Re: [RESEND v5 6/7] remoteproc: qcom: Add support for memory sandbox
From: Srinivasa Rao Mandadapu
Date: Thu Sep 08 2022 - 09:00:27 EST
On 8/25/2022 3:34 PM, Srinivasa Rao Mandadapu wrote:
On 8/23/2022 8:55 AM, Stephen Boyd wrote:
Thanks for your time and valuable suggestions!!!
Quoting Srinivasa Rao Mandadapu (2022-08-22 01:22:02)
diff --git a/drivers/remoteproc/qcom_q6v5_adsp.c
b/drivers/remoteproc/qcom_q6v5_adsp.c
index b0a63a0..d01c97e 100644
--- a/drivers/remoteproc/qcom_q6v5_adsp.c
+++ b/drivers/remoteproc/qcom_q6v5_adsp.c
@@ -333,6 +336,185 @@ static int adsp_load(struct rproc *rproc,
const struct firmware *fw)
return 0;
}
+static void adsp_of_unmap_smmu(struct iommu_domain *iommu_dom,
const __be32 *prop, int len)
+{
+ unsigned long iova;
+ unsigned int mem_size;
+ int i;
+
+ len /= sizeof(__be32);
+ for (i = 0; i < len; i++) {
+ iova = be32_to_cpu(prop[i++]);
+ /* Skip Physical address*/
+ i++;
+ mem_size = be32_to_cpu(prop[i++]);
+ iommu_unmap(iommu_dom, iova, mem_size);
+ }
+}
+
+static void adsp_rproc_unmap_smmu(struct rproc *rproc, int len)
+{
+ struct fw_rsc_devmem *rsc_fw;
+ struct fw_rsc_hdr *hdr;
+ int offset;
+ int i;
+
+ for (i = 0; i < len; i++) {
+ offset = rproc->table_ptr->offset[i];
+ hdr = (void *)rproc->table_ptr + offset;
+ rsc_fw = (struct fw_rsc_devmem *)hdr + sizeof(*hdr);
+
+ iommu_unmap(rproc->domain, rsc_fw->da, rsc_fw->len);
+ }
+}
+
+static void adsp_unmap_smmu(struct rproc *rproc)
+{
+ struct qcom_adsp *adsp = (struct qcom_adsp *)rproc->priv;
+ const __be32 *prop;
+ unsigned int len;
+
+ iommu_unmap(adsp->iommu_dom, adsp->mem_phys, adsp->mem_size);
+
+ prop = of_get_property(adsp->dev->of_node,
"qcom,adsp-memory-regions", &len);
+ if (prop) {
+ adsp_of_unmap_smmu(adsp->iommu_dom, prop, len);
+ } else {
+ if (rproc->table_ptr)
+ adsp_rproc_unmap_smmu(rproc,
rproc->table_ptr->num);
+ }
+
+ iommu_domain_free(adsp->iommu_dom);
+}
+
+static int adsp_of_map_smmu(struct iommu_domain *iommu_dom, const
__be32 *prop, int len)
+{
+ unsigned long mem_phys;
+ unsigned long iova;
+ unsigned int mem_size;
+ unsigned int flag;
+ int access_level;
+ int ret;
+ int i;
+
+ len /= sizeof(__be32);
+ for (i = 0; i < len; i++) {
+ iova = be32_to_cpu(prop[i++]);
+ mem_phys = be32_to_cpu(prop[i++]);
+ mem_size = be32_to_cpu(prop[i++]);
+ access_level = be32_to_cpu(prop[i]);
+
+ if (access_level)
+ flag = IOMMU_READ | IOMMU_WRITE;
+ else
+ flag = IOMMU_READ;
+
+ ret = iommu_map(iommu_dom, iova, mem_phys, mem_size,
flag);
+ if (ret) {
+ pr_err("failed to map addr = %p mem_size =
%x\n", &(mem_phys), mem_size);
Why can't this be dev_err()?
Actually, dev pointer is not available here, hence used pr_err.
+ goto of_smmu_unmap;
+ }
+ }
+ return 0;
+of_smmu_unmap:
+ adsp_of_unmap_smmu(iommu_dom, prop, i);
+ return ret;
+}
+
+static int adsp_rproc_map_smmu(struct rproc *rproc, int len)
+{
+ struct fw_rsc_devmem *rsc_fw;
const?
Okay. will update.
+ struct fw_rsc_hdr *hdr;
const?
Okay. Will update.
+ int offset;
+ int ret;
+ int i;
+
+ if (!rproc->table_ptr)
+ return 0;
+
+ for (i = 0; i < rproc->table_ptr->num; i++) {
+ offset = rproc->table_ptr->offset[i];
+ hdr = (void *)rproc->table_ptr + offset;
+ rsc_fw = (struct fw_rsc_devmem *)hdr + sizeof(*hdr);
+
+ ret = iommu_map(rproc->domain, rsc_fw->da, rsc_fw->pa,
+ rsc_fw->len, rsc_fw->flags);
+ if (ret) {
+ pr_err("failed to map addr = %x mem_size =
%x\n", rsc_fw->pa, rsc_fw->len);
Why can't this be dev_err()?
Okay. Will change it.
+ goto rproc_smmu_unmap;
+ }
+ }
+
+ return 0;
+
+rproc_smmu_unmap:
+ adsp_rproc_unmap_smmu(rproc, i);
Does i need to be incremented? And/or unmap should be done in reverse.
Here i is the upper bound index in mapping failure case, hence it is
used as length. un-mapping is being done from starting till i value.
Please correct me if I am missing some thing.
+ return ret;
+}
+
+static int adsp_map_smmu(struct qcom_adsp *adsp, struct rproc *rproc)
+{
+ struct of_phandle_args args;
+ const __be32 *prop;
+ long long sid;
+ unsigned int len;
+ int ret;
+
+ ret = of_parse_phandle_with_args(adsp->dev->of_node,
"iommus", "#iommu-cells", 0, &args);
+ if (ret < 0)
+ sid = -1;
Is it a good idea to set the sid to -1? Does that mean all stream IDs?
It seems, if sid is -1 iomap fails, because of alignment issues. Any I
will update with return in this case.
+ else
+ sid = args.args[0] & SID_MASK_DEFAULT;
+
+ adsp->iommu_dom = iommu_domain_alloc(adsp->dev->bus);
+ if (!adsp->iommu_dom) {
+ dev_err(adsp->dev, "failed to allocate iommu
domain\n");
+ ret = -ENOMEM;
+ goto domain_free;
+ }
+
+ ret = iommu_attach_device(adsp->iommu_dom, adsp->dev);
+ if (ret) {
+ dev_err(adsp->dev, "could not attach device ret =
%d\n", ret);
+ ret = -EBUSY;
Why do we overwrite the error value?
It seems not required. Will remove it.
+ goto detach_device;
+ }
+
+ /* Add SID configuration for ADSP Firmware to SMMU */
+ adsp->mem_phys = adsp->mem_phys | (sid << 32);
+
+ ret = iommu_map(adsp->iommu_dom, adsp->mem_phys,
adsp->mem_phys,
+ adsp->mem_size, IOMMU_READ | IOMMU_WRITE);
+ if (ret) {
+ dev_err(adsp->dev, "Unable to map ADSP Physical
Memory\n");
+ goto sid_unmap;
+ }
+
+ prop = of_get_property(adsp->dev->of_node,
"qcom,adsp-memory-regions", &len);
I find it odd that we're encoding virtual addresses (iovas) into
Actually from HLOS perspective, same IOVA and physical memory is being
used. Hence will remove virtual address field in DT.
devicetree. Presumably the physical address needs to be in DT as a
carveout, but after that I would think we're free to allocate the
Will try to carveout the physical addresses and handle it accordingly.
If this method is ignored I don't think we need to mention in DT in
anyway.
segments from the carveout however we see fit and then program that into
the SMMU. Maybe DT can be a suggestion, but otherwise can it be ignored?
Yes, it seems it can be ignored. As it is the approach we did for
bringing up AudioReach solution, and used when ADSP binary is without
metadata section header info.
Will check internally and update accordingly.
Your opinion also helps please!!. Is it okay to keep it as backup
option with proper comment, such that this method can be used
internally, with raw ADSP binary in early stage bring-up scenarios?
After internal discussions, decided to remove sandboxing using device
tree and explicitly doing in PIL driver.
Instead, decided to use rproc's parse_fw call back
"rproc_elf_load_rsc_table" function and achieve the sand boxing.
Will re post the patches with corresponding changes.
+ if (prop) {
+ ret = adsp_of_map_smmu(adsp->iommu_dom, prop, len);
+ if (ret) {
+ dev_err(adsp->dev, "Unable to map memory
regions accessed by ADSP\n");
+ goto sid_unmap;
+ }
+ } else {
+ ret = adsp_rproc_map_smmu(rproc, len);
+ if (ret) {
+ dev_err(adsp->dev, "Unable to map memory
regions accessed by ADSP\n");
Maybe this should be a different string in case it is confused with the
above print of the same string.
Okay. Will modify the string.
As explained above will remove both of the above methods.
+ goto sid_unmap;
+ }
+ }
+ return 0;
+
+sid_unmap:
+ iommu_unmap(adsp->iommu_dom, adsp->mem_phys, adsp->mem_size);
+detach_device:
+ iommu_domain_free(adsp->iommu_dom);
+domain_free:
+ return ret;
+}
+
+
static int adsp_start(struct rproc *rproc)
{
struct qcom_adsp *adsp = (struct qcom_adsp *)rproc->priv;
@@ -343,9 +525,16 @@ static int adsp_start(struct rproc *rproc)
if (ret)
return ret;
+ if (adsp->adsp_sandbox_needed) {
+ ret = adsp_map_smmu(adsp, rproc);
+ if (ret) {
+ dev_err(adsp->dev, "ADSP smmu mapping
failed\n");
+ goto disable_irqs;
+ }
+ }
Newline here please.
Okay.
ret = clk_prepare_enable(adsp->xo);
if (ret)
- goto disable_irqs;
+ goto adsp_smmu_unmap;
ret = qcom_rproc_pds_enable(adsp, adsp->proxy_pds,
adsp->proxy_pd_count);