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 OM |
133 | */ |
134 | ||
135 | #include <drm/drmP.h> | |
136 | #include <drm/i915_drm.h> | |
137 | #include "i915_drv.h" | |
3bbaba0c | 138 | #include "intel_mocs.h" |
127f1003 | 139 | |
468c6816 | 140 | #define GEN9_LR_CONTEXT_RENDER_SIZE (22 * PAGE_SIZE) |
8c857917 OM |
141 | #define GEN8_LR_CONTEXT_RENDER_SIZE (20 * PAGE_SIZE) |
142 | #define GEN8_LR_CONTEXT_OTHER_SIZE (2 * PAGE_SIZE) | |
143 | ||
e981e7b1 TD |
144 | #define RING_EXECLIST_QFULL (1 << 0x2) |
145 | #define RING_EXECLIST1_VALID (1 << 0x3) | |
146 | #define RING_EXECLIST0_VALID (1 << 0x4) | |
147 | #define RING_EXECLIST_ACTIVE_STATUS (3 << 0xE) | |
148 | #define RING_EXECLIST1_ACTIVE (1 << 0x11) | |
149 | #define RING_EXECLIST0_ACTIVE (1 << 0x12) | |
150 | ||
151 | #define GEN8_CTX_STATUS_IDLE_ACTIVE (1 << 0) | |
152 | #define GEN8_CTX_STATUS_PREEMPTED (1 << 1) | |
153 | #define GEN8_CTX_STATUS_ELEMENT_SWITCH (1 << 2) | |
154 | #define GEN8_CTX_STATUS_ACTIVE_IDLE (1 << 3) | |
155 | #define GEN8_CTX_STATUS_COMPLETE (1 << 4) | |
156 | #define GEN8_CTX_STATUS_LITE_RESTORE (1 << 15) | |
8670d6f9 OM |
157 | |
158 | #define CTX_LRI_HEADER_0 0x01 | |
159 | #define CTX_CONTEXT_CONTROL 0x02 | |
160 | #define CTX_RING_HEAD 0x04 | |
161 | #define CTX_RING_TAIL 0x06 | |
162 | #define CTX_RING_BUFFER_START 0x08 | |
163 | #define CTX_RING_BUFFER_CONTROL 0x0a | |
164 | #define CTX_BB_HEAD_U 0x0c | |
165 | #define CTX_BB_HEAD_L 0x0e | |
166 | #define CTX_BB_STATE 0x10 | |
167 | #define CTX_SECOND_BB_HEAD_U 0x12 | |
168 | #define CTX_SECOND_BB_HEAD_L 0x14 | |
169 | #define CTX_SECOND_BB_STATE 0x16 | |
170 | #define CTX_BB_PER_CTX_PTR 0x18 | |
171 | #define CTX_RCS_INDIRECT_CTX 0x1a | |
172 | #define CTX_RCS_INDIRECT_CTX_OFFSET 0x1c | |
173 | #define CTX_LRI_HEADER_1 0x21 | |
174 | #define CTX_CTX_TIMESTAMP 0x22 | |
175 | #define CTX_PDP3_UDW 0x24 | |
176 | #define CTX_PDP3_LDW 0x26 | |
177 | #define CTX_PDP2_UDW 0x28 | |
178 | #define CTX_PDP2_LDW 0x2a | |
179 | #define CTX_PDP1_UDW 0x2c | |
180 | #define CTX_PDP1_LDW 0x2e | |
181 | #define CTX_PDP0_UDW 0x30 | |
182 | #define CTX_PDP0_LDW 0x32 | |
183 | #define CTX_LRI_HEADER_2 0x41 | |
184 | #define CTX_R_PWR_CLK_STATE 0x42 | |
185 | #define CTX_GPGPU_CSR_BASE_ADDRESS 0x44 | |
186 | ||
84b790f8 BW |
187 | #define GEN8_CTX_VALID (1<<0) |
188 | #define GEN8_CTX_FORCE_PD_RESTORE (1<<1) | |
189 | #define GEN8_CTX_FORCE_RESTORE (1<<2) | |
190 | #define GEN8_CTX_L3LLC_COHERENT (1<<5) | |
191 | #define GEN8_CTX_PRIVILEGE (1<<8) | |
e5815a2e | 192 | |
0d925ea0 | 193 | #define ASSIGN_CTX_REG(reg_state, pos, reg, val) do { \ |
f0f59a00 | 194 | (reg_state)[(pos)+0] = i915_mmio_reg_offset(reg); \ |
0d925ea0 VS |
195 | (reg_state)[(pos)+1] = (val); \ |
196 | } while (0) | |
197 | ||
198 | #define ASSIGN_CTX_PDP(ppgtt, reg_state, n) do { \ | |
d852c7bf | 199 | const u64 _addr = i915_page_dir_dma_addr((ppgtt), (n)); \ |
e5815a2e MT |
200 | reg_state[CTX_PDP ## n ## _UDW+1] = upper_32_bits(_addr); \ |
201 | reg_state[CTX_PDP ## n ## _LDW+1] = lower_32_bits(_addr); \ | |
9244a817 | 202 | } while (0) |
e5815a2e | 203 | |
9244a817 | 204 | #define ASSIGN_CTX_PML4(ppgtt, reg_state) do { \ |
2dba3239 MT |
205 | reg_state[CTX_PDP0_UDW + 1] = upper_32_bits(px_dma(&ppgtt->pml4)); \ |
206 | reg_state[CTX_PDP0_LDW + 1] = lower_32_bits(px_dma(&ppgtt->pml4)); \ | |
9244a817 | 207 | } while (0) |
2dba3239 | 208 | |
84b790f8 BW |
209 | enum { |
210 | ADVANCED_CONTEXT = 0, | |
2dba3239 | 211 | LEGACY_32B_CONTEXT, |
84b790f8 BW |
212 | ADVANCED_AD_CONTEXT, |
213 | LEGACY_64B_CONTEXT | |
214 | }; | |
2dba3239 MT |
215 | #define GEN8_CTX_ADDRESSING_MODE_SHIFT 3 |
216 | #define GEN8_CTX_ADDRESSING_MODE(dev) (USES_FULL_48BIT_PPGTT(dev) ?\ | |
217 | LEGACY_64B_CONTEXT :\ | |
218 | LEGACY_32B_CONTEXT) | |
84b790f8 BW |
219 | enum { |
220 | FAULT_AND_HANG = 0, | |
221 | FAULT_AND_HALT, /* Debug only */ | |
222 | FAULT_AND_STREAM, | |
223 | FAULT_AND_CONTINUE /* Unsupported */ | |
224 | }; | |
225 | #define GEN8_CTX_ID_SHIFT 32 | |
17ee950d | 226 | #define CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT 0x17 |
84b790f8 | 227 | |
8ba319da | 228 | static int intel_lr_context_pin(struct drm_i915_gem_request *rq); |
e84fe803 NH |
229 | static void lrc_setup_hardware_status_page(struct intel_engine_cs *ring, |
230 | struct drm_i915_gem_object *default_ctx_obj); | |
231 | ||
7ba717cf | 232 | |
73e4d07f OM |
233 | /** |
234 | * intel_sanitize_enable_execlists() - sanitize i915.enable_execlists | |
235 | * @dev: DRM device. | |
236 | * @enable_execlists: value of i915.enable_execlists module parameter. | |
237 | * | |
238 | * Only certain platforms support Execlists (the prerequisites being | |
27401d12 | 239 | * support for Logical Ring Contexts and Aliasing PPGTT or better). |
73e4d07f OM |
240 | * |
241 | * Return: 1 if Execlists is supported and has to be enabled. | |
242 | */ | |
127f1003 OM |
243 | int intel_sanitize_enable_execlists(struct drm_device *dev, int enable_execlists) |
244 | { | |
bd84b1e9 DV |
245 | WARN_ON(i915.enable_ppgtt == -1); |
246 | ||
a0bd6c31 ZL |
247 | /* On platforms with execlist available, vGPU will only |
248 | * support execlist mode, no ring buffer mode. | |
249 | */ | |
250 | if (HAS_LOGICAL_RING_CONTEXTS(dev) && intel_vgpu_active(dev)) | |
251 | return 1; | |
252 | ||
70ee45e1 DL |
253 | if (INTEL_INFO(dev)->gen >= 9) |
254 | return 1; | |
255 | ||
127f1003 OM |
256 | if (enable_execlists == 0) |
257 | return 0; | |
258 | ||
14bf993e OM |
259 | if (HAS_LOGICAL_RING_CONTEXTS(dev) && USES_PPGTT(dev) && |
260 | i915.use_mmio_flip >= 0) | |
127f1003 OM |
261 | return 1; |
262 | ||
263 | return 0; | |
264 | } | |
ede7d42b | 265 | |
ca82580c TU |
266 | static void |
267 | logical_ring_init_platform_invariants(struct intel_engine_cs *ring) | |
268 | { | |
269 | struct drm_device *dev = ring->dev; | |
270 | ||
271 | ring->disable_lite_restore_wa = (IS_SKL_REVID(dev, 0, SKL_REVID_B0) || | |
272 | IS_BXT_REVID(dev, 0, BXT_REVID_A1)) && | |
273 | (ring->id == VCS || ring->id == VCS2); | |
274 | ||
275 | ring->ctx_desc_template = GEN8_CTX_VALID; | |
276 | ring->ctx_desc_template |= GEN8_CTX_ADDRESSING_MODE(dev) << | |
277 | GEN8_CTX_ADDRESSING_MODE_SHIFT; | |
278 | if (IS_GEN8(dev)) | |
279 | ring->ctx_desc_template |= GEN8_CTX_L3LLC_COHERENT; | |
280 | ring->ctx_desc_template |= GEN8_CTX_PRIVILEGE; | |
281 | ||
282 | /* TODO: WaDisableLiteRestore when we start using semaphore | |
283 | * signalling between Command Streamers */ | |
284 | /* ring->ctx_desc_template |= GEN8_CTX_FORCE_RESTORE; */ | |
285 | ||
286 | /* WaEnableForceRestoreInCtxtDescForVCS:skl */ | |
287 | /* WaEnableForceRestoreInCtxtDescForVCS:bxt */ | |
288 | if (ring->disable_lite_restore_wa) | |
289 | ring->ctx_desc_template |= GEN8_CTX_FORCE_RESTORE; | |
290 | } | |
291 | ||
73e4d07f | 292 | /** |
ca82580c TU |
293 | * intel_lr_context_descriptor_update() - calculate & cache the descriptor |
294 | * descriptor for a pinned context | |
73e4d07f | 295 | * |
ca82580c TU |
296 | * @ctx: Context to work on |
297 | * @ring: Engine the descriptor will be used with | |
73e4d07f | 298 | * |
ca82580c TU |
299 | * The context descriptor encodes various attributes of a context, |
300 | * including its GTT address and some flags. Because it's fairly | |
301 | * expensive to calculate, we'll just do it once and cache the result, | |
302 | * which remains valid until the context is unpinned. | |
303 | * | |
304 | * This is what a descriptor looks like, from LSB to MSB: | |
305 | * bits 0-11: flags, GEN8_CTX_* (cached in ctx_desc_template) | |
306 | * bits 12-31: LRCA, GTT address of (the HWSP of) this context | |
307 | * bits 32-51: ctx ID, a globally unique tag (the LRCA again!) | |
308 | * bits 52-63: reserved, may encode the engine ID (for GuC) | |
73e4d07f | 309 | */ |
ca82580c TU |
310 | static void |
311 | intel_lr_context_descriptor_update(struct intel_context *ctx, | |
312 | struct intel_engine_cs *ring) | |
84b790f8 | 313 | { |
ca82580c | 314 | uint64_t lrca, desc; |
84b790f8 | 315 | |
ca82580c TU |
316 | lrca = ctx->engine[ring->id].lrc_vma->node.start + |
317 | LRC_PPHWSP_PN * PAGE_SIZE; | |
84b790f8 | 318 | |
ca82580c TU |
319 | desc = ring->ctx_desc_template; /* bits 0-11 */ |
320 | desc |= lrca; /* bits 12-31 */ | |
321 | desc |= (lrca >> PAGE_SHIFT) << GEN8_CTX_ID_SHIFT; /* bits 32-51 */ | |
5af05fef | 322 | |
ca82580c | 323 | ctx->engine[ring->id].lrc_desc = desc; |
5af05fef MT |
324 | } |
325 | ||
919f1f55 DG |
326 | uint64_t intel_lr_context_descriptor(struct intel_context *ctx, |
327 | struct intel_engine_cs *ring) | |
84b790f8 | 328 | { |
ca82580c TU |
329 | return ctx->engine[ring->id].lrc_desc; |
330 | } | |
203a571b | 331 | |
ca82580c TU |
332 | /** |
333 | * intel_execlists_ctx_id() - get the Execlists Context ID | |
334 | * @ctx: Context to get the ID for | |
335 | * @ring: Engine to get the ID for | |
336 | * | |
337 | * Do not confuse with ctx->id! Unfortunately we have a name overload | |
338 | * here: the old context ID we pass to userspace as a handler so that | |
339 | * they can refer to a context, and the new context ID we pass to the | |
340 | * ELSP so that the GPU can inform us of the context status via | |
341 | * interrupts. | |
342 | * | |
343 | * The context ID is a portion of the context descriptor, so we can | |
344 | * just extract the required part from the cached descriptor. | |
345 | * | |
346 | * Return: 20-bits globally unique context ID. | |
347 | */ | |
348 | u32 intel_execlists_ctx_id(struct intel_context *ctx, | |
349 | struct intel_engine_cs *ring) | |
350 | { | |
351 | return intel_lr_context_descriptor(ctx, ring) >> GEN8_CTX_ID_SHIFT; | |
84b790f8 BW |
352 | } |
353 | ||
cc3c4253 MK |
354 | static void execlists_elsp_write(struct drm_i915_gem_request *rq0, |
355 | struct drm_i915_gem_request *rq1) | |
84b790f8 | 356 | { |
cc3c4253 MK |
357 | |
358 | struct intel_engine_cs *ring = rq0->ring; | |
6e7cc470 TU |
359 | struct drm_device *dev = ring->dev; |
360 | struct drm_i915_private *dev_priv = dev->dev_private; | |
1cff8cc3 | 361 | uint64_t desc[2]; |
84b790f8 | 362 | |
1cff8cc3 | 363 | if (rq1) { |
919f1f55 | 364 | desc[1] = intel_lr_context_descriptor(rq1->ctx, rq1->ring); |
1cff8cc3 MK |
365 | rq1->elsp_submitted++; |
366 | } else { | |
367 | desc[1] = 0; | |
368 | } | |
84b790f8 | 369 | |
919f1f55 | 370 | desc[0] = intel_lr_context_descriptor(rq0->ctx, rq0->ring); |
1cff8cc3 | 371 | rq0->elsp_submitted++; |
84b790f8 | 372 | |
1cff8cc3 | 373 | /* You must always write both descriptors in the order below. */ |
a6111f7b CW |
374 | spin_lock(&dev_priv->uncore.lock); |
375 | intel_uncore_forcewake_get__locked(dev_priv, FORCEWAKE_ALL); | |
1cff8cc3 MK |
376 | I915_WRITE_FW(RING_ELSP(ring), upper_32_bits(desc[1])); |
377 | I915_WRITE_FW(RING_ELSP(ring), lower_32_bits(desc[1])); | |
6daccb0b | 378 | |
1cff8cc3 | 379 | I915_WRITE_FW(RING_ELSP(ring), upper_32_bits(desc[0])); |
84b790f8 | 380 | /* The context is automatically loaded after the following */ |
1cff8cc3 | 381 | I915_WRITE_FW(RING_ELSP(ring), lower_32_bits(desc[0])); |
84b790f8 | 382 | |
1cff8cc3 | 383 | /* ELSP is a wo register, use another nearby reg for posting */ |
83843d84 | 384 | POSTING_READ_FW(RING_EXECLIST_STATUS_LO(ring)); |
a6111f7b CW |
385 | intel_uncore_forcewake_put__locked(dev_priv, FORCEWAKE_ALL); |
386 | spin_unlock(&dev_priv->uncore.lock); | |
84b790f8 BW |
387 | } |
388 | ||
05d9824b | 389 | static int execlists_update_context(struct drm_i915_gem_request *rq) |
ae1250b9 | 390 | { |
05d9824b MK |
391 | struct intel_engine_cs *ring = rq->ring; |
392 | struct i915_hw_ppgtt *ppgtt = rq->ctx->ppgtt; | |
82352e90 | 393 | uint32_t *reg_state = rq->ctx->engine[ring->id].lrc_reg_state; |
ae1250b9 | 394 | |
05d9824b | 395 | reg_state[CTX_RING_TAIL+1] = rq->tail; |
ae1250b9 | 396 | |
2dba3239 MT |
397 | if (ppgtt && !USES_FULL_48BIT_PPGTT(ppgtt->base.dev)) { |
398 | /* True 32b PPGTT with dynamic page allocation: update PDP | |
399 | * registers and point the unallocated PDPs to scratch page. | |
400 | * PML4 is allocated during ppgtt init, so this is not needed | |
401 | * in 48-bit mode. | |
402 | */ | |
d7b2633d MT |
403 | ASSIGN_CTX_PDP(ppgtt, reg_state, 3); |
404 | ASSIGN_CTX_PDP(ppgtt, reg_state, 2); | |
405 | ASSIGN_CTX_PDP(ppgtt, reg_state, 1); | |
406 | ASSIGN_CTX_PDP(ppgtt, reg_state, 0); | |
407 | } | |
408 | ||
ae1250b9 OM |
409 | return 0; |
410 | } | |
411 | ||
d8cb8875 MK |
412 | static void execlists_submit_requests(struct drm_i915_gem_request *rq0, |
413 | struct drm_i915_gem_request *rq1) | |
84b790f8 | 414 | { |
05d9824b | 415 | execlists_update_context(rq0); |
d8cb8875 | 416 | |
cc3c4253 | 417 | if (rq1) |
05d9824b | 418 | execlists_update_context(rq1); |
84b790f8 | 419 | |
cc3c4253 | 420 | execlists_elsp_write(rq0, rq1); |
84b790f8 BW |
421 | } |
422 | ||
acdd884a MT |
423 | static void execlists_context_unqueue(struct intel_engine_cs *ring) |
424 | { | |
6d3d8274 NH |
425 | struct drm_i915_gem_request *req0 = NULL, *req1 = NULL; |
426 | struct drm_i915_gem_request *cursor = NULL, *tmp = NULL; | |
e981e7b1 TD |
427 | |
428 | assert_spin_locked(&ring->execlist_lock); | |
acdd884a | 429 | |
779949f4 PA |
430 | /* |
431 | * If irqs are not active generate a warning as batches that finish | |
432 | * without the irqs may get lost and a GPU Hang may occur. | |
433 | */ | |
434 | WARN_ON(!intel_irqs_enabled(ring->dev->dev_private)); | |
435 | ||
acdd884a MT |
436 | if (list_empty(&ring->execlist_queue)) |
437 | return; | |
438 | ||
439 | /* Try to read in pairs */ | |
440 | list_for_each_entry_safe(cursor, tmp, &ring->execlist_queue, | |
441 | execlist_link) { | |
442 | if (!req0) { | |
443 | req0 = cursor; | |
6d3d8274 | 444 | } else if (req0->ctx == cursor->ctx) { |
acdd884a MT |
445 | /* Same ctx: ignore first request, as second request |
446 | * will update tail past first request's workload */ | |
e1fee72c | 447 | cursor->elsp_submitted = req0->elsp_submitted; |
7eb08a25 TU |
448 | list_move_tail(&req0->execlist_link, |
449 | &ring->execlist_retired_req_list); | |
acdd884a MT |
450 | req0 = cursor; |
451 | } else { | |
452 | req1 = cursor; | |
453 | break; | |
454 | } | |
455 | } | |
456 | ||
53292cdb MT |
457 | if (IS_GEN8(ring->dev) || IS_GEN9(ring->dev)) { |
458 | /* | |
459 | * WaIdleLiteRestore: make sure we never cause a lite | |
460 | * restore with HEAD==TAIL | |
461 | */ | |
d63f820f | 462 | if (req0->elsp_submitted) { |
53292cdb MT |
463 | /* |
464 | * Apply the wa NOOPS to prevent ring:HEAD == req:TAIL | |
465 | * as we resubmit the request. See gen8_emit_request() | |
466 | * for where we prepare the padding after the end of the | |
467 | * request. | |
468 | */ | |
469 | struct intel_ringbuffer *ringbuf; | |
470 | ||
471 | ringbuf = req0->ctx->engine[ring->id].ringbuf; | |
472 | req0->tail += 8; | |
473 | req0->tail &= ringbuf->size - 1; | |
474 | } | |
475 | } | |
476 | ||
e1fee72c OM |
477 | WARN_ON(req1 && req1->elsp_submitted); |
478 | ||
d8cb8875 | 479 | execlists_submit_requests(req0, req1); |
acdd884a MT |
480 | } |
481 | ||
e981e7b1 TD |
482 | static bool execlists_check_remove_request(struct intel_engine_cs *ring, |
483 | u32 request_id) | |
484 | { | |
6d3d8274 | 485 | struct drm_i915_gem_request *head_req; |
e981e7b1 TD |
486 | |
487 | assert_spin_locked(&ring->execlist_lock); | |
488 | ||
489 | head_req = list_first_entry_or_null(&ring->execlist_queue, | |
6d3d8274 | 490 | struct drm_i915_gem_request, |
e981e7b1 TD |
491 | execlist_link); |
492 | ||
493 | if (head_req != NULL) { | |
ca82580c | 494 | if (intel_execlists_ctx_id(head_req->ctx, ring) == request_id) { |
e1fee72c OM |
495 | WARN(head_req->elsp_submitted == 0, |
496 | "Never submitted head request\n"); | |
497 | ||
498 | if (--head_req->elsp_submitted <= 0) { | |
7eb08a25 TU |
499 | list_move_tail(&head_req->execlist_link, |
500 | &ring->execlist_retired_req_list); | |
e1fee72c OM |
501 | return true; |
502 | } | |
e981e7b1 TD |
503 | } |
504 | } | |
505 | ||
506 | return false; | |
507 | } | |
508 | ||
91a41032 BW |
509 | static void get_context_status(struct intel_engine_cs *ring, |
510 | u8 read_pointer, | |
511 | u32 *status, u32 *context_id) | |
512 | { | |
513 | struct drm_i915_private *dev_priv = ring->dev->dev_private; | |
514 | ||
515 | if (WARN_ON(read_pointer >= GEN8_CSB_ENTRIES)) | |
516 | return; | |
517 | ||
518 | *status = I915_READ(RING_CONTEXT_STATUS_BUF_LO(ring, read_pointer)); | |
519 | *context_id = I915_READ(RING_CONTEXT_STATUS_BUF_HI(ring, read_pointer)); | |
520 | } | |
521 | ||
73e4d07f | 522 | /** |
3f7531c3 | 523 | * intel_lrc_irq_handler() - handle Context Switch interrupts |
73e4d07f OM |
524 | * @ring: Engine Command Streamer to handle. |
525 | * | |
526 | * Check the unread Context Status Buffers and manage the submission of new | |
527 | * contexts to the ELSP accordingly. | |
528 | */ | |
3f7531c3 | 529 | void intel_lrc_irq_handler(struct intel_engine_cs *ring) |
e981e7b1 TD |
530 | { |
531 | struct drm_i915_private *dev_priv = ring->dev->dev_private; | |
532 | u32 status_pointer; | |
533 | u8 read_pointer; | |
534 | u8 write_pointer; | |
5af05fef | 535 | u32 status = 0; |
e981e7b1 TD |
536 | u32 status_id; |
537 | u32 submit_contexts = 0; | |
538 | ||
539 | status_pointer = I915_READ(RING_CONTEXT_STATUS_PTR(ring)); | |
540 | ||
541 | read_pointer = ring->next_context_status_buffer; | |
5590a5f0 | 542 | write_pointer = GEN8_CSB_WRITE_PTR(status_pointer); |
e981e7b1 | 543 | if (read_pointer > write_pointer) |
dfc53c5e | 544 | write_pointer += GEN8_CSB_ENTRIES; |
e981e7b1 TD |
545 | |
546 | spin_lock(&ring->execlist_lock); | |
547 | ||
548 | while (read_pointer < write_pointer) { | |
91a41032 BW |
549 | |
550 | get_context_status(ring, ++read_pointer % GEN8_CSB_ENTRIES, | |
551 | &status, &status_id); | |
e981e7b1 | 552 | |
031a8936 MK |
553 | if (status & GEN8_CTX_STATUS_IDLE_ACTIVE) |
554 | continue; | |
555 | ||
e1fee72c OM |
556 | if (status & GEN8_CTX_STATUS_PREEMPTED) { |
557 | if (status & GEN8_CTX_STATUS_LITE_RESTORE) { | |
558 | if (execlists_check_remove_request(ring, status_id)) | |
559 | WARN(1, "Lite Restored request removed from queue\n"); | |
560 | } else | |
561 | WARN(1, "Preemption without Lite Restore\n"); | |
562 | } | |
563 | ||
eba51190 BW |
564 | if ((status & GEN8_CTX_STATUS_ACTIVE_IDLE) || |
565 | (status & GEN8_CTX_STATUS_ELEMENT_SWITCH)) { | |
e981e7b1 TD |
566 | if (execlists_check_remove_request(ring, status_id)) |
567 | submit_contexts++; | |
568 | } | |
569 | } | |
570 | ||
ca82580c | 571 | if (ring->disable_lite_restore_wa) { |
5af05fef MT |
572 | /* Prevent a ctx to preempt itself */ |
573 | if ((status & GEN8_CTX_STATUS_ACTIVE_IDLE) && | |
574 | (submit_contexts != 0)) | |
575 | execlists_context_unqueue(ring); | |
576 | } else if (submit_contexts != 0) { | |
e981e7b1 | 577 | execlists_context_unqueue(ring); |
5af05fef | 578 | } |
e981e7b1 TD |
579 | |
580 | spin_unlock(&ring->execlist_lock); | |
581 | ||
f764a8b1 BW |
582 | if (unlikely(submit_contexts > 2)) |
583 | DRM_ERROR("More than two context complete events?\n"); | |
584 | ||
dfc53c5e | 585 | ring->next_context_status_buffer = write_pointer % GEN8_CSB_ENTRIES; |
e981e7b1 | 586 | |
5590a5f0 BW |
587 | /* Update the read pointer to the old write pointer. Manual ringbuffer |
588 | * management ftw </sarcasm> */ | |
e981e7b1 | 589 | I915_WRITE(RING_CONTEXT_STATUS_PTR(ring), |
5590a5f0 BW |
590 | _MASKED_FIELD(GEN8_CSB_READ_PTR_MASK, |
591 | ring->next_context_status_buffer << 8)); | |
e981e7b1 TD |
592 | } |
593 | ||
ae70797d | 594 | static int execlists_context_queue(struct drm_i915_gem_request *request) |
acdd884a | 595 | { |
ae70797d | 596 | struct intel_engine_cs *ring = request->ring; |
6d3d8274 | 597 | struct drm_i915_gem_request *cursor; |
f1ad5a1f | 598 | int num_elements = 0; |
acdd884a | 599 | |
ed54c1a1 | 600 | if (request->ctx != request->i915->kernel_context) |
af3302b9 DV |
601 | intel_lr_context_pin(request); |
602 | ||
9bb1af44 JH |
603 | i915_gem_request_reference(request); |
604 | ||
b5eba372 | 605 | spin_lock_irq(&ring->execlist_lock); |
acdd884a | 606 | |
f1ad5a1f OM |
607 | list_for_each_entry(cursor, &ring->execlist_queue, execlist_link) |
608 | if (++num_elements > 2) | |
609 | break; | |
610 | ||
611 | if (num_elements > 2) { | |
6d3d8274 | 612 | struct drm_i915_gem_request *tail_req; |
f1ad5a1f OM |
613 | |
614 | tail_req = list_last_entry(&ring->execlist_queue, | |
6d3d8274 | 615 | struct drm_i915_gem_request, |
f1ad5a1f OM |
616 | execlist_link); |
617 | ||
ae70797d | 618 | if (request->ctx == tail_req->ctx) { |
f1ad5a1f | 619 | WARN(tail_req->elsp_submitted != 0, |
7ba717cf | 620 | "More than 2 already-submitted reqs queued\n"); |
7eb08a25 TU |
621 | list_move_tail(&tail_req->execlist_link, |
622 | &ring->execlist_retired_req_list); | |
f1ad5a1f OM |
623 | } |
624 | } | |
625 | ||
6d3d8274 | 626 | list_add_tail(&request->execlist_link, &ring->execlist_queue); |
f1ad5a1f | 627 | if (num_elements == 0) |
acdd884a MT |
628 | execlists_context_unqueue(ring); |
629 | ||
b5eba372 | 630 | spin_unlock_irq(&ring->execlist_lock); |
acdd884a MT |
631 | |
632 | return 0; | |
633 | } | |
634 | ||
2f20055d | 635 | static int logical_ring_invalidate_all_caches(struct drm_i915_gem_request *req) |
ba8b7ccb | 636 | { |
2f20055d | 637 | struct intel_engine_cs *ring = req->ring; |
ba8b7ccb OM |
638 | uint32_t flush_domains; |
639 | int ret; | |
640 | ||
641 | flush_domains = 0; | |
642 | if (ring->gpu_caches_dirty) | |
643 | flush_domains = I915_GEM_GPU_DOMAINS; | |
644 | ||
7deb4d39 | 645 | ret = ring->emit_flush(req, I915_GEM_GPU_DOMAINS, flush_domains); |
ba8b7ccb OM |
646 | if (ret) |
647 | return ret; | |
648 | ||
649 | ring->gpu_caches_dirty = false; | |
650 | return 0; | |
651 | } | |
652 | ||
535fbe82 | 653 | static int execlists_move_to_gpu(struct drm_i915_gem_request *req, |
ba8b7ccb OM |
654 | struct list_head *vmas) |
655 | { | |
535fbe82 | 656 | const unsigned other_rings = ~intel_ring_flag(req->ring); |
ba8b7ccb OM |
657 | struct i915_vma *vma; |
658 | uint32_t flush_domains = 0; | |
659 | bool flush_chipset = false; | |
660 | int ret; | |
661 | ||
662 | list_for_each_entry(vma, vmas, exec_list) { | |
663 | struct drm_i915_gem_object *obj = vma->obj; | |
664 | ||
03ade511 | 665 | if (obj->active & other_rings) { |
91af127f | 666 | ret = i915_gem_object_sync(obj, req->ring, &req); |
03ade511 CW |
667 | if (ret) |
668 | return ret; | |
669 | } | |
ba8b7ccb OM |
670 | |
671 | if (obj->base.write_domain & I915_GEM_DOMAIN_CPU) | |
672 | flush_chipset |= i915_gem_clflush_object(obj, false); | |
673 | ||
674 | flush_domains |= obj->base.write_domain; | |
675 | } | |
676 | ||
677 | if (flush_domains & I915_GEM_DOMAIN_GTT) | |
678 | wmb(); | |
679 | ||
680 | /* Unconditionally invalidate gpu caches and ensure that we do flush | |
681 | * any residual writes from the previous batch. | |
682 | */ | |
2f20055d | 683 | return logical_ring_invalidate_all_caches(req); |
ba8b7ccb OM |
684 | } |
685 | ||
40e895ce | 686 | int intel_logical_ring_alloc_request_extras(struct drm_i915_gem_request *request) |
bc0dce3f | 687 | { |
e28e404c | 688 | int ret = 0; |
bc0dce3f | 689 | |
f3cc01f0 MK |
690 | request->ringbuf = request->ctx->engine[request->ring->id].ringbuf; |
691 | ||
a7e02199 AD |
692 | if (i915.enable_guc_submission) { |
693 | /* | |
694 | * Check that the GuC has space for the request before | |
695 | * going any further, as the i915_add_request() call | |
696 | * later on mustn't fail ... | |
697 | */ | |
698 | struct intel_guc *guc = &request->i915->guc; | |
699 | ||
700 | ret = i915_guc_wq_check_space(guc->execbuf_client); | |
701 | if (ret) | |
702 | return ret; | |
703 | } | |
704 | ||
e28e404c DG |
705 | if (request->ctx != request->i915->kernel_context) |
706 | ret = intel_lr_context_pin(request); | |
707 | ||
708 | return ret; | |
bc0dce3f JH |
709 | } |
710 | ||
ae70797d | 711 | static int logical_ring_wait_for_space(struct drm_i915_gem_request *req, |
595e1eeb | 712 | int bytes) |
bc0dce3f | 713 | { |
ae70797d JH |
714 | struct intel_ringbuffer *ringbuf = req->ringbuf; |
715 | struct intel_engine_cs *ring = req->ring; | |
716 | struct drm_i915_gem_request *target; | |
b4716185 CW |
717 | unsigned space; |
718 | int ret; | |
bc0dce3f JH |
719 | |
720 | if (intel_ring_space(ringbuf) >= bytes) | |
721 | return 0; | |
722 | ||
79bbcc29 JH |
723 | /* The whole point of reserving space is to not wait! */ |
724 | WARN_ON(ringbuf->reserved_in_use); | |
725 | ||
ae70797d | 726 | list_for_each_entry(target, &ring->request_list, list) { |
bc0dce3f JH |
727 | /* |
728 | * The request queue is per-engine, so can contain requests | |
729 | * from multiple ringbuffers. Here, we must ignore any that | |
730 | * aren't from the ringbuffer we're considering. | |
731 | */ | |
ae70797d | 732 | if (target->ringbuf != ringbuf) |
bc0dce3f JH |
733 | continue; |
734 | ||
735 | /* Would completion of this request free enough space? */ | |
ae70797d | 736 | space = __intel_ring_space(target->postfix, ringbuf->tail, |
b4716185 CW |
737 | ringbuf->size); |
738 | if (space >= bytes) | |
bc0dce3f | 739 | break; |
bc0dce3f JH |
740 | } |
741 | ||
ae70797d | 742 | if (WARN_ON(&target->list == &ring->request_list)) |
bc0dce3f JH |
743 | return -ENOSPC; |
744 | ||
ae70797d | 745 | ret = i915_wait_request(target); |
bc0dce3f JH |
746 | if (ret) |
747 | return ret; | |
748 | ||
b4716185 CW |
749 | ringbuf->space = space; |
750 | return 0; | |
bc0dce3f JH |
751 | } |
752 | ||
753 | /* | |
754 | * intel_logical_ring_advance_and_submit() - advance the tail and submit the workload | |
ae70797d | 755 | * @request: Request to advance the logical ringbuffer of. |
bc0dce3f JH |
756 | * |
757 | * The tail is updated in our logical ringbuffer struct, not in the actual context. What | |
758 | * really happens during submission is that the context and current tail will be placed | |
759 | * on a queue waiting for the ELSP to be ready to accept a new context submission. At that | |
760 | * point, the tail *inside* the context is updated and the ELSP written to. | |
761 | */ | |
7c17d377 | 762 | static int |
ae70797d | 763 | intel_logical_ring_advance_and_submit(struct drm_i915_gem_request *request) |
bc0dce3f | 764 | { |
7c17d377 | 765 | struct intel_ringbuffer *ringbuf = request->ringbuf; |
d1675198 | 766 | struct drm_i915_private *dev_priv = request->i915; |
bc0dce3f | 767 | |
7c17d377 CW |
768 | intel_logical_ring_advance(ringbuf); |
769 | request->tail = ringbuf->tail; | |
bc0dce3f | 770 | |
7c17d377 CW |
771 | /* |
772 | * Here we add two extra NOOPs as padding to avoid | |
773 | * lite restore of a context with HEAD==TAIL. | |
774 | * | |
775 | * Caller must reserve WA_TAIL_DWORDS for us! | |
776 | */ | |
777 | intel_logical_ring_emit(ringbuf, MI_NOOP); | |
778 | intel_logical_ring_emit(ringbuf, MI_NOOP); | |
779 | intel_logical_ring_advance(ringbuf); | |
d1675198 | 780 | |
7c17d377 CW |
781 | if (intel_ring_stopped(request->ring)) |
782 | return 0; | |
bc0dce3f | 783 | |
d1675198 AD |
784 | if (dev_priv->guc.execbuf_client) |
785 | i915_guc_submit(dev_priv->guc.execbuf_client, request); | |
786 | else | |
787 | execlists_context_queue(request); | |
7c17d377 CW |
788 | |
789 | return 0; | |
bc0dce3f JH |
790 | } |
791 | ||
79bbcc29 | 792 | static void __wrap_ring_buffer(struct intel_ringbuffer *ringbuf) |
bc0dce3f JH |
793 | { |
794 | uint32_t __iomem *virt; | |
795 | int rem = ringbuf->size - ringbuf->tail; | |
796 | ||
bc0dce3f JH |
797 | virt = ringbuf->virtual_start + ringbuf->tail; |
798 | rem /= 4; | |
799 | while (rem--) | |
800 | iowrite32(MI_NOOP, virt++); | |
801 | ||
802 | ringbuf->tail = 0; | |
803 | intel_ring_update_space(ringbuf); | |
bc0dce3f JH |
804 | } |
805 | ||
ae70797d | 806 | static int logical_ring_prepare(struct drm_i915_gem_request *req, int bytes) |
bc0dce3f | 807 | { |
ae70797d | 808 | struct intel_ringbuffer *ringbuf = req->ringbuf; |
79bbcc29 JH |
809 | int remain_usable = ringbuf->effective_size - ringbuf->tail; |
810 | int remain_actual = ringbuf->size - ringbuf->tail; | |
811 | int ret, total_bytes, wait_bytes = 0; | |
812 | bool need_wrap = false; | |
29b1b415 | 813 | |
79bbcc29 JH |
814 | if (ringbuf->reserved_in_use) |
815 | total_bytes = bytes; | |
816 | else | |
817 | total_bytes = bytes + ringbuf->reserved_size; | |
29b1b415 | 818 | |
79bbcc29 JH |
819 | if (unlikely(bytes > remain_usable)) { |
820 | /* | |
821 | * Not enough space for the basic request. So need to flush | |
822 | * out the remainder and then wait for base + reserved. | |
823 | */ | |
824 | wait_bytes = remain_actual + total_bytes; | |
825 | need_wrap = true; | |
826 | } else { | |
827 | if (unlikely(total_bytes > remain_usable)) { | |
828 | /* | |
829 | * The base request will fit but the reserved space | |
830 | * falls off the end. So only need to to wait for the | |
831 | * reserved size after flushing out the remainder. | |
832 | */ | |
833 | wait_bytes = remain_actual + ringbuf->reserved_size; | |
834 | need_wrap = true; | |
835 | } else if (total_bytes > ringbuf->space) { | |
836 | /* No wrapping required, just waiting. */ | |
837 | wait_bytes = total_bytes; | |
29b1b415 | 838 | } |
bc0dce3f JH |
839 | } |
840 | ||
79bbcc29 JH |
841 | if (wait_bytes) { |
842 | ret = logical_ring_wait_for_space(req, wait_bytes); | |
bc0dce3f JH |
843 | if (unlikely(ret)) |
844 | return ret; | |
79bbcc29 JH |
845 | |
846 | if (need_wrap) | |
847 | __wrap_ring_buffer(ringbuf); | |
bc0dce3f JH |
848 | } |
849 | ||
850 | return 0; | |
851 | } | |
852 | ||
853 | /** | |
854 | * intel_logical_ring_begin() - prepare the logical ringbuffer to accept some commands | |
855 | * | |
374887ba | 856 | * @req: The request to start some new work for |
bc0dce3f JH |
857 | * @num_dwords: number of DWORDs that we plan to write to the ringbuffer. |
858 | * | |
859 | * The ringbuffer might not be ready to accept the commands right away (maybe it needs to | |
860 | * be wrapped, or wait a bit for the tail to be updated). This function takes care of that | |
861 | * and also preallocates a request (every workload submission is still mediated through | |
862 | * requests, same as it did with legacy ringbuffer submission). | |
863 | * | |
864 | * Return: non-zero if the ringbuffer is not ready to be written to. | |
865 | */ | |
3bbaba0c | 866 | int intel_logical_ring_begin(struct drm_i915_gem_request *req, int num_dwords) |
bc0dce3f | 867 | { |
4d616a29 | 868 | struct drm_i915_private *dev_priv; |
bc0dce3f JH |
869 | int ret; |
870 | ||
4d616a29 JH |
871 | WARN_ON(req == NULL); |
872 | dev_priv = req->ring->dev->dev_private; | |
873 | ||
bc0dce3f JH |
874 | ret = i915_gem_check_wedge(&dev_priv->gpu_error, |
875 | dev_priv->mm.interruptible); | |
876 | if (ret) | |
877 | return ret; | |
878 | ||
ae70797d | 879 | ret = logical_ring_prepare(req, num_dwords * sizeof(uint32_t)); |
bc0dce3f JH |
880 | if (ret) |
881 | return ret; | |
882 | ||
4d616a29 | 883 | req->ringbuf->space -= num_dwords * sizeof(uint32_t); |
bc0dce3f JH |
884 | return 0; |
885 | } | |
886 | ||
ccd98fe4 JH |
887 | int intel_logical_ring_reserve_space(struct drm_i915_gem_request *request) |
888 | { | |
889 | /* | |
890 | * The first call merely notes the reserve request and is common for | |
891 | * all back ends. The subsequent localised _begin() call actually | |
892 | * ensures that the reservation is available. Without the begin, if | |
893 | * the request creator immediately submitted the request without | |
894 | * adding any commands to it then there might not actually be | |
895 | * sufficient room for the submission commands. | |
896 | */ | |
897 | intel_ring_reserved_space_reserve(request->ringbuf, MIN_SPACE_FOR_ADD_REQUEST); | |
898 | ||
899 | return intel_logical_ring_begin(request, 0); | |
900 | } | |
901 | ||
73e4d07f OM |
902 | /** |
903 | * execlists_submission() - submit a batchbuffer for execution, Execlists style | |
904 | * @dev: DRM device. | |
905 | * @file: DRM file. | |
906 | * @ring: Engine Command Streamer to submit to. | |
907 | * @ctx: Context to employ for this submission. | |
908 | * @args: execbuffer call arguments. | |
909 | * @vmas: list of vmas. | |
910 | * @batch_obj: the batchbuffer to submit. | |
911 | * @exec_start: batchbuffer start virtual address pointer. | |
8e004efc | 912 | * @dispatch_flags: translated execbuffer call flags. |
73e4d07f OM |
913 | * |
914 | * This is the evil twin version of i915_gem_ringbuffer_submission. It abstracts | |
915 | * away the submission details of the execbuffer ioctl call. | |
916 | * | |
917 | * Return: non-zero if the submission fails. | |
918 | */ | |
5f19e2bf | 919 | int intel_execlists_submission(struct i915_execbuffer_params *params, |
454afebd | 920 | struct drm_i915_gem_execbuffer2 *args, |
5f19e2bf | 921 | struct list_head *vmas) |
454afebd | 922 | { |
5f19e2bf JH |
923 | struct drm_device *dev = params->dev; |
924 | struct intel_engine_cs *ring = params->ring; | |
ba8b7ccb | 925 | struct drm_i915_private *dev_priv = dev->dev_private; |
5f19e2bf JH |
926 | struct intel_ringbuffer *ringbuf = params->ctx->engine[ring->id].ringbuf; |
927 | u64 exec_start; | |
ba8b7ccb OM |
928 | int instp_mode; |
929 | u32 instp_mask; | |
930 | int ret; | |
931 | ||
932 | instp_mode = args->flags & I915_EXEC_CONSTANTS_MASK; | |
933 | instp_mask = I915_EXEC_CONSTANTS_MASK; | |
934 | switch (instp_mode) { | |
935 | case I915_EXEC_CONSTANTS_REL_GENERAL: | |
936 | case I915_EXEC_CONSTANTS_ABSOLUTE: | |
937 | case I915_EXEC_CONSTANTS_REL_SURFACE: | |
938 | if (instp_mode != 0 && ring != &dev_priv->ring[RCS]) { | |
939 | DRM_DEBUG("non-0 rel constants mode on non-RCS\n"); | |
940 | return -EINVAL; | |
941 | } | |
942 | ||
943 | if (instp_mode != dev_priv->relative_constants_mode) { | |
944 | if (instp_mode == I915_EXEC_CONSTANTS_REL_SURFACE) { | |
945 | DRM_DEBUG("rel surface constants mode invalid on gen5+\n"); | |
946 | return -EINVAL; | |
947 | } | |
948 | ||
949 | /* The HW changed the meaning on this bit on gen6 */ | |
950 | instp_mask &= ~I915_EXEC_CONSTANTS_REL_SURFACE; | |
951 | } | |
952 | break; | |
953 | default: | |
954 | DRM_DEBUG("execbuf with unknown constants: %d\n", instp_mode); | |
955 | return -EINVAL; | |
956 | } | |
957 | ||
ba8b7ccb OM |
958 | if (args->flags & I915_EXEC_GEN7_SOL_RESET) { |
959 | DRM_DEBUG("sol reset is gen7 only\n"); | |
960 | return -EINVAL; | |
961 | } | |
962 | ||
535fbe82 | 963 | ret = execlists_move_to_gpu(params->request, vmas); |
ba8b7ccb OM |
964 | if (ret) |
965 | return ret; | |
966 | ||
967 | if (ring == &dev_priv->ring[RCS] && | |
968 | instp_mode != dev_priv->relative_constants_mode) { | |
4d616a29 | 969 | ret = intel_logical_ring_begin(params->request, 4); |
ba8b7ccb OM |
970 | if (ret) |
971 | return ret; | |
972 | ||
973 | intel_logical_ring_emit(ringbuf, MI_NOOP); | |
974 | intel_logical_ring_emit(ringbuf, MI_LOAD_REGISTER_IMM(1)); | |
f92a9162 | 975 | intel_logical_ring_emit_reg(ringbuf, INSTPM); |
ba8b7ccb OM |
976 | intel_logical_ring_emit(ringbuf, instp_mask << 16 | instp_mode); |
977 | intel_logical_ring_advance(ringbuf); | |
978 | ||
979 | dev_priv->relative_constants_mode = instp_mode; | |
980 | } | |
981 | ||
5f19e2bf JH |
982 | exec_start = params->batch_obj_vm_offset + |
983 | args->batch_start_offset; | |
984 | ||
be795fc1 | 985 | ret = ring->emit_bb_start(params->request, exec_start, params->dispatch_flags); |
ba8b7ccb OM |
986 | if (ret) |
987 | return ret; | |
988 | ||
95c24161 | 989 | trace_i915_gem_ring_dispatch(params->request, params->dispatch_flags); |
5e4be7bd | 990 | |
8a8edb59 | 991 | i915_gem_execbuffer_move_to_active(vmas, params->request); |
adeca76d | 992 | i915_gem_execbuffer_retire_commands(params); |
ba8b7ccb | 993 | |
454afebd OM |
994 | return 0; |
995 | } | |
996 | ||
c86ee3a9 TD |
997 | void intel_execlists_retire_requests(struct intel_engine_cs *ring) |
998 | { | |
6d3d8274 | 999 | struct drm_i915_gem_request *req, *tmp; |
c86ee3a9 TD |
1000 | struct list_head retired_list; |
1001 | ||
1002 | WARN_ON(!mutex_is_locked(&ring->dev->struct_mutex)); | |
1003 | if (list_empty(&ring->execlist_retired_req_list)) | |
1004 | return; | |
1005 | ||
1006 | INIT_LIST_HEAD(&retired_list); | |
b5eba372 | 1007 | spin_lock_irq(&ring->execlist_lock); |
c86ee3a9 | 1008 | list_replace_init(&ring->execlist_retired_req_list, &retired_list); |
b5eba372 | 1009 | spin_unlock_irq(&ring->execlist_lock); |
c86ee3a9 TD |
1010 | |
1011 | list_for_each_entry_safe(req, tmp, &retired_list, execlist_link) { | |
af3302b9 DV |
1012 | struct intel_context *ctx = req->ctx; |
1013 | struct drm_i915_gem_object *ctx_obj = | |
1014 | ctx->engine[ring->id].state; | |
1015 | ||
ed54c1a1 | 1016 | if (ctx_obj && (ctx != req->i915->kernel_context)) |
af3302b9 | 1017 | intel_lr_context_unpin(req); |
c86ee3a9 | 1018 | list_del(&req->execlist_link); |
f8210795 | 1019 | i915_gem_request_unreference(req); |
c86ee3a9 TD |
1020 | } |
1021 | } | |
1022 | ||
454afebd OM |
1023 | void intel_logical_ring_stop(struct intel_engine_cs *ring) |
1024 | { | |
9832b9da OM |
1025 | struct drm_i915_private *dev_priv = ring->dev->dev_private; |
1026 | int ret; | |
1027 | ||
1028 | if (!intel_ring_initialized(ring)) | |
1029 | return; | |
1030 | ||
1031 | ret = intel_ring_idle(ring); | |
1032 | if (ret && !i915_reset_in_progress(&to_i915(ring->dev)->gpu_error)) | |
1033 | DRM_ERROR("failed to quiesce %s whilst cleaning up: %d\n", | |
1034 | ring->name, ret); | |
1035 | ||
1036 | /* TODO: Is this correct with Execlists enabled? */ | |
1037 | I915_WRITE_MODE(ring, _MASKED_BIT_ENABLE(STOP_RING)); | |
1038 | if (wait_for_atomic((I915_READ_MODE(ring) & MODE_IDLE) != 0, 1000)) { | |
1039 | DRM_ERROR("%s :timed out trying to stop ring\n", ring->name); | |
1040 | return; | |
1041 | } | |
1042 | I915_WRITE_MODE(ring, _MASKED_BIT_DISABLE(STOP_RING)); | |
454afebd OM |
1043 | } |
1044 | ||
4866d729 | 1045 | int logical_ring_flush_all_caches(struct drm_i915_gem_request *req) |
48e29f55 | 1046 | { |
4866d729 | 1047 | struct intel_engine_cs *ring = req->ring; |
48e29f55 OM |
1048 | int ret; |
1049 | ||
1050 | if (!ring->gpu_caches_dirty) | |
1051 | return 0; | |
1052 | ||
7deb4d39 | 1053 | ret = ring->emit_flush(req, 0, I915_GEM_GPU_DOMAINS); |
48e29f55 OM |
1054 | if (ret) |
1055 | return ret; | |
1056 | ||
1057 | ring->gpu_caches_dirty = false; | |
1058 | return 0; | |
1059 | } | |
1060 | ||
e84fe803 | 1061 | static int intel_lr_context_do_pin(struct intel_engine_cs *ring, |
ca82580c | 1062 | struct intel_context *ctx) |
dcb4c12a | 1063 | { |
e84fe803 NH |
1064 | struct drm_device *dev = ring->dev; |
1065 | struct drm_i915_private *dev_priv = dev->dev_private; | |
ca82580c TU |
1066 | struct drm_i915_gem_object *ctx_obj = ctx->engine[ring->id].state; |
1067 | struct intel_ringbuffer *ringbuf = ctx->engine[ring->id].ringbuf; | |
82352e90 | 1068 | struct page *lrc_state_page; |
77b04a04 | 1069 | uint32_t *lrc_reg_state; |
ca82580c | 1070 | int ret; |
dcb4c12a OM |
1071 | |
1072 | WARN_ON(!mutex_is_locked(&ring->dev->struct_mutex)); | |
ca82580c | 1073 | |
e84fe803 NH |
1074 | ret = i915_gem_obj_ggtt_pin(ctx_obj, GEN8_LR_CONTEXT_ALIGN, |
1075 | PIN_OFFSET_BIAS | GUC_WOPCM_TOP); | |
1076 | if (ret) | |
1077 | return ret; | |
7ba717cf | 1078 | |
82352e90 TU |
1079 | lrc_state_page = i915_gem_object_get_dirty_page(ctx_obj, LRC_STATE_PN); |
1080 | if (WARN_ON(!lrc_state_page)) { | |
1081 | ret = -ENODEV; | |
1082 | goto unpin_ctx_obj; | |
1083 | } | |
1084 | ||
e84fe803 NH |
1085 | ret = intel_pin_and_map_ringbuffer_obj(ring->dev, ringbuf); |
1086 | if (ret) | |
1087 | goto unpin_ctx_obj; | |
d1675198 | 1088 | |
ca82580c TU |
1089 | ctx->engine[ring->id].lrc_vma = i915_gem_obj_to_ggtt(ctx_obj); |
1090 | intel_lr_context_descriptor_update(ctx, ring); | |
77b04a04 TU |
1091 | lrc_reg_state = kmap(lrc_state_page); |
1092 | lrc_reg_state[CTX_RING_BUFFER_START+1] = ringbuf->vma->node.start; | |
1093 | ctx->engine[ring->id].lrc_reg_state = lrc_reg_state; | |
e84fe803 | 1094 | ctx_obj->dirty = true; |
e93c28f3 | 1095 | |
e84fe803 NH |
1096 | /* Invalidate GuC TLB. */ |
1097 | if (i915.enable_guc_submission) | |
1098 | I915_WRITE(GEN8_GTCR, GEN8_GTCR_INVALIDATE); | |
dcb4c12a | 1099 | |
7ba717cf TD |
1100 | return ret; |
1101 | ||
1102 | unpin_ctx_obj: | |
1103 | i915_gem_object_ggtt_unpin(ctx_obj); | |
e84fe803 NH |
1104 | |
1105 | return ret; | |
1106 | } | |
1107 | ||
1108 | static int intel_lr_context_pin(struct drm_i915_gem_request *rq) | |
1109 | { | |
1110 | int ret = 0; | |
1111 | struct intel_engine_cs *ring = rq->ring; | |
e84fe803 NH |
1112 | |
1113 | if (rq->ctx->engine[ring->id].pin_count++ == 0) { | |
ca82580c | 1114 | ret = intel_lr_context_do_pin(ring, rq->ctx); |
e84fe803 NH |
1115 | if (ret) |
1116 | goto reset_pin_count; | |
1117 | } | |
1118 | return ret; | |
1119 | ||
a7cbedec | 1120 | reset_pin_count: |
8ba319da | 1121 | rq->ctx->engine[ring->id].pin_count = 0; |
dcb4c12a OM |
1122 | return ret; |
1123 | } | |
1124 | ||
af3302b9 | 1125 | void intel_lr_context_unpin(struct drm_i915_gem_request *rq) |
dcb4c12a | 1126 | { |
af3302b9 DV |
1127 | struct intel_engine_cs *ring = rq->ring; |
1128 | struct drm_i915_gem_object *ctx_obj = rq->ctx->engine[ring->id].state; | |
1129 | struct intel_ringbuffer *ringbuf = rq->ringbuf; | |
1130 | ||
82352e90 TU |
1131 | WARN_ON(!mutex_is_locked(&ring->dev->struct_mutex)); |
1132 | ||
1133 | if (!ctx_obj) | |
1134 | return; | |
1135 | ||
1136 | if (--rq->ctx->engine[ring->id].pin_count == 0) { | |
1137 | kunmap(kmap_to_page(rq->ctx->engine[ring->id].lrc_reg_state)); | |
1138 | intel_unpin_ringbuffer_obj(ringbuf); | |
1139 | i915_gem_object_ggtt_unpin(ctx_obj); | |
1140 | rq->ctx->engine[ring->id].lrc_vma = NULL; | |
1141 | rq->ctx->engine[ring->id].lrc_desc = 0; | |
1142 | rq->ctx->engine[ring->id].lrc_reg_state = NULL; | |
dcb4c12a OM |
1143 | } |
1144 | } | |
1145 | ||
e2be4faf | 1146 | static int intel_logical_ring_workarounds_emit(struct drm_i915_gem_request *req) |
771b9a53 MT |
1147 | { |
1148 | int ret, i; | |
e2be4faf JH |
1149 | struct intel_engine_cs *ring = req->ring; |
1150 | struct intel_ringbuffer *ringbuf = req->ringbuf; | |
771b9a53 MT |
1151 | struct drm_device *dev = ring->dev; |
1152 | struct drm_i915_private *dev_priv = dev->dev_private; | |
1153 | struct i915_workarounds *w = &dev_priv->workarounds; | |
1154 | ||
cd7feaaa | 1155 | if (w->count == 0) |
771b9a53 MT |
1156 | return 0; |
1157 | ||
1158 | ring->gpu_caches_dirty = true; | |
4866d729 | 1159 | ret = logical_ring_flush_all_caches(req); |
771b9a53 MT |
1160 | if (ret) |
1161 | return ret; | |
1162 | ||
4d616a29 | 1163 | ret = intel_logical_ring_begin(req, w->count * 2 + 2); |
771b9a53 MT |
1164 | if (ret) |
1165 | return ret; | |
1166 | ||
1167 | intel_logical_ring_emit(ringbuf, MI_LOAD_REGISTER_IMM(w->count)); | |
1168 | for (i = 0; i < w->count; i++) { | |
f92a9162 | 1169 | intel_logical_ring_emit_reg(ringbuf, w->reg[i].addr); |
771b9a53 MT |
1170 | intel_logical_ring_emit(ringbuf, w->reg[i].value); |
1171 | } | |
1172 | intel_logical_ring_emit(ringbuf, MI_NOOP); | |
1173 | ||
1174 | intel_logical_ring_advance(ringbuf); | |
1175 | ||
1176 | ring->gpu_caches_dirty = true; | |
4866d729 | 1177 | ret = logical_ring_flush_all_caches(req); |
771b9a53 MT |
1178 | if (ret) |
1179 | return ret; | |
1180 | ||
1181 | return 0; | |
1182 | } | |
1183 | ||
83b8a982 | 1184 | #define wa_ctx_emit(batch, index, cmd) \ |
17ee950d | 1185 | do { \ |
83b8a982 AS |
1186 | int __index = (index)++; \ |
1187 | if (WARN_ON(__index >= (PAGE_SIZE / sizeof(uint32_t)))) { \ | |
17ee950d AS |
1188 | return -ENOSPC; \ |
1189 | } \ | |
83b8a982 | 1190 | batch[__index] = (cmd); \ |
17ee950d AS |
1191 | } while (0) |
1192 | ||
8f40db77 | 1193 | #define wa_ctx_emit_reg(batch, index, reg) \ |
f0f59a00 | 1194 | wa_ctx_emit((batch), (index), i915_mmio_reg_offset(reg)) |
9e000847 AS |
1195 | |
1196 | /* | |
1197 | * In this WA we need to set GEN8_L3SQCREG4[21:21] and reset it after | |
1198 | * PIPE_CONTROL instruction. This is required for the flush to happen correctly | |
1199 | * but there is a slight complication as this is applied in WA batch where the | |
1200 | * values are only initialized once so we cannot take register value at the | |
1201 | * beginning and reuse it further; hence we save its value to memory, upload a | |
1202 | * constant value with bit21 set and then we restore it back with the saved value. | |
1203 | * To simplify the WA, a constant value is formed by using the default value | |
1204 | * of this register. This shouldn't be a problem because we are only modifying | |
1205 | * it for a short period and this batch in non-premptible. We can ofcourse | |
1206 | * use additional instructions that read the actual value of the register | |
1207 | * at that time and set our bit of interest but it makes the WA complicated. | |
1208 | * | |
1209 | * This WA is also required for Gen9 so extracting as a function avoids | |
1210 | * code duplication. | |
1211 | */ | |
1212 | static inline int gen8_emit_flush_coherentl3_wa(struct intel_engine_cs *ring, | |
1213 | uint32_t *const batch, | |
1214 | uint32_t index) | |
1215 | { | |
1216 | uint32_t l3sqc4_flush = (0x40400000 | GEN8_LQSC_FLUSH_COHERENT_LINES); | |
1217 | ||
a4106a78 AS |
1218 | /* |
1219 | * WaDisableLSQCROPERFforOCL:skl | |
1220 | * This WA is implemented in skl_init_clock_gating() but since | |
1221 | * this batch updates GEN8_L3SQCREG4 with default value we need to | |
1222 | * set this bit here to retain the WA during flush. | |
1223 | */ | |
e87a005d | 1224 | if (IS_SKL_REVID(ring->dev, 0, SKL_REVID_E0)) |
a4106a78 AS |
1225 | l3sqc4_flush |= GEN8_LQSC_RO_PERF_DIS; |
1226 | ||
f1afe24f | 1227 | wa_ctx_emit(batch, index, (MI_STORE_REGISTER_MEM_GEN8 | |
83b8a982 | 1228 | MI_SRM_LRM_GLOBAL_GTT)); |
8f40db77 | 1229 | wa_ctx_emit_reg(batch, index, GEN8_L3SQCREG4); |
83b8a982 AS |
1230 | wa_ctx_emit(batch, index, ring->scratch.gtt_offset + 256); |
1231 | wa_ctx_emit(batch, index, 0); | |
1232 | ||
1233 | wa_ctx_emit(batch, index, MI_LOAD_REGISTER_IMM(1)); | |
8f40db77 | 1234 | wa_ctx_emit_reg(batch, index, GEN8_L3SQCREG4); |
83b8a982 AS |
1235 | wa_ctx_emit(batch, index, l3sqc4_flush); |
1236 | ||
1237 | wa_ctx_emit(batch, index, GFX_OP_PIPE_CONTROL(6)); | |
1238 | wa_ctx_emit(batch, index, (PIPE_CONTROL_CS_STALL | | |
1239 | PIPE_CONTROL_DC_FLUSH_ENABLE)); | |
1240 | wa_ctx_emit(batch, index, 0); | |
1241 | wa_ctx_emit(batch, index, 0); | |
1242 | wa_ctx_emit(batch, index, 0); | |
1243 | wa_ctx_emit(batch, index, 0); | |
1244 | ||
f1afe24f | 1245 | wa_ctx_emit(batch, index, (MI_LOAD_REGISTER_MEM_GEN8 | |
83b8a982 | 1246 | MI_SRM_LRM_GLOBAL_GTT)); |
8f40db77 | 1247 | wa_ctx_emit_reg(batch, index, GEN8_L3SQCREG4); |
83b8a982 AS |
1248 | wa_ctx_emit(batch, index, ring->scratch.gtt_offset + 256); |
1249 | wa_ctx_emit(batch, index, 0); | |
9e000847 AS |
1250 | |
1251 | return index; | |
1252 | } | |
1253 | ||
17ee950d AS |
1254 | static inline uint32_t wa_ctx_start(struct i915_wa_ctx_bb *wa_ctx, |
1255 | uint32_t offset, | |
1256 | uint32_t start_alignment) | |
1257 | { | |
1258 | return wa_ctx->offset = ALIGN(offset, start_alignment); | |
1259 | } | |
1260 | ||
1261 | static inline int wa_ctx_end(struct i915_wa_ctx_bb *wa_ctx, | |
1262 | uint32_t offset, | |
1263 | uint32_t size_alignment) | |
1264 | { | |
1265 | wa_ctx->size = offset - wa_ctx->offset; | |
1266 | ||
1267 | WARN(wa_ctx->size % size_alignment, | |
1268 | "wa_ctx_bb failed sanity checks: size %d is not aligned to %d\n", | |
1269 | wa_ctx->size, size_alignment); | |
1270 | return 0; | |
1271 | } | |
1272 | ||
1273 | /** | |
1274 | * gen8_init_indirectctx_bb() - initialize indirect ctx batch with WA | |
1275 | * | |
1276 | * @ring: only applicable for RCS | |
1277 | * @wa_ctx: structure representing wa_ctx | |
1278 | * offset: specifies start of the batch, should be cache-aligned. This is updated | |
1279 | * with the offset value received as input. | |
1280 | * size: size of the batch in DWORDS but HW expects in terms of cachelines | |
1281 | * @batch: page in which WA are loaded | |
1282 | * @offset: This field specifies the start of the batch, it should be | |
1283 | * cache-aligned otherwise it is adjusted accordingly. | |
1284 | * Typically we only have one indirect_ctx and per_ctx batch buffer which are | |
1285 | * initialized at the beginning and shared across all contexts but this field | |
1286 | * helps us to have multiple batches at different offsets and select them based | |
1287 | * on a criteria. At the moment this batch always start at the beginning of the page | |
1288 | * and at this point we don't have multiple wa_ctx batch buffers. | |
1289 | * | |
1290 | * The number of WA applied are not known at the beginning; we use this field | |
1291 | * to return the no of DWORDS written. | |
4d78c8dc | 1292 | * |
17ee950d AS |
1293 | * It is to be noted that this batch does not contain MI_BATCH_BUFFER_END |
1294 | * so it adds NOOPs as padding to make it cacheline aligned. | |
1295 | * MI_BATCH_BUFFER_END will be added to perctx batch and both of them together | |
1296 | * makes a complete batch buffer. | |
1297 | * | |
1298 | * Return: non-zero if we exceed the PAGE_SIZE limit. | |
1299 | */ | |
1300 | ||
1301 | static int gen8_init_indirectctx_bb(struct intel_engine_cs *ring, | |
1302 | struct i915_wa_ctx_bb *wa_ctx, | |
1303 | uint32_t *const batch, | |
1304 | uint32_t *offset) | |
1305 | { | |
0160f055 | 1306 | uint32_t scratch_addr; |
17ee950d AS |
1307 | uint32_t index = wa_ctx_start(wa_ctx, *offset, CACHELINE_DWORDS); |
1308 | ||
7ad00d1a | 1309 | /* WaDisableCtxRestoreArbitration:bdw,chv */ |
83b8a982 | 1310 | wa_ctx_emit(batch, index, MI_ARB_ON_OFF | MI_ARB_DISABLE); |
17ee950d | 1311 | |
c82435bb AS |
1312 | /* WaFlushCoherentL3CacheLinesAtContextSwitch:bdw */ |
1313 | if (IS_BROADWELL(ring->dev)) { | |
604ef734 AH |
1314 | int rc = gen8_emit_flush_coherentl3_wa(ring, batch, index); |
1315 | if (rc < 0) | |
1316 | return rc; | |
1317 | index = rc; | |
c82435bb AS |
1318 | } |
1319 | ||
0160f055 AS |
1320 | /* WaClearSlmSpaceAtContextSwitch:bdw,chv */ |
1321 | /* Actual scratch location is at 128 bytes offset */ | |
1322 | scratch_addr = ring->scratch.gtt_offset + 2*CACHELINE_BYTES; | |
1323 | ||
83b8a982 AS |
1324 | wa_ctx_emit(batch, index, GFX_OP_PIPE_CONTROL(6)); |
1325 | wa_ctx_emit(batch, index, (PIPE_CONTROL_FLUSH_L3 | | |
1326 | PIPE_CONTROL_GLOBAL_GTT_IVB | | |
1327 | PIPE_CONTROL_CS_STALL | | |
1328 | PIPE_CONTROL_QW_WRITE)); | |
1329 | wa_ctx_emit(batch, index, scratch_addr); | |
1330 | wa_ctx_emit(batch, index, 0); | |
1331 | wa_ctx_emit(batch, index, 0); | |
1332 | wa_ctx_emit(batch, index, 0); | |
0160f055 | 1333 | |
17ee950d AS |
1334 | /* Pad to end of cacheline */ |
1335 | while (index % CACHELINE_DWORDS) | |
83b8a982 | 1336 | wa_ctx_emit(batch, index, MI_NOOP); |
17ee950d AS |
1337 | |
1338 | /* | |
1339 | * MI_BATCH_BUFFER_END is not required in Indirect ctx BB because | |
1340 | * execution depends on the length specified in terms of cache lines | |
1341 | * in the register CTX_RCS_INDIRECT_CTX | |
1342 | */ | |
1343 | ||
1344 | return wa_ctx_end(wa_ctx, *offset = index, CACHELINE_DWORDS); | |
1345 | } | |
1346 | ||
1347 | /** | |
1348 | * gen8_init_perctx_bb() - initialize per ctx batch with WA | |
1349 | * | |
1350 | * @ring: only applicable for RCS | |
1351 | * @wa_ctx: structure representing wa_ctx | |
1352 | * offset: specifies start of the batch, should be cache-aligned. | |
1353 | * size: size of the batch in DWORDS but HW expects in terms of cachelines | |
4d78c8dc | 1354 | * @batch: page in which WA are loaded |
17ee950d AS |
1355 | * @offset: This field specifies the start of this batch. |
1356 | * This batch is started immediately after indirect_ctx batch. Since we ensure | |
1357 | * that indirect_ctx ends on a cacheline this batch is aligned automatically. | |
1358 | * | |
1359 | * The number of DWORDS written are returned using this field. | |
1360 | * | |
1361 | * This batch is terminated with MI_BATCH_BUFFER_END and so we need not add padding | |
1362 | * to align it with cacheline as padding after MI_BATCH_BUFFER_END is redundant. | |
1363 | */ | |
1364 | static int gen8_init_perctx_bb(struct intel_engine_cs *ring, | |
1365 | struct i915_wa_ctx_bb *wa_ctx, | |
1366 | uint32_t *const batch, | |
1367 | uint32_t *offset) | |
1368 | { | |
1369 | uint32_t index = wa_ctx_start(wa_ctx, *offset, CACHELINE_DWORDS); | |
1370 | ||
7ad00d1a | 1371 | /* WaDisableCtxRestoreArbitration:bdw,chv */ |
83b8a982 | 1372 | wa_ctx_emit(batch, index, MI_ARB_ON_OFF | MI_ARB_ENABLE); |
7ad00d1a | 1373 | |
83b8a982 | 1374 | wa_ctx_emit(batch, index, MI_BATCH_BUFFER_END); |
17ee950d AS |
1375 | |
1376 | return wa_ctx_end(wa_ctx, *offset = index, 1); | |
1377 | } | |
1378 | ||
0504cffc AS |
1379 | static int gen9_init_indirectctx_bb(struct intel_engine_cs *ring, |
1380 | struct i915_wa_ctx_bb *wa_ctx, | |
1381 | uint32_t *const batch, | |
1382 | uint32_t *offset) | |
1383 | { | |
a4106a78 | 1384 | int ret; |
0907c8f7 | 1385 | struct drm_device *dev = ring->dev; |
0504cffc AS |
1386 | uint32_t index = wa_ctx_start(wa_ctx, *offset, CACHELINE_DWORDS); |
1387 | ||
0907c8f7 | 1388 | /* WaDisableCtxRestoreArbitration:skl,bxt */ |
e87a005d | 1389 | if (IS_SKL_REVID(dev, 0, SKL_REVID_D0) || |
cbdc12a9 | 1390 | IS_BXT_REVID(dev, 0, BXT_REVID_A1)) |
0907c8f7 | 1391 | wa_ctx_emit(batch, index, MI_ARB_ON_OFF | MI_ARB_DISABLE); |
0504cffc | 1392 | |
a4106a78 AS |
1393 | /* WaFlushCoherentL3CacheLinesAtContextSwitch:skl,bxt */ |
1394 | ret = gen8_emit_flush_coherentl3_wa(ring, batch, index); | |
1395 | if (ret < 0) | |
1396 | return ret; | |
1397 | index = ret; | |
1398 | ||
0504cffc AS |
1399 | /* Pad to end of cacheline */ |
1400 | while (index % CACHELINE_DWORDS) | |
1401 | wa_ctx_emit(batch, index, MI_NOOP); | |
1402 | ||
1403 | return wa_ctx_end(wa_ctx, *offset = index, CACHELINE_DWORDS); | |
1404 | } | |
1405 | ||
1406 | static int gen9_init_perctx_bb(struct intel_engine_cs *ring, | |
1407 | struct i915_wa_ctx_bb *wa_ctx, | |
1408 | uint32_t *const batch, | |
1409 | uint32_t *offset) | |
1410 | { | |
0907c8f7 | 1411 | struct drm_device *dev = ring->dev; |
0504cffc AS |
1412 | uint32_t index = wa_ctx_start(wa_ctx, *offset, CACHELINE_DWORDS); |
1413 | ||
9b01435d | 1414 | /* WaSetDisablePixMaskCammingAndRhwoInCommonSliceChicken:skl,bxt */ |
e87a005d | 1415 | if (IS_SKL_REVID(dev, 0, SKL_REVID_B0) || |
cbdc12a9 | 1416 | IS_BXT_REVID(dev, 0, BXT_REVID_A1)) { |
9b01435d | 1417 | wa_ctx_emit(batch, index, MI_LOAD_REGISTER_IMM(1)); |
8f40db77 | 1418 | wa_ctx_emit_reg(batch, index, GEN9_SLICE_COMMON_ECO_CHICKEN0); |
9b01435d AS |
1419 | wa_ctx_emit(batch, index, |
1420 | _MASKED_BIT_ENABLE(DISABLE_PIXEL_MASK_CAMMING)); | |
1421 | wa_ctx_emit(batch, index, MI_NOOP); | |
1422 | } | |
1423 | ||
0907c8f7 | 1424 | /* WaDisableCtxRestoreArbitration:skl,bxt */ |
e87a005d | 1425 | if (IS_SKL_REVID(dev, 0, SKL_REVID_D0) || |
cbdc12a9 | 1426 | IS_BXT_REVID(dev, 0, BXT_REVID_A1)) |
0907c8f7 AS |
1427 | wa_ctx_emit(batch, index, MI_ARB_ON_OFF | MI_ARB_ENABLE); |
1428 | ||
0504cffc AS |
1429 | wa_ctx_emit(batch, index, MI_BATCH_BUFFER_END); |
1430 | ||
1431 | return wa_ctx_end(wa_ctx, *offset = index, 1); | |
1432 | } | |
1433 | ||
17ee950d AS |
1434 | static int lrc_setup_wa_ctx_obj(struct intel_engine_cs *ring, u32 size) |
1435 | { | |
1436 | int ret; | |
1437 | ||
1438 | ring->wa_ctx.obj = i915_gem_alloc_object(ring->dev, PAGE_ALIGN(size)); | |
1439 | if (!ring->wa_ctx.obj) { | |
1440 | DRM_DEBUG_DRIVER("alloc LRC WA ctx backing obj failed.\n"); | |
1441 | return -ENOMEM; | |
1442 | } | |
1443 | ||
1444 | ret = i915_gem_obj_ggtt_pin(ring->wa_ctx.obj, PAGE_SIZE, 0); | |
1445 | if (ret) { | |
1446 | DRM_DEBUG_DRIVER("pin LRC WA ctx backing obj failed: %d\n", | |
1447 | ret); | |
1448 | drm_gem_object_unreference(&ring->wa_ctx.obj->base); | |
1449 | return ret; | |
1450 | } | |
1451 | ||
1452 | return 0; | |
1453 | } | |
1454 | ||
1455 | static void lrc_destroy_wa_ctx_obj(struct intel_engine_cs *ring) | |
1456 | { | |
1457 | if (ring->wa_ctx.obj) { | |
1458 | i915_gem_object_ggtt_unpin(ring->wa_ctx.obj); | |
1459 | drm_gem_object_unreference(&ring->wa_ctx.obj->base); | |
1460 | ring->wa_ctx.obj = NULL; | |
1461 | } | |
1462 | } | |
1463 | ||
1464 | static int intel_init_workaround_bb(struct intel_engine_cs *ring) | |
1465 | { | |
1466 | int ret; | |
1467 | uint32_t *batch; | |
1468 | uint32_t offset; | |
1469 | struct page *page; | |
1470 | struct i915_ctx_workarounds *wa_ctx = &ring->wa_ctx; | |
1471 | ||
1472 | WARN_ON(ring->id != RCS); | |
1473 | ||
5e60d790 | 1474 | /* update this when WA for higher Gen are added */ |
0504cffc AS |
1475 | if (INTEL_INFO(ring->dev)->gen > 9) { |
1476 | DRM_ERROR("WA batch buffer is not initialized for Gen%d\n", | |
1477 | INTEL_INFO(ring->dev)->gen); | |
5e60d790 | 1478 | return 0; |
0504cffc | 1479 | } |
5e60d790 | 1480 | |
c4db7599 AS |
1481 | /* some WA perform writes to scratch page, ensure it is valid */ |
1482 | if (ring->scratch.obj == NULL) { | |
1483 | DRM_ERROR("scratch page not allocated for %s\n", ring->name); | |
1484 | return -EINVAL; | |
1485 | } | |
1486 | ||
17ee950d AS |
1487 | ret = lrc_setup_wa_ctx_obj(ring, PAGE_SIZE); |
1488 | if (ret) { | |
1489 | DRM_DEBUG_DRIVER("Failed to setup context WA page: %d\n", ret); | |
1490 | return ret; | |
1491 | } | |
1492 | ||
033908ae | 1493 | page = i915_gem_object_get_dirty_page(wa_ctx->obj, 0); |
17ee950d AS |
1494 | batch = kmap_atomic(page); |
1495 | offset = 0; | |
1496 | ||
1497 | if (INTEL_INFO(ring->dev)->gen == 8) { | |
1498 | ret = gen8_init_indirectctx_bb(ring, | |
1499 | &wa_ctx->indirect_ctx, | |
1500 | batch, | |
1501 | &offset); | |
1502 | if (ret) | |
1503 | goto out; | |
1504 | ||
1505 | ret = gen8_init_perctx_bb(ring, | |
1506 | &wa_ctx->per_ctx, | |
1507 | batch, | |
1508 | &offset); | |
1509 | if (ret) | |
1510 | goto out; | |
0504cffc AS |
1511 | } else if (INTEL_INFO(ring->dev)->gen == 9) { |
1512 | ret = gen9_init_indirectctx_bb(ring, | |
1513 | &wa_ctx->indirect_ctx, | |
1514 | batch, | |
1515 | &offset); | |
1516 | if (ret) | |
1517 | goto out; | |
1518 | ||
1519 | ret = gen9_init_perctx_bb(ring, | |
1520 | &wa_ctx->per_ctx, | |
1521 | batch, | |
1522 | &offset); | |
1523 | if (ret) | |
1524 | goto out; | |
17ee950d AS |
1525 | } |
1526 | ||
1527 | out: | |
1528 | kunmap_atomic(batch); | |
1529 | if (ret) | |
1530 | lrc_destroy_wa_ctx_obj(ring); | |
1531 | ||
1532 | return ret; | |
1533 | } | |
1534 | ||
9b1136d5 OM |
1535 | static int gen8_init_common_ring(struct intel_engine_cs *ring) |
1536 | { | |
1537 | struct drm_device *dev = ring->dev; | |
1538 | struct drm_i915_private *dev_priv = dev->dev_private; | |
dfc53c5e | 1539 | u8 next_context_status_buffer_hw; |
9b1136d5 | 1540 | |
e84fe803 | 1541 | lrc_setup_hardware_status_page(ring, |
ed54c1a1 | 1542 | dev_priv->kernel_context->engine[ring->id].state); |
e84fe803 | 1543 | |
73d477f6 OM |
1544 | I915_WRITE_IMR(ring, ~(ring->irq_enable_mask | ring->irq_keep_mask)); |
1545 | I915_WRITE(RING_HWSTAM(ring->mmio_base), 0xffffffff); | |
1546 | ||
9b1136d5 OM |
1547 | I915_WRITE(RING_MODE_GEN7(ring), |
1548 | _MASKED_BIT_DISABLE(GFX_REPLAY_MODE) | | |
1549 | _MASKED_BIT_ENABLE(GFX_RUN_LIST_ENABLE)); | |
1550 | POSTING_READ(RING_MODE_GEN7(ring)); | |
dfc53c5e MT |
1551 | |
1552 | /* | |
1553 | * Instead of resetting the Context Status Buffer (CSB) read pointer to | |
1554 | * zero, we need to read the write pointer from hardware and use its | |
1555 | * value because "this register is power context save restored". | |
1556 | * Effectively, these states have been observed: | |
1557 | * | |
1558 | * | Suspend-to-idle (freeze) | Suspend-to-RAM (mem) | | |
1559 | * BDW | CSB regs not reset | CSB regs reset | | |
1560 | * CHT | CSB regs not reset | CSB regs not reset | | |
5590a5f0 BW |
1561 | * SKL | ? | ? | |
1562 | * BXT | ? | ? | | |
dfc53c5e | 1563 | */ |
5590a5f0 BW |
1564 | next_context_status_buffer_hw = |
1565 | GEN8_CSB_WRITE_PTR(I915_READ(RING_CONTEXT_STATUS_PTR(ring))); | |
dfc53c5e MT |
1566 | |
1567 | /* | |
1568 | * When the CSB registers are reset (also after power-up / gpu reset), | |
1569 | * CSB write pointer is set to all 1's, which is not valid, use '5' in | |
1570 | * this special case, so the first element read is CSB[0]. | |
1571 | */ | |
1572 | if (next_context_status_buffer_hw == GEN8_CSB_PTR_MASK) | |
1573 | next_context_status_buffer_hw = (GEN8_CSB_ENTRIES - 1); | |
1574 | ||
1575 | ring->next_context_status_buffer = next_context_status_buffer_hw; | |
9b1136d5 OM |
1576 | DRM_DEBUG_DRIVER("Execlists enabled for %s\n", ring->name); |
1577 | ||
1578 | memset(&ring->hangcheck, 0, sizeof(ring->hangcheck)); | |
1579 | ||
1580 | return 0; | |
1581 | } | |
1582 | ||
1583 | static int gen8_init_render_ring(struct intel_engine_cs *ring) | |
1584 | { | |
1585 | struct drm_device *dev = ring->dev; | |
1586 | struct drm_i915_private *dev_priv = dev->dev_private; | |
1587 | int ret; | |
1588 | ||
1589 | ret = gen8_init_common_ring(ring); | |
1590 | if (ret) | |
1591 | return ret; | |
1592 | ||
1593 | /* We need to disable the AsyncFlip performance optimisations in order | |
1594 | * to use MI_WAIT_FOR_EVENT within the CS. It should already be | |
1595 | * programmed to '1' on all products. | |
1596 | * | |
1597 | * WaDisableAsyncFlipPerfMode:snb,ivb,hsw,vlv,bdw,chv | |
1598 | */ | |
1599 | I915_WRITE(MI_MODE, _MASKED_BIT_ENABLE(ASYNC_FLIP_PERF_DISABLE)); | |
1600 | ||
9b1136d5 OM |
1601 | I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_FORCE_ORDERING)); |
1602 | ||
771b9a53 | 1603 | return init_workarounds_ring(ring); |
9b1136d5 OM |
1604 | } |
1605 | ||
82ef822e DL |
1606 | static int gen9_init_render_ring(struct intel_engine_cs *ring) |
1607 | { | |
1608 | int ret; | |
1609 | ||
1610 | ret = gen8_init_common_ring(ring); | |
1611 | if (ret) | |
1612 | return ret; | |
1613 | ||
1614 | return init_workarounds_ring(ring); | |
1615 | } | |
1616 | ||
7a01a0a2 MT |
1617 | static int intel_logical_ring_emit_pdps(struct drm_i915_gem_request *req) |
1618 | { | |
1619 | struct i915_hw_ppgtt *ppgtt = req->ctx->ppgtt; | |
1620 | struct intel_engine_cs *ring = req->ring; | |
1621 | struct intel_ringbuffer *ringbuf = req->ringbuf; | |
1622 | const int num_lri_cmds = GEN8_LEGACY_PDPES * 2; | |
1623 | int i, ret; | |
1624 | ||
1625 | ret = intel_logical_ring_begin(req, num_lri_cmds * 2 + 2); | |
1626 | if (ret) | |
1627 | return ret; | |
1628 | ||
1629 | intel_logical_ring_emit(ringbuf, MI_LOAD_REGISTER_IMM(num_lri_cmds)); | |
1630 | for (i = GEN8_LEGACY_PDPES - 1; i >= 0; i--) { | |
1631 | const dma_addr_t pd_daddr = i915_page_dir_dma_addr(ppgtt, i); | |
1632 | ||
f92a9162 | 1633 | intel_logical_ring_emit_reg(ringbuf, GEN8_RING_PDP_UDW(ring, i)); |
7a01a0a2 | 1634 | intel_logical_ring_emit(ringbuf, upper_32_bits(pd_daddr)); |
f92a9162 | 1635 | intel_logical_ring_emit_reg(ringbuf, GEN8_RING_PDP_LDW(ring, i)); |
7a01a0a2 MT |
1636 | intel_logical_ring_emit(ringbuf, lower_32_bits(pd_daddr)); |
1637 | } | |
1638 | ||
1639 | intel_logical_ring_emit(ringbuf, MI_NOOP); | |
1640 | intel_logical_ring_advance(ringbuf); | |
1641 | ||
1642 | return 0; | |
1643 | } | |
1644 | ||
be795fc1 | 1645 | static int gen8_emit_bb_start(struct drm_i915_gem_request *req, |
8e004efc | 1646 | u64 offset, unsigned dispatch_flags) |
15648585 | 1647 | { |
be795fc1 | 1648 | struct intel_ringbuffer *ringbuf = req->ringbuf; |
8e004efc | 1649 | bool ppgtt = !(dispatch_flags & I915_DISPATCH_SECURE); |
15648585 OM |
1650 | int ret; |
1651 | ||
7a01a0a2 MT |
1652 | /* Don't rely in hw updating PDPs, specially in lite-restore. |
1653 | * Ideally, we should set Force PD Restore in ctx descriptor, | |
1654 | * but we can't. Force Restore would be a second option, but | |
1655 | * it is unsafe in case of lite-restore (because the ctx is | |
2dba3239 MT |
1656 | * not idle). PML4 is allocated during ppgtt init so this is |
1657 | * not needed in 48-bit.*/ | |
7a01a0a2 MT |
1658 | if (req->ctx->ppgtt && |
1659 | (intel_ring_flag(req->ring) & req->ctx->ppgtt->pd_dirty_rings)) { | |
331f38e7 ZL |
1660 | if (!USES_FULL_48BIT_PPGTT(req->i915) && |
1661 | !intel_vgpu_active(req->i915->dev)) { | |
2dba3239 MT |
1662 | ret = intel_logical_ring_emit_pdps(req); |
1663 | if (ret) | |
1664 | return ret; | |
1665 | } | |
7a01a0a2 MT |
1666 | |
1667 | req->ctx->ppgtt->pd_dirty_rings &= ~intel_ring_flag(req->ring); | |
1668 | } | |
1669 | ||
4d616a29 | 1670 | ret = intel_logical_ring_begin(req, 4); |
15648585 OM |
1671 | if (ret) |
1672 | return ret; | |
1673 | ||
1674 | /* FIXME(BDW): Address space and security selectors. */ | |
6922528a AJ |
1675 | intel_logical_ring_emit(ringbuf, MI_BATCH_BUFFER_START_GEN8 | |
1676 | (ppgtt<<8) | | |
1677 | (dispatch_flags & I915_DISPATCH_RS ? | |
1678 | MI_BATCH_RESOURCE_STREAMER : 0)); | |
15648585 OM |
1679 | intel_logical_ring_emit(ringbuf, lower_32_bits(offset)); |
1680 | intel_logical_ring_emit(ringbuf, upper_32_bits(offset)); | |
1681 | intel_logical_ring_emit(ringbuf, MI_NOOP); | |
1682 | intel_logical_ring_advance(ringbuf); | |
1683 | ||
1684 | return 0; | |
1685 | } | |
1686 | ||
73d477f6 OM |
1687 | static bool gen8_logical_ring_get_irq(struct intel_engine_cs *ring) |
1688 | { | |
1689 | struct drm_device *dev = ring->dev; | |
1690 | struct drm_i915_private *dev_priv = dev->dev_private; | |
1691 | unsigned long flags; | |
1692 | ||
7cd512f1 | 1693 | if (WARN_ON(!intel_irqs_enabled(dev_priv))) |
73d477f6 OM |
1694 | return false; |
1695 | ||
1696 | spin_lock_irqsave(&dev_priv->irq_lock, flags); | |
1697 | if (ring->irq_refcount++ == 0) { | |
1698 | I915_WRITE_IMR(ring, ~(ring->irq_enable_mask | ring->irq_keep_mask)); | |
1699 | POSTING_READ(RING_IMR(ring->mmio_base)); | |
1700 | } | |
1701 | spin_unlock_irqrestore(&dev_priv->irq_lock, flags); | |
1702 | ||
1703 | return true; | |
1704 | } | |
1705 | ||
1706 | static void gen8_logical_ring_put_irq(struct intel_engine_cs *ring) | |
1707 | { | |
1708 | struct drm_device *dev = ring->dev; | |
1709 | struct drm_i915_private *dev_priv = dev->dev_private; | |
1710 | unsigned long flags; | |
1711 | ||
1712 | spin_lock_irqsave(&dev_priv->irq_lock, flags); | |
1713 | if (--ring->irq_refcount == 0) { | |
1714 | I915_WRITE_IMR(ring, ~ring->irq_keep_mask); | |
1715 | POSTING_READ(RING_IMR(ring->mmio_base)); | |
1716 | } | |
1717 | spin_unlock_irqrestore(&dev_priv->irq_lock, flags); | |
1718 | } | |
1719 | ||
7deb4d39 | 1720 | static int gen8_emit_flush(struct drm_i915_gem_request *request, |
4712274c OM |
1721 | u32 invalidate_domains, |
1722 | u32 unused) | |
1723 | { | |
7deb4d39 | 1724 | struct intel_ringbuffer *ringbuf = request->ringbuf; |
4712274c OM |
1725 | struct intel_engine_cs *ring = ringbuf->ring; |
1726 | struct drm_device *dev = ring->dev; | |
1727 | struct drm_i915_private *dev_priv = dev->dev_private; | |
1728 | uint32_t cmd; | |
1729 | int ret; | |
1730 | ||
4d616a29 | 1731 | ret = intel_logical_ring_begin(request, 4); |
4712274c OM |
1732 | if (ret) |
1733 | return ret; | |
1734 | ||
1735 | cmd = MI_FLUSH_DW + 1; | |
1736 | ||
f0a1fb10 CW |
1737 | /* We always require a command barrier so that subsequent |
1738 | * commands, such as breadcrumb interrupts, are strictly ordered | |
1739 | * wrt the contents of the write cache being flushed to memory | |
1740 | * (and thus being coherent from the CPU). | |
1741 | */ | |
1742 | cmd |= MI_FLUSH_DW_STORE_INDEX | MI_FLUSH_DW_OP_STOREDW; | |
1743 | ||
1744 | if (invalidate_domains & I915_GEM_GPU_DOMAINS) { | |
1745 | cmd |= MI_INVALIDATE_TLB; | |
1746 | if (ring == &dev_priv->ring[VCS]) | |
1747 | cmd |= MI_INVALIDATE_BSD; | |
4712274c OM |
1748 | } |
1749 | ||
1750 | intel_logical_ring_emit(ringbuf, cmd); | |
1751 | intel_logical_ring_emit(ringbuf, | |
1752 | I915_GEM_HWS_SCRATCH_ADDR | | |
1753 | MI_FLUSH_DW_USE_GTT); | |
1754 | intel_logical_ring_emit(ringbuf, 0); /* upper addr */ | |
1755 | intel_logical_ring_emit(ringbuf, 0); /* value */ | |
1756 | intel_logical_ring_advance(ringbuf); | |
1757 | ||
1758 | return 0; | |
1759 | } | |
1760 | ||
7deb4d39 | 1761 | static int gen8_emit_flush_render(struct drm_i915_gem_request *request, |
4712274c OM |
1762 | u32 invalidate_domains, |
1763 | u32 flush_domains) | |
1764 | { | |
7deb4d39 | 1765 | struct intel_ringbuffer *ringbuf = request->ringbuf; |
4712274c OM |
1766 | struct intel_engine_cs *ring = ringbuf->ring; |
1767 | u32 scratch_addr = ring->scratch.gtt_offset + 2 * CACHELINE_BYTES; | |
1a5a9ce7 | 1768 | bool vf_flush_wa = false; |
4712274c OM |
1769 | u32 flags = 0; |
1770 | int ret; | |
1771 | ||
1772 | flags |= PIPE_CONTROL_CS_STALL; | |
1773 | ||
1774 | if (flush_domains) { | |
1775 | flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH; | |
1776 | flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH; | |
965fd602 | 1777 | flags |= PIPE_CONTROL_DC_FLUSH_ENABLE; |
40a24488 | 1778 | flags |= PIPE_CONTROL_FLUSH_ENABLE; |
4712274c OM |
1779 | } |
1780 | ||
1781 | if (invalidate_domains) { | |
1782 | flags |= PIPE_CONTROL_TLB_INVALIDATE; | |
1783 | flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE; | |
1784 | flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE; | |
1785 | flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE; | |
1786 | flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE; | |
1787 | flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE; | |
1788 | flags |= PIPE_CONTROL_QW_WRITE; | |
1789 | flags |= PIPE_CONTROL_GLOBAL_GTT_IVB; | |
4712274c | 1790 | |
1a5a9ce7 BW |
1791 | /* |
1792 | * On GEN9: before VF_CACHE_INVALIDATE we need to emit a NULL | |
1793 | * pipe control. | |
1794 | */ | |
1795 | if (IS_GEN9(ring->dev)) | |
1796 | vf_flush_wa = true; | |
1797 | } | |
9647ff36 | 1798 | |
4d616a29 | 1799 | ret = intel_logical_ring_begin(request, vf_flush_wa ? 12 : 6); |
4712274c OM |
1800 | if (ret) |
1801 | return ret; | |
1802 | ||
9647ff36 ID |
1803 | if (vf_flush_wa) { |
1804 | intel_logical_ring_emit(ringbuf, GFX_OP_PIPE_CONTROL(6)); | |
1805 | intel_logical_ring_emit(ringbuf, 0); | |
1806 | intel_logical_ring_emit(ringbuf, 0); | |
1807 | intel_logical_ring_emit(ringbuf, 0); | |
1808 | intel_logical_ring_emit(ringbuf, 0); | |
1809 | intel_logical_ring_emit(ringbuf, 0); | |
1810 | } | |
1811 | ||
4712274c OM |
1812 | intel_logical_ring_emit(ringbuf, GFX_OP_PIPE_CONTROL(6)); |
1813 | intel_logical_ring_emit(ringbuf, flags); | |
1814 | intel_logical_ring_emit(ringbuf, scratch_addr); | |
1815 | intel_logical_ring_emit(ringbuf, 0); | |
1816 | intel_logical_ring_emit(ringbuf, 0); | |
1817 | intel_logical_ring_emit(ringbuf, 0); | |
1818 | intel_logical_ring_advance(ringbuf); | |
1819 | ||
1820 | return 0; | |
1821 | } | |
1822 | ||
e94e37ad OM |
1823 | static u32 gen8_get_seqno(struct intel_engine_cs *ring, bool lazy_coherency) |
1824 | { | |
1825 | return intel_read_status_page(ring, I915_GEM_HWS_INDEX); | |
1826 | } | |
1827 | ||
1828 | static void gen8_set_seqno(struct intel_engine_cs *ring, u32 seqno) | |
1829 | { | |
1830 | intel_write_status_page(ring, I915_GEM_HWS_INDEX, seqno); | |
1831 | } | |
1832 | ||
319404df ID |
1833 | static u32 bxt_a_get_seqno(struct intel_engine_cs *ring, bool lazy_coherency) |
1834 | { | |
1835 | ||
1836 | /* | |
1837 | * On BXT A steppings there is a HW coherency issue whereby the | |
1838 | * MI_STORE_DATA_IMM storing the completed request's seqno | |
1839 | * occasionally doesn't invalidate the CPU cache. Work around this by | |
1840 | * clflushing the corresponding cacheline whenever the caller wants | |
1841 | * the coherency to be guaranteed. Note that this cacheline is known | |
1842 | * to be clean at this point, since we only write it in | |
1843 | * bxt_a_set_seqno(), where we also do a clflush after the write. So | |
1844 | * this clflush in practice becomes an invalidate operation. | |
1845 | */ | |
1846 | ||
1847 | if (!lazy_coherency) | |
1848 | intel_flush_status_page(ring, I915_GEM_HWS_INDEX); | |
1849 | ||
1850 | return intel_read_status_page(ring, I915_GEM_HWS_INDEX); | |
1851 | } | |
1852 | ||
1853 | static void bxt_a_set_seqno(struct intel_engine_cs *ring, u32 seqno) | |
1854 | { | |
1855 | intel_write_status_page(ring, I915_GEM_HWS_INDEX, seqno); | |
1856 | ||
1857 | /* See bxt_a_get_seqno() explaining the reason for the clflush. */ | |
1858 | intel_flush_status_page(ring, I915_GEM_HWS_INDEX); | |
1859 | } | |
1860 | ||
7c17d377 CW |
1861 | /* |
1862 | * Reserve space for 2 NOOPs at the end of each request to be | |
1863 | * used as a workaround for not being allowed to do lite | |
1864 | * restore with HEAD==TAIL (WaIdleLiteRestore). | |
1865 | */ | |
1866 | #define WA_TAIL_DWORDS 2 | |
1867 | ||
1868 | static inline u32 hws_seqno_address(struct intel_engine_cs *engine) | |
1869 | { | |
1870 | return engine->status_page.gfx_addr + I915_GEM_HWS_INDEX_ADDR; | |
1871 | } | |
1872 | ||
c4e76638 | 1873 | static int gen8_emit_request(struct drm_i915_gem_request *request) |
4da46e1e | 1874 | { |
c4e76638 | 1875 | struct intel_ringbuffer *ringbuf = request->ringbuf; |
4da46e1e OM |
1876 | int ret; |
1877 | ||
7c17d377 | 1878 | ret = intel_logical_ring_begin(request, 6 + WA_TAIL_DWORDS); |
4da46e1e OM |
1879 | if (ret) |
1880 | return ret; | |
1881 | ||
7c17d377 CW |
1882 | /* w/a: bit 5 needs to be zero for MI_FLUSH_DW address. */ |
1883 | BUILD_BUG_ON(I915_GEM_HWS_INDEX_ADDR & (1 << 5)); | |
4da46e1e | 1884 | |
4da46e1e | 1885 | intel_logical_ring_emit(ringbuf, |
7c17d377 CW |
1886 | (MI_FLUSH_DW + 1) | MI_FLUSH_DW_OP_STOREDW); |
1887 | intel_logical_ring_emit(ringbuf, | |
1888 | hws_seqno_address(request->ring) | | |
1889 | MI_FLUSH_DW_USE_GTT); | |
4da46e1e | 1890 | intel_logical_ring_emit(ringbuf, 0); |
c4e76638 | 1891 | intel_logical_ring_emit(ringbuf, i915_gem_request_get_seqno(request)); |
4da46e1e OM |
1892 | intel_logical_ring_emit(ringbuf, MI_USER_INTERRUPT); |
1893 | intel_logical_ring_emit(ringbuf, MI_NOOP); | |
7c17d377 CW |
1894 | return intel_logical_ring_advance_and_submit(request); |
1895 | } | |
4da46e1e | 1896 | |
7c17d377 CW |
1897 | static int gen8_emit_request_render(struct drm_i915_gem_request *request) |
1898 | { | |
1899 | struct intel_ringbuffer *ringbuf = request->ringbuf; | |
1900 | int ret; | |
53292cdb | 1901 | |
7c17d377 CW |
1902 | ret = intel_logical_ring_begin(request, 6 + WA_TAIL_DWORDS); |
1903 | if (ret) | |
1904 | return ret; | |
1905 | ||
1906 | /* w/a for post sync ops following a GPGPU operation we | |
1907 | * need a prior CS_STALL, which is emitted by the flush | |
1908 | * following the batch. | |
1909 | */ | |
1910 | intel_logical_ring_emit(ringbuf, GFX_OP_PIPE_CONTROL(5)); | |
1911 | intel_logical_ring_emit(ringbuf, | |
1912 | (PIPE_CONTROL_GLOBAL_GTT_IVB | | |
1913 | PIPE_CONTROL_CS_STALL | | |
1914 | PIPE_CONTROL_QW_WRITE)); | |
1915 | intel_logical_ring_emit(ringbuf, hws_seqno_address(request->ring)); | |
1916 | intel_logical_ring_emit(ringbuf, 0); | |
1917 | intel_logical_ring_emit(ringbuf, i915_gem_request_get_seqno(request)); | |
1918 | intel_logical_ring_emit(ringbuf, MI_USER_INTERRUPT); | |
1919 | return intel_logical_ring_advance_and_submit(request); | |
4da46e1e OM |
1920 | } |
1921 | ||
be01363f | 1922 | static int intel_lr_context_render_state_init(struct drm_i915_gem_request *req) |
cef437ad | 1923 | { |
cef437ad | 1924 | struct render_state so; |
cef437ad DL |
1925 | int ret; |
1926 | ||
be01363f | 1927 | ret = i915_gem_render_state_prepare(req->ring, &so); |
cef437ad DL |
1928 | if (ret) |
1929 | return ret; | |
1930 | ||
1931 | if (so.rodata == NULL) | |
1932 | return 0; | |
1933 | ||
be795fc1 | 1934 | ret = req->ring->emit_bb_start(req, so.ggtt_offset, |
be01363f | 1935 | I915_DISPATCH_SECURE); |
cef437ad DL |
1936 | if (ret) |
1937 | goto out; | |
1938 | ||
84e81020 AS |
1939 | ret = req->ring->emit_bb_start(req, |
1940 | (so.ggtt_offset + so.aux_batch_offset), | |
1941 | I915_DISPATCH_SECURE); | |
1942 | if (ret) | |
1943 | goto out; | |
1944 | ||
b2af0376 | 1945 | i915_vma_move_to_active(i915_gem_obj_to_ggtt(so.obj), req); |
cef437ad | 1946 | |
cef437ad DL |
1947 | out: |
1948 | i915_gem_render_state_fini(&so); | |
1949 | return ret; | |
1950 | } | |
1951 | ||
8753181e | 1952 | static int gen8_init_rcs_context(struct drm_i915_gem_request *req) |
e7778be1 TD |
1953 | { |
1954 | int ret; | |
1955 | ||
e2be4faf | 1956 | ret = intel_logical_ring_workarounds_emit(req); |
e7778be1 TD |
1957 | if (ret) |
1958 | return ret; | |
1959 | ||
3bbaba0c PA |
1960 | ret = intel_rcs_context_init_mocs(req); |
1961 | /* | |
1962 | * Failing to program the MOCS is non-fatal.The system will not | |
1963 | * run at peak performance. So generate an error and carry on. | |
1964 | */ | |
1965 | if (ret) | |
1966 | DRM_ERROR("MOCS failed to program: expect performance issues.\n"); | |
1967 | ||
be01363f | 1968 | return intel_lr_context_render_state_init(req); |
e7778be1 TD |
1969 | } |
1970 | ||
73e4d07f OM |
1971 | /** |
1972 | * intel_logical_ring_cleanup() - deallocate the Engine Command Streamer | |
1973 | * | |
1974 | * @ring: Engine Command Streamer. | |
1975 | * | |
1976 | */ | |
454afebd OM |
1977 | void intel_logical_ring_cleanup(struct intel_engine_cs *ring) |
1978 | { | |
6402c330 | 1979 | struct drm_i915_private *dev_priv; |
9832b9da | 1980 | |
48d82387 OM |
1981 | if (!intel_ring_initialized(ring)) |
1982 | return; | |
1983 | ||
6402c330 JH |
1984 | dev_priv = ring->dev->dev_private; |
1985 | ||
b0366a54 DG |
1986 | if (ring->buffer) { |
1987 | intel_logical_ring_stop(ring); | |
1988 | WARN_ON((I915_READ_MODE(ring) & MODE_IDLE) == 0); | |
1989 | } | |
48d82387 OM |
1990 | |
1991 | if (ring->cleanup) | |
1992 | ring->cleanup(ring); | |
1993 | ||
1994 | i915_cmd_parser_fini_ring(ring); | |
06fbca71 | 1995 | i915_gem_batch_pool_fini(&ring->batch_pool); |
48d82387 OM |
1996 | |
1997 | if (ring->status_page.obj) { | |
1998 | kunmap(sg_page(ring->status_page.obj->pages->sgl)); | |
1999 | ring->status_page.obj = NULL; | |
2000 | } | |
17ee950d | 2001 | |
ca82580c TU |
2002 | ring->disable_lite_restore_wa = false; |
2003 | ring->ctx_desc_template = 0; | |
2004 | ||
17ee950d | 2005 | lrc_destroy_wa_ctx_obj(ring); |
b0366a54 | 2006 | ring->dev = NULL; |
454afebd OM |
2007 | } |
2008 | ||
c9cacf93 TU |
2009 | static void |
2010 | logical_ring_default_vfuncs(struct drm_device *dev, | |
2011 | struct intel_engine_cs *ring) | |
2012 | { | |
2013 | /* Default vfuncs which can be overriden by each engine. */ | |
2014 | ring->init_hw = gen8_init_common_ring; | |
2015 | ring->emit_request = gen8_emit_request; | |
2016 | ring->emit_flush = gen8_emit_flush; | |
2017 | ring->irq_get = gen8_logical_ring_get_irq; | |
2018 | ring->irq_put = gen8_logical_ring_put_irq; | |
2019 | ring->emit_bb_start = gen8_emit_bb_start; | |
2020 | if (IS_BXT_REVID(dev, 0, BXT_REVID_A1)) { | |
2021 | ring->get_seqno = bxt_a_get_seqno; | |
2022 | ring->set_seqno = bxt_a_set_seqno; | |
2023 | } else { | |
2024 | ring->get_seqno = gen8_get_seqno; | |
2025 | ring->set_seqno = gen8_set_seqno; | |
2026 | } | |
2027 | } | |
2028 | ||
d9f3af96 TU |
2029 | static inline void |
2030 | logical_ring_default_irqs(struct intel_engine_cs *ring, unsigned shift) | |
2031 | { | |
2032 | ring->irq_enable_mask = GT_RENDER_USER_INTERRUPT << shift; | |
2033 | ring->irq_keep_mask = GT_CONTEXT_SWITCH_INTERRUPT << shift; | |
2034 | } | |
2035 | ||
c9cacf93 TU |
2036 | static int |
2037 | logical_ring_init(struct drm_device *dev, struct intel_engine_cs *ring) | |
454afebd | 2038 | { |
ed54c1a1 | 2039 | struct intel_context *dctx = to_i915(dev)->kernel_context; |
48d82387 | 2040 | int ret; |
48d82387 OM |
2041 | |
2042 | /* Intentionally left blank. */ | |
2043 | ring->buffer = NULL; | |
2044 | ||
2045 | ring->dev = dev; | |
2046 | INIT_LIST_HEAD(&ring->active_list); | |
2047 | INIT_LIST_HEAD(&ring->request_list); | |
06fbca71 | 2048 | i915_gem_batch_pool_init(dev, &ring->batch_pool); |
48d82387 OM |
2049 | init_waitqueue_head(&ring->irq_queue); |
2050 | ||
608c1a52 | 2051 | INIT_LIST_HEAD(&ring->buffers); |
acdd884a | 2052 | INIT_LIST_HEAD(&ring->execlist_queue); |
c86ee3a9 | 2053 | INIT_LIST_HEAD(&ring->execlist_retired_req_list); |
acdd884a MT |
2054 | spin_lock_init(&ring->execlist_lock); |
2055 | ||
ca82580c TU |
2056 | logical_ring_init_platform_invariants(ring); |
2057 | ||
48d82387 OM |
2058 | ret = i915_cmd_parser_init_ring(ring); |
2059 | if (ret) | |
b0366a54 | 2060 | goto error; |
48d82387 | 2061 | |
ed54c1a1 | 2062 | ret = intel_lr_context_deferred_alloc(dctx, ring); |
e84fe803 | 2063 | if (ret) |
b0366a54 | 2064 | goto error; |
e84fe803 NH |
2065 | |
2066 | /* As this is the default context, always pin it */ | |
ed54c1a1 | 2067 | ret = intel_lr_context_do_pin(ring, dctx); |
e84fe803 NH |
2068 | if (ret) { |
2069 | DRM_ERROR( | |
2070 | "Failed to pin and map ringbuffer %s: %d\n", | |
2071 | ring->name, ret); | |
b0366a54 | 2072 | goto error; |
e84fe803 | 2073 | } |
564ddb2f | 2074 | |
b0366a54 DG |
2075 | return 0; |
2076 | ||
2077 | error: | |
2078 | intel_logical_ring_cleanup(ring); | |
564ddb2f | 2079 | return ret; |
454afebd OM |
2080 | } |
2081 | ||
2082 | static int logical_render_ring_init(struct drm_device *dev) | |
2083 | { | |
2084 | struct drm_i915_private *dev_priv = dev->dev_private; | |
2085 | struct intel_engine_cs *ring = &dev_priv->ring[RCS]; | |
99be1dfe | 2086 | int ret; |
454afebd OM |
2087 | |
2088 | ring->name = "render ring"; | |
2089 | ring->id = RCS; | |
426960be | 2090 | ring->exec_id = I915_EXEC_RENDER; |
397097b0 | 2091 | ring->guc_id = GUC_RENDER_ENGINE; |
454afebd | 2092 | ring->mmio_base = RENDER_RING_BASE; |
d9f3af96 TU |
2093 | |
2094 | logical_ring_default_irqs(ring, GEN8_RCS_IRQ_SHIFT); | |
73d477f6 OM |
2095 | if (HAS_L3_DPF(dev)) |
2096 | ring->irq_keep_mask |= GT_RENDER_L3_PARITY_ERROR_INTERRUPT; | |
454afebd | 2097 | |
c9cacf93 TU |
2098 | logical_ring_default_vfuncs(dev, ring); |
2099 | ||
2100 | /* Override some for render ring. */ | |
82ef822e DL |
2101 | if (INTEL_INFO(dev)->gen >= 9) |
2102 | ring->init_hw = gen9_init_render_ring; | |
2103 | else | |
2104 | ring->init_hw = gen8_init_render_ring; | |
e7778be1 | 2105 | ring->init_context = gen8_init_rcs_context; |
9b1136d5 | 2106 | ring->cleanup = intel_fini_pipe_control; |
4712274c | 2107 | ring->emit_flush = gen8_emit_flush_render; |
7c17d377 | 2108 | ring->emit_request = gen8_emit_request_render; |
9b1136d5 | 2109 | |
99be1dfe | 2110 | ring->dev = dev; |
c4db7599 AS |
2111 | |
2112 | ret = intel_init_pipe_control(ring); | |
99be1dfe DV |
2113 | if (ret) |
2114 | return ret; | |
2115 | ||
17ee950d AS |
2116 | ret = intel_init_workaround_bb(ring); |
2117 | if (ret) { | |
2118 | /* | |
2119 | * We continue even if we fail to initialize WA batch | |
2120 | * because we only expect rare glitches but nothing | |
2121 | * critical to prevent us from using GPU | |
2122 | */ | |
2123 | DRM_ERROR("WA batch buffer initialization failed: %d\n", | |
2124 | ret); | |
2125 | } | |
2126 | ||
c4db7599 AS |
2127 | ret = logical_ring_init(dev, ring); |
2128 | if (ret) { | |
17ee950d | 2129 | lrc_destroy_wa_ctx_obj(ring); |
c4db7599 | 2130 | } |
17ee950d AS |
2131 | |
2132 | return ret; | |
454afebd OM |
2133 | } |
2134 | ||
2135 | static int logical_bsd_ring_init(struct drm_device *dev) | |
2136 | { | |
2137 | struct drm_i915_private *dev_priv = dev->dev_private; | |
2138 | struct intel_engine_cs *ring = &dev_priv->ring[VCS]; | |
2139 | ||
2140 | ring->name = "bsd ring"; | |
2141 | ring->id = VCS; | |
426960be | 2142 | ring->exec_id = I915_EXEC_BSD; |
397097b0 | 2143 | ring->guc_id = GUC_VIDEO_ENGINE; |
454afebd | 2144 | ring->mmio_base = GEN6_BSD_RING_BASE; |
454afebd | 2145 | |
d9f3af96 | 2146 | logical_ring_default_irqs(ring, GEN8_VCS1_IRQ_SHIFT); |
c9cacf93 | 2147 | logical_ring_default_vfuncs(dev, ring); |
9b1136d5 | 2148 | |
454afebd OM |
2149 | return logical_ring_init(dev, ring); |
2150 | } | |
2151 | ||
2152 | static int logical_bsd2_ring_init(struct drm_device *dev) | |
2153 | { | |
2154 | struct drm_i915_private *dev_priv = dev->dev_private; | |
2155 | struct intel_engine_cs *ring = &dev_priv->ring[VCS2]; | |
2156 | ||
ec8a9776 | 2157 | ring->name = "bsd2 ring"; |
454afebd | 2158 | ring->id = VCS2; |
426960be | 2159 | ring->exec_id = I915_EXEC_BSD; |
397097b0 | 2160 | ring->guc_id = GUC_VIDEO_ENGINE2; |
454afebd | 2161 | ring->mmio_base = GEN8_BSD2_RING_BASE; |
454afebd | 2162 | |
d9f3af96 | 2163 | logical_ring_default_irqs(ring, GEN8_VCS2_IRQ_SHIFT); |
c9cacf93 | 2164 | logical_ring_default_vfuncs(dev, ring); |
9b1136d5 | 2165 | |
454afebd OM |
2166 | return logical_ring_init(dev, ring); |
2167 | } | |
2168 | ||
2169 | static int logical_blt_ring_init(struct drm_device *dev) | |
2170 | { | |
2171 | struct drm_i915_private *dev_priv = dev->dev_private; | |
2172 | struct intel_engine_cs *ring = &dev_priv->ring[BCS]; | |
2173 | ||
2174 | ring->name = "blitter ring"; | |
2175 | ring->id = BCS; | |
426960be | 2176 | ring->exec_id = I915_EXEC_BLT; |
397097b0 | 2177 | ring->guc_id = GUC_BLITTER_ENGINE; |
454afebd | 2178 | ring->mmio_base = BLT_RING_BASE; |
454afebd | 2179 | |
d9f3af96 | 2180 | logical_ring_default_irqs(ring, GEN8_BCS_IRQ_SHIFT); |
c9cacf93 | 2181 | logical_ring_default_vfuncs(dev, ring); |
9b1136d5 | 2182 | |
454afebd OM |
2183 | return logical_ring_init(dev, ring); |
2184 | } | |
2185 | ||
2186 | static int logical_vebox_ring_init(struct drm_device *dev) | |
2187 | { | |
2188 | struct drm_i915_private *dev_priv = dev->dev_private; | |
2189 | struct intel_engine_cs *ring = &dev_priv->ring[VECS]; | |
2190 | ||
2191 | ring->name = "video enhancement ring"; | |
2192 | ring->id = VECS; | |
426960be | 2193 | ring->exec_id = I915_EXEC_VEBOX; |
397097b0 | 2194 | ring->guc_id = GUC_VIDEOENHANCE_ENGINE; |
454afebd | 2195 | ring->mmio_base = VEBOX_RING_BASE; |
454afebd | 2196 | |
d9f3af96 | 2197 | logical_ring_default_irqs(ring, GEN8_VECS_IRQ_SHIFT); |
c9cacf93 | 2198 | logical_ring_default_vfuncs(dev, ring); |
9b1136d5 | 2199 | |
454afebd OM |
2200 | return logical_ring_init(dev, ring); |
2201 | } | |
2202 | ||
73e4d07f OM |
2203 | /** |
2204 | * intel_logical_rings_init() - allocate, populate and init the Engine Command Streamers | |
2205 | * @dev: DRM device. | |
2206 | * | |
2207 | * This function inits the engines for an Execlists submission style (the equivalent in the | |
2208 | * legacy ringbuffer submission world would be i915_gem_init_rings). It does it only for | |
2209 | * those engines that are present in the hardware. | |
2210 | * | |
2211 | * Return: non-zero if the initialization failed. | |
2212 | */ | |
454afebd OM |
2213 | int intel_logical_rings_init(struct drm_device *dev) |
2214 | { | |
2215 | struct drm_i915_private *dev_priv = dev->dev_private; | |
2216 | int ret; | |
2217 | ||
2218 | ret = logical_render_ring_init(dev); | |
2219 | if (ret) | |
2220 | return ret; | |
2221 | ||
2222 | if (HAS_BSD(dev)) { | |
2223 | ret = logical_bsd_ring_init(dev); | |
2224 | if (ret) | |
2225 | goto cleanup_render_ring; | |
2226 | } | |
2227 | ||
2228 | if (HAS_BLT(dev)) { | |
2229 | ret = logical_blt_ring_init(dev); | |
2230 | if (ret) | |
2231 | goto cleanup_bsd_ring; | |
2232 | } | |
2233 | ||
2234 | if (HAS_VEBOX(dev)) { | |
2235 | ret = logical_vebox_ring_init(dev); | |
2236 | if (ret) | |
2237 | goto cleanup_blt_ring; | |
2238 | } | |
2239 | ||
2240 | if (HAS_BSD2(dev)) { | |
2241 | ret = logical_bsd2_ring_init(dev); | |
2242 | if (ret) | |
2243 | goto cleanup_vebox_ring; | |
2244 | } | |
2245 | ||
454afebd OM |
2246 | return 0; |
2247 | ||
454afebd OM |
2248 | cleanup_vebox_ring: |
2249 | intel_logical_ring_cleanup(&dev_priv->ring[VECS]); | |
2250 | cleanup_blt_ring: | |
2251 | intel_logical_ring_cleanup(&dev_priv->ring[BCS]); | |
2252 | cleanup_bsd_ring: | |
2253 | intel_logical_ring_cleanup(&dev_priv->ring[VCS]); | |
2254 | cleanup_render_ring: | |
2255 | intel_logical_ring_cleanup(&dev_priv->ring[RCS]); | |
2256 | ||
2257 | return ret; | |
2258 | } | |
2259 | ||
0cea6502 JM |
2260 | static u32 |
2261 | make_rpcs(struct drm_device *dev) | |
2262 | { | |
2263 | u32 rpcs = 0; | |
2264 | ||
2265 | /* | |
2266 | * No explicit RPCS request is needed to ensure full | |
2267 | * slice/subslice/EU enablement prior to Gen9. | |
2268 | */ | |
2269 | if (INTEL_INFO(dev)->gen < 9) | |
2270 | return 0; | |
2271 | ||
2272 | /* | |
2273 | * Starting in Gen9, render power gating can leave | |
2274 | * slice/subslice/EU in a partially enabled state. We | |
2275 | * must make an explicit request through RPCS for full | |
2276 | * enablement. | |
2277 | */ | |
2278 | if (INTEL_INFO(dev)->has_slice_pg) { | |
2279 | rpcs |= GEN8_RPCS_S_CNT_ENABLE; | |
2280 | rpcs |= INTEL_INFO(dev)->slice_total << | |
2281 | GEN8_RPCS_S_CNT_SHIFT; | |
2282 | rpcs |= GEN8_RPCS_ENABLE; | |
2283 | } | |
2284 | ||
2285 | if (INTEL_INFO(dev)->has_subslice_pg) { | |
2286 | rpcs |= GEN8_RPCS_SS_CNT_ENABLE; | |
2287 | rpcs |= INTEL_INFO(dev)->subslice_per_slice << | |
2288 | GEN8_RPCS_SS_CNT_SHIFT; | |
2289 | rpcs |= GEN8_RPCS_ENABLE; | |
2290 | } | |
2291 | ||
2292 | if (INTEL_INFO(dev)->has_eu_pg) { | |
2293 | rpcs |= INTEL_INFO(dev)->eu_per_subslice << | |
2294 | GEN8_RPCS_EU_MIN_SHIFT; | |
2295 | rpcs |= INTEL_INFO(dev)->eu_per_subslice << | |
2296 | GEN8_RPCS_EU_MAX_SHIFT; | |
2297 | rpcs |= GEN8_RPCS_ENABLE; | |
2298 | } | |
2299 | ||
2300 | return rpcs; | |
2301 | } | |
2302 | ||
8670d6f9 OM |
2303 | static int |
2304 | populate_lr_context(struct intel_context *ctx, struct drm_i915_gem_object *ctx_obj, | |
2305 | struct intel_engine_cs *ring, struct intel_ringbuffer *ringbuf) | |
2306 | { | |
2d965536 TD |
2307 | struct drm_device *dev = ring->dev; |
2308 | struct drm_i915_private *dev_priv = dev->dev_private; | |
ae6c4806 | 2309 | struct i915_hw_ppgtt *ppgtt = ctx->ppgtt; |
8670d6f9 OM |
2310 | struct page *page; |
2311 | uint32_t *reg_state; | |
2312 | int ret; | |
2313 | ||
2d965536 TD |
2314 | if (!ppgtt) |
2315 | ppgtt = dev_priv->mm.aliasing_ppgtt; | |
2316 | ||
8670d6f9 OM |
2317 | ret = i915_gem_object_set_to_cpu_domain(ctx_obj, true); |
2318 | if (ret) { | |
2319 | DRM_DEBUG_DRIVER("Could not set to CPU domain\n"); | |
2320 | return ret; | |
2321 | } | |
2322 | ||
2323 | ret = i915_gem_object_get_pages(ctx_obj); | |
2324 | if (ret) { | |
2325 | DRM_DEBUG_DRIVER("Could not get object pages\n"); | |
2326 | return ret; | |
2327 | } | |
2328 | ||
2329 | i915_gem_object_pin_pages(ctx_obj); | |
2330 | ||
2331 | /* The second page of the context object contains some fields which must | |
2332 | * be set up prior to the first execution. */ | |
033908ae | 2333 | page = i915_gem_object_get_dirty_page(ctx_obj, LRC_STATE_PN); |
8670d6f9 OM |
2334 | reg_state = kmap_atomic(page); |
2335 | ||
2336 | /* A context is actually a big batch buffer with several MI_LOAD_REGISTER_IMM | |
2337 | * commands followed by (reg, value) pairs. The values we are setting here are | |
2338 | * only for the first context restore: on a subsequent save, the GPU will | |
2339 | * recreate this batchbuffer with new values (including all the missing | |
2340 | * MI_LOAD_REGISTER_IMM commands that we are not initializing here). */ | |
0d925ea0 VS |
2341 | reg_state[CTX_LRI_HEADER_0] = |
2342 | MI_LOAD_REGISTER_IMM(ring->id == RCS ? 14 : 11) | MI_LRI_FORCE_POSTED; | |
2343 | ASSIGN_CTX_REG(reg_state, CTX_CONTEXT_CONTROL, RING_CONTEXT_CONTROL(ring), | |
2344 | _MASKED_BIT_ENABLE(CTX_CTRL_INHIBIT_SYN_CTX_SWITCH | | |
2345 | CTX_CTRL_ENGINE_CTX_RESTORE_INHIBIT | | |
2346 | CTX_CTRL_RS_CTX_ENABLE)); | |
2347 | ASSIGN_CTX_REG(reg_state, CTX_RING_HEAD, RING_HEAD(ring->mmio_base), 0); | |
2348 | ASSIGN_CTX_REG(reg_state, CTX_RING_TAIL, RING_TAIL(ring->mmio_base), 0); | |
7ba717cf TD |
2349 | /* Ring buffer start address is not known until the buffer is pinned. |
2350 | * It is written to the context image in execlists_update_context() | |
2351 | */ | |
0d925ea0 VS |
2352 | ASSIGN_CTX_REG(reg_state, CTX_RING_BUFFER_START, RING_START(ring->mmio_base), 0); |
2353 | ASSIGN_CTX_REG(reg_state, CTX_RING_BUFFER_CONTROL, RING_CTL(ring->mmio_base), | |
2354 | ((ringbuf->size - PAGE_SIZE) & RING_NR_PAGES) | RING_VALID); | |
2355 | ASSIGN_CTX_REG(reg_state, CTX_BB_HEAD_U, RING_BBADDR_UDW(ring->mmio_base), 0); | |
2356 | ASSIGN_CTX_REG(reg_state, CTX_BB_HEAD_L, RING_BBADDR(ring->mmio_base), 0); | |
2357 | ASSIGN_CTX_REG(reg_state, CTX_BB_STATE, RING_BBSTATE(ring->mmio_base), | |
2358 | RING_BB_PPGTT); | |
2359 | ASSIGN_CTX_REG(reg_state, CTX_SECOND_BB_HEAD_U, RING_SBBADDR_UDW(ring->mmio_base), 0); | |
2360 | ASSIGN_CTX_REG(reg_state, CTX_SECOND_BB_HEAD_L, RING_SBBADDR(ring->mmio_base), 0); | |
2361 | ASSIGN_CTX_REG(reg_state, CTX_SECOND_BB_STATE, RING_SBBSTATE(ring->mmio_base), 0); | |
8670d6f9 | 2362 | if (ring->id == RCS) { |
0d925ea0 VS |
2363 | ASSIGN_CTX_REG(reg_state, CTX_BB_PER_CTX_PTR, RING_BB_PER_CTX_PTR(ring->mmio_base), 0); |
2364 | ASSIGN_CTX_REG(reg_state, CTX_RCS_INDIRECT_CTX, RING_INDIRECT_CTX(ring->mmio_base), 0); | |
2365 | ASSIGN_CTX_REG(reg_state, CTX_RCS_INDIRECT_CTX_OFFSET, RING_INDIRECT_CTX_OFFSET(ring->mmio_base), 0); | |
17ee950d AS |
2366 | if (ring->wa_ctx.obj) { |
2367 | struct i915_ctx_workarounds *wa_ctx = &ring->wa_ctx; | |
2368 | uint32_t ggtt_offset = i915_gem_obj_ggtt_offset(wa_ctx->obj); | |
2369 | ||
2370 | reg_state[CTX_RCS_INDIRECT_CTX+1] = | |
2371 | (ggtt_offset + wa_ctx->indirect_ctx.offset * sizeof(uint32_t)) | | |
2372 | (wa_ctx->indirect_ctx.size / CACHELINE_DWORDS); | |
2373 | ||
2374 | reg_state[CTX_RCS_INDIRECT_CTX_OFFSET+1] = | |
2375 | CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT << 6; | |
2376 | ||
2377 | reg_state[CTX_BB_PER_CTX_PTR+1] = | |
2378 | (ggtt_offset + wa_ctx->per_ctx.offset * sizeof(uint32_t)) | | |
2379 | 0x01; | |
2380 | } | |
8670d6f9 | 2381 | } |
0d925ea0 VS |
2382 | reg_state[CTX_LRI_HEADER_1] = MI_LOAD_REGISTER_IMM(9) | MI_LRI_FORCE_POSTED; |
2383 | ASSIGN_CTX_REG(reg_state, CTX_CTX_TIMESTAMP, RING_CTX_TIMESTAMP(ring->mmio_base), 0); | |
2384 | /* PDP values well be assigned later if needed */ | |
2385 | ASSIGN_CTX_REG(reg_state, CTX_PDP3_UDW, GEN8_RING_PDP_UDW(ring, 3), 0); | |
2386 | ASSIGN_CTX_REG(reg_state, CTX_PDP3_LDW, GEN8_RING_PDP_LDW(ring, 3), 0); | |
2387 | ASSIGN_CTX_REG(reg_state, CTX_PDP2_UDW, GEN8_RING_PDP_UDW(ring, 2), 0); | |
2388 | ASSIGN_CTX_REG(reg_state, CTX_PDP2_LDW, GEN8_RING_PDP_LDW(ring, 2), 0); | |
2389 | ASSIGN_CTX_REG(reg_state, CTX_PDP1_UDW, GEN8_RING_PDP_UDW(ring, 1), 0); | |
2390 | ASSIGN_CTX_REG(reg_state, CTX_PDP1_LDW, GEN8_RING_PDP_LDW(ring, 1), 0); | |
2391 | ASSIGN_CTX_REG(reg_state, CTX_PDP0_UDW, GEN8_RING_PDP_UDW(ring, 0), 0); | |
2392 | ASSIGN_CTX_REG(reg_state, CTX_PDP0_LDW, GEN8_RING_PDP_LDW(ring, 0), 0); | |
d7b2633d | 2393 | |
2dba3239 MT |
2394 | if (USES_FULL_48BIT_PPGTT(ppgtt->base.dev)) { |
2395 | /* 64b PPGTT (48bit canonical) | |
2396 | * PDP0_DESCRIPTOR contains the base address to PML4 and | |
2397 | * other PDP Descriptors are ignored. | |
2398 | */ | |
2399 | ASSIGN_CTX_PML4(ppgtt, reg_state); | |
2400 | } else { | |
2401 | /* 32b PPGTT | |
2402 | * PDP*_DESCRIPTOR contains the base address of space supported. | |
2403 | * With dynamic page allocation, PDPs may not be allocated at | |
2404 | * this point. Point the unallocated PDPs to the scratch page | |
2405 | */ | |
2406 | ASSIGN_CTX_PDP(ppgtt, reg_state, 3); | |
2407 | ASSIGN_CTX_PDP(ppgtt, reg_state, 2); | |
2408 | ASSIGN_CTX_PDP(ppgtt, reg_state, 1); | |
2409 | ASSIGN_CTX_PDP(ppgtt, reg_state, 0); | |
2410 | } | |
2411 | ||
8670d6f9 OM |
2412 | if (ring->id == RCS) { |
2413 | reg_state[CTX_LRI_HEADER_2] = MI_LOAD_REGISTER_IMM(1); | |
0d925ea0 VS |
2414 | ASSIGN_CTX_REG(reg_state, CTX_R_PWR_CLK_STATE, GEN8_R_PWR_CLK_STATE, |
2415 | make_rpcs(dev)); | |
8670d6f9 OM |
2416 | } |
2417 | ||
2418 | kunmap_atomic(reg_state); | |
8670d6f9 OM |
2419 | i915_gem_object_unpin_pages(ctx_obj); |
2420 | ||
2421 | return 0; | |
2422 | } | |
2423 | ||
73e4d07f OM |
2424 | /** |
2425 | * intel_lr_context_free() - free the LRC specific bits of a context | |
2426 | * @ctx: the LR context to free. | |
2427 | * | |
2428 | * The real context freeing is done in i915_gem_context_free: this only | |
2429 | * takes care of the bits that are LRC related: the per-engine backing | |
2430 | * objects and the logical ringbuffer. | |
2431 | */ | |
ede7d42b OM |
2432 | void intel_lr_context_free(struct intel_context *ctx) |
2433 | { | |
8c857917 OM |
2434 | int i; |
2435 | ||
e28e404c DG |
2436 | for (i = I915_NUM_RINGS; --i >= 0; ) { |
2437 | struct intel_ringbuffer *ringbuf = ctx->engine[i].ringbuf; | |
8c857917 | 2438 | struct drm_i915_gem_object *ctx_obj = ctx->engine[i].state; |
84c2377f | 2439 | |
e28e404c DG |
2440 | if (!ctx_obj) |
2441 | continue; | |
dcb4c12a | 2442 | |
e28e404c DG |
2443 | if (ctx == ctx->i915->kernel_context) { |
2444 | intel_unpin_ringbuffer_obj(ringbuf); | |
2445 | i915_gem_object_ggtt_unpin(ctx_obj); | |
8c857917 | 2446 | } |
e28e404c DG |
2447 | |
2448 | WARN_ON(ctx->engine[i].pin_count); | |
2449 | intel_ringbuffer_free(ringbuf); | |
2450 | drm_gem_object_unreference(&ctx_obj->base); | |
8c857917 OM |
2451 | } |
2452 | } | |
2453 | ||
c5d46ee2 DG |
2454 | /** |
2455 | * intel_lr_context_size() - return the size of the context for an engine | |
2456 | * @ring: which engine to find the context size for | |
2457 | * | |
2458 | * Each engine may require a different amount of space for a context image, | |
2459 | * so when allocating (or copying) an image, this function can be used to | |
2460 | * find the right size for the specific engine. | |
2461 | * | |
2462 | * Return: size (in bytes) of an engine-specific context image | |
2463 | * | |
2464 | * Note: this size includes the HWSP, which is part of the context image | |
2465 | * in LRC mode, but does not include the "shared data page" used with | |
2466 | * GuC submission. The caller should account for this if using the GuC. | |
2467 | */ | |
95a66f7e | 2468 | uint32_t intel_lr_context_size(struct intel_engine_cs *ring) |
8c857917 OM |
2469 | { |
2470 | int ret = 0; | |
2471 | ||
468c6816 | 2472 | WARN_ON(INTEL_INFO(ring->dev)->gen < 8); |
8c857917 OM |
2473 | |
2474 | switch (ring->id) { | |
2475 | case RCS: | |
468c6816 MN |
2476 | if (INTEL_INFO(ring->dev)->gen >= 9) |
2477 | ret = GEN9_LR_CONTEXT_RENDER_SIZE; | |
2478 | else | |
2479 | ret = GEN8_LR_CONTEXT_RENDER_SIZE; | |
8c857917 OM |
2480 | break; |
2481 | case VCS: | |
2482 | case BCS: | |
2483 | case VECS: | |
2484 | case VCS2: | |
2485 | ret = GEN8_LR_CONTEXT_OTHER_SIZE; | |
2486 | break; | |
2487 | } | |
2488 | ||
2489 | return ret; | |
ede7d42b OM |
2490 | } |
2491 | ||
70b0ea86 | 2492 | static void lrc_setup_hardware_status_page(struct intel_engine_cs *ring, |
1df06b75 TD |
2493 | struct drm_i915_gem_object *default_ctx_obj) |
2494 | { | |
2495 | struct drm_i915_private *dev_priv = ring->dev->dev_private; | |
d1675198 | 2496 | struct page *page; |
1df06b75 | 2497 | |
d1675198 AD |
2498 | /* The HWSP is part of the default context object in LRC mode. */ |
2499 | ring->status_page.gfx_addr = i915_gem_obj_ggtt_offset(default_ctx_obj) | |
2500 | + LRC_PPHWSP_PN * PAGE_SIZE; | |
2501 | page = i915_gem_object_get_page(default_ctx_obj, LRC_PPHWSP_PN); | |
2502 | ring->status_page.page_addr = kmap(page); | |
1df06b75 TD |
2503 | ring->status_page.obj = default_ctx_obj; |
2504 | ||
2505 | I915_WRITE(RING_HWS_PGA(ring->mmio_base), | |
2506 | (u32)ring->status_page.gfx_addr); | |
2507 | POSTING_READ(RING_HWS_PGA(ring->mmio_base)); | |
1df06b75 TD |
2508 | } |
2509 | ||
73e4d07f | 2510 | /** |
e84fe803 | 2511 | * intel_lr_context_deferred_alloc() - create the LRC specific bits of a context |
73e4d07f OM |
2512 | * @ctx: LR context to create. |
2513 | * @ring: engine to be used with the context. | |
2514 | * | |
2515 | * This function can be called more than once, with different engines, if we plan | |
2516 | * to use the context with them. The context backing objects and the ringbuffers | |
2517 | * (specially the ringbuffer backing objects) suck a lot of memory up, and that's why | |
2518 | * the creation is a deferred call: it's better to make sure first that we need to use | |
2519 | * a given ring with the context. | |
2520 | * | |
32197aab | 2521 | * Return: non-zero on error. |
73e4d07f | 2522 | */ |
e84fe803 NH |
2523 | |
2524 | int intel_lr_context_deferred_alloc(struct intel_context *ctx, | |
e28e404c | 2525 | struct intel_engine_cs *ring) |
ede7d42b | 2526 | { |
8c857917 OM |
2527 | struct drm_device *dev = ring->dev; |
2528 | struct drm_i915_gem_object *ctx_obj; | |
2529 | uint32_t context_size; | |
84c2377f | 2530 | struct intel_ringbuffer *ringbuf; |
8c857917 OM |
2531 | int ret; |
2532 | ||
ede7d42b | 2533 | WARN_ON(ctx->legacy_hw_ctx.rcs_state != NULL); |
bfc882b4 | 2534 | WARN_ON(ctx->engine[ring->id].state); |
ede7d42b | 2535 | |
95a66f7e | 2536 | context_size = round_up(intel_lr_context_size(ring), 4096); |
8c857917 | 2537 | |
d1675198 AD |
2538 | /* One extra page as the sharing data between driver and GuC */ |
2539 | context_size += PAGE_SIZE * LRC_PPHWSP_PN; | |
2540 | ||
149c86e7 | 2541 | ctx_obj = i915_gem_alloc_object(dev, context_size); |
3126a660 DC |
2542 | if (!ctx_obj) { |
2543 | DRM_DEBUG_DRIVER("Alloc LRC backing obj failed.\n"); | |
2544 | return -ENOMEM; | |
8c857917 OM |
2545 | } |
2546 | ||
01101fa7 CW |
2547 | ringbuf = intel_engine_create_ringbuffer(ring, 4 * PAGE_SIZE); |
2548 | if (IS_ERR(ringbuf)) { | |
2549 | ret = PTR_ERR(ringbuf); | |
e84fe803 | 2550 | goto error_deref_obj; |
8670d6f9 OM |
2551 | } |
2552 | ||
2553 | ret = populate_lr_context(ctx, ctx_obj, ring, ringbuf); | |
2554 | if (ret) { | |
2555 | DRM_DEBUG_DRIVER("Failed to populate LRC: %d\n", ret); | |
e84fe803 | 2556 | goto error_ringbuf; |
84c2377f OM |
2557 | } |
2558 | ||
2559 | ctx->engine[ring->id].ringbuf = ringbuf; | |
8c857917 | 2560 | ctx->engine[ring->id].state = ctx_obj; |
ede7d42b | 2561 | |
ed54c1a1 | 2562 | if (ctx != ctx->i915->kernel_context && ring->init_context) { |
e84fe803 | 2563 | struct drm_i915_gem_request *req; |
76c39168 | 2564 | |
26827088 DG |
2565 | req = i915_gem_request_alloc(ring, ctx); |
2566 | if (IS_ERR(req)) { | |
2567 | ret = PTR_ERR(req); | |
2568 | DRM_ERROR("ring create req: %d\n", ret); | |
e84fe803 | 2569 | goto error_ringbuf; |
771b9a53 MT |
2570 | } |
2571 | ||
e84fe803 NH |
2572 | ret = ring->init_context(req); |
2573 | if (ret) { | |
2574 | DRM_ERROR("ring init context: %d\n", | |
2575 | ret); | |
2576 | i915_gem_request_cancel(req); | |
2577 | goto error_ringbuf; | |
2578 | } | |
2579 | i915_add_request_no_flush(req); | |
564ddb2f | 2580 | } |
ede7d42b | 2581 | return 0; |
8670d6f9 | 2582 | |
01101fa7 CW |
2583 | error_ringbuf: |
2584 | intel_ringbuffer_free(ringbuf); | |
e84fe803 | 2585 | error_deref_obj: |
8670d6f9 | 2586 | drm_gem_object_unreference(&ctx_obj->base); |
e84fe803 NH |
2587 | ctx->engine[ring->id].ringbuf = NULL; |
2588 | ctx->engine[ring->id].state = NULL; | |
8670d6f9 | 2589 | return ret; |
ede7d42b | 2590 | } |
3e5b6f05 TD |
2591 | |
2592 | void intel_lr_context_reset(struct drm_device *dev, | |
2593 | struct intel_context *ctx) | |
2594 | { | |
2595 | struct drm_i915_private *dev_priv = dev->dev_private; | |
2596 | struct intel_engine_cs *ring; | |
2597 | int i; | |
2598 | ||
2599 | for_each_ring(ring, dev_priv, i) { | |
2600 | struct drm_i915_gem_object *ctx_obj = | |
2601 | ctx->engine[ring->id].state; | |
2602 | struct intel_ringbuffer *ringbuf = | |
2603 | ctx->engine[ring->id].ringbuf; | |
2604 | uint32_t *reg_state; | |
2605 | struct page *page; | |
2606 | ||
2607 | if (!ctx_obj) | |
2608 | continue; | |
2609 | ||
2610 | if (i915_gem_object_get_pages(ctx_obj)) { | |
2611 | WARN(1, "Failed get_pages for context obj\n"); | |
2612 | continue; | |
2613 | } | |
033908ae | 2614 | page = i915_gem_object_get_dirty_page(ctx_obj, LRC_STATE_PN); |
3e5b6f05 TD |
2615 | reg_state = kmap_atomic(page); |
2616 | ||
2617 | reg_state[CTX_RING_HEAD+1] = 0; | |
2618 | reg_state[CTX_RING_TAIL+1] = 0; | |
2619 | ||
2620 | kunmap_atomic(reg_state); | |
2621 | ||
2622 | ringbuf->head = 0; | |
2623 | ringbuf->tail = 0; | |
2624 | } | |
2625 | } |