[GFS2] Fix ordering bug in lock_dlm
[safe/jmp/linux-2.6] / fs / gfs2 / locking / dlm / thread.c
index e53db6f..f30350a 100644 (file)
 
 #include "lock_dlm.h"
 
-/* A lock placed on this queue is re-submitted to DLM as soon as the lock_dlm
-   thread gets to it. */
-
-static void queue_submit(struct gdlm_lock *lp)
-{
-       struct gdlm_ls *ls = lp->ls;
-
-       spin_lock(&ls->async_lock);
-       list_add_tail(&lp->delay_list, &ls->submit);
-       spin_unlock(&ls->async_lock);
-       wake_up(&ls->thread_wait);
-}
-
-static void process_blocking(struct gdlm_lock *lp, int bast_mode)
-{
-       struct gdlm_ls *ls = lp->ls;
-       unsigned int cb = 0;
-
-       switch (gdlm_make_lmstate(bast_mode)) {
-       case LM_ST_EXCLUSIVE:
-               cb = LM_CB_NEED_E;
-               break;
-       case LM_ST_DEFERRED:
-               cb = LM_CB_NEED_D;
-               break;
-       case LM_ST_SHARED:
-               cb = LM_CB_NEED_S;
-               break;
-       default:
-               gdlm_assert(0, "unknown bast mode %u", lp->bast_mode);
-       }
-
-       ls->fscb(ls->sdp, cb, &lp->lockname);
-}
-
-static void wake_up_ast(struct gdlm_lock *lp)
-{
-       clear_bit(LFL_AST_WAIT, &lp->flags);
-       smp_mb__after_clear_bit();
-       wake_up_bit(&lp->flags, LFL_AST_WAIT);
-}
-
-static void process_complete(struct gdlm_lock *lp)
-{
-       struct gdlm_ls *ls = lp->ls;
-       struct lm_async_cb acb;
-       s16 prev_mode = lp->cur;
-
-       memset(&acb, 0, sizeof(acb));
-
-       if (lp->lksb.sb_status == -DLM_ECANCEL) {
-               log_info("complete dlm cancel %x,%llx flags %lx",
-                        lp->lockname.ln_type,
-                        (unsigned long long)lp->lockname.ln_number,
-                        lp->flags);
-
-               lp->req = lp->cur;
-               acb.lc_ret |= LM_OUT_CANCELED;
-               if (lp->cur == DLM_LOCK_IV)
-                       lp->lksb.sb_lkid = 0;
-               goto out;
-       }
-
-       if (test_and_clear_bit(LFL_DLM_UNLOCK, &lp->flags)) {
-               if (lp->lksb.sb_status != -DLM_EUNLOCK) {
-                       log_info("unlock sb_status %d %x,%llx flags %lx",
-                                lp->lksb.sb_status, lp->lockname.ln_type,
-                                (unsigned long long)lp->lockname.ln_number,
-                                lp->flags);
-                       return;
-               }
-
-               lp->cur = DLM_LOCK_IV;
-               lp->req = DLM_LOCK_IV;
-               lp->lksb.sb_lkid = 0;
-
-               if (test_and_clear_bit(LFL_UNLOCK_DELETE, &lp->flags)) {
-                       gdlm_delete_lp(lp);
-                       return;
-               }
-               goto out;
-       }
-
-       if (lp->lksb.sb_flags & DLM_SBF_VALNOTVALID)
-               memset(lp->lksb.sb_lvbptr, 0, GDLM_LVB_SIZE);
-
-       if (lp->lksb.sb_flags & DLM_SBF_ALTMODE) {
-               if (lp->req == DLM_LOCK_PR)
-                       lp->req = DLM_LOCK_CW;
-               else if (lp->req == DLM_LOCK_CW)
-                       lp->req = DLM_LOCK_PR;
-       }
-
-       /*
-        * A canceled lock request.  The lock was just taken off the delayed
-        * list and was never even submitted to dlm.
-        */
-
-       if (test_and_clear_bit(LFL_CANCEL, &lp->flags)) {
-               log_info("complete internal cancel %x,%llx",
-                        lp->lockname.ln_type,
-                        (unsigned long long)lp->lockname.ln_number);
-               lp->req = lp->cur;
-               acb.lc_ret |= LM_OUT_CANCELED;
-               goto out;
-       }
-
-       /*
-        * An error occured.
-        */
-
-       if (lp->lksb.sb_status) {
-               /* a "normal" error */
-               if ((lp->lksb.sb_status == -EAGAIN) &&
-                   (lp->lkf & DLM_LKF_NOQUEUE)) {
-                       lp->req = lp->cur;
-                       if (lp->cur == DLM_LOCK_IV)
-                               lp->lksb.sb_lkid = 0;
-                       goto out;
-               }
-
-               /* this could only happen with cancels I think */
-               log_info("ast sb_status %d %x,%llx flags %lx",
-                        lp->lksb.sb_status, lp->lockname.ln_type,
-                        (unsigned long long)lp->lockname.ln_number,
-                        lp->flags);
-               if (lp->lksb.sb_status == -EDEADLOCK &&
-                   lp->ls->fsflags & LM_MFLAG_CONV_NODROP) {
-                       lp->req = lp->cur;
-                       acb.lc_ret |= LM_OUT_CONV_DEADLK;
-                       if (lp->cur == DLM_LOCK_IV)
-                               lp->lksb.sb_lkid = 0;
-                       goto out;
-               } else
-                       return;
-       }
-
-       /*
-        * This is an AST for an EX->EX conversion for sync_lvb from GFS.
-        */
-
-       if (test_and_clear_bit(LFL_SYNC_LVB, &lp->flags)) {
-               wake_up_ast(lp);
-               return;
-       }
-
-       /*
-        * A lock has been demoted to NL because it initially completed during
-        * BLOCK_LOCKS.  Now it must be requested in the originally requested
-        * mode.
-        */
-
-       if (test_and_clear_bit(LFL_REREQUEST, &lp->flags)) {
-               gdlm_assert(lp->req == DLM_LOCK_NL, "%x,%llx",
-                           lp->lockname.ln_type,
-                           (unsigned long long)lp->lockname.ln_number);
-               gdlm_assert(lp->prev_req > DLM_LOCK_NL, "%x,%llx",
-                           lp->lockname.ln_type,
-                           (unsigned long long)lp->lockname.ln_number);
-
-               lp->cur = DLM_LOCK_NL;
-               lp->req = lp->prev_req;
-               lp->prev_req = DLM_LOCK_IV;
-               lp->lkf &= ~DLM_LKF_CONVDEADLK;
-
-               set_bit(LFL_NOCACHE, &lp->flags);
-
-               if (test_bit(DFL_BLOCK_LOCKS, &ls->flags) &&
-                   !test_bit(LFL_NOBLOCK, &lp->flags))
-                       gdlm_queue_delayed(lp);
-               else
-                       queue_submit(lp);
-               return;
-       }
-
-       /*
-        * A request is granted during dlm recovery.  It may be granted
-        * because the locks of a failed node were cleared.  In that case,
-        * there may be inconsistent data beneath this lock and we must wait
-        * for recovery to complete to use it.  When gfs recovery is done this
-        * granted lock will be converted to NL and then reacquired in this
-        * granted state.
-        */
-
-       if (test_bit(DFL_BLOCK_LOCKS, &ls->flags) &&
-           !test_bit(LFL_NOBLOCK, &lp->flags) &&
-           lp->req != DLM_LOCK_NL) {
-
-               lp->cur = lp->req;
-               lp->prev_req = lp->req;
-               lp->req = DLM_LOCK_NL;
-               lp->lkf |= DLM_LKF_CONVERT;
-               lp->lkf &= ~DLM_LKF_CONVDEADLK;
-
-               log_debug("rereq %x,%llx id %x %d,%d",
-                         lp->lockname.ln_type,
-                         (unsigned long long)lp->lockname.ln_number,
-                         lp->lksb.sb_lkid, lp->cur, lp->req);
-
-               set_bit(LFL_REREQUEST, &lp->flags);
-               queue_submit(lp);
-               return;
-       }
-
-       /*
-        * DLM demoted the lock to NL before it was granted so GFS must be
-        * told it cannot cache data for this lock.
-        */
-
-       if (lp->lksb.sb_flags & DLM_SBF_DEMOTED)
-               set_bit(LFL_NOCACHE, &lp->flags);
-
-out:
-       /*
-        * This is an internal lock_dlm lock
-        */
-
-       if (test_bit(LFL_INLOCK, &lp->flags)) {
-               clear_bit(LFL_NOBLOCK, &lp->flags);
-               lp->cur = lp->req;
-               wake_up_ast(lp);
-               return;
-       }
-
-       /*
-        * Normal completion of a lock request.  Tell GFS it now has the lock.
-        */
-
-       clear_bit(LFL_NOBLOCK, &lp->flags);
-       lp->cur = lp->req;
-
-       acb.lc_name = lp->lockname;
-       acb.lc_ret |= gdlm_make_lmstate(lp->cur);
-
-       if (!test_and_clear_bit(LFL_NOCACHE, &lp->flags) &&
-           (lp->cur > DLM_LOCK_NL) && (prev_mode > DLM_LOCK_NL))
-               acb.lc_ret |= LM_OUT_CACHEABLE;
-
-       ls->fscb(ls->sdp, LM_CB_ASYNC, &acb);
-}
-
-static inline int no_work(struct gdlm_ls *ls, int blocking)
+static inline int no_work(struct gdlm_ls *ls)
 {
        int ret;
 
        spin_lock(&ls->async_lock);
-       ret = list_empty(&ls->complete) && list_empty(&ls->submit);
-       if (ret && blocking)
-               ret = list_empty(&ls->blocking);
+       ret = list_empty(&ls->submit);
        spin_unlock(&ls->async_lock);
 
        return ret;
@@ -276,100 +33,55 @@ static inline int check_drop(struct gdlm_ls *ls)
        return 0;
 }
 
-static int gdlm_thread(void *data, int blist)
+static int gdlm_thread(void *data)
 {
        struct gdlm_ls *ls = (struct gdlm_ls *) data;
        struct gdlm_lock *lp = NULL;
-       uint8_t complete, blocking, submit, drop;
-
-       /* Only thread1 is allowed to do blocking callbacks since gfs
-          may wait for a completion callback within a blocking cb. */
 
        while (!kthread_should_stop()) {
                wait_event_interruptible(ls->thread_wait,
-                               !no_work(ls, blist) || kthread_should_stop());
-
-               complete = blocking = submit = drop = 0;
+                               !no_work(ls) || kthread_should_stop());
 
                spin_lock(&ls->async_lock);
 
-               if (blist && !list_empty(&ls->blocking)) {
-                       lp = list_entry(ls->blocking.next, struct gdlm_lock,
-                                       blist);
-                       list_del_init(&lp->blist);
-                       blocking = lp->bast_mode;
-                       lp->bast_mode = 0;
-               } else if (!list_empty(&ls->complete)) {
-                       lp = list_entry(ls->complete.next, struct gdlm_lock,
-                                       clist);
-                       list_del_init(&lp->clist);
-                       complete = 1;
-               } else if (!list_empty(&ls->submit)) {
+               if (!list_empty(&ls->submit)) {
                        lp = list_entry(ls->submit.next, struct gdlm_lock,
                                        delay_list);
                        list_del_init(&lp->delay_list);
-                       submit = 1;
-               }
-
-               drop = check_drop(ls);
-               spin_unlock(&ls->async_lock);
-
-               if (complete)
-                       process_complete(lp);
-
-               else if (blocking)
-                       process_blocking(lp, blocking);
-
-               else if (submit)
+                       spin_unlock(&ls->async_lock);
                        gdlm_do_lock(lp);
-
-               if (drop)
+                       spin_lock(&ls->async_lock);
+               }
+               /* Does this ever happen these days? I hope not anyway */
+               if (check_drop(ls)) {
+                       spin_unlock(&ls->async_lock);
                        ls->fscb(ls->sdp, LM_CB_DROPLOCKS, NULL);
-
-               schedule();
+                       spin_lock(&ls->async_lock);
+               }
+               spin_unlock(&ls->async_lock);
        }
 
        return 0;
 }
 
-static int gdlm_thread1(void *data)
-{
-       return gdlm_thread(data, 1);
-}
-
-static int gdlm_thread2(void *data)
-{
-       return gdlm_thread(data, 0);
-}
-
 int gdlm_init_threads(struct gdlm_ls *ls)
 {
        struct task_struct *p;
        int error;
 
-       p = kthread_run(gdlm_thread1, ls, "lock_dlm1");
-       error = IS_ERR(p);
-       if (error) {
-               log_error("can't start lock_dlm1 thread %d", error);
-               return error;
-       }
-       ls->thread1 = p;
-
-       p = kthread_run(gdlm_thread2, ls, "lock_dlm2");
+       p = kthread_run(gdlm_thread, ls, "lock_dlm");
        error = IS_ERR(p);
        if (error) {
-               log_error("can't start lock_dlm2 thread %d", error);
-               kthread_stop(ls->thread1);
+               log_error("can't start lock_dlm thread %d", error);
                return error;
        }
-       ls->thread2 = p;
+       ls->thread = p;
 
        return 0;
 }
 
 void gdlm_release_threads(struct gdlm_ls *ls)
 {
-       kthread_stop(ls->thread1);
-       kthread_stop(ls->thread2);
+       kthread_stop(ls->thread);
 }