"Don't fear the fsync()"

For files which are vital to boot

1. Avoid opening any window where power loss will zero them out or worse.
   I know app developers all coded to the ext3 implementation, but
   the only formal documentation we have says we're broken if we actually
   rely on it.  E.g.

   * `man mount`, search for `auto_da_alloc`.
   * http://www.linux-mtd.infradead.org/faq/ubifs.html#L_atomic_change
   * https://thunk.org/tytso/blog/2009/03/15/dont-fear-the-fsync/

2. If we tell the kernel we're interested in writing them to disk, it will
   tell us if that fails.  So at minimum, this means we play our part in
   notifying the user about errors.

I refactored error-handling in `udevadm-hwdb` a little.  It turns out I did
exactly the same as had already been done in the `systemd-hwdb` version,
i.e. commit d702dcd.
This commit is contained in:
Alan Jenkins
2017-08-17 20:26:36 +01:00
parent dce892acef
commit 0675e94ab5
8 changed files with 100 additions and 47 deletions
+10 -4
View File
@@ -233,9 +233,15 @@ static int make_backup(const char *target, const char *x) {
if (futimens(fileno(dst), ts) < 0)
log_warning_errno(errno, "Failed to fix access and modification time of %s: %m", backup);
if (rename(temp, backup) < 0)
r = fflush_sync_and_check(dst);
if (r < 0)
goto fail;
if (rename(temp, backup) < 0) {
r = -errno;
goto fail;
}
return 0;
fail:
@@ -532,7 +538,7 @@ static int write_temporary_shadow(const char *shadow_path, FILE **tmpfile, char
return errno ? -errno : -EIO;
}
r = fflush_and_check(shadow);
r = fflush_sync_and_check(shadow);
if (r < 0)
return r;
@@ -616,7 +622,7 @@ static int write_temporary_group(const char *group_path, FILE **tmpfile, char **
group_changed = true;
}
r = fflush_and_check(group);
r = fflush_sync_and_check(group);
if (r < 0)
return r;
@@ -693,7 +699,7 @@ static int write_temporary_gshadow(const char * gshadow_path, FILE **tmpfile, ch
group_changed = true;
}
r = fflush_and_check(gshadow);
r = fflush_sync_and_check(gshadow);
if (r < 0)
return r;