Fix amd64->i386 linux syscall restart problem

Fix amd64->i386 linux syscall restart problem

This commit fixes some failures in gdb.base/interrupt.exp
when debugging a 32-bit i386 linux inferior from an amd64 host.

When running the following test...

  make check RUNTESTFLAGS="--target_board unix/-m32 interrupt.exp"

... without this commit, I see the following output:

FAIL: gdb.base/interrupt.exp: continue (the program exited)
FAIL: gdb.base/interrupt.exp: echo data
FAIL: gdb.base/interrupt.exp: Send Control-C, second time
FAIL: gdb.base/interrupt.exp: signal SIGINT (the program is no longer running)
ERROR: Undefined command "".
ERROR: GDB process no longer exists

		=== gdb Summary ===

When the test is run with this commit in place, we see 12 passes
instead.  This is the desired behavior.

Analysis:

On Linux, when a syscall is interrupted by a signal, the syscall
may return -ERESTARTSYS when a signal occurs.  Doing so indicates that
the syscall is restartable.  Then, depending on settings associated
with the signal handler, and after the signal handler is called, the
kernel can then either return -EINTR or can cause the syscall to be
restarted.  In this discussion, we are concerned with the latter
case.

On i386, the kernel returns this status via the EAX register.

When debugging a 32-bit (i386) process from a 64-bit (amd64)
GDB, the debugger fetches 64-bit registers even though the
process being debugged is 32-bit.  Since we're debugging a 32-bit
target, only 32 bits are being saved in the register cache.
Now, ideally, GDB would save all 64-bits in the regcache and
then would be able to restore those same values when it comes
time to continue the target.  I've looked into doing this, but
it's not easy and I don't see many benefits to doing so.  One
benefit, however, would be that EAX would appear as a negative
value for doing syscall restarts.

At the moment, GDB is setting the high 32 bits of RAX (and other
registers too) to 0.  So, when GDB restores EAX just prior to
a syscall restart, the high 32 bits of RAX are zeroed, thus making
it look like a positive value.  For this particular purpose, we
need to sign extend EAX so that RAX will appear as a negative
value when EAX is set to -ERESTARTSYS.  This in turn will cause
the signal handling code in the kernel to recognize -ERESTARTSYS
which will in turn cause the syscall to be restarted.

This commit is based on work by Jan Kratochvil from 2009:

https://sourceware.org/ml/gdb-patches/2009-11/msg00592.html

Jan's patch had the sign extension code in amd64-nat.c.  Several
other native targets make use of this code, so it seemed better
to move the sign extension code to a linux specific file.  I
also added similar code to gdbserver.

Another approach is to fix the problem in the kernel.  Hui Zhu
tried to get a fix into the kernel back in 2014, but it was not
accepted.  Discussion regarding this approach may be found here:

https://lore.kernel.org/patchwork/patch/457841/

Even if a fix were to be put into the kernel, we'd still need
some kind of fix in GDB in order to support older kernels.

Finally, I'll note that Fedora has been carrying a similar patch for
at least nine years.  Other distributions, including RHEL and CentOS
have picked up this change and have been using it too.

gdb/ChangeLog:

	* amd64-linux-nat.c (amd64_linux_collect_native_gregset): New
	function.
	(fill_gregset): Call amd64_linux_collect_native_gregset instead
	of amd64_collect_native_gregset.
	(amd64_linux_nat_target::store_registers): Likewise.

gdb/gdbserver/ChangeLog:

	* linux-x86-low.c (x86_fill_gregset): Sign extend EAX value
	when using a 64-bit gdbserver.