From d584198a6fe4c51f4aa88ad72f258f8961a0f11c Mon Sep 17 00:00:00 2001 From: Lancelot SIX Date: Tue, 28 Jan 2025 19:16:49 +0000 Subject: [PATCH] drm/amdkfd: Ensure consistent barrier state saved in gfx12 trap handler It is possible for some waves in a workgroup to finish their save sequence before the group leader has had time to capture the workgroup barrier state. When this happens, having those waves exit do impact the barrier state. As a consequence, the state captured by the group leader is invalid, and is eventually incorrectly restored. This patch proposes to have all waves in a workgroup wait for each other at the end of their save sequence (just before calling s_endpgm_saved). Signed-off-by: Lancelot SIX Reviewed-by: Jay Cornwall Signed-off-by: Alex Deucher Cc: stable@vger.kernel.org # 6.12.x --- drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler.h | 3 ++- drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx12.asm | 4 ++++ 2 files changed, 6 insertions(+), 1 deletion(-) diff --git a/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler.h b/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler.h index 984f0e705078..651660958e5b 100644 --- a/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler.h +++ b/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler.h @@ -4121,7 +4121,8 @@ static const uint32_t cwsr_trap_gfx12_hex[] = { 0x0000ffff, 0x8bfe7e7e, 0x8bea6a6a, 0xb97af804, 0xbe804ec2, 0xbf94fffe, - 0xbe804a6c, 0xbfb10000, + 0xbe804a6c, 0xbe804ec2, + 0xbf94fffe, 0xbfb10000, 0xbf9f0000, 0xbf9f0000, 0xbf9f0000, 0xbf9f0000, 0xbf9f0000, 0x00000000, diff --git a/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx12.asm b/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx12.asm index 1740e98c6719..7b9d36e5fa43 100644 --- a/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx12.asm +++ b/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx12.asm @@ -1049,6 +1049,10 @@ L_SKIP_BARRIER_RESTORE: s_rfe_b64 s_restore_pc_lo //Return to the main shader program and resume execution L_END_PGM: + // Make sure that no wave of the workgroup can exit the trap handler + // before the workgroup barrier state is saved. + s_barrier_signal -2 + s_barrier_wait -2 s_endpgm_saved end