/** Create a new memcached handle
*
* @param ctx to allocate handle in.
* @param instance data.
*/
static void *mod_conn_create(TALLOC_CTX *ctx, void *instance)
{
rlm_cache_t *inst = instance;
rlm_cache_memcached_t *driver = inst->driver;
rlm_cache_memcached_handle_t *mandle;
memcached_st *sandle;
memcached_return_t ret;
sandle = memcached(driver->options, talloc_array_length(driver->options) -1);
if (!sandle) {
ERROR("rlm_cache_memcached: Failed creating memcached connection");
return NULL;
}
ret = memcached_version(sandle);
if (ret != MEMCACHED_SUCCESS) {
ERROR("rlm_cache_memcached: Failed getting server info: %s: %s", memcached_strerror(sandle, ret),
memcached_last_error_message(sandle));
memcached_free(sandle);
return NULL;
}
mandle = talloc_zero(ctx, rlm_cache_memcached_handle_t);
mandle->handle = sandle;
talloc_set_destructor(mandle, _mod_conn_free);
return mandle;
}
/** Insert a new entry into the data store
*
* @param inst main rlm_cache instance.
* @param request The current request.
* @param handle Pointer to memcached handle.
* @param c entry to insert.
* @return CACHE_OK on success else CACHE_ERROR on error.
*/
static cache_status_t cache_entry_insert(UNUSED rlm_cache_t *inst, REQUEST *request, rlm_cache_handle_t **handle,
rlm_cache_entry_t *c)
{
rlm_cache_memcached_handle_t *mandle = *handle;
memcached_return_t ret;
TALLOC_CTX *pool;
char *to_store;
pool = talloc_pool(NULL, 1024);
if (!pool) return CACHE_ERROR;
if (cache_serialize(pool, &to_store, c) < 0) {
talloc_free(pool);
return CACHE_ERROR;
}
ret = memcached_set(mandle->handle, c->key, talloc_array_length(c->key) - 1,
to_store ? to_store : "",
to_store ? talloc_array_length(to_store) - 1 : 0, c->expires, 0);
talloc_free(pool);
if (ret != MEMCACHED_SUCCESS) {
RERROR("Failed storing entry with key \"%s\": %s: %s", c->key,
memcached_strerror(mandle->handle, ret),
memcached_last_error_message(mandle->handle));
return CACHE_ERROR;
}
return CACHE_OK;
}
const std::string Memcached::GetErrMsg(void)
{
std::string str = m_sErrMsg;
if (m_pMemc != NULL)
{
const char* pMsg = memcached_last_error_message(m_pMemc);
str = str + "\tErrMsg:" + pMsg;
}
return str;
}
/** Call delete the cache entry from memcached
*
* @param inst main rlm_cache instance.
* @param request The current request.
* @param handle Pointer to memcached handle.
* @param c entry to expire.
* @return CACHE_OK on success else CACHE_ERROR.
*/
static cache_status_t cache_entry_expire(UNUSED rlm_cache_t *inst, REQUEST *request, rlm_cache_handle_t **handle,
rlm_cache_entry_t *c)
{
rlm_cache_memcached_handle_t *mandle = *handle;
memcached_return_t ret;
ret = memcached_delete(mandle->handle, c->key, talloc_array_length(c->key) - 1, 0);
if (ret != MEMCACHED_SUCCESS) {
RERROR("Failed deleting entry with key \"%s\": %s", c->key,
memcached_last_error_message(mandle->handle));
return CACHE_ERROR;
}
return CACHE_OK;
}
/** Locate a cache entry in memcached
*
* @param out Where to write the pointer to the cach entry.
* @param inst main rlm_cache instance.
* @param request The current request.
* @param handle Pointer to memcached handle.
* @param key to search for.
* @return CACHE_OK on success CACHE_MISS if no entry found, CACHE_ERROR on error.
*/
static cache_status_t cache_entry_find(rlm_cache_entry_t **out, UNUSED rlm_cache_t *inst, REQUEST *request,
rlm_cache_handle_t **handle, char const *key)
{
rlm_cache_memcached_handle_t *mandle = *handle;
memcached_return_t mret;
size_t len;
int ret;
uint32_t flags;
char *from_store;
rlm_cache_entry_t *c;
from_store = memcached_get(mandle->handle, key, strlen(key), &len, &flags, &mret);
if (!from_store) {
if (mret == MEMCACHED_NOTFOUND) return CACHE_MISS;
RERROR("Failed retrieving entry for key \"%s\": %s: %s", key, memcached_strerror(mandle->handle, mret),
memcached_last_error_message(mandle->handle));
return CACHE_ERROR;
}
RDEBUG2("Retrieved %zu bytes from memcached", len);
RDEBUG2("%s", from_store);
c = talloc_zero(NULL, rlm_cache_entry_t);
ret = cache_deserialize(c, from_store, len);
free(from_store);
if (ret < 0) {
RERROR("%s", fr_strerror());
talloc_free(c);
return CACHE_ERROR;
}
c->key = talloc_strdup(c, key);
*out = c;
return CACHE_OK;
}
/// Update the data for the specified namespace and key. Writes the data
/// atomically, so if the underlying data has changed since it was last
/// read, the update is rejected and this returns Store::Status::CONTENTION.
Store::Status MemcachedStore::set_data(const std::string& table,
const std::string& key,
const std::string& data,
uint64_t cas,
int expiry,
SAS::TrailId trail)
{
Store::Status status = Store::Status::OK;
LOG_DEBUG("Writing %d bytes to table %s key %s, CAS = %ld, expiry = %d",
data.length(), table.c_str(), key.c_str(), cas, expiry);
// Construct the fully qualified key.
std::string fqkey = table + "\\\\" + key;
const char* key_ptr = fqkey.data();
const size_t key_len = fqkey.length();
const std::vector<memcached_st*>& replicas = get_replicas(fqkey, Op::WRITE);
if (trail != 0)
{
SAS::Event start(trail, SASEvent::MEMCACHED_SET_START, 0);
start.add_var_param(fqkey);
start.add_var_param(data);
start.add_static_param(cas);
start.add_static_param(expiry);
SAS::report_event(start);
}
LOG_DEBUG("%d write replicas for key %s", replicas.size(), fqkey.c_str());
// Calculate the rough expected expiry time. We store this in the flags
// as it may be useful in future for read repair function.
uint32_t now = time(NULL);
uint32_t exptime = now + expiry;
// Memcached uses a flexible mechanism for specifying expiration.
// - 0 indicates never expire.
// - <= MEMCACHED_EXPIRATION_MAXDELTA indicates a relative (delta) time.
// - > MEMCACHED_EXPIRATION_MAXDELTA indicates an absolute time.
// Absolute time is the only way to force immediate expiry. Unfortunately,
// it's not reliable - see https://github.com/Metaswitch/cpp-common/issues/160
// for details. Instead, we use relative time for future times (expiry > 0)
// and the earliest absolute time for immediate expiry (expiry == 0).
time_t memcached_expiration =
(time_t)((expiry > 0) ? expiry : MEMCACHED_EXPIRATION_MAXDELTA + 1);
// First try to write the primary data record to the first responding
// server.
memcached_return_t rc = MEMCACHED_ERROR;
size_t ii;
size_t replica_idx;
// If we only have one replica, we should try it twice -
// libmemcached won't notice a dropped TCP connection until it tries
// to make a request on it, and will fail the request then
// reconnect, so the second attempt could still work.
size_t attempts = (replicas.size() == 1) ? 2: replicas.size();
for (ii = 0; ii < attempts; ++ii)
{
if ((replicas.size() == 1) && (ii == 1)) {
if (rc != MEMCACHED_CONNECTION_FAILURE)
{
// This is a legitimate error, not a transient server failure, so we
// shouldn't retry.
break;
}
replica_idx = 0;
LOG_WARNING("Failed to write to sole memcached replica: retrying once");
}
else
{
replica_idx = ii;
}
LOG_DEBUG("Attempt conditional write to replica %d (connection %p), CAS = %ld",
replica_idx,
replicas[replica_idx],
cas);
if (cas == 0)
{
// New record, so attempt to add. This will fail if someone else
// gets there first.
rc = memcached_add(replicas[replica_idx],
key_ptr,
key_len,
data.data(),
data.length(),
memcached_expiration,
exptime);
}
else
{
// This is an update to an existing record, so use memcached_cas
// to make sure it is atomic.
rc = memcached_cas(replicas[replica_idx],
key_ptr,
key_len,
data.data(),
data.length(),
memcached_expiration,
exptime,
cas);
}
if (memcached_success(rc))
{
LOG_DEBUG("Conditional write succeeded to replica %d", replica_idx);
break;
}
else
{
LOG_DEBUG("memcached_%s command for %s failed on replica %d, rc = %d (%s), expiry = %d\n%s",
(cas == 0) ? "add" : "cas",
fqkey.c_str(),
replica_idx,
rc,
memcached_strerror(replicas[replica_idx], rc),
expiry,
memcached_last_error_message(replicas[replica_idx]));
if ((rc == MEMCACHED_NOTSTORED) ||
(rc == MEMCACHED_DATA_EXISTS))
{
if (trail != 0)
{
SAS::Event err(trail, SASEvent::MEMCACHED_SET_CONTENTION, 0);
err.add_var_param(fqkey);
SAS::report_event(err);
}
// A NOT_STORED or EXISTS response indicates a concurrent write failure,
// so return this to the application immediately - don't go on to
// other replicas.
LOG_INFO("Contention writing data for %s to store", fqkey.c_str());
status = Store::Status::DATA_CONTENTION;
break;
}
}
}
if ((rc == MEMCACHED_SUCCESS) &&
(replica_idx < replicas.size()))
{
// Write has succeeded, so write unconditionally (and asynchronously)
// to the replicas.
for (size_t jj = replica_idx + 1; jj < replicas.size(); ++jj)
{
LOG_DEBUG("Attempt unconditional write to replica %d", jj);
memcached_behavior_set(replicas[jj], MEMCACHED_BEHAVIOR_NOREPLY, 1);
memcached_set(replicas[jj],
key_ptr,
key_len,
data.data(),
data.length(),
memcached_expiration,
exptime);
memcached_behavior_set(replicas[jj], MEMCACHED_BEHAVIOR_NOREPLY, 0);
}
}
if ((!memcached_success(rc)) &&
(rc != MEMCACHED_NOTSTORED) &&
(rc != MEMCACHED_DATA_EXISTS))
{
if (trail != 0)
{
SAS::Event err(trail, SASEvent::MEMCACHED_SET_FAILED, 0);
err.add_var_param(fqkey);
SAS::report_event(err);
}
LOG_ERROR("Failed to write data for %s to %d replicas",
fqkey.c_str(), replicas.size());
status = Store::Status::ERROR;
}
return status;
}
memcached_return_t BaseMemcachedStore::add_overwriting_tombstone(memcached_st* replica,
const char* key_ptr,
const size_t key_len,
const uint32_t vbucket,
const std::string& data,
time_t memcached_expiration,
uint32_t flags,
SAS::TrailId trail)
{
memcached_return_t rc;
uint64_t cas = 0;
TRC_DEBUG("Attempting to add data for key %.*s", key_len, key_ptr);
// Convert the key into a std::string (sas-client does not like that
// key_{ptr,len} are constant).
const std::string key(key_ptr, key_len);
while (true)
{
if (cas == 0)
{
TRC_DEBUG("Attempting memcached ADD command");
rc = memcached_add_vb(replica,
key_ptr,
key_len,
_binary ? vbucket : 0,
data.data(),
data.length(),
memcached_expiration,
flags);
}
else
{
TRC_DEBUG("Attempting memcached CAS command (cas = %d)", cas);
rc = memcached_cas_vb(replica,
key_ptr,
key_len,
_binary ? vbucket : 0,
data.data(),
data.length(),
memcached_expiration,
flags,
cas);
}
if ((rc == MEMCACHED_DATA_EXISTS) ||
(rc == MEMCACHED_NOTSTORED))
{
// A record with this key already exists. If it is a tombstone, we need
// to overwrite it. Get the record to see what it is.
memcached_return_t get_rc;
std::string existing_data;
TRC_DEBUG("Existing data prevented the ADD/CAS."
"Issue GET to see if we need to overwrite a tombstone");
get_rc = get_from_replica(replica, key_ptr, key_len, existing_data, cas);
if (memcached_success(get_rc))
{
if (existing_data != TOMBSTONE)
{
// The existing record is not a tombstone. We mustn't overwrite
// this, so break out of the loop and return the original return code
// from the ADD/CAS.
TRC_DEBUG("Found real data. Give up");
break;
}
else
{
// The existing record IS a tombstone. Go round the loop again to
// overwrite it. `cas` has been set to the cas of the tombstone.
TRC_DEBUG("Found a tombstone. Attempt to overwrite");
if (trail != 0)
{
SAS::Event event(trail, SASEvent::MEMCACHED_SET_BLOCKED_BY_TOMBSTONE, 0);
event.add_var_param(key);
event.add_static_param(cas);
SAS::report_event(event);
}
}
}
else if (get_rc == MEMCACHED_NOTFOUND)
{
// The GET returned that there is no record for this key. This can
// happen if the record has expired. We need to try again (it could
// have been a tombstone which should not block adds).
TRC_DEBUG("GET failed with NOT_FOUND");
if (trail != 0)
{
SAS::Event event(trail, SASEvent::MEMCACHED_SET_BLOCKED_BY_EXPIRED, 0);
event.add_var_param(key);
SAS::report_event(event);
}
}
else
{
// The replica failed. Return the return code from the original ADD/CAS.
TRC_DEBUG("GET failed, rc = %d (%s)\n%s",
get_rc,
memcached_strerror(replica, get_rc),
memcached_last_error_message(replica));
break;
}
}
else
{
TRC_DEBUG("ADD/CAS returned rc = %d (%s)\n%s",
rc,
memcached_strerror(replica, rc),
memcached_last_error_message(replica));
break;
}
}
return rc;
}
/// Update the data for the specified namespace and key. Writes the data
/// atomically, so if the underlying data has changed since it was last
/// read, the update is rejected and this returns Store::Status::CONTENTION.
Store::Status BaseMemcachedStore::set_data(const std::string& table,
const std::string& key,
const std::string& data,
uint64_t cas,
int expiry,
SAS::TrailId trail)
{
Store::Status status = Store::Status::OK;
TRC_DEBUG("Writing %d bytes to table %s key %s, CAS = %ld, expiry = %d",
data.length(), table.c_str(), key.c_str(), cas, expiry);
// Construct the fully qualified key.
std::string fqkey = table + "\\\\" + key;
const char* key_ptr = fqkey.data();
const size_t key_len = fqkey.length();
int vbucket = vbucket_for_key(fqkey);
const std::vector<memcached_st*>& replicas = get_replicas(vbucket, Op::WRITE);
if (trail != 0)
{
SAS::Event start(trail, SASEvent::MEMCACHED_SET_START, 0);
start.add_var_param(fqkey);
start.add_var_param(data);
start.add_static_param(cas);
start.add_static_param(expiry);
SAS::report_event(start);
}
TRC_DEBUG("%d write replicas for key %s", replicas.size(), fqkey.c_str());
// Calculate a timestamp (least-significant 32 bits of milliseconds since the
// epoch) for the current time. We store this in the flags field to allow us
// to resolve conflicts when resynchronizing between memcached servers.
struct timespec ts;
(void)clock_gettime(CLOCK_REALTIME, &ts);
uint32_t flags = (uint32_t)((ts.tv_sec * 1000) + (ts.tv_nsec / 1000000));
// Memcached uses a flexible mechanism for specifying expiration.
// - 0 indicates never expire.
// - <= MEMCACHED_EXPIRATION_MAXDELTA indicates a relative (delta) time.
// - > MEMCACHED_EXPIRATION_MAXDELTA indicates an absolute time.
// Absolute time is the only way to force immediate expiry. Unfortunately,
// it's not reliable - see https://github.com/Metaswitch/cpp-common/issues/160
// for details. Instead, we use relative time for future times (expiry > 0)
// and the earliest absolute time for immediate expiry (expiry == 0).
time_t memcached_expiration =
(time_t)((expiry > 0) ? expiry : MEMCACHED_EXPIRATION_MAXDELTA + 1);
// First try to write the primary data record to the first responding
// server.
memcached_return_t rc = MEMCACHED_ERROR;
size_t ii;
size_t replica_idx;
// If we only have one replica, we should try it twice -
// libmemcached won't notice a dropped TCP connection until it tries
// to make a request on it, and will fail the request then
// reconnect, so the second attempt could still work.
size_t attempts = (replicas.size() == 1) ? 2: replicas.size();
for (ii = 0; ii < attempts; ++ii)
{
if ((replicas.size() == 1) && (ii == 1)) {
if (rc != MEMCACHED_CONNECTION_FAILURE)
{
// This is a legitimate error, not a transient server failure, so we
// shouldn't retry.
break;
}
replica_idx = 0;
TRC_WARNING("Failed to write to sole memcached replica: retrying once");
}
else
{
replica_idx = ii;
}
TRC_DEBUG("Attempt conditional write to vbucket %d on replica %d (connection %p), CAS = %ld, expiry = %d",
vbucket,
replica_idx,
replicas[replica_idx],
cas,
expiry);
if (cas == 0)
{
// New record, so attempt to add (but overwrite any tombstones we
// encounter). This will fail if someone else got there first and some
// data already exists in memcached for this key.
rc = add_overwriting_tombstone(replicas[replica_idx],
key_ptr,
key_len,
vbucket,
data,
memcached_expiration,
flags,
trail);
}
else
{
// This is an update to an existing record, so use memcached_cas
// to make sure it is atomic.
rc = memcached_cas_vb(replicas[replica_idx],
key_ptr,
key_len,
_binary ? vbucket : 0,
data.data(),
data.length(),
memcached_expiration,
flags,
cas);
if (!memcached_success(rc))
{
TRC_DEBUG("memcached_cas command failed, rc = %d (%s)\n%s",
rc,
memcached_strerror(replicas[replica_idx], rc),
memcached_last_error_message(replicas[replica_idx]));
}
}
if (memcached_success(rc))
{
TRC_DEBUG("Conditional write succeeded to replica %d", replica_idx);
break;
}
else if ((rc == MEMCACHED_NOTSTORED) ||
(rc == MEMCACHED_DATA_EXISTS))
{
if (trail != 0)
{
SAS::Event err(trail, SASEvent::MEMCACHED_SET_CONTENTION, 0);
err.add_var_param(fqkey);
SAS::report_event(err);
}
// A NOT_STORED or EXISTS response indicates a concurrent write failure,
// so return this to the application immediately - don't go on to
// other replicas.
TRC_INFO("Contention writing data for %s to store", fqkey.c_str());
status = Store::Status::DATA_CONTENTION;
break;
}
}
if ((rc == MEMCACHED_SUCCESS) &&
(replica_idx < replicas.size()))
{
// Write has succeeded, so write unconditionally (and asynchronously)
// to the replicas.
for (size_t jj = replica_idx + 1; jj < replicas.size(); ++jj)
{
TRC_DEBUG("Attempt unconditional write to replica %d", jj);
memcached_behavior_set(replicas[jj], MEMCACHED_BEHAVIOR_NOREPLY, 1);
memcached_set_vb(replicas[jj],
key_ptr,
key_len,
_binary ? vbucket : 0,
data.data(),
data.length(),
memcached_expiration,
flags);
memcached_behavior_set(replicas[jj], MEMCACHED_BEHAVIOR_NOREPLY, 0);
}
}
if ((!memcached_success(rc)) &&
(rc != MEMCACHED_NOTSTORED) &&
(rc != MEMCACHED_DATA_EXISTS))
{
if (trail != 0)
{
SAS::Event err(trail, SASEvent::MEMCACHED_SET_FAILED, 0);
err.add_var_param(fqkey);
SAS::report_event(err);
}
update_vbucket_comm_state(vbucket, FAILED);
if (_comm_monitor)
{
_comm_monitor->inform_failure();
}
TRC_ERROR("Failed to write data for %s to %d replicas",
fqkey.c_str(), replicas.size());
status = Store::Status::ERROR;
}
else
{
update_vbucket_comm_state(vbucket, OK);
if (_comm_monitor)
{
_comm_monitor->inform_success();
}
}
return status;
}
