On Tue, 7 Nov 2017 09:15:21 +0100
Florian Weimer [off-list ref] wrote:

On 11/07/2017 06:07 AM, Nicholas Piggin wrote:

quoted

First of all, using addr and MAP_FIXED to develop our heuristic can
never really give unchanged ABI. It's an in-band signal. brk() is a
good example that steadily keeps incrementing address, so depending
on malloc usage and address space randomization, you will get a brk()
that ends exactly at 128T, then the next one will be >
DEFAULT_MAP_WINDOW, and it will switch you to 56 bit address space.  

Note that this brk phenomenon is only a concern for some currently 
obscure process memory layouts where the heap ends up at the top of the 
address space.  Usually, there is something above it which eliminates 
the possibility that it can cross into the 128 TiB wilderness.  So the 
brk problem only happens on some architectures (e.g., not x86-64), and 
only with strange ways of running programs (explicitly ld.so invocation 
and likely static PIE, too).

That's true, but there was an ABI change and the result is that it
changed behaviour.

And actually if it were not for a powerpc bug that caused a segfault,
the allocation would have worked, and you probably would never have
filed the bug. However the program would have been given a pointer >
128TB, which is what we were trying to avoid -- if your app also put
metadata in the top of the pointers, it would have broken.

So, are any obscure apps that could break? I don't know. Maybe it's
quite unlikely. Is that enough to go ahead with changing behaviour?

I'm not going to put up a big fight about this if others feel it's
not a problem. I raise it because in debugging this and seeing the
changed behaviour, the differences between implementations (x86,
powerpc hash, and powerpc radix, all have different edge case
behaviour), and that none of them seem to conform exactly to the
heuristics described in the changelogs, and there seems to be no man
page updates that I can see, it raises some red flags.

So I'm calling for one last opportunity to 1) agree on the desired
behaviour, and 2) ensure implementations are conforming.

quoted

So unless everyone else thinks I'm crazy and disagrees, I'd ask for
a bit more time to make sure we get this interface right. I would
hope for something like prctl PR_SET_MM which can be used to set
our user virtual address bits on a fine grained basis. Maybe a
sysctl, maybe a personality. Something out-of-band. I don't wan to
get too far into that discussion yet. First we need to agree whether
or not the code in the tree today is a problem.  

There is certainly more demand for similar functionality, like creating 
mappings below 2 GB/4 GB/32 GB, and probably other bit patterns. 
Hotspot would use this to place the heap with compressed oops, instead 
of manually hunting for a suitable place for the mapping.  (Essentially, 
32-bit pointers on 64-bit architectures for sufficiently small heap 
sizes.)  It would perhaps be possible to use the hints address as a 
source of the bit count, for full flexibility.  And the mapping should 
be placed into the upper half of the selected window if possible.

MAP_FIXED is near-impossible to use correctly.  I hope you don't expect 
applications to do that.  If you want address-based opt in, it should 
work without MAP_FIXED.  Sure, in obscure cases, applications might 
still see out-of-range addresses, but I expected a full opt-out based on 
RLIMIT_AS would be sufficient for them.

All good points, and no I don't think MAP_FIXED is a substitute.

I don't want to get too far ahead of ourselves, but I don't see why
some new interfaces with reasonably flexible and extensible ways to
specify VA behaviour is a bad idea. We already went through similar
with the ADDR_COMPAT_LAYOUT, and some various address space
randomization options, and now this, and they're all different...
As you noted, for many years now with compressed pointers and data
in pointers, address space bits have *mattered* to applications and
we can't change that.

Surely we're going to have to solve it *properly* sooner or later,
why not now? We could do both existing heuristic *and* make nicer
interfaces, but the >128TB support seems like a good place to do it
because very few apps will have to change code, and we can leave
behaviour exactly unchanged for 99.99999% that will never need so
much memory.

Thanks,
Nick