PM QOS update
[linux-2.6.git] / kernel / kgdb.c
index 79e3c90..11f3515 100644 (file)
 
 static int kgdb_break_asap;
 
+#define KGDB_MAX_THREAD_QUERY 17
 struct kgdb_state {
        int                     ex_vector;
        int                     signo;
        int                     err_code;
        int                     cpu;
        int                     pass_exception;
+       unsigned long           thr_query;
        unsigned long           threadid;
        long                    kgdb_usethreadid;
        struct pt_regs          *linux_regs;
 };
 
+/* Exception state values */
+#define DCPU_WANT_MASTER 0x1 /* Waiting to become a master kgdb cpu */
+#define DCPU_NEXT_MASTER 0x2 /* Transition from one master cpu to another */
+#define DCPU_IS_SLAVE    0x4 /* Slave cpu enter exception */
+#define DCPU_SSTEP       0x8 /* CPU is single stepping */
+
 static struct debuggerinfo_struct {
        void                    *debuggerinfo;
        struct task_struct      *task;
+       int                     exception_state;
 } kgdb_info[NR_CPUS];
 
 /**
@@ -127,6 +136,7 @@ struct task_struct          *kgdb_usethread;
 struct task_struct             *kgdb_contthread;
 
 int                            kgdb_single_step;
+pid_t                          kgdb_sstep_pid;
 
 /* Our I/O buffers. */
 static char                    remcom_in_buffer[BUFMAX];
@@ -166,13 +176,6 @@ early_param("nokgdbroundup", opt_nokgdbroundup);
  * Weak aliases for breakpoint management,
  * can be overriden by architectures when needed:
  */
-int __weak kgdb_validate_break_address(unsigned long addr)
-{
-       char tmp_variable[BREAK_INSTR_SIZE];
-
-       return probe_kernel_read(tmp_variable, (char *)addr, BREAK_INSTR_SIZE);
-}
-
 int __weak kgdb_arch_set_breakpoint(unsigned long addr, char *saved_instr)
 {
        int err;
@@ -191,6 +194,25 @@ int __weak kgdb_arch_remove_breakpoint(unsigned long addr, char *bundle)
                                  (char *)bundle, BREAK_INSTR_SIZE);
 }
 
+int __weak kgdb_validate_break_address(unsigned long addr)
+{
+       char tmp_variable[BREAK_INSTR_SIZE];
+       int err;
+       /* Validate setting the breakpoint and then removing it.  In the
+        * remove fails, the kernel needs to emit a bad message because we
+        * are deep trouble not being able to put things back the way we
+        * found them.
+        */
+       err = kgdb_arch_set_breakpoint(addr, tmp_variable);
+       if (err)
+               return err;
+       err = kgdb_arch_remove_breakpoint(addr, tmp_variable);
+       if (err)
+               printk(KERN_ERR "KGDB: Critical breakpoint error, kernel "
+                  "memory destroyed at: %lx", addr);
+       return err;
+}
+
 unsigned long __weak kgdb_arch_pc(int exception, struct pt_regs *regs)
 {
        return instruction_pointer(regs);
@@ -376,27 +398,22 @@ int kgdb_mem2hex(char *mem, char *buf, int count)
 
 /*
  * Copy the binary array pointed to by buf into mem.  Fix $, #, and
- * 0x7d escaped with 0x7d.  Return a pointer to the character after
- * the last byte written.
+ * 0x7d escaped with 0x7d. Return -EFAULT on failure or 0 on success.
+ * The input buf is overwitten with the result to write to mem.
  */
 static int kgdb_ebin2mem(char *buf, char *mem, int count)
 {
-       int err = 0;
-       char c;
+       int size = 0;
+       char *c = buf;
 
        while (count-- > 0) {
-               c = *buf++;
-               if (c == 0x7d)
-                       c = *buf++ ^ 0x20;
-
-               err = probe_kernel_write(mem, &c, 1);
-               if (err)
-                       break;
-
-               mem++;
+               c[size] = *buf++;
+               if (c[size] == 0x7d)
+                       c[size] = *buf++ ^ 0x20;
+               size++;
        }
 
-       return err;
+       return probe_kernel_write(mem, c, size);
 }
 
 /*
@@ -433,9 +450,14 @@ int kgdb_hex2long(char **ptr, unsigned long *long_val)
 {
        int hex_val;
        int num = 0;
+       int negate = 0;
 
        *long_val = 0;
 
+       if (**ptr == '-') {
+               negate = 1;
+               (*ptr)++;
+       }
        while (**ptr) {
                hex_val = hex(**ptr);
                if (hex_val < 0)
@@ -446,6 +468,9 @@ int kgdb_hex2long(char **ptr, unsigned long *long_val)
                (*ptr)++;
        }
 
+       if (negate)
+               *long_val = -*long_val;
+
        return num;
 }
 
@@ -466,7 +491,7 @@ static int write_mem_msg(int binary)
                if (err)
                        return err;
                if (CACHE_FLUSH_IS_SAFE)
-                       flush_icache_range(addr, addr + length + 1);
+                       flush_icache_range(addr, addr + length);
                return 0;
        }
 
@@ -515,10 +540,21 @@ static void int_to_threadref(unsigned char *id, int value)
 static struct task_struct *getthread(struct pt_regs *regs, int tid)
 {
        /*
-        * Non-positive TIDs are remapped idle tasks:
+        * Non-positive TIDs are remapped to the cpu shadow information
         */
-       if (tid <= 0)
-               return idle_task(-tid);
+       if (tid == 0 || tid == -1)
+               tid = -atomic_read(&kgdb_active) - 2;
+       if (tid < -1 && tid > -NR_CPUS - 2) {
+               if (kgdb_info[-tid - 2].task)
+                       return kgdb_info[-tid - 2].task;
+               else
+                       return idle_task(-tid - 2);
+       }
+       if (tid <= 0) {
+               printk(KERN_ERR "KGDB: Internal thread select error\n");
+               dump_stack();
+               return NULL;
+       }
 
        /*
         * find_task_by_pid_ns() does not take the tasklist lock anymore
@@ -529,45 +565,6 @@ static struct task_struct *getthread(struct pt_regs *regs, int tid)
 }
 
 /*
- * CPU debug state control:
- */
-
-#ifdef CONFIG_SMP
-static void kgdb_wait(struct pt_regs *regs)
-{
-       unsigned long flags;
-       int cpu;
-
-       local_irq_save(flags);
-       cpu = raw_smp_processor_id();
-       kgdb_info[cpu].debuggerinfo = regs;
-       kgdb_info[cpu].task = current;
-       /*
-        * Make sure the above info reaches the primary CPU before
-        * our cpu_in_kgdb[] flag setting does:
-        */
-       smp_wmb();
-       atomic_set(&cpu_in_kgdb[cpu], 1);
-
-       /* Wait till primary CPU is done with debugging */
-       while (atomic_read(&passive_cpu_wait[cpu]))
-               cpu_relax();
-
-       kgdb_info[cpu].debuggerinfo = NULL;
-       kgdb_info[cpu].task = NULL;
-
-       /* fix up hardware debug registers on local cpu */
-       if (arch_kgdb_ops.correct_hw_break)
-               arch_kgdb_ops.correct_hw_break();
-
-       /* Signal the primary CPU that we are done: */
-       atomic_set(&cpu_in_kgdb[cpu], 0);
-       clocksource_touch_watchdog();
-       local_irq_restore(flags);
-}
-#endif
-
-/*
  * Some architectures need cache flushes when we set/clear a
  * breakpoint:
  */
@@ -590,7 +587,8 @@ static void kgdb_flush_swbreak_addr(unsigned long addr)
 static int kgdb_activate_sw_breakpoints(void)
 {
        unsigned long addr;
-       int error = 0;
+       int error;
+       int ret = 0;
        int i;
 
        for (i = 0; i < KGDB_MAX_BREAKPOINTS; i++) {
@@ -600,13 +598,16 @@ static int kgdb_activate_sw_breakpoints(void)
                addr = kgdb_break[i].bpt_addr;
                error = kgdb_arch_set_breakpoint(addr,
                                kgdb_break[i].saved_instr);
-               if (error)
-                       return error;
+               if (error) {
+                       ret = error;
+                       printk(KERN_INFO "KGDB: BP install failed: %lx", addr);
+                       continue;
+               }
 
                kgdb_flush_swbreak_addr(addr);
                kgdb_break[i].state = BP_ACTIVE;
        }
-       return 0;
+       return ret;
 }
 
 static int kgdb_set_sw_break(unsigned long addr)
@@ -653,7 +654,8 @@ static int kgdb_set_sw_break(unsigned long addr)
 static int kgdb_deactivate_sw_breakpoints(void)
 {
        unsigned long addr;
-       int error = 0;
+       int error;
+       int ret = 0;
        int i;
 
        for (i = 0; i < KGDB_MAX_BREAKPOINTS; i++) {
@@ -662,13 +664,15 @@ static int kgdb_deactivate_sw_breakpoints(void)
                addr = kgdb_break[i].bpt_addr;
                error = kgdb_arch_remove_breakpoint(addr,
                                        kgdb_break[i].saved_instr);
-               if (error)
-                       return error;
+               if (error) {
+                       printk(KERN_INFO "KGDB: BP remove failed: %lx\n", addr);
+                       ret = error;
+               }
 
                kgdb_flush_swbreak_addr(addr);
                kgdb_break[i].state = BP_SET;
        }
-       return 0;
+       return ret;
 }
 
 static int kgdb_remove_sw_break(unsigned long addr)
@@ -725,14 +729,15 @@ setundefined:
 }
 
 /*
- * Remap normal tasks to their real PID, idle tasks to -1 ... -NR_CPUs:
+ * Remap normal tasks to their real PID,
+ * CPU shadow threads are mapped to -CPU - 2
  */
 static inline int shadow_pid(int realpid)
 {
        if (realpid)
                return realpid;
 
-       return -1-raw_smp_processor_id();
+       return -raw_smp_processor_id() - 2;
 }
 
 static char gdbmsgbuf[BUFMAX + 1];
@@ -826,7 +831,7 @@ static void gdb_cmd_getregs(struct kgdb_state *ks)
                local_debuggerinfo = kgdb_info[ks->cpu].debuggerinfo;
        } else {
                local_debuggerinfo = NULL;
-               for (i = 0; i < NR_CPUS; i++) {
+               for_each_online_cpu(i) {
                        /*
                         * Try to find the task on some other
                         * or possibly this node if we do not
@@ -840,7 +845,7 @@ static void gdb_cmd_getregs(struct kgdb_state *ks)
 
        /*
         * All threads that don't have debuggerinfo should be
-        * in __schedule() sleeping, since all other CPUs
+        * in schedule() sleeping, since all other CPUs
         * are in kgdb_wait, and thus have debuggerinfo.
         */
        if (local_debuggerinfo) {
@@ -960,10 +965,13 @@ static int gdb_cmd_reboot(struct kgdb_state *ks)
 /* Handle the 'q' query packets */
 static void gdb_cmd_query(struct kgdb_state *ks)
 {
-       struct task_struct *thread;
+       struct task_struct *g;
+       struct task_struct *p;
        unsigned char thref[8];
        char *ptr;
        int i;
+       int cpu;
+       int finished = 0;
 
        switch (remcom_in_buffer[1]) {
        case 's':
@@ -973,22 +981,34 @@ static void gdb_cmd_query(struct kgdb_state *ks)
                        break;
                }
 
-               if (remcom_in_buffer[1] == 'f')
-                       ks->threadid = 1;
-
+               i = 0;
                remcom_out_buffer[0] = 'm';
                ptr = remcom_out_buffer + 1;
-
-               for (i = 0; i < 17; ks->threadid++) {
-                       thread = getthread(ks->linux_regs, ks->threadid);
-                       if (thread) {
-                               int_to_threadref(thref, ks->threadid);
+               if (remcom_in_buffer[1] == 'f') {
+                       /* Each cpu is a shadow thread */
+                       for_each_online_cpu(cpu) {
+                               ks->thr_query = 0;
+                               int_to_threadref(thref, -cpu - 2);
                                pack_threadid(ptr, thref);
                                ptr += BUF_THREAD_ID_SIZE;
                                *(ptr++) = ',';
                                i++;
                        }
                }
+
+               do_each_thread(g, p) {
+                       if (i >= ks->thr_query && !finished) {
+                               int_to_threadref(thref, p->pid);
+                               pack_threadid(ptr, thref);
+                               ptr += BUF_THREAD_ID_SIZE;
+                               *(ptr++) = ',';
+                               ks->thr_query++;
+                               if (ks->thr_query % KGDB_MAX_THREAD_QUERY == 0)
+                                       finished = 1;
+                       }
+                       i++;
+               } while_each_thread(g, p);
+
                *(--ptr) = '\0';
                break;
 
@@ -1011,15 +1031,15 @@ static void gdb_cmd_query(struct kgdb_state *ks)
                        error_packet(remcom_out_buffer, -EINVAL);
                        break;
                }
-               if (ks->threadid > 0) {
+               if ((int)ks->threadid > 0) {
                        kgdb_mem2hex(getthread(ks->linux_regs,
                                        ks->threadid)->comm,
                                        remcom_out_buffer, 16);
                } else {
                        static char tmpstr[23 + BUF_THREAD_ID_SIZE];
 
-                       sprintf(tmpstr, "Shadow task %d for pid 0",
-                                       (int)(-ks->threadid-1));
+                       sprintf(tmpstr, "shadowCPU%d",
+                                       (int)(-ks->threadid - 2));
                        kgdb_mem2hex(tmpstr, remcom_out_buffer, strlen(tmpstr));
                }
                break;
@@ -1159,8 +1179,10 @@ static int gdb_cmd_exception_pass(struct kgdb_state *ks)
                return 1;
 
        } else {
-               error_packet(remcom_out_buffer, -EINVAL);
-               return 0;
+               kgdb_msg_write("KGDB only knows signal 9 (pass)"
+                       " and 15 (pass and disconnect)\n"
+                       "Executing a continue without signal passing\n", 0);
+               remcom_in_buffer[0] = 'c';
        }
 
        /* Indicate fall through */
@@ -1337,33 +1359,13 @@ static int kgdb_reenter_check(struct kgdb_state *ks)
        return 1;
 }
 
-/*
- * kgdb_handle_exception() - main entry point from a kernel exception
- *
- * Locking hierarchy:
- *     interface locks, if any (begin_session)
- *     kgdb lock (kgdb_active)
- */
-int
-kgdb_handle_exception(int evector, int signo, int ecode, struct pt_regs *regs)
+static int kgdb_cpu_enter(struct kgdb_state *ks, struct pt_regs *regs)
 {
-       struct kgdb_state kgdb_var;
-       struct kgdb_state *ks = &kgdb_var;
        unsigned long flags;
+       int sstep_tries = 100;
        int error = 0;
        int i, cpu;
-
-       ks->cpu                 = raw_smp_processor_id();
-       ks->ex_vector           = evector;
-       ks->signo               = signo;
-       ks->ex_vector           = evector;
-       ks->err_code            = ecode;
-       ks->kgdb_usethreadid    = 0;
-       ks->linux_regs          = regs;
-
-       if (kgdb_reenter_check(ks))
-               return 0; /* Ouch, double exception ! */
-
+       int trace_on = 0;
 acquirelock:
        /*
         * Interrupts will be restored by the 'trap return' code, except when
@@ -1371,23 +1373,55 @@ acquirelock:
         */
        local_irq_save(flags);
 
-       cpu = raw_smp_processor_id();
+       cpu = ks->cpu;
+       kgdb_info[cpu].debuggerinfo = regs;
+       kgdb_info[cpu].task = current;
+       /*
+        * Make sure the above info reaches the primary CPU before
+        * our cpu_in_kgdb[] flag setting does:
+        */
+       atomic_inc(&cpu_in_kgdb[cpu]);
 
        /*
-        * Acquire the kgdb_active lock:
+        * CPU will loop if it is a slave or request to become a kgdb
+        * master cpu and acquire the kgdb_active lock:
         */
-       while (atomic_cmpxchg(&kgdb_active, -1, cpu) != -1)
+       while (1) {
+               if (kgdb_info[cpu].exception_state & DCPU_WANT_MASTER) {
+                       if (atomic_cmpxchg(&kgdb_active, -1, cpu) == cpu)
+                               break;
+               } else if (kgdb_info[cpu].exception_state & DCPU_IS_SLAVE) {
+                       if (!atomic_read(&passive_cpu_wait[cpu]))
+                               goto return_normal;
+               } else {
+return_normal:
+                       /* Return to normal operation by executing any
+                        * hw breakpoint fixup.
+                        */
+                       if (arch_kgdb_ops.correct_hw_break)
+                               arch_kgdb_ops.correct_hw_break();
+                       if (trace_on)
+                               tracing_on();
+                       atomic_dec(&cpu_in_kgdb[cpu]);
+                       touch_softlockup_watchdog_sync();
+                       clocksource_touch_watchdog();
+                       local_irq_restore(flags);
+                       return 0;
+               }
                cpu_relax();
+       }
 
        /*
-        * Do not start the debugger connection on this CPU if the last
-        * instance of the exception handler wanted to come into the
-        * debugger on a different CPU via a single step
+        * For single stepping, try to only enter on the processor
+        * that was single stepping.  To gaurd against a deadlock, the
+        * kernel will only try for the value of sstep_tries before
+        * giving up and continuing on.
         */
        if (atomic_read(&kgdb_cpu_doing_single_step) != -1 &&
-           atomic_read(&kgdb_cpu_doing_single_step) != cpu) {
-
+           (kgdb_info[cpu].task &&
+            kgdb_info[cpu].task->pid != kgdb_sstep_pid) && --sstep_tries) {
                atomic_set(&kgdb_active, -1);
+               touch_softlockup_watchdog_sync();
                clocksource_touch_watchdog();
                local_irq_restore(flags);
 
@@ -1409,29 +1443,20 @@ acquirelock:
        if (kgdb_io_ops->pre_exception)
                kgdb_io_ops->pre_exception();
 
-       kgdb_info[ks->cpu].debuggerinfo = ks->linux_regs;
-       kgdb_info[ks->cpu].task = current;
-
        kgdb_disable_hw_debug(ks->linux_regs);
 
        /*
         * Get the passive CPU lock which will hold all the non-primary
         * CPU in a spin state while the debugger is active
         */
-       if (!kgdb_single_step || !kgdb_contthread) {
+       if (!kgdb_single_step) {
                for (i = 0; i < NR_CPUS; i++)
-                       atomic_set(&passive_cpu_wait[i], 1);
+                       atomic_inc(&passive_cpu_wait[i]);
        }
 
-       /*
-        * spin_lock code is good enough as a barrier so we don't
-        * need one here:
-        */
-       atomic_set(&cpu_in_kgdb[ks->cpu], 1);
-
 #ifdef CONFIG_SMP
        /* Signal the other CPUs to enter kgdb_wait() */
-       if ((!kgdb_single_step || !kgdb_contthread) && kgdb_do_roundup)
+       if ((!kgdb_single_step) && kgdb_do_roundup)
                kgdb_roundup_cpus(flags);
 #endif
 
@@ -1450,8 +1475,11 @@ acquirelock:
        kgdb_post_primary_code(ks->linux_regs, ks->ex_vector, ks->err_code);
        kgdb_deactivate_sw_breakpoints();
        kgdb_single_step = 0;
-       kgdb_contthread = NULL;
+       kgdb_contthread = current;
        exception_level = 0;
+       trace_on = tracing_is_on();
+       if (trace_on)
+               tracing_off();
 
        /* Talk to debugger with gdbserial protocol */
        error = gdb_serial_stub(ks);
@@ -1460,13 +1488,11 @@ acquirelock:
        if (kgdb_io_ops->post_exception)
                kgdb_io_ops->post_exception();
 
-       kgdb_info[ks->cpu].debuggerinfo = NULL;
-       kgdb_info[ks->cpu].task = NULL;
-       atomic_set(&cpu_in_kgdb[ks->cpu], 0);
+       atomic_dec(&cpu_in_kgdb[ks->cpu]);
 
-       if (!kgdb_single_step || !kgdb_contthread) {
+       if (!kgdb_single_step) {
                for (i = NR_CPUS-1; i >= 0; i--)
-                       atomic_set(&passive_cpu_wait[i], 0);
+                       atomic_dec(&passive_cpu_wait[i]);
                /*
                 * Wait till all the CPUs have quit
                 * from the debugger.
@@ -1478,28 +1504,78 @@ acquirelock:
        }
 
 kgdb_restore:
+       if (atomic_read(&kgdb_cpu_doing_single_step) != -1) {
+               int sstep_cpu = atomic_read(&kgdb_cpu_doing_single_step);
+               if (kgdb_info[sstep_cpu].task)
+                       kgdb_sstep_pid = kgdb_info[sstep_cpu].task->pid;
+               else
+                       kgdb_sstep_pid = 0;
+       }
+       if (trace_on)
+               tracing_on();
        /* Free kgdb_active */
        atomic_set(&kgdb_active, -1);
+       touch_softlockup_watchdog_sync();
        clocksource_touch_watchdog();
        local_irq_restore(flags);
 
        return error;
 }
 
+/*
+ * kgdb_handle_exception() - main entry point from a kernel exception
+ *
+ * Locking hierarchy:
+ *     interface locks, if any (begin_session)
+ *     kgdb lock (kgdb_active)
+ */
+int
+kgdb_handle_exception(int evector, int signo, int ecode, struct pt_regs *regs)
+{
+       struct kgdb_state kgdb_var;
+       struct kgdb_state *ks = &kgdb_var;
+       int ret;
+
+       ks->cpu                 = raw_smp_processor_id();
+       ks->ex_vector           = evector;
+       ks->signo               = signo;
+       ks->ex_vector           = evector;
+       ks->err_code            = ecode;
+       ks->kgdb_usethreadid    = 0;
+       ks->linux_regs          = regs;
+
+       if (kgdb_reenter_check(ks))
+               return 0; /* Ouch, double exception ! */
+       kgdb_info[ks->cpu].exception_state |= DCPU_WANT_MASTER;
+       ret = kgdb_cpu_enter(ks, regs);
+       kgdb_info[ks->cpu].exception_state &= ~DCPU_WANT_MASTER;
+       return ret;
+}
+
 int kgdb_nmicallback(int cpu, void *regs)
 {
 #ifdef CONFIG_SMP
+       struct kgdb_state kgdb_var;
+       struct kgdb_state *ks = &kgdb_var;
+
+       memset(ks, 0, sizeof(struct kgdb_state));
+       ks->cpu                 = cpu;
+       ks->linux_regs          = regs;
+
        if (!atomic_read(&cpu_in_kgdb[cpu]) &&
-                       atomic_read(&kgdb_active) != cpu &&
-                       atomic_read(&cpu_in_kgdb[atomic_read(&kgdb_active)])) {
-               kgdb_wait((struct pt_regs *)regs);
+           atomic_read(&kgdb_active) != -1 &&
+           atomic_read(&kgdb_active) != cpu) {
+               kgdb_info[cpu].exception_state |= DCPU_IS_SLAVE;
+               kgdb_cpu_enter(ks, regs);
+               kgdb_info[cpu].exception_state &= ~DCPU_IS_SLAVE;
                return 0;
        }
 #endif
        return 1;
 }
 
-void kgdb_console_write(struct console *co, const char *s, unsigned count)
+static void kgdb_console_write(struct console *co, const char *s,
+   unsigned count)
 {
        unsigned long flags;
 
@@ -1535,8 +1611,8 @@ static void sysrq_handle_gdb(int key, struct tty_struct *tty)
 
 static struct sysrq_key_op sysrq_gdb_op = {
        .handler        = sysrq_handle_gdb,
-       .help_msg       = "Gdb",
-       .action_msg     = "GDB",
+       .help_msg       = "debug(G)",
+       .action_msg     = "DEBUG",
 };
 #endif
 
@@ -1667,11 +1743,11 @@ EXPORT_SYMBOL_GPL(kgdb_unregister_io_module);
  */
 void kgdb_breakpoint(void)
 {
-       atomic_set(&kgdb_setting_breakpoint, 1);
+       atomic_inc(&kgdb_setting_breakpoint);
        wmb(); /* Sync point before breakpoint */
        arch_kgdb_breakpoint();
        wmb(); /* Sync point after breakpoint */
-       atomic_set(&kgdb_setting_breakpoint, 0);
+       atomic_dec(&kgdb_setting_breakpoint);
 }
 EXPORT_SYMBOL_GPL(kgdb_breakpoint);