blob: 700e26b69abca9ec1870e374b6c91d52d8d1e900 [file] [log] [blame]
Googler9398cc32022-12-02 17:21:52 +08001/*
2 * Copyright 2008 Advanced Micro Devices, Inc.
3 * Copyright 2008 Red Hat Inc.
4 * Copyright 2009 Jerome Glisse.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
22 * OTHER DEALINGS IN THE SOFTWARE.
23 *
24 */
25
26#include <linux/kthread.h>
27#include <linux/pci.h>
28#include <linux/uaccess.h>
29
30#include <drm/drm_debugfs.h>
31
32#include "amdgpu.h"
33
34/**
35 * amdgpu_debugfs_add_files - Add simple debugfs entries
36 *
37 * @adev: Device to attach debugfs entries to
38 * @files: Array of function callbacks that respond to reads
39 * @nfiles: Number of callbacks to register
40 *
41 */
42int amdgpu_debugfs_add_files(struct amdgpu_device *adev,
43 const struct drm_info_list *files,
44 unsigned nfiles)
45{
46 unsigned i;
47
48 for (i = 0; i < adev->debugfs_count; i++) {
49 if (adev->debugfs[i].files == files) {
50 /* Already registered */
51 return 0;
52 }
53 }
54
55 i = adev->debugfs_count + 1;
56 if (i > AMDGPU_DEBUGFS_MAX_COMPONENTS) {
57 DRM_ERROR("Reached maximum number of debugfs components.\n");
58 DRM_ERROR("Report so we increase "
59 "AMDGPU_DEBUGFS_MAX_COMPONENTS.\n");
60 return -EINVAL;
61 }
62 adev->debugfs[adev->debugfs_count].files = files;
63 adev->debugfs[adev->debugfs_count].num_files = nfiles;
64 adev->debugfs_count = i;
65#if defined(CONFIG_DEBUG_FS)
66 drm_debugfs_create_files(files, nfiles,
67 adev->ddev->primary->debugfs_root,
68 adev->ddev->primary);
69#endif
70 return 0;
71}
72
73#if defined(CONFIG_DEBUG_FS)
74
75/**
76 * amdgpu_debugfs_process_reg_op - Handle MMIO register reads/writes
77 *
78 * @read: True if reading
79 * @f: open file handle
80 * @buf: User buffer to write/read to
81 * @size: Number of bytes to write/read
82 * @pos: Offset to seek to
83 *
84 * This debugfs entry has special meaning on the offset being sought.
85 * Various bits have different meanings:
86 *
87 * Bit 62: Indicates a GRBM bank switch is needed
88 * Bit 61: Indicates a SRBM bank switch is needed (implies bit 62 is
89 * zero)
90 * Bits 24..33: The SE or ME selector if needed
91 * Bits 34..43: The SH (or SA) or PIPE selector if needed
92 * Bits 44..53: The INSTANCE (or CU/WGP) or QUEUE selector if needed
93 *
94 * Bit 23: Indicates that the PM power gating lock should be held
95 * This is necessary to read registers that might be
96 * unreliable during a power gating transistion.
97 *
98 * The lower bits are the BYTE offset of the register to read. This
99 * allows reading multiple registers in a single call and having
100 * the returned size reflect that.
101 */
102static int amdgpu_debugfs_process_reg_op(bool read, struct file *f,
103 char __user *buf, size_t size, loff_t *pos)
104{
105 struct amdgpu_device *adev = file_inode(f)->i_private;
106 ssize_t result = 0;
107 int r;
108 bool pm_pg_lock, use_bank, use_ring;
109 unsigned instance_bank, sh_bank, se_bank, me, pipe, queue, vmid;
110
111 pm_pg_lock = use_bank = use_ring = false;
112 instance_bank = sh_bank = se_bank = me = pipe = queue = vmid = 0;
113
114 if (size & 0x3 || *pos & 0x3 ||
115 ((*pos & (1ULL << 62)) && (*pos & (1ULL << 61))))
116 return -EINVAL;
117
118 /* are we reading registers for which a PG lock is necessary? */
119 pm_pg_lock = (*pos >> 23) & 1;
120
121 if (*pos & (1ULL << 62)) {
122 se_bank = (*pos & GENMASK_ULL(33, 24)) >> 24;
123 sh_bank = (*pos & GENMASK_ULL(43, 34)) >> 34;
124 instance_bank = (*pos & GENMASK_ULL(53, 44)) >> 44;
125
126 if (se_bank == 0x3FF)
127 se_bank = 0xFFFFFFFF;
128 if (sh_bank == 0x3FF)
129 sh_bank = 0xFFFFFFFF;
130 if (instance_bank == 0x3FF)
131 instance_bank = 0xFFFFFFFF;
132 use_bank = 1;
133 } else if (*pos & (1ULL << 61)) {
134
135 me = (*pos & GENMASK_ULL(33, 24)) >> 24;
136 pipe = (*pos & GENMASK_ULL(43, 34)) >> 34;
137 queue = (*pos & GENMASK_ULL(53, 44)) >> 44;
138 vmid = (*pos & GENMASK_ULL(58, 54)) >> 54;
139
140 use_ring = 1;
141 } else {
142 use_bank = use_ring = 0;
143 }
144
145 *pos &= (1UL << 22) - 1;
146
147 if (use_bank) {
148 if ((sh_bank != 0xFFFFFFFF && sh_bank >= adev->gfx.config.max_sh_per_se) ||
149 (se_bank != 0xFFFFFFFF && se_bank >= adev->gfx.config.max_shader_engines))
150 return -EINVAL;
151 mutex_lock(&adev->grbm_idx_mutex);
152 amdgpu_gfx_select_se_sh(adev, se_bank,
153 sh_bank, instance_bank);
154 } else if (use_ring) {
155 mutex_lock(&adev->srbm_mutex);
156 amdgpu_gfx_select_me_pipe_q(adev, me, pipe, queue, vmid);
157 }
158
159 if (pm_pg_lock)
160 mutex_lock(&adev->pm.mutex);
161
162 while (size) {
163 uint32_t value;
164
165 if (read) {
166 value = RREG32(*pos >> 2);
167 r = put_user(value, (uint32_t *)buf);
168 } else {
169 r = get_user(value, (uint32_t *)buf);
170 if (!r)
171 WREG32(*pos >> 2, value);
172 }
173 if (r) {
174 result = r;
175 goto end;
176 }
177
178 result += 4;
179 buf += 4;
180 *pos += 4;
181 size -= 4;
182 }
183
184end:
185 if (use_bank) {
186 amdgpu_gfx_select_se_sh(adev, 0xffffffff, 0xffffffff, 0xffffffff);
187 mutex_unlock(&adev->grbm_idx_mutex);
188 } else if (use_ring) {
189 amdgpu_gfx_select_me_pipe_q(adev, 0, 0, 0, 0);
190 mutex_unlock(&adev->srbm_mutex);
191 }
192
193 if (pm_pg_lock)
194 mutex_unlock(&adev->pm.mutex);
195
196 return result;
197}
198
199/**
200 * amdgpu_debugfs_regs_read - Callback for reading MMIO registers
201 */
202static ssize_t amdgpu_debugfs_regs_read(struct file *f, char __user *buf,
203 size_t size, loff_t *pos)
204{
205 return amdgpu_debugfs_process_reg_op(true, f, buf, size, pos);
206}
207
208/**
209 * amdgpu_debugfs_regs_write - Callback for writing MMIO registers
210 */
211static ssize_t amdgpu_debugfs_regs_write(struct file *f, const char __user *buf,
212 size_t size, loff_t *pos)
213{
214 return amdgpu_debugfs_process_reg_op(false, f, (char __user *)buf, size, pos);
215}
216
217
218/**
219 * amdgpu_debugfs_regs_pcie_read - Read from a PCIE register
220 *
221 * @f: open file handle
222 * @buf: User buffer to store read data in
223 * @size: Number of bytes to read
224 * @pos: Offset to seek to
225 *
226 * The lower bits are the BYTE offset of the register to read. This
227 * allows reading multiple registers in a single call and having
228 * the returned size reflect that.
229 */
230static ssize_t amdgpu_debugfs_regs_pcie_read(struct file *f, char __user *buf,
231 size_t size, loff_t *pos)
232{
233 struct amdgpu_device *adev = file_inode(f)->i_private;
234 ssize_t result = 0;
235 int r;
236
237 if (size & 0x3 || *pos & 0x3)
238 return -EINVAL;
239
240 while (size) {
241 uint32_t value;
242
Googlerb48fa912023-03-17 12:40:29 +0530243 value = RREG32_PCIE(*pos >> 2);
Googler9398cc32022-12-02 17:21:52 +0800244 r = put_user(value, (uint32_t *)buf);
245 if (r)
246 return r;
247
248 result += 4;
249 buf += 4;
250 *pos += 4;
251 size -= 4;
252 }
253
254 return result;
255}
256
257/**
258 * amdgpu_debugfs_regs_pcie_write - Write to a PCIE register
259 *
260 * @f: open file handle
261 * @buf: User buffer to write data from
262 * @size: Number of bytes to write
263 * @pos: Offset to seek to
264 *
265 * The lower bits are the BYTE offset of the register to write. This
266 * allows writing multiple registers in a single call and having
267 * the returned size reflect that.
268 */
269static ssize_t amdgpu_debugfs_regs_pcie_write(struct file *f, const char __user *buf,
270 size_t size, loff_t *pos)
271{
272 struct amdgpu_device *adev = file_inode(f)->i_private;
273 ssize_t result = 0;
274 int r;
275
276 if (size & 0x3 || *pos & 0x3)
277 return -EINVAL;
278
279 while (size) {
280 uint32_t value;
281
282 r = get_user(value, (uint32_t *)buf);
283 if (r)
284 return r;
285
Googlerb48fa912023-03-17 12:40:29 +0530286 WREG32_PCIE(*pos >> 2, value);
Googler9398cc32022-12-02 17:21:52 +0800287
288 result += 4;
289 buf += 4;
290 *pos += 4;
291 size -= 4;
292 }
293
294 return result;
295}
296
297/**
298 * amdgpu_debugfs_regs_didt_read - Read from a DIDT register
299 *
300 * @f: open file handle
301 * @buf: User buffer to store read data in
302 * @size: Number of bytes to read
303 * @pos: Offset to seek to
304 *
305 * The lower bits are the BYTE offset of the register to read. This
306 * allows reading multiple registers in a single call and having
307 * the returned size reflect that.
308 */
309static ssize_t amdgpu_debugfs_regs_didt_read(struct file *f, char __user *buf,
310 size_t size, loff_t *pos)
311{
312 struct amdgpu_device *adev = file_inode(f)->i_private;
313 ssize_t result = 0;
314 int r;
315
316 if (size & 0x3 || *pos & 0x3)
317 return -EINVAL;
318
319 while (size) {
320 uint32_t value;
321
322 value = RREG32_DIDT(*pos >> 2);
323 r = put_user(value, (uint32_t *)buf);
324 if (r)
325 return r;
326
327 result += 4;
328 buf += 4;
329 *pos += 4;
330 size -= 4;
331 }
332
333 return result;
334}
335
336/**
337 * amdgpu_debugfs_regs_didt_write - Write to a DIDT register
338 *
339 * @f: open file handle
340 * @buf: User buffer to write data from
341 * @size: Number of bytes to write
342 * @pos: Offset to seek to
343 *
344 * The lower bits are the BYTE offset of the register to write. This
345 * allows writing multiple registers in a single call and having
346 * the returned size reflect that.
347 */
348static ssize_t amdgpu_debugfs_regs_didt_write(struct file *f, const char __user *buf,
349 size_t size, loff_t *pos)
350{
351 struct amdgpu_device *adev = file_inode(f)->i_private;
352 ssize_t result = 0;
353 int r;
354
355 if (size & 0x3 || *pos & 0x3)
356 return -EINVAL;
357
358 while (size) {
359 uint32_t value;
360
361 r = get_user(value, (uint32_t *)buf);
362 if (r)
363 return r;
364
365 WREG32_DIDT(*pos >> 2, value);
366
367 result += 4;
368 buf += 4;
369 *pos += 4;
370 size -= 4;
371 }
372
373 return result;
374}
375
376/**
377 * amdgpu_debugfs_regs_smc_read - Read from a SMC register
378 *
379 * @f: open file handle
380 * @buf: User buffer to store read data in
381 * @size: Number of bytes to read
382 * @pos: Offset to seek to
383 *
384 * The lower bits are the BYTE offset of the register to read. This
385 * allows reading multiple registers in a single call and having
386 * the returned size reflect that.
387 */
388static ssize_t amdgpu_debugfs_regs_smc_read(struct file *f, char __user *buf,
389 size_t size, loff_t *pos)
390{
391 struct amdgpu_device *adev = file_inode(f)->i_private;
392 ssize_t result = 0;
393 int r;
394
395 if (size & 0x3 || *pos & 0x3)
396 return -EINVAL;
397
398 while (size) {
399 uint32_t value;
400
401 value = RREG32_SMC(*pos);
402 r = put_user(value, (uint32_t *)buf);
403 if (r)
404 return r;
405
406 result += 4;
407 buf += 4;
408 *pos += 4;
409 size -= 4;
410 }
411
412 return result;
413}
414
415/**
416 * amdgpu_debugfs_regs_smc_write - Write to a SMC register
417 *
418 * @f: open file handle
419 * @buf: User buffer to write data from
420 * @size: Number of bytes to write
421 * @pos: Offset to seek to
422 *
423 * The lower bits are the BYTE offset of the register to write. This
424 * allows writing multiple registers in a single call and having
425 * the returned size reflect that.
426 */
427static ssize_t amdgpu_debugfs_regs_smc_write(struct file *f, const char __user *buf,
428 size_t size, loff_t *pos)
429{
430 struct amdgpu_device *adev = file_inode(f)->i_private;
431 ssize_t result = 0;
432 int r;
433
434 if (size & 0x3 || *pos & 0x3)
435 return -EINVAL;
436
437 while (size) {
438 uint32_t value;
439
440 r = get_user(value, (uint32_t *)buf);
441 if (r)
442 return r;
443
444 WREG32_SMC(*pos, value);
445
446 result += 4;
447 buf += 4;
448 *pos += 4;
449 size -= 4;
450 }
451
452 return result;
453}
454
455/**
456 * amdgpu_debugfs_gca_config_read - Read from gfx config data
457 *
458 * @f: open file handle
459 * @buf: User buffer to store read data in
460 * @size: Number of bytes to read
461 * @pos: Offset to seek to
462 *
463 * This file is used to access configuration data in a somewhat
464 * stable fashion. The format is a series of DWORDs with the first
465 * indicating which revision it is. New content is appended to the
466 * end so that older software can still read the data.
467 */
468
469static ssize_t amdgpu_debugfs_gca_config_read(struct file *f, char __user *buf,
470 size_t size, loff_t *pos)
471{
472 struct amdgpu_device *adev = file_inode(f)->i_private;
473 ssize_t result = 0;
474 int r;
475 uint32_t *config, no_regs = 0;
476
477 if (size & 0x3 || *pos & 0x3)
478 return -EINVAL;
479
480 config = kmalloc_array(256, sizeof(*config), GFP_KERNEL);
481 if (!config)
482 return -ENOMEM;
483
484 /* version, increment each time something is added */
485 config[no_regs++] = 3;
486 config[no_regs++] = adev->gfx.config.max_shader_engines;
487 config[no_regs++] = adev->gfx.config.max_tile_pipes;
488 config[no_regs++] = adev->gfx.config.max_cu_per_sh;
489 config[no_regs++] = adev->gfx.config.max_sh_per_se;
490 config[no_regs++] = adev->gfx.config.max_backends_per_se;
491 config[no_regs++] = adev->gfx.config.max_texture_channel_caches;
492 config[no_regs++] = adev->gfx.config.max_gprs;
493 config[no_regs++] = adev->gfx.config.max_gs_threads;
494 config[no_regs++] = adev->gfx.config.max_hw_contexts;
495 config[no_regs++] = adev->gfx.config.sc_prim_fifo_size_frontend;
496 config[no_regs++] = adev->gfx.config.sc_prim_fifo_size_backend;
497 config[no_regs++] = adev->gfx.config.sc_hiz_tile_fifo_size;
498 config[no_regs++] = adev->gfx.config.sc_earlyz_tile_fifo_size;
499 config[no_regs++] = adev->gfx.config.num_tile_pipes;
500 config[no_regs++] = adev->gfx.config.backend_enable_mask;
501 config[no_regs++] = adev->gfx.config.mem_max_burst_length_bytes;
502 config[no_regs++] = adev->gfx.config.mem_row_size_in_kb;
503 config[no_regs++] = adev->gfx.config.shader_engine_tile_size;
504 config[no_regs++] = adev->gfx.config.num_gpus;
505 config[no_regs++] = adev->gfx.config.multi_gpu_tile_size;
506 config[no_regs++] = adev->gfx.config.mc_arb_ramcfg;
507 config[no_regs++] = adev->gfx.config.gb_addr_config;
508 config[no_regs++] = adev->gfx.config.num_rbs;
509
510 /* rev==1 */
511 config[no_regs++] = adev->rev_id;
512 config[no_regs++] = adev->pg_flags;
513 config[no_regs++] = adev->cg_flags;
514
515 /* rev==2 */
516 config[no_regs++] = adev->family;
517 config[no_regs++] = adev->external_rev_id;
518
519 /* rev==3 */
520 config[no_regs++] = adev->pdev->device;
521 config[no_regs++] = adev->pdev->revision;
522 config[no_regs++] = adev->pdev->subsystem_device;
523 config[no_regs++] = adev->pdev->subsystem_vendor;
524
525 while (size && (*pos < no_regs * 4)) {
526 uint32_t value;
527
528 value = config[*pos >> 2];
529 r = put_user(value, (uint32_t *)buf);
530 if (r) {
531 kfree(config);
532 return r;
533 }
534
535 result += 4;
536 buf += 4;
537 *pos += 4;
538 size -= 4;
539 }
540
541 kfree(config);
542 return result;
543}
544
545/**
546 * amdgpu_debugfs_sensor_read - Read from the powerplay sensors
547 *
548 * @f: open file handle
549 * @buf: User buffer to store read data in
550 * @size: Number of bytes to read
551 * @pos: Offset to seek to
552 *
553 * The offset is treated as the BYTE address of one of the sensors
554 * enumerated in amd/include/kgd_pp_interface.h under the
555 * 'amd_pp_sensors' enumeration. For instance to read the UVD VCLK
556 * you would use the offset 3 * 4 = 12.
557 */
558static ssize_t amdgpu_debugfs_sensor_read(struct file *f, char __user *buf,
559 size_t size, loff_t *pos)
560{
561 struct amdgpu_device *adev = file_inode(f)->i_private;
562 int idx, x, outsize, r, valuesize;
563 uint32_t values[16];
564
565 if (size & 3 || *pos & 0x3)
566 return -EINVAL;
567
568 if (!adev->pm.dpm_enabled)
569 return -EINVAL;
570
571 /* convert offset to sensor number */
572 idx = *pos >> 2;
573
574 valuesize = sizeof(values);
575 r = amdgpu_dpm_read_sensor(adev, idx, &values[0], &valuesize);
576 if (r)
577 return r;
578
579 if (size > valuesize)
580 return -EINVAL;
581
582 outsize = 0;
583 x = 0;
584 if (!r) {
585 while (size) {
586 r = put_user(values[x++], (int32_t *)buf);
587 buf += 4;
588 size -= 4;
589 outsize += 4;
590 }
591 }
592
593 return !r ? outsize : r;
594}
595
596/** amdgpu_debugfs_wave_read - Read WAVE STATUS data
597 *
598 * @f: open file handle
599 * @buf: User buffer to store read data in
600 * @size: Number of bytes to read
601 * @pos: Offset to seek to
602 *
603 * The offset being sought changes which wave that the status data
604 * will be returned for. The bits are used as follows:
605 *
606 * Bits 0..6: Byte offset into data
607 * Bits 7..14: SE selector
608 * Bits 15..22: SH/SA selector
609 * Bits 23..30: CU/{WGP+SIMD} selector
610 * Bits 31..36: WAVE ID selector
611 * Bits 37..44: SIMD ID selector
612 *
613 * The returned data begins with one DWORD of version information
614 * Followed by WAVE STATUS registers relevant to the GFX IP version
615 * being used. See gfx_v8_0_read_wave_data() for an example output.
616 */
617static ssize_t amdgpu_debugfs_wave_read(struct file *f, char __user *buf,
618 size_t size, loff_t *pos)
619{
620 struct amdgpu_device *adev = f->f_inode->i_private;
621 int r, x;
622 ssize_t result=0;
623 uint32_t offset, se, sh, cu, wave, simd, data[32];
624
625 if (size & 3 || *pos & 3)
626 return -EINVAL;
627
628 /* decode offset */
629 offset = (*pos & GENMASK_ULL(6, 0));
630 se = (*pos & GENMASK_ULL(14, 7)) >> 7;
631 sh = (*pos & GENMASK_ULL(22, 15)) >> 15;
632 cu = (*pos & GENMASK_ULL(30, 23)) >> 23;
633 wave = (*pos & GENMASK_ULL(36, 31)) >> 31;
634 simd = (*pos & GENMASK_ULL(44, 37)) >> 37;
635
636 /* switch to the specific se/sh/cu */
637 mutex_lock(&adev->grbm_idx_mutex);
638 amdgpu_gfx_select_se_sh(adev, se, sh, cu);
639
640 x = 0;
641 if (adev->gfx.funcs->read_wave_data)
642 adev->gfx.funcs->read_wave_data(adev, simd, wave, data, &x);
643
644 amdgpu_gfx_select_se_sh(adev, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
645 mutex_unlock(&adev->grbm_idx_mutex);
646
647 if (!x)
648 return -EINVAL;
649
650 while (size && (offset < x * 4)) {
651 uint32_t value;
652
653 value = data[offset >> 2];
654 r = put_user(value, (uint32_t *)buf);
655 if (r)
656 return r;
657
658 result += 4;
659 buf += 4;
660 offset += 4;
661 size -= 4;
662 }
663
664 return result;
665}
666
667/** amdgpu_debugfs_gpr_read - Read wave gprs
668 *
669 * @f: open file handle
670 * @buf: User buffer to store read data in
671 * @size: Number of bytes to read
672 * @pos: Offset to seek to
673 *
674 * The offset being sought changes which wave that the status data
675 * will be returned for. The bits are used as follows:
676 *
677 * Bits 0..11: Byte offset into data
678 * Bits 12..19: SE selector
679 * Bits 20..27: SH/SA selector
680 * Bits 28..35: CU/{WGP+SIMD} selector
681 * Bits 36..43: WAVE ID selector
682 * Bits 37..44: SIMD ID selector
683 * Bits 52..59: Thread selector
684 * Bits 60..61: Bank selector (VGPR=0,SGPR=1)
685 *
686 * The return data comes from the SGPR or VGPR register bank for
687 * the selected operational unit.
688 */
689static ssize_t amdgpu_debugfs_gpr_read(struct file *f, char __user *buf,
690 size_t size, loff_t *pos)
691{
692 struct amdgpu_device *adev = f->f_inode->i_private;
693 int r;
694 ssize_t result = 0;
695 uint32_t offset, se, sh, cu, wave, simd, thread, bank, *data;
696
697 if (size > 4096 || size & 3 || *pos & 3)
698 return -EINVAL;
699
700 /* decode offset */
701 offset = (*pos & GENMASK_ULL(11, 0)) >> 2;
702 se = (*pos & GENMASK_ULL(19, 12)) >> 12;
703 sh = (*pos & GENMASK_ULL(27, 20)) >> 20;
704 cu = (*pos & GENMASK_ULL(35, 28)) >> 28;
705 wave = (*pos & GENMASK_ULL(43, 36)) >> 36;
706 simd = (*pos & GENMASK_ULL(51, 44)) >> 44;
707 thread = (*pos & GENMASK_ULL(59, 52)) >> 52;
708 bank = (*pos & GENMASK_ULL(61, 60)) >> 60;
709
710 data = kcalloc(1024, sizeof(*data), GFP_KERNEL);
711 if (!data)
712 return -ENOMEM;
713
714 /* switch to the specific se/sh/cu */
715 mutex_lock(&adev->grbm_idx_mutex);
716 amdgpu_gfx_select_se_sh(adev, se, sh, cu);
717
718 if (bank == 0) {
719 if (adev->gfx.funcs->read_wave_vgprs)
720 adev->gfx.funcs->read_wave_vgprs(adev, simd, wave, thread, offset, size>>2, data);
721 } else {
722 if (adev->gfx.funcs->read_wave_sgprs)
723 adev->gfx.funcs->read_wave_sgprs(adev, simd, wave, offset, size>>2, data);
724 }
725
726 amdgpu_gfx_select_se_sh(adev, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
727 mutex_unlock(&adev->grbm_idx_mutex);
728
729 while (size) {
730 uint32_t value;
731
732 value = data[result >> 2];
733 r = put_user(value, (uint32_t *)buf);
734 if (r) {
735 result = r;
736 goto err;
737 }
738
739 result += 4;
740 buf += 4;
741 size -= 4;
742 }
743
744err:
745 kfree(data);
746 return result;
747}
748
749static const struct file_operations amdgpu_debugfs_regs_fops = {
750 .owner = THIS_MODULE,
751 .read = amdgpu_debugfs_regs_read,
752 .write = amdgpu_debugfs_regs_write,
753 .llseek = default_llseek
754};
755static const struct file_operations amdgpu_debugfs_regs_didt_fops = {
756 .owner = THIS_MODULE,
757 .read = amdgpu_debugfs_regs_didt_read,
758 .write = amdgpu_debugfs_regs_didt_write,
759 .llseek = default_llseek
760};
761static const struct file_operations amdgpu_debugfs_regs_pcie_fops = {
762 .owner = THIS_MODULE,
763 .read = amdgpu_debugfs_regs_pcie_read,
764 .write = amdgpu_debugfs_regs_pcie_write,
765 .llseek = default_llseek
766};
767static const struct file_operations amdgpu_debugfs_regs_smc_fops = {
768 .owner = THIS_MODULE,
769 .read = amdgpu_debugfs_regs_smc_read,
770 .write = amdgpu_debugfs_regs_smc_write,
771 .llseek = default_llseek
772};
773
774static const struct file_operations amdgpu_debugfs_gca_config_fops = {
775 .owner = THIS_MODULE,
776 .read = amdgpu_debugfs_gca_config_read,
777 .llseek = default_llseek
778};
779
780static const struct file_operations amdgpu_debugfs_sensors_fops = {
781 .owner = THIS_MODULE,
782 .read = amdgpu_debugfs_sensor_read,
783 .llseek = default_llseek
784};
785
786static const struct file_operations amdgpu_debugfs_wave_fops = {
787 .owner = THIS_MODULE,
788 .read = amdgpu_debugfs_wave_read,
789 .llseek = default_llseek
790};
791static const struct file_operations amdgpu_debugfs_gpr_fops = {
792 .owner = THIS_MODULE,
793 .read = amdgpu_debugfs_gpr_read,
794 .llseek = default_llseek
795};
796
797static const struct file_operations *debugfs_regs[] = {
798 &amdgpu_debugfs_regs_fops,
799 &amdgpu_debugfs_regs_didt_fops,
800 &amdgpu_debugfs_regs_pcie_fops,
801 &amdgpu_debugfs_regs_smc_fops,
802 &amdgpu_debugfs_gca_config_fops,
803 &amdgpu_debugfs_sensors_fops,
804 &amdgpu_debugfs_wave_fops,
805 &amdgpu_debugfs_gpr_fops,
806};
807
808static const char *debugfs_regs_names[] = {
809 "amdgpu_regs",
810 "amdgpu_regs_didt",
811 "amdgpu_regs_pcie",
812 "amdgpu_regs_smc",
813 "amdgpu_gca_config",
814 "amdgpu_sensors",
815 "amdgpu_wave",
816 "amdgpu_gpr",
817};
818
819/**
820 * amdgpu_debugfs_regs_init - Initialize debugfs entries that provide
821 * register access.
822 *
823 * @adev: The device to attach the debugfs entries to
824 */
825int amdgpu_debugfs_regs_init(struct amdgpu_device *adev)
826{
827 struct drm_minor *minor = adev->ddev->primary;
828 struct dentry *ent, *root = minor->debugfs_root;
829 unsigned int i;
830
831 for (i = 0; i < ARRAY_SIZE(debugfs_regs); i++) {
832 ent = debugfs_create_file(debugfs_regs_names[i],
833 S_IFREG | S_IRUGO, root,
834 adev, debugfs_regs[i]);
835 if (!i && !IS_ERR_OR_NULL(ent))
836 i_size_write(ent->d_inode, adev->rmmio_size);
837 adev->debugfs_regs[i] = ent;
838 }
839
840 return 0;
841}
842
843void amdgpu_debugfs_regs_cleanup(struct amdgpu_device *adev)
844{
845 unsigned i;
846
847 for (i = 0; i < ARRAY_SIZE(debugfs_regs); i++) {
848 if (adev->debugfs_regs[i]) {
849 debugfs_remove(adev->debugfs_regs[i]);
850 adev->debugfs_regs[i] = NULL;
851 }
852 }
853}
854
855static int amdgpu_debugfs_test_ib(struct seq_file *m, void *data)
856{
857 struct drm_info_node *node = (struct drm_info_node *) m->private;
858 struct drm_device *dev = node->minor->dev;
859 struct amdgpu_device *adev = dev->dev_private;
860 int r = 0, i;
861
862 /* Avoid accidently unparking the sched thread during GPU reset */
863 mutex_lock(&adev->lock_reset);
864
865 /* hold on the scheduler */
866 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
867 struct amdgpu_ring *ring = adev->rings[i];
868
869 if (!ring || !ring->sched.thread)
870 continue;
871 kthread_park(ring->sched.thread);
872 }
873
874 seq_printf(m, "run ib test:\n");
875 r = amdgpu_ib_ring_tests(adev);
876 if (r)
877 seq_printf(m, "ib ring tests failed (%d).\n", r);
878 else
879 seq_printf(m, "ib ring tests passed.\n");
880
881 /* go on the scheduler */
882 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
883 struct amdgpu_ring *ring = adev->rings[i];
884
885 if (!ring || !ring->sched.thread)
886 continue;
887 kthread_unpark(ring->sched.thread);
888 }
889
890 mutex_unlock(&adev->lock_reset);
891
892 return 0;
893}
894
895static int amdgpu_debugfs_get_vbios_dump(struct seq_file *m, void *data)
896{
897 struct drm_info_node *node = (struct drm_info_node *) m->private;
898 struct drm_device *dev = node->minor->dev;
899 struct amdgpu_device *adev = dev->dev_private;
900
901 seq_write(m, adev->bios, adev->bios_size);
902 return 0;
903}
904
905static int amdgpu_debugfs_evict_vram(struct seq_file *m, void *data)
906{
907 struct drm_info_node *node = (struct drm_info_node *)m->private;
908 struct drm_device *dev = node->minor->dev;
909 struct amdgpu_device *adev = dev->dev_private;
910
911 seq_printf(m, "(%d)\n", amdgpu_bo_evict_vram(adev));
912 return 0;
913}
914
915static int amdgpu_debugfs_evict_gtt(struct seq_file *m, void *data)
916{
917 struct drm_info_node *node = (struct drm_info_node *)m->private;
918 struct drm_device *dev = node->minor->dev;
919 struct amdgpu_device *adev = dev->dev_private;
920
921 seq_printf(m, "(%d)\n", ttm_bo_evict_mm(&adev->mman.bdev, TTM_PL_TT));
922 return 0;
923}
924
925static const struct drm_info_list amdgpu_debugfs_list[] = {
926 {"amdgpu_vbios", amdgpu_debugfs_get_vbios_dump},
927 {"amdgpu_test_ib", &amdgpu_debugfs_test_ib},
928 {"amdgpu_evict_vram", &amdgpu_debugfs_evict_vram},
929 {"amdgpu_evict_gtt", &amdgpu_debugfs_evict_gtt},
930};
931
932static void amdgpu_ib_preempt_fences_swap(struct amdgpu_ring *ring,
933 struct dma_fence **fences)
934{
935 struct amdgpu_fence_driver *drv = &ring->fence_drv;
936 uint32_t sync_seq, last_seq;
937
938 last_seq = atomic_read(&ring->fence_drv.last_seq);
939 sync_seq = ring->fence_drv.sync_seq;
940
941 last_seq &= drv->num_fences_mask;
942 sync_seq &= drv->num_fences_mask;
943
944 do {
945 struct dma_fence *fence, **ptr;
946
947 ++last_seq;
948 last_seq &= drv->num_fences_mask;
949 ptr = &drv->fences[last_seq];
950
951 fence = rcu_dereference_protected(*ptr, 1);
952 RCU_INIT_POINTER(*ptr, NULL);
953
954 if (!fence)
955 continue;
956
957 fences[last_seq] = fence;
958
959 } while (last_seq != sync_seq);
960}
961
962static void amdgpu_ib_preempt_signal_fences(struct dma_fence **fences,
963 int length)
964{
965 int i;
966 struct dma_fence *fence;
967
968 for (i = 0; i < length; i++) {
969 fence = fences[i];
970 if (!fence)
971 continue;
972 dma_fence_signal(fence);
973 dma_fence_put(fence);
974 }
975}
976
977static void amdgpu_ib_preempt_job_recovery(struct drm_gpu_scheduler *sched)
978{
979 struct drm_sched_job *s_job;
980 struct dma_fence *fence;
981
982 spin_lock(&sched->job_list_lock);
983 list_for_each_entry(s_job, &sched->ring_mirror_list, node) {
984 fence = sched->ops->run_job(s_job);
985 dma_fence_put(fence);
986 }
987 spin_unlock(&sched->job_list_lock);
988}
989
990static void amdgpu_ib_preempt_mark_partial_job(struct amdgpu_ring *ring)
991{
992 struct amdgpu_job *job;
993 struct drm_sched_job *s_job, *tmp;
994 uint32_t preempt_seq;
995 struct dma_fence *fence, **ptr;
996 struct amdgpu_fence_driver *drv = &ring->fence_drv;
997 struct drm_gpu_scheduler *sched = &ring->sched;
998 bool preempted = true;
999
1000 if (ring->funcs->type != AMDGPU_RING_TYPE_GFX)
1001 return;
1002
1003 preempt_seq = le32_to_cpu(*(drv->cpu_addr + 2));
1004 if (preempt_seq <= atomic_read(&drv->last_seq)) {
1005 preempted = false;
1006 goto no_preempt;
1007 }
1008
1009 preempt_seq &= drv->num_fences_mask;
1010 ptr = &drv->fences[preempt_seq];
1011 fence = rcu_dereference_protected(*ptr, 1);
1012
1013no_preempt:
1014 spin_lock(&sched->job_list_lock);
1015 list_for_each_entry_safe(s_job, tmp, &sched->ring_mirror_list, node) {
1016 if (dma_fence_is_signaled(&s_job->s_fence->finished)) {
1017 /* remove job from ring_mirror_list */
1018 list_del_init(&s_job->node);
1019 sched->ops->free_job(s_job);
1020 continue;
1021 }
1022 job = to_amdgpu_job(s_job);
1023 if (preempted && job->fence == fence)
1024 /* mark the job as preempted */
1025 job->preemption_status |= AMDGPU_IB_PREEMPTED;
1026 }
1027 spin_unlock(&sched->job_list_lock);
1028}
1029
1030static int amdgpu_debugfs_ib_preempt(void *data, u64 val)
1031{
1032 int r, resched, length;
1033 struct amdgpu_ring *ring;
1034 struct dma_fence **fences = NULL;
1035 struct amdgpu_device *adev = (struct amdgpu_device *)data;
1036
1037 if (val >= AMDGPU_MAX_RINGS)
1038 return -EINVAL;
1039
1040 ring = adev->rings[val];
1041
1042 if (!ring || !ring->funcs->preempt_ib || !ring->sched.thread)
1043 return -EINVAL;
1044
1045 /* the last preemption failed */
1046 if (ring->trail_seq != le32_to_cpu(*ring->trail_fence_cpu_addr))
1047 return -EBUSY;
1048
1049 length = ring->fence_drv.num_fences_mask + 1;
1050 fences = kcalloc(length, sizeof(void *), GFP_KERNEL);
1051 if (!fences)
1052 return -ENOMEM;
1053
1054 /* Avoid accidently unparking the sched thread during GPU reset */
1055 mutex_lock(&adev->lock_reset);
1056
1057 /* stop the scheduler */
1058 kthread_park(ring->sched.thread);
1059
1060 resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
1061
1062 /* preempt the IB */
1063 r = amdgpu_ring_preempt_ib(ring);
1064 if (r) {
1065 DRM_WARN("failed to preempt ring %d\n", ring->idx);
1066 goto failure;
1067 }
1068
1069 amdgpu_fence_process(ring);
1070
1071 if (atomic_read(&ring->fence_drv.last_seq) !=
1072 ring->fence_drv.sync_seq) {
1073 DRM_INFO("ring %d was preempted\n", ring->idx);
1074
1075 amdgpu_ib_preempt_mark_partial_job(ring);
1076
1077 /* swap out the old fences */
1078 amdgpu_ib_preempt_fences_swap(ring, fences);
1079
1080 amdgpu_fence_driver_force_completion(ring);
1081
1082 /* resubmit unfinished jobs */
1083 amdgpu_ib_preempt_job_recovery(&ring->sched);
1084
1085 /* wait for jobs finished */
1086 amdgpu_fence_wait_empty(ring);
1087
1088 /* signal the old fences */
1089 amdgpu_ib_preempt_signal_fences(fences, length);
1090 }
1091
1092failure:
1093 /* restart the scheduler */
1094 kthread_unpark(ring->sched.thread);
1095
1096 mutex_unlock(&adev->lock_reset);
1097
1098 ttm_bo_unlock_delayed_workqueue(&adev->mman.bdev, resched);
1099
1100 if (fences)
1101 kfree(fences);
1102
1103 return 0;
1104}
1105
1106DEFINE_SIMPLE_ATTRIBUTE(fops_ib_preempt, NULL,
1107 amdgpu_debugfs_ib_preempt, "%llu\n");
1108
1109int amdgpu_debugfs_init(struct amdgpu_device *adev)
1110{
1111 adev->debugfs_preempt =
1112 debugfs_create_file("amdgpu_preempt_ib", 0600,
1113 adev->ddev->primary->debugfs_root,
1114 (void *)adev, &fops_ib_preempt);
1115 if (!(adev->debugfs_preempt)) {
1116 DRM_ERROR("unable to create amdgpu_preempt_ib debugsfs file\n");
1117 return -EIO;
1118 }
1119
1120 return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_list,
1121 ARRAY_SIZE(amdgpu_debugfs_list));
1122}
1123
1124void amdgpu_debugfs_preempt_cleanup(struct amdgpu_device *adev)
1125{
1126 if (adev->debugfs_preempt)
1127 debugfs_remove(adev->debugfs_preempt);
1128}
1129
1130#else
1131int amdgpu_debugfs_init(struct amdgpu_device *adev)
1132{
1133 return 0;
1134}
1135void amdgpu_debugfs_preempt_cleanup(struct amdgpu_device *adev) { }
1136int amdgpu_debugfs_regs_init(struct amdgpu_device *adev)
1137{
1138 return 0;
1139}
1140void amdgpu_debugfs_regs_cleanup(struct amdgpu_device *adev) { }
1141#endif