atomic_sub(page_count, &obd_unstable_pages);
        LASSERT(atomic_read(&obd_unstable_pages) >= 0);
 
-       spin_lock(&req->rq_lock);
-       req->rq_committed = 1;
-       req->rq_unstable  = 0;
-       spin_unlock(&req->rq_lock);
-
        wake_up_all(&cli->cl_cache->ccc_unstable_waitq);
 }
 
        LASSERT(atomic_read(&obd_unstable_pages) >= 0);
        atomic_add(page_count, &obd_unstable_pages);
 
-       spin_lock(&req->rq_lock);
-
        /*
         * If the request has already been committed (i.e. brw_commit
         * called via rq_commit_cb), we need to undo the unstable page
         * increments we just performed because rq_commit_cb wont be
-        * called again. Otherwise, just set the commit callback so the
-        * unstable page accounting is properly updated when the request
-        * is committed
+        * called again.
         */
-       if (req->rq_committed) {
+       spin_lock(&req->rq_lock);
+       if (unlikely(req->rq_committed)) {
                /* Drop lock before calling osc_dec_unstable_pages */
                spin_unlock(&req->rq_lock);
                osc_dec_unstable_pages(req);
-               spin_lock(&req->rq_lock);
        } else {
                req->rq_unstable = 1;
-               req->rq_commit_cb = osc_dec_unstable_pages;
+               spin_unlock(&req->rq_lock);
        }
-
-       spin_unlock(&req->rq_lock);
 }
 
 /* this must be called holding the loi list lock to give coverage to exit_cache,
 
 
 static void brw_commit(struct ptlrpc_request *req)
 {
-       spin_lock(&req->rq_lock);
        /*
         * If osc_inc_unstable_pages (via osc_extent_finish) races with
         * this called via the rq_commit_cb, I need to ensure
         * osc_dec_unstable_pages is still called. Otherwise unstable
         * pages may be leaked.
         */
-       if (req->rq_unstable) {
+       spin_lock(&req->rq_lock);
+       if (unlikely(req->rq_unstable)) {
+               req->rq_unstable = 0;
                spin_unlock(&req->rq_lock);
                osc_dec_unstable_pages(req);
-               spin_lock(&req->rq_lock);
        } else {
                req->rq_committed = 1;
+               spin_unlock(&req->rq_lock);
        }
-       spin_unlock(&req->rq_lock);
 }
 
 /**