port-arm/55791: software breakpoint skipped on ARMv6

[rokuyama.rk%gmail.com@localhost]

>Number:         55791
>Category:       port-arm
>Synopsis:       software breakpoint skipped on ARMv6
>Confidential:   no
>Severity:       serious
>Priority:       medium
>Responsible:    port-arm-maintainer
>State:          open
>Class:          sw-bug
>Submitter-Id:   net
>Arrival-Date:   Sat Nov 07 00:15:00 +0000 2020
>Originator:     Rin Okuyama
>Release:        9.99.75
>Organization:
Department of Physics, Meiji University
>Environment:
NetBSD rpi0w 9.99.75 NetBSD 9.99.75 (RPI) #68: Sat Nov  7 07:24:01 JST 2020  rin@latipes:/sys/arch/evbarm/compile/RPI evbarm earmv6hf
>Description:
At least on ARM1176 in RPI[01], software breakpoints inserted by GDB in
shared libraries are skipped with a high probability:

----
$ gdb echo
GNU gdb (GDB) 8.3
...
Reading symbols from echo...
Reading symbols from /usr/libdata/debug//bin/echo.debug...
(gdb) b main
Breakpoint 1 at 0x950: file /usr/src/bin/echo/echo.c, line 58.
(gdb) r
Starting program: /bin/echo

Breakpoint 1, main (argc=1, argv=0x7feff32c) at /usr/src/bin/echo/echo.c:58
58      /usr/src/bin/echo/echo.c: No such file or directory.
(gdb) s

[Inferior 1 (process 2354) exited normally]
(gdb)
----

This is odd since this userland is built with MKDEBUGLIB=yes. On the other
hand, on Cortex-A53 in RPI2 rev1.2, breakpoints seem to work as expected:

----
$ gdb echo
GNU gdb (GDB) 8.3
...
Reading symbols from echo...
Reading symbols from /usr/libdata/debug//bin/echo.debug...
(gdb) b main
Breakpoint 1 at 0x950: file /usr/src/bin/echo/echo.c, line 58.
(gdb) r
Starting program: /bin/echo

Breakpoint 1, main (argc=1, argv=0x7ff19804) at /usr/src/bin/echo/echo.c:58
58      /usr/src/bin/echo/echo.c: No such file or directory.
(gdb) s
setprogname (progname=0x7ff19cc0 "/bin/echo")
    at /usr/src/lib/libc/gen/setprogname.c:65
65      /usr/src/lib/libc/gen/setprogname.c: No such file or directory.
(gdb)
main (argc=<optimized out>, argv=0x7ff19804) at /usr/src/bin/echo/echo.c:59
59      /usr/src/bin/echo/echo.c: No such file or directory.
(gdb)
__setlocale50 (category=0, locale=0xc110adc "")
    at /usr/src/lib/libc/locale/setlocale.c:193
193     /usr/src/lib/libc/locale/setlocale.c: No such file or directory.
(gdb)
194     in /usr/src/lib/libc/locale/setlocale.c
(gdb)
__setlocale (category=0, name=0xc110adc "")
    at /usr/src/lib/libc/locale/setlocale.c:175
175     in /usr/src/lib/libc/locale/setlocale.c
...
----

Even on ARM1176, one or two breakpoints sometimes work. Therefore,
I suspect this is due to pmap_procwr():

	https://nxr.netbsd.org/xref/src/sys/arch/arm/arm32/pmap.c#4793

4793 void
4794 pmap_procwr(struct proc *p, vaddr_t va, int len)
4795 {
4796 #ifndef ARM_MMU_EXTENDED
4797 
4798 	/* We only need to do anything if it is the current process. */
4799 	if (p == curproc)
4800 		cpu_icache_sync_range(va, len);
4801 #endif
4802 }

When GDB inserts a software breakpoint, p != curproc. So, we may need to
sync icache even for p != curproc (ARM_MMU_EXTENDED is not defined for
ARM1176). Note that a similar fix had been applied to aarch64:

	http://cvsweb.netbsd.org/bsdweb.cgi/src/sys/arch/aarch64/aarch64/pmap.c#rev1.82
>How-To-Repeat:
Described above; ``gdb echo'' with userland built with MKDEBUGLIB=yes on
Raspberry Pi 1 or Zero.
>Fix:
For !ARM_MMU_EXTENDED case at least, sync icache even for p != curproc.
However, unfortunately, I don't know realizations for arm at the moment...