xref: /linux/drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c (revision cb8fef4b07f71da01843457b66e579ad248a625e)
1 /*
2  * Copyright 2019 Advanced Micro Devices, Inc.
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice shall be included in
12  * all copies or substantial portions of the Software.
13  *
14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20  * OTHER DEALINGS IN THE SOFTWARE.
21  *
22  */
23 
24 #include <linux/delay.h>
25 #include <linux/firmware.h>
26 #include <linux/module.h>
27 #include <linux/pci.h>
28 
29 #include "amdgpu.h"
30 #include "amdgpu_ucode.h"
31 #include "amdgpu_sdma.h"
32 #include "amdgpu_trace.h"
33 
34 #include "gc/gc_10_3_0_offset.h"
35 #include "gc/gc_10_3_0_sh_mask.h"
36 #include "ivsrcid/sdma0/irqsrcs_sdma0_5_0.h"
37 #include "ivsrcid/sdma1/irqsrcs_sdma1_5_0.h"
38 #include "ivsrcid/sdma2/irqsrcs_sdma2_5_0.h"
39 #include "ivsrcid/sdma3/irqsrcs_sdma3_5_0.h"
40 
41 #include "soc15_common.h"
42 #include "soc15.h"
43 #include "navi10_sdma_pkt_open.h"
44 #include "nbio_v2_3.h"
45 #include "sdma_common.h"
46 #include "sdma_v5_2.h"
47 
48 MODULE_FIRMWARE("amdgpu/sienna_cichlid_sdma.bin");
49 MODULE_FIRMWARE("amdgpu/navy_flounder_sdma.bin");
50 MODULE_FIRMWARE("amdgpu/dimgrey_cavefish_sdma.bin");
51 MODULE_FIRMWARE("amdgpu/beige_goby_sdma.bin");
52 
53 MODULE_FIRMWARE("amdgpu/vangogh_sdma.bin");
54 MODULE_FIRMWARE("amdgpu/yellow_carp_sdma.bin");
55 MODULE_FIRMWARE("amdgpu/sdma_5_2_6.bin");
56 MODULE_FIRMWARE("amdgpu/sdma_5_2_7.bin");
57 
58 #define SDMA1_REG_OFFSET 0x600
59 #define SDMA3_REG_OFFSET 0x400
60 #define SDMA0_HYP_DEC_REG_START 0x5880
61 #define SDMA0_HYP_DEC_REG_END 0x5893
62 #define SDMA1_HYP_DEC_REG_OFFSET 0x20
63 
64 static const struct amdgpu_hwip_reg_entry sdma_reg_list_5_2[] = {
65 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_STATUS_REG),
66 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_STATUS1_REG),
67 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_STATUS2_REG),
68 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_STATUS3_REG),
69 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_UCODE_CHECKSUM),
70 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RB_RPTR_FETCH_HI),
71 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RB_RPTR_FETCH),
72 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_UTCL1_RD_STATUS),
73 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_UTCL1_WR_STATUS),
74 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_UTCL1_RD_XNACK0),
75 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_UTCL1_RD_XNACK1),
76 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_UTCL1_WR_XNACK0),
77 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_UTCL1_WR_XNACK1),
78 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_RB_CNTL),
79 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_RB_RPTR),
80 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_RB_RPTR_HI),
81 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_RB_WPTR),
82 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_RB_WPTR_HI),
83 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_IB_OFFSET),
84 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_IB_BASE_LO),
85 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_IB_BASE_HI),
86 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_IB_CNTL),
87 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_IB_RPTR),
88 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_IB_SUB_REMAIN),
89 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_DUMMY_REG),
90 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_RB_CNTL),
91 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_RB_RPTR),
92 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_RB_RPTR_HI),
93 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_RB_WPTR),
94 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_RB_WPTR_HI),
95 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_IB_OFFSET),
96 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_IB_BASE_LO),
97 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_IB_BASE_HI),
98 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_DUMMY_REG),
99 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_RB_CNTL),
100 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_RB_RPTR),
101 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_RB_RPTR_HI),
102 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_RB_WPTR),
103 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_RB_WPTR_HI),
104 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_IB_OFFSET),
105 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_IB_BASE_LO),
106 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_IB_BASE_HI),
107 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_DUMMY_REG),
108 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_INT_STATUS),
109 	SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_VM_CNTL),
110 	SOC15_REG_ENTRY_STR(GC, 0, mmGRBM_STATUS2)
111 };
112 
113 static void sdma_v5_2_set_ring_funcs(struct amdgpu_device *adev);
114 static void sdma_v5_2_set_buffer_funcs(struct amdgpu_device *adev);
115 static void sdma_v5_2_set_irq_funcs(struct amdgpu_device *adev);
116 static int sdma_v5_2_stop_queue(struct amdgpu_ring *ring);
117 static int sdma_v5_2_restore_queue(struct amdgpu_ring *ring);
118 
119 static u32 sdma_v5_2_get_reg_offset(struct amdgpu_device *adev, u32 instance, u32 internal_offset)
120 {
121 	u32 base;
122 
123 	if (internal_offset >= SDMA0_HYP_DEC_REG_START &&
124 	    internal_offset <= SDMA0_HYP_DEC_REG_END) {
125 		base = adev->reg_offset[GC_HWIP][0][1];
126 		if (instance != 0)
127 			internal_offset += SDMA1_HYP_DEC_REG_OFFSET * instance;
128 	} else {
129 		if (instance < 2) {
130 			base = adev->reg_offset[GC_HWIP][0][0];
131 			if (instance == 1)
132 				internal_offset += SDMA1_REG_OFFSET;
133 		} else {
134 			base = adev->reg_offset[GC_HWIP][0][2];
135 			if (instance == 3)
136 				internal_offset += SDMA3_REG_OFFSET;
137 		}
138 	}
139 
140 	return base + internal_offset;
141 }
142 
143 static unsigned sdma_v5_2_ring_init_cond_exec(struct amdgpu_ring *ring,
144 					      uint64_t addr)
145 {
146 	unsigned ret;
147 
148 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_COND_EXE));
149 	amdgpu_ring_write(ring, lower_32_bits(addr));
150 	amdgpu_ring_write(ring, upper_32_bits(addr));
151 	amdgpu_ring_write(ring, 1);
152 	/* this is the offset we need patch later */
153 	ret = ring->wptr & ring->buf_mask;
154 	/* insert dummy here and patch it later */
155 	amdgpu_ring_write(ring, 0);
156 
157 	return ret;
158 }
159 
160 /**
161  * sdma_v5_2_ring_get_rptr - get the current read pointer
162  *
163  * @ring: amdgpu ring pointer
164  *
165  * Get the current rptr from the hardware (NAVI10+).
166  */
167 static uint64_t sdma_v5_2_ring_get_rptr(struct amdgpu_ring *ring)
168 {
169 	u64 *rptr;
170 
171 	/* XXX check if swapping is necessary on BE */
172 	rptr = (u64 *)ring->rptr_cpu_addr;
173 
174 	DRM_DEBUG("rptr before shift == 0x%016llx\n", *rptr);
175 	return ((*rptr) >> 2);
176 }
177 
178 /**
179  * sdma_v5_2_ring_get_wptr - get the current write pointer
180  *
181  * @ring: amdgpu ring pointer
182  *
183  * Get the current wptr from the hardware (NAVI10+).
184  */
185 static uint64_t sdma_v5_2_ring_get_wptr(struct amdgpu_ring *ring)
186 {
187 	struct amdgpu_device *adev = ring->adev;
188 	u64 wptr;
189 
190 	if (ring->use_doorbell) {
191 		/* XXX check if swapping is necessary on BE */
192 		wptr = READ_ONCE(*((u64 *)ring->wptr_cpu_addr));
193 		DRM_DEBUG("wptr/doorbell before shift == 0x%016llx\n", wptr);
194 	} else {
195 		wptr = RREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI));
196 		wptr = wptr << 32;
197 		wptr |= RREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR));
198 		DRM_DEBUG("wptr before shift [%i] wptr == 0x%016llx\n", ring->me, wptr);
199 	}
200 
201 	return wptr >> 2;
202 }
203 
204 /**
205  * sdma_v5_2_ring_set_wptr - commit the write pointer
206  *
207  * @ring: amdgpu ring pointer
208  *
209  * Write the wptr back to the hardware (NAVI10+).
210  */
211 static void sdma_v5_2_ring_set_wptr(struct amdgpu_ring *ring)
212 {
213 	struct amdgpu_device *adev = ring->adev;
214 
215 	DRM_DEBUG("Setting write pointer\n");
216 	if (ring->use_doorbell) {
217 		DRM_DEBUG("Using doorbell -- "
218 				"wptr_offs == 0x%08x "
219 				"lower_32_bits(ring->wptr << 2) == 0x%08x "
220 				"upper_32_bits(ring->wptr << 2) == 0x%08x\n",
221 				ring->wptr_offs,
222 				lower_32_bits(ring->wptr << 2),
223 				upper_32_bits(ring->wptr << 2));
224 		/* XXX check if swapping is necessary on BE */
225 		atomic64_set((atomic64_t *)ring->wptr_cpu_addr,
226 			     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 		if (amdgpu_ip_version(adev, SDMA0_HWIP, 0) == IP_VERSION(5, 2, 1)) {
231 			/* SDMA seems to miss doorbells sometimes when powergating kicks in.
232 			 * Updating the wptr directly will wake it. This is only safe because
233 			 * we disallow gfxoff in begin_use() and then allow it again in end_use().
234 			 */
235 			WREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR),
236 			       lower_32_bits(ring->wptr << 2));
237 			WREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI),
238 			       upper_32_bits(ring->wptr << 2));
239 		}
240 	} else {
241 		DRM_DEBUG("Not using doorbell -- "
242 				"mmSDMA%i_GFX_RB_WPTR == 0x%08x "
243 				"mmSDMA%i_GFX_RB_WPTR_HI == 0x%08x\n",
244 				ring->me,
245 				lower_32_bits(ring->wptr << 2),
246 				ring->me,
247 				upper_32_bits(ring->wptr << 2));
248 		WREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR),
249 			lower_32_bits(ring->wptr << 2));
250 		WREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI),
251 			upper_32_bits(ring->wptr << 2));
252 	}
253 }
254 
255 static void sdma_v5_2_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
256 {
257 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
258 	const u32 nop = ring->funcs->nop;
259 
260 	if (count && sdma->burst_nop) {
261 		--count;
262 		amdgpu_ring_write(ring, nop | SDMA_PKT_NOP_HEADER_COUNT(count));
263 	}
264 
265 	amdgpu_ring_fill(ring, nop, count);
266 }
267 
268 /**
269  * sdma_v5_2_ring_emit_ib - Schedule an IB on the DMA engine
270  *
271  * @ring: amdgpu ring pointer
272  * @job: job to retrieve vmid from
273  * @ib: IB object to schedule
274  * @flags: unused
275  *
276  * Schedule an IB in the DMA ring.
277  */
278 static void sdma_v5_2_ring_emit_ib(struct amdgpu_ring *ring,
279 				   struct amdgpu_job *job,
280 				   struct amdgpu_ib *ib,
281 				   uint32_t flags)
282 {
283 	unsigned vmid = AMDGPU_JOB_GET_VMID(job);
284 	uint64_t csa_mc_addr = amdgpu_sdma_get_csa_mc_addr(ring, vmid);
285 
286 	/* An IB packet must end on a 8 DW boundary--the next dword
287 	 * must be on a 8-dword boundary. Our IB packet below is 6
288 	 * dwords long, thus add x number of NOPs, such that, in
289 	 * modular arithmetic,
290 	 * wptr + 6 + x = 8k, k >= 0, which in C is,
291 	 * (wptr + 6 + x) % 8 = 0.
292 	 * The expression below, is a solution of x.
293 	 */
294 	sdma_v5_2_ring_insert_nop(ring, (2 - lower_32_bits(ring->wptr)) & 7);
295 
296 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) |
297 			  SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf));
298 	/* base must be 32 byte aligned */
299 	amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
300 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
301 	amdgpu_ring_write(ring, ib->length_dw);
302 	amdgpu_ring_write(ring, lower_32_bits(csa_mc_addr));
303 	amdgpu_ring_write(ring, upper_32_bits(csa_mc_addr));
304 }
305 
306 /**
307  * sdma_v5_2_ring_emit_mem_sync - flush the IB by graphics cache rinse
308  *
309  * @ring: amdgpu ring pointer
310  *
311  * flush the IB by graphics cache rinse.
312  */
313 static void sdma_v5_2_ring_emit_mem_sync(struct amdgpu_ring *ring)
314 {
315 	uint32_t gcr_cntl = SDMA_GCR_GL2_INV | SDMA_GCR_GL2_WB |
316 			    SDMA_GCR_GLM_INV | SDMA_GCR_GL1_INV |
317 			    SDMA_GCR_GLV_INV | SDMA_GCR_GLK_INV |
318 			    SDMA_GCR_GLI_INV(1);
319 
320 	/* flush entire cache L0/L1/L2, this can be optimized by performance requirement */
321 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_GCR_REQ));
322 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD1_BASE_VA_31_7(0));
323 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD2_GCR_CONTROL_15_0(gcr_cntl) |
324 			SDMA_PKT_GCR_REQ_PAYLOAD2_BASE_VA_47_32(0));
325 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD3_LIMIT_VA_31_7(0) |
326 			SDMA_PKT_GCR_REQ_PAYLOAD3_GCR_CONTROL_18_16(gcr_cntl >> 16));
327 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD4_LIMIT_VA_47_32(0) |
328 			SDMA_PKT_GCR_REQ_PAYLOAD4_VMID(0));
329 }
330 
331 /**
332  * sdma_v5_2_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
333  *
334  * @ring: amdgpu ring pointer
335  *
336  * Emit an hdp flush packet on the requested DMA ring.
337  */
338 static void sdma_v5_2_ring_emit_hdp_flush(struct amdgpu_ring *ring)
339 {
340 	struct amdgpu_device *adev = ring->adev;
341 	u32 ref_and_mask = 0;
342 	const struct nbio_hdp_flush_reg *nbio_hf_reg = adev->nbio.hdp_flush_reg;
343 
344 	if (ring->me > 1) {
345 		amdgpu_hdp_flush(adev, ring);
346 	} else {
347 		ref_and_mask = nbio_hf_reg->ref_and_mask_sdma0 << ring->me;
348 
349 		amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
350 				  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) |
351 				  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
352 		amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_done_offset(adev)) << 2);
353 		amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_req_offset(adev)) << 2);
354 		amdgpu_ring_write(ring, ref_and_mask); /* reference */
355 		amdgpu_ring_write(ring, ref_and_mask); /* mask */
356 		amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
357 				  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
358 	}
359 }
360 
361 /**
362  * sdma_v5_2_ring_emit_fence - emit a fence on the DMA ring
363  *
364  * @ring: amdgpu ring pointer
365  * @addr: address
366  * @seq: sequence number
367  * @flags: fence related flags
368  *
369  * Add a DMA fence packet to the ring to write
370  * the fence seq number and DMA trap packet to generate
371  * an interrupt if needed.
372  */
373 static void sdma_v5_2_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
374 				      unsigned flags)
375 {
376 	bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
377 	/* write the fence */
378 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) |
379 			  SDMA_PKT_FENCE_HEADER_MTYPE(0x3)); /* Ucached(UC) */
380 	/* zero in first two bits */
381 	WARN_ON(addr & 0x3);
382 	amdgpu_ring_write(ring, lower_32_bits(addr));
383 	amdgpu_ring_write(ring, upper_32_bits(addr));
384 	amdgpu_ring_write(ring, lower_32_bits(seq));
385 
386 	/* optionally write high bits as well */
387 	if (write64bit) {
388 		addr += 4;
389 		amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) |
390 				  SDMA_PKT_FENCE_HEADER_MTYPE(0x3));
391 		/* zero in first two bits */
392 		WARN_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, upper_32_bits(seq));
396 	}
397 
398 	if ((flags & AMDGPU_FENCE_FLAG_INT)) {
399 		/* generate an interrupt */
400 		amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP));
401 		amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(0));
402 	}
403 }
404 
405 
406 /**
407  * sdma_v5_2_gfx_stop - stop the gfx async dma engines
408  *
409  * @adev: amdgpu_device pointer
410  * @inst_mask: mask of dma engine instances to be disabled
411  * Stop the gfx async dma ring buffers.
412  */
413 static void sdma_v5_2_gfx_stop(struct amdgpu_device *adev,  uint32_t inst_mask)
414 {
415 	u32 rb_cntl, ib_cntl;
416 	int i;
417 
418 	for_each_inst(i, inst_mask) {
419 		rb_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL));
420 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 0);
421 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
422 		ib_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL));
423 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 0);
424 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl);
425 	}
426 }
427 
428 /**
429  * sdma_v5_2_ctx_switch_enable - stop the async dma engines context switch
430  *
431  * @adev: amdgpu_device pointer
432  * @enable: enable/disable the DMA MEs context switch.
433  *
434  * Halt or unhalt the async dma engines context switch.
435  */
436 static void sdma_v5_2_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
437 {
438 	u32 f32_cntl, phase_quantum = 0;
439 	int i;
440 
441 	if (amdgpu_sdma_phase_quantum) {
442 		unsigned value = amdgpu_sdma_phase_quantum;
443 		unsigned unit = 0;
444 
445 		while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
446 				SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) {
447 			value = (value + 1) >> 1;
448 			unit++;
449 		}
450 		if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
451 			    SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) {
452 			value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
453 				 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT);
454 			unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
455 				SDMA0_PHASE0_QUANTUM__UNIT__SHIFT);
456 			WARN_ONCE(1,
457 			"clamping sdma_phase_quantum to %uK clock cycles\n",
458 				  value << unit);
459 		}
460 		phase_quantum =
461 			value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
462 			unit  << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
463 	}
464 
465 	for (i = 0; i < adev->sdma.num_instances; i++) {
466 		if (enable && amdgpu_sdma_phase_quantum) {
467 			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE0_QUANTUM),
468 			       phase_quantum);
469 			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE1_QUANTUM),
470 			       phase_quantum);
471 			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE2_QUANTUM),
472 			       phase_quantum);
473 		}
474 
475 		if (!amdgpu_sriov_vf(adev)) {
476 			f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
477 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
478 					AUTO_CTXSW_ENABLE, enable ? 1 : 0);
479 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl);
480 		}
481 	}
482 
483 }
484 
485 /**
486  * sdma_v5_2_enable - stop the async dma engines
487  *
488  * @adev: amdgpu_device pointer
489  * @enable: enable/disable the DMA MEs.
490  *
491  * Halt or unhalt the async dma engines.
492  */
493 static void sdma_v5_2_enable(struct amdgpu_device *adev, bool enable)
494 {
495 	u32 f32_cntl;
496 	int i;
497 	uint32_t inst_mask;
498 
499 	inst_mask = GENMASK(adev->sdma.num_instances - 1, 0);
500 	if (!enable) {
501 		sdma_v5_2_gfx_stop(adev, inst_mask);
502 	}
503 
504 	if (!amdgpu_sriov_vf(adev)) {
505 		for (i = 0; i < adev->sdma.num_instances; i++) {
506 			f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
507 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, enable ? 0 : 1);
508 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl);
509 		}
510 	}
511 }
512 
513 /**
514  * sdma_v5_2_gfx_resume_instance - start/restart a certain sdma engine
515  *
516  * @adev: amdgpu_device pointer
517  * @i: instance
518  * @restore: used to restore wptr when restart
519  *
520  * Set up the gfx DMA ring buffers and enable them. On restart, we will restore wptr and rptr.
521  * Return 0 for success.
522  */
523 
524 static int sdma_v5_2_gfx_resume_instance(struct amdgpu_device *adev, int i, bool restore)
525 {
526 	struct amdgpu_ring *ring;
527 	u32 rb_cntl, ib_cntl;
528 	u32 rb_bufsz;
529 	u32 doorbell;
530 	u32 doorbell_offset;
531 	u32 temp;
532 	u32 wptr_poll_cntl;
533 	u64 wptr_gpu_addr;
534 
535 	ring = &adev->sdma.instance[i].ring;
536 
537 	if (!amdgpu_sriov_vf(adev))
538 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL), 0);
539 
540 	/* Set ring buffer size in dwords */
541 	rb_bufsz = order_base_2(ring->ring_size / 4);
542 	rb_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL));
543 	rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz);
544 #ifdef __BIG_ENDIAN
545 	rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1);
546 	rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL,
547 				RPTR_WRITEBACK_SWAP_ENABLE, 1);
548 #endif
549 	WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
550 
551 	/* Initialize the ring buffer's read and write pointers */
552 	if (restore) {
553 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR), lower_32_bits(ring->wptr << 2));
554 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_HI), upper_32_bits(ring->wptr << 2));
555 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), lower_32_bits(ring->wptr << 2));
556 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), upper_32_bits(ring->wptr << 2));
557 	} else {
558 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR), 0);
559 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_HI), 0);
560 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), 0);
561 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), 0);
562 	}
563 
564 	/* setup the wptr shadow polling */
565 	wptr_gpu_addr = ring->wptr_gpu_addr;
566 	WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_LO),
567 	       lower_32_bits(wptr_gpu_addr));
568 	WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_HI),
569 	       upper_32_bits(wptr_gpu_addr));
570 	wptr_poll_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i,
571 						 mmSDMA0_GFX_RB_WPTR_POLL_CNTL));
572 	wptr_poll_cntl = REG_SET_FIELD(wptr_poll_cntl,
573 				       SDMA0_GFX_RB_WPTR_POLL_CNTL,
574 				       F32_POLL_ENABLE, 1);
575 	WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_CNTL),
576 	       wptr_poll_cntl);
577 
578 	/* set the wb address whether it's enabled or not */
579 	WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_HI),
580 	       upper_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFF);
581 	WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_LO),
582 	       lower_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFC);
583 
584 	rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
585 
586 	WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE), ring->gpu_addr >> 8);
587 	WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE_HI), ring->gpu_addr >> 40);
588 
589 	if (!restore)
590 		ring->wptr = 0;
591 
592 	/* before programing wptr to a less value, need set minor_ptr_update first */
593 	WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 1);
594 
595 	if (!amdgpu_sriov_vf(adev)) { /* only bare-metal use register write for wptr */
596 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), lower_32_bits(ring->wptr << 2));
597 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), upper_32_bits(ring->wptr << 2));
598 	}
599 
600 	doorbell = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL));
601 	doorbell_offset = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL_OFFSET));
602 
603 	if (ring->use_doorbell) {
604 		doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 1);
605 		doorbell_offset = REG_SET_FIELD(doorbell_offset, SDMA0_GFX_DOORBELL_OFFSET,
606 				OFFSET, ring->doorbell_index);
607 	} else {
608 		doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 0);
609 	}
610 	WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL), doorbell);
611 	WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL_OFFSET), doorbell_offset);
612 
613 	adev->nbio.funcs->sdma_doorbell_range(adev, i, ring->use_doorbell,
614 					      ring->doorbell_index,
615 					      adev->doorbell_index.sdma_doorbell_range);
616 
617 	if (amdgpu_sriov_vf(adev))
618 		sdma_v5_2_ring_set_wptr(ring);
619 
620 	/* set minor_ptr_update to 0 after wptr programed */
621 
622 	WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 0);
623 
624 	/* SRIOV VF has no control of any of registers below */
625 	if (!amdgpu_sriov_vf(adev)) {
626 		/* set utc l1 enable flag always to 1 */
627 		temp = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
628 		temp = REG_SET_FIELD(temp, SDMA0_CNTL, UTC_L1_ENABLE, 1);
629 
630 		/* enable MCBP */
631 		temp = REG_SET_FIELD(temp, SDMA0_CNTL, MIDCMD_PREEMPT_ENABLE, 1);
632 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), temp);
633 
634 		/* Set up RESP_MODE to non-copy addresses */
635 		temp = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL));
636 		temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, RESP_MODE, 3);
637 		temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, REDO_DELAY, 9);
638 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL), temp);
639 
640 		/* program default cache read and write policy */
641 		temp = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE));
642 		/* clean read policy and write policy bits */
643 		temp &= 0xFF0FFF;
644 		temp |= ((CACHE_READ_POLICY_L2__DEFAULT << 12) |
645 			 (CACHE_WRITE_POLICY_L2__DEFAULT << 14) |
646 			 SDMA0_UTCL1_PAGE__LLC_NOALLOC_MASK);
647 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE), temp);
648 
649 		/* unhalt engine */
650 		temp = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
651 		temp = REG_SET_FIELD(temp, SDMA0_F32_CNTL, HALT, 0);
652 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), temp);
653 	}
654 
655 	/* enable DMA RB */
656 	rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1);
657 	WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
658 
659 	ib_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL));
660 	ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 1);
661 #ifdef __BIG_ENDIAN
662 	ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1);
663 #endif
664 	/* enable DMA IBs */
665 	WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl);
666 
667 	if (amdgpu_sriov_vf(adev)) { /* bare-metal sequence doesn't need below to lines */
668 		sdma_v5_2_ctx_switch_enable(adev, true);
669 		sdma_v5_2_enable(adev, true);
670 	}
671 
672 	return amdgpu_ring_test_helper(ring);
673 }
674 
675 /**
676  * sdma_v5_2_gfx_resume - setup and start the async dma engines
677  *
678  * @adev: amdgpu_device pointer
679  *
680  * Set up the gfx DMA ring buffers and enable them.
681  * Returns 0 for success, error for failure.
682  */
683 static int sdma_v5_2_gfx_resume(struct amdgpu_device *adev)
684 {
685 	int i, r;
686 
687 	for (i = 0; i < adev->sdma.num_instances; i++) {
688 		r = sdma_v5_2_gfx_resume_instance(adev, i, false);
689 		if (r)
690 			return r;
691 	}
692 
693 	return 0;
694 }
695 
696 /**
697  * sdma_v5_2_load_microcode - load the sDMA ME ucode
698  *
699  * @adev: amdgpu_device pointer
700  *
701  * Loads the sDMA0/1/2/3 ucode.
702  * Returns 0 for success, -EINVAL if the ucode is not available.
703  */
704 static int sdma_v5_2_load_microcode(struct amdgpu_device *adev)
705 {
706 	const struct sdma_firmware_header_v1_0 *hdr;
707 	const __le32 *fw_data;
708 	u32 fw_size;
709 	int i, j;
710 
711 	/* halt the MEs */
712 	sdma_v5_2_enable(adev, false);
713 
714 	for (i = 0; i < adev->sdma.num_instances; i++) {
715 		if (!adev->sdma.instance[i].fw)
716 			return -EINVAL;
717 
718 		hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
719 		amdgpu_ucode_print_sdma_hdr(&hdr->header);
720 		fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
721 
722 		fw_data = (const __le32 *)
723 			(adev->sdma.instance[i].fw->data +
724 				le32_to_cpu(hdr->header.ucode_array_offset_bytes));
725 
726 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), 0);
727 
728 		for (j = 0; j < fw_size; j++) {
729 			if (amdgpu_emu_mode == 1 && j % 500 == 0)
730 				msleep(1);
731 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UCODE_DATA), le32_to_cpup(fw_data++));
732 		}
733 
734 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), adev->sdma.instance[i].fw_version);
735 	}
736 
737 	return 0;
738 }
739 
740 static int sdma_v5_2_soft_reset_engine(struct amdgpu_device *adev, u32 instance_id)
741 {
742 	u32 grbm_soft_reset;
743 	u32 tmp;
744 
745 	grbm_soft_reset = REG_SET_FIELD(0,
746 					GRBM_SOFT_RESET, SOFT_RESET_SDMA0,
747 					1);
748 	grbm_soft_reset <<= instance_id;
749 
750 	tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET);
751 	tmp |= grbm_soft_reset;
752 	DRM_DEBUG("GRBM_SOFT_RESET=0x%08X\n", tmp);
753 	WREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET, tmp);
754 	tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET);
755 
756 	udelay(50);
757 
758 	tmp &= ~grbm_soft_reset;
759 	WREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET, tmp);
760 	tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET);
761 	return 0;
762 }
763 
764 static int sdma_v5_2_soft_reset(struct amdgpu_ip_block *ip_block)
765 {
766 	struct amdgpu_device *adev = ip_block->adev;
767 	int i;
768 
769 	for (i = 0; i < adev->sdma.num_instances; i++) {
770 		sdma_v5_2_soft_reset_engine(adev, i);
771 		udelay(50);
772 	}
773 
774 	return 0;
775 }
776 
777 static const struct amdgpu_sdma_funcs sdma_v5_2_sdma_funcs = {
778 	.stop_kernel_queue = &sdma_v5_2_stop_queue,
779 	.start_kernel_queue = &sdma_v5_2_restore_queue,
780 	.soft_reset_kernel_queue = &sdma_v5_2_soft_reset_engine,
781 };
782 
783 /**
784  * sdma_v5_2_start - setup and start the async dma engines
785  *
786  * @adev: amdgpu_device pointer
787  *
788  * Set up the DMA engines and enable them.
789  * Returns 0 for success, error for failure.
790  */
791 static int sdma_v5_2_start(struct amdgpu_device *adev)
792 {
793 	int r = 0;
794 	struct amdgpu_ip_block *ip_block;
795 
796 	if (amdgpu_sriov_vf(adev)) {
797 		sdma_v5_2_ctx_switch_enable(adev, false);
798 		sdma_v5_2_enable(adev, false);
799 
800 		/* set RB registers */
801 		r = sdma_v5_2_gfx_resume(adev);
802 		return r;
803 	}
804 
805 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
806 		r = sdma_v5_2_load_microcode(adev);
807 		if (r)
808 			return r;
809 
810 		/* The value of mmSDMA_F32_CNTL is invalid the moment after loading fw */
811 		if (amdgpu_emu_mode == 1)
812 			msleep(1000);
813 	}
814 
815 	ip_block = amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_SDMA);
816 	if (!ip_block)
817 		return -EINVAL;
818 
819 	sdma_v5_2_soft_reset(ip_block);
820 	/* unhalt the MEs */
821 	sdma_v5_2_enable(adev, true);
822 	/* enable sdma ring preemption */
823 	sdma_v5_2_ctx_switch_enable(adev, true);
824 
825 	/* start the gfx rings and rlc compute queues */
826 	r = sdma_v5_2_gfx_resume(adev);
827 	if (r)
828 		return r;
829 
830 	return r;
831 }
832 
833 static int sdma_v5_2_mqd_init(struct amdgpu_device *adev, void *mqd,
834 			      struct amdgpu_mqd_prop *prop)
835 {
836 	struct v10_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_RLC0_RB_CNTL__RB_SIZE__SHIFT |
841 		1 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_ENABLE__SHIFT |
842 		6 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_TIMER__SHIFT |
843 		1 << SDMA0_RLC0_RB_CNTL__RB_PRIV__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 	m->sdmax_rlcx_rb_wptr_poll_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, 0,
849 						  mmSDMA0_GFX_RB_WPTR_POLL_CNTL));
850 
851 	wb_gpu_addr = prop->wptr_gpu_addr;
852 	m->sdmax_rlcx_rb_wptr_poll_addr_lo = lower_32_bits(wb_gpu_addr);
853 	m->sdmax_rlcx_rb_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr);
854 
855 	wb_gpu_addr = prop->rptr_gpu_addr;
856 	m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits(wb_gpu_addr);
857 	m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits(wb_gpu_addr);
858 
859 	m->sdmax_rlcx_ib_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, 0,
860 							mmSDMA0_GFX_IB_CNTL));
861 
862 	m->sdmax_rlcx_doorbell_offset =
863 		prop->doorbell_index << SDMA0_RLC0_DOORBELL_OFFSET__OFFSET__SHIFT;
864 
865 	m->sdmax_rlcx_doorbell = REG_SET_FIELD(0, SDMA0_RLC0_DOORBELL, ENABLE, 1);
866 
867 	return 0;
868 }
869 
870 static void sdma_v5_2_set_mqd_funcs(struct amdgpu_device *adev)
871 {
872 	adev->mqds[AMDGPU_HW_IP_DMA].mqd_size = sizeof(struct v10_sdma_mqd);
873 	adev->mqds[AMDGPU_HW_IP_DMA].init_mqd = sdma_v5_2_mqd_init;
874 }
875 
876 /**
877  * sdma_v5_2_ring_test_ring - simple async dma engine test
878  *
879  * @ring: amdgpu_ring structure holding ring information
880  *
881  * Test the DMA engine by writing using it to write an
882  * value to memory.
883  * Returns 0 for success, error for failure.
884  */
885 static int sdma_v5_2_ring_test_ring(struct amdgpu_ring *ring)
886 {
887 	struct amdgpu_device *adev = ring->adev;
888 	unsigned i;
889 	unsigned index;
890 	int r;
891 	u32 tmp;
892 	u64 gpu_addr;
893 
894 	tmp = 0xCAFEDEAD;
895 
896 	r = amdgpu_wb_get(adev, &index);
897 	if (r) {
898 		dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
899 		return r;
900 	}
901 
902 	gpu_addr = adev->wb.gpu_addr + (index * 4);
903 	adev->wb.wb[index] = cpu_to_le32(tmp);
904 
905 	r = amdgpu_ring_alloc(ring, 20);
906 	if (r) {
907 		drm_err(adev_to_drm(adev), "dma failed to lock ring %d (%d).\n", ring->idx, r);
908 		amdgpu_wb_free(adev, index);
909 		return r;
910 	}
911 
912 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
913 			  SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
914 	amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
915 	amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
916 	amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0));
917 	amdgpu_ring_write(ring, 0xDEADBEEF);
918 	amdgpu_ring_commit(ring);
919 
920 	for (i = 0; i < adev->usec_timeout; i++) {
921 		tmp = le32_to_cpu(adev->wb.wb[index]);
922 		if (tmp == 0xDEADBEEF)
923 			break;
924 		if (amdgpu_emu_mode == 1)
925 			msleep(1);
926 		else
927 			udelay(1);
928 	}
929 
930 	if (i >= adev->usec_timeout)
931 		r = -ETIMEDOUT;
932 
933 	amdgpu_wb_free(adev, index);
934 
935 	return r;
936 }
937 
938 /**
939  * sdma_v5_2_ring_test_ib - test an IB on the DMA engine
940  *
941  * @ring: amdgpu_ring structure holding ring information
942  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
943  *
944  * Test a simple IB in the DMA ring.
945  * Returns 0 on success, error on failure.
946  */
947 static int sdma_v5_2_ring_test_ib(struct amdgpu_ring *ring, long timeout)
948 {
949 	struct amdgpu_device *adev = ring->adev;
950 	struct amdgpu_job *job;
951 	struct amdgpu_ib *ib;
952 	struct dma_fence *f = NULL;
953 	unsigned index;
954 	long r;
955 	u32 tmp = 0;
956 	u64 gpu_addr;
957 
958 	tmp = 0xCAFEDEAD;
959 
960 	r = amdgpu_wb_get(adev, &index);
961 	if (r) {
962 		dev_err(adev->dev, "(%ld) failed to allocate wb slot\n", r);
963 		return r;
964 	}
965 
966 	gpu_addr = adev->wb.gpu_addr + (index * 4);
967 	adev->wb.wb[index] = cpu_to_le32(tmp);
968 
969 	r = amdgpu_job_alloc_with_ib(ring->adev, NULL, NULL, 256,
970 				     AMDGPU_IB_POOL_DIRECT,
971 				     AMDGPU_KERNEL_JOB_ID_SDMA_RING_TEST,
972 				     &job);
973 	if (r) {
974 		drm_err(adev_to_drm(adev), "failed to get ib (%ld).\n", r);
975 		goto err0;
976 	}
977 
978 	ib = &job->ibs[0];
979 	ib->ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
980 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
981 	ib->ptr[1] = lower_32_bits(gpu_addr);
982 	ib->ptr[2] = upper_32_bits(gpu_addr);
983 	ib->ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0);
984 	ib->ptr[4] = 0xDEADBEEF;
985 	ib->ptr[5] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
986 	ib->ptr[6] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
987 	ib->ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
988 	ib->length_dw = 8;
989 
990 	r = amdgpu_job_submit_direct(job, ring, &f);
991 	if (r) {
992 		amdgpu_job_free(job);
993 		goto err1;
994 	}
995 
996 	r = dma_fence_wait_timeout(f, false, timeout);
997 	if (r == 0) {
998 		drm_err(adev_to_drm(adev), "IB test timed out\n");
999 		r = -ETIMEDOUT;
1000 		goto err1;
1001 	} else if (r < 0) {
1002 		drm_err(adev_to_drm(adev), "fence wait failed (%ld).\n", r);
1003 		goto err1;
1004 	}
1005 
1006 	tmp = le32_to_cpu(adev->wb.wb[index]);
1007 
1008 	if (tmp == 0xDEADBEEF)
1009 		r = 0;
1010 	else
1011 		r = -EINVAL;
1012 
1013 err1:
1014 	dma_fence_put(f);
1015 err0:
1016 	amdgpu_wb_free(adev, index);
1017 	return r;
1018 }
1019 
1020 
1021 /**
1022  * sdma_v5_2_vm_copy_pte - update PTEs by copying them from the GART
1023  *
1024  * @ib: indirect buffer to fill with commands
1025  * @pe: addr of the page entry
1026  * @src: src addr to copy from
1027  * @count: number of page entries to update
1028  *
1029  * Update PTEs by copying them from the GART using sDMA.
1030  */
1031 static void sdma_v5_2_vm_copy_pte(struct amdgpu_ib *ib,
1032 				  uint64_t pe, uint64_t src,
1033 				  unsigned count)
1034 {
1035 	unsigned bytes = count * 8;
1036 
1037 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1038 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
1039 	ib->ptr[ib->length_dw++] = bytes - 1;
1040 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1041 	ib->ptr[ib->length_dw++] = lower_32_bits(src);
1042 	ib->ptr[ib->length_dw++] = upper_32_bits(src);
1043 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1044 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1045 
1046 }
1047 
1048 /**
1049  * sdma_v5_2_vm_write_pte - update PTEs by writing them manually
1050  *
1051  * @ib: indirect buffer to fill with commands
1052  * @pe: addr of the page entry
1053  * @value: dst addr to write into pe
1054  * @count: number of page entries to update
1055  * @incr: increase next addr by incr bytes
1056  *
1057  * Update PTEs by writing them manually using sDMA.
1058  */
1059 static void sdma_v5_2_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
1060 				   uint64_t value, unsigned count,
1061 				   uint32_t incr)
1062 {
1063 	unsigned ndw = count * 2;
1064 
1065 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
1066 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1067 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1068 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1069 	ib->ptr[ib->length_dw++] = ndw - 1;
1070 	for (; ndw > 0; ndw -= 2) {
1071 		ib->ptr[ib->length_dw++] = lower_32_bits(value);
1072 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
1073 		value += incr;
1074 	}
1075 }
1076 
1077 /**
1078  * sdma_v5_2_vm_set_pte_pde - update the page tables using sDMA
1079  *
1080  * @ib: indirect buffer to fill with commands
1081  * @pe: addr of the page entry
1082  * @addr: dst addr to write into pe
1083  * @count: number of page entries to update
1084  * @incr: increase next addr by incr bytes
1085  * @flags: access flags
1086  *
1087  * Update the page tables using sDMA.
1088  */
1089 static void sdma_v5_2_vm_set_pte_pde(struct amdgpu_ib *ib,
1090 				     uint64_t pe,
1091 				     uint64_t addr, unsigned count,
1092 				     uint32_t incr, uint64_t flags)
1093 {
1094 	/* for physically contiguous pages (vram) */
1095 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_PTEPDE);
1096 	ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
1097 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1098 	ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
1099 	ib->ptr[ib->length_dw++] = upper_32_bits(flags);
1100 	ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
1101 	ib->ptr[ib->length_dw++] = upper_32_bits(addr);
1102 	ib->ptr[ib->length_dw++] = incr; /* increment size */
1103 	ib->ptr[ib->length_dw++] = 0;
1104 	ib->ptr[ib->length_dw++] = count - 1; /* number of entries */
1105 }
1106 
1107 /**
1108  * sdma_v5_2_ring_pad_ib - pad the IB
1109  *
1110  * @ib: indirect buffer to fill with padding
1111  * @ring: amdgpu_ring structure holding ring information
1112  *
1113  * Pad the IB with NOPs to a boundary multiple of 8.
1114  */
1115 static void sdma_v5_2_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
1116 {
1117 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
1118 	const bool burst_nop = sdma->burst_nop;
1119 	u32 pad_count;
1120 	int i;
1121 
1122 	pad_count = (-ib->length_dw) & 0x7;
1123 	for (i = 0; i < pad_count; i++)
1124 		if (i == 0 && burst_nop)
1125 			ib->ptr[ib->length_dw++] =
1126 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP) |
1127 				SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
1128 		else
1129 			ib->ptr[ib->length_dw++] =
1130 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
1131 }
1132 
1133 
1134 /**
1135  * sdma_v5_2_ring_emit_pipeline_sync - sync the pipeline
1136  *
1137  * @ring: amdgpu_ring pointer
1138  *
1139  * Make sure all previous operations are completed (CIK).
1140  */
1141 static void sdma_v5_2_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
1142 {
1143 	uint32_t seq = ring->fence_drv.sync_seq;
1144 	uint64_t addr = ring->fence_drv.gpu_addr;
1145 
1146 	/* wait for idle */
1147 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1148 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1149 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
1150 			  SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1));
1151 	amdgpu_ring_write(ring, addr & 0xfffffffc);
1152 	amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
1153 	amdgpu_ring_write(ring, seq); /* reference */
1154 	amdgpu_ring_write(ring, 0xffffffff); /* mask */
1155 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1156 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */
1157 }
1158 
1159 
1160 /**
1161  * sdma_v5_2_ring_emit_vm_flush - vm flush using sDMA
1162  *
1163  * @ring: amdgpu_ring pointer
1164  * @vmid: vmid number to use
1165  * @pd_addr: address
1166  *
1167  * Update the page table base and flush the VM TLB
1168  * using sDMA.
1169  */
1170 static void sdma_v5_2_ring_emit_vm_flush(struct amdgpu_ring *ring,
1171 					 unsigned vmid, uint64_t pd_addr)
1172 {
1173 	struct amdgpu_vmhub *hub = &ring->adev->vmhub[ring->vm_hub];
1174 	uint32_t req = hub->vmhub_funcs->get_invalidate_req(vmid, 0);
1175 
1176 	/* Update the PD address for this VMID. */
1177 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_lo32 +
1178 			      (hub->ctx_addr_distance * vmid),
1179 			      lower_32_bits(pd_addr));
1180 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_hi32 +
1181 			      (hub->ctx_addr_distance * vmid),
1182 			      upper_32_bits(pd_addr));
1183 
1184 	/* Trigger invalidation. */
1185 	amdgpu_ring_write(ring,
1186 			  SDMA_PKT_VM_INVALIDATION_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1187 			  SDMA_PKT_VM_INVALIDATION_HEADER_SUB_OP(SDMA_SUBOP_VM_INVALIDATION) |
1188 			  SDMA_PKT_VM_INVALIDATION_HEADER_GFX_ENG_ID(ring->vm_inv_eng) |
1189 			  SDMA_PKT_VM_INVALIDATION_HEADER_MM_ENG_ID(0x1f));
1190 	amdgpu_ring_write(ring, req);
1191 	amdgpu_ring_write(ring, 0xFFFFFFFF);
1192 	amdgpu_ring_write(ring,
1193 			  SDMA_PKT_VM_INVALIDATION_ADDRESSRANGEHI_INVALIDATEACK(1 << vmid) |
1194 			  SDMA_PKT_VM_INVALIDATION_ADDRESSRANGEHI_ADDRESSRANGEHI(0x1F));
1195 }
1196 
1197 static void sdma_v5_2_ring_emit_wreg(struct amdgpu_ring *ring,
1198 				     uint32_t reg, uint32_t val)
1199 {
1200 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
1201 			  SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
1202 	amdgpu_ring_write(ring, reg);
1203 	amdgpu_ring_write(ring, val);
1204 }
1205 
1206 static void sdma_v5_2_ring_emit_reg_wait(struct amdgpu_ring *ring, uint32_t reg,
1207 					 uint32_t val, uint32_t mask)
1208 {
1209 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1210 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1211 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* equal */
1212 	amdgpu_ring_write(ring, reg << 2);
1213 	amdgpu_ring_write(ring, 0);
1214 	amdgpu_ring_write(ring, val); /* reference */
1215 	amdgpu_ring_write(ring, mask); /* mask */
1216 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1217 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10));
1218 }
1219 
1220 static void sdma_v5_2_ring_emit_reg_write_reg_wait(struct amdgpu_ring *ring,
1221 						   uint32_t reg0, uint32_t reg1,
1222 						   uint32_t ref, uint32_t mask)
1223 {
1224 	amdgpu_ring_emit_wreg(ring, reg0, ref);
1225 	/* wait for a cycle to reset vm_inv_eng*_ack */
1226 	amdgpu_ring_emit_reg_wait(ring, reg0, 0, 0);
1227 	amdgpu_ring_emit_reg_wait(ring, reg1, mask, mask);
1228 }
1229 
1230 static const struct amdgpu_vm_pte_funcs sdma_v5_2_vm_pte_funcs = {
1231 	.copy_pte_num_dw = 7,
1232 	.copy_pte = sdma_v5_2_vm_copy_pte,
1233 	.write_pte = sdma_v5_2_vm_write_pte,
1234 	.set_pte_pde = sdma_v5_2_vm_set_pte_pde,
1235 };
1236 
1237 static int sdma_v5_2_early_init(struct amdgpu_ip_block *ip_block)
1238 {
1239 	struct amdgpu_device *adev = ip_block->adev;
1240 	int r;
1241 
1242 	r = amdgpu_sdma_init_microcode(adev, 0, true);
1243 	if (r)
1244 		return r;
1245 
1246 	sdma_v5_2_set_ring_funcs(adev);
1247 	amdgpu_sdma_set_vm_pte_scheds(adev, &sdma_v5_2_vm_pte_funcs);
1248 	sdma_v5_2_set_irq_funcs(adev);
1249 	sdma_v5_2_set_mqd_funcs(adev);
1250 
1251 	return 0;
1252 }
1253 
1254 static unsigned sdma_v5_2_seq_to_irq_id(int seq_num)
1255 {
1256 	switch (seq_num) {
1257 	case 0:
1258 		return SOC15_IH_CLIENTID_SDMA0;
1259 	case 1:
1260 		return SOC15_IH_CLIENTID_SDMA1;
1261 	case 2:
1262 		return SOC15_IH_CLIENTID_SDMA2;
1263 	case 3:
1264 		return SOC15_IH_CLIENTID_SDMA3_Sienna_Cichlid;
1265 	default:
1266 		break;
1267 	}
1268 	return -EINVAL;
1269 }
1270 
1271 static unsigned sdma_v5_2_seq_to_trap_id(int seq_num)
1272 {
1273 	switch (seq_num) {
1274 	case 0:
1275 		return SDMA0_5_0__SRCID__SDMA_TRAP;
1276 	case 1:
1277 		return SDMA1_5_0__SRCID__SDMA_TRAP;
1278 	case 2:
1279 		return SDMA2_5_0__SRCID__SDMA_TRAP;
1280 	case 3:
1281 		return SDMA3_5_0__SRCID__SDMA_TRAP;
1282 	default:
1283 		break;
1284 	}
1285 	return -EINVAL;
1286 }
1287 
1288 static int sdma_v5_2_sw_init(struct amdgpu_ip_block *ip_block)
1289 {
1290 	struct amdgpu_ring *ring;
1291 	int r, i;
1292 	struct amdgpu_device *adev = ip_block->adev;
1293 	uint32_t reg_count = ARRAY_SIZE(sdma_reg_list_5_2);
1294 	uint32_t *ptr;
1295 
1296 	/* SDMA trap event */
1297 	for (i = 0; i < adev->sdma.num_instances; i++) {
1298 		r = amdgpu_irq_add_id(adev, sdma_v5_2_seq_to_irq_id(i),
1299 				      sdma_v5_2_seq_to_trap_id(i),
1300 				      &adev->sdma.trap_irq);
1301 		if (r)
1302 			return r;
1303 	}
1304 
1305 	for (i = 0; i < adev->sdma.num_instances; i++) {
1306 		mutex_init(&adev->sdma.instance[i].engine_reset_mutex);
1307 		adev->sdma.instance[i].funcs = &sdma_v5_2_sdma_funcs;
1308 		ring = &adev->sdma.instance[i].ring;
1309 		ring->use_doorbell = true;
1310 		ring->me = i;
1311 
1312 		drm_info(adev_to_drm(adev), "use_doorbell being set to: [%s]\n",
1313 			 ring->use_doorbell?"true":"false");
1314 
1315 		ring->doorbell_index =
1316 			(adev->doorbell_index.sdma_engine[i] << 1); //get DWORD offset
1317 
1318 		ring->vm_hub = AMDGPU_GFXHUB(0);
1319 		r = amdgpu_sdma_ring_init(adev, ring, i, "sdma%d", i);
1320 		if (r)
1321 			return r;
1322 	}
1323 
1324 	adev->sdma.supported_reset =
1325 		amdgpu_get_soft_full_reset_mask(&adev->sdma.instance[0].ring);
1326 	if (!amdgpu_sriov_vf(adev) &&
1327 	    !adev->debug_disable_gpu_ring_reset)
1328 		adev->sdma.supported_reset |= AMDGPU_RESET_TYPE_PER_QUEUE;
1329 
1330 	/* Allocate memory for SDMA IP Dump buffer */
1331 	ptr = kzalloc_objs(*ptr, adev->sdma.num_instances * reg_count);
1332 	if (ptr)
1333 		adev->sdma.ip_dump = ptr;
1334 	else
1335 		DRM_ERROR("Failed to allocated memory for SDMA IP Dump\n");
1336 
1337 	r = amdgpu_sdma_sysfs_reset_mask_init(adev);
1338 	if (r)
1339 		return r;
1340 
1341 	return r;
1342 }
1343 
1344 static int sdma_v5_2_sw_fini(struct amdgpu_ip_block *ip_block)
1345 {
1346 	struct amdgpu_device *adev = ip_block->adev;
1347 	int i;
1348 
1349 	for (i = 0; i < adev->sdma.num_instances; i++)
1350 		amdgpu_ring_fini(&adev->sdma.instance[i].ring);
1351 
1352 	amdgpu_sdma_sysfs_reset_mask_fini(adev);
1353 	amdgpu_sdma_destroy_inst_ctx(adev, true);
1354 
1355 	kfree(adev->sdma.ip_dump);
1356 
1357 	return 0;
1358 }
1359 
1360 static int sdma_v5_2_hw_init(struct amdgpu_ip_block *ip_block)
1361 {
1362 	struct amdgpu_device *adev = ip_block->adev;
1363 	int r;
1364 
1365 	r = sdma_v5_2_start(adev);
1366 	if (r)
1367 		return r;
1368 	sdma_v5_2_set_buffer_funcs(adev);
1369 
1370 	return 0;
1371 }
1372 
1373 static int sdma_v5_2_hw_fini(struct amdgpu_ip_block *ip_block)
1374 {
1375 	struct amdgpu_device *adev = ip_block->adev;
1376 
1377 	if (amdgpu_sriov_vf(adev))
1378 		return 0;
1379 
1380 	sdma_v5_2_ctx_switch_enable(adev, false);
1381 	sdma_v5_2_enable(adev, false);
1382 
1383 	return 0;
1384 }
1385 
1386 static int sdma_v5_2_suspend(struct amdgpu_ip_block *ip_block)
1387 {
1388 	return sdma_v5_2_hw_fini(ip_block);
1389 }
1390 
1391 static int sdma_v5_2_resume(struct amdgpu_ip_block *ip_block)
1392 {
1393 	return sdma_v5_2_hw_init(ip_block);
1394 }
1395 
1396 static int sdma_v5_2_wait_for_idle(struct amdgpu_ip_block *ip_block)
1397 {
1398 	unsigned i;
1399 	u32 sdma0, sdma1, sdma2, sdma3;
1400 	struct amdgpu_device *adev = ip_block->adev;
1401 
1402 	for (i = 0; i < adev->usec_timeout; i++) {
1403 		sdma0 = RREG32(sdma_v5_2_get_reg_offset(adev, 0, mmSDMA0_STATUS_REG));
1404 		sdma1 = RREG32(sdma_v5_2_get_reg_offset(adev, 1, mmSDMA0_STATUS_REG));
1405 		sdma2 = RREG32(sdma_v5_2_get_reg_offset(adev, 2, mmSDMA0_STATUS_REG));
1406 		sdma3 = RREG32(sdma_v5_2_get_reg_offset(adev, 3, mmSDMA0_STATUS_REG));
1407 
1408 		if (sdma0 & sdma1 & sdma2 & sdma3 & SDMA0_STATUS_REG__IDLE_MASK)
1409 			return 0;
1410 		udelay(1);
1411 	}
1412 	return -ETIMEDOUT;
1413 }
1414 
1415 static int sdma_v5_2_stop_queue(struct amdgpu_ring *ring)
1416 {
1417 	u32 f32_cntl, freeze, cntl, stat1_reg;
1418 	struct amdgpu_device *adev = ring->adev;
1419 	int i, j, r = 0;
1420 
1421 	if (amdgpu_sriov_vf(adev))
1422 		return -EINVAL;
1423 
1424 	i = ring->me;
1425 	amdgpu_gfx_rlc_enter_safe_mode(adev, 0);
1426 
1427 	/* stop queue */
1428 	sdma_v5_2_gfx_stop(adev, 1 << i);
1429 
1430 	/*engine stop SDMA1_F32_CNTL.HALT to 1 and SDMAx_FREEZE freeze bit to 1 */
1431 	freeze = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_FREEZE));
1432 	freeze = REG_SET_FIELD(freeze, SDMA0_FREEZE, FREEZE, 1);
1433 	WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_FREEZE), freeze);
1434 
1435 	for (j = 0; j < adev->usec_timeout; j++) {
1436 		freeze = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_FREEZE));
1437 
1438 		if (REG_GET_FIELD(freeze, SDMA0_FREEZE, FROZEN) & 1)
1439 			break;
1440 		udelay(1);
1441 	}
1442 
1443 
1444 	if (j == adev->usec_timeout) {
1445 		stat1_reg = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_STATUS1_REG));
1446 		if ((stat1_reg & 0x3FF) != 0x3FF) {
1447 			DRM_ERROR("cannot soft reset as sdma not idle\n");
1448 			r = -ETIMEDOUT;
1449 			goto err0;
1450 		}
1451 	}
1452 
1453 	f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
1454 	f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 1);
1455 	WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl);
1456 
1457 	cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
1458 	cntl = REG_SET_FIELD(cntl, SDMA0_CNTL, UTC_L1_ENABLE, 0);
1459 	WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), cntl);
1460 
1461 err0:
1462 	amdgpu_gfx_rlc_exit_safe_mode(adev, 0);
1463 	return r;
1464 }
1465 
1466 static int sdma_v5_2_restore_queue(struct amdgpu_ring *ring)
1467 {
1468 	struct amdgpu_device *adev = ring->adev;
1469 	u32 inst_id = ring->me;
1470 	u32 freeze;
1471 	int r;
1472 
1473 	amdgpu_gfx_rlc_enter_safe_mode(adev, 0);
1474 	/* unfreeze and unhalt */
1475 	freeze = RREG32(sdma_v5_2_get_reg_offset(adev, inst_id, mmSDMA0_FREEZE));
1476 	freeze = REG_SET_FIELD(freeze, SDMA0_FREEZE, FREEZE, 0);
1477 	WREG32(sdma_v5_2_get_reg_offset(adev, inst_id, mmSDMA0_FREEZE), freeze);
1478 
1479 	r = sdma_v5_2_gfx_resume_instance(adev, inst_id, true);
1480 
1481 	amdgpu_gfx_rlc_exit_safe_mode(adev, 0);
1482 
1483 	return r;
1484 }
1485 
1486 static int sdma_v5_2_ring_preempt_ib(struct amdgpu_ring *ring)
1487 {
1488 	int i, r = 0;
1489 	struct amdgpu_device *adev = ring->adev;
1490 	u64 sdma_gfx_preempt;
1491 
1492 	sdma_gfx_preempt = sdma_v5_2_get_reg_offset(adev, ring->me,
1493 						    mmSDMA0_GFX_PREEMPT);
1494 
1495 	/* assert preemption condition */
1496 	amdgpu_ring_set_preempt_cond_exec(ring, false);
1497 
1498 	/* emit the trailing fence */
1499 	ring->trail_seq += 1;
1500 	amdgpu_ring_alloc(ring, 10);
1501 	sdma_v5_2_ring_emit_fence(ring, ring->trail_fence_gpu_addr,
1502 				  ring->trail_seq, 0);
1503 	amdgpu_ring_commit(ring);
1504 
1505 	/* assert IB preemption */
1506 	WREG32(sdma_gfx_preempt, 1);
1507 
1508 	/* poll the trailing fence */
1509 	for (i = 0; i < adev->usec_timeout; i++) {
1510 		if (ring->trail_seq ==
1511 		    le32_to_cpu(*(ring->trail_fence_cpu_addr)))
1512 			break;
1513 		udelay(1);
1514 	}
1515 
1516 	if (i >= adev->usec_timeout) {
1517 		r = -EINVAL;
1518 		DRM_ERROR("ring %d failed to be preempted\n", ring->idx);
1519 	}
1520 
1521 	/* deassert IB preemption */
1522 	WREG32(sdma_gfx_preempt, 0);
1523 
1524 	/* deassert the preemption condition */
1525 	amdgpu_ring_set_preempt_cond_exec(ring, true);
1526 	return r;
1527 }
1528 
1529 static int sdma_v5_2_set_trap_irq_state(struct amdgpu_device *adev,
1530 					struct amdgpu_irq_src *source,
1531 					unsigned type,
1532 					enum amdgpu_interrupt_state state)
1533 {
1534 	u32 sdma_cntl;
1535 	u32 reg_offset = sdma_v5_2_get_reg_offset(adev, type, mmSDMA0_CNTL);
1536 
1537 	if (!amdgpu_sriov_vf(adev)) {
1538 		sdma_cntl = RREG32(reg_offset);
1539 		sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE,
1540 			       state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0);
1541 		WREG32(reg_offset, sdma_cntl);
1542 	}
1543 
1544 	return 0;
1545 }
1546 
1547 static int sdma_v5_2_process_trap_irq(struct amdgpu_device *adev,
1548 				      struct amdgpu_irq_src *source,
1549 				      struct amdgpu_iv_entry *entry)
1550 {
1551 	DRM_DEBUG("IH: SDMA trap\n");
1552 
1553 	if (drm_WARN_ON_ONCE(&adev->ddev,
1554 			     adev->enable_mes &&
1555 			     (entry->src_data[0] & AMDGPU_FENCE_MES_QUEUE_FLAG)))
1556 		return 0;
1557 
1558 	switch (entry->client_id) {
1559 	case SOC15_IH_CLIENTID_SDMA0:
1560 		switch (entry->ring_id) {
1561 		case 0:
1562 			amdgpu_fence_process(&adev->sdma.instance[0].ring);
1563 			break;
1564 		case 1:
1565 			/* XXX compute */
1566 			break;
1567 		case 2:
1568 			/* XXX compute */
1569 			break;
1570 		case 3:
1571 			/* XXX page queue*/
1572 			break;
1573 		}
1574 		break;
1575 	case SOC15_IH_CLIENTID_SDMA1:
1576 		switch (entry->ring_id) {
1577 		case 0:
1578 			amdgpu_fence_process(&adev->sdma.instance[1].ring);
1579 			break;
1580 		case 1:
1581 			/* XXX compute */
1582 			break;
1583 		case 2:
1584 			/* XXX compute */
1585 			break;
1586 		case 3:
1587 			/* XXX page queue*/
1588 			break;
1589 		}
1590 		break;
1591 	case SOC15_IH_CLIENTID_SDMA2:
1592 		switch (entry->ring_id) {
1593 		case 0:
1594 			amdgpu_fence_process(&adev->sdma.instance[2].ring);
1595 			break;
1596 		case 1:
1597 			/* XXX compute */
1598 			break;
1599 		case 2:
1600 			/* XXX compute */
1601 			break;
1602 		case 3:
1603 			/* XXX page queue*/
1604 			break;
1605 		}
1606 		break;
1607 	case SOC15_IH_CLIENTID_SDMA3_Sienna_Cichlid:
1608 		switch (entry->ring_id) {
1609 		case 0:
1610 			amdgpu_fence_process(&adev->sdma.instance[3].ring);
1611 			break;
1612 		case 1:
1613 			/* XXX compute */
1614 			break;
1615 		case 2:
1616 			/* XXX compute */
1617 			break;
1618 		case 3:
1619 			/* XXX page queue*/
1620 			break;
1621 		}
1622 		break;
1623 	}
1624 	return 0;
1625 }
1626 
1627 static int sdma_v5_2_process_illegal_inst_irq(struct amdgpu_device *adev,
1628 					      struct amdgpu_irq_src *source,
1629 					      struct amdgpu_iv_entry *entry)
1630 {
1631 	return 0;
1632 }
1633 
1634 static bool sdma_v5_2_firmware_mgcg_support(struct amdgpu_device *adev,
1635 						     int i)
1636 {
1637 	switch (amdgpu_ip_version(adev, SDMA0_HWIP, 0)) {
1638 	case IP_VERSION(5, 2, 1):
1639 		if (adev->sdma.instance[i].fw_version < 70)
1640 			return false;
1641 		break;
1642 	case IP_VERSION(5, 2, 3):
1643 		if (adev->sdma.instance[i].fw_version < 47)
1644 			return false;
1645 		break;
1646 	case IP_VERSION(5, 2, 7):
1647 		if (adev->sdma.instance[i].fw_version < 9)
1648 			return false;
1649 		break;
1650 	default:
1651 		return true;
1652 	}
1653 
1654 	return true;
1655 
1656 }
1657 
1658 static void sdma_v5_2_update_medium_grain_clock_gating(struct amdgpu_device *adev,
1659 						       bool enable)
1660 {
1661 	uint32_t data, def;
1662 	int i;
1663 
1664 	for (i = 0; i < adev->sdma.num_instances; i++) {
1665 
1666 		if (!sdma_v5_2_firmware_mgcg_support(adev, i))
1667 			adev->cg_flags &= ~AMD_CG_SUPPORT_SDMA_MGCG;
1668 
1669 		if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
1670 			/* Enable sdma clock gating */
1671 			def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL));
1672 			data &= ~(SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1673 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1674 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1675 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1676 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK |
1677 				  SDMA0_CLK_CTRL__SOFT_OVERRIDER_REG_MASK);
1678 			if (def != data)
1679 				WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data);
1680 		} else {
1681 			/* Disable sdma clock gating */
1682 			def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL));
1683 			data |= (SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1684 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1685 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1686 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1687 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK |
1688 				 SDMA0_CLK_CTRL__SOFT_OVERRIDER_REG_MASK);
1689 			if (def != data)
1690 				WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data);
1691 		}
1692 	}
1693 }
1694 
1695 static void sdma_v5_2_update_medium_grain_light_sleep(struct amdgpu_device *adev,
1696 						      bool enable)
1697 {
1698 	uint32_t data, def;
1699 	int i;
1700 
1701 	for (i = 0; i < adev->sdma.num_instances; i++) {
1702 		if (adev->sdma.instance[i].fw_version < 70 &&
1703 		    amdgpu_ip_version(adev, SDMA0_HWIP, 0) ==
1704 			    IP_VERSION(5, 2, 1))
1705 			adev->cg_flags &= ~AMD_CG_SUPPORT_SDMA_LS;
1706 
1707 		if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) {
1708 			/* Enable sdma mem light sleep */
1709 			def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL));
1710 			data |= SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1711 			if (def != data)
1712 				WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data);
1713 
1714 		} else {
1715 			/* Disable sdma mem light sleep */
1716 			def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL));
1717 			data &= ~SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1718 			if (def != data)
1719 				WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data);
1720 
1721 		}
1722 	}
1723 }
1724 
1725 static int sdma_v5_2_set_clockgating_state(struct amdgpu_ip_block *ip_block,
1726 					   enum amd_clockgating_state state)
1727 {
1728 	struct amdgpu_device *adev = ip_block->adev;
1729 
1730 	if (amdgpu_sriov_vf(adev))
1731 		return 0;
1732 
1733 	switch (amdgpu_ip_version(adev, SDMA0_HWIP, 0)) {
1734 	case IP_VERSION(5, 2, 0):
1735 	case IP_VERSION(5, 2, 2):
1736 	case IP_VERSION(5, 2, 1):
1737 	case IP_VERSION(5, 2, 4):
1738 	case IP_VERSION(5, 2, 5):
1739 	case IP_VERSION(5, 2, 6):
1740 	case IP_VERSION(5, 2, 3):
1741 	case IP_VERSION(5, 2, 7):
1742 		sdma_v5_2_update_medium_grain_clock_gating(adev,
1743 				state == AMD_CG_STATE_GATE);
1744 		sdma_v5_2_update_medium_grain_light_sleep(adev,
1745 				state == AMD_CG_STATE_GATE);
1746 		break;
1747 	default:
1748 		break;
1749 	}
1750 
1751 	return 0;
1752 }
1753 
1754 static int sdma_v5_2_set_powergating_state(struct amdgpu_ip_block *ip_block,
1755 					  enum amd_powergating_state state)
1756 {
1757 	return 0;
1758 }
1759 
1760 static void sdma_v5_2_get_clockgating_state(struct amdgpu_ip_block *ip_block, u64 *flags)
1761 {
1762 	struct amdgpu_device *adev = ip_block->adev;
1763 	int data;
1764 
1765 	if (amdgpu_sriov_vf(adev))
1766 		*flags = 0;
1767 
1768 	/* AMD_CG_SUPPORT_SDMA_MGCG */
1769 	data = RREG32(sdma_v5_2_get_reg_offset(adev, 0, mmSDMA0_CLK_CTRL));
1770 	if (!(data & SDMA0_CLK_CTRL__CGCG_EN_OVERRIDE_MASK))
1771 		*flags |= AMD_CG_SUPPORT_SDMA_MGCG;
1772 
1773 	/* AMD_CG_SUPPORT_SDMA_LS */
1774 	data = RREG32_KIQ(sdma_v5_2_get_reg_offset(adev, 0, mmSDMA0_POWER_CNTL));
1775 	if (data & SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK)
1776 		*flags |= AMD_CG_SUPPORT_SDMA_LS;
1777 }
1778 
1779 static void sdma_v5_2_ring_begin_use(struct amdgpu_ring *ring)
1780 {
1781 	struct amdgpu_device *adev = ring->adev;
1782 
1783 	/* SDMA 5.2.3 (RMB) FW doesn't seem to properly
1784 	 * disallow GFXOFF in some cases leading to
1785 	 * hangs in SDMA.  Disallow GFXOFF while SDMA is active.
1786 	 * We can probably just limit this to 5.2.3,
1787 	 * but it shouldn't hurt for other parts since
1788 	 * this GFXOFF will be disallowed anyway when SDMA is
1789 	 * active, this just makes it explicit.
1790 	 * sdma_v5_2_ring_set_wptr() takes advantage of this
1791 	 * to update the wptr because sometimes SDMA seems to miss
1792 	 * doorbells when entering PG.  If you remove this, update
1793 	 * sdma_v5_2_ring_set_wptr() as well!
1794 	 */
1795 	amdgpu_gfx_off_ctrl(adev, false);
1796 }
1797 
1798 static void sdma_v5_2_ring_end_use(struct amdgpu_ring *ring)
1799 {
1800 	struct amdgpu_device *adev = ring->adev;
1801 
1802 	/* SDMA 5.2.3 (RMB) FW doesn't seem to properly
1803 	 * disallow GFXOFF in some cases leading to
1804 	 * hangs in SDMA.  Allow GFXOFF when SDMA is complete.
1805 	 */
1806 	amdgpu_gfx_off_ctrl(adev, true);
1807 }
1808 
1809 static void sdma_v5_2_print_ip_state(struct amdgpu_ip_block *ip_block, struct drm_printer *p)
1810 {
1811 	struct amdgpu_device *adev = ip_block->adev;
1812 	int i, j;
1813 	uint32_t reg_count = ARRAY_SIZE(sdma_reg_list_5_2);
1814 	uint32_t instance_offset;
1815 
1816 	if (!adev->sdma.ip_dump)
1817 		return;
1818 
1819 	drm_printf(p, "num_instances:%d\n", adev->sdma.num_instances);
1820 	for (i = 0; i < adev->sdma.num_instances; i++) {
1821 		instance_offset = i * reg_count;
1822 		drm_printf(p, "\nInstance:%d\n", i);
1823 
1824 		for (j = 0; j < reg_count; j++)
1825 			drm_printf(p, "%-50s \t 0x%08x\n", sdma_reg_list_5_2[j].reg_name,
1826 				   adev->sdma.ip_dump[instance_offset + j]);
1827 	}
1828 }
1829 
1830 static void sdma_v5_2_dump_ip_state(struct amdgpu_ip_block *ip_block)
1831 {
1832 	struct amdgpu_device *adev = ip_block->adev;
1833 	int i, j;
1834 	uint32_t instance_offset;
1835 	uint32_t reg_count = ARRAY_SIZE(sdma_reg_list_5_2);
1836 
1837 	if (!adev->sdma.ip_dump)
1838 		return;
1839 
1840 	amdgpu_gfx_off_ctrl(adev, false);
1841 	for (i = 0; i < adev->sdma.num_instances; i++) {
1842 		instance_offset = i * reg_count;
1843 		for (j = 0; j < reg_count; j++)
1844 			adev->sdma.ip_dump[instance_offset + j] =
1845 				RREG32(sdma_v5_2_get_reg_offset(adev, i,
1846 				       sdma_reg_list_5_2[j].reg_offset));
1847 	}
1848 	amdgpu_gfx_off_ctrl(adev, true);
1849 }
1850 
1851 static const struct amd_ip_funcs sdma_v5_2_ip_funcs = {
1852 	.name = "sdma_v5_2",
1853 	.early_init = sdma_v5_2_early_init,
1854 	.sw_init = sdma_v5_2_sw_init,
1855 	.sw_fini = sdma_v5_2_sw_fini,
1856 	.hw_init = sdma_v5_2_hw_init,
1857 	.hw_fini = sdma_v5_2_hw_fini,
1858 	.suspend = sdma_v5_2_suspend,
1859 	.resume = sdma_v5_2_resume,
1860 	.wait_for_idle = sdma_v5_2_wait_for_idle,
1861 	.soft_reset = sdma_v5_2_soft_reset,
1862 	.set_clockgating_state = sdma_v5_2_set_clockgating_state,
1863 	.set_powergating_state = sdma_v5_2_set_powergating_state,
1864 	.get_clockgating_state = sdma_v5_2_get_clockgating_state,
1865 	.dump_ip_state = sdma_v5_2_dump_ip_state,
1866 	.print_ip_state = sdma_v5_2_print_ip_state,
1867 };
1868 
1869 static const struct amdgpu_ring_funcs sdma_v5_2_ring_funcs = {
1870 	.type = AMDGPU_RING_TYPE_SDMA,
1871 	.align_mask = 0xf,
1872 	.nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP),
1873 	.support_64bit_ptrs = true,
1874 	.secure_submission_supported = true,
1875 	.get_rptr = sdma_v5_2_ring_get_rptr,
1876 	.get_wptr = sdma_v5_2_ring_get_wptr,
1877 	.set_wptr = sdma_v5_2_ring_set_wptr,
1878 	.emit_frame_size =
1879 		5 + /* sdma_v5_2_ring_init_cond_exec */
1880 		6 + /* sdma_v5_2_ring_emit_hdp_flush */
1881 		3 + /* hdp_invalidate */
1882 		6 + /* sdma_v5_2_ring_emit_pipeline_sync */
1883 		/* sdma_v5_2_ring_emit_vm_flush */
1884 		SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 +
1885 		SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6 +
1886 		10 + 10 + 10, /* sdma_v5_2_ring_emit_fence x3 for user fence, vm fence */
1887 	.emit_ib_size = 7 + 6, /* sdma_v5_2_ring_emit_ib */
1888 	.emit_ib = sdma_v5_2_ring_emit_ib,
1889 	.emit_mem_sync = sdma_v5_2_ring_emit_mem_sync,
1890 	.emit_fence = sdma_v5_2_ring_emit_fence,
1891 	.emit_pipeline_sync = sdma_v5_2_ring_emit_pipeline_sync,
1892 	.emit_vm_flush = sdma_v5_2_ring_emit_vm_flush,
1893 	.emit_hdp_flush = sdma_v5_2_ring_emit_hdp_flush,
1894 	.test_ring = sdma_v5_2_ring_test_ring,
1895 	.test_ib = sdma_v5_2_ring_test_ib,
1896 	.insert_nop = sdma_v5_2_ring_insert_nop,
1897 	.pad_ib = sdma_v5_2_ring_pad_ib,
1898 	.begin_use = sdma_v5_2_ring_begin_use,
1899 	.end_use = sdma_v5_2_ring_end_use,
1900 	.emit_wreg = sdma_v5_2_ring_emit_wreg,
1901 	.emit_reg_wait = sdma_v5_2_ring_emit_reg_wait,
1902 	.emit_reg_write_reg_wait = sdma_v5_2_ring_emit_reg_write_reg_wait,
1903 	.init_cond_exec = sdma_v5_2_ring_init_cond_exec,
1904 	.preempt_ib = sdma_v5_2_ring_preempt_ib,
1905 	.reset = amdgpu_sdma_reset_queue_legacy,
1906 };
1907 
1908 static void sdma_v5_2_set_ring_funcs(struct amdgpu_device *adev)
1909 {
1910 	int i;
1911 
1912 	for (i = 0; i < adev->sdma.num_instances; i++) {
1913 		adev->sdma.instance[i].ring.funcs = &sdma_v5_2_ring_funcs;
1914 		adev->sdma.instance[i].ring.me = i;
1915 	}
1916 }
1917 
1918 static const struct amdgpu_irq_src_funcs sdma_v5_2_trap_irq_funcs = {
1919 	.set = sdma_v5_2_set_trap_irq_state,
1920 	.process = sdma_v5_2_process_trap_irq,
1921 };
1922 
1923 static const struct amdgpu_irq_src_funcs sdma_v5_2_illegal_inst_irq_funcs = {
1924 	.process = sdma_v5_2_process_illegal_inst_irq,
1925 };
1926 
1927 static void sdma_v5_2_set_irq_funcs(struct amdgpu_device *adev)
1928 {
1929 	adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_INSTANCE0 +
1930 					adev->sdma.num_instances;
1931 	adev->sdma.trap_irq.funcs = &sdma_v5_2_trap_irq_funcs;
1932 	adev->sdma.illegal_inst_irq.funcs = &sdma_v5_2_illegal_inst_irq_funcs;
1933 }
1934 
1935 /**
1936  * sdma_v5_2_emit_copy_buffer - copy buffer using the sDMA engine
1937  *
1938  * @ib: indirect buffer to copy to
1939  * @src_offset: src GPU address
1940  * @dst_offset: dst GPU address
1941  * @byte_count: number of bytes to xfer
1942  * @copy_flags: copy flags for the buffers
1943  *
1944  * Copy GPU buffers using the DMA engine.
1945  * Used by the amdgpu ttm implementation to move pages if
1946  * registered as the asic copy callback.
1947  */
1948 static void sdma_v5_2_emit_copy_buffer(struct amdgpu_ib *ib,
1949 				       uint64_t src_offset,
1950 				       uint64_t dst_offset,
1951 				       uint32_t byte_count,
1952 				       uint32_t copy_flags)
1953 {
1954 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1955 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR) |
1956 		SDMA_PKT_COPY_LINEAR_HEADER_TMZ((copy_flags & AMDGPU_COPY_FLAGS_TMZ) ? 1 : 0);
1957 	ib->ptr[ib->length_dw++] = byte_count - 1;
1958 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1959 	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1960 	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1961 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1962 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1963 }
1964 
1965 /**
1966  * sdma_v5_2_emit_fill_buffer - fill buffer using the sDMA engine
1967  *
1968  * @ib: indirect buffer to fill
1969  * @src_data: value to write to buffer
1970  * @dst_offset: dst GPU address
1971  * @byte_count: number of bytes to xfer
1972  *
1973  * Fill GPU buffers using the DMA engine.
1974  */
1975 static void sdma_v5_2_emit_fill_buffer(struct amdgpu_ib *ib,
1976 				       uint32_t src_data,
1977 				       uint64_t dst_offset,
1978 				       uint32_t byte_count)
1979 {
1980 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL);
1981 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1982 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1983 	ib->ptr[ib->length_dw++] = src_data;
1984 	ib->ptr[ib->length_dw++] = byte_count - 1;
1985 }
1986 
1987 static const struct amdgpu_buffer_funcs sdma_v5_2_buffer_funcs = {
1988 	.copy_max_bytes = 1 << 30,
1989 	.copy_num_dw = 7,
1990 	.emit_copy_buffer = sdma_v5_2_emit_copy_buffer,
1991 
1992 	.fill_max_bytes = 1 << 30, /* HW supports 1 << 30, but PAL uses 1 << 22 */
1993 	.fill_num_dw = 5,
1994 	.emit_fill_buffer = sdma_v5_2_emit_fill_buffer,
1995 };
1996 
1997 static void sdma_v5_2_set_buffer_funcs(struct amdgpu_device *adev)
1998 {
1999 	amdgpu_sdma_set_buffer_funcs_scheds(adev, &sdma_v5_2_buffer_funcs);
2000 }
2001 
2002 const struct amdgpu_ip_block_version sdma_v5_2_ip_block = {
2003 	.type = AMD_IP_BLOCK_TYPE_SDMA,
2004 	.major = 5,
2005 	.minor = 2,
2006 	.rev = 0,
2007 	.funcs = &sdma_v5_2_ip_funcs,
2008 };
2009