On Jul 12 15:57, Brian Ford wrote: > On Wed, 12 Jul 2006, Corinna Vinschen wrote: > > > The problem results from introducing MAP_NORESERVE in 1.5.19. That's > > the reason the same code works on 1.5.18. Your code simply defines > > MAP_NORESERVE to 0 under 1.5.18. > > While that is true, I compiled the STC with MAP_NORESERVE defined under > the 1.5.21s headers, and then ran the binary without recompiling it under > 1.5.[18-20]. So, you must have ignored the MAP_NORESERVE value in the > older releases since it was passed to mmap in my test.
Right, only the correctness of the POSIX flags is checked at the start of mmap. Any non-standard flags are usually ignored. > > This might succeed on Linux, but it's not guaranteed. It certainly > > doesn't work this way on Cygwin. Call something like `mprotect (addr, > > virt_size, PROT_READ|PROT_WRITE)' before accessing the mmap'ed memory. > > Why? I already told mmap it should have these attributes. > > There is a bug in one of our understandings of MAP_NORESERVE symantics. > My understanding of MAP_NORESERVE is from the Solaris man page: > > The MAP_NORESERVE option specifies that no swap space be > reserved for a mapping. Without this flag, the creation of a > writable MAP_PRIVATE mapping reserves swap space equal to > the size of the mapping; when the mapping is written into, > the reserved space is employed to hold private copies of > the data. A write into a MAP_NORESERVE mapping produces > results which depend on the current availability of swap > space in the system. If space is available, the write > succeeds and a private copy of the written page is created; > if space is not available, the write fails and a SIGBUS or > SIGSEGV signal is delivered to the writing process. > MAP_NORESERVE mappings are inherited across fork(); at the > time of the fork(), swap space is reserved in the child for > all private pages that currently exist in the parent; > thereafter the child's mapping behaves as described above. > > So, the only difference with MAP_NORESERVE is that swap space is not > reserved up front, but on reference (read is unclear from this > description, but write is definately clear). If on reference we run out > of swap, SIGBUS is generated. > > What am I missing? I just read the Linux man page (again), and I seem to have missed some important bits. The above semantics are not implemented this way in Cygwin. A private anonymous mmap is actually just an area allocated with VirtualAlloc. A MAP_NORESERVE area is not commited (MEM_COMMIT) but just reserved (MEM_RESERVE). What's missing is the automatic commiting when a page fault on one of these pages occurs. I added an (hopefully) appropriate patch, which commits a page within a MAP_NORESERVE area when memory in this page is accessed by the application. If commiting the memory fails, SIGBUS is generated. Please test CVS HEAD or the next developer snapshot from http://cygwin.com/snapshots/ Note however, that MAP_NORESERVE is only implemented for private anonymous mappings. Shared anonymous maps seem to be possible (and would probably make sense to minimze the swap space footprint), but are somewhat tricky since it's not quite clear what happens to memory which is commited in one process and then accessed in another. This would require some extensive testing which I'm not willing to do yet. File-backed mappings are always ignoring the MAP_NORESERVE flag and are using committed memory, since uncommitted file maps are not supported on Windows. Corinna -- Corinna Vinschen Please, send mails regarding Cygwin to Cygwin Project Co-Leader cygwin AT cygwin DOT com Red Hat -- Unsubscribe info: http://cygwin.com/ml/#unsubscribe-simple Problem reports: http://cygwin.com/problems.html Documentation: http://cygwin.com/docs.html FAQ: http://cygwin.com/faq/
