2 * Hypervisor supplied "24x7" performance counter support
4 * Author: Cody P Schafer <cody@linux.vnet.ibm.com>
5 * Copyright 2014 IBM Corporation.
7 * This program is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU General Public License
9 * as published by the Free Software Foundation; either version
10 * 2 of the License, or (at your option) any later version.
13 #define pr_fmt(fmt) "hv-24x7: " fmt
15 #include <linux/perf_event.h>
16 #include <linux/rbtree.h>
17 #include <linux/module.h>
18 #include <linux/slab.h>
19 #include <linux/vmalloc.h>
21 #include <asm/firmware.h>
22 #include <asm/hvcall.h>
24 #include <linux/byteorder/generic.h>
27 #include "hv-24x7-catalog.h"
28 #include "hv-common.h"
30 static bool domain_is_valid(unsigned domain
)
33 #define DOMAIN(n, v, x, c) \
34 case HV_PERF_DOMAIN_##n: \
36 #include "hv-24x7-domains.h"
44 static bool is_physical_domain(unsigned domain
)
47 #define DOMAIN(n, v, x, c) \
48 case HV_PERF_DOMAIN_##n: \
50 #include "hv-24x7-domains.h"
57 static const char *domain_name(unsigned domain
)
59 if (!domain_is_valid(domain
))
63 case HV_PERF_DOMAIN_PHYS_CHIP
: return "Physical Chip";
64 case HV_PERF_DOMAIN_PHYS_CORE
: return "Physical Core";
65 case HV_PERF_DOMAIN_VCPU_HOME_CORE
: return "VCPU Home Core";
66 case HV_PERF_DOMAIN_VCPU_HOME_CHIP
: return "VCPU Home Chip";
67 case HV_PERF_DOMAIN_VCPU_HOME_NODE
: return "VCPU Home Node";
68 case HV_PERF_DOMAIN_VCPU_REMOTE_NODE
: return "VCPU Remote Node";
75 static bool catalog_entry_domain_is_valid(unsigned domain
)
77 return is_physical_domain(domain
);
81 * TODO: Merging events:
82 * - Think of the hcall as an interface to a 4d array of counters:
84 * - y = indexes in the domain (core, chip, vcpu, node, etc)
85 * - z = offset into the counter space
86 * - w = lpars (guest vms, "logical partitions")
87 * - A single request is: x,y,y_last,z,z_last,w,w_last
88 * - this means we can retrieve a rectangle of counters in y,z for a single x.
90 * - Things to consider (ignoring w):
91 * - input cost_per_request = 16
92 * - output cost_per_result(ys,zs) = 8 + 8 * ys + ys * zs
93 * - limited number of requests per hcall (must fit into 4K bytes)
94 * - 4k = 16 [buffer header] - 16 [request size] * request_count
95 * - 255 requests per hcall
96 * - sometimes it will be more efficient to read extra data and discard
101 * perf stat -e 'hv_24x7/domain=2,offset=8,vcpu=0,lpar=0xffffffff/'
104 /* u3 0-6, one of HV_24X7_PERF_DOMAIN */
105 EVENT_DEFINE_RANGE_FORMAT(domain
, config
, 0, 3);
107 EVENT_DEFINE_RANGE_FORMAT(core
, config
, 16, 31);
108 EVENT_DEFINE_RANGE_FORMAT(chip
, config
, 16, 31);
109 EVENT_DEFINE_RANGE_FORMAT(vcpu
, config
, 16, 31);
110 /* u32, see "data_offset" */
111 EVENT_DEFINE_RANGE_FORMAT(offset
, config
, 32, 63);
113 EVENT_DEFINE_RANGE_FORMAT(lpar
, config1
, 0, 15);
115 EVENT_DEFINE_RANGE(reserved1
, config
, 4, 15);
116 EVENT_DEFINE_RANGE(reserved2
, config1
, 16, 63);
117 EVENT_DEFINE_RANGE(reserved3
, config2
, 0, 63);
119 static struct attribute
*format_attrs
[] = {
120 &format_attr_domain
.attr
,
121 &format_attr_offset
.attr
,
122 &format_attr_core
.attr
,
123 &format_attr_chip
.attr
,
124 &format_attr_vcpu
.attr
,
125 &format_attr_lpar
.attr
,
129 static struct attribute_group format_group
= {
131 .attrs
= format_attrs
,
134 static struct attribute_group event_group
= {
136 /* .attrs is set in init */
139 static struct attribute_group event_desc_group
= {
140 .name
= "event_descs",
141 /* .attrs is set in init */
144 static struct attribute_group event_long_desc_group
= {
145 .name
= "event_long_descs",
146 /* .attrs is set in init */
149 static struct kmem_cache
*hv_page_cache
;
151 DEFINE_PER_CPU(int, hv_24x7_txn_flags
);
152 DEFINE_PER_CPU(int, hv_24x7_txn_err
);
155 struct perf_event
*events
[255];
158 DEFINE_PER_CPU(struct hv_24x7_hw
, hv_24x7_hw
);
161 * request_buffer and result_buffer are not required to be 4k aligned,
162 * but are not allowed to cross any 4k boundary. Aligning them to 4k is
163 * the simplest way to ensure that.
165 #define H24x7_DATA_BUFFER_SIZE 4096
166 DEFINE_PER_CPU(char, hv_24x7_reqb
[H24x7_DATA_BUFFER_SIZE
]) __aligned(4096);
167 DEFINE_PER_CPU(char, hv_24x7_resb
[H24x7_DATA_BUFFER_SIZE
]) __aligned(4096);
169 static char *event_name(struct hv_24x7_event_data
*ev
, int *len
)
171 *len
= be16_to_cpu(ev
->event_name_len
) - 2;
172 return (char *)ev
->remainder
;
175 static char *event_desc(struct hv_24x7_event_data
*ev
, int *len
)
177 unsigned nl
= be16_to_cpu(ev
->event_name_len
);
178 __be16
*desc_len
= (__be16
*)(ev
->remainder
+ nl
- 2);
180 *len
= be16_to_cpu(*desc_len
) - 2;
181 return (char *)ev
->remainder
+ nl
;
184 static char *event_long_desc(struct hv_24x7_event_data
*ev
, int *len
)
186 unsigned nl
= be16_to_cpu(ev
->event_name_len
);
187 __be16
*desc_len_
= (__be16
*)(ev
->remainder
+ nl
- 2);
188 unsigned desc_len
= be16_to_cpu(*desc_len_
);
189 __be16
*long_desc_len
= (__be16
*)(ev
->remainder
+ nl
+ desc_len
- 2);
191 *len
= be16_to_cpu(*long_desc_len
) - 2;
192 return (char *)ev
->remainder
+ nl
+ desc_len
;
195 static bool event_fixed_portion_is_within(struct hv_24x7_event_data
*ev
,
200 return (start
+ offsetof(struct hv_24x7_event_data
, remainder
)) < end
;
204 * Things we don't check:
205 * - padding for desc, name, and long/detailed desc is required to be '\0'
208 * Return NULL if we pass end,
209 * Otherwise return the address of the byte just following the event.
211 static void *event_end(struct hv_24x7_event_data
*ev
, void *end
)
216 unsigned nl
= be16_to_cpu(ev
->event_name_len
);
219 pr_debug("%s: name length too short: %d", __func__
, nl
);
223 if (start
+ nl
> end
) {
224 pr_debug("%s: start=%p + nl=%u > end=%p",
225 __func__
, start
, nl
, end
);
229 dl_
= (__be16
*)(ev
->remainder
+ nl
- 2);
230 if (!IS_ALIGNED((uintptr_t)dl_
, 2))
231 pr_warn("desc len not aligned %p", dl_
);
232 dl
= be16_to_cpu(*dl_
);
234 pr_debug("%s: desc len too short: %d", __func__
, dl
);
238 if (start
+ nl
+ dl
> end
) {
239 pr_debug("%s: (start=%p + nl=%u + dl=%u)=%p > end=%p",
240 __func__
, start
, nl
, dl
, start
+ nl
+ dl
, end
);
244 ldl_
= (__be16
*)(ev
->remainder
+ nl
+ dl
- 2);
245 if (!IS_ALIGNED((uintptr_t)ldl_
, 2))
246 pr_warn("long desc len not aligned %p", ldl_
);
247 ldl
= be16_to_cpu(*ldl_
);
249 pr_debug("%s: long desc len too short (ldl=%u)",
254 if (start
+ nl
+ dl
+ ldl
> end
) {
255 pr_debug("%s: start=%p + nl=%u + dl=%u + ldl=%u > end=%p",
256 __func__
, start
, nl
, dl
, ldl
, end
);
260 return start
+ nl
+ dl
+ ldl
;
263 static unsigned long h_get_24x7_catalog_page_(unsigned long phys_4096
,
264 unsigned long version
,
267 pr_devel("h_get_24x7_catalog_page(0x%lx, %lu, %lu)",
268 phys_4096
, version
, index
);
270 WARN_ON(!IS_ALIGNED(phys_4096
, 4096));
272 return plpar_hcall_norets(H_GET_24X7_CATALOG_PAGE
,
273 phys_4096
, version
, index
);
276 static unsigned long h_get_24x7_catalog_page(char page
[],
277 u64 version
, u32 index
)
279 return h_get_24x7_catalog_page_(virt_to_phys(page
),
284 * Each event we find in the catalog, will have a sysfs entry. Format the
285 * data for this sysfs entry based on the event's domain.
287 * Events belonging to the Chip domain can only be monitored in that domain.
288 * i.e the domain for these events is a fixed/knwon value.
290 * Events belonging to the Core domain can be monitored either in the physical
291 * core or in one of the virtual CPU domains. So the domain value for these
292 * events must be specified by the user (i.e is a required parameter). Format
293 * the Core events with 'domain=?' so the perf-tool can error check required
296 * NOTE: For the Core domain events, rather than making domain a required
297 * parameter we could default it to PHYS_CORE and allowe users to
298 * override the domain to one of the VCPU domains.
300 * However, this can make the interface a little inconsistent.
302 * If we set domain=2 (PHYS_CHIP) and allow user to override this field
303 * the user may be tempted to also modify the "offset=x" field in which
304 * can lead to confusing usage. Consider the HPM_PCYC (offset=0x18) and
305 * HPM_INST (offset=0x20) events. With:
307 * perf stat -e hv_24x7/HPM_PCYC,offset=0x20/
309 * we end up monitoring HPM_INST, while the command line has HPM_PCYC.
311 * By not assigning a default value to the domain for the Core events,
312 * we can have simple guidelines:
314 * - Specifying values for parameters with "=?" is required.
316 * - Specifying (i.e overriding) values for other parameters
319 static char *event_fmt(struct hv_24x7_event_data
*event
, unsigned domain
)
323 const char *domain_str
;
327 case HV_PERF_DOMAIN_PHYS_CHIP
:
328 snprintf(buf
, sizeof(buf
), "%d", domain
);
333 case HV_PERF_DOMAIN_PHYS_CORE
:
344 return kasprintf(GFP_KERNEL
,
345 "domain=%s,offset=0x%x,%s=?,lpar=%s",
347 be16_to_cpu(event
->event_counter_offs
) +
348 be16_to_cpu(event
->event_group_record_offs
),
353 /* Avoid trusting fw to NUL terminate strings */
354 static char *memdup_to_str(char *maybe_str
, int max_len
, gfp_t gfp
)
356 return kasprintf(gfp
, "%.*s", max_len
, maybe_str
);
359 static ssize_t
device_show_string(struct device
*dev
,
360 struct device_attribute
*attr
, char *buf
)
362 struct dev_ext_attribute
*d
;
364 d
= container_of(attr
, struct dev_ext_attribute
, attr
);
366 return sprintf(buf
, "%s\n", (char *)d
->var
);
369 static struct attribute
*device_str_attr_create_(char *name
, char *str
)
371 struct dev_ext_attribute
*attr
= kzalloc(sizeof(*attr
), GFP_KERNEL
);
376 sysfs_attr_init(&attr
->attr
.attr
);
379 attr
->attr
.attr
.name
= name
;
380 attr
->attr
.attr
.mode
= 0444;
381 attr
->attr
.show
= device_show_string
;
383 return &attr
->attr
.attr
;
387 * Allocate and initialize strings representing event attributes.
389 * NOTE: The strings allocated here are never destroyed and continue to
390 * exist till shutdown. This is to allow us to create as many events
391 * from the catalog as possible, even if we encounter errors with some.
392 * In case of changes to error paths in future, these may need to be
393 * freed by the caller.
395 static struct attribute
*device_str_attr_create(char *name
, int name_max
,
397 char *str
, size_t str_max
)
400 char *s
= memdup_to_str(str
, str_max
, GFP_KERNEL
);
407 n
= kasprintf(GFP_KERNEL
, "%.*s", name_max
, name
);
409 n
= kasprintf(GFP_KERNEL
, "%.*s__%d", name_max
, name
,
414 a
= device_str_attr_create_(n
, s
);
426 static struct attribute
*event_to_attr(unsigned ix
,
427 struct hv_24x7_event_data
*event
,
432 char *ev_name
, *a_ev_name
, *val
;
433 struct attribute
*attr
;
435 if (!domain_is_valid(domain
)) {
436 pr_warn("catalog event %u has invalid domain %u\n",
441 val
= event_fmt(event
, domain
);
445 ev_name
= event_name(event
, &event_name_len
);
447 a_ev_name
= kasprintf(GFP_KERNEL
, "%.*s",
448 (int)event_name_len
, ev_name
);
450 a_ev_name
= kasprintf(GFP_KERNEL
, "%.*s__%d",
451 (int)event_name_len
, ev_name
, nonce
);
456 attr
= device_str_attr_create_(a_ev_name
, val
);
468 static struct attribute
*event_to_desc_attr(struct hv_24x7_event_data
*event
,
472 char *name
= event_name(event
, &nl
);
473 char *desc
= event_desc(event
, &dl
);
475 /* If there isn't a description, don't create the sysfs file */
479 return device_str_attr_create(name
, nl
, nonce
, desc
, dl
);
482 static struct attribute
*
483 event_to_long_desc_attr(struct hv_24x7_event_data
*event
, int nonce
)
486 char *name
= event_name(event
, &nl
);
487 char *desc
= event_long_desc(event
, &dl
);
489 /* If there isn't a description, don't create the sysfs file */
493 return device_str_attr_create(name
, nl
, nonce
, desc
, dl
);
496 static int event_data_to_attrs(unsigned ix
, struct attribute
**attrs
,
497 struct hv_24x7_event_data
*event
, int nonce
)
499 *attrs
= event_to_attr(ix
, event
, event
->domain
, nonce
);
515 static int memord(const void *d1
, size_t s1
, const void *d2
, size_t s2
)
522 return memcmp(d1
, d2
, s1
);
525 static int ev_uniq_ord(const void *v1
, size_t s1
, unsigned d1
, const void *v2
,
526 size_t s2
, unsigned d2
)
528 int r
= memord(v1
, s1
, v2
, s2
);
539 static int event_uniq_add(struct rb_root
*root
, const char *name
, int nl
,
542 struct rb_node
**new = &(root
->rb_node
), *parent
= NULL
;
543 struct event_uniq
*data
;
545 /* Figure out where to put new node */
547 struct event_uniq
*it
;
550 it
= container_of(*new, struct event_uniq
, node
);
551 result
= ev_uniq_ord(name
, nl
, domain
, it
->name
, it
->nl
,
556 new = &((*new)->rb_left
);
558 new = &((*new)->rb_right
);
561 pr_info("found a duplicate event %.*s, ct=%u\n", nl
,
567 data
= kmalloc(sizeof(*data
), GFP_KERNEL
);
571 *data
= (struct event_uniq
) {
578 /* Add new node and rebalance tree. */
579 rb_link_node(&data
->node
, parent
, new);
580 rb_insert_color(&data
->node
, root
);
586 static void event_uniq_destroy(struct rb_root
*root
)
589 * the strings we point to are in the giant block of memory filled by
590 * the catalog, and are freed separately.
592 struct event_uniq
*pos
, *n
;
594 rbtree_postorder_for_each_entry_safe(pos
, n
, root
, node
)
600 * ensure the event structure's sizes are self consistent and don't cause us to
601 * read outside of the event
603 * On success, return the event length in bytes.
604 * Otherwise, return -1 (and print as appropriate).
606 static ssize_t
catalog_event_len_validate(struct hv_24x7_event_data
*event
,
608 size_t event_data_bytes
,
609 size_t event_entry_count
,
610 size_t offset
, void *end
)
613 void *ev_end
, *calc_ev_end
;
615 if (offset
>= event_data_bytes
)
618 if (event_idx
>= event_entry_count
) {
619 pr_devel("catalog event data has %zu bytes of padding after last event\n",
620 event_data_bytes
- offset
);
624 if (!event_fixed_portion_is_within(event
, end
)) {
625 pr_warn("event %zu fixed portion is not within range\n",
630 ev_len
= be16_to_cpu(event
->length
);
633 pr_info("event %zu has length %zu not divisible by 16: event=%pK\n",
634 event_idx
, ev_len
, event
);
636 ev_end
= (__u8
*)event
+ ev_len
;
638 pr_warn("event %zu has .length=%zu, ends after buffer end: ev_end=%pK > end=%pK, offset=%zu\n",
639 event_idx
, ev_len
, ev_end
, end
,
644 calc_ev_end
= event_end(event
, end
);
646 pr_warn("event %zu has a calculated length which exceeds buffer length %zu: event=%pK end=%pK, offset=%zu\n",
647 event_idx
, event_data_bytes
, event
, end
,
652 if (calc_ev_end
> ev_end
) {
653 pr_warn("event %zu exceeds it's own length: event=%pK, end=%pK, offset=%zu, calc_ev_end=%pK\n",
654 event_idx
, event
, ev_end
, offset
, calc_ev_end
);
661 #define MAX_4K (SIZE_MAX / 4096)
663 static int create_events_from_catalog(struct attribute
***events_
,
664 struct attribute
***event_descs_
,
665 struct attribute
***event_long_descs_
)
668 size_t catalog_len
, catalog_page_len
, event_entry_count
,
669 event_data_len
, event_data_offs
,
670 event_data_bytes
, junk_events
, event_idx
, event_attr_ct
, i
,
671 attr_max
, event_idx_last
, desc_ct
, long_desc_ct
;
673 uint32_t catalog_version_num
;
674 struct attribute
**events
, **event_descs
, **event_long_descs
;
675 struct hv_24x7_catalog_page_0
*page_0
=
676 kmem_cache_alloc(hv_page_cache
, GFP_KERNEL
);
678 void *event_data
, *end
;
679 struct hv_24x7_event_data
*event
;
680 struct rb_root ev_uniq
= RB_ROOT
;
688 hret
= h_get_24x7_catalog_page(page
, 0, 0);
694 catalog_version_num
= be64_to_cpu(page_0
->version
);
695 catalog_page_len
= be32_to_cpu(page_0
->length
);
697 if (MAX_4K
< catalog_page_len
) {
698 pr_err("invalid page count: %zu\n", catalog_page_len
);
703 catalog_len
= catalog_page_len
* 4096;
705 event_entry_count
= be16_to_cpu(page_0
->event_entry_count
);
706 event_data_offs
= be16_to_cpu(page_0
->event_data_offs
);
707 event_data_len
= be16_to_cpu(page_0
->event_data_len
);
709 pr_devel("cv %zu cl %zu eec %zu edo %zu edl %zu\n",
710 (size_t)catalog_version_num
, catalog_len
,
711 event_entry_count
, event_data_offs
, event_data_len
);
713 if ((MAX_4K
< event_data_len
)
714 || (MAX_4K
< event_data_offs
)
715 || (MAX_4K
- event_data_offs
< event_data_len
)) {
716 pr_err("invalid event data offs %zu and/or len %zu\n",
717 event_data_offs
, event_data_len
);
722 if ((event_data_offs
+ event_data_len
) > catalog_page_len
) {
723 pr_err("event data %zu-%zu does not fit inside catalog 0-%zu\n",
725 event_data_offs
+ event_data_len
,
731 if (SIZE_MAX
- 1 < event_entry_count
) {
732 pr_err("event_entry_count %zu is invalid\n", event_entry_count
);
737 event_data_bytes
= event_data_len
* 4096;
740 * event data can span several pages, events can cross between these
741 * pages. Use vmalloc to make this easier.
743 event_data
= vmalloc(event_data_bytes
);
745 pr_err("could not allocate event data\n");
750 end
= event_data
+ event_data_bytes
;
753 * using vmalloc_to_phys() like this only works if PAGE_SIZE is
756 BUILD_BUG_ON(PAGE_SIZE
% 4096);
758 for (i
= 0; i
< event_data_len
; i
++) {
759 hret
= h_get_24x7_catalog_page_(
760 vmalloc_to_phys(event_data
+ i
* 4096),
762 i
+ event_data_offs
);
764 pr_err("failed to get event data in page %zu\n",
765 i
+ event_data_offs
);
772 * scan the catalog to determine the number of attributes we need, and
773 * verify it at the same time.
775 for (junk_events
= 0, event
= event_data
, event_idx
= 0, attr_max
= 0;
777 event_idx
++, event
= (void *)event
+ ev_len
) {
778 size_t offset
= (void *)event
- (void *)event_data
;
782 ev_len
= catalog_event_len_validate(event
, event_idx
,
789 name
= event_name(event
, &nl
);
791 if (event
->event_group_record_len
== 0) {
792 pr_devel("invalid event %zu (%.*s): group_record_len == 0, skipping\n",
793 event_idx
, nl
, name
);
798 if (!catalog_entry_domain_is_valid(event
->domain
)) {
799 pr_info("event %zu (%.*s) has invalid domain %d\n",
800 event_idx
, nl
, name
, event
->domain
);
808 event_idx_last
= event_idx
;
809 if (event_idx_last
!= event_entry_count
)
810 pr_warn("event buffer ended before listed # of events were parsed (got %zu, wanted %zu, junk %zu)\n",
811 event_idx_last
, event_entry_count
, junk_events
);
813 events
= kmalloc_array(attr_max
+ 1, sizeof(*events
), GFP_KERNEL
);
819 event_descs
= kmalloc_array(event_idx
+ 1, sizeof(*event_descs
),
826 event_long_descs
= kmalloc_array(event_idx
+ 1,
827 sizeof(*event_long_descs
), GFP_KERNEL
);
828 if (!event_long_descs
) {
833 /* Iterate over the catalog filling in the attribute vector */
834 for (junk_events
= 0, event_attr_ct
= 0, desc_ct
= 0, long_desc_ct
= 0,
835 event
= event_data
, event_idx
= 0;
836 event_idx
< event_idx_last
;
837 event_idx
++, ev_len
= be16_to_cpu(event
->length
),
838 event
= (void *)event
+ ev_len
) {
843 * these are the only "bad" events that are intermixed and that
844 * we can ignore without issue. make sure to skip them here
846 if (event
->event_group_record_len
== 0)
848 if (!catalog_entry_domain_is_valid(event
->domain
))
851 name
= event_name(event
, &nl
);
852 nonce
= event_uniq_add(&ev_uniq
, name
, nl
, event
->domain
);
853 ct
= event_data_to_attrs(event_idx
, events
+ event_attr_ct
,
856 pr_warn("event %zu (%.*s) creation failure, skipping\n",
857 event_idx
, nl
, name
);
861 event_descs
[desc_ct
] = event_to_desc_attr(event
, nonce
);
862 if (event_descs
[desc_ct
])
864 event_long_descs
[long_desc_ct
] =
865 event_to_long_desc_attr(event
, nonce
);
866 if (event_long_descs
[long_desc_ct
])
871 pr_info("read %zu catalog entries, created %zu event attrs (%zu failures), %zu descs\n",
872 event_idx
, event_attr_ct
, junk_events
, desc_ct
);
874 events
[event_attr_ct
] = NULL
;
875 event_descs
[desc_ct
] = NULL
;
876 event_long_descs
[long_desc_ct
] = NULL
;
878 event_uniq_destroy(&ev_uniq
);
880 kmem_cache_free(hv_page_cache
, page
);
883 *event_descs_
= event_descs
;
884 *event_long_descs_
= event_long_descs
;
894 kmem_cache_free(hv_page_cache
, page
);
897 *event_descs_
= NULL
;
898 *event_long_descs_
= NULL
;
902 static ssize_t
catalog_read(struct file
*filp
, struct kobject
*kobj
,
903 struct bin_attribute
*bin_attr
, char *buf
,
904 loff_t offset
, size_t count
)
908 size_t catalog_len
= 0, catalog_page_len
= 0;
909 loff_t page_offset
= 0;
910 loff_t offset_in_page
;
912 uint64_t catalog_version_num
= 0;
913 void *page
= kmem_cache_alloc(hv_page_cache
, GFP_USER
);
914 struct hv_24x7_catalog_page_0
*page_0
= page
;
919 hret
= h_get_24x7_catalog_page(page
, 0, 0);
925 catalog_version_num
= be64_to_cpu(page_0
->version
);
926 catalog_page_len
= be32_to_cpu(page_0
->length
);
927 catalog_len
= catalog_page_len
* 4096;
929 page_offset
= offset
/ 4096;
930 offset_in_page
= offset
% 4096;
932 if (page_offset
>= catalog_page_len
)
935 if (page_offset
!= 0) {
936 hret
= h_get_24x7_catalog_page(page
, catalog_version_num
,
944 copy_len
= 4096 - offset_in_page
;
945 if (copy_len
> count
)
948 memcpy(buf
, page
+offset_in_page
, copy_len
);
953 pr_err("h_get_24x7_catalog_page(ver=%lld, page=%lld) failed:"
955 catalog_version_num
, page_offset
, hret
);
956 kmem_cache_free(hv_page_cache
, page
);
958 pr_devel("catalog_read: offset=%lld(%lld) count=%zu "
959 "catalog_len=%zu(%zu) => %zd\n", offset
, page_offset
,
960 count
, catalog_len
, catalog_page_len
, ret
);
965 static ssize_t
domains_show(struct device
*dev
, struct device_attribute
*attr
,
971 for (d
= 0; d
< HV_PERF_DOMAIN_MAX
; d
++) {
972 str
= domain_name(d
);
976 n
= sprintf(page
, "%d: %s\n", d
, str
);
986 #define PAGE_0_ATTR(_name, _fmt, _expr) \
987 static ssize_t _name##_show(struct device *dev, \
988 struct device_attribute *dev_attr, \
991 unsigned long hret; \
993 void *page = kmem_cache_alloc(hv_page_cache, GFP_USER); \
994 struct hv_24x7_catalog_page_0 *page_0 = page; \
997 hret = h_get_24x7_catalog_page(page, 0, 0); \
1002 ret = sprintf(buf, _fmt, _expr); \
1004 kmem_cache_free(hv_page_cache, page); \
1007 static DEVICE_ATTR_RO(_name)
1009 PAGE_0_ATTR(catalog_version
, "%lld\n",
1010 (unsigned long long)be64_to_cpu(page_0
->version
));
1011 PAGE_0_ATTR(catalog_len
, "%lld\n",
1012 (unsigned long long)be32_to_cpu(page_0
->length
) * 4096);
1013 static BIN_ATTR_RO(catalog
, 0/* real length varies */);
1014 static DEVICE_ATTR_RO(domains
);
1016 static struct bin_attribute
*if_bin_attrs
[] = {
1021 static struct attribute
*if_attrs
[] = {
1022 &dev_attr_catalog_len
.attr
,
1023 &dev_attr_catalog_version
.attr
,
1024 &dev_attr_domains
.attr
,
1028 static struct attribute_group if_group
= {
1029 .name
= "interface",
1030 .bin_attrs
= if_bin_attrs
,
1034 static const struct attribute_group
*attr_groups
[] = {
1038 &event_long_desc_group
,
1043 static void log_24x7_hcall(struct hv_24x7_request_buffer
*request_buffer
,
1044 struct hv_24x7_data_result_buffer
*result_buffer
,
1047 struct hv_24x7_request
*req
;
1049 req
= &request_buffer
->requests
[0];
1050 pr_notice_ratelimited("hcall failed: [%d %#x %#x %d] => "
1051 "ret 0x%lx (%ld) detail=0x%x failing ix=%x\n",
1052 req
->performance_domain
, req
->data_offset
,
1053 req
->starting_ix
, req
->starting_lpar_ix
, ret
, ret
,
1054 result_buffer
->detailed_rc
,
1055 result_buffer
->failing_request_ix
);
1059 * Start the process for a new H_GET_24x7_DATA hcall.
1061 static void init_24x7_request(struct hv_24x7_request_buffer
*request_buffer
,
1062 struct hv_24x7_data_result_buffer
*result_buffer
)
1065 memset(request_buffer
, 0, 4096);
1066 memset(result_buffer
, 0, 4096);
1068 request_buffer
->interface_version
= HV_24X7_IF_VERSION_CURRENT
;
1069 /* memset above set request_buffer->num_requests to 0 */
1073 * Commit (i.e perform) the H_GET_24x7_DATA hcall using the data collected
1074 * by 'init_24x7_request()' and 'add_event_to_24x7_request()'.
1076 static int make_24x7_request(struct hv_24x7_request_buffer
*request_buffer
,
1077 struct hv_24x7_data_result_buffer
*result_buffer
)
1082 * NOTE: Due to variable number of array elements in request and
1083 * result buffer(s), sizeof() is not reliable. Use the actual
1084 * allocated buffer size, H24x7_DATA_BUFFER_SIZE.
1086 ret
= plpar_hcall_norets(H_GET_24X7_DATA
,
1087 virt_to_phys(request_buffer
), H24x7_DATA_BUFFER_SIZE
,
1088 virt_to_phys(result_buffer
), H24x7_DATA_BUFFER_SIZE
);
1091 log_24x7_hcall(request_buffer
, result_buffer
, ret
);
1097 * Add the given @event to the next slot in the 24x7 request_buffer.
1099 * Note that H_GET_24X7_DATA hcall allows reading several counters'
1100 * values in a single HCALL. We expect the caller to add events to the
1101 * request buffer one by one, make the HCALL and process the results.
1103 static int add_event_to_24x7_request(struct perf_event
*event
,
1104 struct hv_24x7_request_buffer
*request_buffer
)
1108 struct hv_24x7_request
*req
;
1110 if (request_buffer
->num_requests
> 254) {
1111 pr_devel("Too many requests for 24x7 HCALL %d\n",
1112 request_buffer
->num_requests
);
1116 switch (event_get_domain(event
)) {
1117 case HV_PERF_DOMAIN_PHYS_CHIP
:
1118 idx
= event_get_chip(event
);
1120 case HV_PERF_DOMAIN_PHYS_CORE
:
1121 idx
= event_get_core(event
);
1124 idx
= event_get_vcpu(event
);
1127 i
= request_buffer
->num_requests
++;
1128 req
= &request_buffer
->requests
[i
];
1130 req
->performance_domain
= event_get_domain(event
);
1131 req
->data_size
= cpu_to_be16(8);
1132 req
->data_offset
= cpu_to_be32(event_get_offset(event
));
1133 req
->starting_lpar_ix
= cpu_to_be16(event_get_lpar(event
)),
1134 req
->max_num_lpars
= cpu_to_be16(1);
1135 req
->starting_ix
= cpu_to_be16(idx
);
1136 req
->max_ix
= cpu_to_be16(1);
1141 static unsigned long single_24x7_request(struct perf_event
*event
, u64
*count
)
1144 struct hv_24x7_request_buffer
*request_buffer
;
1145 struct hv_24x7_data_result_buffer
*result_buffer
;
1147 BUILD_BUG_ON(sizeof(*request_buffer
) > 4096);
1148 BUILD_BUG_ON(sizeof(*result_buffer
) > 4096);
1150 request_buffer
= (void *)get_cpu_var(hv_24x7_reqb
);
1151 result_buffer
= (void *)get_cpu_var(hv_24x7_resb
);
1153 init_24x7_request(request_buffer
, result_buffer
);
1155 ret
= add_event_to_24x7_request(event
, request_buffer
);
1159 ret
= make_24x7_request(request_buffer
, result_buffer
);
1161 log_24x7_hcall(request_buffer
, result_buffer
, ret
);
1165 /* process result from hcall */
1166 *count
= be64_to_cpu(result_buffer
->results
[0].elements
[0].element_data
[0]);
1169 put_cpu_var(hv_24x7_reqb
);
1170 put_cpu_var(hv_24x7_resb
);
1175 static int h_24x7_event_init(struct perf_event
*event
)
1177 struct hv_perf_caps caps
;
1183 if (event
->attr
.type
!= event
->pmu
->type
)
1186 /* Unused areas must be 0 */
1187 if (event_get_reserved1(event
) ||
1188 event_get_reserved2(event
) ||
1189 event_get_reserved3(event
)) {
1190 pr_devel("reserved set when forbidden 0x%llx(0x%llx) 0x%llx(0x%llx) 0x%llx(0x%llx)\n",
1192 event_get_reserved1(event
),
1193 event
->attr
.config1
,
1194 event_get_reserved2(event
),
1195 event
->attr
.config2
,
1196 event_get_reserved3(event
));
1200 /* unsupported modes and filters */
1201 if (event
->attr
.exclude_user
||
1202 event
->attr
.exclude_kernel
||
1203 event
->attr
.exclude_hv
||
1204 event
->attr
.exclude_idle
||
1205 event
->attr
.exclude_host
||
1206 event
->attr
.exclude_guest
)
1209 /* no branch sampling */
1210 if (has_branch_stack(event
))
1213 /* offset must be 8 byte aligned */
1214 if (event_get_offset(event
) % 8) {
1215 pr_devel("bad alignment\n");
1219 /* Domains above 6 are invalid */
1220 domain
= event_get_domain(event
);
1222 pr_devel("invalid domain %d\n", domain
);
1226 hret
= hv_perf_caps_get(&caps
);
1228 pr_devel("could not get capabilities: rc=%ld\n", hret
);
1232 /* Physical domains & other lpars require extra capabilities */
1233 if (!caps
.collect_privileged
&& (is_physical_domain(domain
) ||
1234 (event_get_lpar(event
) != event_get_lpar_max()))) {
1235 pr_devel("hv permissions disallow: is_physical_domain:%d, lpar=0x%llx\n",
1236 is_physical_domain(domain
),
1237 event_get_lpar(event
));
1241 /* Get the initial value of the counter for this event */
1242 if (single_24x7_request(event
, &ct
)) {
1243 pr_devel("test hcall failed\n");
1246 (void)local64_xchg(&event
->hw
.prev_count
, ct
);
1251 static u64
h_24x7_get_value(struct perf_event
*event
)
1255 ret
= single_24x7_request(event
, &ct
);
1257 /* We checked this in event init, shouldn't fail here... */
1263 static void update_event_count(struct perf_event
*event
, u64 now
)
1267 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
1268 local64_add(now
- prev
, &event
->count
);
1271 static void h_24x7_event_read(struct perf_event
*event
)
1274 struct hv_24x7_request_buffer
*request_buffer
;
1275 struct hv_24x7_hw
*h24x7hw
;
1278 txn_flags
= __this_cpu_read(hv_24x7_txn_flags
);
1281 * If in a READ transaction, add this counter to the list of
1282 * counters to read during the next HCALL (i.e commit_txn()).
1283 * If not in a READ transaction, go ahead and make the HCALL
1284 * to read this counter by itself.
1287 if (txn_flags
& PERF_PMU_TXN_READ
) {
1291 if (__this_cpu_read(hv_24x7_txn_err
))
1294 request_buffer
= (void *)get_cpu_var(hv_24x7_reqb
);
1296 ret
= add_event_to_24x7_request(event
, request_buffer
);
1298 __this_cpu_write(hv_24x7_txn_err
, ret
);
1301 * Assoicate the event with the HCALL request index,
1302 * so ->commit_txn() can quickly find/update count.
1304 i
= request_buffer
->num_requests
- 1;
1306 h24x7hw
= &get_cpu_var(hv_24x7_hw
);
1307 h24x7hw
->events
[i
] = event
;
1308 put_cpu_var(h24x7hw
);
1310 * Clear the event count so we can compute the _change_
1311 * in the 24x7 raw counter value at the end of the txn.
1313 * Note that we could alternatively read the 24x7 value
1314 * now and save its value in event->hw.prev_count. But
1315 * that would require issuing a hcall, which would then
1316 * defeat the purpose of using the txn interface.
1318 local64_set(&event
->count
, 0);
1321 put_cpu_var(hv_24x7_reqb
);
1323 now
= h_24x7_get_value(event
);
1324 update_event_count(event
, now
);
1328 static void h_24x7_event_start(struct perf_event
*event
, int flags
)
1330 if (flags
& PERF_EF_RELOAD
)
1331 local64_set(&event
->hw
.prev_count
, h_24x7_get_value(event
));
1334 static void h_24x7_event_stop(struct perf_event
*event
, int flags
)
1336 h_24x7_event_read(event
);
1339 static int h_24x7_event_add(struct perf_event
*event
, int flags
)
1341 if (flags
& PERF_EF_START
)
1342 h_24x7_event_start(event
, flags
);
1348 * 24x7 counters only support READ transactions. They are
1349 * always counting and dont need/support ADD transactions.
1350 * Cache the flags, but otherwise ignore transactions that
1351 * are not PERF_PMU_TXN_READ.
1353 static void h_24x7_event_start_txn(struct pmu
*pmu
, unsigned int flags
)
1355 struct hv_24x7_request_buffer
*request_buffer
;
1356 struct hv_24x7_data_result_buffer
*result_buffer
;
1358 /* We should not be called if we are already in a txn */
1359 WARN_ON_ONCE(__this_cpu_read(hv_24x7_txn_flags
));
1361 __this_cpu_write(hv_24x7_txn_flags
, flags
);
1362 if (flags
& ~PERF_PMU_TXN_READ
)
1365 request_buffer
= (void *)get_cpu_var(hv_24x7_reqb
);
1366 result_buffer
= (void *)get_cpu_var(hv_24x7_resb
);
1368 init_24x7_request(request_buffer
, result_buffer
);
1370 put_cpu_var(hv_24x7_resb
);
1371 put_cpu_var(hv_24x7_reqb
);
1375 * Clean up transaction state.
1377 * NOTE: Ignore state of request and result buffers for now.
1378 * We will initialize them during the next read/txn.
1380 static void reset_txn(void)
1382 __this_cpu_write(hv_24x7_txn_flags
, 0);
1383 __this_cpu_write(hv_24x7_txn_err
, 0);
1387 * 24x7 counters only support READ transactions. They are always counting
1388 * and dont need/support ADD transactions. Clear ->txn_flags but otherwise
1389 * ignore transactions that are not of type PERF_PMU_TXN_READ.
1391 * For READ transactions, submit all pending 24x7 requests (i.e requests
1392 * that were queued by h_24x7_event_read()), to the hypervisor and update
1395 static int h_24x7_event_commit_txn(struct pmu
*pmu
)
1397 struct hv_24x7_request_buffer
*request_buffer
;
1398 struct hv_24x7_data_result_buffer
*result_buffer
;
1399 struct hv_24x7_result
*resb
;
1400 struct perf_event
*event
;
1402 int i
, ret
, txn_flags
;
1403 struct hv_24x7_hw
*h24x7hw
;
1405 txn_flags
= __this_cpu_read(hv_24x7_txn_flags
);
1406 WARN_ON_ONCE(!txn_flags
);
1409 if (txn_flags
& ~PERF_PMU_TXN_READ
)
1412 ret
= __this_cpu_read(hv_24x7_txn_err
);
1416 request_buffer
= (void *)get_cpu_var(hv_24x7_reqb
);
1417 result_buffer
= (void *)get_cpu_var(hv_24x7_resb
);
1419 ret
= make_24x7_request(request_buffer
, result_buffer
);
1421 log_24x7_hcall(request_buffer
, result_buffer
, ret
);
1425 h24x7hw
= &get_cpu_var(hv_24x7_hw
);
1427 /* Update event counts from hcall */
1428 for (i
= 0; i
< request_buffer
->num_requests
; i
++) {
1429 resb
= &result_buffer
->results
[i
];
1430 count
= be64_to_cpu(resb
->elements
[0].element_data
[0]);
1431 event
= h24x7hw
->events
[i
];
1432 h24x7hw
->events
[i
] = NULL
;
1433 update_event_count(event
, count
);
1436 put_cpu_var(hv_24x7_hw
);
1439 put_cpu_var(hv_24x7_resb
);
1440 put_cpu_var(hv_24x7_reqb
);
1447 * 24x7 counters only support READ transactions. They are always counting
1448 * and dont need/support ADD transactions. However, regardless of type
1449 * of transaction, all we need to do is cleanup, so we don't have to check
1450 * the type of transaction.
1452 static void h_24x7_event_cancel_txn(struct pmu
*pmu
)
1454 WARN_ON_ONCE(!__this_cpu_read(hv_24x7_txn_flags
));
1458 static struct pmu h_24x7_pmu
= {
1459 .task_ctx_nr
= perf_invalid_context
,
1462 .attr_groups
= attr_groups
,
1463 .event_init
= h_24x7_event_init
,
1464 .add
= h_24x7_event_add
,
1465 .del
= h_24x7_event_stop
,
1466 .start
= h_24x7_event_start
,
1467 .stop
= h_24x7_event_stop
,
1468 .read
= h_24x7_event_read
,
1469 .start_txn
= h_24x7_event_start_txn
,
1470 .commit_txn
= h_24x7_event_commit_txn
,
1471 .cancel_txn
= h_24x7_event_cancel_txn
,
1474 static int hv_24x7_init(void)
1478 struct hv_perf_caps caps
;
1480 if (!firmware_has_feature(FW_FEATURE_LPAR
)) {
1481 pr_debug("not a virtualized system, not enabling\n");
1485 hret
= hv_perf_caps_get(&caps
);
1487 pr_debug("could not obtain capabilities, not enabling, rc=%ld\n",
1492 hv_page_cache
= kmem_cache_create("hv-page-4096", 4096, 4096, 0, NULL
);
1496 /* sampling not supported */
1497 h_24x7_pmu
.capabilities
|= PERF_PMU_CAP_NO_INTERRUPT
;
1499 r
= create_events_from_catalog(&event_group
.attrs
,
1500 &event_desc_group
.attrs
,
1501 &event_long_desc_group
.attrs
);
1506 r
= perf_pmu_register(&h_24x7_pmu
, h_24x7_pmu
.name
, -1);
1513 device_initcall(hv_24x7_init
);