Whamcloud - gitweb
Branch HEAD
[fs/lustre-release.git] / lustre / ptlrpc / sec.c
index 5c24c6c..977ac89 100644 (file)
@@ -1,22 +1,41 @@
 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
  * vim:expandtab:shiftwidth=8:tabstop=8:
  *
- * Copyright (C) 2004-2006 Cluster File Systems, Inc.
+ * GPL HEADER START
  *
- *   This file is part of Lustre, http://www.lustre.org.
+ * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
- *   Lustre is free software; you can redistribute it and/or
- *   modify it under the terms of version 2 of the GNU General Public
- *   License as published by the Free Software Foundation.
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 only,
+ * as published by the Free Software Foundation.
  *
- *   Lustre is distributed in the hope that it will be useful,
- *   but WITHOUT ANY WARRANTY; without even the implied warranty of
- *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
- *   GNU General Public License for more details.
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License version 2 for more details (a copy is included
+ * in the LICENSE file that accompanied this code).
  *
- *   You should have received a copy of the GNU General Public License
- *   along with Lustre; if not, write to the Free Software
- *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ * You should have received a copy of the GNU General Public License
+ * 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
+ * have any questions.
+ *
+ * GPL HEADER END
+ */
+/*
+ * Copyright  2008 Sun Microsystems, Inc. All rights reserved
+ * Use is subject to license terms.
+ */
+/*
+ * This file is part of Lustre, http://www.lustre.org/
+ * Lustre is a trademark of Sun Microsystems, Inc.
+ *
+ * lustre/ptlrpc/sec.c
+ *
+ * Author: Eric Mei <ericm@clusterfs.com>
  */
 
 #ifndef EXPORT_SYMTAB
@@ -30,6 +49,7 @@
 #include <libcfs/list.h>
 #else
 #include <linux/crypto.h>
+#include <linux/key.h>
 #endif
 
 #include <obd.h>
 
 #include "ptlrpc_internal.h"
 
-static void sptlrpc_sec_destroy(struct ptlrpc_sec *sec);
-static int sptlrpc_sec_destroy_ctx(struct ptlrpc_sec *sec,
-                                   struct ptlrpc_cli_ctx *ctx);
-static void sptlrpc_ctx_refresh(struct ptlrpc_cli_ctx *ctx);
-
 /***********************************************
  * policy registers                            *
  ***********************************************/
 
-static rwlock_t policy_lock = RW_LOCK_UNLOCKED;
+static rwlock_t policy_lock;
 static struct ptlrpc_sec_policy *policies[SPTLRPC_POLICY_MAX] = {
         NULL,
 };
 
 int sptlrpc_register_policy(struct ptlrpc_sec_policy *policy)
 {
-        __u32 number = policy->sp_policy;
+        __u16 number = policy->sp_policy;
 
         LASSERT(policy->sp_name);
         LASSERT(policy->sp_cops);
@@ -82,7 +97,7 @@ EXPORT_SYMBOL(sptlrpc_register_policy);
 
 int sptlrpc_unregister_policy(struct ptlrpc_sec_policy *policy)
 {
-        __u32 number = policy->sp_policy;
+        __u16 number = policy->sp_policy;
 
         LASSERT(number < SPTLRPC_POLICY_MAX);
 
@@ -103,55 +118,55 @@ int sptlrpc_unregister_policy(struct ptlrpc_sec_policy *policy)
 EXPORT_SYMBOL(sptlrpc_unregister_policy);
 
 static
-struct ptlrpc_sec_policy * sptlrpc_flavor2policy(ptlrpc_sec_flavor_t flavor)
+struct ptlrpc_sec_policy * sptlrpc_rpcflavor2policy(__u16 flavor)
 {
         static DECLARE_MUTEX(load_mutex);
-        static atomic_t         loaded = ATOMIC_INIT(0);
-        struct                  ptlrpc_sec_policy *policy;
-        __u32                   number = SEC_FLAVOR_POLICY(flavor), flag = 0;
+        static atomic_t           loaded = ATOMIC_INIT(0);
+        struct ptlrpc_sec_policy *policy;
+        __u16                     number = RPC_FLVR_POLICY(flavor), flag = 0;
 
         if (number >= SPTLRPC_POLICY_MAX)
                 return NULL;
 
-again:
-        read_lock(&policy_lock);
-        policy = policies[number];
-        if (policy && !try_module_get(policy->sp_owner))
-                policy = NULL;
-        if (policy == NULL)
-                flag = atomic_read(&loaded);
-        read_unlock(&policy_lock);
-
-#ifdef CONFIG_KMOD
-        /* if failure, try to load gss module, once */
-        if (unlikely(policy == NULL) &&
-            number == SPTLRPC_POLICY_GSS && flag == 0) {
+        while (1) {
+                read_lock(&policy_lock);
+                policy = policies[number];
+                if (policy && !try_module_get(policy->sp_owner))
+                        policy = NULL;
+                if (policy == NULL)
+                        flag = atomic_read(&loaded);
+                read_unlock(&policy_lock);
+
+                if (policy != NULL || flag != 0 ||
+                    number != SPTLRPC_POLICY_GSS)
+                        break;
+
+                /* try to load gss module, once */
                 mutex_down(&load_mutex);
                 if (atomic_read(&loaded) == 0) {
-                        if (request_module("ptlrpc_gss") != 0)
-                                CERROR("Unable to load module ptlrpc_gss\n");
+                        if (request_module("ptlrpc_gss") == 0)
+                                CWARN("module ptlrpc_gss loaded on demand\n");
                         else
-                                CWARN("module ptlrpc_gss loaded\n");
+                                CERROR("Unable to load module ptlrpc_gss\n");
 
                         atomic_set(&loaded, 1);
                 }
                 mutex_up(&load_mutex);
-
-                goto again;
         }
-#endif
 
         return policy;
 }
 
-ptlrpc_sec_flavor_t sptlrpc_name2flavor(const char *name)
+__u16 sptlrpc_name2rpcflavor(const char *name)
 {
         if (!strcmp(name, "null"))
                 return SPTLRPC_FLVR_NULL;
         if (!strcmp(name, "plain"))
                 return SPTLRPC_FLVR_PLAIN;
-        if (!strcmp(name, "krb5"))
-                return SPTLRPC_FLVR_KRB5;
+        if (!strcmp(name, "krb5n"))
+                return SPTLRPC_FLVR_KRB5N;
+        if (!strcmp(name, "krb5a"))
+                return SPTLRPC_FLVR_KRB5A;
         if (!strcmp(name, "krb5i"))
                 return SPTLRPC_FLVR_KRB5I;
         if (!strcmp(name, "krb5p"))
@@ -159,388 +174,69 @@ ptlrpc_sec_flavor_t sptlrpc_name2flavor(const char *name)
 
         return SPTLRPC_FLVR_INVALID;
 }
-EXPORT_SYMBOL(sptlrpc_name2flavor);
+EXPORT_SYMBOL(sptlrpc_name2rpcflavor);
 
-char *sptlrpc_flavor2name(ptlrpc_sec_flavor_t flavor)
+const char *sptlrpc_rpcflavor2name(__u16 flavor)
 {
         switch (flavor) {
         case SPTLRPC_FLVR_NULL:
                 return "null";
         case SPTLRPC_FLVR_PLAIN:
                 return "plain";
-        case SPTLRPC_FLVR_KRB5:
-                return "krb5";
+        case SPTLRPC_FLVR_KRB5N:
+                return "krb5n";
+        case SPTLRPC_FLVR_KRB5A:
+                return "krb5a";
         case SPTLRPC_FLVR_KRB5I:
                 return "krb5i";
         case SPTLRPC_FLVR_KRB5P:
                 return "krb5p";
         default:
-                CERROR("invalid flavor 0x%x(p%u,s%u,v%u)\n", flavor,
-                       SEC_FLAVOR_POLICY(flavor), SEC_FLAVOR_SUBPOLICY(flavor),
-                       SEC_FLAVOR_SVC(flavor));
-        }
-        return "UNKNOWN";
-}
-EXPORT_SYMBOL(sptlrpc_flavor2name);
-
-/***********************************************
- * context helpers                             *
- * internal APIs                               *
- * cache management                            *
- ***********************************************/
-
-static inline
-unsigned long ctx_status(struct ptlrpc_cli_ctx *ctx)
-{
-        smp_mb();
-        return (ctx->cc_flags & PTLRPC_CTX_STATUS_MASK);
-}
-
-static inline
-int ctx_is_uptodate(struct ptlrpc_cli_ctx *ctx)
-{
-        return (ctx_status(ctx) == PTLRPC_CTX_UPTODATE);
-}
-
-static inline
-int ctx_is_refreshed(struct ptlrpc_cli_ctx *ctx)
-{
-        return (ctx_status(ctx) != 0);
-}
-
-static inline
-int ctx_is_dead(struct ptlrpc_cli_ctx *ctx)
-{
-        smp_mb();
-        return ((ctx->cc_flags & (PTLRPC_CTX_DEAD | PTLRPC_CTX_ERROR)) != 0);
-}
-
-static inline
-int ctx_is_eternal(struct ptlrpc_cli_ctx *ctx)
-{
-        smp_mb();
-        return ((ctx->cc_flags & PTLRPC_CTX_ETERNAL) != 0);
-}
-
-static
-int ctx_expire(struct ptlrpc_cli_ctx *ctx)
-{
-        LASSERT(atomic_read(&ctx->cc_refcount));
-
-        if (!test_and_set_bit(PTLRPC_CTX_DEAD_BIT, &ctx->cc_flags)) {
-                cfs_time_t now = cfs_time_current_sec();
-
-                smp_mb();
-                clear_bit(PTLRPC_CTX_UPTODATE_BIT, &ctx->cc_flags);
-
-                if (ctx->cc_expire && cfs_time_aftereq(now, ctx->cc_expire))
-                        CWARN("ctx %p(%u->%s): get expired (%lds exceeds)\n",
-                              ctx, ctx->cc_vcred.vc_uid,
-                              sec2target_str(ctx->cc_sec),
-                              cfs_time_sub(now, ctx->cc_expire));
-                else
-                        CWARN("ctx %p(%u->%s): force to die (%lds remains)\n",
-                              ctx, ctx->cc_vcred.vc_uid,
-                              sec2target_str(ctx->cc_sec),
-                              ctx->cc_expire == 0 ? 0 :
-                              cfs_time_sub(ctx->cc_expire, now));
-
-                return 1;
+                CERROR("invalid rpc flavor 0x%x(p%u,s%u,v%u)\n", flavor,
+                       RPC_FLVR_POLICY(flavor), RPC_FLVR_MECH(flavor),
+                       RPC_FLVR_SVC(flavor));
         }
-        return 0;
-}
-
-static
-void ctx_enhash(struct ptlrpc_cli_ctx *ctx, struct hlist_head *hash)
-{
-        set_bit(PTLRPC_CTX_HASHED_BIT, &ctx->cc_flags);
-        atomic_inc(&ctx->cc_refcount);
-        hlist_add_head(&ctx->cc_hash, hash);
-}
-
-static
-void ctx_unhash(struct ptlrpc_cli_ctx *ctx, struct hlist_head *freelist)
-{
-        LASSERT_SPIN_LOCKED(&ctx->cc_sec->ps_lock);
-        LASSERT(atomic_read(&ctx->cc_refcount) > 0);
-        LASSERT(test_bit(PTLRPC_CTX_HASHED_BIT, &ctx->cc_flags));
-        LASSERT(!hlist_unhashed(&ctx->cc_hash));
-
-        clear_bit(PTLRPC_CTX_HASHED_BIT, &ctx->cc_flags);
-
-        if (atomic_dec_and_test(&ctx->cc_refcount)) {
-                __hlist_del(&ctx->cc_hash);
-                hlist_add_head(&ctx->cc_hash, freelist);
-        } else
-                hlist_del_init(&ctx->cc_hash);
+        return "unknown";
 }
+EXPORT_SYMBOL(sptlrpc_rpcflavor2name);
 
-/*
- * return 1 if the context is dead.
- */
-static
-int ctx_check_death(struct ptlrpc_cli_ctx *ctx, struct hlist_head *freelist)
+int sptlrpc_flavor2name(struct sptlrpc_flavor *sf, char *buf, int bufsize)
 {
-        if (unlikely(ctx_is_dead(ctx)))
-                goto unhash;
+        char           *bulk;
 
-        /* expire is 0 means never expire. a newly created gss context
-         * which during upcall also has 0 expiration
-         */
-        smp_mb();
-        if (ctx->cc_expire == 0)
-                return 0;
-
-        /* check real expiration */
-        smp_mb();
-        if (cfs_time_after(ctx->cc_expire, cfs_time_current_sec()))
-                return 0;
-
-        ctx_expire(ctx);
-
-unhash:
-        if (freelist)
-                ctx_unhash(ctx, freelist);
-
-        return 1;
-}
-
-static inline
-int ctx_check_death_locked(struct ptlrpc_cli_ctx *ctx,
-                           struct hlist_head *freelist)
-{
-        LASSERT(ctx->cc_sec);
-        LASSERT(atomic_read(&ctx->cc_refcount) > 0);
-        LASSERT_SPIN_LOCKED(&ctx->cc_sec->ps_lock);
-        LASSERT(test_bit(PTLRPC_CTX_HASHED_BIT, &ctx->cc_flags));
-
-        return ctx_check_death(ctx, freelist);
-}
-
-static
-int ctx_check_uptodate(struct ptlrpc_cli_ctx *ctx)
-{
-        LASSERT(ctx->cc_sec);
-        LASSERT(atomic_read(&ctx->cc_refcount) > 0);
+        if (sf->sf_bulk_ciph != BULK_CIPH_ALG_NULL)
+                bulk = "bulkp";
+        else if (sf->sf_bulk_hash != BULK_HASH_ALG_NULL)
+                bulk = "bulki";
+        else
+                bulk = "bulkn";
 
-        if (!ctx_check_death(ctx, NULL) && ctx_is_uptodate(ctx))
-                return 1;
+        snprintf(buf, bufsize, "%s-%s:%s/%s",
+                 sptlrpc_rpcflavor2name(sf->sf_rpc), bulk,
+                 sptlrpc_get_hash_name(sf->sf_bulk_hash),
+                 sptlrpc_get_ciph_name(sf->sf_bulk_ciph));
         return 0;
 }
+EXPORT_SYMBOL(sptlrpc_flavor2name);
 
-static inline
-int ctx_match(struct ptlrpc_cli_ctx *ctx, struct vfs_cred *vcred)
-{
-        /* a little bit optimization for null policy */
-        if (!ctx->cc_ops->match)
-                return 1;
-
-        return ctx->cc_ops->match(ctx, vcred);
-}
-
-static
-void ctx_list_destroy(struct hlist_head *head)
-{
-        struct ptlrpc_cli_ctx *ctx;
-
-        while (!hlist_empty(head)) {
-                ctx = hlist_entry(head->first, struct ptlrpc_cli_ctx, cc_hash);
-
-                LASSERT(atomic_read(&ctx->cc_refcount) == 0);
-                LASSERT(test_bit(PTLRPC_CTX_HASHED_BIT, &ctx->cc_flags) == 0);
-
-                hlist_del_init(&ctx->cc_hash);
-                sptlrpc_sec_destroy_ctx(ctx->cc_sec, ctx);
-        }
-}
-
-static
-void ctx_cache_gc(struct ptlrpc_sec *sec, struct hlist_head *freelist)
-{
-        struct ptlrpc_cli_ctx *ctx;
-        struct hlist_node *pos, *next;
-        int i;
-        ENTRY;
-
-        CDEBUG(D_SEC, "do gc on sec %s@%p\n", sec->ps_policy->sp_name, sec);
-
-        for (i = 0; i < sec->ps_ccache_size; i++) {
-                hlist_for_each_entry_safe(ctx, pos, next,
-                                          &sec->ps_ccache[i], cc_hash)
-                        ctx_check_death_locked(ctx, freelist);
-        }
-
-        sec->ps_gc_next = cfs_time_current_sec() + sec->ps_gc_interval;
-        EXIT;
-}
-
-/*
- * @uid: which user. "-1" means flush all.
- * @grace: mark context DEAD, allow graceful destroy like notify
- *         server side, etc.
- * @force: also flush busy entries.
- *
- * return the number of busy context encountered.
- *
- * In any cases, never touch "eternal" contexts.
- */
-static
-int ctx_cache_flush(struct ptlrpc_sec *sec, uid_t uid, int grace, int force)
-{
-        struct ptlrpc_cli_ctx *ctx;
-        struct hlist_node *pos, *next;
-        HLIST_HEAD(freelist);
-        int i, busy = 0;
-        ENTRY;
-
-        might_sleep_if(grace);
-
-        spin_lock(&sec->ps_lock);
-        for (i = 0; i < sec->ps_ccache_size; i++) {
-                hlist_for_each_entry_safe(ctx, pos, next,
-                                          &sec->ps_ccache[i], cc_hash) {
-                        LASSERT(atomic_read(&ctx->cc_refcount) > 0);
-
-                        if (ctx_is_eternal(ctx))
-                                continue;
-                        if (uid != -1 && uid != ctx->cc_vcred.vc_uid)
-                                continue;
-
-                        if (atomic_read(&ctx->cc_refcount) > 1) {
-                                busy++;
-                                if (!force)
-                                        continue;
-
-                                CWARN("flush busy(%d) ctx %p(%u->%s) by force, "
-                                      "grace %d\n",
-                                      atomic_read(&ctx->cc_refcount),
-                                      ctx, ctx->cc_vcred.vc_uid,
-                                      sec2target_str(ctx->cc_sec), grace);
-                        }
-                        ctx_unhash(ctx, &freelist);
-
-                        set_bit(PTLRPC_CTX_DEAD_BIT, &ctx->cc_flags);
-                        if (!grace)
-                                clear_bit(PTLRPC_CTX_UPTODATE_BIT,
-                                          &ctx->cc_flags);
-                }
-        }
-        spin_unlock(&sec->ps_lock);
-
-        ctx_list_destroy(&freelist);
-        RETURN(busy);
-}
-
-static inline
-unsigned int ctx_hash_index(struct ptlrpc_sec *sec, __u64 key)
-{
-        return (unsigned int) (key & (sec->ps_ccache_size - 1));
-}
+/**************************************************
+ * client context APIs                            *
+ **************************************************/
 
-/*
- * return matched context. If it's a newly created one, we also give the
- * first push to refresh. return NULL if error happens.
- */
 static
-struct ptlrpc_cli_ctx * ctx_cache_lookup(struct ptlrpc_sec *sec,
-                                         struct vfs_cred *vcred,
-                                         int create, int remove_dead)
-{
-        struct ptlrpc_cli_ctx *ctx = NULL, *new = NULL;
-        struct hlist_head *hash_head;
-        struct hlist_node *pos, *next;
-        HLIST_HEAD(freelist);
-        unsigned int hash, gc = 0, found = 0;
-        ENTRY;
-
-        might_sleep();
-
-        hash = ctx_hash_index(sec, (__u64) vcred->vc_uid);
-        LASSERT(hash < sec->ps_ccache_size);
-        hash_head = &sec->ps_ccache[hash];
-
-retry:
-        spin_lock(&sec->ps_lock);
-
-        /* gc_next == 0 means never do gc */
-        if (remove_dead && sec->ps_gc_next &&
-            cfs_time_after(cfs_time_current_sec(), sec->ps_gc_next)) {
-                ctx_cache_gc(sec, &freelist);
-                gc = 1;
-        }
-
-        hlist_for_each_entry_safe(ctx, pos, next, hash_head, cc_hash) {
-                if (gc == 0 &&
-                    ctx_check_death_locked(ctx, remove_dead ? &freelist : NULL))
-                        continue;
-
-                if (ctx_match(ctx, vcred)) {
-                        found = 1;
-                        break;
-                }
-        }
-
-        if (found) {
-                if (new && new != ctx) {
-                        /* lost the race, just free it */
-                        hlist_add_head(&new->cc_hash, &freelist);
-                        new = NULL;
-                }
-
-                /* hot node, move to head */
-                if (hash_head->first != &ctx->cc_hash) {
-                        __hlist_del(&ctx->cc_hash);
-                        hlist_add_head(&ctx->cc_hash, hash_head);
-                }
-        } else {
-                /* don't allocate for reverse sec */
-                if (sec->ps_flags & PTLRPC_SEC_FL_REVERSE) {
-                        spin_unlock(&sec->ps_lock);
-                        RETURN(NULL);
-                }
-
-                if (new) {
-                        ctx_enhash(new, hash_head);
-                        ctx = new;
-                } else if (create) {
-                        spin_unlock(&sec->ps_lock);
-                        new = sec->ps_policy->sp_cops->create_ctx(sec, vcred);
-                        if (new) {
-                                atomic_inc(&sec->ps_busy);
-                                goto retry;
-                        }
-                } else
-                        ctx = NULL;
-        }
-
-        /* hold a ref */
-        if (ctx)
-                atomic_inc(&ctx->cc_refcount);
-
-        spin_unlock(&sec->ps_lock);
-
-        /* the allocator of the context must give the first push to refresh */
-        if (new) {
-                LASSERT(new == ctx);
-                sptlrpc_ctx_refresh(new);
-        }
-
-        ctx_list_destroy(&freelist);
-        RETURN(ctx);
-}
-
-static inline
 struct ptlrpc_cli_ctx *get_my_ctx(struct ptlrpc_sec *sec)
 {
         struct vfs_cred vcred;
         int create = 1, remove_dead = 1;
 
         LASSERT(sec);
+        LASSERT(sec->ps_policy->sp_cops->lookup_ctx);
 
-        if (sec->ps_flags & (PTLRPC_SEC_FL_REVERSE | PTLRPC_SEC_FL_ROOTONLY)) {
+        if (sec->ps_flvr.sf_flags & (PTLRPC_SEC_FL_REVERSE |
+                                     PTLRPC_SEC_FL_ROOTONLY)) {
                 vcred.vc_uid = 0;
                 vcred.vc_gid = 0;
-                if (sec->ps_flags & PTLRPC_SEC_FL_REVERSE) {
+                if (sec->ps_flvr.sf_flags & PTLRPC_SEC_FL_REVERSE) {
                         create = 0;
                         remove_dead = 0;
                 }
@@ -549,34 +245,19 @@ struct ptlrpc_cli_ctx *get_my_ctx(struct ptlrpc_sec *sec)
                 vcred.vc_gid = cfs_current()->gid;
         }
 
-        if (sec->ps_policy->sp_cops->lookup_ctx)
-                return sec->ps_policy->sp_cops->lookup_ctx(sec, &vcred);
-        else
-                return ctx_cache_lookup(sec, &vcred, create, remove_dead);
+        return sec->ps_policy->sp_cops->lookup_ctx(sec, &vcred,
+                                                   create, remove_dead);
 }
 
-/**************************************************
- * client context APIs                            *
- **************************************************/
-
-static
-void sptlrpc_ctx_refresh(struct ptlrpc_cli_ctx *ctx)
-{
-        LASSERT(atomic_read(&ctx->cc_refcount) > 0);
-
-        if (!ctx_is_refreshed(ctx) && ctx->cc_ops->refresh)
-                ctx->cc_ops->refresh(ctx);
-}
-
-struct ptlrpc_cli_ctx *sptlrpc_ctx_get(struct ptlrpc_cli_ctx *ctx)
+struct ptlrpc_cli_ctx *sptlrpc_cli_ctx_get(struct ptlrpc_cli_ctx *ctx)
 {
         LASSERT(atomic_read(&ctx->cc_refcount) > 0);
         atomic_inc(&ctx->cc_refcount);
         return ctx;
 }
-EXPORT_SYMBOL(sptlrpc_ctx_get);
+EXPORT_SYMBOL(sptlrpc_cli_ctx_get);
 
-void sptlrpc_ctx_put(struct ptlrpc_cli_ctx *ctx, int sync)
+void sptlrpc_cli_ctx_put(struct ptlrpc_cli_ctx *ctx, int sync)
 {
         struct ptlrpc_sec *sec = ctx->cc_sec;
 
@@ -586,134 +267,108 @@ void sptlrpc_ctx_put(struct ptlrpc_cli_ctx *ctx, int sync)
         if (!atomic_dec_and_test(&ctx->cc_refcount))
                 return;
 
-        LASSERT(test_bit(PTLRPC_CTX_HASHED_BIT, &ctx->cc_flags) == 0);
-        LASSERT(hlist_unhashed(&ctx->cc_hash));
-
-        /* if required async, we must clear the UPTODATE bit to prevent extra
-         * rpcs during destroy procedure.
-         */
-        if (!sync)
-                clear_bit(PTLRPC_CTX_UPTODATE_BIT, &ctx->cc_flags);
-
-        /* destroy this context */
-        if (!sptlrpc_sec_destroy_ctx(sec, ctx))
-                return;
-
-        CWARN("%s@%p: put last ctx, also destroy the sec\n",
-              sec->ps_policy->sp_name, sec);
-
-        sptlrpc_sec_destroy(sec);
+        sec->ps_policy->sp_cops->release_ctx(sec, ctx, sync);
 }
-EXPORT_SYMBOL(sptlrpc_ctx_put);
+EXPORT_SYMBOL(sptlrpc_cli_ctx_put);
 
 /*
- * mark a ctx as DEAD, and pull it out from hash table.
- *
- * NOTE: the caller must hold at least 1 ref on the ctx.
+ * expire the context immediately.
+ * the caller must hold at least 1 ref on the ctx.
  */
-void sptlrpc_ctx_expire(struct ptlrpc_cli_ctx *ctx)
+void sptlrpc_cli_ctx_expire(struct ptlrpc_cli_ctx *ctx)
 {
-        LASSERT(ctx->cc_sec);
-        LASSERT(atomic_read(&ctx->cc_refcount) > 0);
-
-        ctx_expire(ctx);
-
-        spin_lock(&ctx->cc_sec->ps_lock);
+        LASSERT(ctx->cc_ops->die);
+        ctx->cc_ops->die(ctx, 0);
+}
+EXPORT_SYMBOL(sptlrpc_cli_ctx_expire);
 
-        if (test_and_clear_bit(PTLRPC_CTX_HASHED_BIT, &ctx->cc_flags)) {
-                LASSERT(!hlist_unhashed(&ctx->cc_hash));
-                LASSERT(atomic_read(&ctx->cc_refcount) > 1);
+void sptlrpc_cli_ctx_wakeup(struct ptlrpc_cli_ctx *ctx)
+{
+        struct ptlrpc_request *req, *next;
 
-                hlist_del_init(&ctx->cc_hash);
-                if (atomic_dec_and_test(&ctx->cc_refcount))
-                        LBUG();
+        spin_lock(&ctx->cc_lock);
+        list_for_each_entry_safe(req, next, &ctx->cc_req_list, rq_ctx_chain) {
+                list_del_init(&req->rq_ctx_chain);
+                ptlrpc_wake_client_req(req);
         }
-
-        spin_unlock(&ctx->cc_sec->ps_lock);
+        spin_unlock(&ctx->cc_lock);
 }
-EXPORT_SYMBOL(sptlrpc_ctx_expire);
+EXPORT_SYMBOL(sptlrpc_cli_ctx_wakeup);
 
-void sptlrpc_ctx_replace(struct ptlrpc_sec *sec, struct ptlrpc_cli_ctx *new)
+int sptlrpc_cli_ctx_display(struct ptlrpc_cli_ctx *ctx, char *buf, int bufsize)
 {
-        struct ptlrpc_cli_ctx *ctx;
-        struct hlist_node *pos, *next;
-        HLIST_HEAD(freelist);
-        unsigned int hash;
-        ENTRY;
+        LASSERT(ctx->cc_ops);
 
-        hash = ctx_hash_index(sec, (__u64) new->cc_vcred.vc_uid);
-        LASSERT(hash < sec->ps_ccache_size);
+        if (ctx->cc_ops->display == NULL)
+                return 0;
 
-        spin_lock(&sec->ps_lock);
+        return ctx->cc_ops->display(ctx, buf, bufsize);
+}
 
-        hlist_for_each_entry_safe(ctx, pos, next,
-                                  &sec->ps_ccache[hash], cc_hash) {
-                if (!ctx_match(ctx, &new->cc_vcred))
-                        continue;
+static int sptlrpc_import_sec_check_expire(struct obd_import *imp)
+{
+        int     adapt = 0;
 
-                ctx_expire(ctx);
-                ctx_unhash(ctx, &freelist);
-                break;
+        spin_lock(&imp->imp_lock);
+        if (imp->imp_sec_expire &&
+            imp->imp_sec_expire < cfs_time_current_sec()) {
+                adapt = 1;
+                imp->imp_sec_expire = 0;
         }
+        spin_unlock(&imp->imp_lock);
 
-        ctx_enhash(new, &sec->ps_ccache[hash]);
-        atomic_inc(&sec->ps_busy);
-
-        spin_unlock(&sec->ps_lock);
+        if (!adapt)
+                return 0;
 
-        ctx_list_destroy(&freelist);
-        EXIT;
+        CDEBUG(D_SEC, "found delayed sec adapt expired, do it now\n");
+        return sptlrpc_import_sec_adapt(imp, NULL, 0);
 }
-EXPORT_SYMBOL(sptlrpc_ctx_replace);
 
 int sptlrpc_req_get_ctx(struct ptlrpc_request *req)
 {
         struct obd_import *imp = req->rq_import;
+        struct ptlrpc_sec *sec;
+        int                rc;
         ENTRY;
 
         LASSERT(!req->rq_cli_ctx);
         LASSERT(imp);
 
-        if (imp->imp_sec == NULL) {
-                CERROR("import %p (%s) with no sec pointer\n",
+        if (unlikely(imp->imp_sec_expire)) {
+                rc = sptlrpc_import_sec_check_expire(imp);
+                if (rc)
+                        RETURN(rc);
+        }
+
+        sec = sptlrpc_import_sec_ref(imp);
+        if (sec == NULL) {
+                CERROR("import %p (%s) with no ptlrpc_sec\n",
                        imp, ptlrpc_import_state_name(imp->imp_state));
                 RETURN(-EACCES);
         }
 
-        req->rq_cli_ctx = get_my_ctx(imp->imp_sec);
-
-        if (!req->rq_cli_ctx) {
-                CERROR("req %p: fail to get context from cache\n", req);
-                RETURN(-ENOMEM);
+        if (unlikely(sec->ps_dying)) {
+                CERROR("attempt to use dying sec %p\n", sec);
+                return -EACCES;
         }
 
-        RETURN(0);
-}
+        req->rq_cli_ctx = get_my_ctx(sec);
 
-void sptlrpc_ctx_wakeup(struct ptlrpc_cli_ctx *ctx)
-{
-        struct ptlrpc_request *req, *next;
+        sptlrpc_sec_put(sec);
 
-        spin_lock(&ctx->cc_lock);
-        list_for_each_entry_safe(req, next, &ctx->cc_req_list, rq_ctx_chain) {
-                list_del_init(&req->rq_ctx_chain);
-                ptlrpc_wake_client_req(req);
+        if (!req->rq_cli_ctx) {
+                CERROR("req %p: fail to get context\n", req);
+                RETURN(-ENOMEM);
         }
-        spin_unlock(&ctx->cc_lock);
-}
-EXPORT_SYMBOL(sptlrpc_ctx_wakeup);
 
-int sptlrpc_ctx_display(struct ptlrpc_cli_ctx *ctx, char *buf, int bufsize)
-{
-        LASSERT(ctx->cc_ops);
-
-        if (ctx->cc_ops->display == NULL)
-                return 0;
-
-        return ctx->cc_ops->display(ctx, buf, bufsize);
+        RETURN(0);
 }
 
-void sptlrpc_req_put_ctx(struct ptlrpc_request *req)
+/*
+ * if @sync == 0, this function should return quickly without sleep;
+ * otherwise might trigger ctx destroying rpc to server.
+ */
+void sptlrpc_req_put_ctx(struct ptlrpc_request *req, int sync)
 {
         ENTRY;
 
@@ -729,50 +384,127 @@ void sptlrpc_req_put_ctx(struct ptlrpc_request *req)
                 spin_unlock(&req->rq_cli_ctx->cc_lock);
         }
 
-        /* this could be called with spinlock hold, use async mode */
-        sptlrpc_ctx_put(req->rq_cli_ctx, 0);
+        sptlrpc_cli_ctx_put(req->rq_cli_ctx, sync);
         req->rq_cli_ctx = NULL;
         EXIT;
 }
 
+static
+int sptlrpc_req_ctx_switch(struct ptlrpc_request *req,
+                           struct ptlrpc_cli_ctx *oldctx,
+                           struct ptlrpc_cli_ctx *newctx)
+{
+        struct sptlrpc_flavor   old_flvr;
+        char                   *reqmsg;
+        int                     reqmsg_size;
+        int                     rc;
+
+        if (likely(oldctx->cc_sec == newctx->cc_sec))
+                return 0;
+
+        LASSERT(req->rq_reqmsg);
+        LASSERT(req->rq_reqlen);
+        LASSERT(req->rq_replen);
+
+        CWARN("req %p: switch ctx %p -> %p, switch sec %p(%s) -> %p(%s)\n",
+              req, oldctx, newctx,
+              oldctx->cc_sec, oldctx->cc_sec->ps_policy->sp_name,
+              newctx->cc_sec, newctx->cc_sec->ps_policy->sp_name);
+
+        /* save flavor */
+        old_flvr = req->rq_flvr;
+
+        /* save request message */
+        reqmsg_size = req->rq_reqlen;
+        OBD_ALLOC(reqmsg, reqmsg_size);
+        if (reqmsg == NULL)
+                return -ENOMEM;
+        memcpy(reqmsg, req->rq_reqmsg, reqmsg_size);
+
+        /* release old req/rep buf */
+        req->rq_cli_ctx = oldctx;
+        sptlrpc_cli_free_reqbuf(req);
+        sptlrpc_cli_free_repbuf(req);
+        req->rq_cli_ctx = newctx;
+
+        /* recalculate the flavor */
+        sptlrpc_req_set_flavor(req, 0);
+
+        /* alloc new request buffer
+         * we don't need to alloc reply buffer here, leave it to the
+         * rest procedure of ptlrpc
+         */
+        rc = sptlrpc_cli_alloc_reqbuf(req, reqmsg_size);
+        if (!rc) {
+                LASSERT(req->rq_reqmsg);
+                memcpy(req->rq_reqmsg, reqmsg, reqmsg_size);
+        } else {
+                CWARN("failed to alloc reqbuf: %d\n", rc);
+                req->rq_flvr = old_flvr;
+        }
+
+        OBD_FREE(reqmsg, reqmsg_size);
+        return rc;
+}
+
 /*
- * request must have a context. if failed to get new context,
- * just restore the old one
+ * request must have a context. in any case of failure, restore the
+ * restore the old one. a request must have a ctx.
  */
 int sptlrpc_req_replace_dead_ctx(struct ptlrpc_request *req)
 {
-        struct ptlrpc_cli_ctx *ctx = req->rq_cli_ctx;
-        int rc;
+        struct ptlrpc_cli_ctx *oldctx = req->rq_cli_ctx;
+        struct ptlrpc_cli_ctx *newctx;
+        int                    rc;
         ENTRY;
 
-        LASSERT(ctx);
-        LASSERT(test_bit(PTLRPC_CTX_DEAD_BIT, &ctx->cc_flags));
+        LASSERT(oldctx);
+        LASSERT(test_bit(PTLRPC_CTX_DEAD_BIT, &oldctx->cc_flags));
 
-        /* make sure not on context waiting list */
-        spin_lock(&ctx->cc_lock);
-        list_del_init(&req->rq_ctx_chain);
-        spin_unlock(&ctx->cc_lock);
+        sptlrpc_cli_ctx_get(oldctx);
+        sptlrpc_req_put_ctx(req, 0);
 
-        sptlrpc_ctx_get(ctx);
-        sptlrpc_req_put_ctx(req);
         rc = sptlrpc_req_get_ctx(req);
-        if (!rc) {
-                LASSERT(req->rq_cli_ctx);
-                LASSERT(req->rq_cli_ctx != ctx);
-                sptlrpc_ctx_put(ctx, 1);
-        } else {
+        if (unlikely(rc)) {
                 LASSERT(!req->rq_cli_ctx);
-                req->rq_cli_ctx = ctx;
+
+                /* restore old ctx */
+                req->rq_cli_ctx = oldctx;
+                RETURN(rc);
         }
-        RETURN(rc);
+
+        newctx = req->rq_cli_ctx;
+        LASSERT(newctx);
+
+        if (unlikely(newctx == oldctx)) {
+                /*
+                 * still get the old ctx, usually means system busy
+                 */
+                CWARN("ctx (%p, fl %lx) doesn't switch, relax a little bit\n",
+                      newctx, newctx->cc_flags);
+
+                cfs_schedule_timeout(CFS_TASK_INTERRUPTIBLE, HZ);
+        } else {
+                rc = sptlrpc_req_ctx_switch(req, oldctx, newctx);
+                if (rc) {
+                        /* restore old ctx */
+                        sptlrpc_req_put_ctx(req, 0);
+                        req->rq_cli_ctx = oldctx;
+                        RETURN(rc);
+                }
+
+                LASSERT(req->rq_cli_ctx == newctx);
+        }
+
+        sptlrpc_cli_ctx_put(oldctx, 1);
+        RETURN(0);
 }
 EXPORT_SYMBOL(sptlrpc_req_replace_dead_ctx);
 
 static
 int ctx_check_refresh(struct ptlrpc_cli_ctx *ctx)
 {
-        smp_mb();
-        if (ctx_is_refreshed(ctx))
+        if (cli_ctx_is_refreshed(ctx))
                 return 1;
         return 0;
 }
@@ -794,20 +526,33 @@ int ctx_refresh_timeout(void *data)
          * later than the context refresh expire time.
          */
         if (rc == 0)
-                ctx_expire(req->rq_cli_ctx);
+                req->rq_cli_ctx->cc_ops->die(req->rq_cli_ctx, 0);
         return rc;
 }
 
 static
 void ctx_refresh_interrupt(void *data)
 {
-        /* do nothing */
+        struct ptlrpc_request *req = data;
+
+        spin_lock(&req->rq_lock);
+        req->rq_intr = 1;
+        spin_unlock(&req->rq_lock);
+}
+
+static
+void req_off_ctx_list(struct ptlrpc_request *req, struct ptlrpc_cli_ctx *ctx)
+{
+        spin_lock(&ctx->cc_lock);
+        if (!list_empty(&req->rq_ctx_chain))
+                list_del_init(&req->rq_ctx_chain);
+        spin_unlock(&ctx->cc_lock);
 }
 
 /*
  * the status of context could be subject to be changed by other threads at any
  * time. we allow this race. but once we return with 0, the caller will
- * suppose it's uptodated and keep using it until the affected rpc is done.
+ * suppose it's uptodated and keep using it until the owning rpc is done.
  *
  * @timeout:
  *    < 0  - don't wait
@@ -825,28 +570,31 @@ int sptlrpc_req_refresh_ctx(struct ptlrpc_request *req, long timeout)
 
         LASSERT(ctx);
 
-        /* special ctxs */
-        if (ctx_is_eternal(ctx) || req->rq_ctx_init || req->rq_ctx_fini)
+        /*
+         * during the process a request's context might change type even
+         * (e.g. from gss ctx to plain ctx), so each loop we need to re-check
+         * everything
+         */
+again:
+        /* skip special ctxs */
+        if (cli_ctx_is_eternal(ctx) || req->rq_ctx_init || req->rq_ctx_fini)
                 RETURN(0);
 
-        /* reverse ctxs, don't refresh */
-        if (ctx->cc_sec->ps_flags & PTLRPC_SEC_FL_REVERSE)
-                RETURN(0);
+        if (test_bit(PTLRPC_CTX_NEW_BIT, &ctx->cc_flags)) {
+                LASSERT(ctx->cc_ops->refresh);
+                ctx->cc_ops->refresh(ctx);
+        }
+        LASSERT(test_bit(PTLRPC_CTX_NEW_BIT, &ctx->cc_flags) == 0);
 
-        spin_lock(&ctx->cc_lock);
-again:
-        if (ctx_check_uptodate(ctx)) {
-                if (!list_empty(&req->rq_ctx_chain))
-                        list_del_init(&req->rq_ctx_chain);
-                spin_unlock(&ctx->cc_lock);
+        LASSERT(ctx->cc_ops->validate);
+        if (ctx->cc_ops->validate(ctx) == 0) {
+                req_off_ctx_list(req, ctx);
                 RETURN(0);
         }
 
-        if (test_bit(PTLRPC_CTX_ERROR_BIT, &ctx->cc_flags)) {
+        if (unlikely(test_bit(PTLRPC_CTX_ERROR_BIT, &ctx->cc_flags))) {
                 req->rq_err = 1;
-                if (!list_empty(&req->rq_ctx_chain))
-                        list_del_init(&req->rq_ctx_chain);
-                spin_unlock(&ctx->cc_lock);
+                req_off_ctx_list(req, ctx);
                 RETURN(-EPERM);
         }
 
@@ -875,50 +623,40 @@ again:
          *     never really send request with old context before.
          */
         if (test_bit(PTLRPC_CTX_UPTODATE_BIT, &ctx->cc_flags) &&
-            req->rq_reqmsg &&
+            unlikely(req->rq_reqmsg) &&
             lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
-                if (!list_empty(&req->rq_ctx_chain))
-                        list_del_init(&req->rq_ctx_chain);
-                spin_unlock(&ctx->cc_lock);
+                req_off_ctx_list(req, ctx);
                 RETURN(0);
         }
 
         if (unlikely(test_bit(PTLRPC_CTX_DEAD_BIT, &ctx->cc_flags))) {
-                spin_unlock(&ctx->cc_lock);
-
-                /* don't have to, but we don't want to release it too soon */
-                sptlrpc_ctx_get(ctx);
-
                 rc = sptlrpc_req_replace_dead_ctx(req);
                 if (rc) {
                         LASSERT(ctx == req->rq_cli_ctx);
-                        CERROR("req %p: failed to replace dead ctx %p\n",
-                                req, ctx);
+                        CERROR("req %p: failed to replace dead ctx %p: %d\n",
+                                req, ctx, rc);
                         req->rq_err = 1;
                         LASSERT(list_empty(&req->rq_ctx_chain));
-                        sptlrpc_ctx_put(ctx, 1);
-                        RETURN(-ENOMEM);
+                        RETURN(rc);
                 }
 
-                LASSERT(ctx != req->rq_cli_ctx);
-                CWARN("req %p: replace dead ctx %p(%u->%s) => %p\n",
-                      req, ctx, ctx->cc_vcred.vc_uid,
-                      sec2target_str(ctx->cc_sec), req->rq_cli_ctx);
+                CWARN("req %p: replace dead ctx %p => ctx %p (%u->%s)\n",
+                      req, ctx, req->rq_cli_ctx,
+                      req->rq_cli_ctx->cc_vcred.vc_uid,
+                      sec2target_str(req->rq_cli_ctx->cc_sec));
 
-                sptlrpc_ctx_put(ctx, 1);
                 ctx = req->rq_cli_ctx;
                 LASSERT(list_empty(&req->rq_ctx_chain));
 
-                spin_lock(&ctx->cc_lock);
                 goto again;
         }
 
         /* Now we're sure this context is during upcall, add myself into
          * waiting list
          */
+        spin_lock(&ctx->cc_lock);
         if (list_empty(&req->rq_ctx_chain))
                 list_add(&req->rq_ctx_chain, &ctx->cc_req_list);
-
         spin_unlock(&ctx->cc_lock);
 
         if (timeout < 0) {
@@ -934,22 +672,22 @@ again:
         req->rq_restart = 0;
         spin_unlock(&req->rq_lock);
 
-        lwi = LWI_TIMEOUT_INTR(timeout == 0 ? LONG_MAX : timeout * HZ,
-                               ctx_refresh_timeout, ctx_refresh_interrupt, req);
+        lwi = LWI_TIMEOUT_INTR(timeout * HZ, ctx_refresh_timeout,
+                               ctx_refresh_interrupt, req);
         rc = l_wait_event(req->rq_reply_waitq, ctx_check_refresh(ctx), &lwi);
 
-        spin_lock(&ctx->cc_lock);
-        /* five cases we are here:
-         * 1. successfully refreshed;
-         * 2. someone else mark this ctx dead by force;
-         * 3. interruptted;
-         * 4. timedout, and we don't want recover from the failure;
-         * 5. timedout, and waked up upon recovery finished;
+        /* following cases we could be here:
+         * - successfully refreshed;
+         * - interruptted;
+         * - timedout, and we don't want recover from the failure;
+         * - timedout, and waked up upon recovery finished;
+         * - someone else mark this ctx dead by force;
+         * - someone invalidate the req and call wake_client_req(),
+         *   e.g. ptlrpc_abort_inflight();
          */
-        if (!ctx_is_refreshed(ctx)) {
+        if (!cli_ctx_is_refreshed(ctx)) {
                 /* timed out or interruptted */
-                list_del_init(&req->rq_ctx_chain);
-                spin_unlock(&ctx->cc_lock);
+                req_off_ctx_list(req, ctx);
 
                 LASSERT(rc != 0);
                 RETURN(rc);
@@ -958,12 +696,16 @@ again:
         goto again;
 }
 
+/*
+ * Note this could be called in two situations:
+ * - new request from ptlrpc_pre_req(), with proper @opcode
+ * - old request which changed ctx in the middle, with @opcode == 0
+ */
 void sptlrpc_req_set_flavor(struct ptlrpc_request *req, int opcode)
 {
-        struct sec_flavor_config *conf;
+        struct ptlrpc_sec *sec;
 
         LASSERT(req->rq_import);
-        LASSERT(req->rq_import->imp_sec);
         LASSERT(req->rq_cli_ctx);
         LASSERT(req->rq_cli_ctx->cc_sec);
         LASSERT(req->rq_bulk_read == 0 || req->rq_bulk_write == 0);
@@ -982,44 +724,45 @@ void sptlrpc_req_set_flavor(struct ptlrpc_request *req, int opcode)
         case SEC_CTX_FINI:
                 req->rq_ctx_fini = 1;
                 break;
+        case 0:
+                /* init/fini rpc won't be resend, so can't be here */
+                LASSERT(req->rq_ctx_init == 0);
+                LASSERT(req->rq_ctx_fini == 0);
+
+                /* cleanup flags, which should be recalculated */
+                req->rq_pack_udesc = 0;
+                req->rq_pack_bulk = 0;
+                break;
         }
 
-        req->rq_sec_flavor = req->rq_cli_ctx->cc_sec->ps_flavor;
+        sec = req->rq_cli_ctx->cc_sec;
 
-        /* force SVC_NONE for context initiation rpc, SVC_AUTH for context
-         * destruction rpc
-         */
-        if (unlikely(req->rq_ctx_init)) {
-                req->rq_sec_flavor = SEC_MAKE_RPC_FLAVOR(
-                                SEC_FLAVOR_POLICY(req->rq_sec_flavor),
-                                SEC_FLAVOR_SUBPOLICY(req->rq_sec_flavor),
-                                SEC_FLAVOR_SVC(SPTLRPC_SVC_NONE));
-        } else if (unlikely(req->rq_ctx_fini)) {
-                req->rq_sec_flavor = SEC_MAKE_RPC_FLAVOR(
-                                SEC_FLAVOR_POLICY(req->rq_sec_flavor),
-                                SEC_FLAVOR_SUBPOLICY(req->rq_sec_flavor),
-                                SEC_FLAVOR_SVC(SPTLRPC_SVC_AUTH));
-        }
-
-        conf = &req->rq_import->imp_obd->u.cli.cl_sec_conf;
-
-        /* user descriptor flag, except ROOTONLY which don't need, and
-         * null security which can't
-         */
-        if ((conf->sfc_flags & PTLRPC_SEC_FL_ROOTONLY) == 0 &&
-            req->rq_sec_flavor != SPTLRPC_FLVR_NULL)
-                req->rq_sec_flavor |= SEC_FLAVOR_FL_USER;
+        spin_lock(&sec->ps_lock);
+        req->rq_flvr = sec->ps_flvr;
+        spin_unlock(&sec->ps_lock);
+
+        /* force SVC_NULL for context initiation rpc, SVC_INTG for context
+         * destruction rpc */
+        if (unlikely(req->rq_ctx_init))
+                rpc_flvr_set_svc(&req->rq_flvr.sf_rpc, SPTLRPC_SVC_NULL);
+        else if (unlikely(req->rq_ctx_fini))
+                rpc_flvr_set_svc(&req->rq_flvr.sf_rpc, SPTLRPC_SVC_INTG);
+
+        /* user descriptor flag, null security can't do it anyway */
+        if ((sec->ps_flvr.sf_flags & PTLRPC_SEC_FL_UDESC) &&
+            (req->rq_flvr.sf_rpc != SPTLRPC_FLVR_NULL))
+                req->rq_pack_udesc = 1;
 
         /* bulk security flag */
         if ((req->rq_bulk_read || req->rq_bulk_write) &&
-            (conf->sfc_bulk_priv != BULK_PRIV_ALG_NULL ||
-             conf->sfc_bulk_csum != BULK_CSUM_ALG_NULL))
-                req->rq_sec_flavor |= SEC_FLAVOR_FL_BULK;
+            (req->rq_flvr.sf_bulk_ciph != BULK_CIPH_ALG_NULL ||
+             req->rq_flvr.sf_bulk_hash != BULK_HASH_ALG_NULL))
+                req->rq_pack_bulk = 1;
 }
 
 void sptlrpc_request_out_callback(struct ptlrpc_request *req)
 {
-        if (SEC_FLAVOR_SVC(req->rq_sec_flavor) != SPTLRPC_SVC_PRIV)
+        if (RPC_FLVR_SVC(req->rq_flvr.sf_rpc) != SPTLRPC_SVC_PRIV)
                 return;
 
         LASSERT(req->rq_clrbuf);
@@ -1038,6 +781,7 @@ void sptlrpc_request_out_callback(struct ptlrpc_request *req)
  */
 int sptlrpc_import_check_ctx(struct obd_import *imp)
 {
+        struct ptlrpc_sec     *sec;
         struct ptlrpc_cli_ctx *ctx;
         struct ptlrpc_request *req = NULL;
         int rc;
@@ -1045,12 +789,16 @@ int sptlrpc_import_check_ctx(struct obd_import *imp)
 
         might_sleep();
 
-        ctx = get_my_ctx(imp->imp_sec);
+        sec = sptlrpc_import_sec_ref(imp);
+        ctx = get_my_ctx(sec);
+        sptlrpc_sec_put(sec);
+
         if (!ctx)
                 RETURN(1);
 
-        if (ctx_is_eternal(ctx)) {
-                sptlrpc_ctx_put(ctx, 1);
+        if (cli_ctx_is_eternal(ctx) ||
+            ctx->cc_ops->validate(ctx) == 0) {
+                sptlrpc_cli_ctx_put(ctx, 1);
                 RETURN(0);
         }
 
@@ -1060,14 +808,14 @@ int sptlrpc_import_check_ctx(struct obd_import *imp)
 
         spin_lock_init(&req->rq_lock);
         atomic_set(&req->rq_refcount, 10000);
-        INIT_LIST_HEAD(&req->rq_ctx_chain);
-        init_waitqueue_head(&req->rq_reply_waitq);
+        CFS_INIT_LIST_HEAD(&req->rq_ctx_chain);
+        cfs_waitq_init(&req->rq_reply_waitq);
         req->rq_import = imp;
         req->rq_cli_ctx = ctx;
 
         rc = sptlrpc_req_refresh_ctx(req, 0);
         LASSERT(list_empty(&req->rq_ctx_chain));
-        sptlrpc_ctx_put(req->rq_cli_ctx, 1);
+        sptlrpc_cli_ctx_put(req->rq_cli_ctx, 1);
         OBD_FREE_PTR(req);
 
         RETURN(rc);
@@ -1092,9 +840,10 @@ int sptlrpc_cli_wrap_request(struct ptlrpc_request *req)
                         RETURN(rc);
         }
 
-        switch (SEC_FLAVOR_SVC(req->rq_sec_flavor)) {
-        case SPTLRPC_SVC_NONE:
+        switch (RPC_FLVR_SVC(req->rq_flvr.sf_rpc)) {
+        case SPTLRPC_SVC_NULL:
         case SPTLRPC_SVC_AUTH:
+        case SPTLRPC_SVC_INTG:
                 LASSERT(ctx->cc_ops->sign);
                 rc = ctx->cc_ops->sign(ctx, req);
                 break;
@@ -1115,62 +864,49 @@ int sptlrpc_cli_wrap_request(struct ptlrpc_request *req)
         RETURN(rc);
 }
 
-/*
- * rq_nob_received is the actual received data length
- */
-int sptlrpc_cli_unwrap_reply(struct ptlrpc_request *req)
+static int do_cli_unwrap_reply(struct ptlrpc_request *req)
 {
         struct ptlrpc_cli_ctx *ctx = req->rq_cli_ctx;
-        int rc;
+        int                    rc;
+        __u16                  rpc_flvr;
         ENTRY;
 
         LASSERT(ctx);
         LASSERT(ctx->cc_sec);
-        LASSERT(ctx->cc_ops);
         LASSERT(req->rq_repbuf);
+        LASSERT(req->rq_repdata);
+        LASSERT(req->rq_repmsg == NULL);
 
-        req->rq_repdata_len = req->rq_nob_received;
-
-        if (req->rq_nob_received < sizeof(struct lustre_msg)) {
+        if (req->rq_repdata_len < sizeof(struct lustre_msg)) {
                 CERROR("replied data length %d too small\n",
-                       req->rq_nob_received);
+                       req->rq_repdata_len);
                 RETURN(-EPROTO);
         }
 
-        if (req->rq_repbuf->lm_magic == LUSTRE_MSG_MAGIC_V1 ||
-            req->rq_repbuf->lm_magic == LUSTRE_MSG_MAGIC_V1_SWABBED) {
-                /* it's must be null flavor, so our requets also should be
-                 * in null flavor */
-                if (SEC_FLAVOR_POLICY(req->rq_sec_flavor) !=
-                    SPTLRPC_POLICY_NULL) {
-                        CERROR("request flavor is %x but reply with null\n",
-                               req->rq_sec_flavor);
-                        RETURN(-EPROTO);
-                }
-        } else {
-                /* v2 message... */
-                ptlrpc_sec_flavor_t tmpf = req->rq_repbuf->lm_secflvr;
-
-                if (req->rq_repbuf->lm_magic == LUSTRE_MSG_MAGIC_V2_SWABBED)
-                        __swab32s(&tmpf);
-
-                if (SEC_FLAVOR_POLICY(tmpf) !=
-                    SEC_FLAVOR_POLICY(req->rq_sec_flavor)) {
-                        CERROR("request policy %u while reply with %d\n",
-                               SEC_FLAVOR_POLICY(req->rq_sec_flavor),
-                               SEC_FLAVOR_POLICY(tmpf));
-                        RETURN(-EPROTO);
-                }
+        /* v2 message, check request/reply policy match */
+        rpc_flvr = WIRE_FLVR_RPC(req->rq_repdata->lm_secflvr);
 
-                if ((SEC_FLAVOR_POLICY(req->rq_sec_flavor) !=
-                     SPTLRPC_POLICY_NULL) &&
-                    lustre_unpack_msg(req->rq_repbuf, req->rq_nob_received))
-                        RETURN(-EPROTO);
+        if (req->rq_repdata->lm_magic == LUSTRE_MSG_MAGIC_V2_SWABBED)
+                __swab16s(&rpc_flvr);
+
+        if (RPC_FLVR_POLICY(rpc_flvr) !=
+            RPC_FLVR_POLICY(req->rq_flvr.sf_rpc)) {
+                CERROR("request policy was %u while reply with %u\n",
+                       RPC_FLVR_POLICY(req->rq_flvr.sf_rpc),
+                       RPC_FLVR_POLICY(rpc_flvr));
+                RETURN(-EPROTO);
         }
 
-        switch (SEC_FLAVOR_SVC(req->rq_sec_flavor)) {
-        case SPTLRPC_SVC_NONE:
+        /* do nothing if it's null policy; otherwise unpack the
+         * wrapper message */
+        if (RPC_FLVR_POLICY(rpc_flvr) != SPTLRPC_POLICY_NULL &&
+            lustre_unpack_msg(req->rq_repdata, req->rq_repdata_len))
+                RETURN(-EPROTO);
+
+        switch (RPC_FLVR_SVC(req->rq_flvr.sf_rpc)) {
+        case SPTLRPC_SVC_NULL:
         case SPTLRPC_SVC_AUTH:
+        case SPTLRPC_SVC_INTG:
                 LASSERT(ctx->cc_ops->verify);
                 rc = ctx->cc_ops->verify(ctx, req);
                 break;
@@ -1186,122 +922,495 @@ int sptlrpc_cli_unwrap_reply(struct ptlrpc_request *req)
         RETURN(rc);
 }
 
+/*
+ * upon this be called, the reply buffer should have been un-posted,
+ * so nothing is going to change.
+ */
+int sptlrpc_cli_unwrap_reply(struct ptlrpc_request *req)
+{
+        LASSERT(req->rq_repbuf);
+        LASSERT(req->rq_repdata == NULL);
+        LASSERT(req->rq_repmsg == NULL);
+        LASSERT(req->rq_reply_off + req->rq_nob_received <= req->rq_repbuf_len);
+
+        if (req->rq_reply_off == 0 &&
+            (lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
+                CERROR("real reply with offset 0\n");
+                return -EPROTO;
+        }
+
+        if (req->rq_reply_off % 8 != 0) {
+                CERROR("reply at odd offset %u\n", req->rq_reply_off);
+                return -EPROTO;
+        }
+
+        req->rq_repdata = (struct lustre_msg *)
+                                (req->rq_repbuf + req->rq_reply_off);
+        req->rq_repdata_len = req->rq_nob_received;
+
+        return do_cli_unwrap_reply(req);
+}
+
+/**
+ * Upon called, the receive buffer might be still posted, so the reply data
+ * might be changed at any time, no matter we're holding rq_lock or not. we
+ * expect the rq_reply_off be 0, rq_nob_received is the early reply size.
+ *
+ * we allocate separate ptlrpc_request and reply buffer for early reply
+ * processing, return 0 and @req_ret is a duplicated ptlrpc_request. caller
+ * must call sptlrpc_cli_finish_early_reply() on the returned request to
+ * release it. if anything goes wrong @req_ret will not be set.
+ */
+int sptlrpc_cli_unwrap_early_reply(struct ptlrpc_request *req,
+                                   struct ptlrpc_request **req_ret)
+{
+        struct ptlrpc_request  *early_req;
+        char                   *early_buf;
+        int                     early_bufsz, early_size;
+        int                     rc;
+        ENTRY;
+
+        OBD_ALLOC_PTR(early_req);
+        if (early_req == NULL)
+                RETURN(-ENOMEM);
+
+        early_size = req->rq_nob_received;
+        early_bufsz = size_roundup_power2(early_size);
+        OBD_ALLOC(early_buf, early_bufsz);
+        if (early_buf == NULL)
+                GOTO(err_req, rc = -ENOMEM);
+
+        /* sanity checkings and copy data out, do it inside spinlock */
+        spin_lock(&req->rq_lock);
+
+        if (req->rq_replied) {
+                spin_unlock(&req->rq_lock);
+                GOTO(err_buf, rc = -EALREADY);
+        }
+
+        LASSERT(req->rq_repbuf);
+        LASSERT(req->rq_repdata == NULL);
+        LASSERT(req->rq_repmsg == NULL);
+
+        if (req->rq_reply_off != 0) {
+                CERROR("early reply with offset %u\n", req->rq_reply_off);
+                spin_unlock(&req->rq_lock);
+                GOTO(err_buf, rc = -EPROTO);
+        }
+
+        if (req->rq_nob_received != early_size) {
+                /* even another early arrived the size should be the same */
+                CERROR("data size has changed from %u to %u\n",
+                       early_size, req->rq_nob_received);
+                spin_unlock(&req->rq_lock);
+                GOTO(err_buf, rc = -EINVAL);
+        }
+
+        if (req->rq_nob_received < sizeof(struct lustre_msg)) {
+                CERROR("early reply length %d too small\n",
+                       req->rq_nob_received);
+                spin_unlock(&req->rq_lock);
+                GOTO(err_buf, rc = -EALREADY);
+        }
+
+        memcpy(early_buf, req->rq_repbuf, early_size);
+        spin_unlock(&req->rq_lock);
+
+        early_req->rq_cli_ctx = sptlrpc_cli_ctx_get(req->rq_cli_ctx);
+        early_req->rq_flvr = req->rq_flvr;
+        early_req->rq_repbuf = early_buf;
+        early_req->rq_repbuf_len = early_bufsz;
+        early_req->rq_repdata = (struct lustre_msg *) early_buf;
+        early_req->rq_repdata_len = early_size;
+        early_req->rq_early = 1;
+
+        rc = do_cli_unwrap_reply(early_req);
+        if (rc) {
+                DEBUG_REQ(D_ADAPTTO, early_req,
+                          "error %d unwrap early reply", rc);
+                GOTO(err_ctx, rc);
+        }
+
+        LASSERT(early_req->rq_repmsg);
+        *req_ret = early_req;
+        RETURN(0);
+
+err_ctx:
+        sptlrpc_cli_ctx_put(early_req->rq_cli_ctx, 1);
+err_buf:
+        OBD_FREE(early_buf, early_bufsz);
+err_req:
+        OBD_FREE_PTR(early_req);
+        RETURN(rc);
+}
+
+void sptlrpc_cli_finish_early_reply(struct ptlrpc_request *early_req)
+{
+        LASSERT(early_req->rq_repbuf);
+        LASSERT(early_req->rq_repdata);
+        LASSERT(early_req->rq_repmsg);
+
+        sptlrpc_cli_ctx_put(early_req->rq_cli_ctx, 1);
+        OBD_FREE(early_req->rq_repbuf, early_req->rq_repbuf_len);
+        OBD_FREE_PTR(early_req);
+}
+
+/**************************************************
+ * sec ID                                         *
+ **************************************************/
+
+/*
+ * "fixed" sec (e.g. null) use sec_id < 0
+ */
+static atomic_t sptlrpc_sec_id = ATOMIC_INIT(1);
+
+int sptlrpc_get_next_secid(void)
+{
+        return atomic_inc_return(&sptlrpc_sec_id);
+}
+EXPORT_SYMBOL(sptlrpc_get_next_secid);
+
 /**************************************************
- * security APIs                                  *
+ * client side high-level security APIs           *
  **************************************************/
 
+static int sec_cop_flush_ctx_cache(struct ptlrpc_sec *sec, uid_t uid,
+                                   int grace, int force)
+{
+        struct ptlrpc_sec_policy *policy = sec->ps_policy;
+
+        LASSERT(policy->sp_cops);
+        LASSERT(policy->sp_cops->flush_ctx_cache);
+
+        return policy->sp_cops->flush_ctx_cache(sec, uid, grace, force);
+}
+
+static void sec_cop_destroy_sec(struct ptlrpc_sec *sec)
+{
+        struct ptlrpc_sec_policy *policy = sec->ps_policy;
+
+        LASSERT(atomic_read(&sec->ps_refcount) == 0);
+        LASSERT(atomic_read(&sec->ps_nctx) == 0);
+        LASSERT(policy->sp_cops->destroy_sec);
+
+        CDEBUG(D_SEC, "%s@%p: being destroied\n", sec->ps_policy->sp_name, sec);
+
+        policy->sp_cops->destroy_sec(sec);
+        sptlrpc_policy_put(policy);
+}
+
+void sptlrpc_sec_destroy(struct ptlrpc_sec *sec)
+{
+        sec_cop_destroy_sec(sec);
+}
+EXPORT_SYMBOL(sptlrpc_sec_destroy);
+
+static void sptlrpc_sec_kill(struct ptlrpc_sec *sec)
+{
+        LASSERT(atomic_read(&sec->ps_refcount) > 0);
+
+        if (sec->ps_policy->sp_cops->kill_sec) {
+                sec->ps_policy->sp_cops->kill_sec(sec);
+
+                sec_cop_flush_ctx_cache(sec, -1, 1, 1);
+        }
+}
+
+struct ptlrpc_sec *sptlrpc_sec_get(struct ptlrpc_sec *sec)
+{
+        if (sec) {
+                LASSERT(atomic_read(&sec->ps_refcount) > 0);
+                atomic_inc(&sec->ps_refcount);
+        }
+
+        return sec;
+}
+EXPORT_SYMBOL(sptlrpc_sec_get);
+
+void sptlrpc_sec_put(struct ptlrpc_sec *sec)
+{
+        if (sec) {
+                LASSERT(atomic_read(&sec->ps_refcount) > 0);
+
+                if (atomic_dec_and_test(&sec->ps_refcount)) {
+                        LASSERT(atomic_read(&sec->ps_nctx) == 0);
+
+                        sptlrpc_gc_del_sec(sec);
+                        sec_cop_destroy_sec(sec);
+                }
+        }
+}
+EXPORT_SYMBOL(sptlrpc_sec_put);
+
 /*
- * let policy module to determine whether take refrence of
- * import or not.
+ * it's policy module responsible for taking refrence of import
  */
 static
 struct ptlrpc_sec * sptlrpc_sec_create(struct obd_import *imp,
-                                       struct ptlrpc_svc_ctx *ctx,
-                                       __u32 flavor,
-                                       unsigned long flags)
+                                       struct ptlrpc_svc_ctx *svc_ctx,
+                                       struct sptlrpc_flavor *sf,
+                                       enum lustre_sec_part sp)
 {
         struct ptlrpc_sec_policy *policy;
-        struct ptlrpc_sec *sec;
+        struct ptlrpc_sec        *sec;
         ENTRY;
 
-        flavor = SEC_FLAVOR_RPC(flavor);
-
-        if (ctx) {
+        if (svc_ctx) {
                 LASSERT(imp->imp_dlm_fake == 1);
 
                 CDEBUG(D_SEC, "%s %s: reverse sec using flavor %s\n",
                        imp->imp_obd->obd_type->typ_name,
                        imp->imp_obd->obd_name,
-                       sptlrpc_flavor2name(flavor));
+                       sptlrpc_rpcflavor2name(sf->sf_rpc));
 
-                policy = sptlrpc_policy_get(ctx->sc_policy);
-                flags |= PTLRPC_SEC_FL_REVERSE | PTLRPC_SEC_FL_ROOTONLY;
+                policy = sptlrpc_policy_get(svc_ctx->sc_policy);
+                sf->sf_flags |= PTLRPC_SEC_FL_REVERSE | PTLRPC_SEC_FL_ROOTONLY;
         } else {
                 LASSERT(imp->imp_dlm_fake == 0);
 
                 CDEBUG(D_SEC, "%s %s: select security flavor %s\n",
                        imp->imp_obd->obd_type->typ_name,
                        imp->imp_obd->obd_name,
-                       sptlrpc_flavor2name(flavor));
+                       sptlrpc_rpcflavor2name(sf->sf_rpc));
 
-                policy = sptlrpc_flavor2policy(flavor);
+                policy = sptlrpc_rpcflavor2policy(sf->sf_rpc);
                 if (!policy) {
-                        CERROR("invalid flavor 0x%x\n", flavor);
+                        CERROR("invalid flavor 0x%x\n", sf->sf_rpc);
                         RETURN(NULL);
                 }
         }
 
-        sec = policy->sp_cops->create_sec(imp, ctx, flavor, flags);
-        if (sec) {
-                atomic_inc(&sec->ps_refcount);
-
-                /* take 1 busy count on behalf of sec itself,
-                 * balanced in sptlrpc_set_put()
-                 */
-                atomic_inc(&sec->ps_busy);
-        } else
-                sptlrpc_policy_put(policy);
+        sec = policy->sp_cops->create_sec(imp, svc_ctx, sf);
+        if (sec) {
+                atomic_inc(&sec->ps_refcount);
+
+                sec->ps_part = sp;
+
+                if (sec->ps_gc_interval && policy->sp_cops->gc_ctx)
+                        sptlrpc_gc_add_sec(sec);
+        } else {
+                sptlrpc_policy_put(policy);
+        }
+
+        RETURN(sec);
+}
+
+struct ptlrpc_sec *sptlrpc_import_sec_ref(struct obd_import *imp)
+{
+        struct ptlrpc_sec *sec;
+
+        spin_lock(&imp->imp_lock);
+        sec = sptlrpc_sec_get(imp->imp_sec);
+        spin_unlock(&imp->imp_lock);
+
+        return sec;
+}
+EXPORT_SYMBOL(sptlrpc_import_sec_ref);
+
+static void sptlrpc_import_sec_install(struct obd_import *imp,
+                                       struct ptlrpc_sec *sec)
+{
+        struct ptlrpc_sec *old_sec;
+
+        LASSERT(atomic_read(&sec->ps_refcount) > 0);
+
+        spin_lock(&imp->imp_lock);
+        old_sec = imp->imp_sec;
+        imp->imp_sec = sec;
+        spin_unlock(&imp->imp_lock);
+
+        if (old_sec) {
+                sptlrpc_sec_kill(old_sec);
+
+                /* balance the ref taken by this import */
+                sptlrpc_sec_put(old_sec);
+        }
+}
+
+static void sptlrpc_import_sec_adapt_inplace(struct obd_import *imp,
+                                             struct ptlrpc_sec *sec,
+                                             struct sptlrpc_flavor *sf)
+{
+        if (sf->sf_bulk_ciph != sec->ps_flvr.sf_bulk_ciph ||
+            sf->sf_bulk_hash != sec->ps_flvr.sf_bulk_hash) {
+                CWARN("imp %p (%s->%s): changing bulk flavor %s/%s -> %s/%s\n",
+                      imp, imp->imp_obd->obd_name,
+                      obd_uuid2str(&imp->imp_connection->c_remote_uuid),
+                      sptlrpc_get_ciph_name(sec->ps_flvr.sf_bulk_ciph),
+                      sptlrpc_get_hash_name(sec->ps_flvr.sf_bulk_hash),
+                      sptlrpc_get_ciph_name(sf->sf_bulk_ciph),
+                      sptlrpc_get_hash_name(sf->sf_bulk_hash));
+
+                spin_lock(&sec->ps_lock);
+                sec->ps_flvr.sf_bulk_ciph = sf->sf_bulk_ciph;
+                sec->ps_flvr.sf_bulk_hash = sf->sf_bulk_hash;
+                spin_unlock(&sec->ps_lock);
+        }
+
+        if (!equi(sf->sf_flags & PTLRPC_SEC_FL_UDESC,
+                  sec->ps_flvr.sf_flags & PTLRPC_SEC_FL_UDESC)) {
+                CWARN("imp %p (%s->%s): %s shipping user descriptor\n",
+                      imp, imp->imp_obd->obd_name,
+                      obd_uuid2str(&imp->imp_connection->c_remote_uuid),
+                      (sf->sf_flags & PTLRPC_SEC_FL_UDESC) ? "start" : "stop");
+
+                spin_lock(&sec->ps_lock);
+                sec->ps_flvr.sf_flags &= ~PTLRPC_SEC_FL_UDESC;
+                sec->ps_flvr.sf_flags |= sf->sf_flags & PTLRPC_SEC_FL_UDESC;
+                spin_unlock(&sec->ps_lock);
+        }
+}
+
+/*
+ * for normal import, @svc_ctx should be NULL and @rpc_flavor is ignored;
+ * for reverse import, @svc_ctx and @rpc_flavor is from incoming request.
+ */
+int sptlrpc_import_sec_adapt(struct obd_import *imp,
+                             struct ptlrpc_svc_ctx *svc_ctx,
+                             __u16 rpc_flavor)
+{
+        struct ptlrpc_connection   *conn;
+        struct sptlrpc_flavor       sf;
+        struct ptlrpc_sec          *sec, *newsec;
+        enum lustre_sec_part        sp;
+        int                         rc;
+
+        if (imp == NULL)
+                return 0;
+
+        conn = imp->imp_connection;
+
+        if (svc_ctx == NULL) {
+                /* normal import, determine flavor from rule set */
+                sptlrpc_rule_set_choose(&imp->imp_obd->u.cli.cl_sptlrpc_rset,
+                                        LUSTRE_SP_ANY, conn->c_self, &sf);
+
+                sp = imp->imp_obd->u.cli.cl_sec_part;
+        } else {
+                /* reverse import, determine flavor from incoming reqeust */
+                sf.sf_rpc = rpc_flavor;
+                sf.sf_bulk_ciph = BULK_CIPH_ALG_NULL;
+                sf.sf_bulk_hash = BULK_HASH_ALG_NULL;
+                sf.sf_flags = PTLRPC_SEC_FL_REVERSE | PTLRPC_SEC_FL_ROOTONLY;
+
+                sp = sptlrpc_target_sec_part(imp->imp_obd);
+        }
+
+        sec = sptlrpc_import_sec_ref(imp);
+        if (sec) {
+                if (svc_ctx == NULL) {
+                        /* normal import, only check rpc flavor, if just bulk
+                         * flavor or flags changed, we can handle it on the fly
+                         * without switching sec. */
+                        if (sf.sf_rpc == sec->ps_flvr.sf_rpc) {
+                                sptlrpc_import_sec_adapt_inplace(imp, sec, &sf);
+
+                                rc = 0;
+                                goto out;
+                        }
+                } else {
+                        /* reverse import, do not compare bulk flavor */
+                        if (sf.sf_rpc == sec->ps_flvr.sf_rpc) {
+                                rc = 0;
+                                goto out;
+                        }
+                }
+
+                CWARN("%simport %p (%s%s%s): changing flavor "
+                      "(%s, %s/%s) -> (%s, %s/%s)\n",
+                      svc_ctx ? "reverse " : "",
+                      imp, imp->imp_obd->obd_name,
+                      svc_ctx == NULL ? "->" : "<-",
+                      obd_uuid2str(&conn->c_remote_uuid),
+                      sptlrpc_rpcflavor2name(sec->ps_flvr.sf_rpc),
+                      sptlrpc_get_hash_name(sec->ps_flvr.sf_bulk_hash),
+                      sptlrpc_get_ciph_name(sec->ps_flvr.sf_bulk_ciph),
+                      sptlrpc_rpcflavor2name(sf.sf_rpc),
+                      sptlrpc_get_hash_name(sf.sf_bulk_hash),
+                      sptlrpc_get_ciph_name(sf.sf_bulk_ciph));
+        } else {
+                CWARN("%simport %p (%s%s%s) netid %x: "
+                      "select initial flavor (%s, %s/%s)\n",
+                      svc_ctx == NULL ? "" : "reverse ",
+                      imp, imp->imp_obd->obd_name,
+                      svc_ctx == NULL ? "->" : "<-",
+                      obd_uuid2str(&conn->c_remote_uuid),
+                      LNET_NIDNET(conn->c_self),
+                      sptlrpc_rpcflavor2name(sf.sf_rpc),
+                      sptlrpc_get_hash_name(sf.sf_bulk_hash),
+                      sptlrpc_get_ciph_name(sf.sf_bulk_ciph));
+        }
+
+        mutex_down(&imp->imp_sec_mutex);
+
+        newsec = sptlrpc_sec_create(imp, svc_ctx, &sf, sp);
+        if (newsec) {
+                sptlrpc_import_sec_install(imp, newsec);
+                rc = 0;
+        } else {
+                CERROR("%simport %p (%s): failed to create new sec\n",
+                       svc_ctx == NULL ? "" : "reverse ",
+                       imp, obd_uuid2str(&conn->c_remote_uuid));
+                rc = -EPERM;
+        }
+
+        mutex_up(&imp->imp_sec_mutex);
 
-        RETURN(sec);
+out:
+        sptlrpc_sec_put(sec);
+        return 0;
 }
 
-static
-void sptlrpc_sec_destroy(struct ptlrpc_sec *sec)
+void sptlrpc_import_sec_put(struct obd_import *imp)
 {
-        struct ptlrpc_sec_policy *policy = sec->ps_policy;
-
-        LASSERT(policy);
-        LASSERT(atomic_read(&sec->ps_refcount) == 0);
-        LASSERT(atomic_read(&sec->ps_busy) == 0);
-        LASSERT(policy->sp_cops->destroy_sec);
+        if (imp->imp_sec) {
+                sptlrpc_sec_kill(imp->imp_sec);
 
-        policy->sp_cops->destroy_sec(sec);
-        sptlrpc_policy_put(policy);
+                sptlrpc_sec_put(imp->imp_sec);
+                imp->imp_sec = NULL;
+        }
 }
 
-static
-void sptlrpc_sec_put(struct ptlrpc_sec *sec)
+static void import_flush_ctx_common(struct obd_import *imp,
+                                    uid_t uid, int grace, int force)
 {
-        struct ptlrpc_sec_policy *policy = sec->ps_policy;
+        struct ptlrpc_sec *sec;
 
-        if (!atomic_dec_and_test(&sec->ps_refcount)) {
-                sptlrpc_policy_put(policy);
+        if (imp == NULL)
                 return;
-        }
 
-        ctx_cache_flush(sec, -1, 1, 1);
+        sec = sptlrpc_import_sec_ref(imp);
+        if (sec == NULL)
+                return;
 
-        if (atomic_dec_and_test(&sec->ps_busy))
-                sptlrpc_sec_destroy(sec);
-        else
-                CWARN("delay to destroy %s@%p: busy contexts\n",
-                      policy->sp_name, sec);
+        sec_cop_flush_ctx_cache(sec, uid, grace, force);
+        sptlrpc_sec_put(sec);
 }
 
-/*
- * return 1 means we should also destroy the sec structure.
- * normally return 0
- */
-static
-int sptlrpc_sec_destroy_ctx(struct ptlrpc_sec *sec,
-                            struct ptlrpc_cli_ctx *ctx)
+void sptlrpc_import_inval_all_ctx(struct obd_import *imp)
 {
-        LASSERT(sec == ctx->cc_sec);
-        LASSERT(atomic_read(&sec->ps_busy));
-        LASSERT(atomic_read(&ctx->cc_refcount) == 0);
-        LASSERT(hlist_unhashed(&ctx->cc_hash));
-        LASSERT(list_empty(&ctx->cc_req_list));
-        LASSERT(sec->ps_policy->sp_cops->destroy_ctx);
+        /* use grace == 0 */
+        import_flush_ctx_common(imp, -1, 0, 1);
+}
 
-        sec->ps_policy->sp_cops->destroy_ctx(sec, ctx);
+void sptlrpc_import_flush_root_ctx(struct obd_import *imp)
+{
+        /* it's important to use grace mode, see explain in
+         * sptlrpc_req_refresh_ctx() */
+        import_flush_ctx_common(imp, 0, 1, 1);
+}
 
-        if (atomic_dec_and_test(&sec->ps_busy)) {
-                LASSERT(atomic_read(&sec->ps_refcount) == 0);
-                return 1;
-        }
+void sptlrpc_import_flush_my_ctx(struct obd_import *imp)
+{
+        import_flush_ctx_common(imp, cfs_current()->uid, 1, 1);
+}
+EXPORT_SYMBOL(sptlrpc_import_flush_my_ctx);
 
-        return 0;
+void sptlrpc_import_flush_all_ctx(struct obd_import *imp)
+{
+        import_flush_ctx_common(imp, -1, 1, 1);
 }
+EXPORT_SYMBOL(sptlrpc_import_flush_all_ctx);
 
 /*
  * when complete successfully, req->rq_reqmsg should point to the
@@ -1342,7 +1451,9 @@ void sptlrpc_cli_free_reqbuf(struct ptlrpc_request *req)
         LASSERT(atomic_read(&ctx->cc_refcount));
         LASSERT(ctx->cc_sec);
         LASSERT(ctx->cc_sec->ps_policy);
-        LASSERT(req->rq_reqbuf || req->rq_clrbuf);
+
+        if (req->rq_reqbuf == NULL && req->rq_clrbuf == NULL)
+                return;
 
         policy = ctx->cc_sec->ps_policy;
         policy->sp_cops->free_reqbuf(ctx->cc_sec, req);
@@ -1449,96 +1560,303 @@ void sptlrpc_cli_free_repbuf(struct ptlrpc_request *req)
         LASSERT(atomic_read(&ctx->cc_refcount));
         LASSERT(ctx->cc_sec);
         LASSERT(ctx->cc_sec->ps_policy);
-        LASSERT(req->rq_repbuf);
+
+        if (req->rq_repbuf == NULL)
+                return;
+        LASSERT(req->rq_repbuf_len);
 
         policy = ctx->cc_sec->ps_policy;
         policy->sp_cops->free_repbuf(ctx->cc_sec, req);
         EXIT;
 }
 
-int sptlrpc_import_get_sec(struct obd_import *imp,
-                           struct ptlrpc_svc_ctx *ctx,
-                           __u32 flavor,
-                           unsigned long flags)
+int sptlrpc_cli_install_rvs_ctx(struct obd_import *imp,
+                                struct ptlrpc_cli_ctx *ctx)
 {
-        struct obd_device *obd = imp->imp_obd;
-        ENTRY;
-
-        LASSERT(obd);
-        LASSERT(obd->obd_type);
-
-        /* old sec might be still there in reconnecting */
-        if (imp->imp_sec)
-                RETURN(0);
-
-        imp->imp_sec = sptlrpc_sec_create(imp, ctx, flavor, flags);
-        if (!imp->imp_sec)
-                RETURN(-EINVAL);
+        struct ptlrpc_sec_policy *policy = ctx->cc_sec->ps_policy;
 
-        RETURN(0);
+        if (!policy->sp_cops->install_rctx)
+                return 0;
+        return policy->sp_cops->install_rctx(imp, ctx->cc_sec, ctx);
 }
 
-void sptlrpc_import_put_sec(struct obd_import *imp)
+int sptlrpc_svc_install_rvs_ctx(struct obd_import *imp,
+                                struct ptlrpc_svc_ctx *ctx)
 {
-        if (imp->imp_sec == NULL)
-                return;
+        struct ptlrpc_sec_policy *policy = ctx->sc_policy;
 
-        sptlrpc_sec_put(imp->imp_sec);
-        imp->imp_sec = NULL;
+        if (!policy->sp_sops->install_rctx)
+                return 0;
+        return policy->sp_sops->install_rctx(imp, ctx);
 }
 
-void sptlrpc_import_flush_root_ctx(struct obd_import *imp)
+/****************************************
+ * server side security                 *
+ ****************************************/
+
+static int flavor_allowed(struct sptlrpc_flavor *exp,
+                          struct ptlrpc_request *req)
 {
-        if (imp == NULL || imp->imp_sec == NULL)
-                return;
+        struct sptlrpc_flavor *flvr = &req->rq_flvr;
 
-        /* use 'grace' mode, it's crutial see explain in
-         * sptlrpc_req_refresh_ctx()
-         */
-        ctx_cache_flush(imp->imp_sec, 0, 1, 1);
-}
+        if (exp->sf_rpc == flvr->sf_rpc)
+                return 1;
 
-void sptlrpc_import_flush_my_ctx(struct obd_import *imp)
-{
-        if (imp == NULL || imp->imp_sec == NULL)
-                return;
+        if ((req->rq_ctx_init || req->rq_ctx_fini) &&
+            RPC_FLVR_POLICY(exp->sf_rpc) == RPC_FLVR_POLICY(flvr->sf_rpc) &&
+            RPC_FLVR_MECH(exp->sf_rpc) == RPC_FLVR_MECH(flvr->sf_rpc))
+                return 1;
 
-        ctx_cache_flush(imp->imp_sec, cfs_current()->uid, 1, 1);
+        return 0;
 }
-EXPORT_SYMBOL(sptlrpc_import_flush_my_ctx);
 
-void sptlrpc_import_flush_all_ctx(struct obd_import *imp)
+#define EXP_FLVR_UPDATE_EXPIRE      (OBD_TIMEOUT_DEFAULT + 10)
+
+int sptlrpc_target_export_check(struct obd_export *exp,
+                                struct ptlrpc_request *req)
 {
-        if (imp == NULL || imp->imp_sec == NULL)
-                return;
+        struct sptlrpc_flavor   flavor;
 
-        ctx_cache_flush(imp->imp_sec, -1, 0, 1);
-}
-EXPORT_SYMBOL(sptlrpc_import_flush_all_ctx);
+        if (exp == NULL)
+                return 0;
 
-int sptlrpc_cli_install_rvs_ctx(struct obd_import *imp,
-                                struct ptlrpc_cli_ctx *ctx)
-{
-        struct ptlrpc_sec_policy *policy = ctx->cc_sec->ps_policy;
+        /* client side export has no imp_reverse, skip
+         * FIXME maybe we should check flavor this as well??? */
+        if (exp->exp_imp_reverse == NULL)
+                return 0;
 
-        if (!policy->sp_cops->install_rctx)
+        /* don't care about ctx fini rpc */
+        if (req->rq_ctx_fini)
                 return 0;
-        return policy->sp_cops->install_rctx(imp, ctx->cc_sec, ctx);
+
+        spin_lock(&exp->exp_lock);
+
+        /* if flavor just changed (exp->exp_flvr_changed != 0), we wait for
+         * the first req with the new flavor, then treat it as current flavor,
+         * adapt reverse sec according to it.
+         * note the first rpc with new flavor might not be with root ctx, in
+         * which case delay the sec_adapt by leaving exp_flvr_adapt == 1. */
+        if (unlikely(exp->exp_flvr_changed) &&
+            flavor_allowed(&exp->exp_flvr_old[1], req)) {
+                /* make the new flavor as "current", and old ones as
+                 * about-to-expire */
+                CDEBUG(D_SEC, "exp %p: just changed: %x->%x\n", exp,
+                       exp->exp_flvr.sf_rpc, exp->exp_flvr_old[1].sf_rpc);
+                flavor = exp->exp_flvr_old[1];
+                exp->exp_flvr_old[1] = exp->exp_flvr_old[0];
+                exp->exp_flvr_expire[1] = exp->exp_flvr_expire[0];
+                exp->exp_flvr_old[0] = exp->exp_flvr;
+                exp->exp_flvr_expire[0] = cfs_time_current_sec() +
+                                          EXP_FLVR_UPDATE_EXPIRE;
+                exp->exp_flvr = flavor;
+
+                /* flavor change finished */
+                exp->exp_flvr_changed = 0;
+                LASSERT(exp->exp_flvr_adapt == 1);
+
+                /* if it's gss, we only interested in root ctx init */
+                if (req->rq_auth_gss &&
+                    !(req->rq_ctx_init && (req->rq_auth_usr_root ||
+                                           req->rq_auth_usr_mdt))) {
+                        spin_unlock(&exp->exp_lock);
+                        CDEBUG(D_SEC, "is good but not root(%d:%d:%d:%d)\n",
+                               req->rq_auth_gss, req->rq_ctx_init,
+                               req->rq_auth_usr_root, req->rq_auth_usr_mdt);
+                        return 0;
+                }
+
+                exp->exp_flvr_adapt = 0;
+                spin_unlock(&exp->exp_lock);
+
+                return sptlrpc_import_sec_adapt(exp->exp_imp_reverse,
+                                                req->rq_svc_ctx, flavor.sf_rpc);
+        }
+
+        /* if it equals to the current flavor, we accept it, but need to
+         * dealing with reverse sec/ctx */
+        if (likely(flavor_allowed(&exp->exp_flvr, req))) {
+                /* most cases should return here, we only interested in
+                 * gss root ctx init */
+                if (!req->rq_auth_gss || !req->rq_ctx_init ||
+                    (!req->rq_auth_usr_root && !req->rq_auth_usr_mdt)) {
+                        spin_unlock(&exp->exp_lock);
+                        return 0;
+                }
+
+                /* if flavor just changed, we should not proceed, just leave
+                 * it and current flavor will be discovered and replaced
+                 * shortly, and let _this_ rpc pass through */
+                if (exp->exp_flvr_changed) {
+                        LASSERT(exp->exp_flvr_adapt);
+                        spin_unlock(&exp->exp_lock);
+                        return 0;
+                }
+
+                if (exp->exp_flvr_adapt) {
+                        exp->exp_flvr_adapt = 0;
+                        CDEBUG(D_SEC, "exp %p (%x|%x|%x): do delayed adapt\n",
+                               exp, exp->exp_flvr.sf_rpc,
+                               exp->exp_flvr_old[0].sf_rpc,
+                               exp->exp_flvr_old[1].sf_rpc);
+                        flavor = exp->exp_flvr;
+                        spin_unlock(&exp->exp_lock);
+
+                        return sptlrpc_import_sec_adapt(exp->exp_imp_reverse,
+                                                        req->rq_svc_ctx,
+                                                        flavor.sf_rpc);
+                } else {
+                        CDEBUG(D_SEC, "exp %p (%x|%x|%x): is current flavor, "
+                               "install rvs ctx\n", exp, exp->exp_flvr.sf_rpc,
+                               exp->exp_flvr_old[0].sf_rpc,
+                               exp->exp_flvr_old[1].sf_rpc);
+                        spin_unlock(&exp->exp_lock);
+
+                        return sptlrpc_svc_install_rvs_ctx(exp->exp_imp_reverse,
+                                                           req->rq_svc_ctx);
+                }
+        }
+
+        if (exp->exp_flvr_expire[0]) {
+                if (exp->exp_flvr_expire[0] >= cfs_time_current_sec()) {
+                        if (flavor_allowed(&exp->exp_flvr_old[0], req)) {
+                                CDEBUG(D_SEC, "exp %p (%x|%x|%x): match the "
+                                       "middle one ("CFS_DURATION_T")\n", exp,
+                                       exp->exp_flvr.sf_rpc,
+                                       exp->exp_flvr_old[0].sf_rpc,
+                                       exp->exp_flvr_old[1].sf_rpc,
+                                       exp->exp_flvr_expire[0] -
+                                                cfs_time_current_sec());
+                                spin_unlock(&exp->exp_lock);
+                                return 0;
+                        }
+                } else {
+                        CDEBUG(D_SEC, "mark middle expired\n");
+                        exp->exp_flvr_expire[0] = 0;
+                }
+                CDEBUG(D_SEC, "exp %p (%x|%x|%x): %x not match middle\n", exp,
+                       exp->exp_flvr.sf_rpc,
+                       exp->exp_flvr_old[0].sf_rpc, exp->exp_flvr_old[1].sf_rpc,
+                       req->rq_flvr.sf_rpc);
+        }
+
+        /* now it doesn't match the current flavor, the only chance we can
+         * accept it is match the old flavors which is not expired. */
+        if (exp->exp_flvr_changed == 0 && exp->exp_flvr_expire[1]) {
+                if (exp->exp_flvr_expire[1] >= cfs_time_current_sec()) {
+                        if (flavor_allowed(&exp->exp_flvr_old[1], req)) {
+                                CDEBUG(D_SEC, "exp %p (%x|%x|%x): match the "
+                                       "oldest one ("CFS_DURATION_T")\n", exp,
+                                       exp->exp_flvr.sf_rpc,
+                                       exp->exp_flvr_old[0].sf_rpc,
+                                       exp->exp_flvr_old[1].sf_rpc,
+                                       exp->exp_flvr_expire[1] -
+                                                cfs_time_current_sec());
+                                spin_unlock(&exp->exp_lock);
+                                return 0;
+                        }
+                } else {
+                        CDEBUG(D_SEC, "mark oldest expired\n");
+                        exp->exp_flvr_expire[1] = 0;
+                }
+                CDEBUG(D_SEC, "exp %p (%x|%x|%x): %x not match found\n",
+                       exp, exp->exp_flvr.sf_rpc,
+                       exp->exp_flvr_old[0].sf_rpc, exp->exp_flvr_old[1].sf_rpc,
+                       req->rq_flvr.sf_rpc);
+        } else {
+                CDEBUG(D_SEC, "exp %p (%x|%x|%x): skip the last one\n",
+                       exp, exp->exp_flvr.sf_rpc, exp->exp_flvr_old[0].sf_rpc,
+                       exp->exp_flvr_old[1].sf_rpc);
+        }
+
+        spin_unlock(&exp->exp_lock);
+
+        CWARN("req %p: (%u|%u|%u|%u|%u) with unauthorized flavor %x\n",
+              req, req->rq_auth_gss, req->rq_ctx_init, req->rq_ctx_fini,
+              req->rq_auth_usr_root, req->rq_auth_usr_mdt, req->rq_flvr.sf_rpc);
+        return -EACCES;
 }
 
-int sptlrpc_svc_install_rvs_ctx(struct obd_import *imp,
-                                struct ptlrpc_svc_ctx *ctx)
+void sptlrpc_target_update_exp_flavor(struct obd_device *obd,
+                                      struct sptlrpc_rule_set *rset)
 {
-        struct ptlrpc_sec_policy *policy = ctx->sc_policy;
+        struct obd_export       *exp;
+        struct sptlrpc_flavor    new_flvr;
 
-        if (!policy->sp_sops->install_rctx)
-                return 0;
-        return policy->sp_sops->install_rctx(imp, ctx);
+        LASSERT(obd);
+
+        spin_lock(&obd->obd_dev_lock);
+
+        list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
+                if (exp->exp_connection == NULL)
+                        continue;
+
+                /* note if this export had just been updated flavor
+                 * (exp_flvr_changed == 1), this will override the
+                 * previous one. */
+                spin_lock(&exp->exp_lock);
+                sptlrpc_rule_set_choose(rset, exp->exp_sp_peer,
+                                        exp->exp_connection->c_peer.nid,
+                                        &new_flvr);
+                if (exp->exp_flvr_changed ||
+                    memcmp(&new_flvr, &exp->exp_flvr, sizeof(new_flvr))) {
+                        exp->exp_flvr_old[1] = new_flvr;
+                        exp->exp_flvr_expire[1] = 0;
+                        exp->exp_flvr_changed = 1;
+                        exp->exp_flvr_adapt = 1;
+                        CDEBUG(D_SEC, "exp %p (%s): updated flavor %x->%x\n",
+                               exp, sptlrpc_part2name(exp->exp_sp_peer),
+                               exp->exp_flvr.sf_rpc,
+                               exp->exp_flvr_old[1].sf_rpc);
+                }
+                spin_unlock(&exp->exp_lock);
+        }
+
+        spin_unlock(&obd->obd_dev_lock);
 }
+EXPORT_SYMBOL(sptlrpc_target_update_exp_flavor);
 
-/****************************************
- * server side security                 *
- ****************************************/
+static int sptlrpc_svc_check_from(struct ptlrpc_request *req, int svc_rc)
+{
+        if (svc_rc == SECSVC_DROP)
+                return SECSVC_DROP;
+
+        switch (req->rq_sp_from) {
+        case LUSTRE_SP_CLI:
+        case LUSTRE_SP_MDT:
+        case LUSTRE_SP_OST:
+        case LUSTRE_SP_MGS:
+        case LUSTRE_SP_ANY:
+                break;
+        default:
+                DEBUG_REQ(D_ERROR, req, "invalid source %u", req->rq_sp_from);
+                return SECSVC_DROP;
+        }
+
+        if (!req->rq_auth_gss)
+                return svc_rc;
+
+        if (unlikely(req->rq_sp_from == LUSTRE_SP_ANY)) {
+                CERROR("not specific part\n");
+                return SECSVC_DROP;
+        }
+
+        /* from MDT, must be authenticated as MDT */
+        if (unlikely(req->rq_sp_from == LUSTRE_SP_MDT &&
+                     !req->rq_auth_usr_mdt)) {
+                DEBUG_REQ(D_ERROR, req, "fake source MDT");
+                return SECSVC_DROP;
+        }
+
+        /* from OST, must be callback to MDT and CLI, the reverse sec
+         * was from mdt/root keytab, so it should be MDT or root FIXME */
+        if (unlikely(req->rq_sp_from == LUSTRE_SP_OST &&
+                     !req->rq_auth_usr_mdt && !req->rq_auth_usr_root)) {
+                DEBUG_REQ(D_ERROR, req, "fake source OST");
+                return SECSVC_DROP;
+        }
+
+        return svc_rc;
+}
 
 int sptlrpc_svc_unwrap_request(struct ptlrpc_request *req)
 {
@@ -1551,33 +1869,31 @@ int sptlrpc_svc_unwrap_request(struct ptlrpc_request *req)
         LASSERT(req->rq_reqmsg == NULL);
         LASSERT(req->rq_repmsg == NULL);
 
-        /* 
-         * in any case we avoid to call unpack_msg() for request of null flavor
-         * which will later be done by ptlrpc_server_handle_request().
-         */
+        req->rq_sp_from = LUSTRE_SP_ANY;
+        req->rq_auth_uid = INVALID_UID;
+        req->rq_auth_mapped_uid = INVALID_UID;
+
         if (req->rq_reqdata_len < sizeof(struct lustre_msg)) {
                 CERROR("request size %d too small\n", req->rq_reqdata_len);
                 RETURN(SECSVC_DROP);
         }
 
-        if (msg->lm_magic == LUSTRE_MSG_MAGIC_V1 ||
-            msg->lm_magic == LUSTRE_MSG_MAGIC_V1_SWABBED) {
-                req->rq_sec_flavor = SPTLRPC_FLVR_NULL;
-        } else {
-                req->rq_sec_flavor = msg->lm_secflvr;
-
-                if (msg->lm_magic == LUSTRE_MSG_MAGIC_V2_SWABBED)
-                        __swab32s(&req->rq_sec_flavor);
+        /*
+         * v2 message.
+         */
+        if (msg->lm_magic == LUSTRE_MSG_MAGIC_V2)
+                req->rq_flvr.sf_rpc = WIRE_FLVR_RPC(msg->lm_secflvr);
+        else
+                req->rq_flvr.sf_rpc = WIRE_FLVR_RPC(__swab32(msg->lm_secflvr));
 
-                if ((SEC_FLAVOR_POLICY(req->rq_sec_flavor) !=
-                     SPTLRPC_POLICY_NULL) &&
-                    lustre_unpack_msg(msg, req->rq_reqdata_len))
-                        RETURN(SECSVC_DROP);
-        }
+        /* unpack the wrapper message if the policy is not null */
+        if ((RPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) &&
+             lustre_unpack_msg(msg, req->rq_reqdata_len))
+                RETURN(SECSVC_DROP);
 
-        policy = sptlrpc_flavor2policy(req->rq_sec_flavor);
+        policy = sptlrpc_rpcflavor2policy(req->rq_flvr.sf_rpc);
         if (!policy) {
-                CERROR("unsupported security flavor %x\n", req->rq_sec_flavor);
+                CERROR("unsupported rpc flavor %x\n", req->rq_flvr.sf_rpc);
                 RETURN(SECSVC_DROP);
         }
 
@@ -1587,6 +1903,9 @@ int sptlrpc_svc_unwrap_request(struct ptlrpc_request *req)
         LASSERT(req->rq_reqmsg || rc != SECSVC_OK);
         sptlrpc_policy_put(policy);
 
+        /* sanity check for the request source */
+        rc = sptlrpc_svc_check_from(req, rc);
+
         /* FIXME move to proper place */
         if (rc == SECSVC_OK) {
                 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
@@ -1722,7 +2041,7 @@ int sptlrpc_cli_wrap_bulk(struct ptlrpc_request *req,
 {
         struct ptlrpc_cli_ctx *ctx;
 
-        if (!SEC_FLAVOR_HAS_BULK(req->rq_sec_flavor))
+        if (!req->rq_pack_bulk)
                 return 0;
 
         LASSERT(req->rq_bulk_read || req->rq_bulk_write);
@@ -1750,7 +2069,9 @@ void pga_to_bulk_desc(int nob, obd_count pg_count, struct brw_page **pga,
                                            nob : pga[i]->count;
                 desc->bd_iov[i].kiov_offset = pga[i]->off & ~CFS_PAGE_MASK;
 #else
-#warning FIXME for liblustre!
+                /* FIXME currently liblustre doesn't support bulk encryption.
+                 * if we do, check again following may not be right. */
+                LASSERTF(0, "Bulk encryption not implemented for liblustre\n");
                 desc->bd_iov[i].iov_base = pga[i]->pg->addr;
                 desc->bd_iov[i].iov_len = pga[i]->count > nob ?
                                            nob : pga[i]->count;
@@ -1769,7 +2090,7 @@ int sptlrpc_cli_unwrap_bulk_read(struct ptlrpc_request *req,
         struct ptlrpc_cli_ctx *ctx;
         int rc = 0;
 
-        if (!SEC_FLAVOR_HAS_BULK(req->rq_sec_flavor))
+        if (!req->rq_pack_bulk)
                 return 0;
 
         LASSERT(req->rq_bulk_read && !req->rq_bulk_write);
@@ -1797,7 +2118,7 @@ int sptlrpc_cli_unwrap_bulk_write(struct ptlrpc_request *req,
 {
         struct ptlrpc_cli_ctx *ctx;
 
-        if (!SEC_FLAVOR_HAS_BULK(req->rq_sec_flavor))
+        if (!req->rq_pack_bulk)
                 return 0;
 
         LASSERT(!req->rq_bulk_read && req->rq_bulk_write);
@@ -1815,7 +2136,7 @@ int sptlrpc_svc_wrap_bulk(struct ptlrpc_request *req,
 {
         struct ptlrpc_svc_ctx *ctx;
 
-        if (!SEC_FLAVOR_HAS_BULK(req->rq_sec_flavor))
+        if (!req->rq_pack_bulk)
                 return 0;
 
         LASSERT(req->rq_bulk_read || req->rq_bulk_write);
@@ -1833,7 +2154,7 @@ int sptlrpc_svc_unwrap_bulk(struct ptlrpc_request *req,
 {
         struct ptlrpc_svc_ctx *ctx;
 
-        if (!SEC_FLAVOR_HAS_BULK(req->rq_sec_flavor))
+        if (!req->rq_pack_bulk)
                 return 0;
 
         LASSERT(req->rq_bulk_read || req->rq_bulk_write);
@@ -1877,7 +2198,7 @@ int sptlrpc_pack_user_desc(struct lustre_msg *msg, int offset)
         pud->pud_gid = cfs_current()->gid;
         pud->pud_fsuid = cfs_current()->fsuid;
         pud->pud_fsgid = cfs_current()->fsgid;
-        pud->pud_cap = cfs_current()->cap_effective;
+        pud->pud_cap = cfs_curproc_cap_pack();
         pud->pud_ngroups = (msg->lm_buflens[offset] - sizeof(*pud)) / 4;
 
 #ifdef __KERNEL__
@@ -1933,219 +2254,6 @@ int sptlrpc_unpack_user_desc(struct lustre_msg *msg, int offset)
 EXPORT_SYMBOL(sptlrpc_unpack_user_desc);
 
 /****************************************
- * user supplied flavor string parsing  *
- ****************************************/
-
-static
-int get_default_flavor(enum lustre_part to_part, struct sec_flavor_config *conf)
-{
-        conf->sfc_bulk_priv = BULK_PRIV_ALG_NULL;
-        conf->sfc_bulk_csum = BULK_CSUM_ALG_NULL;
-        conf->sfc_flags = 0;
-
-        switch (to_part) {
-        case LUSTRE_MDT:
-                conf->sfc_rpc_flavor = SPTLRPC_FLVR_PLAIN;
-                return 0;
-        case LUSTRE_OST:
-                conf->sfc_rpc_flavor = SPTLRPC_FLVR_NULL;
-                return 0;
-        default:
-                CERROR("Unknown to lustre part %d, apply defaults\n", to_part);
-                conf->sfc_rpc_flavor = SPTLRPC_FLVR_NULL;
-                return -EINVAL;
-        }
-}
-
-static
-void get_flavor_by_rpc(__u32 rpc_flavor, struct sec_flavor_config *conf)
-{
-        conf->sfc_rpc_flavor = rpc_flavor;
-        conf->sfc_bulk_priv = BULK_PRIV_ALG_NULL;
-        conf->sfc_bulk_csum = BULK_CSUM_ALG_NULL;
-        conf->sfc_flags = 0;
-
-        switch (rpc_flavor) {
-        case SPTLRPC_FLVR_NULL:
-        case SPTLRPC_FLVR_PLAIN:
-                break;
-        case SPTLRPC_FLVR_KRB5P:
-                conf->sfc_bulk_priv = BULK_PRIV_ALG_ARC4;
-                /* fall through */
-        case SPTLRPC_FLVR_KRB5I:
-                conf->sfc_bulk_csum = BULK_CSUM_ALG_SHA1;
-                break;
-        default:
-                LBUG();
-        }
-}
-
-static
-void get_flavor_by_rpc_bulk(__u32 rpc_flavor, int bulk_priv,
-                            struct sec_flavor_config *conf)
-{
-        if (bulk_priv)
-                conf->sfc_bulk_priv = BULK_PRIV_ALG_ARC4;
-        else
-                conf->sfc_bulk_priv = BULK_PRIV_ALG_NULL;
-
-        switch (rpc_flavor) {
-        case SPTLRPC_FLVR_PLAIN:
-                conf->sfc_bulk_csum = BULK_CSUM_ALG_MD5;
-                break;
-        case SPTLRPC_FLVR_KRB5I:
-        case SPTLRPC_FLVR_KRB5P:
-                conf->sfc_bulk_csum = BULK_CSUM_ALG_SHA1;
-                break;
-        default:
-                LBUG();
-        }
-}
-
-static __u32 __flavors[] = {
-        SPTLRPC_FLVR_NULL,
-        SPTLRPC_FLVR_PLAIN,
-        SPTLRPC_FLVR_KRB5I,
-        SPTLRPC_FLVR_KRB5P,
-};
-
-#define __nflavors      (sizeof(__flavors)/sizeof(__u32))
-
-/*
- * flavor string format: rpc[-bulk{n|i|p}[:cksum/enc]]
- * for examples:
- *  null
- *  plain-bulki
- *  krb5p-bulkn
- *  krb5i-bulkp
- *  krb5i-bulkp:sha512/arc4
- */
-int sptlrpc_parse_flavor(enum lustre_part from_part, enum lustre_part to_part,
-                         char *str, struct sec_flavor_config *conf)
-{
-        char   *f, *bulk, *alg, *enc;
-        char    buf[64];
-        int     i, bulk_priv;
-        ENTRY;
-
-        if (str == NULL) {
-                if (get_default_flavor(to_part, conf))
-                        return -EINVAL;
-                goto set_flags;
-        }
-
-        for (i = 0; i < __nflavors; i++) {
-                f = sptlrpc_flavor2name(__flavors[i]);
-                if (strncmp(str, f, strlen(f)) == 0)
-                        break;
-        }
-
-        if (i >= __nflavors)
-                GOTO(invalid, -EINVAL);
-
-        /* prepare local buffer thus we can modify it as we want */
-        strncpy(buf, str, 64);
-        buf[64 - 1] = '\0';
-
-        /* find bulk string */
-        bulk = strchr(buf, '-');
-        if (bulk)
-                *bulk++ = '\0';
-
-        /* now the first part must equal to rpc flavor name */
-        if (strcmp(buf, f) != 0)
-                GOTO(invalid, -EINVAL);
-
-        get_flavor_by_rpc(__flavors[i], conf);
-
-        if (bulk == NULL)
-                goto set_flags;
-
-        /* null flavor should not have any suffix */
-        if (__flavors[i] == SPTLRPC_FLVR_NULL)
-                GOTO(invalid, -EINVAL);
-
-        /* find bulk algorithm string */
-        alg = strchr(bulk, ':');
-        if (alg)
-                *alg++ = '\0';
-
-        /* verify bulk section */
-        if (strcmp(bulk, "bulkn") == 0) {
-                conf->sfc_bulk_csum = BULK_CSUM_ALG_NULL;
-                conf->sfc_bulk_priv = BULK_PRIV_ALG_NULL;
-                goto set_flags;
-        }
-
-        if (strcmp(bulk, "bulki") == 0)
-                bulk_priv = 0;
-        else if (strcmp(bulk, "bulkp") == 0)
-                bulk_priv = 1;
-        else
-                GOTO(invalid, -EINVAL);
-
-        /* plain policy dosen't support bulk encryption */
-        if (bulk_priv && __flavors[i] == SPTLRPC_FLVR_PLAIN)
-                GOTO(invalid, -EINVAL);
-
-        get_flavor_by_rpc_bulk(__flavors[i], bulk_priv, conf);
-
-        if (alg == NULL)
-                goto set_flags;
-
-        /* find encryption algorithm string */
-        enc = strchr(alg, '/');
-        if (enc)
-                *enc++ = '\0';
-
-        /* bulk combination sanity check */
-        if ((bulk_priv && enc == NULL) || (bulk_priv == 0 && enc))
-                GOTO(invalid, -EINVAL);
-
-        /* checksum algorithm */
-        for (i = 0; i < BULK_CSUM_ALG_MAX; i++) {
-                if (strcmp(alg, sptlrpc_bulk_csum_alg2name(i)) == 0) {
-                        conf->sfc_bulk_csum = i;
-                        break;
-                }
-        }
-        if (i >= BULK_CSUM_ALG_MAX)
-                GOTO(invalid, -EINVAL);
-
-        /* privacy algorithm */
-        if (enc) {
-                if (strcmp(enc, "arc4") != 0)
-                        GOTO(invalid, -EINVAL);
-                conf->sfc_bulk_priv = BULK_PRIV_ALG_ARC4;
-        }
-
-set_flags:
-        /* * set ROOTONLY flag:
-         *   - to OST
-         *   - from MDT to MDT
-         * * set BULK flag for:
-         *   - from CLI to OST
-         */
-        if (to_part == LUSTRE_OST ||
-            (from_part == LUSTRE_MDT && to_part == LUSTRE_MDT))
-                conf->sfc_flags |= PTLRPC_SEC_FL_ROOTONLY;
-        if (from_part == LUSTRE_CLI && to_part == LUSTRE_OST)
-                conf->sfc_flags |= PTLRPC_SEC_FL_BULK;
-
-#ifdef __BIG_ENDIAN
-        __swab32s(&conf->sfc_rpc_flavor);
-        __swab32s(&conf->sfc_bulk_csum);
-        __swab32s(&conf->sfc_bulk_priv);
-        __swab32s(&conf->sfc_flags);
-#endif
-        return 0;
-invalid:
-        CERROR("invalid flavor string: %s\n", str);
-        return -EINVAL;
-}
-EXPORT_SYMBOL(sptlrpc_parse_flavor);
-
-/****************************************
  * misc helpers                         *
  ****************************************/
 
@@ -2153,24 +2261,63 @@ const char * sec2target_str(struct ptlrpc_sec *sec)
 {
         if (!sec || !sec->ps_import || !sec->ps_import->imp_obd)
                 return "*";
-        if (sec->ps_flags & PTLRPC_SEC_FL_REVERSE)
+        if (sec_is_reverse(sec))
                 return "c";
         return obd_uuid2str(&sec->ps_import->imp_obd->u.cli.cl_target_uuid);
 }
 EXPORT_SYMBOL(sec2target_str);
 
 /****************************************
+ * crypto API helper/alloc blkciper     *
+ ****************************************/
+
+#ifdef __KERNEL__
+#ifndef HAVE_ASYNC_BLOCK_CIPHER
+struct ll_crypto_cipher *ll_crypto_alloc_blkcipher(const char * algname,
+                                                   u32 type, u32 mask)
+{
+        char        buf[CRYPTO_MAX_ALG_NAME + 1];
+        const char *pan = algname;
+        u32         flag = 0; 
+
+        if (strncmp("cbc(", algname, 4) == 0)
+                flag |= CRYPTO_TFM_MODE_CBC;
+        else if (strncmp("ecb(", algname, 4) == 0)
+                flag |= CRYPTO_TFM_MODE_ECB;
+        if (flag) {
+                char *vp = strnchr(algname, CRYPTO_MAX_ALG_NAME, ')');
+                if (vp) {
+                        memcpy(buf, algname + 4, vp - algname - 4);
+                        buf[vp - algname - 4] = '\0';
+                        pan = buf;
+                } else {
+                        flag = 0;
+                }
+        }
+        return crypto_alloc_tfm(pan, flag);
+}
+EXPORT_SYMBOL(ll_crypto_alloc_blkcipher);
+#endif
+#endif
+
+/****************************************
  * initialize/finalize                  *
  ****************************************/
 
-int sptlrpc_init(void)
+int __init sptlrpc_init(void)
 {
         int rc;
 
-        rc = sptlrpc_enc_pool_init();
+        rwlock_init(&policy_lock);
+
+        rc = sptlrpc_gc_start_thread();
         if (rc)
                 goto out;
 
+        rc = sptlrpc_enc_pool_init();
+        if (rc)
+                goto out_gc;
+
         rc = sptlrpc_null_init();
         if (rc)
                 goto out_pool;
@@ -2191,14 +2338,17 @@ out_null:
         sptlrpc_null_fini();
 out_pool:
         sptlrpc_enc_pool_fini();
+out_gc:
+        sptlrpc_gc_stop_thread();
 out:
         return rc;
 }
 
-void sptlrpc_fini(void)
+void __exit sptlrpc_fini(void)
 {
         sptlrpc_lproc_fini();
         sptlrpc_plain_fini();
         sptlrpc_null_fini();
         sptlrpc_enc_pool_fini();
+        sptlrpc_gc_stop_thread();
 }