tech-userlevel archive

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index][Old Index]

Re: Patch to make <stdio.h> reentrant by default

On 26.04.2019 23:36, Kamil Rytarowski wrote:
> <stdio.h> traditionally (at least in the BSD tradition) contained
> optimization for !_REENTRANT and !threaded code in <stdio.h>.
> The header was shadowing real functions such as fileno(3) or getc(3)
> with inlined non-reentrant ones.
> This caused issues with C++ code especially C++11 (notably std::thread)
> and was explicitly disabled by Joerg back years ago.
> Non-reentrant inlined functions also caused damage to sanitizers (and
> other related software) that link transparently with -lpthread without
> defining _REENTRANT in preprocessor.
> This caused mixing threaded and !reentrant code in sanitizers and caused
> breakage in programs using these functions with double variations. The
> most sensitive sanitizer to this is MSan.
> As a workaround we have added in the clang's NetBSD driver explicit
> -D_REENTRANT for all sanitizers to work around the problem on NetBSD and
> pick threading safe functions. The same workaround hasn't been
> propagated to GCC so far.
> We keep detecting that more software is happy to just pick -lpthread
> (like LLVM OpenMP) and prebuilt software works by an accident.
> Catching every -lpthread user worldwide and teaching about -pthread is a
> losing battle. My current observation is that people (in toolchains!)
> are surprised that they need to link with -lpthread for C++ threads.
> I propose a patch that:
> 1. Removes alternative versions of <stdio.h> depending on preprocessor
> namespace definitions. All code since now is assumed to be
> reentrant/threading safe.
> 2. Adds explicit non-reentrant versions of fileno(), ferror(3), feof(3)
> and clearerr(3) called fileno_unlocked(), ferror_unlocked(),
> feof_unlocked(), clearerr_unlocked().
> These routines have users in GNU tools and are implemented in at least
> macOS, Linux and FreeBSD. They are also documented in the GNU world as
> BSD extensions.
> New unlocked() functions are deliberately implemented as libc symbols in
> order to pick them easily from libc and don't break existing software
> calling code from pregenerated configure scripts, such as "#define
> fileno_unlocked fileno" in base's GCC. (With next upgrade of GCC I
> expect GCC to switch to native _unlocked() variations).
> Implementing these routines as light functions inside libc is also
> sanitizer friendly as inlined code that interacts directly with stdio.h
> buffers is more difficult to sanitize. Even if we would want to sanitize
> such scenarios we still want _REENTRANT functions only as the LLVM
> runtimes of sanitizers are threaded.

If some software uses explicitly _unlocked() routine as a libc symbol
(and not inlined code operating on global buffers), sanitizing it is
straightforward without [almost] any extra work. Keeping it as sanitized
as inlined code that operates on global <stdio.h> buffers is too
difficult for too little (/no?) gain to make it worth the effort.

> 3. Not introduces any extra _unlocked() functions, even if they are
> available in other systems, as they weren't available in the original
> NetBSD stdio.h magic macros.
> 4. Adapts surrounding code, documentation, etc.
> 5. Not changes the alternative semantics of _REENTRANT in other headers,
> as there this symbol is used as a feature detection for _r() functions
> (not sure if there are any users of this semantics, but I'm not changing
> this).
> 6. Defines h_errno in <netdb.h> in a reentrant way always.
> 7. Silently drops dead feature check _PTHREAD. It is still generated by
> GCC for -pthread, but it was never worth to get into Clang/LLVM and has
> 0 users in worlwide open-source software (according to Debian Code search).
> There wre recurring internal discussions about this with at least Joerg,
> Christos and Michal (Gorny).
> If we want to optimize code for 1 thread, we can now pick _unlocked()
> functions explicitly in the basesystem software.
> This is something I would like to see in -9.
> All ATF tests pass with the proposed patch.

Attachment: signature.asc
Description: OpenPGP digital signature

Home | Main Index | Thread Index | Old Index