drm/amdgpu: add sysfs node for compute partition mode
authorLe Ma <le.ma@amd.com>
Tue, 17 May 2022 14:20:10 +0000 (22:20 +0800)
committerAlex Deucher <alexander.deucher@amd.com>
Fri, 9 Jun 2023 13:40:25 +0000 (09:40 -0400)
Add current/available compute partitin mode sysfs node.

v2: make the sysfs node as IP independent one in amdgpu_gfx.c

Signed-off-by: Le Ma <le.ma@amd.com>
Reviewed-by: Hawking Zhang <Hawking.Zhang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h
drivers/gpu/drm/amd/amdgpu/amdgpu_nbio.h
drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c
drivers/gpu/drm/amd/amdgpu/nbio_v7_9.c

index e94507a..f432064 100644 (file)
@@ -3634,6 +3634,7 @@ int amdgpu_device_init(struct amdgpu_device *adev,
        mutex_init(&adev->srbm_mutex);
        mutex_init(&adev->gfx.pipe_reserve_mutex);
        mutex_init(&adev->gfx.gfx_off_mutex);
+       mutex_init(&adev->gfx.partition_mutex);
        mutex_init(&adev->grbm_idx_mutex);
        mutex_init(&adev->mn_lock);
        mutex_init(&adev->virt.vf_errors.lock);
index e5ff0bf..4a4d71f 100644 (file)
@@ -1174,3 +1174,135 @@ bool amdgpu_gfx_is_master_xcc(struct amdgpu_device *adev, int xcc_id)
        return !(xcc_id % (adev->gfx.num_xcc_per_xcp ?
                        adev->gfx.num_xcc_per_xcp : 1));
 }
+
+static ssize_t amdgpu_gfx_get_current_compute_partition(struct device *dev,
+                                               struct device_attribute *addr,
+                                               char *buf)
+{
+       struct drm_device *ddev = dev_get_drvdata(dev);
+       struct amdgpu_device *adev = drm_to_adev(ddev);
+       enum amdgpu_gfx_partition mode;
+       char *partition_mode;
+
+       mode = adev->gfx.funcs->query_partition_mode(adev);
+
+       switch (mode) {
+       case AMDGPU_SPX_PARTITION_MODE:
+               partition_mode = "SPX";
+               break;
+       case AMDGPU_DPX_PARTITION_MODE:
+               partition_mode = "DPX";
+               break;
+       case AMDGPU_TPX_PARTITION_MODE:
+               partition_mode = "TPX";
+               break;
+       case AMDGPU_QPX_PARTITION_MODE:
+               partition_mode = "QPX";
+               break;
+       case AMDGPU_CPX_PARTITION_MODE:
+               partition_mode = "CPX";
+               break;
+       default:
+               partition_mode = "UNKNOWN";
+               break;
+       }
+
+       return sysfs_emit(buf, "%s\n", partition_mode);
+}
+
+static ssize_t amdgpu_gfx_set_compute_partition(struct device *dev,
+                                               struct device_attribute *addr,
+                                               const char *buf, size_t count)
+{
+       struct drm_device *ddev = dev_get_drvdata(dev);
+       struct amdgpu_device *adev = drm_to_adev(ddev);
+       enum amdgpu_gfx_partition mode;
+       int ret;
+
+       if (adev->gfx.num_xcd % 2 != 0)
+               return -EINVAL;
+
+       if (!strncasecmp("SPX", buf, strlen("SPX"))) {
+               mode = AMDGPU_SPX_PARTITION_MODE;
+       } else if (!strncasecmp("DPX", buf, strlen("DPX"))) {
+               if (adev->gfx.num_xcd != 4 || adev->gfx.num_xcd != 8)
+                       return -EINVAL;
+               mode = AMDGPU_DPX_PARTITION_MODE;
+       } else if (!strncasecmp("TPX", buf, strlen("TPX"))) {
+               if (adev->gfx.num_xcd != 6)
+                       return -EINVAL;
+               mode = AMDGPU_TPX_PARTITION_MODE;
+       } else if (!strncasecmp("QPX", buf, strlen("QPX"))) {
+               if (adev->gfx.num_xcd != 8)
+                       return -EINVAL;
+               mode = AMDGPU_QPX_PARTITION_MODE;
+       } else if (!strncasecmp("CPX", buf, strlen("CPX"))) {
+               mode = AMDGPU_CPX_PARTITION_MODE;
+       } else {
+               return -EINVAL;
+       }
+
+       mutex_lock(&adev->gfx.partition_mutex);
+
+       ret = adev->gfx.funcs->switch_partition_mode(adev, mode);
+
+       mutex_unlock(&adev->gfx.partition_mutex);
+
+       if (ret)
+               return ret;
+
+       return count;
+}
+
+static ssize_t amdgpu_gfx_get_available_compute_partition(struct device *dev,
+                                               struct device_attribute *addr,
+                                               char *buf)
+{
+       struct drm_device *ddev = dev_get_drvdata(dev);
+       struct amdgpu_device *adev = drm_to_adev(ddev);
+       char *supported_partition;
+
+       /* TBD */
+       switch (adev->gfx.num_xcd) {
+       case 8:
+               supported_partition = "SPX, DPX, QPX, CPX";
+               break;
+       case 6:
+               supported_partition = "SPX, TPX, CPX";
+               break;
+       case 4:
+               supported_partition = "SPX, DPX, CPX";
+               break;
+       /* this seems only existing in emulation phase */
+       case 2:
+               supported_partition = "SPX, CPX";
+               break;
+       default:
+               supported_partition = "Not supported";
+               break;
+       }
+
+       return sysfs_emit(buf, "%s\n", supported_partition);
+}
+
+static DEVICE_ATTR(current_compute_partition, S_IRUGO | S_IWUSR,
+                  amdgpu_gfx_get_current_compute_partition,
+                  amdgpu_gfx_set_compute_partition);
+
+static DEVICE_ATTR(available_compute_partition, S_IRUGO,
+                  amdgpu_gfx_get_available_compute_partition, NULL);
+
+int amdgpu_gfx_sysfs_init(struct amdgpu_device *adev)
+{
+       int r;
+
+       r = device_create_file(adev->dev, &dev_attr_current_compute_partition);
+       if (r)
+               return r;
+
+       r = device_create_file(adev->dev, &dev_attr_available_compute_partition);
+       if (r)
+               return r;
+
+       return 0;
+}
index 2755f00..8df3652 100644 (file)
@@ -259,6 +259,10 @@ struct amdgpu_gfx_funcs {
        void (*update_perfmon_mgcg)(struct amdgpu_device *adev, bool enable);
        int (*get_gfx_shadow_info)(struct amdgpu_device *adev,
                                   struct amdgpu_gfx_shadow_info *shadow_info);
+       enum amdgpu_gfx_partition
+                       (*query_partition_mode)(struct amdgpu_device *adev);
+       int (*switch_partition_mode)(struct amdgpu_device *adev,
+                                    enum amdgpu_gfx_partition mode);
 };
 
 struct sq_work {
@@ -394,6 +398,7 @@ struct amdgpu_gfx {
        enum amdgpu_gfx_partition       partition_mode;
        uint32_t                        num_xcd;
        uint32_t                        num_xcc_per_xcp;
+       struct mutex                    partition_mutex;
 };
 
 #define amdgpu_gfx_get_gpu_clock_counter(adev) (adev)->gfx.funcs->get_gpu_clock_counter((adev))
@@ -478,4 +483,5 @@ int amdgpu_gfx_poison_consumption_handler(struct amdgpu_device *adev,
                                                struct amdgpu_iv_entry *entry);
 
 bool amdgpu_gfx_is_master_xcc(struct amdgpu_device *adev, int xcc_id);
+int amdgpu_gfx_sysfs_init(struct amdgpu_device *adev);
 #endif
index c686ff4..6a636c3 100644 (file)
@@ -95,6 +95,9 @@ struct amdgpu_nbio_funcs {
        void (*apply_l1_link_width_reconfig_wa)(struct amdgpu_device *adev);
        void (*clear_doorbell_interrupt)(struct amdgpu_device *adev);
        u32 (*get_rom_offset)(struct amdgpu_device *adev);
+       u32 (*get_compute_partition_mode)(struct amdgpu_device *adev);
+       void (*set_compute_partition_mode)(struct amdgpu_device *adev,
+                                          enum amdgpu_gfx_partition mode);
 };
 
 struct amdgpu_nbio {
index 064cd02..1d15db9 100644 (file)
@@ -662,6 +662,67 @@ static void gfx_v9_4_3_select_me_pipe_q(struct amdgpu_device *adev,
        soc15_grbm_select(adev, me, pipe, q, vm, 0);
 }
 
+static enum amdgpu_gfx_partition
+gfx_v9_4_3_query_compute_partition(struct amdgpu_device *adev)
+{
+       enum amdgpu_gfx_partition mode = AMDGPU_UNKNOWN_COMPUTE_PARTITION_MODE;
+
+       if (adev->nbio.funcs->get_compute_partition_mode)
+               mode = adev->nbio.funcs->get_compute_partition_mode(adev);
+
+       return mode;
+}
+
+static int gfx_v9_4_3_switch_compute_partition(struct amdgpu_device *adev,
+                                               enum amdgpu_gfx_partition mode)
+{
+       u32 tmp = 0;
+       int num_xcc_per_partition, i;
+
+       if (mode == adev->gfx.partition_mode)
+               return mode;
+
+       switch (mode) {
+       case AMDGPU_SPX_PARTITION_MODE:
+               num_xcc_per_partition = adev->gfx.num_xcd;
+               break;
+       case AMDGPU_DPX_PARTITION_MODE:
+               num_xcc_per_partition = adev->gfx.num_xcd / 2;
+               break;
+       case AMDGPU_TPX_PARTITION_MODE:
+               num_xcc_per_partition = adev->gfx.num_xcd / 3;
+               break;
+       case AMDGPU_QPX_PARTITION_MODE:
+               num_xcc_per_partition = adev->gfx.num_xcd / 4;
+               break;
+       case AMDGPU_CPX_PARTITION_MODE:
+               num_xcc_per_partition = 1;
+               break;
+       default:
+               return -EINVAL;
+       }
+
+       /* TODO:
+        * Stop user queues and threads, and make sure GPU is empty of work.
+        */
+
+       for (i = 0; i < adev->gfx.num_xcd; i++) {
+               tmp = REG_SET_FIELD(tmp, CP_HYP_XCP_CTL, NUM_XCC_IN_XCP,
+                                   num_xcc_per_partition);
+               tmp = REG_SET_FIELD(tmp, CP_HYP_XCP_CTL, VIRTUAL_XCC_ID,
+                                   i % num_xcc_per_partition);
+               WREG32_SOC15(GC, i, regCP_HYP_XCP_CTL, tmp);
+       }
+
+       if (adev->nbio.funcs->set_compute_partition_mode)
+               adev->nbio.funcs->set_compute_partition_mode(adev, mode);
+
+       adev->gfx.num_xcc_per_xcp = num_xcc_per_partition;
+       adev->gfx.partition_mode = mode;
+
+       return 0;
+}
+
 static const struct amdgpu_gfx_funcs gfx_v9_4_3_gfx_funcs = {
        .get_gpu_clock_counter = &gfx_v9_4_3_get_gpu_clock_counter,
        .select_se_sh = &gfx_v9_4_3_select_se_sh,
@@ -669,6 +730,8 @@ static const struct amdgpu_gfx_funcs gfx_v9_4_3_gfx_funcs = {
        .read_wave_sgprs = &gfx_v9_4_3_read_wave_sgprs,
        .read_wave_vgprs = &gfx_v9_4_3_read_wave_vgprs,
        .select_me_pipe_q = &gfx_v9_4_3_select_me_pipe_q,
+       .query_partition_mode = &gfx_v9_4_3_query_compute_partition,
+       .switch_partition_mode = &gfx_v9_4_3_switch_compute_partition,
 };
 
 static int gfx_v9_4_3_gpu_early_init(struct amdgpu_device *adev)
@@ -858,6 +921,10 @@ static int gfx_v9_4_3_sw_init(void *handle)
        if (r)
                return r;
 
+       r = amdgpu_gfx_sysfs_init(adev);
+       if (r)
+               return r;
+
        return 0;
 }
 
index a331a59..bdb84a5 100644 (file)
@@ -347,6 +347,36 @@ static void nbio_v7_9_enable_doorbell_interrupt(struct amdgpu_device *adev,
                              DOORBELL_INTERRUPT_DISABLE, enable ? 0 : 1);
 }
 
+static enum amdgpu_gfx_partition nbio_v7_9_get_compute_partition_mode(struct amdgpu_device *adev)
+{
+       u32 tmp;
+
+       tmp = RREG32_SOC15(NBIO, 0, regBIF_BX_PF0_PARTITION_COMPUTE_CAP);
+
+       if (REG_GET_FIELD(tmp, BIF_BX_PF0_PARTITION_COMPUTE_CAP, SPX_SUPPORT))
+               return AMDGPU_SPX_PARTITION_MODE;
+       else if (REG_GET_FIELD(tmp, BIF_BX_PF0_PARTITION_COMPUTE_CAP, DPX_SUPPORT))
+               return AMDGPU_DPX_PARTITION_MODE;
+       else if (REG_GET_FIELD(tmp, BIF_BX_PF0_PARTITION_COMPUTE_CAP, TPX_SUPPORT))
+               return AMDGPU_TPX_PARTITION_MODE;
+       else if (REG_GET_FIELD(tmp, BIF_BX_PF0_PARTITION_COMPUTE_CAP, CPX_SUPPORT))
+               return AMDGPU_CPX_PARTITION_MODE;
+       else
+               return AMDGPU_UNKNOWN_COMPUTE_PARTITION_MODE;
+}
+
+static void nbio_v7_9_set_compute_partition_mode(struct amdgpu_device *adev,
+                                       enum amdgpu_gfx_partition mode)
+{
+       u32 tmp;
+
+       tmp = RREG32_SOC15(NBIO, 0, regBIF_BX_PF0_PARTITION_COMPUTE_CAP);
+       tmp &= ~0x1f;
+       tmp |= 1 << mode;
+
+       WREG32_SOC15(NBIO, 0, regBIF_BX_PF0_PARTITION_COMPUTE_CAP, tmp);
+}
+
 const struct amdgpu_nbio_funcs nbio_v7_9_funcs = {
        .get_hdp_flush_req_offset = nbio_v7_9_get_hdp_flush_req_offset,
        .get_hdp_flush_done_offset = nbio_v7_9_get_hdp_flush_done_offset,
@@ -366,4 +396,6 @@ const struct amdgpu_nbio_funcs nbio_v7_9_funcs = {
        .get_clockgating_state = nbio_v7_9_get_clockgating_state,
        .ih_control = nbio_v7_9_ih_control,
        .remap_hdp_registers = nbio_v7_9_remap_hdp_registers,
+       .get_compute_partition_mode = nbio_v7_9_get_compute_partition_mode,
+       .set_compute_partition_mode = nbio_v7_9_set_compute_partition_mode,
 };