
+		 * context switch. We thus wait for two switches to be
+		 * sure at least one completed.
+		 */
+		if ((p->nvcsw - nvcsw) > 1)
+			break;
+		if ((p->nivcsw - nivcsw) > 1)
+			break;
+
+		cpu_relax();
+	}
+}
+
+/*
  * wait_task_inactive - wait for a thread to unschedule.
  *
  * If @match_state is nonzero, it's the @p->state value just checked and
diff --git a/kernel/trace/Makefile b/kernel/trace/Makefile
index 7c34cbf..06b8585 100644
--- a/kernel/trace/Makefile
+++ b/kernel/trace/Makefile
@@ -15,12 +15,17 @@ ifdef CONFIG_TRACING_BRANCHES
 KBUILD_CFLAGS += -DDISABLE_BRANCH_PROFILING
 endif
 
+#
+# Make the trace clocks available generally: it's infrastructure
+# relied on by ptrace for example:
+#
+obj-y += trace_clock.o
+
 obj-$(CONFIG_FUNCTION_TRACER) += libftrace.o
 obj-$(CONFIG_RING_BUFFER) += ring_buffer.o
 obj-$(CONFIG_RING_BUFFER_BENCHMARK) += ring_buffer_benchmark.o
 
 obj-$(CONFIG_TRACING) += trace.o
-obj-$(CONFIG_TRACING) += trace_clock.o
 obj-$(CONFIG_TRACING) += trace_output.o
 obj-$(CONFIG_TRACING) += trace_stat.o
 obj-$(CONFIG_TRACING) += trace_printk.o
diff --git a/kernel/trace/trace.h b/kernel/trace/trace.h
index ba25793..6e735d4 100644
--- a/kernel/trace/trace.h
+++ b/kernel/trace/trace.h
@@ -538,6 +538,8 @@ extern int trace_selftest_startup_sysprof(struct tracer *trace,
 					       struct trace_array *tr);
 extern int trace_selftest_startup_branch(struct tracer *trace,
 					 struct trace_array *tr);
+extern int trace_selftest_startup_hw_branches(struct tracer *trace,
+					      struct trace_array *tr);
 #endif /* CONFIG_FTRACE_STARTUP_TEST */
 
 extern void *head_page(struct trace_array_cpu *data);
diff --git a/kernel/trace/trace_hw_branches.c b/kernel/trace/trace_hw_branches.c
index 8683d50..ca7d7c4 100644
--- a/kernel/trace/trace_hw_branches.c
+++ b/kernel/trace/trace_hw_branches.c
@@ -1,10 +1,9 @@
 /*
- * h/w branch tracer for x86 based on bts
+ * h/w branch tracer for x86 based on BTS
  *
  * Copyright (C) 2008-2009 Intel Corporation.
  * Markus Metzger <markus.t.metzger@gmail.com>, 2008-2009
  */
-#include <linux/spinlock.h>
 #include <linux/kallsyms.h>
 #include <linux/debugfs.h>
 #include <linux/ftrace.h>
@@ -15,110 +14,119 @@
 
 #include <asm/ds.h>
 
-#include "trace.h"
 #include "trace_output.h"
+#include "trace.h"
 
 
-#define SIZEOF_BTS (1 << 13)
+#define BTS_BUFFER_SIZE (1 << 13)
 
-/*
- * The tracer lock protects the below per-cpu tracer array.
- * It needs to be held to:
- * - start tracing on all cpus
- * - stop tracing on all cpus
- * - start tracing on a single hotplug cpu
- * - stop tracing on a single hotplug cpu
- * - read the trace from all cpus
- * - read the trace from a single cpu
- */
-static DEFINE_SPINLOCK(bts_tracer_lock);
 static DEFINE_PER_CPU(struct bts_tracer *, tracer);
-static DEFINE_PER_CPU(unsigned char[SIZEOF_BTS], buffer);
+static DEFINE_PER_CPU(unsigned char[BTS_BUFFER_SIZE], buffer);
 
 #define this_tracer per_cpu(tracer, smp_processor_id())
-#define this_buffer per_cpu(buffer, smp_processor_id())
 
-static int __read_mostly trace_hw_branches_enabled;
+static int trace_hw_branches_enabled __read_mostly;
+static int trace_hw_branches_suspended __read_mostly;
 static struct trace_array *hw_branch_trace __read_mostly;
 
 
-/*
- * Start tracing on the current cpu.
- * The argument is ignored.
- *
- * pre: bts_tracer_lock must be locked.
- */
-static void bts_trace_start_cpu(void *arg)
+static void bts_trace_init_cpu(int cpu)
 {
-	if (this_tracer)
-		ds_release_bts(this_tracer);
-
-	this_tracer =
-		ds_request_bts(/* task = */ NULL, this_buffer, SIZEOF_BTS,
-			       /* ovfl = */ NULL, /* th = */ (size_t)-1,
-			       BTS_KERNEL);
-	if (IS_ERR(this_tracer)) {
-		this_tracer = NULL;
-		return;
-	}
+	per_cpu(tracer, cpu) =
+		ds_request_bts_cpu(cpu, per_cpu(buffer, cpu), BTS_BUFFER_SIZE,
+				   NULL, (size_t)-1, BTS_KERNEL);
+
+	if (IS_ERR(per_cpu(tracer, cpu)))
+		per_cpu(tracer, cpu) = NULL;
 }
 
-static void bts_trace_start(struct trace_array *tr)
+static int bts_trace_init(struct trace_array *tr)
 {
-	spin_lock(&bts_tracer_lock);
+	int cpu;
+
+	hw_branch_trace = tr;
+	trace_hw_branches_enabled = 0;
 
-	on_each_cpu(bts_trace_start_cpu, NULL, 1);
-	trace_hw_branches_enabled = 1;
+	get_online_cpus();
+	for_each_online_cpu(cpu) {
+		bts_trace_init_cpu(cpu);
 
-	spin_unlock(&bts_tracer_lock);
+		if (likely(per_cpu(tracer, cpu)))
+			trace_hw_branches_enabled = 1;
+	}
+	trace_hw_branches_suspended = 0;
+	put_online_cpus();
+
+	/* If we could not enable tracing on a single cpu, we fail. */
+	return trace_hw_branches_enabled ? 0 : -EOPNOTSUPP;
 }
 
-/*
- * Stop tracing on the current cpu.
- * The argument is ignored.
- *
- * pre: bts_tracer_lock must be locked.
- */
-static void bts_trace_stop_cpu(void *arg)
+static void bts_trace_reset(struct trace_array *tr)
 {
-	if (this_tracer) {
-		ds_release_bts(this_tracer);
-		this_tracer = NULL;
+	int cpu;
+
+	get_online_cpus();
+	for_each_online_cpu(cpu) {
+		if (likely(per_cpu(tracer, cpu))) {
+			ds_release_bts(per_cpu(tracer, cpu));
+			per_cpu(tracer, cpu) = NULL;
+		}
 	}
+	trace_hw_branches_enabled = 0;
+	trace_hw_branches_suspended = 0;
+	put_online_cpus();
 }
 
-static void bts_trace_stop(struct trace_array *tr)
+static void bts_trace_start(struct trace_array *tr)
 {
-	spin_lock(&bts_tracer_lock);
+	int cpu;
 
-	trace_hw_branches_enabled = 0;
-	on_each_cpu(bts_trace_stop_cpu, NULL, 1);
+	get_online_cpus();
+	for_each_online_cpu(cpu)
+		if (likely(per_cpu(tracer, cpu)))
+			ds_resume_bts(per_cpu(tracer, cpu));
+	trace_hw_branches_suspended = 0;
+	put_online_cpus();
+}
 
-	spin_unlock(&bts_tracer_lock);
+static void bts_trace_stop(struct trace_array *tr)
+{
+	int cpu;
+
+	get_online_cpus();
+	for_each_online_cpu(cpu)
+		if (likely(per_cpu(tracer, cpu)))
+			ds_suspend_bts(per_cpu(tracer, cpu));
+	trace_hw_branches_suspended = 1;
+	put_online_cpus();
 }
 
 static int __cpuinit bts_hotcpu_handler(struct notifier_block *nfb,
 				     unsigned long action, void *hcpu)
 {
-	unsigned int cpu = (unsigned long)hcpu;
-
-	spin_lock(&bts_tracer_lock);
-
-	if (!trace_hw_branches_enabled)
-		goto out;
+	int cpu = (long)hcpu;
 
 	switch (action) {
 	case CPU_ONLINE:
 	case CPU_DOWN_FAILED:
-		smp_call_function_single(cpu, bts_trace_start_cpu, NULL, 1);
+		/* The notification is sent with interrupts enabled. */
+		if (trace_hw_branches_enabled) {
+			bts_trace_init_cpu(cpu);
+
+			if (trace_hw_branches_suspended &&
+			    likely(per_cpu(tracer, cpu)))
+				ds_suspend_bts(per_cpu(tracer, cpu));
+		}
 		break;
+
 	case CPU_DOWN_PREPARE:
-		smp_call_function_single(cpu, bts_trace_stop_cpu, NULL, 1);
-		break;
+		/* The notification is sent with interrupts enabled. */
+		if (likely(per_cpu(tracer, cpu))) {
+			ds_release_bts(per_cpu(tracer, cpu));
+			per_cpu(tracer, cpu) = NULL;
+		}
 	}
 
- out:
-	spin_unlock(&bts_tracer_lock);
 	return NOTIFY_DONE;
 }
 
@@ -126,20 +134,6 @@ static struct notifier_block bts_hotcpu_notifier __cpuinitdata = {
 	.notifier_call = bts_hotcpu_handler
 };
 
-static int bts_trace_init(struct trace_array *tr)
-{
-	hw_branch_trace = tr;
-
-	bts_trace_start(tr);
-
-	return 0;
-}
-
-static void bts_trace_reset(struct trace_array *tr)
-{
-	bts_trace_stop(tr);
-}
-
 static void bts_trace_print_header(struct seq_file *m)
 {
 	seq_puts(m, "# CPU#        TO  <-  FROM\n");
@@ -147,10 +141,10 @@ static void bts_trace_print_header(struct seq_file *m)
 
 static enum print_line_t bts_trace_print_line(struct trace_iterator *iter)
 {
+	unsigned long symflags = TRACE_ITER_SYM_OFFSET;
 	struct trace_entry *entry = iter->ent;
 	struct trace_seq *seq = &iter->seq;
 	struct hw_branch_entry *it;
-	unsigned long symflags = TRACE_ITER_SYM_OFFSET;
 
 	trace_assign_type(it, entry);
 
@@ -226,11 +220,11 @@ static void trace_bts_at(const struct bts_trace *trace, void *at)
 /*
  * Collect the trace on the current cpu and write it into the ftrace buffer.
  *
- * pre: bts_tracer_lock must be locked
+ * pre: tracing must be suspended on the current cpu
  */
 static void trace_bts_cpu(void *arg)
 {
-	struct trace_array *tr = (struct trace_array *) arg;
+	struct trace_array *tr = (struct trace_array *)arg;
 	const struct bts_trace *trace;
 	unsigned char *at;
 
@@ -243,10 +237,9 @@ static void trace_bts_cpu(void *arg)
 	if (unlikely(!this_tracer))
 		return;
 
-	ds_suspend_bts(this_tracer);
 	trace = ds_read_bts(this_tracer);
 	if (!trace)
-		goto out;
+		return;
 
 	for (at = trace->ds.top; (void *)at < trace->ds.end;
 	     at += trace->ds.size)
@@ -255,18 +248,27 @@ static void trace_bts_cpu(void *arg)
 	for (at = trace->ds.begin; (void *)at < trace->ds.top;
 	     at += trace->ds.size)
 		trace_bts_at(trace, at);
-
-out:
-	ds_resume_bts(this_tracer);
 }
 
 static void trace_bts_prepare(struct trace_iterator *iter)
 {
-	spin_lock(&bts_tracer_lock);
+	int cpu;
 
+	get_online_cpus();
+	for_each_online_cpu(cpu)
+		if (likely(per_cpu(tracer, cpu)))
+			ds_suspend_bts(per_cpu(tracer, cpu));
+	/*
+	 * We need to collect the trace on the respective cpu since ftrace
+	 * implicitly adds the record for the current cpu.
+	 * Once that is more flexible, we could collect the data from any cpu.
+	 */
 	on_each_cpu(trace_bts_cpu, iter->tr, 1);
 
-	spin_unlock(&bts_tracer_lock);
+	for_each_online_cpu(cpu)
+		if (likely(per_cpu(tracer, cpu)))
+			ds_resume_bts(per_cpu(tracer, cpu));
+	put_online_cpus();
 }
 
 static void trace_bts_close(struct trace_iterator *iter)
@@ -276,11 +278,11 @@ static void trace_bts_close(struct trace_iterator *iter)
 
 void trace_hw_branch_oops(void)
 {
-	spin_lock(&bts_tracer_lock);
-
-	trace_bts_cpu(hw_branch_trace);
-
-	spin_unlock(&bts_tracer_lock);
+	if (this_tracer) {
+		ds_suspend_bts_noirq(this_tracer);
+		trace_bts_cpu(hw_branch_trace);
+		ds_resume_bts_noirq(this_tracer);
+	}
 }
 
 struct tracer bts_tracer __read_mostly =
@@ -293,7 +295,10 @@ struct tracer bts_tracer __read_mostly =
 	.start		= bts_trace_start,
 	.stop		= bts_trace_stop,
 	.open		= trace_bts_prepare,
-	.close		= trace_bts_close
+	.close		= trace_bts_close,
+#ifdef CONFIG_FTRACE_SELFTEST
+	.selftest	= trace_selftest_startup_hw_branches,
+#endif /* CONFIG_FTRACE_SELFTEST */
 };
 
 __init static int init_bts_trace(void)
diff --git a/kernel/trace/trace_selftest.c b/kernel/trace/trace_selftest.c
index 08f4eb2..00dd648 100644
--- a/kernel/trace/trace_selftest.c
+++ b/kernel/trace/trace_selftest.c
@@ -16,6 +16,7 @@ static inline int trace_valid_entry(struct trace_entry *entry)
 	case TRACE_BRANCH:
 	case TRACE_GRAPH_ENT:
 	case TRACE_GRAPH_RET:
+	case TRACE_HW_BRANCHES:
 		return 1;
 	}
 	return 0;
@@ -188,6 +189,7 @@ int trace_selftest_startup_dynamic_tracing(struct tracer *trace,
 #else
 # define trace_selftest_startup_dynamic_tracing(trace, tr, func) ({ 0; })
 #endif /* CONFIG_DYNAMIC_FTRACE */
+
 /*
  * Simple verification test of ftrace function tracer.
  * Enable ftrace, sleep 1/10 second, and then read the trace
@@ -749,3 +751,59 @@ trace_selftest_startup_branch(struct tracer *trace, struct trace_array *tr)
 	return ret;
 }
 #endif /* CONFIG_BRANCH_TRACER */
+
+#ifdef CONFIG_HW_BRANCH_TRACER
+int
+trace_selftest_startup_hw_branches(struct tracer *trace,
+				   struct trace_array *tr)
+{
+	struct trace_iterator *iter;
+	struct tracer tracer;
+	unsigned long count;
+	int ret;
+
+	if (!trace->open) {
+		printk(KERN_CONT "missing open function...");
+		return -1;
+	}
+
+	ret = tracer_init(trace, tr);
+	if (ret) {
+		warn_failed_init_tracer(trace, ret);
+		return ret;
+	}
+
+	/*
+	 * The hw-branch tracer needs to collect the trace from the various
+	 * cpu trace buffers - before tracing is stopped.
+	 */
+	iter = kzalloc(sizeof(*iter), GFP_KERNEL);
+	if (!iter)
+		return -ENOMEM;
+
+	memcpy(&tracer, trace, sizeof(tracer));
+
+	iter->trace = &tracer;
+	iter->tr = tr;
+	iter->pos = -1;
+	mutex_init(&iter->mutex);
+
+	trace->open(iter);
+
+	mutex_destroy(&iter->mutex);
+	kfree(iter);
+
+	tracing_stop();
+
+	ret = trace_test_buffer(tr, &count);
+	trace->reset(tr);
+	tracing_start();
+
+	if (!ret && !count) {
+		printk(KERN_CONT "no entries found..");
+		ret = -1;
+	}
+
+	return ret;
+}
+#endif /* CONFIG_HW_BRANCH_TRACER */
diff --git a/mm/mlock.c b/mm/mlock.c
index cbe9e05..ac13043 100644
--- a/mm/mlock.c
+++ b/mm/mlock.c
@@ -629,52 +629,43 @@ void user_shm_unlock(size_t size, struct user_struct *user)
 	free_uid(user);
 }
 
-void *alloc_locked_buffer(size_t size)
+int account_locked_memory(struct mm_struct *mm, struct rlimit *rlim,
+			  size_t size)
 {
-	unsigned long rlim, vm, pgsz;
-	void *buffer = NULL;
+	unsigned long lim, vm, pgsz;
+	int error = -ENOMEM;
 
 	pgsz = PAGE_ALIGN(size) >> PAGE_SHIFT;
 
-	down_write(&current->mm->mmap_sem);
-
-	rlim = current->signal->rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT;
-	vm   = current->mm->total_vm + pgsz;
-	if (rlim < vm)
-		goto out;
+	down_write(&mm->mmap_sem);
 
-	rlim = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
-	vm   = current->mm->locked_vm + pgsz;
-	if (rlim < vm)
+	lim = rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT;
+	vm   = mm->total_vm + pgsz;
+	if (lim < vm)
 		goto out;
 
-	buffer = kzalloc(size, GFP_KERNEL);
-	if (!buffer)
+	lim = rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
+	vm   = mm->locked_vm + pgsz;
+	if (lim < vm)
 		goto out;
 
-	current->mm->total_vm  += pgsz;
-	current->mm->locked_vm += pgsz;
+	mm->total_vm  += pgsz;
+	mm->locked_vm += pgsz;
 
+	error = 0;
  out:
-	up_write(&current->mm->mmap_sem);
-	return buffer;
+	up_write(&mm->mmap_sem);
+	return error;
 }
 
-void release_locked_buffer(void *buffer, size_t size)
+void refund_locked_memory(struct mm_struct *mm, size_t size)
 {
 	unsigned long pgsz = PAGE_ALIGN(size) >> PAGE_SHIFT;
 
-	down_write(&current->mm->mmap_sem);
-
-	current->mm->total_vm  -= pgsz;
-	current->mm->locked_vm -= pgsz;
-
-	up_write(&current->mm->mmap_sem);
-}
+	down_write(&mm->mmap_sem);
 
-void free_locked_buffer(void *buffer, size_t size)
-{
-	release_locked_buffer(buffer, size);
+	mm->total_vm  -= pgsz;
+	mm->locked_vm -= pgsz;
 
-	kfree(buffer);
+	up_write(&mm->mmap_sem);
 }