6

I use an anonymous mmap to allocate a giant chunk of memory. There are several contiguous pages in this that I'd like to turn into a ring buffer, using virtual memory mirroring.

This example on Wikipedia shows what I'm meaning by virtual memory mirroring.

Say the first 14 blocks below are the pages in my giant chunk. I'd like to virtually map pages 6 and 7 to another two consecutive locations.

[0][1][2][3][4][5][6][7][8][9][10][11][12][13].......[6][7][6][7]

Mike Ash gives a rundown of what I want to do, but using mach specific APIs.

How can this be done on Linux?

kasperd
  • 1,952
  • 1
  • 20
  • 31
Max
  • 2,760
  • 1
  • 28
  • 47
  • What's wrong with that example on Wikipedia? – pentadecagon Jan 14 '14 at 12:04
  • @pentadecagon I already have a big chunk of memory, a couple of pages from within I want to mirror. The Wikipedia example requires allocating new memory. – Max Jan 14 '14 at 12:06
  • 1
    Not really, all it requires is a file descriptor that it can `mmap` twice, it doesn't care where it came from. The idea being that you want the complete buffer to be mirrored (which is obviously what _most_ people want). If you only want to mirror a part of the memory block, simply use something other than zero for `offset`. The only other possible way of mirroring pages is using `remap_file_pages`, and this, too, requires a memory mapping (so you gain nothing, really). You lose something though... portability. – Damon Jan 14 '14 at 15:29
  • 1
    @Damon thanks, `remap_file_pages` is what I was looking for – Max Jan 14 '14 at 16:10
  • @max @damon Alas, `remap_file_pages` is marked deprecated in recent Linux versions. Probably better to use `mmap`. – Chromatix May 20 '18 at 01:39
  • @Max here is what you were looking for https://github.com/vitalyvch/rng_buf – vitaly.v.ch Feb 02 '20 at 14:26

1 Answers1

0

This is not possible to achieve with anonymous mappings. You can, however, use memfd_create(2) or create a temporary file on tmpfs (so that it won't ever be stored to disk, except by "swapping") and do shared mappings on that file descriptor, adjacent to each other.

remap_file_pages(2) is just a performance hack and not any more powerful than classic mmap/mremap calls, and it can't operate on anonymous mappings either.

Related: Linux mremap without freeing the old mapping?

Jin-oh Kang
  • 317
  • 1
  • 8