xref: /linux/drivers/gpu/drm/amd/amdgpu/sdma_v6_0.c (revision d53b8e36925256097a08d7cb749198d85cbf9b2b)
1 /*
2  * Copyright 2020 Advanced Micro Devices, Inc.
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice shall be included in
12  * all copies or substantial portions of the Software.
13  *
14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20  * OTHER DEALINGS IN THE SOFTWARE.
21  *
22  */
23 
24 #include <linux/delay.h>
25 #include <linux/firmware.h>
26 #include <linux/module.h>
27 #include <linux/pci.h>
28 
29 #include "amdgpu.h"
30 #include "amdgpu_ucode.h"
31 #include "amdgpu_trace.h"
32 
33 #include "gc/gc_11_0_0_offset.h"
34 #include "gc/gc_11_0_0_sh_mask.h"
35 #include "gc/gc_11_0_0_default.h"
36 #include "hdp/hdp_6_0_0_offset.h"
37 #include "ivsrcid/gfx/irqsrcs_gfx_11_0_0.h"
38 
39 #include "soc15_common.h"
40 #include "soc15.h"
41 #include "sdma_v6_0_0_pkt_open.h"
42 #include "nbio_v4_3.h"
43 #include "sdma_common.h"
44 #include "sdma_v6_0.h"
45 #include "v11_structs.h"
46 
47 MODULE_FIRMWARE("amdgpu/sdma_6_0_0.bin");
48 MODULE_FIRMWARE("amdgpu/sdma_6_0_1.bin");
49 MODULE_FIRMWARE("amdgpu/sdma_6_0_2.bin");
50 MODULE_FIRMWARE("amdgpu/sdma_6_0_3.bin");
51 MODULE_FIRMWARE("amdgpu/sdma_6_1_0.bin");
52 MODULE_FIRMWARE("amdgpu/sdma_6_1_1.bin");
53 MODULE_FIRMWARE("amdgpu/sdma_6_1_2.bin");
54 
55 #define SDMA1_REG_OFFSET 0x600
56 #define SDMA0_HYP_DEC_REG_START 0x5880
57 #define SDMA0_HYP_DEC_REG_END 0x589a
58 #define SDMA1_HYP_DEC_REG_OFFSET 0x20
59 
60 static const struct amdgpu_hwip_reg_entry sdma_reg_list_6_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_CHECKSUM),
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, regGRBM_STATUS2),
114 	SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_CHICKEN_BITS),
115 };
116 
117 static void sdma_v6_0_set_ring_funcs(struct amdgpu_device *adev);
118 static void sdma_v6_0_set_buffer_funcs(struct amdgpu_device *adev);
119 static void sdma_v6_0_set_vm_pte_funcs(struct amdgpu_device *adev);
120 static void sdma_v6_0_set_irq_funcs(struct amdgpu_device *adev);
121 static int sdma_v6_0_start(struct amdgpu_device *adev);
122 
123 static u32 sdma_v6_0_get_reg_offset(struct amdgpu_device *adev, u32 instance, u32 internal_offset)
124 {
125 	u32 base;
126 
127 	if (internal_offset >= SDMA0_HYP_DEC_REG_START &&
128 	    internal_offset <= SDMA0_HYP_DEC_REG_END) {
129 		base = adev->reg_offset[GC_HWIP][0][1];
130 		if (instance != 0)
131 			internal_offset += SDMA1_HYP_DEC_REG_OFFSET * instance;
132 	} else {
133 		base = adev->reg_offset[GC_HWIP][0][0];
134 		if (instance == 1)
135 			internal_offset += SDMA1_REG_OFFSET;
136 	}
137 
138 	return base + internal_offset;
139 }
140 
141 static unsigned sdma_v6_0_ring_init_cond_exec(struct amdgpu_ring *ring,
142 					      uint64_t addr)
143 {
144 	unsigned ret;
145 
146 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_COND_EXE));
147 	amdgpu_ring_write(ring, lower_32_bits(addr));
148 	amdgpu_ring_write(ring, upper_32_bits(addr));
149 	amdgpu_ring_write(ring, 1);
150 	/* this is the offset we need patch later */
151 	ret = ring->wptr & ring->buf_mask;
152 	/* insert dummy here and patch it later */
153 	amdgpu_ring_write(ring, 0);
154 
155 	return ret;
156 }
157 
158 /**
159  * sdma_v6_0_ring_get_rptr - get the current read pointer
160  *
161  * @ring: amdgpu ring pointer
162  *
163  * Get the current rptr from the hardware.
164  */
165 static uint64_t sdma_v6_0_ring_get_rptr(struct amdgpu_ring *ring)
166 {
167 	u64 *rptr;
168 
169 	/* XXX check if swapping is necessary on BE */
170 	rptr = (u64 *)ring->rptr_cpu_addr;
171 
172 	DRM_DEBUG("rptr before shift == 0x%016llx\n", *rptr);
173 	return ((*rptr) >> 2);
174 }
175 
176 /**
177  * sdma_v6_0_ring_get_wptr - get the current write pointer
178  *
179  * @ring: amdgpu ring pointer
180  *
181  * Get the current wptr from the hardware.
182  */
183 static uint64_t sdma_v6_0_ring_get_wptr(struct amdgpu_ring *ring)
184 {
185 	u64 wptr = 0;
186 
187 	if (ring->use_doorbell) {
188 		/* XXX check if swapping is necessary on BE */
189 		wptr = READ_ONCE(*((u64 *)ring->wptr_cpu_addr));
190 		DRM_DEBUG("wptr/doorbell before shift == 0x%016llx\n", wptr);
191 	}
192 
193 	return wptr >> 2;
194 }
195 
196 /**
197  * sdma_v6_0_ring_set_wptr - commit the write pointer
198  *
199  * @ring: amdgpu ring pointer
200  *
201  * Write the wptr back to the hardware.
202  */
203 static void sdma_v6_0_ring_set_wptr(struct amdgpu_ring *ring)
204 {
205 	struct amdgpu_device *adev = ring->adev;
206 
207 	if (ring->use_doorbell) {
208 		DRM_DEBUG("Using doorbell -- "
209 			  "wptr_offs == 0x%08x "
210 			  "lower_32_bits(ring->wptr) << 2 == 0x%08x "
211 			  "upper_32_bits(ring->wptr) << 2 == 0x%08x\n",
212 			  ring->wptr_offs,
213 			  lower_32_bits(ring->wptr << 2),
214 			  upper_32_bits(ring->wptr << 2));
215 		/* XXX check if swapping is necessary on BE */
216 		atomic64_set((atomic64_t *)ring->wptr_cpu_addr,
217 			     ring->wptr << 2);
218 		DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016llx)\n",
219 			  ring->doorbell_index, ring->wptr << 2);
220 		WDOORBELL64(ring->doorbell_index, ring->wptr << 2);
221 	} else {
222 		DRM_DEBUG("Not using doorbell -- "
223 			  "regSDMA%i_GFX_RB_WPTR == 0x%08x "
224 			  "regSDMA%i_GFX_RB_WPTR_HI == 0x%08x\n",
225 			  ring->me,
226 			  lower_32_bits(ring->wptr << 2),
227 			  ring->me,
228 			  upper_32_bits(ring->wptr << 2));
229 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev,
230 							     ring->me, regSDMA0_QUEUE0_RB_WPTR),
231 				lower_32_bits(ring->wptr << 2));
232 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev,
233 							     ring->me, regSDMA0_QUEUE0_RB_WPTR_HI),
234 				upper_32_bits(ring->wptr << 2));
235 	}
236 }
237 
238 static void sdma_v6_0_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
239 {
240 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
241 	int i;
242 
243 	for (i = 0; i < count; i++)
244 		if (sdma && sdma->burst_nop && (i == 0))
245 			amdgpu_ring_write(ring, ring->funcs->nop |
246 				SDMA_PKT_NOP_HEADER_COUNT(count - 1));
247 		else
248 			amdgpu_ring_write(ring, ring->funcs->nop);
249 }
250 
251 /*
252  * sdma_v6_0_ring_emit_ib - Schedule an IB on the DMA engine
253  *
254  * @ring: amdgpu ring pointer
255  * @ib: IB object to schedule
256  * @flags: unused
257  * @job: job to retrieve vmid from
258  *
259  * Schedule an IB in the DMA ring.
260  */
261 static void sdma_v6_0_ring_emit_ib(struct amdgpu_ring *ring,
262 				   struct amdgpu_job *job,
263 				   struct amdgpu_ib *ib,
264 				   uint32_t flags)
265 {
266 	unsigned vmid = AMDGPU_JOB_GET_VMID(job);
267 	uint64_t csa_mc_addr = amdgpu_sdma_get_csa_mc_addr(ring, vmid);
268 
269 	/* An IB packet must end on a 8 DW boundary--the next dword
270 	 * must be on a 8-dword boundary. Our IB packet below is 6
271 	 * dwords long, thus add x number of NOPs, such that, in
272 	 * modular arithmetic,
273 	 * wptr + 6 + x = 8k, k >= 0, which in C is,
274 	 * (wptr + 6 + x) % 8 = 0.
275 	 * The expression below, is a solution of x.
276 	 */
277 	sdma_v6_0_ring_insert_nop(ring, (2 - lower_32_bits(ring->wptr)) & 7);
278 
279 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_INDIRECT) |
280 			  SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf));
281 	/* base must be 32 byte aligned */
282 	amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
283 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
284 	amdgpu_ring_write(ring, ib->length_dw);
285 	amdgpu_ring_write(ring, lower_32_bits(csa_mc_addr));
286 	amdgpu_ring_write(ring, upper_32_bits(csa_mc_addr));
287 }
288 
289 /**
290  * sdma_v6_0_ring_emit_mem_sync - flush the IB by graphics cache rinse
291  *
292  * @ring: amdgpu ring pointer
293  *
294  * flush the IB by graphics cache rinse.
295  */
296 static void sdma_v6_0_ring_emit_mem_sync(struct amdgpu_ring *ring)
297 {
298         uint32_t gcr_cntl = SDMA_GCR_GL2_INV | SDMA_GCR_GL2_WB | SDMA_GCR_GLM_INV |
299                             SDMA_GCR_GL1_INV | SDMA_GCR_GLV_INV | SDMA_GCR_GLK_INV |
300                             SDMA_GCR_GLI_INV(1);
301 
302         /* flush entire cache L0/L1/L2, this can be optimized by performance requirement */
303         amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_GCR_REQ));
304         amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD1_BASE_VA_31_7(0));
305         amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD2_GCR_CONTROL_15_0(gcr_cntl) |
306                           SDMA_PKT_GCR_REQ_PAYLOAD2_BASE_VA_47_32(0));
307         amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD3_LIMIT_VA_31_7(0) |
308                           SDMA_PKT_GCR_REQ_PAYLOAD3_GCR_CONTROL_18_16(gcr_cntl >> 16));
309         amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD4_LIMIT_VA_47_32(0) |
310                           SDMA_PKT_GCR_REQ_PAYLOAD4_VMID(0));
311 }
312 
313 
314 /**
315  * sdma_v6_0_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
316  *
317  * @ring: amdgpu ring pointer
318  *
319  * Emit an hdp flush packet on the requested DMA ring.
320  */
321 static void sdma_v6_0_ring_emit_hdp_flush(struct amdgpu_ring *ring)
322 {
323 	struct amdgpu_device *adev = ring->adev;
324 	u32 ref_and_mask = 0;
325 	const struct nbio_hdp_flush_reg *nbio_hf_reg = adev->nbio.hdp_flush_reg;
326 
327 	ref_and_mask = nbio_hf_reg->ref_and_mask_sdma0 << ring->me;
328 
329 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_POLL_REGMEM) |
330 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) |
331 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
332 	amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_done_offset(adev)) << 2);
333 	amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_req_offset(adev)) << 2);
334 	amdgpu_ring_write(ring, ref_and_mask); /* reference */
335 	amdgpu_ring_write(ring, ref_and_mask); /* mask */
336 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
337 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
338 }
339 
340 /**
341  * sdma_v6_0_ring_emit_fence - emit a fence on the DMA ring
342  *
343  * @ring: amdgpu ring pointer
344  * @addr: address
345  * @seq: fence seq number
346  * @flags: fence flags
347  *
348  * Add a DMA fence packet to the ring to write
349  * the fence seq number and DMA trap packet to generate
350  * an interrupt if needed.
351  */
352 static void sdma_v6_0_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
353 				      unsigned flags)
354 {
355 	bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
356 	/* write the fence */
357 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_FENCE) |
358 			  SDMA_PKT_FENCE_HEADER_MTYPE(0x3)); /* Ucached(UC) */
359 	/* zero in first two bits */
360 	BUG_ON(addr & 0x3);
361 	amdgpu_ring_write(ring, lower_32_bits(addr));
362 	amdgpu_ring_write(ring, upper_32_bits(addr));
363 	amdgpu_ring_write(ring, lower_32_bits(seq));
364 
365 	/* optionally write high bits as well */
366 	if (write64bit) {
367 		addr += 4;
368 		amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_FENCE) |
369 				  SDMA_PKT_FENCE_HEADER_MTYPE(0x3));
370 		/* zero in first two bits */
371 		BUG_ON(addr & 0x3);
372 		amdgpu_ring_write(ring, lower_32_bits(addr));
373 		amdgpu_ring_write(ring, upper_32_bits(addr));
374 		amdgpu_ring_write(ring, upper_32_bits(seq));
375 	}
376 
377 	if (flags & AMDGPU_FENCE_FLAG_INT) {
378 		uint32_t ctx = ring->is_mes_queue ?
379 			(ring->hw_queue_id | AMDGPU_FENCE_MES_QUEUE_FLAG) : 0;
380 		/* generate an interrupt */
381 		amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_TRAP));
382 		amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(ctx));
383 	}
384 }
385 
386 /**
387  * sdma_v6_0_gfx_stop - stop the gfx async dma engines
388  *
389  * @adev: amdgpu_device pointer
390  *
391  * Stop the gfx async dma ring buffers.
392  */
393 static void sdma_v6_0_gfx_stop(struct amdgpu_device *adev)
394 {
395 	u32 rb_cntl, ib_cntl;
396 	int i;
397 
398 	for (i = 0; i < adev->sdma.num_instances; i++) {
399 		rb_cntl = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_CNTL));
400 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RB_ENABLE, 0);
401 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_CNTL), rb_cntl);
402 		ib_cntl = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_IB_CNTL));
403 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_QUEUE0_IB_CNTL, IB_ENABLE, 0);
404 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_IB_CNTL), ib_cntl);
405 	}
406 }
407 
408 /**
409  * sdma_v6_0_rlc_stop - stop the compute async dma engines
410  *
411  * @adev: amdgpu_device pointer
412  *
413  * Stop the compute async dma queues.
414  */
415 static void sdma_v6_0_rlc_stop(struct amdgpu_device *adev)
416 {
417 	/* XXX todo */
418 }
419 
420 /**
421  * sdma_v6_0_ctxempty_int_enable - enable or disable context empty interrupts
422  *
423  * @adev: amdgpu_device pointer
424  * @enable: enable/disable context switching due to queue empty conditions
425  *
426  * Enable or disable the async dma engines queue empty context switch.
427  */
428 static void sdma_v6_0_ctxempty_int_enable(struct amdgpu_device *adev, bool enable)
429 {
430 	u32 f32_cntl;
431 	int i;
432 
433 	if (!amdgpu_sriov_vf(adev)) {
434 		for (i = 0; i < adev->sdma.num_instances; i++) {
435 			f32_cntl = RREG32(sdma_v6_0_get_reg_offset(adev, i, regSDMA0_CNTL));
436 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
437 					CTXEMPTY_INT_ENABLE, enable ? 1 : 0);
438 			WREG32(sdma_v6_0_get_reg_offset(adev, i, regSDMA0_CNTL), f32_cntl);
439 		}
440 	}
441 }
442 
443 /**
444  * sdma_v6_0_enable - stop the async dma engines
445  *
446  * @adev: amdgpu_device pointer
447  * @enable: enable/disable the DMA MEs.
448  *
449  * Halt or unhalt the async dma engines.
450  */
451 static void sdma_v6_0_enable(struct amdgpu_device *adev, bool enable)
452 {
453 	u32 f32_cntl;
454 	int i;
455 
456 	if (!enable) {
457 		sdma_v6_0_gfx_stop(adev);
458 		sdma_v6_0_rlc_stop(adev);
459 	}
460 
461 	if (amdgpu_sriov_vf(adev))
462 		return;
463 
464 	for (i = 0; i < adev->sdma.num_instances; i++) {
465 		f32_cntl = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_F32_CNTL));
466 		f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, enable ? 0 : 1);
467 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_F32_CNTL), f32_cntl);
468 	}
469 }
470 
471 /**
472  * sdma_v6_0_gfx_resume - setup and start the async dma engines
473  *
474  * @adev: amdgpu_device pointer
475  *
476  * Set up the gfx DMA ring buffers and enable them.
477  * Returns 0 for success, error for failure.
478  */
479 static int sdma_v6_0_gfx_resume(struct amdgpu_device *adev)
480 {
481 	struct amdgpu_ring *ring;
482 	u32 rb_cntl, ib_cntl;
483 	u32 rb_bufsz;
484 	u32 doorbell;
485 	u32 doorbell_offset;
486 	u32 temp;
487 	u64 wptr_gpu_addr;
488 	int i, r;
489 
490 	for (i = 0; i < adev->sdma.num_instances; i++) {
491 		ring = &adev->sdma.instance[i].ring;
492 
493 		if (!amdgpu_sriov_vf(adev))
494 			WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_SEM_WAIT_FAIL_TIMER_CNTL), 0);
495 
496 		/* Set ring buffer size in dwords */
497 		rb_bufsz = order_base_2(ring->ring_size / 4);
498 		rb_cntl = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_CNTL));
499 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RB_SIZE, rb_bufsz);
500 #ifdef __BIG_ENDIAN
501 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RB_SWAP_ENABLE, 1);
502 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL,
503 					RPTR_WRITEBACK_SWAP_ENABLE, 1);
504 #endif
505 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RB_PRIV, 1);
506 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_CNTL), rb_cntl);
507 
508 		/* Initialize the ring buffer's read and write pointers */
509 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_RPTR), 0);
510 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_RPTR_HI), 0);
511 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR), 0);
512 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR_HI), 0);
513 
514 		/* setup the wptr shadow polling */
515 		wptr_gpu_addr = ring->wptr_gpu_addr;
516 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR_POLL_ADDR_LO),
517 		       lower_32_bits(wptr_gpu_addr));
518 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR_POLL_ADDR_HI),
519 		       upper_32_bits(wptr_gpu_addr));
520 
521 		/* set the wb address whether it's enabled or not */
522 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_RPTR_ADDR_HI),
523 		       upper_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFF);
524 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_RPTR_ADDR_LO),
525 		       lower_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFC);
526 
527 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
528 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, WPTR_POLL_ENABLE, 0);
529 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, F32_WPTR_POLL_ENABLE, 1);
530 
531 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_BASE), ring->gpu_addr >> 8);
532 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_BASE_HI), ring->gpu_addr >> 40);
533 
534 		ring->wptr = 0;
535 
536 		/* before programing wptr to a less value, need set minor_ptr_update first */
537 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_MINOR_PTR_UPDATE), 1);
538 
539 		if (!amdgpu_sriov_vf(adev)) { /* only bare-metal use register write for wptr */
540 			WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR), lower_32_bits(ring->wptr) << 2);
541 			WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR_HI), upper_32_bits(ring->wptr) << 2);
542 		}
543 
544 		doorbell = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_DOORBELL));
545 		doorbell_offset = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_DOORBELL_OFFSET));
546 
547 		if (ring->use_doorbell) {
548 			doorbell = REG_SET_FIELD(doorbell, SDMA0_QUEUE0_DOORBELL, ENABLE, 1);
549 			doorbell_offset = REG_SET_FIELD(doorbell_offset, SDMA0_QUEUE0_DOORBELL_OFFSET,
550 					OFFSET, ring->doorbell_index);
551 		} else {
552 			doorbell = REG_SET_FIELD(doorbell, SDMA0_QUEUE0_DOORBELL, ENABLE, 0);
553 		}
554 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_DOORBELL), doorbell);
555 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_DOORBELL_OFFSET), doorbell_offset);
556 
557 		if (i == 0)
558 			adev->nbio.funcs->sdma_doorbell_range(adev, i, ring->use_doorbell,
559 						      ring->doorbell_index,
560 						      adev->doorbell_index.sdma_doorbell_range * adev->sdma.num_instances);
561 
562 		if (amdgpu_sriov_vf(adev))
563 			sdma_v6_0_ring_set_wptr(ring);
564 
565 		/* set minor_ptr_update to 0 after wptr programed */
566 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_MINOR_PTR_UPDATE), 0);
567 
568 		/* Set up sdma hang watchdog */
569 		temp = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_WATCHDOG_CNTL));
570 		/* 100ms per unit */
571 		temp = REG_SET_FIELD(temp, SDMA0_WATCHDOG_CNTL, QUEUE_HANG_COUNT,
572 				     max(adev->usec_timeout/100000, 1));
573 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_WATCHDOG_CNTL), temp);
574 
575 		/* Set up RESP_MODE to non-copy addresses */
576 		temp = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UTCL1_CNTL));
577 		temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, RESP_MODE, 3);
578 		temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, REDO_DELAY, 9);
579 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UTCL1_CNTL), temp);
580 
581 		/* program default cache read and write policy */
582 		temp = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UTCL1_PAGE));
583 		/* clean read policy and write policy bits */
584 		temp &= 0xFF0FFF;
585 		temp |= ((CACHE_READ_POLICY_L2__DEFAULT << 12) |
586 			 (CACHE_WRITE_POLICY_L2__DEFAULT << 14) |
587 			 SDMA0_UTCL1_PAGE__LLC_NOALLOC_MASK);
588 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UTCL1_PAGE), temp);
589 
590 		if (!amdgpu_sriov_vf(adev)) {
591 			/* unhalt engine */
592 			temp = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_F32_CNTL));
593 			temp = REG_SET_FIELD(temp, SDMA0_F32_CNTL, HALT, 0);
594 			temp = REG_SET_FIELD(temp, SDMA0_F32_CNTL, TH1_RESET, 0);
595 			WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_F32_CNTL), temp);
596 		}
597 
598 		/* enable DMA RB */
599 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RB_ENABLE, 1);
600 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_CNTL), rb_cntl);
601 
602 		ib_cntl = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_IB_CNTL));
603 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_QUEUE0_IB_CNTL, IB_ENABLE, 1);
604 #ifdef __BIG_ENDIAN
605 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_QUEUE0_IB_CNTL, IB_SWAP_ENABLE, 1);
606 #endif
607 		/* enable DMA IBs */
608 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_IB_CNTL), ib_cntl);
609 
610 		if (amdgpu_sriov_vf(adev))
611 			sdma_v6_0_enable(adev, true);
612 
613 		r = amdgpu_ring_test_helper(ring);
614 		if (r)
615 			return r;
616 	}
617 
618 	return 0;
619 }
620 
621 /**
622  * sdma_v6_0_rlc_resume - setup and start the async dma engines
623  *
624  * @adev: amdgpu_device pointer
625  *
626  * Set up the compute DMA queues and enable them.
627  * Returns 0 for success, error for failure.
628  */
629 static int sdma_v6_0_rlc_resume(struct amdgpu_device *adev)
630 {
631 	return 0;
632 }
633 
634 /**
635  * sdma_v6_0_load_microcode - load the sDMA ME ucode
636  *
637  * @adev: amdgpu_device pointer
638  *
639  * Loads the sDMA0/1 ucode.
640  * Returns 0 for success, -EINVAL if the ucode is not available.
641  */
642 static int sdma_v6_0_load_microcode(struct amdgpu_device *adev)
643 {
644 	const struct sdma_firmware_header_v2_0 *hdr;
645 	const __le32 *fw_data;
646 	u32 fw_size;
647 	int i, j;
648 	bool use_broadcast;
649 
650 	/* halt the MEs */
651 	sdma_v6_0_enable(adev, false);
652 
653 	if (!adev->sdma.instance[0].fw)
654 		return -EINVAL;
655 
656 	/* use broadcast mode to load SDMA microcode by default */
657 	use_broadcast = true;
658 
659 	if (use_broadcast) {
660 		dev_info(adev->dev, "Use broadcast method to load SDMA firmware\n");
661 		/* load Control Thread microcode */
662 		hdr = (const struct sdma_firmware_header_v2_0 *)adev->sdma.instance[0].fw->data;
663 		amdgpu_ucode_print_sdma_hdr(&hdr->header);
664 		fw_size = le32_to_cpu(hdr->ctx_jt_offset + hdr->ctx_jt_size) / 4;
665 
666 		fw_data = (const __le32 *)
667 			(adev->sdma.instance[0].fw->data +
668 				le32_to_cpu(hdr->header.ucode_array_offset_bytes));
669 
670 		WREG32(sdma_v6_0_get_reg_offset(adev, 0, regSDMA0_BROADCAST_UCODE_ADDR), 0);
671 
672 		for (j = 0; j < fw_size; j++) {
673 			if (amdgpu_emu_mode == 1 && j % 500 == 0)
674 				msleep(1);
675 			WREG32(sdma_v6_0_get_reg_offset(adev, 0, regSDMA0_BROADCAST_UCODE_DATA), le32_to_cpup(fw_data++));
676 		}
677 
678 		/* load Context Switch microcode */
679 		fw_size = le32_to_cpu(hdr->ctl_jt_offset + hdr->ctl_jt_size) / 4;
680 
681 		fw_data = (const __le32 *)
682 			(adev->sdma.instance[0].fw->data +
683 				le32_to_cpu(hdr->ctl_ucode_offset));
684 
685 		WREG32(sdma_v6_0_get_reg_offset(adev, 0, regSDMA0_BROADCAST_UCODE_ADDR), 0x8000);
686 
687 		for (j = 0; j < fw_size; j++) {
688 			if (amdgpu_emu_mode == 1 && j % 500 == 0)
689 				msleep(1);
690 			WREG32(sdma_v6_0_get_reg_offset(adev, 0, regSDMA0_BROADCAST_UCODE_DATA), le32_to_cpup(fw_data++));
691 		}
692 	} else {
693 		dev_info(adev->dev, "Use legacy method to load SDMA firmware\n");
694 		for (i = 0; i < adev->sdma.num_instances; i++) {
695 			/* load Control Thread microcode */
696 			hdr = (const struct sdma_firmware_header_v2_0 *)adev->sdma.instance[0].fw->data;
697 			amdgpu_ucode_print_sdma_hdr(&hdr->header);
698 			fw_size = le32_to_cpu(hdr->ctx_jt_offset + hdr->ctx_jt_size) / 4;
699 
700 			fw_data = (const __le32 *)
701 				(adev->sdma.instance[0].fw->data +
702 					le32_to_cpu(hdr->header.ucode_array_offset_bytes));
703 
704 			WREG32(sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UCODE_ADDR), 0);
705 
706 			for (j = 0; j < fw_size; j++) {
707 				if (amdgpu_emu_mode == 1 && j % 500 == 0)
708 					msleep(1);
709 				WREG32(sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UCODE_DATA), le32_to_cpup(fw_data++));
710 			}
711 
712 			WREG32(sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UCODE_ADDR), adev->sdma.instance[0].fw_version);
713 
714 			/* load Context Switch microcode */
715 			fw_size = le32_to_cpu(hdr->ctl_jt_offset + hdr->ctl_jt_size) / 4;
716 
717 			fw_data = (const __le32 *)
718 				(adev->sdma.instance[0].fw->data +
719 					le32_to_cpu(hdr->ctl_ucode_offset));
720 
721 			WREG32(sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UCODE_ADDR), 0x8000);
722 
723 			for (j = 0; j < fw_size; j++) {
724 				if (amdgpu_emu_mode == 1 && j % 500 == 0)
725 					msleep(1);
726 				WREG32(sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UCODE_DATA), le32_to_cpup(fw_data++));
727 			}
728 
729 			WREG32(sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UCODE_ADDR), adev->sdma.instance[0].fw_version);
730 		}
731 	}
732 
733 	return 0;
734 }
735 
736 static int sdma_v6_0_soft_reset(void *handle)
737 {
738 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
739 	u32 tmp;
740 	int i;
741 
742 	sdma_v6_0_gfx_stop(adev);
743 
744 	for (i = 0; i < adev->sdma.num_instances; i++) {
745 		tmp = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_FREEZE));
746 		tmp |= SDMA0_FREEZE__FREEZE_MASK;
747 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_FREEZE), tmp);
748 		tmp = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_F32_CNTL));
749 		tmp |= SDMA0_F32_CNTL__HALT_MASK;
750 		tmp |= SDMA0_F32_CNTL__TH1_RESET_MASK;
751 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_F32_CNTL), tmp);
752 
753 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_PREEMPT), 0);
754 
755 		udelay(100);
756 
757 		tmp = GRBM_SOFT_RESET__SOFT_RESET_SDMA0_MASK << i;
758 		WREG32_SOC15(GC, 0, regGRBM_SOFT_RESET, tmp);
759 		tmp = RREG32_SOC15(GC, 0, regGRBM_SOFT_RESET);
760 
761 		udelay(100);
762 
763 		WREG32_SOC15(GC, 0, regGRBM_SOFT_RESET, 0);
764 		tmp = RREG32_SOC15(GC, 0, regGRBM_SOFT_RESET);
765 
766 		udelay(100);
767 	}
768 
769 	return sdma_v6_0_start(adev);
770 }
771 
772 static bool sdma_v6_0_check_soft_reset(void *handle)
773 {
774 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
775 	struct amdgpu_ring *ring;
776 	int i, r;
777 	long tmo = msecs_to_jiffies(1000);
778 
779 	for (i = 0; i < adev->sdma.num_instances; i++) {
780 		ring = &adev->sdma.instance[i].ring;
781 		r = amdgpu_ring_test_ib(ring, tmo);
782 		if (r)
783 			return true;
784 	}
785 
786 	return false;
787 }
788 
789 /**
790  * sdma_v6_0_start - setup and start the async dma engines
791  *
792  * @adev: amdgpu_device pointer
793  *
794  * Set up the DMA engines and enable them.
795  * Returns 0 for success, error for failure.
796  */
797 static int sdma_v6_0_start(struct amdgpu_device *adev)
798 {
799 	int r = 0;
800 
801 	if (amdgpu_sriov_vf(adev)) {
802 		sdma_v6_0_enable(adev, false);
803 
804 		/* set RB registers */
805 		r = sdma_v6_0_gfx_resume(adev);
806 		return r;
807 	}
808 
809 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
810 		r = sdma_v6_0_load_microcode(adev);
811 		if (r)
812 			return r;
813 
814 		/* The value of regSDMA_F32_CNTL is invalid the moment after loading fw */
815 		if (amdgpu_emu_mode == 1)
816 			msleep(1000);
817 	}
818 
819 	/* unhalt the MEs */
820 	sdma_v6_0_enable(adev, true);
821 	/* enable sdma ring preemption */
822 	sdma_v6_0_ctxempty_int_enable(adev, true);
823 
824 	/* start the gfx rings and rlc compute queues */
825 	r = sdma_v6_0_gfx_resume(adev);
826 	if (r)
827 		return r;
828 	r = sdma_v6_0_rlc_resume(adev);
829 
830 	return r;
831 }
832 
833 static int sdma_v6_0_mqd_init(struct amdgpu_device *adev, void *mqd,
834 			      struct amdgpu_mqd_prop *prop)
835 {
836 	struct v11_sdma_mqd *m = mqd;
837 	uint64_t wb_gpu_addr;
838 
839 	m->sdmax_rlcx_rb_cntl =
840 		order_base_2(prop->queue_size / 4) << SDMA0_QUEUE0_RB_CNTL__RB_SIZE__SHIFT |
841 		1 << SDMA0_QUEUE0_RB_CNTL__RPTR_WRITEBACK_ENABLE__SHIFT |
842 		4 << SDMA0_QUEUE0_RB_CNTL__RPTR_WRITEBACK_TIMER__SHIFT |
843 		1 << SDMA0_QUEUE0_RB_CNTL__F32_WPTR_POLL_ENABLE__SHIFT;
844 
845 	m->sdmax_rlcx_rb_base = lower_32_bits(prop->hqd_base_gpu_addr >> 8);
846 	m->sdmax_rlcx_rb_base_hi = upper_32_bits(prop->hqd_base_gpu_addr >> 8);
847 
848 	wb_gpu_addr = prop->wptr_gpu_addr;
849 	m->sdmax_rlcx_rb_wptr_poll_addr_lo = lower_32_bits(wb_gpu_addr);
850 	m->sdmax_rlcx_rb_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr);
851 
852 	wb_gpu_addr = prop->rptr_gpu_addr;
853 	m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits(wb_gpu_addr);
854 	m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits(wb_gpu_addr);
855 
856 	m->sdmax_rlcx_ib_cntl = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, 0,
857 							regSDMA0_QUEUE0_IB_CNTL));
858 
859 	m->sdmax_rlcx_doorbell_offset =
860 		prop->doorbell_index << SDMA0_QUEUE0_DOORBELL_OFFSET__OFFSET__SHIFT;
861 
862 	m->sdmax_rlcx_doorbell = REG_SET_FIELD(0, SDMA0_QUEUE0_DOORBELL, ENABLE, 1);
863 
864 	m->sdmax_rlcx_skip_cntl = 0;
865 	m->sdmax_rlcx_context_status = 0;
866 	m->sdmax_rlcx_doorbell_log = 0;
867 
868 	m->sdmax_rlcx_rb_aql_cntl = regSDMA0_QUEUE0_RB_AQL_CNTL_DEFAULT;
869 	m->sdmax_rlcx_dummy_reg = regSDMA0_QUEUE0_DUMMY_REG_DEFAULT;
870 
871 	return 0;
872 }
873 
874 static void sdma_v6_0_set_mqd_funcs(struct amdgpu_device *adev)
875 {
876 	adev->mqds[AMDGPU_HW_IP_DMA].mqd_size = sizeof(struct v11_sdma_mqd);
877 	adev->mqds[AMDGPU_HW_IP_DMA].init_mqd = sdma_v6_0_mqd_init;
878 }
879 
880 /**
881  * sdma_v6_0_ring_test_ring - simple async dma engine test
882  *
883  * @ring: amdgpu_ring structure holding ring information
884  *
885  * Test the DMA engine by writing using it to write an
886  * value to memory.
887  * Returns 0 for success, error for failure.
888  */
889 static int sdma_v6_0_ring_test_ring(struct amdgpu_ring *ring)
890 {
891 	struct amdgpu_device *adev = ring->adev;
892 	unsigned i;
893 	unsigned index;
894 	int r;
895 	u32 tmp;
896 	u64 gpu_addr;
897 	volatile uint32_t *cpu_ptr = NULL;
898 
899 	tmp = 0xCAFEDEAD;
900 
901 	if (ring->is_mes_queue) {
902 		uint32_t offset = 0;
903 		offset = amdgpu_mes_ctx_get_offs(ring,
904 					 AMDGPU_MES_CTX_PADDING_OFFS);
905 		gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
906 		cpu_ptr = amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
907 		*cpu_ptr = tmp;
908 	} else {
909 		r = amdgpu_device_wb_get(adev, &index);
910 		if (r) {
911 			dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
912 			return r;
913 		}
914 
915 		gpu_addr = adev->wb.gpu_addr + (index * 4);
916 		adev->wb.wb[index] = cpu_to_le32(tmp);
917 	}
918 
919 	r = amdgpu_ring_alloc(ring, 5);
920 	if (r) {
921 		DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
922 		if (!ring->is_mes_queue)
923 			amdgpu_device_wb_free(adev, index);
924 		return r;
925 	}
926 
927 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_WRITE) |
928 			  SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
929 	amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
930 	amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
931 	amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0));
932 	amdgpu_ring_write(ring, 0xDEADBEEF);
933 	amdgpu_ring_commit(ring);
934 
935 	for (i = 0; i < adev->usec_timeout; i++) {
936 		if (ring->is_mes_queue)
937 			tmp = le32_to_cpu(*cpu_ptr);
938 		else
939 			tmp = le32_to_cpu(adev->wb.wb[index]);
940 		if (tmp == 0xDEADBEEF)
941 			break;
942 		if (amdgpu_emu_mode == 1)
943 			msleep(1);
944 		else
945 			udelay(1);
946 	}
947 
948 	if (i >= adev->usec_timeout)
949 		r = -ETIMEDOUT;
950 
951 	if (!ring->is_mes_queue)
952 		amdgpu_device_wb_free(adev, index);
953 
954 	return r;
955 }
956 
957 /*
958  * sdma_v6_0_ring_test_ib - test an IB on the DMA engine
959  *
960  * @ring: amdgpu_ring structure holding ring information
961  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
962  *
963  * Test a simple IB in the DMA ring.
964  * Returns 0 on success, error on failure.
965  */
966 static int sdma_v6_0_ring_test_ib(struct amdgpu_ring *ring, long timeout)
967 {
968 	struct amdgpu_device *adev = ring->adev;
969 	struct amdgpu_ib ib;
970 	struct dma_fence *f = NULL;
971 	unsigned index;
972 	long r;
973 	u32 tmp = 0;
974 	u64 gpu_addr;
975 	volatile uint32_t *cpu_ptr = NULL;
976 
977 	tmp = 0xCAFEDEAD;
978 	memset(&ib, 0, sizeof(ib));
979 
980 	if (ring->is_mes_queue) {
981 		uint32_t offset = 0;
982 		offset = amdgpu_mes_ctx_get_offs(ring, AMDGPU_MES_CTX_IB_OFFS);
983 		ib.gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
984 		ib.ptr = (void *)amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
985 
986 		offset = amdgpu_mes_ctx_get_offs(ring,
987 					 AMDGPU_MES_CTX_PADDING_OFFS);
988 		gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
989 		cpu_ptr = amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
990 		*cpu_ptr = tmp;
991 	} else {
992 		r = amdgpu_device_wb_get(adev, &index);
993 		if (r) {
994 			dev_err(adev->dev, "(%ld) failed to allocate wb slot\n", r);
995 			return r;
996 		}
997 
998 		gpu_addr = adev->wb.gpu_addr + (index * 4);
999 		adev->wb.wb[index] = cpu_to_le32(tmp);
1000 
1001 		r = amdgpu_ib_get(adev, NULL, 256, AMDGPU_IB_POOL_DIRECT, &ib);
1002 		if (r) {
1003 			DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r);
1004 			goto err0;
1005 		}
1006 	}
1007 
1008 	ib.ptr[0] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_WRITE) |
1009 		SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1010 	ib.ptr[1] = lower_32_bits(gpu_addr);
1011 	ib.ptr[2] = upper_32_bits(gpu_addr);
1012 	ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0);
1013 	ib.ptr[4] = 0xDEADBEEF;
1014 	ib.ptr[5] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1015 	ib.ptr[6] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1016 	ib.ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1017 	ib.length_dw = 8;
1018 
1019 	r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
1020 	if (r)
1021 		goto err1;
1022 
1023 	r = dma_fence_wait_timeout(f, false, timeout);
1024 	if (r == 0) {
1025 		DRM_ERROR("amdgpu: IB test timed out\n");
1026 		r = -ETIMEDOUT;
1027 		goto err1;
1028 	} else if (r < 0) {
1029 		DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
1030 		goto err1;
1031 	}
1032 
1033 	if (ring->is_mes_queue)
1034 		tmp = le32_to_cpu(*cpu_ptr);
1035 	else
1036 		tmp = le32_to_cpu(adev->wb.wb[index]);
1037 
1038 	if (tmp == 0xDEADBEEF)
1039 		r = 0;
1040 	else
1041 		r = -EINVAL;
1042 
1043 err1:
1044 	amdgpu_ib_free(adev, &ib, NULL);
1045 	dma_fence_put(f);
1046 err0:
1047 	if (!ring->is_mes_queue)
1048 		amdgpu_device_wb_free(adev, index);
1049 	return r;
1050 }
1051 
1052 
1053 /**
1054  * sdma_v6_0_vm_copy_pte - update PTEs by copying them from the GART
1055  *
1056  * @ib: indirect buffer to fill with commands
1057  * @pe: addr of the page entry
1058  * @src: src addr to copy from
1059  * @count: number of page entries to update
1060  *
1061  * Update PTEs by copying them from the GART using sDMA.
1062  */
1063 static void sdma_v6_0_vm_copy_pte(struct amdgpu_ib *ib,
1064 				  uint64_t pe, uint64_t src,
1065 				  unsigned count)
1066 {
1067 	unsigned bytes = count * 8;
1068 
1069 	ib->ptr[ib->length_dw++] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_COPY) |
1070 		SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
1071 	ib->ptr[ib->length_dw++] = bytes - 1;
1072 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1073 	ib->ptr[ib->length_dw++] = lower_32_bits(src);
1074 	ib->ptr[ib->length_dw++] = upper_32_bits(src);
1075 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1076 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1077 
1078 }
1079 
1080 /**
1081  * sdma_v6_0_vm_write_pte - update PTEs by writing them manually
1082  *
1083  * @ib: indirect buffer to fill with commands
1084  * @pe: addr of the page entry
1085  * @value: dst addr to write into pe
1086  * @count: number of page entries to update
1087  * @incr: increase next addr by incr bytes
1088  *
1089  * Update PTEs by writing them manually using sDMA.
1090  */
1091 static void sdma_v6_0_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
1092 				   uint64_t value, unsigned count,
1093 				   uint32_t incr)
1094 {
1095 	unsigned ndw = count * 2;
1096 
1097 	ib->ptr[ib->length_dw++] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_WRITE) |
1098 		SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1099 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1100 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1101 	ib->ptr[ib->length_dw++] = ndw - 1;
1102 	for (; ndw > 0; ndw -= 2) {
1103 		ib->ptr[ib->length_dw++] = lower_32_bits(value);
1104 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
1105 		value += incr;
1106 	}
1107 }
1108 
1109 /**
1110  * sdma_v6_0_vm_set_pte_pde - update the page tables using sDMA
1111  *
1112  * @ib: indirect buffer to fill with commands
1113  * @pe: addr of the page entry
1114  * @addr: dst addr to write into pe
1115  * @count: number of page entries to update
1116  * @incr: increase next addr by incr bytes
1117  * @flags: access flags
1118  *
1119  * Update the page tables using sDMA.
1120  */
1121 static void sdma_v6_0_vm_set_pte_pde(struct amdgpu_ib *ib,
1122 				     uint64_t pe,
1123 				     uint64_t addr, unsigned count,
1124 				     uint32_t incr, uint64_t flags)
1125 {
1126 	/* for physically contiguous pages (vram) */
1127 	ib->ptr[ib->length_dw++] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_PTEPDE);
1128 	ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
1129 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1130 	ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
1131 	ib->ptr[ib->length_dw++] = upper_32_bits(flags);
1132 	ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
1133 	ib->ptr[ib->length_dw++] = upper_32_bits(addr);
1134 	ib->ptr[ib->length_dw++] = incr; /* increment size */
1135 	ib->ptr[ib->length_dw++] = 0;
1136 	ib->ptr[ib->length_dw++] = count - 1; /* number of entries */
1137 }
1138 
1139 /*
1140  * sdma_v6_0_ring_pad_ib - pad the IB
1141  * @ib: indirect buffer to fill with padding
1142  * @ring: amdgpu ring pointer
1143  *
1144  * Pad the IB with NOPs to a boundary multiple of 8.
1145  */
1146 static void sdma_v6_0_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
1147 {
1148 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
1149 	u32 pad_count;
1150 	int i;
1151 
1152 	pad_count = (-ib->length_dw) & 0x7;
1153 	for (i = 0; i < pad_count; i++)
1154 		if (sdma && sdma->burst_nop && (i == 0))
1155 			ib->ptr[ib->length_dw++] =
1156 				SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_NOP) |
1157 				SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
1158 		else
1159 			ib->ptr[ib->length_dw++] =
1160 				SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_NOP);
1161 }
1162 
1163 /**
1164  * sdma_v6_0_ring_emit_pipeline_sync - sync the pipeline
1165  *
1166  * @ring: amdgpu_ring pointer
1167  *
1168  * Make sure all previous operations are completed (CIK).
1169  */
1170 static void sdma_v6_0_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
1171 {
1172 	uint32_t seq = ring->fence_drv.sync_seq;
1173 	uint64_t addr = ring->fence_drv.gpu_addr;
1174 
1175 	/* wait for idle */
1176 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1177 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1178 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
1179 			  SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1));
1180 	amdgpu_ring_write(ring, addr & 0xfffffffc);
1181 	amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
1182 	amdgpu_ring_write(ring, seq); /* reference */
1183 	amdgpu_ring_write(ring, 0xffffffff); /* mask */
1184 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1185 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */
1186 }
1187 
1188 /*
1189  * sdma_v6_0_ring_emit_vm_flush - vm flush using sDMA
1190  *
1191  * @ring: amdgpu_ring pointer
1192  * @vmid: vmid number to use
1193  * @pd_addr: address
1194  *
1195  * Update the page table base and flush the VM TLB
1196  * using sDMA.
1197  */
1198 static void sdma_v6_0_ring_emit_vm_flush(struct amdgpu_ring *ring,
1199 					 unsigned vmid, uint64_t pd_addr)
1200 {
1201 	struct amdgpu_vmhub *hub = &ring->adev->vmhub[ring->vm_hub];
1202 	uint32_t req = hub->vmhub_funcs->get_invalidate_req(vmid, 0);
1203 
1204 	/* Update the PD address for this VMID. */
1205 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_lo32 +
1206 			      (hub->ctx_addr_distance * vmid),
1207 			      lower_32_bits(pd_addr));
1208 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_hi32 +
1209 			      (hub->ctx_addr_distance * vmid),
1210 			      upper_32_bits(pd_addr));
1211 
1212 	/* Trigger invalidation. */
1213 	amdgpu_ring_write(ring,
1214 			  SDMA_PKT_VM_INVALIDATION_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1215 			  SDMA_PKT_VM_INVALIDATION_HEADER_SUB_OP(SDMA_SUBOP_VM_INVALIDATION) |
1216 			  SDMA_PKT_VM_INVALIDATION_HEADER_GFX_ENG_ID(ring->vm_inv_eng) |
1217 			  SDMA_PKT_VM_INVALIDATION_HEADER_MM_ENG_ID(0x1f));
1218 	amdgpu_ring_write(ring, req);
1219 	amdgpu_ring_write(ring, 0xFFFFFFFF);
1220 	amdgpu_ring_write(ring,
1221 			  SDMA_PKT_VM_INVALIDATION_ADDRESSRANGEHI_INVALIDATEACK(1 << vmid) |
1222 			  SDMA_PKT_VM_INVALIDATION_ADDRESSRANGEHI_ADDRESSRANGEHI(0x1F));
1223 }
1224 
1225 static void sdma_v6_0_ring_emit_wreg(struct amdgpu_ring *ring,
1226 				     uint32_t reg, uint32_t val)
1227 {
1228 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_SRBM_WRITE) |
1229 			  SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
1230 	amdgpu_ring_write(ring, reg);
1231 	amdgpu_ring_write(ring, val);
1232 }
1233 
1234 static void sdma_v6_0_ring_emit_reg_wait(struct amdgpu_ring *ring, uint32_t reg,
1235 					 uint32_t val, uint32_t mask)
1236 {
1237 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1238 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1239 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* equal */
1240 	amdgpu_ring_write(ring, reg << 2);
1241 	amdgpu_ring_write(ring, 0);
1242 	amdgpu_ring_write(ring, val); /* reference */
1243 	amdgpu_ring_write(ring, mask); /* mask */
1244 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1245 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10));
1246 }
1247 
1248 static void sdma_v6_0_ring_emit_reg_write_reg_wait(struct amdgpu_ring *ring,
1249 						   uint32_t reg0, uint32_t reg1,
1250 						   uint32_t ref, uint32_t mask)
1251 {
1252 	amdgpu_ring_emit_wreg(ring, reg0, ref);
1253 	/* wait for a cycle to reset vm_inv_eng*_ack */
1254 	amdgpu_ring_emit_reg_wait(ring, reg0, 0, 0);
1255 	amdgpu_ring_emit_reg_wait(ring, reg1, mask, mask);
1256 }
1257 
1258 static struct amdgpu_sdma_ras sdma_v6_0_3_ras = {
1259 	.ras_block = {
1260 		.ras_late_init = amdgpu_ras_block_late_init,
1261 	},
1262 };
1263 
1264 static void sdma_v6_0_set_ras_funcs(struct amdgpu_device *adev)
1265 {
1266 	switch (amdgpu_ip_version(adev, SDMA0_HWIP, 0)) {
1267 	case IP_VERSION(6, 0, 3):
1268 		adev->sdma.ras = &sdma_v6_0_3_ras;
1269 		break;
1270 	default:
1271 		break;
1272 	}
1273 }
1274 
1275 static int sdma_v6_0_early_init(void *handle)
1276 {
1277 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1278 	int r;
1279 
1280 	r = amdgpu_sdma_init_microcode(adev, 0, true);
1281 	if (r)
1282 		return r;
1283 
1284 	sdma_v6_0_set_ring_funcs(adev);
1285 	sdma_v6_0_set_buffer_funcs(adev);
1286 	sdma_v6_0_set_vm_pte_funcs(adev);
1287 	sdma_v6_0_set_irq_funcs(adev);
1288 	sdma_v6_0_set_mqd_funcs(adev);
1289 	sdma_v6_0_set_ras_funcs(adev);
1290 
1291 	return 0;
1292 }
1293 
1294 static int sdma_v6_0_sw_init(void *handle)
1295 {
1296 	struct amdgpu_ring *ring;
1297 	int r, i;
1298 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1299 	uint32_t reg_count = ARRAY_SIZE(sdma_reg_list_6_0);
1300 	uint32_t *ptr;
1301 
1302 	/* SDMA trap event */
1303 	r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_GFX,
1304 			      GFX_11_0_0__SRCID__SDMA_TRAP,
1305 			      &adev->sdma.trap_irq);
1306 	if (r)
1307 		return r;
1308 
1309 	for (i = 0; i < adev->sdma.num_instances; i++) {
1310 		ring = &adev->sdma.instance[i].ring;
1311 		ring->ring_obj = NULL;
1312 		ring->use_doorbell = true;
1313 		ring->me = i;
1314 
1315 		DRM_DEBUG("SDMA %d use_doorbell being set to: [%s]\n", i,
1316 				ring->use_doorbell?"true":"false");
1317 
1318 		ring->doorbell_index =
1319 			(adev->doorbell_index.sdma_engine[i] << 1); // get DWORD offset
1320 
1321 		ring->vm_hub = AMDGPU_GFXHUB(0);
1322 		sprintf(ring->name, "sdma%d", i);
1323 		r = amdgpu_ring_init(adev, ring, 1024,
1324 				     &adev->sdma.trap_irq,
1325 				     AMDGPU_SDMA_IRQ_INSTANCE0 + i,
1326 				     AMDGPU_RING_PRIO_DEFAULT, NULL);
1327 		if (r)
1328 			return r;
1329 	}
1330 
1331 	if (amdgpu_sdma_ras_sw_init(adev)) {
1332 		dev_err(adev->dev, "Failed to initialize sdma ras block!\n");
1333 		return -EINVAL;
1334 	}
1335 
1336 	/* Allocate memory for SDMA IP Dump buffer */
1337 	ptr = kcalloc(adev->sdma.num_instances * reg_count, sizeof(uint32_t), GFP_KERNEL);
1338 	if (ptr)
1339 		adev->sdma.ip_dump = ptr;
1340 	else
1341 		DRM_ERROR("Failed to allocated memory for SDMA IP Dump\n");
1342 
1343 	return r;
1344 }
1345 
1346 static int sdma_v6_0_sw_fini(void *handle)
1347 {
1348 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1349 	int i;
1350 
1351 	for (i = 0; i < adev->sdma.num_instances; i++)
1352 		amdgpu_ring_fini(&adev->sdma.instance[i].ring);
1353 
1354 	amdgpu_sdma_destroy_inst_ctx(adev, true);
1355 
1356 	kfree(adev->sdma.ip_dump);
1357 
1358 	return 0;
1359 }
1360 
1361 static int sdma_v6_0_hw_init(void *handle)
1362 {
1363 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1364 
1365 	return sdma_v6_0_start(adev);
1366 }
1367 
1368 static int sdma_v6_0_hw_fini(void *handle)
1369 {
1370 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1371 
1372 	if (amdgpu_sriov_vf(adev))
1373 		return 0;
1374 
1375 	sdma_v6_0_ctxempty_int_enable(adev, false);
1376 	sdma_v6_0_enable(adev, false);
1377 
1378 	return 0;
1379 }
1380 
1381 static int sdma_v6_0_suspend(void *handle)
1382 {
1383 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1384 
1385 	return sdma_v6_0_hw_fini(adev);
1386 }
1387 
1388 static int sdma_v6_0_resume(void *handle)
1389 {
1390 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1391 
1392 	return sdma_v6_0_hw_init(adev);
1393 }
1394 
1395 static bool sdma_v6_0_is_idle(void *handle)
1396 {
1397 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1398 	u32 i;
1399 
1400 	for (i = 0; i < adev->sdma.num_instances; i++) {
1401 		u32 tmp = RREG32(sdma_v6_0_get_reg_offset(adev, i, regSDMA0_STATUS_REG));
1402 
1403 		if (!(tmp & SDMA0_STATUS_REG__IDLE_MASK))
1404 			return false;
1405 	}
1406 
1407 	return true;
1408 }
1409 
1410 static int sdma_v6_0_wait_for_idle(void *handle)
1411 {
1412 	unsigned i;
1413 	u32 sdma0, sdma1;
1414 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1415 
1416 	for (i = 0; i < adev->usec_timeout; i++) {
1417 		sdma0 = RREG32(sdma_v6_0_get_reg_offset(adev, 0, regSDMA0_STATUS_REG));
1418 		sdma1 = RREG32(sdma_v6_0_get_reg_offset(adev, 1, regSDMA0_STATUS_REG));
1419 
1420 		if (sdma0 & sdma1 & SDMA0_STATUS_REG__IDLE_MASK)
1421 			return 0;
1422 		udelay(1);
1423 	}
1424 	return -ETIMEDOUT;
1425 }
1426 
1427 static int sdma_v6_0_ring_preempt_ib(struct amdgpu_ring *ring)
1428 {
1429 	int i, r = 0;
1430 	struct amdgpu_device *adev = ring->adev;
1431 	u32 index = 0;
1432 	u64 sdma_gfx_preempt;
1433 
1434 	amdgpu_sdma_get_index_from_ring(ring, &index);
1435 	sdma_gfx_preempt =
1436 		sdma_v6_0_get_reg_offset(adev, index, regSDMA0_QUEUE0_PREEMPT);
1437 
1438 	/* assert preemption condition */
1439 	amdgpu_ring_set_preempt_cond_exec(ring, false);
1440 
1441 	/* emit the trailing fence */
1442 	ring->trail_seq += 1;
1443 	amdgpu_ring_alloc(ring, 10);
1444 	sdma_v6_0_ring_emit_fence(ring, ring->trail_fence_gpu_addr,
1445 				  ring->trail_seq, 0);
1446 	amdgpu_ring_commit(ring);
1447 
1448 	/* assert IB preemption */
1449 	WREG32(sdma_gfx_preempt, 1);
1450 
1451 	/* poll the trailing fence */
1452 	for (i = 0; i < adev->usec_timeout; i++) {
1453 		if (ring->trail_seq ==
1454 		    le32_to_cpu(*(ring->trail_fence_cpu_addr)))
1455 			break;
1456 		udelay(1);
1457 	}
1458 
1459 	if (i >= adev->usec_timeout) {
1460 		r = -EINVAL;
1461 		DRM_ERROR("ring %d failed to be preempted\n", ring->idx);
1462 	}
1463 
1464 	/* deassert IB preemption */
1465 	WREG32(sdma_gfx_preempt, 0);
1466 
1467 	/* deassert the preemption condition */
1468 	amdgpu_ring_set_preempt_cond_exec(ring, true);
1469 	return r;
1470 }
1471 
1472 static int sdma_v6_0_set_trap_irq_state(struct amdgpu_device *adev,
1473 					struct amdgpu_irq_src *source,
1474 					unsigned type,
1475 					enum amdgpu_interrupt_state state)
1476 {
1477 	u32 sdma_cntl;
1478 
1479 	u32 reg_offset = sdma_v6_0_get_reg_offset(adev, type, regSDMA0_CNTL);
1480 
1481 	if (!amdgpu_sriov_vf(adev)) {
1482 		sdma_cntl = RREG32(reg_offset);
1483 		sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE,
1484 				state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0);
1485 		WREG32(reg_offset, sdma_cntl);
1486 	}
1487 
1488 	return 0;
1489 }
1490 
1491 static int sdma_v6_0_process_trap_irq(struct amdgpu_device *adev,
1492 				      struct amdgpu_irq_src *source,
1493 				      struct amdgpu_iv_entry *entry)
1494 {
1495 	int instances, queue;
1496 	uint32_t mes_queue_id = entry->src_data[0];
1497 
1498 	DRM_DEBUG("IH: SDMA trap\n");
1499 
1500 	if (adev->enable_mes && (mes_queue_id & AMDGPU_FENCE_MES_QUEUE_FLAG)) {
1501 		struct amdgpu_mes_queue *queue;
1502 
1503 		mes_queue_id &= AMDGPU_FENCE_MES_QUEUE_ID_MASK;
1504 
1505 		spin_lock(&adev->mes.queue_id_lock);
1506 		queue = idr_find(&adev->mes.queue_id_idr, mes_queue_id);
1507 		if (queue) {
1508 			DRM_DEBUG("process smda queue id = %d\n", mes_queue_id);
1509 			amdgpu_fence_process(queue->ring);
1510 		}
1511 		spin_unlock(&adev->mes.queue_id_lock);
1512 		return 0;
1513 	}
1514 
1515 	queue = entry->ring_id & 0xf;
1516 	instances = (entry->ring_id & 0xf0) >> 4;
1517 	if (instances > 1) {
1518 		DRM_ERROR("IH: wrong ring_ID detected, as wrong sdma instance\n");
1519 		return -EINVAL;
1520 	}
1521 
1522 	switch (entry->client_id) {
1523 	case SOC21_IH_CLIENTID_GFX:
1524 		switch (queue) {
1525 		case 0:
1526 			amdgpu_fence_process(&adev->sdma.instance[instances].ring);
1527 			break;
1528 		default:
1529 			break;
1530 		}
1531 		break;
1532 	}
1533 	return 0;
1534 }
1535 
1536 static int sdma_v6_0_process_illegal_inst_irq(struct amdgpu_device *adev,
1537 					      struct amdgpu_irq_src *source,
1538 					      struct amdgpu_iv_entry *entry)
1539 {
1540 	return 0;
1541 }
1542 
1543 static int sdma_v6_0_set_clockgating_state(void *handle,
1544 					   enum amd_clockgating_state state)
1545 {
1546 	return 0;
1547 }
1548 
1549 static int sdma_v6_0_set_powergating_state(void *handle,
1550 					  enum amd_powergating_state state)
1551 {
1552 	return 0;
1553 }
1554 
1555 static void sdma_v6_0_get_clockgating_state(void *handle, u64 *flags)
1556 {
1557 }
1558 
1559 static void sdma_v6_0_print_ip_state(void *handle, struct drm_printer *p)
1560 {
1561 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1562 	int i, j;
1563 	uint32_t reg_count = ARRAY_SIZE(sdma_reg_list_6_0);
1564 	uint32_t instance_offset;
1565 
1566 	if (!adev->sdma.ip_dump)
1567 		return;
1568 
1569 	drm_printf(p, "num_instances:%d\n", adev->sdma.num_instances);
1570 	for (i = 0; i < adev->sdma.num_instances; i++) {
1571 		instance_offset = i * reg_count;
1572 		drm_printf(p, "\nInstance:%d\n", i);
1573 
1574 		for (j = 0; j < reg_count; j++)
1575 			drm_printf(p, "%-50s \t 0x%08x\n", sdma_reg_list_6_0[j].reg_name,
1576 				   adev->sdma.ip_dump[instance_offset + j]);
1577 	}
1578 }
1579 
1580 static void sdma_v6_0_dump_ip_state(void *handle)
1581 {
1582 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1583 	int i, j;
1584 	uint32_t instance_offset;
1585 	uint32_t reg_count = ARRAY_SIZE(sdma_reg_list_6_0);
1586 
1587 	if (!adev->sdma.ip_dump)
1588 		return;
1589 
1590 	amdgpu_gfx_off_ctrl(adev, false);
1591 	for (i = 0; i < adev->sdma.num_instances; i++) {
1592 		instance_offset = i * reg_count;
1593 		for (j = 0; j < reg_count; j++)
1594 			adev->sdma.ip_dump[instance_offset + j] =
1595 				RREG32(sdma_v6_0_get_reg_offset(adev, i,
1596 				       sdma_reg_list_6_0[j].reg_offset));
1597 	}
1598 	amdgpu_gfx_off_ctrl(adev, true);
1599 }
1600 
1601 const struct amd_ip_funcs sdma_v6_0_ip_funcs = {
1602 	.name = "sdma_v6_0",
1603 	.early_init = sdma_v6_0_early_init,
1604 	.late_init = NULL,
1605 	.sw_init = sdma_v6_0_sw_init,
1606 	.sw_fini = sdma_v6_0_sw_fini,
1607 	.hw_init = sdma_v6_0_hw_init,
1608 	.hw_fini = sdma_v6_0_hw_fini,
1609 	.suspend = sdma_v6_0_suspend,
1610 	.resume = sdma_v6_0_resume,
1611 	.is_idle = sdma_v6_0_is_idle,
1612 	.wait_for_idle = sdma_v6_0_wait_for_idle,
1613 	.soft_reset = sdma_v6_0_soft_reset,
1614 	.check_soft_reset = sdma_v6_0_check_soft_reset,
1615 	.set_clockgating_state = sdma_v6_0_set_clockgating_state,
1616 	.set_powergating_state = sdma_v6_0_set_powergating_state,
1617 	.get_clockgating_state = sdma_v6_0_get_clockgating_state,
1618 	.dump_ip_state = sdma_v6_0_dump_ip_state,
1619 	.print_ip_state = sdma_v6_0_print_ip_state,
1620 };
1621 
1622 static const struct amdgpu_ring_funcs sdma_v6_0_ring_funcs = {
1623 	.type = AMDGPU_RING_TYPE_SDMA,
1624 	.align_mask = 0xf,
1625 	.nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP),
1626 	.support_64bit_ptrs = true,
1627 	.secure_submission_supported = true,
1628 	.get_rptr = sdma_v6_0_ring_get_rptr,
1629 	.get_wptr = sdma_v6_0_ring_get_wptr,
1630 	.set_wptr = sdma_v6_0_ring_set_wptr,
1631 	.emit_frame_size =
1632 		5 + /* sdma_v6_0_ring_init_cond_exec */
1633 		6 + /* sdma_v6_0_ring_emit_hdp_flush */
1634 		6 + /* sdma_v6_0_ring_emit_pipeline_sync */
1635 		/* sdma_v6_0_ring_emit_vm_flush */
1636 		SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 +
1637 		SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6 +
1638 		10 + 10 + 10, /* sdma_v6_0_ring_emit_fence x3 for user fence, vm fence */
1639 	.emit_ib_size = 5 + 7 + 6, /* sdma_v6_0_ring_emit_ib */
1640 	.emit_ib = sdma_v6_0_ring_emit_ib,
1641 	.emit_mem_sync = sdma_v6_0_ring_emit_mem_sync,
1642 	.emit_fence = sdma_v6_0_ring_emit_fence,
1643 	.emit_pipeline_sync = sdma_v6_0_ring_emit_pipeline_sync,
1644 	.emit_vm_flush = sdma_v6_0_ring_emit_vm_flush,
1645 	.emit_hdp_flush = sdma_v6_0_ring_emit_hdp_flush,
1646 	.test_ring = sdma_v6_0_ring_test_ring,
1647 	.test_ib = sdma_v6_0_ring_test_ib,
1648 	.insert_nop = sdma_v6_0_ring_insert_nop,
1649 	.pad_ib = sdma_v6_0_ring_pad_ib,
1650 	.emit_wreg = sdma_v6_0_ring_emit_wreg,
1651 	.emit_reg_wait = sdma_v6_0_ring_emit_reg_wait,
1652 	.emit_reg_write_reg_wait = sdma_v6_0_ring_emit_reg_write_reg_wait,
1653 	.init_cond_exec = sdma_v6_0_ring_init_cond_exec,
1654 	.preempt_ib = sdma_v6_0_ring_preempt_ib,
1655 };
1656 
1657 static void sdma_v6_0_set_ring_funcs(struct amdgpu_device *adev)
1658 {
1659 	int i;
1660 
1661 	for (i = 0; i < adev->sdma.num_instances; i++) {
1662 		adev->sdma.instance[i].ring.funcs = &sdma_v6_0_ring_funcs;
1663 		adev->sdma.instance[i].ring.me = i;
1664 	}
1665 }
1666 
1667 static const struct amdgpu_irq_src_funcs sdma_v6_0_trap_irq_funcs = {
1668 	.set = sdma_v6_0_set_trap_irq_state,
1669 	.process = sdma_v6_0_process_trap_irq,
1670 };
1671 
1672 static const struct amdgpu_irq_src_funcs sdma_v6_0_illegal_inst_irq_funcs = {
1673 	.process = sdma_v6_0_process_illegal_inst_irq,
1674 };
1675 
1676 static void sdma_v6_0_set_irq_funcs(struct amdgpu_device *adev)
1677 {
1678 	adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_INSTANCE0 +
1679 					adev->sdma.num_instances;
1680 	adev->sdma.trap_irq.funcs = &sdma_v6_0_trap_irq_funcs;
1681 	adev->sdma.illegal_inst_irq.funcs = &sdma_v6_0_illegal_inst_irq_funcs;
1682 }
1683 
1684 /**
1685  * sdma_v6_0_emit_copy_buffer - copy buffer using the sDMA engine
1686  *
1687  * @ib: indirect buffer to fill with commands
1688  * @src_offset: src GPU address
1689  * @dst_offset: dst GPU address
1690  * @byte_count: number of bytes to xfer
1691  * @copy_flags: copy flags for the buffers
1692  *
1693  * Copy GPU buffers using the DMA engine.
1694  * Used by the amdgpu ttm implementation to move pages if
1695  * registered as the asic copy callback.
1696  */
1697 static void sdma_v6_0_emit_copy_buffer(struct amdgpu_ib *ib,
1698 				       uint64_t src_offset,
1699 				       uint64_t dst_offset,
1700 				       uint32_t byte_count,
1701 				       uint32_t copy_flags)
1702 {
1703 	ib->ptr[ib->length_dw++] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_COPY) |
1704 		SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR) |
1705 		SDMA_PKT_COPY_LINEAR_HEADER_TMZ((copy_flags & AMDGPU_COPY_FLAGS_TMZ) ? 1 : 0);
1706 	ib->ptr[ib->length_dw++] = byte_count - 1;
1707 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1708 	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1709 	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1710 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1711 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1712 }
1713 
1714 /**
1715  * sdma_v6_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_v6_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_COPY_LINEAR_HEADER_OP(SDMA_OP_CONST_FILL);
1730 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1731 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1732 	ib->ptr[ib->length_dw++] = src_data;
1733 	ib->ptr[ib->length_dw++] = byte_count - 1;
1734 }
1735 
1736 static const struct amdgpu_buffer_funcs sdma_v6_0_buffer_funcs = {
1737 	.copy_max_bytes = 0x400000,
1738 	.copy_num_dw = 7,
1739 	.emit_copy_buffer = sdma_v6_0_emit_copy_buffer,
1740 
1741 	.fill_max_bytes = 0x400000,
1742 	.fill_num_dw = 5,
1743 	.emit_fill_buffer = sdma_v6_0_emit_fill_buffer,
1744 };
1745 
1746 static void sdma_v6_0_set_buffer_funcs(struct amdgpu_device *adev)
1747 {
1748 	adev->mman.buffer_funcs = &sdma_v6_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_v6_0_vm_pte_funcs = {
1753 	.copy_pte_num_dw = 7,
1754 	.copy_pte = sdma_v6_0_vm_copy_pte,
1755 	.write_pte = sdma_v6_0_vm_write_pte,
1756 	.set_pte_pde = sdma_v6_0_vm_set_pte_pde,
1757 };
1758 
1759 static void sdma_v6_0_set_vm_pte_funcs(struct amdgpu_device *adev)
1760 {
1761 	unsigned i;
1762 
1763 	adev->vm_manager.vm_pte_funcs = &sdma_v6_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_v6_0_ip_block = {
1772 	.type = AMD_IP_BLOCK_TYPE_SDMA,
1773 	.major = 6,
1774 	.minor = 0,
1775 	.rev = 0,
1776 	.funcs = &sdma_v6_0_ip_funcs,
1777 };
1778