Message-ID: <20161027092919.GA3666@yury-N73SV>
Date:   Thu, 27 Oct 2016 12:29:19 +0300
From:   Yury Norov <ynorov@...iumnetworks.com>
To:     Arnd Bergmann <arnd@...db.de>
CC:     Chris Metcalf <cmetcalf@...lanox.com>, <catalin.marinas@....com>,
        <linux-arm-kernel@...ts.infradead.org>,
        <linux-kernel@...r.kernel.org>, <linux-doc@...r.kernel.org>,
        <linux-arch@...r.kernel.org>, <schwidefsky@...ibm.com>,
        <heiko.carstens@...ibm.com>, <pinskia@...il.com>,
        <broonie@...nel.org>, <joseph@...esourcery.com>,
        <christoph.muellner@...obroma-systems.com>,
        <bamvor.zhangjian@...wei.com>, <szabolcs.nagy@....com>,
        <klimov.linux@...il.com>, <Nathan_Lynch@...tor.com>,
        <agraf@...e.de>, <Prasun.Kapoor@...iumnetworks.com>,
        <kilobyte@...band.pl>, <geert@...ux-m68k.org>,
        <philipp.tomsich@...obroma-systems.com>,
        <manuel.montezelo@...il.com>, <linyongting@...wei.com>,
        <maxim.kuvyrkov@...aro.org>, <davem@...emloft.net>,
        <zhouchengming1@...wei.com>
Subject: Re: [PATCH 01/18] 32-bit ABI: introduce ARCH_32BIT_OFF_T config
 option

On Tue, Oct 25, 2016 at 12:22:47AM +0200, Arnd Bergmann wrote:
> On Monday, October 24, 2016 12:30:47 PM CEST Chris Metcalf wrote:
> > On 10/21/2016 4:33 PM, Yury Norov wrote:
> > > All new 32-bit architectures should have 64-bit off_t type, but existing
> > > architectures has 32-bit ones.
> > >
> > > [...]
> > > For syscalls sys_openat() and sys_open_by_handle_at() force_o_largefile()
> > > is called, to set O_LARGEFILE flag, and this is the only difference
> > > comparing to compat versions. All compat ABIs are already turned to use
> > > 64-bit off_t, except tile. So, compat versions for this syscalls are not
> > > needed anymore. Tile is handled explicitly.
> > >
> > > [...]
> > > --- a/arch/tile/kernel/compat.c
> > > +++ b/arch/tile/kernel/compat.c
> > > @@ -103,6 +103,9 @@ COMPAT_SYSCALL_DEFINE5(llseek, unsigned int, fd, unsigned int, offset_high,
> > >   #define compat_sys_readahead sys32_readahead
> > >   #define sys_llseek compat_sys_llseek
> > >   
> > > +#define sys_openat             compat_sys_openat
> > > +#define sys_open_by_handle_at  compat_sys_open_by_handle_at
> > > +
> > >   /* Call the assembly trampolines where necessary. */
> > >   #define compat_sys_rt_sigreturn _compat_sys_rt_sigreturn
> > >   #define sys_clone _sys_clone
> > 
> > This patch accomplishes two goals that could be completely separated.
> > It's confusing to have them mixed in the same patch without any
> > discussion of why they are in the same patch.
> > 
> > First, you want to modify the default <asm/unistd.h> behavior for
> > compat syscalls so that the default is sys_openat (etc) rather than
> > the existing compat_sys_openat, and then use that new behavior for
> > arm64 ILP32.  This lets you force O_LARGEFILE for arm64 ILP32 to
> > support having a 64-bit off_t at all times.  To do that, you fix the
> > asm-generic header, and then make tile have a special override.
> > This seems reasonable enough.
> > 
> > Second, you introduce ARCH_32BIT_OFF_T basically as a synonym for
> > "BITS_PER_WORD == 32", so that new 32-bit architectures can choose not
> > to enable it.  This is fine in the abstract, but I'm a bit troubled by
> > the fact that you are not actually introducing a new 32-bit
> > architecture here (just a new 32-bit mode for the arm 64-bit kernel).
> > Shouldn't this part of the change wait until someone actually has a
> > new 32-bit kernel to drive this forward?
> 
> I asked for this specifically because we identified the problem
> during the review of the aarch64 ilp32 code, and it might not
> be noticed in the next architecture submission.
> 
> The most important aspect from my perspective is that the new
> ilp32 ABI on aarch64 behaves the same way that any native 32-bit
> architecture does, and when we change the default, it should
> be done for both compat mode and native mode at the same time.
> 
> > If you want to push forward the ARCH_32BIT_OFF_T change in the absence
> > of an architecture that supports it, I would think it would be a lot
> > less confusing to have these two in separate patches, and make it
> > clear that the ARCH_32BIT_OFF_T change is just laying groundwork
> > for some hypothetical future architecture.
> > 
> > The existing commit language itself is also confusing. You write "All
> > compat ABIs are already turned to use 64-bit off_t, except tile."
> > First, I'm not sure what you mean by "turned" here.  And, tile is just
> > one of many compat ABIs that allow O_LARGEFILE not to be part of the
> > open call: see arm64's AArch32 ABI, MIPS o32, s390 31-bit emulation,
> > sparc64's 32-bit mode, and of course x86's 32-bit compat mode.
> > Presumably your point here is that tile is the only pre-existing
> > architecture that #includes <asm/unistd.h> to create its compat
> > syscall table, and so I think "all except tile" here is particularly
> > confusing, since there are no architectures except tile that use the
> > __SYSCALL_COMPAT functionality in the current tree.
> 
> Agreed, this could be made clearer, and splitting the patch up
> in two also seems reasonable, though I didn't see it as important.
> 
> 	Arnd

In the past it was a separated series of 2 patches, and it was even
acked by Arnd, but not submitted. 
http://lists-archives.com/linux-kernel/28471253-32-bit-abi-introduce-arch_32bit_off_t-config-option.html

I can restore that small series in aarch64/ilp32 for next iteration, or resend
it separately if you think to submit it before aarch64/ilp32 (which is
better, for me).

Yury

Powered by blists - more mailing lists