Hello Simon,

Thanks for the feedback, I'll fix the changes in my v2. and sorry for these delayed answers.

On Sun, 26 Jul 2020 at 16:55, Simon Glass sjg@chromium.org wrote:

Hi Etienne,

On Fri, 17 Jul 2020 at 09:38, Etienne Carriere etienne.carriere@linaro.org wrote:

...

This change introduces SCMI agent driver in U-Boot in the firmware U-class.

SCMI agent driver is designed for platforms that embed a SCMI server in a firmware hosted for example by a companion co-processor or the secure world of the executing processor.

SCMI protocols allow an SCMI agent to discover and access external resources as clock, reset controllers and many more. SCMI agent and server communicate following the SCMI specification [1]. SCMI agent complies with the DT bindings defined in the Linux kernel source tree regarding SCMI agent description since v5.8-rc1.

These bindings describe 2 supported message transport layer: using mailbox uclass devices or using Arm SMC invocation instruction. Both use a piece or shared memory for message data exchange.

In the current state, the SCMI agent driver does not bind to any SCMI protocol to a U-Boot device driver. Former changes will implement dedicated driver (i.e. an SCMI clock driver or an SCMI reset controller driver) and add bind supported SCMI protocols in scmi_agent_bind().

Links: [1] https://developer.arm.com/architectures/system-architectures/software-standa... Signed-off-by: Etienne Carriere etienne.carriere@linaro.org

---

drivers/firmware/Kconfig | 15 ++ drivers/firmware/Makefile | 1 + drivers/firmware/scmi.c | 439 ++++++++++++++++++++++++++++++++++++++ include/scmi.h | 82 +++++++ 4 files changed, 537 insertions(+) create mode 100644 drivers/firmware/scmi.c create mode 100644 include/scmi.h

diff --git a/drivers/firmware/Kconfig b/drivers/firmware/Kconfig index b70a2063551..f7c7ee7a5aa 100644 --- a/drivers/firmware/Kconfig +++ b/drivers/firmware/Kconfig @@ -1,6 +1,21 @@ config FIRMWARE bool "Enable Firmware driver support"

+config SCMI_FIRMWARE

•   bool "Enable SCMI support"
  

•   select FIRMWARE
  

•   select OF_TRANSLATE
  

•   depends on DM_MAILBOX || ARM_SMCCC
  

•   help
  

•     An SCMI agent communicates with a related SCMI server firmware
  

Please write out SCMI in full somewhere and add a link to the spec.

...

•     located in another sub-system, as a companion micro controller
  

•     or a companion host in the CPU system.
  

•     Communications between agent (client) and the SCMI server are
  

•     based on message exchange. Messages can be exchange over tranport
  

•     channels as a mailbox device or an Arm SMCCC service with some
  

•     piece of identified shared memory.
  

config SPL_FIRMWARE bool "Enable Firmware driver support in SPL" depends on FIRMWARE diff --git a/drivers/firmware/Makefile b/drivers/firmware/Makefile index a0c250a473e..3965838179f 100644 --- a/drivers/firmware/Makefile +++ b/drivers/firmware/Makefile @@ -2,4 +2,5 @@ obj-$(CONFIG_FIRMWARE) += firmware-uclass.o obj-$(CONFIG_$(SPL_)ARM_PSCI_FW) += psci.o obj-$(CONFIG_TI_SCI_PROTOCOL) += ti_sci.o obj-$(CONFIG_SANDBOX) += firmware-sandbox.o +obj-$(CONFIG_SCMI_FIRMWARE) += scmi.o obj-$(CONFIG_ZYNQMP_FIRMWARE) += firmware-zynqmp.o diff --git a/drivers/firmware/scmi.c b/drivers/firmware/scmi.c new file mode 100644 index 00000000000..fa8a91c3f3d --- /dev/null +++ b/drivers/firmware/scmi.c @@ -0,0 +1,439 @@ +// SPDX-License-Identifier: GPL-2.0+ +/*

• • Copyright (c) 2015-2019, Arm Limited and Contributors. All rights reserved.
• • Copyright (C) 2019-2020 Linaro Limited.
• */

+#include <common.h> +#include <cpu_func.h> +#include <dm.h> +#include <errno.h> +#include <mailbox.h> +#include <memalign.h> +#include <scmi.h> +#include <asm/system.h> +#include <asm/types.h> +#include <dm/device-internal.h> +#include <dm/devres.h> +#include <dm/lists.h> +#include <dm/ofnode.h> +#include <linux/arm-smccc.h> +#include <linux/compat.h> +#include <linux/errno.h> +#include <linux/io.h> +#include <linux/ioport.h>

+#define TIMEOUT_US_10MS 10000

+struct error_code {

Function comment

...

•   int scmi;
  

•   int errno;
  

+};

+static const struct error_code scmi_linux_errmap[] = {

•   { .scmi = SCMI_NOT_SUPPORTED, .errno = -EOPNOTSUPP, },
  

•   { .scmi = SCMI_INVALID_PARAMETERS, .errno = -EINVAL, },
  

•   { .scmi = SCMI_DENIED, .errno = -EACCES, },
  

•   { .scmi = SCMI_NOT_FOUND, .errno = -ENOENT, },
  

•   { .scmi = SCMI_OUT_OF_RANGE, .errno = -ERANGE, },
  

•   { .scmi = SCMI_BUSY, .errno = -EBUSY, },
  

•   { .scmi = SCMI_COMMS_ERROR, .errno = -ECOMM, },
  

•   { .scmi = SCMI_GENERIC_ERROR, .errno = -EIO, },
  

•   { .scmi = SCMI_HARDWARE_ERROR, .errno = -EREMOTEIO, },
  

•   { .scmi = SCMI_PROTOCOL_ERROR, .errno = -EPROTO, },
  

+};

+int scmi_to_linux_errno(s32 scmi_code) +{

•   int n;
  

•   if (scmi_code == 0)
  

•           return 0;
  

•   for (n = 0; n < ARRAY_SIZE(scmi_linux_errmap); n++)
  

•           if (scmi_code == scmi_linux_errmap[n].scmi)
  

•                   return scmi_linux_errmap[1].errno;
  

•   return -EPROTO;
  

+}

+struct method_ops {

•   int (*process_msg)(struct udevice *dev, struct scmi_msg *msg);
  

•   int (*remove_agent)(struct udevice *dev);
  

Looks like this should be a new uclass and child driver?

I think a uclass is overkilling here. A scmi agent device is a firmware driver that binds to devices of other uclasses (clock, reset, ...). When scmi agent device is probed to get the related clock, reset, ... devices be probed. So it does not need any child. (see also related comment below)

Also we should have a sandbox implementation so this code can be tested.

Ok, I understand the sandbox test stuff. Should it be part of the v2 this series, can it be done in parallel, or a later v<N>?

...

+};

+struct scmi_agent {

•   struct method_ops *method_ops;
  

•   void *method_priv;
  

+};

+/*

• • Shared Memory based Transport (SMT) message buffer management
• • SMT uses 28 byte header prior message payload to handle the state of
• • the communication channel realized by the shared memory area and
• • to define SCMI protocol information the payload relates to.
• */

+struct scmi_smt_header {

•   __le32 reserved;
  

•   __le32 channel_status;
  

+#define SCMI_SHMEM_CHAN_STAT_CHANNEL_ERROR BIT(1) +#define SCMI_SHMEM_CHAN_STAT_CHANNEL_FREE BIT(0)

•   __le32 reserved1[2];
  

•   __le32 flags;
  

+#define SCMI_SHMEM_FLAG_INTR_ENABLED BIT(0)

•   __le32 length;
  

•   __le32 msg_header;
  

•   u8 msg_payload[0];
  

+};

+#define SMT_HEADER_TOKEN(token) (((token) << 18) & GENMASK(31, 18)) +#define SMT_HEADER_PROTOCOL_ID(proto) (((proto) << 10) & GENMASK(17, 10)) +#define SMT_HEADER_MESSAGE_TYPE(type) (((type) << 18) & GENMASK(9, 8)) +#define SMT_HEADER_MESSAGE_ID(id) ((id) & GENMASK(7, 0))

+struct scmi_shm_buf {

comment

Ok. I will add inline comments for all the defined structs and move macros/structs to source file top.

...

•   u8 *buf;
  

•   size_t size;
  

+};

+static int get_shm_buffer(struct udevice *dev, struct scmi_shm_buf *shm) +{

•   int rc;
  

Can you use 'ret' as 'rc' is a bit short for a var, and we mostly use 'ret' with driver model.

Ok, I'll try to remember to adapt my habits :)

...

•   struct ofnode_phandle_args args;
  

•   struct resource resource;
  

•   fdt32_t faddr;
  

•   phys_addr_t paddr;
  

•   rc = dev_read_phandle_with_args(dev, "shmem", NULL, 0, 0, &args);
  

•   if (rc)
  

•           return rc;
  

•   rc = ofnode_read_resource(args.node, 0, &resource);
  

•   if (rc)
  

•           return rc;
  

•   faddr = cpu_to_fdt32(resource.start);
  

•   paddr = ofnode_translate_address(args.node, &faddr);
  

•   shm->size = resource_size(&resource);
  

•   if (shm->size < sizeof(struct scmi_smt_header)) {
  

•           dev_err(dev, "Shared memory buffer too small\n");
  

•           return -EINVAL;
  

•   }
  

•   shm->buf = devm_ioremap(dev, paddr, shm->size);
  

•   if (!shm->buf)
  

•           return -ENOMEM;
  

•   if (dcache_status())
  

•           mmu_set_region_dcache_behaviour((uintptr_t)shm->buf,
  

•                                           shm->size, DCACHE_OFF);
  

•   return 0;
  

+}

+static int write_msg_to_smt(struct udevice *dev, struct scmi_shm_buf *shm_buf,

•                       struct scmi_msg *msg)
  

function comment with errors

...

+{

•   struct scmi_smt_header *hdr = (void *)shm_buf->buf;
  

•   if ((!msg->in_msg && msg->in_msg_sz) ||
  

•       (!msg->out_msg && msg->out_msg_sz))
  

•           return -EINVAL;
  

•   if (!(hdr->channel_status & SCMI_SHMEM_CHAN_STAT_CHANNEL_FREE)) {
  

•           dev_dbg(dev, "Channel busy\n");
  

•           return -EBUSY;
  

•   }
  

•   if (shm_buf->size < (sizeof(*hdr) + msg->in_msg_sz) ||
  

•       shm_buf->size < (sizeof(*hdr) + msg->out_msg_sz)) {
  

•           dev_dbg(dev, "Buffer too small\n");
  

•           return -ETOOSMALL;
  

•   }
  

•   /* Load message in shared memory */
  

•   hdr->channel_status &= ~SCMI_SHMEM_CHAN_STAT_CHANNEL_FREE;
  

•   hdr->length = msg->in_msg_sz + sizeof(hdr->msg_header);
  

•   hdr->msg_header = SMT_HEADER_TOKEN(0) |
  

•                     SMT_HEADER_MESSAGE_TYPE(0) |
  

•                     SMT_HEADER_PROTOCOL_ID(msg->protocol_id) |
  

•                     SMT_HEADER_MESSAGE_ID(msg->message_id);
  

•   memcpy(hdr->msg_payload, msg->in_msg, msg->in_msg_sz);
  

•   return 0;
  

+}

+static int read_resp_from_smt(struct udevice *dev, struct scmi_shm_buf *shm_buf,

•                         struct scmi_msg *msg)
  

+{

•   struct scmi_smt_header *hdr = (void *)shm_buf->buf;
  

•   if (!(hdr->channel_status & SCMI_SHMEM_CHAN_STAT_CHANNEL_FREE)) {
  

•           dev_err(dev, "Channel unexpectedly busy\n");
  

•           return -EBUSY;
  

•   }
  

•   if (hdr->channel_status & SCMI_SHMEM_CHAN_STAT_CHANNEL_ERROR) {
  

•           dev_err(dev, "Channel error reported, reset channel\n");
  

•           return -ECOMM;
  

•   }
  

•   if (hdr->length > msg->out_msg_sz + sizeof(hdr->msg_header)) {
  

•           dev_err(dev, "Buffer to small\n");
  

•           return -ETOOSMALL;
  

•   }
  

•   /* Get the data */
  

•   msg->out_msg_sz = hdr->length - sizeof(hdr->msg_header);
  

•   memcpy(msg->out_msg, hdr->msg_payload, msg->out_msg_sz);
  

•   return 0;
  

+}

+static void clear_smt_channel(struct scmi_shm_buf *shm_buf) +{

•   struct scmi_smt_header *hdr = (void *)shm_buf->buf;
  

•   hdr->channel_status &= ~SCMI_SHMEM_CHAN_STAT_CHANNEL_ERROR;
  

+}

+struct scmi_mbox_channel {

please put structs at top of file

...

•   struct scmi_shm_buf shm_buf;
  

•   struct mbox_chan mbox;
  

•   ulong timeout_us;
  

+};

+static int mbox_process_msg(struct udevice *dev, struct scmi_msg *msg) +{

•   struct scmi_agent *agent = dev_get_priv(dev);
  

•   struct scmi_mbox_channel *chan = agent->method_priv;
  

•   int rc;
  

•   rc = write_msg_to_smt(dev, &chan->shm_buf, msg);
  

•   if (rc)
  

•           return rc;
  

•   /* Give shm addr to mbox in case it is meaningful */
  

•   rc = mbox_send(&chan->mbox, chan->shm_buf.buf);
  

•   if (rc) {
  

•           dev_err(dev, "Message send failed: %d\n", rc);
  

•           goto out;
  

•   }
  

•   /* Receive the response */
  

•   rc = mbox_recv(&chan->mbox, chan->shm_buf.buf, chan->timeout_us);
  

•   if (rc) {
  

•           dev_err(dev, "Response failed: %d, abort\n", rc);
  

•           goto out;
  

•   }
  

•   rc = read_resp_from_smt(dev, &chan->shm_buf, msg);
  

+out:

•   clear_smt_channel(&chan->shm_buf);
  

•   return rc;
  

+}

+struct method_ops mbox_channel_ops = {

•   .process_msg = mbox_process_msg,
  

+};

+static int probe_mailbox_channel(struct udevice *dev) +{

•   struct scmi_agent *agent = dev_get_priv(dev);
  

•   struct scmi_mbox_channel *chan;
  

•   int rc;
  

•   chan = devm_kzalloc(dev, sizeof(*chan), GFP_KERNEL);
  

•   if (!chan)
  

•           return -ENOMEM;
  

•   chan->timeout_us = TIMEOUT_US_10MS;
  

•   rc = mbox_get_by_index(dev, 0, &chan->mbox);
  

•   if (rc) {
  

•           dev_err(dev, "Failed to find mailbox: %d\n", rc);
  

•           goto out;
  

•   }
  

•   rc = get_shm_buffer(dev, &chan->shm_buf);
  

•   if (rc)
  

•           dev_err(dev, "Failed to get shm resources: %d\n", rc);
  

+out:

•   if (rc) {
  

•           devm_kfree(dev, chan);
  

return rc

...

•   } else {
  

•           agent->method_ops = &mbox_channel_ops;
  

•           agent->method_priv = (void *)chan;
  

•   }
  

•   return rc;
  

return 0

(to keep the normal non-error flow as the main one)

...

+}

+struct scmi_arm_smc_channel {

•   ulong func_id;
  

•   struct scmi_shm_buf shm_buf;
  

+};

+#define SMCCC_RET_NOT_SUPPORTED ((unsigned long)-1)

Please struct and #define decls to top of file.

...

+static int arm_smc_process_msg(struct udevice *dev, struct scmi_msg *msg) +{

•   struct scmi_agent *agent = dev_get_priv(dev);
  

•   struct scmi_arm_smc_channel *chan = agent->method_priv;
  

•   struct arm_smccc_res res;
  

•   int rc;
  

•   rc = write_msg_to_smt(dev, &chan->shm_buf, msg);
  

•   if (rc)
  

•           return rc;
  

•   arm_smccc_smc(chan->func_id, 0, 0, 0, 0, 0, 0, 0, &res);
  

•   if (res.a0 == SMCCC_RET_NOT_SUPPORTED)
  

•           rc = -EINVAL;
  

ENOTSUP?

ENOTSUP stands for protocols. Here, it is the communication link that is bugged. Maybe ENXIO?

...

•   else
  

•           rc = read_resp_from_smt(dev, &chan->shm_buf, msg);
  

•   clear_smt_channel(&chan->shm_buf);
  

•   return rc;
  

+}

+struct method_ops arm_smc_channel_ops = {

•   .process_msg = arm_smc_process_msg,
  

+};

+static int probe_arm_smc_channel(struct udevice *dev) +{

•   struct scmi_agent *agent = dev_get_priv(dev);
  

•   struct scmi_arm_smc_channel *chan;
  

•   ofnode node = dev_ofnode(dev);
  

•   u32 func_id;
  

•   int rc;
  

•   chan = devm_kzalloc(dev, sizeof(*chan), GFP_KERNEL);
  

•   if (!chan)
  

•           return -ENOMEM;
  

•   if (ofnode_read_u32(node, "arm,smc-id", &func_id)) {
  

dev_read_u32()

...

•           dev_err(dev, "Missing property func-id\n");
  

•           return -EINVAL;
  

•   }
  

•   chan->func_id = func_id;
  

•   rc = get_shm_buffer(dev, &chan->shm_buf);
  

•   if (rc) {
  

•           dev_err(dev, "Failed to get shm resources: %d\n", rc);
  

•           return rc;
  

•   }
  

•   agent->method_ops = &arm_smc_channel_ops;
  

•   agent->method_priv = (void *)chan;
  

•   return rc;
  

+}

+/*

• • Exported functions by the SCMI agent
• */

+int scmi_send_and_process_msg(struct udevice *dev, struct scmi_msg *msg) +{

•   struct scmi_agent *agent = dev_get_priv(dev);
  

•   return agent->method_ops->process_msg(dev, msg);
  

+}

+static int scmi_remove(struct udevice *dev) +{

•   struct scmi_agent *agent = dev_get_priv(dev);
  

•   if (agent->method_ops->remove_agent)
  

•           return agent->method_ops->remove_agent(dev);
  

method_ops should be the uclass operations. I think you need a new uclass or add something to UCLASS_FIRMWARE.

Hmmm, I see you point. I'll consider. Yet, i think it would be a uclass with a unique type of child: a scmi agent driver. Your thoughts?

...

•   return 0;
  

+}

+enum scmi_transport_channel {

•   SCMI_MAILBOX_TRANSPORT,
  

•   SCMI_ARM_SMCCC_TRANSPORT,
  

+};

+static int scmi_probe(struct udevice *dev) +{

•   switch (dev_get_driver_data(dev)) {
  

•   case SCMI_MAILBOX_TRANSPORT:
  

•           if (IS_ENABLED(CONFIG_DM_MAILBOX))
  

•                   return probe_mailbox_channel(dev);
  

•           break;
  

•   case SCMI_ARM_SMCCC_TRANSPORT:
  

•           if (IS_ENABLED(CONFIG_ARM_SMCCC))
  

•                   return probe_arm_smc_channel(dev);
  

•           break;
  

•   default:
  

•           break;
  

•   }
  

•   return -EINVAL;
  

+}

+static int scmi_bind(struct udevice *dev) +{

•   int rc = 0;
  

•   ofnode node;
  

•   struct driver *drv;
  

•   dev_for_each_subnode(node, dev) {
  

•           u32 protocol_id;
  

•           if (!ofnode_is_available(node))
  

•                   continue;
  

•           if (ofnode_read_u32(node, "reg", &protocol_id))
  

•                   continue;
  

•           switch (protocol_id) {
  

•           default:
  

•                   dev_info(dev, "Ignore unsupported SCMI protocol %u\n",
  

•                            protocol_id);
  

•                   continue;
  

•           }
  

•           rc = device_bind_ofnode(dev, drv, ofnode_get_name(node),
  

•                                   NULL, node, NULL);
  

Can we instead add a compatible string to the device tree so driver model creates these children automatically?

Here the bound device already belongs to an existing uclass driver, at least we expect them to. These are clock, reset, power_domain uclass.

...

•           if (rc)
  

•                   break;
  

•   }
  

•   if (rc)
  

•           device_unbind(dev);
  

You should not unbind the device within its bind() method. Just make sure it refuses to probe.

Ok, thanks.

...

•   return rc;
  

+}

+static const struct udevice_id scmi_ids[] = { +#ifdef CONFIG_DM_MAILBOX

•   { .compatible = "arm,scmi", .data = SCMI_MAILBOX_TRANSPORT },
  

+#endif +#ifdef CONFIG_ARM_SMCCC

•   { .compatible = "arm,scmi-smc", .data = SCMI_ARM_SMCCC_TRANSPORT },
  

+#endif

Seems like you need a uclass for the transport?

...

•   { }
  

+};

+U_BOOT_DRIVER(scmi) = {

•   .name           = "scmi",
  

•   .id             = UCLASS_FIRMWARE,
  

•   .of_match       = scmi_ids,
  

•   .priv_auto_alloc_size = sizeof(struct scmi_agent),
  

•   .bind           = scmi_bind,
  

•   .probe          = scmi_probe,
  

•   .remove         = scmi_remove,
  

•   .flags          = DM_FLAG_OS_PREPARE,
  

+}; diff --git a/include/scmi.h b/include/scmi.h new file mode 100644 index 00000000000..e12e322991d --- /dev/null +++ b/include/scmi.h @@ -0,0 +1,82 @@ +/* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */ +/*

• • Copyright (c) 2015-2019, Arm Limited and Contributors. All rights reserved.
• • Copyright (C) 2019, Linaro Limited
• */

+#ifndef SCMI_H +#define SCMI_H

+#include <asm/types.h>

+/**

• • An SCMI agent represent on communication path from a device driver to
• • the remote SCMI server which driver sends messages to and receives
• • response messages from.
• */

+struct scmi_agent;

+enum scmi_std_protocol {

•   SCMI_PROTOCOL_ID_BASE = 0x10,
  

•   SCMI_PROTOCOL_ID_POWER_DOMAIN = 0x11,
  

•   SCMI_PROTOCOL_ID_SYSTEM = 0x12,
  

•   SCMI_PROTOCOL_ID_PERF = 0x13,
  

•   SCMI_PROTOCOL_ID_CLOCK = 0x14,
  

•   SCMI_PROTOCOL_ID_SENSOR = 0x15,
  

•   SCMI_PROTOCOL_ID_RESET_DOMAIN = 0x16,
  

+};

+enum scmi_status_code {

•   SCMI_SUCCESS =  0,
  

•   SCMI_NOT_SUPPORTED = -1,
  

•   SCMI_INVALID_PARAMETERS = -2,
  

•   SCMI_DENIED = -3,
  

•   SCMI_NOT_FOUND = -4,
  

•   SCMI_OUT_OF_RANGE = -5,
  

•   SCMI_BUSY = -6,
  

•   SCMI_COMMS_ERROR = -7,
  

•   SCMI_GENERIC_ERROR = -8,
  

•   SCMI_HARDWARE_ERROR = -9,
  

•   SCMI_PROTOCOL_ERROR = -10,
  

+};

+/*

• • struct scmi_msg - Context of a SCMI message sent and the response received
• • @protocol_id: SCMI protocol ID
• • @message_id: SCMI message ID for a defined protocol ID
• • @in_msg: pointer to the message payload sent by the driver
• • @in_msg_sz: byte size of the message payload sent
• • @out_msg: pointer to buffer to store response message payload
• • @out_msg_size: Byte size of the response buffer or payload

sz

...

• */

+struct scmi_msg {

•   unsigned int protocol_id;
  

•   unsigned int message_id;
  

Does this need a particular size, e,g. u32?

...

•   u8 *in_msg;
  

•   size_t in_msg_sz;
  

•   u8 *out_msg;
  

•   size_t out_msg_sz;
  

+};

+/**

• • scmi_send_and_process_msg() - send and process a SCMI message
• • Send a message to a SCMI server through a target SCMI agent device.
• • Caller sets scmi_msg::out_msg_sz to the output message buffer size.
• • On return, scmi_msg::out_msg_sz stores the response payload size.
• • @dev: SCMI agent device
• • @msg: Message structure reference
• • @return 0 on success, a negative errno otherwise
• */

+int scmi_send_and_process_msg(struct udevice *dev, struct scmi_msg *msg);

+/**

• • scmi_to_linux_errno() - Convert an SCMI error code into a Linux errno code
• • @scmi_errno: SCMI error code value
• • @return 0 for successful status and a negative errno otherwise
• */

+int scmi_to_linux_errno(s32 scmi_errno);

+#endif /* SCMI_H */

2.17.1

Regards, Simon