xref: /linux/drivers/gpu/drm/xe/xe_rtp.c (revision 2697b79a469b68e3ad3640f55284359c1396278d)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2022 Intel Corporation
4  */
5 
6 #include "xe_rtp.h"
7 
8 #include <kunit/visibility.h>
9 
10 #include <drm/xe_drm.h>
11 
12 #include "xe_gt.h"
13 #include "xe_gt_topology.h"
14 #include "xe_macros.h"
15 #include "xe_reg_sr.h"
16 
17 /**
18  * DOC: Register Table Processing
19  *
20  * Internal infrastructure to define how registers should be updated based on
21  * rules and actions. This can be used to define tables with multiple entries
22  * (one per register) that will be walked over at some point in time to apply
23  * the values to the registers that have matching rules.
24  */
25 
26 static bool has_samedia(const struct xe_device *xe)
27 {
28 	return xe->info.media_verx100 >= 1300;
29 }
30 
31 static bool rule_matches(const struct xe_device *xe,
32 			 struct xe_gt *gt,
33 			 struct xe_hw_engine *hwe,
34 			 const struct xe_rtp_rule *rules,
35 			 unsigned int n_rules)
36 {
37 	const struct xe_rtp_rule *r;
38 	unsigned int i;
39 	bool match;
40 
41 	for (r = rules, i = 0; i < n_rules; r = &rules[++i]) {
42 		switch (r->match_type) {
43 		case XE_RTP_MATCH_PLATFORM:
44 			match = xe->info.platform == r->platform;
45 			break;
46 		case XE_RTP_MATCH_SUBPLATFORM:
47 			match = xe->info.platform == r->platform &&
48 				xe->info.subplatform == r->subplatform;
49 			break;
50 		case XE_RTP_MATCH_GRAPHICS_VERSION:
51 			match = xe->info.graphics_verx100 == r->ver_start &&
52 				(!has_samedia(xe) || !xe_gt_is_media_type(gt));
53 			break;
54 		case XE_RTP_MATCH_GRAPHICS_VERSION_RANGE:
55 			match = xe->info.graphics_verx100 >= r->ver_start &&
56 				xe->info.graphics_verx100 <= r->ver_end &&
57 				(!has_samedia(xe) || !xe_gt_is_media_type(gt));
58 			break;
59 		case XE_RTP_MATCH_GRAPHICS_STEP:
60 			match = xe->info.step.graphics >= r->step_start &&
61 				xe->info.step.graphics < r->step_end &&
62 				(!has_samedia(xe) || !xe_gt_is_media_type(gt));
63 			break;
64 		case XE_RTP_MATCH_MEDIA_VERSION:
65 			match = xe->info.media_verx100 == r->ver_start &&
66 				(!has_samedia(xe) || xe_gt_is_media_type(gt));
67 			break;
68 		case XE_RTP_MATCH_MEDIA_VERSION_RANGE:
69 			match = xe->info.media_verx100 >= r->ver_start &&
70 				xe->info.media_verx100 <= r->ver_end &&
71 				(!has_samedia(xe) || xe_gt_is_media_type(gt));
72 			break;
73 		case XE_RTP_MATCH_MEDIA_STEP:
74 			match = xe->info.step.media >= r->step_start &&
75 				xe->info.step.media < r->step_end &&
76 				(!has_samedia(xe) || xe_gt_is_media_type(gt));
77 			break;
78 		case XE_RTP_MATCH_INTEGRATED:
79 			match = !xe->info.is_dgfx;
80 			break;
81 		case XE_RTP_MATCH_DISCRETE:
82 			match = xe->info.is_dgfx;
83 			break;
84 		case XE_RTP_MATCH_ENGINE_CLASS:
85 			if (drm_WARN_ON(&xe->drm, !hwe))
86 				return false;
87 
88 			match = hwe->class == r->engine_class;
89 			break;
90 		case XE_RTP_MATCH_NOT_ENGINE_CLASS:
91 			if (drm_WARN_ON(&xe->drm, !hwe))
92 				return false;
93 
94 			match = hwe->class != r->engine_class;
95 			break;
96 		case XE_RTP_MATCH_FUNC:
97 			match = r->match_func(gt, hwe);
98 			break;
99 		default:
100 			drm_warn(&xe->drm, "Invalid RTP match %u\n",
101 				 r->match_type);
102 			match = false;
103 		}
104 
105 		if (!match)
106 			return false;
107 	}
108 
109 	return true;
110 }
111 
112 static void rtp_add_sr_entry(const struct xe_rtp_action *action,
113 			     struct xe_gt *gt,
114 			     u32 mmio_base,
115 			     struct xe_reg_sr *sr)
116 {
117 	struct xe_reg_sr_entry sr_entry = {
118 		.reg = action->reg,
119 		.clr_bits = action->clr_bits,
120 		.set_bits = action->set_bits,
121 		.read_mask = action->read_mask,
122 	};
123 
124 	sr_entry.reg.addr += mmio_base;
125 	xe_reg_sr_add(sr, &sr_entry, gt);
126 }
127 
128 static bool rtp_process_one_sr(const struct xe_rtp_entry_sr *entry,
129 			       struct xe_device *xe, struct xe_gt *gt,
130 			       struct xe_hw_engine *hwe, struct xe_reg_sr *sr)
131 {
132 	const struct xe_rtp_action *action;
133 	u32 mmio_base;
134 	unsigned int i;
135 
136 	if (!rule_matches(xe, gt, hwe, entry->rules, entry->n_rules))
137 		return false;
138 
139 	for (i = 0, action = &entry->actions[0]; i < entry->n_actions; action++, i++) {
140 		if ((entry->flags & XE_RTP_ENTRY_FLAG_FOREACH_ENGINE) ||
141 		    (action->flags & XE_RTP_ACTION_FLAG_ENGINE_BASE))
142 			mmio_base = hwe->mmio_base;
143 		else
144 			mmio_base = 0;
145 
146 		rtp_add_sr_entry(action, gt, mmio_base, sr);
147 	}
148 
149 	return true;
150 }
151 
152 static void rtp_get_context(struct xe_rtp_process_ctx *ctx,
153 			    struct xe_hw_engine **hwe,
154 			    struct xe_gt **gt,
155 			    struct xe_device **xe)
156 {
157 	switch (ctx->type) {
158 	case XE_RTP_PROCESS_TYPE_GT:
159 		*hwe = NULL;
160 		*gt = ctx->gt;
161 		*xe = gt_to_xe(*gt);
162 		break;
163 	case XE_RTP_PROCESS_TYPE_ENGINE:
164 		*hwe = ctx->hwe;
165 		*gt = (*hwe)->gt;
166 		*xe = gt_to_xe(*gt);
167 		break;
168 	};
169 }
170 
171 /**
172  * xe_rtp_process_ctx_enable_active_tracking - Enable tracking of active entries
173  *
174  * Set additional metadata to track what entries are considered "active", i.e.
175  * their rules match the condition. Bits are never cleared: entries with
176  * matching rules set the corresponding bit in the bitmap.
177  *
178  * @ctx: The context for processing the table
179  * @active_entries: bitmap to store the active entries
180  * @n_entries: number of entries to be processed
181  */
182 void xe_rtp_process_ctx_enable_active_tracking(struct xe_rtp_process_ctx *ctx,
183 					       unsigned long *active_entries,
184 					       size_t n_entries)
185 {
186 	ctx->active_entries = active_entries;
187 	ctx->n_entries = n_entries;
188 }
189 
190 static void rtp_mark_active(struct xe_device *xe,
191 			    struct xe_rtp_process_ctx *ctx,
192 			    unsigned int first, unsigned int last)
193 {
194 	if (!ctx->active_entries)
195 		return;
196 
197 	if (drm_WARN_ON(&xe->drm, last > ctx->n_entries))
198 		return;
199 
200 	if (first == last)
201 		bitmap_set(ctx->active_entries, first, 1);
202 	else
203 		bitmap_set(ctx->active_entries, first, last - first + 2);
204 }
205 
206 /**
207  * xe_rtp_process_to_sr - Process all rtp @entries, adding the matching ones to
208  *                        the save-restore argument.
209  * @ctx: The context for processing the table, with one of device, gt or hwe
210  * @entries: Table with RTP definitions
211  * @sr: Save-restore struct where matching rules execute the action. This can be
212  *      viewed as the "coalesced view" of multiple the tables. The bits for each
213  *      register set are expected not to collide with previously added entries
214  *
215  * Walk the table pointed by @entries (with an empty sentinel) and add all
216  * entries with matching rules to @sr. If @hwe is not NULL, its mmio_base is
217  * used to calculate the right register offset
218  */
219 void xe_rtp_process_to_sr(struct xe_rtp_process_ctx *ctx,
220 			  const struct xe_rtp_entry_sr *entries,
221 			  struct xe_reg_sr *sr)
222 {
223 	const struct xe_rtp_entry_sr *entry;
224 	struct xe_hw_engine *hwe = NULL;
225 	struct xe_gt *gt = NULL;
226 	struct xe_device *xe = NULL;
227 
228 	rtp_get_context(ctx, &hwe, &gt, &xe);
229 
230 	for (entry = entries; entry && entry->name; entry++) {
231 		bool match = false;
232 
233 		if (entry->flags & XE_RTP_ENTRY_FLAG_FOREACH_ENGINE) {
234 			struct xe_hw_engine *each_hwe;
235 			enum xe_hw_engine_id id;
236 
237 			for_each_hw_engine(each_hwe, gt, id)
238 				match |= rtp_process_one_sr(entry, xe, gt,
239 							    each_hwe, sr);
240 		} else {
241 			match = rtp_process_one_sr(entry, xe, gt, hwe, sr);
242 		}
243 
244 		if (match)
245 			rtp_mark_active(xe, ctx, entry - entries,
246 					entry - entries);
247 	}
248 }
249 EXPORT_SYMBOL_IF_KUNIT(xe_rtp_process_to_sr);
250 
251 /**
252  * xe_rtp_process - Process all rtp @entries, without running any action
253  * @ctx: The context for processing the table, with one of device, gt or hwe
254  * @entries: Table with RTP definitions
255  *
256  * Walk the table pointed by @entries (with an empty sentinel), executing the
257  * rules. A few differences from xe_rtp_process_to_sr():
258  *
259  * 1. There is no action associated with each entry since this uses
260  *    struct xe_rtp_entry. Its main use is for marking active workarounds via
261  *    xe_rtp_process_ctx_enable_active_tracking().
262  * 2. There is support for OR operations by having entries with no name.
263  */
264 void xe_rtp_process(struct xe_rtp_process_ctx *ctx,
265 		    const struct xe_rtp_entry *entries)
266 {
267 	const struct xe_rtp_entry *entry, *first_entry;
268 	struct xe_hw_engine *hwe;
269 	struct xe_gt *gt;
270 	struct xe_device *xe;
271 
272 	rtp_get_context(ctx, &hwe, &gt, &xe);
273 
274 	first_entry = entries;
275 	if (drm_WARN_ON(&xe->drm, !first_entry->name))
276 		return;
277 
278 	for (entry = entries; entry && entry->rules; entry++) {
279 		if (entry->name)
280 			first_entry = entry;
281 
282 		if (!rule_matches(xe, gt, hwe, entry->rules, entry->n_rules))
283 			continue;
284 
285 		/* Fast-forward entry, eliminating the OR'ed entries */
286 		for (entry++; entry && entry->rules; entry++)
287 			if (entry->name)
288 				break;
289 		entry--;
290 
291 		rtp_mark_active(xe, ctx, first_entry - entries,
292 				entry - entries);
293 	}
294 }
295 
296 bool xe_rtp_match_even_instance(const struct xe_gt *gt,
297 				const struct xe_hw_engine *hwe)
298 {
299 	return hwe->instance % 2 == 0;
300 }
301 
302 bool xe_rtp_match_first_render_or_compute(const struct xe_gt *gt,
303 					  const struct xe_hw_engine *hwe)
304 {
305 	u64 render_compute_mask = gt->info.engine_mask &
306 		(XE_HW_ENGINE_CCS_MASK | XE_HW_ENGINE_RCS_MASK);
307 
308 	return render_compute_mask &&
309 		hwe->engine_id == __ffs(render_compute_mask);
310 }
311 
312 bool xe_rtp_match_first_gslice_fused_off(const struct xe_gt *gt,
313 					 const struct xe_hw_engine *hwe)
314 {
315 	unsigned int dss_per_gslice = 4;
316 	unsigned int dss;
317 
318 	if (drm_WARN(&gt_to_xe(gt)->drm, xe_dss_mask_empty(gt->fuse_topo.g_dss_mask),
319 		     "Checking gslice for platform without geometry pipeline\n"))
320 		return false;
321 
322 	dss = xe_dss_mask_group_ffs(gt->fuse_topo.g_dss_mask, 0, 0);
323 
324 	return dss >= dss_per_gslice;
325 }
326