xref: /openbmc/linux/drivers/gpu/drm/amd/amdgpu/cik_sdma.c (revision c4e78957)
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 	for (i = 0; i < adev->sdma.num_instances; i++) {
77 			release_firmware(adev->sdma.instance[i].fw);
78 			adev->sdma.instance[i].fw = NULL;
79 	}
80 }
81 
82 /*
83  * sDMA - System DMA
84  * Starting with CIK, the GPU has new asynchronous
85  * DMA engines.  These engines are used for compute
86  * and gfx.  There are two DMA engines (SDMA0, SDMA1)
87  * and each one supports 1 ring buffer used for gfx
88  * and 2 queues used for compute.
89  *
90  * The programming model is very similar to the CP
91  * (ring buffer, IBs, etc.), but sDMA has it's own
92  * packet format that is different from the PM4 format
93  * used by the CP. sDMA supports copying data, writing
94  * embedded data, solid fills, and a number of other
95  * things.  It also has support for tiling/detiling of
96  * buffers.
97  */
98 
99 /**
100  * cik_sdma_init_microcode - load ucode images from disk
101  *
102  * @adev: amdgpu_device pointer
103  *
104  * Use the firmware interface to load the ucode images into
105  * the driver (not loaded into hw).
106  * Returns 0 on success, error on failure.
107  */
108 static int cik_sdma_init_microcode(struct amdgpu_device *adev)
109 {
110 	const char *chip_name;
111 	char fw_name[30];
112 	int err = 0, i;
113 
114 	DRM_DEBUG("\n");
115 
116 	switch (adev->asic_type) {
117 	case CHIP_BONAIRE:
118 		chip_name = "bonaire";
119 		break;
120 	case CHIP_HAWAII:
121 		chip_name = "hawaii";
122 		break;
123 	case CHIP_KAVERI:
124 		chip_name = "kaveri";
125 		break;
126 	case CHIP_KABINI:
127 		chip_name = "kabini";
128 		break;
129 	case CHIP_MULLINS:
130 		chip_name = "mullins";
131 		break;
132 	default: BUG();
133 	}
134 
135 	for (i = 0; i < adev->sdma.num_instances; i++) {
136 		if (i == 0)
137 			snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma.bin", chip_name);
138 		else
139 			snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma1.bin", chip_name);
140 		err = request_firmware(&adev->sdma.instance[i].fw, fw_name, adev->dev);
141 		if (err)
142 			goto out;
143 		err = amdgpu_ucode_validate(adev->sdma.instance[i].fw);
144 	}
145 out:
146 	if (err) {
147 		pr_err("cik_sdma: Failed to load firmware \"%s\"\n", fw_name);
148 		for (i = 0; i < adev->sdma.num_instances; i++) {
149 			release_firmware(adev->sdma.instance[i].fw);
150 			adev->sdma.instance[i].fw = NULL;
151 		}
152 	}
153 	return err;
154 }
155 
156 /**
157  * cik_sdma_ring_get_rptr - get the current read pointer
158  *
159  * @ring: amdgpu ring pointer
160  *
161  * Get the current rptr from the hardware (CIK+).
162  */
163 static uint64_t cik_sdma_ring_get_rptr(struct amdgpu_ring *ring)
164 {
165 	u32 rptr;
166 
167 	rptr = *ring->rptr_cpu_addr;
168 
169 	return (rptr & 0x3fffc) >> 2;
170 }
171 
172 /**
173  * cik_sdma_ring_get_wptr - get the current write pointer
174  *
175  * @ring: amdgpu ring pointer
176  *
177  * Get the current wptr from the hardware (CIK+).
178  */
179 static uint64_t cik_sdma_ring_get_wptr(struct amdgpu_ring *ring)
180 {
181 	struct amdgpu_device *adev = ring->adev;
182 
183 	return (RREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[ring->me]) & 0x3fffc) >> 2;
184 }
185 
186 /**
187  * cik_sdma_ring_set_wptr - commit the write pointer
188  *
189  * @ring: amdgpu ring pointer
190  *
191  * Write the wptr back to the hardware (CIK+).
192  */
193 static void cik_sdma_ring_set_wptr(struct amdgpu_ring *ring)
194 {
195 	struct amdgpu_device *adev = ring->adev;
196 
197 	WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[ring->me],
198 	       (ring->wptr << 2) & 0x3fffc);
199 }
200 
201 static void cik_sdma_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
202 {
203 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
204 	int i;
205 
206 	for (i = 0; i < count; i++)
207 		if (sdma && sdma->burst_nop && (i == 0))
208 			amdgpu_ring_write(ring, ring->funcs->nop |
209 					  SDMA_NOP_COUNT(count - 1));
210 		else
211 			amdgpu_ring_write(ring, ring->funcs->nop);
212 }
213 
214 /**
215  * cik_sdma_ring_emit_ib - Schedule an IB on the DMA engine
216  *
217  * @ring: amdgpu ring pointer
218  * @job: job to retrive vmid from
219  * @ib: IB object to schedule
220  * @flags: unused
221  *
222  * Schedule an IB in the DMA ring (CIK).
223  */
224 static void cik_sdma_ring_emit_ib(struct amdgpu_ring *ring,
225 				  struct amdgpu_job *job,
226 				  struct amdgpu_ib *ib,
227 				  uint32_t flags)
228 {
229 	unsigned vmid = AMDGPU_JOB_GET_VMID(job);
230 	u32 extra_bits = vmid & 0xf;
231 
232 	/* IB packet must end on a 8 DW boundary */
233 	cik_sdma_ring_insert_nop(ring, (4 - lower_32_bits(ring->wptr)) & 7);
234 
235 	amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_INDIRECT_BUFFER, 0, extra_bits));
236 	amdgpu_ring_write(ring, ib->gpu_addr & 0xffffffe0); /* base must be 32 byte aligned */
237 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr) & 0xffffffff);
238 	amdgpu_ring_write(ring, ib->length_dw);
239 
240 }
241 
242 /**
243  * cik_sdma_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
244  *
245  * @ring: amdgpu ring pointer
246  *
247  * Emit an hdp flush packet on the requested DMA ring.
248  */
249 static void cik_sdma_ring_emit_hdp_flush(struct amdgpu_ring *ring)
250 {
251 	u32 extra_bits = (SDMA_POLL_REG_MEM_EXTRA_OP(1) |
252 			  SDMA_POLL_REG_MEM_EXTRA_FUNC(3)); /* == */
253 	u32 ref_and_mask;
254 
255 	if (ring->me == 0)
256 		ref_and_mask = GPU_HDP_FLUSH_DONE__SDMA0_MASK;
257 	else
258 		ref_and_mask = GPU_HDP_FLUSH_DONE__SDMA1_MASK;
259 
260 	amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0, extra_bits));
261 	amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_DONE << 2);
262 	amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_REQ << 2);
263 	amdgpu_ring_write(ring, ref_and_mask); /* reference */
264 	amdgpu_ring_write(ring, ref_and_mask); /* mask */
265 	amdgpu_ring_write(ring, (0xfff << 16) | 10); /* retry count, poll interval */
266 }
267 
268 /**
269  * cik_sdma_ring_emit_fence - emit a fence on the DMA ring
270  *
271  * @ring: amdgpu ring pointer
272  * @addr: address
273  * @seq: sequence number
274  * @flags: fence related flags
275  *
276  * Add a DMA fence packet to the ring to write
277  * the fence seq number and DMA trap packet to generate
278  * an interrupt if needed (CIK).
279  */
280 static void cik_sdma_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
281 				     unsigned flags)
282 {
283 	bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
284 	/* write the fence */
285 	amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_FENCE, 0, 0));
286 	amdgpu_ring_write(ring, lower_32_bits(addr));
287 	amdgpu_ring_write(ring, upper_32_bits(addr));
288 	amdgpu_ring_write(ring, lower_32_bits(seq));
289 
290 	/* optionally write high bits as well */
291 	if (write64bit) {
292 		addr += 4;
293 		amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_FENCE, 0, 0));
294 		amdgpu_ring_write(ring, lower_32_bits(addr));
295 		amdgpu_ring_write(ring, upper_32_bits(addr));
296 		amdgpu_ring_write(ring, upper_32_bits(seq));
297 	}
298 
299 	/* generate an interrupt */
300 	amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_TRAP, 0, 0));
301 }
302 
303 /**
304  * cik_sdma_gfx_stop - stop the gfx async dma engines
305  *
306  * @adev: amdgpu_device pointer
307  *
308  * Stop the gfx async dma ring buffers (CIK).
309  */
310 static void cik_sdma_gfx_stop(struct amdgpu_device *adev)
311 {
312 	struct amdgpu_ring *sdma0 = &adev->sdma.instance[0].ring;
313 	struct amdgpu_ring *sdma1 = &adev->sdma.instance[1].ring;
314 	u32 rb_cntl;
315 	int i;
316 
317 	if ((adev->mman.buffer_funcs_ring == sdma0) ||
318 	    (adev->mman.buffer_funcs_ring == sdma1))
319 			amdgpu_ttm_set_buffer_funcs_status(adev, false);
320 
321 	for (i = 0; i < adev->sdma.num_instances; i++) {
322 		rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
323 		rb_cntl &= ~SDMA0_GFX_RB_CNTL__RB_ENABLE_MASK;
324 		WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
325 		WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], 0);
326 	}
327 }
328 
329 /**
330  * cik_sdma_rlc_stop - stop the compute async dma engines
331  *
332  * @adev: amdgpu_device pointer
333  *
334  * Stop the compute async dma queues (CIK).
335  */
336 static void cik_sdma_rlc_stop(struct amdgpu_device *adev)
337 {
338 	/* XXX todo */
339 }
340 
341 /**
342  * cik_ctx_switch_enable - stop the async dma engines context switch
343  *
344  * @adev: amdgpu_device pointer
345  * @enable: enable/disable the DMA MEs context switch.
346  *
347  * Halt or unhalt the async dma engines context switch (VI).
348  */
349 static void cik_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
350 {
351 	u32 f32_cntl, phase_quantum = 0;
352 	int i;
353 
354 	if (amdgpu_sdma_phase_quantum) {
355 		unsigned value = amdgpu_sdma_phase_quantum;
356 		unsigned unit = 0;
357 
358 		while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
359 				SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) {
360 			value = (value + 1) >> 1;
361 			unit++;
362 		}
363 		if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
364 			    SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) {
365 			value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
366 				 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT);
367 			unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
368 				SDMA0_PHASE0_QUANTUM__UNIT__SHIFT);
369 			WARN_ONCE(1,
370 			"clamping sdma_phase_quantum to %uK clock cycles\n",
371 				  value << unit);
372 		}
373 		phase_quantum =
374 			value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
375 			unit  << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
376 	}
377 
378 	for (i = 0; i < adev->sdma.num_instances; i++) {
379 		f32_cntl = RREG32(mmSDMA0_CNTL + sdma_offsets[i]);
380 		if (enable) {
381 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
382 					AUTO_CTXSW_ENABLE, 1);
383 			if (amdgpu_sdma_phase_quantum) {
384 				WREG32(mmSDMA0_PHASE0_QUANTUM + sdma_offsets[i],
385 				       phase_quantum);
386 				WREG32(mmSDMA0_PHASE1_QUANTUM + sdma_offsets[i],
387 				       phase_quantum);
388 			}
389 		} else {
390 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
391 					AUTO_CTXSW_ENABLE, 0);
392 		}
393 
394 		WREG32(mmSDMA0_CNTL + sdma_offsets[i], f32_cntl);
395 	}
396 }
397 
398 /**
399  * cik_sdma_enable - stop the async dma engines
400  *
401  * @adev: amdgpu_device pointer
402  * @enable: enable/disable the DMA MEs.
403  *
404  * Halt or unhalt the async dma engines (CIK).
405  */
406 static void cik_sdma_enable(struct amdgpu_device *adev, bool enable)
407 {
408 	u32 me_cntl;
409 	int i;
410 
411 	if (!enable) {
412 		cik_sdma_gfx_stop(adev);
413 		cik_sdma_rlc_stop(adev);
414 	}
415 
416 	for (i = 0; i < adev->sdma.num_instances; i++) {
417 		me_cntl = RREG32(mmSDMA0_F32_CNTL + sdma_offsets[i]);
418 		if (enable)
419 			me_cntl &= ~SDMA0_F32_CNTL__HALT_MASK;
420 		else
421 			me_cntl |= SDMA0_F32_CNTL__HALT_MASK;
422 		WREG32(mmSDMA0_F32_CNTL + sdma_offsets[i], me_cntl);
423 	}
424 }
425 
426 /**
427  * cik_sdma_gfx_resume - setup and start the async dma engines
428  *
429  * @adev: amdgpu_device pointer
430  *
431  * Set up the gfx DMA ring buffers and enable them (CIK).
432  * Returns 0 for success, error for failure.
433  */
434 static int cik_sdma_gfx_resume(struct amdgpu_device *adev)
435 {
436 	struct amdgpu_ring *ring;
437 	u32 rb_cntl, ib_cntl;
438 	u32 rb_bufsz;
439 	int i, j, r;
440 
441 	for (i = 0; i < adev->sdma.num_instances; i++) {
442 		ring = &adev->sdma.instance[i].ring;
443 
444 		mutex_lock(&adev->srbm_mutex);
445 		for (j = 0; j < 16; j++) {
446 			cik_srbm_select(adev, 0, 0, 0, j);
447 			/* SDMA GFX */
448 			WREG32(mmSDMA0_GFX_VIRTUAL_ADDR + sdma_offsets[i], 0);
449 			WREG32(mmSDMA0_GFX_APE1_CNTL + sdma_offsets[i], 0);
450 			/* XXX SDMA RLC - todo */
451 		}
452 		cik_srbm_select(adev, 0, 0, 0, 0);
453 		mutex_unlock(&adev->srbm_mutex);
454 
455 		WREG32(mmSDMA0_TILING_CONFIG + sdma_offsets[i],
456 		       adev->gfx.config.gb_addr_config & 0x70);
457 
458 		WREG32(mmSDMA0_SEM_INCOMPLETE_TIMER_CNTL + sdma_offsets[i], 0);
459 		WREG32(mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i], 0);
460 
461 		/* Set ring buffer size in dwords */
462 		rb_bufsz = order_base_2(ring->ring_size / 4);
463 		rb_cntl = rb_bufsz << 1;
464 #ifdef __BIG_ENDIAN
465 		rb_cntl |= SDMA0_GFX_RB_CNTL__RB_SWAP_ENABLE_MASK |
466 			SDMA0_GFX_RB_CNTL__RPTR_WRITEBACK_SWAP_ENABLE_MASK;
467 #endif
468 		WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
469 
470 		/* Initialize the ring buffer's read and write pointers */
471 		WREG32(mmSDMA0_GFX_RB_RPTR + sdma_offsets[i], 0);
472 		WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], 0);
473 		WREG32(mmSDMA0_GFX_IB_RPTR + sdma_offsets[i], 0);
474 		WREG32(mmSDMA0_GFX_IB_OFFSET + sdma_offsets[i], 0);
475 
476 		/* set the wb address whether it's enabled or not */
477 		WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_HI + sdma_offsets[i],
478 		       upper_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFF);
479 		WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_LO + sdma_offsets[i],
480 		       ((ring->rptr_gpu_addr) & 0xFFFFFFFC));
481 
482 		rb_cntl |= SDMA0_GFX_RB_CNTL__RPTR_WRITEBACK_ENABLE_MASK;
483 
484 		WREG32(mmSDMA0_GFX_RB_BASE + sdma_offsets[i], ring->gpu_addr >> 8);
485 		WREG32(mmSDMA0_GFX_RB_BASE_HI + sdma_offsets[i], ring->gpu_addr >> 40);
486 
487 		ring->wptr = 0;
488 		WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], ring->wptr << 2);
489 
490 		/* enable DMA RB */
491 		WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i],
492 		       rb_cntl | SDMA0_GFX_RB_CNTL__RB_ENABLE_MASK);
493 
494 		ib_cntl = SDMA0_GFX_IB_CNTL__IB_ENABLE_MASK;
495 #ifdef __BIG_ENDIAN
496 		ib_cntl |= SDMA0_GFX_IB_CNTL__IB_SWAP_ENABLE_MASK;
497 #endif
498 		/* enable DMA IBs */
499 		WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
500 
501 		ring->sched.ready = true;
502 	}
503 
504 	cik_sdma_enable(adev, true);
505 
506 	for (i = 0; i < adev->sdma.num_instances; i++) {
507 		ring = &adev->sdma.instance[i].ring;
508 		r = amdgpu_ring_test_helper(ring);
509 		if (r)
510 			return r;
511 
512 		if (adev->mman.buffer_funcs_ring == ring)
513 			amdgpu_ttm_set_buffer_funcs_status(adev, true);
514 	}
515 
516 	return 0;
517 }
518 
519 /**
520  * cik_sdma_rlc_resume - setup and start the async dma engines
521  *
522  * @adev: amdgpu_device pointer
523  *
524  * Set up the compute DMA queues and enable them (CIK).
525  * Returns 0 for success, error for failure.
526  */
527 static int cik_sdma_rlc_resume(struct amdgpu_device *adev)
528 {
529 	/* XXX todo */
530 	return 0;
531 }
532 
533 /**
534  * cik_sdma_load_microcode - load the sDMA ME ucode
535  *
536  * @adev: amdgpu_device pointer
537  *
538  * Loads the sDMA0/1 ucode.
539  * Returns 0 for success, -EINVAL if the ucode is not available.
540  */
541 static int cik_sdma_load_microcode(struct amdgpu_device *adev)
542 {
543 	const struct sdma_firmware_header_v1_0 *hdr;
544 	const __le32 *fw_data;
545 	u32 fw_size;
546 	int i, j;
547 
548 	/* halt the MEs */
549 	cik_sdma_enable(adev, false);
550 
551 	for (i = 0; i < adev->sdma.num_instances; i++) {
552 		if (!adev->sdma.instance[i].fw)
553 			return -EINVAL;
554 		hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
555 		amdgpu_ucode_print_sdma_hdr(&hdr->header);
556 		fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
557 		adev->sdma.instance[i].fw_version = le32_to_cpu(hdr->header.ucode_version);
558 		adev->sdma.instance[i].feature_version = le32_to_cpu(hdr->ucode_feature_version);
559 		if (adev->sdma.instance[i].feature_version >= 20)
560 			adev->sdma.instance[i].burst_nop = true;
561 		fw_data = (const __le32 *)
562 			(adev->sdma.instance[i].fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
563 		WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], 0);
564 		for (j = 0; j < fw_size; j++)
565 			WREG32(mmSDMA0_UCODE_DATA + sdma_offsets[i], le32_to_cpup(fw_data++));
566 		WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], adev->sdma.instance[i].fw_version);
567 	}
568 
569 	return 0;
570 }
571 
572 /**
573  * cik_sdma_start - setup and start the async dma engines
574  *
575  * @adev: amdgpu_device pointer
576  *
577  * Set up the DMA engines and enable them (CIK).
578  * Returns 0 for success, error for failure.
579  */
580 static int cik_sdma_start(struct amdgpu_device *adev)
581 {
582 	int r;
583 
584 	r = cik_sdma_load_microcode(adev);
585 	if (r)
586 		return r;
587 
588 	/* halt the engine before programing */
589 	cik_sdma_enable(adev, false);
590 	/* enable sdma ring preemption */
591 	cik_ctx_switch_enable(adev, true);
592 
593 	/* start the gfx rings and rlc compute queues */
594 	r = cik_sdma_gfx_resume(adev);
595 	if (r)
596 		return r;
597 	r = cik_sdma_rlc_resume(adev);
598 	if (r)
599 		return r;
600 
601 	return 0;
602 }
603 
604 /**
605  * cik_sdma_ring_test_ring - simple async dma engine test
606  *
607  * @ring: amdgpu_ring structure holding ring information
608  *
609  * Test the DMA engine by writing using it to write an
610  * value to memory. (CIK).
611  * Returns 0 for success, error for failure.
612  */
613 static int cik_sdma_ring_test_ring(struct amdgpu_ring *ring)
614 {
615 	struct amdgpu_device *adev = ring->adev;
616 	unsigned i;
617 	unsigned index;
618 	int r;
619 	u32 tmp;
620 	u64 gpu_addr;
621 
622 	r = amdgpu_device_wb_get(adev, &index);
623 	if (r)
624 		return r;
625 
626 	gpu_addr = adev->wb.gpu_addr + (index * 4);
627 	tmp = 0xCAFEDEAD;
628 	adev->wb.wb[index] = cpu_to_le32(tmp);
629 
630 	r = amdgpu_ring_alloc(ring, 5);
631 	if (r)
632 		goto error_free_wb;
633 
634 	amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_WRITE, SDMA_WRITE_SUB_OPCODE_LINEAR, 0));
635 	amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
636 	amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
637 	amdgpu_ring_write(ring, 1); /* number of DWs to follow */
638 	amdgpu_ring_write(ring, 0xDEADBEEF);
639 	amdgpu_ring_commit(ring);
640 
641 	for (i = 0; i < adev->usec_timeout; i++) {
642 		tmp = le32_to_cpu(adev->wb.wb[index]);
643 		if (tmp == 0xDEADBEEF)
644 			break;
645 		udelay(1);
646 	}
647 
648 	if (i >= adev->usec_timeout)
649 		r = -ETIMEDOUT;
650 
651 error_free_wb:
652 	amdgpu_device_wb_free(adev, index);
653 	return r;
654 }
655 
656 /**
657  * cik_sdma_ring_test_ib - test an IB on the DMA engine
658  *
659  * @ring: amdgpu_ring structure holding ring information
660  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
661  *
662  * Test a simple IB in the DMA ring (CIK).
663  * Returns 0 on success, error on failure.
664  */
665 static int cik_sdma_ring_test_ib(struct amdgpu_ring *ring, long timeout)
666 {
667 	struct amdgpu_device *adev = ring->adev;
668 	struct amdgpu_ib ib;
669 	struct dma_fence *f = NULL;
670 	unsigned index;
671 	u32 tmp = 0;
672 	u64 gpu_addr;
673 	long r;
674 
675 	r = amdgpu_device_wb_get(adev, &index);
676 	if (r)
677 		return r;
678 
679 	gpu_addr = adev->wb.gpu_addr + (index * 4);
680 	tmp = 0xCAFEDEAD;
681 	adev->wb.wb[index] = cpu_to_le32(tmp);
682 	memset(&ib, 0, sizeof(ib));
683 	r = amdgpu_ib_get(adev, NULL, 256,
684 					AMDGPU_IB_POOL_DIRECT, &ib);
685 	if (r)
686 		goto err0;
687 
688 	ib.ptr[0] = SDMA_PACKET(SDMA_OPCODE_WRITE,
689 				SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
690 	ib.ptr[1] = lower_32_bits(gpu_addr);
691 	ib.ptr[2] = upper_32_bits(gpu_addr);
692 	ib.ptr[3] = 1;
693 	ib.ptr[4] = 0xDEADBEEF;
694 	ib.length_dw = 5;
695 	r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
696 	if (r)
697 		goto err1;
698 
699 	r = dma_fence_wait_timeout(f, false, timeout);
700 	if (r == 0) {
701 		r = -ETIMEDOUT;
702 		goto err1;
703 	} else if (r < 0) {
704 		goto err1;
705 	}
706 	tmp = le32_to_cpu(adev->wb.wb[index]);
707 	if (tmp == 0xDEADBEEF)
708 		r = 0;
709 	else
710 		r = -EINVAL;
711 
712 err1:
713 	amdgpu_ib_free(adev, &ib, NULL);
714 	dma_fence_put(f);
715 err0:
716 	amdgpu_device_wb_free(adev, index);
717 	return r;
718 }
719 
720 /**
721  * cik_sdma_vm_copy_pte - update PTEs by copying them from the GART
722  *
723  * @ib: indirect buffer to fill with commands
724  * @pe: addr of the page entry
725  * @src: src addr to copy from
726  * @count: number of page entries to update
727  *
728  * Update PTEs by copying them from the GART using sDMA (CIK).
729  */
730 static void cik_sdma_vm_copy_pte(struct amdgpu_ib *ib,
731 				 uint64_t pe, uint64_t src,
732 				 unsigned count)
733 {
734 	unsigned bytes = count * 8;
735 
736 	ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_COPY,
737 		SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
738 	ib->ptr[ib->length_dw++] = bytes;
739 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
740 	ib->ptr[ib->length_dw++] = lower_32_bits(src);
741 	ib->ptr[ib->length_dw++] = upper_32_bits(src);
742 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
743 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
744 }
745 
746 /**
747  * cik_sdma_vm_write_pte - update PTEs by writing them manually
748  *
749  * @ib: indirect buffer to fill with commands
750  * @pe: addr of the page entry
751  * @value: dst addr to write into pe
752  * @count: number of page entries to update
753  * @incr: increase next addr by incr bytes
754  *
755  * Update PTEs by writing them manually using sDMA (CIK).
756  */
757 static void cik_sdma_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
758 				  uint64_t value, unsigned count,
759 				  uint32_t incr)
760 {
761 	unsigned ndw = count * 2;
762 
763 	ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_WRITE,
764 		SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
765 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
766 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
767 	ib->ptr[ib->length_dw++] = ndw;
768 	for (; ndw > 0; ndw -= 2) {
769 		ib->ptr[ib->length_dw++] = lower_32_bits(value);
770 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
771 		value += incr;
772 	}
773 }
774 
775 /**
776  * cik_sdma_vm_set_pte_pde - update the page tables using sDMA
777  *
778  * @ib: indirect buffer to fill with commands
779  * @pe: addr of the page entry
780  * @addr: dst addr to write into pe
781  * @count: number of page entries to update
782  * @incr: increase next addr by incr bytes
783  * @flags: access flags
784  *
785  * Update the page tables using sDMA (CIK).
786  */
787 static void cik_sdma_vm_set_pte_pde(struct amdgpu_ib *ib, uint64_t pe,
788 				    uint64_t addr, unsigned count,
789 				    uint32_t incr, uint64_t flags)
790 {
791 	/* for physically contiguous pages (vram) */
792 	ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_GENERATE_PTE_PDE, 0, 0);
793 	ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
794 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
795 	ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
796 	ib->ptr[ib->length_dw++] = upper_32_bits(flags);
797 	ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
798 	ib->ptr[ib->length_dw++] = upper_32_bits(addr);
799 	ib->ptr[ib->length_dw++] = incr; /* increment size */
800 	ib->ptr[ib->length_dw++] = 0;
801 	ib->ptr[ib->length_dw++] = count; /* number of entries */
802 }
803 
804 /**
805  * cik_sdma_ring_pad_ib - pad the IB to the required number of dw
806  *
807  * @ring: amdgpu_ring structure holding ring information
808  * @ib: indirect buffer to fill with padding
809  *
810  */
811 static void cik_sdma_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
812 {
813 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
814 	u32 pad_count;
815 	int i;
816 
817 	pad_count = (-ib->length_dw) & 7;
818 	for (i = 0; i < pad_count; i++)
819 		if (sdma && sdma->burst_nop && (i == 0))
820 			ib->ptr[ib->length_dw++] =
821 					SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0) |
822 					SDMA_NOP_COUNT(pad_count - 1);
823 		else
824 			ib->ptr[ib->length_dw++] =
825 					SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0);
826 }
827 
828 /**
829  * cik_sdma_ring_emit_pipeline_sync - sync the pipeline
830  *
831  * @ring: amdgpu_ring pointer
832  *
833  * Make sure all previous operations are completed (CIK).
834  */
835 static void cik_sdma_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
836 {
837 	uint32_t seq = ring->fence_drv.sync_seq;
838 	uint64_t addr = ring->fence_drv.gpu_addr;
839 
840 	/* wait for idle */
841 	amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0,
842 					    SDMA_POLL_REG_MEM_EXTRA_OP(0) |
843 					    SDMA_POLL_REG_MEM_EXTRA_FUNC(3) | /* equal */
844 					    SDMA_POLL_REG_MEM_EXTRA_M));
845 	amdgpu_ring_write(ring, addr & 0xfffffffc);
846 	amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
847 	amdgpu_ring_write(ring, seq); /* reference */
848 	amdgpu_ring_write(ring, 0xffffffff); /* mask */
849 	amdgpu_ring_write(ring, (0xfff << 16) | 4); /* retry count, poll interval */
850 }
851 
852 /**
853  * cik_sdma_ring_emit_vm_flush - cik vm flush using sDMA
854  *
855  * @ring: amdgpu_ring pointer
856  * @vmid: vmid number to use
857  * @pd_addr: address
858  *
859  * Update the page table base and flush the VM TLB
860  * using sDMA (CIK).
861  */
862 static void cik_sdma_ring_emit_vm_flush(struct amdgpu_ring *ring,
863 					unsigned vmid, uint64_t pd_addr)
864 {
865 	u32 extra_bits = (SDMA_POLL_REG_MEM_EXTRA_OP(0) |
866 			  SDMA_POLL_REG_MEM_EXTRA_FUNC(0)); /* always */
867 
868 	amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
869 
870 	amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0, extra_bits));
871 	amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST << 2);
872 	amdgpu_ring_write(ring, 0);
873 	amdgpu_ring_write(ring, 0); /* reference */
874 	amdgpu_ring_write(ring, 0); /* mask */
875 	amdgpu_ring_write(ring, (0xfff << 16) | 10); /* retry count, poll interval */
876 }
877 
878 static void cik_sdma_ring_emit_wreg(struct amdgpu_ring *ring,
879 				    uint32_t reg, uint32_t val)
880 {
881 	amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
882 	amdgpu_ring_write(ring, reg);
883 	amdgpu_ring_write(ring, val);
884 }
885 
886 static void cik_enable_sdma_mgcg(struct amdgpu_device *adev,
887 				 bool enable)
888 {
889 	u32 orig, data;
890 
891 	if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
892 		WREG32(mmSDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET, 0x00000100);
893 		WREG32(mmSDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET, 0x00000100);
894 	} else {
895 		orig = data = RREG32(mmSDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET);
896 		data |= 0xff000000;
897 		if (data != orig)
898 			WREG32(mmSDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET, data);
899 
900 		orig = data = RREG32(mmSDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET);
901 		data |= 0xff000000;
902 		if (data != orig)
903 			WREG32(mmSDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET, data);
904 	}
905 }
906 
907 static void cik_enable_sdma_mgls(struct amdgpu_device *adev,
908 				 bool enable)
909 {
910 	u32 orig, data;
911 
912 	if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) {
913 		orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET);
914 		data |= 0x100;
915 		if (orig != data)
916 			WREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET, data);
917 
918 		orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET);
919 		data |= 0x100;
920 		if (orig != data)
921 			WREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET, data);
922 	} else {
923 		orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET);
924 		data &= ~0x100;
925 		if (orig != data)
926 			WREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET, data);
927 
928 		orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET);
929 		data &= ~0x100;
930 		if (orig != data)
931 			WREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET, data);
932 	}
933 }
934 
935 static int cik_sdma_early_init(void *handle)
936 {
937 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
938 
939 	adev->sdma.num_instances = SDMA_MAX_INSTANCE;
940 
941 	cik_sdma_set_ring_funcs(adev);
942 	cik_sdma_set_irq_funcs(adev);
943 	cik_sdma_set_buffer_funcs(adev);
944 	cik_sdma_set_vm_pte_funcs(adev);
945 
946 	return 0;
947 }
948 
949 static int cik_sdma_sw_init(void *handle)
950 {
951 	struct amdgpu_ring *ring;
952 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
953 	int r, i;
954 
955 	r = cik_sdma_init_microcode(adev);
956 	if (r) {
957 		DRM_ERROR("Failed to load sdma firmware!\n");
958 		return r;
959 	}
960 
961 	/* SDMA trap event */
962 	r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 224,
963 			      &adev->sdma.trap_irq);
964 	if (r)
965 		return r;
966 
967 	/* SDMA Privileged inst */
968 	r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 241,
969 			      &adev->sdma.illegal_inst_irq);
970 	if (r)
971 		return r;
972 
973 	/* SDMA Privileged inst */
974 	r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 247,
975 			      &adev->sdma.illegal_inst_irq);
976 	if (r)
977 		return r;
978 
979 	for (i = 0; i < adev->sdma.num_instances; i++) {
980 		ring = &adev->sdma.instance[i].ring;
981 		ring->ring_obj = NULL;
982 		sprintf(ring->name, "sdma%d", i);
983 		r = amdgpu_ring_init(adev, ring, 1024,
984 				     &adev->sdma.trap_irq,
985 				     (i == 0) ? AMDGPU_SDMA_IRQ_INSTANCE0 :
986 				     AMDGPU_SDMA_IRQ_INSTANCE1,
987 				     AMDGPU_RING_PRIO_DEFAULT, NULL);
988 		if (r)
989 			return r;
990 	}
991 
992 	return r;
993 }
994 
995 static int cik_sdma_sw_fini(void *handle)
996 {
997 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
998 	int i;
999 
1000 	for (i = 0; i < adev->sdma.num_instances; i++)
1001 		amdgpu_ring_fini(&adev->sdma.instance[i].ring);
1002 
1003 	cik_sdma_free_microcode(adev);
1004 	return 0;
1005 }
1006 
1007 static int cik_sdma_hw_init(void *handle)
1008 {
1009 	int r;
1010 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1011 
1012 	r = cik_sdma_start(adev);
1013 	if (r)
1014 		return r;
1015 
1016 	return r;
1017 }
1018 
1019 static int cik_sdma_hw_fini(void *handle)
1020 {
1021 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1022 
1023 	cik_ctx_switch_enable(adev, false);
1024 	cik_sdma_enable(adev, false);
1025 
1026 	return 0;
1027 }
1028 
1029 static int cik_sdma_suspend(void *handle)
1030 {
1031 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1032 
1033 	return cik_sdma_hw_fini(adev);
1034 }
1035 
1036 static int cik_sdma_resume(void *handle)
1037 {
1038 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1039 
1040 	cik_sdma_soft_reset(handle);
1041 
1042 	return cik_sdma_hw_init(adev);
1043 }
1044 
1045 static bool cik_sdma_is_idle(void *handle)
1046 {
1047 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1048 	u32 tmp = RREG32(mmSRBM_STATUS2);
1049 
1050 	if (tmp & (SRBM_STATUS2__SDMA_BUSY_MASK |
1051 				SRBM_STATUS2__SDMA1_BUSY_MASK))
1052 	    return false;
1053 
1054 	return true;
1055 }
1056 
1057 static int cik_sdma_wait_for_idle(void *handle)
1058 {
1059 	unsigned i;
1060 	u32 tmp;
1061 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1062 
1063 	for (i = 0; i < adev->usec_timeout; i++) {
1064 		tmp = RREG32(mmSRBM_STATUS2) & (SRBM_STATUS2__SDMA_BUSY_MASK |
1065 				SRBM_STATUS2__SDMA1_BUSY_MASK);
1066 
1067 		if (!tmp)
1068 			return 0;
1069 		udelay(1);
1070 	}
1071 	return -ETIMEDOUT;
1072 }
1073 
1074 static int cik_sdma_soft_reset(void *handle)
1075 {
1076 	u32 srbm_soft_reset = 0;
1077 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1078 	u32 tmp;
1079 
1080 	/* sdma0 */
1081 	tmp = RREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET);
1082 	tmp |= SDMA0_F32_CNTL__HALT_MASK;
1083 	WREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET, tmp);
1084 	srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA_MASK;
1085 
1086 	/* sdma1 */
1087 	tmp = RREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET);
1088 	tmp |= SDMA0_F32_CNTL__HALT_MASK;
1089 	WREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET, tmp);
1090 	srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA1_MASK;
1091 
1092 	if (srbm_soft_reset) {
1093 		tmp = RREG32(mmSRBM_SOFT_RESET);
1094 		tmp |= srbm_soft_reset;
1095 		dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
1096 		WREG32(mmSRBM_SOFT_RESET, tmp);
1097 		tmp = RREG32(mmSRBM_SOFT_RESET);
1098 
1099 		udelay(50);
1100 
1101 		tmp &= ~srbm_soft_reset;
1102 		WREG32(mmSRBM_SOFT_RESET, tmp);
1103 		tmp = RREG32(mmSRBM_SOFT_RESET);
1104 
1105 		/* Wait a little for things to settle down */
1106 		udelay(50);
1107 	}
1108 
1109 	return 0;
1110 }
1111 
1112 static int cik_sdma_set_trap_irq_state(struct amdgpu_device *adev,
1113 				       struct amdgpu_irq_src *src,
1114 				       unsigned type,
1115 				       enum amdgpu_interrupt_state state)
1116 {
1117 	u32 sdma_cntl;
1118 
1119 	switch (type) {
1120 	case AMDGPU_SDMA_IRQ_INSTANCE0:
1121 		switch (state) {
1122 		case AMDGPU_IRQ_STATE_DISABLE:
1123 			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1124 			sdma_cntl &= ~SDMA0_CNTL__TRAP_ENABLE_MASK;
1125 			WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1126 			break;
1127 		case AMDGPU_IRQ_STATE_ENABLE:
1128 			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1129 			sdma_cntl |= SDMA0_CNTL__TRAP_ENABLE_MASK;
1130 			WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1131 			break;
1132 		default:
1133 			break;
1134 		}
1135 		break;
1136 	case AMDGPU_SDMA_IRQ_INSTANCE1:
1137 		switch (state) {
1138 		case AMDGPU_IRQ_STATE_DISABLE:
1139 			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1140 			sdma_cntl &= ~SDMA0_CNTL__TRAP_ENABLE_MASK;
1141 			WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1142 			break;
1143 		case AMDGPU_IRQ_STATE_ENABLE:
1144 			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1145 			sdma_cntl |= SDMA0_CNTL__TRAP_ENABLE_MASK;
1146 			WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1147 			break;
1148 		default:
1149 			break;
1150 		}
1151 		break;
1152 	default:
1153 		break;
1154 	}
1155 	return 0;
1156 }
1157 
1158 static int cik_sdma_process_trap_irq(struct amdgpu_device *adev,
1159 				     struct amdgpu_irq_src *source,
1160 				     struct amdgpu_iv_entry *entry)
1161 {
1162 	u8 instance_id, queue_id;
1163 
1164 	instance_id = (entry->ring_id & 0x3) >> 0;
1165 	queue_id = (entry->ring_id & 0xc) >> 2;
1166 	DRM_DEBUG("IH: SDMA trap\n");
1167 	switch (instance_id) {
1168 	case 0:
1169 		switch (queue_id) {
1170 		case 0:
1171 			amdgpu_fence_process(&adev->sdma.instance[0].ring);
1172 			break;
1173 		case 1:
1174 			/* XXX compute */
1175 			break;
1176 		case 2:
1177 			/* XXX compute */
1178 			break;
1179 		}
1180 		break;
1181 	case 1:
1182 		switch (queue_id) {
1183 		case 0:
1184 			amdgpu_fence_process(&adev->sdma.instance[1].ring);
1185 			break;
1186 		case 1:
1187 			/* XXX compute */
1188 			break;
1189 		case 2:
1190 			/* XXX compute */
1191 			break;
1192 		}
1193 		break;
1194 	}
1195 
1196 	return 0;
1197 }
1198 
1199 static int cik_sdma_process_illegal_inst_irq(struct amdgpu_device *adev,
1200 					     struct amdgpu_irq_src *source,
1201 					     struct amdgpu_iv_entry *entry)
1202 {
1203 	u8 instance_id;
1204 
1205 	DRM_ERROR("Illegal instruction in SDMA command stream\n");
1206 	instance_id = (entry->ring_id & 0x3) >> 0;
1207 	drm_sched_fault(&adev->sdma.instance[instance_id].ring.sched);
1208 	return 0;
1209 }
1210 
1211 static int cik_sdma_set_clockgating_state(void *handle,
1212 					  enum amd_clockgating_state state)
1213 {
1214 	bool gate = false;
1215 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1216 
1217 	if (state == AMD_CG_STATE_GATE)
1218 		gate = true;
1219 
1220 	cik_enable_sdma_mgcg(adev, gate);
1221 	cik_enable_sdma_mgls(adev, gate);
1222 
1223 	return 0;
1224 }
1225 
1226 static int cik_sdma_set_powergating_state(void *handle,
1227 					  enum amd_powergating_state state)
1228 {
1229 	return 0;
1230 }
1231 
1232 static const struct amd_ip_funcs cik_sdma_ip_funcs = {
1233 	.name = "cik_sdma",
1234 	.early_init = cik_sdma_early_init,
1235 	.late_init = NULL,
1236 	.sw_init = cik_sdma_sw_init,
1237 	.sw_fini = cik_sdma_sw_fini,
1238 	.hw_init = cik_sdma_hw_init,
1239 	.hw_fini = cik_sdma_hw_fini,
1240 	.suspend = cik_sdma_suspend,
1241 	.resume = cik_sdma_resume,
1242 	.is_idle = cik_sdma_is_idle,
1243 	.wait_for_idle = cik_sdma_wait_for_idle,
1244 	.soft_reset = cik_sdma_soft_reset,
1245 	.set_clockgating_state = cik_sdma_set_clockgating_state,
1246 	.set_powergating_state = cik_sdma_set_powergating_state,
1247 };
1248 
1249 static const struct amdgpu_ring_funcs cik_sdma_ring_funcs = {
1250 	.type = AMDGPU_RING_TYPE_SDMA,
1251 	.align_mask = 0xf,
1252 	.nop = SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0),
1253 	.support_64bit_ptrs = false,
1254 	.get_rptr = cik_sdma_ring_get_rptr,
1255 	.get_wptr = cik_sdma_ring_get_wptr,
1256 	.set_wptr = cik_sdma_ring_set_wptr,
1257 	.emit_frame_size =
1258 		6 + /* cik_sdma_ring_emit_hdp_flush */
1259 		3 + /* hdp invalidate */
1260 		6 + /* cik_sdma_ring_emit_pipeline_sync */
1261 		CIK_FLUSH_GPU_TLB_NUM_WREG * 3 + 6 + /* cik_sdma_ring_emit_vm_flush */
1262 		9 + 9 + 9, /* cik_sdma_ring_emit_fence x3 for user fence, vm fence */
1263 	.emit_ib_size = 7 + 4, /* cik_sdma_ring_emit_ib */
1264 	.emit_ib = cik_sdma_ring_emit_ib,
1265 	.emit_fence = cik_sdma_ring_emit_fence,
1266 	.emit_pipeline_sync = cik_sdma_ring_emit_pipeline_sync,
1267 	.emit_vm_flush = cik_sdma_ring_emit_vm_flush,
1268 	.emit_hdp_flush = cik_sdma_ring_emit_hdp_flush,
1269 	.test_ring = cik_sdma_ring_test_ring,
1270 	.test_ib = cik_sdma_ring_test_ib,
1271 	.insert_nop = cik_sdma_ring_insert_nop,
1272 	.pad_ib = cik_sdma_ring_pad_ib,
1273 	.emit_wreg = cik_sdma_ring_emit_wreg,
1274 };
1275 
1276 static void cik_sdma_set_ring_funcs(struct amdgpu_device *adev)
1277 {
1278 	int i;
1279 
1280 	for (i = 0; i < adev->sdma.num_instances; i++) {
1281 		adev->sdma.instance[i].ring.funcs = &cik_sdma_ring_funcs;
1282 		adev->sdma.instance[i].ring.me = i;
1283 	}
1284 }
1285 
1286 static const struct amdgpu_irq_src_funcs cik_sdma_trap_irq_funcs = {
1287 	.set = cik_sdma_set_trap_irq_state,
1288 	.process = cik_sdma_process_trap_irq,
1289 };
1290 
1291 static const struct amdgpu_irq_src_funcs cik_sdma_illegal_inst_irq_funcs = {
1292 	.process = cik_sdma_process_illegal_inst_irq,
1293 };
1294 
1295 static void cik_sdma_set_irq_funcs(struct amdgpu_device *adev)
1296 {
1297 	adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST;
1298 	adev->sdma.trap_irq.funcs = &cik_sdma_trap_irq_funcs;
1299 	adev->sdma.illegal_inst_irq.funcs = &cik_sdma_illegal_inst_irq_funcs;
1300 }
1301 
1302 /**
1303  * cik_sdma_emit_copy_buffer - copy buffer using the sDMA engine
1304  *
1305  * @ib: indirect buffer to copy to
1306  * @src_offset: src GPU address
1307  * @dst_offset: dst GPU address
1308  * @byte_count: number of bytes to xfer
1309  * @tmz: is this a secure operation
1310  *
1311  * Copy GPU buffers using the DMA engine (CIK).
1312  * Used by the amdgpu ttm implementation to move pages if
1313  * registered as the asic copy callback.
1314  */
1315 static void cik_sdma_emit_copy_buffer(struct amdgpu_ib *ib,
1316 				      uint64_t src_offset,
1317 				      uint64_t dst_offset,
1318 				      uint32_t byte_count,
1319 				      bool tmz)
1320 {
1321 	ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_COPY, SDMA_COPY_SUB_OPCODE_LINEAR, 0);
1322 	ib->ptr[ib->length_dw++] = byte_count;
1323 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1324 	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1325 	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1326 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1327 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1328 }
1329 
1330 /**
1331  * cik_sdma_emit_fill_buffer - fill buffer using the sDMA engine
1332  *
1333  * @ib: indirect buffer to fill
1334  * @src_data: value to write to buffer
1335  * @dst_offset: dst GPU address
1336  * @byte_count: number of bytes to xfer
1337  *
1338  * Fill GPU buffers using the DMA engine (CIK).
1339  */
1340 static void cik_sdma_emit_fill_buffer(struct amdgpu_ib *ib,
1341 				      uint32_t src_data,
1342 				      uint64_t dst_offset,
1343 				      uint32_t byte_count)
1344 {
1345 	ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_CONSTANT_FILL, 0, 0);
1346 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1347 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1348 	ib->ptr[ib->length_dw++] = src_data;
1349 	ib->ptr[ib->length_dw++] = byte_count;
1350 }
1351 
1352 static const struct amdgpu_buffer_funcs cik_sdma_buffer_funcs = {
1353 	.copy_max_bytes = 0x1fffff,
1354 	.copy_num_dw = 7,
1355 	.emit_copy_buffer = cik_sdma_emit_copy_buffer,
1356 
1357 	.fill_max_bytes = 0x1fffff,
1358 	.fill_num_dw = 5,
1359 	.emit_fill_buffer = cik_sdma_emit_fill_buffer,
1360 };
1361 
1362 static void cik_sdma_set_buffer_funcs(struct amdgpu_device *adev)
1363 {
1364 	adev->mman.buffer_funcs = &cik_sdma_buffer_funcs;
1365 	adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1366 }
1367 
1368 static const struct amdgpu_vm_pte_funcs cik_sdma_vm_pte_funcs = {
1369 	.copy_pte_num_dw = 7,
1370 	.copy_pte = cik_sdma_vm_copy_pte,
1371 
1372 	.write_pte = cik_sdma_vm_write_pte,
1373 	.set_pte_pde = cik_sdma_vm_set_pte_pde,
1374 };
1375 
1376 static void cik_sdma_set_vm_pte_funcs(struct amdgpu_device *adev)
1377 {
1378 	unsigned i;
1379 
1380 	adev->vm_manager.vm_pte_funcs = &cik_sdma_vm_pte_funcs;
1381 	for (i = 0; i < adev->sdma.num_instances; i++) {
1382 		adev->vm_manager.vm_pte_scheds[i] =
1383 			&adev->sdma.instance[i].ring.sched;
1384 	}
1385 	adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
1386 }
1387 
1388 const struct amdgpu_ip_block_version cik_sdma_ip_block =
1389 {
1390 	.type = AMD_IP_BLOCK_TYPE_SDMA,
1391 	.major = 2,
1392 	.minor = 0,
1393 	.rev = 0,
1394 	.funcs = &cik_sdma_ip_funcs,
1395 };
1396