Re: When should ralloc.c be used?

[Daniel Colascione]

On 10/28/2016 12:46 AM, Eli Zaretskii wrote:
> > From: Daniel Colascione <address@hidden>
> > Cc: Stefan Monnier <address@hidden>,  address@hidden,  address@hidden
> > Date: Thu, 27 Oct 2016 23:23:05 -0700
> >
> > We're talking about setting aside address space only, not asking
> > the OS to make guarantees about future memory availability.  All major
> > operating systems, even ones like Windows that don't do overcommit,
> > provide ways to reserve address space without asking the OS to guarantee
> > availability of memory.
>
> Not sure I understand you: if a portion of the address space has been
> reserved, how come these addresses won't be available when we try to
> commit them later?  There might not be physical pages available for
> that, but virtual memory for those addresses must be available, no?

I'm not sure I understand what you're confused about, so I'll try a 
broader explanation.

Say I mmap (anonymously, for simplicity) a page PROT_NONE. After the 
initial mapping, that address space is unavailable for other uses. But 
because the page protections are PROT_NONE, my program has no legal 
right to access that page, so the OS doesn't have to guarantee that it 
can find a physical page to back that page I've mmaped. In this state, 
the memory is reserved.

The 20GB PROT_NONE address space reservation itself requires very little 
memory. It's just a note in the kernel's VM interval tree that says "the 
addresses in range [0x20000, 0x500020000) are reserved". Virtual memory is

Now imagine I change the protections to PROT_READ|PROT_WRITE --- once 
the PROT_READ|PROT_WRITE mprotect succeeds, my program has every right 
to access that page; under a strict accounting scheme (that is, without 
overcommit), the OS has to guarantee that it'll be able to go find a 
physical page to back that virtual page. In this state, the memory is 
committed -- the kernel has committed to finding backing storage for 
that page at some point when the current process tries to access it.

Say you have a strict-accounting system with 1GB of RAM and 1GB of swap. 
I can write a program that reserves 20GB of address space. That's fine. 
The kernel isn't promising to give you 20GB of memory: it's setting 
address space. Now if I attempt to map 20GB PROT_READ|PROT_WRITE, on any 
reasonable (i..e, not overcommit) system, mmap should fail, since 
there's no way a system with 1GB of RAM and 1GB of swap can promise to 
provide 20GB of private memory.

Overcommit confuses the issue: the kernel will _commit_ to as much 
memory as you ask it for and then renege on that commitment when it 
finds it convenient. An overcommit system with 1GB of RAM and 1GB of 
swap will happily let you make that 20GB PROT_READ|PROT_WRITE mapping. 
It'll just kill you after you use more than 2GB of that mapping. A 
non-overcommit system understands how to say "sorry, I can't let you do 
that" up front. On a non-overcommit system, your process will never be 
killed for accessing memory that the kernel told the process in advance 
that it could use.

I think Jérémie is working with a mental model where every memory 
mapping is commit, and only an overcommit system allows you to commit 
more memory than the system actually has. Any system will let you 
reserve more address space than you have available commit: reservations 
are cheap.

(In Windows, the corresponding concepts MEM_RESERVE and MEM_COMMIT. 
Windows is much more explicit about the difference between memory 
reservations and memory commitments than Linux is, because most of the 
time, Linux users don't care about the difference.)