Junio C Hamano <gitster@xxxxxxxxx> writes: > Junio C Hamano <gitster@xxxxxxxxx> writes: > >> Gaah, of course. >> >> This is coming from the cache preload codepath, where multiple threads >> try to run ce_path_match(). >> It used to be OK because pathspec magic never looked at attributes, >> but now it does, and attribute system is not thread-safe. > > The symlink check code has dealt with a similar issue in the past in > the codepath. threaded_has_symlink_leading_path() is called with > per-thread data structure called "cache", because the leading > symbolic link check uses a cache that is shared across different > invocations of the check. > > We need something similar to duplicate the attribute stack per > thread. git_check_attrs() currently uses the singleton instance of > attr_stack chain, that is rooted at the file scope global > "attr_stack", and no wonder that would be clobbered when multiple > threads try to use it. The result of attribute collection is also > accumulated in a file scope global, which should probably be moved > to the "struct git_attr_check" introduced in the jc/attr topic. > > We need to teach the callchain that includes prepare_attr_stack() > and bootstrap_attr_stack() to take a pointer to the attr_stack root, > give git_check_attr_threaded() that takes such root so that threaded > code can use per-thread attr stack, and for non-threaded > applications use &the_default_attr just like the file scope global > "default_cache" is used in symlinks.c. Then a threaded attribute > lookup can maintain its own attr_stack when running more than one > instance of lookup. Another is the global attr dictionary. It is like the global object hash, so the look-up and insertion into it need to be protected the same way with mutex, just like builtin/pack-objects.c serializes the object store access with locks. -- To unsubscribe from this list: send the line "unsubscribe git" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html