Skip to contentSkip to conversation

drm/amdgpu: track faulted gfx user-queue slots

PATCH New Jesse Zhang 1 participant 5 messages Jul 31

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH v4 1/5 drm/amdgpu: track faulted gfx user-queue slots

A gfx priv/bad-op fault IV carries only the HW slot (ring_id), not the faulting user queue's doorbell. Add userq_priv_fault_slots (an atomic bitmap of faulted slots, so concurrent faults are not dropped) and userq_priv_fault_work to struct amdgpu_gfx; a worker drains the bitmap and reads the doorbell back from each HQD to locate and reset the queue.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h | 3 +++
 1 file changed, 3 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h +3 -0
index c15f45a7347c..1023719c18bb 100644
@@ -535,6 +535,9 @@ struct amdgpu_gfx {
 	struct mutex                    userq_sch_mutex;
 	u64				userq_sch_req_count[MAX_XCP];
 	bool				userq_sch_inactive[MAX_XCP];
+	/* atomic bitmap of faulted gfx UQ slots (index = pipe | queue << 2) */
+	unsigned long			userq_priv_fault_slots;
+	struct work_struct		userq_priv_fault_work;
 	unsigned long			enforce_isolation_jiffies[MAX_XCP];
 	unsigned long			enforce_isolation_time[MAX_XCP];
 
Signature · 1 line
2.49.0
#1 lore raw

PATCH v4 2/5 drm/amdgpu/gfx11: add priv-fault user-queue recovery worker

A gfx user-queue priv/bad-op fault is raised by the ME and carries only the HW slot, not the faulting queue's doorbell. Add a per-IP worker that drains adev->gfx.userq_priv_fault_slots, reads the doorbell back from each HQD via soc21_grbm_select (regCP_RB_DOORBELL_CONTROL), looks up the user queue and kicks its per-queue reset.

The shared amdgpu_gfx_handle_priv_fault() helper schedules this worker for a gfx user-queue fault; wiring the helper up is done in a later patch.

v2:
- gate on adev->gfx.disable_uq instead of !adev->enable_mes (Alex)
- document why both the doorbell (compute) and HW-slot (gfx) reset paths are needed (Alex)

v3:
- rebase amd-staging-drm-next. adapt to the deduplicated amdgpu_gfx_handle_priv_fault() helper (commit d8ab7636160e); no functional change

---
 drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c | 39 ++++++++++++++++++++++++++
 1 file changed, 39 insertions(+)
drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c +39 -0
index 9faf31421a15..69776dbe188d 100644
@@ -362,6 +362,7 @@ static void gfx_v11_0_ring_invalidate_tlbs(struct amdgpu_ring *ring,
 					   bool all_hub, uint8_t dst_sel);
 static void gfx_v11_0_set_safe_mode(struct amdgpu_device *adev, int xcc_id);
 static void gfx_v11_0_unset_safe_mode(struct amdgpu_device *adev, int xcc_id);
+static void gfx_v11_0_userq_priv_fault_work(struct work_struct *work);
 static void gfx_v11_0_update_perf_clk(struct amdgpu_device *adev,
 				      bool enable);
 
@@ -1932,6 +1933,8 @@ static int gfx_v11_0_sw_init(struct amdgpu_ip_block *ip_block)
 
 	mutex_init(&adev->gfx.mec.reset_mutex);
 
+	INIT_WORK(&adev->gfx.userq_priv_fault_work, gfx_v11_0_userq_priv_fault_work);
+
 	return 0;
 }
 
@@ -1969,6 +1972,8 @@ static int gfx_v11_0_sw_fini(struct amdgpu_ip_block *ip_block)
 	int i;
 	struct amdgpu_device *adev = ip_block->adev;
 
+	cancel_work_sync(&adev->gfx.userq_priv_fault_work);
+
 	for (i = 0; i < adev->gfx.num_gfx_rings; i++)
 		amdgpu_ring_fini(&adev->gfx.gfx_ring[i]);
 	for (i = 0; i < adev->gfx.num_compute_rings; i++)
@@ -6711,6 +6716,40 @@ static int gfx_v11_0_set_priv_inst_fault_state(struct amdgpu_device *adev,
 	return 0;
 }
 
+/*
+ * A gfx exception IV carries no doorbell. For each faulted slot, read the
+ * doorbell back from the HQD (left in place by the fatal fault), look up the
+ * user queue and kick its per-queue reset.
+ */
+static void gfx_v11_0_userq_priv_fault_work(struct work_struct *work)
+{
+	struct amdgpu_device *adev =
+		container_of(work, struct amdgpu_device, gfx.userq_priv_fault_work);
+	unsigned long slots = xchg(&adev->gfx.userq_priv_fault_slots, 0);
+	unsigned int id;
+
+	for_each_set_bit(id, &slots, BITS_PER_LONG) {
+		u8 pipe = id & 0x3;
+		u8 queue = (id >> 2) & 0x7;
+		struct amdgpu_usermode_queue *q;
+		u32 db_ctrl, doorbell;
+
+		amdgpu_gfx_off_ctrl(adev, false);
+		mutex_lock(&adev->srbm_mutex);
+		soc21_grbm_select(adev, 0, pipe, queue, 0);
+		db_ctrl = RREG32_SOC15(GC, 0, regCP_RB_DOORBELL_CONTROL);
+		soc21_grbm_select(adev, 0, 0, 0, 0);
+		mutex_unlock(&adev->srbm_mutex);
+		amdgpu_gfx_off_ctrl(adev, true);
+
+		doorbell = (db_ctrl & CP_RB_DOORBELL_CONTROL__DOORBELL_OFFSET_MASK) >>
+			   CP_RB_DOORBELL_CONTROL__DOORBELL_OFFSET__SHIFT;
+		q = xa_load(&adev->userq_doorbell_xa, doorbell);
+		if (q)
+			amdgpu_userq_start_hang_detect_work(q);
+	}
+}
+
 static void gfx_v11_0_handle_priv_fault(struct amdgpu_device *adev,
 					struct amdgpu_iv_entry *entry)
 {
Signature · 1 line
2.49.0
#2 lore raw

PATCH v4 3/5 drm/amdgpu/gfx12: add priv-fault user-queue recovery worker

In reply to Jesse Zhang

Mirror the gfx11 priv-fault user-queue recovery worker for GFX12, reading the doorbell back from the HQD via soc24_grbm_select.

The shared amdgpu_gfx_handle_priv_fault() helper schedules this worker for a gfx user-queue fault; wiring the helper up is done in a later patch.

v2:
- gate on adev->gfx.disable_uq instead of !adev->enable_mes (Alex)
- document why both the doorbell (compute) and HW-slot (gfx) reset paths are needed (Alex)

v3:
- rebase amd-staging-drm-next. adapt to the deduplicated amdgpu_gfx_handle_priv_fault() helper (commit d8ab7636160e); no functional change

---
 drivers/gpu/drm/amd/amdgpu/gfx_v12_0.c | 39 ++++++++++++++++++++++++++
 1 file changed, 39 insertions(+)
drivers/gpu/drm/amd/amdgpu/gfx_v12_0.c +39 -0
index 719441dbcc6b..1e5fd1ef8f1d 100644
@@ -286,6 +286,7 @@ static void gfx_v12_0_ring_invalidate_tlbs(struct amdgpu_ring *ring,
 					   bool all_hub, uint8_t dst_sel);
 static void gfx_v12_0_set_safe_mode(struct amdgpu_device *adev, int xcc_id);
 static void gfx_v12_0_unset_safe_mode(struct amdgpu_device *adev, int xcc_id);
+static void gfx_v12_0_userq_priv_fault_work(struct work_struct *work);
 static void gfx_v12_0_update_perf_clk(struct amdgpu_device *adev,
 				      bool enable);
 
@@ -1609,6 +1610,8 @@ static int gfx_v12_0_sw_init(struct amdgpu_ip_block *ip_block)
 
 	mutex_init(&adev->gfx.mec.reset_mutex);
 
+	INIT_WORK(&adev->gfx.userq_priv_fault_work, gfx_v12_0_userq_priv_fault_work);
+
 	return 0;
 }
 
@@ -1646,6 +1649,8 @@ static int gfx_v12_0_sw_fini(struct amdgpu_ip_block *ip_block)
 	int i;
 	struct amdgpu_device *adev = ip_block->adev;
 
+	cancel_work_sync(&adev->gfx.userq_priv_fault_work);
+
 	for (i = 0; i < adev->gfx.num_gfx_rings; i++)
 		amdgpu_ring_fini(&adev->gfx.gfx_ring[i]);
 	for (i = 0; i < adev->gfx.num_compute_rings; i++)
@@ -5048,6 +5053,40 @@ static int gfx_v12_0_set_priv_inst_fault_state(struct amdgpu_device *adev,
 	return 0;
 }
 
+/*
+ * A gfx exception IV carries no doorbell. For each faulted slot, read the
+ * doorbell back from the HQD (left in place by the fatal fault), look up the
+ * user queue and kick its per-queue reset.
+ */
+static void gfx_v12_0_userq_priv_fault_work(struct work_struct *work)
+{
+	struct amdgpu_device *adev =
+		container_of(work, struct amdgpu_device, gfx.userq_priv_fault_work);
+	unsigned long slots = xchg(&adev->gfx.userq_priv_fault_slots, 0);
+	unsigned int id;
+
+	for_each_set_bit(id, &slots, BITS_PER_LONG) {
+		u8 pipe = id & 0x3;
+		u8 queue = (id >> 2) & 0x7;
+		struct amdgpu_usermode_queue *q;
+		u32 db_ctrl, doorbell;
+
+		amdgpu_gfx_off_ctrl(adev, false);
+		mutex_lock(&adev->srbm_mutex);
+		soc24_grbm_select(adev, 0, pipe, queue, 0);
+		db_ctrl = RREG32_SOC15(GC, 0, regCP_RB_DOORBELL_CONTROL);
+		soc24_grbm_select(adev, 0, 0, 0, 0);
+		mutex_unlock(&adev->srbm_mutex);
+		amdgpu_gfx_off_ctrl(adev, true);
+
+		doorbell = (db_ctrl & CP_RB_DOORBELL_CONTROL__DOORBELL_OFFSET_MASK) >>
+			   CP_RB_DOORBELL_CONTROL__DOORBELL_OFFSET__SHIFT;
+		q = xa_load(&adev->userq_doorbell_xa, doorbell);
+		if (q)
+			amdgpu_userq_start_hang_detect_work(q);
+	}
+}
+
 static void gfx_v12_0_handle_priv_fault(struct amdgpu_device *adev,
 					struct amdgpu_iv_entry *entry)
 {
Signature · 1 line
2.49.0
#3 lore raw

PATCH v4 4/5 drm/amdgpu: recover user queues in the shared priv-fault helper

In reply to Jesse Zhang

If a priv/bad-op fault does not match a kernel queue slot, it belongs to a MES-scheduled user queue. Extend the shared amdgpu_gfx_handle_priv_fault() helper introduced by commit d8ab7636160e ("drm/amd/amdgpu: remove duplicated code in gfx_v11 and gfx_v12") to recover it: gate on adev->gfx.disable_uq, reset a compute user queue directly from its doorbell, and for a gfx user queue (whose IV carries no doorbell) record the HW slot and schedule the per-IP recovery worker.

v2:
- gate on adev->gfx.disable_uq instead of !adev->enable_mes (Alex)
- document why both the doorbell (compute) and HW-slot (gfx) reset paths are needed (Alex)

v3:
- rebase amd-staing-drm-next. adapt to the deduplicated amdgpu_gfx_handle_priv_fault() helper (commit d8ab7636160e); no functional change

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c | 20 +++++++++++++++++---
 1 file changed, 17 insertions(+), 3 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +17 -3
index 4d21d83451a4..a6f95ff47d24 100644
@@ -866,7 +866,8 @@ int amdgpu_gfx_enable_kgq(struct amdgpu_device *adev, int xcc_id)
  * @queue_id: queue ID of the faulty ring
  *
  * This function handles privileged instruction faults by identifying
- * the faulty ring (gfx or compute) and triggering a scheduler fault
+ * the faulty ring (gfx or compute) and triggering a scheduler fault, or by
+ * recovering the faulting user queue.
  */
 void amdgpu_gfx_handle_priv_fault(struct amdgpu_device *adev,
 					struct amdgpu_iv_entry *entry,
@@ -901,12 +902,25 @@ void amdgpu_gfx_handle_priv_fault(struct amdgpu_device *adev,
 		}
 	}
 
+	/* No KQ matched: the faulting slot belongs to a user queue. */
+	if (adev->gfx.disable_uq)
+		return;
+
 	doorbell_offset = entry->src_data[0] & AMDGPU_CTXID0_DOORBELL_ID_MASK;
 
-	/* No KQ matched: HW slot is a MES-scheduled user queue. */
-	if (adev->enable_mes && doorbell_offset)
+	/*
+	 * A compute user-queue fault IV carries the doorbell offset, so reset
+	 * the queue directly from it. A gfx user-queue fault is raised by the
+	 * ME and carries only the HW slot (no doorbell); record the slot and
+	 * let the worker read the doorbell back from the HQD.
+	 */
+	if (doorbell_offset) {
 		amdgpu_userq_process_reset_irq(adev, entry->pasid,
 					       doorbell_offset);
+	} else {
+		set_bit(pipe_id | (queue_id << 2), &adev->gfx.userq_priv_fault_slots);
+		schedule_work(&adev->gfx.userq_priv_fault_work);
+	}
 }
 
 static void amdgpu_gfx_do_off_ctrl(struct amdgpu_device *adev, bool enable,
Signature · 1 line
2.49.0
#4 lore raw

PATCH v4 5/5 drm/amdgpu/userq: mark a queue unmapped after a per-queue reset

In reply to Jesse Zhang

mes_userq_reset() unmaps the queue via the low-level mes_userq_unmap() (REMOVE_QUEUE) but does not update queue->state, so the queue still looks MAPPED. The destroy path then issues a second, redundant REMOVE_QUEUE for the already-removed queue; for gfx that unmap waits on an EOP that never arrives, times out (-110) and escalates to a full GPU reset.

Mark the queue UNMAPPED on a successful reset-path unmap so destroy skips the redundant REMOVE_QUEUE.

---
 drivers/gpu/drm/amd/amdgpu/mes_userqueue.c | 12 +++++++++++-
 1 file changed, 11 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/mes_userqueue.c +11 -1
index 639e4a37f872..b5816f9a8ffd 100644
@@ -26,6 +26,7 @@
 #include "amdgpu_gfx.h"
 #include "mes_userqueue.h"
 #include "amdgpu_userq_fence.h"
+#include "amdgpu_trace.h"
 
 #define AMDGPU_USERQ_PROC_CTX_SZ PAGE_SIZE
 #define AMDGPU_USERQ_GANG_CTX_SZ PAGE_SIZE
@@ -205,7 +206,16 @@ int mes_userq_reset(struct amdgpu_usermode_queue *queue)
 	amdgpu_mes_unlock(&adev->mes);
 	if (r)
 		return r;
-	return mes_userq_unmap(queue);
+
+	/* mes_userq_unmap() does not update queue->state; mark it UNMAPPED so the
+	 * destroy path does not issue a second REMOVE_QUEUE for the removed queue.
+	 */
+	r = mes_userq_unmap(queue);
+	if (!r) {
+		trace_amdgpu_userq_state_changed(queue, AMDGPU_USERQ_STATE_UNMAPPED);
+		queue->state = AMDGPU_USERQ_STATE_UNMAPPED;
+	}
+	return r;
 }
 
 int mes_userq_reset_queue(struct amdgpu_device *adev,
Signature · 1 line
2.49.0