dd: work around Linux kernel CLOCK_MONOTONIC bug

Problem reported by Sick Pigs (bug#81269).
Ordinarily we’d say “fix the kernel or the hardware”,
but this bug seems widespread and unlikely to be fixed any time soon,
and the workaround is not a lot of trouble.
* src/dd.c (real_start_time): New static var.
(getrealxtime): New static function.
(print_xfer_stats): Use the max of elapsed monotonic and elapsed
real time.
This commit is contained in:
Paul Eggert committed 2026-06-20 12:20:33 -07:00
1 parent 6a850b9aac
commit 2e7707c23e
2 files changed
+23 -6

No files matched your search

+3
View File
@@ -12,6 +12,9 @@ GNU coreutils NEWS -*- outline -*-
will correctly match the last delimiter specified.
[bug introduced with multi-byte support in coreutils-9.11]
'dd conv=fsync' no longer outputs bogus stats if the monotonic clock freezes.
[This works around a too-common bug in the Linux kernel.]
'head' and 'tail' now quote names in file headers when needed.
[This bug was present in "the beginning".]
+20 -6
View File
@@ -181,8 +181,8 @@ static intmax_t w_bytes = 0;
/* Last-reported number of bytes written, or negative if never reported. */
static intmax_t reported_w_bytes = -1;
/* Time that dd started. */
static xtime_t start_time;
/* Time that dd started, in both monotonic and real time. */
static xtime_t start_time, real_start_time;
/* Next time to report periodic progress. */
static xtime_t next_time;
@@ -740,12 +740,20 @@ abbreviation_lacks_prefix (char const *message)
return message[strlen (message) - 2] == ' ';
}
static xtime_t
getrealxtime (void)
{
struct timespec now = current_timespec ();
return xtime_make (now.tv_sec, now.tv_nsec);
}
/* Print transfer statistics. */
static void
print_xfer_stats (xtime_t progress_time)
{
xtime_t now = progress_time ? progress_time : gethrxtime ();
xtime_t real_now = getrealxtime ();
static char const slash_s[] = "/s";
char hbuf[3][LONGEST_HUMAN_READABLE + sizeof slash_s];
double delta_s;
@@ -754,14 +762,19 @@ print_xfer_stats (xtime_t progress_time)
char const *iec = human_readable (w_bytes, hbuf[1],
human_opts | human_base_1024, 1, 1);
/* Use integer arithmetic to compute the transfer rate,
since that makes it easy to use SI abbreviations. */
/* Use integer arithmetic to compute the transfer rate, since that
makes it easy to use SI abbreviations. Although we want the
elapsed monotonic time, too often the Linux kernel stops its
monotonic clocks during an fsync/fdatasync (coreutils bug#81269),
and perhaps other kernels have similar bugs. So take the maximum
of the elapsed monotonic and real times, even though this can
overestimate if the real clock was set during our run. */
char *bpsbuf = hbuf[2];
int bpsbufsize = sizeof hbuf[2];
if (start_time < now)
xtime_t delta_xtime = MAX (now - start_time, real_now - real_start_time);
if (0 < delta_xtime)
{
double XTIME_PRECISIONe0 = XTIME_PRECISION;
xtime_t delta_xtime = now - start_time;
delta_s = delta_xtime / XTIME_PRECISIONe0;
bytes_per_second = human_readable (w_bytes, bpsbuf, human_opts,
XTIME_PRECISION, delta_xtime);
@@ -2537,6 +2550,7 @@ main (int argc, char **argv)
}
start_time = gethrxtime ();
real_start_time = getrealxtime ();
next_time = start_time + XTIME_PRECISION;
int copy_status = dd_copy ();