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_7_1_d.h"
36 #include "gmc/gmc_7_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 = 0, 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 < adev->sdma.num_instances; 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.instance[i].fw, fw_name, adev->dev);
141 		if (err)
142 			goto out;
143 		err = amdgpu_ucode_validate(adev->sdma.instance[i].fw);
144 		if (err)
145 			goto out;
146 		hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
147 		adev->sdma.instance[i].fw_version = le32_to_cpu(hdr->header.ucode_version);
148 		adev->sdma.instance[i].feature_version = le32_to_cpu(hdr->ucode_feature_version);
149 		if (adev->sdma.instance[i].feature_version >= 20)
150 			adev->sdma.instance[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.instance[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 < adev->sdma.num_instances; i++) {
168 			release_firmware(adev->sdma.instance[i].fw);
169 			adev->sdma.instance[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.instance[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.instance[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_instance *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_id & 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.instance[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 static void sdma_v2_4_ring_emit_hdp_invalidate(struct amdgpu_ring *ring)
304 {
305 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
306 			  SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
307 	amdgpu_ring_write(ring, mmHDP_DEBUG0);
308 	amdgpu_ring_write(ring, 1);
309 }
310 /**
311  * sdma_v2_4_ring_emit_fence - emit a fence on the DMA ring
312  *
313  * @ring: amdgpu ring pointer
314  * @fence: amdgpu fence object
315  *
316  * Add a DMA fence packet to the ring to write
317  * the fence seq number and DMA trap packet to generate
318  * an interrupt if needed (VI).
319  */
320 static void sdma_v2_4_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
321 				      unsigned flags)
322 {
323 	bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
324 	/* write the fence */
325 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE));
326 	amdgpu_ring_write(ring, lower_32_bits(addr));
327 	amdgpu_ring_write(ring, upper_32_bits(addr));
328 	amdgpu_ring_write(ring, lower_32_bits(seq));
329 
330 	/* optionally write high bits as well */
331 	if (write64bit) {
332 		addr += 4;
333 		amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE));
334 		amdgpu_ring_write(ring, lower_32_bits(addr));
335 		amdgpu_ring_write(ring, upper_32_bits(addr));
336 		amdgpu_ring_write(ring, upper_32_bits(seq));
337 	}
338 
339 	/* generate an interrupt */
340 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP));
341 	amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(0));
342 }
343 
344 /**
345  * sdma_v2_4_gfx_stop - stop the gfx async dma engines
346  *
347  * @adev: amdgpu_device pointer
348  *
349  * Stop the gfx async dma ring buffers (VI).
350  */
351 static void sdma_v2_4_gfx_stop(struct amdgpu_device *adev)
352 {
353 	struct amdgpu_ring *sdma0 = &adev->sdma.instance[0].ring;
354 	struct amdgpu_ring *sdma1 = &adev->sdma.instance[1].ring;
355 	u32 rb_cntl, ib_cntl;
356 	int i;
357 
358 	if ((adev->mman.buffer_funcs_ring == sdma0) ||
359 	    (adev->mman.buffer_funcs_ring == sdma1))
360 		amdgpu_ttm_set_active_vram_size(adev, adev->mc.visible_vram_size);
361 
362 	for (i = 0; i < adev->sdma.num_instances; i++) {
363 		rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
364 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 0);
365 		WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
366 		ib_cntl = RREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i]);
367 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 0);
368 		WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
369 	}
370 	sdma0->ready = false;
371 	sdma1->ready = false;
372 }
373 
374 /**
375  * sdma_v2_4_rlc_stop - stop the compute async dma engines
376  *
377  * @adev: amdgpu_device pointer
378  *
379  * Stop the compute async dma queues (VI).
380  */
381 static void sdma_v2_4_rlc_stop(struct amdgpu_device *adev)
382 {
383 	/* XXX todo */
384 }
385 
386 /**
387  * sdma_v2_4_enable - stop the async dma engines
388  *
389  * @adev: amdgpu_device pointer
390  * @enable: enable/disable the DMA MEs.
391  *
392  * Halt or unhalt the async dma engines (VI).
393  */
394 static void sdma_v2_4_enable(struct amdgpu_device *adev, bool enable)
395 {
396 	u32 f32_cntl;
397 	int i;
398 
399 	if (enable == false) {
400 		sdma_v2_4_gfx_stop(adev);
401 		sdma_v2_4_rlc_stop(adev);
402 	}
403 
404 	for (i = 0; i < adev->sdma.num_instances; i++) {
405 		f32_cntl = RREG32(mmSDMA0_F32_CNTL + sdma_offsets[i]);
406 		if (enable)
407 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 0);
408 		else
409 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 1);
410 		WREG32(mmSDMA0_F32_CNTL + sdma_offsets[i], f32_cntl);
411 	}
412 }
413 
414 /**
415  * sdma_v2_4_gfx_resume - setup and start the async dma engines
416  *
417  * @adev: amdgpu_device pointer
418  *
419  * Set up the gfx DMA ring buffers and enable them (VI).
420  * Returns 0 for success, error for failure.
421  */
422 static int sdma_v2_4_gfx_resume(struct amdgpu_device *adev)
423 {
424 	struct amdgpu_ring *ring;
425 	u32 rb_cntl, ib_cntl;
426 	u32 rb_bufsz;
427 	u32 wb_offset;
428 	int i, j, r;
429 
430 	for (i = 0; i < adev->sdma.num_instances; i++) {
431 		ring = &adev->sdma.instance[i].ring;
432 		wb_offset = (ring->rptr_offs * 4);
433 
434 		mutex_lock(&adev->srbm_mutex);
435 		for (j = 0; j < 16; j++) {
436 			vi_srbm_select(adev, 0, 0, 0, j);
437 			/* SDMA GFX */
438 			WREG32(mmSDMA0_GFX_VIRTUAL_ADDR + sdma_offsets[i], 0);
439 			WREG32(mmSDMA0_GFX_APE1_CNTL + sdma_offsets[i], 0);
440 		}
441 		vi_srbm_select(adev, 0, 0, 0, 0);
442 		mutex_unlock(&adev->srbm_mutex);
443 
444 		WREG32(mmSDMA0_TILING_CONFIG + sdma_offsets[i],
445 		       adev->gfx.config.gb_addr_config & 0x70);
446 
447 		WREG32(mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i], 0);
448 
449 		/* Set ring buffer size in dwords */
450 		rb_bufsz = order_base_2(ring->ring_size / 4);
451 		rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
452 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz);
453 #ifdef __BIG_ENDIAN
454 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1);
455 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL,
456 					RPTR_WRITEBACK_SWAP_ENABLE, 1);
457 #endif
458 		WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
459 
460 		/* Initialize the ring buffer's read and write pointers */
461 		WREG32(mmSDMA0_GFX_RB_RPTR + sdma_offsets[i], 0);
462 		WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], 0);
463 
464 		/* set the wb address whether it's enabled or not */
465 		WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_HI + sdma_offsets[i],
466 		       upper_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFF);
467 		WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_LO + sdma_offsets[i],
468 		       lower_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC);
469 
470 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
471 
472 		WREG32(mmSDMA0_GFX_RB_BASE + sdma_offsets[i], ring->gpu_addr >> 8);
473 		WREG32(mmSDMA0_GFX_RB_BASE_HI + sdma_offsets[i], ring->gpu_addr >> 40);
474 
475 		ring->wptr = 0;
476 		WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], ring->wptr << 2);
477 
478 		/* enable DMA RB */
479 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1);
480 		WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
481 
482 		ib_cntl = RREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i]);
483 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 1);
484 #ifdef __BIG_ENDIAN
485 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1);
486 #endif
487 		/* enable DMA IBs */
488 		WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
489 
490 		ring->ready = true;
491 
492 		r = amdgpu_ring_test_ring(ring);
493 		if (r) {
494 			ring->ready = false;
495 			return r;
496 		}
497 
498 		if (adev->mman.buffer_funcs_ring == ring)
499 			amdgpu_ttm_set_active_vram_size(adev, adev->mc.real_vram_size);
500 	}
501 
502 	return 0;
503 }
504 
505 /**
506  * sdma_v2_4_rlc_resume - setup and start the async dma engines
507  *
508  * @adev: amdgpu_device pointer
509  *
510  * Set up the compute DMA queues and enable them (VI).
511  * Returns 0 for success, error for failure.
512  */
513 static int sdma_v2_4_rlc_resume(struct amdgpu_device *adev)
514 {
515 	/* XXX todo */
516 	return 0;
517 }
518 
519 /**
520  * sdma_v2_4_load_microcode - load the sDMA ME ucode
521  *
522  * @adev: amdgpu_device pointer
523  *
524  * Loads the sDMA0/1 ucode.
525  * Returns 0 for success, -EINVAL if the ucode is not available.
526  */
527 static int sdma_v2_4_load_microcode(struct amdgpu_device *adev)
528 {
529 	const struct sdma_firmware_header_v1_0 *hdr;
530 	const __le32 *fw_data;
531 	u32 fw_size;
532 	int i, j;
533 
534 	/* halt the MEs */
535 	sdma_v2_4_enable(adev, false);
536 
537 	for (i = 0; i < adev->sdma.num_instances; i++) {
538 		if (!adev->sdma.instance[i].fw)
539 			return -EINVAL;
540 		hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
541 		amdgpu_ucode_print_sdma_hdr(&hdr->header);
542 		fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
543 		fw_data = (const __le32 *)
544 			(adev->sdma.instance[i].fw->data +
545 			 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
546 		WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], 0);
547 		for (j = 0; j < fw_size; j++)
548 			WREG32(mmSDMA0_UCODE_DATA + sdma_offsets[i], le32_to_cpup(fw_data++));
549 		WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], adev->sdma.instance[i].fw_version);
550 	}
551 
552 	return 0;
553 }
554 
555 /**
556  * sdma_v2_4_start - setup and start the async dma engines
557  *
558  * @adev: amdgpu_device pointer
559  *
560  * Set up the DMA engines and enable them (VI).
561  * Returns 0 for success, error for failure.
562  */
563 static int sdma_v2_4_start(struct amdgpu_device *adev)
564 {
565 	int r;
566 
567 	if (!adev->firmware.smu_load) {
568 		r = sdma_v2_4_load_microcode(adev);
569 		if (r)
570 			return r;
571 	} else {
572 		r = adev->smu.smumgr_funcs->check_fw_load_finish(adev,
573 						AMDGPU_UCODE_ID_SDMA0);
574 		if (r)
575 			return -EINVAL;
576 		r = adev->smu.smumgr_funcs->check_fw_load_finish(adev,
577 						AMDGPU_UCODE_ID_SDMA1);
578 		if (r)
579 			return -EINVAL;
580 	}
581 
582 	/* unhalt the MEs */
583 	sdma_v2_4_enable(adev, true);
584 
585 	/* start the gfx rings and rlc compute queues */
586 	r = sdma_v2_4_gfx_resume(adev);
587 	if (r)
588 		return r;
589 	r = sdma_v2_4_rlc_resume(adev);
590 	if (r)
591 		return r;
592 
593 	return 0;
594 }
595 
596 /**
597  * sdma_v2_4_ring_test_ring - simple async dma engine test
598  *
599  * @ring: amdgpu_ring structure holding ring information
600  *
601  * Test the DMA engine by writing using it to write an
602  * value to memory. (VI).
603  * Returns 0 for success, error for failure.
604  */
605 static int sdma_v2_4_ring_test_ring(struct amdgpu_ring *ring)
606 {
607 	struct amdgpu_device *adev = ring->adev;
608 	unsigned i;
609 	unsigned index;
610 	int r;
611 	u32 tmp;
612 	u64 gpu_addr;
613 
614 	r = amdgpu_wb_get(adev, &index);
615 	if (r) {
616 		dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
617 		return r;
618 	}
619 
620 	gpu_addr = adev->wb.gpu_addr + (index * 4);
621 	tmp = 0xCAFEDEAD;
622 	adev->wb.wb[index] = cpu_to_le32(tmp);
623 
624 	r = amdgpu_ring_alloc(ring, 5);
625 	if (r) {
626 		DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
627 		amdgpu_wb_free(adev, index);
628 		return r;
629 	}
630 
631 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
632 			  SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
633 	amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
634 	amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
635 	amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(1));
636 	amdgpu_ring_write(ring, 0xDEADBEEF);
637 	amdgpu_ring_commit(ring);
638 
639 	for (i = 0; i < adev->usec_timeout; i++) {
640 		tmp = le32_to_cpu(adev->wb.wb[index]);
641 		if (tmp == 0xDEADBEEF)
642 			break;
643 		DRM_UDELAY(1);
644 	}
645 
646 	if (i < adev->usec_timeout) {
647 		DRM_INFO("ring test on %d succeeded in %d usecs\n", ring->idx, i);
648 	} else {
649 		DRM_ERROR("amdgpu: ring %d test failed (0x%08X)\n",
650 			  ring->idx, tmp);
651 		r = -EINVAL;
652 	}
653 	amdgpu_wb_free(adev, index);
654 
655 	return r;
656 }
657 
658 /**
659  * sdma_v2_4_ring_test_ib - test an IB on the DMA engine
660  *
661  * @ring: amdgpu_ring structure holding ring information
662  *
663  * Test a simple IB in the DMA ring (VI).
664  * Returns 0 on success, error on failure.
665  */
666 static int sdma_v2_4_ring_test_ib(struct amdgpu_ring *ring)
667 {
668 	struct amdgpu_device *adev = ring->adev;
669 	struct amdgpu_ib ib;
670 	struct fence *f = NULL;
671 	unsigned i;
672 	unsigned index;
673 	int r;
674 	u32 tmp = 0;
675 	u64 gpu_addr;
676 
677 	r = amdgpu_wb_get(adev, &index);
678 	if (r) {
679 		dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
680 		return r;
681 	}
682 
683 	gpu_addr = adev->wb.gpu_addr + (index * 4);
684 	tmp = 0xCAFEDEAD;
685 	adev->wb.wb[index] = cpu_to_le32(tmp);
686 	memset(&ib, 0, sizeof(ib));
687 	r = amdgpu_ib_get(adev, NULL, 256, &ib);
688 	if (r) {
689 		DRM_ERROR("amdgpu: failed to get ib (%d).\n", r);
690 		goto err0;
691 	}
692 
693 	ib.ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
694 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
695 	ib.ptr[1] = lower_32_bits(gpu_addr);
696 	ib.ptr[2] = upper_32_bits(gpu_addr);
697 	ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(1);
698 	ib.ptr[4] = 0xDEADBEEF;
699 	ib.ptr[5] = SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
700 	ib.ptr[6] = SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
701 	ib.ptr[7] = SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
702 	ib.length_dw = 8;
703 
704 	r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
705 	if (r)
706 		goto err1;
707 
708 	r = fence_wait(f, false);
709 	if (r) {
710 		DRM_ERROR("amdgpu: fence wait failed (%d).\n", r);
711 		goto err1;
712 	}
713 	for (i = 0; i < adev->usec_timeout; i++) {
714 		tmp = le32_to_cpu(adev->wb.wb[index]);
715 		if (tmp == 0xDEADBEEF)
716 			break;
717 		DRM_UDELAY(1);
718 	}
719 	if (i < adev->usec_timeout) {
720 		DRM_INFO("ib test on ring %d succeeded in %u usecs\n",
721 			 ring->idx, i);
722 		goto err1;
723 	} else {
724 		DRM_ERROR("amdgpu: ib test failed (0x%08X)\n", tmp);
725 		r = -EINVAL;
726 	}
727 
728 err1:
729 	fence_put(f);
730 	amdgpu_ib_free(adev, &ib, NULL);
731 	fence_put(f);
732 err0:
733 	amdgpu_wb_free(adev, index);
734 	return r;
735 }
736 
737 /**
738  * sdma_v2_4_vm_copy_pte - update PTEs by copying them from the GART
739  *
740  * @ib: indirect buffer to fill with commands
741  * @pe: addr of the page entry
742  * @src: src addr to copy from
743  * @count: number of page entries to update
744  *
745  * Update PTEs by copying them from the GART using sDMA (CIK).
746  */
747 static void sdma_v2_4_vm_copy_pte(struct amdgpu_ib *ib,
748 				  uint64_t pe, uint64_t src,
749 				  unsigned count)
750 {
751 	while (count) {
752 		unsigned bytes = count * 8;
753 		if (bytes > 0x1FFFF8)
754 			bytes = 0x1FFFF8;
755 
756 		ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
757 			SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
758 		ib->ptr[ib->length_dw++] = bytes;
759 		ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
760 		ib->ptr[ib->length_dw++] = lower_32_bits(src);
761 		ib->ptr[ib->length_dw++] = upper_32_bits(src);
762 		ib->ptr[ib->length_dw++] = lower_32_bits(pe);
763 		ib->ptr[ib->length_dw++] = upper_32_bits(pe);
764 
765 		pe += bytes;
766 		src += bytes;
767 		count -= bytes / 8;
768 	}
769 }
770 
771 /**
772  * sdma_v2_4_vm_write_pte - update PTEs by writing them manually
773  *
774  * @ib: indirect buffer to fill with commands
775  * @pe: addr of the page entry
776  * @addr: dst addr to write into pe
777  * @count: number of page entries to update
778  * @incr: increase next addr by incr bytes
779  * @flags: access flags
780  *
781  * Update PTEs by writing them manually using sDMA (CIK).
782  */
783 static void sdma_v2_4_vm_write_pte(struct amdgpu_ib *ib,
784 				   const dma_addr_t *pages_addr, uint64_t pe,
785 				   uint64_t addr, unsigned count,
786 				   uint32_t incr, uint32_t flags)
787 {
788 	uint64_t value;
789 	unsigned ndw;
790 
791 	while (count) {
792 		ndw = count * 2;
793 		if (ndw > 0xFFFFE)
794 			ndw = 0xFFFFE;
795 
796 		/* for non-physically contiguous pages (system) */
797 		ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
798 			SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
799 		ib->ptr[ib->length_dw++] = pe;
800 		ib->ptr[ib->length_dw++] = upper_32_bits(pe);
801 		ib->ptr[ib->length_dw++] = ndw;
802 		for (; ndw > 0; ndw -= 2, --count, pe += 8) {
803 			value = amdgpu_vm_map_gart(pages_addr, addr);
804 			addr += incr;
805 			value |= flags;
806 			ib->ptr[ib->length_dw++] = value;
807 			ib->ptr[ib->length_dw++] = upper_32_bits(value);
808 		}
809 	}
810 }
811 
812 /**
813  * sdma_v2_4_vm_set_pte_pde - update the page tables using sDMA
814  *
815  * @ib: indirect buffer to fill with commands
816  * @pe: addr of the page entry
817  * @addr: dst addr to write into pe
818  * @count: number of page entries to update
819  * @incr: increase next addr by incr bytes
820  * @flags: access flags
821  *
822  * Update the page tables using sDMA (CIK).
823  */
824 static void sdma_v2_4_vm_set_pte_pde(struct amdgpu_ib *ib,
825 				     uint64_t pe,
826 				     uint64_t addr, unsigned count,
827 				     uint32_t incr, uint32_t flags)
828 {
829 	uint64_t value;
830 	unsigned ndw;
831 
832 	while (count) {
833 		ndw = count;
834 		if (ndw > 0x7FFFF)
835 			ndw = 0x7FFFF;
836 
837 		if (flags & AMDGPU_PTE_VALID)
838 			value = addr;
839 		else
840 			value = 0;
841 
842 		/* for physically contiguous pages (vram) */
843 		ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_GEN_PTEPDE);
844 		ib->ptr[ib->length_dw++] = pe; /* dst addr */
845 		ib->ptr[ib->length_dw++] = upper_32_bits(pe);
846 		ib->ptr[ib->length_dw++] = flags; /* mask */
847 		ib->ptr[ib->length_dw++] = 0;
848 		ib->ptr[ib->length_dw++] = value; /* value */
849 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
850 		ib->ptr[ib->length_dw++] = incr; /* increment size */
851 		ib->ptr[ib->length_dw++] = 0;
852 		ib->ptr[ib->length_dw++] = ndw; /* number of entries */
853 
854 		pe += ndw * 8;
855 		addr += ndw * incr;
856 		count -= ndw;
857 	}
858 }
859 
860 /**
861  * sdma_v2_4_ring_pad_ib - pad the IB to the required number of dw
862  *
863  * @ib: indirect buffer to fill with padding
864  *
865  */
866 static void sdma_v2_4_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
867 {
868 	struct amdgpu_sdma_instance *sdma = amdgpu_get_sdma_instance(ring);
869 	u32 pad_count;
870 	int i;
871 
872 	pad_count = (8 - (ib->length_dw & 0x7)) % 8;
873 	for (i = 0; i < pad_count; i++)
874 		if (sdma && sdma->burst_nop && (i == 0))
875 			ib->ptr[ib->length_dw++] =
876 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP) |
877 				SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
878 		else
879 			ib->ptr[ib->length_dw++] =
880 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
881 }
882 
883 /**
884  * sdma_v2_4_ring_emit_pipeline_sync - sync the pipeline
885  *
886  * @ring: amdgpu_ring pointer
887  *
888  * Make sure all previous operations are completed (CIK).
889  */
890 static void sdma_v2_4_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
891 {
892 	uint32_t seq = ring->fence_drv.sync_seq;
893 	uint64_t addr = ring->fence_drv.gpu_addr;
894 
895 	/* wait for idle */
896 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
897 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
898 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
899 			  SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1));
900 	amdgpu_ring_write(ring, addr & 0xfffffffc);
901 	amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
902 	amdgpu_ring_write(ring, seq); /* reference */
903 	amdgpu_ring_write(ring, 0xfffffff); /* mask */
904 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
905 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */
906 }
907 
908 /**
909  * sdma_v2_4_ring_emit_vm_flush - cik vm flush using sDMA
910  *
911  * @ring: amdgpu_ring pointer
912  * @vm: amdgpu_vm pointer
913  *
914  * Update the page table base and flush the VM TLB
915  * using sDMA (VI).
916  */
917 static void sdma_v2_4_ring_emit_vm_flush(struct amdgpu_ring *ring,
918 					 unsigned vm_id, uint64_t pd_addr)
919 {
920 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
921 			  SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
922 	if (vm_id < 8) {
923 		amdgpu_ring_write(ring, (mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + vm_id));
924 	} else {
925 		amdgpu_ring_write(ring, (mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + vm_id - 8));
926 	}
927 	amdgpu_ring_write(ring, pd_addr >> 12);
928 
929 	/* flush TLB */
930 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
931 			  SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
932 	amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST);
933 	amdgpu_ring_write(ring, 1 << vm_id);
934 
935 	/* wait for flush */
936 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
937 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
938 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(0)); /* always */
939 	amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST << 2);
940 	amdgpu_ring_write(ring, 0);
941 	amdgpu_ring_write(ring, 0); /* reference */
942 	amdgpu_ring_write(ring, 0); /* mask */
943 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
944 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
945 }
946 
947 static int sdma_v2_4_early_init(void *handle)
948 {
949 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
950 
951 	adev->sdma.num_instances = SDMA_MAX_INSTANCE;
952 
953 	sdma_v2_4_set_ring_funcs(adev);
954 	sdma_v2_4_set_buffer_funcs(adev);
955 	sdma_v2_4_set_vm_pte_funcs(adev);
956 	sdma_v2_4_set_irq_funcs(adev);
957 
958 	return 0;
959 }
960 
961 static int sdma_v2_4_sw_init(void *handle)
962 {
963 	struct amdgpu_ring *ring;
964 	int r, i;
965 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
966 
967 	/* SDMA trap event */
968 	r = amdgpu_irq_add_id(adev, 224, &adev->sdma.trap_irq);
969 	if (r)
970 		return r;
971 
972 	/* SDMA Privileged inst */
973 	r = amdgpu_irq_add_id(adev, 241, &adev->sdma.illegal_inst_irq);
974 	if (r)
975 		return r;
976 
977 	/* SDMA Privileged inst */
978 	r = amdgpu_irq_add_id(adev, 247, &adev->sdma.illegal_inst_irq);
979 	if (r)
980 		return r;
981 
982 	r = sdma_v2_4_init_microcode(adev);
983 	if (r) {
984 		DRM_ERROR("Failed to load sdma firmware!\n");
985 		return r;
986 	}
987 
988 	for (i = 0; i < adev->sdma.num_instances; i++) {
989 		ring = &adev->sdma.instance[i].ring;
990 		ring->ring_obj = NULL;
991 		ring->use_doorbell = false;
992 		sprintf(ring->name, "sdma%d", i);
993 		r = amdgpu_ring_init(adev, ring, 256 * 1024,
994 				     SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP), 0xf,
995 				     &adev->sdma.trap_irq,
996 				     (i == 0) ?
997 				     AMDGPU_SDMA_IRQ_TRAP0 : AMDGPU_SDMA_IRQ_TRAP1,
998 				     AMDGPU_RING_TYPE_SDMA);
999 		if (r)
1000 			return r;
1001 	}
1002 
1003 	return r;
1004 }
1005 
1006 static int sdma_v2_4_sw_fini(void *handle)
1007 {
1008 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1009 	int i;
1010 
1011 	for (i = 0; i < adev->sdma.num_instances; i++)
1012 		amdgpu_ring_fini(&adev->sdma.instance[i].ring);
1013 
1014 	return 0;
1015 }
1016 
1017 static int sdma_v2_4_hw_init(void *handle)
1018 {
1019 	int r;
1020 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1021 
1022 	sdma_v2_4_init_golden_registers(adev);
1023 
1024 	r = sdma_v2_4_start(adev);
1025 	if (r)
1026 		return r;
1027 
1028 	return r;
1029 }
1030 
1031 static int sdma_v2_4_hw_fini(void *handle)
1032 {
1033 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1034 
1035 	sdma_v2_4_enable(adev, false);
1036 
1037 	return 0;
1038 }
1039 
1040 static int sdma_v2_4_suspend(void *handle)
1041 {
1042 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1043 
1044 	return sdma_v2_4_hw_fini(adev);
1045 }
1046 
1047 static int sdma_v2_4_resume(void *handle)
1048 {
1049 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1050 
1051 	return sdma_v2_4_hw_init(adev);
1052 }
1053 
1054 static bool sdma_v2_4_is_idle(void *handle)
1055 {
1056 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1057 	u32 tmp = RREG32(mmSRBM_STATUS2);
1058 
1059 	if (tmp & (SRBM_STATUS2__SDMA_BUSY_MASK |
1060 		   SRBM_STATUS2__SDMA1_BUSY_MASK))
1061 	    return false;
1062 
1063 	return true;
1064 }
1065 
1066 static int sdma_v2_4_wait_for_idle(void *handle)
1067 {
1068 	unsigned i;
1069 	u32 tmp;
1070 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1071 
1072 	for (i = 0; i < adev->usec_timeout; i++) {
1073 		tmp = RREG32(mmSRBM_STATUS2) & (SRBM_STATUS2__SDMA_BUSY_MASK |
1074 				SRBM_STATUS2__SDMA1_BUSY_MASK);
1075 
1076 		if (!tmp)
1077 			return 0;
1078 		udelay(1);
1079 	}
1080 	return -ETIMEDOUT;
1081 }
1082 
1083 static void sdma_v2_4_print_status(void *handle)
1084 {
1085 	int i, j;
1086 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1087 
1088 	dev_info(adev->dev, "VI SDMA registers\n");
1089 	dev_info(adev->dev, "  SRBM_STATUS2=0x%08X\n",
1090 		 RREG32(mmSRBM_STATUS2));
1091 	for (i = 0; i < adev->sdma.num_instances; i++) {
1092 		dev_info(adev->dev, "  SDMA%d_STATUS_REG=0x%08X\n",
1093 			 i, RREG32(mmSDMA0_STATUS_REG + sdma_offsets[i]));
1094 		dev_info(adev->dev, "  SDMA%d_F32_CNTL=0x%08X\n",
1095 			 i, RREG32(mmSDMA0_F32_CNTL + sdma_offsets[i]));
1096 		dev_info(adev->dev, "  SDMA%d_CNTL=0x%08X\n",
1097 			 i, RREG32(mmSDMA0_CNTL + sdma_offsets[i]));
1098 		dev_info(adev->dev, "  SDMA%d_SEM_WAIT_FAIL_TIMER_CNTL=0x%08X\n",
1099 			 i, RREG32(mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i]));
1100 		dev_info(adev->dev, "  SDMA%d_GFX_IB_CNTL=0x%08X\n",
1101 			 i, RREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i]));
1102 		dev_info(adev->dev, "  SDMA%d_GFX_RB_CNTL=0x%08X\n",
1103 			 i, RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]));
1104 		dev_info(adev->dev, "  SDMA%d_GFX_RB_RPTR=0x%08X\n",
1105 			 i, RREG32(mmSDMA0_GFX_RB_RPTR + sdma_offsets[i]));
1106 		dev_info(adev->dev, "  SDMA%d_GFX_RB_WPTR=0x%08X\n",
1107 			 i, RREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i]));
1108 		dev_info(adev->dev, "  SDMA%d_GFX_RB_RPTR_ADDR_HI=0x%08X\n",
1109 			 i, RREG32(mmSDMA0_GFX_RB_RPTR_ADDR_HI + sdma_offsets[i]));
1110 		dev_info(adev->dev, "  SDMA%d_GFX_RB_RPTR_ADDR_LO=0x%08X\n",
1111 			 i, RREG32(mmSDMA0_GFX_RB_RPTR_ADDR_LO + sdma_offsets[i]));
1112 		dev_info(adev->dev, "  SDMA%d_GFX_RB_BASE=0x%08X\n",
1113 			 i, RREG32(mmSDMA0_GFX_RB_BASE + sdma_offsets[i]));
1114 		dev_info(adev->dev, "  SDMA%d_GFX_RB_BASE_HI=0x%08X\n",
1115 			 i, RREG32(mmSDMA0_GFX_RB_BASE_HI + sdma_offsets[i]));
1116 		dev_info(adev->dev, "  SDMA%d_TILING_CONFIG=0x%08X\n",
1117 			 i, RREG32(mmSDMA0_TILING_CONFIG + sdma_offsets[i]));
1118 		mutex_lock(&adev->srbm_mutex);
1119 		for (j = 0; j < 16; j++) {
1120 			vi_srbm_select(adev, 0, 0, 0, j);
1121 			dev_info(adev->dev, "  VM %d:\n", j);
1122 			dev_info(adev->dev, "  SDMA%d_GFX_VIRTUAL_ADDR=0x%08X\n",
1123 				 i, RREG32(mmSDMA0_GFX_VIRTUAL_ADDR + sdma_offsets[i]));
1124 			dev_info(adev->dev, "  SDMA%d_GFX_APE1_CNTL=0x%08X\n",
1125 				 i, RREG32(mmSDMA0_GFX_APE1_CNTL + sdma_offsets[i]));
1126 		}
1127 		vi_srbm_select(adev, 0, 0, 0, 0);
1128 		mutex_unlock(&adev->srbm_mutex);
1129 	}
1130 }
1131 
1132 static int sdma_v2_4_soft_reset(void *handle)
1133 {
1134 	u32 srbm_soft_reset = 0;
1135 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1136 	u32 tmp = RREG32(mmSRBM_STATUS2);
1137 
1138 	if (tmp & SRBM_STATUS2__SDMA_BUSY_MASK) {
1139 		/* sdma0 */
1140 		tmp = RREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET);
1141 		tmp = REG_SET_FIELD(tmp, SDMA0_F32_CNTL, HALT, 0);
1142 		WREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET, tmp);
1143 		srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA_MASK;
1144 	}
1145 	if (tmp & SRBM_STATUS2__SDMA1_BUSY_MASK) {
1146 		/* sdma1 */
1147 		tmp = RREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET);
1148 		tmp = REG_SET_FIELD(tmp, SDMA0_F32_CNTL, HALT, 0);
1149 		WREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET, tmp);
1150 		srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA1_MASK;
1151 	}
1152 
1153 	if (srbm_soft_reset) {
1154 		sdma_v2_4_print_status((void *)adev);
1155 
1156 		tmp = RREG32(mmSRBM_SOFT_RESET);
1157 		tmp |= srbm_soft_reset;
1158 		dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
1159 		WREG32(mmSRBM_SOFT_RESET, tmp);
1160 		tmp = RREG32(mmSRBM_SOFT_RESET);
1161 
1162 		udelay(50);
1163 
1164 		tmp &= ~srbm_soft_reset;
1165 		WREG32(mmSRBM_SOFT_RESET, tmp);
1166 		tmp = RREG32(mmSRBM_SOFT_RESET);
1167 
1168 		/* Wait a little for things to settle down */
1169 		udelay(50);
1170 
1171 		sdma_v2_4_print_status((void *)adev);
1172 	}
1173 
1174 	return 0;
1175 }
1176 
1177 static int sdma_v2_4_set_trap_irq_state(struct amdgpu_device *adev,
1178 					struct amdgpu_irq_src *src,
1179 					unsigned type,
1180 					enum amdgpu_interrupt_state state)
1181 {
1182 	u32 sdma_cntl;
1183 
1184 	switch (type) {
1185 	case AMDGPU_SDMA_IRQ_TRAP0:
1186 		switch (state) {
1187 		case AMDGPU_IRQ_STATE_DISABLE:
1188 			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1189 			sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 0);
1190 			WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1191 			break;
1192 		case AMDGPU_IRQ_STATE_ENABLE:
1193 			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1194 			sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 1);
1195 			WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1196 			break;
1197 		default:
1198 			break;
1199 		}
1200 		break;
1201 	case AMDGPU_SDMA_IRQ_TRAP1:
1202 		switch (state) {
1203 		case AMDGPU_IRQ_STATE_DISABLE:
1204 			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1205 			sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 0);
1206 			WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1207 			break;
1208 		case AMDGPU_IRQ_STATE_ENABLE:
1209 			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1210 			sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 1);
1211 			WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1212 			break;
1213 		default:
1214 			break;
1215 		}
1216 		break;
1217 	default:
1218 		break;
1219 	}
1220 	return 0;
1221 }
1222 
1223 static int sdma_v2_4_process_trap_irq(struct amdgpu_device *adev,
1224 				      struct amdgpu_irq_src *source,
1225 				      struct amdgpu_iv_entry *entry)
1226 {
1227 	u8 instance_id, queue_id;
1228 
1229 	instance_id = (entry->ring_id & 0x3) >> 0;
1230 	queue_id = (entry->ring_id & 0xc) >> 2;
1231 	DRM_DEBUG("IH: SDMA trap\n");
1232 	switch (instance_id) {
1233 	case 0:
1234 		switch (queue_id) {
1235 		case 0:
1236 			amdgpu_fence_process(&adev->sdma.instance[0].ring);
1237 			break;
1238 		case 1:
1239 			/* XXX compute */
1240 			break;
1241 		case 2:
1242 			/* XXX compute */
1243 			break;
1244 		}
1245 		break;
1246 	case 1:
1247 		switch (queue_id) {
1248 		case 0:
1249 			amdgpu_fence_process(&adev->sdma.instance[1].ring);
1250 			break;
1251 		case 1:
1252 			/* XXX compute */
1253 			break;
1254 		case 2:
1255 			/* XXX compute */
1256 			break;
1257 		}
1258 		break;
1259 	}
1260 	return 0;
1261 }
1262 
1263 static int sdma_v2_4_process_illegal_inst_irq(struct amdgpu_device *adev,
1264 					      struct amdgpu_irq_src *source,
1265 					      struct amdgpu_iv_entry *entry)
1266 {
1267 	DRM_ERROR("Illegal instruction in SDMA command stream\n");
1268 	schedule_work(&adev->reset_work);
1269 	return 0;
1270 }
1271 
1272 static int sdma_v2_4_set_clockgating_state(void *handle,
1273 					  enum amd_clockgating_state state)
1274 {
1275 	/* XXX handled via the smc on VI */
1276 	return 0;
1277 }
1278 
1279 static int sdma_v2_4_set_powergating_state(void *handle,
1280 					  enum amd_powergating_state state)
1281 {
1282 	return 0;
1283 }
1284 
1285 const struct amd_ip_funcs sdma_v2_4_ip_funcs = {
1286 	.early_init = sdma_v2_4_early_init,
1287 	.late_init = NULL,
1288 	.sw_init = sdma_v2_4_sw_init,
1289 	.sw_fini = sdma_v2_4_sw_fini,
1290 	.hw_init = sdma_v2_4_hw_init,
1291 	.hw_fini = sdma_v2_4_hw_fini,
1292 	.suspend = sdma_v2_4_suspend,
1293 	.resume = sdma_v2_4_resume,
1294 	.is_idle = sdma_v2_4_is_idle,
1295 	.wait_for_idle = sdma_v2_4_wait_for_idle,
1296 	.soft_reset = sdma_v2_4_soft_reset,
1297 	.print_status = sdma_v2_4_print_status,
1298 	.set_clockgating_state = sdma_v2_4_set_clockgating_state,
1299 	.set_powergating_state = sdma_v2_4_set_powergating_state,
1300 };
1301 
1302 static const struct amdgpu_ring_funcs sdma_v2_4_ring_funcs = {
1303 	.get_rptr = sdma_v2_4_ring_get_rptr,
1304 	.get_wptr = sdma_v2_4_ring_get_wptr,
1305 	.set_wptr = sdma_v2_4_ring_set_wptr,
1306 	.parse_cs = NULL,
1307 	.emit_ib = sdma_v2_4_ring_emit_ib,
1308 	.emit_fence = sdma_v2_4_ring_emit_fence,
1309 	.emit_pipeline_sync = sdma_v2_4_ring_emit_pipeline_sync,
1310 	.emit_vm_flush = sdma_v2_4_ring_emit_vm_flush,
1311 	.emit_hdp_flush = sdma_v2_4_ring_emit_hdp_flush,
1312 	.emit_hdp_invalidate = sdma_v2_4_ring_emit_hdp_invalidate,
1313 	.test_ring = sdma_v2_4_ring_test_ring,
1314 	.test_ib = sdma_v2_4_ring_test_ib,
1315 	.insert_nop = sdma_v2_4_ring_insert_nop,
1316 	.pad_ib = sdma_v2_4_ring_pad_ib,
1317 };
1318 
1319 static void sdma_v2_4_set_ring_funcs(struct amdgpu_device *adev)
1320 {
1321 	int i;
1322 
1323 	for (i = 0; i < adev->sdma.num_instances; i++)
1324 		adev->sdma.instance[i].ring.funcs = &sdma_v2_4_ring_funcs;
1325 }
1326 
1327 static const struct amdgpu_irq_src_funcs sdma_v2_4_trap_irq_funcs = {
1328 	.set = sdma_v2_4_set_trap_irq_state,
1329 	.process = sdma_v2_4_process_trap_irq,
1330 };
1331 
1332 static const struct amdgpu_irq_src_funcs sdma_v2_4_illegal_inst_irq_funcs = {
1333 	.process = sdma_v2_4_process_illegal_inst_irq,
1334 };
1335 
1336 static void sdma_v2_4_set_irq_funcs(struct amdgpu_device *adev)
1337 {
1338 	adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST;
1339 	adev->sdma.trap_irq.funcs = &sdma_v2_4_trap_irq_funcs;
1340 	adev->sdma.illegal_inst_irq.funcs = &sdma_v2_4_illegal_inst_irq_funcs;
1341 }
1342 
1343 /**
1344  * sdma_v2_4_emit_copy_buffer - copy buffer using the sDMA engine
1345  *
1346  * @ring: amdgpu_ring structure holding ring information
1347  * @src_offset: src GPU address
1348  * @dst_offset: dst GPU address
1349  * @byte_count: number of bytes to xfer
1350  *
1351  * Copy GPU buffers using the DMA engine (VI).
1352  * Used by the amdgpu ttm implementation to move pages if
1353  * registered as the asic copy callback.
1354  */
1355 static void sdma_v2_4_emit_copy_buffer(struct amdgpu_ib *ib,
1356 				       uint64_t src_offset,
1357 				       uint64_t dst_offset,
1358 				       uint32_t byte_count)
1359 {
1360 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1361 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
1362 	ib->ptr[ib->length_dw++] = byte_count;
1363 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1364 	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1365 	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1366 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1367 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1368 }
1369 
1370 /**
1371  * sdma_v2_4_emit_fill_buffer - fill buffer using the sDMA engine
1372  *
1373  * @ring: amdgpu_ring structure holding ring information
1374  * @src_data: value to write to buffer
1375  * @dst_offset: dst GPU address
1376  * @byte_count: number of bytes to xfer
1377  *
1378  * Fill GPU buffers using the DMA engine (VI).
1379  */
1380 static void sdma_v2_4_emit_fill_buffer(struct amdgpu_ib *ib,
1381 				       uint32_t src_data,
1382 				       uint64_t dst_offset,
1383 				       uint32_t byte_count)
1384 {
1385 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL);
1386 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1387 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1388 	ib->ptr[ib->length_dw++] = src_data;
1389 	ib->ptr[ib->length_dw++] = byte_count;
1390 }
1391 
1392 static const struct amdgpu_buffer_funcs sdma_v2_4_buffer_funcs = {
1393 	.copy_max_bytes = 0x1fffff,
1394 	.copy_num_dw = 7,
1395 	.emit_copy_buffer = sdma_v2_4_emit_copy_buffer,
1396 
1397 	.fill_max_bytes = 0x1fffff,
1398 	.fill_num_dw = 7,
1399 	.emit_fill_buffer = sdma_v2_4_emit_fill_buffer,
1400 };
1401 
1402 static void sdma_v2_4_set_buffer_funcs(struct amdgpu_device *adev)
1403 {
1404 	if (adev->mman.buffer_funcs == NULL) {
1405 		adev->mman.buffer_funcs = &sdma_v2_4_buffer_funcs;
1406 		adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1407 	}
1408 }
1409 
1410 static const struct amdgpu_vm_pte_funcs sdma_v2_4_vm_pte_funcs = {
1411 	.copy_pte = sdma_v2_4_vm_copy_pte,
1412 	.write_pte = sdma_v2_4_vm_write_pte,
1413 	.set_pte_pde = sdma_v2_4_vm_set_pte_pde,
1414 };
1415 
1416 static void sdma_v2_4_set_vm_pte_funcs(struct amdgpu_device *adev)
1417 {
1418 	unsigned i;
1419 
1420 	if (adev->vm_manager.vm_pte_funcs == NULL) {
1421 		adev->vm_manager.vm_pte_funcs = &sdma_v2_4_vm_pte_funcs;
1422 		for (i = 0; i < adev->sdma.num_instances; i++)
1423 			adev->vm_manager.vm_pte_rings[i] =
1424 				&adev->sdma.instance[i].ring;
1425 
1426 		adev->vm_manager.vm_pte_num_rings = adev->sdma.num_instances;
1427 	}
1428 }
1429