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
)
183 for (sz
= PAGE_SIZE
; sz
< 16 * PAGE_SIZE
; sz
+= PAGE_SIZE
) {
184 if (!(p
= (char *)kmalloc(sz
, GFP_KERNEL
)))
186 e
= ia64_sn_ioif_get_pci_topology(__pa(p
), sz
);
190 if (e
== SALRET_OK
|| e
== SALRET_NOT_IMPLEMENTED
)
195 static int sn_topology_show(struct seq_file
*s
, void *d
)
202 const char *slabname
;
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 */
222 seq_printf(s
, "# sn_topology version 2\n");
223 seq_printf(s
, "# objtype ordinal location partition"
224 " [attribute value [, ...]]\n");
226 if (ia64_sn_get_sn_info(0,
227 &shubtype
, &nasid_mask
, &nasid_shift
, &system_size
,
228 &sharing_size
, &partid
, &coher
, ®ion_size
))
230 for (nasid_msb
=63; nasid_msb
> 0; nasid_msb
--) {
231 if (((u64
)nasid_mask
<< nasid_shift
) & (1ULL << nasid_msb
))
234 seq_printf(s
, "partition %u %s local "
236 "nasid_mask 0x%016lx, "
240 "coherency_domain %d, "
243 partid
, system_utsname
.nodename
,
244 shubtype
? "shub2" : "shub1",
245 (u64
)nasid_mask
<< nasid_shift
, nasid_msb
, nasid_shift
,
246 system_size
, sharing_size
, coher
, region_size
);
248 print_pci_topology(s
);
251 if (SN_HWPERF_FOREIGN(obj
)) {
252 /* private in another partition: not interesting */
256 for (i
= 0; i
< SN_HWPERF_MAXSTRING
&& obj
->name
[i
]; i
++) {
257 if (obj
->name
[i
] == ' ')
261 slabname
= sn_hwperf_get_slabname(obj
, objs
, &ordinal
);
262 seq_printf(s
, "%s %d %s %s asic %s", slabname
, ordinal
, obj
->location
,
263 obj
->sn_hwp_this_part
? "local" : "shared", obj
->name
);
265 if (!SN_HWPERF_IS_NODE(obj
) && !SN_HWPERF_IS_IONODE(obj
))
268 seq_printf(s
, ", nasid 0x%x", cnodeid_to_nasid(ordinal
));
269 for (i
=0; i
< numionodes
; i
++) {
270 seq_printf(s
, i
? ":%d" : ", dist %d",
271 node_distance(ordinal
, i
));
276 * CPUs on this node, if any
278 cpumask
= node_to_cpumask(ordinal
);
279 for_each_online_cpu(i
) {
280 if (cpu_isset(i
, cpumask
)) {
281 slice
= 'a' + cpuid_to_slice(i
);
283 seq_printf(s
, "cpu %d %s%c local"
284 " freq %luMHz, arch ia64",
285 i
, obj
->location
, slice
,
286 c
->proc_freq
/ 1000000);
287 for_each_online_cpu(j
) {
288 seq_printf(s
, j
? ":%d" : ", dist %d",
291 cpuid_to_cnodeid(j
)));
302 sz
= obj
->ports
* sizeof(struct sn_hwperf_port_info
);
303 if ((ptdata
= vmalloc(sz
)) == NULL
)
305 e
= ia64_sn_hwperf_op(sn_hwperf_master_nasid
,
306 SN_HWPERF_ENUM_PORTS
, obj
->id
, sz
,
307 (u64
) ptdata
, 0, 0, NULL
);
308 if (e
!= SN_HWPERF_OP_OK
)
310 for (ordinal
=0, p
=objs
; p
!= obj
; p
++) {
311 if (!SN_HWPERF_FOREIGN(p
))
314 for (pt
= 0; pt
< obj
->ports
; pt
++) {
315 for (p
= objs
, i
= 0; i
< sn_hwperf_obj_cnt
; i
++, p
++) {
316 if (ptdata
[pt
].conn_id
== p
->id
) {
320 seq_printf(s
, "numalink %d %s-%d",
321 ordinal
+pt
, obj
->location
, ptdata
[pt
].port
);
323 if (i
>= sn_hwperf_obj_cnt
) {
325 seq_puts(s
, " local endpoint disconnected"
326 ", protocol unknown\n");
330 if (obj
->sn_hwp_this_part
&& p
->sn_hwp_this_part
)
331 /* both ends local to this partition */
332 seq_puts(s
, " local");
333 else if (!obj
->sn_hwp_this_part
&& !p
->sn_hwp_this_part
)
334 /* both ends of the link in foreign partiton */
335 seq_puts(s
, " foreign");
337 /* link straddles a partition */
338 seq_puts(s
, " shared");
341 * Unlikely, but strictly should query the LLP config
342 * registers because an NL4R can be configured to run
343 * NL3 protocol, even when not talking to an NL3 router.
344 * Ditto for node-node.
346 seq_printf(s
, " endpoint %s-%d, protocol %s\n",
347 p
->location
, ptdata
[pt
].conn_port
,
348 (SN_HWPERF_IS_NL3ROUTER(obj
) ||
349 SN_HWPERF_IS_NL3ROUTER(p
)) ? "LLP3" : "LLP4");
357 static void *sn_topology_start(struct seq_file
*s
, loff_t
* pos
)
359 struct sn_hwperf_object_info
*objs
= s
->private;
361 if (*pos
< sn_hwperf_obj_cnt
)
362 return (void *)(objs
+ *pos
);
367 static void *sn_topology_next(struct seq_file
*s
, void *v
, loff_t
* pos
)
370 return sn_topology_start(s
, pos
);
373 static void sn_topology_stop(struct seq_file
*m
, void *v
)
379 * /proc/sgi_sn/sn_topology, read-only using seq_file
381 static struct seq_operations sn_topology_seq_ops
= {
382 .start
= sn_topology_start
,
383 .next
= sn_topology_next
,
384 .stop
= sn_topology_stop
,
385 .show
= sn_topology_show
388 struct sn_hwperf_op_info
{
390 struct sn_hwperf_ioctl_args
*a
;
396 static void sn_hwperf_call_sal(void *info
)
398 struct sn_hwperf_op_info
*op_info
= info
;
401 r
= ia64_sn_hwperf_op(sn_hwperf_master_nasid
, op_info
->op
,
402 op_info
->a
->arg
, op_info
->a
->sz
,
403 (u64
) op_info
->p
, 0, 0, op_info
->v0
);
407 static int sn_hwperf_op_cpu(struct sn_hwperf_op_info
*op_info
)
412 cpumask_t save_allowed
;
414 cpu
= (op_info
->a
->arg
& SN_HWPERF_ARG_CPU_MASK
) >> 32;
415 use_ipi
= op_info
->a
->arg
& SN_HWPERF_ARG_USE_IPI_MASK
;
416 op_info
->a
->arg
&= SN_HWPERF_ARG_OBJID_MASK
;
418 if (cpu
!= SN_HWPERF_ARG_ANY_CPU
) {
419 if (cpu
>= num_online_cpus() || !cpu_online(cpu
)) {
425 if (cpu
== SN_HWPERF_ARG_ANY_CPU
|| cpu
== get_cpu()) {
426 /* don't care, or already on correct cpu */
427 sn_hwperf_call_sal(op_info
);
431 /* use an interprocessor interrupt to call SAL */
432 smp_call_function_single(cpu
, sn_hwperf_call_sal
,
436 /* migrate the task before calling SAL */
437 save_allowed
= current
->cpus_allowed
;
438 set_cpus_allowed(current
, cpumask_of_cpu(cpu
));
439 sn_hwperf_call_sal(op_info
);
440 set_cpus_allowed(current
, save_allowed
);
449 /* map SAL hwperf error code to system error code */
450 static int sn_hwperf_map_err(int hwperf_err
)
455 case SN_HWPERF_OP_OK
:
459 case SN_HWPERF_OP_NOMEM
:
463 case SN_HWPERF_OP_NO_PERM
:
467 case SN_HWPERF_OP_IO_ERROR
:
471 case SN_HWPERF_OP_BUSY
:
475 case SN_HWPERF_OP_RECONFIGURE
:
479 case SN_HWPERF_OP_INVAL
:
489 * ioctl for "sn_hwperf" misc device
492 sn_hwperf_ioctl(struct inode
*in
, struct file
*fp
, u32 op
, u64 arg
)
494 struct sn_hwperf_ioctl_args a
;
495 struct cpuinfo_ia64
*cdata
;
496 struct sn_hwperf_object_info
*objs
;
497 struct sn_hwperf_object_info
*cpuobj
;
498 struct sn_hwperf_op_info op_info
;
510 /* only user requests are allowed here */
511 if ((op
& SN_HWPERF_OP_MASK
) < 10) {
515 r
= copy_from_user(&a
, (const void __user
*)arg
,
516 sizeof(struct sn_hwperf_ioctl_args
));
523 * Allocate memory to hold a kernel copy of the user buffer. The
524 * buffer contents are either copied in or out (or both) of user
525 * space depending on the flags encoded in the requested operation.
535 if (op
& SN_HWPERF_OP_MEM_COPYIN
) {
536 r
= copy_from_user(p
, (const void __user
*)a
.ptr
, a
.sz
);
544 case SN_HWPERF_GET_CPU_INFO
:
545 if (a
.sz
== sizeof(u64
)) {
546 /* special case to get size needed */
547 *(u64
*) p
= (u64
) num_online_cpus() *
548 sizeof(struct sn_hwperf_object_info
);
550 if (a
.sz
< num_online_cpus() * sizeof(struct sn_hwperf_object_info
)) {
554 if ((r
= sn_hwperf_enum_objects(&nobj
, &objs
)) == 0) {
556 for (i
= 0; i
< nobj
; i
++) {
557 node
= sn_hwperf_obj_to_cnode(objs
+ i
);
558 for_each_online_cpu(j
) {
559 if (node
!= cpu_to_node(j
))
561 cpuobj
= (struct sn_hwperf_object_info
*) p
+ j
;
562 slice
= 'a' + cpuid_to_slice(j
);
565 snprintf(cpuobj
->name
,
566 sizeof(cpuobj
->name
),
568 cdata
->proc_freq
/ 1000000,
570 snprintf(cpuobj
->location
,
571 sizeof(cpuobj
->location
),
572 "%s%c", objs
[i
].location
,
581 case SN_HWPERF_GET_NODE_NASID
:
582 if (a
.sz
!= sizeof(u64
) ||
583 (node
= a
.arg
) < 0 || node
>= numionodes
) {
587 *(u64
*)p
= (u64
)cnodeid_to_nasid(node
);
590 case SN_HWPERF_GET_OBJ_NODE
:
591 if (a
.sz
!= sizeof(u64
) || a
.arg
< 0) {
595 if ((r
= sn_hwperf_enum_objects(&nobj
, &objs
)) == 0) {
601 if (objs
[(i
= a
.arg
)].id
!= a
.arg
) {
602 for (i
= 0; i
< nobj
; i
++) {
603 if (objs
[i
].id
== a
.arg
)
612 *(u64
*)p
= (u64
)sn_hwperf_obj_to_cnode(objs
+ i
);
617 case SN_HWPERF_GET_MMRS
:
618 case SN_HWPERF_SET_MMRS
:
619 case SN_HWPERF_OBJECT_DISTANCE
:
624 r
= sn_hwperf_op_cpu(&op_info
);
626 r
= sn_hwperf_map_err(r
);
633 /* all other ops are a direct SAL call */
634 r
= ia64_sn_hwperf_op(sn_hwperf_master_nasid
, op
,
635 a
.arg
, a
.sz
, (u64
) p
, 0, 0, &v0
);
637 r
= sn_hwperf_map_err(r
);
644 if (op
& SN_HWPERF_OP_MEM_COPYOUT
) {
645 r
= copy_to_user((void __user
*)a
.ptr
, p
, a
.sz
);
659 static struct file_operations sn_hwperf_fops
= {
660 .ioctl
= sn_hwperf_ioctl
,
663 static struct miscdevice sn_hwperf_dev
= {
669 static int sn_hwperf_init(void)
675 /* single threaded, once-only initialization */
676 down(&sn_hwperf_init_mutex
);
677 if (sn_hwperf_salheap
) {
678 up(&sn_hwperf_init_mutex
);
683 * The PROM code needs a fixed reference node. For convenience the
684 * same node as the console I/O is used.
686 sn_hwperf_master_nasid
= (nasid_t
) ia64_sn_get_console_nasid();
689 * Request the needed size and install the PROM scratch area.
690 * The PROM keeps various tracking bits in this memory area.
692 salr
= ia64_sn_hwperf_op(sn_hwperf_master_nasid
,
693 (u64
) SN_HWPERF_GET_HEAPSIZE
, 0,
694 (u64
) sizeof(u64
), (u64
) &v
, 0, 0, NULL
);
695 if (salr
!= SN_HWPERF_OP_OK
) {
700 if ((sn_hwperf_salheap
= vmalloc(v
)) == NULL
) {
704 salr
= ia64_sn_hwperf_op(sn_hwperf_master_nasid
,
705 SN_HWPERF_INSTALL_HEAP
, 0, v
,
706 (u64
) sn_hwperf_salheap
, 0, 0, NULL
);
707 if (salr
!= SN_HWPERF_OP_OK
) {
712 salr
= ia64_sn_hwperf_op(sn_hwperf_master_nasid
,
713 SN_HWPERF_OBJECT_COUNT
, 0,
714 sizeof(u64
), (u64
) &v
, 0, 0, NULL
);
715 if (salr
!= SN_HWPERF_OP_OK
) {
719 sn_hwperf_obj_cnt
= (int)v
;
722 if (e
< 0 && sn_hwperf_salheap
) {
723 vfree(sn_hwperf_salheap
);
724 sn_hwperf_salheap
= NULL
;
725 sn_hwperf_obj_cnt
= 0;
730 * Register a dynamic misc device for ioctl. Platforms
731 * supporting hotplug will create /dev/sn_hwperf, else
732 * user can to look up the minor number in /proc/misc.
734 if ((e
= misc_register(&sn_hwperf_dev
)) != 0) {
735 printk(KERN_ERR
"sn_hwperf_init: misc register "
736 "for \"sn_hwperf\" failed, err %d\n", e
);
740 up(&sn_hwperf_init_mutex
);
744 int sn_topology_open(struct inode
*inode
, struct file
*file
)
747 struct seq_file
*seq
;
748 struct sn_hwperf_object_info
*objbuf
;
751 if ((e
= sn_hwperf_enum_objects(&nobj
, &objbuf
)) == 0) {
752 e
= seq_open(file
, &sn_topology_seq_ops
);
753 seq
= file
->private_data
;
754 seq
->private = objbuf
;
760 int sn_topology_release(struct inode
*inode
, struct file
*file
)
762 struct seq_file
*seq
= file
->private_data
;
765 return seq_release(inode
, file
);