Commit | Line | Data |
---|---|---|
a17ae4c3 | 1 | // SPDX-License-Identifier: GPL-2.0 |
8c069ff4 HB |
2 | /* |
3 | * Performance event support for the System z CPU-measurement Sampling Facility | |
4 | * | |
5223c671 | 5 | * Copyright IBM Corp. 2013, 2018 |
8c069ff4 | 6 | * Author(s): Hendrik Brueckner <brueckner@linux.vnet.ibm.com> |
8c069ff4 HB |
7 | */ |
8 | #define KMSG_COMPONENT "cpum_sf" | |
9 | #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt | |
10 | ||
11 | #include <linux/kernel.h> | |
12 | #include <linux/kernel_stat.h> | |
13 | #include <linux/perf_event.h> | |
14 | #include <linux/percpu.h> | |
544e8dd7 | 15 | #include <linux/pid.h> |
8c069ff4 HB |
16 | #include <linux/notifier.h> |
17 | #include <linux/export.h> | |
7e75fc3f | 18 | #include <linux/slab.h> |
69f239ed HB |
19 | #include <linux/mm.h> |
20 | #include <linux/moduleparam.h> | |
8c069ff4 HB |
21 | #include <asm/cpu_mf.h> |
22 | #include <asm/irq.h> | |
23 | #include <asm/debug.h> | |
24 | #include <asm/timex.h> | |
25 | ||
26 | /* Minimum number of sample-data-block-tables: | |
27 | * At least one table is required for the sampling buffer structure. | |
28 | * A single table contains up to 511 pointers to sample-data-blocks. | |
29 | */ | |
69f239ed | 30 | #define CPUM_SF_MIN_SDBT 1 |
8c069ff4 | 31 | |
69f239ed | 32 | /* Number of sample-data-blocks per sample-data-block-table (SDBT): |
7e75fc3f HB |
33 | * A table contains SDB pointers (8 bytes) and one table-link entry |
34 | * that points to the origin of the next SDBT. | |
8c069ff4 | 35 | */ |
69f239ed | 36 | #define CPUM_SF_SDB_PER_TABLE ((PAGE_SIZE - 8) / 8) |
8c069ff4 | 37 | |
69f239ed HB |
38 | /* Maximum page offset for an SDBT table-link entry: |
39 | * If this page offset is reached, a table-link entry to the next SDBT | |
40 | * must be added. | |
41 | */ | |
42 | #define CPUM_SF_SDBT_TL_OFFSET (CPUM_SF_SDB_PER_TABLE * 8) | |
43 | static inline int require_table_link(const void *sdbt) | |
44 | { | |
45 | return ((unsigned long) sdbt & ~PAGE_MASK) == CPUM_SF_SDBT_TL_OFFSET; | |
46 | } | |
47 | ||
48 | /* Minimum and maximum sampling buffer sizes: | |
49 | * | |
7e75fc3f HB |
50 | * This number represents the maximum size of the sampling buffer taking |
51 | * the number of sample-data-block-tables into account. Note that these | |
52 | * numbers apply to the basic-sampling function only. | |
53 | * The maximum number of SDBs is increased by CPUM_SF_SDB_DIAG_FACTOR if | |
54 | * the diagnostic-sampling function is active. | |
8c069ff4 | 55 | * |
69f239ed HB |
56 | * Sampling buffer size Buffer characteristics |
57 | * --------------------------------------------------- | |
58 | * 64KB == 16 pages (4KB per page) | |
59 | * 1 page for SDB-tables | |
60 | * 15 pages for SDBs | |
61 | * | |
62 | * 32MB == 8192 pages (4KB per page) | |
63 | * 16 pages for SDB-tables | |
64 | * 8176 pages for SDBs | |
8c069ff4 | 65 | */ |
69f239ed HB |
66 | static unsigned long __read_mostly CPUM_SF_MIN_SDB = 15; |
67 | static unsigned long __read_mostly CPUM_SF_MAX_SDB = 8176; | |
7e75fc3f | 68 | static unsigned long __read_mostly CPUM_SF_SDB_DIAG_FACTOR = 1; |
8c069ff4 HB |
69 | |
70 | struct sf_buffer { | |
69f239ed | 71 | unsigned long *sdbt; /* Sample-data-block-table origin */ |
8c069ff4 | 72 | /* buffer characteristics (required for buffer increments) */ |
69f239ed HB |
73 | unsigned long num_sdb; /* Number of sample-data-blocks */ |
74 | unsigned long num_sdbt; /* Number of sample-data-block-tables */ | |
75 | unsigned long *tail; /* last sample-data-block-table */ | |
8c069ff4 HB |
76 | }; |
77 | ||
ca5955cd PH |
78 | struct aux_buffer { |
79 | struct sf_buffer sfb; | |
80 | unsigned long head; /* index of SDB of buffer head */ | |
81 | unsigned long alert_mark; /* index of SDB of alert request position */ | |
82 | unsigned long empty_mark; /* mark of SDB not marked full */ | |
83 | unsigned long *sdb_index; /* SDB address for fast lookup */ | |
84 | unsigned long *sdbt_index; /* SDBT address for fast lookup */ | |
85 | }; | |
86 | ||
8c069ff4 HB |
87 | struct cpu_hw_sf { |
88 | /* CPU-measurement sampling information block */ | |
89 | struct hws_qsi_info_block qsi; | |
69f239ed | 90 | /* CPU-measurement sampling control block */ |
8c069ff4 HB |
91 | struct hws_lsctl_request_block lsctl; |
92 | struct sf_buffer sfb; /* Sampling buffer */ | |
93 | unsigned int flags; /* Status flags */ | |
94 | struct perf_event *event; /* Scheduled perf event */ | |
ca5955cd | 95 | struct perf_output_handle handle; /* AUX buffer output handle */ |
8c069ff4 HB |
96 | }; |
97 | static DEFINE_PER_CPU(struct cpu_hw_sf, cpu_hw_sf); | |
98 | ||
99 | /* Debug feature */ | |
100 | static debug_info_t *sfdbg; | |
101 | ||
69f239ed HB |
102 | /* |
103 | * sf_disable() - Switch off sampling facility | |
104 | */ | |
105 | static int sf_disable(void) | |
106 | { | |
107 | struct hws_lsctl_request_block sreq; | |
108 | ||
109 | memset(&sreq, 0, sizeof(sreq)); | |
110 | return lsctl(&sreq); | |
111 | } | |
112 | ||
8c069ff4 HB |
113 | /* |
114 | * sf_buffer_available() - Check for an allocated sampling buffer | |
115 | */ | |
116 | static int sf_buffer_available(struct cpu_hw_sf *cpuhw) | |
117 | { | |
69f239ed | 118 | return !!cpuhw->sfb.sdbt; |
8c069ff4 HB |
119 | } |
120 | ||
121 | /* | |
122 | * deallocate sampling facility buffer | |
123 | */ | |
124 | static void free_sampling_buffer(struct sf_buffer *sfb) | |
125 | { | |
69f239ed | 126 | unsigned long *sdbt, *curr; |
8c069ff4 HB |
127 | |
128 | if (!sfb->sdbt) | |
129 | return; | |
130 | ||
131 | sdbt = sfb->sdbt; | |
69f239ed | 132 | curr = sdbt; |
8c069ff4 | 133 | |
69f239ed | 134 | /* Free the SDBT after all SDBs are processed... */ |
8c069ff4 HB |
135 | while (1) { |
136 | if (!*curr || !sdbt) | |
137 | break; | |
138 | ||
69f239ed | 139 | /* Process table-link entries */ |
8c069ff4 HB |
140 | if (is_link_entry(curr)) { |
141 | curr = get_next_sdbt(curr); | |
142 | if (sdbt) | |
69f239ed | 143 | free_page((unsigned long) sdbt); |
8c069ff4 | 144 | |
69f239ed HB |
145 | /* If the origin is reached, sampling buffer is freed */ |
146 | if (curr == sfb->sdbt) | |
8c069ff4 HB |
147 | break; |
148 | else | |
69f239ed | 149 | sdbt = curr; |
8c069ff4 | 150 | } else { |
69f239ed | 151 | /* Process SDB pointer */ |
8c069ff4 HB |
152 | if (*curr) { |
153 | free_page(*curr); | |
154 | curr++; | |
155 | } | |
156 | } | |
157 | } | |
158 | ||
532da3de TR |
159 | debug_sprintf_event(sfdbg, 5, "%s: freed sdbt %#lx\n", __func__, |
160 | (unsigned long)sfb->sdbt); | |
8c069ff4 HB |
161 | memset(sfb, 0, sizeof(*sfb)); |
162 | } | |
163 | ||
69f239ed HB |
164 | static int alloc_sample_data_block(unsigned long *sdbt, gfp_t gfp_flags) |
165 | { | |
166 | unsigned long sdb, *trailer; | |
167 | ||
168 | /* Allocate and initialize sample-data-block */ | |
169 | sdb = get_zeroed_page(gfp_flags); | |
170 | if (!sdb) | |
171 | return -ENOMEM; | |
172 | trailer = trailer_entry_ptr(sdb); | |
173 | *trailer = SDB_TE_ALERT_REQ_MASK; | |
174 | ||
175 | /* Link SDB into the sample-data-block-table */ | |
176 | *sdbt = sdb; | |
177 | ||
178 | return 0; | |
179 | } | |
180 | ||
181 | /* | |
182 | * realloc_sampling_buffer() - extend sampler memory | |
183 | * | |
184 | * Allocates new sample-data-blocks and adds them to the specified sampling | |
185 | * buffer memory. | |
186 | * | |
187 | * Important: This modifies the sampling buffer and must be called when the | |
188 | * sampling facility is disabled. | |
189 | * | |
190 | * Returns zero on success, non-zero otherwise. | |
191 | */ | |
192 | static int realloc_sampling_buffer(struct sf_buffer *sfb, | |
193 | unsigned long num_sdb, gfp_t gfp_flags) | |
194 | { | |
195 | int i, rc; | |
247f265f | 196 | unsigned long *new, *tail, *tail_prev = NULL; |
69f239ed HB |
197 | |
198 | if (!sfb->sdbt || !sfb->tail) | |
199 | return -EINVAL; | |
200 | ||
201 | if (!is_link_entry(sfb->tail)) | |
202 | return -EINVAL; | |
203 | ||
204 | /* Append to the existing sampling buffer, overwriting the table-link | |
205 | * register. | |
206 | * The tail variables always points to the "tail" (last and table-link) | |
207 | * entry in an SDB-table. | |
208 | */ | |
209 | tail = sfb->tail; | |
210 | ||
211 | /* Do a sanity check whether the table-link entry points to | |
212 | * the sampling buffer origin. | |
213 | */ | |
214 | if (sfb->sdbt != get_next_sdbt(tail)) { | |
c1838834 | 215 | debug_sprintf_event(sfdbg, 3, "%s: " |
532da3de TR |
216 | "sampling buffer is not linked: origin %#lx" |
217 | " tail %#lx\n", __func__, | |
218 | (unsigned long)sfb->sdbt, | |
219 | (unsigned long)tail); | |
69f239ed HB |
220 | return -EINVAL; |
221 | } | |
222 | ||
223 | /* Allocate remaining SDBs */ | |
224 | rc = 0; | |
225 | for (i = 0; i < num_sdb; i++) { | |
226 | /* Allocate a new SDB-table if it is full. */ | |
227 | if (require_table_link(tail)) { | |
228 | new = (unsigned long *) get_zeroed_page(gfp_flags); | |
229 | if (!new) { | |
230 | rc = -ENOMEM; | |
231 | break; | |
232 | } | |
233 | sfb->num_sdbt++; | |
234 | /* Link current page to tail of chain */ | |
235 | *tail = (unsigned long)(void *) new + 1; | |
247f265f | 236 | tail_prev = tail; |
69f239ed HB |
237 | tail = new; |
238 | } | |
239 | ||
240 | /* Allocate a new sample-data-block. | |
241 | * If there is not enough memory, stop the realloc process | |
242 | * and simply use what was allocated. If this is a temporary | |
243 | * issue, a new realloc call (if required) might succeed. | |
244 | */ | |
245 | rc = alloc_sample_data_block(tail, gfp_flags); | |
247f265f TR |
246 | if (rc) { |
247 | /* Undo last SDBT. An SDBT with no SDB at its first | |
248 | * entry but with an SDBT entry instead can not be | |
249 | * handled by the interrupt handler code. | |
250 | * Avoid this situation. | |
251 | */ | |
252 | if (tail_prev) { | |
253 | sfb->num_sdbt--; | |
254 | free_page((unsigned long) new); | |
255 | tail = tail_prev; | |
256 | } | |
69f239ed | 257 | break; |
247f265f | 258 | } |
69f239ed HB |
259 | sfb->num_sdb++; |
260 | tail++; | |
247f265f | 261 | tail_prev = new = NULL; /* Allocated at least one SBD */ |
69f239ed HB |
262 | } |
263 | ||
264 | /* Link sampling buffer to its origin */ | |
265 | *tail = (unsigned long) sfb->sdbt + 1; | |
266 | sfb->tail = tail; | |
267 | ||
532da3de TR |
268 | debug_sprintf_event(sfdbg, 4, "%s: new buffer" |
269 | " settings: sdbt %lu sdb %lu\n", __func__, | |
69f239ed HB |
270 | sfb->num_sdbt, sfb->num_sdb); |
271 | return rc; | |
272 | } | |
273 | ||
8c069ff4 HB |
274 | /* |
275 | * allocate_sampling_buffer() - allocate sampler memory | |
276 | * | |
277 | * Allocates and initializes a sampling buffer structure using the | |
278 | * specified number of sample-data-blocks (SDB). For each allocation, | |
279 | * a 4K page is used. The number of sample-data-block-tables (SDBT) | |
280 | * are calculated from SDBs. | |
281 | * Also set the ALERT_REQ mask in each SDBs trailer. | |
282 | * | |
283 | * Returns zero on success, non-zero otherwise. | |
284 | */ | |
285 | static int alloc_sampling_buffer(struct sf_buffer *sfb, unsigned long num_sdb) | |
286 | { | |
69f239ed | 287 | int rc; |
8c069ff4 HB |
288 | |
289 | if (sfb->sdbt) | |
290 | return -EINVAL; | |
69f239ed HB |
291 | |
292 | /* Allocate the sample-data-block-table origin */ | |
293 | sfb->sdbt = (unsigned long *) get_zeroed_page(GFP_KERNEL); | |
294 | if (!sfb->sdbt) | |
295 | return -ENOMEM; | |
8c069ff4 | 296 | sfb->num_sdb = 0; |
69f239ed | 297 | sfb->num_sdbt = 1; |
8c069ff4 | 298 | |
69f239ed HB |
299 | /* Link the table origin to point to itself to prepare for |
300 | * realloc_sampling_buffer() invocation. | |
301 | */ | |
302 | sfb->tail = sfb->sdbt; | |
303 | *sfb->tail = (unsigned long)(void *) sfb->sdbt + 1; | |
8c069ff4 | 304 | |
69f239ed HB |
305 | /* Allocate requested number of sample-data-blocks */ |
306 | rc = realloc_sampling_buffer(sfb, num_sdb, GFP_KERNEL); | |
307 | if (rc) { | |
308 | free_sampling_buffer(sfb); | |
532da3de TR |
309 | debug_sprintf_event(sfdbg, 4, "%s: " |
310 | "realloc_sampling_buffer failed with rc %i\n", | |
311 | __func__, rc); | |
69f239ed HB |
312 | } else |
313 | debug_sprintf_event(sfdbg, 4, | |
532da3de TR |
314 | "%s: tear %#lx dear %#lx\n", __func__, |
315 | (unsigned long)sfb->sdbt, (unsigned long)*sfb->sdbt); | |
69f239ed HB |
316 | return rc; |
317 | } | |
8c069ff4 | 318 | |
69f239ed HB |
319 | static void sfb_set_limits(unsigned long min, unsigned long max) |
320 | { | |
7e75fc3f HB |
321 | struct hws_qsi_info_block si; |
322 | ||
69f239ed HB |
323 | CPUM_SF_MIN_SDB = min; |
324 | CPUM_SF_MAX_SDB = max; | |
7e75fc3f HB |
325 | |
326 | memset(&si, 0, sizeof(si)); | |
327 | if (!qsi(&si)) | |
328 | CPUM_SF_SDB_DIAG_FACTOR = DIV_ROUND_UP(si.dsdes, si.bsdes); | |
329 | } | |
330 | ||
331 | static unsigned long sfb_max_limit(struct hw_perf_event *hwc) | |
332 | { | |
333 | return SAMPL_DIAG_MODE(hwc) ? CPUM_SF_MAX_SDB * CPUM_SF_SDB_DIAG_FACTOR | |
334 | : CPUM_SF_MAX_SDB; | |
69f239ed | 335 | } |
8c069ff4 | 336 | |
69f239ed HB |
337 | static unsigned long sfb_pending_allocs(struct sf_buffer *sfb, |
338 | struct hw_perf_event *hwc) | |
339 | { | |
340 | if (!sfb->sdbt) | |
341 | return SFB_ALLOC_REG(hwc); | |
342 | if (SFB_ALLOC_REG(hwc) > sfb->num_sdb) | |
343 | return SFB_ALLOC_REG(hwc) - sfb->num_sdb; | |
344 | return 0; | |
345 | } | |
8c069ff4 | 346 | |
69f239ed HB |
347 | static int sfb_has_pending_allocs(struct sf_buffer *sfb, |
348 | struct hw_perf_event *hwc) | |
349 | { | |
350 | return sfb_pending_allocs(sfb, hwc) > 0; | |
351 | } | |
8c069ff4 | 352 | |
69f239ed HB |
353 | static void sfb_account_allocs(unsigned long num, struct hw_perf_event *hwc) |
354 | { | |
7e75fc3f HB |
355 | /* Limit the number of SDBs to not exceed the maximum */ |
356 | num = min_t(unsigned long, num, sfb_max_limit(hwc) - SFB_ALLOC_REG(hwc)); | |
69f239ed HB |
357 | if (num) |
358 | SFB_ALLOC_REG(hwc) += num; | |
8c069ff4 HB |
359 | } |
360 | ||
69f239ed HB |
361 | static void sfb_init_allocs(unsigned long num, struct hw_perf_event *hwc) |
362 | { | |
363 | SFB_ALLOC_REG(hwc) = 0; | |
364 | sfb_account_allocs(num, hwc); | |
365 | } | |
366 | ||
7e75fc3f HB |
367 | static void deallocate_buffers(struct cpu_hw_sf *cpuhw) |
368 | { | |
369 | if (cpuhw->sfb.sdbt) | |
370 | free_sampling_buffer(&cpuhw->sfb); | |
371 | } | |
372 | ||
373 | static int allocate_buffers(struct cpu_hw_sf *cpuhw, struct hw_perf_event *hwc) | |
8c069ff4 | 374 | { |
0d6f1693 | 375 | unsigned long n_sdb, freq; |
3d43b981 | 376 | size_t sample_size; |
8c069ff4 HB |
377 | |
378 | /* Calculate sampling buffers using 4K pages | |
379 | * | |
0d6f1693 TR |
380 | * 1. The sampling size is 32 bytes for basic sampling. This size |
381 | * is the same for all machine types. Diagnostic | |
382 | * sampling uses auxlilary data buffer setup which provides the | |
383 | * memory for SDBs using linux common code auxiliary trace | |
384 | * setup. | |
7e75fc3f | 385 | * |
0d6f1693 | 386 | * 2. Function alloc_sampling_buffer() sets the Alert Request |
7e75fc3f | 387 | * Control indicator to trigger a measurement-alert to harvest |
0d6f1693 TR |
388 | * sample-data-blocks (SDB). This is done per SDB. This |
389 | * measurement alert interrupt fires quick enough to handle | |
390 | * one SDB, on very high frequency and work loads there might | |
391 | * be 2 to 3 SBDs available for sample processing. | |
392 | * Currently there is no need for setup alert request on every | |
393 | * n-th page. This is counterproductive as one IRQ triggers | |
394 | * a very high number of samples to be processed at one IRQ. | |
8c069ff4 | 395 | * |
0d6f1693 TR |
396 | * 3. Use the sampling frequency as input. |
397 | * Compute the number of SDBs and ensure a minimum | |
398 | * of CPUM_SF_MIN_SDB. Depending on frequency add some more | |
399 | * SDBs to handle a higher sampling rate. | |
400 | * Use a minimum of CPUM_SF_MIN_SDB and allow for 100 samples | |
401 | * (one SDB) for every 10000 HZ frequency increment. | |
8c069ff4 | 402 | * |
7e75fc3f HB |
403 | * 4. Compute the number of sample-data-block-tables (SDBT) and |
404 | * ensure a minimum of CPUM_SF_MIN_SDBT (one table can manage up | |
405 | * to 511 SDBs). | |
8c069ff4 | 406 | */ |
3d43b981 | 407 | sample_size = sizeof(struct hws_basic_entry); |
8c069ff4 | 408 | freq = sample_rate_to_freq(&cpuhw->qsi, SAMPL_RATE(hwc)); |
0d6f1693 | 409 | n_sdb = CPUM_SF_MIN_SDB + DIV_ROUND_UP(freq, 10000); |
8c069ff4 | 410 | |
69f239ed HB |
411 | /* If there is already a sampling buffer allocated, it is very likely |
412 | * that the sampling facility is enabled too. If the event to be | |
413 | * initialized requires a greater sampling buffer, the allocation must | |
414 | * be postponed. Changing the sampling buffer requires the sampling | |
415 | * facility to be in the disabled state. So, account the number of | |
416 | * required SDBs and let cpumsf_pmu_enable() resize the buffer just | |
417 | * before the event is started. | |
8c069ff4 | 418 | */ |
69f239ed | 419 | sfb_init_allocs(n_sdb, hwc); |
8c069ff4 HB |
420 | if (sf_buffer_available(cpuhw)) |
421 | return 0; | |
422 | ||
423 | debug_sprintf_event(sfdbg, 3, | |
c1838834 TR |
424 | "%s: rate %lu f %lu sdb %lu/%lu" |
425 | " sample_size %lu cpuhw %p\n", __func__, | |
7e75fc3f HB |
426 | SAMPL_RATE(hwc), freq, n_sdb, sfb_max_limit(hwc), |
427 | sample_size, cpuhw); | |
8c069ff4 HB |
428 | |
429 | return alloc_sampling_buffer(&cpuhw->sfb, | |
69f239ed | 430 | sfb_pending_allocs(&cpuhw->sfb, hwc)); |
8c069ff4 HB |
431 | } |
432 | ||
69f239ed HB |
433 | static unsigned long min_percent(unsigned int percent, unsigned long base, |
434 | unsigned long min) | |
435 | { | |
436 | return min_t(unsigned long, min, DIV_ROUND_UP(percent * base, 100)); | |
437 | } | |
8c069ff4 | 438 | |
69f239ed HB |
439 | static unsigned long compute_sfb_extent(unsigned long ratio, unsigned long base) |
440 | { | |
441 | /* Use a percentage-based approach to extend the sampling facility | |
442 | * buffer. Accept up to 5% sample data loss. | |
443 | * Vary the extents between 1% to 5% of the current number of | |
444 | * sample-data-blocks. | |
445 | */ | |
446 | if (ratio <= 5) | |
447 | return 0; | |
448 | if (ratio <= 25) | |
449 | return min_percent(1, base, 1); | |
450 | if (ratio <= 50) | |
451 | return min_percent(1, base, 1); | |
452 | if (ratio <= 75) | |
453 | return min_percent(2, base, 2); | |
454 | if (ratio <= 100) | |
455 | return min_percent(3, base, 3); | |
456 | if (ratio <= 250) | |
457 | return min_percent(4, base, 4); | |
458 | ||
459 | return min_percent(5, base, 8); | |
460 | } | |
8c069ff4 | 461 | |
69f239ed HB |
462 | static void sfb_account_overflows(struct cpu_hw_sf *cpuhw, |
463 | struct hw_perf_event *hwc) | |
464 | { | |
465 | unsigned long ratio, num; | |
466 | ||
467 | if (!OVERFLOW_REG(hwc)) | |
468 | return; | |
469 | ||
470 | /* The sample_overflow contains the average number of sample data | |
471 | * that has been lost because sample-data-blocks were full. | |
472 | * | |
473 | * Calculate the total number of sample data entries that has been | |
474 | * discarded. Then calculate the ratio of lost samples to total samples | |
475 | * per second in percent. | |
476 | */ | |
477 | ratio = DIV_ROUND_UP(100 * OVERFLOW_REG(hwc) * cpuhw->sfb.num_sdb, | |
478 | sample_rate_to_freq(&cpuhw->qsi, SAMPL_RATE(hwc))); | |
479 | ||
480 | /* Compute number of sample-data-blocks */ | |
481 | num = compute_sfb_extent(ratio, cpuhw->sfb.num_sdb); | |
482 | if (num) | |
483 | sfb_account_allocs(num, hwc); | |
484 | ||
532da3de TR |
485 | debug_sprintf_event(sfdbg, 5, "%s: overflow %llu ratio %lu num %lu\n", |
486 | __func__, OVERFLOW_REG(hwc), ratio, num); | |
69f239ed HB |
487 | OVERFLOW_REG(hwc) = 0; |
488 | } | |
489 | ||
490 | /* extend_sampling_buffer() - Extend sampling buffer | |
491 | * @sfb: Sampling buffer structure (for local CPU) | |
492 | * @hwc: Perf event hardware structure | |
493 | * | |
494 | * Use this function to extend the sampling buffer based on the overflow counter | |
495 | * and postponed allocation extents stored in the specified Perf event hardware. | |
496 | * | |
497 | * Important: This function disables the sampling facility in order to safely | |
498 | * change the sampling buffer structure. Do not call this function | |
499 | * when the PMU is active. | |
8c069ff4 | 500 | */ |
69f239ed HB |
501 | static void extend_sampling_buffer(struct sf_buffer *sfb, |
502 | struct hw_perf_event *hwc) | |
8c069ff4 | 503 | { |
69f239ed HB |
504 | unsigned long num, num_old; |
505 | int rc; | |
8c069ff4 | 506 | |
69f239ed HB |
507 | num = sfb_pending_allocs(sfb, hwc); |
508 | if (!num) | |
509 | return; | |
510 | num_old = sfb->num_sdb; | |
511 | ||
512 | /* Disable the sampling facility to reset any states and also | |
513 | * clear pending measurement alerts. | |
514 | */ | |
515 | sf_disable(); | |
516 | ||
517 | /* Extend the sampling buffer. | |
518 | * This memory allocation typically happens in an atomic context when | |
519 | * called by perf. Because this is a reallocation, it is fine if the | |
520 | * new SDB-request cannot be satisfied immediately. | |
521 | */ | |
522 | rc = realloc_sampling_buffer(sfb, num, GFP_ATOMIC); | |
523 | if (rc) | |
532da3de TR |
524 | debug_sprintf_event(sfdbg, 5, "%s: realloc failed with rc %i\n", |
525 | __func__, rc); | |
69f239ed HB |
526 | |
527 | if (sfb_has_pending_allocs(sfb, hwc)) | |
532da3de | 528 | debug_sprintf_event(sfdbg, 5, "%s: " |
d98b5d07 | 529 | "req %lu alloc %lu remaining %lu\n", |
532da3de | 530 | __func__, num, sfb->num_sdb - num_old, |
69f239ed | 531 | sfb_pending_allocs(sfb, hwc)); |
8c069ff4 HB |
532 | } |
533 | ||
69f239ed HB |
534 | /* Number of perf events counting hardware events */ |
535 | static atomic_t num_events; | |
536 | /* Used to avoid races in calling reserve/release_cpumf_hardware */ | |
537 | static DEFINE_MUTEX(pmc_reserve_mutex); | |
538 | ||
8c069ff4 HB |
539 | #define PMC_INIT 0 |
540 | #define PMC_RELEASE 1 | |
e28bb79d | 541 | #define PMC_FAILURE 2 |
8c069ff4 HB |
542 | static void setup_pmc_cpu(void *flags) |
543 | { | |
544 | int err; | |
eb7e7d76 | 545 | struct cpu_hw_sf *cpusf = this_cpu_ptr(&cpu_hw_sf); |
8c069ff4 | 546 | |
8c069ff4 HB |
547 | err = 0; |
548 | switch (*((int *) flags)) { | |
549 | case PMC_INIT: | |
550 | memset(cpusf, 0, sizeof(*cpusf)); | |
551 | err = qsi(&cpusf->qsi); | |
552 | if (err) | |
553 | break; | |
554 | cpusf->flags |= PMU_F_RESERVED; | |
555 | err = sf_disable(); | |
556 | if (err) | |
557 | pr_err("Switching off the sampling facility failed " | |
d98b5d07 | 558 | "with rc %i\n", err); |
8c069ff4 | 559 | debug_sprintf_event(sfdbg, 5, |
c1838834 | 560 | "%s: initialized: cpuhw %p\n", __func__, |
d98b5d07 | 561 | cpusf); |
8c069ff4 HB |
562 | break; |
563 | case PMC_RELEASE: | |
564 | cpusf->flags &= ~PMU_F_RESERVED; | |
565 | err = sf_disable(); | |
566 | if (err) { | |
567 | pr_err("Switching off the sampling facility failed " | |
d98b5d07 | 568 | "with rc %i\n", err); |
7e75fc3f HB |
569 | } else |
570 | deallocate_buffers(cpusf); | |
8c069ff4 | 571 | debug_sprintf_event(sfdbg, 5, |
c1838834 | 572 | "%s: released: cpuhw %p\n", __func__, |
d98b5d07 | 573 | cpusf); |
8c069ff4 HB |
574 | break; |
575 | } | |
e28bb79d HB |
576 | if (err) |
577 | *((int *) flags) |= PMC_FAILURE; | |
8c069ff4 HB |
578 | } |
579 | ||
580 | static void release_pmc_hardware(void) | |
581 | { | |
582 | int flags = PMC_RELEASE; | |
583 | ||
584 | irq_subclass_unregister(IRQ_SUBCLASS_MEASUREMENT_ALERT); | |
585 | on_each_cpu(setup_pmc_cpu, &flags, 1); | |
586 | } | |
587 | ||
588 | static int reserve_pmc_hardware(void) | |
589 | { | |
590 | int flags = PMC_INIT; | |
591 | ||
592 | on_each_cpu(setup_pmc_cpu, &flags, 1); | |
e28bb79d HB |
593 | if (flags & PMC_FAILURE) { |
594 | release_pmc_hardware(); | |
595 | return -ENODEV; | |
596 | } | |
8c069ff4 HB |
597 | irq_subclass_register(IRQ_SUBCLASS_MEASUREMENT_ALERT); |
598 | ||
599 | return 0; | |
600 | } | |
601 | ||
602 | static void hw_perf_event_destroy(struct perf_event *event) | |
603 | { | |
604 | /* Release PMC if this is the last perf event */ | |
605 | if (!atomic_add_unless(&num_events, -1, 1)) { | |
606 | mutex_lock(&pmc_reserve_mutex); | |
607 | if (atomic_dec_return(&num_events) == 0) | |
608 | release_pmc_hardware(); | |
609 | mutex_unlock(&pmc_reserve_mutex); | |
610 | } | |
611 | } | |
612 | ||
613 | static void hw_init_period(struct hw_perf_event *hwc, u64 period) | |
614 | { | |
615 | hwc->sample_period = period; | |
616 | hwc->last_period = hwc->sample_period; | |
617 | local64_set(&hwc->period_left, hwc->sample_period); | |
618 | } | |
619 | ||
8c069ff4 HB |
620 | static unsigned long hw_limit_rate(const struct hws_qsi_info_block *si, |
621 | unsigned long rate) | |
622 | { | |
69f239ed HB |
623 | return clamp_t(unsigned long, rate, |
624 | si->min_sampl_rate, si->max_sampl_rate); | |
8c069ff4 HB |
625 | } |
626 | ||
544e8dd7 HB |
627 | static u32 cpumsf_pid_type(struct perf_event *event, |
628 | u32 pid, enum pid_type type) | |
629 | { | |
630 | struct task_struct *tsk; | |
631 | ||
632 | /* Idle process */ | |
633 | if (!pid) | |
634 | goto out; | |
635 | ||
636 | tsk = find_task_by_pid_ns(pid, &init_pid_ns); | |
637 | pid = -1; | |
638 | if (tsk) { | |
639 | /* | |
640 | * Only top level events contain the pid namespace in which | |
641 | * they are created. | |
642 | */ | |
643 | if (event->parent) | |
644 | event = event->parent; | |
645 | pid = __task_pid_nr_ns(tsk, type, event->ns); | |
646 | /* | |
647 | * See also 1d953111b648 | |
648 | * "perf/core: Don't report zero PIDs for exiting tasks". | |
649 | */ | |
650 | if (!pid && !pid_alive(tsk)) | |
651 | pid = -1; | |
652 | } | |
653 | out: | |
654 | return pid; | |
655 | } | |
656 | ||
657 | static void cpumsf_output_event_pid(struct perf_event *event, | |
658 | struct perf_sample_data *data, | |
659 | struct pt_regs *regs) | |
660 | { | |
661 | u32 pid; | |
662 | struct perf_event_header header; | |
663 | struct perf_output_handle handle; | |
664 | ||
665 | /* | |
666 | * Obtain the PID from the basic-sampling data entry and | |
667 | * correct the data->tid_entry.pid value. | |
668 | */ | |
669 | pid = data->tid_entry.pid; | |
670 | ||
671 | /* Protect callchain buffers, tasks */ | |
672 | rcu_read_lock(); | |
673 | ||
674 | perf_prepare_sample(&header, data, event, regs); | |
267fb273 | 675 | if (perf_output_begin(&handle, data, event, header.size)) |
544e8dd7 HB |
676 | goto out; |
677 | ||
678 | /* Update the process ID (see also kernel/events/core.c) */ | |
6883f81a | 679 | data->tid_entry.pid = cpumsf_pid_type(event, pid, PIDTYPE_TGID); |
544e8dd7 HB |
680 | data->tid_entry.tid = cpumsf_pid_type(event, pid, PIDTYPE_PID); |
681 | ||
682 | perf_output_sample(&handle, &header, data, event); | |
683 | perf_output_end(&handle); | |
684 | out: | |
685 | rcu_read_unlock(); | |
686 | } | |
687 | ||
2cb549a8 TR |
688 | static unsigned long getrate(bool freq, unsigned long sample, |
689 | struct hws_qsi_info_block *si) | |
690 | { | |
691 | unsigned long rate; | |
692 | ||
693 | if (freq) { | |
694 | rate = freq_to_sample_rate(si, sample); | |
695 | rate = hw_limit_rate(si, rate); | |
696 | } else { | |
697 | /* The min/max sampling rates specifies the valid range | |
698 | * of sample periods. If the specified sample period is | |
699 | * out of range, limit the period to the range boundary. | |
700 | */ | |
701 | rate = hw_limit_rate(si, sample); | |
702 | ||
703 | /* The perf core maintains a maximum sample rate that is | |
704 | * configurable through the sysctl interface. Ensure the | |
705 | * sampling rate does not exceed this value. This also helps | |
706 | * to avoid throttling when pushing samples with | |
707 | * perf_event_overflow(). | |
708 | */ | |
709 | if (sample_rate_to_freq(si, rate) > | |
710 | sysctl_perf_event_sample_rate) { | |
532da3de | 711 | debug_sprintf_event(sfdbg, 1, "%s: " |
2cb549a8 | 712 | "Sampling rate exceeds maximum " |
532da3de | 713 | "perf sample rate\n", __func__); |
2cb549a8 TR |
714 | rate = 0; |
715 | } | |
716 | } | |
717 | return rate; | |
718 | } | |
719 | ||
720 | /* The sampling information (si) contains information about the | |
721 | * min/max sampling intervals and the CPU speed. So calculate the | |
722 | * correct sampling interval and avoid the whole period adjust | |
723 | * feedback loop. | |
724 | * | |
725 | * Since the CPU Measurement sampling facility can not handle frequency | |
726 | * calculate the sampling interval when frequency is specified using | |
727 | * this formula: | |
728 | * interval := cpu_speed * 1000000 / sample_freq | |
729 | * | |
730 | * Returns errno on bad input and zero on success with parameter interval | |
731 | * set to the correct sampling rate. | |
732 | * | |
733 | * Note: This function turns off freq bit to avoid calling function | |
734 | * perf_adjust_period(). This causes frequency adjustment in the common | |
735 | * code part which causes tremendous variations in the counter values. | |
736 | */ | |
737 | static int __hw_perf_event_init_rate(struct perf_event *event, | |
738 | struct hws_qsi_info_block *si) | |
739 | { | |
740 | struct perf_event_attr *attr = &event->attr; | |
741 | struct hw_perf_event *hwc = &event->hw; | |
742 | unsigned long rate; | |
743 | ||
744 | if (attr->freq) { | |
745 | if (!attr->sample_freq) | |
746 | return -EINVAL; | |
747 | rate = getrate(attr->freq, attr->sample_freq, si); | |
748 | attr->freq = 0; /* Don't call perf_adjust_period() */ | |
749 | SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_FREQ_MODE; | |
750 | } else { | |
751 | rate = getrate(attr->freq, attr->sample_period, si); | |
752 | if (!rate) | |
753 | return -EINVAL; | |
754 | } | |
755 | attr->sample_period = rate; | |
756 | SAMPL_RATE(hwc) = rate; | |
757 | hw_init_period(hwc, SAMPL_RATE(hwc)); | |
532da3de TR |
758 | debug_sprintf_event(sfdbg, 4, "%s: cpu %d period %#llx freq %d,%#lx\n", |
759 | __func__, event->cpu, event->attr.sample_period, | |
760 | event->attr.freq, SAMPLE_FREQ_MODE(hwc)); | |
2cb549a8 TR |
761 | return 0; |
762 | } | |
763 | ||
8c069ff4 HB |
764 | static int __hw_perf_event_init(struct perf_event *event) |
765 | { | |
766 | struct cpu_hw_sf *cpuhw; | |
767 | struct hws_qsi_info_block si; | |
768 | struct perf_event_attr *attr = &event->attr; | |
769 | struct hw_perf_event *hwc = &event->hw; | |
8c069ff4 HB |
770 | int cpu, err; |
771 | ||
772 | /* Reserve CPU-measurement sampling facility */ | |
773 | err = 0; | |
774 | if (!atomic_inc_not_zero(&num_events)) { | |
775 | mutex_lock(&pmc_reserve_mutex); | |
776 | if (atomic_read(&num_events) == 0 && reserve_pmc_hardware()) | |
777 | err = -EBUSY; | |
778 | else | |
779 | atomic_inc(&num_events); | |
780 | mutex_unlock(&pmc_reserve_mutex); | |
781 | } | |
782 | event->destroy = hw_perf_event_destroy; | |
783 | ||
784 | if (err) | |
785 | goto out; | |
786 | ||
787 | /* Access per-CPU sampling information (query sampling info) */ | |
788 | /* | |
789 | * The event->cpu value can be -1 to count on every CPU, for example, | |
790 | * when attaching to a task. If this is specified, use the query | |
791 | * sampling info from the current CPU, otherwise use event->cpu to | |
792 | * retrieve the per-CPU information. | |
793 | * Later, cpuhw indicates whether to allocate sampling buffers for a | |
794 | * particular CPU (cpuhw!=NULL) or each online CPU (cpuw==NULL). | |
795 | */ | |
796 | memset(&si, 0, sizeof(si)); | |
797 | cpuhw = NULL; | |
798 | if (event->cpu == -1) | |
799 | qsi(&si); | |
800 | else { | |
801 | /* Event is pinned to a particular CPU, retrieve the per-CPU | |
802 | * sampling structure for accessing the CPU-specific QSI. | |
803 | */ | |
804 | cpuhw = &per_cpu(cpu_hw_sf, event->cpu); | |
805 | si = cpuhw->qsi; | |
806 | } | |
807 | ||
808 | /* Check sampling facility authorization and, if not authorized, | |
809 | * fall back to other PMUs. It is safe to check any CPU because | |
810 | * the authorization is identical for all configured CPUs. | |
811 | */ | |
812 | if (!si.as) { | |
813 | err = -ENOENT; | |
814 | goto out; | |
815 | } | |
816 | ||
932bfc5a TR |
817 | if (si.ribm & CPU_MF_SF_RIBM_NOTAV) { |
818 | pr_warn("CPU Measurement Facility sampling is temporarily not available\n"); | |
819 | err = -EBUSY; | |
820 | goto out; | |
821 | } | |
822 | ||
7e75fc3f HB |
823 | /* Always enable basic sampling */ |
824 | SAMPL_FLAGS(hwc) = PERF_CPUM_SF_BASIC_MODE; | |
825 | ||
826 | /* Check if diagnostic sampling is requested. Deny if the required | |
827 | * sampling authorization is missing. | |
828 | */ | |
829 | if (attr->config == PERF_EVENT_CPUM_SF_DIAG) { | |
830 | if (!si.ad) { | |
831 | err = -EPERM; | |
832 | goto out; | |
833 | } | |
834 | SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_DIAG_MODE; | |
835 | } | |
836 | ||
d7528862 HB |
837 | /* Check and set other sampling flags */ |
838 | if (attr->config1 & PERF_CPUM_SF_FULL_BLOCKS) | |
839 | SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_FULL_BLOCKS; | |
840 | ||
2cb549a8 TR |
841 | err = __hw_perf_event_init_rate(event, &si); |
842 | if (err) | |
843 | goto out; | |
8c069ff4 | 844 | |
69f239ed HB |
845 | /* Initialize sample data overflow accounting */ |
846 | hwc->extra_reg.reg = REG_OVERFLOW; | |
847 | OVERFLOW_REG(hwc) = 0; | |
848 | ||
cbf6948f PH |
849 | /* Use AUX buffer. No need to allocate it by ourself */ |
850 | if (attr->config == PERF_EVENT_CPUM_SF_DIAG) | |
851 | return 0; | |
852 | ||
8c069ff4 HB |
853 | /* Allocate the per-CPU sampling buffer using the CPU information |
854 | * from the event. If the event is not pinned to a particular | |
855 | * CPU (event->cpu == -1; or cpuhw == NULL), allocate sampling | |
856 | * buffers for each online CPU. | |
857 | */ | |
858 | if (cpuhw) | |
859 | /* Event is pinned to a particular CPU */ | |
7e75fc3f | 860 | err = allocate_buffers(cpuhw, hwc); |
8c069ff4 HB |
861 | else { |
862 | /* Event is not pinned, allocate sampling buffer on | |
863 | * each online CPU | |
864 | */ | |
865 | for_each_online_cpu(cpu) { | |
866 | cpuhw = &per_cpu(cpu_hw_sf, cpu); | |
7e75fc3f | 867 | err = allocate_buffers(cpuhw, hwc); |
8c069ff4 HB |
868 | if (err) |
869 | break; | |
870 | } | |
871 | } | |
544e8dd7 HB |
872 | |
873 | /* If PID/TID sampling is active, replace the default overflow | |
874 | * handler to extract and resolve the PIDs from the basic-sampling | |
875 | * data entries. | |
876 | */ | |
877 | if (event->attr.sample_type & PERF_SAMPLE_TID) | |
878 | if (is_default_overflow_handler(event)) | |
879 | event->overflow_handler = cpumsf_output_event_pid; | |
8c069ff4 HB |
880 | out: |
881 | return err; | |
882 | } | |
883 | ||
5aa98879 TR |
884 | static bool is_callchain_event(struct perf_event *event) |
885 | { | |
886 | u64 sample_type = event->attr.sample_type; | |
887 | ||
888 | return sample_type & (PERF_SAMPLE_CALLCHAIN | PERF_SAMPLE_REGS_USER | | |
889 | PERF_SAMPLE_STACK_USER); | |
890 | } | |
891 | ||
8c069ff4 HB |
892 | static int cpumsf_pmu_event_init(struct perf_event *event) |
893 | { | |
894 | int err; | |
895 | ||
55baa2f8 | 896 | /* No support for taken branch sampling */ |
5aa98879 TR |
897 | /* No support for callchain, stacks and registers */ |
898 | if (has_branch_stack(event) || is_callchain_event(event)) | |
55baa2f8 HB |
899 | return -EOPNOTSUPP; |
900 | ||
901 | switch (event->attr.type) { | |
902 | case PERF_TYPE_RAW: | |
7e75fc3f HB |
903 | if ((event->attr.config != PERF_EVENT_CPUM_SF) && |
904 | (event->attr.config != PERF_EVENT_CPUM_SF_DIAG)) | |
55baa2f8 HB |
905 | return -ENOENT; |
906 | break; | |
907 | case PERF_TYPE_HARDWARE: | |
908 | /* Support sampling of CPU cycles in addition to the | |
909 | * counter facility. However, the counter facility | |
910 | * is more precise and, hence, restrict this PMU to | |
911 | * sampling events only. | |
912 | */ | |
913 | if (event->attr.config != PERF_COUNT_HW_CPU_CYCLES) | |
914 | return -ENOENT; | |
915 | if (!is_sampling_event(event)) | |
916 | return -ENOENT; | |
917 | break; | |
918 | default: | |
8c069ff4 | 919 | return -ENOENT; |
55baa2f8 | 920 | } |
8c069ff4 | 921 | |
dd127b3b | 922 | /* Check online status of the CPU to which the event is pinned */ |
19220999 | 923 | if (event->cpu >= 0 && !cpu_online(event->cpu)) |
e14e59c1 | 924 | return -ENODEV; |
8c069ff4 | 925 | |
dd127b3b HB |
926 | /* Force reset of idle/hv excludes regardless of what the |
927 | * user requested. | |
928 | */ | |
929 | if (event->attr.exclude_hv) | |
930 | event->attr.exclude_hv = 0; | |
931 | if (event->attr.exclude_idle) | |
932 | event->attr.exclude_idle = 0; | |
933 | ||
8c069ff4 HB |
934 | err = __hw_perf_event_init(event); |
935 | if (unlikely(err)) | |
936 | if (event->destroy) | |
937 | event->destroy(event); | |
938 | return err; | |
939 | } | |
940 | ||
941 | static void cpumsf_pmu_enable(struct pmu *pmu) | |
942 | { | |
eb7e7d76 | 943 | struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf); |
69f239ed | 944 | struct hw_perf_event *hwc; |
8c069ff4 HB |
945 | int err; |
946 | ||
947 | if (cpuhw->flags & PMU_F_ENABLED) | |
948 | return; | |
949 | ||
950 | if (cpuhw->flags & PMU_F_ERR_MASK) | |
951 | return; | |
952 | ||
69f239ed HB |
953 | /* Check whether to extent the sampling buffer. |
954 | * | |
955 | * Two conditions trigger an increase of the sampling buffer for a | |
956 | * perf event: | |
957 | * 1. Postponed buffer allocations from the event initialization. | |
958 | * 2. Sampling overflows that contribute to pending allocations. | |
959 | * | |
960 | * Note that the extend_sampling_buffer() function disables the sampling | |
961 | * facility, but it can be fully re-enabled using sampling controls that | |
962 | * have been saved in cpumsf_pmu_disable(). | |
963 | */ | |
964 | if (cpuhw->event) { | |
965 | hwc = &cpuhw->event->hw; | |
cbf6948f PH |
966 | if (!(SAMPL_DIAG_MODE(hwc))) { |
967 | /* | |
968 | * Account number of overflow-designated | |
969 | * buffer extents | |
970 | */ | |
971 | sfb_account_overflows(cpuhw, hwc); | |
c17a7c6e | 972 | extend_sampling_buffer(&cpuhw->sfb, hwc); |
cbf6948f | 973 | } |
2cb549a8 TR |
974 | /* Rate may be adjusted with ioctl() */ |
975 | cpuhw->lsctl.interval = SAMPL_RATE(&cpuhw->event->hw); | |
69f239ed HB |
976 | } |
977 | ||
978 | /* (Re)enable the PMU and sampling facility */ | |
8c069ff4 HB |
979 | cpuhw->flags |= PMU_F_ENABLED; |
980 | barrier(); | |
981 | ||
982 | err = lsctl(&cpuhw->lsctl); | |
983 | if (err) { | |
984 | cpuhw->flags &= ~PMU_F_ENABLED; | |
d98b5d07 | 985 | pr_err("Loading sampling controls failed: op %i err %i\n", |
8c069ff4 HB |
986 | 1, err); |
987 | return; | |
988 | } | |
989 | ||
d4c7e649 HB |
990 | /* Load current program parameter */ |
991 | lpp(&S390_lowcore.lpp); | |
992 | ||
532da3de TR |
993 | debug_sprintf_event(sfdbg, 6, "%s: es %i cs %i ed %i cd %i " |
994 | "interval %#lx tear %#lx dear %#lx\n", __func__, | |
2cb549a8 TR |
995 | cpuhw->lsctl.es, cpuhw->lsctl.cs, cpuhw->lsctl.ed, |
996 | cpuhw->lsctl.cd, cpuhw->lsctl.interval, | |
532da3de | 997 | cpuhw->lsctl.tear, cpuhw->lsctl.dear); |
8c069ff4 HB |
998 | } |
999 | ||
1000 | static void cpumsf_pmu_disable(struct pmu *pmu) | |
1001 | { | |
eb7e7d76 | 1002 | struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf); |
8c069ff4 HB |
1003 | struct hws_lsctl_request_block inactive; |
1004 | struct hws_qsi_info_block si; | |
1005 | int err; | |
1006 | ||
1007 | if (!(cpuhw->flags & PMU_F_ENABLED)) | |
1008 | return; | |
1009 | ||
1010 | if (cpuhw->flags & PMU_F_ERR_MASK) | |
1011 | return; | |
1012 | ||
1013 | /* Switch off sampling activation control */ | |
1014 | inactive = cpuhw->lsctl; | |
1015 | inactive.cs = 0; | |
7e75fc3f | 1016 | inactive.cd = 0; |
8c069ff4 HB |
1017 | |
1018 | err = lsctl(&inactive); | |
1019 | if (err) { | |
d98b5d07 | 1020 | pr_err("Loading sampling controls failed: op %i err %i\n", |
8c069ff4 HB |
1021 | 2, err); |
1022 | return; | |
1023 | } | |
1024 | ||
1025 | /* Save state of TEAR and DEAR register contents */ | |
72fbcd05 TR |
1026 | err = qsi(&si); |
1027 | if (!err) { | |
8c069ff4 HB |
1028 | /* TEAR/DEAR values are valid only if the sampling facility is |
1029 | * enabled. Note that cpumsf_pmu_disable() might be called even | |
1030 | * for a disabled sampling facility because cpumsf_pmu_enable() | |
1031 | * controls the enable/disable state. | |
1032 | */ | |
1033 | if (si.es) { | |
1034 | cpuhw->lsctl.tear = si.tear; | |
1035 | cpuhw->lsctl.dear = si.dear; | |
1036 | } | |
1037 | } else | |
532da3de TR |
1038 | debug_sprintf_event(sfdbg, 3, "%s: qsi() failed with err %i\n", |
1039 | __func__, err); | |
8c069ff4 HB |
1040 | |
1041 | cpuhw->flags &= ~PMU_F_ENABLED; | |
1042 | } | |
1043 | ||
dd127b3b HB |
1044 | /* perf_exclude_event() - Filter event |
1045 | * @event: The perf event | |
1046 | * @regs: pt_regs structure | |
1047 | * @sde_regs: Sample-data-entry (sde) regs structure | |
1048 | * | |
1049 | * Filter perf events according to their exclude specification. | |
1050 | * | |
1051 | * Return non-zero if the event shall be excluded. | |
1052 | */ | |
1053 | static int perf_exclude_event(struct perf_event *event, struct pt_regs *regs, | |
1054 | struct perf_sf_sde_regs *sde_regs) | |
1055 | { | |
1056 | if (event->attr.exclude_user && user_mode(regs)) | |
1057 | return 1; | |
1058 | if (event->attr.exclude_kernel && !user_mode(regs)) | |
1059 | return 1; | |
1060 | if (event->attr.exclude_guest && sde_regs->in_guest) | |
1061 | return 1; | |
1062 | if (event->attr.exclude_host && !sde_regs->in_guest) | |
1063 | return 1; | |
1064 | return 0; | |
1065 | } | |
1066 | ||
8c069ff4 HB |
1067 | /* perf_push_sample() - Push samples to perf |
1068 | * @event: The perf event | |
1069 | * @sample: Hardware sample data | |
1070 | * | |
1071 | * Use the hardware sample data to create perf event sample. The sample | |
1072 | * is the pushed to the event subsystem and the function checks for | |
1073 | * possible event overflows. If an event overflow occurs, the PMU is | |
1074 | * stopped. | |
1075 | * | |
1076 | * Return non-zero if an event overflow occurred. | |
1077 | */ | |
3d43b981 PH |
1078 | static int perf_push_sample(struct perf_event *event, |
1079 | struct hws_basic_entry *basic) | |
8c069ff4 HB |
1080 | { |
1081 | int overflow; | |
1082 | struct pt_regs regs; | |
443e802b | 1083 | struct perf_sf_sde_regs *sde_regs; |
8c069ff4 HB |
1084 | struct perf_sample_data data; |
1085 | ||
7e75fc3f | 1086 | /* Setup perf sample */ |
8c069ff4 HB |
1087 | perf_sample_data_init(&data, 0, event->hw.last_period); |
1088 | ||
443e802b HB |
1089 | /* Setup pt_regs to look like an CPU-measurement external interrupt |
1090 | * using the Program Request Alert code. The regs.int_parm_long | |
1091 | * field which is unused contains additional sample-data-entry related | |
1092 | * indicators. | |
1093 | */ | |
8c069ff4 | 1094 | memset(®s, 0, sizeof(regs)); |
443e802b HB |
1095 | regs.int_code = 0x1407; |
1096 | regs.int_parm = CPU_MF_INT_SF_PRA; | |
1097 | sde_regs = (struct perf_sf_sde_regs *) ®s.int_parm_long; | |
1098 | ||
3d43b981 PH |
1099 | psw_bits(regs.psw).ia = basic->ia; |
1100 | psw_bits(regs.psw).dat = basic->T; | |
1101 | psw_bits(regs.psw).wait = basic->W; | |
1102 | psw_bits(regs.psw).pstate = basic->P; | |
1103 | psw_bits(regs.psw).as = basic->AS; | |
8c069ff4 | 1104 | |
e22cf8ca | 1105 | /* |
c19805f8 CB |
1106 | * Use the hardware provided configuration level to decide if the |
1107 | * sample belongs to a guest or host. If that is not available, | |
1108 | * fall back to the following heuristics: | |
1109 | * A non-zero guest program parameter always indicates a guest | |
1110 | * sample. Some early samples or samples from guests without | |
b1685ab9 | 1111 | * lpp usage would be misaccounted to the host. We use the asn |
c19805f8 | 1112 | * value as an addon heuristic to detect most of these guest samples. |
df26c2e8 MS |
1113 | * If the value differs from 0xffff (the host value), we assume to |
1114 | * be a KVM guest. | |
443e802b | 1115 | */ |
3d43b981 | 1116 | switch (basic->CL) { |
c19805f8 CB |
1117 | case 1: /* logical partition */ |
1118 | sde_regs->in_guest = 0; | |
1119 | break; | |
1120 | case 2: /* virtual machine */ | |
443e802b | 1121 | sde_regs->in_guest = 1; |
c19805f8 CB |
1122 | break; |
1123 | default: /* old machine, use heuristics */ | |
3d43b981 | 1124 | if (basic->gpp || basic->prim_asn != 0xffff) |
c19805f8 CB |
1125 | sde_regs->in_guest = 1; |
1126 | break; | |
1127 | } | |
443e802b | 1128 | |
544e8dd7 HB |
1129 | /* |
1130 | * Store the PID value from the sample-data-entry to be | |
1131 | * processed and resolved by cpumsf_output_event_pid(). | |
1132 | */ | |
1133 | data.tid_entry.pid = basic->hpp & LPP_PID_MASK; | |
1134 | ||
8c069ff4 | 1135 | overflow = 0; |
dd127b3b HB |
1136 | if (perf_exclude_event(event, ®s, sde_regs)) |
1137 | goto out; | |
8c069ff4 HB |
1138 | if (perf_event_overflow(event, &data, ®s)) { |
1139 | overflow = 1; | |
1140 | event->pmu->stop(event, 0); | |
8c069ff4 HB |
1141 | } |
1142 | perf_event_update_userpage(event); | |
dd127b3b | 1143 | out: |
8c069ff4 HB |
1144 | return overflow; |
1145 | } | |
1146 | ||
1147 | static void perf_event_count_update(struct perf_event *event, u64 count) | |
1148 | { | |
1149 | local64_add(count, &event->count); | |
1150 | } | |
1151 | ||
1152 | /* hw_collect_samples() - Walk through a sample-data-block and collect samples | |
1153 | * @event: The perf event | |
1154 | * @sdbt: Sample-data-block table | |
1155 | * @overflow: Event overflow counter | |
1156 | * | |
7e75fc3f HB |
1157 | * Walks through a sample-data-block and collects sampling data entries that are |
1158 | * then pushed to the perf event subsystem. Depending on the sampling function, | |
1159 | * there can be either basic-sampling or combined-sampling data entries. A | |
1160 | * combined-sampling data entry consists of a basic- and a diagnostic-sampling | |
1161 | * data entry. The sampling function is determined by the flags in the perf | |
1162 | * event hardware structure. The function always works with a combined-sampling | |
1163 | * data entry but ignores the the diagnostic portion if it is not available. | |
1164 | * | |
1165 | * Note that the implementation focuses on basic-sampling data entries and, if | |
1166 | * such an entry is not valid, the entire combined-sampling data entry is | |
1167 | * ignored. | |
1168 | * | |
1169 | * The overflow variables counts the number of samples that has been discarded | |
1170 | * due to a perf event overflow. | |
8c069ff4 HB |
1171 | */ |
1172 | static void hw_collect_samples(struct perf_event *event, unsigned long *sdbt, | |
1173 | unsigned long long *overflow) | |
1174 | { | |
7e75fc3f | 1175 | struct hws_trailer_entry *te; |
3d43b981 | 1176 | struct hws_basic_entry *sample; |
8c069ff4 | 1177 | |
7e75fc3f | 1178 | te = (struct hws_trailer_entry *) trailer_entry_ptr(*sdbt); |
3d43b981 | 1179 | sample = (struct hws_basic_entry *) *sdbt; |
7e75fc3f | 1180 | while ((unsigned long *) sample < (unsigned long *) te) { |
8c069ff4 | 1181 | /* Check for an empty sample */ |
745f5d20 | 1182 | if (!sample->def || sample->LS) |
8c069ff4 HB |
1183 | break; |
1184 | ||
1185 | /* Update perf event period */ | |
1186 | perf_event_count_update(event, SAMPL_RATE(&event->hw)); | |
1187 | ||
3d43b981 PH |
1188 | /* Check whether sample is valid */ |
1189 | if (sample->def == 0x0001) { | |
8c069ff4 HB |
1190 | /* If an event overflow occurred, the PMU is stopped to |
1191 | * throttle event delivery. Remaining sample data is | |
1192 | * discarded. | |
1193 | */ | |
7e75fc3f | 1194 | if (!*overflow) { |
3d43b981 PH |
1195 | /* Check whether sample is consistent */ |
1196 | if (sample->I == 0 && sample->W == 0) { | |
7e75fc3f | 1197 | /* Deliver sample data to perf */ |
3d43b981 PH |
1198 | *overflow = perf_push_sample(event, |
1199 | sample); | |
7e75fc3f HB |
1200 | } |
1201 | } else | |
8c069ff4 HB |
1202 | /* Count discarded samples */ |
1203 | *overflow += 1; | |
7e75fc3f | 1204 | } else { |
d98b5d07 | 1205 | debug_sprintf_event(sfdbg, 4, |
c1838834 | 1206 | "%s: Found unknown" |
d98b5d07 | 1207 | " sampling data entry: te->f %i" |
c1838834 | 1208 | " basic.def %#4x (%p)\n", __func__, |
d98b5d07 | 1209 | te->f, sample->def, sample); |
7e75fc3f HB |
1210 | /* Sample slot is not yet written or other record. |
1211 | * | |
1212 | * This condition can occur if the buffer was reused | |
1213 | * from a combined basic- and diagnostic-sampling. | |
1214 | * If only basic-sampling is then active, entries are | |
1215 | * written into the larger diagnostic entries. | |
1216 | * This is typically the case for sample-data-blocks | |
1217 | * that are not full. Stop processing if the first | |
1218 | * invalid format was detected. | |
1219 | */ | |
1220 | if (!te->f) | |
1221 | break; | |
1222 | } | |
8c069ff4 HB |
1223 | |
1224 | /* Reset sample slot and advance to next sample */ | |
3d43b981 PH |
1225 | sample->def = 0; |
1226 | sample++; | |
8c069ff4 HB |
1227 | } |
1228 | } | |
1229 | ||
1230 | /* hw_perf_event_update() - Process sampling buffer | |
1231 | * @event: The perf event | |
1232 | * @flush_all: Flag to also flush partially filled sample-data-blocks | |
1233 | * | |
1234 | * Processes the sampling buffer and create perf event samples. | |
1235 | * The sampling buffer position are retrieved and saved in the TEAR_REG | |
1236 | * register of the specified perf event. | |
1237 | * | |
1238 | * Only full sample-data-blocks are processed. Specify the flash_all flag | |
d7528862 HB |
1239 | * to also walk through partially filled sample-data-blocks. It is ignored |
1240 | * if PERF_CPUM_SF_FULL_BLOCKS is set. The PERF_CPUM_SF_FULL_BLOCKS flag | |
1241 | * enforces the processing of full sample-data-blocks only (trailer entries | |
1242 | * with the block-full-indicator bit set). | |
8c069ff4 HB |
1243 | */ |
1244 | static void hw_perf_event_update(struct perf_event *event, int flush_all) | |
1245 | { | |
1246 | struct hw_perf_event *hwc = &event->hw; | |
1247 | struct hws_trailer_entry *te; | |
1248 | unsigned long *sdbt; | |
fcc77f50 | 1249 | unsigned long long event_overflow, sampl_overflow, num_sdb, te_flags; |
8c069ff4 HB |
1250 | int done; |
1251 | ||
cbf6948f PH |
1252 | /* |
1253 | * AUX buffer is used when in diagnostic sampling mode. | |
1254 | * No perf events/samples are created. | |
1255 | */ | |
1256 | if (SAMPL_DIAG_MODE(&event->hw)) | |
1257 | return; | |
1258 | ||
d7528862 HB |
1259 | if (flush_all && SDB_FULL_BLOCKS(hwc)) |
1260 | flush_all = 0; | |
1261 | ||
8c069ff4 | 1262 | sdbt = (unsigned long *) TEAR_REG(hwc); |
69f239ed | 1263 | done = event_overflow = sampl_overflow = num_sdb = 0; |
8c069ff4 HB |
1264 | while (!done) { |
1265 | /* Get the trailer entry of the sample-data-block */ | |
1266 | te = (struct hws_trailer_entry *) trailer_entry_ptr(*sdbt); | |
1267 | ||
1268 | /* Leave loop if no more work to do (block full indicator) */ | |
1269 | if (!te->f) { | |
1270 | done = 1; | |
1271 | if (!flush_all) | |
1272 | break; | |
1273 | } | |
1274 | ||
69f239ed HB |
1275 | /* Check the sample overflow count */ |
1276 | if (te->overflow) | |
1277 | /* Account sample overflows and, if a particular limit | |
1278 | * is reached, extend the sampling buffer. | |
1279 | * For details, see sfb_account_overflows(). | |
8c069ff4 | 1280 | */ |
69f239ed | 1281 | sampl_overflow += te->overflow; |
8c069ff4 HB |
1282 | |
1283 | /* Timestamps are valid for full sample-data-blocks only */ | |
532da3de | 1284 | debug_sprintf_event(sfdbg, 6, "%s: sdbt %#lx " |
d98b5d07 | 1285 | "overflow %llu timestamp %#llx\n", |
532da3de | 1286 | __func__, (unsigned long)sdbt, te->overflow, |
443d4beb | 1287 | (te->f) ? trailer_timestamp(te) : 0ULL); |
8c069ff4 HB |
1288 | |
1289 | /* Collect all samples from a single sample-data-block and | |
1290 | * flag if an (perf) event overflow happened. If so, the PMU | |
1291 | * is stopped and remaining samples will be discarded. | |
1292 | */ | |
1293 | hw_collect_samples(event, sdbt, &event_overflow); | |
69f239ed | 1294 | num_sdb++; |
8c069ff4 | 1295 | |
fcc77f50 HB |
1296 | /* Reset trailer (using compare-double-and-swap) */ |
1297 | do { | |
1298 | te_flags = te->flags & ~SDB_TE_BUFFER_FULL_MASK; | |
1299 | te_flags |= SDB_TE_ALERT_REQ_MASK; | |
1300 | } while (!cmpxchg_double(&te->flags, &te->overflow, | |
1301 | te->flags, te->overflow, | |
1302 | te_flags, 0ULL)); | |
8c069ff4 HB |
1303 | |
1304 | /* Advance to next sample-data-block */ | |
1305 | sdbt++; | |
1306 | if (is_link_entry(sdbt)) | |
1307 | sdbt = get_next_sdbt(sdbt); | |
1308 | ||
1309 | /* Update event hardware registers */ | |
1310 | TEAR_REG(hwc) = (unsigned long) sdbt; | |
1311 | ||
1312 | /* Stop processing sample-data if all samples of the current | |
1313 | * sample-data-block were flushed even if it was not full. | |
1314 | */ | |
1315 | if (flush_all && done) | |
1316 | break; | |
8c069ff4 HB |
1317 | } |
1318 | ||
69f239ed HB |
1319 | /* Account sample overflows in the event hardware structure */ |
1320 | if (sampl_overflow) | |
1321 | OVERFLOW_REG(hwc) = DIV_ROUND_UP(OVERFLOW_REG(hwc) + | |
1322 | sampl_overflow, 1 + num_sdb); | |
39d4a501 TR |
1323 | |
1324 | /* Perf_event_overflow() and perf_event_account_interrupt() limit | |
1325 | * the interrupt rate to an upper limit. Roughly 1000 samples per | |
1326 | * task tick. | |
1327 | * Hitting this limit results in a large number | |
1328 | * of throttled REF_REPORT_THROTTLE entries and the samples | |
1329 | * are dropped. | |
1330 | * Slightly increase the interval to avoid hitting this limit. | |
1331 | */ | |
1332 | if (event_overflow) { | |
1333 | SAMPL_RATE(hwc) += DIV_ROUND_UP(SAMPL_RATE(hwc), 10); | |
1334 | debug_sprintf_event(sfdbg, 1, "%s: rate adjustment %ld\n", | |
1335 | __func__, | |
1336 | DIV_ROUND_UP(SAMPL_RATE(hwc), 10)); | |
1337 | } | |
1338 | ||
8c069ff4 | 1339 | if (sampl_overflow || event_overflow) |
c1838834 | 1340 | debug_sprintf_event(sfdbg, 4, "%s: " |
532da3de TR |
1341 | "overflows: sample %llu event %llu" |
1342 | " total %llu num_sdb %llu\n", | |
1343 | __func__, sampl_overflow, event_overflow, | |
1344 | OVERFLOW_REG(hwc), num_sdb); | |
8c069ff4 HB |
1345 | } |
1346 | ||
ca5955cd PH |
1347 | #define AUX_SDB_INDEX(aux, i) ((i) % aux->sfb.num_sdb) |
1348 | #define AUX_SDB_NUM(aux, start, end) (end >= start ? end - start + 1 : 0) | |
1349 | #define AUX_SDB_NUM_ALERT(aux) AUX_SDB_NUM(aux, aux->head, aux->alert_mark) | |
1350 | #define AUX_SDB_NUM_EMPTY(aux) AUX_SDB_NUM(aux, aux->head, aux->empty_mark) | |
1351 | ||
1352 | /* | |
1353 | * Get trailer entry by index of SDB. | |
1354 | */ | |
1355 | static struct hws_trailer_entry *aux_sdb_trailer(struct aux_buffer *aux, | |
1356 | unsigned long index) | |
1357 | { | |
1358 | unsigned long sdb; | |
1359 | ||
1360 | index = AUX_SDB_INDEX(aux, index); | |
1361 | sdb = aux->sdb_index[index]; | |
1362 | return (struct hws_trailer_entry *)trailer_entry_ptr(sdb); | |
1363 | } | |
1364 | ||
1365 | /* | |
1366 | * Finish sampling on the cpu. Called by cpumsf_pmu_del() with pmu | |
1367 | * disabled. Collect the full SDBs in AUX buffer which have not reached | |
1368 | * the point of alert indicator. And ignore the SDBs which are not | |
1369 | * full. | |
1370 | * | |
1371 | * 1. Scan SDBs to see how much data is there and consume them. | |
1372 | * 2. Remove alert indicator in the buffer. | |
1373 | */ | |
1374 | static void aux_output_end(struct perf_output_handle *handle) | |
1375 | { | |
1376 | unsigned long i, range_scan, idx; | |
1377 | struct aux_buffer *aux; | |
1378 | struct hws_trailer_entry *te; | |
1379 | ||
1380 | aux = perf_get_aux(handle); | |
1381 | if (!aux) | |
1382 | return; | |
1383 | ||
1384 | range_scan = AUX_SDB_NUM_ALERT(aux); | |
1385 | for (i = 0, idx = aux->head; i < range_scan; i++, idx++) { | |
1386 | te = aux_sdb_trailer(aux, idx); | |
1387 | if (!(te->flags & SDB_TE_BUFFER_FULL_MASK)) | |
1388 | break; | |
1389 | } | |
1390 | /* i is num of SDBs which are full */ | |
1391 | perf_aux_output_end(handle, i << PAGE_SHIFT); | |
1392 | ||
1393 | /* Remove alert indicators in the buffer */ | |
1394 | te = aux_sdb_trailer(aux, aux->alert_mark); | |
1395 | te->flags &= ~SDB_TE_ALERT_REQ_MASK; | |
1396 | ||
ee5c4ccf TR |
1397 | debug_sprintf_event(sfdbg, 6, "%s: SDBs %ld range %ld head %ld\n", |
1398 | __func__, i, range_scan, aux->head); | |
ca5955cd PH |
1399 | } |
1400 | ||
1401 | /* | |
1402 | * Start sampling on the CPU. Called by cpumsf_pmu_add() when an event | |
1403 | * is first added to the CPU or rescheduled again to the CPU. It is called | |
1404 | * with pmu disabled. | |
1405 | * | |
1406 | * 1. Reset the trailer of SDBs to get ready for new data. | |
1407 | * 2. Tell the hardware where to put the data by reset the SDBs buffer | |
1408 | * head(tear/dear). | |
1409 | */ | |
1410 | static int aux_output_begin(struct perf_output_handle *handle, | |
1411 | struct aux_buffer *aux, | |
1412 | struct cpu_hw_sf *cpuhw) | |
1413 | { | |
1414 | unsigned long range; | |
1415 | unsigned long i, range_scan, idx; | |
1416 | unsigned long head, base, offset; | |
1417 | struct hws_trailer_entry *te; | |
1418 | ||
1419 | if (WARN_ON_ONCE(handle->head & ~PAGE_MASK)) | |
1420 | return -EINVAL; | |
1421 | ||
1422 | aux->head = handle->head >> PAGE_SHIFT; | |
1423 | range = (handle->size + 1) >> PAGE_SHIFT; | |
1424 | if (range <= 1) | |
1425 | return -ENOMEM; | |
1426 | ||
1427 | /* | |
1428 | * SDBs between aux->head and aux->empty_mark are already ready | |
1429 | * for new data. range_scan is num of SDBs not within them. | |
1430 | */ | |
ee5c4ccf TR |
1431 | debug_sprintf_event(sfdbg, 6, |
1432 | "%s: range %ld head %ld alert %ld empty %ld\n", | |
1433 | __func__, range, aux->head, aux->alert_mark, | |
1434 | aux->empty_mark); | |
ca5955cd PH |
1435 | if (range > AUX_SDB_NUM_EMPTY(aux)) { |
1436 | range_scan = range - AUX_SDB_NUM_EMPTY(aux); | |
1437 | idx = aux->empty_mark + 1; | |
1438 | for (i = 0; i < range_scan; i++, idx++) { | |
1439 | te = aux_sdb_trailer(aux, idx); | |
32dab682 TR |
1440 | te->flags &= ~(SDB_TE_BUFFER_FULL_MASK | |
1441 | SDB_TE_ALERT_REQ_MASK); | |
ca5955cd PH |
1442 | te->overflow = 0; |
1443 | } | |
1444 | /* Save the position of empty SDBs */ | |
1445 | aux->empty_mark = aux->head + range - 1; | |
1446 | } | |
1447 | ||
1448 | /* Set alert indicator */ | |
1449 | aux->alert_mark = aux->head + range/2 - 1; | |
1450 | te = aux_sdb_trailer(aux, aux->alert_mark); | |
1451 | te->flags = te->flags | SDB_TE_ALERT_REQ_MASK; | |
1452 | ||
1453 | /* Reset hardware buffer head */ | |
1454 | head = AUX_SDB_INDEX(aux, aux->head); | |
1455 | base = aux->sdbt_index[head / CPUM_SF_SDB_PER_TABLE]; | |
1456 | offset = head % CPUM_SF_SDB_PER_TABLE; | |
1457 | cpuhw->lsctl.tear = base + offset * sizeof(unsigned long); | |
1458 | cpuhw->lsctl.dear = aux->sdb_index[head]; | |
1459 | ||
ee5c4ccf TR |
1460 | debug_sprintf_event(sfdbg, 6, "%s: head %ld alert %ld empty %ld " |
1461 | "index %ld tear %#lx dear %#lx\n", __func__, | |
ca5955cd | 1462 | aux->head, aux->alert_mark, aux->empty_mark, |
ca5955cd | 1463 | head / CPUM_SF_SDB_PER_TABLE, |
ee5c4ccf | 1464 | cpuhw->lsctl.tear, cpuhw->lsctl.dear); |
ca5955cd PH |
1465 | |
1466 | return 0; | |
1467 | } | |
1468 | ||
1469 | /* | |
1470 | * Set alert indicator on SDB at index @alert_index while sampler is running. | |
1471 | * | |
1472 | * Return true if successfully. | |
1473 | * Return false if full indicator is already set by hardware sampler. | |
1474 | */ | |
1475 | static bool aux_set_alert(struct aux_buffer *aux, unsigned long alert_index, | |
1476 | unsigned long long *overflow) | |
1477 | { | |
1478 | unsigned long long orig_overflow, orig_flags, new_flags; | |
1479 | struct hws_trailer_entry *te; | |
1480 | ||
1481 | te = aux_sdb_trailer(aux, alert_index); | |
1482 | do { | |
1483 | orig_flags = te->flags; | |
32dab682 | 1484 | *overflow = orig_overflow = te->overflow; |
ca5955cd PH |
1485 | if (orig_flags & SDB_TE_BUFFER_FULL_MASK) { |
1486 | /* | |
1487 | * SDB is already set by hardware. | |
1488 | * Abort and try to set somewhere | |
1489 | * behind. | |
1490 | */ | |
1491 | return false; | |
1492 | } | |
1493 | new_flags = orig_flags | SDB_TE_ALERT_REQ_MASK; | |
1494 | } while (!cmpxchg_double(&te->flags, &te->overflow, | |
1495 | orig_flags, orig_overflow, | |
1496 | new_flags, 0ULL)); | |
1497 | return true; | |
1498 | } | |
1499 | ||
1500 | /* | |
1501 | * aux_reset_buffer() - Scan and setup SDBs for new samples | |
1502 | * @aux: The AUX buffer to set | |
1503 | * @range: The range of SDBs to scan started from aux->head | |
1504 | * @overflow: Set to overflow count | |
1505 | * | |
1506 | * Set alert indicator on the SDB at index of aux->alert_mark. If this SDB is | |
1507 | * marked as empty, check if it is already set full by the hardware sampler. | |
1508 | * If yes, that means new data is already there before we can set an alert | |
1509 | * indicator. Caller should try to set alert indicator to some position behind. | |
1510 | * | |
1511 | * Scan the SDBs in AUX buffer from behind aux->empty_mark. They are used | |
1512 | * previously and have already been consumed by user space. Reset these SDBs | |
1513 | * (clear full indicator and alert indicator) for new data. | |
1514 | * If aux->alert_mark fall in this area, just set it. Overflow count is | |
1515 | * recorded while scanning. | |
1516 | * | |
1517 | * SDBs between aux->head and aux->empty_mark are already reset at last time. | |
1518 | * and ready for new samples. So scanning on this area could be skipped. | |
1519 | * | |
1520 | * Return true if alert indicator is set successfully and false if not. | |
1521 | */ | |
1522 | static bool aux_reset_buffer(struct aux_buffer *aux, unsigned long range, | |
1523 | unsigned long long *overflow) | |
1524 | { | |
1525 | unsigned long long orig_overflow, orig_flags, new_flags; | |
ee5c4ccf | 1526 | unsigned long i, range_scan, idx, idx_old; |
ca5955cd PH |
1527 | struct hws_trailer_entry *te; |
1528 | ||
ee5c4ccf TR |
1529 | debug_sprintf_event(sfdbg, 6, "%s: range %ld head %ld alert %ld " |
1530 | "empty %ld\n", __func__, range, aux->head, | |
1531 | aux->alert_mark, aux->empty_mark); | |
ca5955cd PH |
1532 | if (range <= AUX_SDB_NUM_EMPTY(aux)) |
1533 | /* | |
1534 | * No need to scan. All SDBs in range are marked as empty. | |
1535 | * Just set alert indicator. Should check race with hardware | |
1536 | * sampler. | |
1537 | */ | |
1538 | return aux_set_alert(aux, aux->alert_mark, overflow); | |
1539 | ||
1540 | if (aux->alert_mark <= aux->empty_mark) | |
1541 | /* | |
1542 | * Set alert indicator on empty SDB. Should check race | |
1543 | * with hardware sampler. | |
1544 | */ | |
1545 | if (!aux_set_alert(aux, aux->alert_mark, overflow)) | |
1546 | return false; | |
1547 | ||
1548 | /* | |
1549 | * Scan the SDBs to clear full and alert indicator used previously. | |
1550 | * Start scanning from one SDB behind empty_mark. If the new alert | |
1551 | * indicator fall into this range, set it. | |
1552 | */ | |
1553 | range_scan = range - AUX_SDB_NUM_EMPTY(aux); | |
ee5c4ccf | 1554 | idx_old = idx = aux->empty_mark + 1; |
ca5955cd PH |
1555 | for (i = 0; i < range_scan; i++, idx++) { |
1556 | te = aux_sdb_trailer(aux, idx); | |
1557 | do { | |
1558 | orig_flags = te->flags; | |
1559 | orig_overflow = te->overflow; | |
1560 | new_flags = orig_flags & ~SDB_TE_BUFFER_FULL_MASK; | |
1561 | if (idx == aux->alert_mark) | |
1562 | new_flags |= SDB_TE_ALERT_REQ_MASK; | |
1563 | else | |
1564 | new_flags &= ~SDB_TE_ALERT_REQ_MASK; | |
1565 | } while (!cmpxchg_double(&te->flags, &te->overflow, | |
1566 | orig_flags, orig_overflow, | |
1567 | new_flags, 0ULL)); | |
1568 | *overflow += orig_overflow; | |
1569 | } | |
1570 | ||
1571 | /* Update empty_mark to new position */ | |
1572 | aux->empty_mark = aux->head + range - 1; | |
1573 | ||
ee5c4ccf TR |
1574 | debug_sprintf_event(sfdbg, 6, "%s: range_scan %ld idx %ld..%ld " |
1575 | "empty %ld\n", __func__, range_scan, idx_old, | |
1576 | idx - 1, aux->empty_mark); | |
ca5955cd PH |
1577 | return true; |
1578 | } | |
1579 | ||
1580 | /* | |
1581 | * Measurement alert handler for diagnostic mode sampling. | |
1582 | */ | |
1583 | static void hw_collect_aux(struct cpu_hw_sf *cpuhw) | |
1584 | { | |
1585 | struct aux_buffer *aux; | |
1586 | int done = 0; | |
1587 | unsigned long range = 0, size; | |
1588 | unsigned long long overflow = 0; | |
1589 | struct perf_output_handle *handle = &cpuhw->handle; | |
4141b6a5 | 1590 | unsigned long num_sdb; |
ca5955cd PH |
1591 | |
1592 | aux = perf_get_aux(handle); | |
1593 | if (WARN_ON_ONCE(!aux)) | |
1594 | return; | |
1595 | ||
1596 | /* Inform user space new data arrived */ | |
1597 | size = AUX_SDB_NUM_ALERT(aux) << PAGE_SHIFT; | |
ee5c4ccf TR |
1598 | debug_sprintf_event(sfdbg, 6, "%s: #alert %ld\n", __func__, |
1599 | size >> PAGE_SHIFT); | |
ca5955cd | 1600 | perf_aux_output_end(handle, size); |
ca5955cd | 1601 | |
4141b6a5 | 1602 | num_sdb = aux->sfb.num_sdb; |
ca5955cd PH |
1603 | while (!done) { |
1604 | /* Get an output handle */ | |
1605 | aux = perf_aux_output_begin(handle, cpuhw->event); | |
1606 | if (handle->size == 0) { | |
1607 | pr_err("The AUX buffer with %lu pages for the " | |
1608 | "diagnostic-sampling mode is full\n", | |
4141b6a5 | 1609 | num_sdb); |
532da3de TR |
1610 | debug_sprintf_event(sfdbg, 1, |
1611 | "%s: AUX buffer used up\n", | |
1612 | __func__); | |
ca5955cd PH |
1613 | break; |
1614 | } | |
1615 | if (WARN_ON_ONCE(!aux)) | |
1616 | return; | |
1617 | ||
1618 | /* Update head and alert_mark to new position */ | |
1619 | aux->head = handle->head >> PAGE_SHIFT; | |
1620 | range = (handle->size + 1) >> PAGE_SHIFT; | |
1621 | if (range == 1) | |
1622 | aux->alert_mark = aux->head; | |
1623 | else | |
1624 | aux->alert_mark = aux->head + range/2 - 1; | |
1625 | ||
1626 | if (aux_reset_buffer(aux, range, &overflow)) { | |
1627 | if (!overflow) { | |
1628 | done = 1; | |
1629 | break; | |
1630 | } | |
1631 | size = range << PAGE_SHIFT; | |
1632 | perf_aux_output_end(&cpuhw->handle, size); | |
1633 | pr_err("Sample data caused the AUX buffer with %lu " | |
ee5c4ccf TR |
1634 | "pages to overflow\n", aux->sfb.num_sdb); |
1635 | debug_sprintf_event(sfdbg, 1, "%s: head %ld range %ld " | |
1636 | "overflow %lld\n", __func__, | |
ca5955cd PH |
1637 | aux->head, range, overflow); |
1638 | } else { | |
1639 | size = AUX_SDB_NUM_ALERT(aux) << PAGE_SHIFT; | |
1640 | perf_aux_output_end(&cpuhw->handle, size); | |
ee5c4ccf | 1641 | debug_sprintf_event(sfdbg, 6, "%s: head %ld alert %ld " |
ca5955cd | 1642 | "already full, try another\n", |
532da3de | 1643 | __func__, |
ca5955cd PH |
1644 | aux->head, aux->alert_mark); |
1645 | } | |
1646 | } | |
1647 | ||
1648 | if (done) | |
ee5c4ccf TR |
1649 | debug_sprintf_event(sfdbg, 6, "%s: head %ld alert %ld " |
1650 | "empty %ld\n", __func__, aux->head, | |
1651 | aux->alert_mark, aux->empty_mark); | |
ca5955cd PH |
1652 | } |
1653 | ||
1654 | /* | |
1655 | * Callback when freeing AUX buffers. | |
1656 | */ | |
1657 | static void aux_buffer_free(void *data) | |
1658 | { | |
1659 | struct aux_buffer *aux = data; | |
1660 | unsigned long i, num_sdbt; | |
1661 | ||
1662 | if (!aux) | |
1663 | return; | |
1664 | ||
1665 | /* Free SDBT. SDB is freed by the caller */ | |
1666 | num_sdbt = aux->sfb.num_sdbt; | |
1667 | for (i = 0; i < num_sdbt; i++) | |
1668 | free_page(aux->sdbt_index[i]); | |
1669 | ||
1670 | kfree(aux->sdbt_index); | |
1671 | kfree(aux->sdb_index); | |
1672 | kfree(aux); | |
1673 | ||
ee5c4ccf | 1674 | debug_sprintf_event(sfdbg, 4, "%s: SDBTs %lu\n", __func__, num_sdbt); |
ca5955cd PH |
1675 | } |
1676 | ||
5223c671 HB |
1677 | static void aux_sdb_init(unsigned long sdb) |
1678 | { | |
1679 | struct hws_trailer_entry *te; | |
1680 | ||
1681 | te = (struct hws_trailer_entry *)trailer_entry_ptr(sdb); | |
1682 | ||
1683 | /* Save clock base */ | |
1684 | te->clock_base = 1; | |
f8d8977a | 1685 | te->progusage2 = tod_clock_base.tod; |
5223c671 HB |
1686 | } |
1687 | ||
ca5955cd PH |
1688 | /* |
1689 | * aux_buffer_setup() - Setup AUX buffer for diagnostic mode sampling | |
84001866 | 1690 | * @event: Event the buffer is setup for, event->cpu == -1 means current |
ca5955cd PH |
1691 | * @pages: Array of pointers to buffer pages passed from perf core |
1692 | * @nr_pages: Total pages | |
1693 | * @snapshot: Flag for snapshot mode | |
1694 | * | |
1695 | * This is the callback when setup an event using AUX buffer. Perf tool can | |
1696 | * trigger this by an additional mmap() call on the event. Unlike the buffer | |
1697 | * for basic samples, AUX buffer belongs to the event. It is scheduled with | |
1698 | * the task among online cpus when it is a per-thread event. | |
1699 | * | |
1700 | * Return the private AUX buffer structure if success or NULL if fails. | |
1701 | */ | |
84001866 MP |
1702 | static void *aux_buffer_setup(struct perf_event *event, void **pages, |
1703 | int nr_pages, bool snapshot) | |
ca5955cd PH |
1704 | { |
1705 | struct sf_buffer *sfb; | |
1706 | struct aux_buffer *aux; | |
1707 | unsigned long *new, *tail; | |
1708 | int i, n_sdbt; | |
1709 | ||
1710 | if (!nr_pages || !pages) | |
1711 | return NULL; | |
1712 | ||
1713 | if (nr_pages > CPUM_SF_MAX_SDB * CPUM_SF_SDB_DIAG_FACTOR) { | |
1714 | pr_err("AUX buffer size (%i pages) is larger than the " | |
1715 | "maximum sampling buffer limit\n", | |
1716 | nr_pages); | |
1717 | return NULL; | |
1718 | } else if (nr_pages < CPUM_SF_MIN_SDB * CPUM_SF_SDB_DIAG_FACTOR) { | |
1719 | pr_err("AUX buffer size (%i pages) is less than the " | |
1720 | "minimum sampling buffer limit\n", | |
1721 | nr_pages); | |
1722 | return NULL; | |
1723 | } | |
1724 | ||
1725 | /* Allocate aux_buffer struct for the event */ | |
32dab682 | 1726 | aux = kzalloc(sizeof(struct aux_buffer), GFP_KERNEL); |
ca5955cd PH |
1727 | if (!aux) |
1728 | goto no_aux; | |
1729 | sfb = &aux->sfb; | |
1730 | ||
1731 | /* Allocate sdbt_index for fast reference */ | |
ee09c914 | 1732 | n_sdbt = DIV_ROUND_UP(nr_pages, CPUM_SF_SDB_PER_TABLE); |
ca5955cd PH |
1733 | aux->sdbt_index = kmalloc_array(n_sdbt, sizeof(void *), GFP_KERNEL); |
1734 | if (!aux->sdbt_index) | |
1735 | goto no_sdbt_index; | |
1736 | ||
1737 | /* Allocate sdb_index for fast reference */ | |
1738 | aux->sdb_index = kmalloc_array(nr_pages, sizeof(void *), GFP_KERNEL); | |
1739 | if (!aux->sdb_index) | |
1740 | goto no_sdb_index; | |
1741 | ||
1742 | /* Allocate the first SDBT */ | |
1743 | sfb->num_sdbt = 0; | |
1744 | sfb->sdbt = (unsigned long *) get_zeroed_page(GFP_KERNEL); | |
1745 | if (!sfb->sdbt) | |
1746 | goto no_sdbt; | |
1747 | aux->sdbt_index[sfb->num_sdbt++] = (unsigned long)sfb->sdbt; | |
1748 | tail = sfb->tail = sfb->sdbt; | |
1749 | ||
1750 | /* | |
1751 | * Link the provided pages of AUX buffer to SDBT. | |
1752 | * Allocate SDBT if needed. | |
1753 | */ | |
1754 | for (i = 0; i < nr_pages; i++, tail++) { | |
1755 | if (require_table_link(tail)) { | |
1756 | new = (unsigned long *) get_zeroed_page(GFP_KERNEL); | |
1757 | if (!new) | |
1758 | goto no_sdbt; | |
1759 | aux->sdbt_index[sfb->num_sdbt++] = (unsigned long)new; | |
1760 | /* Link current page to tail of chain */ | |
1761 | *tail = (unsigned long)(void *) new + 1; | |
1762 | tail = new; | |
1763 | } | |
1764 | /* Tail is the entry in a SDBT */ | |
1765 | *tail = (unsigned long)pages[i]; | |
1766 | aux->sdb_index[i] = (unsigned long)pages[i]; | |
5223c671 | 1767 | aux_sdb_init((unsigned long)pages[i]); |
ca5955cd PH |
1768 | } |
1769 | sfb->num_sdb = nr_pages; | |
1770 | ||
1771 | /* Link the last entry in the SDBT to the first SDBT */ | |
1772 | *tail = (unsigned long) sfb->sdbt + 1; | |
1773 | sfb->tail = tail; | |
1774 | ||
1775 | /* | |
1776 | * Initial all SDBs are zeroed. Mark it as empty. | |
1777 | * So there is no need to clear the full indicator | |
1778 | * when this event is first added. | |
1779 | */ | |
1780 | aux->empty_mark = sfb->num_sdb - 1; | |
1781 | ||
ee5c4ccf TR |
1782 | debug_sprintf_event(sfdbg, 4, "%s: SDBTs %lu SDBs %lu\n", __func__, |
1783 | sfb->num_sdbt, sfb->num_sdb); | |
ca5955cd PH |
1784 | |
1785 | return aux; | |
1786 | ||
1787 | no_sdbt: | |
1788 | /* SDBs (AUX buffer pages) are freed by caller */ | |
1789 | for (i = 0; i < sfb->num_sdbt; i++) | |
1790 | free_page(aux->sdbt_index[i]); | |
1791 | kfree(aux->sdb_index); | |
1792 | no_sdb_index: | |
1793 | kfree(aux->sdbt_index); | |
1794 | no_sdbt_index: | |
1795 | kfree(aux); | |
1796 | no_aux: | |
1797 | return NULL; | |
1798 | } | |
1799 | ||
8c069ff4 HB |
1800 | static void cpumsf_pmu_read(struct perf_event *event) |
1801 | { | |
1802 | /* Nothing to do ... updates are interrupt-driven */ | |
1803 | } | |
1804 | ||
2cb549a8 TR |
1805 | /* Check if the new sampling period/freqeuncy is appropriate. |
1806 | * | |
1807 | * Return non-zero on error and zero on passed checks. | |
1808 | */ | |
1809 | static int cpumsf_pmu_check_period(struct perf_event *event, u64 value) | |
1810 | { | |
1811 | struct hws_qsi_info_block si; | |
1812 | unsigned long rate; | |
1813 | bool do_freq; | |
1814 | ||
1815 | memset(&si, 0, sizeof(si)); | |
1816 | if (event->cpu == -1) { | |
1817 | if (qsi(&si)) | |
1818 | return -ENODEV; | |
1819 | } else { | |
1820 | /* Event is pinned to a particular CPU, retrieve the per-CPU | |
1821 | * sampling structure for accessing the CPU-specific QSI. | |
1822 | */ | |
1823 | struct cpu_hw_sf *cpuhw = &per_cpu(cpu_hw_sf, event->cpu); | |
1824 | ||
1825 | si = cpuhw->qsi; | |
1826 | } | |
1827 | ||
1828 | do_freq = !!SAMPLE_FREQ_MODE(&event->hw); | |
1829 | rate = getrate(do_freq, value, &si); | |
1830 | if (!rate) | |
1831 | return -EINVAL; | |
1832 | ||
1833 | event->attr.sample_period = rate; | |
1834 | SAMPL_RATE(&event->hw) = rate; | |
1835 | hw_init_period(&event->hw, SAMPL_RATE(&event->hw)); | |
532da3de TR |
1836 | debug_sprintf_event(sfdbg, 4, "%s:" |
1837 | " cpu %d value %#llx period %#llx freq %d\n", | |
1838 | __func__, event->cpu, value, | |
2cb549a8 TR |
1839 | event->attr.sample_period, do_freq); |
1840 | return 0; | |
1841 | } | |
1842 | ||
8c069ff4 HB |
1843 | /* Activate sampling control. |
1844 | * Next call of pmu_enable() starts sampling. | |
1845 | */ | |
1846 | static void cpumsf_pmu_start(struct perf_event *event, int flags) | |
1847 | { | |
eb7e7d76 | 1848 | struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf); |
8c069ff4 HB |
1849 | |
1850 | if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED))) | |
1851 | return; | |
1852 | ||
1853 | if (flags & PERF_EF_RELOAD) | |
1854 | WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE)); | |
1855 | ||
1856 | perf_pmu_disable(event->pmu); | |
1857 | event->hw.state = 0; | |
1858 | cpuhw->lsctl.cs = 1; | |
7e75fc3f HB |
1859 | if (SAMPL_DIAG_MODE(&event->hw)) |
1860 | cpuhw->lsctl.cd = 1; | |
8c069ff4 HB |
1861 | perf_pmu_enable(event->pmu); |
1862 | } | |
1863 | ||
1864 | /* Deactivate sampling control. | |
1865 | * Next call of pmu_enable() stops sampling. | |
1866 | */ | |
1867 | static void cpumsf_pmu_stop(struct perf_event *event, int flags) | |
1868 | { | |
eb7e7d76 | 1869 | struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf); |
8c069ff4 HB |
1870 | |
1871 | if (event->hw.state & PERF_HES_STOPPED) | |
1872 | return; | |
1873 | ||
1874 | perf_pmu_disable(event->pmu); | |
1875 | cpuhw->lsctl.cs = 0; | |
7e75fc3f | 1876 | cpuhw->lsctl.cd = 0; |
8c069ff4 HB |
1877 | event->hw.state |= PERF_HES_STOPPED; |
1878 | ||
1879 | if ((flags & PERF_EF_UPDATE) && !(event->hw.state & PERF_HES_UPTODATE)) { | |
1880 | hw_perf_event_update(event, 1); | |
1881 | event->hw.state |= PERF_HES_UPTODATE; | |
1882 | } | |
1883 | perf_pmu_enable(event->pmu); | |
1884 | } | |
1885 | ||
1886 | static int cpumsf_pmu_add(struct perf_event *event, int flags) | |
1887 | { | |
eb7e7d76 | 1888 | struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf); |
cbf6948f | 1889 | struct aux_buffer *aux; |
8c069ff4 HB |
1890 | int err; |
1891 | ||
1892 | if (cpuhw->flags & PMU_F_IN_USE) | |
1893 | return -EAGAIN; | |
1894 | ||
cbf6948f | 1895 | if (!SAMPL_DIAG_MODE(&event->hw) && !cpuhw->sfb.sdbt) |
8c069ff4 HB |
1896 | return -EINVAL; |
1897 | ||
1898 | err = 0; | |
1899 | perf_pmu_disable(event->pmu); | |
1900 | ||
1901 | event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED; | |
1902 | ||
1903 | /* Set up sampling controls. Always program the sampling register | |
1904 | * using the SDB-table start. Reset TEAR_REG event hardware register | |
1905 | * that is used by hw_perf_event_update() to store the sampling buffer | |
1906 | * position after samples have been flushed. | |
1907 | */ | |
1908 | cpuhw->lsctl.s = 0; | |
1909 | cpuhw->lsctl.h = 1; | |
8c069ff4 | 1910 | cpuhw->lsctl.interval = SAMPL_RATE(&event->hw); |
cbf6948f PH |
1911 | if (!SAMPL_DIAG_MODE(&event->hw)) { |
1912 | cpuhw->lsctl.tear = (unsigned long) cpuhw->sfb.sdbt; | |
1913 | cpuhw->lsctl.dear = *(unsigned long *) cpuhw->sfb.sdbt; | |
7dd6b199 | 1914 | TEAR_REG(&event->hw) = (unsigned long) cpuhw->sfb.sdbt; |
cbf6948f | 1915 | } |
8c069ff4 HB |
1916 | |
1917 | /* Ensure sampling functions are in the disabled state. If disabled, | |
1918 | * switch on sampling enable control. */ | |
7e75fc3f | 1919 | if (WARN_ON_ONCE(cpuhw->lsctl.es == 1 || cpuhw->lsctl.ed == 1)) { |
8c069ff4 HB |
1920 | err = -EAGAIN; |
1921 | goto out; | |
1922 | } | |
cbf6948f PH |
1923 | if (SAMPL_DIAG_MODE(&event->hw)) { |
1924 | aux = perf_aux_output_begin(&cpuhw->handle, event); | |
1925 | if (!aux) { | |
1926 | err = -EINVAL; | |
1927 | goto out; | |
1928 | } | |
1929 | err = aux_output_begin(&cpuhw->handle, aux, cpuhw); | |
1930 | if (err) | |
1931 | goto out; | |
7e75fc3f | 1932 | cpuhw->lsctl.ed = 1; |
cbf6948f PH |
1933 | } |
1934 | cpuhw->lsctl.es = 1; | |
8c069ff4 HB |
1935 | |
1936 | /* Set in_use flag and store event */ | |
8c069ff4 HB |
1937 | cpuhw->event = event; |
1938 | cpuhw->flags |= PMU_F_IN_USE; | |
1939 | ||
1940 | if (flags & PERF_EF_START) | |
1941 | cpumsf_pmu_start(event, PERF_EF_RELOAD); | |
1942 | out: | |
1943 | perf_event_update_userpage(event); | |
1944 | perf_pmu_enable(event->pmu); | |
1945 | return err; | |
1946 | } | |
1947 | ||
1948 | static void cpumsf_pmu_del(struct perf_event *event, int flags) | |
1949 | { | |
eb7e7d76 | 1950 | struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf); |
8c069ff4 HB |
1951 | |
1952 | perf_pmu_disable(event->pmu); | |
1953 | cpumsf_pmu_stop(event, PERF_EF_UPDATE); | |
1954 | ||
1955 | cpuhw->lsctl.es = 0; | |
7e75fc3f | 1956 | cpuhw->lsctl.ed = 0; |
8c069ff4 HB |
1957 | cpuhw->flags &= ~PMU_F_IN_USE; |
1958 | cpuhw->event = NULL; | |
1959 | ||
cbf6948f PH |
1960 | if (SAMPL_DIAG_MODE(&event->hw)) |
1961 | aux_output_end(&cpuhw->handle); | |
8c069ff4 HB |
1962 | perf_event_update_userpage(event); |
1963 | perf_pmu_enable(event->pmu); | |
1964 | } | |
1965 | ||
8c069ff4 | 1966 | CPUMF_EVENT_ATTR(SF, SF_CYCLES_BASIC, PERF_EVENT_CPUM_SF); |
7e75fc3f | 1967 | CPUMF_EVENT_ATTR(SF, SF_CYCLES_BASIC_DIAG, PERF_EVENT_CPUM_SF_DIAG); |
8c069ff4 | 1968 | |
c43e1c5a TR |
1969 | /* Attribute list for CPU_SF. |
1970 | * | |
1971 | * The availablitiy depends on the CPU_MF sampling facility authorization | |
1972 | * for basic + diagnositic samples. This is determined at initialization | |
1973 | * time by the sampling facility device driver. | |
1974 | * If the authorization for basic samples is turned off, it should be | |
1975 | * also turned off for diagnostic sampling. | |
1976 | * | |
1977 | * During initialization of the device driver, check the authorization | |
1978 | * level for diagnostic sampling and installs the attribute | |
1979 | * file for diagnostic sampling if necessary. | |
1980 | * | |
1981 | * For now install a placeholder to reference all possible attributes: | |
1982 | * SF_CYCLES_BASIC and SF_CYCLES_BASIC_DIAG. | |
1983 | * Add another entry for the final NULL pointer. | |
1984 | */ | |
1985 | enum { | |
1986 | SF_CYCLES_BASIC_ATTR_IDX = 0, | |
1987 | SF_CYCLES_BASIC_DIAG_ATTR_IDX, | |
1988 | SF_CYCLES_ATTR_MAX | |
1989 | }; | |
1990 | ||
1991 | static struct attribute *cpumsf_pmu_events_attr[SF_CYCLES_ATTR_MAX + 1] = { | |
1992 | [SF_CYCLES_BASIC_ATTR_IDX] = CPUMF_EVENT_PTR(SF, SF_CYCLES_BASIC) | |
8c069ff4 HB |
1993 | }; |
1994 | ||
1995 | PMU_FORMAT_ATTR(event, "config:0-63"); | |
1996 | ||
1997 | static struct attribute *cpumsf_pmu_format_attr[] = { | |
1998 | &format_attr_event.attr, | |
1999 | NULL, | |
2000 | }; | |
2001 | ||
2002 | static struct attribute_group cpumsf_pmu_events_group = { | |
2003 | .name = "events", | |
2004 | .attrs = cpumsf_pmu_events_attr, | |
2005 | }; | |
03e9e42f | 2006 | |
8c069ff4 HB |
2007 | static struct attribute_group cpumsf_pmu_format_group = { |
2008 | .name = "format", | |
2009 | .attrs = cpumsf_pmu_format_attr, | |
2010 | }; | |
03e9e42f | 2011 | |
8c069ff4 HB |
2012 | static const struct attribute_group *cpumsf_pmu_attr_groups[] = { |
2013 | &cpumsf_pmu_events_group, | |
2014 | &cpumsf_pmu_format_group, | |
2015 | NULL, | |
2016 | }; | |
2017 | ||
2018 | static struct pmu cpumf_sampling = { | |
2019 | .pmu_enable = cpumsf_pmu_enable, | |
2020 | .pmu_disable = cpumsf_pmu_disable, | |
2021 | ||
2022 | .event_init = cpumsf_pmu_event_init, | |
2023 | .add = cpumsf_pmu_add, | |
2024 | .del = cpumsf_pmu_del, | |
2025 | ||
2026 | .start = cpumsf_pmu_start, | |
2027 | .stop = cpumsf_pmu_stop, | |
2028 | .read = cpumsf_pmu_read, | |
2029 | ||
8c069ff4 | 2030 | .attr_groups = cpumsf_pmu_attr_groups, |
ca5955cd PH |
2031 | |
2032 | .setup_aux = aux_buffer_setup, | |
2033 | .free_aux = aux_buffer_free, | |
2cb549a8 TR |
2034 | |
2035 | .check_period = cpumsf_pmu_check_period, | |
8c069ff4 HB |
2036 | }; |
2037 | ||
2038 | static void cpumf_measurement_alert(struct ext_code ext_code, | |
2039 | unsigned int alert, unsigned long unused) | |
2040 | { | |
2041 | struct cpu_hw_sf *cpuhw; | |
2042 | ||
2043 | if (!(alert & CPU_MF_INT_SF_MASK)) | |
2044 | return; | |
2045 | inc_irq_stat(IRQEXT_CMS); | |
eb7e7d76 | 2046 | cpuhw = this_cpu_ptr(&cpu_hw_sf); |
8c069ff4 HB |
2047 | |
2048 | /* Measurement alerts are shared and might happen when the PMU | |
2049 | * is not reserved. Ignore these alerts in this case. */ | |
2050 | if (!(cpuhw->flags & PMU_F_RESERVED)) | |
2051 | return; | |
2052 | ||
2053 | /* The processing below must take care of multiple alert events that | |
2054 | * might be indicated concurrently. */ | |
2055 | ||
2056 | /* Program alert request */ | |
2057 | if (alert & CPU_MF_INT_SF_PRA) { | |
2058 | if (cpuhw->flags & PMU_F_IN_USE) | |
cbf6948f PH |
2059 | if (SAMPL_DIAG_MODE(&cpuhw->event->hw)) |
2060 | hw_collect_aux(cpuhw); | |
2061 | else | |
2062 | hw_perf_event_update(cpuhw->event, 0); | |
8c069ff4 HB |
2063 | else |
2064 | WARN_ON_ONCE(!(cpuhw->flags & PMU_F_IN_USE)); | |
2065 | } | |
2066 | ||
2067 | /* Report measurement alerts only for non-PRA codes */ | |
2068 | if (alert != CPU_MF_INT_SF_PRA) | |
532da3de | 2069 | debug_sprintf_event(sfdbg, 6, "%s: alert %#x\n", __func__, |
03e9e42f | 2070 | alert); |
8c069ff4 HB |
2071 | |
2072 | /* Sampling authorization change request */ | |
2073 | if (alert & CPU_MF_INT_SF_SACA) | |
2074 | qsi(&cpuhw->qsi); | |
2075 | ||
2076 | /* Loss of sample data due to high-priority machine activities */ | |
2077 | if (alert & CPU_MF_INT_SF_LSDA) { | |
2078 | pr_err("Sample data was lost\n"); | |
2079 | cpuhw->flags |= PMU_F_ERR_LSDA; | |
2080 | sf_disable(); | |
2081 | } | |
2082 | ||
2083 | /* Invalid sampling buffer entry */ | |
2084 | if (alert & (CPU_MF_INT_SF_IAE|CPU_MF_INT_SF_ISE)) { | |
2085 | pr_err("A sampling buffer entry is incorrect (alert=0x%x)\n", | |
2086 | alert); | |
2087 | cpuhw->flags |= PMU_F_ERR_IBE; | |
2088 | sf_disable(); | |
2089 | } | |
2090 | } | |
03e9e42f | 2091 | |
e3d617fe | 2092 | static int cpusf_pmu_setup(unsigned int cpu, int flags) |
8c069ff4 | 2093 | { |
8c069ff4 HB |
2094 | /* Ignore the notification if no events are scheduled on the PMU. |
2095 | * This might be racy... | |
2096 | */ | |
2097 | if (!atomic_read(&num_events)) | |
e3d617fe | 2098 | return 0; |
8c069ff4 | 2099 | |
e3d617fe SAS |
2100 | local_irq_disable(); |
2101 | setup_pmc_cpu(&flags); | |
2102 | local_irq_enable(); | |
2103 | return 0; | |
2104 | } | |
2105 | ||
2106 | static int s390_pmu_sf_online_cpu(unsigned int cpu) | |
2107 | { | |
2108 | return cpusf_pmu_setup(cpu, PMC_INIT); | |
2109 | } | |
2110 | ||
2111 | static int s390_pmu_sf_offline_cpu(unsigned int cpu) | |
2112 | { | |
2113 | return cpusf_pmu_setup(cpu, PMC_RELEASE); | |
8c069ff4 HB |
2114 | } |
2115 | ||
69f239ed HB |
2116 | static int param_get_sfb_size(char *buffer, const struct kernel_param *kp) |
2117 | { | |
2118 | if (!cpum_sf_avail()) | |
2119 | return -ENODEV; | |
2120 | return sprintf(buffer, "%lu,%lu", CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB); | |
2121 | } | |
2122 | ||
2123 | static int param_set_sfb_size(const char *val, const struct kernel_param *kp) | |
2124 | { | |
2125 | int rc; | |
2126 | unsigned long min, max; | |
2127 | ||
2128 | if (!cpum_sf_avail()) | |
2129 | return -ENODEV; | |
2130 | if (!val || !strlen(val)) | |
2131 | return -EINVAL; | |
2132 | ||
2133 | /* Valid parameter values: "min,max" or "max" */ | |
2134 | min = CPUM_SF_MIN_SDB; | |
2135 | max = CPUM_SF_MAX_SDB; | |
2136 | if (strchr(val, ',')) | |
2137 | rc = (sscanf(val, "%lu,%lu", &min, &max) == 2) ? 0 : -EINVAL; | |
2138 | else | |
2139 | rc = kstrtoul(val, 10, &max); | |
2140 | ||
2141 | if (min < 2 || min >= max || max > get_num_physpages()) | |
2142 | rc = -EINVAL; | |
2143 | if (rc) | |
2144 | return rc; | |
2145 | ||
2146 | sfb_set_limits(min, max); | |
7e75fc3f | 2147 | pr_info("The sampling buffer limits have changed to: " |
d98b5d07 | 2148 | "min %lu max %lu (diag %lu)\n", |
7e75fc3f | 2149 | CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB, CPUM_SF_SDB_DIAG_FACTOR); |
69f239ed HB |
2150 | return 0; |
2151 | } | |
2152 | ||
2153 | #define param_check_sfb_size(name, p) __param_check(name, p, void) | |
9c27847d | 2154 | static const struct kernel_param_ops param_ops_sfb_size = { |
69f239ed HB |
2155 | .set = param_set_sfb_size, |
2156 | .get = param_get_sfb_size, | |
2157 | }; | |
2158 | ||
7e75fc3f HB |
2159 | #define RS_INIT_FAILURE_QSI 0x0001 |
2160 | #define RS_INIT_FAILURE_BSDES 0x0002 | |
2161 | #define RS_INIT_FAILURE_ALRT 0x0003 | |
2162 | #define RS_INIT_FAILURE_PERF 0x0004 | |
2163 | static void __init pr_cpumsf_err(unsigned int reason) | |
2164 | { | |
2165 | pr_err("Sampling facility support for perf is not available: " | |
d98b5d07 | 2166 | "reason %#x\n", reason); |
7e75fc3f HB |
2167 | } |
2168 | ||
8c069ff4 HB |
2169 | static int __init init_cpum_sampling_pmu(void) |
2170 | { | |
7e75fc3f | 2171 | struct hws_qsi_info_block si; |
8c069ff4 HB |
2172 | int err; |
2173 | ||
2174 | if (!cpum_sf_avail()) | |
2175 | return -ENODEV; | |
2176 | ||
7e75fc3f HB |
2177 | memset(&si, 0, sizeof(si)); |
2178 | if (qsi(&si)) { | |
2179 | pr_cpumsf_err(RS_INIT_FAILURE_QSI); | |
2180 | return -ENODEV; | |
2181 | } | |
2182 | ||
9232c3c7 HB |
2183 | if (!si.as && !si.ad) |
2184 | return -ENODEV; | |
2185 | ||
7e75fc3f HB |
2186 | if (si.bsdes != sizeof(struct hws_basic_entry)) { |
2187 | pr_cpumsf_err(RS_INIT_FAILURE_BSDES); | |
2188 | return -EINVAL; | |
2189 | } | |
2190 | ||
0a648150 | 2191 | if (si.ad) { |
7e75fc3f | 2192 | sfb_set_limits(CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB); |
c43e1c5a TR |
2193 | /* Sampling of diagnostic data authorized, |
2194 | * install event into attribute list of PMU device. | |
2195 | */ | |
2196 | cpumsf_pmu_events_attr[SF_CYCLES_BASIC_DIAG_ATTR_IDX] = | |
0a648150 HB |
2197 | CPUMF_EVENT_PTR(SF, SF_CYCLES_BASIC_DIAG); |
2198 | } | |
7e75fc3f | 2199 | |
8c069ff4 | 2200 | sfdbg = debug_register(KMSG_COMPONENT, 2, 1, 80); |
ec0c0bb4 | 2201 | if (!sfdbg) { |
8c069ff4 | 2202 | pr_err("Registering for s390dbf failed\n"); |
ec0c0bb4 TR |
2203 | return -ENOMEM; |
2204 | } | |
8c069ff4 HB |
2205 | debug_register_view(sfdbg, &debug_sprintf_view); |
2206 | ||
1dad093b TH |
2207 | err = register_external_irq(EXT_IRQ_MEASURE_ALERT, |
2208 | cpumf_measurement_alert); | |
8c069ff4 | 2209 | if (err) { |
7e75fc3f | 2210 | pr_cpumsf_err(RS_INIT_FAILURE_ALRT); |
ec0c0bb4 | 2211 | debug_unregister(sfdbg); |
8c069ff4 HB |
2212 | goto out; |
2213 | } | |
2214 | ||
2215 | err = perf_pmu_register(&cpumf_sampling, "cpum_sf", PERF_TYPE_RAW); | |
2216 | if (err) { | |
7e75fc3f | 2217 | pr_cpumsf_err(RS_INIT_FAILURE_PERF); |
1dad093b TH |
2218 | unregister_external_irq(EXT_IRQ_MEASURE_ALERT, |
2219 | cpumf_measurement_alert); | |
ec0c0bb4 | 2220 | debug_unregister(sfdbg); |
8c069ff4 HB |
2221 | goto out; |
2222 | } | |
e3d617fe | 2223 | |
73c1b41e | 2224 | cpuhp_setup_state(CPUHP_AP_PERF_S390_SF_ONLINE, "perf/s390/sf:online", |
e3d617fe | 2225 | s390_pmu_sf_online_cpu, s390_pmu_sf_offline_cpu); |
8c069ff4 HB |
2226 | out: |
2227 | return err; | |
2228 | } | |
03e9e42f | 2229 | |
8c069ff4 | 2230 | arch_initcall(init_cpum_sampling_pmu); |
78d732e1 | 2231 | core_param(cpum_sfb_size, CPUM_SF_MAX_SDB, sfb_size, 0644); |