* multiple hb threads are watching multiple regions. A node is live
* whenever any of the threads sees activity from the node in its region.
*/
-static spinlock_t o2hb_live_lock = SPIN_LOCK_UNLOCKED;
+static DEFINE_SPINLOCK(o2hb_live_lock);
static struct list_head o2hb_live_slots[O2NM_MAX_NODES];
static unsigned long o2hb_live_node_bitmap[BITS_TO_LONGS(O2NM_MAX_NODES)];
static LIST_HEAD(o2hb_node_events);
* recognizes a node going up and down in one iteration */
u64 hr_generation;
- struct work_struct hr_write_timeout_work;
+ struct delayed_work hr_write_timeout_work;
unsigned long hr_last_timeout_start;
/* Used during o2hb_check_slot to hold a copy of the block
struct o2hb_bio_wait_ctxt {
atomic_t wc_num_reqs;
struct completion wc_io_complete;
+ int wc_error;
};
-static void o2hb_write_timeout(void *arg)
+static void o2hb_write_timeout(struct work_struct *work)
{
- struct o2hb_region *reg = arg;
+ struct o2hb_region *reg =
+ container_of(work, struct o2hb_region,
+ hr_write_timeout_work.work);
mlog(ML_ERROR, "Heartbeat write timeout to device %s after %u "
"milliseconds\n", reg->hr_dev_name,
{
atomic_set(&wc->wc_num_reqs, num_ios);
init_completion(&wc->wc_io_complete);
+ wc->wc_error = 0;
}
/* Used in error paths too */
{
struct o2hb_bio_wait_ctxt *wc = bio->bi_private;
- if (error)
+ if (error) {
mlog(ML_ERROR, "IO Error %d\n", error);
+ wc->wc_error = error;
+ }
if (bio->bi_size)
return 1;
max_pages = q->max_hw_segments;
max_pages--; /* Handle I/Os that straddle a page */
- max_sectors = max_pages << (PAGE_SHIFT - 9);
-
+ if (max_pages) {
+ max_sectors = max_pages << (PAGE_SHIFT - 9);
+ } else {
+ /* If BIO contains 1 or less than 1 page. */
+ max_sectors = q->max_sectors;
+ }
/* Why is fls() 1-based???? */
pow_two_sectors = 1 << (fls(max_sectors) - 1);
bail_and_wait:
o2hb_wait_on_io(reg, &wc);
+ if (wc.wc_error && !status)
+ status = wc.wc_error;
if (bios) {
for(i = 0; i < num_bios; i++)
hb_block->hb_seq = cpu_to_le64(cputime);
hb_block->hb_node = node_num;
hb_block->hb_generation = cpu_to_le64(generation);
+ hb_block->hb_dead_ms = cpu_to_le32(o2hb_dead_threshold * O2HB_REGION_TIMEOUT_MS);
/* This step must always happen last! */
hb_block->hb_cksum = cpu_to_le32(o2hb_compute_block_crc_le(reg,
struct o2nm_node *node;
struct o2hb_disk_heartbeat_block *hb_block = reg->hr_tmp_block;
u64 cputime;
+ unsigned int dead_ms = o2hb_dead_threshold * O2HB_REGION_TIMEOUT_MS;
+ unsigned int slot_dead_ms;
memcpy(hb_block, slot->ds_raw_block, reg->hr_block_bytes);
&o2hb_live_slots[slot->ds_node_num]);
slot->ds_equal_samples = 0;
+
+ /* We want to be sure that all nodes agree on the
+ * number of milliseconds before a node will be
+ * considered dead. The self-fencing timeout is
+ * computed from this value, and a discrepancy might
+ * result in heartbeat calling a node dead when it
+ * hasn't self-fenced yet. */
+ slot_dead_ms = le32_to_cpu(hb_block->hb_dead_ms);
+ if (slot_dead_ms && slot_dead_ms != dead_ms) {
+ /* TODO: Perhaps we can fail the region here. */
+ mlog(ML_ERROR, "Node %d on device %s has a dead count "
+ "of %u ms, but our count is %u ms.\n"
+ "Please double check your configuration values "
+ "for 'O2CB_HEARTBEAT_THRESHOLD'\n",
+ slot->ds_node_num, reg->hr_dev_name, slot_dead_ms,
+ dead_ms);
+ }
goto out;
}
return highest;
}
-static void o2hb_do_disk_heartbeat(struct o2hb_region *reg)
+static int o2hb_do_disk_heartbeat(struct o2hb_region *reg)
{
int i, ret, highest_node, change = 0;
unsigned long configured_nodes[BITS_TO_LONGS(O2NM_MAX_NODES)];
struct bio *write_bio;
struct o2hb_bio_wait_ctxt write_wc;
- if (o2nm_configured_node_map(configured_nodes, sizeof(configured_nodes)))
- return;
+ ret = o2nm_configured_node_map(configured_nodes,
+ sizeof(configured_nodes));
+ if (ret) {
+ mlog_errno(ret);
+ return ret;
+ }
highest_node = o2hb_highest_node(configured_nodes, O2NM_MAX_NODES);
if (highest_node >= O2NM_MAX_NODES) {
mlog(ML_NOTICE, "ocfs2_heartbeat: no configured nodes found!\n");
- return;
+ return -EINVAL;
}
/* No sense in reading the slots of nodes that don't exist
ret = o2hb_read_slots(reg, highest_node + 1);
if (ret < 0) {
mlog_errno(ret);
- return;
+ return ret;
}
/* With an up to date view of the slots, we can check that no
ret = o2hb_issue_node_write(reg, &write_bio, &write_wc);
if (ret < 0) {
mlog_errno(ret);
- return;
+ return ret;
}
i = -1;
*/
o2hb_wait_on_io(reg, &write_wc);
bio_put(write_bio);
+ if (write_wc.wc_error) {
+ /* Do not re-arm the write timeout on I/O error - we
+ * can't be sure that the new block ever made it to
+ * disk */
+ mlog(ML_ERROR, "Write error %d on device \"%s\"\n",
+ write_wc.wc_error, reg->hr_dev_name);
+ return write_wc.wc_error;
+ }
+
o2hb_arm_write_timeout(reg);
/* let the person who launched us know when things are steady */
if (atomic_dec_and_test(®->hr_steady_iterations))
wake_up(&o2hb_steady_queue);
}
+
+ return 0;
}
/* Subtract b from a, storing the result in a. a *must* have a larger
* likely to time itself out. */
do_gettimeofday(&before_hb);
- o2hb_do_disk_heartbeat(reg);
+ i = 0;
+ do {
+ ret = o2hb_do_disk_heartbeat(reg);
+ } while (ret && ++i < 2);
do_gettimeofday(&after_hb);
elapsed_msec = o2hb_elapsed_msecs(&before_hb, &after_hb);
goto out;
}
- INIT_WORK(®->hr_write_timeout_work, o2hb_write_timeout, reg);
+ INIT_DELAYED_WORK(®->hr_write_timeout_work, o2hb_write_timeout);
/*
* A node is considered live after it has beat LIVE_THRESHOLD