[PATCH 2/2] lib: sbi: Make ipi send/clear and access ipi_type non-reentrant

[Xiang W]

在 2023-11-21星期二的 16:56 -0800，Bo Gan写道：
> > > @@ -220,8 +224,12 @@ void sbi_ipi_process(void)
> > >   	u32 hartindex = sbi_hartid_to_hartindex(current_hartid());
> > >   
> > >   	sbi_pmu_ctr_incr_fw(SBI_PMU_FW_IPI_RECVD);
> > > +
> > > +	spin_lock(&ipi_data->lock);
> > >   	sbi_ipi_raw_clear(hartindex);
> > >   	ipi_type = atomic_raw_xchg_ulong(&ipi_data->ipi_type, 0);
> > > +	spin_unlock(&ipi_data->lock);
> > > +
> > >   	ipi_event = 0;
> > >   	while (ipi_type) {
> > >   		if (ipi_type & 1UL) {
> 
> Hi Xiang,
> 
> I think we should drop this, and use Anup's new patch instead. I don't
> think the spinlock here is helping to ensure the ordering between
> 
> sbi_ipi_raw_clear and xchg(). Furthermore, spinlock in OpenSBI doesn't
> order IO, so ipi_raw_clear can leak below spin_unlock. Thus, it won't
> address the race condition in previous discussions.
> 
> With Anip's new patch that uses wmb() in both sender and receiver, it
> fixes the correctness issue, and hopefully has negligible performance
> impact: See the discussion here:
> 
> http://lists.infradead.org/pipermail/opensbi/2023-November/005937.html
> 
> Bo
This bug is not related to reorder. It's one more ipi interrupt is sent.

Regards,
Xiang W