[PATCH 0/8] drm/msm: Assorted fixes - October

[PATCH 0/8] drm/msm: Assorted fixes - October

[Akhil P Oommen]

Few assorted fixes for drm-msm driver lying in my stack.

Signed-off-by: Akhil P Oommen <akhilpo@oss.qualcomm.com>
---
Akhil P Oommen (4):
      drm/msm/a6xx: Check BX dependency vote index
      drm/msm/a6xx: Reset preemption records across suspend
      drm/msm/a6xx: Request CLX enable from secure world
      drm/msm/a8xx: Add AQE firmware to Adreno X2-90 GPU

Jie Zhang (4):
      drm/msm/a6xx: Fix GMU shutdown idle check
      drm/msm/a6xx: Fix stale boot up BW vote
      drm/msm/devfreq: Record idle_time at GPU idle onset
      drm/msm: Fix sched_prio inversion within a ring

 drivers/gpu/drm/msm/adreno/a6xx_gmu.c             | 31 ++++++++++++++-----
 drivers/gpu/drm/msm/adreno/a6xx_gmu.h             |  1 +
 drivers/gpu/drm/msm/adreno/a6xx_gpu.c             |  2 ++
 drivers/gpu/drm/msm/adreno/a6xx_gpu.h             |  1 +
 drivers/gpu/drm/msm/adreno/a6xx_hfi.c             |  2 +-
 drivers/gpu/drm/msm/adreno/a6xx_preempt.c         | 37 +++++++++++++----------
 drivers/gpu/drm/msm/adreno/a8xx_catalog.c         |  1 +
 drivers/gpu/drm/msm/adreno/a8xx_preempt.c         | 11 +++++--
 drivers/gpu/drm/msm/msm_gpu.h                     | 24 ++++++---------
 drivers/gpu/drm/msm/msm_gpu_devfreq.c             |  3 +-
 drivers/gpu/drm/msm/msm_submitqueue.c             |  2 +-
 drivers/gpu/drm/msm/registers/adreno/a6xx_gmu.xml |  1 +
 12 files changed, 73 insertions(+), 43 deletions(-)
---
base-commit: 77cb3534193b758eaecd662a610a3a1bffde43ed
change-id: 20261024-assorted-fixes-oct-1f981e99e677

Best regards,
-- 
Akhil P Oommen <akhilpo@oss.qualcomm.com>

[PATCH 1/8] drm/msm/a6xx: Check BX dependency vote index

[Akhil P Oommen]

a6xx_gmu_rpmh_arc_index() returns a signed error value. Check the result
before passing it to FIELD_MODIFY(), which rejects a compiler-propagated
negative value as too large for the field.

Fixes: 020b8eaa1db8 ("drm/msm/a6xx: Add BX rail support to GMU dependency votes")
Cc: stable@vger.kernel.org
Reported-by: kernel test robot <lkp@intel.com>
Closes: https://lore.kernel.org/oe-kbuild-all/202610070334.bYvBJSpj-lkp@intel.com/
Signed-off-by: Akhil P Oommen <akhilpo@oss.qualcomm.com>
---
 drivers/gpu/drm/msm/adreno/a6xx_gmu.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gmu.c b/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
index 662f09bec9d5..9e02d6f939af 100644
--- a/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
+++ b/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
@@ -1867,6 +1867,9 @@ static int a6xx_gmu_rpmh_dep_votes_init(struct device *dev, u32 *votes,
 		 * back to the highest BX corner if none is high enough.
 		 */
 		bx = a6xx_gmu_rpmh_arc_index(dev, gbx, gbx_count, gx[index], true);
+		if (bx < 0)
+			return bx;
+
 		FIELD_MODIFY(GMU_BX_MASK, &votes[i], bx);
 	}
 

-- 
2.55.0

[PATCH 2/8] drm/msm/a6xx: Fix GMU shutdown idle check

[Akhil P Oommen]

From: Jie Zhang <jie.zhang@oss.qualcomm.com>

GPUBUSYIGNAHB only indicates GPU hardware busy status and does not cover
the whole GPU subsystem. Use CXGXCPUBUSYIGNAHB instead so GMU shutdown
waits until CX/GX/CPU busy activity is cleared.

Also, increase the timeout to 100ms to match the downstream driver.

Fixes: 4b565ca5a2cb ("drm/msm: Add A6XX device support")
Cc: stable@vger.kernel.org
Signed-off-by: Jie Zhang <jie.zhang@oss.qualcomm.com>
Signed-off-by: Akhil P Oommen <akhilpo@oss.qualcomm.com>
---
 drivers/gpu/drm/msm/adreno/a6xx_gmu.c             | 4 ++--
 drivers/gpu/drm/msm/registers/adreno/a6xx_gmu.xml | 1 +
 2 files changed, 3 insertions(+), 2 deletions(-)

diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gmu.c b/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
index 9e02d6f939af..7e80670736a6 100644
--- a/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
+++ b/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
@@ -1461,8 +1461,8 @@ static void a6xx_gmu_shutdown(struct a6xx_gmu *gmu)
 
 	ret = gmu_poll_timeout(gmu,
 		REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS, val,
-		!(val & A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS_GPUBUSYIGNAHB),
-		100, 10000);
+		!(val & A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS_CXGXCPUBUSYIGNAHB),
+		100, 100000);
 
 	/*
 	 * Let the user know we failed to slumber but don't worry too
diff --git a/drivers/gpu/drm/msm/registers/adreno/a6xx_gmu.xml b/drivers/gpu/drm/msm/registers/adreno/a6xx_gmu.xml
index b3082738c99d..786fd7d2f674 100644
--- a/drivers/gpu/drm/msm/registers/adreno/a6xx_gmu.xml
+++ b/drivers/gpu/drm/msm/registers/adreno/a6xx_gmu.xml
@@ -224,6 +224,7 @@ xsi:schemaLocation="https://gitlab.freedesktop.org/freedreno/ rules-fd.xsd">
 	<reg32 offset="0x23b0b" name="GPU_GMU_AO_GMU_CGC_HYST_CNTL"/>
 	<reg32 offset="0x23b0c" name="GPU_GMU_AO_GPU_CX_BUSY_STATUS">
 		<bitfield name = "GPUBUSYIGNAHB" pos="23" type="boolean"/>
+		<bitfield name = "CXGXCPUBUSYIGNAHB" pos="30" type="boolean"/>
 	</reg32>
 	<reg32 offset="0x23b0d" name="GPU_GMU_AO_GPU_CX_BUSY_STATUS2"/>
 	<reg32 offset="0x23b0e" name="GPU_GMU_AO_GPU_CX_BUSY_MASK"/>

-- 
2.55.0

[PATCH 3/8] drm/msm/a6xx: Fix stale boot up BW vote

[Akhil P Oommen]

From: Jie Zhang <jie.zhang@oss.qualcomm.com>

An initial bus bandwidth is voted during boot to bring up the GMU. In GPUs
where the GMU drives bandwidth scaling, this boot-time BW vote is
redundant and can be removed.

Fixes: 7047e655a567 ("drm/msm: adreno: find bandwidth index of OPP and set it along freq index")
Cc: stable@vger.kernel.org
Signed-off-by: Jie Zhang <jie.zhang@oss.qualcomm.com>
Signed-off-by: Akhil P Oommen <akhilpo@oss.qualcomm.com>
---
 drivers/gpu/drm/msm/adreno/a6xx_gmu.c | 16 +++++++++++++---
 drivers/gpu/drm/msm/adreno/a6xx_gmu.h |  1 +
 drivers/gpu/drm/msm/adreno/a6xx_hfi.c |  2 +-
 3 files changed, 15 insertions(+), 4 deletions(-)

diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gmu.c b/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
index 7e80670736a6..991db3c9cd96 100644
--- a/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
+++ b/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
@@ -76,6 +76,13 @@ static irqreturn_t a6xx_hfi_irq(int irq, void *data)
 	return IRQ_HANDLED;
 }
 
+bool a6xx_gmu_scales_bw(struct a6xx_gmu *gmu)
+{
+	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
+
+	return a6xx_gpu->base.info->a6xx->bcms && gmu->nr_gpu_bws > 1;
+}
+
 bool a6xx_gmu_sptprac_is_on(struct a6xx_gmu *gmu)
 {
 	u32 val;
@@ -169,7 +176,6 @@ void a6xx_gmu_set_freq(struct msm_gpu *gpu, struct dev_pm_opp *opp,
 		       bool suspended)
 {
 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
-	const struct a6xx_info *info = adreno_gpu->info->a6xx;
 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
 	u32 perf_index;
@@ -187,7 +193,7 @@ void a6xx_gmu_set_freq(struct msm_gpu *gpu, struct dev_pm_opp *opp,
 			break;
 
 	/* If enabled, find the corresponding DDR bandwidth index */
-	if (info->bcms && gmu->nr_gpu_bws > 1) {
+	if (a6xx_gmu_scales_bw(gmu)) {
 		unsigned int bw = dev_pm_opp_get_bw(opp, true, 0);
 
 		for (bw_index = 0; bw_index < gmu->nr_gpu_bws - 1; bw_index++) {
@@ -1393,6 +1399,10 @@ int a6xx_gmu_resume(struct a6xx_gpu *a6xx_gpu)
 	/* Set the GPU to the current freq */
 	a6xx_gmu_set_initial_freq(gpu, gmu);
 
+	/* Clear the boot bus quota to allow GMU-driven BW scaling */
+	if (a6xx_gmu_scales_bw(gmu))
+		dev_pm_opp_set_opp(&gpu->pdev->dev, NULL);
+
 	return 0;
 
 disable_irq:
@@ -1913,7 +1923,7 @@ static int a6xx_gmu_rpmh_votes_init(struct a6xx_gmu *gmu)
 		gmu->gmu_freqs, gmu->nr_gmu_freqs, "cx.lvl", "mx.lvl");
 
 	/* Build the interconnect votes */
-	if (info->bcms && gmu->nr_gpu_bws > 1)
+	if (a6xx_gmu_scales_bw(gmu))
 		ret |= a6xx_gmu_rpmh_bw_votes_init(adreno_gpu, info, gmu);
 
 	return ret;
diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gmu.h b/drivers/gpu/drm/msm/adreno/a6xx_gmu.h
index d3f6cd8b9741..b9e966e602e2 100644
--- a/drivers/gpu/drm/msm/adreno/a6xx_gmu.h
+++ b/drivers/gpu/drm/msm/adreno/a6xx_gmu.h
@@ -239,6 +239,7 @@ bool a6xx_gmu_gx_is_on(struct adreno_gpu *adreno_gpu);
 bool a7xx_gmu_gx_is_on(struct adreno_gpu *adreno_gpu);
 bool a8xx_gmu_gx_is_on(struct adreno_gpu *adreno_gpu);
 bool a6xx_gmu_sptprac_is_on(struct a6xx_gmu *gmu);
+bool a6xx_gmu_scales_bw(struct a6xx_gmu *gmu);
 void a6xx_sptprac_disable(struct a6xx_gmu *gmu);
 int a6xx_sptprac_enable(struct a6xx_gmu *gmu);
 
diff --git a/drivers/gpu/drm/msm/adreno/a6xx_hfi.c b/drivers/gpu/drm/msm/adreno/a6xx_hfi.c
index 6b34e704a70d..ac684d72d2b4 100644
--- a/drivers/gpu/drm/msm/adreno/a6xx_hfi.c
+++ b/drivers/gpu/drm/msm/adreno/a6xx_hfi.c
@@ -810,7 +810,7 @@ static int a6xx_hfi_send_bw_table(struct a6xx_gmu *gmu)
 	if (!msg)
 		return -ENOMEM;
 
-	if (info->bcms && gmu->nr_gpu_bws > 1)
+	if (a6xx_gmu_scales_bw(gmu))
 		a6xx_generate_bw_table(info, gmu, msg);
 	else if (adreno_is_a618(adreno_gpu))
 		a618_build_bw_table(msg);

-- 
2.55.0

[PATCH 4/8] drm/msm/devfreq: Record idle_time at GPU idle onset

[Akhil P Oommen]

From: Jie Zhang <jie.zhang@oss.qualcomm.com>

Setting df->idle_time in msm_devfreq_idle_work() means a cancelled
idle_work leaves a stale timestamp, causing idle_time to accumulate
across idle cycles and potentially trigger a spurious devfreq boost.

Move the assignment to msm_devfreq_idle() so the timestamp always
reflects the actual GPU idle onset.

Fixes: 658f4c829688 ("drm/msm/devfreq: Add 1ms delay before clamping freq")
Cc: stable@vger.kernel.org
Signed-off-by: Jie Zhang <jie.zhang@oss.qualcomm.com>
Signed-off-by: Akhil P Oommen <akhilpo@oss.qualcomm.com>
---
 drivers/gpu/drm/msm/msm_gpu_devfreq.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/msm/msm_gpu_devfreq.c b/drivers/gpu/drm/msm/msm_gpu_devfreq.c
index 2e1d5c343272..f3c664b20a1b 100644
--- a/drivers/gpu/drm/msm/msm_gpu_devfreq.c
+++ b/drivers/gpu/drm/msm/msm_gpu_devfreq.c
@@ -357,7 +357,6 @@ static void msm_devfreq_idle_work(struct kthread_work *work)
 	if (priv->gpu_clamp_to_idle)
 		msm_devfreq_target(&gpu->pdev->dev, &target_freq, 0);
 
-	df->idle_time = ktime_get();
 	df->idle_freq = idle_freq;
 
 	mutex_unlock(&df->devfreq->lock);
@@ -370,6 +369,8 @@ void msm_devfreq_idle(struct msm_gpu *gpu)
 	if (!has_devfreq(gpu))
 		return;
 
+	df->idle_time = ktime_get();
+
 	msm_hrtimer_queue_work(&df->idle_work, ms_to_ktime(1),
 			       HRTIMER_MODE_REL);
 }

-- 
2.55.0

[PATCH 5/8] drm/msm: Fix sched_prio inversion within a ring

[Akhil P Oommen]

From: Jie Zhang <jie.zhang@oss.qualcomm.com>

Commit 38f922a563aa ("drm/sched: Reverse run-queue priority enumeration")
reversed the values of each sched priority. But it didn't update
msm_gpu_convert_priority() to match the new order. This results in
incorrect sched priorities within a ring.

To fix this, remove the inversion logic and add an offset of
DRM_SCHED_PRIORITY_HIGH to keep the range within [HIGH..LOW], reserving
KERNEL for internal submissions.

Fixes: 38f922a563aa ("drm/sched: Reverse run-queue priority enumeration")
Cc: stable@vger.kernel.org
Signed-off-by: Jie Zhang <jie.zhang@oss.qualcomm.com>
Signed-off-by: Akhil P Oommen <akhilpo@oss.qualcomm.com>
---
 drivers/gpu/drm/msm/msm_gpu.h         | 24 ++++++++++--------------
 drivers/gpu/drm/msm/msm_submitqueue.c |  2 +-
 2 files changed, 11 insertions(+), 15 deletions(-)

diff --git a/drivers/gpu/drm/msm/msm_gpu.h b/drivers/gpu/drm/msm/msm_gpu.h
index d27d54bdb7a7..bc5b31eb1eca 100644
--- a/drivers/gpu/drm/msm/msm_gpu.h
+++ b/drivers/gpu/drm/msm/msm_gpu.h
@@ -522,17 +522,18 @@ msm_context_is_vmbind(struct msm_context *ctx)
  * @sched_prio: [out] the gpu scheduler priority level which the userspace
  *              priority maps to
  *
- * With drm/scheduler providing it's own level of prioritization, our total
+ * With drm/scheduler providing its own level of prioritization, our total
  * number of available priority levels is (nr_rings * NR_SCHED_PRIORITIES).
- * Each ring is associated with it's own scheduler instance.  However, our
- * UABI is that lower numerical values are higher priority.  So mapping the
- * single userspace priority level into ring_nr and sched_prio takes some
- * care.  The userspace provided priority (when a submitqueue is created)
- * is mapped to ring nr and scheduler priority as such:
+ * Each ring is associated with its own scheduler instance.  Both the MSM
+ * UABI and drm/scheduler use lower-numeric = higher-priority, so the mapping
+ * from a single userspace priority level into ring_nr and sched_prio is:
  *
  *   ring_nr    = userspace_prio / NR_SCHED_PRIORITIES
- *   sched_prio = NR_SCHED_PRIORITIES -
- *                (userspace_prio % NR_SCHED_PRIORITIES) - 1
+ *   sched_prio = (userspace_prio % NR_SCHED_PRIORITIES) + DRM_SCHED_PRIORITY_HIGH
+ *
+ * The offset by DRM_SCHED_PRIORITY_HIGH keeps the range within [HIGH..LOW],
+ * intentionally excluding DRM_SCHED_PRIORITY_KERNEL which is reserved for
+ * kernel-internal submissions.
  *
  * This allows generations without preemption (nr_rings==1) to have some
  * amount of prioritization, and provides more priority levels for gens
@@ -547,16 +548,11 @@ static inline int msm_gpu_convert_priority(struct msm_gpu *gpu, int prio,
 
 	rn = div_u64_rem(prio, NR_SCHED_PRIORITIES, &sp);
 
-	/* invert sched priority to map to higher-numeric-is-higher-
-	 * priority convention
-	 */
-	sp = NR_SCHED_PRIORITIES - sp - 1;
-
 	if (rn >= gpu->nr_rings)
 		return -EINVAL;
 
 	*ring_nr = rn;
-	*sched_prio = sp;
+	*sched_prio = sp + DRM_SCHED_PRIORITY_HIGH;
 
 	return 0;
 }
diff --git a/drivers/gpu/drm/msm/msm_submitqueue.c b/drivers/gpu/drm/msm/msm_submitqueue.c
index d307cdd17d90..8d550d281704 100644
--- a/drivers/gpu/drm/msm/msm_submitqueue.c
+++ b/drivers/gpu/drm/msm/msm_submitqueue.c
@@ -136,7 +136,7 @@ get_sched_entity(struct msm_context *ctx, struct msm_ringbuffer *ring,
 		 unsigned ring_nr, enum drm_sched_priority sched_prio)
 {
 	static DEFINE_MUTEX(entity_lock);
-	unsigned idx = (ring_nr * NR_SCHED_PRIORITIES) + sched_prio;
+	unsigned idx = (ring_nr * NR_SCHED_PRIORITIES) + (sched_prio - DRM_SCHED_PRIORITY_HIGH);
 
 	/* We should have already validated that the requested priority is
 	 * valid by the time we get here.

-- 
2.55.0

[PATCH 6/8] drm/msm/a6xx: Reset preemption records across suspend

[Akhil P Oommen]

The preemption records are initialised only once at probe time. On every
resume, a6xx_preempt_hw_init() and a8xx_preempt_hw_init() refresh only a
subset of the fields (rptr, wptr, rptr_addr, info, data, rbase). Whatever
else the CP wrote into the record before the GPU was power-collapsed
is left in place and gets consumed again by the CP on the next preemption,
now carrying stale context state.

To fix this, zero the whole preemption record on suspend and fully
initialise every field in the hw_init path.

Fixes: e7ae83da4a28 ("drm/msm/a6xx: Implement preemption for a7xx targets")
Cc: stable@vger.kernel.org
Signed-off-by: Akhil P Oommen <akhilpo@oss.qualcomm.com>
---
 drivers/gpu/drm/msm/adreno/a6xx_gpu.c     |  2 ++
 drivers/gpu/drm/msm/adreno/a6xx_gpu.h     |  1 +
 drivers/gpu/drm/msm/adreno/a6xx_preempt.c | 37 ++++++++++++++++++-------------
 drivers/gpu/drm/msm/adreno/a8xx_preempt.c | 11 ++++++---
 4 files changed, 32 insertions(+), 19 deletions(-)

diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
index 92e8c8abb86d..3e4afaa5345f 100644
--- a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
+++ b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
@@ -2248,6 +2248,7 @@ static int a6xx_gmu_pm_suspend(struct msm_gpu *gpu)
 		for (i = 0; i < gpu->nr_rings; i++)
 			a6xx_gpu->shadow[i] = 0;
 
+	a6xx_preempt_record_reset(gpu);
 	gpu->suspend_count++;
 
 	return 0;
@@ -2287,6 +2288,7 @@ static int a6xx_pm_suspend(struct msm_gpu *gpu)
 		for (i = 0; i < gpu->nr_rings; i++)
 			a6xx_gpu->shadow[i] = 0;
 
+	a6xx_preempt_record_reset(gpu);
 	gpu->suspend_count++;
 
 	return 0;
diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gpu.h b/drivers/gpu/drm/msm/adreno/a6xx_gpu.h
index 79912a33e17c..62a19e22e230 100644
--- a/drivers/gpu/drm/msm/adreno/a6xx_gpu.h
+++ b/drivers/gpu/drm/msm/adreno/a6xx_gpu.h
@@ -305,6 +305,7 @@ void a6xx_preempt_hw_init(struct msm_gpu *gpu);
 void a6xx_preempt_trigger(struct msm_gpu *gpu);
 void a6xx_preempt_irq(struct msm_gpu *gpu);
 void a6xx_preempt_fini(struct msm_gpu *gpu);
+void a6xx_preempt_record_reset(struct msm_gpu *gpu);
 void a6xx_emit_set_pseudo_reg(struct msm_ringbuffer *ring,
 		struct a6xx_gpu *a6xx_gpu, struct msm_gpu_submitqueue *queue);
 int a6xx_preempt_submitqueue_setup(struct msm_gpu *gpu,
diff --git a/drivers/gpu/drm/msm/adreno/a6xx_preempt.c b/drivers/gpu/drm/msm/adreno/a6xx_preempt.c
index 1e599d4ddea1..df8d4f3e8d4c 100644
--- a/drivers/gpu/drm/msm/adreno/a6xx_preempt.c
+++ b/drivers/gpu/drm/msm/adreno/a6xx_preempt.c
@@ -140,12 +140,17 @@ void a6xx_preempt_hw_init(struct msm_gpu *gpu)
 	for (i = 0; i < gpu->nr_rings; i++) {
 		struct a6xx_preempt_record *record_ptr = a6xx_gpu->preempt[i];
 
-		record_ptr->wptr = 0;
-		record_ptr->rptr = 0;
-		record_ptr->rptr_addr = shadowptr(a6xx_gpu, gpu->rb[i]);
+		record_ptr->magic = A6XX_PREEMPT_RECORD_MAGIC;
 		record_ptr->info = 0;
+		record_ptr->errno = 0;
 		record_ptr->data = 0;
+		record_ptr->cntl = MSM_GPU_RB_CNTL_DEFAULT;
+		record_ptr->rptr = 0;
+		record_ptr->wptr = 0;
+		record_ptr->rptr_addr = shadowptr(a6xx_gpu, gpu->rb[i]);
 		record_ptr->rbase = gpu->rb[i]->iova;
+		record_ptr->counter = 0;
+		record_ptr->bv_rptr_addr = rbmemptr(gpu->rb[i], bv_rptr);
 	}
 
 	/* Write a 0 to signal that we aren't switching pagetables */
@@ -163,6 +168,19 @@ void a6xx_preempt_hw_init(struct msm_gpu *gpu)
 	a6xx_gpu->cur_ring = gpu->rb[0];
 }
 
+void a6xx_preempt_record_reset(struct msm_gpu *gpu)
+{
+	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
+	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
+	int i;
+
+	if (gpu->nr_rings == 1)
+		return;
+
+	for (i = 0; i < gpu->nr_rings; i++)
+		memset(a6xx_gpu->preempt[i], 0, a6xx_gpu->preempt_bo[i]->size);
+}
+
 void a6xx_preempt_trigger(struct msm_gpu *gpu)
 {
 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
@@ -302,8 +320,6 @@ static int preempt_init_ring(struct a6xx_gpu *a6xx_gpu,
 	a6xx_gpu->preempt_iova[ring->id] = iova;
 	a6xx_gpu->preempt[ring->id] = ptr;
 
-	struct a6xx_preempt_record *record_ptr = ptr;
-
 	ptr = msm_gem_kernel_new(gpu->dev,
 		PREEMPT_SMMU_INFO_SIZE,
 		MSM_BO_WC | MSM_BO_MAP_PRIV | MSM_BO_GPU_READONLY,
@@ -329,17 +345,6 @@ static int preempt_init_ring(struct a6xx_gpu *a6xx_gpu,
 	smmu_info_ptr->asid = 0xdecafbad;
 	smmu_info_ptr->context_idr = 0;
 
-	/* Set up the defaults on the preemption record */
-	record_ptr->magic = A6XX_PREEMPT_RECORD_MAGIC;
-	record_ptr->info = 0;
-	record_ptr->data = 0;
-	record_ptr->rptr = 0;
-	record_ptr->wptr = 0;
-	record_ptr->cntl = MSM_GPU_RB_CNTL_DEFAULT;
-	record_ptr->rbase = ring->iova;
-	record_ptr->counter = 0;
-	record_ptr->bv_rptr_addr = rbmemptr(ring, bv_rptr);
-
 	return 0;
 }
 
diff --git a/drivers/gpu/drm/msm/adreno/a8xx_preempt.c b/drivers/gpu/drm/msm/adreno/a8xx_preempt.c
index 6cb53a071801..0e0503ed54a1 100644
--- a/drivers/gpu/drm/msm/adreno/a8xx_preempt.c
+++ b/drivers/gpu/drm/msm/adreno/a8xx_preempt.c
@@ -121,12 +121,17 @@ void a8xx_preempt_hw_init(struct msm_gpu *gpu)
 	for (i = 0; i < gpu->nr_rings; i++) {
 		struct a6xx_preempt_record *record_ptr = a6xx_gpu->preempt[i];
 
-		record_ptr->wptr = 0;
-		record_ptr->rptr = 0;
-		record_ptr->rptr_addr = shadowptr(a6xx_gpu, gpu->rb[i]);
+		record_ptr->magic = A6XX_PREEMPT_RECORD_MAGIC;
 		record_ptr->info = 0;
+		record_ptr->errno = 0;
 		record_ptr->data = 0;
+		record_ptr->cntl = MSM_GPU_RB_CNTL_DEFAULT;
+		record_ptr->rptr = 0;
+		record_ptr->wptr = 0;
+		record_ptr->rptr_addr = shadowptr(a6xx_gpu, gpu->rb[i]);
 		record_ptr->rbase = gpu->rb[i]->iova;
+		record_ptr->counter = 0;
+		record_ptr->bv_rptr_addr = rbmemptr(gpu->rb[i], bv_rptr);
 	}
 
 	/* Write a 0 to signal that we aren't switching pagetables */

-- 
2.55.0

Re: [PATCH 6/8] drm/msm/a6xx: Reset preemption records across suspend

[Connor Abbott]

On Sat, Oct 24, 2026 at 6:31 AM Akhil P Oommen <akhilpo@oss.qualcomm.com> wrote:
>
> The preemption records are initialised only once at probe time. On every
> resume, a6xx_preempt_hw_init() and a8xx_preempt_hw_init() refresh only a
> subset of the fields (rptr, wptr, rptr_addr, info, data, rbase). Whatever
> else the CP wrote into the record before the GPU was power-collapsed
> is left in place and gets consumed again by the CP on the next preemption,
> now carrying stale context state.
>
> To fix this, zero the whole preemption record on suspend and fully
> initialise every field in the hw_init path.
>
> Fixes: e7ae83da4a28 ("drm/msm/a6xx: Implement preemption for a7xx targets")
> Cc: stable@vger.kernel.org
> Signed-off-by: Akhil P Oommen <akhilpo@oss.qualcomm.com>

Thanks, this gets the AQE working on my glymur laptop (Slim 7x).

Tested-by: Connor Abbott <cwabbott0@gmail.com>

> ---
>  drivers/gpu/drm/msm/adreno/a6xx_gpu.c     |  2 ++
>  drivers/gpu/drm/msm/adreno/a6xx_gpu.h     |  1 +
>  drivers/gpu/drm/msm/adreno/a6xx_preempt.c | 37 ++++++++++++++++++-------------
>  drivers/gpu/drm/msm/adreno/a8xx_preempt.c | 11 ++++++---
>  4 files changed, 32 insertions(+), 19 deletions(-)
>
> diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
> index 92e8c8abb86d..3e4afaa5345f 100644
> --- a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
> +++ b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
> @@ -2248,6 +2248,7 @@ static int a6xx_gmu_pm_suspend(struct msm_gpu *gpu)
>                 for (i = 0; i < gpu->nr_rings; i++)
>                         a6xx_gpu->shadow[i] = 0;
>
> +       a6xx_preempt_record_reset(gpu);
>         gpu->suspend_count++;
>
>         return 0;
> @@ -2287,6 +2288,7 @@ static int a6xx_pm_suspend(struct msm_gpu *gpu)
>                 for (i = 0; i < gpu->nr_rings; i++)
>                         a6xx_gpu->shadow[i] = 0;
>
> +       a6xx_preempt_record_reset(gpu);
>         gpu->suspend_count++;
>
>         return 0;
> diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gpu.h b/drivers/gpu/drm/msm/adreno/a6xx_gpu.h
> index 79912a33e17c..62a19e22e230 100644
> --- a/drivers/gpu/drm/msm/adreno/a6xx_gpu.h
> +++ b/drivers/gpu/drm/msm/adreno/a6xx_gpu.h
> @@ -305,6 +305,7 @@ void a6xx_preempt_hw_init(struct msm_gpu *gpu);
>  void a6xx_preempt_trigger(struct msm_gpu *gpu);
>  void a6xx_preempt_irq(struct msm_gpu *gpu);
>  void a6xx_preempt_fini(struct msm_gpu *gpu);
> +void a6xx_preempt_record_reset(struct msm_gpu *gpu);
>  void a6xx_emit_set_pseudo_reg(struct msm_ringbuffer *ring,
>                 struct a6xx_gpu *a6xx_gpu, struct msm_gpu_submitqueue *queue);
>  int a6xx_preempt_submitqueue_setup(struct msm_gpu *gpu,
> diff --git a/drivers/gpu/drm/msm/adreno/a6xx_preempt.c b/drivers/gpu/drm/msm/adreno/a6xx_preempt.c
> index 1e599d4ddea1..df8d4f3e8d4c 100644
> --- a/drivers/gpu/drm/msm/adreno/a6xx_preempt.c
> +++ b/drivers/gpu/drm/msm/adreno/a6xx_preempt.c
> @@ -140,12 +140,17 @@ void a6xx_preempt_hw_init(struct msm_gpu *gpu)
>         for (i = 0; i < gpu->nr_rings; i++) {
>                 struct a6xx_preempt_record *record_ptr = a6xx_gpu->preempt[i];
>
> -               record_ptr->wptr = 0;
> -               record_ptr->rptr = 0;
> -               record_ptr->rptr_addr = shadowptr(a6xx_gpu, gpu->rb[i]);
> +               record_ptr->magic = A6XX_PREEMPT_RECORD_MAGIC;
>                 record_ptr->info = 0;
> +               record_ptr->errno = 0;
>                 record_ptr->data = 0;
> +               record_ptr->cntl = MSM_GPU_RB_CNTL_DEFAULT;
> +               record_ptr->rptr = 0;
> +               record_ptr->wptr = 0;
> +               record_ptr->rptr_addr = shadowptr(a6xx_gpu, gpu->rb[i]);
>                 record_ptr->rbase = gpu->rb[i]->iova;
> +               record_ptr->counter = 0;
> +               record_ptr->bv_rptr_addr = rbmemptr(gpu->rb[i], bv_rptr);
>         }
>
>         /* Write a 0 to signal that we aren't switching pagetables */
> @@ -163,6 +168,19 @@ void a6xx_preempt_hw_init(struct msm_gpu *gpu)
>         a6xx_gpu->cur_ring = gpu->rb[0];
>  }
>
> +void a6xx_preempt_record_reset(struct msm_gpu *gpu)
> +{
> +       struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
> +       struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
> +       int i;
> +
> +       if (gpu->nr_rings == 1)
> +               return;
> +
> +       for (i = 0; i < gpu->nr_rings; i++)
> +               memset(a6xx_gpu->preempt[i], 0, a6xx_gpu->preempt_bo[i]->size);
> +}
> +
>  void a6xx_preempt_trigger(struct msm_gpu *gpu)
>  {
>         struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
> @@ -302,8 +320,6 @@ static int preempt_init_ring(struct a6xx_gpu *a6xx_gpu,
>         a6xx_gpu->preempt_iova[ring->id] = iova;
>         a6xx_gpu->preempt[ring->id] = ptr;
>
> -       struct a6xx_preempt_record *record_ptr = ptr;
> -
>         ptr = msm_gem_kernel_new(gpu->dev,
>                 PREEMPT_SMMU_INFO_SIZE,
>                 MSM_BO_WC | MSM_BO_MAP_PRIV | MSM_BO_GPU_READONLY,
> @@ -329,17 +345,6 @@ static int preempt_init_ring(struct a6xx_gpu *a6xx_gpu,
>         smmu_info_ptr->asid = 0xdecafbad;
>         smmu_info_ptr->context_idr = 0;
>
> -       /* Set up the defaults on the preemption record */
> -       record_ptr->magic = A6XX_PREEMPT_RECORD_MAGIC;
> -       record_ptr->info = 0;
> -       record_ptr->data = 0;
> -       record_ptr->rptr = 0;
> -       record_ptr->wptr = 0;
> -       record_ptr->cntl = MSM_GPU_RB_CNTL_DEFAULT;
> -       record_ptr->rbase = ring->iova;
> -       record_ptr->counter = 0;
> -       record_ptr->bv_rptr_addr = rbmemptr(ring, bv_rptr);
> -
>         return 0;
>  }
>
> diff --git a/drivers/gpu/drm/msm/adreno/a8xx_preempt.c b/drivers/gpu/drm/msm/adreno/a8xx_preempt.c
> index 6cb53a071801..0e0503ed54a1 100644
> --- a/drivers/gpu/drm/msm/adreno/a8xx_preempt.c
> +++ b/drivers/gpu/drm/msm/adreno/a8xx_preempt.c
> @@ -121,12 +121,17 @@ void a8xx_preempt_hw_init(struct msm_gpu *gpu)
>         for (i = 0; i < gpu->nr_rings; i++) {
>                 struct a6xx_preempt_record *record_ptr = a6xx_gpu->preempt[i];
>
> -               record_ptr->wptr = 0;
> -               record_ptr->rptr = 0;
> -               record_ptr->rptr_addr = shadowptr(a6xx_gpu, gpu->rb[i]);
> +               record_ptr->magic = A6XX_PREEMPT_RECORD_MAGIC;
>                 record_ptr->info = 0;
> +               record_ptr->errno = 0;
>                 record_ptr->data = 0;
> +               record_ptr->cntl = MSM_GPU_RB_CNTL_DEFAULT;
> +               record_ptr->rptr = 0;
> +               record_ptr->wptr = 0;
> +               record_ptr->rptr_addr = shadowptr(a6xx_gpu, gpu->rb[i]);
>                 record_ptr->rbase = gpu->rb[i]->iova;
> +               record_ptr->counter = 0;
> +               record_ptr->bv_rptr_addr = rbmemptr(gpu->rb[i], bv_rptr);
>         }
>
>         /* Write a 0 to signal that we aren't switching pagetables */
>
> --
> 2.55.0
>

[PATCH 7/8] drm/msm/a6xx: Request CLX enable from secure world

[Akhil P Oommen]

The CLX (Current Limit Extension) feature also requires the secure
world to enable it via the GPU init regs SCM call. Without the
QCOM_SCM_GPU_CLX_EN_REQ flag, the CLX tables sent to the GMU over HFI
are not effective.

Set QCOM_SCM_GPU_CLX_EN_REQ in the SCM request on GPUs that have a CLX
table in their catalog entry.

Fixes: 967e62fbeecc ("drm/msm/a6xx: Add HFI support for CLX feature")
Signed-off-by: Akhil P Oommen <akhilpo@oss.qualcomm.com>
---
 drivers/gpu/drm/msm/adreno/a6xx_gmu.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)

diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gmu.c b/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
index 991db3c9cd96..a1561ca41a5b 100644
--- a/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
+++ b/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
@@ -1247,8 +1247,12 @@ static int a6xx_gmu_secure_init(struct a6xx_gpu *a6xx_gpu)
 			goto done;
 		}
 
-		ret = qcom_scm_gpu_init_regs(QCOM_SCM_GPU_ALWAYS_EN_REQ |
-					     QCOM_SCM_GPU_TSENSE_EN_REQ);
+		u32 gpu_req = QCOM_SCM_GPU_ALWAYS_EN_REQ | QCOM_SCM_GPU_TSENSE_EN_REQ;
+
+		if (adreno_gpu->info->a6xx->clx_tbl)
+			gpu_req |= QCOM_SCM_GPU_CLX_EN_REQ;
+
+		ret = qcom_scm_gpu_init_regs(gpu_req);
 		if (ret) {
 			dev_warn_once(gpu->dev->dev,
 				"SCM call failed\n");

-- 
2.55.0

[PATCH 8/8] drm/msm/a8xx: Add AQE firmware to Adreno X2-90 GPU

[Akhil P Oommen]

Update the catalog to add AQE firmware to the Adreno X2-90 GPU.

Signed-off-by: Akhil P Oommen <akhilpo@oss.qualcomm.com>
---
 drivers/gpu/drm/msm/adreno/a8xx_catalog.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/drivers/gpu/drm/msm/adreno/a8xx_catalog.c b/drivers/gpu/drm/msm/adreno/a8xx_catalog.c
index 55a805ab438c..775aca1964ee 100644
--- a/drivers/gpu/drm/msm/adreno/a8xx_catalog.c
+++ b/drivers/gpu/drm/msm/adreno/a8xx_catalog.c
@@ -1582,6 +1582,7 @@ static const struct adreno_info a8xx_gpus[] = {
 		.fw = {
 			[ADRENO_FW_SQE] = "gen80100_sqe.fw",
 			[ADRENO_FW_GMU] = "gen80100_gmu.bin",
+			[ADRENO_FW_AQE] = "gen80100_aqe.fw",
 		},
 		.gmem = 21 * SZ_1M,
 		.inactive_period = DRM_MSM_INACTIVE_PERIOD,

-- 
2.55.0