add addr(*mut u8) to set the desired target address

[ghost]

No description provided.

[dtolnay]

Could you look into the CI failure on Windows?

error[E0308]: mismatched types
  --> src\windows.rs:66:47
   |
66 |                                               addr as winapi::DWORD);
   |                                               ^^^^^^^^^^^^^^^^^^^^^ expected *-ptr, found u32
   |
   = note: expected type `*mut std::os::raw::c_void`
              found type `u32`

[danburkert]

I'm not sure that exposing MAP_FIXED is a good idea. My concern is that misuse of the API (which is easy to do) can cause existing mappings to be overwritten, including mappings created by the heap allocator.

It does seem like there is a path towards a less dangerous and cross-platform consistent API, though: we could specify the address without setting MAP_FIXED, and return an error if the address returned by mmap isn't the one requested. I've never had a usecase for MAP_FIXED, though, so I can't say if that will suffice for whatever case you have in mind.

[danburkert]

Mostly style nits, but the MAP_FIXED issue is the bigger issue.

[danburkert]

typo: 'correctly'

[danburkert]

s/needs to be/must be

[danburkert]

uppercase 'Unix' and 'Windows' here and elsewhere

[danburkert]

typo: 'replace an existing mapping'

[danburkert]

reindent

[danburkert]

reindent

[danburkert]

I think this would be more clear as 'The address must be aligned to the system pagesize.'

[ghost]

@danburkert My usecase does require MAP_FIXED. (I need a bigger alignment that the page-size.)
Without MAP_FIXED the returned address is a bit like gambling on Linux. sometimes you get what you requested in the first attempt, sometimes you don't even get it after 10 attempts.

[danburkert]

@s3bk not sure I follow. If you need a bigger alignment than page size, you should over-allocate and then start using the map from the closest offset matching your desired alignment.

I've never used MAP_FIXED, but my interpretation from the docs is that if you used MAP_FIXED to work around the memory address you are requesting already being mapped, there's no way to tell the previous user of that memory space that it's been preempted. Needless to say, this would be very unsafe.

[danburkert]

Note about 1.0: regardless of how we land on whether MAP_FIXED should be exposed, I do think we should allow setting the address, however I'm confident this can be done backwards compatibly with the new builder pattern post 1.0, so I don't consider this feature a blocker.

[ghost]

@danburkert

If you need a bigger alignment than page size, you should over-allocate and then start using the map from the closest offset matching your desired alignment.

That would work, if it wasn't using a file (I don't think negative offsets are allowed)

Creating a dummy mmap to pad the area up to the desired alignment would probably work.