1 /*        $NetBSD: i915_perf_types.h,v 1.6 2021/12/19 11:36:57 riastradh Exp $  */
2 
3 /* SPDX-License-Identifier: MIT */
4 /*
5  * Copyright © 2019 Intel Corporation
6  */
7 
8 #ifndef _I915_PERF_TYPES_H_
9 #define _I915_PERF_TYPES_H_
10 
11 #include <linux/atomic.h>
12 #include <linux/device.h>
13 #include <linux/hrtimer.h>
14 #include <linux/llist.h>
15 #include <linux/poll.h>
16 #include <linux/sysfs.h>
17 #include <linux/types.h>
18 #include <linux/uuid.h>
19 #include <linux/wait.h>
20 
21 #include "i915_reg.h"
22 #include "intel_wakeref.h"
23 
24 struct drm_i915_private;
25 struct file;
26 struct i915_gem_context;
27 struct i915_perf;
28 struct i915_vma;
29 struct intel_context;
30 struct intel_engine_cs;
31 
32 struct i915_oa_format {
33           u32 format;
34           int size;
35 };
36 
37 struct i915_oa_reg {
38           i915_reg_t addr;
39           u32 value;
40 };
41 
42 struct i915_oa_config {
43           struct i915_perf *perf;
44 
45           char uuid[UUID_STRING_LEN + 1];
46           int id;
47 
48           const struct i915_oa_reg *mux_regs;
49           u32 mux_regs_len;
50           const struct i915_oa_reg *b_counter_regs;
51           u32 b_counter_regs_len;
52           const struct i915_oa_reg *flex_regs;
53           u32 flex_regs_len;
54 
55 #ifndef __NetBSD__            /* XXX sysfs */
56           struct attribute_group sysfs_metric;
57           struct attribute *attrs[2];
58           struct device_attribute sysfs_metric_id;
59 #endif
60 
61           struct kref ref;
62           struct rcu_head rcu;
63 };
64 
65 struct i915_perf_stream;
66 
67 /**
68  * struct i915_perf_stream_ops - the OPs to support a specific stream type
69  */
70 struct i915_perf_stream_ops {
71           /**
72            * @enable: Enables the collection of HW samples, either in response to
73            * `I915_PERF_IOCTL_ENABLE` or implicitly called when stream is opened
74            * without `I915_PERF_FLAG_DISABLED`.
75            */
76           void (*enable)(struct i915_perf_stream *stream);
77 
78           /**
79            * @disable: Disables the collection of HW samples, either in response
80            * to `I915_PERF_IOCTL_DISABLE` or implicitly called before destroying
81            * the stream.
82            */
83           void (*disable)(struct i915_perf_stream *stream);
84 
85 #ifndef __NetBSD__
86           /**
87            * @poll_wait: Call poll_wait, passing a wait queue that will be woken
88            * once there is something ready to read() for the stream
89            */
90           void (*poll_wait)(struct i915_perf_stream *stream,
91                                 struct file *file,
92                                 poll_table *wait);
93 #endif
94 
95           /**
96            * @wait_unlocked: For handling a blocking read, wait until there is
97            * something to ready to read() for the stream. E.g. wait on the same
98            * wait queue that would be passed to poll_wait().
99            */
100           int (*wait_unlocked)(struct i915_perf_stream *stream);
101 
102           /**
103            * @read: Copy buffered metrics as records to userspace
104            * **buf**: the userspace, destination buffer
105            * **count**: the number of bytes to copy, requested by userspace
106            * **offset**: zero at the start of the read, updated as the read
107            * proceeds, it represents how many bytes have been copied so far and
108            * the buffer offset for copying the next record.
109            *
110            * Copy as many buffered i915 perf samples and records for this stream
111            * to userspace as will fit in the given buffer.
112            *
113            * Only write complete records; returning -%ENOSPC if there isn't room
114            * for a complete record.
115            *
116            * Return any error condition that results in a short read such as
117            * -%ENOSPC or -%EFAULT, even though these may be squashed before
118            * returning to userspace.
119            */
120 #ifdef __NetBSD__
121           int (*read)(struct i915_perf_stream *stream,
122                         struct uio *buf,
123                         kauth_cred_t count, /* XXX dummy */
124                         int offset);    /* XXX dummy */
125 #else
126           int (*read)(struct i915_perf_stream *stream,
127                         char __user *buf,
128                         size_t count,
129                         size_t *offset);
130 #endif
131 
132           /**
133            * @destroy: Cleanup any stream specific resources.
134            *
135            * The stream will always be disabled before this is called.
136            */
137           void (*destroy)(struct i915_perf_stream *stream);
138 };
139 
140 /**
141  * struct i915_perf_stream - state for a single open stream FD
142  */
143 struct i915_perf_stream {
144           /**
145            * @perf: i915_perf backpointer
146            */
147           struct i915_perf *perf;
148 
149           /**
150            * @uncore: mmio access path
151            */
152           struct intel_uncore *uncore;
153 
154           /**
155            * @engine: Engine associated with this performance stream.
156            */
157           struct intel_engine_cs *engine;
158 
159           /**
160            * @sample_flags: Flags representing the `DRM_I915_PERF_PROP_SAMPLE_*`
161            * properties given when opening a stream, representing the contents
162            * of a single sample as read() by userspace.
163            */
164           u32 sample_flags;
165 
166           /**
167            * @sample_size: Considering the configured contents of a sample
168            * combined with the required header size, this is the total size
169            * of a single sample record.
170            */
171           int sample_size;
172 
173           /**
174            * @ctx: %NULL if measuring system-wide across all contexts or a
175            * specific context that is being monitored.
176            */
177           struct i915_gem_context *ctx;
178 
179           /**
180            * @enabled: Whether the stream is currently enabled, considering
181            * whether the stream was opened in a disabled state and based
182            * on `I915_PERF_IOCTL_ENABLE` and `I915_PERF_IOCTL_DISABLE` calls.
183            */
184           bool enabled;
185 
186           /**
187            * @hold_preemption: Whether preemption is put on hold for command
188            * submissions done on the @ctx. This is useful for some drivers that
189            * cannot easily post process the OA buffer context to subtract delta
190            * of performance counters not associated with @ctx.
191            */
192           bool hold_preemption;
193 
194           /**
195            * @ops: The callbacks providing the implementation of this specific
196            * type of configured stream.
197            */
198           const struct i915_perf_stream_ops *ops;
199 
200           /**
201            * @oa_config: The OA configuration used by the stream.
202            */
203           struct i915_oa_config *oa_config;
204 
205           /**
206            * @oa_config_bos: A list of struct i915_oa_config_bo allocated lazily
207            * each time @oa_config changes.
208            */
209           struct llist_head oa_config_bos;
210 
211           /**
212            * @pinned_ctx: The OA context specific information.
213            */
214           struct intel_context *pinned_ctx;
215 
216           /**
217            * @specific_ctx_id: The id of the specific context.
218            */
219           u32 specific_ctx_id;
220 
221           /**
222            * @specific_ctx_id_mask: The mask used to masking specific_ctx_id bits.
223            */
224           u32 specific_ctx_id_mask;
225 
226           /**
227            * @poll_check_timer: High resolution timer that will periodically
228            * check for data in the circular OA buffer for notifying userspace
229            * (e.g. during a read() or poll()).
230            */
231           struct hrtimer poll_check_timer;
232 
233           /**
234            * @poll_wq: The wait queue that hrtimer callback wakes when it
235            * sees data ready to read in the circular OA buffer.
236            */
237 #ifdef __NetBSD__
238           drm_waitqueue_t poll_wq;
239           struct selinfo poll_selq;
240 #else
241           wait_queue_head_t poll_wq;
242 #endif
243 
244           /**
245            * @pollin: Whether there is data available to read.
246            */
247           bool pollin;
248 
249           /**
250            * @periodic: Whether periodic sampling is currently enabled.
251            */
252           bool periodic;
253 
254           /**
255            * @period_exponent: The OA unit sampling frequency is derived from this.
256            */
257           int period_exponent;
258 
259           /**
260            * @oa_buffer: State of the OA buffer.
261            */
262           struct {
263                     struct i915_vma *vma;
264                     u8 *vaddr;
265                     u32 last_ctx_id;
266                     int format;
267                     int format_size;
268                     int size_exponent;
269 
270                     /**
271                      * @ptr_lock: Locks reads and writes to all head/tail state
272                      *
273                      * Consider: the head and tail pointer state needs to be read
274                      * consistently from a hrtimer callback (atomic context) and
275                      * read() fop (user context) with tail pointer updates happening
276                      * in atomic context and head updates in user context and the
277                      * (unlikely) possibility of read() errors needing to reset all
278                      * head/tail state.
279                      *
280                      * Note: Contention/performance aren't currently a significant
281                      * concern here considering the relatively low frequency of
282                      * hrtimer callbacks (5ms period) and that reads typically only
283                      * happen in response to a hrtimer event and likely complete
284                      * before the next callback.
285                      *
286                      * Note: This lock is not held *while* reading and copying data
287                      * to userspace so the value of head observed in htrimer
288                      * callbacks won't represent any partial consumption of data.
289                      */
290                     spinlock_t ptr_lock;
291 
292                     /**
293                      * @tails: One 'aging' tail pointer and one 'aged' tail pointer ready to
294                      * used for reading.
295                      *
296                      * Initial values of 0xffffffff are invalid and imply that an
297                      * update is required (and should be ignored by an attempted
298                      * read)
299                      */
300                     struct {
301                               u32 offset;
302                     } tails[2];
303 
304                     /**
305                      * @aged_tail_idx: Index for the aged tail ready to read() data up to.
306                      */
307                     unsigned int aged_tail_idx;
308 
309                     /**
310                      * @aging_timestamp: A monotonic timestamp for when the current aging tail pointer
311                      * was read; used to determine when it is old enough to trust.
312                      */
313                     u64 aging_timestamp;
314 
315                     /**
316                      * @head: Although we can always read back the head pointer register,
317                      * we prefer to avoid trusting the HW state, just to avoid any
318                      * risk that some hardware condition could * somehow bump the
319                      * head pointer unpredictably and cause us to forward the wrong
320                      * OA buffer data to userspace.
321                      */
322                     u32 head;
323           } oa_buffer;
324 
325           /**
326            * @noa_wait: A batch buffer doing a wait on the GPU for the NOA logic to be
327            * reprogrammed.
328            */
329           struct i915_vma *noa_wait;
330 };
331 
332 /**
333  * struct i915_oa_ops - Gen specific implementation of an OA unit stream
334  */
335 struct i915_oa_ops {
336           /**
337            * @is_valid_b_counter_reg: Validates register's address for
338            * programming boolean counters for a particular platform.
339            */
340           bool (*is_valid_b_counter_reg)(struct i915_perf *perf, u32 addr);
341 
342           /**
343            * @is_valid_mux_reg: Validates register's address for programming mux
344            * for a particular platform.
345            */
346           bool (*is_valid_mux_reg)(struct i915_perf *perf, u32 addr);
347 
348           /**
349            * @is_valid_flex_reg: Validates register's address for programming
350            * flex EU filtering for a particular platform.
351            */
352           bool (*is_valid_flex_reg)(struct i915_perf *perf, u32 addr);
353 
354           /**
355            * @enable_metric_set: Selects and applies any MUX configuration to set
356            * up the Boolean and Custom (B/C) counters that are part of the
357            * counter reports being sampled. May apply system constraints such as
358            * disabling EU clock gating as required.
359            */
360           int (*enable_metric_set)(struct i915_perf_stream *stream);
361 
362           /**
363            * @disable_metric_set: Remove system constraints associated with using
364            * the OA unit.
365            */
366           void (*disable_metric_set)(struct i915_perf_stream *stream);
367 
368           /**
369            * @oa_enable: Enable periodic sampling
370            */
371           void (*oa_enable)(struct i915_perf_stream *stream);
372 
373           /**
374            * @oa_disable: Disable periodic sampling
375            */
376           void (*oa_disable)(struct i915_perf_stream *stream);
377 
378           /**
379            * @read: Copy data from the circular OA buffer into a given userspace
380            * buffer.
381            */
382 #ifdef __NetBSD__
383           int (*read)(struct i915_perf_stream *stream,
384                         struct uio *buf,
385                         kauth_cred_t count, /* XXX dummy */
386                         int offset);    /* XXX dummy */
387 #else
388           int (*read)(struct i915_perf_stream *stream,
389                         char __user *buf,
390                         size_t count,
391                         size_t *offset);
392 #endif
393 
394           /**
395            * @oa_hw_tail_read: read the OA tail pointer register
396            *
397            * In particular this enables us to share all the fiddly code for
398            * handling the OA unit tail pointer race that affects multiple
399            * generations.
400            */
401           u32 (*oa_hw_tail_read)(struct i915_perf_stream *stream);
402 };
403 
404 struct i915_perf {
405           struct drm_i915_private *i915;
406 
407           struct kobject *metrics_kobj;
408 
409           /*
410            * Lock associated with adding/modifying/removing OA configs
411            * in perf->metrics_idr.
412            */
413           struct mutex metrics_lock;
414 
415           /*
416            * List of dynamic configurations (struct i915_oa_config), you
417            * need to hold perf->metrics_lock to access it.
418            */
419           struct idr metrics_idr;
420 
421           /*
422            * Lock associated with anything below within this structure
423            * except exclusive_stream.
424            */
425           struct mutex lock;
426 
427           /*
428            * The stream currently using the OA unit. If accessed
429            * outside a syscall associated to its file
430            * descriptor.
431            */
432           struct i915_perf_stream *exclusive_stream;
433 
434           /**
435            * For rate limiting any notifications of spurious
436            * invalid OA reports
437            */
438           struct ratelimit_state spurious_report_rs;
439 
440           struct i915_oa_config test_config;
441 
442           u32 gen7_latched_oastatus1;
443           u32 ctx_oactxctrl_offset;
444           u32 ctx_flexeu0_offset;
445 
446           /**
447            * The RPT_ID/reason field for Gen8+ includes a bit
448            * to determine if the CTX ID in the report is valid
449            * but the specific bit differs between Gen 8 and 9
450            */
451           u32 gen8_valid_ctx_bit;
452 
453           struct i915_oa_ops ops;
454           const struct i915_oa_format *oa_formats;
455 
456           atomic64_t noa_programming_delay;
457 };
458 
459 #endif /* _I915_PERF_TYPES_H_ */
460