Commit | Line | Data |
---|---|---|
b20385f1 OM |
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 | ||
73e4d07f OM |
31 | /** |
32 | * DOC: Logical Rings, Logical Ring Contexts and Execlists | |
33 | * | |
34 | * Motivation: | |
b20385f1 OM |
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 | * | |
73e4d07f OM |
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: | |
b20385f1 OM |
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). | |
73e4d07f OM |
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 | * | |
b20385f1 | 133 | */ |
27af5eea | 134 | #include <linux/interrupt.h> |
b20385f1 OM |
135 | |
136 | #include <drm/drmP.h> | |
137 | #include <drm/i915_drm.h> | |
138 | #include "i915_drv.h" | |
3bbaba0c | 139 | #include "intel_mocs.h" |
127f1003 | 140 | |
468c6816 | 141 | #define GEN9_LR_CONTEXT_RENDER_SIZE (22 * PAGE_SIZE) |
8c857917 OM |
142 | #define GEN8_LR_CONTEXT_RENDER_SIZE (20 * PAGE_SIZE) |
143 | #define GEN8_LR_CONTEXT_OTHER_SIZE (2 * PAGE_SIZE) | |
144 | ||
e981e7b1 TD |
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) | |
8670d6f9 | 158 | |
70c2a24d CW |
159 | #define GEN8_CTX_STATUS_COMPLETED_MASK \ |
160 | (GEN8_CTX_STATUS_ACTIVE_IDLE | \ | |
161 | GEN8_CTX_STATUS_PREEMPTED | \ | |
162 | GEN8_CTX_STATUS_ELEMENT_SWITCH) | |
163 | ||
8670d6f9 OM |
164 | #define CTX_LRI_HEADER_0 0x01 |
165 | #define CTX_CONTEXT_CONTROL 0x02 | |
166 | #define CTX_RING_HEAD 0x04 | |
167 | #define CTX_RING_TAIL 0x06 | |
168 | #define CTX_RING_BUFFER_START 0x08 | |
169 | #define CTX_RING_BUFFER_CONTROL 0x0a | |
170 | #define CTX_BB_HEAD_U 0x0c | |
171 | #define CTX_BB_HEAD_L 0x0e | |
172 | #define CTX_BB_STATE 0x10 | |
173 | #define CTX_SECOND_BB_HEAD_U 0x12 | |
174 | #define CTX_SECOND_BB_HEAD_L 0x14 | |
175 | #define CTX_SECOND_BB_STATE 0x16 | |
176 | #define CTX_BB_PER_CTX_PTR 0x18 | |
177 | #define CTX_RCS_INDIRECT_CTX 0x1a | |
178 | #define CTX_RCS_INDIRECT_CTX_OFFSET 0x1c | |
179 | #define CTX_LRI_HEADER_1 0x21 | |
180 | #define CTX_CTX_TIMESTAMP 0x22 | |
181 | #define CTX_PDP3_UDW 0x24 | |
182 | #define CTX_PDP3_LDW 0x26 | |
183 | #define CTX_PDP2_UDW 0x28 | |
184 | #define CTX_PDP2_LDW 0x2a | |
185 | #define CTX_PDP1_UDW 0x2c | |
186 | #define CTX_PDP1_LDW 0x2e | |
187 | #define CTX_PDP0_UDW 0x30 | |
188 | #define CTX_PDP0_LDW 0x32 | |
189 | #define CTX_LRI_HEADER_2 0x41 | |
190 | #define CTX_R_PWR_CLK_STATE 0x42 | |
191 | #define CTX_GPGPU_CSR_BASE_ADDRESS 0x44 | |
192 | ||
84b790f8 BW |
193 | #define GEN8_CTX_VALID (1<<0) |
194 | #define GEN8_CTX_FORCE_PD_RESTORE (1<<1) | |
195 | #define GEN8_CTX_FORCE_RESTORE (1<<2) | |
196 | #define GEN8_CTX_L3LLC_COHERENT (1<<5) | |
197 | #define GEN8_CTX_PRIVILEGE (1<<8) | |
e5815a2e | 198 | |
0d925ea0 | 199 | #define ASSIGN_CTX_REG(reg_state, pos, reg, val) do { \ |
f0f59a00 | 200 | (reg_state)[(pos)+0] = i915_mmio_reg_offset(reg); \ |
0d925ea0 VS |
201 | (reg_state)[(pos)+1] = (val); \ |
202 | } while (0) | |
203 | ||
204 | #define ASSIGN_CTX_PDP(ppgtt, reg_state, n) do { \ | |
d852c7bf | 205 | const u64 _addr = i915_page_dir_dma_addr((ppgtt), (n)); \ |
e5815a2e MT |
206 | reg_state[CTX_PDP ## n ## _UDW+1] = upper_32_bits(_addr); \ |
207 | reg_state[CTX_PDP ## n ## _LDW+1] = lower_32_bits(_addr); \ | |
9244a817 | 208 | } while (0) |
e5815a2e | 209 | |
9244a817 | 210 | #define ASSIGN_CTX_PML4(ppgtt, reg_state) do { \ |
2dba3239 MT |
211 | reg_state[CTX_PDP0_UDW + 1] = upper_32_bits(px_dma(&ppgtt->pml4)); \ |
212 | reg_state[CTX_PDP0_LDW + 1] = lower_32_bits(px_dma(&ppgtt->pml4)); \ | |
9244a817 | 213 | } while (0) |
2dba3239 | 214 | |
84b790f8 BW |
215 | enum { |
216 | FAULT_AND_HANG = 0, | |
217 | FAULT_AND_HALT, /* Debug only */ | |
218 | FAULT_AND_STREAM, | |
219 | FAULT_AND_CONTINUE /* Unsupported */ | |
220 | }; | |
221 | #define GEN8_CTX_ID_SHIFT 32 | |
7069b144 | 222 | #define GEN8_CTX_ID_WIDTH 21 |
71562919 MT |
223 | #define GEN8_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT 0x17 |
224 | #define GEN9_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT 0x26 | |
84b790f8 | 225 | |
0e93cdd4 CW |
226 | /* Typical size of the average request (2 pipecontrols and a MI_BB) */ |
227 | #define EXECLISTS_REQUEST_SIZE 64 /* bytes */ | |
228 | ||
a3aabe86 CW |
229 | #define WA_TAIL_DWORDS 2 |
230 | ||
e2efd130 | 231 | static int execlists_context_deferred_alloc(struct i915_gem_context *ctx, |
978f1e09 | 232 | struct intel_engine_cs *engine); |
e2efd130 | 233 | static int intel_lr_context_pin(struct i915_gem_context *ctx, |
e5292823 | 234 | struct intel_engine_cs *engine); |
a3aabe86 CW |
235 | static void execlists_init_reg_state(u32 *reg_state, |
236 | struct i915_gem_context *ctx, | |
237 | struct intel_engine_cs *engine, | |
238 | struct intel_ring *ring); | |
7ba717cf | 239 | |
73e4d07f OM |
240 | /** |
241 | * intel_sanitize_enable_execlists() - sanitize i915.enable_execlists | |
14bb2c11 | 242 | * @dev_priv: i915 device private |
73e4d07f OM |
243 | * @enable_execlists: value of i915.enable_execlists module parameter. |
244 | * | |
245 | * Only certain platforms support Execlists (the prerequisites being | |
27401d12 | 246 | * support for Logical Ring Contexts and Aliasing PPGTT or better). |
73e4d07f OM |
247 | * |
248 | * Return: 1 if Execlists is supported and has to be enabled. | |
249 | */ | |
c033666a | 250 | int intel_sanitize_enable_execlists(struct drm_i915_private *dev_priv, int enable_execlists) |
127f1003 | 251 | { |
a0bd6c31 ZL |
252 | /* On platforms with execlist available, vGPU will only |
253 | * support execlist mode, no ring buffer mode. | |
254 | */ | |
c033666a | 255 | if (HAS_LOGICAL_RING_CONTEXTS(dev_priv) && intel_vgpu_active(dev_priv)) |
a0bd6c31 ZL |
256 | return 1; |
257 | ||
c033666a | 258 | if (INTEL_GEN(dev_priv) >= 9) |
70ee45e1 DL |
259 | return 1; |
260 | ||
127f1003 OM |
261 | if (enable_execlists == 0) |
262 | return 0; | |
263 | ||
5a21b665 DV |
264 | if (HAS_LOGICAL_RING_CONTEXTS(dev_priv) && |
265 | USES_PPGTT(dev_priv) && | |
266 | i915.use_mmio_flip >= 0) | |
127f1003 OM |
267 | return 1; |
268 | ||
269 | return 0; | |
270 | } | |
ede7d42b | 271 | |
ca82580c | 272 | static void |
0bc40be8 | 273 | logical_ring_init_platform_invariants(struct intel_engine_cs *engine) |
ca82580c | 274 | { |
c033666a | 275 | struct drm_i915_private *dev_priv = engine->i915; |
ca82580c | 276 | |
70c2a24d | 277 | engine->disable_lite_restore_wa = |
a117f378 | 278 | IS_BXT_REVID(dev_priv, 0, BXT_REVID_A1) && |
70c2a24d | 279 | (engine->id == VCS || engine->id == VCS2); |
ca82580c | 280 | |
0bc40be8 | 281 | engine->ctx_desc_template = GEN8_CTX_VALID; |
c033666a | 282 | if (IS_GEN8(dev_priv)) |
0bc40be8 TU |
283 | engine->ctx_desc_template |= GEN8_CTX_L3LLC_COHERENT; |
284 | engine->ctx_desc_template |= GEN8_CTX_PRIVILEGE; | |
ca82580c TU |
285 | |
286 | /* TODO: WaDisableLiteRestore when we start using semaphore | |
287 | * signalling between Command Streamers */ | |
288 | /* ring->ctx_desc_template |= GEN8_CTX_FORCE_RESTORE; */ | |
289 | ||
290 | /* WaEnableForceRestoreInCtxtDescForVCS:skl */ | |
291 | /* WaEnableForceRestoreInCtxtDescForVCS:bxt */ | |
0bc40be8 TU |
292 | if (engine->disable_lite_restore_wa) |
293 | engine->ctx_desc_template |= GEN8_CTX_FORCE_RESTORE; | |
ca82580c TU |
294 | } |
295 | ||
73e4d07f | 296 | /** |
ca82580c TU |
297 | * intel_lr_context_descriptor_update() - calculate & cache the descriptor |
298 | * descriptor for a pinned context | |
ca82580c | 299 | * @ctx: Context to work on |
9021ad03 | 300 | * @engine: Engine the descriptor will be used with |
73e4d07f | 301 | * |
ca82580c TU |
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 | * | |
6e5248b5 DV |
307 | * This is what a descriptor looks like, from LSB to MSB:: |
308 | * | |
309 | * bits 0-11: flags, GEN8_CTX_* (cached in ctx_desc_template) | |
310 | * bits 12-31: LRCA, GTT address of (the HWSP of) this context | |
311 | * bits 32-52: ctx ID, a globally unique tag | |
312 | * bits 53-54: mbz, reserved for use by hardware | |
313 | * bits 55-63: group ID, currently unused and set to 0 | |
73e4d07f | 314 | */ |
ca82580c | 315 | static void |
e2efd130 | 316 | intel_lr_context_descriptor_update(struct i915_gem_context *ctx, |
0bc40be8 | 317 | struct intel_engine_cs *engine) |
84b790f8 | 318 | { |
9021ad03 | 319 | struct intel_context *ce = &ctx->engine[engine->id]; |
7069b144 | 320 | u64 desc; |
84b790f8 | 321 | |
7069b144 | 322 | BUILD_BUG_ON(MAX_CONTEXT_HW_ID > (1<<GEN8_CTX_ID_WIDTH)); |
84b790f8 | 323 | |
c01fc532 ZW |
324 | desc = ctx->desc_template; /* bits 3-4 */ |
325 | desc |= engine->ctx_desc_template; /* bits 0-11 */ | |
bde13ebd | 326 | desc |= i915_ggtt_offset(ce->state) + LRC_PPHWSP_PN * PAGE_SIZE; |
9021ad03 | 327 | /* bits 12-31 */ |
7069b144 | 328 | desc |= (u64)ctx->hw_id << GEN8_CTX_ID_SHIFT; /* bits 32-52 */ |
5af05fef | 329 | |
9021ad03 | 330 | ce->lrc_desc = desc; |
5af05fef MT |
331 | } |
332 | ||
e2efd130 | 333 | uint64_t intel_lr_context_descriptor(struct i915_gem_context *ctx, |
0bc40be8 | 334 | struct intel_engine_cs *engine) |
84b790f8 | 335 | { |
0bc40be8 | 336 | return ctx->engine[engine->id].lrc_desc; |
ca82580c | 337 | } |
203a571b | 338 | |
bbd6c47e CW |
339 | static inline void |
340 | execlists_context_status_change(struct drm_i915_gem_request *rq, | |
341 | unsigned long status) | |
84b790f8 | 342 | { |
bbd6c47e CW |
343 | /* |
344 | * Only used when GVT-g is enabled now. When GVT-g is disabled, | |
345 | * The compiler should eliminate this function as dead-code. | |
346 | */ | |
347 | if (!IS_ENABLED(CONFIG_DRM_I915_GVT)) | |
348 | return; | |
6daccb0b | 349 | |
bbd6c47e | 350 | atomic_notifier_call_chain(&rq->ctx->status_notifier, status, rq); |
84b790f8 BW |
351 | } |
352 | ||
c6a2ac71 TU |
353 | static void |
354 | execlists_update_context_pdps(struct i915_hw_ppgtt *ppgtt, u32 *reg_state) | |
355 | { | |
356 | ASSIGN_CTX_PDP(ppgtt, reg_state, 3); | |
357 | ASSIGN_CTX_PDP(ppgtt, reg_state, 2); | |
358 | ASSIGN_CTX_PDP(ppgtt, reg_state, 1); | |
359 | ASSIGN_CTX_PDP(ppgtt, reg_state, 0); | |
360 | } | |
361 | ||
70c2a24d | 362 | static u64 execlists_update_context(struct drm_i915_gem_request *rq) |
ae1250b9 | 363 | { |
70c2a24d | 364 | struct intel_context *ce = &rq->ctx->engine[rq->engine->id]; |
05d9824b | 365 | struct i915_hw_ppgtt *ppgtt = rq->ctx->ppgtt; |
70c2a24d | 366 | u32 *reg_state = ce->lrc_reg_state; |
ae1250b9 | 367 | |
caddfe71 | 368 | reg_state[CTX_RING_TAIL+1] = rq->tail; |
ae1250b9 | 369 | |
c6a2ac71 TU |
370 | /* True 32b PPGTT with dynamic page allocation: update PDP |
371 | * registers and point the unallocated PDPs to scratch page. | |
372 | * PML4 is allocated during ppgtt init, so this is not needed | |
373 | * in 48-bit mode. | |
374 | */ | |
375 | if (ppgtt && !USES_FULL_48BIT_PPGTT(ppgtt->base.dev)) | |
376 | execlists_update_context_pdps(ppgtt, reg_state); | |
70c2a24d CW |
377 | |
378 | return ce->lrc_desc; | |
ae1250b9 OM |
379 | } |
380 | ||
70c2a24d | 381 | static void execlists_submit_ports(struct intel_engine_cs *engine) |
bbd6c47e | 382 | { |
70c2a24d CW |
383 | struct drm_i915_private *dev_priv = engine->i915; |
384 | struct execlist_port *port = engine->execlist_port; | |
bbd6c47e CW |
385 | u32 __iomem *elsp = |
386 | dev_priv->regs + i915_mmio_reg_offset(RING_ELSP(engine)); | |
387 | u64 desc[2]; | |
388 | ||
70c2a24d CW |
389 | if (!port[0].count) |
390 | execlists_context_status_change(port[0].request, | |
391 | INTEL_CONTEXT_SCHEDULE_IN); | |
392 | desc[0] = execlists_update_context(port[0].request); | |
393 | engine->preempt_wa = port[0].count++; /* bdw only? fixed on skl? */ | |
394 | ||
395 | if (port[1].request) { | |
396 | GEM_BUG_ON(port[1].count); | |
397 | execlists_context_status_change(port[1].request, | |
398 | INTEL_CONTEXT_SCHEDULE_IN); | |
399 | desc[1] = execlists_update_context(port[1].request); | |
400 | port[1].count = 1; | |
bbd6c47e CW |
401 | } else { |
402 | desc[1] = 0; | |
403 | } | |
70c2a24d | 404 | GEM_BUG_ON(desc[0] == desc[1]); |
bbd6c47e CW |
405 | |
406 | /* You must always write both descriptors in the order below. */ | |
407 | writel(upper_32_bits(desc[1]), elsp); | |
408 | writel(lower_32_bits(desc[1]), elsp); | |
409 | ||
410 | writel(upper_32_bits(desc[0]), elsp); | |
411 | /* The context is automatically loaded after the following */ | |
412 | writel(lower_32_bits(desc[0]), elsp); | |
413 | } | |
414 | ||
70c2a24d | 415 | static bool ctx_single_port_submission(const struct i915_gem_context *ctx) |
84b790f8 | 416 | { |
70c2a24d CW |
417 | return (IS_ENABLED(CONFIG_DRM_I915_GVT) && |
418 | ctx->execlists_force_single_submission); | |
419 | } | |
84b790f8 | 420 | |
70c2a24d CW |
421 | static bool can_merge_ctx(const struct i915_gem_context *prev, |
422 | const struct i915_gem_context *next) | |
423 | { | |
424 | if (prev != next) | |
425 | return false; | |
26720ab9 | 426 | |
70c2a24d CW |
427 | if (ctx_single_port_submission(prev)) |
428 | return false; | |
26720ab9 | 429 | |
70c2a24d | 430 | return true; |
84b790f8 BW |
431 | } |
432 | ||
70c2a24d | 433 | static void execlists_dequeue(struct intel_engine_cs *engine) |
acdd884a | 434 | { |
20311bd3 | 435 | struct drm_i915_gem_request *last; |
70c2a24d | 436 | struct execlist_port *port = engine->execlist_port; |
d55ac5bf | 437 | unsigned long flags; |
20311bd3 | 438 | struct rb_node *rb; |
70c2a24d CW |
439 | bool submit = false; |
440 | ||
441 | last = port->request; | |
442 | if (last) | |
443 | /* WaIdleLiteRestore:bdw,skl | |
444 | * Apply the wa NOOPs to prevent ring:HEAD == req:TAIL | |
9b81d556 | 445 | * as we resubmit the request. See gen8_emit_breadcrumb() |
70c2a24d CW |
446 | * for where we prepare the padding after the end of the |
447 | * request. | |
448 | */ | |
449 | last->tail = last->wa_tail; | |
e981e7b1 | 450 | |
70c2a24d | 451 | GEM_BUG_ON(port[1].request); |
acdd884a | 452 | |
70c2a24d CW |
453 | /* Hardware submission is through 2 ports. Conceptually each port |
454 | * has a (RING_START, RING_HEAD, RING_TAIL) tuple. RING_START is | |
455 | * static for a context, and unique to each, so we only execute | |
456 | * requests belonging to a single context from each ring. RING_HEAD | |
457 | * is maintained by the CS in the context image, it marks the place | |
458 | * where it got up to last time, and through RING_TAIL we tell the CS | |
459 | * where we want to execute up to this time. | |
460 | * | |
461 | * In this list the requests are in order of execution. Consecutive | |
462 | * requests from the same context are adjacent in the ringbuffer. We | |
463 | * can combine these requests into a single RING_TAIL update: | |
464 | * | |
465 | * RING_HEAD...req1...req2 | |
466 | * ^- RING_TAIL | |
467 | * since to execute req2 the CS must first execute req1. | |
468 | * | |
469 | * Our goal then is to point each port to the end of a consecutive | |
470 | * sequence of requests as being the most optimal (fewest wake ups | |
471 | * and context switches) submission. | |
779949f4 | 472 | */ |
acdd884a | 473 | |
d55ac5bf | 474 | spin_lock_irqsave(&engine->timeline->lock, flags); |
20311bd3 CW |
475 | rb = engine->execlist_first; |
476 | while (rb) { | |
477 | struct drm_i915_gem_request *cursor = | |
478 | rb_entry(rb, typeof(*cursor), priotree.node); | |
479 | ||
70c2a24d CW |
480 | /* Can we combine this request with the current port? It has to |
481 | * be the same context/ringbuffer and not have any exceptions | |
482 | * (e.g. GVT saying never to combine contexts). | |
c6a2ac71 | 483 | * |
70c2a24d CW |
484 | * If we can combine the requests, we can execute both by |
485 | * updating the RING_TAIL to point to the end of the second | |
486 | * request, and so we never need to tell the hardware about | |
487 | * the first. | |
53292cdb | 488 | */ |
70c2a24d CW |
489 | if (last && !can_merge_ctx(cursor->ctx, last->ctx)) { |
490 | /* If we are on the second port and cannot combine | |
491 | * this request with the last, then we are done. | |
492 | */ | |
493 | if (port != engine->execlist_port) | |
494 | break; | |
495 | ||
496 | /* If GVT overrides us we only ever submit port[0], | |
497 | * leaving port[1] empty. Note that we also have | |
498 | * to be careful that we don't queue the same | |
499 | * context (even though a different request) to | |
500 | * the second port. | |
501 | */ | |
d7ab992c MH |
502 | if (ctx_single_port_submission(last->ctx) || |
503 | ctx_single_port_submission(cursor->ctx)) | |
70c2a24d CW |
504 | break; |
505 | ||
506 | GEM_BUG_ON(last->ctx == cursor->ctx); | |
507 | ||
508 | i915_gem_request_assign(&port->request, last); | |
509 | port++; | |
510 | } | |
d55ac5bf | 511 | |
20311bd3 CW |
512 | rb = rb_next(rb); |
513 | rb_erase(&cursor->priotree.node, &engine->execlist_queue); | |
514 | RB_CLEAR_NODE(&cursor->priotree.node); | |
515 | cursor->priotree.priority = INT_MAX; | |
516 | ||
d55ac5bf CW |
517 | /* We keep the previous context alive until we retire the |
518 | * following request. This ensures that any the context object | |
519 | * is still pinned for any residual writes the HW makes into it | |
520 | * on the context switch into the next object following the | |
521 | * breadcrumb. Otherwise, we may retire the context too early. | |
522 | */ | |
523 | cursor->previous_context = engine->last_context; | |
524 | engine->last_context = cursor->ctx; | |
525 | ||
526 | __i915_gem_request_submit(cursor); | |
70c2a24d CW |
527 | last = cursor; |
528 | submit = true; | |
529 | } | |
530 | if (submit) { | |
70c2a24d | 531 | i915_gem_request_assign(&port->request, last); |
20311bd3 | 532 | engine->execlist_first = rb; |
53292cdb | 533 | } |
d55ac5bf | 534 | spin_unlock_irqrestore(&engine->timeline->lock, flags); |
53292cdb | 535 | |
70c2a24d CW |
536 | if (submit) |
537 | execlists_submit_ports(engine); | |
acdd884a MT |
538 | } |
539 | ||
70c2a24d | 540 | static bool execlists_elsp_idle(struct intel_engine_cs *engine) |
e981e7b1 | 541 | { |
70c2a24d | 542 | return !engine->execlist_port[0].request; |
e981e7b1 TD |
543 | } |
544 | ||
0cb5670b ID |
545 | /** |
546 | * intel_execlists_idle() - Determine if all engine submission ports are idle | |
547 | * @dev_priv: i915 device private | |
548 | * | |
549 | * Return true if there are no requests pending on any of the submission ports | |
550 | * of any engines. | |
551 | */ | |
552 | bool intel_execlists_idle(struct drm_i915_private *dev_priv) | |
553 | { | |
554 | struct intel_engine_cs *engine; | |
555 | enum intel_engine_id id; | |
556 | ||
557 | if (!i915.enable_execlists) | |
558 | return true; | |
559 | ||
560 | for_each_engine(engine, dev_priv, id) | |
561 | if (!execlists_elsp_idle(engine)) | |
562 | return false; | |
563 | ||
564 | return true; | |
565 | } | |
566 | ||
70c2a24d | 567 | static bool execlists_elsp_ready(struct intel_engine_cs *engine) |
91a41032 | 568 | { |
70c2a24d | 569 | int port; |
91a41032 | 570 | |
70c2a24d CW |
571 | port = 1; /* wait for a free slot */ |
572 | if (engine->disable_lite_restore_wa || engine->preempt_wa) | |
573 | port = 0; /* wait for GPU to be idle before continuing */ | |
c6a2ac71 | 574 | |
70c2a24d | 575 | return !engine->execlist_port[port].request; |
91a41032 BW |
576 | } |
577 | ||
6e5248b5 | 578 | /* |
73e4d07f OM |
579 | * Check the unread Context Status Buffers and manage the submission of new |
580 | * contexts to the ELSP accordingly. | |
581 | */ | |
27af5eea | 582 | static void intel_lrc_irq_handler(unsigned long data) |
e981e7b1 | 583 | { |
27af5eea | 584 | struct intel_engine_cs *engine = (struct intel_engine_cs *)data; |
70c2a24d | 585 | struct execlist_port *port = engine->execlist_port; |
c033666a | 586 | struct drm_i915_private *dev_priv = engine->i915; |
c6a2ac71 | 587 | |
3756685a | 588 | intel_uncore_forcewake_get(dev_priv, engine->fw_domains); |
c6a2ac71 | 589 | |
70c2a24d CW |
590 | if (!execlists_elsp_idle(engine)) { |
591 | u32 __iomem *csb_mmio = | |
592 | dev_priv->regs + i915_mmio_reg_offset(RING_CONTEXT_STATUS_PTR(engine)); | |
593 | u32 __iomem *buf = | |
594 | dev_priv->regs + i915_mmio_reg_offset(RING_CONTEXT_STATUS_BUF_LO(engine, 0)); | |
595 | unsigned int csb, head, tail; | |
596 | ||
597 | csb = readl(csb_mmio); | |
598 | head = GEN8_CSB_READ_PTR(csb); | |
599 | tail = GEN8_CSB_WRITE_PTR(csb); | |
600 | if (tail < head) | |
601 | tail += GEN8_CSB_ENTRIES; | |
602 | while (head < tail) { | |
603 | unsigned int idx = ++head % GEN8_CSB_ENTRIES; | |
604 | unsigned int status = readl(buf + 2 * idx); | |
605 | ||
606 | if (!(status & GEN8_CTX_STATUS_COMPLETED_MASK)) | |
607 | continue; | |
608 | ||
609 | GEM_BUG_ON(port[0].count == 0); | |
610 | if (--port[0].count == 0) { | |
611 | GEM_BUG_ON(status & GEN8_CTX_STATUS_PREEMPTED); | |
612 | execlists_context_status_change(port[0].request, | |
613 | INTEL_CONTEXT_SCHEDULE_OUT); | |
614 | ||
615 | i915_gem_request_put(port[0].request); | |
616 | port[0] = port[1]; | |
617 | memset(&port[1], 0, sizeof(port[1])); | |
618 | ||
619 | engine->preempt_wa = false; | |
620 | } | |
26720ab9 | 621 | |
70c2a24d CW |
622 | GEM_BUG_ON(port[0].count == 0 && |
623 | !(status & GEN8_CTX_STATUS_ACTIVE_IDLE)); | |
e1fee72c OM |
624 | } |
625 | ||
70c2a24d CW |
626 | writel(_MASKED_FIELD(GEN8_CSB_READ_PTR_MASK, |
627 | GEN8_CSB_WRITE_PTR(csb) << 8), | |
628 | csb_mmio); | |
e981e7b1 TD |
629 | } |
630 | ||
70c2a24d CW |
631 | if (execlists_elsp_ready(engine)) |
632 | execlists_dequeue(engine); | |
c6a2ac71 | 633 | |
70c2a24d | 634 | intel_uncore_forcewake_put(dev_priv, engine->fw_domains); |
e981e7b1 TD |
635 | } |
636 | ||
20311bd3 CW |
637 | static bool insert_request(struct i915_priotree *pt, struct rb_root *root) |
638 | { | |
639 | struct rb_node **p, *rb; | |
640 | bool first = true; | |
641 | ||
642 | /* most positive priority is scheduled first, equal priorities fifo */ | |
643 | rb = NULL; | |
644 | p = &root->rb_node; | |
645 | while (*p) { | |
646 | struct i915_priotree *pos; | |
647 | ||
648 | rb = *p; | |
649 | pos = rb_entry(rb, typeof(*pos), node); | |
650 | if (pt->priority > pos->priority) { | |
651 | p = &rb->rb_left; | |
652 | } else { | |
653 | p = &rb->rb_right; | |
654 | first = false; | |
655 | } | |
656 | } | |
657 | rb_link_node(&pt->node, rb, p); | |
658 | rb_insert_color(&pt->node, root); | |
659 | ||
660 | return first; | |
661 | } | |
662 | ||
f4ea6bdd | 663 | static void execlists_submit_request(struct drm_i915_gem_request *request) |
acdd884a | 664 | { |
4a570db5 | 665 | struct intel_engine_cs *engine = request->engine; |
5590af3e | 666 | unsigned long flags; |
acdd884a | 667 | |
663f71e7 CW |
668 | /* Will be called from irq-context when using foreign fences. */ |
669 | spin_lock_irqsave(&engine->timeline->lock, flags); | |
acdd884a | 670 | |
20311bd3 CW |
671 | if (insert_request(&request->priotree, &engine->execlist_queue)) |
672 | engine->execlist_first = &request->priotree.node; | |
70c2a24d CW |
673 | if (execlists_elsp_idle(engine)) |
674 | tasklet_hi_schedule(&engine->irq_tasklet); | |
acdd884a | 675 | |
663f71e7 | 676 | spin_unlock_irqrestore(&engine->timeline->lock, flags); |
acdd884a MT |
677 | } |
678 | ||
20311bd3 CW |
679 | static struct intel_engine_cs * |
680 | pt_lock_engine(struct i915_priotree *pt, struct intel_engine_cs *locked) | |
681 | { | |
682 | struct intel_engine_cs *engine; | |
683 | ||
684 | engine = container_of(pt, | |
685 | struct drm_i915_gem_request, | |
686 | priotree)->engine; | |
687 | if (engine != locked) { | |
688 | if (locked) | |
689 | spin_unlock_irq(&locked->timeline->lock); | |
690 | spin_lock_irq(&engine->timeline->lock); | |
691 | } | |
692 | ||
693 | return engine; | |
694 | } | |
695 | ||
696 | static void execlists_schedule(struct drm_i915_gem_request *request, int prio) | |
697 | { | |
27745e82 | 698 | static DEFINE_MUTEX(lock); |
20311bd3 CW |
699 | struct intel_engine_cs *engine = NULL; |
700 | struct i915_dependency *dep, *p; | |
701 | struct i915_dependency stack; | |
702 | LIST_HEAD(dfs); | |
703 | ||
704 | if (prio <= READ_ONCE(request->priotree.priority)) | |
705 | return; | |
706 | ||
27745e82 CW |
707 | /* Need global lock to use the temporary link inside i915_dependency */ |
708 | mutex_lock(&lock); | |
20311bd3 CW |
709 | |
710 | stack.signaler = &request->priotree; | |
711 | list_add(&stack.dfs_link, &dfs); | |
712 | ||
713 | /* Recursively bump all dependent priorities to match the new request. | |
714 | * | |
715 | * A naive approach would be to use recursion: | |
716 | * static void update_priorities(struct i915_priotree *pt, prio) { | |
717 | * list_for_each_entry(dep, &pt->signalers_list, signal_link) | |
718 | * update_priorities(dep->signal, prio) | |
719 | * insert_request(pt); | |
720 | * } | |
721 | * but that may have unlimited recursion depth and so runs a very | |
722 | * real risk of overunning the kernel stack. Instead, we build | |
723 | * a flat list of all dependencies starting with the current request. | |
724 | * As we walk the list of dependencies, we add all of its dependencies | |
725 | * to the end of the list (this may include an already visited | |
726 | * request) and continue to walk onwards onto the new dependencies. The | |
727 | * end result is a topological list of requests in reverse order, the | |
728 | * last element in the list is the request we must execute first. | |
729 | */ | |
730 | list_for_each_entry_safe(dep, p, &dfs, dfs_link) { | |
731 | struct i915_priotree *pt = dep->signaler; | |
732 | ||
733 | list_for_each_entry(p, &pt->signalers_list, signal_link) | |
734 | if (prio > READ_ONCE(p->signaler->priority)) | |
735 | list_move_tail(&p->dfs_link, &dfs); | |
736 | ||
737 | p = list_next_entry(dep, dfs_link); | |
738 | if (!RB_EMPTY_NODE(&pt->node)) | |
739 | continue; | |
740 | ||
741 | engine = pt_lock_engine(pt, engine); | |
742 | ||
743 | /* If it is not already in the rbtree, we can update the | |
744 | * priority inplace and skip over it (and its dependencies) | |
745 | * if it is referenced *again* as we descend the dfs. | |
746 | */ | |
747 | if (prio > pt->priority && RB_EMPTY_NODE(&pt->node)) { | |
748 | pt->priority = prio; | |
749 | list_del_init(&dep->dfs_link); | |
750 | } | |
751 | } | |
752 | ||
753 | /* Fifo and depth-first replacement ensure our deps execute before us */ | |
754 | list_for_each_entry_safe_reverse(dep, p, &dfs, dfs_link) { | |
755 | struct i915_priotree *pt = dep->signaler; | |
756 | ||
757 | INIT_LIST_HEAD(&dep->dfs_link); | |
758 | ||
759 | engine = pt_lock_engine(pt, engine); | |
760 | ||
761 | if (prio <= pt->priority) | |
762 | continue; | |
763 | ||
764 | GEM_BUG_ON(RB_EMPTY_NODE(&pt->node)); | |
765 | ||
766 | pt->priority = prio; | |
767 | rb_erase(&pt->node, &engine->execlist_queue); | |
768 | if (insert_request(pt, &engine->execlist_queue)) | |
769 | engine->execlist_first = &pt->node; | |
770 | } | |
771 | ||
772 | if (engine) | |
773 | spin_unlock_irq(&engine->timeline->lock); | |
774 | ||
27745e82 CW |
775 | mutex_unlock(&lock); |
776 | ||
20311bd3 CW |
777 | /* XXX Do we need to preempt to make room for us and our deps? */ |
778 | } | |
779 | ||
40e895ce | 780 | int intel_logical_ring_alloc_request_extras(struct drm_i915_gem_request *request) |
bc0dce3f | 781 | { |
24f1d3cc | 782 | struct intel_engine_cs *engine = request->engine; |
9021ad03 | 783 | struct intel_context *ce = &request->ctx->engine[engine->id]; |
bfa01200 | 784 | int ret; |
bc0dce3f | 785 | |
6310346e CW |
786 | /* Flush enough space to reduce the likelihood of waiting after |
787 | * we start building the request - in which case we will just | |
788 | * have to repeat work. | |
789 | */ | |
0e93cdd4 | 790 | request->reserved_space += EXECLISTS_REQUEST_SIZE; |
6310346e | 791 | |
9021ad03 | 792 | if (!ce->state) { |
978f1e09 CW |
793 | ret = execlists_context_deferred_alloc(request->ctx, engine); |
794 | if (ret) | |
795 | return ret; | |
796 | } | |
797 | ||
dca33ecc | 798 | request->ring = ce->ring; |
f3cc01f0 | 799 | |
5ba89908 CW |
800 | ret = intel_lr_context_pin(request->ctx, engine); |
801 | if (ret) | |
802 | return ret; | |
803 | ||
a7e02199 AD |
804 | if (i915.enable_guc_submission) { |
805 | /* | |
806 | * Check that the GuC has space for the request before | |
807 | * going any further, as the i915_add_request() call | |
808 | * later on mustn't fail ... | |
809 | */ | |
7a9347f9 | 810 | ret = i915_guc_wq_reserve(request); |
a7e02199 | 811 | if (ret) |
5ba89908 | 812 | goto err_unpin; |
a7e02199 AD |
813 | } |
814 | ||
bfa01200 CW |
815 | ret = intel_ring_begin(request, 0); |
816 | if (ret) | |
5ba89908 | 817 | goto err_unreserve; |
bfa01200 | 818 | |
9021ad03 | 819 | if (!ce->initialised) { |
24f1d3cc CW |
820 | ret = engine->init_context(request); |
821 | if (ret) | |
5ba89908 | 822 | goto err_unreserve; |
24f1d3cc | 823 | |
9021ad03 | 824 | ce->initialised = true; |
24f1d3cc CW |
825 | } |
826 | ||
827 | /* Note that after this point, we have committed to using | |
828 | * this request as it is being used to both track the | |
829 | * state of engine initialisation and liveness of the | |
830 | * golden renderstate above. Think twice before you try | |
831 | * to cancel/unwind this request now. | |
832 | */ | |
833 | ||
0e93cdd4 | 834 | request->reserved_space -= EXECLISTS_REQUEST_SIZE; |
bfa01200 CW |
835 | return 0; |
836 | ||
5ba89908 CW |
837 | err_unreserve: |
838 | if (i915.enable_guc_submission) | |
839 | i915_guc_wq_unreserve(request); | |
bfa01200 | 840 | err_unpin: |
24f1d3cc | 841 | intel_lr_context_unpin(request->ctx, engine); |
e28e404c | 842 | return ret; |
bc0dce3f JH |
843 | } |
844 | ||
e2efd130 | 845 | static int intel_lr_context_pin(struct i915_gem_context *ctx, |
24f1d3cc | 846 | struct intel_engine_cs *engine) |
dcb4c12a | 847 | { |
9021ad03 | 848 | struct intel_context *ce = &ctx->engine[engine->id]; |
7d774cac | 849 | void *vaddr; |
ca82580c | 850 | int ret; |
dcb4c12a | 851 | |
91c8a326 | 852 | lockdep_assert_held(&ctx->i915->drm.struct_mutex); |
ca82580c | 853 | |
9021ad03 | 854 | if (ce->pin_count++) |
24f1d3cc CW |
855 | return 0; |
856 | ||
bf3783e5 CW |
857 | ret = i915_vma_pin(ce->state, 0, GEN8_LR_CONTEXT_ALIGN, |
858 | PIN_OFFSET_BIAS | GUC_WOPCM_TOP | PIN_GLOBAL); | |
e84fe803 | 859 | if (ret) |
24f1d3cc | 860 | goto err; |
7ba717cf | 861 | |
bf3783e5 | 862 | vaddr = i915_gem_object_pin_map(ce->state->obj, I915_MAP_WB); |
7d774cac TU |
863 | if (IS_ERR(vaddr)) { |
864 | ret = PTR_ERR(vaddr); | |
bf3783e5 | 865 | goto unpin_vma; |
82352e90 TU |
866 | } |
867 | ||
aad29fbb | 868 | ret = intel_ring_pin(ce->ring); |
e84fe803 | 869 | if (ret) |
7d774cac | 870 | goto unpin_map; |
d1675198 | 871 | |
0bc40be8 | 872 | intel_lr_context_descriptor_update(ctx, engine); |
9021ad03 | 873 | |
a3aabe86 CW |
874 | ce->lrc_reg_state = vaddr + LRC_STATE_PN * PAGE_SIZE; |
875 | ce->lrc_reg_state[CTX_RING_BUFFER_START+1] = | |
bde13ebd | 876 | i915_ggtt_offset(ce->ring->vma); |
a3aabe86 | 877 | |
a4f5ea64 | 878 | ce->state->obj->mm.dirty = true; |
e93c28f3 | 879 | |
e84fe803 | 880 | /* Invalidate GuC TLB. */ |
bf3783e5 CW |
881 | if (i915.enable_guc_submission) { |
882 | struct drm_i915_private *dev_priv = ctx->i915; | |
e84fe803 | 883 | I915_WRITE(GEN8_GTCR, GEN8_GTCR_INVALIDATE); |
bf3783e5 | 884 | } |
dcb4c12a | 885 | |
9a6feaf0 | 886 | i915_gem_context_get(ctx); |
24f1d3cc | 887 | return 0; |
7ba717cf | 888 | |
7d774cac | 889 | unpin_map: |
bf3783e5 CW |
890 | i915_gem_object_unpin_map(ce->state->obj); |
891 | unpin_vma: | |
892 | __i915_vma_unpin(ce->state); | |
24f1d3cc | 893 | err: |
9021ad03 | 894 | ce->pin_count = 0; |
e84fe803 NH |
895 | return ret; |
896 | } | |
897 | ||
e2efd130 | 898 | void intel_lr_context_unpin(struct i915_gem_context *ctx, |
24f1d3cc | 899 | struct intel_engine_cs *engine) |
e84fe803 | 900 | { |
9021ad03 | 901 | struct intel_context *ce = &ctx->engine[engine->id]; |
e84fe803 | 902 | |
91c8a326 | 903 | lockdep_assert_held(&ctx->i915->drm.struct_mutex); |
9021ad03 | 904 | GEM_BUG_ON(ce->pin_count == 0); |
321fe304 | 905 | |
9021ad03 | 906 | if (--ce->pin_count) |
24f1d3cc | 907 | return; |
e84fe803 | 908 | |
aad29fbb | 909 | intel_ring_unpin(ce->ring); |
dcb4c12a | 910 | |
bf3783e5 CW |
911 | i915_gem_object_unpin_map(ce->state->obj); |
912 | i915_vma_unpin(ce->state); | |
321fe304 | 913 | |
9a6feaf0 | 914 | i915_gem_context_put(ctx); |
dcb4c12a OM |
915 | } |
916 | ||
e2be4faf | 917 | static int intel_logical_ring_workarounds_emit(struct drm_i915_gem_request *req) |
771b9a53 MT |
918 | { |
919 | int ret, i; | |
7e37f889 | 920 | struct intel_ring *ring = req->ring; |
c033666a | 921 | struct i915_workarounds *w = &req->i915->workarounds; |
771b9a53 | 922 | |
cd7feaaa | 923 | if (w->count == 0) |
771b9a53 MT |
924 | return 0; |
925 | ||
7c9cf4e3 | 926 | ret = req->engine->emit_flush(req, EMIT_BARRIER); |
771b9a53 MT |
927 | if (ret) |
928 | return ret; | |
929 | ||
987046ad | 930 | ret = intel_ring_begin(req, w->count * 2 + 2); |
771b9a53 MT |
931 | if (ret) |
932 | return ret; | |
933 | ||
1dae2dfb | 934 | intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(w->count)); |
771b9a53 | 935 | for (i = 0; i < w->count; i++) { |
1dae2dfb CW |
936 | intel_ring_emit_reg(ring, w->reg[i].addr); |
937 | intel_ring_emit(ring, w->reg[i].value); | |
771b9a53 | 938 | } |
1dae2dfb | 939 | intel_ring_emit(ring, MI_NOOP); |
771b9a53 | 940 | |
1dae2dfb | 941 | intel_ring_advance(ring); |
771b9a53 | 942 | |
7c9cf4e3 | 943 | ret = req->engine->emit_flush(req, EMIT_BARRIER); |
771b9a53 MT |
944 | if (ret) |
945 | return ret; | |
946 | ||
947 | return 0; | |
948 | } | |
949 | ||
83b8a982 | 950 | #define wa_ctx_emit(batch, index, cmd) \ |
17ee950d | 951 | do { \ |
83b8a982 AS |
952 | int __index = (index)++; \ |
953 | if (WARN_ON(__index >= (PAGE_SIZE / sizeof(uint32_t)))) { \ | |
17ee950d AS |
954 | return -ENOSPC; \ |
955 | } \ | |
83b8a982 | 956 | batch[__index] = (cmd); \ |
17ee950d AS |
957 | } while (0) |
958 | ||
8f40db77 | 959 | #define wa_ctx_emit_reg(batch, index, reg) \ |
f0f59a00 | 960 | wa_ctx_emit((batch), (index), i915_mmio_reg_offset(reg)) |
9e000847 AS |
961 | |
962 | /* | |
963 | * In this WA we need to set GEN8_L3SQCREG4[21:21] and reset it after | |
964 | * PIPE_CONTROL instruction. This is required for the flush to happen correctly | |
965 | * but there is a slight complication as this is applied in WA batch where the | |
966 | * values are only initialized once so we cannot take register value at the | |
967 | * beginning and reuse it further; hence we save its value to memory, upload a | |
968 | * constant value with bit21 set and then we restore it back with the saved value. | |
969 | * To simplify the WA, a constant value is formed by using the default value | |
970 | * of this register. This shouldn't be a problem because we are only modifying | |
971 | * it for a short period and this batch in non-premptible. We can ofcourse | |
972 | * use additional instructions that read the actual value of the register | |
973 | * at that time and set our bit of interest but it makes the WA complicated. | |
974 | * | |
975 | * This WA is also required for Gen9 so extracting as a function avoids | |
976 | * code duplication. | |
977 | */ | |
0bc40be8 | 978 | static inline int gen8_emit_flush_coherentl3_wa(struct intel_engine_cs *engine, |
6e5248b5 | 979 | uint32_t *batch, |
9e000847 AS |
980 | uint32_t index) |
981 | { | |
982 | uint32_t l3sqc4_flush = (0x40400000 | GEN8_LQSC_FLUSH_COHERENT_LINES); | |
983 | ||
f1afe24f | 984 | wa_ctx_emit(batch, index, (MI_STORE_REGISTER_MEM_GEN8 | |
83b8a982 | 985 | MI_SRM_LRM_GLOBAL_GTT)); |
8f40db77 | 986 | wa_ctx_emit_reg(batch, index, GEN8_L3SQCREG4); |
bde13ebd | 987 | wa_ctx_emit(batch, index, i915_ggtt_offset(engine->scratch) + 256); |
83b8a982 AS |
988 | wa_ctx_emit(batch, index, 0); |
989 | ||
990 | wa_ctx_emit(batch, index, MI_LOAD_REGISTER_IMM(1)); | |
8f40db77 | 991 | wa_ctx_emit_reg(batch, index, GEN8_L3SQCREG4); |
83b8a982 AS |
992 | wa_ctx_emit(batch, index, l3sqc4_flush); |
993 | ||
994 | wa_ctx_emit(batch, index, GFX_OP_PIPE_CONTROL(6)); | |
995 | wa_ctx_emit(batch, index, (PIPE_CONTROL_CS_STALL | | |
996 | PIPE_CONTROL_DC_FLUSH_ENABLE)); | |
997 | wa_ctx_emit(batch, index, 0); | |
998 | wa_ctx_emit(batch, index, 0); | |
999 | wa_ctx_emit(batch, index, 0); | |
1000 | wa_ctx_emit(batch, index, 0); | |
1001 | ||
f1afe24f | 1002 | wa_ctx_emit(batch, index, (MI_LOAD_REGISTER_MEM_GEN8 | |
83b8a982 | 1003 | MI_SRM_LRM_GLOBAL_GTT)); |
8f40db77 | 1004 | wa_ctx_emit_reg(batch, index, GEN8_L3SQCREG4); |
bde13ebd | 1005 | wa_ctx_emit(batch, index, i915_ggtt_offset(engine->scratch) + 256); |
83b8a982 | 1006 | wa_ctx_emit(batch, index, 0); |
9e000847 AS |
1007 | |
1008 | return index; | |
1009 | } | |
1010 | ||
17ee950d AS |
1011 | static inline uint32_t wa_ctx_start(struct i915_wa_ctx_bb *wa_ctx, |
1012 | uint32_t offset, | |
1013 | uint32_t start_alignment) | |
1014 | { | |
1015 | return wa_ctx->offset = ALIGN(offset, start_alignment); | |
1016 | } | |
1017 | ||
1018 | static inline int wa_ctx_end(struct i915_wa_ctx_bb *wa_ctx, | |
1019 | uint32_t offset, | |
1020 | uint32_t size_alignment) | |
1021 | { | |
1022 | wa_ctx->size = offset - wa_ctx->offset; | |
1023 | ||
1024 | WARN(wa_ctx->size % size_alignment, | |
1025 | "wa_ctx_bb failed sanity checks: size %d is not aligned to %d\n", | |
1026 | wa_ctx->size, size_alignment); | |
1027 | return 0; | |
1028 | } | |
1029 | ||
6e5248b5 DV |
1030 | /* |
1031 | * Typically we only have one indirect_ctx and per_ctx batch buffer which are | |
1032 | * initialized at the beginning and shared across all contexts but this field | |
1033 | * helps us to have multiple batches at different offsets and select them based | |
1034 | * on a criteria. At the moment this batch always start at the beginning of the page | |
1035 | * and at this point we don't have multiple wa_ctx batch buffers. | |
4d78c8dc | 1036 | * |
6e5248b5 DV |
1037 | * The number of WA applied are not known at the beginning; we use this field |
1038 | * to return the no of DWORDS written. | |
17ee950d | 1039 | * |
6e5248b5 DV |
1040 | * It is to be noted that this batch does not contain MI_BATCH_BUFFER_END |
1041 | * so it adds NOOPs as padding to make it cacheline aligned. | |
1042 | * MI_BATCH_BUFFER_END will be added to perctx batch and both of them together | |
1043 | * makes a complete batch buffer. | |
17ee950d | 1044 | */ |
0bc40be8 | 1045 | static int gen8_init_indirectctx_bb(struct intel_engine_cs *engine, |
17ee950d | 1046 | struct i915_wa_ctx_bb *wa_ctx, |
6e5248b5 | 1047 | uint32_t *batch, |
17ee950d AS |
1048 | uint32_t *offset) |
1049 | { | |
0160f055 | 1050 | uint32_t scratch_addr; |
17ee950d AS |
1051 | uint32_t index = wa_ctx_start(wa_ctx, *offset, CACHELINE_DWORDS); |
1052 | ||
7ad00d1a | 1053 | /* WaDisableCtxRestoreArbitration:bdw,chv */ |
83b8a982 | 1054 | wa_ctx_emit(batch, index, MI_ARB_ON_OFF | MI_ARB_DISABLE); |
17ee950d | 1055 | |
c82435bb | 1056 | /* WaFlushCoherentL3CacheLinesAtContextSwitch:bdw */ |
c033666a | 1057 | if (IS_BROADWELL(engine->i915)) { |
0bc40be8 | 1058 | int rc = gen8_emit_flush_coherentl3_wa(engine, batch, index); |
604ef734 AH |
1059 | if (rc < 0) |
1060 | return rc; | |
1061 | index = rc; | |
c82435bb AS |
1062 | } |
1063 | ||
0160f055 AS |
1064 | /* WaClearSlmSpaceAtContextSwitch:bdw,chv */ |
1065 | /* Actual scratch location is at 128 bytes offset */ | |
bde13ebd | 1066 | scratch_addr = i915_ggtt_offset(engine->scratch) + 2 * CACHELINE_BYTES; |
0160f055 | 1067 | |
83b8a982 AS |
1068 | wa_ctx_emit(batch, index, GFX_OP_PIPE_CONTROL(6)); |
1069 | wa_ctx_emit(batch, index, (PIPE_CONTROL_FLUSH_L3 | | |
1070 | PIPE_CONTROL_GLOBAL_GTT_IVB | | |
1071 | PIPE_CONTROL_CS_STALL | | |
1072 | PIPE_CONTROL_QW_WRITE)); | |
1073 | wa_ctx_emit(batch, index, scratch_addr); | |
1074 | wa_ctx_emit(batch, index, 0); | |
1075 | wa_ctx_emit(batch, index, 0); | |
1076 | wa_ctx_emit(batch, index, 0); | |
0160f055 | 1077 | |
17ee950d AS |
1078 | /* Pad to end of cacheline */ |
1079 | while (index % CACHELINE_DWORDS) | |
83b8a982 | 1080 | wa_ctx_emit(batch, index, MI_NOOP); |
17ee950d AS |
1081 | |
1082 | /* | |
1083 | * MI_BATCH_BUFFER_END is not required in Indirect ctx BB because | |
1084 | * execution depends on the length specified in terms of cache lines | |
1085 | * in the register CTX_RCS_INDIRECT_CTX | |
1086 | */ | |
1087 | ||
1088 | return wa_ctx_end(wa_ctx, *offset = index, CACHELINE_DWORDS); | |
1089 | } | |
1090 | ||
6e5248b5 DV |
1091 | /* |
1092 | * This batch is started immediately after indirect_ctx batch. Since we ensure | |
1093 | * that indirect_ctx ends on a cacheline this batch is aligned automatically. | |
17ee950d | 1094 | * |
6e5248b5 | 1095 | * The number of DWORDS written are returned using this field. |
17ee950d AS |
1096 | * |
1097 | * This batch is terminated with MI_BATCH_BUFFER_END and so we need not add padding | |
1098 | * to align it with cacheline as padding after MI_BATCH_BUFFER_END is redundant. | |
1099 | */ | |
0bc40be8 | 1100 | static int gen8_init_perctx_bb(struct intel_engine_cs *engine, |
17ee950d | 1101 | struct i915_wa_ctx_bb *wa_ctx, |
6e5248b5 | 1102 | uint32_t *batch, |
17ee950d AS |
1103 | uint32_t *offset) |
1104 | { | |
1105 | uint32_t index = wa_ctx_start(wa_ctx, *offset, CACHELINE_DWORDS); | |
1106 | ||
7ad00d1a | 1107 | /* WaDisableCtxRestoreArbitration:bdw,chv */ |
83b8a982 | 1108 | wa_ctx_emit(batch, index, MI_ARB_ON_OFF | MI_ARB_ENABLE); |
7ad00d1a | 1109 | |
83b8a982 | 1110 | wa_ctx_emit(batch, index, MI_BATCH_BUFFER_END); |
17ee950d AS |
1111 | |
1112 | return wa_ctx_end(wa_ctx, *offset = index, 1); | |
1113 | } | |
1114 | ||
0bc40be8 | 1115 | static int gen9_init_indirectctx_bb(struct intel_engine_cs *engine, |
0504cffc | 1116 | struct i915_wa_ctx_bb *wa_ctx, |
6e5248b5 | 1117 | uint32_t *batch, |
0504cffc AS |
1118 | uint32_t *offset) |
1119 | { | |
a4106a78 | 1120 | int ret; |
5e580523 | 1121 | struct drm_i915_private *dev_priv = engine->i915; |
0504cffc AS |
1122 | uint32_t index = wa_ctx_start(wa_ctx, *offset, CACHELINE_DWORDS); |
1123 | ||
9fc736e8 JN |
1124 | /* WaDisableCtxRestoreArbitration:bxt */ |
1125 | if (IS_BXT_REVID(dev_priv, 0, BXT_REVID_A1)) | |
0907c8f7 | 1126 | wa_ctx_emit(batch, index, MI_ARB_ON_OFF | MI_ARB_DISABLE); |
0504cffc | 1127 | |
a4106a78 | 1128 | /* WaFlushCoherentL3CacheLinesAtContextSwitch:skl,bxt */ |
0bc40be8 | 1129 | ret = gen8_emit_flush_coherentl3_wa(engine, batch, index); |
a4106a78 AS |
1130 | if (ret < 0) |
1131 | return ret; | |
1132 | index = ret; | |
1133 | ||
873e8171 MK |
1134 | /* WaDisableGatherAtSetShaderCommonSlice:skl,bxt,kbl */ |
1135 | wa_ctx_emit(batch, index, MI_LOAD_REGISTER_IMM(1)); | |
1136 | wa_ctx_emit_reg(batch, index, COMMON_SLICE_CHICKEN2); | |
1137 | wa_ctx_emit(batch, index, _MASKED_BIT_DISABLE( | |
1138 | GEN9_DISABLE_GATHER_AT_SET_SHADER_COMMON_SLICE)); | |
1139 | wa_ctx_emit(batch, index, MI_NOOP); | |
1140 | ||
066d4628 MK |
1141 | /* WaClearSlmSpaceAtContextSwitch:kbl */ |
1142 | /* Actual scratch location is at 128 bytes offset */ | |
703d1282 | 1143 | if (IS_KBL_REVID(dev_priv, 0, KBL_REVID_A0)) { |
56c0f1a7 | 1144 | u32 scratch_addr = |
bde13ebd | 1145 | i915_ggtt_offset(engine->scratch) + 2 * CACHELINE_BYTES; |
066d4628 MK |
1146 | |
1147 | wa_ctx_emit(batch, index, GFX_OP_PIPE_CONTROL(6)); | |
1148 | wa_ctx_emit(batch, index, (PIPE_CONTROL_FLUSH_L3 | | |
1149 | PIPE_CONTROL_GLOBAL_GTT_IVB | | |
1150 | PIPE_CONTROL_CS_STALL | | |
1151 | PIPE_CONTROL_QW_WRITE)); | |
1152 | wa_ctx_emit(batch, index, scratch_addr); | |
1153 | wa_ctx_emit(batch, index, 0); | |
1154 | wa_ctx_emit(batch, index, 0); | |
1155 | wa_ctx_emit(batch, index, 0); | |
1156 | } | |
3485d99e TG |
1157 | |
1158 | /* WaMediaPoolStateCmdInWABB:bxt */ | |
1159 | if (HAS_POOLED_EU(engine->i915)) { | |
1160 | /* | |
1161 | * EU pool configuration is setup along with golden context | |
1162 | * during context initialization. This value depends on | |
1163 | * device type (2x6 or 3x6) and needs to be updated based | |
1164 | * on which subslice is disabled especially for 2x6 | |
1165 | * devices, however it is safe to load default | |
1166 | * configuration of 3x6 device instead of masking off | |
1167 | * corresponding bits because HW ignores bits of a disabled | |
1168 | * subslice and drops down to appropriate config. Please | |
1169 | * see render_state_setup() in i915_gem_render_state.c for | |
1170 | * possible configurations, to avoid duplication they are | |
1171 | * not shown here again. | |
1172 | */ | |
1173 | u32 eu_pool_config = 0x00777000; | |
1174 | wa_ctx_emit(batch, index, GEN9_MEDIA_POOL_STATE); | |
1175 | wa_ctx_emit(batch, index, GEN9_MEDIA_POOL_ENABLE); | |
1176 | wa_ctx_emit(batch, index, eu_pool_config); | |
1177 | wa_ctx_emit(batch, index, 0); | |
1178 | wa_ctx_emit(batch, index, 0); | |
1179 | wa_ctx_emit(batch, index, 0); | |
1180 | } | |
1181 | ||
0504cffc AS |
1182 | /* Pad to end of cacheline */ |
1183 | while (index % CACHELINE_DWORDS) | |
1184 | wa_ctx_emit(batch, index, MI_NOOP); | |
1185 | ||
1186 | return wa_ctx_end(wa_ctx, *offset = index, CACHELINE_DWORDS); | |
1187 | } | |
1188 | ||
0bc40be8 | 1189 | static int gen9_init_perctx_bb(struct intel_engine_cs *engine, |
0504cffc | 1190 | struct i915_wa_ctx_bb *wa_ctx, |
6e5248b5 | 1191 | uint32_t *batch, |
0504cffc AS |
1192 | uint32_t *offset) |
1193 | { | |
1194 | uint32_t index = wa_ctx_start(wa_ctx, *offset, CACHELINE_DWORDS); | |
1195 | ||
a117f378 JN |
1196 | /* WaSetDisablePixMaskCammingAndRhwoInCommonSliceChicken:bxt */ |
1197 | if (IS_BXT_REVID(engine->i915, 0, BXT_REVID_A1)) { | |
9b01435d | 1198 | wa_ctx_emit(batch, index, MI_LOAD_REGISTER_IMM(1)); |
8f40db77 | 1199 | wa_ctx_emit_reg(batch, index, GEN9_SLICE_COMMON_ECO_CHICKEN0); |
9b01435d AS |
1200 | wa_ctx_emit(batch, index, |
1201 | _MASKED_BIT_ENABLE(DISABLE_PIXEL_MASK_CAMMING)); | |
1202 | wa_ctx_emit(batch, index, MI_NOOP); | |
1203 | } | |
1204 | ||
b1e429fe | 1205 | /* WaClearTdlStateAckDirtyBits:bxt */ |
c033666a | 1206 | if (IS_BXT_REVID(engine->i915, 0, BXT_REVID_B0)) { |
b1e429fe TG |
1207 | wa_ctx_emit(batch, index, MI_LOAD_REGISTER_IMM(4)); |
1208 | ||
1209 | wa_ctx_emit_reg(batch, index, GEN8_STATE_ACK); | |
1210 | wa_ctx_emit(batch, index, _MASKED_BIT_DISABLE(GEN9_SUBSLICE_TDL_ACK_BITS)); | |
1211 | ||
1212 | wa_ctx_emit_reg(batch, index, GEN9_STATE_ACK_SLICE1); | |
1213 | wa_ctx_emit(batch, index, _MASKED_BIT_DISABLE(GEN9_SUBSLICE_TDL_ACK_BITS)); | |
1214 | ||
1215 | wa_ctx_emit_reg(batch, index, GEN9_STATE_ACK_SLICE2); | |
1216 | wa_ctx_emit(batch, index, _MASKED_BIT_DISABLE(GEN9_SUBSLICE_TDL_ACK_BITS)); | |
1217 | ||
1218 | wa_ctx_emit_reg(batch, index, GEN7_ROW_CHICKEN2); | |
1219 | /* dummy write to CS, mask bits are 0 to ensure the register is not modified */ | |
1220 | wa_ctx_emit(batch, index, 0x0); | |
1221 | wa_ctx_emit(batch, index, MI_NOOP); | |
1222 | } | |
1223 | ||
9fc736e8 JN |
1224 | /* WaDisableCtxRestoreArbitration:bxt */ |
1225 | if (IS_BXT_REVID(engine->i915, 0, BXT_REVID_A1)) | |
0907c8f7 AS |
1226 | wa_ctx_emit(batch, index, MI_ARB_ON_OFF | MI_ARB_ENABLE); |
1227 | ||
0504cffc AS |
1228 | wa_ctx_emit(batch, index, MI_BATCH_BUFFER_END); |
1229 | ||
1230 | return wa_ctx_end(wa_ctx, *offset = index, 1); | |
1231 | } | |
1232 | ||
0bc40be8 | 1233 | static int lrc_setup_wa_ctx_obj(struct intel_engine_cs *engine, u32 size) |
17ee950d | 1234 | { |
48bb74e4 CW |
1235 | struct drm_i915_gem_object *obj; |
1236 | struct i915_vma *vma; | |
1237 | int err; | |
17ee950d | 1238 | |
48bb74e4 CW |
1239 | obj = i915_gem_object_create(&engine->i915->drm, PAGE_ALIGN(size)); |
1240 | if (IS_ERR(obj)) | |
1241 | return PTR_ERR(obj); | |
17ee950d | 1242 | |
48bb74e4 CW |
1243 | vma = i915_vma_create(obj, &engine->i915->ggtt.base, NULL); |
1244 | if (IS_ERR(vma)) { | |
1245 | err = PTR_ERR(vma); | |
1246 | goto err; | |
17ee950d AS |
1247 | } |
1248 | ||
48bb74e4 CW |
1249 | err = i915_vma_pin(vma, 0, PAGE_SIZE, PIN_GLOBAL | PIN_HIGH); |
1250 | if (err) | |
1251 | goto err; | |
1252 | ||
1253 | engine->wa_ctx.vma = vma; | |
17ee950d | 1254 | return 0; |
48bb74e4 CW |
1255 | |
1256 | err: | |
1257 | i915_gem_object_put(obj); | |
1258 | return err; | |
17ee950d AS |
1259 | } |
1260 | ||
0bc40be8 | 1261 | static void lrc_destroy_wa_ctx_obj(struct intel_engine_cs *engine) |
17ee950d | 1262 | { |
19880c4a | 1263 | i915_vma_unpin_and_release(&engine->wa_ctx.vma); |
17ee950d AS |
1264 | } |
1265 | ||
0bc40be8 | 1266 | static int intel_init_workaround_bb(struct intel_engine_cs *engine) |
17ee950d | 1267 | { |
48bb74e4 | 1268 | struct i915_ctx_workarounds *wa_ctx = &engine->wa_ctx; |
17ee950d AS |
1269 | uint32_t *batch; |
1270 | uint32_t offset; | |
1271 | struct page *page; | |
48bb74e4 | 1272 | int ret; |
17ee950d | 1273 | |
0bc40be8 | 1274 | WARN_ON(engine->id != RCS); |
17ee950d | 1275 | |
5e60d790 | 1276 | /* update this when WA for higher Gen are added */ |
c033666a | 1277 | if (INTEL_GEN(engine->i915) > 9) { |
0504cffc | 1278 | DRM_ERROR("WA batch buffer is not initialized for Gen%d\n", |
c033666a | 1279 | INTEL_GEN(engine->i915)); |
5e60d790 | 1280 | return 0; |
0504cffc | 1281 | } |
5e60d790 | 1282 | |
c4db7599 | 1283 | /* some WA perform writes to scratch page, ensure it is valid */ |
56c0f1a7 | 1284 | if (!engine->scratch) { |
0bc40be8 | 1285 | DRM_ERROR("scratch page not allocated for %s\n", engine->name); |
c4db7599 AS |
1286 | return -EINVAL; |
1287 | } | |
1288 | ||
0bc40be8 | 1289 | ret = lrc_setup_wa_ctx_obj(engine, PAGE_SIZE); |
17ee950d AS |
1290 | if (ret) { |
1291 | DRM_DEBUG_DRIVER("Failed to setup context WA page: %d\n", ret); | |
1292 | return ret; | |
1293 | } | |
1294 | ||
48bb74e4 | 1295 | page = i915_gem_object_get_dirty_page(wa_ctx->vma->obj, 0); |
17ee950d AS |
1296 | batch = kmap_atomic(page); |
1297 | offset = 0; | |
1298 | ||
c033666a | 1299 | if (IS_GEN8(engine->i915)) { |
0bc40be8 | 1300 | ret = gen8_init_indirectctx_bb(engine, |
17ee950d AS |
1301 | &wa_ctx->indirect_ctx, |
1302 | batch, | |
1303 | &offset); | |
1304 | if (ret) | |
1305 | goto out; | |
1306 | ||
0bc40be8 | 1307 | ret = gen8_init_perctx_bb(engine, |
17ee950d AS |
1308 | &wa_ctx->per_ctx, |
1309 | batch, | |
1310 | &offset); | |
1311 | if (ret) | |
1312 | goto out; | |
c033666a | 1313 | } else if (IS_GEN9(engine->i915)) { |
0bc40be8 | 1314 | ret = gen9_init_indirectctx_bb(engine, |
0504cffc AS |
1315 | &wa_ctx->indirect_ctx, |
1316 | batch, | |
1317 | &offset); | |
1318 | if (ret) | |
1319 | goto out; | |
1320 | ||
0bc40be8 | 1321 | ret = gen9_init_perctx_bb(engine, |
0504cffc AS |
1322 | &wa_ctx->per_ctx, |
1323 | batch, | |
1324 | &offset); | |
1325 | if (ret) | |
1326 | goto out; | |
17ee950d AS |
1327 | } |
1328 | ||
1329 | out: | |
1330 | kunmap_atomic(batch); | |
1331 | if (ret) | |
0bc40be8 | 1332 | lrc_destroy_wa_ctx_obj(engine); |
17ee950d AS |
1333 | |
1334 | return ret; | |
1335 | } | |
1336 | ||
04794adb TU |
1337 | static void lrc_init_hws(struct intel_engine_cs *engine) |
1338 | { | |
c033666a | 1339 | struct drm_i915_private *dev_priv = engine->i915; |
04794adb TU |
1340 | |
1341 | I915_WRITE(RING_HWS_PGA(engine->mmio_base), | |
57e88531 | 1342 | engine->status_page.ggtt_offset); |
04794adb TU |
1343 | POSTING_READ(RING_HWS_PGA(engine->mmio_base)); |
1344 | } | |
1345 | ||
0bc40be8 | 1346 | static int gen8_init_common_ring(struct intel_engine_cs *engine) |
9b1136d5 | 1347 | { |
c033666a | 1348 | struct drm_i915_private *dev_priv = engine->i915; |
821ed7df CW |
1349 | int ret; |
1350 | ||
1351 | ret = intel_mocs_init_engine(engine); | |
1352 | if (ret) | |
1353 | return ret; | |
9b1136d5 | 1354 | |
04794adb | 1355 | lrc_init_hws(engine); |
e84fe803 | 1356 | |
ad07dfcd | 1357 | intel_engine_reset_breadcrumbs(engine); |
821ed7df | 1358 | |
0bc40be8 | 1359 | I915_WRITE(RING_HWSTAM(engine->mmio_base), 0xffffffff); |
73d477f6 | 1360 | |
0bc40be8 | 1361 | I915_WRITE(RING_MODE_GEN7(engine), |
9b1136d5 OM |
1362 | _MASKED_BIT_DISABLE(GFX_REPLAY_MODE) | |
1363 | _MASKED_BIT_ENABLE(GFX_RUN_LIST_ENABLE)); | |
dfc53c5e | 1364 | |
0bc40be8 | 1365 | DRM_DEBUG_DRIVER("Execlists enabled for %s\n", engine->name); |
9b1136d5 | 1366 | |
fc0768ce | 1367 | intel_engine_init_hangcheck(engine); |
9b1136d5 | 1368 | |
c87d50cc CW |
1369 | /* After a GPU reset, we may have requests to replay */ |
1370 | if (!execlists_elsp_idle(engine)) { | |
1371 | engine->execlist_port[0].count = 0; | |
1372 | engine->execlist_port[1].count = 0; | |
821ed7df | 1373 | execlists_submit_ports(engine); |
c87d50cc | 1374 | } |
821ed7df CW |
1375 | |
1376 | return 0; | |
9b1136d5 OM |
1377 | } |
1378 | ||
0bc40be8 | 1379 | static int gen8_init_render_ring(struct intel_engine_cs *engine) |
9b1136d5 | 1380 | { |
c033666a | 1381 | struct drm_i915_private *dev_priv = engine->i915; |
9b1136d5 OM |
1382 | int ret; |
1383 | ||
0bc40be8 | 1384 | ret = gen8_init_common_ring(engine); |
9b1136d5 OM |
1385 | if (ret) |
1386 | return ret; | |
1387 | ||
1388 | /* We need to disable the AsyncFlip performance optimisations in order | |
1389 | * to use MI_WAIT_FOR_EVENT within the CS. It should already be | |
1390 | * programmed to '1' on all products. | |
1391 | * | |
1392 | * WaDisableAsyncFlipPerfMode:snb,ivb,hsw,vlv,bdw,chv | |
1393 | */ | |
1394 | I915_WRITE(MI_MODE, _MASKED_BIT_ENABLE(ASYNC_FLIP_PERF_DISABLE)); | |
1395 | ||
9b1136d5 OM |
1396 | I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_FORCE_ORDERING)); |
1397 | ||
0bc40be8 | 1398 | return init_workarounds_ring(engine); |
9b1136d5 OM |
1399 | } |
1400 | ||
0bc40be8 | 1401 | static int gen9_init_render_ring(struct intel_engine_cs *engine) |
82ef822e DL |
1402 | { |
1403 | int ret; | |
1404 | ||
0bc40be8 | 1405 | ret = gen8_init_common_ring(engine); |
82ef822e DL |
1406 | if (ret) |
1407 | return ret; | |
1408 | ||
0bc40be8 | 1409 | return init_workarounds_ring(engine); |
82ef822e DL |
1410 | } |
1411 | ||
821ed7df CW |
1412 | static void reset_common_ring(struct intel_engine_cs *engine, |
1413 | struct drm_i915_gem_request *request) | |
1414 | { | |
1415 | struct drm_i915_private *dev_priv = engine->i915; | |
1416 | struct execlist_port *port = engine->execlist_port; | |
1417 | struct intel_context *ce = &request->ctx->engine[engine->id]; | |
1418 | ||
a3aabe86 CW |
1419 | /* We want a simple context + ring to execute the breadcrumb update. |
1420 | * We cannot rely on the context being intact across the GPU hang, | |
1421 | * so clear it and rebuild just what we need for the breadcrumb. | |
1422 | * All pending requests for this context will be zapped, and any | |
1423 | * future request will be after userspace has had the opportunity | |
1424 | * to recreate its own state. | |
1425 | */ | |
1426 | execlists_init_reg_state(ce->lrc_reg_state, | |
1427 | request->ctx, engine, ce->ring); | |
1428 | ||
821ed7df | 1429 | /* Move the RING_HEAD onto the breadcrumb, past the hanging batch */ |
a3aabe86 CW |
1430 | ce->lrc_reg_state[CTX_RING_BUFFER_START+1] = |
1431 | i915_ggtt_offset(ce->ring->vma); | |
821ed7df | 1432 | ce->lrc_reg_state[CTX_RING_HEAD+1] = request->postfix; |
a3aabe86 | 1433 | |
821ed7df CW |
1434 | request->ring->head = request->postfix; |
1435 | request->ring->last_retired_head = -1; | |
1436 | intel_ring_update_space(request->ring); | |
1437 | ||
1438 | if (i915.enable_guc_submission) | |
1439 | return; | |
1440 | ||
1441 | /* Catch up with any missed context-switch interrupts */ | |
1442 | I915_WRITE(RING_CONTEXT_STATUS_PTR(engine), _MASKED_FIELD(0xffff, 0)); | |
1443 | if (request->ctx != port[0].request->ctx) { | |
1444 | i915_gem_request_put(port[0].request); | |
1445 | port[0] = port[1]; | |
1446 | memset(&port[1], 0, sizeof(port[1])); | |
1447 | } | |
1448 | ||
821ed7df | 1449 | GEM_BUG_ON(request->ctx != port[0].request->ctx); |
a3aabe86 CW |
1450 | |
1451 | /* Reset WaIdleLiteRestore:bdw,skl as well */ | |
1452 | request->tail = request->wa_tail - WA_TAIL_DWORDS * sizeof(u32); | |
821ed7df CW |
1453 | } |
1454 | ||
7a01a0a2 MT |
1455 | static int intel_logical_ring_emit_pdps(struct drm_i915_gem_request *req) |
1456 | { | |
1457 | struct i915_hw_ppgtt *ppgtt = req->ctx->ppgtt; | |
7e37f889 | 1458 | struct intel_ring *ring = req->ring; |
4a570db5 | 1459 | struct intel_engine_cs *engine = req->engine; |
7a01a0a2 MT |
1460 | const int num_lri_cmds = GEN8_LEGACY_PDPES * 2; |
1461 | int i, ret; | |
1462 | ||
987046ad | 1463 | ret = intel_ring_begin(req, num_lri_cmds * 2 + 2); |
7a01a0a2 MT |
1464 | if (ret) |
1465 | return ret; | |
1466 | ||
b5321f30 | 1467 | intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(num_lri_cmds)); |
7a01a0a2 MT |
1468 | for (i = GEN8_LEGACY_PDPES - 1; i >= 0; i--) { |
1469 | const dma_addr_t pd_daddr = i915_page_dir_dma_addr(ppgtt, i); | |
1470 | ||
b5321f30 CW |
1471 | intel_ring_emit_reg(ring, GEN8_RING_PDP_UDW(engine, i)); |
1472 | intel_ring_emit(ring, upper_32_bits(pd_daddr)); | |
1473 | intel_ring_emit_reg(ring, GEN8_RING_PDP_LDW(engine, i)); | |
1474 | intel_ring_emit(ring, lower_32_bits(pd_daddr)); | |
7a01a0a2 MT |
1475 | } |
1476 | ||
b5321f30 CW |
1477 | intel_ring_emit(ring, MI_NOOP); |
1478 | intel_ring_advance(ring); | |
7a01a0a2 MT |
1479 | |
1480 | return 0; | |
1481 | } | |
1482 | ||
be795fc1 | 1483 | static int gen8_emit_bb_start(struct drm_i915_gem_request *req, |
803688ba CW |
1484 | u64 offset, u32 len, |
1485 | unsigned int dispatch_flags) | |
15648585 | 1486 | { |
7e37f889 | 1487 | struct intel_ring *ring = req->ring; |
8e004efc | 1488 | bool ppgtt = !(dispatch_flags & I915_DISPATCH_SECURE); |
15648585 OM |
1489 | int ret; |
1490 | ||
7a01a0a2 MT |
1491 | /* Don't rely in hw updating PDPs, specially in lite-restore. |
1492 | * Ideally, we should set Force PD Restore in ctx descriptor, | |
1493 | * but we can't. Force Restore would be a second option, but | |
1494 | * it is unsafe in case of lite-restore (because the ctx is | |
2dba3239 MT |
1495 | * not idle). PML4 is allocated during ppgtt init so this is |
1496 | * not needed in 48-bit.*/ | |
7a01a0a2 | 1497 | if (req->ctx->ppgtt && |
666796da | 1498 | (intel_engine_flag(req->engine) & req->ctx->ppgtt->pd_dirty_rings)) { |
331f38e7 | 1499 | if (!USES_FULL_48BIT_PPGTT(req->i915) && |
c033666a | 1500 | !intel_vgpu_active(req->i915)) { |
2dba3239 MT |
1501 | ret = intel_logical_ring_emit_pdps(req); |
1502 | if (ret) | |
1503 | return ret; | |
1504 | } | |
7a01a0a2 | 1505 | |
666796da | 1506 | req->ctx->ppgtt->pd_dirty_rings &= ~intel_engine_flag(req->engine); |
7a01a0a2 MT |
1507 | } |
1508 | ||
987046ad | 1509 | ret = intel_ring_begin(req, 4); |
15648585 OM |
1510 | if (ret) |
1511 | return ret; | |
1512 | ||
1513 | /* FIXME(BDW): Address space and security selectors. */ | |
b5321f30 CW |
1514 | intel_ring_emit(ring, MI_BATCH_BUFFER_START_GEN8 | |
1515 | (ppgtt<<8) | | |
1516 | (dispatch_flags & I915_DISPATCH_RS ? | |
1517 | MI_BATCH_RESOURCE_STREAMER : 0)); | |
1518 | intel_ring_emit(ring, lower_32_bits(offset)); | |
1519 | intel_ring_emit(ring, upper_32_bits(offset)); | |
1520 | intel_ring_emit(ring, MI_NOOP); | |
1521 | intel_ring_advance(ring); | |
15648585 OM |
1522 | |
1523 | return 0; | |
1524 | } | |
1525 | ||
31bb59cc | 1526 | static void gen8_logical_ring_enable_irq(struct intel_engine_cs *engine) |
73d477f6 | 1527 | { |
c033666a | 1528 | struct drm_i915_private *dev_priv = engine->i915; |
31bb59cc CW |
1529 | I915_WRITE_IMR(engine, |
1530 | ~(engine->irq_enable_mask | engine->irq_keep_mask)); | |
1531 | POSTING_READ_FW(RING_IMR(engine->mmio_base)); | |
73d477f6 OM |
1532 | } |
1533 | ||
31bb59cc | 1534 | static void gen8_logical_ring_disable_irq(struct intel_engine_cs *engine) |
73d477f6 | 1535 | { |
c033666a | 1536 | struct drm_i915_private *dev_priv = engine->i915; |
31bb59cc | 1537 | I915_WRITE_IMR(engine, ~engine->irq_keep_mask); |
73d477f6 OM |
1538 | } |
1539 | ||
7c9cf4e3 | 1540 | static int gen8_emit_flush(struct drm_i915_gem_request *request, u32 mode) |
4712274c | 1541 | { |
7e37f889 CW |
1542 | struct intel_ring *ring = request->ring; |
1543 | u32 cmd; | |
4712274c OM |
1544 | int ret; |
1545 | ||
987046ad | 1546 | ret = intel_ring_begin(request, 4); |
4712274c OM |
1547 | if (ret) |
1548 | return ret; | |
1549 | ||
1550 | cmd = MI_FLUSH_DW + 1; | |
1551 | ||
f0a1fb10 CW |
1552 | /* We always require a command barrier so that subsequent |
1553 | * commands, such as breadcrumb interrupts, are strictly ordered | |
1554 | * wrt the contents of the write cache being flushed to memory | |
1555 | * (and thus being coherent from the CPU). | |
1556 | */ | |
1557 | cmd |= MI_FLUSH_DW_STORE_INDEX | MI_FLUSH_DW_OP_STOREDW; | |
1558 | ||
7c9cf4e3 | 1559 | if (mode & EMIT_INVALIDATE) { |
f0a1fb10 | 1560 | cmd |= MI_INVALIDATE_TLB; |
1dae2dfb | 1561 | if (request->engine->id == VCS) |
f0a1fb10 | 1562 | cmd |= MI_INVALIDATE_BSD; |
4712274c OM |
1563 | } |
1564 | ||
b5321f30 CW |
1565 | intel_ring_emit(ring, cmd); |
1566 | intel_ring_emit(ring, | |
1567 | I915_GEM_HWS_SCRATCH_ADDR | | |
1568 | MI_FLUSH_DW_USE_GTT); | |
1569 | intel_ring_emit(ring, 0); /* upper addr */ | |
1570 | intel_ring_emit(ring, 0); /* value */ | |
1571 | intel_ring_advance(ring); | |
4712274c OM |
1572 | |
1573 | return 0; | |
1574 | } | |
1575 | ||
7deb4d39 | 1576 | static int gen8_emit_flush_render(struct drm_i915_gem_request *request, |
7c9cf4e3 | 1577 | u32 mode) |
4712274c | 1578 | { |
7e37f889 | 1579 | struct intel_ring *ring = request->ring; |
b5321f30 | 1580 | struct intel_engine_cs *engine = request->engine; |
bde13ebd CW |
1581 | u32 scratch_addr = |
1582 | i915_ggtt_offset(engine->scratch) + 2 * CACHELINE_BYTES; | |
0b2d0934 | 1583 | bool vf_flush_wa = false, dc_flush_wa = false; |
4712274c OM |
1584 | u32 flags = 0; |
1585 | int ret; | |
0b2d0934 | 1586 | int len; |
4712274c OM |
1587 | |
1588 | flags |= PIPE_CONTROL_CS_STALL; | |
1589 | ||
7c9cf4e3 | 1590 | if (mode & EMIT_FLUSH) { |
4712274c OM |
1591 | flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH; |
1592 | flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH; | |
965fd602 | 1593 | flags |= PIPE_CONTROL_DC_FLUSH_ENABLE; |
40a24488 | 1594 | flags |= PIPE_CONTROL_FLUSH_ENABLE; |
4712274c OM |
1595 | } |
1596 | ||
7c9cf4e3 | 1597 | if (mode & EMIT_INVALIDATE) { |
4712274c OM |
1598 | flags |= PIPE_CONTROL_TLB_INVALIDATE; |
1599 | flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE; | |
1600 | flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE; | |
1601 | flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE; | |
1602 | flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE; | |
1603 | flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE; | |
1604 | flags |= PIPE_CONTROL_QW_WRITE; | |
1605 | flags |= PIPE_CONTROL_GLOBAL_GTT_IVB; | |
4712274c | 1606 | |
1a5a9ce7 BW |
1607 | /* |
1608 | * On GEN9: before VF_CACHE_INVALIDATE we need to emit a NULL | |
1609 | * pipe control. | |
1610 | */ | |
c033666a | 1611 | if (IS_GEN9(request->i915)) |
1a5a9ce7 | 1612 | vf_flush_wa = true; |
0b2d0934 MK |
1613 | |
1614 | /* WaForGAMHang:kbl */ | |
1615 | if (IS_KBL_REVID(request->i915, 0, KBL_REVID_B0)) | |
1616 | dc_flush_wa = true; | |
1a5a9ce7 | 1617 | } |
9647ff36 | 1618 | |
0b2d0934 MK |
1619 | len = 6; |
1620 | ||
1621 | if (vf_flush_wa) | |
1622 | len += 6; | |
1623 | ||
1624 | if (dc_flush_wa) | |
1625 | len += 12; | |
1626 | ||
1627 | ret = intel_ring_begin(request, len); | |
4712274c OM |
1628 | if (ret) |
1629 | return ret; | |
1630 | ||
9647ff36 | 1631 | if (vf_flush_wa) { |
b5321f30 CW |
1632 | intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(6)); |
1633 | intel_ring_emit(ring, 0); | |
1634 | intel_ring_emit(ring, 0); | |
1635 | intel_ring_emit(ring, 0); | |
1636 | intel_ring_emit(ring, 0); | |
1637 | intel_ring_emit(ring, 0); | |
9647ff36 ID |
1638 | } |
1639 | ||
0b2d0934 | 1640 | if (dc_flush_wa) { |
b5321f30 CW |
1641 | intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(6)); |
1642 | intel_ring_emit(ring, PIPE_CONTROL_DC_FLUSH_ENABLE); | |
1643 | intel_ring_emit(ring, 0); | |
1644 | intel_ring_emit(ring, 0); | |
1645 | intel_ring_emit(ring, 0); | |
1646 | intel_ring_emit(ring, 0); | |
0b2d0934 MK |
1647 | } |
1648 | ||
b5321f30 CW |
1649 | intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(6)); |
1650 | intel_ring_emit(ring, flags); | |
1651 | intel_ring_emit(ring, scratch_addr); | |
1652 | intel_ring_emit(ring, 0); | |
1653 | intel_ring_emit(ring, 0); | |
1654 | intel_ring_emit(ring, 0); | |
0b2d0934 MK |
1655 | |
1656 | if (dc_flush_wa) { | |
b5321f30 CW |
1657 | intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(6)); |
1658 | intel_ring_emit(ring, PIPE_CONTROL_CS_STALL); | |
1659 | intel_ring_emit(ring, 0); | |
1660 | intel_ring_emit(ring, 0); | |
1661 | intel_ring_emit(ring, 0); | |
1662 | intel_ring_emit(ring, 0); | |
0b2d0934 MK |
1663 | } |
1664 | ||
b5321f30 | 1665 | intel_ring_advance(ring); |
4712274c OM |
1666 | |
1667 | return 0; | |
1668 | } | |
1669 | ||
c04e0f3b | 1670 | static void bxt_a_seqno_barrier(struct intel_engine_cs *engine) |
319404df | 1671 | { |
319404df ID |
1672 | /* |
1673 | * On BXT A steppings there is a HW coherency issue whereby the | |
1674 | * MI_STORE_DATA_IMM storing the completed request's seqno | |
1675 | * occasionally doesn't invalidate the CPU cache. Work around this by | |
1676 | * clflushing the corresponding cacheline whenever the caller wants | |
1677 | * the coherency to be guaranteed. Note that this cacheline is known | |
1678 | * to be clean at this point, since we only write it in | |
1679 | * bxt_a_set_seqno(), where we also do a clflush after the write. So | |
1680 | * this clflush in practice becomes an invalidate operation. | |
1681 | */ | |
c04e0f3b | 1682 | intel_flush_status_page(engine, I915_GEM_HWS_INDEX); |
319404df ID |
1683 | } |
1684 | ||
7c17d377 CW |
1685 | /* |
1686 | * Reserve space for 2 NOOPs at the end of each request to be | |
1687 | * used as a workaround for not being allowed to do lite | |
1688 | * restore with HEAD==TAIL (WaIdleLiteRestore). | |
1689 | */ | |
caddfe71 | 1690 | static void gen8_emit_wa_tail(struct drm_i915_gem_request *request, u32 *out) |
4da46e1e | 1691 | { |
caddfe71 CW |
1692 | *out++ = MI_NOOP; |
1693 | *out++ = MI_NOOP; | |
1694 | request->wa_tail = intel_ring_offset(request->ring, out); | |
1695 | } | |
4da46e1e | 1696 | |
caddfe71 CW |
1697 | static void gen8_emit_breadcrumb(struct drm_i915_gem_request *request, |
1698 | u32 *out) | |
1699 | { | |
7c17d377 CW |
1700 | /* w/a: bit 5 needs to be zero for MI_FLUSH_DW address. */ |
1701 | BUILD_BUG_ON(I915_GEM_HWS_INDEX_ADDR & (1 << 5)); | |
4da46e1e | 1702 | |
caddfe71 CW |
1703 | *out++ = (MI_FLUSH_DW + 1) | MI_FLUSH_DW_OP_STOREDW; |
1704 | *out++ = intel_hws_seqno_address(request->engine) | MI_FLUSH_DW_USE_GTT; | |
1705 | *out++ = 0; | |
1706 | *out++ = request->global_seqno; | |
1707 | *out++ = MI_USER_INTERRUPT; | |
1708 | *out++ = MI_NOOP; | |
1709 | request->tail = intel_ring_offset(request->ring, out); | |
1710 | ||
1711 | gen8_emit_wa_tail(request, out); | |
7c17d377 | 1712 | } |
4da46e1e | 1713 | |
98f29e8d CW |
1714 | static const int gen8_emit_breadcrumb_sz = 6 + WA_TAIL_DWORDS; |
1715 | ||
caddfe71 CW |
1716 | static void gen8_emit_breadcrumb_render(struct drm_i915_gem_request *request, |
1717 | u32 *out) | |
7c17d377 | 1718 | { |
ce81a65c MW |
1719 | /* We're using qword write, seqno should be aligned to 8 bytes. */ |
1720 | BUILD_BUG_ON(I915_GEM_HWS_INDEX & 1); | |
1721 | ||
7c17d377 CW |
1722 | /* w/a for post sync ops following a GPGPU operation we |
1723 | * need a prior CS_STALL, which is emitted by the flush | |
1724 | * following the batch. | |
1725 | */ | |
caddfe71 CW |
1726 | *out++ = GFX_OP_PIPE_CONTROL(6); |
1727 | *out++ = (PIPE_CONTROL_GLOBAL_GTT_IVB | | |
1728 | PIPE_CONTROL_CS_STALL | | |
1729 | PIPE_CONTROL_QW_WRITE); | |
1730 | *out++ = intel_hws_seqno_address(request->engine); | |
1731 | *out++ = 0; | |
1732 | *out++ = request->global_seqno; | |
ce81a65c | 1733 | /* We're thrashing one dword of HWS. */ |
caddfe71 CW |
1734 | *out++ = 0; |
1735 | *out++ = MI_USER_INTERRUPT; | |
1736 | *out++ = MI_NOOP; | |
1737 | request->tail = intel_ring_offset(request->ring, out); | |
1738 | ||
1739 | gen8_emit_wa_tail(request, out); | |
4da46e1e OM |
1740 | } |
1741 | ||
98f29e8d CW |
1742 | static const int gen8_emit_breadcrumb_render_sz = 8 + WA_TAIL_DWORDS; |
1743 | ||
8753181e | 1744 | static int gen8_init_rcs_context(struct drm_i915_gem_request *req) |
e7778be1 TD |
1745 | { |
1746 | int ret; | |
1747 | ||
e2be4faf | 1748 | ret = intel_logical_ring_workarounds_emit(req); |
e7778be1 TD |
1749 | if (ret) |
1750 | return ret; | |
1751 | ||
3bbaba0c PA |
1752 | ret = intel_rcs_context_init_mocs(req); |
1753 | /* | |
1754 | * Failing to program the MOCS is non-fatal.The system will not | |
1755 | * run at peak performance. So generate an error and carry on. | |
1756 | */ | |
1757 | if (ret) | |
1758 | DRM_ERROR("MOCS failed to program: expect performance issues.\n"); | |
1759 | ||
4e50f082 | 1760 | return i915_gem_render_state_emit(req); |
e7778be1 TD |
1761 | } |
1762 | ||
73e4d07f OM |
1763 | /** |
1764 | * intel_logical_ring_cleanup() - deallocate the Engine Command Streamer | |
14bb2c11 | 1765 | * @engine: Engine Command Streamer. |
73e4d07f | 1766 | */ |
0bc40be8 | 1767 | void intel_logical_ring_cleanup(struct intel_engine_cs *engine) |
454afebd | 1768 | { |
6402c330 | 1769 | struct drm_i915_private *dev_priv; |
9832b9da | 1770 | |
27af5eea TU |
1771 | /* |
1772 | * Tasklet cannot be active at this point due intel_mark_active/idle | |
1773 | * so this is just for documentation. | |
1774 | */ | |
1775 | if (WARN_ON(test_bit(TASKLET_STATE_SCHED, &engine->irq_tasklet.state))) | |
1776 | tasklet_kill(&engine->irq_tasklet); | |
1777 | ||
c033666a | 1778 | dev_priv = engine->i915; |
6402c330 | 1779 | |
0bc40be8 | 1780 | if (engine->buffer) { |
0bc40be8 | 1781 | WARN_ON((I915_READ_MODE(engine) & MODE_IDLE) == 0); |
b0366a54 | 1782 | } |
48d82387 | 1783 | |
0bc40be8 TU |
1784 | if (engine->cleanup) |
1785 | engine->cleanup(engine); | |
48d82387 | 1786 | |
96a945aa | 1787 | intel_engine_cleanup_common(engine); |
688e6c72 | 1788 | |
57e88531 CW |
1789 | if (engine->status_page.vma) { |
1790 | i915_gem_object_unpin_map(engine->status_page.vma->obj); | |
1791 | engine->status_page.vma = NULL; | |
48d82387 | 1792 | } |
24f1d3cc | 1793 | intel_lr_context_unpin(dev_priv->kernel_context, engine); |
17ee950d | 1794 | |
0bc40be8 | 1795 | lrc_destroy_wa_ctx_obj(engine); |
c033666a | 1796 | engine->i915 = NULL; |
3b3f1650 AG |
1797 | dev_priv->engine[engine->id] = NULL; |
1798 | kfree(engine); | |
454afebd OM |
1799 | } |
1800 | ||
ddd66c51 CW |
1801 | void intel_execlists_enable_submission(struct drm_i915_private *dev_priv) |
1802 | { | |
1803 | struct intel_engine_cs *engine; | |
3b3f1650 | 1804 | enum intel_engine_id id; |
ddd66c51 | 1805 | |
20311bd3 | 1806 | for_each_engine(engine, dev_priv, id) { |
f4ea6bdd | 1807 | engine->submit_request = execlists_submit_request; |
20311bd3 CW |
1808 | engine->schedule = execlists_schedule; |
1809 | } | |
ddd66c51 CW |
1810 | } |
1811 | ||
c9cacf93 | 1812 | static void |
e1382efb | 1813 | logical_ring_default_vfuncs(struct intel_engine_cs *engine) |
c9cacf93 TU |
1814 | { |
1815 | /* Default vfuncs which can be overriden by each engine. */ | |
0bc40be8 | 1816 | engine->init_hw = gen8_init_common_ring; |
821ed7df | 1817 | engine->reset_hw = reset_common_ring; |
0bc40be8 | 1818 | engine->emit_flush = gen8_emit_flush; |
9b81d556 | 1819 | engine->emit_breadcrumb = gen8_emit_breadcrumb; |
98f29e8d | 1820 | engine->emit_breadcrumb_sz = gen8_emit_breadcrumb_sz; |
f4ea6bdd | 1821 | engine->submit_request = execlists_submit_request; |
20311bd3 | 1822 | engine->schedule = execlists_schedule; |
ddd66c51 | 1823 | |
31bb59cc CW |
1824 | engine->irq_enable = gen8_logical_ring_enable_irq; |
1825 | engine->irq_disable = gen8_logical_ring_disable_irq; | |
0bc40be8 | 1826 | engine->emit_bb_start = gen8_emit_bb_start; |
1b7744e7 | 1827 | if (IS_BXT_REVID(engine->i915, 0, BXT_REVID_A1)) |
c04e0f3b | 1828 | engine->irq_seqno_barrier = bxt_a_seqno_barrier; |
c9cacf93 TU |
1829 | } |
1830 | ||
d9f3af96 | 1831 | static inline void |
c2c7f240 | 1832 | logical_ring_default_irqs(struct intel_engine_cs *engine) |
d9f3af96 | 1833 | { |
c2c7f240 | 1834 | unsigned shift = engine->irq_shift; |
0bc40be8 TU |
1835 | engine->irq_enable_mask = GT_RENDER_USER_INTERRUPT << shift; |
1836 | engine->irq_keep_mask = GT_CONTEXT_SWITCH_INTERRUPT << shift; | |
d9f3af96 TU |
1837 | } |
1838 | ||
7d774cac | 1839 | static int |
bf3783e5 | 1840 | lrc_setup_hws(struct intel_engine_cs *engine, struct i915_vma *vma) |
04794adb | 1841 | { |
57e88531 | 1842 | const int hws_offset = LRC_PPHWSP_PN * PAGE_SIZE; |
7d774cac | 1843 | void *hws; |
04794adb TU |
1844 | |
1845 | /* The HWSP is part of the default context object in LRC mode. */ | |
bf3783e5 | 1846 | hws = i915_gem_object_pin_map(vma->obj, I915_MAP_WB); |
7d774cac TU |
1847 | if (IS_ERR(hws)) |
1848 | return PTR_ERR(hws); | |
57e88531 CW |
1849 | |
1850 | engine->status_page.page_addr = hws + hws_offset; | |
bde13ebd | 1851 | engine->status_page.ggtt_offset = i915_ggtt_offset(vma) + hws_offset; |
57e88531 | 1852 | engine->status_page.vma = vma; |
7d774cac TU |
1853 | |
1854 | return 0; | |
04794adb TU |
1855 | } |
1856 | ||
bb45438f TU |
1857 | static void |
1858 | logical_ring_setup(struct intel_engine_cs *engine) | |
1859 | { | |
1860 | struct drm_i915_private *dev_priv = engine->i915; | |
1861 | enum forcewake_domains fw_domains; | |
1862 | ||
019bf277 TU |
1863 | intel_engine_setup_common(engine); |
1864 | ||
bb45438f TU |
1865 | /* Intentionally left blank. */ |
1866 | engine->buffer = NULL; | |
1867 | ||
1868 | fw_domains = intel_uncore_forcewake_for_reg(dev_priv, | |
1869 | RING_ELSP(engine), | |
1870 | FW_REG_WRITE); | |
1871 | ||
1872 | fw_domains |= intel_uncore_forcewake_for_reg(dev_priv, | |
1873 | RING_CONTEXT_STATUS_PTR(engine), | |
1874 | FW_REG_READ | FW_REG_WRITE); | |
1875 | ||
1876 | fw_domains |= intel_uncore_forcewake_for_reg(dev_priv, | |
1877 | RING_CONTEXT_STATUS_BUF_BASE(engine), | |
1878 | FW_REG_READ); | |
1879 | ||
1880 | engine->fw_domains = fw_domains; | |
1881 | ||
bb45438f TU |
1882 | tasklet_init(&engine->irq_tasklet, |
1883 | intel_lrc_irq_handler, (unsigned long)engine); | |
1884 | ||
1885 | logical_ring_init_platform_invariants(engine); | |
1886 | logical_ring_default_vfuncs(engine); | |
1887 | logical_ring_default_irqs(engine); | |
bb45438f TU |
1888 | } |
1889 | ||
a19d6ff2 TU |
1890 | static int |
1891 | logical_ring_init(struct intel_engine_cs *engine) | |
1892 | { | |
1893 | struct i915_gem_context *dctx = engine->i915->kernel_context; | |
1894 | int ret; | |
1895 | ||
019bf277 | 1896 | ret = intel_engine_init_common(engine); |
a19d6ff2 TU |
1897 | if (ret) |
1898 | goto error; | |
1899 | ||
1900 | ret = execlists_context_deferred_alloc(dctx, engine); | |
1901 | if (ret) | |
1902 | goto error; | |
1903 | ||
1904 | /* As this is the default context, always pin it */ | |
1905 | ret = intel_lr_context_pin(dctx, engine); | |
1906 | if (ret) { | |
1907 | DRM_ERROR("Failed to pin context for %s: %d\n", | |
1908 | engine->name, ret); | |
1909 | goto error; | |
1910 | } | |
1911 | ||
1912 | /* And setup the hardware status page. */ | |
1913 | ret = lrc_setup_hws(engine, dctx->engine[engine->id].state); | |
1914 | if (ret) { | |
1915 | DRM_ERROR("Failed to set up hws %s: %d\n", engine->name, ret); | |
1916 | goto error; | |
1917 | } | |
1918 | ||
1919 | return 0; | |
1920 | ||
1921 | error: | |
1922 | intel_logical_ring_cleanup(engine); | |
1923 | return ret; | |
1924 | } | |
1925 | ||
88d2ba2e | 1926 | int logical_render_ring_init(struct intel_engine_cs *engine) |
a19d6ff2 TU |
1927 | { |
1928 | struct drm_i915_private *dev_priv = engine->i915; | |
1929 | int ret; | |
1930 | ||
bb45438f TU |
1931 | logical_ring_setup(engine); |
1932 | ||
a19d6ff2 TU |
1933 | if (HAS_L3_DPF(dev_priv)) |
1934 | engine->irq_keep_mask |= GT_RENDER_L3_PARITY_ERROR_INTERRUPT; | |
1935 | ||
1936 | /* Override some for render ring. */ | |
1937 | if (INTEL_GEN(dev_priv) >= 9) | |
1938 | engine->init_hw = gen9_init_render_ring; | |
1939 | else | |
1940 | engine->init_hw = gen8_init_render_ring; | |
1941 | engine->init_context = gen8_init_rcs_context; | |
a19d6ff2 | 1942 | engine->emit_flush = gen8_emit_flush_render; |
9b81d556 | 1943 | engine->emit_breadcrumb = gen8_emit_breadcrumb_render; |
98f29e8d | 1944 | engine->emit_breadcrumb_sz = gen8_emit_breadcrumb_render_sz; |
a19d6ff2 | 1945 | |
56c0f1a7 | 1946 | ret = intel_engine_create_scratch(engine, 4096); |
a19d6ff2 TU |
1947 | if (ret) |
1948 | return ret; | |
1949 | ||
1950 | ret = intel_init_workaround_bb(engine); | |
1951 | if (ret) { | |
1952 | /* | |
1953 | * We continue even if we fail to initialize WA batch | |
1954 | * because we only expect rare glitches but nothing | |
1955 | * critical to prevent us from using GPU | |
1956 | */ | |
1957 | DRM_ERROR("WA batch buffer initialization failed: %d\n", | |
1958 | ret); | |
1959 | } | |
1960 | ||
d8953c83 | 1961 | return logical_ring_init(engine); |
a19d6ff2 TU |
1962 | } |
1963 | ||
88d2ba2e | 1964 | int logical_xcs_ring_init(struct intel_engine_cs *engine) |
bb45438f TU |
1965 | { |
1966 | logical_ring_setup(engine); | |
1967 | ||
1968 | return logical_ring_init(engine); | |
454afebd OM |
1969 | } |
1970 | ||
0cea6502 | 1971 | static u32 |
c033666a | 1972 | make_rpcs(struct drm_i915_private *dev_priv) |
0cea6502 JM |
1973 | { |
1974 | u32 rpcs = 0; | |
1975 | ||
1976 | /* | |
1977 | * No explicit RPCS request is needed to ensure full | |
1978 | * slice/subslice/EU enablement prior to Gen9. | |
1979 | */ | |
c033666a | 1980 | if (INTEL_GEN(dev_priv) < 9) |
0cea6502 JM |
1981 | return 0; |
1982 | ||
1983 | /* | |
1984 | * Starting in Gen9, render power gating can leave | |
1985 | * slice/subslice/EU in a partially enabled state. We | |
1986 | * must make an explicit request through RPCS for full | |
1987 | * enablement. | |
1988 | */ | |
43b67998 | 1989 | if (INTEL_INFO(dev_priv)->sseu.has_slice_pg) { |
0cea6502 | 1990 | rpcs |= GEN8_RPCS_S_CNT_ENABLE; |
f08a0c92 | 1991 | rpcs |= hweight8(INTEL_INFO(dev_priv)->sseu.slice_mask) << |
0cea6502 JM |
1992 | GEN8_RPCS_S_CNT_SHIFT; |
1993 | rpcs |= GEN8_RPCS_ENABLE; | |
1994 | } | |
1995 | ||
43b67998 | 1996 | if (INTEL_INFO(dev_priv)->sseu.has_subslice_pg) { |
0cea6502 | 1997 | rpcs |= GEN8_RPCS_SS_CNT_ENABLE; |
57ec171e | 1998 | rpcs |= hweight8(INTEL_INFO(dev_priv)->sseu.subslice_mask) << |
0cea6502 JM |
1999 | GEN8_RPCS_SS_CNT_SHIFT; |
2000 | rpcs |= GEN8_RPCS_ENABLE; | |
2001 | } | |
2002 | ||
43b67998 ID |
2003 | if (INTEL_INFO(dev_priv)->sseu.has_eu_pg) { |
2004 | rpcs |= INTEL_INFO(dev_priv)->sseu.eu_per_subslice << | |
0cea6502 | 2005 | GEN8_RPCS_EU_MIN_SHIFT; |
43b67998 | 2006 | rpcs |= INTEL_INFO(dev_priv)->sseu.eu_per_subslice << |
0cea6502 JM |
2007 | GEN8_RPCS_EU_MAX_SHIFT; |
2008 | rpcs |= GEN8_RPCS_ENABLE; | |
2009 | } | |
2010 | ||
2011 | return rpcs; | |
2012 | } | |
2013 | ||
0bc40be8 | 2014 | static u32 intel_lr_indirect_ctx_offset(struct intel_engine_cs *engine) |
71562919 MT |
2015 | { |
2016 | u32 indirect_ctx_offset; | |
2017 | ||
c033666a | 2018 | switch (INTEL_GEN(engine->i915)) { |
71562919 | 2019 | default: |
c033666a | 2020 | MISSING_CASE(INTEL_GEN(engine->i915)); |
71562919 MT |
2021 | /* fall through */ |
2022 | case 9: | |
2023 | indirect_ctx_offset = | |
2024 | GEN9_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT; | |
2025 | break; | |
2026 | case 8: | |
2027 | indirect_ctx_offset = | |
2028 | GEN8_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT; | |
2029 | break; | |
2030 | } | |
2031 | ||
2032 | return indirect_ctx_offset; | |
2033 | } | |
2034 | ||
a3aabe86 CW |
2035 | static void execlists_init_reg_state(u32 *reg_state, |
2036 | struct i915_gem_context *ctx, | |
2037 | struct intel_engine_cs *engine, | |
2038 | struct intel_ring *ring) | |
8670d6f9 | 2039 | { |
a3aabe86 CW |
2040 | struct drm_i915_private *dev_priv = engine->i915; |
2041 | struct i915_hw_ppgtt *ppgtt = ctx->ppgtt ?: dev_priv->mm.aliasing_ppgtt; | |
8670d6f9 OM |
2042 | |
2043 | /* A context is actually a big batch buffer with several MI_LOAD_REGISTER_IMM | |
2044 | * commands followed by (reg, value) pairs. The values we are setting here are | |
2045 | * only for the first context restore: on a subsequent save, the GPU will | |
2046 | * recreate this batchbuffer with new values (including all the missing | |
2047 | * MI_LOAD_REGISTER_IMM commands that we are not initializing here). */ | |
0d925ea0 | 2048 | reg_state[CTX_LRI_HEADER_0] = |
0bc40be8 TU |
2049 | MI_LOAD_REGISTER_IMM(engine->id == RCS ? 14 : 11) | MI_LRI_FORCE_POSTED; |
2050 | ASSIGN_CTX_REG(reg_state, CTX_CONTEXT_CONTROL, | |
2051 | RING_CONTEXT_CONTROL(engine), | |
0d925ea0 VS |
2052 | _MASKED_BIT_ENABLE(CTX_CTRL_INHIBIT_SYN_CTX_SWITCH | |
2053 | CTX_CTRL_ENGINE_CTX_RESTORE_INHIBIT | | |
c033666a | 2054 | (HAS_RESOURCE_STREAMER(dev_priv) ? |
a3aabe86 | 2055 | CTX_CTRL_RS_CTX_ENABLE : 0))); |
0bc40be8 TU |
2056 | ASSIGN_CTX_REG(reg_state, CTX_RING_HEAD, RING_HEAD(engine->mmio_base), |
2057 | 0); | |
2058 | ASSIGN_CTX_REG(reg_state, CTX_RING_TAIL, RING_TAIL(engine->mmio_base), | |
2059 | 0); | |
0bc40be8 TU |
2060 | ASSIGN_CTX_REG(reg_state, CTX_RING_BUFFER_START, |
2061 | RING_START(engine->mmio_base), 0); | |
2062 | ASSIGN_CTX_REG(reg_state, CTX_RING_BUFFER_CONTROL, | |
2063 | RING_CTL(engine->mmio_base), | |
62ae14b1 | 2064 | RING_CTL_SIZE(ring->size) | RING_VALID); |
0bc40be8 TU |
2065 | ASSIGN_CTX_REG(reg_state, CTX_BB_HEAD_U, |
2066 | RING_BBADDR_UDW(engine->mmio_base), 0); | |
2067 | ASSIGN_CTX_REG(reg_state, CTX_BB_HEAD_L, | |
2068 | RING_BBADDR(engine->mmio_base), 0); | |
2069 | ASSIGN_CTX_REG(reg_state, CTX_BB_STATE, | |
2070 | RING_BBSTATE(engine->mmio_base), | |
0d925ea0 | 2071 | RING_BB_PPGTT); |
0bc40be8 TU |
2072 | ASSIGN_CTX_REG(reg_state, CTX_SECOND_BB_HEAD_U, |
2073 | RING_SBBADDR_UDW(engine->mmio_base), 0); | |
2074 | ASSIGN_CTX_REG(reg_state, CTX_SECOND_BB_HEAD_L, | |
2075 | RING_SBBADDR(engine->mmio_base), 0); | |
2076 | ASSIGN_CTX_REG(reg_state, CTX_SECOND_BB_STATE, | |
2077 | RING_SBBSTATE(engine->mmio_base), 0); | |
2078 | if (engine->id == RCS) { | |
2079 | ASSIGN_CTX_REG(reg_state, CTX_BB_PER_CTX_PTR, | |
2080 | RING_BB_PER_CTX_PTR(engine->mmio_base), 0); | |
2081 | ASSIGN_CTX_REG(reg_state, CTX_RCS_INDIRECT_CTX, | |
2082 | RING_INDIRECT_CTX(engine->mmio_base), 0); | |
2083 | ASSIGN_CTX_REG(reg_state, CTX_RCS_INDIRECT_CTX_OFFSET, | |
2084 | RING_INDIRECT_CTX_OFFSET(engine->mmio_base), 0); | |
48bb74e4 | 2085 | if (engine->wa_ctx.vma) { |
0bc40be8 | 2086 | struct i915_ctx_workarounds *wa_ctx = &engine->wa_ctx; |
bde13ebd | 2087 | u32 ggtt_offset = i915_ggtt_offset(wa_ctx->vma); |
17ee950d AS |
2088 | |
2089 | reg_state[CTX_RCS_INDIRECT_CTX+1] = | |
2090 | (ggtt_offset + wa_ctx->indirect_ctx.offset * sizeof(uint32_t)) | | |
2091 | (wa_ctx->indirect_ctx.size / CACHELINE_DWORDS); | |
2092 | ||
2093 | reg_state[CTX_RCS_INDIRECT_CTX_OFFSET+1] = | |
0bc40be8 | 2094 | intel_lr_indirect_ctx_offset(engine) << 6; |
17ee950d AS |
2095 | |
2096 | reg_state[CTX_BB_PER_CTX_PTR+1] = | |
2097 | (ggtt_offset + wa_ctx->per_ctx.offset * sizeof(uint32_t)) | | |
2098 | 0x01; | |
2099 | } | |
8670d6f9 | 2100 | } |
0d925ea0 | 2101 | reg_state[CTX_LRI_HEADER_1] = MI_LOAD_REGISTER_IMM(9) | MI_LRI_FORCE_POSTED; |
0bc40be8 TU |
2102 | ASSIGN_CTX_REG(reg_state, CTX_CTX_TIMESTAMP, |
2103 | RING_CTX_TIMESTAMP(engine->mmio_base), 0); | |
0d925ea0 | 2104 | /* PDP values well be assigned later if needed */ |
0bc40be8 TU |
2105 | ASSIGN_CTX_REG(reg_state, CTX_PDP3_UDW, GEN8_RING_PDP_UDW(engine, 3), |
2106 | 0); | |
2107 | ASSIGN_CTX_REG(reg_state, CTX_PDP3_LDW, GEN8_RING_PDP_LDW(engine, 3), | |
2108 | 0); | |
2109 | ASSIGN_CTX_REG(reg_state, CTX_PDP2_UDW, GEN8_RING_PDP_UDW(engine, 2), | |
2110 | 0); | |
2111 | ASSIGN_CTX_REG(reg_state, CTX_PDP2_LDW, GEN8_RING_PDP_LDW(engine, 2), | |
2112 | 0); | |
2113 | ASSIGN_CTX_REG(reg_state, CTX_PDP1_UDW, GEN8_RING_PDP_UDW(engine, 1), | |
2114 | 0); | |
2115 | ASSIGN_CTX_REG(reg_state, CTX_PDP1_LDW, GEN8_RING_PDP_LDW(engine, 1), | |
2116 | 0); | |
2117 | ASSIGN_CTX_REG(reg_state, CTX_PDP0_UDW, GEN8_RING_PDP_UDW(engine, 0), | |
2118 | 0); | |
2119 | ASSIGN_CTX_REG(reg_state, CTX_PDP0_LDW, GEN8_RING_PDP_LDW(engine, 0), | |
2120 | 0); | |
d7b2633d | 2121 | |
2dba3239 MT |
2122 | if (USES_FULL_48BIT_PPGTT(ppgtt->base.dev)) { |
2123 | /* 64b PPGTT (48bit canonical) | |
2124 | * PDP0_DESCRIPTOR contains the base address to PML4 and | |
2125 | * other PDP Descriptors are ignored. | |
2126 | */ | |
2127 | ASSIGN_CTX_PML4(ppgtt, reg_state); | |
2128 | } else { | |
2129 | /* 32b PPGTT | |
2130 | * PDP*_DESCRIPTOR contains the base address of space supported. | |
2131 | * With dynamic page allocation, PDPs may not be allocated at | |
2132 | * this point. Point the unallocated PDPs to the scratch page | |
2133 | */ | |
c6a2ac71 | 2134 | execlists_update_context_pdps(ppgtt, reg_state); |
2dba3239 MT |
2135 | } |
2136 | ||
0bc40be8 | 2137 | if (engine->id == RCS) { |
8670d6f9 | 2138 | reg_state[CTX_LRI_HEADER_2] = MI_LOAD_REGISTER_IMM(1); |
0d925ea0 | 2139 | ASSIGN_CTX_REG(reg_state, CTX_R_PWR_CLK_STATE, GEN8_R_PWR_CLK_STATE, |
c033666a | 2140 | make_rpcs(dev_priv)); |
8670d6f9 | 2141 | } |
a3aabe86 CW |
2142 | } |
2143 | ||
2144 | static int | |
2145 | populate_lr_context(struct i915_gem_context *ctx, | |
2146 | struct drm_i915_gem_object *ctx_obj, | |
2147 | struct intel_engine_cs *engine, | |
2148 | struct intel_ring *ring) | |
2149 | { | |
2150 | void *vaddr; | |
2151 | int ret; | |
2152 | ||
2153 | ret = i915_gem_object_set_to_cpu_domain(ctx_obj, true); | |
2154 | if (ret) { | |
2155 | DRM_DEBUG_DRIVER("Could not set to CPU domain\n"); | |
2156 | return ret; | |
2157 | } | |
2158 | ||
2159 | vaddr = i915_gem_object_pin_map(ctx_obj, I915_MAP_WB); | |
2160 | if (IS_ERR(vaddr)) { | |
2161 | ret = PTR_ERR(vaddr); | |
2162 | DRM_DEBUG_DRIVER("Could not map object pages! (%d)\n", ret); | |
2163 | return ret; | |
2164 | } | |
a4f5ea64 | 2165 | ctx_obj->mm.dirty = true; |
a3aabe86 CW |
2166 | |
2167 | /* The second page of the context object contains some fields which must | |
2168 | * be set up prior to the first execution. */ | |
2169 | ||
2170 | execlists_init_reg_state(vaddr + LRC_STATE_PN * PAGE_SIZE, | |
2171 | ctx, engine, ring); | |
8670d6f9 | 2172 | |
7d774cac | 2173 | i915_gem_object_unpin_map(ctx_obj); |
8670d6f9 OM |
2174 | |
2175 | return 0; | |
2176 | } | |
2177 | ||
c5d46ee2 DG |
2178 | /** |
2179 | * intel_lr_context_size() - return the size of the context for an engine | |
14bb2c11 | 2180 | * @engine: which engine to find the context size for |
c5d46ee2 DG |
2181 | * |
2182 | * Each engine may require a different amount of space for a context image, | |
2183 | * so when allocating (or copying) an image, this function can be used to | |
2184 | * find the right size for the specific engine. | |
2185 | * | |
2186 | * Return: size (in bytes) of an engine-specific context image | |
2187 | * | |
2188 | * Note: this size includes the HWSP, which is part of the context image | |
2189 | * in LRC mode, but does not include the "shared data page" used with | |
2190 | * GuC submission. The caller should account for this if using the GuC. | |
2191 | */ | |
0bc40be8 | 2192 | uint32_t intel_lr_context_size(struct intel_engine_cs *engine) |
8c857917 OM |
2193 | { |
2194 | int ret = 0; | |
2195 | ||
c033666a | 2196 | WARN_ON(INTEL_GEN(engine->i915) < 8); |
8c857917 | 2197 | |
0bc40be8 | 2198 | switch (engine->id) { |
8c857917 | 2199 | case RCS: |
c033666a | 2200 | if (INTEL_GEN(engine->i915) >= 9) |
468c6816 MN |
2201 | ret = GEN9_LR_CONTEXT_RENDER_SIZE; |
2202 | else | |
2203 | ret = GEN8_LR_CONTEXT_RENDER_SIZE; | |
8c857917 OM |
2204 | break; |
2205 | case VCS: | |
2206 | case BCS: | |
2207 | case VECS: | |
2208 | case VCS2: | |
2209 | ret = GEN8_LR_CONTEXT_OTHER_SIZE; | |
2210 | break; | |
2211 | } | |
2212 | ||
2213 | return ret; | |
ede7d42b OM |
2214 | } |
2215 | ||
e2efd130 | 2216 | static int execlists_context_deferred_alloc(struct i915_gem_context *ctx, |
978f1e09 | 2217 | struct intel_engine_cs *engine) |
ede7d42b | 2218 | { |
8c857917 | 2219 | struct drm_i915_gem_object *ctx_obj; |
9021ad03 | 2220 | struct intel_context *ce = &ctx->engine[engine->id]; |
bf3783e5 | 2221 | struct i915_vma *vma; |
8c857917 | 2222 | uint32_t context_size; |
7e37f889 | 2223 | struct intel_ring *ring; |
8c857917 OM |
2224 | int ret; |
2225 | ||
9021ad03 | 2226 | WARN_ON(ce->state); |
ede7d42b | 2227 | |
0bc40be8 | 2228 | context_size = round_up(intel_lr_context_size(engine), 4096); |
8c857917 | 2229 | |
d1675198 AD |
2230 | /* One extra page as the sharing data between driver and GuC */ |
2231 | context_size += PAGE_SIZE * LRC_PPHWSP_PN; | |
2232 | ||
91c8a326 | 2233 | ctx_obj = i915_gem_object_create(&ctx->i915->drm, context_size); |
fe3db79b | 2234 | if (IS_ERR(ctx_obj)) { |
3126a660 | 2235 | DRM_DEBUG_DRIVER("Alloc LRC backing obj failed.\n"); |
fe3db79b | 2236 | return PTR_ERR(ctx_obj); |
8c857917 OM |
2237 | } |
2238 | ||
bf3783e5 CW |
2239 | vma = i915_vma_create(ctx_obj, &ctx->i915->ggtt.base, NULL); |
2240 | if (IS_ERR(vma)) { | |
2241 | ret = PTR_ERR(vma); | |
2242 | goto error_deref_obj; | |
2243 | } | |
2244 | ||
7e37f889 | 2245 | ring = intel_engine_create_ring(engine, ctx->ring_size); |
dca33ecc CW |
2246 | if (IS_ERR(ring)) { |
2247 | ret = PTR_ERR(ring); | |
e84fe803 | 2248 | goto error_deref_obj; |
8670d6f9 OM |
2249 | } |
2250 | ||
dca33ecc | 2251 | ret = populate_lr_context(ctx, ctx_obj, engine, ring); |
8670d6f9 OM |
2252 | if (ret) { |
2253 | DRM_DEBUG_DRIVER("Failed to populate LRC: %d\n", ret); | |
dca33ecc | 2254 | goto error_ring_free; |
84c2377f OM |
2255 | } |
2256 | ||
dca33ecc | 2257 | ce->ring = ring; |
bf3783e5 | 2258 | ce->state = vma; |
9021ad03 | 2259 | ce->initialised = engine->init_context == NULL; |
ede7d42b OM |
2260 | |
2261 | return 0; | |
8670d6f9 | 2262 | |
dca33ecc | 2263 | error_ring_free: |
7e37f889 | 2264 | intel_ring_free(ring); |
e84fe803 | 2265 | error_deref_obj: |
f8c417cd | 2266 | i915_gem_object_put(ctx_obj); |
8670d6f9 | 2267 | return ret; |
ede7d42b | 2268 | } |
3e5b6f05 | 2269 | |
821ed7df | 2270 | void intel_lr_context_resume(struct drm_i915_private *dev_priv) |
3e5b6f05 | 2271 | { |
e2f80391 | 2272 | struct intel_engine_cs *engine; |
bafb2f7d | 2273 | struct i915_gem_context *ctx; |
3b3f1650 | 2274 | enum intel_engine_id id; |
bafb2f7d CW |
2275 | |
2276 | /* Because we emit WA_TAIL_DWORDS there may be a disparity | |
2277 | * between our bookkeeping in ce->ring->head and ce->ring->tail and | |
2278 | * that stored in context. As we only write new commands from | |
2279 | * ce->ring->tail onwards, everything before that is junk. If the GPU | |
2280 | * starts reading from its RING_HEAD from the context, it may try to | |
2281 | * execute that junk and die. | |
2282 | * | |
2283 | * So to avoid that we reset the context images upon resume. For | |
2284 | * simplicity, we just zero everything out. | |
2285 | */ | |
2286 | list_for_each_entry(ctx, &dev_priv->context_list, link) { | |
3b3f1650 | 2287 | for_each_engine(engine, dev_priv, id) { |
bafb2f7d CW |
2288 | struct intel_context *ce = &ctx->engine[engine->id]; |
2289 | u32 *reg; | |
3e5b6f05 | 2290 | |
bafb2f7d CW |
2291 | if (!ce->state) |
2292 | continue; | |
7d774cac | 2293 | |
bafb2f7d CW |
2294 | reg = i915_gem_object_pin_map(ce->state->obj, |
2295 | I915_MAP_WB); | |
2296 | if (WARN_ON(IS_ERR(reg))) | |
2297 | continue; | |
3e5b6f05 | 2298 | |
bafb2f7d CW |
2299 | reg += LRC_STATE_PN * PAGE_SIZE / sizeof(*reg); |
2300 | reg[CTX_RING_HEAD+1] = 0; | |
2301 | reg[CTX_RING_TAIL+1] = 0; | |
3e5b6f05 | 2302 | |
a4f5ea64 | 2303 | ce->state->obj->mm.dirty = true; |
bafb2f7d | 2304 | i915_gem_object_unpin_map(ce->state->obj); |
3e5b6f05 | 2305 | |
bafb2f7d CW |
2306 | ce->ring->head = ce->ring->tail = 0; |
2307 | ce->ring->last_retired_head = -1; | |
2308 | intel_ring_update_space(ce->ring); | |
2309 | } | |
3e5b6f05 TD |
2310 | } |
2311 | } |