[Crash-utility] Problem in bt for ARM64

[Karlsson, Jan]

Ok. I have seen this change in the pt_regs struct before but did not connect it to this problem. I see these new field in pt_regs in earlier kernel versions than 4.7, but it is probably backports. It really does not matter for the solution of the problem. The following change works for me:
Change:
#define USER_EFRAME_OFFSET (304)
to:
#define USER_EFRAME_OFFSET (STRUCT_SIZE("pt_regs") + 16)

Then you might want to avoid the recalculation of the struct size. I think your patch does the same thing.

Jan

________________________________________
Från: crash-utility-bounces at redhat.com <crash-utility-bounces at redhat.com> för Dave Anderson <anderson at redhat.com>
Skickat: den 21 september 2017 17:38
Till: Discussion list for crash utility usage,  maintenance and development
Ämne: Re: [Crash-utility] Problem in bt for ARM64

----- Original Message -----
>
>
> ----- Original Message -----
> > Hi Dave
> >
> > I have experienced some problems in the bt command for ARM64. It seems that
> > the test in arm64_print_exception_frame in arm64.c if the task is running
> > in
> > 32 or 64-bit mode in userland does not work. It "always" becomes 32-bit
> > mode. Example:
> >
> > crash> bt 1
> > PID: 1      TASK: ffffffe1f90f8000  CPU: 2   COMMAND: "init"
> >  #0 [ffffffe1f9103c80] __switch_to at ffffff85b6a862f8
> >  #1 [ffffffe1f9103ca0] __schedule at ffffff85b7b0d9b0
> >  #2 [ffffffe1f9103d00] schedule at ffffff85b7b0df28
> >  #3 [ffffffe1f9103d20] schedule_hrtimeout_range_clock at ffffff85b7b11308
> >  #4 [ffffffe1f9103da0] schedule_hrtimeout_range at ffffff85b7b11320
> >  #5 [ffffffe1f9103db0] sys_epoll_wait at ffffff85b6c394c8
> >  #6 [ffffffe1f9103e70] sys_epoll_pwait at ffffff85b6c396fc
> >  #7 [ffffffe1f9103ed0] el0_svc_naked at ffffff85b6a8312c
> >      PC: 00000004  LR: 00000000  SP: 00000000  PSTATE: 00000016
> >     X12: 00546694 X11: 3431206c616e6769 X10: 00546338  X9: 00000000
> >      X8: 00000112  X7: dfdab819254dd1e8  X6: 00000016  X5: 0000000a
> >      X4: 00000031  X3: 00000008  X2: 00000000  X1: ffffffff
> >      X0: 00000001
> >
> > The register values are:
> > r0:                 4   r1:        7ff0b27f90
> > r2:                 1   r3:          ffffffff
> > r4:                 0   r5:                 8
> > r6:                31   r7:                 a
> > r8:                16   r9:  dfdab819254dd1e8
> > r10:              112   r11:                0
> > r12:           546338   r13: 3431206c616e6769
> > r14:           546694   r15:                0
> > r16:                0   r17:         f04245b7
> > r18:         51f2a993   r19:           5783c0
> > r20:           415254   r21:           527a5c
> > r22:           527e04   r23:         ffffffff
> > r24:         ffffffff   r25:           576000
> > r26:           578000   r27:           578000
> > r28:              3e8   fp:        7ff0b27ec0
> > lr:            4f4f24   sp:        7ff0b27eb0
> > pc:            4fb8d4   psr:         40000000
> >
> > I have unfortunately not had the time to look for a solution, so I just want
> > to report what I have seen. The kernel running in the example above is
> > 4.4.74 and I have seen the same problem for a 4.9.40 kernel.
> >
> > Jan
>
> Yeah, even the functions that do show 64-bit user mode exception frames
> appear to have invalid contents.  It looks like the pt_regs is not at the
> same fixed location on the stack where it has always been, i.e., as it gets
> set here unconditionally at the bottom of arm64_back_trace_cmd():
>
> complete_user:
>         exception_frame = bt->stacktop - USER_EFRAME_OFFSET;
>         arm64_print_exception_frame(bt, exception_frame, USER_MODE, ofp);
>

Jan,

I don't see the problem in anything before 4.7, and I do have a set
of 4.5 dumpfiles and an early 4.7-rc0 dumpfile that do not show the
problem.  Are you sure about seeing it in a "4.4.74" kernel?

I say that because it does seem to be confluent with v4.7-rc7 commit
e19a6ee2460bdd0d0055a6029383422773f9999a that has this:

  diff --git a/arch/arm64/include/asm/ptrace.h b/arch/arm64/include/asm/ptrace.h
  index a307eb6..7f94755 100644
  --- a/arch/arm64/include/asm/ptrace.h
  +++ b/arch/arm64/include/asm/ptrace.h
  @@ -117,6 +117,8 @@ struct pt_regs {
          };
          u64 orig_x0;
          u64 syscallno;
  +       u64 orig_addr_limit;
  +       u64 unused;     // maintain 16 byte alignment
   };

Does the attached patch work for you?

Thanks,
  Dave



diff --git a/arm64.c b/arm64.c
index 0ab8396..9fe0c87 100644
--- a/arm64.c
+++ b/arm64.c
@@ -393,6 +393,11 @@ arm64_init(int when)
                        get_symbol_data("nr_irqs", sizeof(unsigned int),
                                &machdep->nr_irqs);

+               if (MEMBER_EXISTS("pt_regs", "orig_addr_limit"))
+                       ms->user_eframe_offset = 320;
+               else
+                       ms->user_eframe_offset = 304;
+
                if (!machdep->hz)
                        machdep->hz = 100;

@@ -608,6 +613,7 @@ arm64_dump_machdep_table(ulong arg)
        fprintf(fp, "      exp_entry2_start: %lx\n", ms->exp_entry2_start);
        fprintf(fp, "        exp_entry2_end: %lx\n", ms->exp_entry2_end);
        fprintf(fp, "       panic_task_regs: %lx\n", (ulong)ms->panic_task_regs);
+       fprintf(fp, "    user_eframe_offset: %ld\n", ms->user_eframe_offset);
        fprintf(fp, "         PTE_PROT_NONE: %lx\n", ms->PTE_PROT_NONE);
        fprintf(fp, "              PTE_FILE: ");
        if (ms->PTE_FILE)
@@ -1461,7 +1467,7 @@ arm64_stackframe_init(void)
 #define KERNEL_MODE (1)
 #define USER_MODE   (2)

-#define USER_EFRAME_OFFSET (304)
+#define USER_EFRAME_OFFSET (machdep->machspec->user_eframe_offset)

 /*
  * PSR bits
diff --git a/defs.h b/defs.h
index fd97ffe..8534d40 100644
--- a/defs.h
+++ b/defs.h
@@ -3110,6 +3110,7 @@ struct machine_specific {
        ulong kimage_voffset;
        ulong kimage_text;
        ulong kimage_end;
+       ulong user_eframe_offset;
 };

 struct arm64_stackframe {