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 int sdma_v5_2_init_inst_ctx(struct amdgpu_sdma_instance *sdma_inst)
93 {
94 	int err = 0;
95 	const struct sdma_firmware_header_v1_0 *hdr;
96 
97 	err = amdgpu_ucode_validate(sdma_inst->fw);
98 	if (err)
99 		return err;
100 
101 	hdr = (const struct sdma_firmware_header_v1_0 *)sdma_inst->fw->data;
102 	sdma_inst->fw_version = le32_to_cpu(hdr->header.ucode_version);
103 	sdma_inst->feature_version = le32_to_cpu(hdr->ucode_feature_version);
104 
105 	if (sdma_inst->feature_version >= 20)
106 		sdma_inst->burst_nop = true;
107 
108 	return 0;
109 }
110 
111 static void sdma_v5_2_destroy_inst_ctx(struct amdgpu_device *adev)
112 {
113 	release_firmware(adev->sdma.instance[0].fw);
114 
115 	memset((void *)adev->sdma.instance, 0,
116 	       sizeof(struct amdgpu_sdma_instance) * AMDGPU_MAX_SDMA_INSTANCES);
117 }
118 
119 /**
120  * sdma_v5_2_init_microcode - load ucode images from disk
121  *
122  * @adev: amdgpu_device pointer
123  *
124  * Use the firmware interface to load the ucode images into
125  * the driver (not loaded into hw).
126  * Returns 0 on success, error on failure.
127  */
128 
129 // emulation only, won't work on real chip
130 // navi10 real chip need to use PSP to load firmware
131 static int sdma_v5_2_init_microcode(struct amdgpu_device *adev)
132 {
133 	const char *chip_name;
134 	char fw_name[40];
135 	int err = 0, i;
136 	struct amdgpu_firmware_info *info = NULL;
137 	const struct common_firmware_header *header = NULL;
138 
139 	DRM_DEBUG("\n");
140 
141 	switch (adev->ip_versions[SDMA0_HWIP][0]) {
142 	case IP_VERSION(5, 2, 0):
143 		chip_name = "sienna_cichlid_sdma";
144 		break;
145 	case IP_VERSION(5, 2, 2):
146 		chip_name = "navy_flounder_sdma";
147 		break;
148 	case IP_VERSION(5, 2, 1):
149 		chip_name = "vangogh_sdma";
150 		break;
151 	case IP_VERSION(5, 2, 4):
152 		chip_name = "dimgrey_cavefish_sdma";
153 		break;
154 	case IP_VERSION(5, 2, 5):
155 		chip_name = "beige_goby_sdma";
156 		break;
157 	case IP_VERSION(5, 2, 3):
158 		chip_name = "yellow_carp_sdma";
159 		break;
160 	case IP_VERSION(5, 2, 6):
161 		chip_name = "sdma_5_2_6";
162 		break;
163 	case IP_VERSION(5, 2, 7):
164 		chip_name = "sdma_5_2_7";
165 		break;
166 	default:
167 		BUG();
168 	}
169 
170 	snprintf(fw_name, sizeof(fw_name), "amdgpu/%s.bin", chip_name);
171 
172 	err = request_firmware(&adev->sdma.instance[0].fw, fw_name, adev->dev);
173 	if (err)
174 		goto out;
175 
176 	err = sdma_v5_2_init_inst_ctx(&adev->sdma.instance[0]);
177 	if (err)
178 		goto out;
179 
180 	for (i = 1; i < adev->sdma.num_instances; i++)
181 		memcpy((void *)&adev->sdma.instance[i],
182 		       (void *)&adev->sdma.instance[0],
183 		       sizeof(struct amdgpu_sdma_instance));
184 
185 	if (amdgpu_sriov_vf(adev) && (adev->ip_versions[SDMA0_HWIP][0] == IP_VERSION(5, 2, 0)))
186 		return 0;
187 
188 	DRM_DEBUG("psp_load == '%s'\n",
189 		  adev->firmware.load_type == AMDGPU_FW_LOAD_PSP ? "true" : "false");
190 
191 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
192 		for (i = 0; i < adev->sdma.num_instances; i++) {
193 			info = &adev->firmware.ucode[AMDGPU_UCODE_ID_SDMA0 + i];
194 			info->ucode_id = AMDGPU_UCODE_ID_SDMA0 + i;
195 			info->fw = adev->sdma.instance[i].fw;
196 			header = (const struct common_firmware_header *)info->fw->data;
197 			adev->firmware.fw_size +=
198 				ALIGN(le32_to_cpu(header->ucode_size_bytes), PAGE_SIZE);
199 		}
200 	}
201 
202 out:
203 	if (err) {
204 		DRM_ERROR("sdma_v5_2: Failed to load firmware \"%s\"\n", fw_name);
205 		sdma_v5_2_destroy_inst_ctx(adev);
206 	}
207 	return err;
208 }
209 
210 static unsigned sdma_v5_2_ring_init_cond_exec(struct amdgpu_ring *ring)
211 {
212 	unsigned ret;
213 
214 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_COND_EXE));
215 	amdgpu_ring_write(ring, lower_32_bits(ring->cond_exe_gpu_addr));
216 	amdgpu_ring_write(ring, upper_32_bits(ring->cond_exe_gpu_addr));
217 	amdgpu_ring_write(ring, 1);
218 	ret = ring->wptr & ring->buf_mask;/* this is the offset we need patch later */
219 	amdgpu_ring_write(ring, 0x55aa55aa);/* insert dummy here and patch it later */
220 
221 	return ret;
222 }
223 
224 static void sdma_v5_2_ring_patch_cond_exec(struct amdgpu_ring *ring,
225 					   unsigned offset)
226 {
227 	unsigned cur;
228 
229 	BUG_ON(offset > ring->buf_mask);
230 	BUG_ON(ring->ring[offset] != 0x55aa55aa);
231 
232 	cur = (ring->wptr - 1) & ring->buf_mask;
233 	if (cur > offset)
234 		ring->ring[offset] = cur - offset;
235 	else
236 		ring->ring[offset] = (ring->buf_mask + 1) - offset + cur;
237 }
238 
239 /**
240  * sdma_v5_2_ring_get_rptr - get the current read pointer
241  *
242  * @ring: amdgpu ring pointer
243  *
244  * Get the current rptr from the hardware (NAVI10+).
245  */
246 static uint64_t sdma_v5_2_ring_get_rptr(struct amdgpu_ring *ring)
247 {
248 	u64 *rptr;
249 
250 	/* XXX check if swapping is necessary on BE */
251 	rptr = (u64 *)ring->rptr_cpu_addr;
252 
253 	DRM_DEBUG("rptr before shift == 0x%016llx\n", *rptr);
254 	return ((*rptr) >> 2);
255 }
256 
257 /**
258  * sdma_v5_2_ring_get_wptr - get the current write pointer
259  *
260  * @ring: amdgpu ring pointer
261  *
262  * Get the current wptr from the hardware (NAVI10+).
263  */
264 static uint64_t sdma_v5_2_ring_get_wptr(struct amdgpu_ring *ring)
265 {
266 	struct amdgpu_device *adev = ring->adev;
267 	u64 wptr;
268 
269 	if (ring->use_doorbell) {
270 		/* XXX check if swapping is necessary on BE */
271 		wptr = READ_ONCE(*((u64 *)ring->wptr_cpu_addr));
272 		DRM_DEBUG("wptr/doorbell before shift == 0x%016llx\n", wptr);
273 	} else {
274 		wptr = RREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI));
275 		wptr = wptr << 32;
276 		wptr |= RREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR));
277 		DRM_DEBUG("wptr before shift [%i] wptr == 0x%016llx\n", ring->me, wptr);
278 	}
279 
280 	return wptr >> 2;
281 }
282 
283 /**
284  * sdma_v5_2_ring_set_wptr - commit the write pointer
285  *
286  * @ring: amdgpu ring pointer
287  *
288  * Write the wptr back to the hardware (NAVI10+).
289  */
290 static void sdma_v5_2_ring_set_wptr(struct amdgpu_ring *ring)
291 {
292 	struct amdgpu_device *adev = ring->adev;
293 
294 	DRM_DEBUG("Setting write pointer\n");
295 	if (ring->use_doorbell) {
296 		DRM_DEBUG("Using doorbell -- "
297 				"wptr_offs == 0x%08x "
298 				"lower_32_bits(ring->wptr << 2) == 0x%08x "
299 				"upper_32_bits(ring->wptr << 2) == 0x%08x\n",
300 				ring->wptr_offs,
301 				lower_32_bits(ring->wptr << 2),
302 				upper_32_bits(ring->wptr << 2));
303 		/* XXX check if swapping is necessary on BE */
304 		atomic64_set((atomic64_t *)ring->wptr_cpu_addr,
305 			     ring->wptr << 2);
306 		DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016llx)\n",
307 				ring->doorbell_index, ring->wptr << 2);
308 		WDOORBELL64(ring->doorbell_index, ring->wptr << 2);
309 	} else {
310 		DRM_DEBUG("Not using doorbell -- "
311 				"mmSDMA%i_GFX_RB_WPTR == 0x%08x "
312 				"mmSDMA%i_GFX_RB_WPTR_HI == 0x%08x\n",
313 				ring->me,
314 				lower_32_bits(ring->wptr << 2),
315 				ring->me,
316 				upper_32_bits(ring->wptr << 2));
317 		WREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR),
318 			lower_32_bits(ring->wptr << 2));
319 		WREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI),
320 			upper_32_bits(ring->wptr << 2));
321 	}
322 }
323 
324 static void sdma_v5_2_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
325 {
326 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
327 	int i;
328 
329 	for (i = 0; i < count; i++)
330 		if (sdma && sdma->burst_nop && (i == 0))
331 			amdgpu_ring_write(ring, ring->funcs->nop |
332 				SDMA_PKT_NOP_HEADER_COUNT(count - 1));
333 		else
334 			amdgpu_ring_write(ring, ring->funcs->nop);
335 }
336 
337 /**
338  * sdma_v5_2_ring_emit_ib - Schedule an IB on the DMA engine
339  *
340  * @ring: amdgpu ring pointer
341  * @job: job to retrieve vmid from
342  * @ib: IB object to schedule
343  * @flags: unused
344  *
345  * Schedule an IB in the DMA ring.
346  */
347 static void sdma_v5_2_ring_emit_ib(struct amdgpu_ring *ring,
348 				   struct amdgpu_job *job,
349 				   struct amdgpu_ib *ib,
350 				   uint32_t flags)
351 {
352 	unsigned vmid = AMDGPU_JOB_GET_VMID(job);
353 	uint64_t csa_mc_addr = amdgpu_sdma_get_csa_mc_addr(ring, vmid);
354 
355 	/* An IB packet must end on a 8 DW boundary--the next dword
356 	 * must be on a 8-dword boundary. Our IB packet below is 6
357 	 * dwords long, thus add x number of NOPs, such that, in
358 	 * modular arithmetic,
359 	 * wptr + 6 + x = 8k, k >= 0, which in C is,
360 	 * (wptr + 6 + x) % 8 = 0.
361 	 * The expression below, is a solution of x.
362 	 */
363 	sdma_v5_2_ring_insert_nop(ring, (2 - lower_32_bits(ring->wptr)) & 7);
364 
365 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) |
366 			  SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf));
367 	/* base must be 32 byte aligned */
368 	amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
369 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
370 	amdgpu_ring_write(ring, ib->length_dw);
371 	amdgpu_ring_write(ring, lower_32_bits(csa_mc_addr));
372 	amdgpu_ring_write(ring, upper_32_bits(csa_mc_addr));
373 }
374 
375 /**
376  * sdma_v5_2_ring_emit_mem_sync - flush the IB by graphics cache rinse
377  *
378  * @ring: amdgpu ring pointer
379  *
380  * flush the IB by graphics cache rinse.
381  */
382 static void sdma_v5_2_ring_emit_mem_sync(struct amdgpu_ring *ring)
383 {
384 	uint32_t gcr_cntl = SDMA_GCR_GL2_INV | SDMA_GCR_GL2_WB |
385 			    SDMA_GCR_GLM_INV | SDMA_GCR_GL1_INV |
386 			    SDMA_GCR_GLV_INV | SDMA_GCR_GLK_INV |
387 			    SDMA_GCR_GLI_INV(1);
388 
389 	/* flush entire cache L0/L1/L2, this can be optimized by performance requirement */
390 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_GCR_REQ));
391 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD1_BASE_VA_31_7(0));
392 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD2_GCR_CONTROL_15_0(gcr_cntl) |
393 			SDMA_PKT_GCR_REQ_PAYLOAD2_BASE_VA_47_32(0));
394 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD3_LIMIT_VA_31_7(0) |
395 			SDMA_PKT_GCR_REQ_PAYLOAD3_GCR_CONTROL_18_16(gcr_cntl >> 16));
396 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD4_LIMIT_VA_47_32(0) |
397 			SDMA_PKT_GCR_REQ_PAYLOAD4_VMID(0));
398 }
399 
400 /**
401  * sdma_v5_2_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
402  *
403  * @ring: amdgpu ring pointer
404  *
405  * Emit an hdp flush packet on the requested DMA ring.
406  */
407 static void sdma_v5_2_ring_emit_hdp_flush(struct amdgpu_ring *ring)
408 {
409 	struct amdgpu_device *adev = ring->adev;
410 	u32 ref_and_mask = 0;
411 	const struct nbio_hdp_flush_reg *nbio_hf_reg = adev->nbio.hdp_flush_reg;
412 
413 	ref_and_mask = nbio_hf_reg->ref_and_mask_sdma0 << ring->me;
414 
415 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
416 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) |
417 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
418 	amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_done_offset(adev)) << 2);
419 	amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_req_offset(adev)) << 2);
420 	amdgpu_ring_write(ring, ref_and_mask); /* reference */
421 	amdgpu_ring_write(ring, ref_and_mask); /* mask */
422 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
423 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
424 }
425 
426 /**
427  * sdma_v5_2_ring_emit_fence - emit a fence on the DMA ring
428  *
429  * @ring: amdgpu ring pointer
430  * @addr: address
431  * @seq: sequence number
432  * @flags: fence related flags
433  *
434  * Add a DMA fence packet to the ring to write
435  * the fence seq number and DMA trap packet to generate
436  * an interrupt if needed.
437  */
438 static void sdma_v5_2_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
439 				      unsigned flags)
440 {
441 	bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
442 	/* write the fence */
443 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) |
444 			  SDMA_PKT_FENCE_HEADER_MTYPE(0x3)); /* Ucached(UC) */
445 	/* zero in first two bits */
446 	BUG_ON(addr & 0x3);
447 	amdgpu_ring_write(ring, lower_32_bits(addr));
448 	amdgpu_ring_write(ring, upper_32_bits(addr));
449 	amdgpu_ring_write(ring, lower_32_bits(seq));
450 
451 	/* optionally write high bits as well */
452 	if (write64bit) {
453 		addr += 4;
454 		amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) |
455 				  SDMA_PKT_FENCE_HEADER_MTYPE(0x3));
456 		/* zero in first two bits */
457 		BUG_ON(addr & 0x3);
458 		amdgpu_ring_write(ring, lower_32_bits(addr));
459 		amdgpu_ring_write(ring, upper_32_bits(addr));
460 		amdgpu_ring_write(ring, upper_32_bits(seq));
461 	}
462 
463 	if ((flags & AMDGPU_FENCE_FLAG_INT)) {
464 		uint32_t ctx = ring->is_mes_queue ?
465 			(ring->hw_queue_id | AMDGPU_FENCE_MES_QUEUE_FLAG) : 0;
466 		/* generate an interrupt */
467 		amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP));
468 		amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(ctx));
469 	}
470 }
471 
472 
473 /**
474  * sdma_v5_2_gfx_stop - stop the gfx async dma engines
475  *
476  * @adev: amdgpu_device pointer
477  *
478  * Stop the gfx async dma ring buffers.
479  */
480 static void sdma_v5_2_gfx_stop(struct amdgpu_device *adev)
481 {
482 	struct amdgpu_ring *sdma0 = &adev->sdma.instance[0].ring;
483 	struct amdgpu_ring *sdma1 = &adev->sdma.instance[1].ring;
484 	struct amdgpu_ring *sdma2 = &adev->sdma.instance[2].ring;
485 	struct amdgpu_ring *sdma3 = &adev->sdma.instance[3].ring;
486 	u32 rb_cntl, ib_cntl;
487 	int i;
488 
489 	if ((adev->mman.buffer_funcs_ring == sdma0) ||
490 	    (adev->mman.buffer_funcs_ring == sdma1) ||
491 	    (adev->mman.buffer_funcs_ring == sdma2) ||
492 	    (adev->mman.buffer_funcs_ring == sdma3))
493 		amdgpu_ttm_set_buffer_funcs_status(adev, false);
494 
495 	for (i = 0; i < adev->sdma.num_instances; i++) {
496 		rb_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL));
497 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 0);
498 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
499 		ib_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL));
500 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 0);
501 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl);
502 	}
503 }
504 
505 /**
506  * sdma_v5_2_rlc_stop - stop the compute async dma engines
507  *
508  * @adev: amdgpu_device pointer
509  *
510  * Stop the compute async dma queues.
511  */
512 static void sdma_v5_2_rlc_stop(struct amdgpu_device *adev)
513 {
514 	/* XXX todo */
515 }
516 
517 /**
518  * sdma_v5_2_ctx_switch_enable - stop the async dma engines context switch
519  *
520  * @adev: amdgpu_device pointer
521  * @enable: enable/disable the DMA MEs context switch.
522  *
523  * Halt or unhalt the async dma engines context switch.
524  */
525 static void sdma_v5_2_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
526 {
527 	u32 f32_cntl, phase_quantum = 0;
528 	int i;
529 
530 	if (amdgpu_sdma_phase_quantum) {
531 		unsigned value = amdgpu_sdma_phase_quantum;
532 		unsigned unit = 0;
533 
534 		while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
535 				SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) {
536 			value = (value + 1) >> 1;
537 			unit++;
538 		}
539 		if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
540 			    SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) {
541 			value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
542 				 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT);
543 			unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
544 				SDMA0_PHASE0_QUANTUM__UNIT__SHIFT);
545 			WARN_ONCE(1,
546 			"clamping sdma_phase_quantum to %uK clock cycles\n",
547 				  value << unit);
548 		}
549 		phase_quantum =
550 			value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
551 			unit  << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
552 	}
553 
554 	for (i = 0; i < adev->sdma.num_instances; i++) {
555 		if (enable && amdgpu_sdma_phase_quantum) {
556 			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE0_QUANTUM),
557 			       phase_quantum);
558 			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE1_QUANTUM),
559 			       phase_quantum);
560 			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE2_QUANTUM),
561 			       phase_quantum);
562 		}
563 
564 		if (!amdgpu_sriov_vf(adev)) {
565 			f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
566 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
567 					AUTO_CTXSW_ENABLE, enable ? 1 : 0);
568 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl);
569 		}
570 	}
571 
572 }
573 
574 /**
575  * sdma_v5_2_enable - stop the async dma engines
576  *
577  * @adev: amdgpu_device pointer
578  * @enable: enable/disable the DMA MEs.
579  *
580  * Halt or unhalt the async dma engines.
581  */
582 static void sdma_v5_2_enable(struct amdgpu_device *adev, bool enable)
583 {
584 	u32 f32_cntl;
585 	int i;
586 
587 	if (!enable) {
588 		sdma_v5_2_gfx_stop(adev);
589 		sdma_v5_2_rlc_stop(adev);
590 	}
591 
592 	if (!amdgpu_sriov_vf(adev)) {
593 		for (i = 0; i < adev->sdma.num_instances; i++) {
594 			f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
595 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, enable ? 0 : 1);
596 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl);
597 		}
598 	}
599 }
600 
601 /**
602  * sdma_v5_2_gfx_resume - setup and start the async dma engines
603  *
604  * @adev: amdgpu_device pointer
605  *
606  * Set up the gfx DMA ring buffers and enable them.
607  * Returns 0 for success, error for failure.
608  */
609 static int sdma_v5_2_gfx_resume(struct amdgpu_device *adev)
610 {
611 	struct amdgpu_ring *ring;
612 	u32 rb_cntl, ib_cntl;
613 	u32 rb_bufsz;
614 	u32 doorbell;
615 	u32 doorbell_offset;
616 	u32 temp;
617 	u32 wptr_poll_cntl;
618 	u64 wptr_gpu_addr;
619 	int i, r;
620 
621 	for (i = 0; i < adev->sdma.num_instances; i++) {
622 		ring = &adev->sdma.instance[i].ring;
623 
624 		if (!amdgpu_sriov_vf(adev))
625 			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL), 0);
626 
627 		/* Set ring buffer size in dwords */
628 		rb_bufsz = order_base_2(ring->ring_size / 4);
629 		rb_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL));
630 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz);
631 #ifdef __BIG_ENDIAN
632 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1);
633 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL,
634 					RPTR_WRITEBACK_SWAP_ENABLE, 1);
635 #endif
636 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
637 
638 		/* Initialize the ring buffer's read and write pointers */
639 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR), 0);
640 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_HI), 0);
641 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), 0);
642 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), 0);
643 
644 		/* setup the wptr shadow polling */
645 		wptr_gpu_addr = ring->wptr_gpu_addr;
646 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_LO),
647 		       lower_32_bits(wptr_gpu_addr));
648 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_HI),
649 		       upper_32_bits(wptr_gpu_addr));
650 		wptr_poll_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i,
651 							 mmSDMA0_GFX_RB_WPTR_POLL_CNTL));
652 		wptr_poll_cntl = REG_SET_FIELD(wptr_poll_cntl,
653 					       SDMA0_GFX_RB_WPTR_POLL_CNTL,
654 					       F32_POLL_ENABLE, 1);
655 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_CNTL),
656 		       wptr_poll_cntl);
657 
658 		/* set the wb address whether it's enabled or not */
659 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_HI),
660 		       upper_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFF);
661 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_LO),
662 		       lower_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFC);
663 
664 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
665 
666 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE), ring->gpu_addr >> 8);
667 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE_HI), ring->gpu_addr >> 40);
668 
669 		ring->wptr = 0;
670 
671 		/* before programing wptr to a less value, need set minor_ptr_update first */
672 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 1);
673 
674 		if (!amdgpu_sriov_vf(adev)) { /* only bare-metal use register write for wptr */
675 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), lower_32_bits(ring->wptr << 2));
676 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), upper_32_bits(ring->wptr << 2));
677 		}
678 
679 		doorbell = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL));
680 		doorbell_offset = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL_OFFSET));
681 
682 		if (ring->use_doorbell) {
683 			doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 1);
684 			doorbell_offset = REG_SET_FIELD(doorbell_offset, SDMA0_GFX_DOORBELL_OFFSET,
685 					OFFSET, ring->doorbell_index);
686 		} else {
687 			doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 0);
688 		}
689 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL), doorbell);
690 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL_OFFSET), doorbell_offset);
691 
692 		adev->nbio.funcs->sdma_doorbell_range(adev, i, ring->use_doorbell,
693 						      ring->doorbell_index,
694 						      adev->doorbell_index.sdma_doorbell_range);
695 
696 		if (amdgpu_sriov_vf(adev))
697 			sdma_v5_2_ring_set_wptr(ring);
698 
699 		/* set minor_ptr_update to 0 after wptr programed */
700 
701 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 0);
702 
703 		/* SRIOV VF has no control of any of registers below */
704 		if (!amdgpu_sriov_vf(adev)) {
705 			/* set utc l1 enable flag always to 1 */
706 			temp = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
707 			temp = REG_SET_FIELD(temp, SDMA0_CNTL, UTC_L1_ENABLE, 1);
708 
709 			/* enable MCBP */
710 			temp = REG_SET_FIELD(temp, SDMA0_CNTL, MIDCMD_PREEMPT_ENABLE, 1);
711 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), temp);
712 
713 			/* Set up RESP_MODE to non-copy addresses */
714 			temp = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL));
715 			temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, RESP_MODE, 3);
716 			temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, REDO_DELAY, 9);
717 			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL), temp);
718 
719 			/* program default cache read and write policy */
720 			temp = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE));
721 			/* clean read policy and write policy bits */
722 			temp &= 0xFF0FFF;
723 			temp |= ((CACHE_READ_POLICY_L2__DEFAULT << 12) |
724 				 (CACHE_WRITE_POLICY_L2__DEFAULT << 14) |
725 				 SDMA0_UTCL1_PAGE__LLC_NOALLOC_MASK);
726 			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE), temp);
727 
728 			/* unhalt engine */
729 			temp = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
730 			temp = REG_SET_FIELD(temp, SDMA0_F32_CNTL, HALT, 0);
731 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), temp);
732 		}
733 
734 		/* enable DMA RB */
735 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1);
736 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
737 
738 		ib_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL));
739 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 1);
740 #ifdef __BIG_ENDIAN
741 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1);
742 #endif
743 		/* enable DMA IBs */
744 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl);
745 
746 		ring->sched.ready = true;
747 
748 		if (amdgpu_sriov_vf(adev)) { /* bare-metal sequence doesn't need below to lines */
749 			sdma_v5_2_ctx_switch_enable(adev, true);
750 			sdma_v5_2_enable(adev, true);
751 		}
752 
753 		r = amdgpu_ring_test_ring(ring);
754 		if (r) {
755 			ring->sched.ready = false;
756 			return r;
757 		}
758 
759 		if (adev->mman.buffer_funcs_ring == ring)
760 			amdgpu_ttm_set_buffer_funcs_status(adev, true);
761 	}
762 
763 	return 0;
764 }
765 
766 /**
767  * sdma_v5_2_rlc_resume - setup and start the async dma engines
768  *
769  * @adev: amdgpu_device pointer
770  *
771  * Set up the compute DMA queues and enable them.
772  * Returns 0 for success, error for failure.
773  */
774 static int sdma_v5_2_rlc_resume(struct amdgpu_device *adev)
775 {
776 	return 0;
777 }
778 
779 /**
780  * sdma_v5_2_load_microcode - load the sDMA ME ucode
781  *
782  * @adev: amdgpu_device pointer
783  *
784  * Loads the sDMA0/1/2/3 ucode.
785  * Returns 0 for success, -EINVAL if the ucode is not available.
786  */
787 static int sdma_v5_2_load_microcode(struct amdgpu_device *adev)
788 {
789 	const struct sdma_firmware_header_v1_0 *hdr;
790 	const __le32 *fw_data;
791 	u32 fw_size;
792 	int i, j;
793 
794 	/* halt the MEs */
795 	sdma_v5_2_enable(adev, false);
796 
797 	for (i = 0; i < adev->sdma.num_instances; i++) {
798 		if (!adev->sdma.instance[i].fw)
799 			return -EINVAL;
800 
801 		hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
802 		amdgpu_ucode_print_sdma_hdr(&hdr->header);
803 		fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
804 
805 		fw_data = (const __le32 *)
806 			(adev->sdma.instance[i].fw->data +
807 				le32_to_cpu(hdr->header.ucode_array_offset_bytes));
808 
809 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), 0);
810 
811 		for (j = 0; j < fw_size; j++) {
812 			if (amdgpu_emu_mode == 1 && j % 500 == 0)
813 				msleep(1);
814 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UCODE_DATA), le32_to_cpup(fw_data++));
815 		}
816 
817 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), adev->sdma.instance[i].fw_version);
818 	}
819 
820 	return 0;
821 }
822 
823 static int sdma_v5_2_soft_reset(void *handle)
824 {
825 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
826 	u32 grbm_soft_reset;
827 	u32 tmp;
828 	int i;
829 
830 	for (i = 0; i < adev->sdma.num_instances; i++) {
831 		grbm_soft_reset = REG_SET_FIELD(0,
832 						GRBM_SOFT_RESET, SOFT_RESET_SDMA0,
833 						1);
834 		grbm_soft_reset <<= i;
835 
836 		tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET);
837 		tmp |= grbm_soft_reset;
838 		DRM_DEBUG("GRBM_SOFT_RESET=0x%08X\n", tmp);
839 		WREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET, tmp);
840 		tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET);
841 
842 		udelay(50);
843 
844 		tmp &= ~grbm_soft_reset;
845 		WREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET, tmp);
846 		tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET);
847 
848 		udelay(50);
849 	}
850 
851 	return 0;
852 }
853 
854 /**
855  * sdma_v5_2_start - setup and start the async dma engines
856  *
857  * @adev: amdgpu_device pointer
858  *
859  * Set up the DMA engines and enable them.
860  * Returns 0 for success, error for failure.
861  */
862 static int sdma_v5_2_start(struct amdgpu_device *adev)
863 {
864 	int r = 0;
865 
866 	if (amdgpu_sriov_vf(adev)) {
867 		sdma_v5_2_ctx_switch_enable(adev, false);
868 		sdma_v5_2_enable(adev, false);
869 
870 		/* set RB registers */
871 		r = sdma_v5_2_gfx_resume(adev);
872 		return r;
873 	}
874 
875 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
876 		r = sdma_v5_2_load_microcode(adev);
877 		if (r)
878 			return r;
879 
880 		/* The value of mmSDMA_F32_CNTL is invalid the moment after loading fw */
881 		if (amdgpu_emu_mode == 1)
882 			msleep(1000);
883 	}
884 
885 	/* TODO: check whether can submit a doorbell request to raise
886 	 * a doorbell fence to exit gfxoff.
887 	 */
888 	if (adev->in_s0ix)
889 		amdgpu_gfx_off_ctrl(adev, false);
890 
891 	sdma_v5_2_soft_reset(adev);
892 	/* unhalt the MEs */
893 	sdma_v5_2_enable(adev, true);
894 	/* enable sdma ring preemption */
895 	sdma_v5_2_ctx_switch_enable(adev, true);
896 
897 	/* start the gfx rings and rlc compute queues */
898 	r = sdma_v5_2_gfx_resume(adev);
899 	if (adev->in_s0ix)
900 		amdgpu_gfx_off_ctrl(adev, true);
901 	if (r)
902 		return r;
903 	r = sdma_v5_2_rlc_resume(adev);
904 
905 	return r;
906 }
907 
908 static int sdma_v5_2_mqd_init(struct amdgpu_device *adev, void *mqd,
909 			      struct amdgpu_mqd_prop *prop)
910 {
911 	struct v10_sdma_mqd *m = mqd;
912 	uint64_t wb_gpu_addr;
913 
914 	m->sdmax_rlcx_rb_cntl =
915 		order_base_2(prop->queue_size / 4) << SDMA0_RLC0_RB_CNTL__RB_SIZE__SHIFT |
916 		1 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_ENABLE__SHIFT |
917 		6 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_TIMER__SHIFT |
918 		1 << SDMA0_RLC0_RB_CNTL__RB_PRIV__SHIFT;
919 
920 	m->sdmax_rlcx_rb_base = lower_32_bits(prop->hqd_base_gpu_addr >> 8);
921 	m->sdmax_rlcx_rb_base_hi = upper_32_bits(prop->hqd_base_gpu_addr >> 8);
922 
923 	m->sdmax_rlcx_rb_wptr_poll_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, 0,
924 						  mmSDMA0_GFX_RB_WPTR_POLL_CNTL));
925 
926 	wb_gpu_addr = prop->wptr_gpu_addr;
927 	m->sdmax_rlcx_rb_wptr_poll_addr_lo = lower_32_bits(wb_gpu_addr);
928 	m->sdmax_rlcx_rb_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr);
929 
930 	wb_gpu_addr = prop->rptr_gpu_addr;
931 	m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits(wb_gpu_addr);
932 	m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits(wb_gpu_addr);
933 
934 	m->sdmax_rlcx_ib_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, 0,
935 							mmSDMA0_GFX_IB_CNTL));
936 
937 	m->sdmax_rlcx_doorbell_offset =
938 		prop->doorbell_index << SDMA0_RLC0_DOORBELL_OFFSET__OFFSET__SHIFT;
939 
940 	m->sdmax_rlcx_doorbell = REG_SET_FIELD(0, SDMA0_RLC0_DOORBELL, ENABLE, 1);
941 
942 	return 0;
943 }
944 
945 static void sdma_v5_2_set_mqd_funcs(struct amdgpu_device *adev)
946 {
947 	adev->mqds[AMDGPU_HW_IP_DMA].mqd_size = sizeof(struct v10_sdma_mqd);
948 	adev->mqds[AMDGPU_HW_IP_DMA].init_mqd = sdma_v5_2_mqd_init;
949 }
950 
951 /**
952  * sdma_v5_2_ring_test_ring - simple async dma engine test
953  *
954  * @ring: amdgpu_ring structure holding ring information
955  *
956  * Test the DMA engine by writing using it to write an
957  * value to memory.
958  * Returns 0 for success, error for failure.
959  */
960 static int sdma_v5_2_ring_test_ring(struct amdgpu_ring *ring)
961 {
962 	struct amdgpu_device *adev = ring->adev;
963 	unsigned i;
964 	unsigned index;
965 	int r;
966 	u32 tmp;
967 	u64 gpu_addr;
968 	volatile uint32_t *cpu_ptr = NULL;
969 
970 	tmp = 0xCAFEDEAD;
971 
972 	if (ring->is_mes_queue) {
973 		uint32_t offset = 0;
974 		offset = amdgpu_mes_ctx_get_offs(ring,
975 					 AMDGPU_MES_CTX_PADDING_OFFS);
976 		gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
977 		cpu_ptr = amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
978 		*cpu_ptr = tmp;
979 	} else {
980 		r = amdgpu_device_wb_get(adev, &index);
981 		if (r) {
982 			dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
983 			return r;
984 		}
985 
986 		gpu_addr = adev->wb.gpu_addr + (index * 4);
987 		adev->wb.wb[index] = cpu_to_le32(tmp);
988 	}
989 
990 	r = amdgpu_ring_alloc(ring, 20);
991 	if (r) {
992 		DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
993 		amdgpu_device_wb_free(adev, index);
994 		return r;
995 	}
996 
997 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
998 			  SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
999 	amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
1000 	amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
1001 	amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0));
1002 	amdgpu_ring_write(ring, 0xDEADBEEF);
1003 	amdgpu_ring_commit(ring);
1004 
1005 	for (i = 0; i < adev->usec_timeout; i++) {
1006 		if (ring->is_mes_queue)
1007 			tmp = le32_to_cpu(*cpu_ptr);
1008 		else
1009 			tmp = le32_to_cpu(adev->wb.wb[index]);
1010 		if (tmp == 0xDEADBEEF)
1011 			break;
1012 		if (amdgpu_emu_mode == 1)
1013 			msleep(1);
1014 		else
1015 			udelay(1);
1016 	}
1017 
1018 	if (i >= adev->usec_timeout)
1019 		r = -ETIMEDOUT;
1020 
1021 	if (!ring->is_mes_queue)
1022 		amdgpu_device_wb_free(adev, index);
1023 
1024 	return r;
1025 }
1026 
1027 /**
1028  * sdma_v5_2_ring_test_ib - test an IB on the DMA engine
1029  *
1030  * @ring: amdgpu_ring structure holding ring information
1031  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
1032  *
1033  * Test a simple IB in the DMA ring.
1034  * Returns 0 on success, error on failure.
1035  */
1036 static int sdma_v5_2_ring_test_ib(struct amdgpu_ring *ring, long timeout)
1037 {
1038 	struct amdgpu_device *adev = ring->adev;
1039 	struct amdgpu_ib ib;
1040 	struct dma_fence *f = NULL;
1041 	unsigned index;
1042 	long r;
1043 	u32 tmp = 0;
1044 	u64 gpu_addr;
1045 	volatile uint32_t *cpu_ptr = NULL;
1046 
1047 	tmp = 0xCAFEDEAD;
1048 	memset(&ib, 0, sizeof(ib));
1049 
1050 	if (ring->is_mes_queue) {
1051 		uint32_t offset = 0;
1052 		offset = amdgpu_mes_ctx_get_offs(ring, AMDGPU_MES_CTX_IB_OFFS);
1053 		ib.gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
1054 		ib.ptr = (void *)amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
1055 
1056 		offset = amdgpu_mes_ctx_get_offs(ring,
1057 					 AMDGPU_MES_CTX_PADDING_OFFS);
1058 		gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
1059 		cpu_ptr = amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
1060 		*cpu_ptr = tmp;
1061 	} else {
1062 		r = amdgpu_device_wb_get(adev, &index);
1063 		if (r) {
1064 			dev_err(adev->dev, "(%ld) failed to allocate wb slot\n", r);
1065 			return r;
1066 		}
1067 
1068 		gpu_addr = adev->wb.gpu_addr + (index * 4);
1069 		adev->wb.wb[index] = cpu_to_le32(tmp);
1070 
1071 		r = amdgpu_ib_get(adev, NULL, 256, AMDGPU_IB_POOL_DIRECT, &ib);
1072 		if (r) {
1073 			DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r);
1074 			goto err0;
1075 		}
1076 	}
1077 
1078 	ib.ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
1079 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1080 	ib.ptr[1] = lower_32_bits(gpu_addr);
1081 	ib.ptr[2] = upper_32_bits(gpu_addr);
1082 	ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0);
1083 	ib.ptr[4] = 0xDEADBEEF;
1084 	ib.ptr[5] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1085 	ib.ptr[6] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1086 	ib.ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1087 	ib.length_dw = 8;
1088 
1089 	r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
1090 	if (r)
1091 		goto err1;
1092 
1093 	r = dma_fence_wait_timeout(f, false, timeout);
1094 	if (r == 0) {
1095 		DRM_ERROR("amdgpu: IB test timed out\n");
1096 		r = -ETIMEDOUT;
1097 		goto err1;
1098 	} else if (r < 0) {
1099 		DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
1100 		goto err1;
1101 	}
1102 
1103 	if (ring->is_mes_queue)
1104 		tmp = le32_to_cpu(*cpu_ptr);
1105 	else
1106 		tmp = le32_to_cpu(adev->wb.wb[index]);
1107 
1108 	if (tmp == 0xDEADBEEF)
1109 		r = 0;
1110 	else
1111 		r = -EINVAL;
1112 
1113 err1:
1114 	amdgpu_ib_free(adev, &ib, NULL);
1115 	dma_fence_put(f);
1116 err0:
1117 	if (!ring->is_mes_queue)
1118 		amdgpu_device_wb_free(adev, index);
1119 	return r;
1120 }
1121 
1122 
1123 /**
1124  * sdma_v5_2_vm_copy_pte - update PTEs by copying them from the GART
1125  *
1126  * @ib: indirect buffer to fill with commands
1127  * @pe: addr of the page entry
1128  * @src: src addr to copy from
1129  * @count: number of page entries to update
1130  *
1131  * Update PTEs by copying them from the GART using sDMA.
1132  */
1133 static void sdma_v5_2_vm_copy_pte(struct amdgpu_ib *ib,
1134 				  uint64_t pe, uint64_t src,
1135 				  unsigned count)
1136 {
1137 	unsigned bytes = count * 8;
1138 
1139 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1140 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
1141 	ib->ptr[ib->length_dw++] = bytes - 1;
1142 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1143 	ib->ptr[ib->length_dw++] = lower_32_bits(src);
1144 	ib->ptr[ib->length_dw++] = upper_32_bits(src);
1145 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1146 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1147 
1148 }
1149 
1150 /**
1151  * sdma_v5_2_vm_write_pte - update PTEs by writing them manually
1152  *
1153  * @ib: indirect buffer to fill with commands
1154  * @pe: addr of the page entry
1155  * @value: dst addr to write into pe
1156  * @count: number of page entries to update
1157  * @incr: increase next addr by incr bytes
1158  *
1159  * Update PTEs by writing them manually using sDMA.
1160  */
1161 static void sdma_v5_2_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
1162 				   uint64_t value, unsigned count,
1163 				   uint32_t incr)
1164 {
1165 	unsigned ndw = count * 2;
1166 
1167 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
1168 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1169 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1170 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1171 	ib->ptr[ib->length_dw++] = ndw - 1;
1172 	for (; ndw > 0; ndw -= 2) {
1173 		ib->ptr[ib->length_dw++] = lower_32_bits(value);
1174 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
1175 		value += incr;
1176 	}
1177 }
1178 
1179 /**
1180  * sdma_v5_2_vm_set_pte_pde - update the page tables using sDMA
1181  *
1182  * @ib: indirect buffer to fill with commands
1183  * @pe: addr of the page entry
1184  * @addr: dst addr to write into pe
1185  * @count: number of page entries to update
1186  * @incr: increase next addr by incr bytes
1187  * @flags: access flags
1188  *
1189  * Update the page tables using sDMA.
1190  */
1191 static void sdma_v5_2_vm_set_pte_pde(struct amdgpu_ib *ib,
1192 				     uint64_t pe,
1193 				     uint64_t addr, unsigned count,
1194 				     uint32_t incr, uint64_t flags)
1195 {
1196 	/* for physically contiguous pages (vram) */
1197 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_PTEPDE);
1198 	ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
1199 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1200 	ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
1201 	ib->ptr[ib->length_dw++] = upper_32_bits(flags);
1202 	ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
1203 	ib->ptr[ib->length_dw++] = upper_32_bits(addr);
1204 	ib->ptr[ib->length_dw++] = incr; /* increment size */
1205 	ib->ptr[ib->length_dw++] = 0;
1206 	ib->ptr[ib->length_dw++] = count - 1; /* number of entries */
1207 }
1208 
1209 /**
1210  * sdma_v5_2_ring_pad_ib - pad the IB
1211  *
1212  * @ib: indirect buffer to fill with padding
1213  * @ring: amdgpu_ring structure holding ring information
1214  *
1215  * Pad the IB with NOPs to a boundary multiple of 8.
1216  */
1217 static void sdma_v5_2_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
1218 {
1219 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
1220 	u32 pad_count;
1221 	int i;
1222 
1223 	pad_count = (-ib->length_dw) & 0x7;
1224 	for (i = 0; i < pad_count; i++)
1225 		if (sdma && sdma->burst_nop && (i == 0))
1226 			ib->ptr[ib->length_dw++] =
1227 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP) |
1228 				SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
1229 		else
1230 			ib->ptr[ib->length_dw++] =
1231 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
1232 }
1233 
1234 
1235 /**
1236  * sdma_v5_2_ring_emit_pipeline_sync - sync the pipeline
1237  *
1238  * @ring: amdgpu_ring pointer
1239  *
1240  * Make sure all previous operations are completed (CIK).
1241  */
1242 static void sdma_v5_2_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
1243 {
1244 	uint32_t seq = ring->fence_drv.sync_seq;
1245 	uint64_t addr = ring->fence_drv.gpu_addr;
1246 
1247 	/* wait for idle */
1248 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1249 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1250 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
1251 			  SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1));
1252 	amdgpu_ring_write(ring, addr & 0xfffffffc);
1253 	amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
1254 	amdgpu_ring_write(ring, seq); /* reference */
1255 	amdgpu_ring_write(ring, 0xffffffff); /* mask */
1256 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1257 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */
1258 }
1259 
1260 
1261 /**
1262  * sdma_v5_2_ring_emit_vm_flush - vm flush using sDMA
1263  *
1264  * @ring: amdgpu_ring pointer
1265  * @vmid: vmid number to use
1266  * @pd_addr: address
1267  *
1268  * Update the page table base and flush the VM TLB
1269  * using sDMA.
1270  */
1271 static void sdma_v5_2_ring_emit_vm_flush(struct amdgpu_ring *ring,
1272 					 unsigned vmid, uint64_t pd_addr)
1273 {
1274 	amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
1275 }
1276 
1277 static void sdma_v5_2_ring_emit_wreg(struct amdgpu_ring *ring,
1278 				     uint32_t reg, uint32_t val)
1279 {
1280 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
1281 			  SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
1282 	amdgpu_ring_write(ring, reg);
1283 	amdgpu_ring_write(ring, val);
1284 }
1285 
1286 static void sdma_v5_2_ring_emit_reg_wait(struct amdgpu_ring *ring, uint32_t reg,
1287 					 uint32_t val, uint32_t mask)
1288 {
1289 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1290 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1291 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* equal */
1292 	amdgpu_ring_write(ring, reg << 2);
1293 	amdgpu_ring_write(ring, 0);
1294 	amdgpu_ring_write(ring, val); /* reference */
1295 	amdgpu_ring_write(ring, mask); /* mask */
1296 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1297 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10));
1298 }
1299 
1300 static void sdma_v5_2_ring_emit_reg_write_reg_wait(struct amdgpu_ring *ring,
1301 						   uint32_t reg0, uint32_t reg1,
1302 						   uint32_t ref, uint32_t mask)
1303 {
1304 	amdgpu_ring_emit_wreg(ring, reg0, ref);
1305 	/* wait for a cycle to reset vm_inv_eng*_ack */
1306 	amdgpu_ring_emit_reg_wait(ring, reg0, 0, 0);
1307 	amdgpu_ring_emit_reg_wait(ring, reg1, mask, mask);
1308 }
1309 
1310 static int sdma_v5_2_early_init(void *handle)
1311 {
1312 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1313 
1314 	sdma_v5_2_set_ring_funcs(adev);
1315 	sdma_v5_2_set_buffer_funcs(adev);
1316 	sdma_v5_2_set_vm_pte_funcs(adev);
1317 	sdma_v5_2_set_irq_funcs(adev);
1318 	sdma_v5_2_set_mqd_funcs(adev);
1319 
1320 	return 0;
1321 }
1322 
1323 static unsigned sdma_v5_2_seq_to_irq_id(int seq_num)
1324 {
1325 	switch (seq_num) {
1326 	case 0:
1327 		return SOC15_IH_CLIENTID_SDMA0;
1328 	case 1:
1329 		return SOC15_IH_CLIENTID_SDMA1;
1330 	case 2:
1331 		return SOC15_IH_CLIENTID_SDMA2;
1332 	case 3:
1333 		return SOC15_IH_CLIENTID_SDMA3_Sienna_Cichlid;
1334 	default:
1335 		break;
1336 	}
1337 	return -EINVAL;
1338 }
1339 
1340 static unsigned sdma_v5_2_seq_to_trap_id(int seq_num)
1341 {
1342 	switch (seq_num) {
1343 	case 0:
1344 		return SDMA0_5_0__SRCID__SDMA_TRAP;
1345 	case 1:
1346 		return SDMA1_5_0__SRCID__SDMA_TRAP;
1347 	case 2:
1348 		return SDMA2_5_0__SRCID__SDMA_TRAP;
1349 	case 3:
1350 		return SDMA3_5_0__SRCID__SDMA_TRAP;
1351 	default:
1352 		break;
1353 	}
1354 	return -EINVAL;
1355 }
1356 
1357 static int sdma_v5_2_sw_init(void *handle)
1358 {
1359 	struct amdgpu_ring *ring;
1360 	int r, i;
1361 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1362 
1363 	/* SDMA trap event */
1364 	for (i = 0; i < adev->sdma.num_instances; i++) {
1365 		r = amdgpu_irq_add_id(adev, sdma_v5_2_seq_to_irq_id(i),
1366 				      sdma_v5_2_seq_to_trap_id(i),
1367 				      &adev->sdma.trap_irq);
1368 		if (r)
1369 			return r;
1370 	}
1371 
1372 	r = sdma_v5_2_init_microcode(adev);
1373 	if (r) {
1374 		DRM_ERROR("Failed to load sdma firmware!\n");
1375 		return r;
1376 	}
1377 
1378 	for (i = 0; i < adev->sdma.num_instances; i++) {
1379 		ring = &adev->sdma.instance[i].ring;
1380 		ring->ring_obj = NULL;
1381 		ring->use_doorbell = true;
1382 		ring->me = i;
1383 
1384 		DRM_INFO("use_doorbell being set to: [%s]\n",
1385 				ring->use_doorbell?"true":"false");
1386 
1387 		ring->doorbell_index =
1388 			(adev->doorbell_index.sdma_engine[i] << 1); //get DWORD offset
1389 
1390 		sprintf(ring->name, "sdma%d", i);
1391 		r = amdgpu_ring_init(adev, ring, 1024, &adev->sdma.trap_irq,
1392 				     AMDGPU_SDMA_IRQ_INSTANCE0 + i,
1393 				     AMDGPU_RING_PRIO_DEFAULT, NULL);
1394 		if (r)
1395 			return r;
1396 	}
1397 
1398 	return r;
1399 }
1400 
1401 static int sdma_v5_2_sw_fini(void *handle)
1402 {
1403 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1404 	int i;
1405 
1406 	for (i = 0; i < adev->sdma.num_instances; i++)
1407 		amdgpu_ring_fini(&adev->sdma.instance[i].ring);
1408 
1409 	sdma_v5_2_destroy_inst_ctx(adev);
1410 
1411 	return 0;
1412 }
1413 
1414 static int sdma_v5_2_hw_init(void *handle)
1415 {
1416 	int r;
1417 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1418 
1419 	r = sdma_v5_2_start(adev);
1420 
1421 	return r;
1422 }
1423 
1424 static int sdma_v5_2_hw_fini(void *handle)
1425 {
1426 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1427 
1428 	if (amdgpu_sriov_vf(adev))
1429 		return 0;
1430 
1431 	sdma_v5_2_ctx_switch_enable(adev, false);
1432 	sdma_v5_2_enable(adev, false);
1433 
1434 	return 0;
1435 }
1436 
1437 static int sdma_v5_2_suspend(void *handle)
1438 {
1439 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1440 
1441 	return sdma_v5_2_hw_fini(adev);
1442 }
1443 
1444 static int sdma_v5_2_resume(void *handle)
1445 {
1446 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1447 
1448 	return sdma_v5_2_hw_init(adev);
1449 }
1450 
1451 static bool sdma_v5_2_is_idle(void *handle)
1452 {
1453 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1454 	u32 i;
1455 
1456 	for (i = 0; i < adev->sdma.num_instances; i++) {
1457 		u32 tmp = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_STATUS_REG));
1458 
1459 		if (!(tmp & SDMA0_STATUS_REG__IDLE_MASK))
1460 			return false;
1461 	}
1462 
1463 	return true;
1464 }
1465 
1466 static int sdma_v5_2_wait_for_idle(void *handle)
1467 {
1468 	unsigned i;
1469 	u32 sdma0, sdma1, sdma2, sdma3;
1470 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1471 
1472 	for (i = 0; i < adev->usec_timeout; i++) {
1473 		sdma0 = RREG32(sdma_v5_2_get_reg_offset(adev, 0, mmSDMA0_STATUS_REG));
1474 		sdma1 = RREG32(sdma_v5_2_get_reg_offset(adev, 1, mmSDMA0_STATUS_REG));
1475 		sdma2 = RREG32(sdma_v5_2_get_reg_offset(adev, 2, mmSDMA0_STATUS_REG));
1476 		sdma3 = RREG32(sdma_v5_2_get_reg_offset(adev, 3, mmSDMA0_STATUS_REG));
1477 
1478 		if (sdma0 & sdma1 & sdma2 & sdma3 & SDMA0_STATUS_REG__IDLE_MASK)
1479 			return 0;
1480 		udelay(1);
1481 	}
1482 	return -ETIMEDOUT;
1483 }
1484 
1485 static int sdma_v5_2_ring_preempt_ib(struct amdgpu_ring *ring)
1486 {
1487 	int i, r = 0;
1488 	struct amdgpu_device *adev = ring->adev;
1489 	u32 index = 0;
1490 	u64 sdma_gfx_preempt;
1491 
1492 	amdgpu_sdma_get_index_from_ring(ring, &index);
1493 	sdma_gfx_preempt =
1494 		sdma_v5_2_get_reg_offset(adev, index, mmSDMA0_GFX_PREEMPT);
1495 
1496 	/* assert preemption condition */
1497 	amdgpu_ring_set_preempt_cond_exec(ring, false);
1498 
1499 	/* emit the trailing fence */
1500 	ring->trail_seq += 1;
1501 	amdgpu_ring_alloc(ring, 10);
1502 	sdma_v5_2_ring_emit_fence(ring, ring->trail_fence_gpu_addr,
1503 				  ring->trail_seq, 0);
1504 	amdgpu_ring_commit(ring);
1505 
1506 	/* assert IB preemption */
1507 	WREG32(sdma_gfx_preempt, 1);
1508 
1509 	/* poll the trailing fence */
1510 	for (i = 0; i < adev->usec_timeout; i++) {
1511 		if (ring->trail_seq ==
1512 		    le32_to_cpu(*(ring->trail_fence_cpu_addr)))
1513 			break;
1514 		udelay(1);
1515 	}
1516 
1517 	if (i >= adev->usec_timeout) {
1518 		r = -EINVAL;
1519 		DRM_ERROR("ring %d failed to be preempted\n", ring->idx);
1520 	}
1521 
1522 	/* deassert IB preemption */
1523 	WREG32(sdma_gfx_preempt, 0);
1524 
1525 	/* deassert the preemption condition */
1526 	amdgpu_ring_set_preempt_cond_exec(ring, true);
1527 	return r;
1528 }
1529 
1530 static int sdma_v5_2_set_trap_irq_state(struct amdgpu_device *adev,
1531 					struct amdgpu_irq_src *source,
1532 					unsigned type,
1533 					enum amdgpu_interrupt_state state)
1534 {
1535 	u32 sdma_cntl;
1536 	u32 reg_offset = sdma_v5_2_get_reg_offset(adev, type, mmSDMA0_CNTL);
1537 
1538 	if (!amdgpu_sriov_vf(adev)) {
1539 		sdma_cntl = RREG32(reg_offset);
1540 		sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE,
1541 			       state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0);
1542 		WREG32(reg_offset, sdma_cntl);
1543 	}
1544 
1545 	return 0;
1546 }
1547 
1548 static int sdma_v5_2_process_trap_irq(struct amdgpu_device *adev,
1549 				      struct amdgpu_irq_src *source,
1550 				      struct amdgpu_iv_entry *entry)
1551 {
1552 	uint32_t mes_queue_id = entry->src_data[0];
1553 
1554 	DRM_DEBUG("IH: SDMA trap\n");
1555 
1556 	if (adev->enable_mes && (mes_queue_id & AMDGPU_FENCE_MES_QUEUE_FLAG)) {
1557 		struct amdgpu_mes_queue *queue;
1558 
1559 		mes_queue_id &= AMDGPU_FENCE_MES_QUEUE_ID_MASK;
1560 
1561 		spin_lock(&adev->mes.queue_id_lock);
1562 		queue = idr_find(&adev->mes.queue_id_idr, mes_queue_id);
1563 		if (queue) {
1564 			DRM_DEBUG("process smda queue id = %d\n", mes_queue_id);
1565 			amdgpu_fence_process(queue->ring);
1566 		}
1567 		spin_unlock(&adev->mes.queue_id_lock);
1568 		return 0;
1569 	}
1570 
1571 	switch (entry->client_id) {
1572 	case SOC15_IH_CLIENTID_SDMA0:
1573 		switch (entry->ring_id) {
1574 		case 0:
1575 			amdgpu_fence_process(&adev->sdma.instance[0].ring);
1576 			break;
1577 		case 1:
1578 			/* XXX compute */
1579 			break;
1580 		case 2:
1581 			/* XXX compute */
1582 			break;
1583 		case 3:
1584 			/* XXX page queue*/
1585 			break;
1586 		}
1587 		break;
1588 	case SOC15_IH_CLIENTID_SDMA1:
1589 		switch (entry->ring_id) {
1590 		case 0:
1591 			amdgpu_fence_process(&adev->sdma.instance[1].ring);
1592 			break;
1593 		case 1:
1594 			/* XXX compute */
1595 			break;
1596 		case 2:
1597 			/* XXX compute */
1598 			break;
1599 		case 3:
1600 			/* XXX page queue*/
1601 			break;
1602 		}
1603 		break;
1604 	case SOC15_IH_CLIENTID_SDMA2:
1605 		switch (entry->ring_id) {
1606 		case 0:
1607 			amdgpu_fence_process(&adev->sdma.instance[2].ring);
1608 			break;
1609 		case 1:
1610 			/* XXX compute */
1611 			break;
1612 		case 2:
1613 			/* XXX compute */
1614 			break;
1615 		case 3:
1616 			/* XXX page queue*/
1617 			break;
1618 		}
1619 		break;
1620 	case SOC15_IH_CLIENTID_SDMA3_Sienna_Cichlid:
1621 		switch (entry->ring_id) {
1622 		case 0:
1623 			amdgpu_fence_process(&adev->sdma.instance[3].ring);
1624 			break;
1625 		case 1:
1626 			/* XXX compute */
1627 			break;
1628 		case 2:
1629 			/* XXX compute */
1630 			break;
1631 		case 3:
1632 			/* XXX page queue*/
1633 			break;
1634 		}
1635 		break;
1636 	}
1637 	return 0;
1638 }
1639 
1640 static int sdma_v5_2_process_illegal_inst_irq(struct amdgpu_device *adev,
1641 					      struct amdgpu_irq_src *source,
1642 					      struct amdgpu_iv_entry *entry)
1643 {
1644 	return 0;
1645 }
1646 
1647 static void sdma_v5_2_update_medium_grain_clock_gating(struct amdgpu_device *adev,
1648 						       bool enable)
1649 {
1650 	uint32_t data, def;
1651 	int i;
1652 
1653 	for (i = 0; i < adev->sdma.num_instances; i++) {
1654 
1655 		if (adev->sdma.instance[i].fw_version < 70 && adev->ip_versions[SDMA0_HWIP][0] == IP_VERSION(5, 2, 1))
1656 			adev->cg_flags &= ~AMD_CG_SUPPORT_SDMA_MGCG;
1657 
1658 		if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
1659 			/* Enable sdma clock gating */
1660 			def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL));
1661 			data &= ~(SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1662 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1663 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1664 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1665 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK |
1666 				  SDMA0_CLK_CTRL__SOFT_OVERRIDER_REG_MASK);
1667 			if (def != data)
1668 				WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data);
1669 		} else {
1670 			/* Disable sdma clock gating */
1671 			def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL));
1672 			data |= (SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1673 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1674 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1675 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1676 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK |
1677 				 SDMA0_CLK_CTRL__SOFT_OVERRIDER_REG_MASK);
1678 			if (def != data)
1679 				WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data);
1680 		}
1681 	}
1682 }
1683 
1684 static void sdma_v5_2_update_medium_grain_light_sleep(struct amdgpu_device *adev,
1685 						      bool enable)
1686 {
1687 	uint32_t data, def;
1688 	int i;
1689 
1690 	for (i = 0; i < adev->sdma.num_instances; i++) {
1691 
1692 		if (adev->sdma.instance[i].fw_version < 70 && adev->ip_versions[SDMA0_HWIP][0] == IP_VERSION(5, 2, 1))
1693 			adev->cg_flags &= ~AMD_CG_SUPPORT_SDMA_LS;
1694 
1695 		if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) {
1696 			/* Enable sdma mem light sleep */
1697 			def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL));
1698 			data |= SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1699 			if (def != data)
1700 				WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data);
1701 
1702 		} else {
1703 			/* Disable sdma mem light sleep */
1704 			def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL));
1705 			data &= ~SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1706 			if (def != data)
1707 				WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data);
1708 
1709 		}
1710 	}
1711 }
1712 
1713 static int sdma_v5_2_set_clockgating_state(void *handle,
1714 					   enum amd_clockgating_state state)
1715 {
1716 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1717 
1718 	if (amdgpu_sriov_vf(adev))
1719 		return 0;
1720 
1721 	switch (adev->ip_versions[SDMA0_HWIP][0]) {
1722 	case IP_VERSION(5, 2, 0):
1723 	case IP_VERSION(5, 2, 2):
1724 	case IP_VERSION(5, 2, 1):
1725 	case IP_VERSION(5, 2, 4):
1726 	case IP_VERSION(5, 2, 5):
1727 	case IP_VERSION(5, 2, 6):
1728 	case IP_VERSION(5, 2, 3):
1729 		sdma_v5_2_update_medium_grain_clock_gating(adev,
1730 				state == AMD_CG_STATE_GATE);
1731 		sdma_v5_2_update_medium_grain_light_sleep(adev,
1732 				state == AMD_CG_STATE_GATE);
1733 		break;
1734 	default:
1735 		break;
1736 	}
1737 
1738 	return 0;
1739 }
1740 
1741 static int sdma_v5_2_set_powergating_state(void *handle,
1742 					  enum amd_powergating_state state)
1743 {
1744 	return 0;
1745 }
1746 
1747 static void sdma_v5_2_get_clockgating_state(void *handle, u64 *flags)
1748 {
1749 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1750 	int data;
1751 
1752 	if (amdgpu_sriov_vf(adev))
1753 		*flags = 0;
1754 
1755 	/* AMD_CG_SUPPORT_SDMA_MGCG */
1756 	data = RREG32(sdma_v5_2_get_reg_offset(adev, 0, mmSDMA0_CLK_CTRL));
1757 	if (!(data & SDMA0_CLK_CTRL__CGCG_EN_OVERRIDE_MASK))
1758 		*flags |= AMD_CG_SUPPORT_SDMA_MGCG;
1759 
1760 	/* AMD_CG_SUPPORT_SDMA_LS */
1761 	data = RREG32_KIQ(sdma_v5_2_get_reg_offset(adev, 0, mmSDMA0_POWER_CNTL));
1762 	if (data & SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK)
1763 		*flags |= AMD_CG_SUPPORT_SDMA_LS;
1764 }
1765 
1766 const struct amd_ip_funcs sdma_v5_2_ip_funcs = {
1767 	.name = "sdma_v5_2",
1768 	.early_init = sdma_v5_2_early_init,
1769 	.late_init = NULL,
1770 	.sw_init = sdma_v5_2_sw_init,
1771 	.sw_fini = sdma_v5_2_sw_fini,
1772 	.hw_init = sdma_v5_2_hw_init,
1773 	.hw_fini = sdma_v5_2_hw_fini,
1774 	.suspend = sdma_v5_2_suspend,
1775 	.resume = sdma_v5_2_resume,
1776 	.is_idle = sdma_v5_2_is_idle,
1777 	.wait_for_idle = sdma_v5_2_wait_for_idle,
1778 	.soft_reset = sdma_v5_2_soft_reset,
1779 	.set_clockgating_state = sdma_v5_2_set_clockgating_state,
1780 	.set_powergating_state = sdma_v5_2_set_powergating_state,
1781 	.get_clockgating_state = sdma_v5_2_get_clockgating_state,
1782 };
1783 
1784 static const struct amdgpu_ring_funcs sdma_v5_2_ring_funcs = {
1785 	.type = AMDGPU_RING_TYPE_SDMA,
1786 	.align_mask = 0xf,
1787 	.nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP),
1788 	.support_64bit_ptrs = true,
1789 	.secure_submission_supported = true,
1790 	.vmhub = AMDGPU_GFXHUB_0,
1791 	.get_rptr = sdma_v5_2_ring_get_rptr,
1792 	.get_wptr = sdma_v5_2_ring_get_wptr,
1793 	.set_wptr = sdma_v5_2_ring_set_wptr,
1794 	.emit_frame_size =
1795 		5 + /* sdma_v5_2_ring_init_cond_exec */
1796 		6 + /* sdma_v5_2_ring_emit_hdp_flush */
1797 		3 + /* hdp_invalidate */
1798 		6 + /* sdma_v5_2_ring_emit_pipeline_sync */
1799 		/* sdma_v5_2_ring_emit_vm_flush */
1800 		SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 +
1801 		SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6 +
1802 		10 + 10 + 10, /* sdma_v5_2_ring_emit_fence x3 for user fence, vm fence */
1803 	.emit_ib_size = 7 + 6, /* sdma_v5_2_ring_emit_ib */
1804 	.emit_ib = sdma_v5_2_ring_emit_ib,
1805 	.emit_mem_sync = sdma_v5_2_ring_emit_mem_sync,
1806 	.emit_fence = sdma_v5_2_ring_emit_fence,
1807 	.emit_pipeline_sync = sdma_v5_2_ring_emit_pipeline_sync,
1808 	.emit_vm_flush = sdma_v5_2_ring_emit_vm_flush,
1809 	.emit_hdp_flush = sdma_v5_2_ring_emit_hdp_flush,
1810 	.test_ring = sdma_v5_2_ring_test_ring,
1811 	.test_ib = sdma_v5_2_ring_test_ib,
1812 	.insert_nop = sdma_v5_2_ring_insert_nop,
1813 	.pad_ib = sdma_v5_2_ring_pad_ib,
1814 	.emit_wreg = sdma_v5_2_ring_emit_wreg,
1815 	.emit_reg_wait = sdma_v5_2_ring_emit_reg_wait,
1816 	.emit_reg_write_reg_wait = sdma_v5_2_ring_emit_reg_write_reg_wait,
1817 	.init_cond_exec = sdma_v5_2_ring_init_cond_exec,
1818 	.patch_cond_exec = sdma_v5_2_ring_patch_cond_exec,
1819 	.preempt_ib = sdma_v5_2_ring_preempt_ib,
1820 };
1821 
1822 static void sdma_v5_2_set_ring_funcs(struct amdgpu_device *adev)
1823 {
1824 	int i;
1825 
1826 	for (i = 0; i < adev->sdma.num_instances; i++) {
1827 		adev->sdma.instance[i].ring.funcs = &sdma_v5_2_ring_funcs;
1828 		adev->sdma.instance[i].ring.me = i;
1829 	}
1830 }
1831 
1832 static const struct amdgpu_irq_src_funcs sdma_v5_2_trap_irq_funcs = {
1833 	.set = sdma_v5_2_set_trap_irq_state,
1834 	.process = sdma_v5_2_process_trap_irq,
1835 };
1836 
1837 static const struct amdgpu_irq_src_funcs sdma_v5_2_illegal_inst_irq_funcs = {
1838 	.process = sdma_v5_2_process_illegal_inst_irq,
1839 };
1840 
1841 static void sdma_v5_2_set_irq_funcs(struct amdgpu_device *adev)
1842 {
1843 	adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_INSTANCE0 +
1844 					adev->sdma.num_instances;
1845 	adev->sdma.trap_irq.funcs = &sdma_v5_2_trap_irq_funcs;
1846 	adev->sdma.illegal_inst_irq.funcs = &sdma_v5_2_illegal_inst_irq_funcs;
1847 }
1848 
1849 /**
1850  * sdma_v5_2_emit_copy_buffer - copy buffer using the sDMA engine
1851  *
1852  * @ib: indirect buffer to copy to
1853  * @src_offset: src GPU address
1854  * @dst_offset: dst GPU address
1855  * @byte_count: number of bytes to xfer
1856  * @tmz: if a secure copy should be used
1857  *
1858  * Copy GPU buffers using the DMA engine.
1859  * Used by the amdgpu ttm implementation to move pages if
1860  * registered as the asic copy callback.
1861  */
1862 static void sdma_v5_2_emit_copy_buffer(struct amdgpu_ib *ib,
1863 				       uint64_t src_offset,
1864 				       uint64_t dst_offset,
1865 				       uint32_t byte_count,
1866 				       bool tmz)
1867 {
1868 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1869 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR) |
1870 		SDMA_PKT_COPY_LINEAR_HEADER_TMZ(tmz ? 1 : 0);
1871 	ib->ptr[ib->length_dw++] = byte_count - 1;
1872 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1873 	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1874 	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1875 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1876 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1877 }
1878 
1879 /**
1880  * sdma_v5_2_emit_fill_buffer - fill buffer using the sDMA engine
1881  *
1882  * @ib: indirect buffer to fill
1883  * @src_data: value to write to buffer
1884  * @dst_offset: dst GPU address
1885  * @byte_count: number of bytes to xfer
1886  *
1887  * Fill GPU buffers using the DMA engine.
1888  */
1889 static void sdma_v5_2_emit_fill_buffer(struct amdgpu_ib *ib,
1890 				       uint32_t src_data,
1891 				       uint64_t dst_offset,
1892 				       uint32_t byte_count)
1893 {
1894 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL);
1895 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1896 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1897 	ib->ptr[ib->length_dw++] = src_data;
1898 	ib->ptr[ib->length_dw++] = byte_count - 1;
1899 }
1900 
1901 static const struct amdgpu_buffer_funcs sdma_v5_2_buffer_funcs = {
1902 	.copy_max_bytes = 0x400000,
1903 	.copy_num_dw = 7,
1904 	.emit_copy_buffer = sdma_v5_2_emit_copy_buffer,
1905 
1906 	.fill_max_bytes = 0x400000,
1907 	.fill_num_dw = 5,
1908 	.emit_fill_buffer = sdma_v5_2_emit_fill_buffer,
1909 };
1910 
1911 static void sdma_v5_2_set_buffer_funcs(struct amdgpu_device *adev)
1912 {
1913 	if (adev->mman.buffer_funcs == NULL) {
1914 		adev->mman.buffer_funcs = &sdma_v5_2_buffer_funcs;
1915 		adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1916 	}
1917 }
1918 
1919 static const struct amdgpu_vm_pte_funcs sdma_v5_2_vm_pte_funcs = {
1920 	.copy_pte_num_dw = 7,
1921 	.copy_pte = sdma_v5_2_vm_copy_pte,
1922 	.write_pte = sdma_v5_2_vm_write_pte,
1923 	.set_pte_pde = sdma_v5_2_vm_set_pte_pde,
1924 };
1925 
1926 static void sdma_v5_2_set_vm_pte_funcs(struct amdgpu_device *adev)
1927 {
1928 	unsigned i;
1929 
1930 	if (adev->vm_manager.vm_pte_funcs == NULL) {
1931 		adev->vm_manager.vm_pte_funcs = &sdma_v5_2_vm_pte_funcs;
1932 		for (i = 0; i < adev->sdma.num_instances; i++) {
1933 			adev->vm_manager.vm_pte_scheds[i] =
1934 				&adev->sdma.instance[i].ring.sched;
1935 		}
1936 		adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
1937 	}
1938 }
1939 
1940 const struct amdgpu_ip_block_version sdma_v5_2_ip_block = {
1941 	.type = AMD_IP_BLOCK_TYPE_SDMA,
1942 	.major = 5,
1943 	.minor = 2,
1944 	.rev = 0,
1945 	.funcs = &sdma_v5_2_ip_funcs,
1946 };
1947