[PATCH 0/2] kdump/mmap: Fix mmap of /proc/vmcore for s390

[Eric W. Biederman]

Vivek Goyal <vgoyal at redhat.com> writes:

> On Fri, May 24, 2013 at 05:06:26PM +0200, Michael Holzheu wrote:
>> Hello Vivek,
>> 
>> On Fri, 24 May 2013 10:36:44 -0400
>> Vivek Goyal <vgoyal at redhat.com> wrote:
>> 
>> [snip]
>> 
>> > Sorry, I don't understand the problem. If we swapped low memory and
>> > crash reserved memory, that should have been taken care by prepared
>> > ELF headers so that we map the right pfns. In x86 we swap 640K of low
>> > memory with 640K of memory in reserved and we take care of this by
>> > preparing elf headers accordingly.
>> > 
>> > So why s390 can't do the same thing?
>> 
>> I am not sure if I understand this. Currently we create the ELF
>> header in a way that we have virtual=real. In the copy_oldmem_page() we
>> do the swap so that for the /proc/vmcore code it looks like contiguous
>> non-swapped memory.
>> 
>> One reason why I thought this was necessary was that /dev/oldmem
>> also uses the function and it should provide linear memory access like
>> it is on the live system with /dev/mem.
>> 
>> Is that implementation incorrect?
>
> [ CC Andrew. Keep him in loop for all kernel kdump patches as all kdump
>   patches are routed through him ].
>
> [ CC Eric Biederman ]
>
> Looking at the code, looks like /dev/oldmem is broken. It does not know
> anything about swap of any of the memory areas and it will simply
> return the contents of page frame asked. And this has been like this
> since the beginning.
>
> I have always questioned the utility of /dev/oldmem. Atleast I am not
> aware of any tool making use of it.
>
> If we want to fix it, then somebow all the swapped memory region info
> needs to be communicated to second kernel so that read_oldmem() can
> do the mapping correctly and we really don't have any mechanism for
> that. (I am assuming that in s390 you must have hardcoded the regions
> of memory which are always swapped).
>
> As /proc/vmcore is the most used and useful interface, I prefer that
> we swap memory and put that info in elf headers. For /dev/oldme, I
> don't mind if we leave it as it is. If somebody really cares, then
> I guess we need to write a new command line option which /dev/mem
> can parse and which tells it about swaps so that /dev/oldmem can
> map things correctly. (This is better than hardcoding things).
>
> Eric, do you have any thoughts on this.

I don't think anyone actually uses /dev/oldmem.  I would like to cite
the s390 confusion as proof but I don't think that quite works.

I think the solution is for someone to send a patch removing /dev/oldmem
as an unused piece of code.  That will also move us in the direction of
resolving HPAs concerns.

The function copy_oldmem_page also concerns me.  I don't have a clue why
we duplicate that function on every architecutre in a slightly different
form.  There should be enough abstractions in the kernel to make that
unnecessary.  I would be glad to see that function go, and remove the
possibility of confusion that happened on s390.

Eric