On Wed, Apr 29, 2020 at 11:42 AM Christoph Hellwig [off-list ref] wrote:

On Wed, Apr 29, 2020 at 10:07:11AM +0200, Arnd Bergmann wrote:

quoted

quoted

What do you think of this version?  This one always overrides
copy_siginfo_to_user32 for the x86 compat case to keep the churn down,
and improves the copy_siginfo_to_external32 documentation a bit.

Looks good to me. I preferred checking for X32 explicitly (so we can
find and kill off the #ifdef if we ever remove X32 for good), but there is
little difference in the end.

Is there any realistic chance we'll get rid of x32?

When we discussed it last year, there were a couple of users that replied
saying they actively use it for a full system, and some others said they run
specific programs built as x32 as it results in much faster (10% to 20%)
execution of the same binaries compared to either i686 or x86_64.

I expect both of these to get less common over time as stuff bitrots
and more of the workloads that benefit most from the higher
performance (cross-compilers, hpc) run out of virtual address space.
Debian popcon numbers are too small to be reliable but they do show
a trend at https://popcon.debian.org/stat/sub-x32.png

I would just ask again every few years, and eventually we'll decide
it's not worth keeping any more. I do expect most 32-bit machines
to stop getting kernel updates before 2030 and we can probably
remove a bunch of architectures including x32 before then, though
at least armv7 users will have to get kernel updates for substantially
longer.

      Arnd