dax: coordinate locking for offsets in PMD range
authorRoss Zwisler <ross.zwisler@linux.intel.com>
Tue, 8 Nov 2016 00:32:20 +0000 (11:32 +1100)
committerDave Chinner <david@fromorbit.com>
Tue, 8 Nov 2016 00:32:20 +0000 (11:32 +1100)
DAX radix tree locking currently locks entries based on the unique
combination of the 'mapping' pointer and the pgoff_t 'index' for the entry.
This works for PTEs, but as we move to PMDs we will need to have all the
offsets within the range covered by the PMD to map to the same bit lock.
To accomplish this, for ranges covered by a PMD entry we will instead lock
based on the page offset of the beginning of the PMD entry.  The 'mapping'
pointer is still used in the same way.

Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Dave Chinner <david@fromorbit.com>
fs/dax.c
include/linux/dax.h
mm/filemap.c

index 835e7f082cff4aa6fec11392fd38ef300619bf32..72387023545e34688b60d0bff4efbbe8ad334cfa 100644 (file)
--- a/fs/dax.c
+++ b/fs/dax.c
@@ -64,14 +64,6 @@ static int __init init_dax_wait_table(void)
 }
 fs_initcall(init_dax_wait_table);
 
-static wait_queue_head_t *dax_entry_waitqueue(struct address_space *mapping,
-                                             pgoff_t index)
-{
-       unsigned long hash = hash_long((unsigned long)mapping ^ index,
-                                      DAX_WAIT_TABLE_BITS);
-       return wait_table + hash;
-}
-
 static long dax_map_atomic(struct block_device *bdev, struct blk_dax_ctl *dax)
 {
        struct request_queue *q = bdev->bd_queue;
@@ -285,7 +277,7 @@ EXPORT_SYMBOL_GPL(dax_do_io);
  */
 struct exceptional_entry_key {
        struct address_space *mapping;
-       unsigned long index;
+       pgoff_t entry_start;
 };
 
 struct wait_exceptional_entry_queue {
@@ -293,6 +285,26 @@ struct wait_exceptional_entry_queue {
        struct exceptional_entry_key key;
 };
 
+static wait_queue_head_t *dax_entry_waitqueue(struct address_space *mapping,
+               pgoff_t index, void *entry, struct exceptional_entry_key *key)
+{
+       unsigned long hash;
+
+       /*
+        * If 'entry' is a PMD, align the 'index' that we use for the wait
+        * queue to the start of that PMD.  This ensures that all offsets in
+        * the range covered by the PMD map to the same bit lock.
+        */
+       if (RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD)
+               index &= ~((1UL << (PMD_SHIFT - PAGE_SHIFT)) - 1);
+
+       key->mapping = mapping;
+       key->entry_start = index;
+
+       hash = hash_long((unsigned long)mapping ^ index, DAX_WAIT_TABLE_BITS);
+       return wait_table + hash;
+}
+
 static int wake_exceptional_entry_func(wait_queue_t *wait, unsigned int mode,
                                       int sync, void *keyp)
 {
@@ -301,7 +313,7 @@ static int wake_exceptional_entry_func(wait_queue_t *wait, unsigned int mode,
                container_of(wait, struct wait_exceptional_entry_queue, wait);
 
        if (key->mapping != ewait->key.mapping ||
-           key->index != ewait->key.index)
+           key->entry_start != ewait->key.entry_start)
                return 0;
        return autoremove_wake_function(wait, mode, sync, NULL);
 }
@@ -359,12 +371,10 @@ static void *get_unlocked_mapping_entry(struct address_space *mapping,
 {
        void *entry, **slot;
        struct wait_exceptional_entry_queue ewait;
-       wait_queue_head_t *wq = dax_entry_waitqueue(mapping, index);
+       wait_queue_head_t *wq;
 
        init_wait(&ewait.wait);
        ewait.wait.func = wake_exceptional_entry_func;
-       ewait.key.mapping = mapping;
-       ewait.key.index = index;
 
        for (;;) {
                entry = __radix_tree_lookup(&mapping->page_tree, index, NULL,
@@ -375,6 +385,8 @@ static void *get_unlocked_mapping_entry(struct address_space *mapping,
                                *slotp = slot;
                        return entry;
                }
+
+               wq = dax_entry_waitqueue(mapping, index, entry, &ewait.key);
                prepare_to_wait_exclusive(wq, &ewait.wait,
                                          TASK_UNINTERRUPTIBLE);
                spin_unlock_irq(&mapping->tree_lock);
@@ -447,10 +459,20 @@ restart:
        return entry;
 }
 
+/*
+ * We do not necessarily hold the mapping->tree_lock when we call this
+ * function so it is possible that 'entry' is no longer a valid item in the
+ * radix tree.  This is okay, though, because all we really need to do is to
+ * find the correct waitqueue where tasks might be sleeping waiting for that
+ * old 'entry' and wake them.
+ */
 void dax_wake_mapping_entry_waiter(struct address_space *mapping,
-                                  pgoff_t index, bool wake_all)
+               pgoff_t index, void *entry, bool wake_all)
 {
-       wait_queue_head_t *wq = dax_entry_waitqueue(mapping, index);
+       struct exceptional_entry_key key;
+       wait_queue_head_t *wq;
+
+       wq = dax_entry_waitqueue(mapping, index, entry, &key);
 
        /*
         * Checking for locked entry and prepare_to_wait_exclusive() happens
@@ -458,13 +480,8 @@ void dax_wake_mapping_entry_waiter(struct address_space *mapping,
         * So at this point all tasks that could have seen our entry locked
         * must be in the waitqueue and the following check will see them.
         */
-       if (waitqueue_active(wq)) {
-               struct exceptional_entry_key key;
-
-               key.mapping = mapping;
-               key.index = index;
+       if (waitqueue_active(wq))
                __wake_up(wq, TASK_NORMAL, wake_all ? 0 : 1, &key);
-       }
 }
 
 void dax_unlock_mapping_entry(struct address_space *mapping, pgoff_t index)
@@ -480,7 +497,7 @@ void dax_unlock_mapping_entry(struct address_space *mapping, pgoff_t index)
        }
        unlock_slot(mapping, slot);
        spin_unlock_irq(&mapping->tree_lock);
-       dax_wake_mapping_entry_waiter(mapping, index, false);
+       dax_wake_mapping_entry_waiter(mapping, index, entry, false);
 }
 
 static void put_locked_mapping_entry(struct address_space *mapping,
@@ -505,7 +522,7 @@ static void put_unlocked_mapping_entry(struct address_space *mapping,
                return;
 
        /* We have to wake up next waiter for the radix tree entry lock */
-       dax_wake_mapping_entry_waiter(mapping, index, false);
+       dax_wake_mapping_entry_waiter(mapping, index, entry, false);
 }
 
 /*
@@ -532,7 +549,7 @@ int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index)
        radix_tree_delete(&mapping->page_tree, index);
        mapping->nrexceptional--;
        spin_unlock_irq(&mapping->tree_lock);
-       dax_wake_mapping_entry_waiter(mapping, index, true);
+       dax_wake_mapping_entry_waiter(mapping, index, entry, true);
 
        return 1;
 }
index add6c4bc568f150cea20c058efee0419f9672fab..a41a747d6112010036dcbdc7108f5e6529be2e6c 100644 (file)
@@ -22,7 +22,7 @@ int iomap_dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
 int dax_fault(struct vm_area_struct *, struct vm_fault *, get_block_t);
 int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index);
 void dax_wake_mapping_entry_waiter(struct address_space *mapping,
-                                  pgoff_t index, bool wake_all);
+               pgoff_t index, void *entry, bool wake_all);
 
 #ifdef CONFIG_FS_DAX
 struct page *read_dax_sector(struct block_device *bdev, sector_t n);
index 849f459ad0780e27bc256ff13fd52fa8c9007661..1ffb7dcd1b5d7b7ffd99e91882858d8f5511c558 100644 (file)
@@ -143,7 +143,7 @@ static int page_cache_tree_insert(struct address_space *mapping,
                        if (node)
                                workingset_node_pages_dec(node);
                        /* Wakeup waiters for exceptional entry lock */
-                       dax_wake_mapping_entry_waiter(mapping, page->index,
+                       dax_wake_mapping_entry_waiter(mapping, page->index, p,
                                                      false);
                }
        }