On Mon, 13 Mar 2017, Dmitry Safonov wrote:
On 03/13/2017 12:39 PM, Thomas Gleixner wrote:
On Mon, 6 Mar 2017, Dmitry Safonov wrote:
Result of mmap() calls with MAP_32BIT flag at this moment depends
on thread flag TIF_ADDR32, which is set during exec() for 32-bit apps.
It's broken as the behavior of mmap() shouldn't depend on exec-ed
application's bitness. Instead, it should check the bitness of mmap()
syscall.
How it worked before:
o for 32-bit compatible binaries it is completely ignored. Which was
fine when there were one mmap_base, computed for 32-bit syscalls.
After introducing mmap_compat_base 64-bit syscalls do use computed
for 64-bit syscalls mmap_base, which means that we can allocate 64-bit
address with 64-bit syscall in application launched from 32-bit
compatible binary. And ignoring this flag is not expected behavior.
Well, the real question here is, whether we should allow 32bit applications
to obtain 64bit mappings at all. We can very well force 32bit applications
into the 4GB address space as it was before your mmap base splitup and be
done with it.
Hmm, yes, we could restrict 32bit applications to 32bit mappings only.
But the approach which I tried to follow in the patches set, it was do
not base the logic on the bitness of launched applications
(native/compat) - only base on bitness of the performing syscall.
The idea was suggested by Andy and I made mmap() logic here independent
from original application's bitness.
It also seems to me simpler:
if 32-bit application wants to allocate 64-bit mapping, it should
long-jump with 64-bit segment descriptor and do `syscall` instruction
for 64-bit syscall entry path. So, in my point of view after this dance
the application does not differ much from native 64-bit binary and can
have 64-bit address mapping.
Works for me, but it lacks documentation .....