xref: /linux/drivers/gpu/drm/amd/amdgpu/sdma_v6_0.c (revision 3d0fe49454652117522f60bfbefb978ba0e5300b)
1 /*
2  * Copyright 2020 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_11_0_0_offset.h"
34 #include "gc/gc_11_0_0_sh_mask.h"
35 #include "gc/gc_11_0_0_default.h"
36 #include "hdp/hdp_6_0_0_offset.h"
37 #include "ivsrcid/gfx/irqsrcs_gfx_11_0_0.h"
38 
39 #include "soc15_common.h"
40 #include "soc15.h"
41 #include "sdma_v6_0_0_pkt_open.h"
42 #include "nbio_v4_3.h"
43 #include "sdma_common.h"
44 #include "sdma_v6_0.h"
45 #include "v11_structs.h"
46 
47 MODULE_FIRMWARE("amdgpu/sdma_6_0_0.bin");
48 MODULE_FIRMWARE("amdgpu/sdma_6_0_1.bin");
49 MODULE_FIRMWARE("amdgpu/sdma_6_0_2.bin");
50 MODULE_FIRMWARE("amdgpu/sdma_6_0_3.bin");
51 MODULE_FIRMWARE("amdgpu/sdma_6_1_0.bin");
52 
53 #define SDMA1_REG_OFFSET 0x600
54 #define SDMA0_HYP_DEC_REG_START 0x5880
55 #define SDMA0_HYP_DEC_REG_END 0x589a
56 #define SDMA1_HYP_DEC_REG_OFFSET 0x20
57 
58 static void sdma_v6_0_set_ring_funcs(struct amdgpu_device *adev);
59 static void sdma_v6_0_set_buffer_funcs(struct amdgpu_device *adev);
60 static void sdma_v6_0_set_vm_pte_funcs(struct amdgpu_device *adev);
61 static void sdma_v6_0_set_irq_funcs(struct amdgpu_device *adev);
62 static int sdma_v6_0_start(struct amdgpu_device *adev);
63 
64 static u32 sdma_v6_0_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 		base = adev->reg_offset[GC_HWIP][0][0];
75 		if (instance == 1)
76 			internal_offset += SDMA1_REG_OFFSET;
77 	}
78 
79 	return base + internal_offset;
80 }
81 
82 static unsigned sdma_v6_0_ring_init_cond_exec(struct amdgpu_ring *ring)
83 {
84 	unsigned ret;
85 
86 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_COND_EXE));
87 	amdgpu_ring_write(ring, lower_32_bits(ring->cond_exe_gpu_addr));
88 	amdgpu_ring_write(ring, upper_32_bits(ring->cond_exe_gpu_addr));
89 	amdgpu_ring_write(ring, 1);
90 	ret = ring->wptr & ring->buf_mask;/* this is the offset we need patch later */
91 	amdgpu_ring_write(ring, 0x55aa55aa);/* insert dummy here and patch it later */
92 
93 	return ret;
94 }
95 
96 static void sdma_v6_0_ring_patch_cond_exec(struct amdgpu_ring *ring,
97 					   unsigned offset)
98 {
99 	unsigned cur;
100 
101 	BUG_ON(offset > ring->buf_mask);
102 	BUG_ON(ring->ring[offset] != 0x55aa55aa);
103 
104 	cur = (ring->wptr - 1) & ring->buf_mask;
105 	if (cur > offset)
106 		ring->ring[offset] = cur - offset;
107 	else
108 		ring->ring[offset] = (ring->buf_mask + 1) - offset + cur;
109 }
110 
111 /**
112  * sdma_v6_0_ring_get_rptr - get the current read pointer
113  *
114  * @ring: amdgpu ring pointer
115  *
116  * Get the current rptr from the hardware.
117  */
118 static uint64_t sdma_v6_0_ring_get_rptr(struct amdgpu_ring *ring)
119 {
120 	u64 *rptr;
121 
122 	/* XXX check if swapping is necessary on BE */
123 	rptr = (u64 *)ring->rptr_cpu_addr;
124 
125 	DRM_DEBUG("rptr before shift == 0x%016llx\n", *rptr);
126 	return ((*rptr) >> 2);
127 }
128 
129 /**
130  * sdma_v6_0_ring_get_wptr - get the current write pointer
131  *
132  * @ring: amdgpu ring pointer
133  *
134  * Get the current wptr from the hardware.
135  */
136 static uint64_t sdma_v6_0_ring_get_wptr(struct amdgpu_ring *ring)
137 {
138 	u64 wptr = 0;
139 
140 	if (ring->use_doorbell) {
141 		/* XXX check if swapping is necessary on BE */
142 		wptr = READ_ONCE(*((u64 *)ring->wptr_cpu_addr));
143 		DRM_DEBUG("wptr/doorbell before shift == 0x%016llx\n", wptr);
144 	}
145 
146 	return wptr >> 2;
147 }
148 
149 /**
150  * sdma_v6_0_ring_set_wptr - commit the write pointer
151  *
152  * @ring: amdgpu ring pointer
153  *
154  * Write the wptr back to the hardware.
155  */
156 static void sdma_v6_0_ring_set_wptr(struct amdgpu_ring *ring)
157 {
158 	struct amdgpu_device *adev = ring->adev;
159 
160 	if (ring->use_doorbell) {
161 		DRM_DEBUG("Using doorbell -- "
162 			  "wptr_offs == 0x%08x "
163 			  "lower_32_bits(ring->wptr) << 2 == 0x%08x "
164 			  "upper_32_bits(ring->wptr) << 2 == 0x%08x\n",
165 			  ring->wptr_offs,
166 			  lower_32_bits(ring->wptr << 2),
167 			  upper_32_bits(ring->wptr << 2));
168 		/* XXX check if swapping is necessary on BE */
169 		atomic64_set((atomic64_t *)ring->wptr_cpu_addr,
170 			     ring->wptr << 2);
171 		DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016llx)\n",
172 			  ring->doorbell_index, ring->wptr << 2);
173 		WDOORBELL64(ring->doorbell_index, ring->wptr << 2);
174 	} else {
175 		DRM_DEBUG("Not using doorbell -- "
176 			  "regSDMA%i_GFX_RB_WPTR == 0x%08x "
177 			  "regSDMA%i_GFX_RB_WPTR_HI == 0x%08x\n",
178 			  ring->me,
179 			  lower_32_bits(ring->wptr << 2),
180 			  ring->me,
181 			  upper_32_bits(ring->wptr << 2));
182 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev,
183 							     ring->me, regSDMA0_QUEUE0_RB_WPTR),
184 				lower_32_bits(ring->wptr << 2));
185 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev,
186 							     ring->me, regSDMA0_QUEUE0_RB_WPTR_HI),
187 				upper_32_bits(ring->wptr << 2));
188 	}
189 }
190 
191 static void sdma_v6_0_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
192 {
193 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
194 	int i;
195 
196 	for (i = 0; i < count; i++)
197 		if (sdma && sdma->burst_nop && (i == 0))
198 			amdgpu_ring_write(ring, ring->funcs->nop |
199 				SDMA_PKT_NOP_HEADER_COUNT(count - 1));
200 		else
201 			amdgpu_ring_write(ring, ring->funcs->nop);
202 }
203 
204 /*
205  * sdma_v6_0_ring_emit_ib - Schedule an IB on the DMA engine
206  *
207  * @ring: amdgpu ring pointer
208  * @ib: IB object to schedule
209  * @flags: unused
210  * @job: job to retrieve vmid from
211  *
212  * Schedule an IB in the DMA ring.
213  */
214 static void sdma_v6_0_ring_emit_ib(struct amdgpu_ring *ring,
215 				   struct amdgpu_job *job,
216 				   struct amdgpu_ib *ib,
217 				   uint32_t flags)
218 {
219 	unsigned vmid = AMDGPU_JOB_GET_VMID(job);
220 	uint64_t csa_mc_addr = amdgpu_sdma_get_csa_mc_addr(ring, vmid);
221 
222 	/* An IB packet must end on a 8 DW boundary--the next dword
223 	 * must be on a 8-dword boundary. Our IB packet below is 6
224 	 * dwords long, thus add x number of NOPs, such that, in
225 	 * modular arithmetic,
226 	 * wptr + 6 + x = 8k, k >= 0, which in C is,
227 	 * (wptr + 6 + x) % 8 = 0.
228 	 * The expression below, is a solution of x.
229 	 */
230 	sdma_v6_0_ring_insert_nop(ring, (2 - lower_32_bits(ring->wptr)) & 7);
231 
232 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_INDIRECT) |
233 			  SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf));
234 	/* base must be 32 byte aligned */
235 	amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
236 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
237 	amdgpu_ring_write(ring, ib->length_dw);
238 	amdgpu_ring_write(ring, lower_32_bits(csa_mc_addr));
239 	amdgpu_ring_write(ring, upper_32_bits(csa_mc_addr));
240 }
241 
242 /**
243  * sdma_v6_0_ring_emit_mem_sync - flush the IB by graphics cache rinse
244  *
245  * @ring: amdgpu ring pointer
246  *
247  * flush the IB by graphics cache rinse.
248  */
249 static void sdma_v6_0_ring_emit_mem_sync(struct amdgpu_ring *ring)
250 {
251         uint32_t gcr_cntl = SDMA_GCR_GL2_INV | SDMA_GCR_GL2_WB | SDMA_GCR_GLM_INV |
252                             SDMA_GCR_GL1_INV | SDMA_GCR_GLV_INV | SDMA_GCR_GLK_INV |
253                             SDMA_GCR_GLI_INV(1);
254 
255         /* flush entire cache L0/L1/L2, this can be optimized by performance requirement */
256         amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_GCR_REQ));
257         amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD1_BASE_VA_31_7(0));
258         amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD2_GCR_CONTROL_15_0(gcr_cntl) |
259                           SDMA_PKT_GCR_REQ_PAYLOAD2_BASE_VA_47_32(0));
260         amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD3_LIMIT_VA_31_7(0) |
261                           SDMA_PKT_GCR_REQ_PAYLOAD3_GCR_CONTROL_18_16(gcr_cntl >> 16));
262         amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD4_LIMIT_VA_47_32(0) |
263                           SDMA_PKT_GCR_REQ_PAYLOAD4_VMID(0));
264 }
265 
266 
267 /**
268  * sdma_v6_0_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
269  *
270  * @ring: amdgpu ring pointer
271  *
272  * Emit an hdp flush packet on the requested DMA ring.
273  */
274 static void sdma_v6_0_ring_emit_hdp_flush(struct amdgpu_ring *ring)
275 {
276 	struct amdgpu_device *adev = ring->adev;
277 	u32 ref_and_mask = 0;
278 	const struct nbio_hdp_flush_reg *nbio_hf_reg = adev->nbio.hdp_flush_reg;
279 
280 	ref_and_mask = nbio_hf_reg->ref_and_mask_sdma0 << ring->me;
281 
282 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_POLL_REGMEM) |
283 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) |
284 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
285 	amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_done_offset(adev)) << 2);
286 	amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_req_offset(adev)) << 2);
287 	amdgpu_ring_write(ring, ref_and_mask); /* reference */
288 	amdgpu_ring_write(ring, ref_and_mask); /* mask */
289 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
290 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
291 }
292 
293 /**
294  * sdma_v6_0_ring_emit_fence - emit a fence on the DMA ring
295  *
296  * @ring: amdgpu ring pointer
297  * @addr: address
298  * @seq: fence seq number
299  * @flags: fence flags
300  *
301  * Add a DMA fence packet to the ring to write
302  * the fence seq number and DMA trap packet to generate
303  * an interrupt if needed.
304  */
305 static void sdma_v6_0_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
306 				      unsigned flags)
307 {
308 	bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
309 	/* write the fence */
310 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_FENCE) |
311 			  SDMA_PKT_FENCE_HEADER_MTYPE(0x3)); /* Ucached(UC) */
312 	/* zero in first two bits */
313 	BUG_ON(addr & 0x3);
314 	amdgpu_ring_write(ring, lower_32_bits(addr));
315 	amdgpu_ring_write(ring, upper_32_bits(addr));
316 	amdgpu_ring_write(ring, lower_32_bits(seq));
317 
318 	/* optionally write high bits as well */
319 	if (write64bit) {
320 		addr += 4;
321 		amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_FENCE) |
322 				  SDMA_PKT_FENCE_HEADER_MTYPE(0x3));
323 		/* zero in first two bits */
324 		BUG_ON(addr & 0x3);
325 		amdgpu_ring_write(ring, lower_32_bits(addr));
326 		amdgpu_ring_write(ring, upper_32_bits(addr));
327 		amdgpu_ring_write(ring, upper_32_bits(seq));
328 	}
329 
330 	if (flags & AMDGPU_FENCE_FLAG_INT) {
331 		uint32_t ctx = ring->is_mes_queue ?
332 			(ring->hw_queue_id | AMDGPU_FENCE_MES_QUEUE_FLAG) : 0;
333 		/* generate an interrupt */
334 		amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_TRAP));
335 		amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(ctx));
336 	}
337 }
338 
339 /**
340  * sdma_v6_0_gfx_stop - stop the gfx async dma engines
341  *
342  * @adev: amdgpu_device pointer
343  *
344  * Stop the gfx async dma ring buffers.
345  */
346 static void sdma_v6_0_gfx_stop(struct amdgpu_device *adev)
347 {
348 	u32 rb_cntl, ib_cntl;
349 	int i;
350 
351 	for (i = 0; i < adev->sdma.num_instances; i++) {
352 		rb_cntl = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_CNTL));
353 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RB_ENABLE, 0);
354 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_CNTL), rb_cntl);
355 		ib_cntl = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_IB_CNTL));
356 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_QUEUE0_IB_CNTL, IB_ENABLE, 0);
357 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_IB_CNTL), ib_cntl);
358 	}
359 }
360 
361 /**
362  * sdma_v6_0_rlc_stop - stop the compute async dma engines
363  *
364  * @adev: amdgpu_device pointer
365  *
366  * Stop the compute async dma queues.
367  */
368 static void sdma_v6_0_rlc_stop(struct amdgpu_device *adev)
369 {
370 	/* XXX todo */
371 }
372 
373 /**
374  * sdma_v6_0_ctxempty_int_enable - enable or disable context empty interrupts
375  *
376  * @adev: amdgpu_device pointer
377  * @enable: enable/disable context switching due to queue empty conditions
378  *
379  * Enable or disable the async dma engines queue empty context switch.
380  */
381 static void sdma_v6_0_ctxempty_int_enable(struct amdgpu_device *adev, bool enable)
382 {
383 	u32 f32_cntl;
384 	int i;
385 
386 	if (!amdgpu_sriov_vf(adev)) {
387 		for (i = 0; i < adev->sdma.num_instances; i++) {
388 			f32_cntl = RREG32(sdma_v6_0_get_reg_offset(adev, i, regSDMA0_CNTL));
389 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
390 					CTXEMPTY_INT_ENABLE, enable ? 1 : 0);
391 			WREG32(sdma_v6_0_get_reg_offset(adev, i, regSDMA0_CNTL), f32_cntl);
392 		}
393 	}
394 }
395 
396 /**
397  * sdma_v6_0_enable - stop the async dma engines
398  *
399  * @adev: amdgpu_device pointer
400  * @enable: enable/disable the DMA MEs.
401  *
402  * Halt or unhalt the async dma engines.
403  */
404 static void sdma_v6_0_enable(struct amdgpu_device *adev, bool enable)
405 {
406 	u32 f32_cntl;
407 	int i;
408 
409 	if (!enable) {
410 		sdma_v6_0_gfx_stop(adev);
411 		sdma_v6_0_rlc_stop(adev);
412 	}
413 
414 	if (amdgpu_sriov_vf(adev))
415 		return;
416 
417 	for (i = 0; i < adev->sdma.num_instances; i++) {
418 		f32_cntl = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_F32_CNTL));
419 		f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, enable ? 0 : 1);
420 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_F32_CNTL), f32_cntl);
421 	}
422 }
423 
424 /**
425  * sdma_v6_0_gfx_resume - setup and start the async dma engines
426  *
427  * @adev: amdgpu_device pointer
428  *
429  * Set up the gfx DMA ring buffers and enable them.
430  * Returns 0 for success, error for failure.
431  */
432 static int sdma_v6_0_gfx_resume(struct amdgpu_device *adev)
433 {
434 	struct amdgpu_ring *ring;
435 	u32 rb_cntl, ib_cntl;
436 	u32 rb_bufsz;
437 	u32 doorbell;
438 	u32 doorbell_offset;
439 	u32 temp;
440 	u64 wptr_gpu_addr;
441 	int i, r;
442 
443 	for (i = 0; i < adev->sdma.num_instances; i++) {
444 		ring = &adev->sdma.instance[i].ring;
445 
446 		if (!amdgpu_sriov_vf(adev))
447 			WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_SEM_WAIT_FAIL_TIMER_CNTL), 0);
448 
449 		/* Set ring buffer size in dwords */
450 		rb_bufsz = order_base_2(ring->ring_size / 4);
451 		rb_cntl = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_CNTL));
452 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RB_SIZE, rb_bufsz);
453 #ifdef __BIG_ENDIAN
454 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RB_SWAP_ENABLE, 1);
455 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL,
456 					RPTR_WRITEBACK_SWAP_ENABLE, 1);
457 #endif
458 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RB_PRIV, 1);
459 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_CNTL), rb_cntl);
460 
461 		/* Initialize the ring buffer's read and write pointers */
462 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_RPTR), 0);
463 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_RPTR_HI), 0);
464 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR), 0);
465 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR_HI), 0);
466 
467 		/* setup the wptr shadow polling */
468 		wptr_gpu_addr = ring->wptr_gpu_addr;
469 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR_POLL_ADDR_LO),
470 		       lower_32_bits(wptr_gpu_addr));
471 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR_POLL_ADDR_HI),
472 		       upper_32_bits(wptr_gpu_addr));
473 
474 		/* set the wb address whether it's enabled or not */
475 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_RPTR_ADDR_HI),
476 		       upper_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFF);
477 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_RPTR_ADDR_LO),
478 		       lower_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFC);
479 
480 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
481 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, WPTR_POLL_ENABLE, 0);
482 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, F32_WPTR_POLL_ENABLE, 1);
483 
484 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_BASE), ring->gpu_addr >> 8);
485 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_BASE_HI), ring->gpu_addr >> 40);
486 
487 		ring->wptr = 0;
488 
489 		/* before programing wptr to a less value, need set minor_ptr_update first */
490 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_MINOR_PTR_UPDATE), 1);
491 
492 		if (!amdgpu_sriov_vf(adev)) { /* only bare-metal use register write for wptr */
493 			WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR), lower_32_bits(ring->wptr) << 2);
494 			WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR_HI), upper_32_bits(ring->wptr) << 2);
495 		}
496 
497 		doorbell = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_DOORBELL));
498 		doorbell_offset = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_DOORBELL_OFFSET));
499 
500 		if (ring->use_doorbell) {
501 			doorbell = REG_SET_FIELD(doorbell, SDMA0_QUEUE0_DOORBELL, ENABLE, 1);
502 			doorbell_offset = REG_SET_FIELD(doorbell_offset, SDMA0_QUEUE0_DOORBELL_OFFSET,
503 					OFFSET, ring->doorbell_index);
504 		} else {
505 			doorbell = REG_SET_FIELD(doorbell, SDMA0_QUEUE0_DOORBELL, ENABLE, 0);
506 		}
507 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_DOORBELL), doorbell);
508 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_DOORBELL_OFFSET), doorbell_offset);
509 
510 		if (i == 0)
511 			adev->nbio.funcs->sdma_doorbell_range(adev, i, ring->use_doorbell,
512 						      ring->doorbell_index,
513 						      adev->doorbell_index.sdma_doorbell_range * adev->sdma.num_instances);
514 
515 		if (amdgpu_sriov_vf(adev))
516 			sdma_v6_0_ring_set_wptr(ring);
517 
518 		/* set minor_ptr_update to 0 after wptr programed */
519 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_MINOR_PTR_UPDATE), 0);
520 
521 		/* Set up RESP_MODE to non-copy addresses */
522 		temp = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UTCL1_CNTL));
523 		temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, RESP_MODE, 3);
524 		temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, REDO_DELAY, 9);
525 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UTCL1_CNTL), temp);
526 
527 		/* program default cache read and write policy */
528 		temp = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UTCL1_PAGE));
529 		/* clean read policy and write policy bits */
530 		temp &= 0xFF0FFF;
531 		temp |= ((CACHE_READ_POLICY_L2__DEFAULT << 12) |
532 			 (CACHE_WRITE_POLICY_L2__DEFAULT << 14) |
533 			 SDMA0_UTCL1_PAGE__LLC_NOALLOC_MASK);
534 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UTCL1_PAGE), temp);
535 
536 		if (!amdgpu_sriov_vf(adev)) {
537 			/* unhalt engine */
538 			temp = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_F32_CNTL));
539 			temp = REG_SET_FIELD(temp, SDMA0_F32_CNTL, HALT, 0);
540 			temp = REG_SET_FIELD(temp, SDMA0_F32_CNTL, TH1_RESET, 0);
541 			WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_F32_CNTL), temp);
542 		}
543 
544 		/* enable DMA RB */
545 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RB_ENABLE, 1);
546 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_CNTL), rb_cntl);
547 
548 		ib_cntl = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_IB_CNTL));
549 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_QUEUE0_IB_CNTL, IB_ENABLE, 1);
550 #ifdef __BIG_ENDIAN
551 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_QUEUE0_IB_CNTL, IB_SWAP_ENABLE, 1);
552 #endif
553 		/* enable DMA IBs */
554 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_IB_CNTL), ib_cntl);
555 
556 		if (amdgpu_sriov_vf(adev))
557 			sdma_v6_0_enable(adev, true);
558 
559 		r = amdgpu_ring_test_helper(ring);
560 		if (r)
561 			return r;
562 	}
563 
564 	return 0;
565 }
566 
567 /**
568  * sdma_v6_0_rlc_resume - setup and start the async dma engines
569  *
570  * @adev: amdgpu_device pointer
571  *
572  * Set up the compute DMA queues and enable them.
573  * Returns 0 for success, error for failure.
574  */
575 static int sdma_v6_0_rlc_resume(struct amdgpu_device *adev)
576 {
577 	return 0;
578 }
579 
580 /**
581  * sdma_v6_0_load_microcode - load the sDMA ME ucode
582  *
583  * @adev: amdgpu_device pointer
584  *
585  * Loads the sDMA0/1 ucode.
586  * Returns 0 for success, -EINVAL if the ucode is not available.
587  */
588 static int sdma_v6_0_load_microcode(struct amdgpu_device *adev)
589 {
590 	const struct sdma_firmware_header_v2_0 *hdr;
591 	const __le32 *fw_data;
592 	u32 fw_size;
593 	int i, j;
594 	bool use_broadcast;
595 
596 	/* halt the MEs */
597 	sdma_v6_0_enable(adev, false);
598 
599 	if (!adev->sdma.instance[0].fw)
600 		return -EINVAL;
601 
602 	/* use broadcast mode to load SDMA microcode by default */
603 	use_broadcast = true;
604 
605 	if (use_broadcast) {
606 		dev_info(adev->dev, "Use broadcast method to load SDMA firmware\n");
607 		/* load Control Thread microcode */
608 		hdr = (const struct sdma_firmware_header_v2_0 *)adev->sdma.instance[0].fw->data;
609 		amdgpu_ucode_print_sdma_hdr(&hdr->header);
610 		fw_size = le32_to_cpu(hdr->ctx_jt_offset + hdr->ctx_jt_size) / 4;
611 
612 		fw_data = (const __le32 *)
613 			(adev->sdma.instance[0].fw->data +
614 				le32_to_cpu(hdr->header.ucode_array_offset_bytes));
615 
616 		WREG32(sdma_v6_0_get_reg_offset(adev, 0, regSDMA0_BROADCAST_UCODE_ADDR), 0);
617 
618 		for (j = 0; j < fw_size; j++) {
619 			if (amdgpu_emu_mode == 1 && j % 500 == 0)
620 				msleep(1);
621 			WREG32(sdma_v6_0_get_reg_offset(adev, 0, regSDMA0_BROADCAST_UCODE_DATA), le32_to_cpup(fw_data++));
622 		}
623 
624 		/* load Context Switch microcode */
625 		fw_size = le32_to_cpu(hdr->ctl_jt_offset + hdr->ctl_jt_size) / 4;
626 
627 		fw_data = (const __le32 *)
628 			(adev->sdma.instance[0].fw->data +
629 				le32_to_cpu(hdr->ctl_ucode_offset));
630 
631 		WREG32(sdma_v6_0_get_reg_offset(adev, 0, regSDMA0_BROADCAST_UCODE_ADDR), 0x8000);
632 
633 		for (j = 0; j < fw_size; j++) {
634 			if (amdgpu_emu_mode == 1 && j % 500 == 0)
635 				msleep(1);
636 			WREG32(sdma_v6_0_get_reg_offset(adev, 0, regSDMA0_BROADCAST_UCODE_DATA), le32_to_cpup(fw_data++));
637 		}
638 	} else {
639 		dev_info(adev->dev, "Use legacy method to load SDMA firmware\n");
640 		for (i = 0; i < adev->sdma.num_instances; i++) {
641 			/* load Control Thread microcode */
642 			hdr = (const struct sdma_firmware_header_v2_0 *)adev->sdma.instance[0].fw->data;
643 			amdgpu_ucode_print_sdma_hdr(&hdr->header);
644 			fw_size = le32_to_cpu(hdr->ctx_jt_offset + hdr->ctx_jt_size) / 4;
645 
646 			fw_data = (const __le32 *)
647 				(adev->sdma.instance[0].fw->data +
648 					le32_to_cpu(hdr->header.ucode_array_offset_bytes));
649 
650 			WREG32(sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UCODE_ADDR), 0);
651 
652 			for (j = 0; j < fw_size; j++) {
653 				if (amdgpu_emu_mode == 1 && j % 500 == 0)
654 					msleep(1);
655 				WREG32(sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UCODE_DATA), le32_to_cpup(fw_data++));
656 			}
657 
658 			WREG32(sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UCODE_ADDR), adev->sdma.instance[0].fw_version);
659 
660 			/* load Context Switch microcode */
661 			fw_size = le32_to_cpu(hdr->ctl_jt_offset + hdr->ctl_jt_size) / 4;
662 
663 			fw_data = (const __le32 *)
664 				(adev->sdma.instance[0].fw->data +
665 					le32_to_cpu(hdr->ctl_ucode_offset));
666 
667 			WREG32(sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UCODE_ADDR), 0x8000);
668 
669 			for (j = 0; j < fw_size; j++) {
670 				if (amdgpu_emu_mode == 1 && j % 500 == 0)
671 					msleep(1);
672 				WREG32(sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UCODE_DATA), le32_to_cpup(fw_data++));
673 			}
674 
675 			WREG32(sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UCODE_ADDR), adev->sdma.instance[0].fw_version);
676 		}
677 	}
678 
679 	return 0;
680 }
681 
682 static int sdma_v6_0_soft_reset(void *handle)
683 {
684 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
685 	u32 tmp;
686 	int i;
687 
688 	sdma_v6_0_gfx_stop(adev);
689 
690 	for (i = 0; i < adev->sdma.num_instances; i++) {
691 		tmp = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_FREEZE));
692 		tmp |= SDMA0_FREEZE__FREEZE_MASK;
693 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_FREEZE), tmp);
694 		tmp = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_F32_CNTL));
695 		tmp |= SDMA0_F32_CNTL__HALT_MASK;
696 		tmp |= SDMA0_F32_CNTL__TH1_RESET_MASK;
697 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_F32_CNTL), tmp);
698 
699 		WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_PREEMPT), 0);
700 
701 		udelay(100);
702 
703 		tmp = GRBM_SOFT_RESET__SOFT_RESET_SDMA0_MASK << i;
704 		WREG32_SOC15(GC, 0, regGRBM_SOFT_RESET, tmp);
705 		tmp = RREG32_SOC15(GC, 0, regGRBM_SOFT_RESET);
706 
707 		udelay(100);
708 
709 		WREG32_SOC15(GC, 0, regGRBM_SOFT_RESET, 0);
710 		tmp = RREG32_SOC15(GC, 0, regGRBM_SOFT_RESET);
711 
712 		udelay(100);
713 	}
714 
715 	return sdma_v6_0_start(adev);
716 }
717 
718 static bool sdma_v6_0_check_soft_reset(void *handle)
719 {
720 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
721 	struct amdgpu_ring *ring;
722 	int i, r;
723 	long tmo = msecs_to_jiffies(1000);
724 
725 	for (i = 0; i < adev->sdma.num_instances; i++) {
726 		ring = &adev->sdma.instance[i].ring;
727 		r = amdgpu_ring_test_ib(ring, tmo);
728 		if (r)
729 			return true;
730 	}
731 
732 	return false;
733 }
734 
735 /**
736  * sdma_v6_0_start - setup and start the async dma engines
737  *
738  * @adev: amdgpu_device pointer
739  *
740  * Set up the DMA engines and enable them.
741  * Returns 0 for success, error for failure.
742  */
743 static int sdma_v6_0_start(struct amdgpu_device *adev)
744 {
745 	int r = 0;
746 
747 	if (amdgpu_sriov_vf(adev)) {
748 		sdma_v6_0_enable(adev, false);
749 
750 		/* set RB registers */
751 		r = sdma_v6_0_gfx_resume(adev);
752 		return r;
753 	}
754 
755 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
756 		r = sdma_v6_0_load_microcode(adev);
757 		if (r)
758 			return r;
759 
760 		/* The value of regSDMA_F32_CNTL is invalid the moment after loading fw */
761 		if (amdgpu_emu_mode == 1)
762 			msleep(1000);
763 	}
764 
765 	/* unhalt the MEs */
766 	sdma_v6_0_enable(adev, true);
767 	/* enable sdma ring preemption */
768 	sdma_v6_0_ctxempty_int_enable(adev, true);
769 
770 	/* start the gfx rings and rlc compute queues */
771 	r = sdma_v6_0_gfx_resume(adev);
772 	if (r)
773 		return r;
774 	r = sdma_v6_0_rlc_resume(adev);
775 
776 	return r;
777 }
778 
779 static int sdma_v6_0_mqd_init(struct amdgpu_device *adev, void *mqd,
780 			      struct amdgpu_mqd_prop *prop)
781 {
782 	struct v11_sdma_mqd *m = mqd;
783 	uint64_t wb_gpu_addr;
784 
785 	m->sdmax_rlcx_rb_cntl =
786 		order_base_2(prop->queue_size / 4) << SDMA0_QUEUE0_RB_CNTL__RB_SIZE__SHIFT |
787 		1 << SDMA0_QUEUE0_RB_CNTL__RPTR_WRITEBACK_ENABLE__SHIFT |
788 		4 << SDMA0_QUEUE0_RB_CNTL__RPTR_WRITEBACK_TIMER__SHIFT |
789 		1 << SDMA0_QUEUE0_RB_CNTL__F32_WPTR_POLL_ENABLE__SHIFT;
790 
791 	m->sdmax_rlcx_rb_base = lower_32_bits(prop->hqd_base_gpu_addr >> 8);
792 	m->sdmax_rlcx_rb_base_hi = upper_32_bits(prop->hqd_base_gpu_addr >> 8);
793 
794 	wb_gpu_addr = prop->wptr_gpu_addr;
795 	m->sdmax_rlcx_rb_wptr_poll_addr_lo = lower_32_bits(wb_gpu_addr);
796 	m->sdmax_rlcx_rb_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr);
797 
798 	wb_gpu_addr = prop->rptr_gpu_addr;
799 	m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits(wb_gpu_addr);
800 	m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits(wb_gpu_addr);
801 
802 	m->sdmax_rlcx_ib_cntl = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, 0,
803 							regSDMA0_QUEUE0_IB_CNTL));
804 
805 	m->sdmax_rlcx_doorbell_offset =
806 		prop->doorbell_index << SDMA0_QUEUE0_DOORBELL_OFFSET__OFFSET__SHIFT;
807 
808 	m->sdmax_rlcx_doorbell = REG_SET_FIELD(0, SDMA0_QUEUE0_DOORBELL, ENABLE, 1);
809 
810 	m->sdmax_rlcx_skip_cntl = 0;
811 	m->sdmax_rlcx_context_status = 0;
812 	m->sdmax_rlcx_doorbell_log = 0;
813 
814 	m->sdmax_rlcx_rb_aql_cntl = regSDMA0_QUEUE0_RB_AQL_CNTL_DEFAULT;
815 	m->sdmax_rlcx_dummy_reg = regSDMA0_QUEUE0_DUMMY_REG_DEFAULT;
816 
817 	return 0;
818 }
819 
820 static void sdma_v6_0_set_mqd_funcs(struct amdgpu_device *adev)
821 {
822 	adev->mqds[AMDGPU_HW_IP_DMA].mqd_size = sizeof(struct v11_sdma_mqd);
823 	adev->mqds[AMDGPU_HW_IP_DMA].init_mqd = sdma_v6_0_mqd_init;
824 }
825 
826 /**
827  * sdma_v6_0_ring_test_ring - simple async dma engine test
828  *
829  * @ring: amdgpu_ring structure holding ring information
830  *
831  * Test the DMA engine by writing using it to write an
832  * value to memory.
833  * Returns 0 for success, error for failure.
834  */
835 static int sdma_v6_0_ring_test_ring(struct amdgpu_ring *ring)
836 {
837 	struct amdgpu_device *adev = ring->adev;
838 	unsigned i;
839 	unsigned index;
840 	int r;
841 	u32 tmp;
842 	u64 gpu_addr;
843 	volatile uint32_t *cpu_ptr = NULL;
844 
845 	tmp = 0xCAFEDEAD;
846 
847 	if (ring->is_mes_queue) {
848 		uint32_t offset = 0;
849 		offset = amdgpu_mes_ctx_get_offs(ring,
850 					 AMDGPU_MES_CTX_PADDING_OFFS);
851 		gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
852 		cpu_ptr = amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
853 		*cpu_ptr = tmp;
854 	} else {
855 		r = amdgpu_device_wb_get(adev, &index);
856 		if (r) {
857 			dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
858 			return r;
859 		}
860 
861 		gpu_addr = adev->wb.gpu_addr + (index * 4);
862 		adev->wb.wb[index] = cpu_to_le32(tmp);
863 	}
864 
865 	r = amdgpu_ring_alloc(ring, 5);
866 	if (r) {
867 		DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
868 		amdgpu_device_wb_free(adev, index);
869 		return r;
870 	}
871 
872 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_WRITE) |
873 			  SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
874 	amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
875 	amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
876 	amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0));
877 	amdgpu_ring_write(ring, 0xDEADBEEF);
878 	amdgpu_ring_commit(ring);
879 
880 	for (i = 0; i < adev->usec_timeout; i++) {
881 		if (ring->is_mes_queue)
882 			tmp = le32_to_cpu(*cpu_ptr);
883 		else
884 			tmp = le32_to_cpu(adev->wb.wb[index]);
885 		if (tmp == 0xDEADBEEF)
886 			break;
887 		if (amdgpu_emu_mode == 1)
888 			msleep(1);
889 		else
890 			udelay(1);
891 	}
892 
893 	if (i >= adev->usec_timeout)
894 		r = -ETIMEDOUT;
895 
896 	if (!ring->is_mes_queue)
897 		amdgpu_device_wb_free(adev, index);
898 
899 	return r;
900 }
901 
902 /*
903  * sdma_v6_0_ring_test_ib - test an IB on the DMA engine
904  *
905  * @ring: amdgpu_ring structure holding ring information
906  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
907  *
908  * Test a simple IB in the DMA ring.
909  * Returns 0 on success, error on failure.
910  */
911 static int sdma_v6_0_ring_test_ib(struct amdgpu_ring *ring, long timeout)
912 {
913 	struct amdgpu_device *adev = ring->adev;
914 	struct amdgpu_ib ib;
915 	struct dma_fence *f = NULL;
916 	unsigned index;
917 	long r;
918 	u32 tmp = 0;
919 	u64 gpu_addr;
920 	volatile uint32_t *cpu_ptr = NULL;
921 
922 	tmp = 0xCAFEDEAD;
923 	memset(&ib, 0, sizeof(ib));
924 
925 	if (ring->is_mes_queue) {
926 		uint32_t offset = 0;
927 		offset = amdgpu_mes_ctx_get_offs(ring, AMDGPU_MES_CTX_IB_OFFS);
928 		ib.gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
929 		ib.ptr = (void *)amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
930 
931 		offset = amdgpu_mes_ctx_get_offs(ring,
932 					 AMDGPU_MES_CTX_PADDING_OFFS);
933 		gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
934 		cpu_ptr = amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
935 		*cpu_ptr = tmp;
936 	} else {
937 		r = amdgpu_device_wb_get(adev, &index);
938 		if (r) {
939 			dev_err(adev->dev, "(%ld) failed to allocate wb slot\n", r);
940 			return r;
941 		}
942 
943 		gpu_addr = adev->wb.gpu_addr + (index * 4);
944 		adev->wb.wb[index] = cpu_to_le32(tmp);
945 
946 		r = amdgpu_ib_get(adev, NULL, 256, AMDGPU_IB_POOL_DIRECT, &ib);
947 		if (r) {
948 			DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r);
949 			goto err0;
950 		}
951 	}
952 
953 	ib.ptr[0] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_WRITE) |
954 		SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
955 	ib.ptr[1] = lower_32_bits(gpu_addr);
956 	ib.ptr[2] = upper_32_bits(gpu_addr);
957 	ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0);
958 	ib.ptr[4] = 0xDEADBEEF;
959 	ib.ptr[5] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
960 	ib.ptr[6] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
961 	ib.ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
962 	ib.length_dw = 8;
963 
964 	r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
965 	if (r)
966 		goto err1;
967 
968 	r = dma_fence_wait_timeout(f, false, timeout);
969 	if (r == 0) {
970 		DRM_ERROR("amdgpu: IB test timed out\n");
971 		r = -ETIMEDOUT;
972 		goto err1;
973 	} else if (r < 0) {
974 		DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
975 		goto err1;
976 	}
977 
978 	if (ring->is_mes_queue)
979 		tmp = le32_to_cpu(*cpu_ptr);
980 	else
981 		tmp = le32_to_cpu(adev->wb.wb[index]);
982 
983 	if (tmp == 0xDEADBEEF)
984 		r = 0;
985 	else
986 		r = -EINVAL;
987 
988 err1:
989 	amdgpu_ib_free(adev, &ib, NULL);
990 	dma_fence_put(f);
991 err0:
992 	if (!ring->is_mes_queue)
993 		amdgpu_device_wb_free(adev, index);
994 	return r;
995 }
996 
997 
998 /**
999  * sdma_v6_0_vm_copy_pte - update PTEs by copying them from the GART
1000  *
1001  * @ib: indirect buffer to fill with commands
1002  * @pe: addr of the page entry
1003  * @src: src addr to copy from
1004  * @count: number of page entries to update
1005  *
1006  * Update PTEs by copying them from the GART using sDMA.
1007  */
1008 static void sdma_v6_0_vm_copy_pte(struct amdgpu_ib *ib,
1009 				  uint64_t pe, uint64_t src,
1010 				  unsigned count)
1011 {
1012 	unsigned bytes = count * 8;
1013 
1014 	ib->ptr[ib->length_dw++] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_COPY) |
1015 		SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
1016 	ib->ptr[ib->length_dw++] = bytes - 1;
1017 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1018 	ib->ptr[ib->length_dw++] = lower_32_bits(src);
1019 	ib->ptr[ib->length_dw++] = upper_32_bits(src);
1020 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1021 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1022 
1023 }
1024 
1025 /**
1026  * sdma_v6_0_vm_write_pte - update PTEs by writing them manually
1027  *
1028  * @ib: indirect buffer to fill with commands
1029  * @pe: addr of the page entry
1030  * @value: dst addr to write into pe
1031  * @count: number of page entries to update
1032  * @incr: increase next addr by incr bytes
1033  *
1034  * Update PTEs by writing them manually using sDMA.
1035  */
1036 static void sdma_v6_0_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
1037 				   uint64_t value, unsigned count,
1038 				   uint32_t incr)
1039 {
1040 	unsigned ndw = count * 2;
1041 
1042 	ib->ptr[ib->length_dw++] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_WRITE) |
1043 		SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1044 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1045 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1046 	ib->ptr[ib->length_dw++] = ndw - 1;
1047 	for (; ndw > 0; ndw -= 2) {
1048 		ib->ptr[ib->length_dw++] = lower_32_bits(value);
1049 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
1050 		value += incr;
1051 	}
1052 }
1053 
1054 /**
1055  * sdma_v6_0_vm_set_pte_pde - update the page tables using sDMA
1056  *
1057  * @ib: indirect buffer to fill with commands
1058  * @pe: addr of the page entry
1059  * @addr: dst addr to write into pe
1060  * @count: number of page entries to update
1061  * @incr: increase next addr by incr bytes
1062  * @flags: access flags
1063  *
1064  * Update the page tables using sDMA.
1065  */
1066 static void sdma_v6_0_vm_set_pte_pde(struct amdgpu_ib *ib,
1067 				     uint64_t pe,
1068 				     uint64_t addr, unsigned count,
1069 				     uint32_t incr, uint64_t flags)
1070 {
1071 	/* for physically contiguous pages (vram) */
1072 	ib->ptr[ib->length_dw++] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_PTEPDE);
1073 	ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
1074 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1075 	ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
1076 	ib->ptr[ib->length_dw++] = upper_32_bits(flags);
1077 	ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
1078 	ib->ptr[ib->length_dw++] = upper_32_bits(addr);
1079 	ib->ptr[ib->length_dw++] = incr; /* increment size */
1080 	ib->ptr[ib->length_dw++] = 0;
1081 	ib->ptr[ib->length_dw++] = count - 1; /* number of entries */
1082 }
1083 
1084 /*
1085  * sdma_v6_0_ring_pad_ib - pad the IB
1086  * @ib: indirect buffer to fill with padding
1087  * @ring: amdgpu ring pointer
1088  *
1089  * Pad the IB with NOPs to a boundary multiple of 8.
1090  */
1091 static void sdma_v6_0_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
1092 {
1093 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
1094 	u32 pad_count;
1095 	int i;
1096 
1097 	pad_count = (-ib->length_dw) & 0x7;
1098 	for (i = 0; i < pad_count; i++)
1099 		if (sdma && sdma->burst_nop && (i == 0))
1100 			ib->ptr[ib->length_dw++] =
1101 				SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_NOP) |
1102 				SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
1103 		else
1104 			ib->ptr[ib->length_dw++] =
1105 				SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_NOP);
1106 }
1107 
1108 /**
1109  * sdma_v6_0_ring_emit_pipeline_sync - sync the pipeline
1110  *
1111  * @ring: amdgpu_ring pointer
1112  *
1113  * Make sure all previous operations are completed (CIK).
1114  */
1115 static void sdma_v6_0_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
1116 {
1117 	uint32_t seq = ring->fence_drv.sync_seq;
1118 	uint64_t addr = ring->fence_drv.gpu_addr;
1119 
1120 	/* wait for idle */
1121 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1122 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1123 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
1124 			  SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1));
1125 	amdgpu_ring_write(ring, addr & 0xfffffffc);
1126 	amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
1127 	amdgpu_ring_write(ring, seq); /* reference */
1128 	amdgpu_ring_write(ring, 0xffffffff); /* mask */
1129 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1130 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */
1131 }
1132 
1133 /*
1134  * sdma_v6_0_ring_emit_vm_flush - vm flush using sDMA
1135  *
1136  * @ring: amdgpu_ring pointer
1137  * @vmid: vmid number to use
1138  * @pd_addr: address
1139  *
1140  * Update the page table base and flush the VM TLB
1141  * using sDMA.
1142  */
1143 static void sdma_v6_0_ring_emit_vm_flush(struct amdgpu_ring *ring,
1144 					 unsigned vmid, uint64_t pd_addr)
1145 {
1146 	struct amdgpu_vmhub *hub = &ring->adev->vmhub[ring->vm_hub];
1147 	uint32_t req = hub->vmhub_funcs->get_invalidate_req(vmid, 0);
1148 
1149 	/* Update the PD address for this VMID. */
1150 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_lo32 +
1151 			      (hub->ctx_addr_distance * vmid),
1152 			      lower_32_bits(pd_addr));
1153 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_hi32 +
1154 			      (hub->ctx_addr_distance * vmid),
1155 			      upper_32_bits(pd_addr));
1156 
1157 	/* Trigger invalidation. */
1158 	amdgpu_ring_write(ring,
1159 			  SDMA_PKT_VM_INVALIDATION_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1160 			  SDMA_PKT_VM_INVALIDATION_HEADER_SUB_OP(SDMA_SUBOP_VM_INVALIDATION) |
1161 			  SDMA_PKT_VM_INVALIDATION_HEADER_GFX_ENG_ID(ring->vm_inv_eng) |
1162 			  SDMA_PKT_VM_INVALIDATION_HEADER_MM_ENG_ID(0x1f));
1163 	amdgpu_ring_write(ring, req);
1164 	amdgpu_ring_write(ring, 0xFFFFFFFF);
1165 	amdgpu_ring_write(ring,
1166 			  SDMA_PKT_VM_INVALIDATION_ADDRESSRANGEHI_INVALIDATEACK(1 << vmid) |
1167 			  SDMA_PKT_VM_INVALIDATION_ADDRESSRANGEHI_ADDRESSRANGEHI(0x1F));
1168 }
1169 
1170 static void sdma_v6_0_ring_emit_wreg(struct amdgpu_ring *ring,
1171 				     uint32_t reg, uint32_t val)
1172 {
1173 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_SRBM_WRITE) |
1174 			  SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
1175 	amdgpu_ring_write(ring, reg);
1176 	amdgpu_ring_write(ring, val);
1177 }
1178 
1179 static void sdma_v6_0_ring_emit_reg_wait(struct amdgpu_ring *ring, uint32_t reg,
1180 					 uint32_t val, uint32_t mask)
1181 {
1182 	amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1183 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1184 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* equal */
1185 	amdgpu_ring_write(ring, reg << 2);
1186 	amdgpu_ring_write(ring, 0);
1187 	amdgpu_ring_write(ring, val); /* reference */
1188 	amdgpu_ring_write(ring, mask); /* mask */
1189 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1190 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10));
1191 }
1192 
1193 static void sdma_v6_0_ring_emit_reg_write_reg_wait(struct amdgpu_ring *ring,
1194 						   uint32_t reg0, uint32_t reg1,
1195 						   uint32_t ref, uint32_t mask)
1196 {
1197 	amdgpu_ring_emit_wreg(ring, reg0, ref);
1198 	/* wait for a cycle to reset vm_inv_eng*_ack */
1199 	amdgpu_ring_emit_reg_wait(ring, reg0, 0, 0);
1200 	amdgpu_ring_emit_reg_wait(ring, reg1, mask, mask);
1201 }
1202 
1203 static struct amdgpu_sdma_ras sdma_v6_0_3_ras = {
1204 	.ras_block = {
1205 		.ras_late_init = amdgpu_ras_block_late_init,
1206 	},
1207 };
1208 
1209 static void sdma_v6_0_set_ras_funcs(struct amdgpu_device *adev)
1210 {
1211 	switch (amdgpu_ip_version(adev, SDMA0_HWIP, 0)) {
1212 	case IP_VERSION(6, 0, 3):
1213 		adev->sdma.ras = &sdma_v6_0_3_ras;
1214 		break;
1215 	default:
1216 		break;
1217 	}
1218 }
1219 
1220 static int sdma_v6_0_early_init(void *handle)
1221 {
1222 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1223 	int r;
1224 
1225 	r = amdgpu_sdma_init_microcode(adev, 0, true);
1226 	if (r)
1227 		return r;
1228 
1229 	sdma_v6_0_set_ring_funcs(adev);
1230 	sdma_v6_0_set_buffer_funcs(adev);
1231 	sdma_v6_0_set_vm_pte_funcs(adev);
1232 	sdma_v6_0_set_irq_funcs(adev);
1233 	sdma_v6_0_set_mqd_funcs(adev);
1234 	sdma_v6_0_set_ras_funcs(adev);
1235 
1236 	return 0;
1237 }
1238 
1239 static int sdma_v6_0_sw_init(void *handle)
1240 {
1241 	struct amdgpu_ring *ring;
1242 	int r, i;
1243 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1244 
1245 	/* SDMA trap event */
1246 	r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_GFX,
1247 			      GFX_11_0_0__SRCID__SDMA_TRAP,
1248 			      &adev->sdma.trap_irq);
1249 	if (r)
1250 		return r;
1251 
1252 	for (i = 0; i < adev->sdma.num_instances; i++) {
1253 		ring = &adev->sdma.instance[i].ring;
1254 		ring->ring_obj = NULL;
1255 		ring->use_doorbell = true;
1256 		ring->me = i;
1257 
1258 		DRM_DEBUG("SDMA %d use_doorbell being set to: [%s]\n", i,
1259 				ring->use_doorbell?"true":"false");
1260 
1261 		ring->doorbell_index =
1262 			(adev->doorbell_index.sdma_engine[i] << 1); // get DWORD offset
1263 
1264 		ring->vm_hub = AMDGPU_GFXHUB(0);
1265 		sprintf(ring->name, "sdma%d", i);
1266 		r = amdgpu_ring_init(adev, ring, 1024,
1267 				     &adev->sdma.trap_irq,
1268 				     AMDGPU_SDMA_IRQ_INSTANCE0 + i,
1269 				     AMDGPU_RING_PRIO_DEFAULT, NULL);
1270 		if (r)
1271 			return r;
1272 	}
1273 
1274 	if (amdgpu_sdma_ras_sw_init(adev)) {
1275 		dev_err(adev->dev, "Failed to initialize sdma ras block!\n");
1276 		return -EINVAL;
1277 	}
1278 
1279 	return r;
1280 }
1281 
1282 static int sdma_v6_0_sw_fini(void *handle)
1283 {
1284 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1285 	int i;
1286 
1287 	for (i = 0; i < adev->sdma.num_instances; i++)
1288 		amdgpu_ring_fini(&adev->sdma.instance[i].ring);
1289 
1290 	amdgpu_sdma_destroy_inst_ctx(adev, true);
1291 
1292 	return 0;
1293 }
1294 
1295 static int sdma_v6_0_hw_init(void *handle)
1296 {
1297 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1298 
1299 	return sdma_v6_0_start(adev);
1300 }
1301 
1302 static int sdma_v6_0_hw_fini(void *handle)
1303 {
1304 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1305 
1306 	if (amdgpu_sriov_vf(adev))
1307 		return 0;
1308 
1309 	sdma_v6_0_ctxempty_int_enable(adev, false);
1310 	sdma_v6_0_enable(adev, false);
1311 
1312 	return 0;
1313 }
1314 
1315 static int sdma_v6_0_suspend(void *handle)
1316 {
1317 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1318 
1319 	return sdma_v6_0_hw_fini(adev);
1320 }
1321 
1322 static int sdma_v6_0_resume(void *handle)
1323 {
1324 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1325 
1326 	return sdma_v6_0_hw_init(adev);
1327 }
1328 
1329 static bool sdma_v6_0_is_idle(void *handle)
1330 {
1331 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1332 	u32 i;
1333 
1334 	for (i = 0; i < adev->sdma.num_instances; i++) {
1335 		u32 tmp = RREG32(sdma_v6_0_get_reg_offset(adev, i, regSDMA0_STATUS_REG));
1336 
1337 		if (!(tmp & SDMA0_STATUS_REG__IDLE_MASK))
1338 			return false;
1339 	}
1340 
1341 	return true;
1342 }
1343 
1344 static int sdma_v6_0_wait_for_idle(void *handle)
1345 {
1346 	unsigned i;
1347 	u32 sdma0, sdma1;
1348 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1349 
1350 	for (i = 0; i < adev->usec_timeout; i++) {
1351 		sdma0 = RREG32(sdma_v6_0_get_reg_offset(adev, 0, regSDMA0_STATUS_REG));
1352 		sdma1 = RREG32(sdma_v6_0_get_reg_offset(adev, 1, regSDMA0_STATUS_REG));
1353 
1354 		if (sdma0 & sdma1 & SDMA0_STATUS_REG__IDLE_MASK)
1355 			return 0;
1356 		udelay(1);
1357 	}
1358 	return -ETIMEDOUT;
1359 }
1360 
1361 static int sdma_v6_0_ring_preempt_ib(struct amdgpu_ring *ring)
1362 {
1363 	int i, r = 0;
1364 	struct amdgpu_device *adev = ring->adev;
1365 	u32 index = 0;
1366 	u64 sdma_gfx_preempt;
1367 
1368 	amdgpu_sdma_get_index_from_ring(ring, &index);
1369 	sdma_gfx_preempt =
1370 		sdma_v6_0_get_reg_offset(adev, index, regSDMA0_QUEUE0_PREEMPT);
1371 
1372 	/* assert preemption condition */
1373 	amdgpu_ring_set_preempt_cond_exec(ring, false);
1374 
1375 	/* emit the trailing fence */
1376 	ring->trail_seq += 1;
1377 	amdgpu_ring_alloc(ring, 10);
1378 	sdma_v6_0_ring_emit_fence(ring, ring->trail_fence_gpu_addr,
1379 				  ring->trail_seq, 0);
1380 	amdgpu_ring_commit(ring);
1381 
1382 	/* assert IB preemption */
1383 	WREG32(sdma_gfx_preempt, 1);
1384 
1385 	/* poll the trailing fence */
1386 	for (i = 0; i < adev->usec_timeout; i++) {
1387 		if (ring->trail_seq ==
1388 		    le32_to_cpu(*(ring->trail_fence_cpu_addr)))
1389 			break;
1390 		udelay(1);
1391 	}
1392 
1393 	if (i >= adev->usec_timeout) {
1394 		r = -EINVAL;
1395 		DRM_ERROR("ring %d failed to be preempted\n", ring->idx);
1396 	}
1397 
1398 	/* deassert IB preemption */
1399 	WREG32(sdma_gfx_preempt, 0);
1400 
1401 	/* deassert the preemption condition */
1402 	amdgpu_ring_set_preempt_cond_exec(ring, true);
1403 	return r;
1404 }
1405 
1406 static int sdma_v6_0_set_trap_irq_state(struct amdgpu_device *adev,
1407 					struct amdgpu_irq_src *source,
1408 					unsigned type,
1409 					enum amdgpu_interrupt_state state)
1410 {
1411 	u32 sdma_cntl;
1412 
1413 	u32 reg_offset = sdma_v6_0_get_reg_offset(adev, type, regSDMA0_CNTL);
1414 
1415 	if (!amdgpu_sriov_vf(adev)) {
1416 		sdma_cntl = RREG32(reg_offset);
1417 		sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE,
1418 				state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0);
1419 		WREG32(reg_offset, sdma_cntl);
1420 	}
1421 
1422 	return 0;
1423 }
1424 
1425 static int sdma_v6_0_process_trap_irq(struct amdgpu_device *adev,
1426 				      struct amdgpu_irq_src *source,
1427 				      struct amdgpu_iv_entry *entry)
1428 {
1429 	int instances, queue;
1430 	uint32_t mes_queue_id = entry->src_data[0];
1431 
1432 	DRM_DEBUG("IH: SDMA trap\n");
1433 
1434 	if (adev->enable_mes && (mes_queue_id & AMDGPU_FENCE_MES_QUEUE_FLAG)) {
1435 		struct amdgpu_mes_queue *queue;
1436 
1437 		mes_queue_id &= AMDGPU_FENCE_MES_QUEUE_ID_MASK;
1438 
1439 		spin_lock(&adev->mes.queue_id_lock);
1440 		queue = idr_find(&adev->mes.queue_id_idr, mes_queue_id);
1441 		if (queue) {
1442 			DRM_DEBUG("process smda queue id = %d\n", mes_queue_id);
1443 			amdgpu_fence_process(queue->ring);
1444 		}
1445 		spin_unlock(&adev->mes.queue_id_lock);
1446 		return 0;
1447 	}
1448 
1449 	queue = entry->ring_id & 0xf;
1450 	instances = (entry->ring_id & 0xf0) >> 4;
1451 	if (instances > 1) {
1452 		DRM_ERROR("IH: wrong ring_ID detected, as wrong sdma instance\n");
1453 		return -EINVAL;
1454 	}
1455 
1456 	switch (entry->client_id) {
1457 	case SOC21_IH_CLIENTID_GFX:
1458 		switch (queue) {
1459 		case 0:
1460 			amdgpu_fence_process(&adev->sdma.instance[instances].ring);
1461 			break;
1462 		default:
1463 			break;
1464 		}
1465 		break;
1466 	}
1467 	return 0;
1468 }
1469 
1470 static int sdma_v6_0_process_illegal_inst_irq(struct amdgpu_device *adev,
1471 					      struct amdgpu_irq_src *source,
1472 					      struct amdgpu_iv_entry *entry)
1473 {
1474 	return 0;
1475 }
1476 
1477 static int sdma_v6_0_set_clockgating_state(void *handle,
1478 					   enum amd_clockgating_state state)
1479 {
1480 	return 0;
1481 }
1482 
1483 static int sdma_v6_0_set_powergating_state(void *handle,
1484 					  enum amd_powergating_state state)
1485 {
1486 	return 0;
1487 }
1488 
1489 static void sdma_v6_0_get_clockgating_state(void *handle, u64 *flags)
1490 {
1491 }
1492 
1493 const struct amd_ip_funcs sdma_v6_0_ip_funcs = {
1494 	.name = "sdma_v6_0",
1495 	.early_init = sdma_v6_0_early_init,
1496 	.late_init = NULL,
1497 	.sw_init = sdma_v6_0_sw_init,
1498 	.sw_fini = sdma_v6_0_sw_fini,
1499 	.hw_init = sdma_v6_0_hw_init,
1500 	.hw_fini = sdma_v6_0_hw_fini,
1501 	.suspend = sdma_v6_0_suspend,
1502 	.resume = sdma_v6_0_resume,
1503 	.is_idle = sdma_v6_0_is_idle,
1504 	.wait_for_idle = sdma_v6_0_wait_for_idle,
1505 	.soft_reset = sdma_v6_0_soft_reset,
1506 	.check_soft_reset = sdma_v6_0_check_soft_reset,
1507 	.set_clockgating_state = sdma_v6_0_set_clockgating_state,
1508 	.set_powergating_state = sdma_v6_0_set_powergating_state,
1509 	.get_clockgating_state = sdma_v6_0_get_clockgating_state,
1510 };
1511 
1512 static const struct amdgpu_ring_funcs sdma_v6_0_ring_funcs = {
1513 	.type = AMDGPU_RING_TYPE_SDMA,
1514 	.align_mask = 0xf,
1515 	.nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP),
1516 	.support_64bit_ptrs = true,
1517 	.secure_submission_supported = true,
1518 	.get_rptr = sdma_v6_0_ring_get_rptr,
1519 	.get_wptr = sdma_v6_0_ring_get_wptr,
1520 	.set_wptr = sdma_v6_0_ring_set_wptr,
1521 	.emit_frame_size =
1522 		5 + /* sdma_v6_0_ring_init_cond_exec */
1523 		6 + /* sdma_v6_0_ring_emit_hdp_flush */
1524 		6 + /* sdma_v6_0_ring_emit_pipeline_sync */
1525 		/* sdma_v6_0_ring_emit_vm_flush */
1526 		SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 +
1527 		SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6 +
1528 		10 + 10 + 10, /* sdma_v6_0_ring_emit_fence x3 for user fence, vm fence */
1529 	.emit_ib_size = 5 + 7 + 6, /* sdma_v6_0_ring_emit_ib */
1530 	.emit_ib = sdma_v6_0_ring_emit_ib,
1531 	.emit_mem_sync = sdma_v6_0_ring_emit_mem_sync,
1532 	.emit_fence = sdma_v6_0_ring_emit_fence,
1533 	.emit_pipeline_sync = sdma_v6_0_ring_emit_pipeline_sync,
1534 	.emit_vm_flush = sdma_v6_0_ring_emit_vm_flush,
1535 	.emit_hdp_flush = sdma_v6_0_ring_emit_hdp_flush,
1536 	.test_ring = sdma_v6_0_ring_test_ring,
1537 	.test_ib = sdma_v6_0_ring_test_ib,
1538 	.insert_nop = sdma_v6_0_ring_insert_nop,
1539 	.pad_ib = sdma_v6_0_ring_pad_ib,
1540 	.emit_wreg = sdma_v6_0_ring_emit_wreg,
1541 	.emit_reg_wait = sdma_v6_0_ring_emit_reg_wait,
1542 	.emit_reg_write_reg_wait = sdma_v6_0_ring_emit_reg_write_reg_wait,
1543 	.init_cond_exec = sdma_v6_0_ring_init_cond_exec,
1544 	.patch_cond_exec = sdma_v6_0_ring_patch_cond_exec,
1545 	.preempt_ib = sdma_v6_0_ring_preempt_ib,
1546 };
1547 
1548 static void sdma_v6_0_set_ring_funcs(struct amdgpu_device *adev)
1549 {
1550 	int i;
1551 
1552 	for (i = 0; i < adev->sdma.num_instances; i++) {
1553 		adev->sdma.instance[i].ring.funcs = &sdma_v6_0_ring_funcs;
1554 		adev->sdma.instance[i].ring.me = i;
1555 	}
1556 }
1557 
1558 static const struct amdgpu_irq_src_funcs sdma_v6_0_trap_irq_funcs = {
1559 	.set = sdma_v6_0_set_trap_irq_state,
1560 	.process = sdma_v6_0_process_trap_irq,
1561 };
1562 
1563 static const struct amdgpu_irq_src_funcs sdma_v6_0_illegal_inst_irq_funcs = {
1564 	.process = sdma_v6_0_process_illegal_inst_irq,
1565 };
1566 
1567 static void sdma_v6_0_set_irq_funcs(struct amdgpu_device *adev)
1568 {
1569 	adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_INSTANCE0 +
1570 					adev->sdma.num_instances;
1571 	adev->sdma.trap_irq.funcs = &sdma_v6_0_trap_irq_funcs;
1572 	adev->sdma.illegal_inst_irq.funcs = &sdma_v6_0_illegal_inst_irq_funcs;
1573 }
1574 
1575 /**
1576  * sdma_v6_0_emit_copy_buffer - copy buffer using the sDMA engine
1577  *
1578  * @ib: indirect buffer to fill with commands
1579  * @src_offset: src GPU address
1580  * @dst_offset: dst GPU address
1581  * @byte_count: number of bytes to xfer
1582  * @tmz: if a secure copy should be used
1583  *
1584  * Copy GPU buffers using the DMA engine.
1585  * Used by the amdgpu ttm implementation to move pages if
1586  * registered as the asic copy callback.
1587  */
1588 static void sdma_v6_0_emit_copy_buffer(struct amdgpu_ib *ib,
1589 				       uint64_t src_offset,
1590 				       uint64_t dst_offset,
1591 				       uint32_t byte_count,
1592 				       bool tmz)
1593 {
1594 	ib->ptr[ib->length_dw++] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_COPY) |
1595 		SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR) |
1596 		SDMA_PKT_COPY_LINEAR_HEADER_TMZ(tmz ? 1 : 0);
1597 	ib->ptr[ib->length_dw++] = byte_count - 1;
1598 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1599 	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1600 	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1601 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1602 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1603 }
1604 
1605 /**
1606  * sdma_v6_0_emit_fill_buffer - fill buffer using the sDMA engine
1607  *
1608  * @ib: indirect buffer to fill
1609  * @src_data: value to write to buffer
1610  * @dst_offset: dst GPU address
1611  * @byte_count: number of bytes to xfer
1612  *
1613  * Fill GPU buffers using the DMA engine.
1614  */
1615 static void sdma_v6_0_emit_fill_buffer(struct amdgpu_ib *ib,
1616 				       uint32_t src_data,
1617 				       uint64_t dst_offset,
1618 				       uint32_t byte_count)
1619 {
1620 	ib->ptr[ib->length_dw++] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_CONST_FILL);
1621 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1622 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1623 	ib->ptr[ib->length_dw++] = src_data;
1624 	ib->ptr[ib->length_dw++] = byte_count - 1;
1625 }
1626 
1627 static const struct amdgpu_buffer_funcs sdma_v6_0_buffer_funcs = {
1628 	.copy_max_bytes = 0x400000,
1629 	.copy_num_dw = 7,
1630 	.emit_copy_buffer = sdma_v6_0_emit_copy_buffer,
1631 
1632 	.fill_max_bytes = 0x400000,
1633 	.fill_num_dw = 5,
1634 	.emit_fill_buffer = sdma_v6_0_emit_fill_buffer,
1635 };
1636 
1637 static void sdma_v6_0_set_buffer_funcs(struct amdgpu_device *adev)
1638 {
1639 	adev->mman.buffer_funcs = &sdma_v6_0_buffer_funcs;
1640 	adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1641 }
1642 
1643 static const struct amdgpu_vm_pte_funcs sdma_v6_0_vm_pte_funcs = {
1644 	.copy_pte_num_dw = 7,
1645 	.copy_pte = sdma_v6_0_vm_copy_pte,
1646 	.write_pte = sdma_v6_0_vm_write_pte,
1647 	.set_pte_pde = sdma_v6_0_vm_set_pte_pde,
1648 };
1649 
1650 static void sdma_v6_0_set_vm_pte_funcs(struct amdgpu_device *adev)
1651 {
1652 	unsigned i;
1653 
1654 	adev->vm_manager.vm_pte_funcs = &sdma_v6_0_vm_pte_funcs;
1655 	for (i = 0; i < adev->sdma.num_instances; i++) {
1656 		adev->vm_manager.vm_pte_scheds[i] =
1657 			&adev->sdma.instance[i].ring.sched;
1658 	}
1659 	adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
1660 }
1661 
1662 const struct amdgpu_ip_block_version sdma_v6_0_ip_block = {
1663 	.type = AMD_IP_BLOCK_TYPE_SDMA,
1664 	.major = 6,
1665 	.minor = 0,
1666 	.rev = 0,
1667 	.funcs = &sdma_v6_0_ip_funcs,
1668 };
1669