Re: [Tarantool-patches] [PATCH 1/9] wal: enrich row's meta information with sync replication flags

[Cyrill Gorcunov via Tarantool-patches <tarantool-patches@dev.tarantool.org>]

On Sun, Apr 11, 2021 at 08:55:56PM +0300, Serge Petrenko wrote:
> diff --git a/src/box/iproto_constants.h b/src/box/iproto_constants.h
> index d3738c705..f7f46088f 100644
> --- a/src/box/iproto_constants.h
> +++ b/src/box/iproto_constants.h
> @@ -49,9 +49,14 @@ enum {
>  	XLOG_FIXHEADER_SIZE = 19
>  };
>  
> +/** IPROTO_FLAGS bitfield constants. */
>  enum {
>  	/** Set for the last xrow in a transaction. */
>  	IPROTO_FLAG_COMMIT = 0x01,
> +	/** Set for the last row of a tx residing in limbo. */
> +	IPROTO_FLAG_WAIT_SYNC = 0x02,
> +	/** Set for the last row of a synchronous tx. */
> +	IPROTO_FLAG_WAIT_ACK = 0x04,
>  };

Serge, is there some particular reason why cant we make them 1:1
mapping to txn flags? Ie

	IPROTO_FLAG_WAIT_SYNC	= 0x10,
	IPROTO_FLAG_WAIT_ACK	= 0x20,

this would allow us to eliminate branching in code. While txn flags
are not part of API and iproto flags _are_ I don't expect the former
to be changed anyhow soon?

Not insisting though, just share the idea.

>  
>  enum iproto_key {
> diff --git a/src/box/journal.h b/src/box/journal.h
> index 76c70c19f..9ab8af2c1 100644
> --- a/src/box/journal.h
> +++ b/src/box/journal.h
> @@ -63,6 +63,7 @@ struct journal_entry {
>  	 * A journal entry completion callback argument.
>  	 */
>  	void *complete_data;
> +	uint8_t opt_flags;
>  	/**
>  	 * Asynchronous write completion function.
>  	 */
> @@ -97,6 +98,7 @@ journal_entry_create(struct journal_entry *entry, size_t n_rows,
>  	entry->approx_len	= approx_len;
>  	entry->n_rows		= n_rows;
>  	entry->res		= -1;
> +	entry->opt_flags = 0;
>  }

Please don't ruine alignment here. I know that Vlad prefer dense style
(which is actually hard to parse by eyes at least for me) but here we
have a block which ether should be whole formatted to dense or kept
aligned.

>  
>  /**
> diff --git a/src/box/txn.c b/src/box/txn.c
> index 40061ff09..d65315f58 100644
> --- a/src/box/txn.c
> +++ b/src/box/txn.c
> @@ -76,6 +76,7 @@ txn_add_redo(struct txn *txn, struct txn_stmt *stmt, struct request *request)
>  		row->lsn = 0;
>  		row->sync = 0;
>  		row->tm = 0;
> +		row->opt_flags = 0;
>  	}
>  	/*
>  	 * Group ID should be set both for requests not having a
> @@ -681,6 +682,10 @@ txn_journal_entry_new(struct txn *txn)
>  		--req->n_rows;
>  	}
>  
> +	req->opt_flags |=
> +		(txn_has_flag(txn, TXN_WAIT_SYNC) ? IPROTO_FLAG_WAIT_SYNC : 0) |
> +		(txn_has_flag(txn, TXN_WAIT_ACK) ? IPROTO_FLAG_WAIT_ACK : 0);
> +

When mentioning branching I meant this code. We could use 1:1 mapping
here if we define iproto flags to be the same as txn flags.

Or say some explicit map could be provided (for readability sake)

	static const uint8_t opt_flags_map[] = {
		[TXN_WAIT_SYNC]		= IPROTO_FLAG_WAIT_SYNC,
		[TXN_WAIT_ACK]		= IPROTO_FLAG_WAIT_ACK,
	};

	req->opt_flags |= opt_flags_map[txn->flags & IPROTO_FLAG_WAIT_SYNC];
	req->opt_flags |= opt_flags_map[txn->flags & IPROTO_FLAG_WAIT_ACK];

Such technique is widely used when need some mapping between flags and
instead of number of `if` one simply use read-only memory.

On the other hands I'm fine with branching here because this is not
that hot code and even misprediction should not harm much. So just
to share the ideas.

>  	return req;
>  }
>  
> diff --git a/src/box/wal.c b/src/box/wal.c
> index 34af0bda6..00fcb21b4 100644
> --- a/src/box/wal.c
> +++ b/src/box/wal.c
> @@ -962,14 +962,14 @@ out:
>   */
>  static void
>  wal_assign_lsn(struct vclock *vclock_diff, struct vclock *base,
> -	       struct xrow_header **row,
> -	       struct xrow_header **end)
> +	       struct journal_entry *entry)
>  {
>  	int64_t tsn = 0;
> -	struct xrow_header **start = row;
> -	struct xrow_header **first_glob_row = row;
> +	struct xrow_header **start = entry->rows;
> +	struct xrow_header **end = entry->rows + entry->n_rows;
> +	struct xrow_header **first_glob_row = start;

There is no need to define another dependency, actually I think
the compiler will figure out that @first_glob_row and @start
are initialized from same @entry->rows still previously this
was more clear, iow I propose to leave `first_glob_row = entry->rows;`

>  	/** Assign LSN to all local rows. */
> -	for ( ; row < end; row++) {
> +	for (struct xrow_header **row = start; row < end; row++) {
>  		if ((*row)->replica_id == 0) {
>  			/*
>  			 * All rows representing local space data
> @@ -996,7 +996,13 @@ wal_assign_lsn(struct vclock *vclock_diff, struct vclock *base,
>  				first_glob_row = row;
>  			}
>  			(*row)->tsn = tsn == 0 ? (*start)->lsn : tsn;
> -			(*row)->is_commit = row == end - 1;
> +			if (row < end - 1)
> +				continue;
> +			/* Tx meta is stored in the last tx row. */
> +			if (row == end - 1) {
> +				(*row)->opt_flags = entry->opt_flags;
> +				(*row)->is_commit = true;

Why can't we use
				(*row)->opt_flags = entry->opt_flags | IPROTO_FLAG_COMMIT;
instead?

> +			}
>  		} else {
>  			int64_t diff = (*row)->lsn - vclock_get(base, (*row)->replica_id);
>  			if (diff <= vclock_get(vclock_diff,
> @@ -1020,7 +1026,7 @@ wal_assign_lsn(struct vclock *vclock_diff, struct vclock *base,
>  	 * the first global row. tsn was yet unknown when those
>  	 * rows were processed.
>  	 */
> -	for (row = start; row < first_glob_row; row++)
> +	for (struct xrow_header **row = start; row < first_glob_row; row++)
>  		(*row)->tsn = tsn;
>  }
>  
> @@ -1098,8 +1104,7 @@ wal_write_to_disk(struct cmsg *msg)
>  	struct journal_entry *entry;
>  	struct stailq_entry *last_committed = NULL;
>  	stailq_foreach_entry(entry, &wal_msg->commit, fifo) {
> -		wal_assign_lsn(&vclock_diff, &writer->vclock,
> -			       entry->rows, entry->rows + entry->n_rows);
> +		wal_assign_lsn(&vclock_diff, &writer->vclock, entry);
>  		entry->res = vclock_sum(&vclock_diff) +
>  			     vclock_sum(&writer->vclock);
>  		rc = xlog_write_entry(l, entry);
> @@ -1319,8 +1324,7 @@ wal_write_none_async(struct journal *journal,
>  	struct vclock vclock_diff;
>  
>  	vclock_create(&vclock_diff);
> -	wal_assign_lsn(&vclock_diff, &writer->vclock, entry->rows,
> -		       entry->rows + entry->n_rows);
> +	wal_assign_lsn(&vclock_diff, &writer->vclock, entry);
>  	vclock_merge(&writer->vclock, &vclock_diff);
>  	vclock_copy(&replicaset.vclock, &writer->vclock);
>  	entry->res = vclock_sum(&writer->vclock);
> diff --git a/src/box/xrow.c b/src/box/xrow.c
> index bc06738ad..cc8e43ed4 100644
> --- a/src/box/xrow.c
> +++ b/src/box/xrow.c
> @@ -183,7 +183,7 @@ error:
>  			break;
>  		case IPROTO_FLAGS:
>  			flags = mp_decode_uint(pos);
> -			header->is_commit = flags & IPROTO_FLAG_COMMIT;
> +			header->opt_flags = flags;
>  			break;
>  		default:
>  			/* unknown header */
> @@ -299,6 +299,7 @@ xrow_header_encode(const struct xrow_header *header, uint64_t sync,
>  	 *   flag to find transaction boundary (last row in the
>  	 *   transaction stream).
>  	 */
> +	uint8_t flags_to_encode = header->opt_flags & ~IPROTO_FLAG_COMMIT;
>  	if (header->tsn != 0) {
>  		if (header->tsn != header->lsn || !header->is_commit) {
>  			/*
> @@ -314,12 +315,14 @@ xrow_header_encode(const struct xrow_header *header, uint64_t sync,
>  			map_size++;
>  		}
>  		if (header->is_commit && header->tsn != header->lsn) {
> -			/* Setup last row for multi row transaction. */
> -			d = mp_encode_uint(d, IPROTO_FLAGS);
> -			d = mp_encode_uint(d, IPROTO_FLAG_COMMIT);
> -			map_size++;
> +			flags_to_encode |= IPROTO_FLAG_COMMIT;
>  		}
>  	}
> +	if (flags_to_encode != 0) {
> +		d = mp_encode_uint(d, IPROTO_FLAGS);
> +		d = mp_encode_uint(d, flags_to_encode);
> +		map_size++;
> +	}
>  	assert(d <= data + XROW_HEADER_LEN_MAX);
>  	mp_encode_map(data, map_size);
>  	out->iov_len = d - (char *) out->iov_base;
> diff --git a/src/box/xrow.h b/src/box/xrow.h
> index fde8f9474..2a18733c0 100644
> --- a/src/box/xrow.h
> +++ b/src/box/xrow.h
> @@ -80,14 +80,28 @@ struct xrow_header {
>  	 * transaction.
>  	 */
>  	int64_t tsn;
> -	/**
> -	 * True for the last row in a multi-statement transaction,
> -	 * or single-statement transaction. Is only encoded in the
> -	 * write ahead log for multi-statement transactions.
> -	 * Single-statement transactions do not encode
> -	 * tsn and is_commit flag to save space.
> -	 */
> -	bool is_commit;
> +	/** Transaction meta flags set only in the last transaction row. */
> +	union {
> +		uint8_t opt_flags;
> +		struct {
> +			/**
> +			 * Is only encoded in the write ahead log for
> +			 * multi-statement transactions. Single-statement
> +			 * transactions do not encode tsn and is_commit flag to
> +			 * save space.
> +			 */
> +			bool is_commit : 1;
> +			/**
> +			 * True for a synchronous transaction.
> +			 */
> +			bool wait_sync : 1;
> +			/**
> +			 * True for any transaction that would enter the limbo
> +			 * (not necessarily a synchronous one).
> +			 */
> +			bool wait_ack  : 1;
> +		};
> +	};

Serge, I know that you mention bitfields in the mail in context of our
discussion and I assume you'll rework this otherwise this will be
just a hidden bug which might work for sometime but one day it will
trigger and we won't even be able to figure out why.

>  
>  	int bodycnt;
>  	uint32_t schema_version;
> -- 
> 2.24.3 (Apple Git-128)
> 

	Cyrill