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 
55 #define SDMA1_REG_OFFSET 0x600
56 #define SDMA3_REG_OFFSET 0x400
57 #define SDMA0_HYP_DEC_REG_START 0x5880
58 #define SDMA0_HYP_DEC_REG_END 0x5893
59 #define SDMA1_HYP_DEC_REG_OFFSET 0x20
60 
61 static void sdma_v5_2_set_ring_funcs(struct amdgpu_device *adev);
62 static void sdma_v5_2_set_buffer_funcs(struct amdgpu_device *adev);
63 static void sdma_v5_2_set_vm_pte_funcs(struct amdgpu_device *adev);
64 static void sdma_v5_2_set_irq_funcs(struct amdgpu_device *adev);
65 
66 static u32 sdma_v5_2_get_reg_offset(struct amdgpu_device *adev, u32 instance, u32 internal_offset)
67 {
68 	u32 base;
69 
70 	if (internal_offset >= SDMA0_HYP_DEC_REG_START &&
71 	    internal_offset <= SDMA0_HYP_DEC_REG_END) {
72 		base = adev->reg_offset[GC_HWIP][0][1];
73 		if (instance != 0)
74 			internal_offset += SDMA1_HYP_DEC_REG_OFFSET * instance;
75 	} else {
76 		if (instance < 2) {
77 			base = adev->reg_offset[GC_HWIP][0][0];
78 			if (instance == 1)
79 				internal_offset += SDMA1_REG_OFFSET;
80 		} else {
81 			base = adev->reg_offset[GC_HWIP][0][2];
82 			if (instance == 3)
83 				internal_offset += SDMA3_REG_OFFSET;
84 		}
85 	}
86 
87 	return base + internal_offset;
88 }
89 
90 static int sdma_v5_2_init_inst_ctx(struct amdgpu_sdma_instance *sdma_inst)
91 {
92 	int err = 0;
93 	const struct sdma_firmware_header_v1_0 *hdr;
94 
95 	err = amdgpu_ucode_validate(sdma_inst->fw);
96 	if (err)
97 		return err;
98 
99 	hdr = (const struct sdma_firmware_header_v1_0 *)sdma_inst->fw->data;
100 	sdma_inst->fw_version = le32_to_cpu(hdr->header.ucode_version);
101 	sdma_inst->feature_version = le32_to_cpu(hdr->ucode_feature_version);
102 
103 	if (sdma_inst->feature_version >= 20)
104 		sdma_inst->burst_nop = true;
105 
106 	return 0;
107 }
108 
109 static void sdma_v5_2_destroy_inst_ctx(struct amdgpu_device *adev)
110 {
111 	release_firmware(adev->sdma.instance[0].fw);
112 
113 	memset((void *)adev->sdma.instance, 0,
114 	       sizeof(struct amdgpu_sdma_instance) * AMDGPU_MAX_SDMA_INSTANCES);
115 }
116 
117 /**
118  * sdma_v5_2_init_microcode - load ucode images from disk
119  *
120  * @adev: amdgpu_device pointer
121  *
122  * Use the firmware interface to load the ucode images into
123  * the driver (not loaded into hw).
124  * Returns 0 on success, error on failure.
125  */
126 
127 // emulation only, won't work on real chip
128 // navi10 real chip need to use PSP to load firmware
129 static int sdma_v5_2_init_microcode(struct amdgpu_device *adev)
130 {
131 	const char *chip_name;
132 	char fw_name[40];
133 	int err = 0, i;
134 	struct amdgpu_firmware_info *info = NULL;
135 	const struct common_firmware_header *header = NULL;
136 
137 	DRM_DEBUG("\n");
138 
139 	switch (adev->ip_versions[SDMA0_HWIP][0]) {
140 	case IP_VERSION(5, 2, 0):
141 		chip_name = "sienna_cichlid";
142 		break;
143 	case IP_VERSION(5, 2, 2):
144 		chip_name = "navy_flounder";
145 		break;
146 	case IP_VERSION(5, 2, 1):
147 		chip_name = "vangogh";
148 		break;
149 	case IP_VERSION(5, 2, 4):
150 		chip_name = "dimgrey_cavefish";
151 		break;
152 	case IP_VERSION(5, 2, 5):
153 		chip_name = "beige_goby";
154 		break;
155 	case IP_VERSION(5, 2, 3):
156 		chip_name = "yellow_carp";
157 		break;
158 	default:
159 		BUG();
160 	}
161 
162 	snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma.bin", chip_name);
163 
164 	err = request_firmware(&adev->sdma.instance[0].fw, fw_name, adev->dev);
165 	if (err)
166 		goto out;
167 
168 	err = sdma_v5_2_init_inst_ctx(&adev->sdma.instance[0]);
169 	if (err)
170 		goto out;
171 
172 	for (i = 1; i < adev->sdma.num_instances; i++)
173 		memcpy((void *)&adev->sdma.instance[i],
174 		       (void *)&adev->sdma.instance[0],
175 		       sizeof(struct amdgpu_sdma_instance));
176 
177 	if (amdgpu_sriov_vf(adev) && (adev->ip_versions[SDMA0_HWIP][0] == IP_VERSION(5, 2, 0)))
178 		return 0;
179 
180 	DRM_DEBUG("psp_load == '%s'\n",
181 		  adev->firmware.load_type == AMDGPU_FW_LOAD_PSP ? "true" : "false");
182 
183 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
184 		for (i = 0; i < adev->sdma.num_instances; i++) {
185 			info = &adev->firmware.ucode[AMDGPU_UCODE_ID_SDMA0 + i];
186 			info->ucode_id = AMDGPU_UCODE_ID_SDMA0 + i;
187 			info->fw = adev->sdma.instance[i].fw;
188 			header = (const struct common_firmware_header *)info->fw->data;
189 			adev->firmware.fw_size +=
190 				ALIGN(le32_to_cpu(header->ucode_size_bytes), PAGE_SIZE);
191 		}
192 	}
193 
194 out:
195 	if (err) {
196 		DRM_ERROR("sdma_v5_2: Failed to load firmware \"%s\"\n", fw_name);
197 		sdma_v5_2_destroy_inst_ctx(adev);
198 	}
199 	return err;
200 }
201 
202 static unsigned sdma_v5_2_ring_init_cond_exec(struct amdgpu_ring *ring)
203 {
204 	unsigned ret;
205 
206 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_COND_EXE));
207 	amdgpu_ring_write(ring, lower_32_bits(ring->cond_exe_gpu_addr));
208 	amdgpu_ring_write(ring, upper_32_bits(ring->cond_exe_gpu_addr));
209 	amdgpu_ring_write(ring, 1);
210 	ret = ring->wptr & ring->buf_mask;/* this is the offset we need patch later */
211 	amdgpu_ring_write(ring, 0x55aa55aa);/* insert dummy here and patch it later */
212 
213 	return ret;
214 }
215 
216 static void sdma_v5_2_ring_patch_cond_exec(struct amdgpu_ring *ring,
217 					   unsigned offset)
218 {
219 	unsigned cur;
220 
221 	BUG_ON(offset > ring->buf_mask);
222 	BUG_ON(ring->ring[offset] != 0x55aa55aa);
223 
224 	cur = (ring->wptr - 1) & ring->buf_mask;
225 	if (cur > offset)
226 		ring->ring[offset] = cur - offset;
227 	else
228 		ring->ring[offset] = (ring->buf_mask + 1) - offset + cur;
229 }
230 
231 /**
232  * sdma_v5_2_ring_get_rptr - get the current read pointer
233  *
234  * @ring: amdgpu ring pointer
235  *
236  * Get the current rptr from the hardware (NAVI10+).
237  */
238 static uint64_t sdma_v5_2_ring_get_rptr(struct amdgpu_ring *ring)
239 {
240 	u64 *rptr;
241 
242 	/* XXX check if swapping is necessary on BE */
243 	rptr = ((u64 *)&ring->adev->wb.wb[ring->rptr_offs]);
244 
245 	DRM_DEBUG("rptr before shift == 0x%016llx\n", *rptr);
246 	return ((*rptr) >> 2);
247 }
248 
249 /**
250  * sdma_v5_2_ring_get_wptr - get the current write pointer
251  *
252  * @ring: amdgpu ring pointer
253  *
254  * Get the current wptr from the hardware (NAVI10+).
255  */
256 static uint64_t sdma_v5_2_ring_get_wptr(struct amdgpu_ring *ring)
257 {
258 	struct amdgpu_device *adev = ring->adev;
259 	u64 wptr;
260 
261 	if (ring->use_doorbell) {
262 		/* XXX check if swapping is necessary on BE */
263 		wptr = READ_ONCE(*((u64 *)&adev->wb.wb[ring->wptr_offs]));
264 		DRM_DEBUG("wptr/doorbell before shift == 0x%016llx\n", wptr);
265 	} else {
266 		wptr = RREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI));
267 		wptr = wptr << 32;
268 		wptr |= RREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR));
269 		DRM_DEBUG("wptr before shift [%i] wptr == 0x%016llx\n", ring->me, wptr);
270 	}
271 
272 	return wptr >> 2;
273 }
274 
275 /**
276  * sdma_v5_2_ring_set_wptr - commit the write pointer
277  *
278  * @ring: amdgpu ring pointer
279  *
280  * Write the wptr back to the hardware (NAVI10+).
281  */
282 static void sdma_v5_2_ring_set_wptr(struct amdgpu_ring *ring)
283 {
284 	struct amdgpu_device *adev = ring->adev;
285 
286 	DRM_DEBUG("Setting write pointer\n");
287 	if (ring->use_doorbell) {
288 		DRM_DEBUG("Using doorbell -- "
289 				"wptr_offs == 0x%08x "
290 				"lower_32_bits(ring->wptr) << 2 == 0x%08x "
291 				"upper_32_bits(ring->wptr) << 2 == 0x%08x\n",
292 				ring->wptr_offs,
293 				lower_32_bits(ring->wptr << 2),
294 				upper_32_bits(ring->wptr << 2));
295 		/* XXX check if swapping is necessary on BE */
296 		adev->wb.wb[ring->wptr_offs] = lower_32_bits(ring->wptr << 2);
297 		adev->wb.wb[ring->wptr_offs + 1] = upper_32_bits(ring->wptr << 2);
298 		DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016llx)\n",
299 				ring->doorbell_index, ring->wptr << 2);
300 		WDOORBELL64(ring->doorbell_index, ring->wptr << 2);
301 	} else {
302 		DRM_DEBUG("Not using doorbell -- "
303 				"mmSDMA%i_GFX_RB_WPTR == 0x%08x "
304 				"mmSDMA%i_GFX_RB_WPTR_HI == 0x%08x\n",
305 				ring->me,
306 				lower_32_bits(ring->wptr << 2),
307 				ring->me,
308 				upper_32_bits(ring->wptr << 2));
309 		WREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR),
310 			lower_32_bits(ring->wptr << 2));
311 		WREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI),
312 			upper_32_bits(ring->wptr << 2));
313 	}
314 }
315 
316 static void sdma_v5_2_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
317 {
318 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
319 	int i;
320 
321 	for (i = 0; i < count; i++)
322 		if (sdma && sdma->burst_nop && (i == 0))
323 			amdgpu_ring_write(ring, ring->funcs->nop |
324 				SDMA_PKT_NOP_HEADER_COUNT(count - 1));
325 		else
326 			amdgpu_ring_write(ring, ring->funcs->nop);
327 }
328 
329 /**
330  * sdma_v5_2_ring_emit_ib - Schedule an IB on the DMA engine
331  *
332  * @ring: amdgpu ring pointer
333  * @job: job to retrieve vmid from
334  * @ib: IB object to schedule
335  * @flags: unused
336  *
337  * Schedule an IB in the DMA ring.
338  */
339 static void sdma_v5_2_ring_emit_ib(struct amdgpu_ring *ring,
340 				   struct amdgpu_job *job,
341 				   struct amdgpu_ib *ib,
342 				   uint32_t flags)
343 {
344 	unsigned vmid = AMDGPU_JOB_GET_VMID(job);
345 	uint64_t csa_mc_addr = amdgpu_sdma_get_csa_mc_addr(ring, vmid);
346 
347 	/* An IB packet must end on a 8 DW boundary--the next dword
348 	 * must be on a 8-dword boundary. Our IB packet below is 6
349 	 * dwords long, thus add x number of NOPs, such that, in
350 	 * modular arithmetic,
351 	 * wptr + 6 + x = 8k, k >= 0, which in C is,
352 	 * (wptr + 6 + x) % 8 = 0.
353 	 * The expression below, is a solution of x.
354 	 */
355 	sdma_v5_2_ring_insert_nop(ring, (2 - lower_32_bits(ring->wptr)) & 7);
356 
357 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) |
358 			  SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf));
359 	/* base must be 32 byte aligned */
360 	amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
361 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
362 	amdgpu_ring_write(ring, ib->length_dw);
363 	amdgpu_ring_write(ring, lower_32_bits(csa_mc_addr));
364 	amdgpu_ring_write(ring, upper_32_bits(csa_mc_addr));
365 }
366 
367 /**
368  * sdma_v5_2_ring_emit_mem_sync - flush the IB by graphics cache rinse
369  *
370  * @ring: amdgpu ring pointer
371  * @job: job to retrieve vmid from
372  * @ib: IB object to schedule
373  *
374  * flush the IB by graphics cache rinse.
375  */
376 static void sdma_v5_2_ring_emit_mem_sync(struct amdgpu_ring *ring)
377 {
378 	uint32_t gcr_cntl = SDMA_GCR_GL2_INV | SDMA_GCR_GL2_WB |
379 			    SDMA_GCR_GLM_INV | SDMA_GCR_GL1_INV |
380 			    SDMA_GCR_GLV_INV | SDMA_GCR_GLK_INV |
381 			    SDMA_GCR_GLI_INV(1);
382 
383 	/* flush entire cache L0/L1/L2, this can be optimized by performance requirement */
384 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_GCR_REQ));
385 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD1_BASE_VA_31_7(0));
386 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD2_GCR_CONTROL_15_0(gcr_cntl) |
387 			SDMA_PKT_GCR_REQ_PAYLOAD2_BASE_VA_47_32(0));
388 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD3_LIMIT_VA_31_7(0) |
389 			SDMA_PKT_GCR_REQ_PAYLOAD3_GCR_CONTROL_18_16(gcr_cntl >> 16));
390 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD4_LIMIT_VA_47_32(0) |
391 			SDMA_PKT_GCR_REQ_PAYLOAD4_VMID(0));
392 }
393 
394 /**
395  * sdma_v5_2_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
396  *
397  * @ring: amdgpu ring pointer
398  *
399  * Emit an hdp flush packet on the requested DMA ring.
400  */
401 static void sdma_v5_2_ring_emit_hdp_flush(struct amdgpu_ring *ring)
402 {
403 	struct amdgpu_device *adev = ring->adev;
404 	u32 ref_and_mask = 0;
405 	const struct nbio_hdp_flush_reg *nbio_hf_reg = adev->nbio.hdp_flush_reg;
406 
407 	ref_and_mask = nbio_hf_reg->ref_and_mask_sdma0 << ring->me;
408 
409 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
410 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) |
411 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
412 	amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_done_offset(adev)) << 2);
413 	amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_req_offset(adev)) << 2);
414 	amdgpu_ring_write(ring, ref_and_mask); /* reference */
415 	amdgpu_ring_write(ring, ref_and_mask); /* mask */
416 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
417 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
418 }
419 
420 /**
421  * sdma_v5_2_ring_emit_fence - emit a fence on the DMA ring
422  *
423  * @ring: amdgpu ring pointer
424  * @addr: address
425  * @seq: sequence number
426  * @flags: fence related flags
427  *
428  * Add a DMA fence packet to the ring to write
429  * the fence seq number and DMA trap packet to generate
430  * an interrupt if needed.
431  */
432 static void sdma_v5_2_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
433 				      unsigned flags)
434 {
435 	bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
436 	/* write the fence */
437 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) |
438 			  SDMA_PKT_FENCE_HEADER_MTYPE(0x3)); /* Ucached(UC) */
439 	/* zero in first two bits */
440 	BUG_ON(addr & 0x3);
441 	amdgpu_ring_write(ring, lower_32_bits(addr));
442 	amdgpu_ring_write(ring, upper_32_bits(addr));
443 	amdgpu_ring_write(ring, lower_32_bits(seq));
444 
445 	/* optionally write high bits as well */
446 	if (write64bit) {
447 		addr += 4;
448 		amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) |
449 				  SDMA_PKT_FENCE_HEADER_MTYPE(0x3));
450 		/* zero in first two bits */
451 		BUG_ON(addr & 0x3);
452 		amdgpu_ring_write(ring, lower_32_bits(addr));
453 		amdgpu_ring_write(ring, upper_32_bits(addr));
454 		amdgpu_ring_write(ring, upper_32_bits(seq));
455 	}
456 
457 	if (flags & AMDGPU_FENCE_FLAG_INT) {
458 		/* generate an interrupt */
459 		amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP));
460 		amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(0));
461 	}
462 }
463 
464 
465 /**
466  * sdma_v5_2_gfx_stop - stop the gfx async dma engines
467  *
468  * @adev: amdgpu_device pointer
469  *
470  * Stop the gfx async dma ring buffers.
471  */
472 static void sdma_v5_2_gfx_stop(struct amdgpu_device *adev)
473 {
474 	struct amdgpu_ring *sdma0 = &adev->sdma.instance[0].ring;
475 	struct amdgpu_ring *sdma1 = &adev->sdma.instance[1].ring;
476 	struct amdgpu_ring *sdma2 = &adev->sdma.instance[2].ring;
477 	struct amdgpu_ring *sdma3 = &adev->sdma.instance[3].ring;
478 	u32 rb_cntl, ib_cntl;
479 	int i;
480 
481 	if ((adev->mman.buffer_funcs_ring == sdma0) ||
482 	    (adev->mman.buffer_funcs_ring == sdma1) ||
483 	    (adev->mman.buffer_funcs_ring == sdma2) ||
484 	    (adev->mman.buffer_funcs_ring == sdma3))
485 		amdgpu_ttm_set_buffer_funcs_status(adev, false);
486 
487 	for (i = 0; i < adev->sdma.num_instances; i++) {
488 		rb_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL));
489 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 0);
490 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
491 		ib_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL));
492 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 0);
493 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl);
494 	}
495 }
496 
497 /**
498  * sdma_v5_2_rlc_stop - stop the compute async dma engines
499  *
500  * @adev: amdgpu_device pointer
501  *
502  * Stop the compute async dma queues.
503  */
504 static void sdma_v5_2_rlc_stop(struct amdgpu_device *adev)
505 {
506 	/* XXX todo */
507 }
508 
509 /**
510  * sdma_v5_2_ctx_switch_enable - stop the async dma engines context switch
511  *
512  * @adev: amdgpu_device pointer
513  * @enable: enable/disable the DMA MEs context switch.
514  *
515  * Halt or unhalt the async dma engines context switch.
516  */
517 static void sdma_v5_2_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
518 {
519 	u32 f32_cntl, phase_quantum = 0;
520 	int i;
521 
522 	if (amdgpu_sdma_phase_quantum) {
523 		unsigned value = amdgpu_sdma_phase_quantum;
524 		unsigned unit = 0;
525 
526 		while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
527 				SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) {
528 			value = (value + 1) >> 1;
529 			unit++;
530 		}
531 		if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
532 			    SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) {
533 			value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
534 				 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT);
535 			unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
536 				SDMA0_PHASE0_QUANTUM__UNIT__SHIFT);
537 			WARN_ONCE(1,
538 			"clamping sdma_phase_quantum to %uK clock cycles\n",
539 				  value << unit);
540 		}
541 		phase_quantum =
542 			value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
543 			unit  << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
544 	}
545 
546 	for (i = 0; i < adev->sdma.num_instances; i++) {
547 		f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
548 		f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
549 				AUTO_CTXSW_ENABLE, enable ? 1 : 0);
550 		if (enable && amdgpu_sdma_phase_quantum) {
551 			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE0_QUANTUM),
552 			       phase_quantum);
553 			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE1_QUANTUM),
554 			       phase_quantum);
555 			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE2_QUANTUM),
556 			       phase_quantum);
557 		}
558 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl);
559 	}
560 
561 }
562 
563 /**
564  * sdma_v5_2_enable - stop the async dma engines
565  *
566  * @adev: amdgpu_device pointer
567  * @enable: enable/disable the DMA MEs.
568  *
569  * Halt or unhalt the async dma engines.
570  */
571 static void sdma_v5_2_enable(struct amdgpu_device *adev, bool enable)
572 {
573 	u32 f32_cntl;
574 	int i;
575 
576 	if (!enable) {
577 		sdma_v5_2_gfx_stop(adev);
578 		sdma_v5_2_rlc_stop(adev);
579 	}
580 
581 	for (i = 0; i < adev->sdma.num_instances; i++) {
582 		f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
583 		f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, enable ? 0 : 1);
584 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl);
585 	}
586 }
587 
588 /**
589  * sdma_v5_2_gfx_resume - setup and start the async dma engines
590  *
591  * @adev: amdgpu_device pointer
592  *
593  * Set up the gfx DMA ring buffers and enable them.
594  * Returns 0 for success, error for failure.
595  */
596 static int sdma_v5_2_gfx_resume(struct amdgpu_device *adev)
597 {
598 	struct amdgpu_ring *ring;
599 	u32 rb_cntl, ib_cntl;
600 	u32 rb_bufsz;
601 	u32 wb_offset;
602 	u32 doorbell;
603 	u32 doorbell_offset;
604 	u32 temp;
605 	u32 wptr_poll_cntl;
606 	u64 wptr_gpu_addr;
607 	int i, r;
608 
609 	for (i = 0; i < adev->sdma.num_instances; i++) {
610 		ring = &adev->sdma.instance[i].ring;
611 		wb_offset = (ring->rptr_offs * 4);
612 
613 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL), 0);
614 
615 		/* Set ring buffer size in dwords */
616 		rb_bufsz = order_base_2(ring->ring_size / 4);
617 		rb_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL));
618 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz);
619 #ifdef __BIG_ENDIAN
620 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1);
621 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL,
622 					RPTR_WRITEBACK_SWAP_ENABLE, 1);
623 #endif
624 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
625 
626 		/* Initialize the ring buffer's read and write pointers */
627 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR), 0);
628 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_HI), 0);
629 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), 0);
630 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), 0);
631 
632 		/* setup the wptr shadow polling */
633 		wptr_gpu_addr = adev->wb.gpu_addr + (ring->wptr_offs * 4);
634 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_LO),
635 		       lower_32_bits(wptr_gpu_addr));
636 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_HI),
637 		       upper_32_bits(wptr_gpu_addr));
638 		wptr_poll_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i,
639 							 mmSDMA0_GFX_RB_WPTR_POLL_CNTL));
640 		wptr_poll_cntl = REG_SET_FIELD(wptr_poll_cntl,
641 					       SDMA0_GFX_RB_WPTR_POLL_CNTL,
642 					       F32_POLL_ENABLE, 1);
643 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_CNTL),
644 		       wptr_poll_cntl);
645 
646 		/* set the wb address whether it's enabled or not */
647 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_HI),
648 		       upper_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFF);
649 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_LO),
650 		       lower_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC);
651 
652 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
653 
654 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE), ring->gpu_addr >> 8);
655 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE_HI), ring->gpu_addr >> 40);
656 
657 		ring->wptr = 0;
658 
659 		/* before programing wptr to a less value, need set minor_ptr_update first */
660 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 1);
661 
662 		if (!amdgpu_sriov_vf(adev)) { /* only bare-metal use register write for wptr */
663 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), lower_32_bits(ring->wptr) << 2);
664 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), upper_32_bits(ring->wptr) << 2);
665 		}
666 
667 		doorbell = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL));
668 		doorbell_offset = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL_OFFSET));
669 
670 		if (ring->use_doorbell) {
671 			doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 1);
672 			doorbell_offset = REG_SET_FIELD(doorbell_offset, SDMA0_GFX_DOORBELL_OFFSET,
673 					OFFSET, ring->doorbell_index);
674 		} else {
675 			doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 0);
676 		}
677 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL), doorbell);
678 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL_OFFSET), doorbell_offset);
679 
680 		adev->nbio.funcs->sdma_doorbell_range(adev, i, ring->use_doorbell,
681 						      ring->doorbell_index,
682 						      adev->doorbell_index.sdma_doorbell_range);
683 
684 		if (amdgpu_sriov_vf(adev))
685 			sdma_v5_2_ring_set_wptr(ring);
686 
687 		/* set minor_ptr_update to 0 after wptr programed */
688 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 0);
689 
690 		/* set utc l1 enable flag always to 1 */
691 		temp = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
692 		temp = REG_SET_FIELD(temp, SDMA0_CNTL, UTC_L1_ENABLE, 1);
693 
694 		/* enable MCBP */
695 		temp = REG_SET_FIELD(temp, SDMA0_CNTL, MIDCMD_PREEMPT_ENABLE, 1);
696 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), temp);
697 
698 		/* Set up RESP_MODE to non-copy addresses */
699 		temp = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL));
700 		temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, RESP_MODE, 3);
701 		temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, REDO_DELAY, 9);
702 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL), temp);
703 
704 		/* program default cache read and write policy */
705 		temp = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE));
706 		/* clean read policy and write policy bits */
707 		temp &= 0xFF0FFF;
708 		temp |= ((CACHE_READ_POLICY_L2__DEFAULT << 12) |
709 			 (CACHE_WRITE_POLICY_L2__DEFAULT << 14) |
710 			 SDMA0_UTCL1_PAGE__LLC_NOALLOC_MASK);
711 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE), temp);
712 
713 		if (!amdgpu_sriov_vf(adev)) {
714 			/* unhalt engine */
715 			temp = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
716 			temp = REG_SET_FIELD(temp, SDMA0_F32_CNTL, HALT, 0);
717 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), temp);
718 		}
719 
720 		/* enable DMA RB */
721 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1);
722 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
723 
724 		ib_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL));
725 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 1);
726 #ifdef __BIG_ENDIAN
727 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1);
728 #endif
729 		/* enable DMA IBs */
730 		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl);
731 
732 		ring->sched.ready = true;
733 
734 		if (amdgpu_sriov_vf(adev)) { /* bare-metal sequence doesn't need below to lines */
735 			sdma_v5_2_ctx_switch_enable(adev, true);
736 			sdma_v5_2_enable(adev, true);
737 		}
738 
739 		r = amdgpu_ring_test_ring(ring);
740 		if (r) {
741 			ring->sched.ready = false;
742 			return r;
743 		}
744 
745 		if (adev->mman.buffer_funcs_ring == ring)
746 			amdgpu_ttm_set_buffer_funcs_status(adev, true);
747 	}
748 
749 	return 0;
750 }
751 
752 /**
753  * sdma_v5_2_rlc_resume - setup and start the async dma engines
754  *
755  * @adev: amdgpu_device pointer
756  *
757  * Set up the compute DMA queues and enable them.
758  * Returns 0 for success, error for failure.
759  */
760 static int sdma_v5_2_rlc_resume(struct amdgpu_device *adev)
761 {
762 	return 0;
763 }
764 
765 /**
766  * sdma_v5_2_load_microcode - load the sDMA ME ucode
767  *
768  * @adev: amdgpu_device pointer
769  *
770  * Loads the sDMA0/1/2/3 ucode.
771  * Returns 0 for success, -EINVAL if the ucode is not available.
772  */
773 static int sdma_v5_2_load_microcode(struct amdgpu_device *adev)
774 {
775 	const struct sdma_firmware_header_v1_0 *hdr;
776 	const __le32 *fw_data;
777 	u32 fw_size;
778 	int i, j;
779 
780 	/* halt the MEs */
781 	sdma_v5_2_enable(adev, false);
782 
783 	for (i = 0; i < adev->sdma.num_instances; i++) {
784 		if (!adev->sdma.instance[i].fw)
785 			return -EINVAL;
786 
787 		hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
788 		amdgpu_ucode_print_sdma_hdr(&hdr->header);
789 		fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
790 
791 		fw_data = (const __le32 *)
792 			(adev->sdma.instance[i].fw->data +
793 				le32_to_cpu(hdr->header.ucode_array_offset_bytes));
794 
795 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), 0);
796 
797 		for (j = 0; j < fw_size; j++) {
798 			if (amdgpu_emu_mode == 1 && j % 500 == 0)
799 				msleep(1);
800 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UCODE_DATA), le32_to_cpup(fw_data++));
801 		}
802 
803 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), adev->sdma.instance[i].fw_version);
804 	}
805 
806 	return 0;
807 }
808 
809 static int sdma_v5_2_soft_reset(void *handle)
810 {
811 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
812 	u32 grbm_soft_reset;
813 	u32 tmp;
814 	int i;
815 
816 	for (i = 0; i < adev->sdma.num_instances; i++) {
817 		grbm_soft_reset = REG_SET_FIELD(0,
818 						GRBM_SOFT_RESET, SOFT_RESET_SDMA0,
819 						1);
820 		grbm_soft_reset <<= i;
821 
822 		tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET);
823 		tmp |= grbm_soft_reset;
824 		DRM_DEBUG("GRBM_SOFT_RESET=0x%08X\n", tmp);
825 		WREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET, tmp);
826 		tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET);
827 
828 		udelay(50);
829 
830 		tmp &= ~grbm_soft_reset;
831 		WREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET, tmp);
832 		tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET);
833 
834 		udelay(50);
835 	}
836 
837 	return 0;
838 }
839 
840 /**
841  * sdma_v5_2_start - setup and start the async dma engines
842  *
843  * @adev: amdgpu_device pointer
844  *
845  * Set up the DMA engines and enable them.
846  * Returns 0 for success, error for failure.
847  */
848 static int sdma_v5_2_start(struct amdgpu_device *adev)
849 {
850 	int r = 0;
851 
852 	if (amdgpu_sriov_vf(adev)) {
853 		sdma_v5_2_ctx_switch_enable(adev, false);
854 		sdma_v5_2_enable(adev, false);
855 
856 		/* set RB registers */
857 		r = sdma_v5_2_gfx_resume(adev);
858 		return r;
859 	}
860 
861 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
862 		r = sdma_v5_2_load_microcode(adev);
863 		if (r)
864 			return r;
865 
866 		/* The value of mmSDMA_F32_CNTL is invalid the moment after loading fw */
867 		if (amdgpu_emu_mode == 1)
868 			msleep(1000);
869 	}
870 
871 	/* TODO: check whether can submit a doorbell request to raise
872 	 * a doorbell fence to exit gfxoff.
873 	 */
874 	if (adev->in_s0ix)
875 		amdgpu_gfx_off_ctrl(adev, false);
876 
877 	sdma_v5_2_soft_reset(adev);
878 	/* unhalt the MEs */
879 	sdma_v5_2_enable(adev, true);
880 	/* enable sdma ring preemption */
881 	sdma_v5_2_ctx_switch_enable(adev, true);
882 
883 	/* start the gfx rings and rlc compute queues */
884 	r = sdma_v5_2_gfx_resume(adev);
885 	if (adev->in_s0ix)
886 		amdgpu_gfx_off_ctrl(adev, true);
887 	if (r)
888 		return r;
889 	r = sdma_v5_2_rlc_resume(adev);
890 
891 	return r;
892 }
893 
894 /**
895  * sdma_v5_2_ring_test_ring - simple async dma engine test
896  *
897  * @ring: amdgpu_ring structure holding ring information
898  *
899  * Test the DMA engine by writing using it to write an
900  * value to memory.
901  * Returns 0 for success, error for failure.
902  */
903 static int sdma_v5_2_ring_test_ring(struct amdgpu_ring *ring)
904 {
905 	struct amdgpu_device *adev = ring->adev;
906 	unsigned i;
907 	unsigned index;
908 	int r;
909 	u32 tmp;
910 	u64 gpu_addr;
911 
912 	r = amdgpu_device_wb_get(adev, &index);
913 	if (r) {
914 		dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
915 		return r;
916 	}
917 
918 	gpu_addr = adev->wb.gpu_addr + (index * 4);
919 	tmp = 0xCAFEDEAD;
920 	adev->wb.wb[index] = cpu_to_le32(tmp);
921 
922 	r = amdgpu_ring_alloc(ring, 5);
923 	if (r) {
924 		DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
925 		amdgpu_device_wb_free(adev, index);
926 		return r;
927 	}
928 
929 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
930 			  SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
931 	amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
932 	amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
933 	amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0));
934 	amdgpu_ring_write(ring, 0xDEADBEEF);
935 	amdgpu_ring_commit(ring);
936 
937 	for (i = 0; i < adev->usec_timeout; i++) {
938 		tmp = le32_to_cpu(adev->wb.wb[index]);
939 		if (tmp == 0xDEADBEEF)
940 			break;
941 		if (amdgpu_emu_mode == 1)
942 			msleep(1);
943 		else
944 			udelay(1);
945 	}
946 
947 	if (i >= adev->usec_timeout)
948 		r = -ETIMEDOUT;
949 
950 	amdgpu_device_wb_free(adev, index);
951 
952 	return r;
953 }
954 
955 /**
956  * sdma_v5_2_ring_test_ib - test an IB on the DMA engine
957  *
958  * @ring: amdgpu_ring structure holding ring information
959  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
960  *
961  * Test a simple IB in the DMA ring.
962  * Returns 0 on success, error on failure.
963  */
964 static int sdma_v5_2_ring_test_ib(struct amdgpu_ring *ring, long timeout)
965 {
966 	struct amdgpu_device *adev = ring->adev;
967 	struct amdgpu_ib ib;
968 	struct dma_fence *f = NULL;
969 	unsigned index;
970 	long r;
971 	u32 tmp = 0;
972 	u64 gpu_addr;
973 
974 	r = amdgpu_device_wb_get(adev, &index);
975 	if (r) {
976 		dev_err(adev->dev, "(%ld) failed to allocate wb slot\n", r);
977 		return r;
978 	}
979 
980 	gpu_addr = adev->wb.gpu_addr + (index * 4);
981 	tmp = 0xCAFEDEAD;
982 	adev->wb.wb[index] = cpu_to_le32(tmp);
983 	memset(&ib, 0, sizeof(ib));
984 	r = amdgpu_ib_get(adev, NULL, 256, AMDGPU_IB_POOL_DIRECT, &ib);
985 	if (r) {
986 		DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r);
987 		goto err0;
988 	}
989 
990 	ib.ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
991 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
992 	ib.ptr[1] = lower_32_bits(gpu_addr);
993 	ib.ptr[2] = upper_32_bits(gpu_addr);
994 	ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0);
995 	ib.ptr[4] = 0xDEADBEEF;
996 	ib.ptr[5] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
997 	ib.ptr[6] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
998 	ib.ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
999 	ib.length_dw = 8;
1000 
1001 	r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
1002 	if (r)
1003 		goto err1;
1004 
1005 	r = dma_fence_wait_timeout(f, false, timeout);
1006 	if (r == 0) {
1007 		DRM_ERROR("amdgpu: IB test timed out\n");
1008 		r = -ETIMEDOUT;
1009 		goto err1;
1010 	} else if (r < 0) {
1011 		DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
1012 		goto err1;
1013 	}
1014 	tmp = le32_to_cpu(adev->wb.wb[index]);
1015 	if (tmp == 0xDEADBEEF)
1016 		r = 0;
1017 	else
1018 		r = -EINVAL;
1019 
1020 err1:
1021 	amdgpu_ib_free(adev, &ib, NULL);
1022 	dma_fence_put(f);
1023 err0:
1024 	amdgpu_device_wb_free(adev, index);
1025 	return r;
1026 }
1027 
1028 
1029 /**
1030  * sdma_v5_2_vm_copy_pte - update PTEs by copying them from the GART
1031  *
1032  * @ib: indirect buffer to fill with commands
1033  * @pe: addr of the page entry
1034  * @src: src addr to copy from
1035  * @count: number of page entries to update
1036  *
1037  * Update PTEs by copying them from the GART using sDMA.
1038  */
1039 static void sdma_v5_2_vm_copy_pte(struct amdgpu_ib *ib,
1040 				  uint64_t pe, uint64_t src,
1041 				  unsigned count)
1042 {
1043 	unsigned bytes = count * 8;
1044 
1045 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1046 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
1047 	ib->ptr[ib->length_dw++] = bytes - 1;
1048 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1049 	ib->ptr[ib->length_dw++] = lower_32_bits(src);
1050 	ib->ptr[ib->length_dw++] = upper_32_bits(src);
1051 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1052 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1053 
1054 }
1055 
1056 /**
1057  * sdma_v5_2_vm_write_pte - update PTEs by writing them manually
1058  *
1059  * @ib: indirect buffer to fill with commands
1060  * @pe: addr of the page entry
1061  * @value: dst addr to write into pe
1062  * @count: number of page entries to update
1063  * @incr: increase next addr by incr bytes
1064  *
1065  * Update PTEs by writing them manually using sDMA.
1066  */
1067 static void sdma_v5_2_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
1068 				   uint64_t value, unsigned count,
1069 				   uint32_t incr)
1070 {
1071 	unsigned ndw = count * 2;
1072 
1073 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
1074 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1075 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1076 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1077 	ib->ptr[ib->length_dw++] = ndw - 1;
1078 	for (; ndw > 0; ndw -= 2) {
1079 		ib->ptr[ib->length_dw++] = lower_32_bits(value);
1080 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
1081 		value += incr;
1082 	}
1083 }
1084 
1085 /**
1086  * sdma_v5_2_vm_set_pte_pde - update the page tables using sDMA
1087  *
1088  * @ib: indirect buffer to fill with commands
1089  * @pe: addr of the page entry
1090  * @addr: dst addr to write into pe
1091  * @count: number of page entries to update
1092  * @incr: increase next addr by incr bytes
1093  * @flags: access flags
1094  *
1095  * Update the page tables using sDMA.
1096  */
1097 static void sdma_v5_2_vm_set_pte_pde(struct amdgpu_ib *ib,
1098 				     uint64_t pe,
1099 				     uint64_t addr, unsigned count,
1100 				     uint32_t incr, uint64_t flags)
1101 {
1102 	/* for physically contiguous pages (vram) */
1103 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_PTEPDE);
1104 	ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
1105 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1106 	ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
1107 	ib->ptr[ib->length_dw++] = upper_32_bits(flags);
1108 	ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
1109 	ib->ptr[ib->length_dw++] = upper_32_bits(addr);
1110 	ib->ptr[ib->length_dw++] = incr; /* increment size */
1111 	ib->ptr[ib->length_dw++] = 0;
1112 	ib->ptr[ib->length_dw++] = count - 1; /* number of entries */
1113 }
1114 
1115 /**
1116  * sdma_v5_2_ring_pad_ib - pad the IB
1117  *
1118  * @ib: indirect buffer to fill with padding
1119  * @ring: amdgpu_ring structure holding ring information
1120  *
1121  * Pad the IB with NOPs to a boundary multiple of 8.
1122  */
1123 static void sdma_v5_2_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
1124 {
1125 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
1126 	u32 pad_count;
1127 	int i;
1128 
1129 	pad_count = (-ib->length_dw) & 0x7;
1130 	for (i = 0; i < pad_count; i++)
1131 		if (sdma && sdma->burst_nop && (i == 0))
1132 			ib->ptr[ib->length_dw++] =
1133 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP) |
1134 				SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
1135 		else
1136 			ib->ptr[ib->length_dw++] =
1137 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
1138 }
1139 
1140 
1141 /**
1142  * sdma_v5_2_ring_emit_pipeline_sync - sync the pipeline
1143  *
1144  * @ring: amdgpu_ring pointer
1145  *
1146  * Make sure all previous operations are completed (CIK).
1147  */
1148 static void sdma_v5_2_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
1149 {
1150 	uint32_t seq = ring->fence_drv.sync_seq;
1151 	uint64_t addr = ring->fence_drv.gpu_addr;
1152 
1153 	/* wait for idle */
1154 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1155 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1156 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
1157 			  SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1));
1158 	amdgpu_ring_write(ring, addr & 0xfffffffc);
1159 	amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
1160 	amdgpu_ring_write(ring, seq); /* reference */
1161 	amdgpu_ring_write(ring, 0xffffffff); /* mask */
1162 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1163 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */
1164 }
1165 
1166 
1167 /**
1168  * sdma_v5_2_ring_emit_vm_flush - vm flush using sDMA
1169  *
1170  * @ring: amdgpu_ring pointer
1171  * @vmid: vmid number to use
1172  * @pd_addr: address
1173  *
1174  * Update the page table base and flush the VM TLB
1175  * using sDMA.
1176  */
1177 static void sdma_v5_2_ring_emit_vm_flush(struct amdgpu_ring *ring,
1178 					 unsigned vmid, uint64_t pd_addr)
1179 {
1180 	amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
1181 }
1182 
1183 static void sdma_v5_2_ring_emit_wreg(struct amdgpu_ring *ring,
1184 				     uint32_t reg, uint32_t val)
1185 {
1186 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
1187 			  SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
1188 	amdgpu_ring_write(ring, reg);
1189 	amdgpu_ring_write(ring, val);
1190 }
1191 
1192 static void sdma_v5_2_ring_emit_reg_wait(struct amdgpu_ring *ring, uint32_t reg,
1193 					 uint32_t val, uint32_t mask)
1194 {
1195 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1196 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1197 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* equal */
1198 	amdgpu_ring_write(ring, reg << 2);
1199 	amdgpu_ring_write(ring, 0);
1200 	amdgpu_ring_write(ring, val); /* reference */
1201 	amdgpu_ring_write(ring, mask); /* mask */
1202 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1203 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10));
1204 }
1205 
1206 static void sdma_v5_2_ring_emit_reg_write_reg_wait(struct amdgpu_ring *ring,
1207 						   uint32_t reg0, uint32_t reg1,
1208 						   uint32_t ref, uint32_t mask)
1209 {
1210 	amdgpu_ring_emit_wreg(ring, reg0, ref);
1211 	/* wait for a cycle to reset vm_inv_eng*_ack */
1212 	amdgpu_ring_emit_reg_wait(ring, reg0, 0, 0);
1213 	amdgpu_ring_emit_reg_wait(ring, reg1, mask, mask);
1214 }
1215 
1216 static int sdma_v5_2_early_init(void *handle)
1217 {
1218 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1219 
1220 	sdma_v5_2_set_ring_funcs(adev);
1221 	sdma_v5_2_set_buffer_funcs(adev);
1222 	sdma_v5_2_set_vm_pte_funcs(adev);
1223 	sdma_v5_2_set_irq_funcs(adev);
1224 
1225 	return 0;
1226 }
1227 
1228 static unsigned sdma_v5_2_seq_to_irq_id(int seq_num)
1229 {
1230 	switch (seq_num) {
1231 	case 0:
1232 		return SOC15_IH_CLIENTID_SDMA0;
1233 	case 1:
1234 		return SOC15_IH_CLIENTID_SDMA1;
1235 	case 2:
1236 		return SOC15_IH_CLIENTID_SDMA2;
1237 	case 3:
1238 		return SOC15_IH_CLIENTID_SDMA3_Sienna_Cichlid;
1239 	default:
1240 		break;
1241 	}
1242 	return -EINVAL;
1243 }
1244 
1245 static unsigned sdma_v5_2_seq_to_trap_id(int seq_num)
1246 {
1247 	switch (seq_num) {
1248 	case 0:
1249 		return SDMA0_5_0__SRCID__SDMA_TRAP;
1250 	case 1:
1251 		return SDMA1_5_0__SRCID__SDMA_TRAP;
1252 	case 2:
1253 		return SDMA2_5_0__SRCID__SDMA_TRAP;
1254 	case 3:
1255 		return SDMA3_5_0__SRCID__SDMA_TRAP;
1256 	default:
1257 		break;
1258 	}
1259 	return -EINVAL;
1260 }
1261 
1262 static int sdma_v5_2_sw_init(void *handle)
1263 {
1264 	struct amdgpu_ring *ring;
1265 	int r, i;
1266 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1267 
1268 	/* SDMA trap event */
1269 	for (i = 0; i < adev->sdma.num_instances; i++) {
1270 		r = amdgpu_irq_add_id(adev, sdma_v5_2_seq_to_irq_id(i),
1271 				      sdma_v5_2_seq_to_trap_id(i),
1272 				      &adev->sdma.trap_irq);
1273 		if (r)
1274 			return r;
1275 	}
1276 
1277 	r = sdma_v5_2_init_microcode(adev);
1278 	if (r) {
1279 		DRM_ERROR("Failed to load sdma firmware!\n");
1280 		return r;
1281 	}
1282 
1283 	for (i = 0; i < adev->sdma.num_instances; i++) {
1284 		ring = &adev->sdma.instance[i].ring;
1285 		ring->ring_obj = NULL;
1286 		ring->use_doorbell = true;
1287 		ring->me = i;
1288 
1289 		DRM_INFO("use_doorbell being set to: [%s]\n",
1290 				ring->use_doorbell?"true":"false");
1291 
1292 		ring->doorbell_index =
1293 			(adev->doorbell_index.sdma_engine[i] << 1); //get DWORD offset
1294 
1295 		sprintf(ring->name, "sdma%d", i);
1296 		r = amdgpu_ring_init(adev, ring, 1024, &adev->sdma.trap_irq,
1297 				     AMDGPU_SDMA_IRQ_INSTANCE0 + i,
1298 				     AMDGPU_RING_PRIO_DEFAULT, NULL);
1299 		if (r)
1300 			return r;
1301 	}
1302 
1303 	return r;
1304 }
1305 
1306 static int sdma_v5_2_sw_fini(void *handle)
1307 {
1308 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1309 	int i;
1310 
1311 	for (i = 0; i < adev->sdma.num_instances; i++)
1312 		amdgpu_ring_fini(&adev->sdma.instance[i].ring);
1313 
1314 	sdma_v5_2_destroy_inst_ctx(adev);
1315 
1316 	return 0;
1317 }
1318 
1319 static int sdma_v5_2_hw_init(void *handle)
1320 {
1321 	int r;
1322 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1323 
1324 	r = sdma_v5_2_start(adev);
1325 
1326 	return r;
1327 }
1328 
1329 static int sdma_v5_2_hw_fini(void *handle)
1330 {
1331 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1332 
1333 	if (amdgpu_sriov_vf(adev))
1334 		return 0;
1335 
1336 	sdma_v5_2_ctx_switch_enable(adev, false);
1337 	sdma_v5_2_enable(adev, false);
1338 
1339 	return 0;
1340 }
1341 
1342 static int sdma_v5_2_suspend(void *handle)
1343 {
1344 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1345 
1346 	return sdma_v5_2_hw_fini(adev);
1347 }
1348 
1349 static int sdma_v5_2_resume(void *handle)
1350 {
1351 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1352 
1353 	return sdma_v5_2_hw_init(adev);
1354 }
1355 
1356 static bool sdma_v5_2_is_idle(void *handle)
1357 {
1358 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1359 	u32 i;
1360 
1361 	for (i = 0; i < adev->sdma.num_instances; i++) {
1362 		u32 tmp = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_STATUS_REG));
1363 
1364 		if (!(tmp & SDMA0_STATUS_REG__IDLE_MASK))
1365 			return false;
1366 	}
1367 
1368 	return true;
1369 }
1370 
1371 static int sdma_v5_2_wait_for_idle(void *handle)
1372 {
1373 	unsigned i;
1374 	u32 sdma0, sdma1, sdma2, sdma3;
1375 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1376 
1377 	for (i = 0; i < adev->usec_timeout; i++) {
1378 		sdma0 = RREG32(sdma_v5_2_get_reg_offset(adev, 0, mmSDMA0_STATUS_REG));
1379 		sdma1 = RREG32(sdma_v5_2_get_reg_offset(adev, 1, mmSDMA0_STATUS_REG));
1380 		sdma2 = RREG32(sdma_v5_2_get_reg_offset(adev, 2, mmSDMA0_STATUS_REG));
1381 		sdma3 = RREG32(sdma_v5_2_get_reg_offset(adev, 3, mmSDMA0_STATUS_REG));
1382 
1383 		if (sdma0 & sdma1 & sdma2 & sdma3 & SDMA0_STATUS_REG__IDLE_MASK)
1384 			return 0;
1385 		udelay(1);
1386 	}
1387 	return -ETIMEDOUT;
1388 }
1389 
1390 static int sdma_v5_2_ring_preempt_ib(struct amdgpu_ring *ring)
1391 {
1392 	int i, r = 0;
1393 	struct amdgpu_device *adev = ring->adev;
1394 	u32 index = 0;
1395 	u64 sdma_gfx_preempt;
1396 
1397 	amdgpu_sdma_get_index_from_ring(ring, &index);
1398 	sdma_gfx_preempt =
1399 		sdma_v5_2_get_reg_offset(adev, index, mmSDMA0_GFX_PREEMPT);
1400 
1401 	/* assert preemption condition */
1402 	amdgpu_ring_set_preempt_cond_exec(ring, false);
1403 
1404 	/* emit the trailing fence */
1405 	ring->trail_seq += 1;
1406 	amdgpu_ring_alloc(ring, 10);
1407 	sdma_v5_2_ring_emit_fence(ring, ring->trail_fence_gpu_addr,
1408 				  ring->trail_seq, 0);
1409 	amdgpu_ring_commit(ring);
1410 
1411 	/* assert IB preemption */
1412 	WREG32(sdma_gfx_preempt, 1);
1413 
1414 	/* poll the trailing fence */
1415 	for (i = 0; i < adev->usec_timeout; i++) {
1416 		if (ring->trail_seq ==
1417 		    le32_to_cpu(*(ring->trail_fence_cpu_addr)))
1418 			break;
1419 		udelay(1);
1420 	}
1421 
1422 	if (i >= adev->usec_timeout) {
1423 		r = -EINVAL;
1424 		DRM_ERROR("ring %d failed to be preempted\n", ring->idx);
1425 	}
1426 
1427 	/* deassert IB preemption */
1428 	WREG32(sdma_gfx_preempt, 0);
1429 
1430 	/* deassert the preemption condition */
1431 	amdgpu_ring_set_preempt_cond_exec(ring, true);
1432 	return r;
1433 }
1434 
1435 static int sdma_v5_2_set_trap_irq_state(struct amdgpu_device *adev,
1436 					struct amdgpu_irq_src *source,
1437 					unsigned type,
1438 					enum amdgpu_interrupt_state state)
1439 {
1440 	u32 sdma_cntl;
1441 
1442 	u32 reg_offset = sdma_v5_2_get_reg_offset(adev, type, mmSDMA0_CNTL);
1443 
1444 	sdma_cntl = RREG32(reg_offset);
1445 	sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE,
1446 		       state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0);
1447 	WREG32(reg_offset, sdma_cntl);
1448 
1449 	return 0;
1450 }
1451 
1452 static int sdma_v5_2_process_trap_irq(struct amdgpu_device *adev,
1453 				      struct amdgpu_irq_src *source,
1454 				      struct amdgpu_iv_entry *entry)
1455 {
1456 	DRM_DEBUG("IH: SDMA trap\n");
1457 	switch (entry->client_id) {
1458 	case SOC15_IH_CLIENTID_SDMA0:
1459 		switch (entry->ring_id) {
1460 		case 0:
1461 			amdgpu_fence_process(&adev->sdma.instance[0].ring);
1462 			break;
1463 		case 1:
1464 			/* XXX compute */
1465 			break;
1466 		case 2:
1467 			/* XXX compute */
1468 			break;
1469 		case 3:
1470 			/* XXX page queue*/
1471 			break;
1472 		}
1473 		break;
1474 	case SOC15_IH_CLIENTID_SDMA1:
1475 		switch (entry->ring_id) {
1476 		case 0:
1477 			amdgpu_fence_process(&adev->sdma.instance[1].ring);
1478 			break;
1479 		case 1:
1480 			/* XXX compute */
1481 			break;
1482 		case 2:
1483 			/* XXX compute */
1484 			break;
1485 		case 3:
1486 			/* XXX page queue*/
1487 			break;
1488 		}
1489 		break;
1490 	case SOC15_IH_CLIENTID_SDMA2:
1491 		switch (entry->ring_id) {
1492 		case 0:
1493 			amdgpu_fence_process(&adev->sdma.instance[2].ring);
1494 			break;
1495 		case 1:
1496 			/* XXX compute */
1497 			break;
1498 		case 2:
1499 			/* XXX compute */
1500 			break;
1501 		case 3:
1502 			/* XXX page queue*/
1503 			break;
1504 		}
1505 		break;
1506 	case SOC15_IH_CLIENTID_SDMA3_Sienna_Cichlid:
1507 		switch (entry->ring_id) {
1508 		case 0:
1509 			amdgpu_fence_process(&adev->sdma.instance[3].ring);
1510 			break;
1511 		case 1:
1512 			/* XXX compute */
1513 			break;
1514 		case 2:
1515 			/* XXX compute */
1516 			break;
1517 		case 3:
1518 			/* XXX page queue*/
1519 			break;
1520 		}
1521 		break;
1522 	}
1523 	return 0;
1524 }
1525 
1526 static int sdma_v5_2_process_illegal_inst_irq(struct amdgpu_device *adev,
1527 					      struct amdgpu_irq_src *source,
1528 					      struct amdgpu_iv_entry *entry)
1529 {
1530 	return 0;
1531 }
1532 
1533 static void sdma_v5_2_update_medium_grain_clock_gating(struct amdgpu_device *adev,
1534 						       bool enable)
1535 {
1536 	uint32_t data, def;
1537 	int i;
1538 
1539 	for (i = 0; i < adev->sdma.num_instances; i++) {
1540 
1541 		if (adev->sdma.instance[i].fw_version < 70 && adev->ip_versions[SDMA0_HWIP][0] == IP_VERSION(5, 2, 1))
1542 			adev->cg_flags &= ~AMD_CG_SUPPORT_SDMA_MGCG;
1543 
1544 		if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
1545 			/* Enable sdma clock gating */
1546 			def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL));
1547 			data &= ~(SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1548 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1549 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1550 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1551 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK |
1552 				  SDMA0_CLK_CTRL__SOFT_OVERRIDER_REG_MASK);
1553 			if (def != data)
1554 				WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data);
1555 		} else {
1556 			/* Disable sdma clock gating */
1557 			def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL));
1558 			data |= (SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1559 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1560 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1561 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1562 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK |
1563 				 SDMA0_CLK_CTRL__SOFT_OVERRIDER_REG_MASK);
1564 			if (def != data)
1565 				WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data);
1566 		}
1567 	}
1568 }
1569 
1570 static void sdma_v5_2_update_medium_grain_light_sleep(struct amdgpu_device *adev,
1571 						      bool enable)
1572 {
1573 	uint32_t data, def;
1574 	int i;
1575 
1576 	for (i = 0; i < adev->sdma.num_instances; i++) {
1577 
1578 		if (adev->sdma.instance[i].fw_version < 70 && adev->ip_versions[SDMA0_HWIP][0] == IP_VERSION(5, 2, 1))
1579 			adev->cg_flags &= ~AMD_CG_SUPPORT_SDMA_LS;
1580 
1581 		if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) {
1582 			/* Enable sdma mem light sleep */
1583 			def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL));
1584 			data |= SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1585 			if (def != data)
1586 				WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data);
1587 
1588 		} else {
1589 			/* Disable sdma mem light sleep */
1590 			def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL));
1591 			data &= ~SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1592 			if (def != data)
1593 				WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data);
1594 
1595 		}
1596 	}
1597 }
1598 
1599 static int sdma_v5_2_set_clockgating_state(void *handle,
1600 					   enum amd_clockgating_state state)
1601 {
1602 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1603 
1604 	if (amdgpu_sriov_vf(adev))
1605 		return 0;
1606 
1607 	switch (adev->ip_versions[SDMA0_HWIP][0]) {
1608 	case IP_VERSION(5, 2, 0):
1609 	case IP_VERSION(5, 2, 2):
1610 	case IP_VERSION(5, 2, 1):
1611 	case IP_VERSION(5, 2, 4):
1612 	case IP_VERSION(5, 2, 5):
1613 	case IP_VERSION(5, 2, 3):
1614 		sdma_v5_2_update_medium_grain_clock_gating(adev,
1615 				state == AMD_CG_STATE_GATE);
1616 		sdma_v5_2_update_medium_grain_light_sleep(adev,
1617 				state == AMD_CG_STATE_GATE);
1618 		break;
1619 	default:
1620 		break;
1621 	}
1622 
1623 	return 0;
1624 }
1625 
1626 static int sdma_v5_2_set_powergating_state(void *handle,
1627 					  enum amd_powergating_state state)
1628 {
1629 	return 0;
1630 }
1631 
1632 static void sdma_v5_2_get_clockgating_state(void *handle, u32 *flags)
1633 {
1634 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1635 	int data;
1636 
1637 	if (amdgpu_sriov_vf(adev))
1638 		*flags = 0;
1639 
1640 	/* AMD_CG_SUPPORT_SDMA_LS */
1641 	data = RREG32_KIQ(sdma_v5_2_get_reg_offset(adev, 0, mmSDMA0_POWER_CNTL));
1642 	if (data & SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK)
1643 		*flags |= AMD_CG_SUPPORT_SDMA_LS;
1644 }
1645 
1646 const struct amd_ip_funcs sdma_v5_2_ip_funcs = {
1647 	.name = "sdma_v5_2",
1648 	.early_init = sdma_v5_2_early_init,
1649 	.late_init = NULL,
1650 	.sw_init = sdma_v5_2_sw_init,
1651 	.sw_fini = sdma_v5_2_sw_fini,
1652 	.hw_init = sdma_v5_2_hw_init,
1653 	.hw_fini = sdma_v5_2_hw_fini,
1654 	.suspend = sdma_v5_2_suspend,
1655 	.resume = sdma_v5_2_resume,
1656 	.is_idle = sdma_v5_2_is_idle,
1657 	.wait_for_idle = sdma_v5_2_wait_for_idle,
1658 	.soft_reset = sdma_v5_2_soft_reset,
1659 	.set_clockgating_state = sdma_v5_2_set_clockgating_state,
1660 	.set_powergating_state = sdma_v5_2_set_powergating_state,
1661 	.get_clockgating_state = sdma_v5_2_get_clockgating_state,
1662 };
1663 
1664 static const struct amdgpu_ring_funcs sdma_v5_2_ring_funcs = {
1665 	.type = AMDGPU_RING_TYPE_SDMA,
1666 	.align_mask = 0xf,
1667 	.nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP),
1668 	.support_64bit_ptrs = true,
1669 	.vmhub = AMDGPU_GFXHUB_0,
1670 	.get_rptr = sdma_v5_2_ring_get_rptr,
1671 	.get_wptr = sdma_v5_2_ring_get_wptr,
1672 	.set_wptr = sdma_v5_2_ring_set_wptr,
1673 	.emit_frame_size =
1674 		5 + /* sdma_v5_2_ring_init_cond_exec */
1675 		6 + /* sdma_v5_2_ring_emit_hdp_flush */
1676 		3 + /* hdp_invalidate */
1677 		6 + /* sdma_v5_2_ring_emit_pipeline_sync */
1678 		/* sdma_v5_2_ring_emit_vm_flush */
1679 		SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 +
1680 		SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6 +
1681 		10 + 10 + 10, /* sdma_v5_2_ring_emit_fence x3 for user fence, vm fence */
1682 	.emit_ib_size = 7 + 6, /* sdma_v5_2_ring_emit_ib */
1683 	.emit_ib = sdma_v5_2_ring_emit_ib,
1684 	.emit_mem_sync = sdma_v5_2_ring_emit_mem_sync,
1685 	.emit_fence = sdma_v5_2_ring_emit_fence,
1686 	.emit_pipeline_sync = sdma_v5_2_ring_emit_pipeline_sync,
1687 	.emit_vm_flush = sdma_v5_2_ring_emit_vm_flush,
1688 	.emit_hdp_flush = sdma_v5_2_ring_emit_hdp_flush,
1689 	.test_ring = sdma_v5_2_ring_test_ring,
1690 	.test_ib = sdma_v5_2_ring_test_ib,
1691 	.insert_nop = sdma_v5_2_ring_insert_nop,
1692 	.pad_ib = sdma_v5_2_ring_pad_ib,
1693 	.emit_wreg = sdma_v5_2_ring_emit_wreg,
1694 	.emit_reg_wait = sdma_v5_2_ring_emit_reg_wait,
1695 	.emit_reg_write_reg_wait = sdma_v5_2_ring_emit_reg_write_reg_wait,
1696 	.init_cond_exec = sdma_v5_2_ring_init_cond_exec,
1697 	.patch_cond_exec = sdma_v5_2_ring_patch_cond_exec,
1698 	.preempt_ib = sdma_v5_2_ring_preempt_ib,
1699 };
1700 
1701 static void sdma_v5_2_set_ring_funcs(struct amdgpu_device *adev)
1702 {
1703 	int i;
1704 
1705 	for (i = 0; i < adev->sdma.num_instances; i++) {
1706 		adev->sdma.instance[i].ring.funcs = &sdma_v5_2_ring_funcs;
1707 		adev->sdma.instance[i].ring.me = i;
1708 	}
1709 }
1710 
1711 static const struct amdgpu_irq_src_funcs sdma_v5_2_trap_irq_funcs = {
1712 	.set = sdma_v5_2_set_trap_irq_state,
1713 	.process = sdma_v5_2_process_trap_irq,
1714 };
1715 
1716 static const struct amdgpu_irq_src_funcs sdma_v5_2_illegal_inst_irq_funcs = {
1717 	.process = sdma_v5_2_process_illegal_inst_irq,
1718 };
1719 
1720 static void sdma_v5_2_set_irq_funcs(struct amdgpu_device *adev)
1721 {
1722 	adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_INSTANCE0 +
1723 					adev->sdma.num_instances;
1724 	adev->sdma.trap_irq.funcs = &sdma_v5_2_trap_irq_funcs;
1725 	adev->sdma.illegal_inst_irq.funcs = &sdma_v5_2_illegal_inst_irq_funcs;
1726 }
1727 
1728 /**
1729  * sdma_v5_2_emit_copy_buffer - copy buffer using the sDMA engine
1730  *
1731  * @ib: indirect buffer to copy to
1732  * @src_offset: src GPU address
1733  * @dst_offset: dst GPU address
1734  * @byte_count: number of bytes to xfer
1735  * @tmz: if a secure copy should be used
1736  *
1737  * Copy GPU buffers using the DMA engine.
1738  * Used by the amdgpu ttm implementation to move pages if
1739  * registered as the asic copy callback.
1740  */
1741 static void sdma_v5_2_emit_copy_buffer(struct amdgpu_ib *ib,
1742 				       uint64_t src_offset,
1743 				       uint64_t dst_offset,
1744 				       uint32_t byte_count,
1745 				       bool tmz)
1746 {
1747 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1748 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR) |
1749 		SDMA_PKT_COPY_LINEAR_HEADER_TMZ(tmz ? 1 : 0);
1750 	ib->ptr[ib->length_dw++] = byte_count - 1;
1751 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1752 	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1753 	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1754 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1755 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1756 }
1757 
1758 /**
1759  * sdma_v5_2_emit_fill_buffer - fill buffer using the sDMA engine
1760  *
1761  * @ib: indirect buffer to fill
1762  * @src_data: value to write to buffer
1763  * @dst_offset: dst GPU address
1764  * @byte_count: number of bytes to xfer
1765  *
1766  * Fill GPU buffers using the DMA engine.
1767  */
1768 static void sdma_v5_2_emit_fill_buffer(struct amdgpu_ib *ib,
1769 				       uint32_t src_data,
1770 				       uint64_t dst_offset,
1771 				       uint32_t byte_count)
1772 {
1773 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL);
1774 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1775 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1776 	ib->ptr[ib->length_dw++] = src_data;
1777 	ib->ptr[ib->length_dw++] = byte_count - 1;
1778 }
1779 
1780 static const struct amdgpu_buffer_funcs sdma_v5_2_buffer_funcs = {
1781 	.copy_max_bytes = 0x400000,
1782 	.copy_num_dw = 7,
1783 	.emit_copy_buffer = sdma_v5_2_emit_copy_buffer,
1784 
1785 	.fill_max_bytes = 0x400000,
1786 	.fill_num_dw = 5,
1787 	.emit_fill_buffer = sdma_v5_2_emit_fill_buffer,
1788 };
1789 
1790 static void sdma_v5_2_set_buffer_funcs(struct amdgpu_device *adev)
1791 {
1792 	if (adev->mman.buffer_funcs == NULL) {
1793 		adev->mman.buffer_funcs = &sdma_v5_2_buffer_funcs;
1794 		adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1795 	}
1796 }
1797 
1798 static const struct amdgpu_vm_pte_funcs sdma_v5_2_vm_pte_funcs = {
1799 	.copy_pte_num_dw = 7,
1800 	.copy_pte = sdma_v5_2_vm_copy_pte,
1801 	.write_pte = sdma_v5_2_vm_write_pte,
1802 	.set_pte_pde = sdma_v5_2_vm_set_pte_pde,
1803 };
1804 
1805 static void sdma_v5_2_set_vm_pte_funcs(struct amdgpu_device *adev)
1806 {
1807 	unsigned i;
1808 
1809 	if (adev->vm_manager.vm_pte_funcs == NULL) {
1810 		adev->vm_manager.vm_pte_funcs = &sdma_v5_2_vm_pte_funcs;
1811 		for (i = 0; i < adev->sdma.num_instances; i++) {
1812 			adev->vm_manager.vm_pte_scheds[i] =
1813 				&adev->sdma.instance[i].ring.sched;
1814 		}
1815 		adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
1816 	}
1817 }
1818 
1819 const struct amdgpu_ip_block_version sdma_v5_2_ip_block = {
1820 	.type = AMD_IP_BLOCK_TYPE_SDMA,
1821 	.major = 5,
1822 	.minor = 2,
1823 	.rev = 0,
1824 	.funcs = &sdma_v5_2_ip_funcs,
1825 };
1826