xref: /linux/drivers/gpu/drm/i915/gt/intel_engine.h (revision 2c1ed907520c50326b8f604907a8478b27881a2e)
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  static inline u32
intel_read_status_page(const struct intel_engine_cs * engine,int reg)130  intel_read_status_page(const struct intel_engine_cs *engine, int reg)
131  {
132  	/* Ensure that the compiler doesn't optimize away the load. */
133  	return READ_ONCE(engine->status_page.addr[reg]);
134  }
135  
136  static inline void
intel_write_status_page(struct intel_engine_cs * engine,int reg,u32 value)137  intel_write_status_page(struct intel_engine_cs *engine, int reg, u32 value)
138  {
139  	/* Writing into the status page should be done sparingly. Since
140  	 * we do when we are uncertain of the device state, we take a bit
141  	 * of extra paranoia to try and ensure that the HWS takes the value
142  	 * we give and that it doesn't end up trapped inside the CPU!
143  	 */
144  	drm_clflush_virt_range(&engine->status_page.addr[reg], sizeof(value));
145  	WRITE_ONCE(engine->status_page.addr[reg], value);
146  	drm_clflush_virt_range(&engine->status_page.addr[reg], sizeof(value));
147  }
148  
149  /*
150   * Reads a dword out of the status page, which is written to from the command
151   * queue by automatic updates, MI_REPORT_HEAD, MI_STORE_DATA_INDEX, or
152   * MI_STORE_DATA_IMM.
153   *
154   * The following dwords have a reserved meaning:
155   * 0x00: ISR copy, updated when an ISR bit not set in the HWSTAM changes.
156   * 0x04: ring 0 head pointer
157   * 0x05: ring 1 head pointer (915-class)
158   * 0x06: ring 2 head pointer (915-class)
159   * 0x10-0x1b: Context status DWords (GM45)
160   * 0x1f: Last written status offset. (GM45)
161   * 0x20-0x2f: Reserved (Gen6+)
162   *
163   * The area from dword 0x30 to 0x3ff is available for driver usage.
164   */
165  #define I915_GEM_HWS_PREEMPT		0x32
166  #define I915_GEM_HWS_PREEMPT_ADDR	(I915_GEM_HWS_PREEMPT * sizeof(u32))
167  #define I915_GEM_HWS_SEQNO		0x40
168  #define I915_GEM_HWS_SEQNO_ADDR		(I915_GEM_HWS_SEQNO * sizeof(u32))
169  #define I915_GEM_HWS_MIGRATE		(0x42 * sizeof(u32))
170  #define I915_GEM_HWS_GGTT_BIND		0x46
171  #define I915_GEM_HWS_GGTT_BIND_ADDR	(I915_GEM_HWS_GGTT_BIND * sizeof(u32))
172  #define I915_GEM_HWS_PXP		0x60
173  #define I915_GEM_HWS_PXP_ADDR		(I915_GEM_HWS_PXP * sizeof(u32))
174  #define I915_GEM_HWS_GSC		0x62
175  #define I915_GEM_HWS_GSC_ADDR		(I915_GEM_HWS_GSC * sizeof(u32))
176  #define I915_GEM_HWS_SCRATCH		0x80
177  
178  #define I915_HWS_CSB_BUF0_INDEX		0x10
179  #define I915_HWS_CSB_WRITE_INDEX	0x1f
180  #define ICL_HWS_CSB_WRITE_INDEX		0x2f
181  #define INTEL_HWS_CSB_WRITE_INDEX(__i915) \
182  	(GRAPHICS_VER(__i915) >= 11 ? ICL_HWS_CSB_WRITE_INDEX : I915_HWS_CSB_WRITE_INDEX)
183  
184  void intel_engine_stop(struct intel_engine_cs *engine);
185  void intel_engine_cleanup(struct intel_engine_cs *engine);
186  
187  int intel_engines_init_mmio(struct intel_gt *gt);
188  int intel_engines_init(struct intel_gt *gt);
189  
190  void intel_engine_free_request_pool(struct intel_engine_cs *engine);
191  
192  void intel_engines_release(struct intel_gt *gt);
193  void intel_engines_free(struct intel_gt *gt);
194  
195  int intel_engine_init_common(struct intel_engine_cs *engine);
196  void intel_engine_cleanup_common(struct intel_engine_cs *engine);
197  
198  int intel_engine_resume(struct intel_engine_cs *engine);
199  
200  int intel_ring_submission_setup(struct intel_engine_cs *engine);
201  
202  int intel_engine_stop_cs(struct intel_engine_cs *engine);
203  void intel_engine_cancel_stop_cs(struct intel_engine_cs *engine);
204  
205  void intel_engine_wait_for_pending_mi_fw(struct intel_engine_cs *engine);
206  
207  void intel_engine_set_hwsp_writemask(struct intel_engine_cs *engine, u32 mask);
208  
209  u64 intel_engine_get_active_head(const struct intel_engine_cs *engine);
210  u64 intel_engine_get_last_batch_head(const struct intel_engine_cs *engine);
211  
212  void intel_engine_get_instdone(const struct intel_engine_cs *engine,
213  			       struct intel_instdone *instdone);
214  
215  void intel_engine_init_execlists(struct intel_engine_cs *engine);
216  
217  bool intel_engine_irq_enable(struct intel_engine_cs *engine);
218  void intel_engine_irq_disable(struct intel_engine_cs *engine);
219  
__intel_engine_reset(struct intel_engine_cs * engine,bool stalled)220  static inline void __intel_engine_reset(struct intel_engine_cs *engine,
221  					bool stalled)
222  {
223  	if (engine->reset.rewind)
224  		engine->reset.rewind(engine, stalled);
225  	engine->serial++; /* contexts lost */
226  }
227  
228  bool intel_engines_are_idle(struct intel_gt *gt);
229  bool intel_engine_is_idle(struct intel_engine_cs *engine);
230  
231  void __intel_engine_flush_submission(struct intel_engine_cs *engine, bool sync);
intel_engine_flush_submission(struct intel_engine_cs * engine)232  static inline void intel_engine_flush_submission(struct intel_engine_cs *engine)
233  {
234  	__intel_engine_flush_submission(engine, true);
235  }
236  
237  void intel_engines_reset_default_submission(struct intel_gt *gt);
238  
239  bool intel_engine_can_store_dword(struct intel_engine_cs *engine);
240  
241  __printf(3, 4)
242  void intel_engine_dump(struct intel_engine_cs *engine,
243  		       struct drm_printer *m,
244  		       const char *header, ...);
245  void intel_engine_dump_active_requests(struct list_head *requests,
246  				       struct i915_request *hung_rq,
247  				       struct drm_printer *m);
248  
249  ktime_t intel_engine_get_busy_time(struct intel_engine_cs *engine,
250  				   ktime_t *now);
251  
252  void intel_engine_get_hung_entity(struct intel_engine_cs *engine,
253  				  struct intel_context **ce, struct i915_request **rq);
254  
255  u32 intel_engine_context_size(struct intel_gt *gt, u8 class);
256  struct intel_context *
257  intel_engine_create_pinned_context(struct intel_engine_cs *engine,
258  				   struct i915_address_space *vm,
259  				   unsigned int ring_size,
260  				   unsigned int hwsp,
261  				   struct lock_class_key *key,
262  				   const char *name);
263  
264  void intel_engine_destroy_pinned_context(struct intel_context *ce);
265  
266  void xehp_enable_ccs_engines(struct intel_engine_cs *engine);
267  
268  #define ENGINE_PHYSICAL	0
269  #define ENGINE_MOCK	1
270  #define ENGINE_VIRTUAL	2
271  
intel_engine_uses_guc(const struct intel_engine_cs * engine)272  static inline bool intel_engine_uses_guc(const struct intel_engine_cs *engine)
273  {
274  	return engine->gt->submission_method >= INTEL_SUBMISSION_GUC;
275  }
276  
277  static inline bool
intel_engine_has_preempt_reset(const struct intel_engine_cs * engine)278  intel_engine_has_preempt_reset(const struct intel_engine_cs *engine)
279  {
280  	if (!CONFIG_DRM_I915_PREEMPT_TIMEOUT)
281  		return false;
282  
283  	return intel_engine_has_preemption(engine);
284  }
285  
286  #define FORCE_VIRTUAL	BIT(0)
287  struct intel_context *
288  intel_engine_create_virtual(struct intel_engine_cs **siblings,
289  			    unsigned int count, unsigned long flags);
290  
291  static inline struct intel_context *
intel_engine_create_parallel(struct intel_engine_cs ** engines,unsigned int num_engines,unsigned int width)292  intel_engine_create_parallel(struct intel_engine_cs **engines,
293  			     unsigned int num_engines,
294  			     unsigned int width)
295  {
296  	GEM_BUG_ON(!engines[0]->cops->create_parallel);
297  	return engines[0]->cops->create_parallel(engines, num_engines, width);
298  }
299  
300  static inline bool
intel_virtual_engine_has_heartbeat(const struct intel_engine_cs * engine)301  intel_virtual_engine_has_heartbeat(const struct intel_engine_cs *engine)
302  {
303  	/*
304  	 * For non-GuC submission we expect the back-end to look at the
305  	 * heartbeat status of the actual physical engine that the work
306  	 * has been (or is being) scheduled on, so we should only reach
307  	 * here with GuC submission enabled.
308  	 */
309  	GEM_BUG_ON(!intel_engine_uses_guc(engine));
310  
311  	return intel_guc_virtual_engine_has_heartbeat(engine);
312  }
313  
314  static inline bool
intel_engine_has_heartbeat(const struct intel_engine_cs * engine)315  intel_engine_has_heartbeat(const struct intel_engine_cs *engine)
316  {
317  	if (!CONFIG_DRM_I915_HEARTBEAT_INTERVAL)
318  		return false;
319  
320  	if (intel_engine_is_virtual(engine))
321  		return intel_virtual_engine_has_heartbeat(engine);
322  	else
323  		return READ_ONCE(engine->props.heartbeat_interval_ms);
324  }
325  
326  static inline struct intel_engine_cs *
intel_engine_get_sibling(struct intel_engine_cs * engine,unsigned int sibling)327  intel_engine_get_sibling(struct intel_engine_cs *engine, unsigned int sibling)
328  {
329  	GEM_BUG_ON(!intel_engine_is_virtual(engine));
330  	return engine->cops->get_sibling(engine, sibling);
331  }
332  
333  static inline void
intel_engine_set_hung_context(struct intel_engine_cs * engine,struct intel_context * ce)334  intel_engine_set_hung_context(struct intel_engine_cs *engine,
335  			      struct intel_context *ce)
336  {
337  	engine->hung_ce = ce;
338  }
339  
340  static inline void
intel_engine_clear_hung_context(struct intel_engine_cs * engine)341  intel_engine_clear_hung_context(struct intel_engine_cs *engine)
342  {
343  	intel_engine_set_hung_context(engine, NULL);
344  }
345  
346  static inline struct intel_context *
intel_engine_get_hung_context(struct intel_engine_cs * engine)347  intel_engine_get_hung_context(struct intel_engine_cs *engine)
348  {
349  	return engine->hung_ce;
350  }
351  
352  u64 intel_clamp_heartbeat_interval_ms(struct intel_engine_cs *engine, u64 value);
353  u64 intel_clamp_max_busywait_duration_ns(struct intel_engine_cs *engine, u64 value);
354  u64 intel_clamp_preempt_timeout_ms(struct intel_engine_cs *engine, u64 value);
355  u64 intel_clamp_stop_timeout_ms(struct intel_engine_cs *engine, u64 value);
356  u64 intel_clamp_timeslice_duration_ms(struct intel_engine_cs *engine, u64 value);
357  
358  #endif /* _INTEL_RINGBUFFER_H_ */
359