Re: [PATCH v8 13/20] xenctx: change is_kernel_text() into kernel_addr().

[Don Slutz <dslutz@verizon.com>]

On 04/02/14 06:34, Ian Campbell wrote:
> On Tue, 2014-04-01 at 17:33 -0400, Don Slutz wrote:
>> On 04/01/14 10:10, Ian Campbell wrote:
>>> On Thu, 2014-03-27 at 15:05 -0400, Don Slutz wrote:
>>

>>>> kernel_end or the symbol "__brk_limit".  The symbol "vgettimeofday"
>>>> is used to mark the end of where modules can be loaded.
>>> I can just about see from the vmlinux.lds.S why __brk_limit might be
>>> somewhat relevant, but where does the use of vgettimeofday come from?
>>>
>>> The 3.2 kernel on my desktop doesn't list vgettimeofday in System.map
>>> for example and in any case you can't rely on it not being moved around.
>>> (you can't really rely on __brk_limit either, but it seems less likely
>>> to move).
>> I would guess you are running a 32 bit desktop.
> No. 3.13-1-amd64 and I looked in System.map-3.13-1-amd64
> System.map-3.2.0-4-amd64 and System.map-3.9-1-amd64 which happened to be
> present in /boot.

I only find it in 2.6 kernels.  Since I am dropping the module support
this symbol is no longer checked for.


>> Here is output of the same 3 stack pages in use guest that is paused:
> You keep dropping in these massive output dumps without explaining what
> they are supposed to illustrate. I have no idea what you think this is
> telling me.

This was in response to:

 >>> This shows that a kernel module is using a lot of stack.

 >> Does it? All I see is the four additional entries at the base, consuming
 >> about 0.5k which I suppose is a lot. I suppose those are stack frames
 >> corresponding to some module? Perhaps mark them as such with "unknown
 >> module function"?

I was trying to tell you that the module code was using much more then 0.5k.
It is just over 2 pages.  But forgot to write this sentence 1st.

Clearly what I am writing in e-mails is not at all clear to you. Will try to do better.

>>>> +            return KERNEL_TEXT_ADDR;
>>>> +    }
>>>> +    else
>>>> +    {
>>>> +        if ( kernel_text &&
>>>> +             (addr >= kernel_text &&
>>>> +              addr <= kernel_end) )
>>>> +            return KERNEL_DATA_ADDR;
>>>> +        if ( kernel_mod_start &&
>>>> +             (addr >= kernel_mod_start &&
>>>> +              addr <= kernel_mod_end) )
>>>> +            return KERNEL_MOD_ADDR;
>>>> +    }
>>>> +    return NOT_KERNEL_ADDR;
>>>>    }
>>>>    
>>>> @@ -180,7 +223,14 @@ static void print_symbol(guest_word_t addr)
>>>>        if (addr==s->address)
>>>>            printf(" %s", s->name);
>>>>        else
>>>> -        printf(" %s+%#x", s->name, (unsigned int)(addr - s->address));
>>>> +    {
>>>> +        unsigned long long offset = addr - s->address;
>>>> +
>>>> +        if ( addr_type == KERNEL_MOD_ADDR &&
>>>> +             offset > MAX_MOD_OFFSET )
>>>> +            return;
>>> Shouldn't kernel_addr have not returned KERNEL_MOD_ADDR in that case?
>>>
>> I did.
> You did what?
>

I had meant to type "It". But that is also wrong.  Somehow I dropped the not in
"not returned KERNEL_MOD_ADDR".  What I should have said:

kernel_addr when checking for KERNEL_MOD_ADDR only does a range check.
It does not search the symbols to see if the address is close enough.

With these issues, I see no reason to attempt adding module support at this time.

If you want I can also answer the later questions.

    -Don Slutz