Re: [PATCH v4 1/2] i2c: i2c-xiic: fix clearing of bnb irq

From: Thilo Zirnstein

Date: Thu Nov 19 2026 - 04:35:02 EST


On Wed, Nov 18, 2026 at 02:22:18PM +0100, Andi Shyti wrote:

Hi Andi,

Thank you for your review and feedback

> Hi Thilo,
>
> ...
>
> > + ret = read_poll_timeout_atomic(xiic_getreg8, sr,
> > + sr & XIIC_SR_BUS_BUSY_MASK, 0, XIIC_BB_TIMEOUT_US, 0,
> > + i2c, XIIC_SR_REG_OFFSET);
>
> this is a bit of an overkill: you are forcing tight loops for
> 1ms. If you consider that the minimum wait is 1ns, in 1ms you
> will end up busy looping a lot of time. Please, don't set the
> delay_us to '0', but calculate a proper waiting time between
> reads.
>

I also had doubts about whether this approach is too heavy.
I have general concerns about the busy-waiting in atomic context
in this patches, though I noticed other I2C adapters wait
for relatively long periods there as well.
The tight loops in patch 2/2 are even longer.
Those waits are longer because the code waits for the entire I2C transfer
to finish, but polling is unavoidable there since sleeping or scheduling
deferred work is not possible in atomic context—is that right?
So this kind of builtin if doing i2c transfer in atomic context ?

I will reduce the maximum poll timeout to 4 µs,
which matches the minimum I2C START condition hold time
in standard mode and seems reasonable.

I will also increase delay_us. Out of curiosity,
what is the reasoning behind the delay interval?
I assume we want to avoid hammering shared buses
or registers with continuous I/O reads.

> Andi

kind regards,

Thilo