pg_stat_monitor/hash_query.c

/*-------------------------------------------------------------------------
 *
 * hash_query.c
 *		Track statement execution times across a whole database cluster.
 *
 * Portions Copyright © 2018-2020, Percona LLC and/or its affiliates
 *
 * Portions Copyright (c) 1996-2020, PostgreSQL Global Development Group
 *
 * Portions Copyright (c) 1994, The Regents of the University of California
 *
 * IDENTIFICATION
 *	  contrib/pg_stat_monitor/hash_query.c
 *
 *-------------------------------------------------------------------------
 */
#include "postgres.h"
#include "nodes/pg_list.h"

#include "pg_stat_monitor.h"


static pgssSharedState *pgss;
static HTAB *pgss_hash;
static HTAB *pgss_query_hash;


static HTAB*
hash_init(const char *hash_name, int key_size, int entry_size, int hash_size)
{
	HASHCTL info;
	memset(&info, 0, sizeof(info));
	info.keysize = key_size;
	info.entrysize = entry_size;
	return ShmemInitHash(hash_name, hash_size, hash_size, &info, HASH_ELEM | HASH_BLOBS);
}

void
pgss_startup(void)
{
	bool		found = false;

	/* reset in case this is a restart within the postmaster */

	pgss = NULL;
	pgss_hash = NULL;

	/*
	* Create or attach to the shared memory state, including hash table
	*/
	LWLockAcquire(AddinShmemInitLock, LW_EXCLUSIVE);

	pgss = ShmemInitStruct("pg_stat_monitor", sizeof(pgssSharedState), &found);
	if (!found)
	{
		/* First time through ... */
		pgss->lock = &(GetNamedLWLockTranche("pg_stat_monitor"))->lock;
		SpinLockInit(&pgss->mutex);
		ResetSharedState(pgss);
	}

#ifdef BENCHMARK
	init_hook_stats();
#endif

	set_qbuf((unsigned char *)ShmemAlloc(MAX_QUERY_BUF));

	pgss_hash = hash_init("pg_stat_monitor: bucket hashtable", sizeof(pgssHashKey), sizeof(pgssEntry), MAX_BUCKET_ENTRIES);
	pgss_query_hash = hash_init("pg_stat_monitor: queryID hashtable", sizeof(uint64), sizeof(pgssQueryEntry), MAX_BUCKET_ENTRIES);

	LWLockRelease(AddinShmemInitLock);

	/*
	 * If we're in the postmaster (or a standalone backend...), set up a shmem
	 * exit hook to dump the statistics to disk.
	 */
	on_shmem_exit(pgss_shmem_shutdown, (Datum) 0);
}

pgssSharedState*
pgsm_get_ss(void)
{
	return pgss;
}

HTAB*
pgsm_get_hash(void)
{
	return pgss_hash;
}

HTAB*
pgsm_get_query_hash(void)
{
	return pgss_query_hash;
}

/*
 * shmem_shutdown hook: Dump statistics into file.
 *
 * Note: we don't bother with acquiring lock, because there should be no
 * other processes running when this is called.
 */
void
pgss_shmem_shutdown(int code, Datum arg)
{
	/* Don't try to dump during a crash. */
	if (code)
		return;

	pgss = NULL;
	/* Safety check ... shouldn't get here unless shmem is set up. */
	if (!IsHashInitialize())
		return;
}

Size
hash_memsize(void)
{
	Size	size;

	size = MAXALIGN(sizeof(pgssSharedState));
	size += MAXALIGN(MAX_QUERY_BUF);
	size = add_size(size, hash_estimate_size(MAX_BUCKET_ENTRIES, sizeof(pgssEntry)));
	size = add_size(size, hash_estimate_size(MAX_BUCKET_ENTRIES, sizeof(pgssQueryEntry)));

	return size;
}

pgssEntry *
hash_entry_alloc(pgssSharedState *pgss, pgssHashKey *key, int encoding)
{
	pgssEntry	*entry = NULL;
	bool		found = false;

	if (hash_get_num_entries(pgss_hash) >= MAX_BUCKET_ENTRIES)
	{
		elog(DEBUG1, "%s", "pg_stat_monitor: out of memory");
		return NULL;
	}
	/* Find or create an entry with desired hash code */
	entry = (pgssEntry *) hash_search(pgss_hash, key, HASH_ENTER_NULL, &found);
	if (entry == NULL)
		pgsm_log_error("hash_entry_alloc: OUT OF MEMORY");
	else if (!found)
	{
		pgss->bucket_entry[pg_atomic_read_u64(&pgss->current_wbucket)]++;
		/* New entry, initialize it */
		/* reset the statistics */
		memset(&entry->counters, 0, sizeof(Counters));
		/* set the appropriate initial usage count */
		/* re-initialize the mutex each time ... we assume no one using it */
		SpinLockInit(&entry->mutex);
		/* ... and don't forget the query text metadata */
		entry->encoding = encoding;
	}

	return entry;
}

/*
 * Prepare resources for using the new bucket:
 *    - Deallocate finished hash table entries in new_bucket_id (entries whose
 *      state is PGSS_FINISHED or PGSS_FINISHED).
 *    - Clear query buffer for new_bucket_id.
 *    - If old_bucket_id != -1, move all pending hash table entries in
 *      old_bucket_id to the new bucket id, also move pending queries from the 
 *      previous query buffer (query_buffer[old_bucket_id]) to the new one
 *      (query_buffer[new_bucket_id]).
 *
 * Caller must hold an exclusive lock on pgss->lock.
 */
void
hash_entry_dealloc(int new_bucket_id, int old_bucket_id, unsigned char *query_buffer)
{
	HASH_SEQ_STATUS hash_seq;
	pgssEntry		*entry = NULL;

	/* Store pending query ids from the previous bucket. */
	List        *pending_entries = NIL;
	ListCell    *pending_entry;

	/* Iterate over the hash table. */
	hash_seq_init(&hash_seq, pgss_hash);
	while ((entry = hash_seq_search(&hash_seq)) != NULL)
	{
		/*
		 * Remove all entries if new_bucket_id == -1.
		 * Otherwise remove entry in new_bucket_id if it has finished already.
		 */
		if (new_bucket_id < 0 ||
			(entry->key.bucket_id == new_bucket_id &&
				 (entry->counters.state == PGSS_FINISHED || entry->counters.state == PGSS_ERROR)))
		{
			if (new_bucket_id == -1) {
				/* pg_stat_monitor_reset(), remove entry from query hash table too. */
				hash_search(pgss_query_hash, &(entry->key.queryid), HASH_REMOVE, NULL);
			}

			entry = hash_search(pgss_hash, &entry->key, HASH_REMOVE, NULL);
		}

		/*
		 * If we detect a pending query residing in the previous bucket id,
		 * we add it to a list of pending elements to be moved to the new
		 * bucket id.
		 * Can't update the hash table while iterating it inside this loop,
		 * as this may introduce all sort of problems.
		 */
		if (old_bucket_id != -1 && entry->key.bucket_id == old_bucket_id)
		{
			if (entry->counters.state == PGSS_PARSE ||
				entry->counters.state == PGSS_PLAN ||
				entry->counters.state == PGSS_EXEC)
			{
				pgssEntry *bkp_entry = malloc(sizeof(pgssEntry));
				if (!bkp_entry)
				{
					pgsm_log_error("hash_entry_dealloc: out of memory");
					/*
					 * No memory, If the entry has calls > 1 then we change the state to finished,
					 * as the pending query will likely finish execution during the new bucket
					 * time window. The pending query will vanish in this case, can't list it
					 * until it completes.
					 *
					 * If there is only one call to the query and it's pending, remove the
					 * entry from the previous bucket and allow it to finish in the new bucket,
					 * in order to avoid the query living in the old bucket forever.
					 */
					if (entry->counters.calls.calls > 1)
						entry->counters.state = PGSS_FINISHED;
					else
						entry = hash_search(pgss_hash, &entry->key, HASH_REMOVE, NULL);
					continue;
				}

				/* Save key/data from the previous entry. */
				memcpy(bkp_entry, entry, sizeof(pgssEntry));

				/* Update key to use the new bucket id. */
				bkp_entry->key.bucket_id = new_bucket_id;

				/* Add the entry to a list of nodes to be processed later. */
				pending_entries = lappend(pending_entries, bkp_entry);

				/*
				 * If the entry has calls > 1 then we change the state to finished in
				 * the previous bucket, as the pending query will likely finish execution
				 * during the new bucket time window. Can't remove it from the previous bucket
				 * as it may have many calls and we would lose the query statistics.
				 *
				 * If there is only one call to the query and it's pending, remove the entry
				 * from the previous bucket and allow it to finish in the new bucket,
				 * in order to avoid the query living in the old bucket forever.
				 */
				if (entry->counters.calls.calls > 1)
					entry->counters.state = PGSS_FINISHED;
				else
					entry = hash_search(pgss_hash, &entry->key, HASH_REMOVE, NULL);
			}
		}
	}

	/*
	 * Iterate over the list of pending queries in order
	 * to add them back to the hash table with the updated bucket id.
	 */
	foreach (pending_entry, pending_entries) {
		bool found = false;
		pgssEntry	*new_entry;
		pgssEntry	*old_entry = (pgssEntry *) lfirst(pending_entry);

		new_entry = (pgssEntry *) hash_search(pgss_hash, &old_entry->key, HASH_ENTER_NULL, &found);
		if (new_entry == NULL)
			elog(DEBUG1, "%s", "pg_stat_monitor: out of memory");
		else if (!found)
		{
			/* Restore counters and other data. */
			new_entry->counters = old_entry->counters;
			SpinLockInit(&new_entry->mutex);
			new_entry->encoding = old_entry->encoding;
			new_entry->query_pos = old_entry->query_pos;
		}

		free(old_entry);
	}

	list_free(pending_entries);
}

/*
 * Release all entries.
 */
void
hash_entry_reset()
{
	pgssSharedState *pgss   = pgsm_get_ss();
	HASH_SEQ_STATUS		hash_seq;
	pgssEntry			*entry;

	LWLockAcquire(pgss->lock, LW_EXCLUSIVE);

	hash_seq_init(&hash_seq, pgss_hash);
	while ((entry = hash_seq_search(&hash_seq)) != NULL)
	{
		hash_search(pgss_hash, &entry->key, HASH_REMOVE, NULL);
	}
	pg_atomic_write_u64(&pgss->current_wbucket, 0);
	LWLockRelease(pgss->lock);
}

bool
IsHashInitialize(void)
{
	return (pgss != NULL &&
			pgss_hash != NULL);
}