Skip to contentSkip to conversation

drm/amdgpu: track faulted gfx user-queue slots

PATCH Superseded Jesse Zhang 2 participants 5 messages Jul 30

Subsystems (2)

v3 of 4 Superseded Latest is v4 →

lore
Share & tools
lore
#0 lore raw

PATCH v3 1/4 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 aefd4f03b443..f83afa7e0aa5 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 v3 2/4 drm/amdgpu/gfx11: recover gfx user queues on priv-fault

If a priv/bad-op fault does not match a kernel queue slot, it belongs to a MES-scheduled user queue. Compute IVs carry the doorbell and use the existing path; gfx IVs do not, so record the faulted slot and let a worker read the doorbell back from the HQD (regCP_RB_DOORBELL_CONTROL), look up the user queue and reset it.

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)

---
 drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c | 62 +++++++++++++++++++++++---
 1 file changed, 57 insertions(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c +57 -5
index 0cbbdc3694f4..4c1ba5a9c75d 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);
 
@@ -1931,6 +1932,8 @@ 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;
 }
 
@@ -1968,6 +1971,8 @@ 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++)
@@ -6707,10 +6712,47 @@ 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)
 {
 	u32 doorbell_offset = entry->src_data[0] & AMDGPU_CTXID0_DOORBELL_ID_MASK;
+	u8 me_id = (entry->ring_id & 0x0c) >> 2;
+	u8 pipe_id = (entry->ring_id & 0x03) >> 0;
+	u8 queue_id = (entry->ring_id & 0x70) >> 4;
 
 	/*
 	 * Try KQ first by ring_id (HW slot is authoritative). The
@@ -6718,9 +6760,6 @@ static void gfx_v11_0_handle_priv_fault(struct amdgpu_device *adev,
 	 * never share a HW slot.
 	 */
 	if (!adev->gfx.disable_kq) {
-		u8 me_id = (entry->ring_id & 0x0c) >> 2;
-		u8 pipe_id = (entry->ring_id & 0x03) >> 0;
-		u8 queue_id = (entry->ring_id & 0x70) >> 4;
 		struct amdgpu_ring *ring;
 		int i;
 
@@ -6752,10 +6791,23 @@ static void gfx_v11_0_handle_priv_fault(struct amdgpu_device *adev,
 		}
 	}
 
-	/* No KQ matched: HW slot is a MES-scheduled user queue. */
-	if (adev->enable_mes && doorbell_offset)
+	/* No KQ matched: the faulting slot belongs to a user queue. */
+	if (adev->gfx.disable_uq)
+		return;
+
+	/*
+	 * 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 int gfx_v11_0_priv_reg_irq(struct amdgpu_device *adev,
Signature · 1 line
2.49.0
#2 lore raw

PATCH v3 3/4 drm/amdgpu/gfx12: recover gfx user queues on priv-fault

In reply to Jesse Zhang

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

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)

---
 drivers/gpu/drm/amd/amdgpu/gfx_v12_0.c | 64 +++++++++++++++++++++++---
 1 file changed, 57 insertions(+), 7 deletions(-)
drivers/gpu/drm/amd/amdgpu/gfx_v12_0.c +57 -7
index b3887c5262a0..8fe991787974 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);
 
@@ -1608,6 +1609,8 @@ 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;
 }
 
@@ -1645,6 +1648,8 @@ 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++)
@@ -5038,24 +5043,56 @@ 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)
 {
 	u32 doorbell_offset = entry->src_data[0] & AMDGPU_CTXID0_DOORBELL_ID_MASK;
+	u8 me_id = (entry->ring_id & 0x0c) >> 2;
+	u8 pipe_id = (entry->ring_id & 0x03) >> 0;
+	u8 queue_id = (entry->ring_id & 0x70) >> 4;
 
 	/*
 	 * Try KQ first by ring_id; UQ as fallback. KCQ and UQ never share
 	 * a HW slot (compute_hqd_mask contract).
 	 */
 	if (!adev->gfx.disable_kq) {
-		u8 me_id, pipe_id, queue_id;
 		struct amdgpu_ring *ring;
 		int i;
 
-		me_id = (entry->ring_id & 0x0c) >> 2;
-		pipe_id = (entry->ring_id & 0x03) >> 0;
-		queue_id = (entry->ring_id & 0x70) >> 4;
-
 		switch (me_id) {
 		case 0:
 			for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
@@ -5084,10 +5121,23 @@ static void gfx_v12_0_handle_priv_fault(struct amdgpu_device *adev,
 		}
 	}
 
-	/* No KQ matched: HW slot is a MES-scheduled user queue. */
-	if (adev->enable_mes && doorbell_offset)
+	/* No KQ matched: the faulting slot belongs to a user queue. */
+	if (adev->gfx.disable_uq)
+		return;
+
+	/*
+	 * 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 int gfx_v12_0_priv_reg_irq(struct amdgpu_device *adev,
Signature · 1 line
2.49.0
#3 lore raw

PATCH v3 4/4 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
Reviewed-by on v3 4/4
#4 lore raw

Re: [PATCH v3 4/4] drm/amdgpu/userq: mark a queue unmapped after a per-queue reset

Quoted 11 lines · 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.

Series is:

Quoted 37 lines
---
 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 -2
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,
--

2.49.0