Re: Re: [PATCH] maintenance: fix two memory leaks
From: <hidden>
Date: 2021-05-10 07:50:21
quoted hunk
"lilinchao@oschina.cn" [off-list ref] writes:quoted
quoted
@@ -1999,6 +2000,7 @@ static int update_background_schedule(int enable)die("unknown background scheduler: %s", scheduler); rollback_lock_file(&lk); + free(lock_path);Based on your change, I think when "hold_lock_file_for_update()<0", we should also free local_path ThanksMeaning something like this? builtin/gc.c | 9 +++++++-- 1 file changed, 7 insertions(+), 2 deletions(-)diff --git c/builtin/gc.c w/builtin/gc.c index 98a803196b..50565c37c7 100644 --- c/builtin/gc.c +++ w/builtin/gc.c@@ -1971,8 +1971,10 @@ static int update_background_schedule(int enable)cmd = sep + 1; } - if (hold_lock_file_for_update(&lk, lock_path, LOCK_NO_DEREF) < 0) - return error(_("another process is scheduling background maintenance")); + if (hold_lock_file_for_update(&lk, lock_path, LOCK_NO_DEREF) < 0) { + result = error(_("another process is scheduling background maintenance")); + goto cleanup; + } if (!strcmp(scheduler, "launchctl")) result = launchctl_update_schedule(enable, get_lock_file_fd(&lk), cmd);@@ -1984,6 +1986,9 @@ static int update_background_schedule(int enable)die("unknown background scheduler: %s", scheduler); rollback_lock_file(&lk); + +cleanup: + free(lock_path); free(testing); return result; }
Yes, it's almost like this, and your is better than I thought.
I just referred to this function(L1266-L1321):
static int maintenance_run_tasks(struct maintenance_run_opts *opts)
{
...
if (hold_lock_file_for_update(&lk, lock_path, LOCK_NO_DEREF) < 0) {
...
free(lock_path);
return 0;
}
free(lock_path);
...
...
rollback_lock_file(&lk);
return result;
}
and thought we should also apply it to "update_background_schedule()".