Date:   Fri, 25 Jan 2019 14:28:39 -0800
From:   Florian Fainelli <f.fainelli@...il.com>
To:     René van Dorst <opensource@...rst.com>
Cc:     Andrew Lunn <andrew@...n.ch>,
        Heiner Kallweit <hkallweit1@...il.com>,
        "David S. Miller" <davem@...emloft.net>, netdev@...r.kernel.org
Subject: Re: [PATCH] sfp: sfp_read: split-up request when hw rx buffer is too
 small.

On 1/25/19 1:58 PM, René van Dorst wrote:
> Quoting René van Dorst <opensource@...rst.com>:
> 
>> Quoting Florian Fainelli <f.fainelli@...il.com>:
>>
>>> On 1/23/19 1:20 PM, René van Dorst wrote:
>>>> Without this patch sfp code retries to read the full struct
>>>> sfp_eeprom_id
>>>> id out of the SFP eeprom. Sizeof(id) is 96 bytes.
>>>> My i2c hardware, Mediatek mt7621, has a rx buffer of 64 bytes.
>>>> So sfp_read gets -NOSUPPORTED back on his turn return -EAGAIN.
>>>> Same issue is with the SFP_EXT_STATUS data which is 92 bytes.
>>>>
>>>> By split-up the request in multiple smaller requests with a max size
>>>> of i2c
>>>> max_read_len, we can readout the SFP module successfully.
>>>>
>>>> Tested with MT7621 and two Fiberstore modules SFP-GB-GE-T and
>>>> SFP-GE-BX.
>>>>
>>>> Signed-off-by: René van Dorst <opensource@...rst.com>
>>>> ---
>>>> drivers/net/phy/sfp.c | 23 ++++++++++++++++++++++-
>>>> 1 file changed, 22 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/drivers/net/phy/sfp.c b/drivers/net/phy/sfp.c
>>>> index fd8bb998ae52..1352a19571cd 100644
>>>> --- a/drivers/net/phy/sfp.c
>>>> +++ b/drivers/net/phy/sfp.c
>>>> @@ -367,7 +367,28 @@ static void sfp_set_state(struct sfp *sfp,
>>>> unsigned int state)
>>>>
>>>> static int sfp_read(struct sfp *sfp, bool a2, u8 addr, void *buf,
>>>> size_t len)
>>>> {
>>>> -    return sfp->read(sfp, a2, addr, buf, len);
>>>> +    const struct i2c_adapter_quirks *q = sfp->i2c->quirks;
>>>> +    int ret;
>>>> +    size_t rx_bytes = 0;
>>>> +
>>>> +    /* Many i2c hw have limited rx buffers, split-up request when
>>>> needed. */
>>>> +    while ((q->max_read_len) && (len > q->max_read_len)) {
>>>> +        ret = sfp->read(sfp, a2, addr, buf, q->max_read_len);
>>>> +        if (ret < 0)
>>>> +            return ret;
>>>> +        rx_bytes += ret;
>>>> +        addr += q->max_read_len;
>>>> +        buf += q->max_read_len;
>>>> +        len -= q->max_read_len;
>>>> +    }
>>>
>>> sfp->read defaults to sfp_i2c_read() but it is possible to override that
>>> by registering a custom sfp_bus instance (nothing does it today, but
>>> that could obviously change), so there is no guarantee that
>>> sfp->i2c->quirks is applicable unless sfp_i2c_read() is used.
>>>
>>> sfp_i2c_read() is presumably where the max_read_len splitting should
>>> occur, or better yet, should not i2c_transfer() take care of that on its
>>> own? That way there would be no layering violation of having to fetch
>>> the quirk bitmask for the i2c adapter being used, that is something that
>>> should belong in the core i2c framework.
>>
>> Yes it is better to put it in sfp_i2c_read().
>>
>> I think it is best to handle the split-up within the driver.
>> The driver knows how to talk to the device and may apply device quirks.
>>
>> Also tda1004x [0] and TPM [1] driver also handles it within the driver
>> itself.
>>
>> TPM driver just try to send the want size and split-up request to
>> I2C_SMBUS_BLOCK_MAX when a -EOPNOTSUPP returns, just retries it a
>> number of
>> times.
>>
>> I can do the same but I have to pick a minimum size.
>> Looking in SSF-8472rev12.2.1 they don't limit the way you access the
>> device.
>> So use I2C_SMBUS_BLOCK_MAX of 32 bytes is sufficient or lookup the i2c
>> quirk max_read_len is also an option.
> 
> Hi Florian,
> 
> I did a test in sfp_i2c_read() with reducing the read size when -EOPNOTSUPP
> returns like TPM driver. But this always produces noise in kernel log about
> the "msg too long" if the device is init at boot or when plugged-in.
> Devices with SFP_EXT_STATUS generates also an 2nd error.
> So I prefer to use the quirk max_read_len.

My point still stands, we have an abstraction layer through the i2c core
which sits between clients and adapters.

If you have a simple read into a buffer transaction (like what we have
here), then you can provide a safe accessor function that takes care of
looking at the max_read_len quirk and splitting things into the
appropriate number of transaction. Then we can also eliminate the "msg
too long" annoying message since we are now using an appropriate function.

The fact that there are multiple drivers that need to be patched to
split i2c transaction is an indication that this is not IMHO addressed
at the right level, especially if what they do is as simple as an
i2c_transfer() into a single buffer and require no quirky transaction.

Andrew, Heiner and Russell might have a different view on this.

> 
> dmesg output reduce read size when -EOPNOTSUPP returns.
> [  134.220000] i2c i2c-0: adapter quirk: msg too long (addr 0x0050, size
> 96, read)
> [  135.250000] sfp sfp-lan5: module FiberStore       SFP-GB-GE-T     
> rev B    sn <snip>      dc <snip>
> [  141.150000] sfp sfp-lan5: module removed
> [  150.950000] i2c i2c-0: adapter quirk: msg too long (addr 0x0050, size
> 96, read)
> [  151.980000] sfp sfp-lan5: module FiberStore       SFP-GE-BX       
> rev A0   sn <snip>      dc <snip>
> [  151.990000] i2c i2c-0: adapter quirk: msg too long (addr 0x0051, size
> 92, read)
> 
> dmesg output with read size always reduced to quiek max_read_len.
> [   27.350000] sfp sfp-lan5: module FiberStore       SFP-GE-BX       
> rev A0   sn <snip>      dc <snip>
> [   34.540000] sfp sfp-lan5: module removed
> [   39.360000] sfp sfp-lan5: module FiberStore       SFP-GB-GE-T     
> rev B    sn <snip>      dc <snip>
> 
> So shall I spin v2 with quirk max_read_len as max read size?


-- 
Florian

Powered by blists - more mailing lists