[PATCH] drm/xe: Add DRM_XE_LOW_LEVEL_TRACEPOINTS

Fri Jan 5 17:23:16 UTC 2024

Quoting Matthew Brost (2024-01-05 13:45:30-03:00)
>Trace events are uABI. The Xe trace events currently implemented are
>developer only and subject to change. Hide all Xe behind
>DRM_XE_LOW_LEVEL_TRACEPOINTS so these are free to change going forward.
>
>Cc: Thomas Hellström <thomas.hellstrom at linux.intel.com>
>Signed-off-by: Matthew Brost <matthew.brost at intel.com>
>---
> drivers/gpu/drm/xe/Kconfig.debug |  10 +
> drivers/gpu/drm/xe/Makefile      |   4 +-
> drivers/gpu/drm/xe/xe_trace.h    | 339 +++++++++++++++++++++++++++++++
> 3 files changed, 352 insertions(+), 1 deletion(-)
>
>diff --git a/drivers/gpu/drm/xe/Kconfig.debug b/drivers/gpu/drm/xe/Kconfig.debug
>index 549065f57a78..a8785e71e224 100644
>--- a/drivers/gpu/drm/xe/Kconfig.debug
>+++ b/drivers/gpu/drm/xe/Kconfig.debug
>@@ -105,3 +105,13 @@ config DRM_XE_USERPTR_INVAL_INJECT
> 
>          Recomended for driver developers only.
>          If in doubt, say "N".
>+
>+config DRM_XE_LOW_LEVEL_TRACEPOINTS
>+       bool "Enable low level trace events"
>+       default n
>+       help
>+         Choose this option to enable low level trace events.
>+
>+         Recomended for driver developers only.
>+
>+         If in doubt, say "N".
>diff --git a/drivers/gpu/drm/xe/Makefile b/drivers/gpu/drm/xe/Makefile
>index 53bd2a8ba1ae..9d33e6975ba0 100644
>--- a/drivers/gpu/drm/xe/Makefile
>+++ b/drivers/gpu/drm/xe/Makefile
>@@ -129,7 +129,6 @@ xe-y += xe_bb.o \
>         xe_sync.o \
>         xe_tile.o \
>         xe_tile_sysfs.o \
>-        xe_trace.o \
>         xe_ttm_sys_mgr.o \
>         xe_ttm_stolen_mgr.o \
>         xe_ttm_vram_mgr.o \
>@@ -142,6 +141,9 @@ xe-y += xe_bb.o \
>         xe_wa.o \
>         xe_wopcm.o
> 
>+# trace support
>+xe-$(CONFIG_DRM_XE_LOW_LEVEL_TRACEPOINTS) += xe_trace.o
>+
> # graphics hardware monitoring (HWMON) support
> xe-$(CONFIG_HWMON) += xe_hwmon.o
> 
>diff --git a/drivers/gpu/drm/xe/xe_trace.h b/drivers/gpu/drm/xe/xe_trace.h
>index 95163c303f3e..7163ac9b8eae 100644
>--- a/drivers/gpu/drm/xe/xe_trace.h
>+++ b/drivers/gpu/drm/xe/xe_trace.h
>@@ -21,6 +21,7 @@
> #include "xe_sched_job.h"
> #include "xe_vm.h"
> 
>+#if defined(CONFIG_DRM_XE_LOW_LEVEL_TRACEPOINTS)
> DECLARE_EVENT_CLASS(xe_gt_tlb_invalidation_fence,
>                     TP_PROTO(struct xe_gt_tlb_invalidation_fence *fence),
>                     TP_ARGS(fence),
>@@ -597,6 +598,344 @@ DEFINE_EVENT_PRINT(xe_guc_ctb, xe_guc_ctb_g2h,
>                              __entry->tail, __entry->_head)
> 
> );
>+#else
>+#if !defined(TRACE_HEADER_MULTI_READ)

I wonder if we could have XE_* variants of the macros used above so that
we do not need to declare an empty trace_* function for each event
added.

For example, here's is a foo.h file that I created to test the idea:

    #if !defined(CONFIG_DRM_XE_LOW_LEVEL_TRACEPOINTS)
    #define _XE_TP_PROTO(...) __VA_ARGS__
    #define XE_DEFINE_EVENT(__unused, __fn, __proto, ...) \
        static inline void trace_ ## __fn(_XE_ ## __proto) { }
    #else
    #define XE_DEFINE_EVENT(...) DEFINE_EVENT(__VA_ARGS__)
    #endif

    XE_DEFINE_EVENT(xe_gt_tlb_invalidation_fence, xe_gt_tlb_invalidation_fence_create,
             TP_PROTO(struct xe_gt_tlb_invalidation_fence *fence),
             TP_ARGS(fence)
    );

Output of `gcc -E foo.h`:

    # 0 "/tmp/foo.h"
    # 0 "<built-in>"
    # 0 "<command-line>"
    # 1 "/usr/include/stdc-predef.h" 1 3 4
    # 0 "<command-line>" 2
    # 1 "/tmp/foo.h"
    # 9 "/tmp/foo.h"
    static inline void trace_xe_gt_tlb_invalidation_fence_create(struct xe_gt_tlb_invalidation_fence *fence) { }

     ;

Output of `gcc -E -DCONFIG_DRM_XE_LOW_LEVEL_TRACEPOINTS foo.h`:
    # 0 "/tmp/foo.h"
    # 0 "<built-in>"
    # 0 "<command-line>"
    # 1 "/usr/include/stdc-predef.h" 1 3 4
    # 0 "<command-line>" 2
    # 1 "/tmp/foo.h"
    # 9 "/tmp/foo.h"
    DEFINE_EVENT(xe_gt_tlb_invalidation_fence, xe_gt_tlb_invalidation_fence_create, TP_PROTO(struct xe_gt_tlb_invalidation_fence *fence), TP_ARGS(fence))

     ;

Well, this is an isolated example, not sure it would work with the
kernel code. But I believe it is worth a try.

--
Gustavo Sousa

>+static inline void
>+trace_xe_gt_tlb_invalidation_fence_create(struct xe_gt_tlb_invalidation_fence *fence)
>+{
>+}
>+
>+static inline void
>+trace_xe_gt_tlb_invalidation_fence_work_func(struct xe_gt_tlb_invalidation_fence *fence)
>+{
>+}
>+
>+static inline void
>+trace_xe_gt_tlb_invalidation_fence_cb(struct xe_gt_tlb_invalidation_fence *fence)
>+{
>+}
>+
>+static inline void
>+trace_xe_gt_tlb_invalidation_fence_send(struct xe_gt_tlb_invalidation_fence *fence)
>+{
>+}
>+
>+static inline void
>+trace_xe_gt_tlb_invalidation_fence_recv(struct xe_gt_tlb_invalidation_fence *fence)
>+{
>+}
>+
>+static inline void
>+trace_xe_gt_tlb_invalidation_fence_signal(struct xe_gt_tlb_invalidation_fence *fence)
>+{
>+}
>+
>+static inline void
>+trace_xe_gt_tlb_invalidation_fence_timeout(struct xe_gt_tlb_invalidation_fence *fence)
>+{
>+}
>+
>+static inline void
>+trace_xe_bo_cpu_fault(struct xe_bo *bo)
>+{
>+}
>+
>+static inline void
>+trace_xe_bo_move(struct xe_bo *bo)
>+{
>+}
>+
>+static inline void
>+trace_xe_exec_queue_create(struct xe_exec_queue *q)
>+{
>+}
>+
>+static inline void
>+trace_xe_exec_queue_supress_resume(struct xe_exec_queue *q)
>+{
>+}
>+
>+static inline void
>+trace_xe_exec_queue_submit(struct xe_exec_queue *q)
>+{
>+}
>+
>+static inline void
>+trace_xe_exec_queue_scheduling_enable(struct xe_exec_queue *q)
>+{
>+}
>+
>+static inline void
>+trace_xe_exec_queue_scheduling_disable(struct xe_exec_queue *q)
>+{
>+}
>+
>+static inline void
>+trace_xe_exec_queue_scheduling_done(struct xe_exec_queue *q)
>+{
>+}
>+
>+static inline void
>+trace_xe_exec_queue_register(struct xe_exec_queue *q)
>+{
>+}
>+
>+static inline void
>+trace_xe_exec_queue_deregister(struct xe_exec_queue *q)
>+{
>+}
>+
>+static inline void
>+trace_xe_exec_queue_deregister_done(struct xe_exec_queue *q)
>+{
>+}
>+
>+static inline void
>+trace_xe_exec_queue_close(struct xe_exec_queue *q)
>+{
>+}
>+
>+static inline void
>+trace_xe_exec_queue_kill(struct xe_exec_queue *q)
>+{
>+}
>+
>+static inline void
>+trace_xe_exec_queue_cleanup_entity(struct xe_exec_queue *q)
>+{
>+}
>+
>+static inline void
>+trace_xe_exec_queue_destroy(struct xe_exec_queue *q)
>+{
>+}
>+
>+static inline void
>+trace_xe_exec_queue_reset(struct xe_exec_queue *q)
>+{
>+}
>+
>+static inline void
>+trace_xe_exec_queue_memory_cat_error(struct xe_exec_queue *q)
>+{
>+}
>+
>+static inline void
>+trace_xe_exec_queue_stop(struct xe_exec_queue *q)
>+{
>+}
>+
>+static inline void
>+trace_xe_exec_queue_resubmit(struct xe_exec_queue *q)
>+{
>+}
>+
>+static inline void
>+trace_xe_exec_queue_lr_cleanup(struct xe_exec_queue *q)
>+{
>+}
>+
>+static inline void
>+trace_xe_sched_job_create(struct xe_sched_job *job)
>+{
>+}
>+
>+static inline void
>+trace_xe_sched_job_exec(struct xe_sched_job *job)
>+{
>+}
>+
>+static inline void
>+trace_xe_sched_job_run(struct xe_sched_job *job)
>+{
>+}
>+
>+static inline void
>+trace_xe_sched_job_free(struct xe_sched_job *job)
>+{
>+}
>+
>+static inline void
>+trace_xe_sched_job_timedout(struct xe_sched_job *job)
>+{
>+}
>+
>+static inline void
>+trace_xe_sched_job_set_error(struct xe_sched_job *job)
>+{
>+}
>+
>+static inline void
>+trace_xe_sched_job_ban(struct xe_sched_job *job)
>+{
>+}
>+
>+static inline void
>+trace_xe_sched_msg_add(struct xe_sched_msg *msg)
>+{
>+}
>+
>+static inline void
>+trace_xe_sched_msg_recv(struct xe_sched_msg *msg)
>+{
>+}
>+
>+static inline void
>+trace_xe_hw_fence_create(struct xe_hw_fence *fence)
>+{
>+}
>+
>+static inline void
>+trace_xe_hw_fence_signal(struct xe_hw_fence *fence)
>+{
>+}
>+
>+static inline void
>+trace_xe_hw_fence_try_signal(struct xe_hw_fence *fence)
>+{
>+}
>+
>+static inline void
>+trace_xe_hw_fence_free(struct xe_hw_fence *fence)
>+{
>+}
>+
>+static inline void
>+trace_xe_vma_flush(struct xe_vma *vma)
>+{
>+}
>+
>+static inline void
>+trace_xe_vma_pagefault(struct xe_vma *vma)
>+{
>+}
>+
>+static inline void
>+trace_xe_vma_acc(struct xe_vma *vma)
>+{
>+}
>+
>+static inline void
>+trace_xe_vma_fail(struct xe_vma *vma)
>+{
>+}
>+
>+static inline void
>+trace_xe_vma_bind(struct xe_vma *vma)
>+{
>+}
>+
>+static inline void
>+trace_xe_vma_pf_bind(struct xe_vma *vma)
>+{
>+}
>+
>+static inline void
>+trace_xe_vma_unbind(struct xe_vma *vma)
>+{
>+}
>+
>+static inline void
>+trace_xe_vma_userptr_rebind_worker(struct xe_vma *vma)
>+{
>+}
>+
>+static inline void
>+trace_xe_vma_userptr_rebind_exec(struct xe_vma *vma)
>+{
>+}
>+
>+static inline void
>+trace_xe_vma_rebind_worker(struct xe_vma *vma)
>+{
>+}
>+
>+static inline void
>+trace_xe_vma_rebind_exec(struct xe_vma *vma)
>+{
>+}
>+
>+static inline void
>+trace_xe_vma_userptr_invalidate(struct xe_vma *vma)
>+{
>+}
>+
>+static inline void
>+trace_xe_vma_usm_invalidate(struct xe_vma *vma)
>+{
>+}
>+
>+static inline void
>+trace_xe_vma_evict(struct xe_vma *vma)
>+{
>+}
>+
>+static inline void
>+trace_xe_vma_userptr_invalidate_complete(struct xe_vma *vma)
>+{
>+}
>+
>+static inline void
>+trace_xe_vm_kill(struct xe_vm *vm)
>+{
>+}
>+
>+static inline void
>+trace_xe_vm_create(struct xe_vm *vm)
>+{
>+}
>+
>+static inline void
>+trace_xe_vm_free(struct xe_vm *vm)
>+{
>+}
>+
>+static inline void
>+trace_xe_vm_cpu_bind(struct xe_vm *vm)
>+{
>+}
>+
>+static inline void
>+trace_xe_vm_restart(struct xe_vm *vm)
>+{
>+}
>+
>+static inline void
>+trace_xe_vm_rebind_worker_enter(struct xe_vm *vm)
>+{
>+}
>+
>+static inline void
>+trace_xe_vm_rebind_worker_retry(struct xe_vm *vm)
>+{
>+}
>+
>+static inline void
>+trace_xe_vm_rebind_worker_exit(struct xe_vm *vm)
>+{
>+}
>+
>+static inline void
>+trace_xe_guc_ct_h2g_flow_control(u32 _head, u32 _tail, u32 size, u32 space, u32 len)
>+{
>+}
>+
>+static inline void
>+trace_xe_guc_ct_g2h_flow_control(u32 _head, u32 _tail, u32 size, u32 space, u32 len)
>+{
>+}
>+
>+static inline void
>+trace_xe_guc_ctb_h2g(u8 gt_id, u32 action, u32 len, u32 _head, u32 tail)
>+{
>+}
>+
>+static inline void
>+trace_xe_guc_ctb_g2h(u8 gt_id, u32 action, u32 len, u32 _head, u32 tail)
>+{
>+}
>+#endif
>+#endif
> 
> #endif
> 
>-- 
>2.34.1
>