[KUBE-50] gpuagent: add GPUGetFilter to skip GPU attribute groups on …

sw/nic/gpuagent/api/gpu.cc

@@ -150,19 +150,21 @@ gpu_entry::update_handler(api_params_base *api_params) {
 }
 
 void
-gpu_entry::fill_stats_(aga_gpu_stats_t *stats) {
+gpu_entry::fill_stats_(aga_gpu_stats_t *stats,
+                       const aga_gpu_get_filter_t *filter) {
     // fill stats only for non-parent GPUs
     if (child_gpus_.size()) {
         return;
     }
     // fetch stats from smi apis
     smi_gpu_fill_stats(handle_, &key_, is_partitioned_,
         (partition_id_ == AGA_GPU_INVALID_PARTITION_ID) ? 0 : partition_id_,
-        first_partition_handle_, stats);
+        first_partition_handle_, stats, filter);
 }
 
 void
-gpu_entry::fill_status_(aga_gpu_spec_t *spec, aga_gpu_status_t *status) {
+gpu_entry::fill_status_(aga_gpu_spec_t *spec, aga_gpu_status_t *status,
+                        const aga_gpu_get_filter_t *filter) {
     if (child_gpus_.size()) {
         status->num_gpu_partition = child_gpus_.size();
         // for parent GPUs get uuids of all children
@@ -175,7 +177,7 @@ gpu_entry::fill_status_(aga_gpu_spec_t *spec, aga_gpu_status_t *status) {
         if (parent_gpu_.valid()) {
             status->physical_gpu = parent_gpu_;
         }
-        smi_gpu_fill_status(handle_, &key_, id_, spec, status);
+        smi_gpu_fill_status(handle_, &key_, id_, spec, status, filter);
     }
 }
 
@@ -201,10 +203,10 @@ gpu_entry::fill_spec_(aga_gpu_spec_t *spec) {
 }
 
 sdk_ret_t
-gpu_entry::read(aga_gpu_info_t *info) {
+gpu_entry::read(aga_gpu_info_t *info, const aga_gpu_get_filter_t *filter) {
     fill_spec_(&info->spec);
-    fill_status_(&info->spec, &info->status);
-    fill_stats_(&info->stats);
+    fill_status_(&info->spec, &info->status, filter);
+    fill_stats_(&info->stats, filter);
     return SDK_RET_OK;
 }
 

sw/nic/gpuagent/api/gpu.hpp

@@ -99,9 +99,11 @@ class gpu_entry : public api_base {
     virtual sdk_ret_t delete_handler(api_params_base *api_params) override;
 
     /// \brief          read config
-    /// \param[out]     info pointer to the info object
+    /// \param[out]     info   pointer to the info object
+    /// \param[in]      filter attributes to skip (NULL fetches all attributes)
     /// \return         SDK_RET_OK on success, failure status code on error
-    sdk_ret_t read(aga_gpu_info_t *info);
+    sdk_ret_t read(aga_gpu_info_t *info,
+                   const aga_gpu_get_filter_t *filter = NULL);
 
     /// \brief return stringified key of the object (for debugging)
     virtual string key2str(void) const override {
@@ -296,11 +298,15 @@ class gpu_entry : public api_base {
     /// \brief      fill the gpu operational status
     /// \param[in]  spec config specification
     /// \param[out] status operational status
-    void fill_status_(aga_gpu_spec_t *spec, aga_gpu_status_t *status);
+    /// \param[in]  filter attributes to skip (NULL fetches all attributes)
+    void fill_status_(aga_gpu_spec_t *spec, aga_gpu_status_t *status,
+                      const aga_gpu_get_filter_t *filter);
 
     /// \brief      fill the gpu statistics
     /// \param[out] stats statistics
-    void fill_stats_(aga_gpu_stats_t *stats);
+    /// \param[in]  filter attributes to skip (NULL fetches all attributes)
+    void fill_stats_(aga_gpu_stats_t *stats,
+                     const aga_gpu_get_filter_t *filter);
 
 private:
     /// uuid of the object

sw/nic/gpuagent/api/gpu_api.cc

@@ -83,7 +83,8 @@ aga_gpu_create (_In_ aga_gpu_spec_t *spec)
 }
 
 sdk_ret_t
-aga_gpu_read (_In_ aga_obj_key_t *key, _Out_ aga_gpu_info_t *info)
+aga_gpu_read (_In_ aga_obj_key_t *key, _Out_ aga_gpu_info_t *info,
+              const aga_gpu_get_filter_t *filter)
 {
     sdk_ret_t ret;
     gpu_entry *entry;
@@ -95,12 +96,13 @@ aga_gpu_read (_In_ aga_obj_key_t *key, _Out_ aga_gpu_info_t *info)
     if (unlikely(ret != SDK_RET_OK)) {
         return ret;
     }
-    return entry->read(info);
+    return entry->read(info, filter);
 }
 
 typedef struct aga_gpu_read_args_s {
     void *ctxt;
     gpu_read_cb_t cb;
+    const aga_gpu_get_filter_t *filter;
 } aga_gpu_read_args_t;
 
 static bool
@@ -120,19 +122,21 @@ aga_gpu_info_from_entry (void *entry, void *ctxt)
     }
     memset(&info, 0, sizeof(aga_gpu_info_t));
     // call entry read
-    gpu->read(&info);
+    gpu->read(&info, args->filter);
     // call cb on info
     args->cb(&info, args->ctxt);
     return false;
 }
 
 sdk_ret_t
-aga_gpu_read_all (gpu_read_cb_t gpu_read_cb, void *ctxt)
+aga_gpu_read_all (gpu_read_cb_t gpu_read_cb, void *ctxt,
+                  const aga_gpu_get_filter_t *filter)
 {
     aga_gpu_read_args_t args = { 0 };
 
     args.ctxt = ctxt;
     args.cb = gpu_read_cb;
+    args.filter = filter;
     return gpu_db()->walk(aga_gpu_info_from_entry, &args);
 }
 

sw/nic/gpuagent/api/include/aga_gpu.hpp

@@ -74,6 +74,38 @@ typedef enum aga_gpu_admin_state_e {
     AGA_GPU_ADMIN_STATE_DOWN = 2,
 } aga_gpu_admin_state_t;
 
+/// \brief GPU attributes that can be skipped in a get request;
+///        default (all false) fetches all attributes
+typedef struct aga_gpu_get_filter_s {
+    /// skip clock status
+    bool skip_clock_status;
+    /// skip PCIe status
+    bool skip_pcie_status;
+    /// skip XGMI error status
+    bool skip_xgmi_status;
+    /// skip process list
+    bool skip_process_status;
+    /// skip UALink state
+    bool skip_ualink_status;
+    /// skip VRAM usage
+    bool skip_vram_usage_stats;
+    /// skip ECC error counts
+    bool skip_ecc_stats;
+    /// skip violation stats
+    bool skip_violation_stats;
+    /// skip PCIe stats
+    bool skip_pcie_stats;
+    /// skip XGMI counters
+    bool skip_xgmi_stats;
+    /// skip GPU activity and usage
+    bool skip_activity_stats;
+} aga_gpu_get_filter_t;
+
+/// true if the given attribute should be skipped;
+/// a NULL filter fetches all attributes
+#define AGA_GPU_SKIP(_filter_, _field_) \
+            (((_filter_) != NULL) && ((_filter_)->_field_))
+
 /// \brief GPU clock types
 typedef enum aga_gpu_clock_type_e {
     AGA_GPU_CLOCK_TYPE_NONE   = 0,
@@ -996,18 +1028,22 @@ typedef struct aga_cper_info_s {
 sdk_ret_t aga_gpu_create(_In_ aga_gpu_spec_t *spec);
 
 /// \brief      read gpu
-/// \param[in]  key  key of the gpu object
-/// \param[out] info information
+/// \param[in]  key    key of the gpu object
+/// \param[out] info   information
+/// \param[in]  filter attributes to skip (NULL fetches all attributes)
 /// \return     #SDK_RET_OK on success, failure status code on error
-sdk_ret_t aga_gpu_read(_In_ aga_obj_key_t *key, _Out_ aga_gpu_info_t *info);
+sdk_ret_t aga_gpu_read(_In_ aga_obj_key_t *key, _Out_ aga_gpu_info_t *info,
+                       const aga_gpu_get_filter_t *filter = NULL);
 
 typedef void (*gpu_read_cb_t)(aga_gpu_info_t *info, void *ctxt);
 
 /// \brief    read all gpu information
-/// \param[in]  cb      callback function
-/// \param[in]  ctxt    opaque context passed to cb
+/// \param[in]  cb     callback function
+/// \param[in]  ctxt   opaque context passed to cb
+/// \param[in]  filter attributes to skip (NULL fetches all attributes)
 /// \return #SDK_RET_OK on success, failure status code on error
-sdk_ret_t aga_gpu_read_all(_In_ gpu_read_cb_t gpu_read_cb, _In_ void *ctxt);
+sdk_ret_t aga_gpu_read_all(_In_ gpu_read_cb_t gpu_read_cb, _In_ void *ctxt,
+                           const aga_gpu_get_filter_t *filter = NULL);
 
 /// \brief      function to get compute partition info of a given physical gpu
 ///             which has been partitioned