Loading...
1#ifndef _INTEL_RINGBUFFER_H_
2#define _INTEL_RINGBUFFER_H_
3
4#include <linux/hashtable.h>
5#include "i915_gem_batch_pool.h"
6
7#define I915_CMD_HASH_ORDER 9
8
9/* Early gen2 devices have a cacheline of just 32 bytes, using 64 is overkill,
10 * but keeps the logic simple. Indeed, the whole purpose of this macro is just
11 * to give some inclination as to some of the magic values used in the various
12 * workarounds!
13 */
14#define CACHELINE_BYTES 64
15#define CACHELINE_DWORDS (CACHELINE_BYTES / sizeof(uint32_t))
16
17/*
18 * Gen2 BSpec "1. Programming Environment" / 1.4.4.6 "Ring Buffer Use"
19 * Gen3 BSpec "vol1c Memory Interface Functions" / 2.3.4.5 "Ring Buffer Use"
20 * Gen4+ BSpec "vol1c Memory Interface and Command Stream" / 5.3.4.5 "Ring Buffer Use"
21 *
22 * "If the Ring Buffer Head Pointer and the Tail Pointer are on the same
23 * cacheline, the Head Pointer must not be greater than the Tail
24 * Pointer."
25 */
26#define I915_RING_FREE_SPACE 64
27
28struct intel_hw_status_page {
29 u32 *page_addr;
30 unsigned int gfx_addr;
31 struct drm_i915_gem_object *obj;
32};
33
34#define I915_READ_TAIL(ring) I915_READ(RING_TAIL((ring)->mmio_base))
35#define I915_WRITE_TAIL(ring, val) I915_WRITE(RING_TAIL((ring)->mmio_base), val)
36
37#define I915_READ_START(ring) I915_READ(RING_START((ring)->mmio_base))
38#define I915_WRITE_START(ring, val) I915_WRITE(RING_START((ring)->mmio_base), val)
39
40#define I915_READ_HEAD(ring) I915_READ(RING_HEAD((ring)->mmio_base))
41#define I915_WRITE_HEAD(ring, val) I915_WRITE(RING_HEAD((ring)->mmio_base), val)
42
43#define I915_READ_CTL(ring) I915_READ(RING_CTL((ring)->mmio_base))
44#define I915_WRITE_CTL(ring, val) I915_WRITE(RING_CTL((ring)->mmio_base), val)
45
46#define I915_READ_IMR(ring) I915_READ(RING_IMR((ring)->mmio_base))
47#define I915_WRITE_IMR(ring, val) I915_WRITE(RING_IMR((ring)->mmio_base), val)
48
49#define I915_READ_MODE(ring) I915_READ(RING_MI_MODE((ring)->mmio_base))
50#define I915_WRITE_MODE(ring, val) I915_WRITE(RING_MI_MODE((ring)->mmio_base), val)
51
52/* seqno size is actually only a uint32, but since we plan to use MI_FLUSH_DW to
53 * do the writes, and that must have qw aligned offsets, simply pretend it's 8b.
54 */
55#define i915_semaphore_seqno_size sizeof(uint64_t)
56#define GEN8_SIGNAL_OFFSET(__ring, to) \
57 (i915_gem_obj_ggtt_offset(dev_priv->semaphore_obj) + \
58 ((__ring)->id * I915_NUM_RINGS * i915_semaphore_seqno_size) + \
59 (i915_semaphore_seqno_size * (to)))
60
61#define GEN8_WAIT_OFFSET(__ring, from) \
62 (i915_gem_obj_ggtt_offset(dev_priv->semaphore_obj) + \
63 ((from) * I915_NUM_RINGS * i915_semaphore_seqno_size) + \
64 (i915_semaphore_seqno_size * (__ring)->id))
65
66#define GEN8_RING_SEMAPHORE_INIT do { \
67 if (!dev_priv->semaphore_obj) { \
68 break; \
69 } \
70 ring->semaphore.signal_ggtt[RCS] = GEN8_SIGNAL_OFFSET(ring, RCS); \
71 ring->semaphore.signal_ggtt[VCS] = GEN8_SIGNAL_OFFSET(ring, VCS); \
72 ring->semaphore.signal_ggtt[BCS] = GEN8_SIGNAL_OFFSET(ring, BCS); \
73 ring->semaphore.signal_ggtt[VECS] = GEN8_SIGNAL_OFFSET(ring, VECS); \
74 ring->semaphore.signal_ggtt[VCS2] = GEN8_SIGNAL_OFFSET(ring, VCS2); \
75 ring->semaphore.signal_ggtt[ring->id] = MI_SEMAPHORE_SYNC_INVALID; \
76 } while(0)
77
78enum intel_ring_hangcheck_action {
79 HANGCHECK_IDLE = 0,
80 HANGCHECK_WAIT,
81 HANGCHECK_ACTIVE,
82 HANGCHECK_ACTIVE_LOOP,
83 HANGCHECK_KICK,
84 HANGCHECK_HUNG,
85};
86
87#define HANGCHECK_SCORE_RING_HUNG 31
88
89struct intel_ring_hangcheck {
90 u64 acthd;
91 u64 max_acthd;
92 u32 seqno;
93 int score;
94 enum intel_ring_hangcheck_action action;
95 int deadlock;
96 u32 instdone[I915_NUM_INSTDONE_REG];
97};
98
99struct intel_ringbuffer {
100 struct drm_i915_gem_object *obj;
101 void __iomem *virtual_start;
102 struct i915_vma *vma;
103
104 struct intel_engine_cs *ring;
105 struct list_head link;
106
107 u32 head;
108 u32 tail;
109 int space;
110 int size;
111 int effective_size;
112 int reserved_size;
113 int reserved_tail;
114 bool reserved_in_use;
115
116 /** We track the position of the requests in the ring buffer, and
117 * when each is retired we increment last_retired_head as the GPU
118 * must have finished processing the request and so we know we
119 * can advance the ringbuffer up to that position.
120 *
121 * last_retired_head is set to -1 after the value is consumed so
122 * we can detect new retirements.
123 */
124 u32 last_retired_head;
125};
126
127struct intel_context;
128struct drm_i915_reg_descriptor;
129
130/*
131 * we use a single page to load ctx workarounds so all of these
132 * values are referred in terms of dwords
133 *
134 * struct i915_wa_ctx_bb:
135 * offset: specifies batch starting position, also helpful in case
136 * if we want to have multiple batches at different offsets based on
137 * some criteria. It is not a requirement at the moment but provides
138 * an option for future use.
139 * size: size of the batch in DWORDS
140 */
141struct i915_ctx_workarounds {
142 struct i915_wa_ctx_bb {
143 u32 offset;
144 u32 size;
145 } indirect_ctx, per_ctx;
146 struct drm_i915_gem_object *obj;
147};
148
149struct intel_engine_cs {
150 const char *name;
151 enum intel_ring_id {
152 RCS = 0,
153 BCS,
154 VCS,
155 VCS2, /* Keep instances of the same type engine together. */
156 VECS
157 } id;
158#define I915_NUM_RINGS 5
159#define _VCS(n) (VCS + (n))
160 unsigned int exec_id;
161 unsigned int guc_id;
162 u32 mmio_base;
163 struct drm_device *dev;
164 struct intel_ringbuffer *buffer;
165 struct list_head buffers;
166
167 /*
168 * A pool of objects to use as shadow copies of client batch buffers
169 * when the command parser is enabled. Prevents the client from
170 * modifying the batch contents after software parsing.
171 */
172 struct i915_gem_batch_pool batch_pool;
173
174 struct intel_hw_status_page status_page;
175 struct i915_ctx_workarounds wa_ctx;
176
177 unsigned irq_refcount; /* protected by dev_priv->irq_lock */
178 u32 irq_enable_mask; /* bitmask to enable ring interrupt */
179 struct drm_i915_gem_request *trace_irq_req;
180 bool __must_check (*irq_get)(struct intel_engine_cs *ring);
181 void (*irq_put)(struct intel_engine_cs *ring);
182
183 int (*init_hw)(struct intel_engine_cs *ring);
184
185 int (*init_context)(struct drm_i915_gem_request *req);
186
187 void (*write_tail)(struct intel_engine_cs *ring,
188 u32 value);
189 int __must_check (*flush)(struct drm_i915_gem_request *req,
190 u32 invalidate_domains,
191 u32 flush_domains);
192 int (*add_request)(struct drm_i915_gem_request *req);
193 /* Some chipsets are not quite as coherent as advertised and need
194 * an expensive kick to force a true read of the up-to-date seqno.
195 * However, the up-to-date seqno is not always required and the last
196 * seen value is good enough. Note that the seqno will always be
197 * monotonic, even if not coherent.
198 */
199 u32 (*get_seqno)(struct intel_engine_cs *ring,
200 bool lazy_coherency);
201 void (*set_seqno)(struct intel_engine_cs *ring,
202 u32 seqno);
203 int (*dispatch_execbuffer)(struct drm_i915_gem_request *req,
204 u64 offset, u32 length,
205 unsigned dispatch_flags);
206#define I915_DISPATCH_SECURE 0x1
207#define I915_DISPATCH_PINNED 0x2
208#define I915_DISPATCH_RS 0x4
209 void (*cleanup)(struct intel_engine_cs *ring);
210
211 /* GEN8 signal/wait table - never trust comments!
212 * signal to signal to signal to signal to signal to
213 * RCS VCS BCS VECS VCS2
214 * --------------------------------------------------------------------
215 * RCS | NOP (0x00) | VCS (0x08) | BCS (0x10) | VECS (0x18) | VCS2 (0x20) |
216 * |-------------------------------------------------------------------
217 * VCS | RCS (0x28) | NOP (0x30) | BCS (0x38) | VECS (0x40) | VCS2 (0x48) |
218 * |-------------------------------------------------------------------
219 * BCS | RCS (0x50) | VCS (0x58) | NOP (0x60) | VECS (0x68) | VCS2 (0x70) |
220 * |-------------------------------------------------------------------
221 * VECS | RCS (0x78) | VCS (0x80) | BCS (0x88) | NOP (0x90) | VCS2 (0x98) |
222 * |-------------------------------------------------------------------
223 * VCS2 | RCS (0xa0) | VCS (0xa8) | BCS (0xb0) | VECS (0xb8) | NOP (0xc0) |
224 * |-------------------------------------------------------------------
225 *
226 * Generalization:
227 * f(x, y) := (x->id * NUM_RINGS * seqno_size) + (seqno_size * y->id)
228 * ie. transpose of g(x, y)
229 *
230 * sync from sync from sync from sync from sync from
231 * RCS VCS BCS VECS VCS2
232 * --------------------------------------------------------------------
233 * RCS | NOP (0x00) | VCS (0x28) | BCS (0x50) | VECS (0x78) | VCS2 (0xa0) |
234 * |-------------------------------------------------------------------
235 * VCS | RCS (0x08) | NOP (0x30) | BCS (0x58) | VECS (0x80) | VCS2 (0xa8) |
236 * |-------------------------------------------------------------------
237 * BCS | RCS (0x10) | VCS (0x38) | NOP (0x60) | VECS (0x88) | VCS2 (0xb0) |
238 * |-------------------------------------------------------------------
239 * VECS | RCS (0x18) | VCS (0x40) | BCS (0x68) | NOP (0x90) | VCS2 (0xb8) |
240 * |-------------------------------------------------------------------
241 * VCS2 | RCS (0x20) | VCS (0x48) | BCS (0x70) | VECS (0x98) | NOP (0xc0) |
242 * |-------------------------------------------------------------------
243 *
244 * Generalization:
245 * g(x, y) := (y->id * NUM_RINGS * seqno_size) + (seqno_size * x->id)
246 * ie. transpose of f(x, y)
247 */
248 struct {
249 u32 sync_seqno[I915_NUM_RINGS-1];
250
251 union {
252 struct {
253 /* our mbox written by others */
254 u32 wait[I915_NUM_RINGS];
255 /* mboxes this ring signals to */
256 i915_reg_t signal[I915_NUM_RINGS];
257 } mbox;
258 u64 signal_ggtt[I915_NUM_RINGS];
259 };
260
261 /* AKA wait() */
262 int (*sync_to)(struct drm_i915_gem_request *to_req,
263 struct intel_engine_cs *from,
264 u32 seqno);
265 int (*signal)(struct drm_i915_gem_request *signaller_req,
266 /* num_dwords needed by caller */
267 unsigned int num_dwords);
268 } semaphore;
269
270 /* Execlists */
271 spinlock_t execlist_lock;
272 struct list_head execlist_queue;
273 struct list_head execlist_retired_req_list;
274 u8 next_context_status_buffer;
275 bool disable_lite_restore_wa;
276 u32 ctx_desc_template;
277 u32 irq_keep_mask; /* bitmask for interrupts that should not be masked */
278 int (*emit_request)(struct drm_i915_gem_request *request);
279 int (*emit_flush)(struct drm_i915_gem_request *request,
280 u32 invalidate_domains,
281 u32 flush_domains);
282 int (*emit_bb_start)(struct drm_i915_gem_request *req,
283 u64 offset, unsigned dispatch_flags);
284
285 /**
286 * List of objects currently involved in rendering from the
287 * ringbuffer.
288 *
289 * Includes buffers having the contents of their GPU caches
290 * flushed, not necessarily primitives. last_read_req
291 * represents when the rendering involved will be completed.
292 *
293 * A reference is held on the buffer while on this list.
294 */
295 struct list_head active_list;
296
297 /**
298 * List of breadcrumbs associated with GPU requests currently
299 * outstanding.
300 */
301 struct list_head request_list;
302
303 /**
304 * Seqno of request most recently submitted to request_list.
305 * Used exclusively by hang checker to avoid grabbing lock while
306 * inspecting request list.
307 */
308 u32 last_submitted_seqno;
309
310 bool gpu_caches_dirty;
311
312 wait_queue_head_t irq_queue;
313
314 struct intel_context *last_context;
315
316 struct intel_ring_hangcheck hangcheck;
317
318 struct {
319 struct drm_i915_gem_object *obj;
320 u32 gtt_offset;
321 volatile u32 *cpu_page;
322 } scratch;
323
324 bool needs_cmd_parser;
325
326 /*
327 * Table of commands the command parser needs to know about
328 * for this ring.
329 */
330 DECLARE_HASHTABLE(cmd_hash, I915_CMD_HASH_ORDER);
331
332 /*
333 * Table of registers allowed in commands that read/write registers.
334 */
335 const struct drm_i915_reg_descriptor *reg_table;
336 int reg_count;
337
338 /*
339 * Table of registers allowed in commands that read/write registers, but
340 * only from the DRM master.
341 */
342 const struct drm_i915_reg_descriptor *master_reg_table;
343 int master_reg_count;
344
345 /*
346 * Returns the bitmask for the length field of the specified command.
347 * Return 0 for an unrecognized/invalid command.
348 *
349 * If the command parser finds an entry for a command in the ring's
350 * cmd_tables, it gets the command's length based on the table entry.
351 * If not, it calls this function to determine the per-ring length field
352 * encoding for the command (i.e. certain opcode ranges use certain bits
353 * to encode the command length in the header).
354 */
355 u32 (*get_cmd_length_mask)(u32 cmd_header);
356};
357
358static inline bool
359intel_ring_initialized(struct intel_engine_cs *ring)
360{
361 return ring->dev != NULL;
362}
363
364static inline unsigned
365intel_ring_flag(struct intel_engine_cs *ring)
366{
367 return 1 << ring->id;
368}
369
370static inline u32
371intel_ring_sync_index(struct intel_engine_cs *ring,
372 struct intel_engine_cs *other)
373{
374 int idx;
375
376 /*
377 * rcs -> 0 = vcs, 1 = bcs, 2 = vecs, 3 = vcs2;
378 * vcs -> 0 = bcs, 1 = vecs, 2 = vcs2, 3 = rcs;
379 * bcs -> 0 = vecs, 1 = vcs2. 2 = rcs, 3 = vcs;
380 * vecs -> 0 = vcs2, 1 = rcs, 2 = vcs, 3 = bcs;
381 * vcs2 -> 0 = rcs, 1 = vcs, 2 = bcs, 3 = vecs;
382 */
383
384 idx = (other - ring) - 1;
385 if (idx < 0)
386 idx += I915_NUM_RINGS;
387
388 return idx;
389}
390
391static inline void
392intel_flush_status_page(struct intel_engine_cs *ring, int reg)
393{
394 drm_clflush_virt_range(&ring->status_page.page_addr[reg],
395 sizeof(uint32_t));
396}
397
398static inline u32
399intel_read_status_page(struct intel_engine_cs *ring,
400 int reg)
401{
402 /* Ensure that the compiler doesn't optimize away the load. */
403 barrier();
404 return ring->status_page.page_addr[reg];
405}
406
407static inline void
408intel_write_status_page(struct intel_engine_cs *ring,
409 int reg, u32 value)
410{
411 ring->status_page.page_addr[reg] = value;
412}
413
414/*
415 * Reads a dword out of the status page, which is written to from the command
416 * queue by automatic updates, MI_REPORT_HEAD, MI_STORE_DATA_INDEX, or
417 * MI_STORE_DATA_IMM.
418 *
419 * The following dwords have a reserved meaning:
420 * 0x00: ISR copy, updated when an ISR bit not set in the HWSTAM changes.
421 * 0x04: ring 0 head pointer
422 * 0x05: ring 1 head pointer (915-class)
423 * 0x06: ring 2 head pointer (915-class)
424 * 0x10-0x1b: Context status DWords (GM45)
425 * 0x1f: Last written status offset. (GM45)
426 * 0x20-0x2f: Reserved (Gen6+)
427 *
428 * The area from dword 0x30 to 0x3ff is available for driver usage.
429 */
430#define I915_GEM_HWS_INDEX 0x30
431#define I915_GEM_HWS_INDEX_ADDR (I915_GEM_HWS_INDEX << MI_STORE_DWORD_INDEX_SHIFT)
432#define I915_GEM_HWS_SCRATCH_INDEX 0x40
433#define I915_GEM_HWS_SCRATCH_ADDR (I915_GEM_HWS_SCRATCH_INDEX << MI_STORE_DWORD_INDEX_SHIFT)
434
435struct intel_ringbuffer *
436intel_engine_create_ringbuffer(struct intel_engine_cs *engine, int size);
437int intel_pin_and_map_ringbuffer_obj(struct drm_device *dev,
438 struct intel_ringbuffer *ringbuf);
439void intel_unpin_ringbuffer_obj(struct intel_ringbuffer *ringbuf);
440void intel_ringbuffer_free(struct intel_ringbuffer *ring);
441
442void intel_stop_ring_buffer(struct intel_engine_cs *ring);
443void intel_cleanup_ring_buffer(struct intel_engine_cs *ring);
444
445int intel_ring_alloc_request_extras(struct drm_i915_gem_request *request);
446
447int __must_check intel_ring_begin(struct drm_i915_gem_request *req, int n);
448int __must_check intel_ring_cacheline_align(struct drm_i915_gem_request *req);
449static inline void intel_ring_emit(struct intel_engine_cs *ring,
450 u32 data)
451{
452 struct intel_ringbuffer *ringbuf = ring->buffer;
453 iowrite32(data, ringbuf->virtual_start + ringbuf->tail);
454 ringbuf->tail += 4;
455}
456static inline void intel_ring_emit_reg(struct intel_engine_cs *ring,
457 i915_reg_t reg)
458{
459 intel_ring_emit(ring, i915_mmio_reg_offset(reg));
460}
461static inline void intel_ring_advance(struct intel_engine_cs *ring)
462{
463 struct intel_ringbuffer *ringbuf = ring->buffer;
464 ringbuf->tail &= ringbuf->size - 1;
465}
466int __intel_ring_space(int head, int tail, int size);
467void intel_ring_update_space(struct intel_ringbuffer *ringbuf);
468int intel_ring_space(struct intel_ringbuffer *ringbuf);
469bool intel_ring_stopped(struct intel_engine_cs *ring);
470
471int __must_check intel_ring_idle(struct intel_engine_cs *ring);
472void intel_ring_init_seqno(struct intel_engine_cs *ring, u32 seqno);
473int intel_ring_flush_all_caches(struct drm_i915_gem_request *req);
474int intel_ring_invalidate_all_caches(struct drm_i915_gem_request *req);
475
476void intel_fini_pipe_control(struct intel_engine_cs *ring);
477int intel_init_pipe_control(struct intel_engine_cs *ring);
478
479int intel_init_render_ring_buffer(struct drm_device *dev);
480int intel_init_bsd_ring_buffer(struct drm_device *dev);
481int intel_init_bsd2_ring_buffer(struct drm_device *dev);
482int intel_init_blt_ring_buffer(struct drm_device *dev);
483int intel_init_vebox_ring_buffer(struct drm_device *dev);
484
485u64 intel_ring_get_active_head(struct intel_engine_cs *ring);
486
487int init_workarounds_ring(struct intel_engine_cs *ring);
488
489static inline u32 intel_ring_get_tail(struct intel_ringbuffer *ringbuf)
490{
491 return ringbuf->tail;
492}
493
494/*
495 * Arbitrary size for largest possible 'add request' sequence. The code paths
496 * are complex and variable. Empirical measurement shows that the worst case
497 * is ILK at 136 words. Reserving too much is better than reserving too little
498 * as that allows for corner cases that might have been missed. So the figure
499 * has been rounded up to 160 words.
500 */
501#define MIN_SPACE_FOR_ADD_REQUEST 160
502
503/*
504 * Reserve space in the ring to guarantee that the i915_add_request() call
505 * will always have sufficient room to do its stuff. The request creation
506 * code calls this automatically.
507 */
508void intel_ring_reserved_space_reserve(struct intel_ringbuffer *ringbuf, int size);
509/* Cancel the reservation, e.g. because the request is being discarded. */
510void intel_ring_reserved_space_cancel(struct intel_ringbuffer *ringbuf);
511/* Use the reserved space - for use by i915_add_request() only. */
512void intel_ring_reserved_space_use(struct intel_ringbuffer *ringbuf);
513/* Finish with the reserved space - for use by i915_add_request() only. */
514void intel_ring_reserved_space_end(struct intel_ringbuffer *ringbuf);
515
516/* Legacy ringbuffer specific portion of reservation code: */
517int intel_ring_reserve_space(struct drm_i915_gem_request *request);
518
519#endif /* _INTEL_RINGBUFFER_H_ */
1#ifndef _INTEL_RINGBUFFER_H_
2#define _INTEL_RINGBUFFER_H_
3
4#include <linux/hashtable.h>
5#include "i915_gem_batch_pool.h"
6#include "i915_gem_request.h"
7#include "i915_gem_timeline.h"
8
9#define I915_CMD_HASH_ORDER 9
10
11/* Early gen2 devices have a cacheline of just 32 bytes, using 64 is overkill,
12 * but keeps the logic simple. Indeed, the whole purpose of this macro is just
13 * to give some inclination as to some of the magic values used in the various
14 * workarounds!
15 */
16#define CACHELINE_BYTES 64
17#define CACHELINE_DWORDS (CACHELINE_BYTES / sizeof(uint32_t))
18
19/*
20 * Gen2 BSpec "1. Programming Environment" / 1.4.4.6 "Ring Buffer Use"
21 * Gen3 BSpec "vol1c Memory Interface Functions" / 2.3.4.5 "Ring Buffer Use"
22 * Gen4+ BSpec "vol1c Memory Interface and Command Stream" / 5.3.4.5 "Ring Buffer Use"
23 *
24 * "If the Ring Buffer Head Pointer and the Tail Pointer are on the same
25 * cacheline, the Head Pointer must not be greater than the Tail
26 * Pointer."
27 */
28#define I915_RING_FREE_SPACE 64
29
30struct intel_hw_status_page {
31 struct i915_vma *vma;
32 u32 *page_addr;
33 u32 ggtt_offset;
34};
35
36#define I915_READ_TAIL(engine) I915_READ(RING_TAIL((engine)->mmio_base))
37#define I915_WRITE_TAIL(engine, val) I915_WRITE(RING_TAIL((engine)->mmio_base), val)
38
39#define I915_READ_START(engine) I915_READ(RING_START((engine)->mmio_base))
40#define I915_WRITE_START(engine, val) I915_WRITE(RING_START((engine)->mmio_base), val)
41
42#define I915_READ_HEAD(engine) I915_READ(RING_HEAD((engine)->mmio_base))
43#define I915_WRITE_HEAD(engine, val) I915_WRITE(RING_HEAD((engine)->mmio_base), val)
44
45#define I915_READ_CTL(engine) I915_READ(RING_CTL((engine)->mmio_base))
46#define I915_WRITE_CTL(engine, val) I915_WRITE(RING_CTL((engine)->mmio_base), val)
47
48#define I915_READ_IMR(engine) I915_READ(RING_IMR((engine)->mmio_base))
49#define I915_WRITE_IMR(engine, val) I915_WRITE(RING_IMR((engine)->mmio_base), val)
50
51#define I915_READ_MODE(engine) I915_READ(RING_MI_MODE((engine)->mmio_base))
52#define I915_WRITE_MODE(engine, val) I915_WRITE(RING_MI_MODE((engine)->mmio_base), val)
53
54/* seqno size is actually only a uint32, but since we plan to use MI_FLUSH_DW to
55 * do the writes, and that must have qw aligned offsets, simply pretend it's 8b.
56 */
57#define gen8_semaphore_seqno_size sizeof(uint64_t)
58#define GEN8_SEMAPHORE_OFFSET(__from, __to) \
59 (((__from) * I915_NUM_ENGINES + (__to)) * gen8_semaphore_seqno_size)
60#define GEN8_SIGNAL_OFFSET(__ring, to) \
61 (dev_priv->semaphore->node.start + \
62 GEN8_SEMAPHORE_OFFSET((__ring)->id, (to)))
63#define GEN8_WAIT_OFFSET(__ring, from) \
64 (dev_priv->semaphore->node.start + \
65 GEN8_SEMAPHORE_OFFSET(from, (__ring)->id))
66
67enum intel_engine_hangcheck_action {
68 HANGCHECK_IDLE = 0,
69 HANGCHECK_WAIT,
70 HANGCHECK_ACTIVE,
71 HANGCHECK_KICK,
72 HANGCHECK_HUNG,
73};
74
75#define HANGCHECK_SCORE_RING_HUNG 31
76
77#define I915_MAX_SLICES 3
78#define I915_MAX_SUBSLICES 3
79
80#define instdone_slice_mask(dev_priv__) \
81 (INTEL_GEN(dev_priv__) == 7 ? \
82 1 : INTEL_INFO(dev_priv__)->sseu.slice_mask)
83
84#define instdone_subslice_mask(dev_priv__) \
85 (INTEL_GEN(dev_priv__) == 7 ? \
86 1 : INTEL_INFO(dev_priv__)->sseu.subslice_mask)
87
88#define for_each_instdone_slice_subslice(dev_priv__, slice__, subslice__) \
89 for ((slice__) = 0, (subslice__) = 0; \
90 (slice__) < I915_MAX_SLICES; \
91 (subslice__) = ((subslice__) + 1) < I915_MAX_SUBSLICES ? (subslice__) + 1 : 0, \
92 (slice__) += ((subslice__) == 0)) \
93 for_each_if((BIT(slice__) & instdone_slice_mask(dev_priv__)) && \
94 (BIT(subslice__) & instdone_subslice_mask(dev_priv__)))
95
96struct intel_instdone {
97 u32 instdone;
98 /* The following exist only in the RCS engine */
99 u32 slice_common;
100 u32 sampler[I915_MAX_SLICES][I915_MAX_SUBSLICES];
101 u32 row[I915_MAX_SLICES][I915_MAX_SUBSLICES];
102};
103
104struct intel_engine_hangcheck {
105 u64 acthd;
106 u32 seqno;
107 int score;
108 enum intel_engine_hangcheck_action action;
109 int deadlock;
110 struct intel_instdone instdone;
111};
112
113struct intel_ring {
114 struct i915_vma *vma;
115 void *vaddr;
116
117 struct intel_engine_cs *engine;
118
119 struct list_head request_list;
120
121 u32 head;
122 u32 tail;
123 int space;
124 int size;
125 int effective_size;
126
127 /** We track the position of the requests in the ring buffer, and
128 * when each is retired we increment last_retired_head as the GPU
129 * must have finished processing the request and so we know we
130 * can advance the ringbuffer up to that position.
131 *
132 * last_retired_head is set to -1 after the value is consumed so
133 * we can detect new retirements.
134 */
135 u32 last_retired_head;
136};
137
138struct i915_gem_context;
139struct drm_i915_reg_table;
140
141/*
142 * we use a single page to load ctx workarounds so all of these
143 * values are referred in terms of dwords
144 *
145 * struct i915_wa_ctx_bb:
146 * offset: specifies batch starting position, also helpful in case
147 * if we want to have multiple batches at different offsets based on
148 * some criteria. It is not a requirement at the moment but provides
149 * an option for future use.
150 * size: size of the batch in DWORDS
151 */
152struct i915_ctx_workarounds {
153 struct i915_wa_ctx_bb {
154 u32 offset;
155 u32 size;
156 } indirect_ctx, per_ctx;
157 struct i915_vma *vma;
158};
159
160struct drm_i915_gem_request;
161struct intel_render_state;
162
163struct intel_engine_cs {
164 struct drm_i915_private *i915;
165 const char *name;
166 enum intel_engine_id {
167 RCS = 0,
168 BCS,
169 VCS,
170 VCS2, /* Keep instances of the same type engine together. */
171 VECS
172 } id;
173#define _VCS(n) (VCS + (n))
174 unsigned int exec_id;
175 enum intel_engine_hw_id {
176 RCS_HW = 0,
177 VCS_HW,
178 BCS_HW,
179 VECS_HW,
180 VCS2_HW
181 } hw_id;
182 enum intel_engine_hw_id guc_id; /* XXX same as hw_id? */
183 u32 mmio_base;
184 unsigned int irq_shift;
185 struct intel_ring *buffer;
186 struct intel_timeline *timeline;
187
188 struct intel_render_state *render_state;
189
190 /* Rather than have every client wait upon all user interrupts,
191 * with the herd waking after every interrupt and each doing the
192 * heavyweight seqno dance, we delegate the task (of being the
193 * bottom-half of the user interrupt) to the first client. After
194 * every interrupt, we wake up one client, who does the heavyweight
195 * coherent seqno read and either goes back to sleep (if incomplete),
196 * or wakes up all the completed clients in parallel, before then
197 * transferring the bottom-half status to the next client in the queue.
198 *
199 * Compared to walking the entire list of waiters in a single dedicated
200 * bottom-half, we reduce the latency of the first waiter by avoiding
201 * a context switch, but incur additional coherent seqno reads when
202 * following the chain of request breadcrumbs. Since it is most likely
203 * that we have a single client waiting on each seqno, then reducing
204 * the overhead of waking that client is much preferred.
205 */
206 struct intel_breadcrumbs {
207 struct task_struct __rcu *irq_seqno_bh; /* bh for interrupts */
208 bool irq_posted;
209
210 spinlock_t lock; /* protects the lists of requests; irqsafe */
211 struct rb_root waiters; /* sorted by retirement, priority */
212 struct rb_root signals; /* sorted by retirement */
213 struct intel_wait *first_wait; /* oldest waiter by retirement */
214 struct task_struct *signaler; /* used for fence signalling */
215 struct drm_i915_gem_request *first_signal;
216 struct timer_list fake_irq; /* used after a missed interrupt */
217 struct timer_list hangcheck; /* detect missed interrupts */
218
219 unsigned long timeout;
220
221 bool irq_enabled : 1;
222 bool rpm_wakelock : 1;
223 } breadcrumbs;
224
225 /*
226 * A pool of objects to use as shadow copies of client batch buffers
227 * when the command parser is enabled. Prevents the client from
228 * modifying the batch contents after software parsing.
229 */
230 struct i915_gem_batch_pool batch_pool;
231
232 struct intel_hw_status_page status_page;
233 struct i915_ctx_workarounds wa_ctx;
234 struct i915_vma *scratch;
235
236 u32 irq_keep_mask; /* always keep these interrupts */
237 u32 irq_enable_mask; /* bitmask to enable ring interrupt */
238 void (*irq_enable)(struct intel_engine_cs *engine);
239 void (*irq_disable)(struct intel_engine_cs *engine);
240
241 int (*init_hw)(struct intel_engine_cs *engine);
242 void (*reset_hw)(struct intel_engine_cs *engine,
243 struct drm_i915_gem_request *req);
244
245 int (*init_context)(struct drm_i915_gem_request *req);
246
247 int (*emit_flush)(struct drm_i915_gem_request *request,
248 u32 mode);
249#define EMIT_INVALIDATE BIT(0)
250#define EMIT_FLUSH BIT(1)
251#define EMIT_BARRIER (EMIT_INVALIDATE | EMIT_FLUSH)
252 int (*emit_bb_start)(struct drm_i915_gem_request *req,
253 u64 offset, u32 length,
254 unsigned int dispatch_flags);
255#define I915_DISPATCH_SECURE BIT(0)
256#define I915_DISPATCH_PINNED BIT(1)
257#define I915_DISPATCH_RS BIT(2)
258 void (*emit_breadcrumb)(struct drm_i915_gem_request *req,
259 u32 *out);
260 int emit_breadcrumb_sz;
261
262 /* Pass the request to the hardware queue (e.g. directly into
263 * the legacy ringbuffer or to the end of an execlist).
264 *
265 * This is called from an atomic context with irqs disabled; must
266 * be irq safe.
267 */
268 void (*submit_request)(struct drm_i915_gem_request *req);
269
270 /* Call when the priority on a request has changed and it and its
271 * dependencies may need rescheduling. Note the request itself may
272 * not be ready to run!
273 *
274 * Called under the struct_mutex.
275 */
276 void (*schedule)(struct drm_i915_gem_request *request,
277 int priority);
278
279 /* Some chipsets are not quite as coherent as advertised and need
280 * an expensive kick to force a true read of the up-to-date seqno.
281 * However, the up-to-date seqno is not always required and the last
282 * seen value is good enough. Note that the seqno will always be
283 * monotonic, even if not coherent.
284 */
285 void (*irq_seqno_barrier)(struct intel_engine_cs *engine);
286 void (*cleanup)(struct intel_engine_cs *engine);
287
288 /* GEN8 signal/wait table - never trust comments!
289 * signal to signal to signal to signal to signal to
290 * RCS VCS BCS VECS VCS2
291 * --------------------------------------------------------------------
292 * RCS | NOP (0x00) | VCS (0x08) | BCS (0x10) | VECS (0x18) | VCS2 (0x20) |
293 * |-------------------------------------------------------------------
294 * VCS | RCS (0x28) | NOP (0x30) | BCS (0x38) | VECS (0x40) | VCS2 (0x48) |
295 * |-------------------------------------------------------------------
296 * BCS | RCS (0x50) | VCS (0x58) | NOP (0x60) | VECS (0x68) | VCS2 (0x70) |
297 * |-------------------------------------------------------------------
298 * VECS | RCS (0x78) | VCS (0x80) | BCS (0x88) | NOP (0x90) | VCS2 (0x98) |
299 * |-------------------------------------------------------------------
300 * VCS2 | RCS (0xa0) | VCS (0xa8) | BCS (0xb0) | VECS (0xb8) | NOP (0xc0) |
301 * |-------------------------------------------------------------------
302 *
303 * Generalization:
304 * f(x, y) := (x->id * NUM_RINGS * seqno_size) + (seqno_size * y->id)
305 * ie. transpose of g(x, y)
306 *
307 * sync from sync from sync from sync from sync from
308 * RCS VCS BCS VECS VCS2
309 * --------------------------------------------------------------------
310 * RCS | NOP (0x00) | VCS (0x28) | BCS (0x50) | VECS (0x78) | VCS2 (0xa0) |
311 * |-------------------------------------------------------------------
312 * VCS | RCS (0x08) | NOP (0x30) | BCS (0x58) | VECS (0x80) | VCS2 (0xa8) |
313 * |-------------------------------------------------------------------
314 * BCS | RCS (0x10) | VCS (0x38) | NOP (0x60) | VECS (0x88) | VCS2 (0xb0) |
315 * |-------------------------------------------------------------------
316 * VECS | RCS (0x18) | VCS (0x40) | BCS (0x68) | NOP (0x90) | VCS2 (0xb8) |
317 * |-------------------------------------------------------------------
318 * VCS2 | RCS (0x20) | VCS (0x48) | BCS (0x70) | VECS (0x98) | NOP (0xc0) |
319 * |-------------------------------------------------------------------
320 *
321 * Generalization:
322 * g(x, y) := (y->id * NUM_RINGS * seqno_size) + (seqno_size * x->id)
323 * ie. transpose of f(x, y)
324 */
325 struct {
326 union {
327#define GEN6_SEMAPHORE_LAST VECS_HW
328#define GEN6_NUM_SEMAPHORES (GEN6_SEMAPHORE_LAST + 1)
329#define GEN6_SEMAPHORES_MASK GENMASK(GEN6_SEMAPHORE_LAST, 0)
330 struct {
331 /* our mbox written by others */
332 u32 wait[GEN6_NUM_SEMAPHORES];
333 /* mboxes this ring signals to */
334 i915_reg_t signal[GEN6_NUM_SEMAPHORES];
335 } mbox;
336 u64 signal_ggtt[I915_NUM_ENGINES];
337 };
338
339 /* AKA wait() */
340 int (*sync_to)(struct drm_i915_gem_request *req,
341 struct drm_i915_gem_request *signal);
342 u32 *(*signal)(struct drm_i915_gem_request *req, u32 *out);
343 } semaphore;
344
345 /* Execlists */
346 struct tasklet_struct irq_tasklet;
347 struct execlist_port {
348 struct drm_i915_gem_request *request;
349 unsigned int count;
350 } execlist_port[2];
351 struct rb_root execlist_queue;
352 struct rb_node *execlist_first;
353 unsigned int fw_domains;
354 bool disable_lite_restore_wa;
355 bool preempt_wa;
356 u32 ctx_desc_template;
357
358 struct i915_gem_context *last_context;
359
360 struct intel_engine_hangcheck hangcheck;
361
362 bool needs_cmd_parser;
363
364 /*
365 * Table of commands the command parser needs to know about
366 * for this engine.
367 */
368 DECLARE_HASHTABLE(cmd_hash, I915_CMD_HASH_ORDER);
369
370 /*
371 * Table of registers allowed in commands that read/write registers.
372 */
373 const struct drm_i915_reg_table *reg_tables;
374 int reg_table_count;
375
376 /*
377 * Returns the bitmask for the length field of the specified command.
378 * Return 0 for an unrecognized/invalid command.
379 *
380 * If the command parser finds an entry for a command in the engine's
381 * cmd_tables, it gets the command's length based on the table entry.
382 * If not, it calls this function to determine the per-engine length
383 * field encoding for the command (i.e. different opcode ranges use
384 * certain bits to encode the command length in the header).
385 */
386 u32 (*get_cmd_length_mask)(u32 cmd_header);
387};
388
389static inline unsigned
390intel_engine_flag(const struct intel_engine_cs *engine)
391{
392 return 1 << engine->id;
393}
394
395static inline void
396intel_flush_status_page(struct intel_engine_cs *engine, int reg)
397{
398 mb();
399 clflush(&engine->status_page.page_addr[reg]);
400 mb();
401}
402
403static inline u32
404intel_read_status_page(struct intel_engine_cs *engine, int reg)
405{
406 /* Ensure that the compiler doesn't optimize away the load. */
407 return READ_ONCE(engine->status_page.page_addr[reg]);
408}
409
410static inline void
411intel_write_status_page(struct intel_engine_cs *engine,
412 int reg, u32 value)
413{
414 engine->status_page.page_addr[reg] = value;
415}
416
417/*
418 * Reads a dword out of the status page, which is written to from the command
419 * queue by automatic updates, MI_REPORT_HEAD, MI_STORE_DATA_INDEX, or
420 * MI_STORE_DATA_IMM.
421 *
422 * The following dwords have a reserved meaning:
423 * 0x00: ISR copy, updated when an ISR bit not set in the HWSTAM changes.
424 * 0x04: ring 0 head pointer
425 * 0x05: ring 1 head pointer (915-class)
426 * 0x06: ring 2 head pointer (915-class)
427 * 0x10-0x1b: Context status DWords (GM45)
428 * 0x1f: Last written status offset. (GM45)
429 * 0x20-0x2f: Reserved (Gen6+)
430 *
431 * The area from dword 0x30 to 0x3ff is available for driver usage.
432 */
433#define I915_GEM_HWS_INDEX 0x30
434#define I915_GEM_HWS_INDEX_ADDR (I915_GEM_HWS_INDEX << MI_STORE_DWORD_INDEX_SHIFT)
435#define I915_GEM_HWS_SCRATCH_INDEX 0x40
436#define I915_GEM_HWS_SCRATCH_ADDR (I915_GEM_HWS_SCRATCH_INDEX << MI_STORE_DWORD_INDEX_SHIFT)
437
438struct intel_ring *
439intel_engine_create_ring(struct intel_engine_cs *engine, int size);
440int intel_ring_pin(struct intel_ring *ring);
441void intel_ring_unpin(struct intel_ring *ring);
442void intel_ring_free(struct intel_ring *ring);
443
444void intel_engine_stop(struct intel_engine_cs *engine);
445void intel_engine_cleanup(struct intel_engine_cs *engine);
446
447void intel_legacy_submission_resume(struct drm_i915_private *dev_priv);
448
449int intel_ring_alloc_request_extras(struct drm_i915_gem_request *request);
450
451int __must_check intel_ring_begin(struct drm_i915_gem_request *req, int n);
452int __must_check intel_ring_cacheline_align(struct drm_i915_gem_request *req);
453
454static inline void intel_ring_emit(struct intel_ring *ring, u32 data)
455{
456 *(uint32_t *)(ring->vaddr + ring->tail) = data;
457 ring->tail += 4;
458}
459
460static inline void intel_ring_emit_reg(struct intel_ring *ring, i915_reg_t reg)
461{
462 intel_ring_emit(ring, i915_mmio_reg_offset(reg));
463}
464
465static inline void intel_ring_advance(struct intel_ring *ring)
466{
467 /* Dummy function.
468 *
469 * This serves as a placeholder in the code so that the reader
470 * can compare against the preceding intel_ring_begin() and
471 * check that the number of dwords emitted matches the space
472 * reserved for the command packet (i.e. the value passed to
473 * intel_ring_begin()).
474 */
475}
476
477static inline u32 intel_ring_offset(struct intel_ring *ring, void *addr)
478{
479 /* Don't write ring->size (equivalent to 0) as that hangs some GPUs. */
480 u32 offset = addr - ring->vaddr;
481 return offset & (ring->size - 1);
482}
483
484int __intel_ring_space(int head, int tail, int size);
485void intel_ring_update_space(struct intel_ring *ring);
486
487void intel_engine_init_global_seqno(struct intel_engine_cs *engine, u32 seqno);
488
489void intel_engine_setup_common(struct intel_engine_cs *engine);
490int intel_engine_init_common(struct intel_engine_cs *engine);
491int intel_engine_create_scratch(struct intel_engine_cs *engine, int size);
492void intel_engine_cleanup_common(struct intel_engine_cs *engine);
493
494int intel_init_render_ring_buffer(struct intel_engine_cs *engine);
495int intel_init_bsd_ring_buffer(struct intel_engine_cs *engine);
496int intel_init_bsd2_ring_buffer(struct intel_engine_cs *engine);
497int intel_init_blt_ring_buffer(struct intel_engine_cs *engine);
498int intel_init_vebox_ring_buffer(struct intel_engine_cs *engine);
499
500u64 intel_engine_get_active_head(struct intel_engine_cs *engine);
501u64 intel_engine_get_last_batch_head(struct intel_engine_cs *engine);
502
503static inline u32 intel_engine_get_seqno(struct intel_engine_cs *engine)
504{
505 return intel_read_status_page(engine, I915_GEM_HWS_INDEX);
506}
507
508static inline u32 intel_engine_last_submit(struct intel_engine_cs *engine)
509{
510 /* We are only peeking at the tail of the submit queue (and not the
511 * queue itself) in order to gain a hint as to the current active
512 * state of the engine. Callers are not expected to be taking
513 * engine->timeline->lock, nor are they expected to be concerned
514 * wtih serialising this hint with anything, so document it as
515 * a hint and nothing more.
516 */
517 return READ_ONCE(engine->timeline->last_submitted_seqno);
518}
519
520int init_workarounds_ring(struct intel_engine_cs *engine);
521
522void intel_engine_get_instdone(struct intel_engine_cs *engine,
523 struct intel_instdone *instdone);
524
525/*
526 * Arbitrary size for largest possible 'add request' sequence. The code paths
527 * are complex and variable. Empirical measurement shows that the worst case
528 * is BDW at 192 bytes (6 + 6 + 36 dwords), then ILK at 136 bytes. However,
529 * we need to allocate double the largest single packet within that emission
530 * to account for tail wraparound (so 6 + 6 + 72 dwords for BDW).
531 */
532#define MIN_SPACE_FOR_ADD_REQUEST 336
533
534static inline u32 intel_hws_seqno_address(struct intel_engine_cs *engine)
535{
536 return engine->status_page.ggtt_offset + I915_GEM_HWS_INDEX_ADDR;
537}
538
539/* intel_breadcrumbs.c -- user interrupt bottom-half for waiters */
540int intel_engine_init_breadcrumbs(struct intel_engine_cs *engine);
541
542static inline void intel_wait_init(struct intel_wait *wait, u32 seqno)
543{
544 wait->tsk = current;
545 wait->seqno = seqno;
546}
547
548static inline bool intel_wait_complete(const struct intel_wait *wait)
549{
550 return RB_EMPTY_NODE(&wait->node);
551}
552
553bool intel_engine_add_wait(struct intel_engine_cs *engine,
554 struct intel_wait *wait);
555void intel_engine_remove_wait(struct intel_engine_cs *engine,
556 struct intel_wait *wait);
557void intel_engine_enable_signaling(struct drm_i915_gem_request *request);
558
559static inline bool intel_engine_has_waiter(const struct intel_engine_cs *engine)
560{
561 return rcu_access_pointer(engine->breadcrumbs.irq_seqno_bh);
562}
563
564static inline bool intel_engine_wakeup(const struct intel_engine_cs *engine)
565{
566 bool wakeup = false;
567
568 /* Note that for this not to dangerously chase a dangling pointer,
569 * we must hold the rcu_read_lock here.
570 *
571 * Also note that tsk is likely to be in !TASK_RUNNING state so an
572 * early test for tsk->state != TASK_RUNNING before wake_up_process()
573 * is unlikely to be beneficial.
574 */
575 if (intel_engine_has_waiter(engine)) {
576 struct task_struct *tsk;
577
578 rcu_read_lock();
579 tsk = rcu_dereference(engine->breadcrumbs.irq_seqno_bh);
580 if (tsk)
581 wakeup = wake_up_process(tsk);
582 rcu_read_unlock();
583 }
584
585 return wakeup;
586}
587
588void intel_engine_reset_breadcrumbs(struct intel_engine_cs *engine);
589void intel_engine_fini_breadcrumbs(struct intel_engine_cs *engine);
590unsigned int intel_breadcrumbs_busy(struct drm_i915_private *i915);
591
592#endif /* _INTEL_RINGBUFFER_H_ */