Skip to contentSkip to conversation

Switch master cpu after gdb thread command for SMP (v2)

PATCH In review sonic zhang 2 participants 2 messages Sep 24, 2008 – Oct 21, 2008 KGDB / KDB /debug_core

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 Switch master cpu after gdb thread command for SMP (v2)

In blackfin SMP architecture, different core has its own L1 SRAM and MMR memory, which code running on the other core can't access. In current kgdb impelemntation, cpus are represented by thread with minus prefix.

If user run thread command in gdb to switch to the thread of the other cpu, kgdb should:
1. send IPI signal to master cpu
2. release the specific passive cpu waiting in IPI handler
3. exit kgdb exception loop on master cpu and trap into kgdb wait in IPI handler
4. trap the released passive cpu into kgdb exception in IPI handler

This patch is tested on bf561-ezkit board with SMP enabled.

---
 include/linux/kgdb.h |   14 +++++++++++
 kernel/kgdb.c        |   64 +++++++++++++++++++++++++++++++++++++++++++-------
 2 files changed, 69 insertions(+), 9 deletions(-)
include/linux/kgdb.h +14 -0
index cb8a3d6..4c3fd6c 100644
@@ -107,6 +107,7 @@ struct kgdb_bkpt {
 #endif
 
 #define KGDB_HW_BREAKPOINT	1
+#define KGDB_THR_PROC_SWAP	2
 
 /*
  * Functions each KGDB-supporting architecture must provide:
@@ -203,6 +204,19 @@ kgdb_arch_handle_exception(int vector, int signo, int err_code,
  */
 extern void kgdb_roundup_cpus(unsigned long flags);
 
+/**
+ *	kgdb_roundup_cpu - Get spcific CPU into a holding pattern
+ *	@cpu: Specific cpu id
+ *	@flags: Current IRQ state
+ *
+ *	On SMP systems, we need to switch cpu from current active one to
+ *	the other passive one. This get current active CPU into a known state
+ *	in kgdb_wait(). 
+ *
+ *	On non-SMP systems, this is not called.
+ */
+extern void kgdb_roundup_cpu(int cpu, unsigned long flags);
+
 /* Optional functions. */
 extern int kgdb_validate_break_address(unsigned long addr);
 extern int kgdb_arch_set_breakpoint(unsigned long addr, char *saved_instr);
kernel/kgdb.c +55 -9

159 lines not on this page · raw message

Signature · 1 line
1.6.0
maintainer
#1 lore raw

Re: [PATCH] [kgdb] Switch master cpu after gdb thread command for SMP(v2)

Quoted 2 lines · sonic zhang

In blackfin SMP architecture, different core has its own L1 SRAM and MMR memory, which code running on the other core can't access. In current

kgdb

Quoted 7 lines

impelemntation, cpus are represented by thread with minus prefix.

If user run thread command in gdb to switch to the thread of the other cpu, kgdb should:
1. send IPI signal to master cpu
2. release the specific passive cpu waiting in IPI handler
3. exit kgdb exception loop on master cpu and trap into kgdb wait in

IPI handler

Quoted 4 lines

4. trap the released passive cpu into kgdb exception in IPI handler

This patch is tested on bf561-ezkit board with SMP enabled.

It does look better this time around, but the patch does not apply to the kgdb dev branch.

Hunk #10 FAILED at 1502.
Hunk #11 FAILED at 1518.
Hunk #12 FAILED at 1537.
Hunk #13 FAILED at 1550.
Hunk #14 succeeded at 1584 (offset -12 lines).
4 out of 14 hunks FAILED -- rejects in file kernel/kgdb.c

Quoted 5 lines
---
 include/linux/kgdb.h |   14 +++++++++++
 kernel/kgdb.c        |   64

+++++++++++++++++++++++++++++++++++++++++++-------

include/linux/kgdb.h struct kgdb_bkpt { · line 110

Quoted 11 lines
 2 files changed, 69 insertions(+), 9 deletions(-)
include/linux/kgdb.h +1 -0
index cb8a3d6..4c3fd6c 100644
@@ -107,6 +107,7 @@ struct kgdb_bkpt {
 #endif
 
 #define KGDB_HW_BREAKPOINT    1
+#define KGDB_THR_PROC_SWAP    2

I can fix this later, but it should have always been a bit mask indicated by 0x1 0x2...

Quoted 11 lines
kernel/kgdb.c +1 -0
@@ -566,6 +566,7 @@ static void kgdb_wait(struct pt_regs *regs)
 {
     unsigned long flags;
     int cpu;
+    struct task_struct *thread;
 
     local_irq_save(flags);
     cpu = raw_smp_processor_id();

kernel/kgdb.c static void kgdb_wait(struct pt_regs *regs) · line 583

Quoted 5 lines
kernel/kgdb.c +2 -0
@@ -578,6 +579,8 @@ static void kgdb_wait(struct pt_regs *regs)
     smp_wmb();
     atomic_set(&cpu_in_kgdb[cpu], 1);
 
+    kgdb_disable_hw_debug(regs);
+

Good catch, this is a real arch independent defect, which should be fixed.

kernel/kgdb.c static int gdb_serial_stub(struct kgdb_state *ks) · line 1256

Quoted 6 lines
kernel/kgdb.c +1 -1
@@ -1223,7 +1253,13 @@ static int gdb_serial_stub(struct kgdb_state *ks)
     /* Clear the out buffer. */
     memset(remcom_out_buffer, 0, sizeof(remcom_out_buffer));
 
-    if (kgdb_connected) {
+    if (kgdb_contthread) {

I thik that check needs to be
if (arch_kgdb_ops.flags & KGDB_THR_PROC_SWAP &&
kgdb_contthread)

Although I am not quite certain, and ther was no way to further look at it since your patch did not apply.

The use of contthread changed a bit in the 2.6.27, as it was found to not be doing what it was intended.

See commit d7161a65341556bacb5e6654e133803f46f51063 for the details.

kernel/kgdb.c static int gdb_serial_stub(struct kgdb_state *ks) · line 1265

Quoted 9 lines
kernel/kgdb.c +6 -0
+        remcom_out_buffer[0] = 'O';
+        remcom_out_buffer[1] = 'K';
+        remcom_out_buffer[2] = 0;
+        put_packet(remcom_out_buffer);
+        kgdb_contthread = NULL;
+    } else if (kgdb_connected) {
         unsigned char thref[8];
         char *ptr;
 

Jason.