selftests/size: rework to use main instead of _start

In a -ffreestanding environment without glibc, we shouldn't be able to call either getenv or printf. It seems that printf() happens to work fine, but it ends up using the glibc getenv(), which doesn't work unless we call the glibc _start() function first.
Using _start() was originally meant as an optimization to reduce the memory consumption of the test program itself, in order to get a more accurate representation of the available RAM at system boot time. However, it causes more problems than it helps, especially when run from a shell script that also consumes some memory.

get_size[2838]: segfault at ffffffffffffffd0 ip 000000000040538b
sp 00007ffc41980668 error 5 in get_size[400000+b0000]
audit: type=1701 audit(1521532923.838:4): auid=0 uid=0 gid=0 ses=2
subj=kernel pid=2838 comm=\"get_size\" exe=\"/opt/kselftests/next/size/get_size\" sig=11 res=1
get_size[2840]: segfault at ffffffffffffffd0 ip 000000000040538b
sp 00007ffdace9b378 error 5 in get_size[400000+b0000]
audit: type=1701 audit(1521532932.057:5): auid=0 uid=0 gid=0 ses=2
subj=kernel pid=2840 comm=\"get_size\" exe=\"/opt/kselftests/next/size/get_size\" sig=11 res=1

Rework to get_size test to use main (and printf) instead of _start. All other seftest tests uses main and not _start.

Fixes 0081901af95f ("selftests: size call ksft_print_header() to print TAP header")

Signed-off-by Anders Roxell <anders.roxell at linaro.org>

Quoted 21 lines · Anders Roxell

In a -ffreestanding environment without glibc, we shouldn't be able to call either getenv or printf. It seems that printf() happens to work fine, but it ends up using the glibc getenv(), which doesn't work unless we call the glibc _start() function first.
Using _start() was originally meant as an optimization to reduce the memory consumption of the test program itself, in order to get a more accurate representation of the available RAM at system boot time. However, it causes more problems than it helps, especially when run from a shell script that also consumes some memory.

get_size[2838]: segfault at ffffffffffffffd0 ip 000000000040538b
sp 00007ffc41980668 error 5 in get_size[400000+b0000]
audit: type=1701 audit(1521532923.838:4): auid=0 uid=0 gid=0 ses=2
subj=kernel pid=2838 comm=\"get_size\" exe=\"/opt/kselftests/next/size/get_size\" sig=11 res=1
get_size[2840]: segfault at ffffffffffffffd0 ip 000000000040538b
sp 00007ffdace9b378 error 5 in get_size[400000+b0000]
audit: type=1701 audit(1521532932.057:5): auid=0 uid=0 gid=0 ses=2
subj=kernel pid=2840 comm=\"get_size\" exe=\"/opt/kselftests/next/size/get_size\" sig=11 res=1

Quoted 5 lines

Rework to get_size test to use main (and printf) instead of _start. All other seftest tests uses main and not _start.

Fixes 0081901af95f ("selftests: size call ksft_print_header() to print TAP header")

Signed-off-by Anders Roxell <anders.roxell at linaro.org>

Andres,

Thanks for finding and reporting the problem. This patch is in linux-kselftest next for 4.17-rc1. I would rather revert the patch than fix it a this time.

I knew there was a reason why Tim didn't use printf() in the first, I couldn't remember. I won't be sending

0081901af95f ("selftests: size call ksft_print_header() to print TAP header")

for inclusion into 4.17-rc1