Re: [PATCH] Add VDSO time function support for x86 32-bit kernel

From: "H. Peter Anvin" <hpa@zytor.com>
To: Andy Lutomirski <luto@amacapital.net>
Cc: Stefani Seibold <stefani@seibold.net>,
	linux-kernel@vger.kernel.org, x86@kernel.org, tglx@linutronix.de,
	mingo@redhat.com, ak@linux.intel.com, aarcange@redhat.com,
	john.stultz@linaro.org
Subject: Re: [PATCH] Add VDSO time function support for x86 32-bit kernel
Date: Thu, 13 Dec 2012 16:36:22 -0800	[thread overview]
Message-ID: <50CA7486.2070804@zytor.com> (raw)
In-Reply-To: <CALCETrV1kvhOvDpzTavGb1X9HT+PdEPXJ0H8afs7jzxpcF7Kjg@mail.gmail.com>

On 12/13/2012 04:20 PM, Andy Lutomirski wrote:
>>
>> Whatever data you need you can just map it into the vdso range.  There
>> really shouldn't be anything special about that at all.
>>
>> The fixmap stuff is an x86-64 legacy that you don't have to worry about,
>> obviously.
>
> The fixmap page is new.  It's not ABI -- the layout can freely change,
> but the vdso needs to be able to see.  It would make no sense to cow
> that page, and it would be more complicated to make it be part of the
> (64-bit) vdso, so I left it as a fixmap when I did the vvar cleanups.
>

Well, the vsyscall fixmap is an ABI.  But just because a page is mapped 
into userspace doesn't mean cow.  Think of it as a device mmap, or an 
mmap of a shared file.

> For compat mode, though, I don't think it can be in the fixmap unless
> there are segmentation games that I think are impossible, or unless
> the vdso wants to do a far call (which I would guess is not much
> faster than a system call, although I haven't benchmarked it).  This
> is because there are no addresses at all that can be seen from 32-bit
> code segments and that are in the kernel address range.

Yes, you'd have to nip into 64-bit mode which is not really practical.

> What you could do is probably arrange (using some linker script magic)
> for a symbol to exist that points at the page *before* the vdso
> starts.  Then just map the vvar page there when starting a compat
> task.  You could then address it using a normal symbol reference by
> tweaking the vvar macro.  (I think this'll access it via the GOT.)
> Alternatively, you could just pick an absolute address -- the page is
> NX, so you don't really need to worry about randomization.

IMO it seems this is making it way more complicated than it is.  Just 
make sure you have a section in the vdso where you can map in a data 
page with the symbols in the right offsets.  Extra points for doing 
magic so that it is at the beginning or end, but I think that might be 
harder than necessary.

	-hpa

-- 
H. Peter Anvin, Intel Open Source Technology Center
I work for Intel.  I don't speak on their behalf.