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 78 static unsigned long h_get_24x7_catalog_page_(unsigned long phys_4096, 79 unsigned long version, 80 unsigned long index) 81 { 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)); 87 return plpar_hcall_norets(H_GET_24X7_CATALOG_PAGE, 88 phys_4096, 89 version, 90 index); 91 } 92 93 static unsigned long h_get_24x7_catalog_page(char page[], 94 u64 version, u32 index) 95 { 96 return h_get_24x7_catalog_page_(virt_to_phys(page), 97 version, index); 98 } 99 100 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; 106 size_t catalog_len = 0, catalog_page_len = 0; 107 loff_t page_offset = 0; 108 loff_t offset_in_page; 109 size_t copy_len; 110 uint64_t catalog_version_num = 0; 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 } 121 122 catalog_version_num = be64_to_cpu(page_0->version); 123 catalog_page_len = be32_to_cpu(page_0->length); 124 catalog_len = catalog_page_len * 4096; 125 126 page_offset = offset / 4096; 127 offset_in_page = offset % 4096; 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 } 140 141 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; 147 148 e_free: 149 if (hret) 150 pr_err("h_get_24x7_catalog_page(ver=%lld, page=%lld) failed:" 151 " rc=%ld\n", 152 catalog_version_num, page_offset, hret); 153 kmem_cache_free(hv_page_cache, page); 154 155 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); 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 kmem_cache_free(hv_page_cache, page); \ 181 return ret; \ 182 } \ 183 static DEVICE_ATTR_RO(_name) 184 185 PAGE_0_ATTR(catalog_version, "%lld\n", 186 (unsigned long long)be64_to_cpu(page_0->version)); 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 DEFINE_PER_CPU(char, hv_24x7_reqb[4096]) __aligned(4096); 221 DEFINE_PER_CPU(char, hv_24x7_resb[4096]) __aligned(4096); 222 223 static unsigned long single_24x7_request(u8 domain, u32 offset, u16 ix, 224 u16 lpar, u64 *res, 225 bool success_expected) 226 { 227 unsigned long ret; 228 229 /* 230 * request_buffer and result_buffer are not required to be 4k aligned, 231 * but are not allowed to cross any 4k boundary. Aligning them to 4k is 232 * the simplest way to ensure that. 233 */ 234 struct reqb { 235 struct hv_24x7_request_buffer buf; 236 struct hv_24x7_request req; 237 } __packed *request_buffer; 238 239 struct { 240 struct hv_24x7_data_result_buffer buf; 241 struct hv_24x7_result res; 242 struct hv_24x7_result_element elem; 243 __be64 result; 244 } __packed *result_buffer; 245 246 BUILD_BUG_ON(sizeof(*request_buffer) > 4096); 247 BUILD_BUG_ON(sizeof(*result_buffer) > 4096); 248 249 request_buffer = (void *)get_cpu_var(hv_24x7_reqb); 250 result_buffer = (void *)get_cpu_var(hv_24x7_resb); 251 252 memset(request_buffer, 0, 4096); 253 memset(result_buffer, 0, 4096); 254 255 *request_buffer = (struct reqb) { 256 .buf = { 257 .interface_version = HV_24X7_IF_VERSION_CURRENT, 258 .num_requests = 1, 259 }, 260 .req = { 261 .performance_domain = domain, 262 .data_size = cpu_to_be16(8), 263 .data_offset = cpu_to_be32(offset), 264 .starting_lpar_ix = cpu_to_be16(lpar), 265 .max_num_lpars = cpu_to_be16(1), 266 .starting_ix = cpu_to_be16(ix), 267 .max_ix = cpu_to_be16(1), 268 } 269 }; 270 271 ret = plpar_hcall_norets(H_GET_24X7_DATA, 272 virt_to_phys(request_buffer), sizeof(*request_buffer), 273 virt_to_phys(result_buffer), sizeof(*result_buffer)); 274 275 if (ret) { 276 if (success_expected) 277 pr_err_ratelimited("hcall failed: %d %#x %#x %d => " 278 "0x%lx (%ld) detail=0x%x failing ix=%x\n", 279 domain, offset, ix, lpar, ret, ret, 280 result_buffer->buf.detailed_rc, 281 result_buffer->buf.failing_request_ix); 282 goto out; 283 } 284 285 *res = be64_to_cpu(result_buffer->result); 286 287 out: 288 return ret; 289 } 290 291 static unsigned long event_24x7_request(struct perf_event *event, u64 *res, 292 bool success_expected) 293 { 294 return single_24x7_request(event_get_domain(event), 295 event_get_offset(event), 296 event_get_starting_index(event), 297 event_get_lpar(event), 298 res, 299 success_expected); 300 } 301 302 static int h_24x7_event_init(struct perf_event *event) 303 { 304 struct hv_perf_caps caps; 305 unsigned domain; 306 unsigned long hret; 307 u64 ct; 308 309 /* Not our event */ 310 if (event->attr.type != event->pmu->type) 311 return -ENOENT; 312 313 /* Unused areas must be 0 */ 314 if (event_get_reserved1(event) || 315 event_get_reserved2(event) || 316 event_get_reserved3(event)) { 317 pr_devel("reserved set when forbidden 0x%llx(0x%llx) 0x%llx(0x%llx) 0x%llx(0x%llx)\n", 318 event->attr.config, 319 event_get_reserved1(event), 320 event->attr.config1, 321 event_get_reserved2(event), 322 event->attr.config2, 323 event_get_reserved3(event)); 324 return -EINVAL; 325 } 326 327 /* unsupported modes and filters */ 328 if (event->attr.exclude_user || 329 event->attr.exclude_kernel || 330 event->attr.exclude_hv || 331 event->attr.exclude_idle || 332 event->attr.exclude_host || 333 event->attr.exclude_guest) 334 return -EINVAL; 335 336 /* no branch sampling */ 337 if (has_branch_stack(event)) 338 return -EOPNOTSUPP; 339 340 /* offset must be 8 byte aligned */ 341 if (event_get_offset(event) % 8) { 342 pr_devel("bad alignment\n"); 343 return -EINVAL; 344 } 345 346 /* Domains above 6 are invalid */ 347 domain = event_get_domain(event); 348 if (domain > 6) { 349 pr_devel("invalid domain %d\n", domain); 350 return -EINVAL; 351 } 352 353 hret = hv_perf_caps_get(&caps); 354 if (hret) { 355 pr_devel("could not get capabilities: rc=%ld\n", hret); 356 return -EIO; 357 } 358 359 /* PHYSICAL domains & other lpars require extra capabilities */ 360 if (!caps.collect_privileged && (is_physical_domain(domain) || 361 (event_get_lpar(event) != event_get_lpar_max()))) { 362 pr_devel("hv permisions disallow: is_physical_domain:%d, lpar=0x%llx\n", 363 is_physical_domain(domain), 364 event_get_lpar(event)); 365 return -EACCES; 366 } 367 368 /* see if the event complains */ 369 if (event_24x7_request(event, &ct, false)) { 370 pr_devel("test hcall failed\n"); 371 return -EIO; 372 } 373 374 return 0; 375 } 376 377 static u64 h_24x7_get_value(struct perf_event *event) 378 { 379 unsigned long ret; 380 u64 ct; 381 ret = event_24x7_request(event, &ct, true); 382 if (ret) 383 /* We checked this in event init, shouldn't fail here... */ 384 return 0; 385 386 return ct; 387 } 388 389 static void h_24x7_event_update(struct perf_event *event) 390 { 391 s64 prev; 392 u64 now; 393 now = h_24x7_get_value(event); 394 prev = local64_xchg(&event->hw.prev_count, now); 395 local64_add(now - prev, &event->count); 396 } 397 398 static void h_24x7_event_start(struct perf_event *event, int flags) 399 { 400 if (flags & PERF_EF_RELOAD) 401 local64_set(&event->hw.prev_count, h_24x7_get_value(event)); 402 } 403 404 static void h_24x7_event_stop(struct perf_event *event, int flags) 405 { 406 h_24x7_event_update(event); 407 } 408 409 static int h_24x7_event_add(struct perf_event *event, int flags) 410 { 411 if (flags & PERF_EF_START) 412 h_24x7_event_start(event, flags); 413 414 return 0; 415 } 416 417 static struct pmu h_24x7_pmu = { 418 .task_ctx_nr = perf_invalid_context, 419 420 .name = "hv_24x7", 421 .attr_groups = attr_groups, 422 .event_init = h_24x7_event_init, 423 .add = h_24x7_event_add, 424 .del = h_24x7_event_stop, 425 .start = h_24x7_event_start, 426 .stop = h_24x7_event_stop, 427 .read = h_24x7_event_update, 428 }; 429 430 static int hv_24x7_init(void) 431 { 432 int r; 433 unsigned long hret; 434 struct hv_perf_caps caps; 435 436 if (!firmware_has_feature(FW_FEATURE_LPAR)) { 437 pr_debug("not a virtualized system, not enabling\n"); 438 return -ENODEV; 439 } 440 441 hret = hv_perf_caps_get(&caps); 442 if (hret) { 443 pr_debug("could not obtain capabilities, not enabling, rc=%ld\n", 444 hret); 445 return -ENODEV; 446 } 447 448 hv_page_cache = kmem_cache_create("hv-page-4096", 4096, 4096, 0, NULL); 449 if (!hv_page_cache) 450 return -ENOMEM; 451 452 /* sampling not supported */ 453 h_24x7_pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT; 454 455 r = perf_pmu_register(&h_24x7_pmu, h_24x7_pmu.name, -1); 456 if (r) 457 return r; 458 459 return 0; 460 } 461 462 device_initcall(hv_24x7_init); 463