Re: [RFC ADD HYPERVISOR CALL SUPPORT 1/2] soc: qcom: Add hypervisor stage two translation request support.
From: Stephen Boyd
Date: Sat Jan 21 2017 - 04:09:38 EST
On 01/09, Avaneesh Kumar Dwivedi wrote:
> This patch add hypervisor support for mss bring up on msm8996.
> MSS rproc driver make hypervisor request to add certain region
> in IPA table owned by hepervisor and assign access permission
Please drop the use of IPA here. There's an IPA acronym for the
IP accelerator and that is confused with Intermediate Physical
Address. Instead, say something like "in the stage 2 page tables
of the SMMU". Also hypervisor is misspelled here.
> to modem. These regions are used to load MBA, MDT, FW into DDR.
>
> Signed-off-by: Avaneesh Kumar Dwivedi <akdwived@xxxxxxxxxxxxxx>
> ---
> drivers/firmware/qcom_scm-64.c | 16 +++
> drivers/firmware/qcom_scm.c | 13 +++
> drivers/firmware/qcom_scm.h | 10 ++
> drivers/remoteproc/qcom_q6v5_pil.c | 96 +++++++++++++++-
Please split the remoteproc code off from this patch.
> drivers/soc/qcom/Kconfig | 8 ++
> drivers/soc/qcom/Makefile | 1 +
> drivers/soc/qcom/secure_buffer.c | 229 +++++++++++++++++++++++++++++++++++++
> include/linux/qcom_scm.h | 1 +
> include/soc/qcom/secure_buffer.h | 51 +++++++++
> 9 files changed, 419 insertions(+), 6 deletions(-)
> create mode 100644 drivers/soc/qcom/secure_buffer.c
> create mode 100644 include/soc/qcom/secure_buffer.h
>
> diff --git a/drivers/firmware/qcom_scm-64.c b/drivers/firmware/qcom_scm-64.c
> index 4a0f5ea..63b814c 100644
> --- a/drivers/firmware/qcom_scm-64.c
> +++ b/drivers/firmware/qcom_scm-64.c
> @@ -358,3 +358,19 @@ int __qcom_scm_pas_mss_reset(struct device *dev, bool reset)
>
> return ret ? : res.a1;
> }
> +
> +int __qcom_scm_assign_mem(struct device *dev, void *data)
> +{
> + int ret;
> + struct qcom_scm_desc *desc = (struct qcom_scm_desc *)data;
Useless cast from void.
> + struct arm_smccc_res res;
> +
> + desc->arginfo = QCOM_SCM_ARGS(7, SCM_RO, SCM_VAL, SCM_RO,
> + SCM_VAL, SCM_RO, SCM_VAL, SCM_VAL);
> +
> + ret = qcom_scm_call(dev, QCOM_SCM_SVC_MP,
> + QCOM_MEM_PROT_ASSIGN_ID,
> + desc, &res);
> +
> + return ret ? : res.a1;
> +}
> diff --git a/drivers/firmware/qcom_scm.c b/drivers/firmware/qcom_scm.c
> index 893f953ea..52394ac 100644
> --- a/drivers/firmware/qcom_scm.c
> +++ b/drivers/firmware/qcom_scm.c
> @@ -292,6 +292,19 @@ int qcom_scm_pas_shutdown(u32 peripheral)
> }
> EXPORT_SYMBOL(qcom_scm_pas_shutdown);
>
> +/**
> + * qcom_scm_assign_mem() -
Some words on what the function does?
> + * @desc: descriptor to send to hypervisor
> + *
> + * Return 0 on success.
> + */
> +int qcom_scm_assign_mem(void *desc)
> +{
> +
Nitpick: Drop the extra newline
> + return __qcom_scm_assign_mem(__scm->dev, desc);
> +}
> +EXPORT_SYMBOL(qcom_scm_assign_mem);
> +
> static int qcom_scm_pas_reset_assert(struct reset_controller_dev *rcdev,
> unsigned long idx)
> {
> diff --git a/drivers/firmware/qcom_scm.h b/drivers/firmware/qcom_scm.h
> index 3584b00..f248dc9 100644
> --- a/drivers/firmware/qcom_scm.h
> +++ b/drivers/firmware/qcom_scm.h
> @@ -47,6 +47,8 @@ extern int __qcom_scm_hdcp_req(struct device *dev,
> #define QCOM_SCM_PAS_SHUTDOWN_CMD 0x6
> #define QCOM_SCM_PAS_IS_SUPPORTED_CMD 0x7
> #define QCOM_SCM_PAS_MSS_RESET 0xa
> +#define QCOM_SCM_SVC_MP 0xc
> +#define QCOM_MEM_PROT_ASSIGN_ID 0x16
> extern bool __qcom_scm_pas_supported(struct device *dev, u32 peripheral);
> extern int __qcom_scm_pas_init_image(struct device *dev, u32 peripheral,
> dma_addr_t metadata_phys);
> @@ -55,6 +57,7 @@ extern int __qcom_scm_pas_mem_setup(struct device *dev, u32 peripheral,
> extern int __qcom_scm_pas_auth_and_reset(struct device *dev, u32 peripheral);
> extern int __qcom_scm_pas_shutdown(struct device *dev, u32 peripheral);
> extern int __qcom_scm_pas_mss_reset(struct device *dev, bool reset);
> +extern int __qcom_scm_assign_mem(struct device *dev, void *desc);
>
> /* common error codes */
> #define QCOM_SCM_V2_EBUSY -12
> @@ -83,4 +86,11 @@ static inline int qcom_scm_remap_error(int err)
> return -EINVAL;
> }
>
> +enum scm_arg_types {
> + SCM_VAL,
> + SCM_RO,
> + SCM_RW,
> + SCM_BUFVAL,
> +};
We already have this enum? It's just prepended with QCOM_
instead.
> +
> #endif
> diff --git a/drivers/remoteproc/qcom_q6v5_pil.c b/drivers/remoteproc/qcom_q6v5_pil.c
> index e5edefa..cbf833c 100644
> --- a/drivers/remoteproc/qcom_q6v5_pil.c
> +++ b/drivers/remoteproc/qcom_q6v5_pil.c
> @@ -31,6 +31,7 @@
> #include <linux/soc/qcom/smem.h>
> #include <linux/soc/qcom/smem_state.h>
> #include <linux/of_device.h>
> +#include <soc/qcom/secure_buffer.h>
>
> #include "remoteproc_internal.h"
> #include "qcom_mdt_loader.h"
> @@ -105,12 +106,19 @@ struct qcom_mss_reg_res {
> int uA;
> };
>
> +struct src_dest_vmid {
> + int *srcVM;
> + int *destVM;
> + int *destVMperm;
Why not have an array of these structures instead of a structure
with three arrays inside? Can you map one source VM to multiple
destination VMs?
> +};
> +
> struct rproc_hexagon_res {
> const char *hexagon_mba_image;
> struct qcom_mss_reg_res proxy_supply[4];
> struct qcom_mss_reg_res active_supply[2];
> char **proxy_clk_names;
> char **active_clk_names;
> + int version;
> };
>
> struct q6v5 {
> @@ -127,6 +135,8 @@ struct q6v5 {
>
> struct reset_control *mss_restart;
>
> + struct src_dest_vmid vmid_details;
> +
> struct qcom_smem_state *state;
> unsigned stop_bit;
>
> @@ -152,6 +162,13 @@ struct q6v5 {
> phys_addr_t mpss_reloc;
> void *mpss_region;
> size_t mpss_size;
> + int version;
> +};
> +
> +enum {
> + MSS_MSM8916,
> + MSS_MSM8974,
> + MSS_MSM8996,
> };
>
> static int q6v5_regulator_init(struct device *dev, struct reg_info *regs,
> @@ -273,13 +290,46 @@ 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)
> +{
> + int src_count = 0;
> + int dest_count = 0;
> + int ret;
> + int i;
> +
> + if (qproc->version != MSS_MSM8996)
> + return 0;
> +
> + for (i = 0; i < 2; i++) {
> + if (qproc->vmid_details.srcVM[i] != 0)
> + src_count++;
Bad tabbing here.
> + if (qproc->vmid_details.destVM[i] != 0)
> + dest_count++;
And here.
When is it ever == 0? The hardcoded 2 in the for loop is also
suspect.
> + }
Add newline
> + ret = hyp_assign_phys(qproc->dev, addr, size,
> + qproc->vmid_details.srcVM,
> + src_count, qproc->vmid_details.destVM,
> + qproc->vmid_details.destVMperm, dest_count);
At the least this API could take a vmid_details structure instead
of all these arguments:
struct vm_perms {
u32 vm;
u32 perm;
};
struct vmid_details {
u32 *src;
size_t src_count;
struct vm_perms *dest;
size_t dest_count;
};
> + if (ret)
> + dev_err(qproc->dev,
> + "%s: failed for %pa address of size %zx\n",
> + __func__, &addr, size);
> + return ret;
> +}
> +
[...]
> diff --git a/drivers/soc/qcom/Kconfig b/drivers/soc/qcom/Kconfig
> index 461b387..9459894 100644
> --- a/drivers/soc/qcom/Kconfig
> +++ b/drivers/soc/qcom/Kconfig
> @@ -76,3 +76,11 @@ config QCOM_WCNSS_CTRL
> help
> Client driver for the WCNSS_CTRL SMD channel, used to download nv
> firmware to a newly booted WCNSS chip.
> +
> +config QCOM_SECURE_BUFFER
Please no. Just fold this into the qcom_scm.c code and drop the
new config and new file.
> + bool "Helper functions for securing buffers through TZ"
> + help
> + Say 'Y' here for targets that need to call into TZ to secure
> + memory buffers. This ensures that only the correct clients can
> + use this memory and no unauthorized access is made to the
> + buffer
> diff --git a/drivers/soc/qcom/secure_buffer.c b/drivers/soc/qcom/secure_buffer.c
> new file mode 100644
> index 0000000..54eaa6f
> --- /dev/null
> +++ b/drivers/soc/qcom/secure_buffer.c
> @@ -0,0 +1,229 @@
> +/*
> + * Copyright (c) 2011-2017, The Linux Foundation. All rights reserved.
> + *
> + * This program is free software; you can redistribute it and/or modify
> + * it under the terms of the GNU General Public License version 2 and
> + * only version 2 as published by the Free Software Foundation.
> + *
> + * This program is distributed in the hope that it will be useful,
> + * but WITHOUT ANY WARRANTY; without even the implied warranty of
> + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
> + * GNU General Public License for more details.
> + *
> + */
> +
> +#include <linux/highmem.h>
Used?
> +#include <linux/kernel.h>
> +#include <linux/kref.h>
Used?
> +#include <linux/mutex.h>
> +#include <linux/scatterlist.h>
> +#include <linux/slab.h>
> +#include <linux/dma-mapping.h>
> +#include <linux/qcom_scm.h>
> +#include <linux/dma-mapping.h>
> +#include <soc/qcom/secure_buffer.h>
> +
> +DEFINE_MUTEX(secure_buffer_mutex);
static?
> +
> +struct scm_desc {
> + u32 arginfo;
> + u64 args[10];
> +};
This would be in qcom_scm-64.c already.
> +
> +struct mem_prot_info {
> + phys_addr_t addr;
> + u64 size;
Both should be __le64 presumably.
> +};
> +
> +struct info_list {
> + struct mem_prot_info *list_head;
> + u64 list_size;
> +};
> +
> +struct dest_vm_and_perm_info {
> + u32 vm;
> + u32 perm;
> + u32 *ctx;
Why is this a pointer? Is this structure put into memory and
passed to the firmware? We should use the appropriate __le32 and
__le64 types then.
> + u32 ctx_size;
> +};
> +
> +struct dest_info_list {
> + struct dest_vm_and_perm_info *dest_info;
> + u64 list_size;
> +};
These structures are confusing. From what I can tell we need an
array of struct mem_prot_info and an array of struct
dest_vm_and_perm_info in our pre-allocated buffer. The other
info_list and dest_info_list structures are bookkeeping things
that we pass to the firmware via registers. So let's drop the
info_list structures and have the functions that populate arrays
take pointers to the memory region to populate and return size_t?
static size_t populate_dest_info(struct dest_vm_and_perm_info *info,
struct int *dest_vmids, int nelements,
int *dest_perms);
static size_t populate_prot_info_from_table(struct mem_prot_info *info,
struct sg_table *table);
And then the callers can add the size returned to the memory
chunk pointer.
void *qcom_secure_mem;
struct mem_prot_info *prot_info = qcom_secure_mem;
struct dest_vm_and_perm_info *dest_info;
size_t size;
size = populate_prot_info_from_table(prot_info, ...);
...
dest_info = qcom_secure_mem + size;
size = populate_dest_info(dest_info, ...);
smc_call()...
}
> +
> +static void *qcom_secure_mem;
> +#define QCOM_SECURE_MEM_SIZE (512 * 1024)
> +#define PADDING 32
What is the padding for? cache aligning? 32 seems magical.
> +
> +static void populate_dest_info(int *dest_vmids, int nelements,
> + int *dest_perms, struct dest_info_list **list,
s/list/list_p/
> + void *current_qcom_secure_mem)
> +{
> + struct dest_vm_and_perm_info *dest_info;
> + int i;
> +
> + dest_info = (struct dest_vm_and_perm_info *)current_qcom_secure_mem;
Useless cast, please remove.
> +
> + for (i = 0; i < nelements; i++) {
> + dest_info[i].vm = dest_vmids[i];
> + dest_info[i].perm = dest_perms[i];
> + dest_info[i].ctx = NULL;
> + dest_info[i].ctx_size = 0;
> + }
> +
> + *list = (struct dest_info_list *)&dest_info[i];
Grow a local variable:
struct dest_info_list *list = *list_p;
list = &dest_info[i];
list->dest_info = dest_info;
list->list_size = nelements * sizeof(*dest_info);
Of course this may all change anyway.
> +
> + (*list)->dest_info = dest_info;
> + (*list)->list_size = nelements * sizeof(struct dest_vm_and_perm_info);
> +}
> +
> +static void get_info_list_from_table(struct sg_table *table,
> + struct info_list **list)
> +{
> + int i;
> + struct scatterlist *sg;
> + struct mem_prot_info *info;
> +
> + info = (struct mem_prot_info *)qcom_secure_mem;
Useless cast from void.
> +
> + for_each_sg(table->sgl, sg, table->nents, i) {
> + info[i].addr = page_to_phys(sg_page(sg));
> + info[i].size = sg->length;
> + }
> +
> + *list = (struct info_list *)&(info[i]);
> +
> + (*list)->list_head = info;
> + (*list)->list_size = table->nents * sizeof(struct mem_prot_info);
> +}
> +
> +int hyp_assign_table(struct device *dev, struct sg_table *table,
> + u32 *source_vm_list, int source_nelems,
> + int *dest_vmids, int *dest_perms,
> + int dest_nelems)
> +{
> + int ret;
> + struct info_list *info_list = NULL;
> + struct dest_info_list *dest_info_list = NULL;
> + struct scm_desc desc = {0};
> + u32 *source_vm_copy;
> + void *current_qcom_secure_mem;
> +
> + size_t reqd_size = dest_nelems * sizeof(struct dest_vm_and_perm_info) +
> + table->nents * sizeof(struct mem_prot_info) +
> + sizeof(dest_info_list) + sizeof(info_list) + PADDING;
> +
> + if (!qcom_secure_mem) {
> + pr_err("%s is not functional as qcom_secure_mem is not allocated.\n",
> + __func__);
> + return -ENOMEM;
> + }
> +
> + if (reqd_size > QCOM_SECURE_MEM_SIZE) {
> + pr_err("%s: Not enough memory allocated. Required size %zd\n",
> + __func__, reqd_size);
> + return -EINVAL;
> + }
> +
> + /*
> + * We can only pass cache-aligned sizes to hypervisor, so we need
> + * to kmalloc and memcpy the source_vm_list here.
> + */
> + source_vm_copy = kmalloc_array(
> + source_nelems, sizeof(*source_vm_copy), GFP_KERNEL);
> + if (!source_vm_copy)
> + return -ENOMEM;
> + memcpy(source_vm_copy, source_vm_list,
> + sizeof(*source_vm_list) * source_nelems);
All of these u32s need an endian swap on big-endian platforms
when they're copied.
> +
> + mutex_lock(&secure_buffer_mutex);
> +
> + get_info_list_from_table(table, &info_list);
> +
> + current_qcom_secure_mem = &(info_list[1]);
> + populate_dest_info(dest_vmids, dest_nelems, dest_perms,
> + &dest_info_list, current_qcom_secure_mem);
> +
> + desc.args[0] = virt_to_phys(info_list->list_head);
> + desc.args[1] = info_list->list_size;
> + desc.args[2] = virt_to_phys(source_vm_copy);
> + desc.args[3] = sizeof(*source_vm_copy) * source_nelems;
> + desc.args[4] = virt_to_phys(dest_info_list->dest_info);
> + desc.args[5] = dest_info_list->list_size;
> + desc.args[6] = 0;
> +
> + dma_set_mask(dev, DMA_BIT_MASK(64));
The dev passed here should be the scm->dev because we don't want
to allocate memory from a memory region that may be associated
with some random device using this API, and also we want to be
able to have the right mask set for communicating with the
firmware, which may be different than whatever mask is needed for
DMA with a device.
> + dma_map_single(dev, (void *)source_vm_copy,
Useless cast to void.
> + (size_t)(source_vm_copy + source_nelems),
This looks wrong? source_vm_copy is a pointer and we're adding
source_nelems and then casting that address to a size_t which
would be potentially very large? Shouldn't it just be
desc.args[3]?
> + DMA_TO_DEVICE);
> + dma_map_single(dev, (void *)info_list->list_head,
> + (size_t)(info_list->list_head +
> + (info_list->list_size / sizeof(*info_list->list_head))),
> + DMA_TO_DEVICE);
> + dma_map_single(dev,
> + (void *)dest_info_list->dest_info,
Useless cast to void.
> + (size_t)(dest_info_list->dest_info +
> + (dest_info_list->list_size /
> + sizeof(*dest_info_list->dest_info))),
> + DMA_TO_DEVICE);
> +
> + ret = qcom_scm_assign_mem((void *)&desc);
Useless cast to void.
> + if (ret)
> + pr_info("%s: Failed to assign memory protection, ret = %d\n",
pr_err?
> + __func__, ret);
> +
> + mutex_unlock(&secure_buffer_mutex);
Missing the dma_unmap calls here? Failure to do that could lead
to a leak if we bounce the page.
Also, shouldn't we skip the dma mapping if we have allocated
coherent memory instead of kmalloc for the qcom_secure_mem
buffer? The code seems to ignore the dma allocation case.
> + kfree(source_vm_copy);
> + return ret;
> +}
> +
> +int hyp_assign_phys(struct device *dev, phys_addr_t addr, u64 size,
> + u32 *source_vm_list, int source_nelems,
> + int *dest_vmids, int *dest_perms,
> + int dest_nelems)
> +{
> + struct sg_table *table;
> + int ret;
> +
> + table = kzalloc(sizeof(struct sg_table), GFP_KERNEL);
sizeof(*table)
> + if (!table)
> + return -ENOMEM;
Newline here
> + ret = sg_alloc_table(table, 1, GFP_KERNEL);
> + if (ret)
> + goto err1;
> +
> + sg_set_page(table->sgl, phys_to_page(addr), size, 0);
> +
> + ret = hyp_assign_table(dev, table, source_vm_list,
> + source_nelems, dest_vmids,
> + dest_perms, dest_nelems);
I'd prefer two user facing APIs exist. One that takes a single
address and size argument, and one that takes an sg_table. Both
APIs can then call some common function that passes that info off
to the firmware, but the first one can be used here without
requiring us to make an sg_table for no reason besides undoing the
sg_table in hyp_assign_table().
> + if (ret)
> + goto err2;
> +
> + return ret;
> +err2:
> + sg_free_table(table);
> +err1:
> + kfree(table);
> + return ret;
> +}
> +
> +static int __init alloc_secure_shared_memory(void)
> +{
> + int ret = 0;
ret is 0...
> +
> + qcom_secure_mem = kzalloc(QCOM_SECURE_MEM_SIZE, GFP_KERNEL);
> + if (!qcom_secure_mem) {
> + /* Fallback to CMA-DMA memory */
> + qcom_secure_mem = dma_alloc_coherent(NULL, QCOM_SECURE_MEM_SIZE,
We can't really pass NULL here anymore. Use the scm device.
> + NULL, GFP_KERNEL);
> + if (!qcom_secure_mem) {
> + pr_err("Couldn't allocate memory for secure use-cases. hyp_assign_table will not work\n");
Memory allocation messages are practically useless. Please remove
them.
> + return -ENOMEM;
> + }
> + }
> +
> + return ret;
And that's all for ret. Always 0.
> +}
> +pure_initcall(alloc_secure_shared_memory);
pure initcall? Maybe we should take the lazy approach and
allocate this big chunk once someone calls into the API the first
time? If we merge this with scm device, then we can probably do
the allocation when scm probes and we have a proper device for
dma allocation.
> diff --git a/include/soc/qcom/secure_buffer.h b/include/soc/qcom/secure_buffer.h
> new file mode 100644
> index 0000000..2c7015d
> --- /dev/null
> +++ b/include/soc/qcom/secure_buffer.h
> @@ -0,0 +1,51 @@
> +/*
> + * Copyright (c) 2015-2016, The Linux Foundation. All rights reserved.
> + *
> + * This program is free software; you can redistribute it and/or modify
> + * it under the terms of the GNU General Public License version 2 and
> + * only version 2 as published by the Free Software Foundation.
> + *
> + * This program is distributed in the hope that it will be useful,
> + * but WITHOUT ANY WARRANTY; without even the implied warranty of
> + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
> + * GNU General Public License for more details.
> + *
> + */
> +
> +#ifndef __MSM_SECURE_BUFFER_H__
> +#define __MSM_SECURE_BUFFER_H__
> +
> +enum vmid {
> + VMID_HLOS = 0x3,
> + VMID_MSS_MSA = 0xF,
> + VMID_INVAL = -1
Just drop the enum and use #defines please.
> +};
> +
> +#define PERM_READ 0x4
> +#define PERM_WRITE 0x2
> +#define PERM_EXEC 0x1
--
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project