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 /**
93  * sdma_v5_2_init_microcode - load ucode images from disk
94  *
95  * @adev: amdgpu_device pointer
96  *
97  * Use the firmware interface to load the ucode images into
98  * the driver (not loaded into hw).
99  * Returns 0 on success, error on failure.
100  */
101 
102 // emulation only, won't work on real chip
103 // navi10 real chip need to use PSP to load firmware
104 static int sdma_v5_2_init_microcode(struct amdgpu_device *adev)
105 {
106 	const char *chip_name;
107 	char fw_name[40];
108 
109 	DRM_DEBUG("\n");
110 
111 	switch (adev->ip_versions[SDMA0_HWIP][0]) {
112 	case IP_VERSION(5, 2, 0):
113 		chip_name = "sienna_cichlid_sdma";
114 		break;
115 	case IP_VERSION(5, 2, 2):
116 		chip_name = "navy_flounder_sdma";
117 		break;
118 	case IP_VERSION(5, 2, 1):
119 		chip_name = "vangogh_sdma";
120 		break;
121 	case IP_VERSION(5, 2, 4):
122 		chip_name = "dimgrey_cavefish_sdma";
123 		break;
124 	case IP_VERSION(5, 2, 5):
125 		chip_name = "beige_goby_sdma";
126 		break;
127 	case IP_VERSION(5, 2, 3):
128 		chip_name = "yellow_carp_sdma";
129 		break;
130 	case IP_VERSION(5, 2, 6):
131 		chip_name = "sdma_5_2_6";
132 		break;
133 	case IP_VERSION(5, 2, 7):
134 		chip_name = "sdma_5_2_7";
135 		break;
136 	default:
137 		BUG();
138 	}
139 
140 	snprintf(fw_name, sizeof(fw_name), "amdgpu/%s.bin", chip_name);
141 
142 	return amdgpu_sdma_init_microcode(adev, fw_name, 0, true);
143 }
144 
145 static unsigned sdma_v5_2_ring_init_cond_exec(struct amdgpu_ring *ring)
146 {
147 	unsigned ret;
148 
149 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_COND_EXE));
150 	amdgpu_ring_write(ring, lower_32_bits(ring->cond_exe_gpu_addr));
151 	amdgpu_ring_write(ring, upper_32_bits(ring->cond_exe_gpu_addr));
152 	amdgpu_ring_write(ring, 1);
153 	ret = ring->wptr & ring->buf_mask;/* this is the offset we need patch later */
154 	amdgpu_ring_write(ring, 0x55aa55aa);/* insert dummy here and patch it later */
155 
156 	return ret;
157 }
158 
159 static void sdma_v5_2_ring_patch_cond_exec(struct amdgpu_ring *ring,
160 					   unsigned offset)
161 {
162 	unsigned cur;
163 
164 	BUG_ON(offset > ring->buf_mask);
165 	BUG_ON(ring->ring[offset] != 0x55aa55aa);
166 
167 	cur = (ring->wptr - 1) & ring->buf_mask;
168 	if (cur > offset)
169 		ring->ring[offset] = cur - offset;
170 	else
171 		ring->ring[offset] = (ring->buf_mask + 1) - offset + cur;
172 }
173 
174 /**
175  * sdma_v5_2_ring_get_rptr - get the current read pointer
176  *
177  * @ring: amdgpu ring pointer
178  *
179  * Get the current rptr from the hardware (NAVI10+).
180  */
181 static uint64_t sdma_v5_2_ring_get_rptr(struct amdgpu_ring *ring)
182 {
183 	u64 *rptr;
184 
185 	/* XXX check if swapping is necessary on BE */
186 	rptr = (u64 *)ring->rptr_cpu_addr;
187 
188 	DRM_DEBUG("rptr before shift == 0x%016llx\n", *rptr);
189 	return ((*rptr) >> 2);
190 }
191 
192 /**
193  * sdma_v5_2_ring_get_wptr - get the current write pointer
194  *
195  * @ring: amdgpu ring pointer
196  *
197  * Get the current wptr from the hardware (NAVI10+).
198  */
199 static uint64_t sdma_v5_2_ring_get_wptr(struct amdgpu_ring *ring)
200 {
201 	struct amdgpu_device *adev = ring->adev;
202 	u64 wptr;
203 
204 	if (ring->use_doorbell) {
205 		/* XXX check if swapping is necessary on BE */
206 		wptr = READ_ONCE(*((u64 *)ring->wptr_cpu_addr));
207 		DRM_DEBUG("wptr/doorbell before shift == 0x%016llx\n", wptr);
208 	} else {
209 		wptr = RREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI));
210 		wptr = wptr << 32;
211 		wptr |= RREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR));
212 		DRM_DEBUG("wptr before shift [%i] wptr == 0x%016llx\n", ring->me, wptr);
213 	}
214 
215 	return wptr >> 2;
216 }
217 
218 /**
219  * sdma_v5_2_ring_set_wptr - commit the write pointer
220  *
221  * @ring: amdgpu ring pointer
222  *
223  * Write the wptr back to the hardware (NAVI10+).
224  */
225 static void sdma_v5_2_ring_set_wptr(struct amdgpu_ring *ring)
226 {
227 	struct amdgpu_device *adev = ring->adev;
228 
229 	DRM_DEBUG("Setting write pointer\n");
230 	if (ring->use_doorbell) {
231 		DRM_DEBUG("Using doorbell -- "
232 				"wptr_offs == 0x%08x "
233 				"lower_32_bits(ring->wptr << 2) == 0x%08x "
234 				"upper_32_bits(ring->wptr << 2) == 0x%08x\n",
235 				ring->wptr_offs,
236 				lower_32_bits(ring->wptr << 2),
237 				upper_32_bits(ring->wptr << 2));
238 		/* XXX check if swapping is necessary on BE */
239 		atomic64_set((atomic64_t *)ring->wptr_cpu_addr,
240 			     ring->wptr << 2);
241 		DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016llx)\n",
242 				ring->doorbell_index, ring->wptr << 2);
243 		WDOORBELL64(ring->doorbell_index, ring->wptr << 2);
244 	} else {
245 		DRM_DEBUG("Not using doorbell -- "
246 				"mmSDMA%i_GFX_RB_WPTR == 0x%08x "
247 				"mmSDMA%i_GFX_RB_WPTR_HI == 0x%08x\n",
248 				ring->me,
249 				lower_32_bits(ring->wptr << 2),
250 				ring->me,
251 				upper_32_bits(ring->wptr << 2));
252 		WREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR),
253 			lower_32_bits(ring->wptr << 2));
254 		WREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI),
255 			upper_32_bits(ring->wptr << 2));
256 	}
257 }
258 
259 static void sdma_v5_2_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
260 {
261 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
262 	int i;
263 
264 	for (i = 0; i < count; i++)
265 		if (sdma && sdma->burst_nop && (i == 0))
266 			amdgpu_ring_write(ring, ring->funcs->nop |
267 				SDMA_PKT_NOP_HEADER_COUNT(count - 1));
268 		else
269 			amdgpu_ring_write(ring, ring->funcs->nop);
270 }
271 
272 /**
273  * sdma_v5_2_ring_emit_ib - Schedule an IB on the DMA engine
274  *
275  * @ring: amdgpu ring pointer
276  * @job: job to retrieve vmid from
277  * @ib: IB object to schedule
278  * @flags: unused
279  *
280  * Schedule an IB in the DMA ring.
281  */
282 static void sdma_v5_2_ring_emit_ib(struct amdgpu_ring *ring,
283 				   struct amdgpu_job *job,
284 				   struct amdgpu_ib *ib,
285 				   uint32_t flags)
286 {
287 	unsigned vmid = AMDGPU_JOB_GET_VMID(job);
288 	uint64_t csa_mc_addr = amdgpu_sdma_get_csa_mc_addr(ring, vmid);
289 
290 	/* An IB packet must end on a 8 DW boundary--the next dword
291 	 * must be on a 8-dword boundary. Our IB packet below is 6
292 	 * dwords long, thus add x number of NOPs, such that, in
293 	 * modular arithmetic,
294 	 * wptr + 6 + x = 8k, k >= 0, which in C is,
295 	 * (wptr + 6 + x) % 8 = 0.
296 	 * The expression below, is a solution of x.
297 	 */
298 	sdma_v5_2_ring_insert_nop(ring, (2 - lower_32_bits(ring->wptr)) & 7);
299 
300 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) |
301 			  SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf));
302 	/* base must be 32 byte aligned */
303 	amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
304 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
305 	amdgpu_ring_write(ring, ib->length_dw);
306 	amdgpu_ring_write(ring, lower_32_bits(csa_mc_addr));
307 	amdgpu_ring_write(ring, upper_32_bits(csa_mc_addr));
308 }
309 
310 /**
311  * sdma_v5_2_ring_emit_mem_sync - flush the IB by graphics cache rinse
312  *
313  * @ring: amdgpu ring pointer
314  *
315  * flush the IB by graphics cache rinse.
316  */
317 static void sdma_v5_2_ring_emit_mem_sync(struct amdgpu_ring *ring)
318 {
319 	uint32_t gcr_cntl = SDMA_GCR_GL2_INV | SDMA_GCR_GL2_WB |
320 			    SDMA_GCR_GLM_INV | SDMA_GCR_GL1_INV |
321 			    SDMA_GCR_GLV_INV | SDMA_GCR_GLK_INV |
322 			    SDMA_GCR_GLI_INV(1);
323 
324 	/* flush entire cache L0/L1/L2, this can be optimized by performance requirement */
325 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_GCR_REQ));
326 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD1_BASE_VA_31_7(0));
327 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD2_GCR_CONTROL_15_0(gcr_cntl) |
328 			SDMA_PKT_GCR_REQ_PAYLOAD2_BASE_VA_47_32(0));
329 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD3_LIMIT_VA_31_7(0) |
330 			SDMA_PKT_GCR_REQ_PAYLOAD3_GCR_CONTROL_18_16(gcr_cntl >> 16));
331 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD4_LIMIT_VA_47_32(0) |
332 			SDMA_PKT_GCR_REQ_PAYLOAD4_VMID(0));
333 }
334 
335 /**
336  * sdma_v5_2_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
337  *
338  * @ring: amdgpu ring pointer
339  *
340  * Emit an hdp flush packet on the requested DMA ring.
341  */
342 static void sdma_v5_2_ring_emit_hdp_flush(struct amdgpu_ring *ring)
343 {
344 	struct amdgpu_device *adev = ring->adev;
345 	u32 ref_and_mask = 0;
346 	const struct nbio_hdp_flush_reg *nbio_hf_reg = adev->nbio.hdp_flush_reg;
347 
348 	ref_and_mask = nbio_hf_reg->ref_and_mask_sdma0 << ring->me;
349 
350 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
351 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) |
352 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
353 	amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_done_offset(adev)) << 2);
354 	amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_req_offset(adev)) << 2);
355 	amdgpu_ring_write(ring, ref_and_mask); /* reference */
356 	amdgpu_ring_write(ring, ref_and_mask); /* mask */
357 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
358 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
359 }
360 
361 /**
362  * sdma_v5_2_ring_emit_fence - emit a fence on the DMA ring
363  *
364  * @ring: amdgpu ring pointer
365  * @addr: address
366  * @seq: sequence number
367  * @flags: fence related flags
368  *
369  * Add a DMA fence packet to the ring to write
370  * the fence seq number and DMA trap packet to generate
371  * an interrupt if needed.
372  */
373 static void sdma_v5_2_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
374 				      unsigned flags)
375 {
376 	bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
377 	/* write the fence */
378 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) |
379 			  SDMA_PKT_FENCE_HEADER_MTYPE(0x3)); /* Ucached(UC) */
380 	/* zero in first two bits */
381 	BUG_ON(addr & 0x3);
382 	amdgpu_ring_write(ring, lower_32_bits(addr));
383 	amdgpu_ring_write(ring, upper_32_bits(addr));
384 	amdgpu_ring_write(ring, lower_32_bits(seq));
385 
386 	/* optionally write high bits as well */
387 	if (write64bit) {
388 		addr += 4;
389 		amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) |
390 				  SDMA_PKT_FENCE_HEADER_MTYPE(0x3));
391 		/* zero in first two bits */
392 		BUG_ON(addr & 0x3);
393 		amdgpu_ring_write(ring, lower_32_bits(addr));
394 		amdgpu_ring_write(ring, upper_32_bits(addr));
395 		amdgpu_ring_write(ring, upper_32_bits(seq));
396 	}
397 
398 	if ((flags & AMDGPU_FENCE_FLAG_INT)) {
399 		uint32_t ctx = ring->is_mes_queue ?
400 			(ring->hw_queue_id | AMDGPU_FENCE_MES_QUEUE_FLAG) : 0;
401 		/* generate an interrupt */
402 		amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP));
403 		amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(ctx));
404 	}
405 }
406 
407 
408 /**
409  * sdma_v5_2_gfx_stop - stop the gfx async dma engines
410  *
411  * @adev: amdgpu_device pointer
412  *
413  * Stop the gfx async dma ring buffers.
414  */
415 static void sdma_v5_2_gfx_stop(struct amdgpu_device *adev)
416 {
417 	u32 rb_cntl, ib_cntl;
418 	int i;
419 
420 	amdgpu_sdma_unset_buffer_funcs_helper(adev);
421 
422 	for (i = 0; i < adev->sdma.num_instances; i++) {
423 		rb_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL));
424 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 0);
425 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
426 		ib_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL));
427 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 0);
428 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl);
429 	}
430 }
431 
432 /**
433  * sdma_v5_2_rlc_stop - stop the compute async dma engines
434  *
435  * @adev: amdgpu_device pointer
436  *
437  * Stop the compute async dma queues.
438  */
439 static void sdma_v5_2_rlc_stop(struct amdgpu_device *adev)
440 {
441 	/* XXX todo */
442 }
443 
444 /**
445  * sdma_v5_2_ctx_switch_enable - stop the async dma engines context switch
446  *
447  * @adev: amdgpu_device pointer
448  * @enable: enable/disable the DMA MEs context switch.
449  *
450  * Halt or unhalt the async dma engines context switch.
451  */
452 static void sdma_v5_2_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
453 {
454 	u32 f32_cntl, phase_quantum = 0;
455 	int i;
456 
457 	if (amdgpu_sdma_phase_quantum) {
458 		unsigned value = amdgpu_sdma_phase_quantum;
459 		unsigned unit = 0;
460 
461 		while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
462 				SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) {
463 			value = (value + 1) >> 1;
464 			unit++;
465 		}
466 		if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
467 			    SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) {
468 			value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
469 				 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT);
470 			unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
471 				SDMA0_PHASE0_QUANTUM__UNIT__SHIFT);
472 			WARN_ONCE(1,
473 			"clamping sdma_phase_quantum to %uK clock cycles\n",
474 				  value << unit);
475 		}
476 		phase_quantum =
477 			value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
478 			unit  << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
479 	}
480 
481 	for (i = 0; i < adev->sdma.num_instances; i++) {
482 		if (enable && amdgpu_sdma_phase_quantum) {
483 			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE0_QUANTUM),
484 			       phase_quantum);
485 			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE1_QUANTUM),
486 			       phase_quantum);
487 			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE2_QUANTUM),
488 			       phase_quantum);
489 		}
490 
491 		if (!amdgpu_sriov_vf(adev)) {
492 			f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
493 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
494 					AUTO_CTXSW_ENABLE, enable ? 1 : 0);
495 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl);
496 		}
497 	}
498 
499 }
500 
501 /**
502  * sdma_v5_2_enable - stop the async dma engines
503  *
504  * @adev: amdgpu_device pointer
505  * @enable: enable/disable the DMA MEs.
506  *
507  * Halt or unhalt the async dma engines.
508  */
509 static void sdma_v5_2_enable(struct amdgpu_device *adev, bool enable)
510 {
511 	u32 f32_cntl;
512 	int i;
513 
514 	if (!enable) {
515 		sdma_v5_2_gfx_stop(adev);
516 		sdma_v5_2_rlc_stop(adev);
517 	}
518 
519 	if (!amdgpu_sriov_vf(adev)) {
520 		for (i = 0; i < adev->sdma.num_instances; i++) {
521 			f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
522 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, enable ? 0 : 1);
523 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl);
524 		}
525 	}
526 }
527 
528 /**
529  * sdma_v5_2_gfx_resume - setup and start the async dma engines
530  *
531  * @adev: amdgpu_device pointer
532  *
533  * Set up the gfx DMA ring buffers and enable them.
534  * Returns 0 for success, error for failure.
535  */
536 static int sdma_v5_2_gfx_resume(struct amdgpu_device *adev)
537 {
538 	struct amdgpu_ring *ring;
539 	u32 rb_cntl, ib_cntl;
540 	u32 rb_bufsz;
541 	u32 doorbell;
542 	u32 doorbell_offset;
543 	u32 temp;
544 	u32 wptr_poll_cntl;
545 	u64 wptr_gpu_addr;
546 	int i, r;
547 
548 	for (i = 0; i < adev->sdma.num_instances; i++) {
549 		ring = &adev->sdma.instance[i].ring;
550 
551 		if (!amdgpu_sriov_vf(adev))
552 			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL), 0);
553 
554 		/* Set ring buffer size in dwords */
555 		rb_bufsz = order_base_2(ring->ring_size / 4);
556 		rb_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL));
557 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz);
558 #ifdef __BIG_ENDIAN
559 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1);
560 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL,
561 					RPTR_WRITEBACK_SWAP_ENABLE, 1);
562 #endif
563 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
564 
565 		/* Initialize the ring buffer's read and write pointers */
566 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR), 0);
567 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_HI), 0);
568 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), 0);
569 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), 0);
570 
571 		/* setup the wptr shadow polling */
572 		wptr_gpu_addr = ring->wptr_gpu_addr;
573 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_LO),
574 		       lower_32_bits(wptr_gpu_addr));
575 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_HI),
576 		       upper_32_bits(wptr_gpu_addr));
577 		wptr_poll_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i,
578 							 mmSDMA0_GFX_RB_WPTR_POLL_CNTL));
579 		wptr_poll_cntl = REG_SET_FIELD(wptr_poll_cntl,
580 					       SDMA0_GFX_RB_WPTR_POLL_CNTL,
581 					       F32_POLL_ENABLE, 1);
582 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_CNTL),
583 		       wptr_poll_cntl);
584 
585 		/* set the wb address whether it's enabled or not */
586 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_HI),
587 		       upper_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFF);
588 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_LO),
589 		       lower_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFC);
590 
591 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
592 
593 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE), ring->gpu_addr >> 8);
594 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE_HI), ring->gpu_addr >> 40);
595 
596 		ring->wptr = 0;
597 
598 		/* before programing wptr to a less value, need set minor_ptr_update first */
599 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 1);
600 
601 		if (!amdgpu_sriov_vf(adev)) { /* only bare-metal use register write for wptr */
602 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), lower_32_bits(ring->wptr << 2));
603 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), upper_32_bits(ring->wptr << 2));
604 		}
605 
606 		doorbell = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL));
607 		doorbell_offset = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL_OFFSET));
608 
609 		if (ring->use_doorbell) {
610 			doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 1);
611 			doorbell_offset = REG_SET_FIELD(doorbell_offset, SDMA0_GFX_DOORBELL_OFFSET,
612 					OFFSET, ring->doorbell_index);
613 		} else {
614 			doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 0);
615 		}
616 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL), doorbell);
617 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL_OFFSET), doorbell_offset);
618 
619 		adev->nbio.funcs->sdma_doorbell_range(adev, i, ring->use_doorbell,
620 						      ring->doorbell_index,
621 						      adev->doorbell_index.sdma_doorbell_range);
622 
623 		if (amdgpu_sriov_vf(adev))
624 			sdma_v5_2_ring_set_wptr(ring);
625 
626 		/* set minor_ptr_update to 0 after wptr programed */
627 
628 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 0);
629 
630 		/* SRIOV VF has no control of any of registers below */
631 		if (!amdgpu_sriov_vf(adev)) {
632 			/* set utc l1 enable flag always to 1 */
633 			temp = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
634 			temp = REG_SET_FIELD(temp, SDMA0_CNTL, UTC_L1_ENABLE, 1);
635 
636 			/* enable MCBP */
637 			temp = REG_SET_FIELD(temp, SDMA0_CNTL, MIDCMD_PREEMPT_ENABLE, 1);
638 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), temp);
639 
640 			/* Set up RESP_MODE to non-copy addresses */
641 			temp = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL));
642 			temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, RESP_MODE, 3);
643 			temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, REDO_DELAY, 9);
644 			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL), temp);
645 
646 			/* program default cache read and write policy */
647 			temp = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE));
648 			/* clean read policy and write policy bits */
649 			temp &= 0xFF0FFF;
650 			temp |= ((CACHE_READ_POLICY_L2__DEFAULT << 12) |
651 				 (CACHE_WRITE_POLICY_L2__DEFAULT << 14) |
652 				 SDMA0_UTCL1_PAGE__LLC_NOALLOC_MASK);
653 			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE), temp);
654 
655 			/* unhalt engine */
656 			temp = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
657 			temp = REG_SET_FIELD(temp, SDMA0_F32_CNTL, HALT, 0);
658 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), temp);
659 		}
660 
661 		/* enable DMA RB */
662 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1);
663 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
664 
665 		ib_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL));
666 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 1);
667 #ifdef __BIG_ENDIAN
668 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1);
669 #endif
670 		/* enable DMA IBs */
671 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl);
672 
673 		ring->sched.ready = true;
674 
675 		if (amdgpu_sriov_vf(adev)) { /* bare-metal sequence doesn't need below to lines */
676 			sdma_v5_2_ctx_switch_enable(adev, true);
677 			sdma_v5_2_enable(adev, true);
678 		}
679 
680 		r = amdgpu_ring_test_ring(ring);
681 		if (r) {
682 			ring->sched.ready = false;
683 			return r;
684 		}
685 
686 		if (adev->mman.buffer_funcs_ring == ring)
687 			amdgpu_ttm_set_buffer_funcs_status(adev, true);
688 	}
689 
690 	return 0;
691 }
692 
693 /**
694  * sdma_v5_2_rlc_resume - setup and start the async dma engines
695  *
696  * @adev: amdgpu_device pointer
697  *
698  * Set up the compute DMA queues and enable them.
699  * Returns 0 for success, error for failure.
700  */
701 static int sdma_v5_2_rlc_resume(struct amdgpu_device *adev)
702 {
703 	return 0;
704 }
705 
706 /**
707  * sdma_v5_2_load_microcode - load the sDMA ME ucode
708  *
709  * @adev: amdgpu_device pointer
710  *
711  * Loads the sDMA0/1/2/3 ucode.
712  * Returns 0 for success, -EINVAL if the ucode is not available.
713  */
714 static int sdma_v5_2_load_microcode(struct amdgpu_device *adev)
715 {
716 	const struct sdma_firmware_header_v1_0 *hdr;
717 	const __le32 *fw_data;
718 	u32 fw_size;
719 	int i, j;
720 
721 	/* halt the MEs */
722 	sdma_v5_2_enable(adev, false);
723 
724 	for (i = 0; i < adev->sdma.num_instances; i++) {
725 		if (!adev->sdma.instance[i].fw)
726 			return -EINVAL;
727 
728 		hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
729 		amdgpu_ucode_print_sdma_hdr(&hdr->header);
730 		fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
731 
732 		fw_data = (const __le32 *)
733 			(adev->sdma.instance[i].fw->data +
734 				le32_to_cpu(hdr->header.ucode_array_offset_bytes));
735 
736 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), 0);
737 
738 		for (j = 0; j < fw_size; j++) {
739 			if (amdgpu_emu_mode == 1 && j % 500 == 0)
740 				msleep(1);
741 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UCODE_DATA), le32_to_cpup(fw_data++));
742 		}
743 
744 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), adev->sdma.instance[i].fw_version);
745 	}
746 
747 	return 0;
748 }
749 
750 static int sdma_v5_2_soft_reset(void *handle)
751 {
752 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
753 	u32 grbm_soft_reset;
754 	u32 tmp;
755 	int i;
756 
757 	for (i = 0; i < adev->sdma.num_instances; i++) {
758 		grbm_soft_reset = REG_SET_FIELD(0,
759 						GRBM_SOFT_RESET, SOFT_RESET_SDMA0,
760 						1);
761 		grbm_soft_reset <<= i;
762 
763 		tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET);
764 		tmp |= grbm_soft_reset;
765 		DRM_DEBUG("GRBM_SOFT_RESET=0x%08X\n", tmp);
766 		WREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET, tmp);
767 		tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET);
768 
769 		udelay(50);
770 
771 		tmp &= ~grbm_soft_reset;
772 		WREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET, tmp);
773 		tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET);
774 
775 		udelay(50);
776 	}
777 
778 	return 0;
779 }
780 
781 /**
782  * sdma_v5_2_start - setup and start the async dma engines
783  *
784  * @adev: amdgpu_device pointer
785  *
786  * Set up the DMA engines and enable them.
787  * Returns 0 for success, error for failure.
788  */
789 static int sdma_v5_2_start(struct amdgpu_device *adev)
790 {
791 	int r = 0;
792 
793 	if (amdgpu_sriov_vf(adev)) {
794 		sdma_v5_2_ctx_switch_enable(adev, false);
795 		sdma_v5_2_enable(adev, false);
796 
797 		/* set RB registers */
798 		r = sdma_v5_2_gfx_resume(adev);
799 		return r;
800 	}
801 
802 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
803 		r = sdma_v5_2_load_microcode(adev);
804 		if (r)
805 			return r;
806 
807 		/* The value of mmSDMA_F32_CNTL is invalid the moment after loading fw */
808 		if (amdgpu_emu_mode == 1)
809 			msleep(1000);
810 	}
811 
812 	/* TODO: check whether can submit a doorbell request to raise
813 	 * a doorbell fence to exit gfxoff.
814 	 */
815 	if (adev->in_s0ix)
816 		amdgpu_gfx_off_ctrl(adev, false);
817 
818 	sdma_v5_2_soft_reset(adev);
819 	/* unhalt the MEs */
820 	sdma_v5_2_enable(adev, true);
821 	/* enable sdma ring preemption */
822 	sdma_v5_2_ctx_switch_enable(adev, true);
823 
824 	/* start the gfx rings and rlc compute queues */
825 	r = sdma_v5_2_gfx_resume(adev);
826 	if (adev->in_s0ix)
827 		amdgpu_gfx_off_ctrl(adev, true);
828 	if (r)
829 		return r;
830 	r = sdma_v5_2_rlc_resume(adev);
831 
832 	return r;
833 }
834 
835 static int sdma_v5_2_mqd_init(struct amdgpu_device *adev, void *mqd,
836 			      struct amdgpu_mqd_prop *prop)
837 {
838 	struct v10_sdma_mqd *m = mqd;
839 	uint64_t wb_gpu_addr;
840 
841 	m->sdmax_rlcx_rb_cntl =
842 		order_base_2(prop->queue_size / 4) << SDMA0_RLC0_RB_CNTL__RB_SIZE__SHIFT |
843 		1 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_ENABLE__SHIFT |
844 		6 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_TIMER__SHIFT |
845 		1 << SDMA0_RLC0_RB_CNTL__RB_PRIV__SHIFT;
846 
847 	m->sdmax_rlcx_rb_base = lower_32_bits(prop->hqd_base_gpu_addr >> 8);
848 	m->sdmax_rlcx_rb_base_hi = upper_32_bits(prop->hqd_base_gpu_addr >> 8);
849 
850 	m->sdmax_rlcx_rb_wptr_poll_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, 0,
851 						  mmSDMA0_GFX_RB_WPTR_POLL_CNTL));
852 
853 	wb_gpu_addr = prop->wptr_gpu_addr;
854 	m->sdmax_rlcx_rb_wptr_poll_addr_lo = lower_32_bits(wb_gpu_addr);
855 	m->sdmax_rlcx_rb_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr);
856 
857 	wb_gpu_addr = prop->rptr_gpu_addr;
858 	m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits(wb_gpu_addr);
859 	m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits(wb_gpu_addr);
860 
861 	m->sdmax_rlcx_ib_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, 0,
862 							mmSDMA0_GFX_IB_CNTL));
863 
864 	m->sdmax_rlcx_doorbell_offset =
865 		prop->doorbell_index << SDMA0_RLC0_DOORBELL_OFFSET__OFFSET__SHIFT;
866 
867 	m->sdmax_rlcx_doorbell = REG_SET_FIELD(0, SDMA0_RLC0_DOORBELL, ENABLE, 1);
868 
869 	return 0;
870 }
871 
872 static void sdma_v5_2_set_mqd_funcs(struct amdgpu_device *adev)
873 {
874 	adev->mqds[AMDGPU_HW_IP_DMA].mqd_size = sizeof(struct v10_sdma_mqd);
875 	adev->mqds[AMDGPU_HW_IP_DMA].init_mqd = sdma_v5_2_mqd_init;
876 }
877 
878 /**
879  * sdma_v5_2_ring_test_ring - simple async dma engine test
880  *
881  * @ring: amdgpu_ring structure holding ring information
882  *
883  * Test the DMA engine by writing using it to write an
884  * value to memory.
885  * Returns 0 for success, error for failure.
886  */
887 static int sdma_v5_2_ring_test_ring(struct amdgpu_ring *ring)
888 {
889 	struct amdgpu_device *adev = ring->adev;
890 	unsigned i;
891 	unsigned index;
892 	int r;
893 	u32 tmp;
894 	u64 gpu_addr;
895 	volatile uint32_t *cpu_ptr = NULL;
896 
897 	tmp = 0xCAFEDEAD;
898 
899 	if (ring->is_mes_queue) {
900 		uint32_t offset = 0;
901 		offset = amdgpu_mes_ctx_get_offs(ring,
902 					 AMDGPU_MES_CTX_PADDING_OFFS);
903 		gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
904 		cpu_ptr = amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
905 		*cpu_ptr = tmp;
906 	} else {
907 		r = amdgpu_device_wb_get(adev, &index);
908 		if (r) {
909 			dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
910 			return r;
911 		}
912 
913 		gpu_addr = adev->wb.gpu_addr + (index * 4);
914 		adev->wb.wb[index] = cpu_to_le32(tmp);
915 	}
916 
917 	r = amdgpu_ring_alloc(ring, 20);
918 	if (r) {
919 		DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
920 		amdgpu_device_wb_free(adev, index);
921 		return r;
922 	}
923 
924 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
925 			  SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
926 	amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
927 	amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
928 	amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0));
929 	amdgpu_ring_write(ring, 0xDEADBEEF);
930 	amdgpu_ring_commit(ring);
931 
932 	for (i = 0; i < adev->usec_timeout; i++) {
933 		if (ring->is_mes_queue)
934 			tmp = le32_to_cpu(*cpu_ptr);
935 		else
936 			tmp = le32_to_cpu(adev->wb.wb[index]);
937 		if (tmp == 0xDEADBEEF)
938 			break;
939 		if (amdgpu_emu_mode == 1)
940 			msleep(1);
941 		else
942 			udelay(1);
943 	}
944 
945 	if (i >= adev->usec_timeout)
946 		r = -ETIMEDOUT;
947 
948 	if (!ring->is_mes_queue)
949 		amdgpu_device_wb_free(adev, index);
950 
951 	return r;
952 }
953 
954 /**
955  * sdma_v5_2_ring_test_ib - test an IB on the DMA engine
956  *
957  * @ring: amdgpu_ring structure holding ring information
958  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
959  *
960  * Test a simple IB in the DMA ring.
961  * Returns 0 on success, error on failure.
962  */
963 static int sdma_v5_2_ring_test_ib(struct amdgpu_ring *ring, long timeout)
964 {
965 	struct amdgpu_device *adev = ring->adev;
966 	struct amdgpu_ib ib;
967 	struct dma_fence *f = NULL;
968 	unsigned index;
969 	long r;
970 	u32 tmp = 0;
971 	u64 gpu_addr;
972 	volatile uint32_t *cpu_ptr = NULL;
973 
974 	tmp = 0xCAFEDEAD;
975 	memset(&ib, 0, sizeof(ib));
976 
977 	if (ring->is_mes_queue) {
978 		uint32_t offset = 0;
979 		offset = amdgpu_mes_ctx_get_offs(ring, AMDGPU_MES_CTX_IB_OFFS);
980 		ib.gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
981 		ib.ptr = (void *)amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
982 
983 		offset = amdgpu_mes_ctx_get_offs(ring,
984 					 AMDGPU_MES_CTX_PADDING_OFFS);
985 		gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
986 		cpu_ptr = amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
987 		*cpu_ptr = tmp;
988 	} else {
989 		r = amdgpu_device_wb_get(adev, &index);
990 		if (r) {
991 			dev_err(adev->dev, "(%ld) failed to allocate wb slot\n", r);
992 			return r;
993 		}
994 
995 		gpu_addr = adev->wb.gpu_addr + (index * 4);
996 		adev->wb.wb[index] = cpu_to_le32(tmp);
997 
998 		r = amdgpu_ib_get(adev, NULL, 256, AMDGPU_IB_POOL_DIRECT, &ib);
999 		if (r) {
1000 			DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r);
1001 			goto err0;
1002 		}
1003 	}
1004 
1005 	ib.ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
1006 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1007 	ib.ptr[1] = lower_32_bits(gpu_addr);
1008 	ib.ptr[2] = upper_32_bits(gpu_addr);
1009 	ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0);
1010 	ib.ptr[4] = 0xDEADBEEF;
1011 	ib.ptr[5] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1012 	ib.ptr[6] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1013 	ib.ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1014 	ib.length_dw = 8;
1015 
1016 	r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
1017 	if (r)
1018 		goto err1;
1019 
1020 	r = dma_fence_wait_timeout(f, false, timeout);
1021 	if (r == 0) {
1022 		DRM_ERROR("amdgpu: IB test timed out\n");
1023 		r = -ETIMEDOUT;
1024 		goto err1;
1025 	} else if (r < 0) {
1026 		DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
1027 		goto err1;
1028 	}
1029 
1030 	if (ring->is_mes_queue)
1031 		tmp = le32_to_cpu(*cpu_ptr);
1032 	else
1033 		tmp = le32_to_cpu(adev->wb.wb[index]);
1034 
1035 	if (tmp == 0xDEADBEEF)
1036 		r = 0;
1037 	else
1038 		r = -EINVAL;
1039 
1040 err1:
1041 	amdgpu_ib_free(adev, &ib, NULL);
1042 	dma_fence_put(f);
1043 err0:
1044 	if (!ring->is_mes_queue)
1045 		amdgpu_device_wb_free(adev, index);
1046 	return r;
1047 }
1048 
1049 
1050 /**
1051  * sdma_v5_2_vm_copy_pte - update PTEs by copying them from the GART
1052  *
1053  * @ib: indirect buffer to fill with commands
1054  * @pe: addr of the page entry
1055  * @src: src addr to copy from
1056  * @count: number of page entries to update
1057  *
1058  * Update PTEs by copying them from the GART using sDMA.
1059  */
1060 static void sdma_v5_2_vm_copy_pte(struct amdgpu_ib *ib,
1061 				  uint64_t pe, uint64_t src,
1062 				  unsigned count)
1063 {
1064 	unsigned bytes = count * 8;
1065 
1066 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1067 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
1068 	ib->ptr[ib->length_dw++] = bytes - 1;
1069 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1070 	ib->ptr[ib->length_dw++] = lower_32_bits(src);
1071 	ib->ptr[ib->length_dw++] = upper_32_bits(src);
1072 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1073 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1074 
1075 }
1076 
1077 /**
1078  * sdma_v5_2_vm_write_pte - update PTEs by writing them manually
1079  *
1080  * @ib: indirect buffer to fill with commands
1081  * @pe: addr of the page entry
1082  * @value: dst addr to write into pe
1083  * @count: number of page entries to update
1084  * @incr: increase next addr by incr bytes
1085  *
1086  * Update PTEs by writing them manually using sDMA.
1087  */
1088 static void sdma_v5_2_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
1089 				   uint64_t value, unsigned count,
1090 				   uint32_t incr)
1091 {
1092 	unsigned ndw = count * 2;
1093 
1094 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
1095 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1096 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1097 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1098 	ib->ptr[ib->length_dw++] = ndw - 1;
1099 	for (; ndw > 0; ndw -= 2) {
1100 		ib->ptr[ib->length_dw++] = lower_32_bits(value);
1101 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
1102 		value += incr;
1103 	}
1104 }
1105 
1106 /**
1107  * sdma_v5_2_vm_set_pte_pde - update the page tables using sDMA
1108  *
1109  * @ib: indirect buffer to fill with commands
1110  * @pe: addr of the page entry
1111  * @addr: dst addr to write into pe
1112  * @count: number of page entries to update
1113  * @incr: increase next addr by incr bytes
1114  * @flags: access flags
1115  *
1116  * Update the page tables using sDMA.
1117  */
1118 static void sdma_v5_2_vm_set_pte_pde(struct amdgpu_ib *ib,
1119 				     uint64_t pe,
1120 				     uint64_t addr, unsigned count,
1121 				     uint32_t incr, uint64_t flags)
1122 {
1123 	/* for physically contiguous pages (vram) */
1124 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_PTEPDE);
1125 	ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
1126 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1127 	ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
1128 	ib->ptr[ib->length_dw++] = upper_32_bits(flags);
1129 	ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
1130 	ib->ptr[ib->length_dw++] = upper_32_bits(addr);
1131 	ib->ptr[ib->length_dw++] = incr; /* increment size */
1132 	ib->ptr[ib->length_dw++] = 0;
1133 	ib->ptr[ib->length_dw++] = count - 1; /* number of entries */
1134 }
1135 
1136 /**
1137  * sdma_v5_2_ring_pad_ib - pad the IB
1138  *
1139  * @ib: indirect buffer to fill with padding
1140  * @ring: amdgpu_ring structure holding ring information
1141  *
1142  * Pad the IB with NOPs to a boundary multiple of 8.
1143  */
1144 static void sdma_v5_2_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
1145 {
1146 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
1147 	u32 pad_count;
1148 	int i;
1149 
1150 	pad_count = (-ib->length_dw) & 0x7;
1151 	for (i = 0; i < pad_count; i++)
1152 		if (sdma && sdma->burst_nop && (i == 0))
1153 			ib->ptr[ib->length_dw++] =
1154 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP) |
1155 				SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
1156 		else
1157 			ib->ptr[ib->length_dw++] =
1158 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
1159 }
1160 
1161 
1162 /**
1163  * sdma_v5_2_ring_emit_pipeline_sync - sync the pipeline
1164  *
1165  * @ring: amdgpu_ring pointer
1166  *
1167  * Make sure all previous operations are completed (CIK).
1168  */
1169 static void sdma_v5_2_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
1170 {
1171 	uint32_t seq = ring->fence_drv.sync_seq;
1172 	uint64_t addr = ring->fence_drv.gpu_addr;
1173 
1174 	/* wait for idle */
1175 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1176 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1177 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
1178 			  SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1));
1179 	amdgpu_ring_write(ring, addr & 0xfffffffc);
1180 	amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
1181 	amdgpu_ring_write(ring, seq); /* reference */
1182 	amdgpu_ring_write(ring, 0xffffffff); /* mask */
1183 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1184 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */
1185 }
1186 
1187 
1188 /**
1189  * sdma_v5_2_ring_emit_vm_flush - vm flush using sDMA
1190  *
1191  * @ring: amdgpu_ring pointer
1192  * @vmid: vmid number to use
1193  * @pd_addr: address
1194  *
1195  * Update the page table base and flush the VM TLB
1196  * using sDMA.
1197  */
1198 static void sdma_v5_2_ring_emit_vm_flush(struct amdgpu_ring *ring,
1199 					 unsigned vmid, uint64_t pd_addr)
1200 {
1201 	amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
1202 }
1203 
1204 static void sdma_v5_2_ring_emit_wreg(struct amdgpu_ring *ring,
1205 				     uint32_t reg, uint32_t val)
1206 {
1207 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
1208 			  SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
1209 	amdgpu_ring_write(ring, reg);
1210 	amdgpu_ring_write(ring, val);
1211 }
1212 
1213 static void sdma_v5_2_ring_emit_reg_wait(struct amdgpu_ring *ring, uint32_t reg,
1214 					 uint32_t val, uint32_t mask)
1215 {
1216 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1217 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1218 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* equal */
1219 	amdgpu_ring_write(ring, reg << 2);
1220 	amdgpu_ring_write(ring, 0);
1221 	amdgpu_ring_write(ring, val); /* reference */
1222 	amdgpu_ring_write(ring, mask); /* mask */
1223 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1224 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10));
1225 }
1226 
1227 static void sdma_v5_2_ring_emit_reg_write_reg_wait(struct amdgpu_ring *ring,
1228 						   uint32_t reg0, uint32_t reg1,
1229 						   uint32_t ref, uint32_t mask)
1230 {
1231 	amdgpu_ring_emit_wreg(ring, reg0, ref);
1232 	/* wait for a cycle to reset vm_inv_eng*_ack */
1233 	amdgpu_ring_emit_reg_wait(ring, reg0, 0, 0);
1234 	amdgpu_ring_emit_reg_wait(ring, reg1, mask, mask);
1235 }
1236 
1237 static int sdma_v5_2_early_init(void *handle)
1238 {
1239 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1240 
1241 	sdma_v5_2_set_ring_funcs(adev);
1242 	sdma_v5_2_set_buffer_funcs(adev);
1243 	sdma_v5_2_set_vm_pte_funcs(adev);
1244 	sdma_v5_2_set_irq_funcs(adev);
1245 	sdma_v5_2_set_mqd_funcs(adev);
1246 
1247 	return 0;
1248 }
1249 
1250 static unsigned sdma_v5_2_seq_to_irq_id(int seq_num)
1251 {
1252 	switch (seq_num) {
1253 	case 0:
1254 		return SOC15_IH_CLIENTID_SDMA0;
1255 	case 1:
1256 		return SOC15_IH_CLIENTID_SDMA1;
1257 	case 2:
1258 		return SOC15_IH_CLIENTID_SDMA2;
1259 	case 3:
1260 		return SOC15_IH_CLIENTID_SDMA3_Sienna_Cichlid;
1261 	default:
1262 		break;
1263 	}
1264 	return -EINVAL;
1265 }
1266 
1267 static unsigned sdma_v5_2_seq_to_trap_id(int seq_num)
1268 {
1269 	switch (seq_num) {
1270 	case 0:
1271 		return SDMA0_5_0__SRCID__SDMA_TRAP;
1272 	case 1:
1273 		return SDMA1_5_0__SRCID__SDMA_TRAP;
1274 	case 2:
1275 		return SDMA2_5_0__SRCID__SDMA_TRAP;
1276 	case 3:
1277 		return SDMA3_5_0__SRCID__SDMA_TRAP;
1278 	default:
1279 		break;
1280 	}
1281 	return -EINVAL;
1282 }
1283 
1284 static int sdma_v5_2_sw_init(void *handle)
1285 {
1286 	struct amdgpu_ring *ring;
1287 	int r, i;
1288 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1289 
1290 	/* SDMA trap event */
1291 	for (i = 0; i < adev->sdma.num_instances; i++) {
1292 		r = amdgpu_irq_add_id(adev, sdma_v5_2_seq_to_irq_id(i),
1293 				      sdma_v5_2_seq_to_trap_id(i),
1294 				      &adev->sdma.trap_irq);
1295 		if (r)
1296 			return r;
1297 	}
1298 
1299 	r = sdma_v5_2_init_microcode(adev);
1300 	if (r) {
1301 		DRM_ERROR("Failed to load sdma firmware!\n");
1302 		return r;
1303 	}
1304 
1305 	for (i = 0; i < adev->sdma.num_instances; i++) {
1306 		ring = &adev->sdma.instance[i].ring;
1307 		ring->ring_obj = NULL;
1308 		ring->use_doorbell = true;
1309 		ring->me = i;
1310 
1311 		DRM_INFO("use_doorbell being set to: [%s]\n",
1312 				ring->use_doorbell?"true":"false");
1313 
1314 		ring->doorbell_index =
1315 			(adev->doorbell_index.sdma_engine[i] << 1); //get DWORD offset
1316 
1317 		sprintf(ring->name, "sdma%d", i);
1318 		r = amdgpu_ring_init(adev, ring, 1024, &adev->sdma.trap_irq,
1319 				     AMDGPU_SDMA_IRQ_INSTANCE0 + i,
1320 				     AMDGPU_RING_PRIO_DEFAULT, NULL);
1321 		if (r)
1322 			return r;
1323 	}
1324 
1325 	return r;
1326 }
1327 
1328 static int sdma_v5_2_sw_fini(void *handle)
1329 {
1330 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1331 	int i;
1332 
1333 	for (i = 0; i < adev->sdma.num_instances; i++)
1334 		amdgpu_ring_fini(&adev->sdma.instance[i].ring);
1335 
1336 	amdgpu_sdma_destroy_inst_ctx(adev, true);
1337 
1338 	return 0;
1339 }
1340 
1341 static int sdma_v5_2_hw_init(void *handle)
1342 {
1343 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1344 
1345 	return sdma_v5_2_start(adev);
1346 }
1347 
1348 static int sdma_v5_2_hw_fini(void *handle)
1349 {
1350 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1351 
1352 	if (amdgpu_sriov_vf(adev)) {
1353 		/* disable the scheduler for SDMA */
1354 		amdgpu_sdma_unset_buffer_funcs_helper(adev);
1355 		return 0;
1356 	}
1357 
1358 	sdma_v5_2_ctx_switch_enable(adev, false);
1359 	sdma_v5_2_enable(adev, false);
1360 
1361 	return 0;
1362 }
1363 
1364 static int sdma_v5_2_suspend(void *handle)
1365 {
1366 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1367 
1368 	return sdma_v5_2_hw_fini(adev);
1369 }
1370 
1371 static int sdma_v5_2_resume(void *handle)
1372 {
1373 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1374 
1375 	return sdma_v5_2_hw_init(adev);
1376 }
1377 
1378 static bool sdma_v5_2_is_idle(void *handle)
1379 {
1380 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1381 	u32 i;
1382 
1383 	for (i = 0; i < adev->sdma.num_instances; i++) {
1384 		u32 tmp = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_STATUS_REG));
1385 
1386 		if (!(tmp & SDMA0_STATUS_REG__IDLE_MASK))
1387 			return false;
1388 	}
1389 
1390 	return true;
1391 }
1392 
1393 static int sdma_v5_2_wait_for_idle(void *handle)
1394 {
1395 	unsigned i;
1396 	u32 sdma0, sdma1, sdma2, sdma3;
1397 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1398 
1399 	for (i = 0; i < adev->usec_timeout; i++) {
1400 		sdma0 = RREG32(sdma_v5_2_get_reg_offset(adev, 0, mmSDMA0_STATUS_REG));
1401 		sdma1 = RREG32(sdma_v5_2_get_reg_offset(adev, 1, mmSDMA0_STATUS_REG));
1402 		sdma2 = RREG32(sdma_v5_2_get_reg_offset(adev, 2, mmSDMA0_STATUS_REG));
1403 		sdma3 = RREG32(sdma_v5_2_get_reg_offset(adev, 3, mmSDMA0_STATUS_REG));
1404 
1405 		if (sdma0 & sdma1 & sdma2 & sdma3 & SDMA0_STATUS_REG__IDLE_MASK)
1406 			return 0;
1407 		udelay(1);
1408 	}
1409 	return -ETIMEDOUT;
1410 }
1411 
1412 static int sdma_v5_2_ring_preempt_ib(struct amdgpu_ring *ring)
1413 {
1414 	int i, r = 0;
1415 	struct amdgpu_device *adev = ring->adev;
1416 	u32 index = 0;
1417 	u64 sdma_gfx_preempt;
1418 
1419 	amdgpu_sdma_get_index_from_ring(ring, &index);
1420 	sdma_gfx_preempt =
1421 		sdma_v5_2_get_reg_offset(adev, index, mmSDMA0_GFX_PREEMPT);
1422 
1423 	/* assert preemption condition */
1424 	amdgpu_ring_set_preempt_cond_exec(ring, false);
1425 
1426 	/* emit the trailing fence */
1427 	ring->trail_seq += 1;
1428 	amdgpu_ring_alloc(ring, 10);
1429 	sdma_v5_2_ring_emit_fence(ring, ring->trail_fence_gpu_addr,
1430 				  ring->trail_seq, 0);
1431 	amdgpu_ring_commit(ring);
1432 
1433 	/* assert IB preemption */
1434 	WREG32(sdma_gfx_preempt, 1);
1435 
1436 	/* poll the trailing fence */
1437 	for (i = 0; i < adev->usec_timeout; i++) {
1438 		if (ring->trail_seq ==
1439 		    le32_to_cpu(*(ring->trail_fence_cpu_addr)))
1440 			break;
1441 		udelay(1);
1442 	}
1443 
1444 	if (i >= adev->usec_timeout) {
1445 		r = -EINVAL;
1446 		DRM_ERROR("ring %d failed to be preempted\n", ring->idx);
1447 	}
1448 
1449 	/* deassert IB preemption */
1450 	WREG32(sdma_gfx_preempt, 0);
1451 
1452 	/* deassert the preemption condition */
1453 	amdgpu_ring_set_preempt_cond_exec(ring, true);
1454 	return r;
1455 }
1456 
1457 static int sdma_v5_2_set_trap_irq_state(struct amdgpu_device *adev,
1458 					struct amdgpu_irq_src *source,
1459 					unsigned type,
1460 					enum amdgpu_interrupt_state state)
1461 {
1462 	u32 sdma_cntl;
1463 	u32 reg_offset = sdma_v5_2_get_reg_offset(adev, type, mmSDMA0_CNTL);
1464 
1465 	if (!amdgpu_sriov_vf(adev)) {
1466 		sdma_cntl = RREG32(reg_offset);
1467 		sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE,
1468 			       state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0);
1469 		WREG32(reg_offset, sdma_cntl);
1470 	}
1471 
1472 	return 0;
1473 }
1474 
1475 static int sdma_v5_2_process_trap_irq(struct amdgpu_device *adev,
1476 				      struct amdgpu_irq_src *source,
1477 				      struct amdgpu_iv_entry *entry)
1478 {
1479 	uint32_t mes_queue_id = entry->src_data[0];
1480 
1481 	DRM_DEBUG("IH: SDMA trap\n");
1482 
1483 	if (adev->enable_mes && (mes_queue_id & AMDGPU_FENCE_MES_QUEUE_FLAG)) {
1484 		struct amdgpu_mes_queue *queue;
1485 
1486 		mes_queue_id &= AMDGPU_FENCE_MES_QUEUE_ID_MASK;
1487 
1488 		spin_lock(&adev->mes.queue_id_lock);
1489 		queue = idr_find(&adev->mes.queue_id_idr, mes_queue_id);
1490 		if (queue) {
1491 			DRM_DEBUG("process smda queue id = %d\n", mes_queue_id);
1492 			amdgpu_fence_process(queue->ring);
1493 		}
1494 		spin_unlock(&adev->mes.queue_id_lock);
1495 		return 0;
1496 	}
1497 
1498 	switch (entry->client_id) {
1499 	case SOC15_IH_CLIENTID_SDMA0:
1500 		switch (entry->ring_id) {
1501 		case 0:
1502 			amdgpu_fence_process(&adev->sdma.instance[0].ring);
1503 			break;
1504 		case 1:
1505 			/* XXX compute */
1506 			break;
1507 		case 2:
1508 			/* XXX compute */
1509 			break;
1510 		case 3:
1511 			/* XXX page queue*/
1512 			break;
1513 		}
1514 		break;
1515 	case SOC15_IH_CLIENTID_SDMA1:
1516 		switch (entry->ring_id) {
1517 		case 0:
1518 			amdgpu_fence_process(&adev->sdma.instance[1].ring);
1519 			break;
1520 		case 1:
1521 			/* XXX compute */
1522 			break;
1523 		case 2:
1524 			/* XXX compute */
1525 			break;
1526 		case 3:
1527 			/* XXX page queue*/
1528 			break;
1529 		}
1530 		break;
1531 	case SOC15_IH_CLIENTID_SDMA2:
1532 		switch (entry->ring_id) {
1533 		case 0:
1534 			amdgpu_fence_process(&adev->sdma.instance[2].ring);
1535 			break;
1536 		case 1:
1537 			/* XXX compute */
1538 			break;
1539 		case 2:
1540 			/* XXX compute */
1541 			break;
1542 		case 3:
1543 			/* XXX page queue*/
1544 			break;
1545 		}
1546 		break;
1547 	case SOC15_IH_CLIENTID_SDMA3_Sienna_Cichlid:
1548 		switch (entry->ring_id) {
1549 		case 0:
1550 			amdgpu_fence_process(&adev->sdma.instance[3].ring);
1551 			break;
1552 		case 1:
1553 			/* XXX compute */
1554 			break;
1555 		case 2:
1556 			/* XXX compute */
1557 			break;
1558 		case 3:
1559 			/* XXX page queue*/
1560 			break;
1561 		}
1562 		break;
1563 	}
1564 	return 0;
1565 }
1566 
1567 static int sdma_v5_2_process_illegal_inst_irq(struct amdgpu_device *adev,
1568 					      struct amdgpu_irq_src *source,
1569 					      struct amdgpu_iv_entry *entry)
1570 {
1571 	return 0;
1572 }
1573 
1574 static void sdma_v5_2_update_medium_grain_clock_gating(struct amdgpu_device *adev,
1575 						       bool enable)
1576 {
1577 	uint32_t data, def;
1578 	int i;
1579 
1580 	for (i = 0; i < adev->sdma.num_instances; i++) {
1581 
1582 		if (adev->sdma.instance[i].fw_version < 70 && adev->ip_versions[SDMA0_HWIP][0] == IP_VERSION(5, 2, 1))
1583 			adev->cg_flags &= ~AMD_CG_SUPPORT_SDMA_MGCG;
1584 
1585 		if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
1586 			/* Enable sdma clock gating */
1587 			def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL));
1588 			data &= ~(SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1589 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1590 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1591 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1592 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK |
1593 				  SDMA0_CLK_CTRL__SOFT_OVERRIDER_REG_MASK);
1594 			if (def != data)
1595 				WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data);
1596 		} else {
1597 			/* Disable sdma clock gating */
1598 			def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL));
1599 			data |= (SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1600 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1601 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1602 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1603 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK |
1604 				 SDMA0_CLK_CTRL__SOFT_OVERRIDER_REG_MASK);
1605 			if (def != data)
1606 				WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data);
1607 		}
1608 	}
1609 }
1610 
1611 static void sdma_v5_2_update_medium_grain_light_sleep(struct amdgpu_device *adev,
1612 						      bool enable)
1613 {
1614 	uint32_t data, def;
1615 	int i;
1616 
1617 	for (i = 0; i < adev->sdma.num_instances; i++) {
1618 
1619 		if (adev->sdma.instance[i].fw_version < 70 && adev->ip_versions[SDMA0_HWIP][0] == IP_VERSION(5, 2, 1))
1620 			adev->cg_flags &= ~AMD_CG_SUPPORT_SDMA_LS;
1621 
1622 		if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) {
1623 			/* Enable sdma mem light sleep */
1624 			def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL));
1625 			data |= SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1626 			if (def != data)
1627 				WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data);
1628 
1629 		} else {
1630 			/* Disable sdma mem light sleep */
1631 			def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL));
1632 			data &= ~SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1633 			if (def != data)
1634 				WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data);
1635 
1636 		}
1637 	}
1638 }
1639 
1640 static int sdma_v5_2_set_clockgating_state(void *handle,
1641 					   enum amd_clockgating_state state)
1642 {
1643 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1644 
1645 	if (amdgpu_sriov_vf(adev))
1646 		return 0;
1647 
1648 	switch (adev->ip_versions[SDMA0_HWIP][0]) {
1649 	case IP_VERSION(5, 2, 0):
1650 	case IP_VERSION(5, 2, 2):
1651 	case IP_VERSION(5, 2, 1):
1652 	case IP_VERSION(5, 2, 4):
1653 	case IP_VERSION(5, 2, 5):
1654 	case IP_VERSION(5, 2, 6):
1655 	case IP_VERSION(5, 2, 3):
1656 		sdma_v5_2_update_medium_grain_clock_gating(adev,
1657 				state == AMD_CG_STATE_GATE);
1658 		sdma_v5_2_update_medium_grain_light_sleep(adev,
1659 				state == AMD_CG_STATE_GATE);
1660 		break;
1661 	default:
1662 		break;
1663 	}
1664 
1665 	return 0;
1666 }
1667 
1668 static int sdma_v5_2_set_powergating_state(void *handle,
1669 					  enum amd_powergating_state state)
1670 {
1671 	return 0;
1672 }
1673 
1674 static void sdma_v5_2_get_clockgating_state(void *handle, u64 *flags)
1675 {
1676 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1677 	int data;
1678 
1679 	if (amdgpu_sriov_vf(adev))
1680 		*flags = 0;
1681 
1682 	/* AMD_CG_SUPPORT_SDMA_MGCG */
1683 	data = RREG32(sdma_v5_2_get_reg_offset(adev, 0, mmSDMA0_CLK_CTRL));
1684 	if (!(data & SDMA0_CLK_CTRL__CGCG_EN_OVERRIDE_MASK))
1685 		*flags |= AMD_CG_SUPPORT_SDMA_MGCG;
1686 
1687 	/* AMD_CG_SUPPORT_SDMA_LS */
1688 	data = RREG32_KIQ(sdma_v5_2_get_reg_offset(adev, 0, mmSDMA0_POWER_CNTL));
1689 	if (data & SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK)
1690 		*flags |= AMD_CG_SUPPORT_SDMA_LS;
1691 }
1692 
1693 const struct amd_ip_funcs sdma_v5_2_ip_funcs = {
1694 	.name = "sdma_v5_2",
1695 	.early_init = sdma_v5_2_early_init,
1696 	.late_init = NULL,
1697 	.sw_init = sdma_v5_2_sw_init,
1698 	.sw_fini = sdma_v5_2_sw_fini,
1699 	.hw_init = sdma_v5_2_hw_init,
1700 	.hw_fini = sdma_v5_2_hw_fini,
1701 	.suspend = sdma_v5_2_suspend,
1702 	.resume = sdma_v5_2_resume,
1703 	.is_idle = sdma_v5_2_is_idle,
1704 	.wait_for_idle = sdma_v5_2_wait_for_idle,
1705 	.soft_reset = sdma_v5_2_soft_reset,
1706 	.set_clockgating_state = sdma_v5_2_set_clockgating_state,
1707 	.set_powergating_state = sdma_v5_2_set_powergating_state,
1708 	.get_clockgating_state = sdma_v5_2_get_clockgating_state,
1709 };
1710 
1711 static const struct amdgpu_ring_funcs sdma_v5_2_ring_funcs = {
1712 	.type = AMDGPU_RING_TYPE_SDMA,
1713 	.align_mask = 0xf,
1714 	.nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP),
1715 	.support_64bit_ptrs = true,
1716 	.secure_submission_supported = true,
1717 	.vmhub = AMDGPU_GFXHUB_0,
1718 	.get_rptr = sdma_v5_2_ring_get_rptr,
1719 	.get_wptr = sdma_v5_2_ring_get_wptr,
1720 	.set_wptr = sdma_v5_2_ring_set_wptr,
1721 	.emit_frame_size =
1722 		5 + /* sdma_v5_2_ring_init_cond_exec */
1723 		6 + /* sdma_v5_2_ring_emit_hdp_flush */
1724 		3 + /* hdp_invalidate */
1725 		6 + /* sdma_v5_2_ring_emit_pipeline_sync */
1726 		/* sdma_v5_2_ring_emit_vm_flush */
1727 		SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 +
1728 		SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6 +
1729 		10 + 10 + 10, /* sdma_v5_2_ring_emit_fence x3 for user fence, vm fence */
1730 	.emit_ib_size = 7 + 6, /* sdma_v5_2_ring_emit_ib */
1731 	.emit_ib = sdma_v5_2_ring_emit_ib,
1732 	.emit_mem_sync = sdma_v5_2_ring_emit_mem_sync,
1733 	.emit_fence = sdma_v5_2_ring_emit_fence,
1734 	.emit_pipeline_sync = sdma_v5_2_ring_emit_pipeline_sync,
1735 	.emit_vm_flush = sdma_v5_2_ring_emit_vm_flush,
1736 	.emit_hdp_flush = sdma_v5_2_ring_emit_hdp_flush,
1737 	.test_ring = sdma_v5_2_ring_test_ring,
1738 	.test_ib = sdma_v5_2_ring_test_ib,
1739 	.insert_nop = sdma_v5_2_ring_insert_nop,
1740 	.pad_ib = sdma_v5_2_ring_pad_ib,
1741 	.emit_wreg = sdma_v5_2_ring_emit_wreg,
1742 	.emit_reg_wait = sdma_v5_2_ring_emit_reg_wait,
1743 	.emit_reg_write_reg_wait = sdma_v5_2_ring_emit_reg_write_reg_wait,
1744 	.init_cond_exec = sdma_v5_2_ring_init_cond_exec,
1745 	.patch_cond_exec = sdma_v5_2_ring_patch_cond_exec,
1746 	.preempt_ib = sdma_v5_2_ring_preempt_ib,
1747 };
1748 
1749 static void sdma_v5_2_set_ring_funcs(struct amdgpu_device *adev)
1750 {
1751 	int i;
1752 
1753 	for (i = 0; i < adev->sdma.num_instances; i++) {
1754 		adev->sdma.instance[i].ring.funcs = &sdma_v5_2_ring_funcs;
1755 		adev->sdma.instance[i].ring.me = i;
1756 	}
1757 }
1758 
1759 static const struct amdgpu_irq_src_funcs sdma_v5_2_trap_irq_funcs = {
1760 	.set = sdma_v5_2_set_trap_irq_state,
1761 	.process = sdma_v5_2_process_trap_irq,
1762 };
1763 
1764 static const struct amdgpu_irq_src_funcs sdma_v5_2_illegal_inst_irq_funcs = {
1765 	.process = sdma_v5_2_process_illegal_inst_irq,
1766 };
1767 
1768 static void sdma_v5_2_set_irq_funcs(struct amdgpu_device *adev)
1769 {
1770 	adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_INSTANCE0 +
1771 					adev->sdma.num_instances;
1772 	adev->sdma.trap_irq.funcs = &sdma_v5_2_trap_irq_funcs;
1773 	adev->sdma.illegal_inst_irq.funcs = &sdma_v5_2_illegal_inst_irq_funcs;
1774 }
1775 
1776 /**
1777  * sdma_v5_2_emit_copy_buffer - copy buffer using the sDMA engine
1778  *
1779  * @ib: indirect buffer to copy to
1780  * @src_offset: src GPU address
1781  * @dst_offset: dst GPU address
1782  * @byte_count: number of bytes to xfer
1783  * @tmz: if a secure copy should be used
1784  *
1785  * Copy GPU buffers using the DMA engine.
1786  * Used by the amdgpu ttm implementation to move pages if
1787  * registered as the asic copy callback.
1788  */
1789 static void sdma_v5_2_emit_copy_buffer(struct amdgpu_ib *ib,
1790 				       uint64_t src_offset,
1791 				       uint64_t dst_offset,
1792 				       uint32_t byte_count,
1793 				       bool tmz)
1794 {
1795 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1796 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR) |
1797 		SDMA_PKT_COPY_LINEAR_HEADER_TMZ(tmz ? 1 : 0);
1798 	ib->ptr[ib->length_dw++] = byte_count - 1;
1799 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1800 	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1801 	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1802 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1803 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1804 }
1805 
1806 /**
1807  * sdma_v5_2_emit_fill_buffer - fill buffer using the sDMA engine
1808  *
1809  * @ib: indirect buffer to fill
1810  * @src_data: value to write to buffer
1811  * @dst_offset: dst GPU address
1812  * @byte_count: number of bytes to xfer
1813  *
1814  * Fill GPU buffers using the DMA engine.
1815  */
1816 static void sdma_v5_2_emit_fill_buffer(struct amdgpu_ib *ib,
1817 				       uint32_t src_data,
1818 				       uint64_t dst_offset,
1819 				       uint32_t byte_count)
1820 {
1821 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL);
1822 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1823 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1824 	ib->ptr[ib->length_dw++] = src_data;
1825 	ib->ptr[ib->length_dw++] = byte_count - 1;
1826 }
1827 
1828 static const struct amdgpu_buffer_funcs sdma_v5_2_buffer_funcs = {
1829 	.copy_max_bytes = 0x400000,
1830 	.copy_num_dw = 7,
1831 	.emit_copy_buffer = sdma_v5_2_emit_copy_buffer,
1832 
1833 	.fill_max_bytes = 0x400000,
1834 	.fill_num_dw = 5,
1835 	.emit_fill_buffer = sdma_v5_2_emit_fill_buffer,
1836 };
1837 
1838 static void sdma_v5_2_set_buffer_funcs(struct amdgpu_device *adev)
1839 {
1840 	if (adev->mman.buffer_funcs == NULL) {
1841 		adev->mman.buffer_funcs = &sdma_v5_2_buffer_funcs;
1842 		adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1843 	}
1844 }
1845 
1846 static const struct amdgpu_vm_pte_funcs sdma_v5_2_vm_pte_funcs = {
1847 	.copy_pte_num_dw = 7,
1848 	.copy_pte = sdma_v5_2_vm_copy_pte,
1849 	.write_pte = sdma_v5_2_vm_write_pte,
1850 	.set_pte_pde = sdma_v5_2_vm_set_pte_pde,
1851 };
1852 
1853 static void sdma_v5_2_set_vm_pte_funcs(struct amdgpu_device *adev)
1854 {
1855 	unsigned i;
1856 
1857 	if (adev->vm_manager.vm_pte_funcs == NULL) {
1858 		adev->vm_manager.vm_pte_funcs = &sdma_v5_2_vm_pte_funcs;
1859 		for (i = 0; i < adev->sdma.num_instances; i++) {
1860 			adev->vm_manager.vm_pte_scheds[i] =
1861 				&adev->sdma.instance[i].ring.sched;
1862 		}
1863 		adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
1864 	}
1865 }
1866 
1867 const struct amdgpu_ip_block_version sdma_v5_2_ip_block = {
1868 	.type = AMD_IP_BLOCK_TYPE_SDMA,
1869 	.major = 5,
1870 	.minor = 2,
1871 	.rev = 0,
1872 	.funcs = &sdma_v5_2_ip_funcs,
1873 };
1874