MIPS: oprofile: Add callgraph support
authorDaniel Kalmar <kalmard@homejinni.com>
Fri, 13 May 2011 12:38:05 +0000 (08:38 -0400)
committerRobert Richter <robert.richter@amd.com>
Wed, 15 Jun 2011 12:35:34 +0000 (14:35 +0200)
Stack unwinding is done by code examination. For kernelspace, the
already existing unwind function is utilized that uses kallsyms to
quickly find the beginning of functions. For userspace a new function
was added that examines code at and before the pc.

Signed-off-by: Daniel Kalmar <kalmard@homejinni.com>
Signed-off-by: Gergely Kis <gergely@homejinni.com>
Signed-off-by: Robert Richter <robert.richter@amd.com>
arch/mips/oprofile/Makefile
arch/mips/oprofile/backtrace.c [new file with mode: 0644]
arch/mips/oprofile/common.c
arch/mips/oprofile/op_impl.h

index 4b9d704..29f2f13 100644 (file)
@@ -8,7 +8,7 @@ DRIVER_OBJS = $(addprefix ../../../drivers/oprofile/, \
                oprofilefs.o oprofile_stats.o \
                timer_int.o )
 
-oprofile-y                             := $(DRIVER_OBJS) common.o
+oprofile-y                             := $(DRIVER_OBJS) common.o backtrace.o
 
 oprofile-$(CONFIG_CPU_MIPS32)          += op_model_mipsxx.o
 oprofile-$(CONFIG_CPU_MIPS64)          += op_model_mipsxx.o
diff --git a/arch/mips/oprofile/backtrace.c b/arch/mips/oprofile/backtrace.c
new file mode 100644 (file)
index 0000000..6854ed5
--- /dev/null
@@ -0,0 +1,175 @@
+#include <linux/oprofile.h>
+#include <linux/sched.h>
+#include <linux/mm.h>
+#include <linux/uaccess.h>
+#include <asm/ptrace.h>
+#include <asm/stacktrace.h>
+#include <linux/stacktrace.h>
+#include <linux/kernel.h>
+#include <asm/sections.h>
+#include <asm/inst.h>
+
+struct stackframe {
+       unsigned long sp;
+       unsigned long pc;
+       unsigned long ra;
+};
+
+static inline int get_mem(unsigned long addr, unsigned long *result)
+{
+       unsigned long *address = (unsigned long *) addr;
+       if (!access_ok(VERIFY_READ, addr, sizeof(unsigned long)))
+               return -1;
+       if (__copy_from_user_inatomic(result, address, sizeof(unsigned long)))
+               return -3;
+       return 0;
+}
+
+/*
+ * These two instruction helpers were taken from process.c
+ */
+static inline int is_ra_save_ins(union mips_instruction *ip)
+{
+       /* sw / sd $ra, offset($sp) */
+       return (ip->i_format.opcode == sw_op || ip->i_format.opcode == sd_op)
+               && ip->i_format.rs == 29 && ip->i_format.rt == 31;
+}
+
+static inline int is_sp_move_ins(union mips_instruction *ip)
+{
+       /* addiu/daddiu sp,sp,-imm */
+       if (ip->i_format.rs != 29 || ip->i_format.rt != 29)
+               return 0;
+       if (ip->i_format.opcode == addiu_op || ip->i_format.opcode == daddiu_op)
+               return 1;
+       return 0;
+}
+
+/*
+ * Looks for specific instructions that mark the end of a function.
+ * This usually means we ran into the code area of the previous function.
+ */
+static inline int is_end_of_function_marker(union mips_instruction *ip)
+{
+       /* jr ra */
+       if (ip->r_format.func == jr_op && ip->r_format.rs == 31)
+               return 1;
+       /* lui gp */
+       if (ip->i_format.opcode == lui_op && ip->i_format.rt == 28)
+               return 1;
+       return 0;
+}
+
+/*
+ * TODO for userspace stack unwinding:
+ * - handle cases where the stack is adjusted inside a function
+ *     (generally doesn't happen)
+ * - find optimal value for max_instr_check
+ * - try to find a way to handle leaf functions
+ */
+
+static inline int unwind_user_frame(struct stackframe *old_frame,
+                                   const unsigned int max_instr_check)
+{
+       struct stackframe new_frame = *old_frame;
+       off_t ra_offset = 0;
+       size_t stack_size = 0;
+       unsigned long addr;
+
+       if (old_frame->pc == 0 || old_frame->sp == 0 || old_frame->ra == 0)
+               return -9;
+
+       for (addr = new_frame.pc; (addr + max_instr_check > new_frame.pc)
+               && (!ra_offset || !stack_size); --addr) {
+               union mips_instruction ip;
+
+               if (get_mem(addr, (unsigned long *) &ip))
+                       return -11;
+
+               if (is_sp_move_ins(&ip)) {
+                       int stack_adjustment = ip.i_format.simmediate;
+                       if (stack_adjustment > 0)
+                               /* This marks the end of the previous function,
+                                  which means we overran. */
+                               break;
+                       stack_size = (unsigned) stack_adjustment;
+               } else if (is_ra_save_ins(&ip)) {
+                       int ra_slot = ip.i_format.simmediate;
+                       if (ra_slot < 0)
+                               /* This shouldn't happen. */
+                               break;
+                       ra_offset = ra_slot;
+               } else if (is_end_of_function_marker(&ip))
+                       break;
+       }
+
+       if (!ra_offset || !stack_size)
+               return -1;
+
+       if (ra_offset) {
+               new_frame.ra = old_frame->sp + ra_offset;
+               if (get_mem(new_frame.ra, &(new_frame.ra)))
+                       return -13;
+       }
+
+       if (stack_size) {
+               new_frame.sp = old_frame->sp + stack_size;
+               if (get_mem(new_frame.sp, &(new_frame.sp)))
+                       return -14;
+       }
+
+       if (new_frame.sp > old_frame->sp)
+               return -2;
+
+       new_frame.pc = old_frame->ra;
+       *old_frame = new_frame;
+
+       return 0;
+}
+
+static inline void do_user_backtrace(unsigned long low_addr,
+                                    struct stackframe *frame,
+                                    unsigned int depth)
+{
+       const unsigned int max_instr_check = 512;
+       const unsigned long high_addr = low_addr + THREAD_SIZE;
+
+       while (depth-- && !unwind_user_frame(frame, max_instr_check)) {
+               oprofile_add_trace(frame->ra);
+               if (frame->sp < low_addr || frame->sp > high_addr)
+                       break;
+       }
+}
+
+#ifndef CONFIG_KALLSYMS
+static inline void do_kernel_backtrace(unsigned long low_addr,
+                                      struct stackframe *frame,
+                                      unsigned int depth) { }
+#else
+static inline void do_kernel_backtrace(unsigned long low_addr,
+                                      struct stackframe *frame,
+                                      unsigned int depth)
+{
+       while (depth-- && frame->pc) {
+               frame->pc = unwind_stack_by_address(low_addr,
+                                                   &(frame->sp),
+                                                   frame->pc,
+                                                   &(frame->ra));
+               oprofile_add_trace(frame->ra);
+       }
+}
+#endif
+
+void notrace op_mips_backtrace(struct pt_regs *const regs, unsigned int depth)
+{
+       struct stackframe frame = { .sp = regs->regs[29],
+                                   .pc = regs->cp0_epc,
+                                   .ra = regs->regs[31] };
+       const int userspace = user_mode(regs);
+       const unsigned long low_addr = ALIGN(frame.sp, THREAD_SIZE);
+
+       if (userspace)
+               do_user_backtrace(low_addr, &frame, depth);
+       else
+               do_kernel_backtrace(low_addr, &frame, depth);
+}
index f9eb1ab..d1f2d4c 100644 (file)
@@ -115,6 +115,7 @@ int __init oprofile_arch_init(struct oprofile_operations *ops)
        ops->start              = op_mips_start;
        ops->stop               = op_mips_stop;
        ops->cpu_type           = lmodel->cpu_type;
+       ops->backtrace          = op_mips_backtrace;
 
        printk(KERN_INFO "oprofile: using %s performance monitoring.\n",
               lmodel->cpu_type);
index f04b54f..7c2da27 100644 (file)
@@ -36,4 +36,6 @@ struct op_mips_model {
        unsigned char num_counters;
 };
 
+void op_mips_backtrace(struct pt_regs * const regs, unsigned int depth);
+
 #endif