Re: [PATCH] xfs_io: Make HAVE_MAP_SYNC more robust

[Florian Fainelli <f.fainelli@gmail.com>]

On 7/20/22 14:32, Darrick J. Wong wrote:
> On Wed, Jul 20, 2022 at 01:53:07PM -0700, Florian Fainelli wrote:
>> MIPS platforms building with recent kernel headers and the musl-libc toolchain
>> will expose the following build failure:
>>
>> mmap.c: In function 'mmap_f':
>> mmap.c:196:12: error: 'MAP_SYNC' undeclared (first use in this function); did you mean 'MS_SYNC'?
>>   196 |    flags = MAP_SYNC | MAP_SHARED_VALIDATE;
>>       |            ^~~~~~~~
>>       |            MS_SYNC
>> mmap.c:196:12: note: each undeclared identifier is reported only once for each function it appears in
>> make[4]: *** [../include/buildrules:81: mmap.o] Error 1
> 
> Didn't we already fix this?
> https://lore.kernel.org/linux-xfs/20220508193029.1277260-1-fontaine.fabrice@gmail.com/
> 
> Didn't we already fix this?
> https://lore.kernel.org/linux-xfs/20181116162346.456255382F@mx7.valuehost.ru/

First off, I was not aware of these two proposed solutions so apologies for submitting a third (are there more) one.

The common problem I see with both proposed patches is that they specifically tackle io/mmap.c when really any inclusion order of linux.h and sys/mman.h could and would lead to the the same problem to re-surface somewhere else in a different file. So sure enough there is only mmap.c now, but it has been identified that this specific include order is problematic so we ought to address it in a general way.

> 
> Oh, I guess the maintainer didn't apply either of these patches, so this
> has been broken for years.

Fabrice's patch is only a few months old, and but "info@mobile-stream.com"'s is nearly 4 years old...

> 
> Well... MAP_SYNC has been with us for a while now, perhaps it makes more
> sense to remove all the override cruft and make xfs_io not export -S if
> if neither kernel headers nor libc define it?

Sure that would work too, but we will still end up with some MAP_SYNC conditional code, so the original intent of always defining it seemed laudable.

> 
> --D
> 
>>
>> The reason for that is that the linux.h header file which intends to provide a fallback definition for MAP_SYNC and MAP_SHARED_VALIDATE is included too early through:
>>
>> input.h -> libfrog/projects.h -> xfs.h -> linux.h and this happens
>> *before* sys/mman.h is included.
>>
>> sys/mman.h -> bits/mman.h which has a:
>>   #undef MAP_SYNC
>>
>> see: https://git.musl-libc.org/cgit/musl/tree/arch/mips/bits/mman.h#n21
>>
>> The end result is that sys/mman.h being included for the first time
>> ends-up overriding the HAVE_MAP_SYNC fallbacks.
>>
>> To remedy that, make sure that linux.h is updated to include sys/mman.h
>> such that its fallbacks are independent of the inclusion order. As a
>> consequence this forces us to ensure that we do not re-define
>> accidentally MAP_SYNC or MAP_SHARED_VALIDATE so we protect against that.
>>
>> Fixes: dad796834cb9 ("xfs_io: add MAP_SYNC support to mmap()")
>> Signed-off-by: Florian Fainelli <f.fainelli@gmail.com>
>> ---
>>  include/linux.h | 5 +++++
>>  1 file changed, 5 insertions(+)
>>
>> diff --git a/include/linux.h b/include/linux.h
>> index 3d9f4e3dca80..c3cc8e30c677 100644
>> --- a/include/linux.h
>> +++ b/include/linux.h
>> @@ -252,8 +252,13 @@ struct fsxattr {
>>  #endif
>>  
>>  #ifndef HAVE_MAP_SYNC
>> +#include <sys/mman.h>
>> +#ifndef MAP_SYNC
>>  #define MAP_SYNC 0
>> +#endif
>> +#ifndef MAP_SHARED_VALIDATE
>>  #define MAP_SHARED_VALIDATE 0
>> +#endif
>>  #else
>>  #include <asm-generic/mman.h>
>>  #include <asm-generic/mman-common.h>
>> -- 
>> 2.25.1
>>


-- 
Florian