]> git.kernelconcepts.de Git - karo-tx-linux.git/blob - drivers/gpu/drm/i915/intel_lrc.c
Merge branch 'for-4.8/core' of git://git.kernel.dk/linux-block
[karo-tx-linux.git] / drivers / gpu / drm / i915 / intel_lrc.c
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 "intel_mocs.h"
140
141 #define GEN9_LR_CONTEXT_RENDER_SIZE (22 * PAGE_SIZE)
142 #define GEN8_LR_CONTEXT_RENDER_SIZE (20 * PAGE_SIZE)
143 #define GEN8_LR_CONTEXT_OTHER_SIZE (2 * PAGE_SIZE)
144
145 #define RING_EXECLIST_QFULL             (1 << 0x2)
146 #define RING_EXECLIST1_VALID            (1 << 0x3)
147 #define RING_EXECLIST0_VALID            (1 << 0x4)
148 #define RING_EXECLIST_ACTIVE_STATUS     (3 << 0xE)
149 #define RING_EXECLIST1_ACTIVE           (1 << 0x11)
150 #define RING_EXECLIST0_ACTIVE           (1 << 0x12)
151
152 #define GEN8_CTX_STATUS_IDLE_ACTIVE     (1 << 0)
153 #define GEN8_CTX_STATUS_PREEMPTED       (1 << 1)
154 #define GEN8_CTX_STATUS_ELEMENT_SWITCH  (1 << 2)
155 #define GEN8_CTX_STATUS_ACTIVE_IDLE     (1 << 3)
156 #define GEN8_CTX_STATUS_COMPLETE        (1 << 4)
157 #define GEN8_CTX_STATUS_LITE_RESTORE    (1 << 15)
158
159 #define CTX_LRI_HEADER_0                0x01
160 #define CTX_CONTEXT_CONTROL             0x02
161 #define CTX_RING_HEAD                   0x04
162 #define CTX_RING_TAIL                   0x06
163 #define CTX_RING_BUFFER_START           0x08
164 #define CTX_RING_BUFFER_CONTROL         0x0a
165 #define CTX_BB_HEAD_U                   0x0c
166 #define CTX_BB_HEAD_L                   0x0e
167 #define CTX_BB_STATE                    0x10
168 #define CTX_SECOND_BB_HEAD_U            0x12
169 #define CTX_SECOND_BB_HEAD_L            0x14
170 #define CTX_SECOND_BB_STATE             0x16
171 #define CTX_BB_PER_CTX_PTR              0x18
172 #define CTX_RCS_INDIRECT_CTX            0x1a
173 #define CTX_RCS_INDIRECT_CTX_OFFSET     0x1c
174 #define CTX_LRI_HEADER_1                0x21
175 #define CTX_CTX_TIMESTAMP               0x22
176 #define CTX_PDP3_UDW                    0x24
177 #define CTX_PDP3_LDW                    0x26
178 #define CTX_PDP2_UDW                    0x28
179 #define CTX_PDP2_LDW                    0x2a
180 #define CTX_PDP1_UDW                    0x2c
181 #define CTX_PDP1_LDW                    0x2e
182 #define CTX_PDP0_UDW                    0x30
183 #define CTX_PDP0_LDW                    0x32
184 #define CTX_LRI_HEADER_2                0x41
185 #define CTX_R_PWR_CLK_STATE             0x42
186 #define CTX_GPGPU_CSR_BASE_ADDRESS      0x44
187
188 #define GEN8_CTX_VALID (1<<0)
189 #define GEN8_CTX_FORCE_PD_RESTORE (1<<1)
190 #define GEN8_CTX_FORCE_RESTORE (1<<2)
191 #define GEN8_CTX_L3LLC_COHERENT (1<<5)
192 #define GEN8_CTX_PRIVILEGE (1<<8)
193
194 #define ASSIGN_CTX_REG(reg_state, pos, reg, val) do { \
195         (reg_state)[(pos)+0] = i915_mmio_reg_offset(reg); \
196         (reg_state)[(pos)+1] = (val); \
197 } while (0)
198
199 #define ASSIGN_CTX_PDP(ppgtt, reg_state, n) do {                \
200         const u64 _addr = i915_page_dir_dma_addr((ppgtt), (n)); \
201         reg_state[CTX_PDP ## n ## _UDW+1] = upper_32_bits(_addr); \
202         reg_state[CTX_PDP ## n ## _LDW+1] = lower_32_bits(_addr); \
203 } while (0)
204
205 #define ASSIGN_CTX_PML4(ppgtt, reg_state) do { \
206         reg_state[CTX_PDP0_UDW + 1] = upper_32_bits(px_dma(&ppgtt->pml4)); \
207         reg_state[CTX_PDP0_LDW + 1] = lower_32_bits(px_dma(&ppgtt->pml4)); \
208 } while (0)
209
210 enum {
211         ADVANCED_CONTEXT = 0,
212         LEGACY_32B_CONTEXT,
213         ADVANCED_AD_CONTEXT,
214         LEGACY_64B_CONTEXT
215 };
216 #define GEN8_CTX_ADDRESSING_MODE_SHIFT 3
217 #define GEN8_CTX_ADDRESSING_MODE(dev)  (USES_FULL_48BIT_PPGTT(dev) ?\
218                 LEGACY_64B_CONTEXT :\
219                 LEGACY_32B_CONTEXT)
220 enum {
221         FAULT_AND_HANG = 0,
222         FAULT_AND_HALT, /* Debug only */
223         FAULT_AND_STREAM,
224         FAULT_AND_CONTINUE /* Unsupported */
225 };
226 #define GEN8_CTX_ID_SHIFT 32
227 #define GEN8_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT        0x17
228 #define GEN9_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT        0x26
229
230 static int intel_lr_context_pin(struct intel_context *ctx,
231                                 struct intel_engine_cs *engine);
232
233 /**
234  * intel_sanitize_enable_execlists() - sanitize i915.enable_execlists
235  * @dev: DRM device.
236  * @enable_execlists: value of i915.enable_execlists module parameter.
237  *
238  * Only certain platforms support Execlists (the prerequisites being
239  * support for Logical Ring Contexts and Aliasing PPGTT or better).
240  *
241  * Return: 1 if Execlists is supported and has to be enabled.
242  */
243 int intel_sanitize_enable_execlists(struct drm_device *dev, int enable_execlists)
244 {
245         WARN_ON(i915.enable_ppgtt == -1);
246
247         /* On platforms with execlist available, vGPU will only
248          * support execlist mode, no ring buffer mode.
249          */
250         if (HAS_LOGICAL_RING_CONTEXTS(dev) && intel_vgpu_active(dev))
251                 return 1;
252
253         if (INTEL_INFO(dev)->gen >= 9)
254                 return 1;
255
256         if (enable_execlists == 0)
257                 return 0;
258
259         if (HAS_LOGICAL_RING_CONTEXTS(dev) && USES_PPGTT(dev) &&
260             i915.use_mmio_flip >= 0)
261                 return 1;
262
263         return 0;
264 }
265
266 static void
267 logical_ring_init_platform_invariants(struct intel_engine_cs *engine)
268 {
269         struct drm_device *dev = engine->dev;
270
271         if (IS_GEN8(dev) || IS_GEN9(dev))
272                 engine->idle_lite_restore_wa = ~0;
273
274         engine->disable_lite_restore_wa = (IS_SKL_REVID(dev, 0, SKL_REVID_B0) ||
275                                         IS_BXT_REVID(dev, 0, BXT_REVID_A1)) &&
276                                         (engine->id == VCS || engine->id == VCS2);
277
278         engine->ctx_desc_template = GEN8_CTX_VALID;
279         engine->ctx_desc_template |= GEN8_CTX_ADDRESSING_MODE(dev) <<
280                                    GEN8_CTX_ADDRESSING_MODE_SHIFT;
281         if (IS_GEN8(dev))
282                 engine->ctx_desc_template |= GEN8_CTX_L3LLC_COHERENT;
283         engine->ctx_desc_template |= GEN8_CTX_PRIVILEGE;
284
285         /* TODO: WaDisableLiteRestore when we start using semaphore
286          * signalling between Command Streamers */
287         /* ring->ctx_desc_template |= GEN8_CTX_FORCE_RESTORE; */
288
289         /* WaEnableForceRestoreInCtxtDescForVCS:skl */
290         /* WaEnableForceRestoreInCtxtDescForVCS:bxt */
291         if (engine->disable_lite_restore_wa)
292                 engine->ctx_desc_template |= GEN8_CTX_FORCE_RESTORE;
293 }
294
295 /**
296  * intel_lr_context_descriptor_update() - calculate & cache the descriptor
297  *                                        descriptor for a pinned context
298  *
299  * @ctx: Context to work on
300  * @ring: Engine the descriptor will be used with
301  *
302  * The context descriptor encodes various attributes of a context,
303  * including its GTT address and some flags. Because it's fairly
304  * expensive to calculate, we'll just do it once and cache the result,
305  * which remains valid until the context is unpinned.
306  *
307  * This is what a descriptor looks like, from LSB to MSB:
308  *    bits 0-11:    flags, GEN8_CTX_* (cached in ctx_desc_template)
309  *    bits 12-31:    LRCA, GTT address of (the HWSP of) this context
310  *    bits 32-51:    ctx ID, a globally unique tag (the LRCA again!)
311  *    bits 52-63:    reserved, may encode the engine ID (for GuC)
312  */
313 static void
314 intel_lr_context_descriptor_update(struct intel_context *ctx,
315                                    struct intel_engine_cs *engine)
316 {
317         uint64_t lrca, desc;
318
319         lrca = ctx->engine[engine->id].lrc_vma->node.start +
320                LRC_PPHWSP_PN * PAGE_SIZE;
321
322         desc = engine->ctx_desc_template;                          /* bits  0-11 */
323         desc |= lrca;                                      /* bits 12-31 */
324         desc |= (lrca >> PAGE_SHIFT) << GEN8_CTX_ID_SHIFT; /* bits 32-51 */
325
326         ctx->engine[engine->id].lrc_desc = desc;
327 }
328
329 uint64_t intel_lr_context_descriptor(struct intel_context *ctx,
330                                      struct intel_engine_cs *engine)
331 {
332         return ctx->engine[engine->id].lrc_desc;
333 }
334
335 /**
336  * intel_execlists_ctx_id() - get the Execlists Context ID
337  * @ctx: Context to get the ID for
338  * @ring: Engine to get the ID for
339  *
340  * Do not confuse with ctx->id! Unfortunately we have a name overload
341  * here: the old context ID we pass to userspace as a handler so that
342  * they can refer to a context, and the new context ID we pass to the
343  * ELSP so that the GPU can inform us of the context status via
344  * interrupts.
345  *
346  * The context ID is a portion of the context descriptor, so we can
347  * just extract the required part from the cached descriptor.
348  *
349  * Return: 20-bits globally unique context ID.
350  */
351 u32 intel_execlists_ctx_id(struct intel_context *ctx,
352                            struct intel_engine_cs *engine)
353 {
354         return intel_lr_context_descriptor(ctx, engine) >> GEN8_CTX_ID_SHIFT;
355 }
356
357 static void execlists_elsp_write(struct drm_i915_gem_request *rq0,
358                                  struct drm_i915_gem_request *rq1)
359 {
360
361         struct intel_engine_cs *engine = rq0->engine;
362         struct drm_device *dev = engine->dev;
363         struct drm_i915_private *dev_priv = dev->dev_private;
364         uint64_t desc[2];
365
366         if (rq1) {
367                 desc[1] = intel_lr_context_descriptor(rq1->ctx, rq1->engine);
368                 rq1->elsp_submitted++;
369         } else {
370                 desc[1] = 0;
371         }
372
373         desc[0] = intel_lr_context_descriptor(rq0->ctx, rq0->engine);
374         rq0->elsp_submitted++;
375
376         /* You must always write both descriptors in the order below. */
377         I915_WRITE_FW(RING_ELSP(engine), upper_32_bits(desc[1]));
378         I915_WRITE_FW(RING_ELSP(engine), lower_32_bits(desc[1]));
379
380         I915_WRITE_FW(RING_ELSP(engine), upper_32_bits(desc[0]));
381         /* The context is automatically loaded after the following */
382         I915_WRITE_FW(RING_ELSP(engine), lower_32_bits(desc[0]));
383
384         /* ELSP is a wo register, use another nearby reg for posting */
385         POSTING_READ_FW(RING_EXECLIST_STATUS_LO(engine));
386 }
387
388 static void
389 execlists_update_context_pdps(struct i915_hw_ppgtt *ppgtt, u32 *reg_state)
390 {
391         ASSIGN_CTX_PDP(ppgtt, reg_state, 3);
392         ASSIGN_CTX_PDP(ppgtt, reg_state, 2);
393         ASSIGN_CTX_PDP(ppgtt, reg_state, 1);
394         ASSIGN_CTX_PDP(ppgtt, reg_state, 0);
395 }
396
397 static void execlists_update_context(struct drm_i915_gem_request *rq)
398 {
399         struct intel_engine_cs *engine = rq->engine;
400         struct i915_hw_ppgtt *ppgtt = rq->ctx->ppgtt;
401         uint32_t *reg_state = rq->ctx->engine[engine->id].lrc_reg_state;
402
403         reg_state[CTX_RING_TAIL+1] = rq->tail;
404
405         /* True 32b PPGTT with dynamic page allocation: update PDP
406          * registers and point the unallocated PDPs to scratch page.
407          * PML4 is allocated during ppgtt init, so this is not needed
408          * in 48-bit mode.
409          */
410         if (ppgtt && !USES_FULL_48BIT_PPGTT(ppgtt->base.dev))
411                 execlists_update_context_pdps(ppgtt, reg_state);
412 }
413
414 static void execlists_submit_requests(struct drm_i915_gem_request *rq0,
415                                       struct drm_i915_gem_request *rq1)
416 {
417         struct drm_i915_private *dev_priv = rq0->i915;
418         unsigned int fw_domains = rq0->engine->fw_domains;
419
420         execlists_update_context(rq0);
421
422         if (rq1)
423                 execlists_update_context(rq1);
424
425         spin_lock_irq(&dev_priv->uncore.lock);
426         intel_uncore_forcewake_get__locked(dev_priv, fw_domains);
427
428         execlists_elsp_write(rq0, rq1);
429
430         intel_uncore_forcewake_put__locked(dev_priv, fw_domains);
431         spin_unlock_irq(&dev_priv->uncore.lock);
432 }
433
434 static void execlists_context_unqueue(struct intel_engine_cs *engine)
435 {
436         struct drm_i915_gem_request *req0 = NULL, *req1 = NULL;
437         struct drm_i915_gem_request *cursor, *tmp;
438
439         assert_spin_locked(&engine->execlist_lock);
440
441         /*
442          * If irqs are not active generate a warning as batches that finish
443          * without the irqs may get lost and a GPU Hang may occur.
444          */
445         WARN_ON(!intel_irqs_enabled(engine->dev->dev_private));
446
447         /* Try to read in pairs */
448         list_for_each_entry_safe(cursor, tmp, &engine->execlist_queue,
449                                  execlist_link) {
450                 if (!req0) {
451                         req0 = cursor;
452                 } else if (req0->ctx == cursor->ctx) {
453                         /* Same ctx: ignore first request, as second request
454                          * will update tail past first request's workload */
455                         cursor->elsp_submitted = req0->elsp_submitted;
456                         list_move_tail(&req0->execlist_link,
457                                        &engine->execlist_retired_req_list);
458                         req0 = cursor;
459                 } else {
460                         req1 = cursor;
461                         WARN_ON(req1->elsp_submitted);
462                         break;
463                 }
464         }
465
466         if (unlikely(!req0))
467                 return;
468
469         if (req0->elsp_submitted & engine->idle_lite_restore_wa) {
470                 /*
471                  * WaIdleLiteRestore: make sure we never cause a lite restore
472                  * with HEAD==TAIL.
473                  *
474                  * Apply the wa NOOPS to prevent ring:HEAD == req:TAIL as we
475                  * resubmit the request. See gen8_emit_request() for where we
476                  * prepare the padding after the end of the request.
477                  */
478                 struct intel_ringbuffer *ringbuf;
479
480                 ringbuf = req0->ctx->engine[engine->id].ringbuf;
481                 req0->tail += 8;
482                 req0->tail &= ringbuf->size - 1;
483         }
484
485         execlists_submit_requests(req0, req1);
486 }
487
488 static unsigned int
489 execlists_check_remove_request(struct intel_engine_cs *engine, u32 request_id)
490 {
491         struct drm_i915_gem_request *head_req;
492
493         assert_spin_locked(&engine->execlist_lock);
494
495         head_req = list_first_entry_or_null(&engine->execlist_queue,
496                                             struct drm_i915_gem_request,
497                                             execlist_link);
498
499         if (!head_req)
500                 return 0;
501
502         if (unlikely(intel_execlists_ctx_id(head_req->ctx, engine) != request_id))
503                 return 0;
504
505         WARN(head_req->elsp_submitted == 0, "Never submitted head request\n");
506
507         if (--head_req->elsp_submitted > 0)
508                 return 0;
509
510         list_move_tail(&head_req->execlist_link,
511                        &engine->execlist_retired_req_list);
512
513         return 1;
514 }
515
516 static u32
517 get_context_status(struct intel_engine_cs *engine, unsigned int read_pointer,
518                    u32 *context_id)
519 {
520         struct drm_i915_private *dev_priv = engine->dev->dev_private;
521         u32 status;
522
523         read_pointer %= GEN8_CSB_ENTRIES;
524
525         status = I915_READ_FW(RING_CONTEXT_STATUS_BUF_LO(engine, read_pointer));
526
527         if (status & GEN8_CTX_STATUS_IDLE_ACTIVE)
528                 return 0;
529
530         *context_id = I915_READ_FW(RING_CONTEXT_STATUS_BUF_HI(engine,
531                                                               read_pointer));
532
533         return status;
534 }
535
536 /**
537  * intel_lrc_irq_handler() - handle Context Switch interrupts
538  * @engine: Engine Command Streamer to handle.
539  *
540  * Check the unread Context Status Buffers and manage the submission of new
541  * contexts to the ELSP accordingly.
542  */
543 static void intel_lrc_irq_handler(unsigned long data)
544 {
545         struct intel_engine_cs *engine = (struct intel_engine_cs *)data;
546         struct drm_i915_private *dev_priv = engine->dev->dev_private;
547         u32 status_pointer;
548         unsigned int read_pointer, write_pointer;
549         u32 csb[GEN8_CSB_ENTRIES][2];
550         unsigned int csb_read = 0, i;
551         unsigned int submit_contexts = 0;
552
553         intel_uncore_forcewake_get(dev_priv, engine->fw_domains);
554
555         status_pointer = I915_READ_FW(RING_CONTEXT_STATUS_PTR(engine));
556
557         read_pointer = engine->next_context_status_buffer;
558         write_pointer = GEN8_CSB_WRITE_PTR(status_pointer);
559         if (read_pointer > write_pointer)
560                 write_pointer += GEN8_CSB_ENTRIES;
561
562         while (read_pointer < write_pointer) {
563                 if (WARN_ON_ONCE(csb_read == GEN8_CSB_ENTRIES))
564                         break;
565                 csb[csb_read][0] = get_context_status(engine, ++read_pointer,
566                                                       &csb[csb_read][1]);
567                 csb_read++;
568         }
569
570         engine->next_context_status_buffer = write_pointer % GEN8_CSB_ENTRIES;
571
572         /* Update the read pointer to the old write pointer. Manual ringbuffer
573          * management ftw </sarcasm> */
574         I915_WRITE_FW(RING_CONTEXT_STATUS_PTR(engine),
575                       _MASKED_FIELD(GEN8_CSB_READ_PTR_MASK,
576                                     engine->next_context_status_buffer << 8));
577
578         intel_uncore_forcewake_put(dev_priv, engine->fw_domains);
579
580         spin_lock(&engine->execlist_lock);
581
582         for (i = 0; i < csb_read; i++) {
583                 if (unlikely(csb[i][0] & GEN8_CTX_STATUS_PREEMPTED)) {
584                         if (csb[i][0] & GEN8_CTX_STATUS_LITE_RESTORE) {
585                                 if (execlists_check_remove_request(engine, csb[i][1]))
586                                         WARN(1, "Lite Restored request removed from queue\n");
587                         } else
588                                 WARN(1, "Preemption without Lite Restore\n");
589                 }
590
591                 if (csb[i][0] & (GEN8_CTX_STATUS_ACTIVE_IDLE |
592                     GEN8_CTX_STATUS_ELEMENT_SWITCH))
593                         submit_contexts +=
594                                 execlists_check_remove_request(engine, csb[i][1]);
595         }
596
597         if (submit_contexts) {
598                 if (!engine->disable_lite_restore_wa ||
599                     (csb[i][0] & GEN8_CTX_STATUS_ACTIVE_IDLE))
600                         execlists_context_unqueue(engine);
601         }
602
603         spin_unlock(&engine->execlist_lock);
604
605         if (unlikely(submit_contexts > 2))
606                 DRM_ERROR("More than two context complete events?\n");
607 }
608
609 static void execlists_context_queue(struct drm_i915_gem_request *request)
610 {
611         struct intel_engine_cs *engine = request->engine;
612         struct drm_i915_gem_request *cursor;
613         int num_elements = 0;
614
615         if (request->ctx != request->i915->kernel_context)
616                 intel_lr_context_pin(request->ctx, engine);
617
618         i915_gem_request_reference(request);
619
620         spin_lock_bh(&engine->execlist_lock);
621
622         list_for_each_entry(cursor, &engine->execlist_queue, execlist_link)
623                 if (++num_elements > 2)
624                         break;
625
626         if (num_elements > 2) {
627                 struct drm_i915_gem_request *tail_req;
628
629                 tail_req = list_last_entry(&engine->execlist_queue,
630                                            struct drm_i915_gem_request,
631                                            execlist_link);
632
633                 if (request->ctx == tail_req->ctx) {
634                         WARN(tail_req->elsp_submitted != 0,
635                                 "More than 2 already-submitted reqs queued\n");
636                         list_move_tail(&tail_req->execlist_link,
637                                        &engine->execlist_retired_req_list);
638                 }
639         }
640
641         list_add_tail(&request->execlist_link, &engine->execlist_queue);
642         if (num_elements == 0)
643                 execlists_context_unqueue(engine);
644
645         spin_unlock_bh(&engine->execlist_lock);
646 }
647
648 static int logical_ring_invalidate_all_caches(struct drm_i915_gem_request *req)
649 {
650         struct intel_engine_cs *engine = req->engine;
651         uint32_t flush_domains;
652         int ret;
653
654         flush_domains = 0;
655         if (engine->gpu_caches_dirty)
656                 flush_domains = I915_GEM_GPU_DOMAINS;
657
658         ret = engine->emit_flush(req, I915_GEM_GPU_DOMAINS, flush_domains);
659         if (ret)
660                 return ret;
661
662         engine->gpu_caches_dirty = false;
663         return 0;
664 }
665
666 static int execlists_move_to_gpu(struct drm_i915_gem_request *req,
667                                  struct list_head *vmas)
668 {
669         const unsigned other_rings = ~intel_engine_flag(req->engine);
670         struct i915_vma *vma;
671         uint32_t flush_domains = 0;
672         bool flush_chipset = false;
673         int ret;
674
675         list_for_each_entry(vma, vmas, exec_list) {
676                 struct drm_i915_gem_object *obj = vma->obj;
677
678                 if (obj->active & other_rings) {
679                         ret = i915_gem_object_sync(obj, req->engine, &req);
680                         if (ret)
681                                 return ret;
682                 }
683
684                 if (obj->base.write_domain & I915_GEM_DOMAIN_CPU)
685                         flush_chipset |= i915_gem_clflush_object(obj, false);
686
687                 flush_domains |= obj->base.write_domain;
688         }
689
690         if (flush_domains & I915_GEM_DOMAIN_GTT)
691                 wmb();
692
693         /* Unconditionally invalidate gpu caches and ensure that we do flush
694          * any residual writes from the previous batch.
695          */
696         return logical_ring_invalidate_all_caches(req);
697 }
698
699 int intel_logical_ring_alloc_request_extras(struct drm_i915_gem_request *request)
700 {
701         int ret = 0;
702
703         request->ringbuf = request->ctx->engine[request->engine->id].ringbuf;
704
705         if (i915.enable_guc_submission) {
706                 /*
707                  * Check that the GuC has space for the request before
708                  * going any further, as the i915_add_request() call
709                  * later on mustn't fail ...
710                  */
711                 struct intel_guc *guc = &request->i915->guc;
712
713                 ret = i915_guc_wq_check_space(guc->execbuf_client);
714                 if (ret)
715                         return ret;
716         }
717
718         if (request->ctx != request->i915->kernel_context)
719                 ret = intel_lr_context_pin(request->ctx, request->engine);
720
721         return ret;
722 }
723
724 /*
725  * intel_logical_ring_advance_and_submit() - advance the tail and submit the workload
726  * @request: Request to advance the logical ringbuffer of.
727  *
728  * The tail is updated in our logical ringbuffer struct, not in the actual context. What
729  * really happens during submission is that the context and current tail will be placed
730  * on a queue waiting for the ELSP to be ready to accept a new context submission. At that
731  * point, the tail *inside* the context is updated and the ELSP written to.
732  */
733 static int
734 intel_logical_ring_advance_and_submit(struct drm_i915_gem_request *request)
735 {
736         struct intel_ringbuffer *ringbuf = request->ringbuf;
737         struct drm_i915_private *dev_priv = request->i915;
738         struct intel_engine_cs *engine = request->engine;
739
740         intel_logical_ring_advance(ringbuf);
741         request->tail = ringbuf->tail;
742
743         /*
744          * Here we add two extra NOOPs as padding to avoid
745          * lite restore of a context with HEAD==TAIL.
746          *
747          * Caller must reserve WA_TAIL_DWORDS for us!
748          */
749         intel_logical_ring_emit(ringbuf, MI_NOOP);
750         intel_logical_ring_emit(ringbuf, MI_NOOP);
751         intel_logical_ring_advance(ringbuf);
752
753         if (intel_engine_stopped(engine))
754                 return 0;
755
756         if (engine->last_context != request->ctx) {
757                 if (engine->last_context)
758                         intel_lr_context_unpin(engine->last_context, engine);
759                 if (request->ctx != request->i915->kernel_context) {
760                         intel_lr_context_pin(request->ctx, engine);
761                         engine->last_context = request->ctx;
762                 } else {
763                         engine->last_context = NULL;
764                 }
765         }
766
767         if (dev_priv->guc.execbuf_client)
768                 i915_guc_submit(dev_priv->guc.execbuf_client, request);
769         else
770                 execlists_context_queue(request);
771
772         return 0;
773 }
774
775 int intel_logical_ring_reserve_space(struct drm_i915_gem_request *request)
776 {
777         /*
778          * The first call merely notes the reserve request and is common for
779          * all back ends. The subsequent localised _begin() call actually
780          * ensures that the reservation is available. Without the begin, if
781          * the request creator immediately submitted the request without
782          * adding any commands to it then there might not actually be
783          * sufficient room for the submission commands.
784          */
785         intel_ring_reserved_space_reserve(request->ringbuf, MIN_SPACE_FOR_ADD_REQUEST);
786
787         return intel_ring_begin(request, 0);
788 }
789
790 /**
791  * execlists_submission() - submit a batchbuffer for execution, Execlists style
792  * @dev: DRM device.
793  * @file: DRM file.
794  * @ring: Engine Command Streamer to submit to.
795  * @ctx: Context to employ for this submission.
796  * @args: execbuffer call arguments.
797  * @vmas: list of vmas.
798  * @batch_obj: the batchbuffer to submit.
799  * @exec_start: batchbuffer start virtual address pointer.
800  * @dispatch_flags: translated execbuffer call flags.
801  *
802  * This is the evil twin version of i915_gem_ringbuffer_submission. It abstracts
803  * away the submission details of the execbuffer ioctl call.
804  *
805  * Return: non-zero if the submission fails.
806  */
807 int intel_execlists_submission(struct i915_execbuffer_params *params,
808                                struct drm_i915_gem_execbuffer2 *args,
809                                struct list_head *vmas)
810 {
811         struct drm_device       *dev = params->dev;
812         struct intel_engine_cs *engine = params->engine;
813         struct drm_i915_private *dev_priv = dev->dev_private;
814         struct intel_ringbuffer *ringbuf = params->ctx->engine[engine->id].ringbuf;
815         u64 exec_start;
816         int instp_mode;
817         u32 instp_mask;
818         int ret;
819
820         instp_mode = args->flags & I915_EXEC_CONSTANTS_MASK;
821         instp_mask = I915_EXEC_CONSTANTS_MASK;
822         switch (instp_mode) {
823         case I915_EXEC_CONSTANTS_REL_GENERAL:
824         case I915_EXEC_CONSTANTS_ABSOLUTE:
825         case I915_EXEC_CONSTANTS_REL_SURFACE:
826                 if (instp_mode != 0 && engine != &dev_priv->engine[RCS]) {
827                         DRM_DEBUG("non-0 rel constants mode on non-RCS\n");
828                         return -EINVAL;
829                 }
830
831                 if (instp_mode != dev_priv->relative_constants_mode) {
832                         if (instp_mode == I915_EXEC_CONSTANTS_REL_SURFACE) {
833                                 DRM_DEBUG("rel surface constants mode invalid on gen5+\n");
834                                 return -EINVAL;
835                         }
836
837                         /* The HW changed the meaning on this bit on gen6 */
838                         instp_mask &= ~I915_EXEC_CONSTANTS_REL_SURFACE;
839                 }
840                 break;
841         default:
842                 DRM_DEBUG("execbuf with unknown constants: %d\n", instp_mode);
843                 return -EINVAL;
844         }
845
846         if (args->flags & I915_EXEC_GEN7_SOL_RESET) {
847                 DRM_DEBUG("sol reset is gen7 only\n");
848                 return -EINVAL;
849         }
850
851         ret = execlists_move_to_gpu(params->request, vmas);
852         if (ret)
853                 return ret;
854
855         if (engine == &dev_priv->engine[RCS] &&
856             instp_mode != dev_priv->relative_constants_mode) {
857                 ret = intel_ring_begin(params->request, 4);
858                 if (ret)
859                         return ret;
860
861                 intel_logical_ring_emit(ringbuf, MI_NOOP);
862                 intel_logical_ring_emit(ringbuf, MI_LOAD_REGISTER_IMM(1));
863                 intel_logical_ring_emit_reg(ringbuf, INSTPM);
864                 intel_logical_ring_emit(ringbuf, instp_mask << 16 | instp_mode);
865                 intel_logical_ring_advance(ringbuf);
866
867                 dev_priv->relative_constants_mode = instp_mode;
868         }
869
870         exec_start = params->batch_obj_vm_offset +
871                      args->batch_start_offset;
872
873         ret = engine->emit_bb_start(params->request, exec_start, params->dispatch_flags);
874         if (ret)
875                 return ret;
876
877         trace_i915_gem_ring_dispatch(params->request, params->dispatch_flags);
878
879         i915_gem_execbuffer_move_to_active(vmas, params->request);
880
881         return 0;
882 }
883
884 void intel_execlists_retire_requests(struct intel_engine_cs *engine)
885 {
886         struct drm_i915_gem_request *req, *tmp;
887         struct list_head retired_list;
888
889         WARN_ON(!mutex_is_locked(&engine->dev->struct_mutex));
890         if (list_empty(&engine->execlist_retired_req_list))
891                 return;
892
893         INIT_LIST_HEAD(&retired_list);
894         spin_lock_bh(&engine->execlist_lock);
895         list_replace_init(&engine->execlist_retired_req_list, &retired_list);
896         spin_unlock_bh(&engine->execlist_lock);
897
898         list_for_each_entry_safe(req, tmp, &retired_list, execlist_link) {
899                 struct intel_context *ctx = req->ctx;
900                 struct drm_i915_gem_object *ctx_obj =
901                                 ctx->engine[engine->id].state;
902
903                 if (ctx_obj && (ctx != req->i915->kernel_context))
904                         intel_lr_context_unpin(ctx, engine);
905
906                 list_del(&req->execlist_link);
907                 i915_gem_request_unreference(req);
908         }
909 }
910
911 void intel_logical_ring_stop(struct intel_engine_cs *engine)
912 {
913         struct drm_i915_private *dev_priv = engine->dev->dev_private;
914         int ret;
915
916         if (!intel_engine_initialized(engine))
917                 return;
918
919         ret = intel_engine_idle(engine);
920         if (ret)
921                 DRM_ERROR("failed to quiesce %s whilst cleaning up: %d\n",
922                           engine->name, ret);
923
924         /* TODO: Is this correct with Execlists enabled? */
925         I915_WRITE_MODE(engine, _MASKED_BIT_ENABLE(STOP_RING));
926         if (wait_for((I915_READ_MODE(engine) & MODE_IDLE) != 0, 1000)) {
927                 DRM_ERROR("%s :timed out trying to stop ring\n", engine->name);
928                 return;
929         }
930         I915_WRITE_MODE(engine, _MASKED_BIT_DISABLE(STOP_RING));
931 }
932
933 int logical_ring_flush_all_caches(struct drm_i915_gem_request *req)
934 {
935         struct intel_engine_cs *engine = req->engine;
936         int ret;
937
938         if (!engine->gpu_caches_dirty)
939                 return 0;
940
941         ret = engine->emit_flush(req, 0, I915_GEM_GPU_DOMAINS);
942         if (ret)
943                 return ret;
944
945         engine->gpu_caches_dirty = false;
946         return 0;
947 }
948
949 static int intel_lr_context_do_pin(struct intel_context *ctx,
950                                    struct intel_engine_cs *engine)
951 {
952         struct drm_device *dev = engine->dev;
953         struct drm_i915_private *dev_priv = dev->dev_private;
954         struct drm_i915_gem_object *ctx_obj = ctx->engine[engine->id].state;
955         struct intel_ringbuffer *ringbuf = ctx->engine[engine->id].ringbuf;
956         void *vaddr;
957         u32 *lrc_reg_state;
958         int ret;
959
960         WARN_ON(!mutex_is_locked(&engine->dev->struct_mutex));
961
962         ret = i915_gem_obj_ggtt_pin(ctx_obj, GEN8_LR_CONTEXT_ALIGN,
963                         PIN_OFFSET_BIAS | GUC_WOPCM_TOP);
964         if (ret)
965                 return ret;
966
967         vaddr = i915_gem_object_pin_map(ctx_obj);
968         if (IS_ERR(vaddr)) {
969                 ret = PTR_ERR(vaddr);
970                 goto unpin_ctx_obj;
971         }
972
973         lrc_reg_state = vaddr + LRC_STATE_PN * PAGE_SIZE;
974
975         ret = intel_pin_and_map_ringbuffer_obj(engine->dev, ringbuf);
976         if (ret)
977                 goto unpin_map;
978
979         ctx->engine[engine->id].lrc_vma = i915_gem_obj_to_ggtt(ctx_obj);
980         intel_lr_context_descriptor_update(ctx, engine);
981         lrc_reg_state[CTX_RING_BUFFER_START+1] = ringbuf->vma->node.start;
982         ctx->engine[engine->id].lrc_reg_state = lrc_reg_state;
983         ctx_obj->dirty = true;
984
985         /* Invalidate GuC TLB. */
986         if (i915.enable_guc_submission)
987                 I915_WRITE(GEN8_GTCR, GEN8_GTCR_INVALIDATE);
988
989         return ret;
990
991 unpin_map:
992         i915_gem_object_unpin_map(ctx_obj);
993 unpin_ctx_obj:
994         i915_gem_object_ggtt_unpin(ctx_obj);
995
996         return ret;
997 }
998
999 static int intel_lr_context_pin(struct intel_context *ctx,
1000                                 struct intel_engine_cs *engine)
1001 {
1002         int ret = 0;
1003
1004         if (ctx->engine[engine->id].pin_count++ == 0) {
1005                 ret = intel_lr_context_do_pin(ctx, engine);
1006                 if (ret)
1007                         goto reset_pin_count;
1008
1009                 i915_gem_context_reference(ctx);
1010         }
1011         return ret;
1012
1013 reset_pin_count:
1014         ctx->engine[engine->id].pin_count = 0;
1015         return ret;
1016 }
1017
1018 void intel_lr_context_unpin(struct intel_context *ctx,
1019                             struct intel_engine_cs *engine)
1020 {
1021         struct drm_i915_gem_object *ctx_obj = ctx->engine[engine->id].state;
1022
1023         WARN_ON(!mutex_is_locked(&ctx->i915->dev->struct_mutex));
1024         if (--ctx->engine[engine->id].pin_count == 0) {
1025                 i915_gem_object_unpin_map(ctx_obj);
1026                 intel_unpin_ringbuffer_obj(ctx->engine[engine->id].ringbuf);
1027                 i915_gem_object_ggtt_unpin(ctx_obj);
1028                 ctx->engine[engine->id].lrc_vma = NULL;
1029                 ctx->engine[engine->id].lrc_desc = 0;
1030                 ctx->engine[engine->id].lrc_reg_state = NULL;
1031
1032                 i915_gem_context_unreference(ctx);
1033         }
1034 }
1035
1036 static int intel_logical_ring_workarounds_emit(struct drm_i915_gem_request *req)
1037 {
1038         int ret, i;
1039         struct intel_engine_cs *engine = req->engine;
1040         struct intel_ringbuffer *ringbuf = req->ringbuf;
1041         struct drm_device *dev = engine->dev;
1042         struct drm_i915_private *dev_priv = dev->dev_private;
1043         struct i915_workarounds *w = &dev_priv->workarounds;
1044
1045         if (w->count == 0)
1046                 return 0;
1047
1048         engine->gpu_caches_dirty = true;
1049         ret = logical_ring_flush_all_caches(req);
1050         if (ret)
1051                 return ret;
1052
1053         ret = intel_ring_begin(req, w->count * 2 + 2);
1054         if (ret)
1055                 return ret;
1056
1057         intel_logical_ring_emit(ringbuf, MI_LOAD_REGISTER_IMM(w->count));
1058         for (i = 0; i < w->count; i++) {
1059                 intel_logical_ring_emit_reg(ringbuf, w->reg[i].addr);
1060                 intel_logical_ring_emit(ringbuf, w->reg[i].value);
1061         }
1062         intel_logical_ring_emit(ringbuf, MI_NOOP);
1063
1064         intel_logical_ring_advance(ringbuf);
1065
1066         engine->gpu_caches_dirty = true;
1067         ret = logical_ring_flush_all_caches(req);
1068         if (ret)
1069                 return ret;
1070
1071         return 0;
1072 }
1073
1074 #define wa_ctx_emit(batch, index, cmd)                                  \
1075         do {                                                            \
1076                 int __index = (index)++;                                \
1077                 if (WARN_ON(__index >= (PAGE_SIZE / sizeof(uint32_t)))) { \
1078                         return -ENOSPC;                                 \
1079                 }                                                       \
1080                 batch[__index] = (cmd);                                 \
1081         } while (0)
1082
1083 #define wa_ctx_emit_reg(batch, index, reg) \
1084         wa_ctx_emit((batch), (index), i915_mmio_reg_offset(reg))
1085
1086 /*
1087  * In this WA we need to set GEN8_L3SQCREG4[21:21] and reset it after
1088  * PIPE_CONTROL instruction. This is required for the flush to happen correctly
1089  * but there is a slight complication as this is applied in WA batch where the
1090  * values are only initialized once so we cannot take register value at the
1091  * beginning and reuse it further; hence we save its value to memory, upload a
1092  * constant value with bit21 set and then we restore it back with the saved value.
1093  * To simplify the WA, a constant value is formed by using the default value
1094  * of this register. This shouldn't be a problem because we are only modifying
1095  * it for a short period and this batch in non-premptible. We can ofcourse
1096  * use additional instructions that read the actual value of the register
1097  * at that time and set our bit of interest but it makes the WA complicated.
1098  *
1099  * This WA is also required for Gen9 so extracting as a function avoids
1100  * code duplication.
1101  */
1102 static inline int gen8_emit_flush_coherentl3_wa(struct intel_engine_cs *engine,
1103                                                 uint32_t *const batch,
1104                                                 uint32_t index)
1105 {
1106         struct drm_i915_private *dev_priv = engine->dev->dev_private;
1107         uint32_t l3sqc4_flush = (0x40400000 | GEN8_LQSC_FLUSH_COHERENT_LINES);
1108
1109         /*
1110          * WaDisableLSQCROPERFforOCL:skl,kbl
1111          * This WA is implemented in skl_init_clock_gating() but since
1112          * this batch updates GEN8_L3SQCREG4 with default value we need to
1113          * set this bit here to retain the WA during flush.
1114          */
1115         if (IS_SKL_REVID(dev_priv, 0, SKL_REVID_E0) ||
1116             IS_KBL_REVID(dev_priv, 0, KBL_REVID_E0))
1117                 l3sqc4_flush |= GEN8_LQSC_RO_PERF_DIS;
1118
1119         wa_ctx_emit(batch, index, (MI_STORE_REGISTER_MEM_GEN8 |
1120                                    MI_SRM_LRM_GLOBAL_GTT));
1121         wa_ctx_emit_reg(batch, index, GEN8_L3SQCREG4);
1122         wa_ctx_emit(batch, index, engine->scratch.gtt_offset + 256);
1123         wa_ctx_emit(batch, index, 0);
1124
1125         wa_ctx_emit(batch, index, MI_LOAD_REGISTER_IMM(1));
1126         wa_ctx_emit_reg(batch, index, GEN8_L3SQCREG4);
1127         wa_ctx_emit(batch, index, l3sqc4_flush);
1128
1129         wa_ctx_emit(batch, index, GFX_OP_PIPE_CONTROL(6));
1130         wa_ctx_emit(batch, index, (PIPE_CONTROL_CS_STALL |
1131                                    PIPE_CONTROL_DC_FLUSH_ENABLE));
1132         wa_ctx_emit(batch, index, 0);
1133         wa_ctx_emit(batch, index, 0);
1134         wa_ctx_emit(batch, index, 0);
1135         wa_ctx_emit(batch, index, 0);
1136
1137         wa_ctx_emit(batch, index, (MI_LOAD_REGISTER_MEM_GEN8 |
1138                                    MI_SRM_LRM_GLOBAL_GTT));
1139         wa_ctx_emit_reg(batch, index, GEN8_L3SQCREG4);
1140         wa_ctx_emit(batch, index, engine->scratch.gtt_offset + 256);
1141         wa_ctx_emit(batch, index, 0);
1142
1143         return index;
1144 }
1145
1146 static inline uint32_t wa_ctx_start(struct i915_wa_ctx_bb *wa_ctx,
1147                                     uint32_t offset,
1148                                     uint32_t start_alignment)
1149 {
1150         return wa_ctx->offset = ALIGN(offset, start_alignment);
1151 }
1152
1153 static inline int wa_ctx_end(struct i915_wa_ctx_bb *wa_ctx,
1154                              uint32_t offset,
1155                              uint32_t size_alignment)
1156 {
1157         wa_ctx->size = offset - wa_ctx->offset;
1158
1159         WARN(wa_ctx->size % size_alignment,
1160              "wa_ctx_bb failed sanity checks: size %d is not aligned to %d\n",
1161              wa_ctx->size, size_alignment);
1162         return 0;
1163 }
1164
1165 /**
1166  * gen8_init_indirectctx_bb() - initialize indirect ctx batch with WA
1167  *
1168  * @ring: only applicable for RCS
1169  * @wa_ctx: structure representing wa_ctx
1170  *  offset: specifies start of the batch, should be cache-aligned. This is updated
1171  *    with the offset value received as input.
1172  *  size: size of the batch in DWORDS but HW expects in terms of cachelines
1173  * @batch: page in which WA are loaded
1174  * @offset: This field specifies the start of the batch, it should be
1175  *  cache-aligned otherwise it is adjusted accordingly.
1176  *  Typically we only have one indirect_ctx and per_ctx batch buffer which are
1177  *  initialized at the beginning and shared across all contexts but this field
1178  *  helps us to have multiple batches at different offsets and select them based
1179  *  on a criteria. At the moment this batch always start at the beginning of the page
1180  *  and at this point we don't have multiple wa_ctx batch buffers.
1181  *
1182  *  The number of WA applied are not known at the beginning; we use this field
1183  *  to return the no of DWORDS written.
1184  *
1185  *  It is to be noted that this batch does not contain MI_BATCH_BUFFER_END
1186  *  so it adds NOOPs as padding to make it cacheline aligned.
1187  *  MI_BATCH_BUFFER_END will be added to perctx batch and both of them together
1188  *  makes a complete batch buffer.
1189  *
1190  * Return: non-zero if we exceed the PAGE_SIZE limit.
1191  */
1192
1193 static int gen8_init_indirectctx_bb(struct intel_engine_cs *engine,
1194                                     struct i915_wa_ctx_bb *wa_ctx,
1195                                     uint32_t *const batch,
1196                                     uint32_t *offset)
1197 {
1198         uint32_t scratch_addr;
1199         uint32_t index = wa_ctx_start(wa_ctx, *offset, CACHELINE_DWORDS);
1200
1201         /* WaDisableCtxRestoreArbitration:bdw,chv */
1202         wa_ctx_emit(batch, index, MI_ARB_ON_OFF | MI_ARB_DISABLE);
1203
1204         /* WaFlushCoherentL3CacheLinesAtContextSwitch:bdw */
1205         if (IS_BROADWELL(engine->dev)) {
1206                 int rc = gen8_emit_flush_coherentl3_wa(engine, batch, index);
1207                 if (rc < 0)
1208                         return rc;
1209                 index = rc;
1210         }
1211
1212         /* WaClearSlmSpaceAtContextSwitch:bdw,chv */
1213         /* Actual scratch location is at 128 bytes offset */
1214         scratch_addr = engine->scratch.gtt_offset + 2*CACHELINE_BYTES;
1215
1216         wa_ctx_emit(batch, index, GFX_OP_PIPE_CONTROL(6));
1217         wa_ctx_emit(batch, index, (PIPE_CONTROL_FLUSH_L3 |
1218                                    PIPE_CONTROL_GLOBAL_GTT_IVB |
1219                                    PIPE_CONTROL_CS_STALL |
1220                                    PIPE_CONTROL_QW_WRITE));
1221         wa_ctx_emit(batch, index, scratch_addr);
1222         wa_ctx_emit(batch, index, 0);
1223         wa_ctx_emit(batch, index, 0);
1224         wa_ctx_emit(batch, index, 0);
1225
1226         /* Pad to end of cacheline */
1227         while (index % CACHELINE_DWORDS)
1228                 wa_ctx_emit(batch, index, MI_NOOP);
1229
1230         /*
1231          * MI_BATCH_BUFFER_END is not required in Indirect ctx BB because
1232          * execution depends on the length specified in terms of cache lines
1233          * in the register CTX_RCS_INDIRECT_CTX
1234          */
1235
1236         return wa_ctx_end(wa_ctx, *offset = index, CACHELINE_DWORDS);
1237 }
1238
1239 /**
1240  * gen8_init_perctx_bb() - initialize per ctx batch with WA
1241  *
1242  * @ring: only applicable for RCS
1243  * @wa_ctx: structure representing wa_ctx
1244  *  offset: specifies start of the batch, should be cache-aligned.
1245  *  size: size of the batch in DWORDS but HW expects in terms of cachelines
1246  * @batch: page in which WA are loaded
1247  * @offset: This field specifies the start of this batch.
1248  *   This batch is started immediately after indirect_ctx batch. Since we ensure
1249  *   that indirect_ctx ends on a cacheline this batch is aligned automatically.
1250  *
1251  *   The number of DWORDS written are returned using this field.
1252  *
1253  *  This batch is terminated with MI_BATCH_BUFFER_END and so we need not add padding
1254  *  to align it with cacheline as padding after MI_BATCH_BUFFER_END is redundant.
1255  */
1256 static int gen8_init_perctx_bb(struct intel_engine_cs *engine,
1257                                struct i915_wa_ctx_bb *wa_ctx,
1258                                uint32_t *const batch,
1259                                uint32_t *offset)
1260 {
1261         uint32_t index = wa_ctx_start(wa_ctx, *offset, CACHELINE_DWORDS);
1262
1263         /* WaDisableCtxRestoreArbitration:bdw,chv */
1264         wa_ctx_emit(batch, index, MI_ARB_ON_OFF | MI_ARB_ENABLE);
1265
1266         wa_ctx_emit(batch, index, MI_BATCH_BUFFER_END);
1267
1268         return wa_ctx_end(wa_ctx, *offset = index, 1);
1269 }
1270
1271 static int gen9_init_indirectctx_bb(struct intel_engine_cs *engine,
1272                                     struct i915_wa_ctx_bb *wa_ctx,
1273                                     uint32_t *const batch,
1274                                     uint32_t *offset)
1275 {
1276         int ret;
1277         struct drm_device *dev = engine->dev;
1278         struct drm_i915_private *dev_priv = dev->dev_private;
1279         uint32_t index = wa_ctx_start(wa_ctx, *offset, CACHELINE_DWORDS);
1280
1281         /* WaDisableCtxRestoreArbitration:skl,bxt */
1282         if (IS_SKL_REVID(dev, 0, SKL_REVID_D0) ||
1283             IS_BXT_REVID(dev, 0, BXT_REVID_A1))
1284                 wa_ctx_emit(batch, index, MI_ARB_ON_OFF | MI_ARB_DISABLE);
1285
1286         /* WaFlushCoherentL3CacheLinesAtContextSwitch:skl,bxt */
1287         ret = gen8_emit_flush_coherentl3_wa(engine, batch, index);
1288         if (ret < 0)
1289                 return ret;
1290         index = ret;
1291
1292         /* WaClearSlmSpaceAtContextSwitch:kbl */
1293         /* Actual scratch location is at 128 bytes offset */
1294         if (IS_KBL_REVID(dev_priv, 0, KBL_REVID_A0)) {
1295                 uint32_t scratch_addr
1296                         = engine->scratch.gtt_offset + 2*CACHELINE_BYTES;
1297
1298                 wa_ctx_emit(batch, index, GFX_OP_PIPE_CONTROL(6));
1299                 wa_ctx_emit(batch, index, (PIPE_CONTROL_FLUSH_L3 |
1300                                            PIPE_CONTROL_GLOBAL_GTT_IVB |
1301                                            PIPE_CONTROL_CS_STALL |
1302                                            PIPE_CONTROL_QW_WRITE));
1303                 wa_ctx_emit(batch, index, scratch_addr);
1304                 wa_ctx_emit(batch, index, 0);
1305                 wa_ctx_emit(batch, index, 0);
1306                 wa_ctx_emit(batch, index, 0);
1307         }
1308         /* Pad to end of cacheline */
1309         while (index % CACHELINE_DWORDS)
1310                 wa_ctx_emit(batch, index, MI_NOOP);
1311
1312         return wa_ctx_end(wa_ctx, *offset = index, CACHELINE_DWORDS);
1313 }
1314
1315 static int gen9_init_perctx_bb(struct intel_engine_cs *engine,
1316                                struct i915_wa_ctx_bb *wa_ctx,
1317                                uint32_t *const batch,
1318                                uint32_t *offset)
1319 {
1320         struct drm_device *dev = engine->dev;
1321         uint32_t index = wa_ctx_start(wa_ctx, *offset, CACHELINE_DWORDS);
1322
1323         /* WaSetDisablePixMaskCammingAndRhwoInCommonSliceChicken:skl,bxt */
1324         if (IS_SKL_REVID(dev, 0, SKL_REVID_B0) ||
1325             IS_BXT_REVID(dev, 0, BXT_REVID_A1)) {
1326                 wa_ctx_emit(batch, index, MI_LOAD_REGISTER_IMM(1));
1327                 wa_ctx_emit_reg(batch, index, GEN9_SLICE_COMMON_ECO_CHICKEN0);
1328                 wa_ctx_emit(batch, index,
1329                             _MASKED_BIT_ENABLE(DISABLE_PIXEL_MASK_CAMMING));
1330                 wa_ctx_emit(batch, index, MI_NOOP);
1331         }
1332
1333         /* WaClearTdlStateAckDirtyBits:bxt */
1334         if (IS_BXT_REVID(dev, 0, BXT_REVID_B0)) {
1335                 wa_ctx_emit(batch, index, MI_LOAD_REGISTER_IMM(4));
1336
1337                 wa_ctx_emit_reg(batch, index, GEN8_STATE_ACK);
1338                 wa_ctx_emit(batch, index, _MASKED_BIT_DISABLE(GEN9_SUBSLICE_TDL_ACK_BITS));
1339
1340                 wa_ctx_emit_reg(batch, index, GEN9_STATE_ACK_SLICE1);
1341                 wa_ctx_emit(batch, index, _MASKED_BIT_DISABLE(GEN9_SUBSLICE_TDL_ACK_BITS));
1342
1343                 wa_ctx_emit_reg(batch, index, GEN9_STATE_ACK_SLICE2);
1344                 wa_ctx_emit(batch, index, _MASKED_BIT_DISABLE(GEN9_SUBSLICE_TDL_ACK_BITS));
1345
1346                 wa_ctx_emit_reg(batch, index, GEN7_ROW_CHICKEN2);
1347                 /* dummy write to CS, mask bits are 0 to ensure the register is not modified */
1348                 wa_ctx_emit(batch, index, 0x0);
1349                 wa_ctx_emit(batch, index, MI_NOOP);
1350         }
1351
1352         /* WaDisableCtxRestoreArbitration:skl,bxt */
1353         if (IS_SKL_REVID(dev, 0, SKL_REVID_D0) ||
1354             IS_BXT_REVID(dev, 0, BXT_REVID_A1))
1355                 wa_ctx_emit(batch, index, MI_ARB_ON_OFF | MI_ARB_ENABLE);
1356
1357         wa_ctx_emit(batch, index, MI_BATCH_BUFFER_END);
1358
1359         return wa_ctx_end(wa_ctx, *offset = index, 1);
1360 }
1361
1362 static int lrc_setup_wa_ctx_obj(struct intel_engine_cs *engine, u32 size)
1363 {
1364         int ret;
1365
1366         engine->wa_ctx.obj = i915_gem_alloc_object(engine->dev,
1367                                                    PAGE_ALIGN(size));
1368         if (!engine->wa_ctx.obj) {
1369                 DRM_DEBUG_DRIVER("alloc LRC WA ctx backing obj failed.\n");
1370                 return -ENOMEM;
1371         }
1372
1373         ret = i915_gem_obj_ggtt_pin(engine->wa_ctx.obj, PAGE_SIZE, 0);
1374         if (ret) {
1375                 DRM_DEBUG_DRIVER("pin LRC WA ctx backing obj failed: %d\n",
1376                                  ret);
1377                 drm_gem_object_unreference(&engine->wa_ctx.obj->base);
1378                 return ret;
1379         }
1380
1381         return 0;
1382 }
1383
1384 static void lrc_destroy_wa_ctx_obj(struct intel_engine_cs *engine)
1385 {
1386         if (engine->wa_ctx.obj) {
1387                 i915_gem_object_ggtt_unpin(engine->wa_ctx.obj);
1388                 drm_gem_object_unreference(&engine->wa_ctx.obj->base);
1389                 engine->wa_ctx.obj = NULL;
1390         }
1391 }
1392
1393 static int intel_init_workaround_bb(struct intel_engine_cs *engine)
1394 {
1395         int ret;
1396         uint32_t *batch;
1397         uint32_t offset;
1398         struct page *page;
1399         struct i915_ctx_workarounds *wa_ctx = &engine->wa_ctx;
1400
1401         WARN_ON(engine->id != RCS);
1402
1403         /* update this when WA for higher Gen are added */
1404         if (INTEL_INFO(engine->dev)->gen > 9) {
1405                 DRM_ERROR("WA batch buffer is not initialized for Gen%d\n",
1406                           INTEL_INFO(engine->dev)->gen);
1407                 return 0;
1408         }
1409
1410         /* some WA perform writes to scratch page, ensure it is valid */
1411         if (engine->scratch.obj == NULL) {
1412                 DRM_ERROR("scratch page not allocated for %s\n", engine->name);
1413                 return -EINVAL;
1414         }
1415
1416         ret = lrc_setup_wa_ctx_obj(engine, PAGE_SIZE);
1417         if (ret) {
1418                 DRM_DEBUG_DRIVER("Failed to setup context WA page: %d\n", ret);
1419                 return ret;
1420         }
1421
1422         page = i915_gem_object_get_dirty_page(wa_ctx->obj, 0);
1423         batch = kmap_atomic(page);
1424         offset = 0;
1425
1426         if (INTEL_INFO(engine->dev)->gen == 8) {
1427                 ret = gen8_init_indirectctx_bb(engine,
1428                                                &wa_ctx->indirect_ctx,
1429                                                batch,
1430                                                &offset);
1431                 if (ret)
1432                         goto out;
1433
1434                 ret = gen8_init_perctx_bb(engine,
1435                                           &wa_ctx->per_ctx,
1436                                           batch,
1437                                           &offset);
1438                 if (ret)
1439                         goto out;
1440         } else if (INTEL_INFO(engine->dev)->gen == 9) {
1441                 ret = gen9_init_indirectctx_bb(engine,
1442                                                &wa_ctx->indirect_ctx,
1443                                                batch,
1444                                                &offset);
1445                 if (ret)
1446                         goto out;
1447
1448                 ret = gen9_init_perctx_bb(engine,
1449                                           &wa_ctx->per_ctx,
1450                                           batch,
1451                                           &offset);
1452                 if (ret)
1453                         goto out;
1454         }
1455
1456 out:
1457         kunmap_atomic(batch);
1458         if (ret)
1459                 lrc_destroy_wa_ctx_obj(engine);
1460
1461         return ret;
1462 }
1463
1464 static void lrc_init_hws(struct intel_engine_cs *engine)
1465 {
1466         struct drm_i915_private *dev_priv = engine->dev->dev_private;
1467
1468         I915_WRITE(RING_HWS_PGA(engine->mmio_base),
1469                    (u32)engine->status_page.gfx_addr);
1470         POSTING_READ(RING_HWS_PGA(engine->mmio_base));
1471 }
1472
1473 static int gen8_init_common_ring(struct intel_engine_cs *engine)
1474 {
1475         struct drm_device *dev = engine->dev;
1476         struct drm_i915_private *dev_priv = dev->dev_private;
1477         unsigned int next_context_status_buffer_hw;
1478
1479         lrc_init_hws(engine);
1480
1481         I915_WRITE_IMR(engine,
1482                        ~(engine->irq_enable_mask | engine->irq_keep_mask));
1483         I915_WRITE(RING_HWSTAM(engine->mmio_base), 0xffffffff);
1484
1485         I915_WRITE(RING_MODE_GEN7(engine),
1486                    _MASKED_BIT_DISABLE(GFX_REPLAY_MODE) |
1487                    _MASKED_BIT_ENABLE(GFX_RUN_LIST_ENABLE));
1488         POSTING_READ(RING_MODE_GEN7(engine));
1489
1490         /*
1491          * Instead of resetting the Context Status Buffer (CSB) read pointer to
1492          * zero, we need to read the write pointer from hardware and use its
1493          * value because "this register is power context save restored".
1494          * Effectively, these states have been observed:
1495          *
1496          *      | Suspend-to-idle (freeze) | Suspend-to-RAM (mem) |
1497          * BDW  | CSB regs not reset       | CSB regs reset       |
1498          * CHT  | CSB regs not reset       | CSB regs not reset   |
1499          * SKL  |         ?                |         ?            |
1500          * BXT  |         ?                |         ?            |
1501          */
1502         next_context_status_buffer_hw =
1503                 GEN8_CSB_WRITE_PTR(I915_READ(RING_CONTEXT_STATUS_PTR(engine)));
1504
1505         /*
1506          * When the CSB registers are reset (also after power-up / gpu reset),
1507          * CSB write pointer is set to all 1's, which is not valid, use '5' in
1508          * this special case, so the first element read is CSB[0].
1509          */
1510         if (next_context_status_buffer_hw == GEN8_CSB_PTR_MASK)
1511                 next_context_status_buffer_hw = (GEN8_CSB_ENTRIES - 1);
1512
1513         engine->next_context_status_buffer = next_context_status_buffer_hw;
1514         DRM_DEBUG_DRIVER("Execlists enabled for %s\n", engine->name);
1515
1516         intel_engine_init_hangcheck(engine);
1517
1518         return intel_mocs_init_engine(engine);
1519 }
1520
1521 static int gen8_init_render_ring(struct intel_engine_cs *engine)
1522 {
1523         struct drm_device *dev = engine->dev;
1524         struct drm_i915_private *dev_priv = dev->dev_private;
1525         int ret;
1526
1527         ret = gen8_init_common_ring(engine);
1528         if (ret)
1529                 return ret;
1530
1531         /* We need to disable the AsyncFlip performance optimisations in order
1532          * to use MI_WAIT_FOR_EVENT within the CS. It should already be
1533          * programmed to '1' on all products.
1534          *
1535          * WaDisableAsyncFlipPerfMode:snb,ivb,hsw,vlv,bdw,chv
1536          */
1537         I915_WRITE(MI_MODE, _MASKED_BIT_ENABLE(ASYNC_FLIP_PERF_DISABLE));
1538
1539         I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_FORCE_ORDERING));
1540
1541         return init_workarounds_ring(engine);
1542 }
1543
1544 static int gen9_init_render_ring(struct intel_engine_cs *engine)
1545 {
1546         int ret;
1547
1548         ret = gen8_init_common_ring(engine);
1549         if (ret)
1550                 return ret;
1551
1552         return init_workarounds_ring(engine);
1553 }
1554
1555 static int intel_logical_ring_emit_pdps(struct drm_i915_gem_request *req)
1556 {
1557         struct i915_hw_ppgtt *ppgtt = req->ctx->ppgtt;
1558         struct intel_engine_cs *engine = req->engine;
1559         struct intel_ringbuffer *ringbuf = req->ringbuf;
1560         const int num_lri_cmds = GEN8_LEGACY_PDPES * 2;
1561         int i, ret;
1562
1563         ret = intel_ring_begin(req, num_lri_cmds * 2 + 2);
1564         if (ret)
1565                 return ret;
1566
1567         intel_logical_ring_emit(ringbuf, MI_LOAD_REGISTER_IMM(num_lri_cmds));
1568         for (i = GEN8_LEGACY_PDPES - 1; i >= 0; i--) {
1569                 const dma_addr_t pd_daddr = i915_page_dir_dma_addr(ppgtt, i);
1570
1571                 intel_logical_ring_emit_reg(ringbuf,
1572                                             GEN8_RING_PDP_UDW(engine, i));
1573                 intel_logical_ring_emit(ringbuf, upper_32_bits(pd_daddr));
1574                 intel_logical_ring_emit_reg(ringbuf,
1575                                             GEN8_RING_PDP_LDW(engine, i));
1576                 intel_logical_ring_emit(ringbuf, lower_32_bits(pd_daddr));
1577         }
1578
1579         intel_logical_ring_emit(ringbuf, MI_NOOP);
1580         intel_logical_ring_advance(ringbuf);
1581
1582         return 0;
1583 }
1584
1585 static int gen8_emit_bb_start(struct drm_i915_gem_request *req,
1586                               u64 offset, unsigned dispatch_flags)
1587 {
1588         struct intel_ringbuffer *ringbuf = req->ringbuf;
1589         bool ppgtt = !(dispatch_flags & I915_DISPATCH_SECURE);
1590         int ret;
1591
1592         /* Don't rely in hw updating PDPs, specially in lite-restore.
1593          * Ideally, we should set Force PD Restore in ctx descriptor,
1594          * but we can't. Force Restore would be a second option, but
1595          * it is unsafe in case of lite-restore (because the ctx is
1596          * not idle). PML4 is allocated during ppgtt init so this is
1597          * not needed in 48-bit.*/
1598         if (req->ctx->ppgtt &&
1599             (intel_engine_flag(req->engine) & req->ctx->ppgtt->pd_dirty_rings)) {
1600                 if (!USES_FULL_48BIT_PPGTT(req->i915) &&
1601                     !intel_vgpu_active(req->i915->dev)) {
1602                         ret = intel_logical_ring_emit_pdps(req);
1603                         if (ret)
1604                                 return ret;
1605                 }
1606
1607                 req->ctx->ppgtt->pd_dirty_rings &= ~intel_engine_flag(req->engine);
1608         }
1609
1610         ret = intel_ring_begin(req, 4);
1611         if (ret)
1612                 return ret;
1613
1614         /* FIXME(BDW): Address space and security selectors. */
1615         intel_logical_ring_emit(ringbuf, MI_BATCH_BUFFER_START_GEN8 |
1616                                 (ppgtt<<8) |
1617                                 (dispatch_flags & I915_DISPATCH_RS ?
1618                                  MI_BATCH_RESOURCE_STREAMER : 0));
1619         intel_logical_ring_emit(ringbuf, lower_32_bits(offset));
1620         intel_logical_ring_emit(ringbuf, upper_32_bits(offset));
1621         intel_logical_ring_emit(ringbuf, MI_NOOP);
1622         intel_logical_ring_advance(ringbuf);
1623
1624         return 0;
1625 }
1626
1627 static bool gen8_logical_ring_get_irq(struct intel_engine_cs *engine)
1628 {
1629         struct drm_device *dev = engine->dev;
1630         struct drm_i915_private *dev_priv = dev->dev_private;
1631         unsigned long flags;
1632
1633         if (WARN_ON(!intel_irqs_enabled(dev_priv)))
1634                 return false;
1635
1636         spin_lock_irqsave(&dev_priv->irq_lock, flags);
1637         if (engine->irq_refcount++ == 0) {
1638                 I915_WRITE_IMR(engine,
1639                                ~(engine->irq_enable_mask | engine->irq_keep_mask));
1640                 POSTING_READ(RING_IMR(engine->mmio_base));
1641         }
1642         spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
1643
1644         return true;
1645 }
1646
1647 static void gen8_logical_ring_put_irq(struct intel_engine_cs *engine)
1648 {
1649         struct drm_device *dev = engine->dev;
1650         struct drm_i915_private *dev_priv = dev->dev_private;
1651         unsigned long flags;
1652
1653         spin_lock_irqsave(&dev_priv->irq_lock, flags);
1654         if (--engine->irq_refcount == 0) {
1655                 I915_WRITE_IMR(engine, ~engine->irq_keep_mask);
1656                 POSTING_READ(RING_IMR(engine->mmio_base));
1657         }
1658         spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
1659 }
1660
1661 static int gen8_emit_flush(struct drm_i915_gem_request *request,
1662                            u32 invalidate_domains,
1663                            u32 unused)
1664 {
1665         struct intel_ringbuffer *ringbuf = request->ringbuf;
1666         struct intel_engine_cs *engine = ringbuf->engine;
1667         struct drm_device *dev = engine->dev;
1668         struct drm_i915_private *dev_priv = dev->dev_private;
1669         uint32_t cmd;
1670         int ret;
1671
1672         ret = intel_ring_begin(request, 4);
1673         if (ret)
1674                 return ret;
1675
1676         cmd = MI_FLUSH_DW + 1;
1677
1678         /* We always require a command barrier so that subsequent
1679          * commands, such as breadcrumb interrupts, are strictly ordered
1680          * wrt the contents of the write cache being flushed to memory
1681          * (and thus being coherent from the CPU).
1682          */
1683         cmd |= MI_FLUSH_DW_STORE_INDEX | MI_FLUSH_DW_OP_STOREDW;
1684
1685         if (invalidate_domains & I915_GEM_GPU_DOMAINS) {
1686                 cmd |= MI_INVALIDATE_TLB;
1687                 if (engine == &dev_priv->engine[VCS])
1688                         cmd |= MI_INVALIDATE_BSD;
1689         }
1690
1691         intel_logical_ring_emit(ringbuf, cmd);
1692         intel_logical_ring_emit(ringbuf,
1693                                 I915_GEM_HWS_SCRATCH_ADDR |
1694                                 MI_FLUSH_DW_USE_GTT);
1695         intel_logical_ring_emit(ringbuf, 0); /* upper addr */
1696         intel_logical_ring_emit(ringbuf, 0); /* value */
1697         intel_logical_ring_advance(ringbuf);
1698
1699         return 0;
1700 }
1701
1702 static int gen8_emit_flush_render(struct drm_i915_gem_request *request,
1703                                   u32 invalidate_domains,
1704                                   u32 flush_domains)
1705 {
1706         struct intel_ringbuffer *ringbuf = request->ringbuf;
1707         struct intel_engine_cs *engine = ringbuf->engine;
1708         u32 scratch_addr = engine->scratch.gtt_offset + 2 * CACHELINE_BYTES;
1709         bool vf_flush_wa = false, dc_flush_wa = false;
1710         u32 flags = 0;
1711         int ret;
1712         int len;
1713
1714         flags |= PIPE_CONTROL_CS_STALL;
1715
1716         if (flush_domains) {
1717                 flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
1718                 flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
1719                 flags |= PIPE_CONTROL_DC_FLUSH_ENABLE;
1720                 flags |= PIPE_CONTROL_FLUSH_ENABLE;
1721         }
1722
1723         if (invalidate_domains) {
1724                 flags |= PIPE_CONTROL_TLB_INVALIDATE;
1725                 flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
1726                 flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
1727                 flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
1728                 flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
1729                 flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
1730                 flags |= PIPE_CONTROL_QW_WRITE;
1731                 flags |= PIPE_CONTROL_GLOBAL_GTT_IVB;
1732
1733                 /*
1734                  * On GEN9: before VF_CACHE_INVALIDATE we need to emit a NULL
1735                  * pipe control.
1736                  */
1737                 if (IS_GEN9(engine->dev))
1738                         vf_flush_wa = true;
1739
1740                 /* WaForGAMHang:kbl */
1741                 if (IS_KBL_REVID(request->i915, 0, KBL_REVID_B0))
1742                         dc_flush_wa = true;
1743         }
1744
1745         len = 6;
1746
1747         if (vf_flush_wa)
1748                 len += 6;
1749
1750         if (dc_flush_wa)
1751                 len += 12;
1752
1753         ret = intel_ring_begin(request, len);
1754         if (ret)
1755                 return ret;
1756
1757         if (vf_flush_wa) {
1758                 intel_logical_ring_emit(ringbuf, GFX_OP_PIPE_CONTROL(6));
1759                 intel_logical_ring_emit(ringbuf, 0);
1760                 intel_logical_ring_emit(ringbuf, 0);
1761                 intel_logical_ring_emit(ringbuf, 0);
1762                 intel_logical_ring_emit(ringbuf, 0);
1763                 intel_logical_ring_emit(ringbuf, 0);
1764         }
1765
1766         if (dc_flush_wa) {
1767                 intel_logical_ring_emit(ringbuf, GFX_OP_PIPE_CONTROL(6));
1768                 intel_logical_ring_emit(ringbuf, PIPE_CONTROL_DC_FLUSH_ENABLE);
1769                 intel_logical_ring_emit(ringbuf, 0);
1770                 intel_logical_ring_emit(ringbuf, 0);
1771                 intel_logical_ring_emit(ringbuf, 0);
1772                 intel_logical_ring_emit(ringbuf, 0);
1773         }
1774
1775         intel_logical_ring_emit(ringbuf, GFX_OP_PIPE_CONTROL(6));
1776         intel_logical_ring_emit(ringbuf, flags);
1777         intel_logical_ring_emit(ringbuf, scratch_addr);
1778         intel_logical_ring_emit(ringbuf, 0);
1779         intel_logical_ring_emit(ringbuf, 0);
1780         intel_logical_ring_emit(ringbuf, 0);
1781
1782         if (dc_flush_wa) {
1783                 intel_logical_ring_emit(ringbuf, GFX_OP_PIPE_CONTROL(6));
1784                 intel_logical_ring_emit(ringbuf, PIPE_CONTROL_CS_STALL);
1785                 intel_logical_ring_emit(ringbuf, 0);
1786                 intel_logical_ring_emit(ringbuf, 0);
1787                 intel_logical_ring_emit(ringbuf, 0);
1788                 intel_logical_ring_emit(ringbuf, 0);
1789         }
1790
1791         intel_logical_ring_advance(ringbuf);
1792
1793         return 0;
1794 }
1795
1796 static u32 gen8_get_seqno(struct intel_engine_cs *engine)
1797 {
1798         return intel_read_status_page(engine, I915_GEM_HWS_INDEX);
1799 }
1800
1801 static void gen8_set_seqno(struct intel_engine_cs *engine, u32 seqno)
1802 {
1803         intel_write_status_page(engine, I915_GEM_HWS_INDEX, seqno);
1804 }
1805
1806 static void bxt_a_seqno_barrier(struct intel_engine_cs *engine)
1807 {
1808         /*
1809          * On BXT A steppings there is a HW coherency issue whereby the
1810          * MI_STORE_DATA_IMM storing the completed request's seqno
1811          * occasionally doesn't invalidate the CPU cache. Work around this by
1812          * clflushing the corresponding cacheline whenever the caller wants
1813          * the coherency to be guaranteed. Note that this cacheline is known
1814          * to be clean at this point, since we only write it in
1815          * bxt_a_set_seqno(), where we also do a clflush after the write. So
1816          * this clflush in practice becomes an invalidate operation.
1817          */
1818         intel_flush_status_page(engine, I915_GEM_HWS_INDEX);
1819 }
1820
1821 static void bxt_a_set_seqno(struct intel_engine_cs *engine, u32 seqno)
1822 {
1823         intel_write_status_page(engine, I915_GEM_HWS_INDEX, seqno);
1824
1825         /* See bxt_a_get_seqno() explaining the reason for the clflush. */
1826         intel_flush_status_page(engine, I915_GEM_HWS_INDEX);
1827 }
1828
1829 /*
1830  * Reserve space for 2 NOOPs at the end of each request to be
1831  * used as a workaround for not being allowed to do lite
1832  * restore with HEAD==TAIL (WaIdleLiteRestore).
1833  */
1834 #define WA_TAIL_DWORDS 2
1835
1836 static inline u32 hws_seqno_address(struct intel_engine_cs *engine)
1837 {
1838         return engine->status_page.gfx_addr + I915_GEM_HWS_INDEX_ADDR;
1839 }
1840
1841 static int gen8_emit_request(struct drm_i915_gem_request *request)
1842 {
1843         struct intel_ringbuffer *ringbuf = request->ringbuf;
1844         int ret;
1845
1846         ret = intel_ring_begin(request, 6 + WA_TAIL_DWORDS);
1847         if (ret)
1848                 return ret;
1849
1850         /* w/a: bit 5 needs to be zero for MI_FLUSH_DW address. */
1851         BUILD_BUG_ON(I915_GEM_HWS_INDEX_ADDR & (1 << 5));
1852
1853         intel_logical_ring_emit(ringbuf,
1854                                 (MI_FLUSH_DW + 1) | MI_FLUSH_DW_OP_STOREDW);
1855         intel_logical_ring_emit(ringbuf,
1856                                 hws_seqno_address(request->engine) |
1857                                 MI_FLUSH_DW_USE_GTT);
1858         intel_logical_ring_emit(ringbuf, 0);
1859         intel_logical_ring_emit(ringbuf, i915_gem_request_get_seqno(request));
1860         intel_logical_ring_emit(ringbuf, MI_USER_INTERRUPT);
1861         intel_logical_ring_emit(ringbuf, MI_NOOP);
1862         return intel_logical_ring_advance_and_submit(request);
1863 }
1864
1865 static int gen8_emit_request_render(struct drm_i915_gem_request *request)
1866 {
1867         struct intel_ringbuffer *ringbuf = request->ringbuf;
1868         int ret;
1869
1870         ret = intel_ring_begin(request, 8 + WA_TAIL_DWORDS);
1871         if (ret)
1872                 return ret;
1873
1874         /* We're using qword write, seqno should be aligned to 8 bytes. */
1875         BUILD_BUG_ON(I915_GEM_HWS_INDEX & 1);
1876
1877         /* w/a for post sync ops following a GPGPU operation we
1878          * need a prior CS_STALL, which is emitted by the flush
1879          * following the batch.
1880          */
1881         intel_logical_ring_emit(ringbuf, GFX_OP_PIPE_CONTROL(6));
1882         intel_logical_ring_emit(ringbuf,
1883                                 (PIPE_CONTROL_GLOBAL_GTT_IVB |
1884                                  PIPE_CONTROL_CS_STALL |
1885                                  PIPE_CONTROL_QW_WRITE));
1886         intel_logical_ring_emit(ringbuf, hws_seqno_address(request->engine));
1887         intel_logical_ring_emit(ringbuf, 0);
1888         intel_logical_ring_emit(ringbuf, i915_gem_request_get_seqno(request));
1889         /* We're thrashing one dword of HWS. */
1890         intel_logical_ring_emit(ringbuf, 0);
1891         intel_logical_ring_emit(ringbuf, MI_USER_INTERRUPT);
1892         intel_logical_ring_emit(ringbuf, MI_NOOP);
1893         return intel_logical_ring_advance_and_submit(request);
1894 }
1895
1896 static int intel_lr_context_render_state_init(struct drm_i915_gem_request *req)
1897 {
1898         struct render_state so;
1899         int ret;
1900
1901         ret = i915_gem_render_state_prepare(req->engine, &so);
1902         if (ret)
1903                 return ret;
1904
1905         if (so.rodata == NULL)
1906                 return 0;
1907
1908         ret = req->engine->emit_bb_start(req, so.ggtt_offset,
1909                                        I915_DISPATCH_SECURE);
1910         if (ret)
1911                 goto out;
1912
1913         ret = req->engine->emit_bb_start(req,
1914                                        (so.ggtt_offset + so.aux_batch_offset),
1915                                        I915_DISPATCH_SECURE);
1916         if (ret)
1917                 goto out;
1918
1919         i915_vma_move_to_active(i915_gem_obj_to_ggtt(so.obj), req);
1920
1921 out:
1922         i915_gem_render_state_fini(&so);
1923         return ret;
1924 }
1925
1926 static int gen8_init_rcs_context(struct drm_i915_gem_request *req)
1927 {
1928         int ret;
1929
1930         ret = intel_logical_ring_workarounds_emit(req);
1931         if (ret)
1932                 return ret;
1933
1934         ret = intel_rcs_context_init_mocs(req);
1935         /*
1936          * Failing to program the MOCS is non-fatal.The system will not
1937          * run at peak performance. So generate an error and carry on.
1938          */
1939         if (ret)
1940                 DRM_ERROR("MOCS failed to program: expect performance issues.\n");
1941
1942         return intel_lr_context_render_state_init(req);
1943 }
1944
1945 /**
1946  * intel_logical_ring_cleanup() - deallocate the Engine Command Streamer
1947  *
1948  * @ring: Engine Command Streamer.
1949  *
1950  */
1951 void intel_logical_ring_cleanup(struct intel_engine_cs *engine)
1952 {
1953         struct drm_i915_private *dev_priv;
1954
1955         if (!intel_engine_initialized(engine))
1956                 return;
1957
1958         /*
1959          * Tasklet cannot be active at this point due intel_mark_active/idle
1960          * so this is just for documentation.
1961          */
1962         if (WARN_ON(test_bit(TASKLET_STATE_SCHED, &engine->irq_tasklet.state)))
1963                 tasklet_kill(&engine->irq_tasklet);
1964
1965         dev_priv = engine->dev->dev_private;
1966
1967         if (engine->buffer) {
1968                 intel_logical_ring_stop(engine);
1969                 WARN_ON((I915_READ_MODE(engine) & MODE_IDLE) == 0);
1970         }
1971
1972         if (engine->cleanup)
1973                 engine->cleanup(engine);
1974
1975         i915_cmd_parser_fini_ring(engine);
1976         i915_gem_batch_pool_fini(&engine->batch_pool);
1977
1978         if (engine->status_page.obj) {
1979                 i915_gem_object_unpin_map(engine->status_page.obj);
1980                 engine->status_page.obj = NULL;
1981         }
1982
1983         engine->idle_lite_restore_wa = 0;
1984         engine->disable_lite_restore_wa = false;
1985         engine->ctx_desc_template = 0;
1986
1987         lrc_destroy_wa_ctx_obj(engine);
1988         engine->dev = NULL;
1989 }
1990
1991 static void
1992 logical_ring_default_vfuncs(struct drm_device *dev,
1993                             struct intel_engine_cs *engine)
1994 {
1995         /* Default vfuncs which can be overriden by each engine. */
1996         engine->init_hw = gen8_init_common_ring;
1997         engine->emit_request = gen8_emit_request;
1998         engine->emit_flush = gen8_emit_flush;
1999         engine->irq_get = gen8_logical_ring_get_irq;
2000         engine->irq_put = gen8_logical_ring_put_irq;
2001         engine->emit_bb_start = gen8_emit_bb_start;
2002         engine->get_seqno = gen8_get_seqno;
2003         engine->set_seqno = gen8_set_seqno;
2004         if (IS_BXT_REVID(dev, 0, BXT_REVID_A1)) {
2005                 engine->irq_seqno_barrier = bxt_a_seqno_barrier;
2006                 engine->set_seqno = bxt_a_set_seqno;
2007         }
2008 }
2009
2010 static inline void
2011 logical_ring_default_irqs(struct intel_engine_cs *engine, unsigned shift)
2012 {
2013         engine->irq_enable_mask = GT_RENDER_USER_INTERRUPT << shift;
2014         engine->irq_keep_mask = GT_CONTEXT_SWITCH_INTERRUPT << shift;
2015 }
2016
2017 static int
2018 lrc_setup_hws(struct intel_engine_cs *engine,
2019               struct drm_i915_gem_object *dctx_obj)
2020 {
2021         void *hws;
2022
2023         /* The HWSP is part of the default context object in LRC mode. */
2024         engine->status_page.gfx_addr = i915_gem_obj_ggtt_offset(dctx_obj) +
2025                                        LRC_PPHWSP_PN * PAGE_SIZE;
2026         hws = i915_gem_object_pin_map(dctx_obj);
2027         if (IS_ERR(hws))
2028                 return PTR_ERR(hws);
2029         engine->status_page.page_addr = hws + LRC_PPHWSP_PN * PAGE_SIZE;
2030         engine->status_page.obj = dctx_obj;
2031
2032         return 0;
2033 }
2034
2035 static int
2036 logical_ring_init(struct drm_device *dev, struct intel_engine_cs *engine)
2037 {
2038         struct drm_i915_private *dev_priv = to_i915(dev);
2039         struct intel_context *dctx = dev_priv->kernel_context;
2040         enum forcewake_domains fw_domains;
2041         int ret;
2042
2043         /* Intentionally left blank. */
2044         engine->buffer = NULL;
2045
2046         engine->dev = dev;
2047         INIT_LIST_HEAD(&engine->active_list);
2048         INIT_LIST_HEAD(&engine->request_list);
2049         i915_gem_batch_pool_init(dev, &engine->batch_pool);
2050         init_waitqueue_head(&engine->irq_queue);
2051
2052         INIT_LIST_HEAD(&engine->buffers);
2053         INIT_LIST_HEAD(&engine->execlist_queue);
2054         INIT_LIST_HEAD(&engine->execlist_retired_req_list);
2055         spin_lock_init(&engine->execlist_lock);
2056
2057         tasklet_init(&engine->irq_tasklet,
2058                      intel_lrc_irq_handler, (unsigned long)engine);
2059
2060         logical_ring_init_platform_invariants(engine);
2061
2062         fw_domains = intel_uncore_forcewake_for_reg(dev_priv,
2063                                                     RING_ELSP(engine),
2064                                                     FW_REG_WRITE);
2065
2066         fw_domains |= intel_uncore_forcewake_for_reg(dev_priv,
2067                                                      RING_CONTEXT_STATUS_PTR(engine),
2068                                                      FW_REG_READ | FW_REG_WRITE);
2069
2070         fw_domains |= intel_uncore_forcewake_for_reg(dev_priv,
2071                                                      RING_CONTEXT_STATUS_BUF_BASE(engine),
2072                                                      FW_REG_READ);
2073
2074         engine->fw_domains = fw_domains;
2075
2076         ret = i915_cmd_parser_init_ring(engine);
2077         if (ret)
2078                 goto error;
2079
2080         ret = intel_lr_context_deferred_alloc(dctx, engine);
2081         if (ret)
2082                 goto error;
2083
2084         /* As this is the default context, always pin it */
2085         ret = intel_lr_context_do_pin(dctx, engine);
2086         if (ret) {
2087                 DRM_ERROR(
2088                         "Failed to pin and map ringbuffer %s: %d\n",
2089                         engine->name, ret);
2090                 goto error;
2091         }
2092
2093         /* And setup the hardware status page. */
2094         ret = lrc_setup_hws(engine, dctx->engine[engine->id].state);
2095         if (ret) {
2096                 DRM_ERROR("Failed to set up hws %s: %d\n", engine->name, ret);
2097                 goto error;
2098         }
2099
2100         return 0;
2101
2102 error:
2103         intel_logical_ring_cleanup(engine);
2104         return ret;
2105 }
2106
2107 static int logical_render_ring_init(struct drm_device *dev)
2108 {
2109         struct drm_i915_private *dev_priv = dev->dev_private;
2110         struct intel_engine_cs *engine = &dev_priv->engine[RCS];
2111         int ret;
2112
2113         engine->name = "render ring";
2114         engine->id = RCS;
2115         engine->exec_id = I915_EXEC_RENDER;
2116         engine->guc_id = GUC_RENDER_ENGINE;
2117         engine->mmio_base = RENDER_RING_BASE;
2118
2119         logical_ring_default_irqs(engine, GEN8_RCS_IRQ_SHIFT);
2120         if (HAS_L3_DPF(dev))
2121                 engine->irq_keep_mask |= GT_RENDER_L3_PARITY_ERROR_INTERRUPT;
2122
2123         logical_ring_default_vfuncs(dev, engine);
2124
2125         /* Override some for render ring. */
2126         if (INTEL_INFO(dev)->gen >= 9)
2127                 engine->init_hw = gen9_init_render_ring;
2128         else
2129                 engine->init_hw = gen8_init_render_ring;
2130         engine->init_context = gen8_init_rcs_context;
2131         engine->cleanup = intel_fini_pipe_control;
2132         engine->emit_flush = gen8_emit_flush_render;
2133         engine->emit_request = gen8_emit_request_render;
2134
2135         engine->dev = dev;
2136
2137         ret = intel_init_pipe_control(engine);
2138         if (ret)
2139                 return ret;
2140
2141         ret = intel_init_workaround_bb(engine);
2142         if (ret) {
2143                 /*
2144                  * We continue even if we fail to initialize WA batch
2145                  * because we only expect rare glitches but nothing
2146                  * critical to prevent us from using GPU
2147                  */
2148                 DRM_ERROR("WA batch buffer initialization failed: %d\n",
2149                           ret);
2150         }
2151
2152         ret = logical_ring_init(dev, engine);
2153         if (ret) {
2154                 lrc_destroy_wa_ctx_obj(engine);
2155         }
2156
2157         return ret;
2158 }
2159
2160 static int logical_bsd_ring_init(struct drm_device *dev)
2161 {
2162         struct drm_i915_private *dev_priv = dev->dev_private;
2163         struct intel_engine_cs *engine = &dev_priv->engine[VCS];
2164
2165         engine->name = "bsd ring";
2166         engine->id = VCS;
2167         engine->exec_id = I915_EXEC_BSD;
2168         engine->guc_id = GUC_VIDEO_ENGINE;
2169         engine->mmio_base = GEN6_BSD_RING_BASE;
2170
2171         logical_ring_default_irqs(engine, GEN8_VCS1_IRQ_SHIFT);
2172         logical_ring_default_vfuncs(dev, engine);
2173
2174         return logical_ring_init(dev, engine);
2175 }
2176
2177 static int logical_bsd2_ring_init(struct drm_device *dev)
2178 {
2179         struct drm_i915_private *dev_priv = dev->dev_private;
2180         struct intel_engine_cs *engine = &dev_priv->engine[VCS2];
2181
2182         engine->name = "bsd2 ring";
2183         engine->id = VCS2;
2184         engine->exec_id = I915_EXEC_BSD;
2185         engine->guc_id = GUC_VIDEO_ENGINE2;
2186         engine->mmio_base = GEN8_BSD2_RING_BASE;
2187
2188         logical_ring_default_irqs(engine, GEN8_VCS2_IRQ_SHIFT);
2189         logical_ring_default_vfuncs(dev, engine);
2190
2191         return logical_ring_init(dev, engine);
2192 }
2193
2194 static int logical_blt_ring_init(struct drm_device *dev)
2195 {
2196         struct drm_i915_private *dev_priv = dev->dev_private;
2197         struct intel_engine_cs *engine = &dev_priv->engine[BCS];
2198
2199         engine->name = "blitter ring";
2200         engine->id = BCS;
2201         engine->exec_id = I915_EXEC_BLT;
2202         engine->guc_id = GUC_BLITTER_ENGINE;
2203         engine->mmio_base = BLT_RING_BASE;
2204
2205         logical_ring_default_irqs(engine, GEN8_BCS_IRQ_SHIFT);
2206         logical_ring_default_vfuncs(dev, engine);
2207
2208         return logical_ring_init(dev, engine);
2209 }
2210
2211 static int logical_vebox_ring_init(struct drm_device *dev)
2212 {
2213         struct drm_i915_private *dev_priv = dev->dev_private;
2214         struct intel_engine_cs *engine = &dev_priv->engine[VECS];
2215
2216         engine->name = "video enhancement ring";
2217         engine->id = VECS;
2218         engine->exec_id = I915_EXEC_VEBOX;
2219         engine->guc_id = GUC_VIDEOENHANCE_ENGINE;
2220         engine->mmio_base = VEBOX_RING_BASE;
2221
2222         logical_ring_default_irqs(engine, GEN8_VECS_IRQ_SHIFT);
2223         logical_ring_default_vfuncs(dev, engine);
2224
2225         return logical_ring_init(dev, engine);
2226 }
2227
2228 /**
2229  * intel_logical_rings_init() - allocate, populate and init the Engine Command Streamers
2230  * @dev: DRM device.
2231  *
2232  * This function inits the engines for an Execlists submission style (the equivalent in the
2233  * legacy ringbuffer submission world would be i915_gem_init_engines). It does it only for
2234  * those engines that are present in the hardware.
2235  *
2236  * Return: non-zero if the initialization failed.
2237  */
2238 int intel_logical_rings_init(struct drm_device *dev)
2239 {
2240         struct drm_i915_private *dev_priv = dev->dev_private;
2241         int ret;
2242
2243         ret = logical_render_ring_init(dev);
2244         if (ret)
2245                 return ret;
2246
2247         if (HAS_BSD(dev)) {
2248                 ret = logical_bsd_ring_init(dev);
2249                 if (ret)
2250                         goto cleanup_render_ring;
2251         }
2252
2253         if (HAS_BLT(dev)) {
2254                 ret = logical_blt_ring_init(dev);
2255                 if (ret)
2256                         goto cleanup_bsd_ring;
2257         }
2258
2259         if (HAS_VEBOX(dev)) {
2260                 ret = logical_vebox_ring_init(dev);
2261                 if (ret)
2262                         goto cleanup_blt_ring;
2263         }
2264
2265         if (HAS_BSD2(dev)) {
2266                 ret = logical_bsd2_ring_init(dev);
2267                 if (ret)
2268                         goto cleanup_vebox_ring;
2269         }
2270
2271         return 0;
2272
2273 cleanup_vebox_ring:
2274         intel_logical_ring_cleanup(&dev_priv->engine[VECS]);
2275 cleanup_blt_ring:
2276         intel_logical_ring_cleanup(&dev_priv->engine[BCS]);
2277 cleanup_bsd_ring:
2278         intel_logical_ring_cleanup(&dev_priv->engine[VCS]);
2279 cleanup_render_ring:
2280         intel_logical_ring_cleanup(&dev_priv->engine[RCS]);
2281
2282         return ret;
2283 }
2284
2285 static u32
2286 make_rpcs(struct drm_device *dev)
2287 {
2288         u32 rpcs = 0;
2289
2290         /*
2291          * No explicit RPCS request is needed to ensure full
2292          * slice/subslice/EU enablement prior to Gen9.
2293         */
2294         if (INTEL_INFO(dev)->gen < 9)
2295                 return 0;
2296
2297         /*
2298          * Starting in Gen9, render power gating can leave
2299          * slice/subslice/EU in a partially enabled state. We
2300          * must make an explicit request through RPCS for full
2301          * enablement.
2302         */
2303         if (INTEL_INFO(dev)->has_slice_pg) {
2304                 rpcs |= GEN8_RPCS_S_CNT_ENABLE;
2305                 rpcs |= INTEL_INFO(dev)->slice_total <<
2306                         GEN8_RPCS_S_CNT_SHIFT;
2307                 rpcs |= GEN8_RPCS_ENABLE;
2308         }
2309
2310         if (INTEL_INFO(dev)->has_subslice_pg) {
2311                 rpcs |= GEN8_RPCS_SS_CNT_ENABLE;
2312                 rpcs |= INTEL_INFO(dev)->subslice_per_slice <<
2313                         GEN8_RPCS_SS_CNT_SHIFT;
2314                 rpcs |= GEN8_RPCS_ENABLE;
2315         }
2316
2317         if (INTEL_INFO(dev)->has_eu_pg) {
2318                 rpcs |= INTEL_INFO(dev)->eu_per_subslice <<
2319                         GEN8_RPCS_EU_MIN_SHIFT;
2320                 rpcs |= INTEL_INFO(dev)->eu_per_subslice <<
2321                         GEN8_RPCS_EU_MAX_SHIFT;
2322                 rpcs |= GEN8_RPCS_ENABLE;
2323         }
2324
2325         return rpcs;
2326 }
2327
2328 static u32 intel_lr_indirect_ctx_offset(struct intel_engine_cs *engine)
2329 {
2330         u32 indirect_ctx_offset;
2331
2332         switch (INTEL_INFO(engine->dev)->gen) {
2333         default:
2334                 MISSING_CASE(INTEL_INFO(engine->dev)->gen);
2335                 /* fall through */
2336         case 9:
2337                 indirect_ctx_offset =
2338                         GEN9_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT;
2339                 break;
2340         case 8:
2341                 indirect_ctx_offset =
2342                         GEN8_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT;
2343                 break;
2344         }
2345
2346         return indirect_ctx_offset;
2347 }
2348
2349 static int
2350 populate_lr_context(struct intel_context *ctx,
2351                     struct drm_i915_gem_object *ctx_obj,
2352                     struct intel_engine_cs *engine,
2353                     struct intel_ringbuffer *ringbuf)
2354 {
2355         struct drm_device *dev = engine->dev;
2356         struct drm_i915_private *dev_priv = dev->dev_private;
2357         struct i915_hw_ppgtt *ppgtt = ctx->ppgtt;
2358         void *vaddr;
2359         u32 *reg_state;
2360         int ret;
2361
2362         if (!ppgtt)
2363                 ppgtt = dev_priv->mm.aliasing_ppgtt;
2364
2365         ret = i915_gem_object_set_to_cpu_domain(ctx_obj, true);
2366         if (ret) {
2367                 DRM_DEBUG_DRIVER("Could not set to CPU domain\n");
2368                 return ret;
2369         }
2370
2371         vaddr = i915_gem_object_pin_map(ctx_obj);
2372         if (IS_ERR(vaddr)) {
2373                 ret = PTR_ERR(vaddr);
2374                 DRM_DEBUG_DRIVER("Could not map object pages! (%d)\n", ret);
2375                 return ret;
2376         }
2377         ctx_obj->dirty = true;
2378
2379         /* The second page of the context object contains some fields which must
2380          * be set up prior to the first execution. */
2381         reg_state = vaddr + LRC_STATE_PN * PAGE_SIZE;
2382
2383         /* A context is actually a big batch buffer with several MI_LOAD_REGISTER_IMM
2384          * commands followed by (reg, value) pairs. The values we are setting here are
2385          * only for the first context restore: on a subsequent save, the GPU will
2386          * recreate this batchbuffer with new values (including all the missing
2387          * MI_LOAD_REGISTER_IMM commands that we are not initializing here). */
2388         reg_state[CTX_LRI_HEADER_0] =
2389                 MI_LOAD_REGISTER_IMM(engine->id == RCS ? 14 : 11) | MI_LRI_FORCE_POSTED;
2390         ASSIGN_CTX_REG(reg_state, CTX_CONTEXT_CONTROL,
2391                        RING_CONTEXT_CONTROL(engine),
2392                        _MASKED_BIT_ENABLE(CTX_CTRL_INHIBIT_SYN_CTX_SWITCH |
2393                                           CTX_CTRL_ENGINE_CTX_RESTORE_INHIBIT |
2394                                           (HAS_RESOURCE_STREAMER(dev) ?
2395                                             CTX_CTRL_RS_CTX_ENABLE : 0)));
2396         ASSIGN_CTX_REG(reg_state, CTX_RING_HEAD, RING_HEAD(engine->mmio_base),
2397                        0);
2398         ASSIGN_CTX_REG(reg_state, CTX_RING_TAIL, RING_TAIL(engine->mmio_base),
2399                        0);
2400         /* Ring buffer start address is not known until the buffer is pinned.
2401          * It is written to the context image in execlists_update_context()
2402          */
2403         ASSIGN_CTX_REG(reg_state, CTX_RING_BUFFER_START,
2404                        RING_START(engine->mmio_base), 0);
2405         ASSIGN_CTX_REG(reg_state, CTX_RING_BUFFER_CONTROL,
2406                        RING_CTL(engine->mmio_base),
2407                        ((ringbuf->size - PAGE_SIZE) & RING_NR_PAGES) | RING_VALID);
2408         ASSIGN_CTX_REG(reg_state, CTX_BB_HEAD_U,
2409                        RING_BBADDR_UDW(engine->mmio_base), 0);
2410         ASSIGN_CTX_REG(reg_state, CTX_BB_HEAD_L,
2411                        RING_BBADDR(engine->mmio_base), 0);
2412         ASSIGN_CTX_REG(reg_state, CTX_BB_STATE,
2413                        RING_BBSTATE(engine->mmio_base),
2414                        RING_BB_PPGTT);
2415         ASSIGN_CTX_REG(reg_state, CTX_SECOND_BB_HEAD_U,
2416                        RING_SBBADDR_UDW(engine->mmio_base), 0);
2417         ASSIGN_CTX_REG(reg_state, CTX_SECOND_BB_HEAD_L,
2418                        RING_SBBADDR(engine->mmio_base), 0);
2419         ASSIGN_CTX_REG(reg_state, CTX_SECOND_BB_STATE,
2420                        RING_SBBSTATE(engine->mmio_base), 0);
2421         if (engine->id == RCS) {
2422                 ASSIGN_CTX_REG(reg_state, CTX_BB_PER_CTX_PTR,
2423                                RING_BB_PER_CTX_PTR(engine->mmio_base), 0);
2424                 ASSIGN_CTX_REG(reg_state, CTX_RCS_INDIRECT_CTX,
2425                                RING_INDIRECT_CTX(engine->mmio_base), 0);
2426                 ASSIGN_CTX_REG(reg_state, CTX_RCS_INDIRECT_CTX_OFFSET,
2427                                RING_INDIRECT_CTX_OFFSET(engine->mmio_base), 0);
2428                 if (engine->wa_ctx.obj) {
2429                         struct i915_ctx_workarounds *wa_ctx = &engine->wa_ctx;
2430                         uint32_t ggtt_offset = i915_gem_obj_ggtt_offset(wa_ctx->obj);
2431
2432                         reg_state[CTX_RCS_INDIRECT_CTX+1] =
2433                                 (ggtt_offset + wa_ctx->indirect_ctx.offset * sizeof(uint32_t)) |
2434                                 (wa_ctx->indirect_ctx.size / CACHELINE_DWORDS);
2435
2436                         reg_state[CTX_RCS_INDIRECT_CTX_OFFSET+1] =
2437                                 intel_lr_indirect_ctx_offset(engine) << 6;
2438
2439                         reg_state[CTX_BB_PER_CTX_PTR+1] =
2440                                 (ggtt_offset + wa_ctx->per_ctx.offset * sizeof(uint32_t)) |
2441                                 0x01;
2442                 }
2443         }
2444         reg_state[CTX_LRI_HEADER_1] = MI_LOAD_REGISTER_IMM(9) | MI_LRI_FORCE_POSTED;
2445         ASSIGN_CTX_REG(reg_state, CTX_CTX_TIMESTAMP,
2446                        RING_CTX_TIMESTAMP(engine->mmio_base), 0);
2447         /* PDP values well be assigned later if needed */
2448         ASSIGN_CTX_REG(reg_state, CTX_PDP3_UDW, GEN8_RING_PDP_UDW(engine, 3),
2449                        0);
2450         ASSIGN_CTX_REG(reg_state, CTX_PDP3_LDW, GEN8_RING_PDP_LDW(engine, 3),
2451                        0);
2452         ASSIGN_CTX_REG(reg_state, CTX_PDP2_UDW, GEN8_RING_PDP_UDW(engine, 2),
2453                        0);
2454         ASSIGN_CTX_REG(reg_state, CTX_PDP2_LDW, GEN8_RING_PDP_LDW(engine, 2),
2455                        0);
2456         ASSIGN_CTX_REG(reg_state, CTX_PDP1_UDW, GEN8_RING_PDP_UDW(engine, 1),
2457                        0);
2458         ASSIGN_CTX_REG(reg_state, CTX_PDP1_LDW, GEN8_RING_PDP_LDW(engine, 1),
2459                        0);
2460         ASSIGN_CTX_REG(reg_state, CTX_PDP0_UDW, GEN8_RING_PDP_UDW(engine, 0),
2461                        0);
2462         ASSIGN_CTX_REG(reg_state, CTX_PDP0_LDW, GEN8_RING_PDP_LDW(engine, 0),
2463                        0);
2464
2465         if (USES_FULL_48BIT_PPGTT(ppgtt->base.dev)) {
2466                 /* 64b PPGTT (48bit canonical)
2467                  * PDP0_DESCRIPTOR contains the base address to PML4 and
2468                  * other PDP Descriptors are ignored.
2469                  */
2470                 ASSIGN_CTX_PML4(ppgtt, reg_state);
2471         } else {
2472                 /* 32b PPGTT
2473                  * PDP*_DESCRIPTOR contains the base address of space supported.
2474                  * With dynamic page allocation, PDPs may not be allocated at
2475                  * this point. Point the unallocated PDPs to the scratch page
2476                  */
2477                 execlists_update_context_pdps(ppgtt, reg_state);
2478         }
2479
2480         if (engine->id == RCS) {
2481                 reg_state[CTX_LRI_HEADER_2] = MI_LOAD_REGISTER_IMM(1);
2482                 ASSIGN_CTX_REG(reg_state, CTX_R_PWR_CLK_STATE, GEN8_R_PWR_CLK_STATE,
2483                                make_rpcs(dev));
2484         }
2485
2486         i915_gem_object_unpin_map(ctx_obj);
2487
2488         return 0;
2489 }
2490
2491 /**
2492  * intel_lr_context_free() - free the LRC specific bits of a context
2493  * @ctx: the LR context to free.
2494  *
2495  * The real context freeing is done in i915_gem_context_free: this only
2496  * takes care of the bits that are LRC related: the per-engine backing
2497  * objects and the logical ringbuffer.
2498  */
2499 void intel_lr_context_free(struct intel_context *ctx)
2500 {
2501         int i;
2502
2503         for (i = I915_NUM_ENGINES; --i >= 0; ) {
2504                 struct intel_ringbuffer *ringbuf = ctx->engine[i].ringbuf;
2505                 struct drm_i915_gem_object *ctx_obj = ctx->engine[i].state;
2506
2507                 if (!ctx_obj)
2508                         continue;
2509
2510                 if (ctx == ctx->i915->kernel_context) {
2511                         intel_unpin_ringbuffer_obj(ringbuf);
2512                         i915_gem_object_ggtt_unpin(ctx_obj);
2513                         i915_gem_object_unpin_map(ctx_obj);
2514                 }
2515
2516                 WARN_ON(ctx->engine[i].pin_count);
2517                 intel_ringbuffer_free(ringbuf);
2518                 drm_gem_object_unreference(&ctx_obj->base);
2519         }
2520 }
2521
2522 /**
2523  * intel_lr_context_size() - return the size of the context for an engine
2524  * @ring: which engine to find the context size for
2525  *
2526  * Each engine may require a different amount of space for a context image,
2527  * so when allocating (or copying) an image, this function can be used to
2528  * find the right size for the specific engine.
2529  *
2530  * Return: size (in bytes) of an engine-specific context image
2531  *
2532  * Note: this size includes the HWSP, which is part of the context image
2533  * in LRC mode, but does not include the "shared data page" used with
2534  * GuC submission. The caller should account for this if using the GuC.
2535  */
2536 uint32_t intel_lr_context_size(struct intel_engine_cs *engine)
2537 {
2538         int ret = 0;
2539
2540         WARN_ON(INTEL_INFO(engine->dev)->gen < 8);
2541
2542         switch (engine->id) {
2543         case RCS:
2544                 if (INTEL_INFO(engine->dev)->gen >= 9)
2545                         ret = GEN9_LR_CONTEXT_RENDER_SIZE;
2546                 else
2547                         ret = GEN8_LR_CONTEXT_RENDER_SIZE;
2548                 break;
2549         case VCS:
2550         case BCS:
2551         case VECS:
2552         case VCS2:
2553                 ret = GEN8_LR_CONTEXT_OTHER_SIZE;
2554                 break;
2555         }
2556
2557         return ret;
2558 }
2559
2560 /**
2561  * intel_lr_context_deferred_alloc() - create the LRC specific bits of a context
2562  * @ctx: LR context to create.
2563  * @ring: engine to be used with the context.
2564  *
2565  * This function can be called more than once, with different engines, if we plan
2566  * to use the context with them. The context backing objects and the ringbuffers
2567  * (specially the ringbuffer backing objects) suck a lot of memory up, and that's why
2568  * the creation is a deferred call: it's better to make sure first that we need to use
2569  * a given ring with the context.
2570  *
2571  * Return: non-zero on error.
2572  */
2573
2574 int intel_lr_context_deferred_alloc(struct intel_context *ctx,
2575                                     struct intel_engine_cs *engine)
2576 {
2577         struct drm_device *dev = engine->dev;
2578         struct drm_i915_gem_object *ctx_obj;
2579         uint32_t context_size;
2580         struct intel_ringbuffer *ringbuf;
2581         int ret;
2582
2583         WARN_ON(ctx->legacy_hw_ctx.rcs_state != NULL);
2584         WARN_ON(ctx->engine[engine->id].state);
2585
2586         context_size = round_up(intel_lr_context_size(engine), 4096);
2587
2588         /* One extra page as the sharing data between driver and GuC */
2589         context_size += PAGE_SIZE * LRC_PPHWSP_PN;
2590
2591         ctx_obj = i915_gem_alloc_object(dev, context_size);
2592         if (!ctx_obj) {
2593                 DRM_DEBUG_DRIVER("Alloc LRC backing obj failed.\n");
2594                 return -ENOMEM;
2595         }
2596
2597         ringbuf = intel_engine_create_ringbuffer(engine, 4 * PAGE_SIZE);
2598         if (IS_ERR(ringbuf)) {
2599                 ret = PTR_ERR(ringbuf);
2600                 goto error_deref_obj;
2601         }
2602
2603         ret = populate_lr_context(ctx, ctx_obj, engine, ringbuf);
2604         if (ret) {
2605                 DRM_DEBUG_DRIVER("Failed to populate LRC: %d\n", ret);
2606                 goto error_ringbuf;
2607         }
2608
2609         ctx->engine[engine->id].ringbuf = ringbuf;
2610         ctx->engine[engine->id].state = ctx_obj;
2611
2612         if (ctx != ctx->i915->kernel_context && engine->init_context) {
2613                 struct drm_i915_gem_request *req;
2614
2615                 req = i915_gem_request_alloc(engine, ctx);
2616                 if (IS_ERR(req)) {
2617                         ret = PTR_ERR(req);
2618                         DRM_ERROR("ring create req: %d\n", ret);
2619                         goto error_ringbuf;
2620                 }
2621
2622                 ret = engine->init_context(req);
2623                 i915_add_request_no_flush(req);
2624                 if (ret) {
2625                         DRM_ERROR("ring init context: %d\n",
2626                                 ret);
2627                         goto error_ringbuf;
2628                 }
2629         }
2630         return 0;
2631
2632 error_ringbuf:
2633         intel_ringbuffer_free(ringbuf);
2634 error_deref_obj:
2635         drm_gem_object_unreference(&ctx_obj->base);
2636         ctx->engine[engine->id].ringbuf = NULL;
2637         ctx->engine[engine->id].state = NULL;
2638         return ret;
2639 }
2640
2641 void intel_lr_context_reset(struct drm_i915_private *dev_priv,
2642                             struct intel_context *ctx)
2643 {
2644         struct intel_engine_cs *engine;
2645
2646         for_each_engine(engine, dev_priv) {
2647                 struct drm_i915_gem_object *ctx_obj =
2648                                 ctx->engine[engine->id].state;
2649                 struct intel_ringbuffer *ringbuf =
2650                                 ctx->engine[engine->id].ringbuf;
2651                 void *vaddr;
2652                 uint32_t *reg_state;
2653
2654                 if (!ctx_obj)
2655                         continue;
2656
2657                 vaddr = i915_gem_object_pin_map(ctx_obj);
2658                 if (WARN_ON(IS_ERR(vaddr)))
2659                         continue;
2660
2661                 reg_state = vaddr + LRC_STATE_PN * PAGE_SIZE;
2662                 ctx_obj->dirty = true;
2663
2664                 reg_state[CTX_RING_HEAD+1] = 0;
2665                 reg_state[CTX_RING_TAIL+1] = 0;
2666
2667                 i915_gem_object_unpin_map(ctx_obj);
2668
2669                 ringbuf->head = 0;
2670                 ringbuf->tail = 0;
2671         }
2672 }