[lkml]   [2019]   [Apr]   [24]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
Patch in this message
Subject[PATCH 19/28] locking/lockdep: Optimize irq usage check when marking lock usage bit
The basic rule concerning IRQ usage related deadlock is to find whether
there is a direct or indirect dependency from an IRQ safe lock to an IRQ
unsafe lock, where an IRQ safe lock means the lock is ever used in IRQ
context (i.e., LOCK_USED_IN_*) while an IRQ unsafe lock means the lock
is ever acquired with IRQ enabled (i.e., LOCK_ENABLED_IN_*).

This rule is checked to find violation at two points:

- When marking a new IRQ usage bit
- When adding a new dependency

At each point, a check may take two searches through the dependency

- Forward search: walk forwards the dependency graph, try to find an unsafe lock.
- Backward search: walk backwards the dependency graph, try to find a safe lock.

If graph search/searches are successful, then there is a rule violation.
The current checks are effective albeit they can be significantly

Lock usage

Any lock class can be marked as in three exclusive usage states:

- IRQ safe or IRQ unsafe
- Lock used (with the LOCK_USED bit marked only), which is neither safe
nor unsafe, but can transition to either one above state

A lock can not be in both IRQ safe and unsafe states, if so it is a
simple violation of the rule by the node itself, which can be easily
found when marking IRQ usage. Therefore, this violation quest is left
out in the following.

Note that with read-write locks considered, the exclusiveness is a bit
more complex, which will be discussed in later patches.

Dependency graph

When checking the rule, the current dependency graph is composed of
nodes of the three exclusive types, and of directed edges as follows,
presumably it has no violation:

- Edges from safe node to safe node
- Edges from unsafe node to safe node
- Edges from unsafe node to unsafe node

- Edges from used node to safe node
- Edges from used node to unsafe node
- Edges from used node to used node
- Edges from safe node to used node
- Edges from unsafe node to used node

The first three types are fixed; edges will not change the type over
time. If the graph has only fixed-type edges, checking the rule would be
so easy: just check the direct dependencies involved at the two checking
points, there is even no need to search the entire graph, because a
violation could only be a direct dependency from a safe lock to an
unsafe lock.

The other five types are uncertain. If a used node transitions to other
states, the involved edge types will be altered.

New checking algorithm

First, two exclusive substates are added to the used state (they will be
discussed later):

- Reachable from IRQ safe nodes
- Unreachable from IRQ safe nodes

The new checking algorithm at the two checking points as follows:

1. When adding a new dependency from prev lock to next lock:

- If prev is in unsafe then done. There is no worry to have any
new violation from prev forwards so a forward search is not needed. A
backward search for safe node is not needed either because prev is like
a shield for next since there has been no safe node to prev.

- If the prev is in safe state:
* If next is safe, do a forward search for unsafe node from next
* If next is unsafe, one violation is found
* If next is used, do a forward search for unsafe node from next

- If prev is in used state:
* If prev is reachable from safe state, do a forward search for
unsafe states from next
* If prev is unreachable from safe state, done

A forward search for unsafe node traverses the graph as the current
forward search does. In addition, after a safe-state node or denpendency
is newly added to the graph, if a used node is encountered not marked as
reachable from safe state, then mark it reachable.

2. When marking a new usage bit for a lock

- Used: this happens only when initializing the lock's usage bit. At
this time, this node should have no any dependency so nothing needs
to be done

- Safe from used:
* if the node is reachable from safe node, done
* if the node is unreachable from safe node, do a forward search for unsafe node

- Unsafe from used:
* If the node is reachable from safe node, then a violation is found
* If the node is unreachable from safe node, done

With this new algorithm, the checking efficiency can be improved
significantly. Backward searches are avoided; some forward search cases
are saved as well. As a result, backward dependencies are not needed any

The following patches implement this new IRQ usage checking algorithm.

This starter patch defines and initializes irqsafe_lock and
irqsafe_distance fields in lock_class, which keep track of the irqsafe
locks that reach this lock class. In addition, four bitmaps are declared
to keep track of the irqsafe locks.

Signed-off-by: Yuyang Du <>
include/linux/lockdep.h | 14 ++++++++++++++
kernel/locking/lockdep.c | 14 ++++++++++++++
2 files changed, 28 insertions(+)

diff --git a/include/linux/lockdep.h b/include/linux/lockdep.h
index 7c2fefa..0e209b8 100644
--- a/include/linux/lockdep.h
+++ b/include/linux/lockdep.h
@@ -72,6 +72,11 @@ struct lock_trace {

+ * For hardirq and softirq, each has one for irqsafe lock that reaches
+ * this lock and one for irqsafe-read lock that reaches this lock.
+ */

* The lock-class itself. The order of the structure members matters.
@@ -114,6 +119,15 @@ struct lock_class {
int name_version;
const char *name;

+ /*
+ * irqsafe_lock indicates whether there is an IRQ safe lock
+ * reaches this lock_class in the dependency graph, and if so
+ * points to it; irqsafe_distance measures the distance from the
+ * IRQ safe locks, used for keeping the shortest.
+ */
+ struct lock_class *irqsafe_lock[LOCK_IRQ_SAFE_LOCKS];
+ int irqsafe_distance[LOCK_IRQ_SAFE_LOCKS];
unsigned long contention_point[LOCKSTAT_POINTS];
unsigned long contending_point[LOCKSTAT_POINTS];
diff --git a/kernel/locking/lockdep.c b/kernel/locking/lockdep.c
index 6e79bd6..a3138c9 100644
--- a/kernel/locking/lockdep.c
+++ b/kernel/locking/lockdep.c
@@ -1101,6 +1101,7 @@ static bool is_dynamic_key(const struct lock_class_key *key)
struct lockdep_subclass_key *key;
struct hlist_head *hash_head;
struct lock_class *class;
+ int i;


@@ -1153,6 +1154,9 @@ static bool is_dynamic_key(const struct lock_class_key *key)
class->name_version = count_matching_names(class);
+ for (i = 0; i < ARRAY_SIZE(class->irqsafe_distance); i++)
+ class->irqsafe_distance[i] = INT_MAX;
* We use RCU's safe list-add method to make
* parallel walking of the hash-list safe:
@@ -2896,6 +2900,10 @@ static void print_usage_bug_scenario(struct held_lock *lock)
return 1;

+static DECLARE_BITMAP(lock_classes_hardirq_safe, MAX_LOCKDEP_KEYS);
+static DECLARE_BITMAP(lock_classes_hardirq_safe_read, MAX_LOCKDEP_KEYS);
+static DECLARE_BITMAP(lock_classes_softirq_safe, MAX_LOCKDEP_KEYS);
+static DECLARE_BITMAP(lock_classes_softirq_safe_read, MAX_LOCKDEP_KEYS);

* print irq inversion bug:
@@ -5001,6 +5009,12 @@ void __init lockdep_init(void)
+ sizeof(lock_chains)
+ sizeof(lock_chains_in_use)
+ sizeof(chain_hlocks)
+ + sizeof(lock_classes_hardirq_safe)
+ + sizeof(lock_classes_hardirq_safe_read)
+ + sizeof(lock_classes_softirq_safe)
+ + sizeof(lock_classes_softirq_safe_read)
) / 1024
 \ /
  Last update: 2019-04-24 12:23    [W:0.204 / U:3.980 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site