Re: [PATCH 09/16] sparc64: use the generic get_user_pages_fast code

From: Anatoly Pugachev
Date: Wed Jul 24 2019 - 16:30:04 EST

On Fri, Jul 19, 2019 at 12:14 AM David Miller <davem@xxxxxxxxxxxxx> wrote:
> > So this ended up as commit 7b9afb86b6328f10dc2cad9223d7def12d60e505
> > (thanks to Anatoly for bisecting) and introduced a regression:
> > futex.test from the strace test suite now causes an Oops on sparc64
> > in futex syscall.
> >
> > Here is a heavily stripped down reproducer:
> Does not reproduce for me on a T4-2 machine.
> So this problem might depend on the type of system you are on,
> I suspect it's one of those "pre-Niagara vs. Niagara and later"
> situations because that's the dividing line between two set of
> wildly different TLB and cache management methods.
> What kind of machine are you on?


the first test where it was discovered was done on my test LDOM named
ttip, hardware (hypervisor) is T5-2 server, running under Solaris 11.4
ttip LDOM is debian sparc64 unstable , so with almost all the latest
software (gcc 8.3.0, binutils, debian GLIBC
2.28-10, etc..)

For another test, i also installed LDOM with oracle sparc linux , but I've to install a
more fresh version of gcc on it first, since system installed gcc 4.4
is too old for a git kernel (linux-2.6/Documentation/Changes lists gcc
4.6 as a minimal version), so I choose to install gcc-7.4.0 to /opt/
(leaving system installed gcc 4.4 under /usr/bin). Compiled and
installed git kernel version, i.e. last tag 5.3.0-rc1 and ran the
test. Kernel still produced oops.