mmap(2) with MAP_ANON honouring offset although it shouldn't

[Alexander Best]

Alexander Best schrieb am 2009-11-03:
> John Baldwin schrieb am 2009-11-03:
> > On Monday 02 November 2009 5:14:27 pm Alexander Best wrote:
> > > John Baldwin schrieb am 2009-11-02:
> > > > On Monday 02 November 2009 4:05:56 pm Alexander Best wrote:
> > > > > John Baldwin schrieb am 2009-11-02:
> > > > > > On Friday 30 October 2009 10:38:24 pm Alexander Best wrote:
> > > > > > > John Baldwin schrieb am 2009-10-21:
> > > > > > > > On Wednesday 21 October 2009 11:51:04 am Alexander Best
> > > > > > > > wrote:
> > > > > > > > > although the mmap(2) manual states in section
> > > > > > > > > MAP_ANON:

> > > > > > > > > "The offset argument is ignored."

> > > > > > > > > this doesn't seem to be true. running

> > > > > > > > > printf("%p\n", mmap((void*)0x1000, 0x1000, PROT_NONE,
> > > > > > > > > MAP_ANON,
> > > > > > > > > -1,
> > > > > > > > > 0x12345678));

> > > > > > > > > and

> > > > > > > > > printf("%p\n", mmap((void*)0x1000, 0x1000, PROT_NONE,
> > > > > > > > > MAP_ANON,
> > > > > > > > > -1,
> > > > > > > > > 0));

> > > > > > > > > produces different outputs. i've attached a patch to
> > > > > > > > > solve
> > > > > > > > > the
> > > > > > > > > problem. the
> > > > > > > > > patch is similar to the one proposed in this PR, but
> > > > > > > > > should
> > > > > > > > > apply
> > > > > > > > > cleanly to
> > > > > > > > > CURRENT:
> > > > > > > > >  http://www.freebsd.org/cgi/query-pr.cgi?pr=kern/71258

> > > > > > > > A simpler patch would be to simply set pos = 0 below
> > > > > > > > the
> > > > > > > > MAP_STACK
> > > > > > > > line if
> > > > > > > > MAP_ANON is set.

> > > > > > > how about the following patch. problem seems to be that
> > > > > > > pos
> > > > > > > = 0
> > > > > > > needs to be
> > > > > > > set before pageoff is being calculated.

> > > > > > I think that that patch is fine, but will defer to alc at .  I
> > > > > > think
> > > > > > he
> > > > > > argued
> > > > > > that any non-zero offset passed to MAP_ANON should fail
> > > > > > with
> > > > > > EINVAL.

> > > > > thanks. if that's what the POSIX standard requests that's ok.
> > > > > however in that
> > > > > case we need to change the mmap(2) manual, because right now
> > > > > it
> > > > > says in
> > > > > section MAP_ANON:

> > > > > "The offset argument is ignored."

> > > > > which should be changed to something like:

> > > > > "The offset argument must be zero."

> > > > > also if the behaviour of MAP_ANON changes this also changes
> > > > > the
> > > > > semantics of
> > > > > MAP_STACK since it implies MAP_ANON. so we need to decide if
> > > > > MAP_STACK should
> > > > > silently reset any offset value to zero or like MAP_ANON
> > > > > should
> > > > > fail if offset
> > > > > isn't zero in which case the MAP_STACK section of the mmap(2)
> > > > > manual needs to
> > > > > be changed to someting like:

> > > > > "MAP_STACK implies MAP_ANON, and requires offset to be zero."

> > > > Right now MAP_STACK sets pos to 0 in the current code, and I
> > > > don't
> > > > expect we
> > > > would remove that if we decide to reject non-zero offsets for
> > > > MAP_ANON.  I'd
> > > > probably rather err on the side of leniency and just ignore the
> > > > offset rather
> > > > than rejecting non-zero, but I'm a bit burned from the last
> > > > round
> > > > of
> > > > mmap()
> > > > API changes. :)

> > > hmmm...i think this will require quite a few changes. if i
> > > remember
> > correctly
> > > MAP_STACK at some point does:

> > > flags =| MAP_ANON;

> > > so if we decide MAP_ANON and MAP_STACK should behave differently
> > > this will
> > > require some checks to distinguish between both flags further
> > > down
> > > in the
> > > code.

> > > let's see what alc@ thinks about this one then. API changes are a
> > > nasty
> > nasty
> > > business. ;)

> > Umm, if you revert your change and just add a simple clause that
> > does:

> > if (flags & MAP_ANON && pos != 0)
> >         return (EINVAL);

> > after the MAP_STACK section then I think that would work fine.  It
> > would
> > not require any further magic apart from that.

> oh. you're right. didn't think of that one. indeed this would let
> mmap fail
> with MAP_ANON and pos != 0, but would keep the current MAP_STACK
> behaviour
> (which is ignoring pos).

> sounds like a really clean and useful mmap API change. if alc@ agrees
> i could
> put your change in the form of a patch and together with a mmap(2)
> manual
> change, submit it as followup to kern/71258. it shouldn't be a big
> deal
> mfc'ing the changes to 8-stable (maybe even 8.0-release), 7-stable
> and
> 6-stable. well...better make that 8.1-release. ;) who knows what
> weird mmap
> calls are in the ports. ;)

> i'll try to build universe over the night to see if the changes break
> anything.

just realised that building universe or only world is pretty useless since the
API changes only affect apps during runtime and at compilation time. :)

i've run a few tests. the following app:

#include <sys/types.h>
#include <sys/mman.h>
#include <stdio.h>

main() {
    printf("%p\n", mmap((void*)0x1000, 0x1000, PROT_READ|PROT_WRITE,
    MAP_STACK, -1, 1));
    printf("%p\n", mmap((void*)0x1000, 0x1000, PROT_READ|PROT_WRITE, MAP_ANON,
    -1, 1));
}

outputs:

0x1000
0xffffffff

as expected.

#include <sys/types.h>
#include <sys/mman.h>
#include <stdio.h>

main() {
    printf("%p\n", mmap((void*)0, 0x1000, PROT_READ|PROT_WRITE, MAP_STACK, -1,
    0));
    printf("%p\n", mmap((void*)0, 0x1000, PROT_READ|PROT_WRITE, MAP_ANON, -1,
    0));
}

however produces this output:

0xffffffff
0x28195000

which seems a bit odd. the mmap(2) manual doesn't say anything about MAP_STACK
not working when addr is zero.

i'll see if this is caused by the changes jhb@ suggested or not.

> alex