Re: [PATCH 6/9] tgt: convert ibmvstgt and libsrp to use scsi_data_buffer

[FUJITA Tomonori <tomof@acm.org>]

On Sun, 09 Sep 2007 13:12:03 +0300
Boaz Harrosh <bharrosh@panasas.com> wrote:

> On Fri, Sep 07 2007 at 0:50 +0300, FUJITA Tomonori <tomof@acm.org> wrote:
> > Signed-off-by: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
> > ---
> >  drivers/scsi/ibmvscsi/ibmvstgt.c |    2 +-
> >  drivers/scsi/libsrp.c            |   27 ++++++++++++++-------------
> >  2 files changed, 15 insertions(+), 14 deletions(-)
> > 
> > diff --git a/drivers/scsi/ibmvscsi/ibmvstgt.c b/drivers/scsi/ibmvscsi/ibmvstgt.c
> > index 8ba7dd0..ab173d4 100644
> > --- a/drivers/scsi/ibmvscsi/ibmvstgt.c
> > +++ b/drivers/scsi/ibmvscsi/ibmvstgt.c
> > @@ -286,7 +286,7 @@ static int ibmvstgt_cmd_done(struct scsi_cmnd *sc,
> >  	dprintk("%p %p %x %u\n", iue, target, vio_iu(iue)->srp.cmd.cdb[0],
> >  		cmd->usg_sg);
> >  
> > -	if (sc->use_sg)
> > +	if (scsi_sg_count(sc))
> >  		err = srp_transfer_data(sc, &vio_iu(iue)->srp.cmd, ibmvstgt_rdma, 1, 1);
> >  
> >  	spin_lock_irqsave(&target->lock, flags);
> > diff --git a/drivers/scsi/libsrp.c b/drivers/scsi/libsrp.c
> > index 732446e..32e1529 100644
> > --- a/drivers/scsi/libsrp.c
> > +++ b/drivers/scsi/libsrp.c
> > @@ -192,18 +192,18 @@ static int srp_direct_data(struct scsi_cmnd *sc, struct srp_direct_buf *md,
> >  
> >  	if (dma_map) {
> >  		iue = (struct iu_entry *) sc->SCp.ptr;
> > -		sg = sc->request_buffer;
> > +		sg = scsi_sglist(sc);
> >  
> > -		dprintk("%p %u %u %d\n", iue, sc->request_bufflen,
> > -			md->len, sc->use_sg);
> > +		dprintk("%p %u %u %d\n", iue, scsi_bufflen(sc),
> > +			md->len, scsi_sg_count(sc));
> >  
> > -		nsg = dma_map_sg(iue->target->dev, sg, sc->use_sg,
> > +		nsg = dma_map_sg(iue->target->dev, sg, scsi_sg_count(sc),
> >  				 DMA_BIDIRECTIONAL);
> >  		if (!nsg) {
> > -			printk("fail to map %p %d\n", iue, sc->use_sg);
> > +			printk("fail to map %p %d\n", iue, scsi_sg_count(sc));
> >  			return 0;
> >  		}
> > -		len = min(sc->request_bufflen, md->len);
> > +		len = min(scsi_bufflen(sc), md->len);
> >  	} else
> >  		len = md->len;
> >  
> > @@ -229,10 +229,10 @@ static int srp_indirect_data(struct scsi_cmnd *sc, struct srp_cmd *cmd,
> >  
> >  	if (dma_map || ext_desc) {
> >  		iue = (struct iu_entry *) sc->SCp.ptr;
> > -		sg = sc->request_buffer;
> > +		sg = scsi_sglist(sc);
> >  
> >  		dprintk("%p %u %u %d %d\n",
> > -			iue, sc->request_bufflen, id->len,
> > +			iue, scsi_bufflen(sc), id->len,
> >  			cmd->data_in_desc_cnt, cmd->data_out_desc_cnt);
> >  	}
> >  
> > @@ -268,13 +268,14 @@ static int srp_indirect_data(struct scsi_cmnd *sc, struct srp_cmd *cmd,
> >  
> >  rdma:
> >  	if (dma_map) {
> > -		nsg = dma_map_sg(iue->target->dev, sg, sc->use_sg, DMA_BIDIRECTIONAL);
> > +		nsg = dma_map_sg(iue->target->dev, sg, scsi_sg_count(sc),
> > +				 DMA_BIDIRECTIONAL);
> >  		if (!nsg) {
> > -			eprintk("fail to map %p %d\n", iue, sc->use_sg);
> > +			eprintk("fail to map %p %d\n", iue, scsi_sg_count(sc));
> >  			err = -EIO;
> >  			goto free_mem;
> >  		}
> > -		len = min(sc->request_bufflen, id->len);
> > +		len = min(scsi_bufflen(sc), id->len);
> >  	} else
> >  		len = id->len;
> >  
> > @@ -425,8 +426,8 @@ int srp_cmd_queue(struct Scsi_Host *shost, struct srp_cmd *cmd, void *info,
> >  
> >  	sc->SCp.ptr = info;
> >  	memcpy(sc->cmnd, cmd->cdb, MAX_COMMAND_SIZE);
> > -	sc->request_bufflen = len;
> > -	sc->request_buffer = (void *) (unsigned long) addr;
> > +	sc->sdb.length = len;
> > +	sc->sdb.sglist = (void *) (unsigned long) addr;
> >  	sc->tag = tag;
> >  	err = scsi_tgt_queue_command(sc, (struct scsi_lun *) &cmd->lun, cmd->tag);
> >  	if (err)
> 
> What is done here in srp_cmd_queue() looks scary to me. even today.
> What is that u64 addr that gets truncated to unsigned long and put
> on sglist? What data-buffer "len" is suppose to describe? And "dir"
> is for what data?

No trancated since it's used for an address. addr, data length, and
data transfer direction though they are not used now.


> It is made to look like addr is a linear pointer with a use_sg==0
> issued command, which is no longer allowed. Only we know it is not,
> because of the "(void *)(unsigned long) addr;" which is a bug in
> 64-bit.
>
> If this is a totally private message sent threw the scsi-ml. I would
> rather it was done with DMA_NONE,bufflen=0,sglist=NULL and put all
> the user-info into scsi_cmnd.SCp like above "void* info".

tgt doesn't queue a command to scsi-ml. It queues it to user space.