Subject: Perfmon2: Fix for the occassional system hang when using hardware-sampling

From: Kevin Corry <kevcorry@us.ibm.com>
When using hardware-sampling, when the kernel sampling-buffer fills up, we
have to mask monitoring and let user-space know that the buffer is full so it
can read it into user-space memory. The overall approach is for user-space
to poll on the Perfmon2 context's file-descriptor. Then when the sampling-
buffer fills up, we post a "buffer-full" message in the Perfmon2 context and
wake-up any processes polling on the context.

However, the hardware-sampling handler can be called from a few different
code paths. In addition to being called from the PMU interrupt handler, it
is called during a process context-switch when the monitored process is
switched off the CPU. If the sampling-buffer is full during this code path,
then we will, as usual, try to notify user-space that the buffer is full.
Part of the wake_up() call is to lock the runqueue where the polling process
is waiting. However, since we are the middle of a process context-switch, the
scheduler may have that runqueue locked already. Therefore we hit a spinlock
lockup, which quickly hangs the system.

To fix this, we simply won't make the call to wake up the polling user-space
process, so we won't ever recurse on the runqueue spinlock. Unfortunately,
this means that user-space cannot use the poll() mechanism, and must loop
while reading from the context and sleeping.

Signed-off-by: Kevin Corry <kevcorry@us.ibm.com>
Acked-by: Carl Love <carll@us.ibm.com>
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>

Index: linux-2.6/arch/powerpc/perfmon/perfmon_cell_hw_smpl.c
===================================================================
--- linux-2.6.orig/arch/powerpc/perfmon/perfmon_cell_hw_smpl.c
+++ linux-2.6/arch/powerpc/perfmon/perfmon_cell_hw_smpl.c
@@ -95,17 +95,17 @@ static int pfm_cell_hw_smpl_init(struct 
 }
 
 /**
- * pfm_cell_hw_smpl_notify_user
+ * pfm_cell_hw_smpl_add_msg
  *
- * Add a "buffer full" message to the context and wake up any user-space
- * process that is polling on the context's file descriptor. That process
- * can then read() from the file-descriptor to get a copy of the message.
+ * Add a "buffer full" message to the context. A user-space process can
+ * then read() from the file-descriptor to get a copy of the message.
  **/
-static int pfm_cell_hw_smpl_notify_user(struct pfm_context *ctx)
+static int pfm_cell_hw_smpl_add_msg(struct pfm_context *ctx)
 {
 	union pfarg_msg *msg;
 
 	if (ctx->flags.no_msg) {
+		/* User-space won't be reading any messages. */
 		return 0;
 	}
 
@@ -120,7 +120,7 @@ static int pfm_cell_hw_smpl_notify_user(
 
 	msg->type = PFM_MSG_CELL_HW_SMPL_BUF_FULL;
 
-	return pfm_notify_user(ctx);
+	return 0;
 }
 
 /**
@@ -250,11 +250,8 @@ static int handle_full_buffer(struct pfm
 	 */
 	cbe_write_pm(cpu, pm_interval, set->pmcs[CELL_PMC_PM_INTERVAL]);
 
-	/* Add a message to the context's message queue and wake up any
-	 * user-space program's that are polling on the context's file
-	 * descriptor.
-	 */
-	pfm_cell_hw_smpl_notify_user(ctx);
+	/* Add a message to the context's message queue. */
+	pfm_cell_hw_smpl_add_msg(ctx);
 
 	/* Mask monitoring until a pfm_restart() occurs. */
 	pfm_mask_monitoring(ctx, set);
Index: linux-2.6/include/linux/perfmon.h
===================================================================
--- linux-2.6.orig/include/linux/perfmon.h
+++ linux-2.6/include/linux/perfmon.h
@@ -584,7 +584,6 @@ void pfm_switch_sets(struct pfm_context 
 
 union pfarg_msg *pfm_get_new_msg(struct pfm_context *ctx);
 void pfm_save_pmds(struct pfm_context *ctx, struct pfm_event_set *set);
-int pfm_notify_user(struct pfm_context *ctx);
 int pfm_ovfl_notify_user(struct pfm_context *ctx,
 			struct pfm_event_set *set,
 	     		unsigned long ip);
Index: linux-2.6/perfmon/perfmon.c
===================================================================
--- linux-2.6.orig/perfmon/perfmon.c
+++ linux-2.6/perfmon/perfmon.c
@@ -477,7 +477,7 @@ static void pfm_flush_pmds(struct task_s
 				set->pmds[i].value += 1 + ovfl_mask;
 				num_ovfls--;
 				PFM_DBG("pmd%u overflowed", i);
- 			}
+			}
 			PFM_DBG("pmd%u set=%u val=0x%llx",
 				i,
 				set->id,
@@ -640,7 +640,7 @@ do_zombie:
 		pfm_release_session(0, 0);
 }
 
-int pfm_notify_user(struct pfm_context *ctx)
+static int pfm_notify_user(struct pfm_context *ctx)
 {
 	if (ctx->state == PFM_CTX_ZOMBIE) {
 		PFM_DBG("ignoring overflow notification, owner is zombie");
@@ -662,7 +662,6 @@ int pfm_notify_user(struct pfm_context *
 
 	return 0;
 }
-EXPORT_SYMBOL(pfm_notify_user);
 
 /*
  * send a counter overflow notification message to