On 13/10/2020 21.49, Jens Axboe wrote:
> On 10/13/20 1:46 PM, Linus Torvalds wrote:
>> On Mon, Oct 12, 2020 at 6:46 AM Jens Axboe <axboe@xxxxxxxxx> wrote:
>>>
>>> Here are the io_uring updates for 5.10.
>>
>> Very strange. My clang build gives a warning I've never seen before:
>>
>> /tmp/io_uring-dd40c4.s:26476: Warning: ignoring changed section
>> attributes for .data..read_mostly
>>
>> and looking at what clang generates for the *.s file, it seems to be
>> the "section" line in:
>>
>> .type io_op_defs,@object # @io_op_defs
>> .section .data..read_mostly,"a",@progbits
>> .p2align 4
>>
>> I think it's the combination of "const" and "__read_mostly".
>>
>> I think the warning is sensible: how can a piece of data be both
>> "const" and "__read_mostly"? If it's "const", then it's not "mostly"
>> read - it had better be _always_ read.
>>
>> I'm letting it go, and I've pulled this (gcc doesn't complain), but
>> please have a look.
>
> Huh weird, I'll take a look. FWIW, the construct isn't unique across
> the kernel.
Citation needed. There's lots of "pointer to const foo" stuff declared
as __read_mostly, but I can't find any objects that are themselves both
const and __read_mostly. Other than that io_op_defs and io_uring_fops now.
But... there's something a little weird:
$ grep read_most -- fs/io_uring.s
.section .data..read_mostly,"a",@progbits
$ readelf --wide -S fs/io_uring.o | grep read_most
[32] .data..read_mostly PROGBITS 0000000000000000 01b4e0 000188
00 WA 0 0 32
(this is with gcc/gas). So despite that .section directive not saying
"aw", the section got the W flag anyway. There are lots of
.section "__tracepoints_ptrs", "a"
.pushsection .smp_locks,"a"
in the .s file, and those sections do end up with just the A bit in the
.o file. Does gas maybe somehow special-case a section name starting
with .data?
Rasmus
