2022-05-04 09:52:51

by Byungchul Park

[permalink] [raw]
Subject: [PATCH RFC v6 21/21] dept: Unstage wait when tagging a normal sleep wait

Staging a wait and commit have been introduced to handle conditional
sleeps that can be determined in __schedule() whether it actually goes
to sleep or not. With this feature, actual wait tagging is delayed
until __schedule().

Unfortunately, an ambiguity arises when a normal sleep wait that doesn't
require staging and commit, is involved in the middle of handling a
conditional sleep e.g. between prepare_to_wait_*() and __schedule(),
which is a very rare case tho.

So let it give up handling the conditional sleep by unstaging it
unconditionally when a normal sleep wait gets involved, to avoid the
ambiguity.

Signed-off-by: Byungchul Park <[email protected]>
---
kernel/dependency/dept.c | 55 +++++++++++++++++++++++++++++++++++-------------
1 file changed, 40 insertions(+), 15 deletions(-)

diff --git a/kernel/dependency/dept.c b/kernel/dependency/dept.c
index 14dc33b..ce6d5b3 100644
--- a/kernel/dependency/dept.c
+++ b/kernel/dependency/dept.c
@@ -2166,6 +2166,21 @@ static void __dept_wait(struct dept_map *m, unsigned long w_f,
}
}

+static inline void stage_map(struct dept_task *dt, struct dept_map *m)
+{
+ dt->stage_m = m;
+}
+
+static inline void unstage_map(struct dept_task *dt)
+{
+ dt->stage_m = NULL;
+}
+
+static inline struct dept_map *staged_map(struct dept_task *dt)
+{
+ return dt->stage_m;
+}
+
void dept_wait(struct dept_map *m, unsigned long w_f, unsigned long ip,
const char *w_fn, int ne, bool sleep)
{
@@ -2183,27 +2198,24 @@ void dept_wait(struct dept_map *m, unsigned long w_f, unsigned long ip,

flags = dept_enter();

+ /*
+ * There's no way to distinguish between a staged wait and this
+ * one, in the middle of handling a wait that requires staging
+ * and commit in __schedule().
+ *
+ * The wait that has been tagged dept_wait() with sleep == true
+ * should ignore the staged wait in __schedule() if it exists,
+ * to avoid the ambiguity. It can be done by unstaging it.
+ */
+ if (sleep)
+ unstage_map(dt);
+
__dept_wait(m, w_f, ip, w_fn, ne, sleep);

dept_exit(flags);
}
EXPORT_SYMBOL_GPL(dept_wait);

-static inline void stage_map(struct dept_task *dt, struct dept_map *m)
-{
- dt->stage_m = m;
-}
-
-static inline void unstage_map(struct dept_task *dt)
-{
- dt->stage_m = NULL;
-}
-
-static inline struct dept_map *staged_map(struct dept_task *dt)
-{
- return dt->stage_m;
-}
-
void dept_stage_wait(struct dept_map *m, unsigned long w_f,
const char *w_fn, int ne)
{
@@ -2565,6 +2577,19 @@ void dept_wait_split_map(struct dept_map_each *me,

flags = dept_enter();

+ /*
+ * There's no way to distinguish between a staged wait and this
+ * one, in the middle of handling a wait that requires staging
+ * and commit in __schedule().
+ *
+ * The wait that has been tagged dept_wait_split_map() with
+ * sleep == true should ignore the staged wait in __schedule()
+ * if it exists, to avoid the ambiguity. It can be done by
+ * unstaging it.
+ */
+ if (sleep)
+ unstage_map(dt);
+
k = mc->keys ?: &mc->keys_local;
c = check_new_class(&mc->keys_local, k, 0, 0UL, mc->name);
if (c)
--
1.9.1