Re: [PATCH][SMB3] 3 small multichannel client patches

[Tom Talpey <tom@talpey.com>]

On 5/7/2021 9:13 PM, Steve French wrote:
> 1) we were not setting CAP_MULTICHANNEL on negotiate request

> diff --git a/fs/cifs/smb2pdu.c b/fs/cifs/smb2pdu.c
> index e36c2a867783..a8bf43184773 100644
> --- a/fs/cifs/smb2pdu.c
> +++ b/fs/cifs/smb2pdu.c
> @@ -841,6 +841,8 @@ SMB2_negotiate(const unsigned int xid, struct cifs_ses *ses)
>  		req->SecurityMode = 0;
>  
>  	req->Capabilities = cpu_to_le32(server->vals->req_capabilities);
> +	if (ses->chan_max > 1)
> +		req->Capabilities |= cpu_to_le32(SMB2_GLOBAL_CAP_MULTI_CHANNEL);
>  
>  	/* ClientGUID must be zero for SMB2.02 dialect */
>  	if (server->vals->protocol_id == SMB20_PROT_ID)
> @@ -1032,6 +1034,9 @@ int smb3_validate_negotiate(const unsigned int xid, struct cifs_tcon *tcon)
>  
>  	pneg_inbuf->Capabilities =
>  			cpu_to_le32(server->vals->req_capabilities);
> +	if (tcon->ses->chan_max > 1)
> +		pneg_inbuf->Capabilities |= cpu_to_le32(SMB2_GLOBAL_CAP_MULTI_CHANNEL);
> +

This doesn't look quite right, and it can lead to failed negotiate by
setting CAP_MULTI_CHANNEL when the server didn't actually send the bit.
Have you tested this with servers that don't do multichannel?


> 2) we were ignoring whether the server set CAP_NEGOTIATE in the response

Is this "CAP_NEGOTIATE" a typo? I think you mean CAP_MULTI_CHANNEL.
In any case:

> diff --git a/fs/cifs/sess.c b/fs/cifs/sess.c
> index 63d517b9f2ff..a391ca3166f3 100644
> --- a/fs/cifs/sess.c
> +++ b/fs/cifs/sess.c
> @@ -97,6 +97,12 @@ int cifs_try_adding_channels(struct cifs_sb_info *cifs_sb, struct cifs_ses *ses)
>  		return 0;
>  	}
>  
> +	if ((ses->server->capabilities & SMB2_GLOBAL_CAP_MULTI_CHANNEL) == false) {

This compares a bit to a boolean. "false" should be "0"?

> +		cifs_dbg(VFS, "server does not support CAP_MULTI_CHANNEL, multichannel disabled\n");

The wording could be clearer. Technically speaking, the server does not
support _multichannel_, which it indicated by not setting CAP_MULTI_CHANNEL.
Also, wouldn't it be more useful to add the servername to this message?
	"server %s does not support multichannel, using single channel"
or similar.


> 3) we were silently ignoring multichannel when "max_channels" was > 1
> but the user forgot to include "multichannel" in mount line.

 > diff --git a/fs/cifs/fs_context.c b/fs/cifs/fs_context.c
 > index 3bcf881c3ae9..8f7af6fcdc76 100644
 > --- a/fs/cifs/fs_context.c
 > +++ b/fs/cifs/fs_context.c
 > @@ -1021,6 +1021,9 @@ static int smb3_fs_context_parse_param(struct 
fs_context *fc,
 >  			goto cifs_parse_mount_err;
 >  		}
 >  		ctx->max_channels = result.uint_32;
 > +		/* If more than one channel requested ... they want multichan */
 > +		if ((ctx->multichannel == false) && (result.uint_32 > 1))
 > +			ctx->multichannel = true;

Wouldn't this be clearer and simpler as just "if (result.uint32 > 1)" ?