https://bugs.freedesktop.org/show_bug.cgi?id=60200
Priority: medium Bug ID: 60200 Assignee: dri-devel@lists.freedesktop.org Summary: radeon_bo with virtual address referencing mismatch Severity: normal Classification: Unclassified OS: Linux (All) Reporter: g02maran@gmail.com Hardware: x86-64 (AMD64) Status: NEW Version: git Component: Drivers/Gallium/r600 Product: Mesa
Created attachment 74101 --> https://bugs.freedesktop.org/attachment.cgi?id=74101&action=edit patch
First a disclaimer, this code is unfamiliar to me so there is probably stuff that i have misunderstood.
When i try to start weston it immediatly crashes with the following error: radeon: Failed to allocate a buffer: radeon: size : 1 bytes radeon: alignment : 4096 bytes radeon: domains : 2
and dmesg says: [ 75.092178] radeon 0000:01:00.0: bo ffff8802233f8800 va 0x00000000 conflict with (bo ffff8802219de400 0x0082C000 0x010F6000) [ 75.092195] radeon 0000:01:00.0: bo ffff8802233f8800 don't has a mapping in vm ffff88022275ac00
Everything works fine if i disable the virtual address code, so the issue is in there somewhere.
I tried to analyze it and this is what i came up with:
Consider the following scenario:
radeon_bomgr_create_bo: create gem buffer kernel: create buffer with handle 1 set ref count to 1 radeon_bomgr_create_bo: create radeon_bo with handle 1 set ref count to 1 radeon_bomgr_create_bo: find virtual address for handle 1 radeon_bomgr_find_va: free list empty use offset 2000 radeon_bomgr_create_bo: map virtual address for handle 1 at offset 2000 kernel: map virtual address for handle 1 at offset 2000
somewhere in userspace: create flink for handle 1 kernel: create flink name 1 for handle 1
radeon_winsys_bo_from_handle: open gem buffer with name 1 kernel: create handle 2 increase ref count to 2 radeon_winsys_bo_from_handle: create radeon_bo with handle 2 set ref count to 1 radeon_winsys_bo_from_handle: find virtual address for handle 2 radeon_bomgr_find_va: free list empty use offset 6000 radeon_winsys_bo_from_handle: map virtual address for handle 2 at offset 6000 kernel: virtual address already mapped for handle 2 use offset 2000
somewhere in userspace: destroy handle 2 somewhere in userspace: decrease handle 2 ref count to 0 radeon_bo_destroy: close gem buffer with handle 2 kernel: decrease ref count to 1 radeon_bo_destroy: free virtual address with offset 2000 radeon_bomgr_free_va: add virtual address with offset 2000 to free list
radeon_bomgr_create_bo: create gem buffer kernel: create buffer with handle 3 set ref count to 1 radeon_bomgr_create_bo: create radeon_bo with handle 3 set ref count to 1 radeon_bomgr_create_bo: find virtual address for handle 3 radeon_bomgr_find_va: from free list use offset 2000 radeon_bomgr_create_bo: map virtual address for handle 3 at offset 2000 kernel: virtual address conflict offset 2000 already mapped
This is a simplifed scenario what happens in my case. The issue, at least as i see it, is that the virtual address is freed(added to the free list) even though it is still mapped in the kernel. This is because the userspace referencing code does not track the radeon_bo object created in function radeon_bomgr_create_bo.
The attached patch fixes the issue for me, but is probably not the correct fix.
The issue was found and tested on wayland/weston master, mesa master and kernel 3.8.0-rc6