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 /* 33 * The register defines to be used with the following macros need to accept a 34 * base param, e.g: 35 * 36 * REG_FOO(base) _MMIO((base) + <relative offset>) 37 * ENGINE_READ(engine, REG_FOO); 38 * 39 * register arrays are to be defined and accessed as follows: 40 * 41 * REG_BAR(base, i) _MMIO((base) + <relative offset> + (i) * <shift>) 42 * ENGINE_READ_IDX(engine, REG_BAR, i) 43 */ 44 45 #define __ENGINE_REG_OP(op__, engine__, ...) \ 46 intel_uncore_##op__((engine__)->uncore, __VA_ARGS__) 47 48 #define __ENGINE_READ_OP(op__, engine__, reg__) \ 49 __ENGINE_REG_OP(op__, (engine__), reg__((engine__)->mmio_base)) 50 51 #define ENGINE_READ16(...) __ENGINE_READ_OP(read16, __VA_ARGS__) 52 #define ENGINE_READ(...) __ENGINE_READ_OP(read, __VA_ARGS__) 53 #define ENGINE_READ_FW(...) __ENGINE_READ_OP(read_fw, __VA_ARGS__) 54 #define ENGINE_POSTING_READ(...) __ENGINE_READ_OP(posting_read_fw, __VA_ARGS__) 55 #define ENGINE_POSTING_READ16(...) __ENGINE_READ_OP(posting_read16, __VA_ARGS__) 56 57 #define ENGINE_READ64(engine__, lower_reg__, upper_reg__) \ 58 __ENGINE_REG_OP(read64_2x32, (engine__), \ 59 lower_reg__((engine__)->mmio_base), \ 60 upper_reg__((engine__)->mmio_base)) 61 62 #define ENGINE_READ_IDX(engine__, reg__, idx__) \ 63 __ENGINE_REG_OP(read, (engine__), reg__((engine__)->mmio_base, (idx__))) 64 65 #define __ENGINE_WRITE_OP(op__, engine__, reg__, val__) \ 66 __ENGINE_REG_OP(op__, (engine__), reg__((engine__)->mmio_base), (val__)) 67 68 #define ENGINE_WRITE16(...) __ENGINE_WRITE_OP(write16, __VA_ARGS__) 69 #define ENGINE_WRITE(...) __ENGINE_WRITE_OP(write, __VA_ARGS__) 70 #define ENGINE_WRITE_FW(...) __ENGINE_WRITE_OP(write_fw, __VA_ARGS__) 71 72 #define GEN6_RING_FAULT_REG_READ(engine__) \ 73 intel_uncore_read((engine__)->uncore, RING_FAULT_REG(engine__)) 74 75 #define GEN6_RING_FAULT_REG_POSTING_READ(engine__) \ 76 intel_uncore_posting_read((engine__)->uncore, RING_FAULT_REG(engine__)) 77 78 #define GEN6_RING_FAULT_REG_RMW(engine__, clear__, set__) \ 79 ({ \ 80 u32 __val; \ 81 \ 82 __val = intel_uncore_read((engine__)->uncore, \ 83 RING_FAULT_REG(engine__)); \ 84 __val &= ~(clear__); \ 85 __val |= (set__); \ 86 intel_uncore_write((engine__)->uncore, RING_FAULT_REG(engine__), \ 87 __val); \ 88 }) 89 90 /* seqno size is actually only a uint32, but since we plan to use MI_FLUSH_DW to 91 * do the writes, and that must have qw aligned offsets, simply pretend it's 8b. 92 */ 93 94 static inline unsigned int 95 execlists_num_ports(const struct intel_engine_execlists * const execlists) 96 { 97 return execlists->port_mask + 1; 98 } 99 100 static inline struct i915_request * 101 execlists_active(const struct intel_engine_execlists *execlists) 102 { 103 GEM_BUG_ON(execlists->active - execlists->inflight > 104 execlists_num_ports(execlists)); 105 return READ_ONCE(*execlists->active); 106 } 107 108 static inline void 109 execlists_active_lock_bh(struct intel_engine_execlists *execlists) 110 { 111 local_bh_disable(); /* prevent local softirq and lock recursion */ 112 tasklet_lock(&execlists->tasklet); 113 } 114 115 static inline void 116 execlists_active_unlock_bh(struct intel_engine_execlists *execlists) 117 { 118 tasklet_unlock(&execlists->tasklet); 119 local_bh_enable(); /* restore softirq, and kick ksoftirqd! */ 120 } 121 122 struct i915_request * 123 execlists_unwind_incomplete_requests(struct intel_engine_execlists *execlists); 124 125 static inline u32 126 intel_read_status_page(const struct intel_engine_cs *engine, int reg) 127 { 128 /* Ensure that the compiler doesn't optimize away the load. */ 129 return READ_ONCE(engine->status_page.addr[reg]); 130 } 131 132 static inline void 133 intel_write_status_page(struct intel_engine_cs *engine, int reg, u32 value) 134 { 135 /* Writing into the status page should be done sparingly. Since 136 * we do when we are uncertain of the device state, we take a bit 137 * of extra paranoia to try and ensure that the HWS takes the value 138 * we give and that it doesn't end up trapped inside the CPU! 139 */ 140 if (static_cpu_has(X86_FEATURE_CLFLUSH)) { 141 mb(); 142 clflush(&engine->status_page.addr[reg]); 143 engine->status_page.addr[reg] = value; 144 clflush(&engine->status_page.addr[reg]); 145 mb(); 146 } else { 147 WRITE_ONCE(engine->status_page.addr[reg], value); 148 } 149 } 150 151 /* 152 * Reads a dword out of the status page, which is written to from the command 153 * queue by automatic updates, MI_REPORT_HEAD, MI_STORE_DATA_INDEX, or 154 * MI_STORE_DATA_IMM. 155 * 156 * The following dwords have a reserved meaning: 157 * 0x00: ISR copy, updated when an ISR bit not set in the HWSTAM changes. 158 * 0x04: ring 0 head pointer 159 * 0x05: ring 1 head pointer (915-class) 160 * 0x06: ring 2 head pointer (915-class) 161 * 0x10-0x1b: Context status DWords (GM45) 162 * 0x1f: Last written status offset. (GM45) 163 * 0x20-0x2f: Reserved (Gen6+) 164 * 165 * The area from dword 0x30 to 0x3ff is available for driver usage. 166 */ 167 #define I915_GEM_HWS_PREEMPT 0x32 168 #define I915_GEM_HWS_PREEMPT_ADDR (I915_GEM_HWS_PREEMPT * sizeof(u32)) 169 #define I915_GEM_HWS_SEQNO 0x40 170 #define I915_GEM_HWS_SEQNO_ADDR (I915_GEM_HWS_SEQNO * sizeof(u32)) 171 #define I915_GEM_HWS_SCRATCH 0x80 172 #define I915_GEM_HWS_SCRATCH_ADDR (I915_GEM_HWS_SCRATCH * sizeof(u32)) 173 174 #define I915_HWS_CSB_BUF0_INDEX 0x10 175 #define I915_HWS_CSB_WRITE_INDEX 0x1f 176 #define CNL_HWS_CSB_WRITE_INDEX 0x2f 177 178 void intel_engine_stop(struct intel_engine_cs *engine); 179 void intel_engine_cleanup(struct intel_engine_cs *engine); 180 181 int intel_engines_init_mmio(struct intel_gt *gt); 182 int intel_engines_setup(struct intel_gt *gt); 183 int intel_engines_init(struct intel_gt *gt); 184 void intel_engines_cleanup(struct intel_gt *gt); 185 186 int intel_engine_init_common(struct intel_engine_cs *engine); 187 void intel_engine_cleanup_common(struct intel_engine_cs *engine); 188 189 int intel_ring_submission_setup(struct intel_engine_cs *engine); 190 int intel_ring_submission_init(struct intel_engine_cs *engine); 191 192 int intel_engine_stop_cs(struct intel_engine_cs *engine); 193 void intel_engine_cancel_stop_cs(struct intel_engine_cs *engine); 194 195 void intel_engine_set_hwsp_writemask(struct intel_engine_cs *engine, u32 mask); 196 197 u64 intel_engine_get_active_head(const struct intel_engine_cs *engine); 198 u64 intel_engine_get_last_batch_head(const struct intel_engine_cs *engine); 199 200 void intel_engine_get_instdone(struct intel_engine_cs *engine, 201 struct intel_instdone *instdone); 202 203 void intel_engine_init_execlists(struct intel_engine_cs *engine); 204 205 void intel_engine_init_breadcrumbs(struct intel_engine_cs *engine); 206 void intel_engine_fini_breadcrumbs(struct intel_engine_cs *engine); 207 208 void intel_engine_disarm_breadcrumbs(struct intel_engine_cs *engine); 209 210 static inline void 211 intel_engine_queue_breadcrumbs(struct intel_engine_cs *engine) 212 { 213 irq_work_queue(&engine->breadcrumbs.irq_work); 214 } 215 216 void intel_engine_breadcrumbs_irq(struct intel_engine_cs *engine); 217 218 void intel_engine_reset_breadcrumbs(struct intel_engine_cs *engine); 219 void intel_engine_fini_breadcrumbs(struct intel_engine_cs *engine); 220 221 void intel_engine_print_breadcrumbs(struct intel_engine_cs *engine, 222 struct drm_printer *p); 223 224 static inline u32 *gen8_emit_pipe_control(u32 *batch, u32 flags, u32 offset) 225 { 226 memset(batch, 0, 6 * sizeof(u32)); 227 228 batch[0] = GFX_OP_PIPE_CONTROL(6); 229 batch[1] = flags; 230 batch[2] = offset; 231 232 return batch + 6; 233 } 234 235 static inline u32 * 236 gen8_emit_ggtt_write_rcs(u32 *cs, u32 value, u32 gtt_offset, u32 flags) 237 { 238 /* We're using qword write, offset should be aligned to 8 bytes. */ 239 GEM_BUG_ON(!IS_ALIGNED(gtt_offset, 8)); 240 241 /* w/a for post sync ops following a GPGPU operation we 242 * need a prior CS_STALL, which is emitted by the flush 243 * following the batch. 244 */ 245 *cs++ = GFX_OP_PIPE_CONTROL(6); 246 *cs++ = flags | PIPE_CONTROL_QW_WRITE | PIPE_CONTROL_GLOBAL_GTT_IVB; 247 *cs++ = gtt_offset; 248 *cs++ = 0; 249 *cs++ = value; 250 /* We're thrashing one dword of HWS. */ 251 *cs++ = 0; 252 253 return cs; 254 } 255 256 static inline u32 * 257 gen8_emit_ggtt_write(u32 *cs, u32 value, u32 gtt_offset, u32 flags) 258 { 259 /* w/a: bit 5 needs to be zero for MI_FLUSH_DW address. */ 260 GEM_BUG_ON(gtt_offset & (1 << 5)); 261 /* Offset should be aligned to 8 bytes for both (QW/DW) write types */ 262 GEM_BUG_ON(!IS_ALIGNED(gtt_offset, 8)); 263 264 *cs++ = (MI_FLUSH_DW + 1) | MI_FLUSH_DW_OP_STOREDW | flags; 265 *cs++ = gtt_offset | MI_FLUSH_DW_USE_GTT; 266 *cs++ = 0; 267 *cs++ = value; 268 269 return cs; 270 } 271 272 static inline void __intel_engine_reset(struct intel_engine_cs *engine, 273 bool stalled) 274 { 275 if (engine->reset.reset) 276 engine->reset.reset(engine, stalled); 277 engine->serial++; /* contexts lost */ 278 } 279 280 bool intel_engines_are_idle(struct intel_gt *gt); 281 bool intel_engine_is_idle(struct intel_engine_cs *engine); 282 void intel_engine_flush_submission(struct intel_engine_cs *engine); 283 284 void intel_engines_reset_default_submission(struct intel_gt *gt); 285 286 bool intel_engine_can_store_dword(struct intel_engine_cs *engine); 287 288 __printf(3, 4) 289 void intel_engine_dump(struct intel_engine_cs *engine, 290 struct drm_printer *m, 291 const char *header, ...); 292 293 int intel_enable_engine_stats(struct intel_engine_cs *engine); 294 void intel_disable_engine_stats(struct intel_engine_cs *engine); 295 296 ktime_t intel_engine_get_busy_time(struct intel_engine_cs *engine); 297 298 struct i915_request * 299 intel_engine_find_active_request(struct intel_engine_cs *engine); 300 301 u32 intel_engine_context_size(struct drm_i915_private *i915, u8 class); 302 303 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) 304 305 static inline bool inject_preempt_hang(struct intel_engine_execlists *execlists) 306 { 307 if (!execlists->preempt_hang.inject_hang) 308 return false; 309 310 complete(&execlists->preempt_hang.completion); 311 return true; 312 } 313 314 #else 315 316 static inline bool inject_preempt_hang(struct intel_engine_execlists *execlists) 317 { 318 return false; 319 } 320 321 #endif 322 323 void intel_engine_init_active(struct intel_engine_cs *engine, 324 unsigned int subclass); 325 #define ENGINE_PHYSICAL 0 326 #define ENGINE_MOCK 1 327 #define ENGINE_VIRTUAL 2 328 329 static inline bool 330 intel_engine_has_preempt_reset(const struct intel_engine_cs *engine) 331 { 332 if (!IS_ACTIVE(CONFIG_DRM_I915_PREEMPT_TIMEOUT)) 333 return false; 334 335 return intel_engine_has_preemption(engine); 336 } 337 338 static inline bool 339 intel_engine_has_timeslices(const struct intel_engine_cs *engine) 340 { 341 if (!IS_ACTIVE(CONFIG_DRM_I915_TIMESLICE_DURATION)) 342 return false; 343 344 return intel_engine_has_semaphores(engine); 345 } 346 347 #endif /* _INTEL_RINGBUFFER_H_ */ 348