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         *len = be16_to_cpu(*desc_len) - 2;
 165         return (char *)ev->remainder + nl;
 166 }
 167
 168 static char *event_long_desc(struct hv_24x7_event_data *ev, int *len)
 169 {
 170         unsigned nl = be16_to_cpu(ev->event_name_len);
 171         __be16 *desc_len_ = (__be16 *)(ev->remainder + nl - 2);
 172         unsigned desc_len = be16_to_cpu(*desc_len_);
 173         __be16 *long_desc_len = (__be16 *)(ev->remainder + nl + desc_len - 2);
 174         *len = be16_to_cpu(*long_desc_len) - 2;
 175         return (char *)ev->remainder + nl + desc_len;
 176 }
 177
 178 static bool event_fixed_portion_is_within(struct hv_24x7_event_data *ev,
 179                                           void *end)
 180 {
 181         void *start = ev;
 182
 183         return (start + offsetof(struct hv_24x7_event_data, remainder)) < end;
 184 }
 185
 186 /*
 187  * Things we don't check:
 188  *  - padding for desc, name, and long/detailed desc is required to be '\0'
 189  *    bytes.
 190  *
 191  *  Return NULL if we pass end,
 192  *  Otherwise return the address of the byte just following the event.
 193  */
 194 static void *event_end(struct hv_24x7_event_data *ev, void *end)
 195 {
 196         void *start = ev;
 197         __be16 *dl_, *ldl_;
 198         unsigned dl, ldl;
 199         unsigned nl = be16_to_cpu(ev->event_name_len);
 200
 201         if (nl < 2) {
 202                 pr_debug("%s: name length too short: %d", __func__, nl);
 203                 return NULL;
 204         }
 205
 206         if (start + nl > end) {
 207                 pr_debug("%s: start=%p + nl=%u > end=%p",
 208                                 __func__, start, nl, end);
 209                 return NULL;
 210         }
 211
 212         dl_ = (__be16 *)(ev->remainder + nl - 2);
 213         if (!IS_ALIGNED((uintptr_t)dl_, 2))
 214                 pr_warn("desc len not aligned %p", dl_);
 215         dl = be16_to_cpu(*dl_);
 216         if (dl < 2) {
 217                 pr_debug("%s: desc len too short: %d", __func__, dl);
 218                 return NULL;
 219         }
 220
 221         if (start + nl + dl > end) {
 222                 pr_debug("%s: (start=%p + nl=%u + dl=%u)=%p > end=%p",
 223                                 __func__, start, nl, dl, start + nl + dl, end);
 224                 return NULL;
 225         }
 226
 227         ldl_ = (__be16 *)(ev->remainder + nl + dl - 2);
 228         if (!IS_ALIGNED((uintptr_t)ldl_, 2))
 229                 pr_warn("long desc len not aligned %p", ldl_);
 230         ldl = be16_to_cpu(*ldl_);
 231         if (ldl < 2) {
 232                 pr_debug("%s: long desc len too short (ldl=%u)",
 233                                 __func__, ldl);
 234                 return NULL;
 235         }
 236
 237         if (start + nl + dl + ldl > end) {
 238                 pr_debug("%s: start=%p + nl=%u + dl=%u + ldl=%u > end=%p",
 239                                 __func__, start, nl, dl, ldl, end);
 240                 return NULL;
 241         }
 242
 243         return start + nl + dl + ldl;
 244 }
 245
 246 static unsigned long h_get_24x7_catalog_page_(unsigned long phys_4096,
 247                                               unsigned long version,
 248                                               unsigned long index)
 249 {
 250         pr_devel("h_get_24x7_catalog_page(0x%lx, %lu, %lu)",
 251                         phys_4096,
 252                         version,
 253                         index);
 254         WARN_ON(!IS_ALIGNED(phys_4096, 4096));
 255         return plpar_hcall_norets(H_GET_24X7_CATALOG_PAGE,
 256                         phys_4096,
 257                         version,
 258                         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         return sprintf(buf, "%s\n", (char *)d->var);
 313 }
 314
 315 static struct attribute *device_str_attr_create_(char *name, char *str)
 316 {
 317         struct dev_ext_attribute *attr = kzalloc(sizeof(*attr), GFP_KERNEL);
 318
 319         if (!attr)
 320                 return NULL;
 321
 322         attr->var = str;
 323         attr->attr.attr.name = name;
 324         attr->attr.attr.mode = 0444;
 325         attr->attr.show = device_show_string;
 326         return &attr->attr.attr;
 327 }
 328
 329 static struct attribute *device_str_attr_create(char *name, int name_max,
 330                                                 int name_nonce,
 331                                                 char *str, size_t str_max)
 332 {
 333         char *n;
 334         char *s = memdup_to_str(str, str_max, GFP_KERNEL);
 335         struct attribute *a;
 336
 337         if (!s)
 338                 return NULL;
 339
 340         if (!name_nonce)
 341                 n = kasprintf(GFP_KERNEL, "%.*s", name_max, name);
 342         else
 343                 n = kasprintf(GFP_KERNEL, "%.*s__%d", name_max, name,
 344                                         name_nonce);
 345         if (!n)
 346                 goto out_s;
 347
 348         a = device_str_attr_create_(n, s);
 349         if (!a)
 350                 goto out_n;
 351
 352         return a;
 353 out_n:
 354         kfree(n);
 355 out_s:
 356         kfree(s);
 357         return NULL;
 358 }
 359
 360 static void device_str_attr_destroy(struct attribute *attr)
 361 {
 362         struct dev_ext_attribute *d;
 363
 364         d = container_of(attr, struct dev_ext_attribute, attr.attr);
 365         kfree(d->var);
 366         kfree(d->attr.attr.name);
 367         kfree(d);
 368 }
 369
 370 static struct attribute *event_to_attr(unsigned ix,
 371                                        struct hv_24x7_event_data *event,
 372                                        unsigned domain,
 373                                        int nonce)
 374 {
 375         int event_name_len;
 376         char *ev_name, *a_ev_name, *val;
 377         const char *ev_suffix;
 378         struct attribute *attr;
 379
 380         if (!domain_is_valid(domain)) {
 381                 pr_warn("catalog event %u has invalid domain %u\n",
 382                                 ix, domain);
 383                 return NULL;
 384         }
 385
 386         val = event_fmt(event, domain);
 387         if (!val)
 388                 return NULL;
 389
 390         ev_suffix = event_domain_suffix(domain);
 391         ev_name = event_name(event, &event_name_len);
 392         if (!nonce)
 393                 a_ev_name = kasprintf(GFP_KERNEL, "%.*s%s",
 394                                 (int)event_name_len, ev_name, ev_suffix);
 395         else
 396                 a_ev_name = kasprintf(GFP_KERNEL, "%.*s%s__%d",
 397                                 (int)event_name_len, ev_name, ev_suffix, nonce);
 398
 399
 400         if (!a_ev_name)
 401                 goto out_val;
 402
 403         attr = device_str_attr_create_(a_ev_name, val);
 404         if (!attr)
 405                 goto out_name;
 406
 407         return attr;
 408 out_name:
 409         kfree(a_ev_name);
 410 out_val:
 411         kfree(val);
 412         return NULL;
 413 }
 414
 415 static struct attribute *event_to_desc_attr(struct hv_24x7_event_data *event,
 416                                 int nonce)
 417 {
 418         int nl, dl;
 419         char *name = event_name(event, &nl);
 420         char *desc = event_desc(event, &dl);
 421
 422         /* If there isn't a description, don't create the sysfs file */
 423         if (!dl)
 424                 return NULL;
 425
 426         return device_str_attr_create(name, nl, nonce, desc, dl);
 427 }
 428
 429 static struct attribute *
 430 event_to_long_desc_attr(struct hv_24x7_event_data *event, int nonce)
 431 {
 432         int nl, dl;
 433         char *name = event_name(event, &nl);
 434         char *desc = event_long_desc(event, &dl);
 435
 436         /* If there isn't a description, don't create the sysfs file */
 437         if (!dl)
 438                 return NULL;
 439
 440         return device_str_attr_create(name, nl, nonce, desc, dl);
 441 }
 442
 443 static ssize_t event_data_to_attrs(unsigned ix, struct attribute **attrs,
 444                 struct hv_24x7_event_data *event, int nonce)
 445 {
 446         unsigned i;
 447
 448         switch (event->domain) {
 449         case HV_PERF_DOMAIN_PHYS_CHIP:
 450                 *attrs = event_to_attr(ix, event, event->domain, nonce);
 451                 return 1;
 452         case HV_PERF_DOMAIN_PHYS_CORE:
 453                 for (i = 0; i < ARRAY_SIZE(core_domains); i++) {
 454                         attrs[i] = event_to_attr(ix, event, core_domains[i],
 455                                                 nonce);
 456                         if (!attrs[i]) {
 457                                 pr_warn("catalog event %u: individual attr %u "
 458                                         "creation failure\n", ix, i);
 459                                 for (; i; i--)
 460                                         device_str_attr_destroy(attrs[i - 1]);
 461                                 return -1;
 462                         }
 463                 }
 464                 return i;
 465         default:
 466                 pr_warn("catalog event %u: domain %u is not allowed in the "
 467                                 "catalog\n", ix, event->domain);
 468                 return -1;
 469         }
 470 }
 471
 472 static size_t event_to_attr_ct(struct hv_24x7_event_data *event)
 473 {
 474         switch (event->domain) {
 475         case HV_PERF_DOMAIN_PHYS_CHIP:
 476                 return 1;
 477         case HV_PERF_DOMAIN_PHYS_CORE:
 478                 return ARRAY_SIZE(core_domains);
 479         default:
 480                 return 0;
 481         }
 482 }
 483
 484 static unsigned long vmalloc_to_phys(void *v)
 485 {
 486         struct page *p = vmalloc_to_page(v);
 487
 488         BUG_ON(!p);
 489         return page_to_phys(p) + offset_in_page(v);
 490 }
 491
 492 /* */
 493 struct event_uniq {
 494         struct rb_node node;
 495         const char *name;
 496         int nl;
 497         unsigned ct;
 498         unsigned domain;
 499 };
 500
 501 static int memord(const void *d1, size_t s1, const void *d2, size_t s2)
 502 {
 503         if (s1 < s2)
 504                 return 1;
 505         if (s2 > s1)
 506                 return -1;
 507
 508         return memcmp(d1, d2, s1);
 509 }
 510
 511 static int ev_uniq_ord(const void *v1, size_t s1, unsigned d1, const void *v2,
 512                                         size_t s2, unsigned d2)
 513 {
 514         int r = memord(v1, s1, v2, s2);
 515
 516         if (r)
 517                 return r;
 518         if (d1 > d2)
 519                 return 1;
 520         if (d2 > d1)
 521                 return -1;
 522         return 0;
 523 }
 524
 525 static int event_uniq_add(struct rb_root *root, const char *name, int nl,
 526                                 unsigned domain)
 527 {
 528         struct rb_node **new = &(root->rb_node), *parent = NULL;
 529         struct event_uniq *data;
 530
 531         /* Figure out where to put new node */
 532         while (*new) {
 533                 struct event_uniq *it;
 534                 int result;
 535
 536                 it = container_of(*new, struct event_uniq, node);
 537                 result = ev_uniq_ord(name, nl, domain, it->name, it->nl,
 538                                         it->domain);
 539
 540                 parent = *new;
 541                 if (result < 0)
 542                         new = &((*new)->rb_left);
 543                 else if (result > 0)
 544                         new = &((*new)->rb_right);
 545                 else {
 546                         it->ct++;
 547                         pr_info("found a duplicate event %.*s, ct=%u\n", nl,
 548                                                 name, it->ct);
 549                         return it->ct;
 550                 }
 551         }
 552
 553         data = kmalloc(sizeof(*data), GFP_KERNEL);
 554         if (!data)
 555                 return -ENOMEM;
 556
 557         *data = (struct event_uniq) {
 558                 .name = name,
 559                 .nl = nl,
 560                 .ct = 0,
 561                 .domain = domain,
 562         };
 563
 564         /* Add new node and rebalance tree. */
 565         rb_link_node(&data->node, parent, new);
 566         rb_insert_color(&data->node, root);
 567
 568         /* data->ct */
 569         return 0;
 570 }
 571
 572 static void event_uniq_destroy(struct rb_root *root)
 573 {
 574         /*
 575          * the strings we point to are in the giant block of memory filled by
 576          * the catalog, and are freed separately.
 577          */
 578         struct event_uniq *pos, *n;
 579
 580         rbtree_postorder_for_each_entry_safe(pos, n, root, node)
 581                 kfree(pos);
 582 }
 583
 584
 585 /*
 586  * ensure the event structure's sizes are self consistent and don't cause us to
 587  * read outside of the event
 588  *
 589  * On success, return the event length in bytes.
 590  * Otherwise, return -1 (and print as appropriate).
 591  */
 592 static ssize_t catalog_event_len_validate(struct hv_24x7_event_data *event,
 593                                           size_t event_idx,
 594                                           size_t event_data_bytes,
 595                                           size_t event_entry_count,
 596                                           size_t offset, void *end)
 597 {
 598         ssize_t ev_len;
 599         void *ev_end, *calc_ev_end;
 600
 601         if (offset >= event_data_bytes)
 602                 return -1;
 603
 604         if (event_idx >= event_entry_count) {
 605                 pr_devel("catalog event data has %zu bytes of padding after last event\n",
 606                                 event_data_bytes - offset);
 607                 return -1;
 608         }
 609
 610         if (!event_fixed_portion_is_within(event, end)) {
 611                 pr_warn("event %zu fixed portion is not within range\n",
 612                                 event_idx);
 613                 return -1;
 614         }
 615
 616         ev_len = be16_to_cpu(event->length);
 617
 618         if (ev_len % 16)
 619                 pr_info("event %zu has length %zu not divisible by 16: event=%pK\n",
 620                                 event_idx, ev_len, event);
 621
 622         ev_end = (__u8 *)event + ev_len;
 623         if (ev_end > end) {
 624                 pr_warn("event %zu has .length=%zu, ends after buffer end: ev_end=%pK > end=%pK, offset=%zu\n",
 625                                 event_idx, ev_len, ev_end, end,
 626                                 offset);
 627                 return -1;
 628         }
 629
 630         calc_ev_end = event_end(event, end);
 631         if (!calc_ev_end) {
 632                 pr_warn("event %zu has a calculated length which exceeds buffer length %zu: event=%pK end=%pK, offset=%zu\n",
 633                         event_idx, event_data_bytes, event, end,
 634                         offset);
 635                 return -1;
 636         }
 637
 638         if (calc_ev_end > ev_end) {
 639                 pr_warn("event %zu exceeds it's own length: event=%pK, end=%pK, offset=%zu, calc_ev_end=%pK\n",
 640                         event_idx, event, ev_end, offset, calc_ev_end);
 641                 return -1;
 642         }
 643
 644         return ev_len;
 645 }
 646
 647 #define MAX_4K (SIZE_MAX / 4096)
 648
 649 static void create_events_from_catalog(struct attribute ***events_,
 650                 struct attribute ***event_descs_,
 651                 struct attribute ***event_long_descs_)
 652 {
 653         unsigned long hret;
 654         size_t catalog_len, catalog_page_len, event_entry_count,
 655                event_data_len, event_data_offs,
 656                event_data_bytes, junk_events, event_idx, event_attr_ct, i,
 657                attr_max, event_idx_last, desc_ct, long_desc_ct;
 658         ssize_t ct, ev_len;
 659         uint32_t catalog_version_num;
 660         struct attribute **events, **event_descs, **event_long_descs;
 661         struct hv_24x7_catalog_page_0 *page_0 =
 662                 kmem_cache_alloc(hv_page_cache, GFP_KERNEL);
 663         void *page = page_0;
 664         void *event_data, *end;
 665         struct hv_24x7_event_data *event;
 666         struct rb_root ev_uniq = RB_ROOT;
 667
 668         if (!page)
 669                 goto e_out;
 670
 671         hret = h_get_24x7_catalog_page(page, 0, 0);
 672         if (hret)
 673                 goto e_free;
 674
 675         catalog_version_num = be64_to_cpu(page_0->version);
 676         catalog_page_len = be32_to_cpu(page_0->length);
 677
 678         if (MAX_4K < catalog_page_len) {
 679                 pr_err("invalid page count: %zu\n", catalog_page_len);
 680                 goto e_free;
 681         }
 682
 683         catalog_len = catalog_page_len * 4096;
 684
 685         event_entry_count = be16_to_cpu(page_0->event_entry_count);
 686         event_data_offs   = be16_to_cpu(page_0->event_data_offs);
 687         event_data_len    = be16_to_cpu(page_0->event_data_len);
 688
 689         pr_devel("cv %zu cl %zu eec %zu edo %zu edl %zu\n",
 690                         (size_t)catalog_version_num, catalog_len,
 691                         event_entry_count, event_data_offs, event_data_len);
 692
 693         if ((MAX_4K < event_data_len)
 694                         || (MAX_4K < event_data_offs)
 695                         || (MAX_4K - event_data_offs < event_data_len)) {
 696                 pr_err("invalid event data offs %zu and/or len %zu\n",
 697                                 event_data_offs, event_data_len);
 698                 goto e_free;
 699         }
 700
 701         if ((event_data_offs + event_data_len) > catalog_page_len) {
 702                 pr_err("event data %zu-%zu does not fit inside catalog 0-%zu\n",
 703                                 event_data_offs,
 704                                 event_data_offs + event_data_len,
 705                                 catalog_page_len);
 706                 goto e_free;
 707         }
 708
 709         if (SIZE_MAX / MAX_EVENTS_PER_EVENT_DATA - 1 < event_entry_count) {
 710                 pr_err("event_entry_count %zu is invalid\n",
 711                                 event_entry_count);
 712                 goto e_free;
 713         }
 714
 715         event_data_bytes = event_data_len * 4096;
 716
 717         /*
 718          * event data can span several pages, events can cross between these
 719          * pages. Use vmalloc to make this easier.
 720          */
 721         event_data = vmalloc(event_data_bytes);
 722         if (!event_data) {
 723                 pr_err("could not allocate event data\n");
 724                 goto e_free;
 725         }
 726
 727         end = event_data + event_data_bytes;
 728
 729         /*
 730          * using vmalloc_to_phys() like this only works if PAGE_SIZE is
 731          * divisible by 4096
 732          */
 733         BUILD_BUG_ON(PAGE_SIZE % 4096);
 734
 735         for (i = 0; i < event_data_len; i++) {
 736                 hret = h_get_24x7_catalog_page_(
 737                                 vmalloc_to_phys(event_data + i * 4096),
 738                                 catalog_version_num,
 739                                 i + event_data_offs);
 740                 if (hret) {
 741                         pr_err("failed to get event data in page %zu\n",
 742                                         i + event_data_offs);
 743                         goto e_event_data;
 744                 }
 745         }
 746
 747         /*
 748          * scan the catalog to determine the number of attributes we need, and
 749          * verify it at the same time.
 750          */
 751         for (junk_events = 0, event = event_data, event_idx = 0, attr_max = 0;
 752              ;
 753              event_idx++, event = (void *)event + ev_len) {
 754                 size_t offset = (void *)event - (void *)event_data;
 755                 char *name;
 756                 int nl;
 757
 758                 ev_len = catalog_event_len_validate(event, event_idx,
 759                                                     event_data_bytes,
 760                                                     event_entry_count,
 761                                                     offset, end);
 762                 if (ev_len < 0)
 763                         break;
 764
 765                 name = event_name(event, &nl);
 766
 767                 if (event->event_group_record_len == 0) {
 768                         pr_devel("invalid event %zu (%.*s): group_record_len == 0, skipping\n",
 769                                         event_idx, nl, name);
 770                         junk_events++;
 771                         continue;
 772                 }
 773
 774                 if (!catalog_entry_domain_is_valid(event->domain)) {
 775                         pr_info("event %zu (%.*s) has invalid domain %d\n",
 776                                         event_idx, nl, name, event->domain);
 777                         junk_events++;
 778                         continue;
 779                 }
 780
 781                 attr_max += event_to_attr_ct(event);
 782         }
 783
 784         event_idx_last = event_idx;
 785         if (event_idx_last != event_entry_count)
 786                 pr_warn("event buffer ended before listed # of events were parsed (got %zu, wanted %zu, junk %zu)\n",
 787                                 event_idx_last, event_entry_count, junk_events);
 788
 789         events = kmalloc_array(attr_max + 1, sizeof(*events), GFP_KERNEL);
 790         if (!events)
 791                 goto e_event_data;
 792
 793         event_descs = kmalloc_array(event_idx + 1, sizeof(*event_descs),
 794                                 GFP_KERNEL);
 795         if (!event_descs)
 796                 goto e_event_attrs;
 797
 798         event_long_descs = kmalloc_array(event_idx + 1,
 799                         sizeof(*event_long_descs), GFP_KERNEL);
 800         if (!event_long_descs)
 801                 goto e_event_descs;
 802
 803         /* Iterate over the catalog filling in the attribute vector */
 804         for (junk_events = 0, event_attr_ct = 0, desc_ct = 0, long_desc_ct = 0,
 805                                 event = event_data, event_idx = 0;
 806                         event_idx < event_idx_last;
 807                         event_idx++, ev_len = be16_to_cpu(event->length),
 808                                 event = (void *)event + ev_len) {
 809                 char *name;
 810                 int nl;
 811                 int nonce;
 812                 /*
 813                  * these are the only "bad" events that are intermixed and that
 814                  * we can ignore without issue. make sure to skip them here
 815                  */
 816                 if (event->event_group_record_len == 0)
 817                         continue;
 818                 if (!catalog_entry_domain_is_valid(event->domain))
 819                         continue;
 820
 821                 name  = event_name(event, &nl);
 822                 nonce = event_uniq_add(&ev_uniq, name, nl, event->domain);
 823                 ct    = event_data_to_attrs(event_idx, events + event_attr_ct,
 824                                             event, nonce);
 825                 if (ct <= 0) {
 826                         pr_warn("event %zu (%.*s) creation failure, skipping\n",
 827                                 event_idx, nl, name);
 828                         junk_events++;
 829                 } else {
 830                         event_attr_ct += ct;
 831                         event_descs[desc_ct] = event_to_desc_attr(event, nonce);
 832                         if (event_descs[desc_ct])
 833                                 desc_ct++;
 834                         event_long_descs[long_desc_ct] =
 835                                         event_to_long_desc_attr(event, nonce);
 836                         if (event_long_descs[long_desc_ct])
 837                                 long_desc_ct++;
 838                 }
 839         }
 840
 841         pr_info("read %zu catalog entries, created %zu event attrs (%zu failures), %zu descs\n",
 842                         event_idx, event_attr_ct, junk_events, desc_ct);
 843
 844         events[event_attr_ct] = NULL;
 845         event_descs[desc_ct] = NULL;
 846         event_long_descs[long_desc_ct] = NULL;
 847
 848         event_uniq_destroy(&ev_uniq);
 849         vfree(event_data);
 850         kmem_cache_free(hv_page_cache, page);
 851
 852         *events_ = events;
 853         *event_descs_ = event_descs;
 854         *event_long_descs_ = event_long_descs;
 855         return;
 856
 857 e_event_descs:
 858         kfree(event_descs);
 859 e_event_attrs:
 860         kfree(events);
 861 e_event_data:
 862         vfree(event_data);
 863 e_free:
 864         kmem_cache_free(hv_page_cache, page);
 865 e_out:
 866         *events_ = NULL;
 867         *event_descs_ = NULL;
 868         *event_long_descs_ = NULL;
 869 }
 870
 871 static ssize_t catalog_read(struct file *filp, struct kobject *kobj,
 872                             struct bin_attribute *bin_attr, char *buf,
 873                             loff_t offset, size_t count)
 874 {
 875         unsigned long hret;
 876         ssize_t ret = 0;
 877         size_t catalog_len = 0, catalog_page_len = 0;
 878         loff_t page_offset = 0;
 879         loff_t offset_in_page;
 880         size_t copy_len;
 881         uint64_t catalog_version_num = 0;
 882         void *page = kmem_cache_alloc(hv_page_cache, GFP_USER);
 883         struct hv_24x7_catalog_page_0 *page_0 = page;
 884         if (!page)
 885                 return -ENOMEM;
 886
 887         hret = h_get_24x7_catalog_page(page, 0, 0);
 888         if (hret) {
 889                 ret = -EIO;
 890                 goto e_free;
 891         }
 892
 893         catalog_version_num = be64_to_cpu(page_0->version);
 894         catalog_page_len = be32_to_cpu(page_0->length);
 895         catalog_len = catalog_page_len * 4096;
 896
 897         page_offset = offset / 4096;
 898         offset_in_page = offset % 4096;
 899
 900         if (page_offset >= catalog_page_len)
 901                 goto e_free;
 902
 903         if (page_offset != 0) {
 904                 hret = h_get_24x7_catalog_page(page, catalog_version_num,
 905                                                page_offset);
 906                 if (hret) {
 907                         ret = -EIO;
 908                         goto e_free;
 909                 }
 910         }
 911
 912         copy_len = 4096 - offset_in_page;
 913         if (copy_len > count)
 914                 copy_len = count;
 915
 916         memcpy(buf, page+offset_in_page, copy_len);
 917         ret = copy_len;
 918
 919 e_free:
 920         if (hret)
 921                 pr_err("h_get_24x7_catalog_page(ver=%lld, page=%lld) failed:"
 922                        " rc=%ld\n",
 923                        catalog_version_num, page_offset, hret);
 924         kmem_cache_free(hv_page_cache, page);
 925
 926         pr_devel("catalog_read: offset=%lld(%lld) count=%zu "
 927                         "catalog_len=%zu(%zu) => %zd\n", offset, page_offset,
 928                         count, catalog_len, catalog_page_len, ret);
 929
 930         return ret;
 931 }
 932
 933 #define PAGE_0_ATTR(_name, _fmt, _expr)                         \
 934 static ssize_t _name##_show(struct device *dev,                 \
 935                             struct device_attribute *dev_attr,  \
 936                             char *buf)                          \
 937 {                                                               \
 938         unsigned long hret;                                     \
 939         ssize_t ret = 0;                                        \
 940         void *page = kmem_cache_alloc(hv_page_cache, GFP_USER); \
 941         struct hv_24x7_catalog_page_0 *page_0 = page;           \
 942         if (!page)                                              \
 943                 return -ENOMEM;                                 \
 944         hret = h_get_24x7_catalog_page(page, 0, 0);             \
 945         if (hret) {                                             \
 946                 ret = -EIO;                                     \
 947                 goto e_free;                                    \
 948         }                                                       \
 949         ret = sprintf(buf, _fmt, _expr);                        \
 950 e_free:                                                         \
 951         kmem_cache_free(hv_page_cache, page);                   \
 952         return ret;                                             \
 953 }                                                               \
 954 static DEVICE_ATTR_RO(_name)
 955
 956 PAGE_0_ATTR(catalog_version, "%lld\n",
 957                 (unsigned long long)be64_to_cpu(page_0->version));
 958 PAGE_0_ATTR(catalog_len, "%lld\n",
 959                 (unsigned long long)be32_to_cpu(page_0->length) * 4096);
 960 static BIN_ATTR_RO(catalog, 0/* real length varies */);
 961
 962 static struct bin_attribute *if_bin_attrs[] = {
 963         &bin_attr_catalog,
 964         NULL,
 965 };
 966
 967 static struct attribute *if_attrs[] = {
 968         &dev_attr_catalog_len.attr,
 969         &dev_attr_catalog_version.attr,
 970         NULL,
 971 };
 972
 973 static struct attribute_group if_group = {
 974         .name = "interface",
 975         .bin_attrs = if_bin_attrs,
 976         .attrs = if_attrs,
 977 };
 978
 979 static const struct attribute_group *attr_groups[] = {
 980         &format_group,
 981         &event_group,
 982         &event_desc_group,
 983         &event_long_desc_group,
 984         &if_group,
 985         NULL,
 986 };
 987
 988 static unsigned long single_24x7_request(u8 domain, u32 offset, u16 ix,
 989                                          u16 lpar, u64 *count,
 990                                          bool success_expected)
 991 {
 992         unsigned long ret;
 993
 994         struct hv_24x7_request_buffer *request_buffer;
 995         struct hv_24x7_data_result_buffer *result_buffer;
 996         struct hv_24x7_result *resb;
 997         struct hv_24x7_request *req;
 998
 999         BUILD_BUG_ON(sizeof(*request_buffer) > 4096);
1000         BUILD_BUG_ON(sizeof(*result_buffer) > 4096);
1001
1002         request_buffer = (void *)get_cpu_var(hv_24x7_reqb);
1003         result_buffer = (void *)get_cpu_var(hv_24x7_resb);
1004
1005         memset(request_buffer, 0, 4096);
1006         memset(result_buffer, 0, 4096);
1007
1008         request_buffer->interface_version = HV_24X7_IF_VERSION_CURRENT;
1009         request_buffer->num_requests = 1;
1010
1011         req = &request_buffer->requests[0];
1012
1013         req->performance_domain = domain;
1014         req->data_size = cpu_to_be16(8);
1015         req->data_offset = cpu_to_be32(offset);
1016         req->starting_lpar_ix = cpu_to_be16(lpar),
1017         req->max_num_lpars = cpu_to_be16(1);
1018         req->starting_ix = cpu_to_be16(ix);
1019         req->max_ix = cpu_to_be16(1);
1020
1021         /*
1022          * NOTE: Due to variable number of array elements in request and
1023          *       result buffer(s), sizeof() is not reliable. Use the actual
1024          *       allocated buffer size, H24x7_DATA_BUFFER_SIZE.
1025          */
1026         ret = plpar_hcall_norets(H_GET_24X7_DATA,
1027                         virt_to_phys(request_buffer), H24x7_DATA_BUFFER_SIZE,
1028                         virt_to_phys(result_buffer),  H24x7_DATA_BUFFER_SIZE);
1029
1030         if (ret) {
1031                 if (success_expected)
1032                         pr_err_ratelimited("hcall failed: %d %#x %#x %d => "
1033                                 "0x%lx (%ld) detail=0x%x failing ix=%x\n",
1034                                 domain, offset, ix, lpar, ret, ret,
1035                                 result_buffer->detailed_rc,
1036                                 result_buffer->failing_request_ix);
1037                 goto out;
1038         }
1039
1040         resb = &result_buffer->results[0];
1041
1042         *count = be64_to_cpu(resb->elements[0].element_data[0]);
1043 out:
1044         return ret;
1045 }
1046
1047 static unsigned long event_24x7_request(struct perf_event *event, u64 *res,
1048                 bool success_expected)
1049 {
1050         u16 idx;
1051         unsigned domain = event_get_domain(event);
1052
1053         if (is_physical_domain(domain))
1054                 idx = event_get_core(event);
1055         else
1056                 idx = event_get_vcpu(event);
1057
1058         return single_24x7_request(event_get_domain(event),
1059                                 event_get_offset(event),
1060                                 idx,
1061                                 event_get_lpar(event),
1062                                 res,
1063                                 success_expected);
1064 }
1065
1066 static int h_24x7_event_init(struct perf_event *event)
1067 {
1068         struct hv_perf_caps caps;
1069         unsigned domain;
1070         unsigned long hret;
1071         u64 ct;
1072
1073         /* Not our event */
1074         if (event->attr.type != event->pmu->type)
1075                 return -ENOENT;
1076
1077         /* Unused areas must be 0 */
1078         if (event_get_reserved1(event) ||
1079             event_get_reserved2(event) ||
1080             event_get_reserved3(event)) {
1081                 pr_devel("reserved set when forbidden 0x%llx(0x%llx) 0x%llx(0x%llx) 0x%llx(0x%llx)\n",
1082                                 event->attr.config,
1083                                 event_get_reserved1(event),
1084                                 event->attr.config1,
1085                                 event_get_reserved2(event),
1086                                 event->attr.config2,
1087                                 event_get_reserved3(event));
1088                 return -EINVAL;
1089         }
1090
1091         /* unsupported modes and filters */
1092         if (event->attr.exclude_user   ||
1093             event->attr.exclude_kernel ||
1094             event->attr.exclude_hv     ||
1095             event->attr.exclude_idle   ||
1096             event->attr.exclude_host   ||
1097             event->attr.exclude_guest)
1098                 return -EINVAL;
1099
1100         /* no branch sampling */
1101         if (has_branch_stack(event))
1102                 return -EOPNOTSUPP;
1103
1104         /* offset must be 8 byte aligned */
1105         if (event_get_offset(event) % 8) {
1106                 pr_devel("bad alignment\n");
1107                 return -EINVAL;
1108         }
1109
1110         /* Domains above 6 are invalid */
1111         domain = event_get_domain(event);
1112         if (domain > 6) {
1113                 pr_devel("invalid domain %d\n", domain);
1114                 return -EINVAL;
1115         }
1116
1117         hret = hv_perf_caps_get(&caps);
1118         if (hret) {
1119                 pr_devel("could not get capabilities: rc=%ld\n", hret);
1120                 return -EIO;
1121         }
1122
1123         /* Physical domains & other lpars require extra capabilities */
1124         if (!caps.collect_privileged && (is_physical_domain(domain) ||
1125                 (event_get_lpar(event) != event_get_lpar_max()))) {
1126                 pr_devel("hv permisions disallow: is_physical_domain:%d, lpar=0x%llx\n",
1127                                 is_physical_domain(domain),
1128                                 event_get_lpar(event));
1129                 return -EACCES;
1130         }
1131
1132         /* see if the event complains */
1133         if (event_24x7_request(event, &ct, false)) {
1134                 pr_devel("test hcall failed\n");
1135                 return -EIO;
1136         }
1137
1138         return 0;
1139 }
1140
1141 static u64 h_24x7_get_value(struct perf_event *event)
1142 {
1143         unsigned long ret;
1144         u64 ct;
1145         ret = event_24x7_request(event, &ct, true);
1146         if (ret)
1147                 /* We checked this in event init, shouldn't fail here... */
1148                 return 0;
1149
1150         return ct;
1151 }
1152
1153 static void h_24x7_event_update(struct perf_event *event)
1154 {
1155         s64 prev;
1156         u64 now;
1157         now = h_24x7_get_value(event);
1158         prev = local64_xchg(&event->hw.prev_count, now);
1159         local64_add(now - prev, &event->count);
1160 }
1161
1162 static void h_24x7_event_start(struct perf_event *event, int flags)
1163 {
1164         if (flags & PERF_EF_RELOAD)
1165                 local64_set(&event->hw.prev_count, h_24x7_get_value(event));
1166 }
1167
1168 static void h_24x7_event_stop(struct perf_event *event, int flags)
1169 {
1170         h_24x7_event_update(event);
1171 }
1172
1173 static int h_24x7_event_add(struct perf_event *event, int flags)
1174 {
1175         if (flags & PERF_EF_START)
1176                 h_24x7_event_start(event, flags);
1177
1178         return 0;
1179 }
1180
1181 static struct pmu h_24x7_pmu = {
1182         .task_ctx_nr = perf_invalid_context,
1183
1184         .name = "hv_24x7",
1185         .attr_groups = attr_groups,
1186         .event_init  = h_24x7_event_init,
1187         .add         = h_24x7_event_add,
1188         .del         = h_24x7_event_stop,
1189         .start       = h_24x7_event_start,
1190         .stop        = h_24x7_event_stop,
1191         .read        = h_24x7_event_update,
1192 };
1193
1194 static int hv_24x7_init(void)
1195 {
1196         int r;
1197         unsigned long hret;
1198         struct hv_perf_caps caps;
1199
1200         if (!firmware_has_feature(FW_FEATURE_LPAR)) {
1201                 pr_debug("not a virtualized system, not enabling\n");
1202                 return -ENODEV;
1203         }
1204
1205         hret = hv_perf_caps_get(&caps);
1206         if (hret) {
1207                 pr_debug("could not obtain capabilities, not enabling, rc=%ld\n",
1208                                 hret);
1209                 return -ENODEV;
1210         }
1211
1212         hv_page_cache = kmem_cache_create("hv-page-4096", 4096, 4096, 0, NULL);
1213         if (!hv_page_cache)
1214                 return -ENOMEM;
1215
1216         /* sampling not supported */
1217         h_24x7_pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
1218
1219         create_events_from_catalog(&event_group.attrs,
1220                                    &event_desc_group.attrs,
1221                                    &event_long_desc_group.attrs);
1222
1223         r = perf_pmu_register(&h_24x7_pmu, h_24x7_pmu.name, -1);
1224         if (r)
1225                 return r;
1226
1227         return 0;
1228 }
1229
1230 device_initcall(hv_24x7_init);