xref: /linux/drivers/gpu/drm/amd/amdgpu/sdma_v2_4.c (revision b85d45947951d23cb22d90caecf4c1eb81342c96)
1 /*
2  * Copyright 2014 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  * Authors: Alex Deucher
23  */
24 #include <linux/firmware.h>
25 #include <drm/drmP.h>
26 #include "amdgpu.h"
27 #include "amdgpu_ucode.h"
28 #include "amdgpu_trace.h"
29 #include "vi.h"
30 #include "vid.h"
31 
32 #include "oss/oss_2_4_d.h"
33 #include "oss/oss_2_4_sh_mask.h"
34 
35 #include "gmc/gmc_8_1_d.h"
36 #include "gmc/gmc_8_1_sh_mask.h"
37 
38 #include "gca/gfx_8_0_d.h"
39 #include "gca/gfx_8_0_enum.h"
40 #include "gca/gfx_8_0_sh_mask.h"
41 
42 #include "bif/bif_5_0_d.h"
43 #include "bif/bif_5_0_sh_mask.h"
44 
45 #include "iceland_sdma_pkt_open.h"
46 
47 static void sdma_v2_4_set_ring_funcs(struct amdgpu_device *adev);
48 static void sdma_v2_4_set_buffer_funcs(struct amdgpu_device *adev);
49 static void sdma_v2_4_set_vm_pte_funcs(struct amdgpu_device *adev);
50 static void sdma_v2_4_set_irq_funcs(struct amdgpu_device *adev);
51 
52 MODULE_FIRMWARE("amdgpu/topaz_sdma.bin");
53 MODULE_FIRMWARE("amdgpu/topaz_sdma1.bin");
54 
55 static const u32 sdma_offsets[SDMA_MAX_INSTANCE] =
56 {
57 	SDMA0_REGISTER_OFFSET,
58 	SDMA1_REGISTER_OFFSET
59 };
60 
61 static const u32 golden_settings_iceland_a11[] =
62 {
63 	mmSDMA0_CHICKEN_BITS, 0xfc910007, 0x00810007,
64 	mmSDMA0_CLK_CTRL, 0xff000fff, 0x00000000,
65 	mmSDMA1_CHICKEN_BITS, 0xfc910007, 0x00810007,
66 	mmSDMA1_CLK_CTRL, 0xff000fff, 0x00000000,
67 };
68 
69 static const u32 iceland_mgcg_cgcg_init[] =
70 {
71 	mmSDMA0_CLK_CTRL, 0xff000ff0, 0x00000100,
72 	mmSDMA1_CLK_CTRL, 0xff000ff0, 0x00000100
73 };
74 
75 /*
76  * sDMA - System DMA
77  * Starting with CIK, the GPU has new asynchronous
78  * DMA engines.  These engines are used for compute
79  * and gfx.  There are two DMA engines (SDMA0, SDMA1)
80  * and each one supports 1 ring buffer used for gfx
81  * and 2 queues used for compute.
82  *
83  * The programming model is very similar to the CP
84  * (ring buffer, IBs, etc.), but sDMA has it's own
85  * packet format that is different from the PM4 format
86  * used by the CP. sDMA supports copying data, writing
87  * embedded data, solid fills, and a number of other
88  * things.  It also has support for tiling/detiling of
89  * buffers.
90  */
91 
92 static void sdma_v2_4_init_golden_registers(struct amdgpu_device *adev)
93 {
94 	switch (adev->asic_type) {
95 	case CHIP_TOPAZ:
96 		amdgpu_program_register_sequence(adev,
97 						 iceland_mgcg_cgcg_init,
98 						 (const u32)ARRAY_SIZE(iceland_mgcg_cgcg_init));
99 		amdgpu_program_register_sequence(adev,
100 						 golden_settings_iceland_a11,
101 						 (const u32)ARRAY_SIZE(golden_settings_iceland_a11));
102 		break;
103 	default:
104 		break;
105 	}
106 }
107 
108 /**
109  * sdma_v2_4_init_microcode - load ucode images from disk
110  *
111  * @adev: amdgpu_device pointer
112  *
113  * Use the firmware interface to load the ucode images into
114  * the driver (not loaded into hw).
115  * Returns 0 on success, error on failure.
116  */
117 static int sdma_v2_4_init_microcode(struct amdgpu_device *adev)
118 {
119 	const char *chip_name;
120 	char fw_name[30];
121 	int err, i;
122 	struct amdgpu_firmware_info *info = NULL;
123 	const struct common_firmware_header *header = NULL;
124 	const struct sdma_firmware_header_v1_0 *hdr;
125 
126 	DRM_DEBUG("\n");
127 
128 	switch (adev->asic_type) {
129 	case CHIP_TOPAZ:
130 		chip_name = "topaz";
131 		break;
132 	default: BUG();
133 	}
134 
135 	for (i = 0; i < SDMA_MAX_INSTANCE; i++) {
136 		if (i == 0)
137 			snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma.bin", chip_name);
138 		else
139 			snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma1.bin", chip_name);
140 		err = request_firmware(&adev->sdma[i].fw, fw_name, adev->dev);
141 		if (err)
142 			goto out;
143 		err = amdgpu_ucode_validate(adev->sdma[i].fw);
144 		if (err)
145 			goto out;
146 		hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma[i].fw->data;
147 		adev->sdma[i].fw_version = le32_to_cpu(hdr->header.ucode_version);
148 		adev->sdma[i].feature_version = le32_to_cpu(hdr->ucode_feature_version);
149 		if (adev->sdma[i].feature_version >= 20)
150 			adev->sdma[i].burst_nop = true;
151 
152 		if (adev->firmware.smu_load) {
153 			info = &adev->firmware.ucode[AMDGPU_UCODE_ID_SDMA0 + i];
154 			info->ucode_id = AMDGPU_UCODE_ID_SDMA0 + i;
155 			info->fw = adev->sdma[i].fw;
156 			header = (const struct common_firmware_header *)info->fw->data;
157 			adev->firmware.fw_size +=
158 				ALIGN(le32_to_cpu(header->ucode_size_bytes), PAGE_SIZE);
159 		}
160 	}
161 
162 out:
163 	if (err) {
164 		printk(KERN_ERR
165 		       "sdma_v2_4: Failed to load firmware \"%s\"\n",
166 		       fw_name);
167 		for (i = 0; i < SDMA_MAX_INSTANCE; i++) {
168 			release_firmware(adev->sdma[i].fw);
169 			adev->sdma[i].fw = NULL;
170 		}
171 	}
172 	return err;
173 }
174 
175 /**
176  * sdma_v2_4_ring_get_rptr - get the current read pointer
177  *
178  * @ring: amdgpu ring pointer
179  *
180  * Get the current rptr from the hardware (VI+).
181  */
182 static uint32_t sdma_v2_4_ring_get_rptr(struct amdgpu_ring *ring)
183 {
184 	u32 rptr;
185 
186 	/* XXX check if swapping is necessary on BE */
187 	rptr = ring->adev->wb.wb[ring->rptr_offs] >> 2;
188 
189 	return rptr;
190 }
191 
192 /**
193  * sdma_v2_4_ring_get_wptr - get the current write pointer
194  *
195  * @ring: amdgpu ring pointer
196  *
197  * Get the current wptr from the hardware (VI+).
198  */
199 static uint32_t sdma_v2_4_ring_get_wptr(struct amdgpu_ring *ring)
200 {
201 	struct amdgpu_device *adev = ring->adev;
202 	int me = (ring == &ring->adev->sdma[0].ring) ? 0 : 1;
203 	u32 wptr = RREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[me]) >> 2;
204 
205 	return wptr;
206 }
207 
208 /**
209  * sdma_v2_4_ring_set_wptr - commit the write pointer
210  *
211  * @ring: amdgpu ring pointer
212  *
213  * Write the wptr back to the hardware (VI+).
214  */
215 static void sdma_v2_4_ring_set_wptr(struct amdgpu_ring *ring)
216 {
217 	struct amdgpu_device *adev = ring->adev;
218 	int me = (ring == &ring->adev->sdma[0].ring) ? 0 : 1;
219 
220 	WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[me], ring->wptr << 2);
221 }
222 
223 static void sdma_v2_4_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
224 {
225 	struct amdgpu_sdma *sdma = amdgpu_get_sdma_instance(ring);
226 	int i;
227 
228 	for (i = 0; i < count; i++)
229 		if (sdma && sdma->burst_nop && (i == 0))
230 			amdgpu_ring_write(ring, ring->nop |
231 				SDMA_PKT_NOP_HEADER_COUNT(count - 1));
232 		else
233 			amdgpu_ring_write(ring, ring->nop);
234 }
235 
236 /**
237  * sdma_v2_4_ring_emit_ib - Schedule an IB on the DMA engine
238  *
239  * @ring: amdgpu ring pointer
240  * @ib: IB object to schedule
241  *
242  * Schedule an IB in the DMA ring (VI).
243  */
244 static void sdma_v2_4_ring_emit_ib(struct amdgpu_ring *ring,
245 				   struct amdgpu_ib *ib)
246 {
247 	u32 vmid = (ib->vm ? ib->vm->ids[ring->idx].id : 0) & 0xf;
248 	u32 next_rptr = ring->wptr + 5;
249 
250 	while ((next_rptr & 7) != 2)
251 		next_rptr++;
252 
253 	next_rptr += 6;
254 
255 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
256 			  SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
257 	amdgpu_ring_write(ring, lower_32_bits(ring->next_rptr_gpu_addr) & 0xfffffffc);
258 	amdgpu_ring_write(ring, upper_32_bits(ring->next_rptr_gpu_addr));
259 	amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(1));
260 	amdgpu_ring_write(ring, next_rptr);
261 
262 	/* IB packet must end on a 8 DW boundary */
263 	sdma_v2_4_ring_insert_nop(ring, (10 - (ring->wptr & 7)) % 8);
264 
265 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) |
266 			  SDMA_PKT_INDIRECT_HEADER_VMID(vmid));
267 	/* base must be 32 byte aligned */
268 	amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
269 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
270 	amdgpu_ring_write(ring, ib->length_dw);
271 	amdgpu_ring_write(ring, 0);
272 	amdgpu_ring_write(ring, 0);
273 
274 }
275 
276 /**
277  * sdma_v2_4_hdp_flush_ring_emit - emit an hdp flush on the DMA ring
278  *
279  * @ring: amdgpu ring pointer
280  *
281  * Emit an hdp flush packet on the requested DMA ring.
282  */
283 static void sdma_v2_4_ring_emit_hdp_flush(struct amdgpu_ring *ring)
284 {
285 	u32 ref_and_mask = 0;
286 
287 	if (ring == &ring->adev->sdma[0].ring)
288 		ref_and_mask = REG_SET_FIELD(ref_and_mask, GPU_HDP_FLUSH_DONE, SDMA0, 1);
289 	else
290 		ref_and_mask = REG_SET_FIELD(ref_and_mask, GPU_HDP_FLUSH_DONE, SDMA1, 1);
291 
292 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
293 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) |
294 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
295 	amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_DONE << 2);
296 	amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_REQ << 2);
297 	amdgpu_ring_write(ring, ref_and_mask); /* reference */
298 	amdgpu_ring_write(ring, ref_and_mask); /* mask */
299 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
300 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
301 }
302 
303 /**
304  * sdma_v2_4_ring_emit_fence - emit a fence on the DMA ring
305  *
306  * @ring: amdgpu ring pointer
307  * @fence: amdgpu fence object
308  *
309  * Add a DMA fence packet to the ring to write
310  * the fence seq number and DMA trap packet to generate
311  * an interrupt if needed (VI).
312  */
313 static void sdma_v2_4_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
314 				      unsigned flags)
315 {
316 	bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
317 	/* write the fence */
318 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE));
319 	amdgpu_ring_write(ring, lower_32_bits(addr));
320 	amdgpu_ring_write(ring, upper_32_bits(addr));
321 	amdgpu_ring_write(ring, lower_32_bits(seq));
322 
323 	/* optionally write high bits as well */
324 	if (write64bit) {
325 		addr += 4;
326 		amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE));
327 		amdgpu_ring_write(ring, lower_32_bits(addr));
328 		amdgpu_ring_write(ring, upper_32_bits(addr));
329 		amdgpu_ring_write(ring, upper_32_bits(seq));
330 	}
331 
332 	/* generate an interrupt */
333 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP));
334 	amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(0));
335 }
336 
337 /**
338  * sdma_v2_4_ring_emit_semaphore - emit a semaphore on the dma ring
339  *
340  * @ring: amdgpu_ring structure holding ring information
341  * @semaphore: amdgpu semaphore object
342  * @emit_wait: wait or signal semaphore
343  *
344  * Add a DMA semaphore packet to the ring wait on or signal
345  * other rings (VI).
346  */
347 static bool sdma_v2_4_ring_emit_semaphore(struct amdgpu_ring *ring,
348 					  struct amdgpu_semaphore *semaphore,
349 					  bool emit_wait)
350 {
351 	u64 addr = semaphore->gpu_addr;
352 	u32 sig = emit_wait ? 0 : 1;
353 
354 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SEM) |
355 			  SDMA_PKT_SEMAPHORE_HEADER_SIGNAL(sig));
356 	amdgpu_ring_write(ring, lower_32_bits(addr) & 0xfffffff8);
357 	amdgpu_ring_write(ring, upper_32_bits(addr));
358 
359 	return true;
360 }
361 
362 /**
363  * sdma_v2_4_gfx_stop - stop the gfx async dma engines
364  *
365  * @adev: amdgpu_device pointer
366  *
367  * Stop the gfx async dma ring buffers (VI).
368  */
369 static void sdma_v2_4_gfx_stop(struct amdgpu_device *adev)
370 {
371 	struct amdgpu_ring *sdma0 = &adev->sdma[0].ring;
372 	struct amdgpu_ring *sdma1 = &adev->sdma[1].ring;
373 	u32 rb_cntl, ib_cntl;
374 	int i;
375 
376 	if ((adev->mman.buffer_funcs_ring == sdma0) ||
377 	    (adev->mman.buffer_funcs_ring == sdma1))
378 		amdgpu_ttm_set_active_vram_size(adev, adev->mc.visible_vram_size);
379 
380 	for (i = 0; i < SDMA_MAX_INSTANCE; i++) {
381 		rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
382 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 0);
383 		WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
384 		ib_cntl = RREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i]);
385 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 0);
386 		WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
387 	}
388 	sdma0->ready = false;
389 	sdma1->ready = false;
390 }
391 
392 /**
393  * sdma_v2_4_rlc_stop - stop the compute async dma engines
394  *
395  * @adev: amdgpu_device pointer
396  *
397  * Stop the compute async dma queues (VI).
398  */
399 static void sdma_v2_4_rlc_stop(struct amdgpu_device *adev)
400 {
401 	/* XXX todo */
402 }
403 
404 /**
405  * sdma_v2_4_enable - stop the async dma engines
406  *
407  * @adev: amdgpu_device pointer
408  * @enable: enable/disable the DMA MEs.
409  *
410  * Halt or unhalt the async dma engines (VI).
411  */
412 static void sdma_v2_4_enable(struct amdgpu_device *adev, bool enable)
413 {
414 	u32 f32_cntl;
415 	int i;
416 
417 	if (enable == false) {
418 		sdma_v2_4_gfx_stop(adev);
419 		sdma_v2_4_rlc_stop(adev);
420 	}
421 
422 	for (i = 0; i < SDMA_MAX_INSTANCE; i++) {
423 		f32_cntl = RREG32(mmSDMA0_F32_CNTL + sdma_offsets[i]);
424 		if (enable)
425 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 0);
426 		else
427 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 1);
428 		WREG32(mmSDMA0_F32_CNTL + sdma_offsets[i], f32_cntl);
429 	}
430 }
431 
432 /**
433  * sdma_v2_4_gfx_resume - setup and start the async dma engines
434  *
435  * @adev: amdgpu_device pointer
436  *
437  * Set up the gfx DMA ring buffers and enable them (VI).
438  * Returns 0 for success, error for failure.
439  */
440 static int sdma_v2_4_gfx_resume(struct amdgpu_device *adev)
441 {
442 	struct amdgpu_ring *ring;
443 	u32 rb_cntl, ib_cntl;
444 	u32 rb_bufsz;
445 	u32 wb_offset;
446 	int i, j, r;
447 
448 	for (i = 0; i < SDMA_MAX_INSTANCE; i++) {
449 		ring = &adev->sdma[i].ring;
450 		wb_offset = (ring->rptr_offs * 4);
451 
452 		mutex_lock(&adev->srbm_mutex);
453 		for (j = 0; j < 16; j++) {
454 			vi_srbm_select(adev, 0, 0, 0, j);
455 			/* SDMA GFX */
456 			WREG32(mmSDMA0_GFX_VIRTUAL_ADDR + sdma_offsets[i], 0);
457 			WREG32(mmSDMA0_GFX_APE1_CNTL + sdma_offsets[i], 0);
458 		}
459 		vi_srbm_select(adev, 0, 0, 0, 0);
460 		mutex_unlock(&adev->srbm_mutex);
461 
462 		WREG32(mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i], 0);
463 
464 		/* Set ring buffer size in dwords */
465 		rb_bufsz = order_base_2(ring->ring_size / 4);
466 		rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
467 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz);
468 #ifdef __BIG_ENDIAN
469 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1);
470 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL,
471 					RPTR_WRITEBACK_SWAP_ENABLE, 1);
472 #endif
473 		WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
474 
475 		/* Initialize the ring buffer's read and write pointers */
476 		WREG32(mmSDMA0_GFX_RB_RPTR + sdma_offsets[i], 0);
477 		WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], 0);
478 
479 		/* set the wb address whether it's enabled or not */
480 		WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_HI + sdma_offsets[i],
481 		       upper_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFF);
482 		WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_LO + sdma_offsets[i],
483 		       lower_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC);
484 
485 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
486 
487 		WREG32(mmSDMA0_GFX_RB_BASE + sdma_offsets[i], ring->gpu_addr >> 8);
488 		WREG32(mmSDMA0_GFX_RB_BASE_HI + sdma_offsets[i], ring->gpu_addr >> 40);
489 
490 		ring->wptr = 0;
491 		WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], ring->wptr << 2);
492 
493 		/* enable DMA RB */
494 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1);
495 		WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
496 
497 		ib_cntl = RREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i]);
498 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 1);
499 #ifdef __BIG_ENDIAN
500 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1);
501 #endif
502 		/* enable DMA IBs */
503 		WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
504 
505 		ring->ready = true;
506 
507 		r = amdgpu_ring_test_ring(ring);
508 		if (r) {
509 			ring->ready = false;
510 			return r;
511 		}
512 
513 		if (adev->mman.buffer_funcs_ring == ring)
514 			amdgpu_ttm_set_active_vram_size(adev, adev->mc.real_vram_size);
515 	}
516 
517 	return 0;
518 }
519 
520 /**
521  * sdma_v2_4_rlc_resume - setup and start the async dma engines
522  *
523  * @adev: amdgpu_device pointer
524  *
525  * Set up the compute DMA queues and enable them (VI).
526  * Returns 0 for success, error for failure.
527  */
528 static int sdma_v2_4_rlc_resume(struct amdgpu_device *adev)
529 {
530 	/* XXX todo */
531 	return 0;
532 }
533 
534 /**
535  * sdma_v2_4_load_microcode - load the sDMA ME ucode
536  *
537  * @adev: amdgpu_device pointer
538  *
539  * Loads the sDMA0/1 ucode.
540  * Returns 0 for success, -EINVAL if the ucode is not available.
541  */
542 static int sdma_v2_4_load_microcode(struct amdgpu_device *adev)
543 {
544 	const struct sdma_firmware_header_v1_0 *hdr;
545 	const __le32 *fw_data;
546 	u32 fw_size;
547 	int i, j;
548 	bool smc_loads_fw = false; /* XXX fix me */
549 
550 	if (!adev->sdma[0].fw || !adev->sdma[1].fw)
551 		return -EINVAL;
552 
553 	/* halt the MEs */
554 	sdma_v2_4_enable(adev, false);
555 
556 	if (smc_loads_fw) {
557 		/* XXX query SMC for fw load complete */
558 	} else {
559 		for (i = 0; i < SDMA_MAX_INSTANCE; i++) {
560 			hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma[i].fw->data;
561 			amdgpu_ucode_print_sdma_hdr(&hdr->header);
562 			fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
563 			fw_data = (const __le32 *)
564 				(adev->sdma[i].fw->data +
565 				 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
566 			WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], 0);
567 			for (j = 0; j < fw_size; j++)
568 				WREG32(mmSDMA0_UCODE_DATA + sdma_offsets[i], le32_to_cpup(fw_data++));
569 			WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], adev->sdma[i].fw_version);
570 		}
571 	}
572 
573 	return 0;
574 }
575 
576 /**
577  * sdma_v2_4_start - setup and start the async dma engines
578  *
579  * @adev: amdgpu_device pointer
580  *
581  * Set up the DMA engines and enable them (VI).
582  * Returns 0 for success, error for failure.
583  */
584 static int sdma_v2_4_start(struct amdgpu_device *adev)
585 {
586 	int r;
587 
588 	if (!adev->firmware.smu_load) {
589 		r = sdma_v2_4_load_microcode(adev);
590 		if (r)
591 			return r;
592 	} else {
593 		r = adev->smu.smumgr_funcs->check_fw_load_finish(adev,
594 						AMDGPU_UCODE_ID_SDMA0);
595 		if (r)
596 			return -EINVAL;
597 		r = adev->smu.smumgr_funcs->check_fw_load_finish(adev,
598 						AMDGPU_UCODE_ID_SDMA1);
599 		if (r)
600 			return -EINVAL;
601 	}
602 
603 	/* unhalt the MEs */
604 	sdma_v2_4_enable(adev, true);
605 
606 	/* start the gfx rings and rlc compute queues */
607 	r = sdma_v2_4_gfx_resume(adev);
608 	if (r)
609 		return r;
610 	r = sdma_v2_4_rlc_resume(adev);
611 	if (r)
612 		return r;
613 
614 	return 0;
615 }
616 
617 /**
618  * sdma_v2_4_ring_test_ring - simple async dma engine test
619  *
620  * @ring: amdgpu_ring structure holding ring information
621  *
622  * Test the DMA engine by writing using it to write an
623  * value to memory. (VI).
624  * Returns 0 for success, error for failure.
625  */
626 static int sdma_v2_4_ring_test_ring(struct amdgpu_ring *ring)
627 {
628 	struct amdgpu_device *adev = ring->adev;
629 	unsigned i;
630 	unsigned index;
631 	int r;
632 	u32 tmp;
633 	u64 gpu_addr;
634 
635 	r = amdgpu_wb_get(adev, &index);
636 	if (r) {
637 		dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
638 		return r;
639 	}
640 
641 	gpu_addr = adev->wb.gpu_addr + (index * 4);
642 	tmp = 0xCAFEDEAD;
643 	adev->wb.wb[index] = cpu_to_le32(tmp);
644 
645 	r = amdgpu_ring_lock(ring, 5);
646 	if (r) {
647 		DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
648 		amdgpu_wb_free(adev, index);
649 		return r;
650 	}
651 
652 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
653 			  SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
654 	amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
655 	amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
656 	amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(1));
657 	amdgpu_ring_write(ring, 0xDEADBEEF);
658 	amdgpu_ring_unlock_commit(ring);
659 
660 	for (i = 0; i < adev->usec_timeout; i++) {
661 		tmp = le32_to_cpu(adev->wb.wb[index]);
662 		if (tmp == 0xDEADBEEF)
663 			break;
664 		DRM_UDELAY(1);
665 	}
666 
667 	if (i < adev->usec_timeout) {
668 		DRM_INFO("ring test on %d succeeded in %d usecs\n", ring->idx, i);
669 	} else {
670 		DRM_ERROR("amdgpu: ring %d test failed (0x%08X)\n",
671 			  ring->idx, tmp);
672 		r = -EINVAL;
673 	}
674 	amdgpu_wb_free(adev, index);
675 
676 	return r;
677 }
678 
679 /**
680  * sdma_v2_4_ring_test_ib - test an IB on the DMA engine
681  *
682  * @ring: amdgpu_ring structure holding ring information
683  *
684  * Test a simple IB in the DMA ring (VI).
685  * Returns 0 on success, error on failure.
686  */
687 static int sdma_v2_4_ring_test_ib(struct amdgpu_ring *ring)
688 {
689 	struct amdgpu_device *adev = ring->adev;
690 	struct amdgpu_ib ib;
691 	struct fence *f = NULL;
692 	unsigned i;
693 	unsigned index;
694 	int r;
695 	u32 tmp = 0;
696 	u64 gpu_addr;
697 
698 	r = amdgpu_wb_get(adev, &index);
699 	if (r) {
700 		dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
701 		return r;
702 	}
703 
704 	gpu_addr = adev->wb.gpu_addr + (index * 4);
705 	tmp = 0xCAFEDEAD;
706 	adev->wb.wb[index] = cpu_to_le32(tmp);
707 	memset(&ib, 0, sizeof(ib));
708 	r = amdgpu_ib_get(ring, NULL, 256, &ib);
709 	if (r) {
710 		DRM_ERROR("amdgpu: failed to get ib (%d).\n", r);
711 		goto err0;
712 	}
713 
714 	ib.ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
715 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
716 	ib.ptr[1] = lower_32_bits(gpu_addr);
717 	ib.ptr[2] = upper_32_bits(gpu_addr);
718 	ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(1);
719 	ib.ptr[4] = 0xDEADBEEF;
720 	ib.ptr[5] = SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
721 	ib.ptr[6] = SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
722 	ib.ptr[7] = SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
723 	ib.length_dw = 8;
724 
725 	r = amdgpu_sched_ib_submit_kernel_helper(adev, ring, &ib, 1, NULL,
726 						 AMDGPU_FENCE_OWNER_UNDEFINED,
727 						 &f);
728 	if (r)
729 		goto err1;
730 
731 	r = fence_wait(f, false);
732 	if (r) {
733 		DRM_ERROR("amdgpu: fence wait failed (%d).\n", r);
734 		goto err1;
735 	}
736 	for (i = 0; i < adev->usec_timeout; i++) {
737 		tmp = le32_to_cpu(adev->wb.wb[index]);
738 		if (tmp == 0xDEADBEEF)
739 			break;
740 		DRM_UDELAY(1);
741 	}
742 	if (i < adev->usec_timeout) {
743 		DRM_INFO("ib test on ring %d succeeded in %u usecs\n",
744 			 ring->idx, i);
745 		goto err1;
746 	} else {
747 		DRM_ERROR("amdgpu: ib test failed (0x%08X)\n", tmp);
748 		r = -EINVAL;
749 	}
750 
751 err1:
752 	fence_put(f);
753 	amdgpu_ib_free(adev, &ib);
754 err0:
755 	amdgpu_wb_free(adev, index);
756 	return r;
757 }
758 
759 /**
760  * sdma_v2_4_vm_copy_pte - update PTEs by copying them from the GART
761  *
762  * @ib: indirect buffer to fill with commands
763  * @pe: addr of the page entry
764  * @src: src addr to copy from
765  * @count: number of page entries to update
766  *
767  * Update PTEs by copying them from the GART using sDMA (CIK).
768  */
769 static void sdma_v2_4_vm_copy_pte(struct amdgpu_ib *ib,
770 				  uint64_t pe, uint64_t src,
771 				  unsigned count)
772 {
773 	while (count) {
774 		unsigned bytes = count * 8;
775 		if (bytes > 0x1FFFF8)
776 			bytes = 0x1FFFF8;
777 
778 		ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
779 			SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
780 		ib->ptr[ib->length_dw++] = bytes;
781 		ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
782 		ib->ptr[ib->length_dw++] = lower_32_bits(src);
783 		ib->ptr[ib->length_dw++] = upper_32_bits(src);
784 		ib->ptr[ib->length_dw++] = lower_32_bits(pe);
785 		ib->ptr[ib->length_dw++] = upper_32_bits(pe);
786 
787 		pe += bytes;
788 		src += bytes;
789 		count -= bytes / 8;
790 	}
791 }
792 
793 /**
794  * sdma_v2_4_vm_write_pte - update PTEs by writing them manually
795  *
796  * @ib: indirect buffer to fill with commands
797  * @pe: addr of the page entry
798  * @addr: dst addr to write into pe
799  * @count: number of page entries to update
800  * @incr: increase next addr by incr bytes
801  * @flags: access flags
802  *
803  * Update PTEs by writing them manually using sDMA (CIK).
804  */
805 static void sdma_v2_4_vm_write_pte(struct amdgpu_ib *ib,
806 				   uint64_t pe,
807 				   uint64_t addr, unsigned count,
808 				   uint32_t incr, uint32_t flags)
809 {
810 	uint64_t value;
811 	unsigned ndw;
812 
813 	while (count) {
814 		ndw = count * 2;
815 		if (ndw > 0xFFFFE)
816 			ndw = 0xFFFFE;
817 
818 		/* for non-physically contiguous pages (system) */
819 		ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
820 			SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
821 		ib->ptr[ib->length_dw++] = pe;
822 		ib->ptr[ib->length_dw++] = upper_32_bits(pe);
823 		ib->ptr[ib->length_dw++] = ndw;
824 		for (; ndw > 0; ndw -= 2, --count, pe += 8) {
825 			if (flags & AMDGPU_PTE_SYSTEM) {
826 				value = amdgpu_vm_map_gart(ib->ring->adev, addr);
827 				value &= 0xFFFFFFFFFFFFF000ULL;
828 			} else if (flags & AMDGPU_PTE_VALID) {
829 				value = addr;
830 			} else {
831 				value = 0;
832 			}
833 			addr += incr;
834 			value |= flags;
835 			ib->ptr[ib->length_dw++] = value;
836 			ib->ptr[ib->length_dw++] = upper_32_bits(value);
837 		}
838 	}
839 }
840 
841 /**
842  * sdma_v2_4_vm_set_pte_pde - update the page tables using sDMA
843  *
844  * @ib: indirect buffer to fill with commands
845  * @pe: addr of the page entry
846  * @addr: dst addr to write into pe
847  * @count: number of page entries to update
848  * @incr: increase next addr by incr bytes
849  * @flags: access flags
850  *
851  * Update the page tables using sDMA (CIK).
852  */
853 static void sdma_v2_4_vm_set_pte_pde(struct amdgpu_ib *ib,
854 				     uint64_t pe,
855 				     uint64_t addr, unsigned count,
856 				     uint32_t incr, uint32_t flags)
857 {
858 	uint64_t value;
859 	unsigned ndw;
860 
861 	while (count) {
862 		ndw = count;
863 		if (ndw > 0x7FFFF)
864 			ndw = 0x7FFFF;
865 
866 		if (flags & AMDGPU_PTE_VALID)
867 			value = addr;
868 		else
869 			value = 0;
870 
871 		/* for physically contiguous pages (vram) */
872 		ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_GEN_PTEPDE);
873 		ib->ptr[ib->length_dw++] = pe; /* dst addr */
874 		ib->ptr[ib->length_dw++] = upper_32_bits(pe);
875 		ib->ptr[ib->length_dw++] = flags; /* mask */
876 		ib->ptr[ib->length_dw++] = 0;
877 		ib->ptr[ib->length_dw++] = value; /* value */
878 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
879 		ib->ptr[ib->length_dw++] = incr; /* increment size */
880 		ib->ptr[ib->length_dw++] = 0;
881 		ib->ptr[ib->length_dw++] = ndw; /* number of entries */
882 
883 		pe += ndw * 8;
884 		addr += ndw * incr;
885 		count -= ndw;
886 	}
887 }
888 
889 /**
890  * sdma_v2_4_vm_pad_ib - pad the IB to the required number of dw
891  *
892  * @ib: indirect buffer to fill with padding
893  *
894  */
895 static void sdma_v2_4_vm_pad_ib(struct amdgpu_ib *ib)
896 {
897 	struct amdgpu_sdma *sdma = amdgpu_get_sdma_instance(ib->ring);
898 	u32 pad_count;
899 	int i;
900 
901 	pad_count = (8 - (ib->length_dw & 0x7)) % 8;
902 	for (i = 0; i < pad_count; i++)
903 		if (sdma && sdma->burst_nop && (i == 0))
904 			ib->ptr[ib->length_dw++] =
905 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP) |
906 				SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
907 		else
908 			ib->ptr[ib->length_dw++] =
909 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
910 }
911 
912 /**
913  * sdma_v2_4_ring_emit_vm_flush - cik vm flush using sDMA
914  *
915  * @ring: amdgpu_ring pointer
916  * @vm: amdgpu_vm pointer
917  *
918  * Update the page table base and flush the VM TLB
919  * using sDMA (VI).
920  */
921 static void sdma_v2_4_ring_emit_vm_flush(struct amdgpu_ring *ring,
922 					 unsigned vm_id, uint64_t pd_addr)
923 {
924 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
925 			  SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
926 	if (vm_id < 8) {
927 		amdgpu_ring_write(ring, (mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + vm_id));
928 	} else {
929 		amdgpu_ring_write(ring, (mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + vm_id - 8));
930 	}
931 	amdgpu_ring_write(ring, pd_addr >> 12);
932 
933 	/* flush TLB */
934 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
935 			  SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
936 	amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST);
937 	amdgpu_ring_write(ring, 1 << vm_id);
938 
939 	/* wait for flush */
940 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
941 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
942 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(0)); /* always */
943 	amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST << 2);
944 	amdgpu_ring_write(ring, 0);
945 	amdgpu_ring_write(ring, 0); /* reference */
946 	amdgpu_ring_write(ring, 0); /* mask */
947 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
948 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
949 }
950 
951 static int sdma_v2_4_early_init(void *handle)
952 {
953 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
954 
955 	sdma_v2_4_set_ring_funcs(adev);
956 	sdma_v2_4_set_buffer_funcs(adev);
957 	sdma_v2_4_set_vm_pte_funcs(adev);
958 	sdma_v2_4_set_irq_funcs(adev);
959 
960 	return 0;
961 }
962 
963 static int sdma_v2_4_sw_init(void *handle)
964 {
965 	struct amdgpu_ring *ring;
966 	int r;
967 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
968 
969 	/* SDMA trap event */
970 	r = amdgpu_irq_add_id(adev, 224, &adev->sdma_trap_irq);
971 	if (r)
972 		return r;
973 
974 	/* SDMA Privileged inst */
975 	r = amdgpu_irq_add_id(adev, 241, &adev->sdma_illegal_inst_irq);
976 	if (r)
977 		return r;
978 
979 	/* SDMA Privileged inst */
980 	r = amdgpu_irq_add_id(adev, 247, &adev->sdma_illegal_inst_irq);
981 	if (r)
982 		return r;
983 
984 	r = sdma_v2_4_init_microcode(adev);
985 	if (r) {
986 		DRM_ERROR("Failed to load sdma firmware!\n");
987 		return r;
988 	}
989 
990 	ring = &adev->sdma[0].ring;
991 	ring->ring_obj = NULL;
992 	ring->use_doorbell = false;
993 
994 	ring = &adev->sdma[1].ring;
995 	ring->ring_obj = NULL;
996 	ring->use_doorbell = false;
997 
998 	ring = &adev->sdma[0].ring;
999 	sprintf(ring->name, "sdma0");
1000 	r = amdgpu_ring_init(adev, ring, 256 * 1024,
1001 			     SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP), 0xf,
1002 			     &adev->sdma_trap_irq, AMDGPU_SDMA_IRQ_TRAP0,
1003 			     AMDGPU_RING_TYPE_SDMA);
1004 	if (r)
1005 		return r;
1006 
1007 	ring = &adev->sdma[1].ring;
1008 	sprintf(ring->name, "sdma1");
1009 	r = amdgpu_ring_init(adev, ring, 256 * 1024,
1010 			     SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP), 0xf,
1011 			     &adev->sdma_trap_irq, AMDGPU_SDMA_IRQ_TRAP1,
1012 			     AMDGPU_RING_TYPE_SDMA);
1013 	if (r)
1014 		return r;
1015 
1016 	return r;
1017 }
1018 
1019 static int sdma_v2_4_sw_fini(void *handle)
1020 {
1021 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1022 
1023 	amdgpu_ring_fini(&adev->sdma[0].ring);
1024 	amdgpu_ring_fini(&adev->sdma[1].ring);
1025 
1026 	return 0;
1027 }
1028 
1029 static int sdma_v2_4_hw_init(void *handle)
1030 {
1031 	int r;
1032 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1033 
1034 	sdma_v2_4_init_golden_registers(adev);
1035 
1036 	r = sdma_v2_4_start(adev);
1037 	if (r)
1038 		return r;
1039 
1040 	return r;
1041 }
1042 
1043 static int sdma_v2_4_hw_fini(void *handle)
1044 {
1045 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1046 
1047 	sdma_v2_4_enable(adev, false);
1048 
1049 	return 0;
1050 }
1051 
1052 static int sdma_v2_4_suspend(void *handle)
1053 {
1054 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1055 
1056 	return sdma_v2_4_hw_fini(adev);
1057 }
1058 
1059 static int sdma_v2_4_resume(void *handle)
1060 {
1061 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1062 
1063 	return sdma_v2_4_hw_init(adev);
1064 }
1065 
1066 static bool sdma_v2_4_is_idle(void *handle)
1067 {
1068 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1069 	u32 tmp = RREG32(mmSRBM_STATUS2);
1070 
1071 	if (tmp & (SRBM_STATUS2__SDMA_BUSY_MASK |
1072 		   SRBM_STATUS2__SDMA1_BUSY_MASK))
1073 	    return false;
1074 
1075 	return true;
1076 }
1077 
1078 static int sdma_v2_4_wait_for_idle(void *handle)
1079 {
1080 	unsigned i;
1081 	u32 tmp;
1082 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1083 
1084 	for (i = 0; i < adev->usec_timeout; i++) {
1085 		tmp = RREG32(mmSRBM_STATUS2) & (SRBM_STATUS2__SDMA_BUSY_MASK |
1086 				SRBM_STATUS2__SDMA1_BUSY_MASK);
1087 
1088 		if (!tmp)
1089 			return 0;
1090 		udelay(1);
1091 	}
1092 	return -ETIMEDOUT;
1093 }
1094 
1095 static void sdma_v2_4_print_status(void *handle)
1096 {
1097 	int i, j;
1098 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1099 
1100 	dev_info(adev->dev, "VI SDMA registers\n");
1101 	dev_info(adev->dev, "  SRBM_STATUS2=0x%08X\n",
1102 		 RREG32(mmSRBM_STATUS2));
1103 	for (i = 0; i < SDMA_MAX_INSTANCE; i++) {
1104 		dev_info(adev->dev, "  SDMA%d_STATUS_REG=0x%08X\n",
1105 			 i, RREG32(mmSDMA0_STATUS_REG + sdma_offsets[i]));
1106 		dev_info(adev->dev, "  SDMA%d_F32_CNTL=0x%08X\n",
1107 			 i, RREG32(mmSDMA0_F32_CNTL + sdma_offsets[i]));
1108 		dev_info(adev->dev, "  SDMA%d_CNTL=0x%08X\n",
1109 			 i, RREG32(mmSDMA0_CNTL + sdma_offsets[i]));
1110 		dev_info(adev->dev, "  SDMA%d_SEM_WAIT_FAIL_TIMER_CNTL=0x%08X\n",
1111 			 i, RREG32(mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i]));
1112 		dev_info(adev->dev, "  SDMA%d_GFX_IB_CNTL=0x%08X\n",
1113 			 i, RREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i]));
1114 		dev_info(adev->dev, "  SDMA%d_GFX_RB_CNTL=0x%08X\n",
1115 			 i, RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]));
1116 		dev_info(adev->dev, "  SDMA%d_GFX_RB_RPTR=0x%08X\n",
1117 			 i, RREG32(mmSDMA0_GFX_RB_RPTR + sdma_offsets[i]));
1118 		dev_info(adev->dev, "  SDMA%d_GFX_RB_WPTR=0x%08X\n",
1119 			 i, RREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i]));
1120 		dev_info(adev->dev, "  SDMA%d_GFX_RB_RPTR_ADDR_HI=0x%08X\n",
1121 			 i, RREG32(mmSDMA0_GFX_RB_RPTR_ADDR_HI + sdma_offsets[i]));
1122 		dev_info(adev->dev, "  SDMA%d_GFX_RB_RPTR_ADDR_LO=0x%08X\n",
1123 			 i, RREG32(mmSDMA0_GFX_RB_RPTR_ADDR_LO + sdma_offsets[i]));
1124 		dev_info(adev->dev, "  SDMA%d_GFX_RB_BASE=0x%08X\n",
1125 			 i, RREG32(mmSDMA0_GFX_RB_BASE + sdma_offsets[i]));
1126 		dev_info(adev->dev, "  SDMA%d_GFX_RB_BASE_HI=0x%08X\n",
1127 			 i, RREG32(mmSDMA0_GFX_RB_BASE_HI + sdma_offsets[i]));
1128 		mutex_lock(&adev->srbm_mutex);
1129 		for (j = 0; j < 16; j++) {
1130 			vi_srbm_select(adev, 0, 0, 0, j);
1131 			dev_info(adev->dev, "  VM %d:\n", j);
1132 			dev_info(adev->dev, "  SDMA%d_GFX_VIRTUAL_ADDR=0x%08X\n",
1133 				 i, RREG32(mmSDMA0_GFX_VIRTUAL_ADDR + sdma_offsets[i]));
1134 			dev_info(adev->dev, "  SDMA%d_GFX_APE1_CNTL=0x%08X\n",
1135 				 i, RREG32(mmSDMA0_GFX_APE1_CNTL + sdma_offsets[i]));
1136 		}
1137 		vi_srbm_select(adev, 0, 0, 0, 0);
1138 		mutex_unlock(&adev->srbm_mutex);
1139 	}
1140 }
1141 
1142 static int sdma_v2_4_soft_reset(void *handle)
1143 {
1144 	u32 srbm_soft_reset = 0;
1145 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1146 	u32 tmp = RREG32(mmSRBM_STATUS2);
1147 
1148 	if (tmp & SRBM_STATUS2__SDMA_BUSY_MASK) {
1149 		/* sdma0 */
1150 		tmp = RREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET);
1151 		tmp = REG_SET_FIELD(tmp, SDMA0_F32_CNTL, HALT, 0);
1152 		WREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET, tmp);
1153 		srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA_MASK;
1154 	}
1155 	if (tmp & SRBM_STATUS2__SDMA1_BUSY_MASK) {
1156 		/* sdma1 */
1157 		tmp = RREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET);
1158 		tmp = REG_SET_FIELD(tmp, SDMA0_F32_CNTL, HALT, 0);
1159 		WREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET, tmp);
1160 		srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA1_MASK;
1161 	}
1162 
1163 	if (srbm_soft_reset) {
1164 		sdma_v2_4_print_status((void *)adev);
1165 
1166 		tmp = RREG32(mmSRBM_SOFT_RESET);
1167 		tmp |= srbm_soft_reset;
1168 		dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
1169 		WREG32(mmSRBM_SOFT_RESET, tmp);
1170 		tmp = RREG32(mmSRBM_SOFT_RESET);
1171 
1172 		udelay(50);
1173 
1174 		tmp &= ~srbm_soft_reset;
1175 		WREG32(mmSRBM_SOFT_RESET, tmp);
1176 		tmp = RREG32(mmSRBM_SOFT_RESET);
1177 
1178 		/* Wait a little for things to settle down */
1179 		udelay(50);
1180 
1181 		sdma_v2_4_print_status((void *)adev);
1182 	}
1183 
1184 	return 0;
1185 }
1186 
1187 static int sdma_v2_4_set_trap_irq_state(struct amdgpu_device *adev,
1188 					struct amdgpu_irq_src *src,
1189 					unsigned type,
1190 					enum amdgpu_interrupt_state state)
1191 {
1192 	u32 sdma_cntl;
1193 
1194 	switch (type) {
1195 	case AMDGPU_SDMA_IRQ_TRAP0:
1196 		switch (state) {
1197 		case AMDGPU_IRQ_STATE_DISABLE:
1198 			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1199 			sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 0);
1200 			WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1201 			break;
1202 		case AMDGPU_IRQ_STATE_ENABLE:
1203 			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1204 			sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 1);
1205 			WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1206 			break;
1207 		default:
1208 			break;
1209 		}
1210 		break;
1211 	case AMDGPU_SDMA_IRQ_TRAP1:
1212 		switch (state) {
1213 		case AMDGPU_IRQ_STATE_DISABLE:
1214 			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1215 			sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 0);
1216 			WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1217 			break;
1218 		case AMDGPU_IRQ_STATE_ENABLE:
1219 			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1220 			sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 1);
1221 			WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1222 			break;
1223 		default:
1224 			break;
1225 		}
1226 		break;
1227 	default:
1228 		break;
1229 	}
1230 	return 0;
1231 }
1232 
1233 static int sdma_v2_4_process_trap_irq(struct amdgpu_device *adev,
1234 				      struct amdgpu_irq_src *source,
1235 				      struct amdgpu_iv_entry *entry)
1236 {
1237 	u8 instance_id, queue_id;
1238 
1239 	instance_id = (entry->ring_id & 0x3) >> 0;
1240 	queue_id = (entry->ring_id & 0xc) >> 2;
1241 	DRM_DEBUG("IH: SDMA trap\n");
1242 	switch (instance_id) {
1243 	case 0:
1244 		switch (queue_id) {
1245 		case 0:
1246 			amdgpu_fence_process(&adev->sdma[0].ring);
1247 			break;
1248 		case 1:
1249 			/* XXX compute */
1250 			break;
1251 		case 2:
1252 			/* XXX compute */
1253 			break;
1254 		}
1255 		break;
1256 	case 1:
1257 		switch (queue_id) {
1258 		case 0:
1259 			amdgpu_fence_process(&adev->sdma[1].ring);
1260 			break;
1261 		case 1:
1262 			/* XXX compute */
1263 			break;
1264 		case 2:
1265 			/* XXX compute */
1266 			break;
1267 		}
1268 		break;
1269 	}
1270 	return 0;
1271 }
1272 
1273 static int sdma_v2_4_process_illegal_inst_irq(struct amdgpu_device *adev,
1274 					      struct amdgpu_irq_src *source,
1275 					      struct amdgpu_iv_entry *entry)
1276 {
1277 	DRM_ERROR("Illegal instruction in SDMA command stream\n");
1278 	schedule_work(&adev->reset_work);
1279 	return 0;
1280 }
1281 
1282 static int sdma_v2_4_set_clockgating_state(void *handle,
1283 					  enum amd_clockgating_state state)
1284 {
1285 	/* XXX handled via the smc on VI */
1286 	return 0;
1287 }
1288 
1289 static int sdma_v2_4_set_powergating_state(void *handle,
1290 					  enum amd_powergating_state state)
1291 {
1292 	return 0;
1293 }
1294 
1295 const struct amd_ip_funcs sdma_v2_4_ip_funcs = {
1296 	.early_init = sdma_v2_4_early_init,
1297 	.late_init = NULL,
1298 	.sw_init = sdma_v2_4_sw_init,
1299 	.sw_fini = sdma_v2_4_sw_fini,
1300 	.hw_init = sdma_v2_4_hw_init,
1301 	.hw_fini = sdma_v2_4_hw_fini,
1302 	.suspend = sdma_v2_4_suspend,
1303 	.resume = sdma_v2_4_resume,
1304 	.is_idle = sdma_v2_4_is_idle,
1305 	.wait_for_idle = sdma_v2_4_wait_for_idle,
1306 	.soft_reset = sdma_v2_4_soft_reset,
1307 	.print_status = sdma_v2_4_print_status,
1308 	.set_clockgating_state = sdma_v2_4_set_clockgating_state,
1309 	.set_powergating_state = sdma_v2_4_set_powergating_state,
1310 };
1311 
1312 /**
1313  * sdma_v2_4_ring_is_lockup - Check if the DMA engine is locked up
1314  *
1315  * @ring: amdgpu_ring structure holding ring information
1316  *
1317  * Check if the async DMA engine is locked up (VI).
1318  * Returns true if the engine appears to be locked up, false if not.
1319  */
1320 static bool sdma_v2_4_ring_is_lockup(struct amdgpu_ring *ring)
1321 {
1322 
1323 	if (sdma_v2_4_is_idle(ring->adev)) {
1324 		amdgpu_ring_lockup_update(ring);
1325 		return false;
1326 	}
1327 	return amdgpu_ring_test_lockup(ring);
1328 }
1329 
1330 static const struct amdgpu_ring_funcs sdma_v2_4_ring_funcs = {
1331 	.get_rptr = sdma_v2_4_ring_get_rptr,
1332 	.get_wptr = sdma_v2_4_ring_get_wptr,
1333 	.set_wptr = sdma_v2_4_ring_set_wptr,
1334 	.parse_cs = NULL,
1335 	.emit_ib = sdma_v2_4_ring_emit_ib,
1336 	.emit_fence = sdma_v2_4_ring_emit_fence,
1337 	.emit_semaphore = sdma_v2_4_ring_emit_semaphore,
1338 	.emit_vm_flush = sdma_v2_4_ring_emit_vm_flush,
1339 	.emit_hdp_flush = sdma_v2_4_ring_emit_hdp_flush,
1340 	.test_ring = sdma_v2_4_ring_test_ring,
1341 	.test_ib = sdma_v2_4_ring_test_ib,
1342 	.is_lockup = sdma_v2_4_ring_is_lockup,
1343 	.insert_nop = sdma_v2_4_ring_insert_nop,
1344 };
1345 
1346 static void sdma_v2_4_set_ring_funcs(struct amdgpu_device *adev)
1347 {
1348 	adev->sdma[0].ring.funcs = &sdma_v2_4_ring_funcs;
1349 	adev->sdma[1].ring.funcs = &sdma_v2_4_ring_funcs;
1350 }
1351 
1352 static const struct amdgpu_irq_src_funcs sdma_v2_4_trap_irq_funcs = {
1353 	.set = sdma_v2_4_set_trap_irq_state,
1354 	.process = sdma_v2_4_process_trap_irq,
1355 };
1356 
1357 static const struct amdgpu_irq_src_funcs sdma_v2_4_illegal_inst_irq_funcs = {
1358 	.process = sdma_v2_4_process_illegal_inst_irq,
1359 };
1360 
1361 static void sdma_v2_4_set_irq_funcs(struct amdgpu_device *adev)
1362 {
1363 	adev->sdma_trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST;
1364 	adev->sdma_trap_irq.funcs = &sdma_v2_4_trap_irq_funcs;
1365 	adev->sdma_illegal_inst_irq.funcs = &sdma_v2_4_illegal_inst_irq_funcs;
1366 }
1367 
1368 /**
1369  * sdma_v2_4_emit_copy_buffer - copy buffer using the sDMA engine
1370  *
1371  * @ring: amdgpu_ring structure holding ring information
1372  * @src_offset: src GPU address
1373  * @dst_offset: dst GPU address
1374  * @byte_count: number of bytes to xfer
1375  *
1376  * Copy GPU buffers using the DMA engine (VI).
1377  * Used by the amdgpu ttm implementation to move pages if
1378  * registered as the asic copy callback.
1379  */
1380 static void sdma_v2_4_emit_copy_buffer(struct amdgpu_ib *ib,
1381 				       uint64_t src_offset,
1382 				       uint64_t dst_offset,
1383 				       uint32_t byte_count)
1384 {
1385 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1386 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
1387 	ib->ptr[ib->length_dw++] = byte_count;
1388 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1389 	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1390 	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1391 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1392 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1393 }
1394 
1395 /**
1396  * sdma_v2_4_emit_fill_buffer - fill buffer using the sDMA engine
1397  *
1398  * @ring: amdgpu_ring structure holding ring information
1399  * @src_data: value to write to buffer
1400  * @dst_offset: dst GPU address
1401  * @byte_count: number of bytes to xfer
1402  *
1403  * Fill GPU buffers using the DMA engine (VI).
1404  */
1405 static void sdma_v2_4_emit_fill_buffer(struct amdgpu_ib *ib,
1406 				       uint32_t src_data,
1407 				       uint64_t dst_offset,
1408 				       uint32_t byte_count)
1409 {
1410 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL);
1411 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1412 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1413 	ib->ptr[ib->length_dw++] = src_data;
1414 	ib->ptr[ib->length_dw++] = byte_count;
1415 }
1416 
1417 static const struct amdgpu_buffer_funcs sdma_v2_4_buffer_funcs = {
1418 	.copy_max_bytes = 0x1fffff,
1419 	.copy_num_dw = 7,
1420 	.emit_copy_buffer = sdma_v2_4_emit_copy_buffer,
1421 
1422 	.fill_max_bytes = 0x1fffff,
1423 	.fill_num_dw = 7,
1424 	.emit_fill_buffer = sdma_v2_4_emit_fill_buffer,
1425 };
1426 
1427 static void sdma_v2_4_set_buffer_funcs(struct amdgpu_device *adev)
1428 {
1429 	if (adev->mman.buffer_funcs == NULL) {
1430 		adev->mman.buffer_funcs = &sdma_v2_4_buffer_funcs;
1431 		adev->mman.buffer_funcs_ring = &adev->sdma[0].ring;
1432 	}
1433 }
1434 
1435 static const struct amdgpu_vm_pte_funcs sdma_v2_4_vm_pte_funcs = {
1436 	.copy_pte = sdma_v2_4_vm_copy_pte,
1437 	.write_pte = sdma_v2_4_vm_write_pte,
1438 	.set_pte_pde = sdma_v2_4_vm_set_pte_pde,
1439 	.pad_ib = sdma_v2_4_vm_pad_ib,
1440 };
1441 
1442 static void sdma_v2_4_set_vm_pte_funcs(struct amdgpu_device *adev)
1443 {
1444 	if (adev->vm_manager.vm_pte_funcs == NULL) {
1445 		adev->vm_manager.vm_pte_funcs = &sdma_v2_4_vm_pte_funcs;
1446 		adev->vm_manager.vm_pte_funcs_ring = &adev->sdma[0].ring;
1447 		adev->vm_manager.vm_pte_funcs_ring->is_pte_ring = true;
1448 	}
1449 }
1450