xref: /netbsd-src/sys/external/bsd/drm2/dist/drm/i915/gt/intel_engine.h (revision 39d0ecb7560c909be3d64e62a882efdab11847d2)
1 /*	$NetBSD: intel_engine.h,v 1.5 2021/12/19 12:40:59 riastradh Exp $	*/
2 
3 /* SPDX-License-Identifier: MIT */
4 #ifndef _INTEL_RINGBUFFER_H_
5 #define _INTEL_RINGBUFFER_H_
6 
7 #include <drm/drm_util.h>
8 
9 #include <linux/hashtable.h>
10 #include <linux/irq_work.h>
11 #include <linux/random.h>
12 #include <linux/seqlock.h>
13 
14 #include "i915_pmu.h"
15 #include "i915_reg.h"
16 #include "i915_request.h"
17 #include "i915_selftest.h"
18 #include "gt/intel_timeline.h"
19 #include "intel_engine_types.h"
20 #include "intel_gpu_commands.h"
21 #include "intel_workarounds.h"
22 
23 struct drm_printer;
24 struct intel_gt;
25 
26 /* Early gen2 devices have a cacheline of just 32 bytes, using 64 is overkill,
27  * but keeps the logic simple. Indeed, the whole purpose of this macro is just
28  * to give some inclination as to some of the magic values used in the various
29  * workarounds!
30  */
31 #define CACHELINE_BYTES 64
32 #define CACHELINE_DWORDS (CACHELINE_BYTES / sizeof(u32))
33 
34 #define ENGINE_TRACE(e, fmt, ...) do {					\
35 	const struct intel_engine_cs *e__ __maybe_unused = (e);		\
36 	GEM_TRACE("%s %s: " fmt,					\
37 		  dev_name(e__->i915->drm.dev), e__->name,		\
38 		  ##__VA_ARGS__);					\
39 } while (0)
40 
41 /*
42  * The register defines to be used with the following macros need to accept a
43  * base param, e.g:
44  *
45  * REG_FOO(base) _MMIO((base) + <relative offset>)
46  * ENGINE_READ(engine, REG_FOO);
47  *
48  * register arrays are to be defined and accessed as follows:
49  *
50  * REG_BAR(base, i) _MMIO((base) + <relative offset> + (i) * <shift>)
51  * ENGINE_READ_IDX(engine, REG_BAR, i)
52  */
53 
54 #define __ENGINE_REG_OP(op__, engine__, ...) \
55 	intel_uncore_##op__((engine__)->uncore, __VA_ARGS__)
56 
57 #define __ENGINE_READ_OP(op__, engine__, reg__) \
58 	__ENGINE_REG_OP(op__, (engine__), reg__((engine__)->mmio_base))
59 
60 #define ENGINE_READ16(...)	__ENGINE_READ_OP(read16, __VA_ARGS__)
61 #define ENGINE_READ(...)	__ENGINE_READ_OP(read, __VA_ARGS__)
62 #define ENGINE_READ_FW(...)	__ENGINE_READ_OP(read_fw, __VA_ARGS__)
63 #define ENGINE_POSTING_READ(...) __ENGINE_READ_OP(posting_read_fw, __VA_ARGS__)
64 #define ENGINE_POSTING_READ16(...) __ENGINE_READ_OP(posting_read16, __VA_ARGS__)
65 
66 #define ENGINE_READ64(engine__, lower_reg__, upper_reg__) \
67 	__ENGINE_REG_OP(read64_2x32, (engine__), \
68 			lower_reg__((engine__)->mmio_base), \
69 			upper_reg__((engine__)->mmio_base))
70 
71 #define ENGINE_READ_IDX(engine__, reg__, idx__) \
72 	__ENGINE_REG_OP(read, (engine__), reg__((engine__)->mmio_base, (idx__)))
73 
74 #define __ENGINE_WRITE_OP(op__, engine__, reg__, val__) \
75 	__ENGINE_REG_OP(op__, (engine__), reg__((engine__)->mmio_base), (val__))
76 
77 #define ENGINE_WRITE16(...)	__ENGINE_WRITE_OP(write16, __VA_ARGS__)
78 #define ENGINE_WRITE(...)	__ENGINE_WRITE_OP(write, __VA_ARGS__)
79 #define ENGINE_WRITE_FW(...)	__ENGINE_WRITE_OP(write_fw, __VA_ARGS__)
80 
81 #define GEN6_RING_FAULT_REG_READ(engine__) \
82 	intel_uncore_read((engine__)->uncore, RING_FAULT_REG(engine__))
83 
84 #define GEN6_RING_FAULT_REG_POSTING_READ(engine__) \
85 	intel_uncore_posting_read((engine__)->uncore, RING_FAULT_REG(engine__))
86 
87 #define GEN6_RING_FAULT_REG_RMW(engine__, clear__, set__) \
88 ({ \
89 	u32 __val; \
90 \
91 	__val = intel_uncore_read((engine__)->uncore, \
92 				  RING_FAULT_REG(engine__)); \
93 	__val &= ~(clear__); \
94 	__val |= (set__); \
95 	intel_uncore_write((engine__)->uncore, RING_FAULT_REG(engine__), \
96 			   __val); \
97 })
98 
99 /* seqno size is actually only a uint32, but since we plan to use MI_FLUSH_DW to
100  * do the writes, and that must have qw aligned offsets, simply pretend it's 8b.
101  */
102 
103 static inline unsigned int
execlists_num_ports(const struct intel_engine_execlists * const execlists)104 execlists_num_ports(const struct intel_engine_execlists * const execlists)
105 {
106 	return execlists->port_mask + 1;
107 }
108 
109 static inline struct i915_request *
execlists_active(const struct intel_engine_execlists * execlists)110 execlists_active(const struct intel_engine_execlists *execlists)
111 {
112 	return *READ_ONCE(execlists->active);
113 }
114 
115 #ifdef __NetBSD__
116 static inline int
execlists_active_lock_bh(struct intel_engine_execlists * execlists)117 execlists_active_lock_bh(struct intel_engine_execlists *execlists)
118 {
119 	int s = splsoftserial(); /* prevent local softirq and lock recursion */
120 	tasklet_lock(&execlists->tasklet);
121 	return s;
122 }
123 
124 static inline void
execlists_active_unlock_bh(struct intel_engine_execlists * execlists,int s)125 execlists_active_unlock_bh(struct intel_engine_execlists *execlists, int s)
126 {
127 	tasklet_unlock(&execlists->tasklet);
128 	splx(s); /* restore softirq, and kick ksoftirqd! */
129 }
130 #else
131 static inline void
execlists_active_lock_bh(struct intel_engine_execlists * execlists)132 execlists_active_lock_bh(struct intel_engine_execlists *execlists)
133 {
134 	local_bh_disable(); /* prevent local softirq and lock recursion */
135 	tasklet_lock(&execlists->tasklet);
136 }
137 
138 static inline void
execlists_active_unlock_bh(struct intel_engine_execlists * execlists)139 execlists_active_unlock_bh(struct intel_engine_execlists *execlists)
140 {
141 	tasklet_unlock(&execlists->tasklet);
142 	local_bh_enable(); /* restore softirq, and kick ksoftirqd! */
143 }
144 #endif
145 
146 struct i915_request *
147 execlists_unwind_incomplete_requests(struct intel_engine_execlists *execlists);
148 
149 static inline u32
intel_read_status_page(const struct intel_engine_cs * engine,int reg)150 intel_read_status_page(const struct intel_engine_cs *engine, int reg)
151 {
152 	/* Ensure that the compiler doesn't optimize away the load. */
153 	return READ_ONCE(engine->status_page.addr[reg]);
154 }
155 
156 static inline void
intel_write_status_page(struct intel_engine_cs * engine,int reg,u32 value)157 intel_write_status_page(struct intel_engine_cs *engine, int reg, u32 value)
158 {
159 	/* Writing into the status page should be done sparingly. Since
160 	 * we do when we are uncertain of the device state, we take a bit
161 	 * of extra paranoia to try and ensure that the HWS takes the value
162 	 * we give and that it doesn't end up trapped inside the CPU!
163 	 */
164 	if (static_cpu_has(X86_FEATURE_CLFLUSH)) {
165 		mb();
166 		clflush(&engine->status_page.addr[reg]);
167 		engine->status_page.addr[reg] = value;
168 		clflush(&engine->status_page.addr[reg]);
169 		mb();
170 	} else {
171 		WRITE_ONCE(engine->status_page.addr[reg], value);
172 	}
173 }
174 
175 /*
176  * Reads a dword out of the status page, which is written to from the command
177  * queue by automatic updates, MI_REPORT_HEAD, MI_STORE_DATA_INDEX, or
178  * MI_STORE_DATA_IMM.
179  *
180  * The following dwords have a reserved meaning:
181  * 0x00: ISR copy, updated when an ISR bit not set in the HWSTAM changes.
182  * 0x04: ring 0 head pointer
183  * 0x05: ring 1 head pointer (915-class)
184  * 0x06: ring 2 head pointer (915-class)
185  * 0x10-0x1b: Context status DWords (GM45)
186  * 0x1f: Last written status offset. (GM45)
187  * 0x20-0x2f: Reserved (Gen6+)
188  *
189  * The area from dword 0x30 to 0x3ff is available for driver usage.
190  */
191 #define I915_GEM_HWS_PREEMPT		0x32
192 #define I915_GEM_HWS_PREEMPT_ADDR	(I915_GEM_HWS_PREEMPT * sizeof(u32))
193 #define I915_GEM_HWS_SEQNO		0x40
194 #define I915_GEM_HWS_SEQNO_ADDR		(I915_GEM_HWS_SEQNO * sizeof(u32))
195 #define I915_GEM_HWS_SCRATCH		0x80
196 #define I915_GEM_HWS_SCRATCH_ADDR	(I915_GEM_HWS_SCRATCH * sizeof(u32))
197 
198 #define I915_HWS_CSB_BUF0_INDEX		0x10
199 #define I915_HWS_CSB_WRITE_INDEX	0x1f
200 #define CNL_HWS_CSB_WRITE_INDEX		0x2f
201 
202 void intel_engine_stop(struct intel_engine_cs *engine);
203 void intel_engine_cleanup(struct intel_engine_cs *engine);
204 
205 int intel_engines_init_mmio(struct intel_gt *gt);
206 int intel_engines_init(struct intel_gt *gt);
207 
208 void intel_engines_release(struct intel_gt *gt);
209 void intel_engines_free(struct intel_gt *gt);
210 
211 int intel_engine_init_common(struct intel_engine_cs *engine);
212 void intel_engine_cleanup_common(struct intel_engine_cs *engine);
213 
214 int intel_ring_submission_setup(struct intel_engine_cs *engine);
215 
216 int intel_engine_stop_cs(struct intel_engine_cs *engine);
217 void intel_engine_cancel_stop_cs(struct intel_engine_cs *engine);
218 
219 void intel_engine_set_hwsp_writemask(struct intel_engine_cs *engine, u32 mask);
220 
221 u64 intel_engine_get_active_head(const struct intel_engine_cs *engine);
222 u64 intel_engine_get_last_batch_head(const struct intel_engine_cs *engine);
223 
224 void intel_engine_get_instdone(const struct intel_engine_cs *engine,
225 			       struct intel_instdone *instdone);
226 
227 void intel_engine_init_execlists(struct intel_engine_cs *engine);
228 
229 void intel_engine_init_breadcrumbs(struct intel_engine_cs *engine);
230 void intel_engine_fini_breadcrumbs(struct intel_engine_cs *engine);
231 
232 void intel_engine_disarm_breadcrumbs(struct intel_engine_cs *engine);
233 
234 static inline void
intel_engine_signal_breadcrumbs(struct intel_engine_cs * engine)235 intel_engine_signal_breadcrumbs(struct intel_engine_cs *engine)
236 {
237 	irq_work_queue(&engine->breadcrumbs.irq_work);
238 }
239 
240 void intel_engine_reset_breadcrumbs(struct intel_engine_cs *engine);
241 void intel_engine_fini_breadcrumbs(struct intel_engine_cs *engine);
242 
243 void intel_engine_print_breadcrumbs(struct intel_engine_cs *engine,
244 				    struct drm_printer *p);
245 
gen8_emit_pipe_control(u32 * batch,u32 flags,u32 offset)246 static inline u32 *gen8_emit_pipe_control(u32 *batch, u32 flags, u32 offset)
247 {
248 	memset(batch, 0, 6 * sizeof(u32));
249 
250 	batch[0] = GFX_OP_PIPE_CONTROL(6);
251 	batch[1] = flags;
252 	batch[2] = offset;
253 
254 	return batch + 6;
255 }
256 
257 static inline u32 *
gen8_emit_ggtt_write_rcs(u32 * cs,u32 value,u32 gtt_offset,u32 flags)258 gen8_emit_ggtt_write_rcs(u32 *cs, u32 value, u32 gtt_offset, u32 flags)
259 {
260 	/* We're using qword write, offset should be aligned to 8 bytes. */
261 	GEM_BUG_ON(!IS_ALIGNED(gtt_offset, 8));
262 
263 	/* w/a for post sync ops following a GPGPU operation we
264 	 * need a prior CS_STALL, which is emitted by the flush
265 	 * following the batch.
266 	 */
267 	*cs++ = GFX_OP_PIPE_CONTROL(6);
268 	*cs++ = flags | PIPE_CONTROL_QW_WRITE | PIPE_CONTROL_GLOBAL_GTT_IVB;
269 	*cs++ = gtt_offset;
270 	*cs++ = 0;
271 	*cs++ = value;
272 	/* We're thrashing one dword of HWS. */
273 	*cs++ = 0;
274 
275 	return cs;
276 }
277 
278 static inline u32 *
gen8_emit_ggtt_write(u32 * cs,u32 value,u32 gtt_offset,u32 flags)279 gen8_emit_ggtt_write(u32 *cs, u32 value, u32 gtt_offset, u32 flags)
280 {
281 	/* w/a: bit 5 needs to be zero for MI_FLUSH_DW address. */
282 	GEM_BUG_ON(gtt_offset & (1 << 5));
283 	/* Offset should be aligned to 8 bytes for both (QW/DW) write types */
284 	GEM_BUG_ON(!IS_ALIGNED(gtt_offset, 8));
285 
286 	*cs++ = (MI_FLUSH_DW + 1) | MI_FLUSH_DW_OP_STOREDW | flags;
287 	*cs++ = gtt_offset | MI_FLUSH_DW_USE_GTT;
288 	*cs++ = 0;
289 	*cs++ = value;
290 
291 	return cs;
292 }
293 
__intel_engine_reset(struct intel_engine_cs * engine,bool stalled)294 static inline void __intel_engine_reset(struct intel_engine_cs *engine,
295 					bool stalled)
296 {
297 	if (engine->reset.rewind)
298 		engine->reset.rewind(engine, stalled);
299 	engine->serial++; /* contexts lost */
300 }
301 
302 bool intel_engines_are_idle(struct intel_gt *gt);
303 bool intel_engine_is_idle(struct intel_engine_cs *engine);
304 void intel_engine_flush_submission(struct intel_engine_cs *engine);
305 
306 void intel_engines_reset_default_submission(struct intel_gt *gt);
307 
308 bool intel_engine_can_store_dword(struct intel_engine_cs *engine);
309 
310 __printf(3, 4)
311 void intel_engine_dump(struct intel_engine_cs *engine,
312 		       struct drm_printer *m,
313 		       const char *header, ...);
314 
315 int intel_enable_engine_stats(struct intel_engine_cs *engine);
316 void intel_disable_engine_stats(struct intel_engine_cs *engine);
317 
318 ktime_t intel_engine_get_busy_time(struct intel_engine_cs *engine);
319 
320 struct i915_request *
321 intel_engine_find_active_request(struct intel_engine_cs *engine);
322 
323 u32 intel_engine_context_size(struct intel_gt *gt, u8 class);
324 
325 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
326 
inject_preempt_hang(struct intel_engine_execlists * execlists)327 static inline bool inject_preempt_hang(struct intel_engine_execlists *execlists)
328 {
329 	if (!execlists->preempt_hang.inject_hang)
330 		return false;
331 
332 	complete(&execlists->preempt_hang.completion);
333 	return true;
334 }
335 
336 #else
337 
inject_preempt_hang(struct intel_engine_execlists * execlists)338 static inline bool inject_preempt_hang(struct intel_engine_execlists *execlists)
339 {
340 	return false;
341 }
342 
343 #endif
344 
345 void intel_engine_init_active(struct intel_engine_cs *engine,
346 			      unsigned int subclass);
347 #define ENGINE_PHYSICAL	0
348 #define ENGINE_MOCK	1
349 #define ENGINE_VIRTUAL	2
350 
351 static inline bool
intel_engine_has_preempt_reset(const struct intel_engine_cs * engine)352 intel_engine_has_preempt_reset(const struct intel_engine_cs *engine)
353 {
354 	if (!IS_ACTIVE(CONFIG_DRM_I915_PREEMPT_TIMEOUT))
355 		return false;
356 
357 	return intel_engine_has_preemption(engine);
358 }
359 
360 static inline bool
intel_engine_has_timeslices(const struct intel_engine_cs * engine)361 intel_engine_has_timeslices(const struct intel_engine_cs *engine)
362 {
363 	if (!IS_ACTIVE(CONFIG_DRM_I915_TIMESLICE_DURATION))
364 		return false;
365 
366 	return intel_engine_has_semaphores(engine);
367 }
368 
369 #endif /* _INTEL_RINGBUFFER_H_ */
370