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/module.h>
17 #include <linux/slab.h>
18 #include <asm/firmware.h>
19 #include <asm/hvcall.h>
23 #include "hv-24x7-catalog.h"
24 #include "hv-common.h"
27 * TODO: Merging events:
28 * - Think of the hcall as an interface to a 4d array of counters:
30 * - y = indexes in the domain (core, chip, vcpu, node, etc)
31 * - z = offset into the counter space
32 * - w = lpars (guest vms, "logical partitions")
33 * - A single request is: x,y,y_last,z,z_last,w,w_last
34 * - this means we can retrieve a rectangle of counters in y,z for a single x.
36 * - Things to consider (ignoring w):
37 * - input cost_per_request = 16
38 * - output cost_per_result(ys,zs) = 8 + 8 * ys + ys * zs
39 * - limited number of requests per hcall (must fit into 4K bytes)
40 * - 4k = 16 [buffer header] - 16 [request size] * request_count
41 * - 255 requests per hcall
42 * - sometimes it will be more efficient to read extra data and discard
47 * perf stat -e 'hv_24x7/domain=2,offset=8,starting_index=0,lpar=0xffffffff/'
50 /* u3 0-6, one of HV_24X7_PERF_DOMAIN */
51 EVENT_DEFINE_RANGE_FORMAT(domain
, config
, 0, 3);
53 EVENT_DEFINE_RANGE_FORMAT(starting_index
, config
, 16, 31);
54 /* u32, see "data_offset" */
55 EVENT_DEFINE_RANGE_FORMAT(offset
, config
, 32, 63);
57 EVENT_DEFINE_RANGE_FORMAT(lpar
, config1
, 0, 15);
59 EVENT_DEFINE_RANGE(reserved1
, config
, 4, 15);
60 EVENT_DEFINE_RANGE(reserved2
, config1
, 16, 63);
61 EVENT_DEFINE_RANGE(reserved3
, config2
, 0, 63);
63 static struct attribute
*format_attrs
[] = {
64 &format_attr_domain
.attr
,
65 &format_attr_offset
.attr
,
66 &format_attr_starting_index
.attr
,
67 &format_attr_lpar
.attr
,
71 static struct attribute_group format_group
= {
73 .attrs
= format_attrs
,
76 static struct kmem_cache
*hv_page_cache
;
79 * read_offset_data - copy data from one buffer to another while treating the
80 * source buffer as a small view on the total avaliable
83 * @dest: buffer to copy into
84 * @dest_len: length of @dest in bytes
85 * @requested_offset: the offset within the source data we want. Must be > 0
86 * @src: buffer to copy data from
87 * @src_len: length of @src in bytes
88 * @source_offset: the offset in the sorce data that (src,src_len) refers to.
91 * returns the number of bytes copied.
93 * The following ascii art shows the various buffer possitioning we need to
94 * handle, assigns some arbitrary varibles to points on the buffer, and then
95 * shows how we fiddle with those values to get things we care about (copy
96 * start in src and copy len)
100 * '.' areas in d are written to.
105 * s |----------------------|
110 * s |------------------|
114 * s |------------------|
117 * d |..................|
118 * s |------------------|
123 * s |------------------|
130 * w = requested_offset
131 * z = source_offset + src_len
132 * v = requested_offset + dest_len
134 * w_offset_in_s = w - x = requested_offset - source_offset
135 * z_offset_in_s = z - x = src_len
136 * v_offset_in_s = v - x = request_offset + dest_len - src_len
138 static ssize_t
read_offset_data(void *dest
, size_t dest_len
,
139 loff_t requested_offset
, void *src
,
140 size_t src_len
, loff_t source_offset
)
142 size_t w_offset_in_s
= requested_offset
- source_offset
;
143 size_t z_offset_in_s
= src_len
;
144 size_t v_offset_in_s
= requested_offset
+ dest_len
- src_len
;
145 size_t u_offset_in_s
= min(z_offset_in_s
, v_offset_in_s
);
146 size_t copy_len
= u_offset_in_s
- w_offset_in_s
;
148 if (requested_offset
< 0 || source_offset
< 0)
151 if (z_offset_in_s
<= w_offset_in_s
)
154 memcpy(dest
, src
+ w_offset_in_s
, copy_len
);
158 static unsigned long h_get_24x7_catalog_page_(unsigned long phys_4096
,
159 unsigned long version
,
162 pr_devel("h_get_24x7_catalog_page(0x%lx, %lu, %lu)",
166 WARN_ON(!IS_ALIGNED(phys_4096
, 4096));
167 return plpar_hcall_norets(H_GET_24X7_CATALOG_PAGE
,
173 static unsigned long h_get_24x7_catalog_page(char page
[],
174 u64 version
, u32 index
)
176 return h_get_24x7_catalog_page_(virt_to_phys(page
),
180 static ssize_t
catalog_read(struct file
*filp
, struct kobject
*kobj
,
181 struct bin_attribute
*bin_attr
, char *buf
,
182 loff_t offset
, size_t count
)
186 size_t catalog_len
= 0, catalog_page_len
= 0, page_count
= 0;
187 loff_t page_offset
= 0;
188 uint64_t catalog_version_num
= 0;
189 void *page
= kmem_cache_alloc(hv_page_cache
, GFP_USER
);
190 struct hv_24x7_catalog_page_0
*page_0
= page
;
194 hret
= h_get_24x7_catalog_page(page
, 0, 0);
200 catalog_version_num
= be64_to_cpu(page_0
->version
);
201 catalog_page_len
= be32_to_cpu(page_0
->length
);
202 catalog_len
= catalog_page_len
* 4096;
204 page_offset
= offset
/ 4096;
205 page_count
= count
/ 4096;
207 if (page_offset
>= catalog_page_len
)
210 if (page_offset
!= 0) {
211 hret
= h_get_24x7_catalog_page(page
, catalog_version_num
,
219 ret
= read_offset_data(buf
, count
, offset
,
220 page
, 4096, page_offset
* 4096);
223 pr_err("h_get_24x7_catalog_page(ver=%lld, page=%lld) failed:"
225 catalog_version_num
, page_offset
, hret
);
228 pr_devel("catalog_read: offset=%lld(%lld) count=%zu(%zu) catalog_len=%zu(%zu) => %zd\n",
229 offset
, page_offset
, count
, page_count
, catalog_len
,
230 catalog_page_len
, ret
);
235 #define PAGE_0_ATTR(_name, _fmt, _expr) \
236 static ssize_t _name##_show(struct device *dev, \
237 struct device_attribute *dev_attr, \
240 unsigned long hret; \
242 void *page = kmem_cache_alloc(hv_page_cache, GFP_USER); \
243 struct hv_24x7_catalog_page_0 *page_0 = page; \
246 hret = h_get_24x7_catalog_page(page, 0, 0); \
251 ret = sprintf(buf, _fmt, _expr); \
256 static DEVICE_ATTR_RO(_name)
258 PAGE_0_ATTR(catalog_version
, "%lld\n",
259 (unsigned long long)be64_to_cpu(page_0
->version
));
260 PAGE_0_ATTR(catalog_len
, "%lld\n",
261 (unsigned long long)be32_to_cpu(page_0
->length
) * 4096);
262 static BIN_ATTR_RO(catalog
, 0/* real length varies */);
264 static struct bin_attribute
*if_bin_attrs
[] = {
269 static struct attribute
*if_attrs
[] = {
270 &dev_attr_catalog_len
.attr
,
271 &dev_attr_catalog_version
.attr
,
275 static struct attribute_group if_group
= {
277 .bin_attrs
= if_bin_attrs
,
281 static const struct attribute_group
*attr_groups
[] = {
287 static bool is_physical_domain(int domain
)
289 return domain
== HV_24X7_PERF_DOMAIN_PHYSICAL_CHIP
||
290 domain
== HV_24X7_PERF_DOMAIN_PHYSICAL_CORE
;
293 static unsigned long single_24x7_request(u8 domain
, u32 offset
, u16 ix
,
295 bool success_expected
)
300 * request_buffer and result_buffer are not required to be 4k aligned,
301 * but are not allowed to cross any 4k boundary. Aligning them to 4k is
302 * the simplest way to ensure that.
305 struct hv_24x7_request_buffer buf
;
306 struct hv_24x7_request req
;
307 } __packed
__aligned(4096) request_buffer
= {
309 .interface_version
= HV_24X7_IF_VERSION_CURRENT
,
313 .performance_domain
= domain
,
314 .data_size
= cpu_to_be16(8),
315 .data_offset
= cpu_to_be32(offset
),
316 .starting_lpar_ix
= cpu_to_be16(lpar
),
317 .max_num_lpars
= cpu_to_be16(1),
318 .starting_ix
= cpu_to_be16(ix
),
319 .max_ix
= cpu_to_be16(1),
324 struct hv_24x7_data_result_buffer buf
;
325 struct hv_24x7_result res
;
326 struct hv_24x7_result_element elem
;
328 } __packed
__aligned(4096) result_buffer
= {};
330 ret
= plpar_hcall_norets(H_GET_24X7_DATA
,
331 virt_to_phys(&request_buffer
), sizeof(request_buffer
),
332 virt_to_phys(&result_buffer
), sizeof(result_buffer
));
335 if (success_expected
)
336 pr_err_ratelimited("hcall failed: %d %#x %#x %d => 0x%lx (%ld) detail=0x%x failing ix=%x\n",
337 domain
, offset
, ix
, lpar
,
339 result_buffer
.buf
.detailed_rc
,
340 result_buffer
.buf
.failing_request_ix
);
344 *res
= be64_to_cpu(result_buffer
.result
);
348 static unsigned long event_24x7_request(struct perf_event
*event
, u64
*res
,
349 bool success_expected
)
351 return single_24x7_request(event_get_domain(event
),
352 event_get_offset(event
),
353 event_get_starting_index(event
),
354 event_get_lpar(event
),
359 static int h_24x7_event_init(struct perf_event
*event
)
361 struct hv_perf_caps caps
;
367 if (event
->attr
.type
!= event
->pmu
->type
)
370 /* Unused areas must be 0 */
371 if (event_get_reserved1(event
) ||
372 event_get_reserved2(event
) ||
373 event_get_reserved3(event
)) {
374 pr_devel("reserved set when forbidden 0x%llx(0x%llx) 0x%llx(0x%llx) 0x%llx(0x%llx)\n",
376 event_get_reserved1(event
),
378 event_get_reserved2(event
),
380 event_get_reserved3(event
));
384 /* unsupported modes and filters */
385 if (event
->attr
.exclude_user
||
386 event
->attr
.exclude_kernel
||
387 event
->attr
.exclude_hv
||
388 event
->attr
.exclude_idle
||
389 event
->attr
.exclude_host
||
390 event
->attr
.exclude_guest
||
391 is_sampling_event(event
)) /* no sampling */
394 /* no branch sampling */
395 if (has_branch_stack(event
))
398 /* offset must be 8 byte aligned */
399 if (event_get_offset(event
) % 8) {
400 pr_devel("bad alignment\n");
404 /* Domains above 6 are invalid */
405 domain
= event_get_domain(event
);
407 pr_devel("invalid domain %d\n", domain
);
411 hret
= hv_perf_caps_get(&caps
);
413 pr_devel("could not get capabilities: rc=%ld\n", hret
);
417 /* PHYSICAL domains & other lpars require extra capabilities */
418 if (!caps
.collect_privileged
&& (is_physical_domain(domain
) ||
419 (event_get_lpar(event
) != event_get_lpar_max()))) {
420 pr_devel("hv permisions disallow: is_physical_domain:%d, lpar=0x%llx\n",
421 is_physical_domain(domain
),
422 event_get_lpar(event
));
426 /* see if the event complains */
427 if (event_24x7_request(event
, &ct
, false)) {
428 pr_devel("test hcall failed\n");
435 static u64
h_24x7_get_value(struct perf_event
*event
)
439 ret
= event_24x7_request(event
, &ct
, true);
441 /* We checked this in event init, shouldn't fail here... */
447 static void h_24x7_event_update(struct perf_event
*event
)
451 now
= h_24x7_get_value(event
);
452 prev
= local64_xchg(&event
->hw
.prev_count
, now
);
453 local64_add(now
- prev
, &event
->count
);
456 static void h_24x7_event_start(struct perf_event
*event
, int flags
)
458 if (flags
& PERF_EF_RELOAD
)
459 local64_set(&event
->hw
.prev_count
, h_24x7_get_value(event
));
462 static void h_24x7_event_stop(struct perf_event
*event
, int flags
)
464 h_24x7_event_update(event
);
467 static int h_24x7_event_add(struct perf_event
*event
, int flags
)
469 if (flags
& PERF_EF_START
)
470 h_24x7_event_start(event
, flags
);
475 static int h_24x7_event_idx(struct perf_event
*event
)
480 static struct pmu h_24x7_pmu
= {
481 .task_ctx_nr
= perf_invalid_context
,
484 .attr_groups
= attr_groups
,
485 .event_init
= h_24x7_event_init
,
486 .add
= h_24x7_event_add
,
487 .del
= h_24x7_event_stop
,
488 .start
= h_24x7_event_start
,
489 .stop
= h_24x7_event_stop
,
490 .read
= h_24x7_event_update
,
491 .event_idx
= h_24x7_event_idx
,
494 static int hv_24x7_init(void)
498 struct hv_perf_caps caps
;
500 if (!firmware_has_feature(FW_FEATURE_LPAR
)) {
501 pr_debug("not a virtualized system, not enabling\n");
505 hret
= hv_perf_caps_get(&caps
);
507 pr_debug("could not obtain capabilities, not enabling, rc=%ld\n",
512 hv_page_cache
= kmem_cache_create("hv-page-4096", 4096, 4096, 0, NULL
);
516 r
= perf_pmu_register(&h_24x7_pmu
, h_24x7_pmu
.name
, -1);
523 device_initcall(hv_24x7_init
);