Re: Makefile race between jobs

[Vineet Gupta]

On Monday 10 December 2012 04:29 PM, Arnd Bergmann wrote:
> On Monday 10 December 2012, Vineet Gupta wrote:
>>
>> On Monday 10 December 2012 04:00 PM, Michal Marek wrote:
>>> On 10.12.2012 11:11, Vineet Gupta wrote:
>>>> ARC Port caches current task pointer in a register - thus we have a
>>>> global asm register definition in current.h
>>>> In the past, a customer ran into issue when porting some "really
>>>> portable" code to kernel - such that asm/current.h didn't make it into
>>>> the build of their module - via normal header includes - strange but
>>>> true. Thus forcing current.h via way of -include seemed like a
>>>> safe/sensible way.
>>>
>>> To me that sounds like either an arc header is using the define but
>>> lacking an include of asm/current.h, or the code is lacking asm/current.h.
>>>
>>
>> It was latter - customer code was lacking include of asm/current.h
>> At any rate, independent of above, since we are dealing with a global
>> reg definition, IMHO, forcing the -include for each file built ensures
>> the generated code correctness (gcc reg allocator not fiddling with that
>> reg) - w/o "assuming" it would.
> 
> I'm not convinced that adding the -include to the kernel Makefile
> actually helps here: If a third party module is built outside of
> the kernel sources, it probably also does not use the kernel Makefile,
> and it may not have access to the kernel header files at all.

There might be NDA issues otherwise I would have uploaded the customer
code. It was third party code alright - but built in tree - and
interestingly not using any kernel headers - even linux/types.h. I was
dazzled myself when I found the root cause (due to r25 getting clobbered
mysteriously after some code from that driver had run).
So seriously it was really portable code.


> That aside, can't you just have an arch specific header file that
> reservers the register but does not rely on asm-generic/types.h?
> That would eliminate the need to serialize the build process.

The code in asm/current.h is as follows:

#ifdef CONFIG_ARC_CURR_IN_REG
register struct task_struct *curr_arc asm("r25");
#define current (curr_arc)
#else
#include <asm-generic/current.h>
#endif

The Makefile snippet in 3.2 (which Noam is using) is
LINUXINCLUDE	+=  -include ${src}/arch/arc/include/asm/current.h

I'm ready to change it to anything which helps alleviate the need for
build system change - w/o terribly uglifying this code.

However, in the 3.7 port which I posted to lists, the Makefile fragment
has already been changed due to other reasons.

ifdef CONFIG_ARC_CURR_IN_REG
# Can't do unconditionally because of recursive include
# issues due to <linux/thread_info.h>
LINUXINCLUDE	+=  -include ${src}/arch/arc/include/asm/current.h
endif

Does that make the problem report moot already - since the issue only
seems to happen for !CONFIG_ARC_CURR_IN_REG in which case the header
won't be -include ed

-Vineet

> 
> 	Arnd
> 

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/