drm/msm: Add A6XX device support

Add support for the A6XX family of Adreno GPUs. The biggest addition
is the GMU (Graphics Management Unit) which takes over most of the
power management of the GPU itself but in a ironic twist of fate
needs a goodly amount of management itself. Add support for the
A6XX core code, the GMU and the HFI (hardware firmware interface)
queue that the CPU uses to communicate with the GMU.

Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>

1 files changed, 435 insertions, 0 deletions

diff --git a/drivers/gpu/drm/msm/adreno/a6xx_hfi.c b/drivers/gpu/drm/msm/adreno/a6xx_hfi.c
new file mode 100644
index 000000000000..f19ef4cb6ea4
--- /dev/null
+++ b/drivers/gpu/drm/msm/adreno/a6xx_hfi.c
@@ -0,0 +1,435 @@
+// SPDX-License-Identifier: GPL-2.0
+/* Copyright (c) 2017-2018 The Linux Foundation. All rights reserved. */
+
+#include <linux/completion.h>
+#include <linux/circ_buf.h>
+#include <linux/list.h>
+
+#include "a6xx_gmu.h"
+#include "a6xx_gmu.xml.h"
+
+#define HFI_MSG_ID(val) [val] = #val
+
+static const char * const a6xx_hfi_msg_id[] = {
+	HFI_MSG_ID(HFI_H2F_MSG_INIT),
+	HFI_MSG_ID(HFI_H2F_MSG_FW_VERSION),
+	HFI_MSG_ID(HFI_H2F_MSG_BW_TABLE),
+	HFI_MSG_ID(HFI_H2F_MSG_PERF_TABLE),
+	HFI_MSG_ID(HFI_H2F_MSG_TEST),
+};
+
+static int a6xx_hfi_queue_read(struct a6xx_hfi_queue *queue, u32 *data,
+		u32 dwords)
+{
+	struct a6xx_hfi_queue_header *header = queue->header;
+	u32 i, hdr, index = header->read_index;
+
+	if (header->read_index == header->write_index) {
+		header->rx_request = 1;
+		return 0;
+	}
+
+	hdr = queue->data[index];
+
+	/*
+	 * If we are to assume that the GMU firmware is in fact a rational actor
+	 * and is programmed to not send us a larger response than we expect
+	 * then we can also assume that if the header size is unexpectedly large
+	 * that it is due to memory corruption and/or hardware failure. In this
+	 * case the only reasonable course of action is to BUG() to help harden
+	 * the failure.
+	 */
+
+	BUG_ON(HFI_HEADER_SIZE(hdr) > dwords);
+
+	for (i = 0; i < HFI_HEADER_SIZE(hdr); i++) {
+		data[i] = queue->data[index];
+		index = (index + 1) % header->size;
+	}
+
+	header->read_index = index;
+	return HFI_HEADER_SIZE(hdr);
+}
+
+static int a6xx_hfi_queue_write(struct a6xx_gmu *gmu,
+	struct a6xx_hfi_queue *queue, u32 *data, u32 dwords)
+{
+	struct a6xx_hfi_queue_header *header = queue->header;
+	u32 i, space, index = header->write_index;
+
+	spin_lock(&queue->lock);
+
+	space = CIRC_SPACE(header->write_index, header->read_index,
+		header->size);
+	if (space < dwords) {
+		header->dropped++;
+		spin_unlock(&queue->lock);
+		return -ENOSPC;
+	}
+
+	for (i = 0; i < dwords; i++) {
+		queue->data[index] = data[i];
+		index = (index + 1) % header->size;
+	}
+
+	header->write_index = index;
+	spin_unlock(&queue->lock);
+
+	gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET, 0x01);
+	return 0;
+}
+
+struct a6xx_hfi_response {
+	u32 id;
+	u32 seqnum;
+	struct list_head node;
+	struct completion complete;
+
+	u32 error;
+	u32 payload[16];
+};
+
+/*
+ * Incoming HFI ack messages can come in out of order so we need to store all
+ * the pending messages on a list until they are handled.
+ */
+static spinlock_t hfi_ack_lock = __SPIN_LOCK_UNLOCKED(message_lock);
+static LIST_HEAD(hfi_ack_list);
+
+static void a6xx_hfi_handle_ack(struct a6xx_gmu *gmu,
+		struct a6xx_hfi_msg_response *msg)
+{
+	struct a6xx_hfi_response *resp;
+	u32 id, seqnum;
+
+	/* msg->ret_header contains the header of the message being acked */
+	id = HFI_HEADER_ID(msg->ret_header);
+	seqnum = HFI_HEADER_SEQNUM(msg->ret_header);
+
+	spin_lock(&hfi_ack_lock);
+	list_for_each_entry(resp, &hfi_ack_list, node) {
+		if (resp->id == id && resp->seqnum == seqnum) {
+			resp->error = msg->error;
+			memcpy(resp->payload, msg->payload,
+				sizeof(resp->payload));
+
+			complete(&resp->complete);
+			spin_unlock(&hfi_ack_lock);
+			return;
+		}
+	}
+	spin_unlock(&hfi_ack_lock);
+
+	dev_err(gmu->dev, "Nobody was waiting for HFI message %d\n", seqnum);
+}
+
+static void a6xx_hfi_handle_error(struct a6xx_gmu *gmu,
+		struct a6xx_hfi_msg_response *msg)
+{
+	struct a6xx_hfi_msg_error *error = (struct a6xx_hfi_msg_error *) msg;
+
+	dev_err(gmu->dev, "GMU firmware error %d\n", error->code);
+}
+
+void a6xx_hfi_task(unsigned long data)
+{
+	struct a6xx_gmu *gmu = (struct a6xx_gmu *) data;
+	struct a6xx_hfi_queue *queue = &gmu->queues[HFI_RESPONSE_QUEUE];
+	struct a6xx_hfi_msg_response resp;
+
+	for (;;) {
+		u32 id;
+		int ret = a6xx_hfi_queue_read(queue, (u32 *) &resp,
+			sizeof(resp) >> 2);
+
+		/* Returns the number of bytes copied or negative on error */
+		if (ret <= 0) {
+			if (ret < 0)
+				dev_err(gmu->dev,
+					"Unable to read the HFI message queue\n");
+			break;
+		}
+
+		id = HFI_HEADER_ID(resp.header);
+
+		if (id == HFI_F2H_MSG_ACK)
+			a6xx_hfi_handle_ack(gmu, &resp);
+		else if (id == HFI_F2H_MSG_ERROR)
+			a6xx_hfi_handle_error(gmu, &resp);
+	}
+}
+
+static int a6xx_hfi_send_msg(struct a6xx_gmu *gmu, int id,
+		void *data, u32 size, u32 *payload, u32 payload_size)
+{
+	struct a6xx_hfi_queue *queue = &gmu->queues[HFI_COMMAND_QUEUE];
+	struct a6xx_hfi_response resp = { 0 };
+	int ret, dwords = size >> 2;
+	u32 seqnum;
+
+	seqnum = atomic_inc_return(&queue->seqnum) % 0xfff;
+
+	/* First dword of the message is the message header - fill it in */
+	*((u32 *) data) = (seqnum << 20) | (HFI_MSG_CMD << 16) |
+		(dwords << 8) | id;
+
+	init_completion(&resp.complete);
+	resp.id = id;
+	resp.seqnum = seqnum;
+
+	spin_lock_bh(&hfi_ack_lock);
+	list_add_tail(&resp.node, &hfi_ack_list);
+	spin_unlock_bh(&hfi_ack_lock);
+
+	ret = a6xx_hfi_queue_write(gmu, queue, data, dwords);
+	if (ret) {
+		dev_err(gmu->dev, "Unable to send message %s id %d\n",
+			a6xx_hfi_msg_id[id], seqnum);
+		goto out;
+	}
+
+	/* Wait up to 5 seconds for the response */
+	ret = wait_for_completion_timeout(&resp.complete,
+		msecs_to_jiffies(5000));
+	if (!ret) {
+		dev_err(gmu->dev,
+			"Message %s id %d timed out waiting for response\n",
+			a6xx_hfi_msg_id[id], seqnum);
+		ret = -ETIMEDOUT;
+	} else
+		ret = 0;
+
+out:
+	spin_lock_bh(&hfi_ack_lock);
+	list_del(&resp.node);
+	spin_unlock_bh(&hfi_ack_lock);
+
+	if (ret)
+		return ret;
+
+	if (resp.error) {
+		dev_err(gmu->dev, "Message %s id %d returned error %d\n",
+			a6xx_hfi_msg_id[id], seqnum, resp.error);
+		return -EINVAL;
+	}
+
+	if (payload && payload_size) {
+		int copy = min_t(u32, payload_size, sizeof(resp.payload));
+
+		memcpy(payload, resp.payload, copy);
+	}
+
+	return 0;
+}
+
+static int a6xx_hfi_send_gmu_init(struct a6xx_gmu *gmu, int boot_state)
+{
+	struct a6xx_hfi_msg_gmu_init_cmd msg = { 0 };
+
+	msg.dbg_buffer_addr = (u32) gmu->debug->iova;
+	msg.dbg_buffer_size = (u32) gmu->debug->size;
+	msg.boot_state = boot_state;
+
+	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_INIT, &msg, sizeof(msg),
+		NULL, 0);
+}
+
+static int a6xx_hfi_get_fw_version(struct a6xx_gmu *gmu, u32 *version)
+{
+	struct a6xx_hfi_msg_fw_version msg = { 0 };
+
+	/* Currently supporting version 1.1 */
+	msg.supported_version = (1 << 28) | (1 << 16);
+
+	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_FW_VERSION, &msg, sizeof(msg),
+		version, sizeof(*version));
+}
+
+static int a6xx_hfi_send_perf_table(struct a6xx_gmu *gmu)
+{
+	struct a6xx_hfi_msg_perf_table msg = { 0 };
+	int i;
+
+	msg.num_gpu_levels = gmu->nr_gpu_freqs;
+	msg.num_gmu_levels = gmu->nr_gmu_freqs;
+
+	for (i = 0; i < gmu->nr_gpu_freqs; i++) {
+		msg.gx_votes[i].vote = gmu->gx_arc_votes[i];
+		msg.gx_votes[i].freq = gmu->gpu_freqs[i] / 1000;
+	}
+
+	for (i = 0; i < gmu->nr_gmu_freqs; i++) {
+		msg.cx_votes[i].vote = gmu->cx_arc_votes[i];
+		msg.cx_votes[i].freq = gmu->gmu_freqs[i] / 1000;
+	}
+
+	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_PERF_TABLE, &msg, sizeof(msg),
+		NULL, 0);
+}
+
+static int a6xx_hfi_send_bw_table(struct a6xx_gmu *gmu)
+{
+	struct a6xx_hfi_msg_bw_table msg = { 0 };
+
+	/*
+	 * The sdm845 GMU doesn't do bus frequency scaling on its own but it
+	 * does need at least one entry in the list because it might be accessed
+	 * when the GMU is shutting down. Send a single "off" entry.
+	 */
+
+	msg.bw_level_num = 1;
+
+	msg.ddr_cmds_num = 3;
+	msg.ddr_wait_bitmask = 0x07;
+
+	msg.ddr_cmds_addrs[0] = 0x50000;
+	msg.ddr_cmds_addrs[1] = 0x5005c;
+	msg.ddr_cmds_addrs[2] = 0x5000c;
+
+	msg.ddr_cmds_data[0][0] =  0x40000000;
+	msg.ddr_cmds_data[0][1] =  0x40000000;
+	msg.ddr_cmds_data[0][2] =  0x40000000;
+
+	/*
+	 * These are the CX (CNOC) votes.  This is used but the values for the
+	 * sdm845 GMU are known and fixed so we can hard code them.
+	 */
+
+	msg.cnoc_cmds_num = 3;
+	msg.cnoc_wait_bitmask = 0x05;
+
+	msg.cnoc_cmds_addrs[0] = 0x50034;
+	msg.cnoc_cmds_addrs[1] = 0x5007c;
+	msg.cnoc_cmds_addrs[2] = 0x5004c;
+
+	msg.cnoc_cmds_data[0][0] =  0x40000000;
+	msg.cnoc_cmds_data[0][1] =  0x00000000;
+	msg.cnoc_cmds_data[0][2] =  0x40000000;
+
+	msg.cnoc_cmds_data[1][0] =  0x60000001;
+	msg.cnoc_cmds_data[1][1] =  0x20000001;
+	msg.cnoc_cmds_data[1][2] =  0x60000001;
+
+	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_BW_TABLE, &msg, sizeof(msg),
+		NULL, 0);
+}
+
+static int a6xx_hfi_send_test(struct a6xx_gmu *gmu)
+{
+	struct a6xx_hfi_msg_test msg = { 0 };
+
+	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_TEST, &msg, sizeof(msg),
+		NULL, 0);
+}
+
+int a6xx_hfi_start(struct a6xx_gmu *gmu, int boot_state)
+{
+	int ret;
+
+	ret = a6xx_hfi_send_gmu_init(gmu, boot_state);
+	if (ret)
+		return ret;
+
+	ret = a6xx_hfi_get_fw_version(gmu, NULL);
+	if (ret)
+		return ret;
+
+	/*
+	 * We have to get exchange version numbers per the sequence but at this
+	 * point th kernel driver doesn't need to know the exact version of
+	 * the GMU firmware
+	 */
+
+	ret = a6xx_hfi_send_perf_table(gmu);
+	if (ret)
+		return ret;
+
+	ret = a6xx_hfi_send_bw_table(gmu);
+	if (ret)
+		return ret;
+
+	/*
+	 * Let the GMU know that there won't be any more HFI messages until next
+	 * boot
+	 */
+	a6xx_hfi_send_test(gmu);
+
+	return 0;
+}
+
+void a6xx_hfi_stop(struct a6xx_gmu *gmu)
+{
+	int i;
+
+	for (i = 0; i < ARRAY_SIZE(gmu->queues); i++) {
+		struct a6xx_hfi_queue *queue = &gmu->queues[i];
+
+		if (!queue->header)
+			continue;
+
+		if (queue->header->read_index != queue->header->write_index)
+			dev_err(gmu->dev, "HFI queue %d is not empty\n", i);
+
+		queue->header->read_index = 0;
+		queue->header->write_index = 0;
+	}
+}
+
+static void a6xx_hfi_queue_init(struct a6xx_hfi_queue *queue,
+		struct a6xx_hfi_queue_header *header, void *virt, u64 iova,
+		u32 id)
+{
+	spin_lock_init(&queue->lock);
+	queue->header = header;
+	queue->data = virt;
+	atomic_set(&queue->seqnum, 0);
+
+	/* Set up the shared memory header */
+	header->iova = iova;
+	header->type =  10 << 8 | id;
+	header->status = 1;
+	header->size = SZ_4K >> 2;
+	header->msg_size = 0;
+	header->dropped = 0;
+	header->rx_watermark = 1;
+	header->tx_watermark = 1;
+	header->rx_request = 1;
+	header->tx_request = 0;
+	header->read_index = 0;
+	header->write_index = 0;
+}
+
+void a6xx_hfi_init(struct a6xx_gmu *gmu)
+{
+	struct a6xx_gmu_bo *hfi = gmu->hfi;
+	struct a6xx_hfi_queue_table_header *table = hfi->virt;
+	struct a6xx_hfi_queue_header *headers = hfi->virt + sizeof(*table);
+	u64 offset;
+	int table_size;
+
+	/*
+	 * The table size is the size of the table header plus all of the queue
+	 * headers
+	 */
+	table_size = sizeof(*table);
+	table_size += (ARRAY_SIZE(gmu->queues) *
+		sizeof(struct a6xx_hfi_queue_header));
+
+	table->version = 0;
+	table->size = table_size;
+	/* First queue header is located immediately after the table header */
+	table->qhdr0_offset = sizeof(*table) >> 2;
+	table->qhdr_size = sizeof(struct a6xx_hfi_queue_header) >> 2;
+	table->num_queues = ARRAY_SIZE(gmu->queues);
+	table->active_queues = ARRAY_SIZE(gmu->queues);
+
+	/* Command queue */
+	offset = SZ_4K;
+	a6xx_hfi_queue_init(&gmu->queues[0], &headers[0], hfi->virt + offset,
+		hfi->iova + offset, 0);
+
+	/* GMU response queue */
+	offset += SZ_4K;
+	a6xx_hfi_queue_init(&gmu->queues[1], &headers[1], hfi->virt + offset,
+		hfi->iova + offset, 4);
+}