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 		/* SDMA seems to miss doorbells sometimes when powergating kicks in.
192 		 * Updating the wptr directly will wake it. This is only safe because
193 		 * we disallow gfxoff in begin_use() and then allow it again in end_use().
194 		 */
195 		WREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR),
196 		       lower_32_bits(ring->wptr << 2));
197 		WREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI),
198 		       upper_32_bits(ring->wptr << 2));
199 	} else {
200 		DRM_DEBUG("Not using doorbell -- "
201 				"mmSDMA%i_GFX_RB_WPTR == 0x%08x "
202 				"mmSDMA%i_GFX_RB_WPTR_HI == 0x%08x\n",
203 				ring->me,
204 				lower_32_bits(ring->wptr << 2),
205 				ring->me,
206 				upper_32_bits(ring->wptr << 2));
207 		WREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR),
208 			lower_32_bits(ring->wptr << 2));
209 		WREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI),
210 			upper_32_bits(ring->wptr << 2));
211 	}
212 }
213 
214 static void sdma_v5_2_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
215 {
216 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
217 	int i;
218 
219 	for (i = 0; i < count; i++)
220 		if (sdma && sdma->burst_nop && (i == 0))
221 			amdgpu_ring_write(ring, ring->funcs->nop |
222 				SDMA_PKT_NOP_HEADER_COUNT(count - 1));
223 		else
224 			amdgpu_ring_write(ring, ring->funcs->nop);
225 }
226 
227 /**
228  * sdma_v5_2_ring_emit_ib - Schedule an IB on the DMA engine
229  *
230  * @ring: amdgpu ring pointer
231  * @job: job to retrieve vmid from
232  * @ib: IB object to schedule
233  * @flags: unused
234  *
235  * Schedule an IB in the DMA ring.
236  */
237 static void sdma_v5_2_ring_emit_ib(struct amdgpu_ring *ring,
238 				   struct amdgpu_job *job,
239 				   struct amdgpu_ib *ib,
240 				   uint32_t flags)
241 {
242 	unsigned vmid = AMDGPU_JOB_GET_VMID(job);
243 	uint64_t csa_mc_addr = amdgpu_sdma_get_csa_mc_addr(ring, vmid);
244 
245 	/* An IB packet must end on a 8 DW boundary--the next dword
246 	 * must be on a 8-dword boundary. Our IB packet below is 6
247 	 * dwords long, thus add x number of NOPs, such that, in
248 	 * modular arithmetic,
249 	 * wptr + 6 + x = 8k, k >= 0, which in C is,
250 	 * (wptr + 6 + x) % 8 = 0.
251 	 * The expression below, is a solution of x.
252 	 */
253 	sdma_v5_2_ring_insert_nop(ring, (2 - lower_32_bits(ring->wptr)) & 7);
254 
255 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) |
256 			  SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf));
257 	/* base must be 32 byte aligned */
258 	amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
259 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
260 	amdgpu_ring_write(ring, ib->length_dw);
261 	amdgpu_ring_write(ring, lower_32_bits(csa_mc_addr));
262 	amdgpu_ring_write(ring, upper_32_bits(csa_mc_addr));
263 }
264 
265 /**
266  * sdma_v5_2_ring_emit_mem_sync - flush the IB by graphics cache rinse
267  *
268  * @ring: amdgpu ring pointer
269  *
270  * flush the IB by graphics cache rinse.
271  */
272 static void sdma_v5_2_ring_emit_mem_sync(struct amdgpu_ring *ring)
273 {
274 	uint32_t gcr_cntl = SDMA_GCR_GL2_INV | SDMA_GCR_GL2_WB |
275 			    SDMA_GCR_GLM_INV | SDMA_GCR_GL1_INV |
276 			    SDMA_GCR_GLV_INV | SDMA_GCR_GLK_INV |
277 			    SDMA_GCR_GLI_INV(1);
278 
279 	/* flush entire cache L0/L1/L2, this can be optimized by performance requirement */
280 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_GCR_REQ));
281 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD1_BASE_VA_31_7(0));
282 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD2_GCR_CONTROL_15_0(gcr_cntl) |
283 			SDMA_PKT_GCR_REQ_PAYLOAD2_BASE_VA_47_32(0));
284 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD3_LIMIT_VA_31_7(0) |
285 			SDMA_PKT_GCR_REQ_PAYLOAD3_GCR_CONTROL_18_16(gcr_cntl >> 16));
286 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD4_LIMIT_VA_47_32(0) |
287 			SDMA_PKT_GCR_REQ_PAYLOAD4_VMID(0));
288 }
289 
290 /**
291  * sdma_v5_2_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
292  *
293  * @ring: amdgpu ring pointer
294  *
295  * Emit an hdp flush packet on the requested DMA ring.
296  */
297 static void sdma_v5_2_ring_emit_hdp_flush(struct amdgpu_ring *ring)
298 {
299 	struct amdgpu_device *adev = ring->adev;
300 	u32 ref_and_mask = 0;
301 	const struct nbio_hdp_flush_reg *nbio_hf_reg = adev->nbio.hdp_flush_reg;
302 
303 	if (ring->me > 1) {
304 		amdgpu_asic_flush_hdp(adev, ring);
305 	} else {
306 		ref_and_mask = nbio_hf_reg->ref_and_mask_sdma0 << ring->me;
307 
308 		amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
309 				  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) |
310 				  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
311 		amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_done_offset(adev)) << 2);
312 		amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_req_offset(adev)) << 2);
313 		amdgpu_ring_write(ring, ref_and_mask); /* reference */
314 		amdgpu_ring_write(ring, ref_and_mask); /* mask */
315 		amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
316 				  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
317 	}
318 }
319 
320 /**
321  * sdma_v5_2_ring_emit_fence - emit a fence on the DMA ring
322  *
323  * @ring: amdgpu ring pointer
324  * @addr: address
325  * @seq: sequence number
326  * @flags: fence related flags
327  *
328  * Add a DMA fence packet to the ring to write
329  * the fence seq number and DMA trap packet to generate
330  * an interrupt if needed.
331  */
332 static void sdma_v5_2_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
333 				      unsigned flags)
334 {
335 	bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
336 	/* write the fence */
337 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) |
338 			  SDMA_PKT_FENCE_HEADER_MTYPE(0x3)); /* Ucached(UC) */
339 	/* zero in first two bits */
340 	BUG_ON(addr & 0x3);
341 	amdgpu_ring_write(ring, lower_32_bits(addr));
342 	amdgpu_ring_write(ring, upper_32_bits(addr));
343 	amdgpu_ring_write(ring, lower_32_bits(seq));
344 
345 	/* optionally write high bits as well */
346 	if (write64bit) {
347 		addr += 4;
348 		amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) |
349 				  SDMA_PKT_FENCE_HEADER_MTYPE(0x3));
350 		/* zero in first two bits */
351 		BUG_ON(addr & 0x3);
352 		amdgpu_ring_write(ring, lower_32_bits(addr));
353 		amdgpu_ring_write(ring, upper_32_bits(addr));
354 		amdgpu_ring_write(ring, upper_32_bits(seq));
355 	}
356 
357 	if ((flags & AMDGPU_FENCE_FLAG_INT)) {
358 		uint32_t ctx = ring->is_mes_queue ?
359 			(ring->hw_queue_id | AMDGPU_FENCE_MES_QUEUE_FLAG) : 0;
360 		/* generate an interrupt */
361 		amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP));
362 		amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(ctx));
363 	}
364 }
365 
366 
367 /**
368  * sdma_v5_2_gfx_stop - stop the gfx async dma engines
369  *
370  * @adev: amdgpu_device pointer
371  *
372  * Stop the gfx async dma ring buffers.
373  */
374 static void sdma_v5_2_gfx_stop(struct amdgpu_device *adev)
375 {
376 	u32 rb_cntl, ib_cntl;
377 	int i;
378 
379 	amdgpu_sdma_unset_buffer_funcs_helper(adev);
380 
381 	for (i = 0; i < adev->sdma.num_instances; i++) {
382 		rb_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL));
383 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 0);
384 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
385 		ib_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL));
386 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 0);
387 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl);
388 	}
389 }
390 
391 /**
392  * sdma_v5_2_rlc_stop - stop the compute async dma engines
393  *
394  * @adev: amdgpu_device pointer
395  *
396  * Stop the compute async dma queues.
397  */
398 static void sdma_v5_2_rlc_stop(struct amdgpu_device *adev)
399 {
400 	/* XXX todo */
401 }
402 
403 /**
404  * sdma_v5_2_ctx_switch_enable - stop the async dma engines context switch
405  *
406  * @adev: amdgpu_device pointer
407  * @enable: enable/disable the DMA MEs context switch.
408  *
409  * Halt or unhalt the async dma engines context switch.
410  */
411 static void sdma_v5_2_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
412 {
413 	u32 f32_cntl, phase_quantum = 0;
414 	int i;
415 
416 	if (amdgpu_sdma_phase_quantum) {
417 		unsigned value = amdgpu_sdma_phase_quantum;
418 		unsigned unit = 0;
419 
420 		while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
421 				SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) {
422 			value = (value + 1) >> 1;
423 			unit++;
424 		}
425 		if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
426 			    SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) {
427 			value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
428 				 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT);
429 			unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
430 				SDMA0_PHASE0_QUANTUM__UNIT__SHIFT);
431 			WARN_ONCE(1,
432 			"clamping sdma_phase_quantum to %uK clock cycles\n",
433 				  value << unit);
434 		}
435 		phase_quantum =
436 			value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
437 			unit  << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
438 	}
439 
440 	for (i = 0; i < adev->sdma.num_instances; i++) {
441 		if (enable && amdgpu_sdma_phase_quantum) {
442 			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE0_QUANTUM),
443 			       phase_quantum);
444 			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE1_QUANTUM),
445 			       phase_quantum);
446 			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE2_QUANTUM),
447 			       phase_quantum);
448 		}
449 
450 		if (!amdgpu_sriov_vf(adev)) {
451 			f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
452 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
453 					AUTO_CTXSW_ENABLE, enable ? 1 : 0);
454 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl);
455 		}
456 	}
457 
458 }
459 
460 /**
461  * sdma_v5_2_enable - stop the async dma engines
462  *
463  * @adev: amdgpu_device pointer
464  * @enable: enable/disable the DMA MEs.
465  *
466  * Halt or unhalt the async dma engines.
467  */
468 static void sdma_v5_2_enable(struct amdgpu_device *adev, bool enable)
469 {
470 	u32 f32_cntl;
471 	int i;
472 
473 	if (!enable) {
474 		sdma_v5_2_gfx_stop(adev);
475 		sdma_v5_2_rlc_stop(adev);
476 	}
477 
478 	if (!amdgpu_sriov_vf(adev)) {
479 		for (i = 0; i < adev->sdma.num_instances; i++) {
480 			f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
481 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, enable ? 0 : 1);
482 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl);
483 		}
484 	}
485 }
486 
487 /**
488  * sdma_v5_2_gfx_resume - setup and start the async dma engines
489  *
490  * @adev: amdgpu_device pointer
491  *
492  * Set up the gfx DMA ring buffers and enable them.
493  * Returns 0 for success, error for failure.
494  */
495 static int sdma_v5_2_gfx_resume(struct amdgpu_device *adev)
496 {
497 	struct amdgpu_ring *ring;
498 	u32 rb_cntl, ib_cntl;
499 	u32 rb_bufsz;
500 	u32 doorbell;
501 	u32 doorbell_offset;
502 	u32 temp;
503 	u32 wptr_poll_cntl;
504 	u64 wptr_gpu_addr;
505 	int i, r;
506 
507 	for (i = 0; i < adev->sdma.num_instances; i++) {
508 		ring = &adev->sdma.instance[i].ring;
509 
510 		if (!amdgpu_sriov_vf(adev))
511 			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL), 0);
512 
513 		/* Set ring buffer size in dwords */
514 		rb_bufsz = order_base_2(ring->ring_size / 4);
515 		rb_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL));
516 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz);
517 #ifdef __BIG_ENDIAN
518 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1);
519 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL,
520 					RPTR_WRITEBACK_SWAP_ENABLE, 1);
521 #endif
522 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
523 
524 		/* Initialize the ring buffer's read and write pointers */
525 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR), 0);
526 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_HI), 0);
527 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), 0);
528 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), 0);
529 
530 		/* setup the wptr shadow polling */
531 		wptr_gpu_addr = ring->wptr_gpu_addr;
532 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_LO),
533 		       lower_32_bits(wptr_gpu_addr));
534 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_HI),
535 		       upper_32_bits(wptr_gpu_addr));
536 		wptr_poll_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i,
537 							 mmSDMA0_GFX_RB_WPTR_POLL_CNTL));
538 		wptr_poll_cntl = REG_SET_FIELD(wptr_poll_cntl,
539 					       SDMA0_GFX_RB_WPTR_POLL_CNTL,
540 					       F32_POLL_ENABLE, 1);
541 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_CNTL),
542 		       wptr_poll_cntl);
543 
544 		/* set the wb address whether it's enabled or not */
545 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_HI),
546 		       upper_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFF);
547 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_LO),
548 		       lower_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFC);
549 
550 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
551 
552 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE), ring->gpu_addr >> 8);
553 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE_HI), ring->gpu_addr >> 40);
554 
555 		ring->wptr = 0;
556 
557 		/* before programing wptr to a less value, need set minor_ptr_update first */
558 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 1);
559 
560 		if (!amdgpu_sriov_vf(adev)) { /* only bare-metal use register write for wptr */
561 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), lower_32_bits(ring->wptr << 2));
562 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), upper_32_bits(ring->wptr << 2));
563 		}
564 
565 		doorbell = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL));
566 		doorbell_offset = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL_OFFSET));
567 
568 		if (ring->use_doorbell) {
569 			doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 1);
570 			doorbell_offset = REG_SET_FIELD(doorbell_offset, SDMA0_GFX_DOORBELL_OFFSET,
571 					OFFSET, ring->doorbell_index);
572 		} else {
573 			doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 0);
574 		}
575 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL), doorbell);
576 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL_OFFSET), doorbell_offset);
577 
578 		adev->nbio.funcs->sdma_doorbell_range(adev, i, ring->use_doorbell,
579 						      ring->doorbell_index,
580 						      adev->doorbell_index.sdma_doorbell_range);
581 
582 		if (amdgpu_sriov_vf(adev))
583 			sdma_v5_2_ring_set_wptr(ring);
584 
585 		/* set minor_ptr_update to 0 after wptr programed */
586 
587 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 0);
588 
589 		/* SRIOV VF has no control of any of registers below */
590 		if (!amdgpu_sriov_vf(adev)) {
591 			/* set utc l1 enable flag always to 1 */
592 			temp = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
593 			temp = REG_SET_FIELD(temp, SDMA0_CNTL, UTC_L1_ENABLE, 1);
594 
595 			/* enable MCBP */
596 			temp = REG_SET_FIELD(temp, SDMA0_CNTL, MIDCMD_PREEMPT_ENABLE, 1);
597 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), temp);
598 
599 			/* Set up RESP_MODE to non-copy addresses */
600 			temp = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL));
601 			temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, RESP_MODE, 3);
602 			temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, REDO_DELAY, 9);
603 			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL), temp);
604 
605 			/* program default cache read and write policy */
606 			temp = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE));
607 			/* clean read policy and write policy bits */
608 			temp &= 0xFF0FFF;
609 			temp |= ((CACHE_READ_POLICY_L2__DEFAULT << 12) |
610 				 (CACHE_WRITE_POLICY_L2__DEFAULT << 14) |
611 				 SDMA0_UTCL1_PAGE__LLC_NOALLOC_MASK);
612 			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE), temp);
613 
614 			/* unhalt engine */
615 			temp = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
616 			temp = REG_SET_FIELD(temp, SDMA0_F32_CNTL, HALT, 0);
617 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), temp);
618 		}
619 
620 		/* enable DMA RB */
621 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1);
622 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
623 
624 		ib_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL));
625 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 1);
626 #ifdef __BIG_ENDIAN
627 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1);
628 #endif
629 		/* enable DMA IBs */
630 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl);
631 
632 		if (amdgpu_sriov_vf(adev)) { /* bare-metal sequence doesn't need below to lines */
633 			sdma_v5_2_ctx_switch_enable(adev, true);
634 			sdma_v5_2_enable(adev, true);
635 		}
636 
637 		r = amdgpu_ring_test_helper(ring);
638 		if (r)
639 			return r;
640 
641 		if (adev->mman.buffer_funcs_ring == ring)
642 			amdgpu_ttm_set_buffer_funcs_status(adev, true);
643 	}
644 
645 	return 0;
646 }
647 
648 /**
649  * sdma_v5_2_rlc_resume - setup and start the async dma engines
650  *
651  * @adev: amdgpu_device pointer
652  *
653  * Set up the compute DMA queues and enable them.
654  * Returns 0 for success, error for failure.
655  */
656 static int sdma_v5_2_rlc_resume(struct amdgpu_device *adev)
657 {
658 	return 0;
659 }
660 
661 /**
662  * sdma_v5_2_load_microcode - load the sDMA ME ucode
663  *
664  * @adev: amdgpu_device pointer
665  *
666  * Loads the sDMA0/1/2/3 ucode.
667  * Returns 0 for success, -EINVAL if the ucode is not available.
668  */
669 static int sdma_v5_2_load_microcode(struct amdgpu_device *adev)
670 {
671 	const struct sdma_firmware_header_v1_0 *hdr;
672 	const __le32 *fw_data;
673 	u32 fw_size;
674 	int i, j;
675 
676 	/* halt the MEs */
677 	sdma_v5_2_enable(adev, false);
678 
679 	for (i = 0; i < adev->sdma.num_instances; i++) {
680 		if (!adev->sdma.instance[i].fw)
681 			return -EINVAL;
682 
683 		hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
684 		amdgpu_ucode_print_sdma_hdr(&hdr->header);
685 		fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
686 
687 		fw_data = (const __le32 *)
688 			(adev->sdma.instance[i].fw->data +
689 				le32_to_cpu(hdr->header.ucode_array_offset_bytes));
690 
691 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), 0);
692 
693 		for (j = 0; j < fw_size; j++) {
694 			if (amdgpu_emu_mode == 1 && j % 500 == 0)
695 				msleep(1);
696 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UCODE_DATA), le32_to_cpup(fw_data++));
697 		}
698 
699 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), adev->sdma.instance[i].fw_version);
700 	}
701 
702 	return 0;
703 }
704 
705 static int sdma_v5_2_soft_reset(void *handle)
706 {
707 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
708 	u32 grbm_soft_reset;
709 	u32 tmp;
710 	int i;
711 
712 	for (i = 0; i < adev->sdma.num_instances; i++) {
713 		grbm_soft_reset = REG_SET_FIELD(0,
714 						GRBM_SOFT_RESET, SOFT_RESET_SDMA0,
715 						1);
716 		grbm_soft_reset <<= i;
717 
718 		tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET);
719 		tmp |= grbm_soft_reset;
720 		DRM_DEBUG("GRBM_SOFT_RESET=0x%08X\n", tmp);
721 		WREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET, tmp);
722 		tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET);
723 
724 		udelay(50);
725 
726 		tmp &= ~grbm_soft_reset;
727 		WREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET, tmp);
728 		tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET);
729 
730 		udelay(50);
731 	}
732 
733 	return 0;
734 }
735 
736 /**
737  * sdma_v5_2_start - setup and start the async dma engines
738  *
739  * @adev: amdgpu_device pointer
740  *
741  * Set up the DMA engines and enable them.
742  * Returns 0 for success, error for failure.
743  */
744 static int sdma_v5_2_start(struct amdgpu_device *adev)
745 {
746 	int r = 0;
747 
748 	if (amdgpu_sriov_vf(adev)) {
749 		sdma_v5_2_ctx_switch_enable(adev, false);
750 		sdma_v5_2_enable(adev, false);
751 
752 		/* set RB registers */
753 		r = sdma_v5_2_gfx_resume(adev);
754 		return r;
755 	}
756 
757 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
758 		r = sdma_v5_2_load_microcode(adev);
759 		if (r)
760 			return r;
761 
762 		/* The value of mmSDMA_F32_CNTL is invalid the moment after loading fw */
763 		if (amdgpu_emu_mode == 1)
764 			msleep(1000);
765 	}
766 
767 	sdma_v5_2_soft_reset(adev);
768 	/* unhalt the MEs */
769 	sdma_v5_2_enable(adev, true);
770 	/* enable sdma ring preemption */
771 	sdma_v5_2_ctx_switch_enable(adev, true);
772 
773 	/* start the gfx rings and rlc compute queues */
774 	r = sdma_v5_2_gfx_resume(adev);
775 	if (r)
776 		return r;
777 	r = sdma_v5_2_rlc_resume(adev);
778 
779 	return r;
780 }
781 
782 static int sdma_v5_2_mqd_init(struct amdgpu_device *adev, void *mqd,
783 			      struct amdgpu_mqd_prop *prop)
784 {
785 	struct v10_sdma_mqd *m = mqd;
786 	uint64_t wb_gpu_addr;
787 
788 	m->sdmax_rlcx_rb_cntl =
789 		order_base_2(prop->queue_size / 4) << SDMA0_RLC0_RB_CNTL__RB_SIZE__SHIFT |
790 		1 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_ENABLE__SHIFT |
791 		6 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_TIMER__SHIFT |
792 		1 << SDMA0_RLC0_RB_CNTL__RB_PRIV__SHIFT;
793 
794 	m->sdmax_rlcx_rb_base = lower_32_bits(prop->hqd_base_gpu_addr >> 8);
795 	m->sdmax_rlcx_rb_base_hi = upper_32_bits(prop->hqd_base_gpu_addr >> 8);
796 
797 	m->sdmax_rlcx_rb_wptr_poll_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, 0,
798 						  mmSDMA0_GFX_RB_WPTR_POLL_CNTL));
799 
800 	wb_gpu_addr = prop->wptr_gpu_addr;
801 	m->sdmax_rlcx_rb_wptr_poll_addr_lo = lower_32_bits(wb_gpu_addr);
802 	m->sdmax_rlcx_rb_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr);
803 
804 	wb_gpu_addr = prop->rptr_gpu_addr;
805 	m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits(wb_gpu_addr);
806 	m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits(wb_gpu_addr);
807 
808 	m->sdmax_rlcx_ib_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, 0,
809 							mmSDMA0_GFX_IB_CNTL));
810 
811 	m->sdmax_rlcx_doorbell_offset =
812 		prop->doorbell_index << SDMA0_RLC0_DOORBELL_OFFSET__OFFSET__SHIFT;
813 
814 	m->sdmax_rlcx_doorbell = REG_SET_FIELD(0, SDMA0_RLC0_DOORBELL, ENABLE, 1);
815 
816 	return 0;
817 }
818 
819 static void sdma_v5_2_set_mqd_funcs(struct amdgpu_device *adev)
820 {
821 	adev->mqds[AMDGPU_HW_IP_DMA].mqd_size = sizeof(struct v10_sdma_mqd);
822 	adev->mqds[AMDGPU_HW_IP_DMA].init_mqd = sdma_v5_2_mqd_init;
823 }
824 
825 /**
826  * sdma_v5_2_ring_test_ring - simple async dma engine test
827  *
828  * @ring: amdgpu_ring structure holding ring information
829  *
830  * Test the DMA engine by writing using it to write an
831  * value to memory.
832  * Returns 0 for success, error for failure.
833  */
834 static int sdma_v5_2_ring_test_ring(struct amdgpu_ring *ring)
835 {
836 	struct amdgpu_device *adev = ring->adev;
837 	unsigned i;
838 	unsigned index;
839 	int r;
840 	u32 tmp;
841 	u64 gpu_addr;
842 	volatile uint32_t *cpu_ptr = NULL;
843 
844 	tmp = 0xCAFEDEAD;
845 
846 	if (ring->is_mes_queue) {
847 		uint32_t offset = 0;
848 		offset = amdgpu_mes_ctx_get_offs(ring,
849 					 AMDGPU_MES_CTX_PADDING_OFFS);
850 		gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
851 		cpu_ptr = amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
852 		*cpu_ptr = tmp;
853 	} else {
854 		r = amdgpu_device_wb_get(adev, &index);
855 		if (r) {
856 			dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
857 			return r;
858 		}
859 
860 		gpu_addr = adev->wb.gpu_addr + (index * 4);
861 		adev->wb.wb[index] = cpu_to_le32(tmp);
862 	}
863 
864 	r = amdgpu_ring_alloc(ring, 20);
865 	if (r) {
866 		DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
867 		amdgpu_device_wb_free(adev, index);
868 		return r;
869 	}
870 
871 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
872 			  SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
873 	amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
874 	amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
875 	amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0));
876 	amdgpu_ring_write(ring, 0xDEADBEEF);
877 	amdgpu_ring_commit(ring);
878 
879 	for (i = 0; i < adev->usec_timeout; i++) {
880 		if (ring->is_mes_queue)
881 			tmp = le32_to_cpu(*cpu_ptr);
882 		else
883 			tmp = le32_to_cpu(adev->wb.wb[index]);
884 		if (tmp == 0xDEADBEEF)
885 			break;
886 		if (amdgpu_emu_mode == 1)
887 			msleep(1);
888 		else
889 			udelay(1);
890 	}
891 
892 	if (i >= adev->usec_timeout)
893 		r = -ETIMEDOUT;
894 
895 	if (!ring->is_mes_queue)
896 		amdgpu_device_wb_free(adev, index);
897 
898 	return r;
899 }
900 
901 /**
902  * sdma_v5_2_ring_test_ib - test an IB on the DMA engine
903  *
904  * @ring: amdgpu_ring structure holding ring information
905  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
906  *
907  * Test a simple IB in the DMA ring.
908  * Returns 0 on success, error on failure.
909  */
910 static int sdma_v5_2_ring_test_ib(struct amdgpu_ring *ring, long timeout)
911 {
912 	struct amdgpu_device *adev = ring->adev;
913 	struct amdgpu_ib ib;
914 	struct dma_fence *f = NULL;
915 	unsigned index;
916 	long r;
917 	u32 tmp = 0;
918 	u64 gpu_addr;
919 	volatile uint32_t *cpu_ptr = NULL;
920 
921 	tmp = 0xCAFEDEAD;
922 	memset(&ib, 0, sizeof(ib));
923 
924 	if (ring->is_mes_queue) {
925 		uint32_t offset = 0;
926 		offset = amdgpu_mes_ctx_get_offs(ring, AMDGPU_MES_CTX_IB_OFFS);
927 		ib.gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
928 		ib.ptr = (void *)amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
929 
930 		offset = amdgpu_mes_ctx_get_offs(ring,
931 					 AMDGPU_MES_CTX_PADDING_OFFS);
932 		gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
933 		cpu_ptr = amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
934 		*cpu_ptr = tmp;
935 	} else {
936 		r = amdgpu_device_wb_get(adev, &index);
937 		if (r) {
938 			dev_err(adev->dev, "(%ld) failed to allocate wb slot\n", r);
939 			return r;
940 		}
941 
942 		gpu_addr = adev->wb.gpu_addr + (index * 4);
943 		adev->wb.wb[index] = cpu_to_le32(tmp);
944 
945 		r = amdgpu_ib_get(adev, NULL, 256, AMDGPU_IB_POOL_DIRECT, &ib);
946 		if (r) {
947 			DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r);
948 			goto err0;
949 		}
950 	}
951 
952 	ib.ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
953 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
954 	ib.ptr[1] = lower_32_bits(gpu_addr);
955 	ib.ptr[2] = upper_32_bits(gpu_addr);
956 	ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0);
957 	ib.ptr[4] = 0xDEADBEEF;
958 	ib.ptr[5] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
959 	ib.ptr[6] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
960 	ib.ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
961 	ib.length_dw = 8;
962 
963 	r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
964 	if (r)
965 		goto err1;
966 
967 	r = dma_fence_wait_timeout(f, false, timeout);
968 	if (r == 0) {
969 		DRM_ERROR("amdgpu: IB test timed out\n");
970 		r = -ETIMEDOUT;
971 		goto err1;
972 	} else if (r < 0) {
973 		DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
974 		goto err1;
975 	}
976 
977 	if (ring->is_mes_queue)
978 		tmp = le32_to_cpu(*cpu_ptr);
979 	else
980 		tmp = le32_to_cpu(adev->wb.wb[index]);
981 
982 	if (tmp == 0xDEADBEEF)
983 		r = 0;
984 	else
985 		r = -EINVAL;
986 
987 err1:
988 	amdgpu_ib_free(adev, &ib, NULL);
989 	dma_fence_put(f);
990 err0:
991 	if (!ring->is_mes_queue)
992 		amdgpu_device_wb_free(adev, index);
993 	return r;
994 }
995 
996 
997 /**
998  * sdma_v5_2_vm_copy_pte - update PTEs by copying them from the GART
999  *
1000  * @ib: indirect buffer to fill with commands
1001  * @pe: addr of the page entry
1002  * @src: src addr to copy from
1003  * @count: number of page entries to update
1004  *
1005  * Update PTEs by copying them from the GART using sDMA.
1006  */
1007 static void sdma_v5_2_vm_copy_pte(struct amdgpu_ib *ib,
1008 				  uint64_t pe, uint64_t src,
1009 				  unsigned count)
1010 {
1011 	unsigned bytes = count * 8;
1012 
1013 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1014 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
1015 	ib->ptr[ib->length_dw++] = bytes - 1;
1016 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1017 	ib->ptr[ib->length_dw++] = lower_32_bits(src);
1018 	ib->ptr[ib->length_dw++] = upper_32_bits(src);
1019 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1020 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1021 
1022 }
1023 
1024 /**
1025  * sdma_v5_2_vm_write_pte - update PTEs by writing them manually
1026  *
1027  * @ib: indirect buffer to fill with commands
1028  * @pe: addr of the page entry
1029  * @value: dst addr to write into pe
1030  * @count: number of page entries to update
1031  * @incr: increase next addr by incr bytes
1032  *
1033  * Update PTEs by writing them manually using sDMA.
1034  */
1035 static void sdma_v5_2_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
1036 				   uint64_t value, unsigned count,
1037 				   uint32_t incr)
1038 {
1039 	unsigned ndw = count * 2;
1040 
1041 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
1042 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1043 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1044 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1045 	ib->ptr[ib->length_dw++] = ndw - 1;
1046 	for (; ndw > 0; ndw -= 2) {
1047 		ib->ptr[ib->length_dw++] = lower_32_bits(value);
1048 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
1049 		value += incr;
1050 	}
1051 }
1052 
1053 /**
1054  * sdma_v5_2_vm_set_pte_pde - update the page tables using sDMA
1055  *
1056  * @ib: indirect buffer to fill with commands
1057  * @pe: addr of the page entry
1058  * @addr: dst addr to write into pe
1059  * @count: number of page entries to update
1060  * @incr: increase next addr by incr bytes
1061  * @flags: access flags
1062  *
1063  * Update the page tables using sDMA.
1064  */
1065 static void sdma_v5_2_vm_set_pte_pde(struct amdgpu_ib *ib,
1066 				     uint64_t pe,
1067 				     uint64_t addr, unsigned count,
1068 				     uint32_t incr, uint64_t flags)
1069 {
1070 	/* for physically contiguous pages (vram) */
1071 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_PTEPDE);
1072 	ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
1073 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1074 	ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
1075 	ib->ptr[ib->length_dw++] = upper_32_bits(flags);
1076 	ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
1077 	ib->ptr[ib->length_dw++] = upper_32_bits(addr);
1078 	ib->ptr[ib->length_dw++] = incr; /* increment size */
1079 	ib->ptr[ib->length_dw++] = 0;
1080 	ib->ptr[ib->length_dw++] = count - 1; /* number of entries */
1081 }
1082 
1083 /**
1084  * sdma_v5_2_ring_pad_ib - pad the IB
1085  *
1086  * @ib: indirect buffer to fill with padding
1087  * @ring: amdgpu_ring structure holding ring information
1088  *
1089  * Pad the IB with NOPs to a boundary multiple of 8.
1090  */
1091 static void sdma_v5_2_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
1092 {
1093 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
1094 	u32 pad_count;
1095 	int i;
1096 
1097 	pad_count = (-ib->length_dw) & 0x7;
1098 	for (i = 0; i < pad_count; i++)
1099 		if (sdma && sdma->burst_nop && (i == 0))
1100 			ib->ptr[ib->length_dw++] =
1101 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP) |
1102 				SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
1103 		else
1104 			ib->ptr[ib->length_dw++] =
1105 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
1106 }
1107 
1108 
1109 /**
1110  * sdma_v5_2_ring_emit_pipeline_sync - sync the pipeline
1111  *
1112  * @ring: amdgpu_ring pointer
1113  *
1114  * Make sure all previous operations are completed (CIK).
1115  */
1116 static void sdma_v5_2_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
1117 {
1118 	uint32_t seq = ring->fence_drv.sync_seq;
1119 	uint64_t addr = ring->fence_drv.gpu_addr;
1120 
1121 	/* wait for idle */
1122 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1123 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1124 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
1125 			  SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1));
1126 	amdgpu_ring_write(ring, addr & 0xfffffffc);
1127 	amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
1128 	amdgpu_ring_write(ring, seq); /* reference */
1129 	amdgpu_ring_write(ring, 0xffffffff); /* mask */
1130 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1131 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */
1132 }
1133 
1134 
1135 /**
1136  * sdma_v5_2_ring_emit_vm_flush - vm flush using sDMA
1137  *
1138  * @ring: amdgpu_ring pointer
1139  * @vmid: vmid number to use
1140  * @pd_addr: address
1141  *
1142  * Update the page table base and flush the VM TLB
1143  * using sDMA.
1144  */
1145 static void sdma_v5_2_ring_emit_vm_flush(struct amdgpu_ring *ring,
1146 					 unsigned vmid, uint64_t pd_addr)
1147 {
1148 	amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
1149 }
1150 
1151 static void sdma_v5_2_ring_emit_wreg(struct amdgpu_ring *ring,
1152 				     uint32_t reg, uint32_t val)
1153 {
1154 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
1155 			  SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
1156 	amdgpu_ring_write(ring, reg);
1157 	amdgpu_ring_write(ring, val);
1158 }
1159 
1160 static void sdma_v5_2_ring_emit_reg_wait(struct amdgpu_ring *ring, uint32_t reg,
1161 					 uint32_t val, uint32_t mask)
1162 {
1163 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1164 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1165 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* equal */
1166 	amdgpu_ring_write(ring, reg << 2);
1167 	amdgpu_ring_write(ring, 0);
1168 	amdgpu_ring_write(ring, val); /* reference */
1169 	amdgpu_ring_write(ring, mask); /* mask */
1170 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1171 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10));
1172 }
1173 
1174 static void sdma_v5_2_ring_emit_reg_write_reg_wait(struct amdgpu_ring *ring,
1175 						   uint32_t reg0, uint32_t reg1,
1176 						   uint32_t ref, uint32_t mask)
1177 {
1178 	amdgpu_ring_emit_wreg(ring, reg0, ref);
1179 	/* wait for a cycle to reset vm_inv_eng*_ack */
1180 	amdgpu_ring_emit_reg_wait(ring, reg0, 0, 0);
1181 	amdgpu_ring_emit_reg_wait(ring, reg1, mask, mask);
1182 }
1183 
1184 static int sdma_v5_2_early_init(void *handle)
1185 {
1186 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1187 
1188 	sdma_v5_2_set_ring_funcs(adev);
1189 	sdma_v5_2_set_buffer_funcs(adev);
1190 	sdma_v5_2_set_vm_pte_funcs(adev);
1191 	sdma_v5_2_set_irq_funcs(adev);
1192 	sdma_v5_2_set_mqd_funcs(adev);
1193 
1194 	return 0;
1195 }
1196 
1197 static unsigned sdma_v5_2_seq_to_irq_id(int seq_num)
1198 {
1199 	switch (seq_num) {
1200 	case 0:
1201 		return SOC15_IH_CLIENTID_SDMA0;
1202 	case 1:
1203 		return SOC15_IH_CLIENTID_SDMA1;
1204 	case 2:
1205 		return SOC15_IH_CLIENTID_SDMA2;
1206 	case 3:
1207 		return SOC15_IH_CLIENTID_SDMA3_Sienna_Cichlid;
1208 	default:
1209 		break;
1210 	}
1211 	return -EINVAL;
1212 }
1213 
1214 static unsigned sdma_v5_2_seq_to_trap_id(int seq_num)
1215 {
1216 	switch (seq_num) {
1217 	case 0:
1218 		return SDMA0_5_0__SRCID__SDMA_TRAP;
1219 	case 1:
1220 		return SDMA1_5_0__SRCID__SDMA_TRAP;
1221 	case 2:
1222 		return SDMA2_5_0__SRCID__SDMA_TRAP;
1223 	case 3:
1224 		return SDMA3_5_0__SRCID__SDMA_TRAP;
1225 	default:
1226 		break;
1227 	}
1228 	return -EINVAL;
1229 }
1230 
1231 static int sdma_v5_2_sw_init(void *handle)
1232 {
1233 	struct amdgpu_ring *ring;
1234 	int r, i;
1235 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1236 
1237 	/* SDMA trap event */
1238 	for (i = 0; i < adev->sdma.num_instances; i++) {
1239 		r = amdgpu_irq_add_id(adev, sdma_v5_2_seq_to_irq_id(i),
1240 				      sdma_v5_2_seq_to_trap_id(i),
1241 				      &adev->sdma.trap_irq);
1242 		if (r)
1243 			return r;
1244 	}
1245 
1246 	r = amdgpu_sdma_init_microcode(adev, 0, true);
1247 	if (r) {
1248 		DRM_ERROR("Failed to load sdma firmware!\n");
1249 		return r;
1250 	}
1251 
1252 	for (i = 0; i < adev->sdma.num_instances; i++) {
1253 		ring = &adev->sdma.instance[i].ring;
1254 		ring->ring_obj = NULL;
1255 		ring->use_doorbell = true;
1256 		ring->me = i;
1257 
1258 		DRM_INFO("use_doorbell being set to: [%s]\n",
1259 				ring->use_doorbell?"true":"false");
1260 
1261 		ring->doorbell_index =
1262 			(adev->doorbell_index.sdma_engine[i] << 1); //get DWORD offset
1263 
1264 		ring->vm_hub = AMDGPU_GFXHUB(0);
1265 		sprintf(ring->name, "sdma%d", i);
1266 		r = amdgpu_ring_init(adev, ring, 1024, &adev->sdma.trap_irq,
1267 				     AMDGPU_SDMA_IRQ_INSTANCE0 + i,
1268 				     AMDGPU_RING_PRIO_DEFAULT, NULL);
1269 		if (r)
1270 			return r;
1271 	}
1272 
1273 	return r;
1274 }
1275 
1276 static int sdma_v5_2_sw_fini(void *handle)
1277 {
1278 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1279 	int i;
1280 
1281 	for (i = 0; i < adev->sdma.num_instances; i++)
1282 		amdgpu_ring_fini(&adev->sdma.instance[i].ring);
1283 
1284 	amdgpu_sdma_destroy_inst_ctx(adev, true);
1285 
1286 	return 0;
1287 }
1288 
1289 static int sdma_v5_2_hw_init(void *handle)
1290 {
1291 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1292 
1293 	return sdma_v5_2_start(adev);
1294 }
1295 
1296 static int sdma_v5_2_hw_fini(void *handle)
1297 {
1298 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1299 
1300 	if (amdgpu_sriov_vf(adev)) {
1301 		/* disable the scheduler for SDMA */
1302 		amdgpu_sdma_unset_buffer_funcs_helper(adev);
1303 		return 0;
1304 	}
1305 
1306 	sdma_v5_2_ctx_switch_enable(adev, false);
1307 	sdma_v5_2_enable(adev, false);
1308 
1309 	return 0;
1310 }
1311 
1312 static int sdma_v5_2_suspend(void *handle)
1313 {
1314 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1315 
1316 	return sdma_v5_2_hw_fini(adev);
1317 }
1318 
1319 static int sdma_v5_2_resume(void *handle)
1320 {
1321 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1322 
1323 	return sdma_v5_2_hw_init(adev);
1324 }
1325 
1326 static bool sdma_v5_2_is_idle(void *handle)
1327 {
1328 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1329 	u32 i;
1330 
1331 	for (i = 0; i < adev->sdma.num_instances; i++) {
1332 		u32 tmp = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_STATUS_REG));
1333 
1334 		if (!(tmp & SDMA0_STATUS_REG__IDLE_MASK))
1335 			return false;
1336 	}
1337 
1338 	return true;
1339 }
1340 
1341 static int sdma_v5_2_wait_for_idle(void *handle)
1342 {
1343 	unsigned i;
1344 	u32 sdma0, sdma1, sdma2, sdma3;
1345 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1346 
1347 	for (i = 0; i < adev->usec_timeout; i++) {
1348 		sdma0 = RREG32(sdma_v5_2_get_reg_offset(adev, 0, mmSDMA0_STATUS_REG));
1349 		sdma1 = RREG32(sdma_v5_2_get_reg_offset(adev, 1, mmSDMA0_STATUS_REG));
1350 		sdma2 = RREG32(sdma_v5_2_get_reg_offset(adev, 2, mmSDMA0_STATUS_REG));
1351 		sdma3 = RREG32(sdma_v5_2_get_reg_offset(adev, 3, mmSDMA0_STATUS_REG));
1352 
1353 		if (sdma0 & sdma1 & sdma2 & sdma3 & SDMA0_STATUS_REG__IDLE_MASK)
1354 			return 0;
1355 		udelay(1);
1356 	}
1357 	return -ETIMEDOUT;
1358 }
1359 
1360 static int sdma_v5_2_ring_preempt_ib(struct amdgpu_ring *ring)
1361 {
1362 	int i, r = 0;
1363 	struct amdgpu_device *adev = ring->adev;
1364 	u32 index = 0;
1365 	u64 sdma_gfx_preempt;
1366 
1367 	amdgpu_sdma_get_index_from_ring(ring, &index);
1368 	sdma_gfx_preempt =
1369 		sdma_v5_2_get_reg_offset(adev, index, mmSDMA0_GFX_PREEMPT);
1370 
1371 	/* assert preemption condition */
1372 	amdgpu_ring_set_preempt_cond_exec(ring, false);
1373 
1374 	/* emit the trailing fence */
1375 	ring->trail_seq += 1;
1376 	amdgpu_ring_alloc(ring, 10);
1377 	sdma_v5_2_ring_emit_fence(ring, ring->trail_fence_gpu_addr,
1378 				  ring->trail_seq, 0);
1379 	amdgpu_ring_commit(ring);
1380 
1381 	/* assert IB preemption */
1382 	WREG32(sdma_gfx_preempt, 1);
1383 
1384 	/* poll the trailing fence */
1385 	for (i = 0; i < adev->usec_timeout; i++) {
1386 		if (ring->trail_seq ==
1387 		    le32_to_cpu(*(ring->trail_fence_cpu_addr)))
1388 			break;
1389 		udelay(1);
1390 	}
1391 
1392 	if (i >= adev->usec_timeout) {
1393 		r = -EINVAL;
1394 		DRM_ERROR("ring %d failed to be preempted\n", ring->idx);
1395 	}
1396 
1397 	/* deassert IB preemption */
1398 	WREG32(sdma_gfx_preempt, 0);
1399 
1400 	/* deassert the preemption condition */
1401 	amdgpu_ring_set_preempt_cond_exec(ring, true);
1402 	return r;
1403 }
1404 
1405 static int sdma_v5_2_set_trap_irq_state(struct amdgpu_device *adev,
1406 					struct amdgpu_irq_src *source,
1407 					unsigned type,
1408 					enum amdgpu_interrupt_state state)
1409 {
1410 	u32 sdma_cntl;
1411 	u32 reg_offset = sdma_v5_2_get_reg_offset(adev, type, mmSDMA0_CNTL);
1412 
1413 	if (!amdgpu_sriov_vf(adev)) {
1414 		sdma_cntl = RREG32(reg_offset);
1415 		sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE,
1416 			       state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0);
1417 		WREG32(reg_offset, sdma_cntl);
1418 	}
1419 
1420 	return 0;
1421 }
1422 
1423 static int sdma_v5_2_process_trap_irq(struct amdgpu_device *adev,
1424 				      struct amdgpu_irq_src *source,
1425 				      struct amdgpu_iv_entry *entry)
1426 {
1427 	uint32_t mes_queue_id = entry->src_data[0];
1428 
1429 	DRM_DEBUG("IH: SDMA trap\n");
1430 
1431 	if (adev->enable_mes && (mes_queue_id & AMDGPU_FENCE_MES_QUEUE_FLAG)) {
1432 		struct amdgpu_mes_queue *queue;
1433 
1434 		mes_queue_id &= AMDGPU_FENCE_MES_QUEUE_ID_MASK;
1435 
1436 		spin_lock(&adev->mes.queue_id_lock);
1437 		queue = idr_find(&adev->mes.queue_id_idr, mes_queue_id);
1438 		if (queue) {
1439 			DRM_DEBUG("process smda queue id = %d\n", mes_queue_id);
1440 			amdgpu_fence_process(queue->ring);
1441 		}
1442 		spin_unlock(&adev->mes.queue_id_lock);
1443 		return 0;
1444 	}
1445 
1446 	switch (entry->client_id) {
1447 	case SOC15_IH_CLIENTID_SDMA0:
1448 		switch (entry->ring_id) {
1449 		case 0:
1450 			amdgpu_fence_process(&adev->sdma.instance[0].ring);
1451 			break;
1452 		case 1:
1453 			/* XXX compute */
1454 			break;
1455 		case 2:
1456 			/* XXX compute */
1457 			break;
1458 		case 3:
1459 			/* XXX page queue*/
1460 			break;
1461 		}
1462 		break;
1463 	case SOC15_IH_CLIENTID_SDMA1:
1464 		switch (entry->ring_id) {
1465 		case 0:
1466 			amdgpu_fence_process(&adev->sdma.instance[1].ring);
1467 			break;
1468 		case 1:
1469 			/* XXX compute */
1470 			break;
1471 		case 2:
1472 			/* XXX compute */
1473 			break;
1474 		case 3:
1475 			/* XXX page queue*/
1476 			break;
1477 		}
1478 		break;
1479 	case SOC15_IH_CLIENTID_SDMA2:
1480 		switch (entry->ring_id) {
1481 		case 0:
1482 			amdgpu_fence_process(&adev->sdma.instance[2].ring);
1483 			break;
1484 		case 1:
1485 			/* XXX compute */
1486 			break;
1487 		case 2:
1488 			/* XXX compute */
1489 			break;
1490 		case 3:
1491 			/* XXX page queue*/
1492 			break;
1493 		}
1494 		break;
1495 	case SOC15_IH_CLIENTID_SDMA3_Sienna_Cichlid:
1496 		switch (entry->ring_id) {
1497 		case 0:
1498 			amdgpu_fence_process(&adev->sdma.instance[3].ring);
1499 			break;
1500 		case 1:
1501 			/* XXX compute */
1502 			break;
1503 		case 2:
1504 			/* XXX compute */
1505 			break;
1506 		case 3:
1507 			/* XXX page queue*/
1508 			break;
1509 		}
1510 		break;
1511 	}
1512 	return 0;
1513 }
1514 
1515 static int sdma_v5_2_process_illegal_inst_irq(struct amdgpu_device *adev,
1516 					      struct amdgpu_irq_src *source,
1517 					      struct amdgpu_iv_entry *entry)
1518 {
1519 	return 0;
1520 }
1521 
1522 static bool sdma_v5_2_firmware_mgcg_support(struct amdgpu_device *adev,
1523 						     int i)
1524 {
1525 	switch (adev->ip_versions[SDMA0_HWIP][0]) {
1526 	case IP_VERSION(5, 2, 1):
1527 		if (adev->sdma.instance[i].fw_version < 70)
1528 			return false;
1529 		break;
1530 	case IP_VERSION(5, 2, 3):
1531 		if (adev->sdma.instance[i].fw_version < 47)
1532 			return false;
1533 		break;
1534 	case IP_VERSION(5, 2, 7):
1535 		if (adev->sdma.instance[i].fw_version < 9)
1536 			return false;
1537 		break;
1538 	default:
1539 		return true;
1540 	}
1541 
1542 	return true;
1543 
1544 }
1545 
1546 static void sdma_v5_2_update_medium_grain_clock_gating(struct amdgpu_device *adev,
1547 						       bool enable)
1548 {
1549 	uint32_t data, def;
1550 	int i;
1551 
1552 	for (i = 0; i < adev->sdma.num_instances; i++) {
1553 
1554 		if (!sdma_v5_2_firmware_mgcg_support(adev, i))
1555 			adev->cg_flags &= ~AMD_CG_SUPPORT_SDMA_MGCG;
1556 
1557 		if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
1558 			/* Enable sdma clock gating */
1559 			def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL));
1560 			data &= ~(SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1561 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1562 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1563 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1564 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK |
1565 				  SDMA0_CLK_CTRL__SOFT_OVERRIDER_REG_MASK);
1566 			if (def != data)
1567 				WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data);
1568 		} else {
1569 			/* Disable sdma clock gating */
1570 			def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL));
1571 			data |= (SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1572 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1573 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1574 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1575 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK |
1576 				 SDMA0_CLK_CTRL__SOFT_OVERRIDER_REG_MASK);
1577 			if (def != data)
1578 				WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data);
1579 		}
1580 	}
1581 }
1582 
1583 static void sdma_v5_2_update_medium_grain_light_sleep(struct amdgpu_device *adev,
1584 						      bool enable)
1585 {
1586 	uint32_t data, def;
1587 	int i;
1588 
1589 	for (i = 0; i < adev->sdma.num_instances; i++) {
1590 
1591 		if (adev->sdma.instance[i].fw_version < 70 && adev->ip_versions[SDMA0_HWIP][0] == IP_VERSION(5, 2, 1))
1592 			adev->cg_flags &= ~AMD_CG_SUPPORT_SDMA_LS;
1593 
1594 		if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) {
1595 			/* Enable sdma mem light sleep */
1596 			def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL));
1597 			data |= SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1598 			if (def != data)
1599 				WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data);
1600 
1601 		} else {
1602 			/* Disable sdma mem light sleep */
1603 			def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL));
1604 			data &= ~SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1605 			if (def != data)
1606 				WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data);
1607 
1608 		}
1609 	}
1610 }
1611 
1612 static int sdma_v5_2_set_clockgating_state(void *handle,
1613 					   enum amd_clockgating_state state)
1614 {
1615 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1616 
1617 	if (amdgpu_sriov_vf(adev))
1618 		return 0;
1619 
1620 	switch (adev->ip_versions[SDMA0_HWIP][0]) {
1621 	case IP_VERSION(5, 2, 0):
1622 	case IP_VERSION(5, 2, 2):
1623 	case IP_VERSION(5, 2, 1):
1624 	case IP_VERSION(5, 2, 4):
1625 	case IP_VERSION(5, 2, 5):
1626 	case IP_VERSION(5, 2, 6):
1627 	case IP_VERSION(5, 2, 3):
1628 	case IP_VERSION(5, 2, 7):
1629 		sdma_v5_2_update_medium_grain_clock_gating(adev,
1630 				state == AMD_CG_STATE_GATE);
1631 		sdma_v5_2_update_medium_grain_light_sleep(adev,
1632 				state == AMD_CG_STATE_GATE);
1633 		break;
1634 	default:
1635 		break;
1636 	}
1637 
1638 	return 0;
1639 }
1640 
1641 static int sdma_v5_2_set_powergating_state(void *handle,
1642 					  enum amd_powergating_state state)
1643 {
1644 	return 0;
1645 }
1646 
1647 static void sdma_v5_2_get_clockgating_state(void *handle, u64 *flags)
1648 {
1649 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1650 	int data;
1651 
1652 	if (amdgpu_sriov_vf(adev))
1653 		*flags = 0;
1654 
1655 	/* AMD_CG_SUPPORT_SDMA_MGCG */
1656 	data = RREG32(sdma_v5_2_get_reg_offset(adev, 0, mmSDMA0_CLK_CTRL));
1657 	if (!(data & SDMA0_CLK_CTRL__CGCG_EN_OVERRIDE_MASK))
1658 		*flags |= AMD_CG_SUPPORT_SDMA_MGCG;
1659 
1660 	/* AMD_CG_SUPPORT_SDMA_LS */
1661 	data = RREG32_KIQ(sdma_v5_2_get_reg_offset(adev, 0, mmSDMA0_POWER_CNTL));
1662 	if (data & SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK)
1663 		*flags |= AMD_CG_SUPPORT_SDMA_LS;
1664 }
1665 
1666 static void sdma_v5_2_ring_begin_use(struct amdgpu_ring *ring)
1667 {
1668 	struct amdgpu_device *adev = ring->adev;
1669 
1670 	/* SDMA 5.2.3 (RMB) FW doesn't seem to properly
1671 	 * disallow GFXOFF in some cases leading to
1672 	 * hangs in SDMA.  Disallow GFXOFF while SDMA is active.
1673 	 * We can probably just limit this to 5.2.3,
1674 	 * but it shouldn't hurt for other parts since
1675 	 * this GFXOFF will be disallowed anyway when SDMA is
1676 	 * active, this just makes it explicit.
1677 	 * sdma_v5_2_ring_set_wptr() takes advantage of this
1678 	 * to update the wptr because sometimes SDMA seems to miss
1679 	 * doorbells when entering PG.  If you remove this, update
1680 	 * sdma_v5_2_ring_set_wptr() as well!
1681 	 */
1682 	amdgpu_gfx_off_ctrl(adev, false);
1683 }
1684 
1685 static void sdma_v5_2_ring_end_use(struct amdgpu_ring *ring)
1686 {
1687 	struct amdgpu_device *adev = ring->adev;
1688 
1689 	/* SDMA 5.2.3 (RMB) FW doesn't seem to properly
1690 	 * disallow GFXOFF in some cases leading to
1691 	 * hangs in SDMA.  Allow GFXOFF when SDMA is complete.
1692 	 */
1693 	amdgpu_gfx_off_ctrl(adev, true);
1694 }
1695 
1696 const struct amd_ip_funcs sdma_v5_2_ip_funcs = {
1697 	.name = "sdma_v5_2",
1698 	.early_init = sdma_v5_2_early_init,
1699 	.late_init = NULL,
1700 	.sw_init = sdma_v5_2_sw_init,
1701 	.sw_fini = sdma_v5_2_sw_fini,
1702 	.hw_init = sdma_v5_2_hw_init,
1703 	.hw_fini = sdma_v5_2_hw_fini,
1704 	.suspend = sdma_v5_2_suspend,
1705 	.resume = sdma_v5_2_resume,
1706 	.is_idle = sdma_v5_2_is_idle,
1707 	.wait_for_idle = sdma_v5_2_wait_for_idle,
1708 	.soft_reset = sdma_v5_2_soft_reset,
1709 	.set_clockgating_state = sdma_v5_2_set_clockgating_state,
1710 	.set_powergating_state = sdma_v5_2_set_powergating_state,
1711 	.get_clockgating_state = sdma_v5_2_get_clockgating_state,
1712 };
1713 
1714 static const struct amdgpu_ring_funcs sdma_v5_2_ring_funcs = {
1715 	.type = AMDGPU_RING_TYPE_SDMA,
1716 	.align_mask = 0xf,
1717 	.nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP),
1718 	.support_64bit_ptrs = true,
1719 	.secure_submission_supported = true,
1720 	.get_rptr = sdma_v5_2_ring_get_rptr,
1721 	.get_wptr = sdma_v5_2_ring_get_wptr,
1722 	.set_wptr = sdma_v5_2_ring_set_wptr,
1723 	.emit_frame_size =
1724 		5 + /* sdma_v5_2_ring_init_cond_exec */
1725 		6 + /* sdma_v5_2_ring_emit_hdp_flush */
1726 		3 + /* hdp_invalidate */
1727 		6 + /* sdma_v5_2_ring_emit_pipeline_sync */
1728 		/* sdma_v5_2_ring_emit_vm_flush */
1729 		SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 +
1730 		SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6 +
1731 		10 + 10 + 10, /* sdma_v5_2_ring_emit_fence x3 for user fence, vm fence */
1732 	.emit_ib_size = 7 + 6, /* sdma_v5_2_ring_emit_ib */
1733 	.emit_ib = sdma_v5_2_ring_emit_ib,
1734 	.emit_mem_sync = sdma_v5_2_ring_emit_mem_sync,
1735 	.emit_fence = sdma_v5_2_ring_emit_fence,
1736 	.emit_pipeline_sync = sdma_v5_2_ring_emit_pipeline_sync,
1737 	.emit_vm_flush = sdma_v5_2_ring_emit_vm_flush,
1738 	.emit_hdp_flush = sdma_v5_2_ring_emit_hdp_flush,
1739 	.test_ring = sdma_v5_2_ring_test_ring,
1740 	.test_ib = sdma_v5_2_ring_test_ib,
1741 	.insert_nop = sdma_v5_2_ring_insert_nop,
1742 	.pad_ib = sdma_v5_2_ring_pad_ib,
1743 	.begin_use = sdma_v5_2_ring_begin_use,
1744 	.end_use = sdma_v5_2_ring_end_use,
1745 	.emit_wreg = sdma_v5_2_ring_emit_wreg,
1746 	.emit_reg_wait = sdma_v5_2_ring_emit_reg_wait,
1747 	.emit_reg_write_reg_wait = sdma_v5_2_ring_emit_reg_write_reg_wait,
1748 	.init_cond_exec = sdma_v5_2_ring_init_cond_exec,
1749 	.patch_cond_exec = sdma_v5_2_ring_patch_cond_exec,
1750 	.preempt_ib = sdma_v5_2_ring_preempt_ib,
1751 };
1752 
1753 static void sdma_v5_2_set_ring_funcs(struct amdgpu_device *adev)
1754 {
1755 	int i;
1756 
1757 	for (i = 0; i < adev->sdma.num_instances; i++) {
1758 		adev->sdma.instance[i].ring.funcs = &sdma_v5_2_ring_funcs;
1759 		adev->sdma.instance[i].ring.me = i;
1760 	}
1761 }
1762 
1763 static const struct amdgpu_irq_src_funcs sdma_v5_2_trap_irq_funcs = {
1764 	.set = sdma_v5_2_set_trap_irq_state,
1765 	.process = sdma_v5_2_process_trap_irq,
1766 };
1767 
1768 static const struct amdgpu_irq_src_funcs sdma_v5_2_illegal_inst_irq_funcs = {
1769 	.process = sdma_v5_2_process_illegal_inst_irq,
1770 };
1771 
1772 static void sdma_v5_2_set_irq_funcs(struct amdgpu_device *adev)
1773 {
1774 	adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_INSTANCE0 +
1775 					adev->sdma.num_instances;
1776 	adev->sdma.trap_irq.funcs = &sdma_v5_2_trap_irq_funcs;
1777 	adev->sdma.illegal_inst_irq.funcs = &sdma_v5_2_illegal_inst_irq_funcs;
1778 }
1779 
1780 /**
1781  * sdma_v5_2_emit_copy_buffer - copy buffer using the sDMA engine
1782  *
1783  * @ib: indirect buffer to copy to
1784  * @src_offset: src GPU address
1785  * @dst_offset: dst GPU address
1786  * @byte_count: number of bytes to xfer
1787  * @tmz: if a secure copy should be used
1788  *
1789  * Copy GPU buffers using the DMA engine.
1790  * Used by the amdgpu ttm implementation to move pages if
1791  * registered as the asic copy callback.
1792  */
1793 static void sdma_v5_2_emit_copy_buffer(struct amdgpu_ib *ib,
1794 				       uint64_t src_offset,
1795 				       uint64_t dst_offset,
1796 				       uint32_t byte_count,
1797 				       bool tmz)
1798 {
1799 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1800 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR) |
1801 		SDMA_PKT_COPY_LINEAR_HEADER_TMZ(tmz ? 1 : 0);
1802 	ib->ptr[ib->length_dw++] = byte_count - 1;
1803 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1804 	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1805 	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1806 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1807 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1808 }
1809 
1810 /**
1811  * sdma_v5_2_emit_fill_buffer - fill buffer using the sDMA engine
1812  *
1813  * @ib: indirect buffer to fill
1814  * @src_data: value to write to buffer
1815  * @dst_offset: dst GPU address
1816  * @byte_count: number of bytes to xfer
1817  *
1818  * Fill GPU buffers using the DMA engine.
1819  */
1820 static void sdma_v5_2_emit_fill_buffer(struct amdgpu_ib *ib,
1821 				       uint32_t src_data,
1822 				       uint64_t dst_offset,
1823 				       uint32_t byte_count)
1824 {
1825 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL);
1826 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1827 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1828 	ib->ptr[ib->length_dw++] = src_data;
1829 	ib->ptr[ib->length_dw++] = byte_count - 1;
1830 }
1831 
1832 static const struct amdgpu_buffer_funcs sdma_v5_2_buffer_funcs = {
1833 	.copy_max_bytes = 0x400000,
1834 	.copy_num_dw = 7,
1835 	.emit_copy_buffer = sdma_v5_2_emit_copy_buffer,
1836 
1837 	.fill_max_bytes = 0x400000,
1838 	.fill_num_dw = 5,
1839 	.emit_fill_buffer = sdma_v5_2_emit_fill_buffer,
1840 };
1841 
1842 static void sdma_v5_2_set_buffer_funcs(struct amdgpu_device *adev)
1843 {
1844 	if (adev->mman.buffer_funcs == NULL) {
1845 		adev->mman.buffer_funcs = &sdma_v5_2_buffer_funcs;
1846 		adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1847 	}
1848 }
1849 
1850 static const struct amdgpu_vm_pte_funcs sdma_v5_2_vm_pte_funcs = {
1851 	.copy_pte_num_dw = 7,
1852 	.copy_pte = sdma_v5_2_vm_copy_pte,
1853 	.write_pte = sdma_v5_2_vm_write_pte,
1854 	.set_pte_pde = sdma_v5_2_vm_set_pte_pde,
1855 };
1856 
1857 static void sdma_v5_2_set_vm_pte_funcs(struct amdgpu_device *adev)
1858 {
1859 	unsigned i;
1860 
1861 	if (adev->vm_manager.vm_pte_funcs == NULL) {
1862 		adev->vm_manager.vm_pte_funcs = &sdma_v5_2_vm_pte_funcs;
1863 		for (i = 0; i < adev->sdma.num_instances; i++) {
1864 			adev->vm_manager.vm_pte_scheds[i] =
1865 				&adev->sdma.instance[i].ring.sched;
1866 		}
1867 		adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
1868 	}
1869 }
1870 
1871 const struct amdgpu_ip_block_version sdma_v5_2_ip_block = {
1872 	.type = AMD_IP_BLOCK_TYPE_SDMA,
1873 	.major = 5,
1874 	.minor = 2,
1875 	.rev = 0,
1876 	.funcs = &sdma_v5_2_ip_funcs,
1877 };
1878