Re: [PATCH v2 2/2] hex: make hash_to_hex_algop() and friends thread-safe

From: Matheus Tavares Bernardino <matheus.bernardino@usp.br>
Date: Jun 30, 2020, 20:37 UTC
Message-ID: <CAHd-oW6U9krWYd++5ipGA6e4Xy6ZoBxE+QVFpshrjUKjN+RH0g@mail.gmail.com>
In-Reply-To: <nycvar.QRO.7.76.6.2006291646420.56@tvgsbejvaqbjf.bet>

On Tue, Jun 30, 2020 at 12:01 PM Johannes Schindelin <Johannes.Schindelin@gmx.de> wrote:

Show 15 quoted lines

>
> Hi Matheus,
>
> I am fine with the Windows changes (although I have to admit that I did
> not find time to test things yet).
>
> There is one problem in that I do not necessarily know that the memory is
> released correctly when threads end; You will notice that the
> `pthread_key_create()` shim in `compat/win32/pthread.h` does not use the
> `destructor` parameter at all. The documentation at
>
>       https://docs.microsoft.com/en-us/windows/win32/procthread/using-thread-local-storage
>
> is also not terribly clear _how_ the memory is released that was assigned
> via `TlsSetValue()`.

Yes, I wasn't sure about that either... It would be great if someone familiar with TLS memory on Windows could help us with this.

> A couple more things:
>
> On Fri, 26 Jun 2020, Matheus Tavares wrote:
> >

[...]

Show 5 quoted lines

> > +struct hexbuf_array {
> > +     int idx;
>
> Is there a specific reason why you renamed `bufno` to `idx`? If not, I'd
> rather keep the old name.

OK, I'll change it back in v3.

Show 62 quoted lines

> > +     char bufs[4][GIT_MAX_HEXSZ + 1];
> > +};
> > +
> > +#ifdef HAVE_THREADS
> > +static pthread_key_t hexbuf_array_key;
> > +static pthread_once_t hexbuf_array_once = PTHREAD_ONCE_INIT;
> > +
> > +static void init_hexbuf_array_key(void)
> > +{
> > +     if (pthread_key_create(&hexbuf_array_key, free))
> > +             die(_("failed to initialize threads' key for hash to hex conversion"));
> > +}
> > +
> > +#else
> > +static struct hexbuf_array default_hexbuf_array;
> > +#endif
> > +
> >  char *hash_to_hex_algop(const unsigned char *hash, const struct git_hash_algo *algop)
> >  {
> > -     static int bufno;
> > -     static char hexbuffer[4][GIT_MAX_HEXSZ + 1];
> > -     bufno = (bufno + 1) % ARRAY_SIZE(hexbuffer);
> > -     return hash_to_hex_algop_r(hexbuffer[bufno], hash, algop);
> > +     struct hexbuf_array *ha;
> > +
> > +#ifdef HAVE_THREADS
> > +     void *value;
> > +
> > +     if (pthread_once(&hexbuf_array_once, init_hexbuf_array_key))
> > +             die(_("failed to initialize threads' key for hash to hex conversion"));
> > +
> > +     value = pthread_getspecific(hexbuf_array_key);
> > +     if (value) {
> > +             ha = (struct hexbuf_array *) value;
> > +     } else {
> > +             ha = xmalloc(sizeof(*ha));
> > +             if (pthread_setspecific(hexbuf_array_key, (void *)ha))
> > +                     die(_("failed to set thread buffer for hash to hex conversion"));
> > +     }
> > +#else
> > +     ha = &default_hexbuf_array;
> > +#endif
>
> This introduces two ugly `#ifdef HAVE_THREADS` constructs which are
> problematic because they are the most likely places to introduce compile
> errors.
>
> I wonder whether you considered introducing a function (and probably a
> macro) that transparently gives you a thread-specific instance of a given
> data type? The caller would look something like
>
>         struct hexbuf_array *hex_array;
>
>         GET_THREADSPECIFIC(hex_array);
>
> where the macro would look somewhat like this:
>
>         #define GET_THREADSPECIFIC(var) \
>                 if (get_thread_specific(&var, sizeof(var)) < 0)
>                         die(_("Failed to get thread-specific %s"), #var);
>
> and the function would allocate and assign the variable.

Hmm, we would need two macros, wouldn't we? GET_THREADSPECIFIC(var) and a DECLARE_THREADSPECIFIC(var, destructor), to declare the pthread_once_t and pthread_key_t variables, as well as define a initialization function for the key (i.e. the callback to pthread_once()). Or we could provide these declarations ourselves, but then we would need the "ifdef HAVE_THREADS" guard to avoid calling pthread_key_create() when there is no multithreading.

I think that would work, yes. Alternatively, I think we could adjust the dummy pthread_key_* functions in thread-utils.h to emulate the real ones when HAVE_THREADS == false. Then we could eliminate the `#ifdef HAVE_THREADS` guards and use the same code for both cases here (and everywhere else pthread_key is used). I haven't thought about it carefully enough yet, but I think this *might* be as simple as adding the following defines in the "#ifdef NO_PTHREADS" section of thread-utils.h:

#define pthread_key_t void *
/*
 * The destructor is not used in this case as the main thread will only
 * exit when the program terminates.
 */
#define pthread_key_create(key_ptr, unused) return_0((*key_ptr) = NULL)
#define pthread_setspecific(key, value) return_0((key) = (value))
#define pthread_getspecific(key) (key)
#define pthread_key_delete(key) return_0(NULL)

static inline int return_0(void *unused)
{
        return 0;
}

That's the general idea, but we might as well define a `struct dummy_pthread_key_t` instead of defining the key directly as `void *` (and have functions instead of macros). This way we could store, e.g., an "initialized" flag that could be used to return an error code on double-initializations.

What do you think?