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 22struct drm_printer; 23struct intel_context; 24struct intel_gt; 25struct 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 104static inline unsigned int 105execlists_num_ports(const struct intel_engine_execlists * const execlists) 106{ 107 return execlists->port_mask + 1; 108} 109 110static inline struct i915_request * 111execlists_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 129struct i915_request * 130execlists_unwind_incomplete_requests(struct intel_engine_execlists *execlists); 131 132static inline u32 133intel_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 139static inline void 140intel_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_PXP 0x60 174#define I915_GEM_HWS_PXP_ADDR (I915_GEM_HWS_PXP * sizeof(u32)) 175#define I915_GEM_HWS_GSC 0x62 176#define I915_GEM_HWS_GSC_ADDR (I915_GEM_HWS_GSC * sizeof(u32)) 177#define I915_GEM_HWS_SCRATCH 0x80 178 179#define I915_HWS_CSB_BUF0_INDEX 0x10 180#define I915_HWS_CSB_WRITE_INDEX 0x1f 181#define ICL_HWS_CSB_WRITE_INDEX 0x2f 182#define INTEL_HWS_CSB_WRITE_INDEX(__i915) \ 183 (GRAPHICS_VER(__i915) >= 11 ? ICL_HWS_CSB_WRITE_INDEX : I915_HWS_CSB_WRITE_INDEX) 184 185void intel_engine_stop(struct intel_engine_cs *engine); 186void intel_engine_cleanup(struct intel_engine_cs *engine); 187 188int intel_engines_init_mmio(struct intel_gt *gt); 189int intel_engines_init(struct intel_gt *gt); 190 191void intel_engine_free_request_pool(struct intel_engine_cs *engine); 192 193void intel_engines_release(struct intel_gt *gt); 194void intel_engines_free(struct intel_gt *gt); 195 196int intel_engine_init_common(struct intel_engine_cs *engine); 197void intel_engine_cleanup_common(struct intel_engine_cs *engine); 198 199int intel_engine_resume(struct intel_engine_cs *engine); 200 201int intel_ring_submission_setup(struct intel_engine_cs *engine); 202 203int intel_engine_stop_cs(struct intel_engine_cs *engine); 204void intel_engine_cancel_stop_cs(struct intel_engine_cs *engine); 205 206void intel_engine_wait_for_pending_mi_fw(struct intel_engine_cs *engine); 207 208void intel_engine_set_hwsp_writemask(struct intel_engine_cs *engine, u32 mask); 209 210u64 intel_engine_get_active_head(const struct intel_engine_cs *engine); 211u64 intel_engine_get_last_batch_head(const struct intel_engine_cs *engine); 212 213void intel_engine_get_instdone(const struct intel_engine_cs *engine, 214 struct intel_instdone *instdone); 215 216void intel_engine_init_execlists(struct intel_engine_cs *engine); 217 218bool intel_engine_irq_enable(struct intel_engine_cs *engine); 219void intel_engine_irq_disable(struct intel_engine_cs *engine); 220 221static inline void __intel_engine_reset(struct intel_engine_cs *engine, 222 bool stalled) 223{ 224 if (engine->reset.rewind) 225 engine->reset.rewind(engine, stalled); 226 engine->serial++; /* contexts lost */ 227} 228 229bool intel_engines_are_idle(struct intel_gt *gt); 230bool intel_engine_is_idle(struct intel_engine_cs *engine); 231 232void __intel_engine_flush_submission(struct intel_engine_cs *engine, bool sync); 233static inline void intel_engine_flush_submission(struct intel_engine_cs *engine) 234{ 235 __intel_engine_flush_submission(engine, true); 236} 237 238void intel_engines_reset_default_submission(struct intel_gt *gt); 239 240bool intel_engine_can_store_dword(struct intel_engine_cs *engine); 241 242__printf(3, 4) 243void intel_engine_dump(struct intel_engine_cs *engine, 244 struct drm_printer *m, 245 const char *header, ...); 246void intel_engine_dump_active_requests(struct list_head *requests, 247 struct i915_request *hung_rq, 248 struct drm_printer *m); 249 250ktime_t intel_engine_get_busy_time(struct intel_engine_cs *engine, 251 ktime_t *now); 252 253void intel_engine_get_hung_entity(struct intel_engine_cs *engine, 254 struct intel_context **ce, struct i915_request **rq); 255 256u32 intel_engine_context_size(struct intel_gt *gt, u8 class); 257struct intel_context * 258intel_engine_create_pinned_context(struct intel_engine_cs *engine, 259 struct i915_address_space *vm, 260 unsigned int ring_size, 261 unsigned int hwsp, 262 struct lock_class_key *key, 263 const char *name); 264 265void intel_engine_destroy_pinned_context(struct intel_context *ce); 266 267void xehp_enable_ccs_engines(struct intel_engine_cs *engine); 268 269#define ENGINE_PHYSICAL 0 270#define ENGINE_MOCK 1 271#define ENGINE_VIRTUAL 2 272 273static inline bool intel_engine_uses_guc(const struct intel_engine_cs *engine) 274{ 275 return engine->gt->submission_method >= INTEL_SUBMISSION_GUC; 276} 277 278static inline bool 279intel_engine_has_preempt_reset(const struct intel_engine_cs *engine) 280{ 281 if (!CONFIG_DRM_I915_PREEMPT_TIMEOUT) 282 return false; 283 284 return intel_engine_has_preemption(engine); 285} 286 287#define FORCE_VIRTUAL BIT(0) 288struct intel_context * 289intel_engine_create_virtual(struct intel_engine_cs **siblings, 290 unsigned int count, unsigned long flags); 291 292static inline struct intel_context * 293intel_engine_create_parallel(struct intel_engine_cs **engines, 294 unsigned int num_engines, 295 unsigned int width) 296{ 297 GEM_BUG_ON(!engines[0]->cops->create_parallel); 298 return engines[0]->cops->create_parallel(engines, num_engines, width); 299} 300 301static inline bool 302intel_virtual_engine_has_heartbeat(const struct intel_engine_cs *engine) 303{ 304 /* 305 * For non-GuC submission we expect the back-end to look at the 306 * heartbeat status of the actual physical engine that the work 307 * has been (or is being) scheduled on, so we should only reach 308 * here with GuC submission enabled. 309 */ 310 GEM_BUG_ON(!intel_engine_uses_guc(engine)); 311 312 return intel_guc_virtual_engine_has_heartbeat(engine); 313} 314 315static inline bool 316intel_engine_has_heartbeat(const struct intel_engine_cs *engine) 317{ 318 if (!CONFIG_DRM_I915_HEARTBEAT_INTERVAL) 319 return false; 320 321 if (intel_engine_is_virtual(engine)) 322 return intel_virtual_engine_has_heartbeat(engine); 323 else 324 return READ_ONCE(engine->props.heartbeat_interval_ms); 325} 326 327static inline struct intel_engine_cs * 328intel_engine_get_sibling(struct intel_engine_cs *engine, unsigned int sibling) 329{ 330 GEM_BUG_ON(!intel_engine_is_virtual(engine)); 331 return engine->cops->get_sibling(engine, sibling); 332} 333 334static inline void 335intel_engine_set_hung_context(struct intel_engine_cs *engine, 336 struct intel_context *ce) 337{ 338 engine->hung_ce = ce; 339} 340 341static inline void 342intel_engine_clear_hung_context(struct intel_engine_cs *engine) 343{ 344 intel_engine_set_hung_context(engine, NULL); 345} 346 347static inline struct intel_context * 348intel_engine_get_hung_context(struct intel_engine_cs *engine) 349{ 350 return engine->hung_ce; 351} 352 353u64 intel_clamp_heartbeat_interval_ms(struct intel_engine_cs *engine, u64 value); 354u64 intel_clamp_max_busywait_duration_ns(struct intel_engine_cs *engine, u64 value); 355u64 intel_clamp_preempt_timeout_ms(struct intel_engine_cs *engine, u64 value); 356u64 intel_clamp_stop_timeout_ms(struct intel_engine_cs *engine, u64 value); 357u64 intel_clamp_timeslice_duration_ms(struct intel_engine_cs *engine, u64 value); 358 359#endif /* _INTEL_RINGBUFFER_H_ */ 360