Re: cleanup the walk_page_range interface

[Christoph Hellwig]

On Fri, Aug 09, 2019 at 12:21:24AM +0200, Thomas Hellstrom wrote:
> On 8/8/19 11:56 PM, Christoph Hellwig wrote:
>> On Thu, Aug 08, 2019 at 10:50:37AM -0700, Linus Torvalds wrote:
>>>> Note that both Thomas and Steven have series touching this area pending,
>>>> and there are a couple consumer in flux too - the hmm tree already
>>>> conflicts with this series, and I have potential dma changes on top of
>>>> the consumers in Thomas and Steven's series, so we'll probably need a
>>>> git tree similar to the hmm one to synchronize these updates.
>>> I'd be willing to just merge this now, if that helps. The conversion
>>> is mechanical, and my only slight worry would be that at least for my
>>> original patch I didn't build-test the (few) non-x86
>>> architecture-specific cases. But I did end up looking at them fairly
>>> closely  (basically using some grep/sed scripts to see that the
>>> conversions I did matched the same patterns). And your changes look
>>> like obvious improvements too where any mistake would have been caught
>>> by the compiler.
>> I did cross compile the s390 and powerpc bits, but I do not have an
>> openrisc compiler.
>>
>>> So I'm not all that worried from a functionality standpoint, and if
>>> this will help the next merge window, I'll happily pull now.
>> That would help with this series vs the others, but not with the other
>> series vs each other.
>
> Although my series doesn't touch the pagewalk code, it rather borrowed some 
> concepts from it and used for the apply_to_page_range() interface.
>
> The reason being that the pagewalk code requires the mmap_sem to be held 
> (mainly for trans-huge pages and reading the vma->vm_flags if I understand 
> the code correctly). That is fine when you scan the vmas of a process, but 
> the helpers I wrote need to instead scan all vmas pointing into a struct 
> address_space, and taking the mmap_sem for each vma will create lock 
> inversion problems.


True.  So you'll just need to apply the same lessons there, and we
should probably fine with this series going into 5.3-rc.