Re: [PATCH v4] locking/rwsem: Make handoff bit handling more consistent

From: Waiman Long
Date: Mon Nov 15 2021 - 11:59:09 EST



On 11/15/21 10:45, Aiqun(Maria) Yu wrote:
On 11/12/2021 8:10 PM, Peter Zijlstra wrote:
On Thu, Nov 11, 2021 at 11:07:53PM -0500, Waiman Long wrote:
@@ -889,6 +892,20 @@ rwsem_spin_on_owner(struct rw_semaphore *sem)
  }
  #endif
  +/*
+ * Common code to handle rwsem flags in out_nolock path with wait_lock held.
+ * If there is more than one waiter in the queue and the HANDOFF bit is set,
+ * the next waiter will inherit it if the first waiter is removed.
+ */
+static inline void rwsem_out_nolock_clear_flags(struct rw_semaphore *sem,
+                        struct rwsem_waiter *waiter)

I'm going to rename that, it doesn't just clear flags, it dequeues the
waiter.

Argh, rwsem_mark_wake() doesn't clear HANDOFF when list_empty(), and
write_slowpath() is *far* too clever about all of this.


+{
+    list_del(&waiter->list);
    if (list_empty(&sem->wait_list)) {
+ atomic_long_andnot(RWSEM_FLAG_HANDOFF | RWSEM_FLAG_WAITERS,
+                   &sem->count);
    }
+}



@@ -1098,7 +1110,7 @@ rwsem_down_write_slowpath(struct rw_semaphore *sem, int state)
           * In this case, we attempt to acquire the lock again
           * without sleeping.
           */
-        if (wstate == WRITER_HANDOFF) {
+        if (waiter.handoff_set) {

I'm thinking this wants to be something like:

        if (rwsem_first_waiter(sem) == &waiter && waiter.handoff_set) {

              enum owner_state owner_state;
                preempt_disable();

How's this (on top) then?

---
--- a/kernel/locking/rwsem.c
+++ b/kernel/locking/rwsem.c
@@ -104,11 +104,10 @@
   * atomic_long_fetch_add() is used to obtain reader lock, whereas
   * atomic_long_cmpxchg() will be used to obtain writer lock.
   *
- * There are four places where the lock handoff bit may be set or cleared.
- * 1) rwsem_mark_wake() for readers            -- set, clear
- * 2) rwsem_try_write_lock() for writers       -- set, clear
- * 3) Error path of rwsem_down_write_slowpath() -- clear
- * 4) Error path of rwsem_down_read_slowpath()  -- clear
+ * There are three places where the lock handoff bit may be set or cleared.
+ * 1) rwsem_mark_wake() for readers        -- set, clear
+ * 2) rwsem_try_write_lock() for writers    -- set, clear
+ * 3) rwsem_del_waiter()            -- clear
   *
   * For all the above cases, wait_lock will be held. A writer must also
   * be the first one in the wait_list to be eligible for setting the handoff
@@ -363,6 +362,31 @@ enum rwsem_wake_type {
   */
  #define MAX_READERS_WAKEUP    0x100
  +static inline void
+rwsem_add_waiter(struct rw_semaphore *sem, struct rwsem_waiter *waiter)
+{
+    lockdep_assert_held(&sem->wait_lock);
+    list_add_tail(&waiter->list, &sem->wait_list);
+    /* caller will set RWSEM_FLAG_WAITERS */
    /* each time a waiter is just added in to the list,
     * handoff_set initialed as false. */
    waiter->handoff_set = false;

waiter initialization is done at entry to the slowpath function. This helper function just insert the waiter into the wait list.


+}
+
+/*
+ * Remove a waiter from the wait_list and clear flags.
+ *
+ * Both rwsem_mark_wake() and rwsem_try_write_lock() contain a full 'copy' of
+ * this function. Modify with care.
+ */
+static inline void
+rwsem_del_waiter(struct rw_semaphore *sem, struct rwsem_waiter *waiter)
+{
+    lockdep_assert_held(&sem->wait_lock);
+    list_del(&waiter->list);
what about avoid unnecessary inherit of handoff bit for waiters?

if (waiter->handoff_set)
    atomic_long_andnot(RWSEM_FLAG_HANDOFF, &sem->count);

We have decided that to let the next waiter inherit the handoff bit. So there is no need to clear it unless there is no more waiter in the queue.


+    if (likely(!list_empty(&sem->wait_list)))
+        return;
+
+    atomic_long_andnot(RWSEM_FLAG_HANDOFF | RWSEM_FLAG_WAITERS, &sem->count);
+}
+
  /*
   * handle the lock release when processes blocked on it that can now run
   * - if we come here from up_xxxx(), then the RWSEM_FLAG_WAITERS bit must
@@ -374,6 +398,8 @@ enum rwsem_wake_type {
   *   preferably when the wait_lock is released
   * - woken process blocks are discarded from the list after having task zeroed
   * - writers are only marked woken if downgrading is false
+ *
+ * Implies rwsem_del_waiter() for all woken readers.
   */
  static void rwsem_mark_wake(struct rw_semaphore *sem,
                  enum rwsem_wake_type wake_type,
@@ -488,18 +514,25 @@ static void rwsem_mark_wake(struct rw_se
        adjustment = woken * RWSEM_READER_BIAS - adjustment;
      lockevent_cond_inc(rwsem_wake_reader, woken);
+
+    oldcount = atomic_long_read(&sem->count);
      if (list_empty(&sem->wait_list)) {
-        /* hit end of list above */
+        /*
+         * Combined with list_move_tail() above, this implies
+         * rwsem_del_waiter().
+         */
          adjustment -= RWSEM_FLAG_WAITERS;
+        if (oldcount & RWSEM_FLAG_HANDOFF)
+            adjustment -= RWSEM_FLAG_HANDOFF;
+    } else if (woken) {
+        /*
+         * When we've woken a reader, we no longer need to force
+         * writers to give up the lock and we can clear HANDOFF.
+         */
+        if (oldcount & RWSEM_FLAG_HANDOFF)
+            adjustment -= RWSEM_FLAG_HANDOFF;
      }
  -    /*
-     * When we've woken a reader, we no longer need to force writers
-     * to give up the lock and we can clear HANDOFF.
-     */
-    if (woken && (atomic_long_read(&sem->count) & RWSEM_FLAG_HANDOFF))
-        adjustment -= RWSEM_FLAG_HANDOFF;
-
      if (adjustment)
          atomic_long_add(adjustment, &sem->count);
  @@ -529,6 +562,8 @@ static void rwsem_mark_wake(struct rw_se
   * This function must be called with the sem->wait_lock held to prevent
   * race conditions between checking the rwsem wait list and setting the
   * sem->count accordingly.
+ *
+ * Implies rwsem_del_waiter() on success.
   */
  static inline bool rwsem_try_write_lock(struct rw_semaphore *sem,
                      struct rwsem_waiter *waiter)
@@ -575,6 +610,11 @@ static inline bool rwsem_try_write_lock(
          return false;
      }
  +    /*
+     * Have rwsem_try_write_lock() fully imply rwsem_del_waiter() on
+     * success.
+     */
+    list_del(&waiter->list);
      rwsem_set_owner(sem);
      return true;
  }
@@ -893,20 +933,6 @@ rwsem_spin_on_owner(struct rw_semaphore
  #endif
    /*
- * Common code to handle rwsem flags in out_nolock path with wait_lock held.
- * If there is more than one waiter in the queue and the HANDOFF bit is set,
- * the next waiter will inherit it if the first waiter is removed.
- */
-static inline void rwsem_out_nolock_clear_flags(struct rw_semaphore *sem,
-                        struct rwsem_waiter *waiter)
-{
-    list_del(&waiter->list);
-    if (list_empty(&sem->wait_list))
-        atomic_long_andnot(RWSEM_FLAG_HANDOFF | RWSEM_FLAG_WAITERS,
-                   &sem->count);
-}
-
-/*
   * Wait for the read lock to be granted
   */
  static struct rw_semaphore __sched *
@@ -973,7 +999,7 @@ rwsem_down_read_slowpath(struct rw_semap
          }
          adjustment += RWSEM_FLAG_WAITERS;
      }
-    list_add_tail(&waiter.list, &sem->wait_list);
+    rwsem_add_waiter(sem, &waiter);
        /* we're now waiting on the lock, but no longer actively locking */
      count = atomic_long_add_return(adjustment, &sem->count);
@@ -1019,7 +1045,7 @@ rwsem_down_read_slowpath(struct rw_semap
      return sem;
    out_nolock:
-    rwsem_out_nolock_clear_flags(sem, &waiter);
+    rwsem_del_waiter(sem, &waiter);
      raw_spin_unlock_irq(&sem->wait_lock);
      __set_current_state(TASK_RUNNING);
      lockevent_inc(rwsem_rlock_fail);
@@ -1034,7 +1060,6 @@ rwsem_down_write_slowpath(struct rw_sema
  {
      long count;
      struct rwsem_waiter waiter;
-    struct rw_semaphore *ret = sem;
      DEFINE_WAKE_Q(wake_q);
        /* do optimistic spinning and steal lock if possible */
@@ -1053,7 +1078,7 @@ rwsem_down_write_slowpath(struct rw_sema
      waiter.handoff_set = false;
        raw_spin_lock_irq(&sem->wait_lock);
-    list_add_tail(&waiter.list, &sem->wait_list);
+    rwsem_add_waiter(sem, &waiter);
        /* we're now waiting on the lock */
      if (rwsem_first_waiter(sem) != &waiter) {
@@ -1110,7 +1135,7 @@ rwsem_down_write_slowpath(struct rw_sema
           * In this case, we attempt to acquire the lock again
           * without sleeping.
           */
-        if (waiter.handoff_set) {
+        if (rwsem_first_waiter(sem) == &waiter && waiter.handoff_set) {
              enum owner_state owner_state;
                preempt_disable();
@@ -1128,16 +1153,14 @@ rwsem_down_write_slowpath(struct rw_sema
          raw_spin_lock_irq(&sem->wait_lock);
      }
      __set_current_state(TASK_RUNNING);
-    list_del(&waiter.list);
      raw_spin_unlock_irq(&sem->wait_lock);
      lockevent_inc(rwsem_wlock);
-
-    return ret;
+    return sem;
    out_nolock:
      __set_current_state(TASK_RUNNING);
      raw_spin_lock_irq(&sem->wait_lock);
-    rwsem_out_nolock_clear_flags(sem, &waiter);
+    rwsem_del_waiter(sem, &waiter);
      if (!list_empty(&sem->wait_list))
          rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q);
      raw_spin_unlock_irq(&sem->wait_lock);


what about avoid unnecessary inherit of handoff bit for waiters?
    1. when set handoff bit also set waiter.handoff_set as  true;
    2. when clear handoff bit also set waiter.handoff_set as false.
    3. And rwsem_add_waiter initial as false;
    4. And rwsem_del_waiter also can clear the handoff bit according to waiter.handoff_set.

Because handoff bit can have better performance if correctly set and cleared.

Killing or interrupt a waiter to force it to quit is not considered a fast path operation. It is an exception rather than the rule. So performance consideration is less important here.

Cheers,
Longman