[v5,09/14] firmware: arm_scmi: Add core Raw transmission support

Message ID 20221103195225.1028864-10-cristian.marussi@arm.com
State New
Headers
Series Introduce a unified API for SCMI Server testing |

Commit Message

Cristian Marussi Nov. 3, 2022, 7:52 p.m. UTC
  Add SCMI Raw mode support which exposes a userspace interface to allow for
bare SCMI command injection and snooping from userspace.

Signed-off-by: Cristian Marussi <cristian.marussi@arm.com>
---
v4 --> v5
- avoid using raw_buffers in scmi_raw_error_report, prefer
  dynalloc GFP_ATOMIC
- use custom tags for Raw msg_dump traces
- use xfer->is_raw to enforce segregation of Raw transactions
- add multiple SCMI server instance support. New Raw debugfs is at
	<debugfs>/scmi_raw/0/...
- fix possible race in scmi_raw_message_dequeue between list_empty
  check and effective dequeuing.
- add circular handling of raw buffers queues for errors and
  notifications
v3 --> v4
- fix handling of O_NONBLOCK raw_mode read requests
- add support for polled/nocomnpletion_irq transports
- removed inlines
v2 --> v3
- changed to dev_dbg when retrying inflight reg
- refactored and simplfiied waiters deferred worker
- fixed sparse errors about LE and __poll_t
v1 --> v2
- make use of proper channel depending on protocol at hand (using
  scmi_xfer_raw_channel_get to lookup related DT config)
- refactored scmi_inflight_register call
- added debugfs docs and comments
---
 drivers/firmware/arm_scmi/Kconfig    |   13 +
 drivers/firmware/arm_scmi/Makefile   |    1 +
 drivers/firmware/arm_scmi/driver.c   |    5 +-
 drivers/firmware/arm_scmi/raw_mode.c | 1337 ++++++++++++++++++++++++++
 drivers/firmware/arm_scmi/raw_mode.h |   30 +
 5 files changed, 1384 insertions(+), 2 deletions(-)
 create mode 100644 drivers/firmware/arm_scmi/raw_mode.c
 create mode 100644 drivers/firmware/arm_scmi/raw_mode.h
  

Patch

diff --git a/drivers/firmware/arm_scmi/Kconfig b/drivers/firmware/arm_scmi/Kconfig
index a14f65444b35..ab726a92ac2f 100644
--- a/drivers/firmware/arm_scmi/Kconfig
+++ b/drivers/firmware/arm_scmi/Kconfig
@@ -23,6 +23,19 @@  config ARM_SCMI_PROTOCOL
 
 if ARM_SCMI_PROTOCOL
 
+config ARM_SCMI_RAW_MODE_SUPPORT
+	bool "Enable support for SCMI Raw transmission mode"
+	help
+	  Enable support for SCMI Raw transmission mode.
+
+	  If enabled allows the direct injection and snooping of SCMI bare
+	  messages through a dedicated debugfs interface.
+	  It is meant to be used by SCMI compliance/testing suites.
+
+	  When enabled regular SCMI drivers interactions are inhibited in
+	  order to avoid unexpected interactions with the SCMI Raw message
+	  flow. If unsure say N.
+
 config ARM_SCMI_HAVE_TRANSPORT
 	bool
 	help
diff --git a/drivers/firmware/arm_scmi/Makefile b/drivers/firmware/arm_scmi/Makefile
index 9ea86f8cc8f7..7c1aca60c8ce 100644
--- a/drivers/firmware/arm_scmi/Makefile
+++ b/drivers/firmware/arm_scmi/Makefile
@@ -1,6 +1,7 @@ 
 # SPDX-License-Identifier: GPL-2.0-only
 scmi-bus-y = bus.o
 scmi-driver-y = driver.o notify.o
+scmi-driver-$(CONFIG_ARM_SCMI_RAW_MODE_SUPPORT) += raw_mode.o
 scmi-transport-$(CONFIG_ARM_SCMI_HAVE_SHMEM) = shmem.o
 scmi-transport-$(CONFIG_ARM_SCMI_TRANSPORT_MAILBOX) += mailbox.o
 scmi-transport-$(CONFIG_ARM_SCMI_TRANSPORT_SMC) += smc.o
diff --git a/drivers/firmware/arm_scmi/driver.c b/drivers/firmware/arm_scmi/driver.c
index 9dcfe096cb2b..7c7f1a666d93 100644
--- a/drivers/firmware/arm_scmi/driver.c
+++ b/drivers/firmware/arm_scmi/driver.c
@@ -796,7 +796,8 @@  static void scmi_handle_response(struct scmi_chan_info *cinfo,
 
 	trace_scmi_msg_dump(xfer->hdr.protocol_id, xfer->hdr.id,
 			    xfer->hdr.type == MSG_TYPE_DELAYED_RESP ?
-			    "DLYD" : "RESP",
+			    (!xfer->is_raw ? "DLYD" : "dlyd") :
+			    (!xfer->is_raw ? "RESP" : "resp"),
 			    xfer->hdr.seq, xfer->hdr.status,
 			    xfer->rx.buf, xfer->rx.len);
 
@@ -918,7 +919,7 @@  static int scmi_wait_for_reply(struct device *dev, const struct scmi_desc *desc,
 
 			/* Trace polled replies. */
 			trace_scmi_msg_dump(xfer->hdr.protocol_id, xfer->hdr.id,
-					    "RESP",
+					    !xfer->is_raw ? "RESP" : "resp",
 					    xfer->hdr.seq, xfer->hdr.status,
 					    xfer->rx.buf, xfer->rx.len);
 		}
diff --git a/drivers/firmware/arm_scmi/raw_mode.c b/drivers/firmware/arm_scmi/raw_mode.c
new file mode 100644
index 000000000000..deb345c740c5
--- /dev/null
+++ b/drivers/firmware/arm_scmi/raw_mode.c
@@ -0,0 +1,1337 @@ 
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * System Control and Management Interface (SCMI) Raw mode support
+ *
+ * Copyright (C) 2022 ARM Ltd.
+ */
+/**
+ * DOC: Theory of operation
+ *
+ * When enabled the SCMI Raw mode support exposes a userspace API which allows
+ * to send and receive SCMI commands, replies and notifications from a user
+ * application through injection and snooping of bare SCMI messages in binary
+ * little-endian format.
+ *
+ * Such injected SCMI transactions will then be routed through the SCMI core
+ * stack towards the SCMI backend server using whatever SCMI transport is
+ * currently configured on the system under test.
+ *
+ * It is meant to help in running any sort of SCMI backend server testing, no
+ * matter where the server is placed, as long as it is normally reachable via
+ * the transport configured on the system.
+ *
+ * It is activated by a Kernel configuration option since it is NOT meant to
+ * be used in production but only during development and in CI deployments.
+ *
+ * In order to avoid possible interferences between the SCMI Raw transactions
+ * originated from a test-suite and the normal operations of the SCMI drivers,
+ * when Raw mode is enabled, by default, all the regular SCMI drivers are
+ * inhibited.
+ *
+ * The exposed API is as follows.
+ * All defined SCMI instances are rooted under a common top scmi_raw debugfs
+ * entry with each instance, in turn, rooted under a sequentially numbered
+ * containing subdirectory like:
+ *
+ *	/sys/kernel/debug/scmi_raw/0/
+ *	|-- errors
+ *	|-- instance_name
+ *	|-- message
+ *	|-- message_async
+ *	|-- notification
+ *	|-- reset
+ *	|-- transport_max_msg_size
+ *	|-- transport_rx_timeout_ms
+ *	|-- transport_tx_max_msg
+ *
+ * where:
+ *
+ *  - errors: used to read back timed-out and unexpected replies
+ *  - instance_name: used to identify the underlying SCMI instance which is
+ *		     accessed by this Raw accessors; it corresponds to the
+ *		     full_name of the top SCMI DT node under which the SCMI
+ *		     instance is rooted.
+ *  - message*: used to send sync/async commands and read back immediate and
+ *		delayed reponses (if any)
+ *  - notification: used to read any notification being emitted by the system
+ *		    (if previously enabled by the user app)
+ *  - reset: used to flush the queues of messages (of any kind) still pending
+ *	     to be read; this is useful at test-suite start/stop to get
+ *	     rid of any unread messages from the previous run.
+ *  - transport*: a bunch of configurations useful to setup the user
+ *		  application expectations in terms of timeouts and message
+ *		  characteristics.
+ *
+ * Each write to the message* entries causes one command request to be built
+ * and sent while the replies or delayed response are read back from those same
+ * entries one message at time (receiving an EOF at each message boundary).
+ *
+ * The user application running the test is in charge of handling timeouts
+ * on replies and properly choosing SCMI sequence numbers for the outgoing
+ * requests (using the same sequence number is supported but discouraged).
+ *
+ * Injection of multiple in-flight requests is supported as long as the user
+ * application uses properly distinct sequence numbers for concurrent requests
+ * and takes care to properly manage all the related issues about concurrency
+ * and command/reply pairing. Keep in mind that, anyway, the real level of
+ * parallelism attainable in such scenario is dependent on the characteristics
+ * of the underlying transport being used.
+ *
+ * Since the SCMI core regular stack is partially used to deliver and collect
+ * the messages, late replies arrived after timeouts and any other sort of
+ * unexpected message can be identified by the SCMI core as usual and they will
+ * be reported as messages under "errors" for later analysis.
+ */
+
+#include <linux/atomic.h>
+#include <linux/bitmap.h>
+#include <linux/debugfs.h>
+#include <linux/delay.h>
+#include <linux/device.h>
+#include <linux/export.h>
+#include <linux/idr.h>
+#include <linux/io.h>
+#include <linux/kernel.h>
+#include <linux/fs.h>
+#include <linux/list.h>
+#include <linux/module.h>
+#include <linux/poll.h>
+#include <linux/of.h>
+#include <linux/slab.h>
+
+#include "common.h"
+
+#include "raw_mode.h"
+
+#include <trace/events/scmi.h>
+
+#define SCMI_XFER_RAW_MAX_RETRIES	10
+
+static atomic_t scmi_raw_instance_counter = ATOMIC_INIT(-1);
+
+/**
+ * struct scmi_raw_mode_info  - Structure holding SCMI Raw instance data
+ *
+ * @id: Sequential Raw instance ID.
+ * @name: Full name of the DT node associated to the SCMI instance which is
+ *	  represented by this Raw instance.
+ * @handle: Pointer to SCMI entity handle to use
+ * @desc: Pointer to the transport descriptor to use
+ * @tx_max_msg: Maximum number of concurrent TX in-flight messages
+ * @free_raw_bufs: An array of freelists heads used to keep unused raw buffers
+ * @free_bufs_lock: Spinlocks used to protect access to each @free_raw_bufs
+ * @msg_q: An array of lists to queue snooped messages waiting to be read out
+ * @msg_q_lock: Spinlocks used to protect access to each @msg_q
+ * @wqs: An array of waitqueues used to wait and poll on related @msg_q
+ * @free_waiters: Head of freelist for unused waiters
+ * @free_mtx: A mutex to protect the waiters freelist
+ * @active_waiters: Head of list for currently active and used waiters
+ * @active_mtx: A mutex to protect the active waiters list
+ * @waiters_work: A work descriptor to be used with the workqueue machinery
+ * @wait_wq: A workqueue reference to the created workqueue
+ * @dentry: Top debugfs root dentry for SCMI Raw
+ * @gid: A group ID used for devres accounting
+ *
+ * Note that this descriptor is passed back to the core after SCMI Raw is
+ * initialized as an opaque handle to use by subsequent SCMI Raw call hooks.
+ *
+ */
+struct scmi_raw_mode_info {
+	unsigned int id;
+	const char *name;
+	const struct scmi_handle *handle;
+	const struct scmi_desc *desc;
+	int tx_max_msg;
+	struct list_head free_raw_bufs[SCMI_RAW_MAX_QUEUE];
+	/* Protect free_raw_bufs[] lists */
+	spinlock_t free_bufs_lock[SCMI_RAW_MAX_QUEUE];
+	struct list_head msg_q[SCMI_RAW_MAX_QUEUE];
+	/* Protect msg_q[] lists */
+	spinlock_t msg_q_lock[SCMI_RAW_MAX_QUEUE];
+	wait_queue_head_t wqs[SCMI_RAW_MAX_QUEUE];
+	struct list_head free_waiters;
+	/* Protect free_waiters list */
+	struct mutex free_mtx;
+	struct list_head active_waiters;
+	/* Protect active_waiters list */
+	struct mutex active_mtx;
+	struct work_struct waiters_work;
+	struct workqueue_struct	*wait_wq;
+	struct dentry *dentry;
+	void *gid;
+};
+
+/**
+ * struct scmi_xfer_raw_waiter  - Structure to describe an xfer to be waited for
+ *
+ * @start_jiffies: The timestamp in jiffies of when this structure was queued.
+ * @cinfo: A reference to the channel to use for this transaction
+ * @xfer: A reference to the xfer to be waited for
+ * @async_response: A completion to be, optionally, used for async waits: it
+ *		    will be setup by @scmi_do_xfer_raw_start, if needed, to be
+ *		    pointed at by xfer->async_done.
+ * @node: A list node.
+ */
+struct scmi_xfer_raw_waiter {
+	unsigned long start_jiffies;
+	struct scmi_chan_info *cinfo;
+	struct scmi_xfer *xfer;
+	struct completion async_response;
+	struct list_head node;
+};
+
+/**
+ * struct scmi_raw_buffer  - Structure to hold a full SCMI message
+ *
+ * @max_len: The maximum allowed message size (header included) that can be
+ *	     stored into @msg
+ * @msg: A message buffer used to collect a full message grabbed from an xfer.
+ * @node: A list node.
+ */
+struct scmi_raw_buffer {
+	size_t max_len;
+	struct scmi_msg msg;
+	struct list_head node;
+};
+
+/**
+ * struct scmi_dbg_raw_data  - Structure holding data needed by the debugfs
+ * layer
+ *
+ * @raw: A reference to the Raw instance.
+ * @tx: A message buffer used to collect TX message on write.
+ * @tx_size: The effective size of the TX message.
+ * @tx_req_size: The final expected size of the complete TX message.
+ * @rx: A message buffer to collect RX message on read.
+ * @rx_size: The effective size of the RX message.
+ */
+struct scmi_dbg_raw_data {
+	struct scmi_raw_mode_info *raw;
+	struct scmi_msg tx;
+	size_t tx_size;
+	size_t tx_req_size;
+	struct scmi_msg rx;
+	size_t rx_size;
+};
+
+static struct scmi_raw_buffer *
+scmi_raw_buffer_get(struct scmi_raw_mode_info *raw, unsigned int idx)
+{
+	unsigned long flags;
+	struct scmi_raw_buffer *rb = NULL;
+	struct list_head *head = &raw->free_raw_bufs[idx];
+
+	spin_lock_irqsave(&raw->free_bufs_lock[idx], flags);
+	if (!list_empty(head)) {
+		rb = list_first_entry(head, struct scmi_raw_buffer, node);
+		list_del_init(&rb->node);
+	}
+	spin_unlock_irqrestore(&raw->free_bufs_lock[idx], flags);
+
+	return rb;
+}
+
+static void scmi_raw_buffer_put(struct scmi_raw_mode_info *raw,
+				struct scmi_raw_buffer *rb, unsigned int idx)
+{
+	unsigned long flags;
+
+	/* Reset to full buffer length */
+	rb->msg.len = rb->max_len;
+
+	spin_lock_irqsave(&raw->free_bufs_lock[idx], flags);
+	list_add_tail(&rb->node, &raw->free_raw_bufs[idx]);
+	spin_unlock_irqrestore(&raw->free_bufs_lock[idx], flags);
+}
+
+static void scmi_raw_buffer_enqueue(struct scmi_raw_mode_info *raw,
+				    struct scmi_raw_buffer *rb,
+				    unsigned int idx)
+{
+	unsigned long flags;
+
+	spin_lock_irqsave(&raw->msg_q_lock[idx], flags);
+	list_add_tail(&rb->node, &raw->msg_q[idx]);
+	spin_unlock_irqrestore(&raw->msg_q_lock[idx], flags);
+
+	wake_up_interruptible(&raw->wqs[idx]);
+}
+
+static struct scmi_raw_buffer*
+scmi_raw_buffer_dequeue_unlocked(struct scmi_raw_mode_info *raw,
+				 unsigned int idx)
+{
+	struct scmi_raw_buffer *rb = NULL;
+
+	if (!list_empty(&raw->msg_q[idx])) {
+		rb = list_first_entry(&raw->msg_q[idx],
+				      struct scmi_raw_buffer, node);
+		list_del_init(&rb->node);
+	}
+
+	return rb;
+}
+
+static struct scmi_raw_buffer *
+scmi_raw_buffer_dequeue(struct scmi_raw_mode_info *raw, unsigned int idx)
+{
+	unsigned long flags;
+	struct scmi_raw_buffer *rb;
+
+	spin_lock_irqsave(&raw->msg_q_lock[idx], flags);
+	rb = scmi_raw_buffer_dequeue_unlocked(raw, idx);
+	spin_unlock_irqrestore(&raw->msg_q_lock[idx], flags);
+
+	return rb;
+}
+
+static void scmi_raw_buffer_queue_flush(struct scmi_raw_mode_info *raw,
+					unsigned int idx)
+{
+	struct scmi_raw_buffer *rb;
+
+	do {
+		rb = scmi_raw_buffer_dequeue(raw, idx);
+		if (rb)
+			scmi_raw_buffer_put(raw, rb, idx);
+	} while (rb);
+}
+
+static struct scmi_xfer_raw_waiter *
+scmi_xfer_raw_waiter_get(struct scmi_raw_mode_info *raw, struct scmi_xfer *xfer,
+			 struct scmi_chan_info *cinfo, bool async)
+{
+	struct scmi_xfer_raw_waiter *rw = NULL;
+
+	mutex_lock(&raw->free_mtx);
+	if (!list_empty(&raw->free_waiters)) {
+		rw = list_first_entry(&raw->free_waiters,
+				      struct scmi_xfer_raw_waiter, node);
+		list_del_init(&rw->node);
+
+		if (async) {
+			reinit_completion(&rw->async_response);
+			xfer->async_done = &rw->async_response;
+		}
+
+		rw->cinfo = cinfo;
+		rw->xfer = xfer;
+	}
+	mutex_unlock(&raw->free_mtx);
+
+	return rw;
+}
+
+static void scmi_xfer_raw_waiter_put(struct scmi_raw_mode_info *raw,
+				     struct scmi_xfer_raw_waiter *rw)
+{
+	if (rw->xfer) {
+		rw->xfer->async_done = NULL;
+		rw->xfer = NULL;
+	}
+
+	mutex_lock(&raw->free_mtx);
+	list_add_tail(&rw->node, &raw->free_waiters);
+	mutex_unlock(&raw->free_mtx);
+}
+
+static void scmi_xfer_raw_waiter_enqueue(struct scmi_raw_mode_info *raw,
+					 struct scmi_xfer_raw_waiter *rw)
+{
+	/* A timestamp for the deferred worker to know how much this has aged */
+	rw->start_jiffies = jiffies;
+
+	trace_scmi_xfer_response_wait(rw->xfer->transfer_id, rw->xfer->hdr.id,
+				      rw->xfer->hdr.protocol_id,
+				      rw->xfer->hdr.seq,
+				      raw->desc->max_rx_timeout_ms,
+				      rw->xfer->hdr.poll_completion);
+
+	mutex_lock(&raw->active_mtx);
+	list_add_tail(&rw->node, &raw->active_waiters);
+	mutex_unlock(&raw->active_mtx);
+
+	/* kick waiter work */
+	queue_work(raw->wait_wq, &raw->waiters_work);
+}
+
+static struct scmi_xfer_raw_waiter *
+scmi_xfer_raw_waiter_dequeue(struct scmi_raw_mode_info *raw)
+{
+	struct scmi_xfer_raw_waiter *rw = NULL;
+
+	mutex_lock(&raw->active_mtx);
+	if (!list_empty(&raw->active_waiters)) {
+		rw = list_first_entry(&raw->active_waiters,
+				      struct scmi_xfer_raw_waiter, node);
+		list_del_init(&rw->node);
+	}
+	mutex_unlock(&raw->active_mtx);
+
+	return rw;
+}
+
+/**
+ * scmi_xfer_raw_worker  - Work function to wait for Raw xfers completions
+ *
+ * @work: A reference to the work.
+ *
+ * In SCMI Raw mode, once a user-provided injected SCMI message is sent, we
+ * cannot wait to receive its response (if any) in the context of the injection
+ * routines so as not to leave the userspace write syscall, which delivered the
+ * SCMI message to send, pending till eventually a reply is received.
+ * Userspace should and will poll/wait instead on the read syscalls which will
+ * be in charge of reading a received reply (if any).
+ *
+ * Even though reply messages are collected and reported into the SCMI Raw layer
+ * on the RX path, nonetheless we have to properly wait for their completion as
+ * usual (and async_completion too if needed) in order to properly release the
+ * xfer structure at the end: to do this out of the context of the write/send
+ * these waiting jobs are delegated to this deferred worker.
+ *
+ * Any sent xfer, to be waited for, is timestamped and queued for later
+ * consumption by this worker: queue aging is accounted for while choosing a
+ * timeout for the completion, BUT we do not really care here if we end up
+ * accidentally waiting for a bit too long.
+ */
+static void scmi_xfer_raw_worker(struct work_struct *work)
+{
+	struct scmi_raw_mode_info *raw;
+	struct device *dev;
+	unsigned long max_tmo;
+
+	raw = container_of(work, struct scmi_raw_mode_info, waiters_work);
+	dev = raw->handle->dev;
+	max_tmo = msecs_to_jiffies(raw->desc->max_rx_timeout_ms);
+
+	do {
+		int ret = 0;
+		unsigned int timeout_ms;
+		unsigned long aging;
+		struct scmi_xfer *xfer;
+		struct scmi_xfer_raw_waiter *rw;
+		struct scmi_chan_info *cinfo;
+
+		rw = scmi_xfer_raw_waiter_dequeue(raw);
+		if (!rw)
+			return;
+
+		cinfo = rw->cinfo;
+		xfer = rw->xfer;
+		/*
+		 * Waiters are queued by wait-deadline at the end, so some of
+		 * them could have been already expired when processed, BUT we
+		 * have to check the completion status anyway just in case a
+		 * virtually expired (aged) transaction was indeed completed
+		 * fine and we'll have to wait for the asynchronous part (if
+		 * any): for this reason a 1 ms timeout is used for already
+		 * expired/aged xfers.
+		 */
+		aging = jiffies - rw->start_jiffies;
+		timeout_ms = max_tmo > aging ?
+			jiffies_to_msecs(max_tmo - aging) : 1;
+
+		ret = scmi_xfer_raw_wait_for_message_response(cinfo, xfer,
+							      timeout_ms);
+		if (!ret && xfer->hdr.status)
+			ret = scmi_to_linux_errno(xfer->hdr.status);
+
+		if (raw->desc->ops->mark_txdone)
+			raw->desc->ops->mark_txdone(rw->cinfo, ret, xfer);
+
+		trace_scmi_xfer_end(xfer->transfer_id, xfer->hdr.id,
+				    xfer->hdr.protocol_id, xfer->hdr.seq, ret);
+
+		/* Wait also for an async delayed response if needed */
+		if (!ret && xfer->async_done) {
+			unsigned long tmo = msecs_to_jiffies(SCMI_MAX_RESPONSE_TIMEOUT);
+
+			if (!wait_for_completion_timeout(xfer->async_done, tmo))
+				dev_err(dev,
+					"timed out in RAW delayed resp - HDR:%08X\n",
+					pack_scmi_header(&xfer->hdr));
+		}
+
+		/* Release waiter and xfer */
+		scmi_xfer_raw_put(raw->handle, xfer);
+		scmi_xfer_raw_waiter_put(raw, rw);
+	} while (1);
+}
+
+static void scmi_xfer_raw_reset(struct scmi_raw_mode_info *raw)
+{
+	int i;
+
+	dev_info(raw->handle->dev, "Resetting SCMI Raw stack.\n");
+
+	for (i = 0; i < SCMI_RAW_MAX_QUEUE; i++)
+		scmi_raw_buffer_queue_flush(raw, i);
+}
+
+/**
+ * scmi_xfer_raw_get_init  - An helper to build a valid xfer from the provided
+ * bare SCMI message.
+ *
+ * @raw: A reference to the Raw instance.
+ * @buf: A buffer containing the whole SCMI message to send (including the
+ *	 header) in little-endian binary formmat.
+ * @len: Length of the message in @buf.
+ * @p: A pointer to return the initialized Raw xfer.
+ *
+ * After an xfer is picked from the TX pool and filled in with the message
+ * content, the xfer is registered as pending with the core in the usual way
+ * using the original sequence number provided by the user with the message.
+ *
+ * Note that, in case the testing user application is NOT using distinct
+ * sequence-numbers between successive SCMI messages such registration could
+ * fail temporarily if the previous message, using the same sequence number,
+ * had still not released; in such a case we just wait and retry.
+ *
+ * Return: 0 on Success
+ */
+static int scmi_xfer_raw_get_init(struct scmi_raw_mode_info *raw, void *buf,
+				  size_t len, struct scmi_xfer **p)
+{
+	u32 msg_hdr;
+	size_t tx_size;
+	struct scmi_xfer *xfer;
+	int ret, retry = SCMI_XFER_RAW_MAX_RETRIES;
+	struct device *dev = raw->handle->dev;
+
+	if (!buf || len < sizeof(u32))
+		return -EINVAL;
+
+	tx_size = len - sizeof(u32);
+	/* Ensure we have sane transfer sizes */
+	if (tx_size > raw->desc->max_msg_size)
+		return -ERANGE;
+
+	xfer = scmi_xfer_raw_get(raw->handle);
+	if (IS_ERR(xfer)) {
+		dev_warn(dev, "RAW - Cannot get a free RAW xfer !\n");
+		return PTR_ERR(xfer);
+	}
+
+	/* Build xfer from the provided SCMI bare LE message */
+	msg_hdr = le32_to_cpu(*((__le32 *)buf));
+	unpack_scmi_header(msg_hdr, &xfer->hdr);
+	xfer->hdr.seq = (u16)MSG_XTRACT_TOKEN(msg_hdr);
+	/* Polling not supported */
+	xfer->hdr.poll_completion = false;
+	xfer->hdr.status = SCMI_SUCCESS;
+	xfer->tx.len = tx_size;
+	xfer->rx.len = raw->desc->max_msg_size;
+	/* Clear the whole TX buffer */
+	memset(xfer->tx.buf, 0x00, raw->desc->max_msg_size);
+	if (xfer->tx.len)
+		memcpy(xfer->tx.buf, (u8 *)buf + sizeof(msg_hdr), xfer->tx.len);
+	*p = xfer;
+
+	/*
+	 * In flight registration can temporarily fail in case of Raw messages
+	 * if the user injects messages without using monotonically increasing
+	 * sequence numbers since, in Raw mode, the xfer (and the token) is
+	 * finally released later by a deferred worker. Just retry for a while.
+	 */
+	do {
+		ret = scmi_xfer_raw_inflight_register(raw->handle, xfer);
+		if (ret) {
+			dev_dbg(dev,
+				"...retrying[%d] inflight registration\n",
+				retry);
+			msleep(raw->desc->max_rx_timeout_ms /
+			       SCMI_XFER_RAW_MAX_RETRIES);
+		}
+	} while (ret && --retry);
+
+	if (ret) {
+		dev_warn(dev,
+			 "RAW - Could NOT register xfer %d in-flight HDR:0x%08X\n",
+			 xfer->hdr.seq, msg_hdr);
+		scmi_xfer_raw_put(raw->handle, xfer);
+	}
+
+	return ret;
+}
+
+/**
+ * scmi_do_xfer_raw_start  - An helper to send a valid raw xfer
+ *
+ * @raw: A reference to the Raw instance.
+ * @xfer: The xfer to send
+ * @async: A flag stating if an asynchronous command is required.
+ *
+ * This function send a previously built raw xfer using an appropriate channel
+ * and queues the related waiting work.
+ *
+ * Note that we need to know explicitly if the required command is meant to be
+ * asynchronous in kind since we have to properly setup the waiter.
+ * (and deducing this from the payload is weak and do not scale given there is
+ *  NOT a common header-flag stating if the command is asynchronous or not)
+ *
+ * Return: 0 on Success
+ */
+static int scmi_do_xfer_raw_start(struct scmi_raw_mode_info *raw,
+				  struct scmi_xfer *xfer, bool async)
+{
+	int ret;
+	struct scmi_chan_info *cinfo;
+	struct scmi_xfer_raw_waiter *rw;
+	struct device *dev = raw->handle->dev;
+
+	cinfo = scmi_xfer_raw_channel_get(raw->handle, xfer->hdr.protocol_id);
+	if (IS_ERR(cinfo))
+		return PTR_ERR(cinfo);
+
+	rw = scmi_xfer_raw_waiter_get(raw, xfer, cinfo, async);
+	if (!rw) {
+		dev_warn(dev, "RAW - Cannot get a free waiter !\n");
+		return -ENOMEM;
+	}
+
+	/* True ONLY if also supported by transport. */
+	if (is_polling_enabled(cinfo, raw->desc))
+		xfer->hdr.poll_completion = true;
+
+	reinit_completion(&xfer->done);
+	/* Make sure xfer state update is visible before sending */
+	smp_store_mb(xfer->state, SCMI_XFER_SENT_OK);
+
+	trace_scmi_xfer_begin(xfer->transfer_id, xfer->hdr.id,
+			      xfer->hdr.protocol_id, xfer->hdr.seq,
+			      xfer->hdr.poll_completion);
+
+	ret = raw->desc->ops->send_message(rw->cinfo, xfer);
+	if (ret) {
+		dev_err(dev, "Failed to send RAW message %d\n", ret);
+		scmi_xfer_raw_waiter_put(raw, rw);
+		return ret;
+	}
+
+	trace_scmi_msg_dump(xfer->hdr.protocol_id, xfer->hdr.id, "cmnd",
+			    xfer->hdr.seq, xfer->hdr.status,
+			    xfer->tx.buf, xfer->tx.len);
+
+	scmi_xfer_raw_waiter_enqueue(raw, rw);
+
+	return ret;
+}
+
+/**
+ * scmi_raw_message_send  - An helper to build and send an SCMI command using
+ * the provided SCMI bare message buffer
+ *
+ * @raw: A reference to the Raw instance.
+ * @buf: A buffer containing the whole SCMI message to send (including the
+ *	 header) in little-endian binary format.
+ * @len: Length of the message in @buf.
+ * @async: A flag stating if an asynchronous command is required.
+ *
+ * Return: 0 on Success
+ */
+static int scmi_raw_message_send(struct scmi_raw_mode_info *raw,
+				 void *buf, size_t len, bool async)
+{
+	int ret;
+	struct scmi_xfer *xfer;
+
+	ret = scmi_xfer_raw_get_init(raw, buf, len, &xfer);
+	if (ret)
+		return ret;
+
+	ret = scmi_do_xfer_raw_start(raw, xfer, async);
+	if (ret)
+		scmi_xfer_raw_put(raw->handle, xfer);
+
+	return ret;
+}
+
+static struct scmi_raw_buffer *
+scmi_raw_message_dequeue(struct scmi_raw_mode_info *raw, unsigned int idx,
+			 bool o_nonblock)
+{
+	unsigned long flags;
+	struct scmi_raw_buffer *rb;
+
+	spin_lock_irqsave(&raw->msg_q_lock[idx], flags);
+	while (list_empty(&raw->msg_q[idx])) {
+		spin_unlock_irqrestore(&raw->msg_q_lock[idx], flags);
+
+		if (o_nonblock)
+			return ERR_PTR(-EAGAIN);
+
+		if (wait_event_interruptible(raw->wqs[idx],
+					     !list_empty(&raw->msg_q[idx])))
+			return ERR_PTR(-ERESTARTSYS);
+
+		spin_lock_irqsave(&raw->msg_q_lock[idx], flags);
+	}
+
+	rb = scmi_raw_buffer_dequeue_unlocked(raw, idx);
+
+	spin_unlock_irqrestore(&raw->msg_q_lock[idx], flags);
+
+	return rb;
+}
+
+/**
+ * scmi_raw_message_receive  - An helper to dequeue and report the next
+ * available enqueued raw message payload that has been collected.
+ *
+ * @raw: A reference to the Raw instance.
+ * @buf: A buffer to get hold of the whole SCMI message received and represented
+ *	 in little-endian binary format.
+ * @len: Length of @buf.
+ * @size: The effective size of the message copied into @buf
+ * @idx: The index of the queue to pick the next queued message from.
+ * @o_nonblock: A flag to request a non-blocking message dequeue.
+ *
+ * Return: 0 on Success
+ */
+static int scmi_raw_message_receive(struct scmi_raw_mode_info *raw,
+				    void *buf, size_t len, size_t *size,
+				    unsigned int idx,
+				    bool o_nonblock)
+{
+	int ret = 0;
+	struct scmi_raw_buffer *rb;
+
+	rb = scmi_raw_message_dequeue(raw, idx, o_nonblock);
+	if (IS_ERR(rb)) {
+		dev_dbg(raw->handle->dev, "RAW - No message available!\n");
+		return PTR_ERR(rb);
+	}
+
+	if (rb->msg.len <= len) {
+		memcpy(buf, rb->msg.buf, rb->msg.len);
+		*size = rb->msg.len;
+	} else {
+		ret = -ENOSPC;
+	}
+
+	scmi_raw_buffer_put(raw, rb, idx);
+
+	return ret;
+}
+
+/* SCMI Raw debugfs helpers */
+
+static ssize_t scmi_dbg_raw_mode_common_read(struct file *filp,
+					     char __user *buf,
+					     size_t count, loff_t *ppos,
+					     unsigned int idx)
+{
+	ssize_t cnt;
+	struct scmi_dbg_raw_data *rd = filp->private_data;
+
+	if (!rd->rx_size) {
+		int ret;
+
+		ret = scmi_raw_message_receive(rd->raw, rd->rx.buf, rd->rx.len,
+					       &rd->rx_size, idx,
+					       filp->f_flags & O_NONBLOCK);
+		if (ret) {
+			rd->rx_size = 0;
+			return ret;
+		}
+
+		/* Reset any previous filepos change, including writes */
+		*ppos = 0;
+	} else if (*ppos == rd->rx_size) {
+		/* Return EOF once all the message has been read-out */
+		rd->rx_size = 0;
+		return 0;
+	}
+
+	cnt = simple_read_from_buffer(buf, count, ppos,
+				      rd->rx.buf, rd->rx_size);
+
+	return cnt;
+}
+
+static ssize_t scmi_dbg_raw_mode_common_write(struct file *filp,
+					      const char __user *buf,
+					      size_t count, loff_t *ppos,
+					      bool async)
+{
+	int ret;
+	struct scmi_dbg_raw_data *rd = filp->private_data;
+
+	if (count > rd->tx.len - rd->tx_size)
+		return -ENOSPC;
+
+	/* On first write attempt @count carries the total full message size. */
+	if (!rd->tx_size)
+		rd->tx_req_size = count;
+
+	/*
+	 * Gather a full message, possibly across multiple interrupted wrrtes,
+	 * before sending it with a single RAW xfer.
+	 */
+	if (rd->tx_size < rd->tx_req_size) {
+		size_t cnt;
+
+		cnt = simple_write_to_buffer(rd->tx.buf, rd->tx.len, ppos,
+					     buf, count);
+		rd->tx_size += cnt;
+		if (cnt < count)
+			return cnt;
+	}
+
+	ret = scmi_raw_message_send(rd->raw, rd->tx.buf, rd->tx_size, async);
+
+	/* Reset ppos for next message ... */
+	rd->tx_size = 0;
+
+	return ret ?: count;
+}
+
+static __poll_t scmi_test_dbg_raw_common_poll(struct file *filp,
+					      struct poll_table_struct *wait,
+					      unsigned int idx)
+{
+	unsigned long flags;
+	struct scmi_dbg_raw_data *rd = filp->private_data;
+	__poll_t mask = 0;
+
+	poll_wait(filp, &rd->raw->wqs[idx], wait);
+
+	spin_lock_irqsave(&rd->raw->msg_q_lock[idx], flags);
+	if (!list_empty(&rd->raw->msg_q[idx]))
+		mask = EPOLLIN | EPOLLRDNORM;
+	spin_unlock_irqrestore(&rd->raw->msg_q_lock[idx], flags);
+
+	return mask;
+}
+
+static ssize_t scmi_dbg_raw_mode_message_read(struct file *filp,
+					      char __user *buf,
+					      size_t count, loff_t *ppos)
+{
+	return scmi_dbg_raw_mode_common_read(filp, buf, count, ppos,
+					     SCMI_RAW_REPLY_QUEUE);
+}
+
+static ssize_t scmi_dbg_raw_mode_message_write(struct file *filp,
+					       const char __user *buf,
+					       size_t count, loff_t *ppos)
+{
+	return scmi_dbg_raw_mode_common_write(filp, buf, count, ppos, false);
+}
+
+static __poll_t scmi_dbg_raw_mode_message_poll(struct file *filp,
+					       struct poll_table_struct *wait)
+{
+	return scmi_test_dbg_raw_common_poll(filp, wait, SCMI_RAW_REPLY_QUEUE);
+}
+
+static int scmi_dbg_raw_mode_open(struct inode *inode, struct file *filp)
+{
+	struct scmi_raw_mode_info *raw;
+	struct scmi_dbg_raw_data *rd;
+
+	if (!inode->i_private)
+		return -ENODEV;
+
+	raw = inode->i_private;
+	rd = kzalloc(sizeof(*rd), GFP_KERNEL);
+	if (!rd)
+		return -ENOMEM;
+
+	rd->rx.len = raw->desc->max_msg_size + sizeof(u32);
+	rd->rx.buf = kzalloc(rd->rx.len, GFP_KERNEL);
+	if (!rd->rx.buf) {
+		kfree(rd);
+		return -ENOMEM;
+	}
+
+	rd->tx.len = raw->desc->max_msg_size + sizeof(u32);
+	rd->tx.buf = kzalloc(rd->tx.len, GFP_KERNEL);
+	if (!rd->tx.buf) {
+		kfree(rd->rx.buf);
+		kfree(rd);
+		return -ENOMEM;
+	}
+
+	rd->raw = raw;
+	filp->private_data = rd;
+
+	return 0;
+}
+
+static int scmi_dbg_raw_mode_release(struct inode *inode, struct file *filp)
+{
+	struct scmi_dbg_raw_data *rd = filp->private_data;
+
+	kfree(rd->rx.buf);
+	kfree(rd->tx.buf);
+	kfree(rd);
+
+	return 0;
+}
+
+static ssize_t scmi_dbg_raw_mode_reset_write(struct file *filp,
+					     const char __user *buf,
+					     size_t count, loff_t *ppos)
+{
+	struct scmi_dbg_raw_data *rd = filp->private_data;
+
+	scmi_xfer_raw_reset(rd->raw);
+
+	return count;
+}
+
+static const struct file_operations scmi_dbg_raw_mode_reset_fops = {
+	.open = scmi_dbg_raw_mode_open,
+	.release = scmi_dbg_raw_mode_release,
+	.write = scmi_dbg_raw_mode_reset_write,
+	.owner = THIS_MODULE,
+};
+
+static const struct file_operations scmi_dbg_raw_mode_message_fops = {
+	.open = scmi_dbg_raw_mode_open,
+	.release = scmi_dbg_raw_mode_release,
+	.read = scmi_dbg_raw_mode_message_read,
+	.write = scmi_dbg_raw_mode_message_write,
+	.poll = scmi_dbg_raw_mode_message_poll,
+	.owner = THIS_MODULE,
+};
+
+static ssize_t scmi_dbg_raw_mode_message_async_write(struct file *filp,
+						     const char __user *buf,
+						     size_t count, loff_t *ppos)
+{
+	return scmi_dbg_raw_mode_common_write(filp, buf, count, ppos, true);
+}
+
+static const struct file_operations scmi_dbg_raw_mode_message_async_fops = {
+	.open = scmi_dbg_raw_mode_open,
+	.release = scmi_dbg_raw_mode_release,
+	.read = scmi_dbg_raw_mode_message_read,
+	.write = scmi_dbg_raw_mode_message_async_write,
+	.poll = scmi_dbg_raw_mode_message_poll,
+	.owner = THIS_MODULE,
+};
+
+static ssize_t scmi_test_dbg_raw_mode_notif_read(struct file *filp,
+						 char __user *buf,
+						 size_t count, loff_t *ppos)
+{
+	return scmi_dbg_raw_mode_common_read(filp, buf, count, ppos,
+					     SCMI_RAW_NOTIF_QUEUE);
+}
+
+static __poll_t scmi_test_dbg_raw_mode_notif_poll(struct file *filp,
+						  struct poll_table_struct *wait)
+{
+	return scmi_test_dbg_raw_common_poll(filp, wait, SCMI_RAW_NOTIF_QUEUE);
+}
+
+static const struct file_operations scmi_dbg_raw_mode_notification_fops = {
+	.open = scmi_dbg_raw_mode_open,
+	.release = scmi_dbg_raw_mode_release,
+	.read = scmi_test_dbg_raw_mode_notif_read,
+	.poll = scmi_test_dbg_raw_mode_notif_poll,
+	.owner = THIS_MODULE,
+};
+
+static ssize_t scmi_test_dbg_raw_mode_errors_read(struct file *filp,
+						  char __user *buf,
+						  size_t count, loff_t *ppos)
+{
+	return scmi_dbg_raw_mode_common_read(filp, buf, count, ppos,
+					     SCMI_RAW_ERRS_QUEUE);
+}
+
+static __poll_t scmi_test_dbg_raw_mode_errors_poll(struct file *filp,
+						   struct poll_table_struct *wait)
+{
+	return scmi_test_dbg_raw_common_poll(filp, wait, SCMI_RAW_ERRS_QUEUE);
+}
+
+static const struct file_operations scmi_dbg_raw_mode_errors_fops = {
+	.open = scmi_dbg_raw_mode_open,
+	.release = scmi_dbg_raw_mode_release,
+	.read = scmi_test_dbg_raw_mode_errors_read,
+	.poll = scmi_test_dbg_raw_mode_errors_poll,
+	.owner = THIS_MODULE,
+};
+
+static int scmi_xfer_raw_free_bufs_init(struct scmi_raw_mode_info *raw, int idx)
+{
+	int i;
+	struct scmi_raw_buffer *rb;
+	struct device *dev = raw->handle->dev;
+
+	rb = devm_kcalloc(dev, raw->tx_max_msg, sizeof(*rb), GFP_KERNEL);
+	if (!rb)
+		return -ENOMEM;
+
+	spin_lock_init(&raw->free_bufs_lock[idx]);
+	INIT_LIST_HEAD(&raw->free_raw_bufs[idx]);
+	for (i = 0; i < raw->tx_max_msg; i++, rb++) {
+		rb->max_len = raw->desc->max_msg_size + sizeof(u32);
+		rb->msg.buf = devm_kzalloc(dev, rb->max_len, GFP_KERNEL);
+		if (!rb->msg.buf)
+			return -ENOMEM;
+		scmi_raw_buffer_put(raw, rb, idx);
+	}
+
+	spin_lock_init(&raw->msg_q_lock[idx]);
+	INIT_LIST_HEAD(&raw->msg_q[idx]);
+	init_waitqueue_head(&raw->wqs[idx]);
+
+	return 0;
+}
+
+static int scmi_xfer_raw_worker_init(struct scmi_raw_mode_info *raw)
+{
+	int i;
+	struct scmi_xfer_raw_waiter *rw;
+	struct device *dev = raw->handle->dev;
+
+	rw = devm_kcalloc(dev, raw->tx_max_msg, sizeof(*rw), GFP_KERNEL);
+	if (!rw)
+		return -ENOMEM;
+
+	raw->wait_wq = alloc_workqueue("scmi-raw-wait-wq-%d",
+				       WQ_UNBOUND | WQ_FREEZABLE |
+				       WQ_HIGHPRI, WQ_SYSFS, raw->id);
+	if (!raw->wait_wq)
+		return -ENOMEM;
+
+	mutex_init(&raw->free_mtx);
+	INIT_LIST_HEAD(&raw->free_waiters);
+	mutex_init(&raw->active_mtx);
+	INIT_LIST_HEAD(&raw->active_waiters);
+
+	for (i = 0; i < raw->tx_max_msg; i++, rw++) {
+		init_completion(&rw->async_response);
+		scmi_xfer_raw_waiter_put(raw, rw);
+	}
+	INIT_WORK(&raw->waiters_work, scmi_xfer_raw_worker);
+
+	return 0;
+}
+
+static int scmi_raw_mode_setup(struct scmi_raw_mode_info *raw)
+{
+	int ret, idx;
+	void *gid;
+	struct device *dev = raw->handle->dev;
+
+	gid = devres_open_group(dev, NULL, GFP_KERNEL);
+	if (!gid)
+		return -ENOMEM;
+
+	for (idx = 0; idx < SCMI_RAW_MAX_QUEUE; idx++) {
+		ret = scmi_xfer_raw_free_bufs_init(raw, idx);
+		if (ret)
+			goto err;
+	}
+
+	ret = scmi_xfer_raw_worker_init(raw);
+	if (ret)
+		goto err;
+
+	devres_close_group(dev, gid);
+	raw->gid = gid;
+
+	return 0;
+
+err:
+	devres_release_group(dev, gid);
+	return ret;
+}
+
+/**
+ * scmi_raw_mode_init  - Function to initialize the SCMI Raw stack
+ *
+ * @handle: Pointer to SCMI entity handle
+ * @raw_top_dentry: A reference to the top Raw debugfs dentry
+ * @desc: Reference to the transport operations
+ * @tx_max_msg: Max number of in-flight messages allowed by the transport
+ *
+ * This function prepare the SCMI Raw stack and creates the debugfs API.
+ *
+ * Return: An opaque handle to the Raw instance on Success, an ERR_PTR otherwise
+ */
+void *scmi_raw_mode_init(const struct scmi_handle *handle,
+			 struct dentry *raw_top_dentry,
+			 const struct scmi_desc *desc, int tx_max_msg)
+{
+	int ret;
+	char instance_dir[8];
+	struct scmi_raw_mode_info *raw;
+	struct device *dev;
+
+	if (!handle || !desc || !raw_top_dentry)
+		return ERR_PTR(-EINVAL);
+
+	dev = handle->dev;
+	raw = devm_kzalloc(dev, sizeof(*raw), GFP_KERNEL);
+	if (!raw)
+		return ERR_PTR(-ENOMEM);
+
+	raw->handle = handle;
+	raw->desc = desc;
+	raw->tx_max_msg = tx_max_msg;
+	raw->id = atomic_inc_return(&scmi_raw_instance_counter);
+	raw->name = kstrdup(of_node_full_name(dev->of_node), GFP_KERNEL);
+	if (!raw->name) {
+		devm_kfree(dev, raw);
+		return ERR_PTR(-ENOMEM);
+	}
+
+	ret = scmi_raw_mode_setup(raw);
+	if (ret) {
+		kfree(raw->name);
+		devm_kfree(dev, raw);
+		atomic_dec(&scmi_raw_instance_counter);
+		return ERR_PTR(ret);
+	}
+
+	snprintf(instance_dir, 8, "%u", raw->id);
+	raw->dentry = debugfs_create_dir(instance_dir, raw_top_dentry);
+	if (IS_ERR(raw->dentry)) {
+		ret = PTR_ERR(raw->dentry);
+		devres_release_group(dev, raw->gid);
+		kfree(raw->name);
+		devm_kfree(dev, raw);
+		atomic_dec(&scmi_raw_instance_counter);
+		return ERR_PTR(ret);
+	}
+
+	debugfs_create_str("instance_name", 0400, raw->dentry,
+			   (char **)&raw->name);
+
+	debugfs_create_file("reset", 0200, raw->dentry, raw,
+			    &scmi_dbg_raw_mode_reset_fops);
+
+	debugfs_create_u32("transport_rx_timeout_ms", 0400, raw->dentry,
+			   (u32 *)&raw->desc->max_rx_timeout_ms);
+
+	debugfs_create_u32("transport_max_msg_size", 0400, raw->dentry,
+			   (u32 *)&raw->desc->max_msg_size);
+
+	debugfs_create_u32("transport_tx_max_msg", 0400, raw->dentry,
+			   (u32 *)&raw->tx_max_msg);
+
+	debugfs_create_file("message", 0600, raw->dentry, raw,
+			    &scmi_dbg_raw_mode_message_fops);
+
+	debugfs_create_file("message_async", 0600, raw->dentry, raw,
+			    &scmi_dbg_raw_mode_message_async_fops);
+
+	debugfs_create_file("notification", 0400, raw->dentry, raw,
+			    &scmi_dbg_raw_mode_notification_fops);
+
+	debugfs_create_file("errors", 0400, raw->dentry, raw,
+			    &scmi_dbg_raw_mode_errors_fops);
+
+	dev_info(dev, "SCMI RAW Mode initialized for instance %d - (%s)\n",
+		 raw->id, raw->name);
+
+	return raw;
+}
+
+/**
+ * scmi_raw_mode_cleanup  - Function to cleanup the SCMI Raw stack
+ *
+ * @r: An opaque handle to an initialized SCMI Raw instance
+ */
+void scmi_raw_mode_cleanup(void *r)
+{
+	struct scmi_raw_mode_info *raw = r;
+
+	if (!raw)
+		return;
+
+	debugfs_remove_recursive(raw->dentry);
+
+	cancel_work_sync(&raw->waiters_work);
+	destroy_workqueue(raw->wait_wq);
+
+	kfree(raw->name);
+}
+
+static int scmi_xfer_raw_collect(void *msg, size_t *msg_len,
+				 struct scmi_xfer *xfer)
+{
+	__le32 *m;
+	size_t msg_size;
+
+	if (!xfer || !msg || !msg_len)
+		return -EINVAL;
+
+	/* Account for hdr ...*/
+	msg_size = xfer->rx.len + sizeof(u32);
+	/* ... and status if needed */
+	if (xfer->hdr.type != MSG_TYPE_NOTIFICATION)
+		msg_size += sizeof(u32);
+
+	if (msg_size > *msg_len)
+		return -ENOSPC;
+
+	m = msg;
+	*m = cpu_to_le32(pack_scmi_header(&xfer->hdr));
+	if (xfer->hdr.type != MSG_TYPE_NOTIFICATION)
+		*++m = cpu_to_le32(xfer->hdr.status);
+
+	memcpy(++m, xfer->rx.buf, xfer->rx.len);
+
+	*msg_len = msg_size;
+
+	return 0;
+}
+
+/**
+ * scmi_raw_message_report  - Helper to report back valid reponses/notifications
+ * to raw message requests.
+ *
+ * @r: An opaque reference to the raw instance configuration
+ * @xfer: The xfer containing the message to be reported
+ * @idx: The index of the queue.
+ *
+ * If Raw mode is enabled, this is called from the SCMI core on the regular RX
+ * path to save and enqueue the response/notification payload carried by this
+ * xfer into a dedicated scmi_raw_buffer for later consumption by the user.
+ *
+ * This way the caller can free the related xfer immediately afterwards and the
+ * user can read back the raw message payload at its own pace (if ever) without
+ * holding an xfer for too long.
+ */
+void scmi_raw_message_report(void *r, struct scmi_xfer *xfer, unsigned int idx)
+{
+	int ret;
+	unsigned long flags;
+	struct scmi_raw_buffer *rb;
+	struct device *dev;
+	struct scmi_raw_mode_info *raw = r;
+
+	if (!raw || (idx == SCMI_RAW_REPLY_QUEUE && !xfer->is_raw))
+		return;
+
+	dev = raw->handle->dev;
+
+	/*
+	 * Grab the msg_q_lock upfront to avoid a possible race between
+	 * realizing the free list was empty and effectively picking the next
+	 * buffer to use from the oldest one enqueued and still unread on this
+	 * msg_q.
+	 *
+	 * Note that nowhere else these locks are taken together, so no risk of
+	 * deadlocks du eto inversion.
+	 */
+	spin_lock_irqsave(&raw->msg_q_lock[idx], flags);
+	rb = scmi_raw_buffer_get(raw, idx);
+	if (!rb) {
+		/*
+		 * Immediate and delayed replies to previously injected Raw
+		 * commands MUST be read back from userspace to free the buffers:
+		 * if this is not happening something is seriously broken and
+		 * must be fixed at the application level: complain loudly.
+		 */
+		if (idx == SCMI_RAW_REPLY_QUEUE) {
+			spin_unlock_irqrestore(&raw->msg_q_lock[idx], flags);
+			dev_warn(dev,
+				 "RAW[%d] - Buffers exhausted. Dropping report.\n",
+				 idx);
+			return;
+		}
+
+		/*
+		 * Notifications and errors queues are instead handled in a
+		 * circular manner: unread old buffers are just overwritten by
+		 * newer ones.
+		 *
+		 * The main reason for this is that notifications originated
+		 * by Raw requests cannot be distinguished from normal ones, so
+		 * your Raw buffers queues risk to be flooded and depleted by
+		 * notifications if you left it mistakenly enabled or when in
+		 * coexistence mode.
+		 */
+		rb = scmi_raw_buffer_dequeue_unlocked(raw, idx);
+		if (WARN_ON(!rb)) {
+			spin_unlock_irqrestore(&raw->msg_q_lock[idx], flags);
+			return;
+		}
+
+		/* Reset to full buffer length */
+		rb->msg.len = rb->max_len;
+
+		dev_warn_once(dev,
+			      "RAW[%d] - Buffers exhausted. Re-using oldest.\n",
+			      idx);
+	}
+	spin_unlock_irqrestore(&raw->msg_q_lock[idx], flags);
+
+	ret = scmi_xfer_raw_collect(rb->msg.buf, &rb->msg.len, xfer);
+	if (ret) {
+		dev_warn(dev, "RAW - Cannot collect xfer into buffer !\n");
+		scmi_raw_buffer_put(raw, rb, idx);
+		return;
+	}
+
+	scmi_raw_buffer_enqueue(raw, rb, idx);
+}
+
+static void scmi_xfer_raw_fill(struct scmi_raw_mode_info *raw,
+			       struct scmi_chan_info *cinfo,
+			       struct scmi_xfer *xfer, u32 msg_hdr)
+{
+	/* Unpack received HDR as it is */
+	unpack_scmi_header(msg_hdr, &xfer->hdr);
+	xfer->hdr.seq = MSG_XTRACT_TOKEN(msg_hdr);
+
+	memset(xfer->rx.buf, 0x00, xfer->rx.len);
+
+	raw->desc->ops->fetch_response(cinfo, xfer);
+}
+
+/**
+ * scmi_raw_error_report  - Helper to report back timed-out or generally
+ * unexpected replies.
+ *
+ * @r: An opaque reference to the raw instance configuration
+ * @cinfo: A reference to the channel to use to retrieve the broken xfer
+ * @msg_hdr: The SCMI message header of the message to fetch and report
+ * @priv: Any private data related to the xfer.
+ *
+ * If Raw mode is enabled, this is called from the SCMI core on the RX path in
+ * case of errors to save and enqueue the bad message payload carried by the
+ * message that has just been received.
+ *
+ * Note that we have to manually fetch any available payload into a temporary
+ * xfer to be able to save and enqueue the message, since the regular RX error
+ * path which had called this would have not fetched the message payload having
+ * classified it as an error.
+ */
+void scmi_raw_error_report(void *r, struct scmi_chan_info *cinfo,
+			   u32 msg_hdr, void *priv)
+{
+	struct scmi_xfer xfer;
+	struct scmi_raw_mode_info *raw = r;
+
+	if (!raw)
+		return;
+
+	xfer.rx.len = raw->desc->max_msg_size;
+	xfer.rx.buf = kzalloc(xfer.rx.len, GFP_ATOMIC);
+	if (!xfer.rx.buf) {
+		dev_info(raw->handle->dev,
+			 "Cannot report Raw error for HDR:0x%X - ENOMEM\n",
+			 msg_hdr);
+		return;
+	}
+
+	/* Any transport-provided priv must be passed back down to transport */
+	if (priv)
+		/* Ensure priv is visible */
+		smp_store_mb(xfer.priv, priv);
+
+	scmi_xfer_raw_fill(raw, cinfo, &xfer, msg_hdr);
+	scmi_raw_message_report(raw, &xfer, SCMI_RAW_ERRS_QUEUE);
+
+	kfree(xfer.rx.buf);
+}
diff --git a/drivers/firmware/arm_scmi/raw_mode.h b/drivers/firmware/arm_scmi/raw_mode.h
new file mode 100644
index 000000000000..ca5018c35c2e
--- /dev/null
+++ b/drivers/firmware/arm_scmi/raw_mode.h
@@ -0,0 +1,30 @@ 
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * System Control and Management Interface (SCMI) Message Protocol
+ * Raw mode support header.
+ *
+ * Copyright (C) 2022 ARM Ltd.
+ */
+#ifndef _SCMI_RAW_MODE_H
+#define _SCMI_RAW_MODE_H
+
+#include "common.h"
+
+enum {
+	SCMI_RAW_REPLY_QUEUE,
+	SCMI_RAW_NOTIF_QUEUE,
+	SCMI_RAW_ERRS_QUEUE,
+	SCMI_RAW_MAX_QUEUE
+};
+
+void *scmi_raw_mode_init(const struct scmi_handle *handle,
+			 struct dentry *raw_top_dentry,
+			 const struct scmi_desc *desc, int tx_max_msg);
+void scmi_raw_mode_cleanup(void *raw);
+
+void scmi_raw_message_report(void *raw, struct scmi_xfer *xfer,
+			     unsigned int idx);
+void scmi_raw_error_report(void *raw, struct scmi_chan_info *cinfo,
+			   u32 msg_hdr, void *priv);
+
+#endif /* _SCMI_RAW_MODE_H */