accel/amdxdna: Add hardware resource solver

The AI Engine consists of 2D array of tiles arranged as columns. Provides
the basic column allocation and release functions for the tile columns.

Co-developed-by: Min Ma <min.ma@amd.com>
Signed-off-by: Min Ma <min.ma@amd.com>
Reviewed-by: Jeffrey Hugo <quic_jhugo@quicinc.com>
Signed-off-by: Lizhi Hou <lizhi.hou@amd.com>
Signed-off-by: Jeffrey Hugo <quic_jhugo@quicinc.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20241118172942.2014541-5-lizhi.hou@amd.com

drivers/accel/amdxdna/Makefile

@@ -5,6 +5,7 @@ amdxdna-y := \
 	aie2_pci.o \
 	aie2_psp.o \
 	aie2_smu.o \
+	aie2_solver.o \
 	amdxdna_mailbox.o \
 	amdxdna_mailbox_helper.o \
 	amdxdna_pci_drv.o \

drivers/accel/amdxdna/aie2_pci.c

@@ -14,9 +14,14 @@
 
 #include "aie2_msg_priv.h"
 #include "aie2_pci.h"
+#include "aie2_solver.h"
 #include "amdxdna_mailbox.h"
 #include "amdxdna_pci_drv.h"
 
+int aie2_max_col = XRS_MAX_COL;
+module_param(aie2_max_col, uint, 0600);
+MODULE_PARM_DESC(aie2_max_col, "Maximum column could be used");
+
 /*
  * The management mailbox channel is allocated by firmware.
  * The related register and ring buffer information is on SRAM BAR.
@@ -307,6 +312,7 @@ static int aie2_init(struct amdxdna_dev *xdna)
 {
 	struct pci_dev *pdev = to_pci_dev(xdna->ddev.dev);
 	void __iomem *tbl[PCI_NUM_RESOURCES] = {0};
+	struct init_config xrs_cfg = { 0 };
 	struct amdxdna_dev_hdl *ndev;
 	struct psp_config psp_conf;
 	const struct firmware *fw;
@@ -403,7 +409,22 @@ static int aie2_init(struct amdxdna_dev *xdna)
 		XDNA_ERR(xdna, "Query firmware failed, ret %d", ret);
 		goto stop_hw;
 	}
-	ndev->total_col = ndev->metadata.cols;
+	ndev->total_col = min(aie2_max_col, ndev->metadata.cols);
+
+	xrs_cfg.clk_list.num_levels = 3;
+	xrs_cfg.clk_list.cu_clk_list[0] = 0;
+	xrs_cfg.clk_list.cu_clk_list[1] = 800;
+	xrs_cfg.clk_list.cu_clk_list[2] = 1000;
+	xrs_cfg.sys_eff_factor = 1;
+	xrs_cfg.ddev = &xdna->ddev;
+	xrs_cfg.total_col = ndev->total_col;
+
+	xdna->xrs_hdl = xrsm_init(&xrs_cfg);
+	if (!xdna->xrs_hdl) {
+		XDNA_ERR(xdna, "Initialize resolver failed");
+		ret = -EINVAL;
+		goto stop_hw;
+	}
 
 	release_firmware(fw);
 	return 0;

drivers/accel/amdxdna/aie2_solver.c

@@ -0,0 +1,330 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Copyright (C) 2022-2024, Advanced Micro Devices, Inc.
+ */
+
+#include <drm/drm_device.h>
+#include <drm/drm_managed.h>
+#include <drm/drm_print.h>
+#include <linux/bitops.h>
+#include <linux/bitmap.h>
+
+#include "aie2_solver.h"
+
+struct partition_node {
+	struct list_head	list;
+	u32			nshared;	/* # shared requests */
+	u32			start_col;	/* start column */
+	u32			ncols;		/* # columns */
+	bool			exclusive;	/* can not be shared if set */
+};
+
+struct solver_node {
+	struct list_head	list;
+	u64			rid;		/* Request ID from consumer */
+
+	struct partition_node	*pt_node;
+	void			*cb_arg;
+	u32			cols_len;
+	u32			start_cols[] __counted_by(cols_len);
+};
+
+struct solver_rgroup {
+	u32				rgid;
+	u32				nnode;
+	u32				npartition_node;
+
+	DECLARE_BITMAP(resbit, XRS_MAX_COL);
+	struct list_head		node_list;
+	struct list_head		pt_node_list;
+};
+
+struct solver_state {
+	struct solver_rgroup		rgp;
+	struct init_config		cfg;
+	struct xrs_action_ops		*actions;
+};
+
+static u32 calculate_gops(struct aie_qos *rqos)
+{
+	u32 service_rate = 0;
+
+	if (rqos->latency)
+		service_rate = (1000 / rqos->latency);
+
+	if (rqos->fps > service_rate)
+		return rqos->fps * rqos->gops;
+
+	return service_rate * rqos->gops;
+}
+
+/*
+ * qos_meet() - Check the QOS request can be met.
+ */
+static int qos_meet(struct solver_state *xrs, struct aie_qos *rqos, u32 cgops)
+{
+	u32 request_gops = calculate_gops(rqos) * xrs->cfg.sys_eff_factor;
+
+	if (request_gops <= cgops)
+		return 0;
+
+	return -EINVAL;
+}
+
+/*
+ * sanity_check() - Do a basic sanity check on allocation request.
+ */
+static int sanity_check(struct solver_state *xrs, struct alloc_requests *req)
+{
+	struct cdo_parts *cdop = &req->cdo;
+	struct aie_qos *rqos = &req->rqos;
+	u32 cu_clk_freq;
+
+	if (cdop->ncols > xrs->cfg.total_col)
+		return -EINVAL;
+
+	/*
+	 * We can find at least one CDOs groups that meet the
+	 * GOPs requirement.
+	 */
+	cu_clk_freq = xrs->cfg.clk_list.cu_clk_list[xrs->cfg.clk_list.num_levels - 1];
+
+	if (qos_meet(xrs, rqos, cdop->qos_cap.opc * cu_clk_freq / 1000))
+		return -EINVAL;
+
+	return 0;
+}
+
+static struct solver_node *rg_search_node(struct solver_rgroup *rgp, u64 rid)
+{
+	struct solver_node *node;
+
+	list_for_each_entry(node, &rgp->node_list, list) {
+		if (node->rid == rid)
+			return node;
+	}
+
+	return NULL;
+}
+
+static void remove_partition_node(struct solver_rgroup *rgp,
+				  struct partition_node *pt_node)
+{
+	pt_node->nshared--;
+	if (pt_node->nshared > 0)
+		return;
+
+	list_del(&pt_node->list);
+	rgp->npartition_node--;
+
+	bitmap_clear(rgp->resbit, pt_node->start_col, pt_node->ncols);
+	kfree(pt_node);
+}
+
+static void remove_solver_node(struct solver_rgroup *rgp,
+			       struct solver_node *node)
+{
+	list_del(&node->list);
+	rgp->nnode--;
+
+	if (node->pt_node)
+		remove_partition_node(rgp, node->pt_node);
+
+	kfree(node);
+}
+
+static int get_free_partition(struct solver_state *xrs,
+			      struct solver_node *snode,
+			      struct alloc_requests *req)
+{
+	struct partition_node *pt_node;
+	u32 ncols = req->cdo.ncols;
+	u32 col, i;
+
+	for (i = 0; i < snode->cols_len; i++) {
+		col = snode->start_cols[i];
+		if (find_next_bit(xrs->rgp.resbit, XRS_MAX_COL, col) >= col + ncols)
+			break;
+	}
+
+	if (i == snode->cols_len)
+		return -ENODEV;
+
+	pt_node = kzalloc(sizeof(*pt_node), GFP_KERNEL);
+	if (!pt_node)
+		return -ENOMEM;
+
+	pt_node->nshared = 1;
+	pt_node->start_col = col;
+	pt_node->ncols = ncols;
+
+	/*
+	 * Before fully support latency in QoS, if a request
+	 * specifies a non-zero latency value, it will not share
+	 * the partition with other requests.
+	 */
+	if (req->rqos.latency)
+		pt_node->exclusive = true;
+
+	list_add_tail(&pt_node->list, &xrs->rgp.pt_node_list);
+	xrs->rgp.npartition_node++;
+	bitmap_set(xrs->rgp.resbit, pt_node->start_col, pt_node->ncols);
+
+	snode->pt_node = pt_node;
+
+	return 0;
+}
+
+static int allocate_partition(struct solver_state *xrs,
+			      struct solver_node *snode,
+			      struct alloc_requests *req)
+{
+	struct partition_node *pt_node, *rpt_node = NULL;
+	int idx, ret;
+
+	ret = get_free_partition(xrs, snode, req);
+	if (!ret)
+		return ret;
+
+	/* try to get a share-able partition */
+	list_for_each_entry(pt_node, &xrs->rgp.pt_node_list, list) {
+		if (pt_node->exclusive)
+			continue;
+
+		if (rpt_node && pt_node->nshared >= rpt_node->nshared)
+			continue;
+
+		for (idx = 0; idx < snode->cols_len; idx++) {
+			if (snode->start_cols[idx] != pt_node->start_col)
+				continue;
+
+			if (req->cdo.ncols != pt_node->ncols)
+				continue;
+
+			rpt_node = pt_node;
+			break;
+		}
+	}
+
+	if (!rpt_node)
+		return -ENODEV;
+
+	rpt_node->nshared++;
+	snode->pt_node = rpt_node;
+
+	return 0;
+}
+
+static struct solver_node *create_solver_node(struct solver_state *xrs,
+					      struct alloc_requests *req)
+{
+	struct cdo_parts *cdop = &req->cdo;
+	struct solver_node *node;
+	int ret;
+
+	node = kzalloc(struct_size(node, start_cols, cdop->cols_len), GFP_KERNEL);
+	if (!node)
+		return ERR_PTR(-ENOMEM);
+
+	node->rid = req->rid;
+	node->cols_len = cdop->cols_len;
+	memcpy(node->start_cols, cdop->start_cols, cdop->cols_len * sizeof(u32));
+
+	ret = allocate_partition(xrs, node, req);
+	if (ret)
+		goto free_node;
+
+	list_add_tail(&node->list, &xrs->rgp.node_list);
+	xrs->rgp.nnode++;
+	return node;
+
+free_node:
+	kfree(node);
+	return ERR_PTR(ret);
+}
+
+static void fill_load_action(struct solver_state *xrs,
+			     struct solver_node *snode,
+			     struct xrs_action_load *action)
+{
+	action->rid = snode->rid;
+	action->part.start_col = snode->pt_node->start_col;
+	action->part.ncols = snode->pt_node->ncols;
+}
+
+int xrs_allocate_resource(void *hdl, struct alloc_requests *req, void *cb_arg)
+{
+	struct xrs_action_load load_act;
+	struct solver_node *snode;
+	struct solver_state *xrs;
+	int ret;
+
+	xrs = (struct solver_state *)hdl;
+
+	ret = sanity_check(xrs, req);
+	if (ret) {
+		drm_err(xrs->cfg.ddev, "invalid request");
+		return ret;
+	}
+
+	if (rg_search_node(&xrs->rgp, req->rid)) {
+		drm_err(xrs->cfg.ddev, "rid %lld is in-use", req->rid);
+		return -EEXIST;
+	}
+
+	snode = create_solver_node(xrs, req);
+	if (IS_ERR(snode))
+		return PTR_ERR(snode);
+
+	fill_load_action(xrs, snode, &load_act);
+	ret = xrs->cfg.actions->load(cb_arg, &load_act);
+	if (ret)
+		goto free_node;
+
+	snode->cb_arg = cb_arg;
+
+	drm_dbg(xrs->cfg.ddev, "start col %d ncols %d\n",
+		snode->pt_node->start_col, snode->pt_node->ncols);
+
+	return 0;
+
+free_node:
+	remove_solver_node(&xrs->rgp, snode);
+
+	return ret;
+}
+
+int xrs_release_resource(void *hdl, u64 rid)
+{
+	struct solver_state *xrs = hdl;
+	struct solver_node *node;
+
+	node = rg_search_node(&xrs->rgp, rid);
+	if (!node) {
+		drm_err(xrs->cfg.ddev, "node not exist");
+		return -ENODEV;
+	}
+
+	xrs->cfg.actions->unload(node->cb_arg);
+	remove_solver_node(&xrs->rgp, node);
+
+	return 0;
+}
+
+void *xrsm_init(struct init_config *cfg)
+{
+	struct solver_rgroup *rgp;
+	struct solver_state *xrs;
+
+	xrs = drmm_kzalloc(cfg->ddev, sizeof(*xrs), GFP_KERNEL);
+	if (!xrs)
+		return NULL;
+
+	memcpy(&xrs->cfg, cfg, sizeof(*cfg));
+
+	rgp = &xrs->rgp;
+	INIT_LIST_HEAD(&rgp->node_list);
+	INIT_LIST_HEAD(&rgp->pt_node_list);
+
+	return xrs;
+}

drivers/accel/amdxdna/aie2_solver.h

@@ -0,0 +1,154 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * Copyright (C) 2023-2024, Advanced Micro Devices, Inc.
+ */
+
+#ifndef _AIE2_SOLVER_H
+#define _AIE2_SOLVER_H
+
+#define XRS_MAX_COL 128
+
+/*
+ * Structure used to describe a partition. A partition is column based
+ * allocation unit described by its start column and number of columns.
+ */
+struct aie_part {
+	u32	start_col;
+	u32	ncols;
+};
+
+/*
+ * The QoS capabilities of a given AIE partition.
+ */
+struct aie_qos_cap {
+	u32     opc;            /* operations per cycle */
+	u32     dma_bw;         /* DMA bandwidth */
+};
+
+/*
+ * QoS requirement of a resource allocation.
+ */
+struct aie_qos {
+	u32	gops;		/* Giga operations */
+	u32	fps;		/* Frames per second */
+	u32	dma_bw;		/* DMA bandwidth */
+	u32	latency;	/* Frame response latency */
+	u32	exec_time;	/* Frame execution time */
+	u32	priority;	/* Request priority */
+};
+
+/*
+ * Structure used to describe a relocatable CDO (Configuration Data Object).
+ */
+struct cdo_parts {
+	u32		   *start_cols;		/* Start column array */
+	u32		   cols_len;		/* Length of start column array */
+	u32		   ncols;		/* # of column */
+	struct aie_qos_cap qos_cap;		/* CDO QoS capabilities */
+};
+
+/*
+ * Structure used to describe a request to allocate.
+ */
+struct alloc_requests {
+	u64			rid;
+	struct cdo_parts	cdo;
+	struct aie_qos		rqos;		/* Requested QoS */
+};
+
+/*
+ * Load callback argument
+ */
+struct xrs_action_load {
+	u32                     rid;
+	struct aie_part         part;
+};
+
+/*
+ * Define the power level available
+ *
+ * POWER_LEVEL_MIN:
+ *     Lowest power level. Usually set when all actions are unloaded.
+ *
+ * POWER_LEVEL_n
+ *     Power levels 0 - n, is a step increase in system frequencies
+ */
+enum power_level {
+	POWER_LEVEL_MIN = 0x0,
+	POWER_LEVEL_0   = 0x1,
+	POWER_LEVEL_1   = 0x2,
+	POWER_LEVEL_2   = 0x3,
+	POWER_LEVEL_3   = 0x4,
+	POWER_LEVEL_4   = 0x5,
+	POWER_LEVEL_5   = 0x6,
+	POWER_LEVEL_6   = 0x7,
+	POWER_LEVEL_7   = 0x8,
+	POWER_LEVEL_NUM,
+};
+
+/*
+ * Structure used to describe the frequency table.
+ * Resource solver chooses the frequency from the table
+ * to meet the QOS requirements.
+ */
+struct clk_list_info {
+	u32        num_levels;                     /* available power levels */
+	u32        cu_clk_list[POWER_LEVEL_NUM];   /* available aie clock frequencies in Mhz*/
+};
+
+struct xrs_action_ops {
+	int (*load)(void *cb_arg, struct xrs_action_load *action);
+	int (*unload)(void *cb_arg);
+};
+
+/*
+ * Structure used to describe information for solver during initialization.
+ */
+struct init_config {
+	u32			total_col;
+	u32			sys_eff_factor; /* system efficiency factor */
+	u32			latency_adj;    /* latency adjustment in ms */
+	struct clk_list_info	clk_list;       /* List of frequencies available in system */
+	struct drm_device	*ddev;
+	struct xrs_action_ops	*actions;
+};
+
+/*
+ * xrsm_init() - Register resource solver. Resource solver client needs
+ *              to call this function to register itself.
+ *
+ * @cfg:	The system metrics for resource solver to use
+ *
+ * Return:	A resource solver handle
+ *
+ * Note: We should only create one handle per AIE array to be managed.
+ */
+void *xrsm_init(struct init_config *cfg);
+
+/*
+ * xrs_allocate_resource() - Request to allocate resources for a given context
+ *                           and a partition metadata. (See struct part_meta)
+ *
+ * @hdl:	Resource solver handle obtained from xrs_init()
+ * @req:	Input to the Resource solver including request id
+ *		and partition metadata.
+ * @cb_arg:	callback argument pointer
+ *
+ * Return:	0 when successful.
+ *		Or standard error number when failing
+ *
+ * Note:
+ *      There is no lock mechanism inside resource solver. So it is
+ *      the caller's responsibility to lock down XCLBINs and grab
+ *      necessary lock.
+ */
+int xrs_allocate_resource(void *hdl, struct alloc_requests *req, void *cb_arg);
+
+/*
+ * xrs_release_resource() - Request to free resources for a given context.
+ *
+ * @hdl:	Resource solver handle obtained from xrs_init()
+ * @rid:	The Request ID to identify the requesting context
+ */
+int xrs_release_resource(void *hdl, u64 rid);
+#endif /* _AIE2_SOLVER_H */

drivers/accel/amdxdna/amdxdna_pci_drv.h

@@ -58,6 +58,7 @@ struct amdxdna_dev {
 	struct drm_device		ddev;
 	struct amdxdna_dev_hdl		*dev_handle;
 	const struct amdxdna_dev_info	*dev_info;
+	void				*xrs_hdl;
 
 	struct mutex			dev_lock; /* per device lock */
 	struct amdxdna_fw_ver		fw_ver;