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
6 * Copyright (C) 2004 Silicon Graphics, Inc. All rights reserved.
8 * SGI Altix topology and hardware performance monitoring API.
9 * Mark Goodwin <markgw@sgi.com>.
11 * Creates /proc/sgi_sn/sn_topology (read-only) to export
12 * info about Altix nodes, routers, CPUs and NumaLink
13 * interconnection/topology.
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.
22 * This API is used by SGI Performance Co-Pilot and other
23 * tools, see http://oss.sgi.com/projects/pcp
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 <linux/nodemask.h>
34 #include <asm/processor.h>
35 #include <asm/topology.h>
37 #include <asm/semaphore.h>
38 #include <asm/segment.h>
39 #include <asm/uaccess.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>
47 static void *sn_hwperf_salheap = NULL;
48 static int sn_hwperf_obj_cnt = 0;
49 static nasid_t sn_hwperf_master_nasid = INVALID_NASID;
50 static int sn_hwperf_init(void);
51 static DECLARE_MUTEX(sn_hwperf_init_mutex);
53 static int sn_hwperf_enum_objects(int *nobj, struct sn_hwperf_object_info **ret)
57 struct sn_hwperf_object_info *objbuf = NULL;
59 if ((e = sn_hwperf_init()) < 0) {
60 printk("sn_hwperf_init failed: err %d\n", e);
64 sz = sn_hwperf_obj_cnt * sizeof(struct sn_hwperf_object_info);
65 if ((objbuf = (struct sn_hwperf_object_info *) vmalloc(sz)) == NULL) {
66 printk("sn_hwperf_enum_objects: vmalloc(%d) failed\n", (int)sz);
71 e = ia64_sn_hwperf_op(sn_hwperf_master_nasid, SN_HWPERF_ENUM_OBJECTS,
72 0, sz, (u64) objbuf, 0, 0, NULL);
73 if (e != SN_HWPERF_OP_OK) {
79 *nobj = sn_hwperf_obj_cnt;
84 static int sn_hwperf_geoid_to_cnode(char *location)
91 int this_rack, this_slot, this_slab;
93 if (sscanf(location, "%03d%c%02d#%d", &rack, &type, &slot, &slab) != 4)
96 for (cnode = 0; cnode < numionodes; cnode++) {
97 geoid = cnodeid_get_geoid(cnode);
98 module_id = geo_module(geoid);
99 this_rack = MODULE_GET_RACK(module_id);
100 this_slot = MODULE_GET_BPOS(module_id);
101 this_slab = geo_slab(geoid);
102 if (rack == this_rack && slot == this_slot && slab == this_slab)
106 return cnode < numionodes ? cnode : -1;
109 static int sn_hwperf_obj_to_cnode(struct sn_hwperf_object_info * obj)
111 if (!obj->sn_hwp_this_part)
113 return sn_hwperf_geoid_to_cnode(obj->location);
116 static int sn_hwperf_generic_ordinal(struct sn_hwperf_object_info *obj,
117 struct sn_hwperf_object_info *objs)
120 struct sn_hwperf_object_info *p;
122 for (ordinal=0, p=objs; p != obj; p++) {
123 if (SN_HWPERF_FOREIGN(p))
125 if (SN_HWPERF_SAME_OBJTYPE(p, obj))
132 static const char *slabname_node = "node"; /* SHub asic */
133 static const char *slabname_ionode = "ionode"; /* TIO asic */
134 static const char *slabname_router = "router"; /* NL3R or NL4R */
135 static const char *slabname_other = "other"; /* unknown asic */
137 static const char *sn_hwperf_get_slabname(struct sn_hwperf_object_info *obj,
138 struct sn_hwperf_object_info *objs, int *ordinal)
141 const char *slabname = slabname_other;
143 if ((isnode = SN_HWPERF_IS_NODE(obj)) || SN_HWPERF_IS_IONODE(obj)) {
144 slabname = isnode ? slabname_node : slabname_ionode;
145 *ordinal = sn_hwperf_obj_to_cnode(obj);
148 *ordinal = sn_hwperf_generic_ordinal(obj, objs);
149 if (SN_HWPERF_IS_ROUTER(obj))
150 slabname = slabname_router;
156 static int sn_topology_show(struct seq_file *s, void *d)
163 const char *slabname;
167 struct cpuinfo_ia64 *c;
168 struct sn_hwperf_port_info *ptdata;
169 struct sn_hwperf_object_info *p;
170 struct sn_hwperf_object_info *obj = d; /* this object */
171 struct sn_hwperf_object_info *objs = s->private; /* all objects */
174 seq_printf(s, "# sn_topology version 1\n");
175 seq_printf(s, "# objtype ordinal location partition"
176 " [attribute value [, ...]]\n");
179 if (SN_HWPERF_FOREIGN(obj)) {
180 /* private in another partition: not interesting */
184 for (i = 0; obj->name[i]; i++) {
185 if (obj->name[i] == ' ')
189 slabname = sn_hwperf_get_slabname(obj, objs, &ordinal);
190 seq_printf(s, "%s %d %s %s asic %s", slabname, ordinal, obj->location,
191 obj->sn_hwp_this_part ? "local" : "shared", obj->name);
193 if (!SN_HWPERF_IS_NODE(obj) && !SN_HWPERF_IS_IONODE(obj))
196 seq_printf(s, ", nasid 0x%x", cnodeid_to_nasid(ordinal));
197 for (i=0; i < numionodes; i++) {
198 seq_printf(s, i ? ":%d" : ", dist %d",
199 node_distance(ordinal, i));
204 * CPUs on this node, if any
206 cpumask = node_to_cpumask(ordinal);
207 for_each_online_cpu(i) {
208 if (cpu_isset(i, cpumask)) {
209 slice = 'a' + cpuid_to_slice(i);
211 seq_printf(s, "cpu %d %s%c local"
212 " freq %luMHz, arch ia64",
213 i, obj->location, slice,
214 c->proc_freq / 1000000);
215 for_each_online_cpu(j) {
216 seq_printf(s, j ? ":%d" : ", dist %d",
219 cpuid_to_cnodeid(j)));
230 sz = obj->ports * sizeof(struct sn_hwperf_port_info);
231 if ((ptdata = vmalloc(sz)) == NULL)
233 e = ia64_sn_hwperf_op(sn_hwperf_master_nasid,
234 SN_HWPERF_ENUM_PORTS, obj->id, sz,
235 (u64) ptdata, 0, 0, NULL);
236 if (e != SN_HWPERF_OP_OK)
238 for (ordinal=0, p=objs; p != obj; p++) {
239 if (!SN_HWPERF_FOREIGN(p))
242 for (pt = 0; pt < obj->ports; pt++) {
243 for (p = objs, i = 0; i < sn_hwperf_obj_cnt; i++, p++) {
244 if (ptdata[pt].conn_id == p->id) {
248 seq_printf(s, "numalink %d %s-%d",
249 ordinal+pt, obj->location, ptdata[pt].port);
251 if (i >= sn_hwperf_obj_cnt) {
253 seq_puts(s, " local endpoint disconnected"
254 ", protocol unknown\n");
258 if (obj->sn_hwp_this_part && p->sn_hwp_this_part)
259 /* both ends local to this partition */
260 seq_puts(s, " local");
261 else if (!obj->sn_hwp_this_part && !p->sn_hwp_this_part)
262 /* both ends of the link in foreign partiton */
263 seq_puts(s, " foreign");
265 /* link straddles a partition */
266 seq_puts(s, " shared");
269 * Unlikely, but strictly should query the LLP config
270 * registers because an NL4R can be configured to run
271 * NL3 protocol, even when not talking to an NL3 router.
272 * Ditto for node-node.
274 seq_printf(s, " endpoint %s-%d, protocol %s\n",
275 p->location, ptdata[pt].conn_port,
276 (SN_HWPERF_IS_NL3ROUTER(obj) ||
277 SN_HWPERF_IS_NL3ROUTER(p)) ? "LLP3" : "LLP4");
285 static void *sn_topology_start(struct seq_file *s, loff_t * pos)
287 struct sn_hwperf_object_info *objs = s->private;
289 if (*pos < sn_hwperf_obj_cnt)
290 return (void *)(objs + *pos);
295 static void *sn_topology_next(struct seq_file *s, void *v, loff_t * pos)
298 return sn_topology_start(s, pos);
301 static void sn_topology_stop(struct seq_file *m, void *v)
307 * /proc/sgi_sn/sn_topology, read-only using seq_file
309 static struct seq_operations sn_topology_seq_ops = {
310 .start = sn_topology_start,
311 .next = sn_topology_next,
312 .stop = sn_topology_stop,
313 .show = sn_topology_show
316 struct sn_hwperf_op_info {
318 struct sn_hwperf_ioctl_args *a;
324 static void sn_hwperf_call_sal(void *info)
326 struct sn_hwperf_op_info *op_info = info;
329 r = ia64_sn_hwperf_op(sn_hwperf_master_nasid, op_info->op,
330 op_info->a->arg, op_info->a->sz,
331 (u64) op_info->p, 0, 0, op_info->v0);
335 static int sn_hwperf_op_cpu(struct sn_hwperf_op_info *op_info)
340 cpumask_t save_allowed;
342 cpu = (op_info->a->arg & SN_HWPERF_ARG_CPU_MASK) >> 32;
343 use_ipi = op_info->a->arg & SN_HWPERF_ARG_USE_IPI_MASK;
344 op_info->a->arg &= SN_HWPERF_ARG_OBJID_MASK;
346 if (cpu != SN_HWPERF_ARG_ANY_CPU) {
347 if (cpu >= num_online_cpus() || !cpu_online(cpu)) {
353 if (cpu == SN_HWPERF_ARG_ANY_CPU || cpu == get_cpu()) {
354 /* don't care, or already on correct cpu */
355 sn_hwperf_call_sal(op_info);
359 /* use an interprocessor interrupt to call SAL */
360 smp_call_function_single(cpu, sn_hwperf_call_sal,
364 /* migrate the task before calling SAL */
365 save_allowed = current->cpus_allowed;
366 set_cpus_allowed(current, cpumask_of_cpu(cpu));
367 sn_hwperf_call_sal(op_info);
368 set_cpus_allowed(current, save_allowed);
378 * ioctl for "sn_hwperf" misc device
381 sn_hwperf_ioctl(struct inode *in, struct file *fp, u32 op, u64 arg)
383 struct sn_hwperf_ioctl_args a;
384 struct cpuinfo_ia64 *cdata;
385 struct sn_hwperf_object_info *objs;
386 struct sn_hwperf_object_info *cpuobj;
387 struct sn_hwperf_op_info op_info;
399 /* only user requests are allowed here */
400 if ((op & SN_HWPERF_OP_MASK) < 10) {
404 r = copy_from_user(&a, (const void __user *)arg,
405 sizeof(struct sn_hwperf_ioctl_args));
412 * Allocate memory to hold a kernel copy of the user buffer. The
413 * buffer contents are either copied in or out (or both) of user
414 * space depending on the flags encoded in the requested operation.
424 if (op & SN_HWPERF_OP_MEM_COPYIN) {
425 r = copy_from_user(p, (const void __user *)a.ptr, a.sz);
433 case SN_HWPERF_GET_CPU_INFO:
434 if (a.sz == sizeof(u64)) {
435 /* special case to get size needed */
436 *(u64 *) p = (u64) num_online_cpus() *
437 sizeof(struct sn_hwperf_object_info);
439 if (a.sz < num_online_cpus() * sizeof(struct sn_hwperf_object_info)) {
443 if ((r = sn_hwperf_enum_objects(&nobj, &objs)) == 0) {
445 for (i = 0; i < nobj; i++) {
446 node = sn_hwperf_obj_to_cnode(objs + i);
447 for_each_online_cpu(j) {
448 if (node != cpu_to_node(j))
450 cpuobj = (struct sn_hwperf_object_info *) p + j;
451 slice = 'a' + cpuid_to_slice(j);
454 snprintf(cpuobj->name,
455 sizeof(cpuobj->name),
457 cdata->proc_freq / 1000000,
459 snprintf(cpuobj->location,
460 sizeof(cpuobj->location),
461 "%s%c", objs[i].location,
470 case SN_HWPERF_GET_NODE_NASID:
471 if (a.sz != sizeof(u64) ||
472 (node = a.arg) < 0 || node >= numionodes) {
476 *(u64 *)p = (u64)cnodeid_to_nasid(node);
479 case SN_HWPERF_GET_OBJ_NODE:
480 if (a.sz != sizeof(u64) || a.arg < 0) {
484 if ((r = sn_hwperf_enum_objects(&nobj, &objs)) == 0) {
490 if (objs[(i = a.arg)].id != a.arg) {
491 for (i = 0; i < nobj; i++) {
492 if (objs[i].id == a.arg)
501 *(u64 *)p = (u64)sn_hwperf_obj_to_cnode(objs + i);
506 case SN_HWPERF_GET_MMRS:
507 case SN_HWPERF_SET_MMRS:
508 case SN_HWPERF_OBJECT_DISTANCE:
513 r = sn_hwperf_op_cpu(&op_info);
517 /* all other ops are a direct SAL call */
518 r = ia64_sn_hwperf_op(sn_hwperf_master_nasid, op,
519 a.arg, a.sz, (u64) p, 0, 0, &v0);
524 if (op & SN_HWPERF_OP_MEM_COPYOUT) {
525 r = copy_to_user((void __user *)a.ptr, p, a.sz);
539 static struct file_operations sn_hwperf_fops = {
540 .ioctl = sn_hwperf_ioctl,
543 static struct miscdevice sn_hwperf_dev = {
549 static int sn_hwperf_init(void)
555 /* single threaded, once-only initialization */
556 down(&sn_hwperf_init_mutex);
557 if (sn_hwperf_salheap) {
558 up(&sn_hwperf_init_mutex);
563 * The PROM code needs a fixed reference node. For convenience the
564 * same node as the console I/O is used.
566 sn_hwperf_master_nasid = (nasid_t) ia64_sn_get_console_nasid();
569 * Request the needed size and install the PROM scratch area.
570 * The PROM keeps various tracking bits in this memory area.
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) {
580 if ((sn_hwperf_salheap = vmalloc(v)) == NULL) {
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) {
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) {
599 sn_hwperf_obj_cnt = (int)v;
602 if (e < 0 && sn_hwperf_salheap) {
603 vfree(sn_hwperf_salheap);
604 sn_hwperf_salheap = NULL;
605 sn_hwperf_obj_cnt = 0;
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.
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);
620 up(&sn_hwperf_init_mutex);
624 int sn_topology_open(struct inode *inode, struct file *file)
627 struct seq_file *seq;
628 struct sn_hwperf_object_info *objbuf;
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;
640 int sn_topology_release(struct inode *inode, struct file *file)
642 struct seq_file *seq = file->private_data;
645 return seq_release(inode, file);