xref: /linux/drivers/gpu/drm/xe/xe_hw_engine.c (revision 6812ce4e4379ffc99c52401ec28f0d7ffbc36206)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2021 Intel Corporation
4  */
5 
6 #include "xe_hw_engine.h"
7 
8 #include <linux/nospec.h>
9 
10 #include <drm/drm_managed.h>
11 #include <drm/drm_print.h>
12 #include <uapi/drm/xe_drm.h>
13 #include <generated/xe_wa_oob.h>
14 
15 #include "regs/xe_engine_regs.h"
16 #include "regs/xe_gt_regs.h"
17 #include "regs/xe_irq_regs.h"
18 #include "xe_assert.h"
19 #include "xe_bo.h"
20 #include "xe_configfs.h"
21 #include "xe_device.h"
22 #include "xe_execlist.h"
23 #include "xe_force_wake.h"
24 #include "xe_gsc.h"
25 #include "xe_gt.h"
26 #include "xe_gt_ccs_mode.h"
27 #include "xe_gt_clock.h"
28 #include "xe_gt_printk.h"
29 #include "xe_gt_mcr.h"
30 #include "xe_gt_topology.h"
31 #include "xe_guc_capture.h"
32 #include "xe_hw_engine_group.h"
33 #include "xe_hw_fence.h"
34 #include "xe_irq.h"
35 #include "xe_lrc.h"
36 #include "xe_mmio.h"
37 #include "xe_reg_sr.h"
38 #include "xe_reg_whitelist.h"
39 #include "xe_rtp.h"
40 #include "xe_sched_job.h"
41 #include "xe_sriov.h"
42 #include "xe_tuning.h"
43 #include "xe_uc_fw.h"
44 #include "xe_wa.h"
45 
46 #define MAX_MMIO_BASES 3
47 struct engine_info {
48 	const char *name;
49 	unsigned int class : 8;
50 	unsigned int instance : 8;
51 	unsigned int irq_offset : 8;
52 	enum xe_force_wake_domains domain;
53 	u32 mmio_base;
54 };
55 
56 static const struct engine_info engine_infos[] = {
57 	[XE_HW_ENGINE_RCS0] = {
58 		.name = "rcs0",
59 		.class = XE_ENGINE_CLASS_RENDER,
60 		.instance = 0,
61 		.irq_offset = ilog2(INTR_RCS0),
62 		.domain = XE_FW_RENDER,
63 		.mmio_base = RENDER_RING_BASE,
64 	},
65 	[XE_HW_ENGINE_BCS0] = {
66 		.name = "bcs0",
67 		.class = XE_ENGINE_CLASS_COPY,
68 		.instance = 0,
69 		.irq_offset = ilog2(INTR_BCS(0)),
70 		.domain = XE_FW_GT,
71 		.mmio_base = BLT_RING_BASE,
72 	},
73 	[XE_HW_ENGINE_BCS1] = {
74 		.name = "bcs1",
75 		.class = XE_ENGINE_CLASS_COPY,
76 		.instance = 1,
77 		.irq_offset = ilog2(INTR_BCS(1)),
78 		.domain = XE_FW_GT,
79 		.mmio_base = XEHPC_BCS1_RING_BASE,
80 	},
81 	[XE_HW_ENGINE_BCS2] = {
82 		.name = "bcs2",
83 		.class = XE_ENGINE_CLASS_COPY,
84 		.instance = 2,
85 		.irq_offset = ilog2(INTR_BCS(2)),
86 		.domain = XE_FW_GT,
87 		.mmio_base = XEHPC_BCS2_RING_BASE,
88 	},
89 	[XE_HW_ENGINE_BCS3] = {
90 		.name = "bcs3",
91 		.class = XE_ENGINE_CLASS_COPY,
92 		.instance = 3,
93 		.irq_offset = ilog2(INTR_BCS(3)),
94 		.domain = XE_FW_GT,
95 		.mmio_base = XEHPC_BCS3_RING_BASE,
96 	},
97 	[XE_HW_ENGINE_BCS4] = {
98 		.name = "bcs4",
99 		.class = XE_ENGINE_CLASS_COPY,
100 		.instance = 4,
101 		.irq_offset = ilog2(INTR_BCS(4)),
102 		.domain = XE_FW_GT,
103 		.mmio_base = XEHPC_BCS4_RING_BASE,
104 	},
105 	[XE_HW_ENGINE_BCS5] = {
106 		.name = "bcs5",
107 		.class = XE_ENGINE_CLASS_COPY,
108 		.instance = 5,
109 		.irq_offset = ilog2(INTR_BCS(5)),
110 		.domain = XE_FW_GT,
111 		.mmio_base = XEHPC_BCS5_RING_BASE,
112 	},
113 	[XE_HW_ENGINE_BCS6] = {
114 		.name = "bcs6",
115 		.class = XE_ENGINE_CLASS_COPY,
116 		.instance = 6,
117 		.irq_offset = ilog2(INTR_BCS(6)),
118 		.domain = XE_FW_GT,
119 		.mmio_base = XEHPC_BCS6_RING_BASE,
120 	},
121 	[XE_HW_ENGINE_BCS7] = {
122 		.name = "bcs7",
123 		.class = XE_ENGINE_CLASS_COPY,
124 		.irq_offset = ilog2(INTR_BCS(7)),
125 		.instance = 7,
126 		.domain = XE_FW_GT,
127 		.mmio_base = XEHPC_BCS7_RING_BASE,
128 	},
129 	[XE_HW_ENGINE_BCS8] = {
130 		.name = "bcs8",
131 		.class = XE_ENGINE_CLASS_COPY,
132 		.instance = 8,
133 		.irq_offset = ilog2(INTR_BCS8),
134 		.domain = XE_FW_GT,
135 		.mmio_base = XEHPC_BCS8_RING_BASE,
136 	},
137 
138 	[XE_HW_ENGINE_VCS0] = {
139 		.name = "vcs0",
140 		.class = XE_ENGINE_CLASS_VIDEO_DECODE,
141 		.instance = 0,
142 		.irq_offset = 32 + ilog2(INTR_VCS(0)),
143 		.domain = XE_FW_MEDIA_VDBOX0,
144 		.mmio_base = BSD_RING_BASE,
145 	},
146 	[XE_HW_ENGINE_VCS1] = {
147 		.name = "vcs1",
148 		.class = XE_ENGINE_CLASS_VIDEO_DECODE,
149 		.instance = 1,
150 		.irq_offset = 32 + ilog2(INTR_VCS(1)),
151 		.domain = XE_FW_MEDIA_VDBOX1,
152 		.mmio_base = BSD2_RING_BASE,
153 	},
154 	[XE_HW_ENGINE_VCS2] = {
155 		.name = "vcs2",
156 		.class = XE_ENGINE_CLASS_VIDEO_DECODE,
157 		.instance = 2,
158 		.irq_offset = 32 + ilog2(INTR_VCS(2)),
159 		.domain = XE_FW_MEDIA_VDBOX2,
160 		.mmio_base = BSD3_RING_BASE,
161 	},
162 	[XE_HW_ENGINE_VCS3] = {
163 		.name = "vcs3",
164 		.class = XE_ENGINE_CLASS_VIDEO_DECODE,
165 		.instance = 3,
166 		.irq_offset = 32 + ilog2(INTR_VCS(3)),
167 		.domain = XE_FW_MEDIA_VDBOX3,
168 		.mmio_base = BSD4_RING_BASE,
169 	},
170 	[XE_HW_ENGINE_VCS4] = {
171 		.name = "vcs4",
172 		.class = XE_ENGINE_CLASS_VIDEO_DECODE,
173 		.instance = 4,
174 		.irq_offset = 32 + ilog2(INTR_VCS(4)),
175 		.domain = XE_FW_MEDIA_VDBOX4,
176 		.mmio_base = XEHP_BSD5_RING_BASE,
177 	},
178 	[XE_HW_ENGINE_VCS5] = {
179 		.name = "vcs5",
180 		.class = XE_ENGINE_CLASS_VIDEO_DECODE,
181 		.instance = 5,
182 		.irq_offset = 32 + ilog2(INTR_VCS(5)),
183 		.domain = XE_FW_MEDIA_VDBOX5,
184 		.mmio_base = XEHP_BSD6_RING_BASE,
185 	},
186 	[XE_HW_ENGINE_VCS6] = {
187 		.name = "vcs6",
188 		.class = XE_ENGINE_CLASS_VIDEO_DECODE,
189 		.instance = 6,
190 		.irq_offset = 32 + ilog2(INTR_VCS(6)),
191 		.domain = XE_FW_MEDIA_VDBOX6,
192 		.mmio_base = XEHP_BSD7_RING_BASE,
193 	},
194 	[XE_HW_ENGINE_VCS7] = {
195 		.name = "vcs7",
196 		.class = XE_ENGINE_CLASS_VIDEO_DECODE,
197 		.instance = 7,
198 		.irq_offset = 32 + ilog2(INTR_VCS(7)),
199 		.domain = XE_FW_MEDIA_VDBOX7,
200 		.mmio_base = XEHP_BSD8_RING_BASE,
201 	},
202 	[XE_HW_ENGINE_VECS0] = {
203 		.name = "vecs0",
204 		.class = XE_ENGINE_CLASS_VIDEO_ENHANCE,
205 		.instance = 0,
206 		.irq_offset = 32 + ilog2(INTR_VECS(0)),
207 		.domain = XE_FW_MEDIA_VEBOX0,
208 		.mmio_base = VEBOX_RING_BASE,
209 	},
210 	[XE_HW_ENGINE_VECS1] = {
211 		.name = "vecs1",
212 		.class = XE_ENGINE_CLASS_VIDEO_ENHANCE,
213 		.instance = 1,
214 		.irq_offset = 32 + ilog2(INTR_VECS(1)),
215 		.domain = XE_FW_MEDIA_VEBOX1,
216 		.mmio_base = VEBOX2_RING_BASE,
217 	},
218 	[XE_HW_ENGINE_VECS2] = {
219 		.name = "vecs2",
220 		.class = XE_ENGINE_CLASS_VIDEO_ENHANCE,
221 		.instance = 2,
222 		.irq_offset = 32 + ilog2(INTR_VECS(2)),
223 		.domain = XE_FW_MEDIA_VEBOX2,
224 		.mmio_base = XEHP_VEBOX3_RING_BASE,
225 	},
226 	[XE_HW_ENGINE_VECS3] = {
227 		.name = "vecs3",
228 		.class = XE_ENGINE_CLASS_VIDEO_ENHANCE,
229 		.instance = 3,
230 		.irq_offset = 32 + ilog2(INTR_VECS(3)),
231 		.domain = XE_FW_MEDIA_VEBOX3,
232 		.mmio_base = XEHP_VEBOX4_RING_BASE,
233 	},
234 	[XE_HW_ENGINE_CCS0] = {
235 		.name = "ccs0",
236 		.class = XE_ENGINE_CLASS_COMPUTE,
237 		.instance = 0,
238 		.irq_offset = ilog2(INTR_CCS(0)),
239 		.domain = XE_FW_RENDER,
240 		.mmio_base = COMPUTE0_RING_BASE,
241 	},
242 	[XE_HW_ENGINE_CCS1] = {
243 		.name = "ccs1",
244 		.class = XE_ENGINE_CLASS_COMPUTE,
245 		.instance = 1,
246 		.irq_offset = ilog2(INTR_CCS(1)),
247 		.domain = XE_FW_RENDER,
248 		.mmio_base = COMPUTE1_RING_BASE,
249 	},
250 	[XE_HW_ENGINE_CCS2] = {
251 		.name = "ccs2",
252 		.class = XE_ENGINE_CLASS_COMPUTE,
253 		.instance = 2,
254 		.irq_offset = ilog2(INTR_CCS(2)),
255 		.domain = XE_FW_RENDER,
256 		.mmio_base = COMPUTE2_RING_BASE,
257 	},
258 	[XE_HW_ENGINE_CCS3] = {
259 		.name = "ccs3",
260 		.class = XE_ENGINE_CLASS_COMPUTE,
261 		.instance = 3,
262 		.irq_offset = ilog2(INTR_CCS(3)),
263 		.domain = XE_FW_RENDER,
264 		.mmio_base = COMPUTE3_RING_BASE,
265 	},
266 	[XE_HW_ENGINE_GSCCS0] = {
267 		.name = "gsccs0",
268 		.class = XE_ENGINE_CLASS_OTHER,
269 		.instance = OTHER_GSC_INSTANCE,
270 		.domain = XE_FW_GSC,
271 		.mmio_base = GSCCS_RING_BASE,
272 	},
273 };
274 
hw_engine_fini(void * arg)275 static void hw_engine_fini(void *arg)
276 {
277 	struct xe_hw_engine *hwe = arg;
278 
279 	if (hwe->exl_port)
280 		xe_execlist_port_destroy(hwe->exl_port);
281 
282 	hwe->gt = NULL;
283 }
284 
285 /**
286  * xe_hw_engine_mmio_read32() - Read engine register
287  * @hwe: engine
288  * @reg: register to read from
289  *
290  * This function will read from an engine specific register.
291  * Forcewake must be held by the caller.
292  *
293  * Return: value of the 32-bit register.
294  */
xe_hw_engine_mmio_read32(struct xe_hw_engine * hwe,struct xe_reg reg)295 u32 xe_hw_engine_mmio_read32(struct xe_hw_engine *hwe, struct xe_reg reg)
296 {
297 	xe_gt_assert(hwe->gt, !(reg.addr & hwe->mmio_base));
298 	xe_force_wake_assert_held(gt_to_fw(hwe->gt), hwe->domain);
299 
300 	reg.addr += hwe->mmio_base;
301 
302 	return xe_mmio_read32(&hwe->gt->mmio, reg);
303 }
304 
xe_hw_engine_enable_ring(struct xe_hw_engine * hwe)305 void xe_hw_engine_enable_ring(struct xe_hw_engine *hwe)
306 {
307 	xe_mmio_write32(&hwe->gt->mmio, RING_HWS_PGA(hwe->mmio_base),
308 			xe_bo_ggtt_addr(hwe->hwsp));
309 }
310 
xe_hw_engine_match_fixed_cslice_mode(const struct xe_device * xe,const struct xe_gt * gt,const struct xe_hw_engine * hwe)311 static bool xe_hw_engine_match_fixed_cslice_mode(const struct xe_device *xe,
312 						 const struct xe_gt *gt,
313 						 const struct xe_hw_engine *hwe)
314 {
315 	/*
316 	 * Xe3p no longer supports load balance mode, so "fixed cslice" mode
317 	 * is automatic and no RCU_MODE programming is required.
318 	 */
319 	if (GRAPHICS_VER(gt_to_xe(gt)) >= 35)
320 		return false;
321 
322 	return xe_gt_ccs_mode_enabled(gt) &&
323 	       xe_rtp_match_first_render_or_compute(xe, gt, hwe);
324 }
325 
xe_rtp_cfeg_wmtp_disabled(const struct xe_device * xe,const struct xe_gt * gt,const struct xe_hw_engine * hwe)326 static bool xe_rtp_cfeg_wmtp_disabled(const struct xe_device *xe,
327 				      const struct xe_gt *gt,
328 				      const struct xe_hw_engine *hwe)
329 {
330 	if (GRAPHICS_VER(xe) < 20)
331 		return false;
332 
333 	if (hwe->class != XE_ENGINE_CLASS_COMPUTE &&
334 	    hwe->class != XE_ENGINE_CLASS_RENDER)
335 		return false;
336 
337 	return xe_mmio_read32(&hwe->gt->mmio, XEHP_FUSE4) & CFEG_WMTP_DISABLE;
338 }
339 
blit_cctl_val(struct xe_gt * gt,struct xe_hw_engine * hwe)340 static u32 blit_cctl_val(struct xe_gt *gt, struct xe_hw_engine *hwe)
341 {
342 	return REG_FIELD_PREP(BLIT_CCTL_DST_MOCS_MASK, gt->mocs.uc_index) |
343 		REG_FIELD_PREP(BLIT_CCTL_SRC_MOCS_MASK, gt->mocs.uc_index);
344 }
345 
346 static const struct xe_rtp_table_sr lrc_setup = XE_RTP_TABLE_SR(
347 	/*
348 	 * Some blitter commands do not have a field for MOCS, those
349 	 * commands will use MOCS index pointed by BLIT_CCTL.
350 	 * BLIT_CCTL registers are needed to be programmed to un-cached.
351 	 */
352 	{ XE_RTP_NAME("BLIT_CCTL_default_MOCS"),
353 	  XE_RTP_RULES(GRAPHICS_VERSION_RANGE(1200, 1274),
354 		       ENGINE_CLASS(COPY)),
355 	  XE_RTP_ACTIONS(FIELD_SET_FUNC(BLIT_CCTL(0),
356 					BLIT_CCTL_DST_MOCS_MASK |
357 					BLIT_CCTL_SRC_MOCS_MASK,
358 					blit_cctl_val,
359 					XE_RTP_ACTION_FLAG(ENGINE_BASE)))
360 	},
361 	/* Disable WMTP if HW doesn't support it */
362 	{ XE_RTP_NAME("DISABLE_WMTP_ON_UNSUPPORTED_HW"),
363 	  XE_RTP_RULES(FUNC(xe_rtp_cfeg_wmtp_disabled)),
364 	  XE_RTP_ACTIONS(FIELD_SET(CS_CHICKEN1(0),
365 				   PREEMPT_GPGPU_LEVEL_MASK,
366 				   PREEMPT_GPGPU_THREAD_GROUP_LEVEL)),
367 	  XE_RTP_ENTRY_FLAG(FOREACH_ENGINE)
368 	},
369 );
370 
371 static void
hw_engine_setup_default_lrc_state(struct xe_hw_engine * hwe)372 hw_engine_setup_default_lrc_state(struct xe_hw_engine *hwe)
373 {
374 	struct xe_rtp_process_ctx ctx = XE_RTP_PROCESS_CTX_INITIALIZER(hwe);
375 
376 	xe_rtp_process_to_sr(&ctx, &lrc_setup, &hwe->reg_lrc, true);
377 }
378 
xe_hw_engine_setup_reg_lrc(struct xe_hw_engine * hwe)379 void xe_hw_engine_setup_reg_lrc(struct xe_hw_engine *hwe)
380 {
381 	struct xe_gt *gt = hwe->gt;
382 	struct xe_device *xe = gt_to_xe(gt);
383 
384 	xe_reg_sr_init(&hwe->reg_lrc, hwe->name, xe);
385 	xe_wa_process_lrc(hwe);
386 	hw_engine_setup_default_lrc_state(hwe);
387 	xe_tuning_process_lrc(hwe);
388 }
389 
390 /*
391  * RING_CMD_CCTL specifies the default MOCS entry that will be
392  * used by the command streamer when executing commands that
393  * don't have a way to explicitly specify a MOCS setting.
394  * The default should usually reference whichever MOCS entry
395  * corresponds to uncached behavior, although use of a WB cached
396  * entry is recommended by the spec in certain circumstances on
397  * specific platforms.
398  * Bspec: 72161
399  */
ring_cmd_cctl_val(struct xe_gt * gt,struct xe_hw_engine * hwe)400 static u32 ring_cmd_cctl_val(struct xe_gt *gt, struct xe_hw_engine *hwe)
401 {
402 	struct xe_device *xe = gt_to_xe(gt);
403 	u8 mocs_read_idx = gt->mocs.uc_index;
404 
405 	if (hwe->class == XE_ENGINE_CLASS_COMPUTE && IS_DGFX(xe) &&
406 	    (GRAPHICS_VER(xe) >= 20 || xe->info.platform == XE_PVC))
407 		mocs_read_idx = gt->mocs.wb_index;
408 
409 	return REG_FIELD_PREP(CMD_CCTL_WRITE_OVERRIDE_MASK, gt->mocs.uc_index) |
410 		REG_FIELD_PREP(CMD_CCTL_READ_OVERRIDE_MASK, mocs_read_idx);
411 }
412 
413 static const struct xe_rtp_table_sr engine_sr = XE_RTP_TABLE_SR(
414 	{ XE_RTP_NAME("RING_CMD_CCTL_default_MOCS"),
415 	  XE_RTP_RULES(FUNC(xe_rtp_match_always)),
416 	  XE_RTP_ACTIONS(FIELD_SET_FUNC(RING_CMD_CCTL(0),
417 					CMD_CCTL_WRITE_OVERRIDE_MASK |
418 					CMD_CCTL_READ_OVERRIDE_MASK,
419 					ring_cmd_cctl_val,
420 					XE_RTP_ACTION_FLAG(ENGINE_BASE)))
421 	},
422 	{ XE_RTP_NAME("Disable HW status page updates for interrupts"),
423 	  XE_RTP_RULES(FUNC(xe_rtp_match_always)),
424 	  XE_RTP_ACTIONS(SET(RING_HWSTAM(0), ~0x0,
425 			     XE_RTP_ACTION_FLAG(ENGINE_BASE)))
426 	},
427 	{ XE_RTP_NAME("Disable engine 'legacy' mode"),
428 	  XE_RTP_RULES(FUNC(xe_rtp_match_always)),
429 	  XE_RTP_ACTIONS(SET(GFX_MODE(0), GFX_DISABLE_LEGACY_MODE,
430 			     XE_RTP_ACTION_FLAG(ENGINE_BASE)))
431 	},
432 	/*
433 	 * To allow the GSC engine to go idle on MTL we need to enable
434 	 * idle messaging and set the hysteresis value (we use 0xA=5us
435 	 * as recommended in spec). On platforms after MTL this is
436 	 * enabled by default.
437 	 */
438 	{ XE_RTP_NAME("MTL GSCCS IDLE MSG enable"),
439 	  XE_RTP_RULES(MEDIA_VERSION(1300), ENGINE_CLASS(OTHER)),
440 	  XE_RTP_ACTIONS(CLR(RING_PSMI_CTL(0),
441 			     IDLE_MSG_DISABLE,
442 			     XE_RTP_ACTION_FLAG(ENGINE_BASE)),
443 			 FIELD_SET(RING_PWRCTX_MAXCNT(0),
444 				   IDLE_WAIT_TIME,
445 				   0xA,
446 				   XE_RTP_ACTION_FLAG(ENGINE_BASE)))
447 	},
448 	/* Enable Priority Mem Read */
449 	{ XE_RTP_NAME("Priority_Mem_Read"),
450 	  XE_RTP_RULES(GRAPHICS_VERSION_RANGE(2001, XE_RTP_END_VERSION_UNDEFINED)),
451 	  XE_RTP_ACTIONS(SET(CSFE_CHICKEN1(0), CS_PRIORITY_MEM_READ,
452 			     XE_RTP_ACTION_FLAG(ENGINE_BASE)))
453 	},
454 	{ XE_RTP_NAME("Enable CCS Engine(s)"),
455 	  XE_RTP_RULES(GRAPHICS_VERSION_RANGE(1255, XE_RTP_END_VERSION_UNDEFINED),
456 		       FUNC(xe_rtp_match_first_render_or_compute)),
457 	  XE_RTP_ACTIONS(SET(RCU_MODE, RCU_MODE_CCS_ENABLE))
458 	},
459 	/* Use Fixed slice CCS mode */
460 	{ XE_RTP_NAME("RCU_MODE_FIXED_SLICE_CCS_MODE"),
461 	  XE_RTP_RULES(FUNC(xe_hw_engine_match_fixed_cslice_mode)),
462 	  XE_RTP_ACTIONS(FIELD_SET(RCU_MODE, RCU_MODE_FIXED_SLICE_CCS_MODE,
463 				   RCU_MODE_FIXED_SLICE_CCS_MODE))
464 	},
465 	{ XE_RTP_NAME("Enable MSI-X interrupt support"),
466 	  XE_RTP_RULES(FUNC(xe_rtp_match_has_msix)),
467 	  XE_RTP_ACTIONS(SET(GFX_MODE(0), GFX_MSIX_INTERRUPT_ENABLE,
468 			     XE_RTP_ACTION_FLAG(ENGINE_BASE)))
469 	},
470 );
471 
472 static void
hw_engine_setup_default_state(struct xe_hw_engine * hwe)473 hw_engine_setup_default_state(struct xe_hw_engine *hwe)
474 {
475 	struct xe_rtp_process_ctx ctx = XE_RTP_PROCESS_CTX_INITIALIZER(hwe);
476 
477 	xe_rtp_process_to_sr(&ctx, &engine_sr, &hwe->reg_sr, false);
478 }
479 
find_engine_info(enum xe_engine_class class,int instance)480 static const struct engine_info *find_engine_info(enum xe_engine_class class, int instance)
481 {
482 	const struct engine_info *info;
483 	enum xe_hw_engine_id id;
484 
485 	for (id = 0; id < XE_NUM_HW_ENGINES; ++id) {
486 		info = &engine_infos[id];
487 		if (info->class == class && info->instance == instance)
488 			return info;
489 	}
490 
491 	return NULL;
492 }
493 
get_msix_irq_offset(struct xe_gt * gt,enum xe_engine_class class)494 static u16 get_msix_irq_offset(struct xe_gt *gt, enum xe_engine_class class)
495 {
496 	/* For MSI-X, hw engines report to offset of engine instance zero */
497 	const struct engine_info *info = find_engine_info(class, 0);
498 
499 	xe_gt_assert(gt, info);
500 
501 	return info ? info->irq_offset : 0;
502 }
503 
hw_engine_init_early(struct xe_gt * gt,struct xe_hw_engine * hwe,enum xe_hw_engine_id id)504 static void hw_engine_init_early(struct xe_gt *gt, struct xe_hw_engine *hwe,
505 				 enum xe_hw_engine_id id)
506 {
507 	const struct engine_info *info;
508 
509 	if (WARN_ON(id >= ARRAY_SIZE(engine_infos) || !engine_infos[id].name))
510 		return;
511 
512 	if (!(gt->info.engine_mask & BIT(id)))
513 		return;
514 
515 	info = &engine_infos[id];
516 
517 	xe_gt_assert(gt, !hwe->gt);
518 
519 	hwe->gt = gt;
520 	hwe->class = info->class;
521 	hwe->instance = info->instance;
522 	hwe->mmio_base = info->mmio_base;
523 	if (xe_device_has_msix(gt_to_xe(gt))) {
524 		hwe->irq_offset = get_msix_irq_offset(gt, info->class);
525 		hwe->irq_page = info->instance;
526 
527 	} else {
528 		hwe->irq_offset = info->irq_offset;
529 		hwe->irq_page = 0;
530 	}
531 	hwe->domain = info->domain;
532 	hwe->name = info->name;
533 	hwe->fence_irq = &gt->fence_irq[info->class];
534 	hwe->engine_id = id;
535 
536 	hwe->eclass = &gt->eclass[hwe->class];
537 	if (!hwe->eclass->sched_props.job_timeout_ms) {
538 		hwe->eclass->sched_props.job_timeout_ms = 5 * 1000;
539 		hwe->eclass->sched_props.job_timeout_min = XE_HW_ENGINE_JOB_TIMEOUT_MIN;
540 		hwe->eclass->sched_props.job_timeout_max = XE_HW_ENGINE_JOB_TIMEOUT_MAX;
541 		hwe->eclass->sched_props.timeslice_us = 1 * 1000;
542 		hwe->eclass->sched_props.timeslice_min = XE_HW_ENGINE_TIMESLICE_MIN;
543 		hwe->eclass->sched_props.timeslice_max = XE_HW_ENGINE_TIMESLICE_MAX;
544 		hwe->eclass->sched_props.preempt_timeout_us = XE_HW_ENGINE_PREEMPT_TIMEOUT;
545 		hwe->eclass->sched_props.preempt_timeout_min = XE_HW_ENGINE_PREEMPT_TIMEOUT_MIN;
546 		hwe->eclass->sched_props.preempt_timeout_max = XE_HW_ENGINE_PREEMPT_TIMEOUT_MAX;
547 
548 		/*
549 		 * The GSC engine can accept submissions while the GSC shim is
550 		 * being reset, during which time the submission is stalled. In
551 		 * the worst case, the shim reset can take up to the maximum GSC
552 		 * command execution time (250ms), so the request start can be
553 		 * delayed by that much; the request itself can take that long
554 		 * without being preemptible, which means worst case it can
555 		 * theoretically take up to 500ms for a preemption to go through
556 		 * on the GSC engine. Adding to that an extra 100ms as a safety
557 		 * margin, we get a minimum recommended timeout of 600ms.
558 		 * The preempt_timeout value can't be tuned for OTHER_CLASS
559 		 * because the class is reserved for kernel usage, so we just
560 		 * need to make sure that the starting value is above that
561 		 * threshold; since our default value (640ms) is greater than
562 		 * 600ms, the only way we can go below is via a kconfig setting.
563 		 * If that happens, log it in dmesg and update the value.
564 		 */
565 		if (hwe->class == XE_ENGINE_CLASS_OTHER) {
566 			const u32 min_preempt_timeout = 600 * 1000;
567 			if (hwe->eclass->sched_props.preempt_timeout_us < min_preempt_timeout) {
568 				hwe->eclass->sched_props.preempt_timeout_us = min_preempt_timeout;
569 				xe_gt_notice(gt, "Increasing preempt_timeout for GSC to 600ms\n");
570 			}
571 		}
572 
573 		/* Record default props */
574 		hwe->eclass->defaults = hwe->eclass->sched_props;
575 	}
576 
577 	xe_reg_sr_init(&hwe->reg_sr, hwe->name, gt_to_xe(gt));
578 	xe_tuning_process_engine(hwe);
579 	xe_wa_process_engine(hwe);
580 	hw_engine_setup_default_state(hwe);
581 
582 	xe_reg_sr_init(&hwe->reg_whitelist, hwe->name, gt_to_xe(gt));
583 	xe_reg_sr_init(&hwe->oa_whitelist, hwe->name, gt_to_xe(gt));
584 	xe_reg_sr_init(&hwe->oa_sr, hwe->name, gt_to_xe(gt));
585 	xe_reg_whitelist_process_engine(hwe);
586 }
587 
idledly_floor_ticks(u32 idledly_ns,u32 idledly_units_ps)588 static u32 idledly_floor_ticks(u32 idledly_ns, u32 idledly_units_ps)
589 {
590 	return DIV_ROUND_DOWN_ULL((u64)idledly_ns * 1000, idledly_units_ps);
591 }
592 
adjust_idledly(struct xe_hw_engine * hwe)593 static void adjust_idledly(struct xe_hw_engine *hwe)
594 {
595 	struct xe_gt *gt = hwe->gt;
596 	u32 idledly, idledly_hw, idledly_reg_val, maxcnt;
597 	u32 idledly_units_ps = 8 * gt->info.timestamp_base;
598 	u32 maxcnt_units_ns = 640;
599 	bool inhibit_switch = false;
600 	bool wa_applied = false;
601 	bool clamped_below_maxcnt = false;
602 
603 	if ((!IS_SRIOV_VF(gt_to_xe(gt)) && XE_GT_WA(gt, 16023105232)) ||
604 	    XE_GT_WA(gt, 14025941587)) {
605 		u32 mincnt_idledly_ns = 5000;
606 
607 		/* xe_gt_clock_init() warns and zeroes timestamp_base on unknown crystal clock. */
608 		if (!idledly_units_ps)
609 			return;
610 
611 		idledly_reg_val = xe_mmio_read32(&gt->mmio, RING_IDLEDLY(hwe->mmio_base));
612 		maxcnt = xe_mmio_read32(&gt->mmio, RING_PWRCTX_MAXCNT(hwe->mmio_base));
613 
614 		inhibit_switch = idledly_reg_val & INHIBIT_SWITCH_UNTIL_PREEMPTED;
615 		idledly = REG_FIELD_GET(IDLE_DELAY, idledly_reg_val);
616 		idledly = DIV_ROUND_CLOSEST_ULL((u64)idledly * idledly_units_ps, 1000);
617 		idledly_hw = idledly;
618 		maxcnt = REG_FIELD_GET(IDLE_WAIT_TIME, maxcnt);
619 		maxcnt *= maxcnt_units_ns;
620 
621 		/*
622 		 * Wa_14025941587 is applied before Wa_16023105232, which takes
623 		 * priority if the two ever conflict (not expected in practice).
624 		 */
625 		if (XE_GT_WA(gt, 14025941587) &&
626 		    idledly < mincnt_idledly_ns) {
627 			idledly = mincnt_idledly_ns;
628 			wa_applied = true;
629 		}
630 
631 		if (XE_GT_WA(gt, 16023105232)) {
632 			/* Clear the inhibit switch without disturbing a valid delay. */
633 			if (inhibit_switch) {
634 				idledly_reg_val &= ~INHIBIT_SWITCH_UNTIL_PREEMPTED;
635 				wa_applied = true;
636 			}
637 
638 			/* Warn only on the value read from hardware. */
639 			xe_gt_WARN_ON(gt, idledly_hw >= maxcnt);
640 
641 			if (idledly >= maxcnt) {
642 				/* maxcnt may be 0 if IDLE_WAIT_TIME is unprogrammed. */
643 				idledly = maxcnt ? maxcnt - 1 : 0;
644 				clamped_below_maxcnt = true;
645 				wa_applied = true;
646 			}
647 		}
648 
649 		if (wa_applied) {
650 			u32 idledly_ticks;
651 
652 			/*
653 			 * Wa_16023105232 requires idledly < maxcnt, so floor
654 			 * that clamp; otherwise round up to guarantee the
655 			 * Wa_14025941587 minimum survives tick quantization.
656 			 */
657 			if (clamped_below_maxcnt)
658 				idledly_ticks = idledly_floor_ticks(idledly, idledly_units_ps);
659 			else
660 				idledly_ticks = DIV_ROUND_UP_ULL((u64)idledly * 1000,
661 								 idledly_units_ps);
662 
663 			/*
664 			 * Tick quantization can still push the rounded-up value
665 			 * to/above maxcnt; re-floor here so Wa_16023105232 keeps
666 			 * priority even in that case.
667 			 */
668 			if (!clamped_below_maxcnt && XE_GT_WA(gt, 16023105232) &&
669 			    (u64)idledly_ticks * idledly_units_ps >= (u64)maxcnt * 1000) {
670 				xe_gt_dbg(gt, "idledly %s: %u ticks would exceed maxcnt=%u, so flooring\n",
671 					  hwe->name, idledly_ticks, maxcnt);
672 				idledly = maxcnt ? maxcnt - 1 : 0;
673 				idledly_ticks = idledly_floor_ticks(idledly, idledly_units_ps);
674 			}
675 
676 			idledly_reg_val &= ~IDLE_DELAY;
677 			idledly_reg_val |= REG_FIELD_PREP(IDLE_DELAY, idledly_ticks);
678 			xe_gt_dbg(gt, "idledly %s: set %u max=%u inh=%u ts=%u\n",
679 				  hwe->name, idledly, maxcnt,
680 				  !!inhibit_switch, gt->info.timestamp_base);
681 			xe_mmio_write32(&gt->mmio,
682 					RING_IDLEDLY(hwe->mmio_base),
683 					idledly_reg_val);
684 		}
685 	}
686 }
687 
hw_engine_init(struct xe_gt * gt,struct xe_hw_engine * hwe,enum xe_hw_engine_id id)688 static int hw_engine_init(struct xe_gt *gt, struct xe_hw_engine *hwe,
689 			  enum xe_hw_engine_id id)
690 {
691 	struct xe_device *xe = gt_to_xe(gt);
692 	struct xe_tile *tile = gt_to_tile(gt);
693 	int err;
694 
695 	xe_gt_assert(gt, id < ARRAY_SIZE(engine_infos) && engine_infos[id].name);
696 	xe_gt_assert(gt, gt->info.engine_mask & BIT(id));
697 
698 	xe_reg_sr_apply_mmio(&hwe->reg_sr, gt);
699 
700 	hwe->hwsp = xe_managed_bo_create_pin_map(xe, tile, SZ_4K,
701 						 XE_BO_FLAG_VRAM_IF_DGFX(tile) |
702 						 XE_BO_FLAG_GGTT |
703 						 XE_BO_FLAG_GGTT_INVALIDATE);
704 	if (IS_ERR(hwe->hwsp)) {
705 		err = PTR_ERR(hwe->hwsp);
706 		goto err_name;
707 	}
708 
709 	if (!xe_device_uc_enabled(xe)) {
710 		hwe->exl_port = xe_execlist_port_create(xe, hwe);
711 		if (IS_ERR(hwe->exl_port)) {
712 			err = PTR_ERR(hwe->exl_port);
713 			goto err_name;
714 		}
715 	} else {
716 		/* GSCCS has a special interrupt for reset */
717 		if (hwe->class == XE_ENGINE_CLASS_OTHER)
718 			hwe->irq_handler = xe_gsc_hwe_irq_handler;
719 
720 		if (!IS_SRIOV_VF(xe))
721 			xe_hw_engine_enable_ring(hwe);
722 	}
723 
724 	/* Ensure IDLEDLY is lower than MAXCNT */
725 	adjust_idledly(hwe);
726 
727 	return devm_add_action_or_reset(xe->drm.dev, hw_engine_fini, hwe);
728 
729 err_name:
730 	hwe->name = NULL;
731 
732 	return err;
733 }
734 
hw_engine_setup_logical_and_paging_mapping(struct xe_gt * gt)735 static int hw_engine_setup_logical_and_paging_mapping(struct xe_gt *gt)
736 {
737 	struct xe_device *xe = gt_to_xe(gt);
738 	unsigned int num_copy_engines = 0, num_paging_engines = 0;
739 	unsigned int reserved_logical_bcs_start;
740 	struct xe_hw_engine *hwe;
741 	enum xe_hw_engine_id id;
742 	int class;
743 
744 	for_each_hw_engine(hwe, gt, id)
745 		if (hwe->class == XE_ENGINE_CLASS_COPY)
746 			num_copy_engines++;
747 
748 	if (num_copy_engines && xe->info.has_usm)
749 		num_paging_engines = 1;
750 
751 	if (IS_SRIOV_VF(xe)) {
752 		u32 vf_num_paging_engines;
753 
754 		/*
755 		 * PF could in theory reserve multiple paging engines, which
756 		 * internally the submission/scheduling backend can load balance
757 		 * from. Not something we currently expect, but we are at the
758 		 * mercy of the PF, so we just need try our best to mirror the
759 		 * paging configuration.
760 		 */
761 		vf_num_paging_engines = xe_gt_sriov_vf_paging_engines(gt);
762 		if (vf_num_paging_engines) {
763 			/* This should only be non-zero on NVL-S+ */
764 			if (xe_gt_WARN_ON(gt, xe->info.platform < XE_NOVALAKE_S))
765 				return -EINVAL;
766 
767 			num_paging_engines = vf_num_paging_engines;
768 		}
769 	}
770 
771 	if (xe_gt_WARN_ON(gt, num_paging_engines > num_copy_engines))
772 		return -EINVAL;
773 
774 	/*
775 	 * On PF, we just reserve the highest BCS instance for USM.
776 	 *
777 	 * Note: This is now a requirement going forward. The PF must ALWAYS
778 	 * reserve BCS instances in top-down order, that way the VF has a chance
779 	 * of discovering the physical BCS instance mappings for paging engines,
780 	 * in conjunction with vf_num_paging_engines. In some places we might
781 	 * only have the physical instance, and from hw pov there is no such
782 	 * thing as a paging engine. For example, the page fault descriptor,
783 	 * which comes directly from the hw, will use the physical engine
784 	 * instance.
785 	 */
786 	reserved_logical_bcs_start = num_copy_engines - num_paging_engines;
787 
788 	/* FIXME: Doing a simple logical mapping that works for most hardware */
789 	for (class = 0; class < XE_ENGINE_CLASS_MAX; ++class) {
790 		int logical_instance = 0;
791 
792 		for_each_hw_engine(hwe, gt, id) {
793 			if (hwe->class == class) {
794 				hwe->logical_instance = logical_instance++;
795 
796 				if (class == XE_ENGINE_CLASS_COPY &&
797 				    hwe->logical_instance >=
798 					    reserved_logical_bcs_start) {
799 					if (!gt->usm.paging_hwe0)
800 						gt->usm.paging_hwe0 = hwe;
801 					gt->usm.paging_logical_mask |=
802 						BIT(hwe->logical_instance);
803 				}
804 			}
805 		}
806 	}
807 
808 	return 0;
809 }
810 
read_media_fuses(struct xe_gt * gt)811 static void read_media_fuses(struct xe_gt *gt)
812 {
813 	struct xe_device *xe = gt_to_xe(gt);
814 	u32 media_fuse;
815 	u16 vdbox_mask;
816 	u16 vebox_mask;
817 	int i, j;
818 
819 	xe_force_wake_assert_held(gt_to_fw(gt), XE_FW_GT);
820 
821 	media_fuse = xe_mmio_read32(&gt->mmio, GT_VEBOX_VDBOX_DISABLE);
822 
823 	/*
824 	 * Pre-Xe_HP platforms had register bits representing absent engines,
825 	 * whereas Xe_HP and beyond have bits representing present engines.
826 	 * Invert the polarity on old platforms so that we can use common
827 	 * handling below.
828 	 */
829 	if (GRAPHICS_VERx100(xe) < 1250)
830 		media_fuse = ~media_fuse;
831 
832 	vdbox_mask = REG_FIELD_GET(GT_VDBOX_DISABLE_MASK, media_fuse);
833 	vebox_mask = REG_FIELD_GET(GT_VEBOX_DISABLE_MASK, media_fuse);
834 
835 	for (i = XE_HW_ENGINE_VCS0, j = 0; i <= XE_HW_ENGINE_VCS7; ++i, ++j) {
836 		if (!(gt->info.engine_mask & BIT(i)))
837 			continue;
838 
839 		if (!(BIT(j) & vdbox_mask)) {
840 			gt->info.engine_mask &= ~BIT(i);
841 			xe_gt_info(gt, "vcs%u fused off\n", j);
842 		}
843 	}
844 
845 	for (i = XE_HW_ENGINE_VECS0, j = 0; i <= XE_HW_ENGINE_VECS3; ++i, ++j) {
846 		if (!(gt->info.engine_mask & BIT(i)))
847 			continue;
848 
849 		if (!(BIT(j) & vebox_mask)) {
850 			gt->info.engine_mask &= ~BIT(i);
851 			xe_gt_info(gt, "vecs%u fused off\n", j);
852 		}
853 	}
854 }
855 
infer_svccopy_from_meml3(struct xe_gt * gt)856 static u32 infer_svccopy_from_meml3(struct xe_gt *gt)
857 {
858 	u32 meml3 = REG_FIELD_GET(MEML3_EN_MASK,
859 				  xe_mmio_read32(&gt->mmio, MIRROR_FUSE3));
860 	u32 svccopy_mask = 0;
861 
862 	/*
863 	 * Each of the four meml3 bits determines the fusing of two service
864 	 * copy engines.
865 	 */
866 	for (int i = 0; i < 4; i++)
867 		svccopy_mask |= (meml3 & BIT(i)) ? 0b11 << 2 * i : 0;
868 
869 	return svccopy_mask;
870 }
871 
read_svccopy_fuses(struct xe_gt * gt)872 static u32 read_svccopy_fuses(struct xe_gt *gt)
873 {
874 	return REG_FIELD_GET(FUSE_SERVICE_COPY_ENABLE_MASK,
875 			     xe_mmio_read32(&gt->mmio, SERVICE_COPY_ENABLE));
876 }
877 
read_copy_fuses(struct xe_gt * gt)878 static void read_copy_fuses(struct xe_gt *gt)
879 {
880 	struct xe_device *xe = gt_to_xe(gt);
881 	u32 bcs_mask;
882 
883 	xe_force_wake_assert_held(gt_to_fw(gt), XE_FW_GT);
884 
885 	if (GRAPHICS_VER(xe) >= 35)
886 		bcs_mask = read_svccopy_fuses(gt);
887 	else if (GRAPHICS_VERx100(xe) == 1260)
888 		bcs_mask = infer_svccopy_from_meml3(gt);
889 	else
890 		return;
891 
892 	/* Only BCS1-BCS8 may be fused off */
893 	bcs_mask <<= XE_HW_ENGINE_BCS1;
894 	for (int i = XE_HW_ENGINE_BCS1; i <= XE_HW_ENGINE_BCS8; ++i) {
895 		if (!(gt->info.engine_mask & BIT(i)))
896 			continue;
897 
898 		if (!(bcs_mask & BIT(i))) {
899 			gt->info.engine_mask &= ~BIT(i);
900 			xe_gt_info(gt, "bcs%u fused off\n",
901 				   i - XE_HW_ENGINE_BCS0);
902 		}
903 	}
904 }
905 
read_compute_fuses_from_dss(struct xe_gt * gt)906 static void read_compute_fuses_from_dss(struct xe_gt *gt)
907 {
908 	/*
909 	 * CCS fusing based on DSS masks only applies to platforms that can
910 	 * have more than one CCS.
911 	 */
912 	if (hweight64(gt->info.engine_mask &
913 		      GENMASK_ULL(XE_HW_ENGINE_CCS3, XE_HW_ENGINE_CCS0)) <= 1)
914 		return;
915 
916 	/*
917 	 * CCS availability on Xe_HP is inferred from the presence of DSS in
918 	 * each quadrant.
919 	 */
920 	for (int i = XE_HW_ENGINE_CCS0, j = 0; i <= XE_HW_ENGINE_CCS3; ++i, ++j) {
921 		if (!(gt->info.engine_mask & BIT(i)))
922 			continue;
923 
924 		if (!xe_gt_topology_has_dss_in_quadrant(gt, j)) {
925 			gt->info.engine_mask &= ~BIT(i);
926 			xe_gt_info(gt, "ccs%u fused off\n", j);
927 		}
928 	}
929 }
930 
read_compute_fuses_from_reg(struct xe_gt * gt)931 static void read_compute_fuses_from_reg(struct xe_gt *gt)
932 {
933 	u32 ccs_mask;
934 
935 	ccs_mask = xe_mmio_read32(&gt->mmio, XEHP_FUSE4);
936 	ccs_mask = REG_FIELD_GET(CCS_EN_MASK, ccs_mask);
937 
938 	for (int i = XE_HW_ENGINE_CCS0, j = 0; i <= XE_HW_ENGINE_CCS3; ++i, ++j) {
939 		if (!(gt->info.engine_mask & BIT(i)))
940 			continue;
941 
942 		if ((ccs_mask & BIT(j)) == 0) {
943 			gt->info.engine_mask &= ~BIT(i);
944 			xe_gt_info(gt, "ccs%u fused off\n", j);
945 		}
946 	}
947 }
948 
read_compute_fuses(struct xe_gt * gt)949 static void read_compute_fuses(struct xe_gt *gt)
950 {
951 	if (GRAPHICS_VER(gt_to_xe(gt)) >= 20)
952 		read_compute_fuses_from_reg(gt);
953 	else
954 		read_compute_fuses_from_dss(gt);
955 }
956 
check_gsc_availability(struct xe_gt * gt)957 static void check_gsc_availability(struct xe_gt *gt)
958 {
959 	if (!(gt->info.engine_mask & BIT(XE_HW_ENGINE_GSCCS0)))
960 		return;
961 
962 	/*
963 	 * The GSCCS is only used to communicate with the GSC FW, so if we don't
964 	 * have the FW there is nothing we need the engine for and can therefore
965 	 * skip its initialization.
966 	 */
967 	if (!xe_uc_fw_is_available(&gt->uc.gsc.fw)) {
968 		gt->info.engine_mask &= ~BIT(XE_HW_ENGINE_GSCCS0);
969 
970 		/* interrupts where previously enabled, so turn them off */
971 		xe_mmio_write32(&gt->mmio, GUNIT_GSC_INTR_ENABLE, 0);
972 		xe_mmio_write32(&gt->mmio, GUNIT_GSC_INTR_MASK, ~0);
973 
974 		xe_gt_dbg(gt, "GSC FW not used, disabling gsccs\n");
975 	}
976 }
977 
check_sw_disable(struct xe_gt * gt)978 static void check_sw_disable(struct xe_gt *gt)
979 {
980 	struct xe_device *xe = gt_to_xe(gt);
981 	u64 sw_allowed = xe_configfs_get_engines_allowed(to_pci_dev(xe->drm.dev));
982 	enum xe_hw_engine_id id;
983 
984 	for (id = 0; id < XE_NUM_HW_ENGINES; ++id) {
985 		if (!(gt->info.engine_mask & BIT(id)))
986 			continue;
987 
988 		if (!(sw_allowed & BIT(id))) {
989 			gt->info.engine_mask &= ~BIT(id);
990 			xe_gt_info(gt, "%s disabled via configfs\n",
991 				   engine_infos[id].name);
992 		}
993 	}
994 }
995 
xe_hw_engines_init_early(struct xe_gt * gt)996 int xe_hw_engines_init_early(struct xe_gt *gt)
997 {
998 	int i;
999 
1000 	read_media_fuses(gt);
1001 	read_copy_fuses(gt);
1002 	read_compute_fuses(gt);
1003 	check_gsc_availability(gt);
1004 	check_sw_disable(gt);
1005 
1006 	BUILD_BUG_ON(XE_HW_ENGINE_PREEMPT_TIMEOUT < XE_HW_ENGINE_PREEMPT_TIMEOUT_MIN);
1007 	BUILD_BUG_ON(XE_HW_ENGINE_PREEMPT_TIMEOUT > XE_HW_ENGINE_PREEMPT_TIMEOUT_MAX);
1008 
1009 	for (i = 0; i < ARRAY_SIZE(gt->hw_engines); i++)
1010 		hw_engine_init_early(gt, &gt->hw_engines[i], i);
1011 
1012 	return 0;
1013 }
1014 
xe_hw_engines_init(struct xe_gt * gt)1015 int xe_hw_engines_init(struct xe_gt *gt)
1016 {
1017 	int err;
1018 	struct xe_hw_engine *hwe;
1019 	enum xe_hw_engine_id id;
1020 
1021 	for_each_hw_engine(hwe, gt, id) {
1022 		err = hw_engine_init(gt, hwe, id);
1023 		if (err)
1024 			return err;
1025 	}
1026 
1027 	err = hw_engine_setup_logical_and_paging_mapping(gt);
1028 	if (err)
1029 		return err;
1030 
1031 	err = xe_hw_engine_setup_groups(gt);
1032 	if (err)
1033 		return err;
1034 
1035 	return 0;
1036 }
1037 
xe_hw_engine_handle_irq(struct xe_hw_engine * hwe,u16 intr_vec)1038 void xe_hw_engine_handle_irq(struct xe_hw_engine *hwe, u16 intr_vec)
1039 {
1040 	wake_up_all(&gt_to_xe(hwe->gt)->ufence_wq);
1041 
1042 	if (hwe->irq_handler)
1043 		hwe->irq_handler(hwe, intr_vec);
1044 
1045 	if (intr_vec & GT_MI_USER_INTERRUPT)
1046 		xe_hw_fence_irq_run(hwe->fence_irq);
1047 }
1048 
1049 /**
1050  * xe_hw_engine_snapshot_capture - Take a quick snapshot of the HW Engine.
1051  * @hwe: Xe HW Engine.
1052  * @q: The exec queue object.
1053  *
1054  * This can be printed out in a later stage like during dev_coredump
1055  * analysis.
1056  *
1057  * Returns: a Xe HW Engine snapshot object that must be freed by the
1058  * caller, using `xe_hw_engine_snapshot_free`.
1059  */
1060 struct xe_hw_engine_snapshot *
xe_hw_engine_snapshot_capture(struct xe_hw_engine * hwe,struct xe_exec_queue * q)1061 xe_hw_engine_snapshot_capture(struct xe_hw_engine *hwe, struct xe_exec_queue *q)
1062 {
1063 	struct xe_hw_engine_snapshot *snapshot;
1064 	struct __guc_capture_parsed_output *node;
1065 
1066 	if (!xe_hw_engine_is_valid(hwe))
1067 		return NULL;
1068 
1069 	snapshot = kzalloc_obj(*snapshot, GFP_ATOMIC);
1070 
1071 	if (!snapshot)
1072 		return NULL;
1073 
1074 	snapshot->name = kstrdup(hwe->name, GFP_ATOMIC);
1075 	snapshot->hwe = hwe;
1076 	snapshot->logical_instance = hwe->logical_instance;
1077 	snapshot->forcewake.domain = hwe->domain;
1078 	snapshot->forcewake.ref = xe_force_wake_ref(gt_to_fw(hwe->gt),
1079 						    hwe->domain);
1080 	snapshot->mmio_base = hwe->mmio_base;
1081 	snapshot->kernel_reserved = xe_hw_engine_is_reserved(hwe);
1082 
1083 	/* no more VF accessible data below this point */
1084 	if (IS_SRIOV_VF(gt_to_xe(hwe->gt)))
1085 		return snapshot;
1086 
1087 	if (q) {
1088 		/* If got guc capture, set source to GuC */
1089 		node = xe_guc_capture_get_matching_and_lock(q);
1090 		if (node) {
1091 			struct xe_device *xe = gt_to_xe(hwe->gt);
1092 			struct xe_devcoredump *coredump = &xe->devcoredump;
1093 
1094 			coredump->snapshot.matched_node = node;
1095 			xe_gt_dbg(hwe->gt, "Found and locked GuC-err-capture node");
1096 			return snapshot;
1097 		}
1098 	}
1099 
1100 	/* otherwise, do manual capture */
1101 	xe_engine_manual_capture(hwe, snapshot);
1102 	xe_gt_dbg(hwe->gt, "Proceeding with manual engine snapshot");
1103 
1104 	return snapshot;
1105 }
1106 
1107 /**
1108  * xe_hw_engine_snapshot_free - Free all allocated objects for a given snapshot.
1109  * @snapshot: Xe HW Engine snapshot object.
1110  *
1111  * This function free all the memory that needed to be allocated at capture
1112  * time.
1113  */
xe_hw_engine_snapshot_free(struct xe_hw_engine_snapshot * snapshot)1114 void xe_hw_engine_snapshot_free(struct xe_hw_engine_snapshot *snapshot)
1115 {
1116 	struct xe_gt *gt;
1117 	if (!snapshot)
1118 		return;
1119 
1120 	gt = snapshot->hwe->gt;
1121 	/*
1122 	 * xe_guc_capture_put_matched_nodes is called here and from
1123 	 * xe_devcoredump_snapshot_free, to cover the 2 calling paths
1124 	 * of hw_engines - debugfs and devcoredump free.
1125 	 */
1126 	xe_guc_capture_put_matched_nodes(&gt->uc.guc);
1127 
1128 	kfree(snapshot->name);
1129 	kfree(snapshot);
1130 }
1131 
1132 /**
1133  * xe_hw_engine_print - Xe HW Engine Print.
1134  * @hwe: Hardware Engine.
1135  * @p: drm_printer.
1136  *
1137  * This function quickly capture a snapshot and immediately print it out.
1138  */
xe_hw_engine_print(struct xe_hw_engine * hwe,struct drm_printer * p)1139 void xe_hw_engine_print(struct xe_hw_engine *hwe, struct drm_printer *p)
1140 {
1141 	struct xe_hw_engine_snapshot *snapshot;
1142 
1143 	snapshot = xe_hw_engine_snapshot_capture(hwe, NULL);
1144 	xe_engine_snapshot_print(snapshot, p);
1145 	xe_hw_engine_snapshot_free(snapshot);
1146 }
1147 
xe_hw_engine_mask_per_class(struct xe_gt * gt,enum xe_engine_class engine_class)1148 u32 xe_hw_engine_mask_per_class(struct xe_gt *gt,
1149 				enum xe_engine_class engine_class)
1150 {
1151 	u32 mask = 0;
1152 	enum xe_hw_engine_id id;
1153 
1154 	for (id = 0; id < XE_NUM_HW_ENGINES; ++id) {
1155 		if (engine_infos[id].class == engine_class &&
1156 		    gt->info.engine_mask & BIT(id))
1157 			mask |= BIT(engine_infos[id].instance);
1158 	}
1159 	return mask;
1160 }
1161 
xe_hw_engine_is_reserved(struct xe_hw_engine * hwe)1162 bool xe_hw_engine_is_reserved(struct xe_hw_engine *hwe)
1163 {
1164 	struct xe_gt *gt = hwe->gt;
1165 	struct xe_device *xe = gt_to_xe(gt);
1166 
1167 	if (xe_device_is_admin_only(xe))
1168 		return true;
1169 
1170 	if (hwe->class == XE_ENGINE_CLASS_OTHER)
1171 		return true;
1172 
1173 	/* Check for engines disabled by ccs_mode setting */
1174 	if (xe_gt_ccs_mode_enabled(gt) &&
1175 	    hwe->class == XE_ENGINE_CLASS_COMPUTE &&
1176 	    hwe->logical_instance >= gt->ccs_mode)
1177 		return true;
1178 
1179 	return xe_gt_is_usm_hwe(gt, hwe);
1180 }
1181 
xe_hw_engine_class_to_str(enum xe_engine_class class)1182 const char *xe_hw_engine_class_to_str(enum xe_engine_class class)
1183 {
1184 	switch (class) {
1185 	case XE_ENGINE_CLASS_RENDER:
1186 		return "rcs";
1187 	case XE_ENGINE_CLASS_VIDEO_DECODE:
1188 		return "vcs";
1189 	case XE_ENGINE_CLASS_VIDEO_ENHANCE:
1190 		return "vecs";
1191 	case XE_ENGINE_CLASS_COPY:
1192 		return "bcs";
1193 	case XE_ENGINE_CLASS_OTHER:
1194 		return "other";
1195 	case XE_ENGINE_CLASS_COMPUTE:
1196 		return "ccs";
1197 	case XE_ENGINE_CLASS_MAX:
1198 		break;
1199 	}
1200 
1201 	return NULL;
1202 }
1203 
xe_hw_engine_read_timestamp(struct xe_hw_engine * hwe)1204 u64 xe_hw_engine_read_timestamp(struct xe_hw_engine *hwe)
1205 {
1206 	return xe_mmio_read64_2x32(&hwe->gt->mmio, RING_TIMESTAMP(hwe->mmio_base));
1207 }
1208 
xe_hw_engine_to_fw_domain(struct xe_hw_engine * hwe)1209 enum xe_force_wake_domains xe_hw_engine_to_fw_domain(struct xe_hw_engine *hwe)
1210 {
1211 	return engine_infos[hwe->engine_id].domain;
1212 }
1213 
1214 static const enum xe_engine_class user_to_xe_engine_class[] = {
1215 	[DRM_XE_ENGINE_CLASS_RENDER] = XE_ENGINE_CLASS_RENDER,
1216 	[DRM_XE_ENGINE_CLASS_COPY] = XE_ENGINE_CLASS_COPY,
1217 	[DRM_XE_ENGINE_CLASS_VIDEO_DECODE] = XE_ENGINE_CLASS_VIDEO_DECODE,
1218 	[DRM_XE_ENGINE_CLASS_VIDEO_ENHANCE] = XE_ENGINE_CLASS_VIDEO_ENHANCE,
1219 	[DRM_XE_ENGINE_CLASS_COMPUTE] = XE_ENGINE_CLASS_COMPUTE,
1220 };
1221 
1222 /**
1223  * xe_hw_engine_lookup() - Lookup hardware engine for class:instance
1224  * @xe: xe device
1225  * @eci: engine class and instance
1226  *
1227  * This function will find a hardware engine for given engine
1228  * class and instance.
1229  *
1230  * Return: If found xe_hw_engine pointer, NULL otherwise.
1231  */
1232 struct xe_hw_engine *
xe_hw_engine_lookup(struct xe_device * xe,struct drm_xe_engine_class_instance eci)1233 xe_hw_engine_lookup(struct xe_device *xe,
1234 		    struct drm_xe_engine_class_instance eci)
1235 {
1236 	struct xe_gt *gt = xe_device_get_gt(xe, eci.gt_id);
1237 	unsigned int idx;
1238 
1239 	if (eci.engine_class >= ARRAY_SIZE(user_to_xe_engine_class))
1240 		return NULL;
1241 
1242 	if (!gt)
1243 		return NULL;
1244 
1245 	idx = array_index_nospec(eci.engine_class,
1246 				 ARRAY_SIZE(user_to_xe_engine_class));
1247 
1248 	return xe_gt_hw_engine(xe_device_get_gt(xe, eci.gt_id),
1249 			       user_to_xe_engine_class[idx],
1250 			       eci.engine_instance, true);
1251 }
1252