[sheepdog] [PATCH] dog: fix cluster-wide snapshot inconsistency due to vdi deleted

[Liu Yuan]

On Wed, Aug 20, 2014 at 04:53:46PM +0800, Ruoyu wrote:
> Once vdi snapshot is deleted before executing 'dog cluster snapshot
> save', error messsage just like as below is seen.
> 
> Failed to read object 807c2b2500000000 No object found
> Failed to read inode header
> 
> Once a vdi is deleted, cluster snapshot save will not backup it to
> farm so that cluster snapshot load cannot restore it to new cluster.
> However, the vdi id is marked as in use for safety, vdi list
> will try to read all vdis in use. Therefore, dog cannot read the object.
> 
> The solution is saving the deleted vdi object too. To achieve it,
> a parameter, include_deleted, is added to the function parse_vdi.

s/include_deleted/no_deleted looks better.

> The default value is false. But for cluster snapshot save, it is true.
> 
> Functional test 030 is also updated to simulate the scenario.
> 
> Reported-by: Valerio Pachera <sirio81 at gmail.com>
> Signed-off-by: Ruoyu <liangry at ucweb.com>
> ---
>  dog/cluster.c            |  5 +++--
>  dog/common.c             |  9 ++++++---
>  dog/dog.h                |  3 ++-
>  dog/node.c               |  2 +-
>  dog/vdi.c                | 19 +++++++++++--------
>  tests/functional/030     | 14 ++++++++++++++
>  tests/functional/030.out | 13 +++++++++++++
>  7 files changed, 50 insertions(+), 15 deletions(-)
> 
> diff --git a/dog/cluster.c b/dog/cluster.c
> index 9cf96a8..86188e2 100644
> --- a/dog/cluster.c
> +++ b/dog/cluster.c
> @@ -419,7 +419,8 @@ static int save_snapshot(int argc, char **argv)
>  
>  	opt.count = argc - optind;
>  	opt.name = argv + optind;
> -	if (parse_vdi(fill_object_tree, SD_INODE_SIZE, &opt) != SD_RES_SUCCESS)
> +	if (parse_vdi(fill_object_tree, SD_INODE_SIZE,
> +			&opt, true) != SD_RES_SUCCESS)
>  		goto out;
>  
>  	if (object_tree_size() == 0) {
> @@ -657,7 +658,7 @@ static void cluster_check_cb(uint32_t vid, const char *name, const char *tag,
>  
>  static int cluster_check(int argc, char **argv)
>  {
> -	if (parse_vdi(cluster_check_cb, SD_INODE_SIZE, NULL) < 0)
> +	if (parse_vdi(cluster_check_cb, SD_INODE_SIZE, NULL, false) < 0)
>  		return EXIT_SYSFAIL;
>  
>  	return EXIT_SUCCESS;
> diff --git a/dog/common.c b/dog/common.c
> index 2335e14..8cbf44c 100644
> --- a/dog/common.c
> +++ b/dog/common.c
> @@ -128,7 +128,8 @@ int dog_write_object(uint64_t oid, uint64_t cow_oid, void *data,
>  
>  #define FOR_EACH_VDI(nr, vdis) FOR_EACH_BIT(nr, vdis, SD_NR_VDIS)
>  
> -int parse_vdi(vdi_parser_func_t func, size_t size, void *data)
> +int parse_vdi(vdi_parser_func_t func, size_t size, void *data,
> +		bool include_deleted)
>  {
>  	int ret;
>  	unsigned long nr;
> @@ -163,8 +164,10 @@ int parse_vdi(vdi_parser_func_t func, size_t size, void *data)
>  			continue;
>  		}
>  
> -		if (i->name[0] == '\0') /* this VDI has been deleted */
> -			continue;
> +		if (!include_deleted) {
> +			if (i->name[0] == '\0') /* this VDI has been deleted */
> +				continue;
> +		}

then
	if (no_deleted && i->name[0] == '\0')
		continue;

looks better self-explained.

Thanks
Yuan