X-Git-Url: https://err.no/cgi-bin/gitweb.cgi?a=blobdiff_plain;f=fs%2Focfs2%2Fdlmglue.c;h=394d25a131a54d406b4fd9422639cde7c1d7c320;hb=6924d1ab8b7bbe5ab416713f5701b3316b2df85b;hp=05fd016ba4bf4a327336acb556659ce1b8fbda1d;hpb=6953b4c008628b945bfe0cee97f6e78a98773859;p=linux-2.6 diff --git a/fs/ocfs2/dlmglue.c b/fs/ocfs2/dlmglue.c index 05fd016ba4..394d25a131 100644 --- a/fs/ocfs2/dlmglue.c +++ b/fs/ocfs2/dlmglue.c @@ -311,12 +311,13 @@ static int ocfs2_inode_lock_update(struct inode *inode, struct buffer_head **bh); static void ocfs2_drop_osb_locks(struct ocfs2_super *osb); static inline int ocfs2_highest_compat_lock_level(int level); -static void ocfs2_prepare_downconvert(struct ocfs2_lock_res *lockres, - int new_level); +static unsigned int ocfs2_prepare_downconvert(struct ocfs2_lock_res *lockres, + int new_level); static int ocfs2_downconvert_lock(struct ocfs2_super *osb, struct ocfs2_lock_res *lockres, int new_level, - int lvb); + int lvb, + unsigned int generation); static int ocfs2_prepare_cancel_convert(struct ocfs2_super *osb, struct ocfs2_lock_res *lockres); static int ocfs2_cancel_convert(struct ocfs2_super *osb, @@ -736,6 +737,113 @@ static int ocfs2_generic_handle_bast(struct ocfs2_lock_res *lockres, return needs_downconvert; } +/* + * OCFS2_LOCK_PENDING and l_pending_gen. + * + * Why does OCFS2_LOCK_PENDING exist? To close a race between setting + * OCFS2_LOCK_BUSY and calling ocfs2_dlm_lock(). See ocfs2_unblock_lock() + * for more details on the race. + * + * OCFS2_LOCK_PENDING closes the race quite nicely. However, it introduces + * a race on itself. In o2dlm, we can get the ast before ocfs2_dlm_lock() + * returns. The ast clears OCFS2_LOCK_BUSY, and must therefore clear + * OCFS2_LOCK_PENDING at the same time. When ocfs2_dlm_lock() returns, + * the caller is going to try to clear PENDING again. If nothing else is + * happening, __lockres_clear_pending() sees PENDING is unset and does + * nothing. + * + * But what if another path (eg downconvert thread) has just started a + * new locking action? The other path has re-set PENDING. Our path + * cannot clear PENDING, because that will re-open the original race + * window. + * + * [Example] + * + * ocfs2_meta_lock() + * ocfs2_cluster_lock() + * set BUSY + * set PENDING + * drop l_lock + * ocfs2_dlm_lock() + * ocfs2_locking_ast() ocfs2_downconvert_thread() + * clear PENDING ocfs2_unblock_lock() + * take_l_lock + * !BUSY + * ocfs2_prepare_downconvert() + * set BUSY + * set PENDING + * drop l_lock + * take l_lock + * clear PENDING + * drop l_lock + * + * ocfs2_dlm_lock() + * + * So as you can see, we now have a window where l_lock is not held, + * PENDING is not set, and ocfs2_dlm_lock() has not been called. + * + * The core problem is that ocfs2_cluster_lock() has cleared the PENDING + * set by ocfs2_prepare_downconvert(). That wasn't nice. + * + * To solve this we introduce l_pending_gen. A call to + * lockres_clear_pending() will only do so when it is passed a generation + * number that matches the lockres. lockres_set_pending() will return the + * current generation number. When ocfs2_cluster_lock() goes to clear + * PENDING, it passes the generation it got from set_pending(). In our + * example above, the generation numbers will *not* match. Thus, + * ocfs2_cluster_lock() will not clear the PENDING set by + * ocfs2_prepare_downconvert(). + */ + +/* Unlocked version for ocfs2_locking_ast() */ +static void __lockres_clear_pending(struct ocfs2_lock_res *lockres, + unsigned int generation, + struct ocfs2_super *osb) +{ + assert_spin_locked(&lockres->l_lock); + + /* + * The ast and locking functions can race us here. The winner + * will clear pending, the loser will not. + */ + if (!(lockres->l_flags & OCFS2_LOCK_PENDING) || + (lockres->l_pending_gen != generation)) + return; + + lockres_clear_flags(lockres, OCFS2_LOCK_PENDING); + lockres->l_pending_gen++; + + /* + * The downconvert thread may have skipped us because we + * were PENDING. Wake it up. + */ + if (lockres->l_flags & OCFS2_LOCK_BLOCKED) + ocfs2_wake_downconvert_thread(osb); +} + +/* Locked version for callers of ocfs2_dlm_lock() */ +static void lockres_clear_pending(struct ocfs2_lock_res *lockres, + unsigned int generation, + struct ocfs2_super *osb) +{ + unsigned long flags; + + spin_lock_irqsave(&lockres->l_lock, flags); + __lockres_clear_pending(lockres, generation, osb); + spin_unlock_irqrestore(&lockres->l_lock, flags); +} + +static unsigned int lockres_set_pending(struct ocfs2_lock_res *lockres) +{ + assert_spin_locked(&lockres->l_lock); + BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BUSY)); + + lockres_or_flags(lockres, OCFS2_LOCK_PENDING); + + return lockres->l_pending_gen; +} + + static void ocfs2_blocking_ast(void *opaque, int level) { struct ocfs2_lock_res *lockres = opaque; @@ -770,14 +878,22 @@ static void ocfs2_blocking_ast(void *opaque, int level) static void ocfs2_locking_ast(void *opaque) { struct ocfs2_lock_res *lockres = opaque; + struct ocfs2_super *osb = ocfs2_get_lockres_osb(lockres); unsigned long flags; + int status; spin_lock_irqsave(&lockres->l_lock, flags); - if (ocfs2_dlm_lock_status(&lockres->l_lksb)) { + status = ocfs2_dlm_lock_status(&lockres->l_lksb); + + if (status == -EAGAIN) { + lockres_clear_flags(lockres, OCFS2_LOCK_BUSY); + goto out; + } + + if (status) { mlog(ML_ERROR, "lockres %s: lksb status value of %d!\n", - lockres->l_name, - ocfs2_dlm_lock_status(&lockres->l_lksb)); + lockres->l_name, status); spin_unlock_irqrestore(&lockres->l_lock, flags); return; } @@ -800,11 +916,23 @@ static void ocfs2_locking_ast(void *opaque) lockres->l_unlock_action); BUG(); } - +out: /* set it to something invalid so if we get called again we * can catch it. */ lockres->l_action = OCFS2_AST_INVALID; + /* Did we try to cancel this lock? Clear that state */ + if (lockres->l_unlock_action == OCFS2_UNLOCK_CANCEL_CONVERT) + lockres->l_unlock_action = OCFS2_UNLOCK_INVALID; + + /* + * We may have beaten the locking functions here. We certainly + * know that dlm_lock() has been called :-) + * Because we can't have two lock calls in flight at once, we + * can use lockres->l_pending_gen. + */ + __lockres_clear_pending(lockres, lockres->l_pending_gen, osb); + wake_up(&lockres->l_event); spin_unlock_irqrestore(&lockres->l_lock, flags); } @@ -838,6 +966,7 @@ static int ocfs2_lock_create(struct ocfs2_super *osb, { int ret = 0; unsigned long flags; + unsigned int gen; mlog_entry_void(); @@ -854,6 +983,7 @@ static int ocfs2_lock_create(struct ocfs2_super *osb, lockres->l_action = OCFS2_AST_ATTACH; lockres->l_requested = level; lockres_or_flags(lockres, OCFS2_LOCK_BUSY); + gen = lockres_set_pending(lockres); spin_unlock_irqrestore(&lockres->l_lock, flags); ret = ocfs2_dlm_lock(osb->cconn, @@ -863,6 +993,7 @@ static int ocfs2_lock_create(struct ocfs2_super *osb, lockres->l_name, OCFS2_LOCK_ID_MAX_LEN - 1, lockres); + lockres_clear_pending(lockres, gen, osb); if (ret) { ocfs2_log_dlm_error("ocfs2_dlm_lock", ret, lockres); ocfs2_recover_from_dlm_error(lockres, 1); @@ -988,6 +1119,8 @@ static int ocfs2_cluster_lock(struct ocfs2_super *osb, int wait, catch_signals = !(osb->s_mount_opt & OCFS2_MOUNT_NOINTR); int ret = 0; /* gcc doesn't realize wait = 1 guarantees ret is set */ unsigned long flags; + unsigned int gen; + int noqueue_attempted = 0; mlog_entry_void(); @@ -1032,6 +1165,13 @@ again: } if (level > lockres->l_level) { + if (noqueue_attempted > 0) { + ret = -EAGAIN; + goto unlock; + } + if (lkm_flags & DLM_LKF_NOQUEUE) + noqueue_attempted = 1; + if (lockres->l_action != OCFS2_AST_INVALID) mlog(ML_ERROR, "lockres %s has action %u pending\n", lockres->l_name, lockres->l_action); @@ -1046,6 +1186,7 @@ again: lockres->l_requested = level; lockres_or_flags(lockres, OCFS2_LOCK_BUSY); + gen = lockres_set_pending(lockres); spin_unlock_irqrestore(&lockres->l_lock, flags); BUG_ON(level == DLM_LOCK_IV); @@ -1062,6 +1203,7 @@ again: lockres->l_name, OCFS2_LOCK_ID_MAX_LEN - 1, lockres); + lockres_clear_pending(lockres, gen, osb); if (ret) { if (!(lkm_flags & DLM_LKF_NOQUEUE) || (ret != -EAGAIN)) { @@ -1494,6 +1636,10 @@ int ocfs2_file_lock(struct file *file, int ex, int trylock) * to just bubble sucess back up to the user. */ ret = ocfs2_flock_handle_signal(lockres, level); + } else if (!ret && (level > lockres->l_level)) { + /* Trylock failed asynchronously */ + BUG_ON(!trylock); + ret = -EAGAIN; } out: @@ -1506,6 +1652,7 @@ out: void ocfs2_file_unlock(struct file *file) { int ret; + unsigned int gen; unsigned long flags; struct ocfs2_file_private *fp = file->private_data; struct ocfs2_lock_res *lockres = &fp->fp_flock; @@ -1531,11 +1678,11 @@ void ocfs2_file_unlock(struct file *file) lockres_or_flags(lockres, OCFS2_LOCK_BLOCKED); lockres->l_blocking = DLM_LOCK_EX; - ocfs2_prepare_downconvert(lockres, LKM_NLMODE); + gen = ocfs2_prepare_downconvert(lockres, LKM_NLMODE); lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0); spin_unlock_irqrestore(&lockres->l_lock, flags); - ret = ocfs2_downconvert_lock(osb, lockres, LKM_NLMODE, 0); + ret = ocfs2_downconvert_lock(osb, lockres, LKM_NLMODE, 0, gen); if (ret) { mlog_errno(ret); return; @@ -2459,8 +2606,10 @@ int ocfs2_dlm_init(struct ocfs2_super *osb) mlog_entry_void(); - if (ocfs2_mount_local(osb)) + if (ocfs2_mount_local(osb)) { + osb->node_num = 0; goto local; + } status = ocfs2_dlm_init_debug(osb); if (status < 0) { @@ -2478,7 +2627,8 @@ int ocfs2_dlm_init(struct ocfs2_super *osb) } /* for now, uuid == domain */ - status = ocfs2_cluster_connect(osb->uuid_str, + status = ocfs2_cluster_connect(osb->osb_cluster_stack, + osb->uuid_str, strlen(osb->uuid_str), ocfs2_do_node_down, osb, &conn); @@ -2487,6 +2637,15 @@ int ocfs2_dlm_init(struct ocfs2_super *osb) goto bail; } + status = ocfs2_cluster_this_node(&osb->node_num); + if (status < 0) { + mlog_errno(status); + mlog(ML_ERROR, + "could not find this host's node number\n"); + ocfs2_cluster_disconnect(conn, 0); + goto bail; + } + local: ocfs2_super_lock_res_init(&osb->osb_super_lockres, osb); ocfs2_rename_lock_res_init(&osb->osb_rename_lockres, osb); @@ -2505,7 +2664,8 @@ bail: return status; } -void ocfs2_dlm_shutdown(struct ocfs2_super *osb) +void ocfs2_dlm_shutdown(struct ocfs2_super *osb, + int hangup_pending) { mlog_entry_void(); @@ -2525,7 +2685,7 @@ void ocfs2_dlm_shutdown(struct ocfs2_super *osb) ocfs2_lock_res_free(&osb->osb_super_lockres); ocfs2_lock_res_free(&osb->osb_rename_lockres); - ocfs2_cluster_disconnect(osb->cconn); + ocfs2_cluster_disconnect(osb->cconn, hangup_pending); osb->cconn = NULL; ocfs2_dlm_shutdown_debug(osb); @@ -2544,23 +2704,7 @@ static void ocfs2_unlock_ast(void *opaque, int error) lockres->l_unlock_action); spin_lock_irqsave(&lockres->l_lock, flags); - /* We tried to cancel a convert request, but it was already - * granted. All we want to do here is clear our unlock - * state. The wake_up call done at the bottom is redundant - * (ocfs2_prepare_cancel_convert doesn't sleep on this) but doesn't - * hurt anything anyway */ - if (error == -DLM_ECANCEL && - lockres->l_unlock_action == OCFS2_UNLOCK_CANCEL_CONVERT) { - mlog(0, "Got cancelgrant for %s\n", lockres->l_name); - - /* We don't clear the busy flag in this case as it - * should have been cleared by the ast which the dlm - * has called. */ - goto complete_unlock; - } - - /* DLM_EUNLOCK is the success code for unlock */ - if (error != -DLM_EUNLOCK) { + if (error) { mlog(ML_ERROR, "Dlm passes error %d for lock %s, " "unlock_action %d\n", error, lockres->l_name, lockres->l_unlock_action); @@ -2581,7 +2725,6 @@ static void ocfs2_unlock_ast(void *opaque, int error) } lockres_clear_flags(lockres, OCFS2_LOCK_BUSY); -complete_unlock: lockres->l_unlock_action = OCFS2_UNLOCK_INVALID; spin_unlock_irqrestore(&lockres->l_lock, flags); @@ -2662,8 +2805,7 @@ static int ocfs2_drop_lock(struct ocfs2_super *osb, if (ret) { ocfs2_log_dlm_error("ocfs2_dlm_unlock", ret, lockres); mlog(ML_ERROR, "lockres flags: %lu\n", lockres->l_flags); - /* XXX Need to abstract this */ - dlm_print_one_lock(lockres->l_lksb.lksb_o2dlm.lockid); + ocfs2_dlm_dump_lksb(&lockres->l_lksb); BUG(); } mlog(0, "lock %s, successfull return from ocfs2_dlm_unlock\n", @@ -2757,8 +2899,8 @@ int ocfs2_drop_inode_locks(struct inode *inode) return status; } -static void ocfs2_prepare_downconvert(struct ocfs2_lock_res *lockres, - int new_level) +static unsigned int ocfs2_prepare_downconvert(struct ocfs2_lock_res *lockres, + int new_level) { assert_spin_locked(&lockres->l_lock); @@ -2776,12 +2918,14 @@ static void ocfs2_prepare_downconvert(struct ocfs2_lock_res *lockres, lockres->l_action = OCFS2_AST_DOWNCONVERT; lockres->l_requested = new_level; lockres_or_flags(lockres, OCFS2_LOCK_BUSY); + return lockres_set_pending(lockres); } static int ocfs2_downconvert_lock(struct ocfs2_super *osb, struct ocfs2_lock_res *lockres, int new_level, - int lvb) + int lvb, + unsigned int generation) { int ret; u32 dlm_flags = DLM_LKF_CONVERT; @@ -2798,6 +2942,7 @@ static int ocfs2_downconvert_lock(struct ocfs2_super *osb, lockres->l_name, OCFS2_LOCK_ID_MAX_LEN - 1, lockres); + lockres_clear_pending(lockres, generation, osb); if (ret) { ocfs2_log_dlm_error("ocfs2_dlm_lock", ret, lockres); ocfs2_recover_from_dlm_error(lockres, 1); @@ -2872,6 +3017,7 @@ static int ocfs2_unblock_lock(struct ocfs2_super *osb, int new_level; int ret = 0; int set_lvb = 0; + unsigned int gen; mlog_entry_void(); @@ -2881,6 +3027,32 @@ static int ocfs2_unblock_lock(struct ocfs2_super *osb, recheck: if (lockres->l_flags & OCFS2_LOCK_BUSY) { + /* XXX + * This is a *big* race. The OCFS2_LOCK_PENDING flag + * exists entirely for one reason - another thread has set + * OCFS2_LOCK_BUSY, but has *NOT* yet called dlm_lock(). + * + * If we do ocfs2_cancel_convert() before the other thread + * calls dlm_lock(), our cancel will do nothing. We will + * get no ast, and we will have no way of knowing the + * cancel failed. Meanwhile, the other thread will call + * into dlm_lock() and wait...forever. + * + * Why forever? Because another node has asked for the + * lock first; that's why we're here in unblock_lock(). + * + * The solution is OCFS2_LOCK_PENDING. When PENDING is + * set, we just requeue the unblock. Only when the other + * thread has called dlm_lock() and cleared PENDING will + * we then cancel their request. + * + * All callers of dlm_lock() must set OCFS2_DLM_PENDING + * at the same time they set OCFS2_DLM_BUSY. They must + * clear OCFS2_DLM_PENDING after dlm_lock() returns. + */ + if (lockres->l_flags & OCFS2_LOCK_PENDING) + goto leave_requeue; + ctl->requeue = 1; ret = ocfs2_prepare_cancel_convert(osb, lockres); spin_unlock_irqrestore(&lockres->l_lock, flags); @@ -2960,9 +3132,11 @@ downconvert: lockres->l_ops->set_lvb(lockres); } - ocfs2_prepare_downconvert(lockres, new_level); + gen = ocfs2_prepare_downconvert(lockres, new_level); spin_unlock_irqrestore(&lockres->l_lock, flags); - ret = ocfs2_downconvert_lock(osb, lockres, new_level, set_lvb); + ret = ocfs2_downconvert_lock(osb, lockres, new_level, set_lvb, + gen); + leave: mlog_exit(ret); return ret; @@ -3194,16 +3368,11 @@ static struct ocfs2_locking_protocol lproto = { .lp_unlock_ast = ocfs2_unlock_ast, }; -/* This interface isn't the final one, hence the less-than-perfect names */ -void dlmglue_init_stack(void) +void ocfs2_set_locking_protocol(void) { - o2cb_get_stack(&lproto); + ocfs2_stack_glue_set_locking_protocol(&lproto); } -void dlmglue_exit_stack(void) -{ - o2cb_put_stack(); -} static void ocfs2_process_blocked_lock(struct ocfs2_super *osb, struct ocfs2_lock_res *lockres)