1 /*
2  * Copyright 2020 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_11_0_0_offset.h"
34 #include "gc/gc_11_0_0_sh_mask.h"
35 #include "gc/gc_11_0_0_default.h"
36 #include "hdp/hdp_6_0_0_offset.h"
37 #include "ivsrcid/gfx/irqsrcs_gfx_11_0_0.h"
38 
39 #include "soc15_common.h"
40 #include "soc15.h"
41 #include "sdma_v6_0_0_pkt_open.h"
42 #include "nbio_v4_3.h"
43 #include "sdma_common.h"
44 #include "sdma_v6_0.h"
45 #include "v11_structs.h"
46 
47 MODULE_FIRMWARE("amdgpu/sdma_6_0_0.bin");
48 MODULE_FIRMWARE("amdgpu/sdma_6_0_1.bin");
49 MODULE_FIRMWARE("amdgpu/sdma_6_0_2.bin");
50 
51 #define SDMA1_REG_OFFSET 0x600
52 #define SDMA0_HYP_DEC_REG_START 0x5880
53 #define SDMA0_HYP_DEC_REG_END 0x589a
54 #define SDMA1_HYP_DEC_REG_OFFSET 0x20
55 
56 static void sdma_v6_0_set_ring_funcs(struct amdgpu_device *adev);
57 static void sdma_v6_0_set_buffer_funcs(struct amdgpu_device *adev);
58 static void sdma_v6_0_set_vm_pte_funcs(struct amdgpu_device *adev);
59 static void sdma_v6_0_set_irq_funcs(struct amdgpu_device *adev);
60 
61 static u32 sdma_v6_0_get_reg_offset(struct amdgpu_device *adev, u32 instance, u32 internal_offset)
62 {
63 	u32 base;
64 
65 	if (internal_offset >= SDMA0_HYP_DEC_REG_START &&
66 	    internal_offset <= SDMA0_HYP_DEC_REG_END) {
67 		base = adev->reg_offset[GC_HWIP][0][1];
68 		if (instance != 0)
69 			internal_offset += SDMA1_HYP_DEC_REG_OFFSET * instance;
70 	} else {
71 		base = adev->reg_offset[GC_HWIP][0][0];
72 		if (instance == 1)
73 			internal_offset += SDMA1_REG_OFFSET;
74 	}
75 
76 	return base + internal_offset;
77 }
78 
79 static int sdma_v6_0_init_inst_ctx(struct amdgpu_sdma_instance *sdma_inst)
80 {
81 	int err = 0;
82 	const struct sdma_firmware_header_v2_0 *hdr;
83 
84 	err = amdgpu_ucode_validate(sdma_inst->fw);
85 	if (err)
86 		return err;
87 
88 	hdr = (const struct sdma_firmware_header_v2_0 *)sdma_inst->fw->data;
89 	sdma_inst->fw_version = le32_to_cpu(hdr->header.ucode_version);
90 	sdma_inst->feature_version = le32_to_cpu(hdr->ucode_feature_version);
91 
92 	if (sdma_inst->feature_version >= 20)
93 		sdma_inst->burst_nop = true;
94 
95 	return 0;
96 }
97 
98 static void sdma_v6_0_destroy_inst_ctx(struct amdgpu_device *adev)
99 {
100 	release_firmware(adev->sdma.instance[0].fw);
101 
102 	memset((void*)adev->sdma.instance, 0,
103 	       sizeof(struct amdgpu_sdma_instance) * AMDGPU_MAX_SDMA_INSTANCES);
104 }
105 
106 /**
107  * sdma_v6_0_init_microcode - load ucode images from disk
108  *
109  * @adev: amdgpu_device pointer
110  *
111  * Use the firmware interface to load the ucode images into
112  * the driver (not loaded into hw).
113  * Returns 0 on success, error on failure.
114  */
115 
116 // emulation only, won't work on real chip
117 // sdma 6.0.0 real chip need to use PSP to load firmware
118 static int sdma_v6_0_init_microcode(struct amdgpu_device *adev)
119 {
120 	char fw_name[30];
121 	char ucode_prefix[30];
122 	int err = 0, i;
123 	struct amdgpu_firmware_info *info = NULL;
124 	const struct sdma_firmware_header_v2_0 *sdma_hdr;
125 
126 	DRM_DEBUG("\n");
127 
128 	amdgpu_ucode_ip_version_decode(adev, SDMA0_HWIP, ucode_prefix, sizeof(ucode_prefix));
129 
130 	snprintf(fw_name, sizeof(fw_name), "amdgpu/%s.bin", ucode_prefix);
131 
132 	err = request_firmware(&adev->sdma.instance[0].fw, fw_name, adev->dev);
133 	if (err)
134 		goto out;
135 
136 	err = sdma_v6_0_init_inst_ctx(&adev->sdma.instance[0]);
137 	if (err)
138 		goto out;
139 
140 	for (i = 1; i < adev->sdma.num_instances; i++) {
141 		memcpy((void*)&adev->sdma.instance[i],
142 		       (void*)&adev->sdma.instance[0],
143 		       sizeof(struct amdgpu_sdma_instance));
144 	}
145 
146 	DRM_DEBUG("psp_load == '%s'\n",
147 		  adev->firmware.load_type == AMDGPU_FW_LOAD_PSP ? "true" : "false");
148 
149 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
150 		sdma_hdr = (const struct sdma_firmware_header_v2_0 *)adev->sdma.instance[0].fw->data;
151 		info = &adev->firmware.ucode[AMDGPU_UCODE_ID_SDMA_UCODE_TH0];
152 		info->ucode_id = AMDGPU_UCODE_ID_SDMA_UCODE_TH0;
153 		info->fw = adev->sdma.instance[0].fw;
154 		adev->firmware.fw_size +=
155 			ALIGN(le32_to_cpu(sdma_hdr->ctx_ucode_size_bytes), PAGE_SIZE);
156 		info = &adev->firmware.ucode[AMDGPU_UCODE_ID_SDMA_UCODE_TH1];
157 		info->ucode_id = AMDGPU_UCODE_ID_SDMA_UCODE_TH1;
158 		info->fw = adev->sdma.instance[0].fw;
159 		adev->firmware.fw_size +=
160 			ALIGN(le32_to_cpu(sdma_hdr->ctl_ucode_size_bytes), PAGE_SIZE);
161 	}
162 
163 out:
164 	if (err) {
165 		DRM_ERROR("sdma_v6_0: Failed to load firmware \"%s\"\n", fw_name);
166 		sdma_v6_0_destroy_inst_ctx(adev);
167 	}
168 	return err;
169 }
170 
171 static unsigned sdma_v6_0_ring_init_cond_exec(struct amdgpu_ring *ring)
172 {
173 	unsigned ret;
174 
175 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_COND_EXE));
176 	amdgpu_ring_write(ring, lower_32_bits(ring->cond_exe_gpu_addr));
177 	amdgpu_ring_write(ring, upper_32_bits(ring->cond_exe_gpu_addr));
178 	amdgpu_ring_write(ring, 1);
179 	ret = ring->wptr & ring->buf_mask;/* this is the offset we need patch later */
180 	amdgpu_ring_write(ring, 0x55aa55aa);/* insert dummy here and patch it later */
181 
182 	return ret;
183 }
184 
185 static void sdma_v6_0_ring_patch_cond_exec(struct amdgpu_ring *ring,
186 					   unsigned offset)
187 {
188 	unsigned cur;
189 
190 	BUG_ON(offset > ring->buf_mask);
191 	BUG_ON(ring->ring[offset] != 0x55aa55aa);
192 
193 	cur = (ring->wptr - 1) & ring->buf_mask;
194 	if (cur > offset)
195 		ring->ring[offset] = cur - offset;
196 	else
197 		ring->ring[offset] = (ring->buf_mask + 1) - offset + cur;
198 }
199 
200 /**
201  * sdma_v6_0_ring_get_rptr - get the current read pointer
202  *
203  * @ring: amdgpu ring pointer
204  *
205  * Get the current rptr from the hardware.
206  */
207 static uint64_t sdma_v6_0_ring_get_rptr(struct amdgpu_ring *ring)
208 {
209 	u64 *rptr;
210 
211 	/* XXX check if swapping is necessary on BE */
212 	rptr = (u64 *)ring->rptr_cpu_addr;
213 
214 	DRM_DEBUG("rptr before shift == 0x%016llx\n", *rptr);
215 	return ((*rptr) >> 2);
216 }
217 
218 /**
219  * sdma_v6_0_ring_get_wptr - get the current write pointer
220  *
221  * @ring: amdgpu ring pointer
222  *
223  * Get the current wptr from the hardware.
224  */
225 static uint64_t sdma_v6_0_ring_get_wptr(struct amdgpu_ring *ring)
226 {
227 	u64 wptr = 0;
228 
229 	if (ring->use_doorbell) {
230 		/* XXX check if swapping is necessary on BE */
231 		wptr = READ_ONCE(*((u64 *)ring->wptr_cpu_addr));
232 		DRM_DEBUG("wptr/doorbell before shift == 0x%016llx\n", wptr);
233 	}
234 
235 	return wptr >> 2;
236 }
237 
238 /**
239  * sdma_v6_0_ring_set_wptr - commit the write pointer
240  *
241  * @ring: amdgpu ring pointer
242  *
243  * Write the wptr back to the hardware.
244  */
245 static void sdma_v6_0_ring_set_wptr(struct amdgpu_ring *ring)
246 {
247 	struct amdgpu_device *adev = ring->adev;
248 
249 	DRM_DEBUG("Setting write pointer\n");
250 	if (ring->use_doorbell) {
251 		DRM_DEBUG("Using doorbell -- "
252 				"wptr_offs == 0x%08x "
253 				"lower_32_bits(ring->wptr) << 2 == 0x%08x "
254 				"upper_32_bits(ring->wptr) << 2 == 0x%08x\n",
255 				ring->wptr_offs,
256 				lower_32_bits(ring->wptr << 2),
257 				upper_32_bits(ring->wptr << 2));
258 		/* XXX check if swapping is necessary on BE */
259 		atomic64_set((atomic64_t *)ring->wptr_cpu_addr,
260 			     ring->wptr << 2);
261 		DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016llx)\n",
262 				ring->doorbell_index, ring->wptr << 2);
263 		WDOORBELL64(ring->doorbell_index, ring->wptr << 2);
264 	} else {
265 		DRM_DEBUG("Not using doorbell -- "
266 				"regSDMA%i_GFX_RB_WPTR == 0x%08x "
267 				"regSDMA%i_GFX_RB_WPTR_HI == 0x%08x\n",
268 				ring->me,
269 				lower_32_bits(ring->wptr << 2),
270 				ring->me,
271 				upper_32_bits(ring->wptr << 2));
272 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, ring->me, regSDMA0_QUEUE0_RB_WPTR),
273 			lower_32_bits(ring->wptr << 2));
274 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, ring->me, regSDMA0_QUEUE0_RB_WPTR_HI),
275 			upper_32_bits(ring->wptr << 2));
276 	}
277 }
278 
279 static void sdma_v6_0_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
280 {
281 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
282 	int i;
283 
284 	for (i = 0; i < count; i++)
285 		if (sdma && sdma->burst_nop && (i == 0))
286 			amdgpu_ring_write(ring, ring->funcs->nop |
287 				SDMA_PKT_NOP_HEADER_COUNT(count - 1));
288 		else
289 			amdgpu_ring_write(ring, ring->funcs->nop);
290 }
291 
292 /**
293  * sdma_v6_0_ring_emit_ib - Schedule an IB on the DMA engine
294  *
295  * @ring: amdgpu ring pointer
296  * @ib: IB object to schedule
297  *
298  * Schedule an IB in the DMA ring.
299  */
300 static void sdma_v6_0_ring_emit_ib(struct amdgpu_ring *ring,
301 				   struct amdgpu_job *job,
302 				   struct amdgpu_ib *ib,
303 				   uint32_t flags)
304 {
305 	unsigned vmid = AMDGPU_JOB_GET_VMID(job);
306 	uint64_t csa_mc_addr = amdgpu_sdma_get_csa_mc_addr(ring, vmid);
307 
308 	/* An IB packet must end on a 8 DW boundary--the next dword
309 	 * must be on a 8-dword boundary. Our IB packet below is 6
310 	 * dwords long, thus add x number of NOPs, such that, in
311 	 * modular arithmetic,
312 	 * wptr + 6 + x = 8k, k >= 0, which in C is,
313 	 * (wptr + 6 + x) % 8 = 0.
314 	 * The expression below, is a solution of x.
315 	 */
316 	sdma_v6_0_ring_insert_nop(ring, (2 - lower_32_bits(ring->wptr)) & 7);
317 
318 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_INDIRECT) |
319 			  SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf));
320 	/* base must be 32 byte aligned */
321 	amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
322 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
323 	amdgpu_ring_write(ring, ib->length_dw);
324 	amdgpu_ring_write(ring, lower_32_bits(csa_mc_addr));
325 	amdgpu_ring_write(ring, upper_32_bits(csa_mc_addr));
326 }
327 
328 /**
329  * sdma_v6_0_ring_emit_mem_sync - flush the IB by graphics cache rinse
330  *
331  * @ring: amdgpu ring pointer
332  * @job: job to retrieve vmid from
333  * @ib: IB object to schedule
334  *
335  * flush the IB by graphics cache rinse.
336  */
337 static void sdma_v6_0_ring_emit_mem_sync(struct amdgpu_ring *ring)
338 {
339         uint32_t gcr_cntl = SDMA_GCR_GL2_INV | SDMA_GCR_GL2_WB | SDMA_GCR_GLM_INV |
340                             SDMA_GCR_GL1_INV | SDMA_GCR_GLV_INV | SDMA_GCR_GLK_INV |
341                             SDMA_GCR_GLI_INV(1);
342 
343         /* flush entire cache L0/L1/L2, this can be optimized by performance requirement */
344         amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_GCR_REQ));
345         amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD1_BASE_VA_31_7(0));
346         amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD2_GCR_CONTROL_15_0(gcr_cntl) |
347                           SDMA_PKT_GCR_REQ_PAYLOAD2_BASE_VA_47_32(0));
348         amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD3_LIMIT_VA_31_7(0) |
349                           SDMA_PKT_GCR_REQ_PAYLOAD3_GCR_CONTROL_18_16(gcr_cntl >> 16));
350         amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD4_LIMIT_VA_47_32(0) |
351                           SDMA_PKT_GCR_REQ_PAYLOAD4_VMID(0));
352 }
353 
354 
355 /**
356  * sdma_v6_0_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
357  *
358  * @ring: amdgpu ring pointer
359  *
360  * Emit an hdp flush packet on the requested DMA ring.
361  */
362 static void sdma_v6_0_ring_emit_hdp_flush(struct amdgpu_ring *ring)
363 {
364 	struct amdgpu_device *adev = ring->adev;
365 	u32 ref_and_mask = 0;
366 	const struct nbio_hdp_flush_reg *nbio_hf_reg = adev->nbio.hdp_flush_reg;
367 
368 	ref_and_mask = nbio_hf_reg->ref_and_mask_sdma0 << ring->me;
369 
370 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_POLL_REGMEM) |
371 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) |
372 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
373 	amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_done_offset(adev)) << 2);
374 	amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_req_offset(adev)) << 2);
375 	amdgpu_ring_write(ring, ref_and_mask); /* reference */
376 	amdgpu_ring_write(ring, ref_and_mask); /* mask */
377 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
378 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
379 }
380 
381 /**
382  * sdma_v6_0_ring_emit_fence - emit a fence on the DMA ring
383  *
384  * @ring: amdgpu ring pointer
385  * @fence: amdgpu fence object
386  *
387  * Add a DMA fence packet to the ring to write
388  * the fence seq number and DMA trap packet to generate
389  * an interrupt if needed.
390  */
391 static void sdma_v6_0_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
392 				      unsigned flags)
393 {
394 	bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
395 	/* write the fence */
396 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_FENCE) |
397 			  SDMA_PKT_FENCE_HEADER_MTYPE(0x3)); /* Ucached(UC) */
398 	/* zero in first two bits */
399 	BUG_ON(addr & 0x3);
400 	amdgpu_ring_write(ring, lower_32_bits(addr));
401 	amdgpu_ring_write(ring, upper_32_bits(addr));
402 	amdgpu_ring_write(ring, lower_32_bits(seq));
403 
404 	/* optionally write high bits as well */
405 	if (write64bit) {
406 		addr += 4;
407 		amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_FENCE) |
408 				  SDMA_PKT_FENCE_HEADER_MTYPE(0x3));
409 		/* zero in first two bits */
410 		BUG_ON(addr & 0x3);
411 		amdgpu_ring_write(ring, lower_32_bits(addr));
412 		amdgpu_ring_write(ring, upper_32_bits(addr));
413 		amdgpu_ring_write(ring, upper_32_bits(seq));
414 	}
415 
416 	if (flags & AMDGPU_FENCE_FLAG_INT) {
417 		uint32_t ctx = ring->is_mes_queue ?
418 			(ring->hw_queue_id | AMDGPU_FENCE_MES_QUEUE_FLAG) : 0;
419 		/* generate an interrupt */
420 		amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_TRAP));
421 		amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(ctx));
422 	}
423 }
424 
425 /**
426  * sdma_v6_0_gfx_stop - stop the gfx async dma engines
427  *
428  * @adev: amdgpu_device pointer
429  *
430  * Stop the gfx async dma ring buffers.
431  */
432 static void sdma_v6_0_gfx_stop(struct amdgpu_device *adev)
433 {
434 	struct amdgpu_ring *sdma0 = &adev->sdma.instance[0].ring;
435 	struct amdgpu_ring *sdma1 = &adev->sdma.instance[1].ring;
436 	u32 rb_cntl, ib_cntl;
437 	int i;
438 
439 	if ((adev->mman.buffer_funcs_ring == sdma0) ||
440 	    (adev->mman.buffer_funcs_ring == sdma1))
441 		amdgpu_ttm_set_buffer_funcs_status(adev, false);
442 
443 	for (i = 0; i < adev->sdma.num_instances; i++) {
444 		rb_cntl = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_CNTL));
445 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RB_ENABLE, 0);
446 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_CNTL), rb_cntl);
447 		ib_cntl = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_IB_CNTL));
448 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_QUEUE0_IB_CNTL, IB_ENABLE, 0);
449 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_IB_CNTL), ib_cntl);
450 	}
451 
452 	sdma0->sched.ready = false;
453 	sdma1->sched.ready = false;
454 }
455 
456 /**
457  * sdma_v6_0_rlc_stop - stop the compute async dma engines
458  *
459  * @adev: amdgpu_device pointer
460  *
461  * Stop the compute async dma queues.
462  */
463 static void sdma_v6_0_rlc_stop(struct amdgpu_device *adev)
464 {
465 	/* XXX todo */
466 }
467 
468 /**
469  * sdma_v6_0_ctx_switch_enable - stop the async dma engines context switch
470  *
471  * @adev: amdgpu_device pointer
472  * @enable: enable/disable the DMA MEs context switch.
473  *
474  * Halt or unhalt the async dma engines context switch.
475  */
476 static void sdma_v6_0_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
477 {
478 }
479 
480 /**
481  * sdma_v6_0_enable - stop the async dma engines
482  *
483  * @adev: amdgpu_device pointer
484  * @enable: enable/disable the DMA MEs.
485  *
486  * Halt or unhalt the async dma engines.
487  */
488 static void sdma_v6_0_enable(struct amdgpu_device *adev, bool enable)
489 {
490 	u32 f32_cntl;
491 	int i;
492 
493 	if (!enable) {
494 		sdma_v6_0_gfx_stop(adev);
495 		sdma_v6_0_rlc_stop(adev);
496 	}
497 
498 	for (i = 0; i < adev->sdma.num_instances; i++) {
499 		f32_cntl = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_F32_CNTL));
500 		f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, enable ? 0 : 1);
501 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_F32_CNTL), f32_cntl);
502 	}
503 }
504 
505 /**
506  * sdma_v6_0_gfx_resume - setup and start the async dma engines
507  *
508  * @adev: amdgpu_device pointer
509  *
510  * Set up the gfx DMA ring buffers and enable them.
511  * Returns 0 for success, error for failure.
512  */
513 static int sdma_v6_0_gfx_resume(struct amdgpu_device *adev)
514 {
515 	struct amdgpu_ring *ring;
516 	u32 rb_cntl, ib_cntl;
517 	u32 rb_bufsz;
518 	u32 doorbell;
519 	u32 doorbell_offset;
520 	u32 temp;
521 	u64 wptr_gpu_addr;
522 	int i, r;
523 
524 	for (i = 0; i < adev->sdma.num_instances; i++) {
525 		ring = &adev->sdma.instance[i].ring;
526 
527 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_SEM_WAIT_FAIL_TIMER_CNTL), 0);
528 
529 		/* Set ring buffer size in dwords */
530 		rb_bufsz = order_base_2(ring->ring_size / 4);
531 		rb_cntl = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_CNTL));
532 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RB_SIZE, rb_bufsz);
533 #ifdef __BIG_ENDIAN
534 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RB_SWAP_ENABLE, 1);
535 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL,
536 					RPTR_WRITEBACK_SWAP_ENABLE, 1);
537 #endif
538 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RB_PRIV, 1);
539 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_CNTL), rb_cntl);
540 
541 		/* Initialize the ring buffer's read and write pointers */
542 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_RPTR), 0);
543 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_RPTR_HI), 0);
544 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR), 0);
545 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR_HI), 0);
546 
547 		/* setup the wptr shadow polling */
548 		wptr_gpu_addr = ring->wptr_gpu_addr;
549 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR_POLL_ADDR_LO),
550 		       lower_32_bits(wptr_gpu_addr));
551 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR_POLL_ADDR_HI),
552 		       upper_32_bits(wptr_gpu_addr));
553 
554 		/* set the wb address whether it's enabled or not */
555 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_RPTR_ADDR_HI),
556 		       upper_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFF);
557 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_RPTR_ADDR_LO),
558 		       lower_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFC);
559 
560 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
561 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, WPTR_POLL_ENABLE, 0);
562 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, F32_WPTR_POLL_ENABLE, 1);
563 
564 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_BASE), ring->gpu_addr >> 8);
565 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_BASE_HI), ring->gpu_addr >> 40);
566 
567 		ring->wptr = 0;
568 
569 		/* before programing wptr to a less value, need set minor_ptr_update first */
570 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_MINOR_PTR_UPDATE), 1);
571 
572 		if (!amdgpu_sriov_vf(adev)) { /* only bare-metal use register write for wptr */
573 			WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR), lower_32_bits(ring->wptr) << 2);
574 			WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR_HI), upper_32_bits(ring->wptr) << 2);
575 		}
576 
577 		doorbell = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_DOORBELL));
578 		doorbell_offset = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_DOORBELL_OFFSET));
579 
580 		if (ring->use_doorbell) {
581 			doorbell = REG_SET_FIELD(doorbell, SDMA0_QUEUE0_DOORBELL, ENABLE, 1);
582 			doorbell_offset = REG_SET_FIELD(doorbell_offset, SDMA0_QUEUE0_DOORBELL_OFFSET,
583 					OFFSET, ring->doorbell_index);
584 		} else {
585 			doorbell = REG_SET_FIELD(doorbell, SDMA0_QUEUE0_DOORBELL, ENABLE, 0);
586 		}
587 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_DOORBELL), doorbell);
588 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_DOORBELL_OFFSET), doorbell_offset);
589 
590 		if (i == 0)
591 			adev->nbio.funcs->sdma_doorbell_range(adev, i, ring->use_doorbell,
592 						      ring->doorbell_index,
593 						      adev->doorbell_index.sdma_doorbell_range * adev->sdma.num_instances);
594 
595 		if (amdgpu_sriov_vf(adev))
596 			sdma_v6_0_ring_set_wptr(ring);
597 
598 		/* set minor_ptr_update to 0 after wptr programed */
599 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_MINOR_PTR_UPDATE), 0);
600 
601 		/* Set up RESP_MODE to non-copy addresses */
602 		temp = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UTCL1_CNTL));
603 		temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, RESP_MODE, 3);
604 		temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, REDO_DELAY, 9);
605 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UTCL1_CNTL), temp);
606 
607 		/* program default cache read and write policy */
608 		temp = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UTCL1_PAGE));
609 		/* clean read policy and write policy bits */
610 		temp &= 0xFF0FFF;
611 		temp |= ((CACHE_READ_POLICY_L2__DEFAULT << 12) |
612 			 (CACHE_WRITE_POLICY_L2__DEFAULT << 14) |
613 			 SDMA0_UTCL1_PAGE__LLC_NOALLOC_MASK);
614 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UTCL1_PAGE), temp);
615 
616 		if (!amdgpu_sriov_vf(adev)) {
617 			/* unhalt engine */
618 			temp = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_F32_CNTL));
619 			temp = REG_SET_FIELD(temp, SDMA0_F32_CNTL, HALT, 0);
620 			temp = REG_SET_FIELD(temp, SDMA0_F32_CNTL, TH1_RESET, 0);
621 			WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_F32_CNTL), temp);
622 		}
623 
624 		/* enable DMA RB */
625 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RB_ENABLE, 1);
626 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_CNTL), rb_cntl);
627 
628 		ib_cntl = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_IB_CNTL));
629 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_QUEUE0_IB_CNTL, IB_ENABLE, 1);
630 #ifdef __BIG_ENDIAN
631 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_QUEUE0_IB_CNTL, IB_SWAP_ENABLE, 1);
632 #endif
633 		/* enable DMA IBs */
634 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_IB_CNTL), ib_cntl);
635 
636 		ring->sched.ready = true;
637 
638 		if (amdgpu_sriov_vf(adev)) { /* bare-metal sequence doesn't need below to lines */
639 			sdma_v6_0_ctx_switch_enable(adev, true);
640 			sdma_v6_0_enable(adev, true);
641 		}
642 
643 		r = amdgpu_ring_test_helper(ring);
644 		if (r) {
645 			ring->sched.ready = false;
646 			return r;
647 		}
648 
649 		if (adev->mman.buffer_funcs_ring == ring)
650 			amdgpu_ttm_set_buffer_funcs_status(adev, true);
651 	}
652 
653 	return 0;
654 }
655 
656 /**
657  * sdma_v6_0_rlc_resume - setup and start the async dma engines
658  *
659  * @adev: amdgpu_device pointer
660  *
661  * Set up the compute DMA queues and enable them.
662  * Returns 0 for success, error for failure.
663  */
664 static int sdma_v6_0_rlc_resume(struct amdgpu_device *adev)
665 {
666 	return 0;
667 }
668 
669 /**
670  * sdma_v6_0_load_microcode - load the sDMA ME ucode
671  *
672  * @adev: amdgpu_device pointer
673  *
674  * Loads the sDMA0/1 ucode.
675  * Returns 0 for success, -EINVAL if the ucode is not available.
676  */
677 static int sdma_v6_0_load_microcode(struct amdgpu_device *adev)
678 {
679 	const struct sdma_firmware_header_v2_0 *hdr;
680 	const __le32 *fw_data;
681 	u32 fw_size;
682 	int i, j;
683 	bool use_broadcast;
684 
685 	/* halt the MEs */
686 	sdma_v6_0_enable(adev, false);
687 
688 	if (!adev->sdma.instance[0].fw)
689 		return -EINVAL;
690 
691 	/* use broadcast mode to load SDMA microcode by default */
692 	use_broadcast = true;
693 
694 	if (use_broadcast) {
695 		dev_info(adev->dev, "Use broadcast method to load SDMA firmware\n");
696 		/* load Control Thread microcode */
697 		hdr = (const struct sdma_firmware_header_v2_0 *)adev->sdma.instance[0].fw->data;
698 		amdgpu_ucode_print_sdma_hdr(&hdr->header);
699 		fw_size = le32_to_cpu(hdr->ctx_jt_offset + hdr->ctx_jt_size) / 4;
700 
701 		fw_data = (const __le32 *)
702 			(adev->sdma.instance[0].fw->data +
703 				le32_to_cpu(hdr->header.ucode_array_offset_bytes));
704 
705 		WREG32(sdma_v6_0_get_reg_offset(adev, 0, regSDMA0_BROADCAST_UCODE_ADDR), 0);
706 
707 		for (j = 0; j < fw_size; j++) {
708 			if (amdgpu_emu_mode == 1 && j % 500 == 0)
709 				msleep(1);
710 			WREG32(sdma_v6_0_get_reg_offset(adev, 0, regSDMA0_BROADCAST_UCODE_DATA), le32_to_cpup(fw_data++));
711 		}
712 
713 		/* load Context Switch microcode */
714 		fw_size = le32_to_cpu(hdr->ctl_jt_offset + hdr->ctl_jt_size) / 4;
715 
716 		fw_data = (const __le32 *)
717 			(adev->sdma.instance[0].fw->data +
718 				le32_to_cpu(hdr->ctl_ucode_offset));
719 
720 		WREG32(sdma_v6_0_get_reg_offset(adev, 0, regSDMA0_BROADCAST_UCODE_ADDR), 0x8000);
721 
722 		for (j = 0; j < fw_size; j++) {
723 			if (amdgpu_emu_mode == 1 && j % 500 == 0)
724 				msleep(1);
725 			WREG32(sdma_v6_0_get_reg_offset(adev, 0, regSDMA0_BROADCAST_UCODE_DATA), le32_to_cpup(fw_data++));
726 		}
727 	} else {
728 		dev_info(adev->dev, "Use legacy method to load SDMA firmware\n");
729 		for (i = 0; i < adev->sdma.num_instances; i++) {
730 			/* load Control Thread microcode */
731 			hdr = (const struct sdma_firmware_header_v2_0 *)adev->sdma.instance[0].fw->data;
732 			amdgpu_ucode_print_sdma_hdr(&hdr->header);
733 			fw_size = le32_to_cpu(hdr->ctx_jt_offset + hdr->ctx_jt_size) / 4;
734 
735 			fw_data = (const __le32 *)
736 				(adev->sdma.instance[0].fw->data +
737 					le32_to_cpu(hdr->header.ucode_array_offset_bytes));
738 
739 			WREG32(sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UCODE_ADDR), 0);
740 
741 			for (j = 0; j < fw_size; j++) {
742 				if (amdgpu_emu_mode == 1 && j % 500 == 0)
743 					msleep(1);
744 				WREG32(sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UCODE_DATA), le32_to_cpup(fw_data++));
745 			}
746 
747 			WREG32(sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UCODE_ADDR), adev->sdma.instance[0].fw_version);
748 
749 			/* load Context Switch microcode */
750 			fw_size = le32_to_cpu(hdr->ctl_jt_offset + hdr->ctl_jt_size) / 4;
751 
752 			fw_data = (const __le32 *)
753 				(adev->sdma.instance[0].fw->data +
754 					le32_to_cpu(hdr->ctl_ucode_offset));
755 
756 			WREG32(sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UCODE_ADDR), 0x8000);
757 
758 			for (j = 0; j < fw_size; j++) {
759 				if (amdgpu_emu_mode == 1 && j % 500 == 0)
760 					msleep(1);
761 				WREG32(sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UCODE_DATA), le32_to_cpup(fw_data++));
762 			}
763 
764 			WREG32(sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UCODE_ADDR), adev->sdma.instance[0].fw_version);
765 		}
766 	}
767 
768 	return 0;
769 }
770 
771 static int sdma_v6_0_soft_reset(void *handle)
772 {
773 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
774 	u32 grbm_soft_reset;
775 	u32 tmp;
776 	int i;
777 
778 	for (i = 0; i < adev->sdma.num_instances; i++) {
779 		grbm_soft_reset = REG_SET_FIELD(0,
780 						GRBM_SOFT_RESET, SOFT_RESET_SDMA0,
781 						1);
782 		grbm_soft_reset <<= i;
783 
784 		tmp = RREG32_SOC15(GC, 0, regGRBM_SOFT_RESET);
785 		tmp |= grbm_soft_reset;
786 		DRM_DEBUG("GRBM_SOFT_RESET=0x%08X\n", tmp);
787 		WREG32_SOC15(GC, 0, regGRBM_SOFT_RESET, tmp);
788 		tmp = RREG32_SOC15(GC, 0, regGRBM_SOFT_RESET);
789 
790 		udelay(50);
791 
792 		tmp &= ~grbm_soft_reset;
793 		WREG32_SOC15(GC, 0, regGRBM_SOFT_RESET, tmp);
794 		tmp = RREG32_SOC15(GC, 0, regGRBM_SOFT_RESET);
795 
796 		udelay(50);
797 	}
798 
799 	return 0;
800 }
801 
802 /**
803  * sdma_v6_0_start - setup and start the async dma engines
804  *
805  * @adev: amdgpu_device pointer
806  *
807  * Set up the DMA engines and enable them.
808  * Returns 0 for success, error for failure.
809  */
810 static int sdma_v6_0_start(struct amdgpu_device *adev)
811 {
812 	int r = 0;
813 
814 	if (amdgpu_sriov_vf(adev)) {
815 		sdma_v6_0_ctx_switch_enable(adev, false);
816 		sdma_v6_0_enable(adev, false);
817 
818 		/* set RB registers */
819 		r = sdma_v6_0_gfx_resume(adev);
820 		return r;
821 	}
822 
823 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
824 		r = sdma_v6_0_load_microcode(adev);
825 		if (r)
826 			return r;
827 
828 		/* The value of regSDMA_F32_CNTL is invalid the moment after loading fw */
829 		if (amdgpu_emu_mode == 1)
830 			msleep(1000);
831 	}
832 
833 	sdma_v6_0_soft_reset(adev);
834 	/* unhalt the MEs */
835 	sdma_v6_0_enable(adev, true);
836 	/* enable sdma ring preemption */
837 	sdma_v6_0_ctx_switch_enable(adev, true);
838 
839 	/* start the gfx rings and rlc compute queues */
840 	r = sdma_v6_0_gfx_resume(adev);
841 	if (r)
842 		return r;
843 	r = sdma_v6_0_rlc_resume(adev);
844 
845 	return r;
846 }
847 
848 static int sdma_v6_0_mqd_init(struct amdgpu_device *adev, void *mqd,
849 			      struct amdgpu_mqd_prop *prop)
850 {
851 	struct v11_sdma_mqd *m = mqd;
852 	uint64_t wb_gpu_addr;
853 
854 	m->sdmax_rlcx_rb_cntl =
855 		order_base_2(prop->queue_size / 4) << SDMA0_QUEUE0_RB_CNTL__RB_SIZE__SHIFT |
856 		1 << SDMA0_QUEUE0_RB_CNTL__RPTR_WRITEBACK_ENABLE__SHIFT |
857 		4 << SDMA0_QUEUE0_RB_CNTL__RPTR_WRITEBACK_TIMER__SHIFT;
858 
859 	m->sdmax_rlcx_rb_base = lower_32_bits(prop->hqd_base_gpu_addr >> 8);
860 	m->sdmax_rlcx_rb_base_hi = upper_32_bits(prop->hqd_base_gpu_addr >> 8);
861 
862 	wb_gpu_addr = prop->wptr_gpu_addr;
863 	m->sdmax_rlcx_rb_wptr_poll_addr_lo = lower_32_bits(wb_gpu_addr);
864 	m->sdmax_rlcx_rb_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr);
865 
866 	wb_gpu_addr = prop->rptr_gpu_addr;
867 	m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits(wb_gpu_addr);
868 	m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits(wb_gpu_addr);
869 
870 	m->sdmax_rlcx_ib_cntl = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, 0,
871 							regSDMA0_QUEUE0_IB_CNTL));
872 
873 	m->sdmax_rlcx_doorbell_offset =
874 		prop->doorbell_index << SDMA0_QUEUE0_DOORBELL_OFFSET__OFFSET__SHIFT;
875 
876 	m->sdmax_rlcx_doorbell = REG_SET_FIELD(0, SDMA0_QUEUE0_DOORBELL, ENABLE, 1);
877 
878 	m->sdmax_rlcx_skip_cntl = 0;
879 	m->sdmax_rlcx_context_status = 0;
880 	m->sdmax_rlcx_doorbell_log = 0;
881 
882 	m->sdmax_rlcx_rb_aql_cntl = regSDMA0_QUEUE0_RB_AQL_CNTL_DEFAULT;
883 	m->sdmax_rlcx_dummy_reg = regSDMA0_QUEUE0_DUMMY_REG_DEFAULT;
884 
885 	return 0;
886 }
887 
888 static void sdma_v6_0_set_mqd_funcs(struct amdgpu_device *adev)
889 {
890 	adev->mqds[AMDGPU_HW_IP_DMA].mqd_size = sizeof(struct v11_sdma_mqd);
891 	adev->mqds[AMDGPU_HW_IP_DMA].init_mqd = sdma_v6_0_mqd_init;
892 }
893 
894 /**
895  * sdma_v6_0_ring_test_ring - simple async dma engine test
896  *
897  * @ring: amdgpu_ring structure holding ring information
898  *
899  * Test the DMA engine by writing using it to write an
900  * value to memory.
901  * Returns 0 for success, error for failure.
902  */
903 static int sdma_v6_0_ring_test_ring(struct amdgpu_ring *ring)
904 {
905 	struct amdgpu_device *adev = ring->adev;
906 	unsigned i;
907 	unsigned index;
908 	int r;
909 	u32 tmp;
910 	u64 gpu_addr;
911 	volatile uint32_t *cpu_ptr = NULL;
912 
913 	tmp = 0xCAFEDEAD;
914 
915 	if (ring->is_mes_queue) {
916 		uint32_t offset = 0;
917 		offset = amdgpu_mes_ctx_get_offs(ring,
918 					 AMDGPU_MES_CTX_PADDING_OFFS);
919 		gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
920 		cpu_ptr = amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
921 		*cpu_ptr = tmp;
922 	} else {
923 		r = amdgpu_device_wb_get(adev, &index);
924 		if (r) {
925 			dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
926 			return r;
927 		}
928 
929 		gpu_addr = adev->wb.gpu_addr + (index * 4);
930 		adev->wb.wb[index] = cpu_to_le32(tmp);
931 	}
932 
933 	r = amdgpu_ring_alloc(ring, 5);
934 	if (r) {
935 		DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
936 		amdgpu_device_wb_free(adev, index);
937 		return r;
938 	}
939 
940 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_WRITE) |
941 			  SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
942 	amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
943 	amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
944 	amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0));
945 	amdgpu_ring_write(ring, 0xDEADBEEF);
946 	amdgpu_ring_commit(ring);
947 
948 	for (i = 0; i < adev->usec_timeout; i++) {
949 		if (ring->is_mes_queue)
950 			tmp = le32_to_cpu(*cpu_ptr);
951 		else
952 			tmp = le32_to_cpu(adev->wb.wb[index]);
953 		if (tmp == 0xDEADBEEF)
954 			break;
955 		if (amdgpu_emu_mode == 1)
956 			msleep(1);
957 		else
958 			udelay(1);
959 	}
960 
961 	if (i >= adev->usec_timeout)
962 		r = -ETIMEDOUT;
963 
964 	if (!ring->is_mes_queue)
965 		amdgpu_device_wb_free(adev, index);
966 
967 	return r;
968 }
969 
970 /**
971  * sdma_v6_0_ring_test_ib - test an IB on the DMA engine
972  *
973  * @ring: amdgpu_ring structure holding ring information
974  *
975  * Test a simple IB in the DMA ring.
976  * Returns 0 on success, error on failure.
977  */
978 static int sdma_v6_0_ring_test_ib(struct amdgpu_ring *ring, long timeout)
979 {
980 	struct amdgpu_device *adev = ring->adev;
981 	struct amdgpu_ib ib;
982 	struct dma_fence *f = NULL;
983 	unsigned index;
984 	long r;
985 	u32 tmp = 0;
986 	u64 gpu_addr;
987 	volatile uint32_t *cpu_ptr = NULL;
988 
989 	tmp = 0xCAFEDEAD;
990 	memset(&ib, 0, sizeof(ib));
991 
992 	if (ring->is_mes_queue) {
993 		uint32_t offset = 0;
994 		offset = amdgpu_mes_ctx_get_offs(ring, AMDGPU_MES_CTX_IB_OFFS);
995 		ib.gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
996 		ib.ptr = (void *)amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
997 
998 		offset = amdgpu_mes_ctx_get_offs(ring,
999 					 AMDGPU_MES_CTX_PADDING_OFFS);
1000 		gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
1001 		cpu_ptr = amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
1002 		*cpu_ptr = tmp;
1003 	} else {
1004 		r = amdgpu_device_wb_get(adev, &index);
1005 		if (r) {
1006 			dev_err(adev->dev, "(%ld) failed to allocate wb slot\n", r);
1007 			return r;
1008 		}
1009 
1010 		gpu_addr = adev->wb.gpu_addr + (index * 4);
1011 		adev->wb.wb[index] = cpu_to_le32(tmp);
1012 
1013 		r = amdgpu_ib_get(adev, NULL, 256, AMDGPU_IB_POOL_DIRECT, &ib);
1014 		if (r) {
1015 			DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r);
1016 			goto err0;
1017 		}
1018 	}
1019 
1020 	ib.ptr[0] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_WRITE) |
1021 		SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1022 	ib.ptr[1] = lower_32_bits(gpu_addr);
1023 	ib.ptr[2] = upper_32_bits(gpu_addr);
1024 	ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0);
1025 	ib.ptr[4] = 0xDEADBEEF;
1026 	ib.ptr[5] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1027 	ib.ptr[6] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1028 	ib.ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1029 	ib.length_dw = 8;
1030 
1031 	r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
1032 	if (r)
1033 		goto err1;
1034 
1035 	r = dma_fence_wait_timeout(f, false, timeout);
1036 	if (r == 0) {
1037 		DRM_ERROR("amdgpu: IB test timed out\n");
1038 		r = -ETIMEDOUT;
1039 		goto err1;
1040 	} else if (r < 0) {
1041 		DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
1042 		goto err1;
1043 	}
1044 
1045 	if (ring->is_mes_queue)
1046 		tmp = le32_to_cpu(*cpu_ptr);
1047 	else
1048 		tmp = le32_to_cpu(adev->wb.wb[index]);
1049 
1050 	if (tmp == 0xDEADBEEF)
1051 		r = 0;
1052 	else
1053 		r = -EINVAL;
1054 
1055 err1:
1056 	amdgpu_ib_free(adev, &ib, NULL);
1057 	dma_fence_put(f);
1058 err0:
1059 	if (!ring->is_mes_queue)
1060 		amdgpu_device_wb_free(adev, index);
1061 	return r;
1062 }
1063 
1064 
1065 /**
1066  * sdma_v6_0_vm_copy_pte - update PTEs by copying them from the GART
1067  *
1068  * @ib: indirect buffer to fill with commands
1069  * @pe: addr of the page entry
1070  * @src: src addr to copy from
1071  * @count: number of page entries to update
1072  *
1073  * Update PTEs by copying them from the GART using sDMA.
1074  */
1075 static void sdma_v6_0_vm_copy_pte(struct amdgpu_ib *ib,
1076 				  uint64_t pe, uint64_t src,
1077 				  unsigned count)
1078 {
1079 	unsigned bytes = count * 8;
1080 
1081 	ib->ptr[ib->length_dw++] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_COPY) |
1082 		SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
1083 	ib->ptr[ib->length_dw++] = bytes - 1;
1084 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1085 	ib->ptr[ib->length_dw++] = lower_32_bits(src);
1086 	ib->ptr[ib->length_dw++] = upper_32_bits(src);
1087 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1088 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1089 
1090 }
1091 
1092 /**
1093  * sdma_v6_0_vm_write_pte - update PTEs by writing them manually
1094  *
1095  * @ib: indirect buffer to fill with commands
1096  * @pe: addr of the page entry
1097  * @addr: dst addr to write into pe
1098  * @count: number of page entries to update
1099  * @incr: increase next addr by incr bytes
1100  * @flags: access flags
1101  *
1102  * Update PTEs by writing them manually using sDMA.
1103  */
1104 static void sdma_v6_0_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
1105 				   uint64_t value, unsigned count,
1106 				   uint32_t incr)
1107 {
1108 	unsigned ndw = count * 2;
1109 
1110 	ib->ptr[ib->length_dw++] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_WRITE) |
1111 		SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1112 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1113 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1114 	ib->ptr[ib->length_dw++] = ndw - 1;
1115 	for (; ndw > 0; ndw -= 2) {
1116 		ib->ptr[ib->length_dw++] = lower_32_bits(value);
1117 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
1118 		value += incr;
1119 	}
1120 }
1121 
1122 /**
1123  * sdma_v6_0_vm_set_pte_pde - update the page tables using sDMA
1124  *
1125  * @ib: indirect buffer to fill with commands
1126  * @pe: addr of the page entry
1127  * @addr: dst addr to write into pe
1128  * @count: number of page entries to update
1129  * @incr: increase next addr by incr bytes
1130  * @flags: access flags
1131  *
1132  * Update the page tables using sDMA.
1133  */
1134 static void sdma_v6_0_vm_set_pte_pde(struct amdgpu_ib *ib,
1135 				     uint64_t pe,
1136 				     uint64_t addr, unsigned count,
1137 				     uint32_t incr, uint64_t flags)
1138 {
1139 	/* for physically contiguous pages (vram) */
1140 	ib->ptr[ib->length_dw++] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_PTEPDE);
1141 	ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
1142 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1143 	ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
1144 	ib->ptr[ib->length_dw++] = upper_32_bits(flags);
1145 	ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
1146 	ib->ptr[ib->length_dw++] = upper_32_bits(addr);
1147 	ib->ptr[ib->length_dw++] = incr; /* increment size */
1148 	ib->ptr[ib->length_dw++] = 0;
1149 	ib->ptr[ib->length_dw++] = count - 1; /* number of entries */
1150 }
1151 
1152 /**
1153  * sdma_v6_0_ring_pad_ib - pad the IB
1154  * @ib: indirect buffer to fill with padding
1155  *
1156  * Pad the IB with NOPs to a boundary multiple of 8.
1157  */
1158 static void sdma_v6_0_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
1159 {
1160 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
1161 	u32 pad_count;
1162 	int i;
1163 
1164 	pad_count = (-ib->length_dw) & 0x7;
1165 	for (i = 0; i < pad_count; i++)
1166 		if (sdma && sdma->burst_nop && (i == 0))
1167 			ib->ptr[ib->length_dw++] =
1168 				SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_NOP) |
1169 				SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
1170 		else
1171 			ib->ptr[ib->length_dw++] =
1172 				SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_NOP);
1173 }
1174 
1175 /**
1176  * sdma_v6_0_ring_emit_pipeline_sync - sync the pipeline
1177  *
1178  * @ring: amdgpu_ring pointer
1179  *
1180  * Make sure all previous operations are completed (CIK).
1181  */
1182 static void sdma_v6_0_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
1183 {
1184 	uint32_t seq = ring->fence_drv.sync_seq;
1185 	uint64_t addr = ring->fence_drv.gpu_addr;
1186 
1187 	/* wait for idle */
1188 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1189 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1190 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
1191 			  SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1));
1192 	amdgpu_ring_write(ring, addr & 0xfffffffc);
1193 	amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
1194 	amdgpu_ring_write(ring, seq); /* reference */
1195 	amdgpu_ring_write(ring, 0xffffffff); /* mask */
1196 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1197 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */
1198 }
1199 
1200 /**
1201  * sdma_v6_0_ring_emit_vm_flush - vm flush using sDMA
1202  *
1203  * @ring: amdgpu_ring pointer
1204  * @vm: amdgpu_vm pointer
1205  *
1206  * Update the page table base and flush the VM TLB
1207  * using sDMA.
1208  */
1209 static void sdma_v6_0_ring_emit_vm_flush(struct amdgpu_ring *ring,
1210 					 unsigned vmid, uint64_t pd_addr)
1211 {
1212 	amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
1213 }
1214 
1215 static void sdma_v6_0_ring_emit_wreg(struct amdgpu_ring *ring,
1216 				     uint32_t reg, uint32_t val)
1217 {
1218 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_SRBM_WRITE) |
1219 			  SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
1220 	amdgpu_ring_write(ring, reg);
1221 	amdgpu_ring_write(ring, val);
1222 }
1223 
1224 static void sdma_v6_0_ring_emit_reg_wait(struct amdgpu_ring *ring, uint32_t reg,
1225 					 uint32_t val, uint32_t mask)
1226 {
1227 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1228 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1229 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* equal */
1230 	amdgpu_ring_write(ring, reg << 2);
1231 	amdgpu_ring_write(ring, 0);
1232 	amdgpu_ring_write(ring, val); /* reference */
1233 	amdgpu_ring_write(ring, mask); /* mask */
1234 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1235 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10));
1236 }
1237 
1238 static void sdma_v6_0_ring_emit_reg_write_reg_wait(struct amdgpu_ring *ring,
1239 						   uint32_t reg0, uint32_t reg1,
1240 						   uint32_t ref, uint32_t mask)
1241 {
1242 	amdgpu_ring_emit_wreg(ring, reg0, ref);
1243 	/* wait for a cycle to reset vm_inv_eng*_ack */
1244 	amdgpu_ring_emit_reg_wait(ring, reg0, 0, 0);
1245 	amdgpu_ring_emit_reg_wait(ring, reg1, mask, mask);
1246 }
1247 
1248 static int sdma_v6_0_early_init(void *handle)
1249 {
1250 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1251 
1252 	sdma_v6_0_set_ring_funcs(adev);
1253 	sdma_v6_0_set_buffer_funcs(adev);
1254 	sdma_v6_0_set_vm_pte_funcs(adev);
1255 	sdma_v6_0_set_irq_funcs(adev);
1256 	sdma_v6_0_set_mqd_funcs(adev);
1257 
1258 	return 0;
1259 }
1260 
1261 static int sdma_v6_0_sw_init(void *handle)
1262 {
1263 	struct amdgpu_ring *ring;
1264 	int r, i;
1265 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1266 
1267 	/* SDMA trap event */
1268 	r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_GFX,
1269 			      GFX_11_0_0__SRCID__SDMA_TRAP,
1270 			      &adev->sdma.trap_irq);
1271 	if (r)
1272 		return r;
1273 
1274 	r = sdma_v6_0_init_microcode(adev);
1275 	if (r) {
1276 		DRM_ERROR("Failed to load sdma firmware!\n");
1277 		return r;
1278 	}
1279 
1280 	for (i = 0; i < adev->sdma.num_instances; i++) {
1281 		ring = &adev->sdma.instance[i].ring;
1282 		ring->ring_obj = NULL;
1283 		ring->use_doorbell = true;
1284 		ring->me = i;
1285 
1286 		DRM_DEBUG("SDMA %d use_doorbell being set to: [%s]\n", i,
1287 				ring->use_doorbell?"true":"false");
1288 
1289 		ring->doorbell_index =
1290 			(adev->doorbell_index.sdma_engine[i] << 1); // get DWORD offset
1291 
1292 		sprintf(ring->name, "sdma%d", i);
1293 		r = amdgpu_ring_init(adev, ring, 1024,
1294 				     &adev->sdma.trap_irq,
1295 				     AMDGPU_SDMA_IRQ_INSTANCE0 + i,
1296 				     AMDGPU_RING_PRIO_DEFAULT, NULL);
1297 		if (r)
1298 			return r;
1299 	}
1300 
1301 	return r;
1302 }
1303 
1304 static int sdma_v6_0_sw_fini(void *handle)
1305 {
1306 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1307 	int i;
1308 
1309 	for (i = 0; i < adev->sdma.num_instances; i++)
1310 		amdgpu_ring_fini(&adev->sdma.instance[i].ring);
1311 
1312 	sdma_v6_0_destroy_inst_ctx(adev);
1313 
1314 	return 0;
1315 }
1316 
1317 static int sdma_v6_0_hw_init(void *handle)
1318 {
1319 	int r;
1320 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1321 
1322 	r = sdma_v6_0_start(adev);
1323 
1324 	return r;
1325 }
1326 
1327 static int sdma_v6_0_hw_fini(void *handle)
1328 {
1329 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1330 
1331 	if (amdgpu_sriov_vf(adev))
1332 		return 0;
1333 
1334 	sdma_v6_0_ctx_switch_enable(adev, false);
1335 	sdma_v6_0_enable(adev, false);
1336 
1337 	return 0;
1338 }
1339 
1340 static int sdma_v6_0_suspend(void *handle)
1341 {
1342 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1343 
1344 	return sdma_v6_0_hw_fini(adev);
1345 }
1346 
1347 static int sdma_v6_0_resume(void *handle)
1348 {
1349 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1350 
1351 	return sdma_v6_0_hw_init(adev);
1352 }
1353 
1354 static bool sdma_v6_0_is_idle(void *handle)
1355 {
1356 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1357 	u32 i;
1358 
1359 	for (i = 0; i < adev->sdma.num_instances; i++) {
1360 		u32 tmp = RREG32(sdma_v6_0_get_reg_offset(adev, i, regSDMA0_STATUS_REG));
1361 
1362 		if (!(tmp & SDMA0_STATUS_REG__IDLE_MASK))
1363 			return false;
1364 	}
1365 
1366 	return true;
1367 }
1368 
1369 static int sdma_v6_0_wait_for_idle(void *handle)
1370 {
1371 	unsigned i;
1372 	u32 sdma0, sdma1;
1373 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1374 
1375 	for (i = 0; i < adev->usec_timeout; i++) {
1376 		sdma0 = RREG32(sdma_v6_0_get_reg_offset(adev, 0, regSDMA0_STATUS_REG));
1377 		sdma1 = RREG32(sdma_v6_0_get_reg_offset(adev, 1, regSDMA0_STATUS_REG));
1378 
1379 		if (sdma0 & sdma1 & SDMA0_STATUS_REG__IDLE_MASK)
1380 			return 0;
1381 		udelay(1);
1382 	}
1383 	return -ETIMEDOUT;
1384 }
1385 
1386 static int sdma_v6_0_ring_preempt_ib(struct amdgpu_ring *ring)
1387 {
1388 	int i, r = 0;
1389 	struct amdgpu_device *adev = ring->adev;
1390 	u32 index = 0;
1391 	u64 sdma_gfx_preempt;
1392 
1393 	amdgpu_sdma_get_index_from_ring(ring, &index);
1394 	sdma_gfx_preempt =
1395 		sdma_v6_0_get_reg_offset(adev, index, regSDMA0_QUEUE0_PREEMPT);
1396 
1397 	/* assert preemption condition */
1398 	amdgpu_ring_set_preempt_cond_exec(ring, false);
1399 
1400 	/* emit the trailing fence */
1401 	ring->trail_seq += 1;
1402 	amdgpu_ring_alloc(ring, 10);
1403 	sdma_v6_0_ring_emit_fence(ring, ring->trail_fence_gpu_addr,
1404 				  ring->trail_seq, 0);
1405 	amdgpu_ring_commit(ring);
1406 
1407 	/* assert IB preemption */
1408 	WREG32(sdma_gfx_preempt, 1);
1409 
1410 	/* poll the trailing fence */
1411 	for (i = 0; i < adev->usec_timeout; i++) {
1412 		if (ring->trail_seq ==
1413 		    le32_to_cpu(*(ring->trail_fence_cpu_addr)))
1414 			break;
1415 		udelay(1);
1416 	}
1417 
1418 	if (i >= adev->usec_timeout) {
1419 		r = -EINVAL;
1420 		DRM_ERROR("ring %d failed to be preempted\n", ring->idx);
1421 	}
1422 
1423 	/* deassert IB preemption */
1424 	WREG32(sdma_gfx_preempt, 0);
1425 
1426 	/* deassert the preemption condition */
1427 	amdgpu_ring_set_preempt_cond_exec(ring, true);
1428 	return r;
1429 }
1430 
1431 static int sdma_v6_0_set_trap_irq_state(struct amdgpu_device *adev,
1432 					struct amdgpu_irq_src *source,
1433 					unsigned type,
1434 					enum amdgpu_interrupt_state state)
1435 {
1436 	u32 sdma_cntl;
1437 
1438 	u32 reg_offset = sdma_v6_0_get_reg_offset(adev, type, regSDMA0_CNTL);
1439 
1440 	sdma_cntl = RREG32(reg_offset);
1441 	sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE,
1442 		       state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0);
1443 	WREG32(reg_offset, sdma_cntl);
1444 
1445 	return 0;
1446 }
1447 
1448 static int sdma_v6_0_process_trap_irq(struct amdgpu_device *adev,
1449 				      struct amdgpu_irq_src *source,
1450 				      struct amdgpu_iv_entry *entry)
1451 {
1452 	int instances, queue;
1453 	uint32_t mes_queue_id = entry->src_data[0];
1454 
1455 	DRM_DEBUG("IH: SDMA trap\n");
1456 
1457 	if (adev->enable_mes && (mes_queue_id & AMDGPU_FENCE_MES_QUEUE_FLAG)) {
1458 		struct amdgpu_mes_queue *queue;
1459 
1460 		mes_queue_id &= AMDGPU_FENCE_MES_QUEUE_ID_MASK;
1461 
1462 		spin_lock(&adev->mes.queue_id_lock);
1463 		queue = idr_find(&adev->mes.queue_id_idr, mes_queue_id);
1464 		if (queue) {
1465 			DRM_DEBUG("process smda queue id = %d\n", mes_queue_id);
1466 			amdgpu_fence_process(queue->ring);
1467 		}
1468 		spin_unlock(&adev->mes.queue_id_lock);
1469 		return 0;
1470 	}
1471 
1472 	queue = entry->ring_id & 0xf;
1473 	instances = (entry->ring_id & 0xf0) >> 4;
1474 	if (instances > 1) {
1475 		DRM_ERROR("IH: wrong ring_ID detected, as wrong sdma instance\n");
1476 		return -EINVAL;
1477 	}
1478 
1479 	switch (entry->client_id) {
1480 	case SOC21_IH_CLIENTID_GFX:
1481 		switch (queue) {
1482 		case 0:
1483 			amdgpu_fence_process(&adev->sdma.instance[instances].ring);
1484 			break;
1485 		default:
1486 			break;
1487 		}
1488 		break;
1489 	}
1490 	return 0;
1491 }
1492 
1493 static int sdma_v6_0_process_illegal_inst_irq(struct amdgpu_device *adev,
1494 					      struct amdgpu_irq_src *source,
1495 					      struct amdgpu_iv_entry *entry)
1496 {
1497 	return 0;
1498 }
1499 
1500 static int sdma_v6_0_set_clockgating_state(void *handle,
1501 					   enum amd_clockgating_state state)
1502 {
1503 	return 0;
1504 }
1505 
1506 static int sdma_v6_0_set_powergating_state(void *handle,
1507 					  enum amd_powergating_state state)
1508 {
1509 	return 0;
1510 }
1511 
1512 static void sdma_v6_0_get_clockgating_state(void *handle, u64 *flags)
1513 {
1514 }
1515 
1516 const struct amd_ip_funcs sdma_v6_0_ip_funcs = {
1517 	.name = "sdma_v6_0",
1518 	.early_init = sdma_v6_0_early_init,
1519 	.late_init = NULL,
1520 	.sw_init = sdma_v6_0_sw_init,
1521 	.sw_fini = sdma_v6_0_sw_fini,
1522 	.hw_init = sdma_v6_0_hw_init,
1523 	.hw_fini = sdma_v6_0_hw_fini,
1524 	.suspend = sdma_v6_0_suspend,
1525 	.resume = sdma_v6_0_resume,
1526 	.is_idle = sdma_v6_0_is_idle,
1527 	.wait_for_idle = sdma_v6_0_wait_for_idle,
1528 	.soft_reset = sdma_v6_0_soft_reset,
1529 	.set_clockgating_state = sdma_v6_0_set_clockgating_state,
1530 	.set_powergating_state = sdma_v6_0_set_powergating_state,
1531 	.get_clockgating_state = sdma_v6_0_get_clockgating_state,
1532 };
1533 
1534 static const struct amdgpu_ring_funcs sdma_v6_0_ring_funcs = {
1535 	.type = AMDGPU_RING_TYPE_SDMA,
1536 	.align_mask = 0xf,
1537 	.nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP),
1538 	.support_64bit_ptrs = true,
1539 	.vmhub = AMDGPU_GFXHUB_0,
1540 	.get_rptr = sdma_v6_0_ring_get_rptr,
1541 	.get_wptr = sdma_v6_0_ring_get_wptr,
1542 	.set_wptr = sdma_v6_0_ring_set_wptr,
1543 	.emit_frame_size =
1544 		5 + /* sdma_v6_0_ring_init_cond_exec */
1545 		6 + /* sdma_v6_0_ring_emit_hdp_flush */
1546 		6 + /* sdma_v6_0_ring_emit_pipeline_sync */
1547 		/* sdma_v6_0_ring_emit_vm_flush */
1548 		SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 +
1549 		SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6 +
1550 		10 + 10 + 10, /* sdma_v6_0_ring_emit_fence x3 for user fence, vm fence */
1551 	.emit_ib_size = 5 + 7 + 6, /* sdma_v6_0_ring_emit_ib */
1552 	.emit_ib = sdma_v6_0_ring_emit_ib,
1553 	.emit_mem_sync = sdma_v6_0_ring_emit_mem_sync,
1554 	.emit_fence = sdma_v6_0_ring_emit_fence,
1555 	.emit_pipeline_sync = sdma_v6_0_ring_emit_pipeline_sync,
1556 	.emit_vm_flush = sdma_v6_0_ring_emit_vm_flush,
1557 	.emit_hdp_flush = sdma_v6_0_ring_emit_hdp_flush,
1558 	.test_ring = sdma_v6_0_ring_test_ring,
1559 	.test_ib = sdma_v6_0_ring_test_ib,
1560 	.insert_nop = sdma_v6_0_ring_insert_nop,
1561 	.pad_ib = sdma_v6_0_ring_pad_ib,
1562 	.emit_wreg = sdma_v6_0_ring_emit_wreg,
1563 	.emit_reg_wait = sdma_v6_0_ring_emit_reg_wait,
1564 	.emit_reg_write_reg_wait = sdma_v6_0_ring_emit_reg_write_reg_wait,
1565 	.init_cond_exec = sdma_v6_0_ring_init_cond_exec,
1566 	.patch_cond_exec = sdma_v6_0_ring_patch_cond_exec,
1567 	.preempt_ib = sdma_v6_0_ring_preempt_ib,
1568 };
1569 
1570 static void sdma_v6_0_set_ring_funcs(struct amdgpu_device *adev)
1571 {
1572 	int i;
1573 
1574 	for (i = 0; i < adev->sdma.num_instances; i++) {
1575 		adev->sdma.instance[i].ring.funcs = &sdma_v6_0_ring_funcs;
1576 		adev->sdma.instance[i].ring.me = i;
1577 	}
1578 }
1579 
1580 static const struct amdgpu_irq_src_funcs sdma_v6_0_trap_irq_funcs = {
1581 	.set = sdma_v6_0_set_trap_irq_state,
1582 	.process = sdma_v6_0_process_trap_irq,
1583 };
1584 
1585 static const struct amdgpu_irq_src_funcs sdma_v6_0_illegal_inst_irq_funcs = {
1586 	.process = sdma_v6_0_process_illegal_inst_irq,
1587 };
1588 
1589 static void sdma_v6_0_set_irq_funcs(struct amdgpu_device *adev)
1590 {
1591 	adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_INSTANCE0 +
1592 					adev->sdma.num_instances;
1593 	adev->sdma.trap_irq.funcs = &sdma_v6_0_trap_irq_funcs;
1594 	adev->sdma.illegal_inst_irq.funcs = &sdma_v6_0_illegal_inst_irq_funcs;
1595 }
1596 
1597 /**
1598  * sdma_v6_0_emit_copy_buffer - copy buffer using the sDMA engine
1599  *
1600  * @ring: amdgpu_ring structure holding ring information
1601  * @src_offset: src GPU address
1602  * @dst_offset: dst GPU address
1603  * @byte_count: number of bytes to xfer
1604  *
1605  * Copy GPU buffers using the DMA engine.
1606  * Used by the amdgpu ttm implementation to move pages if
1607  * registered as the asic copy callback.
1608  */
1609 static void sdma_v6_0_emit_copy_buffer(struct amdgpu_ib *ib,
1610 				       uint64_t src_offset,
1611 				       uint64_t dst_offset,
1612 				       uint32_t byte_count,
1613 				       bool tmz)
1614 {
1615 	ib->ptr[ib->length_dw++] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_COPY) |
1616 		SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR) |
1617 		SDMA_PKT_COPY_LINEAR_HEADER_TMZ(tmz ? 1 : 0);
1618 	ib->ptr[ib->length_dw++] = byte_count - 1;
1619 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1620 	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1621 	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1622 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1623 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1624 }
1625 
1626 /**
1627  * sdma_v6_0_emit_fill_buffer - fill buffer using the sDMA engine
1628  *
1629  * @ring: amdgpu_ring structure holding ring information
1630  * @src_data: value to write to buffer
1631  * @dst_offset: dst GPU address
1632  * @byte_count: number of bytes to xfer
1633  *
1634  * Fill GPU buffers using the DMA engine.
1635  */
1636 static void sdma_v6_0_emit_fill_buffer(struct amdgpu_ib *ib,
1637 				       uint32_t src_data,
1638 				       uint64_t dst_offset,
1639 				       uint32_t byte_count)
1640 {
1641 	ib->ptr[ib->length_dw++] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_CONST_FILL);
1642 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1643 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1644 	ib->ptr[ib->length_dw++] = src_data;
1645 	ib->ptr[ib->length_dw++] = byte_count - 1;
1646 }
1647 
1648 static const struct amdgpu_buffer_funcs sdma_v6_0_buffer_funcs = {
1649 	.copy_max_bytes = 0x400000,
1650 	.copy_num_dw = 7,
1651 	.emit_copy_buffer = sdma_v6_0_emit_copy_buffer,
1652 
1653 	.fill_max_bytes = 0x400000,
1654 	.fill_num_dw = 5,
1655 	.emit_fill_buffer = sdma_v6_0_emit_fill_buffer,
1656 };
1657 
1658 static void sdma_v6_0_set_buffer_funcs(struct amdgpu_device *adev)
1659 {
1660 	adev->mman.buffer_funcs = &sdma_v6_0_buffer_funcs;
1661 	adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1662 }
1663 
1664 static const struct amdgpu_vm_pte_funcs sdma_v6_0_vm_pte_funcs = {
1665 	.copy_pte_num_dw = 7,
1666 	.copy_pte = sdma_v6_0_vm_copy_pte,
1667 	.write_pte = sdma_v6_0_vm_write_pte,
1668 	.set_pte_pde = sdma_v6_0_vm_set_pte_pde,
1669 };
1670 
1671 static void sdma_v6_0_set_vm_pte_funcs(struct amdgpu_device *adev)
1672 {
1673 	unsigned i;
1674 
1675 	adev->vm_manager.vm_pte_funcs = &sdma_v6_0_vm_pte_funcs;
1676 	for (i = 0; i < adev->sdma.num_instances; i++) {
1677 		adev->vm_manager.vm_pte_scheds[i] =
1678 			&adev->sdma.instance[i].ring.sched;
1679 	}
1680 	adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
1681 }
1682 
1683 const struct amdgpu_ip_block_version sdma_v6_0_ip_block = {
1684 	.type = AMD_IP_BLOCK_TYPE_SDMA,
1685 	.major = 6,
1686 	.minor = 0,
1687 	.rev = 0,
1688 	.funcs = &sdma_v6_0_ip_funcs,
1689 };
1690