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;
78 static unsigned long h_get_24x7_catalog_page_(unsigned long phys_4096,
79 unsigned long version,
82 pr_devel("h_get_24x7_catalog_page(0x%lx, %lu, %lu)",
86 WARN_ON(!IS_ALIGNED(phys_4096, 4096));
87 return plpar_hcall_norets(H_GET_24X7_CATALOG_PAGE,
93 static unsigned long h_get_24x7_catalog_page(char page[],
94 u64 version, u32 index)
96 return h_get_24x7_catalog_page_(virt_to_phys(page),
100 static ssize_t catalog_read(struct file *filp, struct kobject *kobj,
101 struct bin_attribute *bin_attr, char *buf,
102 loff_t offset, size_t count)
106 size_t catalog_len = 0, catalog_page_len = 0;
107 loff_t page_offset = 0;
108 loff_t offset_in_page;
110 uint64_t catalog_version_num = 0;
111 void *page = kmem_cache_alloc(hv_page_cache, GFP_USER);
112 struct hv_24x7_catalog_page_0 *page_0 = page;
116 hret = h_get_24x7_catalog_page(page, 0, 0);
122 catalog_version_num = be64_to_cpu(page_0->version);
123 catalog_page_len = be32_to_cpu(page_0->length);
124 catalog_len = catalog_page_len * 4096;
126 page_offset = offset / 4096;
127 offset_in_page = offset % 4096;
129 if (page_offset >= catalog_page_len)
132 if (page_offset != 0) {
133 hret = h_get_24x7_catalog_page(page, catalog_version_num,
141 copy_len = 4096 - offset_in_page;
142 if (copy_len > count)
145 memcpy(buf, page+offset_in_page, copy_len);
150 pr_err("h_get_24x7_catalog_page(ver=%lld, page=%lld) failed:"
152 catalog_version_num, page_offset, hret);
153 kmem_cache_free(hv_page_cache, page);
155 pr_devel("catalog_read: offset=%lld(%lld) count=%zu "
156 "catalog_len=%zu(%zu) => %zd\n", offset, page_offset,
157 count, catalog_len, catalog_page_len, ret);
162 #define PAGE_0_ATTR(_name, _fmt, _expr) \
163 static ssize_t _name##_show(struct device *dev, \
164 struct device_attribute *dev_attr, \
167 unsigned long hret; \
169 void *page = kmem_cache_alloc(hv_page_cache, GFP_USER); \
170 struct hv_24x7_catalog_page_0 *page_0 = page; \
173 hret = h_get_24x7_catalog_page(page, 0, 0); \
178 ret = sprintf(buf, _fmt, _expr); \
180 kmem_cache_free(hv_page_cache, page); \
183 static DEVICE_ATTR_RO(_name)
185 PAGE_0_ATTR(catalog_version, "%lld\n",
186 (unsigned long long)be64_to_cpu(page_0->version));
187 PAGE_0_ATTR(catalog_len, "%lld\n",
188 (unsigned long long)be32_to_cpu(page_0->length) * 4096);
189 static BIN_ATTR_RO(catalog, 0/* real length varies */);
191 static struct bin_attribute *if_bin_attrs[] = {
196 static struct attribute *if_attrs[] = {
197 &dev_attr_catalog_len.attr,
198 &dev_attr_catalog_version.attr,
202 static struct attribute_group if_group = {
204 .bin_attrs = if_bin_attrs,
208 static const struct attribute_group *attr_groups[] = {
214 static bool is_physical_domain(int domain)
216 return domain == HV_24X7_PERF_DOMAIN_PHYSICAL_CHIP ||
217 domain == HV_24X7_PERF_DOMAIN_PHYSICAL_CORE;
220 DEFINE_PER_CPU(char, hv_24x7_reqb[4096]) __aligned(4096);
221 DEFINE_PER_CPU(char, hv_24x7_resb[4096]) __aligned(4096);
223 static unsigned long single_24x7_request(u8 domain, u32 offset, u16 ix,
225 bool success_expected)
230 * request_buffer and result_buffer are not required to be 4k aligned,
231 * but are not allowed to cross any 4k boundary. Aligning them to 4k is
232 * the simplest way to ensure that.
235 struct hv_24x7_request_buffer buf;
236 struct hv_24x7_request req;
237 } __packed *request_buffer;
240 struct hv_24x7_data_result_buffer buf;
241 struct hv_24x7_result res;
242 struct hv_24x7_result_element elem;
244 } __packed *result_buffer;
246 BUILD_BUG_ON(sizeof(*request_buffer) > 4096);
247 BUILD_BUG_ON(sizeof(*result_buffer) > 4096);
249 request_buffer = (void *)get_cpu_var(hv_24x7_reqb);
250 result_buffer = (void *)get_cpu_var(hv_24x7_resb);
252 memset(request_buffer, 0, 4096);
253 memset(result_buffer, 0, 4096);
255 *request_buffer = (struct reqb) {
257 .interface_version = HV_24X7_IF_VERSION_CURRENT,
261 .performance_domain = domain,
262 .data_size = cpu_to_be16(8),
263 .data_offset = cpu_to_be32(offset),
264 .starting_lpar_ix = cpu_to_be16(lpar),
265 .max_num_lpars = cpu_to_be16(1),
266 .starting_ix = cpu_to_be16(ix),
267 .max_ix = cpu_to_be16(1),
271 ret = plpar_hcall_norets(H_GET_24X7_DATA,
272 virt_to_phys(request_buffer), sizeof(*request_buffer),
273 virt_to_phys(result_buffer), sizeof(*result_buffer));
276 if (success_expected)
277 pr_err_ratelimited("hcall failed: %d %#x %#x %d => "
278 "0x%lx (%ld) detail=0x%x failing ix=%x\n",
279 domain, offset, ix, lpar, ret, ret,
280 result_buffer->buf.detailed_rc,
281 result_buffer->buf.failing_request_ix);
285 *res = be64_to_cpu(result_buffer->result);
291 static unsigned long event_24x7_request(struct perf_event *event, u64 *res,
292 bool success_expected)
294 return single_24x7_request(event_get_domain(event),
295 event_get_offset(event),
296 event_get_starting_index(event),
297 event_get_lpar(event),
302 static int h_24x7_event_init(struct perf_event *event)
304 struct hv_perf_caps caps;
310 if (event->attr.type != event->pmu->type)
313 /* Unused areas must be 0 */
314 if (event_get_reserved1(event) ||
315 event_get_reserved2(event) ||
316 event_get_reserved3(event)) {
317 pr_devel("reserved set when forbidden 0x%llx(0x%llx) 0x%llx(0x%llx) 0x%llx(0x%llx)\n",
319 event_get_reserved1(event),
321 event_get_reserved2(event),
323 event_get_reserved3(event));
327 /* unsupported modes and filters */
328 if (event->attr.exclude_user ||
329 event->attr.exclude_kernel ||
330 event->attr.exclude_hv ||
331 event->attr.exclude_idle ||
332 event->attr.exclude_host ||
333 event->attr.exclude_guest)
336 /* no branch sampling */
337 if (has_branch_stack(event))
340 /* offset must be 8 byte aligned */
341 if (event_get_offset(event) % 8) {
342 pr_devel("bad alignment\n");
346 /* Domains above 6 are invalid */
347 domain = event_get_domain(event);
349 pr_devel("invalid domain %d\n", domain);
353 hret = hv_perf_caps_get(&caps);
355 pr_devel("could not get capabilities: rc=%ld\n", hret);
359 /* PHYSICAL domains & other lpars require extra capabilities */
360 if (!caps.collect_privileged && (is_physical_domain(domain) ||
361 (event_get_lpar(event) != event_get_lpar_max()))) {
362 pr_devel("hv permisions disallow: is_physical_domain:%d, lpar=0x%llx\n",
363 is_physical_domain(domain),
364 event_get_lpar(event));
368 /* see if the event complains */
369 if (event_24x7_request(event, &ct, false)) {
370 pr_devel("test hcall failed\n");
377 static u64 h_24x7_get_value(struct perf_event *event)
381 ret = event_24x7_request(event, &ct, true);
383 /* We checked this in event init, shouldn't fail here... */
389 static void h_24x7_event_update(struct perf_event *event)
393 now = h_24x7_get_value(event);
394 prev = local64_xchg(&event->hw.prev_count, now);
395 local64_add(now - prev, &event->count);
398 static void h_24x7_event_start(struct perf_event *event, int flags)
400 if (flags & PERF_EF_RELOAD)
401 local64_set(&event->hw.prev_count, h_24x7_get_value(event));
404 static void h_24x7_event_stop(struct perf_event *event, int flags)
406 h_24x7_event_update(event);
409 static int h_24x7_event_add(struct perf_event *event, int flags)
411 if (flags & PERF_EF_START)
412 h_24x7_event_start(event, flags);
417 static struct pmu h_24x7_pmu = {
418 .task_ctx_nr = perf_invalid_context,
421 .attr_groups = attr_groups,
422 .event_init = h_24x7_event_init,
423 .add = h_24x7_event_add,
424 .del = h_24x7_event_stop,
425 .start = h_24x7_event_start,
426 .stop = h_24x7_event_stop,
427 .read = h_24x7_event_update,
430 static int hv_24x7_init(void)
434 struct hv_perf_caps caps;
436 if (!firmware_has_feature(FW_FEATURE_LPAR)) {
437 pr_debug("not a virtualized system, not enabling\n");
441 hret = hv_perf_caps_get(&caps);
443 pr_debug("could not obtain capabilities, not enabling, rc=%ld\n",
448 hv_page_cache = kmem_cache_create("hv-page-4096", 4096, 4096, 0, NULL);
452 /* sampling not supported */
453 h_24x7_pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
455 r = perf_pmu_register(&h_24x7_pmu, h_24x7_pmu.name, -1);
462 device_initcall(hv_24x7_init);