1 /*
2 * Copyright 2019 Advanced Micro Devices, Inc.
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
13 *
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
21 *
22 */
23
24 #include <linux/delay.h>
25 #include <linux/firmware.h>
26 #include <linux/module.h>
27 #include <linux/pci.h>
28
29 #include "amdgpu.h"
30 #include "amdgpu_ucode.h"
31 #include "amdgpu_trace.h"
32
33 #include "gc/gc_10_3_0_offset.h"
34 #include "gc/gc_10_3_0_sh_mask.h"
35 #include "ivsrcid/sdma0/irqsrcs_sdma0_5_0.h"
36 #include "ivsrcid/sdma1/irqsrcs_sdma1_5_0.h"
37 #include "ivsrcid/sdma2/irqsrcs_sdma2_5_0.h"
38 #include "ivsrcid/sdma3/irqsrcs_sdma3_5_0.h"
39
40 #include "soc15_common.h"
41 #include "soc15.h"
42 #include "navi10_sdma_pkt_open.h"
43 #include "nbio_v2_3.h"
44 #include "sdma_common.h"
45 #include "sdma_v5_2.h"
46
47 MODULE_FIRMWARE("amdgpu/sienna_cichlid_sdma.bin");
48 MODULE_FIRMWARE("amdgpu/navy_flounder_sdma.bin");
49 MODULE_FIRMWARE("amdgpu/dimgrey_cavefish_sdma.bin");
50 MODULE_FIRMWARE("amdgpu/beige_goby_sdma.bin");
51
52 MODULE_FIRMWARE("amdgpu/vangogh_sdma.bin");
53 MODULE_FIRMWARE("amdgpu/yellow_carp_sdma.bin");
54 MODULE_FIRMWARE("amdgpu/sdma_5_2_6.bin");
55 MODULE_FIRMWARE("amdgpu/sdma_5_2_7.bin");
56
57 #define SDMA1_REG_OFFSET 0x600
58 #define SDMA3_REG_OFFSET 0x400
59 #define SDMA0_HYP_DEC_REG_START 0x5880
60 #define SDMA0_HYP_DEC_REG_END 0x5893
61 #define SDMA1_HYP_DEC_REG_OFFSET 0x20
62
63 static const struct amdgpu_hwip_reg_entry sdma_reg_list_5_2[] = {
64 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_STATUS_REG),
65 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_STATUS1_REG),
66 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_STATUS2_REG),
67 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_STATUS3_REG),
68 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_UCODE_CHECKSUM),
69 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RB_RPTR_FETCH_HI),
70 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RB_RPTR_FETCH),
71 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_UTCL1_RD_STATUS),
72 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_UTCL1_WR_STATUS),
73 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_UTCL1_RD_XNACK0),
74 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_UTCL1_RD_XNACK1),
75 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_UTCL1_WR_XNACK0),
76 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_UTCL1_WR_XNACK1),
77 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_RB_CNTL),
78 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_RB_RPTR),
79 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_RB_RPTR_HI),
80 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_RB_WPTR),
81 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_RB_WPTR_HI),
82 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_IB_OFFSET),
83 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_IB_BASE_LO),
84 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_IB_BASE_HI),
85 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_IB_CNTL),
86 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_IB_RPTR),
87 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_IB_SUB_REMAIN),
88 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_DUMMY_REG),
89 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_RB_CNTL),
90 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_RB_RPTR),
91 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_RB_RPTR_HI),
92 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_RB_WPTR),
93 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_RB_WPTR_HI),
94 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_IB_OFFSET),
95 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_IB_BASE_LO),
96 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_IB_BASE_HI),
97 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_DUMMY_REG),
98 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_RB_CNTL),
99 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_RB_RPTR),
100 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_RB_RPTR_HI),
101 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_RB_WPTR),
102 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_RB_WPTR_HI),
103 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_IB_OFFSET),
104 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_IB_BASE_LO),
105 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_IB_BASE_HI),
106 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_DUMMY_REG),
107 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_INT_STATUS),
108 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_VM_CNTL),
109 SOC15_REG_ENTRY_STR(GC, 0, mmGRBM_STATUS2)
110 };
111
112 static void sdma_v5_2_set_ring_funcs(struct amdgpu_device *adev);
113 static void sdma_v5_2_set_buffer_funcs(struct amdgpu_device *adev);
114 static void sdma_v5_2_set_vm_pte_funcs(struct amdgpu_device *adev);
115 static void sdma_v5_2_set_irq_funcs(struct amdgpu_device *adev);
116 static int sdma_v5_2_stop_queue(struct amdgpu_ring *ring);
117 static int sdma_v5_2_restore_queue(struct amdgpu_ring *ring);
118
sdma_v5_2_get_reg_offset(struct amdgpu_device * adev,u32 instance,u32 internal_offset)119 static u32 sdma_v5_2_get_reg_offset(struct amdgpu_device *adev, u32 instance, u32 internal_offset)
120 {
121 u32 base;
122
123 if (internal_offset >= SDMA0_HYP_DEC_REG_START &&
124 internal_offset <= SDMA0_HYP_DEC_REG_END) {
125 base = adev->reg_offset[GC_HWIP][0][1];
126 if (instance != 0)
127 internal_offset += SDMA1_HYP_DEC_REG_OFFSET * instance;
128 } else {
129 if (instance < 2) {
130 base = adev->reg_offset[GC_HWIP][0][0];
131 if (instance == 1)
132 internal_offset += SDMA1_REG_OFFSET;
133 } else {
134 base = adev->reg_offset[GC_HWIP][0][2];
135 if (instance == 3)
136 internal_offset += SDMA3_REG_OFFSET;
137 }
138 }
139
140 return base + internal_offset;
141 }
142
sdma_v5_2_ring_init_cond_exec(struct amdgpu_ring * ring,uint64_t addr)143 static unsigned sdma_v5_2_ring_init_cond_exec(struct amdgpu_ring *ring,
144 uint64_t addr)
145 {
146 unsigned ret;
147
148 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_COND_EXE));
149 amdgpu_ring_write(ring, lower_32_bits(addr));
150 amdgpu_ring_write(ring, upper_32_bits(addr));
151 amdgpu_ring_write(ring, 1);
152 /* this is the offset we need patch later */
153 ret = ring->wptr & ring->buf_mask;
154 /* insert dummy here and patch it later */
155 amdgpu_ring_write(ring, 0);
156
157 return ret;
158 }
159
160 /**
161 * sdma_v5_2_ring_get_rptr - get the current read pointer
162 *
163 * @ring: amdgpu ring pointer
164 *
165 * Get the current rptr from the hardware (NAVI10+).
166 */
sdma_v5_2_ring_get_rptr(struct amdgpu_ring * ring)167 static uint64_t sdma_v5_2_ring_get_rptr(struct amdgpu_ring *ring)
168 {
169 u64 *rptr;
170
171 /* XXX check if swapping is necessary on BE */
172 rptr = (u64 *)ring->rptr_cpu_addr;
173
174 DRM_DEBUG("rptr before shift == 0x%016llx\n", *rptr);
175 return ((*rptr) >> 2);
176 }
177
178 /**
179 * sdma_v5_2_ring_get_wptr - get the current write pointer
180 *
181 * @ring: amdgpu ring pointer
182 *
183 * Get the current wptr from the hardware (NAVI10+).
184 */
sdma_v5_2_ring_get_wptr(struct amdgpu_ring * ring)185 static uint64_t sdma_v5_2_ring_get_wptr(struct amdgpu_ring *ring)
186 {
187 struct amdgpu_device *adev = ring->adev;
188 u64 wptr;
189
190 if (ring->use_doorbell) {
191 /* XXX check if swapping is necessary on BE */
192 wptr = READ_ONCE(*((u64 *)ring->wptr_cpu_addr));
193 DRM_DEBUG("wptr/doorbell before shift == 0x%016llx\n", wptr);
194 } else {
195 wptr = RREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI));
196 wptr = wptr << 32;
197 wptr |= RREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR));
198 DRM_DEBUG("wptr before shift [%i] wptr == 0x%016llx\n", ring->me, wptr);
199 }
200
201 return wptr >> 2;
202 }
203
204 /**
205 * sdma_v5_2_ring_set_wptr - commit the write pointer
206 *
207 * @ring: amdgpu ring pointer
208 *
209 * Write the wptr back to the hardware (NAVI10+).
210 */
sdma_v5_2_ring_set_wptr(struct amdgpu_ring * ring)211 static void sdma_v5_2_ring_set_wptr(struct amdgpu_ring *ring)
212 {
213 struct amdgpu_device *adev = ring->adev;
214
215 DRM_DEBUG("Setting write pointer\n");
216 if (ring->use_doorbell) {
217 DRM_DEBUG("Using doorbell -- "
218 "wptr_offs == 0x%08x "
219 "lower_32_bits(ring->wptr << 2) == 0x%08x "
220 "upper_32_bits(ring->wptr << 2) == 0x%08x\n",
221 ring->wptr_offs,
222 lower_32_bits(ring->wptr << 2),
223 upper_32_bits(ring->wptr << 2));
224 /* XXX check if swapping is necessary on BE */
225 atomic64_set((atomic64_t *)ring->wptr_cpu_addr,
226 ring->wptr << 2);
227 DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016llx)\n",
228 ring->doorbell_index, ring->wptr << 2);
229 WDOORBELL64(ring->doorbell_index, ring->wptr << 2);
230 if (amdgpu_ip_version(adev, SDMA0_HWIP, 0) == IP_VERSION(5, 2, 1)) {
231 /* SDMA seems to miss doorbells sometimes when powergating kicks in.
232 * Updating the wptr directly will wake it. This is only safe because
233 * we disallow gfxoff in begin_use() and then allow it again in end_use().
234 */
235 WREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR),
236 lower_32_bits(ring->wptr << 2));
237 WREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI),
238 upper_32_bits(ring->wptr << 2));
239 }
240 } else {
241 DRM_DEBUG("Not using doorbell -- "
242 "mmSDMA%i_GFX_RB_WPTR == 0x%08x "
243 "mmSDMA%i_GFX_RB_WPTR_HI == 0x%08x\n",
244 ring->me,
245 lower_32_bits(ring->wptr << 2),
246 ring->me,
247 upper_32_bits(ring->wptr << 2));
248 WREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR),
249 lower_32_bits(ring->wptr << 2));
250 WREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI),
251 upper_32_bits(ring->wptr << 2));
252 }
253 }
254
sdma_v5_2_ring_insert_nop(struct amdgpu_ring * ring,uint32_t count)255 static void sdma_v5_2_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
256 {
257 struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
258 int i;
259
260 for (i = 0; i < count; i++)
261 if (sdma && sdma->burst_nop && (i == 0))
262 amdgpu_ring_write(ring, ring->funcs->nop |
263 SDMA_PKT_NOP_HEADER_COUNT(count - 1));
264 else
265 amdgpu_ring_write(ring, ring->funcs->nop);
266 }
267
268 /**
269 * sdma_v5_2_ring_emit_ib - Schedule an IB on the DMA engine
270 *
271 * @ring: amdgpu ring pointer
272 * @job: job to retrieve vmid from
273 * @ib: IB object to schedule
274 * @flags: unused
275 *
276 * Schedule an IB in the DMA ring.
277 */
sdma_v5_2_ring_emit_ib(struct amdgpu_ring * ring,struct amdgpu_job * job,struct amdgpu_ib * ib,uint32_t flags)278 static void sdma_v5_2_ring_emit_ib(struct amdgpu_ring *ring,
279 struct amdgpu_job *job,
280 struct amdgpu_ib *ib,
281 uint32_t flags)
282 {
283 unsigned vmid = AMDGPU_JOB_GET_VMID(job);
284 uint64_t csa_mc_addr = amdgpu_sdma_get_csa_mc_addr(ring, vmid);
285
286 /* An IB packet must end on a 8 DW boundary--the next dword
287 * must be on a 8-dword boundary. Our IB packet below is 6
288 * dwords long, thus add x number of NOPs, such that, in
289 * modular arithmetic,
290 * wptr + 6 + x = 8k, k >= 0, which in C is,
291 * (wptr + 6 + x) % 8 = 0.
292 * The expression below, is a solution of x.
293 */
294 sdma_v5_2_ring_insert_nop(ring, (2 - lower_32_bits(ring->wptr)) & 7);
295
296 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) |
297 SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf));
298 /* base must be 32 byte aligned */
299 amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
300 amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
301 amdgpu_ring_write(ring, ib->length_dw);
302 amdgpu_ring_write(ring, lower_32_bits(csa_mc_addr));
303 amdgpu_ring_write(ring, upper_32_bits(csa_mc_addr));
304 }
305
306 /**
307 * sdma_v5_2_ring_emit_mem_sync - flush the IB by graphics cache rinse
308 *
309 * @ring: amdgpu ring pointer
310 *
311 * flush the IB by graphics cache rinse.
312 */
sdma_v5_2_ring_emit_mem_sync(struct amdgpu_ring * ring)313 static void sdma_v5_2_ring_emit_mem_sync(struct amdgpu_ring *ring)
314 {
315 uint32_t gcr_cntl = SDMA_GCR_GL2_INV | SDMA_GCR_GL2_WB |
316 SDMA_GCR_GLM_INV | SDMA_GCR_GL1_INV |
317 SDMA_GCR_GLV_INV | SDMA_GCR_GLK_INV |
318 SDMA_GCR_GLI_INV(1);
319
320 /* flush entire cache L0/L1/L2, this can be optimized by performance requirement */
321 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_GCR_REQ));
322 amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD1_BASE_VA_31_7(0));
323 amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD2_GCR_CONTROL_15_0(gcr_cntl) |
324 SDMA_PKT_GCR_REQ_PAYLOAD2_BASE_VA_47_32(0));
325 amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD3_LIMIT_VA_31_7(0) |
326 SDMA_PKT_GCR_REQ_PAYLOAD3_GCR_CONTROL_18_16(gcr_cntl >> 16));
327 amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD4_LIMIT_VA_47_32(0) |
328 SDMA_PKT_GCR_REQ_PAYLOAD4_VMID(0));
329 }
330
331 /**
332 * sdma_v5_2_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
333 *
334 * @ring: amdgpu ring pointer
335 *
336 * Emit an hdp flush packet on the requested DMA ring.
337 */
sdma_v5_2_ring_emit_hdp_flush(struct amdgpu_ring * ring)338 static void sdma_v5_2_ring_emit_hdp_flush(struct amdgpu_ring *ring)
339 {
340 struct amdgpu_device *adev = ring->adev;
341 u32 ref_and_mask = 0;
342 const struct nbio_hdp_flush_reg *nbio_hf_reg = adev->nbio.hdp_flush_reg;
343
344 if (ring->me > 1) {
345 amdgpu_asic_flush_hdp(adev, ring);
346 } else {
347 ref_and_mask = nbio_hf_reg->ref_and_mask_sdma0 << ring->me;
348
349 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
350 SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) |
351 SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
352 amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_done_offset(adev)) << 2);
353 amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_req_offset(adev)) << 2);
354 amdgpu_ring_write(ring, ref_and_mask); /* reference */
355 amdgpu_ring_write(ring, ref_and_mask); /* mask */
356 amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
357 SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
358 }
359 }
360
361 /**
362 * sdma_v5_2_ring_emit_fence - emit a fence on the DMA ring
363 *
364 * @ring: amdgpu ring pointer
365 * @addr: address
366 * @seq: sequence number
367 * @flags: fence related flags
368 *
369 * Add a DMA fence packet to the ring to write
370 * the fence seq number and DMA trap packet to generate
371 * an interrupt if needed.
372 */
sdma_v5_2_ring_emit_fence(struct amdgpu_ring * ring,u64 addr,u64 seq,unsigned flags)373 static void sdma_v5_2_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
374 unsigned flags)
375 {
376 bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
377 /* write the fence */
378 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) |
379 SDMA_PKT_FENCE_HEADER_MTYPE(0x3)); /* Ucached(UC) */
380 /* zero in first two bits */
381 BUG_ON(addr & 0x3);
382 amdgpu_ring_write(ring, lower_32_bits(addr));
383 amdgpu_ring_write(ring, upper_32_bits(addr));
384 amdgpu_ring_write(ring, lower_32_bits(seq));
385
386 /* optionally write high bits as well */
387 if (write64bit) {
388 addr += 4;
389 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) |
390 SDMA_PKT_FENCE_HEADER_MTYPE(0x3));
391 /* zero in first two bits */
392 BUG_ON(addr & 0x3);
393 amdgpu_ring_write(ring, lower_32_bits(addr));
394 amdgpu_ring_write(ring, upper_32_bits(addr));
395 amdgpu_ring_write(ring, upper_32_bits(seq));
396 }
397
398 if ((flags & AMDGPU_FENCE_FLAG_INT)) {
399 /* generate an interrupt */
400 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP));
401 amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(0));
402 }
403 }
404
405
406 /**
407 * sdma_v5_2_gfx_stop - stop the gfx async dma engines
408 *
409 * @adev: amdgpu_device pointer
410 * @inst_mask: mask of dma engine instances to be disabled
411 * Stop the gfx async dma ring buffers.
412 */
sdma_v5_2_gfx_stop(struct amdgpu_device * adev,uint32_t inst_mask)413 static void sdma_v5_2_gfx_stop(struct amdgpu_device *adev, uint32_t inst_mask)
414 {
415 u32 rb_cntl, ib_cntl;
416 int i;
417
418 for_each_inst(i, inst_mask) {
419 rb_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL));
420 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 0);
421 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
422 ib_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL));
423 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 0);
424 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl);
425 }
426 }
427
428 /**
429 * sdma_v5_2_rlc_stop - stop the compute async dma engines
430 *
431 * @adev: amdgpu_device pointer
432 *
433 * Stop the compute async dma queues.
434 */
sdma_v5_2_rlc_stop(struct amdgpu_device * adev)435 static void sdma_v5_2_rlc_stop(struct amdgpu_device *adev)
436 {
437 /* XXX todo */
438 }
439
440 /**
441 * sdma_v5_2_ctx_switch_enable - stop the async dma engines context switch
442 *
443 * @adev: amdgpu_device pointer
444 * @enable: enable/disable the DMA MEs context switch.
445 *
446 * Halt or unhalt the async dma engines context switch.
447 */
sdma_v5_2_ctx_switch_enable(struct amdgpu_device * adev,bool enable)448 static void sdma_v5_2_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
449 {
450 u32 f32_cntl, phase_quantum = 0;
451 int i;
452
453 if (amdgpu_sdma_phase_quantum) {
454 unsigned value = amdgpu_sdma_phase_quantum;
455 unsigned unit = 0;
456
457 while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
458 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) {
459 value = (value + 1) >> 1;
460 unit++;
461 }
462 if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
463 SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) {
464 value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
465 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT);
466 unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
467 SDMA0_PHASE0_QUANTUM__UNIT__SHIFT);
468 WARN_ONCE(1,
469 "clamping sdma_phase_quantum to %uK clock cycles\n",
470 value << unit);
471 }
472 phase_quantum =
473 value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
474 unit << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
475 }
476
477 for (i = 0; i < adev->sdma.num_instances; i++) {
478 if (enable && amdgpu_sdma_phase_quantum) {
479 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE0_QUANTUM),
480 phase_quantum);
481 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE1_QUANTUM),
482 phase_quantum);
483 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE2_QUANTUM),
484 phase_quantum);
485 }
486
487 if (!amdgpu_sriov_vf(adev)) {
488 f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
489 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
490 AUTO_CTXSW_ENABLE, enable ? 1 : 0);
491 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl);
492 }
493 }
494
495 }
496
497 /**
498 * sdma_v5_2_enable - stop the async dma engines
499 *
500 * @adev: amdgpu_device pointer
501 * @enable: enable/disable the DMA MEs.
502 *
503 * Halt or unhalt the async dma engines.
504 */
sdma_v5_2_enable(struct amdgpu_device * adev,bool enable)505 static void sdma_v5_2_enable(struct amdgpu_device *adev, bool enable)
506 {
507 u32 f32_cntl;
508 int i;
509 uint32_t inst_mask;
510
511 inst_mask = GENMASK(adev->sdma.num_instances - 1, 0);
512 if (!enable) {
513 sdma_v5_2_gfx_stop(adev, inst_mask);
514 sdma_v5_2_rlc_stop(adev);
515 }
516
517 if (!amdgpu_sriov_vf(adev)) {
518 for (i = 0; i < adev->sdma.num_instances; i++) {
519 f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
520 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, enable ? 0 : 1);
521 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl);
522 }
523 }
524 }
525
526 /**
527 * sdma_v5_2_gfx_resume_instance - start/restart a certain sdma engine
528 *
529 * @adev: amdgpu_device pointer
530 * @i: instance
531 * @restore: used to restore wptr when restart
532 *
533 * Set up the gfx DMA ring buffers and enable them. On restart, we will restore wptr and rptr.
534 * Return 0 for success.
535 */
536
sdma_v5_2_gfx_resume_instance(struct amdgpu_device * adev,int i,bool restore)537 static int sdma_v5_2_gfx_resume_instance(struct amdgpu_device *adev, int i, bool restore)
538 {
539 struct amdgpu_ring *ring;
540 u32 rb_cntl, ib_cntl;
541 u32 rb_bufsz;
542 u32 doorbell;
543 u32 doorbell_offset;
544 u32 temp;
545 u32 wptr_poll_cntl;
546 u64 wptr_gpu_addr;
547
548 ring = &adev->sdma.instance[i].ring;
549
550 if (!amdgpu_sriov_vf(adev))
551 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL), 0);
552
553 /* Set ring buffer size in dwords */
554 rb_bufsz = order_base_2(ring->ring_size / 4);
555 rb_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL));
556 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz);
557 #ifdef __BIG_ENDIAN
558 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1);
559 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL,
560 RPTR_WRITEBACK_SWAP_ENABLE, 1);
561 #endif
562 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
563
564 /* Initialize the ring buffer's read and write pointers */
565 if (restore) {
566 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR), lower_32_bits(ring->wptr << 2));
567 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_HI), upper_32_bits(ring->wptr << 2));
568 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), lower_32_bits(ring->wptr << 2));
569 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), upper_32_bits(ring->wptr << 2));
570 } else {
571 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR), 0);
572 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_HI), 0);
573 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), 0);
574 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), 0);
575 }
576
577 /* setup the wptr shadow polling */
578 wptr_gpu_addr = ring->wptr_gpu_addr;
579 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_LO),
580 lower_32_bits(wptr_gpu_addr));
581 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_HI),
582 upper_32_bits(wptr_gpu_addr));
583 wptr_poll_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i,
584 mmSDMA0_GFX_RB_WPTR_POLL_CNTL));
585 wptr_poll_cntl = REG_SET_FIELD(wptr_poll_cntl,
586 SDMA0_GFX_RB_WPTR_POLL_CNTL,
587 F32_POLL_ENABLE, 1);
588 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_CNTL),
589 wptr_poll_cntl);
590
591 /* set the wb address whether it's enabled or not */
592 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_HI),
593 upper_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFF);
594 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_LO),
595 lower_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFC);
596
597 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
598
599 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE), ring->gpu_addr >> 8);
600 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE_HI), ring->gpu_addr >> 40);
601
602 if (!restore)
603 ring->wptr = 0;
604
605 /* before programing wptr to a less value, need set minor_ptr_update first */
606 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 1);
607
608 if (!amdgpu_sriov_vf(adev)) { /* only bare-metal use register write for wptr */
609 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), lower_32_bits(ring->wptr << 2));
610 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), upper_32_bits(ring->wptr << 2));
611 }
612
613 doorbell = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL));
614 doorbell_offset = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL_OFFSET));
615
616 if (ring->use_doorbell) {
617 doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 1);
618 doorbell_offset = REG_SET_FIELD(doorbell_offset, SDMA0_GFX_DOORBELL_OFFSET,
619 OFFSET, ring->doorbell_index);
620 } else {
621 doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 0);
622 }
623 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL), doorbell);
624 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL_OFFSET), doorbell_offset);
625
626 adev->nbio.funcs->sdma_doorbell_range(adev, i, ring->use_doorbell,
627 ring->doorbell_index,
628 adev->doorbell_index.sdma_doorbell_range);
629
630 if (amdgpu_sriov_vf(adev))
631 sdma_v5_2_ring_set_wptr(ring);
632
633 /* set minor_ptr_update to 0 after wptr programed */
634
635 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 0);
636
637 /* SRIOV VF has no control of any of registers below */
638 if (!amdgpu_sriov_vf(adev)) {
639 /* set utc l1 enable flag always to 1 */
640 temp = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
641 temp = REG_SET_FIELD(temp, SDMA0_CNTL, UTC_L1_ENABLE, 1);
642
643 /* enable MCBP */
644 temp = REG_SET_FIELD(temp, SDMA0_CNTL, MIDCMD_PREEMPT_ENABLE, 1);
645 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), temp);
646
647 /* Set up RESP_MODE to non-copy addresses */
648 temp = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL));
649 temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, RESP_MODE, 3);
650 temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, REDO_DELAY, 9);
651 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL), temp);
652
653 /* program default cache read and write policy */
654 temp = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE));
655 /* clean read policy and write policy bits */
656 temp &= 0xFF0FFF;
657 temp |= ((CACHE_READ_POLICY_L2__DEFAULT << 12) |
658 (CACHE_WRITE_POLICY_L2__DEFAULT << 14) |
659 SDMA0_UTCL1_PAGE__LLC_NOALLOC_MASK);
660 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE), temp);
661
662 /* unhalt engine */
663 temp = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
664 temp = REG_SET_FIELD(temp, SDMA0_F32_CNTL, HALT, 0);
665 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), temp);
666 }
667
668 /* enable DMA RB */
669 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1);
670 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
671
672 ib_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL));
673 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 1);
674 #ifdef __BIG_ENDIAN
675 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1);
676 #endif
677 /* enable DMA IBs */
678 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl);
679
680 if (amdgpu_sriov_vf(adev)) { /* bare-metal sequence doesn't need below to lines */
681 sdma_v5_2_ctx_switch_enable(adev, true);
682 sdma_v5_2_enable(adev, true);
683 }
684
685 return amdgpu_ring_test_helper(ring);
686 }
687
688 /**
689 * sdma_v5_2_gfx_resume - setup and start the async dma engines
690 *
691 * @adev: amdgpu_device pointer
692 *
693 * Set up the gfx DMA ring buffers and enable them.
694 * Returns 0 for success, error for failure.
695 */
sdma_v5_2_gfx_resume(struct amdgpu_device * adev)696 static int sdma_v5_2_gfx_resume(struct amdgpu_device *adev)
697 {
698 int i, r;
699
700 for (i = 0; i < adev->sdma.num_instances; i++) {
701 r = sdma_v5_2_gfx_resume_instance(adev, i, false);
702 if (r)
703 return r;
704 }
705
706 return 0;
707 }
708
709 /**
710 * sdma_v5_2_rlc_resume - setup and start the async dma engines
711 *
712 * @adev: amdgpu_device pointer
713 *
714 * Set up the compute DMA queues and enable them.
715 * Returns 0 for success, error for failure.
716 */
sdma_v5_2_rlc_resume(struct amdgpu_device * adev)717 static int sdma_v5_2_rlc_resume(struct amdgpu_device *adev)
718 {
719 return 0;
720 }
721
722 /**
723 * sdma_v5_2_load_microcode - load the sDMA ME ucode
724 *
725 * @adev: amdgpu_device pointer
726 *
727 * Loads the sDMA0/1/2/3 ucode.
728 * Returns 0 for success, -EINVAL if the ucode is not available.
729 */
sdma_v5_2_load_microcode(struct amdgpu_device * adev)730 static int sdma_v5_2_load_microcode(struct amdgpu_device *adev)
731 {
732 const struct sdma_firmware_header_v1_0 *hdr;
733 const __le32 *fw_data;
734 u32 fw_size;
735 int i, j;
736
737 /* halt the MEs */
738 sdma_v5_2_enable(adev, false);
739
740 for (i = 0; i < adev->sdma.num_instances; i++) {
741 if (!adev->sdma.instance[i].fw)
742 return -EINVAL;
743
744 hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
745 amdgpu_ucode_print_sdma_hdr(&hdr->header);
746 fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
747
748 fw_data = (const __le32 *)
749 (adev->sdma.instance[i].fw->data +
750 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
751
752 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), 0);
753
754 for (j = 0; j < fw_size; j++) {
755 if (amdgpu_emu_mode == 1 && j % 500 == 0)
756 msleep(1);
757 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UCODE_DATA), le32_to_cpup(fw_data++));
758 }
759
760 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), adev->sdma.instance[i].fw_version);
761 }
762
763 return 0;
764 }
765
sdma_v5_2_soft_reset_engine(struct amdgpu_device * adev,u32 instance_id)766 static int sdma_v5_2_soft_reset_engine(struct amdgpu_device *adev, u32 instance_id)
767 {
768 u32 grbm_soft_reset;
769 u32 tmp;
770
771 grbm_soft_reset = REG_SET_FIELD(0,
772 GRBM_SOFT_RESET, SOFT_RESET_SDMA0,
773 1);
774 grbm_soft_reset <<= instance_id;
775
776 tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET);
777 tmp |= grbm_soft_reset;
778 DRM_DEBUG("GRBM_SOFT_RESET=0x%08X\n", tmp);
779 WREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET, tmp);
780 tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET);
781
782 udelay(50);
783
784 tmp &= ~grbm_soft_reset;
785 WREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET, tmp);
786 tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET);
787 return 0;
788 }
789
sdma_v5_2_soft_reset(struct amdgpu_ip_block * ip_block)790 static int sdma_v5_2_soft_reset(struct amdgpu_ip_block *ip_block)
791 {
792 struct amdgpu_device *adev = ip_block->adev;
793 int i;
794
795 for (i = 0; i < adev->sdma.num_instances; i++) {
796 sdma_v5_2_soft_reset_engine(adev, i);
797 udelay(50);
798 }
799
800 return 0;
801 }
802
803 static const struct amdgpu_sdma_funcs sdma_v5_2_sdma_funcs = {
804 .stop_kernel_queue = &sdma_v5_2_stop_queue,
805 .start_kernel_queue = &sdma_v5_2_restore_queue,
806 .soft_reset_kernel_queue = &sdma_v5_2_soft_reset_engine,
807 };
808
809 /**
810 * sdma_v5_2_start - setup and start the async dma engines
811 *
812 * @adev: amdgpu_device pointer
813 *
814 * Set up the DMA engines and enable them.
815 * Returns 0 for success, error for failure.
816 */
sdma_v5_2_start(struct amdgpu_device * adev)817 static int sdma_v5_2_start(struct amdgpu_device *adev)
818 {
819 int r = 0;
820 struct amdgpu_ip_block *ip_block;
821
822 if (amdgpu_sriov_vf(adev)) {
823 sdma_v5_2_ctx_switch_enable(adev, false);
824 sdma_v5_2_enable(adev, false);
825
826 /* set RB registers */
827 r = sdma_v5_2_gfx_resume(adev);
828 return r;
829 }
830
831 if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
832 r = sdma_v5_2_load_microcode(adev);
833 if (r)
834 return r;
835
836 /* The value of mmSDMA_F32_CNTL is invalid the moment after loading fw */
837 if (amdgpu_emu_mode == 1)
838 msleep(1000);
839 }
840
841 ip_block = amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_SDMA);
842 if (!ip_block)
843 return -EINVAL;
844
845 sdma_v5_2_soft_reset(ip_block);
846 /* unhalt the MEs */
847 sdma_v5_2_enable(adev, true);
848 /* enable sdma ring preemption */
849 sdma_v5_2_ctx_switch_enable(adev, true);
850
851 /* start the gfx rings and rlc compute queues */
852 r = sdma_v5_2_gfx_resume(adev);
853 if (r)
854 return r;
855 r = sdma_v5_2_rlc_resume(adev);
856
857 return r;
858 }
859
sdma_v5_2_mqd_init(struct amdgpu_device * adev,void * mqd,struct amdgpu_mqd_prop * prop)860 static int sdma_v5_2_mqd_init(struct amdgpu_device *adev, void *mqd,
861 struct amdgpu_mqd_prop *prop)
862 {
863 struct v10_sdma_mqd *m = mqd;
864 uint64_t wb_gpu_addr;
865
866 m->sdmax_rlcx_rb_cntl =
867 order_base_2(prop->queue_size / 4) << SDMA0_RLC0_RB_CNTL__RB_SIZE__SHIFT |
868 1 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_ENABLE__SHIFT |
869 6 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_TIMER__SHIFT |
870 1 << SDMA0_RLC0_RB_CNTL__RB_PRIV__SHIFT;
871
872 m->sdmax_rlcx_rb_base = lower_32_bits(prop->hqd_base_gpu_addr >> 8);
873 m->sdmax_rlcx_rb_base_hi = upper_32_bits(prop->hqd_base_gpu_addr >> 8);
874
875 m->sdmax_rlcx_rb_wptr_poll_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, 0,
876 mmSDMA0_GFX_RB_WPTR_POLL_CNTL));
877
878 wb_gpu_addr = prop->wptr_gpu_addr;
879 m->sdmax_rlcx_rb_wptr_poll_addr_lo = lower_32_bits(wb_gpu_addr);
880 m->sdmax_rlcx_rb_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr);
881
882 wb_gpu_addr = prop->rptr_gpu_addr;
883 m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits(wb_gpu_addr);
884 m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits(wb_gpu_addr);
885
886 m->sdmax_rlcx_ib_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, 0,
887 mmSDMA0_GFX_IB_CNTL));
888
889 m->sdmax_rlcx_doorbell_offset =
890 prop->doorbell_index << SDMA0_RLC0_DOORBELL_OFFSET__OFFSET__SHIFT;
891
892 m->sdmax_rlcx_doorbell = REG_SET_FIELD(0, SDMA0_RLC0_DOORBELL, ENABLE, 1);
893
894 return 0;
895 }
896
sdma_v5_2_set_mqd_funcs(struct amdgpu_device * adev)897 static void sdma_v5_2_set_mqd_funcs(struct amdgpu_device *adev)
898 {
899 adev->mqds[AMDGPU_HW_IP_DMA].mqd_size = sizeof(struct v10_sdma_mqd);
900 adev->mqds[AMDGPU_HW_IP_DMA].init_mqd = sdma_v5_2_mqd_init;
901 }
902
903 /**
904 * sdma_v5_2_ring_test_ring - simple async dma engine test
905 *
906 * @ring: amdgpu_ring structure holding ring information
907 *
908 * Test the DMA engine by writing using it to write an
909 * value to memory.
910 * Returns 0 for success, error for failure.
911 */
sdma_v5_2_ring_test_ring(struct amdgpu_ring * ring)912 static int sdma_v5_2_ring_test_ring(struct amdgpu_ring *ring)
913 {
914 struct amdgpu_device *adev = ring->adev;
915 unsigned i;
916 unsigned index;
917 int r;
918 u32 tmp;
919 u64 gpu_addr;
920
921 tmp = 0xCAFEDEAD;
922
923 r = amdgpu_device_wb_get(adev, &index);
924 if (r) {
925 dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
926 return r;
927 }
928
929 gpu_addr = adev->wb.gpu_addr + (index * 4);
930 adev->wb.wb[index] = cpu_to_le32(tmp);
931
932 r = amdgpu_ring_alloc(ring, 20);
933 if (r) {
934 DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
935 amdgpu_device_wb_free(adev, index);
936 return r;
937 }
938
939 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
940 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
941 amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
942 amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
943 amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0));
944 amdgpu_ring_write(ring, 0xDEADBEEF);
945 amdgpu_ring_commit(ring);
946
947 for (i = 0; i < adev->usec_timeout; i++) {
948 tmp = le32_to_cpu(adev->wb.wb[index]);
949 if (tmp == 0xDEADBEEF)
950 break;
951 if (amdgpu_emu_mode == 1)
952 msleep(1);
953 else
954 udelay(1);
955 }
956
957 if (i >= adev->usec_timeout)
958 r = -ETIMEDOUT;
959
960 amdgpu_device_wb_free(adev, index);
961
962 return r;
963 }
964
965 /**
966 * sdma_v5_2_ring_test_ib - test an IB on the DMA engine
967 *
968 * @ring: amdgpu_ring structure holding ring information
969 * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
970 *
971 * Test a simple IB in the DMA ring.
972 * Returns 0 on success, error on failure.
973 */
sdma_v5_2_ring_test_ib(struct amdgpu_ring * ring,long timeout)974 static int sdma_v5_2_ring_test_ib(struct amdgpu_ring *ring, long timeout)
975 {
976 struct amdgpu_device *adev = ring->adev;
977 struct amdgpu_ib ib;
978 struct dma_fence *f = NULL;
979 unsigned index;
980 long r;
981 u32 tmp = 0;
982 u64 gpu_addr;
983
984 tmp = 0xCAFEDEAD;
985 memset(&ib, 0, sizeof(ib));
986
987 r = amdgpu_device_wb_get(adev, &index);
988 if (r) {
989 dev_err(adev->dev, "(%ld) failed to allocate wb slot\n", r);
990 return r;
991 }
992
993 gpu_addr = adev->wb.gpu_addr + (index * 4);
994 adev->wb.wb[index] = cpu_to_le32(tmp);
995
996 r = amdgpu_ib_get(adev, NULL, 256, AMDGPU_IB_POOL_DIRECT, &ib);
997 if (r) {
998 DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r);
999 goto err0;
1000 }
1001
1002 ib.ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
1003 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1004 ib.ptr[1] = lower_32_bits(gpu_addr);
1005 ib.ptr[2] = upper_32_bits(gpu_addr);
1006 ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0);
1007 ib.ptr[4] = 0xDEADBEEF;
1008 ib.ptr[5] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1009 ib.ptr[6] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1010 ib.ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1011 ib.length_dw = 8;
1012
1013 r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
1014 if (r)
1015 goto err1;
1016
1017 r = dma_fence_wait_timeout(f, false, timeout);
1018 if (r == 0) {
1019 DRM_ERROR("amdgpu: IB test timed out\n");
1020 r = -ETIMEDOUT;
1021 goto err1;
1022 } else if (r < 0) {
1023 DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
1024 goto err1;
1025 }
1026
1027 tmp = le32_to_cpu(adev->wb.wb[index]);
1028
1029 if (tmp == 0xDEADBEEF)
1030 r = 0;
1031 else
1032 r = -EINVAL;
1033
1034 err1:
1035 amdgpu_ib_free(&ib, NULL);
1036 dma_fence_put(f);
1037 err0:
1038 amdgpu_device_wb_free(adev, index);
1039 return r;
1040 }
1041
1042
1043 /**
1044 * sdma_v5_2_vm_copy_pte - update PTEs by copying them from the GART
1045 *
1046 * @ib: indirect buffer to fill with commands
1047 * @pe: addr of the page entry
1048 * @src: src addr to copy from
1049 * @count: number of page entries to update
1050 *
1051 * Update PTEs by copying them from the GART using sDMA.
1052 */
sdma_v5_2_vm_copy_pte(struct amdgpu_ib * ib,uint64_t pe,uint64_t src,unsigned count)1053 static void sdma_v5_2_vm_copy_pte(struct amdgpu_ib *ib,
1054 uint64_t pe, uint64_t src,
1055 unsigned count)
1056 {
1057 unsigned bytes = count * 8;
1058
1059 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1060 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
1061 ib->ptr[ib->length_dw++] = bytes - 1;
1062 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1063 ib->ptr[ib->length_dw++] = lower_32_bits(src);
1064 ib->ptr[ib->length_dw++] = upper_32_bits(src);
1065 ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1066 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1067
1068 }
1069
1070 /**
1071 * sdma_v5_2_vm_write_pte - update PTEs by writing them manually
1072 *
1073 * @ib: indirect buffer to fill with commands
1074 * @pe: addr of the page entry
1075 * @value: dst addr to write into pe
1076 * @count: number of page entries to update
1077 * @incr: increase next addr by incr bytes
1078 *
1079 * Update PTEs by writing them manually using sDMA.
1080 */
sdma_v5_2_vm_write_pte(struct amdgpu_ib * ib,uint64_t pe,uint64_t value,unsigned count,uint32_t incr)1081 static void sdma_v5_2_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
1082 uint64_t value, unsigned count,
1083 uint32_t incr)
1084 {
1085 unsigned ndw = count * 2;
1086
1087 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
1088 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1089 ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1090 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1091 ib->ptr[ib->length_dw++] = ndw - 1;
1092 for (; ndw > 0; ndw -= 2) {
1093 ib->ptr[ib->length_dw++] = lower_32_bits(value);
1094 ib->ptr[ib->length_dw++] = upper_32_bits(value);
1095 value += incr;
1096 }
1097 }
1098
1099 /**
1100 * sdma_v5_2_vm_set_pte_pde - update the page tables using sDMA
1101 *
1102 * @ib: indirect buffer to fill with commands
1103 * @pe: addr of the page entry
1104 * @addr: dst addr to write into pe
1105 * @count: number of page entries to update
1106 * @incr: increase next addr by incr bytes
1107 * @flags: access flags
1108 *
1109 * Update the page tables using sDMA.
1110 */
sdma_v5_2_vm_set_pte_pde(struct amdgpu_ib * ib,uint64_t pe,uint64_t addr,unsigned count,uint32_t incr,uint64_t flags)1111 static void sdma_v5_2_vm_set_pte_pde(struct amdgpu_ib *ib,
1112 uint64_t pe,
1113 uint64_t addr, unsigned count,
1114 uint32_t incr, uint64_t flags)
1115 {
1116 /* for physically contiguous pages (vram) */
1117 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_PTEPDE);
1118 ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
1119 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1120 ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
1121 ib->ptr[ib->length_dw++] = upper_32_bits(flags);
1122 ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
1123 ib->ptr[ib->length_dw++] = upper_32_bits(addr);
1124 ib->ptr[ib->length_dw++] = incr; /* increment size */
1125 ib->ptr[ib->length_dw++] = 0;
1126 ib->ptr[ib->length_dw++] = count - 1; /* number of entries */
1127 }
1128
1129 /**
1130 * sdma_v5_2_ring_pad_ib - pad the IB
1131 *
1132 * @ib: indirect buffer to fill with padding
1133 * @ring: amdgpu_ring structure holding ring information
1134 *
1135 * Pad the IB with NOPs to a boundary multiple of 8.
1136 */
sdma_v5_2_ring_pad_ib(struct amdgpu_ring * ring,struct amdgpu_ib * ib)1137 static void sdma_v5_2_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
1138 {
1139 struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
1140 u32 pad_count;
1141 int i;
1142
1143 pad_count = (-ib->length_dw) & 0x7;
1144 for (i = 0; i < pad_count; i++)
1145 if (sdma && sdma->burst_nop && (i == 0))
1146 ib->ptr[ib->length_dw++] =
1147 SDMA_PKT_HEADER_OP(SDMA_OP_NOP) |
1148 SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
1149 else
1150 ib->ptr[ib->length_dw++] =
1151 SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
1152 }
1153
1154
1155 /**
1156 * sdma_v5_2_ring_emit_pipeline_sync - sync the pipeline
1157 *
1158 * @ring: amdgpu_ring pointer
1159 *
1160 * Make sure all previous operations are completed (CIK).
1161 */
sdma_v5_2_ring_emit_pipeline_sync(struct amdgpu_ring * ring)1162 static void sdma_v5_2_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
1163 {
1164 uint32_t seq = ring->fence_drv.sync_seq;
1165 uint64_t addr = ring->fence_drv.gpu_addr;
1166
1167 /* wait for idle */
1168 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1169 SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1170 SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
1171 SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1));
1172 amdgpu_ring_write(ring, addr & 0xfffffffc);
1173 amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
1174 amdgpu_ring_write(ring, seq); /* reference */
1175 amdgpu_ring_write(ring, 0xffffffff); /* mask */
1176 amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1177 SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */
1178 }
1179
1180
1181 /**
1182 * sdma_v5_2_ring_emit_vm_flush - vm flush using sDMA
1183 *
1184 * @ring: amdgpu_ring pointer
1185 * @vmid: vmid number to use
1186 * @pd_addr: address
1187 *
1188 * Update the page table base and flush the VM TLB
1189 * using sDMA.
1190 */
sdma_v5_2_ring_emit_vm_flush(struct amdgpu_ring * ring,unsigned vmid,uint64_t pd_addr)1191 static void sdma_v5_2_ring_emit_vm_flush(struct amdgpu_ring *ring,
1192 unsigned vmid, uint64_t pd_addr)
1193 {
1194 struct amdgpu_vmhub *hub = &ring->adev->vmhub[ring->vm_hub];
1195 uint32_t req = hub->vmhub_funcs->get_invalidate_req(vmid, 0);
1196
1197 /* Update the PD address for this VMID. */
1198 amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_lo32 +
1199 (hub->ctx_addr_distance * vmid),
1200 lower_32_bits(pd_addr));
1201 amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_hi32 +
1202 (hub->ctx_addr_distance * vmid),
1203 upper_32_bits(pd_addr));
1204
1205 /* Trigger invalidation. */
1206 amdgpu_ring_write(ring,
1207 SDMA_PKT_VM_INVALIDATION_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1208 SDMA_PKT_VM_INVALIDATION_HEADER_SUB_OP(SDMA_SUBOP_VM_INVALIDATION) |
1209 SDMA_PKT_VM_INVALIDATION_HEADER_GFX_ENG_ID(ring->vm_inv_eng) |
1210 SDMA_PKT_VM_INVALIDATION_HEADER_MM_ENG_ID(0x1f));
1211 amdgpu_ring_write(ring, req);
1212 amdgpu_ring_write(ring, 0xFFFFFFFF);
1213 amdgpu_ring_write(ring,
1214 SDMA_PKT_VM_INVALIDATION_ADDRESSRANGEHI_INVALIDATEACK(1 << vmid) |
1215 SDMA_PKT_VM_INVALIDATION_ADDRESSRANGEHI_ADDRESSRANGEHI(0x1F));
1216 }
1217
sdma_v5_2_ring_emit_wreg(struct amdgpu_ring * ring,uint32_t reg,uint32_t val)1218 static void sdma_v5_2_ring_emit_wreg(struct amdgpu_ring *ring,
1219 uint32_t reg, uint32_t val)
1220 {
1221 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
1222 SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
1223 amdgpu_ring_write(ring, reg);
1224 amdgpu_ring_write(ring, val);
1225 }
1226
sdma_v5_2_ring_emit_reg_wait(struct amdgpu_ring * ring,uint32_t reg,uint32_t val,uint32_t mask)1227 static void sdma_v5_2_ring_emit_reg_wait(struct amdgpu_ring *ring, uint32_t reg,
1228 uint32_t val, uint32_t mask)
1229 {
1230 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1231 SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1232 SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* equal */
1233 amdgpu_ring_write(ring, reg << 2);
1234 amdgpu_ring_write(ring, 0);
1235 amdgpu_ring_write(ring, val); /* reference */
1236 amdgpu_ring_write(ring, mask); /* mask */
1237 amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1238 SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10));
1239 }
1240
sdma_v5_2_ring_emit_reg_write_reg_wait(struct amdgpu_ring * ring,uint32_t reg0,uint32_t reg1,uint32_t ref,uint32_t mask)1241 static void sdma_v5_2_ring_emit_reg_write_reg_wait(struct amdgpu_ring *ring,
1242 uint32_t reg0, uint32_t reg1,
1243 uint32_t ref, uint32_t mask)
1244 {
1245 amdgpu_ring_emit_wreg(ring, reg0, ref);
1246 /* wait for a cycle to reset vm_inv_eng*_ack */
1247 amdgpu_ring_emit_reg_wait(ring, reg0, 0, 0);
1248 amdgpu_ring_emit_reg_wait(ring, reg1, mask, mask);
1249 }
1250
sdma_v5_2_early_init(struct amdgpu_ip_block * ip_block)1251 static int sdma_v5_2_early_init(struct amdgpu_ip_block *ip_block)
1252 {
1253 struct amdgpu_device *adev = ip_block->adev;
1254 int r;
1255
1256 r = amdgpu_sdma_init_microcode(adev, 0, true);
1257 if (r)
1258 return r;
1259
1260 sdma_v5_2_set_ring_funcs(adev);
1261 sdma_v5_2_set_buffer_funcs(adev);
1262 sdma_v5_2_set_vm_pte_funcs(adev);
1263 sdma_v5_2_set_irq_funcs(adev);
1264 sdma_v5_2_set_mqd_funcs(adev);
1265
1266 return 0;
1267 }
1268
sdma_v5_2_seq_to_irq_id(int seq_num)1269 static unsigned sdma_v5_2_seq_to_irq_id(int seq_num)
1270 {
1271 switch (seq_num) {
1272 case 0:
1273 return SOC15_IH_CLIENTID_SDMA0;
1274 case 1:
1275 return SOC15_IH_CLIENTID_SDMA1;
1276 case 2:
1277 return SOC15_IH_CLIENTID_SDMA2;
1278 case 3:
1279 return SOC15_IH_CLIENTID_SDMA3_Sienna_Cichlid;
1280 default:
1281 break;
1282 }
1283 return -EINVAL;
1284 }
1285
sdma_v5_2_seq_to_trap_id(int seq_num)1286 static unsigned sdma_v5_2_seq_to_trap_id(int seq_num)
1287 {
1288 switch (seq_num) {
1289 case 0:
1290 return SDMA0_5_0__SRCID__SDMA_TRAP;
1291 case 1:
1292 return SDMA1_5_0__SRCID__SDMA_TRAP;
1293 case 2:
1294 return SDMA2_5_0__SRCID__SDMA_TRAP;
1295 case 3:
1296 return SDMA3_5_0__SRCID__SDMA_TRAP;
1297 default:
1298 break;
1299 }
1300 return -EINVAL;
1301 }
1302
sdma_v5_2_sw_init(struct amdgpu_ip_block * ip_block)1303 static int sdma_v5_2_sw_init(struct amdgpu_ip_block *ip_block)
1304 {
1305 struct amdgpu_ring *ring;
1306 int r, i;
1307 struct amdgpu_device *adev = ip_block->adev;
1308 uint32_t reg_count = ARRAY_SIZE(sdma_reg_list_5_2);
1309 uint32_t *ptr;
1310
1311 /* SDMA trap event */
1312 for (i = 0; i < adev->sdma.num_instances; i++) {
1313 r = amdgpu_irq_add_id(adev, sdma_v5_2_seq_to_irq_id(i),
1314 sdma_v5_2_seq_to_trap_id(i),
1315 &adev->sdma.trap_irq);
1316 if (r)
1317 return r;
1318 }
1319
1320 for (i = 0; i < adev->sdma.num_instances; i++) {
1321 mutex_init(&adev->sdma.instance[i].engine_reset_mutex);
1322 adev->sdma.instance[i].funcs = &sdma_v5_2_sdma_funcs;
1323 ring = &adev->sdma.instance[i].ring;
1324 ring->ring_obj = NULL;
1325 ring->use_doorbell = true;
1326 ring->me = i;
1327
1328 DRM_INFO("use_doorbell being set to: [%s]\n",
1329 ring->use_doorbell?"true":"false");
1330
1331 ring->doorbell_index =
1332 (adev->doorbell_index.sdma_engine[i] << 1); //get DWORD offset
1333
1334 ring->vm_hub = AMDGPU_GFXHUB(0);
1335 sprintf(ring->name, "sdma%d", i);
1336 r = amdgpu_ring_init(adev, ring, 1024, &adev->sdma.trap_irq,
1337 AMDGPU_SDMA_IRQ_INSTANCE0 + i,
1338 AMDGPU_RING_PRIO_DEFAULT, NULL);
1339 if (r)
1340 return r;
1341 }
1342
1343 adev->sdma.supported_reset =
1344 amdgpu_get_soft_full_reset_mask(&adev->sdma.instance[0].ring);
1345 switch (amdgpu_ip_version(adev, SDMA0_HWIP, 0)) {
1346 case IP_VERSION(5, 2, 0):
1347 case IP_VERSION(5, 2, 2):
1348 case IP_VERSION(5, 2, 3):
1349 case IP_VERSION(5, 2, 4):
1350 if ((adev->sdma.instance[0].fw_version >= 76) &&
1351 !amdgpu_sriov_vf(adev))
1352 adev->sdma.supported_reset |= AMDGPU_RESET_TYPE_PER_QUEUE;
1353 break;
1354 case IP_VERSION(5, 2, 5):
1355 if ((adev->sdma.instance[0].fw_version >= 34) &&
1356 !amdgpu_sriov_vf(adev))
1357 adev->sdma.supported_reset |= AMDGPU_RESET_TYPE_PER_QUEUE;
1358 break;
1359 default:
1360 break;
1361 }
1362
1363 /* Allocate memory for SDMA IP Dump buffer */
1364 ptr = kcalloc(adev->sdma.num_instances * reg_count, sizeof(uint32_t), GFP_KERNEL);
1365 if (ptr)
1366 adev->sdma.ip_dump = ptr;
1367 else
1368 DRM_ERROR("Failed to allocated memory for SDMA IP Dump\n");
1369
1370 r = amdgpu_sdma_sysfs_reset_mask_init(adev);
1371 if (r)
1372 return r;
1373
1374 return r;
1375 }
1376
sdma_v5_2_sw_fini(struct amdgpu_ip_block * ip_block)1377 static int sdma_v5_2_sw_fini(struct amdgpu_ip_block *ip_block)
1378 {
1379 struct amdgpu_device *adev = ip_block->adev;
1380 int i;
1381
1382 for (i = 0; i < adev->sdma.num_instances; i++)
1383 amdgpu_ring_fini(&adev->sdma.instance[i].ring);
1384
1385 amdgpu_sdma_sysfs_reset_mask_fini(adev);
1386 amdgpu_sdma_destroy_inst_ctx(adev, true);
1387
1388 kfree(adev->sdma.ip_dump);
1389
1390 return 0;
1391 }
1392
sdma_v5_2_hw_init(struct amdgpu_ip_block * ip_block)1393 static int sdma_v5_2_hw_init(struct amdgpu_ip_block *ip_block)
1394 {
1395 struct amdgpu_device *adev = ip_block->adev;
1396
1397 return sdma_v5_2_start(adev);
1398 }
1399
sdma_v5_2_hw_fini(struct amdgpu_ip_block * ip_block)1400 static int sdma_v5_2_hw_fini(struct amdgpu_ip_block *ip_block)
1401 {
1402 struct amdgpu_device *adev = ip_block->adev;
1403
1404 if (amdgpu_sriov_vf(adev))
1405 return 0;
1406
1407 sdma_v5_2_ctx_switch_enable(adev, false);
1408 sdma_v5_2_enable(adev, false);
1409
1410 return 0;
1411 }
1412
sdma_v5_2_suspend(struct amdgpu_ip_block * ip_block)1413 static int sdma_v5_2_suspend(struct amdgpu_ip_block *ip_block)
1414 {
1415 return sdma_v5_2_hw_fini(ip_block);
1416 }
1417
sdma_v5_2_resume(struct amdgpu_ip_block * ip_block)1418 static int sdma_v5_2_resume(struct amdgpu_ip_block *ip_block)
1419 {
1420 return sdma_v5_2_hw_init(ip_block);
1421 }
1422
sdma_v5_2_is_idle(struct amdgpu_ip_block * ip_block)1423 static bool sdma_v5_2_is_idle(struct amdgpu_ip_block *ip_block)
1424 {
1425 struct amdgpu_device *adev = ip_block->adev;
1426 u32 i;
1427
1428 for (i = 0; i < adev->sdma.num_instances; i++) {
1429 u32 tmp = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_STATUS_REG));
1430
1431 if (!(tmp & SDMA0_STATUS_REG__IDLE_MASK))
1432 return false;
1433 }
1434
1435 return true;
1436 }
1437
sdma_v5_2_wait_for_idle(struct amdgpu_ip_block * ip_block)1438 static int sdma_v5_2_wait_for_idle(struct amdgpu_ip_block *ip_block)
1439 {
1440 unsigned i;
1441 u32 sdma0, sdma1, sdma2, sdma3;
1442 struct amdgpu_device *adev = ip_block->adev;
1443
1444 for (i = 0; i < adev->usec_timeout; i++) {
1445 sdma0 = RREG32(sdma_v5_2_get_reg_offset(adev, 0, mmSDMA0_STATUS_REG));
1446 sdma1 = RREG32(sdma_v5_2_get_reg_offset(adev, 1, mmSDMA0_STATUS_REG));
1447 sdma2 = RREG32(sdma_v5_2_get_reg_offset(adev, 2, mmSDMA0_STATUS_REG));
1448 sdma3 = RREG32(sdma_v5_2_get_reg_offset(adev, 3, mmSDMA0_STATUS_REG));
1449
1450 if (sdma0 & sdma1 & sdma2 & sdma3 & SDMA0_STATUS_REG__IDLE_MASK)
1451 return 0;
1452 udelay(1);
1453 }
1454 return -ETIMEDOUT;
1455 }
1456
sdma_v5_2_reset_queue(struct amdgpu_ring * ring,unsigned int vmid,struct amdgpu_fence * timedout_fence)1457 static int sdma_v5_2_reset_queue(struct amdgpu_ring *ring,
1458 unsigned int vmid,
1459 struct amdgpu_fence *timedout_fence)
1460 {
1461 struct amdgpu_device *adev = ring->adev;
1462 int r;
1463
1464 if (ring->me >= adev->sdma.num_instances) {
1465 dev_err(adev->dev, "sdma instance not found\n");
1466 return -EINVAL;
1467 }
1468
1469 amdgpu_ring_reset_helper_begin(ring, timedout_fence);
1470
1471 amdgpu_amdkfd_suspend(adev, true);
1472 r = amdgpu_sdma_reset_engine(adev, ring->me, true);
1473 amdgpu_amdkfd_resume(adev, true);
1474 if (r)
1475 return r;
1476
1477 return amdgpu_ring_reset_helper_end(ring, timedout_fence);
1478 }
1479
sdma_v5_2_stop_queue(struct amdgpu_ring * ring)1480 static int sdma_v5_2_stop_queue(struct amdgpu_ring *ring)
1481 {
1482 u32 f32_cntl, freeze, cntl, stat1_reg;
1483 struct amdgpu_device *adev = ring->adev;
1484 int i, j, r = 0;
1485
1486 if (amdgpu_sriov_vf(adev))
1487 return -EINVAL;
1488
1489 i = ring->me;
1490 amdgpu_gfx_rlc_enter_safe_mode(adev, 0);
1491
1492 /* stop queue */
1493 sdma_v5_2_gfx_stop(adev, 1 << i);
1494
1495 /*engine stop SDMA1_F32_CNTL.HALT to 1 and SDMAx_FREEZE freeze bit to 1 */
1496 freeze = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_FREEZE));
1497 freeze = REG_SET_FIELD(freeze, SDMA0_FREEZE, FREEZE, 1);
1498 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_FREEZE), freeze);
1499
1500 for (j = 0; j < adev->usec_timeout; j++) {
1501 freeze = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_FREEZE));
1502
1503 if (REG_GET_FIELD(freeze, SDMA0_FREEZE, FROZEN) & 1)
1504 break;
1505 udelay(1);
1506 }
1507
1508
1509 if (j == adev->usec_timeout) {
1510 stat1_reg = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_STATUS1_REG));
1511 if ((stat1_reg & 0x3FF) != 0x3FF) {
1512 DRM_ERROR("cannot soft reset as sdma not idle\n");
1513 r = -ETIMEDOUT;
1514 goto err0;
1515 }
1516 }
1517
1518 f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
1519 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 1);
1520 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl);
1521
1522 cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
1523 cntl = REG_SET_FIELD(cntl, SDMA0_CNTL, UTC_L1_ENABLE, 0);
1524 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), cntl);
1525
1526 err0:
1527 amdgpu_gfx_rlc_exit_safe_mode(adev, 0);
1528 return r;
1529 }
1530
sdma_v5_2_restore_queue(struct amdgpu_ring * ring)1531 static int sdma_v5_2_restore_queue(struct amdgpu_ring *ring)
1532 {
1533 struct amdgpu_device *adev = ring->adev;
1534 u32 inst_id = ring->me;
1535 u32 freeze;
1536 int r;
1537
1538 amdgpu_gfx_rlc_enter_safe_mode(adev, 0);
1539 /* unfreeze and unhalt */
1540 freeze = RREG32(sdma_v5_2_get_reg_offset(adev, inst_id, mmSDMA0_FREEZE));
1541 freeze = REG_SET_FIELD(freeze, SDMA0_FREEZE, FREEZE, 0);
1542 WREG32(sdma_v5_2_get_reg_offset(adev, inst_id, mmSDMA0_FREEZE), freeze);
1543
1544 r = sdma_v5_2_gfx_resume_instance(adev, inst_id, true);
1545
1546 amdgpu_gfx_rlc_exit_safe_mode(adev, 0);
1547
1548 return r;
1549 }
1550
sdma_v5_2_ring_preempt_ib(struct amdgpu_ring * ring)1551 static int sdma_v5_2_ring_preempt_ib(struct amdgpu_ring *ring)
1552 {
1553 int i, r = 0;
1554 struct amdgpu_device *adev = ring->adev;
1555 u32 index = 0;
1556 u64 sdma_gfx_preempt;
1557
1558 amdgpu_sdma_get_index_from_ring(ring, &index);
1559 sdma_gfx_preempt =
1560 sdma_v5_2_get_reg_offset(adev, index, mmSDMA0_GFX_PREEMPT);
1561
1562 /* assert preemption condition */
1563 amdgpu_ring_set_preempt_cond_exec(ring, false);
1564
1565 /* emit the trailing fence */
1566 ring->trail_seq += 1;
1567 amdgpu_ring_alloc(ring, 10);
1568 sdma_v5_2_ring_emit_fence(ring, ring->trail_fence_gpu_addr,
1569 ring->trail_seq, 0);
1570 amdgpu_ring_commit(ring);
1571
1572 /* assert IB preemption */
1573 WREG32(sdma_gfx_preempt, 1);
1574
1575 /* poll the trailing fence */
1576 for (i = 0; i < adev->usec_timeout; i++) {
1577 if (ring->trail_seq ==
1578 le32_to_cpu(*(ring->trail_fence_cpu_addr)))
1579 break;
1580 udelay(1);
1581 }
1582
1583 if (i >= adev->usec_timeout) {
1584 r = -EINVAL;
1585 DRM_ERROR("ring %d failed to be preempted\n", ring->idx);
1586 }
1587
1588 /* deassert IB preemption */
1589 WREG32(sdma_gfx_preempt, 0);
1590
1591 /* deassert the preemption condition */
1592 amdgpu_ring_set_preempt_cond_exec(ring, true);
1593 return r;
1594 }
1595
sdma_v5_2_set_trap_irq_state(struct amdgpu_device * adev,struct amdgpu_irq_src * source,unsigned type,enum amdgpu_interrupt_state state)1596 static int sdma_v5_2_set_trap_irq_state(struct amdgpu_device *adev,
1597 struct amdgpu_irq_src *source,
1598 unsigned type,
1599 enum amdgpu_interrupt_state state)
1600 {
1601 u32 sdma_cntl;
1602 u32 reg_offset = sdma_v5_2_get_reg_offset(adev, type, mmSDMA0_CNTL);
1603
1604 if (!amdgpu_sriov_vf(adev)) {
1605 sdma_cntl = RREG32(reg_offset);
1606 sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE,
1607 state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0);
1608 WREG32(reg_offset, sdma_cntl);
1609 }
1610
1611 return 0;
1612 }
1613
sdma_v5_2_process_trap_irq(struct amdgpu_device * adev,struct amdgpu_irq_src * source,struct amdgpu_iv_entry * entry)1614 static int sdma_v5_2_process_trap_irq(struct amdgpu_device *adev,
1615 struct amdgpu_irq_src *source,
1616 struct amdgpu_iv_entry *entry)
1617 {
1618 uint32_t mes_queue_id = entry->src_data[0];
1619
1620 DRM_DEBUG("IH: SDMA trap\n");
1621
1622 if (adev->enable_mes && (mes_queue_id & AMDGPU_FENCE_MES_QUEUE_FLAG)) {
1623 struct amdgpu_mes_queue *queue;
1624
1625 mes_queue_id &= AMDGPU_FENCE_MES_QUEUE_ID_MASK;
1626
1627 spin_lock(&adev->mes.queue_id_lock);
1628 queue = idr_find(&adev->mes.queue_id_idr, mes_queue_id);
1629 if (queue) {
1630 DRM_DEBUG("process smda queue id = %d\n", mes_queue_id);
1631 amdgpu_fence_process(queue->ring);
1632 }
1633 spin_unlock(&adev->mes.queue_id_lock);
1634 return 0;
1635 }
1636
1637 switch (entry->client_id) {
1638 case SOC15_IH_CLIENTID_SDMA0:
1639 switch (entry->ring_id) {
1640 case 0:
1641 amdgpu_fence_process(&adev->sdma.instance[0].ring);
1642 break;
1643 case 1:
1644 /* XXX compute */
1645 break;
1646 case 2:
1647 /* XXX compute */
1648 break;
1649 case 3:
1650 /* XXX page queue*/
1651 break;
1652 }
1653 break;
1654 case SOC15_IH_CLIENTID_SDMA1:
1655 switch (entry->ring_id) {
1656 case 0:
1657 amdgpu_fence_process(&adev->sdma.instance[1].ring);
1658 break;
1659 case 1:
1660 /* XXX compute */
1661 break;
1662 case 2:
1663 /* XXX compute */
1664 break;
1665 case 3:
1666 /* XXX page queue*/
1667 break;
1668 }
1669 break;
1670 case SOC15_IH_CLIENTID_SDMA2:
1671 switch (entry->ring_id) {
1672 case 0:
1673 amdgpu_fence_process(&adev->sdma.instance[2].ring);
1674 break;
1675 case 1:
1676 /* XXX compute */
1677 break;
1678 case 2:
1679 /* XXX compute */
1680 break;
1681 case 3:
1682 /* XXX page queue*/
1683 break;
1684 }
1685 break;
1686 case SOC15_IH_CLIENTID_SDMA3_Sienna_Cichlid:
1687 switch (entry->ring_id) {
1688 case 0:
1689 amdgpu_fence_process(&adev->sdma.instance[3].ring);
1690 break;
1691 case 1:
1692 /* XXX compute */
1693 break;
1694 case 2:
1695 /* XXX compute */
1696 break;
1697 case 3:
1698 /* XXX page queue*/
1699 break;
1700 }
1701 break;
1702 }
1703 return 0;
1704 }
1705
sdma_v5_2_process_illegal_inst_irq(struct amdgpu_device * adev,struct amdgpu_irq_src * source,struct amdgpu_iv_entry * entry)1706 static int sdma_v5_2_process_illegal_inst_irq(struct amdgpu_device *adev,
1707 struct amdgpu_irq_src *source,
1708 struct amdgpu_iv_entry *entry)
1709 {
1710 return 0;
1711 }
1712
sdma_v5_2_firmware_mgcg_support(struct amdgpu_device * adev,int i)1713 static bool sdma_v5_2_firmware_mgcg_support(struct amdgpu_device *adev,
1714 int i)
1715 {
1716 switch (amdgpu_ip_version(adev, SDMA0_HWIP, 0)) {
1717 case IP_VERSION(5, 2, 1):
1718 if (adev->sdma.instance[i].fw_version < 70)
1719 return false;
1720 break;
1721 case IP_VERSION(5, 2, 3):
1722 if (adev->sdma.instance[i].fw_version < 47)
1723 return false;
1724 break;
1725 case IP_VERSION(5, 2, 7):
1726 if (adev->sdma.instance[i].fw_version < 9)
1727 return false;
1728 break;
1729 default:
1730 return true;
1731 }
1732
1733 return true;
1734
1735 }
1736
sdma_v5_2_update_medium_grain_clock_gating(struct amdgpu_device * adev,bool enable)1737 static void sdma_v5_2_update_medium_grain_clock_gating(struct amdgpu_device *adev,
1738 bool enable)
1739 {
1740 uint32_t data, def;
1741 int i;
1742
1743 for (i = 0; i < adev->sdma.num_instances; i++) {
1744
1745 if (!sdma_v5_2_firmware_mgcg_support(adev, i))
1746 adev->cg_flags &= ~AMD_CG_SUPPORT_SDMA_MGCG;
1747
1748 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
1749 /* Enable sdma clock gating */
1750 def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL));
1751 data &= ~(SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1752 SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1753 SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1754 SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1755 SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK |
1756 SDMA0_CLK_CTRL__SOFT_OVERRIDER_REG_MASK);
1757 if (def != data)
1758 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data);
1759 } else {
1760 /* Disable sdma clock gating */
1761 def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL));
1762 data |= (SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1763 SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1764 SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1765 SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1766 SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK |
1767 SDMA0_CLK_CTRL__SOFT_OVERRIDER_REG_MASK);
1768 if (def != data)
1769 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data);
1770 }
1771 }
1772 }
1773
sdma_v5_2_update_medium_grain_light_sleep(struct amdgpu_device * adev,bool enable)1774 static void sdma_v5_2_update_medium_grain_light_sleep(struct amdgpu_device *adev,
1775 bool enable)
1776 {
1777 uint32_t data, def;
1778 int i;
1779
1780 for (i = 0; i < adev->sdma.num_instances; i++) {
1781 if (adev->sdma.instance[i].fw_version < 70 &&
1782 amdgpu_ip_version(adev, SDMA0_HWIP, 0) ==
1783 IP_VERSION(5, 2, 1))
1784 adev->cg_flags &= ~AMD_CG_SUPPORT_SDMA_LS;
1785
1786 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) {
1787 /* Enable sdma mem light sleep */
1788 def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL));
1789 data |= SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1790 if (def != data)
1791 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data);
1792
1793 } else {
1794 /* Disable sdma mem light sleep */
1795 def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL));
1796 data &= ~SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1797 if (def != data)
1798 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data);
1799
1800 }
1801 }
1802 }
1803
sdma_v5_2_set_clockgating_state(struct amdgpu_ip_block * ip_block,enum amd_clockgating_state state)1804 static int sdma_v5_2_set_clockgating_state(struct amdgpu_ip_block *ip_block,
1805 enum amd_clockgating_state state)
1806 {
1807 struct amdgpu_device *adev = ip_block->adev;
1808
1809 if (amdgpu_sriov_vf(adev))
1810 return 0;
1811
1812 switch (amdgpu_ip_version(adev, SDMA0_HWIP, 0)) {
1813 case IP_VERSION(5, 2, 0):
1814 case IP_VERSION(5, 2, 2):
1815 case IP_VERSION(5, 2, 1):
1816 case IP_VERSION(5, 2, 4):
1817 case IP_VERSION(5, 2, 5):
1818 case IP_VERSION(5, 2, 6):
1819 case IP_VERSION(5, 2, 3):
1820 case IP_VERSION(5, 2, 7):
1821 sdma_v5_2_update_medium_grain_clock_gating(adev,
1822 state == AMD_CG_STATE_GATE);
1823 sdma_v5_2_update_medium_grain_light_sleep(adev,
1824 state == AMD_CG_STATE_GATE);
1825 break;
1826 default:
1827 break;
1828 }
1829
1830 return 0;
1831 }
1832
sdma_v5_2_set_powergating_state(struct amdgpu_ip_block * ip_block,enum amd_powergating_state state)1833 static int sdma_v5_2_set_powergating_state(struct amdgpu_ip_block *ip_block,
1834 enum amd_powergating_state state)
1835 {
1836 return 0;
1837 }
1838
sdma_v5_2_get_clockgating_state(struct amdgpu_ip_block * ip_block,u64 * flags)1839 static void sdma_v5_2_get_clockgating_state(struct amdgpu_ip_block *ip_block, u64 *flags)
1840 {
1841 struct amdgpu_device *adev = ip_block->adev;
1842 int data;
1843
1844 if (amdgpu_sriov_vf(adev))
1845 *flags = 0;
1846
1847 /* AMD_CG_SUPPORT_SDMA_MGCG */
1848 data = RREG32(sdma_v5_2_get_reg_offset(adev, 0, mmSDMA0_CLK_CTRL));
1849 if (!(data & SDMA0_CLK_CTRL__CGCG_EN_OVERRIDE_MASK))
1850 *flags |= AMD_CG_SUPPORT_SDMA_MGCG;
1851
1852 /* AMD_CG_SUPPORT_SDMA_LS */
1853 data = RREG32_KIQ(sdma_v5_2_get_reg_offset(adev, 0, mmSDMA0_POWER_CNTL));
1854 if (data & SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK)
1855 *flags |= AMD_CG_SUPPORT_SDMA_LS;
1856 }
1857
sdma_v5_2_ring_begin_use(struct amdgpu_ring * ring)1858 static void sdma_v5_2_ring_begin_use(struct amdgpu_ring *ring)
1859 {
1860 struct amdgpu_device *adev = ring->adev;
1861
1862 /* SDMA 5.2.3 (RMB) FW doesn't seem to properly
1863 * disallow GFXOFF in some cases leading to
1864 * hangs in SDMA. Disallow GFXOFF while SDMA is active.
1865 * We can probably just limit this to 5.2.3,
1866 * but it shouldn't hurt for other parts since
1867 * this GFXOFF will be disallowed anyway when SDMA is
1868 * active, this just makes it explicit.
1869 * sdma_v5_2_ring_set_wptr() takes advantage of this
1870 * to update the wptr because sometimes SDMA seems to miss
1871 * doorbells when entering PG. If you remove this, update
1872 * sdma_v5_2_ring_set_wptr() as well!
1873 */
1874 amdgpu_gfx_off_ctrl(adev, false);
1875 }
1876
sdma_v5_2_ring_end_use(struct amdgpu_ring * ring)1877 static void sdma_v5_2_ring_end_use(struct amdgpu_ring *ring)
1878 {
1879 struct amdgpu_device *adev = ring->adev;
1880
1881 /* SDMA 5.2.3 (RMB) FW doesn't seem to properly
1882 * disallow GFXOFF in some cases leading to
1883 * hangs in SDMA. Allow GFXOFF when SDMA is complete.
1884 */
1885 amdgpu_gfx_off_ctrl(adev, true);
1886 }
1887
sdma_v5_2_print_ip_state(struct amdgpu_ip_block * ip_block,struct drm_printer * p)1888 static void sdma_v5_2_print_ip_state(struct amdgpu_ip_block *ip_block, struct drm_printer *p)
1889 {
1890 struct amdgpu_device *adev = ip_block->adev;
1891 int i, j;
1892 uint32_t reg_count = ARRAY_SIZE(sdma_reg_list_5_2);
1893 uint32_t instance_offset;
1894
1895 if (!adev->sdma.ip_dump)
1896 return;
1897
1898 drm_printf(p, "num_instances:%d\n", adev->sdma.num_instances);
1899 for (i = 0; i < adev->sdma.num_instances; i++) {
1900 instance_offset = i * reg_count;
1901 drm_printf(p, "\nInstance:%d\n", i);
1902
1903 for (j = 0; j < reg_count; j++)
1904 drm_printf(p, "%-50s \t 0x%08x\n", sdma_reg_list_5_2[j].reg_name,
1905 adev->sdma.ip_dump[instance_offset + j]);
1906 }
1907 }
1908
sdma_v5_2_dump_ip_state(struct amdgpu_ip_block * ip_block)1909 static void sdma_v5_2_dump_ip_state(struct amdgpu_ip_block *ip_block)
1910 {
1911 struct amdgpu_device *adev = ip_block->adev;
1912 int i, j;
1913 uint32_t instance_offset;
1914 uint32_t reg_count = ARRAY_SIZE(sdma_reg_list_5_2);
1915
1916 if (!adev->sdma.ip_dump)
1917 return;
1918
1919 amdgpu_gfx_off_ctrl(adev, false);
1920 for (i = 0; i < adev->sdma.num_instances; i++) {
1921 instance_offset = i * reg_count;
1922 for (j = 0; j < reg_count; j++)
1923 adev->sdma.ip_dump[instance_offset + j] =
1924 RREG32(sdma_v5_2_get_reg_offset(adev, i,
1925 sdma_reg_list_5_2[j].reg_offset));
1926 }
1927 amdgpu_gfx_off_ctrl(adev, true);
1928 }
1929
1930 static const struct amd_ip_funcs sdma_v5_2_ip_funcs = {
1931 .name = "sdma_v5_2",
1932 .early_init = sdma_v5_2_early_init,
1933 .sw_init = sdma_v5_2_sw_init,
1934 .sw_fini = sdma_v5_2_sw_fini,
1935 .hw_init = sdma_v5_2_hw_init,
1936 .hw_fini = sdma_v5_2_hw_fini,
1937 .suspend = sdma_v5_2_suspend,
1938 .resume = sdma_v5_2_resume,
1939 .is_idle = sdma_v5_2_is_idle,
1940 .wait_for_idle = sdma_v5_2_wait_for_idle,
1941 .soft_reset = sdma_v5_2_soft_reset,
1942 .set_clockgating_state = sdma_v5_2_set_clockgating_state,
1943 .set_powergating_state = sdma_v5_2_set_powergating_state,
1944 .get_clockgating_state = sdma_v5_2_get_clockgating_state,
1945 .dump_ip_state = sdma_v5_2_dump_ip_state,
1946 .print_ip_state = sdma_v5_2_print_ip_state,
1947 };
1948
1949 static const struct amdgpu_ring_funcs sdma_v5_2_ring_funcs = {
1950 .type = AMDGPU_RING_TYPE_SDMA,
1951 .align_mask = 0xf,
1952 .nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP),
1953 .support_64bit_ptrs = true,
1954 .secure_submission_supported = true,
1955 .get_rptr = sdma_v5_2_ring_get_rptr,
1956 .get_wptr = sdma_v5_2_ring_get_wptr,
1957 .set_wptr = sdma_v5_2_ring_set_wptr,
1958 .emit_frame_size =
1959 5 + /* sdma_v5_2_ring_init_cond_exec */
1960 6 + /* sdma_v5_2_ring_emit_hdp_flush */
1961 3 + /* hdp_invalidate */
1962 6 + /* sdma_v5_2_ring_emit_pipeline_sync */
1963 /* sdma_v5_2_ring_emit_vm_flush */
1964 SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 +
1965 SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6 +
1966 10 + 10 + 10, /* sdma_v5_2_ring_emit_fence x3 for user fence, vm fence */
1967 .emit_ib_size = 7 + 6, /* sdma_v5_2_ring_emit_ib */
1968 .emit_ib = sdma_v5_2_ring_emit_ib,
1969 .emit_mem_sync = sdma_v5_2_ring_emit_mem_sync,
1970 .emit_fence = sdma_v5_2_ring_emit_fence,
1971 .emit_pipeline_sync = sdma_v5_2_ring_emit_pipeline_sync,
1972 .emit_vm_flush = sdma_v5_2_ring_emit_vm_flush,
1973 .emit_hdp_flush = sdma_v5_2_ring_emit_hdp_flush,
1974 .test_ring = sdma_v5_2_ring_test_ring,
1975 .test_ib = sdma_v5_2_ring_test_ib,
1976 .insert_nop = sdma_v5_2_ring_insert_nop,
1977 .pad_ib = sdma_v5_2_ring_pad_ib,
1978 .begin_use = sdma_v5_2_ring_begin_use,
1979 .end_use = sdma_v5_2_ring_end_use,
1980 .emit_wreg = sdma_v5_2_ring_emit_wreg,
1981 .emit_reg_wait = sdma_v5_2_ring_emit_reg_wait,
1982 .emit_reg_write_reg_wait = sdma_v5_2_ring_emit_reg_write_reg_wait,
1983 .init_cond_exec = sdma_v5_2_ring_init_cond_exec,
1984 .preempt_ib = sdma_v5_2_ring_preempt_ib,
1985 .reset = sdma_v5_2_reset_queue,
1986 };
1987
sdma_v5_2_set_ring_funcs(struct amdgpu_device * adev)1988 static void sdma_v5_2_set_ring_funcs(struct amdgpu_device *adev)
1989 {
1990 int i;
1991
1992 for (i = 0; i < adev->sdma.num_instances; i++) {
1993 adev->sdma.instance[i].ring.funcs = &sdma_v5_2_ring_funcs;
1994 adev->sdma.instance[i].ring.me = i;
1995 }
1996 }
1997
1998 static const struct amdgpu_irq_src_funcs sdma_v5_2_trap_irq_funcs = {
1999 .set = sdma_v5_2_set_trap_irq_state,
2000 .process = sdma_v5_2_process_trap_irq,
2001 };
2002
2003 static const struct amdgpu_irq_src_funcs sdma_v5_2_illegal_inst_irq_funcs = {
2004 .process = sdma_v5_2_process_illegal_inst_irq,
2005 };
2006
sdma_v5_2_set_irq_funcs(struct amdgpu_device * adev)2007 static void sdma_v5_2_set_irq_funcs(struct amdgpu_device *adev)
2008 {
2009 adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_INSTANCE0 +
2010 adev->sdma.num_instances;
2011 adev->sdma.trap_irq.funcs = &sdma_v5_2_trap_irq_funcs;
2012 adev->sdma.illegal_inst_irq.funcs = &sdma_v5_2_illegal_inst_irq_funcs;
2013 }
2014
2015 /**
2016 * sdma_v5_2_emit_copy_buffer - copy buffer using the sDMA engine
2017 *
2018 * @ib: indirect buffer to copy to
2019 * @src_offset: src GPU address
2020 * @dst_offset: dst GPU address
2021 * @byte_count: number of bytes to xfer
2022 * @copy_flags: copy flags for the buffers
2023 *
2024 * Copy GPU buffers using the DMA engine.
2025 * Used by the amdgpu ttm implementation to move pages if
2026 * registered as the asic copy callback.
2027 */
sdma_v5_2_emit_copy_buffer(struct amdgpu_ib * ib,uint64_t src_offset,uint64_t dst_offset,uint32_t byte_count,uint32_t copy_flags)2028 static void sdma_v5_2_emit_copy_buffer(struct amdgpu_ib *ib,
2029 uint64_t src_offset,
2030 uint64_t dst_offset,
2031 uint32_t byte_count,
2032 uint32_t copy_flags)
2033 {
2034 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
2035 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR) |
2036 SDMA_PKT_COPY_LINEAR_HEADER_TMZ((copy_flags & AMDGPU_COPY_FLAGS_TMZ) ? 1 : 0);
2037 ib->ptr[ib->length_dw++] = byte_count - 1;
2038 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
2039 ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
2040 ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
2041 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
2042 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
2043 }
2044
2045 /**
2046 * sdma_v5_2_emit_fill_buffer - fill buffer using the sDMA engine
2047 *
2048 * @ib: indirect buffer to fill
2049 * @src_data: value to write to buffer
2050 * @dst_offset: dst GPU address
2051 * @byte_count: number of bytes to xfer
2052 *
2053 * Fill GPU buffers using the DMA engine.
2054 */
sdma_v5_2_emit_fill_buffer(struct amdgpu_ib * ib,uint32_t src_data,uint64_t dst_offset,uint32_t byte_count)2055 static void sdma_v5_2_emit_fill_buffer(struct amdgpu_ib *ib,
2056 uint32_t src_data,
2057 uint64_t dst_offset,
2058 uint32_t byte_count)
2059 {
2060 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL);
2061 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
2062 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
2063 ib->ptr[ib->length_dw++] = src_data;
2064 ib->ptr[ib->length_dw++] = byte_count - 1;
2065 }
2066
2067 static const struct amdgpu_buffer_funcs sdma_v5_2_buffer_funcs = {
2068 .copy_max_bytes = 0x400000,
2069 .copy_num_dw = 7,
2070 .emit_copy_buffer = sdma_v5_2_emit_copy_buffer,
2071
2072 .fill_max_bytes = 0x400000,
2073 .fill_num_dw = 5,
2074 .emit_fill_buffer = sdma_v5_2_emit_fill_buffer,
2075 };
2076
sdma_v5_2_set_buffer_funcs(struct amdgpu_device * adev)2077 static void sdma_v5_2_set_buffer_funcs(struct amdgpu_device *adev)
2078 {
2079 if (adev->mman.buffer_funcs == NULL) {
2080 adev->mman.buffer_funcs = &sdma_v5_2_buffer_funcs;
2081 adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
2082 }
2083 }
2084
2085 static const struct amdgpu_vm_pte_funcs sdma_v5_2_vm_pte_funcs = {
2086 .copy_pte_num_dw = 7,
2087 .copy_pte = sdma_v5_2_vm_copy_pte,
2088 .write_pte = sdma_v5_2_vm_write_pte,
2089 .set_pte_pde = sdma_v5_2_vm_set_pte_pde,
2090 };
2091
sdma_v5_2_set_vm_pte_funcs(struct amdgpu_device * adev)2092 static void sdma_v5_2_set_vm_pte_funcs(struct amdgpu_device *adev)
2093 {
2094 unsigned i;
2095
2096 if (adev->vm_manager.vm_pte_funcs == NULL) {
2097 adev->vm_manager.vm_pte_funcs = &sdma_v5_2_vm_pte_funcs;
2098 for (i = 0; i < adev->sdma.num_instances; i++) {
2099 adev->vm_manager.vm_pte_scheds[i] =
2100 &adev->sdma.instance[i].ring.sched;
2101 }
2102 adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
2103 }
2104 }
2105
2106 const struct amdgpu_ip_block_version sdma_v5_2_ip_block = {
2107 .type = AMD_IP_BLOCK_TYPE_SDMA,
2108 .major = 5,
2109 .minor = 2,
2110 .rev = 0,
2111 .funcs = &sdma_v5_2_ip_funcs,
2112 };
2113