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         attr->var = str;
 324         attr->attr.attr.name = name;
 325         attr->attr.attr.mode = 0444;
 326         attr->attr.show = device_show_string;
 327
 328         return &attr->attr.attr;
 329 }
 330
 331 static struct attribute *device_str_attr_create(char *name, int name_max,
 332                                                 int name_nonce,
 333                                                 char *str, size_t str_max)
 334 {
 335         char *n;
 336         char *s = memdup_to_str(str, str_max, GFP_KERNEL);
 337         struct attribute *a;
 338
 339         if (!s)
 340                 return NULL;
 341
 342         if (!name_nonce)
 343                 n = kasprintf(GFP_KERNEL, "%.*s", name_max, name);
 344         else
 345                 n = kasprintf(GFP_KERNEL, "%.*s__%d", name_max, name,
 346                                         name_nonce);
 347         if (!n)
 348                 goto out_s;
 349
 350         a = device_str_attr_create_(n, s);
 351         if (!a)
 352                 goto out_n;
 353
 354         return a;
 355 out_n:
 356         kfree(n);
 357 out_s:
 358         kfree(s);
 359         return NULL;
 360 }
 361
 362 static void device_str_attr_destroy(struct attribute *attr)
 363 {
 364         struct dev_ext_attribute *d;
 365
 366         d = container_of(attr, struct dev_ext_attribute, attr.attr);
 367         kfree(d->var);
 368         kfree(d->attr.attr.name);
 369         kfree(d);
 370 }
 371
 372 static struct attribute *event_to_attr(unsigned ix,
 373                                        struct hv_24x7_event_data *event,
 374                                        unsigned domain,
 375                                        int nonce)
 376 {
 377         int event_name_len;
 378         char *ev_name, *a_ev_name, *val;
 379         const char *ev_suffix;
 380         struct attribute *attr;
 381
 382         if (!domain_is_valid(domain)) {
 383                 pr_warn("catalog event %u has invalid domain %u\n",
 384                                 ix, domain);
 385                 return NULL;
 386         }
 387
 388         val = event_fmt(event, domain);
 389         if (!val)
 390                 return NULL;
 391
 392         ev_suffix = event_domain_suffix(domain);
 393         ev_name = event_name(event, &event_name_len);
 394         if (!nonce)
 395                 a_ev_name = kasprintf(GFP_KERNEL, "%.*s%s",
 396                                 (int)event_name_len, ev_name, ev_suffix);
 397         else
 398                 a_ev_name = kasprintf(GFP_KERNEL, "%.*s%s__%d",
 399                                 (int)event_name_len, ev_name, ev_suffix, nonce);
 400
 401         if (!a_ev_name)
 402                 goto out_val;
 403
 404         attr = device_str_attr_create_(a_ev_name, val);
 405         if (!attr)
 406                 goto out_name;
 407
 408         return attr;
 409 out_name:
 410         kfree(a_ev_name);
 411 out_val:
 412         kfree(val);
 413         return NULL;
 414 }
 415
 416 static struct attribute *event_to_desc_attr(struct hv_24x7_event_data *event,
 417                                 int nonce)
 418 {
 419         int nl, dl;
 420         char *name = event_name(event, &nl);
 421         char *desc = event_desc(event, &dl);
 422
 423         /* If there isn't a description, don't create the sysfs file */
 424         if (!dl)
 425                 return NULL;
 426
 427         return device_str_attr_create(name, nl, nonce, desc, dl);
 428 }
 429
 430 static struct attribute *
 431 event_to_long_desc_attr(struct hv_24x7_event_data *event, int nonce)
 432 {
 433         int nl, dl;
 434         char *name = event_name(event, &nl);
 435         char *desc = event_long_desc(event, &dl);
 436
 437         /* If there isn't a description, don't create the sysfs file */
 438         if (!dl)
 439                 return NULL;
 440
 441         return device_str_attr_create(name, nl, nonce, desc, dl);
 442 }
 443
 444 static ssize_t event_data_to_attrs(unsigned ix, struct attribute **attrs,
 445                 struct hv_24x7_event_data *event, int nonce)
 446 {
 447         unsigned i;
 448
 449         switch (event->domain) {
 450         case HV_PERF_DOMAIN_PHYS_CHIP:
 451                 *attrs = event_to_attr(ix, event, event->domain, nonce);
 452                 return 1;
 453         case HV_PERF_DOMAIN_PHYS_CORE:
 454                 for (i = 0; i < ARRAY_SIZE(core_domains); i++) {
 455                         attrs[i] = event_to_attr(ix, event, core_domains[i],
 456                                                 nonce);
 457                         if (!attrs[i]) {
 458                                 pr_warn("catalog event %u: individual attr %u "
 459                                         "creation failure\n", ix, i);
 460                                 for (; i; i--)
 461                                         device_str_attr_destroy(attrs[i - 1]);
 462                                 return -1;
 463                         }
 464                 }
 465                 return i;
 466         default:
 467                 pr_warn("catalog event %u: domain %u is not allowed in the "
 468                                 "catalog\n", ix, event->domain);
 469                 return -1;
 470         }
 471 }
 472
 473 static size_t event_to_attr_ct(struct hv_24x7_event_data *event)
 474 {
 475         switch (event->domain) {
 476         case HV_PERF_DOMAIN_PHYS_CHIP:
 477                 return 1;
 478         case HV_PERF_DOMAIN_PHYS_CORE:
 479                 return ARRAY_SIZE(core_domains);
 480         default:
 481                 return 0;
 482         }
 483 }
 484
 485 static unsigned long vmalloc_to_phys(void *v)
 486 {
 487         struct page *p = vmalloc_to_page(v);
 488
 489         BUG_ON(!p);
 490         return page_to_phys(p) + offset_in_page(v);
 491 }
 492
 493 /* */
 494 struct event_uniq {
 495         struct rb_node node;
 496         const char *name;
 497         int nl;
 498         unsigned ct;
 499         unsigned domain;
 500 };
 501
 502 static int memord(const void *d1, size_t s1, const void *d2, size_t s2)
 503 {
 504         if (s1 < s2)
 505                 return 1;
 506         if (s2 > s1)
 507                 return -1;
 508
 509         return memcmp(d1, d2, s1);
 510 }
 511
 512 static int ev_uniq_ord(const void *v1, size_t s1, unsigned d1, const void *v2,
 513                                         size_t s2, unsigned d2)
 514 {
 515         int r = memord(v1, s1, v2, s2);
 516
 517         if (r)
 518                 return r;
 519         if (d1 > d2)
 520                 return 1;
 521         if (d2 > d1)
 522                 return -1;
 523         return 0;
 524 }
 525
 526 static int event_uniq_add(struct rb_root *root, const char *name, int nl,
 527                                 unsigned domain)
 528 {
 529         struct rb_node **new = &(root->rb_node), *parent = NULL;
 530         struct event_uniq *data;
 531
 532         /* Figure out where to put new node */
 533         while (*new) {
 534                 struct event_uniq *it;
 535                 int result;
 536
 537                 it = container_of(*new, struct event_uniq, node);
 538                 result = ev_uniq_ord(name, nl, domain, it->name, it->nl,
 539                                         it->domain);
 540
 541                 parent = *new;
 542                 if (result < 0)
 543                         new = &((*new)->rb_left);
 544                 else if (result > 0)
 545                         new = &((*new)->rb_right);
 546                 else {
 547                         it->ct++;
 548                         pr_info("found a duplicate event %.*s, ct=%u\n", nl,
 549                                                 name, it->ct);
 550                         return it->ct;
 551                 }
 552         }
 553
 554         data = kmalloc(sizeof(*data), GFP_KERNEL);
 555         if (!data)
 556                 return -ENOMEM;
 557
 558         *data = (struct event_uniq) {
 559                 .name = name,
 560                 .nl = nl,
 561                 .ct = 0,
 562                 .domain = domain,
 563         };
 564
 565         /* Add new node and rebalance tree. */
 566         rb_link_node(&data->node, parent, new);
 567         rb_insert_color(&data->node, root);
 568
 569         /* data->ct */
 570         return 0;
 571 }
 572
 573 static void event_uniq_destroy(struct rb_root *root)
 574 {
 575         /*
 576          * the strings we point to are in the giant block of memory filled by
 577          * the catalog, and are freed separately.
 578          */
 579         struct event_uniq *pos, *n;
 580
 581         rbtree_postorder_for_each_entry_safe(pos, n, root, node)
 582                 kfree(pos);
 583 }
 584
 585
 586 /*
 587  * ensure the event structure's sizes are self consistent and don't cause us to
 588  * read outside of the event
 589  *
 590  * On success, return the event length in bytes.
 591  * Otherwise, return -1 (and print as appropriate).
 592  */
 593 static ssize_t catalog_event_len_validate(struct hv_24x7_event_data *event,
 594                                           size_t event_idx,
 595                                           size_t event_data_bytes,
 596                                           size_t event_entry_count,
 597                                           size_t offset, void *end)
 598 {
 599         ssize_t ev_len;
 600         void *ev_end, *calc_ev_end;
 601
 602         if (offset >= event_data_bytes)
 603                 return -1;
 604
 605         if (event_idx >= event_entry_count) {
 606                 pr_devel("catalog event data has %zu bytes of padding after last event\n",
 607                                 event_data_bytes - offset);
 608                 return -1;
 609         }
 610
 611         if (!event_fixed_portion_is_within(event, end)) {
 612                 pr_warn("event %zu fixed portion is not within range\n",
 613                                 event_idx);
 614                 return -1;
 615         }
 616
 617         ev_len = be16_to_cpu(event->length);
 618
 619         if (ev_len % 16)
 620                 pr_info("event %zu has length %zu not divisible by 16: event=%pK\n",
 621                                 event_idx, ev_len, event);
 622
 623         ev_end = (__u8 *)event + ev_len;
 624         if (ev_end > end) {
 625                 pr_warn("event %zu has .length=%zu, ends after buffer end: ev_end=%pK > end=%pK, offset=%zu\n",
 626                                 event_idx, ev_len, ev_end, end,
 627                                 offset);
 628                 return -1;
 629         }
 630
 631         calc_ev_end = event_end(event, end);
 632         if (!calc_ev_end) {
 633                 pr_warn("event %zu has a calculated length which exceeds buffer length %zu: event=%pK end=%pK, offset=%zu\n",
 634                         event_idx, event_data_bytes, event, end,
 635                         offset);
 636                 return -1;
 637         }
 638
 639         if (calc_ev_end > ev_end) {
 640                 pr_warn("event %zu exceeds it's own length: event=%pK, end=%pK, offset=%zu, calc_ev_end=%pK\n",
 641                         event_idx, event, ev_end, offset, calc_ev_end);
 642                 return -1;
 643         }
 644
 645         return ev_len;
 646 }
 647
 648 #define MAX_4K (SIZE_MAX / 4096)
 649
 650 static void create_events_from_catalog(struct attribute ***events_,
 651                 struct attribute ***event_descs_,
 652                 struct attribute ***event_long_descs_)
 653 {
 654         unsigned long hret;
 655         size_t catalog_len, catalog_page_len, event_entry_count,
 656                event_data_len, event_data_offs,
 657                event_data_bytes, junk_events, event_idx, event_attr_ct, i,
 658                attr_max, event_idx_last, desc_ct, long_desc_ct;
 659         ssize_t ct, ev_len;
 660         uint32_t catalog_version_num;
 661         struct attribute **events, **event_descs, **event_long_descs;
 662         struct hv_24x7_catalog_page_0 *page_0 =
 663                 kmem_cache_alloc(hv_page_cache, GFP_KERNEL);
 664         void *page = page_0;
 665         void *event_data, *end;
 666         struct hv_24x7_event_data *event;
 667         struct rb_root ev_uniq = RB_ROOT;
 668
 669         if (!page)
 670                 goto e_out;
 671
 672         hret = h_get_24x7_catalog_page(page, 0, 0);
 673         if (hret)
 674                 goto e_free;
 675
 676         catalog_version_num = be64_to_cpu(page_0->version);
 677         catalog_page_len = be32_to_cpu(page_0->length);
 678
 679         if (MAX_4K < catalog_page_len) {
 680                 pr_err("invalid page count: %zu\n", catalog_page_len);
 681                 goto e_free;
 682         }
 683
 684         catalog_len = catalog_page_len * 4096;
 685
 686         event_entry_count = be16_to_cpu(page_0->event_entry_count);
 687         event_data_offs   = be16_to_cpu(page_0->event_data_offs);
 688         event_data_len    = be16_to_cpu(page_0->event_data_len);
 689
 690         pr_devel("cv %zu cl %zu eec %zu edo %zu edl %zu\n",
 691                         (size_t)catalog_version_num, catalog_len,
 692                         event_entry_count, event_data_offs, event_data_len);
 693
 694         if ((MAX_4K < event_data_len)
 695                         || (MAX_4K < event_data_offs)
 696                         || (MAX_4K - event_data_offs < event_data_len)) {
 697                 pr_err("invalid event data offs %zu and/or len %zu\n",
 698                                 event_data_offs, event_data_len);
 699                 goto e_free;
 700         }
 701
 702         if ((event_data_offs + event_data_len) > catalog_page_len) {
 703                 pr_err("event data %zu-%zu does not fit inside catalog 0-%zu\n",
 704                                 event_data_offs,
 705                                 event_data_offs + event_data_len,
 706                                 catalog_page_len);
 707                 goto e_free;
 708         }
 709
 710         if (SIZE_MAX / MAX_EVENTS_PER_EVENT_DATA - 1 < event_entry_count) {
 711                 pr_err("event_entry_count %zu is invalid\n",
 712                                 event_entry_count);
 713                 goto e_free;
 714         }
 715
 716         event_data_bytes = event_data_len * 4096;
 717
 718         /*
 719          * event data can span several pages, events can cross between these
 720          * pages. Use vmalloc to make this easier.
 721          */
 722         event_data = vmalloc(event_data_bytes);
 723         if (!event_data) {
 724                 pr_err("could not allocate event data\n");
 725                 goto e_free;
 726         }
 727
 728         end = event_data + event_data_bytes;
 729
 730         /*
 731          * using vmalloc_to_phys() like this only works if PAGE_SIZE is
 732          * divisible by 4096
 733          */
 734         BUILD_BUG_ON(PAGE_SIZE % 4096);
 735
 736         for (i = 0; i < event_data_len; i++) {
 737                 hret = h_get_24x7_catalog_page_(
 738                                 vmalloc_to_phys(event_data + i * 4096),
 739                                 catalog_version_num,
 740                                 i + event_data_offs);
 741                 if (hret) {
 742                         pr_err("failed to get event data in page %zu\n",
 743                                         i + event_data_offs);
 744                         goto e_event_data;
 745                 }
 746         }
 747
 748         /*
 749          * scan the catalog to determine the number of attributes we need, and
 750          * verify it at the same time.
 751          */
 752         for (junk_events = 0, event = event_data, event_idx = 0, attr_max = 0;
 753              ;
 754              event_idx++, event = (void *)event + ev_len) {
 755                 size_t offset = (void *)event - (void *)event_data;
 756                 char *name;
 757                 int nl;
 758
 759                 ev_len = catalog_event_len_validate(event, event_idx,
 760                                                     event_data_bytes,
 761                                                     event_entry_count,
 762                                                     offset, end);
 763                 if (ev_len < 0)
 764                         break;
 765
 766                 name = event_name(event, &nl);
 767
 768                 if (event->event_group_record_len == 0) {
 769                         pr_devel("invalid event %zu (%.*s): group_record_len == 0, skipping\n",
 770                                         event_idx, nl, name);
 771                         junk_events++;
 772                         continue;
 773                 }
 774
 775                 if (!catalog_entry_domain_is_valid(event->domain)) {
 776                         pr_info("event %zu (%.*s) has invalid domain %d\n",
 777                                         event_idx, nl, name, event->domain);
 778                         junk_events++;
 779                         continue;
 780                 }
 781
 782                 attr_max += event_to_attr_ct(event);
 783         }
 784
 785         event_idx_last = event_idx;
 786         if (event_idx_last != event_entry_count)
 787                 pr_warn("event buffer ended before listed # of events were parsed (got %zu, wanted %zu, junk %zu)\n",
 788                                 event_idx_last, event_entry_count, junk_events);
 789
 790         events = kmalloc_array(attr_max + 1, sizeof(*events), GFP_KERNEL);
 791         if (!events)
 792                 goto e_event_data;
 793
 794         event_descs = kmalloc_array(event_idx + 1, sizeof(*event_descs),
 795                                 GFP_KERNEL);
 796         if (!event_descs)
 797                 goto e_event_attrs;
 798
 799         event_long_descs = kmalloc_array(event_idx + 1,
 800                         sizeof(*event_long_descs), GFP_KERNEL);
 801         if (!event_long_descs)
 802                 goto e_event_descs;
 803
 804         /* Iterate over the catalog filling in the attribute vector */
 805         for (junk_events = 0, event_attr_ct = 0, desc_ct = 0, long_desc_ct = 0,
 806                                 event = event_data, event_idx = 0;
 807                         event_idx < event_idx_last;
 808                         event_idx++, ev_len = be16_to_cpu(event->length),
 809                                 event = (void *)event + ev_len) {
 810                 char *name;
 811                 int nl;
 812                 int nonce;
 813                 /*
 814                  * these are the only "bad" events that are intermixed and that
 815                  * we can ignore without issue. make sure to skip them here
 816                  */
 817                 if (event->event_group_record_len == 0)
 818                         continue;
 819                 if (!catalog_entry_domain_is_valid(event->domain))
 820                         continue;
 821
 822                 name  = event_name(event, &nl);
 823                 nonce = event_uniq_add(&ev_uniq, name, nl, event->domain);
 824                 ct    = event_data_to_attrs(event_idx, events + event_attr_ct,
 825                                             event, nonce);
 826                 if (ct <= 0) {
 827                         pr_warn("event %zu (%.*s) creation failure, skipping\n",
 828                                 event_idx, nl, name);
 829                         junk_events++;
 830                 } else {
 831                         event_attr_ct += ct;
 832                         event_descs[desc_ct] = event_to_desc_attr(event, nonce);
 833                         if (event_descs[desc_ct])
 834                                 desc_ct++;
 835                         event_long_descs[long_desc_ct] =
 836                                         event_to_long_desc_attr(event, nonce);
 837                         if (event_long_descs[long_desc_ct])
 838                                 long_desc_ct++;
 839                 }
 840         }
 841
 842         pr_info("read %zu catalog entries, created %zu event attrs (%zu failures), %zu descs\n",
 843                         event_idx, event_attr_ct, junk_events, desc_ct);
 844
 845         events[event_attr_ct] = NULL;
 846         event_descs[desc_ct] = NULL;
 847         event_long_descs[long_desc_ct] = NULL;
 848
 849         event_uniq_destroy(&ev_uniq);
 850         vfree(event_data);
 851         kmem_cache_free(hv_page_cache, page);
 852
 853         *events_ = events;
 854         *event_descs_ = event_descs;
 855         *event_long_descs_ = event_long_descs;
 856         return;
 857
 858 e_event_descs:
 859         kfree(event_descs);
 860 e_event_attrs:
 861         kfree(events);
 862 e_event_data:
 863         vfree(event_data);
 864 e_free:
 865         kmem_cache_free(hv_page_cache, page);
 866 e_out:
 867         *events_ = NULL;
 868         *event_descs_ = NULL;
 869         *event_long_descs_ = NULL;
 870 }
 871
 872 static ssize_t catalog_read(struct file *filp, struct kobject *kobj,
 873                             struct bin_attribute *bin_attr, char *buf,
 874                             loff_t offset, size_t count)
 875 {
 876         unsigned long hret;
 877         ssize_t ret = 0;
 878         size_t catalog_len = 0, catalog_page_len = 0;
 879         loff_t page_offset = 0;
 880         loff_t offset_in_page;
 881         size_t copy_len;
 882         uint64_t catalog_version_num = 0;
 883         void *page = kmem_cache_alloc(hv_page_cache, GFP_USER);
 884         struct hv_24x7_catalog_page_0 *page_0 = page;
 885
 886         if (!page)
 887                 return -ENOMEM;
 888
 889         hret = h_get_24x7_catalog_page(page, 0, 0);
 890         if (hret) {
 891                 ret = -EIO;
 892                 goto e_free;
 893         }
 894
 895         catalog_version_num = be64_to_cpu(page_0->version);
 896         catalog_page_len = be32_to_cpu(page_0->length);
 897         catalog_len = catalog_page_len * 4096;
 898
 899         page_offset = offset / 4096;
 900         offset_in_page = offset % 4096;
 901
 902         if (page_offset >= catalog_page_len)
 903                 goto e_free;
 904
 905         if (page_offset != 0) {
 906                 hret = h_get_24x7_catalog_page(page, catalog_version_num,
 907                                                page_offset);
 908                 if (hret) {
 909                         ret = -EIO;
 910                         goto e_free;
 911                 }
 912         }
 913
 914         copy_len = 4096 - offset_in_page;
 915         if (copy_len > count)
 916                 copy_len = count;
 917
 918         memcpy(buf, page+offset_in_page, copy_len);
 919         ret = copy_len;
 920
 921 e_free:
 922         if (hret)
 923                 pr_err("h_get_24x7_catalog_page(ver=%lld, page=%lld) failed:"
 924                        " rc=%ld\n",
 925                        catalog_version_num, page_offset, hret);
 926         kmem_cache_free(hv_page_cache, page);
 927
 928         pr_devel("catalog_read: offset=%lld(%lld) count=%zu "
 929                         "catalog_len=%zu(%zu) => %zd\n", offset, page_offset,
 930                         count, catalog_len, catalog_page_len, ret);
 931
 932         return ret;
 933 }
 934
 935 #define PAGE_0_ATTR(_name, _fmt, _expr)                         \
 936 static ssize_t _name##_show(struct device *dev,                 \
 937                             struct device_attribute *dev_attr,  \
 938                             char *buf)                          \
 939 {                                                               \
 940         unsigned long hret;                                     \
 941         ssize_t ret = 0;                                        \
 942         void *page = kmem_cache_alloc(hv_page_cache, GFP_USER); \
 943         struct hv_24x7_catalog_page_0 *page_0 = page;           \
 944         if (!page)                                              \
 945                 return -ENOMEM;                                 \
 946         hret = h_get_24x7_catalog_page(page, 0, 0);             \
 947         if (hret) {                                             \
 948                 ret = -EIO;                                     \
 949                 goto e_free;                                    \
 950         }                                                       \
 951         ret = sprintf(buf, _fmt, _expr);                        \
 952 e_free:                                                         \
 953         kmem_cache_free(hv_page_cache, page);                   \
 954         return ret;                                             \
 955 }                                                               \
 956 static DEVICE_ATTR_RO(_name)
 957
 958 PAGE_0_ATTR(catalog_version, "%lld\n",
 959                 (unsigned long long)be64_to_cpu(page_0->version));
 960 PAGE_0_ATTR(catalog_len, "%lld\n",
 961                 (unsigned long long)be32_to_cpu(page_0->length) * 4096);
 962 static BIN_ATTR_RO(catalog, 0/* real length varies */);
 963
 964 static struct bin_attribute *if_bin_attrs[] = {
 965         &bin_attr_catalog,
 966         NULL,
 967 };
 968
 969 static struct attribute *if_attrs[] = {
 970         &dev_attr_catalog_len.attr,
 971         &dev_attr_catalog_version.attr,
 972         NULL,
 973 };
 974
 975 static struct attribute_group if_group = {
 976         .name = "interface",
 977         .bin_attrs = if_bin_attrs,
 978         .attrs = if_attrs,
 979 };
 980
 981 static const struct attribute_group *attr_groups[] = {
 982         &format_group,
 983         &event_group,
 984         &event_desc_group,
 985         &event_long_desc_group,
 986         &if_group,
 987         NULL,
 988 };
 989
 990 static void log_24x7_hcall(struct hv_24x7_request_buffer *request_buffer,
 991                         struct hv_24x7_data_result_buffer *result_buffer,
 992                         unsigned long ret)
 993 {
 994         struct hv_24x7_request *req;
 995
 996         req = &request_buffer->requests[0];
 997         pr_notice_ratelimited("hcall failed: [%d %#x %#x %d] => "
 998                         "ret 0x%lx (%ld) detail=0x%x failing ix=%x\n",
 999                         req->performance_domain, req->data_offset,
1000                         req->starting_ix, req->starting_lpar_ix, ret, ret,
1001                         result_buffer->detailed_rc,
1002                         result_buffer->failing_request_ix);
1003 }
1004
1005 /*
1006  * Start the process for a new H_GET_24x7_DATA hcall.
1007  */
1008 static void init_24x7_request(struct hv_24x7_request_buffer *request_buffer,
1009                         struct hv_24x7_data_result_buffer *result_buffer)
1010 {
1011
1012         memset(request_buffer, 0, 4096);
1013         memset(result_buffer, 0, 4096);
1014
1015         request_buffer->interface_version = HV_24X7_IF_VERSION_CURRENT;
1016         /* memset above set request_buffer->num_requests to 0 */
1017 }
1018
1019 /*
1020  * Commit (i.e perform) the H_GET_24x7_DATA hcall using the data collected
1021  * by 'init_24x7_request()' and 'add_event_to_24x7_request()'.
1022  */
1023 static int make_24x7_request(struct hv_24x7_request_buffer *request_buffer,
1024                         struct hv_24x7_data_result_buffer *result_buffer)
1025 {
1026         unsigned long ret;
1027
1028         /*
1029          * NOTE: Due to variable number of array elements in request and
1030          *       result buffer(s), sizeof() is not reliable. Use the actual
1031          *       allocated buffer size, H24x7_DATA_BUFFER_SIZE.
1032          */
1033         ret = plpar_hcall_norets(H_GET_24X7_DATA,
1034                         virt_to_phys(request_buffer), H24x7_DATA_BUFFER_SIZE,
1035                         virt_to_phys(result_buffer),  H24x7_DATA_BUFFER_SIZE);
1036
1037         if (ret)
1038                 log_24x7_hcall(request_buffer, result_buffer, ret);
1039
1040         return ret;
1041 }
1042
1043 /*
1044  * Add the given @event to the next slot in the 24x7 request_buffer.
1045  *
1046  * Note that H_GET_24X7_DATA hcall allows reading several counters'
1047  * values in a single HCALL. We expect the caller to add events to the
1048  * request buffer one by one, make the HCALL and process the results.
1049  */
1050 static int add_event_to_24x7_request(struct perf_event *event,
1051                                 struct hv_24x7_request_buffer *request_buffer)
1052 {
1053         u16 idx;
1054         int i;
1055         struct hv_24x7_request *req;
1056
1057         if (request_buffer->num_requests > 254) {
1058                 pr_devel("Too many requests for 24x7 HCALL %d\n",
1059                                 request_buffer->num_requests);
1060                 return -EINVAL;
1061         }
1062
1063         if (is_physical_domain(event_get_domain(event)))
1064                 idx = event_get_core(event);
1065         else
1066                 idx = event_get_vcpu(event);
1067
1068         i = request_buffer->num_requests++;
1069         req = &request_buffer->requests[i];
1070
1071         req->performance_domain = event_get_domain(event);
1072         req->data_size = cpu_to_be16(8);
1073         req->data_offset = cpu_to_be32(event_get_offset(event));
1074         req->starting_lpar_ix = cpu_to_be16(event_get_lpar(event)),
1075         req->max_num_lpars = cpu_to_be16(1);
1076         req->starting_ix = cpu_to_be16(idx);
1077         req->max_ix = cpu_to_be16(1);
1078
1079         return 0;
1080 }
1081
1082 static unsigned long single_24x7_request(struct perf_event *event, u64 *count)
1083 {
1084         unsigned long ret;
1085         struct hv_24x7_request_buffer *request_buffer;
1086         struct hv_24x7_data_result_buffer *result_buffer;
1087         struct hv_24x7_result *resb;
1088
1089         BUILD_BUG_ON(sizeof(*request_buffer) > 4096);
1090         BUILD_BUG_ON(sizeof(*result_buffer) > 4096);
1091
1092         request_buffer = (void *)get_cpu_var(hv_24x7_reqb);
1093         result_buffer = (void *)get_cpu_var(hv_24x7_resb);
1094
1095         init_24x7_request(request_buffer, result_buffer);
1096
1097         ret = add_event_to_24x7_request(event, request_buffer);
1098         if (ret)
1099                 return ret;
1100
1101         ret = make_24x7_request(request_buffer, result_buffer);
1102         if (ret) {
1103                 log_24x7_hcall(request_buffer, result_buffer, ret);
1104                 goto out;
1105         }
1106
1107         /* process result from hcall */
1108         resb = &result_buffer->results[0];
1109         *count = be64_to_cpu(resb->elements[0].element_data[0]);
1110
1111 out:
1112         return ret;
1113 }
1114
1115
1116 static int h_24x7_event_init(struct perf_event *event)
1117 {
1118         struct hv_perf_caps caps;
1119         unsigned domain;
1120         unsigned long hret;
1121         u64 ct;
1122
1123         /* Not our event */
1124         if (event->attr.type != event->pmu->type)
1125                 return -ENOENT;
1126
1127         /* Unused areas must be 0 */
1128         if (event_get_reserved1(event) ||
1129             event_get_reserved2(event) ||
1130             event_get_reserved3(event)) {
1131                 pr_devel("reserved set when forbidden 0x%llx(0x%llx) 0x%llx(0x%llx) 0x%llx(0x%llx)\n",
1132                                 event->attr.config,
1133                                 event_get_reserved1(event),
1134                                 event->attr.config1,
1135                                 event_get_reserved2(event),
1136                                 event->attr.config2,
1137                                 event_get_reserved3(event));
1138                 return -EINVAL;
1139         }
1140
1141         /* unsupported modes and filters */
1142         if (event->attr.exclude_user   ||
1143             event->attr.exclude_kernel ||
1144             event->attr.exclude_hv     ||
1145             event->attr.exclude_idle   ||
1146             event->attr.exclude_host   ||
1147             event->attr.exclude_guest)
1148                 return -EINVAL;
1149
1150         /* no branch sampling */
1151         if (has_branch_stack(event))
1152                 return -EOPNOTSUPP;
1153
1154         /* offset must be 8 byte aligned */
1155         if (event_get_offset(event) % 8) {
1156                 pr_devel("bad alignment\n");
1157                 return -EINVAL;
1158         }
1159
1160         /* Domains above 6 are invalid */
1161         domain = event_get_domain(event);
1162         if (domain > 6) {
1163                 pr_devel("invalid domain %d\n", domain);
1164                 return -EINVAL;
1165         }
1166
1167         hret = hv_perf_caps_get(&caps);
1168         if (hret) {
1169                 pr_devel("could not get capabilities: rc=%ld\n", hret);
1170                 return -EIO;
1171         }
1172
1173         /* Physical domains & other lpars require extra capabilities */
1174         if (!caps.collect_privileged && (is_physical_domain(domain) ||
1175                 (event_get_lpar(event) != event_get_lpar_max()))) {
1176                 pr_devel("hv permisions disallow: is_physical_domain:%d, lpar=0x%llx\n",
1177                                 is_physical_domain(domain),
1178                                 event_get_lpar(event));
1179                 return -EACCES;
1180         }
1181
1182         /* see if the event complains */
1183         if (single_24x7_request(event, &ct)) {
1184                 pr_devel("test hcall failed\n");
1185                 return -EIO;
1186         }
1187
1188         return 0;
1189 }
1190
1191 static u64 h_24x7_get_value(struct perf_event *event)
1192 {
1193         unsigned long ret;
1194         u64 ct;
1195         ret = single_24x7_request(event, &ct);
1196         if (ret)
1197                 /* We checked this in event init, shouldn't fail here... */
1198                 return 0;
1199
1200         return ct;
1201 }
1202
1203 static void update_event_count(struct perf_event *event, u64 now)
1204 {
1205         s64 prev;
1206
1207         prev = local64_xchg(&event->hw.prev_count, now);
1208         local64_add(now - prev, &event->count);
1209 }
1210
1211 static void h_24x7_event_read(struct perf_event *event)
1212 {
1213         u64 now;
1214
1215         now = h_24x7_get_value(event);
1216         update_event_count(event, now);
1217 }
1218
1219 static void h_24x7_event_start(struct perf_event *event, int flags)
1220 {
1221         if (flags & PERF_EF_RELOAD)
1222                 local64_set(&event->hw.prev_count, h_24x7_get_value(event));
1223 }
1224
1225 static void h_24x7_event_stop(struct perf_event *event, int flags)
1226 {
1227         h_24x7_event_read(event);
1228 }
1229
1230 static int h_24x7_event_add(struct perf_event *event, int flags)
1231 {
1232         if (flags & PERF_EF_START)
1233                 h_24x7_event_start(event, flags);
1234
1235         return 0;
1236 }
1237
1238 static struct pmu h_24x7_pmu = {
1239         .task_ctx_nr = perf_invalid_context,
1240
1241         .name = "hv_24x7",
1242         .attr_groups = attr_groups,
1243         .event_init  = h_24x7_event_init,
1244         .add         = h_24x7_event_add,
1245         .del         = h_24x7_event_stop,
1246         .start       = h_24x7_event_start,
1247         .stop        = h_24x7_event_stop,
1248         .read        = h_24x7_event_read,
1249 };
1250
1251 static int hv_24x7_init(void)
1252 {
1253         int r;
1254         unsigned long hret;
1255         struct hv_perf_caps caps;
1256
1257         if (!firmware_has_feature(FW_FEATURE_LPAR)) {
1258                 pr_debug("not a virtualized system, not enabling\n");
1259                 return -ENODEV;
1260         }
1261
1262         hret = hv_perf_caps_get(&caps);
1263         if (hret) {
1264                 pr_debug("could not obtain capabilities, not enabling, rc=%ld\n",
1265                                 hret);
1266                 return -ENODEV;
1267         }
1268
1269         hv_page_cache = kmem_cache_create("hv-page-4096", 4096, 4096, 0, NULL);
1270         if (!hv_page_cache)
1271                 return -ENOMEM;
1272
1273         /* sampling not supported */
1274         h_24x7_pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
1275
1276         create_events_from_catalog(&event_group.attrs,
1277                                    &event_desc_group.attrs,
1278                                    &event_long_desc_group.attrs);
1279
1280         r = perf_pmu_register(&h_24x7_pmu, h_24x7_pmu.name, -1);
1281         if (r)
1282                 return r;
1283
1284         return 0;
1285 }
1286
1287 device_initcall(hv_24x7_init);