Whamcloud - gitweb
b=19188
[fs/lustre-release.git] / lustre / osc / osc_lock.c
index 8f2b2b5..c9d8e31 100644 (file)
@@ -210,8 +210,7 @@ static void osc_lock_fini(const struct lu_env *env,
          */
         if (ols->ols_hold)
                 osc_lock_unuse(env, slice);
-        if (ols->ols_lock != NULL)
-                osc_lock_detach(env, ols);
+        LASSERT(ols->ols_lock == NULL);
 
         OBD_SLAB_FREE_PTR(ols, osc_lock_kmem);
 }
@@ -243,6 +242,7 @@ static void osc_lock_build_policy(const struct lu_env *env,
         const struct cl_lock_descr *d = &lock->cll_descr;
 
         osc_index2policy(policy, d->cld_obj, d->cld_start, d->cld_end);
+        policy->l_extent.gid = d->cld_gid;
 }
 
 static int osc_enq2ldlm_flags(__u32 enqflags)
@@ -312,6 +312,8 @@ static void osc_ast_data_put(const struct lu_env *env, struct osc_lock *olck)
  *
  * This can be optimized to not update attributes when lock is a result of a
  * local match.
+ *
+ * Called under lock and resource spin-locks.
  */
 static void osc_lock_lvb_update(const struct lu_env *env, struct osc_lock *olck,
                                 int rc)
@@ -344,6 +346,8 @@ static void osc_lock_lvb_update(const struct lu_env *env, struct osc_lock *olck,
                 dlmlock = olck->ols_lock;
                 LASSERT(dlmlock != NULL);
 
+                /* re-grab LVB from a dlm lock under DLM spin-locks. */
+                *lvb = *(struct ost_lvb *)dlmlock->l_lvb_data;
                 size = lvb->lvb_size;
                 /* Extend KMS up to the end of this lock and no further
                  * A lock on [x,y] means a KMS of up to y + 1 bytes! */
@@ -360,7 +364,7 @@ static void osc_lock_lvb_update(const struct lu_env *env, struct osc_lock *olck,
                                    lvb->lvb_size, oinfo->loi_kms,
                                    dlmlock->l_policy_data.l_extent.end);
                 }
-                ldlm_lock_allow_match(dlmlock);
+                ldlm_lock_allow_match_locked(dlmlock);
         } else if (rc == -ENAVAIL && olck->ols_glimpse) {
                 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
                        " kms="LPU64"\n", lvb->lvb_size, oinfo->loi_kms);
@@ -375,6 +379,13 @@ static void osc_lock_lvb_update(const struct lu_env *env, struct osc_lock *olck,
         EXIT;
 }
 
+/**
+ * Called when a lock is granted, from an upcall (when server returned a
+ * granted lock), or from completion AST, when server returned a blocked lock.
+ *
+ * Called under lock and resource spin-locks, that are released temporarily
+ * here.
+ */
 static void osc_lock_granted(const struct lu_env *env, struct osc_lock *olck,
                              struct ldlm_lock *dlmlock, int rc)
 {
@@ -395,15 +406,24 @@ static void osc_lock_granted(const struct lu_env *env, struct osc_lock *olck,
                 descr->cld_mode  = osc_ldlm2cl_lock(dlmlock->l_granted_mode);
                 descr->cld_start = cl_index(descr->cld_obj, ext->start);
                 descr->cld_end   = cl_index(descr->cld_obj, ext->end);
+                descr->cld_gid   = ext->gid;
                 /*
                  * tell upper layers the extent of the lock that was actually
                  * granted
                  */
-                cl_lock_modify(env, lock, descr);
-                LINVRNT(osc_lock_invariant(olck));
                 olck->ols_state = OLS_GRANTED;
                 osc_lock_lvb_update(env, olck, rc);
+
+                /* release DLM spin-locks to allow cl_lock_{modify,signal}()
+                 * to take a semaphore on a parent lock. This is safe, because
+                 * spin-locks are needed to protect consistency of
+                 * dlmlock->l_*_mode and LVB, and we have finished processing
+                 * them. */
+                unlock_res_and_lock(dlmlock);
+                cl_lock_modify(env, lock, descr);
                 cl_lock_signal(env, lock);
+                LINVRNT(osc_lock_invariant(olck));
+                lock_res_and_lock(dlmlock);
         }
         EXIT;
 }
@@ -424,7 +444,6 @@ static void osc_lock_upcall0(const struct lu_env *env, struct osc_lock *olck)
         LASSERT(olck->ols_lock == NULL);
         olck->ols_lock = dlmlock;
         spin_unlock(&osc_ast_guard);
-        unlock_res_and_lock(dlmlock);
 
         /*
          * Lock might be not yet granted. In this case, completion ast
@@ -433,6 +452,8 @@ static void osc_lock_upcall0(const struct lu_env *env, struct osc_lock *olck)
          */
         if (dlmlock->l_granted_mode == dlmlock->l_req_mode)
                 osc_lock_granted(env, olck, dlmlock, 0);
+        unlock_res_and_lock(dlmlock);
+
         /*
          * osc_enqueue_interpret() decrefs asynchronous locks, counter
          * this.
@@ -452,18 +473,14 @@ static void osc_lock_upcall0(const struct lu_env *env, struct osc_lock *olck)
  */
 static int osc_lock_upcall(void *cookie, int errcode)
 {
-        struct osc_lock      *olck  = cookie;
-        struct cl_lock_slice *slice = &olck->ols_cl;
-        struct cl_lock       *lock  = slice->cls_lock;
-        struct lu_env        *env;
-
-        int refcheck;
+        struct osc_lock         *olck  = cookie;
+        struct cl_lock_slice    *slice = &olck->ols_cl;
+        struct cl_lock          *lock  = slice->cls_lock;
+        struct lu_env           *env;
+        struct cl_env_nest       nest;
 
         ENTRY;
-        /*
-         * XXX environment should be created in ptlrpcd.
-         */
-        env = cl_env_get(&refcheck);
+        env = cl_env_nested_get(&nest);
         if (!IS_ERR(env)) {
                 int rc;
 
@@ -529,7 +546,7 @@ static int osc_lock_upcall(void *cookie, int errcode)
                 /* release cookie reference, acquired by osc_lock_enqueue() */
                 lu_ref_del(&lock->cll_reference, "upcall", lock);
                 cl_lock_put(env, lock);
-                cl_env_put(env, &refcheck);
+                cl_env_nested_put(&nest, env);
         } else
                 /* should never happen, similar to osc_ldlm_blocking_ast(). */
                 LBUG();
@@ -704,9 +721,10 @@ static int osc_ldlm_blocking_ast(struct ldlm_lock *dlmlock,
          * new environment has to be created to not corrupt outer context.
          */
         env = cl_env_nested_get(&nest);
-        if (!IS_ERR(env))
+        if (!IS_ERR(env)) {
                 result = osc_dlm_blocking_ast0(env, dlmlock, data, flag);
-        else {
+                cl_env_nested_put(&nest, env);
+        } else {
                 result = PTR_ERR(env);
                 /*
                  * XXX This should never happen, as cl_lock is
@@ -721,26 +739,23 @@ static int osc_ldlm_blocking_ast(struct ldlm_lock *dlmlock,
                 else
                         CERROR("BAST failed: %d\n", result);
         }
-        cl_env_nested_put(&nest, env);
         return result;
 }
 
 static int osc_ldlm_completion_ast(struct ldlm_lock *dlmlock,
                                    int flags, void *data)
 {
-        struct lu_env   *env;
-        void            *env_cookie;
-        struct osc_lock *olck;
-        struct cl_lock  *lock;
-        int refcheck;
+        struct cl_env_nest nest;
+        struct lu_env     *env;
+        struct osc_lock   *olck;
+        struct cl_lock    *lock;
         int result;
         int dlmrc;
 
         /* first, do dlm part of the work */
         dlmrc = ldlm_completion_ast_async(dlmlock, flags, data);
         /* then, notify cl_lock */
-        env_cookie = cl_env_reenter();
-        env = cl_env_get(&refcheck);
+        env = cl_env_nested_get(&nest);
         if (!IS_ERR(env)) {
                 olck = osc_ast_data_get(dlmlock);
                 if (olck != NULL) {
@@ -751,6 +766,7 @@ static int osc_ldlm_completion_ast(struct ldlm_lock *dlmlock,
                          * to lock->l_lvb_data, store it in osc_lock.
                          */
                         LASSERT(dlmlock->l_lvb_data != NULL);
+                        lock_res_and_lock(dlmlock);
                         olck->ols_lvb = *(struct ost_lvb *)dlmlock->l_lvb_data;
                         if (olck->ols_lock == NULL)
                                 /*
@@ -765,6 +781,7 @@ static int osc_ldlm_completion_ast(struct ldlm_lock *dlmlock,
                                 ;
                         else if (dlmlock->l_granted_mode != LCK_MINMODE)
                                 osc_lock_granted(env, olck, dlmlock, dlmrc);
+                        unlock_res_and_lock(dlmlock);
                         if (dlmrc != 0)
                                 cl_lock_error(env, lock, dlmrc);
                         cl_lock_mutex_put(env, lock);
@@ -772,10 +789,9 @@ static int osc_ldlm_completion_ast(struct ldlm_lock *dlmlock,
                         result = 0;
                 } else
                         result = -ELDLM_NO_LOCK_DATA;
-                cl_env_put(env, &refcheck);
+                cl_env_nested_put(&nest, env);
         } else
                 result = PTR_ERR(env);
-        cl_env_reexit(env_cookie);
         return dlmrc ?: result;
 }
 
@@ -785,15 +801,15 @@ static int osc_ldlm_glimpse_ast(struct ldlm_lock *dlmlock, void *data)
         struct osc_lock        *olck;
         struct cl_lock         *lock;
         struct cl_object       *obj;
+        struct cl_env_nest      nest;
         struct lu_env          *env;
         struct ost_lvb         *lvb;
         struct req_capsule     *cap;
         int                     result;
-        int                     refcheck;
 
         LASSERT(lustre_msg_get_opc(req->rq_reqmsg) == LDLM_GL_CALLBACK);
 
-        env = cl_env_get(&refcheck);
+        env = cl_env_nested_get(&nest);
         if (!IS_ERR(env)) {
                 /*
                  * osc_ast_data_get() has to go after environment is
@@ -826,7 +842,7 @@ static int osc_ldlm_glimpse_ast(struct ldlm_lock *dlmlock, void *data)
                         lustre_pack_reply(req, 1, NULL, NULL);
                         result = -ELDLM_NO_LOCK_DATA;
                 }
-                cl_env_put(env, &refcheck);
+                cl_env_nested_put(&nest, env);
         } else
                 result = PTR_ERR(env);
         req->rq_status = result;
@@ -850,16 +866,14 @@ static unsigned long osc_lock_weigh(const struct lu_env *env,
  */
 static unsigned long osc_ldlm_weigh_ast(struct ldlm_lock *dlmlock)
 {
+        struct cl_env_nest       nest;
         struct lu_env           *env;
-        int                      refcheck;
-        void                    *cookie;
         struct osc_lock         *lock;
         struct cl_lock          *cll;
         unsigned long            weight;
         ENTRY;
 
         might_sleep();
-        cookie = cl_env_reenter();
         /*
          * osc_ldlm_weigh_ast has a complex context since it might be called
          * because of lock canceling, or from user's input. We have to make
@@ -867,12 +881,10 @@ static unsigned long osc_ldlm_weigh_ast(struct ldlm_lock *dlmlock)
          * the upper context because cl_lock_put don't modify environment
          * variables. But in case of ..
          */
-        env = cl_env_get(&refcheck);
-        if (IS_ERR(env)) {
+        env = cl_env_nested_get(&nest);
+        if (IS_ERR(env))
                 /* Mostly because lack of memory, tend to eliminate this lock*/
-                cl_env_reexit(cookie);
                 RETURN(0);
-        }
 
         LASSERT(dlmlock->l_resource->lr_type == LDLM_EXTENT);
         lock = osc_ast_data_get(dlmlock);
@@ -892,8 +904,7 @@ static unsigned long osc_ldlm_weigh_ast(struct ldlm_lock *dlmlock)
         EXIT;
 
 out:
-        cl_env_put(env, &refcheck);
-        cl_env_reexit(cookie);
+        cl_env_nested_put(&nest, env);
         return weight;
 }
 
@@ -922,6 +933,20 @@ static void osc_lock_build_einfo(const struct lu_env *env,
         einfo->ei_cbdata = lock; /* value to be put into ->l_ast_data */
 }
 
+static int osc_lock_delete0(struct cl_lock *conflict)
+{
+        struct cl_env_nest    nest;
+        struct lu_env        *env;
+        int    rc = 0;        
+
+        env = cl_env_nested_get(&nest);
+        if (!IS_ERR(env)) {
+                cl_lock_delete(env, conflict);
+                cl_env_nested_put(&nest, env);
+        } else
+                rc = PTR_ERR(env);
+        return rc; 
+}
 /**
  * Cancels \a conflict lock and waits until it reached CLS_FREEING state. This
  * is called as a part of enqueuing to cancel conflicting locks early.
@@ -950,7 +975,9 @@ static int osc_lock_cancel_wait(const struct lu_env *env, struct cl_lock *lock,
         rc = 0;
         if (conflict->cll_state != CLS_FREEING) {
                 cl_lock_cancel(env, conflict);
-                cl_lock_delete(env, conflict);
+                rc = osc_lock_delete0(conflict);
+                if (rc)
+                        return rc; 
                 if (conflict->cll_flags & (CLF_CANCELPEND|CLF_DOOMED)) {
                         rc = -EWOULDBLOCK;
                         if (cl_lock_nr_mutexed(env) > 2)
@@ -1038,6 +1065,7 @@ static void osc_lock_to_lockless(const struct lu_env *env,
                         slice->cls_ops = &osc_lock_lockless_ops;
                 }
         }
+        LASSERT(ergo(ols->ols_glimpse, !osc_lock_is_lockless(ols)));
 }
 
 /**
@@ -1090,6 +1118,14 @@ static int osc_lock_enqueue_wait(const struct lu_env *env,
                         continue;
 
                 /* overlapped and living locks. */
+
+                /* We're not supposed to give up group lock. */
+                if (scan->cll_descr.cld_mode == CLM_GROUP) {
+                        LASSERT(descr->cld_mode != CLM_GROUP ||
+                                descr->cld_gid != scan->cll_descr.cld_gid);
+                        continue;
+                }
+
                 /* A tricky case for lockless pages:
                  * We need to cancel the compatible locks if we're enqueuing
                  * a lockless lock, for example:
@@ -1273,7 +1309,7 @@ static int osc_lock_enqueue(const struct lu_env *env,
                         ols->ols_state = OLS_GRANTED;
                 }
         }
-
+        LASSERT(ergo(ols->ols_glimpse, !osc_lock_is_lockless(ols)));
         RETURN(result);
 }
 
@@ -1324,7 +1360,6 @@ static int osc_lock_use(const struct lu_env *env,
                 lock = slice->cls_lock;
                 LASSERT(lock->cll_state == CLS_CACHED);
                 LASSERT(lock->cll_users > 0);
-                LASSERT(olck->ols_lock->l_flags & LDLM_FL_CBPENDING);
                 /* set a flag for osc_dlm_blocking_ast0() to signal the
                  * lock.*/
                 olck->ols_ast_wait = 1;
@@ -1371,19 +1406,31 @@ static void osc_lock_cancel(const struct lu_env *env,
         struct cl_lock   *lock    = slice->cls_lock;
         struct osc_lock  *olck    = cl2osc_lock(slice);
         struct ldlm_lock *dlmlock = olck->ols_lock;
-        int               result;
+        int               result  = 0;
         int               discard;
 
         LASSERT(cl_lock_is_mutexed(lock));
         LINVRNT(osc_lock_invariant(olck));
 
         if (dlmlock != NULL) {
+                int do_cancel;
+
                 discard = dlmlock->l_flags & LDLM_FL_DISCARD_DATA;
                 result = osc_lock_flush(olck, discard);
                 if (olck->ols_hold)
                         osc_lock_unuse(env, slice);
-                LASSERT(dlmlock->l_readers == 0 && dlmlock->l_writers == 0);
-                result = ldlm_cli_cancel(&olck->ols_handle);
+
+                lock_res_and_lock(dlmlock);
+                /* Now that we're the only user of dlm read/write reference,
+                 * mostly the ->l_readers + ->l_writers should be zero.
+                 * However, there is a corner case.
+                 * See bug 18829 for details.*/
+                do_cancel = (dlmlock->l_readers == 0 &&
+                             dlmlock->l_writers == 0);
+                dlmlock->l_flags |= LDLM_FL_CBPENDING;
+                unlock_res_and_lock(dlmlock);
+                if (do_cancel)
+                        result = ldlm_cli_cancel(&olck->ols_handle);
                 if (result < 0)
                         CL_LOCK_DEBUG(D_ERROR, env, lock,
                                       "lock %p cancel failure with error(%d)\n",
@@ -1614,7 +1661,7 @@ int osc_lock_init(const struct lu_env *env,
         struct osc_lock *clk;
         int result;
 
-        OBD_SLAB_ALLOC_PTR(clk, osc_lock_kmem);
+        OBD_SLAB_ALLOC_PTR_GFP(clk, osc_lock_kmem, CFS_ALLOC_IO);
         if (clk != NULL) {
                 osc_lock_build_einfo(env, lock, clk, &clk->ols_einfo);
                 clk->ols_state = OLS_NEW;