On 5/15/19 5:47 AM, Arnd Bergmann wrote: > On Sun, May 12, 2019 at 3:25 AM Alex Elder <el...@linaro.org> wrote: > > The per-event interrupt handling seems to be more complex than > necessary:
I just noticed this message. I'll take another look at this whole interrupt control mechanism and will try to streamline it along the lines of what you describe. Thanks. -Alex > >> +/* Enable or disable an event interrupt */ >> +static void >> +_gsi_irq_control_event(struct gsi *gsi, u32 evt_ring_id, bool enable) >> +{ >> + u32 mask = BIT(evt_ring_id); >> + u32 val; >> + >> + if (enable) >> + gsi->event_enable_bitmap |= mask; >> + else >> + gsi->event_enable_bitmap &= ~mask; >> + >> + val = gsi->event_enable_bitmap; >> + iowrite32(val, gsi->virt + GSI_CNTXT_SRC_IEOB_IRQ_MSK_OFFSET); >> +} >> + >> +static void gsi_irq_enable_event(struct gsi *gsi, u32 evt_ring_id) >> +{ >> + _gsi_irq_control_event(gsi, evt_ring_id, true); > > You maintain a bitmap here of the enabled-state, and have > to use a spinlock to ensure that the two are in sync. > >> +/* Returns true if the interrupt state (enabled or not) changed */ >> +static bool gsi_channel_intr(struct gsi_channel *channel, bool enable) >> +{ >> + u32 evt_ring_id = channel->evt_ring_id; >> + struct gsi *gsi = channel->gsi; >> + u32 mask = BIT(evt_ring_id); >> + unsigned long flags; >> + bool different; >> + u32 enabled; >> + >> + spin_lock_irqsave(&gsi->spinlock, flags); >> + >> + enabled = gsi->event_enable_bitmap & mask; >> + different = enable == !enabled; >> + >> + if (different) { >> + if (enabled) >> + gsi_irq_disable_event(channel->gsi, evt_ring_id); >> + else >> + gsi_irq_enable_event(channel->gsi, evt_ring_id); >> + } >> + >> + spin_unlock_irqrestore(&gsi->spinlock, flags); >> + >> + return different; >> +} > > This gets called for each active channel, so you repeatedly > have to get the spinlock and read the irq-enabled state for it. > >> +static void gsi_isr_ieob(struct gsi *gsi) >> +{ >> + u32 evt_mask; >> + >> + evt_mask = ioread32(gsi->virt + GSI_CNTXT_SRC_IEOB_IRQ_OFFSET); >> + evt_mask &= ioread32(gsi->virt + GSI_CNTXT_SRC_IEOB_IRQ_MSK_OFFSET); >> + iowrite32(evt_mask, gsi->virt + GSI_CNTXT_SRC_IEOB_IRQ_CLR_OFFSET); >> + >> + while (evt_mask) { >> + u32 evt_ring_id = __ffs(evt_mask); >> + >> + evt_mask ^= BIT(evt_ring_id); >> + >> + gsi_event_handle(gsi, evt_ring_id); >> + } >> +} > > However, you start out by clearing all bits here. > > Why not skip the clearing and and leave the interrupts enabled, > while moving the GSI_CNTXT_SRC_IEOB_IRQ_CLR_OFFSET > write (for a single channel that was completed) to the end of > gsi_channel_poll()? > > Something like > > static void gsi_isr_ieob(struct gsi *gsi) > { > u32 evt_mask; > > evt_mask = ioread32(gsi->virt + GSI_CNTXT_SRC_IEOB_IRQ_OFFSET); > while (evt_mask) { > u32 evt_ring_id = __ffs(evt_mask); > evt_mask ^= BIT(evt_ring_id); > > napi_schedule(gsi->evt_ring[evt_ring_id].channel.napi); > } > } > > I also removed the GSI_CNTXT_SRC_IEOB_IRQ_MSK_OFFSET > read here, as that is probably more expensive than calling napi_schedule() > for a channel that is already scheduled. Most of the time, I'd expect the > interrupt to only signal a single channel anyway. > > Arnd >