Re: [PATCH 08/10] ipmi: kcs_bmc: Track clients in core

[Andrew Jeffery]

On Fri, 2023-11-03 at 15:05 +0000, Jonathan Cameron wrote:
> On Fri,  3 Nov 2023 16:45:20 +1030
> Andrew Jeffery <andrew@xxxxxxxxxxxxxxxxxxxx> wrote:
> 
> > I ran out of spoons before I could come up with a better client tracking
> > scheme back in the original refactoring series:
> > 
> > https://lore.kernel.org/all/20210608104757.582199-1-andrew@xxxxxxxx/
> > 
> > Jonathan prodded Konstantin about the issue in a review of Konstantin's
> > MCTP patches[1], prompting an attempt to clean it up.
> > 
> > [1]: https://lore.kernel.org/all/20230929120835.0000108e@xxxxxxxxxx/
> > 
> > Prevent client modules from having to track their own instances by
> > requiring they return a pointer to a client object from their
> > add_device() implementation. We can then track this in the core, and
> > provide it as the argument to the remove_device() implementation to save
> > the client module from further work. The usual container_of() pattern
> > gets the client module access to its private data.
> > 
> > Signed-off-by: Andrew Jeffery <andrew@xxxxxxxxxxxxxxxxxxxx>
> 
> Hi Andrew,
> 
> A few comments inline.
> More generally, whilst this is definitely an improvement I'd have been tempted
> to make more use of the linux device model for this with the clients added
> as devices with a parent of the kcs_bmc_device.  That would then allow for
> simple dependency tracking, binding of individual drivers and all that.
> 
> What you have here feels fine though and is a much less invasive change.

Yeah, I had this debate with myself before posting the patches. My
reasoning for the current approach is that the clients don't typically
represent a device, rather a protocol implementation that is
communicated over a KCS device (maybe more like pairing a line
discipline with a UART). It was unclear to me whether associating a
`struct device` with a protocol implementation was stretching the
abstraction a bit, or whether I haven't considered some other
perspective hard enough - maybe we treat the client as the remote
device, similar to e.g. a `struct i2c_client`?

> 
> Jonathan
> 
> 
> > diff --git a/drivers/char/ipmi/kcs_bmc_cdev_ipmi.c b/drivers/char/ipmi/kcs_bmc_cdev_ipmi.c
> > index 98f231f24c26..9fca31f8c7c2 100644
> > --- a/drivers/char/ipmi/kcs_bmc_cdev_ipmi.c
> > +++ b/drivers/char/ipmi/kcs_bmc_cdev_ipmi.c
> > @@ -71,8 +71,6 @@ enum kcs_ipmi_errors {
> 
> 
> 
> > +static struct kcs_bmc_client *
> > +kcs_bmc_ipmi_add_device(struct kcs_bmc_driver *drv, struct kcs_bmc_device *dev)
> >  {
> >  	struct kcs_bmc_ipmi *priv;
> >  	int rc;
> >  
> >  	priv = kzalloc(sizeof(*priv), GFP_KERNEL);
> >  	if (!priv)
> > -		return -ENOMEM;
> > +		return ERR_PTR(ENOMEM);
> As below. I thought it took negatives..

I should have double checked that. It requires negatives. Thanks.

> >  
> >  	spin_lock_init(&priv->lock);
> >  	mutex_init(&priv->mutex);
> >  	init_waitqueue_head(&priv->queue);
> >  
> > -	priv->client.dev = kcs_bmc;
> > -	priv->client.ops = &kcs_bmc_ipmi_client_ops;
> > +	kcs_bmc_client_init(&priv->client, &kcs_bmc_ipmi_client_ops, drv, dev);
> >  
> >  	priv->miscdev.minor = MISC_DYNAMIC_MINOR;
> > -	priv->miscdev.name = kasprintf(GFP_KERNEL, "%s%u", DEVICE_NAME, kcs_bmc->channel);
> > +	priv->miscdev.name = kasprintf(GFP_KERNEL, "%s%u", DEVICE_NAME, dev->channel);
> >  	if (!priv->miscdev.name) {
> >  		rc = -ENOMEM;
> ERR_PTR

I converted it to an ERR_PTR in the return after the cleanup_priv
label. Maybe it's preferable I do the conversion immediately? Easy
enough to change if you think so.

> >  		goto cleanup_priv;
> 
> 
> 
> ...
> 
> > diff --git a/drivers/char/ipmi/kcs_bmc_serio.c b/drivers/char/ipmi/kcs_bmc_serio.c
> > index 0a68c76da955..3cfda39506f6 100644
> > --- a/drivers/char/ipmi/kcs_bmc_serio.c
> > +++ b/drivers/char/ipmi/kcs_bmc_serio.c
> 
> ...
> 
> 
> > +static struct kcs_bmc_client *
> > +kcs_bmc_serio_add_device(struct kcs_bmc_driver *drv, struct kcs_bmc_device *dev)
> >  {
> >  	struct kcs_bmc_serio *priv;
> >  	struct serio *port;
> > @@ -75,12 +71,12 @@ static int kcs_bmc_serio_add_device(struct kcs_bmc_device *kcs_bmc)
> >  
> >  	priv = kzalloc(sizeof(*priv), GFP_KERNEL);
> >  	if (!priv)
> > -		return -ENOMEM;
> > +		return ERR_PTR(ENOMEM);
> >  
> >  	/* Use kzalloc() as the allocation is cleaned up with kfree() via serio_unregister_port() */
> >  	port = kzalloc(sizeof(*port), GFP_KERNEL);
> >  	if (!port) {
> > -		rc = -ENOMEM;
> > +		rc = ENOMEM;
> Why positive?
> Doesn't ERR_PTR() typically get passed negatives? 

Ack, as above.

Thanks for the review,

Andrew