On Fri, 26 Jul 2019, Alexandre Ghiti wrote:
> On 7/26/19 2:20 AM, Paul Walmsley wrote:
> >
> > On Wed, 24 Jul 2019, Alexandre Ghiti wrote:
> >
> > > In order to avoid wasting user address space by using bottom-up mmap
> > > allocation scheme, prefer top-down scheme when possible.
> > >
> > > Before:
> > > root@qemuriscv64:~# cat /proc/self/maps
> > > 00010000-00016000 r-xp 00000000 fe:00 6389 /bin/cat.coreutils
> > > 00016000-00017000 r--p 00005000 fe:00 6389 /bin/cat.coreutils
> > > 00017000-00018000 rw-p 00006000 fe:00 6389 /bin/cat.coreutils
> > > 00018000-00039000 rw-p 00000000 00:00 0 [heap]
> > > 1555556000-155556d000 r-xp 00000000 fe:00 7193 /lib/ld-2.28.so
> > > 155556d000-155556e000 r--p 00016000 fe:00 7193 /lib/ld-2.28.so
> > > 155556e000-155556f000 rw-p 00017000 fe:00 7193 /lib/ld-2.28.so
> > > 155556f000-1555570000 rw-p 00000000 00:00 0
> > > 1555570000-1555572000 r-xp 00000000 00:00 0 [vdso]
> > > 1555574000-1555576000 rw-p 00000000 00:00 0
> > > 1555576000-1555674000 r-xp 00000000 fe:00 7187 /lib/libc-2.28.so
> > > 1555674000-1555678000 r--p 000fd000 fe:00 7187 /lib/libc-2.28.so
> > > 1555678000-155567a000 rw-p 00101000 fe:00 7187 /lib/libc-2.28.so
> > > 155567a000-15556a0000 rw-p 00000000 00:00 0
> > > 3fffb90000-3fffbb1000 rw-p 00000000 00:00 0 [stack]
> > >
> > > After:
> > > root@qemuriscv64:~# cat /proc/self/maps
> > > 00010000-00016000 r-xp 00000000 fe:00 6389 /bin/cat.coreutils
> > > 00016000-00017000 r--p 00005000 fe:00 6389 /bin/cat.coreutils
> > > 00017000-00018000 rw-p 00006000 fe:00 6389 /bin/cat.coreutils
> > > 2de81000-2dea2000 rw-p 00000000 00:00 0 [heap]
> > > 3ff7eb6000-3ff7ed8000 rw-p 00000000 00:00 0
> > > 3ff7ed8000-3ff7fd6000 r-xp 00000000 fe:00 7187 /lib/libc-2.28.so
> > > 3ff7fd6000-3ff7fda000 r--p 000fd000 fe:00 7187 /lib/libc-2.28.so
> > > 3ff7fda000-3ff7fdc000 rw-p 00101000 fe:00 7187 /lib/libc-2.28.so
> > > 3ff7fdc000-3ff7fe2000 rw-p 00000000 00:00 0
> > > 3ff7fe4000-3ff7fe6000 r-xp 00000000 00:00 0 [vdso]
> > > 3ff7fe6000-3ff7ffd000 r-xp 00000000 fe:00 7193 /lib/ld-2.28.so
> > > 3ff7ffd000-3ff7ffe000 r--p 00016000 fe:00 7193 /lib/ld-2.28.so
> > > 3ff7ffe000-3ff7fff000 rw-p 00017000 fe:00 7193 /lib/ld-2.28.so
> > > 3ff7fff000-3ff8000000 rw-p 00000000 00:00 0
> > > 3fff888000-3fff8a9000 rw-p 00000000 00:00 0 [stack]
> > >
> > > Signed-off-by: Alexandre Ghiti <alex@xxxxxxxx>
> > > Reviewed-by: Christoph Hellwig <hch@xxxxxx>
> > > Reviewed-by: Kees Cook <keescook@xxxxxxxxxxxx>
> > > ---
> > > arch/riscv/Kconfig | 11 +++++++++++
> > > 1 file changed, 11 insertions(+)
> > >
> > > diff --git a/arch/riscv/Kconfig b/arch/riscv/Kconfig
> > > index 59a4727ecd6c..6a63973873fd 100644
> > > --- a/arch/riscv/Kconfig
> > > +++ b/arch/riscv/Kconfig
> > > @@ -54,6 +54,17 @@ config RISCV
> > > select EDAC_SUPPORT
> > > select ARCH_HAS_GIGANTIC_PAGE
> > > select ARCH_WANT_HUGE_PMD_SHARE if 64BIT
> > > + select ARCH_WANT_DEFAULT_TOPDOWN_MMAP_LAYOUT if MMU
> > > + select HAVE_ARCH_MMAP_RND_BITS
> > > +
> > > +config ARCH_MMAP_RND_BITS_MIN
> > > + default 18
> > Could you help me understand the rationale behind this constant?
>
>
> Indeed, I took that from arm64 code and I did not think enough about it:
> that's great you spotted this because that's a way too large value for
> 32 bits as it would, at minimum, make mmap random offset go up to 1GB
> (18 + 12), which is a big hole for this small address space :)
>
> arm and mips propose 8 as default value for 32bits systems which is 1MB offset
> at minimum.
8 seems like a fine minimum for Sv32.
> > > +
> > > +# max bits determined by the following formula:
> > > +# VA_BITS - PAGE_SHIFT - 3
> > I realize that these lines are probably copied from arch/arm64/Kconfig.
> > But the rationale behind the "- 3" is not immediately obvious. This
> > apparently originates from commit 8f0d3aa9de57 ("arm64: mm: support
> > ARCH_MMAP_RND_BITS"). Can you provide any additional context here?
>
>
> The formula comes from commit d07e22597d1d ("mm: mmap: add new /proc
> tunable for mmap_base ASLR"), where the author states that "generally a
> 3-4 bits less than the number of bits in the user-space accessible
> virtual address space [allows to] give the greatest flexibility without
> generating an invalid mmap_base address".
>
> In practice, that limits the mmap random offset to at maximum 1/8 (for -
> 3) of the total address space.
OK.
> > > +config ARCH_MMAP_RND_BITS_MAX
> > > + default 33 if 64BIT # SV48 based
> > The rationale here is clear for Sv48, per the above formula:
> >
> > (48 - 12 - 3) = 33
> >
> > > + default 18
> > However, here it is less clear to me. For Sv39, shouldn't this be
> >
> > (39 - 12 - 3) = 24
> >
> > ? And what about Sv32?
>
>
> You're right. Is there a way to distinguish between sv39 and sv48 here ?
This patch has just been posted:
https://lore.kernel.org/linux-riscv/alpine.DEB.2.21.9999.1907261259420.26670@xxxxxxxxxxxxxxxx/T/#u
Assuming there are no negative comments, we'll plan to send it upstream
during v5.3-rc. Your patch should be able to set different minimums and
maximums based on the value of CONFIG_RISCV_VM_SV*
- Paul
