Re: [PATCH v14 11/18] irqchip: Add RISC-V incoming MSI controller early driver

[Anup Patel]

On Thu, Feb 22, 2024 at 6:43 PM Björn Töpel <bjorn@kernelorg> wrote:
>
> Anup Patel <apatel@xxxxxxxxxxxxxxxx> writes:
>
> > diff --git a/drivers/irqchip/irq-riscv-imsic-state.c b/drivers/irqchip/irq-riscv-imsic-state.c
> > new file mode 100644
> > index 000000000000..0c19ffb9ca3e
> > --- /dev/null
> > +++ b/drivers/irqchip/irq-riscv-imsic-state.c
> > @@ -0,0 +1,870 @@
>
> [...]
>
> > +static void __imsic_local_sync(struct imsic_local_priv *lpriv)
> > +{
> > +     struct imsic_local_config *mlocal;
> > +     struct imsic_vector *vec, *mvec;
> > +     int i;
> > +
> > +     lockdep_assert_held(&lpriv->lock);
> > +
> > +     /* This pairs with the barrier in __imsic_remote_sync(). */
> > +     smp_mb();
>
> I'm trying to figure out why this barrier is needed? All the updates are
> done behind the spinlocks. If there're some ordering constraints that
> I'm missing, please document them.
>
> > +
> > +     for_each_set_bit(i, lpriv->dirty_bitmap, imsic->global.nr_ids + 1) {
> > +             if (!i || i == IMSIC_IPI_ID)
> > +                     goto skip;
> > +             vec = &lpriv->vectors[i];
> > +
> > +             if (READ_ONCE(vec->enable))
> > +                     __imsic_id_set_enable(i);
> > +             else
> > +                     __imsic_id_clear_enable(i);
> > +
> > +             /*
> > +              * If the ID was being moved to a new ID on some other CPU
> > +              * then we can get a MSI during the movement so check the
> > +              * ID pending bit and re-trigger the new ID on other CPU
> > +              * using MMIO write.
> > +              */
> > +             mvec = READ_ONCE(vec->move);
> > +             WRITE_ONCE(vec->move, NULL);
>
> mvec = xchg(&vec->move, NULL) ?

The __imsic_local_sync() is called with spinlock held.

>
> > +             if (mvec && mvec != vec) {
> > +                     if (__imsic_id_read_clear_pending(i)) {
> > +                             mlocal = per_cpu_ptr(imsic->global.local, mvec->cpu);
> > +                             writel_relaxed(mvec->local_id, mlocal->msi_va);
> > +                     }
> > +
> > +                     imsic_vector_free(&lpriv->vectors[i]);
> > +             }
> > +
> > +skip:
> > +             bitmap_clear(lpriv->dirty_bitmap, i, 1);
> > +     }
> > +}
> > +
> > +void imsic_local_sync_all(void)
> > +{
> > +     struct imsic_local_priv *lpriv = this_cpu_ptr(imsic->lpriv);
> > +     unsigned long flags;
> > +
> > +     raw_spin_lock_irqsave(&lpriv->lock, flags);
> > +     bitmap_fill(lpriv->dirty_bitmap, imsic->global.nr_ids + 1);
> > +     __imsic_local_sync(lpriv);
> > +     raw_spin_unlock_irqrestore(&lpriv->lock, flags);
> > +}
> > +
> > +void imsic_local_delivery(bool enable)
> > +{
> > +     if (enable) {
> > +             imsic_csr_write(IMSIC_EITHRESHOLD, IMSIC_ENABLE_EITHRESHOLD);
> > +             imsic_csr_write(IMSIC_EIDELIVERY, IMSIC_ENABLE_EIDELIVERY);
> > +             return;
> > +     }
> > +
> > +     imsic_csr_write(IMSIC_EIDELIVERY, IMSIC_DISABLE_EIDELIVERY);
> > +     imsic_csr_write(IMSIC_EITHRESHOLD, IMSIC_DISABLE_EITHRESHOLD);
> > +}
> > +
> > +#ifdef CONFIG_SMP
> > +static void imsic_local_timer_callback(struct timer_list *timer)
> > +{
> > +     struct imsic_local_priv *lpriv = this_cpu_ptr(imsic->lpriv);
> > +     unsigned long flags;
> > +
> > +     raw_spin_lock_irqsave(&lpriv->lock, flags);
> > +     __imsic_local_sync(lpriv);
> > +     raw_spin_unlock_irqrestore(&lpriv->lock, flags);
> > +}
> > +
> > +static void __imsic_remote_sync(struct imsic_local_priv *lpriv, unsigned int cpu)
> > +{
> > +     lockdep_assert_held(&lpriv->lock);
> > +
> > +     /*
> > +      * Ensure that changes to vector enable, vector move and
> > +      * dirty bitmap are visible to the target CPU.
>
> ...which case the spinlock(s) are enough, no?

spinlocks are not fences.

If the timer wheel on the target cpu is already running and we don't
have a fence here then the target cpu might not see the changes
done by this cpu.

>
> > +      *
> > +      * This pairs with the barrier in __imsic_local_sync().
> > +      */
> > +     smp_mb();
> > +
> > +     /*
> > +      * We schedule a timer on the target CPU if the target CPU is not
> > +      * same as the current CPU. An offline CPU will unconditionally
> > +      * synchronize IDs through imsic_starting_cpu() when the
> > +      * CPU is brought up.
> > +      */
> > +     if (cpu_online(cpu)) {
> > +             if (cpu == smp_processor_id()) {
> > +                     __imsic_local_sync(lpriv);
> > +                     return;
> > +             }
>
> Maybe move this if-clause out from the cpu_online(), and only have
> something like
>   if (cpu_online(cpu) && !timer_pending(&lpriv->timer)) { ... }
> inside the CONFIG_SMP guard...
>
> > +
> > +             if (!timer_pending(&lpriv->timer)) {
> > +                     lpriv->timer.expires = jiffies + 1;
> > +                     add_timer_on(&lpriv->timer, cpu);
> > +             }
> > +     }
> > +}
> > +#else
> > +static void __imsic_remote_sync(struct imsic_local_priv *lpriv, unsigned int cpu)
> > +{
> > +     lockdep_assert_held(&lpriv->lock);
> > +     __imsic_local_sync(lpriv);
> > +}
> > +#endif
>
> ...where we can get rid of this special !SMP all together for this
> function.

I failed to understand what is wrong the current code.

>
> > +
> > +void imsic_vector_mask(struct imsic_vector *vec)
> > +{
> > +     struct imsic_local_priv *lpriv;
> > +
> > +     lpriv = per_cpu_ptr(imsic->lpriv, vec->cpu);
> > +     if (WARN_ON_ONCE(&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;
>
> Should have WRITE_ONCE to make the checkers happy.

Okay, I will update.

>
> > +     bitmap_set(lpriv->dirty_bitmap, vec->local_id, 1);
> > +     __imsic_remote_sync(lpriv, vec->cpu);
> > +
> > +     raw_spin_unlock(&lpriv->lock);
> > +}
> > +
> > +void imsic_vector_unmask(struct imsic_vector *vec)
> > +{
> > +     struct imsic_local_priv *lpriv;
> > +
> > +     lpriv = per_cpu_ptr(imsic->lpriv, vec->cpu);
> > +     if (WARN_ON_ONCE(&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 = true;
>
> Dito.

Okay, I will update.

>
> > +     bitmap_set(lpriv->dirty_bitmap, vec->local_id, 1);
> > +     __imsic_remote_sync(lpriv, vec->cpu);
> > +
> > +     raw_spin_unlock(&lpriv->lock);
> > +}
> > +
> > +static bool imsic_vector_move_update(struct imsic_local_priv *lpriv, struct imsic_vector *vec,
> > +                                  bool new_enable, struct imsic_vector *new_move)
> > +{
> > +     unsigned long flags;
> > +     bool enabled;
> > +
> > +     raw_spin_lock_irqsave(&lpriv->lock, flags);
> > +
> > +     /* Update enable and move details */
> > +     enabled = READ_ONCE(vec->enable);
> > +     WRITE_ONCE(vec->enable, new_enable);
>
> Again, xchg() would be easier to read.

why ? spinlock is not enough?

Regards,
Anup