Re: [PATCH v2 1/6] xlog: store prev vclock in xlog header

[Vladimir Davydov <vdavydov.dev@gmail.com>]

On Thu, Jul 05, 2018 at 09:52:41AM +0300, Konstantin Osipov wrote:
> * Vladimir Davydov <vdavydov.dev@gmail.com> [18/06/29 19:49]:
> > This patch adds a new key to xlog header, PrevVclock, which contains the
> > vclock of the previous xlog file in the xlog directory. It is set by
> > xdir_create_xlog() to the last vclock in xdir::index. The new key is
> > only present in XLOG files (it doesn't make sense for SNAP or VYLOG
> > anyway). It will be used to make the check for xlog gaps more thorough.
> 
> What about the following diff:
> 
> ---
> diff --git a/src/box/xlog.c b/src/box/xlog.c
> index 59459b25d..a60f2ffa8 100644
> --- a/src/box/xlog.c
> +++ b/src/box/xlog.c
> @@ -126,12 +126,10 @@ xlog_meta_format(const struct xlog_meta *meta, char *buf, int size)
>  		VCLOCK_KEY ": %s\n",
>  		meta->filetype, v13, PACKAGE_VERSION, instance_uuid, vstr);
>  	free(vstr);
> -	if (meta->has_prev_vclock) {
> -		vstr = vclock_to_string(&meta->prev_vclock);
> -		SNPRINT(total, snprintf, buf, size,
> -			PREV_VCLOCK_KEY ": %s\n", vstr);
> -		free(vstr);
> -	}
> +	vstr = vclock_to_string(&meta->prev_vclock);
> +	SNPRINT(total, snprintf, buf, size,
> +		PREV_VCLOCK_KEY ": %s\n", vstr);
> +	free(vstr);
>  	SNPRINT(total, snprintf, buf, size, "\n");
>  	assert(total > 0);
>  	return total;
> @@ -263,7 +261,6 @@ xlog_meta_parse(struct xlog_meta *meta, const char **data,
>  			 */
>  			if (parse_vclock(val, val_end, &meta->prev_vclock) != 0)
>  				return -1;
> -			meta->has_prev_vclock = true;

We need to know if PrevVclock is available so that we can skip the check
when recovering from an xlog generated by a previous version.

>  		} else if (memcmp(key, VERSION_KEY, key_end - key) == 0) {
>  			/* Ignore Version: for now */
>  		} else {
> @@ -926,12 +923,10 @@ xdir_create_xlog(struct xdir *dir, struct xlog *xlog,
>  	 * For WAL dir: store vclock of the previous xlog file
>  	 * to check for gaps on recovery.
>  	 */
> -	if (dir->type == XLOG && !vclockset_empty(&dir->index)) {
> +	if (!vclockset_empty(&dir->index)) {
>  		vclock_copy(&meta.prev_vclock, vclockset_last(&dir->index));
> -		meta.has_prev_vclock = true;
>  	} else {
>  		vclock_create(&meta.prev_vclock);
> -		meta.has_prev_vclock = false;
>  	}

Storing vclock in snap and vylog files doesn't make any sense IMHO.
I'd prefer to check dir->type explicitly.

If you dislike the extra variable, we can probably introduce a helper to
check if vclock is set, something like

	static inline bool
	vclock_is_set(const struct vclock *vclock)
	{
		return vclock->signature >= 0;
	}