test were suggested by Ian Taylor <ian@airs.com> and Joel did the
hard part of putting it in aclocal and editting all the offending
Makefiles and source code which could use this feature.
<ian@airs.com>:
The pc386 linker scripts omits .gnu.linkonce.r* sections. It's not a
big deal, but they should be treated like .rodata sections. ELF
versions of g++ generate them for static constants defined in template
classes, such as string::npos.
The pc386 linker scripts omits .gnu.linkonce.r* sections. It's not a
big deal, but they should be treated like .rodata sections. ELF
versions of g++ generate them for static constants defined in template
classes, such as string::npos.
I found that my 68040/68360 test programs would not run even after
I fixed the `wrong BSP' problem.
It seems that there's a bug in the interrupt handling code for
processors with hardware interrupt stacks (e.g. 68040). The wrong
status register was getting pushed on the stack for the `return
from exception' to call _ISR__Dispatch. This ended up making
the context switch code run on the interrupt stack, so interrupt-driven
context switches would always fail.
I guess that no one has tried running any of the RTEMS-4.0 snapshots
on a 68040 machine!
Anyhow, here are the patches for
1) gen68360.cfg --- to fix the `wrong-BSP' problem.
2) m68k/cpu_asm.s --- to fix the hardware interrupt stack problem.
With these patches in place, the network demo programs run on my
68040/68360 system. The paranoia program runs with no failures,
defects nor flaws.
Please find enclosed a patch which enables me to build the bare-bsp for
sh-rtems.
Changes:
1. Add preinstall to libbsp/bare/include/Makefile.in
2. Removed FORCEIT, add preinstall to
libbsp/sh/gensh1/include/Makefile.in
3. Disabled support of set_vector from sh code (shared/setvec.c is still
present but isn't used anymore), set_vector replaced with standard rtems
functions.
Problems still present:
1. Support of spin-delays in bare bsp
2. Proper support of cpu frequency
To configure I used:
<srcdir>/configure \
--target=sh-rtems \
--prefix=<instdir>/sh-bare \
--enable-bare-cpu-model=sh7032 \
--enable-bare-cpu-cflags='-Wall -m1 -DMHZ=20
-DCPU_CONSOLE_DEVNAME="\"/dev/null\""'
--enable-rtemsbsp=bare \
--disable-networking \
--disable-cxx \
--disable-posix \
--disable-tests
IMO, if there are no objections to this patch, a similar approach should
be applied to all CPUs/BSPs (esp. hppa1.1, mips64orion, ppc403, because
they apply set_vector inside of libcpu).
Remember the test to see if a socket could be read and written at
the same time by two different tasks? I discovered that if both
tasks attempt to close the socket a panic can occur from inside the
BSD code.
Closing the same socket twice from two different threads is
certainly an error, but a panic is not the greatest error reporting
method :-)
The following small change to the socket close routine should reduce
the chances of the panic.
Please find attached a start.s that includes a cli prior to the hlt
instruction. This ensures that external interrupts cannot restart
the system after returning to the startup code. ( According to the hlt
docs, they will! )
Also find a new timer.c. ( I forgot to update the countdowm value
in the timer when I changed the PSCLK frequency in start.s) . This
improves timer accuracy.
The raw_idt_notify messages are no longer infinite, I tested sp11 and
sp05, both which were bad, and I have seen the message print once in
one test. I think it's ok if it prints out once. In fact, I don't
think you can effectively stop it!
Please find attached the two files that have been changed relative to
980921 . The changes here are in the handling of the counter-timer used
as the basis for the rtems executive clock. For the most part, these
are housekeeping changes.
The PSCLK frequency change in start.s... was a part of several
bug-fixes. The fix improves executive clock and timer accuracy.
changes :
start.s -- All timers are disabled by the initialization routine
-- PSCLK ( used by clock and timers ) frequency changed to 1MHz
The clock_initialize routine now assumes that the PSCLK frequency is
exactly 1 MHz.
ckinit.c
Clock_isr -- removed division by 1000. Now use 'static'
variable -- clock_intial_isr_value -- to reset Clock_isrs variable.
clock_initialize -- moved counter timer initialization here. Values
used to configure the timer are totally dependent on
BSP_configuration.microseconds_per_tick ( and the PSCLK assumption).
Initializes clock_initial_isr_value used by th Clock_isr to reset
Clock_isrs.
clock_on -- no longer configures the timer, just enables it.
Since altering the number of sections in the BSP, I decided to give it a
good "once over" . The clock handling is now cleaner.
1. Finally fixes raw interrupts for pc386
2. Makes some minor cleanup in console and startup
3. Makes rtems_termios_dequeue_characters() to return count of
outstanding chars - it allows to simplify console isrs a little
bit.
4. pc386 uart modified to be friendlier to termios parameter changes,
to have minor performance improvement and to take advantage of
of above termios modification.
Here's a patch to make the rtems_showroute routine a little more
useful. For `host' route table entries the link-level address is now
displayed. This is equivalent to the old `show arp table'
information displayed by the KA9Q code.
Here are patches that bring 980911 back to what I think is a correct
version of raw IDT management as well as a correct initialisation
of video console and rtems managed interrupts.
I fixed the problems noted by Victor Vengerov.
1) Fix typo in cfsetispeed().
2) In rtems_termios_open, ensure that args->iop->data1 is set before calling
device-specific open routine.
I've fixed a few minor probs with the optimised version that Eric put
together for me the other day and sent the fixes back to him. Provided he
doesn't have a problem with it we've got a pretty solid in_cksum for the
ColdFire as well as straight m68k. I've enclosed my updated in_cksum_m68k.c
At the moment my own bottlenecks are elsewhere...as my driver is pulling
16bit data chunks through a libchip-esq access routine from the chip which
for a polled I/O device is never going to be quick.
I have managed to build the bsp ods68302 and the rtti test case I made
with egcs-1.1b and binutils-2.9.1.
I have built our C++ application and got no link errors so it looks like
this is now working. I am yet to test the code but getting the thing to
link was the problem.
Please find a patch attached which removes the -fno-rtti option.
in a previous patch. Comments from Eric:
patch_com2 was before your own fixes of raw IDT vector management
routine.
It mainly use human readable constants for the serial line driver code
+ the patch you gave me for com2 write routine.
I applied your patch for raw IDT mngt on top of that. Then made minor
cleaning in patch_idt_mngt. I hope this helps.
After some good comments from Eric Norum [thanks, Eric !],
I have added some modifications to my previous driver patch :
- wait for transmitter ready before sending a packet,
- new delay management in case of ring-overwritting.
Eric> NB : there is still a bug on PC386 serial line : exit does not
Eric> flush the remaining output queue. As this is not a bug in the
Eric> driver itself but somewhere in PC386 initialization/termios
Eric> relationship it will be part of another patch.
Eric> NB2 : As Emmanuel excerced the exception hanlder code, while
Eric> porting the SMC driver to the new BSD stack, we found a bug
Eric> in the exception handler : it shall not delete the current
Eric> thread in case we are running at interrupt level. This will
Eric> be part of another patch...
So here is the patch. This patch fixes the two problems mentionned above
+ it use vpath mechanism intead of copying the irq related files in
the right directory. This avoid to compile them each time and is
more homogenous with other Makefiles.
Here is a brief description of the attached patch :
1) There was a bug in the code for the COM2 serial line driver. Aleksey
gave me a fix that fixes the driver code itself. I would like to thank
him again publicly,
2) I introduced constants in the serial driver code because I had a hard
time reading the meanning of hexadecimal values in the NS data book :)
3)You can now mix printk and printf on serial line (tested on COM2).
There is a #ifdef PRINTK_ON_SERIAL in console.c that enables to have
printk on console while printf on serial line,
4) Removed call to displayCpuInfo because anyway if was at the wrong
place for serial line console (too early). It can anyway be called at
application level,
5) The original printk was unable to display negative integer values
and was also recursive. It now works corectly,
All the modifications have been tested here on the COM2 port from
a PC running RTEMS to a PC running linux,
NB : there is still a bug on PC386 serial line : exit does not flush the
remaining output queue. As this is not a bug in the driver itself but
somewhere in PC386 initialization/termios relationship it will be part
of another patch.
NB2 : As Emmanuel excerced the exception hanlder code, while porting the
SMC driver to the new BSD stack, we found a bug in the exception
handler : it shall not delete the current thread in case we are running
at interrupt level. This will be part of another patch...
