On Fri, Feb 08, 2019 at 07:13:39AM +0000, Alexey Brodkin wrote: > Hi Greg, > > > -----Original Message----- > > From: gregkh@xxxxxxxxxxxxxxxxxxx <gregkh@xxxxxxxxxxxxxxxxxxx> > > Sent: Thursday, February 7, 2019 6:52 PM > > To: Alexey Brodkin <alexey.brodkin@xxxxxxxxxxxx> > > Cc: David.Laight@xxxxxxxxxx; geert@xxxxxxxxxxxxxx; peterz@xxxxxxxxxxxxx; stable@xxxxxxxxxxxxxxx; > > tglx@xxxxxxxxxxxxx; will.deacon@xxxxxxx; Vineet Gupta <vineet.gupta1@xxxxxxxxxxxx>; linux-snps- > > arc@xxxxxxxxxxxxxxxxxxx > > Subject: Re: patch "devres: Align data[] to ARCH_KMALLOC_MINALIGN" added to driver-core-linus > > [snip] > > > Ah, I was waiting to see if you would notice :) > > Well I was just patiently waiting as I guess there's a long queue > of patches to deal with in your inbox :) > > > See this question from Linus about this patch: > > https://lore.kernel.org/lkml/CAHk-=wj3Q7CkMQYwfZSsqUTqkEhNwVGrRbCwe7AVJ70S8i5sWw@xxxxxxxxxxxxxx/ > > I didn't see that. Though I intentionally sent my patch to most if not all > arch maintainers so they might share their concerns... but IIRC nobody ever > replied with either concerns or acks. > > Also I do agree that it's a trade-off between: > 1. Predictability > I was completely sure devm-allocated buffer is the same as anything kmalloced > except some meta-data stored _separately_ and so supposed alignment > should match as well... but how wrong that feeling was. > > 2. Optimization > Indeed it's so sweet when both devm "meta-data" and real small buffer fit > into 1 cache line. > > > I figured that you all did this for a good reason, and wasting that much > > space was going to be ok. But, I wanted to be sure, so if you never > > noticed it, I figured it was not that pressing of an issue. > > It's not super pressing because: > 1. Fortunately [or unfortunately] this problem happens only in pretty rare cases > like that Etnaviv driver where I first caught it. > > 2. There's a solution and affected parties may apply known patch locally. > > > Anyway, is this really needed to be backported? > > For us poor ARC developers and users it's really needed as our tools ABI > sets 32-bit alignment for 64-bit types. See that's the same optimization - > why wasting precious bytes on useless holes - let's pack data tighter :) > > So having that fix at least in the most recent LTS (i.e. 4.19) would be really good. > As for older kernels I think for now we may not touch them as indeed change is > quite intrusive. Ok, that sounds reasonable. I'll go add the patch there and see if the ARM64 people even notice :) thanks, greg k-h
