Re: [Qemu-devel] [RFC][PATCH 05/16 v6] Add API to get memory mapping

[Jan Kiszka]

On 2012-02-15 05:07, Wen Congyang wrote:
> At 02/15/2012 01:21 AM, Jan Kiszka Wrote:
>> On 2012-02-09 04:22, Wen Congyang wrote:
>>> Add API to get all virtual address and physical address mapping.
>>> If there is no virtual address for some physical address, the virtual
>>> address is 0.
>>>
>>> Signed-off-by: Wen Congyang <address@hidden>
>>> ---
>>>  memory_mapping.c |   65 
>>> ++++++++++++++++++++++++++++++++++++++++++++++++++++++
>>>  memory_mapping.h |    1 +
>>>  2 files changed, 66 insertions(+), 0 deletions(-)
>>>
>>> diff --git a/memory_mapping.c b/memory_mapping.c
>>> index d83b7d7..fc0ddee 100644
>>> --- a/memory_mapping.c
>>> +++ b/memory_mapping.c
>>> @@ -128,3 +128,68 @@ void free_memory_mapping_list(MemoryMappingList *list)
>>>  
>>>      list->num = 0;
>>>  }
>>> +
>>> +void get_memory_mapping(MemoryMappingList *list)
>>> +{
>>> +    CPUState *env;
>>> +    MemoryMapping *memory_mapping;
>>> +    RAMBlock *block;
>>> +    ram_addr_t offset, length;
>>> +
>>> +    last_mapping = NULL;
>>> +
>>> +    for (env = first_cpu; env != NULL; env = env->next_cpu) {
>>> +        cpu_get_memory_mapping(list, env);
>>
>> Hmm, is the CPU number recorded along with the mappings? I mean, how
>> could crash tell them apart afterward if they are contradictory? This
>> way, they are just thrown in the same bucket, correct?
>>
>> Even if crash or gdb aren't prepared for cpu/thread-specific mappings,
>> could we already record that information for later use? Or would it
>> break compatibility with current versions?
> 
> crash does not need this information. It only needs the physical address
> stored in PT_LOAD.

So crash does not support viewing memory through the eyes of different
CPUs? OK.

> 
> gdb needs the virtual address and physical address stored in PT_LOAD.
> 
> If the address is in the kernel space, the virtual address and physical
> address mapping should be the same. I collect the mapping information
> from all vcpus, because the OS may enter the second kernel. In this case,
> IIRC(according to my test result, but I don't remeber clearly), gdb's bt
> can output the backtrace in the first kernel if the OS does not use the
> first vcpu to do kdump. otherwise gdb's bt can output the backtrace in
> the second kernel.

gdb could only make proper use of the additional mappings if they are
not contradictory (which can easily happen with user space processes) or
the cpu context is additionally provided so that views can be switched
via the "thread N" command. So far, QEMU's gdbstub does this for gdb
when it requests some memory over the remote connection. I bet gdb
requires some extension to exploit such information offline from a core
file, but I'm also sure that this will come as the importance of gdb for
system level debugging will rise.

Therefore my question: is there room to encode the mapping relation to a
CPU/thread context?

Jan

-- 
Siemens AG, Corporate Technology, CT T DE IT 1
Corporate Competence Center Embedded Linux