xref: /linux/drivers/gpu/drm/amd/amdgpu/sdma_v7_0.c (revision 566ab427f827b0256d3e8ce0235d088e6a9c28bd)
1 /*
2  * Copyright 2023 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_12_0_0_offset.h"
34 #include "gc/gc_12_0_0_sh_mask.h"
35 #include "hdp/hdp_6_0_0_offset.h"
36 #include "ivsrcid/gfx/irqsrcs_gfx_11_0_0.h"
37 
38 #include "soc15_common.h"
39 #include "soc15.h"
40 #include "sdma_v6_0_0_pkt_open.h"
41 #include "nbio_v4_3.h"
42 #include "sdma_common.h"
43 #include "sdma_v7_0.h"
44 #include "v12_structs.h"
45 
46 MODULE_FIRMWARE("amdgpu/sdma_7_0_0.bin");
47 MODULE_FIRMWARE("amdgpu/sdma_7_0_1.bin");
48 
49 #define SDMA1_REG_OFFSET 0x600
50 #define SDMA0_HYP_DEC_REG_START 0x5880
51 #define SDMA0_HYP_DEC_REG_END 0x589a
52 #define SDMA1_HYP_DEC_REG_OFFSET 0x20
53 
54 static const struct amdgpu_hwip_reg_entry sdma_reg_list_7_0[] = {
55 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_STATUS_REG),
56 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_STATUS1_REG),
57 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_STATUS2_REG),
58 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_STATUS3_REG),
59 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_STATUS4_REG),
60 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_STATUS5_REG),
61 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_STATUS6_REG),
62 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_UCODE_REV),
63 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_RB_RPTR_FETCH_HI),
64 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_RB_RPTR_FETCH),
65 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_UTCL1_RD_STATUS),
66 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_UTCL1_WR_STATUS),
67 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_UTCL1_RD_XNACK0),
68 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_UTCL1_RD_XNACK1),
69 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_UTCL1_WR_XNACK0),
70 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_UTCL1_WR_XNACK1),
71 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE0_RB_CNTL),
72 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE0_RB_RPTR),
73 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE0_RB_RPTR_HI),
74 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE0_RB_WPTR),
75 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE0_RB_WPTR_HI),
76 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE0_IB_OFFSET),
77 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE0_IB_BASE_LO),
78 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE0_IB_BASE_HI),
79 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE0_IB_CNTL),
80 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE0_IB_RPTR),
81 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE0_IB_SUB_REMAIN),
82 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE0_DUMMY_REG),
83 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE_STATUS0),
84 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE1_RB_CNTL),
85 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE1_RB_RPTR),
86 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE1_RB_RPTR_HI),
87 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE1_RB_WPTR),
88 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE1_RB_WPTR_HI),
89 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE1_IB_OFFSET),
90 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE1_IB_BASE_LO),
91 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE1_IB_BASE_HI),
92 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE1_IB_RPTR),
93 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE1_IB_SUB_REMAIN),
94 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE1_DUMMY_REG),
95 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE2_RB_CNTL),
96 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE2_RB_RPTR),
97 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE2_RB_RPTR_HI),
98 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE2_RB_WPTR),
99 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE2_RB_WPTR_HI),
100 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE2_IB_OFFSET),
101 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE2_IB_BASE_LO),
102 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE2_IB_BASE_HI),
103 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE2_IB_RPTR),
104 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE2_IB_SUB_REMAIN),
105 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE2_DUMMY_REG),
106 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_INT_STATUS),
107 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_VM_CNTL),
108 	SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS2),
109 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_CHICKEN_BITS),
110 };
111 
112 static void sdma_v7_0_set_ring_funcs(struct amdgpu_device *adev);
113 static void sdma_v7_0_set_buffer_funcs(struct amdgpu_device *adev);
114 static void sdma_v7_0_set_vm_pte_funcs(struct amdgpu_device *adev);
115 static void sdma_v7_0_set_irq_funcs(struct amdgpu_device *adev);
116 static int sdma_v7_0_start(struct amdgpu_device *adev);
117 
118 static u32 sdma_v7_0_get_reg_offset(struct amdgpu_device *adev, u32 instance, u32 internal_offset)
119 {
120 	u32 base;
121 
122 	if (internal_offset >= SDMA0_HYP_DEC_REG_START &&
123 	    internal_offset <= SDMA0_HYP_DEC_REG_END) {
124 		base = adev->reg_offset[GC_HWIP][0][1];
125 		if (instance != 0)
126 			internal_offset += SDMA1_HYP_DEC_REG_OFFSET * instance;
127 	} else {
128 		base = adev->reg_offset[GC_HWIP][0][0];
129 		if (instance == 1)
130 			internal_offset += SDMA1_REG_OFFSET;
131 	}
132 
133 	return base + internal_offset;
134 }
135 
136 static unsigned sdma_v7_0_ring_init_cond_exec(struct amdgpu_ring *ring,
137 					      uint64_t addr)
138 {
139 	unsigned ret;
140 
141 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_COND_EXE));
142 	amdgpu_ring_write(ring, lower_32_bits(addr));
143 	amdgpu_ring_write(ring, upper_32_bits(addr));
144 	amdgpu_ring_write(ring, 1);
145 	/* this is the offset we need patch later */
146 	ret = ring->wptr & ring->buf_mask;
147 	/* insert dummy here and patch it later */
148 	amdgpu_ring_write(ring, 0);
149 
150 	return ret;
151 }
152 
153 /**
154  * sdma_v7_0_ring_get_rptr - get the current read pointer
155  *
156  * @ring: amdgpu ring pointer
157  *
158  * Get the current rptr from the hardware.
159  */
160 static uint64_t sdma_v7_0_ring_get_rptr(struct amdgpu_ring *ring)
161 {
162 	u64 *rptr;
163 
164 	/* XXX check if swapping is necessary on BE */
165 	rptr = (u64 *)ring->rptr_cpu_addr;
166 
167 	DRM_DEBUG("rptr before shift == 0x%016llx\n", *rptr);
168 	return ((*rptr) >> 2);
169 }
170 
171 /**
172  * sdma_v7_0_ring_get_wptr - get the current write pointer
173  *
174  * @ring: amdgpu ring pointer
175  *
176  * Get the current wptr from the hardware.
177  */
178 static uint64_t sdma_v7_0_ring_get_wptr(struct amdgpu_ring *ring)
179 {
180 	u64 wptr = 0;
181 
182 	if (ring->use_doorbell) {
183 		/* XXX check if swapping is necessary on BE */
184 		wptr = READ_ONCE(*((u64 *)ring->wptr_cpu_addr));
185 		DRM_DEBUG("wptr/doorbell before shift == 0x%016llx\n", wptr);
186 	}
187 
188 	return wptr >> 2;
189 }
190 
191 /**
192  * sdma_v7_0_ring_set_wptr - commit the write pointer
193  *
194  * @ring: amdgpu ring pointer
195  *
196  * Write the wptr back to the hardware.
197  */
198 static void sdma_v7_0_ring_set_wptr(struct amdgpu_ring *ring)
199 {
200 	struct amdgpu_device *adev = ring->adev;
201 	uint32_t *wptr_saved;
202 	uint32_t *is_queue_unmap;
203 	uint64_t aggregated_db_index;
204 	uint32_t mqd_size = adev->mqds[AMDGPU_HW_IP_DMA].mqd_size;
205 
206 	DRM_DEBUG("Setting write pointer\n");
207 
208 	if (ring->is_mes_queue) {
209 		wptr_saved = (uint32_t *)(ring->mqd_ptr + mqd_size);
210 		is_queue_unmap = (uint32_t *)(ring->mqd_ptr + mqd_size +
211 					      sizeof(uint32_t));
212 		aggregated_db_index =
213 			amdgpu_mes_get_aggregated_doorbell_index(adev,
214 							 ring->hw_prio);
215 
216 		atomic64_set((atomic64_t *)ring->wptr_cpu_addr,
217 			     ring->wptr << 2);
218 		*wptr_saved = ring->wptr << 2;
219 		if (*is_queue_unmap) {
220 			WDOORBELL64(aggregated_db_index, ring->wptr << 2);
221 			DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016llx)\n",
222 					ring->doorbell_index, ring->wptr << 2);
223 			WDOORBELL64(ring->doorbell_index, ring->wptr << 2);
224 		} else {
225 			DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016llx)\n",
226 					ring->doorbell_index, ring->wptr << 2);
227 			WDOORBELL64(ring->doorbell_index, ring->wptr << 2);
228 		}
229 	} else {
230 		if (ring->use_doorbell) {
231 			DRM_DEBUG("Using doorbell -- "
232 				  "wptr_offs == 0x%08x "
233 				  "lower_32_bits(ring->wptr) << 2 == 0x%08x "
234 				  "upper_32_bits(ring->wptr) << 2 == 0x%08x\n",
235 				  ring->wptr_offs,
236 				  lower_32_bits(ring->wptr << 2),
237 				  upper_32_bits(ring->wptr << 2));
238 			/* XXX check if swapping is necessary on BE */
239 			atomic64_set((atomic64_t *)ring->wptr_cpu_addr,
240 				     ring->wptr << 2);
241 			DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016llx)\n",
242 				  ring->doorbell_index, ring->wptr << 2);
243 			WDOORBELL64(ring->doorbell_index, ring->wptr << 2);
244 		} else {
245 			DRM_DEBUG("Not using doorbell -- "
246 				  "regSDMA%i_GFX_RB_WPTR == 0x%08x "
247 				  "regSDMA%i_GFX_RB_WPTR_HI == 0x%08x\n",
248 				  ring->me,
249 				  lower_32_bits(ring->wptr << 2),
250 				  ring->me,
251 				  upper_32_bits(ring->wptr << 2));
252 			WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev,
253 								     ring->me,
254 								     regSDMA0_QUEUE0_RB_WPTR),
255 					lower_32_bits(ring->wptr << 2));
256 			WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev,
257 								     ring->me,
258 								     regSDMA0_QUEUE0_RB_WPTR_HI),
259 					upper_32_bits(ring->wptr << 2));
260 		}
261 	}
262 }
263 
264 static void sdma_v7_0_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
265 {
266 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
267 	int i;
268 
269 	for (i = 0; i < count; i++)
270 		if (sdma && sdma->burst_nop && (i == 0))
271 			amdgpu_ring_write(ring, ring->funcs->nop |
272 				SDMA_PKT_NOP_HEADER_COUNT(count - 1));
273 		else
274 			amdgpu_ring_write(ring, ring->funcs->nop);
275 }
276 
277 /**
278  * sdma_v7_0_ring_emit_ib - Schedule an IB on the DMA engine
279  *
280  * @ring: amdgpu ring pointer
281  * @job: job to retrieve vmid from
282  * @ib: IB object to schedule
283  * @flags: unused
284  *
285  * Schedule an IB in the DMA ring.
286  */
287 static void sdma_v7_0_ring_emit_ib(struct amdgpu_ring *ring,
288 				   struct amdgpu_job *job,
289 				   struct amdgpu_ib *ib,
290 				   uint32_t flags)
291 {
292 	unsigned vmid = AMDGPU_JOB_GET_VMID(job);
293 	uint64_t csa_mc_addr = amdgpu_sdma_get_csa_mc_addr(ring, vmid);
294 
295 	/* An IB packet must end on a 8 DW boundary--the next dword
296 	 * must be on a 8-dword boundary. Our IB packet below is 6
297 	 * dwords long, thus add x number of NOPs, such that, in
298 	 * modular arithmetic,
299 	 * wptr + 6 + x = 8k, k >= 0, which in C is,
300 	 * (wptr + 6 + x) % 8 = 0.
301 	 * The expression below, is a solution of x.
302 	 */
303 	sdma_v7_0_ring_insert_nop(ring, (2 - lower_32_bits(ring->wptr)) & 7);
304 
305 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_INDIRECT) |
306 			  SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf));
307 	/* base must be 32 byte aligned */
308 	amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
309 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
310 	amdgpu_ring_write(ring, ib->length_dw);
311 	amdgpu_ring_write(ring, lower_32_bits(csa_mc_addr));
312 	amdgpu_ring_write(ring, upper_32_bits(csa_mc_addr));
313 }
314 
315 /**
316  * sdma_v7_0_ring_emit_mem_sync - flush the IB by graphics cache rinse
317  *
318  * @ring: amdgpu ring pointer
319  *
320  * flush the IB by graphics cache rinse.
321  */
322 static void sdma_v7_0_ring_emit_mem_sync(struct amdgpu_ring *ring)
323 {
324 	uint32_t gcr_cntl = SDMA_GCR_GL2_INV | SDMA_GCR_GL2_WB | SDMA_GCR_GLM_INV |
325 		SDMA_GCR_GL1_INV | SDMA_GCR_GLV_INV | SDMA_GCR_GLK_INV |
326 		SDMA_GCR_GLI_INV(1);
327 
328 	/* flush entire cache L0/L1/L2, this can be optimized by performance requirement */
329 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_GCR_REQ));
330 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD1_BASE_VA_31_7(0));
331 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD2_GCR_CONTROL_15_0(gcr_cntl) |
332 			  SDMA_PKT_GCR_REQ_PAYLOAD2_BASE_VA_47_32(0));
333 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD3_LIMIT_VA_31_7(0) |
334 			  SDMA_PKT_GCR_REQ_PAYLOAD3_GCR_CONTROL_18_16(gcr_cntl >> 16));
335 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD4_LIMIT_VA_47_32(0) |
336 			  SDMA_PKT_GCR_REQ_PAYLOAD4_VMID(0));
337 }
338 
339 
340 /**
341  * sdma_v7_0_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
342  *
343  * @ring: amdgpu ring pointer
344  *
345  * Emit an hdp flush packet on the requested DMA ring.
346  */
347 static void sdma_v7_0_ring_emit_hdp_flush(struct amdgpu_ring *ring)
348 {
349 	struct amdgpu_device *adev = ring->adev;
350 	u32 ref_and_mask = 0;
351 	const struct nbio_hdp_flush_reg *nbio_hf_reg = adev->nbio.hdp_flush_reg;
352 
353 	ref_and_mask = nbio_hf_reg->ref_and_mask_sdma0 << ring->me;
354 
355 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_POLL_REGMEM) |
356 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) |
357 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
358 	amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_done_offset(adev)) << 2);
359 	amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_req_offset(adev)) << 2);
360 	amdgpu_ring_write(ring, ref_and_mask); /* reference */
361 	amdgpu_ring_write(ring, ref_and_mask); /* mask */
362 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
363 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
364 }
365 
366 /**
367  * sdma_v7_0_ring_emit_fence - emit a fence on the DMA ring
368  *
369  * @ring: amdgpu ring pointer
370  * @addr: address
371  * @seq: fence seq number
372  * @flags: fence flags
373  *
374  * Add a DMA fence packet to the ring to write
375  * the fence seq number and DMA trap packet to generate
376  * an interrupt if needed.
377  */
378 static void sdma_v7_0_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
379 				      unsigned flags)
380 {
381 	bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
382 	/* write the fence */
383 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_FENCE) |
384 			  SDMA_PKT_FENCE_HEADER_MTYPE(0x3)); /* Ucached(UC) */
385 	/* zero in first two bits */
386 	BUG_ON(addr & 0x3);
387 	amdgpu_ring_write(ring, lower_32_bits(addr));
388 	amdgpu_ring_write(ring, upper_32_bits(addr));
389 	amdgpu_ring_write(ring, lower_32_bits(seq));
390 
391 	/* optionally write high bits as well */
392 	if (write64bit) {
393 		addr += 4;
394 		amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_FENCE) |
395 				  SDMA_PKT_FENCE_HEADER_MTYPE(0x3));
396 		/* zero in first two bits */
397 		BUG_ON(addr & 0x3);
398 		amdgpu_ring_write(ring, lower_32_bits(addr));
399 		amdgpu_ring_write(ring, upper_32_bits(addr));
400 		amdgpu_ring_write(ring, upper_32_bits(seq));
401 	}
402 
403 	if (flags & AMDGPU_FENCE_FLAG_INT) {
404 		uint32_t ctx = ring->is_mes_queue ?
405 			(ring->hw_queue_id | AMDGPU_FENCE_MES_QUEUE_FLAG) : 0;
406 		/* generate an interrupt */
407 		amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_TRAP));
408 		amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(ctx));
409 	}
410 }
411 
412 /**
413  * sdma_v7_0_gfx_stop - stop the gfx async dma engines
414  *
415  * @adev: amdgpu_device pointer
416  *
417  * Stop the gfx async dma ring buffers.
418  */
419 static void sdma_v7_0_gfx_stop(struct amdgpu_device *adev)
420 {
421 	u32 rb_cntl, ib_cntl;
422 	int i;
423 
424 	for (i = 0; i < adev->sdma.num_instances; i++) {
425 		rb_cntl = RREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_CNTL));
426 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RB_ENABLE, 0);
427 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_CNTL), rb_cntl);
428 		ib_cntl = RREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_IB_CNTL));
429 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_QUEUE0_IB_CNTL, IB_ENABLE, 0);
430 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_IB_CNTL), ib_cntl);
431 	}
432 }
433 
434 /**
435  * sdma_v7_0_rlc_stop - stop the compute async dma engines
436  *
437  * @adev: amdgpu_device pointer
438  *
439  * Stop the compute async dma queues.
440  */
441 static void sdma_v7_0_rlc_stop(struct amdgpu_device *adev)
442 {
443 	/* XXX todo */
444 }
445 
446 /**
447  * sdma_v7_0_ctx_switch_enable - stop the async dma engines context switch
448  *
449  * @adev: amdgpu_device pointer
450  * @enable: enable/disable the DMA MEs context switch.
451  *
452  * Halt or unhalt the async dma engines context switch.
453  */
454 static void sdma_v7_0_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
455 {
456 }
457 
458 /**
459  * sdma_v7_0_enable - stop the async dma engines
460  *
461  * @adev: amdgpu_device pointer
462  * @enable: enable/disable the DMA MEs.
463  *
464  * Halt or unhalt the async dma engines.
465  */
466 static void sdma_v7_0_enable(struct amdgpu_device *adev, bool enable)
467 {
468 	u32 mcu_cntl;
469 	int i;
470 
471 	if (!enable) {
472 		sdma_v7_0_gfx_stop(adev);
473 		sdma_v7_0_rlc_stop(adev);
474 	}
475 
476 	if (amdgpu_sriov_vf(adev))
477 		return;
478 
479 	for (i = 0; i < adev->sdma.num_instances; i++) {
480 		mcu_cntl = RREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_MCU_CNTL));
481 		mcu_cntl = REG_SET_FIELD(mcu_cntl, SDMA0_MCU_CNTL, HALT, enable ? 0 : 1);
482 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_MCU_CNTL), mcu_cntl);
483 	}
484 }
485 
486 /**
487  * sdma_v7_0_gfx_resume - setup and start the async dma engines
488  *
489  * @adev: amdgpu_device pointer
490  *
491  * Set up the gfx DMA ring buffers and enable them.
492  * Returns 0 for success, error for failure.
493  */
494 static int sdma_v7_0_gfx_resume(struct amdgpu_device *adev)
495 {
496 	struct amdgpu_ring *ring;
497 	u32 rb_cntl, ib_cntl;
498 	u32 rb_bufsz;
499 	u32 doorbell;
500 	u32 doorbell_offset;
501 	u32 tmp;
502 	u64 wptr_gpu_addr;
503 	int i, r;
504 
505 	for (i = 0; i < adev->sdma.num_instances; i++) {
506 		ring = &adev->sdma.instance[i].ring;
507 
508 		//if (!amdgpu_sriov_vf(adev))
509 		//	WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_SEM_WAIT_FAIL_TIMER_CNTL), 0);
510 
511 		/* Set ring buffer size in dwords */
512 		rb_bufsz = order_base_2(ring->ring_size / 4);
513 		rb_cntl = RREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_CNTL));
514 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RB_SIZE, rb_bufsz);
515 #ifdef __BIG_ENDIAN
516 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RB_SWAP_ENABLE, 1);
517 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL,
518 					RPTR_WRITEBACK_SWAP_ENABLE, 1);
519 #endif
520 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RB_PRIV, 1);
521 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_CNTL), rb_cntl);
522 
523 		/* Initialize the ring buffer's read and write pointers */
524 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_RPTR), 0);
525 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_RPTR_HI), 0);
526 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR), 0);
527 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR_HI), 0);
528 
529 		/* setup the wptr shadow polling */
530 		wptr_gpu_addr = ring->wptr_gpu_addr;
531 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR_POLL_ADDR_LO),
532 		       lower_32_bits(wptr_gpu_addr));
533 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR_POLL_ADDR_HI),
534 		       upper_32_bits(wptr_gpu_addr));
535 
536 		/* set the wb address whether it's enabled or not */
537 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_RPTR_ADDR_HI),
538 		       upper_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFF);
539 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_RPTR_ADDR_LO),
540 		       lower_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFC);
541 
542 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
543 		if (amdgpu_sriov_vf(adev))
544 			rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, WPTR_POLL_ENABLE, 1);
545 		else
546 			rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, WPTR_POLL_ENABLE, 0);
547 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, MCU_WPTR_POLL_ENABLE, 1);
548 
549 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_BASE), ring->gpu_addr >> 8);
550 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_BASE_HI), ring->gpu_addr >> 40);
551 
552 		ring->wptr = 0;
553 
554 		/* before programing wptr to a less value, need set minor_ptr_update first */
555 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_MINOR_PTR_UPDATE), 1);
556 
557 		if (!amdgpu_sriov_vf(adev)) { /* only bare-metal use register write for wptr */
558 			WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR), lower_32_bits(ring->wptr) << 2);
559 			WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR_HI), upper_32_bits(ring->wptr) << 2);
560 		}
561 
562 		doorbell = RREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_DOORBELL));
563 		doorbell_offset = RREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_DOORBELL_OFFSET));
564 
565 		if (ring->use_doorbell) {
566 			doorbell = REG_SET_FIELD(doorbell, SDMA0_QUEUE0_DOORBELL, ENABLE, 1);
567 			doorbell_offset = REG_SET_FIELD(doorbell_offset, SDMA0_QUEUE0_DOORBELL_OFFSET,
568 					OFFSET, ring->doorbell_index);
569 		} else {
570 			doorbell = REG_SET_FIELD(doorbell, SDMA0_QUEUE0_DOORBELL, ENABLE, 0);
571 		}
572 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_DOORBELL), doorbell);
573 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_DOORBELL_OFFSET), doorbell_offset);
574 
575 		if (i == 0)
576 			adev->nbio.funcs->sdma_doorbell_range(adev, i, ring->use_doorbell,
577 						      ring->doorbell_index,
578 						      adev->doorbell_index.sdma_doorbell_range * adev->sdma.num_instances);
579 
580 		if (amdgpu_sriov_vf(adev))
581 			sdma_v7_0_ring_set_wptr(ring);
582 
583 		/* set minor_ptr_update to 0 after wptr programed */
584 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_MINOR_PTR_UPDATE), 0);
585 
586 		/* Set up sdma hang watchdog */
587 		tmp = RREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_WATCHDOG_CNTL));
588 		/* 100ms per unit */
589 		tmp = REG_SET_FIELD(tmp, SDMA0_WATCHDOG_CNTL, QUEUE_HANG_COUNT,
590 				    max(adev->usec_timeout/100000, 1));
591 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_WATCHDOG_CNTL), tmp);
592 
593 		/* Set up RESP_MODE to non-copy addresses */
594 		tmp = RREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_UTCL1_CNTL));
595 		tmp = REG_SET_FIELD(tmp, SDMA0_UTCL1_CNTL, RESP_MODE, 3);
596 		tmp = REG_SET_FIELD(tmp, SDMA0_UTCL1_CNTL, REDO_DELAY, 9);
597 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_UTCL1_CNTL), tmp);
598 
599 		/* program default cache read and write policy */
600 		tmp = RREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_UTCL1_PAGE));
601 		/* clean read policy and write policy bits */
602 		tmp &= 0xFF0FFF;
603 		tmp |= ((CACHE_READ_POLICY_L2__DEFAULT << 12) |
604 			 (CACHE_WRITE_POLICY_L2__DEFAULT << 14));
605 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_UTCL1_PAGE), tmp);
606 
607 		if (!amdgpu_sriov_vf(adev)) {
608 			/* unhalt engine */
609 			tmp = RREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_MCU_CNTL));
610 			tmp = REG_SET_FIELD(tmp, SDMA0_MCU_CNTL, HALT, 0);
611 			tmp = REG_SET_FIELD(tmp, SDMA0_MCU_CNTL, RESET, 0);
612 			WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_MCU_CNTL), tmp);
613 		}
614 
615 		/* enable DMA RB */
616 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RB_ENABLE, 1);
617 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_CNTL), rb_cntl);
618 
619 		ib_cntl = RREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_IB_CNTL));
620 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_QUEUE0_IB_CNTL, IB_ENABLE, 1);
621 #ifdef __BIG_ENDIAN
622 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_QUEUE0_IB_CNTL, IB_SWAP_ENABLE, 1);
623 #endif
624 		/* enable DMA IBs */
625 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_IB_CNTL), ib_cntl);
626 
627 		ring->sched.ready = true;
628 
629 		if (amdgpu_sriov_vf(adev)) { /* bare-metal sequence doesn't need below to lines */
630 			sdma_v7_0_ctx_switch_enable(adev, true);
631 			sdma_v7_0_enable(adev, true);
632 		}
633 
634 		r = amdgpu_ring_test_helper(ring);
635 		if (r) {
636 			ring->sched.ready = false;
637 			return r;
638 		}
639 
640 	}
641 
642 	return 0;
643 }
644 
645 /**
646  * sdma_v7_0_rlc_resume - setup and start the async dma engines
647  *
648  * @adev: amdgpu_device pointer
649  *
650  * Set up the compute DMA queues and enable them.
651  * Returns 0 for success, error for failure.
652  */
653 static int sdma_v7_0_rlc_resume(struct amdgpu_device *adev)
654 {
655 	return 0;
656 }
657 
658 static void sdma_v12_0_free_ucode_buffer(struct amdgpu_device *adev)
659 {
660 	int i;
661 
662 	for (i = 0; i < adev->sdma.num_instances; i++) {
663 		amdgpu_bo_free_kernel(&adev->sdma.instance[i].sdma_fw_obj,
664 				      &adev->sdma.instance[i].sdma_fw_gpu_addr,
665 				      (void **)&adev->sdma.instance[i].sdma_fw_ptr);
666 	}
667 }
668 
669 /**
670  * sdma_v7_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_v7_0_load_microcode(struct amdgpu_device *adev)
678 {
679 	const struct sdma_firmware_header_v3_0 *hdr;
680 	const __le32 *fw_data;
681 	u32 fw_size;
682 	uint32_t tmp, sdma_status, ic_op_cntl;
683 	int i, r, j;
684 
685 	/* halt the MEs */
686 	sdma_v7_0_enable(adev, false);
687 
688 	if (!adev->sdma.instance[0].fw)
689 		return -EINVAL;
690 
691 	hdr = (const struct sdma_firmware_header_v3_0 *)
692 		adev->sdma.instance[0].fw->data;
693 	amdgpu_ucode_print_sdma_hdr(&hdr->header);
694 
695 	fw_data = (const __le32 *)(adev->sdma.instance[0].fw->data +
696 			le32_to_cpu(hdr->ucode_offset_bytes));
697 	fw_size = le32_to_cpu(hdr->ucode_size_bytes);
698 
699 	for (i = 0; i < adev->sdma.num_instances; i++) {
700 		r = amdgpu_bo_create_reserved(adev, fw_size,
701 					      PAGE_SIZE,
702 					      AMDGPU_GEM_DOMAIN_VRAM,
703 					      &adev->sdma.instance[i].sdma_fw_obj,
704 					      &adev->sdma.instance[i].sdma_fw_gpu_addr,
705 					      (void **)&adev->sdma.instance[i].sdma_fw_ptr);
706 		if (r) {
707 			dev_err(adev->dev, "(%d) failed to create sdma ucode bo\n", r);
708 			return r;
709 		}
710 
711 		memcpy(adev->sdma.instance[i].sdma_fw_ptr, fw_data, fw_size);
712 
713 		amdgpu_bo_kunmap(adev->sdma.instance[i].sdma_fw_obj);
714 		amdgpu_bo_unreserve(adev->sdma.instance[i].sdma_fw_obj);
715 
716 		tmp = RREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_IC_CNTL));
717 		tmp = REG_SET_FIELD(tmp, SDMA0_IC_CNTL, GPA, 0);
718 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_IC_CNTL), tmp);
719 
720 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_IC_BASE_LO),
721 			lower_32_bits(adev->sdma.instance[i].sdma_fw_gpu_addr));
722 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_IC_BASE_HI),
723 			upper_32_bits(adev->sdma.instance[i].sdma_fw_gpu_addr));
724 
725 		tmp = RREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_IC_OP_CNTL));
726 		tmp = REG_SET_FIELD(tmp, SDMA0_IC_OP_CNTL, PRIME_ICACHE, 1);
727 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_IC_OP_CNTL), tmp);
728 
729 		/* Wait for sdma ucode init complete */
730 		for (j = 0; j < adev->usec_timeout; j++) {
731 			ic_op_cntl = RREG32_SOC15_IP(GC,
732 					sdma_v7_0_get_reg_offset(adev, i, regSDMA0_IC_OP_CNTL));
733 			sdma_status = RREG32_SOC15_IP(GC,
734 					sdma_v7_0_get_reg_offset(adev, i, regSDMA0_STATUS_REG));
735 			if ((REG_GET_FIELD(ic_op_cntl, SDMA0_IC_OP_CNTL, ICACHE_PRIMED) == 1) &&
736 			    (REG_GET_FIELD(sdma_status, SDMA0_STATUS_REG, UCODE_INIT_DONE) == 1))
737 				break;
738 			udelay(1);
739 		}
740 
741 		if (j >= adev->usec_timeout) {
742 			dev_err(adev->dev, "failed to init sdma ucode\n");
743 			return -EINVAL;
744 		}
745 	}
746 
747 	return 0;
748 }
749 
750 static int sdma_v7_0_soft_reset(void *handle)
751 {
752 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
753 	u32 tmp;
754 	int i;
755 
756 	sdma_v7_0_gfx_stop(adev);
757 
758 	for (i = 0; i < adev->sdma.num_instances; i++) {
759 		//tmp = RREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_FREEZE));
760 		//tmp |= SDMA0_FREEZE__FREEZE_MASK;
761 		//WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_FREEZE), tmp);
762 		tmp = RREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_MCU_CNTL));
763 		tmp |= SDMA0_MCU_CNTL__HALT_MASK;
764 		tmp |= SDMA0_MCU_CNTL__RESET_MASK;
765 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_MCU_CNTL), tmp);
766 
767 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_PREEMPT), 0);
768 
769 		udelay(100);
770 
771 		tmp = GRBM_SOFT_RESET__SOFT_RESET_SDMA0_MASK << i;
772 		WREG32_SOC15(GC, 0, regGRBM_SOFT_RESET, tmp);
773 		tmp = RREG32_SOC15(GC, 0, regGRBM_SOFT_RESET);
774 
775 		udelay(100);
776 
777 		WREG32_SOC15(GC, 0, regGRBM_SOFT_RESET, 0);
778 		tmp = RREG32_SOC15(GC, 0, regGRBM_SOFT_RESET);
779 
780 		udelay(100);
781 	}
782 
783 	return sdma_v7_0_start(adev);
784 }
785 
786 static bool sdma_v7_0_check_soft_reset(void *handle)
787 {
788 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
789 	struct amdgpu_ring *ring;
790 	int i, r;
791 	long tmo = msecs_to_jiffies(1000);
792 
793 	for (i = 0; i < adev->sdma.num_instances; i++) {
794 		ring = &adev->sdma.instance[i].ring;
795 		r = amdgpu_ring_test_ib(ring, tmo);
796 		if (r)
797 			return true;
798 	}
799 
800 	return false;
801 }
802 
803 /**
804  * sdma_v7_0_start - setup and start the async dma engines
805  *
806  * @adev: amdgpu_device pointer
807  *
808  * Set up the DMA engines and enable them.
809  * Returns 0 for success, error for failure.
810  */
811 static int sdma_v7_0_start(struct amdgpu_device *adev)
812 {
813 	int r = 0;
814 
815 	if (amdgpu_sriov_vf(adev)) {
816 		sdma_v7_0_ctx_switch_enable(adev, false);
817 		sdma_v7_0_enable(adev, false);
818 
819 		/* set RB registers */
820 		r = sdma_v7_0_gfx_resume(adev);
821 		return r;
822 	}
823 
824 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
825 		r = sdma_v7_0_load_microcode(adev);
826 		if (r) {
827 			sdma_v12_0_free_ucode_buffer(adev);
828 			return r;
829 		}
830 
831 		if (amdgpu_emu_mode == 1)
832 			msleep(1000);
833 	}
834 
835 	/* unhalt the MEs */
836 	sdma_v7_0_enable(adev, true);
837 	/* enable sdma ring preemption */
838 	sdma_v7_0_ctx_switch_enable(adev, true);
839 
840 	/* start the gfx rings and rlc compute queues */
841 	r = sdma_v7_0_gfx_resume(adev);
842 	if (r)
843 		return r;
844 	r = sdma_v7_0_rlc_resume(adev);
845 
846 	return r;
847 }
848 
849 static int sdma_v7_0_mqd_init(struct amdgpu_device *adev, void *mqd,
850 			      struct amdgpu_mqd_prop *prop)
851 {
852 	struct v12_sdma_mqd *m = mqd;
853 	uint64_t wb_gpu_addr;
854 
855 	m->sdmax_rlcx_rb_cntl =
856 		order_base_2(prop->queue_size / 4) << SDMA0_QUEUE0_RB_CNTL__RB_SIZE__SHIFT |
857 		1 << SDMA0_QUEUE0_RB_CNTL__RPTR_WRITEBACK_ENABLE__SHIFT |
858 		4 << SDMA0_QUEUE0_RB_CNTL__RPTR_WRITEBACK_TIMER__SHIFT |
859 		1 << SDMA0_QUEUE0_RB_CNTL__MCU_WPTR_POLL_ENABLE__SHIFT;
860 
861 	m->sdmax_rlcx_rb_base = lower_32_bits(prop->hqd_base_gpu_addr >> 8);
862 	m->sdmax_rlcx_rb_base_hi = upper_32_bits(prop->hqd_base_gpu_addr >> 8);
863 
864 	wb_gpu_addr = prop->wptr_gpu_addr;
865 	m->sdmax_rlcx_rb_wptr_poll_addr_lo = lower_32_bits(wb_gpu_addr);
866 	m->sdmax_rlcx_rb_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr);
867 
868 	wb_gpu_addr = prop->rptr_gpu_addr;
869 	m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits(wb_gpu_addr);
870 	m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits(wb_gpu_addr);
871 
872 	m->sdmax_rlcx_ib_cntl = RREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, 0,
873 							regSDMA0_QUEUE0_IB_CNTL));
874 
875 	m->sdmax_rlcx_doorbell_offset =
876 		prop->doorbell_index << SDMA0_QUEUE0_DOORBELL_OFFSET__OFFSET__SHIFT;
877 
878 	m->sdmax_rlcx_doorbell = REG_SET_FIELD(0, SDMA0_QUEUE0_DOORBELL, ENABLE, 1);
879 
880 	m->sdmax_rlcx_doorbell_log = 0;
881 	m->sdmax_rlcx_rb_aql_cntl = 0x4000;	//regSDMA0_QUEUE0_RB_AQL_CNTL_DEFAULT;
882 	m->sdmax_rlcx_dummy_reg = 0xf;	//regSDMA0_QUEUE0_DUMMY_REG_DEFAULT;
883 
884 	return 0;
885 }
886 
887 static void sdma_v7_0_set_mqd_funcs(struct amdgpu_device *adev)
888 {
889 	adev->mqds[AMDGPU_HW_IP_DMA].mqd_size = sizeof(struct v12_sdma_mqd);
890 	adev->mqds[AMDGPU_HW_IP_DMA].init_mqd = sdma_v7_0_mqd_init;
891 }
892 
893 /**
894  * sdma_v7_0_ring_test_ring - simple async dma engine test
895  *
896  * @ring: amdgpu_ring structure holding ring information
897  *
898  * Test the DMA engine by writing using it to write an
899  * value to memory.
900  * Returns 0 for success, error for failure.
901  */
902 static int sdma_v7_0_ring_test_ring(struct amdgpu_ring *ring)
903 {
904 	struct amdgpu_device *adev = ring->adev;
905 	unsigned i;
906 	unsigned index;
907 	int r;
908 	u32 tmp;
909 	u64 gpu_addr;
910 	volatile uint32_t *cpu_ptr = NULL;
911 
912 	tmp = 0xCAFEDEAD;
913 
914 	if (ring->is_mes_queue) {
915 		uint32_t offset = 0;
916 		offset = amdgpu_mes_ctx_get_offs(ring,
917 					 AMDGPU_MES_CTX_PADDING_OFFS);
918 		gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
919 		cpu_ptr = amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
920 		*cpu_ptr = tmp;
921 	} else {
922 		r = amdgpu_device_wb_get(adev, &index);
923 		if (r) {
924 			dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
925 			return r;
926 		}
927 
928 		gpu_addr = adev->wb.gpu_addr + (index * 4);
929 		adev->wb.wb[index] = cpu_to_le32(tmp);
930 	}
931 
932 	r = amdgpu_ring_alloc(ring, 5);
933 	if (r) {
934 		DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
935 		if (!ring->is_mes_queue)
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_v7_0_ring_test_ib - test an IB on the DMA engine
972  *
973  * @ring: amdgpu_ring structure holding ring information
974  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
975  *
976  * Test a simple IB in the DMA ring.
977  * Returns 0 on success, error on failure.
978  */
979 static int sdma_v7_0_ring_test_ib(struct amdgpu_ring *ring, long timeout)
980 {
981 	struct amdgpu_device *adev = ring->adev;
982 	struct amdgpu_ib ib;
983 	struct dma_fence *f = NULL;
984 	unsigned index;
985 	long r;
986 	u32 tmp = 0;
987 	u64 gpu_addr;
988 	volatile uint32_t *cpu_ptr = NULL;
989 
990 	tmp = 0xCAFEDEAD;
991 	memset(&ib, 0, sizeof(ib));
992 
993 	if (ring->is_mes_queue) {
994 		uint32_t offset = 0;
995 		offset = amdgpu_mes_ctx_get_offs(ring, AMDGPU_MES_CTX_IB_OFFS);
996 		ib.gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
997 		ib.ptr = (void *)amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
998 
999 		offset = amdgpu_mes_ctx_get_offs(ring,
1000 					 AMDGPU_MES_CTX_PADDING_OFFS);
1001 		gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
1002 		cpu_ptr = amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
1003 		*cpu_ptr = tmp;
1004 	} else {
1005 		r = amdgpu_device_wb_get(adev, &index);
1006 		if (r) {
1007 			dev_err(adev->dev, "(%ld) failed to allocate wb slot\n", r);
1008 			return r;
1009 		}
1010 
1011 		gpu_addr = adev->wb.gpu_addr + (index * 4);
1012 		adev->wb.wb[index] = cpu_to_le32(tmp);
1013 
1014 		r = amdgpu_ib_get(adev, NULL, 256, AMDGPU_IB_POOL_DIRECT, &ib);
1015 		if (r) {
1016 			DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r);
1017 			goto err0;
1018 		}
1019 	}
1020 
1021 	ib.ptr[0] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_WRITE) |
1022 		SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1023 	ib.ptr[1] = lower_32_bits(gpu_addr);
1024 	ib.ptr[2] = upper_32_bits(gpu_addr);
1025 	ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0);
1026 	ib.ptr[4] = 0xDEADBEEF;
1027 	ib.ptr[5] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1028 	ib.ptr[6] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1029 	ib.ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1030 	ib.length_dw = 8;
1031 
1032 	r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
1033 	if (r)
1034 		goto err1;
1035 
1036 	r = dma_fence_wait_timeout(f, false, timeout);
1037 	if (r == 0) {
1038 		DRM_ERROR("amdgpu: IB test timed out\n");
1039 		r = -ETIMEDOUT;
1040 		goto err1;
1041 	} else if (r < 0) {
1042 		DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
1043 		goto err1;
1044 	}
1045 
1046 	if (ring->is_mes_queue)
1047 		tmp = le32_to_cpu(*cpu_ptr);
1048 	else
1049 		tmp = le32_to_cpu(adev->wb.wb[index]);
1050 
1051 	if (tmp == 0xDEADBEEF)
1052 		r = 0;
1053 	else
1054 		r = -EINVAL;
1055 
1056 err1:
1057 	amdgpu_ib_free(adev, &ib, NULL);
1058 	dma_fence_put(f);
1059 err0:
1060 	if (!ring->is_mes_queue)
1061 		amdgpu_device_wb_free(adev, index);
1062 	return r;
1063 }
1064 
1065 
1066 /**
1067  * sdma_v7_0_vm_copy_pte - update PTEs by copying them from the GART
1068  *
1069  * @ib: indirect buffer to fill with commands
1070  * @pe: addr of the page entry
1071  * @src: src addr to copy from
1072  * @count: number of page entries to update
1073  *
1074  * Update PTEs by copying them from the GART using sDMA.
1075  */
1076 static void sdma_v7_0_vm_copy_pte(struct amdgpu_ib *ib,
1077 				  uint64_t pe, uint64_t src,
1078 				  unsigned count)
1079 {
1080 	unsigned bytes = count * 8;
1081 
1082 	ib->ptr[ib->length_dw++] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_COPY) |
1083 		SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR) |
1084 		SDMA_PKT_COPY_LINEAR_HEADER_CPV(1);
1085 
1086 	ib->ptr[ib->length_dw++] = bytes - 1;
1087 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1088 	ib->ptr[ib->length_dw++] = lower_32_bits(src);
1089 	ib->ptr[ib->length_dw++] = upper_32_bits(src);
1090 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1091 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1092 	ib->ptr[ib->length_dw++] = 0;
1093 
1094 }
1095 
1096 /**
1097  * sdma_v7_0_vm_write_pte - update PTEs by writing them manually
1098  *
1099  * @ib: indirect buffer to fill with commands
1100  * @pe: addr of the page entry
1101  * @value: dst addr to write into pe
1102  * @count: number of page entries to update
1103  * @incr: increase next addr by incr bytes
1104  *
1105  * Update PTEs by writing them manually using sDMA.
1106  */
1107 static void sdma_v7_0_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
1108 				   uint64_t value, unsigned count,
1109 				   uint32_t incr)
1110 {
1111 	unsigned ndw = count * 2;
1112 
1113 	ib->ptr[ib->length_dw++] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_WRITE) |
1114 		SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1115 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1116 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1117 	ib->ptr[ib->length_dw++] = ndw - 1;
1118 	for (; ndw > 0; ndw -= 2) {
1119 		ib->ptr[ib->length_dw++] = lower_32_bits(value);
1120 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
1121 		value += incr;
1122 	}
1123 }
1124 
1125 /**
1126  * sdma_v7_0_vm_set_pte_pde - update the page tables using sDMA
1127  *
1128  * @ib: indirect buffer to fill with commands
1129  * @pe: addr of the page entry
1130  * @addr: dst addr to write into pe
1131  * @count: number of page entries to update
1132  * @incr: increase next addr by incr bytes
1133  * @flags: access flags
1134  *
1135  * Update the page tables using sDMA.
1136  */
1137 static void sdma_v7_0_vm_set_pte_pde(struct amdgpu_ib *ib,
1138 				     uint64_t pe,
1139 				     uint64_t addr, unsigned count,
1140 				     uint32_t incr, uint64_t flags)
1141 {
1142 	/* for physically contiguous pages (vram) */
1143 	ib->ptr[ib->length_dw++] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_PTEPDE);
1144 	ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
1145 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1146 	ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
1147 	ib->ptr[ib->length_dw++] = upper_32_bits(flags);
1148 	ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
1149 	ib->ptr[ib->length_dw++] = upper_32_bits(addr);
1150 	ib->ptr[ib->length_dw++] = incr; /* increment size */
1151 	ib->ptr[ib->length_dw++] = 0;
1152 	ib->ptr[ib->length_dw++] = count - 1; /* number of entries */
1153 }
1154 
1155 /**
1156  * sdma_v7_0_ring_pad_ib - pad the IB
1157  *
1158  * @ring: amdgpu ring pointer
1159  * @ib: indirect buffer to fill with padding
1160  *
1161  * Pad the IB with NOPs to a boundary multiple of 8.
1162  */
1163 static void sdma_v7_0_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
1164 {
1165 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
1166 	u32 pad_count;
1167 	int i;
1168 
1169 	pad_count = (-ib->length_dw) & 0x7;
1170 	for (i = 0; i < pad_count; i++)
1171 		if (sdma && sdma->burst_nop && (i == 0))
1172 			ib->ptr[ib->length_dw++] =
1173 				SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_NOP) |
1174 				SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
1175 		else
1176 			ib->ptr[ib->length_dw++] =
1177 				SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_NOP);
1178 }
1179 
1180 /**
1181  * sdma_v7_0_ring_emit_pipeline_sync - sync the pipeline
1182  *
1183  * @ring: amdgpu_ring pointer
1184  *
1185  * Make sure all previous operations are completed (CIK).
1186  */
1187 static void sdma_v7_0_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
1188 {
1189 	uint32_t seq = ring->fence_drv.sync_seq;
1190 	uint64_t addr = ring->fence_drv.gpu_addr;
1191 
1192 	/* wait for idle */
1193 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1194 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1195 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
1196 			  SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1));
1197 	amdgpu_ring_write(ring, addr & 0xfffffffc);
1198 	amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
1199 	amdgpu_ring_write(ring, seq); /* reference */
1200 	amdgpu_ring_write(ring, 0xffffffff); /* mask */
1201 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1202 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */
1203 }
1204 
1205 /**
1206  * sdma_v7_0_ring_emit_vm_flush - vm flush using sDMA
1207  *
1208  * @ring: amdgpu_ring pointer
1209  * @vmid: vmid number to use
1210  * @pd_addr: address
1211  *
1212  * Update the page table base and flush the VM TLB
1213  * using sDMA.
1214  */
1215 static void sdma_v7_0_ring_emit_vm_flush(struct amdgpu_ring *ring,
1216 					 unsigned vmid, uint64_t pd_addr)
1217 {
1218 	amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
1219 }
1220 
1221 static void sdma_v7_0_ring_emit_wreg(struct amdgpu_ring *ring,
1222 				     uint32_t reg, uint32_t val)
1223 {
1224 	/* SRBM WRITE command will not support on sdma v7.
1225 	 * Use Register WRITE command instead, which OPCODE is same as SRBM WRITE
1226 	 */
1227 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_SRBM_WRITE));
1228 	amdgpu_ring_write(ring, reg << 2);
1229 	amdgpu_ring_write(ring, val);
1230 }
1231 
1232 static void sdma_v7_0_ring_emit_reg_wait(struct amdgpu_ring *ring, uint32_t reg,
1233 					 uint32_t val, uint32_t mask)
1234 {
1235 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1236 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1237 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* equal */
1238 	amdgpu_ring_write(ring, reg << 2);
1239 	amdgpu_ring_write(ring, 0);
1240 	amdgpu_ring_write(ring, val); /* reference */
1241 	amdgpu_ring_write(ring, mask); /* mask */
1242 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1243 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10));
1244 }
1245 
1246 static void sdma_v7_0_ring_emit_reg_write_reg_wait(struct amdgpu_ring *ring,
1247 						   uint32_t reg0, uint32_t reg1,
1248 						   uint32_t ref, uint32_t mask)
1249 {
1250 	amdgpu_ring_emit_wreg(ring, reg0, ref);
1251 	/* wait for a cycle to reset vm_inv_eng*_ack */
1252 	amdgpu_ring_emit_reg_wait(ring, reg0, 0, 0);
1253 	amdgpu_ring_emit_reg_wait(ring, reg1, mask, mask);
1254 }
1255 
1256 static int sdma_v7_0_early_init(void *handle)
1257 {
1258 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1259 	int r;
1260 
1261 	r = amdgpu_sdma_init_microcode(adev, 0, true);
1262 	if (r) {
1263 		DRM_ERROR("Failed to init sdma firmware!\n");
1264 		return r;
1265 	}
1266 
1267 	sdma_v7_0_set_ring_funcs(adev);
1268 	sdma_v7_0_set_buffer_funcs(adev);
1269 	sdma_v7_0_set_vm_pte_funcs(adev);
1270 	sdma_v7_0_set_irq_funcs(adev);
1271 	sdma_v7_0_set_mqd_funcs(adev);
1272 
1273 	return 0;
1274 }
1275 
1276 static int sdma_v7_0_sw_init(void *handle)
1277 {
1278 	struct amdgpu_ring *ring;
1279 	int r, i;
1280 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1281 	uint32_t reg_count = ARRAY_SIZE(sdma_reg_list_7_0);
1282 	uint32_t *ptr;
1283 
1284 	/* SDMA trap event */
1285 	r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_GFX,
1286 			      GFX_11_0_0__SRCID__SDMA_TRAP,
1287 			      &adev->sdma.trap_irq);
1288 	if (r)
1289 		return r;
1290 
1291 	for (i = 0; i < adev->sdma.num_instances; i++) {
1292 		ring = &adev->sdma.instance[i].ring;
1293 		ring->ring_obj = NULL;
1294 		ring->use_doorbell = true;
1295 		ring->me = i;
1296 
1297 		DRM_DEBUG("SDMA %d use_doorbell being set to: [%s]\n", i,
1298 				ring->use_doorbell?"true":"false");
1299 
1300 		ring->doorbell_index =
1301 			(adev->doorbell_index.sdma_engine[i] << 1); // get DWORD offset
1302 
1303 		ring->vm_hub = AMDGPU_GFXHUB(0);
1304 		sprintf(ring->name, "sdma%d", i);
1305 		r = amdgpu_ring_init(adev, ring, 1024,
1306 				     &adev->sdma.trap_irq,
1307 				     AMDGPU_SDMA_IRQ_INSTANCE0 + i,
1308 				     AMDGPU_RING_PRIO_DEFAULT, NULL);
1309 		if (r)
1310 			return r;
1311 	}
1312 
1313 	/* Allocate memory for SDMA IP Dump buffer */
1314 	ptr = kcalloc(adev->sdma.num_instances * reg_count, sizeof(uint32_t), GFP_KERNEL);
1315 	if (ptr)
1316 		adev->sdma.ip_dump = ptr;
1317 	else
1318 		DRM_ERROR("Failed to allocated memory for SDMA IP Dump\n");
1319 
1320 	return r;
1321 }
1322 
1323 static int sdma_v7_0_sw_fini(void *handle)
1324 {
1325 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1326 	int i;
1327 
1328 	for (i = 0; i < adev->sdma.num_instances; i++)
1329 		amdgpu_ring_fini(&adev->sdma.instance[i].ring);
1330 
1331 	amdgpu_sdma_destroy_inst_ctx(adev, true);
1332 
1333 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT)
1334 		sdma_v12_0_free_ucode_buffer(adev);
1335 
1336 	kfree(adev->sdma.ip_dump);
1337 
1338 	return 0;
1339 }
1340 
1341 static int sdma_v7_0_hw_init(void *handle)
1342 {
1343 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1344 
1345 	return sdma_v7_0_start(adev);
1346 }
1347 
1348 static int sdma_v7_0_hw_fini(void *handle)
1349 {
1350 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1351 
1352 	if (amdgpu_sriov_vf(adev))
1353 		return 0;
1354 
1355 	sdma_v7_0_ctx_switch_enable(adev, false);
1356 	sdma_v7_0_enable(adev, false);
1357 
1358 	return 0;
1359 }
1360 
1361 static int sdma_v7_0_suspend(void *handle)
1362 {
1363 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1364 
1365 	return sdma_v7_0_hw_fini(adev);
1366 }
1367 
1368 static int sdma_v7_0_resume(void *handle)
1369 {
1370 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1371 
1372 	return sdma_v7_0_hw_init(adev);
1373 }
1374 
1375 static bool sdma_v7_0_is_idle(void *handle)
1376 {
1377 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1378 	u32 i;
1379 
1380 	for (i = 0; i < adev->sdma.num_instances; i++) {
1381 		u32 tmp = RREG32(sdma_v7_0_get_reg_offset(adev, i, regSDMA0_STATUS_REG));
1382 
1383 		if (!(tmp & SDMA0_STATUS_REG__IDLE_MASK))
1384 			return false;
1385 	}
1386 
1387 	return true;
1388 }
1389 
1390 static int sdma_v7_0_wait_for_idle(void *handle)
1391 {
1392 	unsigned i;
1393 	u32 sdma0, sdma1;
1394 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1395 
1396 	for (i = 0; i < adev->usec_timeout; i++) {
1397 		sdma0 = RREG32(sdma_v7_0_get_reg_offset(adev, 0, regSDMA0_STATUS_REG));
1398 		sdma1 = RREG32(sdma_v7_0_get_reg_offset(adev, 1, regSDMA0_STATUS_REG));
1399 
1400 		if (sdma0 & sdma1 & SDMA0_STATUS_REG__IDLE_MASK)
1401 			return 0;
1402 		udelay(1);
1403 	}
1404 	return -ETIMEDOUT;
1405 }
1406 
1407 static int sdma_v7_0_ring_preempt_ib(struct amdgpu_ring *ring)
1408 {
1409 	int i, r = 0;
1410 	struct amdgpu_device *adev = ring->adev;
1411 	u32 index = 0;
1412 	u64 sdma_gfx_preempt;
1413 
1414 	amdgpu_sdma_get_index_from_ring(ring, &index);
1415 	sdma_gfx_preempt =
1416 		sdma_v7_0_get_reg_offset(adev, index, regSDMA0_QUEUE0_PREEMPT);
1417 
1418 	/* assert preemption condition */
1419 	amdgpu_ring_set_preempt_cond_exec(ring, false);
1420 
1421 	/* emit the trailing fence */
1422 	ring->trail_seq += 1;
1423 	r = amdgpu_ring_alloc(ring, 10);
1424 	if (r) {
1425 		DRM_ERROR("ring %d failed to be allocated \n", ring->idx);
1426 		return r;
1427 	}
1428 	sdma_v7_0_ring_emit_fence(ring, ring->trail_fence_gpu_addr,
1429 				  ring->trail_seq, 0);
1430 	amdgpu_ring_commit(ring);
1431 
1432 	/* assert IB preemption */
1433 	WREG32(sdma_gfx_preempt, 1);
1434 
1435 	/* poll the trailing fence */
1436 	for (i = 0; i < adev->usec_timeout; i++) {
1437 		if (ring->trail_seq ==
1438 		    le32_to_cpu(*(ring->trail_fence_cpu_addr)))
1439 			break;
1440 		udelay(1);
1441 	}
1442 
1443 	if (i >= adev->usec_timeout) {
1444 		r = -EINVAL;
1445 		DRM_ERROR("ring %d failed to be preempted\n", ring->idx);
1446 	}
1447 
1448 	/* deassert IB preemption */
1449 	WREG32(sdma_gfx_preempt, 0);
1450 
1451 	/* deassert the preemption condition */
1452 	amdgpu_ring_set_preempt_cond_exec(ring, true);
1453 	return r;
1454 }
1455 
1456 static int sdma_v7_0_set_trap_irq_state(struct amdgpu_device *adev,
1457 					struct amdgpu_irq_src *source,
1458 					unsigned type,
1459 					enum amdgpu_interrupt_state state)
1460 {
1461 	u32 sdma_cntl;
1462 
1463 	u32 reg_offset = sdma_v7_0_get_reg_offset(adev, type, regSDMA0_CNTL);
1464 
1465 	sdma_cntl = RREG32(reg_offset);
1466 	sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE,
1467 		       state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0);
1468 	WREG32(reg_offset, sdma_cntl);
1469 
1470 	return 0;
1471 }
1472 
1473 static int sdma_v7_0_process_trap_irq(struct amdgpu_device *adev,
1474 				      struct amdgpu_irq_src *source,
1475 				      struct amdgpu_iv_entry *entry)
1476 {
1477 	int instances, queue;
1478 	uint32_t mes_queue_id = entry->src_data[0];
1479 
1480 	DRM_DEBUG("IH: SDMA trap\n");
1481 
1482 	if (adev->enable_mes && (mes_queue_id & AMDGPU_FENCE_MES_QUEUE_FLAG)) {
1483 		struct amdgpu_mes_queue *queue;
1484 
1485 		mes_queue_id &= AMDGPU_FENCE_MES_QUEUE_ID_MASK;
1486 
1487 		spin_lock(&adev->mes.queue_id_lock);
1488 		queue = idr_find(&adev->mes.queue_id_idr, mes_queue_id);
1489 		if (queue) {
1490 			DRM_DEBUG("process smda queue id = %d\n", mes_queue_id);
1491 			amdgpu_fence_process(queue->ring);
1492 		}
1493 		spin_unlock(&adev->mes.queue_id_lock);
1494 		return 0;
1495 	}
1496 
1497 	queue = entry->ring_id & 0xf;
1498 	instances = (entry->ring_id & 0xf0) >> 4;
1499 	if (instances > 1) {
1500 		DRM_ERROR("IH: wrong ring_ID detected, as wrong sdma instance\n");
1501 		return -EINVAL;
1502 	}
1503 
1504 	switch (entry->client_id) {
1505 	case SOC21_IH_CLIENTID_GFX:
1506 		switch (queue) {
1507 		case 0:
1508 			amdgpu_fence_process(&adev->sdma.instance[instances].ring);
1509 			break;
1510 		default:
1511 			break;
1512 		}
1513 		break;
1514 	}
1515 	return 0;
1516 }
1517 
1518 static int sdma_v7_0_process_illegal_inst_irq(struct amdgpu_device *adev,
1519 					      struct amdgpu_irq_src *source,
1520 					      struct amdgpu_iv_entry *entry)
1521 {
1522 	return 0;
1523 }
1524 
1525 static int sdma_v7_0_set_clockgating_state(void *handle,
1526 					   enum amd_clockgating_state state)
1527 {
1528 	return 0;
1529 }
1530 
1531 static int sdma_v7_0_set_powergating_state(void *handle,
1532 					  enum amd_powergating_state state)
1533 {
1534 	return 0;
1535 }
1536 
1537 static void sdma_v7_0_get_clockgating_state(void *handle, u64 *flags)
1538 {
1539 }
1540 
1541 static void sdma_v7_0_print_ip_state(void *handle, struct drm_printer *p)
1542 {
1543 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1544 	int i, j;
1545 	uint32_t reg_count = ARRAY_SIZE(sdma_reg_list_7_0);
1546 	uint32_t instance_offset;
1547 
1548 	if (!adev->sdma.ip_dump)
1549 		return;
1550 
1551 	drm_printf(p, "num_instances:%d\n", adev->sdma.num_instances);
1552 	for (i = 0; i < adev->sdma.num_instances; i++) {
1553 		instance_offset = i * reg_count;
1554 		drm_printf(p, "\nInstance:%d\n", i);
1555 
1556 		for (j = 0; j < reg_count; j++)
1557 			drm_printf(p, "%-50s \t 0x%08x\n", sdma_reg_list_7_0[j].reg_name,
1558 				   adev->sdma.ip_dump[instance_offset + j]);
1559 	}
1560 }
1561 
1562 static void sdma_v7_0_dump_ip_state(void *handle)
1563 {
1564 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1565 	int i, j;
1566 	uint32_t instance_offset;
1567 	uint32_t reg_count = ARRAY_SIZE(sdma_reg_list_7_0);
1568 
1569 	if (!adev->sdma.ip_dump)
1570 		return;
1571 
1572 	amdgpu_gfx_off_ctrl(adev, false);
1573 	for (i = 0; i < adev->sdma.num_instances; i++) {
1574 		instance_offset = i * reg_count;
1575 		for (j = 0; j < reg_count; j++)
1576 			adev->sdma.ip_dump[instance_offset + j] =
1577 				RREG32(sdma_v7_0_get_reg_offset(adev, i,
1578 				       sdma_reg_list_7_0[j].reg_offset));
1579 	}
1580 	amdgpu_gfx_off_ctrl(adev, true);
1581 }
1582 
1583 const struct amd_ip_funcs sdma_v7_0_ip_funcs = {
1584 	.name = "sdma_v7_0",
1585 	.early_init = sdma_v7_0_early_init,
1586 	.late_init = NULL,
1587 	.sw_init = sdma_v7_0_sw_init,
1588 	.sw_fini = sdma_v7_0_sw_fini,
1589 	.hw_init = sdma_v7_0_hw_init,
1590 	.hw_fini = sdma_v7_0_hw_fini,
1591 	.suspend = sdma_v7_0_suspend,
1592 	.resume = sdma_v7_0_resume,
1593 	.is_idle = sdma_v7_0_is_idle,
1594 	.wait_for_idle = sdma_v7_0_wait_for_idle,
1595 	.soft_reset = sdma_v7_0_soft_reset,
1596 	.check_soft_reset = sdma_v7_0_check_soft_reset,
1597 	.set_clockgating_state = sdma_v7_0_set_clockgating_state,
1598 	.set_powergating_state = sdma_v7_0_set_powergating_state,
1599 	.get_clockgating_state = sdma_v7_0_get_clockgating_state,
1600 	.dump_ip_state = sdma_v7_0_dump_ip_state,
1601 	.print_ip_state = sdma_v7_0_print_ip_state,
1602 };
1603 
1604 static const struct amdgpu_ring_funcs sdma_v7_0_ring_funcs = {
1605 	.type = AMDGPU_RING_TYPE_SDMA,
1606 	.align_mask = 0xf,
1607 	.nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP),
1608 	.support_64bit_ptrs = true,
1609 	.secure_submission_supported = true,
1610 	.get_rptr = sdma_v7_0_ring_get_rptr,
1611 	.get_wptr = sdma_v7_0_ring_get_wptr,
1612 	.set_wptr = sdma_v7_0_ring_set_wptr,
1613 	.emit_frame_size =
1614 		5 + /* sdma_v7_0_ring_init_cond_exec */
1615 		6 + /* sdma_v7_0_ring_emit_hdp_flush */
1616 		6 + /* sdma_v7_0_ring_emit_pipeline_sync */
1617 		/* sdma_v7_0_ring_emit_vm_flush */
1618 		SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 +
1619 		SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6 +
1620 		10 + 10 + 10, /* sdma_v7_0_ring_emit_fence x3 for user fence, vm fence */
1621 	.emit_ib_size = 5 + 7 + 6, /* sdma_v7_0_ring_emit_ib */
1622 	.emit_ib = sdma_v7_0_ring_emit_ib,
1623 	.emit_mem_sync = sdma_v7_0_ring_emit_mem_sync,
1624 	.emit_fence = sdma_v7_0_ring_emit_fence,
1625 	.emit_pipeline_sync = sdma_v7_0_ring_emit_pipeline_sync,
1626 	.emit_vm_flush = sdma_v7_0_ring_emit_vm_flush,
1627 	.emit_hdp_flush = sdma_v7_0_ring_emit_hdp_flush,
1628 	.test_ring = sdma_v7_0_ring_test_ring,
1629 	.test_ib = sdma_v7_0_ring_test_ib,
1630 	.insert_nop = sdma_v7_0_ring_insert_nop,
1631 	.pad_ib = sdma_v7_0_ring_pad_ib,
1632 	.emit_wreg = sdma_v7_0_ring_emit_wreg,
1633 	.emit_reg_wait = sdma_v7_0_ring_emit_reg_wait,
1634 	.emit_reg_write_reg_wait = sdma_v7_0_ring_emit_reg_write_reg_wait,
1635 	.init_cond_exec = sdma_v7_0_ring_init_cond_exec,
1636 	.preempt_ib = sdma_v7_0_ring_preempt_ib,
1637 };
1638 
1639 static void sdma_v7_0_set_ring_funcs(struct amdgpu_device *adev)
1640 {
1641 	int i;
1642 
1643 	for (i = 0; i < adev->sdma.num_instances; i++) {
1644 		adev->sdma.instance[i].ring.funcs = &sdma_v7_0_ring_funcs;
1645 		adev->sdma.instance[i].ring.me = i;
1646 	}
1647 }
1648 
1649 static const struct amdgpu_irq_src_funcs sdma_v7_0_trap_irq_funcs = {
1650 	.set = sdma_v7_0_set_trap_irq_state,
1651 	.process = sdma_v7_0_process_trap_irq,
1652 };
1653 
1654 static const struct amdgpu_irq_src_funcs sdma_v7_0_illegal_inst_irq_funcs = {
1655 	.process = sdma_v7_0_process_illegal_inst_irq,
1656 };
1657 
1658 static void sdma_v7_0_set_irq_funcs(struct amdgpu_device *adev)
1659 {
1660 	adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_INSTANCE0 +
1661 					adev->sdma.num_instances;
1662 	adev->sdma.trap_irq.funcs = &sdma_v7_0_trap_irq_funcs;
1663 	adev->sdma.illegal_inst_irq.funcs = &sdma_v7_0_illegal_inst_irq_funcs;
1664 }
1665 
1666 /**
1667  * sdma_v7_0_emit_copy_buffer - copy buffer using the sDMA engine
1668  *
1669  * @ib: indirect buffer to fill with commands
1670  * @src_offset: src GPU address
1671  * @dst_offset: dst GPU address
1672  * @byte_count: number of bytes to xfer
1673  * @copy_flags: copy flags for the buffers
1674  *
1675  * Copy GPU buffers using the DMA engine.
1676  * Used by the amdgpu ttm implementation to move pages if
1677  * registered as the asic copy callback.
1678  */
1679 static void sdma_v7_0_emit_copy_buffer(struct amdgpu_ib *ib,
1680 				       uint64_t src_offset,
1681 				       uint64_t dst_offset,
1682 				       uint32_t byte_count,
1683 				       uint32_t copy_flags)
1684 {
1685 	uint32_t num_type, data_format, max_com;
1686 
1687 	max_com = AMDGPU_COPY_FLAGS_GET(copy_flags, MAX_COMPRESSED);
1688 	data_format = AMDGPU_COPY_FLAGS_GET(copy_flags, DATA_FORMAT);
1689 	num_type = AMDGPU_COPY_FLAGS_GET(copy_flags, NUMBER_TYPE);
1690 
1691 	ib->ptr[ib->length_dw++] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_COPY) |
1692 		SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR) |
1693 		SDMA_PKT_COPY_LINEAR_HEADER_TMZ((copy_flags & AMDGPU_COPY_FLAGS_TMZ) ? 1 : 0) |
1694 		SDMA_PKT_COPY_LINEAR_HEADER_CPV(1);
1695 
1696 	ib->ptr[ib->length_dw++] = byte_count - 1;
1697 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1698 	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1699 	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1700 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1701 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1702 
1703 	if ((copy_flags & (AMDGPU_COPY_FLAGS_READ_DECOMPRESSED | AMDGPU_COPY_FLAGS_WRITE_COMPRESSED)))
1704 		ib->ptr[ib->length_dw++] = SDMA_DCC_DATA_FORMAT(data_format) | SDMA_DCC_NUM_TYPE(num_type) |
1705 			((copy_flags & AMDGPU_COPY_FLAGS_READ_DECOMPRESSED) ? SDMA_DCC_READ_CM(2) : 0) |
1706 			((copy_flags & AMDGPU_COPY_FLAGS_WRITE_COMPRESSED) ? SDMA_DCC_WRITE_CM(1) : 0) |
1707 			SDMA_DCC_MAX_COM(max_com) | SDMA_DCC_MAX_UCOM(1);
1708 	else
1709 		ib->ptr[ib->length_dw++] = 0;
1710 }
1711 
1712 /**
1713  * sdma_v7_0_emit_fill_buffer - fill buffer using the sDMA engine
1714  *
1715  * @ib: indirect buffer to fill
1716  * @src_data: value to write to buffer
1717  * @dst_offset: dst GPU address
1718  * @byte_count: number of bytes to xfer
1719  *
1720  * Fill GPU buffers using the DMA engine.
1721  */
1722 static void sdma_v7_0_emit_fill_buffer(struct amdgpu_ib *ib,
1723 				       uint32_t src_data,
1724 				       uint64_t dst_offset,
1725 				       uint32_t byte_count)
1726 {
1727 	ib->ptr[ib->length_dw++] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_CONST_FILL);
1728 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1729 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1730 	ib->ptr[ib->length_dw++] = src_data;
1731 	ib->ptr[ib->length_dw++] = byte_count - 1;
1732 }
1733 
1734 static const struct amdgpu_buffer_funcs sdma_v7_0_buffer_funcs = {
1735 	.copy_max_bytes = 0x400000,
1736 	.copy_num_dw = 8,
1737 	.emit_copy_buffer = sdma_v7_0_emit_copy_buffer,
1738 	.fill_max_bytes = 0x400000,
1739 	.fill_num_dw = 5,
1740 	.emit_fill_buffer = sdma_v7_0_emit_fill_buffer,
1741 };
1742 
1743 static void sdma_v7_0_set_buffer_funcs(struct amdgpu_device *adev)
1744 {
1745 	adev->mman.buffer_funcs = &sdma_v7_0_buffer_funcs;
1746 	adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1747 }
1748 
1749 static const struct amdgpu_vm_pte_funcs sdma_v7_0_vm_pte_funcs = {
1750 	.copy_pte_num_dw = 8,
1751 	.copy_pte = sdma_v7_0_vm_copy_pte,
1752 	.write_pte = sdma_v7_0_vm_write_pte,
1753 	.set_pte_pde = sdma_v7_0_vm_set_pte_pde,
1754 };
1755 
1756 static void sdma_v7_0_set_vm_pte_funcs(struct amdgpu_device *adev)
1757 {
1758 	unsigned i;
1759 
1760 	adev->vm_manager.vm_pte_funcs = &sdma_v7_0_vm_pte_funcs;
1761 	for (i = 0; i < adev->sdma.num_instances; i++) {
1762 		adev->vm_manager.vm_pte_scheds[i] =
1763 			&adev->sdma.instance[i].ring.sched;
1764 	}
1765 	adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
1766 }
1767 
1768 const struct amdgpu_ip_block_version sdma_v7_0_ip_block = {
1769 	.type = AMD_IP_BLOCK_TYPE_SDMA,
1770 	.major = 7,
1771 	.minor = 0,
1772 	.rev = 0,
1773 	.funcs = &sdma_v7_0_ip_funcs,
1774 };
1775