Hello,
I came along some software modules where I suggested source code adjustments.
Example:
nfs/write: Use common error handling code in nfs_lock_and_join_requests()
https://lkml.org/lkml/2017/11/7/599
https://patchwork.kernel.org/patch/10047013/
https://lkml.kernel.org/r/<[email protected]>
I would like to check corresponding build results then without extra
optimisation applied by the compiler.
But I got surprised by error messages for a command like the following.
elfring@Sonne:~/Projekte/Linux/next-patched> my_cc=/usr/bin/gcc-7 && LANG=C make -j4 CC="${my_cc}" HOSTCC="${my_cc}" EXTRA_CFLAGS='-O0' allmodconfig fs/nfs/write.o
…
In file included from ./include/linux/compiler.h:58:0,
from ./include/uapi/linux/stddef.h:1,
from ./include/linux/stddef.h:4,
from ./include/uapi/linux/posix_types.h:4,
from ./include/uapi/linux/types.h:13,
from ./include/linux/types.h:5,
from fs/nfs/write.c:9:
./arch/x86/include/asm/jump_label.h: In function ‘trace_nfs_writeback_page_enter’:
./include/linux/compiler-gcc.h:275:38: warning: asm operand 0 probably doesn’t match constraints
#define asm_volatile_goto(x...) do { asm goto(x); asm (""); } while (0)
…
How do you think about to improve this software situation anyhow?
Regards,
Markus
On Sun, 2017-12-03 at 15:15 +0100, SF Markus Elfring wrote:
> Hello,
>
> I came along some software modules where I suggested source code
> adjustments.
>
> Example:
> nfs/write: Use common error handling code in
> nfs_lock_and_join_requests()
>
> https://lkml.org/lkml/2017/11/7/599
> https://patchwork.kernel.org/patch/10047013/
> https://lkml.kernel.org/r/<7f072f78-eef4-6d87-d233-cee71dac5a32@users
> .sourceforge.net>;
>
> I would like to check corresponding build results then without extra
> optimisation applied by the compiler.
> But I got surprised by error messages for a command like the
> following.
>
> elfring@Sonne:~/Projekte/Linux/next-patched> my_cc=/usr/bin/gcc-7 &&
> LANG=C make -j4 CC="${my_cc}" HOSTCC="${my_cc}" EXTRA_CFLAGS='-O0'
> allmodconfig fs/nfs/write.o
> …
> In file included from ./include/linux/compiler.h:58:0,
> from ./include/uapi/linux/stddef.h:1,
> from ./include/linux/stddef.h:4,
> from ./include/uapi/linux/posix_types.h:4,
> from ./include/uapi/linux/types.h:13,
> from ./include/linux/types.h:5,
> from fs/nfs/write.c:9:
> ./arch/x86/include/asm/jump_label.h: In function
> ‘trace_nfs_writeback_page_enter’:
> ./include/linux/compiler-gcc.h:275:38: warning: asm operand 0
> probably doesn’t match constraints
> #define asm_volatile_goto(x...) do { asm goto(x); asm (""); } while
> (0)
> …
>
>
> How do you think about to improve this software situation anyhow?
I'm not seeing anything obviously wrong with the NFS use of tracepoints
there, and the warning suggests rather that gcc has an issue with the
inlined assembly code in jump_label.h.
Ccing Peter Zijlstra (who appears to have been the last person to touch
that assembly code) and Steven Rostedt.
--
Trond Myklebust
Linux NFS client maintainer, PrimaryData
[email protected]
On Sun, 3 Dec 2017 15:17:32 +0000
Trond Myklebust <[email protected]> wrote:
> > I would like to check corresponding build results then without extra
> > optimisation applied by the compiler.
> > But I got surprised by error messages for a command like the
> > following.
> >
> > elfring@Sonne:~/Projekte/Linux/next-patched> my_cc=/usr/bin/gcc-7 &&
> > LANG=C make -j4 CC="${my_cc}" HOSTCC="${my_cc}" EXTRA_CFLAGS='-O0'
> > allmodconfig fs/nfs/write.o
Why would you compile the kernel without optimization? There's many
places in the kernel that WILL NOT BUILD without optimization. In fact,
we do a lot of tricks to make sure that things work the way we expect
it to, because we add broken code that only gets compiled out when gcc
optimizes the code the way we expect it to be, and the kernel build
will break otherwise.
-- Steve
> Why would you compile the kernel without optimization?
I would like to see how big an effect finally is in such a build configuration
after specific source code adjustments.
> There's many places in the kernel that WILL NOT BUILD without optimization.
I did not really know this detail so far.
I noticed that the optimised build variants worked during my test comparisons.
> In fact, we do a lot of tricks to make sure that things work the way
> we expect it to, because we add broken code that only gets compiled out
> when gcc optimizes the code the way we expect it to be,
> and the kernel build will break otherwise.
Thanks for your information.
Can the software areas distinguished where such special handling matters?
Regards,
Markus
On Sun, 3 Dec 2017 22:56:51 +0100
SF Markus Elfring <[email protected]> wrote:
> Can the software areas distinguished where such special handling matters?
No idea. That's something you are going to have to figure out on your
own.
-- Steve
> Why would you compile the kernel without optimization?
Can another reason be occasionally still relevant?
Will the compilation be a bit quicker when extra data processing
could be omitted?
> There's many places in the kernel that WILL NOT BUILD without optimization.
Would you like to keep the software situation in this way?
> In fact, we do a lot of tricks to make sure that things work the way
> we expect it to, because we add broken code that only gets compiled out
> when gcc optimizes the code the way we expect it to be,
> and the kernel build will break otherwise.
* Can this goal be also achieved without the addition of “broken code”?
* How do you think about to improve the error handling there?
Regards,
Markus
On Mon, 4 Dec 2017 10:00:54 +0100
SF Markus Elfring <[email protected]> wrote:
> > Why would you compile the kernel without optimization?
>
> Can another reason be occasionally still relevant?
No.
>
> Will the compilation be a bit quicker when extra data processing
> could be omitted?
Why would you care more about the time it takes to compile the kernel,
than the time it takes for executing it? Benchmarks are all about
performance of a running kernel, nobody compares benchmarks of the time
it takes to compile it. Sure, we like to make the compile times quicker
(heck, I wrote "make localmodconfig" for just that purpose), but we
never favor compiler time over execution time. In fact, if we can
improve the execution performance by sacrificing compile time, we are
happy to do that.
>
>
> > There's many places in the kernel that WILL NOT BUILD without optimization.
>
> Would you like to keep the software situation in this way?
Yes.
>
>
> > In fact, we do a lot of tricks to make sure that things work the way
> > we expect it to, because we add broken code that only gets compiled out
> > when gcc optimizes the code the way we expect it to be,
> > and the kernel build will break otherwise.
>
> * Can this goal be also achieved without the addition of “broken code”?
No.
>
> * How do you think about to improve the error handling there?
It works just fine as is. Errors that can be detected at build time are
100 times better than detecting them at execution time.
-- Steve
>> Can the software areas distinguished where such special handling matters?
>
> No idea.
I would like to point another example out.
> That's something you are going to have to figure out on your own.
How do you think about information from an other clarification request
for the topic “caif: Use common error handling code in cfdgml_receive()”?
https://lkml.org/lkml/2017/11/7/486
https://patchwork.kernel.org/patch/10046849/
https://lkml.kernel.org/r/<[email protected]>
Regards,
Markus
>> Will the compilation be a bit quicker when extra data processing
>> could be omitted?
>
> Why would you care more about the time it takes to compile the kernel,
> than the time it takes for executing it?
I am also interested in the evolution of compilation time frames.
> Benchmarks are all about performance of a running kernel,
This is generally reasonable.
> nobody compares benchmarks of the time it takes to compile it.
I guess that the situation can be occasionally different there.
> Sure, we like to make the compile times quicker
Good to know …
> (heck, I wrote "make localmodconfig" for just that purpose),
Thanks.
> but we never favor compiler time over execution time.
I imagine that the speed expectations could be adjusted during software development,
couldn't they?
> In fact, if we can improve the execution performance by sacrificing compile time,
> we are happy to do that.
I guess that you would like to consider some constraints there.
>>> In fact, we do a lot of tricks to make sure that things work the way
>>> we expect it to, because we add broken code that only gets compiled out
>>> when gcc optimizes the code the way we expect it to be,
>>> and the kernel build will break otherwise.
>>
>> * Can this goal be also achieved without the addition of “broken code”?
>
> No.
Will any other contributors take another look?
>> * How do you think about to improve the error handling there?
>
> It works just fine as is.
I hope that further software improvements can be achieved also for this use case.
> Errors that can be detected at build time are 100 times better
> than detecting them at execution time.
I agree to such a general view.
Will an other (or no) error message be more appropriate?
Regards,
Markus