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_trace.h"
32 
33 #include "gc/gc_10_3_0_offset.h"
34 #include "gc/gc_10_3_0_sh_mask.h"
35 #include "ivsrcid/sdma0/irqsrcs_sdma0_5_0.h"
36 #include "ivsrcid/sdma1/irqsrcs_sdma1_5_0.h"
37 #include "ivsrcid/sdma2/irqsrcs_sdma2_5_0.h"
38 #include "ivsrcid/sdma3/irqsrcs_sdma3_5_0.h"
39 
40 #include "soc15_common.h"
41 #include "soc15.h"
42 #include "navi10_sdma_pkt_open.h"
43 #include "nbio_v2_3.h"
44 #include "sdma_common.h"
45 #include "sdma_v5_2.h"
46 
47 MODULE_FIRMWARE("amdgpu/sienna_cichlid_sdma.bin");
48 MODULE_FIRMWARE("amdgpu/navy_flounder_sdma.bin");
49 MODULE_FIRMWARE("amdgpu/dimgrey_cavefish_sdma.bin");
50 MODULE_FIRMWARE("amdgpu/beige_goby_sdma.bin");
51 
52 MODULE_FIRMWARE("amdgpu/vangogh_sdma.bin");
53 MODULE_FIRMWARE("amdgpu/yellow_carp_sdma.bin");
54 MODULE_FIRMWARE("amdgpu/sdma_5_2_6.bin");
55 MODULE_FIRMWARE("amdgpu/sdma_5_2_7.bin");
56 
57 #define SDMA1_REG_OFFSET 0x600
58 #define SDMA3_REG_OFFSET 0x400
59 #define SDMA0_HYP_DEC_REG_START 0x5880
60 #define SDMA0_HYP_DEC_REG_END 0x5893
61 #define SDMA1_HYP_DEC_REG_OFFSET 0x20
62 
63 static void sdma_v5_2_set_ring_funcs(struct amdgpu_device *adev);
64 static void sdma_v5_2_set_buffer_funcs(struct amdgpu_device *adev);
65 static void sdma_v5_2_set_vm_pte_funcs(struct amdgpu_device *adev);
66 static void sdma_v5_2_set_irq_funcs(struct amdgpu_device *adev);
67 
68 static u32 sdma_v5_2_get_reg_offset(struct amdgpu_device *adev, u32 instance, u32 internal_offset)
69 {
70 	u32 base;
71 
72 	if (internal_offset >= SDMA0_HYP_DEC_REG_START &&
73 	    internal_offset <= SDMA0_HYP_DEC_REG_END) {
74 		base = adev->reg_offset[GC_HWIP][0][1];
75 		if (instance != 0)
76 			internal_offset += SDMA1_HYP_DEC_REG_OFFSET * instance;
77 	} else {
78 		if (instance < 2) {
79 			base = adev->reg_offset[GC_HWIP][0][0];
80 			if (instance == 1)
81 				internal_offset += SDMA1_REG_OFFSET;
82 		} else {
83 			base = adev->reg_offset[GC_HWIP][0][2];
84 			if (instance == 3)
85 				internal_offset += SDMA3_REG_OFFSET;
86 		}
87 	}
88 
89 	return base + internal_offset;
90 }
91 
92 static unsigned sdma_v5_2_ring_init_cond_exec(struct amdgpu_ring *ring)
93 {
94 	unsigned ret;
95 
96 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_COND_EXE));
97 	amdgpu_ring_write(ring, lower_32_bits(ring->cond_exe_gpu_addr));
98 	amdgpu_ring_write(ring, upper_32_bits(ring->cond_exe_gpu_addr));
99 	amdgpu_ring_write(ring, 1);
100 	ret = ring->wptr & ring->buf_mask;/* this is the offset we need patch later */
101 	amdgpu_ring_write(ring, 0x55aa55aa);/* insert dummy here and patch it later */
102 
103 	return ret;
104 }
105 
106 static void sdma_v5_2_ring_patch_cond_exec(struct amdgpu_ring *ring,
107 					   unsigned offset)
108 {
109 	unsigned cur;
110 
111 	BUG_ON(offset > ring->buf_mask);
112 	BUG_ON(ring->ring[offset] != 0x55aa55aa);
113 
114 	cur = (ring->wptr - 1) & ring->buf_mask;
115 	if (cur > offset)
116 		ring->ring[offset] = cur - offset;
117 	else
118 		ring->ring[offset] = (ring->buf_mask + 1) - offset + cur;
119 }
120 
121 /**
122  * sdma_v5_2_ring_get_rptr - get the current read pointer
123  *
124  * @ring: amdgpu ring pointer
125  *
126  * Get the current rptr from the hardware (NAVI10+).
127  */
128 static uint64_t sdma_v5_2_ring_get_rptr(struct amdgpu_ring *ring)
129 {
130 	u64 *rptr;
131 
132 	/* XXX check if swapping is necessary on BE */
133 	rptr = (u64 *)ring->rptr_cpu_addr;
134 
135 	DRM_DEBUG("rptr before shift == 0x%016llx\n", *rptr);
136 	return ((*rptr) >> 2);
137 }
138 
139 /**
140  * sdma_v5_2_ring_get_wptr - get the current write pointer
141  *
142  * @ring: amdgpu ring pointer
143  *
144  * Get the current wptr from the hardware (NAVI10+).
145  */
146 static uint64_t sdma_v5_2_ring_get_wptr(struct amdgpu_ring *ring)
147 {
148 	struct amdgpu_device *adev = ring->adev;
149 	u64 wptr;
150 
151 	if (ring->use_doorbell) {
152 		/* XXX check if swapping is necessary on BE */
153 		wptr = READ_ONCE(*((u64 *)ring->wptr_cpu_addr));
154 		DRM_DEBUG("wptr/doorbell before shift == 0x%016llx\n", wptr);
155 	} else {
156 		wptr = RREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI));
157 		wptr = wptr << 32;
158 		wptr |= RREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR));
159 		DRM_DEBUG("wptr before shift [%i] wptr == 0x%016llx\n", ring->me, wptr);
160 	}
161 
162 	return wptr >> 2;
163 }
164 
165 /**
166  * sdma_v5_2_ring_set_wptr - commit the write pointer
167  *
168  * @ring: amdgpu ring pointer
169  *
170  * Write the wptr back to the hardware (NAVI10+).
171  */
172 static void sdma_v5_2_ring_set_wptr(struct amdgpu_ring *ring)
173 {
174 	struct amdgpu_device *adev = ring->adev;
175 
176 	DRM_DEBUG("Setting write pointer\n");
177 	if (ring->use_doorbell) {
178 		DRM_DEBUG("Using doorbell -- "
179 				"wptr_offs == 0x%08x "
180 				"lower_32_bits(ring->wptr << 2) == 0x%08x "
181 				"upper_32_bits(ring->wptr << 2) == 0x%08x\n",
182 				ring->wptr_offs,
183 				lower_32_bits(ring->wptr << 2),
184 				upper_32_bits(ring->wptr << 2));
185 		/* XXX check if swapping is necessary on BE */
186 		atomic64_set((atomic64_t *)ring->wptr_cpu_addr,
187 			     ring->wptr << 2);
188 		DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016llx)\n",
189 				ring->doorbell_index, ring->wptr << 2);
190 		WDOORBELL64(ring->doorbell_index, ring->wptr << 2);
191 	} else {
192 		DRM_DEBUG("Not using doorbell -- "
193 				"mmSDMA%i_GFX_RB_WPTR == 0x%08x "
194 				"mmSDMA%i_GFX_RB_WPTR_HI == 0x%08x\n",
195 				ring->me,
196 				lower_32_bits(ring->wptr << 2),
197 				ring->me,
198 				upper_32_bits(ring->wptr << 2));
199 		WREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR),
200 			lower_32_bits(ring->wptr << 2));
201 		WREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI),
202 			upper_32_bits(ring->wptr << 2));
203 	}
204 }
205 
206 static void sdma_v5_2_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
207 {
208 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
209 	int i;
210 
211 	for (i = 0; i < count; i++)
212 		if (sdma && sdma->burst_nop && (i == 0))
213 			amdgpu_ring_write(ring, ring->funcs->nop |
214 				SDMA_PKT_NOP_HEADER_COUNT(count - 1));
215 		else
216 			amdgpu_ring_write(ring, ring->funcs->nop);
217 }
218 
219 /**
220  * sdma_v5_2_ring_emit_ib - Schedule an IB on the DMA engine
221  *
222  * @ring: amdgpu ring pointer
223  * @job: job to retrieve vmid from
224  * @ib: IB object to schedule
225  * @flags: unused
226  *
227  * Schedule an IB in the DMA ring.
228  */
229 static void sdma_v5_2_ring_emit_ib(struct amdgpu_ring *ring,
230 				   struct amdgpu_job *job,
231 				   struct amdgpu_ib *ib,
232 				   uint32_t flags)
233 {
234 	unsigned vmid = AMDGPU_JOB_GET_VMID(job);
235 	uint64_t csa_mc_addr = amdgpu_sdma_get_csa_mc_addr(ring, vmid);
236 
237 	/* An IB packet must end on a 8 DW boundary--the next dword
238 	 * must be on a 8-dword boundary. Our IB packet below is 6
239 	 * dwords long, thus add x number of NOPs, such that, in
240 	 * modular arithmetic,
241 	 * wptr + 6 + x = 8k, k >= 0, which in C is,
242 	 * (wptr + 6 + x) % 8 = 0.
243 	 * The expression below, is a solution of x.
244 	 */
245 	sdma_v5_2_ring_insert_nop(ring, (2 - lower_32_bits(ring->wptr)) & 7);
246 
247 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) |
248 			  SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf));
249 	/* base must be 32 byte aligned */
250 	amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
251 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
252 	amdgpu_ring_write(ring, ib->length_dw);
253 	amdgpu_ring_write(ring, lower_32_bits(csa_mc_addr));
254 	amdgpu_ring_write(ring, upper_32_bits(csa_mc_addr));
255 }
256 
257 /**
258  * sdma_v5_2_ring_emit_mem_sync - flush the IB by graphics cache rinse
259  *
260  * @ring: amdgpu ring pointer
261  *
262  * flush the IB by graphics cache rinse.
263  */
264 static void sdma_v5_2_ring_emit_mem_sync(struct amdgpu_ring *ring)
265 {
266 	uint32_t gcr_cntl = SDMA_GCR_GL2_INV | SDMA_GCR_GL2_WB |
267 			    SDMA_GCR_GLM_INV | SDMA_GCR_GL1_INV |
268 			    SDMA_GCR_GLV_INV | SDMA_GCR_GLK_INV |
269 			    SDMA_GCR_GLI_INV(1);
270 
271 	/* flush entire cache L0/L1/L2, this can be optimized by performance requirement */
272 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_GCR_REQ));
273 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD1_BASE_VA_31_7(0));
274 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD2_GCR_CONTROL_15_0(gcr_cntl) |
275 			SDMA_PKT_GCR_REQ_PAYLOAD2_BASE_VA_47_32(0));
276 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD3_LIMIT_VA_31_7(0) |
277 			SDMA_PKT_GCR_REQ_PAYLOAD3_GCR_CONTROL_18_16(gcr_cntl >> 16));
278 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD4_LIMIT_VA_47_32(0) |
279 			SDMA_PKT_GCR_REQ_PAYLOAD4_VMID(0));
280 }
281 
282 /**
283  * sdma_v5_2_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
284  *
285  * @ring: amdgpu ring pointer
286  *
287  * Emit an hdp flush packet on the requested DMA ring.
288  */
289 static void sdma_v5_2_ring_emit_hdp_flush(struct amdgpu_ring *ring)
290 {
291 	struct amdgpu_device *adev = ring->adev;
292 	u32 ref_and_mask = 0;
293 	const struct nbio_hdp_flush_reg *nbio_hf_reg = adev->nbio.hdp_flush_reg;
294 
295 	ref_and_mask = nbio_hf_reg->ref_and_mask_sdma0 << ring->me;
296 
297 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
298 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) |
299 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
300 	amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_done_offset(adev)) << 2);
301 	amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_req_offset(adev)) << 2);
302 	amdgpu_ring_write(ring, ref_and_mask); /* reference */
303 	amdgpu_ring_write(ring, ref_and_mask); /* mask */
304 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
305 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
306 }
307 
308 /**
309  * sdma_v5_2_ring_emit_fence - emit a fence on the DMA ring
310  *
311  * @ring: amdgpu ring pointer
312  * @addr: address
313  * @seq: sequence number
314  * @flags: fence related flags
315  *
316  * Add a DMA fence packet to the ring to write
317  * the fence seq number and DMA trap packet to generate
318  * an interrupt if needed.
319  */
320 static void sdma_v5_2_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
321 				      unsigned flags)
322 {
323 	bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
324 	/* write the fence */
325 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) |
326 			  SDMA_PKT_FENCE_HEADER_MTYPE(0x3)); /* Ucached(UC) */
327 	/* zero in first two bits */
328 	BUG_ON(addr & 0x3);
329 	amdgpu_ring_write(ring, lower_32_bits(addr));
330 	amdgpu_ring_write(ring, upper_32_bits(addr));
331 	amdgpu_ring_write(ring, lower_32_bits(seq));
332 
333 	/* optionally write high bits as well */
334 	if (write64bit) {
335 		addr += 4;
336 		amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) |
337 				  SDMA_PKT_FENCE_HEADER_MTYPE(0x3));
338 		/* zero in first two bits */
339 		BUG_ON(addr & 0x3);
340 		amdgpu_ring_write(ring, lower_32_bits(addr));
341 		amdgpu_ring_write(ring, upper_32_bits(addr));
342 		amdgpu_ring_write(ring, upper_32_bits(seq));
343 	}
344 
345 	if ((flags & AMDGPU_FENCE_FLAG_INT)) {
346 		uint32_t ctx = ring->is_mes_queue ?
347 			(ring->hw_queue_id | AMDGPU_FENCE_MES_QUEUE_FLAG) : 0;
348 		/* generate an interrupt */
349 		amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP));
350 		amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(ctx));
351 	}
352 }
353 
354 
355 /**
356  * sdma_v5_2_gfx_stop - stop the gfx async dma engines
357  *
358  * @adev: amdgpu_device pointer
359  *
360  * Stop the gfx async dma ring buffers.
361  */
362 static void sdma_v5_2_gfx_stop(struct amdgpu_device *adev)
363 {
364 	u32 rb_cntl, ib_cntl;
365 	int i;
366 
367 	amdgpu_sdma_unset_buffer_funcs_helper(adev);
368 
369 	for (i = 0; i < adev->sdma.num_instances; i++) {
370 		rb_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL));
371 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 0);
372 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
373 		ib_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL));
374 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 0);
375 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl);
376 	}
377 }
378 
379 /**
380  * sdma_v5_2_rlc_stop - stop the compute async dma engines
381  *
382  * @adev: amdgpu_device pointer
383  *
384  * Stop the compute async dma queues.
385  */
386 static void sdma_v5_2_rlc_stop(struct amdgpu_device *adev)
387 {
388 	/* XXX todo */
389 }
390 
391 /**
392  * sdma_v5_2_ctx_switch_enable - stop the async dma engines context switch
393  *
394  * @adev: amdgpu_device pointer
395  * @enable: enable/disable the DMA MEs context switch.
396  *
397  * Halt or unhalt the async dma engines context switch.
398  */
399 static void sdma_v5_2_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
400 {
401 	u32 f32_cntl, phase_quantum = 0;
402 	int i;
403 
404 	if (amdgpu_sdma_phase_quantum) {
405 		unsigned value = amdgpu_sdma_phase_quantum;
406 		unsigned unit = 0;
407 
408 		while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
409 				SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) {
410 			value = (value + 1) >> 1;
411 			unit++;
412 		}
413 		if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
414 			    SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) {
415 			value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
416 				 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT);
417 			unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
418 				SDMA0_PHASE0_QUANTUM__UNIT__SHIFT);
419 			WARN_ONCE(1,
420 			"clamping sdma_phase_quantum to %uK clock cycles\n",
421 				  value << unit);
422 		}
423 		phase_quantum =
424 			value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
425 			unit  << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
426 	}
427 
428 	for (i = 0; i < adev->sdma.num_instances; i++) {
429 		if (enable && amdgpu_sdma_phase_quantum) {
430 			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE0_QUANTUM),
431 			       phase_quantum);
432 			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE1_QUANTUM),
433 			       phase_quantum);
434 			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE2_QUANTUM),
435 			       phase_quantum);
436 		}
437 
438 		if (!amdgpu_sriov_vf(adev)) {
439 			f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
440 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
441 					AUTO_CTXSW_ENABLE, enable ? 1 : 0);
442 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl);
443 		}
444 	}
445 
446 }
447 
448 /**
449  * sdma_v5_2_enable - stop the async dma engines
450  *
451  * @adev: amdgpu_device pointer
452  * @enable: enable/disable the DMA MEs.
453  *
454  * Halt or unhalt the async dma engines.
455  */
456 static void sdma_v5_2_enable(struct amdgpu_device *adev, bool enable)
457 {
458 	u32 f32_cntl;
459 	int i;
460 
461 	if (!enable) {
462 		sdma_v5_2_gfx_stop(adev);
463 		sdma_v5_2_rlc_stop(adev);
464 	}
465 
466 	if (!amdgpu_sriov_vf(adev)) {
467 		for (i = 0; i < adev->sdma.num_instances; i++) {
468 			f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
469 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, enable ? 0 : 1);
470 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl);
471 		}
472 	}
473 }
474 
475 /**
476  * sdma_v5_2_gfx_resume - setup and start the async dma engines
477  *
478  * @adev: amdgpu_device pointer
479  *
480  * Set up the gfx DMA ring buffers and enable them.
481  * Returns 0 for success, error for failure.
482  */
483 static int sdma_v5_2_gfx_resume(struct amdgpu_device *adev)
484 {
485 	struct amdgpu_ring *ring;
486 	u32 rb_cntl, ib_cntl;
487 	u32 rb_bufsz;
488 	u32 doorbell;
489 	u32 doorbell_offset;
490 	u32 temp;
491 	u32 wptr_poll_cntl;
492 	u64 wptr_gpu_addr;
493 	int i, r;
494 
495 	for (i = 0; i < adev->sdma.num_instances; i++) {
496 		ring = &adev->sdma.instance[i].ring;
497 
498 		if (!amdgpu_sriov_vf(adev))
499 			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL), 0);
500 
501 		/* Set ring buffer size in dwords */
502 		rb_bufsz = order_base_2(ring->ring_size / 4);
503 		rb_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL));
504 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz);
505 #ifdef __BIG_ENDIAN
506 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1);
507 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL,
508 					RPTR_WRITEBACK_SWAP_ENABLE, 1);
509 #endif
510 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
511 
512 		/* Initialize the ring buffer's read and write pointers */
513 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR), 0);
514 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_HI), 0);
515 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), 0);
516 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), 0);
517 
518 		/* setup the wptr shadow polling */
519 		wptr_gpu_addr = ring->wptr_gpu_addr;
520 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_LO),
521 		       lower_32_bits(wptr_gpu_addr));
522 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_HI),
523 		       upper_32_bits(wptr_gpu_addr));
524 		wptr_poll_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i,
525 							 mmSDMA0_GFX_RB_WPTR_POLL_CNTL));
526 		wptr_poll_cntl = REG_SET_FIELD(wptr_poll_cntl,
527 					       SDMA0_GFX_RB_WPTR_POLL_CNTL,
528 					       F32_POLL_ENABLE, 1);
529 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_CNTL),
530 		       wptr_poll_cntl);
531 
532 		/* set the wb address whether it's enabled or not */
533 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_HI),
534 		       upper_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFF);
535 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_LO),
536 		       lower_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFC);
537 
538 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
539 
540 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE), ring->gpu_addr >> 8);
541 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE_HI), ring->gpu_addr >> 40);
542 
543 		ring->wptr = 0;
544 
545 		/* before programing wptr to a less value, need set minor_ptr_update first */
546 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 1);
547 
548 		if (!amdgpu_sriov_vf(adev)) { /* only bare-metal use register write for wptr */
549 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), lower_32_bits(ring->wptr << 2));
550 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), upper_32_bits(ring->wptr << 2));
551 		}
552 
553 		doorbell = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL));
554 		doorbell_offset = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL_OFFSET));
555 
556 		if (ring->use_doorbell) {
557 			doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 1);
558 			doorbell_offset = REG_SET_FIELD(doorbell_offset, SDMA0_GFX_DOORBELL_OFFSET,
559 					OFFSET, ring->doorbell_index);
560 		} else {
561 			doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 0);
562 		}
563 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL), doorbell);
564 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL_OFFSET), doorbell_offset);
565 
566 		adev->nbio.funcs->sdma_doorbell_range(adev, i, ring->use_doorbell,
567 						      ring->doorbell_index,
568 						      adev->doorbell_index.sdma_doorbell_range);
569 
570 		if (amdgpu_sriov_vf(adev))
571 			sdma_v5_2_ring_set_wptr(ring);
572 
573 		/* set minor_ptr_update to 0 after wptr programed */
574 
575 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 0);
576 
577 		/* SRIOV VF has no control of any of registers below */
578 		if (!amdgpu_sriov_vf(adev)) {
579 			/* set utc l1 enable flag always to 1 */
580 			temp = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
581 			temp = REG_SET_FIELD(temp, SDMA0_CNTL, UTC_L1_ENABLE, 1);
582 
583 			/* enable MCBP */
584 			temp = REG_SET_FIELD(temp, SDMA0_CNTL, MIDCMD_PREEMPT_ENABLE, 1);
585 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), temp);
586 
587 			/* Set up RESP_MODE to non-copy addresses */
588 			temp = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL));
589 			temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, RESP_MODE, 3);
590 			temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, REDO_DELAY, 9);
591 			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL), temp);
592 
593 			/* program default cache read and write policy */
594 			temp = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE));
595 			/* clean read policy and write policy bits */
596 			temp &= 0xFF0FFF;
597 			temp |= ((CACHE_READ_POLICY_L2__DEFAULT << 12) |
598 				 (CACHE_WRITE_POLICY_L2__DEFAULT << 14) |
599 				 SDMA0_UTCL1_PAGE__LLC_NOALLOC_MASK);
600 			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE), temp);
601 
602 			/* unhalt engine */
603 			temp = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
604 			temp = REG_SET_FIELD(temp, SDMA0_F32_CNTL, HALT, 0);
605 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), temp);
606 		}
607 
608 		/* enable DMA RB */
609 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1);
610 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
611 
612 		ib_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL));
613 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 1);
614 #ifdef __BIG_ENDIAN
615 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1);
616 #endif
617 		/* enable DMA IBs */
618 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl);
619 
620 		if (amdgpu_sriov_vf(adev)) { /* bare-metal sequence doesn't need below to lines */
621 			sdma_v5_2_ctx_switch_enable(adev, true);
622 			sdma_v5_2_enable(adev, true);
623 		}
624 
625 		r = amdgpu_ring_test_helper(ring);
626 		if (r)
627 			return r;
628 
629 		if (adev->mman.buffer_funcs_ring == ring)
630 			amdgpu_ttm_set_buffer_funcs_status(adev, true);
631 	}
632 
633 	return 0;
634 }
635 
636 /**
637  * sdma_v5_2_rlc_resume - setup and start the async dma engines
638  *
639  * @adev: amdgpu_device pointer
640  *
641  * Set up the compute DMA queues and enable them.
642  * Returns 0 for success, error for failure.
643  */
644 static int sdma_v5_2_rlc_resume(struct amdgpu_device *adev)
645 {
646 	return 0;
647 }
648 
649 /**
650  * sdma_v5_2_load_microcode - load the sDMA ME ucode
651  *
652  * @adev: amdgpu_device pointer
653  *
654  * Loads the sDMA0/1/2/3 ucode.
655  * Returns 0 for success, -EINVAL if the ucode is not available.
656  */
657 static int sdma_v5_2_load_microcode(struct amdgpu_device *adev)
658 {
659 	const struct sdma_firmware_header_v1_0 *hdr;
660 	const __le32 *fw_data;
661 	u32 fw_size;
662 	int i, j;
663 
664 	/* halt the MEs */
665 	sdma_v5_2_enable(adev, false);
666 
667 	for (i = 0; i < adev->sdma.num_instances; i++) {
668 		if (!adev->sdma.instance[i].fw)
669 			return -EINVAL;
670 
671 		hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
672 		amdgpu_ucode_print_sdma_hdr(&hdr->header);
673 		fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
674 
675 		fw_data = (const __le32 *)
676 			(adev->sdma.instance[i].fw->data +
677 				le32_to_cpu(hdr->header.ucode_array_offset_bytes));
678 
679 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), 0);
680 
681 		for (j = 0; j < fw_size; j++) {
682 			if (amdgpu_emu_mode == 1 && j % 500 == 0)
683 				msleep(1);
684 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UCODE_DATA), le32_to_cpup(fw_data++));
685 		}
686 
687 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), adev->sdma.instance[i].fw_version);
688 	}
689 
690 	return 0;
691 }
692 
693 static int sdma_v5_2_soft_reset(void *handle)
694 {
695 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
696 	u32 grbm_soft_reset;
697 	u32 tmp;
698 	int i;
699 
700 	for (i = 0; i < adev->sdma.num_instances; i++) {
701 		grbm_soft_reset = REG_SET_FIELD(0,
702 						GRBM_SOFT_RESET, SOFT_RESET_SDMA0,
703 						1);
704 		grbm_soft_reset <<= i;
705 
706 		tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET);
707 		tmp |= grbm_soft_reset;
708 		DRM_DEBUG("GRBM_SOFT_RESET=0x%08X\n", tmp);
709 		WREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET, tmp);
710 		tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET);
711 
712 		udelay(50);
713 
714 		tmp &= ~grbm_soft_reset;
715 		WREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET, tmp);
716 		tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET);
717 
718 		udelay(50);
719 	}
720 
721 	return 0;
722 }
723 
724 /**
725  * sdma_v5_2_start - setup and start the async dma engines
726  *
727  * @adev: amdgpu_device pointer
728  *
729  * Set up the DMA engines and enable them.
730  * Returns 0 for success, error for failure.
731  */
732 static int sdma_v5_2_start(struct amdgpu_device *adev)
733 {
734 	int r = 0;
735 
736 	if (amdgpu_sriov_vf(adev)) {
737 		sdma_v5_2_ctx_switch_enable(adev, false);
738 		sdma_v5_2_enable(adev, false);
739 
740 		/* set RB registers */
741 		r = sdma_v5_2_gfx_resume(adev);
742 		return r;
743 	}
744 
745 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
746 		r = sdma_v5_2_load_microcode(adev);
747 		if (r)
748 			return r;
749 
750 		/* The value of mmSDMA_F32_CNTL is invalid the moment after loading fw */
751 		if (amdgpu_emu_mode == 1)
752 			msleep(1000);
753 	}
754 
755 	sdma_v5_2_soft_reset(adev);
756 	/* unhalt the MEs */
757 	sdma_v5_2_enable(adev, true);
758 	/* enable sdma ring preemption */
759 	sdma_v5_2_ctx_switch_enable(adev, true);
760 
761 	/* start the gfx rings and rlc compute queues */
762 	r = sdma_v5_2_gfx_resume(adev);
763 	if (r)
764 		return r;
765 	r = sdma_v5_2_rlc_resume(adev);
766 
767 	return r;
768 }
769 
770 static int sdma_v5_2_mqd_init(struct amdgpu_device *adev, void *mqd,
771 			      struct amdgpu_mqd_prop *prop)
772 {
773 	struct v10_sdma_mqd *m = mqd;
774 	uint64_t wb_gpu_addr;
775 
776 	m->sdmax_rlcx_rb_cntl =
777 		order_base_2(prop->queue_size / 4) << SDMA0_RLC0_RB_CNTL__RB_SIZE__SHIFT |
778 		1 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_ENABLE__SHIFT |
779 		6 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_TIMER__SHIFT |
780 		1 << SDMA0_RLC0_RB_CNTL__RB_PRIV__SHIFT;
781 
782 	m->sdmax_rlcx_rb_base = lower_32_bits(prop->hqd_base_gpu_addr >> 8);
783 	m->sdmax_rlcx_rb_base_hi = upper_32_bits(prop->hqd_base_gpu_addr >> 8);
784 
785 	m->sdmax_rlcx_rb_wptr_poll_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, 0,
786 						  mmSDMA0_GFX_RB_WPTR_POLL_CNTL));
787 
788 	wb_gpu_addr = prop->wptr_gpu_addr;
789 	m->sdmax_rlcx_rb_wptr_poll_addr_lo = lower_32_bits(wb_gpu_addr);
790 	m->sdmax_rlcx_rb_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr);
791 
792 	wb_gpu_addr = prop->rptr_gpu_addr;
793 	m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits(wb_gpu_addr);
794 	m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits(wb_gpu_addr);
795 
796 	m->sdmax_rlcx_ib_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, 0,
797 							mmSDMA0_GFX_IB_CNTL));
798 
799 	m->sdmax_rlcx_doorbell_offset =
800 		prop->doorbell_index << SDMA0_RLC0_DOORBELL_OFFSET__OFFSET__SHIFT;
801 
802 	m->sdmax_rlcx_doorbell = REG_SET_FIELD(0, SDMA0_RLC0_DOORBELL, ENABLE, 1);
803 
804 	return 0;
805 }
806 
807 static void sdma_v5_2_set_mqd_funcs(struct amdgpu_device *adev)
808 {
809 	adev->mqds[AMDGPU_HW_IP_DMA].mqd_size = sizeof(struct v10_sdma_mqd);
810 	adev->mqds[AMDGPU_HW_IP_DMA].init_mqd = sdma_v5_2_mqd_init;
811 }
812 
813 /**
814  * sdma_v5_2_ring_test_ring - simple async dma engine test
815  *
816  * @ring: amdgpu_ring structure holding ring information
817  *
818  * Test the DMA engine by writing using it to write an
819  * value to memory.
820  * Returns 0 for success, error for failure.
821  */
822 static int sdma_v5_2_ring_test_ring(struct amdgpu_ring *ring)
823 {
824 	struct amdgpu_device *adev = ring->adev;
825 	unsigned i;
826 	unsigned index;
827 	int r;
828 	u32 tmp;
829 	u64 gpu_addr;
830 	volatile uint32_t *cpu_ptr = NULL;
831 
832 	tmp = 0xCAFEDEAD;
833 
834 	if (ring->is_mes_queue) {
835 		uint32_t offset = 0;
836 		offset = amdgpu_mes_ctx_get_offs(ring,
837 					 AMDGPU_MES_CTX_PADDING_OFFS);
838 		gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
839 		cpu_ptr = amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
840 		*cpu_ptr = tmp;
841 	} else {
842 		r = amdgpu_device_wb_get(adev, &index);
843 		if (r) {
844 			dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
845 			return r;
846 		}
847 
848 		gpu_addr = adev->wb.gpu_addr + (index * 4);
849 		adev->wb.wb[index] = cpu_to_le32(tmp);
850 	}
851 
852 	r = amdgpu_ring_alloc(ring, 20);
853 	if (r) {
854 		DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
855 		amdgpu_device_wb_free(adev, index);
856 		return r;
857 	}
858 
859 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
860 			  SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
861 	amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
862 	amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
863 	amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0));
864 	amdgpu_ring_write(ring, 0xDEADBEEF);
865 	amdgpu_ring_commit(ring);
866 
867 	for (i = 0; i < adev->usec_timeout; i++) {
868 		if (ring->is_mes_queue)
869 			tmp = le32_to_cpu(*cpu_ptr);
870 		else
871 			tmp = le32_to_cpu(adev->wb.wb[index]);
872 		if (tmp == 0xDEADBEEF)
873 			break;
874 		if (amdgpu_emu_mode == 1)
875 			msleep(1);
876 		else
877 			udelay(1);
878 	}
879 
880 	if (i >= adev->usec_timeout)
881 		r = -ETIMEDOUT;
882 
883 	if (!ring->is_mes_queue)
884 		amdgpu_device_wb_free(adev, index);
885 
886 	return r;
887 }
888 
889 /**
890  * sdma_v5_2_ring_test_ib - test an IB on the DMA engine
891  *
892  * @ring: amdgpu_ring structure holding ring information
893  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
894  *
895  * Test a simple IB in the DMA ring.
896  * Returns 0 on success, error on failure.
897  */
898 static int sdma_v5_2_ring_test_ib(struct amdgpu_ring *ring, long timeout)
899 {
900 	struct amdgpu_device *adev = ring->adev;
901 	struct amdgpu_ib ib;
902 	struct dma_fence *f = NULL;
903 	unsigned index;
904 	long r;
905 	u32 tmp = 0;
906 	u64 gpu_addr;
907 	volatile uint32_t *cpu_ptr = NULL;
908 
909 	tmp = 0xCAFEDEAD;
910 	memset(&ib, 0, sizeof(ib));
911 
912 	if (ring->is_mes_queue) {
913 		uint32_t offset = 0;
914 		offset = amdgpu_mes_ctx_get_offs(ring, AMDGPU_MES_CTX_IB_OFFS);
915 		ib.gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
916 		ib.ptr = (void *)amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
917 
918 		offset = amdgpu_mes_ctx_get_offs(ring,
919 					 AMDGPU_MES_CTX_PADDING_OFFS);
920 		gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
921 		cpu_ptr = amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
922 		*cpu_ptr = tmp;
923 	} else {
924 		r = amdgpu_device_wb_get(adev, &index);
925 		if (r) {
926 			dev_err(adev->dev, "(%ld) failed to allocate wb slot\n", r);
927 			return r;
928 		}
929 
930 		gpu_addr = adev->wb.gpu_addr + (index * 4);
931 		adev->wb.wb[index] = cpu_to_le32(tmp);
932 
933 		r = amdgpu_ib_get(adev, NULL, 256, AMDGPU_IB_POOL_DIRECT, &ib);
934 		if (r) {
935 			DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r);
936 			goto err0;
937 		}
938 	}
939 
940 	ib.ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
941 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
942 	ib.ptr[1] = lower_32_bits(gpu_addr);
943 	ib.ptr[2] = upper_32_bits(gpu_addr);
944 	ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0);
945 	ib.ptr[4] = 0xDEADBEEF;
946 	ib.ptr[5] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
947 	ib.ptr[6] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
948 	ib.ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
949 	ib.length_dw = 8;
950 
951 	r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
952 	if (r)
953 		goto err1;
954 
955 	r = dma_fence_wait_timeout(f, false, timeout);
956 	if (r == 0) {
957 		DRM_ERROR("amdgpu: IB test timed out\n");
958 		r = -ETIMEDOUT;
959 		goto err1;
960 	} else if (r < 0) {
961 		DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
962 		goto err1;
963 	}
964 
965 	if (ring->is_mes_queue)
966 		tmp = le32_to_cpu(*cpu_ptr);
967 	else
968 		tmp = le32_to_cpu(adev->wb.wb[index]);
969 
970 	if (tmp == 0xDEADBEEF)
971 		r = 0;
972 	else
973 		r = -EINVAL;
974 
975 err1:
976 	amdgpu_ib_free(adev, &ib, NULL);
977 	dma_fence_put(f);
978 err0:
979 	if (!ring->is_mes_queue)
980 		amdgpu_device_wb_free(adev, index);
981 	return r;
982 }
983 
984 
985 /**
986  * sdma_v5_2_vm_copy_pte - update PTEs by copying them from the GART
987  *
988  * @ib: indirect buffer to fill with commands
989  * @pe: addr of the page entry
990  * @src: src addr to copy from
991  * @count: number of page entries to update
992  *
993  * Update PTEs by copying them from the GART using sDMA.
994  */
995 static void sdma_v5_2_vm_copy_pte(struct amdgpu_ib *ib,
996 				  uint64_t pe, uint64_t src,
997 				  unsigned count)
998 {
999 	unsigned bytes = count * 8;
1000 
1001 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1002 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
1003 	ib->ptr[ib->length_dw++] = bytes - 1;
1004 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1005 	ib->ptr[ib->length_dw++] = lower_32_bits(src);
1006 	ib->ptr[ib->length_dw++] = upper_32_bits(src);
1007 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1008 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1009 
1010 }
1011 
1012 /**
1013  * sdma_v5_2_vm_write_pte - update PTEs by writing them manually
1014  *
1015  * @ib: indirect buffer to fill with commands
1016  * @pe: addr of the page entry
1017  * @value: dst addr to write into pe
1018  * @count: number of page entries to update
1019  * @incr: increase next addr by incr bytes
1020  *
1021  * Update PTEs by writing them manually using sDMA.
1022  */
1023 static void sdma_v5_2_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
1024 				   uint64_t value, unsigned count,
1025 				   uint32_t incr)
1026 {
1027 	unsigned ndw = count * 2;
1028 
1029 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
1030 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1031 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1032 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1033 	ib->ptr[ib->length_dw++] = ndw - 1;
1034 	for (; ndw > 0; ndw -= 2) {
1035 		ib->ptr[ib->length_dw++] = lower_32_bits(value);
1036 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
1037 		value += incr;
1038 	}
1039 }
1040 
1041 /**
1042  * sdma_v5_2_vm_set_pte_pde - update the page tables using sDMA
1043  *
1044  * @ib: indirect buffer to fill with commands
1045  * @pe: addr of the page entry
1046  * @addr: dst addr to write into pe
1047  * @count: number of page entries to update
1048  * @incr: increase next addr by incr bytes
1049  * @flags: access flags
1050  *
1051  * Update the page tables using sDMA.
1052  */
1053 static void sdma_v5_2_vm_set_pte_pde(struct amdgpu_ib *ib,
1054 				     uint64_t pe,
1055 				     uint64_t addr, unsigned count,
1056 				     uint32_t incr, uint64_t flags)
1057 {
1058 	/* for physically contiguous pages (vram) */
1059 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_PTEPDE);
1060 	ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
1061 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1062 	ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
1063 	ib->ptr[ib->length_dw++] = upper_32_bits(flags);
1064 	ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
1065 	ib->ptr[ib->length_dw++] = upper_32_bits(addr);
1066 	ib->ptr[ib->length_dw++] = incr; /* increment size */
1067 	ib->ptr[ib->length_dw++] = 0;
1068 	ib->ptr[ib->length_dw++] = count - 1; /* number of entries */
1069 }
1070 
1071 /**
1072  * sdma_v5_2_ring_pad_ib - pad the IB
1073  *
1074  * @ib: indirect buffer to fill with padding
1075  * @ring: amdgpu_ring structure holding ring information
1076  *
1077  * Pad the IB with NOPs to a boundary multiple of 8.
1078  */
1079 static void sdma_v5_2_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
1080 {
1081 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
1082 	u32 pad_count;
1083 	int i;
1084 
1085 	pad_count = (-ib->length_dw) & 0x7;
1086 	for (i = 0; i < pad_count; i++)
1087 		if (sdma && sdma->burst_nop && (i == 0))
1088 			ib->ptr[ib->length_dw++] =
1089 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP) |
1090 				SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
1091 		else
1092 			ib->ptr[ib->length_dw++] =
1093 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
1094 }
1095 
1096 
1097 /**
1098  * sdma_v5_2_ring_emit_pipeline_sync - sync the pipeline
1099  *
1100  * @ring: amdgpu_ring pointer
1101  *
1102  * Make sure all previous operations are completed (CIK).
1103  */
1104 static void sdma_v5_2_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
1105 {
1106 	uint32_t seq = ring->fence_drv.sync_seq;
1107 	uint64_t addr = ring->fence_drv.gpu_addr;
1108 
1109 	/* wait for idle */
1110 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1111 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1112 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
1113 			  SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1));
1114 	amdgpu_ring_write(ring, addr & 0xfffffffc);
1115 	amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
1116 	amdgpu_ring_write(ring, seq); /* reference */
1117 	amdgpu_ring_write(ring, 0xffffffff); /* mask */
1118 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1119 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */
1120 }
1121 
1122 
1123 /**
1124  * sdma_v5_2_ring_emit_vm_flush - vm flush using sDMA
1125  *
1126  * @ring: amdgpu_ring pointer
1127  * @vmid: vmid number to use
1128  * @pd_addr: address
1129  *
1130  * Update the page table base and flush the VM TLB
1131  * using sDMA.
1132  */
1133 static void sdma_v5_2_ring_emit_vm_flush(struct amdgpu_ring *ring,
1134 					 unsigned vmid, uint64_t pd_addr)
1135 {
1136 	amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
1137 }
1138 
1139 static void sdma_v5_2_ring_emit_wreg(struct amdgpu_ring *ring,
1140 				     uint32_t reg, uint32_t val)
1141 {
1142 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
1143 			  SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
1144 	amdgpu_ring_write(ring, reg);
1145 	amdgpu_ring_write(ring, val);
1146 }
1147 
1148 static void sdma_v5_2_ring_emit_reg_wait(struct amdgpu_ring *ring, uint32_t reg,
1149 					 uint32_t val, uint32_t mask)
1150 {
1151 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1152 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1153 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* equal */
1154 	amdgpu_ring_write(ring, reg << 2);
1155 	amdgpu_ring_write(ring, 0);
1156 	amdgpu_ring_write(ring, val); /* reference */
1157 	amdgpu_ring_write(ring, mask); /* mask */
1158 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1159 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10));
1160 }
1161 
1162 static void sdma_v5_2_ring_emit_reg_write_reg_wait(struct amdgpu_ring *ring,
1163 						   uint32_t reg0, uint32_t reg1,
1164 						   uint32_t ref, uint32_t mask)
1165 {
1166 	amdgpu_ring_emit_wreg(ring, reg0, ref);
1167 	/* wait for a cycle to reset vm_inv_eng*_ack */
1168 	amdgpu_ring_emit_reg_wait(ring, reg0, 0, 0);
1169 	amdgpu_ring_emit_reg_wait(ring, reg1, mask, mask);
1170 }
1171 
1172 static int sdma_v5_2_early_init(void *handle)
1173 {
1174 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1175 
1176 	sdma_v5_2_set_ring_funcs(adev);
1177 	sdma_v5_2_set_buffer_funcs(adev);
1178 	sdma_v5_2_set_vm_pte_funcs(adev);
1179 	sdma_v5_2_set_irq_funcs(adev);
1180 	sdma_v5_2_set_mqd_funcs(adev);
1181 
1182 	return 0;
1183 }
1184 
1185 static unsigned sdma_v5_2_seq_to_irq_id(int seq_num)
1186 {
1187 	switch (seq_num) {
1188 	case 0:
1189 		return SOC15_IH_CLIENTID_SDMA0;
1190 	case 1:
1191 		return SOC15_IH_CLIENTID_SDMA1;
1192 	case 2:
1193 		return SOC15_IH_CLIENTID_SDMA2;
1194 	case 3:
1195 		return SOC15_IH_CLIENTID_SDMA3_Sienna_Cichlid;
1196 	default:
1197 		break;
1198 	}
1199 	return -EINVAL;
1200 }
1201 
1202 static unsigned sdma_v5_2_seq_to_trap_id(int seq_num)
1203 {
1204 	switch (seq_num) {
1205 	case 0:
1206 		return SDMA0_5_0__SRCID__SDMA_TRAP;
1207 	case 1:
1208 		return SDMA1_5_0__SRCID__SDMA_TRAP;
1209 	case 2:
1210 		return SDMA2_5_0__SRCID__SDMA_TRAP;
1211 	case 3:
1212 		return SDMA3_5_0__SRCID__SDMA_TRAP;
1213 	default:
1214 		break;
1215 	}
1216 	return -EINVAL;
1217 }
1218 
1219 static int sdma_v5_2_sw_init(void *handle)
1220 {
1221 	struct amdgpu_ring *ring;
1222 	int r, i;
1223 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1224 
1225 	/* SDMA trap event */
1226 	for (i = 0; i < adev->sdma.num_instances; i++) {
1227 		r = amdgpu_irq_add_id(adev, sdma_v5_2_seq_to_irq_id(i),
1228 				      sdma_v5_2_seq_to_trap_id(i),
1229 				      &adev->sdma.trap_irq);
1230 		if (r)
1231 			return r;
1232 	}
1233 
1234 	r = amdgpu_sdma_init_microcode(adev, 0, true);
1235 	if (r) {
1236 		DRM_ERROR("Failed to load sdma firmware!\n");
1237 		return r;
1238 	}
1239 
1240 	for (i = 0; i < adev->sdma.num_instances; i++) {
1241 		ring = &adev->sdma.instance[i].ring;
1242 		ring->ring_obj = NULL;
1243 		ring->use_doorbell = true;
1244 		ring->me = i;
1245 
1246 		DRM_INFO("use_doorbell being set to: [%s]\n",
1247 				ring->use_doorbell?"true":"false");
1248 
1249 		ring->doorbell_index =
1250 			(adev->doorbell_index.sdma_engine[i] << 1); //get DWORD offset
1251 
1252 		ring->vm_hub = AMDGPU_GFXHUB(0);
1253 		sprintf(ring->name, "sdma%d", i);
1254 		r = amdgpu_ring_init(adev, ring, 1024, &adev->sdma.trap_irq,
1255 				     AMDGPU_SDMA_IRQ_INSTANCE0 + i,
1256 				     AMDGPU_RING_PRIO_DEFAULT, NULL);
1257 		if (r)
1258 			return r;
1259 	}
1260 
1261 	return r;
1262 }
1263 
1264 static int sdma_v5_2_sw_fini(void *handle)
1265 {
1266 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1267 	int i;
1268 
1269 	for (i = 0; i < adev->sdma.num_instances; i++)
1270 		amdgpu_ring_fini(&adev->sdma.instance[i].ring);
1271 
1272 	amdgpu_sdma_destroy_inst_ctx(adev, true);
1273 
1274 	return 0;
1275 }
1276 
1277 static int sdma_v5_2_hw_init(void *handle)
1278 {
1279 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1280 
1281 	return sdma_v5_2_start(adev);
1282 }
1283 
1284 static int sdma_v5_2_hw_fini(void *handle)
1285 {
1286 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1287 
1288 	if (amdgpu_sriov_vf(adev)) {
1289 		/* disable the scheduler for SDMA */
1290 		amdgpu_sdma_unset_buffer_funcs_helper(adev);
1291 		return 0;
1292 	}
1293 
1294 	sdma_v5_2_ctx_switch_enable(adev, false);
1295 	sdma_v5_2_enable(adev, false);
1296 
1297 	return 0;
1298 }
1299 
1300 static int sdma_v5_2_suspend(void *handle)
1301 {
1302 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1303 
1304 	return sdma_v5_2_hw_fini(adev);
1305 }
1306 
1307 static int sdma_v5_2_resume(void *handle)
1308 {
1309 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1310 
1311 	return sdma_v5_2_hw_init(adev);
1312 }
1313 
1314 static bool sdma_v5_2_is_idle(void *handle)
1315 {
1316 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1317 	u32 i;
1318 
1319 	for (i = 0; i < adev->sdma.num_instances; i++) {
1320 		u32 tmp = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_STATUS_REG));
1321 
1322 		if (!(tmp & SDMA0_STATUS_REG__IDLE_MASK))
1323 			return false;
1324 	}
1325 
1326 	return true;
1327 }
1328 
1329 static int sdma_v5_2_wait_for_idle(void *handle)
1330 {
1331 	unsigned i;
1332 	u32 sdma0, sdma1, sdma2, sdma3;
1333 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1334 
1335 	for (i = 0; i < adev->usec_timeout; i++) {
1336 		sdma0 = RREG32(sdma_v5_2_get_reg_offset(adev, 0, mmSDMA0_STATUS_REG));
1337 		sdma1 = RREG32(sdma_v5_2_get_reg_offset(adev, 1, mmSDMA0_STATUS_REG));
1338 		sdma2 = RREG32(sdma_v5_2_get_reg_offset(adev, 2, mmSDMA0_STATUS_REG));
1339 		sdma3 = RREG32(sdma_v5_2_get_reg_offset(adev, 3, mmSDMA0_STATUS_REG));
1340 
1341 		if (sdma0 & sdma1 & sdma2 & sdma3 & SDMA0_STATUS_REG__IDLE_MASK)
1342 			return 0;
1343 		udelay(1);
1344 	}
1345 	return -ETIMEDOUT;
1346 }
1347 
1348 static int sdma_v5_2_ring_preempt_ib(struct amdgpu_ring *ring)
1349 {
1350 	int i, r = 0;
1351 	struct amdgpu_device *adev = ring->adev;
1352 	u32 index = 0;
1353 	u64 sdma_gfx_preempt;
1354 
1355 	amdgpu_sdma_get_index_from_ring(ring, &index);
1356 	sdma_gfx_preempt =
1357 		sdma_v5_2_get_reg_offset(adev, index, mmSDMA0_GFX_PREEMPT);
1358 
1359 	/* assert preemption condition */
1360 	amdgpu_ring_set_preempt_cond_exec(ring, false);
1361 
1362 	/* emit the trailing fence */
1363 	ring->trail_seq += 1;
1364 	amdgpu_ring_alloc(ring, 10);
1365 	sdma_v5_2_ring_emit_fence(ring, ring->trail_fence_gpu_addr,
1366 				  ring->trail_seq, 0);
1367 	amdgpu_ring_commit(ring);
1368 
1369 	/* assert IB preemption */
1370 	WREG32(sdma_gfx_preempt, 1);
1371 
1372 	/* poll the trailing fence */
1373 	for (i = 0; i < adev->usec_timeout; i++) {
1374 		if (ring->trail_seq ==
1375 		    le32_to_cpu(*(ring->trail_fence_cpu_addr)))
1376 			break;
1377 		udelay(1);
1378 	}
1379 
1380 	if (i >= adev->usec_timeout) {
1381 		r = -EINVAL;
1382 		DRM_ERROR("ring %d failed to be preempted\n", ring->idx);
1383 	}
1384 
1385 	/* deassert IB preemption */
1386 	WREG32(sdma_gfx_preempt, 0);
1387 
1388 	/* deassert the preemption condition */
1389 	amdgpu_ring_set_preempt_cond_exec(ring, true);
1390 	return r;
1391 }
1392 
1393 static int sdma_v5_2_set_trap_irq_state(struct amdgpu_device *adev,
1394 					struct amdgpu_irq_src *source,
1395 					unsigned type,
1396 					enum amdgpu_interrupt_state state)
1397 {
1398 	u32 sdma_cntl;
1399 	u32 reg_offset = sdma_v5_2_get_reg_offset(adev, type, mmSDMA0_CNTL);
1400 
1401 	if (!amdgpu_sriov_vf(adev)) {
1402 		sdma_cntl = RREG32(reg_offset);
1403 		sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE,
1404 			       state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0);
1405 		WREG32(reg_offset, sdma_cntl);
1406 	}
1407 
1408 	return 0;
1409 }
1410 
1411 static int sdma_v5_2_process_trap_irq(struct amdgpu_device *adev,
1412 				      struct amdgpu_irq_src *source,
1413 				      struct amdgpu_iv_entry *entry)
1414 {
1415 	uint32_t mes_queue_id = entry->src_data[0];
1416 
1417 	DRM_DEBUG("IH: SDMA trap\n");
1418 
1419 	if (adev->enable_mes && (mes_queue_id & AMDGPU_FENCE_MES_QUEUE_FLAG)) {
1420 		struct amdgpu_mes_queue *queue;
1421 
1422 		mes_queue_id &= AMDGPU_FENCE_MES_QUEUE_ID_MASK;
1423 
1424 		spin_lock(&adev->mes.queue_id_lock);
1425 		queue = idr_find(&adev->mes.queue_id_idr, mes_queue_id);
1426 		if (queue) {
1427 			DRM_DEBUG("process smda queue id = %d\n", mes_queue_id);
1428 			amdgpu_fence_process(queue->ring);
1429 		}
1430 		spin_unlock(&adev->mes.queue_id_lock);
1431 		return 0;
1432 	}
1433 
1434 	switch (entry->client_id) {
1435 	case SOC15_IH_CLIENTID_SDMA0:
1436 		switch (entry->ring_id) {
1437 		case 0:
1438 			amdgpu_fence_process(&adev->sdma.instance[0].ring);
1439 			break;
1440 		case 1:
1441 			/* XXX compute */
1442 			break;
1443 		case 2:
1444 			/* XXX compute */
1445 			break;
1446 		case 3:
1447 			/* XXX page queue*/
1448 			break;
1449 		}
1450 		break;
1451 	case SOC15_IH_CLIENTID_SDMA1:
1452 		switch (entry->ring_id) {
1453 		case 0:
1454 			amdgpu_fence_process(&adev->sdma.instance[1].ring);
1455 			break;
1456 		case 1:
1457 			/* XXX compute */
1458 			break;
1459 		case 2:
1460 			/* XXX compute */
1461 			break;
1462 		case 3:
1463 			/* XXX page queue*/
1464 			break;
1465 		}
1466 		break;
1467 	case SOC15_IH_CLIENTID_SDMA2:
1468 		switch (entry->ring_id) {
1469 		case 0:
1470 			amdgpu_fence_process(&adev->sdma.instance[2].ring);
1471 			break;
1472 		case 1:
1473 			/* XXX compute */
1474 			break;
1475 		case 2:
1476 			/* XXX compute */
1477 			break;
1478 		case 3:
1479 			/* XXX page queue*/
1480 			break;
1481 		}
1482 		break;
1483 	case SOC15_IH_CLIENTID_SDMA3_Sienna_Cichlid:
1484 		switch (entry->ring_id) {
1485 		case 0:
1486 			amdgpu_fence_process(&adev->sdma.instance[3].ring);
1487 			break;
1488 		case 1:
1489 			/* XXX compute */
1490 			break;
1491 		case 2:
1492 			/* XXX compute */
1493 			break;
1494 		case 3:
1495 			/* XXX page queue*/
1496 			break;
1497 		}
1498 		break;
1499 	}
1500 	return 0;
1501 }
1502 
1503 static int sdma_v5_2_process_illegal_inst_irq(struct amdgpu_device *adev,
1504 					      struct amdgpu_irq_src *source,
1505 					      struct amdgpu_iv_entry *entry)
1506 {
1507 	return 0;
1508 }
1509 
1510 static bool sdma_v5_2_firmware_mgcg_support(struct amdgpu_device *adev,
1511 						     int i)
1512 {
1513 	switch (adev->ip_versions[SDMA0_HWIP][0]) {
1514 	case IP_VERSION(5, 2, 1):
1515 		if (adev->sdma.instance[i].fw_version < 70)
1516 			return false;
1517 		break;
1518 	case IP_VERSION(5, 2, 3):
1519 		if (adev->sdma.instance[i].fw_version < 47)
1520 			return false;
1521 		break;
1522 	case IP_VERSION(5, 2, 7):
1523 		if (adev->sdma.instance[i].fw_version < 9)
1524 			return false;
1525 		break;
1526 	default:
1527 		return true;
1528 	}
1529 
1530 	return true;
1531 
1532 }
1533 
1534 static void sdma_v5_2_update_medium_grain_clock_gating(struct amdgpu_device *adev,
1535 						       bool enable)
1536 {
1537 	uint32_t data, def;
1538 	int i;
1539 
1540 	for (i = 0; i < adev->sdma.num_instances; i++) {
1541 
1542 		if (!sdma_v5_2_firmware_mgcg_support(adev, i))
1543 			adev->cg_flags &= ~AMD_CG_SUPPORT_SDMA_MGCG;
1544 
1545 		if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
1546 			/* Enable sdma clock gating */
1547 			def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL));
1548 			data &= ~(SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1549 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1550 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1551 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1552 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK |
1553 				  SDMA0_CLK_CTRL__SOFT_OVERRIDER_REG_MASK);
1554 			if (def != data)
1555 				WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data);
1556 		} else {
1557 			/* Disable sdma clock gating */
1558 			def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL));
1559 			data |= (SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1560 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1561 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1562 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1563 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK |
1564 				 SDMA0_CLK_CTRL__SOFT_OVERRIDER_REG_MASK);
1565 			if (def != data)
1566 				WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data);
1567 		}
1568 	}
1569 }
1570 
1571 static void sdma_v5_2_update_medium_grain_light_sleep(struct amdgpu_device *adev,
1572 						      bool enable)
1573 {
1574 	uint32_t data, def;
1575 	int i;
1576 
1577 	for (i = 0; i < adev->sdma.num_instances; i++) {
1578 
1579 		if (adev->sdma.instance[i].fw_version < 70 && adev->ip_versions[SDMA0_HWIP][0] == IP_VERSION(5, 2, 1))
1580 			adev->cg_flags &= ~AMD_CG_SUPPORT_SDMA_LS;
1581 
1582 		if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) {
1583 			/* Enable sdma mem light sleep */
1584 			def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL));
1585 			data |= SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1586 			if (def != data)
1587 				WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data);
1588 
1589 		} else {
1590 			/* Disable sdma mem light sleep */
1591 			def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL));
1592 			data &= ~SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1593 			if (def != data)
1594 				WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data);
1595 
1596 		}
1597 	}
1598 }
1599 
1600 static int sdma_v5_2_set_clockgating_state(void *handle,
1601 					   enum amd_clockgating_state state)
1602 {
1603 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1604 
1605 	if (amdgpu_sriov_vf(adev))
1606 		return 0;
1607 
1608 	switch (adev->ip_versions[SDMA0_HWIP][0]) {
1609 	case IP_VERSION(5, 2, 0):
1610 	case IP_VERSION(5, 2, 2):
1611 	case IP_VERSION(5, 2, 1):
1612 	case IP_VERSION(5, 2, 4):
1613 	case IP_VERSION(5, 2, 5):
1614 	case IP_VERSION(5, 2, 6):
1615 	case IP_VERSION(5, 2, 3):
1616 	case IP_VERSION(5, 2, 7):
1617 		sdma_v5_2_update_medium_grain_clock_gating(adev,
1618 				state == AMD_CG_STATE_GATE);
1619 		sdma_v5_2_update_medium_grain_light_sleep(adev,
1620 				state == AMD_CG_STATE_GATE);
1621 		break;
1622 	default:
1623 		break;
1624 	}
1625 
1626 	return 0;
1627 }
1628 
1629 static int sdma_v5_2_set_powergating_state(void *handle,
1630 					  enum amd_powergating_state state)
1631 {
1632 	return 0;
1633 }
1634 
1635 static void sdma_v5_2_get_clockgating_state(void *handle, u64 *flags)
1636 {
1637 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1638 	int data;
1639 
1640 	if (amdgpu_sriov_vf(adev))
1641 		*flags = 0;
1642 
1643 	/* AMD_CG_SUPPORT_SDMA_MGCG */
1644 	data = RREG32(sdma_v5_2_get_reg_offset(adev, 0, mmSDMA0_CLK_CTRL));
1645 	if (!(data & SDMA0_CLK_CTRL__CGCG_EN_OVERRIDE_MASK))
1646 		*flags |= AMD_CG_SUPPORT_SDMA_MGCG;
1647 
1648 	/* AMD_CG_SUPPORT_SDMA_LS */
1649 	data = RREG32_KIQ(sdma_v5_2_get_reg_offset(adev, 0, mmSDMA0_POWER_CNTL));
1650 	if (data & SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK)
1651 		*flags |= AMD_CG_SUPPORT_SDMA_LS;
1652 }
1653 
1654 const struct amd_ip_funcs sdma_v5_2_ip_funcs = {
1655 	.name = "sdma_v5_2",
1656 	.early_init = sdma_v5_2_early_init,
1657 	.late_init = NULL,
1658 	.sw_init = sdma_v5_2_sw_init,
1659 	.sw_fini = sdma_v5_2_sw_fini,
1660 	.hw_init = sdma_v5_2_hw_init,
1661 	.hw_fini = sdma_v5_2_hw_fini,
1662 	.suspend = sdma_v5_2_suspend,
1663 	.resume = sdma_v5_2_resume,
1664 	.is_idle = sdma_v5_2_is_idle,
1665 	.wait_for_idle = sdma_v5_2_wait_for_idle,
1666 	.soft_reset = sdma_v5_2_soft_reset,
1667 	.set_clockgating_state = sdma_v5_2_set_clockgating_state,
1668 	.set_powergating_state = sdma_v5_2_set_powergating_state,
1669 	.get_clockgating_state = sdma_v5_2_get_clockgating_state,
1670 };
1671 
1672 static const struct amdgpu_ring_funcs sdma_v5_2_ring_funcs = {
1673 	.type = AMDGPU_RING_TYPE_SDMA,
1674 	.align_mask = 0xf,
1675 	.nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP),
1676 	.support_64bit_ptrs = true,
1677 	.secure_submission_supported = true,
1678 	.get_rptr = sdma_v5_2_ring_get_rptr,
1679 	.get_wptr = sdma_v5_2_ring_get_wptr,
1680 	.set_wptr = sdma_v5_2_ring_set_wptr,
1681 	.emit_frame_size =
1682 		5 + /* sdma_v5_2_ring_init_cond_exec */
1683 		6 + /* sdma_v5_2_ring_emit_hdp_flush */
1684 		3 + /* hdp_invalidate */
1685 		6 + /* sdma_v5_2_ring_emit_pipeline_sync */
1686 		/* sdma_v5_2_ring_emit_vm_flush */
1687 		SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 +
1688 		SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6 +
1689 		10 + 10 + 10, /* sdma_v5_2_ring_emit_fence x3 for user fence, vm fence */
1690 	.emit_ib_size = 7 + 6, /* sdma_v5_2_ring_emit_ib */
1691 	.emit_ib = sdma_v5_2_ring_emit_ib,
1692 	.emit_mem_sync = sdma_v5_2_ring_emit_mem_sync,
1693 	.emit_fence = sdma_v5_2_ring_emit_fence,
1694 	.emit_pipeline_sync = sdma_v5_2_ring_emit_pipeline_sync,
1695 	.emit_vm_flush = sdma_v5_2_ring_emit_vm_flush,
1696 	.emit_hdp_flush = sdma_v5_2_ring_emit_hdp_flush,
1697 	.test_ring = sdma_v5_2_ring_test_ring,
1698 	.test_ib = sdma_v5_2_ring_test_ib,
1699 	.insert_nop = sdma_v5_2_ring_insert_nop,
1700 	.pad_ib = sdma_v5_2_ring_pad_ib,
1701 	.emit_wreg = sdma_v5_2_ring_emit_wreg,
1702 	.emit_reg_wait = sdma_v5_2_ring_emit_reg_wait,
1703 	.emit_reg_write_reg_wait = sdma_v5_2_ring_emit_reg_write_reg_wait,
1704 	.init_cond_exec = sdma_v5_2_ring_init_cond_exec,
1705 	.patch_cond_exec = sdma_v5_2_ring_patch_cond_exec,
1706 	.preempt_ib = sdma_v5_2_ring_preempt_ib,
1707 };
1708 
1709 static void sdma_v5_2_set_ring_funcs(struct amdgpu_device *adev)
1710 {
1711 	int i;
1712 
1713 	for (i = 0; i < adev->sdma.num_instances; i++) {
1714 		adev->sdma.instance[i].ring.funcs = &sdma_v5_2_ring_funcs;
1715 		adev->sdma.instance[i].ring.me = i;
1716 	}
1717 }
1718 
1719 static const struct amdgpu_irq_src_funcs sdma_v5_2_trap_irq_funcs = {
1720 	.set = sdma_v5_2_set_trap_irq_state,
1721 	.process = sdma_v5_2_process_trap_irq,
1722 };
1723 
1724 static const struct amdgpu_irq_src_funcs sdma_v5_2_illegal_inst_irq_funcs = {
1725 	.process = sdma_v5_2_process_illegal_inst_irq,
1726 };
1727 
1728 static void sdma_v5_2_set_irq_funcs(struct amdgpu_device *adev)
1729 {
1730 	adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_INSTANCE0 +
1731 					adev->sdma.num_instances;
1732 	adev->sdma.trap_irq.funcs = &sdma_v5_2_trap_irq_funcs;
1733 	adev->sdma.illegal_inst_irq.funcs = &sdma_v5_2_illegal_inst_irq_funcs;
1734 }
1735 
1736 /**
1737  * sdma_v5_2_emit_copy_buffer - copy buffer using the sDMA engine
1738  *
1739  * @ib: indirect buffer to copy to
1740  * @src_offset: src GPU address
1741  * @dst_offset: dst GPU address
1742  * @byte_count: number of bytes to xfer
1743  * @tmz: if a secure copy should be used
1744  *
1745  * Copy GPU buffers using the DMA engine.
1746  * Used by the amdgpu ttm implementation to move pages if
1747  * registered as the asic copy callback.
1748  */
1749 static void sdma_v5_2_emit_copy_buffer(struct amdgpu_ib *ib,
1750 				       uint64_t src_offset,
1751 				       uint64_t dst_offset,
1752 				       uint32_t byte_count,
1753 				       bool tmz)
1754 {
1755 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1756 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR) |
1757 		SDMA_PKT_COPY_LINEAR_HEADER_TMZ(tmz ? 1 : 0);
1758 	ib->ptr[ib->length_dw++] = byte_count - 1;
1759 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1760 	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1761 	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1762 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1763 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1764 }
1765 
1766 /**
1767  * sdma_v5_2_emit_fill_buffer - fill buffer using the sDMA engine
1768  *
1769  * @ib: indirect buffer to fill
1770  * @src_data: value to write to buffer
1771  * @dst_offset: dst GPU address
1772  * @byte_count: number of bytes to xfer
1773  *
1774  * Fill GPU buffers using the DMA engine.
1775  */
1776 static void sdma_v5_2_emit_fill_buffer(struct amdgpu_ib *ib,
1777 				       uint32_t src_data,
1778 				       uint64_t dst_offset,
1779 				       uint32_t byte_count)
1780 {
1781 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL);
1782 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1783 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1784 	ib->ptr[ib->length_dw++] = src_data;
1785 	ib->ptr[ib->length_dw++] = byte_count - 1;
1786 }
1787 
1788 static const struct amdgpu_buffer_funcs sdma_v5_2_buffer_funcs = {
1789 	.copy_max_bytes = 0x400000,
1790 	.copy_num_dw = 7,
1791 	.emit_copy_buffer = sdma_v5_2_emit_copy_buffer,
1792 
1793 	.fill_max_bytes = 0x400000,
1794 	.fill_num_dw = 5,
1795 	.emit_fill_buffer = sdma_v5_2_emit_fill_buffer,
1796 };
1797 
1798 static void sdma_v5_2_set_buffer_funcs(struct amdgpu_device *adev)
1799 {
1800 	if (adev->mman.buffer_funcs == NULL) {
1801 		adev->mman.buffer_funcs = &sdma_v5_2_buffer_funcs;
1802 		adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1803 	}
1804 }
1805 
1806 static const struct amdgpu_vm_pte_funcs sdma_v5_2_vm_pte_funcs = {
1807 	.copy_pte_num_dw = 7,
1808 	.copy_pte = sdma_v5_2_vm_copy_pte,
1809 	.write_pte = sdma_v5_2_vm_write_pte,
1810 	.set_pte_pde = sdma_v5_2_vm_set_pte_pde,
1811 };
1812 
1813 static void sdma_v5_2_set_vm_pte_funcs(struct amdgpu_device *adev)
1814 {
1815 	unsigned i;
1816 
1817 	if (adev->vm_manager.vm_pte_funcs == NULL) {
1818 		adev->vm_manager.vm_pte_funcs = &sdma_v5_2_vm_pte_funcs;
1819 		for (i = 0; i < adev->sdma.num_instances; i++) {
1820 			adev->vm_manager.vm_pte_scheds[i] =
1821 				&adev->sdma.instance[i].ring.sched;
1822 		}
1823 		adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
1824 	}
1825 }
1826 
1827 const struct amdgpu_ip_block_version sdma_v5_2_ip_block = {
1828 	.type = AMD_IP_BLOCK_TYPE_SDMA,
1829 	.major = 5,
1830 	.minor = 2,
1831 	.rev = 0,
1832 	.funcs = &sdma_v5_2_ip_funcs,
1833 };
1834