qga-win: implement a 'guest-get-load' command

Windows has no native equivalent API, but it would be possible to
simulate it as illustrated here (BSD-3-Clause):

  https://github.com/giampaolo/psutil/pull/1485

Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Tested-by: Dehan Meng <demeng@redhat.com>
Reviewed-by: Yan Vugenfirer <yvugenfi@redhat.com>
Signed-off-by: Konstantin Kostiuk <kkostiuk@redhat.com>

1 files changed, 148 insertions, 0 deletions

diff --git a/qga/commands-win32.c b/qga/commands-win32.c
index d4482538ec..8227480810 100644
--- a/qga/commands-win32.c
+++ b/qga/commands-win32.c
@@ -27,6 +27,7 @@
 #include <lm.h>
 #include <wtsapi32.h>
 #include <wininet.h>
+#include <pdh.h>
 
 #include "guest-agent-core.h"
 #include "vss-win32.h"
@@ -119,6 +120,28 @@ static OpenFlags guest_file_open_modes[] = {
     {"a+b", FILE_GENERIC_APPEND | GENERIC_READ, OPEN_ALWAYS  }
 };
 
+/*
+ * We use an exponentially weighted moving average, just like Unix systems do
+ * https://en.wikipedia.org/wiki/Load_(computing)#Unix-style_load_calculation
+ *
+ * These constants serve as the damping factor and are calculated with
+ * 1 / exp(sampling interval in seconds / window size in seconds)
+ *
+ * This formula comes from linux's include/linux/sched/loadavg.h
+ * https://github.com/torvalds/linux/blob/345671ea0f9258f410eb057b9ced9cefbbe5dc78/include/linux/sched/loadavg.h#L20-L23
+ */
+#define LOADAVG_FACTOR_1F  0.9200444146293232478931553241
+#define LOADAVG_FACTOR_5F  0.9834714538216174894737477501
+#define LOADAVG_FACTOR_15F 0.9944598480048967508795473394
+/*
+ * The time interval in seconds between taking load counts, same as Linux
+ */
+#define LOADAVG_SAMPLING_INTERVAL 5
+
+double load_avg_1m;
+double load_avg_5m;
+double load_avg_15m;
+
 #define debug_error(msg) do { \
     char *suffix = g_win32_error_message(GetLastError()); \
     g_debug("%s: %s", (msg), suffix); \
@@ -2444,3 +2467,128 @@ char *qga_get_host_name(Error **errp)
 
     return g_utf16_to_utf8(tmp, size, NULL, NULL, NULL);
 }
+
+
+static VOID CALLBACK load_avg_callback(PVOID hCounter, BOOLEAN timedOut)
+{
+    PDH_FMT_COUNTERVALUE displayValue;
+    double currentLoad;
+    PDH_STATUS err;
+
+    err = PdhGetFormattedCounterValue(
+        (PDH_HCOUNTER)hCounter, PDH_FMT_DOUBLE, 0, &displayValue);
+    /* Skip updating the load if we can't get the value successfully */
+    if (err != ERROR_SUCCESS) {
+        slog("PdhGetFormattedCounterValue failed to get load value with 0x%lx",
+             err);
+        return;
+    }
+    currentLoad = displayValue.doubleValue;
+
+    load_avg_1m = load_avg_1m * LOADAVG_FACTOR_1F + currentLoad * \
+        (1.0 - LOADAVG_FACTOR_1F);
+    load_avg_5m = load_avg_5m * LOADAVG_FACTOR_5F + currentLoad * \
+        (1.0 - LOADAVG_FACTOR_5F);
+    load_avg_15m = load_avg_15m * LOADAVG_FACTOR_15F + currentLoad * \
+        (1.0 - LOADAVG_FACTOR_15F);
+}
+
+static BOOL init_load_avg_counter(Error **errp)
+{
+    CONST WCHAR *szCounterPath = L"\\System\\Processor Queue Length";
+    PDH_STATUS status;
+    BOOL ret;
+    HQUERY hQuery;
+    HCOUNTER hCounter;
+    HANDLE event;
+    HANDLE waitHandle;
+
+    status = PdhOpenQueryW(NULL, 0, &hQuery);
+    if (status != ERROR_SUCCESS) {
+        /*
+         * If the function fails, the return value is a system error code or
+         * a PDH error code. error_setg_win32 cant translate PDH error code
+         * properly, so just report it as is.
+         */
+        error_setg_win32(errp, (DWORD)status,
+                         "PdhOpenQueryW failed with 0x%lx", status);
+        return FALSE;
+    }
+
+    status = PdhAddEnglishCounterW(hQuery, szCounterPath, 0, &hCounter);
+    if (status != ERROR_SUCCESS) {
+        error_setg_win32(errp, (DWORD)status,
+            "PdhAddEnglishCounterW failed with 0x%lx. Performance counters may be disabled.",
+            status);
+        PdhCloseQuery(hQuery);
+        return FALSE;
+    }
+
+    event = CreateEventW(NULL, FALSE, FALSE, L"LoadUpdateEvent");
+    if (event == NULL) {
+        error_setg_win32(errp, GetLastError(), "Create LoadUpdateEvent failed");
+        PdhCloseQuery(hQuery);
+        return FALSE;
+    }
+
+    status = PdhCollectQueryDataEx(hQuery, LOADAVG_SAMPLING_INTERVAL, event);
+    if (status != ERROR_SUCCESS) {
+        error_setg_win32(errp, (DWORD)status,
+                         "PdhCollectQueryDataEx failed with 0x%lx", status);
+        CloseHandle(event);
+        PdhCloseQuery(hQuery);
+        return FALSE;
+    }
+
+    ret = RegisterWaitForSingleObject(
+        &waitHandle,
+        event,
+        (WAITORTIMERCALLBACK)load_avg_callback,
+        (PVOID)hCounter,
+        INFINITE,
+        WT_EXECUTEDEFAULT);
+
+    if (ret == 0) {
+        error_setg_win32(errp, GetLastError(),
+                         "RegisterWaitForSingleObject failed");
+        CloseHandle(event);
+        PdhCloseQuery(hQuery);
+        return FALSE;
+    }
+
+    ga_set_load_avg_wait_handle(ga_state, waitHandle);
+    ga_set_load_avg_event(ga_state, event);
+    ga_set_load_avg_pdh_query(ga_state, hQuery);
+
+    return TRUE;
+}
+
+GuestLoadAverage *qmp_guest_get_load(Error **errp)
+{
+    /*
+     * The load average logic calls PerformaceCounterAPI, which can result
+     * in a performance penalty. This avoids running the load average logic
+     * until a management application actually requests it. The load average
+     * will not initially be very accurate, but assuming that any interested
+     * management application will request it repeatedly throughout the lifetime
+     * of the VM, this seems like a good mitigation.
+     */
+    if (ga_get_load_avg_pdh_query(ga_state) == NULL) {
+        /* set initial values */
+        load_avg_1m = 0;
+        load_avg_5m = 0;
+        load_avg_15m = 0;
+
+        if (init_load_avg_counter(errp) == false) {
+            return NULL;
+        }
+    }
+
+    GuestLoadAverage *ret = NULL;
+
+    ret = g_new0(GuestLoadAverage, 1);
+    ret->load1m = load_avg_1m;
+    ret->load5m = load_avg_5m;
+    ret->load15m = load_avg_15m;
+    return ret;
+}