Re: [PATCH v5 1/5] drm/xe: Make irq enabled flag atomic

[Rodrigo Vivi <rodrigo.vivi@intel.com>]

On Tue, Dec 03, 2024 at 01:47:42PM +0200, Levi, Ilia wrote:
> On 02/12/2024 20:32, Piotr Piórkowski wrote:
> > Ilia Levi <ilia.levi@intel.com> wrote on czw [2024-lis-28 14:53:41 +0200]:
> >> The irq.enabled flag was protected by a spin lock (irq.lock).
> >> By making it atomic we no longer need to wait for the spin lock in
> >> irq handlers. This will become especially useful for MSI-X irq
> >> handlers to prevent lock contention between different interrupts.
> > I am not convinced that you can simply replace this spin_lock with an atomic.
> > Note that this spin lock is also used for whole blocks in the irq handler
> > (for example gt_irq_handler), and not only to access this flag.
> 
> Yes, I saw that. However, irq.enabled is not accessed within those blocks, so imho there is no need in mutual exclusion between checking the flag and those blocks.
> If I understand correctly, the role of irq.enabled flag is to prevent new irq handlers from running  once the interrupts have been turned off (e. g. as a result of suspension), while synchronize_irq waits for the already running irq handlers to finish. Making the flag atomic should achieve the same goal. I have left the spin lock to protect access to interrupt registers (and there's also some protection in xe_execlist_port_destroy), though I'm not sure when it is useful.
> 
> Adding Rodrigo and Matthew in case I'm missing something.

I believe Ilia is right here and I was over-protective trying to use
the spinlock to also protect the enabled and not just the irq data
itself as its intended doc says.

I was concerned about display code, but then looking again to the code,
on i915 the enabled is not protected as well and not even an atomic,
so an atomic is a bonus here.

But we do need to test this patch here with display enabled.
Please ensure you also convert the irq.enabled at
drivers/gpu/drm/xe/display/ext/i915_irq.c
and kill that XXX comment section that I had written there,
because after this patch that argument makes even less sense.

Thanks,
Rodrigo.

> Thanks,
> Ilia
> 
> >
> > Thanks,
> > Piotr
> >
> >
> >> Signed-off-by: Ilia Levi <ilia.levi@intel.com>
> >> ---
> >>  drivers/gpu/drm/xe/display/ext/i915_irq.c |  2 +-
> >>  drivers/gpu/drm/xe/xe_device_types.h      |  2 +-
> >>  drivers/gpu/drm/xe/xe_irq.c               | 29 ++++++-----------------
> >>  3 files changed, 9 insertions(+), 24 deletions(-)
> >>
> >> diff --git a/drivers/gpu/drm/xe/display/ext/i915_irq.c b/drivers/gpu/drm/xe/display/ext/i915_irq.c
> >> index a7dbc6554d69..0c0f4533c34f 100644
> >> --- a/drivers/gpu/drm/xe/display/ext/i915_irq.c
> >> +++ b/drivers/gpu/drm/xe/display/ext/i915_irq.c
> >> @@ -64,7 +64,7 @@ bool intel_irqs_enabled(struct xe_device *xe)
> >>  	 * But at this point the xe irq is better protected against races,
> >>  	 * although the full solution would be protecting the i915 side.
> >>  	 */
> >> -	return xe->irq.enabled;
> >> +	return atomic_read(&xe->irq.enabled);
> >>  }
> >>  
> >>  void intel_synchronize_irq(struct xe_device *xe)
> >> diff --git a/drivers/gpu/drm/xe/xe_device_types.h b/drivers/gpu/drm/xe/xe_device_types.h
> >> index 6a04f975ec16..7ee114c17552 100644
> >> --- a/drivers/gpu/drm/xe/xe_device_types.h
> >> +++ b/drivers/gpu/drm/xe/xe_device_types.h
> >> @@ -347,7 +347,7 @@ struct xe_device {
> >>  		spinlock_t lock;
> >>  
> >>  		/** @irq.enabled: interrupts enabled on this device */
> >> -		bool enabled;
> >> +		atomic_t enabled;
> >>  	} irq;
> >>  
> >>  	/** @ttm: ttm device */
> >> diff --git a/drivers/gpu/drm/xe/xe_irq.c b/drivers/gpu/drm/xe/xe_irq.c
> >> index 7bf7201529ac..1c509e66694d 100644
> >> --- a/drivers/gpu/drm/xe/xe_irq.c
> >> +++ b/drivers/gpu/drm/xe/xe_irq.c
> >> @@ -348,12 +348,8 @@ static irqreturn_t xelp_irq_handler(int irq, void *arg)
> >>  	unsigned long intr_dw[2];
> >>  	u32 identity[32];
> >>  
> >> -	spin_lock(&xe->irq.lock);
> >> -	if (!xe->irq.enabled) {
> >> -		spin_unlock(&xe->irq.lock);
> >> +	if (!atomic_read(&xe->irq.enabled))
> >>  		return IRQ_NONE;
> >> -	}
> >> -	spin_unlock(&xe->irq.lock);
> >>  
> >>  	master_ctl = xelp_intr_disable(xe);
> >>  	if (!master_ctl) {
> >> @@ -417,12 +413,8 @@ static irqreturn_t dg1_irq_handler(int irq, void *arg)
> >>  
> >>  	/* TODO: This really shouldn't be copied+pasted */
> >>  
> >> -	spin_lock(&xe->irq.lock);
> >> -	if (!xe->irq.enabled) {
> >> -		spin_unlock(&xe->irq.lock);
> >> +	if (!atomic_read(&xe->irq.enabled))
> >>  		return IRQ_NONE;
> >> -	}
> >> -	spin_unlock(&xe->irq.lock);
> >>  
> >>  	master_tile_ctl = dg1_intr_disable(xe);
> >>  	if (!master_tile_ctl) {
> >> @@ -644,12 +636,8 @@ static irqreturn_t vf_mem_irq_handler(int irq, void *arg)
> >>  	struct xe_tile *tile;
> >>  	unsigned int id;
> >>  
> >> -	spin_lock(&xe->irq.lock);
> >> -	if (!xe->irq.enabled) {
> >> -		spin_unlock(&xe->irq.lock);
> >> +	if (!atomic_read(&xe->irq.enabled))
> >>  		return IRQ_NONE;
> >> -	}
> >> -	spin_unlock(&xe->irq.lock);
> >>  
> >>  	for_each_tile(tile, xe, id)
> >>  		xe_memirq_handler(&tile->memirq);
> >> @@ -674,10 +662,9 @@ static void irq_uninstall(void *arg)
> >>  	struct pci_dev *pdev = to_pci_dev(xe->drm.dev);
> >>  	int irq;
> >>  
> >> -	if (!xe->irq.enabled)
> >> +	if (!atomic_xchg(&xe->irq.enabled, 0))
> >>  		return;
> >>  
> >> -	xe->irq.enabled = false;
> >>  	xe_irq_reset(xe);
> >>  
> >>  	irq = pci_irq_vector(pdev, 0);
> >> @@ -724,7 +711,7 @@ int xe_irq_install(struct xe_device *xe)
> >>  		return err;
> >>  	}
> >>  
> >> -	xe->irq.enabled = true;
> >> +	atomic_set(&xe->irq.enabled, 1);
> >>  
> >>  	xe_irq_postinstall(xe);
> >>  
> >> @@ -744,9 +731,7 @@ void xe_irq_suspend(struct xe_device *xe)
> >>  {
> >>  	int irq = to_pci_dev(xe->drm.dev)->irq;
> >>  
> >> -	spin_lock_irq(&xe->irq.lock);
> >> -	xe->irq.enabled = false; /* no new irqs */
> >> -	spin_unlock_irq(&xe->irq.lock);
> >> +	atomic_set(&xe->irq.enabled, 0); /* no new irqs */
> >>  
> >>  	synchronize_irq(irq); /* flush irqs */
> >>  	xe_irq_reset(xe); /* turn irqs off */
> >> @@ -762,7 +747,7 @@ void xe_irq_resume(struct xe_device *xe)
> >>  	 * 1. no irq will arrive before the postinstall
> >>  	 * 2. display is not yet resumed
> >>  	 */
> >> -	xe->irq.enabled = true;
> >> +	atomic_set(&xe->irq.enabled, 1);
> >>  	xe_irq_reset(xe);
> >>  	xe_irq_postinstall(xe); /* turn irqs on */
> >>  
> >> -- 
> >> 2.43.2
> >>
>