From: Alexander Boyko <alexander.boyko@hpe.com>
Date: Wed, 5 Oct 2022 07:06:59 +0000 (-0700)
Subject: LU-14599 osp: limit allocation at osp_sync_process_committed
X-Git-Url: https://git.whamcloud.com/?a=commitdiff_plain;h=989e202f8e9ee6676486450baec81ee4e92a7f9c;p=fs%2Flustre-release.git

LU-14599 osp: limit allocation at osp_sync_process_committed

Sometimes osp cancels very large cookie list with 64K elements.
In this case osp_sync_process_committed() tries to allocate 64 pages
and uses vmalloc.
The fix limits memory allocation size to 4 page with kmalloc, and
reuse it in a loop.

Lustre-change: https://review.whamcloud.com/43250
Lustre-commit: 9b692e2e7d105f4926649ea46007ac65b24c4b6d

HPE-bug-id: LUS-9815
Fixes: 6d7332102 ("LU-11924 osp: combine llog cancel operations")
Signed-off-by: Alexander Boyko <alexander.boyko@hpe.com>
Change-Id: Ic875335a28f78494fdb3cbc4b0145e5a43831ee8
Reviewed-on: https://review.whamcloud.com/c/ex/lustre-release/+/48773
Tested-by: jenkins <devops@whamcloud.com>
Tested-by: Maloo <maloo@whamcloud.com>
Reviewed-by: Andreas Dilger <adilger@whamcloud.com>
---

diff --git a/lustre/osp/osp_sync.c b/lustre/osp/osp_sync.c
index d49190a..bed2cf8 100644
--- a/lustre/osp/osp_sync.c
+++ b/lustre/osp/osp_sync.c
@@ -991,7 +991,7 @@ static void osp_sync_process_committed(const struct lu_env *env,
 	struct ptlrpc_request	*req;
 	struct llog_ctxt	*ctxt;
 	struct llog_handle	*llh;
-	int			*arr;
+	int			*arr, arr_size;
 	LIST_HEAD(list);
 	struct list_head	 *le;
 	struct llog_logid	 lgid;
@@ -1031,10 +1031,16 @@ static void osp_sync_process_committed(const struct lu_env *env,
 
 	list_for_each(le, &list)
 		count++;
-	if (count > 2)
-		OBD_ALLOC_PTR_ARRAY_LARGE(arr, count);
-	else
+	if (count > 2) {
+		arr_size = sizeof(int) * count;
+		/* limit cookie array to order 2 */
+		arr_size = arr_size < (PAGE_SIZE * 4) ? arr_size :
+			(PAGE_SIZE * 4);
+		OBD_ALLOC_LARGE(arr, arr_size);
+	} else {
 		arr = NULL;
+		arr_size = 0;
+	}
 	i = 0;
 	while (!list_empty(&list)) {
 		struct osp_job_req_args	*jra;
@@ -1063,7 +1069,7 @@ static void osp_sync_process_committed(const struct lu_env *env,
 				rc = llog_cat_cancel_records(env, llh, 1,
 							     &jra->jra_lcookie);
 				if (rc)
-					CERROR("%s: can't cancel record: %d\n",
+					CERROR("%s: can't cancel record: rc = %d\n",
 					       obd->obd_name, rc);
 			}
 		} else {
@@ -1072,20 +1078,25 @@ static void osp_sync_process_committed(const struct lu_env *env,
 		}
 		ptlrpc_req_finished(req);
 		done++;
+		if (arr &&
+		    ((i * sizeof(int)) == arr_size ||
+		     (list_empty(&list) && i > 0))) {
+			rc = llog_cat_cancel_arr_rec(env, llh, &lgid, i, arr);
+
+			if (rc)
+				CERROR("%s: can't cancel %d records: rc = %d\n",
+				       obd->obd_name, i, rc);
+			else
+				CDEBUG(D_OTHER, "%s: massive records cancel id "DFID" num %d\n",
+				       obd->obd_name, PFID(&lgid.lgl_oi.oi_fid),
+				       i);
+			i = 0;
+		}
+
 	}
-	if (arr && i > 0) {
-		rc = llog_cat_cancel_arr_rec(env, llh, &lgid, i, arr);
-
-		if (rc)
-			CERROR("%s: can't cancel %d records rc: %d\n",
-			       obd->obd_name, i, rc);
-		else
-			CDEBUG(D_OTHER, "%s: massive records cancel id "DFID\
-			       " num %d\n", obd->obd_name,
-			       PFID(&lgid.lgl_oi.oi_fid), i);
-	}
+
 	if (arr)
-		OBD_FREE_PTR_ARRAY_LARGE(arr, count);
+		OBD_FREE_LARGE(arr, arr_size);
 
 	llog_ctxt_put(ctxt);