X-Git-Url: http://git.onelab.eu/?a=blobdiff_plain;f=drivers%2Foprofile%2Fbuffer_sync.c;h=fd8b050596eeb207a0a343c993876c6b30da8ed8;hb=refs%2Fheads%2Fvserver;hp=e5d06ea62b4c1b7a9e0817a30857c53dc06633ee;hpb=9bf4aaab3e101692164d49b7ca357651eb691cb6;p=linux-2.6.git

diff --git a/drivers/oprofile/buffer_sync.c b/drivers/oprofile/buffer_sync.c
index e5d06ea62..fd8b05059 100644
--- a/drivers/oprofile/buffer_sync.c
+++ b/drivers/oprofile/buffer_sync.c
@@ -6,6 +6,10 @@
  *
  * @author John Levon <levon@movementarian.org>
  *
+ * Modified by Aravind Menon for Xen
+ * These modifications are:
+ * Copyright (C) 2005 Hewlett-Packard Co.
+ *
  * This is the core of the buffer management. Each
  * CPU buffer is processed and entered into the
  * global event buffer. Such processing is necessary
@@ -32,61 +36,68 @@
 #include "cpu_buffer.h"
 #include "buffer_sync.h"
  
-#define DEFAULT_EXPIRE (HZ / 4)
- 
-static void wq_sync_buffers(void *);
-static DECLARE_WORK(sync_wq, wq_sync_buffers, NULL);
- 
-static struct timer_list sync_timer;
-static void timer_ping(unsigned long data);
-static void sync_cpu_buffers(void);
+static LIST_HEAD(dying_tasks);
+static LIST_HEAD(dead_tasks);
+static cpumask_t marked_cpus = CPU_MASK_NONE;
+static DEFINE_SPINLOCK(task_mortuary);
+static void process_task_mortuary(void);
 
- 
-/* We must make sure to process every entry in the CPU buffers
- * before a task got the PF_EXITING flag, otherwise we will hold
- * references to a possibly freed task_struct. We are safe with
- * samples past the PF_EXITING point in do_exit(), because we
- * explicitly check for that in cpu_buffer.c 
+
+/* Take ownership of the task struct and place it on the
+ * list for processing. Only after two full buffer syncs
+ * does the task eventually get freed, because by then
+ * we are sure we will not reference it again.
+ * Can be invoked from softirq via RCU callback due to
+ * call_rcu() of the task struct, hence the _irqsave.
  */
-static int exit_task_notify(struct notifier_block * self, unsigned long val, void * data)
+static int task_free_notify(struct notifier_block * self, unsigned long val, void * data)
 {
-	sync_cpu_buffers();
-	return 0;
+	unsigned long flags;
+	struct task_struct * task = data;
+	spin_lock_irqsave(&task_mortuary, flags);
+	list_add(&task->tasks, &dying_tasks);
+	spin_unlock_irqrestore(&task_mortuary, flags);
+	return NOTIFY_OK;
 }
- 
-/* There are two cases of tasks modifying task->mm->mmap list we
- * must concern ourselves with. First, when a task is about to
- * exit (exit_mmap()), we should process the buffer to deal with
- * any samples in the CPU buffer, before we lose the ->mmap information
- * we need. It is vital to get this case correct, otherwise we can
- * end up trying to access a freed task_struct.
+
+
+/* The task is on its way out. A sync of the buffer means we can catch
+ * any remaining samples for this task.
  */
-static int mm_notify(struct notifier_block * self, unsigned long val, void * data)
+static int task_exit_notify(struct notifier_block * self, unsigned long val, void * data)
 {
-	sync_cpu_buffers();
-	return 0;
+	/* To avoid latency problems, we only process the current CPU,
+	 * hoping that most samples for the task are on this CPU
+	 */
+	sync_buffer(raw_smp_processor_id());
+  	return 0;
 }
 
 
-/* Second, a task may unmap (part of) an executable mmap,
- * so we want to process samples before that happens too. This is merely
- * a QOI issue not a correctness one.
+/* The task is about to try a do_munmap(). We peek at what it's going to
+ * do, and if it's an executable region, process the samples first, so
+ * we don't lose any. This does not have to be exact, it's a QoI issue
+ * only.
  */
 static int munmap_notify(struct notifier_block * self, unsigned long val, void * data)
 {
-	/* Note that we cannot sync the buffers directly, because we might end up
-	 * taking the the mmap_sem that we hold now inside of event_buffer_read()
-	 * on a page fault, whilst holding buffer_sem - deadlock.
-	 *
-	 * This would mean a threaded reader of the event buffer, but we should
-	 * prevent it anyway.
-	 *
-	 * Delaying the work in a context that doesn't hold the mmap_sem means
-	 * that we won't lose samples from other mappings that current() may
-	 * have. Note that either way, we lose any pending samples for what is
-	 * being unmapped.
-	 */
-	schedule_work(&sync_wq);
+	unsigned long addr = (unsigned long)data;
+	struct mm_struct * mm = current->mm;
+	struct vm_area_struct * mpnt;
+
+	down_read(&mm->mmap_sem);
+
+	mpnt = find_vma(mm, addr);
+	if (mpnt && mpnt->vm_file && (mpnt->vm_flags & VM_EXEC)) {
+		up_read(&mm->mmap_sem);
+		/* To avoid latency problems, we only process the current CPU,
+		 * hoping that most samples for the task are on this CPU
+		 */
+		sync_buffer(raw_smp_processor_id());
+		return 0;
+	}
+
+	up_read(&mm->mmap_sem);
 	return 0;
 }
 
@@ -100,26 +111,26 @@ static int module_load_notify(struct notifier_block * self, unsigned long val, v
 	if (val != MODULE_STATE_COMING)
 		return 0;
 
-	sync_cpu_buffers();
-	down(&buffer_sem);
+	/* FIXME: should we process all CPU buffers ? */
+	mutex_lock(&buffer_mutex);
 	add_event_entry(ESCAPE_CODE);
 	add_event_entry(MODULE_LOADED_CODE);
-	up(&buffer_sem);
+	mutex_unlock(&buffer_mutex);
 #endif
 	return 0;
 }
 
  
-static struct notifier_block exit_task_nb = {
-	.notifier_call	= exit_task_notify,
+static struct notifier_block task_free_nb = {
+	.notifier_call	= task_free_notify,
 };
 
-static struct notifier_block exec_unmap_nb = {
-	.notifier_call	= munmap_notify,
+static struct notifier_block task_exit_nb = {
+	.notifier_call	= task_exit_notify,
 };
 
-static struct notifier_block exit_mmap_nb = {
-	.notifier_call	= mm_notify,
+static struct notifier_block munmap_nb = {
+	.notifier_call	= munmap_notify,
 };
 
 static struct notifier_block module_load_nb = {
@@ -127,11 +138,12 @@ static struct notifier_block module_load_nb = {
 };
 
  
-static void end_sync_timer(void)
+static void end_sync(void)
 {
-	del_timer_sync(&sync_timer);
-	/* timer might have queued work, make sure it's completed. */
-	flush_scheduled_work();
+	end_cpu_work();
+	/* make sure we don't leak task structs */
+	process_task_mortuary();
+	process_task_mortuary();
 }
 
 
@@ -139,18 +151,15 @@ int sync_start(void)
 {
 	int err;
 
-	init_timer(&sync_timer);
-	sync_timer.function = timer_ping;
-	sync_timer.expires = jiffies + DEFAULT_EXPIRE;
-	add_timer(&sync_timer);
+	start_cpu_work();
 
-	err = profile_event_register(EXIT_TASK, &exit_task_nb);
+	err = task_handoff_register(&task_free_nb);
 	if (err)
 		goto out1;
-	err = profile_event_register(EXIT_MMAP, &exit_mmap_nb);
+	err = profile_event_register(PROFILE_TASK_EXIT, &task_exit_nb);
 	if (err)
 		goto out2;
-	err = profile_event_register(EXEC_UNMAP, &exec_unmap_nb);
+	err = profile_event_register(PROFILE_MUNMAP, &munmap_nb);
 	if (err)
 		goto out3;
 	err = register_module_notifier(&module_load_nb);
@@ -160,13 +169,13 @@ int sync_start(void)
 out:
 	return err;
 out4:
-	profile_event_unregister(EXEC_UNMAP, &exec_unmap_nb);
+	profile_event_unregister(PROFILE_MUNMAP, &munmap_nb);
 out3:
-	profile_event_unregister(EXIT_MMAP, &exit_mmap_nb);
+	profile_event_unregister(PROFILE_TASK_EXIT, &task_exit_nb);
 out2:
-	profile_event_unregister(EXIT_TASK, &exit_task_nb);
+	task_handoff_unregister(&task_free_nb);
 out1:
-	end_sync_timer();
+	end_sync();
 	goto out;
 }
 
@@ -174,10 +183,10 @@ out1:
 void sync_stop(void)
 {
 	unregister_module_notifier(&module_load_nb);
-	profile_event_unregister(EXIT_TASK, &exit_task_nb);
-	profile_event_unregister(EXIT_MMAP, &exit_mmap_nb);
-	profile_event_unregister(EXEC_UNMAP, &exec_unmap_nb);
-	end_sync_timer();
+	profile_event_unregister(PROFILE_MUNMAP, &munmap_nb);
+	profile_event_unregister(PROFILE_TASK_EXIT, &task_exit_nb);
+	task_handoff_unregister(&task_free_nb);
+	end_sync();
 }
 
  
@@ -204,7 +213,7 @@ static inline unsigned long fast_get_dcookie(struct dentry * dentry,
  */
 static unsigned long get_exec_dcookie(struct mm_struct * mm)
 {
-	unsigned long cookie = 0;
+	unsigned long cookie = NO_COOKIE;
 	struct vm_area_struct * vma;
  
 	if (!mm)
@@ -215,8 +224,8 @@ static unsigned long get_exec_dcookie(struct mm_struct * mm)
 			continue;
 		if (!(vma->vm_flags & VM_EXECUTABLE))
 			continue;
-		cookie = fast_get_dcookie(vma->vm_file->f_dentry,
-			vma->vm_file->f_vfsmnt);
+		cookie = fast_get_dcookie(vma->vm_file->f_path.dentry,
+			vma->vm_file->f_path.mnt);
 		break;
 	}
 
@@ -232,46 +241,69 @@ out:
  */
 static unsigned long lookup_dcookie(struct mm_struct * mm, unsigned long addr, off_t * offset)
 {
-	unsigned long cookie = 0;
+	unsigned long cookie = NO_COOKIE;
 	struct vm_area_struct * vma;
 
 	for (vma = find_vma(mm, addr); vma; vma = vma->vm_next) {
  
-		if (!vma->vm_file)
-			continue;
-
 		if (addr < vma->vm_start || addr >= vma->vm_end)
 			continue;
 
-		cookie = fast_get_dcookie(vma->vm_file->f_dentry,
-			vma->vm_file->f_vfsmnt);
-		*offset = (vma->vm_pgoff << PAGE_SHIFT) + addr - vma->vm_start; 
+		if (vma->vm_file) {
+			cookie = fast_get_dcookie(vma->vm_file->f_path.dentry,
+				vma->vm_file->f_path.mnt);
+			*offset = (vma->vm_pgoff << PAGE_SHIFT) + addr -
+				vma->vm_start;
+		} else {
+			/* must be an anonymous map */
+			*offset = addr;
+		}
+
 		break;
 	}
 
+	if (!vma)
+		cookie = INVALID_COOKIE;
+
 	return cookie;
 }
 
 
-static unsigned long last_cookie = ~0UL;
+static unsigned long last_cookie = INVALID_COOKIE;
  
 static void add_cpu_switch(int i)
 {
 	add_event_entry(ESCAPE_CODE);
 	add_event_entry(CPU_SWITCH_CODE);
 	add_event_entry(i);
-	last_cookie = ~0UL;
+	last_cookie = INVALID_COOKIE;
 }
 
-static void add_kernel_ctx_switch(unsigned int in_kernel)
+static void add_cpu_mode_switch(unsigned int cpu_mode)
 {
 	add_event_entry(ESCAPE_CODE);
-	if (in_kernel)
-		add_event_entry(KERNEL_ENTER_SWITCH_CODE); 
-	else
-		add_event_entry(KERNEL_EXIT_SWITCH_CODE); 
+	switch (cpu_mode) {
+	case CPU_MODE_USER:
+		add_event_entry(USER_ENTER_SWITCH_CODE);
+		break;
+	case CPU_MODE_KERNEL:
+		add_event_entry(KERNEL_ENTER_SWITCH_CODE);
+		break;
+	case CPU_MODE_XEN:
+		add_event_entry(XEN_ENTER_SWITCH_CODE);
+	  	break;
+	default:
+		break;
+	}
 }
- 
+
+static void add_domain_switch(unsigned long domain_id)
+{
+	add_event_entry(ESCAPE_CODE);
+	add_event_entry(DOMAIN_SWITCH_CODE);
+	add_event_entry(domain_id);
+}
+
 static void
 add_user_ctx_switch(struct task_struct const * task, unsigned long cookie)
 {
@@ -294,6 +326,13 @@ static void add_cookie_switch(unsigned long cookie)
 }
 
  
+static void add_trace_begin(void)
+{
+	add_event_entry(ESCAPE_CODE);
+	add_event_entry(TRACE_BEGIN_CODE);
+}
+
+
 static void add_sample_entry(unsigned long offset, unsigned long event)
 {
 	add_event_entry(offset);
@@ -301,16 +340,16 @@ static void add_sample_entry(unsigned long offset, unsigned long event)
 }
 
 
-static void add_us_sample(struct mm_struct * mm, struct op_sample * s)
+static int add_us_sample(struct mm_struct * mm, struct op_sample * s)
 {
 	unsigned long cookie;
 	off_t offset;
  
  	cookie = lookup_dcookie(mm, s->eip, &offset);
  
-	if (!cookie) {
+	if (cookie == INVALID_COOKIE) {
 		atomic_inc(&oprofile_stats.sample_lost_no_mapping);
-		return;
+		return 0;
 	}
 
 	if (cookie != last_cookie) {
@@ -319,6 +358,8 @@ static void add_us_sample(struct mm_struct * mm, struct op_sample * s)
 	}
 
 	add_sample_entry(offset, s->event);
+
+	return 1;
 }
 
  
@@ -326,54 +367,42 @@ static void add_us_sample(struct mm_struct * mm, struct op_sample * s)
  * sample is converted into a persistent dentry/offset pair
  * for later lookup from userspace.
  */
-static void add_sample(struct mm_struct * mm, struct op_sample * s, int in_kernel)
+static int
+add_sample(struct mm_struct * mm, struct op_sample * s, int cpu_mode)
 {
-	if (in_kernel) {
+	if (cpu_mode >= CPU_MODE_KERNEL) {
 		add_sample_entry(s->eip, s->event);
+		return 1;
 	} else if (mm) {
-		add_us_sample(mm, s);
+		return add_us_sample(mm, s);
 	} else {
 		atomic_inc(&oprofile_stats.sample_lost_no_mm);
 	}
+	return 0;
 }
  
- 
+
 static void release_mm(struct mm_struct * mm)
 {
-	if (mm)
-		up_read(&mm->mmap_sem);
+	if (!mm)
+		return;
+	up_read(&mm->mmap_sem);
+	mmput(mm);
 }
 
 
-/* Take the task's mmap_sem to protect ourselves from
- * races when we do lookup_dcookie().
- */
 static struct mm_struct * take_tasks_mm(struct task_struct * task)
 {
-	struct mm_struct * mm;
-       
-	/* Subtle. We don't need to keep a reference to this task's mm,
-	 * because, for the mm to be freed on another CPU, that would have
-	 * to go through the task exit notifier, which ends up sleeping
-	 * on the buffer_sem we hold, so we end up with mutual exclusion
-	 * anyway.
-	 */
-	task_lock(task);
-	mm = task->mm;
-	task_unlock(task);
- 
-	if (mm) {
-		/* needed to walk the task's VMAs */
+	struct mm_struct * mm = get_task_mm(task);
+	if (mm)
 		down_read(&mm->mmap_sem);
-	}
- 
 	return mm;
 }
- 
- 
-static inline int is_ctx_switch(unsigned long val)
+
+
+static inline int is_code(unsigned long val)
 {
-	return val == ~0UL;
+	return val == ESCAPE_CODE;
 }
  
 
@@ -410,39 +439,112 @@ static void increment_tail(struct oprofile_cpu_buffer * b)
 
 	rmb();
 
-	if (new_tail < (b->buffer_size))
+	if (new_tail < b->buffer_size)
 		b->tail_pos = new_tail;
 	else
 		b->tail_pos = 0;
 }
 
 
+/* Move tasks along towards death. Any tasks on dead_tasks
+ * will definitely have no remaining references in any
+ * CPU buffers at this point, because we use two lists,
+ * and to have reached the list, it must have gone through
+ * one full sync already.
+ */
+static void process_task_mortuary(void)
+{
+	unsigned long flags;
+	LIST_HEAD(local_dead_tasks);
+	struct task_struct * task;
+	struct task_struct * ttask;
+
+	spin_lock_irqsave(&task_mortuary, flags);
+
+	list_splice_init(&dead_tasks, &local_dead_tasks);
+	list_splice_init(&dying_tasks, &dead_tasks);
+
+	spin_unlock_irqrestore(&task_mortuary, flags);
+
+	list_for_each_entry_safe(task, ttask, &local_dead_tasks, tasks) {
+		list_del(&task->tasks);
+		free_task(task);
+	}
+}
+
+
+static void mark_done(int cpu)
+{
+	int i;
+
+	cpu_set(cpu, marked_cpus);
+
+	for_each_online_cpu(i) {
+		if (!cpu_isset(i, marked_cpus))
+			return;
+	}
+
+	/* All CPUs have been processed at least once,
+	 * we can process the mortuary once
+	 */
+	process_task_mortuary();
+
+	cpus_clear(marked_cpus);
+}
+
+
+/* FIXME: this is not sufficient if we implement syscall barrier backtrace
+ * traversal, the code switch to sb_sample_start at first kernel enter/exit
+ * switch so we need a fifth state and some special handling in sync_buffer()
+ */
+typedef enum {
+	sb_bt_ignore = -2,
+	sb_buffer_start,
+	sb_bt_start,
+	sb_sample_start,
+} sync_buffer_state;
+
 /* Sync one of the CPU's buffers into the global event buffer.
  * Here we need to go through each batch of samples punctuated
  * by context switch notes, taking the task's mmap_sem and doing
  * lookup in task->mm->mmap to convert EIP into dcookie/offset
  * value.
  */
-static void sync_buffer(struct oprofile_cpu_buffer * cpu_buf)
+void sync_buffer(int cpu)
 {
+	struct oprofile_cpu_buffer * cpu_buf = &cpu_buffer[cpu];
 	struct mm_struct *mm = NULL;
 	struct task_struct * new;
 	unsigned long cookie = 0;
-	int in_kernel = 1;
+	int cpu_mode = 1;
 	unsigned int i;
+	sync_buffer_state state = sb_buffer_start;
+	unsigned long available;
+	int domain_switch = 0;
+
+	mutex_lock(&buffer_mutex);
  
+	add_cpu_switch(cpu);
+
 	/* Remember, only we can modify tail_pos */
 
-	unsigned long const available = get_slots(cpu_buf);
-  
-	for (i=0; i < available; ++i) {
+	available = get_slots(cpu_buf);
+
+	for (i = 0; i < available; ++i) {
 		struct op_sample * s = &cpu_buf->buffer[cpu_buf->tail_pos];
  
-		if (is_ctx_switch(s->eip)) {
-			if (s->event <= 1) {
-				/* kernel/userspace switch */
-				in_kernel = s->event;
-				add_kernel_ctx_switch(s->event);
+		if (is_code(s->eip) && !domain_switch) {
+			if (s->event <= CPU_MODE_XEN) {
+				/* xen/kernel/userspace switch */
+				cpu_mode = s->event;
+				if (state == sb_buffer_start)
+					state = sb_sample_start;
+				add_cpu_mode_switch(s->event);
+			} else if (s->event == CPU_TRACE_BEGIN) {
+				state = sb_bt_start;
+				add_trace_begin();
+			} else if (s->event == CPU_DOMAIN_SWITCH) {
+					domain_switch = 1;				
 			} else {
 				struct mm_struct * oldmm = mm;
 
@@ -456,56 +558,25 @@ static void sync_buffer(struct oprofile_cpu_buffer * cpu_buf)
 				add_user_ctx_switch(new, cookie);
 			}
 		} else {
-			add_sample(mm, s, in_kernel);
+			if (domain_switch) {
+				add_domain_switch(s->eip);
+				domain_switch = 0;
+			} else {
+				if (state >= sb_bt_start &&
+				    !add_sample(mm, s, cpu_mode)) {
+					if (state == sb_bt_start) {
+						state = sb_bt_ignore;
+						atomic_inc(&oprofile_stats.bt_lost_no_mapping);
+					}
+				}
+			}
 		}
 
 		increment_tail(cpu_buf);
 	}
 	release_mm(mm);
-}
- 
- 
-/* Process each CPU's local buffer into the global
- * event buffer.
- */
-static void sync_cpu_buffers(void)
-{
-	int i;
-
-	down(&buffer_sem);
- 
-	for (i = 0; i < NR_CPUS; ++i) {
-		struct oprofile_cpu_buffer * cpu_buf;
- 
-		if (!cpu_possible(i))
-			continue;
- 
-		cpu_buf = &cpu_buffer[i];
- 
-		add_cpu_switch(i);
-		sync_buffer(cpu_buf);
-	}
 
-	up(&buffer_sem);
- 
-	mod_timer(&sync_timer, jiffies + DEFAULT_EXPIRE);
-}
- 
+	mark_done(cpu);
 
-static void wq_sync_buffers(void * data)
-{
-	sync_cpu_buffers();
-}
- 
- 
-/* It is possible that we could have no munmap() or
- * other events for a period of time. This will lead
- * the CPU buffers to overflow and lose samples and
- * context switches. We try to reduce the problem
- * by timing out when nothing happens for a while.
- */
-static void timer_ping(unsigned long data)
-{
-	schedule_work(&sync_wq);
-	/* timer is re-added by the scheduled task */
+	mutex_unlock(&buffer_mutex);
 }