[Qemu-devel] [PATCH] gdbstub improvements for gdb kill/detach/quit

[Qemu-devel] [PATCH] gdbstub improvements for gdb kill/detach/quit

[Jason Wessel]

[-- Attachment #1: Type: text/plain, Size: 112 bytes --]

Attached is a patch to improve the gdb stub 'D' 'k' and '?' packets. 
See the patch header for details.

Jason.

[-- Attachment #2: gdb_k_and_D_packet_support.patch --]
[-- Type: text/x-patch, Size: 2876 bytes --]

From: Jason Wessel <jason.wessel@windriver.com>
Subject: [PATCH] support for gdb "detach/kill/quit"

Implement the 'k' gdbserial packet which kills the qemu instance via
the debugger stub.

Implement the 'D' detach packet for the gdb stub such that you can
disconnect gdb with the "detach" command.  This required implementing
a cpu_breakpoint_remove_all function to cleanup all the breakpoints
prior to leaving the gdb stub else simulation can stop with no
debugger attached.

On a '?' packet remove all the breakpoints.  This is considered more
of a safety net in case you force killed gdb or it crashed and you are
reconnecting.  The identical behavior exists for kgdb in the linux
kernel.

Signed-off-by: Jason Wessel <jason.wessel@windriver.com>


---
 cpu-all.h |    1 +
 exec.c    |   14 ++++++++++++++
 gdbstub.c |   16 ++++++++++++++++
 3 files changed, 31 insertions(+)

--- a/cpu-all.h
+++ b/cpu-all.h
@@ -759,6 +759,7 @@ void cpu_interrupt(CPUState *s, int mask
 void cpu_reset_interrupt(CPUState *env, int mask);
 
 int cpu_watchpoint_insert(CPUState *env, target_ulong addr);
+int cpu_breakpoint_remove_all(CPUState *env);
 int cpu_watchpoint_remove(CPUState *env, target_ulong addr);
 int cpu_breakpoint_insert(CPUState *env, target_ulong pc);
 int cpu_breakpoint_remove(CPUState *env, target_ulong pc);
--- a/gdbstub.c
+++ b/gdbstub.c
@@ -901,6 +901,11 @@ static int gdb_handle_packet(GDBState *s
         /* TODO: Make this return the correct value for user-mode.  */
         snprintf(buf, sizeof(buf), "S%02x", SIGTRAP);
         put_packet(s, buf);
+        /* Remove all the breakpoints when this query is issued,
+         * because gdb is doing and initial connect and the state
+         * should be cleaned up. 
+         */
+        cpu_breakpoint_remove_all(env);
         break;
     case 'c':
         if (*p != '\0') {
@@ -924,6 +929,17 @@ static int gdb_handle_packet(GDBState *s
         }
         gdb_continue(s);
 	return RS_IDLE;
+    case 'k':
+        /* Kill the target */
+        fprintf(stderr, "\nQEMU: Terminated via GDBstub\n");
+        exit(0);
+    case 'D':
+        /* Detach packet */
+        if (!cpu_breakpoint_remove_all(env)) {
+            gdb_continue(s);
+            put_packet(s, "OK");
+            break;
+        }
     case 's':
         if (*p != '\0') {
             addr = strtoull(p, (char **)&p, 16);
--- a/exec.c
+++ b/exec.c
@@ -1150,6 +1150,20 @@ int cpu_breakpoint_insert(CPUState *env,
 #endif
 }
 
+/* remove all breakpoints */
+int cpu_breakpoint_remove_all(CPUState *env) {
+#if defined(TARGET_HAS_ICE)
+    int i;
+    for(i = 0; i < env->nb_breakpoints; i++) {
+        breakpoint_invalidate(env, env->breakpoints[i]);
+    }
+    env->nb_breakpoints = 0;
+    return 0;
+#else
+    return -1;
+#endif
+}
+
 /* remove a breakpoint */
 int cpu_breakpoint_remove(CPUState *env, target_ulong pc)
 {

Re: [Qemu-devel] [PATCH] gdbstub improvements for gdb kill/detach/quit

[Edgar E. Iglesias]

On Fri, May 09, 2008 at 08:51:38AM -0500, Jason Wessel wrote:
> Attached is a patch to improve the gdb stub 'D' 'k' and '?' packets. 
> See the patch header for details.
> 
> Jason.

> From: Jason Wessel <jason.wessel@windriver.com>
> Subject: [PATCH] support for gdb "detach/kill/quit"
> 
> Implement the 'k' gdbserial packet which kills the qemu instance via
> the debugger stub.
> 
> Implement the 'D' detach packet for the gdb stub such that you can
> disconnect gdb with the "detach" command.  This required implementing
> a cpu_breakpoint_remove_all function to cleanup all the breakpoints
> prior to leaving the gdb stub else simulation can stop with no
> debugger attached.
> 
> On a '?' packet remove all the breakpoints.  This is considered more
> of a safety net in case you force killed gdb or it crashed and you are
> reconnecting.  The identical behavior exists for kgdb in the linux
> kernel.
> 
> Signed-off-by: Jason Wessel <jason.wessel@windriver.com>

Thanks, a few comments inlined.

> ---
>  cpu-all.h |    1 +
>  exec.c    |   14 ++++++++++++++
>  gdbstub.c |   16 ++++++++++++++++
>  3 files changed, 31 insertions(+)
> 
> --- a/cpu-all.h
> +++ b/cpu-all.h
> @@ -759,6 +759,7 @@ void cpu_interrupt(CPUState *s, int mask
>  void cpu_reset_interrupt(CPUState *env, int mask);
>  
>  int cpu_watchpoint_insert(CPUState *env, target_ulong addr);
> +int cpu_breakpoint_remove_all(CPUState *env);
>  int cpu_watchpoint_remove(CPUState *env, target_ulong addr);
>  int cpu_breakpoint_insert(CPUState *env, target_ulong pc);
>  int cpu_breakpoint_remove(CPUState *env, target_ulong pc);
> --- a/gdbstub.c
> +++ b/gdbstub.c
> @@ -901,6 +901,11 @@ static int gdb_handle_packet(GDBState *s
>          /* TODO: Make this return the correct value for user-mode.  */
>          snprintf(buf, sizeof(buf), "S%02x", SIGTRAP);
>          put_packet(s, buf);
> +        /* Remove all the breakpoints when this query is issued,
> +         * because gdb is doing and initial connect and the state
> +         * should be cleaned up. 
> +         */
> +        cpu_breakpoint_remove_all(env);
>          break;
>      case 'c':
>          if (*p != '\0') {
> @@ -924,6 +929,17 @@ static int gdb_handle_packet(GDBState *s
>          }
>          gdb_continue(s);
>  	return RS_IDLE;
> +    case 'k':
> +        /* Kill the target */
> +        fprintf(stderr, "\nQEMU: Terminated via GDBstub\n");
> +        exit(0);



Should we consider qemu_system_shutdown_request() here?


> +    case 'D':
> +        /* Detach packet */
> +        if (!cpu_breakpoint_remove_all(env)) {
> +            gdb_continue(s);
> +            put_packet(s, "OK");
> +            break;
> +        }
>      case 's':
>          if (*p != '\0') {
>              addr = strtoull(p, (char **)&p, 16);
> --- a/exec.c
> +++ b/exec.c
> @@ -1150,6 +1150,20 @@ int cpu_breakpoint_insert(CPUState *env,
>  #endif
>  }
>  
> +/* remove all breakpoints */
> +int cpu_breakpoint_remove_all(CPUState *env) {
> +#if defined(TARGET_HAS_ICE)
> +    int i;
> +    for(i = 0; i < env->nb_breakpoints; i++) {
> +        breakpoint_invalidate(env, env->breakpoints[i]);
> +    }
> +    env->nb_breakpoints = 0;
> +    return 0;
> +#else
> +    return -1;
> +#endif
> +}

Why not just drop the return value and make the 'D' command always succeed?

Best regards
-- 
Edgar E. Iglesias
Axis Communications AB

Re: [Qemu-devel] [PATCH] gdbstub improvements for gdb kill/detach/quit

[Jason Wessel]

More comments below.

Edgar E. Iglesias wrote:
> On Fri, May 09, 2008 at 08:51:38AM -0500, Jason Wessel wrote:
>> Attached is a patch to improve the gdb stub 'D' 'k' and '?' packets.
>> See the patch header for details.
>>
>>
>> --- a/gdbstub.c
>> +++ b/gdbstub.c
>> @@ -901,6 +901,11 @@ static int gdb_handle_packet(GDBState *s
>> /* TODO: Make this return the correct value for user-mode. */
>> snprintf(buf, sizeof(buf), "S%02x", SIGTRAP);
>> put_packet(s, buf);
>> + /* Remove all the breakpoints when this query is issued,
>> + * because gdb is doing and initial connect and the state
>> + * should be cleaned up.
>> + */
>> + cpu_breakpoint_remove_all(env);
>> break;
>> case 'c':
>> if (*p != '\0') {
>> @@ -924,6 +929,17 @@ static int gdb_handle_packet(GDBState *s
>> }
>> gdb_continue(s);
>> return RS_IDLE;
>> + case 'k':
>> + /* Kill the target */
>> + fprintf(stderr, "\nQEMU: Terminated via GDBstub\n");
>> + exit(0);
>
>
>
> Should we consider qemu_system_shutdown_request() here?


We desire the same action as if you run "quit" from the monitor.
Today the monitor invokes do_quit() which just does exit(). If there
is a more graceful way to exit the qemu instance which saves/flushes
state, the monitor should probably be changed this way as well.

I figure you can use the monitor commands from the gdbstub or directly
in the monitor if you want to do any kind of saving of the state of
the machine at this point.


>
>
>> + case 'D':
>> + /* Detach packet */
>> + if (!cpu_breakpoint_remove_all(env)) {
>> + gdb_continue(s);
>> + put_packet(s, "OK");
>> + break;
>> + }
>> case 's':
>> if (*p != '\0') {
>> addr = strtoull(p, (char **)&p, 16);
>> --- a/exec.c
>> +++ b/exec.c
>> @@ -1150,6 +1150,20 @@ int cpu_breakpoint_insert(CPUState *env,
>> #endif
>> }
>>
>> +/* remove all breakpoints */
>> +int cpu_breakpoint_remove_all(CPUState *env) {
>> +#if defined(TARGET_HAS_ICE)
>> + int i;
>> + for(i = 0; i < env->nb_breakpoints; i++) {
>> + breakpoint_invalidate(env, env->breakpoints[i]);
>> + }
>> + env->nb_breakpoints = 0;
>> + return 0;
>> +#else
>> + return -1;
>> +#endif
>> +}
>
> Why not just drop the return value and make the 'D' command always
succeed?


This seems reasonable. Previously I had simply used the same
call/return semantics as removing a single break point. Also I
noticed the watch points have the exact same problem as the
breakpoints with respect to detaching the debugger, so I fixed that in
the next version of the patch which will be posted shortly.


Jason.