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 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/cpumask.h>
  32 #include <linux/smp_lock.h>
  33 #include <asm/processor.h>
  34 #include <asm/topology.h>
  35 #include <asm/smp.h>
  36 #include <asm/semaphore.h>
  37 #include <asm/segment.h>
  38 #include <asm/uaccess.h>
  39 #include <asm/sal.h>
  40 #include <asm/sn/io.h>
  41 #include <asm/sn/sn_sal.h>
  42 #include <asm/sn/module.h>
  43 #include <asm/sn/geo.h>
  44 #include <asm/sn/sn2/sn_hwperf.h>
  45
  46 static void *sn_hwperf_salheap = NULL;
  47 static int sn_hwperf_obj_cnt = 0;
  48 static nasid_t sn_hwperf_master_nasid = INVALID_NASID;
  49 static int sn_hwperf_init(void);
  50 static DECLARE_MUTEX(sn_hwperf_init_mutex);
  51
  52 static int sn_hwperf_enum_objects(int *nobj, struct sn_hwperf_object_info **ret)
  53 {
  54         int e;
  55         u64 sz;
  56         struct sn_hwperf_object_info *objbuf = NULL;
  57
  58         if ((e = sn_hwperf_init()) < 0) {
  59                 printk("sn_hwperf_init failed: err %d\n", e);
  60                 goto out;
  61         }
  62
  63         sz = sn_hwperf_obj_cnt * sizeof(struct sn_hwperf_object_info);
  64         if ((objbuf = (struct sn_hwperf_object_info *) vmalloc(sz)) == NULL) {
  65                 printk("sn_hwperf_enum_objects: vmalloc(%d) failed\n", (int)sz);
  66                 e = -ENOMEM;
  67                 goto out;
  68         }
  69
  70         e = ia64_sn_hwperf_op(sn_hwperf_master_nasid, SN_HWPERF_ENUM_OBJECTS,
  71                 0, sz, (u64) objbuf, 0, 0, NULL);
  72         if (e != SN_HWPERF_OP_OK) {
  73                 e = -EINVAL;
  74                 vfree(objbuf);
  75         }
  76
  77 out:
  78         *nobj = sn_hwperf_obj_cnt;
  79         *ret = objbuf;
  80         return e;
  81 }
  82
  83 static int sn_hwperf_geoid_to_cnode(char *location)
  84 {
  85         int cnode;
  86         geoid_t geoid;
  87         moduleid_t module_id;
  88         char type;
  89         int rack, slot, slab;
  90         int this_rack, this_slot, this_slab;
  91
  92         if (sscanf(location, "%03d%c%02d#%d", &rack, &type, &slot, &slab) != 4)
  93                 return -1;
  94
  95         for (cnode = 0; cnode < numionodes; cnode++) {
  96                 geoid = cnodeid_get_geoid(cnode);
  97                 module_id = geo_module(geoid);
  98                 this_rack = MODULE_GET_RACK(module_id);
  99                 this_slot = MODULE_GET_BPOS(module_id);
 100                 this_slab = geo_slab(geoid);
 101                 if (rack == this_rack && slot == this_slot && slab == this_slab)
 102                         break;
 103         }
 104
 105         return cnode < numionodes ? cnode : -1;
 106 }
 107
 108 static int sn_hwperf_obj_to_cnode(struct sn_hwperf_object_info * obj)
 109 {
 110         if (!obj->sn_hwp_this_part)
 111                 return -1;
 112         return sn_hwperf_geoid_to_cnode(obj->location);
 113 }
 114
 115 static int sn_hwperf_generic_ordinal(struct sn_hwperf_object_info *obj,
 116                                 struct sn_hwperf_object_info *objs)
 117 {
 118         int ordinal;
 119         struct sn_hwperf_object_info *p;
 120
 121         for (ordinal=0, p=objs; p != obj; p++) {
 122                 if (SN_HWPERF_FOREIGN(p))
 123                         continue;
 124                 if (SN_HWPERF_SAME_OBJTYPE(p, obj))
 125                         ordinal++;
 126         }
 127
 128         return ordinal;
 129 }
 130
 131 static const char *slabname_node =      "node"; /* SHub asic */
 132 static const char *slabname_ionode =    "ionode"; /* TIO asic */
 133 static const char *slabname_router =    "router"; /* NL3R or NL4R */
 134 static const char *slabname_other =     "other"; /* unknown asic */
 135
 136 static const char *sn_hwperf_get_slabname(struct sn_hwperf_object_info *obj,
 137                         struct sn_hwperf_object_info *objs, int *ordinal)
 138 {
 139         int isnode;
 140         const char *slabname = slabname_other;
 141
 142         if ((isnode = SN_HWPERF_IS_NODE(obj)) || SN_HWPERF_IS_IONODE(obj)) {
 143                 slabname = isnode ? slabname_node : slabname_ionode;
 144                 *ordinal = sn_hwperf_obj_to_cnode(obj);
 145         }
 146         else {
 147                 *ordinal = sn_hwperf_generic_ordinal(obj, objs);
 148                 if (SN_HWPERF_IS_ROUTER(obj))
 149                         slabname = slabname_router;
 150         }
 151
 152         return slabname;
 153 }
 154
 155 static int sn_topology_show(struct seq_file *s, void *d)
 156 {
 157         int sz;
 158         int pt;
 159         int e;
 160         int i;
 161         int j;
 162         const char *slabname;
 163         int ordinal;
 164         cpumask_t cpumask;
 165         char slice;
 166         struct cpuinfo_ia64 *c;
 167         struct sn_hwperf_port_info *ptdata;
 168         struct sn_hwperf_object_info *p;
 169         struct sn_hwperf_object_info *obj = d;  /* this object */
 170         struct sn_hwperf_object_info *objs = s->private; /* all objects */
 171
 172         if (obj == objs) {
 173                 seq_printf(s, "# sn_topology version 1\n");
 174                 seq_printf(s, "# objtype ordinal location partition"
 175                         " [attribute value [, ...]]\n");
 176         }
 177
 178         if (SN_HWPERF_FOREIGN(obj)) {
 179                 /* private in another partition: not interesting */
 180                 return 0;
 181         }
 182
 183         for (i = 0; obj->name[i]; i++) {
 184                 if (obj->name[i] == ' ')
 185                         obj->name[i] = '_';
 186         }
 187
 188         slabname = sn_hwperf_get_slabname(obj, objs, &ordinal);
 189         seq_printf(s, "%s %d %s %s asic %s", slabname, ordinal, obj->location,
 190                 obj->sn_hwp_this_part ? "local" : "shared", obj->name);
 191
 192         if (!SN_HWPERF_IS_NODE(obj) && !SN_HWPERF_IS_IONODE(obj))
 193                 seq_putc(s, '\n');
 194         else {
 195                 seq_printf(s, ", nasid 0x%x", cnodeid_to_nasid(ordinal));
 196                 for (i=0; i < numionodes; i++) {
 197                         seq_printf(s, i ? ":%d" : ", dist %d",
 198                                 node_distance(ordinal, i));
 199                 }
 200                 seq_putc(s, '\n');
 201
 202                 /*
 203                  * CPUs on this node, if any
 204                  */
 205                 cpumask = node_to_cpumask(ordinal);
 206                 for_each_online_cpu(i) {
 207                         if (cpu_isset(i, cpumask)) {
 208                                 slice = 'a' + cpuid_to_slice(i);
 209                                 c = cpu_data(i);
 210                                 seq_printf(s, "cpu %d %s%c local"
 211                                         " freq %luMHz, arch ia64",
 212                                         i, obj->location, slice,
 213                                         c->proc_freq / 1000000);
 214                                 for_each_online_cpu(j) {
 215                                         seq_printf(s, j ? ":%d" : ", dist %d",
 216                                                 node_distance(
 217                                                     cpuid_to_cnodeid(i),
 218                                                     cpuid_to_cnodeid(j)));
 219                                 }
 220                                 seq_putc(s, '\n');
 221                         }
 222                 }
 223         }
 224
 225         if (obj->ports) {
 226                 /*
 227                  * numalink ports
 228                  */
 229                 sz = obj->ports * sizeof(struct sn_hwperf_port_info);
 230                 if ((ptdata = vmalloc(sz)) == NULL)
 231                         return -ENOMEM;
 232                 e = ia64_sn_hwperf_op(sn_hwperf_master_nasid,
 233                                       SN_HWPERF_ENUM_PORTS, obj->id, sz,
 234                                       (u64) ptdata, 0, 0, NULL);
 235                 if (e != SN_HWPERF_OP_OK)
 236                         return -EINVAL;
 237                 for (ordinal=0, p=objs; p != obj; p++) {
 238                         if (!SN_HWPERF_FOREIGN(p))
 239                                 ordinal += p->ports;
 240                 }
 241                 for (pt = 0; pt < obj->ports; pt++) {
 242                         for (p = objs, i = 0; i < sn_hwperf_obj_cnt; i++, p++) {
 243                                 if (ptdata[pt].conn_id == p->id) {
 244                                         break;
 245                                 }
 246                         }
 247                         seq_printf(s, "numalink %d %s-%d",
 248                             ordinal+pt, obj->location, ptdata[pt].port);
 249
 250                         if (i >= sn_hwperf_obj_cnt) {
 251                                 /* no connection */
 252                                 seq_puts(s, " local endpoint disconnected"
 253                                             ", protocol unknown\n");
 254                                 continue;
 255                         }
 256
 257                         if (obj->sn_hwp_this_part && p->sn_hwp_this_part)
 258                                 /* both ends local to this partition */
 259                                 seq_puts(s, " local");
 260                         else if (!obj->sn_hwp_this_part && !p->sn_hwp_this_part)
 261                                 /* both ends of the link in foreign partiton */
 262                                 seq_puts(s, " foreign");
 263                         else
 264                                 /* link straddles a partition */
 265                                 seq_puts(s, " shared");
 266
 267                         /*
 268                          * Unlikely, but strictly should query the LLP config
 269                          * registers because an NL4R can be configured to run
 270                          * NL3 protocol, even when not talking to an NL3 router.
 271                          * Ditto for node-node.
 272                          */
 273                         seq_printf(s, " endpoint %s-%d, protocol %s\n",
 274                                 p->location, ptdata[pt].conn_port,
 275                                 (SN_HWPERF_IS_NL3ROUTER(obj) ||
 276                                 SN_HWPERF_IS_NL3ROUTER(p)) ?  "LLP3" : "LLP4");
 277                 }
 278                 vfree(ptdata);
 279         }
 280
 281         return 0;
 282 }
 283
 284 static void *sn_topology_start(struct seq_file *s, loff_t * pos)
 285 {
 286         struct sn_hwperf_object_info *objs = s->private;
 287
 288         if (*pos < sn_hwperf_obj_cnt)
 289                 return (void *)(objs + *pos);
 290
 291         return NULL;
 292 }
 293
 294 static void *sn_topology_next(struct seq_file *s, void *v, loff_t * pos)
 295 {
 296         ++*pos;
 297         return sn_topology_start(s, pos);
 298 }
 299
 300 static void sn_topology_stop(struct seq_file *m, void *v)
 301 {
 302         return;
 303 }
 304
 305 /*
 306  * /proc/sgi_sn/sn_topology, read-only using seq_file
 307  */
 308 static struct seq_operations sn_topology_seq_ops = {
 309         .start = sn_topology_start,
 310         .next = sn_topology_next,
 311         .stop = sn_topology_stop,
 312         .show = sn_topology_show
 313 };
 314
 315 struct sn_hwperf_op_info {
 316         u64 op;
 317         struct sn_hwperf_ioctl_args *a;
 318         void *p;
 319         int *v0;
 320         int ret;
 321 };
 322
 323 static void sn_hwperf_call_sal(void *info)
 324 {
 325         struct sn_hwperf_op_info *op_info = info;
 326         int r;
 327
 328         r = ia64_sn_hwperf_op(sn_hwperf_master_nasid, op_info->op,
 329                       op_info->a->arg, op_info->a->sz,
 330                       (u64) op_info->p, 0, 0, op_info->v0);
 331         op_info->ret = r;
 332 }
 333
 334 static int sn_hwperf_op_cpu(struct sn_hwperf_op_info *op_info)
 335 {
 336         u32 cpu;
 337         u32 use_ipi;
 338         int r = 0;
 339         cpumask_t save_allowed;
 340
 341         cpu = (op_info->a->arg & SN_HWPERF_ARG_CPU_MASK) >> 32;
 342         use_ipi = op_info->a->arg & SN_HWPERF_ARG_USE_IPI_MASK;
 343         op_info->a->arg &= SN_HWPERF_ARG_OBJID_MASK;
 344
 345         if (cpu != SN_HWPERF_ARG_ANY_CPU) {
 346                 if (cpu >= num_online_cpus() || !cpu_online(cpu)) {
 347                         r = -EINVAL;
 348                         goto out;
 349                 }
 350         }
 351
 352         if (cpu == SN_HWPERF_ARG_ANY_CPU || cpu == get_cpu()) {
 353                 /* don't care, or already on correct cpu */
 354                 sn_hwperf_call_sal(op_info);
 355         }
 356         else {
 357                 if (use_ipi) {
 358                         /* use an interprocessor interrupt to call SAL */
 359                         smp_call_function_single(cpu, sn_hwperf_call_sal,
 360                                 op_info, 1, 1);
 361                 }
 362                 else {
 363                         /* migrate the task before calling SAL */
 364                         save_allowed = current->cpus_allowed;
 365                         set_cpus_allowed(current, cpumask_of_cpu(cpu));
 366                         sn_hwperf_call_sal(op_info);
 367                         set_cpus_allowed(current, save_allowed);
 368                 }
 369         }
 370         r = op_info->ret;
 371
 372 out:
 373         return r;
 374 }
 375
 376 /*
 377  * ioctl for "sn_hwperf" misc device
 378  */
 379 static int
 380 sn_hwperf_ioctl(struct inode *in, struct file *fp, u32 op, u64 arg)
 381 {
 382         struct sn_hwperf_ioctl_args a;
 383         struct cpuinfo_ia64 *cdata;
 384         struct sn_hwperf_object_info *objs;
 385         struct sn_hwperf_object_info *cpuobj;
 386         struct sn_hwperf_op_info op_info;
 387         void *p = NULL;
 388         int nobj;
 389         char slice;
 390         int node;
 391         int r;
 392         int v0;
 393         int i;
 394         int j;
 395
 396         unlock_kernel();
 397
 398         /* only user requests are allowed here */
 399         if ((op & SN_HWPERF_OP_MASK) < 10) {
 400                 r = -EINVAL;
 401                 goto error;
 402         }
 403         r = copy_from_user(&a, (const void __user *)arg,
 404                 sizeof(struct sn_hwperf_ioctl_args));
 405         if (r != 0) {
 406                 r = -EFAULT;
 407                 goto error;
 408         }
 409
 410         /*
 411          * Allocate memory to hold a kernel copy of the user buffer. The
 412          * buffer contents are either copied in or out (or both) of user
 413          * space depending on the flags encoded in the requested operation.
 414          */
 415         if (a.ptr) {
 416                 p = vmalloc(a.sz);
 417                 if (!p) {
 418                         r = -ENOMEM;
 419                         goto error;
 420                 }
 421         }
 422
 423         if (op & SN_HWPERF_OP_MEM_COPYIN) {
 424                 r = copy_from_user(p, (const void __user *)a.ptr, a.sz);
 425                 if (r != 0) {
 426                         r = -EFAULT;
 427                         goto error;
 428                 }
 429         }
 430
 431         switch (op) {
 432         case SN_HWPERF_GET_CPU_INFO:
 433                 if (a.sz == sizeof(u64)) {
 434                         /* special case to get size needed */
 435                         *(u64 *) p = (u64) num_online_cpus() *
 436                                 sizeof(struct sn_hwperf_object_info);
 437                 } else
 438                 if (a.sz < num_online_cpus() * sizeof(struct sn_hwperf_object_info)) {
 439                         r = -ENOMEM;
 440                         goto error;
 441                 } else
 442                 if ((r = sn_hwperf_enum_objects(&nobj, &objs)) == 0) {
 443                         memset(p, 0, a.sz);
 444                         for (i = 0; i < nobj; i++) {
 445                                 node = sn_hwperf_obj_to_cnode(objs + i);
 446                                 for_each_online_cpu(j) {
 447                                         if (node != cpu_to_node(j))
 448                                                 continue;
 449                                         cpuobj = (struct sn_hwperf_object_info *) p + j;
 450                                         slice = 'a' + cpuid_to_slice(j);
 451                                         cdata = cpu_data(j);
 452                                         cpuobj->id = j;
 453                                         snprintf(cpuobj->name,
 454                                                  sizeof(cpuobj->name),
 455                                                  "CPU %luMHz %s",
 456                                                  cdata->proc_freq / 1000000,
 457                                                  cdata->vendor);
 458                                         snprintf(cpuobj->location,
 459                                                  sizeof(cpuobj->location),
 460                                                  "%s%c", objs[i].location,
 461                                                  slice);
 462                                 }
 463                         }
 464
 465                         vfree(objs);
 466                 }
 467                 break;
 468
 469         case SN_HWPERF_GET_NODE_NASID:
 470                 if (a.sz != sizeof(u64) ||
 471                    (node = a.arg) < 0 || node >= numionodes) {
 472                         r = -EINVAL;
 473                         goto error;
 474                 }
 475                 *(u64 *)p = (u64)cnodeid_to_nasid(node);
 476                 break;
 477
 478         case SN_HWPERF_GET_OBJ_NODE:
 479                 if (a.sz != sizeof(u64) || a.arg < 0) {
 480                         r = -EINVAL;
 481                         goto error;
 482                 }
 483                 if ((r = sn_hwperf_enum_objects(&nobj, &objs)) == 0) {
 484                         if (a.arg >= nobj) {
 485                                 r = -EINVAL;
 486                                 vfree(objs);
 487                                 goto error;
 488                         }
 489                         if (objs[(i = a.arg)].id != a.arg) {
 490                                 for (i = 0; i < nobj; i++) {
 491                                         if (objs[i].id == a.arg)
 492                                                 break;
 493                                 }
 494                         }
 495                         if (i == nobj) {
 496                                 r = -EINVAL;
 497                                 vfree(objs);
 498                                 goto error;
 499                         }
 500                         *(u64 *)p = (u64)sn_hwperf_obj_to_cnode(objs + i);
 501                         vfree(objs);
 502                 }
 503                 break;
 504
 505         case SN_HWPERF_GET_MMRS:
 506         case SN_HWPERF_SET_MMRS:
 507         case SN_HWPERF_OBJECT_DISTANCE:
 508                 op_info.p = p;
 509                 op_info.a = &a;
 510                 op_info.v0 = &v0;
 511                 op_info.op = op;
 512                 r = sn_hwperf_op_cpu(&op_info);
 513                 break;
 514
 515         default:
 516                 /* all other ops are a direct SAL call */
 517                 r = ia64_sn_hwperf_op(sn_hwperf_master_nasid, op,
 518                               a.arg, a.sz, (u64) p, 0, 0, &v0);
 519                 a.v0 = v0;
 520                 break;
 521         }
 522
 523         if (op & SN_HWPERF_OP_MEM_COPYOUT) {
 524                 r = copy_to_user((void __user *)a.ptr, p, a.sz);
 525                 if (r != 0) {
 526                         r = -EFAULT;
 527                         goto error;
 528                 }
 529         }
 530
 531 error:
 532         if (p)
 533                 vfree(p);
 534
 535         lock_kernel();
 536         return r;
 537 }
 538
 539 static struct file_operations sn_hwperf_fops = {
 540         .ioctl = sn_hwperf_ioctl,
 541 };
 542
 543 static struct miscdevice sn_hwperf_dev = {
 544         MISC_DYNAMIC_MINOR,
 545         "sn_hwperf",
 546         &sn_hwperf_fops
 547 };
 548
 549 static int sn_hwperf_init(void)
 550 {
 551         u64 v;
 552         int salr;
 553         int e = 0;
 554
 555         /* single threaded, once-only initialization */
 556         down(&sn_hwperf_init_mutex);
 557         if (sn_hwperf_salheap) {
 558                 up(&sn_hwperf_init_mutex);
 559                 return e;
 560         }
 561
 562         /*
 563          * The PROM code needs a fixed reference node. For convenience the
 564          * same node as the console I/O is used.
 565          */
 566         sn_hwperf_master_nasid = (nasid_t) ia64_sn_get_console_nasid();
 567
 568         /*
 569          * Request the needed size and install the PROM scratch area.
 570          * The PROM keeps various tracking bits in this memory area.
 571          */
 572         salr = ia64_sn_hwperf_op(sn_hwperf_master_nasid,
 573                                  (u64) SN_HWPERF_GET_HEAPSIZE, 0,
 574                                  (u64) sizeof(u64), (u64) &v, 0, 0, NULL);
 575         if (salr != SN_HWPERF_OP_OK) {
 576                 e = -EINVAL;
 577                 goto out;
 578         }
 579
 580         if ((sn_hwperf_salheap = vmalloc(v)) == NULL) {
 581                 e = -ENOMEM;
 582                 goto out;
 583         }
 584         salr = ia64_sn_hwperf_op(sn_hwperf_master_nasid,
 585                                  SN_HWPERF_INSTALL_HEAP, 0, v,
 586                                  (u64) sn_hwperf_salheap, 0, 0, NULL);
 587         if (salr != SN_HWPERF_OP_OK) {
 588                 e = -EINVAL;
 589                 goto out;
 590         }
 591
 592         salr = ia64_sn_hwperf_op(sn_hwperf_master_nasid,
 593                                  SN_HWPERF_OBJECT_COUNT, 0,
 594                                  sizeof(u64), (u64) &v, 0, 0, NULL);
 595         if (salr != SN_HWPERF_OP_OK) {
 596                 e = -EINVAL;
 597                 goto out;
 598         }
 599         sn_hwperf_obj_cnt = (int)v;
 600
 601 out:
 602         if (e < 0 && sn_hwperf_salheap) {
 603                 vfree(sn_hwperf_salheap);
 604                 sn_hwperf_salheap = NULL;
 605                 sn_hwperf_obj_cnt = 0;
 606         }
 607
 608         if (!e) {
 609                 /*
 610                  * Register a dynamic misc device for ioctl. Platforms
 611                  * supporting hotplug will create /dev/sn_hwperf, else
 612                  * user can to look up the minor number in /proc/misc.
 613                  */
 614                 if ((e = misc_register(&sn_hwperf_dev)) != 0) {
 615                         printk(KERN_ERR "sn_hwperf_init: misc register "
 616                                "for \"sn_hwperf\" failed, err %d\n", e);
 617                 }
 618         }
 619
 620         up(&sn_hwperf_init_mutex);
 621         return e;
 622 }
 623
 624 int sn_topology_open(struct inode *inode, struct file *file)
 625 {
 626         int e;
 627         struct seq_file *seq;
 628         struct sn_hwperf_object_info *objbuf;
 629         int nobj;
 630
 631         if ((e = sn_hwperf_enum_objects(&nobj, &objbuf)) == 0) {
 632                 e = seq_open(file, &sn_topology_seq_ops);
 633                 seq = file->private_data;
 634                 seq->private = objbuf;
 635         }
 636
 637         return e;
 638 }
 639
 640 int sn_topology_release(struct inode *inode, struct file *file)
 641 {
 642         struct seq_file *seq = file->private_data;
 643
 644         if (seq->private)
 645                 vfree(seq->private);
 646         return seq_release(inode, file);
 647 }