Commit | Line | Data |
---|---|---|
0e93a6ed CS |
1 | /* |
2 | * Hypervisor supplied "24x7" performance counter support | |
3 | * | |
4 | * Author: Cody P Schafer <cody@linux.vnet.ibm.com> | |
5 | * Copyright 2014 IBM Corporation. | |
6 | * | |
7 | * This program is free software; you can redistribute it and/or | |
8 | * modify it under the terms of the GNU General Public License | |
9 | * as published by the Free Software Foundation; either version | |
10 | * 2 of the License, or (at your option) any later version. | |
11 | */ | |
12 | ||
13 | #define pr_fmt(fmt) "hv-24x7: " fmt | |
14 | ||
15 | #include <linux/perf_event.h> | |
16 | #include <linux/module.h> | |
17 | #include <linux/slab.h> | |
18 | #include <asm/firmware.h> | |
19 | #include <asm/hvcall.h> | |
20 | #include <asm/io.h> | |
21 | ||
22 | #include "hv-24x7.h" | |
23 | #include "hv-24x7-catalog.h" | |
24 | #include "hv-common.h" | |
25 | ||
26 | /* | |
27 | * TODO: Merging events: | |
28 | * - Think of the hcall as an interface to a 4d array of counters: | |
29 | * - x = domains | |
30 | * - y = indexes in the domain (core, chip, vcpu, node, etc) | |
31 | * - z = offset into the counter space | |
32 | * - w = lpars (guest vms, "logical partitions") | |
33 | * - A single request is: x,y,y_last,z,z_last,w,w_last | |
34 | * - this means we can retrieve a rectangle of counters in y,z for a single x. | |
35 | * | |
36 | * - Things to consider (ignoring w): | |
37 | * - input cost_per_request = 16 | |
38 | * - output cost_per_result(ys,zs) = 8 + 8 * ys + ys * zs | |
39 | * - limited number of requests per hcall (must fit into 4K bytes) | |
40 | * - 4k = 16 [buffer header] - 16 [request size] * request_count | |
41 | * - 255 requests per hcall | |
42 | * - sometimes it will be more efficient to read extra data and discard | |
43 | */ | |
44 | ||
45 | /* | |
46 | * Example usage: | |
47 | * perf stat -e 'hv_24x7/domain=2,offset=8,starting_index=0,lpar=0xffffffff/' | |
48 | */ | |
49 | ||
50 | /* u3 0-6, one of HV_24X7_PERF_DOMAIN */ | |
51 | EVENT_DEFINE_RANGE_FORMAT(domain, config, 0, 3); | |
52 | /* u16 */ | |
53 | EVENT_DEFINE_RANGE_FORMAT(starting_index, config, 16, 31); | |
54 | /* u32, see "data_offset" */ | |
55 | EVENT_DEFINE_RANGE_FORMAT(offset, config, 32, 63); | |
56 | /* u16 */ | |
57 | EVENT_DEFINE_RANGE_FORMAT(lpar, config1, 0, 15); | |
58 | ||
59 | EVENT_DEFINE_RANGE(reserved1, config, 4, 15); | |
60 | EVENT_DEFINE_RANGE(reserved2, config1, 16, 63); | |
61 | EVENT_DEFINE_RANGE(reserved3, config2, 0, 63); | |
62 | ||
63 | static struct attribute *format_attrs[] = { | |
64 | &format_attr_domain.attr, | |
65 | &format_attr_offset.attr, | |
66 | &format_attr_starting_index.attr, | |
67 | &format_attr_lpar.attr, | |
68 | NULL, | |
69 | }; | |
70 | ||
71 | static struct attribute_group format_group = { | |
72 | .name = "format", | |
73 | .attrs = format_attrs, | |
74 | }; | |
75 | ||
76 | static struct kmem_cache *hv_page_cache; | |
77 | ||
78d13166 CS |
78 | static unsigned long h_get_24x7_catalog_page_(unsigned long phys_4096, |
79 | unsigned long version, | |
80 | unsigned long index) | |
0e93a6ed | 81 | { |
78d13166 CS |
82 | pr_devel("h_get_24x7_catalog_page(0x%lx, %lu, %lu)", |
83 | phys_4096, | |
84 | version, | |
85 | index); | |
86 | WARN_ON(!IS_ALIGNED(phys_4096, 4096)); | |
0e93a6ed | 87 | return plpar_hcall_norets(H_GET_24X7_CATALOG_PAGE, |
78d13166 | 88 | phys_4096, |
0e93a6ed CS |
89 | version, |
90 | index); | |
91 | } | |
92 | ||
1ee9fcc1 | 93 | static unsigned long h_get_24x7_catalog_page(char page[], |
bbad3e50 | 94 | u64 version, u32 index) |
78d13166 CS |
95 | { |
96 | return h_get_24x7_catalog_page_(virt_to_phys(page), | |
97 | version, index); | |
98 | } | |
99 | ||
0e93a6ed CS |
100 | static ssize_t catalog_read(struct file *filp, struct kobject *kobj, |
101 | struct bin_attribute *bin_attr, char *buf, | |
102 | loff_t offset, size_t count) | |
103 | { | |
104 | unsigned long hret; | |
105 | ssize_t ret = 0; | |
56f12bee | 106 | size_t catalog_len = 0, catalog_page_len = 0; |
0e93a6ed | 107 | loff_t page_offset = 0; |
56f12bee | 108 | loff_t offset_in_page; |
109 | size_t copy_len; | |
bbad3e50 | 110 | uint64_t catalog_version_num = 0; |
0e93a6ed CS |
111 | void *page = kmem_cache_alloc(hv_page_cache, GFP_USER); |
112 | struct hv_24x7_catalog_page_0 *page_0 = page; | |
113 | if (!page) | |
114 | return -ENOMEM; | |
115 | ||
116 | hret = h_get_24x7_catalog_page(page, 0, 0); | |
117 | if (hret) { | |
118 | ret = -EIO; | |
119 | goto e_free; | |
120 | } | |
121 | ||
bbad3e50 | 122 | catalog_version_num = be64_to_cpu(page_0->version); |
0e93a6ed CS |
123 | catalog_page_len = be32_to_cpu(page_0->length); |
124 | catalog_len = catalog_page_len * 4096; | |
125 | ||
126 | page_offset = offset / 4096; | |
56f12bee | 127 | offset_in_page = offset % 4096; |
0e93a6ed CS |
128 | |
129 | if (page_offset >= catalog_page_len) | |
130 | goto e_free; | |
131 | ||
132 | if (page_offset != 0) { | |
133 | hret = h_get_24x7_catalog_page(page, catalog_version_num, | |
134 | page_offset); | |
135 | if (hret) { | |
136 | ret = -EIO; | |
137 | goto e_free; | |
138 | } | |
139 | } | |
140 | ||
56f12bee | 141 | copy_len = 4096 - offset_in_page; |
142 | if (copy_len > count) | |
143 | copy_len = count; | |
144 | ||
145 | memcpy(buf, page+offset_in_page, copy_len); | |
146 | ret = copy_len; | |
147 | ||
0e93a6ed CS |
148 | e_free: |
149 | if (hret) | |
bbad3e50 CS |
150 | pr_err("h_get_24x7_catalog_page(ver=%lld, page=%lld) failed:" |
151 | " rc=%ld\n", | |
152 | catalog_version_num, page_offset, hret); | |
d6589722 | 153 | kmem_cache_free(hv_page_cache, page); |
0e93a6ed | 154 | |
56f12bee | 155 | pr_devel("catalog_read: offset=%lld(%lld) count=%zu " |
156 | "catalog_len=%zu(%zu) => %zd\n", offset, page_offset, | |
157 | count, catalog_len, catalog_page_len, ret); | |
0e93a6ed CS |
158 | |
159 | return ret; | |
160 | } | |
161 | ||
162 | #define PAGE_0_ATTR(_name, _fmt, _expr) \ | |
163 | static ssize_t _name##_show(struct device *dev, \ | |
164 | struct device_attribute *dev_attr, \ | |
165 | char *buf) \ | |
166 | { \ | |
167 | unsigned long hret; \ | |
168 | ssize_t ret = 0; \ | |
169 | void *page = kmem_cache_alloc(hv_page_cache, GFP_USER); \ | |
170 | struct hv_24x7_catalog_page_0 *page_0 = page; \ | |
171 | if (!page) \ | |
172 | return -ENOMEM; \ | |
173 | hret = h_get_24x7_catalog_page(page, 0, 0); \ | |
174 | if (hret) { \ | |
175 | ret = -EIO; \ | |
176 | goto e_free; \ | |
177 | } \ | |
178 | ret = sprintf(buf, _fmt, _expr); \ | |
179 | e_free: \ | |
180 | kfree(page); \ | |
181 | return ret; \ | |
182 | } \ | |
183 | static DEVICE_ATTR_RO(_name) | |
184 | ||
185 | PAGE_0_ATTR(catalog_version, "%lld\n", | |
bbad3e50 | 186 | (unsigned long long)be64_to_cpu(page_0->version)); |
0e93a6ed CS |
187 | PAGE_0_ATTR(catalog_len, "%lld\n", |
188 | (unsigned long long)be32_to_cpu(page_0->length) * 4096); | |
189 | static BIN_ATTR_RO(catalog, 0/* real length varies */); | |
190 | ||
191 | static struct bin_attribute *if_bin_attrs[] = { | |
192 | &bin_attr_catalog, | |
193 | NULL, | |
194 | }; | |
195 | ||
196 | static struct attribute *if_attrs[] = { | |
197 | &dev_attr_catalog_len.attr, | |
198 | &dev_attr_catalog_version.attr, | |
199 | NULL, | |
200 | }; | |
201 | ||
202 | static struct attribute_group if_group = { | |
203 | .name = "interface", | |
204 | .bin_attrs = if_bin_attrs, | |
205 | .attrs = if_attrs, | |
206 | }; | |
207 | ||
208 | static const struct attribute_group *attr_groups[] = { | |
209 | &format_group, | |
210 | &if_group, | |
211 | NULL, | |
212 | }; | |
213 | ||
214 | static bool is_physical_domain(int domain) | |
215 | { | |
216 | return domain == HV_24X7_PERF_DOMAIN_PHYSICAL_CHIP || | |
217 | domain == HV_24X7_PERF_DOMAIN_PHYSICAL_CORE; | |
218 | } | |
219 | ||
220 | static unsigned long single_24x7_request(u8 domain, u32 offset, u16 ix, | |
221 | u16 lpar, u64 *res, | |
222 | bool success_expected) | |
223 | { | |
48bee8a6 | 224 | unsigned long ret = -ENOMEM; |
0e93a6ed CS |
225 | |
226 | /* | |
227 | * request_buffer and result_buffer are not required to be 4k aligned, | |
228 | * but are not allowed to cross any 4k boundary. Aligning them to 4k is | |
229 | * the simplest way to ensure that. | |
230 | */ | |
231 | struct reqb { | |
232 | struct hv_24x7_request_buffer buf; | |
233 | struct hv_24x7_request req; | |
48bee8a6 CS |
234 | } __packed *request_buffer; |
235 | ||
236 | struct { | |
237 | struct hv_24x7_data_result_buffer buf; | |
238 | struct hv_24x7_result res; | |
239 | struct hv_24x7_result_element elem; | |
240 | __be64 result; | |
241 | } __packed *result_buffer; | |
242 | ||
243 | BUILD_BUG_ON(sizeof(*request_buffer) > 4096); | |
244 | BUILD_BUG_ON(sizeof(*result_buffer) > 4096); | |
245 | ||
246 | request_buffer = kmem_cache_zalloc(hv_page_cache, GFP_USER); | |
247 | if (!request_buffer) | |
248 | goto out; | |
249 | ||
250 | result_buffer = kmem_cache_zalloc(hv_page_cache, GFP_USER); | |
251 | if (!result_buffer) | |
252 | goto out_free_request_buffer; | |
253 | ||
254 | *request_buffer = (struct reqb) { | |
0e93a6ed CS |
255 | .buf = { |
256 | .interface_version = HV_24X7_IF_VERSION_CURRENT, | |
257 | .num_requests = 1, | |
258 | }, | |
259 | .req = { | |
260 | .performance_domain = domain, | |
261 | .data_size = cpu_to_be16(8), | |
262 | .data_offset = cpu_to_be32(offset), | |
263 | .starting_lpar_ix = cpu_to_be16(lpar), | |
264 | .max_num_lpars = cpu_to_be16(1), | |
265 | .starting_ix = cpu_to_be16(ix), | |
266 | .max_ix = cpu_to_be16(1), | |
267 | } | |
268 | }; | |
269 | ||
0e93a6ed | 270 | ret = plpar_hcall_norets(H_GET_24X7_DATA, |
48bee8a6 CS |
271 | virt_to_phys(request_buffer), sizeof(*request_buffer), |
272 | virt_to_phys(result_buffer), sizeof(*result_buffer)); | |
0e93a6ed CS |
273 | |
274 | if (ret) { | |
275 | if (success_expected) | |
48bee8a6 CS |
276 | pr_err_ratelimited("hcall failed: %d %#x %#x %d => " |
277 | "0x%lx (%ld) detail=0x%x failing ix=%x\n", | |
278 | domain, offset, ix, lpar, ret, ret, | |
279 | result_buffer->buf.detailed_rc, | |
280 | result_buffer->buf.failing_request_ix); | |
281 | goto out_free_result_buffer; | |
0e93a6ed CS |
282 | } |
283 | ||
48bee8a6 CS |
284 | *res = be64_to_cpu(result_buffer->result); |
285 | ||
286 | out_free_result_buffer: | |
287 | kfree(result_buffer); | |
288 | out_free_request_buffer: | |
289 | kfree(request_buffer); | |
290 | out: | |
0e93a6ed CS |
291 | return ret; |
292 | } | |
293 | ||
294 | static unsigned long event_24x7_request(struct perf_event *event, u64 *res, | |
295 | bool success_expected) | |
296 | { | |
297 | return single_24x7_request(event_get_domain(event), | |
298 | event_get_offset(event), | |
299 | event_get_starting_index(event), | |
300 | event_get_lpar(event), | |
301 | res, | |
302 | success_expected); | |
303 | } | |
304 | ||
305 | static int h_24x7_event_init(struct perf_event *event) | |
306 | { | |
307 | struct hv_perf_caps caps; | |
308 | unsigned domain; | |
309 | unsigned long hret; | |
310 | u64 ct; | |
311 | ||
312 | /* Not our event */ | |
313 | if (event->attr.type != event->pmu->type) | |
314 | return -ENOENT; | |
315 | ||
316 | /* Unused areas must be 0 */ | |
317 | if (event_get_reserved1(event) || | |
318 | event_get_reserved2(event) || | |
319 | event_get_reserved3(event)) { | |
320 | pr_devel("reserved set when forbidden 0x%llx(0x%llx) 0x%llx(0x%llx) 0x%llx(0x%llx)\n", | |
321 | event->attr.config, | |
322 | event_get_reserved1(event), | |
323 | event->attr.config1, | |
324 | event_get_reserved2(event), | |
325 | event->attr.config2, | |
326 | event_get_reserved3(event)); | |
327 | return -EINVAL; | |
328 | } | |
329 | ||
330 | /* unsupported modes and filters */ | |
331 | if (event->attr.exclude_user || | |
332 | event->attr.exclude_kernel || | |
333 | event->attr.exclude_hv || | |
334 | event->attr.exclude_idle || | |
335 | event->attr.exclude_host || | |
cc56d673 | 336 | event->attr.exclude_guest) |
0e93a6ed CS |
337 | return -EINVAL; |
338 | ||
339 | /* no branch sampling */ | |
340 | if (has_branch_stack(event)) | |
341 | return -EOPNOTSUPP; | |
342 | ||
343 | /* offset must be 8 byte aligned */ | |
344 | if (event_get_offset(event) % 8) { | |
345 | pr_devel("bad alignment\n"); | |
346 | return -EINVAL; | |
347 | } | |
348 | ||
349 | /* Domains above 6 are invalid */ | |
350 | domain = event_get_domain(event); | |
351 | if (domain > 6) { | |
352 | pr_devel("invalid domain %d\n", domain); | |
353 | return -EINVAL; | |
354 | } | |
355 | ||
356 | hret = hv_perf_caps_get(&caps); | |
357 | if (hret) { | |
358 | pr_devel("could not get capabilities: rc=%ld\n", hret); | |
359 | return -EIO; | |
360 | } | |
361 | ||
362 | /* PHYSICAL domains & other lpars require extra capabilities */ | |
363 | if (!caps.collect_privileged && (is_physical_domain(domain) || | |
364 | (event_get_lpar(event) != event_get_lpar_max()))) { | |
365 | pr_devel("hv permisions disallow: is_physical_domain:%d, lpar=0x%llx\n", | |
366 | is_physical_domain(domain), | |
367 | event_get_lpar(event)); | |
368 | return -EACCES; | |
369 | } | |
370 | ||
371 | /* see if the event complains */ | |
372 | if (event_24x7_request(event, &ct, false)) { | |
373 | pr_devel("test hcall failed\n"); | |
374 | return -EIO; | |
375 | } | |
376 | ||
377 | return 0; | |
378 | } | |
379 | ||
380 | static u64 h_24x7_get_value(struct perf_event *event) | |
381 | { | |
382 | unsigned long ret; | |
383 | u64 ct; | |
384 | ret = event_24x7_request(event, &ct, true); | |
385 | if (ret) | |
386 | /* We checked this in event init, shouldn't fail here... */ | |
387 | return 0; | |
388 | ||
389 | return ct; | |
390 | } | |
391 | ||
392 | static void h_24x7_event_update(struct perf_event *event) | |
393 | { | |
394 | s64 prev; | |
395 | u64 now; | |
396 | now = h_24x7_get_value(event); | |
397 | prev = local64_xchg(&event->hw.prev_count, now); | |
398 | local64_add(now - prev, &event->count); | |
399 | } | |
400 | ||
401 | static void h_24x7_event_start(struct perf_event *event, int flags) | |
402 | { | |
403 | if (flags & PERF_EF_RELOAD) | |
404 | local64_set(&event->hw.prev_count, h_24x7_get_value(event)); | |
405 | } | |
406 | ||
407 | static void h_24x7_event_stop(struct perf_event *event, int flags) | |
408 | { | |
409 | h_24x7_event_update(event); | |
410 | } | |
411 | ||
412 | static int h_24x7_event_add(struct perf_event *event, int flags) | |
413 | { | |
414 | if (flags & PERF_EF_START) | |
415 | h_24x7_event_start(event, flags); | |
416 | ||
417 | return 0; | |
418 | } | |
419 | ||
0e93a6ed CS |
420 | static struct pmu h_24x7_pmu = { |
421 | .task_ctx_nr = perf_invalid_context, | |
422 | ||
423 | .name = "hv_24x7", | |
424 | .attr_groups = attr_groups, | |
425 | .event_init = h_24x7_event_init, | |
426 | .add = h_24x7_event_add, | |
427 | .del = h_24x7_event_stop, | |
428 | .start = h_24x7_event_start, | |
429 | .stop = h_24x7_event_stop, | |
430 | .read = h_24x7_event_update, | |
0e93a6ed CS |
431 | }; |
432 | ||
433 | static int hv_24x7_init(void) | |
434 | { | |
435 | int r; | |
436 | unsigned long hret; | |
437 | struct hv_perf_caps caps; | |
438 | ||
439 | if (!firmware_has_feature(FW_FEATURE_LPAR)) { | |
e98bf005 | 440 | pr_debug("not a virtualized system, not enabling\n"); |
0e93a6ed CS |
441 | return -ENODEV; |
442 | } | |
443 | ||
444 | hret = hv_perf_caps_get(&caps); | |
445 | if (hret) { | |
e98bf005 | 446 | pr_debug("could not obtain capabilities, not enabling, rc=%ld\n", |
0e93a6ed CS |
447 | hret); |
448 | return -ENODEV; | |
449 | } | |
450 | ||
451 | hv_page_cache = kmem_cache_create("hv-page-4096", 4096, 4096, 0, NULL); | |
452 | if (!hv_page_cache) | |
453 | return -ENOMEM; | |
454 | ||
cc56d673 VW |
455 | /* sampling not supported */ |
456 | h_24x7_pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT; | |
457 | ||
0e93a6ed CS |
458 | r = perf_pmu_register(&h_24x7_pmu, h_24x7_pmu.name, -1); |
459 | if (r) | |
460 | return r; | |
461 | ||
462 | return 0; | |
463 | } | |
464 | ||
465 | device_initcall(hv_24x7_init); |