Whamcloud - gitweb
LU-12616 obclass: fix MDS start/stop race
[fs/lustre-release.git] / lustre / ldlm / ldlm_flock.c
index f2a0859..be84993 100644 (file)
  *
  * 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.
+ * http://www.gnu.org/licenses/gpl-2.0.html
  *
  * GPL HEADER END
  */
@@ -31,7 +27,7 @@
  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
  *
- * Copyright (c) 2010, 2014, Intel Corporation.
+ * Copyright (c) 2010, 2017, Intel Corporation.
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
 
 #define DEBUG_SUBSYSTEM S_LDLM
 
+#include <linux/list.h>
 #include <lustre_dlm.h>
 #include <obd_support.h>
 #include <obd_class.h>
 #include <lustre_lib.h>
-#include <libcfs/list.h>
 
 #include "ldlm_internal.h"
 
@@ -108,7 +104,7 @@ static inline void ldlm_flock_blocking_link(struct ldlm_lock *req,
                 lock->l_policy_data.l_flock.owner;
         req->l_policy_data.l_flock.blocking_export =
                lock->l_export;
-       req->l_policy_data.l_flock.blocking_refs = 0;
+       atomic_set(&req->l_policy_data.l_flock.blocking_refs, 0);
 
        cfs_hash_add(req->l_export->exp_flock_hash,
                     &req->l_policy_data.l_flock.owner,
@@ -130,11 +126,11 @@ static inline void ldlm_flock_blocking_unlink(struct ldlm_lock *req)
 }
 
 static inline void
-ldlm_flock_destroy(struct ldlm_lock *lock, ldlm_mode_t mode, __u64 flags)
+ldlm_flock_destroy(struct ldlm_lock *lock, enum ldlm_mode mode, __u64 flags)
 {
        ENTRY;
 
-       LDLM_DEBUG(lock, "ldlm_flock_destroy(mode: %d, flags: "LPX64")",
+       LDLM_DEBUG(lock, "ldlm_flock_destroy(mode: %d, flags: %#llx)",
                   mode, flags);
 
        /* Safe to not lock here, since it should be empty anyway */
@@ -170,7 +166,7 @@ struct ldlm_flock_lookup_cb_data {
        struct obd_export *exp;
 };
 
-static int ldlm_flock_lookup_cb(cfs_hash_t *hs, cfs_hash_bd_t *bd,
+static int ldlm_flock_lookup_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
                                struct hlist_node *hnode, void *data)
 {
        struct ldlm_flock_lookup_cb_data *cb_data = data;
@@ -274,38 +270,31 @@ static void ldlm_flock_cancel_on_deadlock(struct ldlm_lock *lock,
  * This function looks for any conflicts for \a lock in the granted or
  * waiting queues. The lock is granted if no conflicts are found in
  * either queue.
- *
- * It is also responsible for splitting a lock if a portion of the lock
- * is released.
- *
- * If \a first_enq is 0 (ie, called from ldlm_reprocess_queue):
- *   - blocking ASTs have already been sent
- *
- * If \a first_enq is 1 (ie, called from ldlm_lock_enqueue):
- *   - blocking ASTs have not been sent yet, so list of conflicting locks
- *     would be collected and ASTs sent.
  */
 int
-ldlm_process_flock_lock(struct ldlm_lock *req, __u64 *flags, int first_enq,
-                       ldlm_error_t *err, struct list_head *work_list)
+ldlm_process_flock_lock(struct ldlm_lock *req, __u64 *flags,
+                       enum ldlm_process_intention intention,
+                       enum ldlm_error *err, struct list_head *work_list)
 {
-        struct ldlm_resource *res = req->l_resource;
-        struct ldlm_namespace *ns = ldlm_res_to_ns(res);
+       struct ldlm_resource *res = req->l_resource;
+       struct ldlm_namespace *ns = ldlm_res_to_ns(res);
        struct list_head *tmp;
        struct list_head *ownlocks = NULL;
-        struct ldlm_lock *lock = NULL;
-        struct ldlm_lock *new = req;
-        struct ldlm_lock *new2 = NULL;
-        ldlm_mode_t mode = req->l_req_mode;
-        int local = ns_is_client(ns);
-        int added = (mode == LCK_NL);
-        int overlaps = 0;
-        int splitted = 0;
-        const struct ldlm_callback_suite null_cbs = { NULL };
-        ENTRY;
+       struct ldlm_lock *lock = NULL;
+       struct ldlm_lock *new = req;
+       struct ldlm_lock *new2 = NULL;
+       enum ldlm_mode mode = req->l_req_mode;
+       int local = ns_is_client(ns);
+       int added = (mode == LCK_NL);
+       int overlaps = 0;
+       int splitted = 0;
+       const struct ldlm_callback_suite null_cbs = { NULL };
+       struct list_head *grant_work = intention == LDLM_PROCESS_ENQUEUE ?
+                                                       NULL : work_list;
+       ENTRY;
 
-       CDEBUG(D_DLMTRACE, "flags "LPX64" owner "LPU64" pid %u mode %u start "
-              LPU64" end "LPU64"\n", *flags,
+       CDEBUG(D_DLMTRACE, "flags %#llx owner %llu pid %u mode %u start "
+              "%llu end %llu\n", *flags,
               new->l_policy_data.l_flock.owner,
                new->l_policy_data.l_flock.pid, mode,
                req->l_policy_data.l_flock.start,
@@ -357,11 +346,11 @@ reprocess:
                         if (!ldlm_flocks_overlap(lock, req))
                                 continue;
 
-                       if (!first_enq) {
+                       if (intention != LDLM_PROCESS_ENQUEUE) {
                                reprocess_failed = 1;
                                if (ldlm_flock_deadlock(req, lock)) {
                                        ldlm_flock_cancel_on_deadlock(req,
-                                                       work_list);
+                                                       grant_work);
                                        RETURN(LDLM_ITER_CONTINUE);
                                }
                                continue;
@@ -574,7 +563,7 @@ reprocess:
 
         if (*flags != LDLM_FL_WAIT_NOREPROC) {
 #ifdef HAVE_SERVER_SUPPORT
-                if (first_enq) {
+               if (intention == LDLM_PROCESS_ENQUEUE) {
                         /* If this is an unlock, reprocess the waitq and
                          * send completions ASTs for locks that can now be
                          * granted. The only problem with doing this
@@ -582,16 +571,17 @@ reprocess:
                          * newly granted locks will be sent before the unlock
                          * completion is sent. It shouldn't be an issue. Also
                          * note that ldlm_process_flock_lock() will recurse,
-                         * but only once because first_enq will be false from
-                         * ldlm_reprocess_queue. */
+                        * but only once because 'intention' won't be
+                        * LDLM_PROCESS_ENQUEUE from ldlm_reprocess_queue. */
                        if ((mode == LCK_NL) && overlaps) {
                                struct list_head rpc_list;
                                 int rc;
 
                                INIT_LIST_HEAD(&rpc_list);
 restart:
-                                ldlm_reprocess_queue(res, &res->lr_waiting,
-                                                     &rpc_list);
+                               ldlm_reprocess_queue(res, &res->lr_waiting,
+                                                    &rpc_list,
+                                                    LDLM_PROCESS_RESCAN, NULL);
 
                                 unlock_res_and_lock(req);
                                 rc = ldlm_run_ast_work(ns, &rpc_list,
@@ -602,7 +592,7 @@ restart:
                        }
                 } else {
                         LASSERT(req->l_completion_ast);
-                        ldlm_add_ast_work_item(req, NULL, work_list);
+                       ldlm_add_ast_work_item(req, NULL, grant_work);
                 }
 #else /* !HAVE_SERVER_SUPPORT */
                 /* The only one possible case for client-side calls flock
@@ -661,14 +651,14 @@ ldlm_flock_interrupted_wait(void *data)
 int
 ldlm_flock_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
 {
-       struct file_lock                *getlk = lock->l_ast_data;
-        struct obd_device              *obd;
-        struct obd_import              *imp = NULL;
-        struct ldlm_flock_wait_data     fwd;
-        struct l_wait_info              lwi;
-        ldlm_error_t                    err;
-        int                             rc = 0;
-        ENTRY;
+       struct file_lock *getlk = lock->l_ast_data;
+       struct obd_device *obd;
+       struct obd_import *imp = NULL;
+       struct ldlm_flock_wait_data fwd;
+       struct l_wait_info lwi;
+       enum ldlm_error err;
+       int rc = 0;
+       ENTRY;
 
        OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT2, 4);
        if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_CP_CB_WAIT3)) {
@@ -677,7 +667,7 @@ ldlm_flock_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
                unlock_res_and_lock(lock);
                OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT3, 4);
        }
-       CDEBUG(D_DLMTRACE, "flags: "LPX64" data: %p getlk: %p\n",
+       CDEBUG(D_DLMTRACE, "flags: %#llx data: %p getlk: %p\n",
               flags, data, getlk);
 
        LASSERT(flags != LDLM_FL_WAIT_NOREPROC);
@@ -754,7 +744,7 @@ granted:
                RETURN(-EIO);
        }
 
-        /* ldlm_lock_enqueue() has already placed lock on the granted list. */
+       /* ldlm_lock_enqueue() has already placed lock on the granted list. */
        ldlm_resource_unlink_lock(lock);
 
        /* Import invalidation. We need to actually release the lock
@@ -769,7 +759,7 @@ granted:
                        LASSERT(ldlm_is_test_lock(lock));
 
                if (ldlm_is_test_lock(lock) || ldlm_is_flock_deadlock(lock))
-                       mode = flock_type(getlk);
+                       mode = getlk->fl_type;
                else
                        mode = lock->l_granted_mode;
 
@@ -792,27 +782,26 @@ granted:
        LDLM_DEBUG(lock, "client-side enqueue granted");
 
        if (flags & LDLM_FL_TEST_LOCK) {
-                /* fcntl(F_GETLK) request */
-                /* The old mode was saved in getlk->fl_type so that if the mode
-                 * in the lock changes we can decref the appropriate refcount.*/
+               /*
+                * fcntl(F_GETLK) request
+                * The old mode was saved in getlk->fl_type so that if the mode
+                * in the lock changes we can decref the appropriate refcount.
+                */
                LASSERT(ldlm_is_test_lock(lock));
-               ldlm_flock_destroy(lock, flock_type(getlk),
-                                  LDLM_FL_WAIT_NOREPROC);
+               ldlm_flock_destroy(lock, getlk->fl_type, LDLM_FL_WAIT_NOREPROC);
                switch (lock->l_granted_mode) {
                case LCK_PR:
-                       flock_set_type(getlk, F_RDLCK);
+                       getlk->fl_type = F_RDLCK;
                        break;
                case LCK_PW:
-                       flock_set_type(getlk, F_WRLCK);
+                       getlk->fl_type = F_WRLCK;
                        break;
                default:
-                       flock_set_type(getlk, F_UNLCK);
+                       getlk->fl_type = F_UNLCK;
                }
-               flock_set_pid(getlk, (pid_t)lock->l_policy_data.l_flock.pid);
-               flock_set_start(getlk,
-                               (loff_t)lock->l_policy_data.l_flock.start);
-               flock_set_end(getlk,
-                             (loff_t)lock->l_policy_data.l_flock.end);
+               getlk->fl_pid = (pid_t)lock->l_policy_data.l_flock.pid;
+               getlk->fl_start = (loff_t)lock->l_policy_data.l_flock.start;
+               getlk->fl_end = (loff_t)lock->l_policy_data.l_flock.end;
        } else {
                __u64 noreproc = LDLM_FL_WAIT_NOREPROC;
 
@@ -840,45 +829,30 @@ int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
         RETURN(0);
 }
 
-void ldlm_flock_policy_wire18_to_local(const ldlm_wire_policy_data_t *wpolicy,
-                                       ldlm_policy_data_t *lpolicy)
+void ldlm_flock_policy_wire_to_local(const union ldlm_wire_policy_data *wpolicy,
+                                    union ldlm_policy_data *lpolicy)
 {
-        memset(lpolicy, 0, sizeof(*lpolicy));
-        lpolicy->l_flock.start = wpolicy->l_flock.lfw_start;
-        lpolicy->l_flock.end = wpolicy->l_flock.lfw_end;
-        lpolicy->l_flock.pid = wpolicy->l_flock.lfw_pid;
-        /* Compat code, old clients had no idea about owner field and
-         * relied solely on pid for ownership. Introduced in LU-104, 2.1,
-         * April 2011 */
-        lpolicy->l_flock.owner = wpolicy->l_flock.lfw_pid;
+       lpolicy->l_flock.start = wpolicy->l_flock.lfw_start;
+       lpolicy->l_flock.end = wpolicy->l_flock.lfw_end;
+       lpolicy->l_flock.pid = wpolicy->l_flock.lfw_pid;
+       lpolicy->l_flock.owner = wpolicy->l_flock.lfw_owner;
 }
 
-
-void ldlm_flock_policy_wire21_to_local(const ldlm_wire_policy_data_t *wpolicy,
-                                       ldlm_policy_data_t *lpolicy)
+void ldlm_flock_policy_local_to_wire(const union ldlm_policy_data *lpolicy,
+                                    union ldlm_wire_policy_data *wpolicy)
 {
-        memset(lpolicy, 0, sizeof(*lpolicy));
-        lpolicy->l_flock.start = wpolicy->l_flock.lfw_start;
-        lpolicy->l_flock.end = wpolicy->l_flock.lfw_end;
-        lpolicy->l_flock.pid = wpolicy->l_flock.lfw_pid;
-        lpolicy->l_flock.owner = wpolicy->l_flock.lfw_owner;
-}
-
-void ldlm_flock_policy_local_to_wire(const ldlm_policy_data_t *lpolicy,
-                                     ldlm_wire_policy_data_t *wpolicy)
-{
-        memset(wpolicy, 0, sizeof(*wpolicy));
-        wpolicy->l_flock.lfw_start = lpolicy->l_flock.start;
-        wpolicy->l_flock.lfw_end = lpolicy->l_flock.end;
-        wpolicy->l_flock.lfw_pid = lpolicy->l_flock.pid;
-        wpolicy->l_flock.lfw_owner = lpolicy->l_flock.owner;
+       memset(wpolicy, 0, sizeof(*wpolicy));
+       wpolicy->l_flock.lfw_start = lpolicy->l_flock.start;
+       wpolicy->l_flock.lfw_end = lpolicy->l_flock.end;
+       wpolicy->l_flock.lfw_pid = lpolicy->l_flock.pid;
+       wpolicy->l_flock.lfw_owner = lpolicy->l_flock.owner;
 }
 
 /*
  * Export handle<->flock hash operations.
  */
 static unsigned
-ldlm_export_flock_hash(cfs_hash_t *hs, const void *key, unsigned mask)
+ldlm_export_flock_hash(struct cfs_hash *hs, const void *key, unsigned mask)
 {
        return cfs_hash_u64_hash(*(__u64 *)key, mask);
 }
@@ -905,7 +879,7 @@ ldlm_export_flock_object(struct hlist_node *hnode)
 }
 
 static void
-ldlm_export_flock_get(cfs_hash_t *hs, struct hlist_node *hnode)
+ldlm_export_flock_get(struct cfs_hash *hs, struct hlist_node *hnode)
 {
        struct ldlm_lock *lock;
        struct ldlm_flock *flock;
@@ -916,28 +890,28 @@ ldlm_export_flock_get(cfs_hash_t *hs, struct hlist_node *hnode)
        flock = &lock->l_policy_data.l_flock;
        LASSERT(flock->blocking_export != NULL);
        class_export_get(flock->blocking_export);
-       flock->blocking_refs++;
+       atomic_inc(&flock->blocking_refs);
 }
 
 static void
-ldlm_export_flock_put(cfs_hash_t *hs, struct hlist_node *hnode)
+ldlm_export_flock_put(struct cfs_hash *hs, struct hlist_node *hnode)
 {
        struct ldlm_lock *lock;
        struct ldlm_flock *flock;
 
        lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
-       LDLM_LOCK_RELEASE(lock);
 
        flock = &lock->l_policy_data.l_flock;
        LASSERT(flock->blocking_export != NULL);
        class_export_put(flock->blocking_export);
-       if (--flock->blocking_refs == 0) {
+       if (atomic_dec_and_test(&flock->blocking_refs)) {
                flock->blocking_owner = 0;
                flock->blocking_export = NULL;
        }
+       LDLM_LOCK_RELEASE(lock);
 }
 
-static cfs_hash_ops_t ldlm_export_flock_ops = {
+static struct cfs_hash_ops ldlm_export_flock_ops = {
        .hs_hash        = ldlm_export_flock_hash,
        .hs_key         = ldlm_export_flock_key,
        .hs_keycmp      = ldlm_export_flock_keycmp,