The existing gpu_mem_total tracepoint [1] is not currently used by any in-tree consumers, we should add some. In addition, there's a desire to report imported memory via the counters too [2]. To do this, we'll have to redefine the event to: a) Change 'pid' to 'ctx_id' The reason is DRM subsystem is created with GEM objects, DRM devices and DRM files in mind. A GEM object is associated with DRM device, and it may be shared between one or more DRM files. Per-instance (or "context") counters make more sense than per-process counters for DRM. For GPUs that per process counters (kgsl), this change is backwards compatible. b) add an "import_mem_total" field We're just appending a field, so no problem here. Change "size" to "mem_total" as well (name changes are backwards compatible). [1] https://lore.kernel.org/r/20200302234840.57188-1-zzyiwei@xxxxxxxxxx/ [2] https://www.spinics.net/lists/kernel/msg4062769.html Signed-off-by: Gurchetan Singh <gurchetansingh@xxxxxxxxxxxx> --- include/trace/events/gpu_mem.h | 61 ++++++++++++++++++++++++---------- 1 file changed, 43 insertions(+), 18 deletions(-) diff --git a/include/trace/events/gpu_mem.h b/include/trace/events/gpu_mem.h index 26d871f96e94..198b87f50356 100644 --- a/include/trace/events/gpu_mem.h +++ b/include/trace/events/gpu_mem.h @@ -14,41 +14,66 @@ #include <linux/tracepoint.h> /* - * The gpu_memory_total event indicates that there's an update to either the - * global or process total gpu memory counters. + * The gpu_mem_total event indicates that there's an update to local or + * global gpu memory counters. * - * This event should be emitted whenever the kernel device driver allocates, - * frees, imports, unimports memory in the GPU addressable space. + * This event should be emitted whenever a GPU device (ctx_id == 0): * - * @gpu_id: This is the gpu id. + * 1) allocates memory. + * 2) frees memory. + * 3) imports memory from an external exporter. * - * @pid: Put 0 for global total, while positive pid for process total. + * OR when a GPU device instance (ctx_id != 0): * - * @size: Size of the allocation in bytes. + * 1) allocates or acquires a reference to memory from another instance. + * 2) frees or releases a reference to memory from another instance. + * 3) imports memory from another GPU device instance. * + * When ctx_id == 0, both mem_total and import_mem_total total counters + * represent a global total. When ctx_id == 0, these counters represent + * an instance specifical total. + * + * Note allocation does not necessarily mean backing the memory with pages. + * + * @gpu_id: unique ID of the GPU. + * + * @ctx_id: an ID for specific instance of the GPU device. + * + * @mem_total: - total size of memory known to a GPU device, including + * imports (ctx_id == 0) + * - total size of memory known to a GPU device instance + * (ctx_id != 0) + * + * @import_mem_total: - size of memory imported from outside GPU + * device (ctx_id == 0) + * - size of memory imported into GPU device instance. + * (ctx_id == 0) */ TRACE_EVENT(gpu_mem_total, - TP_PROTO(uint32_t gpu_id, uint32_t pid, uint64_t size), + TP_PROTO(u32 gpu_id, u32 ctx_id, u64 mem_total, u64 import_mem_total), - TP_ARGS(gpu_id, pid, size), + TP_ARGS(gpu_id, ctx_id, mem_total, import_mem_total), TP_STRUCT__entry( - __field(uint32_t, gpu_id) - __field(uint32_t, pid) - __field(uint64_t, size) + __field(u32, gpu_id) + __field(u32, ctx_id) + __field(u64, mem_total) + __field(u64, import_mem_total) ), TP_fast_assign( __entry->gpu_id = gpu_id; - __entry->pid = pid; - __entry->size = size; + __entry->ctx_id = ctx_id; + __entry->mem_total = mem_total; + __entry->import_mem_total = import_mem_total; ), - TP_printk("gpu_id=%u pid=%u size=%llu", - __entry->gpu_id, - __entry->pid, - __entry->size) + TP_printk("gpu_id=%u, ctx_id=%u, mem total=%llu, mem import total=%llu", + __entry->gpu_id, + __entry->ctx_id, + __entry->mem_total, + __entry->import_mem_total) ); #endif /* _TRACE_GPU_MEM_H */ -- 2.25.1
