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) 2004-2005 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/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>
38 #include <asm/semaphore.h>
39 #include <asm/segment.h>
40 #include <asm/uaccess.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>
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);
55 static int sn_hwperf_enum_objects(int *nobj, struct sn_hwperf_object_info **ret)
59 struct sn_hwperf_object_info *objbuf = NULL;
61 if ((e = sn_hwperf_init()) < 0) {
62 printk("sn_hwperf_init failed: err %d\n", e);
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);
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) {
81 *nobj = sn_hwperf_obj_cnt;
86 static int sn_hwperf_location_to_bpos(char *location,
87 int *rack, int *bay, int *slot, int *slab)
91 /* first scan for an old style geoid string */
92 if (sscanf(location, "%03d%c%02d#%d",
93 rack, &type, bay, slab) == 4)
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)
103 static int sn_hwperf_geoid_to_cnode(char *location)
107 moduleid_t module_id;
108 int rack, bay, slot, slab;
109 int this_rack, this_bay, this_slot, this_slab;
111 if (sn_hwperf_location_to_bpos(location, &rack, &bay, &slot, &slab))
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 = geo_slot(geoid);
120 this_slab = geo_slab(geoid);
121 if (rack == this_rack && bay == this_bay &&
122 slot == this_slot && slab == this_slab) {
127 return cnode < numionodes ? cnode : -1;
130 static int sn_hwperf_obj_to_cnode(struct sn_hwperf_object_info * obj)
132 if (!obj->sn_hwp_this_part)
134 return sn_hwperf_geoid_to_cnode(obj->location);
137 static int sn_hwperf_generic_ordinal(struct sn_hwperf_object_info *obj,
138 struct sn_hwperf_object_info *objs)
141 struct sn_hwperf_object_info *p;
143 for (ordinal=0, p=objs; p != obj; p++) {
144 if (SN_HWPERF_FOREIGN(p))
146 if (SN_HWPERF_SAME_OBJTYPE(p, obj))
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 */
158 static const char *sn_hwperf_get_slabname(struct sn_hwperf_object_info *obj,
159 struct sn_hwperf_object_info *objs, int *ordinal)
162 const char *slabname = slabname_other;
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);
169 *ordinal = sn_hwperf_generic_ordinal(obj, objs);
170 if (SN_HWPERF_IS_ROUTER(obj))
171 slabname = slabname_router;
177 static void print_pci_topology(struct seq_file *s,
178 struct sn_hwperf_object_info *obj, int *ordinal,
179 u64 rack, u64 bay, u64 slot, u64 slab)
185 if (!(pg = (char *)get_zeroed_page(GFP_KERNEL)))
187 if (ia64_sn_ioif_get_pci_topology(rack, bay, slot, slab,
188 __pa(pg), PAGE_SIZE) == SN_HWPERF_OP_OK) {
189 for (p1=pg; *p1 && p1 < pg + PAGE_SIZE;) {
190 if (!(p2 = strchr(p1, '\n')))
193 seq_printf(s, "pcibus %d %s-%s\n",
194 *ordinal, obj->location, p1);
199 free_page((unsigned long)pg);
202 static int sn_topology_show(struct seq_file *s, void *d)
209 const char *slabname;
213 struct cpuinfo_ia64 *c;
214 struct sn_hwperf_port_info *ptdata;
215 struct sn_hwperf_object_info *p;
216 struct sn_hwperf_object_info *obj = d; /* this object */
217 struct sn_hwperf_object_info *objs = s->private; /* all objects */
218 int rack, bay, slot, slab;
228 int pci_bus_ordinal = 0;
231 seq_printf(s, "# sn_topology version 2\n");
232 seq_printf(s, "# objtype ordinal location partition"
233 " [attribute value [, ...]]\n");
235 if (ia64_sn_get_sn_info(0,
236 &shubtype, &nasid_mask, &nasid_shift, &system_size,
237 &sharing_size, &partid, &coher, ®ion_size))
239 for (nasid_msb=63; nasid_msb > 0; nasid_msb--) {
240 if (((u64)nasid_mask << nasid_shift) & (1ULL << nasid_msb))
243 seq_printf(s, "partition %u %s local "
245 "nasid_mask 0x%016lx, "
249 "coherency_domain %d, "
252 partid, system_utsname.nodename,
253 shubtype ? "shub2" : "shub1",
254 (u64)nasid_mask << nasid_shift, nasid_msb, nasid_shift,
255 system_size, sharing_size, coher, region_size);
258 if (SN_HWPERF_FOREIGN(obj)) {
259 /* private in another partition: not interesting */
263 for (i = 0; i < SN_HWPERF_MAXSTRING && obj->name[i]; i++) {
264 if (obj->name[i] == ' ')
268 slabname = sn_hwperf_get_slabname(obj, objs, &ordinal);
269 seq_printf(s, "%s %d %s %s asic %s", slabname, ordinal, obj->location,
270 obj->sn_hwp_this_part ? "local" : "shared", obj->name);
272 if (!SN_HWPERF_IS_NODE(obj) && !SN_HWPERF_IS_IONODE(obj))
275 seq_printf(s, ", nasid 0x%x", cnodeid_to_nasid(ordinal));
276 for (i=0; i < numionodes; i++) {
277 seq_printf(s, i ? ":%d" : ", dist %d",
278 node_distance(ordinal, i));
283 * CPUs on this node, if any
285 cpumask = node_to_cpumask(ordinal);
286 for_each_online_cpu(i) {
287 if (cpu_isset(i, cpumask)) {
288 slice = 'a' + cpuid_to_slice(i);
290 seq_printf(s, "cpu %d %s%c local"
291 " freq %luMHz, arch ia64",
292 i, obj->location, slice,
293 c->proc_freq / 1000000);
294 for_each_online_cpu(j) {
295 seq_printf(s, j ? ":%d" : ", dist %d",
298 cpuid_to_cnodeid(j)));
305 * PCI busses attached to this node, if any
307 if (sn_hwperf_location_to_bpos(obj->location,
308 &rack, &bay, &slot, &slab)) {
309 /* export pci bus info */
310 print_pci_topology(s, obj, &pci_bus_ordinal,
311 rack, bay, slot, slab);
320 sz = obj->ports * sizeof(struct sn_hwperf_port_info);
321 if ((ptdata = vmalloc(sz)) == NULL)
323 e = ia64_sn_hwperf_op(sn_hwperf_master_nasid,
324 SN_HWPERF_ENUM_PORTS, obj->id, sz,
325 (u64) ptdata, 0, 0, NULL);
326 if (e != SN_HWPERF_OP_OK)
328 for (ordinal=0, p=objs; p != obj; p++) {
329 if (!SN_HWPERF_FOREIGN(p))
332 for (pt = 0; pt < obj->ports; pt++) {
333 for (p = objs, i = 0; i < sn_hwperf_obj_cnt; i++, p++) {
334 if (ptdata[pt].conn_id == p->id) {
338 seq_printf(s, "numalink %d %s-%d",
339 ordinal+pt, obj->location, ptdata[pt].port);
341 if (i >= sn_hwperf_obj_cnt) {
343 seq_puts(s, " local endpoint disconnected"
344 ", protocol unknown\n");
348 if (obj->sn_hwp_this_part && p->sn_hwp_this_part)
349 /* both ends local to this partition */
350 seq_puts(s, " local");
351 else if (!obj->sn_hwp_this_part && !p->sn_hwp_this_part)
352 /* both ends of the link in foreign partiton */
353 seq_puts(s, " foreign");
355 /* link straddles a partition */
356 seq_puts(s, " shared");
359 * Unlikely, but strictly should query the LLP config
360 * registers because an NL4R can be configured to run
361 * NL3 protocol, even when not talking to an NL3 router.
362 * Ditto for node-node.
364 seq_printf(s, " endpoint %s-%d, protocol %s\n",
365 p->location, ptdata[pt].conn_port,
366 (SN_HWPERF_IS_NL3ROUTER(obj) ||
367 SN_HWPERF_IS_NL3ROUTER(p)) ? "LLP3" : "LLP4");
375 static void *sn_topology_start(struct seq_file *s, loff_t * pos)
377 struct sn_hwperf_object_info *objs = s->private;
379 if (*pos < sn_hwperf_obj_cnt)
380 return (void *)(objs + *pos);
385 static void *sn_topology_next(struct seq_file *s, void *v, loff_t * pos)
388 return sn_topology_start(s, pos);
391 static void sn_topology_stop(struct seq_file *m, void *v)
397 * /proc/sgi_sn/sn_topology, read-only using seq_file
399 static struct seq_operations sn_topology_seq_ops = {
400 .start = sn_topology_start,
401 .next = sn_topology_next,
402 .stop = sn_topology_stop,
403 .show = sn_topology_show
406 struct sn_hwperf_op_info {
408 struct sn_hwperf_ioctl_args *a;
414 static void sn_hwperf_call_sal(void *info)
416 struct sn_hwperf_op_info *op_info = info;
419 r = ia64_sn_hwperf_op(sn_hwperf_master_nasid, op_info->op,
420 op_info->a->arg, op_info->a->sz,
421 (u64) op_info->p, 0, 0, op_info->v0);
425 static int sn_hwperf_op_cpu(struct sn_hwperf_op_info *op_info)
430 cpumask_t save_allowed;
432 cpu = (op_info->a->arg & SN_HWPERF_ARG_CPU_MASK) >> 32;
433 use_ipi = op_info->a->arg & SN_HWPERF_ARG_USE_IPI_MASK;
434 op_info->a->arg &= SN_HWPERF_ARG_OBJID_MASK;
436 if (cpu != SN_HWPERF_ARG_ANY_CPU) {
437 if (cpu >= num_online_cpus() || !cpu_online(cpu)) {
443 if (cpu == SN_HWPERF_ARG_ANY_CPU || cpu == get_cpu()) {
444 /* don't care, or already on correct cpu */
445 sn_hwperf_call_sal(op_info);
449 /* use an interprocessor interrupt to call SAL */
450 smp_call_function_single(cpu, sn_hwperf_call_sal,
454 /* migrate the task before calling SAL */
455 save_allowed = current->cpus_allowed;
456 set_cpus_allowed(current, cpumask_of_cpu(cpu));
457 sn_hwperf_call_sal(op_info);
458 set_cpus_allowed(current, save_allowed);
467 /* map SAL hwperf error code to system error code */
468 static int sn_hwperf_map_err(int hwperf_err)
473 case SN_HWPERF_OP_OK:
477 case SN_HWPERF_OP_NOMEM:
481 case SN_HWPERF_OP_NO_PERM:
485 case SN_HWPERF_OP_IO_ERROR:
489 case SN_HWPERF_OP_BUSY:
493 case SN_HWPERF_OP_RECONFIGURE:
497 case SN_HWPERF_OP_INVAL:
507 * ioctl for "sn_hwperf" misc device
510 sn_hwperf_ioctl(struct inode *in, struct file *fp, u32 op, u64 arg)
512 struct sn_hwperf_ioctl_args a;
513 struct cpuinfo_ia64 *cdata;
514 struct sn_hwperf_object_info *objs;
515 struct sn_hwperf_object_info *cpuobj;
516 struct sn_hwperf_op_info op_info;
528 /* only user requests are allowed here */
529 if ((op & SN_HWPERF_OP_MASK) < 10) {
533 r = copy_from_user(&a, (const void __user *)arg,
534 sizeof(struct sn_hwperf_ioctl_args));
541 * Allocate memory to hold a kernel copy of the user buffer. The
542 * buffer contents are either copied in or out (or both) of user
543 * space depending on the flags encoded in the requested operation.
553 if (op & SN_HWPERF_OP_MEM_COPYIN) {
554 r = copy_from_user(p, (const void __user *)a.ptr, a.sz);
562 case SN_HWPERF_GET_CPU_INFO:
563 if (a.sz == sizeof(u64)) {
564 /* special case to get size needed */
565 *(u64 *) p = (u64) num_online_cpus() *
566 sizeof(struct sn_hwperf_object_info);
568 if (a.sz < num_online_cpus() * sizeof(struct sn_hwperf_object_info)) {
572 if ((r = sn_hwperf_enum_objects(&nobj, &objs)) == 0) {
574 for (i = 0; i < nobj; i++) {
575 node = sn_hwperf_obj_to_cnode(objs + i);
576 for_each_online_cpu(j) {
577 if (node != cpu_to_node(j))
579 cpuobj = (struct sn_hwperf_object_info *) p + j;
580 slice = 'a' + cpuid_to_slice(j);
583 snprintf(cpuobj->name,
584 sizeof(cpuobj->name),
586 cdata->proc_freq / 1000000,
588 snprintf(cpuobj->location,
589 sizeof(cpuobj->location),
590 "%s%c", objs[i].location,
599 case SN_HWPERF_GET_NODE_NASID:
600 if (a.sz != sizeof(u64) ||
601 (node = a.arg) < 0 || node >= numionodes) {
605 *(u64 *)p = (u64)cnodeid_to_nasid(node);
608 case SN_HWPERF_GET_OBJ_NODE:
609 if (a.sz != sizeof(u64) || a.arg < 0) {
613 if ((r = sn_hwperf_enum_objects(&nobj, &objs)) == 0) {
619 if (objs[(i = a.arg)].id != a.arg) {
620 for (i = 0; i < nobj; i++) {
621 if (objs[i].id == a.arg)
630 *(u64 *)p = (u64)sn_hwperf_obj_to_cnode(objs + i);
635 case SN_HWPERF_GET_MMRS:
636 case SN_HWPERF_SET_MMRS:
637 case SN_HWPERF_OBJECT_DISTANCE:
642 r = sn_hwperf_op_cpu(&op_info);
644 r = sn_hwperf_map_err(r);
651 /* all other ops are a direct SAL call */
652 r = ia64_sn_hwperf_op(sn_hwperf_master_nasid, op,
653 a.arg, a.sz, (u64) p, 0, 0, &v0);
655 r = sn_hwperf_map_err(r);
662 if (op & SN_HWPERF_OP_MEM_COPYOUT) {
663 r = copy_to_user((void __user *)a.ptr, p, a.sz);
677 static struct file_operations sn_hwperf_fops = {
678 .ioctl = sn_hwperf_ioctl,
681 static struct miscdevice sn_hwperf_dev = {
687 static int sn_hwperf_init(void)
693 /* single threaded, once-only initialization */
694 down(&sn_hwperf_init_mutex);
695 if (sn_hwperf_salheap) {
696 up(&sn_hwperf_init_mutex);
701 * The PROM code needs a fixed reference node. For convenience the
702 * same node as the console I/O is used.
704 sn_hwperf_master_nasid = (nasid_t) ia64_sn_get_console_nasid();
707 * Request the needed size and install the PROM scratch area.
708 * The PROM keeps various tracking bits in this memory area.
710 salr = ia64_sn_hwperf_op(sn_hwperf_master_nasid,
711 (u64) SN_HWPERF_GET_HEAPSIZE, 0,
712 (u64) sizeof(u64), (u64) &v, 0, 0, NULL);
713 if (salr != SN_HWPERF_OP_OK) {
718 if ((sn_hwperf_salheap = vmalloc(v)) == NULL) {
722 salr = ia64_sn_hwperf_op(sn_hwperf_master_nasid,
723 SN_HWPERF_INSTALL_HEAP, 0, v,
724 (u64) sn_hwperf_salheap, 0, 0, NULL);
725 if (salr != SN_HWPERF_OP_OK) {
730 salr = ia64_sn_hwperf_op(sn_hwperf_master_nasid,
731 SN_HWPERF_OBJECT_COUNT, 0,
732 sizeof(u64), (u64) &v, 0, 0, NULL);
733 if (salr != SN_HWPERF_OP_OK) {
737 sn_hwperf_obj_cnt = (int)v;
740 if (e < 0 && sn_hwperf_salheap) {
741 vfree(sn_hwperf_salheap);
742 sn_hwperf_salheap = NULL;
743 sn_hwperf_obj_cnt = 0;
748 * Register a dynamic misc device for ioctl. Platforms
749 * supporting hotplug will create /dev/sn_hwperf, else
750 * user can to look up the minor number in /proc/misc.
752 if ((e = misc_register(&sn_hwperf_dev)) != 0) {
753 printk(KERN_ERR "sn_hwperf_init: misc register "
754 "for \"sn_hwperf\" failed, err %d\n", e);
758 up(&sn_hwperf_init_mutex);
762 int sn_topology_open(struct inode *inode, struct file *file)
765 struct seq_file *seq;
766 struct sn_hwperf_object_info *objbuf;
769 if ((e = sn_hwperf_enum_objects(&nobj, &objbuf)) == 0) {
770 e = seq_open(file, &sn_topology_seq_ops);
771 seq = file->private_data;
772 seq->private = objbuf;
778 int sn_topology_release(struct inode *inode, struct file *file)
780 struct seq_file *seq = file->private_data;
783 return seq_release(inode, file);