1 files changed, 663 insertions, 258 deletions

diff --git a/arch/x86/kernel/ds.c b/arch/x86/kernel/ds.c
index 87b67e3a765a..48bfe1386038 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);