1 /*
2  * Copyright 2019 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  */
23 
24 #include <linux/delay.h>
25 #include <linux/firmware.h>
26 #include <linux/module.h>
27 #include <linux/pci.h>
28 
29 #include "amdgpu.h"
30 #include "amdgpu_ucode.h"
31 #include "amdgpu_trace.h"
32 
33 #include "gc/gc_10_3_0_offset.h"
34 #include "gc/gc_10_3_0_sh_mask.h"
35 #include "ivsrcid/sdma0/irqsrcs_sdma0_5_0.h"
36 #include "ivsrcid/sdma1/irqsrcs_sdma1_5_0.h"
37 #include "ivsrcid/sdma2/irqsrcs_sdma2_5_0.h"
38 #include "ivsrcid/sdma3/irqsrcs_sdma3_5_0.h"
39 
40 #include "soc15_common.h"
41 #include "soc15.h"
42 #include "navi10_sdma_pkt_open.h"
43 #include "nbio_v2_3.h"
44 #include "sdma_common.h"
45 #include "sdma_v5_2.h"
46 
47 MODULE_FIRMWARE("amdgpu/sienna_cichlid_sdma.bin");
48 MODULE_FIRMWARE("amdgpu/navy_flounder_sdma.bin");
49 MODULE_FIRMWARE("amdgpu/dimgrey_cavefish_sdma.bin");
50 MODULE_FIRMWARE("amdgpu/beige_goby_sdma.bin");
51 
52 MODULE_FIRMWARE("amdgpu/vangogh_sdma.bin");
53 MODULE_FIRMWARE("amdgpu/yellow_carp_sdma.bin");
54 MODULE_FIRMWARE("amdgpu/sdma_5_2_6.bin");
55 MODULE_FIRMWARE("amdgpu/sdma_5_2_7.bin");
56 
57 #define SDMA1_REG_OFFSET 0x600
58 #define SDMA3_REG_OFFSET 0x400
59 #define SDMA0_HYP_DEC_REG_START 0x5880
60 #define SDMA0_HYP_DEC_REG_END 0x5893
61 #define SDMA1_HYP_DEC_REG_OFFSET 0x20
62 
63 static void sdma_v5_2_set_ring_funcs(struct amdgpu_device *adev);
64 static void sdma_v5_2_set_buffer_funcs(struct amdgpu_device *adev);
65 static void sdma_v5_2_set_vm_pte_funcs(struct amdgpu_device *adev);
66 static void sdma_v5_2_set_irq_funcs(struct amdgpu_device *adev);
67 
68 static u32 sdma_v5_2_get_reg_offset(struct amdgpu_device *adev, u32 instance, u32 internal_offset)
69 {
70 	u32 base;
71 
72 	if (internal_offset >= SDMA0_HYP_DEC_REG_START &&
73 	    internal_offset <= SDMA0_HYP_DEC_REG_END) {
74 		base = adev->reg_offset[GC_HWIP][0][1];
75 		if (instance != 0)
76 			internal_offset += SDMA1_HYP_DEC_REG_OFFSET * instance;
77 	} else {
78 		if (instance < 2) {
79 			base = adev->reg_offset[GC_HWIP][0][0];
80 			if (instance == 1)
81 				internal_offset += SDMA1_REG_OFFSET;
82 		} else {
83 			base = adev->reg_offset[GC_HWIP][0][2];
84 			if (instance == 3)
85 				internal_offset += SDMA3_REG_OFFSET;
86 		}
87 	}
88 
89 	return base + internal_offset;
90 }
91 
92 static int sdma_v5_2_init_inst_ctx(struct amdgpu_sdma_instance *sdma_inst)
93 {
94 	int err = 0;
95 	const struct sdma_firmware_header_v1_0 *hdr;
96 
97 	err = amdgpu_ucode_validate(sdma_inst->fw);
98 	if (err)
99 		return err;
100 
101 	hdr = (const struct sdma_firmware_header_v1_0 *)sdma_inst->fw->data;
102 	sdma_inst->fw_version = le32_to_cpu(hdr->header.ucode_version);
103 	sdma_inst->feature_version = le32_to_cpu(hdr->ucode_feature_version);
104 
105 	if (sdma_inst->feature_version >= 20)
106 		sdma_inst->burst_nop = true;
107 
108 	return 0;
109 }
110 
111 static void sdma_v5_2_destroy_inst_ctx(struct amdgpu_device *adev)
112 {
113 	release_firmware(adev->sdma.instance[0].fw);
114 
115 	memset((void *)adev->sdma.instance, 0,
116 	       sizeof(struct amdgpu_sdma_instance) * AMDGPU_MAX_SDMA_INSTANCES);
117 }
118 
119 /**
120  * sdma_v5_2_init_microcode - load ucode images from disk
121  *
122  * @adev: amdgpu_device pointer
123  *
124  * Use the firmware interface to load the ucode images into
125  * the driver (not loaded into hw).
126  * Returns 0 on success, error on failure.
127  */
128 
129 // emulation only, won't work on real chip
130 // navi10 real chip need to use PSP to load firmware
131 static int sdma_v5_2_init_microcode(struct amdgpu_device *adev)
132 {
133 	const char *chip_name;
134 	char fw_name[40];
135 	int err = 0, i;
136 	struct amdgpu_firmware_info *info = NULL;
137 	const struct common_firmware_header *header = NULL;
138 
139 	DRM_DEBUG("\n");
140 
141 	switch (adev->ip_versions[SDMA0_HWIP][0]) {
142 	case IP_VERSION(5, 2, 0):
143 		chip_name = "sienna_cichlid_sdma";
144 		break;
145 	case IP_VERSION(5, 2, 2):
146 		chip_name = "navy_flounder_sdma";
147 		break;
148 	case IP_VERSION(5, 2, 1):
149 		chip_name = "vangogh_sdma";
150 		break;
151 	case IP_VERSION(5, 2, 4):
152 		chip_name = "dimgrey_cavefish_sdma";
153 		break;
154 	case IP_VERSION(5, 2, 5):
155 		chip_name = "beige_goby_sdma";
156 		break;
157 	case IP_VERSION(5, 2, 3):
158 		chip_name = "yellow_carp_sdma";
159 		break;
160 	case IP_VERSION(5, 2, 6):
161 		chip_name = "sdma_5_2_6";
162 		break;
163 	case IP_VERSION(5, 2, 7):
164 		chip_name = "sdma_5_2_7";
165 		break;
166 	default:
167 		BUG();
168 	}
169 
170 	snprintf(fw_name, sizeof(fw_name), "amdgpu/%s.bin", chip_name);
171 
172 	err = request_firmware(&adev->sdma.instance[0].fw, fw_name, adev->dev);
173 	if (err)
174 		goto out;
175 
176 	err = sdma_v5_2_init_inst_ctx(&adev->sdma.instance[0]);
177 	if (err)
178 		goto out;
179 
180 	for (i = 1; i < adev->sdma.num_instances; i++)
181 		memcpy((void *)&adev->sdma.instance[i],
182 		       (void *)&adev->sdma.instance[0],
183 		       sizeof(struct amdgpu_sdma_instance));
184 
185 	if (amdgpu_sriov_vf(adev) && (adev->ip_versions[SDMA0_HWIP][0] == IP_VERSION(5, 2, 0)))
186 		return 0;
187 
188 	DRM_DEBUG("psp_load == '%s'\n",
189 		  adev->firmware.load_type == AMDGPU_FW_LOAD_PSP ? "true" : "false");
190 
191 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
192 		for (i = 0; i < adev->sdma.num_instances; i++) {
193 			info = &adev->firmware.ucode[AMDGPU_UCODE_ID_SDMA0 + i];
194 			info->ucode_id = AMDGPU_UCODE_ID_SDMA0 + i;
195 			info->fw = adev->sdma.instance[i].fw;
196 			header = (const struct common_firmware_header *)info->fw->data;
197 			adev->firmware.fw_size +=
198 				ALIGN(le32_to_cpu(header->ucode_size_bytes), PAGE_SIZE);
199 		}
200 	}
201 
202 out:
203 	if (err) {
204 		DRM_ERROR("sdma_v5_2: Failed to load firmware \"%s\"\n", fw_name);
205 		sdma_v5_2_destroy_inst_ctx(adev);
206 	}
207 	return err;
208 }
209 
210 static unsigned sdma_v5_2_ring_init_cond_exec(struct amdgpu_ring *ring)
211 {
212 	unsigned ret;
213 
214 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_COND_EXE));
215 	amdgpu_ring_write(ring, lower_32_bits(ring->cond_exe_gpu_addr));
216 	amdgpu_ring_write(ring, upper_32_bits(ring->cond_exe_gpu_addr));
217 	amdgpu_ring_write(ring, 1);
218 	ret = ring->wptr & ring->buf_mask;/* this is the offset we need patch later */
219 	amdgpu_ring_write(ring, 0x55aa55aa);/* insert dummy here and patch it later */
220 
221 	return ret;
222 }
223 
224 static void sdma_v5_2_ring_patch_cond_exec(struct amdgpu_ring *ring,
225 					   unsigned offset)
226 {
227 	unsigned cur;
228 
229 	BUG_ON(offset > ring->buf_mask);
230 	BUG_ON(ring->ring[offset] != 0x55aa55aa);
231 
232 	cur = (ring->wptr - 1) & ring->buf_mask;
233 	if (cur > offset)
234 		ring->ring[offset] = cur - offset;
235 	else
236 		ring->ring[offset] = (ring->buf_mask + 1) - offset + cur;
237 }
238 
239 /**
240  * sdma_v5_2_ring_get_rptr - get the current read pointer
241  *
242  * @ring: amdgpu ring pointer
243  *
244  * Get the current rptr from the hardware (NAVI10+).
245  */
246 static uint64_t sdma_v5_2_ring_get_rptr(struct amdgpu_ring *ring)
247 {
248 	u64 *rptr;
249 
250 	/* XXX check if swapping is necessary on BE */
251 	rptr = (u64 *)ring->rptr_cpu_addr;
252 
253 	DRM_DEBUG("rptr before shift == 0x%016llx\n", *rptr);
254 	return ((*rptr) >> 2);
255 }
256 
257 /**
258  * sdma_v5_2_ring_get_wptr - get the current write pointer
259  *
260  * @ring: amdgpu ring pointer
261  *
262  * Get the current wptr from the hardware (NAVI10+).
263  */
264 static uint64_t sdma_v5_2_ring_get_wptr(struct amdgpu_ring *ring)
265 {
266 	struct amdgpu_device *adev = ring->adev;
267 	u64 wptr;
268 
269 	if (ring->use_doorbell) {
270 		/* XXX check if swapping is necessary on BE */
271 		wptr = READ_ONCE(*((u64 *)ring->wptr_cpu_addr));
272 		DRM_DEBUG("wptr/doorbell before shift == 0x%016llx\n", wptr);
273 	} else {
274 		wptr = RREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI));
275 		wptr = wptr << 32;
276 		wptr |= RREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR));
277 		DRM_DEBUG("wptr before shift [%i] wptr == 0x%016llx\n", ring->me, wptr);
278 	}
279 
280 	return wptr >> 2;
281 }
282 
283 /**
284  * sdma_v5_2_ring_set_wptr - commit the write pointer
285  *
286  * @ring: amdgpu ring pointer
287  *
288  * Write the wptr back to the hardware (NAVI10+).
289  */
290 static void sdma_v5_2_ring_set_wptr(struct amdgpu_ring *ring)
291 {
292 	struct amdgpu_device *adev = ring->adev;
293 
294 	DRM_DEBUG("Setting write pointer\n");
295 	if (ring->use_doorbell) {
296 		DRM_DEBUG("Using doorbell -- "
297 				"wptr_offs == 0x%08x "
298 				"lower_32_bits(ring->wptr << 2) == 0x%08x "
299 				"upper_32_bits(ring->wptr << 2) == 0x%08x\n",
300 				ring->wptr_offs,
301 				lower_32_bits(ring->wptr << 2),
302 				upper_32_bits(ring->wptr << 2));
303 		/* XXX check if swapping is necessary on BE */
304 		atomic64_set((atomic64_t *)ring->wptr_cpu_addr,
305 			     ring->wptr << 2);
306 		DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016llx)\n",
307 				ring->doorbell_index, ring->wptr << 2);
308 		WDOORBELL64(ring->doorbell_index, ring->wptr << 2);
309 	} else {
310 		DRM_DEBUG("Not using doorbell -- "
311 				"mmSDMA%i_GFX_RB_WPTR == 0x%08x "
312 				"mmSDMA%i_GFX_RB_WPTR_HI == 0x%08x\n",
313 				ring->me,
314 				lower_32_bits(ring->wptr << 2),
315 				ring->me,
316 				upper_32_bits(ring->wptr << 2));
317 		WREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR),
318 			lower_32_bits(ring->wptr << 2));
319 		WREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI),
320 			upper_32_bits(ring->wptr << 2));
321 	}
322 }
323 
324 static void sdma_v5_2_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
325 {
326 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
327 	int i;
328 
329 	for (i = 0; i < count; i++)
330 		if (sdma && sdma->burst_nop && (i == 0))
331 			amdgpu_ring_write(ring, ring->funcs->nop |
332 				SDMA_PKT_NOP_HEADER_COUNT(count - 1));
333 		else
334 			amdgpu_ring_write(ring, ring->funcs->nop);
335 }
336 
337 /**
338  * sdma_v5_2_ring_emit_ib - Schedule an IB on the DMA engine
339  *
340  * @ring: amdgpu ring pointer
341  * @job: job to retrieve vmid from
342  * @ib: IB object to schedule
343  * @flags: unused
344  *
345  * Schedule an IB in the DMA ring.
346  */
347 static void sdma_v5_2_ring_emit_ib(struct amdgpu_ring *ring,
348 				   struct amdgpu_job *job,
349 				   struct amdgpu_ib *ib,
350 				   uint32_t flags)
351 {
352 	unsigned vmid = AMDGPU_JOB_GET_VMID(job);
353 	uint64_t csa_mc_addr = amdgpu_sdma_get_csa_mc_addr(ring, vmid);
354 
355 	/* An IB packet must end on a 8 DW boundary--the next dword
356 	 * must be on a 8-dword boundary. Our IB packet below is 6
357 	 * dwords long, thus add x number of NOPs, such that, in
358 	 * modular arithmetic,
359 	 * wptr + 6 + x = 8k, k >= 0, which in C is,
360 	 * (wptr + 6 + x) % 8 = 0.
361 	 * The expression below, is a solution of x.
362 	 */
363 	sdma_v5_2_ring_insert_nop(ring, (2 - lower_32_bits(ring->wptr)) & 7);
364 
365 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) |
366 			  SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf));
367 	/* base must be 32 byte aligned */
368 	amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
369 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
370 	amdgpu_ring_write(ring, ib->length_dw);
371 	amdgpu_ring_write(ring, lower_32_bits(csa_mc_addr));
372 	amdgpu_ring_write(ring, upper_32_bits(csa_mc_addr));
373 }
374 
375 /**
376  * sdma_v5_2_ring_emit_mem_sync - flush the IB by graphics cache rinse
377  *
378  * @ring: amdgpu ring pointer
379  *
380  * flush the IB by graphics cache rinse.
381  */
382 static void sdma_v5_2_ring_emit_mem_sync(struct amdgpu_ring *ring)
383 {
384 	uint32_t gcr_cntl = SDMA_GCR_GL2_INV | SDMA_GCR_GL2_WB |
385 			    SDMA_GCR_GLM_INV | SDMA_GCR_GL1_INV |
386 			    SDMA_GCR_GLV_INV | SDMA_GCR_GLK_INV |
387 			    SDMA_GCR_GLI_INV(1);
388 
389 	/* flush entire cache L0/L1/L2, this can be optimized by performance requirement */
390 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_GCR_REQ));
391 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD1_BASE_VA_31_7(0));
392 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD2_GCR_CONTROL_15_0(gcr_cntl) |
393 			SDMA_PKT_GCR_REQ_PAYLOAD2_BASE_VA_47_32(0));
394 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD3_LIMIT_VA_31_7(0) |
395 			SDMA_PKT_GCR_REQ_PAYLOAD3_GCR_CONTROL_18_16(gcr_cntl >> 16));
396 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD4_LIMIT_VA_47_32(0) |
397 			SDMA_PKT_GCR_REQ_PAYLOAD4_VMID(0));
398 }
399 
400 /**
401  * sdma_v5_2_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
402  *
403  * @ring: amdgpu ring pointer
404  *
405  * Emit an hdp flush packet on the requested DMA ring.
406  */
407 static void sdma_v5_2_ring_emit_hdp_flush(struct amdgpu_ring *ring)
408 {
409 	struct amdgpu_device *adev = ring->adev;
410 	u32 ref_and_mask = 0;
411 	const struct nbio_hdp_flush_reg *nbio_hf_reg = adev->nbio.hdp_flush_reg;
412 
413 	ref_and_mask = nbio_hf_reg->ref_and_mask_sdma0 << ring->me;
414 
415 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
416 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) |
417 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
418 	amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_done_offset(adev)) << 2);
419 	amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_req_offset(adev)) << 2);
420 	amdgpu_ring_write(ring, ref_and_mask); /* reference */
421 	amdgpu_ring_write(ring, ref_and_mask); /* mask */
422 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
423 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
424 }
425 
426 /**
427  * sdma_v5_2_ring_emit_fence - emit a fence on the DMA ring
428  *
429  * @ring: amdgpu ring pointer
430  * @addr: address
431  * @seq: sequence number
432  * @flags: fence related flags
433  *
434  * Add a DMA fence packet to the ring to write
435  * the fence seq number and DMA trap packet to generate
436  * an interrupt if needed.
437  */
438 static void sdma_v5_2_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
439 				      unsigned flags)
440 {
441 	bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
442 	/* write the fence */
443 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) |
444 			  SDMA_PKT_FENCE_HEADER_MTYPE(0x3)); /* Ucached(UC) */
445 	/* zero in first two bits */
446 	BUG_ON(addr & 0x3);
447 	amdgpu_ring_write(ring, lower_32_bits(addr));
448 	amdgpu_ring_write(ring, upper_32_bits(addr));
449 	amdgpu_ring_write(ring, lower_32_bits(seq));
450 
451 	/* optionally write high bits as well */
452 	if (write64bit) {
453 		addr += 4;
454 		amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) |
455 				  SDMA_PKT_FENCE_HEADER_MTYPE(0x3));
456 		/* zero in first two bits */
457 		BUG_ON(addr & 0x3);
458 		amdgpu_ring_write(ring, lower_32_bits(addr));
459 		amdgpu_ring_write(ring, upper_32_bits(addr));
460 		amdgpu_ring_write(ring, upper_32_bits(seq));
461 	}
462 
463 	if ((flags & AMDGPU_FENCE_FLAG_INT)) {
464 		uint32_t ctx = ring->is_mes_queue ?
465 			(ring->hw_queue_id | AMDGPU_FENCE_MES_QUEUE_FLAG) : 0;
466 		/* generate an interrupt */
467 		amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP));
468 		amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(ctx));
469 	}
470 }
471 
472 /**
473  * sdma_v5_2_gfx_stop - stop the gfx async dma engines
474  *
475  * @adev: amdgpu_device pointer
476  *
477  * Stop the gfx async dma ring buffers.
478  */
479 static void sdma_v5_2_gfx_stop(struct amdgpu_device *adev)
480 {
481 	struct amdgpu_ring *sdma0 = &adev->sdma.instance[0].ring;
482 	struct amdgpu_ring *sdma1 = &adev->sdma.instance[1].ring;
483 	struct amdgpu_ring *sdma2 = &adev->sdma.instance[2].ring;
484 	struct amdgpu_ring *sdma3 = &adev->sdma.instance[3].ring;
485 	u32 rb_cntl, ib_cntl;
486 	int i;
487 
488 	if ((adev->mman.buffer_funcs_ring == sdma0) ||
489 	    (adev->mman.buffer_funcs_ring == sdma1) ||
490 	    (adev->mman.buffer_funcs_ring == sdma2) ||
491 	    (adev->mman.buffer_funcs_ring == sdma3))
492 		amdgpu_ttm_set_buffer_funcs_status(adev, false);
493 
494 	for (i = 0; i < adev->sdma.num_instances; i++) {
495 		rb_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL));
496 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 0);
497 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
498 		ib_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL));
499 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 0);
500 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl);
501 	}
502 }
503 
504 /**
505  * sdma_v5_2_rlc_stop - stop the compute async dma engines
506  *
507  * @adev: amdgpu_device pointer
508  *
509  * Stop the compute async dma queues.
510  */
511 static void sdma_v5_2_rlc_stop(struct amdgpu_device *adev)
512 {
513 	/* XXX todo */
514 }
515 
516 /**
517  * sdma_v5_2_ctx_switch_enable_for_instance - start the async dma engines
518  * context switch for an instance
519  *
520  * @adev: amdgpu_device pointer
521  * @instance_idx: the index of the SDMA instance
522  *
523  * Unhalt the async dma engines context switch.
524  */
525 static void sdma_v5_2_ctx_switch_enable_for_instance(struct amdgpu_device *adev, int instance_idx)
526 {
527 	u32 f32_cntl, phase_quantum = 0;
528 
529 	if (WARN_ON(instance_idx >= adev->sdma.num_instances)) {
530 		return;
531 	}
532 
533 	if (amdgpu_sdma_phase_quantum) {
534 		unsigned value = amdgpu_sdma_phase_quantum;
535 		unsigned unit = 0;
536 
537 		while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
538 				SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) {
539 			value = (value + 1) >> 1;
540 			unit++;
541 		}
542 		if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
543 			    SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) {
544 			value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
545 				 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT);
546 			unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
547 				SDMA0_PHASE0_QUANTUM__UNIT__SHIFT);
548 			WARN_ONCE(1,
549 			"clamping sdma_phase_quantum to %uK clock cycles\n",
550 				  value << unit);
551 		}
552 		phase_quantum =
553 			value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
554 			unit  << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
555 
556 		WREG32_SOC15_IP(GC,
557 			sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_PHASE0_QUANTUM),
558 			phase_quantum);
559 		WREG32_SOC15_IP(GC,
560 			sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_PHASE1_QUANTUM),
561 		    phase_quantum);
562 		WREG32_SOC15_IP(GC,
563 			sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_PHASE2_QUANTUM),
564 		    phase_quantum);
565 	}
566 
567 	if (!amdgpu_sriov_vf(adev)) {
568 		f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_CNTL));
569 		f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
570 				AUTO_CTXSW_ENABLE, 1);
571 		WREG32(sdma_v5_2_get_reg_offset(adev, instance_idx, mmSDMA0_CNTL), f32_cntl);
572 	}
573 }
574 
575 /**
576  * sdma_v5_2_ctx_switch_disable_all - stop the async dma engines context switch
577  *
578  * @adev: amdgpu_device pointer
579  *
580  * Halt the async dma engines context switch.
581  */
582 static void sdma_v5_2_ctx_switch_disable_all(struct amdgpu_device *adev)
583 {
584 	u32 f32_cntl;
585 	int i;
586 
587 	if (amdgpu_sriov_vf(adev))
588 		return;
589 
590 	for (i = 0; i < adev->sdma.num_instances; i++) {
591 		f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
592 		f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
593 				AUTO_CTXSW_ENABLE, 0);
594 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl);
595 	}
596 }
597 
598 /**
599  * sdma_v5_2_halt - stop the async dma engines
600  *
601  * @adev: amdgpu_device pointer
602  *
603  * Halt the async dma engines.
604  */
605 static void sdma_v5_2_halt(struct amdgpu_device *adev)
606 {
607 	int i;
608 	u32 f32_cntl;
609 
610 	sdma_v5_2_gfx_stop(adev);
611 	sdma_v5_2_rlc_stop(adev);
612 
613 	if (!amdgpu_sriov_vf(adev)) {
614 		for (i = 0; i < adev->sdma.num_instances; i++) {
615 			f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
616 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 1);
617 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl);
618 		}
619 	}
620 }
621 
622 /**
623  * sdma_v5_2_gfx_resume - setup and start the async dma engines
624  *
625  * @adev: amdgpu_device pointer
626  *
627  * Set up the gfx DMA ring buffers and enable them.
628  * It assumes that the dma engine is stopped for each instance.
629  * The function enables the engine and preemptions sequentially for each instance.
630  *
631  * Returns 0 for success, error for failure.
632  */
633 static int sdma_v5_2_gfx_resume(struct amdgpu_device *adev)
634 {
635 	struct amdgpu_ring *ring;
636 	u32 rb_cntl, ib_cntl;
637 	u32 rb_bufsz;
638 	u32 doorbell;
639 	u32 doorbell_offset;
640 	u32 temp;
641 	u32 wptr_poll_cntl;
642 	u64 wptr_gpu_addr;
643 	int i, r;
644 
645 	for (i = 0; i < adev->sdma.num_instances; i++) {
646 		ring = &adev->sdma.instance[i].ring;
647 
648 		if (!amdgpu_sriov_vf(adev))
649 			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL), 0);
650 
651 		/* Set ring buffer size in dwords */
652 		rb_bufsz = order_base_2(ring->ring_size / 4);
653 		rb_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL));
654 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz);
655 #ifdef __BIG_ENDIAN
656 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1);
657 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL,
658 					RPTR_WRITEBACK_SWAP_ENABLE, 1);
659 #endif
660 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
661 
662 		/* Initialize the ring buffer's read and write pointers */
663 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR), 0);
664 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_HI), 0);
665 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), 0);
666 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), 0);
667 
668 		/* setup the wptr shadow polling */
669 		wptr_gpu_addr = ring->wptr_gpu_addr;
670 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_LO),
671 		       lower_32_bits(wptr_gpu_addr));
672 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_HI),
673 		       upper_32_bits(wptr_gpu_addr));
674 		wptr_poll_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i,
675 							 mmSDMA0_GFX_RB_WPTR_POLL_CNTL));
676 		wptr_poll_cntl = REG_SET_FIELD(wptr_poll_cntl,
677 					       SDMA0_GFX_RB_WPTR_POLL_CNTL,
678 					       F32_POLL_ENABLE, 1);
679 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_CNTL),
680 		       wptr_poll_cntl);
681 
682 		/* set the wb address whether it's enabled or not */
683 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_HI),
684 		       upper_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFF);
685 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_LO),
686 		       lower_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFC);
687 
688 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
689 
690 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE), ring->gpu_addr >> 8);
691 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE_HI), ring->gpu_addr >> 40);
692 
693 		ring->wptr = 0;
694 
695 		/* before programing wptr to a less value, need set minor_ptr_update first */
696 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 1);
697 
698 		if (!amdgpu_sriov_vf(adev)) { /* only bare-metal use register write for wptr */
699 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), lower_32_bits(ring->wptr << 2));
700 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), upper_32_bits(ring->wptr << 2));
701 		}
702 
703 		doorbell = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL));
704 		doorbell_offset = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL_OFFSET));
705 
706 		if (ring->use_doorbell) {
707 			doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 1);
708 			doorbell_offset = REG_SET_FIELD(doorbell_offset, SDMA0_GFX_DOORBELL_OFFSET,
709 					OFFSET, ring->doorbell_index);
710 		} else {
711 			doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 0);
712 		}
713 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL), doorbell);
714 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL_OFFSET), doorbell_offset);
715 
716 		adev->nbio.funcs->sdma_doorbell_range(adev, i, ring->use_doorbell,
717 						      ring->doorbell_index,
718 						      adev->doorbell_index.sdma_doorbell_range);
719 
720 		if (amdgpu_sriov_vf(adev))
721 			sdma_v5_2_ring_set_wptr(ring);
722 
723 		/* set minor_ptr_update to 0 after wptr programed */
724 
725 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 0);
726 
727 		/* SRIOV VF has no control of any of registers below */
728 		if (!amdgpu_sriov_vf(adev)) {
729 			/* set utc l1 enable flag always to 1 */
730 			temp = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
731 			temp = REG_SET_FIELD(temp, SDMA0_CNTL, UTC_L1_ENABLE, 1);
732 
733 			/* enable MCBP */
734 			temp = REG_SET_FIELD(temp, SDMA0_CNTL, MIDCMD_PREEMPT_ENABLE, 1);
735 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), temp);
736 
737 			/* Set up RESP_MODE to non-copy addresses */
738 			temp = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL));
739 			temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, RESP_MODE, 3);
740 			temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, REDO_DELAY, 9);
741 			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL), temp);
742 
743 			/* program default cache read and write policy */
744 			temp = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE));
745 			/* clean read policy and write policy bits */
746 			temp &= 0xFF0FFF;
747 			temp |= ((CACHE_READ_POLICY_L2__DEFAULT << 12) |
748 				 (CACHE_WRITE_POLICY_L2__DEFAULT << 14) |
749 				 SDMA0_UTCL1_PAGE__LLC_NOALLOC_MASK);
750 			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE), temp);
751 
752 			/* unhalt engine */
753 			temp = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
754 			temp = REG_SET_FIELD(temp, SDMA0_F32_CNTL, HALT, 0);
755 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), temp);
756 		}
757 
758 		/* enable DMA RB */
759 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1);
760 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
761 
762 		ib_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL));
763 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 1);
764 #ifdef __BIG_ENDIAN
765 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1);
766 #endif
767 		/* enable DMA IBs */
768 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl);
769 
770 		ring->sched.ready = true;
771 
772 		sdma_v5_2_ctx_switch_enable_for_instance(adev, i);
773 
774 		r = amdgpu_ring_test_ring(ring);
775 		if (r) {
776 			ring->sched.ready = false;
777 			return r;
778 		}
779 
780 		if (adev->mman.buffer_funcs_ring == ring)
781 			amdgpu_ttm_set_buffer_funcs_status(adev, true);
782 	}
783 
784 	return 0;
785 }
786 
787 /**
788  * sdma_v5_2_rlc_resume - setup and start the async dma engines
789  *
790  * @adev: amdgpu_device pointer
791  *
792  * Set up the compute DMA queues and enable them.
793  * Returns 0 for success, error for failure.
794  */
795 static int sdma_v5_2_rlc_resume(struct amdgpu_device *adev)
796 {
797 	return 0;
798 }
799 
800 /**
801  * sdma_v5_2_load_microcode - load the sDMA ME ucode
802  *
803  * @adev: amdgpu_device pointer
804  *
805  * Loads the sDMA0/1/2/3 ucode.
806  * Returns 0 for success, -EINVAL if the ucode is not available.
807  */
808 static int sdma_v5_2_load_microcode(struct amdgpu_device *adev)
809 {
810 	const struct sdma_firmware_header_v1_0 *hdr;
811 	const __le32 *fw_data;
812 	u32 fw_size;
813 	int i, j;
814 
815 	/* halt the MEs */
816 	sdma_v5_2_halt(adev);
817 
818 	for (i = 0; i < adev->sdma.num_instances; i++) {
819 		if (!adev->sdma.instance[i].fw)
820 			return -EINVAL;
821 
822 		hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
823 		amdgpu_ucode_print_sdma_hdr(&hdr->header);
824 		fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
825 
826 		fw_data = (const __le32 *)
827 			(adev->sdma.instance[i].fw->data +
828 				le32_to_cpu(hdr->header.ucode_array_offset_bytes));
829 
830 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), 0);
831 
832 		for (j = 0; j < fw_size; j++) {
833 			if (amdgpu_emu_mode == 1 && j % 500 == 0)
834 				msleep(1);
835 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UCODE_DATA), le32_to_cpup(fw_data++));
836 		}
837 
838 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), adev->sdma.instance[i].fw_version);
839 	}
840 
841 	return 0;
842 }
843 
844 static int sdma_v5_2_soft_reset(void *handle)
845 {
846 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
847 	u32 grbm_soft_reset;
848 	u32 tmp;
849 	int i;
850 
851 	for (i = 0; i < adev->sdma.num_instances; i++) {
852 		grbm_soft_reset = REG_SET_FIELD(0,
853 						GRBM_SOFT_RESET, SOFT_RESET_SDMA0,
854 						1);
855 		grbm_soft_reset <<= i;
856 
857 		tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET);
858 		tmp |= grbm_soft_reset;
859 		DRM_DEBUG("GRBM_SOFT_RESET=0x%08X\n", tmp);
860 		WREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET, tmp);
861 		tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET);
862 
863 		udelay(50);
864 
865 		tmp &= ~grbm_soft_reset;
866 		WREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET, tmp);
867 		tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET);
868 
869 		udelay(50);
870 	}
871 
872 	return 0;
873 }
874 
875 /**
876  * sdma_v5_2_start - setup and start the async dma engines
877  *
878  * @adev: amdgpu_device pointer
879  *
880  * Set up the DMA engines and enable them.
881  * Returns 0 for success, error for failure.
882  */
883 static int sdma_v5_2_start(struct amdgpu_device *adev)
884 {
885 	int r = 0;
886 
887 	if (amdgpu_sriov_vf(adev)) {
888 		sdma_v5_2_ctx_switch_disable_all(adev);
889 		sdma_v5_2_halt(adev);
890 
891 		/* set RB registers */
892 		r = sdma_v5_2_gfx_resume(adev);
893 		return r;
894 	}
895 
896 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
897 		r = sdma_v5_2_load_microcode(adev);
898 		if (r)
899 			return r;
900 
901 		/* The value of mmSDMA_F32_CNTL is invalid the moment after loading fw */
902 		if (amdgpu_emu_mode == 1)
903 			msleep(1000);
904 	}
905 
906 	/* TODO: check whether can submit a doorbell request to raise
907 	 * a doorbell fence to exit gfxoff.
908 	 */
909 	if (adev->in_s0ix)
910 		amdgpu_gfx_off_ctrl(adev, false);
911 
912 	sdma_v5_2_soft_reset(adev);
913 
914 	/* Soft reset supposes to disable the dma engine and preemption.
915 	 * Now start the gfx rings and rlc compute queues.
916 	 */
917 	r = sdma_v5_2_gfx_resume(adev);
918 	if (adev->in_s0ix)
919 		amdgpu_gfx_off_ctrl(adev, true);
920 	if (r)
921 		return r;
922 	r = sdma_v5_2_rlc_resume(adev);
923 
924 	return r;
925 }
926 
927 static int sdma_v5_2_mqd_init(struct amdgpu_device *adev, void *mqd,
928 			      struct amdgpu_mqd_prop *prop)
929 {
930 	struct v10_sdma_mqd *m = mqd;
931 	uint64_t wb_gpu_addr;
932 
933 	m->sdmax_rlcx_rb_cntl =
934 		order_base_2(prop->queue_size / 4) << SDMA0_RLC0_RB_CNTL__RB_SIZE__SHIFT |
935 		1 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_ENABLE__SHIFT |
936 		6 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_TIMER__SHIFT |
937 		1 << SDMA0_RLC0_RB_CNTL__RB_PRIV__SHIFT;
938 
939 	m->sdmax_rlcx_rb_base = lower_32_bits(prop->hqd_base_gpu_addr >> 8);
940 	m->sdmax_rlcx_rb_base_hi = upper_32_bits(prop->hqd_base_gpu_addr >> 8);
941 
942 	m->sdmax_rlcx_rb_wptr_poll_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, 0,
943 						  mmSDMA0_GFX_RB_WPTR_POLL_CNTL));
944 
945 	wb_gpu_addr = prop->wptr_gpu_addr;
946 	m->sdmax_rlcx_rb_wptr_poll_addr_lo = lower_32_bits(wb_gpu_addr);
947 	m->sdmax_rlcx_rb_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr);
948 
949 	wb_gpu_addr = prop->rptr_gpu_addr;
950 	m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits(wb_gpu_addr);
951 	m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits(wb_gpu_addr);
952 
953 	m->sdmax_rlcx_ib_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, 0,
954 							mmSDMA0_GFX_IB_CNTL));
955 
956 	m->sdmax_rlcx_doorbell_offset =
957 		prop->doorbell_index << SDMA0_RLC0_DOORBELL_OFFSET__OFFSET__SHIFT;
958 
959 	m->sdmax_rlcx_doorbell = REG_SET_FIELD(0, SDMA0_RLC0_DOORBELL, ENABLE, 1);
960 
961 	return 0;
962 }
963 
964 static void sdma_v5_2_set_mqd_funcs(struct amdgpu_device *adev)
965 {
966 	adev->mqds[AMDGPU_HW_IP_DMA].mqd_size = sizeof(struct v10_sdma_mqd);
967 	adev->mqds[AMDGPU_HW_IP_DMA].init_mqd = sdma_v5_2_mqd_init;
968 }
969 
970 /**
971  * sdma_v5_2_ring_test_ring - simple async dma engine test
972  *
973  * @ring: amdgpu_ring structure holding ring information
974  *
975  * Test the DMA engine by writing using it to write an
976  * value to memory.
977  * Returns 0 for success, error for failure.
978  */
979 static int sdma_v5_2_ring_test_ring(struct amdgpu_ring *ring)
980 {
981 	struct amdgpu_device *adev = ring->adev;
982 	unsigned i;
983 	unsigned index;
984 	int r;
985 	u32 tmp;
986 	u64 gpu_addr;
987 	volatile uint32_t *cpu_ptr = NULL;
988 
989 	tmp = 0xCAFEDEAD;
990 
991 	if (ring->is_mes_queue) {
992 		uint32_t offset = 0;
993 		offset = amdgpu_mes_ctx_get_offs(ring,
994 					 AMDGPU_MES_CTX_PADDING_OFFS);
995 		gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
996 		cpu_ptr = amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
997 		*cpu_ptr = tmp;
998 	} else {
999 		r = amdgpu_device_wb_get(adev, &index);
1000 		if (r) {
1001 			dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
1002 			return r;
1003 		}
1004 
1005 		gpu_addr = adev->wb.gpu_addr + (index * 4);
1006 		adev->wb.wb[index] = cpu_to_le32(tmp);
1007 	}
1008 
1009 	r = amdgpu_ring_alloc(ring, 20);
1010 	if (r) {
1011 		DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
1012 		amdgpu_device_wb_free(adev, index);
1013 		return r;
1014 	}
1015 
1016 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
1017 			  SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
1018 	amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
1019 	amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
1020 	amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0));
1021 	amdgpu_ring_write(ring, 0xDEADBEEF);
1022 	amdgpu_ring_commit(ring);
1023 
1024 	for (i = 0; i < adev->usec_timeout; i++) {
1025 		if (ring->is_mes_queue)
1026 			tmp = le32_to_cpu(*cpu_ptr);
1027 		else
1028 			tmp = le32_to_cpu(adev->wb.wb[index]);
1029 		if (tmp == 0xDEADBEEF)
1030 			break;
1031 		if (amdgpu_emu_mode == 1)
1032 			msleep(1);
1033 		else
1034 			udelay(1);
1035 	}
1036 
1037 	if (i >= adev->usec_timeout)
1038 		r = -ETIMEDOUT;
1039 
1040 	if (!ring->is_mes_queue)
1041 		amdgpu_device_wb_free(adev, index);
1042 
1043 	return r;
1044 }
1045 
1046 /**
1047  * sdma_v5_2_ring_test_ib - test an IB on the DMA engine
1048  *
1049  * @ring: amdgpu_ring structure holding ring information
1050  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
1051  *
1052  * Test a simple IB in the DMA ring.
1053  * Returns 0 on success, error on failure.
1054  */
1055 static int sdma_v5_2_ring_test_ib(struct amdgpu_ring *ring, long timeout)
1056 {
1057 	struct amdgpu_device *adev = ring->adev;
1058 	struct amdgpu_ib ib;
1059 	struct dma_fence *f = NULL;
1060 	unsigned index;
1061 	long r;
1062 	u32 tmp = 0;
1063 	u64 gpu_addr;
1064 	volatile uint32_t *cpu_ptr = NULL;
1065 
1066 	tmp = 0xCAFEDEAD;
1067 	memset(&ib, 0, sizeof(ib));
1068 
1069 	if (ring->is_mes_queue) {
1070 		uint32_t offset = 0;
1071 		offset = amdgpu_mes_ctx_get_offs(ring, AMDGPU_MES_CTX_IB_OFFS);
1072 		ib.gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
1073 		ib.ptr = (void *)amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
1074 
1075 		offset = amdgpu_mes_ctx_get_offs(ring,
1076 					 AMDGPU_MES_CTX_PADDING_OFFS);
1077 		gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
1078 		cpu_ptr = amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
1079 		*cpu_ptr = tmp;
1080 	} else {
1081 		r = amdgpu_device_wb_get(adev, &index);
1082 		if (r) {
1083 			dev_err(adev->dev, "(%ld) failed to allocate wb slot\n", r);
1084 			return r;
1085 		}
1086 
1087 		gpu_addr = adev->wb.gpu_addr + (index * 4);
1088 		adev->wb.wb[index] = cpu_to_le32(tmp);
1089 
1090 		r = amdgpu_ib_get(adev, NULL, 256, AMDGPU_IB_POOL_DIRECT, &ib);
1091 		if (r) {
1092 			DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r);
1093 			goto err0;
1094 		}
1095 	}
1096 
1097 	ib.ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
1098 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1099 	ib.ptr[1] = lower_32_bits(gpu_addr);
1100 	ib.ptr[2] = upper_32_bits(gpu_addr);
1101 	ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0);
1102 	ib.ptr[4] = 0xDEADBEEF;
1103 	ib.ptr[5] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1104 	ib.ptr[6] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1105 	ib.ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1106 	ib.length_dw = 8;
1107 
1108 	r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
1109 	if (r)
1110 		goto err1;
1111 
1112 	r = dma_fence_wait_timeout(f, false, timeout);
1113 	if (r == 0) {
1114 		DRM_ERROR("amdgpu: IB test timed out\n");
1115 		r = -ETIMEDOUT;
1116 		goto err1;
1117 	} else if (r < 0) {
1118 		DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
1119 		goto err1;
1120 	}
1121 
1122 	if (ring->is_mes_queue)
1123 		tmp = le32_to_cpu(*cpu_ptr);
1124 	else
1125 		tmp = le32_to_cpu(adev->wb.wb[index]);
1126 
1127 	if (tmp == 0xDEADBEEF)
1128 		r = 0;
1129 	else
1130 		r = -EINVAL;
1131 
1132 err1:
1133 	amdgpu_ib_free(adev, &ib, NULL);
1134 	dma_fence_put(f);
1135 err0:
1136 	if (!ring->is_mes_queue)
1137 		amdgpu_device_wb_free(adev, index);
1138 	return r;
1139 }
1140 
1141 
1142 /**
1143  * sdma_v5_2_vm_copy_pte - update PTEs by copying them from the GART
1144  *
1145  * @ib: indirect buffer to fill with commands
1146  * @pe: addr of the page entry
1147  * @src: src addr to copy from
1148  * @count: number of page entries to update
1149  *
1150  * Update PTEs by copying them from the GART using sDMA.
1151  */
1152 static void sdma_v5_2_vm_copy_pte(struct amdgpu_ib *ib,
1153 				  uint64_t pe, uint64_t src,
1154 				  unsigned count)
1155 {
1156 	unsigned bytes = count * 8;
1157 
1158 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1159 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
1160 	ib->ptr[ib->length_dw++] = bytes - 1;
1161 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1162 	ib->ptr[ib->length_dw++] = lower_32_bits(src);
1163 	ib->ptr[ib->length_dw++] = upper_32_bits(src);
1164 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1165 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1166 
1167 }
1168 
1169 /**
1170  * sdma_v5_2_vm_write_pte - update PTEs by writing them manually
1171  *
1172  * @ib: indirect buffer to fill with commands
1173  * @pe: addr of the page entry
1174  * @value: dst addr to write into pe
1175  * @count: number of page entries to update
1176  * @incr: increase next addr by incr bytes
1177  *
1178  * Update PTEs by writing them manually using sDMA.
1179  */
1180 static void sdma_v5_2_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
1181 				   uint64_t value, unsigned count,
1182 				   uint32_t incr)
1183 {
1184 	unsigned ndw = count * 2;
1185 
1186 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
1187 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1188 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1189 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1190 	ib->ptr[ib->length_dw++] = ndw - 1;
1191 	for (; ndw > 0; ndw -= 2) {
1192 		ib->ptr[ib->length_dw++] = lower_32_bits(value);
1193 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
1194 		value += incr;
1195 	}
1196 }
1197 
1198 /**
1199  * sdma_v5_2_vm_set_pte_pde - update the page tables using sDMA
1200  *
1201  * @ib: indirect buffer to fill with commands
1202  * @pe: addr of the page entry
1203  * @addr: dst addr to write into pe
1204  * @count: number of page entries to update
1205  * @incr: increase next addr by incr bytes
1206  * @flags: access flags
1207  *
1208  * Update the page tables using sDMA.
1209  */
1210 static void sdma_v5_2_vm_set_pte_pde(struct amdgpu_ib *ib,
1211 				     uint64_t pe,
1212 				     uint64_t addr, unsigned count,
1213 				     uint32_t incr, uint64_t flags)
1214 {
1215 	/* for physically contiguous pages (vram) */
1216 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_PTEPDE);
1217 	ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
1218 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1219 	ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
1220 	ib->ptr[ib->length_dw++] = upper_32_bits(flags);
1221 	ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
1222 	ib->ptr[ib->length_dw++] = upper_32_bits(addr);
1223 	ib->ptr[ib->length_dw++] = incr; /* increment size */
1224 	ib->ptr[ib->length_dw++] = 0;
1225 	ib->ptr[ib->length_dw++] = count - 1; /* number of entries */
1226 }
1227 
1228 /**
1229  * sdma_v5_2_ring_pad_ib - pad the IB
1230  *
1231  * @ib: indirect buffer to fill with padding
1232  * @ring: amdgpu_ring structure holding ring information
1233  *
1234  * Pad the IB with NOPs to a boundary multiple of 8.
1235  */
1236 static void sdma_v5_2_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
1237 {
1238 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
1239 	u32 pad_count;
1240 	int i;
1241 
1242 	pad_count = (-ib->length_dw) & 0x7;
1243 	for (i = 0; i < pad_count; i++)
1244 		if (sdma && sdma->burst_nop && (i == 0))
1245 			ib->ptr[ib->length_dw++] =
1246 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP) |
1247 				SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
1248 		else
1249 			ib->ptr[ib->length_dw++] =
1250 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
1251 }
1252 
1253 
1254 /**
1255  * sdma_v5_2_ring_emit_pipeline_sync - sync the pipeline
1256  *
1257  * @ring: amdgpu_ring pointer
1258  *
1259  * Make sure all previous operations are completed (CIK).
1260  */
1261 static void sdma_v5_2_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
1262 {
1263 	uint32_t seq = ring->fence_drv.sync_seq;
1264 	uint64_t addr = ring->fence_drv.gpu_addr;
1265 
1266 	/* wait for idle */
1267 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1268 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1269 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
1270 			  SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1));
1271 	amdgpu_ring_write(ring, addr & 0xfffffffc);
1272 	amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
1273 	amdgpu_ring_write(ring, seq); /* reference */
1274 	amdgpu_ring_write(ring, 0xffffffff); /* mask */
1275 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1276 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */
1277 }
1278 
1279 
1280 /**
1281  * sdma_v5_2_ring_emit_vm_flush - vm flush using sDMA
1282  *
1283  * @ring: amdgpu_ring pointer
1284  * @vmid: vmid number to use
1285  * @pd_addr: address
1286  *
1287  * Update the page table base and flush the VM TLB
1288  * using sDMA.
1289  */
1290 static void sdma_v5_2_ring_emit_vm_flush(struct amdgpu_ring *ring,
1291 					 unsigned vmid, uint64_t pd_addr)
1292 {
1293 	amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
1294 }
1295 
1296 static void sdma_v5_2_ring_emit_wreg(struct amdgpu_ring *ring,
1297 				     uint32_t reg, uint32_t val)
1298 {
1299 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
1300 			  SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
1301 	amdgpu_ring_write(ring, reg);
1302 	amdgpu_ring_write(ring, val);
1303 }
1304 
1305 static void sdma_v5_2_ring_emit_reg_wait(struct amdgpu_ring *ring, uint32_t reg,
1306 					 uint32_t val, uint32_t mask)
1307 {
1308 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1309 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1310 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* equal */
1311 	amdgpu_ring_write(ring, reg << 2);
1312 	amdgpu_ring_write(ring, 0);
1313 	amdgpu_ring_write(ring, val); /* reference */
1314 	amdgpu_ring_write(ring, mask); /* mask */
1315 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1316 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10));
1317 }
1318 
1319 static void sdma_v5_2_ring_emit_reg_write_reg_wait(struct amdgpu_ring *ring,
1320 						   uint32_t reg0, uint32_t reg1,
1321 						   uint32_t ref, uint32_t mask)
1322 {
1323 	amdgpu_ring_emit_wreg(ring, reg0, ref);
1324 	/* wait for a cycle to reset vm_inv_eng*_ack */
1325 	amdgpu_ring_emit_reg_wait(ring, reg0, 0, 0);
1326 	amdgpu_ring_emit_reg_wait(ring, reg1, mask, mask);
1327 }
1328 
1329 static int sdma_v5_2_early_init(void *handle)
1330 {
1331 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1332 
1333 	sdma_v5_2_set_ring_funcs(adev);
1334 	sdma_v5_2_set_buffer_funcs(adev);
1335 	sdma_v5_2_set_vm_pte_funcs(adev);
1336 	sdma_v5_2_set_irq_funcs(adev);
1337 	sdma_v5_2_set_mqd_funcs(adev);
1338 
1339 	return 0;
1340 }
1341 
1342 static unsigned sdma_v5_2_seq_to_irq_id(int seq_num)
1343 {
1344 	switch (seq_num) {
1345 	case 0:
1346 		return SOC15_IH_CLIENTID_SDMA0;
1347 	case 1:
1348 		return SOC15_IH_CLIENTID_SDMA1;
1349 	case 2:
1350 		return SOC15_IH_CLIENTID_SDMA2;
1351 	case 3:
1352 		return SOC15_IH_CLIENTID_SDMA3_Sienna_Cichlid;
1353 	default:
1354 		break;
1355 	}
1356 	return -EINVAL;
1357 }
1358 
1359 static unsigned sdma_v5_2_seq_to_trap_id(int seq_num)
1360 {
1361 	switch (seq_num) {
1362 	case 0:
1363 		return SDMA0_5_0__SRCID__SDMA_TRAP;
1364 	case 1:
1365 		return SDMA1_5_0__SRCID__SDMA_TRAP;
1366 	case 2:
1367 		return SDMA2_5_0__SRCID__SDMA_TRAP;
1368 	case 3:
1369 		return SDMA3_5_0__SRCID__SDMA_TRAP;
1370 	default:
1371 		break;
1372 	}
1373 	return -EINVAL;
1374 }
1375 
1376 static int sdma_v5_2_sw_init(void *handle)
1377 {
1378 	struct amdgpu_ring *ring;
1379 	int r, i;
1380 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1381 
1382 	/* SDMA trap event */
1383 	for (i = 0; i < adev->sdma.num_instances; i++) {
1384 		r = amdgpu_irq_add_id(adev, sdma_v5_2_seq_to_irq_id(i),
1385 				      sdma_v5_2_seq_to_trap_id(i),
1386 				      &adev->sdma.trap_irq);
1387 		if (r)
1388 			return r;
1389 	}
1390 
1391 	r = sdma_v5_2_init_microcode(adev);
1392 	if (r) {
1393 		DRM_ERROR("Failed to load sdma firmware!\n");
1394 		return r;
1395 	}
1396 
1397 	for (i = 0; i < adev->sdma.num_instances; i++) {
1398 		ring = &adev->sdma.instance[i].ring;
1399 		ring->ring_obj = NULL;
1400 		ring->use_doorbell = true;
1401 		ring->me = i;
1402 
1403 		DRM_INFO("use_doorbell being set to: [%s]\n",
1404 				ring->use_doorbell?"true":"false");
1405 
1406 		ring->doorbell_index =
1407 			(adev->doorbell_index.sdma_engine[i] << 1); //get DWORD offset
1408 
1409 		sprintf(ring->name, "sdma%d", i);
1410 		r = amdgpu_ring_init(adev, ring, 1024, &adev->sdma.trap_irq,
1411 				     AMDGPU_SDMA_IRQ_INSTANCE0 + i,
1412 				     AMDGPU_RING_PRIO_DEFAULT, NULL);
1413 		if (r)
1414 			return r;
1415 	}
1416 
1417 	return r;
1418 }
1419 
1420 static int sdma_v5_2_sw_fini(void *handle)
1421 {
1422 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1423 	int i;
1424 
1425 	for (i = 0; i < adev->sdma.num_instances; i++)
1426 		amdgpu_ring_fini(&adev->sdma.instance[i].ring);
1427 
1428 	sdma_v5_2_destroy_inst_ctx(adev);
1429 
1430 	return 0;
1431 }
1432 
1433 static int sdma_v5_2_hw_init(void *handle)
1434 {
1435 	int r;
1436 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1437 
1438 	r = sdma_v5_2_start(adev);
1439 
1440 	return r;
1441 }
1442 
1443 static int sdma_v5_2_hw_fini(void *handle)
1444 {
1445 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1446 
1447 	if (amdgpu_sriov_vf(adev))
1448 		return 0;
1449 
1450 	sdma_v5_2_ctx_switch_disable_all(adev);
1451 	sdma_v5_2_halt(adev);
1452 
1453 	return 0;
1454 }
1455 
1456 static int sdma_v5_2_suspend(void *handle)
1457 {
1458 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1459 
1460 	return sdma_v5_2_hw_fini(adev);
1461 }
1462 
1463 static int sdma_v5_2_resume(void *handle)
1464 {
1465 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1466 
1467 	return sdma_v5_2_hw_init(adev);
1468 }
1469 
1470 static bool sdma_v5_2_is_idle(void *handle)
1471 {
1472 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1473 	u32 i;
1474 
1475 	for (i = 0; i < adev->sdma.num_instances; i++) {
1476 		u32 tmp = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_STATUS_REG));
1477 
1478 		if (!(tmp & SDMA0_STATUS_REG__IDLE_MASK))
1479 			return false;
1480 	}
1481 
1482 	return true;
1483 }
1484 
1485 static int sdma_v5_2_wait_for_idle(void *handle)
1486 {
1487 	unsigned i;
1488 	u32 sdma0, sdma1, sdma2, sdma3;
1489 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1490 
1491 	for (i = 0; i < adev->usec_timeout; i++) {
1492 		sdma0 = RREG32(sdma_v5_2_get_reg_offset(adev, 0, mmSDMA0_STATUS_REG));
1493 		sdma1 = RREG32(sdma_v5_2_get_reg_offset(adev, 1, mmSDMA0_STATUS_REG));
1494 		sdma2 = RREG32(sdma_v5_2_get_reg_offset(adev, 2, mmSDMA0_STATUS_REG));
1495 		sdma3 = RREG32(sdma_v5_2_get_reg_offset(adev, 3, mmSDMA0_STATUS_REG));
1496 
1497 		if (sdma0 & sdma1 & sdma2 & sdma3 & SDMA0_STATUS_REG__IDLE_MASK)
1498 			return 0;
1499 		udelay(1);
1500 	}
1501 	return -ETIMEDOUT;
1502 }
1503 
1504 static int sdma_v5_2_ring_preempt_ib(struct amdgpu_ring *ring)
1505 {
1506 	int i, r = 0;
1507 	struct amdgpu_device *adev = ring->adev;
1508 	u32 index = 0;
1509 	u64 sdma_gfx_preempt;
1510 
1511 	amdgpu_sdma_get_index_from_ring(ring, &index);
1512 	sdma_gfx_preempt =
1513 		sdma_v5_2_get_reg_offset(adev, index, mmSDMA0_GFX_PREEMPT);
1514 
1515 	/* assert preemption condition */
1516 	amdgpu_ring_set_preempt_cond_exec(ring, false);
1517 
1518 	/* emit the trailing fence */
1519 	ring->trail_seq += 1;
1520 	amdgpu_ring_alloc(ring, 10);
1521 	sdma_v5_2_ring_emit_fence(ring, ring->trail_fence_gpu_addr,
1522 				  ring->trail_seq, 0);
1523 	amdgpu_ring_commit(ring);
1524 
1525 	/* assert IB preemption */
1526 	WREG32(sdma_gfx_preempt, 1);
1527 
1528 	/* poll the trailing fence */
1529 	for (i = 0; i < adev->usec_timeout; i++) {
1530 		if (ring->trail_seq ==
1531 		    le32_to_cpu(*(ring->trail_fence_cpu_addr)))
1532 			break;
1533 		udelay(1);
1534 	}
1535 
1536 	if (i >= adev->usec_timeout) {
1537 		r = -EINVAL;
1538 		DRM_ERROR("ring %d failed to be preempted\n", ring->idx);
1539 	}
1540 
1541 	/* deassert IB preemption */
1542 	WREG32(sdma_gfx_preempt, 0);
1543 
1544 	/* deassert the preemption condition */
1545 	amdgpu_ring_set_preempt_cond_exec(ring, true);
1546 	return r;
1547 }
1548 
1549 static int sdma_v5_2_set_trap_irq_state(struct amdgpu_device *adev,
1550 					struct amdgpu_irq_src *source,
1551 					unsigned type,
1552 					enum amdgpu_interrupt_state state)
1553 {
1554 	u32 sdma_cntl;
1555 	u32 reg_offset = sdma_v5_2_get_reg_offset(adev, type, mmSDMA0_CNTL);
1556 
1557 	if (!amdgpu_sriov_vf(adev)) {
1558 		sdma_cntl = RREG32(reg_offset);
1559 		sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE,
1560 			       state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0);
1561 		WREG32(reg_offset, sdma_cntl);
1562 	}
1563 
1564 	return 0;
1565 }
1566 
1567 static int sdma_v5_2_process_trap_irq(struct amdgpu_device *adev,
1568 				      struct amdgpu_irq_src *source,
1569 				      struct amdgpu_iv_entry *entry)
1570 {
1571 	uint32_t mes_queue_id = entry->src_data[0];
1572 
1573 	DRM_DEBUG("IH: SDMA trap\n");
1574 
1575 	if (adev->enable_mes && (mes_queue_id & AMDGPU_FENCE_MES_QUEUE_FLAG)) {
1576 		struct amdgpu_mes_queue *queue;
1577 
1578 		mes_queue_id &= AMDGPU_FENCE_MES_QUEUE_ID_MASK;
1579 
1580 		spin_lock(&adev->mes.queue_id_lock);
1581 		queue = idr_find(&adev->mes.queue_id_idr, mes_queue_id);
1582 		if (queue) {
1583 			DRM_DEBUG("process smda queue id = %d\n", mes_queue_id);
1584 			amdgpu_fence_process(queue->ring);
1585 		}
1586 		spin_unlock(&adev->mes.queue_id_lock);
1587 		return 0;
1588 	}
1589 
1590 	switch (entry->client_id) {
1591 	case SOC15_IH_CLIENTID_SDMA0:
1592 		switch (entry->ring_id) {
1593 		case 0:
1594 			amdgpu_fence_process(&adev->sdma.instance[0].ring);
1595 			break;
1596 		case 1:
1597 			/* XXX compute */
1598 			break;
1599 		case 2:
1600 			/* XXX compute */
1601 			break;
1602 		case 3:
1603 			/* XXX page queue*/
1604 			break;
1605 		}
1606 		break;
1607 	case SOC15_IH_CLIENTID_SDMA1:
1608 		switch (entry->ring_id) {
1609 		case 0:
1610 			amdgpu_fence_process(&adev->sdma.instance[1].ring);
1611 			break;
1612 		case 1:
1613 			/* XXX compute */
1614 			break;
1615 		case 2:
1616 			/* XXX compute */
1617 			break;
1618 		case 3:
1619 			/* XXX page queue*/
1620 			break;
1621 		}
1622 		break;
1623 	case SOC15_IH_CLIENTID_SDMA2:
1624 		switch (entry->ring_id) {
1625 		case 0:
1626 			amdgpu_fence_process(&adev->sdma.instance[2].ring);
1627 			break;
1628 		case 1:
1629 			/* XXX compute */
1630 			break;
1631 		case 2:
1632 			/* XXX compute */
1633 			break;
1634 		case 3:
1635 			/* XXX page queue*/
1636 			break;
1637 		}
1638 		break;
1639 	case SOC15_IH_CLIENTID_SDMA3_Sienna_Cichlid:
1640 		switch (entry->ring_id) {
1641 		case 0:
1642 			amdgpu_fence_process(&adev->sdma.instance[3].ring);
1643 			break;
1644 		case 1:
1645 			/* XXX compute */
1646 			break;
1647 		case 2:
1648 			/* XXX compute */
1649 			break;
1650 		case 3:
1651 			/* XXX page queue*/
1652 			break;
1653 		}
1654 		break;
1655 	}
1656 	return 0;
1657 }
1658 
1659 static int sdma_v5_2_process_illegal_inst_irq(struct amdgpu_device *adev,
1660 					      struct amdgpu_irq_src *source,
1661 					      struct amdgpu_iv_entry *entry)
1662 {
1663 	return 0;
1664 }
1665 
1666 static void sdma_v5_2_update_medium_grain_clock_gating(struct amdgpu_device *adev,
1667 						       bool enable)
1668 {
1669 	uint32_t data, def;
1670 	int i;
1671 
1672 	for (i = 0; i < adev->sdma.num_instances; i++) {
1673 
1674 		if (adev->sdma.instance[i].fw_version < 70 && adev->ip_versions[SDMA0_HWIP][0] == IP_VERSION(5, 2, 1))
1675 			adev->cg_flags &= ~AMD_CG_SUPPORT_SDMA_MGCG;
1676 
1677 		if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
1678 			/* Enable sdma clock gating */
1679 			def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL));
1680 			data &= ~(SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1681 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1682 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1683 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1684 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK |
1685 				  SDMA0_CLK_CTRL__SOFT_OVERRIDER_REG_MASK);
1686 			if (def != data)
1687 				WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data);
1688 		} else {
1689 			/* Disable sdma clock gating */
1690 			def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL));
1691 			data |= (SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1692 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1693 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1694 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1695 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK |
1696 				 SDMA0_CLK_CTRL__SOFT_OVERRIDER_REG_MASK);
1697 			if (def != data)
1698 				WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data);
1699 		}
1700 	}
1701 }
1702 
1703 static void sdma_v5_2_update_medium_grain_light_sleep(struct amdgpu_device *adev,
1704 						      bool enable)
1705 {
1706 	uint32_t data, def;
1707 	int i;
1708 
1709 	for (i = 0; i < adev->sdma.num_instances; i++) {
1710 
1711 		if (adev->sdma.instance[i].fw_version < 70 && adev->ip_versions[SDMA0_HWIP][0] == IP_VERSION(5, 2, 1))
1712 			adev->cg_flags &= ~AMD_CG_SUPPORT_SDMA_LS;
1713 
1714 		if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) {
1715 			/* Enable sdma mem light sleep */
1716 			def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL));
1717 			data |= SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1718 			if (def != data)
1719 				WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data);
1720 
1721 		} else {
1722 			/* Disable sdma mem light sleep */
1723 			def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL));
1724 			data &= ~SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1725 			if (def != data)
1726 				WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data);
1727 
1728 		}
1729 	}
1730 }
1731 
1732 static int sdma_v5_2_set_clockgating_state(void *handle,
1733 					   enum amd_clockgating_state state)
1734 {
1735 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1736 
1737 	if (amdgpu_sriov_vf(adev))
1738 		return 0;
1739 
1740 	switch (adev->ip_versions[SDMA0_HWIP][0]) {
1741 	case IP_VERSION(5, 2, 0):
1742 	case IP_VERSION(5, 2, 2):
1743 	case IP_VERSION(5, 2, 1):
1744 	case IP_VERSION(5, 2, 4):
1745 	case IP_VERSION(5, 2, 5):
1746 	case IP_VERSION(5, 2, 6):
1747 	case IP_VERSION(5, 2, 3):
1748 		sdma_v5_2_update_medium_grain_clock_gating(adev,
1749 				state == AMD_CG_STATE_GATE);
1750 		sdma_v5_2_update_medium_grain_light_sleep(adev,
1751 				state == AMD_CG_STATE_GATE);
1752 		break;
1753 	default:
1754 		break;
1755 	}
1756 
1757 	return 0;
1758 }
1759 
1760 static int sdma_v5_2_set_powergating_state(void *handle,
1761 					  enum amd_powergating_state state)
1762 {
1763 	return 0;
1764 }
1765 
1766 static void sdma_v5_2_get_clockgating_state(void *handle, u64 *flags)
1767 {
1768 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1769 	int data;
1770 
1771 	if (amdgpu_sriov_vf(adev))
1772 		*flags = 0;
1773 
1774 	/* AMD_CG_SUPPORT_SDMA_MGCG */
1775 	data = RREG32(sdma_v5_2_get_reg_offset(adev, 0, mmSDMA0_CLK_CTRL));
1776 	if (!(data & SDMA0_CLK_CTRL__CGCG_EN_OVERRIDE_MASK))
1777 		*flags |= AMD_CG_SUPPORT_SDMA_MGCG;
1778 
1779 	/* AMD_CG_SUPPORT_SDMA_LS */
1780 	data = RREG32_KIQ(sdma_v5_2_get_reg_offset(adev, 0, mmSDMA0_POWER_CNTL));
1781 	if (data & SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK)
1782 		*flags |= AMD_CG_SUPPORT_SDMA_LS;
1783 }
1784 
1785 const struct amd_ip_funcs sdma_v5_2_ip_funcs = {
1786 	.name = "sdma_v5_2",
1787 	.early_init = sdma_v5_2_early_init,
1788 	.late_init = NULL,
1789 	.sw_init = sdma_v5_2_sw_init,
1790 	.sw_fini = sdma_v5_2_sw_fini,
1791 	.hw_init = sdma_v5_2_hw_init,
1792 	.hw_fini = sdma_v5_2_hw_fini,
1793 	.suspend = sdma_v5_2_suspend,
1794 	.resume = sdma_v5_2_resume,
1795 	.is_idle = sdma_v5_2_is_idle,
1796 	.wait_for_idle = sdma_v5_2_wait_for_idle,
1797 	.soft_reset = sdma_v5_2_soft_reset,
1798 	.set_clockgating_state = sdma_v5_2_set_clockgating_state,
1799 	.set_powergating_state = sdma_v5_2_set_powergating_state,
1800 	.get_clockgating_state = sdma_v5_2_get_clockgating_state,
1801 };
1802 
1803 static const struct amdgpu_ring_funcs sdma_v5_2_ring_funcs = {
1804 	.type = AMDGPU_RING_TYPE_SDMA,
1805 	.align_mask = 0xf,
1806 	.nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP),
1807 	.support_64bit_ptrs = true,
1808 	.secure_submission_supported = true,
1809 	.vmhub = AMDGPU_GFXHUB_0,
1810 	.get_rptr = sdma_v5_2_ring_get_rptr,
1811 	.get_wptr = sdma_v5_2_ring_get_wptr,
1812 	.set_wptr = sdma_v5_2_ring_set_wptr,
1813 	.emit_frame_size =
1814 		5 + /* sdma_v5_2_ring_init_cond_exec */
1815 		6 + /* sdma_v5_2_ring_emit_hdp_flush */
1816 		3 + /* hdp_invalidate */
1817 		6 + /* sdma_v5_2_ring_emit_pipeline_sync */
1818 		/* sdma_v5_2_ring_emit_vm_flush */
1819 		SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 +
1820 		SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6 +
1821 		10 + 10 + 10, /* sdma_v5_2_ring_emit_fence x3 for user fence, vm fence */
1822 	.emit_ib_size = 7 + 6, /* sdma_v5_2_ring_emit_ib */
1823 	.emit_ib = sdma_v5_2_ring_emit_ib,
1824 	.emit_mem_sync = sdma_v5_2_ring_emit_mem_sync,
1825 	.emit_fence = sdma_v5_2_ring_emit_fence,
1826 	.emit_pipeline_sync = sdma_v5_2_ring_emit_pipeline_sync,
1827 	.emit_vm_flush = sdma_v5_2_ring_emit_vm_flush,
1828 	.emit_hdp_flush = sdma_v5_2_ring_emit_hdp_flush,
1829 	.test_ring = sdma_v5_2_ring_test_ring,
1830 	.test_ib = sdma_v5_2_ring_test_ib,
1831 	.insert_nop = sdma_v5_2_ring_insert_nop,
1832 	.pad_ib = sdma_v5_2_ring_pad_ib,
1833 	.emit_wreg = sdma_v5_2_ring_emit_wreg,
1834 	.emit_reg_wait = sdma_v5_2_ring_emit_reg_wait,
1835 	.emit_reg_write_reg_wait = sdma_v5_2_ring_emit_reg_write_reg_wait,
1836 	.init_cond_exec = sdma_v5_2_ring_init_cond_exec,
1837 	.patch_cond_exec = sdma_v5_2_ring_patch_cond_exec,
1838 	.preempt_ib = sdma_v5_2_ring_preempt_ib,
1839 };
1840 
1841 static void sdma_v5_2_set_ring_funcs(struct amdgpu_device *adev)
1842 {
1843 	int i;
1844 
1845 	for (i = 0; i < adev->sdma.num_instances; i++) {
1846 		adev->sdma.instance[i].ring.funcs = &sdma_v5_2_ring_funcs;
1847 		adev->sdma.instance[i].ring.me = i;
1848 	}
1849 }
1850 
1851 static const struct amdgpu_irq_src_funcs sdma_v5_2_trap_irq_funcs = {
1852 	.set = sdma_v5_2_set_trap_irq_state,
1853 	.process = sdma_v5_2_process_trap_irq,
1854 };
1855 
1856 static const struct amdgpu_irq_src_funcs sdma_v5_2_illegal_inst_irq_funcs = {
1857 	.process = sdma_v5_2_process_illegal_inst_irq,
1858 };
1859 
1860 static void sdma_v5_2_set_irq_funcs(struct amdgpu_device *adev)
1861 {
1862 	adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_INSTANCE0 +
1863 					adev->sdma.num_instances;
1864 	adev->sdma.trap_irq.funcs = &sdma_v5_2_trap_irq_funcs;
1865 	adev->sdma.illegal_inst_irq.funcs = &sdma_v5_2_illegal_inst_irq_funcs;
1866 }
1867 
1868 /**
1869  * sdma_v5_2_emit_copy_buffer - copy buffer using the sDMA engine
1870  *
1871  * @ib: indirect buffer to copy to
1872  * @src_offset: src GPU address
1873  * @dst_offset: dst GPU address
1874  * @byte_count: number of bytes to xfer
1875  * @tmz: if a secure copy should be used
1876  *
1877  * Copy GPU buffers using the DMA engine.
1878  * Used by the amdgpu ttm implementation to move pages if
1879  * registered as the asic copy callback.
1880  */
1881 static void sdma_v5_2_emit_copy_buffer(struct amdgpu_ib *ib,
1882 				       uint64_t src_offset,
1883 				       uint64_t dst_offset,
1884 				       uint32_t byte_count,
1885 				       bool tmz)
1886 {
1887 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1888 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR) |
1889 		SDMA_PKT_COPY_LINEAR_HEADER_TMZ(tmz ? 1 : 0);
1890 	ib->ptr[ib->length_dw++] = byte_count - 1;
1891 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1892 	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1893 	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1894 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1895 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1896 }
1897 
1898 /**
1899  * sdma_v5_2_emit_fill_buffer - fill buffer using the sDMA engine
1900  *
1901  * @ib: indirect buffer to fill
1902  * @src_data: value to write to buffer
1903  * @dst_offset: dst GPU address
1904  * @byte_count: number of bytes to xfer
1905  *
1906  * Fill GPU buffers using the DMA engine.
1907  */
1908 static void sdma_v5_2_emit_fill_buffer(struct amdgpu_ib *ib,
1909 				       uint32_t src_data,
1910 				       uint64_t dst_offset,
1911 				       uint32_t byte_count)
1912 {
1913 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL);
1914 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1915 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1916 	ib->ptr[ib->length_dw++] = src_data;
1917 	ib->ptr[ib->length_dw++] = byte_count - 1;
1918 }
1919 
1920 static const struct amdgpu_buffer_funcs sdma_v5_2_buffer_funcs = {
1921 	.copy_max_bytes = 0x400000,
1922 	.copy_num_dw = 7,
1923 	.emit_copy_buffer = sdma_v5_2_emit_copy_buffer,
1924 
1925 	.fill_max_bytes = 0x400000,
1926 	.fill_num_dw = 5,
1927 	.emit_fill_buffer = sdma_v5_2_emit_fill_buffer,
1928 };
1929 
1930 static void sdma_v5_2_set_buffer_funcs(struct amdgpu_device *adev)
1931 {
1932 	if (adev->mman.buffer_funcs == NULL) {
1933 		adev->mman.buffer_funcs = &sdma_v5_2_buffer_funcs;
1934 		adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1935 	}
1936 }
1937 
1938 static const struct amdgpu_vm_pte_funcs sdma_v5_2_vm_pte_funcs = {
1939 	.copy_pte_num_dw = 7,
1940 	.copy_pte = sdma_v5_2_vm_copy_pte,
1941 	.write_pte = sdma_v5_2_vm_write_pte,
1942 	.set_pte_pde = sdma_v5_2_vm_set_pte_pde,
1943 };
1944 
1945 static void sdma_v5_2_set_vm_pte_funcs(struct amdgpu_device *adev)
1946 {
1947 	unsigned i;
1948 
1949 	if (adev->vm_manager.vm_pte_funcs == NULL) {
1950 		adev->vm_manager.vm_pte_funcs = &sdma_v5_2_vm_pte_funcs;
1951 		for (i = 0; i < adev->sdma.num_instances; i++) {
1952 			adev->vm_manager.vm_pte_scheds[i] =
1953 				&adev->sdma.instance[i].ring.sched;
1954 		}
1955 		adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
1956 	}
1957 }
1958 
1959 const struct amdgpu_ip_block_version sdma_v5_2_ip_block = {
1960 	.type = AMD_IP_BLOCK_TYPE_SDMA,
1961 	.major = 5,
1962 	.minor = 2,
1963 	.rev = 0,
1964 	.funcs = &sdma_v5_2_ip_funcs,
1965 };
1966