| Message ID | 20230530193436.3833889-1-quic_bjorande@quicinc.com |
|---|---|
| Headers | show |
| Series | soc: qcom: rmtfs: Support dynamic allocation | expand |
On 30.05.2023 21:34, Bjorn Andersson wrote: > In some configurations, the exact placement of the rmtfs shared memory > region isn't so strict. In the current implementation the author of the > DeviceTree source is forced to make up a memory region. IIUC the test here would be... "works" / "doesn't", just as if one misplaced the fixed region? Does the downstream sharedmem-uio driver do any additional cryptic magic or does it simply rely on the vendor's cma/dma pool settings? Can we replicate its behavior to stop hardcoding rmtfs, period? > > Extend the rmtfs memory driver to relieve the author of this > responsibility by introducing support for using dynamic allocation in > the driver. > > Signed-off-by: Bjorn Andersson <quic_bjorande@quicinc.com> > --- > arch/arm64/boot/dts/qcom/sdm845-mtp.dts | 10 ++++ > drivers/soc/qcom/rmtfs_mem.c | 66 +++++++++++++++++++------ > 2 files changed, 61 insertions(+), 15 deletions(-) > > diff --git a/arch/arm64/boot/dts/qcom/sdm845-mtp.dts b/arch/arm64/boot/dts/qcom/sdm845-mtp.dts > index d1440b790fa6..e6191b8ba4c6 100644 > --- a/arch/arm64/boot/dts/qcom/sdm845-mtp.dts > +++ b/arch/arm64/boot/dts/qcom/sdm845-mtp.dts > @@ -12,6 +12,8 @@ > #include "pm8998.dtsi" > #include "pmi8998.dtsi" > > +/delete-node/ &rmtfs_mem; > + > / { > model = "Qualcomm Technologies, Inc. SDM845 MTP"; > compatible = "qcom,sdm845-mtp", "qcom,sdm845"; > @@ -48,6 +50,14 @@ vreg_s4a_1p8: pm8998-smps4 { > vin-supply = <&vph_pwr>; > }; > > + rmtfs { > + compatible = "qcom,rmtfs-mem"; > + > + qcom,alloc-size = <(2*1024*1024)>; > + qcom,client-id = <1>; > + qcom,vmid = <15>; > + }; This should have been a separate patch. > + > thermal-zones { > xo_thermal: xo-thermal { > polling-delay-passive = <0>; > diff --git a/drivers/soc/qcom/rmtfs_mem.c b/drivers/soc/qcom/rmtfs_mem.c > index f83811f51175..5f56ded9f905 100644 > --- a/drivers/soc/qcom/rmtfs_mem.c > +++ b/drivers/soc/qcom/rmtfs_mem.c > @@ -3,6 +3,8 @@ > * Copyright (c) 2017 Linaro Ltd. > */ > > +#include "linux/gfp_types.h" > +#include "linux/sizes.h" <>? > #include <linux/kernel.h> > #include <linux/cdev.h> > #include <linux/err.h> > @@ -168,23 +170,63 @@ static void qcom_rmtfs_mem_release_device(struct device *dev) > kfree(rmtfs_mem); > } > > +static int qcom_rmtfs_acquire_mem(struct device *dev, struct qcom_rmtfs_mem *rmtfs_mem) > +{ > + struct device_node *node = dev->of_node; > + struct reserved_mem *rmem; > + dma_addr_t dma_addr; > + void *mem; > + u32 size; > + int ret; > + > + rmem = of_reserved_mem_lookup(node); > + if (rmem) { > + rmtfs_mem->addr = rmem->base; > + rmtfs_mem->size = rmem->size; > + > + rmtfs_mem->base = devm_memremap(&rmtfs_mem->dev, rmtfs_mem->addr, > + rmtfs_mem->size, MEMREMAP_WC); > + if (IS_ERR(rmtfs_mem->base)) { > + dev_err(dev, "failed to remap rmtfs_mem region\n"); > + return PTR_ERR(rmtfs_mem->base); > + } > + > + return 0; > + } > + > + ret = of_property_read_u32(node, "qcom,alloc-size", &size); > + if (ret < 0) { > + dev_err(dev, "rmtfs of unknown size\n"); > + return -EINVAL; > + } > + > + /* > + * Ensure that the protected region isn't adjacent to other protected > + * regions by allocating an empty page on either side. > + */ > + mem = dma_alloc_coherent(dev, size + 2 * SZ_4K, &dma_addr, GFP_KERNEL); Should this be made pagesize-independent? Can we even run non-4K kernels on msm? Konrad > + if (mem) { > + rmtfs_mem->base = mem + SZ_4K; > + rmtfs_mem->addr = dma_addr + SZ_4K; > + rmtfs_mem->size = size; > + > + return 0; > + } > + > + dev_err(dev, "unable to allocate memory for rmtfs mem\n"); > + return -ENOMEM; > +} > + > static int qcom_rmtfs_mem_probe(struct platform_device *pdev) > { > struct device_node *node = pdev->dev.of_node; > struct qcom_scm_vmperm perms[NUM_MAX_VMIDS + 1]; > - struct reserved_mem *rmem; > struct qcom_rmtfs_mem *rmtfs_mem; > u32 client_id; > u32 vmid[NUM_MAX_VMIDS]; > int num_vmids; > int ret, i; > > - rmem = of_reserved_mem_lookup(node); > - if (!rmem) { > - dev_err(&pdev->dev, "failed to acquire memory region\n"); > - return -EINVAL; > - } > - > ret = of_property_read_u32(node, "qcom,client-id", &client_id); > if (ret) { > dev_err(&pdev->dev, "failed to parse \"qcom,client-id\"\n"); > @@ -196,22 +238,16 @@ static int qcom_rmtfs_mem_probe(struct platform_device *pdev) > if (!rmtfs_mem) > return -ENOMEM; > > - rmtfs_mem->addr = rmem->base; > rmtfs_mem->client_id = client_id; > - rmtfs_mem->size = rmem->size; > > device_initialize(&rmtfs_mem->dev); > rmtfs_mem->dev.parent = &pdev->dev; > rmtfs_mem->dev.groups = qcom_rmtfs_mem_groups; > rmtfs_mem->dev.release = qcom_rmtfs_mem_release_device; > > - rmtfs_mem->base = devm_memremap(&rmtfs_mem->dev, rmtfs_mem->addr, > - rmtfs_mem->size, MEMREMAP_WC); > - if (IS_ERR(rmtfs_mem->base)) { > - dev_err(&pdev->dev, "failed to remap rmtfs_mem region\n"); > - ret = PTR_ERR(rmtfs_mem->base); > + ret = qcom_rmtfs_acquire_mem(&pdev->dev, rmtfs_mem); > + if (ret < 0) > goto put_device; > - } > > cdev_init(&rmtfs_mem->cdev, &qcom_rmtfs_mem_fops); > rmtfs_mem->cdev.owner = THIS_MODULE;
On Tue, May 30, 2023 at 12:34:36PM -0700, Bjorn Andersson wrote: > In some configurations, the exact placement of the rmtfs shared memory > region isn't so strict. In the current implementation the author of the > DeviceTree source is forced to make up a memory region. > > Extend the rmtfs memory driver to relieve the author of this > responsibility by introducing support for using dynamic allocation in > the driver. > > Signed-off-by: Bjorn Andersson <quic_bjorande@quicinc.com> > --- > arch/arm64/boot/dts/qcom/sdm845-mtp.dts | 10 ++++ > drivers/soc/qcom/rmtfs_mem.c | 66 +++++++++++++++++++------ > 2 files changed, 61 insertions(+), 15 deletions(-) > > diff --git a/arch/arm64/boot/dts/qcom/sdm845-mtp.dts b/arch/arm64/boot/dts/qcom/sdm845-mtp.dts > index d1440b790fa6..e6191b8ba4c6 100644 > --- a/arch/arm64/boot/dts/qcom/sdm845-mtp.dts > +++ b/arch/arm64/boot/dts/qcom/sdm845-mtp.dts > @@ -12,6 +12,8 @@ > #include "pm8998.dtsi" > #include "pmi8998.dtsi" > > +/delete-node/ &rmtfs_mem; > + > / { > model = "Qualcomm Technologies, Inc. SDM845 MTP"; > compatible = "qcom,sdm845-mtp", "qcom,sdm845"; > @@ -48,6 +50,14 @@ vreg_s4a_1p8: pm8998-smps4 { > vin-supply = <&vph_pwr>; > }; > > + rmtfs { > + compatible = "qcom,rmtfs-mem"; > + > + qcom,alloc-size = <(2*1024*1024)>; > + qcom,client-id = <1>; > + qcom,vmid = <15>; > + }; > + Couldn't you just use the existing dynamic allocation of reserved-memory, without any driver changes? / { reserved-memory { rmtfs { compatible = "qcom,rmtfs-mem"; size = <0x0 (2*1024*1024)>; alignment = <0x0 ...>; // if you want a special one no-map; // don't we want to map this actually? qcom,client-id = <1>; qcom,vmid = <15>; }; }; }; You won't get the 4K empty pages but I guess you just have them because you allocate the memory without proper alignment? Related patch series where I propose using it for most firmware memory regions: https://lore.kernel.org/linux-arm-msm/20230510-dt-resv-bottom-up-v1-5-3bf68873dbed@gerhold.net/ Thanks, Stephan
On Tue, May 30, 2023 at 09:45:10PM +0200, Konrad Dybcio wrote: > > > On 30.05.2023 21:34, Bjorn Andersson wrote: > > In some configurations, the exact placement of the rmtfs shared memory > > region isn't so strict. In the current implementation the author of the > > DeviceTree source is forced to make up a memory region. > IIUC the test here would be... "works" / "doesn't", just as if one > misplaced the fixed region? > The patch makes no effort to clarify this part. > Does the downstream sharedmem-uio driver do any additional cryptic > magic or does it simply rely on the vendor's cma/dma pool settings? > Can we replicate its behavior to stop hardcoding rmtfs, period? > Alignment on that is the intention with this patchset. > > > > Extend the rmtfs memory driver to relieve the author of this > > responsibility by introducing support for using dynamic allocation in > > the driver. > > > > Signed-off-by: Bjorn Andersson <quic_bjorande@quicinc.com> > > --- > > arch/arm64/boot/dts/qcom/sdm845-mtp.dts | 10 ++++ > > drivers/soc/qcom/rmtfs_mem.c | 66 +++++++++++++++++++------ > > 2 files changed, 61 insertions(+), 15 deletions(-) > > > > diff --git a/arch/arm64/boot/dts/qcom/sdm845-mtp.dts b/arch/arm64/boot/dts/qcom/sdm845-mtp.dts > > index d1440b790fa6..e6191b8ba4c6 100644 > > --- a/arch/arm64/boot/dts/qcom/sdm845-mtp.dts > > +++ b/arch/arm64/boot/dts/qcom/sdm845-mtp.dts > > @@ -12,6 +12,8 @@ > > #include "pm8998.dtsi" > > #include "pmi8998.dtsi" > > > > +/delete-node/ &rmtfs_mem; > > + > > / { > > model = "Qualcomm Technologies, Inc. SDM845 MTP"; > > compatible = "qcom,sdm845-mtp", "qcom,sdm845"; > > @@ -48,6 +50,14 @@ vreg_s4a_1p8: pm8998-smps4 { > > vin-supply = <&vph_pwr>; > > }; > > > > + rmtfs { > > + compatible = "qcom,rmtfs-mem"; > > + > > + qcom,alloc-size = <(2*1024*1024)>; > > + qcom,client-id = <1>; > > + qcom,vmid = <15>; > > + }; > This should have been a separate patch. > Of course, I should have paid more attention when I did the last git add, to not include test code... > > + > > thermal-zones { > > xo_thermal: xo-thermal { > > polling-delay-passive = <0>; > > diff --git a/drivers/soc/qcom/rmtfs_mem.c b/drivers/soc/qcom/rmtfs_mem.c > > index f83811f51175..5f56ded9f905 100644 > > --- a/drivers/soc/qcom/rmtfs_mem.c > > +++ b/drivers/soc/qcom/rmtfs_mem.c > > @@ -3,6 +3,8 @@ > > * Copyright (c) 2017 Linaro Ltd. > > */ > > > > +#include "linux/gfp_types.h" > > +#include "linux/sizes.h" > <>? > > > #include <linux/kernel.h> > > #include <linux/cdev.h> > > #include <linux/err.h> > > @@ -168,23 +170,63 @@ static void qcom_rmtfs_mem_release_device(struct device *dev) > > kfree(rmtfs_mem); > > } > > > > +static int qcom_rmtfs_acquire_mem(struct device *dev, struct qcom_rmtfs_mem *rmtfs_mem) > > +{ > > + struct device_node *node = dev->of_node; > > + struct reserved_mem *rmem; > > + dma_addr_t dma_addr; > > + void *mem; > > + u32 size; > > + int ret; > > + > > + rmem = of_reserved_mem_lookup(node); > > + if (rmem) { > > + rmtfs_mem->addr = rmem->base; > > + rmtfs_mem->size = rmem->size; > > + > > + rmtfs_mem->base = devm_memremap(&rmtfs_mem->dev, rmtfs_mem->addr, > > + rmtfs_mem->size, MEMREMAP_WC); > > + if (IS_ERR(rmtfs_mem->base)) { > > + dev_err(dev, "failed to remap rmtfs_mem region\n"); > > + return PTR_ERR(rmtfs_mem->base); > > + } > > + > > + return 0; > > + } > > + > > + ret = of_property_read_u32(node, "qcom,alloc-size", &size); > > + if (ret < 0) { > > + dev_err(dev, "rmtfs of unknown size\n"); > > + return -EINVAL; > > + } > > + > > + /* > > + * Ensure that the protected region isn't adjacent to other protected > > + * regions by allocating an empty page on either side. > > + */ > > + mem = dma_alloc_coherent(dev, size + 2 * SZ_4K, &dma_addr, GFP_KERNEL); > Should this be made pagesize-independent? Can we even run non-4K kernels on msm? > Yes, I fixed the issue in UFS and I believe Alex corrected the bug in IPA. With that I've been able to boot the few platforms where I've tried it with 16KB PAGE_SIZE. That's however the Linux page size, the numbers here relates to things on the secure side. Regards, Bjorn > Konrad > > + if (mem) { > > + rmtfs_mem->base = mem + SZ_4K; > > + rmtfs_mem->addr = dma_addr + SZ_4K; > > + rmtfs_mem->size = size; > > + > > + return 0; > > + } > > + > > + dev_err(dev, "unable to allocate memory for rmtfs mem\n"); > > + return -ENOMEM; > > +} > > + > > static int qcom_rmtfs_mem_probe(struct platform_device *pdev) > > { > > struct device_node *node = pdev->dev.of_node; > > struct qcom_scm_vmperm perms[NUM_MAX_VMIDS + 1]; > > - struct reserved_mem *rmem; > > struct qcom_rmtfs_mem *rmtfs_mem; > > u32 client_id; > > u32 vmid[NUM_MAX_VMIDS]; > > int num_vmids; > > int ret, i; > > > > - rmem = of_reserved_mem_lookup(node); > > - if (!rmem) { > > - dev_err(&pdev->dev, "failed to acquire memory region\n"); > > - return -EINVAL; > > - } > > - > > ret = of_property_read_u32(node, "qcom,client-id", &client_id); > > if (ret) { > > dev_err(&pdev->dev, "failed to parse \"qcom,client-id\"\n"); > > @@ -196,22 +238,16 @@ static int qcom_rmtfs_mem_probe(struct platform_device *pdev) > > if (!rmtfs_mem) > > return -ENOMEM; > > > > - rmtfs_mem->addr = rmem->base; > > rmtfs_mem->client_id = client_id; > > - rmtfs_mem->size = rmem->size; > > > > device_initialize(&rmtfs_mem->dev); > > rmtfs_mem->dev.parent = &pdev->dev; > > rmtfs_mem->dev.groups = qcom_rmtfs_mem_groups; > > rmtfs_mem->dev.release = qcom_rmtfs_mem_release_device; > > > > - rmtfs_mem->base = devm_memremap(&rmtfs_mem->dev, rmtfs_mem->addr, > > - rmtfs_mem->size, MEMREMAP_WC); > > - if (IS_ERR(rmtfs_mem->base)) { > > - dev_err(&pdev->dev, "failed to remap rmtfs_mem region\n"); > > - ret = PTR_ERR(rmtfs_mem->base); > > + ret = qcom_rmtfs_acquire_mem(&pdev->dev, rmtfs_mem); > > + if (ret < 0) > > goto put_device; > > - } > > > > cdev_init(&rmtfs_mem->cdev, &qcom_rmtfs_mem_fops); > > rmtfs_mem->cdev.owner = THIS_MODULE;
On Tue, May 30, 2023 at 09:48:46PM +0200, Stephan Gerhold wrote: > On Tue, May 30, 2023 at 12:34:36PM -0700, Bjorn Andersson wrote: > > In some configurations, the exact placement of the rmtfs shared memory > > region isn't so strict. In the current implementation the author of the > > DeviceTree source is forced to make up a memory region. > > > > Extend the rmtfs memory driver to relieve the author of this > > responsibility by introducing support for using dynamic allocation in > > the driver. > > > > Signed-off-by: Bjorn Andersson <quic_bjorande@quicinc.com> > > --- > > arch/arm64/boot/dts/qcom/sdm845-mtp.dts | 10 ++++ > > drivers/soc/qcom/rmtfs_mem.c | 66 +++++++++++++++++++------ > > 2 files changed, 61 insertions(+), 15 deletions(-) > > > > diff --git a/arch/arm64/boot/dts/qcom/sdm845-mtp.dts b/arch/arm64/boot/dts/qcom/sdm845-mtp.dts > > index d1440b790fa6..e6191b8ba4c6 100644 > > --- a/arch/arm64/boot/dts/qcom/sdm845-mtp.dts > > +++ b/arch/arm64/boot/dts/qcom/sdm845-mtp.dts > > @@ -12,6 +12,8 @@ > > #include "pm8998.dtsi" > > #include "pmi8998.dtsi" > > > > +/delete-node/ &rmtfs_mem; > > + > > / { > > model = "Qualcomm Technologies, Inc. SDM845 MTP"; > > compatible = "qcom,sdm845-mtp", "qcom,sdm845"; > > @@ -48,6 +50,14 @@ vreg_s4a_1p8: pm8998-smps4 { > > vin-supply = <&vph_pwr>; > > }; > > > > + rmtfs { > > + compatible = "qcom,rmtfs-mem"; > > + > > + qcom,alloc-size = <(2*1024*1024)>; > > + qcom,client-id = <1>; > > + qcom,vmid = <15>; > > + }; > > + > > Couldn't you just use the existing dynamic allocation of > reserved-memory, without any driver changes? > That should give us a similar end result, but we have alloc-ranges as well, if the placement needs to be further refined... > / { > reserved-memory { > rmtfs { > compatible = "qcom,rmtfs-mem"; > size = <0x0 (2*1024*1024)>; > alignment = <0x0 ...>; // if you want a special one > no-map; // don't we want to map this actually? > > qcom,client-id = <1>; > qcom,vmid = <15>; > }; > }; > }; > > You won't get the 4K empty pages but I guess you just have them because > you allocate the memory without proper alignment? > With dynamic placement there's no guarantee that the region isn't physically adjacent to another protected region, so this needs to be handled somehow. Perhaps the intention to include guard pages can be derived from the size... > Related patch series where I propose using it for most firmware memory > regions: > https://lore.kernel.org/linux-arm-msm/20230510-dt-resv-bottom-up-v1-5-3bf68873dbed@gerhold.net/ > Thanks for the suggestion, Bjorn > Thanks, > Stephan