Re: [GIT PULL] Block fixes for 4.5-final

[Jens Axboe <axboe@fb.com>]

On 03/03/2016 02:20 PM, Linus Torvalds wrote:
> On Thu, Mar 3, 2016 at 1:11 PM, Jens Axboe <axboe@fb.com> wrote:
>>
>> It does fix a regression - the change is that NVMe now uses the block layer
>> for these types of requests, and they don't have to adhere to the regular fs
>> limits of sizing. Hence we broke real use cases, of (for instance) pulling
>> logs off devices. Both of the referenced commits were added yesterday, not
>> today. And they should have been folded, but I had already committed the
>> first one. I don't think that should preclude doing it much cleaner than the
>> first one.
>
> Why does this affect only NVMe, and not all the other drivers that
> have been around forever?  What is that magical case that breaks?
> Details, please.

Development around NVMe is a lot more active than any other driver. And 
that tends to drive a lot more testing, and find a lot of other bugs. 
That, and the fact that NVMe is still fairly young. On top of that, NVMe 
has been driving/utilizing some parts of blk-mq, and exercising things 
like surprise hot removal that haven't seen a ton of testing.

>> Fair enough, I can boil it down somewhat. But honestly, the only stuff I'd
>> feel comfortable pulling out now would be the lightnvm changes which aren't
>> that critical due to the user base, though that's also why it would be fine
>> to shove it in now. And the cgroup writeback enable, which can wait. The two
>> commits referenced above could be folded, but they'd still be in the new
>> pull request.
>>
>> So let me know if you want that, or we can proceed with the current branch,
>> because most of it should really go in as-is.
>
> I basically want for every commit an explanation of why it's so
> critical by now. I want to make you have to *think* and explain before
> you send stuff at this stage, and I want to understand why each commit
> is so important.
>
> Because really, this has been going on far too long, and this pull
> request looked singularly pointless.
>
> No way do I want things like cgroup writeback changes outside the
> merge window, for example, unless it's a major performance regression
> (with numbers) or something like that.
>
> No way do I want any lightnvm stuff.
>
> No way do I want big "cleanup" patches.

I'll boil it down.

-- 
Jens Axboe