Re: [PATCH 3/3] vdso: preallocate new vmas

[Davidlohr Bueso]

On Wed, 2013-10-23 at 02:53 -0700, walken@xxxxxxxxxx wrote:
> On Sun, Oct 20, 2013 at 08:26:15PM -0700, Davidlohr Bueso wrote:
> > From: Davidlohr Bueso <davidlohr@xxxxxx>
> > Subject: [PATCH v2 3/3] vdso: preallocate new vmas
> > 
> > With the exception of um and tile, architectures that use
> > the install_special_mapping() function, when setting up a
> > new vma at program startup, do so with the mmap_sem lock
> > held for writing. Unless there's an error, this process
> > ends up allocating a new vma through kmem_cache_zalloc,
> > and inserting it in the task's address space.
> > 
> > This patch moves the vma's space allocation outside of
> > install_special_mapping(), and leaves the callers to do so
> > explicitly, without depending on mmap_sem. The same goes for
> > freeing: if the new vma isn't used (and thus the process fails
> > at some point), it's caller's responsibility to free it -
> > currently this is done inside install_special_mapping.
> > 
> > Furthermore, uprobes behaves exactly the same and thus now the
> > xol_add_vma() function also preallocates the new vma.
> > 
> > While the changes to x86 vdso handling have been tested on both
> > large and small 64-bit systems, the rest of the architectures
> > are totally *untested*. Note that all changes are quite similar
> > from architecture to architecture.
> > 
> > Signed-off-by: Davidlohr Bueso <davidlohr@xxxxxx>
> > Cc: Russell King <linux@xxxxxxxxxxxxxxxx>
> > Cc: Catalin Marinas <catalin.marinas@xxxxxxx>
> > Cc: Will Deacon <will.deacon@xxxxxxx>
> > Cc: Richard Kuo <rkuo@xxxxxxxxxxxxxx>
> > Cc: Ralf Baechle <ralf@xxxxxxxxxxxxxx>
> > Cc: Benjamin Herrenschmidt <benh@xxxxxxxxxxxxxxxxxxx>
> > Cc: Paul Mackerras <paulus@xxxxxxxxx>
> > Cc: Martin Schwidefsky <schwidefsky@xxxxxxxxxx>
> > Cc: Heiko Carstens <heiko.carstens@xxxxxxxxxx>
> > Cc: Paul Mundt <lethal@xxxxxxxxxxxx>
> > Cc: Chris Metcalf <cmetcalf@xxxxxxxxxx>
> > Cc: Jeff Dike <jdike@xxxxxxxxxxx>
> > Cc: Richard Weinberger <richard@xxxxxx>
> > Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> > Cc: Ingo Molnar <mingo@xxxxxxxxxx>
> > Cc: "H. Peter Anvin" <hpa@xxxxxxxxx>
> > Cc: Peter Zijlstra <a.p.zijlstra@xxxxxxxxx>
> > ---
> > v2:
> > - Simplify install_special_mapping interface (Linus Torvalds)
> > - Fix return for uml_setup_stubs when mem allocation fails (Richard Weinberger)
> 
> I'm still confused as to why you're seeing any gains with this
> one. This code runs during exec when mm isn't shared with any other
> threads yet, so why does it matter how long the mmap_sem is held since
> nobody else can contend on it ? (well, except for accesses from
> /fs/proc/base.c, but I don't see why these would matter in your
> benchmarks either).

Yeah, that's why I dropped the performance numbers from the changelog in
v2, of course any differences are within the noise range. When I did the
initial runs I was scratching my head as to why I was seeing benefits,
but it was most likely a matter of clock frequency differences, and I no
longer see such boosts.

Thanks,
Davidlohr


--
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/