diff options
Diffstat (limited to 'fs/ocfs2/dlm/dlmmaster.c')
| -rw-r--r-- | fs/ocfs2/dlm/dlmmaster.c | 1636 |
1 files changed, 1047 insertions, 589 deletions
diff --git a/fs/ocfs2/dlm/dlmmaster.c b/fs/ocfs2/dlm/dlmmaster.c index f784177b624..82abf0cc9a1 100644 --- a/fs/ocfs2/dlm/dlmmaster.c +++ b/fs/ocfs2/dlm/dlmmaster.c @@ -30,7 +30,6 @@ #include <linux/types.h> #include <linux/slab.h> #include <linux/highmem.h> -#include <linux/utsname.h> #include <linux/init.h> #include <linux/sysctl.h> #include <linux/random.h> @@ -48,47 +47,11 @@ #include "dlmapi.h" #include "dlmcommon.h" #include "dlmdomain.h" +#include "dlmdebug.h" #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_MASTER) #include "cluster/masklog.h" -enum dlm_mle_type { - DLM_MLE_BLOCK, - DLM_MLE_MASTER, - DLM_MLE_MIGRATION -}; - -struct dlm_lock_name -{ - u8 len; - u8 name[DLM_LOCKID_NAME_MAX]; -}; - -struct dlm_master_list_entry -{ - struct list_head list; - struct list_head hb_events; - struct dlm_ctxt *dlm; - spinlock_t spinlock; - wait_queue_head_t wq; - atomic_t woken; - struct kref mle_refs; - int inuse; - unsigned long maybe_map[BITS_TO_LONGS(O2NM_MAX_NODES)]; - unsigned long vote_map[BITS_TO_LONGS(O2NM_MAX_NODES)]; - unsigned long response_map[BITS_TO_LONGS(O2NM_MAX_NODES)]; - unsigned long node_map[BITS_TO_LONGS(O2NM_MAX_NODES)]; - u8 master; - u8 new_master; - enum dlm_mle_type type; - struct o2hb_callback_func mle_hb_up; - struct o2hb_callback_func mle_hb_down; - union { - struct dlm_lock_resource *res; - struct dlm_lock_name name; - } u; -}; - static void dlm_mle_node_down(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle, struct o2nm_node *node, @@ -99,130 +62,29 @@ static void dlm_mle_node_up(struct dlm_ctxt *dlm, int idx); static void dlm_assert_master_worker(struct dlm_work_item *item, void *data); -static int dlm_do_assert_master(struct dlm_ctxt *dlm, const char *lockname, - unsigned int namelen, void *nodemap, - u32 flags); +static int dlm_do_assert_master(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + void *nodemap, u32 flags); +static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data); static inline int dlm_mle_equal(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle, const char *name, unsigned int namelen) { - struct dlm_lock_resource *res; - if (dlm != mle->dlm) return 0; - if (mle->type == DLM_MLE_BLOCK || - mle->type == DLM_MLE_MIGRATION) { - if (namelen != mle->u.name.len || - memcmp(name, mle->u.name.name, namelen)!=0) - return 0; - } else { - res = mle->u.res; - if (namelen != res->lockname.len || - memcmp(res->lockname.name, name, namelen) != 0) - return 0; - } - return 1; -} - -#define dlm_print_nodemap(m) _dlm_print_nodemap(m,#m) -static void _dlm_print_nodemap(unsigned long *map, const char *mapname) -{ - int i; - printk("%s=[ ", mapname); - for (i=0; i<O2NM_MAX_NODES; i++) - if (test_bit(i, map)) - printk("%d ", i); - printk("]"); -} - -static void dlm_print_one_mle(struct dlm_master_list_entry *mle) -{ - int refs; - char *type; - char attached; - u8 master; - unsigned int namelen; - const char *name; - struct kref *k; - unsigned long *maybe = mle->maybe_map, - *vote = mle->vote_map, - *resp = mle->response_map, - *node = mle->node_map; - - k = &mle->mle_refs; - if (mle->type == DLM_MLE_BLOCK) - type = "BLK"; - else if (mle->type == DLM_MLE_MASTER) - type = "MAS"; - else - type = "MIG"; - refs = atomic_read(&k->refcount); - master = mle->master; - attached = (list_empty(&mle->hb_events) ? 'N' : 'Y'); - - if (mle->type != DLM_MLE_MASTER) { - namelen = mle->u.name.len; - name = mle->u.name.name; - } else { - namelen = mle->u.res->lockname.len; - name = mle->u.res->lockname.name; - } - - mlog(ML_NOTICE, "%.*s: %3s refs=%3d mas=%3u new=%3u evt=%c inuse=%d ", - namelen, name, type, refs, master, mle->new_master, attached, - mle->inuse); - dlm_print_nodemap(maybe); - printk(", "); - dlm_print_nodemap(vote); - printk(", "); - dlm_print_nodemap(resp); - printk(", "); - dlm_print_nodemap(node); - printk(", "); - printk("\n"); -} - -#if 0 -/* Code here is included but defined out as it aids debugging */ - -static void dlm_dump_mles(struct dlm_ctxt *dlm) -{ - struct dlm_master_list_entry *mle; - struct list_head *iter; - - mlog(ML_NOTICE, "dumping all mles for domain %s:\n", dlm->name); - spin_lock(&dlm->master_lock); - list_for_each(iter, &dlm->master_list) { - mle = list_entry(iter, struct dlm_master_list_entry, list); - dlm_print_one_mle(mle); - } - spin_unlock(&dlm->master_lock); -} - -int dlm_dump_all_mles(const char __user *data, unsigned int len) -{ - struct list_head *iter; - struct dlm_ctxt *dlm; + if (namelen != mle->mnamelen || + memcmp(name, mle->mname, namelen) != 0) + return 0; - spin_lock(&dlm_domain_lock); - list_for_each(iter, &dlm_domains) { - dlm = list_entry (iter, struct dlm_ctxt, list); - mlog(ML_NOTICE, "found dlm: %p, name=%s\n", dlm, dlm->name); - dlm_dump_mles(dlm); - } - spin_unlock(&dlm_domain_lock); - return len; + return 1; } -EXPORT_SYMBOL_GPL(dlm_dump_all_mles); - -#endif /* 0 */ - - -static kmem_cache_t *dlm_mle_cache = NULL; +static struct kmem_cache *dlm_lockres_cache; +static struct kmem_cache *dlm_lockname_cache; +static struct kmem_cache *dlm_mle_cache; static void dlm_mle_release(struct kref *kref); static void dlm_init_mle(struct dlm_master_list_entry *mle, @@ -237,7 +99,8 @@ static int dlm_find_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry **mle, char *name, unsigned int namelen); -static int dlm_do_master_request(struct dlm_master_list_entry *mle, int to); +static int dlm_do_master_request(struct dlm_lock_resource *res, + struct dlm_master_list_entry *mle, int to); static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm, @@ -410,7 +273,7 @@ static void dlm_init_mle(struct dlm_master_list_entry *mle, mle->dlm = dlm; mle->type = type; - INIT_LIST_HEAD(&mle->list); + INIT_HLIST_NODE(&mle->master_hash_node); INIT_LIST_HEAD(&mle->hb_events); memset(mle->maybe_map, 0, sizeof(mle->maybe_map)); spin_lock_init(&mle->spinlock); @@ -422,19 +285,27 @@ static void dlm_init_mle(struct dlm_master_list_entry *mle, mle->new_master = O2NM_MAX_NODES; mle->inuse = 0; + BUG_ON(mle->type != DLM_MLE_BLOCK && + mle->type != DLM_MLE_MASTER && + mle->type != DLM_MLE_MIGRATION); + if (mle->type == DLM_MLE_MASTER) { BUG_ON(!res); - mle->u.res = res; - } else if (mle->type == DLM_MLE_BLOCK) { - BUG_ON(!name); - memcpy(mle->u.name.name, name, namelen); - mle->u.name.len = namelen; - } else /* DLM_MLE_MIGRATION */ { + mle->mleres = res; + memcpy(mle->mname, res->lockname.name, res->lockname.len); + mle->mnamelen = res->lockname.len; + mle->mnamehash = res->lockname.hash; + } else { BUG_ON(!name); - memcpy(mle->u.name.name, name, namelen); - mle->u.name.len = namelen; + mle->mleres = NULL; + memcpy(mle->mname, name, namelen); + mle->mnamelen = namelen; + mle->mnamehash = dlm_lockid_hash(name, namelen); } + atomic_inc(&dlm->mle_tot_count[mle->type]); + atomic_inc(&dlm->mle_cur_count[mle->type]); + /* copy off the node_map and register hb callbacks on our copy */ memcpy(mle->node_map, dlm->domain_map, sizeof(mle->node_map)); memcpy(mle->vote_map, dlm->domain_map, sizeof(mle->vote_map)); @@ -445,6 +316,24 @@ static void dlm_init_mle(struct dlm_master_list_entry *mle, __dlm_mle_attach_hb_events(dlm, mle); } +void __dlm_unlink_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle) +{ + assert_spin_locked(&dlm->spinlock); + assert_spin_locked(&dlm->master_lock); + + if (!hlist_unhashed(&mle->master_hash_node)) + hlist_del_init(&mle->master_hash_node); +} + +void __dlm_insert_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle) +{ + struct hlist_head *bucket; + + assert_spin_locked(&dlm->master_lock); + + bucket = dlm_master_hash(dlm, mle->mnamehash); + hlist_add_head(&mle->master_hash_node, bucket); +} /* returns 1 if found, 0 if not */ static int dlm_find_mle(struct dlm_ctxt *dlm, @@ -452,12 +341,14 @@ static int dlm_find_mle(struct dlm_ctxt *dlm, char *name, unsigned int namelen) { struct dlm_master_list_entry *tmpmle; - struct list_head *iter; + struct hlist_head *bucket; + unsigned int hash; assert_spin_locked(&dlm->master_lock); - list_for_each(iter, &dlm->master_list) { - tmpmle = list_entry(iter, struct dlm_master_list_entry, list); + hash = dlm_lockid_hash(name, namelen); + bucket = dlm_master_hash(dlm, hash); + hlist_for_each_entry(tmpmle, bucket, master_hash_node) { if (!dlm_mle_equal(dlm, tmpmle, name, namelen)) continue; dlm_get_mle(tmpmle); @@ -470,13 +361,10 @@ static int dlm_find_mle(struct dlm_ctxt *dlm, void dlm_hb_event_notify_attached(struct dlm_ctxt *dlm, int idx, int node_up) { struct dlm_master_list_entry *mle; - struct list_head *iter; assert_spin_locked(&dlm->spinlock); - - list_for_each(iter, &dlm->mle_hb_events) { - mle = list_entry(iter, struct dlm_master_list_entry, - hb_events); + + list_for_each_entry(mle, &dlm->mle_hb_events, hb_events) { if (node_up) dlm_mle_node_up(dlm, mle, NULL, idx); else @@ -515,10 +403,10 @@ static void dlm_mle_node_up(struct dlm_ctxt *dlm, int dlm_init_mle_cache(void) { - dlm_mle_cache = kmem_cache_create("dlm_mle_cache", + dlm_mle_cache = kmem_cache_create("o2dlm_mle", sizeof(struct dlm_master_list_entry), 0, SLAB_HWCACHE_ALIGN, - NULL, NULL); + NULL); if (dlm_mle_cache == NULL) return -ENOMEM; return 0; @@ -535,29 +423,23 @@ static void dlm_mle_release(struct kref *kref) struct dlm_master_list_entry *mle; struct dlm_ctxt *dlm; - mlog_entry_void(); - mle = container_of(kref, struct dlm_master_list_entry, mle_refs); dlm = mle->dlm; - if (mle->type != DLM_MLE_MASTER) { - mlog(0, "calling mle_release for %.*s, type %d\n", - mle->u.name.len, mle->u.name.name, mle->type); - } else { - mlog(0, "calling mle_release for %.*s, type %d\n", - mle->u.res->lockname.len, - mle->u.res->lockname.name, mle->type); - } assert_spin_locked(&dlm->spinlock); assert_spin_locked(&dlm->master_lock); + mlog(0, "Releasing mle for %.*s, type %d\n", mle->mnamelen, mle->mname, + mle->type); + /* remove from list if not already */ - if (!list_empty(&mle->list)) - list_del_init(&mle->list); + __dlm_unlink_mle(dlm, mle); /* detach the mle from the domain node up/down events */ __dlm_mle_detach_hb_events(dlm, mle); + atomic_dec(&dlm->mle_cur_count[mle->type]); + /* NOTE: kfree under spinlock here. * if this is bad, we can move this to a freelist. */ kmem_cache_free(dlm_mle_cache, mle); @@ -568,48 +450,46 @@ static void dlm_mle_release(struct kref *kref) * LOCK RESOURCE FUNCTIONS */ -static void dlm_set_lockres_owner(struct dlm_ctxt *dlm, - struct dlm_lock_resource *res, - u8 owner) +int dlm_init_master_caches(void) { - assert_spin_locked(&res->spinlock); - - mlog_entry("%.*s, %u\n", res->lockname.len, res->lockname.name, owner); + dlm_lockres_cache = kmem_cache_create("o2dlm_lockres", + sizeof(struct dlm_lock_resource), + 0, SLAB_HWCACHE_ALIGN, NULL); + if (!dlm_lockres_cache) + goto bail; + + dlm_lockname_cache = kmem_cache_create("o2dlm_lockname", + DLM_LOCKID_NAME_MAX, 0, + SLAB_HWCACHE_ALIGN, NULL); + if (!dlm_lockname_cache) + goto bail; - if (owner == dlm->node_num) - atomic_inc(&dlm->local_resources); - else if (owner == DLM_LOCK_RES_OWNER_UNKNOWN) - atomic_inc(&dlm->unknown_resources); - else - atomic_inc(&dlm->remote_resources); - - res->owner = owner; + return 0; +bail: + dlm_destroy_master_caches(); + return -ENOMEM; } -void dlm_change_lockres_owner(struct dlm_ctxt *dlm, - struct dlm_lock_resource *res, u8 owner) +void dlm_destroy_master_caches(void) { - assert_spin_locked(&res->spinlock); - - if (owner == res->owner) - return; - - if (res->owner == dlm->node_num) - atomic_dec(&dlm->local_resources); - else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) - atomic_dec(&dlm->unknown_resources); - else - atomic_dec(&dlm->remote_resources); + if (dlm_lockname_cache) { + kmem_cache_destroy(dlm_lockname_cache); + dlm_lockname_cache = NULL; + } - dlm_set_lockres_owner(dlm, res, owner); + if (dlm_lockres_cache) { + kmem_cache_destroy(dlm_lockres_cache); + dlm_lockres_cache = NULL; + } } - static void dlm_lockres_release(struct kref *kref) { struct dlm_lock_resource *res; + struct dlm_ctxt *dlm; res = container_of(kref, struct dlm_lock_resource, refs); + dlm = res->dlm; /* This should not happen -- all lockres' have a name * associated with them at init time. */ @@ -618,6 +498,18 @@ static void dlm_lockres_release(struct kref *kref) mlog(0, "destroying lockres %.*s\n", res->lockname.len, res->lockname.name); + spin_lock(&dlm->track_lock); + if (!list_empty(&res->tracking)) + list_del_init(&res->tracking); + else { + mlog(ML_ERROR, "Resource %.*s not on the Tracking list\n", + res->lockname.len, res->lockname.name); + dlm_print_one_lock_resource(res); + } + spin_unlock(&dlm->track_lock); + + atomic_dec(&dlm->res_cur_count); + if (!hlist_unhashed(&res->hash_node) || !list_empty(&res->granted) || !list_empty(&res->converting) || @@ -650,9 +542,9 @@ static void dlm_lockres_release(struct kref *kref) BUG_ON(!list_empty(&res->recovering)); BUG_ON(!list_empty(&res->purge)); - kfree(res->lockname.name); + kmem_cache_free(dlm_lockname_cache, (void *)res->lockname.name); - kfree(res); + kmem_cache_free(dlm_lockres_cache, res); } void dlm_lockres_put(struct dlm_lock_resource *res) @@ -685,11 +577,19 @@ static void dlm_init_lockres(struct dlm_ctxt *dlm, INIT_LIST_HEAD(&res->dirty); INIT_LIST_HEAD(&res->recovering); INIT_LIST_HEAD(&res->purge); + INIT_LIST_HEAD(&res->tracking); atomic_set(&res->asts_reserved, 0); res->migration_pending = 0; + res->inflight_locks = 0; + res->inflight_assert_workers = 0; + + res->dlm = dlm; kref_init(&res->refs); + atomic_inc(&dlm->res_tot_count); + atomic_inc(&dlm->res_cur_count); + /* just for consistency */ spin_lock(&res->spinlock); dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN); @@ -699,27 +599,126 @@ static void dlm_init_lockres(struct dlm_ctxt *dlm, res->last_used = 0; + spin_lock(&dlm->spinlock); + list_add_tail(&res->tracking, &dlm->tracking_list); + spin_unlock(&dlm->spinlock); + memset(res->lvb, 0, DLM_LVB_LEN); + memset(res->refmap, 0, sizeof(res->refmap)); } struct dlm_lock_resource *dlm_new_lockres(struct dlm_ctxt *dlm, const char *name, unsigned int namelen) { - struct dlm_lock_resource *res; + struct dlm_lock_resource *res = NULL; - res = kmalloc(sizeof(struct dlm_lock_resource), GFP_NOFS); + res = kmem_cache_zalloc(dlm_lockres_cache, GFP_NOFS); if (!res) - return NULL; + goto error; - res->lockname.name = kmalloc(namelen, GFP_NOFS); - if (!res->lockname.name) { - kfree(res); - return NULL; - } + res->lockname.name = kmem_cache_zalloc(dlm_lockname_cache, GFP_NOFS); + if (!res->lockname.name) + goto error; dlm_init_lockres(dlm, res, name, namelen); return res; + +error: + if (res && res->lockname.name) + kmem_cache_free(dlm_lockname_cache, (void *)res->lockname.name); + + if (res) + kmem_cache_free(dlm_lockres_cache, res); + return NULL; +} + +void dlm_lockres_set_refmap_bit(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, int bit) +{ + assert_spin_locked(&res->spinlock); + + mlog(0, "res %.*s, set node %u, %ps()\n", res->lockname.len, + res->lockname.name, bit, __builtin_return_address(0)); + + set_bit(bit, res->refmap); +} + +void dlm_lockres_clear_refmap_bit(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, int bit) +{ + assert_spin_locked(&res->spinlock); + + mlog(0, "res %.*s, clr node %u, %ps()\n", res->lockname.len, + res->lockname.name, bit, __builtin_return_address(0)); + + clear_bit(bit, res->refmap); +} + + +void dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res) +{ + assert_spin_locked(&res->spinlock); + + res->inflight_locks++; + + mlog(0, "%s: res %.*s, inflight++: now %u, %ps()\n", dlm->name, + res->lockname.len, res->lockname.name, res->inflight_locks, + __builtin_return_address(0)); +} + +void dlm_lockres_drop_inflight_ref(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res) +{ + assert_spin_locked(&res->spinlock); + + BUG_ON(res->inflight_locks == 0); + + res->inflight_locks--; + + mlog(0, "%s: res %.*s, inflight--: now %u, %ps()\n", dlm->name, + res->lockname.len, res->lockname.name, res->inflight_locks, + __builtin_return_address(0)); + + wake_up(&res->wq); +} + +void __dlm_lockres_grab_inflight_worker(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res) +{ + assert_spin_locked(&res->spinlock); + res->inflight_assert_workers++; + mlog(0, "%s:%.*s: inflight assert worker++: now %u\n", + dlm->name, res->lockname.len, res->lockname.name, + res->inflight_assert_workers); +} + +static void dlm_lockres_grab_inflight_worker(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res) +{ + spin_lock(&res->spinlock); + __dlm_lockres_grab_inflight_worker(dlm, res); + spin_unlock(&res->spinlock); +} + +static void __dlm_lockres_drop_inflight_worker(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res) +{ + assert_spin_locked(&res->spinlock); + BUG_ON(res->inflight_assert_workers == 0); + res->inflight_assert_workers--; + mlog(0, "%s:%.*s: inflight assert worker--: now %u\n", + dlm->name, res->lockname.len, res->lockname.name, + res->inflight_assert_workers); +} + +static void dlm_lockres_drop_inflight_worker(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res) +{ + spin_lock(&res->spinlock); + __dlm_lockres_drop_inflight_worker(dlm, res); + spin_unlock(&res->spinlock); } /* @@ -761,10 +760,35 @@ struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm, lookup: spin_lock(&dlm->spinlock); - tmpres = __dlm_lookup_lockres(dlm, lockid, namelen, hash); + tmpres = __dlm_lookup_lockres_full(dlm, lockid, namelen, hash); if (tmpres) { spin_unlock(&dlm->spinlock); - mlog(0, "found in hash!\n"); + spin_lock(&tmpres->spinlock); + /* Wait on the thread that is mastering the resource */ + if (tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN) { + __dlm_wait_on_lockres(tmpres); + BUG_ON(tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN); + spin_unlock(&tmpres->spinlock); + dlm_lockres_put(tmpres); + tmpres = NULL; + goto lookup; + } + + /* Wait on the resource purge to complete before continuing */ + if (tmpres->state & DLM_LOCK_RES_DROPPING_REF) { + BUG_ON(tmpres->owner == dlm->node_num); + __dlm_wait_on_lockres_flags(tmpres, + DLM_LOCK_RES_DROPPING_REF); + spin_unlock(&tmpres->spinlock); + dlm_lockres_put(tmpres); + tmpres = NULL; + goto lookup; + } + + /* Grab inflight ref to pin the resource */ + dlm_lockres_grab_inflight_ref(dlm, tmpres); + + spin_unlock(&tmpres->spinlock); if (res) dlm_lockres_put(res); res = tmpres; @@ -775,8 +799,7 @@ lookup: spin_unlock(&dlm->spinlock); mlog(0, "allocating a new resource\n"); /* nothing found and we need to allocate one. */ - alloc_mle = (struct dlm_master_list_entry *) - kmem_cache_alloc(dlm_mle_cache, GFP_NOFS); + alloc_mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS); if (!alloc_mle) goto leave; res = dlm_new_lockres(dlm, lockid, namelen); @@ -793,6 +816,7 @@ lookup: spin_lock(&res->spinlock); dlm_change_lockres_owner(dlm, res, dlm->node_num); __dlm_insert_lockres(dlm, res); + dlm_lockres_grab_inflight_ref(dlm, res); spin_unlock(&res->spinlock); spin_unlock(&dlm->spinlock); /* lockres still marked IN_PROGRESS */ @@ -805,29 +829,40 @@ lookup: /* if we found a block, wait for lock to be mastered by another node */ blocked = dlm_find_mle(dlm, &mle, (char *)lockid, namelen); if (blocked) { + int mig; if (mle->type == DLM_MLE_MASTER) { mlog(ML_ERROR, "master entry for nonexistent lock!\n"); BUG(); - } else if (mle->type == DLM_MLE_MIGRATION) { - /* migration is in progress! */ - /* the good news is that we now know the - * "current" master (mle->master). */ - + } + mig = (mle->type == DLM_MLE_MIGRATION); + /* if there is a migration in progress, let the migration + * finish before continuing. we can wait for the absence + * of the MIGRATION mle: either the migrate finished or + * one of the nodes died and the mle was cleaned up. + * if there is a BLOCK here, but it already has a master + * set, we are too late. the master does not have a ref + * for us in the refmap. detach the mle and drop it. + * either way, go back to the top and start over. */ + if (mig || mle->master != O2NM_MAX_NODES) { + BUG_ON(mig && mle->master == dlm->node_num); + /* we arrived too late. the master does not + * have a ref for us. retry. */ + mlog(0, "%s:%.*s: late on %s\n", + dlm->name, namelen, lockid, + mig ? "MIGRATION" : "BLOCK"); spin_unlock(&dlm->master_lock); - assert_spin_locked(&dlm->spinlock); - - /* set the lockres owner and hash it */ - spin_lock(&res->spinlock); - dlm_set_lockres_owner(dlm, res, mle->master); - __dlm_insert_lockres(dlm, res); - spin_unlock(&res->spinlock); spin_unlock(&dlm->spinlock); /* master is known, detach */ - dlm_mle_detach_hb_events(dlm, mle); + if (!mig) + dlm_mle_detach_hb_events(dlm, mle); dlm_put_mle(mle); mle = NULL; - goto wake_waiters; + /* this is lame, but we can't wait on either + * the mle or lockres waitqueue here */ + if (mig) + msleep(100); + goto lookup; } } else { /* go ahead and try to master lock on this node */ @@ -836,16 +871,16 @@ lookup: alloc_mle = NULL; dlm_init_mle(mle, DLM_MLE_MASTER, dlm, res, NULL, 0); set_bit(dlm->node_num, mle->maybe_map); - list_add(&mle->list, &dlm->master_list); + __dlm_insert_mle(dlm, mle); /* still holding the dlm spinlock, check the recovery map - * to see if there are any nodes that still need to be + * to see if there are any nodes that still need to be * considered. these will not appear in the mle nodemap * but they might own this lockres. wait on them. */ bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0); if (bit < O2NM_MAX_NODES) { - mlog(ML_NOTICE, "%s:%.*s: at least one node (%d) to" - "recover before lock mastery can begin\n", + mlog(0, "%s: res %.*s, At least one node (%d) " + "to recover before lock mastery can begin\n", dlm->name, namelen, (char *)lockid, bit); wait_on_recovery = 1; } @@ -858,6 +893,12 @@ lookup: /* finally add the lockres to its hash bucket */ __dlm_insert_lockres(dlm, res); + + /* Grab inflight ref to pin the resource */ + spin_lock(&res->spinlock); + dlm_lockres_grab_inflight_ref(dlm, res); + spin_unlock(&res->spinlock); + /* get an extra ref on the mle in case this is a BLOCK * if so, the creator of the BLOCK may try to put the last * ref at this time in the assert master handler, so we @@ -872,8 +913,8 @@ redo_request: * dlm spinlock would be detectable be a change on the mle, * so we only need to clear out the recovery map once. */ if (dlm_is_recovery_lock(lockid, namelen)) { - mlog(ML_NOTICE, "%s: recovery map is not empty, but " - "must master $RECOVERY lock now\n", dlm->name); + mlog(0, "%s: Recovery map is not empty, but must " + "master $RECOVERY lock now\n", dlm->name); if (!dlm_pre_master_reco_lockres(dlm, res)) wait_on_recovery = 0; else { @@ -882,7 +923,7 @@ redo_request: msleep(500); } continue; - } + } dlm_kick_recovery_thread(dlm); msleep(1000); @@ -891,8 +932,8 @@ redo_request: spin_lock(&dlm->spinlock); bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0); if (bit < O2NM_MAX_NODES) { - mlog(ML_NOTICE, "%s:%.*s: at least one node (%d) to" - "recover before lock mastery can begin\n", + mlog(0, "%s: res %.*s, At least one node (%d) " + "to recover before lock mastery can begin\n", dlm->name, namelen, (char *)lockid, bit); wait_on_recovery = 1; } else @@ -910,7 +951,7 @@ redo_request: ret = -EINVAL; dlm_node_iter_init(mle->vote_map, &iter); while ((nodenum = dlm_node_iter_next(&iter)) >= 0) { - ret = dlm_do_master_request(mle, nodenum); + ret = dlm_do_master_request(res, mle, nodenum); if (ret < 0) mlog_errno(ret); if (mle->master != O2NM_MAX_NODES) { @@ -921,8 +962,8 @@ redo_request: * yet, keep going until it does. this is how the * master will know that asserts are needed back to * the lower nodes. */ - mlog(0, "%s:%.*s: requests only up to %u but master " - "is %u, keep going\n", dlm->name, namelen, + mlog(0, "%s: res %.*s, Requests only up to %u but " + "master is %u, keep going\n", dlm->name, namelen, lockid, nodenum, mle->master); } } @@ -932,14 +973,13 @@ wait: ret = dlm_wait_for_lock_mastery(dlm, res, mle, &blocked); if (ret < 0) { wait_on_recovery = 1; - mlog(0, "%s:%.*s: node map changed, redo the " - "master request now, blocked=%d\n", - dlm->name, res->lockname.len, + mlog(0, "%s: res %.*s, Node map changed, redo the master " + "request now, blocked=%d\n", dlm->name, res->lockname.len, res->lockname.name, blocked); if (++tries > 20) { - mlog(ML_ERROR, "%s:%.*s: spinning on " - "dlm_wait_for_lock_mastery, blocked=%d\n", - dlm->name, res->lockname.len, + mlog(ML_ERROR, "%s: res %.*s, Spinning on " + "dlm_wait_for_lock_mastery, blocked = %d\n", + dlm->name, res->lockname.len, res->lockname.name, blocked); dlm_print_one_lock_resource(res); dlm_print_one_mle(mle); @@ -948,7 +988,8 @@ wait: goto redo_request; } - mlog(0, "lockres mastered by %u\n", res->owner); + mlog(0, "%s: res %.*s, Mastered by %u\n", dlm->name, res->lockname.len, + res->lockname.name, res->owner); /* make sure we never continue without this */ BUG_ON(res->owner == O2NM_MAX_NODES); @@ -998,7 +1039,7 @@ recheck: /* this will cause the master to re-assert across * the whole cluster, freeing up mles */ if (res->owner != dlm->node_num) { - ret = dlm_do_master_request(mle, res->owner); + ret = dlm_do_master_request(res, mle, res->owner); if (ret < 0) { /* give recovery a chance to run */ mlog(ML_ERROR, "link to %u went down?: %d\n", res->owner, ret); @@ -1026,7 +1067,7 @@ recheck: ret = dlm_restart_lock_mastery(dlm, res, mle, *blocked); b = (mle->type == DLM_MLE_BLOCK); if ((*blocked && !b) || (!*blocked && b)) { - mlog(0, "%s:%.*s: status change: old=%d new=%d\n", + mlog(0, "%s:%.*s: status change: old=%d new=%d\n", dlm->name, res->lockname.len, res->lockname.name, *blocked, b); *blocked = b; @@ -1062,6 +1103,8 @@ recheck: * now tell other nodes that I am * mastering this. */ mle->master = dlm->node_num; + /* ref was grabbed in get_lock_resource + * will be dropped in dlmlock_master */ assert = 1; sleep = 0; } @@ -1087,7 +1130,8 @@ recheck: (atomic_read(&mle->woken) == 1), timeo); if (res->owner == O2NM_MAX_NODES) { - mlog(0, "waiting again\n"); + mlog(0, "%s:%.*s: waiting again\n", dlm->name, + res->lockname.len, res->lockname.name); goto recheck; } mlog(0, "done waiting, master is %u\n", res->owner); @@ -1100,8 +1144,7 @@ recheck: m = dlm->node_num; mlog(0, "about to master %.*s here, this=%u\n", res->lockname.len, res->lockname.name, m); - ret = dlm_do_assert_master(dlm, res->lockname.name, - res->lockname.len, mle->vote_map, 0); + ret = dlm_do_assert_master(dlm, res, mle->vote_map, 0); if (ret) { /* This is a failure in the network path, * not in the response to the assert_master @@ -1117,6 +1160,8 @@ recheck: /* set the lockres owner */ spin_lock(&res->spinlock); + /* mastery reference obtained either during + * assert_master_handler or in get_lock_resource */ dlm_change_lockres_owner(dlm, res, m); spin_unlock(&res->spinlock); @@ -1250,7 +1295,7 @@ static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm, res->lockname.len, res->lockname.name); mle->type = DLM_MLE_MASTER; - mle->u.res = res; + mle->mleres = res; } } } @@ -1283,7 +1328,8 @@ static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm, * */ -static int dlm_do_master_request(struct dlm_master_list_entry *mle, int to) +static int dlm_do_master_request(struct dlm_lock_resource *res, + struct dlm_master_list_entry *mle, int to) { struct dlm_ctxt *dlm = mle->dlm; struct dlm_master_request request; @@ -1294,14 +1340,8 @@ static int dlm_do_master_request(struct dlm_master_list_entry *mle, int to) BUG_ON(mle->type == DLM_MLE_MIGRATION); - if (mle->type != DLM_MLE_MASTER) { - request.namelen = mle->u.name.len; - memcpy(request.name, mle->u.name.name, request.namelen); - } else { - request.namelen = mle->u.res->lockname.len; - memcpy(request.name, mle->u.res->lockname.name, - request.namelen); - } + request.namelen = (u8)mle->mnamelen; + memcpy(request.name, mle->mname, request.namelen); again: ret = o2net_send_message(DLM_MASTER_REQUEST_MSG, dlm->key, &request, @@ -1339,6 +1379,9 @@ again: case DLM_MASTER_RESP_YES: set_bit(to, mle->response_map); mlog(0, "node %u is the master, response=YES\n", to); + mlog(0, "%s:%.*s: master node %u now knows I have a " + "reference\n", dlm->name, res->lockname.len, + res->lockname.name, to); mle->master = to; break; case DLM_MASTER_RESP_NO: @@ -1379,7 +1422,8 @@ out: * * if possible, TRIM THIS DOWN!!! */ -int dlm_master_request_handler(struct o2net_msg *msg, u32 len, void *data) +int dlm_master_request_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data) { u8 response = DLM_MASTER_RESP_MAYBE; struct dlm_ctxt *dlm = data; @@ -1417,10 +1461,11 @@ way_up_top: /* take care of the easy cases up front */ spin_lock(&res->spinlock); - if (res->state & DLM_LOCK_RES_RECOVERING) { + if (res->state & (DLM_LOCK_RES_RECOVERING| + DLM_LOCK_RES_MIGRATING)) { spin_unlock(&res->spinlock); mlog(0, "returning DLM_MASTER_RESP_ERROR since res is " - "being recovered\n"); + "being recovered/migrated\n"); response = DLM_MASTER_RESP_ERROR; if (mle) kmem_cache_free(dlm_mle_cache, mle); @@ -1428,8 +1473,8 @@ way_up_top: } if (res->owner == dlm->node_num) { + dlm_lockres_set_refmap_bit(dlm, res, request->node_idx); spin_unlock(&res->spinlock); - // mlog(0, "this node is the master\n"); response = DLM_MASTER_RESP_YES; if (mle) kmem_cache_free(dlm_mle_cache, mle); @@ -1477,7 +1522,6 @@ way_up_top: mlog(0, "node %u is master, but trying to migrate to " "node %u.\n", tmpmle->master, tmpmle->new_master); if (tmpmle->master == dlm->node_num) { - response = DLM_MASTER_RESP_YES; mlog(ML_ERROR, "no owner on lockres, but this " "node is trying to migrate it to %u?!\n", tmpmle->new_master); @@ -1494,6 +1538,8 @@ way_up_top: * go back and clean the mles on any * other nodes */ dispatch_assert = 1; + dlm_lockres_set_refmap_bit(dlm, res, + request->node_idx); } else response = DLM_MASTER_RESP_NO; } else { @@ -1530,8 +1576,7 @@ way_up_top: spin_unlock(&dlm->master_lock); spin_unlock(&dlm->spinlock); - mle = (struct dlm_master_list_entry *) - kmem_cache_alloc(dlm_mle_cache, GFP_NOFS); + mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS); if (!mle) { response = DLM_MASTER_RESP_ERROR; mlog_errno(-ENOMEM); @@ -1544,7 +1589,7 @@ way_up_top: // "add the block.\n"); dlm_init_mle(mle, DLM_MLE_BLOCK, dlm, NULL, name, namelen); set_bit(request->node_idx, mle->maybe_map); - list_add(&mle->list, &dlm->master_list); + __dlm_insert_mle(dlm, mle); response = DLM_MASTER_RESP_NO; } else { // mlog(0, "mle was found\n"); @@ -1575,7 +1620,12 @@ way_up_top: dlm_put_mle(tmpmle); } send_response: - + /* + * __dlm_lookup_lockres() grabbed a reference to this lockres. + * The reference is released by dlm_assert_master_worker() under + * the call to dlm_dispatch_assert_master(). If + * dlm_assert_master_worker() isn't called, we drop it here. + */ if (dispatch_assert) { if (response != DLM_MASTER_RESP_YES) mlog(ML_ERROR, "invalid response %d\n", response); @@ -1585,12 +1635,17 @@ send_response: } mlog(0, "%u is the owner of %.*s, cleaning everyone else\n", dlm->node_num, res->lockname.len, res->lockname.name); - ret = dlm_dispatch_assert_master(dlm, res, 0, request->node_idx, + ret = dlm_dispatch_assert_master(dlm, res, 0, request->node_idx, DLM_ASSERT_MASTER_MLE_CLEANUP); if (ret < 0) { mlog(ML_ERROR, "failed to dispatch assert master work\n"); response = DLM_MASTER_RESP_ERROR; - } + dlm_lockres_put(res); + } else + dlm_lockres_grab_inflight_worker(dlm, res); + } else { + if (res) + dlm_lockres_put(res); } dlm_put(dlm); @@ -1607,17 +1662,24 @@ send_response: * can periodically run all locks owned by this node * and re-assert across the cluster... */ -static int dlm_do_assert_master(struct dlm_ctxt *dlm, const char *lockname, - unsigned int namelen, void *nodemap, - u32 flags) +static int dlm_do_assert_master(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, + void *nodemap, u32 flags) { struct dlm_assert_master assert; int to, tmpret; struct dlm_node_iter iter; int ret = 0; int reassert; + const char *lockname = res->lockname.name; + unsigned int namelen = res->lockname.len; BUG_ON(namelen > O2NM_MAX_NAME_LEN); + + spin_lock(&res->spinlock); + res->state |= DLM_LOCK_RES_SETREF_INPROG; + spin_unlock(&res->spinlock); + again: reassert = 0; @@ -1638,7 +1700,9 @@ again: tmpret = o2net_send_message(DLM_ASSERT_MASTER_MSG, dlm->key, &assert, sizeof(assert), to, &r); if (tmpret < 0) { - mlog(0, "assert_master returned %d!\n", tmpret); + mlog(ML_ERROR, "Error %d when sending message %u (key " + "0x%x) to node %u\n", tmpret, + DLM_ASSERT_MASTER_MSG, dlm->key, to); if (!dlm_is_host_down(tmpret)) { mlog(ML_ERROR, "unhandled error=%d!\n", tmpret); BUG(); @@ -1647,6 +1711,7 @@ again: mlog(0, "link to %d went down!\n", to); /* any nonzero status return will do */ ret = tmpret; + r = 0; } else if (r < 0) { /* ok, something horribly messed. kill thyself. */ mlog(ML_ERROR,"during assert master of %.*s to %u, " @@ -1661,17 +1726,39 @@ again: spin_unlock(&dlm->master_lock); spin_unlock(&dlm->spinlock); BUG(); - } else if (r == EAGAIN) { + } + + if (r & DLM_ASSERT_RESPONSE_REASSERT && + !(r & DLM_ASSERT_RESPONSE_MASTERY_REF)) { + mlog(ML_ERROR, "%.*s: very strange, " + "master MLE but no lockres on %u\n", + namelen, lockname, to); + } + + if (r & DLM_ASSERT_RESPONSE_REASSERT) { mlog(0, "%.*s: node %u create mles on other " - "nodes and requests a re-assert\n", + "nodes and requests a re-assert\n", namelen, lockname, to); reassert = 1; } + if (r & DLM_ASSERT_RESPONSE_MASTERY_REF) { + mlog(0, "%.*s: node %u has a reference to this " + "lockres, set the bit in the refmap\n", + namelen, lockname, to); + spin_lock(&res->spinlock); + dlm_lockres_set_refmap_bit(dlm, res, to); + spin_unlock(&res->spinlock); + } } if (reassert) goto again; + spin_lock(&res->spinlock); + res->state &= ~DLM_LOCK_RES_SETREF_INPROG; + spin_unlock(&res->spinlock); + wake_up(&res->wq); + return ret; } @@ -1684,7 +1771,8 @@ again: * * if possible, TRIM THIS DOWN!!! */ -int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data) +int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data) { struct dlm_ctxt *dlm = data; struct dlm_master_list_entry *mle = NULL; @@ -1693,7 +1781,7 @@ int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data) char *name; unsigned int namelen, hash; u32 flags; - int master_request = 0; + int master_request = 0, have_lockres_ref = 0; int ret = 0; if (!dlm_grab(dlm)) @@ -1760,7 +1848,7 @@ int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data) spin_unlock(&dlm->master_lock); spin_unlock(&dlm->spinlock); goto done; - } + } } } spin_unlock(&dlm->master_lock); @@ -1778,12 +1866,12 @@ int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data) if (!mle) { if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN && res->owner != assert->node_idx) { - mlog(ML_ERROR, "assert_master from " - "%u, but current owner is " - "%u! (%.*s)\n", - assert->node_idx, res->owner, - namelen, name); - goto kill; + mlog(ML_ERROR, "DIE! Mastery assert from %u, " + "but current owner is %u! (%.*s)\n", + assert->node_idx, res->owner, namelen, + name); + __dlm_print_one_lock_resource(res); + BUG(); } } else if (mle->type != DLM_MLE_MIGRATION) { if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) { @@ -1823,7 +1911,6 @@ int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data) ok: spin_unlock(&res->spinlock); } - spin_unlock(&dlm->spinlock); // mlog(0, "woo! got an assert_master from node %u!\n", // assert->node_idx); @@ -1831,7 +1918,7 @@ ok: int extra_ref = 0; int nn = -1; int rr, err = 0; - + spin_lock(&mle->spinlock); if (mle->type == DLM_MLE_BLOCK || mle->type == DLM_MLE_MIGRATION) extra_ref = 1; @@ -1839,10 +1926,12 @@ ok: /* MASTER mle: if any bits set in the response map * then the calling node needs to re-assert to clear * up nodes that this node contacted */ - while ((nn = find_next_bit (mle->response_map, O2NM_MAX_NODES, + while ((nn = find_next_bit (mle->response_map, O2NM_MAX_NODES, nn+1)) < O2NM_MAX_NODES) { - if (nn != dlm->node_num && nn != assert->node_idx) + if (nn != dlm->node_num && nn != assert->node_idx) { master_request = 1; + break; + } } } mle->master = assert->node_idx; @@ -1851,6 +1940,7 @@ ok: spin_unlock(&mle->spinlock); if (res) { + int wake = 0; spin_lock(&res->spinlock); if (mle->type == DLM_MLE_MIGRATION) { mlog(0, "finishing off migration of lockres %.*s, " @@ -1858,18 +1948,21 @@ ok: res->lockname.len, res->lockname.name, dlm->node_num, mle->new_master); res->state &= ~DLM_LOCK_RES_MIGRATING; + wake = 1; dlm_change_lockres_owner(dlm, res, mle->new_master); BUG_ON(res->state & DLM_LOCK_RES_DIRTY); } else { dlm_change_lockres_owner(dlm, res, mle->master); } spin_unlock(&res->spinlock); + have_lockres_ref = 1; + if (wake) + wake_up(&res->wq); } /* master is known, detach if not already detached. * ensures that only one assert_master call will happen * on this mle. */ - spin_lock(&dlm->spinlock); spin_lock(&dlm->master_lock); rr = atomic_read(&mle->mle_refs.refcount); @@ -1891,7 +1984,7 @@ ok: assert->node_idx, rr, extra_ref, mle->inuse); dlm_print_one_mle(mle); } - list_del_init(&mle->list); + __dlm_unlink_mle(dlm, mle); __dlm_mle_detach_hb_events(dlm, mle); __dlm_put_mle(mle); if (extra_ref) { @@ -1902,7 +1995,6 @@ ok: __dlm_put_mle(mle); } spin_unlock(&dlm->master_lock); - spin_unlock(&dlm->spinlock); } else if (res) { if (res->owner != assert->node_idx) { mlog(0, "assert_master from %u, but current " @@ -1910,15 +2002,32 @@ ok: res->owner, namelen, name); } } + spin_unlock(&dlm->spinlock); done: ret = 0; - if (res) - dlm_lockres_put(res); + if (res) { + spin_lock(&res->spinlock); + res->state |= DLM_LOCK_RES_SETREF_INPROG; + spin_unlock(&res->spinlock); + *ret_data = (void *)res; + } dlm_put(dlm); if (master_request) { mlog(0, "need to tell master to reassert\n"); - ret = EAGAIN; // positive. negative would shoot down the node. + /* positive. negative would shoot down the node. */ + ret |= DLM_ASSERT_RESPONSE_REASSERT; + if (!have_lockres_ref) { + mlog(ML_ERROR, "strange, got assert from %u, MASTER " + "mle present here for %s:%.*s, but no lockres!\n", + assert->node_idx, dlm->name, namelen, name); + } + } + if (have_lockres_ref) { + /* let the master know we have a reference to the lockres */ + ret |= DLM_ASSERT_RESPONSE_MASTERY_REF; + mlog(0, "%s:%.*s: got assert from %u, need a ref\n", + dlm->name, namelen, name, assert->node_idx); } return ret; @@ -1929,17 +2038,31 @@ kill: __dlm_print_one_lock_resource(res); spin_unlock(&res->spinlock); spin_unlock(&dlm->spinlock); - dlm_lockres_put(res); + *ret_data = (void *)res; dlm_put(dlm); return -EINVAL; } +void dlm_assert_master_post_handler(int status, void *data, void *ret_data) +{ + struct dlm_lock_resource *res = (struct dlm_lock_resource *)ret_data; + + if (ret_data) { + spin_lock(&res->spinlock); + res->state &= ~DLM_LOCK_RES_SETREF_INPROG; + spin_unlock(&res->spinlock); + wake_up(&res->wq); + dlm_lockres_put(res); + } + return; +} + int dlm_dispatch_assert_master(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, int ignore_higher, u8 request_from, u32 flags) { struct dlm_work_item *item; - item = kcalloc(1, sizeof(*item), GFP_NOFS); + item = kzalloc(sizeof(*item), GFP_ATOMIC); if (!item) return -ENOMEM; @@ -1953,10 +2076,10 @@ int dlm_dispatch_assert_master(struct dlm_ctxt *dlm, item->u.am.request_from = request_from; item->u.am.flags = flags; - if (ignore_higher) - mlog(0, "IGNORE HIGHER: %.*s\n", res->lockname.len, + if (ignore_higher) + mlog(0, "IGNORE HIGHER: %.*s\n", res->lockname.len, res->lockname.name); - + spin_lock(&dlm->work_lock); list_add_tail(&item->list, &dlm->work_list); spin_unlock(&dlm->work_lock); @@ -2023,9 +2146,7 @@ static void dlm_assert_master_worker(struct dlm_work_item *item, void *data) * even if one or more nodes die */ mlog(0, "worker about to master %.*s here, this=%u\n", res->lockname.len, res->lockname.name, dlm->node_num); - ret = dlm_do_assert_master(dlm, res->lockname.name, - res->lockname.len, - nodemap, flags); + ret = dlm_do_assert_master(dlm, res, nodemap, flags); if (ret < 0) { /* no need to restart, we are done */ if (!dlm_is_host_down(ret)) @@ -2036,6 +2157,8 @@ static void dlm_assert_master_worker(struct dlm_work_item *item, void *data) dlm_lockres_release_ast(dlm, res); put: + dlm_lockres_drop_inflight_worker(dlm, res); + dlm_lockres_put(res); mlog(0, "finished with dlm_assert_master_worker\n"); @@ -2048,7 +2171,7 @@ put: * think that $RECOVERY is currently mastered by a dead node. If so, * we wait a short time to allow that node to get notified by its own * heartbeat stack, then check again. All $RECOVERY lock resources - * mastered by dead nodes are purged when the hearbeat callback is + * mastered by dead nodes are purged when the hearbeat callback is * fired, so we can know for sure that it is safe to continue once * the node returns a live node or no node. */ static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm, @@ -2089,7 +2212,7 @@ static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm, ret = -EAGAIN; } spin_unlock(&dlm->spinlock); - mlog(0, "%s: reco lock master is %u\n", dlm->name, + mlog(0, "%s: reco lock master is %u\n", dlm->name, master); break; } @@ -2097,83 +2220,253 @@ static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm, return ret; } - /* - * DLM_MIGRATE_LOCKRES + * DLM_DEREF_LOCKRES_MSG */ +int dlm_drop_lockres_ref(struct dlm_ctxt *dlm, struct dlm_lock_resource *res) +{ + struct dlm_deref_lockres deref; + int ret = 0, r; + const char *lockname; + unsigned int namelen; + + lockname = res->lockname.name; + namelen = res->lockname.len; + BUG_ON(namelen > O2NM_MAX_NAME_LEN); + + memset(&deref, 0, sizeof(deref)); + deref.node_idx = dlm->node_num; + deref.namelen = namelen; + memcpy(deref.name, lockname, namelen); + + ret = o2net_send_message(DLM_DEREF_LOCKRES_MSG, dlm->key, + &deref, sizeof(deref), res->owner, &r); + if (ret < 0) + mlog(ML_ERROR, "%s: res %.*s, error %d send DEREF to node %u\n", + dlm->name, namelen, lockname, ret, res->owner); + else if (r < 0) { + /* BAD. other node says I did not have a ref. */ + mlog(ML_ERROR, "%s: res %.*s, DEREF to node %u got %d\n", + dlm->name, namelen, lockname, res->owner, r); + dlm_print_one_lock_resource(res); + BUG(); + } + return ret; +} -int dlm_migrate_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, - u8 target) +int dlm_deref_lockres_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data) { - struct dlm_master_list_entry *mle = NULL; - struct dlm_master_list_entry *oldmle = NULL; - struct dlm_migratable_lockres *mres = NULL; - int ret = -EINVAL; - const char *name; + struct dlm_ctxt *dlm = data; + struct dlm_deref_lockres *deref = (struct dlm_deref_lockres *)msg->buf; + struct dlm_lock_resource *res = NULL; + char *name; unsigned int namelen; - int mle_added = 0; - struct list_head *queue, *iter; - int i; - struct dlm_lock *lock; - int empty = 1; + int ret = -EINVAL; + u8 node; + unsigned int hash; + struct dlm_work_item *item; + int cleared = 0; + int dispatch = 0; if (!dlm_grab(dlm)) - return -EINVAL; + return 0; - name = res->lockname.name; - namelen = res->lockname.len; + name = deref->name; + namelen = deref->namelen; + node = deref->node_idx; - mlog(0, "migrating %.*s to %u\n", namelen, name, target); + if (namelen > DLM_LOCKID_NAME_MAX) { + mlog(ML_ERROR, "Invalid name length!"); + goto done; + } + if (deref->node_idx >= O2NM_MAX_NODES) { + mlog(ML_ERROR, "Invalid node number: %u\n", node); + goto done; + } + + hash = dlm_lockid_hash(name, namelen); + + spin_lock(&dlm->spinlock); + res = __dlm_lookup_lockres_full(dlm, name, namelen, hash); + if (!res) { + spin_unlock(&dlm->spinlock); + mlog(ML_ERROR, "%s:%.*s: bad lockres name\n", + dlm->name, namelen, name); + goto done; + } + spin_unlock(&dlm->spinlock); - /* - * ensure this lockres is a proper candidate for migration - */ spin_lock(&res->spinlock); - if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) { - mlog(0, "cannot migrate lockres with unknown owner!\n"); - spin_unlock(&res->spinlock); - goto leave; + if (res->state & DLM_LOCK_RES_SETREF_INPROG) + dispatch = 1; + else { + BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF); + if (test_bit(node, res->refmap)) { + dlm_lockres_clear_refmap_bit(dlm, res, node); + cleared = 1; + } } - if (res->owner != dlm->node_num) { - mlog(0, "cannot migrate lockres this node doesn't own!\n"); - spin_unlock(&res->spinlock); - goto leave; + spin_unlock(&res->spinlock); + + if (!dispatch) { + if (cleared) + dlm_lockres_calc_usage(dlm, res); + else { + mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref " + "but it is already dropped!\n", dlm->name, + res->lockname.len, res->lockname.name, node); + dlm_print_one_lock_resource(res); + } + ret = 0; + goto done; } - mlog(0, "checking queues...\n"); - queue = &res->granted; - for (i=0; i<3; i++) { - list_for_each(iter, queue) { - lock = list_entry (iter, struct dlm_lock, list); - empty = 0; - if (lock->ml.node == dlm->node_num) { - mlog(0, "found a lock owned by this node " - "still on the %s queue! will not " - "migrate this lockres\n", - i==0 ? "granted" : - (i==1 ? "converting" : "blocked")); - spin_unlock(&res->spinlock); - ret = -ENOTEMPTY; - goto leave; + + item = kzalloc(sizeof(*item), GFP_NOFS); + if (!item) { + ret = -ENOMEM; + mlog_errno(ret); + goto done; + } + + dlm_init_work_item(dlm, item, dlm_deref_lockres_worker, NULL); + item->u.dl.deref_res = res; + item->u.dl.deref_node = node; + + spin_lock(&dlm->work_lock); + list_add_tail(&item->list, &dlm->work_list); + spin_unlock(&dlm->work_lock); + + queue_work(dlm->dlm_worker, &dlm->dispatched_work); + return 0; + +done: + if (res) + dlm_lockres_put(res); + dlm_put(dlm); + + return ret; +} + +static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data) +{ + struct dlm_ctxt *dlm; + struct dlm_lock_resource *res; + u8 node; + u8 cleared = 0; + + dlm = item->dlm; + res = item->u.dl.deref_res; + node = item->u.dl.deref_node; + + spin_lock(&res->spinlock); + BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF); + if (test_bit(node, res->refmap)) { + __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG); + dlm_lockres_clear_refmap_bit(dlm, res, node); + cleared = 1; + } + spin_unlock(&res->spinlock); + + if (cleared) { + mlog(0, "%s:%.*s node %u ref dropped in dispatch\n", + dlm->name, res->lockname.len, res->lockname.name, node); + dlm_lockres_calc_usage(dlm, res); + } else { + mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref " + "but it is already dropped!\n", dlm->name, + res->lockname.len, res->lockname.name, node); + dlm_print_one_lock_resource(res); + } + + dlm_lockres_put(res); +} + +/* + * A migrateable resource is one that is : + * 1. locally mastered, and, + * 2. zero local locks, and, + * 3. one or more non-local locks, or, one or more references + * Returns 1 if yes, 0 if not. + */ +static int dlm_is_lockres_migrateable(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res) +{ + enum dlm_lockres_list idx; + int nonlocal = 0, node_ref; + struct list_head *queue; + struct dlm_lock *lock; + u64 cookie; + + assert_spin_locked(&res->spinlock); + + /* delay migration when the lockres is in MIGRATING state */ + if (res->state & DLM_LOCK_RES_MIGRATING) + return 0; + + if (res->owner != dlm->node_num) + return 0; + + for (idx = DLM_GRANTED_LIST; idx <= DLM_BLOCKED_LIST; idx++) { + queue = dlm_list_idx_to_ptr(res, idx); + list_for_each_entry(lock, queue, list) { + if (lock->ml.node != dlm->node_num) { + nonlocal++; + continue; } + cookie = be64_to_cpu(lock->ml.cookie); + mlog(0, "%s: Not migrateable res %.*s, lock %u:%llu on " + "%s list\n", dlm->name, res->lockname.len, + res->lockname.name, + dlm_get_lock_cookie_node(cookie), + dlm_get_lock_cookie_seq(cookie), + dlm_list_in_text(idx)); + return 0; } - queue++; } - mlog(0, "all locks on this lockres are nonlocal. continuing\n"); - spin_unlock(&res->spinlock); - /* no work to do */ - if (empty) { - mlog(0, "no locks were found on this lockres! done!\n"); - ret = 0; - goto leave; + if (!nonlocal) { + node_ref = find_next_bit(res->refmap, O2NM_MAX_NODES, 0); + if (node_ref >= O2NM_MAX_NODES) + return 0; } - /* - * preallocate up front - * if this fails, abort - */ + mlog(0, "%s: res %.*s, Migrateable\n", dlm->name, res->lockname.len, + res->lockname.name); + + return 1; +} +/* + * DLM_MIGRATE_LOCKRES + */ + + +static int dlm_migrate_lockres(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res, u8 target) +{ + struct dlm_master_list_entry *mle = NULL; + struct dlm_master_list_entry *oldmle = NULL; + struct dlm_migratable_lockres *mres = NULL; + int ret = 0; + const char *name; + unsigned int namelen; + int mle_added = 0; + int wake = 0; + + if (!dlm_grab(dlm)) + return -EINVAL; + + BUG_ON(target == O2NM_MAX_NODES); + + name = res->lockname.name; + namelen = res->lockname.len; + + mlog(0, "%s: Migrating %.*s to node %u\n", dlm->name, namelen, name, + target); + + /* preallocate up front. if this fails, abort */ ret = -ENOMEM; mres = (struct dlm_migratable_lockres *) __get_free_page(GFP_NOFS); if (!mres) { @@ -2181,8 +2474,7 @@ int dlm_migrate_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, goto leave; } - mle = (struct dlm_master_list_entry *) kmem_cache_alloc(dlm_mle_cache, - GFP_NOFS); + mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS); if (!mle) { mlog_errno(ret); goto leave; @@ -2190,35 +2482,10 @@ int dlm_migrate_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, ret = 0; /* - * find a node to migrate the lockres to - */ - - mlog(0, "picking a migration node\n"); - spin_lock(&dlm->spinlock); - /* pick a new node */ - if (!test_bit(target, dlm->domain_map) || - target >= O2NM_MAX_NODES) { - target = dlm_pick_migration_target(dlm, res); - } - mlog(0, "node %u chosen for migration\n", target); - - if (target >= O2NM_MAX_NODES || - !test_bit(target, dlm->domain_map)) { - /* target chosen is not alive */ - ret = -EINVAL; - } - - if (ret) { - spin_unlock(&dlm->spinlock); - goto fail; - } - - mlog(0, "continuing with target = %u\n", target); - - /* * clear any existing master requests and * add the migration mle to the list */ + spin_lock(&dlm->spinlock); spin_lock(&dlm->master_lock); ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, name, namelen, target, dlm->node_num); @@ -2241,6 +2508,7 @@ int dlm_migrate_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, res->lockname.name, target); spin_lock(&res->spinlock); res->state &= ~DLM_LOCK_RES_MIGRATING; + wake = 1; spin_unlock(&res->spinlock); ret = -EINVAL; } @@ -2258,6 +2526,7 @@ fail: dlm_put_mle(mle); } else if (mle) { kmem_cache_free(dlm_mle_cache, mle); + mle = NULL; } goto leave; } @@ -2268,6 +2537,9 @@ fail: * the lockres */ + /* now that remote nodes are spinning on the MIGRATING flag, + * ensure that all assert_master work is flushed. */ + flush_workqueue(dlm->dlm_worker); /* get an extra reference on the mle. * otherwise the assert_master from the new @@ -2296,7 +2568,11 @@ fail: dlm_put_mle_inuse(mle); spin_lock(&res->spinlock); res->state &= ~DLM_LOCK_RES_MIGRATING; + wake = 1; spin_unlock(&res->spinlock); + if (dlm_is_host_down(ret)) + dlm_wait_for_node_death(dlm, target, + DLM_NODE_DEATH_WAIT_MAX); goto leave; } @@ -2308,7 +2584,7 @@ fail: * is complete everywhere. if the target dies while this is * going on, some nodes could potentially see the target as the * master, so it is important that my recovery finds the migration - * mle and sets the master to UNKNONWN. */ + * mle and sets the master to UNKNOWN. */ /* wait for new node to assert master */ @@ -2322,28 +2598,29 @@ fail: res->owner == target) break; - mlog(0, "timed out during migration\n"); - /* avoid hang during shutdown when migrating lockres + mlog(0, "%s:%.*s: timed out during migration\n", + dlm->name, res->lockname.len, res->lockname.name); + /* avoid hang during shutdown when migrating lockres * to a node which also goes down */ if (dlm_is_node_dead(dlm, target)) { mlog(0, "%s:%.*s: expected migration " "target %u is no longer up, restarting\n", dlm->name, res->lockname.len, res->lockname.name, target); - ret = -ERESTARTSYS; + ret = -EINVAL; + /* migration failed, detach and clean up mle */ + dlm_mle_detach_hb_events(dlm, mle); + dlm_put_mle(mle); + dlm_put_mle_inuse(mle); + spin_lock(&res->spinlock); + res->state &= ~DLM_LOCK_RES_MIGRATING; + wake = 1; + spin_unlock(&res->spinlock); + goto leave; } - } - if (ret == -ERESTARTSYS) { - /* migration failed, detach and clean up mle */ - dlm_mle_detach_hb_events(dlm, mle); - dlm_put_mle(mle); - dlm_put_mle_inuse(mle); - spin_lock(&res->spinlock); - res->state &= ~DLM_LOCK_RES_MIGRATING; - spin_unlock(&res->spinlock); - goto leave; - } - /* TODO: if node died: stop, clean up, return error */ + } else + mlog(0, "%s:%.*s: caught signal during migration\n", + dlm->name, res->lockname.len, res->lockname.name); } /* all done, set the owner, clear the flag */ @@ -2366,16 +2643,62 @@ leave: if (ret < 0) dlm_kick_thread(dlm, res); - /* TODO: cleanup */ + /* wake up waiters if the MIGRATING flag got set + * but migration failed */ + if (wake) + wake_up(&res->wq); + if (mres) free_page((unsigned long)mres); dlm_put(dlm); - mlog(0, "returning %d\n", ret); + mlog(0, "%s: Migrating %.*s to %u, returns %d\n", dlm->name, namelen, + name, target, ret); return ret; } +#define DLM_MIGRATION_RETRY_MS 100 + +/* + * Should be called only after beginning the domain leave process. + * There should not be any remaining locks on nonlocal lock resources, + * and there should be no local locks left on locally mastered resources. + * + * Called with the dlm spinlock held, may drop it to do migration, but + * will re-acquire before exit. + * + * Returns: 1 if dlm->spinlock was dropped/retaken, 0 if never dropped + */ +int dlm_empty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res) +{ + int ret; + int lock_dropped = 0; + u8 target = O2NM_MAX_NODES; + + assert_spin_locked(&dlm->spinlock); + + spin_lock(&res->spinlock); + if (dlm_is_lockres_migrateable(dlm, res)) + target = dlm_pick_migration_target(dlm, res); + spin_unlock(&res->spinlock); + + if (target == O2NM_MAX_NODES) + goto leave; + + /* Wheee! Migrate lockres here! Will sleep so drop spinlock. */ + spin_unlock(&dlm->spinlock); + lock_dropped = 1; + ret = dlm_migrate_lockres(dlm, res, target); + if (ret) + mlog(0, "%s: res %.*s, Migrate to node %u failed with %d\n", + dlm->name, res->lockname.len, res->lockname.name, + target, ret); + spin_lock(&dlm->spinlock); +leave: + return lock_dropped; +} + int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock) { int ret; @@ -2396,7 +2719,7 @@ static int dlm_migration_can_proceed(struct dlm_ctxt *dlm, can_proceed = !!(res->state & DLM_LOCK_RES_MIGRATING); spin_unlock(&res->spinlock); - /* target has died, so make the caller break out of the + /* target has died, so make the caller break out of the * wait_event, but caller must recheck the domain_map */ spin_lock(&dlm->spinlock); if (!test_bit(mig_target, dlm->domain_map)) @@ -2405,7 +2728,8 @@ static int dlm_migration_can_proceed(struct dlm_ctxt *dlm, return can_proceed; } -int dlm_lockres_is_dirty(struct dlm_ctxt *dlm, struct dlm_lock_resource *res) +static int dlm_lockres_is_dirty(struct dlm_ctxt *dlm, + struct dlm_lock_resource *res) { int ret; spin_lock(&res->spinlock); @@ -2434,8 +2758,15 @@ static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm, __dlm_lockres_reserve_ast(res); spin_unlock(&res->spinlock); - /* now flush all the pending asts.. hang out for a bit */ + /* now flush all the pending asts */ dlm_kick_thread(dlm, res); + /* before waiting on DIRTY, block processes which may + * try to dirty the lockres before MIGRATING is set */ + spin_lock(&res->spinlock); + BUG_ON(res->state & DLM_LOCK_RES_BLOCK_DIRTY); + res->state |= DLM_LOCK_RES_BLOCK_DIRTY; + spin_unlock(&res->spinlock); + /* now wait on any pending asts and the DIRTY state */ wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res)); dlm_lockres_release_ast(dlm, res); @@ -2462,6 +2793,7 @@ again: goto again; } + ret = 0; /* did the target go down or die? */ spin_lock(&dlm->spinlock); if (!test_bit(target, dlm->domain_map)) { @@ -2472,9 +2804,21 @@ again: spin_unlock(&dlm->spinlock); /* + * if target is down, we need to clear DLM_LOCK_RES_BLOCK_DIRTY for + * another try; otherwise, we are sure the MIGRATING state is there, + * drop the unneded state which blocked threads trying to DIRTY + */ + spin_lock(&res->spinlock); + BUG_ON(!(res->state & DLM_LOCK_RES_BLOCK_DIRTY)); + res->state &= ~DLM_LOCK_RES_BLOCK_DIRTY; + if (!ret) + BUG_ON(!(res->state & DLM_LOCK_RES_MIGRATING)); + spin_unlock(&res->spinlock); + + /* * at this point: * - * o the DLM_LOCK_RES_MIGRATING flag is set + * o the DLM_LOCK_RES_MIGRATING flag is set if target not down * o there are no pending asts on this lockres * o all processes trying to reserve an ast on this * lockres must wait for the MIGRATING flag to clear @@ -2488,18 +2832,16 @@ again: static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm, struct dlm_lock_resource *res) { - struct list_head *iter, *iter2; struct list_head *queue = &res->granted; - int i; - struct dlm_lock *lock; + int i, bit; + struct dlm_lock *lock, *next; assert_spin_locked(&res->spinlock); BUG_ON(res->owner == dlm->node_num); for (i=0; i<3; i++) { - list_for_each_safe(iter, iter2, queue) { - lock = list_entry (iter, struct dlm_lock, list); + list_for_each_entry_safe(lock, next, queue, list) { if (lock->ml.node != dlm->node_num) { mlog(0, "putting lock for node %u\n", lock->ml.node); @@ -2508,65 +2850,83 @@ static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm, BUG_ON(!list_empty(&lock->bast_list)); BUG_ON(lock->ast_pending); BUG_ON(lock->bast_pending); + dlm_lockres_clear_refmap_bit(dlm, res, + lock->ml.node); list_del_init(&lock->list); dlm_lock_put(lock); + /* In a normal unlock, we would have added a + * DLM_UNLOCK_FREE_LOCK action. Force it. */ + dlm_lock_put(lock); } } queue++; } + bit = 0; + while (1) { + bit = find_next_bit(res->refmap, O2NM_MAX_NODES, bit); + if (bit >= O2NM_MAX_NODES) + break; + /* do not clear the local node reference, if there is a + * process holding this, let it drop the ref itself */ + if (bit != dlm->node_num) { + mlog(0, "%s:%.*s: node %u had a ref to this " + "migrating lockres, clearing\n", dlm->name, + res->lockname.len, res->lockname.name, bit); + dlm_lockres_clear_refmap_bit(dlm, res, bit); + } + bit++; + } } -/* for now this is not too intelligent. we will - * need stats to make this do the right thing. - * this just finds the first lock on one of the - * queues and uses that node as the target. */ +/* + * Pick a node to migrate the lock resource to. This function selects a + * potential target based first on the locks and then on refmap. It skips + * nodes that are in the process of exiting the domain. + */ static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm, struct dlm_lock_resource *res) { - int i; + enum dlm_lockres_list idx; struct list_head *queue = &res->granted; - struct list_head *iter; struct dlm_lock *lock; - int nodenum; + int noderef; + u8 nodenum = O2NM_MAX_NODES; assert_spin_locked(&dlm->spinlock); + assert_spin_locked(&res->spinlock); - spin_lock(&res->spinlock); - for (i=0; i<3; i++) { - list_for_each(iter, queue) { - /* up to the caller to make sure this node - * is alive */ - lock = list_entry (iter, struct dlm_lock, list); - if (lock->ml.node != dlm->node_num) { - spin_unlock(&res->spinlock); - return lock->ml.node; - } + /* Go through all the locks */ + for (idx = DLM_GRANTED_LIST; idx <= DLM_BLOCKED_LIST; idx++) { + queue = dlm_list_idx_to_ptr(res, idx); + list_for_each_entry(lock, queue, list) { + if (lock->ml.node == dlm->node_num) + continue; + if (test_bit(lock->ml.node, dlm->exit_domain_map)) + continue; + nodenum = lock->ml.node; + goto bail; } - queue++; } - spin_unlock(&res->spinlock); - mlog(0, "have not found a suitable target yet! checking domain map\n"); - /* ok now we're getting desperate. pick anyone alive. */ - nodenum = -1; + /* Go thru the refmap */ + noderef = -1; while (1) { - nodenum = find_next_bit(dlm->domain_map, - O2NM_MAX_NODES, nodenum+1); - mlog(0, "found %d in domain map\n", nodenum); - if (nodenum >= O2NM_MAX_NODES) + noderef = find_next_bit(res->refmap, O2NM_MAX_NODES, + noderef + 1); + if (noderef >= O2NM_MAX_NODES) break; - if (nodenum != dlm->node_num) { - mlog(0, "picking %d\n", nodenum); - return nodenum; - } + if (noderef == dlm->node_num) + continue; + if (test_bit(noderef, dlm->exit_domain_map)) + continue; + nodenum = noderef; + goto bail; } - mlog(0, "giving up. no master to migrate to\n"); - return DLM_LOCK_RES_OWNER_UNKNOWN; +bail: + return nodenum; } - - /* this is called by the new master once all lockres * data has been received */ static int dlm_do_migrate_request(struct dlm_ctxt *dlm, @@ -2575,7 +2935,7 @@ static int dlm_do_migrate_request(struct dlm_ctxt *dlm, struct dlm_node_iter *iter) { struct dlm_migrate_request migrate; - int ret, status = 0; + int ret, skip, status = 0; int nodenum; memset(&migrate, 0, sizeof(migrate)); @@ -2592,15 +2952,42 @@ static int dlm_do_migrate_request(struct dlm_ctxt *dlm, nodenum == new_master) continue; + /* We could race exit domain. If exited, skip. */ + spin_lock(&dlm->spinlock); + skip = (!test_bit(nodenum, dlm->domain_map)); + spin_unlock(&dlm->spinlock); + if (skip) { + clear_bit(nodenum, iter->node_map); + continue; + } + ret = o2net_send_message(DLM_MIGRATE_REQUEST_MSG, dlm->key, &migrate, sizeof(migrate), nodenum, &status); - if (ret < 0) - mlog_errno(ret); - else if (status < 0) { + if (ret < 0) { + mlog(ML_ERROR, "%s: res %.*s, Error %d send " + "MIGRATE_REQUEST to node %u\n", dlm->name, + migrate.namelen, migrate.name, ret, nodenum); + if (!dlm_is_host_down(ret)) { + mlog(ML_ERROR, "unhandled error=%d!\n", ret); + BUG(); + } + clear_bit(nodenum, iter->node_map); + ret = 0; + } else if (status < 0) { mlog(0, "migrate request (node %u) returned %d!\n", nodenum, status); ret = status; + } else if (status == DLM_MIGRATE_RESPONSE_MASTERY_REF) { + /* during the migration request we short-circuited + * the mastery of the lockres. make sure we have + * a mastery ref for nodenum */ + mlog(0, "%s:%.*s: need ref for node %u\n", + dlm->name, res->lockname.len, res->lockname.name, + nodenum); + spin_lock(&res->spinlock); + dlm_lockres_set_refmap_bit(dlm, res, nodenum); + spin_unlock(&res->spinlock); } } @@ -2619,7 +3006,8 @@ static int dlm_do_migrate_request(struct dlm_ctxt *dlm, * we will have no mle in the list to start with. now we can add an mle for * the migration and this should be the only one found for those scanning the * list. */ -int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data) +int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data, + void **ret_data) { struct dlm_ctxt *dlm = data; struct dlm_lock_resource *res = NULL; @@ -2637,8 +3025,7 @@ int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data) hash = dlm_lockid_hash(name, namelen); /* preallocate.. if this fails, abort */ - mle = (struct dlm_master_list_entry *) kmem_cache_alloc(dlm_mle_cache, - GFP_NOFS); + mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS); if (!mle) { ret = -ENOMEM; @@ -2648,8 +3035,6 @@ int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data) /* check for pre-existing lock */ spin_lock(&dlm->spinlock); res = __dlm_lookup_lockres(dlm, name, namelen, hash); - spin_lock(&dlm->master_lock); - if (res) { spin_lock(&res->spinlock); if (res->state & DLM_LOCK_RES_RECOVERING) { @@ -2667,14 +3052,15 @@ int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data) spin_unlock(&res->spinlock); } + spin_lock(&dlm->master_lock); /* ignore status. only nonzero status would BUG. */ ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, name, namelen, migrate->new_master, migrate->master); -unlock: spin_unlock(&dlm->master_lock); +unlock: spin_unlock(&dlm->spinlock); if (oldmle) { @@ -2709,8 +3095,6 @@ static int dlm_add_migration_mle(struct dlm_ctxt *dlm, *oldmle = NULL; - mlog_entry_void(); - assert_spin_locked(&dlm->spinlock); assert_spin_locked(&dlm->master_lock); @@ -2742,10 +3126,18 @@ static int dlm_add_migration_mle(struct dlm_ctxt *dlm, tmp->master = master; atomic_set(&tmp->woken, 1); wake_up(&tmp->wq); - /* remove it from the list so that only one - * mle will be found */ - list_del_init(&tmp->list); - __dlm_mle_detach_hb_events(dlm, mle); + /* remove it so that only one mle will be found */ + __dlm_unlink_mle(dlm, tmp); + __dlm_mle_detach_hb_events(dlm, tmp); + if (tmp->type == DLM_MLE_MASTER) { + ret = DLM_MIGRATE_RESPONSE_MASTERY_REF; + mlog(0, "%s:%.*s: master=%u, newmaster=%u, " + "telling master to get ref " + "for cleared out mle during " + "migration\n", dlm->name, + namelen, name, master, + new_master); + } } spin_unlock(&tmp->spinlock); } @@ -2753,143 +3145,166 @@ static int dlm_add_migration_mle(struct dlm_ctxt *dlm, /* now add a migration mle to the tail of the list */ dlm_init_mle(mle, DLM_MLE_MIGRATION, dlm, res, name, namelen); mle->new_master = new_master; + /* the new master will be sending an assert master for this. + * at that point we will get the refmap reference */ mle->master = master; /* do this for consistency with other mle types */ set_bit(new_master, mle->maybe_map); - list_add(&mle->list, &dlm->master_list); + __dlm_insert_mle(dlm, mle); return ret; } - -void dlm_clean_master_list(struct dlm_ctxt *dlm, u8 dead_node) +/* + * Sets the owner of the lockres, associated to the mle, to UNKNOWN + */ +static struct dlm_lock_resource *dlm_reset_mleres_owner(struct dlm_ctxt *dlm, + struct dlm_master_list_entry *mle) { - struct list_head *iter, *iter2; - struct dlm_master_list_entry *mle; struct dlm_lock_resource *res; - unsigned int hash; - mlog_entry("dlm=%s, dead node=%u\n", dlm->name, dead_node); -top: - assert_spin_locked(&dlm->spinlock); + /* Find the lockres associated to the mle and set its owner to UNK */ + res = __dlm_lookup_lockres(dlm, mle->mname, mle->mnamelen, + mle->mnamehash); + if (res) { + spin_unlock(&dlm->master_lock); - /* clean the master list */ - spin_lock(&dlm->master_lock); - list_for_each_safe(iter, iter2, &dlm->master_list) { - mle = list_entry(iter, struct dlm_master_list_entry, list); + /* move lockres onto recovery list */ + spin_lock(&res->spinlock); + dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN); + dlm_move_lockres_to_recovery_list(dlm, res); + spin_unlock(&res->spinlock); + dlm_lockres_put(res); - BUG_ON(mle->type != DLM_MLE_BLOCK && - mle->type != DLM_MLE_MASTER && - mle->type != DLM_MLE_MIGRATION); + /* about to get rid of mle, detach from heartbeat */ + __dlm_mle_detach_hb_events(dlm, mle); - /* MASTER mles are initiated locally. the waiting - * process will notice the node map change - * shortly. let that happen as normal. */ - if (mle->type == DLM_MLE_MASTER) - continue; + /* dump the mle */ + spin_lock(&dlm->master_lock); + __dlm_put_mle(mle); + spin_unlock(&dlm->master_lock); + } + return res; +} - /* BLOCK mles are initiated by other nodes. - * need to clean up if the dead node would have - * been the master. */ - if (mle->type == DLM_MLE_BLOCK) { - int bit; +static void dlm_clean_migration_mle(struct dlm_ctxt *dlm, + struct dlm_master_list_entry *mle) +{ + __dlm_mle_detach_hb_events(dlm, mle); - spin_lock(&mle->spinlock); - bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0); - if (bit != dead_node) { - mlog(0, "mle found, but dead node %u would " - "not have been master\n", dead_node); - spin_unlock(&mle->spinlock); - } else { - /* must drop the refcount by one since the - * assert_master will never arrive. this - * may result in the mle being unlinked and - * freed, but there may still be a process - * waiting in the dlmlock path which is fine. */ - mlog(0, "node %u was expected master\n", - dead_node); - atomic_set(&mle->woken, 1); - spin_unlock(&mle->spinlock); - wake_up(&mle->wq); - /* do not need events any longer, so detach - * from heartbeat */ - __dlm_mle_detach_hb_events(dlm, mle); - __dlm_put_mle(mle); - } - continue; - } + spin_lock(&mle->spinlock); + __dlm_unlink_mle(dlm, mle); + atomic_set(&mle->woken, 1); + spin_unlock(&mle->spinlock); - /* everything else is a MIGRATION mle */ - - /* the rule for MIGRATION mles is that the master - * becomes UNKNOWN if *either* the original or - * the new master dies. all UNKNOWN lockreses - * are sent to whichever node becomes the recovery - * master. the new master is responsible for - * determining if there is still a master for - * this lockres, or if he needs to take over - * mastery. either way, this node should expect - * another message to resolve this. */ - if (mle->master != dead_node && - mle->new_master != dead_node) - continue; + wake_up(&mle->wq); +} - /* if we have reached this point, this mle needs to - * be removed from the list and freed. */ +static void dlm_clean_block_mle(struct dlm_ctxt *dlm, + struct dlm_master_list_entry *mle, u8 dead_node) +{ + int bit; - /* remove from the list early. NOTE: unlinking - * list_head while in list_for_each_safe */ - __dlm_mle_detach_hb_events(dlm, mle); - spin_lock(&mle->spinlock); - list_del_init(&mle->list); + BUG_ON(mle->type != DLM_MLE_BLOCK); + + spin_lock(&mle->spinlock); + bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0); + if (bit != dead_node) { + mlog(0, "mle found, but dead node %u would not have been " + "master\n", dead_node); + spin_unlock(&mle->spinlock); + } else { + /* Must drop the refcount by one since the assert_master will + * never arrive. This may result in the mle being unlinked and + * freed, but there may still be a process waiting in the + * dlmlock path which is fine. */ + mlog(0, "node %u was expected master\n", dead_node); atomic_set(&mle->woken, 1); spin_unlock(&mle->spinlock); wake_up(&mle->wq); - mlog(0, "%s: node %u died during migration from " - "%u to %u!\n", dlm->name, dead_node, - mle->master, mle->new_master); - /* if there is a lockres associated with this - * mle, find it and set its owner to UNKNOWN */ - hash = dlm_lockid_hash(mle->u.name.name, mle->u.name.len); - res = __dlm_lookup_lockres(dlm, mle->u.name.name, - mle->u.name.len, hash); - if (res) { - /* unfortunately if we hit this rare case, our - * lock ordering is messed. we need to drop - * the master lock so that we can take the - * lockres lock, meaning that we will have to - * restart from the head of list. */ - spin_unlock(&dlm->master_lock); + /* Do not need events any longer, so detach from heartbeat */ + __dlm_mle_detach_hb_events(dlm, mle); + __dlm_put_mle(mle); + } +} - /* move lockres onto recovery list */ - spin_lock(&res->spinlock); - dlm_set_lockres_owner(dlm, res, - DLM_LOCK_RES_OWNER_UNKNOWN); - dlm_move_lockres_to_recovery_list(dlm, res); - spin_unlock(&res->spinlock); - dlm_lockres_put(res); +void dlm_clean_master_list(struct dlm_ctxt *dlm, u8 dead_node) +{ + struct dlm_master_list_entry *mle; + struct dlm_lock_resource *res; + struct hlist_head *bucket; + struct hlist_node *tmp; + unsigned int i; - /* about to get rid of mle, detach from heartbeat */ - __dlm_mle_detach_hb_events(dlm, mle); + mlog(0, "dlm=%s, dead node=%u\n", dlm->name, dead_node); +top: + assert_spin_locked(&dlm->spinlock); - /* dump the mle */ - spin_lock(&dlm->master_lock); - __dlm_put_mle(mle); - spin_unlock(&dlm->master_lock); + /* clean the master list */ + spin_lock(&dlm->master_lock); + for (i = 0; i < DLM_HASH_BUCKETS; i++) { + bucket = dlm_master_hash(dlm, i); + hlist_for_each_entry_safe(mle, tmp, bucket, master_hash_node) { + BUG_ON(mle->type != DLM_MLE_BLOCK && + mle->type != DLM_MLE_MASTER && + mle->type != DLM_MLE_MIGRATION); + + /* MASTER mles are initiated locally. The waiting + * process will notice the node map change shortly. + * Let that happen as normal. */ + if (mle->type == DLM_MLE_MASTER) + continue; + + /* BLOCK mles are initiated by other nodes. Need to + * clean up if the dead node would have been the + * master. */ + if (mle->type == DLM_MLE_BLOCK) { + dlm_clean_block_mle(dlm, mle, dead_node); + continue; + } - /* restart */ - goto top; - } + /* Everything else is a MIGRATION mle */ + + /* The rule for MIGRATION mles is that the master + * becomes UNKNOWN if *either* the original or the new + * master dies. All UNKNOWN lockres' are sent to + * whichever node becomes the recovery master. The new + * master is responsible for determining if there is + * still a master for this lockres, or if he needs to + * take over mastery. Either way, this node should + * expect another message to resolve this. */ + + if (mle->master != dead_node && + mle->new_master != dead_node) + continue; + + /* If we have reached this point, this mle needs to be + * removed from the list and freed. */ + dlm_clean_migration_mle(dlm, mle); + + mlog(0, "%s: node %u died during migration from " + "%u to %u!\n", dlm->name, dead_node, mle->master, + mle->new_master); + + /* If we find a lockres associated with the mle, we've + * hit this rare case that messes up our lock ordering. + * If so, we need to drop the master lock so that we can + * take the lockres lock, meaning that we will have to + * restart from the head of list. */ + res = dlm_reset_mleres_owner(dlm, mle); + if (res) + /* restart */ + goto top; - /* this may be the last reference */ - __dlm_put_mle(mle); + /* This may be the last reference */ + __dlm_put_mle(mle); + } } spin_unlock(&dlm->master_lock); } - int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, u8 old_master) { @@ -2902,6 +3317,13 @@ int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, clear_bit(dlm->node_num, iter.node_map); spin_unlock(&dlm->spinlock); + /* ownership of the lockres is changing. account for the + * mastery reference here since old_master will briefly have + * a reference after the migration completes */ + spin_lock(&res->spinlock); + dlm_lockres_set_refmap_bit(dlm, res, old_master); + spin_unlock(&res->spinlock); + mlog(0, "now time to do a migrate request to other nodes\n"); ret = dlm_do_migrate_request(dlm, res, old_master, dlm->node_num, &iter); @@ -2914,8 +3336,7 @@ int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, res->lockname.len, res->lockname.name); /* this call now finishes out the nodemap * even if one or more nodes die */ - ret = dlm_do_assert_master(dlm, res->lockname.name, - res->lockname.len, iter.node_map, + ret = dlm_do_assert_master(dlm, res, iter.node_map, DLM_ASSERT_MASTER_FINISH_MIGRATION); if (ret < 0) { /* no longer need to retry. all living nodes contacted. */ @@ -2927,8 +3348,7 @@ int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, set_bit(old_master, iter.node_map); mlog(0, "doing assert master of %.*s back to %u\n", res->lockname.len, res->lockname.name, old_master); - ret = dlm_do_assert_master(dlm, res->lockname.name, - res->lockname.len, iter.node_map, + ret = dlm_do_assert_master(dlm, res, iter.node_map, DLM_ASSERT_MASTER_FINISH_MIGRATION); if (ret < 0) { mlog(0, "assert master to original master failed " @@ -3001,3 +3421,41 @@ void dlm_lockres_release_ast(struct dlm_ctxt *dlm, wake_up(&res->wq); wake_up(&dlm->migration_wq); } + +void dlm_force_free_mles(struct dlm_ctxt *dlm) +{ + int i; + struct hlist_head *bucket; + struct dlm_master_list_entry *mle; + struct hlist_node *tmp; + + /* + * We notified all other nodes that we are exiting the domain and + * marked the dlm state to DLM_CTXT_LEAVING. If any mles are still + * around we force free them and wake any processes that are waiting + * on the mles + */ + spin_lock(&dlm->spinlock); + spin_lock(&dlm->master_lock); + + BUG_ON(dlm->dlm_state != DLM_CTXT_LEAVING); + BUG_ON((find_next_bit(dlm->domain_map, O2NM_MAX_NODES, 0) < O2NM_MAX_NODES)); + + for (i = 0; i < DLM_HASH_BUCKETS; i++) { + bucket = dlm_master_hash(dlm, i); + hlist_for_each_entry_safe(mle, tmp, bucket, master_hash_node) { + if (mle->type != DLM_MLE_BLOCK) { + mlog(ML_ERROR, "bad mle: %p\n", mle); + dlm_print_one_mle(mle); + } + atomic_set(&mle->woken, 1); + wake_up(&mle->wq); + + __dlm_unlink_mle(dlm, mle); + __dlm_mle_detach_hb_events(dlm, mle); + __dlm_put_mle(mle); + } + } + spin_unlock(&dlm->master_lock); + spin_unlock(&dlm->spinlock); +} |