projects / deliverable / linux.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
drm/amdgpu: fix user ptr race condition
[deliverable/linux.git] / drivers / gpu / drm / amd / amdgpu / cik_sdma.c
CommitLineData
a2e73f56
AD		 1/*
2 * Copyright 2013 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 "cikd.h"
30#include "cik.h"
31
32#include "bif/bif_4_1_d.h"
33#include "bif/bif_4_1_sh_mask.h"
34
35#include "gca/gfx_7_2_d.h"
74a5d165
JX		 36#include "gca/gfx_7_2_enum.h"
37#include "gca/gfx_7_2_sh_mask.h"
a2e73f56
AD		 38
39#include "gmc/gmc_7_1_d.h"
40#include "gmc/gmc_7_1_sh_mask.h"
41
42#include "oss/oss_2_0_d.h"
43#include "oss/oss_2_0_sh_mask.h"
44
45static const u32 sdma_offsets[SDMA_MAX_INSTANCE] =
46{
47 SDMA0_REGISTER_OFFSET,
48 SDMA1_REGISTER_OFFSET
49};
50
51static void cik_sdma_set_ring_funcs(struct amdgpu_device *adev);
52static void cik_sdma_set_irq_funcs(struct amdgpu_device *adev);
53static void cik_sdma_set_buffer_funcs(struct amdgpu_device *adev);
54static void cik_sdma_set_vm_pte_funcs(struct amdgpu_device *adev);
55
56MODULE_FIRMWARE("radeon/bonaire_sdma.bin");
57MODULE_FIRMWARE("radeon/bonaire_sdma1.bin");
58MODULE_FIRMWARE("radeon/hawaii_sdma.bin");
59MODULE_FIRMWARE("radeon/hawaii_sdma1.bin");
60MODULE_FIRMWARE("radeon/kaveri_sdma.bin");
61MODULE_FIRMWARE("radeon/kaveri_sdma1.bin");
62MODULE_FIRMWARE("radeon/kabini_sdma.bin");
63MODULE_FIRMWARE("radeon/kabini_sdma1.bin");
64MODULE_FIRMWARE("radeon/mullins_sdma.bin");
65MODULE_FIRMWARE("radeon/mullins_sdma1.bin");
66
67u32 amdgpu_cik_gpu_check_soft_reset(struct amdgpu_device *adev);
68
69/*
70 * sDMA - System DMA
71 * Starting with CIK, the GPU has new asynchronous
72 * DMA engines. These engines are used for compute
73 * and gfx. There are two DMA engines (SDMA0, SDMA1)
74 * and each one supports 1 ring buffer used for gfx
75 * and 2 queues used for compute.
76 *
77 * The programming model is very similar to the CP
78 * (ring buffer, IBs, etc.), but sDMA has it's own
79 * packet format that is different from the PM4 format
80 * used by the CP. sDMA supports copying data, writing
81 * embedded data, solid fills, and a number of other
82 * things. It also has support for tiling/detiling of
83 * buffers.
84 */
85
86/**
87 * cik_sdma_init_microcode - load ucode images from disk
88 *
89 * @adev: amdgpu_device pointer
90 *
91 * Use the firmware interface to load the ucode images into
92 * the driver (not loaded into hw).
93 * Returns 0 on success, error on failure.
94 */
95static int cik_sdma_init_microcode(struct amdgpu_device *adev)
96{
97 const char *chip_name;
98 char fw_name[30];
99 int err, i;
100
101 DRM_DEBUG("\n");
102
103 switch (adev->asic_type) {
104 case CHIP_BONAIRE:
105 chip_name = "bonaire";
106 break;
107 case CHIP_HAWAII:
108 chip_name = "hawaii";
109 break;
110 case CHIP_KAVERI:
111 chip_name = "kaveri";
112 break;
113 case CHIP_KABINI:
114 chip_name = "kabini";
115 break;
116 case CHIP_MULLINS:
117 chip_name = "mullins";
118 break;
119 default: BUG();
120 }
121
122 for (i = 0; i < SDMA_MAX_INSTANCE; i++) {
123 if (i == 0)
124 snprintf(fw_name, sizeof(fw_name), "radeon/%s_sdma.bin", chip_name);
125 else
126 snprintf(fw_name, sizeof(fw_name), "radeon/%s_sdma1.bin", chip_name);
127 err = request_firmware(&adev->sdma[i].fw, fw_name, adev->dev);
128 if (err)
129 goto out;
130 err = amdgpu_ucode_validate(adev->sdma[i].fw);
131 }
132out:
133 if (err) {
134 printk(KERN_ERR
135 "cik_sdma: Failed to load firmware \"%s\"\n",
136 fw_name);
137 for (i = 0; i < SDMA_MAX_INSTANCE; i++) {
138 release_firmware(adev->sdma[i].fw);
139 adev->sdma[i].fw = NULL;
140 }
141 }
142 return err;
143}
144
145/**
146 * cik_sdma_ring_get_rptr - get the current read pointer
147 *
148 * @ring: amdgpu ring pointer
149 *
150 * Get the current rptr from the hardware (CIK+).
151 */
152static uint32_t cik_sdma_ring_get_rptr(struct amdgpu_ring *ring)
153{
154 u32 rptr;
155
156 rptr = ring->adev->wb.wb[ring->rptr_offs];
157
158 return (rptr & 0x3fffc) >> 2;
159}
160
161/**
162 * cik_sdma_ring_get_wptr - get the current write pointer
163 *
164 * @ring: amdgpu ring pointer
165 *
166 * Get the current wptr from the hardware (CIK+).
167 */
168static uint32_t cik_sdma_ring_get_wptr(struct amdgpu_ring *ring)
169{
170 struct amdgpu_device *adev = ring->adev;
171 u32 me = (ring == &adev->sdma[0].ring) ? 0 : 1;
172
173 return (RREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[me]) & 0x3fffc) >> 2;
174}
175
176/**
177 * cik_sdma_ring_set_wptr - commit the write pointer
178 *
179 * @ring: amdgpu ring pointer
180 *
181 * Write the wptr back to the hardware (CIK+).
182 */
183static void cik_sdma_ring_set_wptr(struct amdgpu_ring *ring)
184{
185 struct amdgpu_device *adev = ring->adev;
186 u32 me = (ring == &adev->sdma[0].ring) ? 0 : 1;
187
188 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[me], (ring->wptr << 2) & 0x3fffc);
189}
190
a2e73f56
AD		 191/**
192 * cik_sdma_ring_emit_ib - Schedule an IB on the DMA engine
193 *
194 * @ring: amdgpu ring pointer
195 * @ib: IB object to schedule
196 *
197 * Schedule an IB in the DMA ring (CIK).
198 */
199static void cik_sdma_ring_emit_ib(struct amdgpu_ring *ring,
200 struct amdgpu_ib *ib)
201{
202 u32 extra_bits = (ib->vm ? ib->vm->ids[ring->idx].id : 0) & 0xf;
203 u32 next_rptr = ring->wptr + 5;
204
a2e73f56
AD		 205 while ((next_rptr & 7) != 4)
206 next_rptr++;
207
208 next_rptr += 4;
209 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_WRITE, SDMA_WRITE_SUB_OPCODE_LINEAR, 0));
210 amdgpu_ring_write(ring, ring->next_rptr_gpu_addr & 0xfffffffc);
211 amdgpu_ring_write(ring, upper_32_bits(ring->next_rptr_gpu_addr) & 0xffffffff);
212 amdgpu_ring_write(ring, 1); /* number of DWs to follow */
213 amdgpu_ring_write(ring, next_rptr);
214
a2e73f56
AD		 215 /* IB packet must end on a 8 DW boundary */
216 while ((ring->wptr & 7) != 4)
217 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0));
218 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_INDIRECT_BUFFER, 0, extra_bits));
219 amdgpu_ring_write(ring, ib->gpu_addr & 0xffffffe0); /* base must be 32 byte aligned */
220 amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr) & 0xffffffff);
221 amdgpu_ring_write(ring, ib->length_dw);
222
223}
224
225/**
d2edb07b 226 * cik_sdma_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
a2e73f56
AD		 227 *
228 * @ring: amdgpu ring pointer
229 *
230 * Emit an hdp flush packet on the requested DMA ring.
231 */
d2edb07b 232static void cik_sdma_ring_emit_hdp_flush(struct amdgpu_ring *ring)
a2e73f56
AD		 233{
234 u32 extra_bits = (SDMA_POLL_REG_MEM_EXTRA_OP(1) |
235 SDMA_POLL_REG_MEM_EXTRA_FUNC(3)); /* == */
236 u32 ref_and_mask;
237
238 if (ring == &ring->adev->sdma[0].ring)
239 ref_and_mask = GPU_HDP_FLUSH_DONE__SDMA0_MASK;
240 else
241 ref_and_mask = GPU_HDP_FLUSH_DONE__SDMA1_MASK;
242
243 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0, extra_bits));
244 amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_DONE << 2);
245 amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_REQ << 2);
246 amdgpu_ring_write(ring, ref_and_mask); /* reference */
247 amdgpu_ring_write(ring, ref_and_mask); /* mask */
248 amdgpu_ring_write(ring, (0xfff << 16) | 10); /* retry count, poll interval */
249}
250
251/**
252 * cik_sdma_ring_emit_fence - emit a fence on the DMA ring
253 *
254 * @ring: amdgpu ring pointer
255 * @fence: amdgpu fence object
256 *
257 * Add a DMA fence packet to the ring to write
258 * the fence seq number and DMA trap packet to generate
259 * an interrupt if needed (CIK).
260 */
261static void cik_sdma_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
262 bool write64bit)
263{
264 /* write the fence */
265 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_FENCE, 0, 0));
266 amdgpu_ring_write(ring, lower_32_bits(addr));
267 amdgpu_ring_write(ring, upper_32_bits(addr));
268 amdgpu_ring_write(ring, lower_32_bits(seq));
269
270 /* optionally write high bits as well */
271 if (write64bit) {
272 addr += 4;
273 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_FENCE, 0, 0));
274 amdgpu_ring_write(ring, lower_32_bits(addr));
275 amdgpu_ring_write(ring, upper_32_bits(addr));
276 amdgpu_ring_write(ring, upper_32_bits(seq));
277 }
278
279 /* generate an interrupt */
280 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_TRAP, 0, 0));
281}
282
283/**
284 * cik_sdma_ring_emit_semaphore - emit a semaphore on the dma ring
285 *
286 * @ring: amdgpu_ring structure holding ring information
287 * @semaphore: amdgpu semaphore object
288 * @emit_wait: wait or signal semaphore
289 *
290 * Add a DMA semaphore packet to the ring wait on or signal
291 * other rings (CIK).
292 */
293static bool cik_sdma_ring_emit_semaphore(struct amdgpu_ring *ring,
294 struct amdgpu_semaphore *semaphore,
295 bool emit_wait)
296{
297 u64 addr = semaphore->gpu_addr;
298 u32 extra_bits = emit_wait ? 0 : SDMA_SEMAPHORE_EXTRA_S;
299
300 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SEMAPHORE, 0, extra_bits));
301 amdgpu_ring_write(ring, addr & 0xfffffff8);
302 amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
303
304 return true;
305}
306
307/**
308 * cik_sdma_gfx_stop - stop the gfx async dma engines
309 *
310 * @adev: amdgpu_device pointer
311 *
312 * Stop the gfx async dma ring buffers (CIK).
313 */
314static void cik_sdma_gfx_stop(struct amdgpu_device *adev)
315{
316 struct amdgpu_ring *sdma0 = &adev->sdma[0].ring;
317 struct amdgpu_ring *sdma1 = &adev->sdma[1].ring;
318 u32 rb_cntl;
319 int i;
320
321 if ((adev->mman.buffer_funcs_ring == sdma0) ||
322 (adev->mman.buffer_funcs_ring == sdma1))
323 amdgpu_ttm_set_active_vram_size(adev, adev->mc.visible_vram_size);
324
325 for (i = 0; i < SDMA_MAX_INSTANCE; i++) {
326 rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
327 rb_cntl &= ~SDMA0_GFX_RB_CNTL__RB_ENABLE_MASK;
328 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
329 WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], 0);
330 }
331 sdma0->ready = false;
332 sdma1->ready = false;
333}
334
335/**
336 * cik_sdma_rlc_stop - stop the compute async dma engines
337 *
338 * @adev: amdgpu_device pointer
339 *
340 * Stop the compute async dma queues (CIK).
341 */
342static void cik_sdma_rlc_stop(struct amdgpu_device *adev)
343{
344 /* XXX todo */
345}
346
347/**
348 * cik_sdma_enable - stop the async dma engines
349 *
350 * @adev: amdgpu_device pointer
351 * @enable: enable/disable the DMA MEs.
352 *
353 * Halt or unhalt the async dma engines (CIK).
354 */
355static void cik_sdma_enable(struct amdgpu_device *adev, bool enable)
356{
357 u32 me_cntl;
358 int i;
359
360 if (enable == false) {
361 cik_sdma_gfx_stop(adev);
362 cik_sdma_rlc_stop(adev);
363 }
364
365 for (i = 0; i < SDMA_MAX_INSTANCE; i++) {
366 me_cntl = RREG32(mmSDMA0_F32_CNTL + sdma_offsets[i]);
367 if (enable)
368 me_cntl &= ~SDMA0_F32_CNTL__HALT_MASK;
369 else
370 me_cntl |= SDMA0_F32_CNTL__HALT_MASK;
371 WREG32(mmSDMA0_F32_CNTL + sdma_offsets[i], me_cntl);
372 }
373}
374
375/**
376 * cik_sdma_gfx_resume - setup and start the async dma engines
377 *
378 * @adev: amdgpu_device pointer
379 *
380 * Set up the gfx DMA ring buffers and enable them (CIK).
381 * Returns 0 for success, error for failure.
382 */
383static int cik_sdma_gfx_resume(struct amdgpu_device *adev)
384{
385 struct amdgpu_ring *ring;
386 u32 rb_cntl, ib_cntl;
387 u32 rb_bufsz;
388 u32 wb_offset;
389 int i, j, r;
390
391 for (i = 0; i < SDMA_MAX_INSTANCE; i++) {
392 ring = &adev->sdma[i].ring;
393 wb_offset = (ring->rptr_offs * 4);
394
395 mutex_lock(&adev->srbm_mutex);
396 for (j = 0; j < 16; j++) {
397 cik_srbm_select(adev, 0, 0, 0, j);
398 /* SDMA GFX */
399 WREG32(mmSDMA0_GFX_VIRTUAL_ADDR + sdma_offsets[i], 0);
400 WREG32(mmSDMA0_GFX_APE1_CNTL + sdma_offsets[i], 0);
401 /* XXX SDMA RLC - todo */
402 }
403 cik_srbm_select(adev, 0, 0, 0, 0);
404 mutex_unlock(&adev->srbm_mutex);
405
406 WREG32(mmSDMA0_SEM_INCOMPLETE_TIMER_CNTL + sdma_offsets[i], 0);
407 WREG32(mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i], 0);
408
409 /* Set ring buffer size in dwords */
410 rb_bufsz = order_base_2(ring->ring_size / 4);
411 rb_cntl = rb_bufsz << 1;
412#ifdef __BIG_ENDIAN
413 rb_cntl |= SDMA_RB_SWAP_ENABLE | SDMA_RPTR_WRITEBACK_SWAP_ENABLE;
414#endif
415 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
416
417 /* Initialize the ring buffer's read and write pointers */
418 WREG32(mmSDMA0_GFX_RB_RPTR + sdma_offsets[i], 0);
419 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], 0);
420
421 /* set the wb address whether it's enabled or not */
422 WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_HI + sdma_offsets[i],
423 upper_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFF);
424 WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_LO + sdma_offsets[i],
425 ((adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC));
426
427 rb_cntl |= SDMA0_GFX_RB_CNTL__RPTR_WRITEBACK_ENABLE_MASK;
428
429 WREG32(mmSDMA0_GFX_RB_BASE + sdma_offsets[i], ring->gpu_addr >> 8);
430 WREG32(mmSDMA0_GFX_RB_BASE_HI + sdma_offsets[i], ring->gpu_addr >> 40);
431
432 ring->wptr = 0;
433 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], ring->wptr << 2);
434
435 /* enable DMA RB */
436 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i],
437 rb_cntl | SDMA0_GFX_RB_CNTL__RB_ENABLE_MASK);
438
439 ib_cntl = SDMA0_GFX_IB_CNTL__IB_ENABLE_MASK;
440#ifdef __BIG_ENDIAN
441 ib_cntl |= SDMA0_GFX_IB_CNTL__IB_SWAP_ENABLE_MASK;
442#endif
443 /* enable DMA IBs */
444 WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
445
446 ring->ready = true;
447
448 r = amdgpu_ring_test_ring(ring);
449 if (r) {
450 ring->ready = false;
451 return r;
452 }
453
454 if (adev->mman.buffer_funcs_ring == ring)
455 amdgpu_ttm_set_active_vram_size(adev, adev->mc.real_vram_size);
456 }
457
458 return 0;
459}
460
461/**
462 * cik_sdma_rlc_resume - setup and start the async dma engines
463 *
464 * @adev: amdgpu_device pointer
465 *
466 * Set up the compute DMA queues and enable them (CIK).
467 * Returns 0 for success, error for failure.
468 */
469static int cik_sdma_rlc_resume(struct amdgpu_device *adev)
470{
471 /* XXX todo */
472 return 0;
473}
474
475/**
476 * cik_sdma_load_microcode - load the sDMA ME ucode
477 *
478 * @adev: amdgpu_device pointer
479 *
480 * Loads the sDMA0/1 ucode.
481 * Returns 0 for success, -EINVAL if the ucode is not available.
482 */
483static int cik_sdma_load_microcode(struct amdgpu_device *adev)
484{
485 const struct sdma_firmware_header_v1_0 *hdr;
486 const __le32 *fw_data;
487 u32 fw_size;
488 int i, j;
489
490 if (!adev->sdma[0].fw || !adev->sdma[1].fw)
491 return -EINVAL;
492
493 /* halt the MEs */
494 cik_sdma_enable(adev, false);
495
496 for (i = 0; i < SDMA_MAX_INSTANCE; i++) {
497 hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma[i].fw->data;
498 amdgpu_ucode_print_sdma_hdr(&hdr->header);
499 fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
500 adev->sdma[i].fw_version = le32_to_cpu(hdr->header.ucode_version);
501 fw_data = (const __le32 *)
502 (adev->sdma[i].fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
503 WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], 0);
504 for (j = 0; j < fw_size; j++)
505 WREG32(mmSDMA0_UCODE_DATA + sdma_offsets[i], le32_to_cpup(fw_data++));
506 WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], adev->sdma[i].fw_version);
507 }
508
509 return 0;
510}
511
512/**
513 * cik_sdma_start - setup and start the async dma engines
514 *
515 * @adev: amdgpu_device pointer
516 *
517 * Set up the DMA engines and enable them (CIK).
518 * Returns 0 for success, error for failure.
519 */
520static int cik_sdma_start(struct amdgpu_device *adev)
521{
522 int r;
523
524 r = cik_sdma_load_microcode(adev);
525 if (r)
526 return r;
527
528 /* unhalt the MEs */
529 cik_sdma_enable(adev, true);
530
531 /* start the gfx rings and rlc compute queues */
532 r = cik_sdma_gfx_resume(adev);
533 if (r)
534 return r;
535 r = cik_sdma_rlc_resume(adev);
536 if (r)
537 return r;
538
539 return 0;
540}
541
542/**
543 * cik_sdma_ring_test_ring - simple async dma engine test
544 *
545 * @ring: amdgpu_ring structure holding ring information
546 *
547 * Test the DMA engine by writing using it to write an
548 * value to memory. (CIK).
549 * Returns 0 for success, error for failure.
550 */
551static int cik_sdma_ring_test_ring(struct amdgpu_ring *ring)
552{
553 struct amdgpu_device *adev = ring->adev;
554 unsigned i;
555 unsigned index;
556 int r;
557 u32 tmp;
558 u64 gpu_addr;
559
560 r = amdgpu_wb_get(adev, &index);
561 if (r) {
562 dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
563 return r;
564 }
565
566 gpu_addr = adev->wb.gpu_addr + (index * 4);
567 tmp = 0xCAFEDEAD;
568 adev->wb.wb[index] = cpu_to_le32(tmp);
569
570 r = amdgpu_ring_lock(ring, 5);
571 if (r) {
572 DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
573 amdgpu_wb_free(adev, index);
574 return r;
575 }
576 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_WRITE, SDMA_WRITE_SUB_OPCODE_LINEAR, 0));
577 amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
578 amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
579 amdgpu_ring_write(ring, 1); /* number of DWs to follow */
580 amdgpu_ring_write(ring, 0xDEADBEEF);
581 amdgpu_ring_unlock_commit(ring);
582
583 for (i = 0; i < adev->usec_timeout; i++) {
584 tmp = le32_to_cpu(adev->wb.wb[index]);
585 if (tmp == 0xDEADBEEF)
586 break;
587 DRM_UDELAY(1);
588 }
589
590 if (i < adev->usec_timeout) {
591 DRM_INFO("ring test on %d succeeded in %d usecs\n", ring->idx, i);
592 } else {
593 DRM_ERROR("amdgpu: ring %d test failed (0x%08X)\n",
594 ring->idx, tmp);
595 r = -EINVAL;
596 }
597 amdgpu_wb_free(adev, index);
598
599 return r;
600}
601
602/**
603 * cik_sdma_ring_test_ib - test an IB on the DMA engine
604 *
605 * @ring: amdgpu_ring structure holding ring information
606 *
607 * Test a simple IB in the DMA ring (CIK).
608 * Returns 0 on success, error on failure.
609 */
610static int cik_sdma_ring_test_ib(struct amdgpu_ring *ring)
611{
612 struct amdgpu_device *adev = ring->adev;
613 struct amdgpu_ib ib;
614 unsigned i;
615 unsigned index;
616 int r;
617 u32 tmp = 0;
618 u64 gpu_addr;
619
620 r = amdgpu_wb_get(adev, &index);
621 if (r) {
622 dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
623 return r;
624 }
625
626 gpu_addr = adev->wb.gpu_addr + (index * 4);
627 tmp = 0xCAFEDEAD;
628 adev->wb.wb[index] = cpu_to_le32(tmp);
629
630 r = amdgpu_ib_get(ring, NULL, 256, &ib);
631 if (r) {
632 amdgpu_wb_free(adev, index);
633 DRM_ERROR("amdgpu: failed to get ib (%d).\n", r);
634 return r;
635 }
636
637 ib.ptr[0] = SDMA_PACKET(SDMA_OPCODE_WRITE, SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
638 ib.ptr[1] = lower_32_bits(gpu_addr);
639 ib.ptr[2] = upper_32_bits(gpu_addr);
640 ib.ptr[3] = 1;
641 ib.ptr[4] = 0xDEADBEEF;
642 ib.length_dw = 5;
643
644 r = amdgpu_ib_schedule(adev, 1, &ib, AMDGPU_FENCE_OWNER_UNDEFINED);
645 if (r) {
646 amdgpu_ib_free(adev, &ib);
647 amdgpu_wb_free(adev, index);
648 DRM_ERROR("amdgpu: failed to schedule ib (%d).\n", r);
649 return r;
650 }
651 r = amdgpu_fence_wait(ib.fence, false);
652 if (r) {
653 amdgpu_ib_free(adev, &ib);
654 amdgpu_wb_free(adev, index);
655 DRM_ERROR("amdgpu: fence wait failed (%d).\n", r);
656 return r;
657 }
658 for (i = 0; i < adev->usec_timeout; i++) {
659 tmp = le32_to_cpu(adev->wb.wb[index]);
660 if (tmp == 0xDEADBEEF)
661 break;
662 DRM_UDELAY(1);
663 }
664 if (i < adev->usec_timeout) {
665 DRM_INFO("ib test on ring %d succeeded in %u usecs\n",
666 ib.fence->ring->idx, i);
667 } else {
668 DRM_ERROR("amdgpu: ib test failed (0x%08X)\n", tmp);
669 r = -EINVAL;
670 }
671 amdgpu_ib_free(adev, &ib);
672 amdgpu_wb_free(adev, index);
673 return r;
674}
675
676/**
677 * cik_sdma_vm_copy_pages - update PTEs by copying them from the GART
678 *
679 * @ib: indirect buffer to fill with commands
680 * @pe: addr of the page entry
681 * @src: src addr to copy from
682 * @count: number of page entries to update
683 *
684 * Update PTEs by copying them from the GART using sDMA (CIK).
685 */
686static void cik_sdma_vm_copy_pte(struct amdgpu_ib *ib,
687 uint64_t pe, uint64_t src,
688 unsigned count)
689{
690 while (count) {
691 unsigned bytes = count * 8;
692 if (bytes > 0x1FFFF8)
693 bytes = 0x1FFFF8;
694
695 ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_COPY,
696 SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
697 ib->ptr[ib->length_dw++] = bytes;
698 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
699 ib->ptr[ib->length_dw++] = lower_32_bits(src);
700 ib->ptr[ib->length_dw++] = upper_32_bits(src);
701 ib->ptr[ib->length_dw++] = lower_32_bits(pe);
702 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
703
704 pe += bytes;
705 src += bytes;
706 count -= bytes / 8;
707 }
708}
709
710/**
711 * cik_sdma_vm_write_pages - update PTEs by writing them manually
712 *
713 * @ib: indirect buffer to fill with commands
714 * @pe: addr of the page entry
715 * @addr: dst addr to write into pe
716 * @count: number of page entries to update
717 * @incr: increase next addr by incr bytes
718 * @flags: access flags
719 *
720 * Update PTEs by writing them manually using sDMA (CIK).
721 */
722static void cik_sdma_vm_write_pte(struct amdgpu_ib *ib,
723 uint64_t pe,
724 uint64_t addr, unsigned count,
725 uint32_t incr, uint32_t flags)
726{
727 uint64_t value;
728 unsigned ndw;
729
730 while (count) {
731 ndw = count * 2;
732 if (ndw > 0xFFFFE)
733 ndw = 0xFFFFE;
734
735 /* for non-physically contiguous pages (system) */
736 ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_WRITE,
737 SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
738 ib->ptr[ib->length_dw++] = pe;
739 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
740 ib->ptr[ib->length_dw++] = ndw;
741 for (; ndw > 0; ndw -= 2, --count, pe += 8) {
742 if (flags & AMDGPU_PTE_SYSTEM) {
743 value = amdgpu_vm_map_gart(ib->ring->adev, addr);
744 value &= 0xFFFFFFFFFFFFF000ULL;
745 } else if (flags & AMDGPU_PTE_VALID) {
746 value = addr;
747 } else {
748 value = 0;
749 }
750 addr += incr;
751 value |= flags;
752 ib->ptr[ib->length_dw++] = value;
753 ib->ptr[ib->length_dw++] = upper_32_bits(value);
754 }
755 }
756}
757
758/**
759 * cik_sdma_vm_set_pages - update the page tables using sDMA
760 *
761 * @ib: indirect buffer to fill with commands
762 * @pe: addr of the page entry
763 * @addr: dst addr to write into pe
764 * @count: number of page entries to update
765 * @incr: increase next addr by incr bytes
766 * @flags: access flags
767 *
768 * Update the page tables using sDMA (CIK).
769 */
770static void cik_sdma_vm_set_pte_pde(struct amdgpu_ib *ib,
771 uint64_t pe,
772 uint64_t addr, unsigned count,
773 uint32_t incr, uint32_t flags)
774{
775 uint64_t value;
776 unsigned ndw;
777
778 while (count) {
779 ndw = count;
780 if (ndw > 0x7FFFF)
781 ndw = 0x7FFFF;
782
783 if (flags & AMDGPU_PTE_VALID)
784 value = addr;
785 else
786 value = 0;
787
788 /* for physically contiguous pages (vram) */
789 ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_GENERATE_PTE_PDE, 0, 0);
790 ib->ptr[ib->length_dw++] = pe; /* dst addr */
791 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
792 ib->ptr[ib->length_dw++] = flags; /* mask */
793 ib->ptr[ib->length_dw++] = 0;
794 ib->ptr[ib->length_dw++] = value; /* value */
795 ib->ptr[ib->length_dw++] = upper_32_bits(value);
796 ib->ptr[ib->length_dw++] = incr; /* increment size */
797 ib->ptr[ib->length_dw++] = 0;
798 ib->ptr[ib->length_dw++] = ndw; /* number of entries */
799
800 pe += ndw * 8;
801 addr += ndw * incr;
802 count -= ndw;
803 }
804}
805
806/**
807 * cik_sdma_vm_pad_ib - pad the IB to the required number of dw
808 *
809 * @ib: indirect buffer to fill with padding
810 *
811 */
812static void cik_sdma_vm_pad_ib(struct amdgpu_ib *ib)
813{
814 while (ib->length_dw & 0x7)
815 ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0);
816}
817
818/**
819 * cik_sdma_ring_emit_vm_flush - cik vm flush using sDMA
820 *
821 * @ring: amdgpu_ring pointer
822 * @vm: amdgpu_vm pointer
823 *
824 * Update the page table base and flush the VM TLB
825 * using sDMA (CIK).
826 */
827static void cik_sdma_ring_emit_vm_flush(struct amdgpu_ring *ring,
828 unsigned vm_id, uint64_t pd_addr)
829{
830 u32 extra_bits = (SDMA_POLL_REG_MEM_EXTRA_OP(0) |
831 SDMA_POLL_REG_MEM_EXTRA_FUNC(0)); /* always */
832
833 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
834 if (vm_id < 8) {
835 amdgpu_ring_write(ring, (mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + vm_id));
836 } else {
837 amdgpu_ring_write(ring, (mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + vm_id - 8));
838 }
839 amdgpu_ring_write(ring, pd_addr >> 12);
840
a2e73f56
AD		 841 /* flush TLB */
842 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
843 amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST);
844 amdgpu_ring_write(ring, 1 << vm_id);
845
846 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0, extra_bits));
847 amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST << 2);
848 amdgpu_ring_write(ring, 0);
849 amdgpu_ring_write(ring, 0); /* reference */
850 amdgpu_ring_write(ring, 0); /* mask */
851 amdgpu_ring_write(ring, (0xfff << 16) | 10); /* retry count, poll interval */
852}
853
854static void cik_enable_sdma_mgcg(struct amdgpu_device *adev,
855 bool enable)
856{
857 u32 orig, data;
858
859 if (enable && (adev->cg_flags & AMDGPU_CG_SUPPORT_SDMA_MGCG)) {
860 WREG32(mmSDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET, 0x00000100);
861 WREG32(mmSDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET, 0x00000100);
862 } else {
863 orig = data = RREG32(mmSDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET);
864 data |= 0xff000000;
865 if (data != orig)
866 WREG32(mmSDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET, data);
867
868 orig = data = RREG32(mmSDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET);
869 data |= 0xff000000;
870 if (data != orig)
871 WREG32(mmSDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET, data);
872 }
873}
874
875static void cik_enable_sdma_mgls(struct amdgpu_device *adev,
876 bool enable)
877{
878 u32 orig, data;
879
880 if (enable && (adev->cg_flags & AMDGPU_CG_SUPPORT_SDMA_LS)) {
881 orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET);
882 data |= 0x100;
883 if (orig != data)
884 WREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET, data);
885
886 orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET);
887 data |= 0x100;
888 if (orig != data)
889 WREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET, data);
890 } else {
891 orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET);
892 data &= ~0x100;
893 if (orig != data)
894 WREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET, data);
895
896 orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET);
897 data &= ~0x100;
898 if (orig != data)
899 WREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET, data);
900 }
901}
902
5fc3aeeb 903static int cik_sdma_early_init(void *handle)
a2e73f56 904{
5fc3aeeb 905 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
906
a2e73f56
AD		 907 cik_sdma_set_ring_funcs(adev);
908 cik_sdma_set_irq_funcs(adev);
909 cik_sdma_set_buffer_funcs(adev);
910 cik_sdma_set_vm_pte_funcs(adev);
911
912 return 0;
913}
914
5fc3aeeb 915static int cik_sdma_sw_init(void *handle)
a2e73f56
AD		 916{
917 struct amdgpu_ring *ring;
5fc3aeeb 918 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
a2e73f56
AD		 919 int r;
920
921 r = cik_sdma_init_microcode(adev);
922 if (r) {
923 DRM_ERROR("Failed to load sdma firmware!\n");
924 return r;
925 }
926
927 /* SDMA trap event */
928 r = amdgpu_irq_add_id(adev, 224, &adev->sdma_trap_irq);
929 if (r)
930 return r;
931
932 /* SDMA Privileged inst */
933 r = amdgpu_irq_add_id(adev, 241, &adev->sdma_illegal_inst_irq);
934 if (r)
935 return r;
936
937 /* SDMA Privileged inst */
938 r = amdgpu_irq_add_id(adev, 247, &adev->sdma_illegal_inst_irq);
939 if (r)
940 return r;
941
942 ring = &adev->sdma[0].ring;
943 ring->ring_obj = NULL;
944
945 ring = &adev->sdma[1].ring;
946 ring->ring_obj = NULL;
947
948 ring = &adev->sdma[0].ring;
949 sprintf(ring->name, "sdma0");
950 r = amdgpu_ring_init(adev, ring, 256 * 1024,
951 SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0), 0xf,
952 &adev->sdma_trap_irq, AMDGPU_SDMA_IRQ_TRAP0,
953 AMDGPU_RING_TYPE_SDMA);
954 if (r)
955 return r;
956
957 ring = &adev->sdma[1].ring;
958 sprintf(ring->name, "sdma1");
959 r = amdgpu_ring_init(adev, ring, 256 * 1024,
960 SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0), 0xf,
961 &adev->sdma_trap_irq, AMDGPU_SDMA_IRQ_TRAP1,
962 AMDGPU_RING_TYPE_SDMA);
963 if (r)
964 return r;
965
966 return r;
967}
968
5fc3aeeb 969static int cik_sdma_sw_fini(void *handle)
a2e73f56 970{
5fc3aeeb 971 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
972
a2e73f56
AD		 973 amdgpu_ring_fini(&adev->sdma[0].ring);
974 amdgpu_ring_fini(&adev->sdma[1].ring);
975
976 return 0;
977}
978
5fc3aeeb 979static int cik_sdma_hw_init(void *handle)
a2e73f56
AD		 980{
981 int r;
5fc3aeeb 982 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
a2e73f56
AD		 983
984 r = cik_sdma_start(adev);
985 if (r)
986 return r;
987
988 return r;
989}
990
5fc3aeeb 991static int cik_sdma_hw_fini(void *handle)
a2e73f56 992{
5fc3aeeb 993 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
994
a2e73f56
AD		 995 cik_sdma_enable(adev, false);
996
997 return 0;
998}
999
5fc3aeeb 1000static int cik_sdma_suspend(void *handle)
a2e73f56 1001{
5fc3aeeb 1002 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
a2e73f56
AD		 1003
1004 return cik_sdma_hw_fini(adev);
1005}
1006
5fc3aeeb 1007static int cik_sdma_resume(void *handle)
a2e73f56 1008{
5fc3aeeb 1009 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
a2e73f56
AD		 1010
1011 return cik_sdma_hw_init(adev);
1012}
1013
5fc3aeeb 1014static bool cik_sdma_is_idle(void *handle)
a2e73f56 1015{
5fc3aeeb 1016 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
a2e73f56
AD		 1017 u32 tmp = RREG32(mmSRBM_STATUS2);
1018
1019 if (tmp & (SRBM_STATUS2__SDMA_BUSY_MASK |
1020 SRBM_STATUS2__SDMA1_BUSY_MASK))
1021 return false;
1022
1023 return true;
1024}
1025
5fc3aeeb 1026static int cik_sdma_wait_for_idle(void *handle)
a2e73f56
AD		 1027{
1028 unsigned i;
1029 u32 tmp;
5fc3aeeb 1030 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
a2e73f56
AD		 1031
1032 for (i = 0; i < adev->usec_timeout; i++) {
1033 tmp = RREG32(mmSRBM_STATUS2) & (SRBM_STATUS2__SDMA_BUSY_MASK |
1034 SRBM_STATUS2__SDMA1_BUSY_MASK);
1035
1036 if (!tmp)
1037 return 0;
1038 udelay(1);
1039 }
1040 return -ETIMEDOUT;
1041}
1042
5fc3aeeb 1043static void cik_sdma_print_status(void *handle)
a2e73f56
AD		 1044{
1045 int i, j;
5fc3aeeb 1046 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
a2e73f56
AD		 1047
1048 dev_info(adev->dev, "CIK SDMA registers\n");
1049 dev_info(adev->dev, " SRBM_STATUS2=0x%08X\n",
1050 RREG32(mmSRBM_STATUS2));
1051 for (i = 0; i < SDMA_MAX_INSTANCE; i++) {
1052 dev_info(adev->dev, " SDMA%d_STATUS_REG=0x%08X\n",
1053 i, RREG32(mmSDMA0_STATUS_REG + sdma_offsets[i]));
1054 dev_info(adev->dev, " SDMA%d_ME_CNTL=0x%08X\n",
1055 i, RREG32(mmSDMA0_F32_CNTL + sdma_offsets[i]));
1056 dev_info(adev->dev, " SDMA%d_CNTL=0x%08X\n",
1057 i, RREG32(mmSDMA0_CNTL + sdma_offsets[i]));
1058 dev_info(adev->dev, " SDMA%d_SEM_INCOMPLETE_TIMER_CNTL=0x%08X\n",
1059 i, RREG32(mmSDMA0_SEM_INCOMPLETE_TIMER_CNTL + sdma_offsets[i]));
1060 dev_info(adev->dev, " SDMA%d_SEM_WAIT_FAIL_TIMER_CNTL=0x%08X\n",
1061 i, RREG32(mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i]));
1062 dev_info(adev->dev, " SDMA%d_GFX_IB_CNTL=0x%08X\n",
1063 i, RREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i]));
1064 dev_info(adev->dev, " SDMA%d_GFX_RB_CNTL=0x%08X\n",
1065 i, RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]));
1066 dev_info(adev->dev, " SDMA%d_GFX_RB_RPTR=0x%08X\n",
1067 i, RREG32(mmSDMA0_GFX_RB_RPTR + sdma_offsets[i]));
1068 dev_info(adev->dev, " SDMA%d_GFX_RB_WPTR=0x%08X\n",
1069 i, RREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i]));
1070 dev_info(adev->dev, " SDMA%d_GFX_RB_RPTR_ADDR_HI=0x%08X\n",
1071 i, RREG32(mmSDMA0_GFX_RB_RPTR_ADDR_HI + sdma_offsets[i]));
1072 dev_info(adev->dev, " SDMA%d_GFX_RB_RPTR_ADDR_LO=0x%08X\n",
1073 i, RREG32(mmSDMA0_GFX_RB_RPTR_ADDR_LO + sdma_offsets[i]));
1074 dev_info(adev->dev, " SDMA%d_GFX_RB_BASE=0x%08X\n",
1075 i, RREG32(mmSDMA0_GFX_RB_BASE + sdma_offsets[i]));
1076 dev_info(adev->dev, " SDMA%d_GFX_RB_BASE_HI=0x%08X\n",
1077 i, RREG32(mmSDMA0_GFX_RB_BASE_HI + sdma_offsets[i]));
1078 mutex_lock(&adev->srbm_mutex);
1079 for (j = 0; j < 16; j++) {
1080 cik_srbm_select(adev, 0, 0, 0, j);
1081 dev_info(adev->dev, " VM %d:\n", j);
1082 dev_info(adev->dev, " SDMA0_GFX_VIRTUAL_ADDR=0x%08X\n",
1083 RREG32(mmSDMA0_GFX_VIRTUAL_ADDR + sdma_offsets[i]));
1084 dev_info(adev->dev, " SDMA0_GFX_APE1_CNTL=0x%08X\n",
1085 RREG32(mmSDMA0_GFX_APE1_CNTL + sdma_offsets[i]));
1086 }
1087 cik_srbm_select(adev, 0, 0, 0, 0);
1088 mutex_unlock(&adev->srbm_mutex);
1089 }
1090}
1091
5fc3aeeb 1092static int cik_sdma_soft_reset(void *handle)
a2e73f56
AD		 1093{
1094 u32 srbm_soft_reset = 0;
5fc3aeeb 1095 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
a2e73f56
AD		 1096 u32 tmp = RREG32(mmSRBM_STATUS2);
1097
1098 if (tmp & SRBM_STATUS2__SDMA_BUSY_MASK) {
1099 /* sdma0 */
1100 tmp = RREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET);
1101 tmp |= SDMA0_F32_CNTL__HALT_MASK;
1102 WREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET, tmp);
1103 srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA_MASK;
1104 }
1105 if (tmp & SRBM_STATUS2__SDMA1_BUSY_MASK) {
1106 /* sdma1 */
1107 tmp = RREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET);
1108 tmp |= SDMA0_F32_CNTL__HALT_MASK;
1109 WREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET, tmp);
1110 srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA1_MASK;
1111 }
1112
1113 if (srbm_soft_reset) {
5fc3aeeb 1114 cik_sdma_print_status((void *)adev);
a2e73f56
AD		 1115
1116 tmp = RREG32(mmSRBM_SOFT_RESET);
1117 tmp |= srbm_soft_reset;
1118 dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
1119 WREG32(mmSRBM_SOFT_RESET, tmp);
1120 tmp = RREG32(mmSRBM_SOFT_RESET);
1121
1122 udelay(50);
1123
1124 tmp &= ~srbm_soft_reset;
1125 WREG32(mmSRBM_SOFT_RESET, tmp);
1126 tmp = RREG32(mmSRBM_SOFT_RESET);
1127
1128 /* Wait a little for things to settle down */
1129 udelay(50);
1130
5fc3aeeb 1131 cik_sdma_print_status((void *)adev);
a2e73f56
AD		 1132 }
1133
1134 return 0;
1135}
1136
1137static int cik_sdma_set_trap_irq_state(struct amdgpu_device *adev,
1138 struct amdgpu_irq_src *src,
1139 unsigned type,
1140 enum amdgpu_interrupt_state state)
1141{
1142 u32 sdma_cntl;
1143
1144 switch (type) {
1145 case AMDGPU_SDMA_IRQ_TRAP0:
1146 switch (state) {
1147 case AMDGPU_IRQ_STATE_DISABLE:
1148 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1149 sdma_cntl &= ~SDMA0_CNTL__TRAP_ENABLE_MASK;
1150 WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1151 break;
1152 case AMDGPU_IRQ_STATE_ENABLE:
1153 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1154 sdma_cntl |= SDMA0_CNTL__TRAP_ENABLE_MASK;
1155 WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1156 break;
1157 default:
1158 break;
1159 }
1160 break;
1161 case AMDGPU_SDMA_IRQ_TRAP1:
1162 switch (state) {
1163 case AMDGPU_IRQ_STATE_DISABLE:
1164 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1165 sdma_cntl &= ~SDMA0_CNTL__TRAP_ENABLE_MASK;
1166 WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1167 break;
1168 case AMDGPU_IRQ_STATE_ENABLE:
1169 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1170 sdma_cntl |= SDMA0_CNTL__TRAP_ENABLE_MASK;
1171 WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1172 break;
1173 default:
1174 break;
1175 }
1176 break;
1177 default:
1178 break;
1179 }
1180 return 0;
1181}
1182
1183static int cik_sdma_process_trap_irq(struct amdgpu_device *adev,
1184 struct amdgpu_irq_src *source,
1185 struct amdgpu_iv_entry *entry)
1186{
1187 u8 instance_id, queue_id;
1188
1189 instance_id = (entry->ring_id & 0x3) >> 0;
1190 queue_id = (entry->ring_id & 0xc) >> 2;
1191 DRM_DEBUG("IH: SDMA trap\n");
1192 switch (instance_id) {
1193 case 0:
1194 switch (queue_id) {
1195 case 0:
1196 amdgpu_fence_process(&adev->sdma[0].ring);
1197 break;
1198 case 1:
1199 /* XXX compute */
1200 break;
1201 case 2:
1202 /* XXX compute */
1203 break;
1204 }
1205 break;
1206 case 1:
1207 switch (queue_id) {
1208 case 0:
1209 amdgpu_fence_process(&adev->sdma[1].ring);
1210 break;
1211 case 1:
1212 /* XXX compute */
1213 break;
1214 case 2:
1215 /* XXX compute */
1216 break;
1217 }
1218 break;
1219 }
1220
1221 return 0;
1222}
1223
1224static int cik_sdma_process_illegal_inst_irq(struct amdgpu_device *adev,
1225 struct amdgpu_irq_src *source,
1226 struct amdgpu_iv_entry *entry)
1227{
1228 DRM_ERROR("Illegal instruction in SDMA command stream\n");
1229 schedule_work(&adev->reset_work);
1230 return 0;
1231}
1232
5fc3aeeb 1233static int cik_sdma_set_clockgating_state(void *handle,
1234 enum amd_clockgating_state state)
a2e73f56
AD		 1235{
1236 bool gate = false;
5fc3aeeb 1237 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
a2e73f56 1238
5fc3aeeb 1239 if (state == AMD_CG_STATE_GATE)
a2e73f56
AD		 1240 gate = true;
1241
1242 cik_enable_sdma_mgcg(adev, gate);
1243 cik_enable_sdma_mgls(adev, gate);
1244
1245 return 0;
1246}
1247
5fc3aeeb 1248static int cik_sdma_set_powergating_state(void *handle,
1249 enum amd_powergating_state state)
a2e73f56
AD		 1250{
1251 return 0;
1252}
1253
5fc3aeeb 1254const struct amd_ip_funcs cik_sdma_ip_funcs = {
a2e73f56
AD		 1255 .early_init = cik_sdma_early_init,
1256 .late_init = NULL,
1257 .sw_init = cik_sdma_sw_init,
1258 .sw_fini = cik_sdma_sw_fini,
1259 .hw_init = cik_sdma_hw_init,
1260 .hw_fini = cik_sdma_hw_fini,
1261 .suspend = cik_sdma_suspend,
1262 .resume = cik_sdma_resume,
1263 .is_idle = cik_sdma_is_idle,
1264 .wait_for_idle = cik_sdma_wait_for_idle,
1265 .soft_reset = cik_sdma_soft_reset,
1266 .print_status = cik_sdma_print_status,
1267 .set_clockgating_state = cik_sdma_set_clockgating_state,
1268 .set_powergating_state = cik_sdma_set_powergating_state,
1269};
1270
1271/**
1272 * cik_sdma_ring_is_lockup - Check if the DMA engine is locked up
1273 *
1274 * @ring: amdgpu_ring structure holding ring information
1275 *
1276 * Check if the async DMA engine is locked up (CIK).
1277 * Returns true if the engine appears to be locked up, false if not.
1278 */
1279static bool cik_sdma_ring_is_lockup(struct amdgpu_ring *ring)
1280{
1281
1282 if (cik_sdma_is_idle(ring->adev)) {
1283 amdgpu_ring_lockup_update(ring);
1284 return false;
1285 }
1286 return amdgpu_ring_test_lockup(ring);
1287}
1288
1289static const struct amdgpu_ring_funcs cik_sdma_ring_funcs = {
1290 .get_rptr = cik_sdma_ring_get_rptr,
1291 .get_wptr = cik_sdma_ring_get_wptr,
1292 .set_wptr = cik_sdma_ring_set_wptr,
1293 .parse_cs = NULL,
1294 .emit_ib = cik_sdma_ring_emit_ib,
1295 .emit_fence = cik_sdma_ring_emit_fence,
1296 .emit_semaphore = cik_sdma_ring_emit_semaphore,
1297 .emit_vm_flush = cik_sdma_ring_emit_vm_flush,
d2edb07b 1298 .emit_hdp_flush = cik_sdma_ring_emit_hdp_flush,
a2e73f56
AD		 1299 .test_ring = cik_sdma_ring_test_ring,
1300 .test_ib = cik_sdma_ring_test_ib,
1301 .is_lockup = cik_sdma_ring_is_lockup,
1302};
1303
1304static void cik_sdma_set_ring_funcs(struct amdgpu_device *adev)
1305{
1306 adev->sdma[0].ring.funcs = &cik_sdma_ring_funcs;
1307 adev->sdma[1].ring.funcs = &cik_sdma_ring_funcs;
1308}
1309
1310static const struct amdgpu_irq_src_funcs cik_sdma_trap_irq_funcs = {
1311 .set = cik_sdma_set_trap_irq_state,
1312 .process = cik_sdma_process_trap_irq,
1313};
1314
1315static const struct amdgpu_irq_src_funcs cik_sdma_illegal_inst_irq_funcs = {
1316 .process = cik_sdma_process_illegal_inst_irq,
1317};
1318
1319static void cik_sdma_set_irq_funcs(struct amdgpu_device *adev)
1320{
1321 adev->sdma_trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST;
1322 adev->sdma_trap_irq.funcs = &cik_sdma_trap_irq_funcs;
1323 adev->sdma_illegal_inst_irq.funcs = &cik_sdma_illegal_inst_irq_funcs;
1324}
1325
1326/**
1327 * cik_sdma_emit_copy_buffer - copy buffer using the sDMA engine
1328 *
1329 * @ring: amdgpu_ring structure holding ring information
1330 * @src_offset: src GPU address
1331 * @dst_offset: dst GPU address
1332 * @byte_count: number of bytes to xfer
1333 *
1334 * Copy GPU buffers using the DMA engine (CIK).
1335 * Used by the amdgpu ttm implementation to move pages if
1336 * registered as the asic copy callback.
1337 */
1338static void cik_sdma_emit_copy_buffer(struct amdgpu_ring *ring,
1339 uint64_t src_offset,
1340 uint64_t dst_offset,
1341 uint32_t byte_count)
1342{
1343 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_COPY, SDMA_COPY_SUB_OPCODE_LINEAR, 0));
1344 amdgpu_ring_write(ring, byte_count);
1345 amdgpu_ring_write(ring, 0); /* src/dst endian swap */
1346 amdgpu_ring_write(ring, lower_32_bits(src_offset));
1347 amdgpu_ring_write(ring, upper_32_bits(src_offset));
1348 amdgpu_ring_write(ring, lower_32_bits(dst_offset));
1349 amdgpu_ring_write(ring, upper_32_bits(dst_offset));
1350}
1351
1352/**
1353 * cik_sdma_emit_fill_buffer - fill buffer using the sDMA engine
1354 *
1355 * @ring: amdgpu_ring structure holding ring information
1356 * @src_data: value to write to buffer
1357 * @dst_offset: dst GPU address
1358 * @byte_count: number of bytes to xfer
1359 *
1360 * Fill GPU buffers using the DMA engine (CIK).
1361 */
1362static void cik_sdma_emit_fill_buffer(struct amdgpu_ring *ring,
1363 uint32_t src_data,
1364 uint64_t dst_offset,
1365 uint32_t byte_count)
1366{
1367 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_CONSTANT_FILL, 0, 0));
1368 amdgpu_ring_write(ring, lower_32_bits(dst_offset));
1369 amdgpu_ring_write(ring, upper_32_bits(dst_offset));
1370 amdgpu_ring_write(ring, src_data);
1371 amdgpu_ring_write(ring, byte_count);
1372}
1373
1374static const struct amdgpu_buffer_funcs cik_sdma_buffer_funcs = {
1375 .copy_max_bytes = 0x1fffff,
1376 .copy_num_dw = 7,
1377 .emit_copy_buffer = cik_sdma_emit_copy_buffer,
1378
1379 .fill_max_bytes = 0x1fffff,
1380 .fill_num_dw = 5,
1381 .emit_fill_buffer = cik_sdma_emit_fill_buffer,
1382};
1383
1384static void cik_sdma_set_buffer_funcs(struct amdgpu_device *adev)
1385{
1386 if (adev->mman.buffer_funcs == NULL) {
1387 adev->mman.buffer_funcs = &cik_sdma_buffer_funcs;
1388 adev->mman.buffer_funcs_ring = &adev->sdma[0].ring;
1389 }
1390}
1391
1392static const struct amdgpu_vm_pte_funcs cik_sdma_vm_pte_funcs = {
1393 .copy_pte = cik_sdma_vm_copy_pte,
1394 .write_pte = cik_sdma_vm_write_pte,
1395 .set_pte_pde = cik_sdma_vm_set_pte_pde,
1396 .pad_ib = cik_sdma_vm_pad_ib,
1397};
1398
1399static void cik_sdma_set_vm_pte_funcs(struct amdgpu_device *adev)
1400{
1401 if (adev->vm_manager.vm_pte_funcs == NULL) {
1402 adev->vm_manager.vm_pte_funcs = &cik_sdma_vm_pte_funcs;
1403 adev->vm_manager.vm_pte_funcs_ring = &adev->sdma[0].ring;
1404 }
1405}
Linux kernel - Deliverables
This page took 0.088801 seconds and 5 git commands to generate.