void dbpf_open_cache_initialize(void)
{
int i = 0, ret = 0;
gen_mutex_lock(&cache_mutex);
/* run through preallocated cache elements to initialize
* and put them on the free list
*/
if (OPEN_CACHE_SIZE == 0)
{
gossip_err("Warning: dbpf_open_cache disabled.\n");
}
for (i = 0; i < OPEN_CACHE_SIZE; i++)
{
prealloc[i].fd = -1;
qlist_add(&prealloc[i].queue_link, &free_list);
}
gen_mutex_unlock(&cache_mutex);
/* Initialize and create the worker thread for threaded deletes */
INIT_QLIST_HEAD(&dbpf_unlink_context.global_list);
pthread_mutex_init(&dbpf_unlink_context.mutex, NULL);
pthread_cond_init(&dbpf_unlink_context.data_available, NULL);
ret = pthread_create(&dbpf_unlink_context.thread_id, NULL, unlink_bstream, (void*)&dbpf_unlink_context);
if(ret)
{
gossip_err("dbpf_open_cache_initialize: failed [%d]\n", ret);
return;
}
}
void fcfs_next_rc(
void * sched,
void * rc_event_save,
model_net_sched_rc * rc,
tw_lp * lp){
mn_sched_queue *s = sched;
if (rc->rtn == -1){
// no op
}
else{
if (s->is_recv_queue){
dprintf("%lu (mn): rc receiving message\n", lp->gid);
s->method->model_net_method_recv_msg_event_rc(lp);
}
else {
dprintf("%lu (mn): rc issuing packet\n", lp->gid);
s->method->model_net_method_packet_event_rc(lp);
}
if (rc->rtn == 0){
// just get the front and increment rem
mn_sched_qitem *q = qlist_entry(s->reqs.next, mn_sched_qitem, ql);
// just increment rem
q->rem += q->req.packet_size;
}
else if (rc->rtn == 1){
// re-create the q item
mn_sched_qitem *q = malloc(sizeof(mn_sched_qitem));
assert(q);
q->req = rc->req;
q->sched_params = rc->sched_params;
q->rem = (q->req.is_pull ? PULL_MSG_SIZE : q->req.msg_size) %
q->req.packet_size;
if (q->rem == 0){ // processed exactly a packet's worth of data
q->rem = q->req.packet_size;
}
void * e_dat = rc_event_save;
if (q->req.remote_event_size > 0){
q->remote_event = malloc(q->req.remote_event_size);
memcpy(q->remote_event, e_dat, q->req.remote_event_size);
e_dat = (char*) e_dat + q->req.remote_event_size;
}
else { q->remote_event = NULL; }
if (q->req.self_event_size > 0) {
q->local_event = malloc(q->req.self_event_size);
memcpy(q->local_event, e_dat, q->req.self_event_size);
}
else { q->local_event = NULL; }
// add back to front of list
qlist_add(&q->ql, &s->reqs);
s->queue_len++;
}
else {
assert(0);
}
}
}
void rr_next_rc (
void * sched,
void * rc_event_save,
model_net_sched_rc * rc,
tw_lp * lp){
// only time we need to do something apart from fcfs is on a successful
// rr_next that didn't remove the item from the queue
if (rc->rtn == 0){
mn_sched_queue *s = sched;
qlist_add(qlist_pop_back(&s->reqs), &s->reqs);
}
fcfs_next_rc(sched, rc_event_save, rc, lp);
}
static void qlist_set(t_qlist *x, t_symbol *s, int ac, t_atom *av)
{
qlist_clear(x);
qlist_add(x, s, ac, av);
}
/**
* The dbpf open cache is used primarily to manage open
* file descriptors to bstream files on IO servers. PVFS
* currently uses a lazy style of creating the actual datafiles for
* bstreams. Only on the first write to a bstream is the file
* actually created (opened with O_CREAT). This means that if a
* read of a bstream that hasn't been written should somehow occur,
* an ENOENT error will be returned immediately, instead of allowing
* a read to EOF (of a zero-byte file). For us, this is ok, since
* the client gets the size of the bstream in the getattr before doing
* any IO. All that being said, the open_cache_get call needs to
* behave differently based on the desired operation: reads on
* files that don't exist should return ENOENT, but writes on files
* that don't exist should create and open the file.
*/
int dbpf_open_cache_get(
TROVE_coll_id coll_id,
TROVE_handle handle,
enum open_cache_open_type type,
struct open_cache_ref* out_ref)
{
struct qlist_head *tmp_link;
struct open_cache_entry* tmp_entry = NULL;
int found = 0;
int ret = 0;
gossip_debug(GOSSIP_DBPF_OPEN_CACHE_DEBUG,
"dbpf_open_cache_get: called\n");
gen_mutex_lock(&cache_mutex);
/* check already opened objects first, reuse ref if possible */
tmp_entry = dbpf_open_cache_find_entry(
&used_list, "used list", coll_id, handle);
if(!tmp_entry)
{
tmp_entry = dbpf_open_cache_find_entry(
&unused_list, "unused list", coll_id, handle);
}
out_ref->fd = -1;
if (tmp_entry)
{
if (tmp_entry->fd < 0)
{
ret = open_fd(&(tmp_entry->fd), coll_id, handle, type);
if (ret < 0)
{
gen_mutex_unlock(&cache_mutex);
return ret;
}
tmp_entry->type = type;
}
out_ref->fd = tmp_entry->fd;
out_ref->type = type;
out_ref->internal = tmp_entry;
tmp_entry->ref_ct++;
/* remove the entry and place it at the used head (assuming it
* will be referenced again soon)
*/
gossip_debug(GOSSIP_DBPF_OPEN_CACHE_DEBUG, "dbpf_open_cache_get: "
"moving to (or reordering in) used list.\n");
qlist_del(&tmp_entry->queue_link);
qlist_add(&tmp_entry->queue_link, &used_list);
gen_mutex_unlock(&cache_mutex);
assert(out_ref->fd > 0);
return 0;
}
/* if we fall through to this point, then the object was not found
* in the cache. In order of priority we will now try: free list,
* unused_list, and then bypass cache
*/
if (!qlist_empty(&free_list))
{
tmp_link = free_list.next;
tmp_entry = qlist_entry(tmp_link, struct open_cache_entry,
queue_link);
qlist_del(&tmp_entry->queue_link);
found = 1;
gossip_debug(GOSSIP_DBPF_OPEN_CACHE_DEBUG,
"dbpf_open_cache_get: resetting entry from free list.\n");
}