On 10/19/24 00:47, Faith Ekstrand wrote:
The timing here isn't great, unfortunately. I'd love to contribute more to the discussion but I'm going on leave starting next week until mid-Febuary so I won't be able to participate much until then. I'll try to leave a few thoughts, though.
Thanks for the comments! I'll actually also be quite busy until then (Uni), but maybe by February they'll be a consensus on the path forward 😀.
On Fri, Oct 18, 2024 at 5:10 PM Derek Lesho dlesho@codeweavers.com wrote:
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.It's an interesting approach, to be sure. I don't mean that as a bad or good thing as I haven't given this enough thought with GL in mind to have a better, more well thought out plan.
The most obvious issue that jumps out to me is that we really want that callback to be set before anyone ever maps a buffer that might possibly get exposed to the client and we want it to never change. If this were Vulkan, we'd have you provide it at vkCreateDevice() time. But this is GL where everybody loves a big mutable state object. If we do go with callbacks (and it's still not 100% clear to me what the right choice is), we'd want them to be somehow set-once and set before any buffers are created. I'm not 100% sure how you'd spec that or how we'd enforce it. There may be some precedent for this somewhere in GL (no_error, maybe?) but I'm not sure.
Right, in the case of Zink I was just lucky it doesn't happen to map anything upon context creation. If I understand what you mean by using a no_error like approach correctly, I think that should definitely work, I think we would then just somewhat-awkwardly want to pass through the callback address as two context attributes, one for the lower and one for the higher part of the address.
However, even if we find a way to make the mapping callback's global on the context level, that then still requires from the driver that they relegate mappings in these contexts to dedicated memory pools. This might be desired, in order to allow other GL clients in the process to continue using the full 64-bit address space, although in practice I don't think Wine uses any libraries that create GL contexts (*other than potentially GStreamer which we are moving away from).
If we just want to keep it simple and workaround the GL context, maybe Wine could just export its allocator in ntdll.so in a way that where Mesa could then directly call it when present. That way we could even avoid the need for a GL extension that will at the end of the day probably only be used by Wine, plus avoid the awkwardness of not wanting to map anything in the driver until we get mapping callback.
The other question that comes to mind is when exactly we'd be expected to use these things. Obviously, we need to do so for any map that may be exposed to the client. However, it's not always easy to do that because you don't know at buffer create time whether or not it will be persistently mapped. A driver is likely to have all sorts of internal mappings for things and, while those can come from one of those ranges, it'll burn more of that precious 32-bit address space than needed. This gets worse when you take sub-allocation into account. If we're okay with all buffer mappings going down the client-request path then it's probably okay. The driver just might need an extra bit in its buffer cache key.
Yeah, it would definitely be a nice bonus to the new path if drivers kept their internal mappings outside of 32-bit address space, but as far as I can see I don't think this affects our choice of interface, as mesa should be able to handle client visible buffers differently from driver-internal ones without the help of the client.
I'm also sitting here trying to come up with some plan that would let us do this more like Vulkan and I'm having trouble coming up with one that works. GL has no concept of "create time". We could theoretically do something where we flush everything, copy the data to a new placed-mappable buffer and then continue on but that's gonna suck.
Yeah, and then you also still have to define how you create this new placed-mappable buffer, and there isn't really a small-neat solution for this. You could define a GL_OUT_OF_PLACED_MEMORY error in glMapBuffer, and then have another entry point by which the app feeds the driver pages, but then you run into the problem of unmapping, which as far as I can see is often performed asynchronously.
I think that's all that comes to mind immediately. As I said at the top, I'm happy to talk more in a few months. Best of luck until then!
~Faith
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.For cases where Zink is being used on the host (this is the current plan for Nouveau going forward), doing Zink in Windows may not be a bad idea. However, I agree that it may not be the best idea to rely on that plan.
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 <https://marc.info/?t=172883260300002&r=1&w=2>
-
Derek Lesho