Ross Zwisler [off-list ref] writes:

To properly handle fsync/msync in an efficient way DAX needs to track dirty
pages so it is able to flush them durably to media on demand.

Please see coments below

IMHO it would be sane to call that function as dax_radix_entry_insert() 

+		sector_t sector, bool pmd_entry, bool dirty)
+{
+	struct radix_tree_root *page_tree = &mapping->page_tree;
+	int type, error = 0;
+	void *entry;
+
+	__mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
+
+	spin_lock_irq(&mapping->tree_lock);
+	entry = radix_tree_lookup(page_tree, index);
+
+	if (entry) {
+		type = RADIX_DAX_TYPE(entry);
+		if (WARN_ON_ONCE(type != RADIX_DAX_PTE &&
+					type != RADIX_DAX_PMD)) {
+			error = -EIO;
+			goto unlock;
+		}
+
+		if (!pmd_entry || type == RADIX_DAX_PMD)
+			goto dirty;
+		radix_tree_delete(&mapping->page_tree, index);
+		mapping->nrexceptional--;
+	}
+
+	if (sector == NO_SECTOR) {
+		/*
+		 * This can happen during correct operation if our pfn_mkwrite
+		 * fault raced against a hole punch operation.  If this
+		 * happens the pte that was hole punched will have been
+		 * unmapped and the radix tree entry will have been removed by
+		 * the time we are called, but the call will still happen.  We
+		 * will return all the way up to wp_pfn_shared(), where the
+		 * pte_same() check will fail, eventually causing page fault
+		 * to be retried by the CPU.
+		 */
+		goto unlock;
+	}
+
+	error = radix_tree_insert(page_tree, index,
+			RADIX_DAX_ENTRY(sector, pmd_entry));
+	if (error)
+		goto unlock;
+
+	mapping->nrexceptional++;
+ dirty:
+	if (dirty)
+		radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY);
+ unlock:
+	spin_unlock_irq(&mapping->tree_lock);
+	return error;
+}
+
+static int dax_writeback_one(struct block_device *bdev,
+		struct address_space *mapping, pgoff_t index, void *entry)
+{
+	struct radix_tree_root *page_tree = &mapping->page_tree;
+	int type = RADIX_DAX_TYPE(entry);
+	struct radix_tree_node *node;
+	struct blk_dax_ctl dax;
+	void **slot;
+	int ret = 0;
+
+	spin_lock_irq(&mapping->tree_lock);
+	/*
+	 * Regular page slots are stabilized by the page lock even
+	 * without the tree itself locked.  These unlocked entries
+	 * need verification under the tree lock.
+	 */
+	if (!__radix_tree_lookup(page_tree, index, &node, &slot))
+		goto unlock;
+	if (*slot != entry)
+		goto unlock;
+
+	/* another fsync thread may have already written back this entry */
+	if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))
+		goto unlock;
+
+	radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE);
+
+	if (WARN_ON_ONCE(type != RADIX_DAX_PTE && type != RADIX_DAX_PMD)) {
+		ret = -EIO;
+		goto unlock;
+	}
+
+	dax.sector = RADIX_DAX_SECTOR(entry);
+	dax.size = (type == RADIX_DAX_PMD ? PMD_SIZE : PAGE_SIZE);
+	spin_unlock_irq(&mapping->tree_lock);
+
+	/*
+	 * We cannot hold tree_lock while calling dax_map_atomic() because it
+	 * eventually calls cond_resched().
+	 */
+	ret = dax_map_atomic(bdev, &dax);
+	if (ret < 0)
+		return ret;
+
+	if (WARN_ON_ONCE(ret < dax.size)) {
+		ret = -EIO;
+		goto unmap;
+	}
+
+	wb_cache_pmem(dax.addr, dax.size);
+ unmap:
+	dax_unmap_atomic(bdev, &dax);
+	return ret;
+
+ unlock:
+	spin_unlock_irq(&mapping->tree_lock);
+	return ret;
+}
+
+/*
+ * Flush the mapping to the persistent domain within the byte range of [start,
+ * end]. This is required by data integrity operations to ensure file data is
+ * on persistent storage prior to completion of the operation.
+ */
+int dax_writeback_mapping_range(struct address_space *mapping, loff_t start,
+		loff_t end)
+{
+	struct inode *inode = mapping->host;
+	struct block_device *bdev = inode->i_sb->s_bdev;
+	pgoff_t indices[PAGEVEC_SIZE];
+	pgoff_t start_page, end_page;
+	struct pagevec pvec;
+	void *entry;
+	int i, ret = 0;
+
+	if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
+		return -EIO;
+
+	rcu_read_lock();
+	entry = radix_tree_lookup(&mapping->page_tree, start & PMD_MASK);
+	rcu_read_unlock();
+
+	/* see if the start of our range is covered by a PMD entry */
+	if (entry && RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD)
+		start &= PMD_MASK;
+
+	start_page = start >> PAGE_CACHE_SHIFT;
+	end_page = end >> PAGE_CACHE_SHIFT;
+
+	tag_pages_for_writeback(mapping, start_page, end_page);
+
+	pagevec_init(&pvec, 0);
+	while (1) {
+		pvec.nr = find_get_entries_tag(mapping, start_page,
+				PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE,
+				pvec.pages, indices);
+
+		if (pvec.nr == 0)
+			break;
+
+		for (i = 0; i < pvec.nr; i++) {
+			ret = dax_writeback_one(bdev, mapping, indices[i],
+					pvec.pages[i]);
+			if (ret < 0)
+				return ret;
+		}

I think it would be more efficient to use batched locking like follows:
                spin_lock_irq(&mapping->tree_lock);
		for (i = 0; i < pvec.nr; i++) {
                    struct blk_dax_ctl dax[PAGEVEC_SIZE];                
                    radix_tree_tag_clear(page_tree, indices[i], PAGECACHE_TAG_TOWRITE);
                    /* It is also reasonable to merge adjacent dax
                     * regions in to one */
                    dax[i].sector = RADIX_DAX_SECTOR(entry);
                    dax[i].size = (type == RADIX_DAX_PMD ? PMD_SIZE : PAGE_SIZE);                    

                }
                spin_unlock_irq(&mapping->tree_lock);
               	if (blk_queue_enter(q, true) != 0)
                    goto error;
                for (i = 0; i < pvec.nr; i++) {
                    rc = bdev_direct_access(bdev, dax[i]);
                    wb_cache_pmem(dax[i].addr, dax[i].size);
                }
                ret = blk_queue_exit(q, true)

I've realized that I do not understand why dax_fault code works at all.
During dax_fault we want to remove page from mapping and insert dax-entry
 Basically code looks like follows:
0 page = find_get_page()
1 lock_page(page)
2 delete_from_page_cache(page);
3 unlock_page(page);
4 dax_insert_mapping(inode, &bh, vma, vmf);

BUT what on earth protects us from other process to reinsert page again
after step(2) but before (4)?
Imagine we do write to file-hole which result in to dax_fault(write), but
another task also does read fault and reinsert deleted page via dax_hole_load
As result dax_tree_entry will fail with EIO
Testcase looks very trivial, but i can not reproduce this.