Hi Bob, You prompted me to google around a bit more for valid use cases of MAP_FIXED and here is an enlightening StackOverflow post: https://stackoverflow.com/questions/28575893/when-would-one-use-mmap-map-fixed So apparently the Autoconf's test doesn’t do anything wrong, but valid use cases for MAP_FIXED are indeed very special. Is it fair to discard the mmap implementation completely, just because MAP_FIXED is failing? Maybe define something like HAVE_MMAP_MAP_FIXED instead (or not?)… Happy to do any tests on AIX if helpful. All the best, Yury > On 28. Jul 2024, at 21:18, Bob Friesenhahn <graphicsmagick.project@xxxxxxxxx> wrote: > > I would not expect MAP_FIXED to be portable or a normal requirement. Most common uses of memory mapping (e.g. file mapping) do not require a fixed mapping. Perhaps device drivers with a user-space component (or applications which store pointers in shared memory) might desire a fixed mapping > > Bob > On Jul 28, 2024, at 8:11 AM, "Yury V. Zaytsev" <yury@xxxxxxxxxx> wrote: Hi there, > > I’m trying to wrap my head around AC_FUNC_MMAP failing on AIX (specifically, version 7.3 is what I tried on GCC compile farm). It appears that it has been failing for decades though, so the problem isn’t exactly new... > > It returns exit code 10, which means that the following comparison fails: > > data2 != mmap (data2, pagesize, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_FIXED, fd, 0L); > > If I check the return value of mmap it equals to MAP_FAILED. > > Now, the question is, why does it fail and is it supposed to fail or not? Here are the IBM docs on the subject: > > https://www.ibm.com/docs/en/aix/7.3?topic=m-mmap-mmap64-subroutine > > It says > > “A region is never placed ... at an address where it would overlap an existing region.” > > (Linux has a similar comment in the man pages: “If the memory region ... overlaps pages of any existing mapping(s), then the overlapped part of the existing > mapping(s) will be discarded.” - macOS simply says MAP_FIXED is discouraged.) > > This is the part that is confusing me. The test says explicitly: > > “Next, try to mmap the file at a fixed address which already has something else allocated at it.” > > To me this sounds like it’s testing something, which actually shouldn’t work. > > However, on macOS at least it works (and obviously on other common systems, otherwise people would have noticed a long time ago) - so maybe I’m misunderstanding the comment. > > If I remove MAP_FIXED, then the rest works (including data consistency checks). So apparently AIX has a working mmap implementation, except for this particular issue / flag. > > Also it doesn’t just do some random nonsense, but returns MAP_FAILED. Thus well-written software should be checking for it anyway and act accordingly. The errorno is set to ENOMEM (“Not enough space”). > > How would you go about that? I see the check has been adjusted several times lately, so hopefully Paul or Zack could chime in? > > Is it a bug that IBM should fix? Obviously, didn’t happen so far, even though it has been the case for a very long while. Anyone has contacts and/or works for IBM? > > Or maybe the test is not good and should be adjusted by either removing the MAP_FIXED flag or still succeeding upon MAP_FAILED? > > P.S. I’m not subscribed to the list, so please keep me on CC. > > All the best, > Yury > > > > >
