[linuxsh-dev] KGDB for SH; SCI/F fixes

[Jeremy S. <js...@mv...>]

A couple months ago I sent an email about moving the gdb stub from sh-ipl
into the kernel for a SW-only kgdb, and heard back from Henry Bell at ST
Micro who was working on the same thing -- we've since exchanged some code
and I've played around with it some more; the resulting patch is
attached.  I'd like to add this to the CVS tree in hopes that it will help
to have some sort of KGDB available for developers w/o sh-ipl+g or other
debug aids; any objections?  (It is clearly far from perfect, and is not
exhaustively tested, but I've been using it or its cousins for a while and
it doesn't seem to break anything too horribly.)

Credit for most of this patch goes to Henry; his patch seemed more
complete than mine (included threads, sysrq, and gdb-console support) so I
went with his... but of course everything is based on existing copies of
the stub (such as sh-ipl+g) so thanks to everyone else too!

Current major defect: not set up to use ^C or break to get into the
running system; if you don't request a startup halt to set breakpoints,
you need to use sysrq-g or NMI to get in.  (Other aspects of the serial
port handling could use serious improvement as well.)

While trying this stuff, a couple problems with SCIF handling popped up
(esp. w.r.t. break on SH3); this patch includes code which attempt to
address these.  Comments on those would be appreciated, since they are NOT
under KGDB configuration flags.  In brief:

1) Flow-control.  The sci_throttle() function turns off Rx interrupts, and
nothing turns them on again... added an interrupt enable line in
sci_unthrottle(), which seemed appropriate?
2) On the SH3 (at least the SH7709A) a <BREAK> caused repeated NUL/FE and
BRK interrupts, resulting in a vareity of problems from hanging the port
to hanging the kernel [in continuous interrupts].  The best results seemed
to be adding a break "hold-down": a flag set on the first BRK interrupt,
which causes subsequent BRK an NUL/FE events to be ignored until the first
non-NUL.  (That char is also discarded if it's 1's then 0's, e.g. 0xF8 or
0xC0 -- the space-to-mark transition ending the break, interpreted as a
char.)  It's still occasionally flaky: misses a break, gets one extra, or
something -- but seems a lot better.  I put it under __SH3__  flags,
assuming this is common to all SH3 SCI/Fs, though I only have a 7709A and
so couldn't test any others.

Thanks!
  --Jeremy Siegel