xref: /linux/drivers/gpu/drm/i915/gt/intel_engine.h (revision 0ea5c948cb64bab5bc7a5516774eb8536f05aa0d)
1 /* SPDX-License-Identifier: MIT */
2 #ifndef _INTEL_RINGBUFFER_H_
3 #define _INTEL_RINGBUFFER_H_
4 
5 #include <asm/cacheflush.h>
6 #include <drm/drm_util.h>
7 #include <drm/drm_cache.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_request.h"
16 #include "i915_selftest.h"
17 #include "intel_engine_types.h"
18 #include "intel_gt_types.h"
19 #include "intel_timeline.h"
20 #include "intel_workarounds.h"
21 
22 struct drm_printer;
23 struct intel_context;
24 struct intel_gt;
25 struct lock_class_key;
26 
27 /* Early gen2 devices have a cacheline of just 32 bytes, using 64 is overkill,
28  * but keeps the logic simple. Indeed, the whole purpose of this macro is just
29  * to give some inclination as to some of the magic values used in the various
30  * workarounds!
31  */
32 #define CACHELINE_BYTES 64
33 #define CACHELINE_DWORDS (CACHELINE_BYTES / sizeof(u32))
34 
35 #define ENGINE_TRACE(e, fmt, ...) do {					\
36 	const struct intel_engine_cs *e__ __maybe_unused = (e);		\
37 	GEM_TRACE("%s %s: " fmt,					\
38 		  dev_name(e__->i915->drm.dev), e__->name,		\
39 		  ##__VA_ARGS__);					\
40 } while (0)
41 
42 /*
43  * The register defines to be used with the following macros need to accept a
44  * base param, e.g:
45  *
46  * REG_FOO(base) _MMIO((base) + <relative offset>)
47  * ENGINE_READ(engine, REG_FOO);
48  *
49  * register arrays are to be defined and accessed as follows:
50  *
51  * REG_BAR(base, i) _MMIO((base) + <relative offset> + (i) * <shift>)
52  * ENGINE_READ_IDX(engine, REG_BAR, i)
53  */
54 
55 #define __ENGINE_REG_OP(op__, engine__, ...) \
56 	intel_uncore_##op__((engine__)->uncore, __VA_ARGS__)
57 
58 #define __ENGINE_READ_OP(op__, engine__, reg__) \
59 	__ENGINE_REG_OP(op__, (engine__), reg__((engine__)->mmio_base))
60 
61 #define ENGINE_READ16(...)	__ENGINE_READ_OP(read16, __VA_ARGS__)
62 #define ENGINE_READ(...)	__ENGINE_READ_OP(read, __VA_ARGS__)
63 #define ENGINE_READ_FW(...)	__ENGINE_READ_OP(read_fw, __VA_ARGS__)
64 #define ENGINE_POSTING_READ(...) __ENGINE_READ_OP(posting_read_fw, __VA_ARGS__)
65 #define ENGINE_POSTING_READ16(...) __ENGINE_READ_OP(posting_read16, __VA_ARGS__)
66 
67 #define ENGINE_READ64(engine__, lower_reg__, upper_reg__) \
68 	__ENGINE_REG_OP(read64_2x32, (engine__), \
69 			lower_reg__((engine__)->mmio_base), \
70 			upper_reg__((engine__)->mmio_base))
71 
72 #define ENGINE_READ_IDX(engine__, reg__, idx__) \
73 	__ENGINE_REG_OP(read, (engine__), reg__((engine__)->mmio_base, (idx__)))
74 
75 #define __ENGINE_WRITE_OP(op__, engine__, reg__, val__) \
76 	__ENGINE_REG_OP(op__, (engine__), reg__((engine__)->mmio_base), (val__))
77 
78 #define ENGINE_WRITE16(...)	__ENGINE_WRITE_OP(write16, __VA_ARGS__)
79 #define ENGINE_WRITE(...)	__ENGINE_WRITE_OP(write, __VA_ARGS__)
80 #define ENGINE_WRITE_FW(...)	__ENGINE_WRITE_OP(write_fw, __VA_ARGS__)
81 
82 #define GEN6_RING_FAULT_REG_READ(engine__) \
83 	intel_uncore_read((engine__)->uncore, RING_FAULT_REG(engine__))
84 
85 #define GEN6_RING_FAULT_REG_POSTING_READ(engine__) \
86 	intel_uncore_posting_read((engine__)->uncore, RING_FAULT_REG(engine__))
87 
88 #define GEN6_RING_FAULT_REG_RMW(engine__, clear__, set__) \
89 ({ \
90 	u32 __val; \
91 \
92 	__val = intel_uncore_read((engine__)->uncore, \
93 				  RING_FAULT_REG(engine__)); \
94 	__val &= ~(clear__); \
95 	__val |= (set__); \
96 	intel_uncore_write((engine__)->uncore, RING_FAULT_REG(engine__), \
97 			   __val); \
98 })
99 
100 /* seqno size is actually only a uint32, but since we plan to use MI_FLUSH_DW to
101  * do the writes, and that must have qw aligned offsets, simply pretend it's 8b.
102  */
103 
104 static inline unsigned int
execlists_num_ports(const struct intel_engine_execlists * const execlists)105 execlists_num_ports(const struct intel_engine_execlists * const execlists)
106 {
107 	return execlists->port_mask + 1;
108 }
109 
110 static inline struct i915_request *
execlists_active(const struct intel_engine_execlists * execlists)111 execlists_active(const struct intel_engine_execlists *execlists)
112 {
113 	struct i915_request * const *cur, * const *old, *active;
114 
115 	cur = READ_ONCE(execlists->active);
116 	smp_rmb(); /* pairs with overwrite protection in process_csb() */
117 	do {
118 		old = cur;
119 
120 		active = READ_ONCE(*cur);
121 		cur = READ_ONCE(execlists->active);
122 
123 		smp_rmb(); /* and complete the seqlock retry */
124 	} while (unlikely(cur != old));
125 
126 	return active;
127 }
128 
129 struct i915_request *
130 execlists_unwind_incomplete_requests(struct intel_engine_execlists *execlists);
131 
132 static inline u32
intel_read_status_page(const struct intel_engine_cs * engine,int reg)133 intel_read_status_page(const struct intel_engine_cs *engine, int reg)
134 {
135 	/* Ensure that the compiler doesn't optimize away the load. */
136 	return READ_ONCE(engine->status_page.addr[reg]);
137 }
138 
139 static inline void
intel_write_status_page(struct intel_engine_cs * engine,int reg,u32 value)140 intel_write_status_page(struct intel_engine_cs *engine, int reg, u32 value)
141 {
142 	/* Writing into the status page should be done sparingly. Since
143 	 * we do when we are uncertain of the device state, we take a bit
144 	 * of extra paranoia to try and ensure that the HWS takes the value
145 	 * we give and that it doesn't end up trapped inside the CPU!
146 	 */
147 	drm_clflush_virt_range(&engine->status_page.addr[reg], sizeof(value));
148 	WRITE_ONCE(engine->status_page.addr[reg], value);
149 	drm_clflush_virt_range(&engine->status_page.addr[reg], sizeof(value));
150 }
151 
152 /*
153  * Reads a dword out of the status page, which is written to from the command
154  * queue by automatic updates, MI_REPORT_HEAD, MI_STORE_DATA_INDEX, or
155  * MI_STORE_DATA_IMM.
156  *
157  * The following dwords have a reserved meaning:
158  * 0x00: ISR copy, updated when an ISR bit not set in the HWSTAM changes.
159  * 0x04: ring 0 head pointer
160  * 0x05: ring 1 head pointer (915-class)
161  * 0x06: ring 2 head pointer (915-class)
162  * 0x10-0x1b: Context status DWords (GM45)
163  * 0x1f: Last written status offset. (GM45)
164  * 0x20-0x2f: Reserved (Gen6+)
165  *
166  * The area from dword 0x30 to 0x3ff is available for driver usage.
167  */
168 #define I915_GEM_HWS_PREEMPT		0x32
169 #define I915_GEM_HWS_PREEMPT_ADDR	(I915_GEM_HWS_PREEMPT * sizeof(u32))
170 #define I915_GEM_HWS_SEQNO		0x40
171 #define I915_GEM_HWS_SEQNO_ADDR		(I915_GEM_HWS_SEQNO * sizeof(u32))
172 #define I915_GEM_HWS_MIGRATE		(0x42 * sizeof(u32))
173 #define I915_GEM_HWS_GGTT_BIND		0x46
174 #define I915_GEM_HWS_GGTT_BIND_ADDR	(I915_GEM_HWS_GGTT_BIND * sizeof(u32))
175 #define I915_GEM_HWS_PXP		0x60
176 #define I915_GEM_HWS_PXP_ADDR		(I915_GEM_HWS_PXP * sizeof(u32))
177 #define I915_GEM_HWS_GSC		0x62
178 #define I915_GEM_HWS_GSC_ADDR		(I915_GEM_HWS_GSC * sizeof(u32))
179 #define I915_GEM_HWS_SCRATCH		0x80
180 
181 #define I915_HWS_CSB_BUF0_INDEX		0x10
182 #define I915_HWS_CSB_WRITE_INDEX	0x1f
183 #define ICL_HWS_CSB_WRITE_INDEX		0x2f
184 #define INTEL_HWS_CSB_WRITE_INDEX(__i915) \
185 	(GRAPHICS_VER(__i915) >= 11 ? ICL_HWS_CSB_WRITE_INDEX : I915_HWS_CSB_WRITE_INDEX)
186 
187 void intel_engine_stop(struct intel_engine_cs *engine);
188 void intel_engine_cleanup(struct intel_engine_cs *engine);
189 
190 int intel_engines_init_mmio(struct intel_gt *gt);
191 int intel_engines_init(struct intel_gt *gt);
192 
193 void intel_engine_free_request_pool(struct intel_engine_cs *engine);
194 
195 void intel_engines_release(struct intel_gt *gt);
196 void intel_engines_free(struct intel_gt *gt);
197 
198 int intel_engine_init_common(struct intel_engine_cs *engine);
199 void intel_engine_cleanup_common(struct intel_engine_cs *engine);
200 
201 int intel_engine_resume(struct intel_engine_cs *engine);
202 
203 int intel_ring_submission_setup(struct intel_engine_cs *engine);
204 
205 int intel_engine_stop_cs(struct intel_engine_cs *engine);
206 void intel_engine_cancel_stop_cs(struct intel_engine_cs *engine);
207 
208 void intel_engine_wait_for_pending_mi_fw(struct intel_engine_cs *engine);
209 
210 void intel_engine_set_hwsp_writemask(struct intel_engine_cs *engine, u32 mask);
211 
212 u64 intel_engine_get_active_head(const struct intel_engine_cs *engine);
213 u64 intel_engine_get_last_batch_head(const struct intel_engine_cs *engine);
214 
215 void intel_engine_get_instdone(const struct intel_engine_cs *engine,
216 			       struct intel_instdone *instdone);
217 
218 void intel_engine_init_execlists(struct intel_engine_cs *engine);
219 
220 bool intel_engine_irq_enable(struct intel_engine_cs *engine);
221 void intel_engine_irq_disable(struct intel_engine_cs *engine);
222 
__intel_engine_reset(struct intel_engine_cs * engine,bool stalled)223 static inline void __intel_engine_reset(struct intel_engine_cs *engine,
224 					bool stalled)
225 {
226 	if (engine->reset.rewind)
227 		engine->reset.rewind(engine, stalled);
228 	engine->serial++; /* contexts lost */
229 }
230 
231 bool intel_engines_are_idle(struct intel_gt *gt);
232 bool intel_engine_is_idle(struct intel_engine_cs *engine);
233 
234 void __intel_engine_flush_submission(struct intel_engine_cs *engine, bool sync);
intel_engine_flush_submission(struct intel_engine_cs * engine)235 static inline void intel_engine_flush_submission(struct intel_engine_cs *engine)
236 {
237 	__intel_engine_flush_submission(engine, true);
238 }
239 
240 void intel_engines_reset_default_submission(struct intel_gt *gt);
241 
242 bool intel_engine_can_store_dword(struct intel_engine_cs *engine);
243 
244 __printf(3, 4)
245 void intel_engine_dump(struct intel_engine_cs *engine,
246 		       struct drm_printer *m,
247 		       const char *header, ...);
248 void intel_engine_dump_active_requests(struct list_head *requests,
249 				       struct i915_request *hung_rq,
250 				       struct drm_printer *m);
251 
252 ktime_t intel_engine_get_busy_time(struct intel_engine_cs *engine,
253 				   ktime_t *now);
254 
255 void intel_engine_get_hung_entity(struct intel_engine_cs *engine,
256 				  struct intel_context **ce, struct i915_request **rq);
257 
258 u32 intel_engine_context_size(struct intel_gt *gt, u8 class);
259 struct intel_context *
260 intel_engine_create_pinned_context(struct intel_engine_cs *engine,
261 				   struct i915_address_space *vm,
262 				   unsigned int ring_size,
263 				   unsigned int hwsp,
264 				   struct lock_class_key *key,
265 				   const char *name);
266 
267 void intel_engine_destroy_pinned_context(struct intel_context *ce);
268 
269 void xehp_enable_ccs_engines(struct intel_engine_cs *engine);
270 
271 #define ENGINE_PHYSICAL	0
272 #define ENGINE_MOCK	1
273 #define ENGINE_VIRTUAL	2
274 
intel_engine_uses_guc(const struct intel_engine_cs * engine)275 static inline bool intel_engine_uses_guc(const struct intel_engine_cs *engine)
276 {
277 	return engine->gt->submission_method >= INTEL_SUBMISSION_GUC;
278 }
279 
280 static inline bool
intel_engine_has_preempt_reset(const struct intel_engine_cs * engine)281 intel_engine_has_preempt_reset(const struct intel_engine_cs *engine)
282 {
283 	if (!CONFIG_DRM_I915_PREEMPT_TIMEOUT)
284 		return false;
285 
286 	return intel_engine_has_preemption(engine);
287 }
288 
289 #define FORCE_VIRTUAL	BIT(0)
290 struct intel_context *
291 intel_engine_create_virtual(struct intel_engine_cs **siblings,
292 			    unsigned int count, unsigned long flags);
293 
294 static inline struct intel_context *
intel_engine_create_parallel(struct intel_engine_cs ** engines,unsigned int num_engines,unsigned int width)295 intel_engine_create_parallel(struct intel_engine_cs **engines,
296 			     unsigned int num_engines,
297 			     unsigned int width)
298 {
299 	GEM_BUG_ON(!engines[0]->cops->create_parallel);
300 	return engines[0]->cops->create_parallel(engines, num_engines, width);
301 }
302 
303 static inline bool
intel_virtual_engine_has_heartbeat(const struct intel_engine_cs * engine)304 intel_virtual_engine_has_heartbeat(const struct intel_engine_cs *engine)
305 {
306 	/*
307 	 * For non-GuC submission we expect the back-end to look at the
308 	 * heartbeat status of the actual physical engine that the work
309 	 * has been (or is being) scheduled on, so we should only reach
310 	 * here with GuC submission enabled.
311 	 */
312 	GEM_BUG_ON(!intel_engine_uses_guc(engine));
313 
314 	return intel_guc_virtual_engine_has_heartbeat(engine);
315 }
316 
317 static inline bool
intel_engine_has_heartbeat(const struct intel_engine_cs * engine)318 intel_engine_has_heartbeat(const struct intel_engine_cs *engine)
319 {
320 	if (!CONFIG_DRM_I915_HEARTBEAT_INTERVAL)
321 		return false;
322 
323 	if (intel_engine_is_virtual(engine))
324 		return intel_virtual_engine_has_heartbeat(engine);
325 	else
326 		return READ_ONCE(engine->props.heartbeat_interval_ms);
327 }
328 
329 static inline struct intel_engine_cs *
intel_engine_get_sibling(struct intel_engine_cs * engine,unsigned int sibling)330 intel_engine_get_sibling(struct intel_engine_cs *engine, unsigned int sibling)
331 {
332 	GEM_BUG_ON(!intel_engine_is_virtual(engine));
333 	return engine->cops->get_sibling(engine, sibling);
334 }
335 
336 static inline void
intel_engine_set_hung_context(struct intel_engine_cs * engine,struct intel_context * ce)337 intel_engine_set_hung_context(struct intel_engine_cs *engine,
338 			      struct intel_context *ce)
339 {
340 	engine->hung_ce = ce;
341 }
342 
343 static inline void
intel_engine_clear_hung_context(struct intel_engine_cs * engine)344 intel_engine_clear_hung_context(struct intel_engine_cs *engine)
345 {
346 	intel_engine_set_hung_context(engine, NULL);
347 }
348 
349 static inline struct intel_context *
intel_engine_get_hung_context(struct intel_engine_cs * engine)350 intel_engine_get_hung_context(struct intel_engine_cs *engine)
351 {
352 	return engine->hung_ce;
353 }
354 
355 u64 intel_clamp_heartbeat_interval_ms(struct intel_engine_cs *engine, u64 value);
356 u64 intel_clamp_max_busywait_duration_ns(struct intel_engine_cs *engine, u64 value);
357 u64 intel_clamp_preempt_timeout_ms(struct intel_engine_cs *engine, u64 value);
358 u64 intel_clamp_stop_timeout_ms(struct intel_engine_cs *engine, u64 value);
359 u64 intel_clamp_timeslice_duration_ms(struct intel_engine_cs *engine, u64 value);
360 
361 #endif /* _INTEL_RINGBUFFER_H_ */
362