drm/amdkfd: Update SDMA queue management for GFX9.4.3

This patch updates SDMA queue management for multi XCC in GFX9.4.3.
- Allocate/deallocate SDMA queues from the correct SDMA engines
  based on the partition mode.
- Updates the kgd2kfd interface to fetch the correct SDMA register
  addresses.
- It also fixes dumping correct SDMA queue info in debugfs.

v2: squash in fix "drm/amdkfd: Fix XGMI SDMA user-mode queue allocation"
Signed-off-by: Mukul Joshi <mukul.joshi@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gc_9_4_3.c

@@ -31,6 +31,192 @@
 #include "oss/osssys_4_0_sh_mask.h"
 #include "v9_structs.h"
 #include "soc15.h"
+#include "sdma/sdma_4_4_2_offset.h"
+#include "sdma/sdma_4_4_2_sh_mask.h"
+
+static inline struct v9_sdma_mqd *get_sdma_mqd(void *mqd)
+{
+	return (struct v9_sdma_mqd *)mqd;
+}
+
+static uint32_t get_sdma_rlc_reg_offset(struct amdgpu_device *adev,
+					unsigned int engine_id,
+					unsigned int queue_id)
+{
+	uint32_t sdma_engine_reg_base = SOC15_REG_OFFSET(SDMA0, engine_id,
+					regSDMA_RLC0_RB_CNTL) -
+					regSDMA_RLC0_RB_CNTL;
+	uint32_t retval = sdma_engine_reg_base +
+		  queue_id * (regSDMA_RLC1_RB_CNTL - regSDMA_RLC0_RB_CNTL);
+
+	pr_debug("RLC register offset for SDMA%d RLC%d: 0x%x\n", engine_id,
+							queue_id, retval);
+	return retval;
+}
+
+int kgd_gfx_v9_4_3_hqd_sdma_load(struct amdgpu_device *adev, void *mqd,
+				 uint32_t __user *wptr, struct mm_struct *mm)
+{
+	struct v9_sdma_mqd *m;
+	uint32_t sdma_rlc_reg_offset;
+	unsigned long end_jiffies;
+	uint32_t data;
+	uint64_t data64;
+	uint64_t __user *wptr64 = (uint64_t __user *)wptr;
+
+	m = get_sdma_mqd(mqd);
+	sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id,
+							m->sdma_queue_id);
+
+	WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_CNTL,
+		m->sdmax_rlcx_rb_cntl & (~SDMA_RLC0_RB_CNTL__RB_ENABLE_MASK));
+
+	end_jiffies = msecs_to_jiffies(2000) + jiffies;
+	while (true) {
+		data = RREG32(sdma_rlc_reg_offset + regSDMA_RLC0_CONTEXT_STATUS);
+		if (data & SDMA_RLC0_CONTEXT_STATUS__IDLE_MASK)
+			break;
+		if (time_after(jiffies, end_jiffies)) {
+			pr_err("SDMA RLC not idle in %s\n", __func__);
+			return -ETIME;
+		}
+		usleep_range(500, 1000);
+	}
+
+	WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_DOORBELL_OFFSET,
+		m->sdmax_rlcx_doorbell_offset);
+
+	data = REG_SET_FIELD(m->sdmax_rlcx_doorbell, SDMA_RLC0_DOORBELL,
+				ENABLE, 1);
+	WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_DOORBELL, data);
+	WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_RPTR,
+					m->sdmax_rlcx_rb_rptr);
+	WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_RPTR_HI,
+					m->sdmax_rlcx_rb_rptr_hi);
+
+	WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_MINOR_PTR_UPDATE, 1);
+	if (read_user_wptr(mm, wptr64, data64)) {
+		WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_WPTR,
+			lower_32_bits(data64));
+		WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_WPTR_HI,
+			upper_32_bits(data64));
+	} else {
+		WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_WPTR,
+			m->sdmax_rlcx_rb_rptr);
+		WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_WPTR_HI,
+			m->sdmax_rlcx_rb_rptr_hi);
+	}
+	WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_MINOR_PTR_UPDATE, 0);
+
+	WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_BASE, m->sdmax_rlcx_rb_base);
+	WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_BASE_HI,
+			m->sdmax_rlcx_rb_base_hi);
+	WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_RPTR_ADDR_LO,
+			m->sdmax_rlcx_rb_rptr_addr_lo);
+	WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_RPTR_ADDR_HI,
+			m->sdmax_rlcx_rb_rptr_addr_hi);
+
+	data = REG_SET_FIELD(m->sdmax_rlcx_rb_cntl, SDMA_RLC0_RB_CNTL,
+				RB_ENABLE, 1);
+	WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_CNTL, data);
+
+	return 0;
+}
+
+int kgd_gfx_v9_4_3_hqd_sdma_dump(struct amdgpu_device *adev,
+				 uint32_t engine_id, uint32_t queue_id,
+				 uint32_t (**dump)[2], uint32_t *n_regs)
+{
+	uint32_t sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev,
+							engine_id, queue_id);
+	uint32_t i = 0, reg;
+#undef HQD_N_REGS
+#define HQD_N_REGS (19+6+7+12)
+#define DUMP_REG(addr) do {				\
+		if (WARN_ON_ONCE(i >= HQD_N_REGS))      \
+			break;				\
+		(*dump)[i][0] = (addr) << 2;            \
+		(*dump)[i++][1] = RREG32(addr);         \
+	} while (0)
+
+	*dump = kmalloc_array(HQD_N_REGS * 2, sizeof(uint32_t), GFP_KERNEL);
+	if (*dump == NULL)
+		return -ENOMEM;
+
+	for (reg = regSDMA_RLC0_RB_CNTL; reg <= regSDMA_RLC0_DOORBELL; reg++)
+		DUMP_REG(sdma_rlc_reg_offset + reg);
+	for (reg = regSDMA_RLC0_STATUS; reg <= regSDMA_RLC0_CSA_ADDR_HI; reg++)
+		DUMP_REG(sdma_rlc_reg_offset + reg);
+	for (reg = regSDMA_RLC0_IB_SUB_REMAIN;
+	     reg <= regSDMA_RLC0_MINOR_PTR_UPDATE; reg++)
+		DUMP_REG(sdma_rlc_reg_offset + reg);
+	for (reg = regSDMA_RLC0_MIDCMD_DATA0;
+	     reg <= regSDMA_RLC0_MIDCMD_CNTL; reg++)
+		DUMP_REG(sdma_rlc_reg_offset + reg);
+
+	WARN_ON_ONCE(i != HQD_N_REGS);
+	*n_regs = i;
+
+	return 0;
+}
+
+bool kgd_gfx_v9_4_3_hqd_sdma_is_occupied(struct amdgpu_device *adev, void *mqd)
+{
+	struct v9_sdma_mqd *m;
+	uint32_t sdma_rlc_reg_offset;
+	uint32_t sdma_rlc_rb_cntl;
+
+	m = get_sdma_mqd(mqd);
+	sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id,
+							m->sdma_queue_id);
+
+	sdma_rlc_rb_cntl = RREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_CNTL);
+
+	if (sdma_rlc_rb_cntl & SDMA_RLC0_RB_CNTL__RB_ENABLE_MASK)
+		return true;
+
+	return false;
+}
+
+int kgd_gfx_v9_4_3_hqd_sdma_destroy(struct amdgpu_device *adev, void *mqd,
+				    unsigned int utimeout)
+{
+	struct v9_sdma_mqd *m;
+	uint32_t sdma_rlc_reg_offset;
+	uint32_t temp;
+	unsigned long end_jiffies = (utimeout * HZ / 1000) + jiffies;
+
+	m = get_sdma_mqd(mqd);
+	sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id,
+							m->sdma_queue_id);
+
+	temp = RREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_CNTL);
+	temp = temp & ~SDMA_RLC0_RB_CNTL__RB_ENABLE_MASK;
+	WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_CNTL, temp);
+
+	while (true) {
+		temp = RREG32(sdma_rlc_reg_offset + regSDMA_RLC0_CONTEXT_STATUS);
+		if (temp & SDMA_RLC0_CONTEXT_STATUS__IDLE_MASK)
+			break;
+		if (time_after(jiffies, end_jiffies)) {
+			pr_err("SDMA RLC not idle in %s\n", __func__);
+			return -ETIME;
+		}
+		usleep_range(500, 1000);
+	}
+
+	WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_DOORBELL, 0);
+	WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_CNTL,
+		RREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_CNTL) |
+		SDMA_RLC0_RB_CNTL__RB_ENABLE_MASK);
+
+	m->sdmax_rlcx_rb_rptr =
+			RREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_RPTR);
+	m->sdmax_rlcx_rb_rptr_hi =
+			RREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_RPTR_HI);
+
+	return 0;
+}
 
 static int kgd_gfx_v9_4_3_set_pasid_vmid_mapping(struct amdgpu_device *adev,
 				u32 pasid, unsigned int vmid, uint32_t inst)
@@ -166,13 +352,13 @@ const struct kfd2kgd_calls gc_9_4_3_kfd2kgd = {
 	.init_interrupts = kgd_gfx_v9_init_interrupts,
 	.hqd_load = kgd_gfx_v9_4_3_hqd_load,
 	.hiq_mqd_load = kgd_gfx_v9_hiq_mqd_load,
-	.hqd_sdma_load = kgd_arcturus_hqd_sdma_load,
+	.hqd_sdma_load = kgd_gfx_v9_4_3_hqd_sdma_load,
 	.hqd_dump = kgd_gfx_v9_hqd_dump,
-	.hqd_sdma_dump = kgd_arcturus_hqd_sdma_dump,
+	.hqd_sdma_dump = kgd_gfx_v9_4_3_hqd_sdma_dump,
 	.hqd_is_occupied = kgd_gfx_v9_hqd_is_occupied,
-	.hqd_sdma_is_occupied = kgd_arcturus_hqd_sdma_is_occupied,
+	.hqd_sdma_is_occupied = kgd_gfx_v9_4_3_hqd_sdma_is_occupied,
 	.hqd_destroy = kgd_gfx_v9_hqd_destroy,
-	.hqd_sdma_destroy = kgd_arcturus_hqd_sdma_destroy,
+	.hqd_sdma_destroy = kgd_gfx_v9_4_3_hqd_sdma_destroy,
 	.wave_control_execute = kgd_gfx_v9_wave_control_execute,
 	.get_atc_vmid_pasid_mapping_info =
 				kgd_gfx_v9_get_atc_vmid_pasid_mapping_info,

drivers/gpu/drm/amd/amdkfd/kfd_device.c

@@ -741,6 +741,7 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd,
 		if (!node)
 			goto node_alloc_error;
 
+		node->node_id = i;
 		node->adev = kfd->adev;
 		node->kfd = kfd;
 		node->kfd2kgd = kfd->kfd2kgd;
@@ -1323,15 +1324,16 @@ unsigned int kfd_get_num_sdma_engines(struct kfd_node *node)
 {
 	/* If XGMI is not supported, all SDMA engines are PCIe */
 	if (!node->adev->gmc.xgmi.supported)
-		return node->adev->sdma.num_instances;
+		return node->adev->sdma.num_instances/(int)node->kfd->num_nodes;
 
-	return min(node->adev->sdma.num_instances, 2);
+	return min(node->adev->sdma.num_instances/(int)node->kfd->num_nodes, 2);
 }
 
 unsigned int kfd_get_num_xgmi_sdma_engines(struct kfd_node *node)
 {
 	/* After reserved for PCIe, the rest of engines are XGMI */
-	return node->adev->sdma.num_instances - kfd_get_num_sdma_engines(node);
+	return node->adev->sdma.num_instances/(int)node->kfd->num_nodes -
+		kfd_get_num_sdma_engines(node);
 }
 
 #if defined(CONFIG_DEBUG_FS)

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c

@@ -124,6 +124,15 @@ static inline uint64_t get_reserved_sdma_queues_bitmap(struct device_queue_manag
 	return dqm->dev->kfd->device_info.reserved_sdma_queues_bitmap;
 }
 
+static void init_sdma_bitmaps(struct device_queue_manager *dqm)
+{
+	bitmap_zero(dqm->sdma_bitmap, KFD_MAX_SDMA_QUEUES);
+	bitmap_set(dqm->sdma_bitmap, 0, get_num_sdma_queues(dqm));
+
+	bitmap_zero(dqm->xgmi_sdma_bitmap, KFD_MAX_SDMA_QUEUES);
+	bitmap_set(dqm->xgmi_sdma_bitmap, 0, get_num_xgmi_sdma_queues(dqm));
+}
+
 void program_sh_mem_settings(struct device_queue_manager *dqm,
 					struct qcm_process_device *qpd)
 {
@@ -1268,24 +1277,6 @@ static void init_interrupts(struct device_queue_manager *dqm)
 	}
 }
 
-static void init_sdma_bitmaps(struct device_queue_manager *dqm)
-{
-	unsigned int num_sdma_queues =
-		min_t(unsigned int, sizeof(dqm->sdma_bitmap)*8,
-		      get_num_sdma_queues(dqm));
-	unsigned int num_xgmi_sdma_queues =
-		min_t(unsigned int, sizeof(dqm->xgmi_sdma_bitmap)*8,
-		      get_num_xgmi_sdma_queues(dqm));
-
-	if (num_sdma_queues)
-		dqm->sdma_bitmap = GENMASK_ULL(num_sdma_queues-1, 0);
-	if (num_xgmi_sdma_queues)
-		dqm->xgmi_sdma_bitmap = GENMASK_ULL(num_xgmi_sdma_queues-1, 0);
-
-	dqm->sdma_bitmap &= ~get_reserved_sdma_queues_bitmap(dqm);
-	pr_info("sdma_bitmap: %llx\n", dqm->sdma_bitmap);
-}
-
 static int initialize_nocpsch(struct device_queue_manager *dqm)
 {
 	int pipe, queue;
@@ -1375,46 +1366,49 @@ static int allocate_sdma_queue(struct device_queue_manager *dqm,
 	int bit;
 
 	if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
-		if (dqm->sdma_bitmap == 0) {
+		if (bitmap_empty(dqm->sdma_bitmap, KFD_MAX_SDMA_QUEUES)) {
 			pr_err("No more SDMA queue to allocate\n");
 			return -ENOMEM;
 		}
 
 		if (restore_sdma_id) {
 			/* Re-use existing sdma_id */
-			if (!(dqm->sdma_bitmap & (1ULL << *restore_sdma_id))) {
+			if (!test_bit(*restore_sdma_id, dqm->sdma_bitmap)) {
 				pr_err("SDMA queue already in use\n");
 				return -EBUSY;
 			}
-			dqm->sdma_bitmap &= ~(1ULL << *restore_sdma_id);
+			clear_bit(*restore_sdma_id, dqm->sdma_bitmap);
 			q->sdma_id = *restore_sdma_id;
 		} else {
 			/* Find first available sdma_id */
-			bit = __ffs64(dqm->sdma_bitmap);
-			dqm->sdma_bitmap &= ~(1ULL << bit);
+			bit = find_first_bit(dqm->sdma_bitmap,
+					     get_num_sdma_queues(dqm));
+			clear_bit(bit, dqm->sdma_bitmap);
 			q->sdma_id = bit;
 		}
 
-		q->properties.sdma_engine_id = q->sdma_id %
-				kfd_get_num_sdma_engines(dqm->dev);
+		q->properties.sdma_engine_id =
+			dqm->dev->node_id * get_num_all_sdma_engines(dqm) +
+			q->sdma_id % kfd_get_num_sdma_engines(dqm->dev);
 		q->properties.sdma_queue_id = q->sdma_id /
 				kfd_get_num_sdma_engines(dqm->dev);
 	} else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
-		if (dqm->xgmi_sdma_bitmap == 0) {
+		if (bitmap_empty(dqm->xgmi_sdma_bitmap, KFD_MAX_SDMA_QUEUES)) {
 			pr_err("No more XGMI SDMA queue to allocate\n");
 			return -ENOMEM;
 		}
 		if (restore_sdma_id) {
 			/* Re-use existing sdma_id */
-			if (!(dqm->xgmi_sdma_bitmap & (1ULL << *restore_sdma_id))) {
+			if (!test_bit(*restore_sdma_id, dqm->xgmi_sdma_bitmap)) {
 				pr_err("SDMA queue already in use\n");
 				return -EBUSY;
 			}
-			dqm->xgmi_sdma_bitmap &= ~(1ULL << *restore_sdma_id);
+			clear_bit(*restore_sdma_id, dqm->xgmi_sdma_bitmap);
 			q->sdma_id = *restore_sdma_id;
 		} else {
-			bit = __ffs64(dqm->xgmi_sdma_bitmap);
-			dqm->xgmi_sdma_bitmap &= ~(1ULL << bit);
+			bit = find_first_bit(dqm->xgmi_sdma_bitmap,
+					     get_num_xgmi_sdma_queues(dqm));
+			clear_bit(bit, dqm->xgmi_sdma_bitmap);
 			q->sdma_id = bit;
 		}
 		/* sdma_engine_id is sdma id including
@@ -1424,6 +1418,7 @@ static int allocate_sdma_queue(struct device_queue_manager *dqm,
 		 * PCIe-optimized ones
 		 */
 		q->properties.sdma_engine_id =
+			dqm->dev->node_id * get_num_all_sdma_engines(dqm) +
 			kfd_get_num_sdma_engines(dqm->dev) +
 			q->sdma_id % kfd_get_num_xgmi_sdma_engines(dqm->dev);
 		q->properties.sdma_queue_id = q->sdma_id /
@@ -1442,11 +1437,11 @@ static void deallocate_sdma_queue(struct device_queue_manager *dqm,
 	if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
 		if (q->sdma_id >= get_num_sdma_queues(dqm))
 			return;
-		dqm->sdma_bitmap |= (1ULL << q->sdma_id);
+		set_bit(q->sdma_id, dqm->sdma_bitmap);
 	} else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
 		if (q->sdma_id >= get_num_xgmi_sdma_queues(dqm))
 			return;
-		dqm->xgmi_sdma_bitmap |= (1ULL << q->sdma_id);
+		set_bit(q->sdma_id, dqm->xgmi_sdma_bitmap);
 	}
 }
 

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h

@@ -239,8 +239,8 @@ struct device_queue_manager {
 	unsigned int		total_queue_count;
 	unsigned int		next_pipe_to_allocate;
 	unsigned int		*allocated_queues;
-	uint64_t		sdma_bitmap;
-	uint64_t		xgmi_sdma_bitmap;
+	DECLARE_BITMAP(sdma_bitmap, KFD_MAX_SDMA_QUEUES);
+	DECLARE_BITMAP(xgmi_sdma_bitmap, KFD_MAX_SDMA_QUEUES);
 	/* the pasid mapping for each kfd vmid */
 	uint16_t		vmid_pasid[VMID_NUM];
 	uint64_t		pipelines_addr;

drivers/gpu/drm/amd/amdkfd/kfd_priv.h

@@ -113,6 +113,8 @@
 
 #define KFD_UNMAP_LATENCY_MS	(4000)
 
+#define KFD_MAX_SDMA_QUEUES	128
+
 /*
  * 512 = 0x200
  * The doorbell index distance between SDMA RLC (2*i) and (2*i+1) in the
@@ -260,6 +262,7 @@ struct kfd_vmid_info {
 struct kfd_dev;
 
 struct kfd_node {
+	unsigned int node_id;
 	struct amdgpu_device *adev;     /* Duplicated here along with keeping
 					 * a copy in kfd_dev to save a hop
 					 */