Re: Recent problems with -current on alpha...

On Thu, 26 Feb 2004, Ken Smith wrote:

> With lots of help from Andrew Gallatin and Marcel Moolenaar I think I've
> located the problem with -current on alpha and I have a very rough
> patch that gets the cluster's Alpha reference machine to the point
> it boots and runs.  In a nutshell the promcons stuff used to allow
> debugging info to be printed to the console from a very early point
> in the boot doesn't work any more.  Things like make_dev() require
> the machine to have progressed to the point that more of the VM
> system is functional than before (this seems to have begun with the
> device megapatches).

make_dev() doesn't need vm for the first DEVT_STASH (default 50)
allocations.  syscons console initialization depends on this.  The
unlogged reason that rev.1.412 of syscons.c didn't work is that it was
committed before make_dev() could handle it, so the system crashed
early iff syscons was the console.  Perhaps the promcons initialization
is even earlier than i386 console initialization.  I'm not familiar
with it, but it is hard for it to be much earlier -- i386 console
initialization is done in the first C function called from locore.

Another thing that can cause problems is using printf() before the
console has been initialized.  I hade a printf() in make_dev().  When
syscons started calling make_dev() early, this printf() caused much
the same crash as rev.1.412 of syscons.c.  Hmm, it was actually the
same crash.  Both crashes were cause by a printf() in make_dev(), and
both were fixed by removing the printf().  In -current, the removal
was in rev.1.139 of kern_conf.c.

> My patch moves the initialization of the promcons stuff until after
> the VM system is initialzed and comments out all the printf()'s I
> could find that might have happened before the console is initialized.

This would break use of ddb until after the vm system is initialized
(unless it is already broken).  i386 console initialization tries hard
to initiatialize the console as early as possible so that it can be
used for debugging.  This is less needed with gdb, but early printfs
still need a console.

The make_dev() line in promcons.c was last changed in Y2K, so it console
initialization is apparently already done late for alphas (which is OK
iff the firmware console is used early).  The problem might be simply a
stray printf done early.  Try avoiding such printfs, and this works then
debug why early printfs cause crashes.  They should just go to a buffer
that gets printed later, or at least get counter so that they can be
warned about later.  The buffer can't be the message buffer, since the
message buffer is not initialized until later, at least on i386's.
printf already takes care not to use a nonexistent message buffer but
doesn't take care to warn about dropped messages.

Bruce