Virtualization - Jul 2021 - [PATCH v12] i2c: virtio: add a virtio i2c frontend driver

On 2021/7/5 10:40, Viresh Kumar wrote:
> I think we missed the first deadline for 5.14-rc1 as Wolfram should be out
of
> office now. Anyway lets make sure we fix all the pending bits before he is
back
> and see if we can still pull it off by rc2.
>
> On 02-07-21, 16:46, Jie Deng wrote:
>> diff --git a/drivers/i2c/busses/i2c-virtio.c
b/drivers/i2c/busses/i2c-virtio.c
>> +static int virtio_i2c_send_reqs(struct virtqueue *vq,
> It would be better to rename this to virtio_i2c_prepare_reqs instead, as
this
> doesn't send anything.

That's a better name. I'm OK with that.

>
>> +				struct virtio_i2c_req *reqs,
>> +				struct i2c_msg *msgs, int nr)
>> +{
>> +	struct scatterlist *sgs[3], out_hdr, msg_buf, in_hdr;
>> +	int i, outcnt, incnt, err = 0;
>> +
>> +	for (i = 0; i < nr; i++) {
>> +		/*
>> +		 * Only 7-bit mode supported for this moment. For the address
format,
>> +		 * Please check the Virtio I2C Specification.
>> +		 */
>> +		reqs[i].out_hdr.addr = cpu_to_le16(msgs[i].addr << 1);
>> +
>> +		if (i != nr - 1)
>> +			reqs[i].out_hdr.flags = cpu_to_le32(VIRTIO_I2C_FLAGS_FAIL_NEXT);
>> +
>> +		outcnt = incnt = 0;
>> +		sg_init_one(&out_hdr, &reqs[i].out_hdr,
sizeof(reqs[i].out_hdr));
>> +		sgs[outcnt++] = &out_hdr;
>> +
>> +		if (msgs[i].len) {
>> +			reqs[i].buf = i2c_get_dma_safe_msg_buf(&msgs[i], 1);
>> +			if (!reqs[i].buf)
>> +				break;
>> +
>> +			sg_init_one(&msg_buf, reqs[i].buf, msgs[i].len);
>> +
>> +			if (msgs[i].flags & I2C_M_RD)
>> +				sgs[outcnt + incnt++] = &msg_buf;
>> +			else
>> +				sgs[outcnt++] = &msg_buf;
>> +		}
>> +
>> +		sg_init_one(&in_hdr, &reqs[i].in_hdr,
sizeof(reqs[i].in_hdr));
>> +		sgs[outcnt + incnt++] = &in_hdr;
>> +
>> +		err = virtqueue_add_sgs(vq, sgs, outcnt, incnt, &reqs[i],
GFP_KERNEL);
>> +		if (err < 0) {
>> +			i2c_put_dma_safe_msg_buf(reqs[i].buf, &msgs[i], false);
>> +			break;
>> +		}
>> +	}
>> +
>> +	return i;
>> +}
> The right way of doing this is is making this function return - Error on
failure
> and 0 on success. There is no point returning number of successful
additions
> here.

We need the number for virtio_i2c_complete_reqs to do cleanup. We don't 
have to

do cleanup "num" times every time. Just do it as needed.

>
> Moreover, on failures this needs to clean up (free the dmabufs) itself,
just
> like you did i2c_put_dma_safe_msg_buf() at the end. The caller
shouldn't be
> required to handle the error cases by freeing up resources.

This function will return the number of requests being successfully 
prepared and make sure

resources of the failed request being freed. And 
virtio_i2c_complete_reqs will free the

resources of those successful request.

>> +static int virtio_i2c_complete_reqs(struct virtqueue *vq,
>> +				    struct virtio_i2c_req *reqs,
>> +				    struct i2c_msg *msgs, int nr,
>> +				    bool fail)
>> +{
>> +	struct virtio_i2c_req *req;
>> +	bool failed = fail;
>> +	unsigned int len;
>> +	int i, j = 0;
>> +
>> +	for (i = 0; i < nr; i++) {
>> +		/* Detach the ith request from the vq */
>> +		req = virtqueue_get_buf(vq, &len);
>> +
>> +		/*
>> +		 * Condition (req && req == &reqs[i]) should always meet
since
>> +		 * we have total nr requests in the vq.
>> +		 */
>> +		if (!failed && (WARN_ON(!(req && req ==
&reqs[i])) ||
>> +		    (req->in_hdr.status != VIRTIO_I2C_MSG_OK)))
> What about writing this as:
>
> 		if (!failed && (WARN_ON(req != &reqs[i]) ||
> 		    (req->in_hdr.status != VIRTIO_I2C_MSG_OK)))
>
> We don't need to check req here since if req is NULL, we will not do
req->in_hdr
> at all.

It's right here just because the &reqs[i] will never be NULL in our 
case. But if you see

"virtio_i2c_complete_reqs" as an independent function, you need to
check the

req. From the perspective of the callee, you can't ask the caller always 
give you

the non-NULL parameters. And some tools may give warnings.

>> +			failed = true;
>> +
>> +		i2c_put_dma_safe_msg_buf(reqs[i].buf, &msgs[i], !failed);
>> +		if (!failed)
>> +			++j;
>> +	}
>> +
>> +	return (fail ? -ETIMEDOUT : j);
>> +}
>> +

On 05-07-21, 11:45, Jie Deng wrote:
> On 2021/7/5 10:40, Viresh Kumar wrote:
> > On 02-07-21, 16:46, Jie Deng wrote:
> > The right way of doing this is is making this function return - Error
on failure
> > and 0 on success. There is no point returning number of successful
additions
> > here.
> 
> 
> We need the number for virtio_i2c_complete_reqs to do cleanup. We don't
have
> to
> 
> do cleanup "num" times every time. Just do it as needed.
If you do full cleanup here, then you won't required that at the caller
site.
> > Moreover, on failures this needs to clean up (free the dmabufs)
itself, just
> > like you did i2c_put_dma_safe_msg_buf() at the end. The caller
shouldn't be
> > required to handle the error cases by freeing up resources.
> 
> 
> This function will return the number of requests being successfully
prepared
> and make sure
> 
> resources of the failed request being freed. And virtio_i2c_complete_reqs
> will free the
> 
> resources of those successful request.
It just looks cleaner to give such responsibility to each and every function,
i.e. if they fail, they should clean stuff up instead of the caller. That's
the
normal philosophy you will find across kernel in most of the cases.
 
> > > +static int virtio_i2c_complete_reqs(struct virtqueue *vq,
> > > +				    struct virtio_i2c_req *reqs,
> > > +				    struct i2c_msg *msgs, int nr,
> > > +				    bool fail)
> > > +{
> > > +	struct virtio_i2c_req *req;
> > > +	bool failed = fail;
> > > +	unsigned int len;
> > > +	int i, j = 0;
> > > +
> > > +	for (i = 0; i < nr; i++) {
> > > +		/* Detach the ith request from the vq */
> > > +		req = virtqueue_get_buf(vq, &len);
> > > +
> > > +		/*
> > > +		 * Condition (req && req == &reqs[i]) should
always meet since
> > > +		 * we have total nr requests in the vq.
> > > +		 */
> > > +		if (!failed && (WARN_ON(!(req && req ==
&reqs[i])) ||
> > > +		    (req->in_hdr.status != VIRTIO_I2C_MSG_OK)))
> > What about writing this as:
> > 
> > 		if (!failed && (WARN_ON(req != &reqs[i]) ||
> > 		    (req->in_hdr.status != VIRTIO_I2C_MSG_OK)))
> > 
> > We don't need to check req here since if req is NULL, we will not
do req->in_hdr
> > at all.
> 
> 
> It's right here just because the &reqs[i] will never be NULL in our
case.
> But if you see
> 
> "virtio_i2c_complete_reqs" as an independent function, you need
to check the
> 
> req. From the perspective of the callee, you can't ask the caller
always
> give you
> 
> the non-NULL parameters.
We need to keep this driver optimized in its current form. If you see your own
argument here, then why don't you test vq or msgs for a valid pointer ? And
even
reqs.

If we know for certain that this will never happen, then it should be optimized.
But if you see a case where reqs[i] can be NULL here, then it would be fine.
> And some tools may give warnings.
I don't see why a tool will raise an error here and if it does, then the
tool is
buggy and not the driver. And we don't need to take care of that.

-- 
viresh