Hi Christian,
Thanks for exploring this option, it's great to have a better
understanding of the capabilities/limits of mremap.
First, a few practical questions:
- How would Wine detect the updated mremap, I suppose we would want to
create a test persistent mapping on init and see if mremap works on that
page?
- In my experience Mesa will sometimes return malloc'd pointers from
glMapBuffer when using the transfer helper, which I assume has to do
with buffer textures. If wine blindly uses remap for all glMapBuffer
calls, we'll then end up hitting the private anonymous path, where from
what I understand page faults are use to preserve the old page,
potentially slowing things down, thoughts?
Second, in regards to going forward:
I don't have the final call on what path we end up taking 😅. I think it
would be good if others from wine-devel pitched into what they think,
but here's my opinion:
I think the mremap approach should definitely be pursued since it looks
like such a simple Kernel patch, but it may also be good to pursue a
simple Mesa-Integrated path as well, maybe with Mesa calling into a Wine
library to allocate 32-bit pages, as we support non-linux OS's as well.
Thanks,
Derek
Am 10/30/24 um 14:03 schrieb Christian König:
Hi guys,
so I looked a bit deeper into the problem of duplicating graphics
driver mappings with mremap().
This use case of duplicating a mapping into a fixed address is already
supported quite well using mremap(). This is used by a couple of
different emulators to re-create the address space like you would find
it in the specific environment.
The only problem is that this only works for files and shared memory
at the moment. Graphic driver mappings on the other hand have the
VM_DONTEXPAND and VM_PFNMAP flag set because their mappings shouldn't
grow and can also include VRAM.
The attached patch changes this restriction for the mremap() function
and so also allows duplicating the VMAs of graphics drivers into the
lower 32bit address space managed by Wine.
I've tested this with some of AMD's GPU unit tests and it actually
seems to work quite fine.
Derek please let me know if that solution works for you and if you're
interested in using it. If yes I would go ahead and send the patch to
the Linux memory management folks for discussion.
Regards,
Christian.
Am 24.10.24 um 17:06 schrieb Christian König:
Darek we are unfortunately both partially right.
Linux supports cloning VMAs using mremap() from userspace by using a
zero old size, but unfortunately only for SHM areas.
See the code in mm/mremap.c:
       /*
        * We allow a zero old-len as a special case
        * for DOS-emu "duplicate shm area" thing. But
        * a zero new-len is nonsensical.
        */
       if (!new_len)
               return ret;
Going to take a closer look to figure out what would be necessary to
solve that for GPU drivers as well.
Regards,
Christian.
Am 24.10.24 um 14:56 schrieb Christian König:
I haven't tested it but as far as I know that isn't correct.
As far as I know you can map the same VMA at a different location
even without MREMAP_DONTUNMAP. And yes MREMAP_DONTUNMAP only work
with private mappings, but that isn't needed here.
Give me a moment to test this.
Regards,
Christian.
Am 24.10.24 um 10:03 schrieb Derek Lesho:
In my last mail I responded to this approach all the way at the
bottom, so it probably got lost: mremap on Linux as it exists now
won't work as it only supports private anonymous mappings (in
conjunction with MREMAP_DONTUNMAP), which GPU mappings are not.
Am 10/24/24 um 01:06 schrieb James Jones:
That makes sense. Reading the man page myself, it does seem like:
-If the drivers can guarantee they set MAP_SHARED when creating
their initial mapping.
-If WINE is fine rounding down to page boundaries to deal with
mappings of suballocations and either using some lookup structure
to avoid duplicate remappings (probably needed to handle unmap
anyway per below) or just living with the perf cost and address
space overconsumption for duplicate remappings.
-If mremap() preserves the cache attributes of the original mapping.
Then no GL API change would be needed. WINE would just have to do
an if (addrAbove4G) { mremapStuff() } on map and presumably add
some tracking to perform an equivalent munmap() when unmapping. I
assume WINE already has a bunch of vaddr tracking logic in use to
manage the <4G address space as described elsewhere in the thread.
That would be pretty ideal from a driver vendor perspective.
Does that work?
Thanks,
-James
On 10/23/24 06:12, Christian König wrote:
I haven't read through the whole mail thread, but if you manage
the address space using mmap() then you always run into this issue.
If you manage the whole 4GiB address space by Wine then you never
run into this issue. You would just allocate some address range
internally and mremap() into that.
Regards,
Christian.
Am 22.10.24 um 19:32 schrieb James Jones:
This sounds interesting, but does it come with the same "Only
gets 2GB VA" downside Derek pointed out in the thread fork where
he was responding to Michel?
Thanks,
-James
On 10/22/24 07:14, Christian König wrote:
Hi guys,
one theoretical alternative not mentioned in this thread is the
use of mremap().
In other words you reserve some address space below 2G by using
mmap(NULL, length, PROT_NONE, MAP_32BIT | MAP_ANONYMOUS, 0, 0)
and then use mremap(addr64bit, 0, length, MREMAP_FIXED,
reserved_addr).
I haven't tested this but at least in theory it should give you
a duplicate of the 64bit mapping in the lower 2G of the address
space.
Important is that you give 0 as oldsize to mremap() so that the
old mapping isn't unmapped but rather just a new mapping of the
existing VMA created.
Regards,
Christian.
Am 18.10.24 um 23:55 schrieb Derek Lesho:
Hey everyone 👋,
I'm Derek from the Wine project, and wanted to start a
discussion with y'all about potentially extending the Mesa OGL
drivers to help us with a functionality gap we're facing.
Problem Space:
In the last few years Wine's support for running 32-bit
windows apps in a 64-bit host environment (wow64) has almost
reached feature completion, but there remains a pain point
with OpenGL applications: Namely that Wine can't return a
64-bit GL implementation's buffer mappings to a 32 bit
application when the address is outside of the 32-bit range.
Currently, we have a workaround that will copy any changes to
the mapping back to the host upon glBufferUnmap, but this of
course is slow when the implementation directly returns mapped
memory, and doesn't work for GL_PERSISTENT_BIT, where directly
mapped memory is required.
A few years ago we also faced this problem with Vulkan's,
which was solved through the VK_EXT_map_memory_placed
extension Faith drafted, allowing us to use our Wine-internal
allocator to provide the pages the driver maps to. I'm now
wondering if an GL equivalent would also be seen as feasible
amongst the devs here.
Proposed solution:
As the GL backend handles host mapping in its own code, only
giving suballocations from its mappings back to the App, the
problem is a little bit less straight forward in comparison to
our Vulkan solution: If we just allowed the application to set
its own placed mapping when calling glMapBuffer, the driver
might then have to handle moving buffers out of already mapped
ranges, and would lose control over its own memory management
schemes.
Therefore, I propose a GL extension that allows the GL client
to provide a mapping and unmapping callback to the
implementation, to be used whenever the driver needs to
perform such operations. This way the driver remains in full
control of its memory management affairs, and the amount of
work for an implementation as well as potential for bugs is
kept minimal. I've written a draft implementation in Zink
using map_memory_placed [1] and a corresponding Wine MR
utilizing it [2], and would be curious to hear your thoughts.
I don't have experience in the Mesa codebase, so I apologize
if the branch is a tad messy.
In theory, the only requirement from drivers from the
extension would be that glMapBuffer always return a pointer
from within a page allocated through the provided callbacks,
so that it can be guaranteed to be positioned within the
required address space. Wine would then use it's existing
workaround for other types of buffers, but as Mesa seems to
often return directly mapped buffers in other cases as well,
Wine could also avoid the slowdown that comes with copying in
these cases as well.
Why not use Zink?:
There's also a proposal to use a 32-bit PE build of Zink in
Wine bypassing the need for an extension; I brought this to
discussion in this Wine-Devel thread last week [3], which has
some arguments against this approach.
If any of you have thoughts, concerns, or questions about this
potential approach, please let me know, thanks!
1:
https://gitlab.freedesktop.org/Guy1524/mesa/-/commits/placed_allocation
2: https://gitlab.winehq.org/wine/wine/-/merge_requests/6663
3: https://marc.info/?t=172883260300002&r=1&w=2
