LU-2722 clio: directIO thread races with completion thread

ll_direct_IO_26()) ASSERTION( obj->cob_transient_pages == 0 ) failed

The directio thread puts pages to the transfer in osc_io_submit(), and
waits for its completion. Upon completion, osc_completion() is called,
which notifies waiters IO is completed and only after that put the
page. the original directIO thread wakes up and may succeed to assert
on the last page which is not yet put by osc_completion().

Signed-off-by: Vitaly Fertman <vitaly_fertman@xyratex.com>
Change-Id: Ib912e913885a9ff2cb1e9c865ae2003a676c2b5b
Xyratex-bug-id: MRP-838
Reviewed-on: http://review.whamcloud.com/5223
Reviewed-by: Jinshan Xiong <jinshan.xiong@intel.com>
Tested-by: Hudson
Tested-by: Maloo <whamcloud.maloo@gmail.com>
Reviewed-by: Niu Yawei <yawei.niu@intel.com>
Reviewed-by: Oleg Drokin <oleg.drokin@intel.com>

diff --git a/lustre/obdclass/cl_page.c b/lustre/obdclass/cl_page.c
index b34555a..29a570c 100644 (file)
--- a/lustre/obdclass/cl_page.c
+++ b/lustre/obdclass/cl_page.c
@@ -1319,8 +1319,17 @@ void cl_page_completion(const struct lu_env *env,
                 LASSERT(cl_page_is_vmlocked(env, pg));
                 LASSERT(pg->cp_sync_io == anchor);
                 pg->cp_sync_io = NULL;
+       }
+       /*
+        * As page->cp_obj is pinned by a reference from page->cp_req, it is
+        * safe to call cl_page_put() without risking object destruction in a
+        * non-blocking context.
+        */
+       cl_page_put(env, pg);
+
+       if (anchor)
                 cl_sync_io_note(anchor, ioret);
-        }
+
         EXIT;
 }
 EXPORT_SYMBOL(cl_page_completion);

diff --git a/lustre/osc/osc_cache.c b/lustre/osc/osc_cache.c
index fc896eb..37b249f 100644 (file)
--- a/lustre/osc/osc_cache.c
+++ b/lustre/osc/osc_cache.c
@@ -1283,8 +1283,6 @@ static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
        opg->ops_submit_time = 0;
        srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
 
-       cl_page_completion(env, page, crt, rc);
-
        /* statistic */
        if (rc == 0 && srvlock) {
                struct lu_device *ld    = opg->ops_cl.cpl_obj->co_lu.lo_dev;
@@ -1303,12 +1301,9 @@ static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
         * reference counter protects page from concurrent reclaim.
         */
        lu_ref_del(&page->cp_reference, "transfer", page);
-       /*
-        * As page->cp_obj is pinned by a reference from page->cp_req, it is
-        * safe to call cl_page_put() without risking object destruction in a
-        * non-blocking context.
-        */
-       cl_page_put(env, page);
+
+       cl_page_completion(env, page, crt, rc);
+
        RETURN(0);
 }