arch/ia64/sn/kernel/sn2/sn_hwperf.c

   1 /*
   2  * This file is subject to the terms and conditions of the GNU General Public
   3  * License.  See the file "COPYING" in the main directory of this archive
   4  * for more details.
   5  *
   6  * Copyright (C) 2004-2006 Silicon Graphics, Inc. All rights reserved.
   7  *
   8  * SGI Altix topology and hardware performance monitoring API.
   9  * Mark Goodwin <markgw@sgi.com>.
  10  *
  11  * Creates /proc/sgi_sn/sn_topology (read-only) to export
  12  * info about Altix nodes, routers, CPUs and NumaLink
  13  * interconnection/topology.
  14  *
  15  * Also creates a dynamic misc device named "sn_hwperf"
  16  * that supports an ioctl interface to call down into SAL
  17  * to discover hw objects, topology and to read/write
  18  * memory mapped registers, e.g. for performance monitoring.
  19  * The "sn_hwperf" device is registered only after the procfs
  20  * file is first opened, i.e. only if/when it's needed.
  21  *
  22  * This API is used by SGI Performance Co-Pilot and other
  23  * tools, see http://oss.sgi.com/projects/pcp
  24  */
  25
  26 #include <linux/fs.h>
  27 #include <linux/slab.h>
  28 #include <linux/export.h>
  29 #include <linux/vmalloc.h>
  30 #include <linux/seq_file.h>
  31 #include <linux/miscdevice.h>
  32 #include <linux/utsname.h>
  33 #include <linux/cpumask.h>
  34 #include <linux/nodemask.h>
  35 #include <linux/smp.h>
  36 #include <linux/mutex.h>
  37
  38 #include <asm/processor.h>
  39 #include <asm/topology.h>
  40 #include <linux/uaccess.h>
  41 #include <asm/sal.h>
  42 #include <asm/sn/io.h>
  43 #include <asm/sn/sn_sal.h>
  44 #include <asm/sn/module.h>
  45 #include <asm/sn/geo.h>
  46 #include <asm/sn/sn2/sn_hwperf.h>
  47 #include <asm/sn/addrs.h>
  48
  49 static void *sn_hwperf_salheap = NULL;
  50 static int sn_hwperf_obj_cnt = 0;
  51 static nasid_t sn_hwperf_master_nasid = INVALID_NASID;
  52 static int sn_hwperf_init(void);
  53 static DEFINE_MUTEX(sn_hwperf_init_mutex);
  54
  55 #define cnode_possible(n)       ((n) < num_cnodes)
  56
  57 static int sn_hwperf_enum_objects(int *nobj, struct sn_hwperf_object_info **ret)
  58 {
  59         int e;
  60         u64 sz;
  61         struct sn_hwperf_object_info *objbuf = NULL;
  62
  63         if ((e = sn_hwperf_init()) < 0) {
  64                 printk(KERN_ERR "sn_hwperf_init failed: err %d\n", e);
  65                 goto out;
  66         }
  67
  68         sz = sn_hwperf_obj_cnt * sizeof(struct sn_hwperf_object_info);
  69         objbuf = vmalloc(sz);
  70         if (objbuf == NULL) {
  71                 printk("sn_hwperf_enum_objects: vmalloc(%d) failed\n", (int)sz);
  72                 e = -ENOMEM;
  73                 goto out;
  74         }
  75
  76         e = ia64_sn_hwperf_op(sn_hwperf_master_nasid, SN_HWPERF_ENUM_OBJECTS,
  77                 0, sz, (u64) objbuf, 0, 0, NULL);
  78         if (e != SN_HWPERF_OP_OK) {
  79                 e = -EINVAL;
  80                 vfree(objbuf);
  81         }
  82
  83 out:
  84         *nobj = sn_hwperf_obj_cnt;
  85         *ret = objbuf;
  86         return e;
  87 }
  88
  89 static int sn_hwperf_location_to_bpos(char *location,
  90         int *rack, int *bay, int *slot, int *slab)
  91 {
  92         char type;
  93
  94         /* first scan for an old style geoid string */
  95         if (sscanf(location, "%03d%c%02d#%d",
  96                 rack, &type, bay, slab) == 4)
  97                 *slot = 0;
  98         else /* scan for a new bladed geoid string */
  99         if (sscanf(location, "%03d%c%02d^%02d#%d",
 100                 rack, &type, bay, slot, slab) != 5)
 101                 return -1;
 102         /* success */
 103         return 0;
 104 }
 105
 106 static int sn_hwperf_geoid_to_cnode(char *location)
 107 {
 108         int cnode;
 109         geoid_t geoid;
 110         moduleid_t module_id;
 111         int rack, bay, slot, slab;
 112         int this_rack, this_bay, this_slot, this_slab;
 113
 114         if (sn_hwperf_location_to_bpos(location, &rack, &bay, &slot, &slab))
 115                 return -1;
 116
 117         /*
 118          * FIXME: replace with cleaner for_each_XXX macro which addresses
 119          * both compute and IO nodes once ACPI3.0 is available.
 120          */
 121         for (cnode = 0; cnode < num_cnodes; cnode++) {
 122                 geoid = cnodeid_get_geoid(cnode);
 123                 module_id = geo_module(geoid);
 124                 this_rack = MODULE_GET_RACK(module_id);
 125                 this_bay = MODULE_GET_BPOS(module_id);
 126                 this_slot = geo_slot(geoid);
 127                 this_slab = geo_slab(geoid);
 128                 if (rack == this_rack && bay == this_bay &&
 129                         slot == this_slot && slab == this_slab) {
 130                         break;
 131                 }
 132         }
 133
 134         return cnode_possible(cnode) ? cnode : -1;
 135 }
 136
 137 static int sn_hwperf_obj_to_cnode(struct sn_hwperf_object_info * obj)
 138 {
 139         if (!SN_HWPERF_IS_NODE(obj) && !SN_HWPERF_IS_IONODE(obj))
 140                 BUG();
 141         if (SN_HWPERF_FOREIGN(obj))
 142                 return -1;
 143         return sn_hwperf_geoid_to_cnode(obj->location);
 144 }
 145
 146 static int sn_hwperf_generic_ordinal(struct sn_hwperf_object_info *obj,
 147                                 struct sn_hwperf_object_info *objs)
 148 {
 149         int ordinal;
 150         struct sn_hwperf_object_info *p;
 151
 152         for (ordinal=0, p=objs; p != obj; p++) {
 153                 if (SN_HWPERF_FOREIGN(p))
 154                         continue;
 155                 if (SN_HWPERF_SAME_OBJTYPE(p, obj))
 156                         ordinal++;
 157         }
 158
 159         return ordinal;
 160 }
 161
 162 static const char *slabname_node =      "node"; /* SHub asic */
 163 static const char *slabname_ionode =    "ionode"; /* TIO asic */
 164 static const char *slabname_router =    "router"; /* NL3R or NL4R */
 165 static const char *slabname_other =     "other"; /* unknown asic */
 166
 167 static const char *sn_hwperf_get_slabname(struct sn_hwperf_object_info *obj,
 168                         struct sn_hwperf_object_info *objs, int *ordinal)
 169 {
 170         int isnode;
 171         const char *slabname = slabname_other;
 172
 173         if ((isnode = SN_HWPERF_IS_NODE(obj)) || SN_HWPERF_IS_IONODE(obj)) {
 174                 slabname = isnode ? slabname_node : slabname_ionode;
 175                 *ordinal = sn_hwperf_obj_to_cnode(obj);
 176         }
 177         else {
 178                 *ordinal = sn_hwperf_generic_ordinal(obj, objs);
 179                 if (SN_HWPERF_IS_ROUTER(obj))
 180                         slabname = slabname_router;
 181         }
 182
 183         return slabname;
 184 }
 185
 186 static void print_pci_topology(struct seq_file *s)
 187 {
 188         char *p;
 189         size_t sz;
 190         int e;
 191
 192         for (sz = PAGE_SIZE; sz < 16 * PAGE_SIZE; sz += PAGE_SIZE) {
 193                 if (!(p = kmalloc(sz, GFP_KERNEL)))
 194                         break;
 195                 e = ia64_sn_ioif_get_pci_topology(__pa(p), sz);
 196                 if (e == SALRET_OK)
 197                         seq_puts(s, p);
 198                 kfree(p);
 199                 if (e == SALRET_OK || e == SALRET_NOT_IMPLEMENTED)
 200                         break;
 201         }
 202 }
 203
 204 static inline int sn_hwperf_has_cpus(cnodeid_t node)
 205 {
 206         return node < MAX_NUMNODES && node_online(node) && nr_cpus_node(node);
 207 }
 208
 209 static inline int sn_hwperf_has_mem(cnodeid_t node)
 210 {
 211         return node < MAX_NUMNODES && node_online(node) && NODE_DATA(node)->node_present_pages;
 212 }
 213
 214 static struct sn_hwperf_object_info *
 215 sn_hwperf_findobj_id(struct sn_hwperf_object_info *objbuf,
 216         int nobj, int id)
 217 {
 218         int i;
 219         struct sn_hwperf_object_info *p = objbuf;
 220
 221         for (i=0; i < nobj; i++, p++) {
 222                 if (p->id == id)
 223                         return p;
 224         }
 225
 226         return NULL;
 227
 228 }
 229
 230 static int sn_hwperf_get_nearest_node_objdata(struct sn_hwperf_object_info *objbuf,
 231         int nobj, cnodeid_t node, cnodeid_t *near_mem_node, cnodeid_t *near_cpu_node)
 232 {
 233         int e;
 234         struct sn_hwperf_object_info *nodeobj = NULL;
 235         struct sn_hwperf_object_info *op;
 236         struct sn_hwperf_object_info *dest;
 237         struct sn_hwperf_object_info *router;
 238         struct sn_hwperf_port_info ptdata[16];
 239         int sz, i, j;
 240         cnodeid_t c;
 241         int found_mem = 0;
 242         int found_cpu = 0;
 243
 244         if (!cnode_possible(node))
 245                 return -EINVAL;
 246
 247         if (sn_hwperf_has_cpus(node)) {
 248                 if (near_cpu_node)
 249                         *near_cpu_node = node;
 250                 found_cpu++;
 251         }
 252
 253         if (sn_hwperf_has_mem(node)) {
 254                 if (near_mem_node)
 255                         *near_mem_node = node;
 256                 found_mem++;
 257         }
 258
 259         if (found_cpu && found_mem)
 260                 return 0; /* trivially successful */
 261
 262         /* find the argument node object */
 263         for (i=0, op=objbuf; i < nobj; i++, op++) {
 264                 if (!SN_HWPERF_IS_NODE(op) && !SN_HWPERF_IS_IONODE(op))
 265                         continue;
 266                 if (node == sn_hwperf_obj_to_cnode(op)) {
 267                         nodeobj = op;
 268                         break;
 269                 }
 270         }
 271         if (!nodeobj) {
 272                 e = -ENOENT;
 273                 goto err;
 274         }
 275
 276         /* get it's interconnect topology */
 277         sz = op->ports * sizeof(struct sn_hwperf_port_info);
 278         BUG_ON(sz > sizeof(ptdata));
 279         e = ia64_sn_hwperf_op(sn_hwperf_master_nasid,
 280                               SN_HWPERF_ENUM_PORTS, nodeobj->id, sz,
 281                               (u64)&ptdata, 0, 0, NULL);
 282         if (e != SN_HWPERF_OP_OK) {
 283                 e = -EINVAL;
 284                 goto err;
 285         }
 286
 287         /* find nearest node with cpus and nearest memory */
 288         for (router=NULL, j=0; j < op->ports; j++) {
 289                 dest = sn_hwperf_findobj_id(objbuf, nobj, ptdata[j].conn_id);
 290                 if (dest && SN_HWPERF_IS_ROUTER(dest))
 291                         router = dest;
 292                 if (!dest || SN_HWPERF_FOREIGN(dest) ||
 293                     !SN_HWPERF_IS_NODE(dest) || SN_HWPERF_IS_IONODE(dest)) {
 294                         continue;
 295                 }
 296                 c = sn_hwperf_obj_to_cnode(dest);
 297                 if (!found_cpu && sn_hwperf_has_cpus(c)) {
 298                         if (near_cpu_node)
 299                                 *near_cpu_node = c;
 300                         found_cpu++;
 301                 }
 302                 if (!found_mem && sn_hwperf_has_mem(c)) {
 303                         if (near_mem_node)
 304                                 *near_mem_node = c;
 305                         found_mem++;
 306                 }
 307         }
 308
 309         if (router && (!found_cpu || !found_mem)) {
 310                 /* search for a node connected to the same router */
 311                 sz = router->ports * sizeof(struct sn_hwperf_port_info);
 312                 BUG_ON(sz > sizeof(ptdata));
 313                 e = ia64_sn_hwperf_op(sn_hwperf_master_nasid,
 314                                       SN_HWPERF_ENUM_PORTS, router->id, sz,
 315                                       (u64)&ptdata, 0, 0, NULL);
 316                 if (e != SN_HWPERF_OP_OK) {
 317                         e = -EINVAL;
 318                         goto err;
 319                 }
 320                 for (j=0; j < router->ports; j++) {
 321                         dest = sn_hwperf_findobj_id(objbuf, nobj,
 322                                 ptdata[j].conn_id);
 323                         if (!dest || dest->id == node ||
 324                             SN_HWPERF_FOREIGN(dest) ||
 325                             !SN_HWPERF_IS_NODE(dest) ||
 326                             SN_HWPERF_IS_IONODE(dest)) {
 327                                 continue;
 328                         }
 329                         c = sn_hwperf_obj_to_cnode(dest);
 330                         if (!found_cpu && sn_hwperf_has_cpus(c)) {
 331                                 if (near_cpu_node)
 332                                         *near_cpu_node = c;
 333                                 found_cpu++;
 334                         }
 335                         if (!found_mem && sn_hwperf_has_mem(c)) {
 336                                 if (near_mem_node)
 337                                         *near_mem_node = c;
 338                                 found_mem++;
 339                         }
 340                         if (found_cpu && found_mem)
 341                                 break;
 342                 }
 343         }
 344
 345         if (!found_cpu || !found_mem) {
 346                 /* resort to _any_ node with CPUs and memory */
 347                 for (i=0, op=objbuf; i < nobj; i++, op++) {
 348                         if (SN_HWPERF_FOREIGN(op) ||
 349                             SN_HWPERF_IS_IONODE(op) ||
 350                             !SN_HWPERF_IS_NODE(op)) {
 351                                 continue;
 352                         }
 353                         c = sn_hwperf_obj_to_cnode(op);
 354                         if (!found_cpu && sn_hwperf_has_cpus(c)) {
 355                                 if (near_cpu_node)
 356                                         *near_cpu_node = c;
 357                                 found_cpu++;
 358                         }
 359                         if (!found_mem && sn_hwperf_has_mem(c)) {
 360                                 if (near_mem_node)
 361                                         *near_mem_node = c;
 362                                 found_mem++;
 363                         }
 364                         if (found_cpu && found_mem)
 365                                 break;
 366                 }
 367         }
 368
 369         if (!found_cpu || !found_mem)
 370                 e = -ENODATA;
 371
 372 err:
 373         return e;
 374 }
 375
 376
 377 static int sn_topology_show(struct seq_file *s, void *d)
 378 {
 379         int sz;
 380         int pt;
 381         int e = 0;
 382         int i;
 383         int j;
 384         const char *slabname;
 385         int ordinal;
 386         char slice;
 387         struct cpuinfo_ia64 *c;
 388         struct sn_hwperf_port_info *ptdata;
 389         struct sn_hwperf_object_info *p;
 390         struct sn_hwperf_object_info *obj = d;  /* this object */
 391         struct sn_hwperf_object_info *objs = s->private; /* all objects */
 392         u8 shubtype;
 393         u8 system_size;
 394         u8 sharing_size;
 395         u8 partid;
 396         u8 coher;
 397         u8 nasid_shift;
 398         u8 region_size;
 399         u16 nasid_mask;
 400         int nasid_msb;
 401
 402         if (obj == objs) {
 403                 seq_printf(s, "# sn_topology version 2\n");
 404                 seq_printf(s, "# objtype ordinal location partition"
 405                         " [attribute value [, ...]]\n");
 406
 407                 if (ia64_sn_get_sn_info(0,
 408                         &shubtype, &nasid_mask, &nasid_shift, &system_size,
 409                         &sharing_size, &partid, &coher, &region_size))
 410                         BUG();
 411                 for (nasid_msb=63; nasid_msb > 0; nasid_msb--) {
 412                         if (((u64)nasid_mask << nasid_shift) & (1ULL << nasid_msb))
 413                                 break;
 414                 }
 415                 seq_printf(s, "partition %u %s local "
 416                         "shubtype %s, "
 417                         "nasid_mask 0x%016llx, "
 418                         "nasid_bits %d:%d, "
 419                         "system_size %d, "
 420                         "sharing_size %d, "
 421                         "coherency_domain %d, "
 422                         "region_size %d\n",
 423
 424                         partid, utsname()->nodename,
 425                         shubtype ? "shub2" : "shub1",
 426                         (u64)nasid_mask << nasid_shift, nasid_msb, nasid_shift,
 427                         system_size, sharing_size, coher, region_size);
 428
 429                 print_pci_topology(s);
 430         }
 431
 432         if (SN_HWPERF_FOREIGN(obj)) {
 433                 /* private in another partition: not interesting */
 434                 return 0;
 435         }
 436
 437         for (i = 0; i < SN_HWPERF_MAXSTRING && obj->name[i]; i++) {
 438                 if (obj->name[i] == ' ')
 439                         obj->name[i] = '_';
 440         }
 441
 442         slabname = sn_hwperf_get_slabname(obj, objs, &ordinal);
 443         seq_printf(s, "%s %d %s %s asic %s", slabname, ordinal, obj->location,
 444                 obj->sn_hwp_this_part ? "local" : "shared", obj->name);
 445
 446         if (ordinal < 0 || (!SN_HWPERF_IS_NODE(obj) && !SN_HWPERF_IS_IONODE(obj)))
 447                 seq_putc(s, '\n');
 448         else {
 449                 cnodeid_t near_mem = -1;
 450                 cnodeid_t near_cpu = -1;
 451
 452                 seq_printf(s, ", nasid 0x%x", cnodeid_to_nasid(ordinal));
 453
 454                 if (sn_hwperf_get_nearest_node_objdata(objs, sn_hwperf_obj_cnt,
 455                         ordinal, &near_mem, &near_cpu) == 0) {
 456                         seq_printf(s, ", near_mem_nodeid %d, near_cpu_nodeid %d",
 457                                 near_mem, near_cpu);
 458                 }
 459
 460                 if (!SN_HWPERF_IS_IONODE(obj)) {
 461                         for_each_online_node(i) {
 462                                 seq_printf(s, i ? ":%d" : ", dist %d",
 463                                         node_distance(ordinal, i));
 464                         }
 465                 }
 466
 467                 seq_putc(s, '\n');
 468
 469                 /*
 470                  * CPUs on this node, if any
 471                  */
 472                 if (!SN_HWPERF_IS_IONODE(obj)) {
 473                         for_each_cpu_and(i, cpu_online_mask,
 474                                          cpumask_of_node(ordinal)) {
 475                                 slice = 'a' + cpuid_to_slice(i);
 476                                 c = cpu_data(i);
 477                                 seq_printf(s, "cpu %d %s%c local"
 478                                            " freq %luMHz, arch ia64",
 479                                            i, obj->location, slice,
 480                                            c->proc_freq / 1000000);
 481                                 for_each_online_cpu(j) {
 482                                         seq_printf(s, j ? ":%d" : ", dist %d",
 483                                                    node_distance(
 484                                                         cpu_to_node(i),
 485                                                         cpu_to_node(j)));
 486                                 }
 487                                 seq_putc(s, '\n');
 488                         }
 489                 }
 490         }
 491
 492         if (obj->ports) {
 493                 /*
 494                  * numalink ports
 495                  */
 496                 sz = obj->ports * sizeof(struct sn_hwperf_port_info);
 497                 if ((ptdata = kmalloc(sz, GFP_KERNEL)) == NULL)
 498                         return -ENOMEM;
 499                 e = ia64_sn_hwperf_op(sn_hwperf_master_nasid,
 500                                       SN_HWPERF_ENUM_PORTS, obj->id, sz,
 501                                       (u64) ptdata, 0, 0, NULL);
 502                 if (e != SN_HWPERF_OP_OK)
 503                         return -EINVAL;
 504                 for (ordinal=0, p=objs; p != obj; p++) {
 505                         if (!SN_HWPERF_FOREIGN(p))
 506                                 ordinal += p->ports;
 507                 }
 508                 for (pt = 0; pt < obj->ports; pt++) {
 509                         for (p = objs, i = 0; i < sn_hwperf_obj_cnt; i++, p++) {
 510                                 if (ptdata[pt].conn_id == p->id) {
 511                                         break;
 512                                 }
 513                         }
 514                         seq_printf(s, "numalink %d %s-%d",
 515                             ordinal+pt, obj->location, ptdata[pt].port);
 516
 517                         if (i >= sn_hwperf_obj_cnt) {
 518                                 /* no connection */
 519                                 seq_puts(s, " local endpoint disconnected"
 520                                             ", protocol unknown\n");
 521                                 continue;
 522                         }
 523
 524                         if (obj->sn_hwp_this_part && p->sn_hwp_this_part)
 525                                 /* both ends local to this partition */
 526                                 seq_puts(s, " local");
 527                         else if (SN_HWPERF_FOREIGN(p))
 528                                 /* both ends of the link in foreign partition */
 529                                 seq_puts(s, " foreign");
 530                         else
 531                                 /* link straddles a partition */
 532                                 seq_puts(s, " shared");
 533
 534                         /*
 535                          * Unlikely, but strictly should query the LLP config
 536                          * registers because an NL4R can be configured to run
 537                          * NL3 protocol, even when not talking to an NL3 router.
 538                          * Ditto for node-node.
 539                          */
 540                         seq_printf(s, " endpoint %s-%d, protocol %s\n",
 541                                 p->location, ptdata[pt].conn_port,
 542                                 (SN_HWPERF_IS_NL3ROUTER(obj) ||
 543                                 SN_HWPERF_IS_NL3ROUTER(p)) ?  "LLP3" : "LLP4");
 544                 }
 545                 kfree(ptdata);
 546         }
 547
 548         return 0;
 549 }
 550
 551 static void *sn_topology_start(struct seq_file *s, loff_t * pos)
 552 {
 553         struct sn_hwperf_object_info *objs = s->private;
 554
 555         if (*pos < sn_hwperf_obj_cnt)
 556                 return (void *)(objs + *pos);
 557
 558         return NULL;
 559 }
 560
 561 static void *sn_topology_next(struct seq_file *s, void *v, loff_t * pos)
 562 {
 563         ++*pos;
 564         return sn_topology_start(s, pos);
 565 }
 566
 567 static void sn_topology_stop(struct seq_file *m, void *v)
 568 {
 569         return;
 570 }
 571
 572 /*
 573  * /proc/sgi_sn/sn_topology, read-only using seq_file
 574  */
 575 static const struct seq_operations sn_topology_seq_ops = {
 576         .start = sn_topology_start,
 577         .next = sn_topology_next,
 578         .stop = sn_topology_stop,
 579         .show = sn_topology_show
 580 };
 581
 582 struct sn_hwperf_op_info {
 583         u64 op;
 584         struct sn_hwperf_ioctl_args *a;
 585         void *p;
 586         int *v0;
 587         int ret;
 588 };
 589
 590 static void sn_hwperf_call_sal(void *info)
 591 {
 592         struct sn_hwperf_op_info *op_info = info;
 593         int r;
 594
 595         r = ia64_sn_hwperf_op(sn_hwperf_master_nasid, op_info->op,
 596                       op_info->a->arg, op_info->a->sz,
 597                       (u64) op_info->p, 0, 0, op_info->v0);
 598         op_info->ret = r;
 599 }
 600
 601 static int sn_hwperf_op_cpu(struct sn_hwperf_op_info *op_info)
 602 {
 603         u32 cpu;
 604         u32 use_ipi;
 605         int r = 0;
 606         cpumask_t save_allowed;
 607
 608         cpu = (op_info->a->arg & SN_HWPERF_ARG_CPU_MASK) >> 32;
 609         use_ipi = op_info->a->arg & SN_HWPERF_ARG_USE_IPI_MASK;
 610         op_info->a->arg &= SN_HWPERF_ARG_OBJID_MASK;
 611
 612         if (cpu != SN_HWPERF_ARG_ANY_CPU) {
 613                 if (cpu >= nr_cpu_ids || !cpu_online(cpu)) {
 614                         r = -EINVAL;
 615                         goto out;
 616                 }
 617         }
 618
 619         if (cpu == SN_HWPERF_ARG_ANY_CPU) {
 620                 /* don't care which cpu */
 621                 sn_hwperf_call_sal(op_info);
 622         } else if (cpu == get_cpu()) {
 623                 /* already on correct cpu */
 624                 sn_hwperf_call_sal(op_info);
 625                 put_cpu();
 626         } else {
 627                 put_cpu();
 628                 if (use_ipi) {
 629                         /* use an interprocessor interrupt to call SAL */
 630                         smp_call_function_single(cpu, sn_hwperf_call_sal,
 631                                 op_info, 1);
 632                 }
 633                 else {
 634                         /* migrate the task before calling SAL */
 635                         save_allowed = current->cpus_allowed;
 636                         set_cpus_allowed_ptr(current, cpumask_of(cpu));
 637                         sn_hwperf_call_sal(op_info);
 638                         set_cpus_allowed_ptr(current, &save_allowed);
 639                 }
 640         }
 641         r = op_info->ret;
 642
 643 out:
 644         return r;
 645 }
 646
 647 /* map SAL hwperf error code to system error code */
 648 static int sn_hwperf_map_err(int hwperf_err)
 649 {
 650         int e;
 651
 652         switch(hwperf_err) {
 653         case SN_HWPERF_OP_OK:
 654                 e = 0;
 655                 break;
 656
 657         case SN_HWPERF_OP_NOMEM:
 658                 e = -ENOMEM;
 659                 break;
 660
 661         case SN_HWPERF_OP_NO_PERM:
 662                 e = -EPERM;
 663                 break;
 664
 665         case SN_HWPERF_OP_IO_ERROR:
 666                 e = -EIO;
 667                 break;
 668
 669         case SN_HWPERF_OP_BUSY:
 670                 e = -EBUSY;
 671                 break;
 672
 673         case SN_HWPERF_OP_RECONFIGURE:
 674                 e = -EAGAIN;
 675                 break;
 676
 677         case SN_HWPERF_OP_INVAL:
 678         default:
 679                 e = -EINVAL;
 680                 break;
 681         }
 682
 683         return e;
 684 }
 685
 686 /*
 687  * ioctl for "sn_hwperf" misc device
 688  */
 689 static long sn_hwperf_ioctl(struct file *fp, u32 op, unsigned long arg)
 690 {
 691         struct sn_hwperf_ioctl_args a;
 692         struct cpuinfo_ia64 *cdata;
 693         struct sn_hwperf_object_info *objs;
 694         struct sn_hwperf_object_info *cpuobj;
 695         struct sn_hwperf_op_info op_info;
 696         void *p = NULL;
 697         int nobj;
 698         char slice;
 699         int node;
 700         int r;
 701         int v0;
 702         int i;
 703         int j;
 704
 705         /* only user requests are allowed here */
 706         if ((op & SN_HWPERF_OP_MASK) < 10) {
 707                 r = -EINVAL;
 708                 goto error;
 709         }
 710         r = copy_from_user(&a, (const void __user *)arg,
 711                 sizeof(struct sn_hwperf_ioctl_args));
 712         if (r != 0) {
 713                 r = -EFAULT;
 714                 goto error;
 715         }
 716
 717         /*
 718          * Allocate memory to hold a kernel copy of the user buffer. The
 719          * buffer contents are either copied in or out (or both) of user
 720          * space depending on the flags encoded in the requested operation.
 721          */
 722         if (a.ptr) {
 723                 p = vmalloc(a.sz);
 724                 if (!p) {
 725                         r = -ENOMEM;
 726                         goto error;
 727                 }
 728         }
 729
 730         if (op & SN_HWPERF_OP_MEM_COPYIN) {
 731                 r = copy_from_user(p, (const void __user *)a.ptr, a.sz);
 732                 if (r != 0) {
 733                         r = -EFAULT;
 734                         goto error;
 735                 }
 736         }
 737
 738         switch (op) {
 739         case SN_HWPERF_GET_CPU_INFO:
 740                 if (a.sz == sizeof(u64)) {
 741                         /* special case to get size needed */
 742                         *(u64 *) p = (u64) num_online_cpus() *
 743                                 sizeof(struct sn_hwperf_object_info);
 744                 } else
 745                 if (a.sz < num_online_cpus() * sizeof(struct sn_hwperf_object_info)) {
 746                         r = -ENOMEM;
 747                         goto error;
 748                 } else
 749                 if ((r = sn_hwperf_enum_objects(&nobj, &objs)) == 0) {
 750                         int cpuobj_index = 0;
 751
 752                         memset(p, 0, a.sz);
 753                         for (i = 0; i < nobj; i++) {
 754                                 if (!SN_HWPERF_IS_NODE(objs + i))
 755                                         continue;
 756                                 node = sn_hwperf_obj_to_cnode(objs + i);
 757                                 for_each_online_cpu(j) {
 758                                         if (node != cpu_to_node(j))
 759                                                 continue;
 760                                         cpuobj = (struct sn_hwperf_object_info *) p + cpuobj_index++;
 761                                         slice = 'a' + cpuid_to_slice(j);
 762                                         cdata = cpu_data(j);
 763                                         cpuobj->id = j;
 764                                         snprintf(cpuobj->name,
 765                                                  sizeof(cpuobj->name),
 766                                                  "CPU %luMHz %s",
 767                                                  cdata->proc_freq / 1000000,
 768                                                  cdata->vendor);
 769                                         snprintf(cpuobj->location,
 770                                                  sizeof(cpuobj->location),
 771                                                  "%s%c", objs[i].location,
 772                                                  slice);
 773                                 }
 774                         }
 775
 776                         vfree(objs);
 777                 }
 778                 break;
 779
 780         case SN_HWPERF_GET_NODE_NASID:
 781                 if (a.sz != sizeof(u64) ||
 782                    (node = a.arg) < 0 || !cnode_possible(node)) {
 783                         r = -EINVAL;
 784                         goto error;
 785                 }
 786                 *(u64 *)p = (u64)cnodeid_to_nasid(node);
 787                 break;
 788
 789         case SN_HWPERF_GET_OBJ_NODE:
 790                 i = a.arg;
 791                 if (a.sz != sizeof(u64) || i < 0) {
 792                         r = -EINVAL;
 793                         goto error;
 794                 }
 795                 if ((r = sn_hwperf_enum_objects(&nobj, &objs)) == 0) {
 796                         if (i >= nobj) {
 797                                 r = -EINVAL;
 798                                 vfree(objs);
 799                                 goto error;
 800                         }
 801                         if (objs[i].id != a.arg) {
 802                                 for (i = 0; i < nobj; i++) {
 803                                         if (objs[i].id == a.arg)
 804                                                 break;
 805                                 }
 806                         }
 807                         if (i == nobj) {
 808                                 r = -EINVAL;
 809                                 vfree(objs);
 810                                 goto error;
 811                         }
 812
 813                         if (!SN_HWPERF_IS_NODE(objs + i) &&
 814                             !SN_HWPERF_IS_IONODE(objs + i)) {
 815                                 r = -ENOENT;
 816                                 vfree(objs);
 817                                 goto error;
 818                         }
 819
 820                         *(u64 *)p = (u64)sn_hwperf_obj_to_cnode(objs + i);
 821                         vfree(objs);
 822                 }
 823                 break;
 824
 825         case SN_HWPERF_GET_MMRS:
 826         case SN_HWPERF_SET_MMRS:
 827         case SN_HWPERF_OBJECT_DISTANCE:
 828                 op_info.p = p;
 829                 op_info.a = &a;
 830                 op_info.v0 = &v0;
 831                 op_info.op = op;
 832                 r = sn_hwperf_op_cpu(&op_info);
 833                 if (r) {
 834                         r = sn_hwperf_map_err(r);
 835                         a.v0 = v0;
 836                         goto error;
 837                 }
 838                 break;
 839
 840         default:
 841                 /* all other ops are a direct SAL call */
 842                 r = ia64_sn_hwperf_op(sn_hwperf_master_nasid, op,
 843                               a.arg, a.sz, (u64) p, 0, 0, &v0);
 844                 if (r) {
 845                         r = sn_hwperf_map_err(r);
 846                         goto error;
 847                 }
 848                 a.v0 = v0;
 849                 break;
 850         }
 851
 852         if (op & SN_HWPERF_OP_MEM_COPYOUT) {
 853                 r = copy_to_user((void __user *)a.ptr, p, a.sz);
 854                 if (r != 0) {
 855                         r = -EFAULT;
 856                         goto error;
 857                 }
 858         }
 859
 860 error:
 861         vfree(p);
 862
 863         return r;
 864 }
 865
 866 static const struct file_operations sn_hwperf_fops = {
 867         .unlocked_ioctl = sn_hwperf_ioctl,
 868         .llseek = noop_llseek,
 869 };
 870
 871 static struct miscdevice sn_hwperf_dev = {
 872         MISC_DYNAMIC_MINOR,
 873         "sn_hwperf",
 874         &sn_hwperf_fops
 875 };
 876
 877 static int sn_hwperf_init(void)
 878 {
 879         u64 v;
 880         int salr;
 881         int e = 0;
 882
 883         /* single threaded, once-only initialization */
 884         mutex_lock(&sn_hwperf_init_mutex);
 885
 886         if (sn_hwperf_salheap) {
 887                 mutex_unlock(&sn_hwperf_init_mutex);
 888                 return e;
 889         }
 890
 891         /*
 892          * The PROM code needs a fixed reference node. For convenience the
 893          * same node as the console I/O is used.
 894          */
 895         sn_hwperf_master_nasid = (nasid_t) ia64_sn_get_console_nasid();
 896
 897         /*
 898          * Request the needed size and install the PROM scratch area.
 899          * The PROM keeps various tracking bits in this memory area.
 900          */
 901         salr = ia64_sn_hwperf_op(sn_hwperf_master_nasid,
 902                                  (u64) SN_HWPERF_GET_HEAPSIZE, 0,
 903                                  (u64) sizeof(u64), (u64) &v, 0, 0, NULL);
 904         if (salr != SN_HWPERF_OP_OK) {
 905                 e = -EINVAL;
 906                 goto out;
 907         }
 908
 909         if ((sn_hwperf_salheap = vmalloc(v)) == NULL) {
 910                 e = -ENOMEM;
 911                 goto out;
 912         }
 913         salr = ia64_sn_hwperf_op(sn_hwperf_master_nasid,
 914                                  SN_HWPERF_INSTALL_HEAP, 0, v,
 915                                  (u64) sn_hwperf_salheap, 0, 0, NULL);
 916         if (salr != SN_HWPERF_OP_OK) {
 917                 e = -EINVAL;
 918                 goto out;
 919         }
 920
 921         salr = ia64_sn_hwperf_op(sn_hwperf_master_nasid,
 922                                  SN_HWPERF_OBJECT_COUNT, 0,
 923                                  sizeof(u64), (u64) &v, 0, 0, NULL);
 924         if (salr != SN_HWPERF_OP_OK) {
 925                 e = -EINVAL;
 926                 goto out;
 927         }
 928         sn_hwperf_obj_cnt = (int)v;
 929
 930 out:
 931         if (e < 0 && sn_hwperf_salheap) {
 932                 vfree(sn_hwperf_salheap);
 933                 sn_hwperf_salheap = NULL;
 934                 sn_hwperf_obj_cnt = 0;
 935         }
 936         mutex_unlock(&sn_hwperf_init_mutex);
 937         return e;
 938 }
 939
 940 int sn_topology_open(struct inode *inode, struct file *file)
 941 {
 942         int e;
 943         struct seq_file *seq;
 944         struct sn_hwperf_object_info *objbuf;
 945         int nobj;
 946
 947         if ((e = sn_hwperf_enum_objects(&nobj, &objbuf)) == 0) {
 948                 e = seq_open(file, &sn_topology_seq_ops);
 949                 seq = file->private_data;
 950                 seq->private = objbuf;
 951         }
 952
 953         return e;
 954 }
 955
 956 int sn_topology_release(struct inode *inode, struct file *file)
 957 {
 958         struct seq_file *seq = file->private_data;
 959
 960         vfree(seq->private);
 961         return seq_release(inode, file);
 962 }
 963
 964 int sn_hwperf_get_nearest_node(cnodeid_t node,
 965         cnodeid_t *near_mem_node, cnodeid_t *near_cpu_node)
 966 {
 967         int e;
 968         int nobj;
 969         struct sn_hwperf_object_info *objbuf;
 970
 971         if ((e = sn_hwperf_enum_objects(&nobj, &objbuf)) == 0) {
 972                 e = sn_hwperf_get_nearest_node_objdata(objbuf, nobj,
 973                         node, near_mem_node, near_cpu_node);
 974                 vfree(objbuf);
 975         }
 976
 977         return e;
 978 }
 979
 980 static int sn_hwperf_misc_register_init(void)
 981 {
 982         int e;
 983
 984         if (!ia64_platform_is("sn2"))
 985                 return 0;
 986
 987         sn_hwperf_init();
 988
 989         /*
 990          * Register a dynamic misc device for hwperf ioctls. Platforms
 991          * supporting hotplug will create /dev/sn_hwperf, else user
 992          * can to look up the minor number in /proc/misc.
 993          */
 994         if ((e = misc_register(&sn_hwperf_dev)) != 0) {
 995                 printk(KERN_ERR "sn_hwperf_misc_register_init: failed to "
 996                 "register misc device for \"%s\"\n", sn_hwperf_dev.name);
 997         }
 998
 999         return e;
1000 }
1001
1002 device_initcall(sn_hwperf_misc_register_init); /* after misc_init() */
1003 EXPORT_SYMBOL(sn_hwperf_get_nearest_node);