MIME-Version: 1.0
References: <YZvIlz7J6vOEY+Xu@yuki> <CAK8P3a0x5Bw7=0ng-s+KsUywqJYa0tk9cSWmZhx+cZRBOR87ZA@mail.gmail.com>
 <YZyw56flmdQnBIuh@yuki>
In-Reply-To: <YZyw56flmdQnBIuh@yuki>
From:   Arnd Bergmann <arnd@arndb.de>
Date:   Tue, 23 Nov 2021 15:18:39 +0100
Message-ID: <CAK8P3a2RU8XJp_hS0JkO9mPJctAHHKBobV97gced6pMXcwzWow@mail.gmail.com>
Subject: Re: [PATCH] uapi: Make __{u,s}64 match {u,}int64_t in userspace
To:     Cyril Hrubis <chrubis@suse.cz>
Cc:     Arnd Bergmann <arnd@arndb.de>,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
        Linux API <linux-api@vger.kernel.org>,
        LTP List <ltp@lists.linux.it>,
        GNU C Library <libc-alpha@sourceware.org>,
        linux-arch <linux-arch@vger.kernel.org>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Tue, Nov 23, 2021 at 10:14 AM Cyril Hrubis <chrubis@suse.cz> wrote:
> > I don't think this is correct on all 64-bit architectures, as far as I
> > remember the
> > definition can use either 'long' or 'long long' depending on the user space
> > toolchain.
>
> As far as I can tell the userspace bits/types.h does exactly the same
> check in order to define uint64_t and int64_t, i.e.:
>
> #if __WORDSIZE == 64
> typedef signed long int __int64_t;
> typedef unsigned long int __uint64_t;
> #else
> __extension__ typedef signed long long int __int64_t;
> __extension__ typedef unsigned long long int __uint64_t;
> #endif
>
> The macro __WORDSIZE is defined per architecture, and it looks like the
> defintions in glibc sources in bits/wordsize.h match the uapi
> asm/bitsperlong.h. But I may have missed something, the code in glibc is
> not exactly easy to read.

It's possible that the only difference between the two files was the
'__u32'/'__s32' definition, which could be either 'int' or 'long'. We used
to try matching the user space types for these, but not use 'int'
everywhere in the kernel.

> > Out of the ten supported 64-bit architectures, there are four that already
> > use asm-generic/int-l64.h conditionally, and six that don't, and I
> > think at least
> > some of those are intentional.
> >
> > I think it would be safer to do this one architecture at a time to make
> > sure this doesn't regress on those that require the int-ll64.h version.
>
> I'm still trying to understand what exactly can go wrong here. As long
> as __BITS_PER_LONG is correctly defined the __u64 and __s64 will be
> correctly sized as well. The only visible change is that one 'long' is
> dropped from the type when it's not needed.

Correct, I'm not worried about getting incorrectly-sized types here,
but using the wrong type can cause compile-time warnings when
they are mismatched against format strings or assigning pointers
to the wrong types. With the kernel types, one would always use
%d for __u32 and %lld for __u64, while with the user space types,
one has to resort to using macros.

       Arnd