xref: /linux/drivers/gpu/drm/xe/xe_gt_ccs_mode.c (revision ad112b3a759cefc245695f7e612f17e6f38d2cc1)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2023 Intel Corporation
4  */
5 
6 #include <drm/drm_managed.h>
7 
8 #include "regs/xe_gt_regs.h"
9 #include "xe_assert.h"
10 #include "xe_gt.h"
11 #include "xe_gt_ccs_mode.h"
12 #include "xe_gt_sysfs.h"
13 #include "xe_mmio.h"
14 
15 static void __xe_gt_apply_ccs_mode(struct xe_gt *gt, u32 num_engines)
16 {
17 	u32 mode = CCS_MODE_CSLICE_0_3_MASK; /* disable all by default */
18 	int num_slices = hweight32(CCS_MASK(gt));
19 	struct xe_device *xe = gt_to_xe(gt);
20 	int width, cslice = 0;
21 	u32 config = 0;
22 
23 	xe_assert(xe, xe_gt_ccs_mode_enabled(gt));
24 
25 	xe_assert(xe, num_engines && num_engines <= num_slices);
26 	xe_assert(xe, !(num_slices % num_engines));
27 
28 	/*
29 	 * Loop over all available slices and assign each a user engine.
30 	 * For example, if there are four compute slices available, the
31 	 * assignment of compute slices to compute engines would be,
32 	 *
33 	 * With 1 engine (ccs0):
34 	 *   slice 0, 1, 2, 3: ccs0
35 	 *
36 	 * With 2 engines (ccs0, ccs1):
37 	 *   slice 0, 2: ccs0
38 	 *   slice 1, 3: ccs1
39 	 *
40 	 * With 4 engines (ccs0, ccs1, ccs2, ccs3):
41 	 *   slice 0: ccs0
42 	 *   slice 1: ccs1
43 	 *   slice 2: ccs2
44 	 *   slice 3: ccs3
45 	 */
46 	for (width = num_slices / num_engines; width; width--) {
47 		struct xe_hw_engine *hwe;
48 		enum xe_hw_engine_id id;
49 
50 		for_each_hw_engine(hwe, gt, id) {
51 			if (hwe->class != XE_ENGINE_CLASS_COMPUTE)
52 				continue;
53 
54 			if (hwe->logical_instance >= num_engines)
55 				break;
56 
57 			config |= BIT(hwe->instance) << XE_HW_ENGINE_CCS0;
58 
59 			/* If a slice is fused off, leave disabled */
60 			while ((CCS_MASK(gt) & BIT(cslice)) == 0)
61 				cslice++;
62 
63 			mode &= ~CCS_MODE_CSLICE(cslice, CCS_MODE_CSLICE_MASK);
64 			mode |= CCS_MODE_CSLICE(cslice, hwe->instance);
65 			cslice++;
66 		}
67 	}
68 
69 	xe_mmio_write32(gt, CCS_MODE, mode);
70 
71 	xe_gt_info(gt, "CCS_MODE=%x config:%08x, num_engines:%d, num_slices:%d\n",
72 		   mode, config, num_engines, num_slices);
73 }
74 
75 void xe_gt_apply_ccs_mode(struct xe_gt *gt)
76 {
77 	if (!gt->ccs_mode)
78 		return;
79 
80 	__xe_gt_apply_ccs_mode(gt, gt->ccs_mode);
81 }
82 
83 static ssize_t
84 num_cslices_show(struct device *kdev,
85 		 struct device_attribute *attr, char *buf)
86 {
87 	struct xe_gt *gt = kobj_to_gt(&kdev->kobj);
88 
89 	return sysfs_emit(buf, "%u\n", hweight32(CCS_MASK(gt)));
90 }
91 
92 static DEVICE_ATTR_RO(num_cslices);
93 
94 static ssize_t
95 ccs_mode_show(struct device *kdev,
96 	      struct device_attribute *attr, char *buf)
97 {
98 	struct xe_gt *gt = kobj_to_gt(&kdev->kobj);
99 
100 	return sysfs_emit(buf, "%u\n", gt->ccs_mode);
101 }
102 
103 static ssize_t
104 ccs_mode_store(struct device *kdev, struct device_attribute *attr,
105 	       const char *buff, size_t count)
106 {
107 	struct xe_gt *gt = kobj_to_gt(&kdev->kobj);
108 	struct xe_device *xe = gt_to_xe(gt);
109 	u32 num_engines, num_slices;
110 	int ret;
111 
112 	ret = kstrtou32(buff, 0, &num_engines);
113 	if (ret)
114 		return ret;
115 
116 	/*
117 	 * Ensure number of engines specified is valid and there is an
118 	 * exact multiple of engines for slices.
119 	 */
120 	num_slices = hweight32(CCS_MASK(gt));
121 	if (!num_engines || num_engines > num_slices || num_slices % num_engines) {
122 		xe_gt_dbg(gt, "Invalid compute config, %d engines %d slices\n",
123 			  num_engines, num_slices);
124 		return -EINVAL;
125 	}
126 
127 	/* CCS mode can only be updated when there are no drm clients */
128 	spin_lock(&xe->clients.lock);
129 	if (xe->clients.count) {
130 		spin_unlock(&xe->clients.lock);
131 		return -EBUSY;
132 	}
133 
134 	if (gt->ccs_mode != num_engines) {
135 		xe_gt_info(gt, "Setting compute mode to %d\n", num_engines);
136 		gt->ccs_mode = num_engines;
137 		xe_gt_reset_async(gt);
138 	}
139 
140 	spin_unlock(&xe->clients.lock);
141 
142 	return count;
143 }
144 
145 static DEVICE_ATTR_RW(ccs_mode);
146 
147 static const struct attribute *gt_ccs_mode_attrs[] = {
148 	&dev_attr_ccs_mode.attr,
149 	&dev_attr_num_cslices.attr,
150 	NULL,
151 };
152 
153 static void xe_gt_ccs_mode_sysfs_fini(struct drm_device *drm, void *arg)
154 {
155 	struct xe_gt *gt = arg;
156 
157 	sysfs_remove_files(gt->sysfs, gt_ccs_mode_attrs);
158 }
159 
160 /**
161  * xe_gt_ccs_mode_sysfs_init - Initialize CCS mode sysfs interfaces
162  * @gt: GT structure
163  *
164  * Through a per-gt 'ccs_mode' sysfs interface, the user can enable a fixed
165  * number of compute hardware engines to which the available compute slices
166  * are to be allocated. This user configuration change triggers a gt reset
167  * and it is expected that there are no open drm clients while doing so.
168  * The number of available compute slices is exposed to user through a per-gt
169  * 'num_cslices' sysfs interface.
170  *
171  * Returns: Returns error value for failure and 0 for success.
172  */
173 int xe_gt_ccs_mode_sysfs_init(struct xe_gt *gt)
174 {
175 	struct xe_device *xe = gt_to_xe(gt);
176 	int err;
177 
178 	if (!xe_gt_ccs_mode_enabled(gt))
179 		return 0;
180 
181 	err = sysfs_create_files(gt->sysfs, gt_ccs_mode_attrs);
182 	if (err)
183 		return err;
184 
185 	return drmm_add_action_or_reset(&xe->drm, xe_gt_ccs_mode_sysfs_fini, gt);
186 }
187