Re: [PATCH 1/1] stp: rt: replace spin_lock with stp style lock and use STP_ALLOC_FLAGS

[David Smith <dsmith at redhat dot com>]

On 10/27/2015 09:30 PM, yzhu1 wrote:
> On 10/27/2015 05:28 AM, David Smith wrote:
>> On 10/25/2015 10:19 PM, yzhu1 wrote:
>>> Hi, David
>>>
>>> I do not have this backtrace. Because STP_ALLOC_SLEEP_FLAGS will cause
>>> kmalloc sleep,
>>> so I replaced it with STP_ALLOC_FLAGS.
>> Evidently I'm not explaining myself well, I'll try again in more detail.
>>
>> We've got 2 sets of allocation flags in systemtap:
>>
>> STP_ALLOC_FLAGS: These are the default flags and allocations using this
>> set of flags will not sleep.
>>
>> STP_ALLOC_SLEEP_FLAGS: These flags are only supposed to be used from
>> contexts where it is safe to sleep (like begin probes).
>>
>> Here's the big question - on the realtime kernel, is it ever safe to
>> sleep?
>>
>> If the answer to the previous question is "no", then your change is
>> correct.
>>
>> If the answer to the previous question is "yes", then your change isn't
>> correct. Instead, we need to look at uses of STP_ALLOC_SLEEP_FLAGS,
>> because we're using the wrong set of flags somewhere. We're using
>> STP_ALLOC_SLEEP_FLAGS in an allocation that should be using
>> STP_ALLOC_FLAGS.
>>
>> I see 6 uses of STP_ALLOC_SLEEP_FLAGS in the systemtap source. If we
>> can't get a backtrace, then I'll need you to change the 6 uses of
>> STP_ALLOC_SLEEP_FLAGS to STP_ALLOC_FLAGS one at a time to figure out
>> which use of STP_ALLOC_SLEEP_FLAGS isn't correct.
>>
>> Thanks for continuing to work this problem.
> Hi, All
> 
> OK. I will do.

I had an idea which might help. I've attached a patch that adds a call
to 'might_sleep()' to all the functions that use STP_ALLOC_SLEEP_FLAGS.
Hopefully that might help point out where the problem lies.

Please give it a try.

-- 
David Smith
dsmith@redhat.com
Red Hat
http://www.redhat.com
256.217.0141 (direct)
256.837.0057 (fax)

diff --git a/runtime/linux/stat_runtime.h b/runtime/linux/stat_runtime.h
index 5e27026..9491a1c 100644
--- a/runtime/linux/stat_runtime.h
+++ b/runtime/linux/stat_runtime.h
@@ -48,6 +48,7 @@ static Stat _stp_stat_alloc(size_t stat_data_size)
 		return NULL;
 
 	/* Called from module_init, so user context, may sleep alloc. */
+	might_sleep();
 	st = _stp_kmalloc_gfp (sizeof(struct _Stat), STP_ALLOC_SLEEP_FLAGS);
 	if (st == NULL)
 		return NULL;
diff --git a/runtime/linux/stp_tracepoint.c b/runtime/linux/stp_tracepoint.c
index 33fe27a..95c4e95 100644
--- a/runtime/linux/stp_tracepoint.c
+++ b/runtime/linux/stp_tracepoint.c
@@ -80,6 +80,7 @@ int add_probe(struct tracepoint_entry *e, void *probe, void *data)
 	}
 	if (found)
 		return -EEXIST;
+	might_sleep();
 	p = _stp_kmalloc_gfp(sizeof(struct stp_tp_probe),
 			STP_ALLOC_SLEEP_FLAGS);
 	if (!p)
@@ -153,6 +154,7 @@ struct tracepoint_entry *add_tracepoint(const char *name)
 	 * Using kmalloc here to allocate a variable length element. Could
 	 * cause some memory fragmentation if overused.
 	 */
+	might_sleep();
 	e = _stp_kmalloc_gfp(sizeof(struct tracepoint_entry) + name_len,
 			STP_ALLOC_SLEEP_FLAGS);
 	if (!e)
diff --git a/runtime/task_finder_vma.c b/runtime/task_finder_vma.c
index f0a4db9..354d424 100644
--- a/runtime/task_finder_vma.c
+++ b/runtime/task_finder_vma.c
@@ -53,6 +53,7 @@ __stp_tf_vma_new_entry(void)
 	struct __stp_tf_vma_entry *entry;
 	size_t size = sizeof (struct __stp_tf_vma_entry);
 #ifdef CONFIG_UTRACE
+	might_sleep();
 	entry = (struct __stp_tf_vma_entry *) _stp_kmalloc_gfp(size,
                                                          STP_ALLOC_SLEEP_FLAGS);
 #else
@@ -78,8 +79,11 @@ static int
 stap_initialize_vma_map(void)
 {
 	size_t size = sizeof(struct hlist_head) * __STP_TF_TABLE_SIZE;
-	struct hlist_head *map = (struct hlist_head *) _stp_kzalloc_gfp(size,
-							STP_ALLOC_SLEEP_FLAGS);
+	struct hlist_head *map;
+
+	might_sleep();
+	map = (struct hlist_head *) _stp_kzalloc_gfp(size,
+						     STP_ALLOC_SLEEP_FLAGS);
 	if (map == NULL)
 		return -ENOMEM;