Add NVIDIA self-heal tools and per-GPU SAT status

This commit is contained in:
Mikhail Chusavitin
2026-04-07 20:20:05 +03:00
parent 93cfa78e8c
commit 531d1ca366
13 changed files with 863 additions and 14 deletions

View File

@@ -122,6 +122,8 @@ type satRunner interface {
RunNvidiaPulseTestPack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, logFunc func(string)) (string, error)
RunNvidiaBandwidthPack(ctx context.Context, baseDir string, gpuIndices []int, logFunc func(string)) (string, error)
RunNvidiaStressPack(ctx context.Context, baseDir string, opts platform.NvidiaStressOptions, logFunc func(string)) (string, error)
ListNvidiaGPUStatuses() ([]platform.NvidiaGPUStatus, error)
ResetNvidiaGPU(index int) (string, error)
RunMemoryAcceptancePack(ctx context.Context, baseDir string, logFunc func(string)) (string, error)
RunStorageAcceptancePack(ctx context.Context, baseDir string, logFunc func(string)) (string, error)
RunCPUAcceptancePack(ctx context.Context, baseDir string, durationSec int, logFunc func(string)) (string, error)
@@ -521,6 +523,15 @@ func (a *App) ListNvidiaGPUs() ([]platform.NvidiaGPU, error) {
return a.sat.ListNvidiaGPUs()
}
func (a *App) ListNvidiaGPUStatuses() ([]platform.NvidiaGPUStatus, error) {
return a.sat.ListNvidiaGPUStatuses()
}
func (a *App) ResetNvidiaGPU(index int) (ActionResult, error) {
out, err := a.sat.ResetNvidiaGPU(index)
return ActionResult{Title: fmt.Sprintf("Reset NVIDIA GPU %d", index), Body: strings.TrimSpace(out)}, err
}
func (a *App) RunNvidiaAcceptancePackWithOptions(ctx context.Context, baseDir string, diagLevel int, gpuIndices []int, logFunc func(string)) (ActionResult, error) {
if strings.TrimSpace(baseDir) == "" {
baseDir = DefaultSATBaseDir

View File

@@ -135,6 +135,8 @@ type fakeSAT struct {
listAMDGPUsFn func() ([]platform.AMDGPUInfo, error)
runAMDPackFn func(string) (string, error)
listNvidiaGPUsFn func() ([]platform.NvidiaGPU, error)
listNvidiaGPUStatusesFn func() ([]platform.NvidiaGPUStatus, error)
resetNvidiaGPUFn func(int) (string, error)
}
func (f fakeSAT) RunNvidiaAcceptancePack(baseDir string, _ func(string)) (string, error) {
@@ -201,6 +203,20 @@ func (f fakeSAT) ListNvidiaGPUs() ([]platform.NvidiaGPU, error) {
return nil, nil
}
func (f fakeSAT) ListNvidiaGPUStatuses() ([]platform.NvidiaGPUStatus, error) {
if f.listNvidiaGPUStatusesFn != nil {
return f.listNvidiaGPUStatusesFn()
}
return nil, nil
}
func (f fakeSAT) ResetNvidiaGPU(index int) (string, error) {
if f.resetNvidiaGPUFn != nil {
return f.resetNvidiaGPUFn(index)
}
return "", nil
}
func (f fakeSAT) RunMemoryAcceptancePack(_ context.Context, baseDir string, _ func(string)) (string, error) {
return f.runMemoryFn(baseDir)
}
@@ -805,6 +821,9 @@ func TestBuildSupportBundleIncludesExportDirContents(t *testing.T) {
for _, want := range []string{
"/system/ip-link.txt",
"/system/ip-link-stats.txt",
"/system/kernel-aer-nvidia.txt",
"/system/lspci-nvidia-bridges-vv.txt",
"/system/pcie-aer-sysfs.txt",
"/system/ethtool-info.txt",
"/system/ethtool-link.txt",
"/system/ethtool-module.txt",

View File

@@ -3,6 +3,7 @@ package app
import (
"os"
"path/filepath"
"strconv"
"sort"
"strings"
@@ -18,6 +19,7 @@ func applyLatestSATStatuses(snap *schema.HardwareSnapshot, baseDir string, db *C
}
if summary, ok := loadLatestSATSummary(baseDir, "gpu-nvidia-"); ok {
applyGPUVendorSAT(snap.PCIeDevices, "nvidia", summary)
applyNvidiaPerGPUStatus(snap.PCIeDevices, baseDir)
}
if summary, ok := loadLatestSATSummary(baseDir, "memory-"); ok {
applyMemorySAT(snap.Memory, summary)
@@ -32,6 +34,100 @@ func applyLatestSATStatuses(snap *schema.HardwareSnapshot, baseDir string, db *C
applyComponentStatusDB(snap, db)
}
type nvidiaPerGPUStatus struct {
runStatus string
reason string
}
func applyNvidiaPerGPUStatus(devs []schema.HardwarePCIeDevice, baseDir string) {
statusByIndex, ts, ok := loadLatestNvidiaPerGPUStatus(baseDir)
if !ok {
return
}
for i := range devs {
if devs[i].Telemetry == nil {
continue
}
rawIdx, ok := devs[i].Telemetry["nvidia_gpu_index"]
if !ok {
continue
}
idx, ok := telemetryInt(rawIdx)
if !ok {
continue
}
st, ok := statusByIndex[idx]
if !ok {
continue
}
status, description, ok := satKeyStatus(st.runStatus, firstNonEmpty(strings.TrimSpace(st.reason), "nvidia GPU SAT"))
if !ok {
continue
}
mergeComponentStatusPreferDetail(&devs[i].HardwareComponentStatus, ts, status, description)
}
}
func loadLatestNvidiaPerGPUStatus(baseDir string) (map[int]nvidiaPerGPUStatus, string, bool) {
matches, err := filepath.Glob(filepath.Join(baseDir, "gpu-nvidia-*"))
if err != nil || len(matches) == 0 {
return nil, "", false
}
sort.Strings(matches)
runDir := matches[len(matches)-1]
summaryRaw, err := os.ReadFile(filepath.Join(runDir, "summary.txt"))
if err != nil {
return nil, "", false
}
summaryKV := parseKeyValueSummary(string(summaryRaw))
runAtUTC := strings.TrimSpace(summaryKV["run_at_utc"])
files, err := filepath.Glob(filepath.Join(runDir, "gpu-*-status.txt"))
if err != nil || len(files) == 0 {
return nil, "", false
}
out := make(map[int]nvidiaPerGPUStatus, len(files))
for _, file := range files {
raw, err := os.ReadFile(file)
if err != nil {
continue
}
kv := parseKeyValueSummary(string(raw))
idx, err := strconv.Atoi(strings.TrimSpace(kv["gpu_index"]))
if err != nil {
continue
}
out[idx] = nvidiaPerGPUStatus{
runStatus: strings.ToUpper(strings.TrimSpace(kv["run_status"])),
reason: strings.TrimSpace(kv["reason"]),
}
}
if len(out) == 0 {
return nil, "", false
}
return out, runAtUTC, true
}
func telemetryInt(v any) (int, bool) {
switch value := v.(type) {
case int:
return value, true
case int32:
return int(value), true
case int64:
return int(value), true
case float64:
return int(value), true
case string:
n, err := strconv.Atoi(strings.TrimSpace(value))
if err != nil {
return 0, false
}
return n, true
default:
return 0, false
}
}
type satSummary struct {
runAtUTC string
overall string
@@ -176,6 +272,31 @@ func mergeComponentStatus(component *schema.HardwareComponentStatus, changedAt,
}
}
func mergeComponentStatusPreferDetail(component *schema.HardwareComponentStatus, changedAt, satStatus, description string) {
if component == nil || satStatus == "" {
return
}
current := strings.TrimSpace(ptrString(component.Status))
newSeverity := statusSeverity(satStatus)
currentSeverity := statusSeverity(current)
if current == "" || current == "Unknown" || newSeverity > currentSeverity {
mergeComponentStatus(component, changedAt, satStatus, description)
return
}
if newSeverity == currentSeverity && strings.TrimSpace(description) != "" {
component.Status = appStringPtr(satStatus)
component.ErrorDescription = appStringPtr(description)
if strings.TrimSpace(changedAt) != "" {
component.StatusChangedAt = appStringPtr(changedAt)
component.StatusHistory = append(component.StatusHistory, schema.HardwareStatusHistory{
Status: satStatus,
ChangedAt: changedAt,
Details: appStringPtr(description),
})
}
}
}
func statusSeverity(status string) int {
switch strings.TrimSpace(status) {
case "Critical":

View File

@@ -59,3 +59,51 @@ func TestApplyLatestSATStatusesMarksAMDGPUs(t *testing.T) {
t.Fatalf("gpu status=%v want Critical", snap.PCIeDevices[0].Status)
}
}
func TestApplyLatestSATStatusesMarksNvidiaGPUByPerGPUStatusFile(t *testing.T) {
baseDir := t.TempDir()
runDir := filepath.Join(baseDir, "gpu-nvidia-20260407-162123")
if err := os.MkdirAll(runDir, 0755); err != nil {
t.Fatal(err)
}
if err := os.WriteFile(filepath.Join(runDir, "summary.txt"), []byte("run_at_utc=2026-04-07T16:21:23Z\noverall_status=FAILED\n"), 0644); err != nil {
t.Fatal(err)
}
if err := os.WriteFile(filepath.Join(runDir, "gpu-1-status.txt"), []byte("gpu_index=1\ngpu_name=NVIDIA H100 PCIe\nrun_status=FAILED\nreason=GPU requires reset\n"), 0644); err != nil {
t.Fatal(err)
}
class := "VideoController"
manufacturer := "NVIDIA Corporation"
bdf0 := "0000:4b:00.0"
bdf1 := "0000:4f:00.0"
snap := schema.HardwareSnapshot{
PCIeDevices: []schema.HardwarePCIeDevice{
{
DeviceClass: &class,
Manufacturer: &manufacturer,
BDF: &bdf0,
Telemetry: map[string]any{"nvidia_gpu_index": 0},
},
{
DeviceClass: &class,
Manufacturer: &manufacturer,
BDF: &bdf1,
Telemetry: map[string]any{"nvidia_gpu_index": 1},
},
},
}
applyLatestSATStatuses(&snap, baseDir, nil)
if snap.PCIeDevices[1].Status == nil || *snap.PCIeDevices[1].Status != "Critical" {
t.Fatalf("gpu1 status=%v want Critical", snap.PCIeDevices[1].Status)
}
if snap.PCIeDevices[1].ErrorDescription == nil || *snap.PCIeDevices[1].ErrorDescription != "GPU requires reset failed" {
got := "<nil>"
if snap.PCIeDevices[1].ErrorDescription != nil {
got = *snap.PCIeDevices[1].ErrorDescription
}
t.Fatalf("gpu1 error=%q want per-gpu reason", got)
}
}

View File

@@ -40,7 +40,36 @@ var supportBundleCommands = []struct {
{name: "system/mount.txt", cmd: []string{"mount"}},
{name: "system/df-h.txt", cmd: []string{"df", "-h"}},
{name: "system/dmesg.txt", cmd: []string{"dmesg"}},
{name: "system/kernel-aer-nvidia.txt", cmd: []string{"sh", "-c", `
if command -v dmesg >/dev/null 2>&1; then
dmesg | grep -iE 'AER|NVRM|Xid|pcieport|nvidia' || echo "no AER/NVRM/Xid kernel messages found"
else
echo "dmesg not found"
fi
`}},
{name: "system/nvidia-smi-q.txt", cmd: []string{"nvidia-smi", "-q"}},
{name: "system/lspci-nvidia-bridges-vv.txt", cmd: []string{"sh", "-c", `
if ! command -v lspci >/dev/null 2>&1; then
echo "lspci not found"
exit 0
fi
found=0
for gpu in $(lspci -Dn | awk '$3 ~ /^10de:/ {print $1}'); do
found=1
echo "=== GPU $gpu ==="
lspci -s "$gpu" -vv 2>&1 || true
bridge=$(basename "$(readlink -f "/sys/bus/pci/devices/$gpu/.." 2>/dev/null)" 2>/dev/null)
if [ -n "$bridge" ] && [ "$bridge" != "$gpu" ]; then
echo
echo "=== UPSTREAM $bridge for $gpu ==="
lspci -s "$bridge" -vv 2>&1 || true
fi
echo
done
if [ "$found" -eq 0 ]; then
echo "no NVIDIA PCI devices found"
fi
`}},
{name: "system/pcie-nvidia-link.txt", cmd: []string{"sh", "-c", `
for d in /sys/bus/pci/devices/*/; do
vendor=$(cat "$d/vendor" 2>/dev/null)
@@ -51,6 +80,30 @@ for d in /sys/bus/pci/devices/*/; do
printf " %-22s %s\n" "$f" "$(cat "$d/$f" 2>/dev/null)"
done
done
`}},
{name: "system/pcie-aer-sysfs.txt", cmd: []string{"sh", "-c", `
found=0
for dev in /sys/bus/pci/devices/*; do
[ -e "$dev" ] || continue
bdf=$(basename "$dev")
block=""
for f in aer_dev_correctable aer_dev_fatal aer_dev_nonfatal aer_rootport_total_err_cor aer_rootport_total_err_fatal aer_rootport_total_err_nonfatal; do
if [ -r "$dev/$f" ]; then
if [ -z "$block" ]; then
block=1
found=1
echo "=== $bdf ==="
fi
printf " %-30s %s\n" "$f" "$(cat "$dev/$f" 2>/dev/null)"
fi
done
if [ -n "$block" ]; then
echo
fi
done
if [ "$found" -eq 0 ]; then
echo "no PCIe AER sysfs counters found"
fi
`}},
{name: "system/ethtool-info.txt", cmd: []string{"sh", "-c", `
if ! command -v ethtool >/dev/null 2>&1; then