Merge branch 'stable/for-linus-5.15' of git://git.kernel.org/pub/scm/linux/kernel...
[linux-2.6-microblaze.git] / drivers / gpu / drm / i915 / gt / intel_engine.h
1 /* SPDX-License-Identifier: MIT */
2 #ifndef _INTEL_RINGBUFFER_H_
3 #define _INTEL_RINGBUFFER_H_
4
5 #include <drm/drm_util.h>
6
7 #include <linux/hashtable.h>
8 #include <linux/irq_work.h>
9 #include <linux/random.h>
10 #include <linux/seqlock.h>
11
12 #include "i915_pmu.h"
13 #include "i915_reg.h"
14 #include "i915_request.h"
15 #include "i915_selftest.h"
16 #include "intel_engine_types.h"
17 #include "intel_gt_types.h"
18 #include "intel_timeline.h"
19 #include "intel_workarounds.h"
20
21 struct drm_printer;
22 struct intel_context;
23 struct intel_gt;
24 struct lock_class_key;
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
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 *
110 execlists_active(const struct intel_engine_execlists *execlists)
111 {
112         struct i915_request * const *cur, * const *old, *active;
113
114         cur = READ_ONCE(execlists->active);
115         smp_rmb(); /* pairs with overwrite protection in process_csb() */
116         do {
117                 old = cur;
118
119                 active = READ_ONCE(*cur);
120                 cur = READ_ONCE(execlists->active);
121
122                 smp_rmb(); /* and complete the seqlock retry */
123         } while (unlikely(cur != old));
124
125         return active;
126 }
127
128 struct i915_request *
129 execlists_unwind_incomplete_requests(struct intel_engine_execlists *execlists);
130
131 static inline u32
132 intel_read_status_page(const struct intel_engine_cs *engine, int reg)
133 {
134         /* Ensure that the compiler doesn't optimize away the load. */
135         return READ_ONCE(engine->status_page.addr[reg]);
136 }
137
138 static inline void
139 intel_write_status_page(struct intel_engine_cs *engine, int reg, u32 value)
140 {
141         /* Writing into the status page should be done sparingly. Since
142          * we do when we are uncertain of the device state, we take a bit
143          * of extra paranoia to try and ensure that the HWS takes the value
144          * we give and that it doesn't end up trapped inside the CPU!
145          */
146         if (static_cpu_has(X86_FEATURE_CLFLUSH)) {
147                 mb();
148                 clflush(&engine->status_page.addr[reg]);
149                 engine->status_page.addr[reg] = value;
150                 clflush(&engine->status_page.addr[reg]);
151                 mb();
152         } else {
153                 WRITE_ONCE(engine->status_page.addr[reg], value);
154         }
155 }
156
157 /*
158  * Reads a dword out of the status page, which is written to from the command
159  * queue by automatic updates, MI_REPORT_HEAD, MI_STORE_DATA_INDEX, or
160  * MI_STORE_DATA_IMM.
161  *
162  * The following dwords have a reserved meaning:
163  * 0x00: ISR copy, updated when an ISR bit not set in the HWSTAM changes.
164  * 0x04: ring 0 head pointer
165  * 0x05: ring 1 head pointer (915-class)
166  * 0x06: ring 2 head pointer (915-class)
167  * 0x10-0x1b: Context status DWords (GM45)
168  * 0x1f: Last written status offset. (GM45)
169  * 0x20-0x2f: Reserved (Gen6+)
170  *
171  * The area from dword 0x30 to 0x3ff is available for driver usage.
172  */
173 #define I915_GEM_HWS_PREEMPT            0x32
174 #define I915_GEM_HWS_PREEMPT_ADDR       (I915_GEM_HWS_PREEMPT * sizeof(u32))
175 #define I915_GEM_HWS_SEQNO              0x40
176 #define I915_GEM_HWS_SEQNO_ADDR         (I915_GEM_HWS_SEQNO * sizeof(u32))
177 #define I915_GEM_HWS_MIGRATE            (0x42 * sizeof(u32))
178 #define I915_GEM_HWS_SCRATCH            0x80
179
180 #define I915_HWS_CSB_BUF0_INDEX         0x10
181 #define I915_HWS_CSB_WRITE_INDEX        0x1f
182 #define ICL_HWS_CSB_WRITE_INDEX         0x2f
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_set_hwsp_writemask(struct intel_engine_cs *engine, u32 mask);
206
207 u64 intel_engine_get_active_head(const struct intel_engine_cs *engine);
208 u64 intel_engine_get_last_batch_head(const struct intel_engine_cs *engine);
209
210 void intel_engine_get_instdone(const struct intel_engine_cs *engine,
211                                struct intel_instdone *instdone);
212
213 void intel_engine_init_execlists(struct intel_engine_cs *engine);
214
215 bool intel_engine_irq_enable(struct intel_engine_cs *engine);
216 void intel_engine_irq_disable(struct intel_engine_cs *engine);
217
218 static inline void __intel_engine_reset(struct intel_engine_cs *engine,
219                                         bool stalled)
220 {
221         if (engine->reset.rewind)
222                 engine->reset.rewind(engine, stalled);
223         engine->serial++; /* contexts lost */
224 }
225
226 bool intel_engines_are_idle(struct intel_gt *gt);
227 bool intel_engine_is_idle(struct intel_engine_cs *engine);
228
229 void __intel_engine_flush_submission(struct intel_engine_cs *engine, bool sync);
230 static inline void intel_engine_flush_submission(struct intel_engine_cs *engine)
231 {
232         __intel_engine_flush_submission(engine, true);
233 }
234
235 void intel_engines_reset_default_submission(struct intel_gt *gt);
236
237 bool intel_engine_can_store_dword(struct intel_engine_cs *engine);
238
239 __printf(3, 4)
240 void intel_engine_dump(struct intel_engine_cs *engine,
241                        struct drm_printer *m,
242                        const char *header, ...);
243 void intel_engine_dump_active_requests(struct list_head *requests,
244                                        struct i915_request *hung_rq,
245                                        struct drm_printer *m);
246
247 ktime_t intel_engine_get_busy_time(struct intel_engine_cs *engine,
248                                    ktime_t *now);
249
250 struct i915_request *
251 intel_engine_execlist_find_hung_request(struct intel_engine_cs *engine);
252
253 u32 intel_engine_context_size(struct intel_gt *gt, u8 class);
254 struct intel_context *
255 intel_engine_create_pinned_context(struct intel_engine_cs *engine,
256                                    struct i915_address_space *vm,
257                                    unsigned int ring_size,
258                                    unsigned int hwsp,
259                                    struct lock_class_key *key,
260                                    const char *name);
261
262 void intel_engine_destroy_pinned_context(struct intel_context *ce);
263
264 #define ENGINE_PHYSICAL 0
265 #define ENGINE_MOCK     1
266 #define ENGINE_VIRTUAL  2
267
268 static inline bool intel_engine_uses_guc(const struct intel_engine_cs *engine)
269 {
270         return engine->gt->submission_method >= INTEL_SUBMISSION_GUC;
271 }
272
273 static inline bool
274 intel_engine_has_preempt_reset(const struct intel_engine_cs *engine)
275 {
276         if (!IS_ACTIVE(CONFIG_DRM_I915_PREEMPT_TIMEOUT))
277                 return false;
278
279         return intel_engine_has_preemption(engine);
280 }
281
282 struct intel_context *
283 intel_engine_create_virtual(struct intel_engine_cs **siblings,
284                             unsigned int count);
285
286 static inline bool
287 intel_virtual_engine_has_heartbeat(const struct intel_engine_cs *engine)
288 {
289         /*
290          * For non-GuC submission we expect the back-end to look at the
291          * heartbeat status of the actual physical engine that the work
292          * has been (or is being) scheduled on, so we should only reach
293          * here with GuC submission enabled.
294          */
295         GEM_BUG_ON(!intel_engine_uses_guc(engine));
296
297         return intel_guc_virtual_engine_has_heartbeat(engine);
298 }
299
300 static inline bool
301 intel_engine_has_heartbeat(const struct intel_engine_cs *engine)
302 {
303         if (!IS_ACTIVE(CONFIG_DRM_I915_HEARTBEAT_INTERVAL))
304                 return false;
305
306         if (intel_engine_is_virtual(engine))
307                 return intel_virtual_engine_has_heartbeat(engine);
308         else
309                 return READ_ONCE(engine->props.heartbeat_interval_ms);
310 }
311
312 static inline struct intel_engine_cs *
313 intel_engine_get_sibling(struct intel_engine_cs *engine, unsigned int sibling)
314 {
315         GEM_BUG_ON(!intel_engine_is_virtual(engine));
316         return engine->cops->get_sibling(engine, sibling);
317 }
318
319 static inline void
320 intel_engine_set_hung_context(struct intel_engine_cs *engine,
321                               struct intel_context *ce)
322 {
323         engine->hung_ce = ce;
324 }
325
326 static inline void
327 intel_engine_clear_hung_context(struct intel_engine_cs *engine)
328 {
329         intel_engine_set_hung_context(engine, NULL);
330 }
331
332 static inline struct intel_context *
333 intel_engine_get_hung_context(struct intel_engine_cs *engine)
334 {
335         return engine->hung_ce;
336 }
337
338 #endif /* _INTEL_RINGBUFFER_H_ */