[sheepdog] [PATCH v2] sheep: improve durability when creating object file.

[Liu Yuan]

On Wed, Mar 04, 2015 at 04:04:05PM +0800, Xu Yifeng wrote:
> When creating a new file, make sure all data is written to disk, this is
> implemented in two steps:
> 
> 1. Temporary file is written and committed to disk, this includes file system
>    meta data and application data.
> 
> 2. The temporary file is renamed, the action should also be commited to disk
>    by calling fsync() on its directory.
> 
> Signed-off-by: Xu Yifeng <skypexu at gmail.com>
> ---
>  sheep/plain_store.c | 39 ++++++++++++++++++++++++++++++++-------
>  1 file changed, 32 insertions(+), 7 deletions(-)
> 
> diff --git a/sheep/plain_store.c b/sheep/plain_store.c
> index 5ea2946..0fb6247 100644
> --- a/sheep/plain_store.c
> +++ b/sheep/plain_store.c
> @@ -27,10 +27,11 @@ static inline bool iocb_is_aligned(const struct siocb *iocb)
>  
>  static int prepare_iocb(uint64_t oid, const struct siocb *iocb, bool create)
>  {
> -	int flags = O_DSYNC | O_RDWR;
> +	int syncflag = create ? O_SYNC : O_DSYNC;
> +	int flags = syncflag | O_RDWR;
>  
>  	if (uatomic_is_true(&sys->use_journal) || sys->nosync == true)
> -		flags &= ~O_DSYNC;
> +		flags &= ~syncflag;
>  
>  	if (sys->backend_dio && iocb_is_aligned(iocb)) {
>  		if (!is_aligned_to_pagesize(iocb->buf))
> @@ -391,7 +392,7 @@ int default_read(uint64_t oid, const struct siocb *iocb)
>  
>  int default_create_and_write(uint64_t oid, const struct siocb *iocb)
>  {
> -	char path[PATH_MAX], tmp_path[PATH_MAX];
> +	char path[PATH_MAX], tmp_path[PATH_MAX], *dir;
>  	int flags = prepare_iocb(oid, iocb, true);
>  	int ret, fd;
>  	uint32_t len = iocb->length;
> @@ -409,7 +410,7 @@ int default_create_and_write(uint64_t oid, const struct siocb *iocb)
>  	    != SD_RES_SUCCESS) {
>  		sd_err("turn off journaling");
>  		uatomic_set_false(&sys->use_journal);
> -		flags |= O_DSYNC;
> +		flags |= O_SYNC;
>  		sync();
>  	}
>  
> @@ -462,11 +463,35 @@ int default_create_and_write(uint64_t oid, const struct siocb *iocb)
>  		goto out;
>  	}
>  
> -	ret = SD_RES_SUCCESS;
> +	close(fd);
> +
> +	if (uatomic_is_true(&sys->use_journal)) {

This should be if (uatomic_is_true(&sys->use_journal) || sys->nosync == true)

Thanks,
Yuan