Whamcloud - gitweb
LU-5496 ldlm: reconstruct proper flags on enqueue resend
[fs/lustre-release.git] / lustre / ldlm / ldlm_flock.c
index 26a1bf4..dc98cc0 100644 (file)
 
 #define DEBUG_SUBSYSTEM S_LDLM
 
-#ifdef __KERNEL__
 #include <lustre_dlm.h>
 #include <obd_support.h>
 #include <obd_class.h>
 #include <lustre_lib.h>
 #include <libcfs/list.h>
-#else
-#include <liblustre.h>
-#include <obd_class.h>
-#endif
 
 #include "ldlm_internal.h"
 
@@ -107,7 +102,7 @@ static inline void ldlm_flock_blocking_link(struct ldlm_lock *req,
         if (req->l_export == NULL)
                return;
 
-       LASSERT(cfs_hlist_unhashed(&req->l_exp_flock_hash));
+       LASSERT(hlist_unhashed(&req->l_exp_flock_hash));
 
         req->l_policy_data.l_flock.blocking_owner =
                 lock->l_policy_data.l_flock.owner;
@@ -128,7 +123,7 @@ static inline void ldlm_flock_blocking_unlink(struct ldlm_lock *req)
 
        check_res_locked(req->l_resource);
        if (req->l_export->exp_flock_hash != NULL &&
-           !cfs_hlist_unhashed(&req->l_exp_flock_hash))
+           !hlist_unhashed(&req->l_exp_flock_hash))
                cfs_hash_del(req->l_export->exp_flock_hash,
                             &req->l_policy_data.l_flock.owner,
                             &req->l_exp_flock_hash);
@@ -139,14 +134,14 @@ ldlm_flock_destroy(struct ldlm_lock *lock, ldlm_mode_t mode, __u64 flags)
 {
        ENTRY;
 
-       LDLM_DEBUG(lock, "ldlm_flock_destroy(mode: %d, flags: 0x%llx)",
+       LDLM_DEBUG(lock, "ldlm_flock_destroy(mode: %d, flags: "LPX64")",
                   mode, flags);
 
        /* Safe to not lock here, since it should be empty anyway */
-       LASSERT(cfs_hlist_unhashed(&lock->l_exp_flock_hash));
+       LASSERT(hlist_unhashed(&lock->l_exp_flock_hash));
 
-       cfs_list_del_init(&lock->l_res_link);
-       if (flags == LDLM_FL_WAIT_NOREPROC && !ldlm_is_failed(lock)) {
+       list_del_init(&lock->l_res_link);
+       if (flags == LDLM_FL_WAIT_NOREPROC) {
                /* client side - set a flag to prevent sending a CANCEL */
                lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_CBPENDING;
 
@@ -176,7 +171,7 @@ struct ldlm_flock_lookup_cb_data {
 };
 
 static int ldlm_flock_lookup_cb(cfs_hash_t *hs, cfs_hash_bd_t *bd,
-                               cfs_hlist_node_t *hnode, void *data)
+                               struct hlist_node *hnode, void *data)
 {
        struct ldlm_flock_lookup_cb_data *cb_data = data;
        struct obd_export *exp = cfs_hash_object(hs, hnode);
@@ -253,7 +248,7 @@ ldlm_flock_deadlock(struct ldlm_lock *req, struct ldlm_lock *bl_lock)
 }
 
 static void ldlm_flock_cancel_on_deadlock(struct ldlm_lock *lock,
-                                                cfs_list_t *work_list)
+                                         struct list_head *work_list)
 {
        CDEBUG(D_INFO, "reprocess deadlock req=%p\n", lock);
 
@@ -292,12 +287,12 @@ static void ldlm_flock_cancel_on_deadlock(struct ldlm_lock *lock,
  */
 int
 ldlm_process_flock_lock(struct ldlm_lock *req, __u64 *flags, int first_enq,
-                       ldlm_error_t *err, cfs_list_t *work_list)
+                       ldlm_error_t *err, struct list_head *work_list)
 {
         struct ldlm_resource *res = req->l_resource;
         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
-        cfs_list_t *tmp;
-        cfs_list_t *ownlocks = NULL;
+       struct list_head *tmp;
+       struct list_head *ownlocks = NULL;
         struct ldlm_lock *lock = NULL;
         struct ldlm_lock *new = req;
         struct ldlm_lock *new2 = NULL;
@@ -309,7 +304,7 @@ ldlm_process_flock_lock(struct ldlm_lock *req, __u64 *flags, int first_enq,
         const struct ldlm_callback_suite null_cbs = { NULL };
         ENTRY;
 
-       CDEBUG(D_DLMTRACE, "flags %#llx owner "LPU64" pid %u mode %u start "
+       CDEBUG(D_DLMTRACE, "flags "LPX64" owner "LPU64" pid %u mode %u start "
               LPU64" end "LPU64"\n", *flags,
               new->l_policy_data.l_flock.owner,
                new->l_policy_data.l_flock.pid, mode,
@@ -331,8 +326,8 @@ reprocess:
         if ((*flags == LDLM_FL_WAIT_NOREPROC) || (mode == LCK_NL)) {
                 /* This loop determines where this processes locks start
                  * in the resource lr_granted list. */
-                cfs_list_for_each(tmp, &res->lr_granted) {
-                        lock = cfs_list_entry(tmp, struct ldlm_lock,
+               list_for_each(tmp, &res->lr_granted) {
+                       lock = list_entry(tmp, struct ldlm_lock,
                                               l_res_link);
                         if (ldlm_same_flock_owner(lock, req)) {
                                 ownlocks = tmp;
@@ -345,8 +340,8 @@ reprocess:
 
                 /* This loop determines if there are existing locks
                  * that conflict with the new lock request. */
-                cfs_list_for_each(tmp, &res->lr_granted) {
-                        lock = cfs_list_entry(tmp, struct ldlm_lock,
+               list_for_each(tmp, &res->lr_granted) {
+                       lock = list_entry(tmp, struct ldlm_lock,
                                               l_res_link);
 
                         if (ldlm_same_flock_owner(lock, req)) {
@@ -428,7 +423,7 @@ reprocess:
                 ownlocks = &res->lr_granted;
 
         list_for_remaining_safe(ownlocks, tmp, &res->lr_granted) {
-                lock = cfs_list_entry(ownlocks, struct ldlm_lock, l_res_link);
+               lock = list_entry(ownlocks, struct ldlm_lock, l_res_link);
 
                 if (!ldlm_same_flock_owner(lock, new))
                         break;
@@ -516,20 +511,20 @@ reprocess:
                 /* XXX - if ldlm_lock_new() can sleep we should
                  * release the lr_lock, allocate the new lock,
                  * and restart processing this lock. */
-                if (!new2) {
-                        unlock_res_and_lock(req);
+               if (new2 == NULL) {
+                       unlock_res_and_lock(req);
                        new2 = ldlm_lock_create(ns, &res->lr_name, LDLM_FLOCK,
                                                lock->l_granted_mode, &null_cbs,
                                                NULL, 0, LVB_T_NONE);
-                        lock_res_and_lock(req);
-                        if (!new2) {
-                                ldlm_flock_destroy(req, lock->l_granted_mode,
-                                                   *flags);
-                                *err = -ENOLCK;
-                                RETURN(LDLM_ITER_STOP);
-                        }
-                        goto reprocess;
-                }
+                       lock_res_and_lock(req);
+                       if (IS_ERR(new2)) {
+                               ldlm_flock_destroy(req, lock->l_granted_mode,
+                                                  *flags);
+                               *err = PTR_ERR(new2);
+                               RETURN(LDLM_ITER_STOP);
+                       }
+                       goto reprocess;
+               }
 
                 splitted = 1;
 
@@ -548,7 +543,7 @@ reprocess:
                 if (lock->l_export != NULL) {
                         new2->l_export = class_export_lock_get(lock->l_export, new2);
                         if (new2->l_export->exp_lock_hash &&
-                            cfs_hlist_unhashed(&new2->l_exp_hash))
+                           hlist_unhashed(&new2->l_exp_hash))
                                 cfs_hash_add(new2->l_export->exp_lock_hash,
                                              &new2->l_remote_handle,
                                              &new2->l_exp_hash);
@@ -572,7 +567,7 @@ reprocess:
 
         /* Add req to the granted queue before calling ldlm_reprocess_all(). */
         if (!added) {
-                cfs_list_del_init(&req->l_res_link);
+               list_del_init(&req->l_res_link);
                 /* insert new lock before ownlocks in list. */
                 ldlm_resource_add_lock(res, ownlocks, req);
         }
@@ -589,9 +584,11 @@ reprocess:
                          * note that ldlm_process_flock_lock() will recurse,
                          * but only once because first_enq will be false from
                          * ldlm_reprocess_queue. */
-                        if ((mode == LCK_NL) && overlaps) {
-                                CFS_LIST_HEAD(rpc_list);
+                       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);
@@ -673,29 +670,22 @@ ldlm_flock_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
         int                             rc = 0;
         ENTRY;
 
-       CDEBUG(D_DLMTRACE, "flags: 0x%llx data: %p getlk: %p\n",
-               flags, data, getlk);
-
-        /* Import invalidation. We need to actually release the lock
-         * references being held, so that it can go away. No point in
-         * holding the lock even if app still believes it has it, since
-         * server already dropped it anyway. Only for granted locks too. */
-       if ((lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_LOCAL_ONLY)) ==
-           (LDLM_FL_FAILED|LDLM_FL_LOCAL_ONLY)) {
-               if (lock->l_req_mode == lock->l_granted_mode &&
-                   lock->l_granted_mode != LCK_NL &&
-                   NULL == data)
-                       ldlm_lock_decref_internal(lock, lock->l_req_mode);
-
-               /* Need to wake up the waiter if we were evicted */
-               wake_up(&lock->l_waitq);
-               RETURN(0);
+       OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT2, 4);
+       if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_CP_CB_WAIT3)) {
+               lock_res_and_lock(lock);
+               lock->l_flags |= LDLM_FL_FAIL_LOC;
+               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",
+              flags, data, getlk);
 
        LASSERT(flags != LDLM_FL_WAIT_NOREPROC);
 
-       if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
-                      LDLM_FL_BLOCK_CONV))) {
+       if (flags & LDLM_FL_FAILED)
+               goto granted;
+
+       if (!(flags & LDLM_FL_BLOCKED_MASK)) {
                if (NULL == data)
                        /* mds granted the lock in the reply */
                        goto granted;
@@ -733,12 +723,21 @@ ldlm_flock_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
 granted:
         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT, 10);
 
-        if (ldlm_is_failed(lock)) {
-                LDLM_DEBUG(lock, "client-side enqueue waking up: failed");
-                RETURN(-EIO);
-        }
-
-        LDLM_DEBUG(lock, "client-side enqueue granted");
+       if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_CP_CB_WAIT4)) {
+               lock_res_and_lock(lock);
+               /* DEADLOCK is always set with CBPENDING */
+               lock->l_flags |= LDLM_FL_FLOCK_DEADLOCK | LDLM_FL_CBPENDING;
+               unlock_res_and_lock(lock);
+               OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT4, 4);
+       }
+       if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_CP_CB_WAIT5)) {
+               lock_res_and_lock(lock);
+               /* DEADLOCK is always set with CBPENDING */
+               lock->l_flags |= LDLM_FL_FAIL_LOC |
+                                LDLM_FL_FLOCK_DEADLOCK | LDLM_FL_CBPENDING;
+               unlock_res_and_lock(lock);
+               OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT5, 4);
+       }
 
        lock_res_and_lock(lock);
 
@@ -749,22 +748,54 @@ granted:
        if (ldlm_is_destroyed(lock)) {
                unlock_res_and_lock(lock);
                LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
-               RETURN(0);
-       }
 
-       /* take lock off the deadlock detection hash list. */
-        ldlm_flock_blocking_unlink(lock);
+               /* An error is still to be returned, to propagate it up to
+                * ldlm_cli_enqueue_fini() caller. */
+               RETURN(-EIO);
+       }
 
         /* ldlm_lock_enqueue() has already placed lock on the granted list. */
-        cfs_list_del_init(&lock->l_res_link);
+       ldlm_resource_unlink_lock(lock);
+
+       /* Import invalidation. We need to actually release the lock
+        * references being held, so that it can go away. No point in
+        * holding the lock even if app still believes it has it, since
+        * server already dropped it anyway. Only for granted locks too. */
+       /* Do the same for DEADLOCK'ed locks. */
+       if (ldlm_is_failed(lock) || ldlm_is_flock_deadlock(lock)) {
+               int mode;
+
+               if (flags & LDLM_FL_TEST_LOCK)
+                       LASSERT(ldlm_is_test_lock(lock));
+
+               if (ldlm_is_test_lock(lock) || ldlm_is_flock_deadlock(lock))
+                       mode = flock_type(getlk);
+               else
+                       mode = lock->l_granted_mode;
+
+               if (ldlm_is_flock_deadlock(lock)) {
+                       LDLM_DEBUG(lock, "client-side enqueue deadlock "
+                                  "received");
+                       rc = -EDEADLK;
+               }
+               ldlm_flock_destroy(lock, mode, LDLM_FL_WAIT_NOREPROC);
+               unlock_res_and_lock(lock);
+
+               /* Need to wake up the waiter if we were evicted */
+               wake_up(&lock->l_waitq);
+
+               /* An error is still to be returned, to propagate it up to
+                * ldlm_cli_enqueue_fini() caller. */
+               RETURN(rc ? : -EIO);
+       }
+
+       LDLM_DEBUG(lock, "client-side enqueue granted");
 
-       if (ldlm_is_flock_deadlock(lock)) {
-               LDLM_DEBUG(lock, "client-side enqueue deadlock received");
-               rc = -EDEADLK;
-       } else if (flags & LDLM_FL_TEST_LOCK) {
+       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.*/
+               LASSERT(ldlm_is_test_lock(lock));
                ldlm_flock_destroy(lock, flock_type(getlk),
                                   LDLM_FL_WAIT_NOREPROC);
                switch (lock->l_granted_mode) {
@@ -853,33 +884,33 @@ ldlm_export_flock_hash(cfs_hash_t *hs, const void *key, unsigned mask)
 }
 
 static void *
-ldlm_export_flock_key(cfs_hlist_node_t *hnode)
+ldlm_export_flock_key(struct hlist_node *hnode)
 {
        struct ldlm_lock *lock;
 
-       lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
+       lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
        return &lock->l_policy_data.l_flock.owner;
 }
 
 static int
-ldlm_export_flock_keycmp(const void *key, cfs_hlist_node_t *hnode)
+ldlm_export_flock_keycmp(const void *key, struct hlist_node *hnode)
 {
        return !memcmp(ldlm_export_flock_key(hnode), key, sizeof(__u64));
 }
 
 static void *
-ldlm_export_flock_object(cfs_hlist_node_t *hnode)
+ldlm_export_flock_object(struct hlist_node *hnode)
 {
-       return cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
+       return hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 }
 
 static void
-ldlm_export_flock_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
+ldlm_export_flock_get(cfs_hash_t *hs, struct hlist_node *hnode)
 {
        struct ldlm_lock *lock;
        struct ldlm_flock *flock;
 
-       lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
+       lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
        LDLM_LOCK_GET(lock);
 
        flock = &lock->l_policy_data.l_flock;
@@ -889,12 +920,12 @@ ldlm_export_flock_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 }
 
 static void
-ldlm_export_flock_put(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
+ldlm_export_flock_put(cfs_hash_t *hs, struct hlist_node *hnode)
 {
        struct ldlm_lock *lock;
        struct ldlm_flock *flock;
 
-       lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
+       lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
        LDLM_LOCK_RELEASE(lock);
 
        flock = &lock->l_policy_data.l_flock;