On Mon, Feb 20, 2017 at 9:21 PM, Michael Pratt <linux@pratt.im> wrote:

On Fri, Feb 17, 2017 at 3:02 PM, Andy Lutomirski <luto@amacapital.net> wrote:
> On Fri, Feb 17, 2017 at 1:01 PM, Linus Torvalds
> <torvalds@linux-foundation.org> wrote:
>> On Fri, Feb 17, 2017 at 12:12 PM, Andy Lutomirski <luto@amacapital.net> wrote:
>>>
>>> At the very least, I'd want to see
>>> MAP_FIXED_BUT_DONT_BLOODY_UNMAP_ANYTHING. I *hate* the current
>>> interface.
>>
>> That's unrelated, but I guess w could add a MAP_NOUNMAP flag, and then
>> you can use MAP_FIXED | MAP_NOUNMAP or something.
>>
>> But that has nothing to do with the 47-vs-56 bit issue.
>>
>>> How about MAP_LIMIT where the address passed in is interpreted as an
>>> upper bound instead of a fixed address?
>>
>> Again, that's a unrelated semantic issue. Right now - if you don't
>> pass in MAP_FIXED at all, the "addr" argument is used as a starting
>> value for deciding where to find an unmapped area. But there is no way
>> to specify the end. That would basically be what the process control
>> thing would be (not per-system-call, but per-thread ).
>>
>
> What I'm trying to say is: if we're going to do the route of 48-bit
> limit unless a specific mmap call requests otherwise, can we at least
> have an interface that doesn't suck?

I've got a set of patches that I've meant to send out as an RFC for a while that tries to address userspace control of address space layout and covers many of these ideas.

There is a new syscall and set of prctls for controlling the "mmap layout" (i.e., get_unmapped_area search range) that look something like this:

struct mmap_layout {

unsigned long start;

unsigned long end;

* These are equivalent to mmap_legacy_base and mmap_base,

* but are not really needed in this proposal.

unsigned long low_base;

unsigned long high_base;

unsigned long flags;

};

/* For flags */

#define MMAP_TOPDOWN 1

struct layout_mmap_args {

unsigned long addr;

unsigned long len;

unsigned long prot;

unsigned long flags;

unsigned long fd;

unsigned long off;

struct mmap_layout layout;

};

void *layout_mmap(struct layout_mmap_args *args);

int prctl(PR_GET_MMAP_LAYOUT, struct mmap_layout *layout);

int prctl(PR_SET_MMAP_LAYOUT, struct mmap_layout *layout);

The prctls control the default range that mmap and friends will allocate. For 56-bit user address space, it could default to [mmap_min_addr, 1<<47), as Linus suggests. Applications that want the full address space can increase it to cover the entire range.

The layout_mmap syscall allows one-off mappings that fall outside the default layout, and nicely solves the "MAP_FIXED but don't unmap anything problem" by passing an explicit range to check without actually setting MAP_FIXED.

This idea is quite similar to the MAX_VADDR + default get_unmapped_area behavior ides, just more generalized to give userspace more control over the ultimate behavior of get_unmapped_area.

PS. Apologies if my email client screwed up this message. I didn't have this thread in my client and have tried to import it from another account.