Message-ID: <20201102072558.GA12588@infradead.org>
Date:   Mon, 2 Nov 2020 07:25:58 +0000
From:   Christoph Hellwig <hch@...radead.org>
To:     Linus Torvalds <torvalds@...ux-foundation.org>
Cc:     Christoph Hellwig <hch@...radead.org>,
        iommu <iommu@...ts.linux-foundation.org>,
        Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
        Geert Uytterhoeven <geert+renesas@...der.be>,
        Rob Herring <robh@...nel.org>
Subject: Re: [GIT PULL] dma-mapping fix for 5.10

On Sat, Oct 31, 2020 at 12:50:44PM -0700, Linus Torvalds wrote:
> So this is just a stylistic nit, and has no impact on this pull (which
> I've done). But looking at the patch, it triggers one of my "this is
> wrong" patterns.

Adding the author and maintainer of that code so that they can sort it
out.

> 
> In particular, this:
> 
>         u64 dma_start = 0;
>         ...
>         for (dma_start = ~0ULL; r->size; r++) {
> 
> is actually completely bogus in theory, and it's a horribly horribly
> bad pattern to have.
> 
> The thing that I hate about that parttern is "~0ULL", which is simply wrong.
> 
> The correct pattern for "all bits set" is ~0. NOTHING ELSE. No extra
> letters at the end.
> 
> Why? Because using an unsigned type is wrong, and will not extend the
> bits up to a potentially bigger size.
> 
> So adding that "ULL" is not just three extra characters to type, it
> actually _detracts_ from the code and makes it more fragile and
> potentially wrong.
> 
> It so happens, that yes, in the kernel, "ull" us 64-bit, and you get
> the right results. So the code _works_. But it's wrong, and it now
> requires that the types match exactly (ie it would not be broken if
> somebody ever were to say "I want to use use 128-bit dma addresses and
> u128").
> 
> Another example is using "~0ul", which would give different results on
> a 32-bit kernel and a 64-bit kernel. Again: DON'T DO THAT.
> 
> I repeat: the right thing to do for "all bits set" is just a plain ~0
> or -1. Either of those are fine (technically assumes a 2's complement
> machine, but let's just be honest: that's a perfectly fine assumption,
> and -1 might be preferred by some because it makes that sign extension
> behavior of the integer constant more obvious).
> 
> Don't try to do anything clever or anything else, because it's going
> to be strictly worse.
> 
> The old code that that patch removed was "technically correct", but
> just pointless, and actually shows the problem:
> 
>         for (dma_start = ~(dma_addr_t)0; r->size; r++) {
> 
> the above is indeed a correct way to say "I want all bits set in a
> dma_addr_t", but while correct, it is - once again - strictly inferior
> to just using "~0".
> 
> Why? Because "~0" works regardless of type. IOW, exactly *because*
> people used the wrong pattern for "all bits set", that patch was now
> (a) bigger than necessary and (b) much more ilkely to cause bugs (ie I
> could have imagined people changing just the type of the variable
> without changing the initialization).
> 
> So in that tiny three-line patch there were actually several examples
> of why "~0" is the right pattern to use for "all bits set". Because it
> JustWorks(tm) in ways other patterns do not.
> 
> And if you have a compiler that complains about assigning -1 or ~0 to
> an unsigned variable, get rid of that piece of garbage. You're almost
> certainly either using some warning flag that you shouldn't be using,
> or the compiler writer didn't know what they were doing.
> 
>             Linus
> _______________________________________________
> iommu mailing list
> iommu@...ts.linux-foundation.org
> https://lists.linuxfoundation.org/mailman/listinfo/iommu
---end quoted text---

Powered by blists - more mailing lists