xref: /linux/drivers/gpu/drm/amd/amdgpu/sdma_v7_0.c (revision 7f71507851fc7764b36a3221839607d3a45c2025)
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 /*define for compression field for sdma7*/
55 #define SDMA_PKT_CONSTANT_FILL_HEADER_compress_offset 0
56 #define SDMA_PKT_CONSTANT_FILL_HEADER_compress_mask   0x00000001
57 #define SDMA_PKT_CONSTANT_FILL_HEADER_compress_shift  16
58 #define SDMA_PKT_CONSTANT_FILL_HEADER_COMPRESS(x) (((x) & SDMA_PKT_CONSTANT_FILL_HEADER_compress_mask) << SDMA_PKT_CONSTANT_FILL_HEADER_compress_shift)
59 
60 static const struct amdgpu_hwip_reg_entry sdma_reg_list_7_0[] = {
61 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_STATUS_REG),
62 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_STATUS1_REG),
63 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_STATUS2_REG),
64 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_STATUS3_REG),
65 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_STATUS4_REG),
66 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_STATUS5_REG),
67 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_STATUS6_REG),
68 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_UCODE_REV),
69 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_RB_RPTR_FETCH_HI),
70 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_RB_RPTR_FETCH),
71 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_UTCL1_RD_STATUS),
72 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_UTCL1_WR_STATUS),
73 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_UTCL1_RD_XNACK0),
74 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_UTCL1_RD_XNACK1),
75 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_UTCL1_WR_XNACK0),
76 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_UTCL1_WR_XNACK1),
77 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE0_RB_CNTL),
78 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE0_RB_RPTR),
79 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE0_RB_RPTR_HI),
80 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE0_RB_WPTR),
81 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE0_RB_WPTR_HI),
82 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE0_IB_OFFSET),
83 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE0_IB_BASE_LO),
84 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE0_IB_BASE_HI),
85 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE0_IB_CNTL),
86 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE0_IB_RPTR),
87 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE0_IB_SUB_REMAIN),
88 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE0_DUMMY_REG),
89 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE_STATUS0),
90 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE1_RB_CNTL),
91 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE1_RB_RPTR),
92 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE1_RB_RPTR_HI),
93 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE1_RB_WPTR),
94 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE1_RB_WPTR_HI),
95 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE1_IB_OFFSET),
96 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE1_IB_BASE_LO),
97 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE1_IB_BASE_HI),
98 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE1_IB_RPTR),
99 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE1_IB_SUB_REMAIN),
100 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE1_DUMMY_REG),
101 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE2_RB_CNTL),
102 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE2_RB_RPTR),
103 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE2_RB_RPTR_HI),
104 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE2_RB_WPTR),
105 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE2_RB_WPTR_HI),
106 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE2_IB_OFFSET),
107 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE2_IB_BASE_LO),
108 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE2_IB_BASE_HI),
109 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE2_IB_RPTR),
110 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE2_IB_SUB_REMAIN),
111 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_QUEUE2_DUMMY_REG),
112 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_INT_STATUS),
113 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_VM_CNTL),
114 	SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS2),
115 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_CHICKEN_BITS),
116 };
117 
118 static void sdma_v7_0_set_ring_funcs(struct amdgpu_device *adev);
119 static void sdma_v7_0_set_buffer_funcs(struct amdgpu_device *adev);
120 static void sdma_v7_0_set_vm_pte_funcs(struct amdgpu_device *adev);
121 static void sdma_v7_0_set_irq_funcs(struct amdgpu_device *adev);
122 static int sdma_v7_0_start(struct amdgpu_device *adev);
123 
124 static u32 sdma_v7_0_get_reg_offset(struct amdgpu_device *adev, u32 instance, u32 internal_offset)
125 {
126 	u32 base;
127 
128 	if (internal_offset >= SDMA0_HYP_DEC_REG_START &&
129 	    internal_offset <= SDMA0_HYP_DEC_REG_END) {
130 		base = adev->reg_offset[GC_HWIP][0][1];
131 		if (instance != 0)
132 			internal_offset += SDMA1_HYP_DEC_REG_OFFSET * instance;
133 	} else {
134 		base = adev->reg_offset[GC_HWIP][0][0];
135 		if (instance == 1)
136 			internal_offset += SDMA1_REG_OFFSET;
137 	}
138 
139 	return base + internal_offset;
140 }
141 
142 static unsigned sdma_v7_0_ring_init_cond_exec(struct amdgpu_ring *ring,
143 					      uint64_t addr)
144 {
145 	unsigned ret;
146 
147 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_COND_EXE));
148 	amdgpu_ring_write(ring, lower_32_bits(addr));
149 	amdgpu_ring_write(ring, upper_32_bits(addr));
150 	amdgpu_ring_write(ring, 1);
151 	/* this is the offset we need patch later */
152 	ret = ring->wptr & ring->buf_mask;
153 	/* insert dummy here and patch it later */
154 	amdgpu_ring_write(ring, 0);
155 
156 	return ret;
157 }
158 
159 /**
160  * sdma_v7_0_ring_get_rptr - get the current read pointer
161  *
162  * @ring: amdgpu ring pointer
163  *
164  * Get the current rptr from the hardware.
165  */
166 static uint64_t sdma_v7_0_ring_get_rptr(struct amdgpu_ring *ring)
167 {
168 	u64 *rptr;
169 
170 	/* XXX check if swapping is necessary on BE */
171 	rptr = (u64 *)ring->rptr_cpu_addr;
172 
173 	DRM_DEBUG("rptr before shift == 0x%016llx\n", *rptr);
174 	return ((*rptr) >> 2);
175 }
176 
177 /**
178  * sdma_v7_0_ring_get_wptr - get the current write pointer
179  *
180  * @ring: amdgpu ring pointer
181  *
182  * Get the current wptr from the hardware.
183  */
184 static uint64_t sdma_v7_0_ring_get_wptr(struct amdgpu_ring *ring)
185 {
186 	u64 wptr = 0;
187 
188 	if (ring->use_doorbell) {
189 		/* XXX check if swapping is necessary on BE */
190 		wptr = READ_ONCE(*((u64 *)ring->wptr_cpu_addr));
191 		DRM_DEBUG("wptr/doorbell before shift == 0x%016llx\n", wptr);
192 	}
193 
194 	return wptr >> 2;
195 }
196 
197 /**
198  * sdma_v7_0_ring_set_wptr - commit the write pointer
199  *
200  * @ring: amdgpu ring pointer
201  *
202  * Write the wptr back to the hardware.
203  */
204 static void sdma_v7_0_ring_set_wptr(struct amdgpu_ring *ring)
205 {
206 	struct amdgpu_device *adev = ring->adev;
207 	uint32_t *wptr_saved;
208 	uint32_t *is_queue_unmap;
209 	uint64_t aggregated_db_index;
210 	uint32_t mqd_size = adev->mqds[AMDGPU_HW_IP_DMA].mqd_size;
211 
212 	DRM_DEBUG("Setting write pointer\n");
213 
214 	if (ring->is_mes_queue) {
215 		wptr_saved = (uint32_t *)(ring->mqd_ptr + mqd_size);
216 		is_queue_unmap = (uint32_t *)(ring->mqd_ptr + mqd_size +
217 					      sizeof(uint32_t));
218 		aggregated_db_index =
219 			amdgpu_mes_get_aggregated_doorbell_index(adev,
220 							 ring->hw_prio);
221 
222 		atomic64_set((atomic64_t *)ring->wptr_cpu_addr,
223 			     ring->wptr << 2);
224 		*wptr_saved = ring->wptr << 2;
225 		if (*is_queue_unmap) {
226 			WDOORBELL64(aggregated_db_index, ring->wptr << 2);
227 			DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016llx)\n",
228 					ring->doorbell_index, ring->wptr << 2);
229 			WDOORBELL64(ring->doorbell_index, ring->wptr << 2);
230 		} else {
231 			DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016llx)\n",
232 					ring->doorbell_index, ring->wptr << 2);
233 			WDOORBELL64(ring->doorbell_index, ring->wptr << 2);
234 		}
235 	} else {
236 		if (ring->use_doorbell) {
237 			DRM_DEBUG("Using doorbell -- "
238 				  "wptr_offs == 0x%08x "
239 				  "lower_32_bits(ring->wptr) << 2 == 0x%08x "
240 				  "upper_32_bits(ring->wptr) << 2 == 0x%08x\n",
241 				  ring->wptr_offs,
242 				  lower_32_bits(ring->wptr << 2),
243 				  upper_32_bits(ring->wptr << 2));
244 			/* XXX check if swapping is necessary on BE */
245 			atomic64_set((atomic64_t *)ring->wptr_cpu_addr,
246 				     ring->wptr << 2);
247 			DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016llx)\n",
248 				  ring->doorbell_index, ring->wptr << 2);
249 			WDOORBELL64(ring->doorbell_index, ring->wptr << 2);
250 		} else {
251 			DRM_DEBUG("Not using doorbell -- "
252 				  "regSDMA%i_GFX_RB_WPTR == 0x%08x "
253 				  "regSDMA%i_GFX_RB_WPTR_HI == 0x%08x\n",
254 				  ring->me,
255 				  lower_32_bits(ring->wptr << 2),
256 				  ring->me,
257 				  upper_32_bits(ring->wptr << 2));
258 			WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev,
259 								     ring->me,
260 								     regSDMA0_QUEUE0_RB_WPTR),
261 					lower_32_bits(ring->wptr << 2));
262 			WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev,
263 								     ring->me,
264 								     regSDMA0_QUEUE0_RB_WPTR_HI),
265 					upper_32_bits(ring->wptr << 2));
266 		}
267 	}
268 }
269 
270 static void sdma_v7_0_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
271 {
272 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
273 	int i;
274 
275 	for (i = 0; i < count; i++)
276 		if (sdma && sdma->burst_nop && (i == 0))
277 			amdgpu_ring_write(ring, ring->funcs->nop |
278 				SDMA_PKT_NOP_HEADER_COUNT(count - 1));
279 		else
280 			amdgpu_ring_write(ring, ring->funcs->nop);
281 }
282 
283 /**
284  * sdma_v7_0_ring_emit_ib - Schedule an IB on the DMA engine
285  *
286  * @ring: amdgpu ring pointer
287  * @job: job to retrieve vmid from
288  * @ib: IB object to schedule
289  * @flags: unused
290  *
291  * Schedule an IB in the DMA ring.
292  */
293 static void sdma_v7_0_ring_emit_ib(struct amdgpu_ring *ring,
294 				   struct amdgpu_job *job,
295 				   struct amdgpu_ib *ib,
296 				   uint32_t flags)
297 {
298 	unsigned vmid = AMDGPU_JOB_GET_VMID(job);
299 	uint64_t csa_mc_addr = amdgpu_sdma_get_csa_mc_addr(ring, vmid);
300 
301 	/* An IB packet must end on a 8 DW boundary--the next dword
302 	 * must be on a 8-dword boundary. Our IB packet below is 6
303 	 * dwords long, thus add x number of NOPs, such that, in
304 	 * modular arithmetic,
305 	 * wptr + 6 + x = 8k, k >= 0, which in C is,
306 	 * (wptr + 6 + x) % 8 = 0.
307 	 * The expression below, is a solution of x.
308 	 */
309 	sdma_v7_0_ring_insert_nop(ring, (2 - lower_32_bits(ring->wptr)) & 7);
310 
311 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_INDIRECT) |
312 			  SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf));
313 	/* base must be 32 byte aligned */
314 	amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
315 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
316 	amdgpu_ring_write(ring, ib->length_dw);
317 	amdgpu_ring_write(ring, lower_32_bits(csa_mc_addr));
318 	amdgpu_ring_write(ring, upper_32_bits(csa_mc_addr));
319 }
320 
321 /**
322  * sdma_v7_0_ring_emit_mem_sync - flush the IB by graphics cache rinse
323  *
324  * @ring: amdgpu ring pointer
325  *
326  * flush the IB by graphics cache rinse.
327  */
328 static void sdma_v7_0_ring_emit_mem_sync(struct amdgpu_ring *ring)
329 {
330 	uint32_t gcr_cntl = SDMA_GCR_GL2_INV | SDMA_GCR_GL2_WB | SDMA_GCR_GLM_INV |
331 		SDMA_GCR_GL1_INV | SDMA_GCR_GLV_INV | SDMA_GCR_GLK_INV |
332 		SDMA_GCR_GLI_INV(1);
333 
334 	/* flush entire cache L0/L1/L2, this can be optimized by performance requirement */
335 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_GCR_REQ));
336 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD1_BASE_VA_31_7(0));
337 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD2_GCR_CONTROL_15_0(gcr_cntl) |
338 			  SDMA_PKT_GCR_REQ_PAYLOAD2_BASE_VA_47_32(0));
339 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD3_LIMIT_VA_31_7(0) |
340 			  SDMA_PKT_GCR_REQ_PAYLOAD3_GCR_CONTROL_18_16(gcr_cntl >> 16));
341 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD4_LIMIT_VA_47_32(0) |
342 			  SDMA_PKT_GCR_REQ_PAYLOAD4_VMID(0));
343 }
344 
345 
346 /**
347  * sdma_v7_0_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
348  *
349  * @ring: amdgpu ring pointer
350  *
351  * Emit an hdp flush packet on the requested DMA ring.
352  */
353 static void sdma_v7_0_ring_emit_hdp_flush(struct amdgpu_ring *ring)
354 {
355 	struct amdgpu_device *adev = ring->adev;
356 	u32 ref_and_mask = 0;
357 	const struct nbio_hdp_flush_reg *nbio_hf_reg = adev->nbio.hdp_flush_reg;
358 
359 	ref_and_mask = nbio_hf_reg->ref_and_mask_sdma0 << ring->me;
360 
361 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_POLL_REGMEM) |
362 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) |
363 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
364 	amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_done_offset(adev)) << 2);
365 	amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_req_offset(adev)) << 2);
366 	amdgpu_ring_write(ring, ref_and_mask); /* reference */
367 	amdgpu_ring_write(ring, ref_and_mask); /* mask */
368 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
369 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
370 }
371 
372 /**
373  * sdma_v7_0_ring_emit_fence - emit a fence on the DMA ring
374  *
375  * @ring: amdgpu ring pointer
376  * @addr: address
377  * @seq: fence seq number
378  * @flags: fence flags
379  *
380  * Add a DMA fence packet to the ring to write
381  * the fence seq number and DMA trap packet to generate
382  * an interrupt if needed.
383  */
384 static void sdma_v7_0_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
385 				      unsigned flags)
386 {
387 	bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
388 	/* write the fence */
389 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_FENCE) |
390 			  SDMA_PKT_FENCE_HEADER_MTYPE(0x3)); /* Ucached(UC) */
391 	/* zero in first two bits */
392 	BUG_ON(addr & 0x3);
393 	amdgpu_ring_write(ring, lower_32_bits(addr));
394 	amdgpu_ring_write(ring, upper_32_bits(addr));
395 	amdgpu_ring_write(ring, lower_32_bits(seq));
396 
397 	/* optionally write high bits as well */
398 	if (write64bit) {
399 		addr += 4;
400 		amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_FENCE) |
401 				  SDMA_PKT_FENCE_HEADER_MTYPE(0x3));
402 		/* zero in first two bits */
403 		BUG_ON(addr & 0x3);
404 		amdgpu_ring_write(ring, lower_32_bits(addr));
405 		amdgpu_ring_write(ring, upper_32_bits(addr));
406 		amdgpu_ring_write(ring, upper_32_bits(seq));
407 	}
408 
409 	if (flags & AMDGPU_FENCE_FLAG_INT) {
410 		uint32_t ctx = ring->is_mes_queue ?
411 			(ring->hw_queue_id | AMDGPU_FENCE_MES_QUEUE_FLAG) : 0;
412 		/* generate an interrupt */
413 		amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_TRAP));
414 		amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(ctx));
415 	}
416 }
417 
418 /**
419  * sdma_v7_0_gfx_stop - stop the gfx async dma engines
420  *
421  * @adev: amdgpu_device pointer
422  *
423  * Stop the gfx async dma ring buffers.
424  */
425 static void sdma_v7_0_gfx_stop(struct amdgpu_device *adev)
426 {
427 	u32 rb_cntl, ib_cntl;
428 	int i;
429 
430 	for (i = 0; i < adev->sdma.num_instances; i++) {
431 		rb_cntl = RREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_CNTL));
432 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RB_ENABLE, 0);
433 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_CNTL), rb_cntl);
434 		ib_cntl = RREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_IB_CNTL));
435 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_QUEUE0_IB_CNTL, IB_ENABLE, 0);
436 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_IB_CNTL), ib_cntl);
437 	}
438 }
439 
440 /**
441  * sdma_v7_0_rlc_stop - stop the compute async dma engines
442  *
443  * @adev: amdgpu_device pointer
444  *
445  * Stop the compute async dma queues.
446  */
447 static void sdma_v7_0_rlc_stop(struct amdgpu_device *adev)
448 {
449 	/* XXX todo */
450 }
451 
452 /**
453  * sdma_v7_0_ctx_switch_enable - stop the async dma engines context switch
454  *
455  * @adev: amdgpu_device pointer
456  * @enable: enable/disable the DMA MEs context switch.
457  *
458  * Halt or unhalt the async dma engines context switch.
459  */
460 static void sdma_v7_0_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
461 {
462 }
463 
464 /**
465  * sdma_v7_0_enable - stop the async dma engines
466  *
467  * @adev: amdgpu_device pointer
468  * @enable: enable/disable the DMA MEs.
469  *
470  * Halt or unhalt the async dma engines.
471  */
472 static void sdma_v7_0_enable(struct amdgpu_device *adev, bool enable)
473 {
474 	u32 mcu_cntl;
475 	int i;
476 
477 	if (!enable) {
478 		sdma_v7_0_gfx_stop(adev);
479 		sdma_v7_0_rlc_stop(adev);
480 	}
481 
482 	if (amdgpu_sriov_vf(adev))
483 		return;
484 
485 	for (i = 0; i < adev->sdma.num_instances; i++) {
486 		mcu_cntl = RREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_MCU_CNTL));
487 		mcu_cntl = REG_SET_FIELD(mcu_cntl, SDMA0_MCU_CNTL, HALT, enable ? 0 : 1);
488 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_MCU_CNTL), mcu_cntl);
489 	}
490 }
491 
492 /**
493  * sdma_v7_0_gfx_resume - setup and start the async dma engines
494  *
495  * @adev: amdgpu_device pointer
496  *
497  * Set up the gfx DMA ring buffers and enable them.
498  * Returns 0 for success, error for failure.
499  */
500 static int sdma_v7_0_gfx_resume(struct amdgpu_device *adev)
501 {
502 	struct amdgpu_ring *ring;
503 	u32 rb_cntl, ib_cntl;
504 	u32 rb_bufsz;
505 	u32 doorbell;
506 	u32 doorbell_offset;
507 	u32 tmp;
508 	u64 wptr_gpu_addr;
509 	int i, r;
510 
511 	for (i = 0; i < adev->sdma.num_instances; i++) {
512 		ring = &adev->sdma.instance[i].ring;
513 
514 		//if (!amdgpu_sriov_vf(adev))
515 		//	WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_SEM_WAIT_FAIL_TIMER_CNTL), 0);
516 
517 		/* Set ring buffer size in dwords */
518 		rb_bufsz = order_base_2(ring->ring_size / 4);
519 		rb_cntl = RREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_CNTL));
520 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RB_SIZE, rb_bufsz);
521 #ifdef __BIG_ENDIAN
522 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RB_SWAP_ENABLE, 1);
523 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL,
524 					RPTR_WRITEBACK_SWAP_ENABLE, 1);
525 #endif
526 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RB_PRIV, 1);
527 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_CNTL), rb_cntl);
528 
529 		/* Initialize the ring buffer's read and write pointers */
530 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_RPTR), 0);
531 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_RPTR_HI), 0);
532 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR), 0);
533 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR_HI), 0);
534 
535 		/* setup the wptr shadow polling */
536 		wptr_gpu_addr = ring->wptr_gpu_addr;
537 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR_POLL_ADDR_LO),
538 		       lower_32_bits(wptr_gpu_addr));
539 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR_POLL_ADDR_HI),
540 		       upper_32_bits(wptr_gpu_addr));
541 
542 		/* set the wb address whether it's enabled or not */
543 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_RPTR_ADDR_HI),
544 		       upper_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFF);
545 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_RPTR_ADDR_LO),
546 		       lower_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFC);
547 
548 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
549 		if (amdgpu_sriov_vf(adev))
550 			rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, WPTR_POLL_ENABLE, 1);
551 		else
552 			rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, WPTR_POLL_ENABLE, 0);
553 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, MCU_WPTR_POLL_ENABLE, 1);
554 
555 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_BASE), ring->gpu_addr >> 8);
556 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_BASE_HI), ring->gpu_addr >> 40);
557 
558 		ring->wptr = 0;
559 
560 		/* before programing wptr to a less value, need set minor_ptr_update first */
561 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_MINOR_PTR_UPDATE), 1);
562 
563 		if (!amdgpu_sriov_vf(adev)) { /* only bare-metal use register write for wptr */
564 			WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR), lower_32_bits(ring->wptr) << 2);
565 			WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR_HI), upper_32_bits(ring->wptr) << 2);
566 		}
567 
568 		doorbell = RREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_DOORBELL));
569 		doorbell_offset = RREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_DOORBELL_OFFSET));
570 
571 		if (ring->use_doorbell) {
572 			doorbell = REG_SET_FIELD(doorbell, SDMA0_QUEUE0_DOORBELL, ENABLE, 1);
573 			doorbell_offset = REG_SET_FIELD(doorbell_offset, SDMA0_QUEUE0_DOORBELL_OFFSET,
574 					OFFSET, ring->doorbell_index);
575 		} else {
576 			doorbell = REG_SET_FIELD(doorbell, SDMA0_QUEUE0_DOORBELL, ENABLE, 0);
577 		}
578 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_DOORBELL), doorbell);
579 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_DOORBELL_OFFSET), doorbell_offset);
580 
581 		if (i == 0)
582 			adev->nbio.funcs->sdma_doorbell_range(adev, i, ring->use_doorbell,
583 						      ring->doorbell_index,
584 						      adev->doorbell_index.sdma_doorbell_range * adev->sdma.num_instances);
585 
586 		if (amdgpu_sriov_vf(adev))
587 			sdma_v7_0_ring_set_wptr(ring);
588 
589 		/* set minor_ptr_update to 0 after wptr programed */
590 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_MINOR_PTR_UPDATE), 0);
591 
592 		/* Set up sdma hang watchdog */
593 		tmp = RREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_WATCHDOG_CNTL));
594 		/* 100ms per unit */
595 		tmp = REG_SET_FIELD(tmp, SDMA0_WATCHDOG_CNTL, QUEUE_HANG_COUNT,
596 				    max(adev->usec_timeout/100000, 1));
597 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_WATCHDOG_CNTL), tmp);
598 
599 		/* Set up RESP_MODE to non-copy addresses */
600 		tmp = RREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_UTCL1_CNTL));
601 		tmp = REG_SET_FIELD(tmp, SDMA0_UTCL1_CNTL, RESP_MODE, 3);
602 		tmp = REG_SET_FIELD(tmp, SDMA0_UTCL1_CNTL, REDO_DELAY, 9);
603 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_UTCL1_CNTL), tmp);
604 
605 		/* program default cache read and write policy */
606 		tmp = RREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_UTCL1_PAGE));
607 		/* clean read policy and write policy bits */
608 		tmp &= 0xFF0FFF;
609 		tmp |= ((CACHE_READ_POLICY_L2__DEFAULT << 12) |
610 			 (CACHE_WRITE_POLICY_L2__DEFAULT << 14));
611 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_UTCL1_PAGE), tmp);
612 
613 		if (!amdgpu_sriov_vf(adev)) {
614 			/* unhalt engine */
615 			tmp = RREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_MCU_CNTL));
616 			tmp = REG_SET_FIELD(tmp, SDMA0_MCU_CNTL, HALT, 0);
617 			tmp = REG_SET_FIELD(tmp, SDMA0_MCU_CNTL, RESET, 0);
618 			WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_MCU_CNTL), tmp);
619 		}
620 
621 		/* enable DMA RB */
622 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RB_ENABLE, 1);
623 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_CNTL), rb_cntl);
624 
625 		ib_cntl = RREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_IB_CNTL));
626 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_QUEUE0_IB_CNTL, IB_ENABLE, 1);
627 #ifdef __BIG_ENDIAN
628 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_QUEUE0_IB_CNTL, IB_SWAP_ENABLE, 1);
629 #endif
630 		/* enable DMA IBs */
631 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_IB_CNTL), ib_cntl);
632 
633 		ring->sched.ready = true;
634 
635 		if (amdgpu_sriov_vf(adev)) { /* bare-metal sequence doesn't need below to lines */
636 			sdma_v7_0_ctx_switch_enable(adev, true);
637 			sdma_v7_0_enable(adev, true);
638 		}
639 
640 		r = amdgpu_ring_test_helper(ring);
641 		if (r) {
642 			ring->sched.ready = false;
643 			return r;
644 		}
645 
646 	}
647 
648 	return 0;
649 }
650 
651 /**
652  * sdma_v7_0_rlc_resume - setup and start the async dma engines
653  *
654  * @adev: amdgpu_device pointer
655  *
656  * Set up the compute DMA queues and enable them.
657  * Returns 0 for success, error for failure.
658  */
659 static int sdma_v7_0_rlc_resume(struct amdgpu_device *adev)
660 {
661 	return 0;
662 }
663 
664 static void sdma_v12_0_free_ucode_buffer(struct amdgpu_device *adev)
665 {
666 	int i;
667 
668 	for (i = 0; i < adev->sdma.num_instances; i++) {
669 		amdgpu_bo_free_kernel(&adev->sdma.instance[i].sdma_fw_obj,
670 				      &adev->sdma.instance[i].sdma_fw_gpu_addr,
671 				      (void **)&adev->sdma.instance[i].sdma_fw_ptr);
672 	}
673 }
674 
675 /**
676  * sdma_v7_0_load_microcode - load the sDMA ME ucode
677  *
678  * @adev: amdgpu_device pointer
679  *
680  * Loads the sDMA0/1 ucode.
681  * Returns 0 for success, -EINVAL if the ucode is not available.
682  */
683 static int sdma_v7_0_load_microcode(struct amdgpu_device *adev)
684 {
685 	const struct sdma_firmware_header_v3_0 *hdr;
686 	const __le32 *fw_data;
687 	u32 fw_size;
688 	uint32_t tmp, sdma_status, ic_op_cntl;
689 	int i, r, j;
690 
691 	/* halt the MEs */
692 	sdma_v7_0_enable(adev, false);
693 
694 	if (!adev->sdma.instance[0].fw)
695 		return -EINVAL;
696 
697 	hdr = (const struct sdma_firmware_header_v3_0 *)
698 		adev->sdma.instance[0].fw->data;
699 	amdgpu_ucode_print_sdma_hdr(&hdr->header);
700 
701 	fw_data = (const __le32 *)(adev->sdma.instance[0].fw->data +
702 			le32_to_cpu(hdr->ucode_offset_bytes));
703 	fw_size = le32_to_cpu(hdr->ucode_size_bytes);
704 
705 	for (i = 0; i < adev->sdma.num_instances; i++) {
706 		r = amdgpu_bo_create_reserved(adev, fw_size,
707 					      PAGE_SIZE,
708 					      AMDGPU_GEM_DOMAIN_VRAM,
709 					      &adev->sdma.instance[i].sdma_fw_obj,
710 					      &adev->sdma.instance[i].sdma_fw_gpu_addr,
711 					      (void **)&adev->sdma.instance[i].sdma_fw_ptr);
712 		if (r) {
713 			dev_err(adev->dev, "(%d) failed to create sdma ucode bo\n", r);
714 			return r;
715 		}
716 
717 		memcpy(adev->sdma.instance[i].sdma_fw_ptr, fw_data, fw_size);
718 
719 		amdgpu_bo_kunmap(adev->sdma.instance[i].sdma_fw_obj);
720 		amdgpu_bo_unreserve(adev->sdma.instance[i].sdma_fw_obj);
721 
722 		tmp = RREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_IC_CNTL));
723 		tmp = REG_SET_FIELD(tmp, SDMA0_IC_CNTL, GPA, 0);
724 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_IC_CNTL), tmp);
725 
726 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_IC_BASE_LO),
727 			lower_32_bits(adev->sdma.instance[i].sdma_fw_gpu_addr));
728 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_IC_BASE_HI),
729 			upper_32_bits(adev->sdma.instance[i].sdma_fw_gpu_addr));
730 
731 		tmp = RREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_IC_OP_CNTL));
732 		tmp = REG_SET_FIELD(tmp, SDMA0_IC_OP_CNTL, PRIME_ICACHE, 1);
733 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_IC_OP_CNTL), tmp);
734 
735 		/* Wait for sdma ucode init complete */
736 		for (j = 0; j < adev->usec_timeout; j++) {
737 			ic_op_cntl = RREG32_SOC15_IP(GC,
738 					sdma_v7_0_get_reg_offset(adev, i, regSDMA0_IC_OP_CNTL));
739 			sdma_status = RREG32_SOC15_IP(GC,
740 					sdma_v7_0_get_reg_offset(adev, i, regSDMA0_STATUS_REG));
741 			if ((REG_GET_FIELD(ic_op_cntl, SDMA0_IC_OP_CNTL, ICACHE_PRIMED) == 1) &&
742 			    (REG_GET_FIELD(sdma_status, SDMA0_STATUS_REG, UCODE_INIT_DONE) == 1))
743 				break;
744 			udelay(1);
745 		}
746 
747 		if (j >= adev->usec_timeout) {
748 			dev_err(adev->dev, "failed to init sdma ucode\n");
749 			return -EINVAL;
750 		}
751 	}
752 
753 	return 0;
754 }
755 
756 static int sdma_v7_0_soft_reset(struct amdgpu_ip_block *ip_block)
757 {
758 	struct amdgpu_device *adev = ip_block->adev;
759 	u32 tmp;
760 	int i;
761 
762 	sdma_v7_0_gfx_stop(adev);
763 
764 	for (i = 0; i < adev->sdma.num_instances; i++) {
765 		//tmp = RREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_FREEZE));
766 		//tmp |= SDMA0_FREEZE__FREEZE_MASK;
767 		//WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_FREEZE), tmp);
768 		tmp = RREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_MCU_CNTL));
769 		tmp |= SDMA0_MCU_CNTL__HALT_MASK;
770 		tmp |= SDMA0_MCU_CNTL__RESET_MASK;
771 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_MCU_CNTL), tmp);
772 
773 		WREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_PREEMPT), 0);
774 
775 		udelay(100);
776 
777 		tmp = GRBM_SOFT_RESET__SOFT_RESET_SDMA0_MASK << i;
778 		WREG32_SOC15(GC, 0, regGRBM_SOFT_RESET, tmp);
779 		tmp = RREG32_SOC15(GC, 0, regGRBM_SOFT_RESET);
780 
781 		udelay(100);
782 
783 		WREG32_SOC15(GC, 0, regGRBM_SOFT_RESET, 0);
784 		tmp = RREG32_SOC15(GC, 0, regGRBM_SOFT_RESET);
785 
786 		udelay(100);
787 	}
788 
789 	return sdma_v7_0_start(adev);
790 }
791 
792 static bool sdma_v7_0_check_soft_reset(struct amdgpu_ip_block *ip_block)
793 {
794 	struct amdgpu_device *adev = ip_block->adev;
795 	struct amdgpu_ring *ring;
796 	int i, r;
797 	long tmo = msecs_to_jiffies(1000);
798 
799 	for (i = 0; i < adev->sdma.num_instances; i++) {
800 		ring = &adev->sdma.instance[i].ring;
801 		r = amdgpu_ring_test_ib(ring, tmo);
802 		if (r)
803 			return true;
804 	}
805 
806 	return false;
807 }
808 
809 /**
810  * sdma_v7_0_start - setup and start the async dma engines
811  *
812  * @adev: amdgpu_device pointer
813  *
814  * Set up the DMA engines and enable them.
815  * Returns 0 for success, error for failure.
816  */
817 static int sdma_v7_0_start(struct amdgpu_device *adev)
818 {
819 	int r = 0;
820 
821 	if (amdgpu_sriov_vf(adev)) {
822 		sdma_v7_0_ctx_switch_enable(adev, false);
823 		sdma_v7_0_enable(adev, false);
824 
825 		/* set RB registers */
826 		r = sdma_v7_0_gfx_resume(adev);
827 		return r;
828 	}
829 
830 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
831 		r = sdma_v7_0_load_microcode(adev);
832 		if (r) {
833 			sdma_v12_0_free_ucode_buffer(adev);
834 			return r;
835 		}
836 
837 		if (amdgpu_emu_mode == 1)
838 			msleep(1000);
839 	}
840 
841 	/* unhalt the MEs */
842 	sdma_v7_0_enable(adev, true);
843 	/* enable sdma ring preemption */
844 	sdma_v7_0_ctx_switch_enable(adev, true);
845 
846 	/* start the gfx rings and rlc compute queues */
847 	r = sdma_v7_0_gfx_resume(adev);
848 	if (r)
849 		return r;
850 	r = sdma_v7_0_rlc_resume(adev);
851 
852 	return r;
853 }
854 
855 static int sdma_v7_0_mqd_init(struct amdgpu_device *adev, void *mqd,
856 			      struct amdgpu_mqd_prop *prop)
857 {
858 	struct v12_sdma_mqd *m = mqd;
859 	uint64_t wb_gpu_addr;
860 
861 	m->sdmax_rlcx_rb_cntl =
862 		order_base_2(prop->queue_size / 4) << SDMA0_QUEUE0_RB_CNTL__RB_SIZE__SHIFT |
863 		1 << SDMA0_QUEUE0_RB_CNTL__RPTR_WRITEBACK_ENABLE__SHIFT |
864 		4 << SDMA0_QUEUE0_RB_CNTL__RPTR_WRITEBACK_TIMER__SHIFT |
865 		1 << SDMA0_QUEUE0_RB_CNTL__MCU_WPTR_POLL_ENABLE__SHIFT;
866 
867 	m->sdmax_rlcx_rb_base = lower_32_bits(prop->hqd_base_gpu_addr >> 8);
868 	m->sdmax_rlcx_rb_base_hi = upper_32_bits(prop->hqd_base_gpu_addr >> 8);
869 
870 	wb_gpu_addr = prop->wptr_gpu_addr;
871 	m->sdmax_rlcx_rb_wptr_poll_addr_lo = lower_32_bits(wb_gpu_addr);
872 	m->sdmax_rlcx_rb_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr);
873 
874 	wb_gpu_addr = prop->rptr_gpu_addr;
875 	m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits(wb_gpu_addr);
876 	m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits(wb_gpu_addr);
877 
878 	m->sdmax_rlcx_ib_cntl = RREG32_SOC15_IP(GC, sdma_v7_0_get_reg_offset(adev, 0,
879 							regSDMA0_QUEUE0_IB_CNTL));
880 
881 	m->sdmax_rlcx_doorbell_offset =
882 		prop->doorbell_index << SDMA0_QUEUE0_DOORBELL_OFFSET__OFFSET__SHIFT;
883 
884 	m->sdmax_rlcx_doorbell = REG_SET_FIELD(0, SDMA0_QUEUE0_DOORBELL, ENABLE, 1);
885 
886 	m->sdmax_rlcx_doorbell_log = 0;
887 	m->sdmax_rlcx_rb_aql_cntl = 0x4000;	//regSDMA0_QUEUE0_RB_AQL_CNTL_DEFAULT;
888 	m->sdmax_rlcx_dummy_reg = 0xf;	//regSDMA0_QUEUE0_DUMMY_REG_DEFAULT;
889 
890 	return 0;
891 }
892 
893 static void sdma_v7_0_set_mqd_funcs(struct amdgpu_device *adev)
894 {
895 	adev->mqds[AMDGPU_HW_IP_DMA].mqd_size = sizeof(struct v12_sdma_mqd);
896 	adev->mqds[AMDGPU_HW_IP_DMA].init_mqd = sdma_v7_0_mqd_init;
897 }
898 
899 /**
900  * sdma_v7_0_ring_test_ring - simple async dma engine test
901  *
902  * @ring: amdgpu_ring structure holding ring information
903  *
904  * Test the DMA engine by writing using it to write an
905  * value to memory.
906  * Returns 0 for success, error for failure.
907  */
908 static int sdma_v7_0_ring_test_ring(struct amdgpu_ring *ring)
909 {
910 	struct amdgpu_device *adev = ring->adev;
911 	unsigned i;
912 	unsigned index;
913 	int r;
914 	u32 tmp;
915 	u64 gpu_addr;
916 	volatile uint32_t *cpu_ptr = NULL;
917 
918 	tmp = 0xCAFEDEAD;
919 
920 	if (ring->is_mes_queue) {
921 		uint32_t offset = 0;
922 		offset = amdgpu_mes_ctx_get_offs(ring,
923 					 AMDGPU_MES_CTX_PADDING_OFFS);
924 		gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
925 		cpu_ptr = amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
926 		*cpu_ptr = tmp;
927 	} else {
928 		r = amdgpu_device_wb_get(adev, &index);
929 		if (r) {
930 			dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
931 			return r;
932 		}
933 
934 		gpu_addr = adev->wb.gpu_addr + (index * 4);
935 		adev->wb.wb[index] = cpu_to_le32(tmp);
936 	}
937 
938 	r = amdgpu_ring_alloc(ring, 5);
939 	if (r) {
940 		DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
941 		if (!ring->is_mes_queue)
942 			amdgpu_device_wb_free(adev, index);
943 		return r;
944 	}
945 
946 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_WRITE) |
947 			  SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
948 	amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
949 	amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
950 	amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0));
951 	amdgpu_ring_write(ring, 0xDEADBEEF);
952 	amdgpu_ring_commit(ring);
953 
954 	for (i = 0; i < adev->usec_timeout; i++) {
955 		if (ring->is_mes_queue)
956 			tmp = le32_to_cpu(*cpu_ptr);
957 		else
958 			tmp = le32_to_cpu(adev->wb.wb[index]);
959 		if (tmp == 0xDEADBEEF)
960 			break;
961 		if (amdgpu_emu_mode == 1)
962 			msleep(1);
963 		else
964 			udelay(1);
965 	}
966 
967 	if (i >= adev->usec_timeout)
968 		r = -ETIMEDOUT;
969 
970 	if (!ring->is_mes_queue)
971 		amdgpu_device_wb_free(adev, index);
972 
973 	return r;
974 }
975 
976 /**
977  * sdma_v7_0_ring_test_ib - test an IB on the DMA engine
978  *
979  * @ring: amdgpu_ring structure holding ring information
980  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
981  *
982  * Test a simple IB in the DMA ring.
983  * Returns 0 on success, error on failure.
984  */
985 static int sdma_v7_0_ring_test_ib(struct amdgpu_ring *ring, long timeout)
986 {
987 	struct amdgpu_device *adev = ring->adev;
988 	struct amdgpu_ib ib;
989 	struct dma_fence *f = NULL;
990 	unsigned index;
991 	long r;
992 	u32 tmp = 0;
993 	u64 gpu_addr;
994 	volatile uint32_t *cpu_ptr = NULL;
995 
996 	tmp = 0xCAFEDEAD;
997 	memset(&ib, 0, sizeof(ib));
998 
999 	if (ring->is_mes_queue) {
1000 		uint32_t offset = 0;
1001 		offset = amdgpu_mes_ctx_get_offs(ring, AMDGPU_MES_CTX_IB_OFFS);
1002 		ib.gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
1003 		ib.ptr = (void *)amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
1004 
1005 		offset = amdgpu_mes_ctx_get_offs(ring,
1006 					 AMDGPU_MES_CTX_PADDING_OFFS);
1007 		gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
1008 		cpu_ptr = amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
1009 		*cpu_ptr = tmp;
1010 	} else {
1011 		r = amdgpu_device_wb_get(adev, &index);
1012 		if (r) {
1013 			dev_err(adev->dev, "(%ld) failed to allocate wb slot\n", r);
1014 			return r;
1015 		}
1016 
1017 		gpu_addr = adev->wb.gpu_addr + (index * 4);
1018 		adev->wb.wb[index] = cpu_to_le32(tmp);
1019 
1020 		r = amdgpu_ib_get(adev, NULL, 256, AMDGPU_IB_POOL_DIRECT, &ib);
1021 		if (r) {
1022 			DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r);
1023 			goto err0;
1024 		}
1025 	}
1026 
1027 	ib.ptr[0] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_WRITE) |
1028 		SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1029 	ib.ptr[1] = lower_32_bits(gpu_addr);
1030 	ib.ptr[2] = upper_32_bits(gpu_addr);
1031 	ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0);
1032 	ib.ptr[4] = 0xDEADBEEF;
1033 	ib.ptr[5] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1034 	ib.ptr[6] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1035 	ib.ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1036 	ib.length_dw = 8;
1037 
1038 	r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
1039 	if (r)
1040 		goto err1;
1041 
1042 	r = dma_fence_wait_timeout(f, false, timeout);
1043 	if (r == 0) {
1044 		DRM_ERROR("amdgpu: IB test timed out\n");
1045 		r = -ETIMEDOUT;
1046 		goto err1;
1047 	} else if (r < 0) {
1048 		DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
1049 		goto err1;
1050 	}
1051 
1052 	if (ring->is_mes_queue)
1053 		tmp = le32_to_cpu(*cpu_ptr);
1054 	else
1055 		tmp = le32_to_cpu(adev->wb.wb[index]);
1056 
1057 	if (tmp == 0xDEADBEEF)
1058 		r = 0;
1059 	else
1060 		r = -EINVAL;
1061 
1062 err1:
1063 	amdgpu_ib_free(adev, &ib, NULL);
1064 	dma_fence_put(f);
1065 err0:
1066 	if (!ring->is_mes_queue)
1067 		amdgpu_device_wb_free(adev, index);
1068 	return r;
1069 }
1070 
1071 
1072 /**
1073  * sdma_v7_0_vm_copy_pte - update PTEs by copying them from the GART
1074  *
1075  * @ib: indirect buffer to fill with commands
1076  * @pe: addr of the page entry
1077  * @src: src addr to copy from
1078  * @count: number of page entries to update
1079  *
1080  * Update PTEs by copying them from the GART using sDMA.
1081  */
1082 static void sdma_v7_0_vm_copy_pte(struct amdgpu_ib *ib,
1083 				  uint64_t pe, uint64_t src,
1084 				  unsigned count)
1085 {
1086 	unsigned bytes = count * 8;
1087 
1088 	ib->ptr[ib->length_dw++] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_COPY) |
1089 		SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR) |
1090 		SDMA_PKT_COPY_LINEAR_HEADER_CPV(1);
1091 
1092 	ib->ptr[ib->length_dw++] = bytes - 1;
1093 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1094 	ib->ptr[ib->length_dw++] = lower_32_bits(src);
1095 	ib->ptr[ib->length_dw++] = upper_32_bits(src);
1096 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1097 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1098 	ib->ptr[ib->length_dw++] = 0;
1099 
1100 }
1101 
1102 /**
1103  * sdma_v7_0_vm_write_pte - update PTEs by writing them manually
1104  *
1105  * @ib: indirect buffer to fill with commands
1106  * @pe: addr of the page entry
1107  * @value: dst addr to write into pe
1108  * @count: number of page entries to update
1109  * @incr: increase next addr by incr bytes
1110  *
1111  * Update PTEs by writing them manually using sDMA.
1112  */
1113 static void sdma_v7_0_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
1114 				   uint64_t value, unsigned count,
1115 				   uint32_t incr)
1116 {
1117 	unsigned ndw = count * 2;
1118 
1119 	ib->ptr[ib->length_dw++] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_WRITE) |
1120 		SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1121 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1122 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1123 	ib->ptr[ib->length_dw++] = ndw - 1;
1124 	for (; ndw > 0; ndw -= 2) {
1125 		ib->ptr[ib->length_dw++] = lower_32_bits(value);
1126 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
1127 		value += incr;
1128 	}
1129 }
1130 
1131 /**
1132  * sdma_v7_0_vm_set_pte_pde - update the page tables using sDMA
1133  *
1134  * @ib: indirect buffer to fill with commands
1135  * @pe: addr of the page entry
1136  * @addr: dst addr to write into pe
1137  * @count: number of page entries to update
1138  * @incr: increase next addr by incr bytes
1139  * @flags: access flags
1140  *
1141  * Update the page tables using sDMA.
1142  */
1143 static void sdma_v7_0_vm_set_pte_pde(struct amdgpu_ib *ib,
1144 				     uint64_t pe,
1145 				     uint64_t addr, unsigned count,
1146 				     uint32_t incr, uint64_t flags)
1147 {
1148 	/* for physically contiguous pages (vram) */
1149 	ib->ptr[ib->length_dw++] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_PTEPDE);
1150 	ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
1151 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1152 	ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
1153 	ib->ptr[ib->length_dw++] = upper_32_bits(flags);
1154 	ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
1155 	ib->ptr[ib->length_dw++] = upper_32_bits(addr);
1156 	ib->ptr[ib->length_dw++] = incr; /* increment size */
1157 	ib->ptr[ib->length_dw++] = 0;
1158 	ib->ptr[ib->length_dw++] = count - 1; /* number of entries */
1159 }
1160 
1161 /**
1162  * sdma_v7_0_ring_pad_ib - pad the IB
1163  *
1164  * @ring: amdgpu ring pointer
1165  * @ib: indirect buffer to fill with padding
1166  *
1167  * Pad the IB with NOPs to a boundary multiple of 8.
1168  */
1169 static void sdma_v7_0_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
1170 {
1171 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
1172 	u32 pad_count;
1173 	int i;
1174 
1175 	pad_count = (-ib->length_dw) & 0x7;
1176 	for (i = 0; i < pad_count; i++)
1177 		if (sdma && sdma->burst_nop && (i == 0))
1178 			ib->ptr[ib->length_dw++] =
1179 				SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_NOP) |
1180 				SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
1181 		else
1182 			ib->ptr[ib->length_dw++] =
1183 				SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_NOP);
1184 }
1185 
1186 /**
1187  * sdma_v7_0_ring_emit_pipeline_sync - sync the pipeline
1188  *
1189  * @ring: amdgpu_ring pointer
1190  *
1191  * Make sure all previous operations are completed (CIK).
1192  */
1193 static void sdma_v7_0_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
1194 {
1195 	uint32_t seq = ring->fence_drv.sync_seq;
1196 	uint64_t addr = ring->fence_drv.gpu_addr;
1197 
1198 	/* wait for idle */
1199 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1200 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1201 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
1202 			  SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1));
1203 	amdgpu_ring_write(ring, addr & 0xfffffffc);
1204 	amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
1205 	amdgpu_ring_write(ring, seq); /* reference */
1206 	amdgpu_ring_write(ring, 0xffffffff); /* mask */
1207 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1208 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */
1209 }
1210 
1211 /**
1212  * sdma_v7_0_ring_emit_vm_flush - vm flush using sDMA
1213  *
1214  * @ring: amdgpu_ring pointer
1215  * @vmid: vmid number to use
1216  * @pd_addr: address
1217  *
1218  * Update the page table base and flush the VM TLB
1219  * using sDMA.
1220  */
1221 static void sdma_v7_0_ring_emit_vm_flush(struct amdgpu_ring *ring,
1222 					 unsigned vmid, uint64_t pd_addr)
1223 {
1224 	amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
1225 }
1226 
1227 static void sdma_v7_0_ring_emit_wreg(struct amdgpu_ring *ring,
1228 				     uint32_t reg, uint32_t val)
1229 {
1230 	/* SRBM WRITE command will not support on sdma v7.
1231 	 * Use Register WRITE command instead, which OPCODE is same as SRBM WRITE
1232 	 */
1233 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_SRBM_WRITE));
1234 	amdgpu_ring_write(ring, reg << 2);
1235 	amdgpu_ring_write(ring, val);
1236 }
1237 
1238 static void sdma_v7_0_ring_emit_reg_wait(struct amdgpu_ring *ring, uint32_t reg,
1239 					 uint32_t val, uint32_t mask)
1240 {
1241 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1242 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1243 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* equal */
1244 	amdgpu_ring_write(ring, reg << 2);
1245 	amdgpu_ring_write(ring, 0);
1246 	amdgpu_ring_write(ring, val); /* reference */
1247 	amdgpu_ring_write(ring, mask); /* mask */
1248 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1249 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10));
1250 }
1251 
1252 static void sdma_v7_0_ring_emit_reg_write_reg_wait(struct amdgpu_ring *ring,
1253 						   uint32_t reg0, uint32_t reg1,
1254 						   uint32_t ref, uint32_t mask)
1255 {
1256 	amdgpu_ring_emit_wreg(ring, reg0, ref);
1257 	/* wait for a cycle to reset vm_inv_eng*_ack */
1258 	amdgpu_ring_emit_reg_wait(ring, reg0, 0, 0);
1259 	amdgpu_ring_emit_reg_wait(ring, reg1, mask, mask);
1260 }
1261 
1262 static int sdma_v7_0_early_init(struct amdgpu_ip_block *ip_block)
1263 {
1264 	struct amdgpu_device *adev = ip_block->adev;
1265 	int r;
1266 
1267 	r = amdgpu_sdma_init_microcode(adev, 0, true);
1268 	if (r) {
1269 		DRM_ERROR("Failed to init sdma firmware!\n");
1270 		return r;
1271 	}
1272 
1273 	sdma_v7_0_set_ring_funcs(adev);
1274 	sdma_v7_0_set_buffer_funcs(adev);
1275 	sdma_v7_0_set_vm_pte_funcs(adev);
1276 	sdma_v7_0_set_irq_funcs(adev);
1277 	sdma_v7_0_set_mqd_funcs(adev);
1278 
1279 	return 0;
1280 }
1281 
1282 static int sdma_v7_0_sw_init(struct amdgpu_ip_block *ip_block)
1283 {
1284 	struct amdgpu_ring *ring;
1285 	int r, i;
1286 	struct amdgpu_device *adev = ip_block->adev;
1287 	uint32_t reg_count = ARRAY_SIZE(sdma_reg_list_7_0);
1288 	uint32_t *ptr;
1289 
1290 	/* SDMA trap event */
1291 	r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_GFX,
1292 			      GFX_11_0_0__SRCID__SDMA_TRAP,
1293 			      &adev->sdma.trap_irq);
1294 	if (r)
1295 		return r;
1296 
1297 	for (i = 0; i < adev->sdma.num_instances; i++) {
1298 		ring = &adev->sdma.instance[i].ring;
1299 		ring->ring_obj = NULL;
1300 		ring->use_doorbell = true;
1301 		ring->me = i;
1302 
1303 		DRM_DEBUG("SDMA %d use_doorbell being set to: [%s]\n", i,
1304 				ring->use_doorbell?"true":"false");
1305 
1306 		ring->doorbell_index =
1307 			(adev->doorbell_index.sdma_engine[i] << 1); // get DWORD offset
1308 
1309 		ring->vm_hub = AMDGPU_GFXHUB(0);
1310 		sprintf(ring->name, "sdma%d", i);
1311 		r = amdgpu_ring_init(adev, ring, 1024,
1312 				     &adev->sdma.trap_irq,
1313 				     AMDGPU_SDMA_IRQ_INSTANCE0 + i,
1314 				     AMDGPU_RING_PRIO_DEFAULT, NULL);
1315 		if (r)
1316 			return r;
1317 	}
1318 
1319 	/* Allocate memory for SDMA IP Dump buffer */
1320 	ptr = kcalloc(adev->sdma.num_instances * reg_count, sizeof(uint32_t), GFP_KERNEL);
1321 	if (ptr)
1322 		adev->sdma.ip_dump = ptr;
1323 	else
1324 		DRM_ERROR("Failed to allocated memory for SDMA IP Dump\n");
1325 
1326 	return r;
1327 }
1328 
1329 static int sdma_v7_0_sw_fini(struct amdgpu_ip_block *ip_block)
1330 {
1331 	struct amdgpu_device *adev = ip_block->adev;
1332 	int i;
1333 
1334 	for (i = 0; i < adev->sdma.num_instances; i++)
1335 		amdgpu_ring_fini(&adev->sdma.instance[i].ring);
1336 
1337 	amdgpu_sdma_destroy_inst_ctx(adev, true);
1338 
1339 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT)
1340 		sdma_v12_0_free_ucode_buffer(adev);
1341 
1342 	kfree(adev->sdma.ip_dump);
1343 
1344 	return 0;
1345 }
1346 
1347 static int sdma_v7_0_hw_init(struct amdgpu_ip_block *ip_block)
1348 {
1349 	struct amdgpu_device *adev = ip_block->adev;
1350 
1351 	return sdma_v7_0_start(adev);
1352 }
1353 
1354 static int sdma_v7_0_hw_fini(struct amdgpu_ip_block *ip_block)
1355 {
1356 	struct amdgpu_device *adev = ip_block->adev;
1357 
1358 	if (amdgpu_sriov_vf(adev))
1359 		return 0;
1360 
1361 	sdma_v7_0_ctx_switch_enable(adev, false);
1362 	sdma_v7_0_enable(adev, false);
1363 
1364 	return 0;
1365 }
1366 
1367 static int sdma_v7_0_suspend(struct amdgpu_ip_block *ip_block)
1368 {
1369 	return sdma_v7_0_hw_fini(ip_block);
1370 }
1371 
1372 static int sdma_v7_0_resume(struct amdgpu_ip_block *ip_block)
1373 {
1374 	return sdma_v7_0_hw_init(ip_block);
1375 }
1376 
1377 static bool sdma_v7_0_is_idle(void *handle)
1378 {
1379 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1380 	u32 i;
1381 
1382 	for (i = 0; i < adev->sdma.num_instances; i++) {
1383 		u32 tmp = RREG32(sdma_v7_0_get_reg_offset(adev, i, regSDMA0_STATUS_REG));
1384 
1385 		if (!(tmp & SDMA0_STATUS_REG__IDLE_MASK))
1386 			return false;
1387 	}
1388 
1389 	return true;
1390 }
1391 
1392 static int sdma_v7_0_wait_for_idle(struct amdgpu_ip_block *ip_block)
1393 {
1394 	unsigned i;
1395 	u32 sdma0, sdma1;
1396 	struct amdgpu_device *adev = ip_block->adev;
1397 
1398 	for (i = 0; i < adev->usec_timeout; i++) {
1399 		sdma0 = RREG32(sdma_v7_0_get_reg_offset(adev, 0, regSDMA0_STATUS_REG));
1400 		sdma1 = RREG32(sdma_v7_0_get_reg_offset(adev, 1, regSDMA0_STATUS_REG));
1401 
1402 		if (sdma0 & sdma1 & SDMA0_STATUS_REG__IDLE_MASK)
1403 			return 0;
1404 		udelay(1);
1405 	}
1406 	return -ETIMEDOUT;
1407 }
1408 
1409 static int sdma_v7_0_ring_preempt_ib(struct amdgpu_ring *ring)
1410 {
1411 	int i, r = 0;
1412 	struct amdgpu_device *adev = ring->adev;
1413 	u32 index = 0;
1414 	u64 sdma_gfx_preempt;
1415 
1416 	amdgpu_sdma_get_index_from_ring(ring, &index);
1417 	sdma_gfx_preempt =
1418 		sdma_v7_0_get_reg_offset(adev, index, regSDMA0_QUEUE0_PREEMPT);
1419 
1420 	/* assert preemption condition */
1421 	amdgpu_ring_set_preempt_cond_exec(ring, false);
1422 
1423 	/* emit the trailing fence */
1424 	ring->trail_seq += 1;
1425 	r = amdgpu_ring_alloc(ring, 10);
1426 	if (r) {
1427 		DRM_ERROR("ring %d failed to be allocated \n", ring->idx);
1428 		return r;
1429 	}
1430 	sdma_v7_0_ring_emit_fence(ring, ring->trail_fence_gpu_addr,
1431 				  ring->trail_seq, 0);
1432 	amdgpu_ring_commit(ring);
1433 
1434 	/* assert IB preemption */
1435 	WREG32(sdma_gfx_preempt, 1);
1436 
1437 	/* poll the trailing fence */
1438 	for (i = 0; i < adev->usec_timeout; i++) {
1439 		if (ring->trail_seq ==
1440 		    le32_to_cpu(*(ring->trail_fence_cpu_addr)))
1441 			break;
1442 		udelay(1);
1443 	}
1444 
1445 	if (i >= adev->usec_timeout) {
1446 		r = -EINVAL;
1447 		DRM_ERROR("ring %d failed to be preempted\n", ring->idx);
1448 	}
1449 
1450 	/* deassert IB preemption */
1451 	WREG32(sdma_gfx_preempt, 0);
1452 
1453 	/* deassert the preemption condition */
1454 	amdgpu_ring_set_preempt_cond_exec(ring, true);
1455 	return r;
1456 }
1457 
1458 static int sdma_v7_0_set_trap_irq_state(struct amdgpu_device *adev,
1459 					struct amdgpu_irq_src *source,
1460 					unsigned type,
1461 					enum amdgpu_interrupt_state state)
1462 {
1463 	u32 sdma_cntl;
1464 
1465 	u32 reg_offset = sdma_v7_0_get_reg_offset(adev, type, regSDMA0_CNTL);
1466 
1467 	sdma_cntl = RREG32(reg_offset);
1468 	sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE,
1469 		       state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0);
1470 	WREG32(reg_offset, sdma_cntl);
1471 
1472 	return 0;
1473 }
1474 
1475 static int sdma_v7_0_process_trap_irq(struct amdgpu_device *adev,
1476 				      struct amdgpu_irq_src *source,
1477 				      struct amdgpu_iv_entry *entry)
1478 {
1479 	int instances, queue;
1480 	uint32_t mes_queue_id = entry->src_data[0];
1481 
1482 	DRM_DEBUG("IH: SDMA trap\n");
1483 
1484 	if (adev->enable_mes && (mes_queue_id & AMDGPU_FENCE_MES_QUEUE_FLAG)) {
1485 		struct amdgpu_mes_queue *queue;
1486 
1487 		mes_queue_id &= AMDGPU_FENCE_MES_QUEUE_ID_MASK;
1488 
1489 		spin_lock(&adev->mes.queue_id_lock);
1490 		queue = idr_find(&adev->mes.queue_id_idr, mes_queue_id);
1491 		if (queue) {
1492 			DRM_DEBUG("process smda queue id = %d\n", mes_queue_id);
1493 			amdgpu_fence_process(queue->ring);
1494 		}
1495 		spin_unlock(&adev->mes.queue_id_lock);
1496 		return 0;
1497 	}
1498 
1499 	queue = entry->ring_id & 0xf;
1500 	instances = (entry->ring_id & 0xf0) >> 4;
1501 	if (instances > 1) {
1502 		DRM_ERROR("IH: wrong ring_ID detected, as wrong sdma instance\n");
1503 		return -EINVAL;
1504 	}
1505 
1506 	switch (entry->client_id) {
1507 	case SOC21_IH_CLIENTID_GFX:
1508 		switch (queue) {
1509 		case 0:
1510 			amdgpu_fence_process(&adev->sdma.instance[instances].ring);
1511 			break;
1512 		default:
1513 			break;
1514 		}
1515 		break;
1516 	}
1517 	return 0;
1518 }
1519 
1520 static int sdma_v7_0_process_illegal_inst_irq(struct amdgpu_device *adev,
1521 					      struct amdgpu_irq_src *source,
1522 					      struct amdgpu_iv_entry *entry)
1523 {
1524 	return 0;
1525 }
1526 
1527 static int sdma_v7_0_set_clockgating_state(void *handle,
1528 					   enum amd_clockgating_state state)
1529 {
1530 	return 0;
1531 }
1532 
1533 static int sdma_v7_0_set_powergating_state(void *handle,
1534 					  enum amd_powergating_state state)
1535 {
1536 	return 0;
1537 }
1538 
1539 static void sdma_v7_0_get_clockgating_state(void *handle, u64 *flags)
1540 {
1541 }
1542 
1543 static void sdma_v7_0_print_ip_state(struct amdgpu_ip_block *ip_block, struct drm_printer *p)
1544 {
1545 	struct amdgpu_device *adev = ip_block->adev;
1546 	int i, j;
1547 	uint32_t reg_count = ARRAY_SIZE(sdma_reg_list_7_0);
1548 	uint32_t instance_offset;
1549 
1550 	if (!adev->sdma.ip_dump)
1551 		return;
1552 
1553 	drm_printf(p, "num_instances:%d\n", adev->sdma.num_instances);
1554 	for (i = 0; i < adev->sdma.num_instances; i++) {
1555 		instance_offset = i * reg_count;
1556 		drm_printf(p, "\nInstance:%d\n", i);
1557 
1558 		for (j = 0; j < reg_count; j++)
1559 			drm_printf(p, "%-50s \t 0x%08x\n", sdma_reg_list_7_0[j].reg_name,
1560 				   adev->sdma.ip_dump[instance_offset + j]);
1561 	}
1562 }
1563 
1564 static void sdma_v7_0_dump_ip_state(struct amdgpu_ip_block *ip_block)
1565 {
1566 	struct amdgpu_device *adev = ip_block->adev;
1567 	int i, j;
1568 	uint32_t instance_offset;
1569 	uint32_t reg_count = ARRAY_SIZE(sdma_reg_list_7_0);
1570 
1571 	if (!adev->sdma.ip_dump)
1572 		return;
1573 
1574 	amdgpu_gfx_off_ctrl(adev, false);
1575 	for (i = 0; i < adev->sdma.num_instances; i++) {
1576 		instance_offset = i * reg_count;
1577 		for (j = 0; j < reg_count; j++)
1578 			adev->sdma.ip_dump[instance_offset + j] =
1579 				RREG32(sdma_v7_0_get_reg_offset(adev, i,
1580 				       sdma_reg_list_7_0[j].reg_offset));
1581 	}
1582 	amdgpu_gfx_off_ctrl(adev, true);
1583 }
1584 
1585 const struct amd_ip_funcs sdma_v7_0_ip_funcs = {
1586 	.name = "sdma_v7_0",
1587 	.early_init = sdma_v7_0_early_init,
1588 	.late_init = NULL,
1589 	.sw_init = sdma_v7_0_sw_init,
1590 	.sw_fini = sdma_v7_0_sw_fini,
1591 	.hw_init = sdma_v7_0_hw_init,
1592 	.hw_fini = sdma_v7_0_hw_fini,
1593 	.suspend = sdma_v7_0_suspend,
1594 	.resume = sdma_v7_0_resume,
1595 	.is_idle = sdma_v7_0_is_idle,
1596 	.wait_for_idle = sdma_v7_0_wait_for_idle,
1597 	.soft_reset = sdma_v7_0_soft_reset,
1598 	.check_soft_reset = sdma_v7_0_check_soft_reset,
1599 	.set_clockgating_state = sdma_v7_0_set_clockgating_state,
1600 	.set_powergating_state = sdma_v7_0_set_powergating_state,
1601 	.get_clockgating_state = sdma_v7_0_get_clockgating_state,
1602 	.dump_ip_state = sdma_v7_0_dump_ip_state,
1603 	.print_ip_state = sdma_v7_0_print_ip_state,
1604 };
1605 
1606 static const struct amdgpu_ring_funcs sdma_v7_0_ring_funcs = {
1607 	.type = AMDGPU_RING_TYPE_SDMA,
1608 	.align_mask = 0xf,
1609 	.nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP),
1610 	.support_64bit_ptrs = true,
1611 	.secure_submission_supported = true,
1612 	.get_rptr = sdma_v7_0_ring_get_rptr,
1613 	.get_wptr = sdma_v7_0_ring_get_wptr,
1614 	.set_wptr = sdma_v7_0_ring_set_wptr,
1615 	.emit_frame_size =
1616 		5 + /* sdma_v7_0_ring_init_cond_exec */
1617 		6 + /* sdma_v7_0_ring_emit_hdp_flush */
1618 		6 + /* sdma_v7_0_ring_emit_pipeline_sync */
1619 		/* sdma_v7_0_ring_emit_vm_flush */
1620 		SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 +
1621 		SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6 +
1622 		10 + 10 + 10, /* sdma_v7_0_ring_emit_fence x3 for user fence, vm fence */
1623 	.emit_ib_size = 5 + 7 + 6, /* sdma_v7_0_ring_emit_ib */
1624 	.emit_ib = sdma_v7_0_ring_emit_ib,
1625 	.emit_mem_sync = sdma_v7_0_ring_emit_mem_sync,
1626 	.emit_fence = sdma_v7_0_ring_emit_fence,
1627 	.emit_pipeline_sync = sdma_v7_0_ring_emit_pipeline_sync,
1628 	.emit_vm_flush = sdma_v7_0_ring_emit_vm_flush,
1629 	.emit_hdp_flush = sdma_v7_0_ring_emit_hdp_flush,
1630 	.test_ring = sdma_v7_0_ring_test_ring,
1631 	.test_ib = sdma_v7_0_ring_test_ib,
1632 	.insert_nop = sdma_v7_0_ring_insert_nop,
1633 	.pad_ib = sdma_v7_0_ring_pad_ib,
1634 	.emit_wreg = sdma_v7_0_ring_emit_wreg,
1635 	.emit_reg_wait = sdma_v7_0_ring_emit_reg_wait,
1636 	.emit_reg_write_reg_wait = sdma_v7_0_ring_emit_reg_write_reg_wait,
1637 	.init_cond_exec = sdma_v7_0_ring_init_cond_exec,
1638 	.preempt_ib = sdma_v7_0_ring_preempt_ib,
1639 };
1640 
1641 static void sdma_v7_0_set_ring_funcs(struct amdgpu_device *adev)
1642 {
1643 	int i;
1644 
1645 	for (i = 0; i < adev->sdma.num_instances; i++) {
1646 		adev->sdma.instance[i].ring.funcs = &sdma_v7_0_ring_funcs;
1647 		adev->sdma.instance[i].ring.me = i;
1648 	}
1649 }
1650 
1651 static const struct amdgpu_irq_src_funcs sdma_v7_0_trap_irq_funcs = {
1652 	.set = sdma_v7_0_set_trap_irq_state,
1653 	.process = sdma_v7_0_process_trap_irq,
1654 };
1655 
1656 static const struct amdgpu_irq_src_funcs sdma_v7_0_illegal_inst_irq_funcs = {
1657 	.process = sdma_v7_0_process_illegal_inst_irq,
1658 };
1659 
1660 static void sdma_v7_0_set_irq_funcs(struct amdgpu_device *adev)
1661 {
1662 	adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_INSTANCE0 +
1663 					adev->sdma.num_instances;
1664 	adev->sdma.trap_irq.funcs = &sdma_v7_0_trap_irq_funcs;
1665 	adev->sdma.illegal_inst_irq.funcs = &sdma_v7_0_illegal_inst_irq_funcs;
1666 }
1667 
1668 /**
1669  * sdma_v7_0_emit_copy_buffer - copy buffer using the sDMA engine
1670  *
1671  * @ib: indirect buffer to fill with commands
1672  * @src_offset: src GPU address
1673  * @dst_offset: dst GPU address
1674  * @byte_count: number of bytes to xfer
1675  * @copy_flags: copy flags for the buffers
1676  *
1677  * Copy GPU buffers using the DMA engine.
1678  * Used by the amdgpu ttm implementation to move pages if
1679  * registered as the asic copy callback.
1680  */
1681 static void sdma_v7_0_emit_copy_buffer(struct amdgpu_ib *ib,
1682 				       uint64_t src_offset,
1683 				       uint64_t dst_offset,
1684 				       uint32_t byte_count,
1685 				       uint32_t copy_flags)
1686 {
1687 	uint32_t num_type, data_format, max_com;
1688 
1689 	max_com = AMDGPU_COPY_FLAGS_GET(copy_flags, MAX_COMPRESSED);
1690 	data_format = AMDGPU_COPY_FLAGS_GET(copy_flags, DATA_FORMAT);
1691 	num_type = AMDGPU_COPY_FLAGS_GET(copy_flags, NUMBER_TYPE);
1692 
1693 	ib->ptr[ib->length_dw++] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_COPY) |
1694 		SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR) |
1695 		SDMA_PKT_COPY_LINEAR_HEADER_TMZ((copy_flags & AMDGPU_COPY_FLAGS_TMZ) ? 1 : 0) |
1696 		SDMA_PKT_COPY_LINEAR_HEADER_CPV(1);
1697 
1698 	ib->ptr[ib->length_dw++] = byte_count - 1;
1699 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1700 	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1701 	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1702 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1703 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1704 
1705 	if ((copy_flags & (AMDGPU_COPY_FLAGS_READ_DECOMPRESSED | AMDGPU_COPY_FLAGS_WRITE_COMPRESSED)))
1706 		ib->ptr[ib->length_dw++] = SDMA_DCC_DATA_FORMAT(data_format) | SDMA_DCC_NUM_TYPE(num_type) |
1707 			((copy_flags & AMDGPU_COPY_FLAGS_READ_DECOMPRESSED) ? SDMA_DCC_READ_CM(2) : 0) |
1708 			((copy_flags & AMDGPU_COPY_FLAGS_WRITE_COMPRESSED) ? SDMA_DCC_WRITE_CM(1) : 0) |
1709 			SDMA_DCC_MAX_COM(max_com) | SDMA_DCC_MAX_UCOM(1);
1710 	else
1711 		ib->ptr[ib->length_dw++] = 0;
1712 }
1713 
1714 /**
1715  * sdma_v7_0_emit_fill_buffer - fill buffer using the sDMA engine
1716  *
1717  * @ib: indirect buffer to fill
1718  * @src_data: value to write to buffer
1719  * @dst_offset: dst GPU address
1720  * @byte_count: number of bytes to xfer
1721  *
1722  * Fill GPU buffers using the DMA engine.
1723  */
1724 static void sdma_v7_0_emit_fill_buffer(struct amdgpu_ib *ib,
1725 				       uint32_t src_data,
1726 				       uint64_t dst_offset,
1727 				       uint32_t byte_count)
1728 {
1729 	ib->ptr[ib->length_dw++] = SDMA_PKT_CONSTANT_FILL_HEADER_OP(SDMA_OP_CONST_FILL) |
1730 		SDMA_PKT_CONSTANT_FILL_HEADER_COMPRESS(1);
1731 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1732 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1733 	ib->ptr[ib->length_dw++] = src_data;
1734 	ib->ptr[ib->length_dw++] = byte_count - 1;
1735 }
1736 
1737 static const struct amdgpu_buffer_funcs sdma_v7_0_buffer_funcs = {
1738 	.copy_max_bytes = 0x400000,
1739 	.copy_num_dw = 8,
1740 	.emit_copy_buffer = sdma_v7_0_emit_copy_buffer,
1741 	.fill_max_bytes = 0x400000,
1742 	.fill_num_dw = 5,
1743 	.emit_fill_buffer = sdma_v7_0_emit_fill_buffer,
1744 };
1745 
1746 static void sdma_v7_0_set_buffer_funcs(struct amdgpu_device *adev)
1747 {
1748 	adev->mman.buffer_funcs = &sdma_v7_0_buffer_funcs;
1749 	adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1750 }
1751 
1752 static const struct amdgpu_vm_pte_funcs sdma_v7_0_vm_pte_funcs = {
1753 	.copy_pte_num_dw = 8,
1754 	.copy_pte = sdma_v7_0_vm_copy_pte,
1755 	.write_pte = sdma_v7_0_vm_write_pte,
1756 	.set_pte_pde = sdma_v7_0_vm_set_pte_pde,
1757 };
1758 
1759 static void sdma_v7_0_set_vm_pte_funcs(struct amdgpu_device *adev)
1760 {
1761 	unsigned i;
1762 
1763 	adev->vm_manager.vm_pte_funcs = &sdma_v7_0_vm_pte_funcs;
1764 	for (i = 0; i < adev->sdma.num_instances; i++) {
1765 		adev->vm_manager.vm_pte_scheds[i] =
1766 			&adev->sdma.instance[i].ring.sched;
1767 	}
1768 	adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
1769 }
1770 
1771 const struct amdgpu_ip_block_version sdma_v7_0_ip_block = {
1772 	.type = AMD_IP_BLOCK_TYPE_SDMA,
1773 	.major = 7,
1774 	.minor = 0,
1775 	.rev = 0,
1776 	.funcs = &sdma_v7_0_ip_funcs,
1777 };
1778