Re: [parisc-linux] Re: NPTL for hppa-linux is not backwards compatible with Linuxthreads.

On 2/22/07, Roland McGrath <[email protected]> wrote:
> > All statically initialized locks are broken. We made locks smaller,
> > and changed the value of the static initialization.
>
> Smaller?  Smaller is easy.  And you didn't actually reduce
> __SIZEOF_PTHREAD_MUTEX_T, did you?  This seems like it would not be at all
> hard to accommodate just with symbol versioning and not break any
> compatibility.  Off hand, it seems like it would be some simple code in one
> new file in your sysdeps/ tree and a smattering of sysdeps Versions files.

Thanks for your feedback Roland.

Yes, our locks are smaller, we went from a 4 word structure to a 1
word integer. I did not reduce __SIZEOF_PTHREAD_MUTEX_T. NPTL and
Linuxthreads pthread_mutex_t are the same size. Everything has been
kept the same size, but not in exactly the same position (probably a
bug on my part).

In Linuxthreads an initialized _pthread_fastlock is {{1,1,1,1},0} e.g.
{__spinlock, __status}

Old RW Locks
116 /* Read-write locks.  */
117 typedef struct _pthread_rwlock_t
118 {
119   struct _pthread_fastlock __rw_lock; /* Lock to guarantee mutual
exclusion */
120   int __rw_readers;                   /* Number of readers */
121   _pthread_descr __rw_writer;         /* Identity of writer, or
NULL if none */
122   _pthread_descr __rw_read_waiting;   /* Threads waiting for reading */
123   _pthread_descr __rw_write_waiting;  /* Threads waiting for writing */
124   int __rw_kind;                      /* Reader/Writer preference
selection */
125   int __rw_pshared;                   /* Shared between processes or not */
126 } pthread_rwlock_t;

New RW Locks
137 /* Data structure for read-write lock variable handling.  The
138    structure of the attribute type is not exposed on purpose.  */
139 typedef union
140 {
141   struct
142   {
143     int __lock;
144     unsigned int __nr_readers;
145     unsigned int __readers_wakeup;
146     unsigned int __writer_wakeup;
147     unsigned int __nr_readers_queued;
148     unsigned int __nr_writers_queued;
149     /* FLAGS must stay at this position in the structure to maintain
150        binary compatibility.  */
151     unsigned int __flags;
152     int __writer;
153   } __data;
154   char __size[__SIZEOF_PTHREAD_RWLOCK_T];
155   long int __align;
156 } pthread_rwlock_t;

In the new structure we have shifted everything up because __lock is
now an integer, instead of a _pthread_fastlock with a 4 word lock
structure. Should I add padding after "__lock" e.g. int pad[3]?

In an old executable the following static initializers:
PTHREAD_MUTEX_INITIALIZER, PTHREAD_COND_INTIALIZER,
PTHREAD_RWLOCK_INITIALIZER will setup the __spinlock structure to
{1,1,1,1} e.g. unlocked in our Linuxthreads implementation.

To be clear are you suggesting I write compat wrappers for the
pthread_mutex_*, pthread_cond_*, and pthread_rwlock_* functions to
detect old style initialized locks, and reinitialize the lock word?

> But if your community is happy with a soname change and not worrying about
> its various impacts, by all means just do that.  It's a lovely thing to be
> able to stop compiling in all the compat stuff.
>
> To do that, all you need is sysdeps/unix/sysv/linux/hppa/nptl/shlib-versions:
>
> hppa.*-.*-linux.*       libc=6.1                GLIBC_2.6
> hppa.*-.*-linux.*       libpthread=0.1          GLIBC_2.6

I would like to avoid breaking backwards compatibility, but I don't
quite understand all the issues. Your feedback is much appreciated.

Cheers,
Carlos.