arch/powerpc/perf/hv-24x7.c

   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/rbtree.h>
  17 #include <linux/module.h>
  18 #include <linux/slab.h>
  19 #include <linux/vmalloc.h>
  20
  21 #include <asm/firmware.h>
  22 #include <asm/hvcall.h>
  23 #include <asm/io.h>
  24 #include <linux/byteorder/generic.h>
  25
  26 #include "hv-24x7.h"
  27 #include "hv-24x7-catalog.h"
  28 #include "hv-common.h"
  29
  30 static const char *event_domain_suffix(unsigned domain)
  31 {
  32         switch (domain) {
  33 #define DOMAIN(n, v, x, c)              \
  34         case HV_PERF_DOMAIN_##n:        \
  35                 return "__" #n;
  36 #include "hv-24x7-domains.h"
  37 #undef DOMAIN
  38         default:
  39                 WARN(1, "unknown domain %d\n", domain);
  40                 return "__UNKNOWN_DOMAIN_SUFFIX";
  41         }
  42 }
  43
  44 static bool domain_is_valid(unsigned domain)
  45 {
  46         switch (domain) {
  47 #define DOMAIN(n, v, x, c)              \
  48         case HV_PERF_DOMAIN_##n:        \
  49                 /* fall through */
  50 #include "hv-24x7-domains.h"
  51 #undef DOMAIN
  52                 return true;
  53         default:
  54                 return false;
  55         }
  56 }
  57
  58 static bool is_physical_domain(unsigned domain)
  59 {
  60         switch (domain) {
  61 #define DOMAIN(n, v, x, c)              \
  62         case HV_PERF_DOMAIN_##n:        \
  63                 return c;
  64 #include "hv-24x7-domains.h"
  65 #undef DOMAIN
  66         default:
  67                 return false;
  68         }
  69 }
  70
  71 static bool catalog_entry_domain_is_valid(unsigned domain)
  72 {
  73         return is_physical_domain(domain);
  74 }
  75
  76 /*
  77  * TODO: Merging events:
  78  * - Think of the hcall as an interface to a 4d array of counters:
  79  *   - x = domains
  80  *   - y = indexes in the domain (core, chip, vcpu, node, etc)
  81  *   - z = offset into the counter space
  82  *   - w = lpars (guest vms, "logical partitions")
  83  * - A single request is: x,y,y_last,z,z_last,w,w_last
  84  *   - this means we can retrieve a rectangle of counters in y,z for a single x.
  85  *
  86  * - Things to consider (ignoring w):
  87  *   - input  cost_per_request = 16
  88  *   - output cost_per_result(ys,zs)  = 8 + 8 * ys + ys * zs
  89  *   - limited number of requests per hcall (must fit into 4K bytes)
  90  *     - 4k = 16 [buffer header] - 16 [request size] * request_count
  91  *     - 255 requests per hcall
  92  *   - sometimes it will be more efficient to read extra data and discard
  93  */
  94
  95 /*
  96  * Example usage:
  97  *  perf stat -e 'hv_24x7/domain=2,offset=8,vcpu=0,lpar=0xffffffff/'
  98  */
  99
 100 /* u3 0-6, one of HV_24X7_PERF_DOMAIN */
 101 EVENT_DEFINE_RANGE_FORMAT(domain, config, 0, 3);
 102 /* u16 */
 103 EVENT_DEFINE_RANGE_FORMAT(core, config, 16, 31);
 104 EVENT_DEFINE_RANGE_FORMAT(vcpu, config, 16, 31);
 105 /* u32, see "data_offset" */
 106 EVENT_DEFINE_RANGE_FORMAT(offset, config, 32, 63);
 107 /* u16 */
 108 EVENT_DEFINE_RANGE_FORMAT(lpar, config1, 0, 15);
 109
 110 EVENT_DEFINE_RANGE(reserved1, config,   4, 15);
 111 EVENT_DEFINE_RANGE(reserved2, config1, 16, 63);
 112 EVENT_DEFINE_RANGE(reserved3, config2,  0, 63);
 113
 114 static struct attribute *format_attrs[] = {
 115         &format_attr_domain.attr,
 116         &format_attr_offset.attr,
 117         &format_attr_core.attr,
 118         &format_attr_vcpu.attr,
 119         &format_attr_lpar.attr,
 120         NULL,
 121 };
 122
 123 static struct attribute_group format_group = {
 124         .name = "format",
 125         .attrs = format_attrs,
 126 };
 127
 128 static struct attribute_group event_group = {
 129         .name = "events",
 130         /* .attrs is set in init */
 131 };
 132
 133 static struct attribute_group event_desc_group = {
 134         .name = "event_descs",
 135         /* .attrs is set in init */
 136 };
 137
 138 static struct attribute_group event_long_desc_group = {
 139         .name = "event_long_descs",
 140         /* .attrs is set in init */
 141 };
 142
 143 static struct kmem_cache *hv_page_cache;
 144
 145 DEFINE_PER_CPU(int, hv_24x7_txn_flags);
 146 DEFINE_PER_CPU(int, hv_24x7_txn_err);
 147
 148 struct hv_24x7_hw {
 149         struct perf_event *events[255];
 150 };
 151
 152 DEFINE_PER_CPU(struct hv_24x7_hw, hv_24x7_hw);
 153
 154 /*
 155  * request_buffer and result_buffer are not required to be 4k aligned,
 156  * but are not allowed to cross any 4k boundary. Aligning them to 4k is
 157  * the simplest way to ensure that.
 158  */
 159 #define H24x7_DATA_BUFFER_SIZE  4096
 160 DEFINE_PER_CPU(char, hv_24x7_reqb[H24x7_DATA_BUFFER_SIZE]) __aligned(4096);
 161 DEFINE_PER_CPU(char, hv_24x7_resb[H24x7_DATA_BUFFER_SIZE]) __aligned(4096);
 162
 163 static char *event_name(struct hv_24x7_event_data *ev, int *len)
 164 {
 165         *len = be16_to_cpu(ev->event_name_len) - 2;
 166         return (char *)ev->remainder;
 167 }
 168
 169 static char *event_desc(struct hv_24x7_event_data *ev, int *len)
 170 {
 171         unsigned nl = be16_to_cpu(ev->event_name_len);
 172         __be16 *desc_len = (__be16 *)(ev->remainder + nl - 2);
 173
 174         *len = be16_to_cpu(*desc_len) - 2;
 175         return (char *)ev->remainder + nl;
 176 }
 177
 178 static char *event_long_desc(struct hv_24x7_event_data *ev, int *len)
 179 {
 180         unsigned nl = be16_to_cpu(ev->event_name_len);
 181         __be16 *desc_len_ = (__be16 *)(ev->remainder + nl - 2);
 182         unsigned desc_len = be16_to_cpu(*desc_len_);
 183         __be16 *long_desc_len = (__be16 *)(ev->remainder + nl + desc_len - 2);
 184
 185         *len = be16_to_cpu(*long_desc_len) - 2;
 186         return (char *)ev->remainder + nl + desc_len;
 187 }
 188
 189 static bool event_fixed_portion_is_within(struct hv_24x7_event_data *ev,
 190                                           void *end)
 191 {
 192         void *start = ev;
 193
 194         return (start + offsetof(struct hv_24x7_event_data, remainder)) < end;
 195 }
 196
 197 /*
 198  * Things we don't check:
 199  *  - padding for desc, name, and long/detailed desc is required to be '\0'
 200  *    bytes.
 201  *
 202  *  Return NULL if we pass end,
 203  *  Otherwise return the address of the byte just following the event.
 204  */
 205 static void *event_end(struct hv_24x7_event_data *ev, void *end)
 206 {
 207         void *start = ev;
 208         __be16 *dl_, *ldl_;
 209         unsigned dl, ldl;
 210         unsigned nl = be16_to_cpu(ev->event_name_len);
 211
 212         if (nl < 2) {
 213                 pr_debug("%s: name length too short: %d", __func__, nl);
 214                 return NULL;
 215         }
 216
 217         if (start + nl > end) {
 218                 pr_debug("%s: start=%p + nl=%u > end=%p",
 219                                 __func__, start, nl, end);
 220                 return NULL;
 221         }
 222
 223         dl_ = (__be16 *)(ev->remainder + nl - 2);
 224         if (!IS_ALIGNED((uintptr_t)dl_, 2))
 225                 pr_warn("desc len not aligned %p", dl_);
 226         dl = be16_to_cpu(*dl_);
 227         if (dl < 2) {
 228                 pr_debug("%s: desc len too short: %d", __func__, dl);
 229                 return NULL;
 230         }
 231
 232         if (start + nl + dl > end) {
 233                 pr_debug("%s: (start=%p + nl=%u + dl=%u)=%p > end=%p",
 234                                 __func__, start, nl, dl, start + nl + dl, end);
 235                 return NULL;
 236         }
 237
 238         ldl_ = (__be16 *)(ev->remainder + nl + dl - 2);
 239         if (!IS_ALIGNED((uintptr_t)ldl_, 2))
 240                 pr_warn("long desc len not aligned %p", ldl_);
 241         ldl = be16_to_cpu(*ldl_);
 242         if (ldl < 2) {
 243                 pr_debug("%s: long desc len too short (ldl=%u)",
 244                                 __func__, ldl);
 245                 return NULL;
 246         }
 247
 248         if (start + nl + dl + ldl > end) {
 249                 pr_debug("%s: start=%p + nl=%u + dl=%u + ldl=%u > end=%p",
 250                                 __func__, start, nl, dl, ldl, end);
 251                 return NULL;
 252         }
 253
 254         return start + nl + dl + ldl;
 255 }
 256
 257 static unsigned long h_get_24x7_catalog_page_(unsigned long phys_4096,
 258                                               unsigned long version,
 259                                               unsigned long index)
 260 {
 261         pr_devel("h_get_24x7_catalog_page(0x%lx, %lu, %lu)",
 262                         phys_4096, version, index);
 263
 264         WARN_ON(!IS_ALIGNED(phys_4096, 4096));
 265
 266         return plpar_hcall_norets(H_GET_24X7_CATALOG_PAGE,
 267                         phys_4096, version, index);
 268 }
 269
 270 static unsigned long h_get_24x7_catalog_page(char page[],
 271                                              u64 version, u32 index)
 272 {
 273         return h_get_24x7_catalog_page_(virt_to_phys(page),
 274                                         version, index);
 275 }
 276
 277 static unsigned core_domains[] = {
 278         HV_PERF_DOMAIN_PHYS_CORE,
 279         HV_PERF_DOMAIN_VCPU_HOME_CORE,
 280         HV_PERF_DOMAIN_VCPU_HOME_CHIP,
 281         HV_PERF_DOMAIN_VCPU_HOME_NODE,
 282         HV_PERF_DOMAIN_VCPU_REMOTE_NODE,
 283 };
 284 /* chip event data always yeilds a single event, core yeilds multiple */
 285 #define MAX_EVENTS_PER_EVENT_DATA ARRAY_SIZE(core_domains)
 286
 287 static char *event_fmt(struct hv_24x7_event_data *event, unsigned domain)
 288 {
 289         const char *sindex;
 290         const char *lpar;
 291
 292         if (is_physical_domain(domain)) {
 293                 lpar = "0x0";
 294                 sindex = "core";
 295         } else {
 296                 lpar = "?";
 297                 sindex = "vcpu";
 298         }
 299
 300         return kasprintf(GFP_KERNEL,
 301                         "domain=0x%x,offset=0x%x,%s=?,lpar=%s",
 302                         domain,
 303                         be16_to_cpu(event->event_counter_offs) +
 304                                 be16_to_cpu(event->event_group_record_offs),
 305                         sindex,
 306                         lpar);
 307 }
 308
 309 /* Avoid trusting fw to NUL terminate strings */
 310 static char *memdup_to_str(char *maybe_str, int max_len, gfp_t gfp)
 311 {
 312         return kasprintf(gfp, "%.*s", max_len, maybe_str);
 313 }
 314
 315 static ssize_t device_show_string(struct device *dev,
 316                 struct device_attribute *attr, char *buf)
 317 {
 318         struct dev_ext_attribute *d;
 319
 320         d = container_of(attr, struct dev_ext_attribute, attr);
 321
 322         return sprintf(buf, "%s\n", (char *)d->var);
 323 }
 324
 325 static struct attribute *device_str_attr_create_(char *name, char *str)
 326 {
 327         struct dev_ext_attribute *attr = kzalloc(sizeof(*attr), GFP_KERNEL);
 328
 329         if (!attr)
 330                 return NULL;
 331
 332         sysfs_attr_init(&attr->attr.attr);
 333
 334         attr->var = str;
 335         attr->attr.attr.name = name;
 336         attr->attr.attr.mode = 0444;
 337         attr->attr.show = device_show_string;
 338
 339         return &attr->attr.attr;
 340 }
 341
 342 static struct attribute *device_str_attr_create(char *name, int name_max,
 343                                                 int name_nonce,
 344                                                 char *str, size_t str_max)
 345 {
 346         char *n;
 347         char *s = memdup_to_str(str, str_max, GFP_KERNEL);
 348         struct attribute *a;
 349
 350         if (!s)
 351                 return NULL;
 352
 353         if (!name_nonce)
 354                 n = kasprintf(GFP_KERNEL, "%.*s", name_max, name);
 355         else
 356                 n = kasprintf(GFP_KERNEL, "%.*s__%d", name_max, name,
 357                                         name_nonce);
 358         if (!n)
 359                 goto out_s;
 360
 361         a = device_str_attr_create_(n, s);
 362         if (!a)
 363                 goto out_n;
 364
 365         return a;
 366 out_n:
 367         kfree(n);
 368 out_s:
 369         kfree(s);
 370         return NULL;
 371 }
 372
 373 static void device_str_attr_destroy(struct attribute *attr)
 374 {
 375         struct dev_ext_attribute *d;
 376
 377         d = container_of(attr, struct dev_ext_attribute, attr.attr);
 378         kfree(d->var);
 379         kfree(d->attr.attr.name);
 380         kfree(d);
 381 }
 382
 383 static struct attribute *event_to_attr(unsigned ix,
 384                                        struct hv_24x7_event_data *event,
 385                                        unsigned domain,
 386                                        int nonce)
 387 {
 388         int event_name_len;
 389         char *ev_name, *a_ev_name, *val;
 390         const char *ev_suffix;
 391         struct attribute *attr;
 392
 393         if (!domain_is_valid(domain)) {
 394                 pr_warn("catalog event %u has invalid domain %u\n",
 395                                 ix, domain);
 396                 return NULL;
 397         }
 398
 399         val = event_fmt(event, domain);
 400         if (!val)
 401                 return NULL;
 402
 403         ev_suffix = event_domain_suffix(domain);
 404         ev_name = event_name(event, &event_name_len);
 405         if (!nonce)
 406                 a_ev_name = kasprintf(GFP_KERNEL, "%.*s%s",
 407                                 (int)event_name_len, ev_name, ev_suffix);
 408         else
 409                 a_ev_name = kasprintf(GFP_KERNEL, "%.*s%s__%d",
 410                                 (int)event_name_len, ev_name, ev_suffix, nonce);
 411
 412         if (!a_ev_name)
 413                 goto out_val;
 414
 415         attr = device_str_attr_create_(a_ev_name, val);
 416         if (!attr)
 417                 goto out_name;
 418
 419         return attr;
 420 out_name:
 421         kfree(a_ev_name);
 422 out_val:
 423         kfree(val);
 424         return NULL;
 425 }
 426
 427 static struct attribute *event_to_desc_attr(struct hv_24x7_event_data *event,
 428                                             int nonce)
 429 {
 430         int nl, dl;
 431         char *name = event_name(event, &nl);
 432         char *desc = event_desc(event, &dl);
 433
 434         /* If there isn't a description, don't create the sysfs file */
 435         if (!dl)
 436                 return NULL;
 437
 438         return device_str_attr_create(name, nl, nonce, desc, dl);
 439 }
 440
 441 static struct attribute *
 442 event_to_long_desc_attr(struct hv_24x7_event_data *event, int nonce)
 443 {
 444         int nl, dl;
 445         char *name = event_name(event, &nl);
 446         char *desc = event_long_desc(event, &dl);
 447
 448         /* If there isn't a description, don't create the sysfs file */
 449         if (!dl)
 450                 return NULL;
 451
 452         return device_str_attr_create(name, nl, nonce, desc, dl);
 453 }
 454
 455 static ssize_t event_data_to_attrs(unsigned ix, struct attribute **attrs,
 456                                    struct hv_24x7_event_data *event, int nonce)
 457 {
 458         unsigned i;
 459
 460         switch (event->domain) {
 461         case HV_PERF_DOMAIN_PHYS_CHIP:
 462                 *attrs = event_to_attr(ix, event, event->domain, nonce);
 463                 return 1;
 464         case HV_PERF_DOMAIN_PHYS_CORE:
 465                 for (i = 0; i < ARRAY_SIZE(core_domains); i++) {
 466                         attrs[i] = event_to_attr(ix, event, core_domains[i],
 467                                                 nonce);
 468                         if (!attrs[i]) {
 469                                 pr_warn("catalog event %u: individual attr %u "
 470                                         "creation failure\n", ix, i);
 471                                 for (; i; i--)
 472                                         device_str_attr_destroy(attrs[i - 1]);
 473                                 return -1;
 474                         }
 475                 }
 476                 return i;
 477         default:
 478                 pr_warn("catalog event %u: domain %u is not allowed in the "
 479                                 "catalog\n", ix, event->domain);
 480                 return -1;
 481         }
 482 }
 483
 484 static size_t event_to_attr_ct(struct hv_24x7_event_data *event)
 485 {
 486         switch (event->domain) {
 487         case HV_PERF_DOMAIN_PHYS_CHIP:
 488                 return 1;
 489         case HV_PERF_DOMAIN_PHYS_CORE:
 490                 return ARRAY_SIZE(core_domains);
 491         default:
 492                 return 0;
 493         }
 494 }
 495
 496 static unsigned long vmalloc_to_phys(void *v)
 497 {
 498         struct page *p = vmalloc_to_page(v);
 499
 500         BUG_ON(!p);
 501         return page_to_phys(p) + offset_in_page(v);
 502 }
 503
 504 /* */
 505 struct event_uniq {
 506         struct rb_node node;
 507         const char *name;
 508         int nl;
 509         unsigned ct;
 510         unsigned domain;
 511 };
 512
 513 static int memord(const void *d1, size_t s1, const void *d2, size_t s2)
 514 {
 515         if (s1 < s2)
 516                 return 1;
 517         if (s2 > s1)
 518                 return -1;
 519
 520         return memcmp(d1, d2, s1);
 521 }
 522
 523 static int ev_uniq_ord(const void *v1, size_t s1, unsigned d1, const void *v2,
 524                        size_t s2, unsigned d2)
 525 {
 526         int r = memord(v1, s1, v2, s2);
 527
 528         if (r)
 529                 return r;
 530         if (d1 > d2)
 531                 return 1;
 532         if (d2 > d1)
 533                 return -1;
 534         return 0;
 535 }
 536
 537 static int event_uniq_add(struct rb_root *root, const char *name, int nl,
 538                           unsigned domain)
 539 {
 540         struct rb_node **new = &(root->rb_node), *parent = NULL;
 541         struct event_uniq *data;
 542
 543         /* Figure out where to put new node */
 544         while (*new) {
 545                 struct event_uniq *it;
 546                 int result;
 547
 548                 it = container_of(*new, struct event_uniq, node);
 549                 result = ev_uniq_ord(name, nl, domain, it->name, it->nl,
 550                                         it->domain);
 551
 552                 parent = *new;
 553                 if (result < 0)
 554                         new = &((*new)->rb_left);
 555                 else if (result > 0)
 556                         new = &((*new)->rb_right);
 557                 else {
 558                         it->ct++;
 559                         pr_info("found a duplicate event %.*s, ct=%u\n", nl,
 560                                                 name, it->ct);
 561                         return it->ct;
 562                 }
 563         }
 564
 565         data = kmalloc(sizeof(*data), GFP_KERNEL);
 566         if (!data)
 567                 return -ENOMEM;
 568
 569         *data = (struct event_uniq) {
 570                 .name = name,
 571                 .nl = nl,
 572                 .ct = 0,
 573                 .domain = domain,
 574         };
 575
 576         /* Add new node and rebalance tree. */
 577         rb_link_node(&data->node, parent, new);
 578         rb_insert_color(&data->node, root);
 579
 580         /* data->ct */
 581         return 0;
 582 }
 583
 584 static void event_uniq_destroy(struct rb_root *root)
 585 {
 586         /*
 587          * the strings we point to are in the giant block of memory filled by
 588          * the catalog, and are freed separately.
 589          */
 590         struct event_uniq *pos, *n;
 591
 592         rbtree_postorder_for_each_entry_safe(pos, n, root, node)
 593                 kfree(pos);
 594 }
 595
 596
 597 /*
 598  * ensure the event structure's sizes are self consistent and don't cause us to
 599  * read outside of the event
 600  *
 601  * On success, return the event length in bytes.
 602  * Otherwise, return -1 (and print as appropriate).
 603  */
 604 static ssize_t catalog_event_len_validate(struct hv_24x7_event_data *event,
 605                                           size_t event_idx,
 606                                           size_t event_data_bytes,
 607                                           size_t event_entry_count,
 608                                           size_t offset, void *end)
 609 {
 610         ssize_t ev_len;
 611         void *ev_end, *calc_ev_end;
 612
 613         if (offset >= event_data_bytes)
 614                 return -1;
 615
 616         if (event_idx >= event_entry_count) {
 617                 pr_devel("catalog event data has %zu bytes of padding after last event\n",
 618                                 event_data_bytes - offset);
 619                 return -1;
 620         }
 621
 622         if (!event_fixed_portion_is_within(event, end)) {
 623                 pr_warn("event %zu fixed portion is not within range\n",
 624                                 event_idx);
 625                 return -1;
 626         }
 627
 628         ev_len = be16_to_cpu(event->length);
 629
 630         if (ev_len % 16)
 631                 pr_info("event %zu has length %zu not divisible by 16: event=%pK\n",
 632                                 event_idx, ev_len, event);
 633
 634         ev_end = (__u8 *)event + ev_len;
 635         if (ev_end > end) {
 636                 pr_warn("event %zu has .length=%zu, ends after buffer end: ev_end=%pK > end=%pK, offset=%zu\n",
 637                                 event_idx, ev_len, ev_end, end,
 638                                 offset);
 639                 return -1;
 640         }
 641
 642         calc_ev_end = event_end(event, end);
 643         if (!calc_ev_end) {
 644                 pr_warn("event %zu has a calculated length which exceeds buffer length %zu: event=%pK end=%pK, offset=%zu\n",
 645                         event_idx, event_data_bytes, event, end,
 646                         offset);
 647                 return -1;
 648         }
 649
 650         if (calc_ev_end > ev_end) {
 651                 pr_warn("event %zu exceeds it's own length: event=%pK, end=%pK, offset=%zu, calc_ev_end=%pK\n",
 652                         event_idx, event, ev_end, offset, calc_ev_end);
 653                 return -1;
 654         }
 655
 656         return ev_len;
 657 }
 658
 659 #define MAX_4K (SIZE_MAX / 4096)
 660
 661 static int create_events_from_catalog(struct attribute ***events_,
 662                                       struct attribute ***event_descs_,
 663                                       struct attribute ***event_long_descs_)
 664 {
 665         unsigned long hret;
 666         size_t catalog_len, catalog_page_len, event_entry_count,
 667                event_data_len, event_data_offs,
 668                event_data_bytes, junk_events, event_idx, event_attr_ct, i,
 669                attr_max, event_idx_last, desc_ct, long_desc_ct;
 670         ssize_t ct, ev_len;
 671         uint32_t catalog_version_num;
 672         struct attribute **events, **event_descs, **event_long_descs;
 673         struct hv_24x7_catalog_page_0 *page_0 =
 674                 kmem_cache_alloc(hv_page_cache, GFP_KERNEL);
 675         void *page = page_0;
 676         void *event_data, *end;
 677         struct hv_24x7_event_data *event;
 678         struct rb_root ev_uniq = RB_ROOT;
 679         int ret = 0;
 680
 681         if (!page) {
 682                 ret = -ENOMEM;
 683                 goto e_out;
 684         }
 685
 686         hret = h_get_24x7_catalog_page(page, 0, 0);
 687         if (hret) {
 688                 ret = -EIO;
 689                 goto e_free;
 690         }
 691
 692         catalog_version_num = be64_to_cpu(page_0->version);
 693         catalog_page_len = be32_to_cpu(page_0->length);
 694
 695         if (MAX_4K < catalog_page_len) {
 696                 pr_err("invalid page count: %zu\n", catalog_page_len);
 697                 ret = -EIO;
 698                 goto e_free;
 699         }
 700
 701         catalog_len = catalog_page_len * 4096;
 702
 703         event_entry_count = be16_to_cpu(page_0->event_entry_count);
 704         event_data_offs   = be16_to_cpu(page_0->event_data_offs);
 705         event_data_len    = be16_to_cpu(page_0->event_data_len);
 706
 707         pr_devel("cv %zu cl %zu eec %zu edo %zu edl %zu\n",
 708                         (size_t)catalog_version_num, catalog_len,
 709                         event_entry_count, event_data_offs, event_data_len);
 710
 711         if ((MAX_4K < event_data_len)
 712                         || (MAX_4K < event_data_offs)
 713                         || (MAX_4K - event_data_offs < event_data_len)) {
 714                 pr_err("invalid event data offs %zu and/or len %zu\n",
 715                                 event_data_offs, event_data_len);
 716                 ret = -EIO;
 717                 goto e_free;
 718         }
 719
 720         if ((event_data_offs + event_data_len) > catalog_page_len) {
 721                 pr_err("event data %zu-%zu does not fit inside catalog 0-%zu\n",
 722                                 event_data_offs,
 723                                 event_data_offs + event_data_len,
 724                                 catalog_page_len);
 725                 ret = -EIO;
 726                 goto e_free;
 727         }
 728
 729         if (SIZE_MAX / MAX_EVENTS_PER_EVENT_DATA - 1 < event_entry_count) {
 730                 pr_err("event_entry_count %zu is invalid\n",
 731                                 event_entry_count);
 732                 ret = -EIO;
 733                 goto e_free;
 734         }
 735
 736         event_data_bytes = event_data_len * 4096;
 737
 738         /*
 739          * event data can span several pages, events can cross between these
 740          * pages. Use vmalloc to make this easier.
 741          */
 742         event_data = vmalloc(event_data_bytes);
 743         if (!event_data) {
 744                 pr_err("could not allocate event data\n");
 745                 ret = -ENOMEM;
 746                 goto e_free;
 747         }
 748
 749         end = event_data + event_data_bytes;
 750
 751         /*
 752          * using vmalloc_to_phys() like this only works if PAGE_SIZE is
 753          * divisible by 4096
 754          */
 755         BUILD_BUG_ON(PAGE_SIZE % 4096);
 756
 757         for (i = 0; i < event_data_len; i++) {
 758                 hret = h_get_24x7_catalog_page_(
 759                                 vmalloc_to_phys(event_data + i * 4096),
 760                                 catalog_version_num,
 761                                 i + event_data_offs);
 762                 if (hret) {
 763                         pr_err("failed to get event data in page %zu\n",
 764                                         i + event_data_offs);
 765                         ret = -EIO;
 766                         goto e_event_data;
 767                 }
 768         }
 769
 770         /*
 771          * scan the catalog to determine the number of attributes we need, and
 772          * verify it at the same time.
 773          */
 774         for (junk_events = 0, event = event_data, event_idx = 0, attr_max = 0;
 775              ;
 776              event_idx++, event = (void *)event + ev_len) {
 777                 size_t offset = (void *)event - (void *)event_data;
 778                 char *name;
 779                 int nl;
 780
 781                 ev_len = catalog_event_len_validate(event, event_idx,
 782                                                     event_data_bytes,
 783                                                     event_entry_count,
 784                                                     offset, end);
 785                 if (ev_len < 0)
 786                         break;
 787
 788                 name = event_name(event, &nl);
 789
 790                 if (event->event_group_record_len == 0) {
 791                         pr_devel("invalid event %zu (%.*s): group_record_len == 0, skipping\n",
 792                                         event_idx, nl, name);
 793                         junk_events++;
 794                         continue;
 795                 }
 796
 797                 if (!catalog_entry_domain_is_valid(event->domain)) {
 798                         pr_info("event %zu (%.*s) has invalid domain %d\n",
 799                                         event_idx, nl, name, event->domain);
 800                         junk_events++;
 801                         continue;
 802                 }
 803
 804                 attr_max += event_to_attr_ct(event);
 805         }
 806
 807         event_idx_last = event_idx;
 808         if (event_idx_last != event_entry_count)
 809                 pr_warn("event buffer ended before listed # of events were parsed (got %zu, wanted %zu, junk %zu)\n",
 810                                 event_idx_last, event_entry_count, junk_events);
 811
 812         events = kmalloc_array(attr_max + 1, sizeof(*events), GFP_KERNEL);
 813         if (!events) {
 814                 ret = -ENOMEM;
 815                 goto e_event_data;
 816         }
 817
 818         event_descs = kmalloc_array(event_idx + 1, sizeof(*event_descs),
 819                                 GFP_KERNEL);
 820         if (!event_descs) {
 821                 ret = -ENOMEM;
 822                 goto e_event_attrs;
 823         }
 824
 825         event_long_descs = kmalloc_array(event_idx + 1,
 826                         sizeof(*event_long_descs), GFP_KERNEL);
 827         if (!event_long_descs) {
 828                 ret = -ENOMEM;
 829                 goto e_event_descs;
 830         }
 831
 832         /* Iterate over the catalog filling in the attribute vector */
 833         for (junk_events = 0, event_attr_ct = 0, desc_ct = 0, long_desc_ct = 0,
 834                                 event = event_data, event_idx = 0;
 835                         event_idx < event_idx_last;
 836                         event_idx++, ev_len = be16_to_cpu(event->length),
 837                                 event = (void *)event + ev_len) {
 838                 char *name;
 839                 int nl;
 840                 int nonce;
 841                 /*
 842                  * these are the only "bad" events that are intermixed and that
 843                  * we can ignore without issue. make sure to skip them here
 844                  */
 845                 if (event->event_group_record_len == 0)
 846                         continue;
 847                 if (!catalog_entry_domain_is_valid(event->domain))
 848                         continue;
 849
 850                 name  = event_name(event, &nl);
 851                 nonce = event_uniq_add(&ev_uniq, name, nl, event->domain);
 852                 ct    = event_data_to_attrs(event_idx, events + event_attr_ct,
 853                                             event, nonce);
 854                 if (ct <= 0) {
 855                         pr_warn("event %zu (%.*s) creation failure, skipping\n",
 856                                 event_idx, nl, name);
 857                         junk_events++;
 858                 } else {
 859                         event_attr_ct += ct;
 860                         event_descs[desc_ct] = event_to_desc_attr(event, nonce);
 861                         if (event_descs[desc_ct])
 862                                 desc_ct++;
 863                         event_long_descs[long_desc_ct] =
 864                                         event_to_long_desc_attr(event, nonce);
 865                         if (event_long_descs[long_desc_ct])
 866                                 long_desc_ct++;
 867                 }
 868         }
 869
 870         pr_info("read %zu catalog entries, created %zu event attrs (%zu failures), %zu descs\n",
 871                         event_idx, event_attr_ct, junk_events, desc_ct);
 872
 873         events[event_attr_ct] = NULL;
 874         event_descs[desc_ct] = NULL;
 875         event_long_descs[long_desc_ct] = NULL;
 876
 877         event_uniq_destroy(&ev_uniq);
 878         vfree(event_data);
 879         kmem_cache_free(hv_page_cache, page);
 880
 881         *events_ = events;
 882         *event_descs_ = event_descs;
 883         *event_long_descs_ = event_long_descs;
 884         return 0;
 885
 886 e_event_descs:
 887         kfree(event_descs);
 888 e_event_attrs:
 889         kfree(events);
 890 e_event_data:
 891         vfree(event_data);
 892 e_free:
 893         kmem_cache_free(hv_page_cache, page);
 894 e_out:
 895         *events_ = NULL;
 896         *event_descs_ = NULL;
 897         *event_long_descs_ = NULL;
 898         return ret;
 899 }
 900
 901 static ssize_t catalog_read(struct file *filp, struct kobject *kobj,
 902                             struct bin_attribute *bin_attr, char *buf,
 903                             loff_t offset, size_t count)
 904 {
 905         unsigned long hret;
 906         ssize_t ret = 0;
 907         size_t catalog_len = 0, catalog_page_len = 0;
 908         loff_t page_offset = 0;
 909         loff_t offset_in_page;
 910         size_t copy_len;
 911         uint64_t catalog_version_num = 0;
 912         void *page = kmem_cache_alloc(hv_page_cache, GFP_USER);
 913         struct hv_24x7_catalog_page_0 *page_0 = page;
 914
 915         if (!page)
 916                 return -ENOMEM;
 917
 918         hret = h_get_24x7_catalog_page(page, 0, 0);
 919         if (hret) {
 920                 ret = -EIO;
 921                 goto e_free;
 922         }
 923
 924         catalog_version_num = be64_to_cpu(page_0->version);
 925         catalog_page_len = be32_to_cpu(page_0->length);
 926         catalog_len = catalog_page_len * 4096;
 927
 928         page_offset = offset / 4096;
 929         offset_in_page = offset % 4096;
 930
 931         if (page_offset >= catalog_page_len)
 932                 goto e_free;
 933
 934         if (page_offset != 0) {
 935                 hret = h_get_24x7_catalog_page(page, catalog_version_num,
 936                                                page_offset);
 937                 if (hret) {
 938                         ret = -EIO;
 939                         goto e_free;
 940                 }
 941         }
 942
 943         copy_len = 4096 - offset_in_page;
 944         if (copy_len > count)
 945                 copy_len = count;
 946
 947         memcpy(buf, page+offset_in_page, copy_len);
 948         ret = copy_len;
 949
 950 e_free:
 951         if (hret)
 952                 pr_err("h_get_24x7_catalog_page(ver=%lld, page=%lld) failed:"
 953                        " rc=%ld\n",
 954                        catalog_version_num, page_offset, hret);
 955         kmem_cache_free(hv_page_cache, page);
 956
 957         pr_devel("catalog_read: offset=%lld(%lld) count=%zu "
 958                         "catalog_len=%zu(%zu) => %zd\n", offset, page_offset,
 959                         count, catalog_len, catalog_page_len, ret);
 960
 961         return ret;
 962 }
 963
 964 #define PAGE_0_ATTR(_name, _fmt, _expr)                         \
 965 static ssize_t _name##_show(struct device *dev,                 \
 966                             struct device_attribute *dev_attr,  \
 967                             char *buf)                          \
 968 {                                                               \
 969         unsigned long hret;                                     \
 970         ssize_t ret = 0;                                        \
 971         void *page = kmem_cache_alloc(hv_page_cache, GFP_USER); \
 972         struct hv_24x7_catalog_page_0 *page_0 = page;           \
 973         if (!page)                                              \
 974                 return -ENOMEM;                                 \
 975         hret = h_get_24x7_catalog_page(page, 0, 0);             \
 976         if (hret) {                                             \
 977                 ret = -EIO;                                     \
 978                 goto e_free;                                    \
 979         }                                                       \
 980         ret = sprintf(buf, _fmt, _expr);                        \
 981 e_free:                                                         \
 982         kmem_cache_free(hv_page_cache, page);                   \
 983         return ret;                                             \
 984 }                                                               \
 985 static DEVICE_ATTR_RO(_name)
 986
 987 PAGE_0_ATTR(catalog_version, "%lld\n",
 988                 (unsigned long long)be64_to_cpu(page_0->version));
 989 PAGE_0_ATTR(catalog_len, "%lld\n",
 990                 (unsigned long long)be32_to_cpu(page_0->length) * 4096);
 991 static BIN_ATTR_RO(catalog, 0/* real length varies */);
 992
 993 static struct bin_attribute *if_bin_attrs[] = {
 994         &bin_attr_catalog,
 995         NULL,
 996 };
 997
 998 static struct attribute *if_attrs[] = {
 999         &dev_attr_catalog_len.attr,
1000         &dev_attr_catalog_version.attr,
1001         NULL,
1002 };
1003
1004 static struct attribute_group if_group = {
1005         .name = "interface",
1006         .bin_attrs = if_bin_attrs,
1007         .attrs = if_attrs,
1008 };
1009
1010 static const struct attribute_group *attr_groups[] = {
1011         &format_group,
1012         &event_group,
1013         &event_desc_group,
1014         &event_long_desc_group,
1015         &if_group,
1016         NULL,
1017 };
1018
1019 static void log_24x7_hcall(struct hv_24x7_request_buffer *request_buffer,
1020                            struct hv_24x7_data_result_buffer *result_buffer,
1021                            unsigned long ret)
1022 {
1023         struct hv_24x7_request *req;
1024
1025         req = &request_buffer->requests[0];
1026         pr_notice_ratelimited("hcall failed: [%d %#x %#x %d] => "
1027                         "ret 0x%lx (%ld) detail=0x%x failing ix=%x\n",
1028                         req->performance_domain, req->data_offset,
1029                         req->starting_ix, req->starting_lpar_ix, ret, ret,
1030                         result_buffer->detailed_rc,
1031                         result_buffer->failing_request_ix);
1032 }
1033
1034 /*
1035  * Start the process for a new H_GET_24x7_DATA hcall.
1036  */
1037 static void init_24x7_request(struct hv_24x7_request_buffer *request_buffer,
1038                               struct hv_24x7_data_result_buffer *result_buffer)
1039 {
1040
1041         memset(request_buffer, 0, 4096);
1042         memset(result_buffer, 0, 4096);
1043
1044         request_buffer->interface_version = HV_24X7_IF_VERSION_CURRENT;
1045         /* memset above set request_buffer->num_requests to 0 */
1046 }
1047
1048 /*
1049  * Commit (i.e perform) the H_GET_24x7_DATA hcall using the data collected
1050  * by 'init_24x7_request()' and 'add_event_to_24x7_request()'.
1051  */
1052 static int make_24x7_request(struct hv_24x7_request_buffer *request_buffer,
1053                              struct hv_24x7_data_result_buffer *result_buffer)
1054 {
1055         unsigned long ret;
1056
1057         /*
1058          * NOTE: Due to variable number of array elements in request and
1059          *       result buffer(s), sizeof() is not reliable. Use the actual
1060          *       allocated buffer size, H24x7_DATA_BUFFER_SIZE.
1061          */
1062         ret = plpar_hcall_norets(H_GET_24X7_DATA,
1063                         virt_to_phys(request_buffer), H24x7_DATA_BUFFER_SIZE,
1064                         virt_to_phys(result_buffer),  H24x7_DATA_BUFFER_SIZE);
1065
1066         if (ret)
1067                 log_24x7_hcall(request_buffer, result_buffer, ret);
1068
1069         return ret;
1070 }
1071
1072 /*
1073  * Add the given @event to the next slot in the 24x7 request_buffer.
1074  *
1075  * Note that H_GET_24X7_DATA hcall allows reading several counters'
1076  * values in a single HCALL. We expect the caller to add events to the
1077  * request buffer one by one, make the HCALL and process the results.
1078  */
1079 static int add_event_to_24x7_request(struct perf_event *event,
1080                                 struct hv_24x7_request_buffer *request_buffer)
1081 {
1082         u16 idx;
1083         int i;
1084         struct hv_24x7_request *req;
1085
1086         if (request_buffer->num_requests > 254) {
1087                 pr_devel("Too many requests for 24x7 HCALL %d\n",
1088                                 request_buffer->num_requests);
1089                 return -EINVAL;
1090         }
1091
1092         if (is_physical_domain(event_get_domain(event)))
1093                 idx = event_get_core(event);
1094         else
1095                 idx = event_get_vcpu(event);
1096
1097         i = request_buffer->num_requests++;
1098         req = &request_buffer->requests[i];
1099
1100         req->performance_domain = event_get_domain(event);
1101         req->data_size = cpu_to_be16(8);
1102         req->data_offset = cpu_to_be32(event_get_offset(event));
1103         req->starting_lpar_ix = cpu_to_be16(event_get_lpar(event)),
1104         req->max_num_lpars = cpu_to_be16(1);
1105         req->starting_ix = cpu_to_be16(idx);
1106         req->max_ix = cpu_to_be16(1);
1107
1108         return 0;
1109 }
1110
1111 static unsigned long single_24x7_request(struct perf_event *event, u64 *count)
1112 {
1113         unsigned long ret;
1114         struct hv_24x7_request_buffer *request_buffer;
1115         struct hv_24x7_data_result_buffer *result_buffer;
1116
1117         BUILD_BUG_ON(sizeof(*request_buffer) > 4096);
1118         BUILD_BUG_ON(sizeof(*result_buffer) > 4096);
1119
1120         request_buffer = (void *)get_cpu_var(hv_24x7_reqb);
1121         result_buffer = (void *)get_cpu_var(hv_24x7_resb);
1122
1123         init_24x7_request(request_buffer, result_buffer);
1124
1125         ret = add_event_to_24x7_request(event, request_buffer);
1126         if (ret)
1127                 goto out;
1128
1129         ret = make_24x7_request(request_buffer, result_buffer);
1130         if (ret) {
1131                 log_24x7_hcall(request_buffer, result_buffer, ret);
1132                 goto out;
1133         }
1134
1135         /* process result from hcall */
1136         *count = be64_to_cpu(result_buffer->results[0].elements[0].element_data[0]);
1137
1138 out:
1139         put_cpu_var(hv_24x7_reqb);
1140         put_cpu_var(hv_24x7_resb);
1141         return ret;
1142 }
1143
1144
1145 static int h_24x7_event_init(struct perf_event *event)
1146 {
1147         struct hv_perf_caps caps;
1148         unsigned domain;
1149         unsigned long hret;
1150         u64 ct;
1151
1152         /* Not our event */
1153         if (event->attr.type != event->pmu->type)
1154                 return -ENOENT;
1155
1156         /* Unused areas must be 0 */
1157         if (event_get_reserved1(event) ||
1158             event_get_reserved2(event) ||
1159             event_get_reserved3(event)) {
1160                 pr_devel("reserved set when forbidden 0x%llx(0x%llx) 0x%llx(0x%llx) 0x%llx(0x%llx)\n",
1161                                 event->attr.config,
1162                                 event_get_reserved1(event),
1163                                 event->attr.config1,
1164                                 event_get_reserved2(event),
1165                                 event->attr.config2,
1166                                 event_get_reserved3(event));
1167                 return -EINVAL;
1168         }
1169
1170         /* unsupported modes and filters */
1171         if (event->attr.exclude_user   ||
1172             event->attr.exclude_kernel ||
1173             event->attr.exclude_hv     ||
1174             event->attr.exclude_idle   ||
1175             event->attr.exclude_host   ||
1176             event->attr.exclude_guest)
1177                 return -EINVAL;
1178
1179         /* no branch sampling */
1180         if (has_branch_stack(event))
1181                 return -EOPNOTSUPP;
1182
1183         /* offset must be 8 byte aligned */
1184         if (event_get_offset(event) % 8) {
1185                 pr_devel("bad alignment\n");
1186                 return -EINVAL;
1187         }
1188
1189         /* Domains above 6 are invalid */
1190         domain = event_get_domain(event);
1191         if (domain > 6) {
1192                 pr_devel("invalid domain %d\n", domain);
1193                 return -EINVAL;
1194         }
1195
1196         hret = hv_perf_caps_get(&caps);
1197         if (hret) {
1198                 pr_devel("could not get capabilities: rc=%ld\n", hret);
1199                 return -EIO;
1200         }
1201
1202         /* Physical domains & other lpars require extra capabilities */
1203         if (!caps.collect_privileged && (is_physical_domain(domain) ||
1204                 (event_get_lpar(event) != event_get_lpar_max()))) {
1205                 pr_devel("hv permissions disallow: is_physical_domain:%d, lpar=0x%llx\n",
1206                                 is_physical_domain(domain),
1207                                 event_get_lpar(event));
1208                 return -EACCES;
1209         }
1210
1211         /* see if the event complains */
1212         if (single_24x7_request(event, &ct)) {
1213                 pr_devel("test hcall failed\n");
1214                 return -EIO;
1215         }
1216
1217         return 0;
1218 }
1219
1220 static u64 h_24x7_get_value(struct perf_event *event)
1221 {
1222         unsigned long ret;
1223         u64 ct;
1224         ret = single_24x7_request(event, &ct);
1225         if (ret)
1226                 /* We checked this in event init, shouldn't fail here... */
1227                 return 0;
1228
1229         return ct;
1230 }
1231
1232 static void update_event_count(struct perf_event *event, u64 now)
1233 {
1234         s64 prev;
1235
1236         prev = local64_xchg(&event->hw.prev_count, now);
1237         local64_add(now - prev, &event->count);
1238 }
1239
1240 static void h_24x7_event_read(struct perf_event *event)
1241 {
1242         u64 now;
1243         struct hv_24x7_request_buffer *request_buffer;
1244         struct hv_24x7_hw *h24x7hw;
1245         int txn_flags;
1246
1247         txn_flags = __this_cpu_read(hv_24x7_txn_flags);
1248
1249         /*
1250          * If in a READ transaction, add this counter to the list of
1251          * counters to read during the next HCALL (i.e commit_txn()).
1252          * If not in a READ transaction, go ahead and make the HCALL
1253          * to read this counter by itself.
1254          */
1255
1256         if (txn_flags & PERF_PMU_TXN_READ) {
1257                 int i;
1258                 int ret;
1259
1260                 if (__this_cpu_read(hv_24x7_txn_err))
1261                         return;
1262
1263                 request_buffer = (void *)get_cpu_var(hv_24x7_reqb);
1264
1265                 ret = add_event_to_24x7_request(event, request_buffer);
1266                 if (ret) {
1267                         __this_cpu_write(hv_24x7_txn_err, ret);
1268                 } else {
1269                         /*
1270                          * Assoicate the event with the HCALL request index,
1271                          * so ->commit_txn() can quickly find/update count.
1272                          */
1273                         i = request_buffer->num_requests - 1;
1274
1275                         h24x7hw = &get_cpu_var(hv_24x7_hw);
1276                         h24x7hw->events[i] = event;
1277                         put_cpu_var(h24x7hw);
1278                 }
1279
1280                 put_cpu_var(hv_24x7_reqb);
1281         } else {
1282                 now = h_24x7_get_value(event);
1283                 update_event_count(event, now);
1284         }
1285 }
1286
1287 static void h_24x7_event_start(struct perf_event *event, int flags)
1288 {
1289         if (flags & PERF_EF_RELOAD)
1290                 local64_set(&event->hw.prev_count, h_24x7_get_value(event));
1291 }
1292
1293 static void h_24x7_event_stop(struct perf_event *event, int flags)
1294 {
1295         h_24x7_event_read(event);
1296 }
1297
1298 static int h_24x7_event_add(struct perf_event *event, int flags)
1299 {
1300         if (flags & PERF_EF_START)
1301                 h_24x7_event_start(event, flags);
1302
1303         return 0;
1304 }
1305
1306 /*
1307  * 24x7 counters only support READ transactions. They are
1308  * always counting and dont need/support ADD transactions.
1309  * Cache the flags, but otherwise ignore transactions that
1310  * are not PERF_PMU_TXN_READ.
1311  */
1312 static void h_24x7_event_start_txn(struct pmu *pmu, unsigned int flags)
1313 {
1314         struct hv_24x7_request_buffer *request_buffer;
1315         struct hv_24x7_data_result_buffer *result_buffer;
1316
1317         /* We should not be called if we are already in a txn */
1318         WARN_ON_ONCE(__this_cpu_read(hv_24x7_txn_flags));
1319
1320         __this_cpu_write(hv_24x7_txn_flags, flags);
1321         if (flags & ~PERF_PMU_TXN_READ)
1322                 return;
1323
1324         request_buffer = (void *)get_cpu_var(hv_24x7_reqb);
1325         result_buffer = (void *)get_cpu_var(hv_24x7_resb);
1326
1327         init_24x7_request(request_buffer, result_buffer);
1328
1329         put_cpu_var(hv_24x7_resb);
1330         put_cpu_var(hv_24x7_reqb);
1331 }
1332
1333 /*
1334  * Clean up transaction state.
1335  *
1336  * NOTE: Ignore state of request and result buffers for now.
1337  *       We will initialize them during the next read/txn.
1338  */
1339 static void reset_txn(void)
1340 {
1341         __this_cpu_write(hv_24x7_txn_flags, 0);
1342         __this_cpu_write(hv_24x7_txn_err, 0);
1343 }
1344
1345 /*
1346  * 24x7 counters only support READ transactions. They are always counting
1347  * and dont need/support ADD transactions. Clear ->txn_flags but otherwise
1348  * ignore transactions that are not of type PERF_PMU_TXN_READ.
1349  *
1350  * For READ transactions, submit all pending 24x7 requests (i.e requests
1351  * that were queued by h_24x7_event_read()), to the hypervisor and update
1352  * the event counts.
1353  */
1354 static int h_24x7_event_commit_txn(struct pmu *pmu)
1355 {
1356         struct hv_24x7_request_buffer *request_buffer;
1357         struct hv_24x7_data_result_buffer *result_buffer;
1358         struct hv_24x7_result *resb;
1359         struct perf_event *event;
1360         u64 count;
1361         int i, ret, txn_flags;
1362         struct hv_24x7_hw *h24x7hw;
1363
1364         txn_flags = __this_cpu_read(hv_24x7_txn_flags);
1365         WARN_ON_ONCE(!txn_flags);
1366
1367         ret = 0;
1368         if (txn_flags & ~PERF_PMU_TXN_READ)
1369                 goto out;
1370
1371         ret = __this_cpu_read(hv_24x7_txn_err);
1372         if (ret)
1373                 goto out;
1374
1375         request_buffer = (void *)get_cpu_var(hv_24x7_reqb);
1376         result_buffer = (void *)get_cpu_var(hv_24x7_resb);
1377
1378         ret = make_24x7_request(request_buffer, result_buffer);
1379         if (ret) {
1380                 log_24x7_hcall(request_buffer, result_buffer, ret);
1381                 goto put_reqb;
1382         }
1383
1384         h24x7hw = &get_cpu_var(hv_24x7_hw);
1385
1386         /* Update event counts from hcall */
1387         for (i = 0; i < request_buffer->num_requests; i++) {
1388                 resb = &result_buffer->results[i];
1389                 count = be64_to_cpu(resb->elements[0].element_data[0]);
1390                 event = h24x7hw->events[i];
1391                 h24x7hw->events[i] = NULL;
1392                 update_event_count(event, count);
1393         }
1394
1395         put_cpu_var(hv_24x7_hw);
1396
1397 put_reqb:
1398         put_cpu_var(hv_24x7_resb);
1399         put_cpu_var(hv_24x7_reqb);
1400 out:
1401         reset_txn();
1402         return ret;
1403 }
1404
1405 /*
1406  * 24x7 counters only support READ transactions. They are always counting
1407  * and dont need/support ADD transactions. However, regardless of type
1408  * of transaction, all we need to do is cleanup, so we don't have to check
1409  * the type of transaction.
1410  */
1411 static void h_24x7_event_cancel_txn(struct pmu *pmu)
1412 {
1413         WARN_ON_ONCE(!__this_cpu_read(hv_24x7_txn_flags));
1414         reset_txn();
1415 }
1416
1417 static struct pmu h_24x7_pmu = {
1418         .task_ctx_nr = perf_invalid_context,
1419
1420         .name = "hv_24x7",
1421         .attr_groups = attr_groups,
1422         .event_init  = h_24x7_event_init,
1423         .add         = h_24x7_event_add,
1424         .del         = h_24x7_event_stop,
1425         .start       = h_24x7_event_start,
1426         .stop        = h_24x7_event_stop,
1427         .read        = h_24x7_event_read,
1428         .start_txn   = h_24x7_event_start_txn,
1429         .commit_txn  = h_24x7_event_commit_txn,
1430         .cancel_txn  = h_24x7_event_cancel_txn,
1431 };
1432
1433 static int hv_24x7_init(void)
1434 {
1435         int r;
1436         unsigned long hret;
1437         struct hv_perf_caps caps;
1438
1439         if (!firmware_has_feature(FW_FEATURE_LPAR)) {
1440                 pr_debug("not a virtualized system, not enabling\n");
1441                 return -ENODEV;
1442         }
1443
1444         hret = hv_perf_caps_get(&caps);
1445         if (hret) {
1446                 pr_debug("could not obtain capabilities, not enabling, rc=%ld\n",
1447                                 hret);
1448                 return -ENODEV;
1449         }
1450
1451         hv_page_cache = kmem_cache_create("hv-page-4096", 4096, 4096, 0, NULL);
1452         if (!hv_page_cache)
1453                 return -ENOMEM;
1454
1455         /* sampling not supported */
1456         h_24x7_pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
1457
1458         r = create_events_from_catalog(&event_group.attrs,
1459                                    &event_desc_group.attrs,
1460                                    &event_long_desc_group.attrs);
1461
1462         if (r)
1463                 return r;
1464
1465         r = perf_pmu_register(&h_24x7_pmu, h_24x7_pmu.name, -1);
1466         if (r)
1467                 return r;
1468
1469         return 0;
1470 }
1471
1472 device_initcall(hv_24x7_init);