void
cl_cluster_scan_shutdown(cl_cluster* asc)
{
// Check whether we ever (lazily) initialized scan machinery.
if (cf_atomic32_get(asc->scan_initialized) == 0 && ! asc->scan_q) {
return;
}
// This tells the worker threads to stop. We do this (instead of using a
// "running" flag) to allow the workers to "wait forever" on processing the
// work dispatch queue, which has minimum impact when the queue is empty.
// This also means all queued requests get processed when shutting down.
for (int i = 0; i < NUM_SCAN_THREADS; i++) {
cl_scan_task task;
task.asc = NULL;
cf_queue_push(asc->scan_q, &task);
}
for (int i = 0; i < NUM_SCAN_THREADS; i++) {
pthread_join(asc->scan_threads[i], NULL);
}
cf_queue_destroy(asc->scan_q);
asc->scan_q = NULL;
cf_atomic32_set(&asc->scan_initialized, 0);
}
static int
do_scan_monte(cl_cluster *asc, char *node_name, uint operation_info, uint operation_info2, const char *ns, const char *set,
cl_bin *bins, int n_bins, uint8_t scan_pct,
citrusleaf_get_many_cb cb, void *udata, cl_scan_parameters *scan_opt)
{
int rv = -1;
uint8_t rd_stack_buf[STACK_BUF_SZ];
uint8_t *rd_buf = 0;
size_t rd_buf_sz = 0;
uint8_t wr_stack_buf[STACK_BUF_SZ];
uint8_t *wr_buf = wr_stack_buf;
size_t wr_buf_sz = sizeof(wr_stack_buf);
cl_scan_param_field scan_param_field;
if (scan_opt) {
scan_param_field.scan_pct = scan_pct>100? 100:scan_pct;
scan_param_field.byte1 = (scan_opt->priority<<4) | (scan_opt->fail_on_cluster_change<<3);
}
// we have a single namespace and/or set to get
if (cl_compile(operation_info, operation_info2, 0, ns, set, 0, 0, 0, 0, 0, 0, &wr_buf, &wr_buf_sz, 0, NULL, 0,
scan_opt ? &scan_param_field : NULL)) {
return(rv);
}
#ifdef DEBUG_VERBOSE
dump_buf("sending request to cluster:", wr_buf, wr_buf_sz);
#endif
int fd;
cl_cluster_node *node = 0;
// Get an FD from a cluster
if (node_name) {
node = cl_cluster_node_get_byname(asc,node_name);
// grab a reservation
if (node)
cl_cluster_node_reserve(node, "T+");
} else {
node = cl_cluster_node_get_random(asc);
}
if (!node) {
#ifdef DEBUG
cf_debug("warning: no healthy nodes in cluster, failing");
#endif
return(-1);
}
fd = cl_cluster_node_fd_get(node, false, asc->nbconnect);
if (fd == -1) {
#ifdef DEBUG
cf_debug("warning: node %s has no file descriptors, retrying transaction", node->name);
#endif
return(-1);
}
// send it to the cluster - non blocking socket, but we're blocking
if (0 != cf_socket_write_forever(fd, wr_buf, wr_buf_sz)) {
#ifdef DEBUG
cf_debug("Citrusleaf: write timeout or error when writing header to server - %d fd %d errno %d", rv, fd, errno);
#endif
close(fd);
return(-1);
}
cl_proto proto;
bool done = false;
do { // multiple CL proto per response
// Now turn around and read a fine cl_pro - that's the first 8 bytes that has types and lengths
if ((rv = cf_socket_read_forever(fd, (uint8_t *) &proto, sizeof(cl_proto) ) ) ) {
cf_error("network error: errno %d fd %d",rv, fd);
close(fd);
return(-1);
}
#ifdef DEBUG_VERBOSE
dump_buf("read proto header from cluster", (uint8_t *) &proto, sizeof(cl_proto));
#endif
cl_proto_swap(&proto);
if (proto.version != CL_PROTO_VERSION) {
cf_error("network error: received protocol message of wrong version %d", proto.version);
close(fd);
return(-1);
}
if (proto.type != CL_PROTO_TYPE_CL_MSG) {
cf_error("network error: received incorrect message version %d", proto.type);
close(fd);
return(-1);
}
// second read for the remainder of the message - expect this to cover lots of data, many lines
//
// if there's no error
rd_buf_sz = proto.sz;
if (rd_buf_sz > 0) {
// cf_debug("message read: size %u",(uint)proto.sz);
if (rd_buf_sz > sizeof(rd_stack_buf))
rd_buf = malloc(rd_buf_sz);
else
rd_buf = rd_stack_buf;
if (rd_buf == NULL) {
close(fd);
return (-1);
}
if ((rv = cf_socket_read_forever(fd, rd_buf, rd_buf_sz))) {
cf_error("network error: errno %d fd %d", rv, fd);
if (rd_buf != rd_stack_buf) { free(rd_buf); }
close(fd);
return(-1);
}
// this one's a little much: printing the entire body before printing the other bits
#ifdef DEBUG_VERBOSE
dump_buf("read msg body header (multiple msgs)", rd_buf, rd_buf_sz);
#endif
}
// process all the cl_msg in this proto
uint8_t *buf = rd_buf;
uint pos = 0;
cl_bin stack_bins[STACK_BINS];
cl_bin *bins_local;
while (pos < rd_buf_sz) {
#ifdef DEBUG_VERBOSE
dump_buf("individual message header", buf, sizeof(cl_msg));
#endif
uint8_t *buf_start = buf;
cl_msg *msg = (cl_msg *) buf;
cl_msg_swap_header(msg);
buf += sizeof(cl_msg);
if (msg->header_sz != sizeof(cl_msg)) {
cf_error("received cl msg of unexpected size: expecting %zd found %d, internal error",
sizeof(cl_msg),msg->header_sz);
close(fd);
return(-1);
}
// parse through the fields
cf_digest *keyd = 0;
char ns_ret[33] = {0};
char *set_ret = NULL;
cl_msg_field *mf = (cl_msg_field *)buf;
for (int i=0;i<msg->n_fields;i++) {
cl_msg_swap_field(mf);
if (mf->type == CL_MSG_FIELD_TYPE_KEY) {
cf_error("read: found a key - unexpected");
}
else if (mf->type == CL_MSG_FIELD_TYPE_DIGEST_RIPE) {
keyd = (cf_digest *) mf->data;
}
else if (mf->type == CL_MSG_FIELD_TYPE_NAMESPACE) {
memcpy(ns_ret, mf->data, cl_msg_field_get_value_sz(mf));
ns_ret[ cl_msg_field_get_value_sz(mf) ] = 0;
}
else if (mf->type == CL_MSG_FIELD_TYPE_SET) {
uint32_t set_name_len = cl_msg_field_get_value_sz(mf);
set_ret = (char *)malloc(set_name_len + 1);
memcpy(set_ret, mf->data, set_name_len);
set_ret[ set_name_len ] = '\0';
}
mf = cl_msg_field_get_next(mf);
}
buf = (uint8_t *) mf;
#ifdef DEBUG_VERBOSE
cf_debug("message header fields: nfields %u nops %u", msg->n_fields, msg->n_ops);
#endif
if (msg->n_ops > STACK_BINS) {
bins_local = malloc(sizeof(cl_bin) * msg->n_ops);
}
else {
bins_local = stack_bins;
}
if (bins_local == NULL) {
if (set_ret) {
free(set_ret);
}
close(fd);
return (-1);
}
// parse through the bins/ops
cl_msg_op *op = (cl_msg_op *)buf;
for (int i=0;i<msg->n_ops;i++) {
cl_msg_swap_op(op);
#ifdef DEBUG_VERBOSE
cf_debug("op receive: %p size %d op %d ptype %d pversion %d namesz %d",
op,op->op_sz, op->op, op->particle_type, op->version, op->name_sz);
#endif
#ifdef DEBUG_VERBOSE
dump_buf("individual op (host order)", (uint8_t *) op, op->op_sz + sizeof(uint32_t));
#endif
cl_set_value_particular(op, &bins_local[i]);
op = cl_msg_op_get_next(op);
}
buf = (uint8_t *) op;
if (msg->result_code != CL_RESULT_OK) {
// Special case - if we scan a set name that doesn't exist on a
// node, it will return "not found" - we unify this with the
// case where OK is returned and no callbacks were made. [AKG]
if (msg->result_code == CL_RESULT_NOTFOUND) {
msg->result_code = CL_RESULT_OK;
}
rv = (int)msg->result_code;
done = true;
}
else if (msg->info3 & CL_MSG_INFO3_LAST) {
#ifdef DEBUG
cf_debug("received final message");
#endif
done = true;
}
else if ((msg->n_ops) || (operation_info & CL_MSG_INFO1_NOBINDATA)) {
// got one good value? call it a success!
(*cb) ( ns_ret, keyd, set_ret, msg->generation, msg->record_ttl, bins_local, msg->n_ops, false /*islast*/, udata);
rv = 0;
}
// else
// cf_debug("received message with no bins, signal of an error");
if (bins_local != stack_bins) {
free(bins_local);
bins_local = 0;
}
if (set_ret) {
free(set_ret);
set_ret = NULL;
}
// don't have to free object internals. They point into the read buffer, where
// a pointer is required
pos += buf - buf_start;
}
if (rd_buf && (rd_buf != rd_stack_buf)) {
free(rd_buf);
rd_buf = 0;
}
} while ( done == false );
if (wr_buf != wr_stack_buf) {
free(wr_buf);
wr_buf = 0;
}
cf_atomic32_set(&node->intervals_unreachable, 0);
cl_cluster_node_fd_put(node, fd, false);
cl_cluster_node_put(node);
node = 0;
#ifdef DEBUG_VERBOSE
cf_debug("exited loop: rv %d", rv );
#endif
return(rv);
}
//Same as do_the_full_monte, but only till the command is sent to the node.
//Most of the code is duplicated. Bad.
int
cl_do_async_monte(cl_cluster *asc, int info1, int info2, const char *ns, const char *set, const cl_object *key,
const cf_digest *digest, cl_bin **values, cl_operator operator, cl_operation **operations,
int *n_values, uint32_t *cl_gen, const cl_write_parameters *cl_w_p, uint64_t *trid, void *udata)
{
cl_async_work *workitem = NULL;
uint8_t wr_stack_buf[STACK_BUF_SZ];
uint8_t *wr_buf = wr_stack_buf;
size_t wr_buf_sz = sizeof(wr_stack_buf);
int progress_timeout_ms;
uint64_t deadline_ms;
uint64_t starttime, endtime;
bool network_error;
int fd = -1;
int rv = CITRUSLEAF_FAIL_CLIENT; //Assume that this is a failure;
// as_msg msg;
cf_digest d_ret;
cl_cluster_node *node = 0;
#if ONEASYNCFD
if (shash_get_size(g_cl_async_hashtab) >= g_async_h_szlimit) {
//cf_error("Async hashtab is full. Cannot insert any more elements");
return CITRUSLEAF_FAIL_ASYNCQ_FULL;
}
#else
//If the async buffer is at the max limit, do not entertain more requests.
if (cf_queue_sz(g_cl_async_q) >= cf_atomic32_get(g_async_q_szlimit)) {
//cf_error("Async buffer is full. Cannot insert any more elements");
return CITRUSLEAF_FAIL_ASYNCQ_FULL;
}
#endif
//Allocate memory for work item that will be added to the async work list
if (cf_queue_sz(g_cl_workitems_freepool_q) > 0) {
cf_queue_pop(g_cl_workitems_freepool_q, &workitem, CF_QUEUE_FOREVER);
} else {
workitem = malloc(sizeof(cl_async_work));
if (workitem == NULL) {
return CITRUSLEAF_FAIL_CLIENT;
}
}
//Compile the write buffer to be sent to the cluster
if (n_values && ( values || operations) ){
cl_compile(info1, info2, 0, ns, set, key, digest, values?*values:NULL, operator, operations?*operations:NULL,
*n_values , &wr_buf, &wr_buf_sz, cl_w_p, &d_ret, *trid,NULL,NULL, 0 /*udf_type*/);
}else{
cl_compile(info1, info2, 0, ns, set, key, digest, 0, 0, 0, 0, &wr_buf, &wr_buf_sz, cl_w_p, &d_ret, *trid,NULL,NULL, 0 /*udf_type*/);
}
deadline_ms = 0;
progress_timeout_ms = 0;
if (cl_w_p && cl_w_p->timeout_ms) {
deadline_ms = cf_getms() + cl_w_p->timeout_ms;
// policy: if asking for a long timeout, give enough time to try twice
if (cl_w_p->timeout_ms > 700) {
progress_timeout_ms = cl_w_p->timeout_ms / 2;
}
else {
progress_timeout_ms = cl_w_p->timeout_ms;
}
}
else {
progress_timeout_ms = g_async_nw_progress_timeout;
}
//Initialize the async work unit
workitem->trid = *trid;
workitem->deadline = deadline_ms;
workitem->starttime = cf_getms();
workitem->udata = udata;
as_msg *msgp;
// Hate special cases, but we have to clear the verify bit on delete verify
if ( (info2 & CL_MSG_INFO2_DELETE) && (info1 & CL_MSG_INFO1_VERIFY))
{
msgp = (as_msg *)wr_buf;
msgp->m.info1 &= ~CL_MSG_INFO1_VERIFY;
}
if (asc->compression_stat.compression_threshold > 0
&& wr_buf_sz > (size_t)asc->compression_stat.compression_threshold)
{
/* Compression is enabled.
* Packet size is above threshold.
* Compress the data
*/
uint8_t *compressed_buf = NULL;
size_t compressed_buf_sz = 0;
// Contstruct packet for compressed data.
cf_packet_compression (wr_buf, wr_buf_sz, &compressed_buf, &compressed_buf_sz);
if (compressed_buf)
{
// If original packet size is > 16k, cl_compile had allocated memory for it.
// Free that memory.
// cf_packet_compression will allocate memory for compressed packet
if (wr_buf != wr_stack_buf) {
free(wr_buf);
}
// Update stats.
citrusleaf_cluster_put_compression_stat(asc, wr_buf_sz, compressed_buf_sz);
wr_buf = compressed_buf;
wr_buf_sz = compressed_buf_sz;
//memcpy (wr_buf, compressed_buf, compressed_buf_sz);
//wr_buf_sz = compressed_buf_sz;
//free (compressed_buf);
}
//else compression failed, continue with uncompressed packet
else
{
// Set compression stat
citrusleaf_cluster_put_compression_stat(asc, wr_buf_sz, wr_buf_sz);
}
}
int try = 0;
// retry request based on the write_policy
do {
network_error = false;
try++;
#ifdef DEBUG
if (try > 1) {
cf_debug("request retrying try %d tid %zu", try, (uint64_t)pthread_self());
}
#endif
// Get an FD from a cluster. First get the probable node for the given digest.
node = cl_cluster_node_get(asc, ns, &d_ret, info2 & CL_MSG_INFO2_WRITE ? true : false);
if (!node) {
#ifdef DEBUG
cf_debug("warning: no healthy nodes in cluster, retrying");
#endif
usleep(10000); //Sleep for 10ms
goto Retry;
}
// Now get the dedicated async FD of this node
starttime = cf_getms();
fd = cl_cluster_node_fd_get(node, true);
endtime = cf_getms();
if ((endtime - starttime) > 10) {
cf_debug("Time to get FD for a node (>10ms)=%"PRIu64, (endtime - starttime));
}
if (fd == -1) {
#ifdef DEBUG
cf_debug("warning: node %s has no async file descriptors, retrying transaction (tid %zu)",node->name,(uint64_t)pthread_self() );
#endif
usleep(1000);
goto Retry;
}
// Send the command to the node
starttime = cf_getms();
rv = cf_socket_write_timeout(fd, wr_buf, wr_buf_sz, deadline_ms, progress_timeout_ms);
endtime = cf_getms();
if ((endtime - starttime) > 10) {
cf_debug("Time to write to the socket (>10ms)=%"PRIu64, (endtime - starttime));
}
if (rv != 0) {
cf_debug("Citrusleaf: write timeout or error when writing header to server - %d fd %d errno %d (tid %zu)",
rv,fd,errno,(uint64_t)pthread_self());
if (rv != ETIMEDOUT)
network_error = true;
goto Retry;
}
goto Ok;
Retry:
if (network_error == true) {
/*
* In case of Async work (for XDS), it may be extreme to
* dun a node in case of network error. We just cleanup
* things and retry to connect to the remote cluster.
* The network error may be a transient one. As this is a
* network error, its is better to wait for some significant
* time before retrying.
*/
sleep(1); //Sleep for 1sec
#if ONEASYNCFD
//Do not close the FD
#else
cf_error("async sender: Closing the fd %d because of network error", fd);
cf_close(fd);
fd = -1;
#endif
}
if (fd != -1) {
cf_error("async sender: Closing the fd %d because of retry", fd);
cf_close(fd);
fd = -1;
}
if (node) {
cl_cluster_node_put(node);
node = 0;
}
if (deadline_ms && (deadline_ms < cf_getms() ) ) {
#ifdef DEBUG
cf_debug("async sender: out of time : deadline %"PRIu64" now %"PRIu64, deadline_ms, cf_getms());
#endif
rv = CITRUSLEAF_FAIL_TIMEOUT;
goto Error;
}
} while ( (cl_w_p == 0) || (cl_w_p->w_pol == CL_WRITE_RETRY) );
Error:
#ifdef DEBUG
cf_debug("exiting with failure: network_error %d wpol %d timeleft %d rv %d",
(int)network_error, (int)(cl_w_p ? cl_w_p->w_pol : 0),
(int)(deadline_ms - cf_getms() ), rv );
#endif
if (wr_buf != wr_stack_buf) {
free(wr_buf);
}
#if ONEASYNCFD
//Do not close the FD
#else
//If it is a network error, the fd would be closed and set to -1.
//So, we reach this place with a valid FD in case of timeout.
if (fd != -1) {
cf_error("async sender: Closing the fd %d because of timeout", fd);
cf_close(fd);
}
#endif
return(rv);
Ok:
/*
* We cannot release the node here as the asyc FD associated
* with this node may get closed. We should do it only when
* we got back the ack for the async command that we just did.
*/
//As we sent the command successfully, add it to the async work list
workitem->node = node;
workitem->fd = fd;
//We are storing only the pointer to the workitem
#if ONEASYNCFD
if (shash_put_unique(g_cl_async_hashtab, trid, &workitem) != SHASH_OK) {
//This should always succeed.
cf_error("Unable to add unique entry into the hash table");
}
cf_queue_push(node->asyncwork_q, &workitem); //Also put in the node's q
#else
cf_queue_push(g_cl_async_q, &workitem);
#endif
if (wr_buf != wr_stack_buf) {
free(wr_buf);
}
rv = CITRUSLEAF_OK;
return rv;
}
int citrusleaf_async_reinit(int size_limit, unsigned int num_receiver_threads)
{
// int num_threads;
if (0 == cf_atomic32_get(g_async_initialized)) {
cf_error("Async client not initialized cannot reinit");
return -1;
}
if (num_receiver_threads > MAX_ASYNC_RECEIVER_THREADS) {
//Limit the threads to the max value even if caller asks for it
num_receiver_threads = MAX_ASYNC_RECEIVER_THREADS;
}
// If number of thread is increased create more threads
if (num_receiver_threads > g_async_num_threads) {
unsigned int i;
for (i = g_async_num_threads; i < num_receiver_threads; i++) {
pthread_create(&g_async_reciever[i], 0, async_receiver_fn, NULL);
}
}
else {
// else just reset the number the async threads will kill themselves
cf_atomic32_set(&g_async_num_threads, num_receiver_threads);
}
cf_atomic32_set(&g_async_q_szlimit , size_limit);
return ( 0 );
}
int citrusleaf_async_init(int size_limit, int num_receiver_threads, cl_async_fail_cb fail_cb_fn, cl_async_success_cb success_cb_fn)
{
int i, num_threads;
//Make sure that we do the initialization only once
if (1 == cf_atomic32_incr(&g_async_initialized)) {
// Start the receiver threads
num_threads = num_receiver_threads;
if (num_threads > MAX_ASYNC_RECEIVER_THREADS) {
//Limit the threads to the max value even if caller asks for it
num_threads = MAX_ASYNC_RECEIVER_THREADS;
}
#if ONEASYNCFD
g_async_h_szlimit = size_limit * 3; //Max number of elements in the hash table
g_async_h_buckets = g_async_h_szlimit/10;//Number of buckets in the hash table
if (shash_create(&g_cl_async_hashtab, async_trid_hash, sizeof(uint64_t), sizeof(cl_async_work *),
g_async_h_buckets, SHASH_CR_MT_BIGLOCK) != SHASH_OK) {
cf_error("Failed to initialize the async work hastable");
cf_atomic32_decr(&g_async_initialized);
return -1;
}
#else
// create work queue
g_async_q_szlimit = size_limit;
if ((g_cl_async_q = cf_queue_create(sizeof(cl_async_work *), true)) == NULL) {
cf_error("Failed to initialize the async work queue");
cf_atomic32_decr(&g_async_initialized);
return -1;
}
for (i=0; i<num_threads; i++) {
pthread_create(&g_async_reciever[i], 0, async_receiver_fn, NULL);
}
g_async_num_threads = num_threads;
#endif
if ((g_cl_workitems_freepool_q = cf_queue_create(sizeof(cl_async_work *), true)) == NULL) {
cf_error("Failed to create memory pool for workitems");
return -1;
}
g_fail_cb_fn = fail_cb_fn;
g_success_cb_fn = success_cb_fn;
// Initialize the stats
g_async_stats.retries = 0;
g_async_stats.dropouts = 0;
}
return(0);
}
