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