summaryrefslogtreecommitdiff
path: root/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
diff options
context:
space:
mode:
Diffstat (limited to 'drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c')
-rw-r--r--drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c301
1 files changed, 279 insertions, 22 deletions
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
index b21285afa4ea..b3b6dab71638 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
@@ -21,10 +21,11 @@
*
*/
+#include <linux/ratelimit.h>
+#include <linux/printk.h>
#include <linux/slab.h>
#include <linux/list.h>
#include <linux/types.h>
-#include <linux/printk.h>
#include <linux/bitops.h>
#include <linux/sched.h>
#include "kfd_priv.h"
@@ -118,9 +119,8 @@ static int allocate_vmid(struct device_queue_manager *dqm,
if (dqm->vmid_bitmap == 0)
return -ENOMEM;
- bit = find_first_bit((unsigned long *)&dqm->vmid_bitmap,
- dqm->dev->vm_info.vmid_num_kfd);
- clear_bit(bit, (unsigned long *)&dqm->vmid_bitmap);
+ bit = ffs(dqm->vmid_bitmap) - 1;
+ dqm->vmid_bitmap &= ~(1 << bit);
allocated_vmid = bit + dqm->dev->vm_info.first_vmid_kfd;
pr_debug("vmid allocation %d\n", allocated_vmid);
@@ -130,6 +130,15 @@ static int allocate_vmid(struct device_queue_manager *dqm,
set_pasid_vmid_mapping(dqm, q->process->pasid, q->properties.vmid);
program_sh_mem_settings(dqm, qpd);
+ /* qpd->page_table_base is set earlier when register_process()
+ * is called, i.e. when the first queue is created.
+ */
+ dqm->dev->kfd2kgd->set_vm_context_page_table_base(dqm->dev->kgd,
+ qpd->vmid,
+ qpd->page_table_base);
+ /* invalidate the VM context after pasid and vmid mapping is set up */
+ kfd_flush_tlb(qpd_to_pdd(qpd));
+
return 0;
}
@@ -139,10 +148,12 @@ static void deallocate_vmid(struct device_queue_manager *dqm,
{
int bit = qpd->vmid - dqm->dev->vm_info.first_vmid_kfd;
+ kfd_flush_tlb(qpd_to_pdd(qpd));
+
/* Release the vmid mapping */
set_pasid_vmid_mapping(dqm, 0, qpd->vmid);
- set_bit(bit, (unsigned long *)&dqm->vmid_bitmap);
+ dqm->vmid_bitmap |= (1 << bit);
qpd->vmid = 0;
q->properties.vmid = 0;
}
@@ -170,6 +181,14 @@ static int create_queue_nocpsch(struct device_queue_manager *dqm,
goto out_unlock;
}
q->properties.vmid = qpd->vmid;
+ /*
+ * Eviction state logic: we only mark active queues as evicted
+ * to avoid the overhead of restoring inactive queues later
+ */
+ if (qpd->evicted)
+ q->properties.is_evicted = (q->properties.queue_size > 0 &&
+ q->properties.queue_percent > 0 &&
+ q->properties.queue_address != 0);
q->properties.tba_addr = qpd->tba_addr;
q->properties.tma_addr = qpd->tma_addr;
@@ -223,12 +242,8 @@ static int allocate_hqd(struct device_queue_manager *dqm, struct queue *q)
continue;
if (dqm->allocated_queues[pipe] != 0) {
- bit = find_first_bit(
- (unsigned long *)&dqm->allocated_queues[pipe],
- get_queues_per_pipe(dqm));
-
- clear_bit(bit,
- (unsigned long *)&dqm->allocated_queues[pipe]);
+ bit = ffs(dqm->allocated_queues[pipe]) - 1;
+ dqm->allocated_queues[pipe] &= ~(1 << bit);
q->pipe = pipe;
q->queue = bit;
set = true;
@@ -249,7 +264,7 @@ static int allocate_hqd(struct device_queue_manager *dqm, struct queue *q)
static inline void deallocate_hqd(struct device_queue_manager *dqm,
struct queue *q)
{
- set_bit(q->queue, (unsigned long *)&dqm->allocated_queues[q->pipe]);
+ dqm->allocated_queues[q->pipe] |= (1 << q->queue);
}
static int create_compute_queue_nocpsch(struct device_queue_manager *dqm,
@@ -371,21 +386,35 @@ static int update_queue(struct device_queue_manager *dqm, struct queue *q)
{
int retval;
struct mqd_manager *mqd;
+ struct kfd_process_device *pdd;
bool prev_active = false;
mutex_lock(&dqm->lock);
+ pdd = kfd_get_process_device_data(q->device, q->process);
+ if (!pdd) {
+ retval = -ENODEV;
+ goto out_unlock;
+ }
mqd = dqm->ops.get_mqd_manager(dqm,
get_mqd_type_from_queue_type(q->properties.type));
if (!mqd) {
retval = -ENOMEM;
goto out_unlock;
}
+ /*
+ * Eviction state logic: we only mark active queues as evicted
+ * to avoid the overhead of restoring inactive queues later
+ */
+ if (pdd->qpd.evicted)
+ q->properties.is_evicted = (q->properties.queue_size > 0 &&
+ q->properties.queue_percent > 0 &&
+ q->properties.queue_address != 0);
/* Save previous activity state for counters */
prev_active = q->properties.is_active;
/* Make sure the queue is unmapped before updating the MQD */
- if (sched_policy != KFD_SCHED_POLICY_NO_HWS) {
+ if (dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) {
retval = unmap_queues_cpsch(dqm,
KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
if (retval) {
@@ -417,7 +446,7 @@ static int update_queue(struct device_queue_manager *dqm, struct queue *q)
else if (!q->properties.is_active && prev_active)
dqm->queue_count--;
- if (sched_policy != KFD_SCHED_POLICY_NO_HWS)
+ if (dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS)
retval = map_queues_cpsch(dqm);
else if (q->properties.is_active &&
(q->properties.type == KFD_QUEUE_TYPE_COMPUTE ||
@@ -451,10 +480,193 @@ static struct mqd_manager *get_mqd_manager(
return mqd;
}
+static int evict_process_queues_nocpsch(struct device_queue_manager *dqm,
+ struct qcm_process_device *qpd)
+{
+ struct queue *q;
+ struct mqd_manager *mqd;
+ struct kfd_process_device *pdd;
+ int retval = 0;
+
+ mutex_lock(&dqm->lock);
+ if (qpd->evicted++ > 0) /* already evicted, do nothing */
+ goto out;
+
+ pdd = qpd_to_pdd(qpd);
+ pr_info_ratelimited("Evicting PASID %u queues\n",
+ pdd->process->pasid);
+
+ /* unactivate all active queues on the qpd */
+ list_for_each_entry(q, &qpd->queues_list, list) {
+ if (!q->properties.is_active)
+ continue;
+ mqd = dqm->ops.get_mqd_manager(dqm,
+ get_mqd_type_from_queue_type(q->properties.type));
+ if (!mqd) { /* should not be here */
+ pr_err("Cannot evict queue, mqd mgr is NULL\n");
+ retval = -ENOMEM;
+ goto out;
+ }
+ q->properties.is_evicted = true;
+ q->properties.is_active = false;
+ retval = mqd->destroy_mqd(mqd, q->mqd,
+ KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN,
+ KFD_UNMAP_LATENCY_MS, q->pipe, q->queue);
+ if (retval)
+ goto out;
+ dqm->queue_count--;
+ }
+
+out:
+ mutex_unlock(&dqm->lock);
+ return retval;
+}
+
+static int evict_process_queues_cpsch(struct device_queue_manager *dqm,
+ struct qcm_process_device *qpd)
+{
+ struct queue *q;
+ struct kfd_process_device *pdd;
+ int retval = 0;
+
+ mutex_lock(&dqm->lock);
+ if (qpd->evicted++ > 0) /* already evicted, do nothing */
+ goto out;
+
+ pdd = qpd_to_pdd(qpd);
+ pr_info_ratelimited("Evicting PASID %u queues\n",
+ pdd->process->pasid);
+
+ /* unactivate all active queues on the qpd */
+ list_for_each_entry(q, &qpd->queues_list, list) {
+ if (!q->properties.is_active)
+ continue;
+ q->properties.is_evicted = true;
+ q->properties.is_active = false;
+ dqm->queue_count--;
+ }
+ retval = execute_queues_cpsch(dqm,
+ qpd->is_debug ?
+ KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES :
+ KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
+
+out:
+ mutex_unlock(&dqm->lock);
+ return retval;
+}
+
+static int restore_process_queues_nocpsch(struct device_queue_manager *dqm,
+ struct qcm_process_device *qpd)
+{
+ struct queue *q;
+ struct mqd_manager *mqd;
+ struct kfd_process_device *pdd;
+ uint32_t pd_base;
+ int retval = 0;
+
+ pdd = qpd_to_pdd(qpd);
+ /* Retrieve PD base */
+ pd_base = dqm->dev->kfd2kgd->get_process_page_dir(pdd->vm);
+
+ mutex_lock(&dqm->lock);
+ if (WARN_ON_ONCE(!qpd->evicted)) /* already restored, do nothing */
+ goto out;
+ if (qpd->evicted > 1) { /* ref count still > 0, decrement & quit */
+ qpd->evicted--;
+ goto out;
+ }
+
+ pr_info_ratelimited("Restoring PASID %u queues\n",
+ pdd->process->pasid);
+
+ /* Update PD Base in QPD */
+ qpd->page_table_base = pd_base;
+ pr_debug("Updated PD address to 0x%08x\n", pd_base);
+
+ if (!list_empty(&qpd->queues_list)) {
+ dqm->dev->kfd2kgd->set_vm_context_page_table_base(
+ dqm->dev->kgd,
+ qpd->vmid,
+ qpd->page_table_base);
+ kfd_flush_tlb(pdd);
+ }
+
+ /* activate all active queues on the qpd */
+ list_for_each_entry(q, &qpd->queues_list, list) {
+ if (!q->properties.is_evicted)
+ continue;
+ mqd = dqm->ops.get_mqd_manager(dqm,
+ get_mqd_type_from_queue_type(q->properties.type));
+ if (!mqd) { /* should not be here */
+ pr_err("Cannot restore queue, mqd mgr is NULL\n");
+ retval = -ENOMEM;
+ goto out;
+ }
+ q->properties.is_evicted = false;
+ q->properties.is_active = true;
+ retval = mqd->load_mqd(mqd, q->mqd, q->pipe,
+ q->queue, &q->properties,
+ q->process->mm);
+ if (retval)
+ goto out;
+ dqm->queue_count++;
+ }
+ qpd->evicted = 0;
+out:
+ mutex_unlock(&dqm->lock);
+ return retval;
+}
+
+static int restore_process_queues_cpsch(struct device_queue_manager *dqm,
+ struct qcm_process_device *qpd)
+{
+ struct queue *q;
+ struct kfd_process_device *pdd;
+ uint32_t pd_base;
+ int retval = 0;
+
+ pdd = qpd_to_pdd(qpd);
+ /* Retrieve PD base */
+ pd_base = dqm->dev->kfd2kgd->get_process_page_dir(pdd->vm);
+
+ mutex_lock(&dqm->lock);
+ if (WARN_ON_ONCE(!qpd->evicted)) /* already restored, do nothing */
+ goto out;
+ if (qpd->evicted > 1) { /* ref count still > 0, decrement & quit */
+ qpd->evicted--;
+ goto out;
+ }
+
+ pr_info_ratelimited("Restoring PASID %u queues\n",
+ pdd->process->pasid);
+
+ /* Update PD Base in QPD */
+ qpd->page_table_base = pd_base;
+ pr_debug("Updated PD address to 0x%08x\n", pd_base);
+
+ /* activate all active queues on the qpd */
+ list_for_each_entry(q, &qpd->queues_list, list) {
+ if (!q->properties.is_evicted)
+ continue;
+ q->properties.is_evicted = false;
+ q->properties.is_active = true;
+ dqm->queue_count++;
+ }
+ retval = execute_queues_cpsch(dqm,
+ KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
+ if (!retval)
+ qpd->evicted = 0;
+out:
+ mutex_unlock(&dqm->lock);
+ return retval;
+}
+
static int register_process(struct device_queue_manager *dqm,
struct qcm_process_device *qpd)
{
struct device_process_node *n;
+ struct kfd_process_device *pdd;
+ uint32_t pd_base;
int retval;
n = kzalloc(sizeof(*n), GFP_KERNEL);
@@ -463,9 +675,16 @@ static int register_process(struct device_queue_manager *dqm,
n->qpd = qpd;
+ pdd = qpd_to_pdd(qpd);
+ /* Retrieve PD base */
+ pd_base = dqm->dev->kfd2kgd->get_process_page_dir(pdd->vm);
+
mutex_lock(&dqm->lock);
list_add(&n->list, &dqm->queues);
+ /* Update PD Base in QPD */
+ qpd->page_table_base = pd_base;
+
retval = dqm->asic_ops.update_qpd(dqm, qpd);
dqm->processes_count++;
@@ -589,10 +808,8 @@ static int allocate_sdma_queue(struct device_queue_manager *dqm,
if (dqm->sdma_bitmap == 0)
return -ENOMEM;
- bit = find_first_bit((unsigned long *)&dqm->sdma_bitmap,
- CIK_SDMA_QUEUES);
-
- clear_bit(bit, (unsigned long *)&dqm->sdma_bitmap);
+ bit = ffs(dqm->sdma_bitmap) - 1;
+ dqm->sdma_bitmap &= ~(1 << bit);
*sdma_queue_id = bit;
return 0;
@@ -603,7 +820,7 @@ static void deallocate_sdma_queue(struct device_queue_manager *dqm,
{
if (sdma_queue_id >= CIK_SDMA_QUEUES)
return;
- set_bit(sdma_queue_id, (unsigned long *)&dqm->sdma_bitmap);
+ dqm->sdma_bitmap |= (1 << sdma_queue_id);
}
static int create_sdma_queue_nocpsch(struct device_queue_manager *dqm,
@@ -840,6 +1057,14 @@ static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q,
retval = -ENOMEM;
goto out;
}
+ /*
+ * Eviction state logic: we only mark active queues as evicted
+ * to avoid the overhead of restoring inactive queues later
+ */
+ if (qpd->evicted)
+ q->properties.is_evicted = (q->properties.queue_size > 0 &&
+ q->properties.queue_percent > 0 &&
+ q->properties.queue_address != 0);
dqm->asic_ops.init_sdma_vm(dqm, q, qpd);
@@ -1097,7 +1322,7 @@ static bool set_cache_memory_policy(struct device_queue_manager *dqm,
alternate_aperture_base,
alternate_aperture_size);
- if ((sched_policy == KFD_SCHED_POLICY_NO_HWS) && (qpd->vmid != 0))
+ if ((dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) && (qpd->vmid != 0))
program_sh_mem_settings(dqm, qpd);
pr_debug("sh_mem_config: 0x%x, ape1_base: 0x%x, ape1_limit: 0x%x\n",
@@ -1242,8 +1467,24 @@ struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev)
if (!dqm)
return NULL;
+ switch (dev->device_info->asic_family) {
+ /* HWS is not available on Hawaii. */
+ case CHIP_HAWAII:
+ /* HWS depends on CWSR for timely dequeue. CWSR is not
+ * available on Tonga.
+ *
+ * FIXME: This argument also applies to Kaveri.
+ */
+ case CHIP_TONGA:
+ dqm->sched_policy = KFD_SCHED_POLICY_NO_HWS;
+ break;
+ default:
+ dqm->sched_policy = sched_policy;
+ break;
+ }
+
dqm->dev = dev;
- switch (sched_policy) {
+ switch (dqm->sched_policy) {
case KFD_SCHED_POLICY_HWS:
case KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION:
/* initialize dqm for cp scheduling */
@@ -1262,6 +1503,8 @@ struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev)
dqm->ops.set_cache_memory_policy = set_cache_memory_policy;
dqm->ops.set_trap_handler = set_trap_handler;
dqm->ops.process_termination = process_termination_cpsch;
+ dqm->ops.evict_process_queues = evict_process_queues_cpsch;
+ dqm->ops.restore_process_queues = restore_process_queues_cpsch;
break;
case KFD_SCHED_POLICY_NO_HWS:
/* initialize dqm for no cp scheduling */
@@ -1278,9 +1521,12 @@ struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev)
dqm->ops.set_cache_memory_policy = set_cache_memory_policy;
dqm->ops.set_trap_handler = set_trap_handler;
dqm->ops.process_termination = process_termination_nocpsch;
+ dqm->ops.evict_process_queues = evict_process_queues_nocpsch;
+ dqm->ops.restore_process_queues =
+ restore_process_queues_nocpsch;
break;
default:
- pr_err("Invalid scheduling policy %d\n", sched_policy);
+ pr_err("Invalid scheduling policy %d\n", dqm->sched_policy);
goto out_free;
}
@@ -1292,6 +1538,17 @@ struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev)
case CHIP_KAVERI:
device_queue_manager_init_cik(&dqm->asic_ops);
break;
+
+ case CHIP_HAWAII:
+ device_queue_manager_init_cik_hawaii(&dqm->asic_ops);
+ break;
+
+ case CHIP_TONGA:
+ case CHIP_FIJI:
+ case CHIP_POLARIS10:
+ case CHIP_POLARIS11:
+ device_queue_manager_init_vi_tonga(&dqm->asic_ops);
+ break;
default:
WARN(1, "Unexpected ASIC family %u",
dev->device_info->asic_family);