err.no Git - linux-2.6/blob - 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) 2004-2005 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/smp_lock.h>
  34 #include <linux/nodemask.h>
  35 #include <asm/processor.h>
  36 #include <asm/topology.h>
  37 #include <asm/smp.h>
  38 #include <asm/semaphore.h>
  39 #include <asm/segment.h>
  40 #include <asm/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 DECLARE_MUTEX(sn_hwperf_init_mutex);
  54
  55 static int sn_hwperf_enum_objects(int *nobj, struct sn_hwperf_object_info **ret)
  56 {
  57         int e;
  58         u64 sz;
  59         struct sn_hwperf_object_info *objbuf = NULL;
  60
  61         if ((e = sn_hwperf_init()) < 0) {
  62                 printk(KERN_ERR "sn_hwperf_init failed: err %d\n", e);
  63                 goto out;
  64         }
  65
  66         sz = sn_hwperf_obj_cnt * sizeof(struct sn_hwperf_object_info);
  67         if ((objbuf = (struct sn_hwperf_object_info *) vmalloc(sz)) == NULL) {
  68                 printk("sn_hwperf_enum_objects: vmalloc(%d) failed\n", (int)sz);
  69                 e = -ENOMEM;
  70                 goto out;
  71         }
  72
  73         e = ia64_sn_hwperf_op(sn_hwperf_master_nasid, SN_HWPERF_ENUM_OBJECTS,
  74                 0, sz, (u64) objbuf, 0, 0, NULL);
  75         if (e != SN_HWPERF_OP_OK) {
  76                 e = -EINVAL;
  77                 vfree(objbuf);
  78         }
  79
  80 out:
  81         *nobj = sn_hwperf_obj_cnt;
  82         *ret = objbuf;
  83         return e;
  84 }
  85
  86 static int sn_hwperf_location_to_bpos(char *location,
  87         int *rack, int *bay, int *slot, int *slab)
  88 {
  89         char type;
  90
  91         /* first scan for an old style geoid string */
  92         if (sscanf(location, "%03d%c%02d#%d",
  93                 rack, &type, bay, slab) == 4)
  94                 *slot = 0;
  95         else /* scan for a new bladed geoid string */
  96         if (sscanf(location, "%03d%c%02d^%02d#%d",
  97                 rack, &type, bay, slot, slab) != 5)
  98                 return -1;
  99         /* success */
 100         return 0;
 101 }
 102
 103 static int sn_hwperf_geoid_to_cnode(char *location)
 104 {
 105         int cnode;
 106         geoid_t geoid;
 107         moduleid_t module_id;
 108         int rack, bay, slot, slab;
 109         int this_rack, this_bay, this_slot, this_slab;
 110
 111         if (sn_hwperf_location_to_bpos(location, &rack, &bay, &slot, &slab))
 112                 return -1;
 113
 114         for_each_node(cnode) {
 115                 geoid = cnodeid_get_geoid(cnode);
 116                 module_id = geo_module(geoid);
 117                 this_rack = MODULE_GET_RACK(module_id);
 118                 this_bay = MODULE_GET_BPOS(module_id);
 119                 this_slot = geo_slot(geoid);
 120                 this_slab = geo_slab(geoid);
 121                 if (rack == this_rack && bay == this_bay &&
 122                         slot == this_slot && slab == this_slab) {
 123                         break;
 124                 }
 125         }
 126
 127         return node_possible(cnode) ? cnode : -1;
 128 }
 129
 130 static int sn_hwperf_obj_to_cnode(struct sn_hwperf_object_info * obj)
 131 {
 132         if (!SN_HWPERF_IS_NODE(obj) && !SN_HWPERF_IS_IONODE(obj))
 133                 BUG();
 134         if (!obj->sn_hwp_this_part)
 135                 return -1;
 136         return sn_hwperf_geoid_to_cnode(obj->location);
 137 }
 138
 139 static int sn_hwperf_generic_ordinal(struct sn_hwperf_object_info *obj,
 140                                 struct sn_hwperf_object_info *objs)
 141 {
 142         int ordinal;
 143         struct sn_hwperf_object_info *p;
 144
 145         for (ordinal=0, p=objs; p != obj; p++) {
 146                 if (SN_HWPERF_FOREIGN(p))
 147                         continue;
 148                 if (SN_HWPERF_SAME_OBJTYPE(p, obj))
 149                         ordinal++;
 150         }
 151
 152         return ordinal;
 153 }
 154
 155 static const char *slabname_node =      "node"; /* SHub asic */
 156 static const char *slabname_ionode =    "ionode"; /* TIO asic */
 157 static const char *slabname_router =    "router"; /* NL3R or NL4R */
 158 static const char *slabname_other =     "other"; /* unknown asic */
 159
 160 static const char *sn_hwperf_get_slabname(struct sn_hwperf_object_info *obj,
 161                         struct sn_hwperf_object_info *objs, int *ordinal)
 162 {
 163         int isnode;
 164         const char *slabname = slabname_other;
 165
 166         if ((isnode = SN_HWPERF_IS_NODE(obj)) || SN_HWPERF_IS_IONODE(obj)) {
 167                 slabname = isnode ? slabname_node : slabname_ionode;
 168                 *ordinal = sn_hwperf_obj_to_cnode(obj);
 169         }
 170         else {
 171                 *ordinal = sn_hwperf_generic_ordinal(obj, objs);
 172                 if (SN_HWPERF_IS_ROUTER(obj))
 173                         slabname = slabname_router;
 174         }
 175
 176         return slabname;
 177 }
 178
 179 static void print_pci_topology(struct seq_file *s)
 180 {
 181         char *p;
 182         size_t sz;
 183         int e;
 184
 185         for (sz = PAGE_SIZE; sz < 16 * PAGE_SIZE; sz += PAGE_SIZE) {
 186                 if (!(p = (char *)kmalloc(sz, GFP_KERNEL)))
 187                         break;
 188                 e = ia64_sn_ioif_get_pci_topology(__pa(p), sz);
 189                 if (e == SALRET_OK)
 190                         seq_puts(s, p);
 191                 kfree(p);
 192                 if (e == SALRET_OK || e == SALRET_NOT_IMPLEMENTED)
 193                         break;
 194         }
 195 }
 196
 197 static inline int sn_hwperf_has_cpus(cnodeid_t node)
 198 {
 199         return node_online(node) && nr_cpus_node(node);
 200 }
 201
 202 static inline int sn_hwperf_has_mem(cnodeid_t node)
 203 {
 204         return node_online(node) && NODE_DATA(node)->node_present_pages;
 205 }
 206
 207 static struct sn_hwperf_object_info *
 208 sn_hwperf_findobj_id(struct sn_hwperf_object_info *objbuf,
 209         int nobj, int id)
 210 {
 211         int i;
 212         struct sn_hwperf_object_info *p = objbuf;
 213
 214         for (i=0; i < nobj; i++, p++) {
 215                 if (p->id == id)
 216                         return p;
 217         }
 218
 219         return NULL;
 220
 221 }
 222
 223 static int sn_hwperf_get_nearest_node_objdata(struct sn_hwperf_object_info *objbuf,
 224         int nobj, cnodeid_t node, cnodeid_t *near_mem_node, cnodeid_t *near_cpu_node)
 225 {
 226         int e;
 227         struct sn_hwperf_object_info *nodeobj = NULL;
 228         struct sn_hwperf_object_info *op;
 229         struct sn_hwperf_object_info *dest;
 230         struct sn_hwperf_object_info *router;
 231         struct sn_hwperf_port_info ptdata[16];
 232         int sz, i, j;
 233         cnodeid_t c;
 234         int found_mem = 0;
 235         int found_cpu = 0;
 236
 237         if (!node_possible(node))
 238                 return -EINVAL;
 239
 240         if (sn_hwperf_has_cpus(node)) {
 241                 if (near_cpu_node)
 242                         *near_cpu_node = node;
 243                 found_cpu++;
 244         }
 245
 246         if (sn_hwperf_has_mem(node)) {
 247                 if (near_mem_node)
 248                         *near_mem_node = node;
 249                 found_mem++;
 250         }
 251
 252         if (found_cpu && found_mem)
 253                 return 0; /* trivially successful */
 254
 255         /* find the argument node object */
 256         for (i=0, op=objbuf; i < nobj; i++, op++) {
 257                 if (!SN_HWPERF_IS_NODE(op) && !SN_HWPERF_IS_IONODE(op))
 258                         continue;
 259                 if (node == sn_hwperf_obj_to_cnode(op)) {
 260                         nodeobj = op;
 261                         break;
 262                 }
 263         }
 264         if (!nodeobj) {
 265                 e = -ENOENT;
 266                 goto err;
 267         }
 268
 269         /* get it's interconnect topology */
 270         sz = op->ports * sizeof(struct sn_hwperf_port_info);
 271         if (sz > sizeof(ptdata))
 272                 BUG();
 273         e = ia64_sn_hwperf_op(sn_hwperf_master_nasid,
 274                               SN_HWPERF_ENUM_PORTS, nodeobj->id, sz,
 275                               (u64)&ptdata, 0, 0, NULL);
 276         if (e != SN_HWPERF_OP_OK) {
 277                 e = -EINVAL;
 278                 goto err;
 279         }
 280
 281         /* find nearest node with cpus and nearest memory */
 282         for (router=NULL, j=0; j < op->ports; j++) {
 283                 dest = sn_hwperf_findobj_id(objbuf, nobj, ptdata[j].conn_id);
 284                 if (!dest || SN_HWPERF_FOREIGN(dest) ||
 285                     !SN_HWPERF_IS_NODE(dest) || SN_HWPERF_IS_IONODE(dest)) {
 286                         continue;
 287                 }
 288                 c = sn_hwperf_obj_to_cnode(dest);
 289                 if (!found_cpu && sn_hwperf_has_cpus(c)) {
 290                         if (near_cpu_node)
 291                                 *near_cpu_node = c;
 292                         found_cpu++;
 293                 }
 294                 if (!found_mem && sn_hwperf_has_mem(c)) {
 295                         if (near_mem_node)
 296                                 *near_mem_node = c;
 297                         found_mem++;
 298                 }
 299                 if (SN_HWPERF_IS_ROUTER(dest))
 300                         router = dest;
 301         }
 302
 303         if (router && (!found_cpu || !found_mem)) {
 304                 /* search for a node connected to the same router */
 305                 sz = router->ports * sizeof(struct sn_hwperf_port_info);
 306                 if (sz > sizeof(ptdata))
 307                         BUG();
 308                 e = ia64_sn_hwperf_op(sn_hwperf_master_nasid,
 309                                       SN_HWPERF_ENUM_PORTS, router->id, sz,
 310                                       (u64)&ptdata, 0, 0, NULL);
 311                 if (e != SN_HWPERF_OP_OK) {
 312                         e = -EINVAL;
 313                         goto err;
 314                 }
 315                 for (j=0; j < router->ports; j++) {
 316                         dest = sn_hwperf_findobj_id(objbuf, nobj,
 317                                 ptdata[j].conn_id);
 318                         if (!dest || dest->id == node ||
 319                             SN_HWPERF_FOREIGN(dest) ||
 320                             !SN_HWPERF_IS_NODE(dest) ||
 321                             SN_HWPERF_IS_IONODE(dest)) {
 322                                 continue;
 323                         }
 324                         c = sn_hwperf_obj_to_cnode(dest);
 325                         if (!found_cpu && sn_hwperf_has_cpus(c)) {
 326                                 if (near_cpu_node)
 327                                         *near_cpu_node = c;
 328                                 found_cpu++;
 329                         }
 330                         if (!found_mem && sn_hwperf_has_mem(c)) {
 331                                 if (near_mem_node)
 332                                         *near_mem_node = c;
 333                                 found_mem++;
 334                         }
 335                         if (found_cpu && found_mem)
 336                                 break;
 337                 }
 338         }
 339
 340         if (!found_cpu || !found_mem) {
 341                 /* resort to _any_ node with CPUs and memory */
 342                 for (i=0, op=objbuf; i < nobj; i++, op++) {
 343                         if (SN_HWPERF_FOREIGN(op) ||
 344                             SN_HWPERF_IS_IONODE(op) ||
 345                             !SN_HWPERF_IS_NODE(op)) {
 346                                 continue;
 347                         }
 348                         c = sn_hwperf_obj_to_cnode(op);
 349                         if (!found_cpu && sn_hwperf_has_cpus(c)) {
 350                                 if (near_cpu_node)
 351                                         *near_cpu_node = c;
 352                                 found_cpu++;
 353                         }
 354                         if (!found_mem && sn_hwperf_has_mem(c)) {
 355                                 if (near_mem_node)
 356                                         *near_mem_node = c;
 357                                 found_mem++;
 358                         }
 359                         if (found_cpu && found_mem)
 360                                 break;
 361                 }
 362         }
 363
 364         if (!found_cpu || !found_mem)
 365                 e = -ENODATA;
 366
 367 err:
 368         return e;
 369 }
 370
 371
 372 static int sn_topology_show(struct seq_file *s, void *d)
 373 {
 374         int sz;
 375         int pt;
 376         int e = 0;
 377         int i;
 378         int j;
 379         const char *slabname;
 380         int ordinal;
 381         cpumask_t cpumask;
 382         char slice;
 383         struct cpuinfo_ia64 *c;
 384         struct sn_hwperf_port_info *ptdata;
 385         struct sn_hwperf_object_info *p;
 386         struct sn_hwperf_object_info *obj = d;  /* this object */
 387         struct sn_hwperf_object_info *objs = s->private; /* all objects */
 388         u8 shubtype;
 389         u8 system_size;
 390         u8 sharing_size;
 391         u8 partid;
 392         u8 coher;
 393         u8 nasid_shift;
 394         u8 region_size;
 395         u16 nasid_mask;
 396         int nasid_msb;
 397
 398         if (obj == objs) {
 399                 seq_printf(s, "# sn_topology version 2\n");
 400                 seq_printf(s, "# objtype ordinal location partition"
 401                         " [attribute value [, ...]]\n");
 402
 403                 if (ia64_sn_get_sn_info(0,
 404                         &shubtype, &nasid_mask, &nasid_shift, &system_size,
 405                         &sharing_size, &partid, &coher, &region_size))
 406                         BUG();
 407                 for (nasid_msb=63; nasid_msb > 0; nasid_msb--) {
 408                         if (((u64)nasid_mask << nasid_shift) & (1ULL << nasid_msb))
 409                                 break;
 410                 }
 411                 seq_printf(s, "partition %u %s local "
 412                         "shubtype %s, "
 413                         "nasid_mask 0x%016lx, "
 414                         "nasid_bits %d:%d, "
 415                         "system_size %d, "
 416                         "sharing_size %d, "
 417                         "coherency_domain %d, "
 418                         "region_size %d\n",
 419
 420                         partid, system_utsname.nodename,
 421                         shubtype ? "shub2" : "shub1",
 422                         (u64)nasid_mask << nasid_shift, nasid_msb, nasid_shift,
 423                         system_size, sharing_size, coher, region_size);
 424
 425                 print_pci_topology(s);
 426         }
 427
 428         if (SN_HWPERF_FOREIGN(obj)) {
 429                 /* private in another partition: not interesting */
 430                 return 0;
 431         }
 432
 433         for (i = 0; i < SN_HWPERF_MAXSTRING && obj->name[i]; i++) {
 434                 if (obj->name[i] == ' ')
 435                         obj->name[i] = '_';
 436         }
 437
 438         slabname = sn_hwperf_get_slabname(obj, objs, &ordinal);
 439         seq_printf(s, "%s %d %s %s asic %s", slabname, ordinal, obj->location,
 440                 obj->sn_hwp_this_part ? "local" : "shared", obj->name);
 441
 442         if (!SN_HWPERF_IS_NODE(obj) && !SN_HWPERF_IS_IONODE(obj))
 443                 seq_putc(s, '\n');
 444         else {
 445                 cnodeid_t near_mem = -1;
 446                 cnodeid_t near_cpu = -1;
 447
 448                 seq_printf(s, ", nasid 0x%x", cnodeid_to_nasid(ordinal));
 449
 450                 if (sn_hwperf_get_nearest_node_objdata(objs, sn_hwperf_obj_cnt,
 451                         ordinal, &near_mem, &near_cpu) == 0) {
 452                         seq_printf(s, ", near_mem_nodeid %d, near_cpu_nodeid %d",
 453                                 near_mem, near_cpu);
 454                 }
 455
 456                 if (!SN_HWPERF_IS_IONODE(obj)) {
 457                         for_each_online_node(i) {
 458                                 seq_printf(s, i ? ":%d" : ", dist %d",
 459                                         node_distance(ordinal, i));
 460                         }
 461                 }
 462
 463                 seq_putc(s, '\n');
 464
 465                 /*
 466                  * CPUs on this node, if any
 467                  */
 468                 cpumask = node_to_cpumask(ordinal);
 469                 for_each_online_cpu(i) {
 470                         if (cpu_isset(i, cpumask)) {
 471                                 slice = 'a' + cpuid_to_slice(i);
 472                                 c = cpu_data(i);
 473                                 seq_printf(s, "cpu %d %s%c local"
 474                                         " freq %luMHz, arch ia64",
 475                                         i, obj->location, slice,
 476                                         c->proc_freq / 1000000);
 477                                 for_each_online_cpu(j) {
 478                                         seq_printf(s, j ? ":%d" : ", dist %d",
 479                                                 node_distance(
 480                                                     cpuid_to_cnodeid(i),
 481                                                     cpuid_to_cnodeid(j)));
 482                                 }
 483                                 seq_putc(s, '\n');
 484                         }
 485                 }
 486         }
 487
 488         if (obj->ports) {
 489                 /*
 490                  * numalink ports
 491                  */
 492                 sz = obj->ports * sizeof(struct sn_hwperf_port_info);
 493                 if ((ptdata = vmalloc(sz)) == NULL)
 494                         return -ENOMEM;
 495                 e = ia64_sn_hwperf_op(sn_hwperf_master_nasid,
 496                                       SN_HWPERF_ENUM_PORTS, obj->id, sz,
 497                                       (u64) ptdata, 0, 0, NULL);
 498                 if (e != SN_HWPERF_OP_OK)
 499                         return -EINVAL;
 500                 for (ordinal=0, p=objs; p != obj; p++) {
 501                         if (!SN_HWPERF_FOREIGN(p))
 502                                 ordinal += p->ports;
 503                 }
 504                 for (pt = 0; pt < obj->ports; pt++) {
 505                         for (p = objs, i = 0; i < sn_hwperf_obj_cnt; i++, p++) {
 506                                 if (ptdata[pt].conn_id == p->id) {
 507                                         break;
 508                                 }
 509                         }
 510                         seq_printf(s, "numalink %d %s-%d",
 511                             ordinal+pt, obj->location, ptdata[pt].port);
 512
 513                         if (i >= sn_hwperf_obj_cnt) {
 514                                 /* no connection */
 515                                 seq_puts(s, " local endpoint disconnected"
 516                                             ", protocol unknown\n");
 517                                 continue;
 518                         }
 519
 520                         if (obj->sn_hwp_this_part && p->sn_hwp_this_part)
 521                                 /* both ends local to this partition */
 522                                 seq_puts(s, " local");
 523                         else if (!obj->sn_hwp_this_part && !p->sn_hwp_this_part)
 524                                 /* both ends of the link in foreign partiton */
 525                                 seq_puts(s, " foreign");
 526                         else
 527                                 /* link straddles a partition */
 528                                 seq_puts(s, " shared");
 529
 530                         /*
 531                          * Unlikely, but strictly should query the LLP config
 532                          * registers because an NL4R can be configured to run
 533                          * NL3 protocol, even when not talking to an NL3 router.
 534                          * Ditto for node-node.
 535                          */
 536                         seq_printf(s, " endpoint %s-%d, protocol %s\n",
 537                                 p->location, ptdata[pt].conn_port,
 538                                 (SN_HWPERF_IS_NL3ROUTER(obj) ||
 539                                 SN_HWPERF_IS_NL3ROUTER(p)) ?  "LLP3" : "LLP4");
 540                 }
 541                 vfree(ptdata);
 542         }
 543
 544         return 0;
 545 }
 546
 547 static void *sn_topology_start(struct seq_file *s, loff_t * pos)
 548 {
 549         struct sn_hwperf_object_info *objs = s->private;
 550
 551         if (*pos < sn_hwperf_obj_cnt)
 552                 return (void *)(objs + *pos);
 553
 554         return NULL;
 555 }
 556
 557 static void *sn_topology_next(struct seq_file *s, void *v, loff_t * pos)
 558 {
 559         ++*pos;
 560         return sn_topology_start(s, pos);
 561 }
 562
 563 static void sn_topology_stop(struct seq_file *m, void *v)
 564 {
 565         return;
 566 }
 567
 568 /*
 569  * /proc/sgi_sn/sn_topology, read-only using seq_file
 570  */
 571 static struct seq_operations sn_topology_seq_ops = {
 572         .start = sn_topology_start,
 573         .next = sn_topology_next,
 574         .stop = sn_topology_stop,
 575         .show = sn_topology_show
 576 };
 577
 578 struct sn_hwperf_op_info {
 579         u64 op;
 580         struct sn_hwperf_ioctl_args *a;
 581         void *p;
 582         int *v0;
 583         int ret;
 584 };
 585
 586 static void sn_hwperf_call_sal(void *info)
 587 {
 588         struct sn_hwperf_op_info *op_info = info;
 589         int r;
 590
 591         r = ia64_sn_hwperf_op(sn_hwperf_master_nasid, op_info->op,
 592                       op_info->a->arg, op_info->a->sz,
 593                       (u64) op_info->p, 0, 0, op_info->v0);
 594         op_info->ret = r;
 595 }
 596
 597 static int sn_hwperf_op_cpu(struct sn_hwperf_op_info *op_info)
 598 {
 599         u32 cpu;
 600         u32 use_ipi;
 601         int r = 0;
 602         cpumask_t save_allowed;
 603
 604         cpu = (op_info->a->arg & SN_HWPERF_ARG_CPU_MASK) >> 32;
 605         use_ipi = op_info->a->arg & SN_HWPERF_ARG_USE_IPI_MASK;
 606         op_info->a->arg &= SN_HWPERF_ARG_OBJID_MASK;
 607
 608         if (cpu != SN_HWPERF_ARG_ANY_CPU) {
 609                 if (cpu >= num_online_cpus() || !cpu_online(cpu)) {
 610                         r = -EINVAL;
 611                         goto out;
 612                 }
 613         }
 614
 615         if (cpu == SN_HWPERF_ARG_ANY_CPU || cpu == get_cpu()) {
 616                 /* don't care, or already on correct cpu */
 617                 sn_hwperf_call_sal(op_info);
 618         }
 619         else {
 620                 if (use_ipi) {
 621                         /* use an interprocessor interrupt to call SAL */
 622                         smp_call_function_single(cpu, sn_hwperf_call_sal,
 623                                 op_info, 1, 1);
 624                 }
 625                 else {
 626                         /* migrate the task before calling SAL */
 627                         save_allowed = current->cpus_allowed;
 628                         set_cpus_allowed(current, cpumask_of_cpu(cpu));
 629                         sn_hwperf_call_sal(op_info);
 630                         set_cpus_allowed(current, save_allowed);
 631                 }
 632         }
 633         r = op_info->ret;
 634
 635 out:
 636         return r;
 637 }
 638
 639 /* map SAL hwperf error code to system error code */
 640 static int sn_hwperf_map_err(int hwperf_err)
 641 {
 642         int e;
 643
 644         switch(hwperf_err) {
 645         case SN_HWPERF_OP_OK:
 646                 e = 0;
 647                 break;
 648
 649         case SN_HWPERF_OP_NOMEM:
 650                 e = -ENOMEM;
 651                 break;
 652
 653         case SN_HWPERF_OP_NO_PERM:
 654                 e = -EPERM;
 655                 break;
 656
 657         case SN_HWPERF_OP_IO_ERROR:
 658                 e = -EIO;
 659                 break;
 660
 661         case SN_HWPERF_OP_BUSY:
 662                 e = -EBUSY;
 663                 break;
 664
 665         case SN_HWPERF_OP_RECONFIGURE:
 666                 e = -EAGAIN;
 667                 break;
 668
 669         case SN_HWPERF_OP_INVAL:
 670         default:
 671                 e = -EINVAL;
 672                 break;
 673         }
 674
 675         return e;
 676 }
 677
 678 /*
 679  * ioctl for "sn_hwperf" misc device
 680  */
 681 static int
 682 sn_hwperf_ioctl(struct inode *in, struct file *fp, u32 op, u64 arg)
 683 {
 684         struct sn_hwperf_ioctl_args a;
 685         struct cpuinfo_ia64 *cdata;
 686         struct sn_hwperf_object_info *objs;
 687         struct sn_hwperf_object_info *cpuobj;
 688         struct sn_hwperf_op_info op_info;
 689         void *p = NULL;
 690         int nobj;
 691         char slice;
 692         int node;
 693         int r;
 694         int v0;
 695         int i;
 696         int j;
 697
 698         unlock_kernel();
 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                         memset(p, 0, a.sz);
 746                         for (i = 0; i < nobj; i++) {
 747                                 if (!SN_HWPERF_IS_NODE(objs + i))
 748                                         continue;
 749                                 node = sn_hwperf_obj_to_cnode(objs + i);
 750                                 for_each_online_cpu(j) {
 751                                         if (node != cpu_to_node(j))
 752                                                 continue;
 753                                         cpuobj = (struct sn_hwperf_object_info *) p + j;
 754                                         slice = 'a' + cpuid_to_slice(j);
 755                                         cdata = cpu_data(j);
 756                                         cpuobj->id = j;
 757                                         snprintf(cpuobj->name,
 758                                                  sizeof(cpuobj->name),
 759                                                  "CPU %luMHz %s",
 760                                                  cdata->proc_freq / 1000000,
 761                                                  cdata->vendor);
 762                                         snprintf(cpuobj->location,
 763                                                  sizeof(cpuobj->location),
 764                                                  "%s%c", objs[i].location,
 765                                                  slice);
 766                                 }
 767                         }
 768
 769                         vfree(objs);
 770                 }
 771                 break;
 772
 773         case SN_HWPERF_GET_NODE_NASID:
 774                 if (a.sz != sizeof(u64) ||
 775                    (node = a.arg) < 0 || !node_possible(node)) {
 776                         r = -EINVAL;
 777                         goto error;
 778                 }
 779                 *(u64 *)p = (u64)cnodeid_to_nasid(node);
 780                 break;
 781
 782         case SN_HWPERF_GET_OBJ_NODE:
 783                 if (a.sz != sizeof(u64) || a.arg < 0) {
 784                         r = -EINVAL;
 785                         goto error;
 786                 }
 787                 if ((r = sn_hwperf_enum_objects(&nobj, &objs)) == 0) {
 788                         if (a.arg >= nobj) {
 789                                 r = -EINVAL;
 790                                 vfree(objs);
 791                                 goto error;
 792                         }
 793                         if (objs[(i = a.arg)].id != a.arg) {
 794                                 for (i = 0; i < nobj; i++) {
 795                                         if (objs[i].id == a.arg)
 796                                                 break;
 797                                 }
 798                         }
 799                         if (i == nobj) {
 800                                 r = -EINVAL;
 801                                 vfree(objs);
 802                                 goto error;
 803                         }
 804
 805                         if (!SN_HWPERF_IS_NODE(objs + i) &&
 806                             !SN_HWPERF_IS_IONODE(objs + i)) {
 807                                 r = -ENOENT;
 808                                 vfree(objs);
 809                                 goto error;
 810                         }
 811
 812                         *(u64 *)p = (u64)sn_hwperf_obj_to_cnode(objs + i);
 813                         vfree(objs);
 814                 }
 815                 break;
 816
 817         case SN_HWPERF_GET_MMRS:
 818         case SN_HWPERF_SET_MMRS:
 819         case SN_HWPERF_OBJECT_DISTANCE:
 820                 op_info.p = p;
 821                 op_info.a = &a;
 822                 op_info.v0 = &v0;
 823                 op_info.op = op;
 824                 r = sn_hwperf_op_cpu(&op_info);
 825                 if (r) {
 826                         r = sn_hwperf_map_err(r);
 827                         a.v0 = v0;
 828                         goto error;
 829                 }
 830                 break;
 831
 832         default:
 833                 /* all other ops are a direct SAL call */
 834                 r = ia64_sn_hwperf_op(sn_hwperf_master_nasid, op,
 835                               a.arg, a.sz, (u64) p, 0, 0, &v0);
 836                 if (r) {
 837                         r = sn_hwperf_map_err(r);
 838                         goto error;
 839                 }
 840                 a.v0 = v0;
 841                 break;
 842         }
 843
 844         if (op & SN_HWPERF_OP_MEM_COPYOUT) {
 845                 r = copy_to_user((void __user *)a.ptr, p, a.sz);
 846                 if (r != 0) {
 847                         r = -EFAULT;
 848                         goto error;
 849                 }
 850         }
 851
 852 error:
 853         vfree(p);
 854
 855         lock_kernel();
 856         return r;
 857 }
 858
 859 static struct file_operations sn_hwperf_fops = {
 860         .ioctl = sn_hwperf_ioctl,
 861 };
 862
 863 static struct miscdevice sn_hwperf_dev = {
 864         MISC_DYNAMIC_MINOR,
 865         "sn_hwperf",
 866         &sn_hwperf_fops
 867 };
 868
 869 static int sn_hwperf_init(void)
 870 {
 871         u64 v;
 872         int salr;
 873         int e = 0;
 874
 875         /* single threaded, once-only initialization */
 876         down(&sn_hwperf_init_mutex);
 877
 878         if (sn_hwperf_salheap) {
 879                 up(&sn_hwperf_init_mutex);
 880                 return e;
 881         }
 882
 883         /*
 884          * The PROM code needs a fixed reference node. For convenience the
 885          * same node as the console I/O is used.
 886          */
 887         sn_hwperf_master_nasid = (nasid_t) ia64_sn_get_console_nasid();
 888
 889         /*
 890          * Request the needed size and install the PROM scratch area.
 891          * The PROM keeps various tracking bits in this memory area.
 892          */
 893         salr = ia64_sn_hwperf_op(sn_hwperf_master_nasid,
 894                                  (u64) SN_HWPERF_GET_HEAPSIZE, 0,
 895                                  (u64) sizeof(u64), (u64) &v, 0, 0, NULL);
 896         if (salr != SN_HWPERF_OP_OK) {
 897                 e = -EINVAL;
 898                 goto out;
 899         }
 900
 901         if ((sn_hwperf_salheap = vmalloc(v)) == NULL) {
 902                 e = -ENOMEM;
 903                 goto out;
 904         }
 905         salr = ia64_sn_hwperf_op(sn_hwperf_master_nasid,
 906                                  SN_HWPERF_INSTALL_HEAP, 0, v,
 907                                  (u64) sn_hwperf_salheap, 0, 0, NULL);
 908         if (salr != SN_HWPERF_OP_OK) {
 909                 e = -EINVAL;
 910                 goto out;
 911         }
 912
 913         salr = ia64_sn_hwperf_op(sn_hwperf_master_nasid,
 914                                  SN_HWPERF_OBJECT_COUNT, 0,
 915                                  sizeof(u64), (u64) &v, 0, 0, NULL);
 916         if (salr != SN_HWPERF_OP_OK) {
 917                 e = -EINVAL;
 918                 goto out;
 919         }
 920         sn_hwperf_obj_cnt = (int)v;
 921
 922 out:
 923         if (e < 0 && sn_hwperf_salheap) {
 924                 vfree(sn_hwperf_salheap);
 925                 sn_hwperf_salheap = NULL;
 926                 sn_hwperf_obj_cnt = 0;
 927         }
 928         up(&sn_hwperf_init_mutex);
 929         return e;
 930 }
 931
 932 int sn_topology_open(struct inode *inode, struct file *file)
 933 {
 934         int e;
 935         struct seq_file *seq;
 936         struct sn_hwperf_object_info *objbuf;
 937         int nobj;
 938
 939         if ((e = sn_hwperf_enum_objects(&nobj, &objbuf)) == 0) {
 940                 e = seq_open(file, &sn_topology_seq_ops);
 941                 seq = file->private_data;
 942                 seq->private = objbuf;
 943         }
 944
 945         return e;
 946 }
 947
 948 int sn_topology_release(struct inode *inode, struct file *file)
 949 {
 950         struct seq_file *seq = file->private_data;
 951
 952         vfree(seq->private);
 953         return seq_release(inode, file);
 954 }
 955
 956 int sn_hwperf_get_nearest_node(cnodeid_t node,
 957         cnodeid_t *near_mem_node, cnodeid_t *near_cpu_node)
 958 {
 959         int e;
 960         int nobj;
 961         struct sn_hwperf_object_info *objbuf;
 962
 963         if ((e = sn_hwperf_enum_objects(&nobj, &objbuf)) == 0) {
 964                 e = sn_hwperf_get_nearest_node_objdata(objbuf, nobj,
 965                         node, near_mem_node, near_cpu_node);
 966                 vfree(objbuf);
 967         }
 968
 969         return e;
 970 }
 971
 972 static int __devinit sn_hwperf_misc_register_init(void)
 973 {
 974         int e;
 975
 976         sn_hwperf_init();
 977
 978         /*
 979          * Register a dynamic misc device for hwperf ioctls. Platforms
 980          * supporting hotplug will create /dev/sn_hwperf, else user
 981          * can to look up the minor number in /proc/misc.
 982          */
 983         if ((e = misc_register(&sn_hwperf_dev)) != 0) {
 984                 printk(KERN_ERR "sn_hwperf_misc_register_init: failed to "
 985                 "register misc device for \"%s\"\n", sn_hwperf_dev.name);
 986         }
 987
 988         return e;
 989 }
 990
 991 device_initcall(sn_hwperf_misc_register_init); /* after misc_init() */
 992 EXPORT_SYMBOL(sn_hwperf_get_nearest_node);