arm64 kexec hang

arm64 kexec hang

[Pratyush Anand]

Hi Geoff,

On Friday 27 March 2015 10:53 PM, Geoff Levand wrote:
> Hi Pratyush,
>
> On Wed, 2015-03-25 at 15:55 +0530, Pratyush Anand wrote:
>> So with following changes kexec load seems to complete without any
>> error. However, kexec reboot does not work yet, Nothing after bye
>> message :( (1st kernel booted with maxcpus=1)
>
> 'Bye!' doesn't mean much, other than the first kernel has
> almost shutdown.  I recommend for debugging you either define
> ARM64_DEBUG_PORT for the kexec-tools build, or have a suitable
> earlyprintk= on the kernel command line.  See the read_sink()
> routine in kexec-arm64.c.
>

Well, I was already using earlycon and it did not print anything after 
*Bye*.

I got OCD-H and is able to track the flow atleast and to see when it dies.

So, it all goes like machine_kexec ->..... -> arm64_kexec_kimage_start 
-> purgatory_start -> purgatory -> printf ... -> putchar

Here it dies (goes to el2_sync_invalid) as arm64_sink does not seem to 
be correct. arm64_sink is (-1) 0xffffffffffffffff. I also tried by just 
returning from printf without doing anything, still it goes to 
el2_sync_invalid (may be from some other location).

I will try to see tomorrow why arm64_sink is wrong and why did not it 
work if printf was null.

just for info: is_hyp_mode_available is true in my case.

~Pratyush


> Either of these will output something like this from the
> purgatory code:
>
>    I'm in purgatory
>    purgatory: kernel_entry: 0000000080080000
>    purgatory: dtb:          0000000080760000
>
> Another option is to add my 'debugging: early putc' patch,
> build the kernel with ARM64_DEBUG_PORT defined for your
>   board.  This will output 'stext' when the second
> stage kernel is entered
>
> The kexec --lite option will bypass purgatory and should just
> show 'stext'.
>
> -Geoff
>
>
> _______________________________________________
> kexec mailing list
> kexec at lists.infradead.org
> http://lists.infradead.org/mailman/listinfo/kexec
>

arm64 kexec hang

[Pratyush Anand]

Hi Geoff,

On Saturday 28 March 2015 02:41 AM, Pratyush Anand wrote:
> Hi Geoff,
>
> On Friday 27 March 2015 10:53 PM, Geoff Levand wrote:
>> Hi Pratyush,
>>
>> On Wed, 2015-03-25 at 15:55 +0530, Pratyush Anand wrote:
>>> So with following changes kexec load seems to complete without any
>>> error. However, kexec reboot does not work yet, Nothing after bye
>>> message :( (1st kernel booted with maxcpus=1)
>>
>> 'Bye!' doesn't mean much, other than the first kernel has
>> almost shutdown.  I recommend for debugging you either define
>> ARM64_DEBUG_PORT for the kexec-tools build, or have a suitable
>> earlyprintk= on the kernel command line.  See the read_sink()
>> routine in kexec-arm64.c.
>>
>

Problem seems to be related to compilation of purgatory code.

For example see here:

   70 0000000000000120 <purgatory>:
   71      120:       a9bf7bfd        stp     x29, x30, [sp,#-16]!
   72      124:       910003fd        mov     x29, sp
   73      128:       58000100        ldr     x0, 148 <purgatory+0x28>
   74      12c:       94000000        bl      544 <printf>


So, when it executes instruction at address 0x128 in above code, it does 
not contain correct address of data where "I'm in purgatory" is located. 
It seems that code has been compiled as 32 bit.

PC: 0x400415012c , PSTATE: 0x400003c9
(gdb) monitor reg x0
x0 (/64): 0x00000000041568F8
(gdb) monitor mdw 0x00000000041568F8 4
0x41568f8: fffefffe fffefffe fffefffe fffefffe  : ................

The buffer location passed to printf is not correct.

Similarly, even if ARM64_DEBUG_PORT is programmed with correct UART TX 
register, arm64_sink is not modified correctly when elf_rel_set_symbol 
is called.

So correct data is at 0x00000040041568F8 and not 0x00000000041568F8

(gdb) monitor mdw 0x00000040041568F8 4
0x40041568f8: 206d2749 70206e69 61677275 79726f74  : I'm in purgatory


should n't the pc relative instruction like adrp be generated instead of 
ldr.

What does -mcmodel=large cflag do? May be it is causing to generate such 
instruction.


~Pratyush

arm64 kexec hang

[Pratyush Anand]

On Monday 30 March 2015 05:56 PM, Pratyush Anand wrote:
> Hi Geoff,
>
> On Saturday 28 March 2015 02:41 AM, Pratyush Anand wrote:
>> Hi Geoff,
>>
>> On Friday 27 March 2015 10:53 PM, Geoff Levand wrote:
>>> Hi Pratyush,
>>>
>>> On Wed, 2015-03-25 at 15:55 +0530, Pratyush Anand wrote:
>>>> So with following changes kexec load seems to complete without any
>>>> error. However, kexec reboot does not work yet, Nothing after bye
>>>> message :( (1st kernel booted with maxcpus=1)
>>>
>>> 'Bye!' doesn't mean much, other than the first kernel has
>>> almost shutdown.  I recommend for debugging you either define
>>> ARM64_DEBUG_PORT for the kexec-tools build, or have a suitable
>>> earlyprintk= on the kernel command line.  See the read_sink()
>>> routine in kexec-arm64.c.
>>>
>>
>
> Problem seems to be related to compilation of purgatory code.
>
> For example see here:
>
>    70 0000000000000120 <purgatory>:
>    71      120:       a9bf7bfd        stp     x29, x30, [sp,#-16]!
>    72      124:       910003fd        mov     x29, sp
>    73      128:       58000100        ldr     x0, 148 <purgatory+0x28>
>    74      12c:       94000000        bl      544 <printf>
>
>
> So, when it executes instruction at address 0x128 in above code, it does
> not contain correct address of data where "I'm in purgatory" is located.
> It seems that code has been compiled as 32 bit.
>
> PC: 0x400415012c , PSTATE: 0x400003c9
> (gdb) monitor reg x0
> x0 (/64): 0x00000000041568F8
> (gdb) monitor mdw 0x00000000041568F8 4
> 0x41568f8: fffefffe fffefffe fffefffe fffefffe  : ................
>
> The buffer location passed to printf is not correct.
>
> Similarly, even if ARM64_DEBUG_PORT is programmed with correct UART TX
> register, arm64_sink is not modified correctly when elf_rel_set_symbol
> is called.
>
> So correct data is at 0x00000040041568F8 and not 0x00000000041568F8
>
> (gdb) monitor mdw 0x00000040041568F8 4
> 0x40041568f8: 206d2749 70206e69 61677275 79726f74  : I'm in purgatory
>
>
> should n't the pc relative instruction like adrp be generated instead of
> ldr.
>
> What does -mcmodel=large cflag do? May be it is causing to generate such
> instruction.
>

Following changes allows purgatory to execute.


diff --git a/kexec/arch/arm64/kexec-arm64.c b/kexec/arch/arm64/kexec-arm64.c
index 8df66f5c8273..4365bb4087ad 100644
--- a/kexec/arch/arm64/kexec-arm64.c
+++ b/kexec/arch/arm64/kexec-arm64.c
@@ -993,8 +993,8 @@ void machine_apply_elf_rel(struct mem_ehdr *ehdr, 
unsigned long r_type,
  # define R_AARCH64_CALL26 283
  #endif

-       uint32_t *location = (uint32_t *)ptr;
-       uint32_t data = *location;
+       uint64_t *location = (uint64_t *)ptr;
+       uint64_t data = *location;
         const char *type = NULL;

         switch(r_type) {
@@ -1026,7 +1026,7 @@ void machine_apply_elf_rel(struct mem_ehdr *ehdr, 
unsigned long r_type,
                 break;
         }

-       dbgprintf("%s: %s %x->%x\n", __func__, type, data, *location);
+       dbgprintf("%s: %s %lx->%lx\n", __func__, type, data, *location);
  }


So I get following and then nothing..

[  162.087569] Bye!
I'm in purgatory
purgatory: kernel_entry: 0000004000280000
purg

Hopefully, I will be able to debug it further.will come back.

~Pratyush

arm64 kexec hang

[Pratyush Anand]

Hi Geoff,


On Monday 30 March 2015 08:08 PM, Pratyush Anand wrote:
> diff --git a/kexec/arch/arm64/kexec-arm64.c
> b/kexec/arch/arm64/kexec-arm64.c
> index 8df66f5c8273..4365bb4087ad 100644
> --- a/kexec/arch/arm64/kexec-arm64.c
> +++ b/kexec/arch/arm64/kexec-arm64.c
> @@ -993,8 +993,8 @@ void machine_apply_elf_rel(struct mem_ehdr *ehdr,
> unsigned long r_type,
>   # define R_AARCH64_CALL26 283
>   #endif
>
> -       uint32_t *location = (uint32_t *)ptr;
> -       uint32_t data = *location;
> +       uint64_t *location = (uint64_t *)ptr;
> +       uint64_t data = *location;
>          const char *type = NULL;
>
>          switch(r_type) {
> @@ -1026,7 +1026,7 @@ void machine_apply_elf_rel(struct mem_ehdr *ehdr,
> unsigned long r_type,
>                  break;
>          }
>
> -       dbgprintf("%s: %s %x->%x\n", __func__, type, data, *location);
> +       dbgprintf("%s: %s %lx->%lx\n", __func__, type, data, *location);
>   }

Thanks for your help/pointer.

So, this was it. With above changes kexec reboot works fine with 
purgatory too.

By the way, how much execution time expected for verify_sha256_digest? I 
have not quantified, but it seems that it takes minutes to execute that.

~Pratyush

arm64 kexec hang

[Geoff Levand]

Hi Pratyush,

On Mon, 2015-03-30 at 20:55 +0530, Pratyush Anand wrote:
> So, this was it. With above changes kexec reboot works fine with 
> purgatory too.

OK, great, thanks for working through it.

> By the way, how much execution time expected for verify_sha256_digest? I 
> have not quantified, but it seems that it takes minutes to execute that.

The digest takes a long time, but it doesn't take minutes on
the fast models.  Maybe you could put some print statements
in the code to see if it is the digest that takes the time,
or something else.  It could be a driver or something waiting
for an event, etc.

-Geoff