summaryrefslogtreecommitdiff
path: root/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c
diff options
context:
space:
mode:
Diffstat (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c')
-rw-r--r--drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c1088
1 files changed, 1003 insertions, 85 deletions
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c
index 6ddc8e3360e2..c1f35ded684e 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c
@@ -24,10 +24,13 @@
*/
#include <linux/firmware.h>
+#include <linux/pm_runtime.h>
+
#include "amdgpu.h"
#include "amdgpu_gfx.h"
#include "amdgpu_rlc.h"
#include "amdgpu_ras.h"
+#include "amdgpu_reset.h"
#include "amdgpu_xcp.h"
#include "amdgpu_xgmi.h"
@@ -84,16 +87,6 @@ int amdgpu_gfx_me_queue_to_bit(struct amdgpu_device *adev,
return bit;
}
-void amdgpu_gfx_bit_to_me_queue(struct amdgpu_device *adev, int bit,
- int *me, int *pipe, int *queue)
-{
- *queue = bit % adev->gfx.me.num_queue_per_pipe;
- *pipe = (bit / adev->gfx.me.num_queue_per_pipe)
- % adev->gfx.me.num_pipe_per_me;
- *me = (bit / adev->gfx.me.num_queue_per_pipe)
- / adev->gfx.me.num_pipe_per_me;
-}
-
bool amdgpu_gfx_is_me_queue_enabled(struct amdgpu_device *adev,
int me, int pipe, int queue)
{
@@ -304,11 +297,11 @@ static int amdgpu_gfx_kiq_acquire(struct amdgpu_device *adev,
return -EINVAL;
}
-int amdgpu_gfx_kiq_init_ring(struct amdgpu_device *adev,
- struct amdgpu_ring *ring,
- struct amdgpu_irq_src *irq, int xcc_id)
+int amdgpu_gfx_kiq_init_ring(struct amdgpu_device *adev, int xcc_id)
{
struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
+ struct amdgpu_irq_src *irq = &kiq->irq;
+ struct amdgpu_ring *ring = &kiq->ring;
int r = 0;
spin_lock_init(&kiq->ring_lock);
@@ -329,7 +322,9 @@ int amdgpu_gfx_kiq_init_ring(struct amdgpu_device *adev,
ring->eop_gpu_addr = kiq->eop_gpu_addr;
ring->no_scheduler = true;
- sprintf(ring->name, "kiq_%d.%d.%d.%d", xcc_id, ring->me, ring->pipe, ring->queue);
+ snprintf(ring->name, sizeof(ring->name), "kiq_%hhu.%hhu.%hhu.%hhu",
+ (unsigned char)xcc_id, (unsigned char)ring->me,
+ (unsigned char)ring->pipe, (unsigned char)ring->queue);
r = amdgpu_ring_init(adev, ring, 1024, irq, AMDGPU_CP_KIQ_IRQ_DRIVER0,
AMDGPU_RING_PRIO_DEFAULT, NULL);
if (r)
@@ -410,7 +405,7 @@ int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,
}
/* prepare MQD backup */
- kiq->mqd_backup = kmalloc(mqd_size, GFP_KERNEL);
+ kiq->mqd_backup = kzalloc(mqd_size, GFP_KERNEL);
if (!kiq->mqd_backup) {
dev_warn(adev->dev,
"no memory to create MQD backup for ring %s\n", ring->name);
@@ -433,7 +428,7 @@ int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,
ring->mqd_size = mqd_size;
/* prepare MQD backup */
- adev->gfx.me.mqd_backup[i] = kmalloc(mqd_size, GFP_KERNEL);
+ adev->gfx.me.mqd_backup[i] = kzalloc(mqd_size, GFP_KERNEL);
if (!adev->gfx.me.mqd_backup[i]) {
dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
return -ENOMEM;
@@ -457,7 +452,7 @@ int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,
ring->mqd_size = mqd_size;
/* prepare MQD backup */
- adev->gfx.mec.mqd_backup[j] = kmalloc(mqd_size, GFP_KERNEL);
+ adev->gfx.mec.mqd_backup[j] = kzalloc(mqd_size, GFP_KERNEL);
if (!adev->gfx.mec.mqd_backup[j]) {
dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
return -ENOMEM;
@@ -504,15 +499,25 @@ int amdgpu_gfx_disable_kcq(struct amdgpu_device *adev, int xcc_id)
{
struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
struct amdgpu_ring *kiq_ring = &kiq->ring;
- struct amdgpu_hive_info *hive;
- struct amdgpu_ras *ras;
- int hive_ras_recovery = 0;
int i, r = 0;
int j;
+ if (adev->enable_mes) {
+ for (i = 0; i < adev->gfx.num_compute_rings; i++) {
+ j = i + xcc_id * adev->gfx.num_compute_rings;
+ amdgpu_mes_unmap_legacy_queue(adev,
+ &adev->gfx.compute_ring[j],
+ RESET_QUEUES, 0, 0);
+ }
+ return 0;
+ }
+
if (!kiq->pmf || !kiq->pmf->kiq_unmap_queues)
return -EINVAL;
+ if (!kiq_ring->sched.ready || amdgpu_in_reset(adev))
+ return 0;
+
spin_lock(&kiq->ring_lock);
if (amdgpu_ring_alloc(kiq_ring, kiq->pmf->unmap_queues_size *
adev->gfx.num_compute_rings)) {
@@ -526,26 +531,15 @@ int amdgpu_gfx_disable_kcq(struct amdgpu_device *adev, int xcc_id)
&adev->gfx.compute_ring[j],
RESET_QUEUES, 0, 0);
}
-
- /**
- * This is workaround: only skip kiq_ring test
- * during ras recovery in suspend stage for gfx9.4.3
+ /* Submit unmap queue packet */
+ amdgpu_ring_commit(kiq_ring);
+ /*
+ * Ring test will do a basic scratch register change check. Just run
+ * this to ensure that unmap queues that is submitted before got
+ * processed successfully before returning.
*/
- hive = amdgpu_get_xgmi_hive(adev);
- if (hive) {
- hive_ras_recovery = atomic_read(&hive->ras_recovery);
- amdgpu_put_xgmi_hive(hive);
- }
-
- ras = amdgpu_ras_get_context(adev);
- if ((amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3)) &&
- ras && (atomic_read(&ras->in_recovery) || hive_ras_recovery)) {
- spin_unlock(&kiq->ring_lock);
- return 0;
- }
+ r = amdgpu_ring_test_helper(kiq_ring);
- if (kiq_ring->sched.ready && !adev->job_hang)
- r = amdgpu_ring_test_helper(kiq_ring);
spin_unlock(&kiq->ring_lock);
return r;
@@ -558,11 +552,26 @@ int amdgpu_gfx_disable_kgq(struct amdgpu_device *adev, int xcc_id)
int i, r = 0;
int j;
+ if (adev->enable_mes) {
+ if (amdgpu_gfx_is_master_xcc(adev, xcc_id)) {
+ for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
+ j = i + xcc_id * adev->gfx.num_gfx_rings;
+ amdgpu_mes_unmap_legacy_queue(adev,
+ &adev->gfx.gfx_ring[j],
+ PREEMPT_QUEUES, 0, 0);
+ }
+ }
+ return 0;
+ }
+
if (!kiq->pmf || !kiq->pmf->kiq_unmap_queues)
return -EINVAL;
- spin_lock(&kiq->ring_lock);
+ if (!adev->gfx.kiq[0].ring.sched.ready || amdgpu_in_reset(adev))
+ return 0;
+
if (amdgpu_gfx_is_master_xcc(adev, xcc_id)) {
+ spin_lock(&kiq->ring_lock);
if (amdgpu_ring_alloc(kiq_ring, kiq->pmf->unmap_queues_size *
adev->gfx.num_gfx_rings)) {
spin_unlock(&kiq->ring_lock);
@@ -575,11 +584,17 @@ int amdgpu_gfx_disable_kgq(struct amdgpu_device *adev, int xcc_id)
&adev->gfx.gfx_ring[j],
PREEMPT_QUEUES, 0, 0);
}
- }
+ /* Submit unmap queue packet */
+ amdgpu_ring_commit(kiq_ring);
- if (adev->gfx.kiq[0].ring.sched.ready && !adev->job_hang)
+ /*
+ * Ring test will do a basic scratch register change check.
+ * Just run this to ensure that unmap queues that is submitted
+ * before got processed successfully before returning.
+ */
r = amdgpu_ring_test_helper(kiq_ring);
- spin_unlock(&kiq->ring_lock);
+ spin_unlock(&kiq->ring_lock);
+ }
return r;
}
@@ -597,6 +612,44 @@ int amdgpu_queue_mask_bit_to_set_resource_bit(struct amdgpu_device *adev,
return set_resource_bit;
}
+static int amdgpu_gfx_mes_enable_kcq(struct amdgpu_device *adev, int xcc_id)
+{
+ struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
+ struct amdgpu_ring *kiq_ring = &kiq->ring;
+ uint64_t queue_mask = ~0ULL;
+ int r, i, j;
+
+ amdgpu_device_flush_hdp(adev, NULL);
+
+ if (!adev->enable_uni_mes) {
+ spin_lock(&kiq->ring_lock);
+ r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->set_resources_size);
+ if (r) {
+ dev_err(adev->dev, "Failed to lock KIQ (%d).\n", r);
+ spin_unlock(&kiq->ring_lock);
+ return r;
+ }
+
+ kiq->pmf->kiq_set_resources(kiq_ring, queue_mask);
+ r = amdgpu_ring_test_helper(kiq_ring);
+ spin_unlock(&kiq->ring_lock);
+ if (r)
+ dev_err(adev->dev, "KIQ failed to set resources\n");
+ }
+
+ for (i = 0; i < adev->gfx.num_compute_rings; i++) {
+ j = i + xcc_id * adev->gfx.num_compute_rings;
+ r = amdgpu_mes_map_legacy_queue(adev,
+ &adev->gfx.compute_ring[j]);
+ if (r) {
+ dev_err(adev->dev, "failed to map compute queue\n");
+ return r;
+ }
+ }
+
+ return 0;
+}
+
int amdgpu_gfx_enable_kcq(struct amdgpu_device *adev, int xcc_id)
{
struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
@@ -604,6 +657,9 @@ int amdgpu_gfx_enable_kcq(struct amdgpu_device *adev, int xcc_id)
uint64_t queue_mask = 0;
int r, i, j;
+ if (adev->mes.enable_legacy_queue_map)
+ return amdgpu_gfx_mes_enable_kcq(adev, xcc_id);
+
if (!kiq->pmf || !kiq->pmf->kiq_map_queues || !kiq->pmf->kiq_set_resources)
return -EINVAL;
@@ -622,10 +678,11 @@ int amdgpu_gfx_enable_kcq(struct amdgpu_device *adev, int xcc_id)
queue_mask |= (1ull << amdgpu_queue_mask_bit_to_set_resource_bit(adev, i));
}
- DRM_INFO("kiq ring mec %d pipe %d q %d\n", kiq_ring->me, kiq_ring->pipe,
- kiq_ring->queue);
amdgpu_device_flush_hdp(adev, NULL);
+ DRM_INFO("kiq ring mec %d pipe %d q %d\n", kiq_ring->me, kiq_ring->pipe,
+ kiq_ring->queue);
+
spin_lock(&kiq->ring_lock);
r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->map_queues_size *
adev->gfx.num_compute_rings +
@@ -636,16 +693,19 @@ int amdgpu_gfx_enable_kcq(struct amdgpu_device *adev, int xcc_id)
return r;
}
- if (adev->enable_mes)
- queue_mask = ~0ULL;
-
kiq->pmf->kiq_set_resources(kiq_ring, queue_mask);
for (i = 0; i < adev->gfx.num_compute_rings; i++) {
j = i + xcc_id * adev->gfx.num_compute_rings;
- kiq->pmf->kiq_map_queues(kiq_ring,
- &adev->gfx.compute_ring[j]);
+ kiq->pmf->kiq_map_queues(kiq_ring,
+ &adev->gfx.compute_ring[j]);
}
-
+ /* Submit map queue packet */
+ amdgpu_ring_commit(kiq_ring);
+ /*
+ * Ring test will do a basic scratch register change check. Just run
+ * this to ensure that map queues that is submitted before got
+ * processed successfully before returning.
+ */
r = amdgpu_ring_test_helper(kiq_ring);
spin_unlock(&kiq->ring_lock);
if (r)
@@ -665,6 +725,20 @@ int amdgpu_gfx_enable_kgq(struct amdgpu_device *adev, int xcc_id)
amdgpu_device_flush_hdp(adev, NULL);
+ if (adev->mes.enable_legacy_queue_map) {
+ for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
+ j = i + xcc_id * adev->gfx.num_gfx_rings;
+ r = amdgpu_mes_map_legacy_queue(adev,
+ &adev->gfx.gfx_ring[j]);
+ if (r) {
+ DRM_ERROR("failed to map gfx queue\n");
+ return r;
+ }
+ }
+
+ return 0;
+ }
+
spin_lock(&kiq->ring_lock);
/* No need to map kcq on the slave */
if (amdgpu_gfx_is_master_xcc(adev, xcc_id)) {
@@ -682,11 +756,17 @@ int amdgpu_gfx_enable_kgq(struct amdgpu_device *adev, int xcc_id)
&adev->gfx.gfx_ring[j]);
}
}
-
+ /* Submit map queue packet */
+ amdgpu_ring_commit(kiq_ring);
+ /*
+ * Ring test will do a basic scratch register change check. Just run
+ * this to ensure that map queues that is submitted before got
+ * processed successfully before returning.
+ */
r = amdgpu_ring_test_helper(kiq_ring);
spin_unlock(&kiq->ring_lock);
if (r)
- DRM_ERROR("KCQ enable failed\n");
+ DRM_ERROR("KGQ enable failed\n");
return r;
}
@@ -726,7 +806,7 @@ void amdgpu_gfx_off_ctrl(struct amdgpu_device *adev, bool enable)
/* If going to s2idle, no need to wait */
if (adev->in_s0ix) {
if (!amdgpu_dpm_set_powergating_by_smu(adev,
- AMD_IP_BLOCK_TYPE_GFX, true))
+ AMD_IP_BLOCK_TYPE_GFX, true, 0))
adev->gfx.gfx_off_state = true;
} else {
schedule_delayed_work(&adev->gfx.gfx_off_delay_work,
@@ -738,7 +818,7 @@ void amdgpu_gfx_off_ctrl(struct amdgpu_device *adev, bool enable)
cancel_delayed_work_sync(&adev->gfx.gfx_off_delay_work);
if (adev->gfx.gfx_off_state &&
- !amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, false)) {
+ !amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, false, 0)) {
adev->gfx.gfx_off_state = false;
if (adev->gfx.funcs->init_spm_golden) {
@@ -814,13 +894,19 @@ int amdgpu_gfx_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *r
int r;
if (amdgpu_ras_is_supported(adev, ras_block->block)) {
- if (!amdgpu_persistent_edc_harvesting_supported(adev))
- amdgpu_ras_reset_error_status(adev, AMDGPU_RAS_BLOCK__GFX);
+ if (!amdgpu_persistent_edc_harvesting_supported(adev)) {
+ r = amdgpu_ras_reset_error_status(adev, AMDGPU_RAS_BLOCK__GFX);
+ if (r)
+ return r;
+ }
r = amdgpu_ras_block_late_init(adev, ras_block);
if (r)
return r;
+ if (amdgpu_sriov_vf(adev))
+ return r;
+
if (adev->gfx.cp_ecc_error_irq.funcs) {
r = amdgpu_irq_get(adev, &adev->gfx.cp_ecc_error_irq, 0);
if (r)
@@ -949,7 +1035,7 @@ uint32_t amdgpu_kiq_rreg(struct amdgpu_device *adev, uint32_t reg, uint32_t xcc_
if (amdgpu_device_skip_hw_access(adev))
return 0;
- if (adev->mes.ring.sched.ready)
+ if (adev->mes.ring[0].sched.ready)
return amdgpu_mes_rreg(adev, reg);
BUG_ON(!ring->funcs->emit_rreg);
@@ -959,7 +1045,10 @@ uint32_t amdgpu_kiq_rreg(struct amdgpu_device *adev, uint32_t reg, uint32_t xcc_
pr_err("critical bug! too many kiq readers\n");
goto failed_unlock;
}
- amdgpu_ring_alloc(ring, 32);
+ r = amdgpu_ring_alloc(ring, 32);
+ if (r)
+ goto failed_unlock;
+
amdgpu_ring_emit_rreg(ring, reg, reg_val_offs);
r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
if (r)
@@ -1019,13 +1108,16 @@ void amdgpu_kiq_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v, uint3
if (amdgpu_device_skip_hw_access(adev))
return;
- if (adev->mes.ring.sched.ready) {
+ if (adev->mes.ring[0].sched.ready) {
amdgpu_mes_wreg(adev, reg, v);
return;
}
spin_lock_irqsave(&kiq->ring_lock, flags);
- amdgpu_ring_alloc(ring, 32);
+ r = amdgpu_ring_alloc(ring, 32);
+ if (r)
+ goto failed_unlock;
+
amdgpu_ring_emit_wreg(ring, reg, v);
r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
if (r)
@@ -1061,6 +1153,7 @@ void amdgpu_kiq_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v, uint3
failed_undo:
amdgpu_ring_undo(ring);
+failed_unlock:
spin_unlock_irqrestore(&kiq->ring_lock, flags);
failed_kiq_write:
dev_err(adev->dev, "failed to write reg:%x\n", reg);
@@ -1205,7 +1298,8 @@ void amdgpu_gfx_cp_init_microcode(struct amdgpu_device *adev,
fw_size = le32_to_cpu(cp_hdr_v2_0->data_size_bytes);
break;
default:
- break;
+ dev_err(adev->dev, "Invalid ucode id %u\n", ucode_id);
+ return;
}
if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
@@ -1281,59 +1375,883 @@ static ssize_t amdgpu_gfx_set_compute_partition(struct device *dev,
return count;
}
+static const char *xcp_desc[] = {
+ [AMDGPU_SPX_PARTITION_MODE] = "SPX",
+ [AMDGPU_DPX_PARTITION_MODE] = "DPX",
+ [AMDGPU_TPX_PARTITION_MODE] = "TPX",
+ [AMDGPU_QPX_PARTITION_MODE] = "QPX",
+ [AMDGPU_CPX_PARTITION_MODE] = "CPX",
+};
+
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;
+ struct amdgpu_xcp_mgr *xcp_mgr = adev->xcp_mgr;
+ int size = 0, mode;
+ char *sep = "";
- /* TBD */
- switch (NUM_XCC(adev->gfx.xcc_mask)) {
- 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;
+ if (!xcp_mgr || !xcp_mgr->avail_xcp_modes)
+ return sysfs_emit(buf, "Not supported\n");
+
+ for_each_inst(mode, xcp_mgr->avail_xcp_modes) {
+ size += sysfs_emit_at(buf, size, "%s%s", sep, xcp_desc[mode]);
+ sep = ", ";
+ }
+
+ size += sysfs_emit_at(buf, size, "\n");
+
+ return size;
+}
+
+static int amdgpu_gfx_run_cleaner_shader_job(struct amdgpu_ring *ring)
+{
+ struct amdgpu_device *adev = ring->adev;
+ struct drm_gpu_scheduler *sched = &ring->sched;
+ struct drm_sched_entity entity;
+ struct dma_fence *f;
+ struct amdgpu_job *job;
+ struct amdgpu_ib *ib;
+ int i, r;
+
+ /* Initialize the scheduler entity */
+ r = drm_sched_entity_init(&entity, DRM_SCHED_PRIORITY_NORMAL,
+ &sched, 1, NULL);
+ if (r) {
+ dev_err(adev->dev, "Failed setting up GFX kernel entity.\n");
+ goto err;
}
- return sysfs_emit(buf, "%s\n", supported_partition);
+ r = amdgpu_job_alloc_with_ib(ring->adev, &entity, NULL,
+ 64, 0,
+ &job);
+ if (r)
+ goto err;
+
+ job->enforce_isolation = true;
+
+ ib = &job->ibs[0];
+ for (i = 0; i <= ring->funcs->align_mask; ++i)
+ ib->ptr[i] = ring->funcs->nop;
+ ib->length_dw = ring->funcs->align_mask + 1;
+
+ f = amdgpu_job_submit(job);
+
+ r = dma_fence_wait(f, false);
+ if (r)
+ goto err;
+
+ dma_fence_put(f);
+
+ /* Clean up the scheduler entity */
+ drm_sched_entity_destroy(&entity);
+ return 0;
+
+err:
+ return r;
+}
+
+static int amdgpu_gfx_run_cleaner_shader(struct amdgpu_device *adev, int xcp_id)
+{
+ int num_xcc = NUM_XCC(adev->gfx.xcc_mask);
+ struct amdgpu_ring *ring;
+ int num_xcc_to_clear;
+ int i, r, xcc_id;
+
+ if (adev->gfx.num_xcc_per_xcp)
+ num_xcc_to_clear = adev->gfx.num_xcc_per_xcp;
+ else
+ num_xcc_to_clear = 1;
+
+ for (xcc_id = 0; xcc_id < num_xcc; xcc_id++) {
+ for (i = 0; i < adev->gfx.num_compute_rings; i++) {
+ ring = &adev->gfx.compute_ring[i + xcc_id * adev->gfx.num_compute_rings];
+ if ((ring->xcp_id == xcp_id) && ring->sched.ready) {
+ r = amdgpu_gfx_run_cleaner_shader_job(ring);
+ if (r)
+ return r;
+ num_xcc_to_clear--;
+ break;
+ }
+ }
+ }
+
+ if (num_xcc_to_clear)
+ return -ENOENT;
+
+ return 0;
+}
+
+/**
+ * amdgpu_gfx_set_run_cleaner_shader - Execute the AMDGPU GFX Cleaner Shader
+ * @dev: The device structure
+ * @attr: The device attribute structure
+ * @buf: The buffer containing the input data
+ * @count: The size of the input data
+ *
+ * Provides the sysfs interface to manually run a cleaner shader, which is
+ * used to clear the GPU state between different tasks. Writing a value to the
+ * 'run_cleaner_shader' sysfs file triggers the cleaner shader execution.
+ * The value written corresponds to the partition index on multi-partition
+ * devices. On single-partition devices, the value should be '0'.
+ *
+ * The cleaner shader clears the Local Data Store (LDS) and General Purpose
+ * Registers (GPRs) to ensure data isolation between GPU workloads.
+ *
+ * Return: The number of bytes written to the sysfs file.
+ */
+static ssize_t amdgpu_gfx_set_run_cleaner_shader(struct device *dev,
+ struct device_attribute *attr,
+ const char *buf,
+ size_t count)
+{
+ struct drm_device *ddev = dev_get_drvdata(dev);
+ struct amdgpu_device *adev = drm_to_adev(ddev);
+ int ret;
+ long value;
+
+ if (amdgpu_in_reset(adev))
+ return -EPERM;
+ if (adev->in_suspend && !adev->in_runpm)
+ return -EPERM;
+
+ ret = kstrtol(buf, 0, &value);
+
+ if (ret)
+ return -EINVAL;
+
+ if (value < 0)
+ return -EINVAL;
+
+ if (adev->xcp_mgr) {
+ if (value >= adev->xcp_mgr->num_xcps)
+ return -EINVAL;
+ } else {
+ if (value > 1)
+ return -EINVAL;
+ }
+
+ ret = pm_runtime_get_sync(ddev->dev);
+ if (ret < 0) {
+ pm_runtime_put_autosuspend(ddev->dev);
+ return ret;
+ }
+
+ ret = amdgpu_gfx_run_cleaner_shader(adev, value);
+
+ pm_runtime_mark_last_busy(ddev->dev);
+ pm_runtime_put_autosuspend(ddev->dev);
+
+ if (ret)
+ return ret;
+
+ return count;
+}
+
+/**
+ * amdgpu_gfx_get_enforce_isolation - Query AMDGPU GFX Enforce Isolation Settings
+ * @dev: The device structure
+ * @attr: The device attribute structure
+ * @buf: The buffer to store the output data
+ *
+ * Provides the sysfs read interface to get the current settings of the 'enforce_isolation'
+ * feature for each GPU partition. Reading from the 'enforce_isolation'
+ * sysfs file returns the isolation settings for all partitions, where '0'
+ * indicates disabled and '1' indicates enabled.
+ *
+ * Return: The number of bytes read from the sysfs file.
+ */
+static ssize_t amdgpu_gfx_get_enforce_isolation(struct device *dev,
+ struct device_attribute *attr,
+ char *buf)
+{
+ struct drm_device *ddev = dev_get_drvdata(dev);
+ struct amdgpu_device *adev = drm_to_adev(ddev);
+ int i;
+ ssize_t size = 0;
+
+ if (adev->xcp_mgr) {
+ for (i = 0; i < adev->xcp_mgr->num_xcps; i++) {
+ size += sysfs_emit_at(buf, size, "%u", adev->enforce_isolation[i]);
+ if (i < (adev->xcp_mgr->num_xcps - 1))
+ size += sysfs_emit_at(buf, size, " ");
+ }
+ buf[size++] = '\n';
+ } else {
+ size = sysfs_emit_at(buf, 0, "%u\n", adev->enforce_isolation[0]);
+ }
+
+ return size;
+}
+
+/**
+ * amdgpu_gfx_set_enforce_isolation - Control AMDGPU GFX Enforce Isolation
+ * @dev: The device structure
+ * @attr: The device attribute structure
+ * @buf: The buffer containing the input data
+ * @count: The size of the input data
+ *
+ * This function allows control over the 'enforce_isolation' feature, which
+ * serializes access to the graphics engine. Writing '1' or '0' to the
+ * 'enforce_isolation' sysfs file enables or disables process isolation for
+ * each partition. The input should specify the setting for all partitions.
+ *
+ * Return: The number of bytes written to the sysfs file.
+ */
+static ssize_t amdgpu_gfx_set_enforce_isolation(struct device *dev,
+ struct device_attribute *attr,
+ const char *buf, size_t count)
+{
+ struct drm_device *ddev = dev_get_drvdata(dev);
+ struct amdgpu_device *adev = drm_to_adev(ddev);
+ long partition_values[MAX_XCP] = {0};
+ int ret, i, num_partitions;
+ const char *input_buf = buf;
+
+ for (i = 0; i < (adev->xcp_mgr ? adev->xcp_mgr->num_xcps : 1); i++) {
+ ret = sscanf(input_buf, "%ld", &partition_values[i]);
+ if (ret <= 0)
+ break;
+
+ /* Move the pointer to the next value in the string */
+ input_buf = strchr(input_buf, ' ');
+ if (input_buf) {
+ input_buf++;
+ } else {
+ i++;
+ break;
+ }
+ }
+ num_partitions = i;
+
+ if (adev->xcp_mgr && num_partitions != adev->xcp_mgr->num_xcps)
+ return -EINVAL;
+
+ if (!adev->xcp_mgr && num_partitions != 1)
+ return -EINVAL;
+
+ for (i = 0; i < num_partitions; i++) {
+ if (partition_values[i] != 0 && partition_values[i] != 1)
+ return -EINVAL;
+ }
+
+ mutex_lock(&adev->enforce_isolation_mutex);
+ for (i = 0; i < num_partitions; i++) {
+ if (adev->enforce_isolation[i] && !partition_values[i])
+ /* Going from enabled to disabled */
+ amdgpu_vmid_free_reserved(adev, AMDGPU_GFXHUB(i));
+ else if (!adev->enforce_isolation[i] && partition_values[i])
+ /* Going from disabled to enabled */
+ amdgpu_vmid_alloc_reserved(adev, AMDGPU_GFXHUB(i));
+ adev->enforce_isolation[i] = partition_values[i];
+ }
+ mutex_unlock(&adev->enforce_isolation_mutex);
+
+ amdgpu_mes_update_enforce_isolation(adev);
+
+ return count;
+}
+
+static ssize_t amdgpu_gfx_get_gfx_reset_mask(struct device *dev,
+ struct device_attribute *attr,
+ char *buf)
+{
+ struct drm_device *ddev = dev_get_drvdata(dev);
+ struct amdgpu_device *adev = drm_to_adev(ddev);
+
+ if (!adev)
+ return -ENODEV;
+
+ return amdgpu_show_reset_mask(buf, adev->gfx.gfx_supported_reset);
+}
+
+static ssize_t amdgpu_gfx_get_compute_reset_mask(struct device *dev,
+ struct device_attribute *attr,
+ char *buf)
+{
+ struct drm_device *ddev = dev_get_drvdata(dev);
+ struct amdgpu_device *adev = drm_to_adev(ddev);
+
+ if (!adev)
+ return -ENODEV;
+
+ return amdgpu_show_reset_mask(buf, adev->gfx.compute_supported_reset);
}
+static DEVICE_ATTR(run_cleaner_shader, 0200,
+ NULL, amdgpu_gfx_set_run_cleaner_shader);
+
+static DEVICE_ATTR(enforce_isolation, 0644,
+ amdgpu_gfx_get_enforce_isolation,
+ amdgpu_gfx_set_enforce_isolation);
+
static DEVICE_ATTR(current_compute_partition, 0644,
amdgpu_gfx_get_current_compute_partition,
amdgpu_gfx_set_compute_partition);
static DEVICE_ATTR(available_compute_partition, 0444,
amdgpu_gfx_get_available_compute_partition, NULL);
+static DEVICE_ATTR(gfx_reset_mask, 0444,
+ amdgpu_gfx_get_gfx_reset_mask, NULL);
-int amdgpu_gfx_sysfs_init(struct amdgpu_device *adev)
+static DEVICE_ATTR(compute_reset_mask, 0444,
+ amdgpu_gfx_get_compute_reset_mask, NULL);
+
+static int amdgpu_gfx_sysfs_xcp_init(struct amdgpu_device *adev)
{
+ struct amdgpu_xcp_mgr *xcp_mgr = adev->xcp_mgr;
+ bool xcp_switch_supported;
int r;
+ if (!xcp_mgr)
+ return 0;
+
+ xcp_switch_supported =
+ (xcp_mgr->funcs && xcp_mgr->funcs->switch_partition_mode);
+
+ if (!xcp_switch_supported)
+ dev_attr_current_compute_partition.attr.mode &=
+ ~(S_IWUSR | S_IWGRP | S_IWOTH);
+
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 (xcp_switch_supported)
+ r = device_create_file(adev->dev,
+ &dev_attr_available_compute_partition);
return r;
}
-void amdgpu_gfx_sysfs_fini(struct amdgpu_device *adev)
+static void amdgpu_gfx_sysfs_xcp_fini(struct amdgpu_device *adev)
{
+ struct amdgpu_xcp_mgr *xcp_mgr = adev->xcp_mgr;
+ bool xcp_switch_supported;
+
+ if (!xcp_mgr)
+ return;
+
+ xcp_switch_supported =
+ (xcp_mgr->funcs && xcp_mgr->funcs->switch_partition_mode);
device_remove_file(adev->dev, &dev_attr_current_compute_partition);
- device_remove_file(adev->dev, &dev_attr_available_compute_partition);
+
+ if (xcp_switch_supported)
+ device_remove_file(adev->dev,
+ &dev_attr_available_compute_partition);
+}
+
+static int amdgpu_gfx_sysfs_isolation_shader_init(struct amdgpu_device *adev)
+{
+ int r;
+
+ r = device_create_file(adev->dev, &dev_attr_enforce_isolation);
+ if (r)
+ return r;
+ if (adev->gfx.enable_cleaner_shader)
+ r = device_create_file(adev->dev, &dev_attr_run_cleaner_shader);
+
+ return r;
+}
+
+static void amdgpu_gfx_sysfs_isolation_shader_fini(struct amdgpu_device *adev)
+{
+ device_remove_file(adev->dev, &dev_attr_enforce_isolation);
+ if (adev->gfx.enable_cleaner_shader)
+ device_remove_file(adev->dev, &dev_attr_run_cleaner_shader);
+}
+
+static int amdgpu_gfx_sysfs_reset_mask_init(struct amdgpu_device *adev)
+{
+ int r = 0;
+
+ if (!amdgpu_gpu_recovery)
+ return r;
+
+ if (adev->gfx.num_gfx_rings) {
+ r = device_create_file(adev->dev, &dev_attr_gfx_reset_mask);
+ if (r)
+ return r;
+ }
+
+ if (adev->gfx.num_compute_rings) {
+ r = device_create_file(adev->dev, &dev_attr_compute_reset_mask);
+ if (r)
+ return r;
+ }
+
+ return r;
+}
+
+static void amdgpu_gfx_sysfs_reset_mask_fini(struct amdgpu_device *adev)
+{
+ if (!amdgpu_gpu_recovery)
+ return;
+
+ if (adev->gfx.num_gfx_rings)
+ device_remove_file(adev->dev, &dev_attr_gfx_reset_mask);
+
+ if (adev->gfx.num_compute_rings)
+ device_remove_file(adev->dev, &dev_attr_compute_reset_mask);
+}
+
+int amdgpu_gfx_sysfs_init(struct amdgpu_device *adev)
+{
+ int r;
+
+ r = amdgpu_gfx_sysfs_xcp_init(adev);
+ if (r) {
+ dev_err(adev->dev, "failed to create xcp sysfs files");
+ return r;
+ }
+
+ r = amdgpu_gfx_sysfs_isolation_shader_init(adev);
+ if (r)
+ dev_err(adev->dev, "failed to create isolation sysfs files");
+
+ r = amdgpu_gfx_sysfs_reset_mask_init(adev);
+ if (r)
+ dev_err(adev->dev, "failed to create reset mask sysfs files");
+
+ return r;
+}
+
+void amdgpu_gfx_sysfs_fini(struct amdgpu_device *adev)
+{
+ if (adev->dev->kobj.sd) {
+ amdgpu_gfx_sysfs_xcp_fini(adev);
+ amdgpu_gfx_sysfs_isolation_shader_fini(adev);
+ amdgpu_gfx_sysfs_reset_mask_fini(adev);
+ }
+}
+
+int amdgpu_gfx_cleaner_shader_sw_init(struct amdgpu_device *adev,
+ unsigned int cleaner_shader_size)
+{
+ if (!adev->gfx.enable_cleaner_shader)
+ return -EOPNOTSUPP;
+
+ return amdgpu_bo_create_kernel(adev, cleaner_shader_size, PAGE_SIZE,
+ AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT,
+ &adev->gfx.cleaner_shader_obj,
+ &adev->gfx.cleaner_shader_gpu_addr,
+ (void **)&adev->gfx.cleaner_shader_cpu_ptr);
+}
+
+void amdgpu_gfx_cleaner_shader_sw_fini(struct amdgpu_device *adev)
+{
+ if (!adev->gfx.enable_cleaner_shader)
+ return;
+
+ amdgpu_bo_free_kernel(&adev->gfx.cleaner_shader_obj,
+ &adev->gfx.cleaner_shader_gpu_addr,
+ (void **)&adev->gfx.cleaner_shader_cpu_ptr);
+}
+
+void amdgpu_gfx_cleaner_shader_init(struct amdgpu_device *adev,
+ unsigned int cleaner_shader_size,
+ const void *cleaner_shader_ptr)
+{
+ if (!adev->gfx.enable_cleaner_shader)
+ return;
+
+ if (adev->gfx.cleaner_shader_cpu_ptr && cleaner_shader_ptr)
+ memcpy_toio(adev->gfx.cleaner_shader_cpu_ptr, cleaner_shader_ptr,
+ cleaner_shader_size);
+}
+
+/**
+ * amdgpu_gfx_kfd_sch_ctrl - Control the KFD scheduler from the KGD (Graphics Driver)
+ * @adev: amdgpu_device pointer
+ * @idx: Index of the scheduler to control
+ * @enable: Whether to enable or disable the KFD scheduler
+ *
+ * This function is used to control the KFD (Kernel Fusion Driver) scheduler
+ * from the KGD. It is part of the cleaner shader feature. This function plays
+ * a key role in enforcing process isolation on the GPU.
+ *
+ * The function uses a reference count mechanism (kfd_sch_req_count) to keep
+ * track of the number of requests to enable the KFD scheduler. When a request
+ * to enable the KFD scheduler is made, the reference count is decremented.
+ * When the reference count reaches zero, a delayed work is scheduled to
+ * enforce isolation after a delay of GFX_SLICE_PERIOD.
+ *
+ * When a request to disable the KFD scheduler is made, the function first
+ * checks if the reference count is zero. If it is, it cancels the delayed work
+ * for enforcing isolation and checks if the KFD scheduler is active. If the
+ * KFD scheduler is active, it sends a request to stop the KFD scheduler and
+ * sets the KFD scheduler state to inactive. Then, it increments the reference
+ * count.
+ *
+ * The function is synchronized using the kfd_sch_mutex to ensure that the KFD
+ * scheduler state and reference count are updated atomically.
+ *
+ * Note: If the reference count is already zero when a request to enable the
+ * KFD scheduler is made, it means there's an imbalance bug somewhere. The
+ * function triggers a warning in this case.
+ */
+static void amdgpu_gfx_kfd_sch_ctrl(struct amdgpu_device *adev, u32 idx,
+ bool enable)
+{
+ mutex_lock(&adev->gfx.kfd_sch_mutex);
+
+ if (enable) {
+ /* If the count is already 0, it means there's an imbalance bug somewhere.
+ * Note that the bug may be in a different caller than the one which triggers the
+ * WARN_ON_ONCE.
+ */
+ if (WARN_ON_ONCE(adev->gfx.kfd_sch_req_count[idx] == 0)) {
+ dev_err(adev->dev, "Attempted to enable KFD scheduler when reference count is already zero\n");
+ goto unlock;
+ }
+
+ adev->gfx.kfd_sch_req_count[idx]--;
+
+ if (adev->gfx.kfd_sch_req_count[idx] == 0 &&
+ adev->gfx.kfd_sch_inactive[idx]) {
+ schedule_delayed_work(&adev->gfx.enforce_isolation[idx].work,
+ msecs_to_jiffies(adev->gfx.enforce_isolation_time[idx]));
+ }
+ } else {
+ if (adev->gfx.kfd_sch_req_count[idx] == 0) {
+ cancel_delayed_work_sync(&adev->gfx.enforce_isolation[idx].work);
+ if (!adev->gfx.kfd_sch_inactive[idx]) {
+ amdgpu_amdkfd_stop_sched(adev, idx);
+ adev->gfx.kfd_sch_inactive[idx] = true;
+ }
+ }
+
+ adev->gfx.kfd_sch_req_count[idx]++;
+ }
+
+unlock:
+ mutex_unlock(&adev->gfx.kfd_sch_mutex);
+}
+
+/**
+ * amdgpu_gfx_enforce_isolation_handler - work handler for enforcing shader isolation
+ *
+ * @work: work_struct.
+ *
+ * This function is the work handler for enforcing shader isolation on AMD GPUs.
+ * It counts the number of emitted fences for each GFX and compute ring. If there
+ * are any fences, it schedules the `enforce_isolation_work` to be run after a
+ * delay of `GFX_SLICE_PERIOD`. If there are no fences, it signals the Kernel Fusion
+ * Driver (KFD) to resume the runqueue. The function is synchronized using the
+ * `enforce_isolation_mutex`.
+ */
+void amdgpu_gfx_enforce_isolation_handler(struct work_struct *work)
+{
+ struct amdgpu_isolation_work *isolation_work =
+ container_of(work, struct amdgpu_isolation_work, work.work);
+ struct amdgpu_device *adev = isolation_work->adev;
+ u32 i, idx, fences = 0;
+
+ if (isolation_work->xcp_id == AMDGPU_XCP_NO_PARTITION)
+ idx = 0;
+ else
+ idx = isolation_work->xcp_id;
+
+ if (idx >= MAX_XCP)
+ return;
+
+ mutex_lock(&adev->enforce_isolation_mutex);
+ for (i = 0; i < AMDGPU_MAX_GFX_RINGS; ++i) {
+ if (isolation_work->xcp_id == adev->gfx.gfx_ring[i].xcp_id)
+ fences += amdgpu_fence_count_emitted(&adev->gfx.gfx_ring[i]);
+ }
+ for (i = 0; i < (AMDGPU_MAX_COMPUTE_RINGS * AMDGPU_MAX_GC_INSTANCES); ++i) {
+ if (isolation_work->xcp_id == adev->gfx.compute_ring[i].xcp_id)
+ fences += amdgpu_fence_count_emitted(&adev->gfx.compute_ring[i]);
+ }
+ if (fences) {
+ /* we've already had our timeslice, so let's wrap this up */
+ schedule_delayed_work(&adev->gfx.enforce_isolation[idx].work,
+ msecs_to_jiffies(1));
+ } else {
+ /* Tell KFD to resume the runqueue */
+ if (adev->kfd.init_complete) {
+ WARN_ON_ONCE(!adev->gfx.kfd_sch_inactive[idx]);
+ WARN_ON_ONCE(adev->gfx.kfd_sch_req_count[idx]);
+ amdgpu_amdkfd_start_sched(adev, idx);
+ adev->gfx.kfd_sch_inactive[idx] = false;
+ }
+ }
+ mutex_unlock(&adev->enforce_isolation_mutex);
+}
+
+/**
+ * amdgpu_gfx_enforce_isolation_wait_for_kfd - Manage KFD wait period for process isolation
+ * @adev: amdgpu_device pointer
+ * @idx: Index of the GPU partition
+ *
+ * When kernel submissions come in, the jobs are given a time slice and once
+ * that time slice is up, if there are KFD user queues active, kernel
+ * submissions are blocked until KFD has had its time slice. Once the KFD time
+ * slice is up, KFD user queues are preempted and kernel submissions are
+ * unblocked and allowed to run again.
+ */
+static void
+amdgpu_gfx_enforce_isolation_wait_for_kfd(struct amdgpu_device *adev,
+ u32 idx)
+{
+ unsigned long cjiffies;
+ bool wait = false;
+
+ mutex_lock(&adev->enforce_isolation_mutex);
+ if (adev->enforce_isolation[idx]) {
+ /* set the initial values if nothing is set */
+ if (!adev->gfx.enforce_isolation_jiffies[idx]) {
+ adev->gfx.enforce_isolation_jiffies[idx] = jiffies;
+ adev->gfx.enforce_isolation_time[idx] = GFX_SLICE_PERIOD_MS;
+ }
+ /* Make sure KFD gets a chance to run */
+ if (amdgpu_amdkfd_compute_active(adev, idx)) {
+ cjiffies = jiffies;
+ if (time_after(cjiffies, adev->gfx.enforce_isolation_jiffies[idx])) {
+ cjiffies -= adev->gfx.enforce_isolation_jiffies[idx];
+ if ((jiffies_to_msecs(cjiffies) >= GFX_SLICE_PERIOD_MS)) {
+ /* if our time is up, let KGD work drain before scheduling more */
+ wait = true;
+ /* reset the timer period */
+ adev->gfx.enforce_isolation_time[idx] = GFX_SLICE_PERIOD_MS;
+ } else {
+ /* set the timer period to what's left in our time slice */
+ adev->gfx.enforce_isolation_time[idx] =
+ GFX_SLICE_PERIOD_MS - jiffies_to_msecs(cjiffies);
+ }
+ } else {
+ /* if jiffies wrap around we will just wait a little longer */
+ adev->gfx.enforce_isolation_jiffies[idx] = jiffies;
+ }
+ } else {
+ /* if there is no KFD work, then set the full slice period */
+ adev->gfx.enforce_isolation_jiffies[idx] = jiffies;
+ adev->gfx.enforce_isolation_time[idx] = GFX_SLICE_PERIOD_MS;
+ }
+ }
+ mutex_unlock(&adev->enforce_isolation_mutex);
+
+ if (wait)
+ msleep(GFX_SLICE_PERIOD_MS);
+}
+
+/**
+ * amdgpu_gfx_enforce_isolation_ring_begin_use - Begin use of a ring with enforced isolation
+ * @ring: Pointer to the amdgpu_ring structure
+ *
+ * Ring begin_use helper implementation for gfx which serializes access to the
+ * gfx IP between kernel submission IOCTLs and KFD user queues when isolation
+ * enforcement is enabled. The kernel submission IOCTLs and KFD user queues
+ * each get a time slice when both are active.
+ */
+void amdgpu_gfx_enforce_isolation_ring_begin_use(struct amdgpu_ring *ring)
+{
+ struct amdgpu_device *adev = ring->adev;
+ u32 idx;
+ bool sched_work = false;
+
+ if (!adev->gfx.enable_cleaner_shader)
+ return;
+
+ if (ring->xcp_id == AMDGPU_XCP_NO_PARTITION)
+ idx = 0;
+ else
+ idx = ring->xcp_id;
+
+ if (idx >= MAX_XCP)
+ return;
+
+ /* Don't submit more work until KFD has had some time */
+ amdgpu_gfx_enforce_isolation_wait_for_kfd(adev, idx);
+
+ mutex_lock(&adev->enforce_isolation_mutex);
+ if (adev->enforce_isolation[idx]) {
+ if (adev->kfd.init_complete)
+ sched_work = true;
+ }
+ mutex_unlock(&adev->enforce_isolation_mutex);
+
+ if (sched_work)
+ amdgpu_gfx_kfd_sch_ctrl(adev, idx, false);
+}
+
+/**
+ * amdgpu_gfx_enforce_isolation_ring_end_use - End use of a ring with enforced isolation
+ * @ring: Pointer to the amdgpu_ring structure
+ *
+ * Ring end_use helper implementation for gfx which serializes access to the
+ * gfx IP between kernel submission IOCTLs and KFD user queues when isolation
+ * enforcement is enabled. The kernel submission IOCTLs and KFD user queues
+ * each get a time slice when both are active.
+ */
+void amdgpu_gfx_enforce_isolation_ring_end_use(struct amdgpu_ring *ring)
+{
+ struct amdgpu_device *adev = ring->adev;
+ u32 idx;
+ bool sched_work = false;
+
+ if (!adev->gfx.enable_cleaner_shader)
+ return;
+
+ if (ring->xcp_id == AMDGPU_XCP_NO_PARTITION)
+ idx = 0;
+ else
+ idx = ring->xcp_id;
+
+ if (idx >= MAX_XCP)
+ return;
+
+ mutex_lock(&adev->enforce_isolation_mutex);
+ if (adev->enforce_isolation[idx]) {
+ if (adev->kfd.init_complete)
+ sched_work = true;
+ }
+ mutex_unlock(&adev->enforce_isolation_mutex);
+
+ if (sched_work)
+ amdgpu_gfx_kfd_sch_ctrl(adev, idx, true);
+}
+
+/*
+ * debugfs for to enable/disable gfx job submission to specific core.
+ */
+#if defined(CONFIG_DEBUG_FS)
+static int amdgpu_debugfs_gfx_sched_mask_set(void *data, u64 val)
+{
+ struct amdgpu_device *adev = (struct amdgpu_device *)data;
+ u32 i;
+ u64 mask = 0;
+ struct amdgpu_ring *ring;
+
+ if (!adev)
+ return -ENODEV;
+
+ mask = (1ULL << adev->gfx.num_gfx_rings) - 1;
+ if ((val & mask) == 0)
+ return -EINVAL;
+
+ for (i = 0; i < adev->gfx.num_gfx_rings; ++i) {
+ ring = &adev->gfx.gfx_ring[i];
+ if (val & (1 << i))
+ ring->sched.ready = true;
+ else
+ ring->sched.ready = false;
+ }
+ /* publish sched.ready flag update effective immediately across smp */
+ smp_rmb();
+ return 0;
+}
+
+static int amdgpu_debugfs_gfx_sched_mask_get(void *data, u64 *val)
+{
+ struct amdgpu_device *adev = (struct amdgpu_device *)data;
+ u32 i;
+ u64 mask = 0;
+ struct amdgpu_ring *ring;
+
+ if (!adev)
+ return -ENODEV;
+ for (i = 0; i < adev->gfx.num_gfx_rings; ++i) {
+ ring = &adev->gfx.gfx_ring[i];
+ if (ring->sched.ready)
+ mask |= 1ULL << i;
+ }
+
+ *val = mask;
+ return 0;
+}
+
+DEFINE_DEBUGFS_ATTRIBUTE(amdgpu_debugfs_gfx_sched_mask_fops,
+ amdgpu_debugfs_gfx_sched_mask_get,
+ amdgpu_debugfs_gfx_sched_mask_set, "%llx\n");
+
+#endif
+
+void amdgpu_debugfs_gfx_sched_mask_init(struct amdgpu_device *adev)
+{
+#if defined(CONFIG_DEBUG_FS)
+ struct drm_minor *minor = adev_to_drm(adev)->primary;
+ struct dentry *root = minor->debugfs_root;
+ char name[32];
+
+ if (!(adev->gfx.num_gfx_rings > 1))
+ return;
+ sprintf(name, "amdgpu_gfx_sched_mask");
+ debugfs_create_file(name, 0600, root, adev,
+ &amdgpu_debugfs_gfx_sched_mask_fops);
+#endif
+}
+
+/*
+ * debugfs for to enable/disable compute job submission to specific core.
+ */
+#if defined(CONFIG_DEBUG_FS)
+static int amdgpu_debugfs_compute_sched_mask_set(void *data, u64 val)
+{
+ struct amdgpu_device *adev = (struct amdgpu_device *)data;
+ u32 i;
+ u64 mask = 0;
+ struct amdgpu_ring *ring;
+
+ if (!adev)
+ return -ENODEV;
+
+ mask = (1ULL << adev->gfx.num_compute_rings) - 1;
+ if ((val & mask) == 0)
+ return -EINVAL;
+
+ for (i = 0; i < adev->gfx.num_compute_rings; ++i) {
+ ring = &adev->gfx.compute_ring[i];
+ if (val & (1 << i))
+ ring->sched.ready = true;
+ else
+ ring->sched.ready = false;
+ }
+
+ /* publish sched.ready flag update effective immediately across smp */
+ smp_rmb();
+ return 0;
+}
+
+static int amdgpu_debugfs_compute_sched_mask_get(void *data, u64 *val)
+{
+ struct amdgpu_device *adev = (struct amdgpu_device *)data;
+ u32 i;
+ u64 mask = 0;
+ struct amdgpu_ring *ring;
+
+ if (!adev)
+ return -ENODEV;
+ for (i = 0; i < adev->gfx.num_compute_rings; ++i) {
+ ring = &adev->gfx.compute_ring[i];
+ if (ring->sched.ready)
+ mask |= 1ULL << i;
+ }
+
+ *val = mask;
+ return 0;
+}
+
+DEFINE_DEBUGFS_ATTRIBUTE(amdgpu_debugfs_compute_sched_mask_fops,
+ amdgpu_debugfs_compute_sched_mask_get,
+ amdgpu_debugfs_compute_sched_mask_set, "%llx\n");
+
+#endif
+
+void amdgpu_debugfs_compute_sched_mask_init(struct amdgpu_device *adev)
+{
+#if defined(CONFIG_DEBUG_FS)
+ struct drm_minor *minor = adev_to_drm(adev)->primary;
+ struct dentry *root = minor->debugfs_root;
+ char name[32];
+
+ if (!(adev->gfx.num_compute_rings > 1))
+ return;
+ sprintf(name, "amdgpu_compute_sched_mask");
+ debugfs_create_file(name, 0600, root, adev,
+ &amdgpu_debugfs_compute_sched_mask_fops);
+#endif
}