Re: mmap and MAP_FIXED

[Daniel Jour <daniel.oertwig%gmail.com@localhost>]

Michael wrote:
> I don't see a reason to do it at all - the pointer to it will sit in a
> variable anyway, so what does it matter where in the address space it
> is? Maybe they're trying to force the start to be aligned to something
> bigger than PAGE_SIZE?

It doesn't have to be in a variable, in theory it could be optimized
away by the compiler, improving speed of access to contents with also
known offset. I don't know (yet) why it's done in CLISP, but that's a
somewhat plausible reason.


J. Hannken-Illjes wrote:
> You try a value above VM_MAXUSER_ADDRESS:
>
>    0x0004000000000000 > 0x00007f8000000000

Ah, thank you a lot. Are these addresses (VM_MAXUSER_ADDRESS and
VM_MIN_ADDRESS) somehow exposed in a header or queryable somehow?
Then I could let configure give the job to decide which address to
choose.


Dan wrote:
> MMAP_FIXED forces mmap to use the addr given or fail; without it, it
> can act as if addr=0, and provide its own address.  So, result might
> != addr.  Don't know if it would make a difference, but you could try
> w/o fixed, and changing to see if result != addr,  (and unmapping )
> Im guessing CLISP is specifying the addr so it can use the same address
> each run, for space for cons-cells and such?  And machine code refers?
> to the address? Could you pass 0 as address, letting mmap pick the address,
> then save result address somehow then, recompile with FIXED, and
> that address?

That's probably what I do in case the relevant address limits aren't
somehow exposed.


Johnny wrote:
> I think that is hardware dependent. The OS might be somewhere in the
> address space as well, just not accessible to user mode, for example.
> I would not expect that to show up in the map. There might also be
> limitations to what addresses actually can be mapped by the hardware.
>
> Which are additional reasons why it's probably better to let the OS
> give you an address instead you you trying to pick one. Depending on the
> hardware, there might actually be different ranges of addresses that you
> will have issues with.

That's a very good point.


Kamil wrote:
> I think you are probably aware of it, but CLISP already exists in
> pkgsrc/lang/clist and works (builds) on NetBSD.

Yes, I'm looking at what patches to CLISP are available "downstream",
because ...


Kamil wrote:
> Are you working on upstreaming of the local patches? There has been
> accepted a GSoC project to make the CLISP release (the latest one 2.49
> was in 2010), it should be the appropriate time for it.

... I'm the student working on the new release :)


Robert wrote:
> Both CMUCL and SBCL use the MAP_FIXED option to mmap(2) on NetBSD, it
> works fine.
>
> Maybe try similar addresses to SBCL for CLISP, the core file for SBCL
> is mapped starting at 0x1000000000.

Thank you. That would have been the next thing to look at for me. As said
above, I hope to be able to use VM_MIN_ADDRESS and VM_MAXUSER_ADDRESS.


Best,
Daniel