Whamcloud - gitweb
LU-354 test: Change dev_set_rdonly() check to warning
[fs/lustre-release.git] / lustre / include / obd_support.h
index a5cc8a3..4c96565 100644 (file)
@@ -16,8 +16,8 @@
  * in the LICENSE file that accompanied this code).
  *
  * You should have received a copy of the GNU General Public License
- * version 2 along with this program; If not, see [sun.com URL with a
- * copy of GPLv2].
+ * version 2 along with this program; If not, see
+ * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  *
  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  * CA 95054 USA or visit www.sun.com if you need additional information or
  * GPL HEADER END
  */
 /*
- * Copyright  2008 Sun Microsystems, Inc. All rights reserved
+ * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
+ *
+ * Copyright (c) 2011, 2012, Whamcloud, Inc.
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
@@ -59,14 +61,6 @@ enum {
         OBD_STATS_NUM,
 };
 
-enum {
-        OBD_FAIL_LOC_NOSET      = 0,
-        OBD_FAIL_LOC_ORSET      = 1,
-        OBD_FAIL_LOC_RESET      = 2
-};
-
-extern unsigned long obd_fail_loc;
-extern unsigned int obd_fail_val;
 extern unsigned int obd_debug_peer_on_timeout;
 extern unsigned int obd_dump_on_timeout;
 extern unsigned int obd_dump_on_eviction;
@@ -74,29 +68,59 @@ extern unsigned int obd_dump_on_eviction;
    networking / disk / timings affected by load (use Adaptive Timeouts) */
 extern unsigned int obd_timeout;          /* seconds */
 extern unsigned int ldlm_timeout;         /* seconds */
+extern unsigned int at_min;
+extern unsigned int at_max;
+extern unsigned int at_history;
+extern int at_early_margin;
+extern int at_extra;
 extern unsigned int obd_sync_filter;
 extern unsigned int obd_max_dirty_pages;
-extern atomic_t obd_dirty_pages;
-extern cfs_waitq_t obd_race_waitq;
-extern int obd_race_state;
+extern cfs_atomic_t obd_dirty_pages;
+extern cfs_atomic_t obd_dirty_transit_pages;
 extern unsigned int obd_alloc_fail_rate;
 
-int __obd_fail_check_set(__u32 id, __u32 value, int set);
-int __obd_fail_timeout_set(__u32 id, __u32 value, int ms, int set);
-
 /* lvfs.c */
 int obd_alloc_fail(const void *ptr, const char *name, const char *type,
                    size_t size, const char *file, int line);
 
+/* Some hash init argument constants */
+#define HASH_POOLS_BKT_BITS 3
+#define HASH_POOLS_CUR_BITS 3
+#define HASH_POOLS_MAX_BITS 7
+#define HASH_UUID_BKT_BITS 5
+#define HASH_UUID_CUR_BITS 7
+#define HASH_UUID_MAX_BITS 12
+#define HASH_NID_BKT_BITS 5
+#define HASH_NID_CUR_BITS 7
+#define HASH_NID_MAX_BITS 12
+#define HASH_NID_STATS_BKT_BITS 5
+#define HASH_NID_STATS_CUR_BITS 7
+#define HASH_NID_STATS_MAX_BITS 12
+#define HASH_LQS_BKT_BITS 5
+#define HASH_LQS_CUR_BITS 7
+#define HASH_LQS_MAX_BITS 12
+#define HASH_CONN_BKT_BITS 5
+#define HASH_CONN_CUR_BITS 5
+#define HASH_CONN_MAX_BITS 15
+#define HASH_EXP_LOCK_BKT_BITS  5
+#define HASH_EXP_LOCK_CUR_BITS  7
+#define HASH_EXP_LOCK_MAX_BITS  16
+#define HASH_CL_ENV_BKT_BITS    5
+#define HASH_CL_ENV_BITS        10
+
 /* Timeout definitions */
 #define OBD_TIMEOUT_DEFAULT             100
 #define LDLM_TIMEOUT_DEFAULT            20
 #define MDS_LDLM_TIMEOUT_DEFAULT        6
-/* Time to wait for all clients to reconnect during recovery */
+/* Time to wait for all clients to reconnect during recovery (hard limit) */
+#define OBD_RECOVERY_TIME_HARD          (obd_timeout * 9)
+/* Time to wait for all clients to reconnect during recovery (soft limit) */
 /* Should be very conservative; must catch the first reconnect after reboot */
-#define OBD_RECOVERY_FACTOR (3) /* times obd_timeout */
+#define OBD_RECOVERY_TIME_SOFT          (obd_timeout * 3)
 /* Change recovery-small 26b time if you change this */
 #define PING_INTERVAL max(obd_timeout / 4, 1U)
+/* a bit more than maximal journal commit time in seconds */
+#define PING_INTERVAL_SHORT min(PING_INTERVAL, 7U)
 /* Client may skip 1 ping; we must wait at least 2.5. But for multiple
  * failover targets the client only pings one server at a time, and pings
  * can be lost on a loaded network. Since eviction has serious consequences,
@@ -122,8 +146,23 @@ int obd_alloc_fail(const void *ptr, const char *name, const char *type,
    chance to generate adaptive timeout data. */
 #define INITIAL_CONNECT_TIMEOUT max(CONNECTION_SWITCH_MIN,obd_timeout/2)
 #endif
+/* The max delay between connects is SWITCH_MAX + SWITCH_INC + INITIAL */
+#define RECONNECT_DELAY_MAX (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC + \
+                             INITIAL_CONNECT_TIMEOUT)
+/* The min time a target should wait for clients to reconnect in recovery */
+#define OBD_RECOVERY_TIME_MIN    (2*RECONNECT_DELAY_MAX)
+#define OBD_IR_FACTOR_MIN         1
+#define OBD_IR_FACTOR_MAX         10
+#define OBD_IR_FACTOR_DEFAULT    (OBD_IR_FACTOR_MAX/2)
+/* default timeout for the MGS to become IR_FULL */
+#define OBD_IR_MGS_TIMEOUT       (4*obd_timeout)
 #define LONG_UNLINK 300          /* Unlink should happen before now */
 
+/**
+ * Time interval of shrink, if the client is "idle" more than this interval,
+ * then the ll_grant thread will return the requested grant space to filter
+ */
+#define GRANT_SHRINK_INTERVAL            1200/*20 minutes*/
 
 #define OBD_FAIL_MDS                     0x100
 #define OBD_FAIL_MDS_HANDLE_UNPACK       0x101
@@ -180,16 +219,32 @@ int obd_alloc_fail(const void *ptr, const char *name, const char *type,
 #define OBD_FAIL_MDS_SETXATTR_WRITE      0x134
 #define OBD_FAIL_MDS_FS_SETUP            0x135
 #define OBD_FAIL_MDS_RESEND              0x136
-#define OBD_FAIL_MDS_IS_SUBDIR_NET       0x137
-#define OBD_FAIL_MDS_IS_SUBDIR_PACK      0x138
-#define OBD_FAIL_MDS_SET_INFO_NET        0x139
-#define OBD_FAIL_MDS_WRITEPAGE_NET       0x13a
-#define OBD_FAIL_MDS_WRITEPAGE_PACK      0x13b
-#define OBD_FAIL_MDS_LLOG_CREATE_FAILED  0x13c
-#define OBD_FAIL_MDS_OSC_PRECREATE       0x13d
-#define OBD_FAIL_MDS_LOV_SYNC_RACE       0x13e
-#define OBD_FAIL_MDS_CLOSE_NET_REP       0x13f
-#define OBD_FAIL_MDS_LLOG_SYNC_TIMEOUT   0x140
+#define OBD_FAIL_MDS_LLOG_CREATE_FAILED  0x137
+#define OBD_FAIL_MDS_LOV_SYNC_RACE       0x138
+#define OBD_FAIL_MDS_OSC_PRECREATE       0x139
+#define OBD_FAIL_MDS_LLOG_SYNC_TIMEOUT   0x13a
+#define OBD_FAIL_MDS_CLOSE_NET_REP       0x13b
+#define OBD_FAIL_MDS_BLOCK_QUOTA_REQ     0x13c
+#define OBD_FAIL_MDS_DROP_QUOTA_REQ      0x13d
+#define OBD_FAIL_MDS_REMOVE_COMMON_EA    0x13e
+#define OBD_FAIL_MDS_ALLOW_COMMON_EA_SETTING   0x13f
+#define OBD_FAIL_MDS_FAIL_LOV_LOG_ADD    0x140
+#define OBD_FAIL_MDS_LOV_PREP_CREATE     0x141
+#define OBD_FAIL_MDS_REINT_DELAY         0x142
+#define OBD_FAIL_MDS_READLINK_EPROTO     0x143
+#define OBD_FAIL_MDS_OPEN_WAIT_CREATE    0x144
+#define OBD_FAIL_MDS_PDO_LOCK            0x145
+#define OBD_FAIL_MDS_PDO_LOCK2           0x146
+#define OBD_FAIL_MDS_OSC_CREATE_FAIL     0x147
+
+/* CMD */
+#define OBD_FAIL_MDS_IS_SUBDIR_NET       0x180
+#define OBD_FAIL_MDS_IS_SUBDIR_PACK      0x181
+#define OBD_FAIL_MDS_SET_INFO_NET        0x182
+#define OBD_FAIL_MDS_WRITEPAGE_NET       0x183
+#define OBD_FAIL_MDS_WRITEPAGE_PACK      0x184
+#define OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS 0x185
+#define OBD_FAIL_MDS_GET_INFO_NET        0x186
 
 #define OBD_FAIL_OST                     0x200
 #define OBD_FAIL_OST_CONNECT_NET         0x201
@@ -229,6 +284,10 @@ int obd_alloc_fail(const void *ptr, const char *name, const char *type,
 #define OBD_FAIL_OST_PAUSE_CREATE        0x223
 #define OBD_FAIL_OST_BRW_PAUSE_PACK      0x224
 #define OBD_FAIL_OST_CONNECT_NET2        0x225
+#define OBD_FAIL_OST_NOMEM               0x226
+#define OBD_FAIL_OST_BRW_PAUSE_BULK2     0x227
+#define OBD_FAIL_OST_MAPBLK_ENOSPC       0x228
+#define OBD_FAIL_OST_ENOINO              0x229
 
 #define OBD_FAIL_LDLM                    0x300
 #define OBD_FAIL_LDLM_NAMESPACE_NEW      0x301
@@ -251,6 +310,16 @@ int obd_alloc_fail(const void *ptr, const char *name, const char *type,
 #define OBD_FAIL_LDLM_PAUSE_CANCEL       0x312
 #define OBD_FAIL_LDLM_CLOSE_THREAD       0x313
 #define OBD_FAIL_LDLM_CANCEL_BL_CB_RACE  0x314
+#define OBD_FAIL_LDLM_CP_CB_WAIT         0x315
+#define OBD_FAIL_LDLM_OST_FAIL_RACE      0x316
+#define OBD_FAIL_LDLM_INTR_CP_AST        0x317
+#define OBD_FAIL_LDLM_CP_BL_RACE         0x318
+#define OBD_FAIL_LDLM_NEW_LOCK           0x319
+#define OBD_FAIL_LDLM_AGL_DELAY          0x31a
+#define OBD_FAIL_LDLM_AGL_NOLOCK         0x31b
+
+/* LOCKLESS IO */
+#define OBD_FAIL_LDLM_SET_CONTENTION     0x385
 
 #define OBD_FAIL_OSC                     0x400
 #define OBD_FAIL_OSC_BRW_READ_BULK       0x401
@@ -266,6 +335,10 @@ int obd_alloc_fail(const void *ptr, const char *name, const char *type,
 #define OBD_FAIL_OSC_CONNECT_CKSUM       0x40b
 #define OBD_FAIL_OSC_CKSUM_ADLER_ONLY    0x40c
 #define OBD_FAIL_OSC_DIO_PAUSE           0x40d
+#define OBD_FAIL_OSC_OBJECT_CONTENTION   0x40e
+#define OBD_FAIL_OSC_CP_CANCEL_RACE      0x40f
+#define OBD_FAIL_OSC_CP_ENQ_RACE         0x410
+#define OBD_FAIL_OSC_NO_GRANT            0x411
 
 #define OBD_FAIL_PTLRPC                  0x500
 #define OBD_FAIL_PTLRPC_ACK              0x501
@@ -278,12 +351,23 @@ int obd_alloc_fail(const void *ptr, const char *name, const char *type,
 #define OBD_FAIL_PTLRPC_CLIENT_BULK_CB   0x508
 #define OBD_FAIL_PTLRPC_PAUSE_REQ        0x50a
 #define OBD_FAIL_PTLRPC_PAUSE_REP        0x50c
+#define OBD_FAIL_PTLRPC_IMP_DEACTIVE     0x50d
+#define OBD_FAIL_PTLRPC_DUMP_LOG         0x50e
+#define OBD_FAIL_PTLRPC_LONG_REPL_UNLINK 0x50f
+#define OBD_FAIL_PTLRPC_LONG_BULK_UNLINK 0x510
+#define OBD_FAIL_PTLRPC_HPREQ_TIMEOUT    0x511
+#define OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT  0x512
+#define OBD_FAIL_PTLRPC_DROP_REQ_OPC     0x513
+#define OBD_FAIL_PTLRPC_FINISH_REPLAY    0x514
+#define OBD_FAIL_PTLRPC_CLIENT_BULK_CB2  0x515
 
 #define OBD_FAIL_OBD_PING_NET            0x600
 #define OBD_FAIL_OBD_LOG_CANCEL_NET      0x601
 #define OBD_FAIL_OBD_LOGD_NET            0x602
 #define OBD_FAIL_OBD_QC_CALLBACK_NET     0x603
 #define OBD_FAIL_OBD_DQACQ               0x604
+#define OBD_FAIL_OBD_LLOG_SETUP          0x605
+#define OBD_FAIL_OBD_LOG_CANCEL_REP      0x606
 
 #define OBD_FAIL_TGT_REPLY_NET           0x700
 #define OBD_FAIL_TGT_CONN_RACE           0x701
@@ -293,6 +377,11 @@ int obd_alloc_fail(const void *ptr, const char *name, const char *type,
 #define OBD_FAIL_TGT_DELAY_PRECREATE     0x705
 #define OBD_FAIL_TGT_TOOMANY_THREADS     0x706
 #define OBD_FAIL_TGT_REPLAY_DROP         0x707
+#define OBD_FAIL_TGT_FAKE_EXP            0x708
+#define OBD_FAIL_TGT_REPLAY_DELAY        0x709
+#define OBD_FAIL_TGT_LAST_REPLAY         0x710
+#define OBD_FAIL_TGT_CLIENT_ADD          0x711
+#define OBD_FAIL_TGT_RCVG_FLAG           0x712
 
 #define OBD_FAIL_MDC_REVALIDATE_PAUSE    0x800
 #define OBD_FAIL_MDC_ENQUEUE_PAUSE       0x801
@@ -306,7 +395,9 @@ int obd_alloc_fail(const void *ptr, const char *name, const char *type,
 #define OBD_FAIL_MGS_PAUSE_REQ           0x904
 #define OBD_FAIL_MGS_PAUSE_TARGET_REG    0x905
 
-#define OBD_FAIL_QUOTA_QD_COUNT_32BIT    0xA00
+#define OBD_FAIL_QUOTA_RET_QDATA         0xA02
+#define OBD_FAIL_QUOTA_DELAY_REL         0xA03
+#define OBD_FAIL_QUOTA_DELAY_SD          0xA04
 
 #define OBD_FAIL_LPROC_REMOVE            0xB00
 
@@ -324,120 +415,37 @@ int obd_alloc_fail(const void *ptr, const char *name, const char *type,
 #define OBD_FAIL_SEC_CTX_FINI_NET        0x1203
 #define OBD_FAIL_SEC_CTX_HDL_PAUSE       0x1204
 
-/* Failure injection control */
-#define OBD_FAIL_MASK_SYS    0x0000FF00
-#define OBD_FAIL_MASK_LOC   (0x000000FF | OBD_FAIL_MASK_SYS)
-
-#define OBD_FAILED_BIT       30
-/* OBD_FAILED is 0x40000000 */
-#define OBD_FAILED          (1 << OBD_FAILED_BIT)
-
-#define OBD_FAIL_ONCE_BIT    31
-/* OBD_FAIL_ONCE is 0x80000000 */
-#define OBD_FAIL_ONCE       (1 << OBD_FAIL_ONCE_BIT)
-
-/* The following flags aren't made to be combined */
-#define OBD_FAIL_SKIP        0x20000000 /* skip N times then fail */
-#define OBD_FAIL_SOME        0x10000000 /* only fail N times */
-#define OBD_FAIL_RAND        0x08000000 /* fail 1/N of the times */
-#define OBD_FAIL_USR1        0x04000000 /* user flag */
-
-#define OBD_FAIL_PRECHECK(id) (obd_fail_loc &&                                \
-                              (obd_fail_loc & OBD_FAIL_MASK_LOC) ==           \
-                              ((id) & OBD_FAIL_MASK_LOC))
-
-static inline int obd_fail_check_set(__u32 id, __u32 value, int set)
-{
-        int ret = 0;
-        if (unlikely(OBD_FAIL_PRECHECK(id) &&
-            (ret = __obd_fail_check_set(id, value, set)))) {
-                CERROR("*** obd_fail_loc=%x ***\n", id);
-        }
-        return ret;
-}
-
-/* If id hit obd_fail_loc, return 1, otherwise return 0 */
-#define OBD_FAIL_CHECK(id) \
-        obd_fail_check_set(id, 0, OBD_FAIL_LOC_NOSET)
-
-/* If id hit obd_fail_loc, obd_fail_loc |= value and return 1,
- * otherwise return 0 */
-#define OBD_FAIL_CHECK_ORSET(id, value) \
-        obd_fail_check_set(id, value, OBD_FAIL_LOC_ORSET)
-
-/* If id hit obd_fail_loc, obd_fail_loc = value and return 1,
- * otherwise return 0 */
-#define OBD_FAIL_CHECK_RESET(id, value) \
-        obd_fail_check_set(id, value, OBD_FAIL_LOC_RESET)
-
-
-static inline int obd_fail_timeout_set(__u32 id, __u32 value, int ms, int set)
-{
-        if (unlikely(OBD_FAIL_PRECHECK(id)))
-                return __obd_fail_timeout_set(id, value, ms, set);
-        else
-                return 0;
-}
-
-/* If id hit obd_fail_loc, sleep for seconds or milliseconds */
-#define OBD_FAIL_TIMEOUT(id, secs) \
-        obd_fail_timeout_set(id, 0, secs * 1000, OBD_FAIL_LOC_NOSET)
-
-#define OBD_FAIL_TIMEOUT_MS(id, ms) \
-        obd_fail_timeout_set(id, 0, ms, OBD_FAIL_LOC_NOSET)
-
-/* If id hit obd_fail_loc, obd_fail_loc |= value and
- * sleep seconds or milliseconds */
-#define OBD_FAIL_TIMEOUT_ORSET(id, value, secs) \
-        obd_fail_timeout_set(id, value, secs * 1000, OBD_FAIL_LOC_ORSET)
-
-#define OBD_FAIL_TIMEOUT_MS_ORSET(id, value, ms) \
-        obd_fail_timeout_set(id, value, ms, OBD_FAIL_LOC_ORSET)
-
-#ifdef __KERNEL__
-static inline void obd_fail_write(int id, struct super_block *sb)
-{
-        /* We set FAIL_ONCE because we never "un-fail" a device */
-        if (OBD_FAIL_CHECK_ORSET(id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
-#ifdef LIBCFS_DEBUG
-                BDEVNAME_DECLARE_STORAGE(tmp);
-                CERROR("obd_fail_loc=%x, fail write operation on %s\n",
-                       id, ll_bdevname(sb, tmp));
-#endif
-                /* TODO-CMD: fix getting jdev */
-                __lvfs_set_rdonly(lvfs_sbdev(sb), (lvfs_sbdev_type)0);
-        }
-}
-#define OBD_FAIL_WRITE(id, sb) obd_fail_write(id, sb)
-
-/* The idea here is to synchronise two threads to force a race. The
- * first thread that calls this with a matching fail_loc is put to
- * sleep. The next thread that calls with the same fail_loc wakes up
- * the first and continues. */
-static inline void obd_race(__u32 id)
-{
-        if (OBD_FAIL_PRECHECK(id)) {
-                if (unlikely(__obd_fail_check_set(id, 0, OBD_FAIL_LOC_NOSET))) {
-                        obd_race_state = 0;
-                        CERROR("obd_race id %x sleeping\n", id);
-                        OBD_SLEEP_ON(obd_race_waitq, obd_race_state != 0);
-                        CERROR("obd_fail_race id %x awake\n", id);
-                } else {
-                        CERROR("obd_fail_race id %x waking\n", id);
-                        obd_race_state = 1;
-                        wake_up(&obd_race_waitq);
-                }
-        }
-}
-#define OBD_RACE(id) obd_race(id)
-#else
-/* sigh.  an expedient fix until OBD_RACE is fixed up */
-#define OBD_RACE(foo) do {} while(0)
-#endif
-
-#define fixme() CDEBUG(D_OTHER, "FIXME\n");
-
-extern atomic_t libcfs_kmemory;
+#define OBD_FAIL_LLOG                               0x1300
+#define OBD_FAIL_LLOG_ORIGIN_CONNECT_NET            0x1301
+#define OBD_FAIL_LLOG_ORIGIN_HANDLE_CREATE_NET      0x1302
+#define OBD_FAIL_LLOG_ORIGIN_HANDLE_DESTROY_NET     0x1303
+#define OBD_FAIL_LLOG_ORIGIN_HANDLE_READ_HEADER_NET 0x1304
+#define OBD_FAIL_LLOG_ORIGIN_HANDLE_NEXT_BLOCK_NET  0x1305
+#define OBD_FAIL_LLOG_ORIGIN_HANDLE_PREV_BLOCK_NET  0x1306
+#define OBD_FAIL_LLOG_ORIGIN_HANDLE_WRITE_REC_NET   0x1307
+#define OBD_FAIL_LLOG_ORIGIN_HANDLE_CLOSE_NET       0x1308
+#define OBD_FAIL_LLOG_CATINFO_NET                   0x1309
+#define OBD_FAIL_MDS_SYNC_CAPA_SL                   0x1310
+#define OBD_FAIL_SEQ_ALLOC                          0x1311
+
+#define OBD_FAIL_LLITE                              0x1400
+#define OBD_FAIL_LLITE_FAULT_TRUNC_RACE             0x1401
+
+/* Assign references to moved code to reduce code changes */
+#define OBD_FAIL_PRECHECK(id)                   CFS_FAIL_PRECHECK(id)
+#define OBD_FAIL_CHECK(id)                      CFS_FAIL_CHECK(id)
+#define OBD_FAIL_CHECK_VALUE(id, value)         CFS_FAIL_CHECK_VALUE(id, value)
+#define OBD_FAIL_CHECK_ORSET(id, value)         CFS_FAIL_CHECK_ORSET(id, value)
+#define OBD_FAIL_CHECK_RESET(id, value)         CFS_FAIL_CHECK_RESET(id, value)
+#define OBD_FAIL_RETURN(id, ret)                CFS_FAIL_RETURN(id, ret)
+#define OBD_FAIL_TIMEOUT(id, secs)              CFS_FAIL_TIMEOUT(id, secs)
+#define OBD_FAIL_TIMEOUT_MS(id, ms)             CFS_FAIL_TIMEOUT_MS(id, ms)
+#define OBD_FAIL_TIMEOUT_ORSET(id, value, secs) CFS_FAIL_TIMEOUT_ORSET(id, value, secs)
+#define OBD_RACE(id)                            CFS_RACE(id)
+#define OBD_FAIL_ONCE                           CFS_FAIL_ONCE
+#define OBD_FAILED                              CFS_FAILED
+
+extern cfs_atomic_t libcfs_kmemory;
 
 #ifdef LPROCFS
 #define obd_memory_add(size)                                                  \
@@ -586,7 +594,7 @@ do {                                                                          \
                 CERROR("vmalloc of '" #ptr "' (%d bytes) failed\n",           \
                        (int)(size));                                          \
                 CERROR(LPU64" total bytes allocated by Lustre, %d by LNET\n", \
-                       obd_memory_sum(), atomic_read(&libcfs_kmemory));       \
+                       obd_memory_sum(), cfs_atomic_read(&libcfs_kmemory));   \
         } else {                                                              \
                 memset(ptr, 0, size);                                         \
                 OBD_ALLOC_POST(ptr, size, "vmalloced");                       \
@@ -594,6 +602,37 @@ do {                                                                          \
 } while(0)
 #endif
 
+#ifdef __KERNEL__
+
+/* Allocations above this size are considered too big and could not be done
+ * atomically. 
+ *
+ * Be very careful when changing this value, especially when decreasing it,
+ * since vmalloc in Linux doesn't perform well on multi-cores system, calling 
+ * vmalloc in critical path would hurt peformance badly. See LU-66.
+ */
+#define OBD_ALLOC_BIG (4 * CFS_PAGE_SIZE)
+
+#define OBD_ALLOC_LARGE(ptr, size)                                            \
+do {                                                                          \
+        if (size > OBD_ALLOC_BIG)                                             \
+                OBD_VMALLOC(ptr, size);                                       \
+        else                                                                  \
+                OBD_ALLOC(ptr, size);                                         \
+} while (0)
+
+#define OBD_FREE_LARGE(ptr, size)                                             \
+do {                                                                          \
+        if (size > OBD_ALLOC_BIG)                                             \
+                OBD_VFREE(ptr, size);                                         \
+        else                                                                  \
+                OBD_FREE(ptr, size);                                          \
+} while (0)
+#else
+#define OBD_ALLOC_LARGE(ptr, size) OBD_ALLOC(ptr, size)
+#define OBD_FREE_LARGE(ptr, size) OBD_FREE(ptr,size)
+#endif
+
 #ifdef CONFIG_DEBUG_SLAB
 #define POISON(ptr, c, s) do {} while (0)
 #define POISON_PTR(ptr)  ((void)0)
@@ -668,7 +707,7 @@ do {                                                                          \
 })
 #define OBD_SLAB_ALLOC(ptr, slab, type, size)                                 \
 do {                                                                          \
-        LASSERT(!in_interrupt());                                             \
+        LASSERT(ergo(type != CFS_ALLOC_ATOMIC, !cfs_in_interrupt()));         \
         (ptr) = cfs_mem_cache_alloc(slab, (type));                            \
         if (likely((ptr) != NULL &&                                           \
                    (!HAS_FAIL_ALLOC_FLAG || obd_alloc_fail_rate == 0 ||       \
@@ -693,49 +732,46 @@ do {                                                                          \
         OBD_SLAB_ALLOC((ptr), (slab), CFS_ALLOC_STD, sizeof *(ptr))
 #define OBD_SLAB_FREE_PTR(ptr, slab)                                          \
         OBD_SLAB_FREE((ptr), (slab), sizeof *(ptr))
+#define OBD_SLAB_ALLOC_PTR_GFP(ptr, slab, gfp)                              \
+        OBD_SLAB_ALLOC((ptr), (slab), (gfp), sizeof *(ptr))
 
 #define KEY_IS(str) \
         (keylen >= (sizeof(str)-1) && memcmp(key, str, (sizeof(str)-1)) == 0)
 
 /* Wrapper for contiguous page frame allocation */
-#define OBD_PAGES_ALLOC(ptr, order, gfp_mask)                                 \
+#define OBD_PAGE_ALLOC(ptr, gfp_mask)                                         \
 do {                                                                          \
-        (ptr) = cfs_alloc_pages(gfp_mask, order);                             \
+        (ptr) = cfs_alloc_page(gfp_mask);                                     \
         if (unlikely((ptr) == NULL)) {                                        \
                 CERROR("alloc_pages of '" #ptr "' %d page(s) / "LPU64" bytes "\
-                       "failed\n", (int)(1 << (order)),                       \
-                       (__u64)((1 << (order)) << CFS_PAGE_SHIFT));            \
+                       "failed\n", (int)1,                                    \
+                       (__u64)(1 << CFS_PAGE_SHIFT));                         \
                 CERROR(LPU64" total bytes and "LPU64" total pages "           \
                        "("LPU64" bytes) allocated by Lustre, "                \
                        "%d total bytes by LNET\n",                            \
                        obd_memory_sum(),                                      \
                        obd_pages_sum() << CFS_PAGE_SHIFT,                     \
                        obd_pages_sum(),                                       \
-                       atomic_read(&libcfs_kmemory));                         \
+                       cfs_atomic_read(&libcfs_kmemory));                     \
         } else {                                                              \
-                obd_pages_add(order);                                         \
+                obd_pages_add(0);                                             \
                 CDEBUG(D_MALLOC, "alloc_pages '" #ptr "': %d page(s) / "      \
                        LPU64" bytes at %p.\n",                                \
-                       (int)(1 << (order)),                                   \
-                       (__u64)((1 << (order)) << CFS_PAGE_SHIFT), ptr);       \
+                       (int)1,                                                \
+                       (__u64)(1 << CFS_PAGE_SHIFT), ptr);                    \
         }                                                                     \
 } while (0)
 
-#define OBD_PAGE_ALLOC(ptr, gfp_mask)                                         \
-        OBD_PAGES_ALLOC(ptr, 0, gfp_mask)
-
-#define OBD_PAGES_FREE(ptr, order)                                            \
+#define OBD_PAGE_FREE(ptr)                                                    \
 do {                                                                          \
         LASSERT(ptr);                                                         \
-        obd_pages_sub(order);                                                 \
+        obd_pages_sub(0);                                                     \
         CDEBUG(D_MALLOC, "free_pages '" #ptr "': %d page(s) / "LPU64" bytes " \
                "at %p.\n",                                                    \
-               (int)(1 << (order)), (__u64)((1 << (order)) << CFS_PAGE_SHIFT),\
+               (int)1, (__u64)(1 << CFS_PAGE_SHIFT),                          \
                ptr);                                                          \
-        __cfs_free_pages(ptr, order);                                         \
+        cfs_free_page(ptr);                                                   \
         (ptr) = (void *)0xdeadbeef;                                           \
 } while (0)
 
-#define OBD_PAGE_FREE(ptr) OBD_PAGES_FREE(ptr, 0)
-
 #endif