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/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 DECLARE_MUTEX(sn_hwperf_init_mutex);
  53
  54 static int sn_hwperf_enum_objects(int *nobj, struct sn_hwperf_object_info **ret)
  55 {
  56         int e;
  57         u64 sz;
  58         struct sn_hwperf_object_info *objbuf = NULL;
  59
  60         if ((e = sn_hwperf_init()) < 0) {
  61                 printk("sn_hwperf_init failed: err %d\n", e);
  62                 goto out;
  63         }
  64
  65         sz = sn_hwperf_obj_cnt * sizeof(struct sn_hwperf_object_info);
  66         if ((objbuf = (struct sn_hwperf_object_info *) vmalloc(sz)) == NULL) {
  67                 printk("sn_hwperf_enum_objects: vmalloc(%d) failed\n", (int)sz);
  68                 e = -ENOMEM;
  69                 goto out;
  70         }
  71
  72         e = ia64_sn_hwperf_op(sn_hwperf_master_nasid, SN_HWPERF_ENUM_OBJECTS,
  73                 0, sz, (u64) objbuf, 0, 0, NULL);
  74         if (e != SN_HWPERF_OP_OK) {
  75                 e = -EINVAL;
  76                 vfree(objbuf);
  77         }
  78
  79 out:
  80         *nobj = sn_hwperf_obj_cnt;
  81         *ret = objbuf;
  82         return e;
  83 }
  84
  85 static int sn_hwperf_location_to_bpos(char *location,
  86         int *rack, int *bay, int *slot, int *slab)
  87 {
  88         char type;
  89
  90         /* first scan for an old style geoid string */
  91         if (sscanf(location, "%03d%c%02d#%d",
  92                 rack, &type, bay, slab) == 4)
  93                 *slot = 0;
  94         else /* scan for a new bladed geoid string */
  95         if (sscanf(location, "%03d%c%02d^%02d#%d",
  96                 rack, &type, bay, slot, slab) != 5)
  97                 return -1;
  98         /* success */
  99         return 0;
 100 }
 101
 102 static int sn_hwperf_geoid_to_cnode(char *location)
 103 {
 104         int cnode;
 105         geoid_t geoid;
 106         moduleid_t module_id;
 107         int rack, bay, slot, slab;
 108         int this_rack, this_bay, this_slot, this_slab;
 109
 110         if (sn_hwperf_location_to_bpos(location, &rack, &bay, &slot, &slab))
 111                 return -1;
 112
 113         for (cnode = 0; cnode < numionodes; cnode++) {
 114                 geoid = cnodeid_get_geoid(cnode);
 115                 module_id = geo_module(geoid);
 116                 this_rack = MODULE_GET_RACK(module_id);
 117                 this_bay = MODULE_GET_BPOS(module_id);
 118                 this_slot = geo_slot(geoid);
 119                 this_slab = geo_slab(geoid);
 120                 if (rack == this_rack && bay == this_bay &&
 121                         slot == this_slot && slab == this_slab) {
 122                         break;
 123                 }
 124         }
 125
 126         return cnode < numionodes ? cnode : -1;
 127 }
 128
 129 static int sn_hwperf_obj_to_cnode(struct sn_hwperf_object_info * obj)
 130 {
 131         if (!obj->sn_hwp_this_part)
 132                 return -1;
 133         return sn_hwperf_geoid_to_cnode(obj->location);
 134 }
 135
 136 static int sn_hwperf_generic_ordinal(struct sn_hwperf_object_info *obj,
 137                                 struct sn_hwperf_object_info *objs)
 138 {
 139         int ordinal;
 140         struct sn_hwperf_object_info *p;
 141
 142         for (ordinal=0, p=objs; p != obj; p++) {
 143                 if (SN_HWPERF_FOREIGN(p))
 144                         continue;
 145                 if (SN_HWPERF_SAME_OBJTYPE(p, obj))
 146                         ordinal++;
 147         }
 148
 149         return ordinal;
 150 }
 151
 152 static const char *slabname_node =      "node"; /* SHub asic */
 153 static const char *slabname_ionode =    "ionode"; /* TIO asic */
 154 static const char *slabname_router =    "router"; /* NL3R or NL4R */
 155 static const char *slabname_other =     "other"; /* unknown asic */
 156
 157 static const char *sn_hwperf_get_slabname(struct sn_hwperf_object_info *obj,
 158                         struct sn_hwperf_object_info *objs, int *ordinal)
 159 {
 160         int isnode;
 161         const char *slabname = slabname_other;
 162
 163         if ((isnode = SN_HWPERF_IS_NODE(obj)) || SN_HWPERF_IS_IONODE(obj)) {
 164                 slabname = isnode ? slabname_node : slabname_ionode;
 165                 *ordinal = sn_hwperf_obj_to_cnode(obj);
 166         }
 167         else {
 168                 *ordinal = sn_hwperf_generic_ordinal(obj, objs);
 169                 if (SN_HWPERF_IS_ROUTER(obj))
 170                         slabname = slabname_router;
 171         }
 172
 173         return slabname;
 174 }
 175
 176 static void print_pci_topology(struct seq_file *s,
 177         struct sn_hwperf_object_info *obj, int *ordinal,
 178         u64 rack, u64 bay, u64 slot, u64 slab)
 179 {
 180         char *p1;
 181         char *p2;
 182         char *pg;
 183
 184         if (!(pg = (char *)get_zeroed_page(GFP_KERNEL)))
 185                 return; /* ignore */
 186         if (ia64_sn_ioif_get_pci_topology(rack, bay, slot, slab,
 187                 __pa(pg), PAGE_SIZE) == SN_HWPERF_OP_OK) {
 188                 for (p1=pg; *p1 && p1 < pg + PAGE_SIZE;) {
 189                         if (!(p2 = strchr(p1, '\n')))
 190                                 break;
 191                         *p2 = '\0';
 192                         seq_printf(s, "pcibus %d %s-%s\n",
 193                                 *ordinal, obj->location, p1);
 194                         (*ordinal)++;
 195                         p1 = p2 + 1;
 196                 }
 197         }
 198         free_page((unsigned long)pg);
 199 }
 200
 201 static int sn_topology_show(struct seq_file *s, void *d)
 202 {
 203         int sz;
 204         int pt;
 205         int e = 0;
 206         int i;
 207         int j;
 208         const char *slabname;
 209         int ordinal;
 210         cpumask_t cpumask;
 211         char slice;
 212         struct cpuinfo_ia64 *c;
 213         struct sn_hwperf_port_info *ptdata;
 214         struct sn_hwperf_object_info *p;
 215         struct sn_hwperf_object_info *obj = d;  /* this object */
 216         struct sn_hwperf_object_info *objs = s->private; /* all objects */
 217         int rack, bay, slot, slab;
 218         u8 shubtype;
 219         u8 system_size;
 220         u8 sharing_size;
 221         u8 partid;
 222         u8 coher;
 223         u8 nasid_shift;
 224         u8 region_size;
 225         u16 nasid_mask;
 226         int nasid_msb;
 227         int pci_bus_ordinal = 0;
 228
 229         if (obj == objs) {
 230                 seq_printf(s, "# sn_topology version 2\n");
 231                 seq_printf(s, "# objtype ordinal location partition"
 232                         " [attribute value [, ...]]\n");
 233
 234                 if (ia64_sn_get_sn_info(0,
 235                         &shubtype, &nasid_mask, &nasid_shift, &system_size,
 236                         &sharing_size, &partid, &coher, &region_size))
 237                         BUG();
 238                 for (nasid_msb=63; nasid_msb > 0; nasid_msb--) {
 239                         if (((u64)nasid_mask << nasid_shift) & (1ULL << nasid_msb))
 240                                 break;
 241                 }
 242                 seq_printf(s, "partition %u %s local "
 243                         "shubtype %s, "
 244                         "nasid_mask 0x%016lx, "
 245                         "nasid_bits %d:%d, "
 246                         "system_size %d, "
 247                         "sharing_size %d, "
 248                         "coherency_domain %d, "
 249                         "region_size %d\n",
 250
 251                         partid, system_utsname.nodename,
 252                         shubtype ? "shub2" : "shub1",
 253                         (u64)nasid_mask << nasid_shift, nasid_msb, nasid_shift,
 254                         system_size, sharing_size, coher, region_size);
 255         }
 256
 257         if (SN_HWPERF_FOREIGN(obj)) {
 258                 /* private in another partition: not interesting */
 259                 return 0;
 260         }
 261
 262         for (i = 0; i < SN_HWPERF_MAXSTRING && obj->name[i]; i++) {
 263                 if (obj->name[i] == ' ')
 264                         obj->name[i] = '_';
 265         }
 266
 267         slabname = sn_hwperf_get_slabname(obj, objs, &ordinal);
 268         seq_printf(s, "%s %d %s %s asic %s", slabname, ordinal, obj->location,
 269                 obj->sn_hwp_this_part ? "local" : "shared", obj->name);
 270
 271         if (!SN_HWPERF_IS_NODE(obj) && !SN_HWPERF_IS_IONODE(obj))
 272                 seq_putc(s, '\n');
 273         else {
 274                 seq_printf(s, ", nasid 0x%x", cnodeid_to_nasid(ordinal));
 275                 for (i=0; i < numionodes; i++) {
 276                         seq_printf(s, i ? ":%d" : ", dist %d",
 277                                 node_distance(ordinal, i));
 278                 }
 279                 seq_putc(s, '\n');
 280
 281                 /*
 282                  * CPUs on this node, if any
 283                  */
 284                 cpumask = node_to_cpumask(ordinal);
 285                 for_each_online_cpu(i) {
 286                         if (cpu_isset(i, cpumask)) {
 287                                 slice = 'a' + cpuid_to_slice(i);
 288                                 c = cpu_data(i);
 289                                 seq_printf(s, "cpu %d %s%c local"
 290                                         " freq %luMHz, arch ia64",
 291                                         i, obj->location, slice,
 292                                         c->proc_freq / 1000000);
 293                                 for_each_online_cpu(j) {
 294                                         seq_printf(s, j ? ":%d" : ", dist %d",
 295                                                 node_distance(
 296                                                     cpuid_to_cnodeid(i),
 297                                                     cpuid_to_cnodeid(j)));
 298                                 }
 299                                 seq_putc(s, '\n');
 300                         }
 301                 }
 302
 303                 /*
 304                  * PCI busses attached to this node, if any
 305                  */
 306                 if (sn_hwperf_location_to_bpos(obj->location,
 307                         &rack, &bay, &slot, &slab)) {
 308                         /* export pci bus info */
 309                         print_pci_topology(s, obj, &pci_bus_ordinal,
 310                                 rack, bay, slot, slab);
 311
 312                 }
 313         }
 314
 315         if (obj->ports) {
 316                 /*
 317                  * numalink ports
 318                  */
 319                 sz = obj->ports * sizeof(struct sn_hwperf_port_info);
 320                 if ((ptdata = vmalloc(sz)) == NULL)
 321                         return -ENOMEM;
 322                 e = ia64_sn_hwperf_op(sn_hwperf_master_nasid,
 323                                       SN_HWPERF_ENUM_PORTS, obj->id, sz,
 324                                       (u64) ptdata, 0, 0, NULL);
 325                 if (e != SN_HWPERF_OP_OK)
 326                         return -EINVAL;
 327                 for (ordinal=0, p=objs; p != obj; p++) {
 328                         if (!SN_HWPERF_FOREIGN(p))
 329                                 ordinal += p->ports;
 330                 }
 331                 for (pt = 0; pt < obj->ports; pt++) {
 332                         for (p = objs, i = 0; i < sn_hwperf_obj_cnt; i++, p++) {
 333                                 if (ptdata[pt].conn_id == p->id) {
 334                                         break;
 335                                 }
 336                         }
 337                         seq_printf(s, "numalink %d %s-%d",
 338                             ordinal+pt, obj->location, ptdata[pt].port);
 339
 340                         if (i >= sn_hwperf_obj_cnt) {
 341                                 /* no connection */
 342                                 seq_puts(s, " local endpoint disconnected"
 343                                             ", protocol unknown\n");
 344                                 continue;
 345                         }
 346
 347                         if (obj->sn_hwp_this_part && p->sn_hwp_this_part)
 348                                 /* both ends local to this partition */
 349                                 seq_puts(s, " local");
 350                         else if (!obj->sn_hwp_this_part && !p->sn_hwp_this_part)
 351                                 /* both ends of the link in foreign partiton */
 352                                 seq_puts(s, " foreign");
 353                         else
 354                                 /* link straddles a partition */
 355                                 seq_puts(s, " shared");
 356
 357                         /*
 358                          * Unlikely, but strictly should query the LLP config
 359                          * registers because an NL4R can be configured to run
 360                          * NL3 protocol, even when not talking to an NL3 router.
 361                          * Ditto for node-node.
 362                          */
 363                         seq_printf(s, " endpoint %s-%d, protocol %s\n",
 364                                 p->location, ptdata[pt].conn_port,
 365                                 (SN_HWPERF_IS_NL3ROUTER(obj) ||
 366                                 SN_HWPERF_IS_NL3ROUTER(p)) ?  "LLP3" : "LLP4");
 367                 }
 368                 vfree(ptdata);
 369         }
 370
 371         return 0;
 372 }
 373
 374 static void *sn_topology_start(struct seq_file *s, loff_t * pos)
 375 {
 376         struct sn_hwperf_object_info *objs = s->private;
 377
 378         if (*pos < sn_hwperf_obj_cnt)
 379                 return (void *)(objs + *pos);
 380
 381         return NULL;
 382 }
 383
 384 static void *sn_topology_next(struct seq_file *s, void *v, loff_t * pos)
 385 {
 386         ++*pos;
 387         return sn_topology_start(s, pos);
 388 }
 389
 390 static void sn_topology_stop(struct seq_file *m, void *v)
 391 {
 392         return;
 393 }
 394
 395 /*
 396  * /proc/sgi_sn/sn_topology, read-only using seq_file
 397  */
 398 static struct seq_operations sn_topology_seq_ops = {
 399         .start = sn_topology_start,
 400         .next = sn_topology_next,
 401         .stop = sn_topology_stop,
 402         .show = sn_topology_show
 403 };
 404
 405 struct sn_hwperf_op_info {
 406         u64 op;
 407         struct sn_hwperf_ioctl_args *a;
 408         void *p;
 409         int *v0;
 410         int ret;
 411 };
 412
 413 static void sn_hwperf_call_sal(void *info)
 414 {
 415         struct sn_hwperf_op_info *op_info = info;
 416         int r;
 417
 418         r = ia64_sn_hwperf_op(sn_hwperf_master_nasid, op_info->op,
 419                       op_info->a->arg, op_info->a->sz,
 420                       (u64) op_info->p, 0, 0, op_info->v0);
 421         op_info->ret = r;
 422 }
 423
 424 static int sn_hwperf_op_cpu(struct sn_hwperf_op_info *op_info)
 425 {
 426         u32 cpu;
 427         u32 use_ipi;
 428         int r = 0;
 429         cpumask_t save_allowed;
 430
 431         cpu = (op_info->a->arg & SN_HWPERF_ARG_CPU_MASK) >> 32;
 432         use_ipi = op_info->a->arg & SN_HWPERF_ARG_USE_IPI_MASK;
 433         op_info->a->arg &= SN_HWPERF_ARG_OBJID_MASK;
 434
 435         if (cpu != SN_HWPERF_ARG_ANY_CPU) {
 436                 if (cpu >= num_online_cpus() || !cpu_online(cpu)) {
 437                         r = -EINVAL;
 438                         goto out;
 439                 }
 440         }
 441
 442         if (cpu == SN_HWPERF_ARG_ANY_CPU || cpu == get_cpu()) {
 443                 /* don't care, or already on correct cpu */
 444                 sn_hwperf_call_sal(op_info);
 445         }
 446         else {
 447                 if (use_ipi) {
 448                         /* use an interprocessor interrupt to call SAL */
 449                         smp_call_function_single(cpu, sn_hwperf_call_sal,
 450                                 op_info, 1, 1);
 451                 }
 452                 else {
 453                         /* migrate the task before calling SAL */
 454                         save_allowed = current->cpus_allowed;
 455                         set_cpus_allowed(current, cpumask_of_cpu(cpu));
 456                         sn_hwperf_call_sal(op_info);
 457                         set_cpus_allowed(current, save_allowed);
 458                 }
 459         }
 460         r = op_info->ret;
 461
 462 out:
 463         return r;
 464 }
 465
 466 /* map SAL hwperf error code to system error code */
 467 static int sn_hwperf_map_err(int hwperf_err)
 468 {
 469         int e;
 470
 471         switch(hwperf_err) {
 472         case SN_HWPERF_OP_OK:
 473                 e = 0;
 474                 break;
 475
 476         case SN_HWPERF_OP_NOMEM:
 477                 e = -ENOMEM;
 478                 break;
 479
 480         case SN_HWPERF_OP_NO_PERM:
 481                 e = -EPERM;
 482                 break;
 483
 484         case SN_HWPERF_OP_IO_ERROR:
 485                 e = -EIO;
 486                 break;
 487
 488         case SN_HWPERF_OP_BUSY:
 489                 e = -EBUSY;
 490                 break;
 491
 492         case SN_HWPERF_OP_RECONFIGURE:
 493                 e = -EAGAIN;
 494                 break;
 495
 496         case SN_HWPERF_OP_INVAL:
 497         default:
 498                 e = -EINVAL;
 499                 break;
 500         }
 501
 502         return e;
 503 }
 504
 505 /*
 506  * ioctl for "sn_hwperf" misc device
 507  */
 508 static int
 509 sn_hwperf_ioctl(struct inode *in, struct file *fp, u32 op, u64 arg)
 510 {
 511         struct sn_hwperf_ioctl_args a;
 512         struct cpuinfo_ia64 *cdata;
 513         struct sn_hwperf_object_info *objs;
 514         struct sn_hwperf_object_info *cpuobj;
 515         struct sn_hwperf_op_info op_info;
 516         void *p = NULL;
 517         int nobj;
 518         char slice;
 519         int node;
 520         int r;
 521         int v0;
 522         int i;
 523         int j;
 524
 525         unlock_kernel();
 526
 527         /* only user requests are allowed here */
 528         if ((op & SN_HWPERF_OP_MASK) < 10) {
 529                 r = -EINVAL;
 530                 goto error;
 531         }
 532         r = copy_from_user(&a, (const void __user *)arg,
 533                 sizeof(struct sn_hwperf_ioctl_args));
 534         if (r != 0) {
 535                 r = -EFAULT;
 536                 goto error;
 537         }
 538
 539         /*
 540          * Allocate memory to hold a kernel copy of the user buffer. The
 541          * buffer contents are either copied in or out (or both) of user
 542          * space depending on the flags encoded in the requested operation.
 543          */
 544         if (a.ptr) {
 545                 p = vmalloc(a.sz);
 546                 if (!p) {
 547                         r = -ENOMEM;
 548                         goto error;
 549                 }
 550         }
 551
 552         if (op & SN_HWPERF_OP_MEM_COPYIN) {
 553                 r = copy_from_user(p, (const void __user *)a.ptr, a.sz);
 554                 if (r != 0) {
 555                         r = -EFAULT;
 556                         goto error;
 557                 }
 558         }
 559
 560         switch (op) {
 561         case SN_HWPERF_GET_CPU_INFO:
 562                 if (a.sz == sizeof(u64)) {
 563                         /* special case to get size needed */
 564                         *(u64 *) p = (u64) num_online_cpus() *
 565                                 sizeof(struct sn_hwperf_object_info);
 566                 } else
 567                 if (a.sz < num_online_cpus() * sizeof(struct sn_hwperf_object_info)) {
 568                         r = -ENOMEM;
 569                         goto error;
 570                 } else
 571                 if ((r = sn_hwperf_enum_objects(&nobj, &objs)) == 0) {
 572                         memset(p, 0, a.sz);
 573                         for (i = 0; i < nobj; i++) {
 574                                 node = sn_hwperf_obj_to_cnode(objs + i);
 575                                 for_each_online_cpu(j) {
 576                                         if (node != cpu_to_node(j))
 577                                                 continue;
 578                                         cpuobj = (struct sn_hwperf_object_info *) p + j;
 579                                         slice = 'a' + cpuid_to_slice(j);
 580                                         cdata = cpu_data(j);
 581                                         cpuobj->id = j;
 582                                         snprintf(cpuobj->name,
 583                                                  sizeof(cpuobj->name),
 584                                                  "CPU %luMHz %s",
 585                                                  cdata->proc_freq / 1000000,
 586                                                  cdata->vendor);
 587                                         snprintf(cpuobj->location,
 588                                                  sizeof(cpuobj->location),
 589                                                  "%s%c", objs[i].location,
 590                                                  slice);
 591                                 }
 592                         }
 593
 594                         vfree(objs);
 595                 }
 596                 break;
 597
 598         case SN_HWPERF_GET_NODE_NASID:
 599                 if (a.sz != sizeof(u64) ||
 600                    (node = a.arg) < 0 || node >= numionodes) {
 601                         r = -EINVAL;
 602                         goto error;
 603                 }
 604                 *(u64 *)p = (u64)cnodeid_to_nasid(node);
 605                 break;
 606
 607         case SN_HWPERF_GET_OBJ_NODE:
 608                 if (a.sz != sizeof(u64) || a.arg < 0) {
 609                         r = -EINVAL;
 610                         goto error;
 611                 }
 612                 if ((r = sn_hwperf_enum_objects(&nobj, &objs)) == 0) {
 613                         if (a.arg >= nobj) {
 614                                 r = -EINVAL;
 615                                 vfree(objs);
 616                                 goto error;
 617                         }
 618                         if (objs[(i = a.arg)].id != a.arg) {
 619                                 for (i = 0; i < nobj; i++) {
 620                                         if (objs[i].id == a.arg)
 621                                                 break;
 622                                 }
 623                         }
 624                         if (i == nobj) {
 625                                 r = -EINVAL;
 626                                 vfree(objs);
 627                                 goto error;
 628                         }
 629                         *(u64 *)p = (u64)sn_hwperf_obj_to_cnode(objs + i);
 630                         vfree(objs);
 631                 }
 632                 break;
 633
 634         case SN_HWPERF_GET_MMRS:
 635         case SN_HWPERF_SET_MMRS:
 636         case SN_HWPERF_OBJECT_DISTANCE:
 637                 op_info.p = p;
 638                 op_info.a = &a;
 639                 op_info.v0 = &v0;
 640                 op_info.op = op;
 641                 r = sn_hwperf_op_cpu(&op_info);
 642                 if (r) {
 643                         r = sn_hwperf_map_err(r);
 644                         a.v0 = v0;
 645                         goto error;
 646                 }
 647                 break;
 648
 649         default:
 650                 /* all other ops are a direct SAL call */
 651                 r = ia64_sn_hwperf_op(sn_hwperf_master_nasid, op,
 652                               a.arg, a.sz, (u64) p, 0, 0, &v0);
 653                 if (r) {
 654                         r = sn_hwperf_map_err(r);
 655                         goto error;
 656                 }
 657                 a.v0 = v0;
 658                 break;
 659         }
 660
 661         if (op & SN_HWPERF_OP_MEM_COPYOUT) {
 662                 r = copy_to_user((void __user *)a.ptr, p, a.sz);
 663                 if (r != 0) {
 664                         r = -EFAULT;
 665                         goto error;
 666                 }
 667         }
 668
 669 error:
 670         vfree(p);
 671
 672         lock_kernel();
 673         return r;
 674 }
 675
 676 static struct file_operations sn_hwperf_fops = {
 677         .ioctl = sn_hwperf_ioctl,
 678 };
 679
 680 static struct miscdevice sn_hwperf_dev = {
 681         MISC_DYNAMIC_MINOR,
 682         "sn_hwperf",
 683         &sn_hwperf_fops
 684 };
 685
 686 static int sn_hwperf_init(void)
 687 {
 688         u64 v;
 689         int salr;
 690         int e = 0;
 691
 692         /* single threaded, once-only initialization */
 693         down(&sn_hwperf_init_mutex);
 694         if (sn_hwperf_salheap) {
 695                 up(&sn_hwperf_init_mutex);
 696                 return e;
 697         }
 698
 699         /*
 700          * The PROM code needs a fixed reference node. For convenience the
 701          * same node as the console I/O is used.
 702          */
 703         sn_hwperf_master_nasid = (nasid_t) ia64_sn_get_console_nasid();
 704
 705         /*
 706          * Request the needed size and install the PROM scratch area.
 707          * The PROM keeps various tracking bits in this memory area.
 708          */
 709         salr = ia64_sn_hwperf_op(sn_hwperf_master_nasid,
 710                                  (u64) SN_HWPERF_GET_HEAPSIZE, 0,
 711                                  (u64) sizeof(u64), (u64) &v, 0, 0, NULL);
 712         if (salr != SN_HWPERF_OP_OK) {
 713                 e = -EINVAL;
 714                 goto out;
 715         }
 716
 717         if ((sn_hwperf_salheap = vmalloc(v)) == NULL) {
 718                 e = -ENOMEM;
 719                 goto out;
 720         }
 721         salr = ia64_sn_hwperf_op(sn_hwperf_master_nasid,
 722                                  SN_HWPERF_INSTALL_HEAP, 0, v,
 723                                  (u64) sn_hwperf_salheap, 0, 0, NULL);
 724         if (salr != SN_HWPERF_OP_OK) {
 725                 e = -EINVAL;
 726                 goto out;
 727         }
 728
 729         salr = ia64_sn_hwperf_op(sn_hwperf_master_nasid,
 730                                  SN_HWPERF_OBJECT_COUNT, 0,
 731                                  sizeof(u64), (u64) &v, 0, 0, NULL);
 732         if (salr != SN_HWPERF_OP_OK) {
 733                 e = -EINVAL;
 734                 goto out;
 735         }
 736         sn_hwperf_obj_cnt = (int)v;
 737
 738 out:
 739         if (e < 0 && sn_hwperf_salheap) {
 740                 vfree(sn_hwperf_salheap);
 741                 sn_hwperf_salheap = NULL;
 742                 sn_hwperf_obj_cnt = 0;
 743         }
 744
 745         if (!e) {
 746                 /*
 747                  * Register a dynamic misc device for ioctl. Platforms
 748                  * supporting hotplug will create /dev/sn_hwperf, else
 749                  * user can to look up the minor number in /proc/misc.
 750                  */
 751                 if ((e = misc_register(&sn_hwperf_dev)) != 0) {
 752                         printk(KERN_ERR "sn_hwperf_init: misc register "
 753                                "for \"sn_hwperf\" failed, err %d\n", e);
 754                 }
 755         }
 756
 757         up(&sn_hwperf_init_mutex);
 758         return e;
 759 }
 760
 761 int sn_topology_open(struct inode *inode, struct file *file)
 762 {
 763         int e;
 764         struct seq_file *seq;
 765         struct sn_hwperf_object_info *objbuf;
 766         int nobj;
 767
 768         if ((e = sn_hwperf_enum_objects(&nobj, &objbuf)) == 0) {
 769                 e = seq_open(file, &sn_topology_seq_ops);
 770                 seq = file->private_data;
 771                 seq->private = objbuf;
 772         }
 773
 774         return e;
 775 }
 776
 777 int sn_topology_release(struct inode *inode, struct file *file)
 778 {
 779         struct seq_file *seq = file->private_data;
 780
 781         vfree(seq->private);
 782         return seq_release(inode, file);
 783 }