dlm: fix race between remove and lookup
authorDavid Teigland <teigland@redhat.com>
Thu, 14 Jun 2012 17:17:32 +0000 (12:17 -0500)
committerDavid Teigland <teigland@redhat.com>
Mon, 16 Jul 2012 19:18:01 +0000 (14:18 -0500)
It was possible for a remove message on an old
rsb to be sent after a lookup message on a new
rsb, where the rsbs were for the same resource
name.  This could lead to a missing directory
entry for the new rsb.

It is fixed by keeping a copy of the resource
name being removed until after the remove has
been sent.  A lookup checks if this in-progress
remove matches the name it is looking up.

Signed-off-by: David Teigland <teigland@redhat.com>
fs/dlm/dlm_internal.h
fs/dlm/lock.c
fs/dlm/lockspace.c

index a5f82d5..9d3e485 100644 (file)
@@ -498,6 +498,13 @@ struct rcom_lock {
        char                    rl_lvb[0];
 };
 
+/*
+ * The max number of resources per rsbtbl bucket that shrink will attempt
+ * to remove in each iteration.
+ */
+
+#define DLM_REMOVE_NAMES_MAX 8
+
 struct dlm_ls {
        struct list_head        ls_list;        /* list of lockspaces */
        dlm_lockspace_t         *ls_local_handle;
@@ -531,6 +538,12 @@ struct dlm_ls {
        int                     ls_new_rsb_count;
        struct list_head        ls_new_rsb;     /* new rsb structs */
 
+       spinlock_t              ls_remove_spin;
+       char                    ls_remove_name[DLM_RESNAME_MAXLEN+1];
+       char                    *ls_remove_names[DLM_REMOVE_NAMES_MAX];
+       int                     ls_remove_len;
+       int                     ls_remove_lens[DLM_REMOVE_NAMES_MAX];
+
        struct list_head        ls_nodes;       /* current nodes in ls */
        struct list_head        ls_nodes_gone;  /* dead node list, recovery */
        int                     ls_num_nodes;   /* number of nodes in ls */
index d9ee1b9..c7c6cf9 100644 (file)
@@ -1624,65 +1624,170 @@ static int remove_from_waiters_ms(struct dlm_lkb *lkb, struct dlm_message *ms)
        return error;
 }
 
-/* FIXME: make this more efficient */
+/* If there's an rsb for the same resource being removed, ensure
+   that the remove message is sent before the new lookup message.
+   It should be rare to need a delay here, but if not, then it may
+   be worthwhile to add a proper wait mechanism rather than a delay. */
 
-static int shrink_bucket(struct dlm_ls *ls, int b)
+static void wait_pending_remove(struct dlm_rsb *r)
 {
-       struct rb_node *n;
+       struct dlm_ls *ls = r->res_ls;
+ restart:
+       spin_lock(&ls->ls_remove_spin);
+       if (ls->ls_remove_len &&
+           !rsb_cmp(r, ls->ls_remove_name, ls->ls_remove_len)) {
+               log_debug(ls, "delay lookup for remove dir %d %s",
+                         r->res_dir_nodeid, r->res_name);
+               spin_unlock(&ls->ls_remove_spin);
+               msleep(1);
+               goto restart;
+       }
+       spin_unlock(&ls->ls_remove_spin);
+}
+
+/*
+ * ls_remove_spin protects ls_remove_name and ls_remove_len which are
+ * read by other threads in wait_pending_remove.  ls_remove_names
+ * and ls_remove_lens are only used by the scan thread, so they do
+ * not need protection.
+ */
+
+static void shrink_bucket(struct dlm_ls *ls, int b)
+{
+       struct rb_node *n, *next;
        struct dlm_rsb *r;
+       char *name;
        int our_nodeid = dlm_our_nodeid();
-       int count = 0, found;
+       int remote_count = 0;
+       int i, len, rv;
 
-       for (;;) {
-               found = 0;
-               spin_lock(&ls->ls_rsbtbl[b].lock);
-               for (n = rb_first(&ls->ls_rsbtbl[b].toss); n; n = rb_next(n)) {
-                       r = rb_entry(n, struct dlm_rsb, res_hashnode);
+       memset(&ls->ls_remove_lens, 0, sizeof(int) * DLM_REMOVE_NAMES_MAX);
 
-                       /* If we're the directory record for this rsb, and
-                          we're not the master of it, then we need to wait
-                          for the master node to send us a dir remove for
-                          before removing the dir record. */
+       spin_lock(&ls->ls_rsbtbl[b].lock);
+       for (n = rb_first(&ls->ls_rsbtbl[b].toss); n; n = next) {
+               next = rb_next(n);
+               r = rb_entry(n, struct dlm_rsb, res_hashnode);
 
-                       if (!dlm_no_directory(ls) && !is_master(r) &&
-                           (dlm_dir_nodeid(r) == our_nodeid)) {
-                               continue;
-                       }
+               /* If we're the directory record for this rsb, and
+                  we're not the master of it, then we need to wait
+                  for the master node to send us a dir remove for
+                  before removing the dir record. */
 
-                       if (!time_after_eq(jiffies, r->res_toss_time +
-                                          dlm_config.ci_toss_secs * HZ))
-                               continue;
-                       found = 1;
-                       break;
+               if (!dlm_no_directory(ls) &&
+                   (r->res_master_nodeid != our_nodeid) &&
+                   (dlm_dir_nodeid(r) == our_nodeid)) {
+                       continue;
                }
 
-               if (!found) {
-                       spin_unlock(&ls->ls_rsbtbl[b].lock);
-                       break;
+               if (!time_after_eq(jiffies, r->res_toss_time +
+                                  dlm_config.ci_toss_secs * HZ)) {
+                       continue;
                }
 
-               if (kref_put(&r->res_ref, kill_rsb)) {
-                       rb_erase(&r->res_hashnode, &ls->ls_rsbtbl[b].toss);
-                       spin_unlock(&ls->ls_rsbtbl[b].lock);
+               if (!dlm_no_directory(ls) &&
+                   (r->res_master_nodeid == our_nodeid) &&
+                   (dlm_dir_nodeid(r) != our_nodeid)) {
 
                        /* We're the master of this rsb but we're not
                           the directory record, so we need to tell the
                           dir node to remove the dir record. */
 
-                       if (!dlm_no_directory(ls) && is_master(r) &&
-                           (dlm_dir_nodeid(r) != our_nodeid)) {
-                               send_remove(r);
-                       }
+                       ls->ls_remove_lens[remote_count] = r->res_length;
+                       memcpy(ls->ls_remove_names[remote_count], r->res_name,
+                              DLM_RESNAME_MAXLEN);
+                       remote_count++;
 
-                       dlm_free_rsb(r);
-                       count++;
-               } else {
-                       spin_unlock(&ls->ls_rsbtbl[b].lock);
+                       if (remote_count >= DLM_REMOVE_NAMES_MAX)
+                               break;
+                       continue;
+               }
+
+               if (!kref_put(&r->res_ref, kill_rsb)) {
                        log_error(ls, "tossed rsb in use %s", r->res_name);
+                       continue;
                }
+
+               rb_erase(&r->res_hashnode, &ls->ls_rsbtbl[b].toss);
+               dlm_free_rsb(r);
        }
+       spin_unlock(&ls->ls_rsbtbl[b].lock);
 
-       return count;
+       /*
+        * While searching for rsb's to free, we found some that require
+        * remote removal.  We leave them in place and find them again here
+        * so there is a very small gap between removing them from the toss
+        * list and sending the removal.  Keeping this gap small is
+        * important to keep us (the master node) from being out of sync
+        * with the remote dir node for very long.
+        *
+        * From the time the rsb is removed from toss until just after
+        * send_remove, the rsb name is saved in ls_remove_name.  A new
+        * lookup checks this to ensure that a new lookup message for the
+        * same resource name is not sent just before the remove message.
+        */
+
+       for (i = 0; i < remote_count; i++) {
+               name = ls->ls_remove_names[i];
+               len = ls->ls_remove_lens[i];
+
+               spin_lock(&ls->ls_rsbtbl[b].lock);
+               rv = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].toss, name, len, &r);
+               if (rv) {
+                       spin_unlock(&ls->ls_rsbtbl[b].lock);
+                       log_debug(ls, "remove_name not toss %s", name);
+                       continue;
+               }
+
+               if (r->res_master_nodeid != our_nodeid) {
+                       spin_unlock(&ls->ls_rsbtbl[b].lock);
+                       log_debug(ls, "remove_name master %d dir %d our %d %s",
+                                 r->res_master_nodeid, r->res_dir_nodeid,
+                                 our_nodeid, name);
+                       continue;
+               }
+
+               if (r->res_dir_nodeid == our_nodeid) {
+                       /* should never happen */
+                       spin_unlock(&ls->ls_rsbtbl[b].lock);
+                       log_error(ls, "remove_name dir %d master %d our %d %s",
+                                 r->res_dir_nodeid, r->res_master_nodeid,
+                                 our_nodeid, name);
+                       continue;
+               }
+
+               if (!time_after_eq(jiffies, r->res_toss_time +
+                                  dlm_config.ci_toss_secs * HZ)) {
+                       spin_unlock(&ls->ls_rsbtbl[b].lock);
+                       log_debug(ls, "remove_name toss_time %lu now %lu %s",
+                                 r->res_toss_time, jiffies, name);
+                       continue;
+               }
+
+               if (!kref_put(&r->res_ref, kill_rsb)) {
+                       spin_unlock(&ls->ls_rsbtbl[b].lock);
+                       log_error(ls, "remove_name in use %s", name);
+                       continue;
+               }
+
+               rb_erase(&r->res_hashnode, &ls->ls_rsbtbl[b].toss);
+
+               /* block lookup of same name until we've sent remove */
+               spin_lock(&ls->ls_remove_spin);
+               ls->ls_remove_len = len;
+               memcpy(ls->ls_remove_name, name, DLM_RESNAME_MAXLEN);
+               spin_unlock(&ls->ls_remove_spin);
+               spin_unlock(&ls->ls_rsbtbl[b].lock);
+
+               send_remove(r);
+
+               /* allow lookup of name again */
+               spin_lock(&ls->ls_remove_spin);
+               ls->ls_remove_len = 0;
+               memset(ls->ls_remove_name, 0, DLM_RESNAME_MAXLEN);
+               spin_unlock(&ls->ls_remove_spin);
+
+               dlm_free_rsb(r);
+       }
 }
 
 void dlm_scan_rsbs(struct dlm_ls *ls)
@@ -2608,6 +2713,8 @@ static int set_master(struct dlm_rsb *r, struct dlm_lkb *lkb)
                return 0;
        }
 
+       wait_pending_remove(r);
+
        r->res_first_lkid = lkb->lkb_id;
        send_lookup(r, lkb);
        return 1;
index d4d3b31..952557d 100644 (file)
@@ -506,6 +506,15 @@ static int new_lockspace(const char *name, const char *cluster,
                spin_lock_init(&ls->ls_rsbtbl[i].lock);
        }
 
+       spin_lock_init(&ls->ls_remove_spin);
+
+       for (i = 0; i < DLM_REMOVE_NAMES_MAX; i++) {
+               ls->ls_remove_names[i] = kzalloc(DLM_RESNAME_MAXLEN+1,
+                                                GFP_KERNEL);
+               if (!ls->ls_remove_names[i])
+                       goto out_rsbtbl;
+       }
+
        idr_init(&ls->ls_lkbidr);
        spin_lock_init(&ls->ls_lkbidr_spin);
 
@@ -556,7 +565,7 @@ static int new_lockspace(const char *name, const char *cluster,
 
        ls->ls_recover_buf = kmalloc(dlm_config.ci_buffer_size, GFP_NOFS);
        if (!ls->ls_recover_buf)
-               goto out_lkbfree;
+               goto out_lkbidr;
 
        ls->ls_slot = 0;
        ls->ls_num_slots = 0;
@@ -640,8 +649,13 @@ static int new_lockspace(const char *name, const char *cluster,
        spin_unlock(&lslist_lock);
        idr_destroy(&ls->ls_recover_idr);
        kfree(ls->ls_recover_buf);
- out_lkbfree:
+ out_lkbidr:
        idr_destroy(&ls->ls_lkbidr);
+       for (i = 0; i < DLM_REMOVE_NAMES_MAX; i++) {
+               if (ls->ls_remove_names[i])
+                       kfree(ls->ls_remove_names[i]);
+       }
+ out_rsbtbl:
        vfree(ls->ls_rsbtbl);
  out_lsfree:
        if (do_unreg)
@@ -796,6 +810,9 @@ static int release_lockspace(struct dlm_ls *ls, int force)
 
        vfree(ls->ls_rsbtbl);
 
+       for (i = 0; i < DLM_REMOVE_NAMES_MAX; i++)
+               kfree(ls->ls_remove_names[i]);
+
        while (!list_empty(&ls->ls_new_rsb)) {
                rsb = list_first_entry(&ls->ls_new_rsb, struct dlm_rsb,
                                       res_hashchain);