]> git.sesse.net Git - bcachefs-tools-debian/blobdiff - linux/six.c
Update bcachefs sources to 31c09369cd six locks: Fix an unitialized var
[bcachefs-tools-debian] / linux / six.c
index 3d366a843eb55b243c0f0c07b9245fcc76c40ae6..a47cd6d008c501fb4c021150671fde3e93b0fccb 100644 (file)
@@ -14,9 +14,9 @@
 #include <trace/events/lock.h>
 
 #ifdef DEBUG
-#define EBUG_ON(cond)          BUG_ON(cond)
+#define EBUG_ON(cond)                  BUG_ON(cond)
 #else
-#define EBUG_ON(cond)          do {} while (0)
+#define EBUG_ON(cond)                  do {} while (0)
 #endif
 
 #define six_acquire(l, t, r, ip)       lock_acquire(l, 0, t, r, 1, NULL, ip)
 
 static void do_six_unlock_type(struct six_lock *lock, enum six_lock_type type);
 
+#define SIX_LOCK_HELD_read_OFFSET      0
+#define SIX_LOCK_HELD_read             ~(~0U << 26)
+#define SIX_LOCK_HELD_intent           (1U << 26)
+#define SIX_LOCK_HELD_write            (1U << 27)
+#define SIX_LOCK_WAITING_read          (1U << (28 + SIX_LOCK_read))
+#define SIX_LOCK_WAITING_intent                (1U << (28 + SIX_LOCK_intent))
+#define SIX_LOCK_WAITING_write         (1U << (28 + SIX_LOCK_write))
+#define SIX_LOCK_NOSPIN                        (1U << 31)
+
 struct six_lock_vals {
        /* Value we add to the lock in order to take the lock: */
-       u64                     lock_val;
+       u32                     lock_val;
 
        /* If the lock has this value (used as a mask), taking the lock fails: */
-       u64                     lock_fail;
-
-       /* Value we add to the lock in order to release the lock: */
-       u64                     unlock_val;
+       u32                     lock_fail;
 
        /* Mask that indicates lock is held for this type: */
-       u64                     held_mask;
+       u32                     held_mask;
 
        /* Waitlist we wakeup when releasing the lock: */
        enum six_lock_type      unlock_wakeup;
 };
 
-#define __SIX_LOCK_HELD_read   __SIX_VAL(read_lock, ~0)
-#define __SIX_LOCK_HELD_intent __SIX_VAL(intent_lock, ~0)
-#define __SIX_LOCK_HELD_write  __SIX_VAL(seq, 1)
-
 #define LOCK_VALS {                                                    \
        [SIX_LOCK_read] = {                                             \
-               .lock_val       = __SIX_VAL(read_lock, 1),              \
-               .lock_fail      = __SIX_LOCK_HELD_write + __SIX_VAL(write_locking, 1),\
-               .unlock_val     = -__SIX_VAL(read_lock, 1),             \
-               .held_mask      = __SIX_LOCK_HELD_read,                 \
+               .lock_val       = 1U << SIX_LOCK_HELD_read_OFFSET,      \
+               .lock_fail      = SIX_LOCK_HELD_write,                  \
+               .held_mask      = SIX_LOCK_HELD_read,                   \
                .unlock_wakeup  = SIX_LOCK_write,                       \
        },                                                              \
        [SIX_LOCK_intent] = {                                           \
-               .lock_val       = __SIX_VAL(intent_lock, 1),            \
-               .lock_fail      = __SIX_LOCK_HELD_intent,               \
-               .unlock_val     = -__SIX_VAL(intent_lock, 1),           \
-               .held_mask      = __SIX_LOCK_HELD_intent,               \
+               .lock_val       = SIX_LOCK_HELD_intent,                 \
+               .lock_fail      = SIX_LOCK_HELD_intent,                 \
+               .held_mask      = SIX_LOCK_HELD_intent,                 \
                .unlock_wakeup  = SIX_LOCK_intent,                      \
        },                                                              \
        [SIX_LOCK_write] = {                                            \
-               .lock_val       = __SIX_VAL(seq, 1),                    \
-               .lock_fail      = __SIX_LOCK_HELD_read,                 \
-               .unlock_val     = __SIX_VAL(seq, 1),                    \
-               .held_mask      = __SIX_LOCK_HELD_write,                \
+               .lock_val       = SIX_LOCK_HELD_write,                  \
+               .lock_fail      = SIX_LOCK_HELD_read,                   \
+               .held_mask      = SIX_LOCK_HELD_write,                  \
                .unlock_wakeup  = SIX_LOCK_read,                        \
        },                                                              \
 }
 
+static inline void six_set_bitmask(struct six_lock *lock, u32 mask)
+{
+       if ((atomic_read(&lock->state) & mask) != mask)
+               atomic_or(mask, &lock->state);
+}
+
+static inline void six_clear_bitmask(struct six_lock *lock, u32 mask)
+{
+       if (atomic_read(&lock->state) & mask)
+               atomic_and(~mask, &lock->state);
+}
+
 static inline void six_set_owner(struct six_lock *lock, enum six_lock_type type,
-                                union six_lock_state old,
-                                struct task_struct *owner)
+                                u32 old, struct task_struct *owner)
 {
        if (type != SIX_LOCK_intent)
                return;
 
-       if (!old.intent_lock) {
+       if (!(old & SIX_LOCK_HELD_intent)) {
                EBUG_ON(lock->owner);
                lock->owner = owner;
        } else {
@@ -94,22 +104,25 @@ static inline unsigned pcpu_read_count(struct six_lock *lock)
        return read_count;
 }
 
-/* This is probably up there with the more evil things I've done */
-#define waitlist_bitnr(id) ilog2((((union six_lock_state) { .waiters = 1 << (id) }).l))
-
-static int __do_six_trylock_type(struct six_lock *lock,
-                                enum six_lock_type type,
-                                struct task_struct *task,
-                                bool try)
+/*
+ * __do_six_trylock() - main trylock routine
+ *
+ * Returns 1 on success, 0 on failure
+ *
+ * In percpu reader mode, a failed trylock may cause a spurious trylock failure
+ * for anoter thread taking the competing lock type, and we may havve to do a
+ * wakeup: when a wakeup is required, we return -1 - wakeup_type.
+ */
+static int __do_six_trylock(struct six_lock *lock, enum six_lock_type type,
+                           struct task_struct *task, bool try)
 {
        const struct six_lock_vals l[] = LOCK_VALS;
-       union six_lock_state old, new;
        int ret;
-       u64 v;
+       u32 old, new, v;
 
        EBUG_ON(type == SIX_LOCK_write && lock->owner != task);
-       EBUG_ON(type == SIX_LOCK_write && (lock->state.seq & 1));
-       EBUG_ON(type == SIX_LOCK_write && (try != !(lock->state.write_locking)));
+       EBUG_ON(type == SIX_LOCK_write &&
+               (try != !(atomic_read(&lock->state) & SIX_LOCK_HELD_write)));
 
        /*
         * Percpu reader mode:
@@ -124,101 +137,75 @@ static int __do_six_trylock_type(struct six_lock *lock,
         * the lock, then issues a full memory barrier, then reads from the
         * other thread's variable to check if the other thread thinks it has
         * the lock. If we raced, we backoff and retry/sleep.
+        *
+        * Failure to take the lock may cause a spurious trylock failure in
+        * another thread, because we temporarily set the lock to indicate that
+        * we held it. This would be a problem for a thread in six_lock(), when
+        * they are calling trylock after adding themself to the waitlist and
+        * prior to sleeping.
+        *
+        * Therefore, if we fail to get the lock, and there were waiters of the
+        * type we conflict with, we will have to issue a wakeup.
+        *
+        * Since we may be called under wait_lock (and by the wakeup code
+        * itself), we return that the wakeup has to be done instead of doing it
+        * here.
         */
-
        if (type == SIX_LOCK_read && lock->readers) {
                preempt_disable();
                this_cpu_inc(*lock->readers); /* signal that we own lock */
 
                smp_mb();
 
-               old.v = READ_ONCE(lock->state.v);
-               ret = !(old.v & l[type].lock_fail);
+               old = atomic_read(&lock->state);
+               ret = !(old & l[type].lock_fail);
 
                this_cpu_sub(*lock->readers, !ret);
                preempt_enable();
 
-               /*
-                * If we failed because a writer was trying to take the
-                * lock, issue a wakeup because we might have caused a
-                * spurious trylock failure:
-                */
-#if 0
-               /*
-                * This code should be sufficient, but we're seeing unexplained
-                * lost wakeups:
-                */
-               if (old.write_locking)
+               if (!ret && (old & SIX_LOCK_WAITING_write))
                        ret = -1 - SIX_LOCK_write;
-#else
-               if (!ret)
-                       ret = -1 - SIX_LOCK_write;
-#endif
        } else if (type == SIX_LOCK_write && lock->readers) {
                if (try) {
-                       atomic64_add(__SIX_VAL(write_locking, 1),
-                                    &lock->state.counter);
-                       smp_mb__after_atomic();
-               } else if (!(lock->state.waiters & (1 << SIX_LOCK_write))) {
-                       atomic64_add(__SIX_VAL(waiters, 1 << SIX_LOCK_write),
-                                    &lock->state.counter);
-                       /*
-                        * pairs with barrier after unlock and before checking
-                        * for readers in unlock path
-                        */
+                       atomic_add(SIX_LOCK_HELD_write, &lock->state);
                        smp_mb__after_atomic();
                }
 
                ret = !pcpu_read_count(lock);
 
-               /*
-                * On success, we increment lock->seq; also we clear
-                * write_locking unless we failed from the lock path:
-                */
-               v = 0;
-               if (ret)
-                       v += __SIX_VAL(seq, 1);
-               if (ret || try)
-                       v -= __SIX_VAL(write_locking, 1);
-
                if (try && !ret) {
-                       old.v = atomic64_add_return(v, &lock->state.counter);
-                       if (old.waiters & (1 << SIX_LOCK_read))
+                       old = atomic_sub_return(SIX_LOCK_HELD_write, &lock->state);
+                       if (old & SIX_LOCK_WAITING_read)
                                ret = -1 - SIX_LOCK_read;
-               } else {
-                       atomic64_add(v, &lock->state.counter);
                }
        } else {
-               v = READ_ONCE(lock->state.v);
+               v = atomic_read(&lock->state);
                do {
-                       new.v = old.v = v;
+                       new = old = v;
 
-                       if (!(old.v & l[type].lock_fail)) {
-                               new.v += l[type].lock_val;
+                       ret = !(old & l[type].lock_fail);
 
-                               if (type == SIX_LOCK_write)
-                                       new.write_locking = 0;
-                       } else if (!try && !(new.waiters & (1 << type)))
-                               new.waiters |= 1 << type;
-                       else
-                               break; /* waiting bit already set */
-               } while ((v = atomic64_cmpxchg_acquire(&lock->state.counter,
-                                       old.v, new.v)) != old.v);
+                       if (!ret || (type == SIX_LOCK_write && !try)) {
+                               smp_mb();
+                               break;
+                       }
 
-               ret = !(old.v & l[type].lock_fail);
+                       new += l[type].lock_val;
+               } while ((v = atomic_cmpxchg_acquire(&lock->state, old, new)) != old);
 
-               EBUG_ON(ret && !(lock->state.v & l[type].held_mask));
+               EBUG_ON(ret && !(atomic_read(&lock->state) & l[type].held_mask));
        }
 
        if (ret > 0)
                six_set_owner(lock, type, old, task);
 
-       EBUG_ON(type == SIX_LOCK_write && (try || ret > 0) && (lock->state.write_locking));
+       EBUG_ON(type == SIX_LOCK_write && try && ret <= 0 &&
+               (atomic_read(&lock->state) & SIX_LOCK_HELD_write));
 
        return ret;
 }
 
-static inline void __six_lock_wakeup(struct six_lock *lock, enum six_lock_type lock_type)
+static void __six_lock_wakeup(struct six_lock *lock, enum six_lock_type lock_type)
 {
        struct six_lock_waiter *w, *next;
        struct task_struct *task;
@@ -237,7 +224,7 @@ again:
                        goto unlock;
                saw_one = true;
 
-               ret = __do_six_trylock_type(lock, lock_type, w->task, false);
+               ret = __do_six_trylock(lock, lock_type, w->task, false);
                if (ret <= 0)
                        goto unlock;
 
@@ -252,7 +239,7 @@ again:
                wake_up_process(task);
        }
 
-       clear_bit(waitlist_bitnr(lock_type), (unsigned long *) &lock->state.v);
+       six_clear_bitmask(lock, SIX_LOCK_WAITING_read << lock_type);
 unlock:
        raw_spin_unlock(&lock->wait_lock);
 
@@ -262,96 +249,74 @@ unlock:
        }
 }
 
-static inline void six_lock_wakeup(struct six_lock *lock,
-                                  union six_lock_state state,
-                                  enum six_lock_type lock_type)
+__always_inline
+static void six_lock_wakeup(struct six_lock *lock, u32 state,
+                           enum six_lock_type lock_type)
 {
-       if (lock_type == SIX_LOCK_write && state.read_lock)
+       if (lock_type == SIX_LOCK_write && (state & SIX_LOCK_HELD_read))
                return;
 
-       if (!(state.waiters & (1 << lock_type)))
+       if (!(state & (SIX_LOCK_WAITING_read << lock_type)))
                return;
 
        __six_lock_wakeup(lock, lock_type);
 }
 
-static bool do_six_trylock_type(struct six_lock *lock,
-                               enum six_lock_type type,
-                               bool try)
+__always_inline
+static bool do_six_trylock(struct six_lock *lock, enum six_lock_type type, bool try)
 {
        int ret;
 
-       ret = __do_six_trylock_type(lock, type, current, try);
+       ret = __do_six_trylock(lock, type, current, try);
        if (ret < 0)
                __six_lock_wakeup(lock, -ret - 1);
 
        return ret > 0;
 }
 
-__always_inline __flatten
-static bool __six_trylock_type(struct six_lock *lock, enum six_lock_type type,
-                              unsigned long ip)
+/**
+ * six_trylock_ip - attempt to take a six lock without blocking
+ * @lock:      lock to take
+ * @type:      SIX_LOCK_read, SIX_LOCK_intent, or SIX_LOCK_write
+ * @ip:                ip parameter for lockdep/lockstat, i.e. _THIS_IP_
+ *
+ * Return: true on success, false on failure.
+ */
+bool six_trylock_ip(struct six_lock *lock, enum six_lock_type type, unsigned long ip)
 {
-       if (!do_six_trylock_type(lock, type, true))
+       if (!do_six_trylock(lock, type, true))
                return false;
 
        if (type != SIX_LOCK_write)
                six_acquire(&lock->dep_map, 1, type == SIX_LOCK_read, ip);
        return true;
 }
-
-__always_inline __flatten
-static bool __six_relock_type(struct six_lock *lock, enum six_lock_type type,
-                             unsigned seq, unsigned long ip)
+EXPORT_SYMBOL_GPL(six_trylock_ip);
+
+/**
+ * six_relock_ip - attempt to re-take a lock that was held previously
+ * @lock:      lock to take
+ * @type:      SIX_LOCK_read, SIX_LOCK_intent, or SIX_LOCK_write
+ * @seq:       lock sequence number obtained from six_lock_seq() while lock was
+ *             held previously
+ * @ip:                ip parameter for lockdep/lockstat, i.e. _THIS_IP_
+ *
+ * Return: true on success, false on failure.
+ */
+bool six_relock_ip(struct six_lock *lock, enum six_lock_type type,
+                  unsigned seq, unsigned long ip)
 {
-       const struct six_lock_vals l[] = LOCK_VALS;
-       union six_lock_state old;
-       u64 v;
-
-       EBUG_ON(type == SIX_LOCK_write);
-
-       if (type == SIX_LOCK_read &&
-           lock->readers) {
-               bool ret;
-
-               preempt_disable();
-               this_cpu_inc(*lock->readers);
-
-               smp_mb();
-
-               old.v = READ_ONCE(lock->state.v);
-               ret = !(old.v & l[type].lock_fail) && old.seq == seq;
-
-               this_cpu_sub(*lock->readers, !ret);
-               preempt_enable();
-
-               /*
-                * Similar to the lock path, we may have caused a spurious write
-                * lock fail and need to issue a wakeup:
-                */
-               if (ret)
-                       six_acquire(&lock->dep_map, 1, type == SIX_LOCK_read, ip);
-               else
-                       six_lock_wakeup(lock, old, SIX_LOCK_write);
+       if (lock->seq != seq || !six_trylock_ip(lock, type, ip))
+               return false;
 
-               return ret;
+       if (lock->seq != seq) {
+               six_unlock_ip(lock, type, ip);
+               return false;
        }
 
-       v = READ_ONCE(lock->state.v);
-       do {
-               old.v = v;
-
-               if (old.seq != seq || old.v & l[type].lock_fail)
-                       return false;
-       } while ((v = atomic64_cmpxchg_acquire(&lock->state.counter,
-                               old.v,
-                               old.v + l[type].lock_val)) != old.v);
-
-       six_set_owner(lock, type, old, current);
-       if (type != SIX_LOCK_write)
-               six_acquire(&lock->dep_map, 1, type == SIX_LOCK_read, ip);
        return true;
 }
+EXPORT_SYMBOL_GPL(six_relock_ip);
 
 #ifdef CONFIG_LOCK_SPIN_ON_OWNER
 
@@ -371,17 +336,6 @@ static inline bool six_can_spin_on_owner(struct six_lock *lock)
        return ret;
 }
 
-static inline void six_set_nospin(struct six_lock *lock)
-{
-       union six_lock_state old, new;
-       u64 v = READ_ONCE(lock->state.v);
-
-       do {
-               new.v = old.v = v;
-               new.nospin = true;
-       } while ((v = atomic64_cmpxchg(&lock->state.counter, old.v, new.v)) != old.v);
-}
-
 static inline bool six_spin_on_owner(struct six_lock *lock,
                                     struct task_struct *owner,
                                     u64 end_time)
@@ -405,7 +359,7 @@ static inline bool six_spin_on_owner(struct six_lock *lock,
                }
 
                if (!(++loop & 0xf) && (time_after64(sched_clock(), end_time))) {
-                       six_set_nospin(lock);
+                       six_set_bitmask(lock, SIX_LOCK_NOSPIN);
                        ret = false;
                        break;
                }
@@ -445,7 +399,7 @@ static inline bool six_optimistic_spin(struct six_lock *lock, enum six_lock_type
                if (owner && !six_spin_on_owner(lock, owner, end_time))
                        break;
 
-               if (do_six_trylock_type(lock, type, false)) {
+               if (do_six_trylock(lock, type, false)) {
                        osq_unlock(&lock->osq);
                        preempt_enable();
                        return true;
@@ -494,17 +448,16 @@ static inline bool six_optimistic_spin(struct six_lock *lock, enum six_lock_type
 #endif
 
 noinline
-static int __six_lock_type_slowpath(struct six_lock *lock, enum six_lock_type type,
-                                   struct six_lock_waiter *wait,
-                                   six_lock_should_sleep_fn should_sleep_fn, void *p,
-                                   unsigned long ip)
+static int six_lock_slowpath(struct six_lock *lock, enum six_lock_type type,
+                            struct six_lock_waiter *wait,
+                            six_lock_should_sleep_fn should_sleep_fn, void *p,
+                            unsigned long ip)
 {
-       union six_lock_state old;
        int ret = 0;
 
        if (type == SIX_LOCK_write) {
-               EBUG_ON(lock->state.write_locking);
-               atomic64_add(__SIX_VAL(write_locking, 1), &lock->state.counter);
+               EBUG_ON(atomic_read(&lock->state) & SIX_LOCK_HELD_write);
+               atomic_add(SIX_LOCK_HELD_write, &lock->state);
                smp_mb__after_atomic();
        }
 
@@ -519,13 +472,12 @@ static int __six_lock_type_slowpath(struct six_lock *lock, enum six_lock_type ty
        wait->lock_acquired     = false;
 
        raw_spin_lock(&lock->wait_lock);
-       if (!(lock->state.waiters & (1 << type)))
-               set_bit(waitlist_bitnr(type), (unsigned long *) &lock->state.v);
+       six_set_bitmask(lock, SIX_LOCK_WAITING_read << type);
        /*
-        * Retry taking the lock after taking waitlist lock, have raced with an
-        * unlock:
+        * Retry taking the lock after taking waitlist lock, in case we raced
+        * with an unlock:
         */
-       ret = __do_six_trylock_type(lock, type, current, false);
+       ret = __do_six_trylock(lock, type, current, false);
        if (ret <= 0) {
                wait->start_time = local_clock();
 
@@ -565,7 +517,7 @@ static int __six_lock_type_slowpath(struct six_lock *lock, enum six_lock_type ty
                                list_del(&wait->list);
                        raw_spin_unlock(&lock->wait_lock);
 
-                       if (wait->lock_acquired)
+                       if (unlikely(wait->lock_acquired))
                                do_six_unlock_type(lock, type);
                        break;
                }
@@ -575,21 +527,49 @@ static int __six_lock_type_slowpath(struct six_lock *lock, enum six_lock_type ty
 
        __set_current_state(TASK_RUNNING);
 out:
-       if (ret && type == SIX_LOCK_write && lock->state.write_locking) {
-               old.v = atomic64_sub_return(__SIX_VAL(write_locking, 1),
-                                           &lock->state.counter);
-               six_lock_wakeup(lock, old, SIX_LOCK_read);
+       if (ret && type == SIX_LOCK_write) {
+               six_clear_bitmask(lock, SIX_LOCK_HELD_write);
+               six_lock_wakeup(lock, atomic_read(&lock->state), SIX_LOCK_read);
        }
        trace_contention_end(lock, 0);
 
        return ret;
 }
 
-__always_inline __flatten
-static int __six_lock_type_waiter(struct six_lock *lock, enum six_lock_type type,
-                        struct six_lock_waiter *wait,
-                        six_lock_should_sleep_fn should_sleep_fn, void *p,
-                        unsigned long ip)
+/**
+ * six_lock_ip_waiter - take a lock, with full waitlist interface
+ * @lock:      lock to take
+ * @type:      SIX_LOCK_read, SIX_LOCK_intent, or SIX_LOCK_write
+ * @wait:      pointer to wait object, which will be added to lock's waitlist
+ * @should_sleep_fn: callback run after adding to waitlist, immediately prior
+ *             to scheduling
+ * @p:         passed through to @should_sleep_fn
+ * @ip:                ip parameter for lockdep/lockstat, i.e. _THIS_IP_
+ *
+ * This is the most general six_lock() variant, with parameters to support full
+ * cycle detection for deadlock avoidance.
+ *
+ * The code calling this function must implement tracking of held locks, and the
+ * @wait object should be embedded into the struct that tracks held locks -
+ * which must also be accessible in a thread-safe way.
+ *
+ * @should_sleep_fn should invoke the cycle detector; it should walk each
+ * lock's waiters, and for each waiter recursively walk their held locks.
+ *
+ * When this function must block, @wait will be added to @lock's waitlist before
+ * calling trylock, and before calling @should_sleep_fn, and @wait will not be
+ * removed from the lock waitlist until the lock has been successfully acquired,
+ * or we abort.
+ *
+ * @wait.start_time will be monotonically increasing for any given waitlist, and
+ * thus may be used as a loop cursor.
+ *
+ * Return: 0 on success, or the return code from @should_sleep_fn on failure.
+ */
+int six_lock_ip_waiter(struct six_lock *lock, enum six_lock_type type,
+                      struct six_lock_waiter *wait,
+                      six_lock_should_sleep_fn should_sleep_fn, void *p,
+                      unsigned long ip)
 {
        int ret;
 
@@ -598,8 +578,8 @@ static int __six_lock_type_waiter(struct six_lock *lock, enum six_lock_type type
        if (type != SIX_LOCK_write)
                six_acquire(&lock->dep_map, 0, type == SIX_LOCK_read, ip);
 
-       ret = do_six_trylock_type(lock, type, true) ? 0
-               : __six_lock_type_slowpath(lock, type, wait, should_sleep_fn, p, ip);
+       ret = do_six_trylock(lock, type, true) ? 0
+               : six_lock_slowpath(lock, type, wait, should_sleep_fn, p, ip);
 
        if (ret && type != SIX_LOCK_write)
                six_release(&lock->dep_map, ip);
@@ -608,22 +588,13 @@ static int __six_lock_type_waiter(struct six_lock *lock, enum six_lock_type type
 
        return ret;
 }
+EXPORT_SYMBOL_GPL(six_lock_ip_waiter);
 
 __always_inline
-static int __six_lock_type(struct six_lock *lock, enum six_lock_type type,
-                          six_lock_should_sleep_fn should_sleep_fn, void *p,
-                          unsigned long ip)
-{
-       struct six_lock_waiter wait;
-
-       return __six_lock_type_waiter(lock, type, &wait, should_sleep_fn, p, ip);
-}
-
-__always_inline __flatten
 static void do_six_unlock_type(struct six_lock *lock, enum six_lock_type type)
 {
        const struct six_lock_vals l[] = LOCK_VALS;
-       union six_lock_state state;
+       u32 state;
 
        if (type == SIX_LOCK_intent)
                lock->owner = NULL;
@@ -633,26 +604,39 @@ static void do_six_unlock_type(struct six_lock *lock, enum six_lock_type type)
                smp_mb(); /* unlock barrier */
                this_cpu_dec(*lock->readers);
                smp_mb(); /* between unlocking and checking for waiters */
-               state.v = READ_ONCE(lock->state.v);
+               state = atomic_read(&lock->state);
        } else {
-               u64 v = l[type].unlock_val;
+               u32 v = l[type].lock_val;
 
                if (type != SIX_LOCK_read)
-                       v -= lock->state.v & __SIX_VAL(nospin, 1);
+                       v += atomic_read(&lock->state) & SIX_LOCK_NOSPIN;
 
-               EBUG_ON(!(lock->state.v & l[type].held_mask));
-               state.v = atomic64_add_return_release(v, &lock->state.counter);
+               EBUG_ON(!(atomic_read(&lock->state) & l[type].held_mask));
+               state = atomic_sub_return_release(v, &lock->state);
        }
 
        six_lock_wakeup(lock, state, l[type].unlock_wakeup);
 }
 
-__always_inline __flatten
-static void __six_unlock_type(struct six_lock *lock, enum six_lock_type type,
-                             unsigned long ip)
+/**
+ * six_unlock_ip - drop a six lock
+ * @lock:      lock to unlock
+ * @type:      SIX_LOCK_read, SIX_LOCK_intent, or SIX_LOCK_write
+ * @ip:                ip parameter for lockdep/lockstat, i.e. _THIS_IP_
+ *
+ * When a lock is held multiple times (because six_lock_incement()) was used),
+ * this decrements the 'lock held' counter by one.
+ *
+ * For example:
+ * six_lock_read(&foo->lock);                          read count 1
+ * six_lock_increment(&foo->lock, SIX_LOCK_read);      read count 2
+ * six_lock_unlock(&foo->lock, SIX_LOCK_read);         read count 1
+ * six_lock_unlock(&foo->lock, SIX_LOCK_read);         read count 0
+ */
+void six_unlock_ip(struct six_lock *lock, enum six_lock_type type, unsigned long ip)
 {
        EBUG_ON(type == SIX_LOCK_write &&
-               !(lock->state.v & __SIX_LOCK_HELD_intent));
+               !(atomic_read(&lock->state) & SIX_LOCK_HELD_intent));
        EBUG_ON((type == SIX_LOCK_write ||
                 type == SIX_LOCK_intent) &&
                lock->owner != current);
@@ -666,52 +650,18 @@ static void __six_unlock_type(struct six_lock *lock, enum six_lock_type type,
                return;
        }
 
+       lock->seq += type == SIX_LOCK_write;
+
        do_six_unlock_type(lock, type);
 }
+EXPORT_SYMBOL_GPL(six_unlock_ip);
 
-#define __SIX_LOCK(type)                                               \
-bool six_trylock_ip_##type(struct six_lock *lock, unsigned long ip)    \
-{                                                                      \
-       return __six_trylock_type(lock, SIX_LOCK_##type, ip);           \
-}                                                                      \
-EXPORT_SYMBOL_GPL(six_trylock_ip_##type);                              \
-                                                                       \
-bool six_relock_ip_##type(struct six_lock *lock, u32 seq, unsigned long ip)\
-{                                                                      \
-       return __six_relock_type(lock, SIX_LOCK_##type, seq, ip);       \
-}                                                                      \
-EXPORT_SYMBOL_GPL(six_relock_ip_##type);                               \
-                                                                       \
-int six_lock_ip_##type(struct six_lock *lock,                          \
-                   six_lock_should_sleep_fn should_sleep_fn, void *p,  \
-                   unsigned long ip)                                   \
-{                                                                      \
-       return __six_lock_type(lock, SIX_LOCK_##type, should_sleep_fn, p, ip);\
-}                                                                      \
-EXPORT_SYMBOL_GPL(six_lock_ip_##type);                                 \
-                                                                       \
-int six_lock_ip_waiter_##type(struct six_lock *lock,                   \
-                          struct six_lock_waiter *wait,                \
-                          six_lock_should_sleep_fn should_sleep_fn, void *p,\
-                          unsigned long ip)                            \
-{                                                                      \
-       return __six_lock_type_waiter(lock, SIX_LOCK_##type, wait, should_sleep_fn, p, ip);\
-}                                                                      \
-EXPORT_SYMBOL_GPL(six_lock_ip_waiter_##type);                          \
-                                                                       \
-void six_unlock_ip_##type(struct six_lock *lock, unsigned long ip)     \
-{                                                                      \
-       __six_unlock_type(lock, SIX_LOCK_##type, ip);                   \
-}                                                                      \
-EXPORT_SYMBOL_GPL(six_unlock_ip_##type);
-
-__SIX_LOCK(read)
-__SIX_LOCK(intent)
-__SIX_LOCK(write)
-
-#undef __SIX_LOCK
-
-/* Convert from intent to read: */
+/**
+ * six_lock_downgrade - convert an intent lock to a read lock
+ * @lock:      lock to dowgrade
+ *
+ * @lock will have read count incremented and intent count decremented
+ */
 void six_lock_downgrade(struct six_lock *lock)
 {
        six_lock_increment(lock, SIX_LOCK_read);
@@ -719,25 +669,33 @@ void six_lock_downgrade(struct six_lock *lock)
 }
 EXPORT_SYMBOL_GPL(six_lock_downgrade);
 
+/**
+ * six_lock_tryupgrade - attempt to convert read lock to an intent lock
+ * @lock:      lock to upgrade
+ *
+ * On success, @lock will have intent count incremented and read count
+ * decremented
+ *
+ * Return: true on success, false on failure
+ */
 bool six_lock_tryupgrade(struct six_lock *lock)
 {
-       union six_lock_state old, new;
-       u64 v = READ_ONCE(lock->state.v);
+       const struct six_lock_vals l[] = LOCK_VALS;
+       u32 old, new, v = atomic_read(&lock->state);
 
        do {
-               new.v = old.v = v;
+               new = old = v;
 
-               if (new.intent_lock)
+               if (new & SIX_LOCK_HELD_intent)
                        return false;
 
                if (!lock->readers) {
-                       EBUG_ON(!new.read_lock);
-                       new.read_lock--;
+                       EBUG_ON(!(new & SIX_LOCK_HELD_read));
+                       new -= l[SIX_LOCK_read].lock_val;
                }
 
-               new.intent_lock = 1;
-       } while ((v = atomic64_cmpxchg_acquire(&lock->state.counter,
-                               old.v, new.v)) != old.v);
+               new |= SIX_LOCK_HELD_intent;
+       } while ((v = atomic_cmpxchg_acquire(&lock->state, old, new)) != old);
 
        if (lock->readers)
                this_cpu_dec(*lock->readers);
@@ -748,6 +706,17 @@ bool six_lock_tryupgrade(struct six_lock *lock)
 }
 EXPORT_SYMBOL_GPL(six_lock_tryupgrade);
 
+/**
+ * six_trylock_convert - attempt to convert a held lock from one type to another
+ * @lock:      lock to upgrade
+ * @from:      SIX_LOCK_read or SIX_LOCK_intent
+ * @to:                SIX_LOCK_read or SIX_LOCK_intent
+ *
+ * On success, @lock will have intent count incremented and read count
+ * decremented
+ *
+ * Return: true on success, false on failure
+ */
 bool six_trylock_convert(struct six_lock *lock,
                         enum six_lock_type from,
                         enum six_lock_type to)
@@ -766,9 +735,16 @@ bool six_trylock_convert(struct six_lock *lock,
 }
 EXPORT_SYMBOL_GPL(six_trylock_convert);
 
-/*
- * Increment read/intent lock count, assuming we already have it read or intent
- * locked:
+/**
+ * six_lock_increment - increase held lock count on a lock that is already held
+ * @lock:      lock to increment
+ * @type:      SIX_LOCK_read or SIX_LOCK_intent
+ *
+ * @lock must already be held, with a lock type that is greater than or equal to
+ * @type
+ *
+ * A corresponding six_unlock_type() call will be required for @lock to be fully
+ * unlocked.
  */
 void six_lock_increment(struct six_lock *lock, enum six_lock_type type)
 {
@@ -783,13 +759,14 @@ void six_lock_increment(struct six_lock *lock, enum six_lock_type type)
                if (lock->readers) {
                        this_cpu_inc(*lock->readers);
                } else {
-                       EBUG_ON(!lock->state.read_lock &&
-                               !lock->state.intent_lock);
-                       atomic64_add(l[type].lock_val, &lock->state.counter);
+                       EBUG_ON(!(atomic_read(&lock->state) &
+                                 (SIX_LOCK_HELD_read|
+                                  SIX_LOCK_HELD_intent)));
+                       atomic_add(l[type].lock_val, &lock->state);
                }
                break;
        case SIX_LOCK_intent:
-               EBUG_ON(!lock->state.intent_lock);
+               EBUG_ON(!(atomic_read(&lock->state) & SIX_LOCK_HELD_intent));
                lock->intent_lock_recurse++;
                break;
        case SIX_LOCK_write:
@@ -799,9 +776,19 @@ void six_lock_increment(struct six_lock *lock, enum six_lock_type type)
 }
 EXPORT_SYMBOL_GPL(six_lock_increment);
 
+/**
+ * six_lock_wakeup_all - wake up all waiters on @lock
+ * @lock:      lock to wake up waiters for
+ *
+ * Wakeing up waiters will cause them to re-run should_sleep_fn, which may then
+ * abort the lock operation.
+ *
+ * This function is never needed in a bug-free program; it's only useful in
+ * debug code, e.g. to determine if a cycle detector is at fault.
+ */
 void six_lock_wakeup_all(struct six_lock *lock)
 {
-       union six_lock_state state = lock->state;
+       u32 state = atomic_read(&lock->state);
        struct six_lock_waiter *w;
 
        six_lock_wakeup(lock, state, SIX_LOCK_read);
@@ -815,38 +802,96 @@ void six_lock_wakeup_all(struct six_lock *lock)
 }
 EXPORT_SYMBOL_GPL(six_lock_wakeup_all);
 
-void six_lock_pcpu_free(struct six_lock *lock)
-{
-       BUG_ON(lock->readers && pcpu_read_count(lock));
-       BUG_ON(lock->state.read_lock);
-
-       free_percpu(lock->readers);
-       lock->readers = NULL;
-}
-EXPORT_SYMBOL_GPL(six_lock_pcpu_free);
-
-void six_lock_pcpu_alloc(struct six_lock *lock)
-{
-#ifdef __KERNEL__
-       if (!lock->readers)
-               lock->readers = alloc_percpu(unsigned);
-#endif
-}
-EXPORT_SYMBOL_GPL(six_lock_pcpu_alloc);
-
-/*
- * Returns lock held counts, for both read and intent
+/**
+ * six_lock_counts - return held lock counts, for each lock type
+ * @lock:      lock to return counters for
+ *
+ * Return: the number of times a lock is held for read, intent and write.
  */
 struct six_lock_count six_lock_counts(struct six_lock *lock)
 {
        struct six_lock_count ret;
 
        ret.n[SIX_LOCK_read]    = !lock->readers
-               ? lock->state.read_lock
+               ? atomic_read(&lock->state) & SIX_LOCK_HELD_read
                : pcpu_read_count(lock);
-       ret.n[SIX_LOCK_intent]  = lock->state.intent_lock + lock->intent_lock_recurse;
-       ret.n[SIX_LOCK_write]   = lock->state.seq & 1;
+       ret.n[SIX_LOCK_intent]  = !!(atomic_read(&lock->state) & SIX_LOCK_HELD_intent) +
+               lock->intent_lock_recurse;
+       ret.n[SIX_LOCK_write]   = !!(atomic_read(&lock->state) & SIX_LOCK_HELD_write);
 
        return ret;
 }
 EXPORT_SYMBOL_GPL(six_lock_counts);
+
+/**
+ * six_lock_readers_add - directly manipulate reader count of a lock
+ * @lock:      lock to add/subtract readers for
+ * @nr:                reader count to add/subtract
+ *
+ * When an upper layer is implementing lock reentrency, we may have both read
+ * and intent locks on the same lock.
+ *
+ * When we need to take a write lock, the read locks will cause self-deadlock,
+ * because six locks themselves do not track which read locks are held by the
+ * current thread and which are held by a different thread - it does no
+ * per-thread tracking of held locks.
+ *
+ * The upper layer that is tracking held locks may however, if trylock() has
+ * failed, count up its own read locks, subtract them, take the write lock, and
+ * then re-add them.
+ *
+ * As in any other situation when taking a write lock, @lock must be held for
+ * intent one (or more) times, so @lock will never be left unlocked.
+ */
+void six_lock_readers_add(struct six_lock *lock, int nr)
+{
+       if (lock->readers) {
+               this_cpu_add(*lock->readers, nr);
+       } else {
+               EBUG_ON((int) (atomic_read(&lock->state) & SIX_LOCK_HELD_read) + nr < 0);
+               /* reader count starts at bit 0 */
+               atomic_add(nr, &lock->state);
+       }
+}
+EXPORT_SYMBOL_GPL(six_lock_readers_add);
+
+/**
+ * six_lock_exit - release resources held by a lock prior to freeing
+ * @lock:      lock to exit
+ *
+ * When a lock was initialized in percpu mode (SIX_OLCK_INIT_PCPU), this is
+ * required to free the percpu read counts.
+ */
+void six_lock_exit(struct six_lock *lock)
+{
+       WARN_ON(lock->readers && pcpu_read_count(lock));
+       WARN_ON(atomic_read(&lock->state) & SIX_LOCK_HELD_read);
+
+       free_percpu(lock->readers);
+       lock->readers = NULL;
+}
+EXPORT_SYMBOL_GPL(six_lock_exit);
+
+void __six_lock_init(struct six_lock *lock, const char *name,
+                    struct lock_class_key *key, enum six_lock_init_flags flags)
+{
+       atomic_set(&lock->state, 0);
+       raw_spin_lock_init(&lock->wait_lock);
+       INIT_LIST_HEAD(&lock->wait_list);
+#ifdef CONFIG_DEBUG_LOCK_ALLOC
+       debug_check_no_locks_freed((void *) lock, sizeof(*lock));
+       lockdep_init_map(&lock->dep_map, name, key, 0);
+#endif
+
+       if (flags & SIX_LOCK_INIT_PCPU) {
+               /*
+                * We don't return an error here on memory allocation failure
+                * since percpu is an optimization, and locks will work with the
+                * same semantics in non-percpu mode: callers can check for
+                * failure if they wish by checking lock->readers, but generally
+                * will not want to treat it as an error.
+                */
+               lock->readers = alloc_percpu(unsigned);
+       }
+}
+EXPORT_SYMBOL_GPL(__six_lock_init);