On Mon, Jul 29, 2024, at 9:17 AM, Bob Friesenhahn wrote: >> I suppose that an app that needs mmap but not MAP_FIXED (which is >> allowed by POSIX sans XSI) could write its own special-purpose >> Autoconf macro for that. That being said, if you're planning to use >> mmap on AIX good luck; you may need it[5]. > The complaints seem to primarily be about archaic systems. > GraphicsMagick optionally uses mmap() for some use cases, where there is > substantial benefit. There have not been any complaints about it, > including under IBM's AIX. At one time (long ago) I noticed that Linux > was the worst performer out of the then current OSs when using mmap() > for file I/O. Behavior is subject to change. > > Regardless, using mmap() for file I/O is no magic bullet. It transfers > all control to the OS, and the hardware. The program may be > de-scheduled (waiting for page-in, or even page-out) when it could > otherwise be doing something useful. It's not really any of autoconf's business what a program wants to use mmap() for, _except_ insofar as it affects how autoconf probes for mmap. I'd guess that modern software uses mmap _without_ MAP_FIXED much more often than it uses it _with_ MAP_FIXED, which makes autoconf's insistence on checking for XSI semantics of MAP_FIXED in its only "check for mmap" macro a little silly. I don't want to change what AC_FUNC_MMAP does because it seems like that could break under-maintained old programs that _do_ expect AC_FUNC_MMAP to check for MAP_FIXED. However, perhaps we could have a new family of AC_FUNC_MMAP_* macros that check for specific subsets of mmap functionality that are generally useful, off the top of my head - check for the basic functionality of mapping files at kernel-selected locations (this can probably just invoke AC_CHECK_HEADERS_ONCE([sys/mman.h]) and AC_CHECK_FUNCS_ONCE([mmap])) - figure out the right way to call it to request anonymous memory at a kernel-selected location (MAP_ANON vs MAP_ANONYMOUS vs /dev/zero) - detect whether MAP_FIXED will replace an existing mapping, and/or whether MAP_FIXED_NOREPLACE exists We should probably also document the AIX issues in the "writing portable C programs" section of the manual if we don't already. Can any of you think of other subsets of mmap functionality that are worth supplying a test for in autoconf? Do any of you have time to work on patches for any or all of this? zw
