Re: [PATCH 1/2] mm/vmalloc: export is_vmalloc_or_module_addr

[Dave Kleikamp <dave.kleikamp@oracle.com>]

On 02/10/2014 06:10 AM, Mel Gorman wrote:
> On Sat, Feb 08, 2014 at 11:45:53AM -0800, Linus Torvalds wrote:
>> On Sat, Feb 8, 2014 at 11:12 AM, Richard Yao <ryao@gentoo.org> wrote:
>>> 9p-virtio needs is_vmalloc_or_module_addr exported before a patch can be
>>> merged to prevent the virtio zero-copy routines from triggering a
>>> hypervisor page fault when loading kernel modules:
>>>
>>> https://groups.google.com/forum/#!topic/linux.kernel/eRR7AyLE29Y
>>>
>>> Without this export, the kernel build breaks with that patch applied and
>>> CONFIG_NET_9P_VIRTIO=m. With this export in place, all is well.
>>
>> I absolutely *detest* this patch.
>>
>> Not so much because is_vmalloc_or_module_addr() should never be used,
>> but because the particular use in question is pure and utter garbage.
>>
>> There are valid reasons to use "is_vmalloc_or_module_addr()", but
>> those are along the lines of drivers/char/mem.c, which says "return
>> -ENXIO for this crap".
>>
>> And btw, that horrid crap called "kmap_to_page()" needs to die too.
>> When is it *ever* valid to use a kmap'ed page for IO? Here's a clue:
>> never.
>>
>> <SNIP>
>>
>> Those *disgusting* get_kernel_page[s]() functions came with a
>> commentary about "The initial user is expected to be NFS.." and that
>> is still the *only* user. The fact that *everybody* else has been able
>> to avoid that crap should tell us something.
> 
> kmap_to_page, get_kernel_page and the NFS user for it are already scheduled
> for death. Dave Kleikamp has been working on a series "Issue O_DIRECT aio
> using bio_vec" that replaces what NFS currently does with something far
> more sane. The series does not remove the helpers on top, probably because
> of the 9P use of it, but if that was changed then the helpers would finally
> die. I'm not sure why Dave's series never got merged so added him to the
> cc to find out.

The patchset I submitted was too ugly to live. Linus and Al Viro both
had numerous issues with the implementation, but not the overall concept.

Al has the same idea for a io-vector container that can hold either an
iovec or a kernel vector and is starting to work on his own
implementation. Once he settles on the particulars of new
aio_write/aio_read methods, I should be able to get the rest of my
patchset to fall into line with it.

http://www.spinics.net/lists/linux-fsdevel/msg71938.html

Shaggy