On Wed, Feb 21, 2024 at 5:29 PM Björn Töpel <bjorn@xxxxxxxxxx> wrote:
>
> Anup Patel <apatel@xxxxxxxxxxxxxxxx> writes:
>
> >> > +void imsic_vector_mask(struct imsic_vector *vec)
> >> > +{
> >> > + struct imsic_local_priv *lpriv;
> >> > +
> >> > + lpriv = per_cpu_ptr(imsic->lpriv, vec->cpu);
> >> > + if (WARN_ON(&lpriv->vectors[vec->local_id] != vec))
> >> > + return;
> >> > +
> >> > + /*
> >> > + * This function is called through Linux irq subsystem with
> >> > + * irqs disabled so no need to save/restore irq flags.
> >> > + */
> >> > +
> >> > + raw_spin_lock(&lpriv->lock);
> >> > +
> >> > + vec->enable = false;
> >> > + bitmap_set(lpriv->dirty_bitmap, vec->local_id, 1);
> >> > + __imsic_remote_sync(lpriv, vec->cpu);
> >> > +
> >> > + raw_spin_unlock(&lpriv->lock);
> >> > +}
> >>
> >> Really nice that you're using a timer for the vector affinity change,
> >> and got rid of the special/weird IMSIC/sync IPI. Can you really use a
> >> timer for mask/unmask? That makes the mask/unmask operation
> >> asynchronous!
> >>
> >> That was what I was trying to get though with this comment:
> >> https://lore.kernel.org/linux-riscv/87sf24mo1g.fsf@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxx/
> >>
> >> Also, using the smp_* IPI functions, you can pass arguments, so you
> >> don't need the dirty_bitmap tracking the changes.
> >
> > The mask/unmask operations are called with irqs disabled so if
> > CPU X does synchronous IPI to another CPU Y from mask/unmask
> > operation then while CPU X is waiting for IPI to complete it cannot
> > receive IPI from other CPUs which can lead to crashes and stalls.
> >
> > In general, we should not do any block/busy-wait work in
> > mask/unmask operation of an irqchip driver.
>
> Hmm, OK. Still, a bit odd that when the .irq_mask callback return, the
> masking is not actually completed.
>
> 1. CPU 0 tries to mask an interrupt tried to CPU 1.
> 2. The timer is queued on CPU 1.
> 3. The call irq_mask returns on CPU 0
> 4. ...the irq is masked at some future point, determined by the callback
> at CPU 1
>
> Is that the expected outcome?
Yes, that's right.
>
> There are .irq_mask implementation that does seem to go at length
> (blocking) to perform the mask, e.g.: gic_mask_irq() which calls
> gic_{re,}dist_wait_for_rwp that have sleep/retry loops. The GIC3 ITS
> code has similar things going on.
The gic_{re,}dist_wait_for_rwp() polls on a HW register for completion
which will certainly complete in a predictable time whereas waiting
for IPI to be executed by another CPU is not predictable and fragile.
>
> I'm not saying you're wrong, I'm just trying to wrap my head around the
> masking semantics.
>
> > The AIA IMSIC spec allows setting ID pending bit using MSI write
> > irrespective whether ID is enabled or not but the interrupt will be
> > taken only after ID is enabled. In other words, there will be no
> > loss of interrupt with delayed mask/unmask using async IPI or
> > lazy timer.
>
> No loss, but we might *get* an interrupt when we explicitly asked not to
> get any. Maybe that's ok?
>
The delayed spurious interrupt after masking is avoided by additional
masking at the source of interrupt. For wired-to-MSI interrupts, we have
additional masking on the APLIC MSI-mode. For PCI MSI interrupts, we
have additional masking at PCI device level using pci_msi_mask_irq().
Regards,
Anup
