xref: /linux/drivers/gpu/drm/amd/amdgpu/cik_sdma.c (revision 4b132aacb0768ac1e652cf517097ea6f237214b9)
1 /*
2  * Copyright 2013 Advanced Micro Devices, Inc.
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 shall be included in
12  * all copies or substantial portions of the Software.
13  *
14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20  * OTHER DEALINGS IN THE SOFTWARE.
21  *
22  * Authors: Alex Deucher
23  */
24 
25 #include <linux/firmware.h>
26 #include <linux/module.h>
27 
28 #include "amdgpu.h"
29 #include "amdgpu_ucode.h"
30 #include "amdgpu_trace.h"
31 #include "cikd.h"
32 #include "cik.h"
33 
34 #include "bif/bif_4_1_d.h"
35 #include "bif/bif_4_1_sh_mask.h"
36 
37 #include "gca/gfx_7_2_d.h"
38 #include "gca/gfx_7_2_enum.h"
39 #include "gca/gfx_7_2_sh_mask.h"
40 
41 #include "gmc/gmc_7_1_d.h"
42 #include "gmc/gmc_7_1_sh_mask.h"
43 
44 #include "oss/oss_2_0_d.h"
45 #include "oss/oss_2_0_sh_mask.h"
46 
47 static const u32 sdma_offsets[SDMA_MAX_INSTANCE] =
48 {
49 	SDMA0_REGISTER_OFFSET,
50 	SDMA1_REGISTER_OFFSET
51 };
52 
53 static void cik_sdma_set_ring_funcs(struct amdgpu_device *adev);
54 static void cik_sdma_set_irq_funcs(struct amdgpu_device *adev);
55 static void cik_sdma_set_buffer_funcs(struct amdgpu_device *adev);
56 static void cik_sdma_set_vm_pte_funcs(struct amdgpu_device *adev);
57 static int cik_sdma_soft_reset(void *handle);
58 
59 MODULE_FIRMWARE("amdgpu/bonaire_sdma.bin");
60 MODULE_FIRMWARE("amdgpu/bonaire_sdma1.bin");
61 MODULE_FIRMWARE("amdgpu/hawaii_sdma.bin");
62 MODULE_FIRMWARE("amdgpu/hawaii_sdma1.bin");
63 MODULE_FIRMWARE("amdgpu/kaveri_sdma.bin");
64 MODULE_FIRMWARE("amdgpu/kaveri_sdma1.bin");
65 MODULE_FIRMWARE("amdgpu/kabini_sdma.bin");
66 MODULE_FIRMWARE("amdgpu/kabini_sdma1.bin");
67 MODULE_FIRMWARE("amdgpu/mullins_sdma.bin");
68 MODULE_FIRMWARE("amdgpu/mullins_sdma1.bin");
69 
70 u32 amdgpu_cik_gpu_check_soft_reset(struct amdgpu_device *adev);
71 
72 
73 static void cik_sdma_free_microcode(struct amdgpu_device *adev)
74 {
75 	int i;
76 
77 	for (i = 0; i < adev->sdma.num_instances; i++)
78 		amdgpu_ucode_release(&adev->sdma.instance[i].fw);
79 }
80 
81 /*
82  * sDMA - System DMA
83  * Starting with CIK, the GPU has new asynchronous
84  * DMA engines.  These engines are used for compute
85  * and gfx.  There are two DMA engines (SDMA0, SDMA1)
86  * and each one supports 1 ring buffer used for gfx
87  * and 2 queues used for compute.
88  *
89  * The programming model is very similar to the CP
90  * (ring buffer, IBs, etc.), but sDMA has it's own
91  * packet format that is different from the PM4 format
92  * used by the CP. sDMA supports copying data, writing
93  * embedded data, solid fills, and a number of other
94  * things.  It also has support for tiling/detiling of
95  * buffers.
96  */
97 
98 /**
99  * cik_sdma_init_microcode - load ucode images from disk
100  *
101  * @adev: amdgpu_device pointer
102  *
103  * Use the firmware interface to load the ucode images into
104  * the driver (not loaded into hw).
105  * Returns 0 on success, error on failure.
106  */
107 static int cik_sdma_init_microcode(struct amdgpu_device *adev)
108 {
109 	const char *chip_name;
110 	int err = 0, i;
111 
112 	DRM_DEBUG("\n");
113 
114 	switch (adev->asic_type) {
115 	case CHIP_BONAIRE:
116 		chip_name = "bonaire";
117 		break;
118 	case CHIP_HAWAII:
119 		chip_name = "hawaii";
120 		break;
121 	case CHIP_KAVERI:
122 		chip_name = "kaveri";
123 		break;
124 	case CHIP_KABINI:
125 		chip_name = "kabini";
126 		break;
127 	case CHIP_MULLINS:
128 		chip_name = "mullins";
129 		break;
130 	default: BUG();
131 	}
132 
133 	for (i = 0; i < adev->sdma.num_instances; i++) {
134 		if (i == 0)
135 			err = amdgpu_ucode_request(adev, &adev->sdma.instance[i].fw,
136 						   "amdgpu/%s_sdma.bin", chip_name);
137 		else
138 			err = amdgpu_ucode_request(adev, &adev->sdma.instance[i].fw,
139 						   "amdgpu/%s_sdma1.bin", chip_name);
140 		if (err)
141 			goto out;
142 	}
143 out:
144 	if (err) {
145 		pr_err("cik_sdma: Failed to load firmware \"%s_sdma%s.bin\"\n",
146 		       chip_name, i == 0 ? "" : "1");
147 		for (i = 0; i < adev->sdma.num_instances; i++)
148 			amdgpu_ucode_release(&adev->sdma.instance[i].fw);
149 	}
150 	return err;
151 }
152 
153 /**
154  * cik_sdma_ring_get_rptr - get the current read pointer
155  *
156  * @ring: amdgpu ring pointer
157  *
158  * Get the current rptr from the hardware (CIK+).
159  */
160 static uint64_t cik_sdma_ring_get_rptr(struct amdgpu_ring *ring)
161 {
162 	u32 rptr;
163 
164 	rptr = *ring->rptr_cpu_addr;
165 
166 	return (rptr & 0x3fffc) >> 2;
167 }
168 
169 /**
170  * cik_sdma_ring_get_wptr - get the current write pointer
171  *
172  * @ring: amdgpu ring pointer
173  *
174  * Get the current wptr from the hardware (CIK+).
175  */
176 static uint64_t cik_sdma_ring_get_wptr(struct amdgpu_ring *ring)
177 {
178 	struct amdgpu_device *adev = ring->adev;
179 
180 	return (RREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[ring->me]) & 0x3fffc) >> 2;
181 }
182 
183 /**
184  * cik_sdma_ring_set_wptr - commit the write pointer
185  *
186  * @ring: amdgpu ring pointer
187  *
188  * Write the wptr back to the hardware (CIK+).
189  */
190 static void cik_sdma_ring_set_wptr(struct amdgpu_ring *ring)
191 {
192 	struct amdgpu_device *adev = ring->adev;
193 
194 	WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[ring->me],
195 	       (ring->wptr << 2) & 0x3fffc);
196 }
197 
198 static void cik_sdma_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
199 {
200 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
201 	int i;
202 
203 	for (i = 0; i < count; i++)
204 		if (sdma && sdma->burst_nop && (i == 0))
205 			amdgpu_ring_write(ring, ring->funcs->nop |
206 					  SDMA_NOP_COUNT(count - 1));
207 		else
208 			amdgpu_ring_write(ring, ring->funcs->nop);
209 }
210 
211 /**
212  * cik_sdma_ring_emit_ib - Schedule an IB on the DMA engine
213  *
214  * @ring: amdgpu ring pointer
215  * @job: job to retrive vmid from
216  * @ib: IB object to schedule
217  * @flags: unused
218  *
219  * Schedule an IB in the DMA ring (CIK).
220  */
221 static void cik_sdma_ring_emit_ib(struct amdgpu_ring *ring,
222 				  struct amdgpu_job *job,
223 				  struct amdgpu_ib *ib,
224 				  uint32_t flags)
225 {
226 	unsigned vmid = AMDGPU_JOB_GET_VMID(job);
227 	u32 extra_bits = vmid & 0xf;
228 
229 	/* IB packet must end on a 8 DW boundary */
230 	cik_sdma_ring_insert_nop(ring, (4 - lower_32_bits(ring->wptr)) & 7);
231 
232 	amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_INDIRECT_BUFFER, 0, extra_bits));
233 	amdgpu_ring_write(ring, ib->gpu_addr & 0xffffffe0); /* base must be 32 byte aligned */
234 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr) & 0xffffffff);
235 	amdgpu_ring_write(ring, ib->length_dw);
236 
237 }
238 
239 /**
240  * cik_sdma_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
241  *
242  * @ring: amdgpu ring pointer
243  *
244  * Emit an hdp flush packet on the requested DMA ring.
245  */
246 static void cik_sdma_ring_emit_hdp_flush(struct amdgpu_ring *ring)
247 {
248 	u32 extra_bits = (SDMA_POLL_REG_MEM_EXTRA_OP(1) |
249 			  SDMA_POLL_REG_MEM_EXTRA_FUNC(3)); /* == */
250 	u32 ref_and_mask;
251 
252 	if (ring->me == 0)
253 		ref_and_mask = GPU_HDP_FLUSH_DONE__SDMA0_MASK;
254 	else
255 		ref_and_mask = GPU_HDP_FLUSH_DONE__SDMA1_MASK;
256 
257 	amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0, extra_bits));
258 	amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_DONE << 2);
259 	amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_REQ << 2);
260 	amdgpu_ring_write(ring, ref_and_mask); /* reference */
261 	amdgpu_ring_write(ring, ref_and_mask); /* mask */
262 	amdgpu_ring_write(ring, (0xfff << 16) | 10); /* retry count, poll interval */
263 }
264 
265 /**
266  * cik_sdma_ring_emit_fence - emit a fence on the DMA ring
267  *
268  * @ring: amdgpu ring pointer
269  * @addr: address
270  * @seq: sequence number
271  * @flags: fence related flags
272  *
273  * Add a DMA fence packet to the ring to write
274  * the fence seq number and DMA trap packet to generate
275  * an interrupt if needed (CIK).
276  */
277 static void cik_sdma_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
278 				     unsigned flags)
279 {
280 	bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
281 	/* write the fence */
282 	amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_FENCE, 0, 0));
283 	amdgpu_ring_write(ring, lower_32_bits(addr));
284 	amdgpu_ring_write(ring, upper_32_bits(addr));
285 	amdgpu_ring_write(ring, lower_32_bits(seq));
286 
287 	/* optionally write high bits as well */
288 	if (write64bit) {
289 		addr += 4;
290 		amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_FENCE, 0, 0));
291 		amdgpu_ring_write(ring, lower_32_bits(addr));
292 		amdgpu_ring_write(ring, upper_32_bits(addr));
293 		amdgpu_ring_write(ring, upper_32_bits(seq));
294 	}
295 
296 	/* generate an interrupt */
297 	amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_TRAP, 0, 0));
298 }
299 
300 /**
301  * cik_sdma_gfx_stop - stop the gfx async dma engines
302  *
303  * @adev: amdgpu_device pointer
304  *
305  * Stop the gfx async dma ring buffers (CIK).
306  */
307 static void cik_sdma_gfx_stop(struct amdgpu_device *adev)
308 {
309 	u32 rb_cntl;
310 	int i;
311 
312 	for (i = 0; i < adev->sdma.num_instances; i++) {
313 		rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
314 		rb_cntl &= ~SDMA0_GFX_RB_CNTL__RB_ENABLE_MASK;
315 		WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
316 		WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], 0);
317 	}
318 }
319 
320 /**
321  * cik_sdma_rlc_stop - stop the compute async dma engines
322  *
323  * @adev: amdgpu_device pointer
324  *
325  * Stop the compute async dma queues (CIK).
326  */
327 static void cik_sdma_rlc_stop(struct amdgpu_device *adev)
328 {
329 	/* XXX todo */
330 }
331 
332 /**
333  * cik_ctx_switch_enable - stop the async dma engines context switch
334  *
335  * @adev: amdgpu_device pointer
336  * @enable: enable/disable the DMA MEs context switch.
337  *
338  * Halt or unhalt the async dma engines context switch (VI).
339  */
340 static void cik_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
341 {
342 	u32 f32_cntl, phase_quantum = 0;
343 	int i;
344 
345 	if (amdgpu_sdma_phase_quantum) {
346 		unsigned value = amdgpu_sdma_phase_quantum;
347 		unsigned unit = 0;
348 
349 		while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
350 				SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) {
351 			value = (value + 1) >> 1;
352 			unit++;
353 		}
354 		if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
355 			    SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) {
356 			value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
357 				 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT);
358 			unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
359 				SDMA0_PHASE0_QUANTUM__UNIT__SHIFT);
360 			WARN_ONCE(1,
361 			"clamping sdma_phase_quantum to %uK clock cycles\n",
362 				  value << unit);
363 		}
364 		phase_quantum =
365 			value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
366 			unit  << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
367 	}
368 
369 	for (i = 0; i < adev->sdma.num_instances; i++) {
370 		f32_cntl = RREG32(mmSDMA0_CNTL + sdma_offsets[i]);
371 		if (enable) {
372 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
373 					AUTO_CTXSW_ENABLE, 1);
374 			if (amdgpu_sdma_phase_quantum) {
375 				WREG32(mmSDMA0_PHASE0_QUANTUM + sdma_offsets[i],
376 				       phase_quantum);
377 				WREG32(mmSDMA0_PHASE1_QUANTUM + sdma_offsets[i],
378 				       phase_quantum);
379 			}
380 		} else {
381 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
382 					AUTO_CTXSW_ENABLE, 0);
383 		}
384 
385 		WREG32(mmSDMA0_CNTL + sdma_offsets[i], f32_cntl);
386 	}
387 }
388 
389 /**
390  * cik_sdma_enable - stop the async dma engines
391  *
392  * @adev: amdgpu_device pointer
393  * @enable: enable/disable the DMA MEs.
394  *
395  * Halt or unhalt the async dma engines (CIK).
396  */
397 static void cik_sdma_enable(struct amdgpu_device *adev, bool enable)
398 {
399 	u32 me_cntl;
400 	int i;
401 
402 	if (!enable) {
403 		cik_sdma_gfx_stop(adev);
404 		cik_sdma_rlc_stop(adev);
405 	}
406 
407 	for (i = 0; i < adev->sdma.num_instances; i++) {
408 		me_cntl = RREG32(mmSDMA0_F32_CNTL + sdma_offsets[i]);
409 		if (enable)
410 			me_cntl &= ~SDMA0_F32_CNTL__HALT_MASK;
411 		else
412 			me_cntl |= SDMA0_F32_CNTL__HALT_MASK;
413 		WREG32(mmSDMA0_F32_CNTL + sdma_offsets[i], me_cntl);
414 	}
415 }
416 
417 /**
418  * cik_sdma_gfx_resume - setup and start the async dma engines
419  *
420  * @adev: amdgpu_device pointer
421  *
422  * Set up the gfx DMA ring buffers and enable them (CIK).
423  * Returns 0 for success, error for failure.
424  */
425 static int cik_sdma_gfx_resume(struct amdgpu_device *adev)
426 {
427 	struct amdgpu_ring *ring;
428 	u32 rb_cntl, ib_cntl;
429 	u32 rb_bufsz;
430 	int i, j, r;
431 
432 	for (i = 0; i < adev->sdma.num_instances; i++) {
433 		ring = &adev->sdma.instance[i].ring;
434 
435 		mutex_lock(&adev->srbm_mutex);
436 		for (j = 0; j < 16; j++) {
437 			cik_srbm_select(adev, 0, 0, 0, j);
438 			/* SDMA GFX */
439 			WREG32(mmSDMA0_GFX_VIRTUAL_ADDR + sdma_offsets[i], 0);
440 			WREG32(mmSDMA0_GFX_APE1_CNTL + sdma_offsets[i], 0);
441 			/* XXX SDMA RLC - todo */
442 		}
443 		cik_srbm_select(adev, 0, 0, 0, 0);
444 		mutex_unlock(&adev->srbm_mutex);
445 
446 		WREG32(mmSDMA0_TILING_CONFIG + sdma_offsets[i],
447 		       adev->gfx.config.gb_addr_config & 0x70);
448 
449 		WREG32(mmSDMA0_SEM_INCOMPLETE_TIMER_CNTL + sdma_offsets[i], 0);
450 		WREG32(mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i], 0);
451 
452 		/* Set ring buffer size in dwords */
453 		rb_bufsz = order_base_2(ring->ring_size / 4);
454 		rb_cntl = rb_bufsz << 1;
455 #ifdef __BIG_ENDIAN
456 		rb_cntl |= SDMA0_GFX_RB_CNTL__RB_SWAP_ENABLE_MASK |
457 			SDMA0_GFX_RB_CNTL__RPTR_WRITEBACK_SWAP_ENABLE_MASK;
458 #endif
459 		WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
460 
461 		/* Initialize the ring buffer's read and write pointers */
462 		WREG32(mmSDMA0_GFX_RB_RPTR + sdma_offsets[i], 0);
463 		WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], 0);
464 		WREG32(mmSDMA0_GFX_IB_RPTR + sdma_offsets[i], 0);
465 		WREG32(mmSDMA0_GFX_IB_OFFSET + sdma_offsets[i], 0);
466 
467 		/* set the wb address whether it's enabled or not */
468 		WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_HI + sdma_offsets[i],
469 		       upper_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFF);
470 		WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_LO + sdma_offsets[i],
471 		       ((ring->rptr_gpu_addr) & 0xFFFFFFFC));
472 
473 		rb_cntl |= SDMA0_GFX_RB_CNTL__RPTR_WRITEBACK_ENABLE_MASK;
474 
475 		WREG32(mmSDMA0_GFX_RB_BASE + sdma_offsets[i], ring->gpu_addr >> 8);
476 		WREG32(mmSDMA0_GFX_RB_BASE_HI + sdma_offsets[i], ring->gpu_addr >> 40);
477 
478 		ring->wptr = 0;
479 		WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], ring->wptr << 2);
480 
481 		/* enable DMA RB */
482 		WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i],
483 		       rb_cntl | SDMA0_GFX_RB_CNTL__RB_ENABLE_MASK);
484 
485 		ib_cntl = SDMA0_GFX_IB_CNTL__IB_ENABLE_MASK;
486 #ifdef __BIG_ENDIAN
487 		ib_cntl |= SDMA0_GFX_IB_CNTL__IB_SWAP_ENABLE_MASK;
488 #endif
489 		/* enable DMA IBs */
490 		WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
491 	}
492 
493 	cik_sdma_enable(adev, true);
494 
495 	for (i = 0; i < adev->sdma.num_instances; i++) {
496 		ring = &adev->sdma.instance[i].ring;
497 		r = amdgpu_ring_test_helper(ring);
498 		if (r)
499 			return r;
500 	}
501 
502 	return 0;
503 }
504 
505 /**
506  * cik_sdma_rlc_resume - setup and start the async dma engines
507  *
508  * @adev: amdgpu_device pointer
509  *
510  * Set up the compute DMA queues and enable them (CIK).
511  * Returns 0 for success, error for failure.
512  */
513 static int cik_sdma_rlc_resume(struct amdgpu_device *adev)
514 {
515 	/* XXX todo */
516 	return 0;
517 }
518 
519 /**
520  * cik_sdma_load_microcode - load the sDMA ME ucode
521  *
522  * @adev: amdgpu_device pointer
523  *
524  * Loads the sDMA0/1 ucode.
525  * Returns 0 for success, -EINVAL if the ucode is not available.
526  */
527 static int cik_sdma_load_microcode(struct amdgpu_device *adev)
528 {
529 	const struct sdma_firmware_header_v1_0 *hdr;
530 	const __le32 *fw_data;
531 	u32 fw_size;
532 	int i, j;
533 
534 	/* halt the MEs */
535 	cik_sdma_enable(adev, false);
536 
537 	for (i = 0; i < adev->sdma.num_instances; i++) {
538 		if (!adev->sdma.instance[i].fw)
539 			return -EINVAL;
540 		hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
541 		amdgpu_ucode_print_sdma_hdr(&hdr->header);
542 		fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
543 		adev->sdma.instance[i].fw_version = le32_to_cpu(hdr->header.ucode_version);
544 		adev->sdma.instance[i].feature_version = le32_to_cpu(hdr->ucode_feature_version);
545 		if (adev->sdma.instance[i].feature_version >= 20)
546 			adev->sdma.instance[i].burst_nop = true;
547 		fw_data = (const __le32 *)
548 			(adev->sdma.instance[i].fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
549 		WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], 0);
550 		for (j = 0; j < fw_size; j++)
551 			WREG32(mmSDMA0_UCODE_DATA + sdma_offsets[i], le32_to_cpup(fw_data++));
552 		WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], adev->sdma.instance[i].fw_version);
553 	}
554 
555 	return 0;
556 }
557 
558 /**
559  * cik_sdma_start - setup and start the async dma engines
560  *
561  * @adev: amdgpu_device pointer
562  *
563  * Set up the DMA engines and enable them (CIK).
564  * Returns 0 for success, error for failure.
565  */
566 static int cik_sdma_start(struct amdgpu_device *adev)
567 {
568 	int r;
569 
570 	r = cik_sdma_load_microcode(adev);
571 	if (r)
572 		return r;
573 
574 	/* halt the engine before programing */
575 	cik_sdma_enable(adev, false);
576 	/* enable sdma ring preemption */
577 	cik_ctx_switch_enable(adev, true);
578 
579 	/* start the gfx rings and rlc compute queues */
580 	r = cik_sdma_gfx_resume(adev);
581 	if (r)
582 		return r;
583 	r = cik_sdma_rlc_resume(adev);
584 	if (r)
585 		return r;
586 
587 	return 0;
588 }
589 
590 /**
591  * cik_sdma_ring_test_ring - simple async dma engine test
592  *
593  * @ring: amdgpu_ring structure holding ring information
594  *
595  * Test the DMA engine by writing using it to write an
596  * value to memory. (CIK).
597  * Returns 0 for success, error for failure.
598  */
599 static int cik_sdma_ring_test_ring(struct amdgpu_ring *ring)
600 {
601 	struct amdgpu_device *adev = ring->adev;
602 	unsigned i;
603 	unsigned index;
604 	int r;
605 	u32 tmp;
606 	u64 gpu_addr;
607 
608 	r = amdgpu_device_wb_get(adev, &index);
609 	if (r)
610 		return r;
611 
612 	gpu_addr = adev->wb.gpu_addr + (index * 4);
613 	tmp = 0xCAFEDEAD;
614 	adev->wb.wb[index] = cpu_to_le32(tmp);
615 
616 	r = amdgpu_ring_alloc(ring, 5);
617 	if (r)
618 		goto error_free_wb;
619 
620 	amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_WRITE, SDMA_WRITE_SUB_OPCODE_LINEAR, 0));
621 	amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
622 	amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
623 	amdgpu_ring_write(ring, 1); /* number of DWs to follow */
624 	amdgpu_ring_write(ring, 0xDEADBEEF);
625 	amdgpu_ring_commit(ring);
626 
627 	for (i = 0; i < adev->usec_timeout; i++) {
628 		tmp = le32_to_cpu(adev->wb.wb[index]);
629 		if (tmp == 0xDEADBEEF)
630 			break;
631 		udelay(1);
632 	}
633 
634 	if (i >= adev->usec_timeout)
635 		r = -ETIMEDOUT;
636 
637 error_free_wb:
638 	amdgpu_device_wb_free(adev, index);
639 	return r;
640 }
641 
642 /**
643  * cik_sdma_ring_test_ib - test an IB on the DMA engine
644  *
645  * @ring: amdgpu_ring structure holding ring information
646  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
647  *
648  * Test a simple IB in the DMA ring (CIK).
649  * Returns 0 on success, error on failure.
650  */
651 static int cik_sdma_ring_test_ib(struct amdgpu_ring *ring, long timeout)
652 {
653 	struct amdgpu_device *adev = ring->adev;
654 	struct amdgpu_ib ib;
655 	struct dma_fence *f = NULL;
656 	unsigned index;
657 	u32 tmp = 0;
658 	u64 gpu_addr;
659 	long r;
660 
661 	r = amdgpu_device_wb_get(adev, &index);
662 	if (r)
663 		return r;
664 
665 	gpu_addr = adev->wb.gpu_addr + (index * 4);
666 	tmp = 0xCAFEDEAD;
667 	adev->wb.wb[index] = cpu_to_le32(tmp);
668 	memset(&ib, 0, sizeof(ib));
669 	r = amdgpu_ib_get(adev, NULL, 256,
670 					AMDGPU_IB_POOL_DIRECT, &ib);
671 	if (r)
672 		goto err0;
673 
674 	ib.ptr[0] = SDMA_PACKET(SDMA_OPCODE_WRITE,
675 				SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
676 	ib.ptr[1] = lower_32_bits(gpu_addr);
677 	ib.ptr[2] = upper_32_bits(gpu_addr);
678 	ib.ptr[3] = 1;
679 	ib.ptr[4] = 0xDEADBEEF;
680 	ib.length_dw = 5;
681 	r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
682 	if (r)
683 		goto err1;
684 
685 	r = dma_fence_wait_timeout(f, false, timeout);
686 	if (r == 0) {
687 		r = -ETIMEDOUT;
688 		goto err1;
689 	} else if (r < 0) {
690 		goto err1;
691 	}
692 	tmp = le32_to_cpu(adev->wb.wb[index]);
693 	if (tmp == 0xDEADBEEF)
694 		r = 0;
695 	else
696 		r = -EINVAL;
697 
698 err1:
699 	amdgpu_ib_free(adev, &ib, NULL);
700 	dma_fence_put(f);
701 err0:
702 	amdgpu_device_wb_free(adev, index);
703 	return r;
704 }
705 
706 /**
707  * cik_sdma_vm_copy_pte - update PTEs by copying them from the GART
708  *
709  * @ib: indirect buffer to fill with commands
710  * @pe: addr of the page entry
711  * @src: src addr to copy from
712  * @count: number of page entries to update
713  *
714  * Update PTEs by copying them from the GART using sDMA (CIK).
715  */
716 static void cik_sdma_vm_copy_pte(struct amdgpu_ib *ib,
717 				 uint64_t pe, uint64_t src,
718 				 unsigned count)
719 {
720 	unsigned bytes = count * 8;
721 
722 	ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_COPY,
723 		SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
724 	ib->ptr[ib->length_dw++] = bytes;
725 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
726 	ib->ptr[ib->length_dw++] = lower_32_bits(src);
727 	ib->ptr[ib->length_dw++] = upper_32_bits(src);
728 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
729 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
730 }
731 
732 /**
733  * cik_sdma_vm_write_pte - update PTEs by writing them manually
734  *
735  * @ib: indirect buffer to fill with commands
736  * @pe: addr of the page entry
737  * @value: dst addr to write into pe
738  * @count: number of page entries to update
739  * @incr: increase next addr by incr bytes
740  *
741  * Update PTEs by writing them manually using sDMA (CIK).
742  */
743 static void cik_sdma_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
744 				  uint64_t value, unsigned count,
745 				  uint32_t incr)
746 {
747 	unsigned ndw = count * 2;
748 
749 	ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_WRITE,
750 		SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
751 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
752 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
753 	ib->ptr[ib->length_dw++] = ndw;
754 	for (; ndw > 0; ndw -= 2) {
755 		ib->ptr[ib->length_dw++] = lower_32_bits(value);
756 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
757 		value += incr;
758 	}
759 }
760 
761 /**
762  * cik_sdma_vm_set_pte_pde - update the page tables using sDMA
763  *
764  * @ib: indirect buffer to fill with commands
765  * @pe: addr of the page entry
766  * @addr: dst addr to write into pe
767  * @count: number of page entries to update
768  * @incr: increase next addr by incr bytes
769  * @flags: access flags
770  *
771  * Update the page tables using sDMA (CIK).
772  */
773 static void cik_sdma_vm_set_pte_pde(struct amdgpu_ib *ib, uint64_t pe,
774 				    uint64_t addr, unsigned count,
775 				    uint32_t incr, uint64_t flags)
776 {
777 	/* for physically contiguous pages (vram) */
778 	ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_GENERATE_PTE_PDE, 0, 0);
779 	ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
780 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
781 	ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
782 	ib->ptr[ib->length_dw++] = upper_32_bits(flags);
783 	ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
784 	ib->ptr[ib->length_dw++] = upper_32_bits(addr);
785 	ib->ptr[ib->length_dw++] = incr; /* increment size */
786 	ib->ptr[ib->length_dw++] = 0;
787 	ib->ptr[ib->length_dw++] = count; /* number of entries */
788 }
789 
790 /**
791  * cik_sdma_ring_pad_ib - pad the IB to the required number of dw
792  *
793  * @ring: amdgpu_ring structure holding ring information
794  * @ib: indirect buffer to fill with padding
795  *
796  */
797 static void cik_sdma_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
798 {
799 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
800 	u32 pad_count;
801 	int i;
802 
803 	pad_count = (-ib->length_dw) & 7;
804 	for (i = 0; i < pad_count; i++)
805 		if (sdma && sdma->burst_nop && (i == 0))
806 			ib->ptr[ib->length_dw++] =
807 					SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0) |
808 					SDMA_NOP_COUNT(pad_count - 1);
809 		else
810 			ib->ptr[ib->length_dw++] =
811 					SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0);
812 }
813 
814 /**
815  * cik_sdma_ring_emit_pipeline_sync - sync the pipeline
816  *
817  * @ring: amdgpu_ring pointer
818  *
819  * Make sure all previous operations are completed (CIK).
820  */
821 static void cik_sdma_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
822 {
823 	uint32_t seq = ring->fence_drv.sync_seq;
824 	uint64_t addr = ring->fence_drv.gpu_addr;
825 
826 	/* wait for idle */
827 	amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0,
828 					    SDMA_POLL_REG_MEM_EXTRA_OP(0) |
829 					    SDMA_POLL_REG_MEM_EXTRA_FUNC(3) | /* equal */
830 					    SDMA_POLL_REG_MEM_EXTRA_M));
831 	amdgpu_ring_write(ring, addr & 0xfffffffc);
832 	amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
833 	amdgpu_ring_write(ring, seq); /* reference */
834 	amdgpu_ring_write(ring, 0xffffffff); /* mask */
835 	amdgpu_ring_write(ring, (0xfff << 16) | 4); /* retry count, poll interval */
836 }
837 
838 /**
839  * cik_sdma_ring_emit_vm_flush - cik vm flush using sDMA
840  *
841  * @ring: amdgpu_ring pointer
842  * @vmid: vmid number to use
843  * @pd_addr: address
844  *
845  * Update the page table base and flush the VM TLB
846  * using sDMA (CIK).
847  */
848 static void cik_sdma_ring_emit_vm_flush(struct amdgpu_ring *ring,
849 					unsigned vmid, uint64_t pd_addr)
850 {
851 	u32 extra_bits = (SDMA_POLL_REG_MEM_EXTRA_OP(0) |
852 			  SDMA_POLL_REG_MEM_EXTRA_FUNC(0)); /* always */
853 
854 	amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
855 
856 	amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0, extra_bits));
857 	amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST << 2);
858 	amdgpu_ring_write(ring, 0);
859 	amdgpu_ring_write(ring, 0); /* reference */
860 	amdgpu_ring_write(ring, 0); /* mask */
861 	amdgpu_ring_write(ring, (0xfff << 16) | 10); /* retry count, poll interval */
862 }
863 
864 static void cik_sdma_ring_emit_wreg(struct amdgpu_ring *ring,
865 				    uint32_t reg, uint32_t val)
866 {
867 	amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
868 	amdgpu_ring_write(ring, reg);
869 	amdgpu_ring_write(ring, val);
870 }
871 
872 static void cik_enable_sdma_mgcg(struct amdgpu_device *adev,
873 				 bool enable)
874 {
875 	u32 orig, data;
876 
877 	if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
878 		WREG32(mmSDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET, 0x00000100);
879 		WREG32(mmSDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET, 0x00000100);
880 	} else {
881 		orig = data = RREG32(mmSDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET);
882 		data |= 0xff000000;
883 		if (data != orig)
884 			WREG32(mmSDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET, data);
885 
886 		orig = data = RREG32(mmSDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET);
887 		data |= 0xff000000;
888 		if (data != orig)
889 			WREG32(mmSDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET, data);
890 	}
891 }
892 
893 static void cik_enable_sdma_mgls(struct amdgpu_device *adev,
894 				 bool enable)
895 {
896 	u32 orig, data;
897 
898 	if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) {
899 		orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET);
900 		data |= 0x100;
901 		if (orig != data)
902 			WREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET, data);
903 
904 		orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET);
905 		data |= 0x100;
906 		if (orig != data)
907 			WREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET, data);
908 	} else {
909 		orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET);
910 		data &= ~0x100;
911 		if (orig != data)
912 			WREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET, data);
913 
914 		orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET);
915 		data &= ~0x100;
916 		if (orig != data)
917 			WREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET, data);
918 	}
919 }
920 
921 static int cik_sdma_early_init(void *handle)
922 {
923 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
924 	int r;
925 
926 	adev->sdma.num_instances = SDMA_MAX_INSTANCE;
927 
928 	r = cik_sdma_init_microcode(adev);
929 	if (r)
930 		return r;
931 
932 	cik_sdma_set_ring_funcs(adev);
933 	cik_sdma_set_irq_funcs(adev);
934 	cik_sdma_set_buffer_funcs(adev);
935 	cik_sdma_set_vm_pte_funcs(adev);
936 
937 	return 0;
938 }
939 
940 static int cik_sdma_sw_init(void *handle)
941 {
942 	struct amdgpu_ring *ring;
943 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
944 	int r, i;
945 
946 	/* SDMA trap event */
947 	r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 224,
948 			      &adev->sdma.trap_irq);
949 	if (r)
950 		return r;
951 
952 	/* SDMA Privileged inst */
953 	r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 241,
954 			      &adev->sdma.illegal_inst_irq);
955 	if (r)
956 		return r;
957 
958 	/* SDMA Privileged inst */
959 	r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 247,
960 			      &adev->sdma.illegal_inst_irq);
961 	if (r)
962 		return r;
963 
964 	for (i = 0; i < adev->sdma.num_instances; i++) {
965 		ring = &adev->sdma.instance[i].ring;
966 		ring->ring_obj = NULL;
967 		sprintf(ring->name, "sdma%d", i);
968 		r = amdgpu_ring_init(adev, ring, 1024,
969 				     &adev->sdma.trap_irq,
970 				     (i == 0) ? AMDGPU_SDMA_IRQ_INSTANCE0 :
971 				     AMDGPU_SDMA_IRQ_INSTANCE1,
972 				     AMDGPU_RING_PRIO_DEFAULT, NULL);
973 		if (r)
974 			return r;
975 	}
976 
977 	return r;
978 }
979 
980 static int cik_sdma_sw_fini(void *handle)
981 {
982 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
983 	int i;
984 
985 	for (i = 0; i < adev->sdma.num_instances; i++)
986 		amdgpu_ring_fini(&adev->sdma.instance[i].ring);
987 
988 	cik_sdma_free_microcode(adev);
989 	return 0;
990 }
991 
992 static int cik_sdma_hw_init(void *handle)
993 {
994 	int r;
995 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
996 
997 	r = cik_sdma_start(adev);
998 	if (r)
999 		return r;
1000 
1001 	return r;
1002 }
1003 
1004 static int cik_sdma_hw_fini(void *handle)
1005 {
1006 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1007 
1008 	cik_ctx_switch_enable(adev, false);
1009 	cik_sdma_enable(adev, false);
1010 
1011 	return 0;
1012 }
1013 
1014 static int cik_sdma_suspend(void *handle)
1015 {
1016 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1017 
1018 	return cik_sdma_hw_fini(adev);
1019 }
1020 
1021 static int cik_sdma_resume(void *handle)
1022 {
1023 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1024 
1025 	cik_sdma_soft_reset(handle);
1026 
1027 	return cik_sdma_hw_init(adev);
1028 }
1029 
1030 static bool cik_sdma_is_idle(void *handle)
1031 {
1032 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1033 	u32 tmp = RREG32(mmSRBM_STATUS2);
1034 
1035 	if (tmp & (SRBM_STATUS2__SDMA_BUSY_MASK |
1036 				SRBM_STATUS2__SDMA1_BUSY_MASK))
1037 	    return false;
1038 
1039 	return true;
1040 }
1041 
1042 static int cik_sdma_wait_for_idle(void *handle)
1043 {
1044 	unsigned i;
1045 	u32 tmp;
1046 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1047 
1048 	for (i = 0; i < adev->usec_timeout; i++) {
1049 		tmp = RREG32(mmSRBM_STATUS2) & (SRBM_STATUS2__SDMA_BUSY_MASK |
1050 				SRBM_STATUS2__SDMA1_BUSY_MASK);
1051 
1052 		if (!tmp)
1053 			return 0;
1054 		udelay(1);
1055 	}
1056 	return -ETIMEDOUT;
1057 }
1058 
1059 static int cik_sdma_soft_reset(void *handle)
1060 {
1061 	u32 srbm_soft_reset = 0;
1062 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1063 	u32 tmp;
1064 
1065 	/* sdma0 */
1066 	tmp = RREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET);
1067 	tmp |= SDMA0_F32_CNTL__HALT_MASK;
1068 	WREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET, tmp);
1069 	srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA_MASK;
1070 
1071 	/* sdma1 */
1072 	tmp = RREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET);
1073 	tmp |= SDMA0_F32_CNTL__HALT_MASK;
1074 	WREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET, tmp);
1075 	srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA1_MASK;
1076 
1077 	if (srbm_soft_reset) {
1078 		tmp = RREG32(mmSRBM_SOFT_RESET);
1079 		tmp |= srbm_soft_reset;
1080 		dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
1081 		WREG32(mmSRBM_SOFT_RESET, tmp);
1082 		tmp = RREG32(mmSRBM_SOFT_RESET);
1083 
1084 		udelay(50);
1085 
1086 		tmp &= ~srbm_soft_reset;
1087 		WREG32(mmSRBM_SOFT_RESET, tmp);
1088 		tmp = RREG32(mmSRBM_SOFT_RESET);
1089 
1090 		/* Wait a little for things to settle down */
1091 		udelay(50);
1092 	}
1093 
1094 	return 0;
1095 }
1096 
1097 static int cik_sdma_set_trap_irq_state(struct amdgpu_device *adev,
1098 				       struct amdgpu_irq_src *src,
1099 				       unsigned type,
1100 				       enum amdgpu_interrupt_state state)
1101 {
1102 	u32 sdma_cntl;
1103 
1104 	switch (type) {
1105 	case AMDGPU_SDMA_IRQ_INSTANCE0:
1106 		switch (state) {
1107 		case AMDGPU_IRQ_STATE_DISABLE:
1108 			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1109 			sdma_cntl &= ~SDMA0_CNTL__TRAP_ENABLE_MASK;
1110 			WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1111 			break;
1112 		case AMDGPU_IRQ_STATE_ENABLE:
1113 			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1114 			sdma_cntl |= SDMA0_CNTL__TRAP_ENABLE_MASK;
1115 			WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1116 			break;
1117 		default:
1118 			break;
1119 		}
1120 		break;
1121 	case AMDGPU_SDMA_IRQ_INSTANCE1:
1122 		switch (state) {
1123 		case AMDGPU_IRQ_STATE_DISABLE:
1124 			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1125 			sdma_cntl &= ~SDMA0_CNTL__TRAP_ENABLE_MASK;
1126 			WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1127 			break;
1128 		case AMDGPU_IRQ_STATE_ENABLE:
1129 			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1130 			sdma_cntl |= SDMA0_CNTL__TRAP_ENABLE_MASK;
1131 			WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1132 			break;
1133 		default:
1134 			break;
1135 		}
1136 		break;
1137 	default:
1138 		break;
1139 	}
1140 	return 0;
1141 }
1142 
1143 static int cik_sdma_process_trap_irq(struct amdgpu_device *adev,
1144 				     struct amdgpu_irq_src *source,
1145 				     struct amdgpu_iv_entry *entry)
1146 {
1147 	u8 instance_id, queue_id;
1148 
1149 	instance_id = (entry->ring_id & 0x3) >> 0;
1150 	queue_id = (entry->ring_id & 0xc) >> 2;
1151 	DRM_DEBUG("IH: SDMA trap\n");
1152 	switch (instance_id) {
1153 	case 0:
1154 		switch (queue_id) {
1155 		case 0:
1156 			amdgpu_fence_process(&adev->sdma.instance[0].ring);
1157 			break;
1158 		case 1:
1159 			/* XXX compute */
1160 			break;
1161 		case 2:
1162 			/* XXX compute */
1163 			break;
1164 		}
1165 		break;
1166 	case 1:
1167 		switch (queue_id) {
1168 		case 0:
1169 			amdgpu_fence_process(&adev->sdma.instance[1].ring);
1170 			break;
1171 		case 1:
1172 			/* XXX compute */
1173 			break;
1174 		case 2:
1175 			/* XXX compute */
1176 			break;
1177 		}
1178 		break;
1179 	}
1180 
1181 	return 0;
1182 }
1183 
1184 static int cik_sdma_process_illegal_inst_irq(struct amdgpu_device *adev,
1185 					     struct amdgpu_irq_src *source,
1186 					     struct amdgpu_iv_entry *entry)
1187 {
1188 	u8 instance_id;
1189 
1190 	DRM_ERROR("Illegal instruction in SDMA command stream\n");
1191 	instance_id = (entry->ring_id & 0x3) >> 0;
1192 	drm_sched_fault(&adev->sdma.instance[instance_id].ring.sched);
1193 	return 0;
1194 }
1195 
1196 static int cik_sdma_set_clockgating_state(void *handle,
1197 					  enum amd_clockgating_state state)
1198 {
1199 	bool gate = false;
1200 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1201 
1202 	if (state == AMD_CG_STATE_GATE)
1203 		gate = true;
1204 
1205 	cik_enable_sdma_mgcg(adev, gate);
1206 	cik_enable_sdma_mgls(adev, gate);
1207 
1208 	return 0;
1209 }
1210 
1211 static int cik_sdma_set_powergating_state(void *handle,
1212 					  enum amd_powergating_state state)
1213 {
1214 	return 0;
1215 }
1216 
1217 static const struct amd_ip_funcs cik_sdma_ip_funcs = {
1218 	.name = "cik_sdma",
1219 	.early_init = cik_sdma_early_init,
1220 	.late_init = NULL,
1221 	.sw_init = cik_sdma_sw_init,
1222 	.sw_fini = cik_sdma_sw_fini,
1223 	.hw_init = cik_sdma_hw_init,
1224 	.hw_fini = cik_sdma_hw_fini,
1225 	.suspend = cik_sdma_suspend,
1226 	.resume = cik_sdma_resume,
1227 	.is_idle = cik_sdma_is_idle,
1228 	.wait_for_idle = cik_sdma_wait_for_idle,
1229 	.soft_reset = cik_sdma_soft_reset,
1230 	.set_clockgating_state = cik_sdma_set_clockgating_state,
1231 	.set_powergating_state = cik_sdma_set_powergating_state,
1232 	.dump_ip_state = NULL,
1233 	.print_ip_state = NULL,
1234 };
1235 
1236 static const struct amdgpu_ring_funcs cik_sdma_ring_funcs = {
1237 	.type = AMDGPU_RING_TYPE_SDMA,
1238 	.align_mask = 0xf,
1239 	.nop = SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0),
1240 	.support_64bit_ptrs = false,
1241 	.get_rptr = cik_sdma_ring_get_rptr,
1242 	.get_wptr = cik_sdma_ring_get_wptr,
1243 	.set_wptr = cik_sdma_ring_set_wptr,
1244 	.emit_frame_size =
1245 		6 + /* cik_sdma_ring_emit_hdp_flush */
1246 		3 + /* hdp invalidate */
1247 		6 + /* cik_sdma_ring_emit_pipeline_sync */
1248 		CIK_FLUSH_GPU_TLB_NUM_WREG * 3 + 6 + /* cik_sdma_ring_emit_vm_flush */
1249 		9 + 9 + 9, /* cik_sdma_ring_emit_fence x3 for user fence, vm fence */
1250 	.emit_ib_size = 7 + 4, /* cik_sdma_ring_emit_ib */
1251 	.emit_ib = cik_sdma_ring_emit_ib,
1252 	.emit_fence = cik_sdma_ring_emit_fence,
1253 	.emit_pipeline_sync = cik_sdma_ring_emit_pipeline_sync,
1254 	.emit_vm_flush = cik_sdma_ring_emit_vm_flush,
1255 	.emit_hdp_flush = cik_sdma_ring_emit_hdp_flush,
1256 	.test_ring = cik_sdma_ring_test_ring,
1257 	.test_ib = cik_sdma_ring_test_ib,
1258 	.insert_nop = cik_sdma_ring_insert_nop,
1259 	.pad_ib = cik_sdma_ring_pad_ib,
1260 	.emit_wreg = cik_sdma_ring_emit_wreg,
1261 };
1262 
1263 static void cik_sdma_set_ring_funcs(struct amdgpu_device *adev)
1264 {
1265 	int i;
1266 
1267 	for (i = 0; i < adev->sdma.num_instances; i++) {
1268 		adev->sdma.instance[i].ring.funcs = &cik_sdma_ring_funcs;
1269 		adev->sdma.instance[i].ring.me = i;
1270 	}
1271 }
1272 
1273 static const struct amdgpu_irq_src_funcs cik_sdma_trap_irq_funcs = {
1274 	.set = cik_sdma_set_trap_irq_state,
1275 	.process = cik_sdma_process_trap_irq,
1276 };
1277 
1278 static const struct amdgpu_irq_src_funcs cik_sdma_illegal_inst_irq_funcs = {
1279 	.process = cik_sdma_process_illegal_inst_irq,
1280 };
1281 
1282 static void cik_sdma_set_irq_funcs(struct amdgpu_device *adev)
1283 {
1284 	adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST;
1285 	adev->sdma.trap_irq.funcs = &cik_sdma_trap_irq_funcs;
1286 	adev->sdma.illegal_inst_irq.funcs = &cik_sdma_illegal_inst_irq_funcs;
1287 }
1288 
1289 /**
1290  * cik_sdma_emit_copy_buffer - copy buffer using the sDMA engine
1291  *
1292  * @ib: indirect buffer to copy to
1293  * @src_offset: src GPU address
1294  * @dst_offset: dst GPU address
1295  * @byte_count: number of bytes to xfer
1296  * @copy_flags: unused
1297  *
1298  * Copy GPU buffers using the DMA engine (CIK).
1299  * Used by the amdgpu ttm implementation to move pages if
1300  * registered as the asic copy callback.
1301  */
1302 static void cik_sdma_emit_copy_buffer(struct amdgpu_ib *ib,
1303 				      uint64_t src_offset,
1304 				      uint64_t dst_offset,
1305 				      uint32_t byte_count,
1306 				      uint32_t copy_flags)
1307 {
1308 	ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_COPY, SDMA_COPY_SUB_OPCODE_LINEAR, 0);
1309 	ib->ptr[ib->length_dw++] = byte_count;
1310 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1311 	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1312 	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1313 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1314 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1315 }
1316 
1317 /**
1318  * cik_sdma_emit_fill_buffer - fill buffer using the sDMA engine
1319  *
1320  * @ib: indirect buffer to fill
1321  * @src_data: value to write to buffer
1322  * @dst_offset: dst GPU address
1323  * @byte_count: number of bytes to xfer
1324  *
1325  * Fill GPU buffers using the DMA engine (CIK).
1326  */
1327 static void cik_sdma_emit_fill_buffer(struct amdgpu_ib *ib,
1328 				      uint32_t src_data,
1329 				      uint64_t dst_offset,
1330 				      uint32_t byte_count)
1331 {
1332 	ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_CONSTANT_FILL, 0, 0);
1333 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1334 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1335 	ib->ptr[ib->length_dw++] = src_data;
1336 	ib->ptr[ib->length_dw++] = byte_count;
1337 }
1338 
1339 static const struct amdgpu_buffer_funcs cik_sdma_buffer_funcs = {
1340 	.copy_max_bytes = 0x1fffff,
1341 	.copy_num_dw = 7,
1342 	.emit_copy_buffer = cik_sdma_emit_copy_buffer,
1343 
1344 	.fill_max_bytes = 0x1fffff,
1345 	.fill_num_dw = 5,
1346 	.emit_fill_buffer = cik_sdma_emit_fill_buffer,
1347 };
1348 
1349 static void cik_sdma_set_buffer_funcs(struct amdgpu_device *adev)
1350 {
1351 	adev->mman.buffer_funcs = &cik_sdma_buffer_funcs;
1352 	adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1353 }
1354 
1355 static const struct amdgpu_vm_pte_funcs cik_sdma_vm_pte_funcs = {
1356 	.copy_pte_num_dw = 7,
1357 	.copy_pte = cik_sdma_vm_copy_pte,
1358 
1359 	.write_pte = cik_sdma_vm_write_pte,
1360 	.set_pte_pde = cik_sdma_vm_set_pte_pde,
1361 };
1362 
1363 static void cik_sdma_set_vm_pte_funcs(struct amdgpu_device *adev)
1364 {
1365 	unsigned i;
1366 
1367 	adev->vm_manager.vm_pte_funcs = &cik_sdma_vm_pte_funcs;
1368 	for (i = 0; i < adev->sdma.num_instances; i++) {
1369 		adev->vm_manager.vm_pte_scheds[i] =
1370 			&adev->sdma.instance[i].ring.sched;
1371 	}
1372 	adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
1373 }
1374 
1375 const struct amdgpu_ip_block_version cik_sdma_ip_block =
1376 {
1377 	.type = AMD_IP_BLOCK_TYPE_SDMA,
1378 	.major = 2,
1379 	.minor = 0,
1380 	.rev = 0,
1381 	.funcs = &cik_sdma_ip_funcs,
1382 };
1383