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