AMD General
> -----Original Message-----
> From: Liu, Shaoyun <[email protected]>
> Sent: Tuesday, November 17, 2026 12:31 AM
> To: Liang, Prike <[email protected]>; [email protected]
> Cc: Deucher, Alexander <[email protected]>; Koenig, Christian
> <[email protected]>; Chen, Michael <[email protected]>
> Subject: RE: [PATCH 2/6] drm/amdkfd: allocate the MES process context slot
> with
> its buffer
>
> AMD General
>
> You can add me as acked . The code changes looks ok to me , but I'm not sure
> whether the behavior will be changed since according to your description
> original
> driver always use slot 0 , now it has its own slot for each process . Not
> sure whether
> it's matter .
Previously, kfd_dbg_set_mes_debug_mode() did not allocate an MES context index
slot, defaulting to slot 0. Consequently, when the KFD debugger freed slot 0 on
teardown, it inadvertently released a slot that another process was actively
using.
Fix this by ensuring a valid MES context index slot is properly allocated for
the shader debugger.
Regards,
Prike
>
> Regards
> Shaoyun.liu
>
> -----Original Message-----
> From: Liang, Prike <[email protected]>
> Sent: Tuesday, November 10, 2026 8:05 AM
> To: [email protected]
> Cc: Deucher, Alexander <[email protected]>; Koenig, Christian
> <[email protected]>; Chen, Michael <[email protected]>; Liu,
> Shaoyun <[email protected]>; Liang, Prike <[email protected]>
> Subject: [PATCH 2/6] drm/amdkfd: allocate the MES process context slot with
> its
> buffer
>
> In MES RS64 mode, pqm_create_queue() allocates a process context slot right
> after
> the process context buffer, but only on the path where it allocates the
> buffer itself:
> - kfd_dbg_set_mes_debug_mode() allocates the buffer as well, without a
> slot. A process that a debugger attaches to before it creates its
> first queue therefore never gets a slot: its queues use slot 0, and
> it frees slot 0 when it exits, while another process may hold it.
> - If no slot is available, the buffer is kept without a slot, and the
> process's later queues use slot 0 the same way.
>
> Allocate the buffer and the slot together in one helper that both paths use,
> and free
> the buffer again if no slot is available.
>
> Fixes: e32b68c6b135 ("drm/amdkfd: enable rs64mem for kfd queue")
> Signed-off-by: Prike Liang <[email protected]>
> ---
> drivers/gpu/drm/amd/amdkfd/kfd_debug.c | 19 ++--------
> drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 1 +
> drivers/gpu/drm/amd/amdkfd/kfd_process.c | 35 +++++++++++++++++++
> .../amd/amdkfd/kfd_process_queue_manager.c | 26 ++------------
> 4 files changed, 42 insertions(+), 39 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c
> b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c
> index cb830f985536..d5c87de519ac 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c
> @@ -350,27 +350,14 @@ int kfd_dbg_set_mes_debug_mode(struct
> kfd_process_device *pdd, bool sq_trap_en) {
> uint32_t spi_dbg_cntl = pdd->spi_dbg_override |
> pdd->spi_dbg_launch_mode;
> uint32_t flags = pdd->process->dbg_flags;
> - struct amdgpu_device *adev = pdd->dev->adev;
> int r;
>
> if (!kfd_dbg_is_per_vmid_supported(pdd->dev))
> return 0;
>
> - if (!pdd->proc_ctx_cpu_ptr) {
> - r = amdgpu_amdkfd_alloc_kernel_mem(adev,
> - AMDGPU_MES_PROC_CTX_SIZE,
> - AMDGPU_GEM_DOMAIN_GTT,
> - &pdd->proc_ctx_bo,
> - &pdd->proc_ctx_gpu_addr,
> - &pdd->proc_ctx_cpu_ptr,
> - false);
> - if (r) {
> - dev_err(adev->dev,
> - "failed to allocate process context bo\n");
> - return r;
> - }
> - memset(pdd->proc_ctx_cpu_ptr, 0, AMDGPU_MES_PROC_CTX_SIZE);
> - }
> + r = kfd_process_device_alloc_proc_ctx(pdd);
> + if (r)
> + return r;
>
> return amdgpu_mes_set_shader_debugger(pdd->dev->adev,
> pdd->proc_ctx_gpu_addr, diff --git
> a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
> b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
> index 158fa70a1a8e..0a4e8639c4b5 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
> @@ -1140,6 +1140,7 @@ struct kfd_process_device
> *kfd_process_device_data_by_id(struct kfd_process *pro
>
> int kfd_process_get_user_gpu_id(struct kfd_process *p, uint32_t
> actual_gpu_id);
>
> +int kfd_process_device_alloc_proc_ctx(struct kfd_process_device *pdd);
> int kfd_process_device_init_vm(struct kfd_process_device *pdd,
> struct file *drm_file); struct
> kfd_process_device
> *kfd_bind_process_to_device(struct kfd_node *dev, diff --git
> a/drivers/gpu/drm/amd/amdkfd/kfd_process.c
> b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
> index 546721b4d59a..102fcb401646 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
> @@ -1204,6 +1204,41 @@ static void kfd_process_profiler_release(struct
> kfd_process *p, struct kfd_proce
> up_read(&adev->reset_domain->sem);
> }
>
> +/* Allocate the MES process context and, in RS64 mode, its slot
> +together. */ int kfd_process_device_alloc_proc_ctx(struct
> +kfd_process_device *pdd) {
> + struct amdgpu_device *adev = pdd->dev->adev;
> + int r;
> +
> + if (pdd->proc_ctx_cpu_ptr)
> + return 0;
> +
> + r = amdgpu_amdkfd_alloc_kernel_mem(adev,
> AMDGPU_MES_PROC_CTX_SIZE,
> + AMDGPU_GEM_DOMAIN_GTT,
> + &pdd->proc_ctx_bo,
> + &pdd->proc_ctx_gpu_addr,
> + &pdd->proc_ctx_cpu_ptr, false);
> + if (r) {
> + dev_err(adev->dev, "failed to allocate process context bo\n");
> + return r;
> + }
> + memset(pdd->proc_ctx_cpu_ptr, 0, AMDGPU_MES_PROC_CTX_SIZE);
> +
> + if (!adev->mes.use_rs64mem)
> + return 0;
> +
> + r = amdgpu_mes_alloc_proc_ctx_index(&adev->mes,
> + &pdd->proc_ctx_array_index);
> + if (r) {
> + dev_err(adev->dev, "failed to allocate process context
> index\n");
> + amdgpu_amdkfd_free_kernel_mem(adev, &pdd->proc_ctx_bo);
> + pdd->proc_ctx_gpu_addr = 0;
> + pdd->proc_ctx_cpu_ptr = NULL;
> + }
> +
> + return r;
> +}
> +
> static void kfd_process_destroy_pdds(struct kfd_process *p) {
> int i;
> 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 d3149e6d40e4..4646dfbabe75 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c
> @@ -387,30 +387,10 @@ int pqm_create_queue(struct process_queue_manager
> *pqm,
> dev->dqm->ops.register_process(dev->dqm, &pdd->qpd);
>
> /* Allocate proc_ctx_bo only if MES is enabled and this is the first
> queue */
> - if (!pdd->proc_ctx_cpu_ptr && dev->kfd->shared_resources.enable_mes) {
> - retval = amdgpu_amdkfd_alloc_kernel_mem(dev->adev,
> - AMDGPU_MES_PROC_CTX_SIZE,
> - AMDGPU_GEM_DOMAIN_GTT,
> - &pdd->proc_ctx_bo,
> - &pdd->proc_ctx_gpu_addr,
> - &pdd->proc_ctx_cpu_ptr,
> - false);
> - if (retval) {
> - dev_err(dev->adev->dev, "failed to allocate process
> context bo\n");
> + if (dev->kfd->shared_resources.enable_mes) {
> + retval = kfd_process_device_alloc_proc_ctx(pdd);
> + if (retval)
> goto err_allocate_pqn;
> - }
> - memset(pdd->proc_ctx_cpu_ptr, 0, AMDGPU_MES_PROC_CTX_SIZE);
> - /* Bind one MES process context slot to the whole process
> - * (per device); every queue of this process reuses it.
> - */
> - if (dev->adev->mes.use_rs64mem) {
> - retval =
> amdgpu_mes_alloc_proc_ctx_index(&dev->adev->mes,
> - &pdd->proc_ctx_array_index);
> - if (retval) {
> - dev_err(dev->adev->dev, "failed to allocate
> process context
> index\n");
> - goto err_allocate_pqn;
> - }
> - }
> }
>
> pqn = kzalloc_obj(*pqn);
> --
> 2.34.1
>