[PATCH 16/25] drm/amdkfd: Implement KFD process eviction/restore
Felix Kuehling
Felix.Kuehling at amd.com
Sat Jan 27 01:09:32 UTC 2018
When the TTM memory manager in KGD evicts BOs, all user mode queues
potentially accessing these BOs must be evicted temporarily. Once
user mode queues are evicted, the eviction fence is signaled,
allowing the migration of the BO to proceed.
A delayed worker is scheduled to restore all the BOs belonging to
the evicted process and restart its queues.
During suspend/resume of the GPU we also evict all processes to allow
KGD to save BOs in system memory, since VRAM will be lost.
Signed-off-by: Harish Kasiviswanathan <Harish.Kasiviswanathan at amd.com>
Signed-off-by: Felix Kuehling <Felix.Kuehling at amd.com>
---
drivers/gpu/drm/amd/amdkfd/kfd_device.c | 65 +++++-
.../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 219 ++++++++++++++++++++-
.../gpu/drm/amd/amdkfd/kfd_device_queue_manager.h | 9 +
drivers/gpu/drm/amd/amdkfd/kfd_module.c | 2 +
drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 32 ++-
drivers/gpu/drm/amd/amdkfd/kfd_process.c | 213 ++++++++++++++++++++
6 files changed, 537 insertions(+), 3 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
index 612afaf..9299a91 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
@@ -32,6 +32,7 @@
#include "cwsr_trap_handler_gfx8.asm"
#define MQD_SIZE_ALIGNED 768
+static atomic_t kfd_device_suspended = ATOMIC_INIT(0);
#if defined(CONFIG_AMD_IOMMU_V2_MODULE) || defined(CONFIG_AMD_IOMMU_V2)
static const struct kfd_device_info kaveri_device_info = {
@@ -545,6 +546,10 @@ void kgd2kfd_suspend(struct kfd_dev *kfd)
if (!kfd->init_complete)
return;
+ /* For first KFD device suspend all the KFD processes */
+ if (atomic_inc_return(&kfd_device_suspended) == 1)
+ kfd_suspend_all_processes();
+
kfd->dqm->ops.stop(kfd->dqm);
#if defined(CONFIG_AMD_IOMMU_V2_MODULE) || defined(CONFIG_AMD_IOMMU_V2)
@@ -561,11 +566,21 @@ void kgd2kfd_suspend(struct kfd_dev *kfd)
int kgd2kfd_resume(struct kfd_dev *kfd)
{
+ int ret, count;
+
if (!kfd->init_complete)
return 0;
- return kfd_resume(kfd);
+ ret = kfd_resume(kfd);
+ if (ret)
+ return ret;
+
+ count = atomic_dec_return(&kfd_device_suspended);
+ WARN_ONCE(count < 0, "KFD suspend / resume ref. error");
+ if (count == 0)
+ ret = kfd_resume_all_processes();
+ return ret;
}
static int kfd_resume(struct kfd_dev *kfd)
@@ -625,6 +640,54 @@ void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry)
spin_unlock(&kfd->interrupt_lock);
}
+/** kgd2kfd_schedule_evict_and_restore_process - Schedules work queue that will
+ * prepare for safe eviction of KFD BOs that belong to the specified
+ * process.
+ *
+ * @mm: mm_struct that identifies the specified KFD process
+ * @fence: eviction fence attached to KFD process BOs
+ *
+ */
+int kgd2kfd_schedule_evict_and_restore_process(struct mm_struct *mm,
+ struct dma_fence *fence)
+{
+ struct kfd_process *p;
+ unsigned long active_time;
+ unsigned long delay_jiffies = msecs_to_jiffies(PROCESS_ACTIVE_TIME_MS);
+
+ if (!fence)
+ return -EINVAL;
+
+ if (dma_fence_is_signaled(fence))
+ return 0;
+
+ p = kfd_lookup_process_by_mm(mm);
+ if (!p)
+ return -ENODEV;
+
+ if (fence->seqno == p->last_eviction_seqno)
+ goto out;
+
+ p->last_eviction_seqno = fence->seqno;
+
+ /* Avoid KFD process starvation. Wait for at least
+ * PROCESS_ACTIVE_TIME_MS before evicting the process again
+ */
+ active_time = get_jiffies_64() - p->last_restore_timestamp;
+ if (delay_jiffies > active_time)
+ delay_jiffies -= active_time;
+ else
+ delay_jiffies = 0;
+
+ /* During process initialization eviction_work.dwork is initialized
+ * to kfd_evict_bo_worker
+ */
+ schedule_delayed_work(&p->eviction_work, delay_jiffies);
+out:
+ kfd_unref_process(p);
+ return 0;
+}
+
static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size,
unsigned int chunk_size)
{
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 b7d0639..b3b6dab 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"
@@ -180,6 +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;
@@ -377,15 +386,29 @@ 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;
@@ -457,6 +480,187 @@ 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)
{
@@ -853,6 +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);
@@ -1291,6 +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 */
@@ -1307,6 +1521,9 @@ 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", dqm->sched_policy);
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h
index 68be0aa..412beff 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h
@@ -79,6 +79,10 @@ struct device_process_node {
*
* @process_termination: Clears all process queues belongs to that device.
*
+ * @evict_process_queues: Evict all active queues of a process
+ *
+ * @restore_process_queues: Restore all evicted queues queues of a process
+ *
*/
struct device_queue_manager_ops {
@@ -129,6 +133,11 @@ struct device_queue_manager_ops {
int (*process_termination)(struct device_queue_manager *dqm,
struct qcm_process_device *qpd);
+
+ int (*evict_process_queues)(struct device_queue_manager *dqm,
+ struct qcm_process_device *qpd);
+ int (*restore_process_queues)(struct device_queue_manager *dqm,
+ struct qcm_process_device *qpd);
};
struct device_queue_manager_asic_ops {
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_module.c b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
index 3ac72be..65574c6 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_module.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
@@ -43,6 +43,8 @@ static const struct kgd2kfd_calls kgd2kfd = {
.interrupt = kgd2kfd_interrupt,
.suspend = kgd2kfd_suspend,
.resume = kgd2kfd_resume,
+ .schedule_evict_and_restore_process =
+ kgd2kfd_schedule_evict_and_restore_process,
};
int sched_policy = KFD_SCHED_POLICY_HWS;
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
index 0687161..785161e 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
@@ -335,7 +335,11 @@ enum kfd_queue_format {
* @is_interop: Defines if this is a interop queue. Interop queue means that
* the queue can access both graphics and compute resources.
*
- * @is_active: Defines if the queue is active or not.
+ * @is_evicted: Defines if the queue is evicted. Only active queues
+ * are evicted, rendering them inactive.
+ *
+ * @is_active: Defines if the queue is active or not. @is_active and
+ * @is_evicted are protected by the DQM lock.
*
* @vmid: If the scheduling mode is no cp scheduling the field defines the vmid
* of the queue.
@@ -357,6 +361,7 @@ struct queue_properties {
uint32_t __iomem *doorbell_ptr;
uint32_t doorbell_off;
bool is_interop;
+ bool is_evicted;
bool is_active;
/* Not relevant for user mode queues in cp scheduling */
unsigned int vmid;
@@ -460,6 +465,7 @@ struct qcm_process_device {
unsigned int queue_count;
unsigned int vmid;
bool is_debug;
+ unsigned int evicted; /* eviction counter, 0=active */
/* This flag tells if we should reset all wavefronts on
* process termination
@@ -486,6 +492,17 @@ struct qcm_process_device {
uint64_t tma_addr;
};
+/* KFD Memory Eviction */
+
+/* Approx. wait time before attempting to restore evicted BOs */
+#define PROCESS_RESTORE_TIME_MS 100
+/* Approx. back off time if restore fails due to lack of memory */
+#define PROCESS_BACK_OFF_TIME_MS 100
+/* Approx. time before evicting the process again */
+#define PROCESS_ACTIVE_TIME_MS 10
+
+int kgd2kfd_schedule_evict_and_restore_process(struct mm_struct *mm,
+ struct dma_fence *fence);
enum kfd_pdd_bound {
PDD_UNBOUND = 0,
@@ -600,6 +617,16 @@ struct kfd_process {
* during restore
*/
struct dma_fence *ef;
+
+ /* Work items for evicting and restoring BOs */
+ struct delayed_work eviction_work;
+ struct delayed_work restore_work;
+ /* seqno of the last scheduled eviction */
+ unsigned int last_eviction_seqno;
+ /* Approx. the last timestamp (in jiffies) when the process was
+ * restored after an eviction
+ */
+ unsigned long last_restore_timestamp;
};
/**
@@ -625,7 +652,10 @@ void kfd_process_destroy_wq(void);
struct kfd_process *kfd_create_process(struct file *filep);
struct kfd_process *kfd_get_process(const struct task_struct *);
struct kfd_process *kfd_lookup_process_by_pasid(unsigned int pasid);
+struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm);
void kfd_unref_process(struct kfd_process *p);
+void kfd_suspend_all_processes(void);
+int kfd_resume_all_processes(void);
struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev,
struct kfd_process *p);
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
index e82f4ac..7eeadfe 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
@@ -55,6 +55,9 @@ static struct kfd_process *create_process(const struct task_struct *thread,
struct file *filep);
static int kfd_process_init_cwsr(struct kfd_process *p, struct file *filep);
+static void evict_process_worker(struct work_struct *work);
+static void restore_process_worker(struct work_struct *work);
+
void kfd_process_create_wq(void)
{
@@ -239,6 +242,9 @@ static void kfd_process_notifier_release(struct mmu_notifier *mn,
mutex_unlock(&kfd_processes_mutex);
synchronize_srcu(&kfd_processes_srcu);
+ cancel_delayed_work_sync(&p->eviction_work);
+ cancel_delayed_work_sync(&p->restore_work);
+
mutex_lock(&p->mutex);
/* Iterate over all process device data structures and if the
@@ -360,6 +366,10 @@ static struct kfd_process *create_process(const struct task_struct *thread,
if (err != 0)
goto err_init_apertures;
+ INIT_DELAYED_WORK(&process->eviction_work, evict_process_worker);
+ INIT_DELAYED_WORK(&process->restore_work, restore_process_worker);
+ process->last_restore_timestamp = get_jiffies_64();
+
err = kfd_process_init_cwsr(process, filep);
if (err)
goto err_init_cwsr;
@@ -411,6 +421,7 @@ struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev,
INIT_LIST_HEAD(&pdd->qpd.priv_queue_list);
pdd->qpd.dqm = dev->dqm;
pdd->qpd.pqm = &p->pqm;
+ pdd->qpd.evicted = 0;
pdd->process = p;
pdd->bound = PDD_UNBOUND;
pdd->already_dequeued = false;
@@ -625,6 +636,208 @@ struct kfd_process *kfd_lookup_process_by_pasid(unsigned int pasid)
return ret_p;
}
+/* This increments the process->ref counter. */
+struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm)
+{
+ struct kfd_process *p;
+
+ int idx = srcu_read_lock(&kfd_processes_srcu);
+
+ p = find_process_by_mm(mm);
+ if (p)
+ kref_get(&p->ref);
+
+ srcu_read_unlock(&kfd_processes_srcu, idx);
+
+ return p;
+}
+
+/* process_evict_queues - Evict all user queues of a process
+ *
+ * Eviction is reference-counted per process-device. This means multiple
+ * evictions from different sources can be nested safely.
+ */
+static int process_evict_queues(struct kfd_process *p)
+{
+ struct kfd_process_device *pdd;
+ int r = 0;
+ unsigned int n_evicted = 0;
+
+ list_for_each_entry(pdd, &p->per_device_data, per_device_list) {
+ r = pdd->dev->dqm->ops.evict_process_queues(pdd->dev->dqm,
+ &pdd->qpd);
+ if (r) {
+ pr_err("Failed to evict process queues\n");
+ goto fail;
+ }
+ n_evicted++;
+ }
+
+ return r;
+
+fail:
+ /* To keep state consistent, roll back partial eviction by
+ * restoring queues
+ */
+ list_for_each_entry(pdd, &p->per_device_data, per_device_list) {
+ if (n_evicted == 0)
+ break;
+ if (pdd->dev->dqm->ops.restore_process_queues(pdd->dev->dqm,
+ &pdd->qpd))
+ pr_err("Failed to restore queues\n");
+
+ n_evicted--;
+ }
+
+ return r;
+}
+
+/* process_restore_queues - Restore all user queues of a process */
+static int process_restore_queues(struct kfd_process *p)
+{
+ struct kfd_process_device *pdd;
+ int r, ret = 0;
+
+ list_for_each_entry(pdd, &p->per_device_data, per_device_list) {
+ r = pdd->dev->dqm->ops.restore_process_queues(pdd->dev->dqm,
+ &pdd->qpd);
+ if (r) {
+ pr_err("Failed to restore process queues\n");
+ if (!ret)
+ ret = r;
+ }
+ }
+
+ return ret;
+}
+
+static void evict_process_worker(struct work_struct *work)
+{
+ int ret;
+ struct kfd_process *p;
+ struct delayed_work *dwork;
+
+ dwork = to_delayed_work(work);
+
+ /* Process termination destroys this worker thread. So during the
+ * lifetime of this thread, kfd_process p will be valid
+ */
+ p = container_of(dwork, struct kfd_process, eviction_work);
+ WARN_ONCE(p->last_eviction_seqno != p->ef->seqno,
+ "Eviction fence mismatch\n");
+
+ /* Narrow window of overlap between restore and evict work
+ * item is possible. Once amdgpu_amdkfd_gpuvm_restore_process_bos
+ * unreserves KFD BOs, it is possible to evicted again. But
+ * restore has few more steps of finish. So lets wait for any
+ * previous restore work to complete
+ */
+ flush_delayed_work(&p->restore_work);
+
+ pr_debug("Started evicting pasid %d\n", p->pasid);
+ ret = process_evict_queues(p);
+ if (!ret) {
+ dma_fence_signal(p->ef);
+ dma_fence_put(p->ef);
+ p->ef = NULL;
+ schedule_delayed_work(&p->restore_work,
+ msecs_to_jiffies(PROCESS_RESTORE_TIME_MS));
+
+ pr_debug("Finished evicting pasid %d\n", p->pasid);
+ } else
+ pr_err("Failed to evict queues of pasid %d\n", p->pasid);
+}
+
+static void restore_process_worker(struct work_struct *work)
+{
+ struct delayed_work *dwork;
+ struct kfd_process *p;
+ struct kfd_process_device *pdd;
+ int ret = 0;
+
+ dwork = to_delayed_work(work);
+
+ /* Process termination destroys this worker thread. So during the
+ * lifetime of this thread, kfd_process p will be valid
+ */
+ p = container_of(dwork, struct kfd_process, restore_work);
+
+ /* Call restore_process_bos on the first KGD device. This function
+ * takes care of restoring the whole process including other devices.
+ * Restore can fail if enough memory is not available. If so,
+ * reschedule again.
+ */
+ pdd = list_first_entry(&p->per_device_data,
+ struct kfd_process_device,
+ per_device_list);
+
+ pr_debug("Started restoring pasid %d\n", p->pasid);
+
+ /* Setting last_restore_timestamp before successful restoration.
+ * Otherwise this would have to be set by KGD (restore_process_bos)
+ * before KFD BOs are unreserved. If not, the process can be evicted
+ * again before the timestamp is set.
+ * If restore fails, the timestamp will be set again in the next
+ * attempt. This would mean that the minimum GPU quanta would be
+ * PROCESS_ACTIVE_TIME_MS - (time to execute the following two
+ * functions)
+ */
+
+ p->last_restore_timestamp = get_jiffies_64();
+ ret = pdd->dev->kfd2kgd->restore_process_bos(p->kgd_process_info,
+ &p->ef);
+ if (ret) {
+ pr_debug("Failed to restore BOs of pasid %d, retry after %d ms\n",
+ p->pasid, PROCESS_BACK_OFF_TIME_MS);
+ ret = schedule_delayed_work(&p->restore_work,
+ msecs_to_jiffies(PROCESS_BACK_OFF_TIME_MS));
+ WARN(!ret, "reschedule restore work failed\n");
+ return;
+ }
+
+ ret = process_restore_queues(p);
+ if (!ret)
+ pr_debug("Finished restoring pasid %d\n", p->pasid);
+ else
+ pr_err("Failed to restore queues of pasid %d\n", p->pasid);
+}
+
+void kfd_suspend_all_processes(void)
+{
+ struct kfd_process *p;
+ unsigned int temp;
+ int idx = srcu_read_lock(&kfd_processes_srcu);
+
+ hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
+ cancel_delayed_work_sync(&p->eviction_work);
+ cancel_delayed_work_sync(&p->restore_work);
+
+ if (process_evict_queues(p))
+ pr_err("Failed to suspend process %d\n", p->pasid);
+ dma_fence_signal(p->ef);
+ dma_fence_put(p->ef);
+ p->ef = NULL;
+ }
+ srcu_read_unlock(&kfd_processes_srcu, idx);
+}
+
+int kfd_resume_all_processes(void)
+{
+ struct kfd_process *p;
+ unsigned int temp;
+ int ret = 0, idx = srcu_read_lock(&kfd_processes_srcu);
+
+ hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
+ if (!schedule_delayed_work(&p->restore_work, 0)) {
+ pr_err("Restore process %d failed during resume\n",
+ p->pasid);
+ ret = -EFAULT;
+ }
+ }
+ srcu_read_unlock(&kfd_processes_srcu, idx);
+ return ret;
+}
+
int kfd_reserved_mem_mmap(struct kfd_process *process,
struct vm_area_struct *vma)
{
--
2.7.4
More information about the amd-gfx
mailing list