Message ID | 1480361317-9937-13-git-send-email-jcrouse@codeaurora.org (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
On Mon 28 Nov 11:28 PST 2016, Jordan Crouse wrote: > The A5XX GPU powers on in "secure" mode. In secure mode the GPU can > only render to buffers that are marked as secure and inaccessible > to the kernel and user through a series of hardware protections. In > practice secure mode is used to draw things like a UI on a secure > video frame. > > In order to switch out of secure mode the GPU executes a special > shader that clears out the GMEM and other sensitve registers and > then writes a register. Because the kernel can't be trusted the > shader binary is signed and verified and programmed by the > secure world. To do this we need to read the MDT header and the > segments from the firmware location and put them in memory and > present them for approval. > > For targets without secure support there is an out: if the > secure world doesn't support secure then there are no hardware > protections and we can freely write the SECVID_TRUST register from > the CPU. We don't have 100% confidence that we can query the > secure capabilities at run time but we have enough calls that > need to go right to give us some confidence that we're at least doing > something useful. > > Of course if we guess wrong you trigger a permissions violation > which usually ends up in a system crash but thats a problem > that shows up immediately. > > Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org> > --- > drivers/gpu/drm/msm/adreno/a5xx_gpu.c | 72 ++++++++++++++++++++++++++++++++++- > 1 file changed, 70 insertions(+), 2 deletions(-) > > diff --git a/drivers/gpu/drm/msm/adreno/a5xx_gpu.c b/drivers/gpu/drm/msm/adreno/a5xx_gpu.c > index eefe197..a7a58ec 100644 > --- a/drivers/gpu/drm/msm/adreno/a5xx_gpu.c > +++ b/drivers/gpu/drm/msm/adreno/a5xx_gpu.c > @@ -469,6 +469,55 @@ static int a5xx_ucode_init(struct msm_gpu *gpu) > return 0; > } > > +static int a5xx_zap_shader_resume(struct msm_gpu *gpu) > +{ > + int ret; > + > + ret = qcom_scm_gpu_zap_resume(); > + if (ret) > + DRM_ERROR("%s: zap-shader resume failed: %d\n", > + gpu->name, ret); > + > + return ret; > +} > + > +static int a5xx_zap_shader_init(struct msm_gpu *gpu) > +{ > + static bool loaded; > + struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu); > + struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu); > + struct platform_device *pdev = a5xx_gpu->pdev; > + struct device_node *node; > + int ret; > + > + /* > + * If the zap shader is already loaded into memory we just need to kick > + * the remote processor to reinitialize it > + */ > + if (loaded) Why is this handling needed? Why can init be called multiple times? > + return a5xx_zap_shader_resume(gpu); > + > + /* Populate the sub-nodes if they haven't already been done */ > + of_platform_populate(pdev->dev.of_node, NULL, NULL, &pdev->dev); I haven't been able to find the qcom,zap-shader platform driver, but I presume you have something like: adreno { qcom,zap-shader { compatible = "qcom,zap-shader"; firmware = "zapfw"; memory-region = <&zap_region>; }; }; I presume this is done to not "taint" the adreno device's with the zap memory region, but I don't think you should (ab)use a platform driver for this. You should rather add a struct device zap_dev to your adreno context, do minimal initialization (name and a parent I think is enough), call device_register(&zap_dev);, of_reserved_mem_device_init() and then use that for your dma allocation. This saves you from creating a platform_driver, instantiating a platform_device and the worry of the race between the creation of that device and the of_find_device_by_node() below. > + > + /* Find the sub-node for the zap shader */ > + node = of_find_node_by_name(pdev->dev.of_node, "qcom,zap-shader"); If you're looking for immediate children use of_get_child_by_name() And no "qcom," in node names please. > + if (!node) { > + DRM_ERROR("%s: qcom,zap-shader not found in device tree\n", > + gpu->name); > + return -ENODEV; > + } > + > + ret = _pil_tz_load_image(of_find_device_by_node(node)); > + if (ret) > + DRM_ERROR("%s: Unable to load the zap shader\n", > + gpu->name); > + > + loaded = !ret; > + > + return ret; > +} Regards, Bjorn
On Mon 05 Dec 11:57 PST 2016, Bjorn Andersson wrote: [..] > You should rather add a struct device zap_dev to your adreno context, do > minimal initialization (name and a parent I think is enough), call > device_register(&zap_dev);, of_reserved_mem_device_init() and then use > that for your dma allocation. You would of course need to set the of_node on the zap_dev for of_reserved_mem_device_init() to work. Sorry for missing that. Regards, Bjorn
On Mon, Dec 05, 2016 at 11:57:12AM -0800, Bjorn Andersson wrote: > On Mon 28 Nov 11:28 PST 2016, Jordan Crouse wrote: > > > The A5XX GPU powers on in "secure" mode. In secure mode the GPU can > > only render to buffers that are marked as secure and inaccessible > > to the kernel and user through a series of hardware protections. In > > practice secure mode is used to draw things like a UI on a secure > > video frame. > > > > In order to switch out of secure mode the GPU executes a special > > shader that clears out the GMEM and other sensitve registers and > > then writes a register. Because the kernel can't be trusted the > > shader binary is signed and verified and programmed by the > > secure world. To do this we need to read the MDT header and the > > segments from the firmware location and put them in memory and > > present them for approval. > > > > For targets without secure support there is an out: if the > > secure world doesn't support secure then there are no hardware > > protections and we can freely write the SECVID_TRUST register from > > the CPU. We don't have 100% confidence that we can query the > > secure capabilities at run time but we have enough calls that > > need to go right to give us some confidence that we're at least doing > > something useful. > > > > Of course if we guess wrong you trigger a permissions violation > > which usually ends up in a system crash but thats a problem > > that shows up immediately. > > > > Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org> > > --- > > drivers/gpu/drm/msm/adreno/a5xx_gpu.c | 72 ++++++++++++++++++++++++++++++++++- > > 1 file changed, 70 insertions(+), 2 deletions(-) > > > > diff --git a/drivers/gpu/drm/msm/adreno/a5xx_gpu.c b/drivers/gpu/drm/msm/adreno/a5xx_gpu.c > > index eefe197..a7a58ec 100644 > > --- a/drivers/gpu/drm/msm/adreno/a5xx_gpu.c > > +++ b/drivers/gpu/drm/msm/adreno/a5xx_gpu.c > > @@ -469,6 +469,55 @@ static int a5xx_ucode_init(struct msm_gpu *gpu) > > return 0; > > } > > > > +static int a5xx_zap_shader_resume(struct msm_gpu *gpu) > > +{ > > + int ret; > > + > > + ret = qcom_scm_gpu_zap_resume(); > > + if (ret) > > + DRM_ERROR("%s: zap-shader resume failed: %d\n", > > + gpu->name, ret); > > + > > + return ret; > > +} > > + > > +static int a5xx_zap_shader_init(struct msm_gpu *gpu) > > +{ > > + static bool loaded; > > + struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu); > > + struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu); > > + struct platform_device *pdev = a5xx_gpu->pdev; > > + struct device_node *node; > > + int ret; > > + > > + /* > > + * If the zap shader is already loaded into memory we just need to kick > > + * the remote processor to reinitialize it > > + */ > > + if (loaded) > > Why is this handling needed? Why can init be called multiple times? This is for resume - if we suspend and resume the device without losing state the secure zone we can't load it again, so we have to call a different operation to "resume" it. This will be much more heavily used when we have more aggressive power management. > > + return a5xx_zap_shader_resume(gpu); > > + > > + /* Populate the sub-nodes if they haven't already been done */ > > + of_platform_populate(pdev->dev.of_node, NULL, NULL, &pdev->dev); > > I haven't been able to find the qcom,zap-shader platform driver, but I > presume you have something like: > > adreno { > qcom,zap-shader { > compatible = "qcom,zap-shader"; > > firmware = "zapfw"; > memory-region = <&zap_region>; > }; > }; > > I presume this is done to not "taint" the adreno device's with the zap > memory region, but I don't think you should (ab)use a platform driver > for this. > > You should rather add a struct device zap_dev to your adreno context, do > minimal initialization (name and a parent I think is enough), call > device_register(&zap_dev);, of_reserved_mem_device_init() and then use > that for your dma allocation. > > This saves you from creating a platform_driver, instantiating a > platform_device and the worry of the race between the creation of that > device and the of_find_device_by_node() below. As far as I know, of_platform_populate() just fleshed out the platform devices for the sub-nodes. We are not creating a new platform driver or doing any sort of probe code, we're just setting up the useful memory to attach the memory-region to. As far as I can tell, using of_platform_populate() + of_find_device_by_node() does a lot of the heavy lifting of what you describe. > > + > > + /* Find the sub-node for the zap shader */ > > + node = of_find_node_by_name(pdev->dev.of_node, "qcom,zap-shader"); > > If you're looking for immediate children use of_get_child_by_name() > > And no "qcom," in node names please. Okay, can do. > > + if (!node) { > > + DRM_ERROR("%s: qcom,zap-shader not found in device tree\n", > > + gpu->name); > > + return -ENODEV; > > + } > > + > > + ret = _pil_tz_load_image(of_find_device_by_node(node)); > > + if (ret) > > + DRM_ERROR("%s: Unable to load the zap shader\n", > > + gpu->name); > > + > > + loaded = !ret; > > + > > + return ret; > > +} > > Regards, > Bjorn Jordan
On Tue, Dec 6, 2016 at 10:35 AM, Jordan Crouse <jcrouse@codeaurora.org> wrote: > On Mon, Dec 05, 2016 at 11:57:12AM -0800, Bjorn Andersson wrote: >> On Mon 28 Nov 11:28 PST 2016, Jordan Crouse wrote: >> >> > The A5XX GPU powers on in "secure" mode. In secure mode the GPU can >> > only render to buffers that are marked as secure and inaccessible >> > to the kernel and user through a series of hardware protections. In >> > practice secure mode is used to draw things like a UI on a secure >> > video frame. >> > >> > In order to switch out of secure mode the GPU executes a special >> > shader that clears out the GMEM and other sensitve registers and >> > then writes a register. Because the kernel can't be trusted the >> > shader binary is signed and verified and programmed by the >> > secure world. To do this we need to read the MDT header and the >> > segments from the firmware location and put them in memory and >> > present them for approval. >> > >> > For targets without secure support there is an out: if the >> > secure world doesn't support secure then there are no hardware >> > protections and we can freely write the SECVID_TRUST register from >> > the CPU. We don't have 100% confidence that we can query the >> > secure capabilities at run time but we have enough calls that >> > need to go right to give us some confidence that we're at least doing >> > something useful. >> > >> > Of course if we guess wrong you trigger a permissions violation >> > which usually ends up in a system crash but thats a problem >> > that shows up immediately. >> > >> > Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org> >> > --- >> > drivers/gpu/drm/msm/adreno/a5xx_gpu.c | 72 ++++++++++++++++++++++++++++++++++- >> > 1 file changed, 70 insertions(+), 2 deletions(-) >> > >> > diff --git a/drivers/gpu/drm/msm/adreno/a5xx_gpu.c b/drivers/gpu/drm/msm/adreno/a5xx_gpu.c >> > index eefe197..a7a58ec 100644 >> > --- a/drivers/gpu/drm/msm/adreno/a5xx_gpu.c >> > +++ b/drivers/gpu/drm/msm/adreno/a5xx_gpu.c >> > @@ -469,6 +469,55 @@ static int a5xx_ucode_init(struct msm_gpu *gpu) >> > return 0; >> > } >> > >> > +static int a5xx_zap_shader_resume(struct msm_gpu *gpu) >> > +{ >> > + int ret; >> > + >> > + ret = qcom_scm_gpu_zap_resume(); >> > + if (ret) >> > + DRM_ERROR("%s: zap-shader resume failed: %d\n", >> > + gpu->name, ret); >> > + >> > + return ret; >> > +} >> > + >> > +static int a5xx_zap_shader_init(struct msm_gpu *gpu) >> > +{ >> > + static bool loaded; >> > + struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu); >> > + struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu); >> > + struct platform_device *pdev = a5xx_gpu->pdev; >> > + struct device_node *node; >> > + int ret; >> > + >> > + /* >> > + * If the zap shader is already loaded into memory we just need to kick >> > + * the remote processor to reinitialize it >> > + */ >> > + if (loaded) >> >> Why is this handling needed? Why can init be called multiple times? > > This is for resume - if we suspend and resume the device without losing state > the secure zone we can't load it again, so we have to call a different > operation to "resume" it. This will be much more heavily used when we have more > aggressive power management. also for recovery when I crash the gpu ;-) BR, -R
On Tue 06 Dec 07:35 PST 2016, Jordan Crouse wrote: > On Mon, Dec 05, 2016 at 11:57:12AM -0800, Bjorn Andersson wrote: > > On Mon 28 Nov 11:28 PST 2016, Jordan Crouse wrote: > > > > > The A5XX GPU powers on in "secure" mode. In secure mode the GPU can > > > only render to buffers that are marked as secure and inaccessible > > > to the kernel and user through a series of hardware protections. In > > > practice secure mode is used to draw things like a UI on a secure > > > video frame. > > > > > > In order to switch out of secure mode the GPU executes a special > > > shader that clears out the GMEM and other sensitve registers and > > > then writes a register. Because the kernel can't be trusted the > > > shader binary is signed and verified and programmed by the > > > secure world. To do this we need to read the MDT header and the > > > segments from the firmware location and put them in memory and > > > present them for approval. > > > > > > For targets without secure support there is an out: if the > > > secure world doesn't support secure then there are no hardware > > > protections and we can freely write the SECVID_TRUST register from > > > the CPU. We don't have 100% confidence that we can query the > > > secure capabilities at run time but we have enough calls that > > > need to go right to give us some confidence that we're at least doing > > > something useful. > > > > > > Of course if we guess wrong you trigger a permissions violation > > > which usually ends up in a system crash but thats a problem > > > that shows up immediately. > > > > > > Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org> > > > --- > > > drivers/gpu/drm/msm/adreno/a5xx_gpu.c | 72 ++++++++++++++++++++++++++++++++++- > > > 1 file changed, 70 insertions(+), 2 deletions(-) > > > > > > diff --git a/drivers/gpu/drm/msm/adreno/a5xx_gpu.c b/drivers/gpu/drm/msm/adreno/a5xx_gpu.c > > > index eefe197..a7a58ec 100644 > > > --- a/drivers/gpu/drm/msm/adreno/a5xx_gpu.c > > > +++ b/drivers/gpu/drm/msm/adreno/a5xx_gpu.c > > > @@ -469,6 +469,55 @@ static int a5xx_ucode_init(struct msm_gpu *gpu) > > > return 0; > > > } > > > > > > +static int a5xx_zap_shader_resume(struct msm_gpu *gpu) > > > +{ > > > + int ret; > > > + > > > + ret = qcom_scm_gpu_zap_resume(); > > > + if (ret) > > > + DRM_ERROR("%s: zap-shader resume failed: %d\n", > > > + gpu->name, ret); > > > + > > > + return ret; > > > +} > > > + > > > +static int a5xx_zap_shader_init(struct msm_gpu *gpu) > > > +{ > > > + static bool loaded; > > > + struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu); > > > + struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu); > > > + struct platform_device *pdev = a5xx_gpu->pdev; > > > + struct device_node *node; > > > + int ret; > > > + > > > + /* > > > + * If the zap shader is already loaded into memory we just need to kick > > > + * the remote processor to reinitialize it > > > + */ > > > + if (loaded) > > > > Why is this handling needed? Why can init be called multiple times? > > This is for resume - if we suspend and resume the device without > losing state the secure zone we can't load it again, so we have to > call a different operation to "resume" it. This will be much more > heavily used when we have more aggressive power management. > I presume then that the rest of the resume path is the same as the initialization. > > > + return a5xx_zap_shader_resume(gpu); > > > + > > > + /* Populate the sub-nodes if they haven't already been done */ > > > + of_platform_populate(pdev->dev.of_node, NULL, NULL, &pdev->dev); > > > > I haven't been able to find the qcom,zap-shader platform driver, but I > > presume you have something like: > > > > adreno { > > qcom,zap-shader { > > compatible = "qcom,zap-shader"; > > > > firmware = "zapfw"; > > memory-region = <&zap_region>; > > }; > > }; > > > > I presume this is done to not "taint" the adreno device's with the zap > > memory region, but I don't think you should (ab)use a platform driver > > for this. > > > > You should rather add a struct device zap_dev to your adreno context, do > > minimal initialization (name and a parent I think is enough), call > > device_register(&zap_dev);, of_reserved_mem_device_init() and then use > > that for your dma allocation. > > > > This saves you from creating a platform_driver, instantiating a > > platform_device and the worry of the race between the creation of that > > device and the of_find_device_by_node() below. > > As far as I know, of_platform_populate() just fleshed out the platform > devices for the sub-nodes. Correct, it will scan all subnodes and try to match their compatible with platform_drivers. For each match it finds it probes that driver. > We are not creating a new platform driver > or doing any sort of probe code, we're just setting up the useful > memory to attach the memory-region to. As far as I can tell, using > of_platform_populate() + of_find_device_by_node() does a lot of the > heavy lifting of what you describe. But as far as I know the compatible of node that you name "qcom,zap-shader" must actually match a platform_driver for there to be a device that of_find_device_by_node() will be able to match against. The heavy lifting I'm talking about is (untested): struct device zap_dev; zap_dev.of_node = of_get_child_by_name("zap-shader"); zap_dev.parent = &pdev->dev; dev_set_name(&zap_dev, "zap-shader"); ret = device_register(&zap_dev); if (ret < 0) { dev_err(); return ret; } of_reserved_mem_device_init(&zap_dev); And then you need a device_put(&zap_dev) when you're done with the device and if you set zap_dev.release you can have that clean things up... Regards, Bjorn
diff --git a/drivers/gpu/drm/msm/adreno/a5xx_gpu.c b/drivers/gpu/drm/msm/adreno/a5xx_gpu.c index eefe197..a7a58ec 100644 --- a/drivers/gpu/drm/msm/adreno/a5xx_gpu.c +++ b/drivers/gpu/drm/msm/adreno/a5xx_gpu.c @@ -469,6 +469,55 @@ static int a5xx_ucode_init(struct msm_gpu *gpu) return 0; } +static int a5xx_zap_shader_resume(struct msm_gpu *gpu) +{ + int ret; + + ret = qcom_scm_gpu_zap_resume(); + if (ret) + DRM_ERROR("%s: zap-shader resume failed: %d\n", + gpu->name, ret); + + return ret; +} + +static int a5xx_zap_shader_init(struct msm_gpu *gpu) +{ + static bool loaded; + struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu); + struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu); + struct platform_device *pdev = a5xx_gpu->pdev; + struct device_node *node; + int ret; + + /* + * If the zap shader is already loaded into memory we just need to kick + * the remote processor to reinitialize it + */ + if (loaded) + return a5xx_zap_shader_resume(gpu); + + /* Populate the sub-nodes if they haven't already been done */ + of_platform_populate(pdev->dev.of_node, NULL, NULL, &pdev->dev); + + /* Find the sub-node for the zap shader */ + node = of_find_node_by_name(pdev->dev.of_node, "qcom,zap-shader"); + if (!node) { + DRM_ERROR("%s: qcom,zap-shader not found in device tree\n", + gpu->name); + return -ENODEV; + } + + ret = _pil_tz_load_image(of_find_device_by_node(node)); + if (ret) + DRM_ERROR("%s: Unable to load the zap shader\n", + gpu->name); + + loaded = !ret; + + return ret; +} + #define A5XX_INT_MASK (A5XX_RBBM_INT_0_MASK_RBBM_AHB_ERROR | \ A5XX_RBBM_INT_0_MASK_RBBM_TRANSFER_TIMEOUT | \ A5XX_RBBM_INT_0_MASK_RBBM_ME_MS_TIMEOUT | \ @@ -653,8 +702,27 @@ static int a5xx_hw_init(struct msm_gpu *gpu) return -EINVAL; } - /* Put the GPU into unsecure mode */ - gpu_write(gpu, REG_A5XX_RBBM_SECVID_TRUST_CNTL, 0x0); + /* + * Try to load a zap shader into the secure world. If successful + * we can use the CP to switch out of secure mode. If not then we + * have no resource but to try to switch ourselves out manually. If we + * guessed wrong then access to the RBBM_SECVID_TRUST_CNTL register will + * be blocked and a permissions violation will soon follow. + */ + ret = a5xx_zap_shader_init(gpu); + if (!ret) { + OUT_PKT7(gpu->rb, CP_SET_SECURE_MODE, 1); + OUT_RING(gpu->rb, 0x00000000); + + gpu->funcs->flush(gpu); + if (!gpu->funcs->idle(gpu)) + return -EINVAL; + } else { + /* Print a warning so if we die, we know why */ + dev_warn_once(gpu->dev->dev, + "Zap shader not enabled - using SECVID_TRUST_CNTL instead\n"); + gpu_write(gpu, REG_A5XX_RBBM_SECVID_TRUST_CNTL, 0x0); + } return 0; }
The A5XX GPU powers on in "secure" mode. In secure mode the GPU can only render to buffers that are marked as secure and inaccessible to the kernel and user through a series of hardware protections. In practice secure mode is used to draw things like a UI on a secure video frame. In order to switch out of secure mode the GPU executes a special shader that clears out the GMEM and other sensitve registers and then writes a register. Because the kernel can't be trusted the shader binary is signed and verified and programmed by the secure world. To do this we need to read the MDT header and the segments from the firmware location and put them in memory and present them for approval. For targets without secure support there is an out: if the secure world doesn't support secure then there are no hardware protections and we can freely write the SECVID_TRUST register from the CPU. We don't have 100% confidence that we can query the secure capabilities at run time but we have enough calls that need to go right to give us some confidence that we're at least doing something useful. Of course if we guess wrong you trigger a permissions violation which usually ends up in a system crash but thats a problem that shows up immediately. Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org> --- drivers/gpu/drm/msm/adreno/a5xx_gpu.c | 72 ++++++++++++++++++++++++++++++++++- 1 file changed, 70 insertions(+), 2 deletions(-)