Whamcloud - gitweb
LU-1422 cray: Cray platforms set the initial connection value to high
[fs/lustre-release.git] / lustre / include / obd_support.h
index 4661626..3fe05d4 100644 (file)
@@ -1,6 +1,4 @@
-/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
- * vim:expandtab:shiftwidth=8:tabstop=8:
- *
+/*
  * GPL HEADER START
  *
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  * 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 +59,6 @@ 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;
@@ -81,44 +73,56 @@ 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 atomic_t obd_dirty_transit_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);
+extern char obd_jobid_var[];
 
 /* 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
+#define HASH_JOB_STATS_BKT_BITS 5
+#define HASH_JOB_STATS_CUR_BITS 7
+#define HASH_JOB_STATS_MAX_BITS 12
 
 /* 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,
@@ -132,21 +136,20 @@ int obd_alloc_fail(const void *ptr, const char *name, const char *type,
     miss the recovery window */
 #define CONNECTION_SWITCH_MAX min(50U, max(CONNECTION_SWITCH_MIN,obd_timeout))
 #define CONNECTION_SWITCH_INC 5  /* Connection timeout backoff */
-#ifndef CRAY_XT3
 /* In general this should be low to have quick detection of a system
    running on a backup server. (If it's too low, import_select_connection
    will increase the timeout anyhow.)  */
 #define INITIAL_CONNECT_TIMEOUT max(CONNECTION_SWITCH_MIN,obd_timeout/20)
-#else
-/* ...but for very large systems (e.g. CRAY) we need to keep the initial
-   connect t.o. high (bz 10803), because they will nearly ALWAYS be doing the
-   connects for the first time (clients "reboot" after every process, so no
-   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 */
 
 /**
@@ -222,8 +225,11 @@ int obd_alloc_fail(const void *ptr, const char *name, const char *type,
 #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_OPEN_WAIT_CREATE    0x143
-#define OBD_FAIL_MDS_READLINK_EPROTO     0x144
+#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
@@ -233,6 +239,7 @@ int obd_alloc_fail(const void *ptr, const char *name, const char *type,
 #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_MDS_DQACQ_NET           0x187
 
 #define OBD_FAIL_OST                     0x200
 #define OBD_FAIL_OST_CONNECT_NET         0x201
@@ -273,6 +280,10 @@ int obd_alloc_fail(const void *ptr, const char *name, const char *type,
 #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_OST_DQACQ_NET           0x230
 
 #define OBD_FAIL_LDLM                    0x300
 #define OBD_FAIL_LDLM_NAMESPACE_NEW      0x301
@@ -299,6 +310,9 @@ int obd_alloc_fail(const void *ptr, const char *name, const char *type,
 #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
@@ -319,6 +333,8 @@ int obd_alloc_fail(const void *ptr, const char *name, const char *type,
 #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
@@ -338,6 +354,8 @@ int obd_alloc_fail(const void *ptr, const char *name, const char *type,
 #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
@@ -358,6 +376,8 @@ int obd_alloc_fail(const void *ptr, const char *name, const char *type,
 #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
@@ -401,122 +421,28 @@ int obd_alloc_fail(const void *ptr, const char *name, const char *type,
 #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
-
-
-/* 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_MDS_SYNC_CAPA_SL                   0x1310
+#define OBD_FAIL_SEQ_ALLOC                          0x1311
+
+#define OBD_FAIL_LLITE                              0x1400
+#define OBD_FAIL_LLITE_FAULT_TRUNC_RACE             0x1401
+#define OBD_FAIL_LOCK_STATE_WAIT_INTR               0x1402
+
+/* 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)                                                  \
@@ -613,18 +539,21 @@ static inline void obd_pages_sub(int order)
 #define OBD_ALLOC_FAIL_MULT (OBD_ALLOC_FAIL_MASK / 100)
 
 #if defined(LUSTRE_UTILS) /* this version is for utils only */
-#define OBD_ALLOC_GFP(ptr, size, gfp_mask)                                    \
-do {                                                                          \
-        (ptr) = cfs_alloc(size, (gfp_mask));                                  \
-        if (unlikely((ptr) == NULL)) {                                        \
-                CERROR("kmalloc of '" #ptr "' (%d bytes) failed at %s:%d\n",  \
-                       (int)(size), __FILE__, __LINE__);                      \
-        } else {                                                              \
-                memset(ptr, 0, size);                                         \
-                CDEBUG(D_MALLOC, "kmalloced '" #ptr "': %d at %p\n",          \
-                       (int)(size), ptr);                                     \
-        }                                                                     \
+#define __OBD_MALLOC_VERBOSE(ptr, cptab, cpt, size, flags)                   \
+do {                                                                         \
+       (ptr) = (cptab) == NULL ?                                             \
+               cfs_alloc(size, flags) :                                      \
+               cfs_cpt_malloc(cptab, cpt, size, flags);                      \
+       if (unlikely((ptr) == NULL)) {                                        \
+               CERROR("kmalloc of '" #ptr "' (%d bytes) failed at %s:%d\n",  \
+                      (int)(size), __FILE__, __LINE__);                      \
+       } else {                                                              \
+               memset(ptr, 0, size);                                         \
+               CDEBUG(D_MALLOC, "kmalloced '" #ptr "': %d at %p\n",          \
+                      (int)(size), ptr);                                     \
+       }                                                                     \
 } while (0)
+
 #else /* this version is for the kernel and liblustre */
 #define OBD_FREE_RTN0(ptr)                                                    \
 ({                                                                            \
@@ -632,9 +561,12 @@ do {                                                                          \
         (ptr) = NULL;                                                         \
         0;                                                                    \
 })
-#define OBD_ALLOC_GFP(ptr, size, gfp_mask)                                    \
-do {                                                                          \
-        (ptr) = cfs_alloc(size, (gfp_mask));                                  \
+
+#define __OBD_MALLOC_VERBOSE(ptr, cptab, cpt, size, flags)                   \
+do {                                                                         \
+       (ptr) = (cptab) == NULL ?                                             \
+               cfs_alloc(size, flags) :                                      \
+               cfs_cpt_malloc(cptab, cpt, size, flags);                      \
         if (likely((ptr) != NULL &&                                           \
                    (!HAS_FAIL_ALLOC_FLAG || obd_alloc_fail_rate == 0 ||       \
                     !obd_alloc_fail(ptr, #ptr, "km", size,                    \
@@ -646,33 +578,101 @@ do {                                                                          \
 } while (0)
 #endif
 
-#ifndef OBD_ALLOC_MASK
-# define OBD_ALLOC_MASK CFS_ALLOC_IO
-#endif
+#define OBD_ALLOC_GFP(ptr, size, gfp_mask)                                   \
+       __OBD_MALLOC_VERBOSE(ptr, NULL, 0, size, gfp_mask)
 
-#define OBD_ALLOC(ptr, size) OBD_ALLOC_GFP(ptr, size, OBD_ALLOC_MASK)
+#define OBD_ALLOC(ptr, size) OBD_ALLOC_GFP(ptr, size, CFS_ALLOC_IO)
 #define OBD_ALLOC_WAIT(ptr, size) OBD_ALLOC_GFP(ptr, size, CFS_ALLOC_STD)
 #define OBD_ALLOC_PTR(ptr) OBD_ALLOC(ptr, sizeof *(ptr))
 #define OBD_ALLOC_PTR_WAIT(ptr) OBD_ALLOC_WAIT(ptr, sizeof *(ptr))
 
+#define OBD_CPT_ALLOC_GFP(ptr, cptab, cpt, size, gfp_mask)                   \
+       __OBD_MALLOC_VERBOSE(ptr, cptab, cpt, size, gfp_mask)
+
+#define OBD_CPT_ALLOC(ptr, cptab, cpt, size)                                 \
+       OBD_CPT_ALLOC_GFP(ptr, cptab, cpt, size, CFS_ALLOC_IO)
+
+#define OBD_CPT_ALLOC_PTR(ptr, cptab, cpt)                                   \
+       OBD_CPT_ALLOC(ptr, cptab, cpt, sizeof *(ptr))
+
 #ifdef __arch_um__
-# define OBD_VMALLOC(ptr, size) OBD_ALLOC(ptr, size)
-#else
-# define OBD_VMALLOC(ptr, size)                                               \
-do {                                                                          \
-        (ptr) = cfs_alloc_large(size);                                        \
+
+# define OBD_VMALLOC(ptr, size)                                                      \
+        OBD_ALLOC(ptr, size)
+# define OBD_CPT_VMALLOC(ptr, cptab, cpt, size)                                      \
+        OBD_CPT_ALLOC(ptr, cptab, cpt, size)
+
+#else /* !__arch_um__ */
+
+# define __OBD_VMALLOC_VEROBSE(ptr, cptab, cpt, size)                        \
+do {                                                                         \
+       (ptr) = cptab == NULL ?                                               \
+               cfs_alloc_large(size) :                                       \
+               cfs_cpt_vmalloc(cptab, cpt, size);                            \
         if (unlikely((ptr) == NULL)) {                                        \
                 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);                                         \
+               memset(ptr, 0, size);                                         \
                 OBD_ALLOC_POST(ptr, size, "vmalloced");                       \
         }                                                                     \
 } while(0)
+
+# define OBD_VMALLOC(ptr, size)                                                      \
+        __OBD_VMALLOC_VEROBSE(ptr, NULL, 0, size)
+# define OBD_CPT_VMALLOC(ptr, cptab, cpt, size)                                      \
+        __OBD_VMALLOC_VEROBSE(ptr, cptab, cpt, size)
+
 #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_CPT_ALLOC_LARGE(ptr, cptab, cpt, size)                           \
+do {                                                                         \
+       if (size > OBD_ALLOC_BIG)                                             \
+               OBD_CPT_VMALLOC(ptr, cptab, cpt, size);                       \
+       else                                                                  \
+               OBD_CPT_ALLOC(ptr, cptab, cpt, 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 /* !__KERNEL__ */
+
+#define OBD_ALLOC_LARGE(ptr, size)                                           \
+       OBD_ALLOC(ptr, size)
+#define OBD_CPT_ALLOC_LARGE(ptr, cptab, cpt, size)                           \
+       OBD_ALLOC(ptr, size)
+#define OBD_FREE_LARGE(ptr, size)                                            \
+       OBD_FREE(ptr, size)
+
+#endif /* __KERNEL__ */
+
 #ifdef CONFIG_DEBUG_SLAB
 #define POISON(ptr, c, s) do {} while (0)
 #define POISON_PTR(ptr)  ((void)0)
@@ -707,22 +707,20 @@ do {                                                                          \
 # define my_call_rcu(rcu, cb)             (cb)(rcu)
 #endif
 
-#define OBD_FREE_RCU_CB(ptr, size, handle, free_cb)                           \
-do {                                                                          \
-        struct portals_handle *__h = (handle);                                \
-        LASSERT(handle);                                                      \
-        __h->h_ptr = (ptr);                                                   \
-        __h->h_size = (size);                                                 \
-        __h->h_free_cb = (void (*)(void *, size_t))(free_cb);                 \
-        my_call_rcu(&__h->h_rcu, class_handle_free_cb);                       \
-        POISON_PTR(ptr);                                                      \
+#define OBD_FREE_RCU(ptr, size, handle)                                              \
+do {                                                                         \
+       struct portals_handle *__h = (handle);                                \
+                                                                             \
+       LASSERT(handle != NULL);                                              \
+       __h->h_cookie = (unsigned long)(ptr);                                 \
+       __h->h_size = (size);                                                 \
+       my_call_rcu(&__h->h_rcu, class_handle_free_cb);                       \
+       POISON_PTR(ptr);                                                      \
 } while(0)
-#define OBD_FREE_RCU(ptr, size, handle) OBD_FREE_RCU_CB(ptr, size, handle, NULL)
 
 #else
 #define OBD_FREE(ptr, size) ((void)(size), free((ptr)))
 #define OBD_FREE_RCU(ptr, size, handle) (OBD_FREE(ptr, size))
-#define OBD_FREE_RCU_CB(ptr, size, handle, cb)     ((*(cb))(ptr, size))
 #endif /* ifdef __KERNEL__ */
 
 #ifdef __arch_um__
@@ -745,10 +743,13 @@ do {                                                                          \
         (ptr) = NULL;                                                         \
         0;                                                                    \
 })
-#define OBD_SLAB_ALLOC(ptr, slab, type, size)                                 \
-do {                                                                          \
-        LASSERT(ergo(type != CFS_ALLOC_ATOMIC, !in_interrupt()));             \
-        (ptr) = cfs_mem_cache_alloc(slab, (type));                            \
+
+#define __OBD_SLAB_ALLOC_VERBOSE(ptr, slab, cptab, cpt, size, type)          \
+do {                                                                         \
+        LASSERT(ergo(type != CFS_ALLOC_ATOMIC, !cfs_in_interrupt()));         \
+       (ptr) = (cptab) == NULL ?                                             \
+               cfs_mem_cache_alloc(slab, type) :                             \
+               cfs_mem_cache_cpt_alloc(slab, cptab, cpt, type);              \
         if (likely((ptr) != NULL &&                                           \
                    (!HAS_FAIL_ALLOC_FLAG || obd_alloc_fail_rate == 0 ||       \
                     !obd_alloc_fail(ptr, #ptr, "slab-", size,                 \
@@ -759,6 +760,11 @@ do {                                                                          \
         }                                                                     \
 } while(0)
 
+#define OBD_SLAB_ALLOC_GFP(ptr, slab, size, flags)                           \
+       __OBD_SLAB_ALLOC_VERBOSE(ptr, slab, NULL, 0, size, flags)
+#define OBD_SLAB_CPT_ALLOC_GFP(ptr, slab, cptab, cpt, size, flags)           \
+       __OBD_SLAB_ALLOC_VERBOSE(ptr, slab, cptab, cpt, size, flags)
+
 #define OBD_FREE_PTR(ptr) OBD_FREE(ptr, sizeof *(ptr))
 
 #define OBD_SLAB_FREE(ptr, slab, size)                                        \
@@ -768,55 +774,71 @@ do {                                                                          \
         POISON_PTR(ptr);                                                      \
 } while(0)
 
-#define OBD_SLAB_ALLOC_PTR(ptr, slab)                                         \
-        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 OBD_SLAB_ALLOC(ptr, slab, size)                                              \
+       OBD_SLAB_ALLOC_GFP(ptr, slab, size, CFS_ALLOC_IO)
+
+#define OBD_SLAB_CPT_ALLOC(ptr, slab, cptab, cpt, size)                              \
+       OBD_SLAB_CPT_ALLOC_GFP(ptr, slab, cptab, cpt, size, CFS_ALLOC_IO)
+
+#define OBD_SLAB_ALLOC_PTR(ptr, slab)                                        \
+       OBD_SLAB_ALLOC(ptr, slab, sizeof *(ptr))
+
+#define OBD_SLAB_CPT_ALLOC_PTR(ptr, slab, cptab, cpt)                        \
+       OBD_SLAB_CPT_ALLOC(ptr, slab, cptab, cpt, sizeof *(ptr))
+
+#define OBD_SLAB_ALLOC_PTR_GFP(ptr, slab, flags)                             \
+       OBD_SLAB_ALLOC_GFP(ptr, slab, sizeof *(ptr), flags)
+
+#define OBD_SLAB_CPT_ALLOC_PTR_GFP(ptr, slab, ctab, cpt, flags)                      \
+       OBD_SLAB_CPT_ALLOC_GFP(ptr, slab, cptab, cpt, sizeof *(ptr), flags)
+
+#define OBD_SLAB_FREE_PTR(ptr, slab)                                         \
+       OBD_SLAB_FREE((ptr), (slab), 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)                                 \
-do {                                                                          \
-        (ptr) = cfs_alloc_pages(gfp_mask, order);                             \
+#define __OBD_PAGE_ALLOC_VERBOSE(ptr, cptab, cpt, gfp_mask)                  \
+do {                                                                         \
+       (ptr) = (cptab) == NULL ?                                             \
+               cfs_alloc_page(gfp_mask) :                                    \
+               cfs_page_cpt_alloc(cptab, cpt, 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_PAGE_ALLOC(ptr, gfp_mask)                                        \
+       __OBD_PAGE_ALLOC_VERBOSE(ptr, NULL, 0, gfp_mask)
+#define OBD_PAGE_CPT_ALLOC(ptr, cptab, cpt, gfp_mask)                        \
+       __OBD_PAGE_ALLOC_VERBOSE(ptr, cptab, cpt, 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