Message-ID: <20170228071632.GG25384@codeaurora.org>
Date:   Mon, 27 Feb 2017 23:16:32 -0800
From:   Stephen Boyd <sboyd@...eaurora.org>
To:     Avaneesh Kumar Dwivedi <akdwived@...eaurora.org>
Cc:     bjorn.andersson@...aro.org, agross@...eaurora.org,
        linux-arm-msm@...r.kernel.org, linux-kernel@...r.kernel.org,
        linux-remoteproc@...r.kernel.org
Subject: Re: [PATCH v2 2/3] remoteproc: qcom: Add scm call to protect modem
 mem in mss rproc drv.

On 01/30, Avaneesh Kumar Dwivedi wrote:
> This patch add hypervisor call support for second stage translation from
> mss remoteproc driver, this is required so that modem on msm8996 which is
> based on armv8 architecture can access DDR region where modem firmware
> are loaded.
> 
> Signed-off-by: Avaneesh Kumar Dwivedi <akdwived@...eaurora.org>
> ---

Most of this should be combined with patch 1 from this series.

>  drivers/remoteproc/qcom_q6v5_pil.c | 202 ++++++++++++++++++++++++++++++++++++-
>  1 file changed, 198 insertions(+), 4 deletions(-)
> 
> diff --git a/drivers/remoteproc/qcom_q6v5_pil.c b/drivers/remoteproc/qcom_q6v5_pil.c
> index e5edefa..35eee68 100644
> --- a/drivers/remoteproc/qcom_q6v5_pil.c
> +++ b/drivers/remoteproc/qcom_q6v5_pil.c
> @@ -93,6 +93,23 @@
>  #define QDSS_BHS_ON			BIT(21)
>  #define QDSS_LDO_BYP			BIT(22)
>  
> +struct dest_vm_and_perm_info {
> +	__le32 vm;
> +	__le32 perm;
> +	__le32 *ctx;
> +	__le32 ctx_size;
> +};
> +
> +struct mem_prot_info {
> +	__le64 addr;
> +	__le64 size;
> +};
> +
> +struct scm_desc {
> +	__le32 arginfo;
> +	__le64 args[10];
> +};

These are all firmware specific things that should live in scm
code, not PIL code.

> +
>  struct reg_info {
>  	struct regulator *reg;
>  	int uV;
> @@ -111,6 +128,7 @@ struct rproc_hexagon_res {
>  	struct qcom_mss_reg_res active_supply[2];
>  	char **proxy_clk_names;
>  	char **active_clk_names;
> +	int version;
>  };
>  
>  struct q6v5 {
> @@ -152,8 +170,29 @@ struct q6v5 {
>  	phys_addr_t mpss_reloc;
>  	void *mpss_region;
>  	size_t mpss_size;
> +	int version;
> +	int protection_cmd;
> +};
> +
> +enum {
> +	MSS_MSM8916,
> +	MSS_MSM8974,
> +	MSS_MSM8996,
>  };
>  
> +enum {
> +	ASSIG_ACCESS_MSA,
> +	REMOV_ACCESS_MSA,
> +	SHARE_ACCESS_MSA,
> +	REMOV_SHARE_MSA,
> +};
> +
> +#define VMID_HLOS	0x3
> +#define VMID_MSS_MSA	0xF
> +#define PERM_READ	0x4
> +#define PERM_WRITE	0x2
> +#define PERM_EXEC	0x1
> +#define PERM_RW	(0x4 | 0x2)

Please USE PERM_READ | PERM_WRITE here instead.

>  static int q6v5_regulator_init(struct device *dev, struct reg_info *regs,
>  				const struct qcom_mss_reg_res *reg_res)
>  {
> @@ -273,12 +312,123 @@ static void q6v5_clk_disable(struct device *dev,
>  		clk_disable_unprepare(clks[i]);
>  }
>  
> +int hyp_mem_access(struct q6v5 *qproc, phys_addr_t addr, size_t size)

This could be the scm API for now. But instead of taking qproc,
it would take the protection_cmd variable (which looks sort of
like a state machine BTW). To be more generic, perhaps it should
take the src vmids + num src vmids, dst vmids + num dst vmids,
and protection flags (which looks to be same size as dst vmid
array). If we can get the right SCM API here everything else
should fall into place.

> +{
> +	unsigned long dma_attrs = DMA_ATTR_FORCE_CONTIGUOUS;
> +	struct dest_vm_and_perm_info *dest_info;
> +	struct mem_prot_info *mem_prot_info;
> +	struct scm_desc desc = {0};
> +	__le32 *source_vm_copy;
> +	__le64 mem_prot_phy;
> +	int dest_count = 1;
> +	int src_count = 1;
> +	__le32 *perm = {0};
> +	__le32 *dest = {0};
> +	__le32 *src = {0};

src/dst seem like pretty confusing names. It makes it sound like
a memcpy operation. Perhaps from/to is better? Or current/next.

> +	__le64 phys_src;
> +	__le64 phy_dest;
> +	int ret;
> +	int i;
> +
> +	if (qproc->version != MSS_MSM8996)
> +		return 0;
> +
> +	switch (qproc->protection_cmd) {
> +		case ASSIG_ACCESS_MSA: {
> +			src = (int[2]) {VMID_HLOS, 0};
> +			dest = (int[2]) {VMID_MSS_MSA, 0};
> +			perm = (int[2]) {PERM_READ | PERM_WRITE, 0};

Why have two element arrays when we're only using the first
element? Relying on default src_count of 1 is very confusing.

> +			break;
> +		}
> +		case REMOV_ACCESS_MSA: {
> +			src = (int[2]) {VMID_MSS_MSA, 0};
> +			dest = (int[2]) {VMID_HLOS, 0};
> +			perm = (int[2]) {PERM_READ | PERM_WRITE | PERM_EXEC, 0};

Same here.

> +			break;
> +		}
> +		case SHARE_ACCESS_MSA: {
> +			src = (int[2]) {VMID_HLOS, 0};
> +			dest = (int[2]) {VMID_HLOS, VMID_MSS_MSA};
> +			perm = (int[2]) {PERM_RW, PERM_RW};

Please add spaces around curly braces like { this }

> +			dest_count = 2;
> +			break;
> +		}
> +		case REMOV_SHARE_MSA: {

And write REMOVE_SHARE_MSA, ASSIGN_ACCESS_MSA, etc. instead of
dropping letters.

> +			src = (int[2]) {VMID_HLOS, VMID_MSS_MSA};
> +			dest = (int[2]) {VMID_HLOS, 0};
> +			perm = (int[2]) {PERM_READ | PERM_WRITE | PERM_EXEC, 0};
> +			src_count = 2;
> +			break;
> +		}
> +		default: {

And also drop curly braces around case statements.

> +			break;
> +		}
> +	}
> +
> +	source_vm_copy = dma_alloc_attrs(qproc->dev,

This should really allocate from the scm->dev instead of qproc.
We don't know if qproc has the same DMA attributes that the
firmware has, but we know that we're trying to relay information
to the firmware here, not the qproc.

> +				src_count*sizeof(*source_vm_copy), &phys_src,
> +				GFP_KERNEL, dma_attrs);
> +	if (!source_vm_copy) {
> +		dev_err(qproc->dev,
> +			"failed to allocate buffer to pass source vmid detail\n");
> +		return -ENOMEM;
> +	}
> +	memcpy(source_vm_copy, src, sizeof(*source_vm_copy) * src_count);
> +
> +	dest_info = dma_alloc_attrs(qproc->dev,
> +			dest_count*sizeof(*dest_info), &phy_dest,
> +			GFP_KERNEL, dma_attrs);
> +	if (!dest_info) {
> +		dev_err(qproc->dev,
> +			"failed to allocate buffer to pass destination vmid detail\n");
> +		return -ENOMEM;
> +	}
> +	for (i = 0; i < dest_count; i++) {
> +		dest_info[i].vm = dest[i];
> +		dest_info[i].perm = perm[i];

Needs to do a cpu_to_le32() somewhere. Please run sparse.

> +		dest_info[i].ctx = NULL;
> +		dest_info[i].ctx_size = 0;
> +	}

Perhaps we should just allocate these first (one or two elements
isn't a big difference) and then fill the details in directly.

> +
> +	mem_prot_info = dma_alloc_attrs(qproc->dev,
> +				sizeof(*mem_prot_info), &mem_prot_phy,
> +				GFP_KERNEL, dma_attrs);
> +	if (!dest_info) {
> +		dev_err(qproc->dev,
> +				"failed to allocate buffer to pass protected mem detail\n");
> +		return -ENOMEM;
> +	}
> +	mem_prot_info->addr = addr;
> +	mem_prot_info->size = size;
> +
> +	desc.args[0] = mem_prot_phy;
> +	desc.args[1] = sizeof(*mem_prot_info);
> +	desc.args[2] = phys_src;
> +	desc.args[3] = sizeof(*source_vm_copy) * src_count;
> +	desc.args[4] = phy_dest;
> +	desc.args[5] = dest_count*sizeof(*dest_info);

Please add spaces around '*'. Run checkpatch.

> +	desc.args[6] = 0;
> +
> +	ret = qcom_scm_assign_mem(&desc);
> +	if (ret)
> +		pr_info("%s: Failed to assign memory protection, ret = %d\n",

pr_err? dev_err?

> +			__func__, ret);
> +
> +	/* SCM call has returned free up buffers passed to secure domain */
> +	dma_free_attrs(qproc->dev, src_count*sizeof(*source_vm_copy),
> +		source_vm_copy, phys_src, dma_attrs);
> +	dma_free_attrs(qproc->dev, dest_count*sizeof(*dest_info),
> +		dest_info, phy_dest, dma_attrs);
> +	dma_free_attrs(qproc->dev, sizeof(*mem_prot_info), mem_prot_info,
> +		mem_prot_phy, dma_attrs);
> +	return ret;
> +}
> +
>  static int q6v5_load(struct rproc *rproc, const struct firmware *fw)
>  {
>  	struct q6v5 *qproc = rproc->priv;
>  
>  	memcpy(qproc->mba_region, fw->data, fw->size);
> -

Please remove this patch noise.

>  	return 0;
>  }

-- 
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project

Powered by blists - more mailing lists