1 /*
2  * Copyright 2014 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 "vi.h"
30 #include "vid.h"
31 
32 #include "oss/oss_2_4_d.h"
33 #include "oss/oss_2_4_sh_mask.h"
34 
35 #include "gmc/gmc_7_1_d.h"
36 #include "gmc/gmc_7_1_sh_mask.h"
37 
38 #include "gca/gfx_8_0_d.h"
39 #include "gca/gfx_8_0_enum.h"
40 #include "gca/gfx_8_0_sh_mask.h"
41 
42 #include "bif/bif_5_0_d.h"
43 #include "bif/bif_5_0_sh_mask.h"
44 
45 #include "iceland_sdma_pkt_open.h"
46 
47 static void sdma_v2_4_set_ring_funcs(struct amdgpu_device *adev);
48 static void sdma_v2_4_set_buffer_funcs(struct amdgpu_device *adev);
49 static void sdma_v2_4_set_vm_pte_funcs(struct amdgpu_device *adev);
50 static void sdma_v2_4_set_irq_funcs(struct amdgpu_device *adev);
51 
52 MODULE_FIRMWARE("amdgpu/topaz_sdma.bin");
53 MODULE_FIRMWARE("amdgpu/topaz_sdma1.bin");
54 
55 static const u32 sdma_offsets[SDMA_MAX_INSTANCE] =
56 {
57 	SDMA0_REGISTER_OFFSET,
58 	SDMA1_REGISTER_OFFSET
59 };
60 
61 static const u32 golden_settings_iceland_a11[] =
62 {
63 	mmSDMA0_CHICKEN_BITS, 0xfc910007, 0x00810007,
64 	mmSDMA0_CLK_CTRL, 0xff000fff, 0x00000000,
65 	mmSDMA1_CHICKEN_BITS, 0xfc910007, 0x00810007,
66 	mmSDMA1_CLK_CTRL, 0xff000fff, 0x00000000,
67 };
68 
69 static const u32 iceland_mgcg_cgcg_init[] =
70 {
71 	mmSDMA0_CLK_CTRL, 0xff000ff0, 0x00000100,
72 	mmSDMA1_CLK_CTRL, 0xff000ff0, 0x00000100
73 };
74 
75 /*
76  * sDMA - System DMA
77  * Starting with CIK, the GPU has new asynchronous
78  * DMA engines.  These engines are used for compute
79  * and gfx.  There are two DMA engines (SDMA0, SDMA1)
80  * and each one supports 1 ring buffer used for gfx
81  * and 2 queues used for compute.
82  *
83  * The programming model is very similar to the CP
84  * (ring buffer, IBs, etc.), but sDMA has it's own
85  * packet format that is different from the PM4 format
86  * used by the CP. sDMA supports copying data, writing
87  * embedded data, solid fills, and a number of other
88  * things.  It also has support for tiling/detiling of
89  * buffers.
90  */
91 
92 static void sdma_v2_4_init_golden_registers(struct amdgpu_device *adev)
93 {
94 	switch (adev->asic_type) {
95 	case CHIP_TOPAZ:
96 		amdgpu_program_register_sequence(adev,
97 						 iceland_mgcg_cgcg_init,
98 						 (const u32)ARRAY_SIZE(iceland_mgcg_cgcg_init));
99 		amdgpu_program_register_sequence(adev,
100 						 golden_settings_iceland_a11,
101 						 (const u32)ARRAY_SIZE(golden_settings_iceland_a11));
102 		break;
103 	default:
104 		break;
105 	}
106 }
107 
108 static void sdma_v2_4_free_microcode(struct amdgpu_device *adev)
109 {
110 	int i;
111 	for (i = 0; i < adev->sdma.num_instances; i++) {
112 		release_firmware(adev->sdma.instance[i].fw);
113 		adev->sdma.instance[i].fw = NULL;
114 	}
115 }
116 
117 /**
118  * sdma_v2_4_init_microcode - load ucode images from disk
119  *
120  * @adev: amdgpu_device pointer
121  *
122  * Use the firmware interface to load the ucode images into
123  * the driver (not loaded into hw).
124  * Returns 0 on success, error on failure.
125  */
126 static int sdma_v2_4_init_microcode(struct amdgpu_device *adev)
127 {
128 	const char *chip_name;
129 	char fw_name[30];
130 	int err = 0, i;
131 	struct amdgpu_firmware_info *info = NULL;
132 	const struct common_firmware_header *header = NULL;
133 	const struct sdma_firmware_header_v1_0 *hdr;
134 
135 	DRM_DEBUG("\n");
136 
137 	switch (adev->asic_type) {
138 	case CHIP_TOPAZ:
139 		chip_name = "topaz";
140 		break;
141 	default: BUG();
142 	}
143 
144 	for (i = 0; i < adev->sdma.num_instances; i++) {
145 		if (i == 0)
146 			snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma.bin", chip_name);
147 		else
148 			snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma1.bin", chip_name);
149 		err = request_firmware(&adev->sdma.instance[i].fw, fw_name, adev->dev);
150 		if (err)
151 			goto out;
152 		err = amdgpu_ucode_validate(adev->sdma.instance[i].fw);
153 		if (err)
154 			goto out;
155 		hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
156 		adev->sdma.instance[i].fw_version = le32_to_cpu(hdr->header.ucode_version);
157 		adev->sdma.instance[i].feature_version = le32_to_cpu(hdr->ucode_feature_version);
158 		if (adev->sdma.instance[i].feature_version >= 20)
159 			adev->sdma.instance[i].burst_nop = true;
160 
161 		if (adev->firmware.smu_load) {
162 			info = &adev->firmware.ucode[AMDGPU_UCODE_ID_SDMA0 + i];
163 			info->ucode_id = AMDGPU_UCODE_ID_SDMA0 + i;
164 			info->fw = adev->sdma.instance[i].fw;
165 			header = (const struct common_firmware_header *)info->fw->data;
166 			adev->firmware.fw_size +=
167 				ALIGN(le32_to_cpu(header->ucode_size_bytes), PAGE_SIZE);
168 		}
169 	}
170 
171 out:
172 	if (err) {
173 		printk(KERN_ERR
174 		       "sdma_v2_4: Failed to load firmware \"%s\"\n",
175 		       fw_name);
176 		for (i = 0; i < adev->sdma.num_instances; i++) {
177 			release_firmware(adev->sdma.instance[i].fw);
178 			adev->sdma.instance[i].fw = NULL;
179 		}
180 	}
181 	return err;
182 }
183 
184 /**
185  * sdma_v2_4_ring_get_rptr - get the current read pointer
186  *
187  * @ring: amdgpu ring pointer
188  *
189  * Get the current rptr from the hardware (VI+).
190  */
191 static uint32_t sdma_v2_4_ring_get_rptr(struct amdgpu_ring *ring)
192 {
193 	u32 rptr;
194 
195 	/* XXX check if swapping is necessary on BE */
196 	rptr = ring->adev->wb.wb[ring->rptr_offs] >> 2;
197 
198 	return rptr;
199 }
200 
201 /**
202  * sdma_v2_4_ring_get_wptr - get the current write pointer
203  *
204  * @ring: amdgpu ring pointer
205  *
206  * Get the current wptr from the hardware (VI+).
207  */
208 static uint32_t sdma_v2_4_ring_get_wptr(struct amdgpu_ring *ring)
209 {
210 	struct amdgpu_device *adev = ring->adev;
211 	int me = (ring == &ring->adev->sdma.instance[0].ring) ? 0 : 1;
212 	u32 wptr = RREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[me]) >> 2;
213 
214 	return wptr;
215 }
216 
217 /**
218  * sdma_v2_4_ring_set_wptr - commit the write pointer
219  *
220  * @ring: amdgpu ring pointer
221  *
222  * Write the wptr back to the hardware (VI+).
223  */
224 static void sdma_v2_4_ring_set_wptr(struct amdgpu_ring *ring)
225 {
226 	struct amdgpu_device *adev = ring->adev;
227 	int me = (ring == &ring->adev->sdma.instance[0].ring) ? 0 : 1;
228 
229 	WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[me], ring->wptr << 2);
230 }
231 
232 static void sdma_v2_4_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
233 {
234 	struct amdgpu_sdma_instance *sdma = amdgpu_get_sdma_instance(ring);
235 	int i;
236 
237 	for (i = 0; i < count; i++)
238 		if (sdma && sdma->burst_nop && (i == 0))
239 			amdgpu_ring_write(ring, ring->nop |
240 				SDMA_PKT_NOP_HEADER_COUNT(count - 1));
241 		else
242 			amdgpu_ring_write(ring, ring->nop);
243 }
244 
245 /**
246  * sdma_v2_4_ring_emit_ib - Schedule an IB on the DMA engine
247  *
248  * @ring: amdgpu ring pointer
249  * @ib: IB object to schedule
250  *
251  * Schedule an IB in the DMA ring (VI).
252  */
253 static void sdma_v2_4_ring_emit_ib(struct amdgpu_ring *ring,
254 				   struct amdgpu_ib *ib,
255 				   unsigned vm_id, bool ctx_switch)
256 {
257 	u32 vmid = vm_id & 0xf;
258 
259 	/* IB packet must end on a 8 DW boundary */
260 	sdma_v2_4_ring_insert_nop(ring, (10 - (ring->wptr & 7)) % 8);
261 
262 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) |
263 			  SDMA_PKT_INDIRECT_HEADER_VMID(vmid));
264 	/* base must be 32 byte aligned */
265 	amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
266 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
267 	amdgpu_ring_write(ring, ib->length_dw);
268 	amdgpu_ring_write(ring, 0);
269 	amdgpu_ring_write(ring, 0);
270 
271 }
272 
273 /**
274  * sdma_v2_4_hdp_flush_ring_emit - emit an hdp flush on the DMA ring
275  *
276  * @ring: amdgpu ring pointer
277  *
278  * Emit an hdp flush packet on the requested DMA ring.
279  */
280 static void sdma_v2_4_ring_emit_hdp_flush(struct amdgpu_ring *ring)
281 {
282 	u32 ref_and_mask = 0;
283 
284 	if (ring == &ring->adev->sdma.instance[0].ring)
285 		ref_and_mask = REG_SET_FIELD(ref_and_mask, GPU_HDP_FLUSH_DONE, SDMA0, 1);
286 	else
287 		ref_and_mask = REG_SET_FIELD(ref_and_mask, GPU_HDP_FLUSH_DONE, SDMA1, 1);
288 
289 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
290 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) |
291 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
292 	amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_DONE << 2);
293 	amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_REQ << 2);
294 	amdgpu_ring_write(ring, ref_and_mask); /* reference */
295 	amdgpu_ring_write(ring, ref_and_mask); /* mask */
296 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
297 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
298 }
299 
300 static void sdma_v2_4_ring_emit_hdp_invalidate(struct amdgpu_ring *ring)
301 {
302 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
303 			  SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
304 	amdgpu_ring_write(ring, mmHDP_DEBUG0);
305 	amdgpu_ring_write(ring, 1);
306 }
307 /**
308  * sdma_v2_4_ring_emit_fence - emit a fence on the DMA ring
309  *
310  * @ring: amdgpu ring pointer
311  * @fence: amdgpu fence object
312  *
313  * Add a DMA fence packet to the ring to write
314  * the fence seq number and DMA trap packet to generate
315  * an interrupt if needed (VI).
316  */
317 static void sdma_v2_4_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
318 				      unsigned flags)
319 {
320 	bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
321 	/* write the fence */
322 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE));
323 	amdgpu_ring_write(ring, lower_32_bits(addr));
324 	amdgpu_ring_write(ring, upper_32_bits(addr));
325 	amdgpu_ring_write(ring, lower_32_bits(seq));
326 
327 	/* optionally write high bits as well */
328 	if (write64bit) {
329 		addr += 4;
330 		amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE));
331 		amdgpu_ring_write(ring, lower_32_bits(addr));
332 		amdgpu_ring_write(ring, upper_32_bits(addr));
333 		amdgpu_ring_write(ring, upper_32_bits(seq));
334 	}
335 
336 	/* generate an interrupt */
337 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP));
338 	amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(0));
339 }
340 
341 /**
342  * sdma_v2_4_gfx_stop - stop the gfx async dma engines
343  *
344  * @adev: amdgpu_device pointer
345  *
346  * Stop the gfx async dma ring buffers (VI).
347  */
348 static void sdma_v2_4_gfx_stop(struct amdgpu_device *adev)
349 {
350 	struct amdgpu_ring *sdma0 = &adev->sdma.instance[0].ring;
351 	struct amdgpu_ring *sdma1 = &adev->sdma.instance[1].ring;
352 	u32 rb_cntl, ib_cntl;
353 	int i;
354 
355 	if ((adev->mman.buffer_funcs_ring == sdma0) ||
356 	    (adev->mman.buffer_funcs_ring == sdma1))
357 		amdgpu_ttm_set_active_vram_size(adev, adev->mc.visible_vram_size);
358 
359 	for (i = 0; i < adev->sdma.num_instances; i++) {
360 		rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
361 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 0);
362 		WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
363 		ib_cntl = RREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i]);
364 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 0);
365 		WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
366 	}
367 	sdma0->ready = false;
368 	sdma1->ready = false;
369 }
370 
371 /**
372  * sdma_v2_4_rlc_stop - stop the compute async dma engines
373  *
374  * @adev: amdgpu_device pointer
375  *
376  * Stop the compute async dma queues (VI).
377  */
378 static void sdma_v2_4_rlc_stop(struct amdgpu_device *adev)
379 {
380 	/* XXX todo */
381 }
382 
383 /**
384  * sdma_v2_4_enable - stop the async dma engines
385  *
386  * @adev: amdgpu_device pointer
387  * @enable: enable/disable the DMA MEs.
388  *
389  * Halt or unhalt the async dma engines (VI).
390  */
391 static void sdma_v2_4_enable(struct amdgpu_device *adev, bool enable)
392 {
393 	u32 f32_cntl;
394 	int i;
395 
396 	if (!enable) {
397 		sdma_v2_4_gfx_stop(adev);
398 		sdma_v2_4_rlc_stop(adev);
399 	}
400 
401 	for (i = 0; i < adev->sdma.num_instances; i++) {
402 		f32_cntl = RREG32(mmSDMA0_F32_CNTL + sdma_offsets[i]);
403 		if (enable)
404 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 0);
405 		else
406 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 1);
407 		WREG32(mmSDMA0_F32_CNTL + sdma_offsets[i], f32_cntl);
408 	}
409 }
410 
411 /**
412  * sdma_v2_4_gfx_resume - setup and start the async dma engines
413  *
414  * @adev: amdgpu_device pointer
415  *
416  * Set up the gfx DMA ring buffers and enable them (VI).
417  * Returns 0 for success, error for failure.
418  */
419 static int sdma_v2_4_gfx_resume(struct amdgpu_device *adev)
420 {
421 	struct amdgpu_ring *ring;
422 	u32 rb_cntl, ib_cntl;
423 	u32 rb_bufsz;
424 	u32 wb_offset;
425 	int i, j, r;
426 
427 	for (i = 0; i < adev->sdma.num_instances; i++) {
428 		ring = &adev->sdma.instance[i].ring;
429 		wb_offset = (ring->rptr_offs * 4);
430 
431 		mutex_lock(&adev->srbm_mutex);
432 		for (j = 0; j < 16; j++) {
433 			vi_srbm_select(adev, 0, 0, 0, j);
434 			/* SDMA GFX */
435 			WREG32(mmSDMA0_GFX_VIRTUAL_ADDR + sdma_offsets[i], 0);
436 			WREG32(mmSDMA0_GFX_APE1_CNTL + sdma_offsets[i], 0);
437 		}
438 		vi_srbm_select(adev, 0, 0, 0, 0);
439 		mutex_unlock(&adev->srbm_mutex);
440 
441 		WREG32(mmSDMA0_TILING_CONFIG + sdma_offsets[i],
442 		       adev->gfx.config.gb_addr_config & 0x70);
443 
444 		WREG32(mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i], 0);
445 
446 		/* Set ring buffer size in dwords */
447 		rb_bufsz = order_base_2(ring->ring_size / 4);
448 		rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
449 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz);
450 #ifdef __BIG_ENDIAN
451 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1);
452 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL,
453 					RPTR_WRITEBACK_SWAP_ENABLE, 1);
454 #endif
455 		WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
456 
457 		/* Initialize the ring buffer's read and write pointers */
458 		WREG32(mmSDMA0_GFX_RB_RPTR + sdma_offsets[i], 0);
459 		WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], 0);
460 		WREG32(mmSDMA0_GFX_IB_RPTR + sdma_offsets[i], 0);
461 		WREG32(mmSDMA0_GFX_IB_OFFSET + sdma_offsets[i], 0);
462 
463 		/* set the wb address whether it's enabled or not */
464 		WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_HI + sdma_offsets[i],
465 		       upper_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFF);
466 		WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_LO + sdma_offsets[i],
467 		       lower_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC);
468 
469 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
470 
471 		WREG32(mmSDMA0_GFX_RB_BASE + sdma_offsets[i], ring->gpu_addr >> 8);
472 		WREG32(mmSDMA0_GFX_RB_BASE_HI + sdma_offsets[i], ring->gpu_addr >> 40);
473 
474 		ring->wptr = 0;
475 		WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], ring->wptr << 2);
476 
477 		/* enable DMA RB */
478 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1);
479 		WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
480 
481 		ib_cntl = RREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i]);
482 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 1);
483 #ifdef __BIG_ENDIAN
484 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1);
485 #endif
486 		/* enable DMA IBs */
487 		WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
488 
489 		ring->ready = true;
490 	}
491 
492 	sdma_v2_4_enable(adev, true);
493 	for (i = 0; i < adev->sdma.num_instances; i++) {
494 		ring = &adev->sdma.instance[i].ring;
495 		r = amdgpu_ring_test_ring(ring);
496 		if (r) {
497 			ring->ready = false;
498 			return r;
499 		}
500 
501 		if (adev->mman.buffer_funcs_ring == ring)
502 			amdgpu_ttm_set_active_vram_size(adev, adev->mc.real_vram_size);
503 	}
504 
505 	return 0;
506 }
507 
508 /**
509  * sdma_v2_4_rlc_resume - setup and start the async dma engines
510  *
511  * @adev: amdgpu_device pointer
512  *
513  * Set up the compute DMA queues and enable them (VI).
514  * Returns 0 for success, error for failure.
515  */
516 static int sdma_v2_4_rlc_resume(struct amdgpu_device *adev)
517 {
518 	/* XXX todo */
519 	return 0;
520 }
521 
522 /**
523  * sdma_v2_4_load_microcode - load the sDMA ME ucode
524  *
525  * @adev: amdgpu_device pointer
526  *
527  * Loads the sDMA0/1 ucode.
528  * Returns 0 for success, -EINVAL if the ucode is not available.
529  */
530 static int sdma_v2_4_load_microcode(struct amdgpu_device *adev)
531 {
532 	const struct sdma_firmware_header_v1_0 *hdr;
533 	const __le32 *fw_data;
534 	u32 fw_size;
535 	int i, j;
536 
537 	/* halt the MEs */
538 	sdma_v2_4_enable(adev, false);
539 
540 	for (i = 0; i < adev->sdma.num_instances; i++) {
541 		if (!adev->sdma.instance[i].fw)
542 			return -EINVAL;
543 		hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
544 		amdgpu_ucode_print_sdma_hdr(&hdr->header);
545 		fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
546 		fw_data = (const __le32 *)
547 			(adev->sdma.instance[i].fw->data +
548 			 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
549 		WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], 0);
550 		for (j = 0; j < fw_size; j++)
551 			WREG32(mmSDMA0_UCODE_DATA + sdma_offsets[i], le32_to_cpup(fw_data++));
552 		WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], adev->sdma.instance[i].fw_version);
553 	}
554 
555 	return 0;
556 }
557 
558 /**
559  * sdma_v2_4_start - setup and start the async dma engines
560  *
561  * @adev: amdgpu_device pointer
562  *
563  * Set up the DMA engines and enable them (VI).
564  * Returns 0 for success, error for failure.
565  */
566 static int sdma_v2_4_start(struct amdgpu_device *adev)
567 {
568 	int r;
569 
570 	if (!adev->pp_enabled) {
571 		if (!adev->firmware.smu_load) {
572 			r = sdma_v2_4_load_microcode(adev);
573 			if (r)
574 				return r;
575 		} else {
576 			r = adev->smu.smumgr_funcs->check_fw_load_finish(adev,
577 							AMDGPU_UCODE_ID_SDMA0);
578 			if (r)
579 				return -EINVAL;
580 			r = adev->smu.smumgr_funcs->check_fw_load_finish(adev,
581 							AMDGPU_UCODE_ID_SDMA1);
582 			if (r)
583 				return -EINVAL;
584 		}
585 	}
586 
587 	/* halt the engine before programing */
588 	sdma_v2_4_enable(adev, false);
589 
590 	/* start the gfx rings and rlc compute queues */
591 	r = sdma_v2_4_gfx_resume(adev);
592 	if (r)
593 		return r;
594 	r = sdma_v2_4_rlc_resume(adev);
595 	if (r)
596 		return r;
597 
598 	return 0;
599 }
600 
601 /**
602  * sdma_v2_4_ring_test_ring - simple async dma engine test
603  *
604  * @ring: amdgpu_ring structure holding ring information
605  *
606  * Test the DMA engine by writing using it to write an
607  * value to memory. (VI).
608  * Returns 0 for success, error for failure.
609  */
610 static int sdma_v2_4_ring_test_ring(struct amdgpu_ring *ring)
611 {
612 	struct amdgpu_device *adev = ring->adev;
613 	unsigned i;
614 	unsigned index;
615 	int r;
616 	u32 tmp;
617 	u64 gpu_addr;
618 
619 	r = amdgpu_wb_get(adev, &index);
620 	if (r) {
621 		dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
622 		return r;
623 	}
624 
625 	gpu_addr = adev->wb.gpu_addr + (index * 4);
626 	tmp = 0xCAFEDEAD;
627 	adev->wb.wb[index] = cpu_to_le32(tmp);
628 
629 	r = amdgpu_ring_alloc(ring, 5);
630 	if (r) {
631 		DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
632 		amdgpu_wb_free(adev, index);
633 		return r;
634 	}
635 
636 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
637 			  SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
638 	amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
639 	amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
640 	amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(1));
641 	amdgpu_ring_write(ring, 0xDEADBEEF);
642 	amdgpu_ring_commit(ring);
643 
644 	for (i = 0; i < adev->usec_timeout; i++) {
645 		tmp = le32_to_cpu(adev->wb.wb[index]);
646 		if (tmp == 0xDEADBEEF)
647 			break;
648 		DRM_UDELAY(1);
649 	}
650 
651 	if (i < adev->usec_timeout) {
652 		DRM_INFO("ring test on %d succeeded in %d usecs\n", ring->idx, i);
653 	} else {
654 		DRM_ERROR("amdgpu: ring %d test failed (0x%08X)\n",
655 			  ring->idx, tmp);
656 		r = -EINVAL;
657 	}
658 	amdgpu_wb_free(adev, index);
659 
660 	return r;
661 }
662 
663 /**
664  * sdma_v2_4_ring_test_ib - test an IB on the DMA engine
665  *
666  * @ring: amdgpu_ring structure holding ring information
667  *
668  * Test a simple IB in the DMA ring (VI).
669  * Returns 0 on success, error on failure.
670  */
671 static int sdma_v2_4_ring_test_ib(struct amdgpu_ring *ring, long timeout)
672 {
673 	struct amdgpu_device *adev = ring->adev;
674 	struct amdgpu_ib ib;
675 	struct fence *f = NULL;
676 	unsigned index;
677 	u32 tmp = 0;
678 	u64 gpu_addr;
679 	long r;
680 
681 	r = amdgpu_wb_get(adev, &index);
682 	if (r) {
683 		dev_err(adev->dev, "(%ld) failed to allocate wb slot\n", r);
684 		return r;
685 	}
686 
687 	gpu_addr = adev->wb.gpu_addr + (index * 4);
688 	tmp = 0xCAFEDEAD;
689 	adev->wb.wb[index] = cpu_to_le32(tmp);
690 	memset(&ib, 0, sizeof(ib));
691 	r = amdgpu_ib_get(adev, NULL, 256, &ib);
692 	if (r) {
693 		DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r);
694 		goto err0;
695 	}
696 
697 	ib.ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
698 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
699 	ib.ptr[1] = lower_32_bits(gpu_addr);
700 	ib.ptr[2] = upper_32_bits(gpu_addr);
701 	ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(1);
702 	ib.ptr[4] = 0xDEADBEEF;
703 	ib.ptr[5] = SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
704 	ib.ptr[6] = SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
705 	ib.ptr[7] = SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
706 	ib.length_dw = 8;
707 
708 	r = amdgpu_ib_schedule(ring, 1, &ib, NULL, NULL, &f);
709 	if (r)
710 		goto err1;
711 
712 	r = fence_wait_timeout(f, false, timeout);
713 	if (r == 0) {
714 		DRM_ERROR("amdgpu: IB test timed out\n");
715 		r = -ETIMEDOUT;
716 		goto err1;
717 	} else if (r) {
718 		DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
719 		goto err1;
720 	}
721 	tmp = le32_to_cpu(adev->wb.wb[index]);
722 	if (tmp == 0xDEADBEEF) {
723 		DRM_INFO("ib test on ring %d succeeded\n", ring->idx);
724 		r = 0;
725 	} else {
726 		DRM_ERROR("amdgpu: ib test failed (0x%08X)\n", tmp);
727 		r = -EINVAL;
728 	}
729 
730 err1:
731 	amdgpu_ib_free(adev, &ib, NULL);
732 	fence_put(f);
733 err0:
734 	amdgpu_wb_free(adev, index);
735 	return r;
736 }
737 
738 /**
739  * sdma_v2_4_vm_copy_pte - update PTEs by copying them from the GART
740  *
741  * @ib: indirect buffer to fill with commands
742  * @pe: addr of the page entry
743  * @src: src addr to copy from
744  * @count: number of page entries to update
745  *
746  * Update PTEs by copying them from the GART using sDMA (CIK).
747  */
748 static void sdma_v2_4_vm_copy_pte(struct amdgpu_ib *ib,
749 				  uint64_t pe, uint64_t src,
750 				  unsigned count)
751 {
752 	while (count) {
753 		unsigned bytes = count * 8;
754 		if (bytes > 0x1FFFF8)
755 			bytes = 0x1FFFF8;
756 
757 		ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
758 			SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
759 		ib->ptr[ib->length_dw++] = bytes;
760 		ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
761 		ib->ptr[ib->length_dw++] = lower_32_bits(src);
762 		ib->ptr[ib->length_dw++] = upper_32_bits(src);
763 		ib->ptr[ib->length_dw++] = lower_32_bits(pe);
764 		ib->ptr[ib->length_dw++] = upper_32_bits(pe);
765 
766 		pe += bytes;
767 		src += bytes;
768 		count -= bytes / 8;
769 	}
770 }
771 
772 /**
773  * sdma_v2_4_vm_write_pte - update PTEs by writing them manually
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 PTEs by writing them manually using sDMA (CIK).
783  */
784 static void sdma_v2_4_vm_write_pte(struct amdgpu_ib *ib,
785 				   const dma_addr_t *pages_addr, uint64_t pe,
786 				   uint64_t addr, unsigned count,
787 				   uint32_t incr, uint32_t flags)
788 {
789 	uint64_t value;
790 	unsigned ndw;
791 
792 	while (count) {
793 		ndw = count * 2;
794 		if (ndw > 0xFFFFE)
795 			ndw = 0xFFFFE;
796 
797 		/* for non-physically contiguous pages (system) */
798 		ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
799 			SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
800 		ib->ptr[ib->length_dw++] = pe;
801 		ib->ptr[ib->length_dw++] = upper_32_bits(pe);
802 		ib->ptr[ib->length_dw++] = ndw;
803 		for (; ndw > 0; ndw -= 2, --count, pe += 8) {
804 			value = amdgpu_vm_map_gart(pages_addr, addr);
805 			addr += incr;
806 			value |= flags;
807 			ib->ptr[ib->length_dw++] = value;
808 			ib->ptr[ib->length_dw++] = upper_32_bits(value);
809 		}
810 	}
811 }
812 
813 /**
814  * sdma_v2_4_vm_set_pte_pde - update the page tables using sDMA
815  *
816  * @ib: indirect buffer to fill with commands
817  * @pe: addr of the page entry
818  * @addr: dst addr to write into pe
819  * @count: number of page entries to update
820  * @incr: increase next addr by incr bytes
821  * @flags: access flags
822  *
823  * Update the page tables using sDMA (CIK).
824  */
825 static void sdma_v2_4_vm_set_pte_pde(struct amdgpu_ib *ib,
826 				     uint64_t pe,
827 				     uint64_t addr, unsigned count,
828 				     uint32_t incr, uint32_t flags)
829 {
830 	uint64_t value;
831 	unsigned ndw;
832 
833 	while (count) {
834 		ndw = count;
835 		if (ndw > 0x7FFFF)
836 			ndw = 0x7FFFF;
837 
838 		if (flags & AMDGPU_PTE_VALID)
839 			value = addr;
840 		else
841 			value = 0;
842 
843 		/* for physically contiguous pages (vram) */
844 		ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_GEN_PTEPDE);
845 		ib->ptr[ib->length_dw++] = pe; /* dst addr */
846 		ib->ptr[ib->length_dw++] = upper_32_bits(pe);
847 		ib->ptr[ib->length_dw++] = flags; /* mask */
848 		ib->ptr[ib->length_dw++] = 0;
849 		ib->ptr[ib->length_dw++] = value; /* value */
850 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
851 		ib->ptr[ib->length_dw++] = incr; /* increment size */
852 		ib->ptr[ib->length_dw++] = 0;
853 		ib->ptr[ib->length_dw++] = ndw; /* number of entries */
854 
855 		pe += ndw * 8;
856 		addr += ndw * incr;
857 		count -= ndw;
858 	}
859 }
860 
861 /**
862  * sdma_v2_4_ring_pad_ib - pad the IB to the required number of dw
863  *
864  * @ib: indirect buffer to fill with padding
865  *
866  */
867 static void sdma_v2_4_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
868 {
869 	struct amdgpu_sdma_instance *sdma = amdgpu_get_sdma_instance(ring);
870 	u32 pad_count;
871 	int i;
872 
873 	pad_count = (8 - (ib->length_dw & 0x7)) % 8;
874 	for (i = 0; i < pad_count; i++)
875 		if (sdma && sdma->burst_nop && (i == 0))
876 			ib->ptr[ib->length_dw++] =
877 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP) |
878 				SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
879 		else
880 			ib->ptr[ib->length_dw++] =
881 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
882 }
883 
884 /**
885  * sdma_v2_4_ring_emit_pipeline_sync - sync the pipeline
886  *
887  * @ring: amdgpu_ring pointer
888  *
889  * Make sure all previous operations are completed (CIK).
890  */
891 static void sdma_v2_4_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
892 {
893 	uint32_t seq = ring->fence_drv.sync_seq;
894 	uint64_t addr = ring->fence_drv.gpu_addr;
895 
896 	/* wait for idle */
897 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
898 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
899 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
900 			  SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1));
901 	amdgpu_ring_write(ring, addr & 0xfffffffc);
902 	amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
903 	amdgpu_ring_write(ring, seq); /* reference */
904 	amdgpu_ring_write(ring, 0xfffffff); /* mask */
905 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
906 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */
907 }
908 
909 /**
910  * sdma_v2_4_ring_emit_vm_flush - cik vm flush using sDMA
911  *
912  * @ring: amdgpu_ring pointer
913  * @vm: amdgpu_vm pointer
914  *
915  * Update the page table base and flush the VM TLB
916  * using sDMA (VI).
917  */
918 static void sdma_v2_4_ring_emit_vm_flush(struct amdgpu_ring *ring,
919 					 unsigned vm_id, uint64_t pd_addr)
920 {
921 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
922 			  SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
923 	if (vm_id < 8) {
924 		amdgpu_ring_write(ring, (mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + vm_id));
925 	} else {
926 		amdgpu_ring_write(ring, (mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + vm_id - 8));
927 	}
928 	amdgpu_ring_write(ring, pd_addr >> 12);
929 
930 	/* flush TLB */
931 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
932 			  SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
933 	amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST);
934 	amdgpu_ring_write(ring, 1 << vm_id);
935 
936 	/* wait for flush */
937 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
938 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
939 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(0)); /* always */
940 	amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST << 2);
941 	amdgpu_ring_write(ring, 0);
942 	amdgpu_ring_write(ring, 0); /* reference */
943 	amdgpu_ring_write(ring, 0); /* mask */
944 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
945 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
946 }
947 
948 static int sdma_v2_4_early_init(void *handle)
949 {
950 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
951 
952 	adev->sdma.num_instances = SDMA_MAX_INSTANCE;
953 
954 	sdma_v2_4_set_ring_funcs(adev);
955 	sdma_v2_4_set_buffer_funcs(adev);
956 	sdma_v2_4_set_vm_pte_funcs(adev);
957 	sdma_v2_4_set_irq_funcs(adev);
958 
959 	return 0;
960 }
961 
962 static int sdma_v2_4_sw_init(void *handle)
963 {
964 	struct amdgpu_ring *ring;
965 	int r, i;
966 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
967 
968 	/* SDMA trap event */
969 	r = amdgpu_irq_add_id(adev, 224, &adev->sdma.trap_irq);
970 	if (r)
971 		return r;
972 
973 	/* SDMA Privileged inst */
974 	r = amdgpu_irq_add_id(adev, 241, &adev->sdma.illegal_inst_irq);
975 	if (r)
976 		return r;
977 
978 	/* SDMA Privileged inst */
979 	r = amdgpu_irq_add_id(adev, 247, &adev->sdma.illegal_inst_irq);
980 	if (r)
981 		return r;
982 
983 	r = sdma_v2_4_init_microcode(adev);
984 	if (r) {
985 		DRM_ERROR("Failed to load sdma firmware!\n");
986 		return r;
987 	}
988 
989 	for (i = 0; i < adev->sdma.num_instances; i++) {
990 		ring = &adev->sdma.instance[i].ring;
991 		ring->ring_obj = NULL;
992 		ring->use_doorbell = false;
993 		sprintf(ring->name, "sdma%d", i);
994 		r = amdgpu_ring_init(adev, ring, 1024,
995 				     SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP), 0xf,
996 				     &adev->sdma.trap_irq,
997 				     (i == 0) ?
998 				     AMDGPU_SDMA_IRQ_TRAP0 : AMDGPU_SDMA_IRQ_TRAP1,
999 				     AMDGPU_RING_TYPE_SDMA);
1000 		if (r)
1001 			return r;
1002 	}
1003 
1004 	return r;
1005 }
1006 
1007 static int sdma_v2_4_sw_fini(void *handle)
1008 {
1009 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1010 	int i;
1011 
1012 	for (i = 0; i < adev->sdma.num_instances; i++)
1013 		amdgpu_ring_fini(&adev->sdma.instance[i].ring);
1014 
1015 	sdma_v2_4_free_microcode(adev);
1016 	return 0;
1017 }
1018 
1019 static int sdma_v2_4_hw_init(void *handle)
1020 {
1021 	int r;
1022 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1023 
1024 	sdma_v2_4_init_golden_registers(adev);
1025 
1026 	r = sdma_v2_4_start(adev);
1027 	if (r)
1028 		return r;
1029 
1030 	return r;
1031 }
1032 
1033 static int sdma_v2_4_hw_fini(void *handle)
1034 {
1035 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1036 
1037 	sdma_v2_4_enable(adev, false);
1038 
1039 	return 0;
1040 }
1041 
1042 static int sdma_v2_4_suspend(void *handle)
1043 {
1044 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1045 
1046 	return sdma_v2_4_hw_fini(adev);
1047 }
1048 
1049 static int sdma_v2_4_resume(void *handle)
1050 {
1051 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1052 
1053 	return sdma_v2_4_hw_init(adev);
1054 }
1055 
1056 static bool sdma_v2_4_is_idle(void *handle)
1057 {
1058 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1059 	u32 tmp = RREG32(mmSRBM_STATUS2);
1060 
1061 	if (tmp & (SRBM_STATUS2__SDMA_BUSY_MASK |
1062 		   SRBM_STATUS2__SDMA1_BUSY_MASK))
1063 	    return false;
1064 
1065 	return true;
1066 }
1067 
1068 static int sdma_v2_4_wait_for_idle(void *handle)
1069 {
1070 	unsigned i;
1071 	u32 tmp;
1072 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1073 
1074 	for (i = 0; i < adev->usec_timeout; i++) {
1075 		tmp = RREG32(mmSRBM_STATUS2) & (SRBM_STATUS2__SDMA_BUSY_MASK |
1076 				SRBM_STATUS2__SDMA1_BUSY_MASK);
1077 
1078 		if (!tmp)
1079 			return 0;
1080 		udelay(1);
1081 	}
1082 	return -ETIMEDOUT;
1083 }
1084 
1085 static int sdma_v2_4_soft_reset(void *handle)
1086 {
1087 	u32 srbm_soft_reset = 0;
1088 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1089 	u32 tmp = RREG32(mmSRBM_STATUS2);
1090 
1091 	if (tmp & SRBM_STATUS2__SDMA_BUSY_MASK) {
1092 		/* sdma0 */
1093 		tmp = RREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET);
1094 		tmp = REG_SET_FIELD(tmp, SDMA0_F32_CNTL, HALT, 0);
1095 		WREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET, tmp);
1096 		srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA_MASK;
1097 	}
1098 	if (tmp & SRBM_STATUS2__SDMA1_BUSY_MASK) {
1099 		/* sdma1 */
1100 		tmp = RREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET);
1101 		tmp = REG_SET_FIELD(tmp, SDMA0_F32_CNTL, HALT, 0);
1102 		WREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET, tmp);
1103 		srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA1_MASK;
1104 	}
1105 
1106 	if (srbm_soft_reset) {
1107 		tmp = RREG32(mmSRBM_SOFT_RESET);
1108 		tmp |= srbm_soft_reset;
1109 		dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
1110 		WREG32(mmSRBM_SOFT_RESET, tmp);
1111 		tmp = RREG32(mmSRBM_SOFT_RESET);
1112 
1113 		udelay(50);
1114 
1115 		tmp &= ~srbm_soft_reset;
1116 		WREG32(mmSRBM_SOFT_RESET, tmp);
1117 		tmp = RREG32(mmSRBM_SOFT_RESET);
1118 
1119 		/* Wait a little for things to settle down */
1120 		udelay(50);
1121 	}
1122 
1123 	return 0;
1124 }
1125 
1126 static int sdma_v2_4_set_trap_irq_state(struct amdgpu_device *adev,
1127 					struct amdgpu_irq_src *src,
1128 					unsigned type,
1129 					enum amdgpu_interrupt_state state)
1130 {
1131 	u32 sdma_cntl;
1132 
1133 	switch (type) {
1134 	case AMDGPU_SDMA_IRQ_TRAP0:
1135 		switch (state) {
1136 		case AMDGPU_IRQ_STATE_DISABLE:
1137 			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1138 			sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 0);
1139 			WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1140 			break;
1141 		case AMDGPU_IRQ_STATE_ENABLE:
1142 			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1143 			sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 1);
1144 			WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1145 			break;
1146 		default:
1147 			break;
1148 		}
1149 		break;
1150 	case AMDGPU_SDMA_IRQ_TRAP1:
1151 		switch (state) {
1152 		case AMDGPU_IRQ_STATE_DISABLE:
1153 			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1154 			sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 0);
1155 			WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1156 			break;
1157 		case AMDGPU_IRQ_STATE_ENABLE:
1158 			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1159 			sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 1);
1160 			WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1161 			break;
1162 		default:
1163 			break;
1164 		}
1165 		break;
1166 	default:
1167 		break;
1168 	}
1169 	return 0;
1170 }
1171 
1172 static int sdma_v2_4_process_trap_irq(struct amdgpu_device *adev,
1173 				      struct amdgpu_irq_src *source,
1174 				      struct amdgpu_iv_entry *entry)
1175 {
1176 	u8 instance_id, queue_id;
1177 
1178 	instance_id = (entry->ring_id & 0x3) >> 0;
1179 	queue_id = (entry->ring_id & 0xc) >> 2;
1180 	DRM_DEBUG("IH: SDMA trap\n");
1181 	switch (instance_id) {
1182 	case 0:
1183 		switch (queue_id) {
1184 		case 0:
1185 			amdgpu_fence_process(&adev->sdma.instance[0].ring);
1186 			break;
1187 		case 1:
1188 			/* XXX compute */
1189 			break;
1190 		case 2:
1191 			/* XXX compute */
1192 			break;
1193 		}
1194 		break;
1195 	case 1:
1196 		switch (queue_id) {
1197 		case 0:
1198 			amdgpu_fence_process(&adev->sdma.instance[1].ring);
1199 			break;
1200 		case 1:
1201 			/* XXX compute */
1202 			break;
1203 		case 2:
1204 			/* XXX compute */
1205 			break;
1206 		}
1207 		break;
1208 	}
1209 	return 0;
1210 }
1211 
1212 static int sdma_v2_4_process_illegal_inst_irq(struct amdgpu_device *adev,
1213 					      struct amdgpu_irq_src *source,
1214 					      struct amdgpu_iv_entry *entry)
1215 {
1216 	DRM_ERROR("Illegal instruction in SDMA command stream\n");
1217 	schedule_work(&adev->reset_work);
1218 	return 0;
1219 }
1220 
1221 static int sdma_v2_4_set_clockgating_state(void *handle,
1222 					  enum amd_clockgating_state state)
1223 {
1224 	/* XXX handled via the smc on VI */
1225 	return 0;
1226 }
1227 
1228 static int sdma_v2_4_set_powergating_state(void *handle,
1229 					  enum amd_powergating_state state)
1230 {
1231 	return 0;
1232 }
1233 
1234 const struct amd_ip_funcs sdma_v2_4_ip_funcs = {
1235 	.name = "sdma_v2_4",
1236 	.early_init = sdma_v2_4_early_init,
1237 	.late_init = NULL,
1238 	.sw_init = sdma_v2_4_sw_init,
1239 	.sw_fini = sdma_v2_4_sw_fini,
1240 	.hw_init = sdma_v2_4_hw_init,
1241 	.hw_fini = sdma_v2_4_hw_fini,
1242 	.suspend = sdma_v2_4_suspend,
1243 	.resume = sdma_v2_4_resume,
1244 	.is_idle = sdma_v2_4_is_idle,
1245 	.wait_for_idle = sdma_v2_4_wait_for_idle,
1246 	.soft_reset = sdma_v2_4_soft_reset,
1247 	.set_clockgating_state = sdma_v2_4_set_clockgating_state,
1248 	.set_powergating_state = sdma_v2_4_set_powergating_state,
1249 };
1250 
1251 static const struct amdgpu_ring_funcs sdma_v2_4_ring_funcs = {
1252 	.get_rptr = sdma_v2_4_ring_get_rptr,
1253 	.get_wptr = sdma_v2_4_ring_get_wptr,
1254 	.set_wptr = sdma_v2_4_ring_set_wptr,
1255 	.parse_cs = NULL,
1256 	.emit_ib = sdma_v2_4_ring_emit_ib,
1257 	.emit_fence = sdma_v2_4_ring_emit_fence,
1258 	.emit_pipeline_sync = sdma_v2_4_ring_emit_pipeline_sync,
1259 	.emit_vm_flush = sdma_v2_4_ring_emit_vm_flush,
1260 	.emit_hdp_flush = sdma_v2_4_ring_emit_hdp_flush,
1261 	.emit_hdp_invalidate = sdma_v2_4_ring_emit_hdp_invalidate,
1262 	.test_ring = sdma_v2_4_ring_test_ring,
1263 	.test_ib = sdma_v2_4_ring_test_ib,
1264 	.insert_nop = sdma_v2_4_ring_insert_nop,
1265 	.pad_ib = sdma_v2_4_ring_pad_ib,
1266 };
1267 
1268 static void sdma_v2_4_set_ring_funcs(struct amdgpu_device *adev)
1269 {
1270 	int i;
1271 
1272 	for (i = 0; i < adev->sdma.num_instances; i++)
1273 		adev->sdma.instance[i].ring.funcs = &sdma_v2_4_ring_funcs;
1274 }
1275 
1276 static const struct amdgpu_irq_src_funcs sdma_v2_4_trap_irq_funcs = {
1277 	.set = sdma_v2_4_set_trap_irq_state,
1278 	.process = sdma_v2_4_process_trap_irq,
1279 };
1280 
1281 static const struct amdgpu_irq_src_funcs sdma_v2_4_illegal_inst_irq_funcs = {
1282 	.process = sdma_v2_4_process_illegal_inst_irq,
1283 };
1284 
1285 static void sdma_v2_4_set_irq_funcs(struct amdgpu_device *adev)
1286 {
1287 	adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST;
1288 	adev->sdma.trap_irq.funcs = &sdma_v2_4_trap_irq_funcs;
1289 	adev->sdma.illegal_inst_irq.funcs = &sdma_v2_4_illegal_inst_irq_funcs;
1290 }
1291 
1292 /**
1293  * sdma_v2_4_emit_copy_buffer - copy buffer using the sDMA engine
1294  *
1295  * @ring: amdgpu_ring structure holding ring information
1296  * @src_offset: src GPU address
1297  * @dst_offset: dst GPU address
1298  * @byte_count: number of bytes to xfer
1299  *
1300  * Copy GPU buffers using the DMA engine (VI).
1301  * Used by the amdgpu ttm implementation to move pages if
1302  * registered as the asic copy callback.
1303  */
1304 static void sdma_v2_4_emit_copy_buffer(struct amdgpu_ib *ib,
1305 				       uint64_t src_offset,
1306 				       uint64_t dst_offset,
1307 				       uint32_t byte_count)
1308 {
1309 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1310 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
1311 	ib->ptr[ib->length_dw++] = byte_count;
1312 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1313 	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1314 	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1315 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1316 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1317 }
1318 
1319 /**
1320  * sdma_v2_4_emit_fill_buffer - fill buffer using the sDMA engine
1321  *
1322  * @ring: amdgpu_ring structure holding ring information
1323  * @src_data: value to write to buffer
1324  * @dst_offset: dst GPU address
1325  * @byte_count: number of bytes to xfer
1326  *
1327  * Fill GPU buffers using the DMA engine (VI).
1328  */
1329 static void sdma_v2_4_emit_fill_buffer(struct amdgpu_ib *ib,
1330 				       uint32_t src_data,
1331 				       uint64_t dst_offset,
1332 				       uint32_t byte_count)
1333 {
1334 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL);
1335 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1336 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1337 	ib->ptr[ib->length_dw++] = src_data;
1338 	ib->ptr[ib->length_dw++] = byte_count;
1339 }
1340 
1341 static const struct amdgpu_buffer_funcs sdma_v2_4_buffer_funcs = {
1342 	.copy_max_bytes = 0x1fffff,
1343 	.copy_num_dw = 7,
1344 	.emit_copy_buffer = sdma_v2_4_emit_copy_buffer,
1345 
1346 	.fill_max_bytes = 0x1fffff,
1347 	.fill_num_dw = 7,
1348 	.emit_fill_buffer = sdma_v2_4_emit_fill_buffer,
1349 };
1350 
1351 static void sdma_v2_4_set_buffer_funcs(struct amdgpu_device *adev)
1352 {
1353 	if (adev->mman.buffer_funcs == NULL) {
1354 		adev->mman.buffer_funcs = &sdma_v2_4_buffer_funcs;
1355 		adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1356 	}
1357 }
1358 
1359 static const struct amdgpu_vm_pte_funcs sdma_v2_4_vm_pte_funcs = {
1360 	.copy_pte = sdma_v2_4_vm_copy_pte,
1361 	.write_pte = sdma_v2_4_vm_write_pte,
1362 	.set_pte_pde = sdma_v2_4_vm_set_pte_pde,
1363 };
1364 
1365 static void sdma_v2_4_set_vm_pte_funcs(struct amdgpu_device *adev)
1366 {
1367 	unsigned i;
1368 
1369 	if (adev->vm_manager.vm_pte_funcs == NULL) {
1370 		adev->vm_manager.vm_pte_funcs = &sdma_v2_4_vm_pte_funcs;
1371 		for (i = 0; i < adev->sdma.num_instances; i++)
1372 			adev->vm_manager.vm_pte_rings[i] =
1373 				&adev->sdma.instance[i].ring;
1374 
1375 		adev->vm_manager.vm_pte_num_rings = adev->sdma.num_instances;
1376 	}
1377 }
1378