[PATCH v2] drm/amd: Add a workaround for GFX11 systems that fail to flush TLB

[Huang, Tim]

[Public]

Hi Mario,


-----Original Message-----
From: Limonciello, Mario <Mario.Limonciello at amd.com>
Sent: Thursday, December 14, 2023 4:31 AM
To: amd-gfx at lists.freedesktop.org
Cc: Limonciello, Mario <Mario.Limonciello at amd.com>; stable at vger.kernel.org; Huang, Tim <Tim.Huang at amd.com>
Subject: [PATCH v2] drm/amd: Add a workaround for GFX11 systems that fail to flush TLB

Some systems with MP1 13.0.4 or 13.0.11 have a firmware bug that causes the first MES packet after resume to fail. Typically this packet is used to flush the TLB when GART is enabled.

This issue is fixed in newer firmware, but as OEMs may not roll this out to the field, introduce a workaround that will add an extra dummy read on resume that the result is discarded.

Cc: stable at vger.kernel.org # 6.1+
Cc: Tim Huang <Tim.Huang at amd.com>
Closes: https://gitlab.freedesktop.org/drm/amd/-/issues/3045
Signed-off-by: Mario Limonciello <mario.limonciello at amd.com>
---
v1->v2:
 * Add a dummy read callback instead and use that.
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c | 19 +++++++++++++++++++  drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h |  3 +++  drivers/gpu/drm/amd/amdgpu/gmc_v11_0.c  | 11 +++++++++++  drivers/gpu/drm/amd/amdgpu/mes_v11_0.c  |  8 ++++++--
 4 files changed, 39 insertions(+), 2 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c
index 9ddbf1494326..cd5e1a027bdf 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c
@@ -868,6 +868,25 @@ int amdgpu_mes_reg_wait(struct amdgpu_device *adev, uint32_t reg,
        return r;
 }

+void amdgpu_mes_reg_dummy_read(struct amdgpu_device *adev) {
+       struct mes_misc_op_input op_input = {
+               .op = MES_MISC_OP_READ_REG,
+               .read_reg.reg_offset = 0,
+               .read_reg.buffer_addr = adev->mes.read_val_gpu_addr,
+       };
+
+       if (!adev->mes.funcs->misc_op) {
+               DRM_ERROR("mes misc op is not supported!\n");
+               return;
+       }
+
+       adev->mes.silent_errors = true;
+       if (adev->mes.funcs->misc_op(&adev->mes, &op_input))
+               DRM_DEBUG("failed to amdgpu_mes_reg_dummy_read\n");
+       adev->mes.silent_errors = false;
+}
+
 int amdgpu_mes_set_shader_debugger(struct amdgpu_device *adev,
                                uint64_t process_context_addr,
                                uint32_t spi_gdbg_per_vmid_cntl,
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h
index a27b424ffe00..d208e60c1d99 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h
@@ -135,6 +135,8 @@ struct amdgpu_mes {

        /* ip specific functions */
        const struct amdgpu_mes_funcs   *funcs;
+
+       bool                            silent_errors;
 };

 struct amdgpu_mes_process {
@@ -356,6 +358,7 @@ int amdgpu_mes_unmap_legacy_queue(struct amdgpu_device *adev,
                                  u64 gpu_addr, u64 seq);

 uint32_t amdgpu_mes_rreg(struct amdgpu_device *adev, uint32_t reg);
+void amdgpu_mes_reg_dummy_read(struct amdgpu_device *adev);
 int amdgpu_mes_wreg(struct amdgpu_device *adev,
                    uint32_t reg, uint32_t val);
 int amdgpu_mes_reg_wait(struct amdgpu_device *adev, uint32_t reg, diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v11_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v11_0.c
index 23d7b548d13f..a2ba45f859ea 100644
--- a/drivers/gpu/drm/amd/amdgpu/gmc_v11_0.c
+++ b/drivers/gpu/drm/amd/amdgpu/gmc_v11_0.c
@@ -960,6 +960,17 @@ static int gmc_v11_0_resume(void *handle)
        int r;
        struct amdgpu_device *adev = (struct amdgpu_device *)handle;

+       switch (amdgpu_ip_version(adev, MP1_HWIP, 0)) {
+       case IP_VERSION(13, 0, 4):
+       case IP_VERSION(13, 0, 11):
+               /* avoid a lost packet @ first GFXOFF exit after resume */
+               if ((adev->pm.fw_version & 0x00FFFFFF) < 0x004c4900 && adev->in_s0ix)
+                       amdgpu_mes_reg_dummy_read(adev);
+               break;
+       default:
+               break;
+       }
+

I tried this patch on my device, but it not working. The situation is this dummy reading not hit the MES timeout error but after that still hit the same error in the amdgpu_virt_kiq_reg_write_reg_wait. Maybe the failed case is not just the first GFXOFF exit.

        r = gmc_v11_0_hw_init(adev);
        if (r)
                return r;
diff --git a/drivers/gpu/drm/amd/amdgpu/mes_v11_0.c b/drivers/gpu/drm/amd/amdgpu/mes_v11_0.c
index 4dfec56e1b7f..71df5cb65485 100644
--- a/drivers/gpu/drm/amd/amdgpu/mes_v11_0.c
+++ b/drivers/gpu/drm/amd/amdgpu/mes_v11_0.c
@@ -137,8 +137,12 @@ static int mes_v11_0_submit_pkt_and_poll_completion(struct amdgpu_mes *mes,
        r = amdgpu_fence_wait_polling(ring, ring->fence_drv.sync_seq,
                      timeout);
        if (r < 1) {
-               DRM_ERROR("MES failed to response msg=%d\n",
-                         x_pkt->header.opcode);
+               if (mes->silent_errors)
+                       DRM_DEBUG("MES failed to response msg=%d\n",
+                                 x_pkt->header.opcode);
+               else
+                       DRM_ERROR("MES failed to response msg=%d\n",
+                                 x_pkt->header.opcode);

                while (halt_if_hws_hang)
                        schedule();
--
2.34.1