Whamcloud - gitweb
LU-1194 llog: fix for not sync llcd at thread stop
[fs/lustre-release.git] / lustre / ptlrpc / recov_thread.c
index 188574b..8076420 100644 (file)
@@ -1,6 +1,4 @@
-/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
- * vim:expandtab:shiftwidth=8:tabstop=8:
- *
+/*
  * GPL HEADER START
  *
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  * GPL HEADER END
  */
 /*
- * Copyright  2008 Sun Microsystems, Inc. All rights reserved
+ * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
+ *
+ * Copyright (c) 2011, Whamcloud, Inc.
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
 
 #define DEBUG_SUBSYSTEM S_LOG
 
-#ifndef EXPORT_SYMTAB
-# define EXPORT_SYMTAB
-#endif
-
 #ifdef __KERNEL__
 # include <libcfs/libcfs.h>
 #else
@@ -67,7 +63,7 @@
 #include <lustre_log.h>
 #include "ptlrpc_internal.h"
 
-static atomic_t                   llcd_count = ATOMIC_INIT(0);
+static cfs_atomic_t               llcd_count = CFS_ATOMIC_INIT(0);
 static cfs_mem_cache_t           *llcd_cache = NULL;
 
 #ifdef __KERNEL__
@@ -76,28 +72,56 @@ enum {
         LLOG_LCM_FL_EXIT        = 1 << 1
 };
 
-/** 
+struct llcd_async_args {
+        struct llog_canceld_ctxt *la_ctxt;
+};
+
+static void llcd_print(struct llog_canceld_ctxt *llcd,
+                       const char *func, int line)
+{
+        CDEBUG(D_RPCTRACE, "Llcd (%p) at %s:%d:\n", llcd, func, line);
+        CDEBUG(D_RPCTRACE, "  size: %d\n", llcd->llcd_size);
+        CDEBUG(D_RPCTRACE, "  ctxt: %p\n", llcd->llcd_ctxt);
+        CDEBUG(D_RPCTRACE, "  lcm : %p\n", llcd->llcd_lcm);
+        CDEBUG(D_RPCTRACE, "  cookiebytes : %d\n", llcd->llcd_cookiebytes);
+}
+
+/**
  * Allocate new llcd from cache, init it and return to caller.
  * Bumps number of objects allocated.
  */
-static struct llog_canceld_ctxt *llcd_alloc(void)
+static struct llog_canceld_ctxt *llcd_alloc(struct llog_commit_master *lcm)
 {
         struct llog_canceld_ctxt *llcd;
-        int llcd_size;
+        int size, overhead;
 
-        /* 
-         * Payload of lustre_msg V2 is bigger.
+        LASSERT(lcm != NULL);
+
+        /*
+         * We want to send one page of cookies with rpc header. This buffer
+         * will be assigned later to the rpc, this is why we preserve the
+         * space for rpc header.
          */
-        llcd_size = CFS_PAGE_SIZE - 
-                lustre_msg_size(LUSTRE_MSG_MAGIC_V2, 1, NULL);
-        llcd_size += offsetof(struct llog_canceld_ctxt, llcd_cookies);
-        OBD_SLAB_ALLOC(llcd, llcd_cache, CFS_ALLOC_STD, llcd_size);
+        size = CFS_PAGE_SIZE - lustre_msg_size(LUSTRE_MSG_MAGIC_V2, 1, NULL);
+        overhead =  offsetof(struct llog_canceld_ctxt, llcd_cookies);
+       OBD_SLAB_ALLOC_GFP(llcd, llcd_cache, size + overhead, CFS_ALLOC_STD);
         if (!llcd)
                 return NULL;
 
-        llcd->llcd_size = llcd_size;
+        CFS_INIT_LIST_HEAD(&llcd->llcd_list);
         llcd->llcd_cookiebytes = 0;
-        atomic_inc(&llcd_count);
+        llcd->llcd_size = size;
+
+        cfs_spin_lock(&lcm->lcm_lock);
+        llcd->llcd_lcm = lcm;
+        cfs_atomic_inc(&lcm->lcm_count);
+        cfs_list_add_tail(&llcd->llcd_list, &lcm->lcm_llcds);
+        cfs_spin_unlock(&lcm->lcm_lock);
+        cfs_atomic_inc(&llcd_count);
+
+        CDEBUG(D_RPCTRACE, "Alloc llcd %p on lcm %p (%d)\n",
+               llcd, lcm, cfs_atomic_read(&lcm->lcm_count));
+
         return llcd;
 }
 
@@ -106,40 +130,53 @@ static struct llog_canceld_ctxt *llcd_alloc(void)
  */
 static void llcd_free(struct llog_canceld_ctxt *llcd)
 {
-        OBD_SLAB_FREE(llcd, llcd_cache, llcd->llcd_size);
-        atomic_dec(&llcd_count);
-}
+        struct llog_commit_master *lcm = llcd->llcd_lcm;
+        int size;
+
+        if (lcm) {
+                if (cfs_atomic_read(&lcm->lcm_count) == 0) {
+                        CERROR("Invalid llcd free %p\n", llcd);
+                        llcd_print(llcd, __FUNCTION__, __LINE__);
+                        LBUG();
+                }
+                cfs_spin_lock(&lcm->lcm_lock);
+                LASSERT(!cfs_list_empty(&llcd->llcd_list));
+                cfs_list_del_init(&llcd->llcd_list);
+                cfs_atomic_dec(&lcm->lcm_count);
+                cfs_spin_unlock(&lcm->lcm_lock);
+
+                CDEBUG(D_RPCTRACE, "Free llcd %p on lcm %p (%d)\n",
+                       llcd, lcm, cfs_atomic_read(&lcm->lcm_count));
+        }
 
-/**
- * Copy passed @cookies to @llcd.
- */
-static void llcd_copy(struct llog_canceld_ctxt *llcd, 
-                      struct llog_cookie *cookies)
-{
-        memcpy((char *)llcd->llcd_cookies + llcd->llcd_cookiebytes, 
-              cookies, sizeof(*cookies));
-        llcd->llcd_cookiebytes += sizeof(*cookies);
+        LASSERT(cfs_atomic_read(&llcd_count) > 0);
+        cfs_atomic_dec(&llcd_count);
+
+        size = offsetof(struct llog_canceld_ctxt, llcd_cookies) +
+            llcd->llcd_size;
+        OBD_SLAB_FREE(llcd, llcd_cache, size);
 }
 
 /**
  * Checks if passed cookie fits into llcd free space buffer. Returns
  * 1 if yes and 0 otherwise.
  */
-static int llcd_fit(struct llog_canceld_ctxt *llcd,
-                 struct llog_cookie *cookies)
+static inline int
+llcd_fit(struct llog_canceld_ctxt *llcd, struct llog_cookie *cookies)
 {
-        return (llcd->llcd_size - 
-                llcd->llcd_cookiebytes) >= sizeof(*cookies);
+        return (llcd->llcd_size - llcd->llcd_cookiebytes >= sizeof(*cookies));
 }
 
-static void llcd_print(struct llog_canceld_ctxt *llcd, 
-                       const char *func, int line) 
+/**
+ * Copy passed @cookies to @llcd.
+ */
+static inline void
+llcd_copy(struct llog_canceld_ctxt *llcd, struct llog_cookie *cookies)
 {
-        CDEBUG(D_RPCTRACE, "Llcd (%p) at %s:%d:\n", llcd, func, line);
-        CDEBUG(D_RPCTRACE, "  size: %d\n", llcd->llcd_size);
-        CDEBUG(D_RPCTRACE, "  ctxt: %p\n", llcd->llcd_ctxt);
-        CDEBUG(D_RPCTRACE, "  lcm : %p\n", llcd->llcd_lcm);
-        CDEBUG(D_RPCTRACE, "  cookiebytes : %d\n", llcd->llcd_cookiebytes);
+        LASSERT(llcd_fit(llcd, cookies));
+        memcpy((char *)llcd->llcd_cookies + llcd->llcd_cookiebytes,
+              cookies, sizeof(*cookies));
+        llcd->llcd_cookiebytes += sizeof(*cookies);
 }
 
 /**
@@ -147,18 +184,21 @@ static void llcd_print(struct llog_canceld_ctxt *llcd,
  * sending result. Error is passed in @rc. Note, that this will be called
  * in cleanup time when all inflight rpcs aborted.
  */
-static int 
-llcd_interpret(struct ptlrpc_request *req, void *noused, int rc)
+static int
+llcd_interpret(const struct lu_env *env,
+               struct ptlrpc_request *req, void *args, int rc)
 {
-        struct llog_canceld_ctxt *llcd = req->rq_async_args.pointer_arg[0];
-        CDEBUG(D_RPCTRACE, "Sent llcd %p (%d)\n", llcd, rc);
+        struct llcd_async_args *la = args;
+        struct llog_canceld_ctxt *llcd = la->la_ctxt;
+
+        CDEBUG(D_RPCTRACE, "Sent llcd %p (%d) - killing it\n", llcd, rc);
         llcd_free(llcd);
         return 0;
 }
+
 /**
  * Send @llcd to remote node. Free llcd uppon completion or error. Sending
- * is performed in async style so this function will return asap without 
+ * is performed in async style so this function will return asap without
  * blocking.
  */
 static int llcd_send(struct llog_canceld_ctxt *llcd)
@@ -166,6 +206,7 @@ static int llcd_send(struct llog_canceld_ctxt *llcd)
         char *bufs[2] = { NULL, (char *)llcd->llcd_cookies };
         struct obd_import *import = NULL;
         struct llog_commit_master *lcm;
+        struct llcd_async_args *la;
         struct ptlrpc_request *req;
         struct llog_ctxt *ctxt;
         int rc;
@@ -173,31 +214,31 @@ static int llcd_send(struct llog_canceld_ctxt *llcd)
 
         ctxt = llcd->llcd_ctxt;
         if (!ctxt) {
-                CERROR("Invalid llcd with NULL ctxt found (%p)\n", 
+                CERROR("Invalid llcd with NULL ctxt found (%p)\n",
                        llcd);
                 llcd_print(llcd, __FUNCTION__, __LINE__);
                 LBUG();
         }
-        LASSERT_SEM_LOCKED(&ctxt->loc_sem);
+        LASSERT_MUTEX_LOCKED(&ctxt->loc_mutex);
 
         if (llcd->llcd_cookiebytes == 0)
                 GOTO(exit, rc = 0);
 
         lcm = llcd->llcd_lcm;
-        
-        /* 
+
+        /*
          * Check if we're in exit stage. Do not send llcd in
-         * this case. 
+         * this case.
          */
-        if (test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags))
+        if (cfs_test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags))
                 GOTO(exit, rc = -ENODEV);
 
         CDEBUG(D_RPCTRACE, "Sending llcd %p\n", llcd);
 
         import = llcd->llcd_ctxt->loc_imp;
-        if (!import || (import == LP_POISON) || 
+        if (!import || (import == LP_POISON) ||
             (import->imp_client == LP_POISON)) {
-                CERROR("Invalid import %p for llcd %p\n", 
+                CERROR("Invalid import %p for llcd %p\n",
                        import, llcd);
                 GOTO(exit, rc = -ENODEV);
         }
@@ -205,12 +246,12 @@ static int llcd_send(struct llog_canceld_ctxt *llcd)
         OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_RECOV, 10);
 
         /*
-         * No need to get import here as it is already done in 
+         * No need to get import here as it is already done in
          * llog_receptor_accept().
          */
         req = ptlrpc_request_alloc(import, &RQF_LOG_CANCEL);
         if (req == NULL) {
-                CERROR("Can't allocate request for sending llcd %p\n", 
+                CERROR("Can't allocate request for sending llcd %p\n",
                        llcd);
                 GOTO(exit, rc = -ENOMEM);
         }
@@ -230,12 +271,18 @@ static int llcd_send(struct llog_canceld_ctxt *llcd)
         /* bug 5515 */
         req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
         req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
-        req->rq_interpret_reply = llcd_interpret;
-        req->rq_async_args.pointer_arg[0] = llcd;
-        rc = ptlrpc_set_add_new_req(&lcm->lcm_pc, req);
-        if (rc)
-                GOTO(exit, rc);
 
+        req->rq_interpret_reply = (ptlrpc_interpterer_t)llcd_interpret;
+
+        CLASSERT(sizeof(*la) <= sizeof(req->rq_async_args));
+        la = ptlrpc_req_async_args(req);
+        la->la_ctxt = llcd;
+
+        /* llog cancels will be replayed after reconnect so this will do twice
+         * first from replay llog, second for resended rpc */
+        req->rq_no_delay = req->rq_no_resend = 1;
+
+        ptlrpc_set_add_new_req(&lcm->lcm_pc, req);
         RETURN(0);
 exit:
         CDEBUG(D_RPCTRACE, "Refused llcd %p\n", llcd);
@@ -250,18 +297,15 @@ exit:
 static int
 llcd_attach(struct llog_ctxt *ctxt, struct llog_canceld_ctxt *llcd)
 {
-        struct llog_commit_master *lcm;
-
         LASSERT(ctxt != NULL && llcd != NULL);
-        LASSERT_SEM_LOCKED(&ctxt->loc_sem);
+        LASSERT_MUTEX_LOCKED(&ctxt->loc_mutex);
         LASSERT(ctxt->loc_llcd == NULL);
-        lcm = ctxt->loc_lcm;
-        atomic_inc(&lcm->lcm_count);
-        CDEBUG(D_RPCTRACE, "Attach llcd %p to ctxt %p (%d)\n",
-               llcd, ctxt, atomic_read(&lcm->lcm_count));
         llcd->llcd_ctxt = llog_ctxt_get(ctxt);
-        llcd->llcd_lcm = ctxt->loc_lcm;
         ctxt->loc_llcd = llcd;
+
+        CDEBUG(D_RPCTRACE, "Attach llcd %p to ctxt %p\n",
+               llcd, ctxt);
+
         return 0;
 }
 
@@ -271,28 +315,19 @@ llcd_attach(struct llog_ctxt *ctxt, struct llog_canceld_ctxt *llcd)
  */
 static struct llog_canceld_ctxt *llcd_detach(struct llog_ctxt *ctxt)
 {
-        struct llog_commit_master *lcm;
         struct llog_canceld_ctxt *llcd;
 
         LASSERT(ctxt != NULL);
-        LASSERT_SEM_LOCKED(&ctxt->loc_sem);
+        LASSERT_MUTEX_LOCKED(&ctxt->loc_mutex);
 
         llcd = ctxt->loc_llcd;
         if (!llcd)
                 return NULL;
 
-        lcm = ctxt->loc_lcm;
-        if (atomic_read(&lcm->lcm_count) == 0) {
-                CERROR("Invalid detach occured %p:%p\n", ctxt, llcd);
-                llcd_print(llcd, __FUNCTION__, __LINE__);
-                LBUG();
-        }
-        atomic_dec(&lcm->lcm_count);
-        ctxt->loc_llcd = NULL;
-        
-        CDEBUG(D_RPCTRACE, "Detach llcd %p from ctxt %p (%d)\n", 
-               llcd, ctxt, atomic_read(&lcm->lcm_count));
+        CDEBUG(D_RPCTRACE, "Detach llcd %p from ctxt %p\n",
+               llcd, ctxt);
 
+        ctxt->loc_llcd = NULL;
         llog_ctxt_put(ctxt);
         return llcd;
 }
@@ -304,10 +339,10 @@ static struct llog_canceld_ctxt *llcd_detach(struct llog_ctxt *ctxt)
 static struct llog_canceld_ctxt *llcd_get(struct llog_ctxt *ctxt)
 {
         struct llog_canceld_ctxt *llcd;
-
-        llcd = llcd_alloc();
+        LASSERT(ctxt);
+        llcd = llcd_alloc(ctxt->loc_lcm);
         if (!llcd) {
-                CERROR("Couldn't alloc an llcd for ctxt %p\n", ctxt);
+                CERROR("Can't alloc an llcd for ctxt %p\n", ctxt);
                 return NULL;
         }
         llcd_attach(ctxt, llcd);
@@ -336,7 +371,7 @@ static int llcd_push(struct llog_ctxt *ctxt)
         int rc;
 
         /*
-         * Make sure that this llcd will not be sent again as we detach 
+         * Make sure that this llcd will not be sent again as we detach
          * it from ctxt.
          */
         llcd = llcd_detach(ctxt);
@@ -345,15 +380,13 @@ static int llcd_push(struct llog_ctxt *ctxt)
                 llcd_print(llcd, __FUNCTION__, __LINE__);
                 LBUG();
         }
-        
+
         rc = llcd_send(llcd);
         if (rc)
                 CERROR("Couldn't send llcd %p (%d)\n", llcd, rc);
         return rc;
 }
 
-static atomic_t llog_tcount = ATOMIC_INIT(0);
-
 /**
  * Start recovery thread which actually deals llcd sending. This
  * is all ptlrpc standard thread based so there is not much of work
@@ -364,15 +397,12 @@ int llog_recov_thread_start(struct llog_commit_master *lcm)
         int rc;
         ENTRY;
 
-        rc = ptlrpcd_start(lcm->lcm_name, &lcm->lcm_pc);
+        rc = ptlrpcd_start(-1, 1, lcm->lcm_name, &lcm->lcm_pc);
         if (rc) {
-                CERROR("Error %d while starting recovery thread %s\n", 
+                CERROR("Error %d while starting recovery thread %s\n",
                        rc, lcm->lcm_name);
                 RETURN(rc);
         }
-        lcm->lcm_set = lcm->lcm_pc.pc_set;
-        atomic_inc(&llog_tcount);
-
         RETURN(rc);
 }
 EXPORT_SYMBOL(llog_recov_thread_start);
@@ -383,25 +413,51 @@ EXPORT_SYMBOL(llog_recov_thread_start);
 void llog_recov_thread_stop(struct llog_commit_master *lcm, int force)
 {
         ENTRY;
-        
-        /**
-         * Let all know that we're stopping. This will also make 
+
+        /*
+         * Let all know that we're stopping. This will also make
          * llcd_send() refuse any new llcds.
          */
-        set_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags);
+        cfs_set_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags);
 
-        /**
+        /*
          * Stop processing thread. No new rpcs will be accepted for
          * for processing now.
          */
         ptlrpcd_stop(&lcm->lcm_pc, force);
-        
+
         /*
-         * No llcds on this @lcm should left.
+         * By this point no alive inflight llcds should be left. Only
+         * those forgotten in sync may still be attached to ctxt. Let's
+         * print them.
          */
-        LASSERTF(atomic_read(&lcm->lcm_count) == 0, 
-                 "Busy llcds found on lcm %p - (%d)\n", 
-                 lcm, atomic_read(&lcm->lcm_count));
+        if (cfs_atomic_read(&lcm->lcm_count) != 0) {
+                struct llog_canceld_ctxt *llcd;
+                cfs_list_t               *tmp;
+
+                CERROR("Busy llcds found (%d) on lcm %p\n",
+                       cfs_atomic_read(&lcm->lcm_count), lcm);
+
+                cfs_spin_lock(&lcm->lcm_lock);
+                cfs_list_for_each(tmp, &lcm->lcm_llcds) {
+                        llcd = cfs_list_entry(tmp, struct llog_canceld_ctxt,
+                                              llcd_list);
+                        llcd_print(llcd, __FUNCTION__, __LINE__);
+                }
+                cfs_spin_unlock(&lcm->lcm_lock);
+
+                /*
+                 * No point to go further with busy llcds at this point
+                 * as this is clear bug. It might mean we got hanging
+                 * rpc which holds import ref and this means we will not
+                 * be able to cleanup anyways.
+                 *
+                 * Or we just missed to kill them when they were not
+                 * attached to ctxt. In this case our slab will remind
+                 * us about this a bit later.
+                 */
+                LBUG();
+        }
         EXIT;
 }
 EXPORT_SYMBOL(llog_recov_thread_stop);
@@ -420,14 +476,15 @@ struct llog_commit_master *llog_recov_thread_init(char *name)
                 RETURN(NULL);
 
         /*
-         * Try to create threads with unique names and user id.
+         * Try to create threads with unique names.
          */
-        snprintf(lcm->lcm_name, sizeof(lcm->lcm_name), 
-                 "ll_log_commit_%s_%02d", name, 
-                 atomic_read(&llog_tcount));
+        snprintf(lcm->lcm_name, sizeof(lcm->lcm_name),
+                 "lcm_%s", name);
 
-        strncpy(lcm->lcm_name, name, sizeof(lcm->lcm_name));
-        atomic_set(&lcm->lcm_count, 0);
+        cfs_atomic_set(&lcm->lcm_count, 0);
+        cfs_atomic_set(&lcm->lcm_refcount, 1);
+        cfs_spin_lock_init(&lcm->lcm_lock);
+        CFS_INIT_LIST_HEAD(&lcm->lcm_llcds);
         rc = llog_recov_thread_start(lcm);
         if (rc) {
                 CERROR("Can't start commit thread, rc %d\n", rc);
@@ -447,13 +504,13 @@ void llog_recov_thread_fini(struct llog_commit_master *lcm, int force)
 {
         ENTRY;
         llog_recov_thread_stop(lcm, force);
-        OBD_FREE_PTR(lcm);
+        lcm_put(lcm);
         EXIT;
 }
 EXPORT_SYMBOL(llog_recov_thread_fini);
 
-static int llog_obd_repl_generic(struct llog_ctxt *ctxt, 
-                                 void *handle, void *arg)
+static int llog_recov_thread_replay(struct llog_ctxt *ctxt,
+                                    void *cb, void *arg)
 {
         struct obd_device *obd = ctxt->loc_obd;
         struct llog_process_cat_args *lpca;
@@ -470,7 +527,7 @@ static int llog_obd_repl_generic(struct llog_ctxt *ctxt,
         if (!lpca)
                 RETURN(-ENOMEM);
 
-        lpca->lpca_cb = handle;
+        lpca->lpca_cb = cb;
         lpca->lpca_arg = arg;
 
         /*
@@ -481,8 +538,7 @@ static int llog_obd_repl_generic(struct llog_ctxt *ctxt,
                 OBD_FREE_PTR(lpca);
                 RETURN(-ENODEV);
         }
-        rc = cfs_kernel_thread(llog_cat_process_thread, lpca, 
-                               CLONE_VM | CLONE_FILES);
+        rc = cfs_create_thread(llog_cat_process_thread, lpca, CFS_DAEMON_FLAGS);
         if (rc < 0) {
                 CERROR("Error starting llog_cat_process_thread(): %d\n", rc);
                 OBD_FREE_PTR(lpca);
@@ -495,45 +551,32 @@ static int llog_obd_repl_generic(struct llog_ctxt *ctxt,
         RETURN(rc);
 }
 
-int llog_obd_repl_connect(struct llog_ctxt *ctxt, int count,
+int llog_obd_repl_connect(struct llog_ctxt *ctxt,
                           struct llog_logid *logid, struct llog_gen *gen,
                           struct obd_uuid *uuid)
 {
-        struct llog_canceld_ctxt *llcd;
         int rc;
         ENTRY;
 
-        mutex_down(&ctxt->loc_sem);
-
-        /* 
-         * Send back cached llcd before recovery from llog if we have any. 
+        /*
+         * Send back cached llcd from llog before recovery if we have any.
+         * This is void is nothing cached is found there.
          */
-        if (ctxt->loc_llcd) {
-                CWARN("Llcd %p:%p is not empty\n", ctxt->loc_llcd, ctxt);
-                mutex_up(&ctxt->loc_sem);
-                llog_sync(ctxt, NULL);
-                mutex_down(&ctxt->loc_sem);
-        }
-
-        llcd = llcd_get(ctxt);
-        if (!llcd) {
-                mutex_up(&ctxt->loc_sem);
-                RETURN(-ENOMEM);
-        }
+       llog_sync(ctxt, NULL, 0);
 
+        /*
+         * Start recovery in separate thread.
+         */
+        cfs_mutex_lock(&ctxt->loc_mutex);
         ctxt->loc_gen = *gen;
+        rc = llog_recov_thread_replay(ctxt, ctxt->llog_proc_cb, logid);
+        cfs_mutex_unlock(&ctxt->loc_mutex);
 
-        rc = llog_obd_repl_generic(ctxt, ctxt->llog_proc_cb, logid);
-        if (rc != 0) {
-                llcd_put(ctxt);
-                CERROR("Error recovery process: %d\n", rc);
-        }
-        mutex_up(&ctxt->loc_sem);
         RETURN(rc);
 }
 EXPORT_SYMBOL(llog_obd_repl_connect);
 
-/** 
+/**
  * Deleted objects have a commit callback that cancels the MDS
  * log record for the deletion. The commit callback calls this
  * function.
@@ -542,30 +585,37 @@ int llog_obd_repl_cancel(struct llog_ctxt *ctxt,
                          struct lov_stripe_md *lsm, int count,
                          struct llog_cookie *cookies, int flags)
 {
+        struct llog_commit_master *lcm;
         struct llog_canceld_ctxt *llcd;
         int rc = 0;
         ENTRY;
 
         LASSERT(ctxt != NULL);
 
-        mutex_down(&ctxt->loc_sem);
-        
+        cfs_mutex_lock(&ctxt->loc_mutex);
+        if (!ctxt->loc_lcm) {
+                CDEBUG(D_RPCTRACE, "No lcm for ctxt %p\n", ctxt);
+                GOTO(out, rc = -ENODEV);
+        }
+        lcm = ctxt->loc_lcm;
+        CDEBUG(D_INFO, "cancel on lsm %p\n", lcm);
+
         /*
          * Let's check if we have all structures alive. We also check for
          * possible shutdown. Do nothing if we're stopping.
          */
-        if (ctxt->loc_imp == NULL) {
-                CDEBUG(D_RPCTRACE, "No import for ctxt %p\n", ctxt);
-                GOTO(out, rc = -ENODEV);
-        }
+       if (ctxt->loc_flags & LLOG_CTXT_FLAG_STOP) {
+               CDEBUG(D_RPCTRACE, "Last sync was done for ctxt %p\n", ctxt);
+               GOTO(out, rc = -ENODEV);
+       }
 
-        if (ctxt->loc_obd->obd_stopping) {
-                CDEBUG(D_RPCTRACE, "Obd is stopping for ctxt %p\n", ctxt);
+       if (ctxt->loc_imp == NULL) {
+                CDEBUG(D_RPCTRACE, "No import for ctxt %p\n", ctxt);
                 GOTO(out, rc = -ENODEV);
         }
 
-        if (test_bit(LLOG_LCM_FL_EXIT, &ctxt->loc_lcm->lcm_flags)) {
-                CDEBUG(D_RPCTRACE, "Commit thread is stopping for ctxt %p\n", 
+        if (cfs_test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags)) {
+                CDEBUG(D_RPCTRACE, "Commit thread is stopping for ctxt %p\n",
                        ctxt);
                 GOTO(out, rc = -ENODEV);
         }
@@ -574,17 +624,23 @@ int llog_obd_repl_cancel(struct llog_ctxt *ctxt,
 
         if (count > 0 && cookies != NULL) {
                 /*
-                 * Get new llcd from ctxt if required. 
+                 * Get new llcd from ctxt if required.
                  */
                 if (!llcd) {
                         llcd = llcd_get(ctxt);
                         if (!llcd)
                                 GOTO(out, rc = -ENOMEM);
+                        /*
+                         * Allocation is successful, let's check for stop
+                         * flag again to fall back as soon as possible.
+                         */
+                        if (cfs_test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags))
+                                GOTO(out, rc = -ENODEV);
                 }
 
                 /*
-                 * Llcd does not have enough room for @cookies. Let's push 
-                 * it out and allocate new one. 
+                 * Llcd does not have enough room for @cookies. Let's push
+                 * it out and allocate new one.
                  */
                 if (!llcd_fit(llcd, cookies)) {
                         rc = llcd_push(ctxt);
@@ -593,45 +649,76 @@ int llog_obd_repl_cancel(struct llog_ctxt *ctxt,
                         llcd = llcd_get(ctxt);
                         if (!llcd)
                                 GOTO(out, rc = -ENOMEM);
+                        /*
+                         * Allocation is successful, let's check for stop
+                         * flag again to fall back as soon as possible.
+                         */
+                        if (cfs_test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags))
+                                GOTO(out, rc = -ENODEV);
                 }
 
                 /*
-                 * Copy cookies to @llcd, no matter old or new allocated one.
+                 * Copy cookies to @llcd, no matter old or new allocated
+                 * one.
                  */
                 llcd_copy(llcd, cookies);
         }
 
         /*
-         * Let's check if we need to send copied @cookies asap. If yes - do it.
+         * Let's check if we need to send copied @cookies asap. If yes
+         * then do it.
          */
         if (llcd && (flags & OBD_LLOG_FL_SENDNOW)) {
+                CDEBUG(D_RPCTRACE, "Sync llcd %p\n", llcd);
                 rc = llcd_push(ctxt);
                 if (rc)
                         GOTO(out, rc);
         }
         EXIT;
 out:
-        mutex_up(&ctxt->loc_sem);
+        if (rc)
+                llcd_put(ctxt);
+
+       if (flags & OBD_LLOG_FL_EXIT)
+               ctxt->loc_flags = LLOG_CTXT_FLAG_STOP;
+
+        cfs_mutex_unlock(&ctxt->loc_mutex);
         return rc;
 }
 EXPORT_SYMBOL(llog_obd_repl_cancel);
 
-int llog_obd_repl_sync(struct llog_ctxt *ctxt, struct obd_export *exp)
+int llog_obd_repl_sync(struct llog_ctxt *ctxt, struct obd_export *exp,
+                      int flags)
 {
         int rc = 0;
         ENTRY;
 
+        /*
+         * Flush any remaining llcd.
+         */
+        cfs_mutex_lock(&ctxt->loc_mutex);
         if (exp && (ctxt->loc_imp == exp->exp_imp_reverse)) {
-                CDEBUG(D_RPCTRACE, "Reverse import disconnect\n");
                 /*
-                 * Check for llcd which might be left attached to @ctxt.
-                 * Let's kill it.
+                 * This is ost->mds connection, we can't be sure that mds
+                 * can still receive cookies, let's killed the cached llcd.
                  */
-                mutex_down(&ctxt->loc_sem);
+                CDEBUG(D_RPCTRACE, "Kill cached llcd\n");
                 llcd_put(ctxt);
-                mutex_up(&ctxt->loc_sem);
+
+               if (flags & OBD_LLOG_FL_EXIT)
+                       ctxt->loc_flags = LLOG_CTXT_FLAG_STOP;
+
+                cfs_mutex_unlock(&ctxt->loc_mutex);
         } else {
-                rc = llog_cancel(ctxt, NULL, 0, NULL, OBD_LLOG_FL_SENDNOW);
+                /*
+                 * This is either llog_sync() from generic llog code or sync
+                 * on client disconnect. In either way let's do it and send
+                 * llcds to the target with waiting for completion.
+                 */
+                CDEBUG(D_RPCTRACE, "Sync cached llcd\n");
+                cfs_mutex_unlock(&ctxt->loc_mutex);
+               rc = llog_cancel(ctxt, NULL, 0, NULL, OBD_LLOG_FL_SENDNOW |
+                                flags);
         }
         RETURN(rc);
 }
@@ -654,7 +741,7 @@ int llog_recov_init(void)
 {
         int llcd_size;
 
-        llcd_size = CFS_PAGE_SIZE - 
+        llcd_size = CFS_PAGE_SIZE -
                 lustre_msg_size(LUSTRE_MSG_MAGIC_V2, 1, NULL);
         llcd_size += offsetof(struct llog_canceld_ctxt, llcd_cookies);
         llcd_cache = cfs_mem_cache_create("llcd_cache", llcd_size, 0, 0);
@@ -670,10 +757,8 @@ int llog_recov_init(void)
  */
 void llog_recov_fini(void)
 {
-        int count;
-
         /*
-         * Kill llcd cache when thread is stopped and we're sure no 
+         * Kill llcd cache when thread is stopped and we're sure no
          * llcd in use left.
          */
         if (llcd_cache) {
@@ -681,9 +766,9 @@ void llog_recov_fini(void)
                  * In 2.6.22 cfs_mem_cache_destroy() will not return error
                  * for busy resources. Let's check it another way.
                  */
-                count = atomic_read(&llcd_count);
-                LASSERTF(count == 0, "Can't destroy llcd cache! Number of "
-                         "busy llcds: %d\n", count);
+                LASSERTF(cfs_atomic_read(&llcd_count) == 0,
+                         "Can't destroy llcd cache! Number of "
+                         "busy llcds: %d\n", cfs_atomic_read(&llcd_count));
                 cfs_mem_cache_destroy(llcd_cache);
                 llcd_cache = NULL;
         }