On Sun, Apr 02, 2023 at 12:04:48PM +0200, Benjamin Bara wrote:
On Wed, 29 Mar 2023 at 21:50, Wolfram Sang wsa@kernel.org wrote:
Could you make sure please?
Sure, I'll try. The check before bae1d3a was: in_atomic() || irqs_disabled() which boils down to: (preempt_count() != 0) || irqs_disabled() preemptible() is defined as: (preempt_count() == 0 && !irqs_disabled())
so this patch should behave the same as pre-v5.2, but with the additional system state check. From my point of view, the additional value of the in_atomic() check was that it activated atomic i2c xfers when preemption is disabled, like in the case of panic(). So reverting that commit would also re-activate atomic i2c transfers during emergency restarts. However, I think considering the system state makes sense here.
From my understanding, non-atomic i2c transfers require enabled IRQs, but atomic i2c transfers do not have any "requirements". So the irqs_disabled() check is not here to ensure that the following atomic i2c transfer works correctly, but to use non-atomic i2c xfer as long/often as possible.
Unfortunately, I am not sure yet about !CONFIG_PREEMPTION. I looked into some i2c-bus implementations which implement both, atomic and non-atomic. As far as I saw, the basic difference is that the non-atomic variants usually utilize the DMA and then call a variant of wait_for_completion(), like in i2c_imx_dma_write() [1]. However, the documentation of wait_for_completion [2] states that: "wait_for_completion() and its variants are only safe in process context (as they can sleep) but not (...) [if] preemption is disabled". Therefore, I am not quite sure yet if !CONFIG_PREEMPTION uses the non-atomic variant at all or if this case is handled differently.
Asking Peter Zijlstra might be a good idea. He helped me with the current implementation.
Thanks for the hint! I wrote an extra email to him and added him to CC.
So yeah, can't call schedule() if non preemptible (which is either preempt_disable(), local_bh_disable() (true for bh handlers) or local_irq_disable() (true for IRQ handlers) and mostly rcu_read_lock()).
You can mostly forget about CONFIG_PREEMPT=n (or more specifically CONFIG_PREMPT_COUNT=n) things that work for PREEMPT typically also work for !PREEMPT.
The question here seems to be if i2c_in_atomic_xfer_mode() should have an in_atomic() / !preemptible() check, right? IIUC Wolfram doesn't like it being used outside of extra special cicumstances?