summaryrefslogtreecommitdiff
path: root/subversion/libsvn_subr/io.c
diff options
context:
space:
mode:
Diffstat (limited to 'subversion/libsvn_subr/io.c')
-rw-r--r--subversion/libsvn_subr/io.c38
1 files changed, 25 insertions, 13 deletions
diff --git a/subversion/libsvn_subr/io.c b/subversion/libsvn_subr/io.c
index f0956e2aa8e94..08f2e326d7fe6 100644
--- a/subversion/libsvn_subr/io.c
+++ b/subversion/libsvn_subr/io.c
@@ -1243,32 +1243,44 @@ svn_io_sleep_for_timestamps(const char *path, apr_pool_t *pool)
{
/* Very simplistic but safe approach:
If the filesystem has < sec mtime we can be reasonably sure
- that the filesystem has <= millisecond precision.
+ that the filesystem has some sub-second resolution. On Windows
+ it is likely to be sub-millisecond; on Linux systems it depends
+ on the filesystem, ext4 is typically 1ms, 4ms or 10ms resolution.
## Perhaps find a better algorithm here. This will fail once
- in every 1000 cases on a millisecond precision filesystem.
+ in every 1000 cases on a millisecond precision filesystem
+ if the mtime happens to be an exact second.
But better to fail once in every thousand cases than every
time, like we did before.
- (All tested filesystems I know have at least microsecond precision.)
Note for further research on algorithm:
- FAT32 has < 1 sec precision on ctime, but 2 sec on mtime */
+ FAT32 has < 1 sec precision on ctime, but 2 sec on mtime.
- /* Sleep for at least 1 millisecond.
- (t < 1000 will be round to 0 in apr) */
- apr_sleep(1000);
+ Linux/ext4 with CONFIG_HZ=250 has high resolution
+ apr_time_now and although the filesystem timestamps
+ have similar high precision they are only updated with
+ a coarser 4ms resolution. */
- return;
+ /* 10 milliseconds after now. */
+#ifndef SVN_HI_RES_SLEEP_MS
+#define SVN_HI_RES_SLEEP_MS 10
+#endif
+ then = now + apr_time_from_msec(SVN_HI_RES_SLEEP_MS);
}
- now = apr_time_now(); /* Extract the time used for the path stat */
-
- if (now >= then)
- return; /* Passing negative values may suspend indefinitely (Windows) */
+ /* Remove time taken to do stat() from sleep. */
+ now = apr_time_now();
}
- apr_sleep(then - now);
+ if (now >= then)
+ return; /* Passing negative values may suspend indefinitely (Windows) */
+
+ /* (t < 1000 will be round to 0 in apr) */
+ if (then - now < 1000)
+ apr_sleep(1000);
+ else
+ apr_sleep(then - now);
}