Creating swapchain gives SIGSEGV / segfault

[Clyybber]

EDIT: There never was an issue, I was dumb enough to pass a physical device to vkCreateSwapchainKHR

If trying to create a swapchain with vkCreateSwapchainKHR(device, addr createInfo, nil, addr swapchain) where swapchain is of type VkSwapchainKHR I get SIGSEGV: Illegal Storage access. I suppose this is just me being dumb but I really can't figure out what I am doing wrong here.

[Clyybber]

I tried to allocate my swapchain variable manually as described here: https://nim-lang.org/docs/manual.html#types-reference-and-pointer-types but that didn't prevent the SIGSEGV either. For windows devs: SIGSEGV is a segfault.

[Clyybber]

Vulkan wrappers for other languages also implement it differently, for example this haskell binding: http://hackage.haskell.org/package/vulkan-2.1.0.0/docs/src/Graphics.Vulkan.Extensions.VK_KHR_swapchain.html#VkSwapchainKHR_T implements it as a pointer to VkSwapchainKHR_T, while this rust binding: https://docs.rs/vks/0.20.0/src/vks/khr_swapchain.rs.html#26 implements it as a struct to be defined by VK_DEFINE_NON_DISPATCHABLE_HANDLE.

[Clyybber]

So according to https://stackoverflow.com/questions/43741998/what-does-vk-define-non-dispatchable-handle-mean and https://github.com/KhronosGroup/Vulkan-Docs/blob/1.0/src/vulkan/vulkan.h#L56 , VkSwapchainKHR is a VkNonDispatchableHandle or something that is 64 bits in size. So type VkSwapchainKHR* = pointer and type VkSwapchainKHR* = VkNonDispatchableHandle should work. But they don't. I don't think its my drivers, since all the vulkan demos I've tested work. Maybe the bug is in the compiler?

[Clyybber]

If it helps here is the relevant code: https://gitlab.com/n0tknot/wyven/blob/master/src/engine/window.nim#L221 The last 3 lines of that function are where the relevant things happen.

[gmpreussner]

Hey Clyybber, thanks for looking into this - I didn't have time yet to take a look myself, but will try to do so soon!

[Clyybber]

In the generated C code the type VkSwapchainKHR gets translated to void*. Then we assign the swapchain variable to alloc0(8), that should be 64 bits in size. The address of that void pointer is then passed to vkCreateSwapchainKHR(). I really can't figure this out. Maybe alloc0(), alloc() or my driver is bugged. EDIT: For reference here is the generated C code for the relevant proc: https://gist.github.com/Clyybber/4aa9835769496cebc9ac5ba13568a079

[gmpreussner]

Hey Clybber, sorry, I still haven't had any time to look into this. Too much going on right now :/

[Clyybber]

No worries, thank you anyways. I wish I could fix it myself, but I can't figure it out.

[gmpreussner]

Had a quick look.

First, my definition of VkSwapchainKHR is incorrect. What the Vulkan API specifies is that this type must be 64-bit on all platforms and unique if possible, so on 64-bit platforms they declare it as a pointer, and on all other platforms as a 64-bit unsigned integer.

Now, the swap chain values that the KHR functions take and return are opaque handles, meaning that we don't actually have to know what they mean under the hood.... could be pointers to structs, could be index numbers into a container, could be a hash value, could be a key into a map, etc.

I'm thinking that we should get rid of the pointer representation altogether and just declare the handle as an int64. That way you are only passing around integers and don't have to worry about (platform-specific) casting and other things.

Can you give that a try?

[Clyybber]

Just tried that, still gives a segfault. Maybe this is an upstream bug, but I can't confirm it since I can't rule out buggy drivers(the examples and demos work though).

[Clyybber]

I'm dumb as fuck :D

Because I used my own function wrapping the creation of a swapchain wrongly, a physical device was passed to vkCreateSwapchainKHR instead of a logical one. The verification layers (and my brain) didnt catch that.

Sry that I have bothered you with this so long

[Clyybber]

Maybe we should make some of those alias types for VkHandle distinct so that such dumbness can be prevented, though on second thought prevention of that might also be the responsibility of the validation layers.

[awr1]

@Clyybber make an issue at https://github.com/KhronosGroup/Vulkan-ValidationLayers ?

[Clyybber]

Done: https://github.com/KhronosGroup/Vulkan-ValidationLayers/issues/288