Re: [HACKERS] Compiling 6.4 on NetBSD-current/pc532

> >
> > Applied.
> >
> > > OK,
> > >
> > > I have a current version of PostgreSQL running on my pc532 now.
> > > (It's a NS32K based machine. Somewhat of an antique really...)
> > >
> > >
> > > *** /usr/local/pgsql/src/include/storage/s_lock.h Fri Sep 11 19:00:55 1998
> > > --- s_lock.h Sat Sep 12 00:27:51 1998
> > > ***************
> > > *** 213,219 ****
> > > #endif /* NEED_I386_TAS_ASM */
> > >
> > >
> > > ! /* NS32K code is in s_lock.c */
> > >
> > > #endif /* defined(__GNUC__) */
> > >
> > > --- 213,234 ----
> > > #endif /* NEED_I386_TAS_ASM */
> > >
> > >
> > > !
> > > ! #if defined(NEED_NS32K_TAS_ASM)
> > > !
> > > ! #define S_LOCK(lock) \
> > > ! { \
> > > ! slock_t res = 1; \
> > > ! while (res) { \
> > > ! __asm__("movqd 0, r0"); \
> > > ! __asm__("sbitd r0, %0" : "=m"(*lock)); \
> > > ! __asm__("sprb us, %0" : "=r" (res)); \
> > > ! res = ((res >> 5) & 1); \
> > > ! } \
> > > ! }
> > > !
> > > ! #endif /* NEED_NS32K_TAS_ASM */
> > > !
> > >
> > > #endif /* defined(__GNUC__) */
> > >
> > > *** /usr/local/pgsql/src/backend/storage/buffer/s_lock.c Thu Sep 10 23:08:00 1998
> > > --- s_lock.c Sat Sep 12 00:23:04 1998
> > > ***************
> > > *** 118,134 ****
> > > #endif /* PPC */
> > >
> > >
> > > - #if defined(__ns32k__)
> > > - int
> > > - tas(volatile slock_t *lock)
> > > - {
> > > - int res;
> > > - __asm__("sbitb 0, %0" : "=m"(*lock));
> > > - __asm__("sprb us, %0" : "=r"(res));
> > > - res = (res >> 5) & 1;
> > > - return res;
> > > - }
> > > - #endif
> > >
> > > #else /* defined(__GNUC__) */
> > > /***************************************************************************
> > > --- 118,123 ----
> > >
> > >
> > > BTW, does the spinlock code need a TAS function so it can spin for
> > > a while and then declare itself stuck, or does a second process/thread
> > > take care of that. It would be simpler for the NS32K to just make
> > > the whole lock function be 2 lines of inline assembler, but that
> > > would contain an infinite loop if the lock was stuck...
> > >
> > > Jon Buller <jonb(at)metronet(dot)com>
>
> I wish I had noticed this before Bruce applied it.
>
> The TAS function is needed so that stuck spinlocks can be recovered from.
> Also, it enables the pseudo random back off which helps performance when
> there are many backends.
>
> In any case, this patch does not "follow the one true path" that I tried
> to outline in s_lock.c and s_lock.h. In fact it is exactly backwards.
>
> Basically the preferred way is:
>
> - in s_lock.h do nothing, the defaults should take care of you.
>
> -in s_lock.c define a TAS function that sets the spinlock and returns the
> previous state of the lock.
>
> I see from your asm()s that you are using gcc. In this case, your TAS function
> should be called tas(), and should be defined inside the __GNUC__ section.

The patch had been on the list for a while, so I figured it was safe.
I will back it out, and the user will have to study what is already
there and resubmit a new patch.

I belive the problem is that 6.3 users are trying to get their platforms
into the new locking code, and naturally using their old code to do it.

--
Bruce Momjian | 830 Blythe Avenue
maillist(at)candle(dot)pha(dot)pa(dot)us | Drexel Hill, Pennsylvania 19026
http://www.op.net/~candle | (610) 353-9879(w)
+ If your life is a hard drive, | (610) 853-3000(h)
+ Christ can be your backup. |