Loading...
1/*
2 * Copyright © 2014 Intel Corporation
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 *
23 * Authors:
24 * Ben Widawsky <ben@bwidawsk.net>
25 * Michel Thierry <michel.thierry@intel.com>
26 * Thomas Daniel <thomas.daniel@intel.com>
27 * Oscar Mateo <oscar.mateo@intel.com>
28 *
29 */
30
31/**
32 * DOC: Logical Rings, Logical Ring Contexts and Execlists
33 *
34 * Motivation:
35 * GEN8 brings an expansion of the HW contexts: "Logical Ring Contexts".
36 * These expanded contexts enable a number of new abilities, especially
37 * "Execlists" (also implemented in this file).
38 *
39 * One of the main differences with the legacy HW contexts is that logical
40 * ring contexts incorporate many more things to the context's state, like
41 * PDPs or ringbuffer control registers:
42 *
43 * The reason why PDPs are included in the context is straightforward: as
44 * PPGTTs (per-process GTTs) are actually per-context, having the PDPs
45 * contained there mean you don't need to do a ppgtt->switch_mm yourself,
46 * instead, the GPU will do it for you on the context switch.
47 *
48 * But, what about the ringbuffer control registers (head, tail, etc..)?
49 * shouldn't we just need a set of those per engine command streamer? This is
50 * where the name "Logical Rings" starts to make sense: by virtualizing the
51 * rings, the engine cs shifts to a new "ring buffer" with every context
52 * switch. When you want to submit a workload to the GPU you: A) choose your
53 * context, B) find its appropriate virtualized ring, C) write commands to it
54 * and then, finally, D) tell the GPU to switch to that context.
55 *
56 * Instead of the legacy MI_SET_CONTEXT, the way you tell the GPU to switch
57 * to a contexts is via a context execution list, ergo "Execlists".
58 *
59 * LRC implementation:
60 * Regarding the creation of contexts, we have:
61 *
62 * - One global default context.
63 * - One local default context for each opened fd.
64 * - One local extra context for each context create ioctl call.
65 *
66 * Now that ringbuffers belong per-context (and not per-engine, like before)
67 * and that contexts are uniquely tied to a given engine (and not reusable,
68 * like before) we need:
69 *
70 * - One ringbuffer per-engine inside each context.
71 * - One backing object per-engine inside each context.
72 *
73 * The global default context starts its life with these new objects fully
74 * allocated and populated. The local default context for each opened fd is
75 * more complex, because we don't know at creation time which engine is going
76 * to use them. To handle this, we have implemented a deferred creation of LR
77 * contexts:
78 *
79 * The local context starts its life as a hollow or blank holder, that only
80 * gets populated for a given engine once we receive an execbuffer. If later
81 * on we receive another execbuffer ioctl for the same context but a different
82 * engine, we allocate/populate a new ringbuffer and context backing object and
83 * so on.
84 *
85 * Finally, regarding local contexts created using the ioctl call: as they are
86 * only allowed with the render ring, we can allocate & populate them right
87 * away (no need to defer anything, at least for now).
88 *
89 * Execlists implementation:
90 * Execlists are the new method by which, on gen8+ hardware, workloads are
91 * submitted for execution (as opposed to the legacy, ringbuffer-based, method).
92 * This method works as follows:
93 *
94 * When a request is committed, its commands (the BB start and any leading or
95 * trailing commands, like the seqno breadcrumbs) are placed in the ringbuffer
96 * for the appropriate context. The tail pointer in the hardware context is not
97 * updated at this time, but instead, kept by the driver in the ringbuffer
98 * structure. A structure representing this request is added to a request queue
99 * for the appropriate engine: this structure contains a copy of the context's
100 * tail after the request was written to the ring buffer and a pointer to the
101 * context itself.
102 *
103 * If the engine's request queue was empty before the request was added, the
104 * queue is processed immediately. Otherwise the queue will be processed during
105 * a context switch interrupt. In any case, elements on the queue will get sent
106 * (in pairs) to the GPU's ExecLists Submit Port (ELSP, for short) with a
107 * globally unique 20-bits submission ID.
108 *
109 * When execution of a request completes, the GPU updates the context status
110 * buffer with a context complete event and generates a context switch interrupt.
111 * During the interrupt handling, the driver examines the events in the buffer:
112 * for each context complete event, if the announced ID matches that on the head
113 * of the request queue, then that request is retired and removed from the queue.
114 *
115 * After processing, if any requests were retired and the queue is not empty
116 * then a new execution list can be submitted. The two requests at the front of
117 * the queue are next to be submitted but since a context may not occur twice in
118 * an execution list, if subsequent requests have the same ID as the first then
119 * the two requests must be combined. This is done simply by discarding requests
120 * at the head of the queue until either only one requests is left (in which case
121 * we use a NULL second context) or the first two requests have unique IDs.
122 *
123 * By always executing the first two requests in the queue the driver ensures
124 * that the GPU is kept as busy as possible. In the case where a single context
125 * completes but a second context is still executing, the request for this second
126 * context will be at the head of the queue when we remove the first one. This
127 * request will then be resubmitted along with a new request for a different context,
128 * which will cause the hardware to continue executing the second request and queue
129 * the new request (the GPU detects the condition of a context getting preempted
130 * with the same context and optimizes the context switch flow by not doing
131 * preemption, but just sampling the new tail pointer).
132 *
133 */
134
135#include <drm/drmP.h>
136#include <drm/i915_drm.h>
137#include "i915_drv.h"
138#include "intel_mocs.h"
139
140#define GEN9_LR_CONTEXT_RENDER_SIZE (22 * PAGE_SIZE)
141#define GEN8_LR_CONTEXT_RENDER_SIZE (20 * PAGE_SIZE)
142#define GEN8_LR_CONTEXT_OTHER_SIZE (2 * PAGE_SIZE)
143
144#define RING_EXECLIST_QFULL (1 << 0x2)
145#define RING_EXECLIST1_VALID (1 << 0x3)
146#define RING_EXECLIST0_VALID (1 << 0x4)
147#define RING_EXECLIST_ACTIVE_STATUS (3 << 0xE)
148#define RING_EXECLIST1_ACTIVE (1 << 0x11)
149#define RING_EXECLIST0_ACTIVE (1 << 0x12)
150
151#define GEN8_CTX_STATUS_IDLE_ACTIVE (1 << 0)
152#define GEN8_CTX_STATUS_PREEMPTED (1 << 1)
153#define GEN8_CTX_STATUS_ELEMENT_SWITCH (1 << 2)
154#define GEN8_CTX_STATUS_ACTIVE_IDLE (1 << 3)
155#define GEN8_CTX_STATUS_COMPLETE (1 << 4)
156#define GEN8_CTX_STATUS_LITE_RESTORE (1 << 15)
157
158#define CTX_LRI_HEADER_0 0x01
159#define CTX_CONTEXT_CONTROL 0x02
160#define CTX_RING_HEAD 0x04
161#define CTX_RING_TAIL 0x06
162#define CTX_RING_BUFFER_START 0x08
163#define CTX_RING_BUFFER_CONTROL 0x0a
164#define CTX_BB_HEAD_U 0x0c
165#define CTX_BB_HEAD_L 0x0e
166#define CTX_BB_STATE 0x10
167#define CTX_SECOND_BB_HEAD_U 0x12
168#define CTX_SECOND_BB_HEAD_L 0x14
169#define CTX_SECOND_BB_STATE 0x16
170#define CTX_BB_PER_CTX_PTR 0x18
171#define CTX_RCS_INDIRECT_CTX 0x1a
172#define CTX_RCS_INDIRECT_CTX_OFFSET 0x1c
173#define CTX_LRI_HEADER_1 0x21
174#define CTX_CTX_TIMESTAMP 0x22
175#define CTX_PDP3_UDW 0x24
176#define CTX_PDP3_LDW 0x26
177#define CTX_PDP2_UDW 0x28
178#define CTX_PDP2_LDW 0x2a
179#define CTX_PDP1_UDW 0x2c
180#define CTX_PDP1_LDW 0x2e
181#define CTX_PDP0_UDW 0x30
182#define CTX_PDP0_LDW 0x32
183#define CTX_LRI_HEADER_2 0x41
184#define CTX_R_PWR_CLK_STATE 0x42
185#define CTX_GPGPU_CSR_BASE_ADDRESS 0x44
186
187#define GEN8_CTX_VALID (1<<0)
188#define GEN8_CTX_FORCE_PD_RESTORE (1<<1)
189#define GEN8_CTX_FORCE_RESTORE (1<<2)
190#define GEN8_CTX_L3LLC_COHERENT (1<<5)
191#define GEN8_CTX_PRIVILEGE (1<<8)
192
193#define ASSIGN_CTX_REG(reg_state, pos, reg, val) do { \
194 (reg_state)[(pos)+0] = i915_mmio_reg_offset(reg); \
195 (reg_state)[(pos)+1] = (val); \
196} while (0)
197
198#define ASSIGN_CTX_PDP(ppgtt, reg_state, n) do { \
199 const u64 _addr = i915_page_dir_dma_addr((ppgtt), (n)); \
200 reg_state[CTX_PDP ## n ## _UDW+1] = upper_32_bits(_addr); \
201 reg_state[CTX_PDP ## n ## _LDW+1] = lower_32_bits(_addr); \
202} while (0)
203
204#define ASSIGN_CTX_PML4(ppgtt, reg_state) do { \
205 reg_state[CTX_PDP0_UDW + 1] = upper_32_bits(px_dma(&ppgtt->pml4)); \
206 reg_state[CTX_PDP0_LDW + 1] = lower_32_bits(px_dma(&ppgtt->pml4)); \
207} while (0)
208
209enum {
210 ADVANCED_CONTEXT = 0,
211 LEGACY_32B_CONTEXT,
212 ADVANCED_AD_CONTEXT,
213 LEGACY_64B_CONTEXT
214};
215#define GEN8_CTX_ADDRESSING_MODE_SHIFT 3
216#define GEN8_CTX_ADDRESSING_MODE(dev) (USES_FULL_48BIT_PPGTT(dev) ?\
217 LEGACY_64B_CONTEXT :\
218 LEGACY_32B_CONTEXT)
219enum {
220 FAULT_AND_HANG = 0,
221 FAULT_AND_HALT, /* Debug only */
222 FAULT_AND_STREAM,
223 FAULT_AND_CONTINUE /* Unsupported */
224};
225#define GEN8_CTX_ID_SHIFT 32
226#define GEN8_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT 0x17
227#define GEN9_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT 0x26
228
229static int intel_lr_context_pin(struct intel_context *ctx,
230 struct intel_engine_cs *engine);
231static void lrc_setup_hardware_status_page(struct intel_engine_cs *ring,
232 struct drm_i915_gem_object *default_ctx_obj);
233
234
235/**
236 * intel_sanitize_enable_execlists() - sanitize i915.enable_execlists
237 * @dev: DRM device.
238 * @enable_execlists: value of i915.enable_execlists module parameter.
239 *
240 * Only certain platforms support Execlists (the prerequisites being
241 * support for Logical Ring Contexts and Aliasing PPGTT or better).
242 *
243 * Return: 1 if Execlists is supported and has to be enabled.
244 */
245int intel_sanitize_enable_execlists(struct drm_device *dev, int enable_execlists)
246{
247 WARN_ON(i915.enable_ppgtt == -1);
248
249 /* On platforms with execlist available, vGPU will only
250 * support execlist mode, no ring buffer mode.
251 */
252 if (HAS_LOGICAL_RING_CONTEXTS(dev) && intel_vgpu_active(dev))
253 return 1;
254
255 if (INTEL_INFO(dev)->gen >= 9)
256 return 1;
257
258 if (enable_execlists == 0)
259 return 0;
260
261 if (HAS_LOGICAL_RING_CONTEXTS(dev) && USES_PPGTT(dev) &&
262 i915.use_mmio_flip >= 0)
263 return 1;
264
265 return 0;
266}
267
268static void
269logical_ring_init_platform_invariants(struct intel_engine_cs *ring)
270{
271 struct drm_device *dev = ring->dev;
272
273 ring->disable_lite_restore_wa = (IS_SKL_REVID(dev, 0, SKL_REVID_B0) ||
274 IS_BXT_REVID(dev, 0, BXT_REVID_A1)) &&
275 (ring->id == VCS || ring->id == VCS2);
276
277 ring->ctx_desc_template = GEN8_CTX_VALID;
278 ring->ctx_desc_template |= GEN8_CTX_ADDRESSING_MODE(dev) <<
279 GEN8_CTX_ADDRESSING_MODE_SHIFT;
280 if (IS_GEN8(dev))
281 ring->ctx_desc_template |= GEN8_CTX_L3LLC_COHERENT;
282 ring->ctx_desc_template |= GEN8_CTX_PRIVILEGE;
283
284 /* TODO: WaDisableLiteRestore when we start using semaphore
285 * signalling between Command Streamers */
286 /* ring->ctx_desc_template |= GEN8_CTX_FORCE_RESTORE; */
287
288 /* WaEnableForceRestoreInCtxtDescForVCS:skl */
289 /* WaEnableForceRestoreInCtxtDescForVCS:bxt */
290 if (ring->disable_lite_restore_wa)
291 ring->ctx_desc_template |= GEN8_CTX_FORCE_RESTORE;
292}
293
294/**
295 * intel_lr_context_descriptor_update() - calculate & cache the descriptor
296 * descriptor for a pinned context
297 *
298 * @ctx: Context to work on
299 * @ring: Engine the descriptor will be used with
300 *
301 * The context descriptor encodes various attributes of a context,
302 * including its GTT address and some flags. Because it's fairly
303 * expensive to calculate, we'll just do it once and cache the result,
304 * which remains valid until the context is unpinned.
305 *
306 * This is what a descriptor looks like, from LSB to MSB:
307 * bits 0-11: flags, GEN8_CTX_* (cached in ctx_desc_template)
308 * bits 12-31: LRCA, GTT address of (the HWSP of) this context
309 * bits 32-51: ctx ID, a globally unique tag (the LRCA again!)
310 * bits 52-63: reserved, may encode the engine ID (for GuC)
311 */
312static void
313intel_lr_context_descriptor_update(struct intel_context *ctx,
314 struct intel_engine_cs *ring)
315{
316 uint64_t lrca, desc;
317
318 lrca = ctx->engine[ring->id].lrc_vma->node.start +
319 LRC_PPHWSP_PN * PAGE_SIZE;
320
321 desc = ring->ctx_desc_template; /* bits 0-11 */
322 desc |= lrca; /* bits 12-31 */
323 desc |= (lrca >> PAGE_SHIFT) << GEN8_CTX_ID_SHIFT; /* bits 32-51 */
324
325 ctx->engine[ring->id].lrc_desc = desc;
326}
327
328uint64_t intel_lr_context_descriptor(struct intel_context *ctx,
329 struct intel_engine_cs *ring)
330{
331 return ctx->engine[ring->id].lrc_desc;
332}
333
334/**
335 * intel_execlists_ctx_id() - get the Execlists Context ID
336 * @ctx: Context to get the ID for
337 * @ring: Engine to get the ID for
338 *
339 * Do not confuse with ctx->id! Unfortunately we have a name overload
340 * here: the old context ID we pass to userspace as a handler so that
341 * they can refer to a context, and the new context ID we pass to the
342 * ELSP so that the GPU can inform us of the context status via
343 * interrupts.
344 *
345 * The context ID is a portion of the context descriptor, so we can
346 * just extract the required part from the cached descriptor.
347 *
348 * Return: 20-bits globally unique context ID.
349 */
350u32 intel_execlists_ctx_id(struct intel_context *ctx,
351 struct intel_engine_cs *ring)
352{
353 return intel_lr_context_descriptor(ctx, ring) >> GEN8_CTX_ID_SHIFT;
354}
355
356static void execlists_elsp_write(struct drm_i915_gem_request *rq0,
357 struct drm_i915_gem_request *rq1)
358{
359
360 struct intel_engine_cs *ring = rq0->ring;
361 struct drm_device *dev = ring->dev;
362 struct drm_i915_private *dev_priv = dev->dev_private;
363 uint64_t desc[2];
364
365 if (rq1) {
366 desc[1] = intel_lr_context_descriptor(rq1->ctx, rq1->ring);
367 rq1->elsp_submitted++;
368 } else {
369 desc[1] = 0;
370 }
371
372 desc[0] = intel_lr_context_descriptor(rq0->ctx, rq0->ring);
373 rq0->elsp_submitted++;
374
375 /* You must always write both descriptors in the order below. */
376 spin_lock(&dev_priv->uncore.lock);
377 intel_uncore_forcewake_get__locked(dev_priv, FORCEWAKE_ALL);
378 I915_WRITE_FW(RING_ELSP(ring), upper_32_bits(desc[1]));
379 I915_WRITE_FW(RING_ELSP(ring), lower_32_bits(desc[1]));
380
381 I915_WRITE_FW(RING_ELSP(ring), upper_32_bits(desc[0]));
382 /* The context is automatically loaded after the following */
383 I915_WRITE_FW(RING_ELSP(ring), lower_32_bits(desc[0]));
384
385 /* ELSP is a wo register, use another nearby reg for posting */
386 POSTING_READ_FW(RING_EXECLIST_STATUS_LO(ring));
387 intel_uncore_forcewake_put__locked(dev_priv, FORCEWAKE_ALL);
388 spin_unlock(&dev_priv->uncore.lock);
389}
390
391static int execlists_update_context(struct drm_i915_gem_request *rq)
392{
393 struct intel_engine_cs *ring = rq->ring;
394 struct i915_hw_ppgtt *ppgtt = rq->ctx->ppgtt;
395 uint32_t *reg_state = rq->ctx->engine[ring->id].lrc_reg_state;
396
397 reg_state[CTX_RING_TAIL+1] = rq->tail;
398
399 if (ppgtt && !USES_FULL_48BIT_PPGTT(ppgtt->base.dev)) {
400 /* True 32b PPGTT with dynamic page allocation: update PDP
401 * registers and point the unallocated PDPs to scratch page.
402 * PML4 is allocated during ppgtt init, so this is not needed
403 * in 48-bit mode.
404 */
405 ASSIGN_CTX_PDP(ppgtt, reg_state, 3);
406 ASSIGN_CTX_PDP(ppgtt, reg_state, 2);
407 ASSIGN_CTX_PDP(ppgtt, reg_state, 1);
408 ASSIGN_CTX_PDP(ppgtt, reg_state, 0);
409 }
410
411 return 0;
412}
413
414static void execlists_submit_requests(struct drm_i915_gem_request *rq0,
415 struct drm_i915_gem_request *rq1)
416{
417 execlists_update_context(rq0);
418
419 if (rq1)
420 execlists_update_context(rq1);
421
422 execlists_elsp_write(rq0, rq1);
423}
424
425static void execlists_context_unqueue(struct intel_engine_cs *ring)
426{
427 struct drm_i915_gem_request *req0 = NULL, *req1 = NULL;
428 struct drm_i915_gem_request *cursor = NULL, *tmp = NULL;
429
430 assert_spin_locked(&ring->execlist_lock);
431
432 /*
433 * If irqs are not active generate a warning as batches that finish
434 * without the irqs may get lost and a GPU Hang may occur.
435 */
436 WARN_ON(!intel_irqs_enabled(ring->dev->dev_private));
437
438 if (list_empty(&ring->execlist_queue))
439 return;
440
441 /* Try to read in pairs */
442 list_for_each_entry_safe(cursor, tmp, &ring->execlist_queue,
443 execlist_link) {
444 if (!req0) {
445 req0 = cursor;
446 } else if (req0->ctx == cursor->ctx) {
447 /* Same ctx: ignore first request, as second request
448 * will update tail past first request's workload */
449 cursor->elsp_submitted = req0->elsp_submitted;
450 list_move_tail(&req0->execlist_link,
451 &ring->execlist_retired_req_list);
452 req0 = cursor;
453 } else {
454 req1 = cursor;
455 break;
456 }
457 }
458
459 if (IS_GEN8(ring->dev) || IS_GEN9(ring->dev)) {
460 /*
461 * WaIdleLiteRestore: make sure we never cause a lite
462 * restore with HEAD==TAIL
463 */
464 if (req0->elsp_submitted) {
465 /*
466 * Apply the wa NOOPS to prevent ring:HEAD == req:TAIL
467 * as we resubmit the request. See gen8_emit_request()
468 * for where we prepare the padding after the end of the
469 * request.
470 */
471 struct intel_ringbuffer *ringbuf;
472
473 ringbuf = req0->ctx->engine[ring->id].ringbuf;
474 req0->tail += 8;
475 req0->tail &= ringbuf->size - 1;
476 }
477 }
478
479 WARN_ON(req1 && req1->elsp_submitted);
480
481 execlists_submit_requests(req0, req1);
482}
483
484static bool execlists_check_remove_request(struct intel_engine_cs *ring,
485 u32 request_id)
486{
487 struct drm_i915_gem_request *head_req;
488
489 assert_spin_locked(&ring->execlist_lock);
490
491 head_req = list_first_entry_or_null(&ring->execlist_queue,
492 struct drm_i915_gem_request,
493 execlist_link);
494
495 if (head_req != NULL) {
496 if (intel_execlists_ctx_id(head_req->ctx, ring) == request_id) {
497 WARN(head_req->elsp_submitted == 0,
498 "Never submitted head request\n");
499
500 if (--head_req->elsp_submitted <= 0) {
501 list_move_tail(&head_req->execlist_link,
502 &ring->execlist_retired_req_list);
503 return true;
504 }
505 }
506 }
507
508 return false;
509}
510
511static void get_context_status(struct intel_engine_cs *ring,
512 u8 read_pointer,
513 u32 *status, u32 *context_id)
514{
515 struct drm_i915_private *dev_priv = ring->dev->dev_private;
516
517 if (WARN_ON(read_pointer >= GEN8_CSB_ENTRIES))
518 return;
519
520 *status = I915_READ(RING_CONTEXT_STATUS_BUF_LO(ring, read_pointer));
521 *context_id = I915_READ(RING_CONTEXT_STATUS_BUF_HI(ring, read_pointer));
522}
523
524/**
525 * intel_lrc_irq_handler() - handle Context Switch interrupts
526 * @ring: Engine Command Streamer to handle.
527 *
528 * Check the unread Context Status Buffers and manage the submission of new
529 * contexts to the ELSP accordingly.
530 */
531void intel_lrc_irq_handler(struct intel_engine_cs *ring)
532{
533 struct drm_i915_private *dev_priv = ring->dev->dev_private;
534 u32 status_pointer;
535 u8 read_pointer;
536 u8 write_pointer;
537 u32 status = 0;
538 u32 status_id;
539 u32 submit_contexts = 0;
540
541 status_pointer = I915_READ(RING_CONTEXT_STATUS_PTR(ring));
542
543 read_pointer = ring->next_context_status_buffer;
544 write_pointer = GEN8_CSB_WRITE_PTR(status_pointer);
545 if (read_pointer > write_pointer)
546 write_pointer += GEN8_CSB_ENTRIES;
547
548 spin_lock(&ring->execlist_lock);
549
550 while (read_pointer < write_pointer) {
551
552 get_context_status(ring, ++read_pointer % GEN8_CSB_ENTRIES,
553 &status, &status_id);
554
555 if (status & GEN8_CTX_STATUS_IDLE_ACTIVE)
556 continue;
557
558 if (status & GEN8_CTX_STATUS_PREEMPTED) {
559 if (status & GEN8_CTX_STATUS_LITE_RESTORE) {
560 if (execlists_check_remove_request(ring, status_id))
561 WARN(1, "Lite Restored request removed from queue\n");
562 } else
563 WARN(1, "Preemption without Lite Restore\n");
564 }
565
566 if ((status & GEN8_CTX_STATUS_ACTIVE_IDLE) ||
567 (status & GEN8_CTX_STATUS_ELEMENT_SWITCH)) {
568 if (execlists_check_remove_request(ring, status_id))
569 submit_contexts++;
570 }
571 }
572
573 if (ring->disable_lite_restore_wa) {
574 /* Prevent a ctx to preempt itself */
575 if ((status & GEN8_CTX_STATUS_ACTIVE_IDLE) &&
576 (submit_contexts != 0))
577 execlists_context_unqueue(ring);
578 } else if (submit_contexts != 0) {
579 execlists_context_unqueue(ring);
580 }
581
582 spin_unlock(&ring->execlist_lock);
583
584 if (unlikely(submit_contexts > 2))
585 DRM_ERROR("More than two context complete events?\n");
586
587 ring->next_context_status_buffer = write_pointer % GEN8_CSB_ENTRIES;
588
589 /* Update the read pointer to the old write pointer. Manual ringbuffer
590 * management ftw </sarcasm> */
591 I915_WRITE(RING_CONTEXT_STATUS_PTR(ring),
592 _MASKED_FIELD(GEN8_CSB_READ_PTR_MASK,
593 ring->next_context_status_buffer << 8));
594}
595
596static int execlists_context_queue(struct drm_i915_gem_request *request)
597{
598 struct intel_engine_cs *ring = request->ring;
599 struct drm_i915_gem_request *cursor;
600 int num_elements = 0;
601
602 if (request->ctx != request->i915->kernel_context)
603 intel_lr_context_pin(request->ctx, ring);
604
605 i915_gem_request_reference(request);
606
607 spin_lock_irq(&ring->execlist_lock);
608
609 list_for_each_entry(cursor, &ring->execlist_queue, execlist_link)
610 if (++num_elements > 2)
611 break;
612
613 if (num_elements > 2) {
614 struct drm_i915_gem_request *tail_req;
615
616 tail_req = list_last_entry(&ring->execlist_queue,
617 struct drm_i915_gem_request,
618 execlist_link);
619
620 if (request->ctx == tail_req->ctx) {
621 WARN(tail_req->elsp_submitted != 0,
622 "More than 2 already-submitted reqs queued\n");
623 list_move_tail(&tail_req->execlist_link,
624 &ring->execlist_retired_req_list);
625 }
626 }
627
628 list_add_tail(&request->execlist_link, &ring->execlist_queue);
629 if (num_elements == 0)
630 execlists_context_unqueue(ring);
631
632 spin_unlock_irq(&ring->execlist_lock);
633
634 return 0;
635}
636
637static int logical_ring_invalidate_all_caches(struct drm_i915_gem_request *req)
638{
639 struct intel_engine_cs *ring = req->ring;
640 uint32_t flush_domains;
641 int ret;
642
643 flush_domains = 0;
644 if (ring->gpu_caches_dirty)
645 flush_domains = I915_GEM_GPU_DOMAINS;
646
647 ret = ring->emit_flush(req, I915_GEM_GPU_DOMAINS, flush_domains);
648 if (ret)
649 return ret;
650
651 ring->gpu_caches_dirty = false;
652 return 0;
653}
654
655static int execlists_move_to_gpu(struct drm_i915_gem_request *req,
656 struct list_head *vmas)
657{
658 const unsigned other_rings = ~intel_ring_flag(req->ring);
659 struct i915_vma *vma;
660 uint32_t flush_domains = 0;
661 bool flush_chipset = false;
662 int ret;
663
664 list_for_each_entry(vma, vmas, exec_list) {
665 struct drm_i915_gem_object *obj = vma->obj;
666
667 if (obj->active & other_rings) {
668 ret = i915_gem_object_sync(obj, req->ring, &req);
669 if (ret)
670 return ret;
671 }
672
673 if (obj->base.write_domain & I915_GEM_DOMAIN_CPU)
674 flush_chipset |= i915_gem_clflush_object(obj, false);
675
676 flush_domains |= obj->base.write_domain;
677 }
678
679 if (flush_domains & I915_GEM_DOMAIN_GTT)
680 wmb();
681
682 /* Unconditionally invalidate gpu caches and ensure that we do flush
683 * any residual writes from the previous batch.
684 */
685 return logical_ring_invalidate_all_caches(req);
686}
687
688int intel_logical_ring_alloc_request_extras(struct drm_i915_gem_request *request)
689{
690 int ret = 0;
691
692 request->ringbuf = request->ctx->engine[request->ring->id].ringbuf;
693
694 if (i915.enable_guc_submission) {
695 /*
696 * Check that the GuC has space for the request before
697 * going any further, as the i915_add_request() call
698 * later on mustn't fail ...
699 */
700 struct intel_guc *guc = &request->i915->guc;
701
702 ret = i915_guc_wq_check_space(guc->execbuf_client);
703 if (ret)
704 return ret;
705 }
706
707 if (request->ctx != request->i915->kernel_context)
708 ret = intel_lr_context_pin(request->ctx, request->ring);
709
710 return ret;
711}
712
713static int logical_ring_wait_for_space(struct drm_i915_gem_request *req,
714 int bytes)
715{
716 struct intel_ringbuffer *ringbuf = req->ringbuf;
717 struct intel_engine_cs *ring = req->ring;
718 struct drm_i915_gem_request *target;
719 unsigned space;
720 int ret;
721
722 if (intel_ring_space(ringbuf) >= bytes)
723 return 0;
724
725 /* The whole point of reserving space is to not wait! */
726 WARN_ON(ringbuf->reserved_in_use);
727
728 list_for_each_entry(target, &ring->request_list, list) {
729 /*
730 * The request queue is per-engine, so can contain requests
731 * from multiple ringbuffers. Here, we must ignore any that
732 * aren't from the ringbuffer we're considering.
733 */
734 if (target->ringbuf != ringbuf)
735 continue;
736
737 /* Would completion of this request free enough space? */
738 space = __intel_ring_space(target->postfix, ringbuf->tail,
739 ringbuf->size);
740 if (space >= bytes)
741 break;
742 }
743
744 if (WARN_ON(&target->list == &ring->request_list))
745 return -ENOSPC;
746
747 ret = i915_wait_request(target);
748 if (ret)
749 return ret;
750
751 ringbuf->space = space;
752 return 0;
753}
754
755/*
756 * intel_logical_ring_advance_and_submit() - advance the tail and submit the workload
757 * @request: Request to advance the logical ringbuffer of.
758 *
759 * The tail is updated in our logical ringbuffer struct, not in the actual context. What
760 * really happens during submission is that the context and current tail will be placed
761 * on a queue waiting for the ELSP to be ready to accept a new context submission. At that
762 * point, the tail *inside* the context is updated and the ELSP written to.
763 */
764static int
765intel_logical_ring_advance_and_submit(struct drm_i915_gem_request *request)
766{
767 struct intel_ringbuffer *ringbuf = request->ringbuf;
768 struct drm_i915_private *dev_priv = request->i915;
769 struct intel_engine_cs *engine = request->ring;
770
771 intel_logical_ring_advance(ringbuf);
772 request->tail = ringbuf->tail;
773
774 /*
775 * Here we add two extra NOOPs as padding to avoid
776 * lite restore of a context with HEAD==TAIL.
777 *
778 * Caller must reserve WA_TAIL_DWORDS for us!
779 */
780 intel_logical_ring_emit(ringbuf, MI_NOOP);
781 intel_logical_ring_emit(ringbuf, MI_NOOP);
782 intel_logical_ring_advance(ringbuf);
783
784 if (intel_ring_stopped(engine))
785 return 0;
786
787 if (engine->last_context != request->ctx) {
788 if (engine->last_context)
789 intel_lr_context_unpin(engine->last_context, engine);
790 if (request->ctx != request->i915->kernel_context) {
791 intel_lr_context_pin(request->ctx, engine);
792 engine->last_context = request->ctx;
793 } else {
794 engine->last_context = NULL;
795 }
796 }
797
798 if (dev_priv->guc.execbuf_client)
799 i915_guc_submit(dev_priv->guc.execbuf_client, request);
800 else
801 execlists_context_queue(request);
802
803 return 0;
804}
805
806static void __wrap_ring_buffer(struct intel_ringbuffer *ringbuf)
807{
808 uint32_t __iomem *virt;
809 int rem = ringbuf->size - ringbuf->tail;
810
811 virt = ringbuf->virtual_start + ringbuf->tail;
812 rem /= 4;
813 while (rem--)
814 iowrite32(MI_NOOP, virt++);
815
816 ringbuf->tail = 0;
817 intel_ring_update_space(ringbuf);
818}
819
820static int logical_ring_prepare(struct drm_i915_gem_request *req, int bytes)
821{
822 struct intel_ringbuffer *ringbuf = req->ringbuf;
823 int remain_usable = ringbuf->effective_size - ringbuf->tail;
824 int remain_actual = ringbuf->size - ringbuf->tail;
825 int ret, total_bytes, wait_bytes = 0;
826 bool need_wrap = false;
827
828 if (ringbuf->reserved_in_use)
829 total_bytes = bytes;
830 else
831 total_bytes = bytes + ringbuf->reserved_size;
832
833 if (unlikely(bytes > remain_usable)) {
834 /*
835 * Not enough space for the basic request. So need to flush
836 * out the remainder and then wait for base + reserved.
837 */
838 wait_bytes = remain_actual + total_bytes;
839 need_wrap = true;
840 } else {
841 if (unlikely(total_bytes > remain_usable)) {
842 /*
843 * The base request will fit but the reserved space
844 * falls off the end. So don't need an immediate wrap
845 * and only need to effectively wait for the reserved
846 * size space from the start of ringbuffer.
847 */
848 wait_bytes = remain_actual + ringbuf->reserved_size;
849 } else if (total_bytes > ringbuf->space) {
850 /* No wrapping required, just waiting. */
851 wait_bytes = total_bytes;
852 }
853 }
854
855 if (wait_bytes) {
856 ret = logical_ring_wait_for_space(req, wait_bytes);
857 if (unlikely(ret))
858 return ret;
859
860 if (need_wrap)
861 __wrap_ring_buffer(ringbuf);
862 }
863
864 return 0;
865}
866
867/**
868 * intel_logical_ring_begin() - prepare the logical ringbuffer to accept some commands
869 *
870 * @req: The request to start some new work for
871 * @num_dwords: number of DWORDs that we plan to write to the ringbuffer.
872 *
873 * The ringbuffer might not be ready to accept the commands right away (maybe it needs to
874 * be wrapped, or wait a bit for the tail to be updated). This function takes care of that
875 * and also preallocates a request (every workload submission is still mediated through
876 * requests, same as it did with legacy ringbuffer submission).
877 *
878 * Return: non-zero if the ringbuffer is not ready to be written to.
879 */
880int intel_logical_ring_begin(struct drm_i915_gem_request *req, int num_dwords)
881{
882 struct drm_i915_private *dev_priv;
883 int ret;
884
885 WARN_ON(req == NULL);
886 dev_priv = req->ring->dev->dev_private;
887
888 ret = i915_gem_check_wedge(&dev_priv->gpu_error,
889 dev_priv->mm.interruptible);
890 if (ret)
891 return ret;
892
893 ret = logical_ring_prepare(req, num_dwords * sizeof(uint32_t));
894 if (ret)
895 return ret;
896
897 req->ringbuf->space -= num_dwords * sizeof(uint32_t);
898 return 0;
899}
900
901int intel_logical_ring_reserve_space(struct drm_i915_gem_request *request)
902{
903 /*
904 * The first call merely notes the reserve request and is common for
905 * all back ends. The subsequent localised _begin() call actually
906 * ensures that the reservation is available. Without the begin, if
907 * the request creator immediately submitted the request without
908 * adding any commands to it then there might not actually be
909 * sufficient room for the submission commands.
910 */
911 intel_ring_reserved_space_reserve(request->ringbuf, MIN_SPACE_FOR_ADD_REQUEST);
912
913 return intel_logical_ring_begin(request, 0);
914}
915
916/**
917 * execlists_submission() - submit a batchbuffer for execution, Execlists style
918 * @dev: DRM device.
919 * @file: DRM file.
920 * @ring: Engine Command Streamer to submit to.
921 * @ctx: Context to employ for this submission.
922 * @args: execbuffer call arguments.
923 * @vmas: list of vmas.
924 * @batch_obj: the batchbuffer to submit.
925 * @exec_start: batchbuffer start virtual address pointer.
926 * @dispatch_flags: translated execbuffer call flags.
927 *
928 * This is the evil twin version of i915_gem_ringbuffer_submission. It abstracts
929 * away the submission details of the execbuffer ioctl call.
930 *
931 * Return: non-zero if the submission fails.
932 */
933int intel_execlists_submission(struct i915_execbuffer_params *params,
934 struct drm_i915_gem_execbuffer2 *args,
935 struct list_head *vmas)
936{
937 struct drm_device *dev = params->dev;
938 struct intel_engine_cs *ring = params->ring;
939 struct drm_i915_private *dev_priv = dev->dev_private;
940 struct intel_ringbuffer *ringbuf = params->ctx->engine[ring->id].ringbuf;
941 u64 exec_start;
942 int instp_mode;
943 u32 instp_mask;
944 int ret;
945
946 instp_mode = args->flags & I915_EXEC_CONSTANTS_MASK;
947 instp_mask = I915_EXEC_CONSTANTS_MASK;
948 switch (instp_mode) {
949 case I915_EXEC_CONSTANTS_REL_GENERAL:
950 case I915_EXEC_CONSTANTS_ABSOLUTE:
951 case I915_EXEC_CONSTANTS_REL_SURFACE:
952 if (instp_mode != 0 && ring != &dev_priv->ring[RCS]) {
953 DRM_DEBUG("non-0 rel constants mode on non-RCS\n");
954 return -EINVAL;
955 }
956
957 if (instp_mode != dev_priv->relative_constants_mode) {
958 if (instp_mode == I915_EXEC_CONSTANTS_REL_SURFACE) {
959 DRM_DEBUG("rel surface constants mode invalid on gen5+\n");
960 return -EINVAL;
961 }
962
963 /* The HW changed the meaning on this bit on gen6 */
964 instp_mask &= ~I915_EXEC_CONSTANTS_REL_SURFACE;
965 }
966 break;
967 default:
968 DRM_DEBUG("execbuf with unknown constants: %d\n", instp_mode);
969 return -EINVAL;
970 }
971
972 if (args->flags & I915_EXEC_GEN7_SOL_RESET) {
973 DRM_DEBUG("sol reset is gen7 only\n");
974 return -EINVAL;
975 }
976
977 ret = execlists_move_to_gpu(params->request, vmas);
978 if (ret)
979 return ret;
980
981 if (ring == &dev_priv->ring[RCS] &&
982 instp_mode != dev_priv->relative_constants_mode) {
983 ret = intel_logical_ring_begin(params->request, 4);
984 if (ret)
985 return ret;
986
987 intel_logical_ring_emit(ringbuf, MI_NOOP);
988 intel_logical_ring_emit(ringbuf, MI_LOAD_REGISTER_IMM(1));
989 intel_logical_ring_emit_reg(ringbuf, INSTPM);
990 intel_logical_ring_emit(ringbuf, instp_mask << 16 | instp_mode);
991 intel_logical_ring_advance(ringbuf);
992
993 dev_priv->relative_constants_mode = instp_mode;
994 }
995
996 exec_start = params->batch_obj_vm_offset +
997 args->batch_start_offset;
998
999 ret = ring->emit_bb_start(params->request, exec_start, params->dispatch_flags);
1000 if (ret)
1001 return ret;
1002
1003 trace_i915_gem_ring_dispatch(params->request, params->dispatch_flags);
1004
1005 i915_gem_execbuffer_move_to_active(vmas, params->request);
1006 i915_gem_execbuffer_retire_commands(params);
1007
1008 return 0;
1009}
1010
1011void intel_execlists_retire_requests(struct intel_engine_cs *ring)
1012{
1013 struct drm_i915_gem_request *req, *tmp;
1014 struct list_head retired_list;
1015
1016 WARN_ON(!mutex_is_locked(&ring->dev->struct_mutex));
1017 if (list_empty(&ring->execlist_retired_req_list))
1018 return;
1019
1020 INIT_LIST_HEAD(&retired_list);
1021 spin_lock_irq(&ring->execlist_lock);
1022 list_replace_init(&ring->execlist_retired_req_list, &retired_list);
1023 spin_unlock_irq(&ring->execlist_lock);
1024
1025 list_for_each_entry_safe(req, tmp, &retired_list, execlist_link) {
1026 struct intel_context *ctx = req->ctx;
1027 struct drm_i915_gem_object *ctx_obj =
1028 ctx->engine[ring->id].state;
1029
1030 if (ctx_obj && (ctx != req->i915->kernel_context))
1031 intel_lr_context_unpin(ctx, ring);
1032
1033 list_del(&req->execlist_link);
1034 i915_gem_request_unreference(req);
1035 }
1036}
1037
1038void intel_logical_ring_stop(struct intel_engine_cs *ring)
1039{
1040 struct drm_i915_private *dev_priv = ring->dev->dev_private;
1041 int ret;
1042
1043 if (!intel_ring_initialized(ring))
1044 return;
1045
1046 ret = intel_ring_idle(ring);
1047 if (ret && !i915_reset_in_progress(&to_i915(ring->dev)->gpu_error))
1048 DRM_ERROR("failed to quiesce %s whilst cleaning up: %d\n",
1049 ring->name, ret);
1050
1051 /* TODO: Is this correct with Execlists enabled? */
1052 I915_WRITE_MODE(ring, _MASKED_BIT_ENABLE(STOP_RING));
1053 if (wait_for_atomic((I915_READ_MODE(ring) & MODE_IDLE) != 0, 1000)) {
1054 DRM_ERROR("%s :timed out trying to stop ring\n", ring->name);
1055 return;
1056 }
1057 I915_WRITE_MODE(ring, _MASKED_BIT_DISABLE(STOP_RING));
1058}
1059
1060int logical_ring_flush_all_caches(struct drm_i915_gem_request *req)
1061{
1062 struct intel_engine_cs *ring = req->ring;
1063 int ret;
1064
1065 if (!ring->gpu_caches_dirty)
1066 return 0;
1067
1068 ret = ring->emit_flush(req, 0, I915_GEM_GPU_DOMAINS);
1069 if (ret)
1070 return ret;
1071
1072 ring->gpu_caches_dirty = false;
1073 return 0;
1074}
1075
1076static int intel_lr_context_do_pin(struct intel_context *ctx,
1077 struct intel_engine_cs *ring)
1078{
1079 struct drm_device *dev = ring->dev;
1080 struct drm_i915_private *dev_priv = dev->dev_private;
1081 struct drm_i915_gem_object *ctx_obj = ctx->engine[ring->id].state;
1082 struct intel_ringbuffer *ringbuf = ctx->engine[ring->id].ringbuf;
1083 struct page *lrc_state_page;
1084 uint32_t *lrc_reg_state;
1085 int ret;
1086
1087 WARN_ON(!mutex_is_locked(&ring->dev->struct_mutex));
1088
1089 ret = i915_gem_obj_ggtt_pin(ctx_obj, GEN8_LR_CONTEXT_ALIGN,
1090 PIN_OFFSET_BIAS | GUC_WOPCM_TOP);
1091 if (ret)
1092 return ret;
1093
1094 lrc_state_page = i915_gem_object_get_dirty_page(ctx_obj, LRC_STATE_PN);
1095 if (WARN_ON(!lrc_state_page)) {
1096 ret = -ENODEV;
1097 goto unpin_ctx_obj;
1098 }
1099
1100 ret = intel_pin_and_map_ringbuffer_obj(ring->dev, ringbuf);
1101 if (ret)
1102 goto unpin_ctx_obj;
1103
1104 ctx->engine[ring->id].lrc_vma = i915_gem_obj_to_ggtt(ctx_obj);
1105 intel_lr_context_descriptor_update(ctx, ring);
1106 lrc_reg_state = kmap(lrc_state_page);
1107 lrc_reg_state[CTX_RING_BUFFER_START+1] = ringbuf->vma->node.start;
1108 ctx->engine[ring->id].lrc_reg_state = lrc_reg_state;
1109 ctx_obj->dirty = true;
1110
1111 /* Invalidate GuC TLB. */
1112 if (i915.enable_guc_submission)
1113 I915_WRITE(GEN8_GTCR, GEN8_GTCR_INVALIDATE);
1114
1115 return ret;
1116
1117unpin_ctx_obj:
1118 i915_gem_object_ggtt_unpin(ctx_obj);
1119
1120 return ret;
1121}
1122
1123static int intel_lr_context_pin(struct intel_context *ctx,
1124 struct intel_engine_cs *engine)
1125{
1126 int ret = 0;
1127
1128 if (ctx->engine[engine->id].pin_count++ == 0) {
1129 ret = intel_lr_context_do_pin(ctx, engine);
1130 if (ret)
1131 goto reset_pin_count;
1132
1133 i915_gem_context_reference(ctx);
1134 }
1135 return ret;
1136
1137reset_pin_count:
1138 ctx->engine[engine->id].pin_count = 0;
1139 return ret;
1140}
1141
1142void intel_lr_context_unpin(struct intel_context *ctx,
1143 struct intel_engine_cs *engine)
1144{
1145 struct drm_i915_gem_object *ctx_obj = ctx->engine[engine->id].state;
1146
1147 WARN_ON(!mutex_is_locked(&ctx->i915->dev->struct_mutex));
1148 if (--ctx->engine[engine->id].pin_count == 0) {
1149 kunmap(kmap_to_page(ctx->engine[engine->id].lrc_reg_state));
1150 intel_unpin_ringbuffer_obj(ctx->engine[engine->id].ringbuf);
1151 i915_gem_object_ggtt_unpin(ctx_obj);
1152 ctx->engine[engine->id].lrc_vma = NULL;
1153 ctx->engine[engine->id].lrc_desc = 0;
1154 ctx->engine[engine->id].lrc_reg_state = NULL;
1155
1156 i915_gem_context_unreference(ctx);
1157 }
1158}
1159
1160static int intel_logical_ring_workarounds_emit(struct drm_i915_gem_request *req)
1161{
1162 int ret, i;
1163 struct intel_engine_cs *ring = req->ring;
1164 struct intel_ringbuffer *ringbuf = req->ringbuf;
1165 struct drm_device *dev = ring->dev;
1166 struct drm_i915_private *dev_priv = dev->dev_private;
1167 struct i915_workarounds *w = &dev_priv->workarounds;
1168
1169 if (w->count == 0)
1170 return 0;
1171
1172 ring->gpu_caches_dirty = true;
1173 ret = logical_ring_flush_all_caches(req);
1174 if (ret)
1175 return ret;
1176
1177 ret = intel_logical_ring_begin(req, w->count * 2 + 2);
1178 if (ret)
1179 return ret;
1180
1181 intel_logical_ring_emit(ringbuf, MI_LOAD_REGISTER_IMM(w->count));
1182 for (i = 0; i < w->count; i++) {
1183 intel_logical_ring_emit_reg(ringbuf, w->reg[i].addr);
1184 intel_logical_ring_emit(ringbuf, w->reg[i].value);
1185 }
1186 intel_logical_ring_emit(ringbuf, MI_NOOP);
1187
1188 intel_logical_ring_advance(ringbuf);
1189
1190 ring->gpu_caches_dirty = true;
1191 ret = logical_ring_flush_all_caches(req);
1192 if (ret)
1193 return ret;
1194
1195 return 0;
1196}
1197
1198#define wa_ctx_emit(batch, index, cmd) \
1199 do { \
1200 int __index = (index)++; \
1201 if (WARN_ON(__index >= (PAGE_SIZE / sizeof(uint32_t)))) { \
1202 return -ENOSPC; \
1203 } \
1204 batch[__index] = (cmd); \
1205 } while (0)
1206
1207#define wa_ctx_emit_reg(batch, index, reg) \
1208 wa_ctx_emit((batch), (index), i915_mmio_reg_offset(reg))
1209
1210/*
1211 * In this WA we need to set GEN8_L3SQCREG4[21:21] and reset it after
1212 * PIPE_CONTROL instruction. This is required for the flush to happen correctly
1213 * but there is a slight complication as this is applied in WA batch where the
1214 * values are only initialized once so we cannot take register value at the
1215 * beginning and reuse it further; hence we save its value to memory, upload a
1216 * constant value with bit21 set and then we restore it back with the saved value.
1217 * To simplify the WA, a constant value is formed by using the default value
1218 * of this register. This shouldn't be a problem because we are only modifying
1219 * it for a short period and this batch in non-premptible. We can ofcourse
1220 * use additional instructions that read the actual value of the register
1221 * at that time and set our bit of interest but it makes the WA complicated.
1222 *
1223 * This WA is also required for Gen9 so extracting as a function avoids
1224 * code duplication.
1225 */
1226static inline int gen8_emit_flush_coherentl3_wa(struct intel_engine_cs *ring,
1227 uint32_t *const batch,
1228 uint32_t index)
1229{
1230 uint32_t l3sqc4_flush = (0x40400000 | GEN8_LQSC_FLUSH_COHERENT_LINES);
1231
1232 /*
1233 * WaDisableLSQCROPERFforOCL:skl
1234 * This WA is implemented in skl_init_clock_gating() but since
1235 * this batch updates GEN8_L3SQCREG4 with default value we need to
1236 * set this bit here to retain the WA during flush.
1237 */
1238 if (IS_SKL_REVID(ring->dev, 0, SKL_REVID_E0))
1239 l3sqc4_flush |= GEN8_LQSC_RO_PERF_DIS;
1240
1241 wa_ctx_emit(batch, index, (MI_STORE_REGISTER_MEM_GEN8 |
1242 MI_SRM_LRM_GLOBAL_GTT));
1243 wa_ctx_emit_reg(batch, index, GEN8_L3SQCREG4);
1244 wa_ctx_emit(batch, index, ring->scratch.gtt_offset + 256);
1245 wa_ctx_emit(batch, index, 0);
1246
1247 wa_ctx_emit(batch, index, MI_LOAD_REGISTER_IMM(1));
1248 wa_ctx_emit_reg(batch, index, GEN8_L3SQCREG4);
1249 wa_ctx_emit(batch, index, l3sqc4_flush);
1250
1251 wa_ctx_emit(batch, index, GFX_OP_PIPE_CONTROL(6));
1252 wa_ctx_emit(batch, index, (PIPE_CONTROL_CS_STALL |
1253 PIPE_CONTROL_DC_FLUSH_ENABLE));
1254 wa_ctx_emit(batch, index, 0);
1255 wa_ctx_emit(batch, index, 0);
1256 wa_ctx_emit(batch, index, 0);
1257 wa_ctx_emit(batch, index, 0);
1258
1259 wa_ctx_emit(batch, index, (MI_LOAD_REGISTER_MEM_GEN8 |
1260 MI_SRM_LRM_GLOBAL_GTT));
1261 wa_ctx_emit_reg(batch, index, GEN8_L3SQCREG4);
1262 wa_ctx_emit(batch, index, ring->scratch.gtt_offset + 256);
1263 wa_ctx_emit(batch, index, 0);
1264
1265 return index;
1266}
1267
1268static inline uint32_t wa_ctx_start(struct i915_wa_ctx_bb *wa_ctx,
1269 uint32_t offset,
1270 uint32_t start_alignment)
1271{
1272 return wa_ctx->offset = ALIGN(offset, start_alignment);
1273}
1274
1275static inline int wa_ctx_end(struct i915_wa_ctx_bb *wa_ctx,
1276 uint32_t offset,
1277 uint32_t size_alignment)
1278{
1279 wa_ctx->size = offset - wa_ctx->offset;
1280
1281 WARN(wa_ctx->size % size_alignment,
1282 "wa_ctx_bb failed sanity checks: size %d is not aligned to %d\n",
1283 wa_ctx->size, size_alignment);
1284 return 0;
1285}
1286
1287/**
1288 * gen8_init_indirectctx_bb() - initialize indirect ctx batch with WA
1289 *
1290 * @ring: only applicable for RCS
1291 * @wa_ctx: structure representing wa_ctx
1292 * offset: specifies start of the batch, should be cache-aligned. This is updated
1293 * with the offset value received as input.
1294 * size: size of the batch in DWORDS but HW expects in terms of cachelines
1295 * @batch: page in which WA are loaded
1296 * @offset: This field specifies the start of the batch, it should be
1297 * cache-aligned otherwise it is adjusted accordingly.
1298 * Typically we only have one indirect_ctx and per_ctx batch buffer which are
1299 * initialized at the beginning and shared across all contexts but this field
1300 * helps us to have multiple batches at different offsets and select them based
1301 * on a criteria. At the moment this batch always start at the beginning of the page
1302 * and at this point we don't have multiple wa_ctx batch buffers.
1303 *
1304 * The number of WA applied are not known at the beginning; we use this field
1305 * to return the no of DWORDS written.
1306 *
1307 * It is to be noted that this batch does not contain MI_BATCH_BUFFER_END
1308 * so it adds NOOPs as padding to make it cacheline aligned.
1309 * MI_BATCH_BUFFER_END will be added to perctx batch and both of them together
1310 * makes a complete batch buffer.
1311 *
1312 * Return: non-zero if we exceed the PAGE_SIZE limit.
1313 */
1314
1315static int gen8_init_indirectctx_bb(struct intel_engine_cs *ring,
1316 struct i915_wa_ctx_bb *wa_ctx,
1317 uint32_t *const batch,
1318 uint32_t *offset)
1319{
1320 uint32_t scratch_addr;
1321 uint32_t index = wa_ctx_start(wa_ctx, *offset, CACHELINE_DWORDS);
1322
1323 /* WaDisableCtxRestoreArbitration:bdw,chv */
1324 wa_ctx_emit(batch, index, MI_ARB_ON_OFF | MI_ARB_DISABLE);
1325
1326 /* WaFlushCoherentL3CacheLinesAtContextSwitch:bdw */
1327 if (IS_BROADWELL(ring->dev)) {
1328 int rc = gen8_emit_flush_coherentl3_wa(ring, batch, index);
1329 if (rc < 0)
1330 return rc;
1331 index = rc;
1332 }
1333
1334 /* WaClearSlmSpaceAtContextSwitch:bdw,chv */
1335 /* Actual scratch location is at 128 bytes offset */
1336 scratch_addr = ring->scratch.gtt_offset + 2*CACHELINE_BYTES;
1337
1338 wa_ctx_emit(batch, index, GFX_OP_PIPE_CONTROL(6));
1339 wa_ctx_emit(batch, index, (PIPE_CONTROL_FLUSH_L3 |
1340 PIPE_CONTROL_GLOBAL_GTT_IVB |
1341 PIPE_CONTROL_CS_STALL |
1342 PIPE_CONTROL_QW_WRITE));
1343 wa_ctx_emit(batch, index, scratch_addr);
1344 wa_ctx_emit(batch, index, 0);
1345 wa_ctx_emit(batch, index, 0);
1346 wa_ctx_emit(batch, index, 0);
1347
1348 /* Pad to end of cacheline */
1349 while (index % CACHELINE_DWORDS)
1350 wa_ctx_emit(batch, index, MI_NOOP);
1351
1352 /*
1353 * MI_BATCH_BUFFER_END is not required in Indirect ctx BB because
1354 * execution depends on the length specified in terms of cache lines
1355 * in the register CTX_RCS_INDIRECT_CTX
1356 */
1357
1358 return wa_ctx_end(wa_ctx, *offset = index, CACHELINE_DWORDS);
1359}
1360
1361/**
1362 * gen8_init_perctx_bb() - initialize per ctx batch with WA
1363 *
1364 * @ring: only applicable for RCS
1365 * @wa_ctx: structure representing wa_ctx
1366 * offset: specifies start of the batch, should be cache-aligned.
1367 * size: size of the batch in DWORDS but HW expects in terms of cachelines
1368 * @batch: page in which WA are loaded
1369 * @offset: This field specifies the start of this batch.
1370 * This batch is started immediately after indirect_ctx batch. Since we ensure
1371 * that indirect_ctx ends on a cacheline this batch is aligned automatically.
1372 *
1373 * The number of DWORDS written are returned using this field.
1374 *
1375 * This batch is terminated with MI_BATCH_BUFFER_END and so we need not add padding
1376 * to align it with cacheline as padding after MI_BATCH_BUFFER_END is redundant.
1377 */
1378static int gen8_init_perctx_bb(struct intel_engine_cs *ring,
1379 struct i915_wa_ctx_bb *wa_ctx,
1380 uint32_t *const batch,
1381 uint32_t *offset)
1382{
1383 uint32_t index = wa_ctx_start(wa_ctx, *offset, CACHELINE_DWORDS);
1384
1385 /* WaDisableCtxRestoreArbitration:bdw,chv */
1386 wa_ctx_emit(batch, index, MI_ARB_ON_OFF | MI_ARB_ENABLE);
1387
1388 wa_ctx_emit(batch, index, MI_BATCH_BUFFER_END);
1389
1390 return wa_ctx_end(wa_ctx, *offset = index, 1);
1391}
1392
1393static int gen9_init_indirectctx_bb(struct intel_engine_cs *ring,
1394 struct i915_wa_ctx_bb *wa_ctx,
1395 uint32_t *const batch,
1396 uint32_t *offset)
1397{
1398 int ret;
1399 struct drm_device *dev = ring->dev;
1400 uint32_t index = wa_ctx_start(wa_ctx, *offset, CACHELINE_DWORDS);
1401
1402 /* WaDisableCtxRestoreArbitration:skl,bxt */
1403 if (IS_SKL_REVID(dev, 0, SKL_REVID_D0) ||
1404 IS_BXT_REVID(dev, 0, BXT_REVID_A1))
1405 wa_ctx_emit(batch, index, MI_ARB_ON_OFF | MI_ARB_DISABLE);
1406
1407 /* WaFlushCoherentL3CacheLinesAtContextSwitch:skl,bxt */
1408 ret = gen8_emit_flush_coherentl3_wa(ring, batch, index);
1409 if (ret < 0)
1410 return ret;
1411 index = ret;
1412
1413 /* Pad to end of cacheline */
1414 while (index % CACHELINE_DWORDS)
1415 wa_ctx_emit(batch, index, MI_NOOP);
1416
1417 return wa_ctx_end(wa_ctx, *offset = index, CACHELINE_DWORDS);
1418}
1419
1420static int gen9_init_perctx_bb(struct intel_engine_cs *ring,
1421 struct i915_wa_ctx_bb *wa_ctx,
1422 uint32_t *const batch,
1423 uint32_t *offset)
1424{
1425 struct drm_device *dev = ring->dev;
1426 uint32_t index = wa_ctx_start(wa_ctx, *offset, CACHELINE_DWORDS);
1427
1428 /* WaSetDisablePixMaskCammingAndRhwoInCommonSliceChicken:skl,bxt */
1429 if (IS_SKL_REVID(dev, 0, SKL_REVID_B0) ||
1430 IS_BXT_REVID(dev, 0, BXT_REVID_A1)) {
1431 wa_ctx_emit(batch, index, MI_LOAD_REGISTER_IMM(1));
1432 wa_ctx_emit_reg(batch, index, GEN9_SLICE_COMMON_ECO_CHICKEN0);
1433 wa_ctx_emit(batch, index,
1434 _MASKED_BIT_ENABLE(DISABLE_PIXEL_MASK_CAMMING));
1435 wa_ctx_emit(batch, index, MI_NOOP);
1436 }
1437
1438 /* WaDisableCtxRestoreArbitration:skl,bxt */
1439 if (IS_SKL_REVID(dev, 0, SKL_REVID_D0) ||
1440 IS_BXT_REVID(dev, 0, BXT_REVID_A1))
1441 wa_ctx_emit(batch, index, MI_ARB_ON_OFF | MI_ARB_ENABLE);
1442
1443 wa_ctx_emit(batch, index, MI_BATCH_BUFFER_END);
1444
1445 return wa_ctx_end(wa_ctx, *offset = index, 1);
1446}
1447
1448static int lrc_setup_wa_ctx_obj(struct intel_engine_cs *ring, u32 size)
1449{
1450 int ret;
1451
1452 ring->wa_ctx.obj = i915_gem_alloc_object(ring->dev, PAGE_ALIGN(size));
1453 if (!ring->wa_ctx.obj) {
1454 DRM_DEBUG_DRIVER("alloc LRC WA ctx backing obj failed.\n");
1455 return -ENOMEM;
1456 }
1457
1458 ret = i915_gem_obj_ggtt_pin(ring->wa_ctx.obj, PAGE_SIZE, 0);
1459 if (ret) {
1460 DRM_DEBUG_DRIVER("pin LRC WA ctx backing obj failed: %d\n",
1461 ret);
1462 drm_gem_object_unreference(&ring->wa_ctx.obj->base);
1463 return ret;
1464 }
1465
1466 return 0;
1467}
1468
1469static void lrc_destroy_wa_ctx_obj(struct intel_engine_cs *ring)
1470{
1471 if (ring->wa_ctx.obj) {
1472 i915_gem_object_ggtt_unpin(ring->wa_ctx.obj);
1473 drm_gem_object_unreference(&ring->wa_ctx.obj->base);
1474 ring->wa_ctx.obj = NULL;
1475 }
1476}
1477
1478static int intel_init_workaround_bb(struct intel_engine_cs *ring)
1479{
1480 int ret;
1481 uint32_t *batch;
1482 uint32_t offset;
1483 struct page *page;
1484 struct i915_ctx_workarounds *wa_ctx = &ring->wa_ctx;
1485
1486 WARN_ON(ring->id != RCS);
1487
1488 /* update this when WA for higher Gen are added */
1489 if (INTEL_INFO(ring->dev)->gen > 9) {
1490 DRM_ERROR("WA batch buffer is not initialized for Gen%d\n",
1491 INTEL_INFO(ring->dev)->gen);
1492 return 0;
1493 }
1494
1495 /* some WA perform writes to scratch page, ensure it is valid */
1496 if (ring->scratch.obj == NULL) {
1497 DRM_ERROR("scratch page not allocated for %s\n", ring->name);
1498 return -EINVAL;
1499 }
1500
1501 ret = lrc_setup_wa_ctx_obj(ring, PAGE_SIZE);
1502 if (ret) {
1503 DRM_DEBUG_DRIVER("Failed to setup context WA page: %d\n", ret);
1504 return ret;
1505 }
1506
1507 page = i915_gem_object_get_dirty_page(wa_ctx->obj, 0);
1508 batch = kmap_atomic(page);
1509 offset = 0;
1510
1511 if (INTEL_INFO(ring->dev)->gen == 8) {
1512 ret = gen8_init_indirectctx_bb(ring,
1513 &wa_ctx->indirect_ctx,
1514 batch,
1515 &offset);
1516 if (ret)
1517 goto out;
1518
1519 ret = gen8_init_perctx_bb(ring,
1520 &wa_ctx->per_ctx,
1521 batch,
1522 &offset);
1523 if (ret)
1524 goto out;
1525 } else if (INTEL_INFO(ring->dev)->gen == 9) {
1526 ret = gen9_init_indirectctx_bb(ring,
1527 &wa_ctx->indirect_ctx,
1528 batch,
1529 &offset);
1530 if (ret)
1531 goto out;
1532
1533 ret = gen9_init_perctx_bb(ring,
1534 &wa_ctx->per_ctx,
1535 batch,
1536 &offset);
1537 if (ret)
1538 goto out;
1539 }
1540
1541out:
1542 kunmap_atomic(batch);
1543 if (ret)
1544 lrc_destroy_wa_ctx_obj(ring);
1545
1546 return ret;
1547}
1548
1549static int gen8_init_common_ring(struct intel_engine_cs *ring)
1550{
1551 struct drm_device *dev = ring->dev;
1552 struct drm_i915_private *dev_priv = dev->dev_private;
1553 u8 next_context_status_buffer_hw;
1554
1555 lrc_setup_hardware_status_page(ring,
1556 dev_priv->kernel_context->engine[ring->id].state);
1557
1558 I915_WRITE_IMR(ring, ~(ring->irq_enable_mask | ring->irq_keep_mask));
1559 I915_WRITE(RING_HWSTAM(ring->mmio_base), 0xffffffff);
1560
1561 I915_WRITE(RING_MODE_GEN7(ring),
1562 _MASKED_BIT_DISABLE(GFX_REPLAY_MODE) |
1563 _MASKED_BIT_ENABLE(GFX_RUN_LIST_ENABLE));
1564 POSTING_READ(RING_MODE_GEN7(ring));
1565
1566 /*
1567 * Instead of resetting the Context Status Buffer (CSB) read pointer to
1568 * zero, we need to read the write pointer from hardware and use its
1569 * value because "this register is power context save restored".
1570 * Effectively, these states have been observed:
1571 *
1572 * | Suspend-to-idle (freeze) | Suspend-to-RAM (mem) |
1573 * BDW | CSB regs not reset | CSB regs reset |
1574 * CHT | CSB regs not reset | CSB regs not reset |
1575 * SKL | ? | ? |
1576 * BXT | ? | ? |
1577 */
1578 next_context_status_buffer_hw =
1579 GEN8_CSB_WRITE_PTR(I915_READ(RING_CONTEXT_STATUS_PTR(ring)));
1580
1581 /*
1582 * When the CSB registers are reset (also after power-up / gpu reset),
1583 * CSB write pointer is set to all 1's, which is not valid, use '5' in
1584 * this special case, so the first element read is CSB[0].
1585 */
1586 if (next_context_status_buffer_hw == GEN8_CSB_PTR_MASK)
1587 next_context_status_buffer_hw = (GEN8_CSB_ENTRIES - 1);
1588
1589 ring->next_context_status_buffer = next_context_status_buffer_hw;
1590 DRM_DEBUG_DRIVER("Execlists enabled for %s\n", ring->name);
1591
1592 memset(&ring->hangcheck, 0, sizeof(ring->hangcheck));
1593
1594 return 0;
1595}
1596
1597static int gen8_init_render_ring(struct intel_engine_cs *ring)
1598{
1599 struct drm_device *dev = ring->dev;
1600 struct drm_i915_private *dev_priv = dev->dev_private;
1601 int ret;
1602
1603 ret = gen8_init_common_ring(ring);
1604 if (ret)
1605 return ret;
1606
1607 /* We need to disable the AsyncFlip performance optimisations in order
1608 * to use MI_WAIT_FOR_EVENT within the CS. It should already be
1609 * programmed to '1' on all products.
1610 *
1611 * WaDisableAsyncFlipPerfMode:snb,ivb,hsw,vlv,bdw,chv
1612 */
1613 I915_WRITE(MI_MODE, _MASKED_BIT_ENABLE(ASYNC_FLIP_PERF_DISABLE));
1614
1615 I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_FORCE_ORDERING));
1616
1617 return init_workarounds_ring(ring);
1618}
1619
1620static int gen9_init_render_ring(struct intel_engine_cs *ring)
1621{
1622 int ret;
1623
1624 ret = gen8_init_common_ring(ring);
1625 if (ret)
1626 return ret;
1627
1628 return init_workarounds_ring(ring);
1629}
1630
1631static int intel_logical_ring_emit_pdps(struct drm_i915_gem_request *req)
1632{
1633 struct i915_hw_ppgtt *ppgtt = req->ctx->ppgtt;
1634 struct intel_engine_cs *ring = req->ring;
1635 struct intel_ringbuffer *ringbuf = req->ringbuf;
1636 const int num_lri_cmds = GEN8_LEGACY_PDPES * 2;
1637 int i, ret;
1638
1639 ret = intel_logical_ring_begin(req, num_lri_cmds * 2 + 2);
1640 if (ret)
1641 return ret;
1642
1643 intel_logical_ring_emit(ringbuf, MI_LOAD_REGISTER_IMM(num_lri_cmds));
1644 for (i = GEN8_LEGACY_PDPES - 1; i >= 0; i--) {
1645 const dma_addr_t pd_daddr = i915_page_dir_dma_addr(ppgtt, i);
1646
1647 intel_logical_ring_emit_reg(ringbuf, GEN8_RING_PDP_UDW(ring, i));
1648 intel_logical_ring_emit(ringbuf, upper_32_bits(pd_daddr));
1649 intel_logical_ring_emit_reg(ringbuf, GEN8_RING_PDP_LDW(ring, i));
1650 intel_logical_ring_emit(ringbuf, lower_32_bits(pd_daddr));
1651 }
1652
1653 intel_logical_ring_emit(ringbuf, MI_NOOP);
1654 intel_logical_ring_advance(ringbuf);
1655
1656 return 0;
1657}
1658
1659static int gen8_emit_bb_start(struct drm_i915_gem_request *req,
1660 u64 offset, unsigned dispatch_flags)
1661{
1662 struct intel_ringbuffer *ringbuf = req->ringbuf;
1663 bool ppgtt = !(dispatch_flags & I915_DISPATCH_SECURE);
1664 int ret;
1665
1666 /* Don't rely in hw updating PDPs, specially in lite-restore.
1667 * Ideally, we should set Force PD Restore in ctx descriptor,
1668 * but we can't. Force Restore would be a second option, but
1669 * it is unsafe in case of lite-restore (because the ctx is
1670 * not idle). PML4 is allocated during ppgtt init so this is
1671 * not needed in 48-bit.*/
1672 if (req->ctx->ppgtt &&
1673 (intel_ring_flag(req->ring) & req->ctx->ppgtt->pd_dirty_rings)) {
1674 if (!USES_FULL_48BIT_PPGTT(req->i915) &&
1675 !intel_vgpu_active(req->i915->dev)) {
1676 ret = intel_logical_ring_emit_pdps(req);
1677 if (ret)
1678 return ret;
1679 }
1680
1681 req->ctx->ppgtt->pd_dirty_rings &= ~intel_ring_flag(req->ring);
1682 }
1683
1684 ret = intel_logical_ring_begin(req, 4);
1685 if (ret)
1686 return ret;
1687
1688 /* FIXME(BDW): Address space and security selectors. */
1689 intel_logical_ring_emit(ringbuf, MI_BATCH_BUFFER_START_GEN8 |
1690 (ppgtt<<8) |
1691 (dispatch_flags & I915_DISPATCH_RS ?
1692 MI_BATCH_RESOURCE_STREAMER : 0));
1693 intel_logical_ring_emit(ringbuf, lower_32_bits(offset));
1694 intel_logical_ring_emit(ringbuf, upper_32_bits(offset));
1695 intel_logical_ring_emit(ringbuf, MI_NOOP);
1696 intel_logical_ring_advance(ringbuf);
1697
1698 return 0;
1699}
1700
1701static bool gen8_logical_ring_get_irq(struct intel_engine_cs *ring)
1702{
1703 struct drm_device *dev = ring->dev;
1704 struct drm_i915_private *dev_priv = dev->dev_private;
1705 unsigned long flags;
1706
1707 if (WARN_ON(!intel_irqs_enabled(dev_priv)))
1708 return false;
1709
1710 spin_lock_irqsave(&dev_priv->irq_lock, flags);
1711 if (ring->irq_refcount++ == 0) {
1712 I915_WRITE_IMR(ring, ~(ring->irq_enable_mask | ring->irq_keep_mask));
1713 POSTING_READ(RING_IMR(ring->mmio_base));
1714 }
1715 spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
1716
1717 return true;
1718}
1719
1720static void gen8_logical_ring_put_irq(struct intel_engine_cs *ring)
1721{
1722 struct drm_device *dev = ring->dev;
1723 struct drm_i915_private *dev_priv = dev->dev_private;
1724 unsigned long flags;
1725
1726 spin_lock_irqsave(&dev_priv->irq_lock, flags);
1727 if (--ring->irq_refcount == 0) {
1728 I915_WRITE_IMR(ring, ~ring->irq_keep_mask);
1729 POSTING_READ(RING_IMR(ring->mmio_base));
1730 }
1731 spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
1732}
1733
1734static int gen8_emit_flush(struct drm_i915_gem_request *request,
1735 u32 invalidate_domains,
1736 u32 unused)
1737{
1738 struct intel_ringbuffer *ringbuf = request->ringbuf;
1739 struct intel_engine_cs *ring = ringbuf->ring;
1740 struct drm_device *dev = ring->dev;
1741 struct drm_i915_private *dev_priv = dev->dev_private;
1742 uint32_t cmd;
1743 int ret;
1744
1745 ret = intel_logical_ring_begin(request, 4);
1746 if (ret)
1747 return ret;
1748
1749 cmd = MI_FLUSH_DW + 1;
1750
1751 /* We always require a command barrier so that subsequent
1752 * commands, such as breadcrumb interrupts, are strictly ordered
1753 * wrt the contents of the write cache being flushed to memory
1754 * (and thus being coherent from the CPU).
1755 */
1756 cmd |= MI_FLUSH_DW_STORE_INDEX | MI_FLUSH_DW_OP_STOREDW;
1757
1758 if (invalidate_domains & I915_GEM_GPU_DOMAINS) {
1759 cmd |= MI_INVALIDATE_TLB;
1760 if (ring == &dev_priv->ring[VCS])
1761 cmd |= MI_INVALIDATE_BSD;
1762 }
1763
1764 intel_logical_ring_emit(ringbuf, cmd);
1765 intel_logical_ring_emit(ringbuf,
1766 I915_GEM_HWS_SCRATCH_ADDR |
1767 MI_FLUSH_DW_USE_GTT);
1768 intel_logical_ring_emit(ringbuf, 0); /* upper addr */
1769 intel_logical_ring_emit(ringbuf, 0); /* value */
1770 intel_logical_ring_advance(ringbuf);
1771
1772 return 0;
1773}
1774
1775static int gen8_emit_flush_render(struct drm_i915_gem_request *request,
1776 u32 invalidate_domains,
1777 u32 flush_domains)
1778{
1779 struct intel_ringbuffer *ringbuf = request->ringbuf;
1780 struct intel_engine_cs *ring = ringbuf->ring;
1781 u32 scratch_addr = ring->scratch.gtt_offset + 2 * CACHELINE_BYTES;
1782 bool vf_flush_wa = false;
1783 u32 flags = 0;
1784 int ret;
1785
1786 flags |= PIPE_CONTROL_CS_STALL;
1787
1788 if (flush_domains) {
1789 flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
1790 flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
1791 flags |= PIPE_CONTROL_DC_FLUSH_ENABLE;
1792 flags |= PIPE_CONTROL_FLUSH_ENABLE;
1793 }
1794
1795 if (invalidate_domains) {
1796 flags |= PIPE_CONTROL_TLB_INVALIDATE;
1797 flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
1798 flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
1799 flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
1800 flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
1801 flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
1802 flags |= PIPE_CONTROL_QW_WRITE;
1803 flags |= PIPE_CONTROL_GLOBAL_GTT_IVB;
1804
1805 /*
1806 * On GEN9: before VF_CACHE_INVALIDATE we need to emit a NULL
1807 * pipe control.
1808 */
1809 if (IS_GEN9(ring->dev))
1810 vf_flush_wa = true;
1811 }
1812
1813 ret = intel_logical_ring_begin(request, vf_flush_wa ? 12 : 6);
1814 if (ret)
1815 return ret;
1816
1817 if (vf_flush_wa) {
1818 intel_logical_ring_emit(ringbuf, GFX_OP_PIPE_CONTROL(6));
1819 intel_logical_ring_emit(ringbuf, 0);
1820 intel_logical_ring_emit(ringbuf, 0);
1821 intel_logical_ring_emit(ringbuf, 0);
1822 intel_logical_ring_emit(ringbuf, 0);
1823 intel_logical_ring_emit(ringbuf, 0);
1824 }
1825
1826 intel_logical_ring_emit(ringbuf, GFX_OP_PIPE_CONTROL(6));
1827 intel_logical_ring_emit(ringbuf, flags);
1828 intel_logical_ring_emit(ringbuf, scratch_addr);
1829 intel_logical_ring_emit(ringbuf, 0);
1830 intel_logical_ring_emit(ringbuf, 0);
1831 intel_logical_ring_emit(ringbuf, 0);
1832 intel_logical_ring_advance(ringbuf);
1833
1834 return 0;
1835}
1836
1837static u32 gen8_get_seqno(struct intel_engine_cs *ring, bool lazy_coherency)
1838{
1839 return intel_read_status_page(ring, I915_GEM_HWS_INDEX);
1840}
1841
1842static void gen8_set_seqno(struct intel_engine_cs *ring, u32 seqno)
1843{
1844 intel_write_status_page(ring, I915_GEM_HWS_INDEX, seqno);
1845}
1846
1847static u32 bxt_a_get_seqno(struct intel_engine_cs *ring, bool lazy_coherency)
1848{
1849
1850 /*
1851 * On BXT A steppings there is a HW coherency issue whereby the
1852 * MI_STORE_DATA_IMM storing the completed request's seqno
1853 * occasionally doesn't invalidate the CPU cache. Work around this by
1854 * clflushing the corresponding cacheline whenever the caller wants
1855 * the coherency to be guaranteed. Note that this cacheline is known
1856 * to be clean at this point, since we only write it in
1857 * bxt_a_set_seqno(), where we also do a clflush after the write. So
1858 * this clflush in practice becomes an invalidate operation.
1859 */
1860
1861 if (!lazy_coherency)
1862 intel_flush_status_page(ring, I915_GEM_HWS_INDEX);
1863
1864 return intel_read_status_page(ring, I915_GEM_HWS_INDEX);
1865}
1866
1867static void bxt_a_set_seqno(struct intel_engine_cs *ring, u32 seqno)
1868{
1869 intel_write_status_page(ring, I915_GEM_HWS_INDEX, seqno);
1870
1871 /* See bxt_a_get_seqno() explaining the reason for the clflush. */
1872 intel_flush_status_page(ring, I915_GEM_HWS_INDEX);
1873}
1874
1875/*
1876 * Reserve space for 2 NOOPs at the end of each request to be
1877 * used as a workaround for not being allowed to do lite
1878 * restore with HEAD==TAIL (WaIdleLiteRestore).
1879 */
1880#define WA_TAIL_DWORDS 2
1881
1882static inline u32 hws_seqno_address(struct intel_engine_cs *engine)
1883{
1884 return engine->status_page.gfx_addr + I915_GEM_HWS_INDEX_ADDR;
1885}
1886
1887static int gen8_emit_request(struct drm_i915_gem_request *request)
1888{
1889 struct intel_ringbuffer *ringbuf = request->ringbuf;
1890 int ret;
1891
1892 ret = intel_logical_ring_begin(request, 6 + WA_TAIL_DWORDS);
1893 if (ret)
1894 return ret;
1895
1896 /* w/a: bit 5 needs to be zero for MI_FLUSH_DW address. */
1897 BUILD_BUG_ON(I915_GEM_HWS_INDEX_ADDR & (1 << 5));
1898
1899 intel_logical_ring_emit(ringbuf,
1900 (MI_FLUSH_DW + 1) | MI_FLUSH_DW_OP_STOREDW);
1901 intel_logical_ring_emit(ringbuf,
1902 hws_seqno_address(request->ring) |
1903 MI_FLUSH_DW_USE_GTT);
1904 intel_logical_ring_emit(ringbuf, 0);
1905 intel_logical_ring_emit(ringbuf, i915_gem_request_get_seqno(request));
1906 intel_logical_ring_emit(ringbuf, MI_USER_INTERRUPT);
1907 intel_logical_ring_emit(ringbuf, MI_NOOP);
1908 return intel_logical_ring_advance_and_submit(request);
1909}
1910
1911static int gen8_emit_request_render(struct drm_i915_gem_request *request)
1912{
1913 struct intel_ringbuffer *ringbuf = request->ringbuf;
1914 int ret;
1915
1916 ret = intel_logical_ring_begin(request, 8 + WA_TAIL_DWORDS);
1917 if (ret)
1918 return ret;
1919
1920 /* We're using qword write, seqno should be aligned to 8 bytes. */
1921 BUILD_BUG_ON(I915_GEM_HWS_INDEX & 1);
1922
1923 /* w/a for post sync ops following a GPGPU operation we
1924 * need a prior CS_STALL, which is emitted by the flush
1925 * following the batch.
1926 */
1927 intel_logical_ring_emit(ringbuf, GFX_OP_PIPE_CONTROL(6));
1928 intel_logical_ring_emit(ringbuf,
1929 (PIPE_CONTROL_GLOBAL_GTT_IVB |
1930 PIPE_CONTROL_CS_STALL |
1931 PIPE_CONTROL_QW_WRITE));
1932 intel_logical_ring_emit(ringbuf, hws_seqno_address(request->ring));
1933 intel_logical_ring_emit(ringbuf, 0);
1934 intel_logical_ring_emit(ringbuf, i915_gem_request_get_seqno(request));
1935 /* We're thrashing one dword of HWS. */
1936 intel_logical_ring_emit(ringbuf, 0);
1937 intel_logical_ring_emit(ringbuf, MI_USER_INTERRUPT);
1938 intel_logical_ring_emit(ringbuf, MI_NOOP);
1939 return intel_logical_ring_advance_and_submit(request);
1940}
1941
1942static int intel_lr_context_render_state_init(struct drm_i915_gem_request *req)
1943{
1944 struct render_state so;
1945 int ret;
1946
1947 ret = i915_gem_render_state_prepare(req->ring, &so);
1948 if (ret)
1949 return ret;
1950
1951 if (so.rodata == NULL)
1952 return 0;
1953
1954 ret = req->ring->emit_bb_start(req, so.ggtt_offset,
1955 I915_DISPATCH_SECURE);
1956 if (ret)
1957 goto out;
1958
1959 ret = req->ring->emit_bb_start(req,
1960 (so.ggtt_offset + so.aux_batch_offset),
1961 I915_DISPATCH_SECURE);
1962 if (ret)
1963 goto out;
1964
1965 i915_vma_move_to_active(i915_gem_obj_to_ggtt(so.obj), req);
1966
1967out:
1968 i915_gem_render_state_fini(&so);
1969 return ret;
1970}
1971
1972static int gen8_init_rcs_context(struct drm_i915_gem_request *req)
1973{
1974 int ret;
1975
1976 ret = intel_logical_ring_workarounds_emit(req);
1977 if (ret)
1978 return ret;
1979
1980 ret = intel_rcs_context_init_mocs(req);
1981 /*
1982 * Failing to program the MOCS is non-fatal.The system will not
1983 * run at peak performance. So generate an error and carry on.
1984 */
1985 if (ret)
1986 DRM_ERROR("MOCS failed to program: expect performance issues.\n");
1987
1988 return intel_lr_context_render_state_init(req);
1989}
1990
1991/**
1992 * intel_logical_ring_cleanup() - deallocate the Engine Command Streamer
1993 *
1994 * @ring: Engine Command Streamer.
1995 *
1996 */
1997void intel_logical_ring_cleanup(struct intel_engine_cs *ring)
1998{
1999 struct drm_i915_private *dev_priv;
2000
2001 if (!intel_ring_initialized(ring))
2002 return;
2003
2004 dev_priv = ring->dev->dev_private;
2005
2006 if (ring->buffer) {
2007 intel_logical_ring_stop(ring);
2008 WARN_ON((I915_READ_MODE(ring) & MODE_IDLE) == 0);
2009 }
2010
2011 if (ring->cleanup)
2012 ring->cleanup(ring);
2013
2014 i915_cmd_parser_fini_ring(ring);
2015 i915_gem_batch_pool_fini(&ring->batch_pool);
2016
2017 if (ring->status_page.obj) {
2018 kunmap(sg_page(ring->status_page.obj->pages->sgl));
2019 ring->status_page.obj = NULL;
2020 }
2021
2022 ring->disable_lite_restore_wa = false;
2023 ring->ctx_desc_template = 0;
2024
2025 lrc_destroy_wa_ctx_obj(ring);
2026 ring->dev = NULL;
2027}
2028
2029static void
2030logical_ring_default_vfuncs(struct drm_device *dev,
2031 struct intel_engine_cs *ring)
2032{
2033 /* Default vfuncs which can be overriden by each engine. */
2034 ring->init_hw = gen8_init_common_ring;
2035 ring->emit_request = gen8_emit_request;
2036 ring->emit_flush = gen8_emit_flush;
2037 ring->irq_get = gen8_logical_ring_get_irq;
2038 ring->irq_put = gen8_logical_ring_put_irq;
2039 ring->emit_bb_start = gen8_emit_bb_start;
2040 if (IS_BXT_REVID(dev, 0, BXT_REVID_A1)) {
2041 ring->get_seqno = bxt_a_get_seqno;
2042 ring->set_seqno = bxt_a_set_seqno;
2043 } else {
2044 ring->get_seqno = gen8_get_seqno;
2045 ring->set_seqno = gen8_set_seqno;
2046 }
2047}
2048
2049static inline void
2050logical_ring_default_irqs(struct intel_engine_cs *ring, unsigned shift)
2051{
2052 ring->irq_enable_mask = GT_RENDER_USER_INTERRUPT << shift;
2053 ring->irq_keep_mask = GT_CONTEXT_SWITCH_INTERRUPT << shift;
2054}
2055
2056static int
2057logical_ring_init(struct drm_device *dev, struct intel_engine_cs *ring)
2058{
2059 struct intel_context *dctx = to_i915(dev)->kernel_context;
2060 int ret;
2061
2062 /* Intentionally left blank. */
2063 ring->buffer = NULL;
2064
2065 ring->dev = dev;
2066 INIT_LIST_HEAD(&ring->active_list);
2067 INIT_LIST_HEAD(&ring->request_list);
2068 i915_gem_batch_pool_init(dev, &ring->batch_pool);
2069 init_waitqueue_head(&ring->irq_queue);
2070
2071 INIT_LIST_HEAD(&ring->buffers);
2072 INIT_LIST_HEAD(&ring->execlist_queue);
2073 INIT_LIST_HEAD(&ring->execlist_retired_req_list);
2074 spin_lock_init(&ring->execlist_lock);
2075
2076 logical_ring_init_platform_invariants(ring);
2077
2078 ret = i915_cmd_parser_init_ring(ring);
2079 if (ret)
2080 goto error;
2081
2082 ret = intel_lr_context_deferred_alloc(dctx, ring);
2083 if (ret)
2084 goto error;
2085
2086 /* As this is the default context, always pin it */
2087 ret = intel_lr_context_do_pin(dctx, ring);
2088 if (ret) {
2089 DRM_ERROR(
2090 "Failed to pin and map ringbuffer %s: %d\n",
2091 ring->name, ret);
2092 goto error;
2093 }
2094
2095 return 0;
2096
2097error:
2098 intel_logical_ring_cleanup(ring);
2099 return ret;
2100}
2101
2102static int logical_render_ring_init(struct drm_device *dev)
2103{
2104 struct drm_i915_private *dev_priv = dev->dev_private;
2105 struct intel_engine_cs *ring = &dev_priv->ring[RCS];
2106 int ret;
2107
2108 ring->name = "render ring";
2109 ring->id = RCS;
2110 ring->exec_id = I915_EXEC_RENDER;
2111 ring->guc_id = GUC_RENDER_ENGINE;
2112 ring->mmio_base = RENDER_RING_BASE;
2113
2114 logical_ring_default_irqs(ring, GEN8_RCS_IRQ_SHIFT);
2115 if (HAS_L3_DPF(dev))
2116 ring->irq_keep_mask |= GT_RENDER_L3_PARITY_ERROR_INTERRUPT;
2117
2118 logical_ring_default_vfuncs(dev, ring);
2119
2120 /* Override some for render ring. */
2121 if (INTEL_INFO(dev)->gen >= 9)
2122 ring->init_hw = gen9_init_render_ring;
2123 else
2124 ring->init_hw = gen8_init_render_ring;
2125 ring->init_context = gen8_init_rcs_context;
2126 ring->cleanup = intel_fini_pipe_control;
2127 ring->emit_flush = gen8_emit_flush_render;
2128 ring->emit_request = gen8_emit_request_render;
2129
2130 ring->dev = dev;
2131
2132 ret = intel_init_pipe_control(ring);
2133 if (ret)
2134 return ret;
2135
2136 ret = intel_init_workaround_bb(ring);
2137 if (ret) {
2138 /*
2139 * We continue even if we fail to initialize WA batch
2140 * because we only expect rare glitches but nothing
2141 * critical to prevent us from using GPU
2142 */
2143 DRM_ERROR("WA batch buffer initialization failed: %d\n",
2144 ret);
2145 }
2146
2147 ret = logical_ring_init(dev, ring);
2148 if (ret) {
2149 lrc_destroy_wa_ctx_obj(ring);
2150 }
2151
2152 return ret;
2153}
2154
2155static int logical_bsd_ring_init(struct drm_device *dev)
2156{
2157 struct drm_i915_private *dev_priv = dev->dev_private;
2158 struct intel_engine_cs *ring = &dev_priv->ring[VCS];
2159
2160 ring->name = "bsd ring";
2161 ring->id = VCS;
2162 ring->exec_id = I915_EXEC_BSD;
2163 ring->guc_id = GUC_VIDEO_ENGINE;
2164 ring->mmio_base = GEN6_BSD_RING_BASE;
2165
2166 logical_ring_default_irqs(ring, GEN8_VCS1_IRQ_SHIFT);
2167 logical_ring_default_vfuncs(dev, ring);
2168
2169 return logical_ring_init(dev, ring);
2170}
2171
2172static int logical_bsd2_ring_init(struct drm_device *dev)
2173{
2174 struct drm_i915_private *dev_priv = dev->dev_private;
2175 struct intel_engine_cs *ring = &dev_priv->ring[VCS2];
2176
2177 ring->name = "bsd2 ring";
2178 ring->id = VCS2;
2179 ring->exec_id = I915_EXEC_BSD;
2180 ring->guc_id = GUC_VIDEO_ENGINE2;
2181 ring->mmio_base = GEN8_BSD2_RING_BASE;
2182
2183 logical_ring_default_irqs(ring, GEN8_VCS2_IRQ_SHIFT);
2184 logical_ring_default_vfuncs(dev, ring);
2185
2186 return logical_ring_init(dev, ring);
2187}
2188
2189static int logical_blt_ring_init(struct drm_device *dev)
2190{
2191 struct drm_i915_private *dev_priv = dev->dev_private;
2192 struct intel_engine_cs *ring = &dev_priv->ring[BCS];
2193
2194 ring->name = "blitter ring";
2195 ring->id = BCS;
2196 ring->exec_id = I915_EXEC_BLT;
2197 ring->guc_id = GUC_BLITTER_ENGINE;
2198 ring->mmio_base = BLT_RING_BASE;
2199
2200 logical_ring_default_irqs(ring, GEN8_BCS_IRQ_SHIFT);
2201 logical_ring_default_vfuncs(dev, ring);
2202
2203 return logical_ring_init(dev, ring);
2204}
2205
2206static int logical_vebox_ring_init(struct drm_device *dev)
2207{
2208 struct drm_i915_private *dev_priv = dev->dev_private;
2209 struct intel_engine_cs *ring = &dev_priv->ring[VECS];
2210
2211 ring->name = "video enhancement ring";
2212 ring->id = VECS;
2213 ring->exec_id = I915_EXEC_VEBOX;
2214 ring->guc_id = GUC_VIDEOENHANCE_ENGINE;
2215 ring->mmio_base = VEBOX_RING_BASE;
2216
2217 logical_ring_default_irqs(ring, GEN8_VECS_IRQ_SHIFT);
2218 logical_ring_default_vfuncs(dev, ring);
2219
2220 return logical_ring_init(dev, ring);
2221}
2222
2223/**
2224 * intel_logical_rings_init() - allocate, populate and init the Engine Command Streamers
2225 * @dev: DRM device.
2226 *
2227 * This function inits the engines for an Execlists submission style (the equivalent in the
2228 * legacy ringbuffer submission world would be i915_gem_init_rings). It does it only for
2229 * those engines that are present in the hardware.
2230 *
2231 * Return: non-zero if the initialization failed.
2232 */
2233int intel_logical_rings_init(struct drm_device *dev)
2234{
2235 struct drm_i915_private *dev_priv = dev->dev_private;
2236 int ret;
2237
2238 ret = logical_render_ring_init(dev);
2239 if (ret)
2240 return ret;
2241
2242 if (HAS_BSD(dev)) {
2243 ret = logical_bsd_ring_init(dev);
2244 if (ret)
2245 goto cleanup_render_ring;
2246 }
2247
2248 if (HAS_BLT(dev)) {
2249 ret = logical_blt_ring_init(dev);
2250 if (ret)
2251 goto cleanup_bsd_ring;
2252 }
2253
2254 if (HAS_VEBOX(dev)) {
2255 ret = logical_vebox_ring_init(dev);
2256 if (ret)
2257 goto cleanup_blt_ring;
2258 }
2259
2260 if (HAS_BSD2(dev)) {
2261 ret = logical_bsd2_ring_init(dev);
2262 if (ret)
2263 goto cleanup_vebox_ring;
2264 }
2265
2266 return 0;
2267
2268cleanup_vebox_ring:
2269 intel_logical_ring_cleanup(&dev_priv->ring[VECS]);
2270cleanup_blt_ring:
2271 intel_logical_ring_cleanup(&dev_priv->ring[BCS]);
2272cleanup_bsd_ring:
2273 intel_logical_ring_cleanup(&dev_priv->ring[VCS]);
2274cleanup_render_ring:
2275 intel_logical_ring_cleanup(&dev_priv->ring[RCS]);
2276
2277 return ret;
2278}
2279
2280static u32
2281make_rpcs(struct drm_device *dev)
2282{
2283 u32 rpcs = 0;
2284
2285 /*
2286 * No explicit RPCS request is needed to ensure full
2287 * slice/subslice/EU enablement prior to Gen9.
2288 */
2289 if (INTEL_INFO(dev)->gen < 9)
2290 return 0;
2291
2292 /*
2293 * Starting in Gen9, render power gating can leave
2294 * slice/subslice/EU in a partially enabled state. We
2295 * must make an explicit request through RPCS for full
2296 * enablement.
2297 */
2298 if (INTEL_INFO(dev)->has_slice_pg) {
2299 rpcs |= GEN8_RPCS_S_CNT_ENABLE;
2300 rpcs |= INTEL_INFO(dev)->slice_total <<
2301 GEN8_RPCS_S_CNT_SHIFT;
2302 rpcs |= GEN8_RPCS_ENABLE;
2303 }
2304
2305 if (INTEL_INFO(dev)->has_subslice_pg) {
2306 rpcs |= GEN8_RPCS_SS_CNT_ENABLE;
2307 rpcs |= INTEL_INFO(dev)->subslice_per_slice <<
2308 GEN8_RPCS_SS_CNT_SHIFT;
2309 rpcs |= GEN8_RPCS_ENABLE;
2310 }
2311
2312 if (INTEL_INFO(dev)->has_eu_pg) {
2313 rpcs |= INTEL_INFO(dev)->eu_per_subslice <<
2314 GEN8_RPCS_EU_MIN_SHIFT;
2315 rpcs |= INTEL_INFO(dev)->eu_per_subslice <<
2316 GEN8_RPCS_EU_MAX_SHIFT;
2317 rpcs |= GEN8_RPCS_ENABLE;
2318 }
2319
2320 return rpcs;
2321}
2322
2323static u32 intel_lr_indirect_ctx_offset(struct intel_engine_cs *ring)
2324{
2325 u32 indirect_ctx_offset;
2326
2327 switch (INTEL_INFO(ring->dev)->gen) {
2328 default:
2329 MISSING_CASE(INTEL_INFO(ring->dev)->gen);
2330 /* fall through */
2331 case 9:
2332 indirect_ctx_offset =
2333 GEN9_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT;
2334 break;
2335 case 8:
2336 indirect_ctx_offset =
2337 GEN8_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT;
2338 break;
2339 }
2340
2341 return indirect_ctx_offset;
2342}
2343
2344static int
2345populate_lr_context(struct intel_context *ctx, struct drm_i915_gem_object *ctx_obj,
2346 struct intel_engine_cs *ring, struct intel_ringbuffer *ringbuf)
2347{
2348 struct drm_device *dev = ring->dev;
2349 struct drm_i915_private *dev_priv = dev->dev_private;
2350 struct i915_hw_ppgtt *ppgtt = ctx->ppgtt;
2351 struct page *page;
2352 uint32_t *reg_state;
2353 int ret;
2354
2355 if (!ppgtt)
2356 ppgtt = dev_priv->mm.aliasing_ppgtt;
2357
2358 ret = i915_gem_object_set_to_cpu_domain(ctx_obj, true);
2359 if (ret) {
2360 DRM_DEBUG_DRIVER("Could not set to CPU domain\n");
2361 return ret;
2362 }
2363
2364 ret = i915_gem_object_get_pages(ctx_obj);
2365 if (ret) {
2366 DRM_DEBUG_DRIVER("Could not get object pages\n");
2367 return ret;
2368 }
2369
2370 i915_gem_object_pin_pages(ctx_obj);
2371
2372 /* The second page of the context object contains some fields which must
2373 * be set up prior to the first execution. */
2374 page = i915_gem_object_get_dirty_page(ctx_obj, LRC_STATE_PN);
2375 reg_state = kmap_atomic(page);
2376
2377 /* A context is actually a big batch buffer with several MI_LOAD_REGISTER_IMM
2378 * commands followed by (reg, value) pairs. The values we are setting here are
2379 * only for the first context restore: on a subsequent save, the GPU will
2380 * recreate this batchbuffer with new values (including all the missing
2381 * MI_LOAD_REGISTER_IMM commands that we are not initializing here). */
2382 reg_state[CTX_LRI_HEADER_0] =
2383 MI_LOAD_REGISTER_IMM(ring->id == RCS ? 14 : 11) | MI_LRI_FORCE_POSTED;
2384 ASSIGN_CTX_REG(reg_state, CTX_CONTEXT_CONTROL, RING_CONTEXT_CONTROL(ring),
2385 _MASKED_BIT_ENABLE(CTX_CTRL_INHIBIT_SYN_CTX_SWITCH |
2386 CTX_CTRL_ENGINE_CTX_RESTORE_INHIBIT |
2387 (HAS_RESOURCE_STREAMER(dev) ?
2388 CTX_CTRL_RS_CTX_ENABLE : 0)));
2389 ASSIGN_CTX_REG(reg_state, CTX_RING_HEAD, RING_HEAD(ring->mmio_base), 0);
2390 ASSIGN_CTX_REG(reg_state, CTX_RING_TAIL, RING_TAIL(ring->mmio_base), 0);
2391 /* Ring buffer start address is not known until the buffer is pinned.
2392 * It is written to the context image in execlists_update_context()
2393 */
2394 ASSIGN_CTX_REG(reg_state, CTX_RING_BUFFER_START, RING_START(ring->mmio_base), 0);
2395 ASSIGN_CTX_REG(reg_state, CTX_RING_BUFFER_CONTROL, RING_CTL(ring->mmio_base),
2396 ((ringbuf->size - PAGE_SIZE) & RING_NR_PAGES) | RING_VALID);
2397 ASSIGN_CTX_REG(reg_state, CTX_BB_HEAD_U, RING_BBADDR_UDW(ring->mmio_base), 0);
2398 ASSIGN_CTX_REG(reg_state, CTX_BB_HEAD_L, RING_BBADDR(ring->mmio_base), 0);
2399 ASSIGN_CTX_REG(reg_state, CTX_BB_STATE, RING_BBSTATE(ring->mmio_base),
2400 RING_BB_PPGTT);
2401 ASSIGN_CTX_REG(reg_state, CTX_SECOND_BB_HEAD_U, RING_SBBADDR_UDW(ring->mmio_base), 0);
2402 ASSIGN_CTX_REG(reg_state, CTX_SECOND_BB_HEAD_L, RING_SBBADDR(ring->mmio_base), 0);
2403 ASSIGN_CTX_REG(reg_state, CTX_SECOND_BB_STATE, RING_SBBSTATE(ring->mmio_base), 0);
2404 if (ring->id == RCS) {
2405 ASSIGN_CTX_REG(reg_state, CTX_BB_PER_CTX_PTR, RING_BB_PER_CTX_PTR(ring->mmio_base), 0);
2406 ASSIGN_CTX_REG(reg_state, CTX_RCS_INDIRECT_CTX, RING_INDIRECT_CTX(ring->mmio_base), 0);
2407 ASSIGN_CTX_REG(reg_state, CTX_RCS_INDIRECT_CTX_OFFSET, RING_INDIRECT_CTX_OFFSET(ring->mmio_base), 0);
2408 if (ring->wa_ctx.obj) {
2409 struct i915_ctx_workarounds *wa_ctx = &ring->wa_ctx;
2410 uint32_t ggtt_offset = i915_gem_obj_ggtt_offset(wa_ctx->obj);
2411
2412 reg_state[CTX_RCS_INDIRECT_CTX+1] =
2413 (ggtt_offset + wa_ctx->indirect_ctx.offset * sizeof(uint32_t)) |
2414 (wa_ctx->indirect_ctx.size / CACHELINE_DWORDS);
2415
2416 reg_state[CTX_RCS_INDIRECT_CTX_OFFSET+1] =
2417 intel_lr_indirect_ctx_offset(ring) << 6;
2418
2419 reg_state[CTX_BB_PER_CTX_PTR+1] =
2420 (ggtt_offset + wa_ctx->per_ctx.offset * sizeof(uint32_t)) |
2421 0x01;
2422 }
2423 }
2424 reg_state[CTX_LRI_HEADER_1] = MI_LOAD_REGISTER_IMM(9) | MI_LRI_FORCE_POSTED;
2425 ASSIGN_CTX_REG(reg_state, CTX_CTX_TIMESTAMP, RING_CTX_TIMESTAMP(ring->mmio_base), 0);
2426 /* PDP values well be assigned later if needed */
2427 ASSIGN_CTX_REG(reg_state, CTX_PDP3_UDW, GEN8_RING_PDP_UDW(ring, 3), 0);
2428 ASSIGN_CTX_REG(reg_state, CTX_PDP3_LDW, GEN8_RING_PDP_LDW(ring, 3), 0);
2429 ASSIGN_CTX_REG(reg_state, CTX_PDP2_UDW, GEN8_RING_PDP_UDW(ring, 2), 0);
2430 ASSIGN_CTX_REG(reg_state, CTX_PDP2_LDW, GEN8_RING_PDP_LDW(ring, 2), 0);
2431 ASSIGN_CTX_REG(reg_state, CTX_PDP1_UDW, GEN8_RING_PDP_UDW(ring, 1), 0);
2432 ASSIGN_CTX_REG(reg_state, CTX_PDP1_LDW, GEN8_RING_PDP_LDW(ring, 1), 0);
2433 ASSIGN_CTX_REG(reg_state, CTX_PDP0_UDW, GEN8_RING_PDP_UDW(ring, 0), 0);
2434 ASSIGN_CTX_REG(reg_state, CTX_PDP0_LDW, GEN8_RING_PDP_LDW(ring, 0), 0);
2435
2436 if (USES_FULL_48BIT_PPGTT(ppgtt->base.dev)) {
2437 /* 64b PPGTT (48bit canonical)
2438 * PDP0_DESCRIPTOR contains the base address to PML4 and
2439 * other PDP Descriptors are ignored.
2440 */
2441 ASSIGN_CTX_PML4(ppgtt, reg_state);
2442 } else {
2443 /* 32b PPGTT
2444 * PDP*_DESCRIPTOR contains the base address of space supported.
2445 * With dynamic page allocation, PDPs may not be allocated at
2446 * this point. Point the unallocated PDPs to the scratch page
2447 */
2448 ASSIGN_CTX_PDP(ppgtt, reg_state, 3);
2449 ASSIGN_CTX_PDP(ppgtt, reg_state, 2);
2450 ASSIGN_CTX_PDP(ppgtt, reg_state, 1);
2451 ASSIGN_CTX_PDP(ppgtt, reg_state, 0);
2452 }
2453
2454 if (ring->id == RCS) {
2455 reg_state[CTX_LRI_HEADER_2] = MI_LOAD_REGISTER_IMM(1);
2456 ASSIGN_CTX_REG(reg_state, CTX_R_PWR_CLK_STATE, GEN8_R_PWR_CLK_STATE,
2457 make_rpcs(dev));
2458 }
2459
2460 kunmap_atomic(reg_state);
2461 i915_gem_object_unpin_pages(ctx_obj);
2462
2463 return 0;
2464}
2465
2466/**
2467 * intel_lr_context_free() - free the LRC specific bits of a context
2468 * @ctx: the LR context to free.
2469 *
2470 * The real context freeing is done in i915_gem_context_free: this only
2471 * takes care of the bits that are LRC related: the per-engine backing
2472 * objects and the logical ringbuffer.
2473 */
2474void intel_lr_context_free(struct intel_context *ctx)
2475{
2476 int i;
2477
2478 for (i = I915_NUM_RINGS; --i >= 0; ) {
2479 struct intel_ringbuffer *ringbuf = ctx->engine[i].ringbuf;
2480 struct drm_i915_gem_object *ctx_obj = ctx->engine[i].state;
2481
2482 if (!ctx_obj)
2483 continue;
2484
2485 if (ctx == ctx->i915->kernel_context) {
2486 intel_unpin_ringbuffer_obj(ringbuf);
2487 i915_gem_object_ggtt_unpin(ctx_obj);
2488 }
2489
2490 WARN_ON(ctx->engine[i].pin_count);
2491 intel_ringbuffer_free(ringbuf);
2492 drm_gem_object_unreference(&ctx_obj->base);
2493 }
2494}
2495
2496/**
2497 * intel_lr_context_size() - return the size of the context for an engine
2498 * @ring: which engine to find the context size for
2499 *
2500 * Each engine may require a different amount of space for a context image,
2501 * so when allocating (or copying) an image, this function can be used to
2502 * find the right size for the specific engine.
2503 *
2504 * Return: size (in bytes) of an engine-specific context image
2505 *
2506 * Note: this size includes the HWSP, which is part of the context image
2507 * in LRC mode, but does not include the "shared data page" used with
2508 * GuC submission. The caller should account for this if using the GuC.
2509 */
2510uint32_t intel_lr_context_size(struct intel_engine_cs *ring)
2511{
2512 int ret = 0;
2513
2514 WARN_ON(INTEL_INFO(ring->dev)->gen < 8);
2515
2516 switch (ring->id) {
2517 case RCS:
2518 if (INTEL_INFO(ring->dev)->gen >= 9)
2519 ret = GEN9_LR_CONTEXT_RENDER_SIZE;
2520 else
2521 ret = GEN8_LR_CONTEXT_RENDER_SIZE;
2522 break;
2523 case VCS:
2524 case BCS:
2525 case VECS:
2526 case VCS2:
2527 ret = GEN8_LR_CONTEXT_OTHER_SIZE;
2528 break;
2529 }
2530
2531 return ret;
2532}
2533
2534static void lrc_setup_hardware_status_page(struct intel_engine_cs *ring,
2535 struct drm_i915_gem_object *default_ctx_obj)
2536{
2537 struct drm_i915_private *dev_priv = ring->dev->dev_private;
2538 struct page *page;
2539
2540 /* The HWSP is part of the default context object in LRC mode. */
2541 ring->status_page.gfx_addr = i915_gem_obj_ggtt_offset(default_ctx_obj)
2542 + LRC_PPHWSP_PN * PAGE_SIZE;
2543 page = i915_gem_object_get_page(default_ctx_obj, LRC_PPHWSP_PN);
2544 ring->status_page.page_addr = kmap(page);
2545 ring->status_page.obj = default_ctx_obj;
2546
2547 I915_WRITE(RING_HWS_PGA(ring->mmio_base),
2548 (u32)ring->status_page.gfx_addr);
2549 POSTING_READ(RING_HWS_PGA(ring->mmio_base));
2550}
2551
2552/**
2553 * intel_lr_context_deferred_alloc() - create the LRC specific bits of a context
2554 * @ctx: LR context to create.
2555 * @ring: engine to be used with the context.
2556 *
2557 * This function can be called more than once, with different engines, if we plan
2558 * to use the context with them. The context backing objects and the ringbuffers
2559 * (specially the ringbuffer backing objects) suck a lot of memory up, and that's why
2560 * the creation is a deferred call: it's better to make sure first that we need to use
2561 * a given ring with the context.
2562 *
2563 * Return: non-zero on error.
2564 */
2565
2566int intel_lr_context_deferred_alloc(struct intel_context *ctx,
2567 struct intel_engine_cs *ring)
2568{
2569 struct drm_device *dev = ring->dev;
2570 struct drm_i915_gem_object *ctx_obj;
2571 uint32_t context_size;
2572 struct intel_ringbuffer *ringbuf;
2573 int ret;
2574
2575 WARN_ON(ctx->legacy_hw_ctx.rcs_state != NULL);
2576 WARN_ON(ctx->engine[ring->id].state);
2577
2578 context_size = round_up(intel_lr_context_size(ring), 4096);
2579
2580 /* One extra page as the sharing data between driver and GuC */
2581 context_size += PAGE_SIZE * LRC_PPHWSP_PN;
2582
2583 ctx_obj = i915_gem_alloc_object(dev, context_size);
2584 if (!ctx_obj) {
2585 DRM_DEBUG_DRIVER("Alloc LRC backing obj failed.\n");
2586 return -ENOMEM;
2587 }
2588
2589 ringbuf = intel_engine_create_ringbuffer(ring, 4 * PAGE_SIZE);
2590 if (IS_ERR(ringbuf)) {
2591 ret = PTR_ERR(ringbuf);
2592 goto error_deref_obj;
2593 }
2594
2595 ret = populate_lr_context(ctx, ctx_obj, ring, ringbuf);
2596 if (ret) {
2597 DRM_DEBUG_DRIVER("Failed to populate LRC: %d\n", ret);
2598 goto error_ringbuf;
2599 }
2600
2601 ctx->engine[ring->id].ringbuf = ringbuf;
2602 ctx->engine[ring->id].state = ctx_obj;
2603
2604 if (ctx != ctx->i915->kernel_context && ring->init_context) {
2605 struct drm_i915_gem_request *req;
2606
2607 req = i915_gem_request_alloc(ring, ctx);
2608 if (IS_ERR(req)) {
2609 ret = PTR_ERR(req);
2610 DRM_ERROR("ring create req: %d\n", ret);
2611 goto error_ringbuf;
2612 }
2613
2614 ret = ring->init_context(req);
2615 if (ret) {
2616 DRM_ERROR("ring init context: %d\n",
2617 ret);
2618 i915_gem_request_cancel(req);
2619 goto error_ringbuf;
2620 }
2621 i915_add_request_no_flush(req);
2622 }
2623 return 0;
2624
2625error_ringbuf:
2626 intel_ringbuffer_free(ringbuf);
2627error_deref_obj:
2628 drm_gem_object_unreference(&ctx_obj->base);
2629 ctx->engine[ring->id].ringbuf = NULL;
2630 ctx->engine[ring->id].state = NULL;
2631 return ret;
2632}
2633
2634void intel_lr_context_reset(struct drm_device *dev,
2635 struct intel_context *ctx)
2636{
2637 struct drm_i915_private *dev_priv = dev->dev_private;
2638 struct intel_engine_cs *ring;
2639 int i;
2640
2641 for_each_ring(ring, dev_priv, i) {
2642 struct drm_i915_gem_object *ctx_obj =
2643 ctx->engine[ring->id].state;
2644 struct intel_ringbuffer *ringbuf =
2645 ctx->engine[ring->id].ringbuf;
2646 uint32_t *reg_state;
2647 struct page *page;
2648
2649 if (!ctx_obj)
2650 continue;
2651
2652 if (i915_gem_object_get_pages(ctx_obj)) {
2653 WARN(1, "Failed get_pages for context obj\n");
2654 continue;
2655 }
2656 page = i915_gem_object_get_dirty_page(ctx_obj, LRC_STATE_PN);
2657 reg_state = kmap_atomic(page);
2658
2659 reg_state[CTX_RING_HEAD+1] = 0;
2660 reg_state[CTX_RING_TAIL+1] = 0;
2661
2662 kunmap_atomic(reg_state);
2663
2664 ringbuf->head = 0;
2665 ringbuf->tail = 0;
2666 }
2667}
1/*
2 * Copyright © 2014 Intel Corporation
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 *
23 * Authors:
24 * Ben Widawsky <ben@bwidawsk.net>
25 * Michel Thierry <michel.thierry@intel.com>
26 * Thomas Daniel <thomas.daniel@intel.com>
27 * Oscar Mateo <oscar.mateo@intel.com>
28 *
29 */
30
31/**
32 * DOC: Logical Rings, Logical Ring Contexts and Execlists
33 *
34 * Motivation:
35 * GEN8 brings an expansion of the HW contexts: "Logical Ring Contexts".
36 * These expanded contexts enable a number of new abilities, especially
37 * "Execlists" (also implemented in this file).
38 *
39 * One of the main differences with the legacy HW contexts is that logical
40 * ring contexts incorporate many more things to the context's state, like
41 * PDPs or ringbuffer control registers:
42 *
43 * The reason why PDPs are included in the context is straightforward: as
44 * PPGTTs (per-process GTTs) are actually per-context, having the PDPs
45 * contained there mean you don't need to do a ppgtt->switch_mm yourself,
46 * instead, the GPU will do it for you on the context switch.
47 *
48 * But, what about the ringbuffer control registers (head, tail, etc..)?
49 * shouldn't we just need a set of those per engine command streamer? This is
50 * where the name "Logical Rings" starts to make sense: by virtualizing the
51 * rings, the engine cs shifts to a new "ring buffer" with every context
52 * switch. When you want to submit a workload to the GPU you: A) choose your
53 * context, B) find its appropriate virtualized ring, C) write commands to it
54 * and then, finally, D) tell the GPU to switch to that context.
55 *
56 * Instead of the legacy MI_SET_CONTEXT, the way you tell the GPU to switch
57 * to a contexts is via a context execution list, ergo "Execlists".
58 *
59 * LRC implementation:
60 * Regarding the creation of contexts, we have:
61 *
62 * - One global default context.
63 * - One local default context for each opened fd.
64 * - One local extra context for each context create ioctl call.
65 *
66 * Now that ringbuffers belong per-context (and not per-engine, like before)
67 * and that contexts are uniquely tied to a given engine (and not reusable,
68 * like before) we need:
69 *
70 * - One ringbuffer per-engine inside each context.
71 * - One backing object per-engine inside each context.
72 *
73 * The global default context starts its life with these new objects fully
74 * allocated and populated. The local default context for each opened fd is
75 * more complex, because we don't know at creation time which engine is going
76 * to use them. To handle this, we have implemented a deferred creation of LR
77 * contexts:
78 *
79 * The local context starts its life as a hollow or blank holder, that only
80 * gets populated for a given engine once we receive an execbuffer. If later
81 * on we receive another execbuffer ioctl for the same context but a different
82 * engine, we allocate/populate a new ringbuffer and context backing object and
83 * so on.
84 *
85 * Finally, regarding local contexts created using the ioctl call: as they are
86 * only allowed with the render ring, we can allocate & populate them right
87 * away (no need to defer anything, at least for now).
88 *
89 * Execlists implementation:
90 * Execlists are the new method by which, on gen8+ hardware, workloads are
91 * submitted for execution (as opposed to the legacy, ringbuffer-based, method).
92 * This method works as follows:
93 *
94 * When a request is committed, its commands (the BB start and any leading or
95 * trailing commands, like the seqno breadcrumbs) are placed in the ringbuffer
96 * for the appropriate context. The tail pointer in the hardware context is not
97 * updated at this time, but instead, kept by the driver in the ringbuffer
98 * structure. A structure representing this request is added to a request queue
99 * for the appropriate engine: this structure contains a copy of the context's
100 * tail after the request was written to the ring buffer and a pointer to the
101 * context itself.
102 *
103 * If the engine's request queue was empty before the request was added, the
104 * queue is processed immediately. Otherwise the queue will be processed during
105 * a context switch interrupt. In any case, elements on the queue will get sent
106 * (in pairs) to the GPU's ExecLists Submit Port (ELSP, for short) with a
107 * globally unique 20-bits submission ID.
108 *
109 * When execution of a request completes, the GPU updates the context status
110 * buffer with a context complete event and generates a context switch interrupt.
111 * During the interrupt handling, the driver examines the events in the buffer:
112 * for each context complete event, if the announced ID matches that on the head
113 * of the request queue, then that request is retired and removed from the queue.
114 *
115 * After processing, if any requests were retired and the queue is not empty
116 * then a new execution list can be submitted. The two requests at the front of
117 * the queue are next to be submitted but since a context may not occur twice in
118 * an execution list, if subsequent requests have the same ID as the first then
119 * the two requests must be combined. This is done simply by discarding requests
120 * at the head of the queue until either only one requests is left (in which case
121 * we use a NULL second context) or the first two requests have unique IDs.
122 *
123 * By always executing the first two requests in the queue the driver ensures
124 * that the GPU is kept as busy as possible. In the case where a single context
125 * completes but a second context is still executing, the request for this second
126 * context will be at the head of the queue when we remove the first one. This
127 * request will then be resubmitted along with a new request for a different context,
128 * which will cause the hardware to continue executing the second request and queue
129 * the new request (the GPU detects the condition of a context getting preempted
130 * with the same context and optimizes the context switch flow by not doing
131 * preemption, but just sampling the new tail pointer).
132 *
133 */
134#include <linux/interrupt.h>
135
136#include <drm/drmP.h>
137#include <drm/i915_drm.h>
138#include "i915_drv.h"
139#include "i915_gem_render_state.h"
140#include "intel_lrc_reg.h"
141#include "intel_mocs.h"
142
143#define RING_EXECLIST_QFULL (1 << 0x2)
144#define RING_EXECLIST1_VALID (1 << 0x3)
145#define RING_EXECLIST0_VALID (1 << 0x4)
146#define RING_EXECLIST_ACTIVE_STATUS (3 << 0xE)
147#define RING_EXECLIST1_ACTIVE (1 << 0x11)
148#define RING_EXECLIST0_ACTIVE (1 << 0x12)
149
150#define GEN8_CTX_STATUS_IDLE_ACTIVE (1 << 0)
151#define GEN8_CTX_STATUS_PREEMPTED (1 << 1)
152#define GEN8_CTX_STATUS_ELEMENT_SWITCH (1 << 2)
153#define GEN8_CTX_STATUS_ACTIVE_IDLE (1 << 3)
154#define GEN8_CTX_STATUS_COMPLETE (1 << 4)
155#define GEN8_CTX_STATUS_LITE_RESTORE (1 << 15)
156
157#define GEN8_CTX_STATUS_COMPLETED_MASK \
158 (GEN8_CTX_STATUS_COMPLETE | GEN8_CTX_STATUS_PREEMPTED)
159
160/* Typical size of the average request (2 pipecontrols and a MI_BB) */
161#define EXECLISTS_REQUEST_SIZE 64 /* bytes */
162#define WA_TAIL_DWORDS 2
163#define WA_TAIL_BYTES (sizeof(u32) * WA_TAIL_DWORDS)
164
165static int execlists_context_deferred_alloc(struct i915_gem_context *ctx,
166 struct intel_engine_cs *engine);
167static void execlists_init_reg_state(u32 *reg_state,
168 struct i915_gem_context *ctx,
169 struct intel_engine_cs *engine,
170 struct intel_ring *ring);
171
172static inline struct i915_priolist *to_priolist(struct rb_node *rb)
173{
174 return rb_entry(rb, struct i915_priolist, node);
175}
176
177static inline int rq_prio(const struct i915_request *rq)
178{
179 return rq->priotree.priority;
180}
181
182static inline bool need_preempt(const struct intel_engine_cs *engine,
183 const struct i915_request *last,
184 int prio)
185{
186 return engine->i915->preempt_context && prio > max(rq_prio(last), 0);
187}
188
189/**
190 * intel_lr_context_descriptor_update() - calculate & cache the descriptor
191 * descriptor for a pinned context
192 * @ctx: Context to work on
193 * @engine: Engine the descriptor will be used with
194 *
195 * The context descriptor encodes various attributes of a context,
196 * including its GTT address and some flags. Because it's fairly
197 * expensive to calculate, we'll just do it once and cache the result,
198 * which remains valid until the context is unpinned.
199 *
200 * This is what a descriptor looks like, from LSB to MSB::
201 *
202 * bits 0-11: flags, GEN8_CTX_* (cached in ctx->desc_template)
203 * bits 12-31: LRCA, GTT address of (the HWSP of) this context
204 * bits 32-52: ctx ID, a globally unique tag
205 * bits 53-54: mbz, reserved for use by hardware
206 * bits 55-63: group ID, currently unused and set to 0
207 *
208 * Starting from Gen11, the upper dword of the descriptor has a new format:
209 *
210 * bits 32-36: reserved
211 * bits 37-47: SW context ID
212 * bits 48:53: engine instance
213 * bit 54: mbz, reserved for use by hardware
214 * bits 55-60: SW counter
215 * bits 61-63: engine class
216 *
217 * engine info, SW context ID and SW counter need to form a unique number
218 * (Context ID) per lrc.
219 */
220static void
221intel_lr_context_descriptor_update(struct i915_gem_context *ctx,
222 struct intel_engine_cs *engine)
223{
224 struct intel_context *ce = &ctx->engine[engine->id];
225 u64 desc;
226
227 BUILD_BUG_ON(MAX_CONTEXT_HW_ID > (BIT(GEN8_CTX_ID_WIDTH)));
228 BUILD_BUG_ON(GEN11_MAX_CONTEXT_HW_ID > (BIT(GEN11_SW_CTX_ID_WIDTH)));
229
230 desc = ctx->desc_template; /* bits 0-11 */
231 GEM_BUG_ON(desc & GENMASK_ULL(63, 12));
232
233 desc |= i915_ggtt_offset(ce->state) + LRC_HEADER_PAGES * PAGE_SIZE;
234 /* bits 12-31 */
235 GEM_BUG_ON(desc & GENMASK_ULL(63, 32));
236
237 if (INTEL_GEN(ctx->i915) >= 11) {
238 GEM_BUG_ON(ctx->hw_id >= BIT(GEN11_SW_CTX_ID_WIDTH));
239 desc |= (u64)ctx->hw_id << GEN11_SW_CTX_ID_SHIFT;
240 /* bits 37-47 */
241
242 desc |= (u64)engine->instance << GEN11_ENGINE_INSTANCE_SHIFT;
243 /* bits 48-53 */
244
245 /* TODO: decide what to do with SW counter (bits 55-60) */
246
247 desc |= (u64)engine->class << GEN11_ENGINE_CLASS_SHIFT;
248 /* bits 61-63 */
249 } else {
250 GEM_BUG_ON(ctx->hw_id >= BIT(GEN8_CTX_ID_WIDTH));
251 desc |= (u64)ctx->hw_id << GEN8_CTX_ID_SHIFT; /* bits 32-52 */
252 }
253
254 ce->lrc_desc = desc;
255}
256
257static struct i915_priolist *
258lookup_priolist(struct intel_engine_cs *engine,
259 struct i915_priotree *pt,
260 int prio)
261{
262 struct intel_engine_execlists * const execlists = &engine->execlists;
263 struct i915_priolist *p;
264 struct rb_node **parent, *rb;
265 bool first = true;
266
267 if (unlikely(execlists->no_priolist))
268 prio = I915_PRIORITY_NORMAL;
269
270find_priolist:
271 /* most positive priority is scheduled first, equal priorities fifo */
272 rb = NULL;
273 parent = &execlists->queue.rb_node;
274 while (*parent) {
275 rb = *parent;
276 p = to_priolist(rb);
277 if (prio > p->priority) {
278 parent = &rb->rb_left;
279 } else if (prio < p->priority) {
280 parent = &rb->rb_right;
281 first = false;
282 } else {
283 return p;
284 }
285 }
286
287 if (prio == I915_PRIORITY_NORMAL) {
288 p = &execlists->default_priolist;
289 } else {
290 p = kmem_cache_alloc(engine->i915->priorities, GFP_ATOMIC);
291 /* Convert an allocation failure to a priority bump */
292 if (unlikely(!p)) {
293 prio = I915_PRIORITY_NORMAL; /* recurses just once */
294
295 /* To maintain ordering with all rendering, after an
296 * allocation failure we have to disable all scheduling.
297 * Requests will then be executed in fifo, and schedule
298 * will ensure that dependencies are emitted in fifo.
299 * There will be still some reordering with existing
300 * requests, so if userspace lied about their
301 * dependencies that reordering may be visible.
302 */
303 execlists->no_priolist = true;
304 goto find_priolist;
305 }
306 }
307
308 p->priority = prio;
309 INIT_LIST_HEAD(&p->requests);
310 rb_link_node(&p->node, rb, parent);
311 rb_insert_color(&p->node, &execlists->queue);
312
313 if (first)
314 execlists->first = &p->node;
315
316 return p;
317}
318
319static void unwind_wa_tail(struct i915_request *rq)
320{
321 rq->tail = intel_ring_wrap(rq->ring, rq->wa_tail - WA_TAIL_BYTES);
322 assert_ring_tail_valid(rq->ring, rq->tail);
323}
324
325static void __unwind_incomplete_requests(struct intel_engine_cs *engine)
326{
327 struct i915_request *rq, *rn;
328 struct i915_priolist *uninitialized_var(p);
329 int last_prio = I915_PRIORITY_INVALID;
330
331 lockdep_assert_held(&engine->timeline->lock);
332
333 list_for_each_entry_safe_reverse(rq, rn,
334 &engine->timeline->requests,
335 link) {
336 if (i915_request_completed(rq))
337 return;
338
339 __i915_request_unsubmit(rq);
340 unwind_wa_tail(rq);
341
342 GEM_BUG_ON(rq_prio(rq) == I915_PRIORITY_INVALID);
343 if (rq_prio(rq) != last_prio) {
344 last_prio = rq_prio(rq);
345 p = lookup_priolist(engine, &rq->priotree, last_prio);
346 }
347
348 list_add(&rq->priotree.link, &p->requests);
349 }
350}
351
352void
353execlists_unwind_incomplete_requests(struct intel_engine_execlists *execlists)
354{
355 struct intel_engine_cs *engine =
356 container_of(execlists, typeof(*engine), execlists);
357
358 spin_lock_irq(&engine->timeline->lock);
359 __unwind_incomplete_requests(engine);
360 spin_unlock_irq(&engine->timeline->lock);
361}
362
363static inline void
364execlists_context_status_change(struct i915_request *rq, unsigned long status)
365{
366 /*
367 * Only used when GVT-g is enabled now. When GVT-g is disabled,
368 * The compiler should eliminate this function as dead-code.
369 */
370 if (!IS_ENABLED(CONFIG_DRM_I915_GVT))
371 return;
372
373 atomic_notifier_call_chain(&rq->engine->context_status_notifier,
374 status, rq);
375}
376
377static inline void
378execlists_context_schedule_in(struct i915_request *rq)
379{
380 execlists_context_status_change(rq, INTEL_CONTEXT_SCHEDULE_IN);
381 intel_engine_context_in(rq->engine);
382}
383
384static inline void
385execlists_context_schedule_out(struct i915_request *rq)
386{
387 intel_engine_context_out(rq->engine);
388 execlists_context_status_change(rq, INTEL_CONTEXT_SCHEDULE_OUT);
389}
390
391static void
392execlists_update_context_pdps(struct i915_hw_ppgtt *ppgtt, u32 *reg_state)
393{
394 ASSIGN_CTX_PDP(ppgtt, reg_state, 3);
395 ASSIGN_CTX_PDP(ppgtt, reg_state, 2);
396 ASSIGN_CTX_PDP(ppgtt, reg_state, 1);
397 ASSIGN_CTX_PDP(ppgtt, reg_state, 0);
398}
399
400static u64 execlists_update_context(struct i915_request *rq)
401{
402 struct intel_context *ce = &rq->ctx->engine[rq->engine->id];
403 struct i915_hw_ppgtt *ppgtt =
404 rq->ctx->ppgtt ?: rq->i915->mm.aliasing_ppgtt;
405 u32 *reg_state = ce->lrc_reg_state;
406
407 reg_state[CTX_RING_TAIL+1] = intel_ring_set_tail(rq->ring, rq->tail);
408
409 /* True 32b PPGTT with dynamic page allocation: update PDP
410 * registers and point the unallocated PDPs to scratch page.
411 * PML4 is allocated during ppgtt init, so this is not needed
412 * in 48-bit mode.
413 */
414 if (ppgtt && !i915_vm_is_48bit(&ppgtt->base))
415 execlists_update_context_pdps(ppgtt, reg_state);
416
417 return ce->lrc_desc;
418}
419
420static inline void write_desc(struct intel_engine_execlists *execlists, u64 desc, u32 port)
421{
422 if (execlists->ctrl_reg) {
423 writel(lower_32_bits(desc), execlists->submit_reg + port * 2);
424 writel(upper_32_bits(desc), execlists->submit_reg + port * 2 + 1);
425 } else {
426 writel(upper_32_bits(desc), execlists->submit_reg);
427 writel(lower_32_bits(desc), execlists->submit_reg);
428 }
429}
430
431static void execlists_submit_ports(struct intel_engine_cs *engine)
432{
433 struct intel_engine_execlists *execlists = &engine->execlists;
434 struct execlist_port *port = execlists->port;
435 unsigned int n;
436
437 /*
438 * ELSQ note: the submit queue is not cleared after being submitted
439 * to the HW so we need to make sure we always clean it up. This is
440 * currently ensured by the fact that we always write the same number
441 * of elsq entries, keep this in mind before changing the loop below.
442 */
443 for (n = execlists_num_ports(execlists); n--; ) {
444 struct i915_request *rq;
445 unsigned int count;
446 u64 desc;
447
448 rq = port_unpack(&port[n], &count);
449 if (rq) {
450 GEM_BUG_ON(count > !n);
451 if (!count++)
452 execlists_context_schedule_in(rq);
453 port_set(&port[n], port_pack(rq, count));
454 desc = execlists_update_context(rq);
455 GEM_DEBUG_EXEC(port[n].context_id = upper_32_bits(desc));
456
457 GEM_TRACE("%s in[%d]: ctx=%d.%d, seqno=%x, prio=%d\n",
458 engine->name, n,
459 port[n].context_id, count,
460 rq->global_seqno,
461 rq_prio(rq));
462 } else {
463 GEM_BUG_ON(!n);
464 desc = 0;
465 }
466
467 write_desc(execlists, desc, n);
468 }
469
470 /* we need to manually load the submit queue */
471 if (execlists->ctrl_reg)
472 writel(EL_CTRL_LOAD, execlists->ctrl_reg);
473
474 execlists_clear_active(execlists, EXECLISTS_ACTIVE_HWACK);
475}
476
477static bool ctx_single_port_submission(const struct i915_gem_context *ctx)
478{
479 return (IS_ENABLED(CONFIG_DRM_I915_GVT) &&
480 i915_gem_context_force_single_submission(ctx));
481}
482
483static bool can_merge_ctx(const struct i915_gem_context *prev,
484 const struct i915_gem_context *next)
485{
486 if (prev != next)
487 return false;
488
489 if (ctx_single_port_submission(prev))
490 return false;
491
492 return true;
493}
494
495static void port_assign(struct execlist_port *port, struct i915_request *rq)
496{
497 GEM_BUG_ON(rq == port_request(port));
498
499 if (port_isset(port))
500 i915_request_put(port_request(port));
501
502 port_set(port, port_pack(i915_request_get(rq), port_count(port)));
503}
504
505static void inject_preempt_context(struct intel_engine_cs *engine)
506{
507 struct intel_engine_execlists *execlists = &engine->execlists;
508 struct intel_context *ce =
509 &engine->i915->preempt_context->engine[engine->id];
510 unsigned int n;
511
512 GEM_BUG_ON(execlists->preempt_complete_status !=
513 upper_32_bits(ce->lrc_desc));
514 GEM_BUG_ON((ce->lrc_reg_state[CTX_CONTEXT_CONTROL + 1] &
515 _MASKED_BIT_ENABLE(CTX_CTRL_ENGINE_CTX_RESTORE_INHIBIT |
516 CTX_CTRL_ENGINE_CTX_SAVE_INHIBIT)) !=
517 _MASKED_BIT_ENABLE(CTX_CTRL_ENGINE_CTX_RESTORE_INHIBIT |
518 CTX_CTRL_ENGINE_CTX_SAVE_INHIBIT));
519
520 /*
521 * Switch to our empty preempt context so
522 * the state of the GPU is known (idle).
523 */
524 GEM_TRACE("%s\n", engine->name);
525 for (n = execlists_num_ports(execlists); --n; )
526 write_desc(execlists, 0, n);
527
528 write_desc(execlists, ce->lrc_desc, n);
529
530 /* we need to manually load the submit queue */
531 if (execlists->ctrl_reg)
532 writel(EL_CTRL_LOAD, execlists->ctrl_reg);
533
534 execlists_clear_active(&engine->execlists, EXECLISTS_ACTIVE_HWACK);
535 execlists_set_active(&engine->execlists, EXECLISTS_ACTIVE_PREEMPT);
536}
537
538static void execlists_dequeue(struct intel_engine_cs *engine)
539{
540 struct intel_engine_execlists * const execlists = &engine->execlists;
541 struct execlist_port *port = execlists->port;
542 const struct execlist_port * const last_port =
543 &execlists->port[execlists->port_mask];
544 struct i915_request *last = port_request(port);
545 struct rb_node *rb;
546 bool submit = false;
547
548 /* Hardware submission is through 2 ports. Conceptually each port
549 * has a (RING_START, RING_HEAD, RING_TAIL) tuple. RING_START is
550 * static for a context, and unique to each, so we only execute
551 * requests belonging to a single context from each ring. RING_HEAD
552 * is maintained by the CS in the context image, it marks the place
553 * where it got up to last time, and through RING_TAIL we tell the CS
554 * where we want to execute up to this time.
555 *
556 * In this list the requests are in order of execution. Consecutive
557 * requests from the same context are adjacent in the ringbuffer. We
558 * can combine these requests into a single RING_TAIL update:
559 *
560 * RING_HEAD...req1...req2
561 * ^- RING_TAIL
562 * since to execute req2 the CS must first execute req1.
563 *
564 * Our goal then is to point each port to the end of a consecutive
565 * sequence of requests as being the most optimal (fewest wake ups
566 * and context switches) submission.
567 */
568
569 spin_lock_irq(&engine->timeline->lock);
570 rb = execlists->first;
571 GEM_BUG_ON(rb_first(&execlists->queue) != rb);
572
573 if (last) {
574 /*
575 * Don't resubmit or switch until all outstanding
576 * preemptions (lite-restore) are seen. Then we
577 * know the next preemption status we see corresponds
578 * to this ELSP update.
579 */
580 GEM_BUG_ON(!execlists_is_active(execlists,
581 EXECLISTS_ACTIVE_USER));
582 GEM_BUG_ON(!port_count(&port[0]));
583 if (port_count(&port[0]) > 1)
584 goto unlock;
585
586 /*
587 * If we write to ELSP a second time before the HW has had
588 * a chance to respond to the previous write, we can confuse
589 * the HW and hit "undefined behaviour". After writing to ELSP,
590 * we must then wait until we see a context-switch event from
591 * the HW to indicate that it has had a chance to respond.
592 */
593 if (!execlists_is_active(execlists, EXECLISTS_ACTIVE_HWACK))
594 goto unlock;
595
596 if (need_preempt(engine, last, execlists->queue_priority)) {
597 inject_preempt_context(engine);
598 goto unlock;
599 }
600
601 /*
602 * In theory, we could coalesce more requests onto
603 * the second port (the first port is active, with
604 * no preemptions pending). However, that means we
605 * then have to deal with the possible lite-restore
606 * of the second port (as we submit the ELSP, there
607 * may be a context-switch) but also we may complete
608 * the resubmission before the context-switch. Ergo,
609 * coalescing onto the second port will cause a
610 * preemption event, but we cannot predict whether
611 * that will affect port[0] or port[1].
612 *
613 * If the second port is already active, we can wait
614 * until the next context-switch before contemplating
615 * new requests. The GPU will be busy and we should be
616 * able to resubmit the new ELSP before it idles,
617 * avoiding pipeline bubbles (momentary pauses where
618 * the driver is unable to keep up the supply of new
619 * work). However, we have to double check that the
620 * priorities of the ports haven't been switch.
621 */
622 if (port_count(&port[1]))
623 goto unlock;
624
625 /*
626 * WaIdleLiteRestore:bdw,skl
627 * Apply the wa NOOPs to prevent
628 * ring:HEAD == rq:TAIL as we resubmit the
629 * request. See gen8_emit_breadcrumb() for
630 * where we prepare the padding after the
631 * end of the request.
632 */
633 last->tail = last->wa_tail;
634 }
635
636 while (rb) {
637 struct i915_priolist *p = to_priolist(rb);
638 struct i915_request *rq, *rn;
639
640 list_for_each_entry_safe(rq, rn, &p->requests, priotree.link) {
641 /*
642 * Can we combine this request with the current port?
643 * It has to be the same context/ringbuffer and not
644 * have any exceptions (e.g. GVT saying never to
645 * combine contexts).
646 *
647 * If we can combine the requests, we can execute both
648 * by updating the RING_TAIL to point to the end of the
649 * second request, and so we never need to tell the
650 * hardware about the first.
651 */
652 if (last && !can_merge_ctx(rq->ctx, last->ctx)) {
653 /*
654 * If we are on the second port and cannot
655 * combine this request with the last, then we
656 * are done.
657 */
658 if (port == last_port) {
659 __list_del_many(&p->requests,
660 &rq->priotree.link);
661 goto done;
662 }
663
664 /*
665 * If GVT overrides us we only ever submit
666 * port[0], leaving port[1] empty. Note that we
667 * also have to be careful that we don't queue
668 * the same context (even though a different
669 * request) to the second port.
670 */
671 if (ctx_single_port_submission(last->ctx) ||
672 ctx_single_port_submission(rq->ctx)) {
673 __list_del_many(&p->requests,
674 &rq->priotree.link);
675 goto done;
676 }
677
678 GEM_BUG_ON(last->ctx == rq->ctx);
679
680 if (submit)
681 port_assign(port, last);
682 port++;
683
684 GEM_BUG_ON(port_isset(port));
685 }
686
687 INIT_LIST_HEAD(&rq->priotree.link);
688 __i915_request_submit(rq);
689 trace_i915_request_in(rq, port_index(port, execlists));
690 last = rq;
691 submit = true;
692 }
693
694 rb = rb_next(rb);
695 rb_erase(&p->node, &execlists->queue);
696 INIT_LIST_HEAD(&p->requests);
697 if (p->priority != I915_PRIORITY_NORMAL)
698 kmem_cache_free(engine->i915->priorities, p);
699 }
700done:
701 execlists->queue_priority = rb ? to_priolist(rb)->priority : INT_MIN;
702 execlists->first = rb;
703 if (submit)
704 port_assign(port, last);
705
706 /* We must always keep the beast fed if we have work piled up */
707 GEM_BUG_ON(execlists->first && !port_isset(execlists->port));
708
709unlock:
710 spin_unlock_irq(&engine->timeline->lock);
711
712 if (submit) {
713 execlists_set_active(execlists, EXECLISTS_ACTIVE_USER);
714 execlists_submit_ports(engine);
715 }
716
717 GEM_BUG_ON(port_isset(execlists->port) &&
718 !execlists_is_active(execlists, EXECLISTS_ACTIVE_USER));
719}
720
721void
722execlists_cancel_port_requests(struct intel_engine_execlists * const execlists)
723{
724 struct execlist_port *port = execlists->port;
725 unsigned int num_ports = execlists_num_ports(execlists);
726
727 while (num_ports-- && port_isset(port)) {
728 struct i915_request *rq = port_request(port);
729
730 GEM_BUG_ON(!execlists->active);
731 intel_engine_context_out(rq->engine);
732
733 execlists_context_status_change(rq,
734 i915_request_completed(rq) ?
735 INTEL_CONTEXT_SCHEDULE_OUT :
736 INTEL_CONTEXT_SCHEDULE_PREEMPTED);
737
738 i915_request_put(rq);
739
740 memset(port, 0, sizeof(*port));
741 port++;
742 }
743
744 execlists_clear_active(execlists, EXECLISTS_ACTIVE_USER);
745}
746
747static void execlists_cancel_requests(struct intel_engine_cs *engine)
748{
749 struct intel_engine_execlists * const execlists = &engine->execlists;
750 struct i915_request *rq, *rn;
751 struct rb_node *rb;
752 unsigned long flags;
753
754 GEM_TRACE("%s\n", engine->name);
755
756 /*
757 * Before we call engine->cancel_requests(), we should have exclusive
758 * access to the submission state. This is arranged for us by the
759 * caller disabling the interrupt generation, the tasklet and other
760 * threads that may then access the same state, giving us a free hand
761 * to reset state. However, we still need to let lockdep be aware that
762 * we know this state may be accessed in hardirq context, so we
763 * disable the irq around this manipulation and we want to keep
764 * the spinlock focused on its duties and not accidentally conflate
765 * coverage to the submission's irq state. (Similarly, although we
766 * shouldn't need to disable irq around the manipulation of the
767 * submission's irq state, we also wish to remind ourselves that
768 * it is irq state.)
769 */
770 local_irq_save(flags);
771
772 /* Cancel the requests on the HW and clear the ELSP tracker. */
773 execlists_cancel_port_requests(execlists);
774
775 spin_lock(&engine->timeline->lock);
776
777 /* Mark all executing requests as skipped. */
778 list_for_each_entry(rq, &engine->timeline->requests, link) {
779 GEM_BUG_ON(!rq->global_seqno);
780 if (!i915_request_completed(rq))
781 dma_fence_set_error(&rq->fence, -EIO);
782 }
783
784 /* Flush the queued requests to the timeline list (for retiring). */
785 rb = execlists->first;
786 while (rb) {
787 struct i915_priolist *p = to_priolist(rb);
788
789 list_for_each_entry_safe(rq, rn, &p->requests, priotree.link) {
790 INIT_LIST_HEAD(&rq->priotree.link);
791
792 dma_fence_set_error(&rq->fence, -EIO);
793 __i915_request_submit(rq);
794 }
795
796 rb = rb_next(rb);
797 rb_erase(&p->node, &execlists->queue);
798 INIT_LIST_HEAD(&p->requests);
799 if (p->priority != I915_PRIORITY_NORMAL)
800 kmem_cache_free(engine->i915->priorities, p);
801 }
802
803 /* Remaining _unready_ requests will be nop'ed when submitted */
804
805 execlists->queue_priority = INT_MIN;
806 execlists->queue = RB_ROOT;
807 execlists->first = NULL;
808 GEM_BUG_ON(port_isset(execlists->port));
809
810 spin_unlock(&engine->timeline->lock);
811
812 /*
813 * The port is checked prior to scheduling a tasklet, but
814 * just in case we have suspended the tasklet to do the
815 * wedging make sure that when it wakes, it decides there
816 * is no work to do by clearing the irq_posted bit.
817 */
818 clear_bit(ENGINE_IRQ_EXECLIST, &engine->irq_posted);
819
820 /* Mark all CS interrupts as complete */
821 execlists->active = 0;
822
823 local_irq_restore(flags);
824}
825
826/*
827 * Check the unread Context Status Buffers and manage the submission of new
828 * contexts to the ELSP accordingly.
829 */
830static void execlists_submission_tasklet(unsigned long data)
831{
832 struct intel_engine_cs * const engine = (struct intel_engine_cs *)data;
833 struct intel_engine_execlists * const execlists = &engine->execlists;
834 struct execlist_port * const port = execlists->port;
835 struct drm_i915_private *dev_priv = engine->i915;
836 bool fw = false;
837
838 /*
839 * We can skip acquiring intel_runtime_pm_get() here as it was taken
840 * on our behalf by the request (see i915_gem_mark_busy()) and it will
841 * not be relinquished until the device is idle (see
842 * i915_gem_idle_work_handler()). As a precaution, we make sure
843 * that all ELSP are drained i.e. we have processed the CSB,
844 * before allowing ourselves to idle and calling intel_runtime_pm_put().
845 */
846 GEM_BUG_ON(!dev_priv->gt.awake);
847
848 /*
849 * Prefer doing test_and_clear_bit() as a two stage operation to avoid
850 * imposing the cost of a locked atomic transaction when submitting a
851 * new request (outside of the context-switch interrupt).
852 */
853 while (test_bit(ENGINE_IRQ_EXECLIST, &engine->irq_posted)) {
854 /* The HWSP contains a (cacheable) mirror of the CSB */
855 const u32 *buf =
856 &engine->status_page.page_addr[I915_HWS_CSB_BUF0_INDEX];
857 unsigned int head, tail;
858
859 if (unlikely(execlists->csb_use_mmio)) {
860 buf = (u32 * __force)
861 (dev_priv->regs + i915_mmio_reg_offset(RING_CONTEXT_STATUS_BUF_LO(engine, 0)));
862 execlists->csb_head = -1; /* force mmio read of CSB ptrs */
863 }
864
865 /* Clear before reading to catch new interrupts */
866 clear_bit(ENGINE_IRQ_EXECLIST, &engine->irq_posted);
867 smp_mb__after_atomic();
868
869 if (unlikely(execlists->csb_head == -1)) { /* following a reset */
870 if (!fw) {
871 intel_uncore_forcewake_get(dev_priv,
872 execlists->fw_domains);
873 fw = true;
874 }
875
876 head = readl(dev_priv->regs + i915_mmio_reg_offset(RING_CONTEXT_STATUS_PTR(engine)));
877 tail = GEN8_CSB_WRITE_PTR(head);
878 head = GEN8_CSB_READ_PTR(head);
879 execlists->csb_head = head;
880 } else {
881 const int write_idx =
882 intel_hws_csb_write_index(dev_priv) -
883 I915_HWS_CSB_BUF0_INDEX;
884
885 head = execlists->csb_head;
886 tail = READ_ONCE(buf[write_idx]);
887 rmb(); /* Hopefully paired with a wmb() in HW */
888 }
889 GEM_TRACE("%s cs-irq head=%d [%d%s], tail=%d [%d%s]\n",
890 engine->name,
891 head, GEN8_CSB_READ_PTR(readl(dev_priv->regs + i915_mmio_reg_offset(RING_CONTEXT_STATUS_PTR(engine)))), fw ? "" : "?",
892 tail, GEN8_CSB_WRITE_PTR(readl(dev_priv->regs + i915_mmio_reg_offset(RING_CONTEXT_STATUS_PTR(engine)))), fw ? "" : "?");
893
894 while (head != tail) {
895 struct i915_request *rq;
896 unsigned int status;
897 unsigned int count;
898
899 if (++head == GEN8_CSB_ENTRIES)
900 head = 0;
901
902 /* We are flying near dragons again.
903 *
904 * We hold a reference to the request in execlist_port[]
905 * but no more than that. We are operating in softirq
906 * context and so cannot hold any mutex or sleep. That
907 * prevents us stopping the requests we are processing
908 * in port[] from being retired simultaneously (the
909 * breadcrumb will be complete before we see the
910 * context-switch). As we only hold the reference to the
911 * request, any pointer chasing underneath the request
912 * is subject to a potential use-after-free. Thus we
913 * store all of the bookkeeping within port[] as
914 * required, and avoid using unguarded pointers beneath
915 * request itself. The same applies to the atomic
916 * status notifier.
917 */
918
919 status = READ_ONCE(buf[2 * head]); /* maybe mmio! */
920 GEM_TRACE("%s csb[%d]: status=0x%08x:0x%08x, active=0x%x\n",
921 engine->name, head,
922 status, buf[2*head + 1],
923 execlists->active);
924
925 if (status & (GEN8_CTX_STATUS_IDLE_ACTIVE |
926 GEN8_CTX_STATUS_PREEMPTED))
927 execlists_set_active(execlists,
928 EXECLISTS_ACTIVE_HWACK);
929 if (status & GEN8_CTX_STATUS_ACTIVE_IDLE)
930 execlists_clear_active(execlists,
931 EXECLISTS_ACTIVE_HWACK);
932
933 if (!(status & GEN8_CTX_STATUS_COMPLETED_MASK))
934 continue;
935
936 /* We should never get a COMPLETED | IDLE_ACTIVE! */
937 GEM_BUG_ON(status & GEN8_CTX_STATUS_IDLE_ACTIVE);
938
939 if (status & GEN8_CTX_STATUS_COMPLETE &&
940 buf[2*head + 1] == execlists->preempt_complete_status) {
941 GEM_TRACE("%s preempt-idle\n", engine->name);
942
943 execlists_cancel_port_requests(execlists);
944 execlists_unwind_incomplete_requests(execlists);
945
946 GEM_BUG_ON(!execlists_is_active(execlists,
947 EXECLISTS_ACTIVE_PREEMPT));
948 execlists_clear_active(execlists,
949 EXECLISTS_ACTIVE_PREEMPT);
950 continue;
951 }
952
953 if (status & GEN8_CTX_STATUS_PREEMPTED &&
954 execlists_is_active(execlists,
955 EXECLISTS_ACTIVE_PREEMPT))
956 continue;
957
958 GEM_BUG_ON(!execlists_is_active(execlists,
959 EXECLISTS_ACTIVE_USER));
960
961 rq = port_unpack(port, &count);
962 GEM_TRACE("%s out[0]: ctx=%d.%d, seqno=%x, prio=%d\n",
963 engine->name,
964 port->context_id, count,
965 rq ? rq->global_seqno : 0,
966 rq ? rq_prio(rq) : 0);
967
968 /* Check the context/desc id for this event matches */
969 GEM_DEBUG_BUG_ON(buf[2 * head + 1] != port->context_id);
970
971 GEM_BUG_ON(count == 0);
972 if (--count == 0) {
973 GEM_BUG_ON(status & GEN8_CTX_STATUS_PREEMPTED);
974 GEM_BUG_ON(port_isset(&port[1]) &&
975 !(status & GEN8_CTX_STATUS_ELEMENT_SWITCH));
976 GEM_BUG_ON(!i915_request_completed(rq));
977 execlists_context_schedule_out(rq);
978 trace_i915_request_out(rq);
979 i915_request_put(rq);
980
981 GEM_TRACE("%s completed ctx=%d\n",
982 engine->name, port->context_id);
983
984 execlists_port_complete(execlists, port);
985 } else {
986 port_set(port, port_pack(rq, count));
987 }
988
989 /* After the final element, the hw should be idle */
990 GEM_BUG_ON(port_count(port) == 0 &&
991 !(status & GEN8_CTX_STATUS_ACTIVE_IDLE));
992 if (port_count(port) == 0)
993 execlists_clear_active(execlists,
994 EXECLISTS_ACTIVE_USER);
995 }
996
997 if (head != execlists->csb_head) {
998 execlists->csb_head = head;
999 writel(_MASKED_FIELD(GEN8_CSB_READ_PTR_MASK, head << 8),
1000 dev_priv->regs + i915_mmio_reg_offset(RING_CONTEXT_STATUS_PTR(engine)));
1001 }
1002 }
1003
1004 if (!execlists_is_active(execlists, EXECLISTS_ACTIVE_PREEMPT))
1005 execlists_dequeue(engine);
1006
1007 if (fw)
1008 intel_uncore_forcewake_put(dev_priv, execlists->fw_domains);
1009
1010 /* If the engine is now idle, so should be the flag; and vice versa. */
1011 GEM_BUG_ON(execlists_is_active(&engine->execlists,
1012 EXECLISTS_ACTIVE_USER) ==
1013 !port_isset(engine->execlists.port));
1014}
1015
1016static void queue_request(struct intel_engine_cs *engine,
1017 struct i915_priotree *pt,
1018 int prio)
1019{
1020 list_add_tail(&pt->link, &lookup_priolist(engine, pt, prio)->requests);
1021}
1022
1023static void submit_queue(struct intel_engine_cs *engine, int prio)
1024{
1025 if (prio > engine->execlists.queue_priority) {
1026 engine->execlists.queue_priority = prio;
1027 tasklet_hi_schedule(&engine->execlists.tasklet);
1028 }
1029}
1030
1031static void execlists_submit_request(struct i915_request *request)
1032{
1033 struct intel_engine_cs *engine = request->engine;
1034 unsigned long flags;
1035
1036 /* Will be called from irq-context when using foreign fences. */
1037 spin_lock_irqsave(&engine->timeline->lock, flags);
1038
1039 queue_request(engine, &request->priotree, rq_prio(request));
1040 submit_queue(engine, rq_prio(request));
1041
1042 GEM_BUG_ON(!engine->execlists.first);
1043 GEM_BUG_ON(list_empty(&request->priotree.link));
1044
1045 spin_unlock_irqrestore(&engine->timeline->lock, flags);
1046}
1047
1048static struct i915_request *pt_to_request(struct i915_priotree *pt)
1049{
1050 return container_of(pt, struct i915_request, priotree);
1051}
1052
1053static struct intel_engine_cs *
1054pt_lock_engine(struct i915_priotree *pt, struct intel_engine_cs *locked)
1055{
1056 struct intel_engine_cs *engine = pt_to_request(pt)->engine;
1057
1058 GEM_BUG_ON(!locked);
1059
1060 if (engine != locked) {
1061 spin_unlock(&locked->timeline->lock);
1062 spin_lock(&engine->timeline->lock);
1063 }
1064
1065 return engine;
1066}
1067
1068static void execlists_schedule(struct i915_request *request, int prio)
1069{
1070 struct intel_engine_cs *engine;
1071 struct i915_dependency *dep, *p;
1072 struct i915_dependency stack;
1073 LIST_HEAD(dfs);
1074
1075 GEM_BUG_ON(prio == I915_PRIORITY_INVALID);
1076
1077 if (i915_request_completed(request))
1078 return;
1079
1080 if (prio <= READ_ONCE(request->priotree.priority))
1081 return;
1082
1083 /* Need BKL in order to use the temporary link inside i915_dependency */
1084 lockdep_assert_held(&request->i915->drm.struct_mutex);
1085
1086 stack.signaler = &request->priotree;
1087 list_add(&stack.dfs_link, &dfs);
1088
1089 /*
1090 * Recursively bump all dependent priorities to match the new request.
1091 *
1092 * A naive approach would be to use recursion:
1093 * static void update_priorities(struct i915_priotree *pt, prio) {
1094 * list_for_each_entry(dep, &pt->signalers_list, signal_link)
1095 * update_priorities(dep->signal, prio)
1096 * queue_request(pt);
1097 * }
1098 * but that may have unlimited recursion depth and so runs a very
1099 * real risk of overunning the kernel stack. Instead, we build
1100 * a flat list of all dependencies starting with the current request.
1101 * As we walk the list of dependencies, we add all of its dependencies
1102 * to the end of the list (this may include an already visited
1103 * request) and continue to walk onwards onto the new dependencies. The
1104 * end result is a topological list of requests in reverse order, the
1105 * last element in the list is the request we must execute first.
1106 */
1107 list_for_each_entry(dep, &dfs, dfs_link) {
1108 struct i915_priotree *pt = dep->signaler;
1109
1110 /*
1111 * Within an engine, there can be no cycle, but we may
1112 * refer to the same dependency chain multiple times
1113 * (redundant dependencies are not eliminated) and across
1114 * engines.
1115 */
1116 list_for_each_entry(p, &pt->signalers_list, signal_link) {
1117 GEM_BUG_ON(p == dep); /* no cycles! */
1118
1119 if (i915_priotree_signaled(p->signaler))
1120 continue;
1121
1122 GEM_BUG_ON(p->signaler->priority < pt->priority);
1123 if (prio > READ_ONCE(p->signaler->priority))
1124 list_move_tail(&p->dfs_link, &dfs);
1125 }
1126 }
1127
1128 /*
1129 * If we didn't need to bump any existing priorities, and we haven't
1130 * yet submitted this request (i.e. there is no potential race with
1131 * execlists_submit_request()), we can set our own priority and skip
1132 * acquiring the engine locks.
1133 */
1134 if (request->priotree.priority == I915_PRIORITY_INVALID) {
1135 GEM_BUG_ON(!list_empty(&request->priotree.link));
1136 request->priotree.priority = prio;
1137 if (stack.dfs_link.next == stack.dfs_link.prev)
1138 return;
1139 __list_del_entry(&stack.dfs_link);
1140 }
1141
1142 engine = request->engine;
1143 spin_lock_irq(&engine->timeline->lock);
1144
1145 /* Fifo and depth-first replacement ensure our deps execute before us */
1146 list_for_each_entry_safe_reverse(dep, p, &dfs, dfs_link) {
1147 struct i915_priotree *pt = dep->signaler;
1148
1149 INIT_LIST_HEAD(&dep->dfs_link);
1150
1151 engine = pt_lock_engine(pt, engine);
1152
1153 if (prio <= pt->priority)
1154 continue;
1155
1156 pt->priority = prio;
1157 if (!list_empty(&pt->link)) {
1158 __list_del_entry(&pt->link);
1159 queue_request(engine, pt, prio);
1160 }
1161 submit_queue(engine, prio);
1162 }
1163
1164 spin_unlock_irq(&engine->timeline->lock);
1165}
1166
1167static int __context_pin(struct i915_gem_context *ctx, struct i915_vma *vma)
1168{
1169 unsigned int flags;
1170 int err;
1171
1172 /*
1173 * Clear this page out of any CPU caches for coherent swap-in/out.
1174 * We only want to do this on the first bind so that we do not stall
1175 * on an active context (which by nature is already on the GPU).
1176 */
1177 if (!(vma->flags & I915_VMA_GLOBAL_BIND)) {
1178 err = i915_gem_object_set_to_gtt_domain(vma->obj, true);
1179 if (err)
1180 return err;
1181 }
1182
1183 flags = PIN_GLOBAL | PIN_HIGH;
1184 if (ctx->ggtt_offset_bias)
1185 flags |= PIN_OFFSET_BIAS | ctx->ggtt_offset_bias;
1186
1187 return i915_vma_pin(vma, 0, GEN8_LR_CONTEXT_ALIGN, flags);
1188}
1189
1190static struct intel_ring *
1191execlists_context_pin(struct intel_engine_cs *engine,
1192 struct i915_gem_context *ctx)
1193{
1194 struct intel_context *ce = &ctx->engine[engine->id];
1195 void *vaddr;
1196 int ret;
1197
1198 lockdep_assert_held(&ctx->i915->drm.struct_mutex);
1199
1200 if (likely(ce->pin_count++))
1201 goto out;
1202 GEM_BUG_ON(!ce->pin_count); /* no overflow please! */
1203
1204 ret = execlists_context_deferred_alloc(ctx, engine);
1205 if (ret)
1206 goto err;
1207 GEM_BUG_ON(!ce->state);
1208
1209 ret = __context_pin(ctx, ce->state);
1210 if (ret)
1211 goto err;
1212
1213 vaddr = i915_gem_object_pin_map(ce->state->obj, I915_MAP_WB);
1214 if (IS_ERR(vaddr)) {
1215 ret = PTR_ERR(vaddr);
1216 goto unpin_vma;
1217 }
1218
1219 ret = intel_ring_pin(ce->ring, ctx->i915, ctx->ggtt_offset_bias);
1220 if (ret)
1221 goto unpin_map;
1222
1223 intel_lr_context_descriptor_update(ctx, engine);
1224
1225 ce->lrc_reg_state = vaddr + LRC_STATE_PN * PAGE_SIZE;
1226 ce->lrc_reg_state[CTX_RING_BUFFER_START+1] =
1227 i915_ggtt_offset(ce->ring->vma);
1228
1229 ce->state->obj->pin_global++;
1230 i915_gem_context_get(ctx);
1231out:
1232 return ce->ring;
1233
1234unpin_map:
1235 i915_gem_object_unpin_map(ce->state->obj);
1236unpin_vma:
1237 __i915_vma_unpin(ce->state);
1238err:
1239 ce->pin_count = 0;
1240 return ERR_PTR(ret);
1241}
1242
1243static void execlists_context_unpin(struct intel_engine_cs *engine,
1244 struct i915_gem_context *ctx)
1245{
1246 struct intel_context *ce = &ctx->engine[engine->id];
1247
1248 lockdep_assert_held(&ctx->i915->drm.struct_mutex);
1249 GEM_BUG_ON(ce->pin_count == 0);
1250
1251 if (--ce->pin_count)
1252 return;
1253
1254 intel_ring_unpin(ce->ring);
1255
1256 ce->state->obj->pin_global--;
1257 i915_gem_object_unpin_map(ce->state->obj);
1258 i915_vma_unpin(ce->state);
1259
1260 i915_gem_context_put(ctx);
1261}
1262
1263static int execlists_request_alloc(struct i915_request *request)
1264{
1265 struct intel_engine_cs *engine = request->engine;
1266 struct intel_context *ce = &request->ctx->engine[engine->id];
1267 int ret;
1268
1269 GEM_BUG_ON(!ce->pin_count);
1270
1271 /* Flush enough space to reduce the likelihood of waiting after
1272 * we start building the request - in which case we will just
1273 * have to repeat work.
1274 */
1275 request->reserved_space += EXECLISTS_REQUEST_SIZE;
1276
1277 ret = intel_ring_wait_for_space(request->ring, request->reserved_space);
1278 if (ret)
1279 return ret;
1280
1281 /* Note that after this point, we have committed to using
1282 * this request as it is being used to both track the
1283 * state of engine initialisation and liveness of the
1284 * golden renderstate above. Think twice before you try
1285 * to cancel/unwind this request now.
1286 */
1287
1288 request->reserved_space -= EXECLISTS_REQUEST_SIZE;
1289 return 0;
1290}
1291
1292/*
1293 * In this WA we need to set GEN8_L3SQCREG4[21:21] and reset it after
1294 * PIPE_CONTROL instruction. This is required for the flush to happen correctly
1295 * but there is a slight complication as this is applied in WA batch where the
1296 * values are only initialized once so we cannot take register value at the
1297 * beginning and reuse it further; hence we save its value to memory, upload a
1298 * constant value with bit21 set and then we restore it back with the saved value.
1299 * To simplify the WA, a constant value is formed by using the default value
1300 * of this register. This shouldn't be a problem because we are only modifying
1301 * it for a short period and this batch in non-premptible. We can ofcourse
1302 * use additional instructions that read the actual value of the register
1303 * at that time and set our bit of interest but it makes the WA complicated.
1304 *
1305 * This WA is also required for Gen9 so extracting as a function avoids
1306 * code duplication.
1307 */
1308static u32 *
1309gen8_emit_flush_coherentl3_wa(struct intel_engine_cs *engine, u32 *batch)
1310{
1311 *batch++ = MI_STORE_REGISTER_MEM_GEN8 | MI_SRM_LRM_GLOBAL_GTT;
1312 *batch++ = i915_mmio_reg_offset(GEN8_L3SQCREG4);
1313 *batch++ = i915_ggtt_offset(engine->scratch) + 256;
1314 *batch++ = 0;
1315
1316 *batch++ = MI_LOAD_REGISTER_IMM(1);
1317 *batch++ = i915_mmio_reg_offset(GEN8_L3SQCREG4);
1318 *batch++ = 0x40400000 | GEN8_LQSC_FLUSH_COHERENT_LINES;
1319
1320 batch = gen8_emit_pipe_control(batch,
1321 PIPE_CONTROL_CS_STALL |
1322 PIPE_CONTROL_DC_FLUSH_ENABLE,
1323 0);
1324
1325 *batch++ = MI_LOAD_REGISTER_MEM_GEN8 | MI_SRM_LRM_GLOBAL_GTT;
1326 *batch++ = i915_mmio_reg_offset(GEN8_L3SQCREG4);
1327 *batch++ = i915_ggtt_offset(engine->scratch) + 256;
1328 *batch++ = 0;
1329
1330 return batch;
1331}
1332
1333/*
1334 * Typically we only have one indirect_ctx and per_ctx batch buffer which are
1335 * initialized at the beginning and shared across all contexts but this field
1336 * helps us to have multiple batches at different offsets and select them based
1337 * on a criteria. At the moment this batch always start at the beginning of the page
1338 * and at this point we don't have multiple wa_ctx batch buffers.
1339 *
1340 * The number of WA applied are not known at the beginning; we use this field
1341 * to return the no of DWORDS written.
1342 *
1343 * It is to be noted that this batch does not contain MI_BATCH_BUFFER_END
1344 * so it adds NOOPs as padding to make it cacheline aligned.
1345 * MI_BATCH_BUFFER_END will be added to perctx batch and both of them together
1346 * makes a complete batch buffer.
1347 */
1348static u32 *gen8_init_indirectctx_bb(struct intel_engine_cs *engine, u32 *batch)
1349{
1350 /* WaDisableCtxRestoreArbitration:bdw,chv */
1351 *batch++ = MI_ARB_ON_OFF | MI_ARB_DISABLE;
1352
1353 /* WaFlushCoherentL3CacheLinesAtContextSwitch:bdw */
1354 if (IS_BROADWELL(engine->i915))
1355 batch = gen8_emit_flush_coherentl3_wa(engine, batch);
1356
1357 /* WaClearSlmSpaceAtContextSwitch:bdw,chv */
1358 /* Actual scratch location is at 128 bytes offset */
1359 batch = gen8_emit_pipe_control(batch,
1360 PIPE_CONTROL_FLUSH_L3 |
1361 PIPE_CONTROL_GLOBAL_GTT_IVB |
1362 PIPE_CONTROL_CS_STALL |
1363 PIPE_CONTROL_QW_WRITE,
1364 i915_ggtt_offset(engine->scratch) +
1365 2 * CACHELINE_BYTES);
1366
1367 *batch++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
1368
1369 /* Pad to end of cacheline */
1370 while ((unsigned long)batch % CACHELINE_BYTES)
1371 *batch++ = MI_NOOP;
1372
1373 /*
1374 * MI_BATCH_BUFFER_END is not required in Indirect ctx BB because
1375 * execution depends on the length specified in terms of cache lines
1376 * in the register CTX_RCS_INDIRECT_CTX
1377 */
1378
1379 return batch;
1380}
1381
1382static u32 *gen9_init_indirectctx_bb(struct intel_engine_cs *engine, u32 *batch)
1383{
1384 *batch++ = MI_ARB_ON_OFF | MI_ARB_DISABLE;
1385
1386 /* WaFlushCoherentL3CacheLinesAtContextSwitch:skl,bxt,glk */
1387 batch = gen8_emit_flush_coherentl3_wa(engine, batch);
1388
1389 /* WaDisableGatherAtSetShaderCommonSlice:skl,bxt,kbl,glk */
1390 *batch++ = MI_LOAD_REGISTER_IMM(1);
1391 *batch++ = i915_mmio_reg_offset(COMMON_SLICE_CHICKEN2);
1392 *batch++ = _MASKED_BIT_DISABLE(
1393 GEN9_DISABLE_GATHER_AT_SET_SHADER_COMMON_SLICE);
1394 *batch++ = MI_NOOP;
1395
1396 /* WaClearSlmSpaceAtContextSwitch:kbl */
1397 /* Actual scratch location is at 128 bytes offset */
1398 if (IS_KBL_REVID(engine->i915, 0, KBL_REVID_A0)) {
1399 batch = gen8_emit_pipe_control(batch,
1400 PIPE_CONTROL_FLUSH_L3 |
1401 PIPE_CONTROL_GLOBAL_GTT_IVB |
1402 PIPE_CONTROL_CS_STALL |
1403 PIPE_CONTROL_QW_WRITE,
1404 i915_ggtt_offset(engine->scratch)
1405 + 2 * CACHELINE_BYTES);
1406 }
1407
1408 /* WaMediaPoolStateCmdInWABB:bxt,glk */
1409 if (HAS_POOLED_EU(engine->i915)) {
1410 /*
1411 * EU pool configuration is setup along with golden context
1412 * during context initialization. This value depends on
1413 * device type (2x6 or 3x6) and needs to be updated based
1414 * on which subslice is disabled especially for 2x6
1415 * devices, however it is safe to load default
1416 * configuration of 3x6 device instead of masking off
1417 * corresponding bits because HW ignores bits of a disabled
1418 * subslice and drops down to appropriate config. Please
1419 * see render_state_setup() in i915_gem_render_state.c for
1420 * possible configurations, to avoid duplication they are
1421 * not shown here again.
1422 */
1423 *batch++ = GEN9_MEDIA_POOL_STATE;
1424 *batch++ = GEN9_MEDIA_POOL_ENABLE;
1425 *batch++ = 0x00777000;
1426 *batch++ = 0;
1427 *batch++ = 0;
1428 *batch++ = 0;
1429 }
1430
1431 *batch++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
1432
1433 /* Pad to end of cacheline */
1434 while ((unsigned long)batch % CACHELINE_BYTES)
1435 *batch++ = MI_NOOP;
1436
1437 return batch;
1438}
1439
1440static u32 *
1441gen10_init_indirectctx_bb(struct intel_engine_cs *engine, u32 *batch)
1442{
1443 int i;
1444
1445 /*
1446 * WaPipeControlBefore3DStateSamplePattern: cnl
1447 *
1448 * Ensure the engine is idle prior to programming a
1449 * 3DSTATE_SAMPLE_PATTERN during a context restore.
1450 */
1451 batch = gen8_emit_pipe_control(batch,
1452 PIPE_CONTROL_CS_STALL,
1453 0);
1454 /*
1455 * WaPipeControlBefore3DStateSamplePattern says we need 4 dwords for
1456 * the PIPE_CONTROL followed by 12 dwords of 0x0, so 16 dwords in
1457 * total. However, a PIPE_CONTROL is 6 dwords long, not 4, which is
1458 * confusing. Since gen8_emit_pipe_control() already advances the
1459 * batch by 6 dwords, we advance the other 10 here, completing a
1460 * cacheline. It's not clear if the workaround requires this padding
1461 * before other commands, or if it's just the regular padding we would
1462 * already have for the workaround bb, so leave it here for now.
1463 */
1464 for (i = 0; i < 10; i++)
1465 *batch++ = MI_NOOP;
1466
1467 /* Pad to end of cacheline */
1468 while ((unsigned long)batch % CACHELINE_BYTES)
1469 *batch++ = MI_NOOP;
1470
1471 return batch;
1472}
1473
1474#define CTX_WA_BB_OBJ_SIZE (PAGE_SIZE)
1475
1476static int lrc_setup_wa_ctx(struct intel_engine_cs *engine)
1477{
1478 struct drm_i915_gem_object *obj;
1479 struct i915_vma *vma;
1480 int err;
1481
1482 obj = i915_gem_object_create(engine->i915, CTX_WA_BB_OBJ_SIZE);
1483 if (IS_ERR(obj))
1484 return PTR_ERR(obj);
1485
1486 vma = i915_vma_instance(obj, &engine->i915->ggtt.base, NULL);
1487 if (IS_ERR(vma)) {
1488 err = PTR_ERR(vma);
1489 goto err;
1490 }
1491
1492 err = i915_vma_pin(vma, 0, PAGE_SIZE, PIN_GLOBAL | PIN_HIGH);
1493 if (err)
1494 goto err;
1495
1496 engine->wa_ctx.vma = vma;
1497 return 0;
1498
1499err:
1500 i915_gem_object_put(obj);
1501 return err;
1502}
1503
1504static void lrc_destroy_wa_ctx(struct intel_engine_cs *engine)
1505{
1506 i915_vma_unpin_and_release(&engine->wa_ctx.vma);
1507}
1508
1509typedef u32 *(*wa_bb_func_t)(struct intel_engine_cs *engine, u32 *batch);
1510
1511static int intel_init_workaround_bb(struct intel_engine_cs *engine)
1512{
1513 struct i915_ctx_workarounds *wa_ctx = &engine->wa_ctx;
1514 struct i915_wa_ctx_bb *wa_bb[2] = { &wa_ctx->indirect_ctx,
1515 &wa_ctx->per_ctx };
1516 wa_bb_func_t wa_bb_fn[2];
1517 struct page *page;
1518 void *batch, *batch_ptr;
1519 unsigned int i;
1520 int ret;
1521
1522 if (GEM_WARN_ON(engine->id != RCS))
1523 return -EINVAL;
1524
1525 switch (INTEL_GEN(engine->i915)) {
1526 case 10:
1527 wa_bb_fn[0] = gen10_init_indirectctx_bb;
1528 wa_bb_fn[1] = NULL;
1529 break;
1530 case 9:
1531 wa_bb_fn[0] = gen9_init_indirectctx_bb;
1532 wa_bb_fn[1] = NULL;
1533 break;
1534 case 8:
1535 wa_bb_fn[0] = gen8_init_indirectctx_bb;
1536 wa_bb_fn[1] = NULL;
1537 break;
1538 default:
1539 MISSING_CASE(INTEL_GEN(engine->i915));
1540 return 0;
1541 }
1542
1543 ret = lrc_setup_wa_ctx(engine);
1544 if (ret) {
1545 DRM_DEBUG_DRIVER("Failed to setup context WA page: %d\n", ret);
1546 return ret;
1547 }
1548
1549 page = i915_gem_object_get_dirty_page(wa_ctx->vma->obj, 0);
1550 batch = batch_ptr = kmap_atomic(page);
1551
1552 /*
1553 * Emit the two workaround batch buffers, recording the offset from the
1554 * start of the workaround batch buffer object for each and their
1555 * respective sizes.
1556 */
1557 for (i = 0; i < ARRAY_SIZE(wa_bb_fn); i++) {
1558 wa_bb[i]->offset = batch_ptr - batch;
1559 if (GEM_WARN_ON(!IS_ALIGNED(wa_bb[i]->offset,
1560 CACHELINE_BYTES))) {
1561 ret = -EINVAL;
1562 break;
1563 }
1564 if (wa_bb_fn[i])
1565 batch_ptr = wa_bb_fn[i](engine, batch_ptr);
1566 wa_bb[i]->size = batch_ptr - (batch + wa_bb[i]->offset);
1567 }
1568
1569 BUG_ON(batch_ptr - batch > CTX_WA_BB_OBJ_SIZE);
1570
1571 kunmap_atomic(batch);
1572 if (ret)
1573 lrc_destroy_wa_ctx(engine);
1574
1575 return ret;
1576}
1577
1578static u8 gtiir[] = {
1579 [RCS] = 0,
1580 [BCS] = 0,
1581 [VCS] = 1,
1582 [VCS2] = 1,
1583 [VECS] = 3,
1584};
1585
1586static void enable_execlists(struct intel_engine_cs *engine)
1587{
1588 struct drm_i915_private *dev_priv = engine->i915;
1589
1590 I915_WRITE(RING_HWSTAM(engine->mmio_base), 0xffffffff);
1591
1592 /*
1593 * Make sure we're not enabling the new 12-deep CSB
1594 * FIFO as that requires a slightly updated handling
1595 * in the ctx switch irq. Since we're currently only
1596 * using only 2 elements of the enhanced execlists the
1597 * deeper FIFO it's not needed and it's not worth adding
1598 * more statements to the irq handler to support it.
1599 */
1600 if (INTEL_GEN(dev_priv) >= 11)
1601 I915_WRITE(RING_MODE_GEN7(engine),
1602 _MASKED_BIT_DISABLE(GEN11_GFX_DISABLE_LEGACY_MODE));
1603 else
1604 I915_WRITE(RING_MODE_GEN7(engine),
1605 _MASKED_BIT_ENABLE(GFX_RUN_LIST_ENABLE));
1606
1607 I915_WRITE(RING_HWS_PGA(engine->mmio_base),
1608 engine->status_page.ggtt_offset);
1609 POSTING_READ(RING_HWS_PGA(engine->mmio_base));
1610
1611 /* Following the reset, we need to reload the CSB read/write pointers */
1612 engine->execlists.csb_head = -1;
1613}
1614
1615static int gen8_init_common_ring(struct intel_engine_cs *engine)
1616{
1617 struct intel_engine_execlists * const execlists = &engine->execlists;
1618 int ret;
1619
1620 ret = intel_mocs_init_engine(engine);
1621 if (ret)
1622 return ret;
1623
1624 intel_engine_reset_breadcrumbs(engine);
1625 intel_engine_init_hangcheck(engine);
1626
1627 enable_execlists(engine);
1628
1629 /* After a GPU reset, we may have requests to replay */
1630 if (execlists->first)
1631 tasklet_schedule(&execlists->tasklet);
1632
1633 return 0;
1634}
1635
1636static int gen8_init_render_ring(struct intel_engine_cs *engine)
1637{
1638 struct drm_i915_private *dev_priv = engine->i915;
1639 int ret;
1640
1641 ret = gen8_init_common_ring(engine);
1642 if (ret)
1643 return ret;
1644
1645 /* We need to disable the AsyncFlip performance optimisations in order
1646 * to use MI_WAIT_FOR_EVENT within the CS. It should already be
1647 * programmed to '1' on all products.
1648 *
1649 * WaDisableAsyncFlipPerfMode:snb,ivb,hsw,vlv,bdw,chv
1650 */
1651 I915_WRITE(MI_MODE, _MASKED_BIT_ENABLE(ASYNC_FLIP_PERF_DISABLE));
1652
1653 I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_FORCE_ORDERING));
1654
1655 return init_workarounds_ring(engine);
1656}
1657
1658static int gen9_init_render_ring(struct intel_engine_cs *engine)
1659{
1660 int ret;
1661
1662 ret = gen8_init_common_ring(engine);
1663 if (ret)
1664 return ret;
1665
1666 return init_workarounds_ring(engine);
1667}
1668
1669static void reset_irq(struct intel_engine_cs *engine)
1670{
1671 struct drm_i915_private *dev_priv = engine->i915;
1672 int i;
1673
1674 GEM_BUG_ON(engine->id >= ARRAY_SIZE(gtiir));
1675
1676 /*
1677 * Clear any pending interrupt state.
1678 *
1679 * We do it twice out of paranoia that some of the IIR are double
1680 * buffered, and if we only reset it once there may still be
1681 * an interrupt pending.
1682 */
1683 for (i = 0; i < 2; i++) {
1684 I915_WRITE(GEN8_GT_IIR(gtiir[engine->id]),
1685 GT_CONTEXT_SWITCH_INTERRUPT << engine->irq_shift);
1686 POSTING_READ(GEN8_GT_IIR(gtiir[engine->id]));
1687 }
1688 GEM_BUG_ON(I915_READ(GEN8_GT_IIR(gtiir[engine->id])) &
1689 (GT_CONTEXT_SWITCH_INTERRUPT << engine->irq_shift));
1690
1691 clear_bit(ENGINE_IRQ_EXECLIST, &engine->irq_posted);
1692}
1693
1694static void reset_common_ring(struct intel_engine_cs *engine,
1695 struct i915_request *request)
1696{
1697 struct intel_engine_execlists * const execlists = &engine->execlists;
1698 struct intel_context *ce;
1699 unsigned long flags;
1700
1701 GEM_TRACE("%s seqno=%x\n",
1702 engine->name, request ? request->global_seqno : 0);
1703
1704 /* See execlists_cancel_requests() for the irq/spinlock split. */
1705 local_irq_save(flags);
1706
1707 reset_irq(engine);
1708
1709 /*
1710 * Catch up with any missed context-switch interrupts.
1711 *
1712 * Ideally we would just read the remaining CSB entries now that we
1713 * know the gpu is idle. However, the CSB registers are sometimes^W
1714 * often trashed across a GPU reset! Instead we have to rely on
1715 * guessing the missed context-switch events by looking at what
1716 * requests were completed.
1717 */
1718 execlists_cancel_port_requests(execlists);
1719
1720 /* Push back any incomplete requests for replay after the reset. */
1721 spin_lock(&engine->timeline->lock);
1722 __unwind_incomplete_requests(engine);
1723 spin_unlock(&engine->timeline->lock);
1724
1725 /* Mark all CS interrupts as complete */
1726 execlists->active = 0;
1727
1728 local_irq_restore(flags);
1729
1730 /*
1731 * If the request was innocent, we leave the request in the ELSP
1732 * and will try to replay it on restarting. The context image may
1733 * have been corrupted by the reset, in which case we may have
1734 * to service a new GPU hang, but more likely we can continue on
1735 * without impact.
1736 *
1737 * If the request was guilty, we presume the context is corrupt
1738 * and have to at least restore the RING register in the context
1739 * image back to the expected values to skip over the guilty request.
1740 */
1741 if (!request || request->fence.error != -EIO)
1742 return;
1743
1744 /*
1745 * We want a simple context + ring to execute the breadcrumb update.
1746 * We cannot rely on the context being intact across the GPU hang,
1747 * so clear it and rebuild just what we need for the breadcrumb.
1748 * All pending requests for this context will be zapped, and any
1749 * future request will be after userspace has had the opportunity
1750 * to recreate its own state.
1751 */
1752 ce = &request->ctx->engine[engine->id];
1753 execlists_init_reg_state(ce->lrc_reg_state,
1754 request->ctx, engine, ce->ring);
1755
1756 /* Move the RING_HEAD onto the breadcrumb, past the hanging batch */
1757 ce->lrc_reg_state[CTX_RING_BUFFER_START+1] =
1758 i915_ggtt_offset(ce->ring->vma);
1759 ce->lrc_reg_state[CTX_RING_HEAD+1] = request->postfix;
1760
1761 request->ring->head = request->postfix;
1762 intel_ring_update_space(request->ring);
1763
1764 /* Reset WaIdleLiteRestore:bdw,skl as well */
1765 unwind_wa_tail(request);
1766}
1767
1768static int intel_logical_ring_emit_pdps(struct i915_request *rq)
1769{
1770 struct i915_hw_ppgtt *ppgtt = rq->ctx->ppgtt;
1771 struct intel_engine_cs *engine = rq->engine;
1772 const int num_lri_cmds = GEN8_3LVL_PDPES * 2;
1773 u32 *cs;
1774 int i;
1775
1776 cs = intel_ring_begin(rq, num_lri_cmds * 2 + 2);
1777 if (IS_ERR(cs))
1778 return PTR_ERR(cs);
1779
1780 *cs++ = MI_LOAD_REGISTER_IMM(num_lri_cmds);
1781 for (i = GEN8_3LVL_PDPES - 1; i >= 0; i--) {
1782 const dma_addr_t pd_daddr = i915_page_dir_dma_addr(ppgtt, i);
1783
1784 *cs++ = i915_mmio_reg_offset(GEN8_RING_PDP_UDW(engine, i));
1785 *cs++ = upper_32_bits(pd_daddr);
1786 *cs++ = i915_mmio_reg_offset(GEN8_RING_PDP_LDW(engine, i));
1787 *cs++ = lower_32_bits(pd_daddr);
1788 }
1789
1790 *cs++ = MI_NOOP;
1791 intel_ring_advance(rq, cs);
1792
1793 return 0;
1794}
1795
1796static int gen8_emit_bb_start(struct i915_request *rq,
1797 u64 offset, u32 len,
1798 const unsigned int flags)
1799{
1800 u32 *cs;
1801 int ret;
1802
1803 /* Don't rely in hw updating PDPs, specially in lite-restore.
1804 * Ideally, we should set Force PD Restore in ctx descriptor,
1805 * but we can't. Force Restore would be a second option, but
1806 * it is unsafe in case of lite-restore (because the ctx is
1807 * not idle). PML4 is allocated during ppgtt init so this is
1808 * not needed in 48-bit.*/
1809 if (rq->ctx->ppgtt &&
1810 (intel_engine_flag(rq->engine) & rq->ctx->ppgtt->pd_dirty_rings) &&
1811 !i915_vm_is_48bit(&rq->ctx->ppgtt->base) &&
1812 !intel_vgpu_active(rq->i915)) {
1813 ret = intel_logical_ring_emit_pdps(rq);
1814 if (ret)
1815 return ret;
1816
1817 rq->ctx->ppgtt->pd_dirty_rings &= ~intel_engine_flag(rq->engine);
1818 }
1819
1820 cs = intel_ring_begin(rq, 4);
1821 if (IS_ERR(cs))
1822 return PTR_ERR(cs);
1823
1824 /*
1825 * WaDisableCtxRestoreArbitration:bdw,chv
1826 *
1827 * We don't need to perform MI_ARB_ENABLE as often as we do (in
1828 * particular all the gen that do not need the w/a at all!), if we
1829 * took care to make sure that on every switch into this context
1830 * (both ordinary and for preemption) that arbitrartion was enabled
1831 * we would be fine. However, there doesn't seem to be a downside to
1832 * being paranoid and making sure it is set before each batch and
1833 * every context-switch.
1834 *
1835 * Note that if we fail to enable arbitration before the request
1836 * is complete, then we do not see the context-switch interrupt and
1837 * the engine hangs (with RING_HEAD == RING_TAIL).
1838 *
1839 * That satisfies both the GPGPU w/a and our heavy-handed paranoia.
1840 */
1841 *cs++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
1842
1843 /* FIXME(BDW): Address space and security selectors. */
1844 *cs++ = MI_BATCH_BUFFER_START_GEN8 |
1845 (flags & I915_DISPATCH_SECURE ? 0 : BIT(8)) |
1846 (flags & I915_DISPATCH_RS ? MI_BATCH_RESOURCE_STREAMER : 0);
1847 *cs++ = lower_32_bits(offset);
1848 *cs++ = upper_32_bits(offset);
1849 intel_ring_advance(rq, cs);
1850
1851 return 0;
1852}
1853
1854static void gen8_logical_ring_enable_irq(struct intel_engine_cs *engine)
1855{
1856 struct drm_i915_private *dev_priv = engine->i915;
1857 I915_WRITE_IMR(engine,
1858 ~(engine->irq_enable_mask | engine->irq_keep_mask));
1859 POSTING_READ_FW(RING_IMR(engine->mmio_base));
1860}
1861
1862static void gen8_logical_ring_disable_irq(struct intel_engine_cs *engine)
1863{
1864 struct drm_i915_private *dev_priv = engine->i915;
1865 I915_WRITE_IMR(engine, ~engine->irq_keep_mask);
1866}
1867
1868static int gen8_emit_flush(struct i915_request *request, u32 mode)
1869{
1870 u32 cmd, *cs;
1871
1872 cs = intel_ring_begin(request, 4);
1873 if (IS_ERR(cs))
1874 return PTR_ERR(cs);
1875
1876 cmd = MI_FLUSH_DW + 1;
1877
1878 /* We always require a command barrier so that subsequent
1879 * commands, such as breadcrumb interrupts, are strictly ordered
1880 * wrt the contents of the write cache being flushed to memory
1881 * (and thus being coherent from the CPU).
1882 */
1883 cmd |= MI_FLUSH_DW_STORE_INDEX | MI_FLUSH_DW_OP_STOREDW;
1884
1885 if (mode & EMIT_INVALIDATE) {
1886 cmd |= MI_INVALIDATE_TLB;
1887 if (request->engine->id == VCS)
1888 cmd |= MI_INVALIDATE_BSD;
1889 }
1890
1891 *cs++ = cmd;
1892 *cs++ = I915_GEM_HWS_SCRATCH_ADDR | MI_FLUSH_DW_USE_GTT;
1893 *cs++ = 0; /* upper addr */
1894 *cs++ = 0; /* value */
1895 intel_ring_advance(request, cs);
1896
1897 return 0;
1898}
1899
1900static int gen8_emit_flush_render(struct i915_request *request,
1901 u32 mode)
1902{
1903 struct intel_engine_cs *engine = request->engine;
1904 u32 scratch_addr =
1905 i915_ggtt_offset(engine->scratch) + 2 * CACHELINE_BYTES;
1906 bool vf_flush_wa = false, dc_flush_wa = false;
1907 u32 *cs, flags = 0;
1908 int len;
1909
1910 flags |= PIPE_CONTROL_CS_STALL;
1911
1912 if (mode & EMIT_FLUSH) {
1913 flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
1914 flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
1915 flags |= PIPE_CONTROL_DC_FLUSH_ENABLE;
1916 flags |= PIPE_CONTROL_FLUSH_ENABLE;
1917 }
1918
1919 if (mode & EMIT_INVALIDATE) {
1920 flags |= PIPE_CONTROL_TLB_INVALIDATE;
1921 flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
1922 flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
1923 flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
1924 flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
1925 flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
1926 flags |= PIPE_CONTROL_QW_WRITE;
1927 flags |= PIPE_CONTROL_GLOBAL_GTT_IVB;
1928
1929 /*
1930 * On GEN9: before VF_CACHE_INVALIDATE we need to emit a NULL
1931 * pipe control.
1932 */
1933 if (IS_GEN9(request->i915))
1934 vf_flush_wa = true;
1935
1936 /* WaForGAMHang:kbl */
1937 if (IS_KBL_REVID(request->i915, 0, KBL_REVID_B0))
1938 dc_flush_wa = true;
1939 }
1940
1941 len = 6;
1942
1943 if (vf_flush_wa)
1944 len += 6;
1945
1946 if (dc_flush_wa)
1947 len += 12;
1948
1949 cs = intel_ring_begin(request, len);
1950 if (IS_ERR(cs))
1951 return PTR_ERR(cs);
1952
1953 if (vf_flush_wa)
1954 cs = gen8_emit_pipe_control(cs, 0, 0);
1955
1956 if (dc_flush_wa)
1957 cs = gen8_emit_pipe_control(cs, PIPE_CONTROL_DC_FLUSH_ENABLE,
1958 0);
1959
1960 cs = gen8_emit_pipe_control(cs, flags, scratch_addr);
1961
1962 if (dc_flush_wa)
1963 cs = gen8_emit_pipe_control(cs, PIPE_CONTROL_CS_STALL, 0);
1964
1965 intel_ring_advance(request, cs);
1966
1967 return 0;
1968}
1969
1970/*
1971 * Reserve space for 2 NOOPs at the end of each request to be
1972 * used as a workaround for not being allowed to do lite
1973 * restore with HEAD==TAIL (WaIdleLiteRestore).
1974 */
1975static void gen8_emit_wa_tail(struct i915_request *request, u32 *cs)
1976{
1977 /* Ensure there's always at least one preemption point per-request. */
1978 *cs++ = MI_ARB_CHECK;
1979 *cs++ = MI_NOOP;
1980 request->wa_tail = intel_ring_offset(request, cs);
1981}
1982
1983static void gen8_emit_breadcrumb(struct i915_request *request, u32 *cs)
1984{
1985 /* w/a: bit 5 needs to be zero for MI_FLUSH_DW address. */
1986 BUILD_BUG_ON(I915_GEM_HWS_INDEX_ADDR & (1 << 5));
1987
1988 cs = gen8_emit_ggtt_write(cs, request->global_seqno,
1989 intel_hws_seqno_address(request->engine));
1990 *cs++ = MI_USER_INTERRUPT;
1991 *cs++ = MI_NOOP;
1992 request->tail = intel_ring_offset(request, cs);
1993 assert_ring_tail_valid(request->ring, request->tail);
1994
1995 gen8_emit_wa_tail(request, cs);
1996}
1997static const int gen8_emit_breadcrumb_sz = 6 + WA_TAIL_DWORDS;
1998
1999static void gen8_emit_breadcrumb_rcs(struct i915_request *request, u32 *cs)
2000{
2001 /* We're using qword write, seqno should be aligned to 8 bytes. */
2002 BUILD_BUG_ON(I915_GEM_HWS_INDEX & 1);
2003
2004 cs = gen8_emit_ggtt_write_rcs(cs, request->global_seqno,
2005 intel_hws_seqno_address(request->engine));
2006 *cs++ = MI_USER_INTERRUPT;
2007 *cs++ = MI_NOOP;
2008 request->tail = intel_ring_offset(request, cs);
2009 assert_ring_tail_valid(request->ring, request->tail);
2010
2011 gen8_emit_wa_tail(request, cs);
2012}
2013static const int gen8_emit_breadcrumb_rcs_sz = 8 + WA_TAIL_DWORDS;
2014
2015static int gen8_init_rcs_context(struct i915_request *rq)
2016{
2017 int ret;
2018
2019 ret = intel_ring_workarounds_emit(rq);
2020 if (ret)
2021 return ret;
2022
2023 ret = intel_rcs_context_init_mocs(rq);
2024 /*
2025 * Failing to program the MOCS is non-fatal.The system will not
2026 * run at peak performance. So generate an error and carry on.
2027 */
2028 if (ret)
2029 DRM_ERROR("MOCS failed to program: expect performance issues.\n");
2030
2031 return i915_gem_render_state_emit(rq);
2032}
2033
2034/**
2035 * intel_logical_ring_cleanup() - deallocate the Engine Command Streamer
2036 * @engine: Engine Command Streamer.
2037 */
2038void intel_logical_ring_cleanup(struct intel_engine_cs *engine)
2039{
2040 struct drm_i915_private *dev_priv;
2041
2042 /*
2043 * Tasklet cannot be active at this point due intel_mark_active/idle
2044 * so this is just for documentation.
2045 */
2046 if (WARN_ON(test_bit(TASKLET_STATE_SCHED,
2047 &engine->execlists.tasklet.state)))
2048 tasklet_kill(&engine->execlists.tasklet);
2049
2050 dev_priv = engine->i915;
2051
2052 if (engine->buffer) {
2053 WARN_ON((I915_READ_MODE(engine) & MODE_IDLE) == 0);
2054 }
2055
2056 if (engine->cleanup)
2057 engine->cleanup(engine);
2058
2059 intel_engine_cleanup_common(engine);
2060
2061 lrc_destroy_wa_ctx(engine);
2062
2063 engine->i915 = NULL;
2064 dev_priv->engine[engine->id] = NULL;
2065 kfree(engine);
2066}
2067
2068static void execlists_set_default_submission(struct intel_engine_cs *engine)
2069{
2070 engine->submit_request = execlists_submit_request;
2071 engine->cancel_requests = execlists_cancel_requests;
2072 engine->schedule = execlists_schedule;
2073 engine->execlists.tasklet.func = execlists_submission_tasklet;
2074
2075 engine->park = NULL;
2076 engine->unpark = NULL;
2077
2078 engine->flags |= I915_ENGINE_SUPPORTS_STATS;
2079
2080 engine->i915->caps.scheduler =
2081 I915_SCHEDULER_CAP_ENABLED |
2082 I915_SCHEDULER_CAP_PRIORITY;
2083 if (engine->i915->preempt_context)
2084 engine->i915->caps.scheduler |= I915_SCHEDULER_CAP_PREEMPTION;
2085}
2086
2087static void
2088logical_ring_default_vfuncs(struct intel_engine_cs *engine)
2089{
2090 /* Default vfuncs which can be overriden by each engine. */
2091 engine->init_hw = gen8_init_common_ring;
2092 engine->reset_hw = reset_common_ring;
2093
2094 engine->context_pin = execlists_context_pin;
2095 engine->context_unpin = execlists_context_unpin;
2096
2097 engine->request_alloc = execlists_request_alloc;
2098
2099 engine->emit_flush = gen8_emit_flush;
2100 engine->emit_breadcrumb = gen8_emit_breadcrumb;
2101 engine->emit_breadcrumb_sz = gen8_emit_breadcrumb_sz;
2102
2103 engine->set_default_submission = execlists_set_default_submission;
2104
2105 if (INTEL_GEN(engine->i915) < 11) {
2106 engine->irq_enable = gen8_logical_ring_enable_irq;
2107 engine->irq_disable = gen8_logical_ring_disable_irq;
2108 } else {
2109 /*
2110 * TODO: On Gen11 interrupt masks need to be clear
2111 * to allow C6 entry. Keep interrupts enabled at
2112 * and take the hit of generating extra interrupts
2113 * until a more refined solution exists.
2114 */
2115 }
2116 engine->emit_bb_start = gen8_emit_bb_start;
2117}
2118
2119static inline void
2120logical_ring_default_irqs(struct intel_engine_cs *engine)
2121{
2122 unsigned shift = engine->irq_shift;
2123 engine->irq_enable_mask = GT_RENDER_USER_INTERRUPT << shift;
2124 engine->irq_keep_mask = GT_CONTEXT_SWITCH_INTERRUPT << shift;
2125}
2126
2127static void
2128logical_ring_setup(struct intel_engine_cs *engine)
2129{
2130 struct drm_i915_private *dev_priv = engine->i915;
2131 enum forcewake_domains fw_domains;
2132
2133 intel_engine_setup_common(engine);
2134
2135 /* Intentionally left blank. */
2136 engine->buffer = NULL;
2137
2138 fw_domains = intel_uncore_forcewake_for_reg(dev_priv,
2139 RING_ELSP(engine),
2140 FW_REG_WRITE);
2141
2142 fw_domains |= intel_uncore_forcewake_for_reg(dev_priv,
2143 RING_CONTEXT_STATUS_PTR(engine),
2144 FW_REG_READ | FW_REG_WRITE);
2145
2146 fw_domains |= intel_uncore_forcewake_for_reg(dev_priv,
2147 RING_CONTEXT_STATUS_BUF_BASE(engine),
2148 FW_REG_READ);
2149
2150 engine->execlists.fw_domains = fw_domains;
2151
2152 tasklet_init(&engine->execlists.tasklet,
2153 execlists_submission_tasklet, (unsigned long)engine);
2154
2155 logical_ring_default_vfuncs(engine);
2156 logical_ring_default_irqs(engine);
2157}
2158
2159static int logical_ring_init(struct intel_engine_cs *engine)
2160{
2161 int ret;
2162
2163 ret = intel_engine_init_common(engine);
2164 if (ret)
2165 goto error;
2166
2167 if (HAS_LOGICAL_RING_ELSQ(engine->i915)) {
2168 engine->execlists.submit_reg = engine->i915->regs +
2169 i915_mmio_reg_offset(RING_EXECLIST_SQ_CONTENTS(engine));
2170 engine->execlists.ctrl_reg = engine->i915->regs +
2171 i915_mmio_reg_offset(RING_EXECLIST_CONTROL(engine));
2172 } else {
2173 engine->execlists.submit_reg = engine->i915->regs +
2174 i915_mmio_reg_offset(RING_ELSP(engine));
2175 }
2176
2177 engine->execlists.preempt_complete_status = ~0u;
2178 if (engine->i915->preempt_context)
2179 engine->execlists.preempt_complete_status =
2180 upper_32_bits(engine->i915->preempt_context->engine[engine->id].lrc_desc);
2181
2182 return 0;
2183
2184error:
2185 intel_logical_ring_cleanup(engine);
2186 return ret;
2187}
2188
2189int logical_render_ring_init(struct intel_engine_cs *engine)
2190{
2191 struct drm_i915_private *dev_priv = engine->i915;
2192 int ret;
2193
2194 logical_ring_setup(engine);
2195
2196 if (HAS_L3_DPF(dev_priv))
2197 engine->irq_keep_mask |= GT_RENDER_L3_PARITY_ERROR_INTERRUPT;
2198
2199 /* Override some for render ring. */
2200 if (INTEL_GEN(dev_priv) >= 9)
2201 engine->init_hw = gen9_init_render_ring;
2202 else
2203 engine->init_hw = gen8_init_render_ring;
2204 engine->init_context = gen8_init_rcs_context;
2205 engine->emit_flush = gen8_emit_flush_render;
2206 engine->emit_breadcrumb = gen8_emit_breadcrumb_rcs;
2207 engine->emit_breadcrumb_sz = gen8_emit_breadcrumb_rcs_sz;
2208
2209 ret = intel_engine_create_scratch(engine, PAGE_SIZE);
2210 if (ret)
2211 return ret;
2212
2213 ret = intel_init_workaround_bb(engine);
2214 if (ret) {
2215 /*
2216 * We continue even if we fail to initialize WA batch
2217 * because we only expect rare glitches but nothing
2218 * critical to prevent us from using GPU
2219 */
2220 DRM_ERROR("WA batch buffer initialization failed: %d\n",
2221 ret);
2222 }
2223
2224 return logical_ring_init(engine);
2225}
2226
2227int logical_xcs_ring_init(struct intel_engine_cs *engine)
2228{
2229 logical_ring_setup(engine);
2230
2231 return logical_ring_init(engine);
2232}
2233
2234static u32
2235make_rpcs(struct drm_i915_private *dev_priv)
2236{
2237 u32 rpcs = 0;
2238
2239 /*
2240 * No explicit RPCS request is needed to ensure full
2241 * slice/subslice/EU enablement prior to Gen9.
2242 */
2243 if (INTEL_GEN(dev_priv) < 9)
2244 return 0;
2245
2246 /*
2247 * Starting in Gen9, render power gating can leave
2248 * slice/subslice/EU in a partially enabled state. We
2249 * must make an explicit request through RPCS for full
2250 * enablement.
2251 */
2252 if (INTEL_INFO(dev_priv)->sseu.has_slice_pg) {
2253 rpcs |= GEN8_RPCS_S_CNT_ENABLE;
2254 rpcs |= hweight8(INTEL_INFO(dev_priv)->sseu.slice_mask) <<
2255 GEN8_RPCS_S_CNT_SHIFT;
2256 rpcs |= GEN8_RPCS_ENABLE;
2257 }
2258
2259 if (INTEL_INFO(dev_priv)->sseu.has_subslice_pg) {
2260 rpcs |= GEN8_RPCS_SS_CNT_ENABLE;
2261 rpcs |= hweight8(INTEL_INFO(dev_priv)->sseu.subslice_mask[0]) <<
2262 GEN8_RPCS_SS_CNT_SHIFT;
2263 rpcs |= GEN8_RPCS_ENABLE;
2264 }
2265
2266 if (INTEL_INFO(dev_priv)->sseu.has_eu_pg) {
2267 rpcs |= INTEL_INFO(dev_priv)->sseu.eu_per_subslice <<
2268 GEN8_RPCS_EU_MIN_SHIFT;
2269 rpcs |= INTEL_INFO(dev_priv)->sseu.eu_per_subslice <<
2270 GEN8_RPCS_EU_MAX_SHIFT;
2271 rpcs |= GEN8_RPCS_ENABLE;
2272 }
2273
2274 return rpcs;
2275}
2276
2277static u32 intel_lr_indirect_ctx_offset(struct intel_engine_cs *engine)
2278{
2279 u32 indirect_ctx_offset;
2280
2281 switch (INTEL_GEN(engine->i915)) {
2282 default:
2283 MISSING_CASE(INTEL_GEN(engine->i915));
2284 /* fall through */
2285 case 11:
2286 indirect_ctx_offset =
2287 GEN11_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT;
2288 break;
2289 case 10:
2290 indirect_ctx_offset =
2291 GEN10_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT;
2292 break;
2293 case 9:
2294 indirect_ctx_offset =
2295 GEN9_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT;
2296 break;
2297 case 8:
2298 indirect_ctx_offset =
2299 GEN8_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT;
2300 break;
2301 }
2302
2303 return indirect_ctx_offset;
2304}
2305
2306static void execlists_init_reg_state(u32 *regs,
2307 struct i915_gem_context *ctx,
2308 struct intel_engine_cs *engine,
2309 struct intel_ring *ring)
2310{
2311 struct drm_i915_private *dev_priv = engine->i915;
2312 struct i915_hw_ppgtt *ppgtt = ctx->ppgtt ?: dev_priv->mm.aliasing_ppgtt;
2313 u32 base = engine->mmio_base;
2314 bool rcs = engine->id == RCS;
2315
2316 /* A context is actually a big batch buffer with several
2317 * MI_LOAD_REGISTER_IMM commands followed by (reg, value) pairs. The
2318 * values we are setting here are only for the first context restore:
2319 * on a subsequent save, the GPU will recreate this batchbuffer with new
2320 * values (including all the missing MI_LOAD_REGISTER_IMM commands that
2321 * we are not initializing here).
2322 */
2323 regs[CTX_LRI_HEADER_0] = MI_LOAD_REGISTER_IMM(rcs ? 14 : 11) |
2324 MI_LRI_FORCE_POSTED;
2325
2326 CTX_REG(regs, CTX_CONTEXT_CONTROL, RING_CONTEXT_CONTROL(engine),
2327 _MASKED_BIT_DISABLE(CTX_CTRL_ENGINE_CTX_RESTORE_INHIBIT |
2328 CTX_CTRL_ENGINE_CTX_SAVE_INHIBIT) |
2329 _MASKED_BIT_ENABLE(CTX_CTRL_INHIBIT_SYN_CTX_SWITCH |
2330 (HAS_RESOURCE_STREAMER(dev_priv) ?
2331 CTX_CTRL_RS_CTX_ENABLE : 0)));
2332 CTX_REG(regs, CTX_RING_HEAD, RING_HEAD(base), 0);
2333 CTX_REG(regs, CTX_RING_TAIL, RING_TAIL(base), 0);
2334 CTX_REG(regs, CTX_RING_BUFFER_START, RING_START(base), 0);
2335 CTX_REG(regs, CTX_RING_BUFFER_CONTROL, RING_CTL(base),
2336 RING_CTL_SIZE(ring->size) | RING_VALID);
2337 CTX_REG(regs, CTX_BB_HEAD_U, RING_BBADDR_UDW(base), 0);
2338 CTX_REG(regs, CTX_BB_HEAD_L, RING_BBADDR(base), 0);
2339 CTX_REG(regs, CTX_BB_STATE, RING_BBSTATE(base), RING_BB_PPGTT);
2340 CTX_REG(regs, CTX_SECOND_BB_HEAD_U, RING_SBBADDR_UDW(base), 0);
2341 CTX_REG(regs, CTX_SECOND_BB_HEAD_L, RING_SBBADDR(base), 0);
2342 CTX_REG(regs, CTX_SECOND_BB_STATE, RING_SBBSTATE(base), 0);
2343 if (rcs) {
2344 struct i915_ctx_workarounds *wa_ctx = &engine->wa_ctx;
2345
2346 CTX_REG(regs, CTX_RCS_INDIRECT_CTX, RING_INDIRECT_CTX(base), 0);
2347 CTX_REG(regs, CTX_RCS_INDIRECT_CTX_OFFSET,
2348 RING_INDIRECT_CTX_OFFSET(base), 0);
2349 if (wa_ctx->indirect_ctx.size) {
2350 u32 ggtt_offset = i915_ggtt_offset(wa_ctx->vma);
2351
2352 regs[CTX_RCS_INDIRECT_CTX + 1] =
2353 (ggtt_offset + wa_ctx->indirect_ctx.offset) |
2354 (wa_ctx->indirect_ctx.size / CACHELINE_BYTES);
2355
2356 regs[CTX_RCS_INDIRECT_CTX_OFFSET + 1] =
2357 intel_lr_indirect_ctx_offset(engine) << 6;
2358 }
2359
2360 CTX_REG(regs, CTX_BB_PER_CTX_PTR, RING_BB_PER_CTX_PTR(base), 0);
2361 if (wa_ctx->per_ctx.size) {
2362 u32 ggtt_offset = i915_ggtt_offset(wa_ctx->vma);
2363
2364 regs[CTX_BB_PER_CTX_PTR + 1] =
2365 (ggtt_offset + wa_ctx->per_ctx.offset) | 0x01;
2366 }
2367 }
2368
2369 regs[CTX_LRI_HEADER_1] = MI_LOAD_REGISTER_IMM(9) | MI_LRI_FORCE_POSTED;
2370
2371 CTX_REG(regs, CTX_CTX_TIMESTAMP, RING_CTX_TIMESTAMP(base), 0);
2372 /* PDP values well be assigned later if needed */
2373 CTX_REG(regs, CTX_PDP3_UDW, GEN8_RING_PDP_UDW(engine, 3), 0);
2374 CTX_REG(regs, CTX_PDP3_LDW, GEN8_RING_PDP_LDW(engine, 3), 0);
2375 CTX_REG(regs, CTX_PDP2_UDW, GEN8_RING_PDP_UDW(engine, 2), 0);
2376 CTX_REG(regs, CTX_PDP2_LDW, GEN8_RING_PDP_LDW(engine, 2), 0);
2377 CTX_REG(regs, CTX_PDP1_UDW, GEN8_RING_PDP_UDW(engine, 1), 0);
2378 CTX_REG(regs, CTX_PDP1_LDW, GEN8_RING_PDP_LDW(engine, 1), 0);
2379 CTX_REG(regs, CTX_PDP0_UDW, GEN8_RING_PDP_UDW(engine, 0), 0);
2380 CTX_REG(regs, CTX_PDP0_LDW, GEN8_RING_PDP_LDW(engine, 0), 0);
2381
2382 if (ppgtt && i915_vm_is_48bit(&ppgtt->base)) {
2383 /* 64b PPGTT (48bit canonical)
2384 * PDP0_DESCRIPTOR contains the base address to PML4 and
2385 * other PDP Descriptors are ignored.
2386 */
2387 ASSIGN_CTX_PML4(ppgtt, regs);
2388 }
2389
2390 if (rcs) {
2391 regs[CTX_LRI_HEADER_2] = MI_LOAD_REGISTER_IMM(1);
2392 CTX_REG(regs, CTX_R_PWR_CLK_STATE, GEN8_R_PWR_CLK_STATE,
2393 make_rpcs(dev_priv));
2394
2395 i915_oa_init_reg_state(engine, ctx, regs);
2396 }
2397}
2398
2399static int
2400populate_lr_context(struct i915_gem_context *ctx,
2401 struct drm_i915_gem_object *ctx_obj,
2402 struct intel_engine_cs *engine,
2403 struct intel_ring *ring)
2404{
2405 void *vaddr;
2406 u32 *regs;
2407 int ret;
2408
2409 ret = i915_gem_object_set_to_cpu_domain(ctx_obj, true);
2410 if (ret) {
2411 DRM_DEBUG_DRIVER("Could not set to CPU domain\n");
2412 return ret;
2413 }
2414
2415 vaddr = i915_gem_object_pin_map(ctx_obj, I915_MAP_WB);
2416 if (IS_ERR(vaddr)) {
2417 ret = PTR_ERR(vaddr);
2418 DRM_DEBUG_DRIVER("Could not map object pages! (%d)\n", ret);
2419 return ret;
2420 }
2421 ctx_obj->mm.dirty = true;
2422
2423 if (engine->default_state) {
2424 /*
2425 * We only want to copy over the template context state;
2426 * skipping over the headers reserved for GuC communication,
2427 * leaving those as zero.
2428 */
2429 const unsigned long start = LRC_HEADER_PAGES * PAGE_SIZE;
2430 void *defaults;
2431
2432 defaults = i915_gem_object_pin_map(engine->default_state,
2433 I915_MAP_WB);
2434 if (IS_ERR(defaults))
2435 return PTR_ERR(defaults);
2436
2437 memcpy(vaddr + start, defaults + start, engine->context_size);
2438 i915_gem_object_unpin_map(engine->default_state);
2439 }
2440
2441 /* The second page of the context object contains some fields which must
2442 * be set up prior to the first execution. */
2443 regs = vaddr + LRC_STATE_PN * PAGE_SIZE;
2444 execlists_init_reg_state(regs, ctx, engine, ring);
2445 if (!engine->default_state)
2446 regs[CTX_CONTEXT_CONTROL + 1] |=
2447 _MASKED_BIT_ENABLE(CTX_CTRL_ENGINE_CTX_RESTORE_INHIBIT);
2448 if (ctx == ctx->i915->preempt_context && INTEL_GEN(engine->i915) < 11)
2449 regs[CTX_CONTEXT_CONTROL + 1] |=
2450 _MASKED_BIT_ENABLE(CTX_CTRL_ENGINE_CTX_RESTORE_INHIBIT |
2451 CTX_CTRL_ENGINE_CTX_SAVE_INHIBIT);
2452
2453 i915_gem_object_unpin_map(ctx_obj);
2454
2455 return 0;
2456}
2457
2458static int execlists_context_deferred_alloc(struct i915_gem_context *ctx,
2459 struct intel_engine_cs *engine)
2460{
2461 struct drm_i915_gem_object *ctx_obj;
2462 struct intel_context *ce = &ctx->engine[engine->id];
2463 struct i915_vma *vma;
2464 uint32_t context_size;
2465 struct intel_ring *ring;
2466 int ret;
2467
2468 if (ce->state)
2469 return 0;
2470
2471 context_size = round_up(engine->context_size, I915_GTT_PAGE_SIZE);
2472
2473 /*
2474 * Before the actual start of the context image, we insert a few pages
2475 * for our own use and for sharing with the GuC.
2476 */
2477 context_size += LRC_HEADER_PAGES * PAGE_SIZE;
2478
2479 ctx_obj = i915_gem_object_create(ctx->i915, context_size);
2480 if (IS_ERR(ctx_obj)) {
2481 DRM_DEBUG_DRIVER("Alloc LRC backing obj failed.\n");
2482 return PTR_ERR(ctx_obj);
2483 }
2484
2485 vma = i915_vma_instance(ctx_obj, &ctx->i915->ggtt.base, NULL);
2486 if (IS_ERR(vma)) {
2487 ret = PTR_ERR(vma);
2488 goto error_deref_obj;
2489 }
2490
2491 ring = intel_engine_create_ring(engine, ctx->ring_size);
2492 if (IS_ERR(ring)) {
2493 ret = PTR_ERR(ring);
2494 goto error_deref_obj;
2495 }
2496
2497 ret = populate_lr_context(ctx, ctx_obj, engine, ring);
2498 if (ret) {
2499 DRM_DEBUG_DRIVER("Failed to populate LRC: %d\n", ret);
2500 goto error_ring_free;
2501 }
2502
2503 ce->ring = ring;
2504 ce->state = vma;
2505
2506 return 0;
2507
2508error_ring_free:
2509 intel_ring_free(ring);
2510error_deref_obj:
2511 i915_gem_object_put(ctx_obj);
2512 return ret;
2513}
2514
2515void intel_lr_context_resume(struct drm_i915_private *dev_priv)
2516{
2517 struct intel_engine_cs *engine;
2518 struct i915_gem_context *ctx;
2519 enum intel_engine_id id;
2520
2521 /* Because we emit WA_TAIL_DWORDS there may be a disparity
2522 * between our bookkeeping in ce->ring->head and ce->ring->tail and
2523 * that stored in context. As we only write new commands from
2524 * ce->ring->tail onwards, everything before that is junk. If the GPU
2525 * starts reading from its RING_HEAD from the context, it may try to
2526 * execute that junk and die.
2527 *
2528 * So to avoid that we reset the context images upon resume. For
2529 * simplicity, we just zero everything out.
2530 */
2531 list_for_each_entry(ctx, &dev_priv->contexts.list, link) {
2532 for_each_engine(engine, dev_priv, id) {
2533 struct intel_context *ce = &ctx->engine[engine->id];
2534 u32 *reg;
2535
2536 if (!ce->state)
2537 continue;
2538
2539 reg = i915_gem_object_pin_map(ce->state->obj,
2540 I915_MAP_WB);
2541 if (WARN_ON(IS_ERR(reg)))
2542 continue;
2543
2544 reg += LRC_STATE_PN * PAGE_SIZE / sizeof(*reg);
2545 reg[CTX_RING_HEAD+1] = 0;
2546 reg[CTX_RING_TAIL+1] = 0;
2547
2548 ce->state->obj->mm.dirty = true;
2549 i915_gem_object_unpin_map(ce->state->obj);
2550
2551 intel_ring_reset(ce->ring, 0);
2552 }
2553 }
2554}