[PATCH] drm/amd/display: only clear the hotplug sources the ASIC has

[Mattia Tadini]

amdgpu_dm_hpd_init() clears every hotplug source in the enum before
enabling the ones it needs:

	for (i = DC_IRQ_SOURCE_HPD1; i <= DC_IRQ_SOURCE_HPD6RX; i++)

DC fills the unused slots of its interrupt table with dummy handlers
that assert when touched, so on an ASIC with fewer than six HPD lines
that sweep warns and taints the kernel. DCN 2.0.1 has two:
mmHPD0_DC_HPD_INT_CONTROL and mmHPD1_DC_HPD_INT_CONTROL are the only
hotplug registers in dcn_2_0_1_offset.h, against five in
dcn_2_1_0_offset.h, and irq_service_dcn201.c defines hpd_int_entry(0)
and (1) accordingly.

On an AMD BC-250, which appears to be the only GPU using DCN 201, every
boot produces this for sources 3 to 6:

  [drm] *ERROR* dal_irq_service_dummy_ack: called for
      non-implemented irq source, src_id=0, ext_id=0
  [drm] dal_irq_service_set: src: 3, st: 0
  WARNING: .../display/dc/irq/irq_service.c:129
      at dal_irq_service_set.cold+0x2d/0x76 [amdgpu]
   dal_irq_service_set+0x4c/0x100 [amdgpu]
   amdgpu_dm_hpd_init.cold+0x15/0x12e [amdgpu]
   dm_hw_init+0x4c/0x170 [amdgpu]
   amdgpu_device_init.cold+0x1727/0x1ec0 [amdgpu]

and the machine comes up tainted W, which makes every unrelated bug
report from it harder to read.

Walk the links this ASIC has instead of the whole enum, and skip a link
with no hotplug line of its own, which is the same guard the enable loop
further down already applies. A source owned by a real link is
implemented by definition, so the dummy handlers are no longer reached.
Parts with six HPD lines are unaffected.

Signed-off-by: Mattia Tadini <info@xxxxxxxxxxxx>
---

Tested on an ASRock AMD BC-250 (Cyan Skillfish, DCN 2.0.1, two HPD lines),
kernel 7.2.6, across reboots before and after the change:

  before:  2 WARN splats and 8 error lines per boot, for sources 3 to 6
           /proc/sys/kernel/tainted = 12864
  after:   no dummy_ack, no irq_service warning, no call trace at all
           /proc/sys/kernel/tainted = 12352

The 512 that goes away is the W bit. What is left comes from an unrelated
out-of-tree module on that machine.

Nothing regressed: both DisplayPort connectors still report the right status,
the driver logs no HPD failure, and amdgpu keeps servicing interrupts
normally.

I do not have an ASIC with six HPD lines to test the unchanged path on.
 .../drm/amd/display/amdgpu_dm/amdgpu_dm_irq.c | 22 ++++++++++++++-----
 1 file changed, 17 insertions(+), 5 deletions(-)

diff --git a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_irq.c b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_irq.c
index e49803a..b0ede5d 100644
--- a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_irq.c
+++ b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_irq.c
@@ -903,11 +903,23 @@ void amdgpu_dm_hpd_init(struct amdgpu_device *adev)
 	int i;
 	bool use_polling = false;
 
-	/* First, clear all hpd and hpdrx interrupts */
-	for (i = DC_IRQ_SOURCE_HPD1; i <= DC_IRQ_SOURCE_HPD6RX; i++) {
-		if (!dc_interrupt_set(adev->dm.dc, i, false))
-			drm_err(dev, "Failed to clear hpd(rx) source=%d on init\n",
-				i);
+	/*
+	 * First, clear all hpd and hpdrx interrupts, over the links this ASIC
+	 * has rather than over the whole enum: the unused slots of the DC
+	 * interrupt table hold dummy handlers that assert when touched.
+	 */
+	for (i = 0; i < adev->dm.dc->link_count; i++) {
+		const struct dc_link *link = adev->dm.dc->links[i];
+
+		if (link->irq_source_hpd != DC_IRQ_SOURCE_INVALID &&
+		    !dc_interrupt_set(adev->dm.dc, link->irq_source_hpd, false))
+			drm_err(dev, "Failed to clear hpd source=%d on init\n",
+				link->irq_source_hpd);
+
+		if (link->irq_source_hpd_rx != DC_IRQ_SOURCE_INVALID &&
+		    !dc_interrupt_set(adev->dm.dc, link->irq_source_hpd_rx, false))
+			drm_err(dev, "Failed to clear hpdrx source=%d on init\n",
+				link->irq_source_hpd_rx);
 	}
 
 	drm_connector_list_iter_begin(dev, &iter);
-- 
2.55.0