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 
51 MODULE_FIRMWARE("amdgpu/vangogh_sdma.bin");
52 
53 #define SDMA1_REG_OFFSET 0x600
54 #define SDMA3_REG_OFFSET 0x400
55 #define SDMA0_HYP_DEC_REG_START 0x5880
56 #define SDMA0_HYP_DEC_REG_END 0x5893
57 #define SDMA1_HYP_DEC_REG_OFFSET 0x20
58 
59 static void sdma_v5_2_set_ring_funcs(struct amdgpu_device *adev);
60 static void sdma_v5_2_set_buffer_funcs(struct amdgpu_device *adev);
61 static void sdma_v5_2_set_vm_pte_funcs(struct amdgpu_device *adev);
62 static void sdma_v5_2_set_irq_funcs(struct amdgpu_device *adev);
63 
64 static u32 sdma_v5_2_get_reg_offset(struct amdgpu_device *adev, u32 instance, u32 internal_offset)
65 {
66 	u32 base;
67 
68 	if (internal_offset >= SDMA0_HYP_DEC_REG_START &&
69 	    internal_offset <= SDMA0_HYP_DEC_REG_END) {
70 		base = adev->reg_offset[GC_HWIP][0][1];
71 		if (instance != 0)
72 			internal_offset += SDMA1_HYP_DEC_REG_OFFSET * instance;
73 	} else {
74 		if (instance < 2) {
75 			base = adev->reg_offset[GC_HWIP][0][0];
76 			if (instance == 1)
77 				internal_offset += SDMA1_REG_OFFSET;
78 		} else {
79 			base = adev->reg_offset[GC_HWIP][0][2];
80 			if (instance == 3)
81 				internal_offset += SDMA3_REG_OFFSET;
82 		}
83 	}
84 
85 	return base + internal_offset;
86 }
87 
88 static void sdma_v5_2_init_golden_registers(struct amdgpu_device *adev)
89 {
90 	switch (adev->asic_type) {
91 	case CHIP_SIENNA_CICHLID:
92 	case CHIP_NAVY_FLOUNDER:
93 	case CHIP_VANGOGH:
94 	case CHIP_DIMGREY_CAVEFISH:
95 		break;
96 	default:
97 		break;
98 	}
99 }
100 
101 static int sdma_v5_2_init_inst_ctx(struct amdgpu_sdma_instance *sdma_inst)
102 {
103 	int err = 0;
104 	const struct sdma_firmware_header_v1_0 *hdr;
105 
106 	err = amdgpu_ucode_validate(sdma_inst->fw);
107 	if (err)
108 		return err;
109 
110 	hdr = (const struct sdma_firmware_header_v1_0 *)sdma_inst->fw->data;
111 	sdma_inst->fw_version = le32_to_cpu(hdr->header.ucode_version);
112 	sdma_inst->feature_version = le32_to_cpu(hdr->ucode_feature_version);
113 
114 	if (sdma_inst->feature_version >= 20)
115 		sdma_inst->burst_nop = true;
116 
117 	return 0;
118 }
119 
120 static void sdma_v5_2_destroy_inst_ctx(struct amdgpu_device *adev)
121 {
122 	int i;
123 
124 	for (i = 0; i < adev->sdma.num_instances; i++) {
125 		release_firmware(adev->sdma.instance[i].fw);
126 		adev->sdma.instance[i].fw = NULL;
127 
128 		if (adev->asic_type == CHIP_SIENNA_CICHLID)
129 			break;
130 	}
131 
132 	memset((void *)adev->sdma.instance, 0,
133 	       sizeof(struct amdgpu_sdma_instance) * AMDGPU_MAX_SDMA_INSTANCES);
134 }
135 
136 /**
137  * sdma_v5_2_init_microcode - load ucode images from disk
138  *
139  * @adev: amdgpu_device pointer
140  *
141  * Use the firmware interface to load the ucode images into
142  * the driver (not loaded into hw).
143  * Returns 0 on success, error on failure.
144  */
145 
146 // emulation only, won't work on real chip
147 // navi10 real chip need to use PSP to load firmware
148 static int sdma_v5_2_init_microcode(struct amdgpu_device *adev)
149 {
150 	const char *chip_name;
151 	char fw_name[40];
152 	int err = 0, i;
153 	struct amdgpu_firmware_info *info = NULL;
154 	const struct common_firmware_header *header = NULL;
155 
156 	DRM_DEBUG("\n");
157 
158 	switch (adev->asic_type) {
159 	case CHIP_SIENNA_CICHLID:
160 		chip_name = "sienna_cichlid";
161 		break;
162 	case CHIP_NAVY_FLOUNDER:
163 		chip_name = "navy_flounder";
164 		break;
165 	case CHIP_VANGOGH:
166 		chip_name = "vangogh";
167 		break;
168 	case CHIP_DIMGREY_CAVEFISH:
169 		chip_name = "dimgrey_cavefish";
170 		break;
171 	default:
172 		BUG();
173 	}
174 
175 	snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma.bin", chip_name);
176 
177 	err = request_firmware(&adev->sdma.instance[0].fw, fw_name, adev->dev);
178 	if (err)
179 		goto out;
180 
181 	err = sdma_v5_2_init_inst_ctx(&adev->sdma.instance[0]);
182 	if (err)
183 		goto out;
184 
185 	for (i = 1; i < adev->sdma.num_instances; i++) {
186 		if (adev->asic_type >= CHIP_SIENNA_CICHLID &&
187 		    adev->asic_type <= CHIP_DIMGREY_CAVEFISH) {
188 			memcpy((void *)&adev->sdma.instance[i],
189 			       (void *)&adev->sdma.instance[0],
190 			       sizeof(struct amdgpu_sdma_instance));
191 		} else {
192 			snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma%d.bin", chip_name, i);
193 			err = request_firmware(&adev->sdma.instance[i].fw, fw_name, adev->dev);
194 			if (err)
195 				goto out;
196 
197 			err = sdma_v5_2_init_inst_ctx(&adev->sdma.instance[i]);
198 			if (err)
199 				goto out;
200 		}
201 	}
202 
203 	DRM_DEBUG("psp_load == '%s'\n",
204 		  adev->firmware.load_type == AMDGPU_FW_LOAD_PSP ? "true" : "false");
205 
206 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
207 		for (i = 0; i < adev->sdma.num_instances; i++) {
208 			info = &adev->firmware.ucode[AMDGPU_UCODE_ID_SDMA0 + i];
209 			info->ucode_id = AMDGPU_UCODE_ID_SDMA0 + i;
210 			info->fw = adev->sdma.instance[i].fw;
211 			header = (const struct common_firmware_header *)info->fw->data;
212 			adev->firmware.fw_size +=
213 				ALIGN(le32_to_cpu(header->ucode_size_bytes), PAGE_SIZE);
214 		}
215 	}
216 
217 out:
218 	if (err) {
219 		DRM_ERROR("sdma_v5_2: Failed to load firmware \"%s\"\n", fw_name);
220 		sdma_v5_2_destroy_inst_ctx(adev);
221 	}
222 	return err;
223 }
224 
225 static unsigned sdma_v5_2_ring_init_cond_exec(struct amdgpu_ring *ring)
226 {
227 	unsigned ret;
228 
229 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_COND_EXE));
230 	amdgpu_ring_write(ring, lower_32_bits(ring->cond_exe_gpu_addr));
231 	amdgpu_ring_write(ring, upper_32_bits(ring->cond_exe_gpu_addr));
232 	amdgpu_ring_write(ring, 1);
233 	ret = ring->wptr & ring->buf_mask;/* this is the offset we need patch later */
234 	amdgpu_ring_write(ring, 0x55aa55aa);/* insert dummy here and patch it later */
235 
236 	return ret;
237 }
238 
239 static void sdma_v5_2_ring_patch_cond_exec(struct amdgpu_ring *ring,
240 					   unsigned offset)
241 {
242 	unsigned cur;
243 
244 	BUG_ON(offset > ring->buf_mask);
245 	BUG_ON(ring->ring[offset] != 0x55aa55aa);
246 
247 	cur = (ring->wptr - 1) & ring->buf_mask;
248 	if (cur > offset)
249 		ring->ring[offset] = cur - offset;
250 	else
251 		ring->ring[offset] = (ring->buf_mask + 1) - offset + cur;
252 }
253 
254 /**
255  * sdma_v5_2_ring_get_rptr - get the current read pointer
256  *
257  * @ring: amdgpu ring pointer
258  *
259  * Get the current rptr from the hardware (NAVI10+).
260  */
261 static uint64_t sdma_v5_2_ring_get_rptr(struct amdgpu_ring *ring)
262 {
263 	u64 *rptr;
264 
265 	/* XXX check if swapping is necessary on BE */
266 	rptr = ((u64 *)&ring->adev->wb.wb[ring->rptr_offs]);
267 
268 	DRM_DEBUG("rptr before shift == 0x%016llx\n", *rptr);
269 	return ((*rptr) >> 2);
270 }
271 
272 /**
273  * sdma_v5_2_ring_get_wptr - get the current write pointer
274  *
275  * @ring: amdgpu ring pointer
276  *
277  * Get the current wptr from the hardware (NAVI10+).
278  */
279 static uint64_t sdma_v5_2_ring_get_wptr(struct amdgpu_ring *ring)
280 {
281 	struct amdgpu_device *adev = ring->adev;
282 	u64 wptr;
283 
284 	if (ring->use_doorbell) {
285 		/* XXX check if swapping is necessary on BE */
286 		wptr = READ_ONCE(*((u64 *)&adev->wb.wb[ring->wptr_offs]));
287 		DRM_DEBUG("wptr/doorbell before shift == 0x%016llx\n", wptr);
288 	} else {
289 		wptr = RREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI));
290 		wptr = wptr << 32;
291 		wptr |= RREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR));
292 		DRM_DEBUG("wptr before shift [%i] wptr == 0x%016llx\n", ring->me, wptr);
293 	}
294 
295 	return wptr >> 2;
296 }
297 
298 /**
299  * sdma_v5_2_ring_set_wptr - commit the write pointer
300  *
301  * @ring: amdgpu ring pointer
302  *
303  * Write the wptr back to the hardware (NAVI10+).
304  */
305 static void sdma_v5_2_ring_set_wptr(struct amdgpu_ring *ring)
306 {
307 	struct amdgpu_device *adev = ring->adev;
308 
309 	DRM_DEBUG("Setting write pointer\n");
310 	if (ring->use_doorbell) {
311 		DRM_DEBUG("Using doorbell -- "
312 				"wptr_offs == 0x%08x "
313 				"lower_32_bits(ring->wptr) << 2 == 0x%08x "
314 				"upper_32_bits(ring->wptr) << 2 == 0x%08x\n",
315 				ring->wptr_offs,
316 				lower_32_bits(ring->wptr << 2),
317 				upper_32_bits(ring->wptr << 2));
318 		/* XXX check if swapping is necessary on BE */
319 		adev->wb.wb[ring->wptr_offs] = lower_32_bits(ring->wptr << 2);
320 		adev->wb.wb[ring->wptr_offs + 1] = upper_32_bits(ring->wptr << 2);
321 		DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016llx)\n",
322 				ring->doorbell_index, ring->wptr << 2);
323 		WDOORBELL64(ring->doorbell_index, ring->wptr << 2);
324 	} else {
325 		DRM_DEBUG("Not using doorbell -- "
326 				"mmSDMA%i_GFX_RB_WPTR == 0x%08x "
327 				"mmSDMA%i_GFX_RB_WPTR_HI == 0x%08x\n",
328 				ring->me,
329 				lower_32_bits(ring->wptr << 2),
330 				ring->me,
331 				upper_32_bits(ring->wptr << 2));
332 		WREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR),
333 			lower_32_bits(ring->wptr << 2));
334 		WREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI),
335 			upper_32_bits(ring->wptr << 2));
336 	}
337 }
338 
339 static void sdma_v5_2_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
340 {
341 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
342 	int i;
343 
344 	for (i = 0; i < count; i++)
345 		if (sdma && sdma->burst_nop && (i == 0))
346 			amdgpu_ring_write(ring, ring->funcs->nop |
347 				SDMA_PKT_NOP_HEADER_COUNT(count - 1));
348 		else
349 			amdgpu_ring_write(ring, ring->funcs->nop);
350 }
351 
352 /**
353  * sdma_v5_2_ring_emit_ib - Schedule an IB on the DMA engine
354  *
355  * @ring: amdgpu ring pointer
356  * @job: job to retrieve vmid from
357  * @ib: IB object to schedule
358  * @flags: unused
359  *
360  * Schedule an IB in the DMA ring.
361  */
362 static void sdma_v5_2_ring_emit_ib(struct amdgpu_ring *ring,
363 				   struct amdgpu_job *job,
364 				   struct amdgpu_ib *ib,
365 				   uint32_t flags)
366 {
367 	unsigned vmid = AMDGPU_JOB_GET_VMID(job);
368 	uint64_t csa_mc_addr = amdgpu_sdma_get_csa_mc_addr(ring, vmid);
369 
370 	/* An IB packet must end on a 8 DW boundary--the next dword
371 	 * must be on a 8-dword boundary. Our IB packet below is 6
372 	 * dwords long, thus add x number of NOPs, such that, in
373 	 * modular arithmetic,
374 	 * wptr + 6 + x = 8k, k >= 0, which in C is,
375 	 * (wptr + 6 + x) % 8 = 0.
376 	 * The expression below, is a solution of x.
377 	 */
378 	sdma_v5_2_ring_insert_nop(ring, (2 - lower_32_bits(ring->wptr)) & 7);
379 
380 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) |
381 			  SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf));
382 	/* base must be 32 byte aligned */
383 	amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
384 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
385 	amdgpu_ring_write(ring, ib->length_dw);
386 	amdgpu_ring_write(ring, lower_32_bits(csa_mc_addr));
387 	amdgpu_ring_write(ring, upper_32_bits(csa_mc_addr));
388 }
389 
390 /**
391  * sdma_v5_2_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
392  *
393  * @ring: amdgpu ring pointer
394  *
395  * Emit an hdp flush packet on the requested DMA ring.
396  */
397 static void sdma_v5_2_ring_emit_hdp_flush(struct amdgpu_ring *ring)
398 {
399 	struct amdgpu_device *adev = ring->adev;
400 	u32 ref_and_mask = 0;
401 	const struct nbio_hdp_flush_reg *nbio_hf_reg = adev->nbio.hdp_flush_reg;
402 
403 	ref_and_mask = nbio_hf_reg->ref_and_mask_sdma0 << ring->me;
404 
405 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
406 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) |
407 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
408 	amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_done_offset(adev)) << 2);
409 	amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_req_offset(adev)) << 2);
410 	amdgpu_ring_write(ring, ref_and_mask); /* reference */
411 	amdgpu_ring_write(ring, ref_and_mask); /* mask */
412 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
413 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
414 }
415 
416 /**
417  * sdma_v5_2_ring_emit_fence - emit a fence on the DMA ring
418  *
419  * @ring: amdgpu ring pointer
420  * @addr: address
421  * @seq: sequence number
422  * @flags: fence related flags
423  *
424  * Add a DMA fence packet to the ring to write
425  * the fence seq number and DMA trap packet to generate
426  * an interrupt if needed.
427  */
428 static void sdma_v5_2_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
429 				      unsigned flags)
430 {
431 	bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
432 	/* write the fence */
433 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) |
434 			  SDMA_PKT_FENCE_HEADER_MTYPE(0x3)); /* Ucached(UC) */
435 	/* zero in first two bits */
436 	BUG_ON(addr & 0x3);
437 	amdgpu_ring_write(ring, lower_32_bits(addr));
438 	amdgpu_ring_write(ring, upper_32_bits(addr));
439 	amdgpu_ring_write(ring, lower_32_bits(seq));
440 
441 	/* optionally write high bits as well */
442 	if (write64bit) {
443 		addr += 4;
444 		amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) |
445 				  SDMA_PKT_FENCE_HEADER_MTYPE(0x3));
446 		/* zero in first two bits */
447 		BUG_ON(addr & 0x3);
448 		amdgpu_ring_write(ring, lower_32_bits(addr));
449 		amdgpu_ring_write(ring, upper_32_bits(addr));
450 		amdgpu_ring_write(ring, upper_32_bits(seq));
451 	}
452 
453 	if (flags & AMDGPU_FENCE_FLAG_INT) {
454 		/* generate an interrupt */
455 		amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP));
456 		amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(0));
457 	}
458 }
459 
460 
461 /**
462  * sdma_v5_2_gfx_stop - stop the gfx async dma engines
463  *
464  * @adev: amdgpu_device pointer
465  *
466  * Stop the gfx async dma ring buffers.
467  */
468 static void sdma_v5_2_gfx_stop(struct amdgpu_device *adev)
469 {
470 	struct amdgpu_ring *sdma0 = &adev->sdma.instance[0].ring;
471 	struct amdgpu_ring *sdma1 = &adev->sdma.instance[1].ring;
472 	struct amdgpu_ring *sdma2 = &adev->sdma.instance[2].ring;
473 	struct amdgpu_ring *sdma3 = &adev->sdma.instance[3].ring;
474 	u32 rb_cntl, ib_cntl;
475 	int i;
476 
477 	if ((adev->mman.buffer_funcs_ring == sdma0) ||
478 	    (adev->mman.buffer_funcs_ring == sdma1) ||
479 	    (adev->mman.buffer_funcs_ring == sdma2) ||
480 	    (adev->mman.buffer_funcs_ring == sdma3))
481 		amdgpu_ttm_set_buffer_funcs_status(adev, false);
482 
483 	for (i = 0; i < adev->sdma.num_instances; i++) {
484 		rb_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL));
485 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 0);
486 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
487 		ib_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL));
488 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 0);
489 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl);
490 	}
491 
492 	sdma0->sched.ready = false;
493 	sdma1->sched.ready = false;
494 	sdma2->sched.ready = false;
495 	sdma3->sched.ready = false;
496 }
497 
498 /**
499  * sdma_v5_2_rlc_stop - stop the compute async dma engines
500  *
501  * @adev: amdgpu_device pointer
502  *
503  * Stop the compute async dma queues.
504  */
505 static void sdma_v5_2_rlc_stop(struct amdgpu_device *adev)
506 {
507 	/* XXX todo */
508 }
509 
510 /**
511  * sdma_v_0_ctx_switch_enable - stop the async dma engines context switch
512  *
513  * @adev: amdgpu_device pointer
514  * @enable: enable/disable the DMA MEs context switch.
515  *
516  * Halt or unhalt the async dma engines context switch.
517  */
518 static void sdma_v5_2_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
519 {
520 	u32 f32_cntl, phase_quantum = 0;
521 	int i;
522 
523 	if (amdgpu_sdma_phase_quantum) {
524 		unsigned value = amdgpu_sdma_phase_quantum;
525 		unsigned unit = 0;
526 
527 		while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
528 				SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) {
529 			value = (value + 1) >> 1;
530 			unit++;
531 		}
532 		if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
533 			    SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) {
534 			value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
535 				 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT);
536 			unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
537 				SDMA0_PHASE0_QUANTUM__UNIT__SHIFT);
538 			WARN_ONCE(1,
539 			"clamping sdma_phase_quantum to %uK clock cycles\n",
540 				  value << unit);
541 		}
542 		phase_quantum =
543 			value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
544 			unit  << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
545 	}
546 
547 	for (i = 0; i < adev->sdma.num_instances; i++) {
548 		f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
549 		f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
550 				AUTO_CTXSW_ENABLE, enable ? 1 : 0);
551 		if (enable && amdgpu_sdma_phase_quantum) {
552 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE0_QUANTUM),
553 			       phase_quantum);
554 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE1_QUANTUM),
555 			       phase_quantum);
556 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE2_QUANTUM),
557 			       phase_quantum);
558 		}
559 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl);
560 	}
561 
562 }
563 
564 /**
565  * sdma_v5_2_enable - stop the async dma engines
566  *
567  * @adev: amdgpu_device pointer
568  * @enable: enable/disable the DMA MEs.
569  *
570  * Halt or unhalt the async dma engines.
571  */
572 static void sdma_v5_2_enable(struct amdgpu_device *adev, bool enable)
573 {
574 	u32 f32_cntl;
575 	int i;
576 
577 	if (!enable) {
578 		sdma_v5_2_gfx_stop(adev);
579 		sdma_v5_2_rlc_stop(adev);
580 	}
581 
582 	for (i = 0; i < adev->sdma.num_instances; i++) {
583 		f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
584 		f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, enable ? 0 : 1);
585 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl);
586 	}
587 }
588 
589 /**
590  * sdma_v5_2_gfx_resume - setup and start the async dma engines
591  *
592  * @adev: amdgpu_device pointer
593  *
594  * Set up the gfx DMA ring buffers and enable them.
595  * Returns 0 for success, error for failure.
596  */
597 static int sdma_v5_2_gfx_resume(struct amdgpu_device *adev)
598 {
599 	struct amdgpu_ring *ring;
600 	u32 rb_cntl, ib_cntl;
601 	u32 rb_bufsz;
602 	u32 wb_offset;
603 	u32 doorbell;
604 	u32 doorbell_offset;
605 	u32 temp;
606 	u32 wptr_poll_cntl;
607 	u64 wptr_gpu_addr;
608 	int i, r;
609 
610 	for (i = 0; i < adev->sdma.num_instances; i++) {
611 		ring = &adev->sdma.instance[i].ring;
612 		wb_offset = (ring->rptr_offs * 4);
613 
614 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL), 0);
615 
616 		/* Set ring buffer size in dwords */
617 		rb_bufsz = order_base_2(ring->ring_size / 4);
618 		rb_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL));
619 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz);
620 #ifdef __BIG_ENDIAN
621 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1);
622 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL,
623 					RPTR_WRITEBACK_SWAP_ENABLE, 1);
624 #endif
625 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
626 
627 		/* Initialize the ring buffer's read and write pointers */
628 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR), 0);
629 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_HI), 0);
630 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), 0);
631 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), 0);
632 
633 		/* setup the wptr shadow polling */
634 		wptr_gpu_addr = adev->wb.gpu_addr + (ring->wptr_offs * 4);
635 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_LO),
636 		       lower_32_bits(wptr_gpu_addr));
637 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_HI),
638 		       upper_32_bits(wptr_gpu_addr));
639 		wptr_poll_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i,
640 							 mmSDMA0_GFX_RB_WPTR_POLL_CNTL));
641 		wptr_poll_cntl = REG_SET_FIELD(wptr_poll_cntl,
642 					       SDMA0_GFX_RB_WPTR_POLL_CNTL,
643 					       F32_POLL_ENABLE, 1);
644 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_CNTL),
645 		       wptr_poll_cntl);
646 
647 		/* set the wb address whether it's enabled or not */
648 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_HI),
649 		       upper_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFF);
650 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_LO),
651 		       lower_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC);
652 
653 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
654 
655 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE), ring->gpu_addr >> 8);
656 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE_HI), ring->gpu_addr >> 40);
657 
658 		ring->wptr = 0;
659 
660 		/* before programing wptr to a less value, need set minor_ptr_update first */
661 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 1);
662 
663 		if (!amdgpu_sriov_vf(adev)) { /* only bare-metal use register write for wptr */
664 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), lower_32_bits(ring->wptr) << 2);
665 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), upper_32_bits(ring->wptr) << 2);
666 		}
667 
668 		doorbell = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL));
669 		doorbell_offset = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL_OFFSET));
670 
671 		if (ring->use_doorbell) {
672 			doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 1);
673 			doorbell_offset = REG_SET_FIELD(doorbell_offset, SDMA0_GFX_DOORBELL_OFFSET,
674 					OFFSET, ring->doorbell_index);
675 		} else {
676 			doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 0);
677 		}
678 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL), doorbell);
679 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL_OFFSET), doorbell_offset);
680 
681 		adev->nbio.funcs->sdma_doorbell_range(adev, i, ring->use_doorbell,
682 						      ring->doorbell_index,
683 						      adev->doorbell_index.sdma_doorbell_range);
684 
685 		if (amdgpu_sriov_vf(adev))
686 			sdma_v5_2_ring_set_wptr(ring);
687 
688 		/* set minor_ptr_update to 0 after wptr programed */
689 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 0);
690 
691 		/* set utc l1 enable flag always to 1 */
692 		temp = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
693 		temp = REG_SET_FIELD(temp, SDMA0_CNTL, UTC_L1_ENABLE, 1);
694 
695 		/* enable MCBP */
696 		temp = REG_SET_FIELD(temp, SDMA0_CNTL, MIDCMD_PREEMPT_ENABLE, 1);
697 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), temp);
698 
699 		/* Set up RESP_MODE to non-copy addresses */
700 		temp = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL));
701 		temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, RESP_MODE, 3);
702 		temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, REDO_DELAY, 9);
703 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL), temp);
704 
705 		/* program default cache read and write policy */
706 		temp = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE));
707 		/* clean read policy and write policy bits */
708 		temp &= 0xFF0FFF;
709 		temp |= ((CACHE_READ_POLICY_L2__DEFAULT << 12) |
710 			 (CACHE_WRITE_POLICY_L2__DEFAULT << 14) |
711 			 SDMA0_UTCL1_PAGE__LLC_NOALLOC_MASK);
712 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE), temp);
713 
714 		if (!amdgpu_sriov_vf(adev)) {
715 			/* unhalt engine */
716 			temp = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
717 			temp = REG_SET_FIELD(temp, SDMA0_F32_CNTL, HALT, 0);
718 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), temp);
719 		}
720 
721 		/* enable DMA RB */
722 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1);
723 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
724 
725 		ib_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL));
726 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 1);
727 #ifdef __BIG_ENDIAN
728 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1);
729 #endif
730 		/* enable DMA IBs */
731 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl);
732 
733 		ring->sched.ready = true;
734 
735 		if (amdgpu_sriov_vf(adev)) { /* bare-metal sequence doesn't need below to lines */
736 			sdma_v5_2_ctx_switch_enable(adev, true);
737 			sdma_v5_2_enable(adev, true);
738 		}
739 
740 		r = amdgpu_ring_test_ring(ring);
741 		if (r) {
742 			ring->sched.ready = false;
743 			return r;
744 		}
745 
746 		if (adev->mman.buffer_funcs_ring == ring)
747 			amdgpu_ttm_set_buffer_funcs_status(adev, true);
748 	}
749 
750 	return 0;
751 }
752 
753 /**
754  * sdma_v5_2_rlc_resume - setup and start the async dma engines
755  *
756  * @adev: amdgpu_device pointer
757  *
758  * Set up the compute DMA queues and enable them.
759  * Returns 0 for success, error for failure.
760  */
761 static int sdma_v5_2_rlc_resume(struct amdgpu_device *adev)
762 {
763 	return 0;
764 }
765 
766 /**
767  * sdma_v5_2_load_microcode - load the sDMA ME ucode
768  *
769  * @adev: amdgpu_device pointer
770  *
771  * Loads the sDMA0/1/2/3 ucode.
772  * Returns 0 for success, -EINVAL if the ucode is not available.
773  */
774 static int sdma_v5_2_load_microcode(struct amdgpu_device *adev)
775 {
776 	const struct sdma_firmware_header_v1_0 *hdr;
777 	const __le32 *fw_data;
778 	u32 fw_size;
779 	int i, j;
780 
781 	/* halt the MEs */
782 	sdma_v5_2_enable(adev, false);
783 
784 	for (i = 0; i < adev->sdma.num_instances; i++) {
785 		if (!adev->sdma.instance[i].fw)
786 			return -EINVAL;
787 
788 		hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
789 		amdgpu_ucode_print_sdma_hdr(&hdr->header);
790 		fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
791 
792 		fw_data = (const __le32 *)
793 			(adev->sdma.instance[i].fw->data +
794 				le32_to_cpu(hdr->header.ucode_array_offset_bytes));
795 
796 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), 0);
797 
798 		for (j = 0; j < fw_size; j++) {
799 			if (amdgpu_emu_mode == 1 && j % 500 == 0)
800 				msleep(1);
801 			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UCODE_DATA), le32_to_cpup(fw_data++));
802 		}
803 
804 		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), adev->sdma.instance[i].fw_version);
805 	}
806 
807 	return 0;
808 }
809 
810 /**
811  * sdma_v5_2_start - setup and start the async dma engines
812  *
813  * @adev: amdgpu_device pointer
814  *
815  * Set up the DMA engines and enable them.
816  * Returns 0 for success, error for failure.
817  */
818 static int sdma_v5_2_start(struct amdgpu_device *adev)
819 {
820 	int r = 0;
821 
822 	if (amdgpu_sriov_vf(adev)) {
823 		sdma_v5_2_ctx_switch_enable(adev, false);
824 		sdma_v5_2_enable(adev, false);
825 
826 		/* set RB registers */
827 		r = sdma_v5_2_gfx_resume(adev);
828 		return r;
829 	}
830 
831 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
832 		r = sdma_v5_2_load_microcode(adev);
833 		if (r)
834 			return r;
835 
836 		/* The value of mmSDMA_F32_CNTL is invalid the moment after loading fw */
837 		if (amdgpu_emu_mode == 1)
838 			msleep(1000);
839 	}
840 
841 	/* unhalt the MEs */
842 	sdma_v5_2_enable(adev, true);
843 	/* enable sdma ring preemption */
844 	sdma_v5_2_ctx_switch_enable(adev, true);
845 
846 	/* start the gfx rings and rlc compute queues */
847 	r = sdma_v5_2_gfx_resume(adev);
848 	if (r)
849 		return r;
850 	r = sdma_v5_2_rlc_resume(adev);
851 
852 	return r;
853 }
854 
855 /**
856  * sdma_v5_2_ring_test_ring - simple async dma engine test
857  *
858  * @ring: amdgpu_ring structure holding ring information
859  *
860  * Test the DMA engine by writing using it to write an
861  * value to memory.
862  * Returns 0 for success, error for failure.
863  */
864 static int sdma_v5_2_ring_test_ring(struct amdgpu_ring *ring)
865 {
866 	struct amdgpu_device *adev = ring->adev;
867 	unsigned i;
868 	unsigned index;
869 	int r;
870 	u32 tmp;
871 	u64 gpu_addr;
872 
873 	r = amdgpu_device_wb_get(adev, &index);
874 	if (r) {
875 		dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
876 		return r;
877 	}
878 
879 	gpu_addr = adev->wb.gpu_addr + (index * 4);
880 	tmp = 0xCAFEDEAD;
881 	adev->wb.wb[index] = cpu_to_le32(tmp);
882 
883 	r = amdgpu_ring_alloc(ring, 5);
884 	if (r) {
885 		DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
886 		amdgpu_device_wb_free(adev, index);
887 		return r;
888 	}
889 
890 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
891 			  SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
892 	amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
893 	amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
894 	amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0));
895 	amdgpu_ring_write(ring, 0xDEADBEEF);
896 	amdgpu_ring_commit(ring);
897 
898 	for (i = 0; i < adev->usec_timeout; i++) {
899 		tmp = le32_to_cpu(adev->wb.wb[index]);
900 		if (tmp == 0xDEADBEEF)
901 			break;
902 		if (amdgpu_emu_mode == 1)
903 			msleep(1);
904 		else
905 			udelay(1);
906 	}
907 
908 	if (i >= adev->usec_timeout)
909 		r = -ETIMEDOUT;
910 
911 	amdgpu_device_wb_free(adev, index);
912 
913 	return r;
914 }
915 
916 /**
917  * sdma_v5_2_ring_test_ib - test an IB on the DMA engine
918  *
919  * @ring: amdgpu_ring structure holding ring information
920  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
921  *
922  * Test a simple IB in the DMA ring.
923  * Returns 0 on success, error on failure.
924  */
925 static int sdma_v5_2_ring_test_ib(struct amdgpu_ring *ring, long timeout)
926 {
927 	struct amdgpu_device *adev = ring->adev;
928 	struct amdgpu_ib ib;
929 	struct dma_fence *f = NULL;
930 	unsigned index;
931 	long r;
932 	u32 tmp = 0;
933 	u64 gpu_addr;
934 
935 	r = amdgpu_device_wb_get(adev, &index);
936 	if (r) {
937 		dev_err(adev->dev, "(%ld) failed to allocate wb slot\n", r);
938 		return r;
939 	}
940 
941 	gpu_addr = adev->wb.gpu_addr + (index * 4);
942 	tmp = 0xCAFEDEAD;
943 	adev->wb.wb[index] = cpu_to_le32(tmp);
944 	memset(&ib, 0, sizeof(ib));
945 	r = amdgpu_ib_get(adev, NULL, 256, AMDGPU_IB_POOL_DIRECT, &ib);
946 	if (r) {
947 		DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r);
948 		goto err0;
949 	}
950 
951 	ib.ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
952 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
953 	ib.ptr[1] = lower_32_bits(gpu_addr);
954 	ib.ptr[2] = upper_32_bits(gpu_addr);
955 	ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0);
956 	ib.ptr[4] = 0xDEADBEEF;
957 	ib.ptr[5] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
958 	ib.ptr[6] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
959 	ib.ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
960 	ib.length_dw = 8;
961 
962 	r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
963 	if (r)
964 		goto err1;
965 
966 	r = dma_fence_wait_timeout(f, false, timeout);
967 	if (r == 0) {
968 		DRM_ERROR("amdgpu: IB test timed out\n");
969 		r = -ETIMEDOUT;
970 		goto err1;
971 	} else if (r < 0) {
972 		DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
973 		goto err1;
974 	}
975 	tmp = le32_to_cpu(adev->wb.wb[index]);
976 	if (tmp == 0xDEADBEEF)
977 		r = 0;
978 	else
979 		r = -EINVAL;
980 
981 err1:
982 	amdgpu_ib_free(adev, &ib, NULL);
983 	dma_fence_put(f);
984 err0:
985 	amdgpu_device_wb_free(adev, index);
986 	return r;
987 }
988 
989 
990 /**
991  * sdma_v5_2_vm_copy_pte - update PTEs by copying them from the GART
992  *
993  * @ib: indirect buffer to fill with commands
994  * @pe: addr of the page entry
995  * @src: src addr to copy from
996  * @count: number of page entries to update
997  *
998  * Update PTEs by copying them from the GART using sDMA.
999  */
1000 static void sdma_v5_2_vm_copy_pte(struct amdgpu_ib *ib,
1001 				  uint64_t pe, uint64_t src,
1002 				  unsigned count)
1003 {
1004 	unsigned bytes = count * 8;
1005 
1006 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1007 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
1008 	ib->ptr[ib->length_dw++] = bytes - 1;
1009 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1010 	ib->ptr[ib->length_dw++] = lower_32_bits(src);
1011 	ib->ptr[ib->length_dw++] = upper_32_bits(src);
1012 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1013 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1014 
1015 }
1016 
1017 /**
1018  * sdma_v5_2_vm_write_pte - update PTEs by writing them manually
1019  *
1020  * @ib: indirect buffer to fill with commands
1021  * @pe: addr of the page entry
1022  * @value: dst addr to write into pe
1023  * @count: number of page entries to update
1024  * @incr: increase next addr by incr bytes
1025  *
1026  * Update PTEs by writing them manually using sDMA.
1027  */
1028 static void sdma_v5_2_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
1029 				   uint64_t value, unsigned count,
1030 				   uint32_t incr)
1031 {
1032 	unsigned ndw = count * 2;
1033 
1034 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
1035 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1036 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1037 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1038 	ib->ptr[ib->length_dw++] = ndw - 1;
1039 	for (; ndw > 0; ndw -= 2) {
1040 		ib->ptr[ib->length_dw++] = lower_32_bits(value);
1041 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
1042 		value += incr;
1043 	}
1044 }
1045 
1046 /**
1047  * sdma_v5_2_vm_set_pte_pde - update the page tables using sDMA
1048  *
1049  * @ib: indirect buffer to fill with commands
1050  * @pe: addr of the page entry
1051  * @addr: dst addr to write into pe
1052  * @count: number of page entries to update
1053  * @incr: increase next addr by incr bytes
1054  * @flags: access flags
1055  *
1056  * Update the page tables using sDMA.
1057  */
1058 static void sdma_v5_2_vm_set_pte_pde(struct amdgpu_ib *ib,
1059 				     uint64_t pe,
1060 				     uint64_t addr, unsigned count,
1061 				     uint32_t incr, uint64_t flags)
1062 {
1063 	/* for physically contiguous pages (vram) */
1064 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_PTEPDE);
1065 	ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
1066 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1067 	ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
1068 	ib->ptr[ib->length_dw++] = upper_32_bits(flags);
1069 	ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
1070 	ib->ptr[ib->length_dw++] = upper_32_bits(addr);
1071 	ib->ptr[ib->length_dw++] = incr; /* increment size */
1072 	ib->ptr[ib->length_dw++] = 0;
1073 	ib->ptr[ib->length_dw++] = count - 1; /* number of entries */
1074 }
1075 
1076 /**
1077  * sdma_v5_2_ring_pad_ib - pad the IB
1078  *
1079  * @ib: indirect buffer to fill with padding
1080  * @ring: amdgpu_ring structure holding ring information
1081  *
1082  * Pad the IB with NOPs to a boundary multiple of 8.
1083  */
1084 static void sdma_v5_2_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
1085 {
1086 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
1087 	u32 pad_count;
1088 	int i;
1089 
1090 	pad_count = (-ib->length_dw) & 0x7;
1091 	for (i = 0; i < pad_count; i++)
1092 		if (sdma && sdma->burst_nop && (i == 0))
1093 			ib->ptr[ib->length_dw++] =
1094 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP) |
1095 				SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
1096 		else
1097 			ib->ptr[ib->length_dw++] =
1098 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
1099 }
1100 
1101 
1102 /**
1103  * sdma_v5_2_ring_emit_pipeline_sync - sync the pipeline
1104  *
1105  * @ring: amdgpu_ring pointer
1106  *
1107  * Make sure all previous operations are completed (CIK).
1108  */
1109 static void sdma_v5_2_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
1110 {
1111 	uint32_t seq = ring->fence_drv.sync_seq;
1112 	uint64_t addr = ring->fence_drv.gpu_addr;
1113 
1114 	/* wait for idle */
1115 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1116 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1117 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
1118 			  SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1));
1119 	amdgpu_ring_write(ring, addr & 0xfffffffc);
1120 	amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
1121 	amdgpu_ring_write(ring, seq); /* reference */
1122 	amdgpu_ring_write(ring, 0xffffffff); /* mask */
1123 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1124 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */
1125 }
1126 
1127 
1128 /**
1129  * sdma_v5_2_ring_emit_vm_flush - vm flush using sDMA
1130  *
1131  * @ring: amdgpu_ring pointer
1132  * @vmid: vmid number to use
1133  * @pd_addr: address
1134  *
1135  * Update the page table base and flush the VM TLB
1136  * using sDMA.
1137  */
1138 static void sdma_v5_2_ring_emit_vm_flush(struct amdgpu_ring *ring,
1139 					 unsigned vmid, uint64_t pd_addr)
1140 {
1141 	amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
1142 }
1143 
1144 static void sdma_v5_2_ring_emit_wreg(struct amdgpu_ring *ring,
1145 				     uint32_t reg, uint32_t val)
1146 {
1147 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
1148 			  SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
1149 	amdgpu_ring_write(ring, reg);
1150 	amdgpu_ring_write(ring, val);
1151 }
1152 
1153 static void sdma_v5_2_ring_emit_reg_wait(struct amdgpu_ring *ring, uint32_t reg,
1154 					 uint32_t val, uint32_t mask)
1155 {
1156 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1157 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1158 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* equal */
1159 	amdgpu_ring_write(ring, reg << 2);
1160 	amdgpu_ring_write(ring, 0);
1161 	amdgpu_ring_write(ring, val); /* reference */
1162 	amdgpu_ring_write(ring, mask); /* mask */
1163 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1164 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10));
1165 }
1166 
1167 static void sdma_v5_2_ring_emit_reg_write_reg_wait(struct amdgpu_ring *ring,
1168 						   uint32_t reg0, uint32_t reg1,
1169 						   uint32_t ref, uint32_t mask)
1170 {
1171 	amdgpu_ring_emit_wreg(ring, reg0, ref);
1172 	/* wait for a cycle to reset vm_inv_eng*_ack */
1173 	amdgpu_ring_emit_reg_wait(ring, reg0, 0, 0);
1174 	amdgpu_ring_emit_reg_wait(ring, reg1, mask, mask);
1175 }
1176 
1177 static int sdma_v5_2_early_init(void *handle)
1178 {
1179 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1180 
1181 	switch (adev->asic_type) {
1182 	case CHIP_SIENNA_CICHLID:
1183 		adev->sdma.num_instances = 4;
1184 		break;
1185 	case CHIP_NAVY_FLOUNDER:
1186 	case CHIP_DIMGREY_CAVEFISH:
1187 		adev->sdma.num_instances = 2;
1188 		break;
1189 	case CHIP_VANGOGH:
1190 		adev->sdma.num_instances = 1;
1191 		break;
1192 	default:
1193 		break;
1194 	}
1195 
1196 	sdma_v5_2_set_ring_funcs(adev);
1197 	sdma_v5_2_set_buffer_funcs(adev);
1198 	sdma_v5_2_set_vm_pte_funcs(adev);
1199 	sdma_v5_2_set_irq_funcs(adev);
1200 
1201 	return 0;
1202 }
1203 
1204 static unsigned sdma_v5_2_seq_to_irq_id(int seq_num)
1205 {
1206 	switch (seq_num) {
1207 	case 0:
1208 		return SOC15_IH_CLIENTID_SDMA0;
1209 	case 1:
1210 		return SOC15_IH_CLIENTID_SDMA1;
1211 	case 2:
1212 		return SOC15_IH_CLIENTID_SDMA2;
1213 	case 3:
1214 		return SOC15_IH_CLIENTID_SDMA3_Sienna_Cichlid;
1215 	default:
1216 		break;
1217 	}
1218 	return -EINVAL;
1219 }
1220 
1221 static unsigned sdma_v5_2_seq_to_trap_id(int seq_num)
1222 {
1223 	switch (seq_num) {
1224 	case 0:
1225 		return SDMA0_5_0__SRCID__SDMA_TRAP;
1226 	case 1:
1227 		return SDMA1_5_0__SRCID__SDMA_TRAP;
1228 	case 2:
1229 		return SDMA2_5_0__SRCID__SDMA_TRAP;
1230 	case 3:
1231 		return SDMA3_5_0__SRCID__SDMA_TRAP;
1232 	default:
1233 		break;
1234 	}
1235 	return -EINVAL;
1236 }
1237 
1238 static int sdma_v5_2_sw_init(void *handle)
1239 {
1240 	struct amdgpu_ring *ring;
1241 	int r, i;
1242 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1243 
1244 	/* SDMA trap event */
1245 	for (i = 0; i < adev->sdma.num_instances; i++) {
1246 		r = amdgpu_irq_add_id(adev, sdma_v5_2_seq_to_irq_id(i),
1247 				      sdma_v5_2_seq_to_trap_id(i),
1248 				      &adev->sdma.trap_irq);
1249 		if (r)
1250 			return r;
1251 	}
1252 
1253 	r = sdma_v5_2_init_microcode(adev);
1254 	if (r) {
1255 		DRM_ERROR("Failed to load sdma firmware!\n");
1256 		return r;
1257 	}
1258 
1259 	for (i = 0; i < adev->sdma.num_instances; i++) {
1260 		ring = &adev->sdma.instance[i].ring;
1261 		ring->ring_obj = NULL;
1262 		ring->use_doorbell = true;
1263 		ring->me = i;
1264 
1265 		DRM_INFO("use_doorbell being set to: [%s]\n",
1266 				ring->use_doorbell?"true":"false");
1267 
1268 		ring->doorbell_index =
1269 			(adev->doorbell_index.sdma_engine[i] << 1); //get DWORD offset
1270 
1271 		sprintf(ring->name, "sdma%d", i);
1272 		r = amdgpu_ring_init(adev, ring, 1024,
1273 				     &adev->sdma.trap_irq,
1274 				     AMDGPU_SDMA_IRQ_INSTANCE0 + i,
1275 				     AMDGPU_RING_PRIO_DEFAULT);
1276 		if (r)
1277 			return r;
1278 	}
1279 
1280 	return r;
1281 }
1282 
1283 static int sdma_v5_2_sw_fini(void *handle)
1284 {
1285 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1286 	int i;
1287 
1288 	for (i = 0; i < adev->sdma.num_instances; i++)
1289 		amdgpu_ring_fini(&adev->sdma.instance[i].ring);
1290 
1291 	sdma_v5_2_destroy_inst_ctx(adev);
1292 
1293 	return 0;
1294 }
1295 
1296 static int sdma_v5_2_hw_init(void *handle)
1297 {
1298 	int r;
1299 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1300 
1301 	sdma_v5_2_init_golden_registers(adev);
1302 
1303 	r = sdma_v5_2_start(adev);
1304 
1305 	return r;
1306 }
1307 
1308 static int sdma_v5_2_hw_fini(void *handle)
1309 {
1310 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1311 
1312 	if (amdgpu_sriov_vf(adev))
1313 		return 0;
1314 
1315 	sdma_v5_2_ctx_switch_enable(adev, false);
1316 	sdma_v5_2_enable(adev, false);
1317 
1318 	return 0;
1319 }
1320 
1321 static int sdma_v5_2_suspend(void *handle)
1322 {
1323 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1324 
1325 	return sdma_v5_2_hw_fini(adev);
1326 }
1327 
1328 static int sdma_v5_2_resume(void *handle)
1329 {
1330 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1331 
1332 	return sdma_v5_2_hw_init(adev);
1333 }
1334 
1335 static bool sdma_v5_2_is_idle(void *handle)
1336 {
1337 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1338 	u32 i;
1339 
1340 	for (i = 0; i < adev->sdma.num_instances; i++) {
1341 		u32 tmp = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_STATUS_REG));
1342 
1343 		if (!(tmp & SDMA0_STATUS_REG__IDLE_MASK))
1344 			return false;
1345 	}
1346 
1347 	return true;
1348 }
1349 
1350 static int sdma_v5_2_wait_for_idle(void *handle)
1351 {
1352 	unsigned i;
1353 	u32 sdma0, sdma1, sdma2, sdma3;
1354 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1355 
1356 	for (i = 0; i < adev->usec_timeout; i++) {
1357 		sdma0 = RREG32(sdma_v5_2_get_reg_offset(adev, 0, mmSDMA0_STATUS_REG));
1358 		sdma1 = RREG32(sdma_v5_2_get_reg_offset(adev, 1, mmSDMA0_STATUS_REG));
1359 		sdma2 = RREG32(sdma_v5_2_get_reg_offset(adev, 2, mmSDMA0_STATUS_REG));
1360 		sdma3 = RREG32(sdma_v5_2_get_reg_offset(adev, 3, mmSDMA0_STATUS_REG));
1361 
1362 		if (sdma0 & sdma1 & sdma2 & sdma3 & SDMA0_STATUS_REG__IDLE_MASK)
1363 			return 0;
1364 		udelay(1);
1365 	}
1366 	return -ETIMEDOUT;
1367 }
1368 
1369 static int sdma_v5_2_soft_reset(void *handle)
1370 {
1371 	/* todo */
1372 
1373 	return 0;
1374 }
1375 
1376 static int sdma_v5_2_ring_preempt_ib(struct amdgpu_ring *ring)
1377 {
1378 	int i, r = 0;
1379 	struct amdgpu_device *adev = ring->adev;
1380 	u32 index = 0;
1381 	u64 sdma_gfx_preempt;
1382 
1383 	amdgpu_sdma_get_index_from_ring(ring, &index);
1384 	sdma_gfx_preempt =
1385 		sdma_v5_2_get_reg_offset(adev, index, mmSDMA0_GFX_PREEMPT);
1386 
1387 	/* assert preemption condition */
1388 	amdgpu_ring_set_preempt_cond_exec(ring, false);
1389 
1390 	/* emit the trailing fence */
1391 	ring->trail_seq += 1;
1392 	amdgpu_ring_alloc(ring, 10);
1393 	sdma_v5_2_ring_emit_fence(ring, ring->trail_fence_gpu_addr,
1394 				  ring->trail_seq, 0);
1395 	amdgpu_ring_commit(ring);
1396 
1397 	/* assert IB preemption */
1398 	WREG32(sdma_gfx_preempt, 1);
1399 
1400 	/* poll the trailing fence */
1401 	for (i = 0; i < adev->usec_timeout; i++) {
1402 		if (ring->trail_seq ==
1403 		    le32_to_cpu(*(ring->trail_fence_cpu_addr)))
1404 			break;
1405 		udelay(1);
1406 	}
1407 
1408 	if (i >= adev->usec_timeout) {
1409 		r = -EINVAL;
1410 		DRM_ERROR("ring %d failed to be preempted\n", ring->idx);
1411 	}
1412 
1413 	/* deassert IB preemption */
1414 	WREG32(sdma_gfx_preempt, 0);
1415 
1416 	/* deassert the preemption condition */
1417 	amdgpu_ring_set_preempt_cond_exec(ring, true);
1418 	return r;
1419 }
1420 
1421 static int sdma_v5_2_set_trap_irq_state(struct amdgpu_device *adev,
1422 					struct amdgpu_irq_src *source,
1423 					unsigned type,
1424 					enum amdgpu_interrupt_state state)
1425 {
1426 	u32 sdma_cntl;
1427 
1428 	u32 reg_offset = sdma_v5_2_get_reg_offset(adev, type, mmSDMA0_CNTL);
1429 
1430 	sdma_cntl = RREG32(reg_offset);
1431 	sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE,
1432 		       state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0);
1433 	WREG32(reg_offset, sdma_cntl);
1434 
1435 	return 0;
1436 }
1437 
1438 static int sdma_v5_2_process_trap_irq(struct amdgpu_device *adev,
1439 				      struct amdgpu_irq_src *source,
1440 				      struct amdgpu_iv_entry *entry)
1441 {
1442 	DRM_DEBUG("IH: SDMA trap\n");
1443 	switch (entry->client_id) {
1444 	case SOC15_IH_CLIENTID_SDMA0:
1445 		switch (entry->ring_id) {
1446 		case 0:
1447 			amdgpu_fence_process(&adev->sdma.instance[0].ring);
1448 			break;
1449 		case 1:
1450 			/* XXX compute */
1451 			break;
1452 		case 2:
1453 			/* XXX compute */
1454 			break;
1455 		case 3:
1456 			/* XXX page queue*/
1457 			break;
1458 		}
1459 		break;
1460 	case SOC15_IH_CLIENTID_SDMA1:
1461 		switch (entry->ring_id) {
1462 		case 0:
1463 			amdgpu_fence_process(&adev->sdma.instance[1].ring);
1464 			break;
1465 		case 1:
1466 			/* XXX compute */
1467 			break;
1468 		case 2:
1469 			/* XXX compute */
1470 			break;
1471 		case 3:
1472 			/* XXX page queue*/
1473 			break;
1474 		}
1475 		break;
1476 	case SOC15_IH_CLIENTID_SDMA2:
1477 		switch (entry->ring_id) {
1478 		case 0:
1479 			amdgpu_fence_process(&adev->sdma.instance[2].ring);
1480 			break;
1481 		case 1:
1482 			/* XXX compute */
1483 			break;
1484 		case 2:
1485 			/* XXX compute */
1486 			break;
1487 		case 3:
1488 			/* XXX page queue*/
1489 			break;
1490 		}
1491 		break;
1492 	case SOC15_IH_CLIENTID_SDMA3_Sienna_Cichlid:
1493 		switch (entry->ring_id) {
1494 		case 0:
1495 			amdgpu_fence_process(&adev->sdma.instance[3].ring);
1496 			break;
1497 		case 1:
1498 			/* XXX compute */
1499 			break;
1500 		case 2:
1501 			/* XXX compute */
1502 			break;
1503 		case 3:
1504 			/* XXX page queue*/
1505 			break;
1506 		}
1507 		break;
1508 	}
1509 	return 0;
1510 }
1511 
1512 static int sdma_v5_2_process_illegal_inst_irq(struct amdgpu_device *adev,
1513 					      struct amdgpu_irq_src *source,
1514 					      struct amdgpu_iv_entry *entry)
1515 {
1516 	return 0;
1517 }
1518 
1519 static void sdma_v5_2_update_medium_grain_clock_gating(struct amdgpu_device *adev,
1520 						       bool enable)
1521 {
1522 	uint32_t data, def;
1523 	int i;
1524 
1525 	for (i = 0; i < adev->sdma.num_instances; i++) {
1526 		if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
1527 			/* Enable sdma clock gating */
1528 			def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL));
1529 			data &= ~(SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1530 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1531 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1532 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1533 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK |
1534 				  SDMA0_CLK_CTRL__SOFT_OVERRIDER_REG_MASK);
1535 			if (def != data)
1536 				WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data);
1537 		} else {
1538 			/* Disable sdma clock gating */
1539 			def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL));
1540 			data |= (SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1541 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1542 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1543 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1544 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK |
1545 				 SDMA0_CLK_CTRL__SOFT_OVERRIDER_REG_MASK);
1546 			if (def != data)
1547 				WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data);
1548 		}
1549 	}
1550 }
1551 
1552 static void sdma_v5_2_update_medium_grain_light_sleep(struct amdgpu_device *adev,
1553 						      bool enable)
1554 {
1555 	uint32_t data, def;
1556 	int i;
1557 
1558 	for (i = 0; i < adev->sdma.num_instances; i++) {
1559 		if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) {
1560 			/* Enable sdma mem light sleep */
1561 			def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL));
1562 			data |= SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1563 			if (def != data)
1564 				WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data);
1565 
1566 		} else {
1567 			/* Disable sdma mem light sleep */
1568 			def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL));
1569 			data &= ~SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1570 			if (def != data)
1571 				WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data);
1572 
1573 		}
1574 	}
1575 }
1576 
1577 static int sdma_v5_2_set_clockgating_state(void *handle,
1578 					   enum amd_clockgating_state state)
1579 {
1580 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1581 
1582 	if (amdgpu_sriov_vf(adev))
1583 		return 0;
1584 
1585 	switch (adev->asic_type) {
1586 	case CHIP_SIENNA_CICHLID:
1587 	case CHIP_NAVY_FLOUNDER:
1588 	case CHIP_VANGOGH:
1589 	case CHIP_DIMGREY_CAVEFISH:
1590 		sdma_v5_2_update_medium_grain_clock_gating(adev,
1591 				state == AMD_CG_STATE_GATE ? true : false);
1592 		sdma_v5_2_update_medium_grain_light_sleep(adev,
1593 				state == AMD_CG_STATE_GATE ? true : false);
1594 		break;
1595 	default:
1596 		break;
1597 	}
1598 
1599 	return 0;
1600 }
1601 
1602 static int sdma_v5_2_set_powergating_state(void *handle,
1603 					  enum amd_powergating_state state)
1604 {
1605 	return 0;
1606 }
1607 
1608 static void sdma_v5_2_get_clockgating_state(void *handle, u32 *flags)
1609 {
1610 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1611 	int data;
1612 
1613 	if (amdgpu_sriov_vf(adev))
1614 		*flags = 0;
1615 
1616 	/* AMD_CG_SUPPORT_SDMA_LS */
1617 	data = RREG32_KIQ(sdma_v5_2_get_reg_offset(adev, 0, mmSDMA0_POWER_CNTL));
1618 	if (data & SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK)
1619 		*flags |= AMD_CG_SUPPORT_SDMA_LS;
1620 }
1621 
1622 const struct amd_ip_funcs sdma_v5_2_ip_funcs = {
1623 	.name = "sdma_v5_2",
1624 	.early_init = sdma_v5_2_early_init,
1625 	.late_init = NULL,
1626 	.sw_init = sdma_v5_2_sw_init,
1627 	.sw_fini = sdma_v5_2_sw_fini,
1628 	.hw_init = sdma_v5_2_hw_init,
1629 	.hw_fini = sdma_v5_2_hw_fini,
1630 	.suspend = sdma_v5_2_suspend,
1631 	.resume = sdma_v5_2_resume,
1632 	.is_idle = sdma_v5_2_is_idle,
1633 	.wait_for_idle = sdma_v5_2_wait_for_idle,
1634 	.soft_reset = sdma_v5_2_soft_reset,
1635 	.set_clockgating_state = sdma_v5_2_set_clockgating_state,
1636 	.set_powergating_state = sdma_v5_2_set_powergating_state,
1637 	.get_clockgating_state = sdma_v5_2_get_clockgating_state,
1638 };
1639 
1640 static const struct amdgpu_ring_funcs sdma_v5_2_ring_funcs = {
1641 	.type = AMDGPU_RING_TYPE_SDMA,
1642 	.align_mask = 0xf,
1643 	.nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP),
1644 	.support_64bit_ptrs = true,
1645 	.vmhub = AMDGPU_GFXHUB_0,
1646 	.get_rptr = sdma_v5_2_ring_get_rptr,
1647 	.get_wptr = sdma_v5_2_ring_get_wptr,
1648 	.set_wptr = sdma_v5_2_ring_set_wptr,
1649 	.emit_frame_size =
1650 		5 + /* sdma_v5_2_ring_init_cond_exec */
1651 		6 + /* sdma_v5_2_ring_emit_hdp_flush */
1652 		3 + /* hdp_invalidate */
1653 		6 + /* sdma_v5_2_ring_emit_pipeline_sync */
1654 		/* sdma_v5_2_ring_emit_vm_flush */
1655 		SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 +
1656 		SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6 +
1657 		10 + 10 + 10, /* sdma_v5_2_ring_emit_fence x3 for user fence, vm fence */
1658 	.emit_ib_size = 7 + 6, /* sdma_v5_2_ring_emit_ib */
1659 	.emit_ib = sdma_v5_2_ring_emit_ib,
1660 	.emit_fence = sdma_v5_2_ring_emit_fence,
1661 	.emit_pipeline_sync = sdma_v5_2_ring_emit_pipeline_sync,
1662 	.emit_vm_flush = sdma_v5_2_ring_emit_vm_flush,
1663 	.emit_hdp_flush = sdma_v5_2_ring_emit_hdp_flush,
1664 	.test_ring = sdma_v5_2_ring_test_ring,
1665 	.test_ib = sdma_v5_2_ring_test_ib,
1666 	.insert_nop = sdma_v5_2_ring_insert_nop,
1667 	.pad_ib = sdma_v5_2_ring_pad_ib,
1668 	.emit_wreg = sdma_v5_2_ring_emit_wreg,
1669 	.emit_reg_wait = sdma_v5_2_ring_emit_reg_wait,
1670 	.emit_reg_write_reg_wait = sdma_v5_2_ring_emit_reg_write_reg_wait,
1671 	.init_cond_exec = sdma_v5_2_ring_init_cond_exec,
1672 	.patch_cond_exec = sdma_v5_2_ring_patch_cond_exec,
1673 	.preempt_ib = sdma_v5_2_ring_preempt_ib,
1674 };
1675 
1676 static void sdma_v5_2_set_ring_funcs(struct amdgpu_device *adev)
1677 {
1678 	int i;
1679 
1680 	for (i = 0; i < adev->sdma.num_instances; i++) {
1681 		adev->sdma.instance[i].ring.funcs = &sdma_v5_2_ring_funcs;
1682 		adev->sdma.instance[i].ring.me = i;
1683 	}
1684 }
1685 
1686 static const struct amdgpu_irq_src_funcs sdma_v5_2_trap_irq_funcs = {
1687 	.set = sdma_v5_2_set_trap_irq_state,
1688 	.process = sdma_v5_2_process_trap_irq,
1689 };
1690 
1691 static const struct amdgpu_irq_src_funcs sdma_v5_2_illegal_inst_irq_funcs = {
1692 	.process = sdma_v5_2_process_illegal_inst_irq,
1693 };
1694 
1695 static void sdma_v5_2_set_irq_funcs(struct amdgpu_device *adev)
1696 {
1697 	adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_INSTANCE0 +
1698 					adev->sdma.num_instances;
1699 	adev->sdma.trap_irq.funcs = &sdma_v5_2_trap_irq_funcs;
1700 	adev->sdma.illegal_inst_irq.funcs = &sdma_v5_2_illegal_inst_irq_funcs;
1701 }
1702 
1703 /**
1704  * sdma_v5_2_emit_copy_buffer - copy buffer using the sDMA engine
1705  *
1706  * @ib: indirect buffer to copy to
1707  * @src_offset: src GPU address
1708  * @dst_offset: dst GPU address
1709  * @byte_count: number of bytes to xfer
1710  * @tmz: if a secure copy should be used
1711  *
1712  * Copy GPU buffers using the DMA engine.
1713  * Used by the amdgpu ttm implementation to move pages if
1714  * registered as the asic copy callback.
1715  */
1716 static void sdma_v5_2_emit_copy_buffer(struct amdgpu_ib *ib,
1717 				       uint64_t src_offset,
1718 				       uint64_t dst_offset,
1719 				       uint32_t byte_count,
1720 				       bool tmz)
1721 {
1722 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1723 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR) |
1724 		SDMA_PKT_COPY_LINEAR_HEADER_TMZ(tmz ? 1 : 0);
1725 	ib->ptr[ib->length_dw++] = byte_count - 1;
1726 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1727 	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1728 	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1729 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1730 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1731 }
1732 
1733 /**
1734  * sdma_v5_2_emit_fill_buffer - fill buffer using the sDMA engine
1735  *
1736  * @ib: indirect buffer to fill
1737  * @src_data: value to write to buffer
1738  * @dst_offset: dst GPU address
1739  * @byte_count: number of bytes to xfer
1740  *
1741  * Fill GPU buffers using the DMA engine.
1742  */
1743 static void sdma_v5_2_emit_fill_buffer(struct amdgpu_ib *ib,
1744 				       uint32_t src_data,
1745 				       uint64_t dst_offset,
1746 				       uint32_t byte_count)
1747 {
1748 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL);
1749 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1750 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1751 	ib->ptr[ib->length_dw++] = src_data;
1752 	ib->ptr[ib->length_dw++] = byte_count - 1;
1753 }
1754 
1755 static const struct amdgpu_buffer_funcs sdma_v5_2_buffer_funcs = {
1756 	.copy_max_bytes = 0x400000,
1757 	.copy_num_dw = 7,
1758 	.emit_copy_buffer = sdma_v5_2_emit_copy_buffer,
1759 
1760 	.fill_max_bytes = 0x400000,
1761 	.fill_num_dw = 5,
1762 	.emit_fill_buffer = sdma_v5_2_emit_fill_buffer,
1763 };
1764 
1765 static void sdma_v5_2_set_buffer_funcs(struct amdgpu_device *adev)
1766 {
1767 	if (adev->mman.buffer_funcs == NULL) {
1768 		adev->mman.buffer_funcs = &sdma_v5_2_buffer_funcs;
1769 		adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1770 	}
1771 }
1772 
1773 static const struct amdgpu_vm_pte_funcs sdma_v5_2_vm_pte_funcs = {
1774 	.copy_pte_num_dw = 7,
1775 	.copy_pte = sdma_v5_2_vm_copy_pte,
1776 	.write_pte = sdma_v5_2_vm_write_pte,
1777 	.set_pte_pde = sdma_v5_2_vm_set_pte_pde,
1778 };
1779 
1780 static void sdma_v5_2_set_vm_pte_funcs(struct amdgpu_device *adev)
1781 {
1782 	unsigned i;
1783 
1784 	if (adev->vm_manager.vm_pte_funcs == NULL) {
1785 		adev->vm_manager.vm_pte_funcs = &sdma_v5_2_vm_pte_funcs;
1786 		for (i = 0; i < adev->sdma.num_instances; i++) {
1787 			adev->vm_manager.vm_pte_scheds[i] =
1788 				&adev->sdma.instance[i].ring.sched;
1789 		}
1790 		adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
1791 	}
1792 }
1793 
1794 const struct amdgpu_ip_block_version sdma_v5_2_ip_block = {
1795 	.type = AMD_IP_BLOCK_TYPE_SDMA,
1796 	.major = 5,
1797 	.minor = 2,
1798 	.rev = 0,
1799 	.funcs = &sdma_v5_2_ip_funcs,
1800 };
1801