Re: [PATCH 3/3] pnfs: add a new mechanism to select a layout driver according to an ordered list

["J. Bruce Fields" <bfields@fieldses.org>]

On Tue, Jun 07, 2016 at 06:38:11AM -0400, Jeff Layton wrote:
> Currently, the layout driver selection code attempts to divine meaning
> from the order of the layout driver list sent by the server.
> Unfortunately, the spec doesn't place any significance on the order
> and the server is free to send them in any order it likes.
> 
> Instead, set a list of preferred driver types in the kernel and have
> the selection code try them in order until it finds one that can be
> loaded.
> 
> If we go through the whole list of preferred drivers and don't find one,
> then try any that weren't recognized in the first scan. This should
> allow the use of 3rd party and experimental drivers without needing to
> muck with the order of preference.
> 
> For now, the order of preference is hardcoded, but it should be possible
> to make this configurable (via module param perhaps?).
> 
> Signed-off-by: Jeff Layton <jeff.layton@primarydata.com>
> ---
>  fs/nfs/pnfs.c | 71 +++++++++++++++++++++++++++++++++++++++++------------------
>  1 file changed, 50 insertions(+), 21 deletions(-)
> 
> diff --git a/fs/nfs/pnfs.c b/fs/nfs/pnfs.c
> index b02cad9c04bf..3ec5f2b392b6 100644
> --- a/fs/nfs/pnfs.c
> +++ b/fs/nfs/pnfs.c
> @@ -99,6 +99,21 @@ unset_pnfs_layoutdriver(struct nfs_server *nfss)
>  }
>  
>  /*
> + * When the server sends a list of layout types, we choose one in the order
> + * given in the list below.
> + *
> + * FIXME: should this list be configurable via module_param or something?
> + */
> +static const u32 ld_prefs[] = {
> +	LAYOUT_SCSI,
> +	LAYOUT_BLOCK_VOLUME,
> +	LAYOUT_OSD2_OBJECTS,
> +	LAYOUT_FLEX_FILES,
> +	LAYOUT_NFSV4_1_FILES,
> +	0
> +};
> +
> +/*
>   * Try to set the server's pnfs module to the pnfs layout type specified by id.
>   * Currently only one pNFS layout driver per filesystem is supported.
>   *
> @@ -110,7 +125,7 @@ set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
>  {
>  	struct pnfs_layoutdriver_type *ld_type = NULL;
>  	u32 id;
> -	int i;
> +	int i, j;
>  
>  	if (!(server->nfs_client->cl_exchange_flags &
>  		 (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
> @@ -118,31 +133,45 @@ set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
>  			__func__, server->nfs_client->cl_exchange_flags);
>  		goto out_no_driver;
>  	}
> -	/*
> -	 * If server supports more than one layout types.
> -	 * By assuming, that server will put 'common default' as the first
> -	 * entry, try all following entries ibefore and fall back to the default
> -	 * if we did not found a matching one.
> -	 */
> -	for(i = 1; i < NFS_MAX_LAYOUT_TYPES && ids[i] != 0; i++) {
> -		id = ids[i];
> -		request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX, id);
> -		ld_type = find_pnfs_driver(id);
> -		if(ld_type)
> -			goto found_module;
>  
> -		dprintk("%s: No pNFS module found for %u.\n", __func__, id);
> +	/* scan the list for each layout type in order of preference */
> +	for (j = 0; ld_prefs[j] != 0; j++) {
> +		for (i = 0; i < NFS_MAX_LAYOUT_TYPES && ids[i] != 0; i++) {
> +			id = ids[i];
> +
> +			if (ld_prefs[j] == id) {
> +				request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX, id);
> +				ld_type = find_pnfs_driver(id);
> +				if (ld_type)
> +					goto found_module;
> +			}
> +		}
>  	}
>  
> -	/*
> -	 * no other layout types found. Try default one.
> -	 */
> -	id = ids[0];
> -	request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX, id);
> -	ld_type = find_pnfs_driver(id);
> +	/* didn't find one -- make sure we try any that weren't in ld_prefs */
> +	for (i = 0; i < NFS_MAX_LAYOUT_TYPES && ids[i] != 0; i++) {
> +		bool	match = false;
> +
> +		id = ids[i];
> +
> +		/* Was it in the prefs list? */
> +		for (j = 0; ld_prefs[j] != 0; j++) {
> +			if (ld_prefs[j] != id) {
> +				match = true;
> +				break;
> +			}
> +		}

This logic feels more complicated than necessary.

We're going out of our way to support (at this point purely theoretical)
3rd-party layout modules.  When new layouts are develop, the chance
they'll be implementable with *no* changes to kernel code outside that
module seem small, and once you have to touch other code you may as well
update ld_prefs.

But anyway, if we want this, it might be easier to follow with logic
like:

	1. sort the ids[] array so the known layouts are at the top, in
	   ld_prefs order.
	2. try request_module() and find_pnfs_driver() in order on
	   the sorted ids[] array.

--b.

> +
> +		if (!match) {
> +			request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX, id);
> +			ld_type = find_pnfs_driver(id);
> +			if (ld_type)
> +				goto found_module;
> +		}
> +	}
>  
>  	if (!ld_type) {
> -		dprintk("%s: No pNFS module found for %u.\n", __func__, id);
> +		dprintk("%s: No pNFS module found!\n", __func__);
>  		goto out_no_driver;
>  	}
>  
> -- 
> 2.5.5