On 8/21/20 5:29 PM, Linus Torvalds wrote: > On Fri, Aug 21, 2020 at 4:01 PM Jens Axboe <axboe@xxxxxxxxx> wrote: >> >> The tree is totally warning clean right now on build, but that's usually >> not the case, and missing a single warning isn't that hard. > > Actually, my tree is generally entirely warning-free and has been for > years, because I absolutely hate warnings. > > You're not the only person I haven't pulled from because I saw a > single warning. This is the second one just this release window. I'm > not singling you out. A warning to me is simply a hard "No". > > Because I know very well that if there is even a single warning, > people will then ignore the other warnings. > > So there is no such thing as a harmless warning - it doesn't matter > _what_ it warns about, it's a bad thing, because even one bogus > warning will then mean that people ignore warnings that _aren't_ > bogus, and it goes all downhill from there. We're in violent agreement on this part. None of the projects I maintain throw any warnings, exactly for the same reason. My test box is using gcc-10 and it definitely hasn't been "clean for years". It is now, and has been for probably a release or two. The warnings do condition you to ignore new warnings, which is of course a sad outcome that shouldn't happen. I don't think anyone sane would disagree on that. > So the reason I notice is exactly because I keep my build tree 100% > clean. I do allmodconfig builds that are entirely clean, and I check > it after each and every pull I do, and each patch series I apply. > Mistakes can happen, but I try _very_ hard to keep my tree clean. > > Now, that said, I'll miss warnings too, because I don't do > cross-builds for other architectures, and I really only do a couple of > fairly basic configurations. > > But perhaps more importantly - I don't build with a lot of different > compilers. I do in fact test two different compilers (both gcc and > clang), but I use two different configurations, so it's not really > overlapping testing, it's more that these days I make sure the clang > build is warning free at least for the core kernel too. > > So "warning free" is always relative to some compiler base (and some > config base). But the baseline should absolutely be "zero warnings". > > Generally, new warnings in my tree is because of a compiler upgrade or > similar congifuration issues. > > And it's sadly the "different compilers" thing that then means that > *my* clean tree might not then necessarily be clean for others. It's > also why I can't just enable "-Werror" to make it harder to miss. > > But I've been tempted. Many times. There's been a couple of times I've > added that "-Werror" to the Makefile, only to never commit it. > > But if you have warnings that you see due to different compilers or > configurations, holler. We'll get them fixed. I obviously don't disagree with any of this, it'd be hard to, and it wasn't really what any of my replies were about. It was about how this was dealt with. But I won't keep beating the dead horse. As I said, I'll rebase the branch, and in fact I already did. I likely won't be able to get it out before after -rc2, which is pretty sad since there's an nvme regression in there that would have been nice to have. There's an off chance that I'll send it out tomorrow, depends on timing. -- Jens Axboe
