forked from ~ljy/RK356X_SDK_RELEASE
blame | history | patch | commit | commitdiff
hc
2023-12-11 1f93a7dfd1f8d5ff7a5c53246c7534fe2332d6f4 
 kernel/include/uapi/gpu/arm/bifrost/mali_kbase_hwcnt_reader.h
....@@ -1,7 +1,7 @@
11 /* SPDX-License-Identifier: GPL-2.0 WITH Linux-syscall-note */
22 /*
33  *
4
- * (C) COPYRIGHT 2015, 2020-2021 ARM Limited. All rights reserved.
4
+ * (C) COPYRIGHT 2015, 2020-2022 ARM Limited. All rights reserved.
55  *
66  * This program is free software and is provided to you under the terms of the
77  * GNU General Public License version 2 as published by the Free Software
....@@ -22,7 +22,7 @@
2222 #ifndef _UAPI_KBASE_HWCNT_READER_H_
2323 #define _UAPI_KBASE_HWCNT_READER_H_
2424 
25
-#include <stddef.h>
25
+#include <linux/stddef.h>
2626 #include <linux/types.h>
2727 
2828 /* The ids of ioctl commands. */
....@@ -91,6 +91,7 @@
9191 #define KBASE_HWCNT_READER_API_VERSION_NO_FEATURE (0)
9292 #define KBASE_HWCNT_READER_API_VERSION_FEATURE_CYCLES_TOP (1 << 0)
9393 #define KBASE_HWCNT_READER_API_VERSION_FEATURE_CYCLES_SHADER_CORES (1 << 1)
94
+
9495 /**
9596  * struct kbase_hwcnt_reader_api_version - hwcnt reader API version
9697  * @version:  API version
....@@ -101,5 +102,402 @@
101102 	__u32 features;
102103 };
103104 
104
-#endif /* _UAPI_KBASE_HWCNT_READER_H_ */
105
+/** Hardware counters reader API version */
106
+#define PRFCNT_READER_API_VERSION (0)
105107 
108
+/**
109
+ * enum prfcnt_list_type - Type of list item
110
+ * @PRFCNT_LIST_TYPE_ENUM:        Enumeration of performance counters.
111
+ * @PRFCNT_LIST_TYPE_REQUEST:     Request for configuration setup.
112
+ * @PRFCNT_LIST_TYPE_SAMPLE_META: Sample metadata.
113
+ */
114
+enum prfcnt_list_type {
115
+	PRFCNT_LIST_TYPE_ENUM,
116
+	PRFCNT_LIST_TYPE_REQUEST,
117
+	PRFCNT_LIST_TYPE_SAMPLE_META,
118
+};
119
+
120
+#define FLEX_LIST_TYPE(type, subtype)                                          \
121
+	((__u16)(((type & 0xf) << 12) | (subtype & 0xfff)))
122
+#define FLEX_LIST_TYPE_NONE FLEX_LIST_TYPE(0, 0)
123
+
124
+#define PRFCNT_ENUM_TYPE_BLOCK FLEX_LIST_TYPE(PRFCNT_LIST_TYPE_ENUM, 0)
125
+#define PRFCNT_ENUM_TYPE_REQUEST FLEX_LIST_TYPE(PRFCNT_LIST_TYPE_ENUM, 1)
126
+#define PRFCNT_ENUM_TYPE_SAMPLE_INFO FLEX_LIST_TYPE(PRFCNT_LIST_TYPE_ENUM, 2)
127
+
128
+#define PRFCNT_REQUEST_TYPE_MODE FLEX_LIST_TYPE(PRFCNT_LIST_TYPE_REQUEST, 0)
129
+#define PRFCNT_REQUEST_TYPE_ENABLE FLEX_LIST_TYPE(PRFCNT_LIST_TYPE_REQUEST, 1)
130
+#define PRFCNT_REQUEST_TYPE_SCOPE FLEX_LIST_TYPE(PRFCNT_LIST_TYPE_REQUEST, 2)
131
+
132
+#define PRFCNT_SAMPLE_META_TYPE_SAMPLE                                         \
133
+	FLEX_LIST_TYPE(PRFCNT_LIST_TYPE_SAMPLE_META, 0)
134
+#define PRFCNT_SAMPLE_META_TYPE_CLOCK                                          \
135
+	FLEX_LIST_TYPE(PRFCNT_LIST_TYPE_SAMPLE_META, 1)
136
+#define PRFCNT_SAMPLE_META_TYPE_BLOCK                                          \
137
+	FLEX_LIST_TYPE(PRFCNT_LIST_TYPE_SAMPLE_META, 2)
138
+
139
+/**
140
+ * struct prfcnt_item_header - Header for an item of the list.
141
+ * @item_type:    Type of item.
142
+ * @item_version: Protocol version.
143
+ */
144
+struct prfcnt_item_header {
145
+	__u16 item_type;
146
+	__u16 item_version;
147
+};
148
+
149
+/**
150
+ * enum prfcnt_block_type - Type of performance counter block.
151
+ * @PRFCNT_BLOCK_TYPE_FE:          Front End.
152
+ * @PRFCNT_BLOCK_TYPE_TILER:       Tiler.
153
+ * @PRFCNT_BLOCK_TYPE_MEMORY:      Memory System.
154
+ * @PRFCNT_BLOCK_TYPE_SHADER_CORE: Shader Core.
155
+ * @PRFCNT_BLOCK_TYPE_RESERVED:    Reserved.
156
+ */
157
+enum prfcnt_block_type {
158
+	PRFCNT_BLOCK_TYPE_FE,
159
+	PRFCNT_BLOCK_TYPE_TILER,
160
+	PRFCNT_BLOCK_TYPE_MEMORY,
161
+	PRFCNT_BLOCK_TYPE_SHADER_CORE,
162
+	PRFCNT_BLOCK_TYPE_RESERVED = 255,
163
+};
164
+
165
+/**
166
+ * enum prfcnt_set - Type of performance counter block set.
167
+ * @PRFCNT_SET_PRIMARY:   Primary.
168
+ * @PRFCNT_SET_SECONDARY: Secondary.
169
+ * @PRFCNT_SET_TERTIARY:  Tertiary.
170
+ * @PRFCNT_SET_RESERVED:  Reserved.
171
+ */
172
+enum prfcnt_set {
173
+	PRFCNT_SET_PRIMARY,
174
+	PRFCNT_SET_SECONDARY,
175
+	PRFCNT_SET_TERTIARY,
176
+	PRFCNT_SET_RESERVED = 255,
177
+};
178
+
179
+/**
180
+ * struct prfcnt_enum_block_counter - Performance counter block descriptor.
181
+ * @block_type:    Type of performance counter block.
182
+ * @set:           Which SET this represents: primary, secondary or tertiary.
183
+ * @pad:           Padding bytes.
184
+ * @num_instances: How many instances of this block type exist in the hardware.
185
+ * @num_values:    How many entries in the values array there are for samples
186
+ *                 from this block.
187
+ * @counter_mask:  Bitmask that indicates counter availability in this block.
188
+ *                 A '0' indicates that a counter is not available at that
189
+ *                 index and will always return zeroes if requested.
190
+ */
191
+struct prfcnt_enum_block_counter {
192
+	__u8 block_type;
193
+	__u8 set;
194
+	__u8 pad[2];
195
+	__u16 num_instances;
196
+	__u16 num_values;
197
+	__u64 counter_mask[2];
198
+};
199
+
200
+/**
201
+ * struct prfcnt_enum_request - Request descriptor.
202
+ * @request_item_type:       Type of request.
203
+ * @pad:                     Padding bytes.
204
+ * @versions_mask: Bitmask of versions that support this request.
205
+ */
206
+struct prfcnt_enum_request {
207
+	__u16 request_item_type;
208
+	__u16 pad;
209
+	__u32 versions_mask;
210
+};
211
+
212
+/**
213
+ * struct prfcnt_enum_sample_info - Sample information descriptor.
214
+ * @num_clock_domains:       Number of clock domains of the GPU.
215
+ * @pad:                     Padding bytes.
216
+ */
217
+struct prfcnt_enum_sample_info {
218
+	__u32 num_clock_domains;
219
+	__u32 pad;
220
+};
221
+
222
+/**
223
+ * struct prfcnt_enum_item - Performance counter enumeration item.
224
+ * @padding:         Padding bytes.
225
+ * @hdr:             Header describing the type of item in the list.
226
+ * @u:               Structure containing discriptor for enumeration item type.
227
+ * @u.block_counter: Performance counter block descriptor.
228
+ * @u.request:       Request descriptor.
229
+ * @u.sample_info:   Performance counter sample information descriptor.
230
+ */
231
+struct prfcnt_enum_item {
232
+	struct prfcnt_item_header hdr;
233
+	__u8 padding[4];
234
+	/** union u - union of block_counter and request */
235
+	union {
236
+		struct prfcnt_enum_block_counter block_counter;
237
+		struct prfcnt_enum_request request;
238
+		struct prfcnt_enum_sample_info sample_info;
239
+	} u;
240
+};
241
+
242
+/**
243
+ * enum prfcnt_mode - Capture mode for counter sampling.
244
+ * @PRFCNT_MODE_MANUAL:   Manual sampling mode.
245
+ * @PRFCNT_MODE_PERIODIC: Periodic sampling mode.
246
+ * @PRFCNT_MODE_RESERVED: Reserved.
247
+ */
248
+enum prfcnt_mode {
249
+	PRFCNT_MODE_MANUAL,
250
+	PRFCNT_MODE_PERIODIC,
251
+	PRFCNT_MODE_RESERVED = 255,
252
+};
253
+
254
+/**
255
+ * struct prfcnt_request_mode - Mode request descriptor.
256
+ * @mode:                           Capture mode for the session, either manual or periodic.
257
+ * @pad:                            Padding bytes.
258
+ * @mode_config:                    Structure containing configuration for periodic mode.
259
+ * @mode_config.periodic:           Periodic config.
260
+ * @mode_config.periodic.period_ns: Period in nanoseconds, for periodic mode.
261
+ */
262
+struct prfcnt_request_mode {
263
+	__u8 mode;
264
+	__u8 pad[7];
265
+	/** union mode_config - request mode configuration*/
266
+	union {
267
+		struct {
268
+			__u64 period_ns;
269
+		} periodic;
270
+	} mode_config;
271
+};
272
+
273
+/**
274
+ * struct prfcnt_request_enable - Enable request descriptor.
275
+ * @block_type:  Type of performance counter block.
276
+ * @set:         Which SET to use: primary, secondary or tertiary.
277
+ * @pad:         Padding bytes.
278
+ * @enable_mask: Bitmask that indicates which performance counters to enable.
279
+ *               Unavailable counters will be ignored.
280
+ */
281
+struct prfcnt_request_enable {
282
+	__u8 block_type;
283
+	__u8 set;
284
+	__u8 pad[6];
285
+	__u64 enable_mask[2];
286
+};
287
+
288
+/**
289
+ * enum prfcnt_scope - Scope of performance counters.
290
+ * @PRFCNT_SCOPE_GLOBAL:   Global scope.
291
+ * @PRFCNT_SCOPE_RESERVED: Reserved.
292
+ */
293
+enum prfcnt_scope {
294
+	PRFCNT_SCOPE_GLOBAL,
295
+	PRFCNT_SCOPE_RESERVED = 255,
296
+};
297
+
298
+/**
299
+ * struct prfcnt_request_scope - Scope request descriptor.
300
+ * @scope: Scope of the performance counters to capture.
301
+ * @pad:   Padding bytes.
302
+ */
303
+struct prfcnt_request_scope {
304
+	__u8 scope;
305
+	__u8 pad[7];
306
+};
307
+
308
+/**
309
+ * struct prfcnt_request_item - Performance counter request item.
310
+ * @padding:      Padding bytes.
311
+ * @hdr:          Header describing the type of item in the list.
312
+ * @u:            Structure containing descriptor for request type.
313
+ * @u.req_mode:   Mode request descriptor.
314
+ * @u.req_enable: Enable request descriptor.
315
+ * @u.req_scope:  Scope request descriptor.
316
+ */
317
+struct prfcnt_request_item {
318
+	struct prfcnt_item_header hdr;
319
+	__u8 padding[4];
320
+	/** union u - union on req_mode and req_enable */
321
+	union {
322
+		struct prfcnt_request_mode req_mode;
323
+		struct prfcnt_request_enable req_enable;
324
+		struct prfcnt_request_scope req_scope;
325
+	} u;
326
+};
327
+
328
+/**
329
+ * enum prfcnt_request_type - Type of request descriptor.
330
+ * @PRFCNT_REQUEST_MODE:   Specify the capture mode to be used for the session.
331
+ * @PRFCNT_REQUEST_ENABLE: Specify which performance counters to capture.
332
+ * @PRFCNT_REQUEST_SCOPE:  Specify the scope of the performance counters.
333
+ */
334
+enum prfcnt_request_type {
335
+	PRFCNT_REQUEST_MODE,
336
+	PRFCNT_REQUEST_ENABLE,
337
+	PRFCNT_REQUEST_SCOPE,
338
+};
339
+
340
+/* This sample contains overflows from dump duration stretch because the sample buffer was full */
341
+#define SAMPLE_FLAG_OVERFLOW (1u << 0)
342
+/* This sample has had an error condition for sample duration */
343
+#define SAMPLE_FLAG_ERROR (1u << 30)
344
+
345
+/**
346
+ * struct prfcnt_sample_metadata - Metadata for counter sample data.
347
+ * @timestamp_start: Earliest timestamp that values in this sample represent.
348
+ * @timestamp_end:   Latest timestamp that values in this sample represent.
349
+ * @seq:             Sequence number of this sample. Must match the value from
350
+ *                   GET_SAMPLE.
351
+ * @user_data:       User data provided to HWC_CMD_START or HWC_CMD_SAMPLE_*
352
+ * @flags:           Property flags.
353
+ * @pad:             Padding bytes.
354
+ */
355
+struct prfcnt_sample_metadata {
356
+	__u64 timestamp_start;
357
+	__u64 timestamp_end;
358
+	__u64 seq;
359
+	__u64 user_data;
360
+	__u32 flags;
361
+	__u32 pad;
362
+};
363
+
364
+/* Maximum number of domains a metadata for clock cycles can refer to */
365
+#define MAX_REPORTED_DOMAINS (4)
366
+
367
+/**
368
+ * struct prfcnt_clock_metadata - Metadata for clock cycles.
369
+ * @num_domains: Number of domains this metadata refers to.
370
+ * @pad:         Padding bytes.
371
+ * @cycles:      Number of cycles elapsed in each counter domain between
372
+ *               timestamp_start and timestamp_end. Valid only for the
373
+ *               first @p num_domains.
374
+ */
375
+struct prfcnt_clock_metadata {
376
+	__u32 num_domains;
377
+	__u32 pad;
378
+	__u64 cycles[MAX_REPORTED_DOMAINS];
379
+};
380
+
381
+/* This block state is unknown */
382
+#define BLOCK_STATE_UNKNOWN (0)
383
+/* This block was powered on for at least some portion of the sample */
384
+#define BLOCK_STATE_ON (1 << 0)
385
+/* This block was powered off for at least some portion of the sample */
386
+#define BLOCK_STATE_OFF (1 << 1)
387
+/* This block was available to this VM for at least some portion of the sample */
388
+#define BLOCK_STATE_AVAILABLE (1 << 2)
389
+/* This block was not available to this VM for at least some portion of the sample
390
+ *  Note that no data is collected when the block is not available to the VM.
391
+ */
392
+#define BLOCK_STATE_UNAVAILABLE (1 << 3)
393
+/* This block was operating in "normal" (non-protected) mode for at least some portion of the sample */
394
+#define BLOCK_STATE_NORMAL (1 << 4)
395
+/* This block was operating in "protected" mode for at least some portion of the sample.
396
+ * Note that no data is collected when the block is in protected mode.
397
+ */
398
+#define BLOCK_STATE_PROTECTED (1 << 5)
399
+
400
+/**
401
+ * struct prfcnt_block_metadata - Metadata for counter block.
402
+ * @block_type:    Type of performance counter block.
403
+ * @block_idx:     Index of performance counter block.
404
+ * @set:           Set of performance counter block.
405
+ * @pad_u8:        Padding bytes.
406
+ * @block_state:   Bits set indicate the states which the block is known
407
+ *                 to have operated in during this sample.
408
+ * @values_offset: Offset from the start of the mmapped region, to the values
409
+ *                 for this block. The values themselves are an array of __u64.
410
+ * @pad_u32:       Padding bytes.
411
+ */
412
+struct prfcnt_block_metadata {
413
+	__u8 block_type;
414
+	__u8 block_idx;
415
+	__u8 set;
416
+	__u8 pad_u8;
417
+	__u32 block_state;
418
+	__u32 values_offset;
419
+	__u32 pad_u32;
420
+};
421
+
422
+/**
423
+ * struct prfcnt_metadata - Performance counter metadata item.
424
+ * @padding:     Padding bytes.
425
+ * @hdr:         Header describing the type of item in the list.
426
+ * @u:           Structure containing descriptor for metadata type.
427
+ * @u.sample_md: Counter sample data metadata descriptor.
428
+ * @u.clock_md:  Clock cycles metadata descriptor.
429
+ * @u.block_md:  Counter block metadata descriptor.
430
+ */
431
+struct prfcnt_metadata {
432
+	struct prfcnt_item_header hdr;
433
+	__u8 padding[4];
434
+	union {
435
+		struct prfcnt_sample_metadata sample_md;
436
+		struct prfcnt_clock_metadata clock_md;
437
+		struct prfcnt_block_metadata block_md;
438
+	} u;
439
+};
440
+
441
+/**
442
+ * enum prfcnt_control_cmd_code - Control command code for client session.
443
+ * @PRFCNT_CONTROL_CMD_START:        Start the counter data dump run for
444
+ *                                   the calling client session.
445
+ * @PRFCNT_CONTROL_CMD_STOP:         Stop the counter data dump run for the
446
+ *                                   calling client session.
447
+ * @PRFCNT_CONTROL_CMD_SAMPLE_SYNC:  Trigger a synchronous manual sample.
448
+ * @PRFCNT_CONTROL_CMD_RESERVED:     Previously SAMPLE_ASYNC not supported any more.
449
+ * @PRFCNT_CONTROL_CMD_DISCARD:      Discard all samples which have not yet
450
+ *                                   been consumed by userspace. Note that
451
+ *                                   this can race with new samples if
452
+ *                                   HWC_CMD_STOP is not called first.
453
+ */
454
+enum prfcnt_control_cmd_code {
455
+	PRFCNT_CONTROL_CMD_START = 1,
456
+	PRFCNT_CONTROL_CMD_STOP,
457
+	PRFCNT_CONTROL_CMD_SAMPLE_SYNC,
458
+	PRFCNT_CONTROL_CMD_RESERVED,
459
+	PRFCNT_CONTROL_CMD_DISCARD,
460
+};
461
+
462
+/** struct prfcnt_control_cmd - Control command
463
+ * @cmd:       Control command for the session.
464
+ * @pad:       Padding bytes.
465
+ * @user_data: Pointer to user data, which will be returned as part of
466
+ *             sample metadata. It only affects a single sample if used
467
+ *             with CMD_SAMPLE_SYNC or CMD_SAMPLE_ASYNC. It affects all
468
+ *             samples between CMD_START and CMD_STOP if used with the
469
+ *             periodic sampling.
470
+ */
471
+struct prfcnt_control_cmd {
472
+	__u16 cmd;
473
+	__u16 pad[3];
474
+	__u64 user_data;
475
+};
476
+
477
+/** struct prfcnt_sample_access - Metadata to access a sample.
478
+ * @sequence:            Sequence number for the sample.
479
+ *                       For GET_SAMPLE, it will be set by the kernel.
480
+ *                       For PUT_SAMPLE, it shall be equal to the same value
481
+ *                       provided by the kernel for GET_SAMPLE.
482
+ * @sample_offset_bytes: Offset from the start of the mapped area to the first
483
+ *                       entry in the metadata list (sample_metadata) for this
484
+ *                       sample.
485
+ */
486
+struct prfcnt_sample_access {
487
+	__u64 sequence;
488
+	__u64 sample_offset_bytes;
489
+};
490
+
491
+/* The ids of ioctl commands, on a reader file descriptor, magic number */
492
+#define KBASE_KINSTR_PRFCNT_READER 0xBF
493
+/* Ioctl ID for issuing a session operational command */
494
+#define KBASE_IOCTL_KINSTR_PRFCNT_CMD                                          \
495
+	_IOW(KBASE_KINSTR_PRFCNT_READER, 0x00, struct prfcnt_control_cmd)
496
+/* Ioctl ID for fetching a dumpped sample */
497
+#define KBASE_IOCTL_KINSTR_PRFCNT_GET_SAMPLE                                   \
498
+	_IOR(KBASE_KINSTR_PRFCNT_READER, 0x01, struct prfcnt_sample_access)
499
+/* Ioctl ID for release internal buffer of the previously fetched sample */
500
+#define KBASE_IOCTL_KINSTR_PRFCNT_PUT_SAMPLE                                   \
501
+	_IOW(KBASE_KINSTR_PRFCNT_READER, 0x10, struct prfcnt_sample_access)
502
+
503
+#endif /* _UAPI_KBASE_HWCNT_READER_H_ */