1/*
2 * Hypervisor supplied "gpci" ("get performance counter info") performance
3 * counter support
4 *
5 * Author: Cody P Schafer <cody@linux.vnet.ibm.com>
6 * Copyright 2014 IBM Corporation.
7 *
8 * This program is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU General Public License
10 * as published by the Free Software Foundation; either version
11 * 2 of the License, or (at your option) any later version.
12 */
13
14#define pr_fmt(fmt) "hv-gpci: " fmt
15
16#include <linux/init.h>
17#include <linux/perf_event.h>
18#include <asm/firmware.h>
19#include <asm/hvcall.h>
20#include <asm/io.h>
21
22#include "hv-gpci.h"
23#include "hv-common.h"
24
25/*
26 * Example usage:
27 *  perf stat -e 'hv_gpci/counter_info_version=3,offset=0,length=8,
28 *		  secondary_index=0,starting_index=0xffffffff,request=0x10/' ...
29 */
30
31/* u32 */
32EVENT_DEFINE_RANGE_FORMAT(request, config, 0, 31);
33/* u32 */
34/*
35 * Note that starting_index, phys_processor_idx, sibling_part_id,
36 * hw_chip_id, partition_id all refer to the same bit range. They
37 * are basically aliases for the starting_index. The specific alias
38 * used depends on the event. See REQUEST_IDX_KIND in hv-gpci-requests.h
39 */
40EVENT_DEFINE_RANGE_FORMAT(starting_index, config, 32, 63);
41EVENT_DEFINE_RANGE_FORMAT_LITE(phys_processor_idx, config, 32, 63);
42EVENT_DEFINE_RANGE_FORMAT_LITE(sibling_part_id, config, 32, 63);
43EVENT_DEFINE_RANGE_FORMAT_LITE(hw_chip_id, config, 32, 63);
44EVENT_DEFINE_RANGE_FORMAT_LITE(partition_id, config, 32, 63);
45
46/* u16 */
47EVENT_DEFINE_RANGE_FORMAT(secondary_index, config1, 0, 15);
48/* u8 */
49EVENT_DEFINE_RANGE_FORMAT(counter_info_version, config1, 16, 23);
50/* u8, bytes of data (1-8) */
51EVENT_DEFINE_RANGE_FORMAT(length, config1, 24, 31);
52/* u32, byte offset */
53EVENT_DEFINE_RANGE_FORMAT(offset, config1, 32, 63);
54
55static struct attribute *format_attrs[] = {
56	&format_attr_request.attr,
57	&format_attr_starting_index.attr,
58	&format_attr_phys_processor_idx.attr,
59	&format_attr_sibling_part_id.attr,
60	&format_attr_hw_chip_id.attr,
61	&format_attr_partition_id.attr,
62	&format_attr_secondary_index.attr,
63	&format_attr_counter_info_version.attr,
64
65	&format_attr_offset.attr,
66	&format_attr_length.attr,
67	NULL,
68};
69
70static struct attribute_group format_group = {
71	.name = "format",
72	.attrs = format_attrs,
73};
74
75static struct attribute_group event_group = {
76	.name  = "events",
77	.attrs = hv_gpci_event_attrs,
78};
79
80#define HV_CAPS_ATTR(_name, _format)				\
81static ssize_t _name##_show(struct device *dev,			\
82			    struct device_attribute *attr,	\
83			    char *page)				\
84{								\
85	struct hv_perf_caps caps;				\
86	unsigned long hret = hv_perf_caps_get(&caps);		\
87	if (hret)						\
88		return -EIO;					\
89								\
90	return sprintf(page, _format, caps._name);		\
91}								\
92static struct device_attribute hv_caps_attr_##_name = __ATTR_RO(_name)
93
94static ssize_t kernel_version_show(struct device *dev,
95				   struct device_attribute *attr,
96				   char *page)
97{
98	return sprintf(page, "0x%x\n", COUNTER_INFO_VERSION_CURRENT);
99}
100
101static DEVICE_ATTR_RO(kernel_version);
102HV_CAPS_ATTR(version, "0x%x\n");
103HV_CAPS_ATTR(ga, "%d\n");
104HV_CAPS_ATTR(expanded, "%d\n");
105HV_CAPS_ATTR(lab, "%d\n");
106HV_CAPS_ATTR(collect_privileged, "%d\n");
107
108static struct attribute *interface_attrs[] = {
109	&dev_attr_kernel_version.attr,
110	&hv_caps_attr_version.attr,
111	&hv_caps_attr_ga.attr,
112	&hv_caps_attr_expanded.attr,
113	&hv_caps_attr_lab.attr,
114	&hv_caps_attr_collect_privileged.attr,
115	NULL,
116};
117
118static struct attribute_group interface_group = {
119	.name = "interface",
120	.attrs = interface_attrs,
121};
122
123static const struct attribute_group *attr_groups[] = {
124	&format_group,
125	&event_group,
126	&interface_group,
127	NULL,
128};
129
130#define GPCI_MAX_DATA_BYTES \
131	(1024 - sizeof(struct hv_get_perf_counter_info_params))
132
133static unsigned long single_gpci_request(u32 req, u32 starting_index,
134		u16 secondary_index, u8 version_in, u32 offset, u8 length,
135		u64 *value)
136{
137	unsigned long ret;
138	size_t i;
139	u64 count;
140
141	struct {
142		struct hv_get_perf_counter_info_params params;
143		uint8_t bytes[GPCI_MAX_DATA_BYTES];
144	} __packed __aligned(sizeof(uint64_t)) arg = {
145		.params = {
146			.counter_request = cpu_to_be32(req),
147			.starting_index = cpu_to_be32(starting_index),
148			.secondary_index = cpu_to_be16(secondary_index),
149			.counter_info_version_in = version_in,
150		}
151	};
152
153	ret = plpar_hcall_norets(H_GET_PERF_COUNTER_INFO,
154			virt_to_phys(&arg), sizeof(arg));
155	if (ret) {
156		pr_devel("hcall failed: 0x%lx\n", ret);
157		return ret;
158	}
159
160	/*
161	 * we verify offset and length are within the zeroed buffer at event
162	 * init.
163	 */
164	count = 0;
165	for (i = offset; i < offset + length; i++)
166		count |= arg.bytes[i] << (i - offset);
167
168	*value = count;
169	return ret;
170}
171
172static u64 h_gpci_get_value(struct perf_event *event)
173{
174	u64 count;
175	unsigned long ret = single_gpci_request(event_get_request(event),
176					event_get_starting_index(event),
177					event_get_secondary_index(event),
178					event_get_counter_info_version(event),
179					event_get_offset(event),
180					event_get_length(event),
181					&count);
182	if (ret)
183		return 0;
184	return count;
185}
186
187static void h_gpci_event_update(struct perf_event *event)
188{
189	s64 prev;
190	u64 now = h_gpci_get_value(event);
191	prev = local64_xchg(&event->hw.prev_count, now);
192	local64_add(now - prev, &event->count);
193}
194
195static void h_gpci_event_start(struct perf_event *event, int flags)
196{
197	local64_set(&event->hw.prev_count, h_gpci_get_value(event));
198}
199
200static void h_gpci_event_stop(struct perf_event *event, int flags)
201{
202	h_gpci_event_update(event);
203}
204
205static int h_gpci_event_add(struct perf_event *event, int flags)
206{
207	if (flags & PERF_EF_START)
208		h_gpci_event_start(event, flags);
209
210	return 0;
211}
212
213static int h_gpci_event_init(struct perf_event *event)
214{
215	u64 count;
216	u8 length;
217
218	/* Not our event */
219	if (event->attr.type != event->pmu->type)
220		return -ENOENT;
221
222	/* config2 is unused */
223	if (event->attr.config2) {
224		pr_devel("config2 set when reserved\n");
225		return -EINVAL;
226	}
227
228	/* unsupported modes and filters */
229	if (event->attr.exclude_user   ||
230	    event->attr.exclude_kernel ||
231	    event->attr.exclude_hv     ||
232	    event->attr.exclude_idle   ||
233	    event->attr.exclude_host   ||
234	    event->attr.exclude_guest)
235		return -EINVAL;
236
237	/* no branch sampling */
238	if (has_branch_stack(event))
239		return -EOPNOTSUPP;
240
241	length = event_get_length(event);
242	if (length < 1 || length > 8) {
243		pr_devel("length invalid\n");
244		return -EINVAL;
245	}
246
247	/* last byte within the buffer? */
248	if ((event_get_offset(event) + length) > GPCI_MAX_DATA_BYTES) {
249		pr_devel("request outside of buffer: %zu > %zu\n",
250				(size_t)event_get_offset(event) + length,
251				GPCI_MAX_DATA_BYTES);
252		return -EINVAL;
253	}
254
255	/* check if the request works... */
256	if (single_gpci_request(event_get_request(event),
257				event_get_starting_index(event),
258				event_get_secondary_index(event),
259				event_get_counter_info_version(event),
260				event_get_offset(event),
261				length,
262				&count)) {
263		pr_devel("gpci hcall failed\n");
264		return -EINVAL;
265	}
266
267	return 0;
268}
269
270static struct pmu h_gpci_pmu = {
271	.task_ctx_nr = perf_invalid_context,
272
273	.name = "hv_gpci",
274	.attr_groups = attr_groups,
275	.event_init  = h_gpci_event_init,
276	.add         = h_gpci_event_add,
277	.del         = h_gpci_event_stop,
278	.start       = h_gpci_event_start,
279	.stop        = h_gpci_event_stop,
280	.read        = h_gpci_event_update,
281};
282
283static int hv_gpci_init(void)
284{
285	int r;
286	unsigned long hret;
287	struct hv_perf_caps caps;
288
289	hv_gpci_assert_offsets_correct();
290
291	if (!firmware_has_feature(FW_FEATURE_LPAR)) {
292		pr_debug("not a virtualized system, not enabling\n");
293		return -ENODEV;
294	}
295
296	hret = hv_perf_caps_get(&caps);
297	if (hret) {
298		pr_debug("could not obtain capabilities, not enabling, rc=%ld\n",
299				hret);
300		return -ENODEV;
301	}
302
303	/* sampling not supported */
304	h_gpci_pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
305
306	r = perf_pmu_register(&h_gpci_pmu, h_gpci_pmu.name, -1);
307	if (r)
308		return r;
309
310	return 0;
311}
312
313device_initcall(hv_gpci_init);
314