#include <linux/buffer_head.h>
 #include <linux/blkdev.h>
 #include <linux/random.h>
+#include <linux/iocontext.h>
 #include <asm/div64.h>
 #include "compat.h"
 #include "ctree.h"
        int again = 0;
        unsigned long num_run = 0;
        unsigned long limit;
+       unsigned long last_waited = 0;
 
        bdi = device->bdev->bd_inode->i_mapping->backing_dev_info;
        fs_info = device->dev_root->fs_info;
                if (pending && bdi_write_congested(bdi) && num_run > 16 &&
                    fs_info->fs_devices->open_devices > 1) {
                        struct bio *old_head;
+                       struct io_context *ioc;
 
+                       ioc = current->io_context;
+
+                       /*
+                        * the main goal here is that we don't want to
+                        * block if we're going to be able to submit
+                        * more requests without blocking.
+                        *
+                        * This code does two great things, it pokes into
+                        * the elevator code from a filesystem _and_
+                        * it makes assumptions about how batching works.
+                        */
+                       if (ioc && ioc->nr_batch_requests > 0 &&
+                           time_before(jiffies, ioc->last_waited + HZ/50UL) &&
+                           (last_waited == 0 ||
+                            ioc->last_waited == last_waited)) {
+                               /*
+                                * we want to go through our batch of
+                                * requests and stop.  So, we copy out
+                                * the ioc->last_waited time and test
+                                * against it before looping
+                                */
+                               last_waited = ioc->last_waited;
+                               continue;
+                       }
                        spin_lock(&device->io_lock);
 
                        old_head = device->pending_bios;