for (i = 0; i < req->num_pages; i++)
                __free_page(req->pages[i]);
-       fuse_file_put(req->ff, false);
+
+       if (req->ff)
+               fuse_file_put(req->ff, false);
 }
 
 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
 
        mapping_set_error(inode->i_mapping, req->out.h.error);
        spin_lock(&fc->lock);
+       while (req->misc.write.next) {
+               struct fuse_req *next = req->misc.write.next;
+               req->misc.write.next = next->misc.write.next;
+               next->misc.write.next = NULL;
+               list_add(&next->writepages_entry, &fi->writepages);
+               list_add_tail(&next->list, &fi->queued_writes);
+               fuse_flush_writepages(inode);
+       }
        fi->writectr--;
        fuse_writepage_finish(fc, req);
        spin_unlock(&fc->lock);
 
        copy_highpage(tmp_page, page);
        req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
+       req->misc.write.next = NULL;
        req->in.argpages = 1;
        req->num_pages = 1;
        req->pages[0] = tmp_page;
                end_page_writeback(data->orig_pages[i]);
 }
 
+static bool fuse_writepage_in_flight(struct fuse_req *new_req,
+                                    struct page *page)
+{
+       struct fuse_conn *fc = get_fuse_conn(new_req->inode);
+       struct fuse_inode *fi = get_fuse_inode(new_req->inode);
+       struct fuse_req *tmp;
+       struct fuse_req *old_req;
+       bool found = false;
+       pgoff_t curr_index;
+
+       BUG_ON(new_req->num_pages != 0);
+
+       spin_lock(&fc->lock);
+       list_del(&new_req->writepages_entry);
+       new_req->num_pages = 1;
+       list_for_each_entry(old_req, &fi->writepages, writepages_entry) {
+               BUG_ON(old_req->inode != new_req->inode);
+               curr_index = old_req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
+               if (curr_index <= page->index &&
+                   page->index < curr_index + old_req->num_pages) {
+                       found = true;
+                       break;
+               }
+       }
+       if (!found)
+               goto out_unlock;
+
+       for (tmp = old_req; tmp != NULL; tmp = tmp->misc.write.next) {
+               BUG_ON(tmp->inode != new_req->inode);
+               curr_index = tmp->misc.write.in.offset >> PAGE_CACHE_SHIFT;
+               if (tmp->num_pages == 1 &&
+                   curr_index == page->index) {
+                       old_req = tmp;
+               }
+       }
+
+       if (old_req->num_pages == 1 && (old_req->state == FUSE_REQ_INIT ||
+                                       old_req->state == FUSE_REQ_PENDING)) {
+               copy_highpage(old_req->pages[0], page);
+               spin_unlock(&fc->lock);
+
+               dec_bdi_stat(page->mapping->backing_dev_info, BDI_WRITEBACK);
+               dec_zone_page_state(page, NR_WRITEBACK_TEMP);
+               fuse_writepage_free(fc, new_req);
+               fuse_request_free(new_req);
+               goto out;
+       } else {
+               new_req->misc.write.next = old_req->misc.write.next;
+               old_req->misc.write.next = new_req;
+       }
+out_unlock:
+       spin_unlock(&fc->lock);
+out:
+       return found;
+}
+
 static int fuse_writepages_fill(struct page *page,
                struct writeback_control *wbc, void *_data)
 {
        struct inode *inode = data->inode;
        struct fuse_conn *fc = get_fuse_conn(inode);
        struct page *tmp_page;
+       bool is_writeback;
        int err;
 
        if (!data->ff) {
                        goto out_unlock;
        }
 
-       if (req) {
-               BUG_ON(!req->num_pages);
-               if (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
-                   (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_write ||
-                   data->orig_pages[req->num_pages - 1]->index + 1 != page->index) {
+       /*
+        * Being under writeback is unlikely but possible.  For example direct
+        * read to an mmaped fuse file will set the page dirty twice; once when
+        * the pages are faulted with get_user_pages(), and then after the read
+        * completed.
+        */
+       is_writeback = fuse_page_is_writeback(inode, page->index);
 
-                       fuse_writepages_send(data);
-                       data->req = NULL;
-               }
+       if (req && req->num_pages &&
+           (is_writeback || req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
+            (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_write ||
+            data->orig_pages[req->num_pages - 1]->index + 1 != page->index)) {
+               fuse_writepages_send(data);
+               data->req = NULL;
        }
        err = -ENOMEM;
        tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
 
                fuse_write_fill(req, data->ff, page_offset(page), 0);
                req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
+               req->misc.write.next = NULL;
                req->in.argpages = 1;
                req->background = 1;
                req->num_pages = 0;
 
        inc_bdi_stat(page->mapping->backing_dev_info, BDI_WRITEBACK);
        inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
+
+       err = 0;
+       if (is_writeback && fuse_writepage_in_flight(req, page)) {
+               end_page_writeback(page);
+               data->req = NULL;
+               goto out_unlock;
+       }
        data->orig_pages[req->num_pages] = page;
 
        /*
        req->num_pages++;
        spin_unlock(&fc->lock);
 
-       err = 0;
 out_unlock:
        unlock_page(page);