xref: /openbmc/linux/drivers/gpu/drm/amd/amdgpu/cik_sdma.c (revision 002dff36)
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->adev->wb.wb[ring->rptr_offs];
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 		       	(lower_32_bits(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  * @ib: IB object to schedule
219  *
220  * Schedule an IB in the DMA ring (CIK).
221  */
222 static void cik_sdma_ring_emit_ib(struct amdgpu_ring *ring,
223 				  struct amdgpu_job *job,
224 				  struct amdgpu_ib *ib,
225 				  uint32_t flags)
226 {
227 	unsigned vmid = AMDGPU_JOB_GET_VMID(job);
228 	u32 extra_bits = vmid & 0xf;
229 
230 	/* IB packet must end on a 8 DW boundary */
231 	cik_sdma_ring_insert_nop(ring, (4 - lower_32_bits(ring->wptr)) & 7);
232 
233 	amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_INDIRECT_BUFFER, 0, extra_bits));
234 	amdgpu_ring_write(ring, ib->gpu_addr & 0xffffffe0); /* base must be 32 byte aligned */
235 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr) & 0xffffffff);
236 	amdgpu_ring_write(ring, ib->length_dw);
237 
238 }
239 
240 /**
241  * cik_sdma_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
242  *
243  * @ring: amdgpu ring pointer
244  *
245  * Emit an hdp flush packet on the requested DMA ring.
246  */
247 static void cik_sdma_ring_emit_hdp_flush(struct amdgpu_ring *ring)
248 {
249 	u32 extra_bits = (SDMA_POLL_REG_MEM_EXTRA_OP(1) |
250 			  SDMA_POLL_REG_MEM_EXTRA_FUNC(3)); /* == */
251 	u32 ref_and_mask;
252 
253 	if (ring->me == 0)
254 		ref_and_mask = GPU_HDP_FLUSH_DONE__SDMA0_MASK;
255 	else
256 		ref_and_mask = GPU_HDP_FLUSH_DONE__SDMA1_MASK;
257 
258 	amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0, extra_bits));
259 	amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_DONE << 2);
260 	amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_REQ << 2);
261 	amdgpu_ring_write(ring, ref_and_mask); /* reference */
262 	amdgpu_ring_write(ring, ref_and_mask); /* mask */
263 	amdgpu_ring_write(ring, (0xfff << 16) | 10); /* retry count, poll interval */
264 }
265 
266 /**
267  * cik_sdma_ring_emit_fence - emit a fence on the DMA ring
268  *
269  * @ring: amdgpu ring pointer
270  * @fence: amdgpu fence object
271  *
272  * Add a DMA fence packet to the ring to write
273  * the fence seq number and DMA trap packet to generate
274  * an interrupt if needed (CIK).
275  */
276 static void cik_sdma_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
277 				     unsigned flags)
278 {
279 	bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
280 	/* write the fence */
281 	amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_FENCE, 0, 0));
282 	amdgpu_ring_write(ring, lower_32_bits(addr));
283 	amdgpu_ring_write(ring, upper_32_bits(addr));
284 	amdgpu_ring_write(ring, lower_32_bits(seq));
285 
286 	/* optionally write high bits as well */
287 	if (write64bit) {
288 		addr += 4;
289 		amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_FENCE, 0, 0));
290 		amdgpu_ring_write(ring, lower_32_bits(addr));
291 		amdgpu_ring_write(ring, upper_32_bits(addr));
292 		amdgpu_ring_write(ring, upper_32_bits(seq));
293 	}
294 
295 	/* generate an interrupt */
296 	amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_TRAP, 0, 0));
297 }
298 
299 /**
300  * cik_sdma_gfx_stop - stop the gfx async dma engines
301  *
302  * @adev: amdgpu_device pointer
303  *
304  * Stop the gfx async dma ring buffers (CIK).
305  */
306 static void cik_sdma_gfx_stop(struct amdgpu_device *adev)
307 {
308 	struct amdgpu_ring *sdma0 = &adev->sdma.instance[0].ring;
309 	struct amdgpu_ring *sdma1 = &adev->sdma.instance[1].ring;
310 	u32 rb_cntl;
311 	int i;
312 
313 	if ((adev->mman.buffer_funcs_ring == sdma0) ||
314 	    (adev->mman.buffer_funcs_ring == sdma1))
315 			amdgpu_ttm_set_buffer_funcs_status(adev, false);
316 
317 	for (i = 0; i < adev->sdma.num_instances; i++) {
318 		rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
319 		rb_cntl &= ~SDMA0_GFX_RB_CNTL__RB_ENABLE_MASK;
320 		WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
321 		WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], 0);
322 	}
323 }
324 
325 /**
326  * cik_sdma_rlc_stop - stop the compute async dma engines
327  *
328  * @adev: amdgpu_device pointer
329  *
330  * Stop the compute async dma queues (CIK).
331  */
332 static void cik_sdma_rlc_stop(struct amdgpu_device *adev)
333 {
334 	/* XXX todo */
335 }
336 
337 /**
338  * cik_ctx_switch_enable - stop the async dma engines context switch
339  *
340  * @adev: amdgpu_device pointer
341  * @enable: enable/disable the DMA MEs context switch.
342  *
343  * Halt or unhalt the async dma engines context switch (VI).
344  */
345 static void cik_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
346 {
347 	u32 f32_cntl, phase_quantum = 0;
348 	int i;
349 
350 	if (amdgpu_sdma_phase_quantum) {
351 		unsigned value = amdgpu_sdma_phase_quantum;
352 		unsigned unit = 0;
353 
354 		while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
355 				SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) {
356 			value = (value + 1) >> 1;
357 			unit++;
358 		}
359 		if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
360 			    SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) {
361 			value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
362 				 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT);
363 			unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
364 				SDMA0_PHASE0_QUANTUM__UNIT__SHIFT);
365 			WARN_ONCE(1,
366 			"clamping sdma_phase_quantum to %uK clock cycles\n",
367 				  value << unit);
368 		}
369 		phase_quantum =
370 			value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
371 			unit  << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
372 	}
373 
374 	for (i = 0; i < adev->sdma.num_instances; i++) {
375 		f32_cntl = RREG32(mmSDMA0_CNTL + sdma_offsets[i]);
376 		if (enable) {
377 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
378 					AUTO_CTXSW_ENABLE, 1);
379 			if (amdgpu_sdma_phase_quantum) {
380 				WREG32(mmSDMA0_PHASE0_QUANTUM + sdma_offsets[i],
381 				       phase_quantum);
382 				WREG32(mmSDMA0_PHASE1_QUANTUM + sdma_offsets[i],
383 				       phase_quantum);
384 			}
385 		} else {
386 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
387 					AUTO_CTXSW_ENABLE, 0);
388 		}
389 
390 		WREG32(mmSDMA0_CNTL + sdma_offsets[i], f32_cntl);
391 	}
392 }
393 
394 /**
395  * cik_sdma_enable - stop the async dma engines
396  *
397  * @adev: amdgpu_device pointer
398  * @enable: enable/disable the DMA MEs.
399  *
400  * Halt or unhalt the async dma engines (CIK).
401  */
402 static void cik_sdma_enable(struct amdgpu_device *adev, bool enable)
403 {
404 	u32 me_cntl;
405 	int i;
406 
407 	if (!enable) {
408 		cik_sdma_gfx_stop(adev);
409 		cik_sdma_rlc_stop(adev);
410 	}
411 
412 	for (i = 0; i < adev->sdma.num_instances; i++) {
413 		me_cntl = RREG32(mmSDMA0_F32_CNTL + sdma_offsets[i]);
414 		if (enable)
415 			me_cntl &= ~SDMA0_F32_CNTL__HALT_MASK;
416 		else
417 			me_cntl |= SDMA0_F32_CNTL__HALT_MASK;
418 		WREG32(mmSDMA0_F32_CNTL + sdma_offsets[i], me_cntl);
419 	}
420 }
421 
422 /**
423  * cik_sdma_gfx_resume - setup and start the async dma engines
424  *
425  * @adev: amdgpu_device pointer
426  *
427  * Set up the gfx DMA ring buffers and enable them (CIK).
428  * Returns 0 for success, error for failure.
429  */
430 static int cik_sdma_gfx_resume(struct amdgpu_device *adev)
431 {
432 	struct amdgpu_ring *ring;
433 	u32 rb_cntl, ib_cntl;
434 	u32 rb_bufsz;
435 	u32 wb_offset;
436 	int i, j, r;
437 
438 	for (i = 0; i < adev->sdma.num_instances; i++) {
439 		ring = &adev->sdma.instance[i].ring;
440 		wb_offset = (ring->rptr_offs * 4);
441 
442 		mutex_lock(&adev->srbm_mutex);
443 		for (j = 0; j < 16; j++) {
444 			cik_srbm_select(adev, 0, 0, 0, j);
445 			/* SDMA GFX */
446 			WREG32(mmSDMA0_GFX_VIRTUAL_ADDR + sdma_offsets[i], 0);
447 			WREG32(mmSDMA0_GFX_APE1_CNTL + sdma_offsets[i], 0);
448 			/* XXX SDMA RLC - todo */
449 		}
450 		cik_srbm_select(adev, 0, 0, 0, 0);
451 		mutex_unlock(&adev->srbm_mutex);
452 
453 		WREG32(mmSDMA0_TILING_CONFIG + sdma_offsets[i],
454 		       adev->gfx.config.gb_addr_config & 0x70);
455 
456 		WREG32(mmSDMA0_SEM_INCOMPLETE_TIMER_CNTL + sdma_offsets[i], 0);
457 		WREG32(mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i], 0);
458 
459 		/* Set ring buffer size in dwords */
460 		rb_bufsz = order_base_2(ring->ring_size / 4);
461 		rb_cntl = rb_bufsz << 1;
462 #ifdef __BIG_ENDIAN
463 		rb_cntl |= SDMA0_GFX_RB_CNTL__RB_SWAP_ENABLE_MASK |
464 			SDMA0_GFX_RB_CNTL__RPTR_WRITEBACK_SWAP_ENABLE_MASK;
465 #endif
466 		WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
467 
468 		/* Initialize the ring buffer's read and write pointers */
469 		WREG32(mmSDMA0_GFX_RB_RPTR + sdma_offsets[i], 0);
470 		WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], 0);
471 		WREG32(mmSDMA0_GFX_IB_RPTR + sdma_offsets[i], 0);
472 		WREG32(mmSDMA0_GFX_IB_OFFSET + sdma_offsets[i], 0);
473 
474 		/* set the wb address whether it's enabled or not */
475 		WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_HI + sdma_offsets[i],
476 		       upper_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFF);
477 		WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_LO + sdma_offsets[i],
478 		       ((adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC));
479 
480 		rb_cntl |= SDMA0_GFX_RB_CNTL__RPTR_WRITEBACK_ENABLE_MASK;
481 
482 		WREG32(mmSDMA0_GFX_RB_BASE + sdma_offsets[i], ring->gpu_addr >> 8);
483 		WREG32(mmSDMA0_GFX_RB_BASE_HI + sdma_offsets[i], ring->gpu_addr >> 40);
484 
485 		ring->wptr = 0;
486 		WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], lower_32_bits(ring->wptr) << 2);
487 
488 		/* enable DMA RB */
489 		WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i],
490 		       rb_cntl | SDMA0_GFX_RB_CNTL__RB_ENABLE_MASK);
491 
492 		ib_cntl = SDMA0_GFX_IB_CNTL__IB_ENABLE_MASK;
493 #ifdef __BIG_ENDIAN
494 		ib_cntl |= SDMA0_GFX_IB_CNTL__IB_SWAP_ENABLE_MASK;
495 #endif
496 		/* enable DMA IBs */
497 		WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
498 
499 		ring->sched.ready = true;
500 	}
501 
502 	cik_sdma_enable(adev, true);
503 
504 	for (i = 0; i < adev->sdma.num_instances; i++) {
505 		ring = &adev->sdma.instance[i].ring;
506 		r = amdgpu_ring_test_helper(ring);
507 		if (r)
508 			return r;
509 
510 		if (adev->mman.buffer_funcs_ring == ring)
511 			amdgpu_ttm_set_buffer_funcs_status(adev, true);
512 	}
513 
514 	return 0;
515 }
516 
517 /**
518  * cik_sdma_rlc_resume - setup and start the async dma engines
519  *
520  * @adev: amdgpu_device pointer
521  *
522  * Set up the compute DMA queues and enable them (CIK).
523  * Returns 0 for success, error for failure.
524  */
525 static int cik_sdma_rlc_resume(struct amdgpu_device *adev)
526 {
527 	/* XXX todo */
528 	return 0;
529 }
530 
531 /**
532  * cik_sdma_load_microcode - load the sDMA ME ucode
533  *
534  * @adev: amdgpu_device pointer
535  *
536  * Loads the sDMA0/1 ucode.
537  * Returns 0 for success, -EINVAL if the ucode is not available.
538  */
539 static int cik_sdma_load_microcode(struct amdgpu_device *adev)
540 {
541 	const struct sdma_firmware_header_v1_0 *hdr;
542 	const __le32 *fw_data;
543 	u32 fw_size;
544 	int i, j;
545 
546 	/* halt the MEs */
547 	cik_sdma_enable(adev, false);
548 
549 	for (i = 0; i < adev->sdma.num_instances; i++) {
550 		if (!adev->sdma.instance[i].fw)
551 			return -EINVAL;
552 		hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
553 		amdgpu_ucode_print_sdma_hdr(&hdr->header);
554 		fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
555 		adev->sdma.instance[i].fw_version = le32_to_cpu(hdr->header.ucode_version);
556 		adev->sdma.instance[i].feature_version = le32_to_cpu(hdr->ucode_feature_version);
557 		if (adev->sdma.instance[i].feature_version >= 20)
558 			adev->sdma.instance[i].burst_nop = true;
559 		fw_data = (const __le32 *)
560 			(adev->sdma.instance[i].fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
561 		WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], 0);
562 		for (j = 0; j < fw_size; j++)
563 			WREG32(mmSDMA0_UCODE_DATA + sdma_offsets[i], le32_to_cpup(fw_data++));
564 		WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], adev->sdma.instance[i].fw_version);
565 	}
566 
567 	return 0;
568 }
569 
570 /**
571  * cik_sdma_start - setup and start the async dma engines
572  *
573  * @adev: amdgpu_device pointer
574  *
575  * Set up the DMA engines and enable them (CIK).
576  * Returns 0 for success, error for failure.
577  */
578 static int cik_sdma_start(struct amdgpu_device *adev)
579 {
580 	int r;
581 
582 	r = cik_sdma_load_microcode(adev);
583 	if (r)
584 		return r;
585 
586 	/* halt the engine before programing */
587 	cik_sdma_enable(adev, false);
588 	/* enable sdma ring preemption */
589 	cik_ctx_switch_enable(adev, true);
590 
591 	/* start the gfx rings and rlc compute queues */
592 	r = cik_sdma_gfx_resume(adev);
593 	if (r)
594 		return r;
595 	r = cik_sdma_rlc_resume(adev);
596 	if (r)
597 		return r;
598 
599 	return 0;
600 }
601 
602 /**
603  * cik_sdma_ring_test_ring - simple async dma engine test
604  *
605  * @ring: amdgpu_ring structure holding ring information
606  *
607  * Test the DMA engine by writing using it to write an
608  * value to memory. (CIK).
609  * Returns 0 for success, error for failure.
610  */
611 static int cik_sdma_ring_test_ring(struct amdgpu_ring *ring)
612 {
613 	struct amdgpu_device *adev = ring->adev;
614 	unsigned i;
615 	unsigned index;
616 	int r;
617 	u32 tmp;
618 	u64 gpu_addr;
619 
620 	r = amdgpu_device_wb_get(adev, &index);
621 	if (r)
622 		return r;
623 
624 	gpu_addr = adev->wb.gpu_addr + (index * 4);
625 	tmp = 0xCAFEDEAD;
626 	adev->wb.wb[index] = cpu_to_le32(tmp);
627 
628 	r = amdgpu_ring_alloc(ring, 5);
629 	if (r)
630 		goto error_free_wb;
631 
632 	amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_WRITE, SDMA_WRITE_SUB_OPCODE_LINEAR, 0));
633 	amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
634 	amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
635 	amdgpu_ring_write(ring, 1); /* number of DWs to follow */
636 	amdgpu_ring_write(ring, 0xDEADBEEF);
637 	amdgpu_ring_commit(ring);
638 
639 	for (i = 0; i < adev->usec_timeout; i++) {
640 		tmp = le32_to_cpu(adev->wb.wb[index]);
641 		if (tmp == 0xDEADBEEF)
642 			break;
643 		udelay(1);
644 	}
645 
646 	if (i >= adev->usec_timeout)
647 		r = -ETIMEDOUT;
648 
649 error_free_wb:
650 	amdgpu_device_wb_free(adev, index);
651 	return r;
652 }
653 
654 /**
655  * cik_sdma_ring_test_ib - test an IB on the DMA engine
656  *
657  * @ring: amdgpu_ring structure holding ring information
658  *
659  * Test a simple IB in the DMA ring (CIK).
660  * Returns 0 on success, error on failure.
661  */
662 static int cik_sdma_ring_test_ib(struct amdgpu_ring *ring, long timeout)
663 {
664 	struct amdgpu_device *adev = ring->adev;
665 	struct amdgpu_ib ib;
666 	struct dma_fence *f = NULL;
667 	unsigned index;
668 	u32 tmp = 0;
669 	u64 gpu_addr;
670 	long r;
671 
672 	r = amdgpu_device_wb_get(adev, &index);
673 	if (r)
674 		return r;
675 
676 	gpu_addr = adev->wb.gpu_addr + (index * 4);
677 	tmp = 0xCAFEDEAD;
678 	adev->wb.wb[index] = cpu_to_le32(tmp);
679 	memset(&ib, 0, sizeof(ib));
680 	r = amdgpu_ib_get(adev, NULL, 256,
681 					AMDGPU_IB_POOL_DIRECT, &ib);
682 	if (r)
683 		goto err0;
684 
685 	ib.ptr[0] = SDMA_PACKET(SDMA_OPCODE_WRITE,
686 				SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
687 	ib.ptr[1] = lower_32_bits(gpu_addr);
688 	ib.ptr[2] = upper_32_bits(gpu_addr);
689 	ib.ptr[3] = 1;
690 	ib.ptr[4] = 0xDEADBEEF;
691 	ib.length_dw = 5;
692 	r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
693 	if (r)
694 		goto err1;
695 
696 	r = dma_fence_wait_timeout(f, false, timeout);
697 	if (r == 0) {
698 		r = -ETIMEDOUT;
699 		goto err1;
700 	} else if (r < 0) {
701 		goto err1;
702 	}
703 	tmp = le32_to_cpu(adev->wb.wb[index]);
704 	if (tmp == 0xDEADBEEF)
705 		r = 0;
706 	else
707 		r = -EINVAL;
708 
709 err1:
710 	amdgpu_ib_free(adev, &ib, NULL);
711 	dma_fence_put(f);
712 err0:
713 	amdgpu_device_wb_free(adev, index);
714 	return r;
715 }
716 
717 /**
718  * cik_sdma_vm_copy_pages - update PTEs by copying them from the GART
719  *
720  * @ib: indirect buffer to fill with commands
721  * @pe: addr of the page entry
722  * @src: src addr to copy from
723  * @count: number of page entries to update
724  *
725  * Update PTEs by copying them from the GART using sDMA (CIK).
726  */
727 static void cik_sdma_vm_copy_pte(struct amdgpu_ib *ib,
728 				 uint64_t pe, uint64_t src,
729 				 unsigned count)
730 {
731 	unsigned bytes = count * 8;
732 
733 	ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_COPY,
734 		SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
735 	ib->ptr[ib->length_dw++] = bytes;
736 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
737 	ib->ptr[ib->length_dw++] = lower_32_bits(src);
738 	ib->ptr[ib->length_dw++] = upper_32_bits(src);
739 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
740 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
741 }
742 
743 /**
744  * cik_sdma_vm_write_pages - update PTEs by writing them manually
745  *
746  * @ib: indirect buffer to fill with commands
747  * @pe: addr of the page entry
748  * @value: dst addr to write into pe
749  * @count: number of page entries to update
750  * @incr: increase next addr by incr bytes
751  *
752  * Update PTEs by writing them manually using sDMA (CIK).
753  */
754 static void cik_sdma_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
755 				  uint64_t value, unsigned count,
756 				  uint32_t incr)
757 {
758 	unsigned ndw = count * 2;
759 
760 	ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_WRITE,
761 		SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
762 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
763 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
764 	ib->ptr[ib->length_dw++] = ndw;
765 	for (; ndw > 0; ndw -= 2) {
766 		ib->ptr[ib->length_dw++] = lower_32_bits(value);
767 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
768 		value += incr;
769 	}
770 }
771 
772 /**
773  * cik_sdma_vm_set_pages - update the page tables using sDMA
774  *
775  * @ib: indirect buffer to fill with commands
776  * @pe: addr of the page entry
777  * @addr: dst addr to write into pe
778  * @count: number of page entries to update
779  * @incr: increase next addr by incr bytes
780  * @flags: access flags
781  *
782  * Update the page tables using sDMA (CIK).
783  */
784 static void cik_sdma_vm_set_pte_pde(struct amdgpu_ib *ib, uint64_t pe,
785 				    uint64_t addr, unsigned count,
786 				    uint32_t incr, uint64_t flags)
787 {
788 	/* for physically contiguous pages (vram) */
789 	ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_GENERATE_PTE_PDE, 0, 0);
790 	ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
791 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
792 	ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
793 	ib->ptr[ib->length_dw++] = upper_32_bits(flags);
794 	ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
795 	ib->ptr[ib->length_dw++] = upper_32_bits(addr);
796 	ib->ptr[ib->length_dw++] = incr; /* increment size */
797 	ib->ptr[ib->length_dw++] = 0;
798 	ib->ptr[ib->length_dw++] = count; /* number of entries */
799 }
800 
801 /**
802  * cik_sdma_vm_pad_ib - pad the IB to the required number of dw
803  *
804  * @ib: indirect buffer to fill with padding
805  *
806  */
807 static void cik_sdma_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
808 {
809 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
810 	u32 pad_count;
811 	int i;
812 
813 	pad_count = (-ib->length_dw) & 7;
814 	for (i = 0; i < pad_count; i++)
815 		if (sdma && sdma->burst_nop && (i == 0))
816 			ib->ptr[ib->length_dw++] =
817 					SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0) |
818 					SDMA_NOP_COUNT(pad_count - 1);
819 		else
820 			ib->ptr[ib->length_dw++] =
821 					SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0);
822 }
823 
824 /**
825  * cik_sdma_ring_emit_pipeline_sync - sync the pipeline
826  *
827  * @ring: amdgpu_ring pointer
828  *
829  * Make sure all previous operations are completed (CIK).
830  */
831 static void cik_sdma_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
832 {
833 	uint32_t seq = ring->fence_drv.sync_seq;
834 	uint64_t addr = ring->fence_drv.gpu_addr;
835 
836 	/* wait for idle */
837 	amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0,
838 					    SDMA_POLL_REG_MEM_EXTRA_OP(0) |
839 					    SDMA_POLL_REG_MEM_EXTRA_FUNC(3) | /* equal */
840 					    SDMA_POLL_REG_MEM_EXTRA_M));
841 	amdgpu_ring_write(ring, addr & 0xfffffffc);
842 	amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
843 	amdgpu_ring_write(ring, seq); /* reference */
844 	amdgpu_ring_write(ring, 0xffffffff); /* mask */
845 	amdgpu_ring_write(ring, (0xfff << 16) | 4); /* retry count, poll interval */
846 }
847 
848 /**
849  * cik_sdma_ring_emit_vm_flush - cik vm flush using sDMA
850  *
851  * @ring: amdgpu_ring pointer
852  * @vm: amdgpu_vm pointer
853  *
854  * Update the page table base and flush the VM TLB
855  * using sDMA (CIK).
856  */
857 static void cik_sdma_ring_emit_vm_flush(struct amdgpu_ring *ring,
858 					unsigned vmid, uint64_t pd_addr)
859 {
860 	u32 extra_bits = (SDMA_POLL_REG_MEM_EXTRA_OP(0) |
861 			  SDMA_POLL_REG_MEM_EXTRA_FUNC(0)); /* always */
862 
863 	amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
864 
865 	amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0, extra_bits));
866 	amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST << 2);
867 	amdgpu_ring_write(ring, 0);
868 	amdgpu_ring_write(ring, 0); /* reference */
869 	amdgpu_ring_write(ring, 0); /* mask */
870 	amdgpu_ring_write(ring, (0xfff << 16) | 10); /* retry count, poll interval */
871 }
872 
873 static void cik_sdma_ring_emit_wreg(struct amdgpu_ring *ring,
874 				    uint32_t reg, uint32_t val)
875 {
876 	amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
877 	amdgpu_ring_write(ring, reg);
878 	amdgpu_ring_write(ring, val);
879 }
880 
881 static void cik_enable_sdma_mgcg(struct amdgpu_device *adev,
882 				 bool enable)
883 {
884 	u32 orig, data;
885 
886 	if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
887 		WREG32(mmSDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET, 0x00000100);
888 		WREG32(mmSDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET, 0x00000100);
889 	} else {
890 		orig = data = RREG32(mmSDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET);
891 		data |= 0xff000000;
892 		if (data != orig)
893 			WREG32(mmSDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET, data);
894 
895 		orig = data = RREG32(mmSDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET);
896 		data |= 0xff000000;
897 		if (data != orig)
898 			WREG32(mmSDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET, data);
899 	}
900 }
901 
902 static void cik_enable_sdma_mgls(struct amdgpu_device *adev,
903 				 bool enable)
904 {
905 	u32 orig, data;
906 
907 	if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) {
908 		orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET);
909 		data |= 0x100;
910 		if (orig != data)
911 			WREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET, data);
912 
913 		orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET);
914 		data |= 0x100;
915 		if (orig != data)
916 			WREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET, data);
917 	} else {
918 		orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET);
919 		data &= ~0x100;
920 		if (orig != data)
921 			WREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET, data);
922 
923 		orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET);
924 		data &= ~0x100;
925 		if (orig != data)
926 			WREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET, data);
927 	}
928 }
929 
930 static int cik_sdma_early_init(void *handle)
931 {
932 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
933 
934 	adev->sdma.num_instances = SDMA_MAX_INSTANCE;
935 
936 	cik_sdma_set_ring_funcs(adev);
937 	cik_sdma_set_irq_funcs(adev);
938 	cik_sdma_set_buffer_funcs(adev);
939 	cik_sdma_set_vm_pte_funcs(adev);
940 
941 	return 0;
942 }
943 
944 static int cik_sdma_sw_init(void *handle)
945 {
946 	struct amdgpu_ring *ring;
947 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
948 	int r, i;
949 
950 	r = cik_sdma_init_microcode(adev);
951 	if (r) {
952 		DRM_ERROR("Failed to load sdma firmware!\n");
953 		return r;
954 	}
955 
956 	/* SDMA trap event */
957 	r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 224,
958 			      &adev->sdma.trap_irq);
959 	if (r)
960 		return r;
961 
962 	/* SDMA Privileged inst */
963 	r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 241,
964 			      &adev->sdma.illegal_inst_irq);
965 	if (r)
966 		return r;
967 
968 	/* SDMA Privileged inst */
969 	r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 247,
970 			      &adev->sdma.illegal_inst_irq);
971 	if (r)
972 		return r;
973 
974 	for (i = 0; i < adev->sdma.num_instances; i++) {
975 		ring = &adev->sdma.instance[i].ring;
976 		ring->ring_obj = NULL;
977 		sprintf(ring->name, "sdma%d", i);
978 		r = amdgpu_ring_init(adev, ring, 1024,
979 				     &adev->sdma.trap_irq,
980 				     (i == 0) ?
981 				     AMDGPU_SDMA_IRQ_INSTANCE0 :
982 				     AMDGPU_SDMA_IRQ_INSTANCE1,
983 				     AMDGPU_RING_PRIO_DEFAULT);
984 		if (r)
985 			return r;
986 	}
987 
988 	return r;
989 }
990 
991 static int cik_sdma_sw_fini(void *handle)
992 {
993 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
994 	int i;
995 
996 	for (i = 0; i < adev->sdma.num_instances; i++)
997 		amdgpu_ring_fini(&adev->sdma.instance[i].ring);
998 
999 	cik_sdma_free_microcode(adev);
1000 	return 0;
1001 }
1002 
1003 static int cik_sdma_hw_init(void *handle)
1004 {
1005 	int r;
1006 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1007 
1008 	r = cik_sdma_start(adev);
1009 	if (r)
1010 		return r;
1011 
1012 	return r;
1013 }
1014 
1015 static int cik_sdma_hw_fini(void *handle)
1016 {
1017 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1018 
1019 	cik_ctx_switch_enable(adev, false);
1020 	cik_sdma_enable(adev, false);
1021 
1022 	return 0;
1023 }
1024 
1025 static int cik_sdma_suspend(void *handle)
1026 {
1027 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1028 
1029 	return cik_sdma_hw_fini(adev);
1030 }
1031 
1032 static int cik_sdma_resume(void *handle)
1033 {
1034 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1035 
1036 	cik_sdma_soft_reset(handle);
1037 
1038 	return cik_sdma_hw_init(adev);
1039 }
1040 
1041 static bool cik_sdma_is_idle(void *handle)
1042 {
1043 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1044 	u32 tmp = RREG32(mmSRBM_STATUS2);
1045 
1046 	if (tmp & (SRBM_STATUS2__SDMA_BUSY_MASK |
1047 				SRBM_STATUS2__SDMA1_BUSY_MASK))
1048 	    return false;
1049 
1050 	return true;
1051 }
1052 
1053 static int cik_sdma_wait_for_idle(void *handle)
1054 {
1055 	unsigned i;
1056 	u32 tmp;
1057 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1058 
1059 	for (i = 0; i < adev->usec_timeout; i++) {
1060 		tmp = RREG32(mmSRBM_STATUS2) & (SRBM_STATUS2__SDMA_BUSY_MASK |
1061 				SRBM_STATUS2__SDMA1_BUSY_MASK);
1062 
1063 		if (!tmp)
1064 			return 0;
1065 		udelay(1);
1066 	}
1067 	return -ETIMEDOUT;
1068 }
1069 
1070 static int cik_sdma_soft_reset(void *handle)
1071 {
1072 	u32 srbm_soft_reset = 0;
1073 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1074 	u32 tmp = RREG32(mmSRBM_STATUS2);
1075 
1076 	if (tmp & SRBM_STATUS2__SDMA_BUSY_MASK) {
1077 		/* sdma0 */
1078 		tmp = RREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET);
1079 		tmp |= SDMA0_F32_CNTL__HALT_MASK;
1080 		WREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET, tmp);
1081 		srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA_MASK;
1082 	}
1083 	if (tmp & SRBM_STATUS2__SDMA1_BUSY_MASK) {
1084 		/* sdma1 */
1085 		tmp = RREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET);
1086 		tmp |= SDMA0_F32_CNTL__HALT_MASK;
1087 		WREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET, tmp);
1088 		srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA1_MASK;
1089 	}
1090 
1091 	if (srbm_soft_reset) {
1092 		tmp = RREG32(mmSRBM_SOFT_RESET);
1093 		tmp |= srbm_soft_reset;
1094 		dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
1095 		WREG32(mmSRBM_SOFT_RESET, tmp);
1096 		tmp = RREG32(mmSRBM_SOFT_RESET);
1097 
1098 		udelay(50);
1099 
1100 		tmp &= ~srbm_soft_reset;
1101 		WREG32(mmSRBM_SOFT_RESET, tmp);
1102 		tmp = RREG32(mmSRBM_SOFT_RESET);
1103 
1104 		/* Wait a little for things to settle down */
1105 		udelay(50);
1106 	}
1107 
1108 	return 0;
1109 }
1110 
1111 static int cik_sdma_set_trap_irq_state(struct amdgpu_device *adev,
1112 				       struct amdgpu_irq_src *src,
1113 				       unsigned type,
1114 				       enum amdgpu_interrupt_state state)
1115 {
1116 	u32 sdma_cntl;
1117 
1118 	switch (type) {
1119 	case AMDGPU_SDMA_IRQ_INSTANCE0:
1120 		switch (state) {
1121 		case AMDGPU_IRQ_STATE_DISABLE:
1122 			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1123 			sdma_cntl &= ~SDMA0_CNTL__TRAP_ENABLE_MASK;
1124 			WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1125 			break;
1126 		case AMDGPU_IRQ_STATE_ENABLE:
1127 			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1128 			sdma_cntl |= SDMA0_CNTL__TRAP_ENABLE_MASK;
1129 			WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1130 			break;
1131 		default:
1132 			break;
1133 		}
1134 		break;
1135 	case AMDGPU_SDMA_IRQ_INSTANCE1:
1136 		switch (state) {
1137 		case AMDGPU_IRQ_STATE_DISABLE:
1138 			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1139 			sdma_cntl &= ~SDMA0_CNTL__TRAP_ENABLE_MASK;
1140 			WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1141 			break;
1142 		case AMDGPU_IRQ_STATE_ENABLE:
1143 			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1144 			sdma_cntl |= SDMA0_CNTL__TRAP_ENABLE_MASK;
1145 			WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1146 			break;
1147 		default:
1148 			break;
1149 		}
1150 		break;
1151 	default:
1152 		break;
1153 	}
1154 	return 0;
1155 }
1156 
1157 static int cik_sdma_process_trap_irq(struct amdgpu_device *adev,
1158 				     struct amdgpu_irq_src *source,
1159 				     struct amdgpu_iv_entry *entry)
1160 {
1161 	u8 instance_id, queue_id;
1162 
1163 	instance_id = (entry->ring_id & 0x3) >> 0;
1164 	queue_id = (entry->ring_id & 0xc) >> 2;
1165 	DRM_DEBUG("IH: SDMA trap\n");
1166 	switch (instance_id) {
1167 	case 0:
1168 		switch (queue_id) {
1169 		case 0:
1170 			amdgpu_fence_process(&adev->sdma.instance[0].ring);
1171 			break;
1172 		case 1:
1173 			/* XXX compute */
1174 			break;
1175 		case 2:
1176 			/* XXX compute */
1177 			break;
1178 		}
1179 		break;
1180 	case 1:
1181 		switch (queue_id) {
1182 		case 0:
1183 			amdgpu_fence_process(&adev->sdma.instance[1].ring);
1184 			break;
1185 		case 1:
1186 			/* XXX compute */
1187 			break;
1188 		case 2:
1189 			/* XXX compute */
1190 			break;
1191 		}
1192 		break;
1193 	}
1194 
1195 	return 0;
1196 }
1197 
1198 static int cik_sdma_process_illegal_inst_irq(struct amdgpu_device *adev,
1199 					     struct amdgpu_irq_src *source,
1200 					     struct amdgpu_iv_entry *entry)
1201 {
1202 	u8 instance_id;
1203 
1204 	DRM_ERROR("Illegal instruction in SDMA command stream\n");
1205 	instance_id = (entry->ring_id & 0x3) >> 0;
1206 	drm_sched_fault(&adev->sdma.instance[instance_id].ring.sched);
1207 	return 0;
1208 }
1209 
1210 static int cik_sdma_set_clockgating_state(void *handle,
1211 					  enum amd_clockgating_state state)
1212 {
1213 	bool gate = false;
1214 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1215 
1216 	if (state == AMD_CG_STATE_GATE)
1217 		gate = true;
1218 
1219 	cik_enable_sdma_mgcg(adev, gate);
1220 	cik_enable_sdma_mgls(adev, gate);
1221 
1222 	return 0;
1223 }
1224 
1225 static int cik_sdma_set_powergating_state(void *handle,
1226 					  enum amd_powergating_state state)
1227 {
1228 	return 0;
1229 }
1230 
1231 static const struct amd_ip_funcs cik_sdma_ip_funcs = {
1232 	.name = "cik_sdma",
1233 	.early_init = cik_sdma_early_init,
1234 	.late_init = NULL,
1235 	.sw_init = cik_sdma_sw_init,
1236 	.sw_fini = cik_sdma_sw_fini,
1237 	.hw_init = cik_sdma_hw_init,
1238 	.hw_fini = cik_sdma_hw_fini,
1239 	.suspend = cik_sdma_suspend,
1240 	.resume = cik_sdma_resume,
1241 	.is_idle = cik_sdma_is_idle,
1242 	.wait_for_idle = cik_sdma_wait_for_idle,
1243 	.soft_reset = cik_sdma_soft_reset,
1244 	.set_clockgating_state = cik_sdma_set_clockgating_state,
1245 	.set_powergating_state = cik_sdma_set_powergating_state,
1246 };
1247 
1248 static const struct amdgpu_ring_funcs cik_sdma_ring_funcs = {
1249 	.type = AMDGPU_RING_TYPE_SDMA,
1250 	.align_mask = 0xf,
1251 	.nop = SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0),
1252 	.support_64bit_ptrs = false,
1253 	.get_rptr = cik_sdma_ring_get_rptr,
1254 	.get_wptr = cik_sdma_ring_get_wptr,
1255 	.set_wptr = cik_sdma_ring_set_wptr,
1256 	.emit_frame_size =
1257 		6 + /* cik_sdma_ring_emit_hdp_flush */
1258 		3 + /* hdp invalidate */
1259 		6 + /* cik_sdma_ring_emit_pipeline_sync */
1260 		CIK_FLUSH_GPU_TLB_NUM_WREG * 3 + 6 + /* cik_sdma_ring_emit_vm_flush */
1261 		9 + 9 + 9, /* cik_sdma_ring_emit_fence x3 for user fence, vm fence */
1262 	.emit_ib_size = 7 + 4, /* cik_sdma_ring_emit_ib */
1263 	.emit_ib = cik_sdma_ring_emit_ib,
1264 	.emit_fence = cik_sdma_ring_emit_fence,
1265 	.emit_pipeline_sync = cik_sdma_ring_emit_pipeline_sync,
1266 	.emit_vm_flush = cik_sdma_ring_emit_vm_flush,
1267 	.emit_hdp_flush = cik_sdma_ring_emit_hdp_flush,
1268 	.test_ring = cik_sdma_ring_test_ring,
1269 	.test_ib = cik_sdma_ring_test_ib,
1270 	.insert_nop = cik_sdma_ring_insert_nop,
1271 	.pad_ib = cik_sdma_ring_pad_ib,
1272 	.emit_wreg = cik_sdma_ring_emit_wreg,
1273 };
1274 
1275 static void cik_sdma_set_ring_funcs(struct amdgpu_device *adev)
1276 {
1277 	int i;
1278 
1279 	for (i = 0; i < adev->sdma.num_instances; i++) {
1280 		adev->sdma.instance[i].ring.funcs = &cik_sdma_ring_funcs;
1281 		adev->sdma.instance[i].ring.me = i;
1282 	}
1283 }
1284 
1285 static const struct amdgpu_irq_src_funcs cik_sdma_trap_irq_funcs = {
1286 	.set = cik_sdma_set_trap_irq_state,
1287 	.process = cik_sdma_process_trap_irq,
1288 };
1289 
1290 static const struct amdgpu_irq_src_funcs cik_sdma_illegal_inst_irq_funcs = {
1291 	.process = cik_sdma_process_illegal_inst_irq,
1292 };
1293 
1294 static void cik_sdma_set_irq_funcs(struct amdgpu_device *adev)
1295 {
1296 	adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST;
1297 	adev->sdma.trap_irq.funcs = &cik_sdma_trap_irq_funcs;
1298 	adev->sdma.illegal_inst_irq.funcs = &cik_sdma_illegal_inst_irq_funcs;
1299 }
1300 
1301 /**
1302  * cik_sdma_emit_copy_buffer - copy buffer using the sDMA engine
1303  *
1304  * @ring: amdgpu_ring structure holding ring information
1305  * @src_offset: src GPU address
1306  * @dst_offset: dst GPU address
1307  * @byte_count: number of bytes to xfer
1308  *
1309  * Copy GPU buffers using the DMA engine (CIK).
1310  * Used by the amdgpu ttm implementation to move pages if
1311  * registered as the asic copy callback.
1312  */
1313 static void cik_sdma_emit_copy_buffer(struct amdgpu_ib *ib,
1314 				      uint64_t src_offset,
1315 				      uint64_t dst_offset,
1316 				      uint32_t byte_count,
1317 				      bool tmz)
1318 {
1319 	ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_COPY, SDMA_COPY_SUB_OPCODE_LINEAR, 0);
1320 	ib->ptr[ib->length_dw++] = byte_count;
1321 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1322 	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1323 	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1324 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1325 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1326 }
1327 
1328 /**
1329  * cik_sdma_emit_fill_buffer - fill buffer using the sDMA engine
1330  *
1331  * @ring: amdgpu_ring structure holding ring information
1332  * @src_data: value to write to buffer
1333  * @dst_offset: dst GPU address
1334  * @byte_count: number of bytes to xfer
1335  *
1336  * Fill GPU buffers using the DMA engine (CIK).
1337  */
1338 static void cik_sdma_emit_fill_buffer(struct amdgpu_ib *ib,
1339 				      uint32_t src_data,
1340 				      uint64_t dst_offset,
1341 				      uint32_t byte_count)
1342 {
1343 	ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_CONSTANT_FILL, 0, 0);
1344 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1345 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1346 	ib->ptr[ib->length_dw++] = src_data;
1347 	ib->ptr[ib->length_dw++] = byte_count;
1348 }
1349 
1350 static const struct amdgpu_buffer_funcs cik_sdma_buffer_funcs = {
1351 	.copy_max_bytes = 0x1fffff,
1352 	.copy_num_dw = 7,
1353 	.emit_copy_buffer = cik_sdma_emit_copy_buffer,
1354 
1355 	.fill_max_bytes = 0x1fffff,
1356 	.fill_num_dw = 5,
1357 	.emit_fill_buffer = cik_sdma_emit_fill_buffer,
1358 };
1359 
1360 static void cik_sdma_set_buffer_funcs(struct amdgpu_device *adev)
1361 {
1362 	adev->mman.buffer_funcs = &cik_sdma_buffer_funcs;
1363 	adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1364 }
1365 
1366 static const struct amdgpu_vm_pte_funcs cik_sdma_vm_pte_funcs = {
1367 	.copy_pte_num_dw = 7,
1368 	.copy_pte = cik_sdma_vm_copy_pte,
1369 
1370 	.write_pte = cik_sdma_vm_write_pte,
1371 	.set_pte_pde = cik_sdma_vm_set_pte_pde,
1372 };
1373 
1374 static void cik_sdma_set_vm_pte_funcs(struct amdgpu_device *adev)
1375 {
1376 	unsigned i;
1377 
1378 	adev->vm_manager.vm_pte_funcs = &cik_sdma_vm_pte_funcs;
1379 	for (i = 0; i < adev->sdma.num_instances; i++) {
1380 		adev->vm_manager.vm_pte_scheds[i] =
1381 			&adev->sdma.instance[i].ring.sched;
1382 	}
1383 	adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
1384 }
1385 
1386 const struct amdgpu_ip_block_version cik_sdma_ip_block =
1387 {
1388 	.type = AMD_IP_BLOCK_TYPE_SDMA,
1389 	.major = 2,
1390 	.minor = 0,
1391 	.rev = 0,
1392 	.funcs = &cik_sdma_ip_funcs,
1393 };
1394