Re: [PATCH 3/4] i2c: i801: Improve i801_block_transaction_byte_by_byte

[Jean Delvare <jdelvare@suse.de>]

Hi Heiner,

On Sun, 27 Aug 2023 19:14:38 +0200, Heiner Kallweit wrote:
> On 27.06.2023 15:46, Jean Delvare wrote:
> > Hi Heiner, Andi,
> > 
> > On Sat, 04 Mar 2023 22:36:34 +0100, Heiner Kallweit wrote:  
> >> Here we don't have to write SMBHSTCNT in each iteration of the loop.
> >> Bit SMBHSTCNT_START is internally cleared immediately, therefore
> >> we don't have to touch the value of SMBHSTCNT until the last byte.
> >>
> >> Signed-off-by: Heiner Kallweit <hkallweit1@gmail.com>
> >> ---
> >>  drivers/i2c/busses/i2c-i801.c | 6 +++---
> >>  1 file changed, 3 insertions(+), 3 deletions(-)
> >>
> >> diff --git a/drivers/i2c/busses/i2c-i801.c b/drivers/i2c/busses/i2c-i801.c
> >> index 7641bd0ac..e1350a8cc 100644
> >> --- a/drivers/i2c/busses/i2c-i801.c
> >> +++ b/drivers/i2c/busses/i2c-i801.c
> >> @@ -677,11 +677,11 @@ static int i801_block_transaction_byte_by_byte(struct i801_priv *priv,
> >>  	for (i = 1; i <= len; i++) {
> >>  		if (i == len && read_write == I2C_SMBUS_READ)
> >>  			smbcmd |= SMBHSTCNT_LAST_BYTE;
> >> -		outb_p(smbcmd, SMBHSTCNT(priv));
> >>  
> >>  		if (i == 1)
> >> -			outb_p(inb(SMBHSTCNT(priv)) | SMBHSTCNT_START,
> >> -			       SMBHSTCNT(priv));
> >> +			outb_p(smbcmd | SMBHSTCNT_START, SMBHSTCNT(priv));
> >> +		else if (smbcmd & SMBHSTCNT_LAST_BYTE)
> >> +			outb_p(smbcmd, SMBHSTCNT(priv));
> >>  
> >>  		status = i801_wait_byte_done(priv);
> >>  		if (status)  
> > 
> > I tested this and it works, but I don't understand how.
> > 
> > I thought that writing to SMBHSTCNT was what was telling the host
> > controller to proceed with the next byte. If writing to SMBHSTCNT for
> > each byte isn't needed, then what causes the next byte to be processed?
> > Does this happen as soon as SMBHSTSTS_BYTE_DONE is written? If so, then
> > what guarantees that we set SMBHSTCNT_LAST_BYTE *before* the last byte
> > is actually processed?
>
> It's my understanding that writing SMBHSTSTS_BYTE_DONE tells the host to
> continue with the next byte.

That's indeed possible, and quite likely, considering that your patch
works.

> We set SMBHSTCNT_LAST_BYTE whilst the host is receiving the last byte.
> Apparently the host checks for SMBHSTCNT_LAST_BYTE once it received
> a byte, in order to determine whether to ack the byte or not.
> So SMBHSTCNT_LAST_BYTE doesn't have to be set before the host starts
> receiving the last byte.

How is this not racy?

In the interrupt-driven case, at the end of a block read transaction,
we set SMBHSTCNT_LAST_BYTE at the end of i801_isr_byte_done(), then
return to i801_isr() where we write 1 to SMBHSTSTS_BYTE_DONE to clear
it. This lets the controller handle the last byte with the knowledge
that this is the last byte.

However, in the poll-driven case, SMBHSTSTS_BYTE_DONE is being cleared
at the end of the loop in i801_block_transaction_byte_by_byte(), then
at the beginning of the next iteration, we write SMBHSTCNT_LAST_BYTE,
then wait for completion. If the controller is super fast (or, to be
more realistic, the i2c-i801 driver gets preempted between writing
SMBHSTSTS_BYTE_DONE and writing SMBHSTCNT_LAST_BYTE) then the byte may
have been already read and acked, before we have the time to let the
controller know that no ACK should be sent. This looks racy. Am I
missing something?

If nothing else, the fact that the order is different between the
interrupt-driven and poll-driven cases is fishy.

I must add that the problem is not related to your patch, I just
happened to notice it while reviewing your patch.

> For writes SMBHSTCNT_LAST_BYTE isn't used.

Agreed.

-- 
Jean Delvare
SUSE L3 Support