[support bundle] Add health checks to support bundles

Author: smklein

nexus/src/app/background/tasks/support_bundle_collector.rs

@@ -662,6 +662,12 @@ impl BundleCollection<'_> {
                         sled_client.support_zpool_info(),
                     )
                     .boxed(),
+                    save_diag_cmd_output_or_error(
+                        &sled_path,
+                        "health-check",
+                        sled_client.support_health_check(),
+                    )
+                    .boxed(),
                 ])
                 // Currently we execute up to 10 commands concurrently which
                 // might be doing their own concurrent work, for example

openapi/sled-agent.json

@@ -659,6 +659,33 @@
         }
       }
     },
+    "/support/health-check": {
+      "get": {
+        "operationId": "support_health_check",
+        "responses": {
+          "200": {
+            "description": "successful operation",
+            "content": {
+              "application/json": {
+                "schema": {
+                  "title": "Array_of_SledDiagnosticsQueryOutput",
+                  "type": "array",
+                  "items": {
+                    "$ref": "#/components/schemas/SledDiagnosticsQueryOutput"
+                  }
+                }
+              }
+            }
+          },
+          "4XX": {
+            "$ref": "#/components/responses/Error"
+          },
+          "5XX": {
+            "$ref": "#/components/responses/Error"
+          }
+        }
+      }
+    },
     "/support/ipadm-info": {
       "get": {
         "operationId": "support_ipadm_info",

sled-agent/api/src/lib.rs

@@ -673,6 +673,14 @@ pub trait SledAgentApi {
         request_context: RequestContext<Self::Context>,
     ) -> Result<HttpResponseOk<SledDiagnosticsQueryOutput>, HttpError>;
 
+    #[endpoint {
+        method = GET,
+        path = "/support/health-check",
+    }]
+    async fn support_health_check(
+        request_context: RequestContext<Self::Context>,
+    ) -> Result<HttpResponseOk<Vec<SledDiagnosticsQueryOutput>>, HttpError>;
+
     /// This endpoint returns a list of known zones on a sled that have service
     /// logs that can be collected into a support bundle.
     #[endpoint {

sled-agent/src/http_entrypoints.rs

@@ -1036,6 +1036,20 @@ impl SledAgentApi for SledAgentImpl {
         Ok(HttpResponseOk(res.get_output()))
     }
 
+    async fn support_health_check(
+        request_context: RequestContext<Self::Context>,
+    ) -> Result<HttpResponseOk<Vec<SledDiagnosticsQueryOutput>>, HttpError>
+    {
+        let sa = request_context.context();
+        Ok(HttpResponseOk(
+            sa.support_health_check()
+                .await
+                .into_iter()
+                .map(|cmd| cmd.get_output())
+                .collect::<Vec<_>>(),
+        ))
+    }
+
     async fn support_logs(
         request_context: RequestContext<Self::Context>,
     ) -> Result<HttpResponseOk<Vec<String>>, HttpError> {

sled-agent/src/sim/http_entrypoints.rs

@@ -745,6 +745,13 @@ impl SledAgentApi for SledAgentSimImpl {
         method_unimplemented()
     }
 
+    async fn support_health_check(
+        _request_context: RequestContext<Self::Context>,
+    ) -> Result<HttpResponseOk<Vec<SledDiagnosticsQueryOutput>>, HttpError>
+    {
+        method_unimplemented()
+    }
+
     async fn support_logs(
         _request_context: RequestContext<Self::Context>,
     ) -> Result<HttpResponseOk<Vec<String>>, HttpError> {

sled-agent/src/sled_agent.rs

@@ -67,7 +67,8 @@ use sled_agent_types::zone_bundle::{
     BundleUtilization, CleanupContext, CleanupCount, CleanupPeriod,
     PriorityOrder, StorageLimit, ZoneBundleMetadata,
 };
-use sled_diagnostics::{SledDiagnosticsCmdError, SledDiagnosticsCmdOutput};
+use sled_diagnostics::SledDiagnosticsCmdError;
+use sled_diagnostics::SledDiagnosticsCmdOutput;
 use sled_hardware::{HardwareManager, MemoryReservations, underlay};
 use sled_hardware_types::Baseboard;
 use sled_hardware_types::underlay::BootstrapInterface;
@@ -1474,6 +1475,12 @@ impl SledAgent {
     ) -> Result<SledDiagnosticsCmdOutput, SledDiagnosticsCmdError> {
         sled_diagnostics::zpool_info().await
     }
+
+    pub(crate) async fn support_health_check(
+        &self,
+    ) -> Vec<Result<SledDiagnosticsCmdOutput, SledDiagnosticsCmdError>> {
+        sled_diagnostics::health_check().await
+    }
 }
 
 #[derive(From, thiserror::Error, Debug)]