Fix #81142: memory leak when unserialize()ing associative array #7160
Conversation
We should not internalize unserialized array keys, because that easily wastes a lot of memory for a small gain. Instead, we check whether the string is already internalized, and use that; otherwise we use a new string. This patch is based on an idea by @nikic.
| @@ -947,7 +947,13 @@ use_double: | |||
| } else if (len == 1) { | |||
| ZVAL_INTERNED_STR(rval, ZSTR_CHAR((zend_uchar)*str)); | |||
| } else if (as_key) { | |||
| ZVAL_STR(rval, zend_string_init_interned(str, len, 0)); | |||
| zend_string *key = zend_string_init(str, len, 0); | |||
| zend_string *ikey = zend_interned_string_find_permanent(key); | |||
There was a problem hiding this comment.
This will only work for strings that are created during startup, which is probably not terribly useful. We'd be mainly interested in strings that are either interned per-request or part of opcache. I think this would require adding a new hook for this purpose that can be replaced by opcache.
There was a problem hiding this comment.
I don't think this ought to use interned strings at all.
It looks simpler to add a table to unserialize data for keys.
There was a problem hiding this comment.
I think the idea here is specifically to reuse normal interned strings for things like object property names, which are usually available as interned strings.
Maybe @dstogov can comment.
There was a problem hiding this comment.
With opcache zend_string_init_interned() doesn't create new interned strings (just search for existing ones).
To implement the same behavior without opcache, we need to introduce a new callback in zend_string API.
There was a problem hiding this comment.
I don't think that we want to introduce a new callback in the stable PHP versions. So, postpone the fix to PHP 8.1?
|
Sorry, I don't have the time to work on this. |
We should not internalize unserialized array keys, because that easily
wastes a lot of memory for a small gain. Instead, we check whether the
string is already internalized, and use that; otherwise we use a new
string.
This patch is based on an idea by @nikic.