xref: /openbmc/linux/drivers/gpu/drm/amd/amdgpu/cik_sdma.c (revision bbecb07f)
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 #include <linux/firmware.h>
25 #include <drm/drmP.h>
26 #include "amdgpu.h"
27 #include "amdgpu_ucode.h"
28 #include "amdgpu_trace.h"
29 #include "cikd.h"
30 #include "cik.h"
31 
32 #include "bif/bif_4_1_d.h"
33 #include "bif/bif_4_1_sh_mask.h"
34 
35 #include "gca/gfx_7_2_d.h"
36 #include "gca/gfx_7_2_enum.h"
37 #include "gca/gfx_7_2_sh_mask.h"
38 
39 #include "gmc/gmc_7_1_d.h"
40 #include "gmc/gmc_7_1_sh_mask.h"
41 
42 #include "oss/oss_2_0_d.h"
43 #include "oss/oss_2_0_sh_mask.h"
44 
45 static const u32 sdma_offsets[SDMA_MAX_INSTANCE] =
46 {
47 	SDMA0_REGISTER_OFFSET,
48 	SDMA1_REGISTER_OFFSET
49 };
50 
51 static void cik_sdma_set_ring_funcs(struct amdgpu_device *adev);
52 static void cik_sdma_set_irq_funcs(struct amdgpu_device *adev);
53 static void cik_sdma_set_buffer_funcs(struct amdgpu_device *adev);
54 static void cik_sdma_set_vm_pte_funcs(struct amdgpu_device *adev);
55 static int cik_sdma_soft_reset(void *handle);
56 
57 MODULE_FIRMWARE("radeon/bonaire_sdma.bin");
58 MODULE_FIRMWARE("radeon/bonaire_sdma1.bin");
59 MODULE_FIRMWARE("radeon/hawaii_sdma.bin");
60 MODULE_FIRMWARE("radeon/hawaii_sdma1.bin");
61 MODULE_FIRMWARE("radeon/kaveri_sdma.bin");
62 MODULE_FIRMWARE("radeon/kaveri_sdma1.bin");
63 MODULE_FIRMWARE("radeon/kabini_sdma.bin");
64 MODULE_FIRMWARE("radeon/kabini_sdma1.bin");
65 MODULE_FIRMWARE("radeon/mullins_sdma.bin");
66 MODULE_FIRMWARE("radeon/mullins_sdma1.bin");
67 
68 u32 amdgpu_cik_gpu_check_soft_reset(struct amdgpu_device *adev);
69 
70 
71 static void cik_sdma_free_microcode(struct amdgpu_device *adev)
72 {
73 	int i;
74 	for (i = 0; i < adev->sdma.num_instances; i++) {
75 			release_firmware(adev->sdma.instance[i].fw);
76 			adev->sdma.instance[i].fw = NULL;
77 	}
78 }
79 
80 /*
81  * sDMA - System DMA
82  * Starting with CIK, the GPU has new asynchronous
83  * DMA engines.  These engines are used for compute
84  * and gfx.  There are two DMA engines (SDMA0, SDMA1)
85  * and each one supports 1 ring buffer used for gfx
86  * and 2 queues used for compute.
87  *
88  * The programming model is very similar to the CP
89  * (ring buffer, IBs, etc.), but sDMA has it's own
90  * packet format that is different from the PM4 format
91  * used by the CP. sDMA supports copying data, writing
92  * embedded data, solid fills, and a number of other
93  * things.  It also has support for tiling/detiling of
94  * buffers.
95  */
96 
97 /**
98  * cik_sdma_init_microcode - load ucode images from disk
99  *
100  * @adev: amdgpu_device pointer
101  *
102  * Use the firmware interface to load the ucode images into
103  * the driver (not loaded into hw).
104  * Returns 0 on success, error on failure.
105  */
106 static int cik_sdma_init_microcode(struct amdgpu_device *adev)
107 {
108 	const char *chip_name;
109 	char fw_name[30];
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 			snprintf(fw_name, sizeof(fw_name), "radeon/%s_sdma.bin", chip_name);
136 		else
137 			snprintf(fw_name, sizeof(fw_name), "radeon/%s_sdma1.bin", chip_name);
138 		err = request_firmware(&adev->sdma.instance[i].fw, fw_name, adev->dev);
139 		if (err)
140 			goto out;
141 		err = amdgpu_ucode_validate(adev->sdma.instance[i].fw);
142 	}
143 out:
144 	if (err) {
145 		pr_err("cik_sdma: Failed to load firmware \"%s\"\n", fw_name);
146 		for (i = 0; i < adev->sdma.num_instances; i++) {
147 			release_firmware(adev->sdma.instance[i].fw);
148 			adev->sdma.instance[i].fw = NULL;
149 		}
150 	}
151 	return err;
152 }
153 
154 /**
155  * cik_sdma_ring_get_rptr - get the current read pointer
156  *
157  * @ring: amdgpu ring pointer
158  *
159  * Get the current rptr from the hardware (CIK+).
160  */
161 static uint64_t cik_sdma_ring_get_rptr(struct amdgpu_ring *ring)
162 {
163 	u32 rptr;
164 
165 	rptr = ring->adev->wb.wb[ring->rptr_offs];
166 
167 	return (rptr & 0x3fffc) >> 2;
168 }
169 
170 /**
171  * cik_sdma_ring_get_wptr - get the current write pointer
172  *
173  * @ring: amdgpu ring pointer
174  *
175  * Get the current wptr from the hardware (CIK+).
176  */
177 static uint64_t cik_sdma_ring_get_wptr(struct amdgpu_ring *ring)
178 {
179 	struct amdgpu_device *adev = ring->adev;
180 	u32 me = (ring == &adev->sdma.instance[0].ring) ? 0 : 1;
181 
182 	return (RREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[me]) & 0x3fffc) >> 2;
183 }
184 
185 /**
186  * cik_sdma_ring_set_wptr - commit the write pointer
187  *
188  * @ring: amdgpu ring pointer
189  *
190  * Write the wptr back to the hardware (CIK+).
191  */
192 static void cik_sdma_ring_set_wptr(struct amdgpu_ring *ring)
193 {
194 	struct amdgpu_device *adev = ring->adev;
195 	u32 me = (ring == &adev->sdma.instance[0].ring) ? 0 : 1;
196 
197 	WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[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_get_sdma_instance(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_ib *ib,
224 				  unsigned vm_id, bool ctx_switch)
225 {
226 	u32 extra_bits = vm_id & 0xf;
227 
228 	/* IB packet must end on a 8 DW boundary */
229 	cik_sdma_ring_insert_nop(ring, (12 - (lower_32_bits(ring->wptr) & 7)) % 8);
230 
231 	amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_INDIRECT_BUFFER, 0, extra_bits));
232 	amdgpu_ring_write(ring, ib->gpu_addr & 0xffffffe0); /* base must be 32 byte aligned */
233 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr) & 0xffffffff);
234 	amdgpu_ring_write(ring, ib->length_dw);
235 
236 }
237 
238 /**
239  * cik_sdma_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
240  *
241  * @ring: amdgpu ring pointer
242  *
243  * Emit an hdp flush packet on the requested DMA ring.
244  */
245 static void cik_sdma_ring_emit_hdp_flush(struct amdgpu_ring *ring)
246 {
247 	u32 extra_bits = (SDMA_POLL_REG_MEM_EXTRA_OP(1) |
248 			  SDMA_POLL_REG_MEM_EXTRA_FUNC(3)); /* == */
249 	u32 ref_and_mask;
250 
251 	if (ring == &ring->adev->sdma.instance[0].ring)
252 		ref_and_mask = GPU_HDP_FLUSH_DONE__SDMA0_MASK;
253 	else
254 		ref_and_mask = GPU_HDP_FLUSH_DONE__SDMA1_MASK;
255 
256 	amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0, extra_bits));
257 	amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_DONE << 2);
258 	amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_REQ << 2);
259 	amdgpu_ring_write(ring, ref_and_mask); /* reference */
260 	amdgpu_ring_write(ring, ref_and_mask); /* mask */
261 	amdgpu_ring_write(ring, (0xfff << 16) | 10); /* retry count, poll interval */
262 }
263 
264 static void cik_sdma_ring_emit_hdp_invalidate(struct amdgpu_ring *ring)
265 {
266 	amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
267 	amdgpu_ring_write(ring, mmHDP_DEBUG0);
268 	amdgpu_ring_write(ring, 1);
269 }
270 
271 /**
272  * cik_sdma_ring_emit_fence - emit a fence on the DMA ring
273  *
274  * @ring: amdgpu ring pointer
275  * @fence: amdgpu fence object
276  *
277  * Add a DMA fence packet to the ring to write
278  * the fence seq number and DMA trap packet to generate
279  * an interrupt if needed (CIK).
280  */
281 static void cik_sdma_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
282 				     unsigned flags)
283 {
284 	bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
285 	/* write the fence */
286 	amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_FENCE, 0, 0));
287 	amdgpu_ring_write(ring, lower_32_bits(addr));
288 	amdgpu_ring_write(ring, upper_32_bits(addr));
289 	amdgpu_ring_write(ring, lower_32_bits(seq));
290 
291 	/* optionally write high bits as well */
292 	if (write64bit) {
293 		addr += 4;
294 		amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_FENCE, 0, 0));
295 		amdgpu_ring_write(ring, lower_32_bits(addr));
296 		amdgpu_ring_write(ring, upper_32_bits(addr));
297 		amdgpu_ring_write(ring, upper_32_bits(seq));
298 	}
299 
300 	/* generate an interrupt */
301 	amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_TRAP, 0, 0));
302 }
303 
304 /**
305  * cik_sdma_gfx_stop - stop the gfx async dma engines
306  *
307  * @adev: amdgpu_device pointer
308  *
309  * Stop the gfx async dma ring buffers (CIK).
310  */
311 static void cik_sdma_gfx_stop(struct amdgpu_device *adev)
312 {
313 	struct amdgpu_ring *sdma0 = &adev->sdma.instance[0].ring;
314 	struct amdgpu_ring *sdma1 = &adev->sdma.instance[1].ring;
315 	u32 rb_cntl;
316 	int i;
317 
318 	if ((adev->mman.buffer_funcs_ring == sdma0) ||
319 	    (adev->mman.buffer_funcs_ring == sdma1))
320 		amdgpu_ttm_set_active_vram_size(adev, adev->mc.visible_vram_size);
321 
322 	for (i = 0; i < adev->sdma.num_instances; i++) {
323 		rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
324 		rb_cntl &= ~SDMA0_GFX_RB_CNTL__RB_ENABLE_MASK;
325 		WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
326 		WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], 0);
327 	}
328 	sdma0->ready = false;
329 	sdma1->ready = false;
330 }
331 
332 /**
333  * cik_sdma_rlc_stop - stop the compute async dma engines
334  *
335  * @adev: amdgpu_device pointer
336  *
337  * Stop the compute async dma queues (CIK).
338  */
339 static void cik_sdma_rlc_stop(struct amdgpu_device *adev)
340 {
341 	/* XXX todo */
342 }
343 
344 /**
345  * cik_ctx_switch_enable - stop the async dma engines context switch
346  *
347  * @adev: amdgpu_device pointer
348  * @enable: enable/disable the DMA MEs context switch.
349  *
350  * Halt or unhalt the async dma engines context switch (VI).
351  */
352 static void cik_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
353 {
354 	u32 f32_cntl, phase_quantum = 0;
355 	int i;
356 
357 	if (amdgpu_sdma_phase_quantum) {
358 		unsigned value = amdgpu_sdma_phase_quantum;
359 		unsigned unit = 0;
360 
361 		while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
362 				SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) {
363 			value = (value + 1) >> 1;
364 			unit++;
365 		}
366 		if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
367 			    SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) {
368 			value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
369 				 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT);
370 			unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
371 				SDMA0_PHASE0_QUANTUM__UNIT__SHIFT);
372 			WARN_ONCE(1,
373 			"clamping sdma_phase_quantum to %uK clock cycles\n",
374 				  value << unit);
375 		}
376 		phase_quantum =
377 			value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
378 			unit  << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
379 	}
380 
381 	for (i = 0; i < adev->sdma.num_instances; i++) {
382 		f32_cntl = RREG32(mmSDMA0_CNTL + sdma_offsets[i]);
383 		if (enable) {
384 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
385 					AUTO_CTXSW_ENABLE, 1);
386 			if (amdgpu_sdma_phase_quantum) {
387 				WREG32(mmSDMA0_PHASE0_QUANTUM + sdma_offsets[i],
388 				       phase_quantum);
389 				WREG32(mmSDMA0_PHASE1_QUANTUM + sdma_offsets[i],
390 				       phase_quantum);
391 			}
392 		} else {
393 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
394 					AUTO_CTXSW_ENABLE, 0);
395 		}
396 
397 		WREG32(mmSDMA0_CNTL + sdma_offsets[i], f32_cntl);
398 	}
399 }
400 
401 /**
402  * cik_sdma_enable - stop the async dma engines
403  *
404  * @adev: amdgpu_device pointer
405  * @enable: enable/disable the DMA MEs.
406  *
407  * Halt or unhalt the async dma engines (CIK).
408  */
409 static void cik_sdma_enable(struct amdgpu_device *adev, bool enable)
410 {
411 	u32 me_cntl;
412 	int i;
413 
414 	if (!enable) {
415 		cik_sdma_gfx_stop(adev);
416 		cik_sdma_rlc_stop(adev);
417 	}
418 
419 	for (i = 0; i < adev->sdma.num_instances; i++) {
420 		me_cntl = RREG32(mmSDMA0_F32_CNTL + sdma_offsets[i]);
421 		if (enable)
422 			me_cntl &= ~SDMA0_F32_CNTL__HALT_MASK;
423 		else
424 			me_cntl |= SDMA0_F32_CNTL__HALT_MASK;
425 		WREG32(mmSDMA0_F32_CNTL + sdma_offsets[i], me_cntl);
426 	}
427 }
428 
429 /**
430  * cik_sdma_gfx_resume - setup and start the async dma engines
431  *
432  * @adev: amdgpu_device pointer
433  *
434  * Set up the gfx DMA ring buffers and enable them (CIK).
435  * Returns 0 for success, error for failure.
436  */
437 static int cik_sdma_gfx_resume(struct amdgpu_device *adev)
438 {
439 	struct amdgpu_ring *ring;
440 	u32 rb_cntl, ib_cntl;
441 	u32 rb_bufsz;
442 	u32 wb_offset;
443 	int i, j, r;
444 
445 	for (i = 0; i < adev->sdma.num_instances; i++) {
446 		ring = &adev->sdma.instance[i].ring;
447 		wb_offset = (ring->rptr_offs * 4);
448 
449 		mutex_lock(&adev->srbm_mutex);
450 		for (j = 0; j < 16; j++) {
451 			cik_srbm_select(adev, 0, 0, 0, j);
452 			/* SDMA GFX */
453 			WREG32(mmSDMA0_GFX_VIRTUAL_ADDR + sdma_offsets[i], 0);
454 			WREG32(mmSDMA0_GFX_APE1_CNTL + sdma_offsets[i], 0);
455 			/* XXX SDMA RLC - todo */
456 		}
457 		cik_srbm_select(adev, 0, 0, 0, 0);
458 		mutex_unlock(&adev->srbm_mutex);
459 
460 		WREG32(mmSDMA0_TILING_CONFIG + sdma_offsets[i],
461 		       adev->gfx.config.gb_addr_config & 0x70);
462 
463 		WREG32(mmSDMA0_SEM_INCOMPLETE_TIMER_CNTL + sdma_offsets[i], 0);
464 		WREG32(mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i], 0);
465 
466 		/* Set ring buffer size in dwords */
467 		rb_bufsz = order_base_2(ring->ring_size / 4);
468 		rb_cntl = rb_bufsz << 1;
469 #ifdef __BIG_ENDIAN
470 		rb_cntl |= SDMA0_GFX_RB_CNTL__RB_SWAP_ENABLE_MASK |
471 			SDMA0_GFX_RB_CNTL__RPTR_WRITEBACK_SWAP_ENABLE_MASK;
472 #endif
473 		WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
474 
475 		/* Initialize the ring buffer's read and write pointers */
476 		WREG32(mmSDMA0_GFX_RB_RPTR + sdma_offsets[i], 0);
477 		WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], 0);
478 		WREG32(mmSDMA0_GFX_IB_RPTR + sdma_offsets[i], 0);
479 		WREG32(mmSDMA0_GFX_IB_OFFSET + sdma_offsets[i], 0);
480 
481 		/* set the wb address whether it's enabled or not */
482 		WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_HI + sdma_offsets[i],
483 		       upper_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFF);
484 		WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_LO + sdma_offsets[i],
485 		       ((adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC));
486 
487 		rb_cntl |= SDMA0_GFX_RB_CNTL__RPTR_WRITEBACK_ENABLE_MASK;
488 
489 		WREG32(mmSDMA0_GFX_RB_BASE + sdma_offsets[i], ring->gpu_addr >> 8);
490 		WREG32(mmSDMA0_GFX_RB_BASE_HI + sdma_offsets[i], ring->gpu_addr >> 40);
491 
492 		ring->wptr = 0;
493 		WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], lower_32_bits(ring->wptr) << 2);
494 
495 		/* enable DMA RB */
496 		WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i],
497 		       rb_cntl | SDMA0_GFX_RB_CNTL__RB_ENABLE_MASK);
498 
499 		ib_cntl = SDMA0_GFX_IB_CNTL__IB_ENABLE_MASK;
500 #ifdef __BIG_ENDIAN
501 		ib_cntl |= SDMA0_GFX_IB_CNTL__IB_SWAP_ENABLE_MASK;
502 #endif
503 		/* enable DMA IBs */
504 		WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
505 
506 		ring->ready = true;
507 	}
508 
509 	cik_sdma_enable(adev, true);
510 
511 	for (i = 0; i < adev->sdma.num_instances; i++) {
512 		ring = &adev->sdma.instance[i].ring;
513 		r = amdgpu_ring_test_ring(ring);
514 		if (r) {
515 			ring->ready = false;
516 			return r;
517 		}
518 
519 		if (adev->mman.buffer_funcs_ring == ring)
520 			amdgpu_ttm_set_active_vram_size(adev, adev->mc.real_vram_size);
521 	}
522 
523 	return 0;
524 }
525 
526 /**
527  * cik_sdma_rlc_resume - setup and start the async dma engines
528  *
529  * @adev: amdgpu_device pointer
530  *
531  * Set up the compute DMA queues and enable them (CIK).
532  * Returns 0 for success, error for failure.
533  */
534 static int cik_sdma_rlc_resume(struct amdgpu_device *adev)
535 {
536 	/* XXX todo */
537 	return 0;
538 }
539 
540 /**
541  * cik_sdma_load_microcode - load the sDMA ME ucode
542  *
543  * @adev: amdgpu_device pointer
544  *
545  * Loads the sDMA0/1 ucode.
546  * Returns 0 for success, -EINVAL if the ucode is not available.
547  */
548 static int cik_sdma_load_microcode(struct amdgpu_device *adev)
549 {
550 	const struct sdma_firmware_header_v1_0 *hdr;
551 	const __le32 *fw_data;
552 	u32 fw_size;
553 	int i, j;
554 
555 	/* halt the MEs */
556 	cik_sdma_enable(adev, false);
557 
558 	for (i = 0; i < adev->sdma.num_instances; i++) {
559 		if (!adev->sdma.instance[i].fw)
560 			return -EINVAL;
561 		hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
562 		amdgpu_ucode_print_sdma_hdr(&hdr->header);
563 		fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
564 		adev->sdma.instance[i].fw_version = le32_to_cpu(hdr->header.ucode_version);
565 		adev->sdma.instance[i].feature_version = le32_to_cpu(hdr->ucode_feature_version);
566 		if (adev->sdma.instance[i].feature_version >= 20)
567 			adev->sdma.instance[i].burst_nop = true;
568 		fw_data = (const __le32 *)
569 			(adev->sdma.instance[i].fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
570 		WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], 0);
571 		for (j = 0; j < fw_size; j++)
572 			WREG32(mmSDMA0_UCODE_DATA + sdma_offsets[i], le32_to_cpup(fw_data++));
573 		WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], adev->sdma.instance[i].fw_version);
574 	}
575 
576 	return 0;
577 }
578 
579 /**
580  * cik_sdma_start - setup and start the async dma engines
581  *
582  * @adev: amdgpu_device pointer
583  *
584  * Set up the DMA engines and enable them (CIK).
585  * Returns 0 for success, error for failure.
586  */
587 static int cik_sdma_start(struct amdgpu_device *adev)
588 {
589 	int r;
590 
591 	r = cik_sdma_load_microcode(adev);
592 	if (r)
593 		return r;
594 
595 	/* halt the engine before programing */
596 	cik_sdma_enable(adev, false);
597 	/* enable sdma ring preemption */
598 	cik_ctx_switch_enable(adev, true);
599 
600 	/* start the gfx rings and rlc compute queues */
601 	r = cik_sdma_gfx_resume(adev);
602 	if (r)
603 		return r;
604 	r = cik_sdma_rlc_resume(adev);
605 	if (r)
606 		return r;
607 
608 	return 0;
609 }
610 
611 /**
612  * cik_sdma_ring_test_ring - simple async dma engine test
613  *
614  * @ring: amdgpu_ring structure holding ring information
615  *
616  * Test the DMA engine by writing using it to write an
617  * value to memory. (CIK).
618  * Returns 0 for success, error for failure.
619  */
620 static int cik_sdma_ring_test_ring(struct amdgpu_ring *ring)
621 {
622 	struct amdgpu_device *adev = ring->adev;
623 	unsigned i;
624 	unsigned index;
625 	int r;
626 	u32 tmp;
627 	u64 gpu_addr;
628 
629 	r = amdgpu_wb_get(adev, &index);
630 	if (r) {
631 		dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
632 		return r;
633 	}
634 
635 	gpu_addr = adev->wb.gpu_addr + (index * 4);
636 	tmp = 0xCAFEDEAD;
637 	adev->wb.wb[index] = cpu_to_le32(tmp);
638 
639 	r = amdgpu_ring_alloc(ring, 5);
640 	if (r) {
641 		DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
642 		amdgpu_wb_free(adev, index);
643 		return r;
644 	}
645 	amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_WRITE, SDMA_WRITE_SUB_OPCODE_LINEAR, 0));
646 	amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
647 	amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
648 	amdgpu_ring_write(ring, 1); /* number of DWs to follow */
649 	amdgpu_ring_write(ring, 0xDEADBEEF);
650 	amdgpu_ring_commit(ring);
651 
652 	for (i = 0; i < adev->usec_timeout; i++) {
653 		tmp = le32_to_cpu(adev->wb.wb[index]);
654 		if (tmp == 0xDEADBEEF)
655 			break;
656 		DRM_UDELAY(1);
657 	}
658 
659 	if (i < adev->usec_timeout) {
660 		DRM_INFO("ring test on %d succeeded in %d usecs\n", ring->idx, i);
661 	} else {
662 		DRM_ERROR("amdgpu: ring %d test failed (0x%08X)\n",
663 			  ring->idx, tmp);
664 		r = -EINVAL;
665 	}
666 	amdgpu_wb_free(adev, index);
667 
668 	return r;
669 }
670 
671 /**
672  * cik_sdma_ring_test_ib - test an IB on the DMA engine
673  *
674  * @ring: amdgpu_ring structure holding ring information
675  *
676  * Test a simple IB in the DMA ring (CIK).
677  * Returns 0 on success, error on failure.
678  */
679 static int cik_sdma_ring_test_ib(struct amdgpu_ring *ring, long timeout)
680 {
681 	struct amdgpu_device *adev = ring->adev;
682 	struct amdgpu_ib ib;
683 	struct dma_fence *f = NULL;
684 	unsigned index;
685 	u32 tmp = 0;
686 	u64 gpu_addr;
687 	long r;
688 
689 	r = amdgpu_wb_get(adev, &index);
690 	if (r) {
691 		dev_err(adev->dev, "(%ld) failed to allocate wb slot\n", r);
692 		return r;
693 	}
694 
695 	gpu_addr = adev->wb.gpu_addr + (index * 4);
696 	tmp = 0xCAFEDEAD;
697 	adev->wb.wb[index] = cpu_to_le32(tmp);
698 	memset(&ib, 0, sizeof(ib));
699 	r = amdgpu_ib_get(adev, NULL, 256, &ib);
700 	if (r) {
701 		DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r);
702 		goto err0;
703 	}
704 
705 	ib.ptr[0] = SDMA_PACKET(SDMA_OPCODE_WRITE,
706 				SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
707 	ib.ptr[1] = lower_32_bits(gpu_addr);
708 	ib.ptr[2] = upper_32_bits(gpu_addr);
709 	ib.ptr[3] = 1;
710 	ib.ptr[4] = 0xDEADBEEF;
711 	ib.length_dw = 5;
712 	r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
713 	if (r)
714 		goto err1;
715 
716 	r = dma_fence_wait_timeout(f, false, timeout);
717 	if (r == 0) {
718 		DRM_ERROR("amdgpu: IB test timed out\n");
719 		r = -ETIMEDOUT;
720 		goto err1;
721 	} else if (r < 0) {
722 		DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
723 		goto err1;
724 	}
725 	tmp = le32_to_cpu(adev->wb.wb[index]);
726 	if (tmp == 0xDEADBEEF) {
727 		DRM_INFO("ib test on ring %d succeeded\n", ring->idx);
728 		r = 0;
729 	} else {
730 		DRM_ERROR("amdgpu: ib test failed (0x%08X)\n", tmp);
731 		r = -EINVAL;
732 	}
733 
734 err1:
735 	amdgpu_ib_free(adev, &ib, NULL);
736 	dma_fence_put(f);
737 err0:
738 	amdgpu_wb_free(adev, index);
739 	return r;
740 }
741 
742 /**
743  * cik_sdma_vm_copy_pages - update PTEs by copying them from the GART
744  *
745  * @ib: indirect buffer to fill with commands
746  * @pe: addr of the page entry
747  * @src: src addr to copy from
748  * @count: number of page entries to update
749  *
750  * Update PTEs by copying them from the GART using sDMA (CIK).
751  */
752 static void cik_sdma_vm_copy_pte(struct amdgpu_ib *ib,
753 				 uint64_t pe, uint64_t src,
754 				 unsigned count)
755 {
756 	unsigned bytes = count * 8;
757 
758 	ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_COPY,
759 		SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
760 	ib->ptr[ib->length_dw++] = bytes;
761 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
762 	ib->ptr[ib->length_dw++] = lower_32_bits(src);
763 	ib->ptr[ib->length_dw++] = upper_32_bits(src);
764 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
765 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
766 }
767 
768 /**
769  * cik_sdma_vm_write_pages - update PTEs by writing them manually
770  *
771  * @ib: indirect buffer to fill with commands
772  * @pe: addr of the page entry
773  * @value: dst addr to write into pe
774  * @count: number of page entries to update
775  * @incr: increase next addr by incr bytes
776  *
777  * Update PTEs by writing them manually using sDMA (CIK).
778  */
779 static void cik_sdma_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
780 				  uint64_t value, unsigned count,
781 				  uint32_t incr)
782 {
783 	unsigned ndw = count * 2;
784 
785 	ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_WRITE,
786 		SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
787 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
788 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
789 	ib->ptr[ib->length_dw++] = ndw;
790 	for (; ndw > 0; ndw -= 2) {
791 		ib->ptr[ib->length_dw++] = lower_32_bits(value);
792 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
793 		value += incr;
794 	}
795 }
796 
797 /**
798  * cik_sdma_vm_set_pages - update the page tables using sDMA
799  *
800  * @ib: indirect buffer to fill with commands
801  * @pe: addr of the page entry
802  * @addr: dst addr to write into pe
803  * @count: number of page entries to update
804  * @incr: increase next addr by incr bytes
805  * @flags: access flags
806  *
807  * Update the page tables using sDMA (CIK).
808  */
809 static void cik_sdma_vm_set_pte_pde(struct amdgpu_ib *ib, uint64_t pe,
810 				    uint64_t addr, unsigned count,
811 				    uint32_t incr, uint64_t flags)
812 {
813 	/* for physically contiguous pages (vram) */
814 	ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_GENERATE_PTE_PDE, 0, 0);
815 	ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
816 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
817 	ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
818 	ib->ptr[ib->length_dw++] = upper_32_bits(flags);
819 	ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
820 	ib->ptr[ib->length_dw++] = upper_32_bits(addr);
821 	ib->ptr[ib->length_dw++] = incr; /* increment size */
822 	ib->ptr[ib->length_dw++] = 0;
823 	ib->ptr[ib->length_dw++] = count; /* number of entries */
824 }
825 
826 /**
827  * cik_sdma_vm_pad_ib - pad the IB to the required number of dw
828  *
829  * @ib: indirect buffer to fill with padding
830  *
831  */
832 static void cik_sdma_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
833 {
834 	struct amdgpu_sdma_instance *sdma = amdgpu_get_sdma_instance(ring);
835 	u32 pad_count;
836 	int i;
837 
838 	pad_count = (8 - (ib->length_dw & 0x7)) % 8;
839 	for (i = 0; i < pad_count; i++)
840 		if (sdma && sdma->burst_nop && (i == 0))
841 			ib->ptr[ib->length_dw++] =
842 					SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0) |
843 					SDMA_NOP_COUNT(pad_count - 1);
844 		else
845 			ib->ptr[ib->length_dw++] =
846 					SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0);
847 }
848 
849 /**
850  * cik_sdma_ring_emit_pipeline_sync - sync the pipeline
851  *
852  * @ring: amdgpu_ring pointer
853  *
854  * Make sure all previous operations are completed (CIK).
855  */
856 static void cik_sdma_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
857 {
858 	uint32_t seq = ring->fence_drv.sync_seq;
859 	uint64_t addr = ring->fence_drv.gpu_addr;
860 
861 	/* wait for idle */
862 	amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0,
863 					    SDMA_POLL_REG_MEM_EXTRA_OP(0) |
864 					    SDMA_POLL_REG_MEM_EXTRA_FUNC(3) | /* equal */
865 					    SDMA_POLL_REG_MEM_EXTRA_M));
866 	amdgpu_ring_write(ring, addr & 0xfffffffc);
867 	amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
868 	amdgpu_ring_write(ring, seq); /* reference */
869 	amdgpu_ring_write(ring, 0xfffffff); /* mask */
870 	amdgpu_ring_write(ring, (0xfff << 16) | 4); /* retry count, poll interval */
871 }
872 
873 /**
874  * cik_sdma_ring_emit_vm_flush - cik vm flush using sDMA
875  *
876  * @ring: amdgpu_ring pointer
877  * @vm: amdgpu_vm pointer
878  *
879  * Update the page table base and flush the VM TLB
880  * using sDMA (CIK).
881  */
882 static void cik_sdma_ring_emit_vm_flush(struct amdgpu_ring *ring,
883 					unsigned vm_id, uint64_t pd_addr)
884 {
885 	u32 extra_bits = (SDMA_POLL_REG_MEM_EXTRA_OP(0) |
886 			  SDMA_POLL_REG_MEM_EXTRA_FUNC(0)); /* always */
887 
888 	amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
889 	if (vm_id < 8) {
890 		amdgpu_ring_write(ring, (mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + vm_id));
891 	} else {
892 		amdgpu_ring_write(ring, (mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + vm_id - 8));
893 	}
894 	amdgpu_ring_write(ring, pd_addr >> 12);
895 
896 	/* flush TLB */
897 	amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
898 	amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST);
899 	amdgpu_ring_write(ring, 1 << vm_id);
900 
901 	amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0, extra_bits));
902 	amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST << 2);
903 	amdgpu_ring_write(ring, 0);
904 	amdgpu_ring_write(ring, 0); /* reference */
905 	amdgpu_ring_write(ring, 0); /* mask */
906 	amdgpu_ring_write(ring, (0xfff << 16) | 10); /* retry count, poll interval */
907 }
908 
909 static void cik_enable_sdma_mgcg(struct amdgpu_device *adev,
910 				 bool enable)
911 {
912 	u32 orig, data;
913 
914 	if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
915 		WREG32(mmSDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET, 0x00000100);
916 		WREG32(mmSDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET, 0x00000100);
917 	} else {
918 		orig = data = RREG32(mmSDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET);
919 		data |= 0xff000000;
920 		if (data != orig)
921 			WREG32(mmSDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET, data);
922 
923 		orig = data = RREG32(mmSDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET);
924 		data |= 0xff000000;
925 		if (data != orig)
926 			WREG32(mmSDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET, data);
927 	}
928 }
929 
930 static void cik_enable_sdma_mgls(struct amdgpu_device *adev,
931 				 bool enable)
932 {
933 	u32 orig, data;
934 
935 	if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) {
936 		orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET);
937 		data |= 0x100;
938 		if (orig != data)
939 			WREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET, data);
940 
941 		orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET);
942 		data |= 0x100;
943 		if (orig != data)
944 			WREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET, data);
945 	} else {
946 		orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET);
947 		data &= ~0x100;
948 		if (orig != data)
949 			WREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET, data);
950 
951 		orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET);
952 		data &= ~0x100;
953 		if (orig != data)
954 			WREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET, data);
955 	}
956 }
957 
958 static int cik_sdma_early_init(void *handle)
959 {
960 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
961 
962 	adev->sdma.num_instances = SDMA_MAX_INSTANCE;
963 
964 	cik_sdma_set_ring_funcs(adev);
965 	cik_sdma_set_irq_funcs(adev);
966 	cik_sdma_set_buffer_funcs(adev);
967 	cik_sdma_set_vm_pte_funcs(adev);
968 
969 	return 0;
970 }
971 
972 static int cik_sdma_sw_init(void *handle)
973 {
974 	struct amdgpu_ring *ring;
975 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
976 	int r, i;
977 
978 	r = cik_sdma_init_microcode(adev);
979 	if (r) {
980 		DRM_ERROR("Failed to load sdma firmware!\n");
981 		return r;
982 	}
983 
984 	/* SDMA trap event */
985 	r = amdgpu_irq_add_id(adev, AMDGPU_IH_CLIENTID_LEGACY, 224,
986 			      &adev->sdma.trap_irq);
987 	if (r)
988 		return r;
989 
990 	/* SDMA Privileged inst */
991 	r = amdgpu_irq_add_id(adev, AMDGPU_IH_CLIENTID_LEGACY, 241,
992 			      &adev->sdma.illegal_inst_irq);
993 	if (r)
994 		return r;
995 
996 	/* SDMA Privileged inst */
997 	r = amdgpu_irq_add_id(adev, AMDGPU_IH_CLIENTID_LEGACY, 247,
998 			      &adev->sdma.illegal_inst_irq);
999 	if (r)
1000 		return r;
1001 
1002 	for (i = 0; i < adev->sdma.num_instances; i++) {
1003 		ring = &adev->sdma.instance[i].ring;
1004 		ring->ring_obj = NULL;
1005 		sprintf(ring->name, "sdma%d", i);
1006 		r = amdgpu_ring_init(adev, ring, 1024,
1007 				     &adev->sdma.trap_irq,
1008 				     (i == 0) ?
1009 				     AMDGPU_SDMA_IRQ_TRAP0 :
1010 				     AMDGPU_SDMA_IRQ_TRAP1);
1011 		if (r)
1012 			return r;
1013 	}
1014 
1015 	return r;
1016 }
1017 
1018 static int cik_sdma_sw_fini(void *handle)
1019 {
1020 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1021 	int i;
1022 
1023 	for (i = 0; i < adev->sdma.num_instances; i++)
1024 		amdgpu_ring_fini(&adev->sdma.instance[i].ring);
1025 
1026 	cik_sdma_free_microcode(adev);
1027 	return 0;
1028 }
1029 
1030 static int cik_sdma_hw_init(void *handle)
1031 {
1032 	int r;
1033 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1034 
1035 	r = cik_sdma_start(adev);
1036 	if (r)
1037 		return r;
1038 
1039 	return r;
1040 }
1041 
1042 static int cik_sdma_hw_fini(void *handle)
1043 {
1044 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1045 
1046 	cik_ctx_switch_enable(adev, false);
1047 	cik_sdma_enable(adev, false);
1048 
1049 	return 0;
1050 }
1051 
1052 static int cik_sdma_suspend(void *handle)
1053 {
1054 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1055 
1056 	return cik_sdma_hw_fini(adev);
1057 }
1058 
1059 static int cik_sdma_resume(void *handle)
1060 {
1061 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1062 
1063 	cik_sdma_soft_reset(handle);
1064 
1065 	return cik_sdma_hw_init(adev);
1066 }
1067 
1068 static bool cik_sdma_is_idle(void *handle)
1069 {
1070 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1071 	u32 tmp = RREG32(mmSRBM_STATUS2);
1072 
1073 	if (tmp & (SRBM_STATUS2__SDMA_BUSY_MASK |
1074 				SRBM_STATUS2__SDMA1_BUSY_MASK))
1075 	    return false;
1076 
1077 	return true;
1078 }
1079 
1080 static int cik_sdma_wait_for_idle(void *handle)
1081 {
1082 	unsigned i;
1083 	u32 tmp;
1084 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1085 
1086 	for (i = 0; i < adev->usec_timeout; i++) {
1087 		tmp = RREG32(mmSRBM_STATUS2) & (SRBM_STATUS2__SDMA_BUSY_MASK |
1088 				SRBM_STATUS2__SDMA1_BUSY_MASK);
1089 
1090 		if (!tmp)
1091 			return 0;
1092 		udelay(1);
1093 	}
1094 	return -ETIMEDOUT;
1095 }
1096 
1097 static int cik_sdma_soft_reset(void *handle)
1098 {
1099 	u32 srbm_soft_reset = 0;
1100 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1101 	u32 tmp = RREG32(mmSRBM_STATUS2);
1102 
1103 	if (tmp & SRBM_STATUS2__SDMA_BUSY_MASK) {
1104 		/* sdma0 */
1105 		tmp = RREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET);
1106 		tmp |= SDMA0_F32_CNTL__HALT_MASK;
1107 		WREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET, tmp);
1108 		srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA_MASK;
1109 	}
1110 	if (tmp & SRBM_STATUS2__SDMA1_BUSY_MASK) {
1111 		/* sdma1 */
1112 		tmp = RREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET);
1113 		tmp |= SDMA0_F32_CNTL__HALT_MASK;
1114 		WREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET, tmp);
1115 		srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA1_MASK;
1116 	}
1117 
1118 	if (srbm_soft_reset) {
1119 		tmp = RREG32(mmSRBM_SOFT_RESET);
1120 		tmp |= srbm_soft_reset;
1121 		dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
1122 		WREG32(mmSRBM_SOFT_RESET, tmp);
1123 		tmp = RREG32(mmSRBM_SOFT_RESET);
1124 
1125 		udelay(50);
1126 
1127 		tmp &= ~srbm_soft_reset;
1128 		WREG32(mmSRBM_SOFT_RESET, tmp);
1129 		tmp = RREG32(mmSRBM_SOFT_RESET);
1130 
1131 		/* Wait a little for things to settle down */
1132 		udelay(50);
1133 	}
1134 
1135 	return 0;
1136 }
1137 
1138 static int cik_sdma_set_trap_irq_state(struct amdgpu_device *adev,
1139 				       struct amdgpu_irq_src *src,
1140 				       unsigned type,
1141 				       enum amdgpu_interrupt_state state)
1142 {
1143 	u32 sdma_cntl;
1144 
1145 	switch (type) {
1146 	case AMDGPU_SDMA_IRQ_TRAP0:
1147 		switch (state) {
1148 		case AMDGPU_IRQ_STATE_DISABLE:
1149 			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1150 			sdma_cntl &= ~SDMA0_CNTL__TRAP_ENABLE_MASK;
1151 			WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1152 			break;
1153 		case AMDGPU_IRQ_STATE_ENABLE:
1154 			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1155 			sdma_cntl |= SDMA0_CNTL__TRAP_ENABLE_MASK;
1156 			WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1157 			break;
1158 		default:
1159 			break;
1160 		}
1161 		break;
1162 	case AMDGPU_SDMA_IRQ_TRAP1:
1163 		switch (state) {
1164 		case AMDGPU_IRQ_STATE_DISABLE:
1165 			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1166 			sdma_cntl &= ~SDMA0_CNTL__TRAP_ENABLE_MASK;
1167 			WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1168 			break;
1169 		case AMDGPU_IRQ_STATE_ENABLE:
1170 			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1171 			sdma_cntl |= SDMA0_CNTL__TRAP_ENABLE_MASK;
1172 			WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1173 			break;
1174 		default:
1175 			break;
1176 		}
1177 		break;
1178 	default:
1179 		break;
1180 	}
1181 	return 0;
1182 }
1183 
1184 static int cik_sdma_process_trap_irq(struct amdgpu_device *adev,
1185 				     struct amdgpu_irq_src *source,
1186 				     struct amdgpu_iv_entry *entry)
1187 {
1188 	u8 instance_id, queue_id;
1189 
1190 	instance_id = (entry->ring_id & 0x3) >> 0;
1191 	queue_id = (entry->ring_id & 0xc) >> 2;
1192 	DRM_DEBUG("IH: SDMA trap\n");
1193 	switch (instance_id) {
1194 	case 0:
1195 		switch (queue_id) {
1196 		case 0:
1197 			amdgpu_fence_process(&adev->sdma.instance[0].ring);
1198 			break;
1199 		case 1:
1200 			/* XXX compute */
1201 			break;
1202 		case 2:
1203 			/* XXX compute */
1204 			break;
1205 		}
1206 		break;
1207 	case 1:
1208 		switch (queue_id) {
1209 		case 0:
1210 			amdgpu_fence_process(&adev->sdma.instance[1].ring);
1211 			break;
1212 		case 1:
1213 			/* XXX compute */
1214 			break;
1215 		case 2:
1216 			/* XXX compute */
1217 			break;
1218 		}
1219 		break;
1220 	}
1221 
1222 	return 0;
1223 }
1224 
1225 static int cik_sdma_process_illegal_inst_irq(struct amdgpu_device *adev,
1226 					     struct amdgpu_irq_src *source,
1227 					     struct amdgpu_iv_entry *entry)
1228 {
1229 	DRM_ERROR("Illegal instruction in SDMA command stream\n");
1230 	schedule_work(&adev->reset_work);
1231 	return 0;
1232 }
1233 
1234 static int cik_sdma_set_clockgating_state(void *handle,
1235 					  enum amd_clockgating_state state)
1236 {
1237 	bool gate = false;
1238 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1239 
1240 	if (state == AMD_CG_STATE_GATE)
1241 		gate = true;
1242 
1243 	cik_enable_sdma_mgcg(adev, gate);
1244 	cik_enable_sdma_mgls(adev, gate);
1245 
1246 	return 0;
1247 }
1248 
1249 static int cik_sdma_set_powergating_state(void *handle,
1250 					  enum amd_powergating_state state)
1251 {
1252 	return 0;
1253 }
1254 
1255 static const struct amd_ip_funcs cik_sdma_ip_funcs = {
1256 	.name = "cik_sdma",
1257 	.early_init = cik_sdma_early_init,
1258 	.late_init = NULL,
1259 	.sw_init = cik_sdma_sw_init,
1260 	.sw_fini = cik_sdma_sw_fini,
1261 	.hw_init = cik_sdma_hw_init,
1262 	.hw_fini = cik_sdma_hw_fini,
1263 	.suspend = cik_sdma_suspend,
1264 	.resume = cik_sdma_resume,
1265 	.is_idle = cik_sdma_is_idle,
1266 	.wait_for_idle = cik_sdma_wait_for_idle,
1267 	.soft_reset = cik_sdma_soft_reset,
1268 	.set_clockgating_state = cik_sdma_set_clockgating_state,
1269 	.set_powergating_state = cik_sdma_set_powergating_state,
1270 };
1271 
1272 static const struct amdgpu_ring_funcs cik_sdma_ring_funcs = {
1273 	.type = AMDGPU_RING_TYPE_SDMA,
1274 	.align_mask = 0xf,
1275 	.nop = SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0),
1276 	.support_64bit_ptrs = false,
1277 	.get_rptr = cik_sdma_ring_get_rptr,
1278 	.get_wptr = cik_sdma_ring_get_wptr,
1279 	.set_wptr = cik_sdma_ring_set_wptr,
1280 	.emit_frame_size =
1281 		6 + /* cik_sdma_ring_emit_hdp_flush */
1282 		3 + /* cik_sdma_ring_emit_hdp_invalidate */
1283 		6 + /* cik_sdma_ring_emit_pipeline_sync */
1284 		12 + /* cik_sdma_ring_emit_vm_flush */
1285 		9 + 9 + 9, /* cik_sdma_ring_emit_fence x3 for user fence, vm fence */
1286 	.emit_ib_size = 7 + 4, /* cik_sdma_ring_emit_ib */
1287 	.emit_ib = cik_sdma_ring_emit_ib,
1288 	.emit_fence = cik_sdma_ring_emit_fence,
1289 	.emit_pipeline_sync = cik_sdma_ring_emit_pipeline_sync,
1290 	.emit_vm_flush = cik_sdma_ring_emit_vm_flush,
1291 	.emit_hdp_flush = cik_sdma_ring_emit_hdp_flush,
1292 	.emit_hdp_invalidate = cik_sdma_ring_emit_hdp_invalidate,
1293 	.test_ring = cik_sdma_ring_test_ring,
1294 	.test_ib = cik_sdma_ring_test_ib,
1295 	.insert_nop = cik_sdma_ring_insert_nop,
1296 	.pad_ib = cik_sdma_ring_pad_ib,
1297 };
1298 
1299 static void cik_sdma_set_ring_funcs(struct amdgpu_device *adev)
1300 {
1301 	int i;
1302 
1303 	for (i = 0; i < adev->sdma.num_instances; i++)
1304 		adev->sdma.instance[i].ring.funcs = &cik_sdma_ring_funcs;
1305 }
1306 
1307 static const struct amdgpu_irq_src_funcs cik_sdma_trap_irq_funcs = {
1308 	.set = cik_sdma_set_trap_irq_state,
1309 	.process = cik_sdma_process_trap_irq,
1310 };
1311 
1312 static const struct amdgpu_irq_src_funcs cik_sdma_illegal_inst_irq_funcs = {
1313 	.process = cik_sdma_process_illegal_inst_irq,
1314 };
1315 
1316 static void cik_sdma_set_irq_funcs(struct amdgpu_device *adev)
1317 {
1318 	adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST;
1319 	adev->sdma.trap_irq.funcs = &cik_sdma_trap_irq_funcs;
1320 	adev->sdma.illegal_inst_irq.funcs = &cik_sdma_illegal_inst_irq_funcs;
1321 }
1322 
1323 /**
1324  * cik_sdma_emit_copy_buffer - copy buffer using the sDMA engine
1325  *
1326  * @ring: amdgpu_ring structure holding ring information
1327  * @src_offset: src GPU address
1328  * @dst_offset: dst GPU address
1329  * @byte_count: number of bytes to xfer
1330  *
1331  * Copy GPU buffers using the DMA engine (CIK).
1332  * Used by the amdgpu ttm implementation to move pages if
1333  * registered as the asic copy callback.
1334  */
1335 static void cik_sdma_emit_copy_buffer(struct amdgpu_ib *ib,
1336 				      uint64_t src_offset,
1337 				      uint64_t dst_offset,
1338 				      uint32_t byte_count)
1339 {
1340 	ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_COPY, SDMA_COPY_SUB_OPCODE_LINEAR, 0);
1341 	ib->ptr[ib->length_dw++] = byte_count;
1342 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1343 	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1344 	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1345 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1346 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1347 }
1348 
1349 /**
1350  * cik_sdma_emit_fill_buffer - fill buffer using the sDMA engine
1351  *
1352  * @ring: amdgpu_ring structure holding ring information
1353  * @src_data: value to write to buffer
1354  * @dst_offset: dst GPU address
1355  * @byte_count: number of bytes to xfer
1356  *
1357  * Fill GPU buffers using the DMA engine (CIK).
1358  */
1359 static void cik_sdma_emit_fill_buffer(struct amdgpu_ib *ib,
1360 				      uint32_t src_data,
1361 				      uint64_t dst_offset,
1362 				      uint32_t byte_count)
1363 {
1364 	ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_CONSTANT_FILL, 0, 0);
1365 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1366 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1367 	ib->ptr[ib->length_dw++] = src_data;
1368 	ib->ptr[ib->length_dw++] = byte_count;
1369 }
1370 
1371 static const struct amdgpu_buffer_funcs cik_sdma_buffer_funcs = {
1372 	.copy_max_bytes = 0x1fffff,
1373 	.copy_num_dw = 7,
1374 	.emit_copy_buffer = cik_sdma_emit_copy_buffer,
1375 
1376 	.fill_max_bytes = 0x1fffff,
1377 	.fill_num_dw = 5,
1378 	.emit_fill_buffer = cik_sdma_emit_fill_buffer,
1379 };
1380 
1381 static void cik_sdma_set_buffer_funcs(struct amdgpu_device *adev)
1382 {
1383 	if (adev->mman.buffer_funcs == NULL) {
1384 		adev->mman.buffer_funcs = &cik_sdma_buffer_funcs;
1385 		adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1386 	}
1387 }
1388 
1389 static const struct amdgpu_vm_pte_funcs cik_sdma_vm_pte_funcs = {
1390 	.copy_pte_num_dw = 7,
1391 	.copy_pte = cik_sdma_vm_copy_pte,
1392 
1393 	.write_pte = cik_sdma_vm_write_pte,
1394 
1395 	.set_max_nums_pte_pde = 0x1fffff >> 3,
1396 	.set_pte_pde_num_dw = 10,
1397 	.set_pte_pde = cik_sdma_vm_set_pte_pde,
1398 };
1399 
1400 static void cik_sdma_set_vm_pte_funcs(struct amdgpu_device *adev)
1401 {
1402 	unsigned i;
1403 
1404 	if (adev->vm_manager.vm_pte_funcs == NULL) {
1405 		adev->vm_manager.vm_pte_funcs = &cik_sdma_vm_pte_funcs;
1406 		for (i = 0; i < adev->sdma.num_instances; i++)
1407 			adev->vm_manager.vm_pte_rings[i] =
1408 				&adev->sdma.instance[i].ring;
1409 
1410 		adev->vm_manager.vm_pte_num_rings = adev->sdma.num_instances;
1411 	}
1412 }
1413 
1414 const struct amdgpu_ip_block_version cik_sdma_ip_block =
1415 {
1416 	.type = AMD_IP_BLOCK_TYPE_SDMA,
1417 	.major = 2,
1418 	.minor = 0,
1419 	.rev = 0,
1420 	.funcs = &cik_sdma_ip_funcs,
1421 };
1422