* [PATCH V5 01/18] amdkfd: enlarge the hashtable of kfd_process
2025-10-17 8:42 [PATCH V5 00/18] [PATCH V4 00/18] amdkfd: Implement kfd multiple contexts Zhu Lingshan
@ 2025-10-17 8:42 ` Zhu Lingshan
2025-10-17 8:42 ` [PATCH V5 02/18] amdkfd: mark the first kfd_process as the primary one Zhu Lingshan
` (17 subsequent siblings)
18 siblings, 0 replies; 27+ messages in thread
From: Zhu Lingshan @ 2025-10-17 8:42 UTC (permalink / raw)
To: felix.kuehling, alexander.deucher; +Cc: ray.huang, amd-gfx, Zhu Lingshan
This commit enlarges the hashtable size of
kfd_process to 256, because of the multiple
contexts feature allowing each application
create multiple kfd_processes
Signed-off-by: Zhu Lingshan <lingshan.zhu@amd.com>
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
---
drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
index 70ef051511bb..d943ecf62a3b 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
@@ -1020,7 +1020,7 @@ struct kfd_process {
bool gpu_page_fault;
};
-#define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */
+#define KFD_PROCESS_TABLE_SIZE 8 /* bits: 256 entries */
extern DECLARE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE);
extern struct srcu_struct kfd_processes_srcu;
--
2.51.0
^ permalink raw reply related [flat|nested] 27+ messages in thread* [PATCH V5 02/18] amdkfd: mark the first kfd_process as the primary one
2025-10-17 8:42 [PATCH V5 00/18] [PATCH V4 00/18] amdkfd: Implement kfd multiple contexts Zhu Lingshan
2025-10-17 8:42 ` [PATCH V5 01/18] amdkfd: enlarge the hashtable of kfd_process Zhu Lingshan
@ 2025-10-17 8:42 ` Zhu Lingshan
2025-10-17 8:42 ` [PATCH V5 03/18] amdkfd: find_process_by_mm always return the primary context Zhu Lingshan
` (16 subsequent siblings)
18 siblings, 0 replies; 27+ messages in thread
From: Zhu Lingshan @ 2025-10-17 8:42 UTC (permalink / raw)
To: felix.kuehling, alexander.deucher; +Cc: ray.huang, amd-gfx, Zhu Lingshan
The first kfd_process is created through open(),
this commit marks it as the primary kfd_process
by assigning a primary id for its context_id.
Only the primary process should register the mmu_notifier.
Signed-off-by: Zhu Lingshan <lingshan.zhu@amd.com>
---
drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 5 +++++
drivers/gpu/drm/amd/amdkfd/kfd_process.c | 20 ++++++++++++--------
2 files changed, 17 insertions(+), 8 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
index d943ecf62a3b..729330f8a384 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
@@ -1018,9 +1018,14 @@ struct kfd_process {
/* if gpu page fault sent to KFD */
bool gpu_page_fault;
+
+ /*kfd context id */
+ u16 context_id;
};
#define KFD_PROCESS_TABLE_SIZE 8 /* bits: 256 entries */
+#define KFD_CONTEXT_ID_PRIMARY 0xFFFF
+
extern DECLARE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE);
extern struct srcu_struct kfd_processes_srcu;
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
index ddfe30c13e9d..2a22c718ee94 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
@@ -68,7 +68,7 @@ static struct workqueue_struct *kfd_restore_wq;
static struct kfd_process *find_process(const struct task_struct *thread,
bool ref);
static void kfd_process_ref_release(struct kref *ref);
-static struct kfd_process *create_process(const struct task_struct *thread);
+static struct kfd_process *create_process(const struct task_struct *thread, bool primary);
static void evict_process_worker(struct work_struct *work);
static void restore_process_worker(struct work_struct *work);
@@ -867,7 +867,7 @@ struct kfd_process *kfd_create_process(struct task_struct *thread)
if (process) {
pr_debug("Process already found\n");
} else {
- process = create_process(thread);
+ process = create_process(thread, true);
if (IS_ERR(process))
goto out;
@@ -1512,7 +1512,7 @@ void kfd_process_set_trap_debug_flag(struct qcm_process_device *qpd,
* On return the kfd_process is fully operational and will be freed when the
* mm is released
*/
-static struct kfd_process *create_process(const struct task_struct *thread)
+static struct kfd_process *create_process(const struct task_struct *thread, bool primary)
{
struct kfd_process *process;
struct mmu_notifier *mn;
@@ -1528,6 +1528,7 @@ static struct kfd_process *create_process(const struct task_struct *thread)
process->lead_thread = thread->group_leader;
process->n_pdds = 0;
process->queues_paused = false;
+
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();
@@ -1571,12 +1572,15 @@ static struct kfd_process *create_process(const struct task_struct *thread)
* After this point, mmu_notifier_put will trigger the cleanup by
* dropping the last process reference in the free_notifier.
*/
- mn = mmu_notifier_get(&kfd_process_mmu_notifier_ops, process->mm);
- if (IS_ERR(mn)) {
- err = PTR_ERR(mn);
- goto err_register_notifier;
+ if (primary) {
+ process->context_id = KFD_CONTEXT_ID_PRIMARY;
+ mn = mmu_notifier_get(&kfd_process_mmu_notifier_ops, process->mm);
+ if (IS_ERR(mn)) {
+ err = PTR_ERR(mn);
+ goto err_register_notifier;
+ }
+ BUG_ON(mn != &process->mmu_notifier);
}
- BUG_ON(mn != &process->mmu_notifier);
kfd_unref_process(process);
get_task_struct(process->lead_thread);
--
2.51.0
^ permalink raw reply related [flat|nested] 27+ messages in thread* [PATCH V5 03/18] amdkfd: find_process_by_mm always return the primary context
2025-10-17 8:42 [PATCH V5 00/18] [PATCH V4 00/18] amdkfd: Implement kfd multiple contexts Zhu Lingshan
2025-10-17 8:42 ` [PATCH V5 01/18] amdkfd: enlarge the hashtable of kfd_process Zhu Lingshan
2025-10-17 8:42 ` [PATCH V5 02/18] amdkfd: mark the first kfd_process as the primary one Zhu Lingshan
@ 2025-10-17 8:42 ` Zhu Lingshan
2025-10-17 8:42 ` [PATCH V5 04/18] amdkfd: Introduce kfd_create_process_sysfs as a separate function Zhu Lingshan
` (15 subsequent siblings)
18 siblings, 0 replies; 27+ messages in thread
From: Zhu Lingshan @ 2025-10-17 8:42 UTC (permalink / raw)
To: felix.kuehling, alexander.deucher; +Cc: ray.huang, amd-gfx, Zhu Lingshan
Up until this commit, the kfd multiple contexts feature has
not been fully implemented in mainline kernel yet.
For backawrd compatibility, not break existing use cases,
this commit changes function find_process_by_mm, let it
always return the primary kfd_process.
Signed-off-by: Zhu Lingshan <lingshan.zhu@amd.com>
---
drivers/gpu/drm/amd/amdkfd/kfd_process.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
index 2a22c718ee94..051000e79e95 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
@@ -935,7 +935,7 @@ static struct kfd_process *find_process_by_mm(const struct mm_struct *mm)
hash_for_each_possible_rcu(kfd_processes_table, process,
kfd_processes, (uintptr_t)mm)
- if (process->mm == mm)
+ if (process->mm == mm && process->context_id == KFD_CONTEXT_ID_PRIMARY)
return process;
return NULL;
--
2.51.0
^ permalink raw reply related [flat|nested] 27+ messages in thread* [PATCH V5 04/18] amdkfd: Introduce kfd_create_process_sysfs as a separate function
2025-10-17 8:42 [PATCH V5 00/18] [PATCH V4 00/18] amdkfd: Implement kfd multiple contexts Zhu Lingshan
` (2 preceding siblings ...)
2025-10-17 8:42 ` [PATCH V5 03/18] amdkfd: find_process_by_mm always return the primary context Zhu Lingshan
@ 2025-10-17 8:42 ` Zhu Lingshan
2025-10-17 8:42 ` [PATCH V5 05/18] amdkfd: destroy kfd secondary contexts through fd close Zhu Lingshan
` (14 subsequent siblings)
18 siblings, 0 replies; 27+ messages in thread
From: Zhu Lingshan @ 2025-10-17 8:42 UTC (permalink / raw)
To: felix.kuehling, alexander.deucher; +Cc: ray.huang, amd-gfx, Zhu Lingshan
KFD creates sysfs entries for a kfd_process in
function kfd_create_process when creating it.
This commit extracts the code creating sysfs
entries to a separate function because it
would be invoked in other code path like
creating secondary kfd contexts (kfd_process).
Signed-off-by: Zhu Lingshan <lingshan.zhu@amd.com>
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
---
drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 1 +
drivers/gpu/drm/amd/amdkfd/kfd_process.c | 66 +++++++++++++++---------
2 files changed, 42 insertions(+), 25 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
index 729330f8a384..cfedc0c5892c 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
@@ -1054,6 +1054,7 @@ int kfd_process_create_wq(void);
void kfd_process_destroy_wq(void);
void kfd_cleanup_processes(void);
struct kfd_process *kfd_create_process(struct task_struct *thread);
+int kfd_create_process_sysfs(struct kfd_process *process);
struct kfd_process *kfd_get_process(const struct task_struct *task);
struct kfd_process *kfd_lookup_process_by_pasid(u32 pasid,
struct kfd_process_device **pdd);
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
index 051000e79e95..83fe571b801c 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
@@ -825,6 +825,44 @@ static void kfd_process_device_destroy_ib_mem(struct kfd_process_device *pdd)
kfd_process_free_gpuvm(qpd->ib_mem, pdd, &qpd->ib_kaddr);
}
+int kfd_create_process_sysfs(struct kfd_process *process)
+{
+ int ret;
+
+ if (process->kobj) {
+ pr_warn("kobject already exsists for the kfd_process\n");
+ return -EINVAL;
+ }
+
+ process->kobj = kfd_alloc_struct(process->kobj);
+ if (!process->kobj) {
+ pr_warn("Creating procfs kobject failed");
+ return -ENOMEM;
+ }
+ ret = kobject_init_and_add(process->kobj, &procfs_type,
+ procfs.kobj, "%d",
+ (int)process->lead_thread->pid);
+ if (ret) {
+ pr_warn("Creating procfs pid directory failed");
+ kobject_put(process->kobj);
+ return ret;
+ }
+
+ kfd_sysfs_create_file(process->kobj, &process->attr_pasid,
+ "pasid");
+
+ process->kobj_queues = kobject_create_and_add("queues",
+ process->kobj);
+ if (!process->kobj_queues)
+ pr_warn("Creating KFD proc/queues folder failed");
+
+ kfd_procfs_add_sysfs_stats(process);
+ kfd_procfs_add_sysfs_files(process);
+ kfd_procfs_add_sysfs_counters(process);
+
+ return 0;
+}
+
struct kfd_process *kfd_create_process(struct task_struct *thread)
{
struct kfd_process *process;
@@ -874,31 +912,9 @@ struct kfd_process *kfd_create_process(struct task_struct *thread)
if (!procfs.kobj)
goto out;
- process->kobj = kfd_alloc_struct(process->kobj);
- if (!process->kobj) {
- pr_warn("Creating procfs kobject failed");
- goto out;
- }
- ret = kobject_init_and_add(process->kobj, &procfs_type,
- procfs.kobj, "%d",
- (int)process->lead_thread->pid);
- if (ret) {
- pr_warn("Creating procfs pid directory failed");
- kobject_put(process->kobj);
- goto out;
- }
-
- kfd_sysfs_create_file(process->kobj, &process->attr_pasid,
- "pasid");
-
- process->kobj_queues = kobject_create_and_add("queues",
- process->kobj);
- if (!process->kobj_queues)
- pr_warn("Creating KFD proc/queues folder failed");
-
- kfd_procfs_add_sysfs_stats(process);
- kfd_procfs_add_sysfs_files(process);
- kfd_procfs_add_sysfs_counters(process);
+ ret = kfd_create_process_sysfs(process);
+ if (ret)
+ pr_warn("Failed to create sysfs entry for the kfd_process");
kfd_debugfs_add_process(process);
--
2.51.0
^ permalink raw reply related [flat|nested] 27+ messages in thread* [PATCH V5 05/18] amdkfd: destroy kfd secondary contexts through fd close
2025-10-17 8:42 [PATCH V5 00/18] [PATCH V4 00/18] amdkfd: Implement kfd multiple contexts Zhu Lingshan
` (3 preceding siblings ...)
2025-10-17 8:42 ` [PATCH V5 04/18] amdkfd: Introduce kfd_create_process_sysfs as a separate function Zhu Lingshan
@ 2025-10-17 8:42 ` Zhu Lingshan
2025-10-17 8:42 ` [PATCH V5 06/18] amdkfd: process svm ioctl only on the primary kfd process Zhu Lingshan
` (13 subsequent siblings)
18 siblings, 0 replies; 27+ messages in thread
From: Zhu Lingshan @ 2025-10-17 8:42 UTC (permalink / raw)
To: felix.kuehling, alexander.deucher; +Cc: ray.huang, amd-gfx, Zhu Lingshan
Life cycle of a KFD secondary context(kfd_process) is tied
to the opened file. Therefore this commit destroy a kfd
secondary context when close the fd it belonging to.
This commit extracts the code removing the kfd_process
from the kfd_process_table to a separate function and
call it in kfd_process_notifier_release_internal unconditionally.
Signed-off-by: Zhu Lingshan <lingshan.zhu@amd.com>
---
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 9 ++++--
drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 1 +
drivers/gpu/drm/amd/amdkfd/kfd_process.c | 41 +++++++++++++-----------
3 files changed, 31 insertions(+), 20 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
index 0f0719528bcc..158e75975557 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
@@ -164,8 +164,13 @@ static int kfd_release(struct inode *inode, struct file *filep)
{
struct kfd_process *process = filep->private_data;
- if (process)
- kfd_unref_process(process);
+ if (!process)
+ return 0;
+
+ if (process->context_id != KFD_CONTEXT_ID_PRIMARY)
+ kfd_process_notifier_release_internal(process);
+
+ kfd_unref_process(process);
return 0;
}
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
index cfedc0c5892c..919510f18249 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
@@ -1096,6 +1096,7 @@ bool kfd_process_xnack_mode(struct kfd_process *p, bool supported);
int kfd_reserved_mem_mmap(struct kfd_node *dev, struct kfd_process *process,
struct vm_area_struct *vma);
+void kfd_process_notifier_release_internal(struct kfd_process *p);
/* KFD process API for creating and translating handles */
int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd,
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
index 83fe571b801c..995d27be06e3 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
@@ -1235,10 +1235,30 @@ static void kfd_process_free_notifier(struct mmu_notifier *mn)
kfd_unref_process(container_of(mn, struct kfd_process, mmu_notifier));
}
-static void kfd_process_notifier_release_internal(struct kfd_process *p)
+static void kfd_process_table_remove(struct kfd_process *p)
+{
+ mutex_lock(&kfd_processes_mutex);
+ /*
+ * Do early return if table is empty.
+ *
+ * This could potentially happen if this function is called concurrently
+ * by mmu_notifier and by kfd_cleanup_pocesses.
+ *
+ */
+ if (hash_empty(kfd_processes_table)) {
+ mutex_unlock(&kfd_processes_mutex);
+ return;
+ }
+ hash_del_rcu(&p->kfd_processes);
+ mutex_unlock(&kfd_processes_mutex);
+ synchronize_srcu(&kfd_processes_srcu);
+}
+
+void kfd_process_notifier_release_internal(struct kfd_process *p)
{
int i;
+ kfd_process_table_remove(p);
cancel_delayed_work_sync(&p->eviction_work);
cancel_delayed_work_sync(&p->restore_work);
@@ -1272,7 +1292,8 @@ static void kfd_process_notifier_release_internal(struct kfd_process *p)
srcu_read_unlock(&kfd_processes_srcu, idx);
}
- mmu_notifier_put(&p->mmu_notifier);
+ if (p->context_id == KFD_CONTEXT_ID_PRIMARY)
+ mmu_notifier_put(&p->mmu_notifier);
}
static void kfd_process_notifier_release(struct mmu_notifier *mn,
@@ -1288,22 +1309,6 @@ static void kfd_process_notifier_release(struct mmu_notifier *mn,
if (WARN_ON(p->mm != mm))
return;
- mutex_lock(&kfd_processes_mutex);
- /*
- * Do early return if table is empty.
- *
- * This could potentially happen if this function is called concurrently
- * by mmu_notifier and by kfd_cleanup_pocesses.
- *
- */
- if (hash_empty(kfd_processes_table)) {
- mutex_unlock(&kfd_processes_mutex);
- return;
- }
- hash_del_rcu(&p->kfd_processes);
- mutex_unlock(&kfd_processes_mutex);
- synchronize_srcu(&kfd_processes_srcu);
-
kfd_process_notifier_release_internal(p);
}
--
2.51.0
^ permalink raw reply related [flat|nested] 27+ messages in thread* [PATCH V5 06/18] amdkfd: process svm ioctl only on the primary kfd process
2025-10-17 8:42 [PATCH V5 00/18] [PATCH V4 00/18] amdkfd: Implement kfd multiple contexts Zhu Lingshan
` (4 preceding siblings ...)
2025-10-17 8:42 ` [PATCH V5 05/18] amdkfd: destroy kfd secondary contexts through fd close Zhu Lingshan
@ 2025-10-17 8:42 ` Zhu Lingshan
2025-10-17 8:42 ` [PATCH V5 07/18] amdkfd: process USERPTR allocation " Zhu Lingshan
` (12 subsequent siblings)
18 siblings, 0 replies; 27+ messages in thread
From: Zhu Lingshan @ 2025-10-17 8:42 UTC (permalink / raw)
To: felix.kuehling, alexander.deucher; +Cc: ray.huang, amd-gfx, Zhu Lingshan
svm ioctl should only be processed on the primary
kfd process because only the lifecycle of the
primary kfd process is tied to the user space
applicaiton.
Another reason is in virtualization the hypervisor owns
the primary kfd process as a privileged one.
Signed-off-by: Zhu Lingshan <lingshan.zhu@amd.com>
---
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 6 ++++++
1 file changed, 6 insertions(+)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
index 158e75975557..58d90bebdff7 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
@@ -1718,6 +1718,12 @@ static int kfd_ioctl_svm(struct file *filep, struct kfd_process *p, void *data)
struct kfd_ioctl_svm_args *args = data;
int r = 0;
+ if (p->context_id != KFD_CONTEXT_ID_PRIMARY) {
+ pr_debug("SVM ioctl not supported on non-primary kfd process\n");
+
+ return -EOPNOTSUPP;
+ }
+
pr_debug("start 0x%llx size 0x%llx op 0x%x nattr 0x%x\n",
args->start_addr, args->size, args->op, args->nattr);
--
2.51.0
^ permalink raw reply related [flat|nested] 27+ messages in thread* [PATCH V5 07/18] amdkfd: process USERPTR allocation only on the primary kfd process
2025-10-17 8:42 [PATCH V5 00/18] [PATCH V4 00/18] amdkfd: Implement kfd multiple contexts Zhu Lingshan
` (5 preceding siblings ...)
2025-10-17 8:42 ` [PATCH V5 06/18] amdkfd: process svm ioctl only on the primary kfd process Zhu Lingshan
@ 2025-10-17 8:42 ` Zhu Lingshan
2025-10-17 8:42 ` [PATCH V5 08/18] amdkfd: identify a secondary kfd process by its id Zhu Lingshan
` (11 subsequent siblings)
18 siblings, 0 replies; 27+ messages in thread
From: Zhu Lingshan @ 2025-10-17 8:42 UTC (permalink / raw)
To: felix.kuehling, alexander.deucher; +Cc: ray.huang, amd-gfx, Zhu Lingshan
The lifecycle of the primary kfd process is tied to
the user space program, all secondary kfd process
would be destroyed when fd close. Thus only the primary
kfd process should process USERPTR memory allocation.
Signed-off-by: Zhu Lingshan <lingshan.zhu@amd.com>
---
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 6 ++++++
1 file changed, 6 insertions(+)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
index 58d90bebdff7..d1cb22f48b00 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
@@ -1062,6 +1062,12 @@ static int kfd_ioctl_alloc_memory_of_gpu(struct file *filep,
if (args->size == 0)
return -EINVAL;
+ if (p->context_id != KFD_CONTEXT_ID_PRIMARY && (flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR)) {
+ pr_debug("USERPTR is not supported on non-primary kfd_process\n");
+
+ return -EOPNOTSUPP;
+ }
+
#if IS_ENABLED(CONFIG_HSA_AMD_SVM)
/* Flush pending deferred work to avoid racing with deferred actions
* from previous memory map changes (e.g. munmap).
--
2.51.0
^ permalink raw reply related [flat|nested] 27+ messages in thread* [PATCH V5 08/18] amdkfd: identify a secondary kfd process by its id
2025-10-17 8:42 [PATCH V5 00/18] [PATCH V4 00/18] amdkfd: Implement kfd multiple contexts Zhu Lingshan
` (6 preceding siblings ...)
2025-10-17 8:42 ` [PATCH V5 07/18] amdkfd: process USERPTR allocation " Zhu Lingshan
@ 2025-10-17 8:42 ` Zhu Lingshan
2025-10-17 23:16 ` Felix Kuehling
2025-10-17 8:42 ` [PATCH V5 09/18] amdkfd: find kfd_process by filep->private_data in kfd_mmap Zhu Lingshan
` (10 subsequent siblings)
18 siblings, 1 reply; 27+ messages in thread
From: Zhu Lingshan @ 2025-10-17 8:42 UTC (permalink / raw)
To: felix.kuehling, alexander.deucher; +Cc: ray.huang, amd-gfx, Zhu Lingshan
This commit introduces a new id field for
struct kfd process, which helps identify
a kfd process among multiple contexts that
all belong to a single user space program.
The sysfs entry of a secondary kfd process
is placed under the sysfs entry folder of
its primary kfd process.
The naming format of the sysfs entry of a secondary
kfd process is "context_%u" where %u is the process id.
Signed-off-by: Zhu Lingshan <lingshan.zhu@amd.com>
---
drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 6 ++
drivers/gpu/drm/amd/amdkfd/kfd_process.c | 78 +++++++++++++++++++++++-
2 files changed, 81 insertions(+), 3 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
index 919510f18249..9de658119cd9 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
@@ -1021,10 +1021,16 @@ struct kfd_process {
/*kfd context id */
u16 context_id;
+
+ /* The primary kfd_process allocating IDs for its secondary kfd_process, 0 for primary kfd_process */
+ struct ida id_table;
+
};
#define KFD_PROCESS_TABLE_SIZE 8 /* bits: 256 entries */
#define KFD_CONTEXT_ID_PRIMARY 0xFFFF
+#define KFD_CONTEXT_ID_MIN 0
+#define KFD_CONTEXT_ID_WIDTH 16
extern DECLARE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE);
extern struct srcu_struct kfd_processes_srcu;
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
index 995d27be06e3..157145c94314 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
@@ -827,6 +827,7 @@ static void kfd_process_device_destroy_ib_mem(struct kfd_process_device *pdd)
int kfd_create_process_sysfs(struct kfd_process *process)
{
+ struct kfd_process *primary_process;
int ret;
if (process->kobj) {
@@ -839,9 +840,22 @@ int kfd_create_process_sysfs(struct kfd_process *process)
pr_warn("Creating procfs kobject failed");
return -ENOMEM;
}
- ret = kobject_init_and_add(process->kobj, &procfs_type,
- procfs.kobj, "%d",
- (int)process->lead_thread->pid);
+
+ if (process->context_id == KFD_CONTEXT_ID_PRIMARY)
+ ret = kobject_init_and_add(process->kobj, &procfs_type,
+ procfs.kobj, "%d",
+ (int)process->lead_thread->pid);
+ else {
+ primary_process = kfd_lookup_process_by_mm(process->lead_thread->mm);
+ if (!primary_process)
+ return -ESRCH;
+
+ ret = kobject_init_and_add(process->kobj, &procfs_type,
+ primary_process->kobj, "context_%u",
+ process->context_id);
+ kfd_unref_process(primary_process);
+ }
+
if (ret) {
pr_warn("Creating procfs pid directory failed");
kobject_put(process->kobj);
@@ -863,6 +877,50 @@ int kfd_create_process_sysfs(struct kfd_process *process)
return 0;
}
+static int kfd_process_alloc_id(struct kfd_process *process)
+{
+ int ret;
+ struct kfd_process *primary_process;
+
+ /* already assign 0xFFFF when create */
+ if (process->context_id == KFD_CONTEXT_ID_PRIMARY)
+ return 0;
+
+ primary_process = kfd_lookup_process_by_mm(process->lead_thread->mm);
+ if (!primary_process)
+ return -ESRCH;
+
+ /* id range: KFD_CONTEXT_ID_MIN to 0xFFFE */
+ ret = ida_alloc_range(&primary_process->id_table, KFD_CONTEXT_ID_MIN,
+ (1 << KFD_CONTEXT_ID_WIDTH) - 2, GFP_KERNEL);
+ if (ret < 0)
+ goto out;
+
+ process->context_id = ret;
+ ret = 0;
+
+out:
+ kfd_unref_process(primary_process);
+
+ return ret;
+}
+
+static void kfd_process_free_id(struct kfd_process *process)
+{
+ struct kfd_process *primary_process;
+
+ if (process->context_id != KFD_CONTEXT_ID_PRIMARY)
+ return;
+
+ primary_process = kfd_lookup_process_by_mm(process->lead_thread->mm);
+ if (!primary_process)
+ return;
+
+ ida_free(&primary_process->id_table, process->context_id);
+
+ kfd_unref_process(primary_process);
+}
+
struct kfd_process *kfd_create_process(struct task_struct *thread)
{
struct kfd_process *process;
@@ -1195,6 +1253,11 @@ static void kfd_process_wq_release(struct work_struct *work)
if (ef)
dma_fence_signal(ef);
+ if (p->context_id != KFD_CONTEXT_ID_PRIMARY)
+ kfd_process_free_id(p);
+ else
+ ida_destroy(&p->id_table);
+
kfd_process_remove_sysfs(p);
kfd_debugfs_remove_process(p);
@@ -1601,6 +1664,13 @@ static struct kfd_process *create_process(const struct task_struct *thread, bool
goto err_register_notifier;
}
BUG_ON(mn != &process->mmu_notifier);
+ ida_init(&process->id_table);
+ }
+
+ err = kfd_process_alloc_id(process);
+ if (err) {
+ pr_err("Creating kfd process: failed to alloc an id\n");
+ goto err_alloc_id;
}
kfd_unref_process(process);
@@ -1610,6 +1680,8 @@ static struct kfd_process *create_process(const struct task_struct *thread, bool
return process;
+err_alloc_id:
+ kfd_process_free_id(process);
err_register_notifier:
hash_del_rcu(&process->kfd_processes);
svm_range_list_fini(process);
--
2.51.0
^ permalink raw reply related [flat|nested] 27+ messages in thread* Re: [PATCH V5 08/18] amdkfd: identify a secondary kfd process by its id
2025-10-17 8:42 ` [PATCH V5 08/18] amdkfd: identify a secondary kfd process by its id Zhu Lingshan
@ 2025-10-17 23:16 ` Felix Kuehling
2025-10-22 6:56 ` Zhu, Lingshan
0 siblings, 1 reply; 27+ messages in thread
From: Felix Kuehling @ 2025-10-17 23:16 UTC (permalink / raw)
To: Zhu Lingshan, alexander.deucher; +Cc: ray.huang, amd-gfx
On 2025-10-17 04:42, Zhu Lingshan wrote:
> This commit introduces a new id field for
> struct kfd process, which helps identify
> a kfd process among multiple contexts that
> all belong to a single user space program.
>
> The sysfs entry of a secondary kfd process
> is placed under the sysfs entry folder of
> its primary kfd process.
>
> The naming format of the sysfs entry of a secondary
> kfd process is "context_%u" where %u is the process id.
>
> Signed-off-by: Zhu Lingshan <lingshan.zhu@amd.com>
> ---
> drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 6 ++
> drivers/gpu/drm/amd/amdkfd/kfd_process.c | 78 +++++++++++++++++++++++-
> 2 files changed, 81 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
> index 919510f18249..9de658119cd9 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
> @@ -1021,10 +1021,16 @@ struct kfd_process {
>
> /*kfd context id */
> u16 context_id;
> +
> + /* The primary kfd_process allocating IDs for its secondary kfd_process, 0 for primary kfd_process */
> + struct ida id_table;
> +
> };
>
> #define KFD_PROCESS_TABLE_SIZE 8 /* bits: 256 entries */
> #define KFD_CONTEXT_ID_PRIMARY 0xFFFF
> +#define KFD_CONTEXT_ID_MIN 0
> +#define KFD_CONTEXT_ID_WIDTH 16
>
> extern DECLARE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE);
> extern struct srcu_struct kfd_processes_srcu;
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
> index 995d27be06e3..157145c94314 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
> @@ -827,6 +827,7 @@ static void kfd_process_device_destroy_ib_mem(struct kfd_process_device *pdd)
>
> int kfd_create_process_sysfs(struct kfd_process *process)
> {
> + struct kfd_process *primary_process;
> int ret;
>
> if (process->kobj) {
> @@ -839,9 +840,22 @@ int kfd_create_process_sysfs(struct kfd_process *process)
> pr_warn("Creating procfs kobject failed");
> return -ENOMEM;
> }
> - ret = kobject_init_and_add(process->kobj, &procfs_type,
> - procfs.kobj, "%d",
> - (int)process->lead_thread->pid);
> +
> + if (process->context_id == KFD_CONTEXT_ID_PRIMARY)
> + ret = kobject_init_and_add(process->kobj, &procfs_type,
> + procfs.kobj, "%d",
> + (int)process->lead_thread->pid);
> + else {
> + primary_process = kfd_lookup_process_by_mm(process->lead_thread->mm);
> + if (!primary_process)
> + return -ESRCH;
> +
> + ret = kobject_init_and_add(process->kobj, &procfs_type,
> + primary_process->kobj, "context_%u",
> + process->context_id);
> + kfd_unref_process(primary_process);
> + }
> +
> if (ret) {
> pr_warn("Creating procfs pid directory failed");
> kobject_put(process->kobj);
> @@ -863,6 +877,50 @@ int kfd_create_process_sysfs(struct kfd_process *process)
> return 0;
> }
>
> +static int kfd_process_alloc_id(struct kfd_process *process)
> +{
> + int ret;
> + struct kfd_process *primary_process;
> +
> + /* already assign 0xFFFF when create */
> + if (process->context_id == KFD_CONTEXT_ID_PRIMARY)
> + return 0;
> +
> + primary_process = kfd_lookup_process_by_mm(process->lead_thread->mm);
> + if (!primary_process)
> + return -ESRCH;
> +
> + /* id range: KFD_CONTEXT_ID_MIN to 0xFFFE */
> + ret = ida_alloc_range(&primary_process->id_table, KFD_CONTEXT_ID_MIN,
> + (1 << KFD_CONTEXT_ID_WIDTH) - 2, GFP_KERNEL);
This would be safer and more obvious if you just set the upper limit as
KFD_CONTEXT_ID_PRIMARY - 1. Then you don't need KFD_CONTEXT_ID_WIDTH at all.
Regards,
Felix
> + if (ret < 0)
> + goto out;
> +
> + process->context_id = ret;
> + ret = 0;
> +
> +out:
> + kfd_unref_process(primary_process);
> +
> + return ret;
> +}
> +
> +static void kfd_process_free_id(struct kfd_process *process)
> +{
> + struct kfd_process *primary_process;
> +
> + if (process->context_id != KFD_CONTEXT_ID_PRIMARY)
> + return;
> +
> + primary_process = kfd_lookup_process_by_mm(process->lead_thread->mm);
> + if (!primary_process)
> + return;
> +
> + ida_free(&primary_process->id_table, process->context_id);
> +
> + kfd_unref_process(primary_process);
> +}
> +
> struct kfd_process *kfd_create_process(struct task_struct *thread)
> {
> struct kfd_process *process;
> @@ -1195,6 +1253,11 @@ static void kfd_process_wq_release(struct work_struct *work)
> if (ef)
> dma_fence_signal(ef);
>
> + if (p->context_id != KFD_CONTEXT_ID_PRIMARY)
> + kfd_process_free_id(p);
> + else
> + ida_destroy(&p->id_table);
> +
> kfd_process_remove_sysfs(p);
> kfd_debugfs_remove_process(p);
>
> @@ -1601,6 +1664,13 @@ static struct kfd_process *create_process(const struct task_struct *thread, bool
> goto err_register_notifier;
> }
> BUG_ON(mn != &process->mmu_notifier);
> + ida_init(&process->id_table);
> + }
> +
> + err = kfd_process_alloc_id(process);
> + if (err) {
> + pr_err("Creating kfd process: failed to alloc an id\n");
> + goto err_alloc_id;
> }
>
> kfd_unref_process(process);
> @@ -1610,6 +1680,8 @@ static struct kfd_process *create_process(const struct task_struct *thread, bool
>
> return process;
>
> +err_alloc_id:
> + kfd_process_free_id(process);
> err_register_notifier:
> hash_del_rcu(&process->kfd_processes);
> svm_range_list_fini(process);
^ permalink raw reply [flat|nested] 27+ messages in thread* Re: [PATCH V5 08/18] amdkfd: identify a secondary kfd process by its id
2025-10-17 23:16 ` Felix Kuehling
@ 2025-10-22 6:56 ` Zhu, Lingshan
0 siblings, 0 replies; 27+ messages in thread
From: Zhu, Lingshan @ 2025-10-22 6:56 UTC (permalink / raw)
To: Felix Kuehling, alexander.deucher; +Cc: ray.huang, amd-gfx
[-- Attachment #1: Type: text/plain, Size: 6619 bytes --]
On 10/18/2025 7:16 AM, Felix Kuehling wrote:
> On 2025-10-17 04:42, Zhu Lingshan wrote:
>> This commit introduces a new id field for
>> struct kfd process, which helps identify
>> a kfd process among multiple contexts that
>> all belong to a single user space program.
>>
>> The sysfs entry of a secondary kfd process
>> is placed under the sysfs entry folder of
>> its primary kfd process.
>>
>> The naming format of the sysfs entry of a secondary
>> kfd process is "context_%u" where %u is the process id.
>>
>> Signed-off-by: Zhu Lingshan <lingshan.zhu@amd.com>
>> ---
>> drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 6 ++
>> drivers/gpu/drm/amd/amdkfd/kfd_process.c | 78 +++++++++++++++++++++++-
>> 2 files changed, 81 insertions(+), 3 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
>> b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
>> index 919510f18249..9de658119cd9 100644
>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
>> @@ -1021,10 +1021,16 @@ struct kfd_process {
>> /*kfd context id */
>> u16 context_id;
>> +
>> + /* The primary kfd_process allocating IDs for its secondary
>> kfd_process, 0 for primary kfd_process */
>> + struct ida id_table;
>> +
>> };
>> #define KFD_PROCESS_TABLE_SIZE 8 /* bits: 256 entries */
>> #define KFD_CONTEXT_ID_PRIMARY 0xFFFF
>> +#define KFD_CONTEXT_ID_MIN 0
>> +#define KFD_CONTEXT_ID_WIDTH 16
>> extern DECLARE_HASHTABLE(kfd_processes_table,
>> KFD_PROCESS_TABLE_SIZE);
>> extern struct srcu_struct kfd_processes_srcu;
>> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c
>> b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
>> index 995d27be06e3..157145c94314 100644
>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c
>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
>> @@ -827,6 +827,7 @@ static void
>> kfd_process_device_destroy_ib_mem(struct kfd_process_device *pdd)
>> int kfd_create_process_sysfs(struct kfd_process *process)
>> {
>> + struct kfd_process *primary_process;
>> int ret;
>> if (process->kobj) {
>> @@ -839,9 +840,22 @@ int kfd_create_process_sysfs(struct kfd_process
>> *process)
>> pr_warn("Creating procfs kobject failed");
>> return -ENOMEM;
>> }
>> - ret = kobject_init_and_add(process->kobj, &procfs_type,
>> - procfs.kobj, "%d",
>> - (int)process->lead_thread->pid);
>> +
>> + if (process->context_id == KFD_CONTEXT_ID_PRIMARY)
>> + ret = kobject_init_and_add(process->kobj, &procfs_type,
>> + procfs.kobj, "%d",
>> + (int)process->lead_thread->pid);
>> + else {
>> + primary_process =
>> kfd_lookup_process_by_mm(process->lead_thread->mm);
>> + if (!primary_process)
>> + return -ESRCH;
>> +
>> + ret = kobject_init_and_add(process->kobj, &procfs_type,
>> + primary_process->kobj, "context_%u",
>> + process->context_id);
>> + kfd_unref_process(primary_process);
>> + }
>> +
>> if (ret) {
>> pr_warn("Creating procfs pid directory failed");
>> kobject_put(process->kobj);
>> @@ -863,6 +877,50 @@ int kfd_create_process_sysfs(struct kfd_process
>> *process)
>> return 0;
>> }
>> +static int kfd_process_alloc_id(struct kfd_process *process)
>> +{
>> + int ret;
>> + struct kfd_process *primary_process;
>> +
>> + /* already assign 0xFFFF when create */
>> + if (process->context_id == KFD_CONTEXT_ID_PRIMARY)
>> + return 0;
>> +
>> + primary_process =
>> kfd_lookup_process_by_mm(process->lead_thread->mm);
>> + if (!primary_process)
>> + return -ESRCH;
>> +
>> + /* id range: KFD_CONTEXT_ID_MIN to 0xFFFE */
>> + ret = ida_alloc_range(&primary_process->id_table,
>> KFD_CONTEXT_ID_MIN,
>> + (1 << KFD_CONTEXT_ID_WIDTH) - 2, GFP_KERNEL);
>
> This would be safer and more obvious if you just set the upper limit
> as KFD_CONTEXT_ID_PRIMARY - 1. Then you don't need
> KFD_CONTEXT_ID_WIDTH at all.
will fix in V6, thanks!
>
> Regards,
> Felix
>
>
>> + if (ret < 0)
>> + goto out;
>> +
>> + process->context_id = ret;
>> + ret = 0;
>> +
>> +out:
>> + kfd_unref_process(primary_process);
>> +
>> + return ret;
>> +}
>> +
>> +static void kfd_process_free_id(struct kfd_process *process)
>> +{
>> + struct kfd_process *primary_process;
>> +
>> + if (process->context_id != KFD_CONTEXT_ID_PRIMARY)
>> + return;
>> +
>> + primary_process =
>> kfd_lookup_process_by_mm(process->lead_thread->mm);
>> + if (!primary_process)
>> + return;
>> +
>> + ida_free(&primary_process->id_table, process->context_id);
>> +
>> + kfd_unref_process(primary_process);
>> +}
>> +
>> struct kfd_process *kfd_create_process(struct task_struct *thread)
>> {
>> struct kfd_process *process;
>> @@ -1195,6 +1253,11 @@ static void kfd_process_wq_release(struct
>> work_struct *work)
>> if (ef)
>> dma_fence_signal(ef);
>> + if (p->context_id != KFD_CONTEXT_ID_PRIMARY)
>> + kfd_process_free_id(p);
>> + else
>> + ida_destroy(&p->id_table);
>> +
>> kfd_process_remove_sysfs(p);
>> kfd_debugfs_remove_process(p);
>> @@ -1601,6 +1664,13 @@ static struct kfd_process
>> *create_process(const struct task_struct *thread, bool
>> goto err_register_notifier;
>> }
>> BUG_ON(mn != &process->mmu_notifier);
>> + ida_init(&process->id_table);
>> + }
>> +
>> + err = kfd_process_alloc_id(process);
>> + if (err) {
>> + pr_err("Creating kfd process: failed to alloc an id\n");
>> + goto err_alloc_id;
>> }
>> kfd_unref_process(process);
>> @@ -1610,6 +1680,8 @@ static struct kfd_process *create_process(const
>> struct task_struct *thread, bool
>> return process;
>> +err_alloc_id:
>> + kfd_process_free_id(process);
>> err_register_notifier:
>> hash_del_rcu(&process->kfd_processes);
>> svm_range_list_fini(process);
[-- Attachment #2: Type: text/html, Size: 12403 bytes --]
^ permalink raw reply [flat|nested] 27+ messages in thread
* [PATCH V5 09/18] amdkfd: find kfd_process by filep->private_data in kfd_mmap
2025-10-17 8:42 [PATCH V5 00/18] [PATCH V4 00/18] amdkfd: Implement kfd multiple contexts Zhu Lingshan
` (7 preceding siblings ...)
2025-10-17 8:42 ` [PATCH V5 08/18] amdkfd: identify a secondary kfd process by its id Zhu Lingshan
@ 2025-10-17 8:42 ` Zhu Lingshan
2025-10-17 8:42 ` [PATCH V5 10/18] amdkfd: remove DIQ support Zhu Lingshan
` (9 subsequent siblings)
18 siblings, 0 replies; 27+ messages in thread
From: Zhu Lingshan @ 2025-10-17 8:42 UTC (permalink / raw)
To: felix.kuehling, alexander.deucher; +Cc: ray.huang, amd-gfx, Zhu Lingshan
This commit finds the proper kfd_process by
filep->private_data in kfd_mmap,
because the function kfd_get_process()
can not locate a specific kfd process among
multiple contexts.
Signed-off-by: Zhu Lingshan <lingshan.zhu@amd.com>
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
---
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 11 +++++++----
1 file changed, 7 insertions(+), 4 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
index d1cb22f48b00..73de2de8be0f 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
@@ -3410,16 +3410,19 @@ static int kfd_mmio_mmap(struct kfd_node *dev, struct kfd_process *process,
}
-static int kfd_mmap(struct file *filp, struct vm_area_struct *vma)
+static int kfd_mmap(struct file *filep, struct vm_area_struct *vma)
{
struct kfd_process *process;
struct kfd_node *dev = NULL;
unsigned long mmap_offset;
unsigned int gpu_id;
- process = kfd_get_process(current);
- if (IS_ERR(process))
- return PTR_ERR(process);
+ process = filep->private_data;
+ if (!process)
+ return -ESRCH;
+
+ if (process->lead_thread != current->group_leader)
+ return -EBADF;
mmap_offset = vma->vm_pgoff << PAGE_SHIFT;
gpu_id = KFD_MMAP_GET_GPU_ID(mmap_offset);
--
2.51.0
^ permalink raw reply related [flat|nested] 27+ messages in thread* [PATCH V5 10/18] amdkfd: remove DIQ support
2025-10-17 8:42 [PATCH V5 00/18] [PATCH V4 00/18] amdkfd: Implement kfd multiple contexts Zhu Lingshan
` (8 preceding siblings ...)
2025-10-17 8:42 ` [PATCH V5 09/18] amdkfd: find kfd_process by filep->private_data in kfd_mmap Zhu Lingshan
@ 2025-10-17 8:42 ` Zhu Lingshan
2025-10-17 8:42 ` [PATCH V5 11/18] amdkfd: process pointer of a HIQ should be NULL Zhu Lingshan
` (8 subsequent siblings)
18 siblings, 0 replies; 27+ messages in thread
From: Zhu Lingshan @ 2025-10-17 8:42 UTC (permalink / raw)
To: felix.kuehling, alexander.deucher; +Cc: ray.huang, amd-gfx, Zhu Lingshan
This commit remove DIQ support because it has been
marked as DEPRECATED since 2022 in commit 5bdd3eb2
Signed-off-by: Zhu Lingshan <lingshan.zhu@amd.com>
---
.../drm/amd/amdkfd/kfd_device_queue_manager.c | 6 +--
drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c | 29 +-------------
.../drm/amd/amdkfd/kfd_packet_manager_v9.c | 4 --
.../drm/amd/amdkfd/kfd_packet_manager_vi.c | 4 --
drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 1 -
.../amd/amdkfd/kfd_process_queue_manager.c | 39 +------------------
6 files changed, 5 insertions(+), 78 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 6c5c7c1bf5ed..ac6052d7a4cb 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
@@ -399,8 +399,7 @@ static void increment_queue_count(struct device_queue_manager *dqm,
struct queue *q)
{
dqm->active_queue_count++;
- if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE ||
- q->properties.type == KFD_QUEUE_TYPE_DIQ)
+ if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE)
dqm->active_cp_queue_count++;
if (q->properties.is_gws) {
@@ -414,8 +413,7 @@ static void decrement_queue_count(struct device_queue_manager *dqm,
struct queue *q)
{
dqm->active_queue_count--;
- if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE ||
- q->properties.type == KFD_QUEUE_TYPE_DIQ)
+ if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE)
dqm->active_cp_queue_count--;
if (q->properties.is_gws) {
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c b/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c
index fb3129883a4c..7ce4531ec3bc 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c
@@ -57,19 +57,7 @@ static bool kq_initialize(struct kernel_queue *kq, struct kfd_node *dev,
kq->dev = dev;
kq->nop_packet = nop.u32all;
- switch (type) {
- case KFD_QUEUE_TYPE_DIQ:
- kq->mqd_mgr = dev->dqm->mqd_mgrs[KFD_MQD_TYPE_DIQ];
- break;
- case KFD_QUEUE_TYPE_HIQ:
- kq->mqd_mgr = dev->dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ];
- break;
- default:
- WARN(1, "Invalid queue type %d\n", type);
- dev_err(dev->adev->dev, "Invalid queue type %d\n", type);
- return false;
- }
-
+ kq->mqd_mgr = dev->dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ];
if (!kq->mqd_mgr)
return false;
@@ -159,24 +147,11 @@ static bool kq_initialize(struct kernel_queue *kq, struct kfd_node *dev,
kq->mqd_mgr->load_mqd(kq->mqd_mgr, kq->queue->mqd,
kq->queue->pipe, kq->queue->queue,
&kq->queue->properties, NULL);
- } else {
- /* allocate fence for DIQ */
-
- retval = kfd_gtt_sa_allocate(dev, sizeof(uint32_t),
- &kq->fence_mem_obj);
-
- if (retval != 0)
- goto err_alloc_fence;
-
- kq->fence_kernel_address = kq->fence_mem_obj->cpu_ptr;
- kq->fence_gpu_addr = kq->fence_mem_obj->gpu_addr;
}
print_queue(kq->queue);
return true;
-err_alloc_fence:
- kq->mqd_mgr->free_mqd(kq->mqd_mgr, kq->queue->mqd, kq->queue->mqd_mem_obj);
err_allocate_mqd:
uninit_queue(kq->queue);
err_init_queue:
@@ -206,8 +181,6 @@ static void kq_uninitialize(struct kernel_queue *kq)
kq->queue->queue);
up_read(&kq->dev->adev->reset_domain->sem);
}
- else if (kq->queue->properties.type == KFD_QUEUE_TYPE_DIQ)
- kfd_gtt_sa_free(kq->dev, kq->fence_mem_obj);
kq->mqd_mgr->free_mqd(kq->mqd_mgr, kq->queue->mqd,
kq->queue->mqd_mem_obj);
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c
index 505036968a77..3d2375817c3e 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c
@@ -252,10 +252,6 @@ static int pm_map_queues_v9(struct packet_manager *pm, uint32_t *buffer,
packet->bitfields2.queue_type =
queue_type__mes_map_queues__normal_latency_static_queue_vi;
break;
- case KFD_QUEUE_TYPE_DIQ:
- packet->bitfields2.queue_type =
- queue_type__mes_map_queues__debug_interface_queue_vi;
- break;
case KFD_QUEUE_TYPE_SDMA:
case KFD_QUEUE_TYPE_SDMA_XGMI:
if (q->properties.sdma_engine_id < 2 &&
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_vi.c b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_vi.c
index a1de5d7e173a..60086e7cc258 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_vi.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_vi.c
@@ -166,10 +166,6 @@ static int pm_map_queues_vi(struct packet_manager *pm, uint32_t *buffer,
packet->bitfields2.queue_type =
queue_type__mes_map_queues__normal_latency_static_queue_vi;
break;
- case KFD_QUEUE_TYPE_DIQ:
- packet->bitfields2.queue_type =
- queue_type__mes_map_queues__debug_interface_queue_vi;
- break;
case KFD_QUEUE_TYPE_SDMA:
case KFD_QUEUE_TYPE_SDMA_XGMI:
packet->bitfields2.engine_sel = q->properties.sdma_engine_id +
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
index 9de658119cd9..71c0cefb87ea 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
@@ -434,7 +434,6 @@ enum kfd_queue_type {
KFD_QUEUE_TYPE_COMPUTE,
KFD_QUEUE_TYPE_SDMA,
KFD_QUEUE_TYPE_HIQ,
- KFD_QUEUE_TYPE_DIQ,
KFD_QUEUE_TYPE_SDMA_XGMI,
KFD_QUEUE_TYPE_SDMA_BY_ENG_ID
};
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c
index 7fbb5c274ccc..8ddc33abf230 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c
@@ -345,7 +345,7 @@ int pqm_create_queue(struct process_queue_manager *pqm,
* If we are just about to create DIQ, the is_debug flag is not set yet
* Hence we also check the type as well
*/
- if ((pdd->qpd.is_debug) || (type == KFD_QUEUE_TYPE_DIQ))
+ if ((pdd->qpd.is_debug))
max_queues = dev->kfd->device_info.max_no_of_hqd/2;
if (pdd->qpd.queue_count >= max_queues)
@@ -426,22 +426,6 @@ int pqm_create_queue(struct process_queue_manager *pqm,
restore_mqd, restore_ctl_stack);
print_queue(q);
break;
- case KFD_QUEUE_TYPE_DIQ:
- kq = kernel_queue_init(dev, KFD_QUEUE_TYPE_DIQ);
- if (!kq) {
- retval = -ENOMEM;
- goto err_create_queue;
- }
- kq->queue->properties.queue_id = *qid;
- pqn->kq = kq;
- pqn->q = NULL;
- retval = kfd_process_drain_interrupts(pdd);
- if (retval)
- break;
-
- retval = dev->dqm->ops.create_kernel_queue(dev->dqm,
- kq, &pdd->qpd);
- break;
default:
WARN(1, "Invalid queue type %d", type);
retval = -EINVAL;
@@ -1131,32 +1115,13 @@ int pqm_debugfs_mqds(struct seq_file *m, void *data)
break;
default:
seq_printf(m,
- " Bad user queue type %d on device %x\n",
+ " Qeueu node with bad user queue type %d on device %x\n",
q->properties.type, q->device->id);
continue;
}
mqd_mgr = q->device->dqm->mqd_mgrs[mqd_type];
size = mqd_mgr->mqd_stride(mqd_mgr,
&q->properties);
- } else if (pqn->kq) {
- q = pqn->kq->queue;
- mqd_mgr = pqn->kq->mqd_mgr;
- switch (q->properties.type) {
- case KFD_QUEUE_TYPE_DIQ:
- seq_printf(m, " DIQ on device %x\n",
- pqn->kq->dev->id);
- break;
- default:
- seq_printf(m,
- " Bad kernel queue type %d on device %x\n",
- q->properties.type,
- pqn->kq->dev->id);
- continue;
- }
- } else {
- seq_printf(m,
- " Weird: Queue node with neither kernel nor user queue\n");
- continue;
}
for (xcc = 0; xcc < num_xccs; xcc++) {
--
2.51.0
^ permalink raw reply related [flat|nested] 27+ messages in thread* [PATCH V5 11/18] amdkfd: process pointer of a HIQ should be NULL
2025-10-17 8:42 [PATCH V5 00/18] [PATCH V4 00/18] amdkfd: Implement kfd multiple contexts Zhu Lingshan
` (9 preceding siblings ...)
2025-10-17 8:42 ` [PATCH V5 10/18] amdkfd: remove DIQ support Zhu Lingshan
@ 2025-10-17 8:42 ` Zhu Lingshan
2025-10-17 8:42 ` [PATCH V5 12/18] amdkfd: remove test_kq Zhu Lingshan
` (7 subsequent siblings)
18 siblings, 0 replies; 27+ messages in thread
From: Zhu Lingshan @ 2025-10-17 8:42 UTC (permalink / raw)
To: felix.kuehling, alexander.deucher; +Cc: ray.huang, amd-gfx, Zhu Lingshan
In kq_initialize, queue->process of a HIQ should
be NULL as initialized, because it does not belong
to any kfd_process.
This commit decommisions the function kfd_get_process() because
it can not locate a specific kfd_process among multiple
contexts and not any code path calls it after this commit.
Signed-off-by: Zhu Lingshan <lingshan.zhu@amd.com>
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
---
drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c | 1 -
drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 1 -
drivers/gpu/drm/amd/amdkfd/kfd_process.c | 18 ------------------
3 files changed, 20 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c b/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c
index 7ce4531ec3bc..19ea759568a6 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c
@@ -129,7 +129,6 @@ static bool kq_initialize(struct kernel_queue *kq, struct kfd_node *dev,
goto err_init_queue;
kq->queue->device = dev;
- kq->queue->process = kfd_get_process(current);
kq->queue->mqd_mem_obj = kq->mqd_mgr->allocate_mqd(kq->mqd_mgr->dev,
&kq->queue->properties);
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
index 71c0cefb87ea..42b6492d7682 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
@@ -1060,7 +1060,6 @@ void kfd_process_destroy_wq(void);
void kfd_cleanup_processes(void);
struct kfd_process *kfd_create_process(struct task_struct *thread);
int kfd_create_process_sysfs(struct kfd_process *process);
-struct kfd_process *kfd_get_process(const struct task_struct *task);
struct kfd_process *kfd_lookup_process_by_pasid(u32 pasid,
struct kfd_process_device **pdd);
struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm);
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
index 157145c94314..35a9e91650ca 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
@@ -985,24 +985,6 @@ struct kfd_process *kfd_create_process(struct task_struct *thread)
return process;
}
-struct kfd_process *kfd_get_process(const struct task_struct *thread)
-{
- struct kfd_process *process;
-
- if (!thread->mm)
- return ERR_PTR(-EINVAL);
-
- /* Only the pthreads threading model is supported. */
- if (thread->group_leader->mm != thread->mm)
- return ERR_PTR(-EINVAL);
-
- process = find_process(thread, false);
- if (!process)
- return ERR_PTR(-EINVAL);
-
- return process;
-}
-
static struct kfd_process *find_process_by_mm(const struct mm_struct *mm)
{
struct kfd_process *process;
--
2.51.0
^ permalink raw reply related [flat|nested] 27+ messages in thread* [PATCH V5 12/18] amdkfd: remove test_kq
2025-10-17 8:42 [PATCH V5 00/18] [PATCH V4 00/18] amdkfd: Implement kfd multiple contexts Zhu Lingshan
` (10 preceding siblings ...)
2025-10-17 8:42 ` [PATCH V5 11/18] amdkfd: process pointer of a HIQ should be NULL Zhu Lingshan
@ 2025-10-17 8:42 ` Zhu Lingshan
2025-10-17 8:42 ` [PATCH V5 13/18] amdkfd: introduce new helper kfd_lookup_process_by_id Zhu Lingshan
` (6 subsequent siblings)
18 siblings, 0 replies; 27+ messages in thread
From: Zhu Lingshan @ 2025-10-17 8:42 UTC (permalink / raw)
To: felix.kuehling, alexander.deucher; +Cc: ray.huang, amd-gfx, Zhu Lingshan
This commit removes test_kq() function becuse it has been
marked as unused since 2014 and no other functions calls it.
Signed-off-by: Zhu Lingshan <lingshan.zhu@amd.com>
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
---
drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c | 31 -------------------
1 file changed, 31 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c b/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c
index 19ea759568a6..baeabe3feef8 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c
@@ -327,34 +327,3 @@ void kernel_queue_uninit(struct kernel_queue *kq)
kq_uninitialize(kq);
kfree(kq);
}
-
-/* FIXME: Can this test be removed? */
-static __attribute__((unused)) void test_kq(struct kfd_node *dev)
-{
- struct kernel_queue *kq;
- uint32_t *buffer, i;
- int retval;
-
- dev_err(dev->adev->dev, "Starting kernel queue test\n");
-
- kq = kernel_queue_init(dev, KFD_QUEUE_TYPE_HIQ);
- if (unlikely(!kq)) {
- dev_err(dev->adev->dev, " Failed to initialize HIQ\n");
- dev_err(dev->adev->dev, "Kernel queue test failed\n");
- return;
- }
-
- retval = kq_acquire_packet_buffer(kq, 5, &buffer);
- if (unlikely(retval != 0)) {
- dev_err(dev->adev->dev, " Failed to acquire packet buffer\n");
- dev_err(dev->adev->dev, "Kernel queue test failed\n");
- return;
- }
- for (i = 0; i < 5; i++)
- buffer[i] = kq->nop_packet;
- kq_submit_packet(kq);
-
- dev_err(dev->adev->dev, "Ending kernel queue test\n");
-}
-
-
--
2.51.0
^ permalink raw reply related [flat|nested] 27+ messages in thread* [PATCH V5 13/18] amdkfd: introduce new helper kfd_lookup_process_by_id
2025-10-17 8:42 [PATCH V5 00/18] [PATCH V4 00/18] amdkfd: Implement kfd multiple contexts Zhu Lingshan
` (11 preceding siblings ...)
2025-10-17 8:42 ` [PATCH V5 12/18] amdkfd: remove test_kq Zhu Lingshan
@ 2025-10-17 8:42 ` Zhu Lingshan
2025-10-17 8:42 ` [PATCH V5 14/18] amdkfd: record kfd context id into kfd process_info Zhu Lingshan
` (5 subsequent siblings)
18 siblings, 0 replies; 27+ messages in thread
From: Zhu Lingshan @ 2025-10-17 8:42 UTC (permalink / raw)
To: felix.kuehling, alexander.deucher; +Cc: ray.huang, amd-gfx, Zhu Lingshan
This commit introduces a new helper function
kfd_lookup_process_by_id which can find a
kfd process that identified by its context id from
the kfd process table
Signed-off-by: Zhu Lingshan <lingshan.zhu@amd.com>
---
drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 1 +
drivers/gpu/drm/amd/amdkfd/kfd_process.c | 21 +++++++++++++++++++++
2 files changed, 22 insertions(+)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
index 42b6492d7682..4237c859050d 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
@@ -1063,6 +1063,7 @@ int kfd_create_process_sysfs(struct kfd_process *process);
struct kfd_process *kfd_lookup_process_by_pasid(u32 pasid,
struct kfd_process_device **pdd);
struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm);
+struct kfd_process *kfd_lookup_process_by_id(const struct mm_struct *mm, u16 id);
int kfd_process_gpuidx_from_gpuid(struct kfd_process *p, uint32_t gpu_id);
int kfd_process_gpuid_from_node(struct kfd_process *p, struct kfd_node *node,
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
index 35a9e91650ca..629a706e2a13 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
@@ -1955,6 +1955,27 @@ struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm)
return p;
}
+/* This increments the process->ref counter. */
+struct kfd_process *kfd_lookup_process_by_id(const struct mm_struct *mm, u16 id)
+{
+ struct kfd_process *p, *ret_p = NULL;
+ unsigned int temp;
+
+ int idx = srcu_read_lock(&kfd_processes_srcu);
+
+ hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
+ if (p->mm == mm && p->context_id == id) {
+ kref_get(&p->ref);
+ ret_p = p;
+ break;
+ }
+ }
+
+ srcu_read_unlock(&kfd_processes_srcu, idx);
+
+ return ret_p;
+}
+
/* kfd_process_evict_queues - Evict all user queues of a process
*
* Eviction is reference-counted per process-device. This means multiple
--
2.51.0
^ permalink raw reply related [flat|nested] 27+ messages in thread* [PATCH V5 14/18] amdkfd: record kfd context id into kfd process_info
2025-10-17 8:42 [PATCH V5 00/18] [PATCH V4 00/18] amdkfd: Implement kfd multiple contexts Zhu Lingshan
` (12 preceding siblings ...)
2025-10-17 8:42 ` [PATCH V5 13/18] amdkfd: introduce new helper kfd_lookup_process_by_id Zhu Lingshan
@ 2025-10-17 8:42 ` Zhu Lingshan
2025-10-17 8:42 ` [PATCH V5 15/18] amdkfd: record kfd context id in amdkfd_fence Zhu Lingshan
` (4 subsequent siblings)
18 siblings, 0 replies; 27+ messages in thread
From: Zhu Lingshan @ 2025-10-17 8:42 UTC (permalink / raw)
To: felix.kuehling, alexander.deucher; +Cc: ray.huang, amd-gfx, Zhu Lingshan
This commit records the context id of the owner
kfd_process into a kfd process_info when
create it.
Signed-off-by: Zhu Lingshan <lingshan.zhu@amd.com>
---
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 2 ++
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 4 ++++
2 files changed, 6 insertions(+)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
index 9e120c934cc1..28b54d7ee1f5 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
@@ -148,6 +148,8 @@ struct amdkfd_process_info {
/* MMU-notifier related fields */
struct mutex notifier_lock;
uint32_t evicted_bos;
+ /* kfd context id */
+ u16 context_id;
struct delayed_work restore_userptr_work;
struct pid *pid;
bool block_mmu_notifications;
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
index 83020963dfde..0e40cdd8c5ec 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
@@ -1386,8 +1386,10 @@ static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info,
struct dma_fence **ef)
{
struct amdkfd_process_info *info = NULL;
+ struct kfd_process *process = NULL;
int ret;
+ process = container_of(process_info, struct kfd_process, kgd_process_info);
if (!*process_info) {
info = kzalloc(sizeof(*info), GFP_KERNEL);
if (!info)
@@ -1414,6 +1416,8 @@ static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info,
INIT_DELAYED_WORK(&info->restore_userptr_work,
amdgpu_amdkfd_restore_userptr_worker);
+ info->context_id = process->context_id;
+
*process_info = info;
}
--
2.51.0
^ permalink raw reply related [flat|nested] 27+ messages in thread* [PATCH V5 15/18] amdkfd: record kfd context id in amdkfd_fence
2025-10-17 8:42 [PATCH V5 00/18] [PATCH V4 00/18] amdkfd: Implement kfd multiple contexts Zhu Lingshan
` (13 preceding siblings ...)
2025-10-17 8:42 ` [PATCH V5 14/18] amdkfd: record kfd context id into kfd process_info Zhu Lingshan
@ 2025-10-17 8:42 ` Zhu Lingshan
2025-10-17 8:42 ` [PATCH V5 16/18] amdkfd: fence handler evict and restore a kfd process by its context id Zhu Lingshan
` (3 subsequent siblings)
18 siblings, 0 replies; 27+ messages in thread
From: Zhu Lingshan @ 2025-10-17 8:42 UTC (permalink / raw)
To: felix.kuehling, alexander.deucher; +Cc: ray.huang, amd-gfx, Zhu Lingshan
This commit records the context id of the
cooresponding kfd process in amdkfd_fence
Signed-off-by: Zhu Lingshan <lingshan.zhu@amd.com>
---
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 4 +++-
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_fence.c | 4 +++-
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 4 ++--
drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 2 +-
4 files changed, 9 insertions(+), 5 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
index 28b54d7ee1f5..087e8fe2c077 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
@@ -98,6 +98,7 @@ struct amdgpu_amdkfd_fence {
spinlock_t lock;
char timeline_name[TASK_COMM_LEN];
struct svm_range_bo *svm_bo;
+ uint16_t context_id;
};
struct amdgpu_kfd_dev {
@@ -190,7 +191,8 @@ int amdgpu_queue_mask_bit_to_set_resource_bit(struct amdgpu_device *adev,
struct amdgpu_amdkfd_fence *amdgpu_amdkfd_fence_create(u64 context,
struct mm_struct *mm,
- struct svm_range_bo *svm_bo);
+ struct svm_range_bo *svm_bo,
+ u16 context_id);
int amdgpu_amdkfd_drm_client_create(struct amdgpu_device *adev);
#if defined(CONFIG_DEBUG_FS)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_fence.c
index 1ef758ac5076..4119d0a9235e 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_fence.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_fence.c
@@ -62,7 +62,8 @@ static atomic_t fence_seq = ATOMIC_INIT(0);
struct amdgpu_amdkfd_fence *amdgpu_amdkfd_fence_create(u64 context,
struct mm_struct *mm,
- struct svm_range_bo *svm_bo)
+ struct svm_range_bo *svm_bo,
+ u16 context_id)
{
struct amdgpu_amdkfd_fence *fence;
@@ -76,6 +77,7 @@ struct amdgpu_amdkfd_fence *amdgpu_amdkfd_fence_create(u64 context,
get_task_comm(fence->timeline_name, current);
spin_lock_init(&fence->lock);
fence->svm_bo = svm_bo;
+ fence->context_id = context_id;
dma_fence_init(&fence->base, &amdkfd_fence_ops, &fence->lock,
context, atomic_inc_return(&fence_seq));
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
index 0e40cdd8c5ec..58d85149fecf 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
@@ -1405,7 +1405,7 @@ static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info,
info->eviction_fence =
amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1),
current->mm,
- NULL);
+ NULL, process->context_id);
if (!info->eviction_fence) {
pr_err("Failed to create eviction fence\n");
ret = -ENOMEM;
@@ -3057,7 +3057,7 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence __rcu *
amdgpu_amdkfd_fence_create(
process_info->eviction_fence->base.context,
process_info->eviction_fence->mm,
- NULL);
+ NULL, process_info->context_id);
if (!new_fence) {
pr_err("Failed to create eviction fence\n");
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
index 9d72411c3379..04582aef1b41 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
@@ -585,7 +585,7 @@ svm_range_vram_node_new(struct kfd_node *node, struct svm_range *prange,
svm_bo->eviction_fence =
amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1),
mm,
- svm_bo);
+ svm_bo, p->context_id);
mmput(mm);
INIT_WORK(&svm_bo->eviction_work, svm_range_evict_svm_bo_worker);
svm_bo->evicting = 0;
--
2.51.0
^ permalink raw reply related [flat|nested] 27+ messages in thread* [PATCH V5 16/18] amdkfd: fence handler evict and restore a kfd process by its context id
2025-10-17 8:42 [PATCH V5 00/18] [PATCH V4 00/18] amdkfd: Implement kfd multiple contexts Zhu Lingshan
` (14 preceding siblings ...)
2025-10-17 8:42 ` [PATCH V5 15/18] amdkfd: record kfd context id in amdkfd_fence Zhu Lingshan
@ 2025-10-17 8:42 ` Zhu Lingshan
2025-10-17 8:42 ` [PATCH V5 17/18] amdkfd: process debug trap ioctl only on a primary context Zhu Lingshan
` (2 subsequent siblings)
18 siblings, 0 replies; 27+ messages in thread
From: Zhu Lingshan @ 2025-10-17 8:42 UTC (permalink / raw)
To: felix.kuehling, alexander.deucher; +Cc: ray.huang, amd-gfx, Zhu Lingshan
In fence enable signaling handler, kfd evicts
and restores the corresponding kfd_process,
this commit helps find the kfd_process by
both its mm and context id.
Signed-off-by: Zhu Lingshan <lingshan.zhu@amd.com>
---
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 2 +-
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_fence.c | 6 +++++-
drivers/gpu/drm/amd/amdkfd/kfd_device.c | 7 ++++---
3 files changed, 10 insertions(+), 5 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
index 087e8fe2c077..321cbf9a1528 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
@@ -411,7 +411,7 @@ int kgd2kfd_init_zone_device(struct amdgpu_device *adev)
int kgd2kfd_quiesce_mm(struct mm_struct *mm, uint32_t trigger);
int kgd2kfd_resume_mm(struct mm_struct *mm);
int kgd2kfd_schedule_evict_and_restore_process(struct mm_struct *mm,
- struct dma_fence *fence);
+ u16 context_id, struct dma_fence *fence);
#if IS_ENABLED(CONFIG_HSA_AMD)
int kgd2kfd_init(void);
void kgd2kfd_exit(void);
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_fence.c
index 4119d0a9235e..193ed8becab8 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_fence.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_fence.c
@@ -128,8 +128,12 @@ static bool amdkfd_fence_enable_signaling(struct dma_fence *f)
if (dma_fence_is_signaled(f))
return true;
+ /* if fence->svm_bo is NULL, means this fence is created through
+ * init_kfd_vm() or amdgpu_amdkfd_gpuvm_restore_process_bos().
+ * Therefore, this fence is amdgpu_amdkfd_fence->eviction_fence.
+ */
if (!fence->svm_bo) {
- if (!kgd2kfd_schedule_evict_and_restore_process(fence->mm, f))
+ if (!kgd2kfd_schedule_evict_and_restore_process(fence->mm, fence->context_id, f))
return true;
} else {
if (!svm_range_schedule_evict_svm_bo(fence))
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
index e9cfb80bd436..0b9a3c198c14 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
@@ -1202,12 +1202,13 @@ int kgd2kfd_resume_mm(struct mm_struct *mm)
* prepare for safe eviction of KFD BOs that belong to the specified
* process.
*
- * @mm: mm_struct that identifies the specified KFD process
+ * @mm: mm_struct that identifies a group of KFD processes
+ * @context_id: an id that identifies a specific KFD context in the above kfd process group
* @fence: eviction fence attached to KFD process BOs
*
*/
int kgd2kfd_schedule_evict_and_restore_process(struct mm_struct *mm,
- struct dma_fence *fence)
+ u16 context_id, struct dma_fence *fence)
{
struct kfd_process *p;
unsigned long active_time;
@@ -1219,7 +1220,7 @@ int kgd2kfd_schedule_evict_and_restore_process(struct mm_struct *mm,
if (dma_fence_is_signaled(fence))
return 0;
- p = kfd_lookup_process_by_mm(mm);
+ p = kfd_lookup_process_by_id(mm, context_id);
if (!p)
return -ENODEV;
--
2.51.0
^ permalink raw reply related [flat|nested] 27+ messages in thread* [PATCH V5 17/18] amdkfd: process debug trap ioctl only on a primary context
2025-10-17 8:42 [PATCH V5 00/18] [PATCH V4 00/18] amdkfd: Implement kfd multiple contexts Zhu Lingshan
` (15 preceding siblings ...)
2025-10-17 8:42 ` [PATCH V5 16/18] amdkfd: fence handler evict and restore a kfd process by its context id Zhu Lingshan
@ 2025-10-17 8:42 ` Zhu Lingshan
2025-10-17 23:32 ` Felix Kuehling
2025-10-17 8:42 ` [PATCH V5 18/18] amdkfd: introduce new ioctl AMDKFD_IOC_CREATE_PROCESS Zhu Lingshan
2025-10-17 23:58 ` [PATCH V5 00/18] [PATCH V4 00/18] amdkfd: Implement kfd multiple contexts Felix Kuehling
18 siblings, 1 reply; 27+ messages in thread
From: Zhu Lingshan @ 2025-10-17 8:42 UTC (permalink / raw)
To: felix.kuehling, alexander.deucher; +Cc: ray.huang, amd-gfx, Zhu Lingshan
Set_debug_trap ioctl should work on a specific kfd_process
even when multiple contexts feature is implemented.
For consistency, this commit allow set_debug_trap ioctl only work on the
primary kfd process of a user space program
Signed-off-by: Zhu Lingshan <lingshan.zhu@amd.com>
---
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 6 ++++++
1 file changed, 6 insertions(+)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
index 73de2de8be0f..7c02e8473622 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
@@ -2964,6 +2964,12 @@ static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, v
target = kfd_lookup_process_by_pid(pid);
}
+ if (target->context_id != KFD_CONTEXT_ID_PRIMARY) {
+ pr_debug("Set debug trap ioctl not supported on non-primary kfd process\n");
+ r = -EOPNOTSUPP;
+ goto out;
+ }
+
if (IS_ERR_OR_NULL(target)) {
pr_debug("Cannot find process PID %i to debug\n", args->pid);
r = target ? PTR_ERR(target) : -ESRCH;
--
2.51.0
^ permalink raw reply related [flat|nested] 27+ messages in thread* Re: [PATCH V5 17/18] amdkfd: process debug trap ioctl only on a primary context
2025-10-17 8:42 ` [PATCH V5 17/18] amdkfd: process debug trap ioctl only on a primary context Zhu Lingshan
@ 2025-10-17 23:32 ` Felix Kuehling
2025-10-22 6:55 ` Zhu, Lingshan
0 siblings, 1 reply; 27+ messages in thread
From: Felix Kuehling @ 2025-10-17 23:32 UTC (permalink / raw)
To: Zhu Lingshan, alexander.deucher; +Cc: ray.huang, amd-gfx
On 2025-10-17 04:42, Zhu Lingshan wrote:
> Set_debug_trap ioctl should work on a specific kfd_process
> even when multiple contexts feature is implemented.
>
> For consistency, this commit allow set_debug_trap ioctl only work on the
> primary kfd process of a user space program
>
> Signed-off-by: Zhu Lingshan <lingshan.zhu@amd.com>
> ---
> drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 6 ++++++
> 1 file changed, 6 insertions(+)
>
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
> index 73de2de8be0f..7c02e8473622 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
> @@ -2964,6 +2964,12 @@ static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, v
> target = kfd_lookup_process_by_pid(pid);
> }
>
> + if (target->context_id != KFD_CONTEXT_ID_PRIMARY) {
> + pr_debug("Set debug trap ioctl not supported on non-primary kfd process\n");
> + r = -EOPNOTSUPP;
> + goto out;
> + }
> +
This check should be after the IS_ERR_OR_NULL(target) check below.
Otherwise you dereference a bogus pointer.
We should also check that the process calling the ioctl is a primary
context (p->context_id == KFD_CONTEXT_ID_PRIMARY). You don't want to
allow a guest process (using a secondary context) to debug another
process running on the host, using any ptrace the privileges of the QEMU
host process.
Regards,
Felix
> if (IS_ERR_OR_NULL(target)) {
> pr_debug("Cannot find process PID %i to debug\n", args->pid);
> r = target ? PTR_ERR(target) : -ESRCH;
^ permalink raw reply [flat|nested] 27+ messages in thread* Re: [PATCH V5 17/18] amdkfd: process debug trap ioctl only on a primary context
2025-10-17 23:32 ` Felix Kuehling
@ 2025-10-22 6:55 ` Zhu, Lingshan
0 siblings, 0 replies; 27+ messages in thread
From: Zhu, Lingshan @ 2025-10-22 6:55 UTC (permalink / raw)
To: Felix Kuehling, alexander.deucher; +Cc: ray.huang, amd-gfx
[-- Attachment #1: Type: text/plain, Size: 1836 bytes --]
On 10/18/2025 7:32 AM, Felix Kuehling wrote:
> On 2025-10-17 04:42, Zhu Lingshan wrote:
>> Set_debug_trap ioctl should work on a specific kfd_process
>> even when multiple contexts feature is implemented.
>>
>> For consistency, this commit allow set_debug_trap ioctl only work on the
>> primary kfd process of a user space program
>>
>> Signed-off-by: Zhu Lingshan <lingshan.zhu@amd.com>
>> ---
>> drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 6 ++++++
>> 1 file changed, 6 insertions(+)
>>
>> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
>> b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
>> index 73de2de8be0f..7c02e8473622 100644
>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
>> @@ -2964,6 +2964,12 @@ static int kfd_ioctl_set_debug_trap(struct
>> file *filep, struct kfd_process *p, v
>> target = kfd_lookup_process_by_pid(pid);
>> }
>> + if (target->context_id != KFD_CONTEXT_ID_PRIMARY) {
>> + pr_debug("Set debug trap ioctl not supported on non-primary
>> kfd process\n");
>> + r = -EOPNOTSUPP;
>> + goto out;
>> + }
>> +
>
> This check should be after the IS_ERR_OR_NULL(target) check below.
> Otherwise you dereference a bogus pointer.
>
> We should also check that the process calling the ioctl is a primary
> context (p->context_id == KFD_CONTEXT_ID_PRIMARY). You don't want to
> allow a guest process (using a secondary context) to debug another
> process running on the host, using any ptrace the privileges of the
> QEMU host process.
will fix, thanks!
>
> Regards,
> Felix
>
>
>> if (IS_ERR_OR_NULL(target)) {
>> pr_debug("Cannot find process PID %i to debug\n", args->pid);
>> r = target ? PTR_ERR(target) : -ESRCH;
[-- Attachment #2: Type: text/html, Size: 3300 bytes --]
^ permalink raw reply [flat|nested] 27+ messages in thread
* [PATCH V5 18/18] amdkfd: introduce new ioctl AMDKFD_IOC_CREATE_PROCESS
2025-10-17 8:42 [PATCH V5 00/18] [PATCH V4 00/18] amdkfd: Implement kfd multiple contexts Zhu Lingshan
` (16 preceding siblings ...)
2025-10-17 8:42 ` [PATCH V5 17/18] amdkfd: process debug trap ioctl only on a primary context Zhu Lingshan
@ 2025-10-17 8:42 ` Zhu Lingshan
2025-10-17 23:55 ` Felix Kuehling
2025-10-17 23:58 ` [PATCH V5 00/18] [PATCH V4 00/18] amdkfd: Implement kfd multiple contexts Felix Kuehling
18 siblings, 1 reply; 27+ messages in thread
From: Zhu Lingshan @ 2025-10-17 8:42 UTC (permalink / raw)
To: felix.kuehling, alexander.deucher; +Cc: ray.huang, amd-gfx, Zhu Lingshan
This commit implemetns a new ioctl AMDKFD_IOC_CREATE_PROCESS
that creates a new secondary kfd_progress on the FD.
To keep backward compatibility, userspace programs need to invoke
this ioctl explicitly on a FD to create a secondary
kfd_process which replacing its primary kfd_process.
This commit bumps ioctl minor version.
Signed-off-by: Zhu Lingshan <lingshan.zhu@amd.com>
---
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 41 ++++++++++++++++++++++++
drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 1 +
drivers/gpu/drm/amd/amdkfd/kfd_process.c | 3 +-
include/uapi/linux/kfd_ioctl.h | 8 +++--
4 files changed, 49 insertions(+), 4 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
index 7c02e8473622..8fe58526bc3a 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
@@ -3140,6 +3140,44 @@ static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, v
return r;
}
+/* userspace programs need to invoke this ioctl explicitly on a FD to
+ * create a secondary kfd_process which replacing its primary kfd_process
+ */
+static int kfd_ioctl_create_process(struct file *filep, struct kfd_process *p, void *data)
+{
+ struct kfd_process *process;
+ int ret;
+
+ if (!filep->private_data || !p)
+ return -EINVAL;
+
+ if (p != filep->private_data)
+ return -EINVAL;
+
+ /* Each FD owns only one kfd_process */
+ if (p->context_id != KFD_CONTEXT_ID_PRIMARY)
+ return -EINVAL;
+
+ mutex_lock(&kfd_processes_mutex);
+ process = create_process(current, false);
+ mutex_unlock(&kfd_processes_mutex);
+
+ if (IS_ERR(process))
+ return PTR_ERR(process);
+
+ /* Each open() increases kref of the primary kfd_process,
+ * so we need to reduce it here before we create a new secondary process replacing it
+ */
+ kfd_unref_process(p);
+
+ filep->private_data = process;
+ ret = kfd_create_process_sysfs(process);
+ if (ret)
+ pr_warn("Failed to create sysfs entry for the kfd_process");
+
+ return 0;
+}
+
#define AMDKFD_IOCTL_DEF(ioctl, _func, _flags) \
[_IOC_NR(ioctl)] = {.cmd = ioctl, .func = _func, .flags = _flags, \
.cmd_drv = 0, .name = #ioctl}
@@ -3258,6 +3296,9 @@ static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = {
AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_TRAP,
kfd_ioctl_set_debug_trap, 0),
+
+ AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_PROCESS,
+ kfd_ioctl_create_process, 0),
};
#define AMDKFD_CORE_IOCTL_COUNT ARRAY_SIZE(amdkfd_ioctls)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
index 4237c859050d..578f9f217e16 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
@@ -1055,6 +1055,7 @@ struct amdkfd_ioctl_desc {
};
bool kfd_dev_is_large_bar(struct kfd_node *dev);
+struct kfd_process *create_process(const struct task_struct *thread, bool primary);
int kfd_process_create_wq(void);
void kfd_process_destroy_wq(void);
void kfd_cleanup_processes(void);
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
index 629a706e2a13..bbb72a77bed5 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
@@ -68,7 +68,6 @@ static struct workqueue_struct *kfd_restore_wq;
static struct kfd_process *find_process(const struct task_struct *thread,
bool ref);
static void kfd_process_ref_release(struct kref *ref);
-static struct kfd_process *create_process(const struct task_struct *thread, bool primary);
static void evict_process_worker(struct work_struct *work);
static void restore_process_worker(struct work_struct *work);
@@ -1578,7 +1577,7 @@ void kfd_process_set_trap_debug_flag(struct qcm_process_device *qpd,
* On return the kfd_process is fully operational and will be freed when the
* mm is released
*/
-static struct kfd_process *create_process(const struct task_struct *thread, bool primary)
+struct kfd_process *create_process(const struct task_struct *thread, bool primary)
{
struct kfd_process *process;
struct mmu_notifier *mn;
diff --git a/include/uapi/linux/kfd_ioctl.h b/include/uapi/linux/kfd_ioctl.h
index 5d1727a6d040..84aa24c02715 100644
--- a/include/uapi/linux/kfd_ioctl.h
+++ b/include/uapi/linux/kfd_ioctl.h
@@ -44,9 +44,10 @@
* - 1.16 - Add contiguous VRAM allocation flag
* - 1.17 - Add SDMA queue creation with target SDMA engine ID
* - 1.18 - Rename pad in set_memory_policy_args to misc_process_flag
+ * - 1.19 - Add a new ioctl to craete secondary kfd processes
*/
#define KFD_IOCTL_MAJOR_VERSION 1
-#define KFD_IOCTL_MINOR_VERSION 18
+#define KFD_IOCTL_MINOR_VERSION 19
struct kfd_ioctl_get_version_args {
__u32 major_version; /* from KFD */
@@ -1671,7 +1672,10 @@ struct kfd_ioctl_dbg_trap_args {
#define AMDKFD_IOC_DBG_TRAP \
AMDKFD_IOWR(0x26, struct kfd_ioctl_dbg_trap_args)
+#define AMDKFD_IOC_CREATE_PROCESS \
+ AMDKFD_IO(0x27)
+
#define AMDKFD_COMMAND_START 0x01
-#define AMDKFD_COMMAND_END 0x27
+#define AMDKFD_COMMAND_END 0x28
#endif
--
2.51.0
^ permalink raw reply related [flat|nested] 27+ messages in thread* Re: [PATCH V5 18/18] amdkfd: introduce new ioctl AMDKFD_IOC_CREATE_PROCESS
2025-10-17 8:42 ` [PATCH V5 18/18] amdkfd: introduce new ioctl AMDKFD_IOC_CREATE_PROCESS Zhu Lingshan
@ 2025-10-17 23:55 ` Felix Kuehling
2025-10-22 6:59 ` Zhu, Lingshan
0 siblings, 1 reply; 27+ messages in thread
From: Felix Kuehling @ 2025-10-17 23:55 UTC (permalink / raw)
To: Zhu Lingshan, alexander.deucher; +Cc: ray.huang, amd-gfx
[-- Attachment #1: Type: text/plain, Size: 6407 bytes --]
On 2025-10-17 04:42, Zhu Lingshan wrote:
> This commit implemetns a new ioctl AMDKFD_IOC_CREATE_PROCESS
> that creates a new secondary kfd_progress on the FD.
>
> To keep backward compatibility, userspace programs need to invoke
> this ioctl explicitly on a FD to create a secondary
> kfd_process which replacing its primary kfd_process.
>
> This commit bumps ioctl minor version.
>
> Signed-off-by: Zhu Lingshan<lingshan.zhu@amd.com>
> ---
> drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 41 ++++++++++++++++++++++++
> drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 1 +
> drivers/gpu/drm/amd/amdkfd/kfd_process.c | 3 +-
> include/uapi/linux/kfd_ioctl.h | 8 +++--
> 4 files changed, 49 insertions(+), 4 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
> index 7c02e8473622..8fe58526bc3a 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
> @@ -3140,6 +3140,44 @@ static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, v
> return r;
> }
>
> +/* userspace programs need to invoke this ioctl explicitly on a FD to
> + * create a secondary kfd_process which replacing its primary kfd_process
> + */
> +static int kfd_ioctl_create_process(struct file *filep, struct kfd_process *p, void *data)
> +{
> + struct kfd_process *process;
> + int ret;
> +
> + if (!filep->private_data || !p)
> + return -EINVAL;
> +
> + if (p != filep->private_data)
> + return -EINVAL;
> +
> + /* Each FD owns only one kfd_process */
> + if (p->context_id != KFD_CONTEXT_ID_PRIMARY)
> + return -EINVAL;
> +
> + mutex_lock(&kfd_processes_mutex);
> + process = create_process(current, false);
> + mutex_unlock(&kfd_processes_mutex);
There could be race conditions between multiple threads calling this
ioctl on the same primary fd concurrently. We could use the
kfd_processes_mutex to protect against this. You'd need to do some of
the above checks and the assignment to filep->private_data under the
lock to be safe, to allow only one thread to successfully create the
secondary context. I think this would work:
/* Atomically create the secondary context and assign it to filep->private_data */
mutex_lock(&kfd_processes_mutex);
if (p != filep->private_data) {
/* Another thread already created a secondary context */
mutex_unlock(&kfd_processes_mutex);
return -EINVAL;
}
secondary = create_process(current, false);
if (!IS_ERR(process))
filep->private_data = secondary;
mutex_unlock(&kfd_processes_mutex);
if (IS_ERR(secondary))
return PTR_ERR(secondary);
/* Release the process reference that was held by filep->private_data */
kfd_unref_process(p);
...
Regards,
Felix
> +
> + if (IS_ERR(process))
> + return PTR_ERR(process);
> +
> + /* Each open() increases kref of the primary kfd_process,
> + * so we need to reduce it here before we create a new secondary process replacing it
> + */
> + kfd_unref_process(p);
> +
> + filep->private_data = process;
> + ret = kfd_create_process_sysfs(process);
> + if (ret)
> + pr_warn("Failed to create sysfs entry for the kfd_process");
> +
> + return 0;
> +}
> +
> #define AMDKFD_IOCTL_DEF(ioctl, _func, _flags) \
> [_IOC_NR(ioctl)] = {.cmd = ioctl, .func = _func, .flags = _flags, \
> .cmd_drv = 0, .name = #ioctl}
> @@ -3258,6 +3296,9 @@ static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = {
>
> AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_TRAP,
> kfd_ioctl_set_debug_trap, 0),
> +
> + AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_PROCESS,
> + kfd_ioctl_create_process, 0),
> };
>
> #define AMDKFD_CORE_IOCTL_COUNT ARRAY_SIZE(amdkfd_ioctls)
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
> index 4237c859050d..578f9f217e16 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
> @@ -1055,6 +1055,7 @@ struct amdkfd_ioctl_desc {
> };
> bool kfd_dev_is_large_bar(struct kfd_node *dev);
>
> +struct kfd_process *create_process(const struct task_struct *thread, bool primary);
> int kfd_process_create_wq(void);
> void kfd_process_destroy_wq(void);
> void kfd_cleanup_processes(void);
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
> index 629a706e2a13..bbb72a77bed5 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
> @@ -68,7 +68,6 @@ static struct workqueue_struct *kfd_restore_wq;
> static struct kfd_process *find_process(const struct task_struct *thread,
> bool ref);
> static void kfd_process_ref_release(struct kref *ref);
> -static struct kfd_process *create_process(const struct task_struct *thread, bool primary);
>
> static void evict_process_worker(struct work_struct *work);
> static void restore_process_worker(struct work_struct *work);
> @@ -1578,7 +1577,7 @@ void kfd_process_set_trap_debug_flag(struct qcm_process_device *qpd,
> * On return the kfd_process is fully operational and will be freed when the
> * mm is released
> */
> -static struct kfd_process *create_process(const struct task_struct *thread, bool primary)
> +struct kfd_process *create_process(const struct task_struct *thread, bool primary)
> {
> struct kfd_process *process;
> struct mmu_notifier *mn;
> diff --git a/include/uapi/linux/kfd_ioctl.h b/include/uapi/linux/kfd_ioctl.h
> index 5d1727a6d040..84aa24c02715 100644
> --- a/include/uapi/linux/kfd_ioctl.h
> +++ b/include/uapi/linux/kfd_ioctl.h
> @@ -44,9 +44,10 @@
> * - 1.16 - Add contiguous VRAM allocation flag
> * - 1.17 - Add SDMA queue creation with target SDMA engine ID
> * - 1.18 - Rename pad in set_memory_policy_args to misc_process_flag
> + * - 1.19 - Add a new ioctl to craete secondary kfd processes
> */
> #define KFD_IOCTL_MAJOR_VERSION 1
> -#define KFD_IOCTL_MINOR_VERSION 18
> +#define KFD_IOCTL_MINOR_VERSION 19
>
> struct kfd_ioctl_get_version_args {
> __u32 major_version; /* from KFD */
> @@ -1671,7 +1672,10 @@ struct kfd_ioctl_dbg_trap_args {
> #define AMDKFD_IOC_DBG_TRAP \
> AMDKFD_IOWR(0x26, struct kfd_ioctl_dbg_trap_args)
>
> +#define AMDKFD_IOC_CREATE_PROCESS \
> + AMDKFD_IO(0x27)
> +
> #define AMDKFD_COMMAND_START 0x01
> -#define AMDKFD_COMMAND_END 0x27
> +#define AMDKFD_COMMAND_END 0x28
>
> #endif
[-- Attachment #2: Type: text/html, Size: 6813 bytes --]
^ permalink raw reply [flat|nested] 27+ messages in thread* Re: [PATCH V5 18/18] amdkfd: introduce new ioctl AMDKFD_IOC_CREATE_PROCESS
2025-10-17 23:55 ` Felix Kuehling
@ 2025-10-22 6:59 ` Zhu, Lingshan
0 siblings, 0 replies; 27+ messages in thread
From: Zhu, Lingshan @ 2025-10-22 6:59 UTC (permalink / raw)
To: Felix Kuehling, alexander.deucher; +Cc: ray.huang, amd-gfx
[-- Attachment #1: Type: text/plain, Size: 7340 bytes --]
On 10/18/2025 7:55 AM, Felix Kuehling wrote:
> On 2025-10-17 04:42, Zhu Lingshan wrote:
>> This commit implemetns a new ioctl AMDKFD_IOC_CREATE_PROCESS
>> that creates a new secondary kfd_progress on the FD.
>>
>> To keep backward compatibility, userspace programs need to invoke
>> this ioctl explicitly on a FD to create a secondary
>> kfd_process which replacing its primary kfd_process.
>>
>> This commit bumps ioctl minor version.
>>
>> Signed-off-by: Zhu Lingshan<lingshan.zhu@amd.com>
>> ---
>> drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 41 ++++++++++++++++++++++++
>> drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 1 +
>> drivers/gpu/drm/amd/amdkfd/kfd_process.c | 3 +-
>> include/uapi/linux/kfd_ioctl.h | 8 +++--
>> 4 files changed, 49 insertions(+), 4 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
>> b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
>> index 7c02e8473622..8fe58526bc3a 100644
>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
>> @@ -3140,6 +3140,44 @@ static int kfd_ioctl_set_debug_trap(struct
>> file *filep, struct kfd_process *p, v
>> return r;
>> }
>> +/* userspace programs need to invoke this ioctl explicitly on a FD to
>> + * create a secondary kfd_process which replacing its primary
>> kfd_process
>> + */
>> +static int kfd_ioctl_create_process(struct file *filep, struct
>> kfd_process *p, void *data)
>> +{
>> + struct kfd_process *process;
>> + int ret;
>> +
>> + if (!filep->private_data || !p)
>> + return -EINVAL;
>> +
>> + if (p != filep->private_data)
>> + return -EINVAL;
>> +
>> + /* Each FD owns only one kfd_process */
>> + if (p->context_id != KFD_CONTEXT_ID_PRIMARY)
>> + return -EINVAL;
>> +
>> + mutex_lock(&kfd_processes_mutex);
>> + process = create_process(current, false);
>> + mutex_unlock(&kfd_processes_mutex);
>
> There could be race conditions between multiple threads calling this
> ioctl on the same primary fd concurrently. We could use the
> kfd_processes_mutex to protect against this. You'd need to do some of
> the above checks and the assignment to filep->private_data under the
> lock to be safe, to allow only one thread to successfully create the
> secondary context. I think this would work:
>
> /* Atomically create the secondary context and assign it to
> filep->private_data */
> mutex_lock(&kfd_processes_mutex);
> if (p != filep->private_data) {
> /* Another thread already created a secondary context */
> mutex_unlock(&kfd_processes_mutex);
> return -EINVAL;
> }
> secondary = create_process(current, false);
> if (!IS_ERR(process))
> filep->private_data = secondary;
> mutex_unlock(&kfd_processes_mutex);
>
> if (IS_ERR(secondary))
> return PTR_ERR(secondary);
>
> /* Release the process reference that was held by
> filep->private_data */
> kfd_unref_process(p);
>
> ...
>
will fix in V6, thanks!
> Regards,
> Felix
>
>
>> +
>> + if (IS_ERR(process))
>> + return PTR_ERR(process);
>> +
>> + /* Each open() increases kref of the primary kfd_process,
>> + * so we need to reduce it here before we create a new secondary
>> process replacing it
>> + */
>> + kfd_unref_process(p);
>> +
>> + filep->private_data = process;
>> + ret = kfd_create_process_sysfs(process);
>> + if (ret)
>> + pr_warn("Failed to create sysfs entry for the kfd_process");
>> +
>> + return 0;
>> +}
>> +
>> #define AMDKFD_IOCTL_DEF(ioctl, _func, _flags) \
>> [_IOC_NR(ioctl)] = {.cmd = ioctl, .func = _func, .flags =
>> _flags, \
>> .cmd_drv = 0, .name = #ioctl}
>> @@ -3258,6 +3296,9 @@ static const struct amdkfd_ioctl_desc
>> amdkfd_ioctls[] = {
>> AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_TRAP,
>> kfd_ioctl_set_debug_trap, 0),
>> +
>> + AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_PROCESS,
>> + kfd_ioctl_create_process, 0),
>> };
>> #define AMDKFD_CORE_IOCTL_COUNT ARRAY_SIZE(amdkfd_ioctls)
>> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
>> b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
>> index 4237c859050d..578f9f217e16 100644
>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
>> @@ -1055,6 +1055,7 @@ struct amdkfd_ioctl_desc {
>> };
>> bool kfd_dev_is_large_bar(struct kfd_node *dev);
>> +struct kfd_process *create_process(const struct task_struct
>> *thread, bool primary);
>> int kfd_process_create_wq(void);
>> void kfd_process_destroy_wq(void);
>> void kfd_cleanup_processes(void);
>> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c
>> b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
>> index 629a706e2a13..bbb72a77bed5 100644
>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c
>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
>> @@ -68,7 +68,6 @@ static struct workqueue_struct *kfd_restore_wq;
>> static struct kfd_process *find_process(const struct task_struct
>> *thread,
>> bool ref);
>> static void kfd_process_ref_release(struct kref *ref);
>> -static struct kfd_process *create_process(const struct task_struct
>> *thread, bool primary);
>> static void evict_process_worker(struct work_struct *work);
>> static void restore_process_worker(struct work_struct *work);
>> @@ -1578,7 +1577,7 @@ void kfd_process_set_trap_debug_flag(struct
>> qcm_process_device *qpd,
>> * On return the kfd_process is fully operational and will be freed
>> when the
>> * mm is released
>> */
>> -static struct kfd_process *create_process(const struct task_struct
>> *thread, bool primary)
>> +struct kfd_process *create_process(const struct task_struct *thread,
>> bool primary)
>> {
>> struct kfd_process *process;
>> struct mmu_notifier *mn;
>> diff --git a/include/uapi/linux/kfd_ioctl.h
>> b/include/uapi/linux/kfd_ioctl.h
>> index 5d1727a6d040..84aa24c02715 100644
>> --- a/include/uapi/linux/kfd_ioctl.h
>> +++ b/include/uapi/linux/kfd_ioctl.h
>> @@ -44,9 +44,10 @@
>> * - 1.16 - Add contiguous VRAM allocation flag
>> * - 1.17 - Add SDMA queue creation with target SDMA engine ID
>> * - 1.18 - Rename pad in set_memory_policy_args to misc_process_flag
>> + * - 1.19 - Add a new ioctl to craete secondary kfd processes
>> */
>> #define KFD_IOCTL_MAJOR_VERSION 1
>> -#define KFD_IOCTL_MINOR_VERSION 18
>> +#define KFD_IOCTL_MINOR_VERSION 19
>> struct kfd_ioctl_get_version_args {
>> __u32 major_version; /* from KFD */
>> @@ -1671,7 +1672,10 @@ struct kfd_ioctl_dbg_trap_args {
>> #define AMDKFD_IOC_DBG_TRAP \
>> AMDKFD_IOWR(0x26, struct kfd_ioctl_dbg_trap_args)
>> +#define AMDKFD_IOC_CREATE_PROCESS \
>> + AMDKFD_IO(0x27)
>> +
>> #define AMDKFD_COMMAND_START 0x01
>> -#define AMDKFD_COMMAND_END 0x27
>> +#define AMDKFD_COMMAND_END 0x28
>> #endif
[-- Attachment #2: Type: text/html, Size: 12307 bytes --]
^ permalink raw reply [flat|nested] 27+ messages in thread
* Re: [PATCH V5 00/18] [PATCH V4 00/18] amdkfd: Implement kfd multiple contexts
2025-10-17 8:42 [PATCH V5 00/18] [PATCH V4 00/18] amdkfd: Implement kfd multiple contexts Zhu Lingshan
` (17 preceding siblings ...)
2025-10-17 8:42 ` [PATCH V5 18/18] amdkfd: introduce new ioctl AMDKFD_IOC_CREATE_PROCESS Zhu Lingshan
@ 2025-10-17 23:58 ` Felix Kuehling
2025-10-22 7:01 ` Zhu, Lingshan
18 siblings, 1 reply; 27+ messages in thread
From: Felix Kuehling @ 2025-10-17 23:58 UTC (permalink / raw)
To: Zhu Lingshan, alexander.deucher; +Cc: ray.huang, amd-gfx
See my comments on patches 8, 17 and 18. The remaining patches look good
and are
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
I think this patch series is nearly ready, pending actual user mode code
that uses it.
Regards,
Felix
On 2025-10-17 04:42, Zhu Lingshan wrote:
> Currently kfd manages kfd_process in a one context (kfd_process)
> per program manner, thus each user space program
> only onws one kfd context (kfd_process).
>
> This model works fine for most of the programs, but imperfect
> for a hypervisor like QEMU. Because all programs in the guest
> user space share the same only one kfd context, which is
> problematic, including but not limited to:
>
> As illustrated in Figure 1, all guest user space programs share the same fd of /dev/kfd
> and the same kfd_process, and the same PASID leading to the same
> GPU_VM address space. Therefore the IOVA range of each
> guest user space programs are not isolated,
> they can attack each other through GPU DMA.
>
>
> +----------------------------------------------------------------------------------+
> | |
> | +-----------+ +-----------+ +------------+ +------------+ |
> | | | | | | | | | |
> | | Program 1 | | Program 2 | | Program 3 | | Program N | |
> | | | | | | | | | |
> | +----+------+ +--------+--+ +--+---------+ +-----+------+ |
> | | | | | |
> | | | | | Guest |
> | | | | | |
> +-------+----------------------+------------+----------------------+---------------+
> | | | |
> | | | |
> | | | |
> | | | |
> | +--+------------+---+ |
> | | file descriptor | |
> +-------------------+ of /dev/kfd +------------------+
> | opened by QEMU |
> | |
> +---------+---------+ User Space
> | QEMU
> |
> ---------------------------------------+-----------------------------------------------------
> | Kernel Space
> | KFD Module
> |
> +--------+--------+
> | |
> | kfd_process |<------------------The only one KFD context
> | |
> +--------+--------+
> |
> +--------+--------+
> | PASID |
> +--------+--------+
> |
> +--------+--------+
> | GPU_VM |
> +-----------------+
>
> Fiture 1
>
>
> This series implements a multiple contexts solution:
> - Allow each program to create and hold multiple contexts (kfd processes)
> - Each context has its own fd of /dev/kfd and an exclusive kfd_process,
> which is a secondary kfd context. So that PASID/GPU VM isolates their IOVA address spaces.
> Therefore, they can not attack each other through GPU DMA.
>
> The design is illustrated in Figure 2 below:
>
> +---------------------------------------------------------------------------------------------------------+
> | |
> | |
> | |
> | +----------------------------------------------------------------------------------+ |
> | | | |
> | | +-----------+ +-----------+ +-----------+ +-----------+ | |
> | | | | | | | | | | | |
> | | | Program 1 | | Program 2 | | Program 3 | | Program N | | |
> | | | | | | | | | | | |
> | | +-----+-----+ +-----+-----+ +-----+-----+ +-----+-----+ | |
> | | | | | | | |
> | | | | | | Guest | |
> | | | | | | | |
> | +-------+------------------+-----------------+----------------+--------------------+ |
> | | | | | QEMU |
> | | | | | |
> +---------------+------------------+-----------------+----------------+--------------------------+--------+
> | | | | |
> | | | | |
> | | | | |
> +---+----+ +---+----+ +---+----+ +---+----+ +---+-----+
> | | | | | | | | | Primary |
> | FD 1 | | FD 2 | | FD 3 | | FD 4 | | FD |
> | | | | | | | | | |
> +---+----+ +---+----+ +---+----+ +----+---+ +----+----+
> | | | | | User Space
> | | | | |
> -------------------+------------------+-----------------+-----------------+--------------------------+----------------------------
> | | | | | Kernel SPace
> | | | | |
> | | | | |
> +--------------------------------------------------------------------------------------------------------------------------+
> | +------+------+ +------+------+ +------+------+ +------+------+ +------+------+ |
> | | Secondary | | Secondary | | Secondary | | Secondary | | Primary | KFD Module |
> | |kfd_process 1| |kfd_process 2| |kfd_process 3| |kfd_process 4| | kfd_process | |
> | | | | | | | | | | | |
> | +------+------+ +------+------+ +------+------+ +------+------+ +------+------+ |
> | | | | | | |
> | +------+------+ +------+------+ +------+------+ +------+------+ +------+------+ |
> | | PASID | | PASID | | PASID | | PASID | | PASID | |
> | +------+------+ +------+------+ +------+------+ +------+------+ +------+------+ |
> | | | | | | |
> | | | | | | |
> | +------+------+ +------+------+ +------+------+ +------+------+ +------+------+ |
> | | GPU_VM | | GPU_VM | | GPU_VM | | GPU_VM | | GPU_VM | |
> | +-------------+ +-------------+ +-------------+ +-------------+ +-------------+ |
> | |
> +--------------------------------------------------------------------------------------------------------------------------+
>
> Figure 2
>
> The relevant reference user space rocm changes could be found at:
> https://github.com/AMD-ROCm-Internal/rocm-systems/pull/78
> https://github.com/AMD-ROCm-Internal/rocm-systems/pull/110
>
> Changes from V4:
> 1) rename process_id to context_id in struct kfd_process
> 2) remove primary flag in struct kfd_process
> 3) reject set_debug_trap ioctl request when
> the target kfd_process is non-primary
> 4) other than default 0, assign context_id 0xFFFF to the primary kfd process
>
> Most of the patches have been changed, so I removed
> their signed-off-by tag.
>
> Please help review
>
> Thanks!
>
> Zhu Lingshan (18):
> amdkfd: enlarge the hashtable of kfd_process
> amdkfd: mark the first kfd_process as the primary one
> amdkfd: find_process_by_mm always return the primary context
> amdkfd: Introduce kfd_create_process_sysfs as a separate function
> amdkfd: destroy kfd secondary contexts through fd close
> amdkfd: process svm ioctl only on the primary kfd process
> amdkfd: process USERPTR allocation only on the primary kfd process
> amdkfd: identify a secondary kfd process by its id
> amdkfd: find kfd_process by filep->private_data in kfd_mmap
> amdkfd: remove DIQ support
> amdkfd: process pointer of a HIQ should be NULL
> amdkfd: remove test_kq
> amdkfd: introduce new helper kfd_lookup_process_by_id
> amdkfd: record kfd context id into kfd process_info
> amdkfd: record kfd context id in amdkfd_fence
> amdkfd: fence handler evict and restore a kfd process by its context
> id
> amdkfd: process debug trap ioctl only on a primary context
> amdkfd: introduce new ioctl AMDKFD_IOC_CREATE_PROCESS
>
> drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 8 +-
> .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_fence.c | 10 +-
> .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 8 +-
> drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 79 +++++-
> drivers/gpu/drm/amd/amdkfd/kfd_device.c | 7 +-
> .../drm/amd/amdkfd/kfd_device_queue_manager.c | 6 +-
> drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c | 61 +----
> .../drm/amd/amdkfd/kfd_packet_manager_v9.c | 4 -
> .../drm/amd/amdkfd/kfd_packet_manager_vi.c | 4 -
> drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 19 +-
> drivers/gpu/drm/amd/amdkfd/kfd_process.c | 239 +++++++++++++-----
> .../amd/amdkfd/kfd_process_queue_manager.c | 39 +--
> drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 2 +-
> include/uapi/linux/kfd_ioctl.h | 8 +-
> 14 files changed, 294 insertions(+), 200 deletions(-)
>
^ permalink raw reply [flat|nested] 27+ messages in thread* Re: [PATCH V5 00/18] [PATCH V4 00/18] amdkfd: Implement kfd multiple contexts
2025-10-17 23:58 ` [PATCH V5 00/18] [PATCH V4 00/18] amdkfd: Implement kfd multiple contexts Felix Kuehling
@ 2025-10-22 7:01 ` Zhu, Lingshan
0 siblings, 0 replies; 27+ messages in thread
From: Zhu, Lingshan @ 2025-10-22 7:01 UTC (permalink / raw)
To: Felix Kuehling, alexander.deucher; +Cc: ray.huang, amd-gfx
[-- Attachment #1: Type: text/plain, Size: 19014 bytes --]
On 10/18/2025 7:58 AM, Felix Kuehling wrote:
> See my comments on patches 8, 17 and 18. The remaining patches look
> good and are
>
> Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
>
> I think this patch series is nearly ready, pending actual user mode
> code that uses it.
Thanks, do you think it is a must to merge user space thunk/kfdtest code first?
Thanks
Lingshan
>
> Regards,
> Felix
>
>
> On 2025-10-17 04:42, Zhu Lingshan wrote:
>> Currently kfd manages kfd_process in a one context (kfd_process)
>> per program manner, thus each user space program
>> only onws one kfd context (kfd_process).
>>
>> This model works fine for most of the programs, but imperfect
>> for a hypervisor like QEMU. Because all programs in the guest
>> user space share the same only one kfd context, which is
>> problematic, including but not limited to:
>>
>> As illustrated in Figure 1, all guest user space programs share the
>> same fd of /dev/kfd
>> and the same kfd_process, and the same PASID leading to the same
>> GPU_VM address space. Therefore the IOVA range of each
>> guest user space programs are not isolated,
>> they can attack each other through GPU DMA.
>>
>>
>>
>> +----------------------------------------------------------------------------------+
>>
>> |
>> |
>> | +-----------+ +-----------+ +------------+
>> +------------+ |
>> | | | | | | |
>> | | |
>> | | Program 1 | | Program 2 | | Program 3 | |
>> Program N | |
>> | | | | | | |
>> | | |
>> | +----+------+ +--------+--+ +--+---------+
>> +-----+------+ |
>> | | | |
>> | |
>> | | | |
>> | Guest |
>> | | | |
>> | |
>>
>> +-------+----------------------+------------+----------------------+---------------+
>> | | | |
>> | | | |
>> | | | |
>> | | | |
>> | +--+------------+---+ |
>> | | file descriptor | |
>> +-------------------+ of /dev/kfd +------------------+
>> | opened by QEMU |
>> | |
>> +---------+---------+
>> User Space
>> |
>> QEMU
>> |
>> ---------------------------------------+-----------------------------------------------------
>>
>> |
>> Kernel Space
>> |
>> KFD Module
>> |
>> +--------+--------+
>> | |
>> | kfd_process
>> |<------------------The only one KFD context
>> | |
>> +--------+--------+
>> |
>> +--------+--------+
>> | PASID |
>> +--------+--------+
>> |
>> +--------+--------+
>> | GPU_VM |
>> +-----------------+
>>
>> Fiture 1
>>
>>
>> This series implements a multiple contexts solution:
>> - Allow each program to create and hold multiple contexts (kfd
>> processes)
>> - Each context has its own fd of /dev/kfd and an exclusive kfd_process,
>> which is a secondary kfd context. So that PASID/GPU VM isolates
>> their IOVA address spaces.
>> Therefore, they can not attack each other through GPU DMA.
>>
>> The design is illustrated in Figure 2 below:
>>
>>
>> +---------------------------------------------------------------------------------------------------------+
>>
>> |
>> |
>>
>> |
>> |
>>
>> |
>> |
>> |
>> +----------------------------------------------------------------------------------+
>> |
>> |
>> |
>> | |
>> | | +-----------+ +-----------+ +-----------+
>> +-----------+ | |
>> | | | | | | | |
>> | | | |
>> | | | Program 1 | | Program 2 | | Program 3 | |
>> Program N | | |
>> | | | | | | | |
>> | | | |
>> | | +-----+-----+ +-----+-----+ +-----+-----+
>> +-----+-----+ | |
>> | | | |
>> | | | |
>> | | | |
>> | | Guest | |
>> | | | |
>> | | | |
>> |
>> +-------+------------------+-----------------+----------------+--------------------+
>> |
>> | | |
>> | | QEMU |
>> | | |
>> | | |
>>
>> +---------------+------------------+-----------------+----------------+--------------------------+--------+
>> | |
>> | | |
>> | |
>> | | |
>> | |
>> | | |
>> +---+----+ +---+----+ +---+----+
>> +---+----+ +---+-----+
>> | | | | | |
>> | | | Primary |
>> | FD 1 | | FD 2 | | FD 3 |
>> | FD 4 | | FD |
>> | | | | | |
>> | | | |
>> +---+----+ +---+----+ +---+----+
>> +----+---+ +----+----+
>> | |
>> | | | User Space
>> | |
>> | | |
>> -------------------+------------------+-----------------+-----------------+--------------------------+----------------------------
>>
>> | |
>> | | | Kernel SPace
>> | |
>> | | |
>> | |
>> | | |
>>
>> +--------------------------------------------------------------------------------------------------------------------------+
>> | +------+------+ +------+------+ +------+------+
>> +------+------+ +------+------+ |
>> | | Secondary | | Secondary | | Secondary | |
>> Secondary | | Primary | KFD Module |
>> | |kfd_process 1| |kfd_process 2| |kfd_process 3|
>> |kfd_process 4| | kfd_process | |
>> | | | | | | |
>> | | | | |
>> | +------+------+ +------+------+ +------+------+
>> +------+------+ +------+------+ |
>> | | |
>> | | | |
>> | +------+------+ +------+------+ +------+------+
>> +------+------+ +------+------+ |
>> | | PASID | | PASID | | PASID |
>> | PASID | | PASID | |
>> | +------+------+ +------+------+ +------+------+
>> +------+------+ +------+------+ |
>> | | |
>> | | | |
>> | | |
>> | | | |
>> | +------+------+ +------+------+ +------+------+
>> +------+------+ +------+------+ |
>> | | GPU_VM | | GPU_VM | | GPU_VM |
>> | GPU_VM | | GPU_VM | |
>> | +-------------+ +-------------+ +-------------+
>> +-------------+ +-------------+ |
>>
>> |
>> |
>>
>> +--------------------------------------------------------------------------------------------------------------------------+
>>
>> Figure 2
>>
>> The relevant reference user space rocm changes could be found at:
>> https://github.com/AMD-ROCm-Internal/rocm-systems/pull/78
>> https://github.com/AMD-ROCm-Internal/rocm-systems/pull/110
>>
>> Changes from V4:
>> 1) rename process_id to context_id in struct kfd_process
>> 2) remove primary flag in struct kfd_process
>> 3) reject set_debug_trap ioctl request when
>> the target kfd_process is non-primary
>> 4) other than default 0, assign context_id 0xFFFF to the primary kfd
>> process
>>
>> Most of the patches have been changed, so I removed
>> their signed-off-by tag.
>>
>> Please help review
>>
>> Thanks!
>>
>> Zhu Lingshan (18):
>> amdkfd: enlarge the hashtable of kfd_process
>> amdkfd: mark the first kfd_process as the primary one
>> amdkfd: find_process_by_mm always return the primary context
>> amdkfd: Introduce kfd_create_process_sysfs as a separate function
>> amdkfd: destroy kfd secondary contexts through fd close
>> amdkfd: process svm ioctl only on the primary kfd process
>> amdkfd: process USERPTR allocation only on the primary kfd process
>> amdkfd: identify a secondary kfd process by its id
>> amdkfd: find kfd_process by filep->private_data in kfd_mmap
>> amdkfd: remove DIQ support
>> amdkfd: process pointer of a HIQ should be NULL
>> amdkfd: remove test_kq
>> amdkfd: introduce new helper kfd_lookup_process_by_id
>> amdkfd: record kfd context id into kfd process_info
>> amdkfd: record kfd context id in amdkfd_fence
>> amdkfd: fence handler evict and restore a kfd process by its context
>> id
>> amdkfd: process debug trap ioctl only on a primary context
>> amdkfd: introduce new ioctl AMDKFD_IOC_CREATE_PROCESS
>>
>> drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 8 +-
>> .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_fence.c | 10 +-
>> .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 8 +-
>> drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 79 +++++-
>> drivers/gpu/drm/amd/amdkfd/kfd_device.c | 7 +-
>> .../drm/amd/amdkfd/kfd_device_queue_manager.c | 6 +-
>> drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c | 61 +----
>> .../drm/amd/amdkfd/kfd_packet_manager_v9.c | 4 -
>> .../drm/amd/amdkfd/kfd_packet_manager_vi.c | 4 -
>> drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 19 +-
>> drivers/gpu/drm/amd/amdkfd/kfd_process.c | 239 +++++++++++++-----
>> .../amd/amdkfd/kfd_process_queue_manager.c | 39 +--
>> drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 2 +-
>> include/uapi/linux/kfd_ioctl.h | 8 +-
>> 14 files changed, 294 insertions(+), 200 deletions(-)
>>
[-- Attachment #2: Type: text/html, Size: 45309 bytes --]
^ permalink raw reply [flat|nested] 27+ messages in thread