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("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 (cnode = 0; cnode < numionodes; 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 = 0; /* XXX */
 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 cnode < numionodes ? cnode : -1;
 128 }
 129
 130 static int sn_hwperf_obj_to_cnode(struct sn_hwperf_object_info * obj)
 131 {
 132         if (!obj->sn_hwp_this_part)
 133                 return -1;
 134         return sn_hwperf_geoid_to_cnode(obj->location);
 135 }
 136
 137 static int sn_hwperf_generic_ordinal(struct sn_hwperf_object_info *obj,
 138                                 struct sn_hwperf_object_info *objs)
 139 {
 140         int ordinal;
 141         struct sn_hwperf_object_info *p;
 142
 143         for (ordinal=0, p=objs; p != obj; p++) {
 144                 if (SN_HWPERF_FOREIGN(p))
 145                         continue;
 146                 if (SN_HWPERF_SAME_OBJTYPE(p, obj))
 147                         ordinal++;
 148         }
 149
 150         return ordinal;
 151 }
 152
 153 static const char *slabname_node =      "node"; /* SHub asic */
 154 static const char *slabname_ionode =    "ionode"; /* TIO asic */
 155 static const char *slabname_router =    "router"; /* NL3R or NL4R */
 156 static const char *slabname_other =     "other"; /* unknown asic */
 157
 158 static const char *sn_hwperf_get_slabname(struct sn_hwperf_object_info *obj,
 159                         struct sn_hwperf_object_info *objs, int *ordinal)
 160 {
 161         int isnode;
 162         const char *slabname = slabname_other;
 163
 164         if ((isnode = SN_HWPERF_IS_NODE(obj)) || SN_HWPERF_IS_IONODE(obj)) {
 165                 slabname = isnode ? slabname_node : slabname_ionode;
 166                 *ordinal = sn_hwperf_obj_to_cnode(obj);
 167         }
 168         else {
 169                 *ordinal = sn_hwperf_generic_ordinal(obj, objs);
 170                 if (SN_HWPERF_IS_ROUTER(obj))
 171                         slabname = slabname_router;
 172         }
 173
 174         return slabname;
 175 }
 176
 177 static void print_pci_topology(struct seq_file *s,
 178         struct sn_hwperf_object_info *obj, int *ordinal,
 179         char *pci_topo_buf, int len)
 180 {
 181         char *p1;
 182         char *p2;
 183
 184         for (p1=pci_topo_buf; *p1 && p1 < pci_topo_buf + len;) {
 185                 if (!(p2 = strchr(p1, '\n')))
 186                         break;
 187                 *p2 = '\0';
 188                 seq_printf(s, "pcibus %d %s-%s\n",
 189                         *ordinal, obj->location, p1);
 190                 (*ordinal)++;
 191                 p1 = p2 + 1;
 192         }
 193 }
 194
 195 static int sn_topology_show(struct seq_file *s, void *d)
 196 {
 197         int sz;
 198         int pt;
 199         int e = 0;
 200         int i;
 201         int j;
 202         const char *slabname;
 203         int ordinal;
 204         cpumask_t cpumask;
 205         char slice;
 206         struct cpuinfo_ia64 *c;
 207         struct sn_hwperf_port_info *ptdata;
 208         struct sn_hwperf_object_info *p;
 209         struct sn_hwperf_object_info *obj = d;  /* this object */
 210         struct sn_hwperf_object_info *objs = s->private; /* all objects */
 211         int rack, bay, slot, slab;
 212         u8 shubtype;
 213         u8 system_size;
 214         u8 sharing_size;
 215         u8 partid;
 216         u8 coher;
 217         u8 nasid_shift;
 218         u8 region_size;
 219         u16 nasid_mask;
 220         int nasid_msb;
 221         char *pci_topo_buf;
 222         int pci_bus_ordinal = 0;
 223         static int pci_topo_buf_len = 256;
 224
 225         if (obj == objs) {
 226                 seq_printf(s, "# sn_topology version 2\n");
 227                 seq_printf(s, "# objtype ordinal location partition"
 228                         " [attribute value [, ...]]\n");
 229
 230                 if (ia64_sn_get_sn_info(0,
 231                         &shubtype, &nasid_mask, &nasid_shift, &system_size,
 232                         &sharing_size, &partid, &coher, &region_size))
 233                         BUG();
 234                 for (nasid_msb=63; nasid_msb > 0; nasid_msb--) {
 235                         if (((u64)nasid_mask << nasid_shift) & (1ULL << nasid_msb))
 236                                 break;
 237                 }
 238                 seq_printf(s, "partition %u %s local "
 239                         "shubtype %s, "
 240                         "nasid_mask 0x%016lx, "
 241                         "nasid_bits %d:%d, "
 242                         "system_size %d, "
 243                         "sharing_size %d, "
 244                         "coherency_domain %d, "
 245                         "region_size %d\n",
 246
 247                         partid, system_utsname.nodename,
 248                         shubtype ? "shub2" : "shub1",
 249                         (u64)nasid_mask << nasid_shift, nasid_msb, nasid_shift,
 250                         system_size, sharing_size, coher, region_size);
 251         }
 252
 253         if (SN_HWPERF_FOREIGN(obj)) {
 254                 /* private in another partition: not interesting */
 255                 return 0;
 256         }
 257
 258         for (i = 0; i < SN_HWPERF_MAXSTRING && obj->name[i]; i++) {
 259                 if (obj->name[i] == ' ')
 260                         obj->name[i] = '_';
 261         }
 262
 263         slabname = sn_hwperf_get_slabname(obj, objs, &ordinal);
 264         seq_printf(s, "%s %d %s %s asic %s", slabname, ordinal, obj->location,
 265                 obj->sn_hwp_this_part ? "local" : "shared", obj->name);
 266
 267         if (!SN_HWPERF_IS_NODE(obj) && !SN_HWPERF_IS_IONODE(obj))
 268                 seq_putc(s, '\n');
 269         else {
 270                 seq_printf(s, ", nasid 0x%x", cnodeid_to_nasid(ordinal));
 271                 for (i=0; i < numionodes; i++) {
 272                         seq_printf(s, i ? ":%d" : ", dist %d",
 273                                 node_distance(ordinal, i));
 274                 }
 275                 seq_putc(s, '\n');
 276
 277                 /*
 278                  * CPUs on this node, if any
 279                  */
 280                 cpumask = node_to_cpumask(ordinal);
 281                 for_each_online_cpu(i) {
 282                         if (cpu_isset(i, cpumask)) {
 283                                 slice = 'a' + cpuid_to_slice(i);
 284                                 c = cpu_data(i);
 285                                 seq_printf(s, "cpu %d %s%c local"
 286                                         " freq %luMHz, arch ia64",
 287                                         i, obj->location, slice,
 288                                         c->proc_freq / 1000000);
 289                                 for_each_online_cpu(j) {
 290                                         seq_printf(s, j ? ":%d" : ", dist %d",
 291                                                 node_distance(
 292                                                     cpuid_to_cnodeid(i),
 293                                                     cpuid_to_cnodeid(j)));
 294                                 }
 295                                 seq_putc(s, '\n');
 296                         }
 297                 }
 298
 299                 /*
 300                  * PCI busses attached to this node, if any
 301                  */
 302                 do {
 303                         if (!(pci_topo_buf = vmalloc(pci_topo_buf_len))) {
 304                                 printk("sn_topology_show: kmalloc failed\n");
 305                                 break;
 306                         }
 307
 308                         if (sn_hwperf_location_to_bpos(obj->location,
 309                                 &rack, &bay, &slot, &slab) != 0)
 310                                 continue;
 311
 312                         e = ia64_sn_ioif_get_pci_topology(rack, bay, slot, slab,
 313                                 pci_topo_buf, pci_topo_buf_len);
 314
 315                         switch (e) {
 316                         case SALRET_NOT_IMPLEMENTED:
 317                         case SALRET_INVALID_ARG:
 318                                 /* ignore, don't print anything */
 319                                 e = SN_HWPERF_OP_OK;
 320                                 break;
 321
 322                         case SALRET_ERROR:
 323                                 /* retry with a bigger buffer */
 324                                 pci_topo_buf_len += 256;
 325                                 break;
 326
 327                         case SN_HWPERF_OP_OK:
 328                                 /* export pci bus info */
 329                                 print_pci_topology(s, obj, &pci_bus_ordinal,
 330                                         pci_topo_buf, pci_topo_buf_len);
 331                                 break;
 332                         }
 333                         vfree(pci_topo_buf);
 334                 } while (e != SN_HWPERF_OP_OK && pci_topo_buf_len < 0x200000);
 335         }
 336
 337         if (obj->ports) {
 338                 /*
 339                  * numalink ports
 340                  */
 341                 sz = obj->ports * sizeof(struct sn_hwperf_port_info);
 342                 if ((ptdata = vmalloc(sz)) == NULL)
 343                         return -ENOMEM;
 344                 e = ia64_sn_hwperf_op(sn_hwperf_master_nasid,
 345                                       SN_HWPERF_ENUM_PORTS, obj->id, sz,
 346                                       (u64) ptdata, 0, 0, NULL);
 347                 if (e != SN_HWPERF_OP_OK)
 348                         return -EINVAL;
 349                 for (ordinal=0, p=objs; p != obj; p++) {
 350                         if (!SN_HWPERF_FOREIGN(p))
 351                                 ordinal += p->ports;
 352                 }
 353                 for (pt = 0; pt < obj->ports; pt++) {
 354                         for (p = objs, i = 0; i < sn_hwperf_obj_cnt; i++, p++) {
 355                                 if (ptdata[pt].conn_id == p->id) {
 356                                         break;
 357                                 }
 358                         }
 359                         seq_printf(s, "numalink %d %s-%d",
 360                             ordinal+pt, obj->location, ptdata[pt].port);
 361
 362                         if (i >= sn_hwperf_obj_cnt) {
 363                                 /* no connection */
 364                                 seq_puts(s, " local endpoint disconnected"
 365                                             ", protocol unknown\n");
 366                                 continue;
 367                         }
 368
 369                         if (obj->sn_hwp_this_part && p->sn_hwp_this_part)
 370                                 /* both ends local to this partition */
 371                                 seq_puts(s, " local");
 372                         else if (!obj->sn_hwp_this_part && !p->sn_hwp_this_part)
 373                                 /* both ends of the link in foreign partiton */
 374                                 seq_puts(s, " foreign");
 375                         else
 376                                 /* link straddles a partition */
 377                                 seq_puts(s, " shared");
 378
 379                         /*
 380                          * Unlikely, but strictly should query the LLP config
 381                          * registers because an NL4R can be configured to run
 382                          * NL3 protocol, even when not talking to an NL3 router.
 383                          * Ditto for node-node.
 384                          */
 385                         seq_printf(s, " endpoint %s-%d, protocol %s\n",
 386                                 p->location, ptdata[pt].conn_port,
 387                                 (SN_HWPERF_IS_NL3ROUTER(obj) ||
 388                                 SN_HWPERF_IS_NL3ROUTER(p)) ?  "LLP3" : "LLP4");
 389                 }
 390                 vfree(ptdata);
 391         }
 392
 393         return 0;
 394 }
 395
 396 static void *sn_topology_start(struct seq_file *s, loff_t * pos)
 397 {
 398         struct sn_hwperf_object_info *objs = s->private;
 399
 400         if (*pos < sn_hwperf_obj_cnt)
 401                 return (void *)(objs + *pos);
 402
 403         return NULL;
 404 }
 405
 406 static void *sn_topology_next(struct seq_file *s, void *v, loff_t * pos)
 407 {
 408         ++*pos;
 409         return sn_topology_start(s, pos);
 410 }
 411
 412 static void sn_topology_stop(struct seq_file *m, void *v)
 413 {
 414         return;
 415 }
 416
 417 /*
 418  * /proc/sgi_sn/sn_topology, read-only using seq_file
 419  */
 420 static struct seq_operations sn_topology_seq_ops = {
 421         .start = sn_topology_start,
 422         .next = sn_topology_next,
 423         .stop = sn_topology_stop,
 424         .show = sn_topology_show
 425 };
 426
 427 struct sn_hwperf_op_info {
 428         u64 op;
 429         struct sn_hwperf_ioctl_args *a;
 430         void *p;
 431         int *v0;
 432         int ret;
 433 };
 434
 435 static void sn_hwperf_call_sal(void *info)
 436 {
 437         struct sn_hwperf_op_info *op_info = info;
 438         int r;
 439
 440         r = ia64_sn_hwperf_op(sn_hwperf_master_nasid, op_info->op,
 441                       op_info->a->arg, op_info->a->sz,
 442                       (u64) op_info->p, 0, 0, op_info->v0);
 443         op_info->ret = r;
 444 }
 445
 446 static int sn_hwperf_op_cpu(struct sn_hwperf_op_info *op_info)
 447 {
 448         u32 cpu;
 449         u32 use_ipi;
 450         int r = 0;
 451         cpumask_t save_allowed;
 452
 453         cpu = (op_info->a->arg & SN_HWPERF_ARG_CPU_MASK) >> 32;
 454         use_ipi = op_info->a->arg & SN_HWPERF_ARG_USE_IPI_MASK;
 455         op_info->a->arg &= SN_HWPERF_ARG_OBJID_MASK;
 456
 457         if (cpu != SN_HWPERF_ARG_ANY_CPU) {
 458                 if (cpu >= num_online_cpus() || !cpu_online(cpu)) {
 459                         r = -EINVAL;
 460                         goto out;
 461                 }
 462         }
 463
 464         if (cpu == SN_HWPERF_ARG_ANY_CPU || cpu == get_cpu()) {
 465                 /* don't care, or already on correct cpu */
 466                 sn_hwperf_call_sal(op_info);
 467         }
 468         else {
 469                 if (use_ipi) {
 470                         /* use an interprocessor interrupt to call SAL */
 471                         smp_call_function_single(cpu, sn_hwperf_call_sal,
 472                                 op_info, 1, 1);
 473                 }
 474                 else {
 475                         /* migrate the task before calling SAL */
 476                         save_allowed = current->cpus_allowed;
 477                         set_cpus_allowed(current, cpumask_of_cpu(cpu));
 478                         sn_hwperf_call_sal(op_info);
 479                         set_cpus_allowed(current, save_allowed);
 480                 }
 481         }
 482         r = op_info->ret;
 483
 484 out:
 485         return r;
 486 }
 487
 488 /* map SAL hwperf error code to system error code */
 489 static int sn_hwperf_map_err(int hwperf_err)
 490 {
 491         int e;
 492
 493         switch(hwperf_err) {
 494         case SN_HWPERF_OP_OK:
 495                 e = 0;
 496                 break;
 497
 498         case SN_HWPERF_OP_NOMEM:
 499                 e = -ENOMEM;
 500                 break;
 501
 502         case SN_HWPERF_OP_NO_PERM:
 503                 e = -EPERM;
 504                 break;
 505
 506         case SN_HWPERF_OP_IO_ERROR:
 507                 e = -EIO;
 508                 break;
 509
 510         case SN_HWPERF_OP_BUSY:
 511                 e = -EBUSY;
 512                 break;
 513
 514         case SN_HWPERF_OP_RECONFIGURE:
 515                 e = -EAGAIN;
 516                 break;
 517
 518         case SN_HWPERF_OP_INVAL:
 519         default:
 520                 e = -EINVAL;
 521                 break;
 522         }
 523
 524         return e;
 525 }
 526
 527 /*
 528  * ioctl for "sn_hwperf" misc device
 529  */
 530 static int
 531 sn_hwperf_ioctl(struct inode *in, struct file *fp, u32 op, u64 arg)
 532 {
 533         struct sn_hwperf_ioctl_args a;
 534         struct cpuinfo_ia64 *cdata;
 535         struct sn_hwperf_object_info *objs;
 536         struct sn_hwperf_object_info *cpuobj;
 537         struct sn_hwperf_op_info op_info;
 538         void *p = NULL;
 539         int nobj;
 540         char slice;
 541         int node;
 542         int r;
 543         int v0;
 544         int i;
 545         int j;
 546
 547         unlock_kernel();
 548
 549         /* only user requests are allowed here */
 550         if ((op & SN_HWPERF_OP_MASK) < 10) {
 551                 r = -EINVAL;
 552                 goto error;
 553         }
 554         r = copy_from_user(&a, (const void __user *)arg,
 555                 sizeof(struct sn_hwperf_ioctl_args));
 556         if (r != 0) {
 557                 r = -EFAULT;
 558                 goto error;
 559         }
 560
 561         /*
 562          * Allocate memory to hold a kernel copy of the user buffer. The
 563          * buffer contents are either copied in or out (or both) of user
 564          * space depending on the flags encoded in the requested operation.
 565          */
 566         if (a.ptr) {
 567                 p = vmalloc(a.sz);
 568                 if (!p) {
 569                         r = -ENOMEM;
 570                         goto error;
 571                 }
 572         }
 573
 574         if (op & SN_HWPERF_OP_MEM_COPYIN) {
 575                 r = copy_from_user(p, (const void __user *)a.ptr, a.sz);
 576                 if (r != 0) {
 577                         r = -EFAULT;
 578                         goto error;
 579                 }
 580         }
 581
 582         switch (op) {
 583         case SN_HWPERF_GET_CPU_INFO:
 584                 if (a.sz == sizeof(u64)) {
 585                         /* special case to get size needed */
 586                         *(u64 *) p = (u64) num_online_cpus() *
 587                                 sizeof(struct sn_hwperf_object_info);
 588                 } else
 589                 if (a.sz < num_online_cpus() * sizeof(struct sn_hwperf_object_info)) {
 590                         r = -ENOMEM;
 591                         goto error;
 592                 } else
 593                 if ((r = sn_hwperf_enum_objects(&nobj, &objs)) == 0) {
 594                         memset(p, 0, a.sz);
 595                         for (i = 0; i < nobj; i++) {
 596                                 node = sn_hwperf_obj_to_cnode(objs + i);
 597                                 for_each_online_cpu(j) {
 598                                         if (node != cpu_to_node(j))
 599                                                 continue;
 600                                         cpuobj = (struct sn_hwperf_object_info *) p + j;
 601                                         slice = 'a' + cpuid_to_slice(j);
 602                                         cdata = cpu_data(j);
 603                                         cpuobj->id = j;
 604                                         snprintf(cpuobj->name,
 605                                                  sizeof(cpuobj->name),
 606                                                  "CPU %luMHz %s",
 607                                                  cdata->proc_freq / 1000000,
 608                                                  cdata->vendor);
 609                                         snprintf(cpuobj->location,
 610                                                  sizeof(cpuobj->location),
 611                                                  "%s%c", objs[i].location,
 612                                                  slice);
 613                                 }
 614                         }
 615
 616                         vfree(objs);
 617                 }
 618                 break;
 619
 620         case SN_HWPERF_GET_NODE_NASID:
 621                 if (a.sz != sizeof(u64) ||
 622                    (node = a.arg) < 0 || node >= numionodes) {
 623                         r = -EINVAL;
 624                         goto error;
 625                 }
 626                 *(u64 *)p = (u64)cnodeid_to_nasid(node);
 627                 break;
 628
 629         case SN_HWPERF_GET_OBJ_NODE:
 630                 if (a.sz != sizeof(u64) || a.arg < 0) {
 631                         r = -EINVAL;
 632                         goto error;
 633                 }
 634                 if ((r = sn_hwperf_enum_objects(&nobj, &objs)) == 0) {
 635                         if (a.arg >= nobj) {
 636                                 r = -EINVAL;
 637                                 vfree(objs);
 638                                 goto error;
 639                         }
 640                         if (objs[(i = a.arg)].id != a.arg) {
 641                                 for (i = 0; i < nobj; i++) {
 642                                         if (objs[i].id == a.arg)
 643                                                 break;
 644                                 }
 645                         }
 646                         if (i == nobj) {
 647                                 r = -EINVAL;
 648                                 vfree(objs);
 649                                 goto error;
 650                         }
 651                         *(u64 *)p = (u64)sn_hwperf_obj_to_cnode(objs + i);
 652                         vfree(objs);
 653                 }
 654                 break;
 655
 656         case SN_HWPERF_GET_MMRS:
 657         case SN_HWPERF_SET_MMRS:
 658         case SN_HWPERF_OBJECT_DISTANCE:
 659                 op_info.p = p;
 660                 op_info.a = &a;
 661                 op_info.v0 = &v0;
 662                 op_info.op = op;
 663                 r = sn_hwperf_op_cpu(&op_info);
 664                 if (r) {
 665                         r = sn_hwperf_map_err(r);
 666                         a.v0 = v0;
 667                         goto error;
 668                 }
 669                 break;
 670
 671         default:
 672                 /* all other ops are a direct SAL call */
 673                 r = ia64_sn_hwperf_op(sn_hwperf_master_nasid, op,
 674                               a.arg, a.sz, (u64) p, 0, 0, &v0);
 675                 if (r) {
 676                         r = sn_hwperf_map_err(r);
 677                         goto error;
 678                 }
 679                 a.v0 = v0;
 680                 break;
 681         }
 682
 683         if (op & SN_HWPERF_OP_MEM_COPYOUT) {
 684                 r = copy_to_user((void __user *)a.ptr, p, a.sz);
 685                 if (r != 0) {
 686                         r = -EFAULT;
 687                         goto error;
 688                 }
 689         }
 690
 691 error:
 692         vfree(p);
 693
 694         lock_kernel();
 695         return r;
 696 }
 697
 698 static struct file_operations sn_hwperf_fops = {
 699         .ioctl = sn_hwperf_ioctl,
 700 };
 701
 702 static struct miscdevice sn_hwperf_dev = {
 703         MISC_DYNAMIC_MINOR,
 704         "sn_hwperf",
 705         &sn_hwperf_fops
 706 };
 707
 708 static int sn_hwperf_init(void)
 709 {
 710         u64 v;
 711         int salr;
 712         int e = 0;
 713
 714         /* single threaded, once-only initialization */
 715         down(&sn_hwperf_init_mutex);
 716         if (sn_hwperf_salheap) {
 717                 up(&sn_hwperf_init_mutex);
 718                 return e;
 719         }
 720
 721         /*
 722          * The PROM code needs a fixed reference node. For convenience the
 723          * same node as the console I/O is used.
 724          */
 725         sn_hwperf_master_nasid = (nasid_t) ia64_sn_get_console_nasid();
 726
 727         /*
 728          * Request the needed size and install the PROM scratch area.
 729          * The PROM keeps various tracking bits in this memory area.
 730          */
 731         salr = ia64_sn_hwperf_op(sn_hwperf_master_nasid,
 732                                  (u64) SN_HWPERF_GET_HEAPSIZE, 0,
 733                                  (u64) sizeof(u64), (u64) &v, 0, 0, NULL);
 734         if (salr != SN_HWPERF_OP_OK) {
 735                 e = -EINVAL;
 736                 goto out;
 737         }
 738
 739         if ((sn_hwperf_salheap = vmalloc(v)) == NULL) {
 740                 e = -ENOMEM;
 741                 goto out;
 742         }
 743         salr = ia64_sn_hwperf_op(sn_hwperf_master_nasid,
 744                                  SN_HWPERF_INSTALL_HEAP, 0, v,
 745                                  (u64) sn_hwperf_salheap, 0, 0, NULL);
 746         if (salr != SN_HWPERF_OP_OK) {
 747                 e = -EINVAL;
 748                 goto out;
 749         }
 750
 751         salr = ia64_sn_hwperf_op(sn_hwperf_master_nasid,
 752                                  SN_HWPERF_OBJECT_COUNT, 0,
 753                                  sizeof(u64), (u64) &v, 0, 0, NULL);
 754         if (salr != SN_HWPERF_OP_OK) {
 755                 e = -EINVAL;
 756                 goto out;
 757         }
 758         sn_hwperf_obj_cnt = (int)v;
 759
 760 out:
 761         if (e < 0 && sn_hwperf_salheap) {
 762                 vfree(sn_hwperf_salheap);
 763                 sn_hwperf_salheap = NULL;
 764                 sn_hwperf_obj_cnt = 0;
 765         }
 766
 767         if (!e) {
 768                 /*
 769                  * Register a dynamic misc device for ioctl. Platforms
 770                  * supporting hotplug will create /dev/sn_hwperf, else
 771                  * user can to look up the minor number in /proc/misc.
 772                  */
 773                 if ((e = misc_register(&sn_hwperf_dev)) != 0) {
 774                         printk(KERN_ERR "sn_hwperf_init: misc register "
 775                                "for \"sn_hwperf\" failed, err %d\n", e);
 776                 }
 777         }
 778
 779         up(&sn_hwperf_init_mutex);
 780         return e;
 781 }
 782
 783 int sn_topology_open(struct inode *inode, struct file *file)
 784 {
 785         int e;
 786         struct seq_file *seq;
 787         struct sn_hwperf_object_info *objbuf;
 788         int nobj;
 789
 790         if ((e = sn_hwperf_enum_objects(&nobj, &objbuf)) == 0) {
 791                 e = seq_open(file, &sn_topology_seq_ops);
 792                 seq = file->private_data;
 793                 seq->private = objbuf;
 794         }
 795
 796         return e;
 797 }
 798
 799 int sn_topology_release(struct inode *inode, struct file *file)
 800 {
 801         struct seq_file *seq = file->private_data;
 802
 803         vfree(seq->private);
 804         return seq_release(inode, file);
 805 }