Merge tag 'xfs-4.21-merge-3' of git://git.kernel.org/pub/scm/fs/xfs/xfs-linux
[linux-2.6-block.git] / lib / sbitmap.c
index fdd1b8aa8ac63c99f5165e4795043e653b821f7a..65c2d06250a63acd453614011fa42ccd55836beb 100644 (file)
 #include <linux/sbitmap.h>
 #include <linux/seq_file.h>
 
+/*
+ * See if we have deferred clears that we can batch move
+ */
+static inline bool sbitmap_deferred_clear(struct sbitmap *sb, int index)
+{
+       unsigned long mask, val;
+       unsigned long __maybe_unused flags;
+       bool ret = false;
+
+       /* Silence bogus lockdep warning */
+#if defined(CONFIG_LOCKDEP)
+       local_irq_save(flags);
+#endif
+       spin_lock(&sb->map[index].swap_lock);
+
+       if (!sb->map[index].cleared)
+               goto out_unlock;
+
+       /*
+        * First get a stable cleared mask, setting the old mask to 0.
+        */
+       do {
+               mask = sb->map[index].cleared;
+       } while (cmpxchg(&sb->map[index].cleared, mask, 0) != mask);
+
+       /*
+        * Now clear the masked bits in our free word
+        */
+       do {
+               val = sb->map[index].word;
+       } while (cmpxchg(&sb->map[index].word, val, val & ~mask) != val);
+
+       ret = true;
+out_unlock:
+       spin_unlock(&sb->map[index].swap_lock);
+#if defined(CONFIG_LOCKDEP)
+       local_irq_restore(flags);
+#endif
+       return ret;
+}
+
 int sbitmap_init_node(struct sbitmap *sb, unsigned int depth, int shift,
                      gfp_t flags, int node)
 {
@@ -59,6 +100,7 @@ int sbitmap_init_node(struct sbitmap *sb, unsigned int depth, int shift,
        for (i = 0; i < sb->map_nr; i++) {
                sb->map[i].depth = min(depth, bits_per_word);
                depth -= sb->map[i].depth;
+               spin_lock_init(&sb->map[i].swap_lock);
        }
        return 0;
 }
@@ -69,6 +111,9 @@ void sbitmap_resize(struct sbitmap *sb, unsigned int depth)
        unsigned int bits_per_word = 1U << sb->shift;
        unsigned int i;
 
+       for (i = 0; i < sb->map_nr; i++)
+               sbitmap_deferred_clear(sb, i);
+
        sb->depth = depth;
        sb->map_nr = DIV_ROUND_UP(sb->depth, bits_per_word);
 
@@ -111,6 +156,24 @@ static int __sbitmap_get_word(unsigned long *word, unsigned long depth,
        return nr;
 }
 
+static int sbitmap_find_bit_in_index(struct sbitmap *sb, int index,
+                                    unsigned int alloc_hint, bool round_robin)
+{
+       int nr;
+
+       do {
+               nr = __sbitmap_get_word(&sb->map[index].word,
+                                       sb->map[index].depth, alloc_hint,
+                                       !round_robin);
+               if (nr != -1)
+                       break;
+               if (!sbitmap_deferred_clear(sb, index))
+                       break;
+       } while (1);
+
+       return nr;
+}
+
 int sbitmap_get(struct sbitmap *sb, unsigned int alloc_hint, bool round_robin)
 {
        unsigned int i, index;
@@ -118,24 +181,28 @@ int sbitmap_get(struct sbitmap *sb, unsigned int alloc_hint, bool round_robin)
 
        index = SB_NR_TO_INDEX(sb, alloc_hint);
 
+       /*
+        * Unless we're doing round robin tag allocation, just use the
+        * alloc_hint to find the right word index. No point in looping
+        * twice in find_next_zero_bit() for that case.
+        */
+       if (round_robin)
+               alloc_hint = SB_NR_TO_BIT(sb, alloc_hint);
+       else
+               alloc_hint = 0;
+
        for (i = 0; i < sb->map_nr; i++) {
-               nr = __sbitmap_get_word(&sb->map[index].word,
-                                       sb->map[index].depth,
-                                       SB_NR_TO_BIT(sb, alloc_hint),
-                                       !round_robin);
+               nr = sbitmap_find_bit_in_index(sb, index, alloc_hint,
+                                               round_robin);
                if (nr != -1) {
                        nr += index << sb->shift;
                        break;
                }
 
                /* Jump to next index. */
-               index++;
-               alloc_hint = index << sb->shift;
-
-               if (index >= sb->map_nr) {
+               alloc_hint = 0;
+               if (++index >= sb->map_nr)
                        index = 0;
-                       alloc_hint = 0;
-               }
        }
 
        return nr;
@@ -151,6 +218,7 @@ int sbitmap_get_shallow(struct sbitmap *sb, unsigned int alloc_hint,
        index = SB_NR_TO_INDEX(sb, alloc_hint);
 
        for (i = 0; i < sb->map_nr; i++) {
+again:
                nr = __sbitmap_get_word(&sb->map[index].word,
                                        min(sb->map[index].depth, shallow_depth),
                                        SB_NR_TO_BIT(sb, alloc_hint), true);
@@ -159,6 +227,9 @@ int sbitmap_get_shallow(struct sbitmap *sb, unsigned int alloc_hint,
                        break;
                }
 
+               if (sbitmap_deferred_clear(sb, index))
+                       goto again;
+
                /* Jump to next index. */
                index++;
                alloc_hint = index << sb->shift;
@@ -178,7 +249,7 @@ bool sbitmap_any_bit_set(const struct sbitmap *sb)
        unsigned int i;
 
        for (i = 0; i < sb->map_nr; i++) {
-               if (sb->map[i].word)
+               if (sb->map[i].word & ~sb->map[i].cleared)
                        return true;
        }
        return false;
@@ -191,9 +262,10 @@ bool sbitmap_any_bit_clear(const struct sbitmap *sb)
 
        for (i = 0; i < sb->map_nr; i++) {
                const struct sbitmap_word *word = &sb->map[i];
+               unsigned long mask = word->word & ~word->cleared;
                unsigned long ret;
 
-               ret = find_first_zero_bit(&word->word, word->depth);
+               ret = find_first_zero_bit(&mask, word->depth);
                if (ret < word->depth)
                        return true;
        }
@@ -201,23 +273,36 @@ bool sbitmap_any_bit_clear(const struct sbitmap *sb)
 }
 EXPORT_SYMBOL_GPL(sbitmap_any_bit_clear);
 
-unsigned int sbitmap_weight(const struct sbitmap *sb)
+static unsigned int __sbitmap_weight(const struct sbitmap *sb, bool set)
 {
        unsigned int i, weight = 0;
 
        for (i = 0; i < sb->map_nr; i++) {
                const struct sbitmap_word *word = &sb->map[i];
 
-               weight += bitmap_weight(&word->word, word->depth);
+               if (set)
+                       weight += bitmap_weight(&word->word, word->depth);
+               else
+                       weight += bitmap_weight(&word->cleared, word->depth);
        }
        return weight;
 }
-EXPORT_SYMBOL_GPL(sbitmap_weight);
+
+static unsigned int sbitmap_weight(const struct sbitmap *sb)
+{
+       return __sbitmap_weight(sb, true);
+}
+
+static unsigned int sbitmap_cleared(const struct sbitmap *sb)
+{
+       return __sbitmap_weight(sb, false);
+}
 
 void sbitmap_show(struct sbitmap *sb, struct seq_file *m)
 {
        seq_printf(m, "depth=%u\n", sb->depth);
-       seq_printf(m, "busy=%u\n", sbitmap_weight(sb));
+       seq_printf(m, "busy=%u\n", sbitmap_weight(sb) - sbitmap_cleared(sb));
+       seq_printf(m, "cleared=%u\n", sbitmap_cleared(sb));
        seq_printf(m, "bits_per_word=%u\n", 1U << sb->shift);
        seq_printf(m, "map_nr=%u\n", sb->map_nr);
 }
@@ -325,6 +410,7 @@ int sbitmap_queue_init_node(struct sbitmap_queue *sbq, unsigned int depth,
        sbq->min_shallow_depth = UINT_MAX;
        sbq->wake_batch = sbq_calc_wake_batch(sbq, depth);
        atomic_set(&sbq->wake_index, 0);
+       atomic_set(&sbq->ws_active, 0);
 
        sbq->ws = kzalloc_node(SBQ_WAIT_QUEUES * sizeof(*sbq->ws), flags, node);
        if (!sbq->ws) {
@@ -440,6 +526,9 @@ static struct sbq_wait_state *sbq_wake_ptr(struct sbitmap_queue *sbq)
 {
        int i, wake_index;
 
+       if (!atomic_read(&sbq->ws_active))
+               return NULL;
+
        wake_index = atomic_read(&sbq->wake_index);
        for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
                struct sbq_wait_state *ws = &sbq->ws[wake_index];
@@ -509,7 +598,8 @@ EXPORT_SYMBOL_GPL(sbitmap_queue_wake_up);
 void sbitmap_queue_clear(struct sbitmap_queue *sbq, unsigned int nr,
                         unsigned int cpu)
 {
-       sbitmap_clear_bit_unlock(&sbq->sb, nr);
+       sbitmap_deferred_clear_bit(&sbq->sb, nr);
+
        /*
         * Pairs with the memory barrier in set_current_state() to ensure the
         * proper ordering of clear_bit_unlock()/waitqueue_active() in the waker
@@ -564,6 +654,7 @@ void sbitmap_queue_show(struct sbitmap_queue *sbq, struct seq_file *m)
 
        seq_printf(m, "wake_batch=%u\n", sbq->wake_batch);
        seq_printf(m, "wake_index=%d\n", atomic_read(&sbq->wake_index));
+       seq_printf(m, "ws_active=%d\n", atomic_read(&sbq->ws_active));
 
        seq_puts(m, "ws={\n");
        for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
@@ -579,3 +670,48 @@ void sbitmap_queue_show(struct sbitmap_queue *sbq, struct seq_file *m)
        seq_printf(m, "min_shallow_depth=%u\n", sbq->min_shallow_depth);
 }
 EXPORT_SYMBOL_GPL(sbitmap_queue_show);
+
+void sbitmap_add_wait_queue(struct sbitmap_queue *sbq,
+                           struct sbq_wait_state *ws,
+                           struct sbq_wait *sbq_wait)
+{
+       if (!sbq_wait->sbq) {
+               sbq_wait->sbq = sbq;
+               atomic_inc(&sbq->ws_active);
+       }
+       add_wait_queue(&ws->wait, &sbq_wait->wait);
+}
+EXPORT_SYMBOL_GPL(sbitmap_add_wait_queue);
+
+void sbitmap_del_wait_queue(struct sbq_wait *sbq_wait)
+{
+       list_del_init(&sbq_wait->wait.entry);
+       if (sbq_wait->sbq) {
+               atomic_dec(&sbq_wait->sbq->ws_active);
+               sbq_wait->sbq = NULL;
+       }
+}
+EXPORT_SYMBOL_GPL(sbitmap_del_wait_queue);
+
+void sbitmap_prepare_to_wait(struct sbitmap_queue *sbq,
+                            struct sbq_wait_state *ws,
+                            struct sbq_wait *sbq_wait, int state)
+{
+       if (!sbq_wait->sbq) {
+               atomic_inc(&sbq->ws_active);
+               sbq_wait->sbq = sbq;
+       }
+       prepare_to_wait_exclusive(&ws->wait, &sbq_wait->wait, state);
+}
+EXPORT_SYMBOL_GPL(sbitmap_prepare_to_wait);
+
+void sbitmap_finish_wait(struct sbitmap_queue *sbq, struct sbq_wait_state *ws,
+                        struct sbq_wait *sbq_wait)
+{
+       finish_wait(&ws->wait, &sbq_wait->wait);
+       if (sbq_wait->sbq) {
+               atomic_dec(&sbq->ws_active);
+               sbq_wait->sbq = NULL;
+       }
+}
+EXPORT_SYMBOL_GPL(sbitmap_finish_wait);