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 "gt/intel_timeline.h"
  17 #include "intel_engine_types.h"
  18 #include "intel_gpu_commands.h"
  19 #include "intel_workarounds.h"
  20
  21 struct drm_printer;
  22 struct intel_gt;
  23
  24 /* Early gen2 devices have a cacheline of just 32 bytes, using 64 is overkill,
  25  * but keeps the logic simple. Indeed, the whole purpose of this macro is just
  26  * to give some inclination as to some of the magic values used in the various
  27  * workarounds!
  28  */
  29 #define CACHELINE_BYTES 64
  30 #define CACHELINE_DWORDS (CACHELINE_BYTES / sizeof(u32))
  31
  32 #define ENGINE_TRACE(e, fmt, ...) do {                                  \
  33         const struct intel_engine_cs *e__ __maybe_unused = (e);         \
  34         GEM_TRACE("%s %s: " fmt,                                        \
  35                   dev_name(e__->i915->drm.dev), e__->name,              \
  36                   ##__VA_ARGS__);                                       \
  37 } while (0)
  38
  39 /*
  40  * The register defines to be used with the following macros need to accept a
  41  * base param, e.g:
  42  *
  43  * REG_FOO(base) _MMIO((base) + <relative offset>)
  44  * ENGINE_READ(engine, REG_FOO);
  45  *
  46  * register arrays are to be defined and accessed as follows:
  47  *
  48  * REG_BAR(base, i) _MMIO((base) + <relative offset> + (i) * <shift>)
  49  * ENGINE_READ_IDX(engine, REG_BAR, i)
  50  */
  51
  52 #define __ENGINE_REG_OP(op__, engine__, ...) \
  53         intel_uncore_##op__((engine__)->uncore, __VA_ARGS__)
  54
  55 #define __ENGINE_READ_OP(op__, engine__, reg__) \
  56         __ENGINE_REG_OP(op__, (engine__), reg__((engine__)->mmio_base))
  57
  58 #define ENGINE_READ16(...)      __ENGINE_READ_OP(read16, __VA_ARGS__)
  59 #define ENGINE_READ(...)        __ENGINE_READ_OP(read, __VA_ARGS__)
  60 #define ENGINE_READ_FW(...)     __ENGINE_READ_OP(read_fw, __VA_ARGS__)
  61 #define ENGINE_POSTING_READ(...) __ENGINE_READ_OP(posting_read_fw, __VA_ARGS__)
  62 #define ENGINE_POSTING_READ16(...) __ENGINE_READ_OP(posting_read16, __VA_ARGS__)
  63
  64 #define ENGINE_READ64(engine__, lower_reg__, upper_reg__) \
  65         __ENGINE_REG_OP(read64_2x32, (engine__), \
  66                         lower_reg__((engine__)->mmio_base), \
  67                         upper_reg__((engine__)->mmio_base))
  68
  69 #define ENGINE_READ_IDX(engine__, reg__, idx__) \
  70         __ENGINE_REG_OP(read, (engine__), reg__((engine__)->mmio_base, (idx__)))
  71
  72 #define __ENGINE_WRITE_OP(op__, engine__, reg__, val__) \
  73         __ENGINE_REG_OP(op__, (engine__), reg__((engine__)->mmio_base), (val__))
  74
  75 #define ENGINE_WRITE16(...)     __ENGINE_WRITE_OP(write16, __VA_ARGS__)
  76 #define ENGINE_WRITE(...)       __ENGINE_WRITE_OP(write, __VA_ARGS__)
  77 #define ENGINE_WRITE_FW(...)    __ENGINE_WRITE_OP(write_fw, __VA_ARGS__)
  78
  79 #define GEN6_RING_FAULT_REG_READ(engine__) \
  80         intel_uncore_read((engine__)->uncore, RING_FAULT_REG(engine__))
  81
  82 #define GEN6_RING_FAULT_REG_POSTING_READ(engine__) \
  83         intel_uncore_posting_read((engine__)->uncore, RING_FAULT_REG(engine__))
  84
  85 #define GEN6_RING_FAULT_REG_RMW(engine__, clear__, set__) \
  86 ({ \
  87         u32 __val; \
  88 \
  89         __val = intel_uncore_read((engine__)->uncore, \
  90                                   RING_FAULT_REG(engine__)); \
  91         __val &= ~(clear__); \
  92         __val |= (set__); \
  93         intel_uncore_write((engine__)->uncore, RING_FAULT_REG(engine__), \
  94                            __val); \
  95 })
  96
  97 /* seqno size is actually only a uint32, but since we plan to use MI_FLUSH_DW to
  98  * do the writes, and that must have qw aligned offsets, simply pretend it's 8b.
  99  */
 100
 101 static inline unsigned int
 102 execlists_num_ports(const struct intel_engine_execlists * const execlists)
 103 {
 104         return execlists->port_mask + 1;
 105 }
 106
 107 static inline struct i915_request *
 108 execlists_active(const struct intel_engine_execlists *execlists)
 109 {
 110         struct i915_request * const *cur, * const *old, *active;
 111
 112         cur = READ_ONCE(execlists->active);
 113         smp_rmb(); /* pairs with overwrite protection in process_csb() */
 114         do {
 115                 old = cur;
 116
 117                 active = READ_ONCE(*cur);
 118                 cur = READ_ONCE(execlists->active);
 119
 120                 smp_rmb(); /* and complete the seqlock retry */
 121         } while (unlikely(cur != old));
 122
 123         return active;
 124 }
 125
 126 static inline void
 127 execlists_active_lock_bh(struct intel_engine_execlists *execlists)
 128 {
 129         local_bh_disable(); /* prevent local softirq and lock recursion */
 130         tasklet_lock(&execlists->tasklet);
 131 }
 132
 133 static inline void
 134 execlists_active_unlock_bh(struct intel_engine_execlists *execlists)
 135 {
 136         tasklet_unlock(&execlists->tasklet);
 137         local_bh_enable(); /* restore softirq, and kick ksoftirqd! */
 138 }
 139
 140 struct i915_request *
 141 execlists_unwind_incomplete_requests(struct intel_engine_execlists *execlists);
 142
 143 static inline u32
 144 intel_read_status_page(const struct intel_engine_cs *engine, int reg)
 145 {
 146         /* Ensure that the compiler doesn't optimize away the load. */
 147         return READ_ONCE(engine->status_page.addr[reg]);
 148 }
 149
 150 static inline void
 151 intel_write_status_page(struct intel_engine_cs *engine, int reg, u32 value)
 152 {
 153         /* Writing into the status page should be done sparingly. Since
 154          * we do when we are uncertain of the device state, we take a bit
 155          * of extra paranoia to try and ensure that the HWS takes the value
 156          * we give and that it doesn't end up trapped inside the CPU!
 157          */
 158         if (static_cpu_has(X86_FEATURE_CLFLUSH)) {
 159                 mb();
 160                 clflush(&engine->status_page.addr[reg]);
 161                 engine->status_page.addr[reg] = value;
 162                 clflush(&engine->status_page.addr[reg]);
 163                 mb();
 164         } else {
 165                 WRITE_ONCE(engine->status_page.addr[reg], value);
 166         }
 167 }
 168
 169 /*
 170  * Reads a dword out of the status page, which is written to from the command
 171  * queue by automatic updates, MI_REPORT_HEAD, MI_STORE_DATA_INDEX, or
 172  * MI_STORE_DATA_IMM.
 173  *
 174  * The following dwords have a reserved meaning:
 175  * 0x00: ISR copy, updated when an ISR bit not set in the HWSTAM changes.
 176  * 0x04: ring 0 head pointer
 177  * 0x05: ring 1 head pointer (915-class)
 178  * 0x06: ring 2 head pointer (915-class)
 179  * 0x10-0x1b: Context status DWords (GM45)
 180  * 0x1f: Last written status offset. (GM45)
 181  * 0x20-0x2f: Reserved (Gen6+)
 182  *
 183  * The area from dword 0x30 to 0x3ff is available for driver usage.
 184  */
 185 #define I915_GEM_HWS_PREEMPT            0x32
 186 #define I915_GEM_HWS_PREEMPT_ADDR       (I915_GEM_HWS_PREEMPT * sizeof(u32))
 187 #define I915_GEM_HWS_SEQNO              0x40
 188 #define I915_GEM_HWS_SEQNO_ADDR         (I915_GEM_HWS_SEQNO * sizeof(u32))
 189 #define I915_GEM_HWS_SCRATCH            0x80
 190
 191 #define I915_HWS_CSB_BUF0_INDEX         0x10
 192 #define I915_HWS_CSB_WRITE_INDEX        0x1f
 193 #define CNL_HWS_CSB_WRITE_INDEX         0x2f
 194
 195 void intel_engine_stop(struct intel_engine_cs *engine);
 196 void intel_engine_cleanup(struct intel_engine_cs *engine);
 197
 198 int intel_engines_init_mmio(struct intel_gt *gt);
 199 int intel_engines_init(struct intel_gt *gt);
 200
 201 void intel_engine_free_request_pool(struct intel_engine_cs *engine);
 202
 203 void intel_engines_release(struct intel_gt *gt);
 204 void intel_engines_free(struct intel_gt *gt);
 205
 206 int intel_engine_init_common(struct intel_engine_cs *engine);
 207 void intel_engine_cleanup_common(struct intel_engine_cs *engine);
 208
 209 int intel_engine_resume(struct intel_engine_cs *engine);
 210
 211 int intel_ring_submission_setup(struct intel_engine_cs *engine);
 212
 213 int intel_engine_stop_cs(struct intel_engine_cs *engine);
 214 void intel_engine_cancel_stop_cs(struct intel_engine_cs *engine);
 215
 216 void intel_engine_set_hwsp_writemask(struct intel_engine_cs *engine, u32 mask);
 217
 218 u64 intel_engine_get_active_head(const struct intel_engine_cs *engine);
 219 u64 intel_engine_get_last_batch_head(const struct intel_engine_cs *engine);
 220
 221 void intel_engine_get_instdone(const struct intel_engine_cs *engine,
 222                                struct intel_instdone *instdone);
 223
 224 void intel_engine_init_execlists(struct intel_engine_cs *engine);
 225
 226 static inline u32 *__gen8_emit_pipe_control(u32 *batch, u32 flags0, u32 flags1, u32 offset)
 227 {
 228         memset(batch, 0, 6 * sizeof(u32));
 229
 230         batch[0] = GFX_OP_PIPE_CONTROL(6) | flags0;
 231         batch[1] = flags1;
 232         batch[2] = offset;
 233
 234         return batch + 6;
 235 }
 236
 237 static inline u32 *gen8_emit_pipe_control(u32 *batch, u32 flags, u32 offset)
 238 {
 239         return __gen8_emit_pipe_control(batch, 0, flags, offset);
 240 }
 241
 242 static inline u32 *gen12_emit_pipe_control(u32 *batch, u32 flags0, u32 flags1, u32 offset)
 243 {
 244         return __gen8_emit_pipe_control(batch, flags0, flags1, offset);
 245 }
 246
 247 static inline u32 *
 248 __gen8_emit_ggtt_write_rcs(u32 *cs, u32 value, u32 gtt_offset, u32 flags0, u32 flags1)
 249 {
 250         /* We're using qword write, offset should be aligned to 8 bytes. */
 251         GEM_BUG_ON(!IS_ALIGNED(gtt_offset, 8));
 252
 253         /* w/a for post sync ops following a GPGPU operation we
 254          * need a prior CS_STALL, which is emitted by the flush
 255          * following the batch.
 256          */
 257         *cs++ = GFX_OP_PIPE_CONTROL(6) | flags0;
 258         *cs++ = flags1 | PIPE_CONTROL_QW_WRITE | PIPE_CONTROL_GLOBAL_GTT_IVB;
 259         *cs++ = gtt_offset;
 260         *cs++ = 0;
 261         *cs++ = value;
 262         /* We're thrashing one dword of HWS. */
 263         *cs++ = 0;
 264
 265         return cs;
 266 }
 267
 268 static inline u32*
 269 gen8_emit_ggtt_write_rcs(u32 *cs, u32 value, u32 gtt_offset, u32 flags)
 270 {
 271         return __gen8_emit_ggtt_write_rcs(cs, value, gtt_offset, 0, flags);
 272 }
 273
 274 static inline u32*
 275 gen12_emit_ggtt_write_rcs(u32 *cs, u32 value, u32 gtt_offset, u32 flags0, u32 flags1)
 276 {
 277         return __gen8_emit_ggtt_write_rcs(cs, value, gtt_offset, flags0, flags1);
 278 }
 279
 280 static inline u32 *
 281 gen8_emit_ggtt_write(u32 *cs, u32 value, u32 gtt_offset, u32 flags)
 282 {
 283         /* w/a: bit 5 needs to be zero for MI_FLUSH_DW address. */
 284         GEM_BUG_ON(gtt_offset & (1 << 5));
 285         /* Offset should be aligned to 8 bytes for both (QW/DW) write types */
 286         GEM_BUG_ON(!IS_ALIGNED(gtt_offset, 8));
 287
 288         *cs++ = (MI_FLUSH_DW + 1) | MI_FLUSH_DW_OP_STOREDW | flags;
 289         *cs++ = gtt_offset | MI_FLUSH_DW_USE_GTT;
 290         *cs++ = 0;
 291         *cs++ = value;
 292
 293         return cs;
 294 }
 295
 296 static inline void __intel_engine_reset(struct intel_engine_cs *engine,
 297                                         bool stalled)
 298 {
 299         if (engine->reset.rewind)
 300                 engine->reset.rewind(engine, stalled);
 301         engine->serial++; /* contexts lost */
 302 }
 303
 304 bool intel_engines_are_idle(struct intel_gt *gt);
 305 bool intel_engine_is_idle(struct intel_engine_cs *engine);
 306 void intel_engine_flush_submission(struct intel_engine_cs *engine);
 307
 308 void intel_engines_reset_default_submission(struct intel_gt *gt);
 309
 310 bool intel_engine_can_store_dword(struct intel_engine_cs *engine);
 311
 312 __printf(3, 4)
 313 void intel_engine_dump(struct intel_engine_cs *engine,
 314                        struct drm_printer *m,
 315                        const char *header, ...);
 316
 317 ktime_t intel_engine_get_busy_time(struct intel_engine_cs *engine,
 318                                    ktime_t *now);
 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 void intel_engine_init_active(struct intel_engine_cs *engine,
 326                               unsigned int subclass);
 327 #define ENGINE_PHYSICAL 0
 328 #define ENGINE_MOCK     1
 329 #define ENGINE_VIRTUAL  2
 330
 331 static inline bool
 332 intel_engine_has_preempt_reset(const struct intel_engine_cs *engine)
 333 {
 334         if (!IS_ACTIVE(CONFIG_DRM_I915_PREEMPT_TIMEOUT))
 335                 return false;
 336
 337         return intel_engine_has_preemption(engine);
 338 }
 339
 340 static inline bool
 341 intel_engine_has_heartbeat(const struct intel_engine_cs *engine)
 342 {
 343         if (!IS_ACTIVE(CONFIG_DRM_I915_HEARTBEAT_INTERVAL))
 344                 return false;
 345
 346         return READ_ONCE(engine->props.heartbeat_interval_ms);
 347 }
 348
 349 #endif /* _INTEL_RINGBUFFER_H_ */