Re: [PATCH v12 9/9] Add 32 bit VDSO time support for 64 bit kernel

[Stefani Seibold]

Am Mittwoch, den 05.02.2014, 14:01 -0800 schrieb Andy Lutomirski:
> On Wed, Feb 5, 2014 at 12:20 AM,  <stefani@xxxxxxxxxxx> wrote:
> > From: Stefani Seibold <stefani@xxxxxxxxxxx>
> >
> > This patch add the VDSO time support for the IA32 Emulation Layer.
> >
> > Due the nature of the kernel headers and the LP64 compiler where the
> > size of a long and a pointer differs against a 32 bit compiler, there
> > is some type hacking necessary.
> >
> > The vsyscall_gtod_data struture must be a little bit rearranged, to
> > serve 32- and 64-bit code access:
> >
> > - The seqcount_t was replaced by an unsigned, this makes the
> >   vsyscall_gtod_data intedepend of kernel configuration and internal functions.
> > - The structure is now packed, so it can accessed from 32- und 64- bit
> >   code at the same time.
> > - The inner struct clock was removed, to make packing of the while
> >   struct easier.
> >
> > The "unsigned seq" would be handled by functions derivated from seqcount_t.
> >
> > Signed-off-by: Stefani Seibold <stefani@xxxxxxxxxxx>
> > ---
> >  arch/x86/include/asm/vgtod.h          |  20 +++---
> >  arch/x86/kernel/vsyscall_gtod.c       |  26 +++++--
> >  arch/x86/vdso/vclock_gettime.c        | 129 ++++++++++++++++++++++++----------
> >  arch/x86/vdso/vdso32/vclock_gettime.c |  11 +++
> >  include/uapi/linux/time.h             |   2 +-
> >  5 files changed, 132 insertions(+), 56 deletions(-)
> >
> > diff --git a/arch/x86/include/asm/vgtod.h b/arch/x86/include/asm/vgtod.h
> > index 46e24d3..2567b02 100644
> > --- a/arch/x86/include/asm/vgtod.h
> > +++ b/arch/x86/include/asm/vgtod.h
> > @@ -4,16 +4,18 @@
> >  #include <asm/vsyscall.h>
> >  #include <linux/clocksource.h>
> >
> > -struct vsyscall_gtod_data {
> > -       seqcount_t      seq;
> > +/*
> > + * vsyscall_gtod_data will be accessed by 32 and 64 bit code at the same time
> > + * so the structure must be packed
> > + */
> > +struct __attribute__((packed)) vsyscall_gtod_data {
> > +       unsigned seq;
> 
> Is that actually true?  At least in the part you're changing,
> everything looks like it's aligned correctly.
> 

Yes, i know. But for convenient this is less error prone when modifying
the structure. I can kick it out if you insist.
 
> It's almost certainly having some kind of BUILD_BUG_ON that will catch
> the case where this structure's size changes.  I suspect that some
> kind of asm-offsets magic can be used for this.
> 

This kind af ASM magic wan't work, because the code will be compiled
with -m32 for a 32 bit VDSO but will access a structure generated with a
64 bit compiler.

> >
> > -       struct { /* extract of a clocksource struct */
> > -               int vclock_mode;
> > -               cycle_t cycle_last;
> > -               cycle_t mask;
> > -               u32     mult;
> > -               u32     shift;
> > -       } clock;
> > +       int vclock_mode;
> > +       cycle_t cycle_last;
> > +       cycle_t mask;
> > +       u32     mult;
> > +       u32     shift;
> >
> >         /* open coded 'struct timespec' */
> >         time_t          wall_time_sec;
> > diff --git a/arch/x86/kernel/vsyscall_gtod.c b/arch/x86/kernel/vsyscall_gtod.c
> > index 91862a4..ca48248 100644
> > --- a/arch/x86/kernel/vsyscall_gtod.c
> > +++ b/arch/x86/kernel/vsyscall_gtod.c
> > @@ -16,6 +16,18 @@
> >
> >  DEFINE_VVAR(struct vsyscall_gtod_data, vsyscall_gtod_data);
> >
> > +static inline void gtod_write_begin(unsigned *s)
> > +{
> > +       ++*s;
> > +       smp_wmb();
> > +}
> > +
> > +static inline void gtod_write_end(unsigned *s)
> > +{
> > +       smp_wmb();
> > +       ++*s;
> > +}
> > +
> 
> Someone else should probably comment on the style for this.  Maybe
> this should live in a header somewhere.  IMO if it's called
> gtod_write_begin, it should take a pointer to gtod as a parameter.
> 

Fixed in the next version.

> 
> > +struct api_timeval {
> > +       long    tv_sec;         /* seconds */
> > +       long    tv_usec;        /* microseconds */
> > +};
> > +
> > +struct api_timespec {
> > +       long    tv_sec;         /* seconds */
> > +       long    tv_nsec;        /* nanoseconds */
> > +};
> 
> Did you address my question about why there are two versions of this?
> Shouldn't it just match the userspace headers regardless of what the
> host kernel is?
> 

Again, -m32 compilation. The timeval and timespec are __kernel_long_t,
which is defined as "long long" in a case of 32 bit VDSO for 64 bit
kernel.

The vsyscall_gtod_data use the kernel defined timeval and timespec, so
for 32 bit a conversation is needed.

The only way to prevent this kind of hack is to replace the struct
timespec and time_t members of vsyscall_gtod_data by a compiler independ
definition.

In this case it would be also necessary to have an own copy of
timespec_add_ns() function for the VDSO.

> You have *still* not responded to my objection to the unconditional fixmaps.
> 

I have responded this objection on monday. If i enable the sysctl
interface i need this mapping too.

- Stefani


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/