Question about mmap syscall and POSIX standard on mips arch

Question about mmap syscall and POSIX standard on mips arch

[Hongzhi, Song]

Hi all,

Ltp has a POSIX teatcase about mmap, 24-2.c.

https://github.com/linux-test-project/ltp/blob/e816127e5d8efbff5ae53e9c2292fae22f36838b/testcases/open_posix_testsuite/conformance/interfaces/mmap/24-2.c#L94

-----part of code-----

     pa = mmap(addr, len, PROT_READ | PROT_WRITE, MAP_FIXED | 
MAP_SHARED, fd,
0);
     if (pa == MAP_FAILED && errno == ENOMEM) {
         printf("Got ENOMEM: %s\nTest PASSED\n", strerror(errno));
exit(PTS_PASS);
     }

-----end----------------


Under POSIX standard, the expected errno should be ENOMEM

when the specific [addr+len] exceeds the bound of memory.


But mips returns EINVAL.

https://github.com/torvalds/linux/blob/bab5c80b211035739997ebd361a679fa85b39465/arch/mips/mm/mmap.c#L69

-------part of code-------

     if (flags & MAP_FIXED) {
         /* Even MAP_FIXED mappings must reside within TASK_SIZE */
         if (TASK_SIZE - len < addr)
             return -EINVAL;

-------end------------------


So, can we change EINVAL to ENOMEM to follow POSIX standard?


--Hongzhi

Re: Question about mmap syscall and POSIX standard on mips arch

[Al Viro]

[mips folks Cc'd]

On Thu, Oct 18, 2018 at 11:26:02AM +0800, Hongzhi, Song wrote:
> Hi all,
> 
> Ltp has a POSIX teatcase about mmap, 24-2.c.
> 
> https://github.com/linux-test-project/ltp/blob/e816127e5d8efbff5ae53e9c2292fae22f36838b/testcases/open_posix_testsuite/conformance/interfaces/mmap/24-2.c#L94

[basically, MAP_FIXED mmap with addr + len > TASK_SIZE fails with
-EINVAL on mips and -ENOMEM elsewhere]
 
> Under POSIX standard, the expected errno should be ENOMEM
> 
> when the specific [addr+len] exceeds the bound of memory.

The mmap() function may fail if:

[EINVAL]
The addr argument (if MAP_FIXED was specified) or off is not a multiple
of the page size as returned by sysconf(), or is considered invalid by
                                           ^^^^^^^^^^^^^^^^^^^^^^^^^^^
the implementation.
^^^^^^^^^^^^^^^^^^^

So that behaviour gets past POSIX.  That part is mostly about the
things like cache aliasing constraints, etc., but it leaves enough
space to weasel out.  Said that, this

[ENOMEM]
MAP_FIXED was specified, and the range [addr,addr+len) exceeds that allowed
for the address space of a process; or, if MAP_FIXED was not specified and
there is insufficient room in the address space to effect the mapping.

is a lot more specific, so switching to -ENOMEM there might be a good idea,
especially since on other architectures we do get -ENOMEM in that case,
AFAICS.

Re: Question about mmap syscall and POSIX standard on mips arch

[Paul Burton]

Hi Al,

On Thu, Oct 18, 2018 at 05:32:00AM +0100, Al Viro wrote:
> [mips folks Cc'd]
> 
> On Thu, Oct 18, 2018 at 11:26:02AM +0800, Hongzhi, Song wrote:
> > Hi all,
> > 
> > Ltp has a POSIX teatcase about mmap, 24-2.c.
> > 
> > https://github.com/linux-test-project/ltp/blob/e816127e5d8efbff5ae53e9c2292fae22f36838b/testcases/open_posix_testsuite/conformance/interfaces/mmap/24-2.c#L94
> 
> [basically, MAP_FIXED mmap with addr + len > TASK_SIZE fails with
> -EINVAL on mips and -ENOMEM elsewhere]
>  
> > Under POSIX standard, the expected errno should be ENOMEM
> > 
> > when the specific [addr+len] exceeds the bound of memory.
> 
> The mmap() function may fail if:
> 
> [EINVAL]
> The addr argument (if MAP_FIXED was specified) or off is not a multiple
> of the page size as returned by sysconf(), or is considered invalid by
>                                            ^^^^^^^^^^^^^^^^^^^^^^^^^^^
> the implementation.
> ^^^^^^^^^^^^^^^^^^^
> 
> So that behaviour gets past POSIX.  That part is mostly about the
> things like cache aliasing constraints, etc., but it leaves enough
> space to weasel out.  Said that, this
> 
> [ENOMEM]
> MAP_FIXED was specified, and the range [addr,addr+len) exceeds that allowed
> for the address space of a process; or, if MAP_FIXED was not specified and
> there is insufficient room in the address space to effect the mapping.
> 
> is a lot more specific, so switching to -ENOMEM there might be a good idea,
> especially since on other architectures we do get -ENOMEM in that case,
> AFAICS.

Thanks for the heads up - that does sound like reasonably clear
non-compliance. I'll make a note to put together a patch & test it out,
likely next week, if nobody submits one first.

Thanks,
    Paul