mirror of
https://github.com/neovim/neovim.git
synced 2025-09-24 12:08:33 +00:00
fix(autocmds): once=true Lua event-handler may call itself #29544
Problem: Event handler declared with `once=true` can re-trigger itself (i.e. more than once!) by calling `nvim_exec_autocmds` or `:doautocmd`. Analysis: This happens because the callback is executed before deletion/cleanup (`aucmd_del`). And calling `aucmd_del` before `call_autocmd_callback` breaks the autocmd execution... Solution: Set `ac->pat=NULL` to temporarily "delete" the autocmd, then restore it after executing the callback. Fix #25526 Co-authored-by: Justin M. Keyes <justinkz@gmail.com>
This commit is contained in:
@@ -2025,7 +2025,8 @@ static void aucmd_next(AutoPatCmd *apc)
|
||||
apc->auidx = SIZE_MAX;
|
||||
}
|
||||
|
||||
static bool call_autocmd_callback(const AutoCmd *ac, const AutoPatCmd *apc)
|
||||
/// Executes an autocmd callback function (as opposed to an Ex command).
|
||||
static bool au_callback(const AutoCmd *ac, const AutoPatCmd *apc)
|
||||
{
|
||||
Callback callback = ac->exec.callable.cb;
|
||||
if (callback.type == kCallbackLua) {
|
||||
@@ -2106,16 +2107,24 @@ char *getnextac(int c, void *cookie, int indent, bool do_concat)
|
||||
apc->script_ctx = current_sctx;
|
||||
|
||||
char *retval;
|
||||
if (ac->exec.type == CALLABLE_CB) {
|
||||
// Can potentially reallocate kvec_t data and invalidate the ac pointer
|
||||
if (call_autocmd_callback(ac, apc)) {
|
||||
// If an autocommand callback returns true, delete the autocommand
|
||||
oneshot = true;
|
||||
switch (ac->exec.type) {
|
||||
case CALLABLE_EX:
|
||||
retval = xstrdup(ac->exec.callable.cmd);
|
||||
break;
|
||||
case CALLABLE_CB: {
|
||||
AutoCmd ac_copy = *ac;
|
||||
// Mark oneshot handler as "removed" now, to prevent recursion by e.g. `:doautocmd`. #25526
|
||||
ac->pat = oneshot ? NULL : ac->pat;
|
||||
// May reallocate `acs` kvec_t data and invalidate the `ac` pointer.
|
||||
bool rv = au_callback(&ac_copy, apc);
|
||||
if (oneshot) {
|
||||
// Restore `pat`. Use `acs` because `ac` may have been invalidated by the callback.
|
||||
kv_A(*acs, apc->auidx).pat = ac_copy.pat;
|
||||
}
|
||||
// If an autocommand callback returns true, delete the autocommand
|
||||
oneshot = oneshot || rv;
|
||||
|
||||
// TODO(tjdevries):
|
||||
//
|
||||
// Major Hack Alert:
|
||||
// HACK(tjdevries):
|
||||
// We just return "not-null" and continue going.
|
||||
// This would be a good candidate for a refactor. You would need to refactor:
|
||||
// 1. do_cmdline to accept something besides a string
|
||||
@@ -2124,8 +2133,11 @@ char *getnextac(int c, void *cookie, int indent, bool do_concat)
|
||||
// and instead we loop over all the matches and just execute one-by-one.
|
||||
// However, my expectation would be that could be expensive.
|
||||
retval = xcalloc(1, 1);
|
||||
} else {
|
||||
retval = xstrdup(ac->exec.callable.cmd);
|
||||
break;
|
||||
}
|
||||
case CALLABLE_NONE:
|
||||
default:
|
||||
abort();
|
||||
}
|
||||
|
||||
// Remove one-shot ("once") autocmd in anticipation of its execution.
|
||||
|
Reference in New Issue
Block a user