Re: [PATCH 3/4] Intel PTI implementaiton of MIPI 1149.7.

[J Freyensee <james_p_freyensee@linux.intel.com>]

On Sun, 2011-04-24 at 02:55 +0200, Jesper Juhl wrote:
> On Fri, 22 Apr 2011, james_p_freyensee@linux.intel.com wrote:
> 
> > From: J Freyensee <james_p_freyensee@linux.intel.com>
> > 
> > The PTI (Parallel Trace Interface) driver directs
> > trace data routed from various parts in the system out
> > through an Intel Penwell PTI port and out of the mobile
> > device for analysis with a debugging tool (Lauterbach or Fido).
> > Though n_tracesink and n_tracerouter line discipline drivers
> > are used to extract modem tracing data to the PTI driver
> > and other parts of an Intel mobile solution, the PTI driver
> > can be used independent of n_tracesink and n_tracerouter.
> > 
> > You should select this driver if the target kernel is meant for
> > an Intel Atom (non-netbook) mobile device containing a MIPI
> > P1149.7 standard implementation.
> > 
> > Signed-off-by: J Freyensee <james_p_freyensee@linux.intel.com>
> 
> A few comments below.
> 
> ...
> > +#define DRIVERNAME		"pti"
> > +#define PCINAME			"pciPTI"
> > +#define TTYNAME			"ttyPTI"
> > +#define CHARNAME		"pti"
> > +#define PTITTY_MINOR_START	0
> > +#define PTITTY_MINOR_NUM	2
> > +#define MAX_APP_IDS		16   /* 128 channel ids / u8 bit size */
> > +#define MAX_OS_IDS		16   /* 128 channel ids / u8 bit size */
> > +#define MAX_MODEM_IDS		16   /* 128 channel ids / u8 bit size */
> > +#define MODEM_BASE_ID		71   /* modem master ID address    */
> ...
> 
> Would be nice if the values of these defines would line up nicely.
> 
> 
> ...
> > +static struct pci_device_id pci_ids[] __devinitconst = {
> > +		{ PCI_DEVICE(PCI_VENDOR_ID_INTEL, 0x82B) },
> > +		{0}
> ...
> 
> Why are there spaces after the opening { and before the closing } for the 
> first entry, but not the second. Looks like you need to pick a 
> consistent style.
> 
> 
> > + *  regroup the appropriate message segments together reconstituting each
> > + *  message.
> > + */
> > +static void pti_write_to_aperture(struct pti_masterchannel *mc,
> > +				  u8 *buf,
> > +				  int len)
> > +{
> > +	int dwordcnt, final, i;
> > +	u32 ptiword;
> > +	u8 *p;
> > +	u32 __iomem *aperture;
> > +
> > +	p = buf;
> ...
> 
> Perhaps save a few lines by doing
> 
> static void pti_write_to_aperture(struct pti_masterchannel *mc,
>                                u8 *buf,
>                                int len)
> {
>      int dwordcnt, final, i;
>      u32 ptiword;
>      u32 __iomem *aperture;
>      u8 *p = buf;
> 
> 

I can make the tweak.

> ...
> > +void pti_writedata(struct pti_masterchannel *mc, u8 *buf, int count)
> > +{
> > +	/*
> > +	 * since this function is exported, this is treated like an
> > +	 * API function, thus, all parameters should
> > +	 * be checked for validity.
> > +	 */
> > +	if ((mc != NULL) && (buf != NULL) && (count > 0))
> > +		pti_write_to_aperture(mc, buf, count);
> > +	return;
> ...
> 
> Pointless return; statement.
> 
> 
> ...
> > +static void __devexit pti_pci_remove(struct pci_dev *pdev)
> > +{
> > +	struct pti_dev *drv_data;
> > +
> > +	drv_data = pci_get_drvdata(pdev);
> > +	if (drv_data != NULL) {
> 
> Perhaps
> 
> static void __devexit pti_pci_remove(struct pci_dev *pdev)
> {
>      struct pti_dev *drv_data = pci_get_drvdata(pdev);
>      
>      if (drv_data) {
> 
> 

I'd rather keep my way.  Just easier to read and more self-explanatory.
I realize everyone on this list are expert programmers, but I tend to
default to code that is dead-simple to read and understand.

> ...
> > +static int pti_tty_driver_open(struct tty_struct *tty, struct file *filp)
> > +{
> > +	int ret = 0;
> > +
> > +	/*
> > +	 * we actually want to allocate a new channel per open, per
> > +	 * system arch.  HW gives more than plenty channels for a single
> > +	 * system task to have its own channel to write trace data. This
> > +	 * also removes a locking requirement for the actual write
> > +	 * procedure.
> > +	 */
> > +	ret = tty_port_open(&drv_data->port, tty, filp);
> > +
> > +	return ret;
> > +}
> ...
> 
> Why not get rid of the pointless 'ret' variable and simplify this down to
> 
> static int pti_tty_driver_open(struct tty_struct *tty, struct file *filp)
> {     
>      /*
>       * we actually want to allocate a new channel per open, per
>       * system arch.  HW gives more than plenty channels for a single
>       * system task to have its own channel to write trace data. This
>       * also removes a locking requirement for the actual write
>       * procedure.
>       */
>      return tty_port_open(&drv_data->port, tty, filp);
> }
> 
> ??
> 

Sure, I can change it.

> 
> ...
> > +static void pti_tty_driver_close(struct tty_struct *tty, struct file *filp)
> > +{
> > +	tty_port_close(&drv_data->port, tty, filp);
> > +
> > +	return;
> > +}
> 
> Just kill that superfluous return statement.
> 
> 
Dido here too.

> ...
> > +static void pti_tty_cleanup(struct tty_struct *tty)
> > +{
> > +	struct pti_tty *pti_tty_data;
> > +	struct pti_masterchannel *mc;
> > +
> > +	pti_tty_data = tty->driver_data;
> > +
> > +	if (pti_tty_data != NULL) {
> > +		mc = pti_tty_data->mc;
> > +		pti_release_masterchannel(mc);
> > +		pti_tty_data->mc = NULL;
> > +	}
> > +
> > +	if (pti_tty_data != NULL)
> > +		kfree(pti_tty_data);
> > +
> > +	tty->driver_data = NULL;
> > +}
> 
> How about this instead?
> 
> static void pti_tty_cleanup(struct tty_struct *tty)
> {
>      if (!tty->driver_data)
>              return;
>      pti_release_masterchannel(tty->driver_data->mc);
>      kfree(tty->driver_data);
> }
> 

I think I answered this already; I like the suggestion and will tweak.

> ...
> > +static int pti_tty_driver_write(struct tty_struct *tty,
> > +	const unsigned char *buf, int len)
> > +{
> > +	struct pti_masterchannel *mc;
> > +	struct pti_tty *pti_tty_data;
> > +
> > +	pti_tty_data = tty->driver_data;
> > +	mc = pti_tty_data->mc;
> > +	pti_write_to_aperture(mc, (u8 *)buf, len);
> > +
> > +	return len;
> > +}
> 
> I'd like to suggest this as an alternative:
> 
> static int pti_tty_driver_write(struct tty_struct *tty,
>      const unsigned char *buf, int len)
> {
>      pti_write_to_aperture(tty->driver_data->mc, (u8 *)buf, len);
>      return len;
> }
> 
> 

If there is no objections I will do it.  What I've coded is the observed
coding style I've seen, if for no other reason that to shorten up the
number of '->' used in accessing a member of driver_data.  But this
doesn't look so bad/ugly.

> ..
> > +static int pti_char_open(struct inode *inode, struct file *filp)
> > +{
> > +	struct pti_masterchannel *mc;
> > +
> > +	mc = pti_request_masterchannel(0);
> > +	if (mc == NULL)
> > +		return -ENOMEM;
> > +	filp->private_data = mc;
> > +	return 0;
> > +}
> 
> Ok, so I admit that I haven't looked to check if it's actually important 
> that filp->private_data does not get overwritten if 
> pti_request_masterchannel() returns NULL, but if we assume that it is not 
> important, then this would be an improvement IMHO:
> 
> static int pti_char_open(struct inode *inode, struct file *filp)
> {
>      filp->private_data = pti_request_masterchannel(0);
>      if (!filp->private_data)
>              return -ENOMEM; 
>      return 0;
> }
> 
> 

I'll play with this with a debugging tool, but I may want to leave the
code the way I have it.

> ...
> > +
> > +/**
> > + * pti_char_release()-  Close a char channel to the PTI device. Part
> > + * of the misc device implementation.
> > + *
> > + * @inode: Not used in this implementaiton.
> > + * @filp:  Contains private_data that contains the master, channel
> > + *         ID to be released by the PTI device.
> > + *
> > + * Returns:
> > + *	always 0
> 
> Why not void then?

Because the prototype for struct file_definitions calls for returning an
int value.

> 
> 
> > +	pti_release_masterchannel(filp->private_data);
> > +	return 0;
> > +}
> > +
> > +/**
> > + * pti_char_write()-  Write trace debugging data through the char
> > + * interface to the PTI HW.  Part of the misc device implementation.
> > + *
> > + * @filp:  Contains private data which is used to obtain
> > + *         master, channel write ID.
> > + * @data:  trace data to be written.
> > + * @len:   # of byte to write.
> > + * @ppose: Not used in this function implementation.
> > + *
> > + * Returns:
> > + *	int, # of bytes written
> > + *	otherwise, error value
> > + *
> > + * Notes: From side discussions with Alan Cox and experimenting
> > + * with PTI debug HW like Nokia's Fido box and Lauterbach
> > + * devices, 8192 byte write buffer used by USER_COPY_SIZE was
> > + * deemed an appropriate size for this type of usage with
> > + * debugging HW.
> > + */
> > +static ssize_t pti_char_write(struct file *filp, const char __user *data,
> > +			      size_t len, loff_t *ppose)
> > +{
> > +	struct pti_masterchannel *mc;
> > +	void *kbuf;
> > +	const char __user *tmp;
> > +	size_t size = USER_COPY_SIZE, n = 0;
> 
> It would be nice to declare these two variables on two sepperate lines 
> IMO.

K, I can fix.

> 
> > +
> > +	tmp = data;
> > +	mc = filp->private_data;
> > +
> > +	kbuf = kmalloc(size, GFP_KERNEL);
> > +	if (kbuf == NULL)  {
> > +		pr_err("%s(%d): buf allocation failed\n",
> > +			__func__, __LINE__);
> > +		return 0;
> 
> Shouldn't you be returning -ENOMEM here?
> 
> > +	}
> > +
> > +	do {
> > +		if (len - n > USER_COPY_SIZE)
> > +			size = USER_COPY_SIZE;
> > +		else
> > +			size = len - n;
> > +
> > +		if (copy_from_user(kbuf, tmp, size)) {
> > +			kfree(kbuf);
> > +			return n ? n : -EFAULT;
> > +		}
> > +
> > +		pti_write_to_aperture(mc, kbuf, size);
> > +		n  += size;
> > +		tmp += size;
> > +
> > +	} while (len > n);
> > +
> > +	kfree(kbuf);
> > +	kbuf = NULL;
> > +
> 
>  kbuff is a local variable. What's the point in assigning NULL to it just 
> before you return? Just get rid of that silly assignment.

I err on the side of paranoia and default to attempting to use good
programming practices and rather receiving comments like this, than the
alternative where I should have assigned something to NULL/0 and I
introduce a security flaw in the driver/kernel.

> 
> 
> ...
> > + * pti_char_release()-  Close a char channel to the PTI device. Part
> > + * of the misc device implementation.
> > + *
> > + * @inode: Not used in this implementaiton.
> > + * @filp:  Contains private_data that contains the master, channel
> > + *         ID to be released by the PTI device.
> > + *
> > + * Returns:
> > + *	always 0
> 
> So why not void?

Same reason; struct file_operations calls for the function prototype
returning an int.

> 
> ...
> > + * pti_console_setup()-  Initialize console variables used by the driver.
> > + *
> > + * @c:     Not used.
> > + * @opts:  Not used.
> > + *
> > + * Returns:
> > + *	always 0.
> 
> Why not void?

Same reason; the kernel driver prototype function calls for an int
return value.

> 
> 
> ...
> > + * pti_port_activate()- Used to start/initialize any items upon
> > + * first opening of tty_port().
> > + *
> > + * @port- The tty port number of the PTI device.
> > + * @tty-  The tty struct associated with this device.
> > + *
> > + * Returns:
> > + *	always returns 0
> 
> Shouldn't it just return void then?

Same reason as above.

> 
>