bug#65255: uptime's boot time is inconsistent after VM sleep & resume

[Bruno Haible <bruno@clisp.org>]

[-- Attachment #1: Type: text/plain, Size: 2754 bytes --]

On the following platforms:
  - Linux (that includes glibc-based systems, Alpine Linux, Raspbian),
  - NetBSD,
  - Cygwin,
  - Minix,

the "up" duration is inconsistent after the VM in which the OS is running
has been
  1. put into sleep / saved / pause mode (terminology depends on the
     hypervisor),
  2. resumed,
  3. a date bump has been done in the VM (either automatically or manually).

VM sleep and a date adjustment after resume is a common operation, as can
be seen
  - from the fact that it's automatic in VirtualBox machines with the
    "guest extensions" installed [1],
  - from the fact that it's automatic in KVM [2],
  - from the fact that it's automatic in Haiku as guest,
  - from a remark in the libvirt documentation [3],
  - from stackoverflow questions such as [4].

After such a VM sleep and a date adjustment, coreutils' "uptime" is
inconsistent in three ways:

  * On Linux, the displayed "up" duration includes active time and
    hardware suspend time [5], but excludes VM sleep time.

    Such a figure is not useful for estimating the boot time, since
    (time now) - (that "up" time) is not the boot time.

    It is also not useful for estimating when e.g. a file system check
    would be necessary on file systems without a journal, or how much
    electricity was consumed — since the hardware suspend time was included.

    So this figure is useless.

  * The displayed "up" duration is inconsistent with the boot time displayed
    by "who -a". (Since the Gnulib module 'readutmp' makes efforts to find
    the boot time in a way that does not change when the date is adjusted.)

  * The behaviour is inconsistent among platforms: On platforms which have
    a /proc/uptime file, the displayed "up" time _excludes_ VM sleep time.
    On the other platforms it _includes_ VM sleep time.

Seen on Linux (Fedora Rawhide, Alpine Linux 3.18, Raspbian), NetBSD 9.3,
Cygwin 2.9.0, Minix 3.3.

I see this as a bug. Find attached a fix for this bug. I'll provide a NEWS
entry afterwards, that summarizes the changes in coreutils + gnulib on the
various platforms.

Bruno


[1] https://docs.oracle.com/en/virtualization/virtualbox/6.1/user/guestadditions.html#4.1.-Introduction-to-Guest-Additions
[2] https://bugzilla.redhat.com/show_bug.cgi?id=1115340
[3] https://libvirt.gitlab.io/libvirt-appdev-guide-python/libvirt_application_development_guide_using_python-Guest_Domains-Lifecycle-Save.html
[4] https://serverfault.com/questions/334698/
[5] This is indicated by glibc's <bits/time.h>:
      /* Monotonic system-wide clock that includes time spent in suspension.  */
      # define CLOCK_BOOTTIME                 7
    I also verified this on a laptop.

[-- Attachment #2: 0001-uptime-Include-VM-sleep-time-in-the-up-duration.patch --]
[-- Type: text/x-patch, Size: 2725 bytes --]

From ff6e92f38b560ade54df4c94e9d91657f740bb4b Mon Sep 17 00:00:00 2001
From: Bruno Haible <bruno@clisp.org>
Date: Sun, 13 Aug 2023 00:57:23 +0200
Subject: [PATCH] uptime: Include VM sleep time in the "up" duration

* src/uptime.c: Don't include c-strtod.h.
(print_uptime): Don't read /proc/uptime, because the value it provides
does not change when a date adjustment occurs.
* bootstrap.conf (gnulib_modules): Remove 'uptime'.
---
 bootstrap.conf |  1 -
 src/uptime.c   | 36 ++++++------------------------------
 2 files changed, 6 insertions(+), 31 deletions(-)

diff --git a/bootstrap.conf b/bootstrap.conf
index 6ce27d5dd..d758a9ff6 100644
--- a/bootstrap.conf
+++ b/bootstrap.conf
@@ -285,7 +285,6 @@ gnulib_modules="
   unlocked-io
   unsetenv
   update-copyright
-  uptime
   useless-if-before-free
   userspec
   utimecmp
diff --git a/src/uptime.c b/src/uptime.c
index c29798d32..0cc1dedba 100644
--- a/src/uptime.c
+++ b/src/uptime.c
@@ -22,7 +22,6 @@
 #include <sys/types.h>
 #include "system.h"
 
-#include "c-strtod.h"
 #include "long-options.h"
 #include "quote.h"
 #include "readutmp.h"
@@ -42,33 +41,13 @@ print_uptime (idx_t n, struct gl_utmp const *this)
   idx_t entries = 0;
   time_t boot_time = 0;
   time_t time_now;
-  time_t uptime = 0;
+  time_t uptime;
   intmax_t updays;
   int uphours;
   int upmins;
   struct tm *tmn;
   double avg[3];
   int loads;
-#ifdef HAVE_PROC_UPTIME
-  FILE *fp;
-
-  fp = fopen ("/proc/uptime", "r");
-  if (fp != nullptr)
-    {
-      char buf[BUFSIZ];
-      char *b = fgets (buf, BUFSIZ, fp);
-      if (b == buf)
-        {
-          char *end_ptr;
-          double upsecs = c_strtod (buf, &end_ptr);
-          if (buf != end_ptr)
-            uptime = (0 <= upsecs && upsecs < TYPE_MAXIMUM (time_t)
-                      ? upsecs : -1);
-        }
-
-      fclose (fp);
-    }
-#endif /* HAVE_PROC_UPTIME */
 
   /* Loop through all the utmp entries we just read and count up the valid
      ones, also in the process possibly gleaning boottime. */
@@ -79,16 +58,13 @@ print_uptime (idx_t n, struct gl_utmp const *this)
         boot_time = this->ut_ts.tv_sec;
       ++this;
     }
+  /* The gnulib module 'readutmp' is supposed to provide a BOOT_TIME entry
+     on all platforms.  */
+  if (boot_time == 0)
+    error (EXIT_FAILURE, errno, _("couldn't get boot time"));
 
   time_now = time (nullptr);
-#if defined HAVE_PROC_UPTIME
-  if (uptime == 0)
-#endif
-    {
-      if (boot_time == 0)
-        error (EXIT_FAILURE, errno, _("couldn't get boot time"));
-      uptime = time_now - boot_time;
-    }
+  uptime = time_now - boot_time;
   updays = uptime / 86400;
   uphours = uptime % 86400 / 3600;
   upmins = uptime % 86400 % 3600 / 60;
-- 
2.34.1