xref: /linux/arch/powerpc/perf/hv-24x7.c (revision 5e4e38446a62a4f50d77b0dd11d4b379dee08988)
1 /*
2  * Hypervisor supplied "24x7" performance counter support
3  *
4  * Author: Cody P Schafer <cody@linux.vnet.ibm.com>
5  * Copyright 2014 IBM Corporation.
6  *
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.
11  */
12 
13 #define pr_fmt(fmt) "hv-24x7: " fmt
14 
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>
20 
21 #include <asm/firmware.h>
22 #include <asm/hvcall.h>
23 #include <asm/io.h>
24 #include <linux/byteorder/generic.h>
25 
26 #include "hv-24x7.h"
27 #include "hv-24x7-catalog.h"
28 #include "hv-common.h"
29 
30 static const char *event_domain_suffix(unsigned domain)
31 {
32 	switch (domain) {
33 #define DOMAIN(n, v, x, c)		\
34 	case HV_PERF_DOMAIN_##n:	\
35 		return "__" #n;
36 #include "hv-24x7-domains.h"
37 #undef DOMAIN
38 	default:
39 		WARN(1, "unknown domain %d\n", domain);
40 		return "__UNKNOWN_DOMAIN_SUFFIX";
41 	}
42 }
43 
44 static bool domain_is_valid(unsigned domain)
45 {
46 	switch (domain) {
47 #define DOMAIN(n, v, x, c)		\
48 	case HV_PERF_DOMAIN_##n:	\
49 		/* fall through */
50 #include "hv-24x7-domains.h"
51 #undef DOMAIN
52 		return true;
53 	default:
54 		return false;
55 	}
56 }
57 
58 static bool is_physical_domain(unsigned domain)
59 {
60 	switch (domain) {
61 #define DOMAIN(n, v, x, c)		\
62 	case HV_PERF_DOMAIN_##n:	\
63 		return c;
64 #include "hv-24x7-domains.h"
65 #undef DOMAIN
66 	default:
67 		return false;
68 	}
69 }
70 
71 static bool catalog_entry_domain_is_valid(unsigned domain)
72 {
73 	return is_physical_domain(domain);
74 }
75 
76 /*
77  * TODO: Merging events:
78  * - Think of the hcall as an interface to a 4d array of counters:
79  *   - x = domains
80  *   - y = indexes in the domain (core, chip, vcpu, node, etc)
81  *   - z = offset into the counter space
82  *   - w = lpars (guest vms, "logical partitions")
83  * - A single request is: x,y,y_last,z,z_last,w,w_last
84  *   - this means we can retrieve a rectangle of counters in y,z for a single x.
85  *
86  * - Things to consider (ignoring w):
87  *   - input  cost_per_request = 16
88  *   - output cost_per_result(ys,zs)  = 8 + 8 * ys + ys * zs
89  *   - limited number of requests per hcall (must fit into 4K bytes)
90  *     - 4k = 16 [buffer header] - 16 [request size] * request_count
91  *     - 255 requests per hcall
92  *   - sometimes it will be more efficient to read extra data and discard
93  */
94 
95 /*
96  * Example usage:
97  *  perf stat -e 'hv_24x7/domain=2,offset=8,vcpu=0,lpar=0xffffffff/'
98  */
99 
100 /* u3 0-6, one of HV_24X7_PERF_DOMAIN */
101 EVENT_DEFINE_RANGE_FORMAT(domain, config, 0, 3);
102 /* u16 */
103 EVENT_DEFINE_RANGE_FORMAT(core, config, 16, 31);
104 EVENT_DEFINE_RANGE_FORMAT(vcpu, config, 16, 31);
105 /* u32, see "data_offset" */
106 EVENT_DEFINE_RANGE_FORMAT(offset, config, 32, 63);
107 /* u16 */
108 EVENT_DEFINE_RANGE_FORMAT(lpar, config1, 0, 15);
109 
110 EVENT_DEFINE_RANGE(reserved1, config,   4, 15);
111 EVENT_DEFINE_RANGE(reserved2, config1, 16, 63);
112 EVENT_DEFINE_RANGE(reserved3, config2,  0, 63);
113 
114 static struct attribute *format_attrs[] = {
115 	&format_attr_domain.attr,
116 	&format_attr_offset.attr,
117 	&format_attr_core.attr,
118 	&format_attr_vcpu.attr,
119 	&format_attr_lpar.attr,
120 	NULL,
121 };
122 
123 static struct attribute_group format_group = {
124 	.name = "format",
125 	.attrs = format_attrs,
126 };
127 
128 static struct attribute_group event_group = {
129 	.name = "events",
130 	/* .attrs is set in init */
131 };
132 
133 static struct attribute_group event_desc_group = {
134 	.name = "event_descs",
135 	/* .attrs is set in init */
136 };
137 
138 static struct attribute_group event_long_desc_group = {
139 	.name = "event_long_descs",
140 	/* .attrs is set in init */
141 };
142 
143 static struct kmem_cache *hv_page_cache;
144 
145 DEFINE_PER_CPU(int, hv_24x7_txn_flags);
146 DEFINE_PER_CPU(int, hv_24x7_txn_err);
147 
148 struct hv_24x7_hw {
149 	struct perf_event *events[255];
150 };
151 
152 DEFINE_PER_CPU(struct hv_24x7_hw, hv_24x7_hw);
153 
154 /*
155  * request_buffer and result_buffer are not required to be 4k aligned,
156  * but are not allowed to cross any 4k boundary. Aligning them to 4k is
157  * the simplest way to ensure that.
158  */
159 #define H24x7_DATA_BUFFER_SIZE	4096
160 DEFINE_PER_CPU(char, hv_24x7_reqb[H24x7_DATA_BUFFER_SIZE]) __aligned(4096);
161 DEFINE_PER_CPU(char, hv_24x7_resb[H24x7_DATA_BUFFER_SIZE]) __aligned(4096);
162 
163 static char *event_name(struct hv_24x7_event_data *ev, int *len)
164 {
165 	*len = be16_to_cpu(ev->event_name_len) - 2;
166 	return (char *)ev->remainder;
167 }
168 
169 static char *event_desc(struct hv_24x7_event_data *ev, int *len)
170 {
171 	unsigned nl = be16_to_cpu(ev->event_name_len);
172 	__be16 *desc_len = (__be16 *)(ev->remainder + nl - 2);
173 
174 	*len = be16_to_cpu(*desc_len) - 2;
175 	return (char *)ev->remainder + nl;
176 }
177 
178 static char *event_long_desc(struct hv_24x7_event_data *ev, int *len)
179 {
180 	unsigned nl = be16_to_cpu(ev->event_name_len);
181 	__be16 *desc_len_ = (__be16 *)(ev->remainder + nl - 2);
182 	unsigned desc_len = be16_to_cpu(*desc_len_);
183 	__be16 *long_desc_len = (__be16 *)(ev->remainder + nl + desc_len - 2);
184 
185 	*len = be16_to_cpu(*long_desc_len) - 2;
186 	return (char *)ev->remainder + nl + desc_len;
187 }
188 
189 static bool event_fixed_portion_is_within(struct hv_24x7_event_data *ev,
190 					  void *end)
191 {
192 	void *start = ev;
193 
194 	return (start + offsetof(struct hv_24x7_event_data, remainder)) < end;
195 }
196 
197 /*
198  * Things we don't check:
199  *  - padding for desc, name, and long/detailed desc is required to be '\0'
200  *    bytes.
201  *
202  *  Return NULL if we pass end,
203  *  Otherwise return the address of the byte just following the event.
204  */
205 static void *event_end(struct hv_24x7_event_data *ev, void *end)
206 {
207 	void *start = ev;
208 	__be16 *dl_, *ldl_;
209 	unsigned dl, ldl;
210 	unsigned nl = be16_to_cpu(ev->event_name_len);
211 
212 	if (nl < 2) {
213 		pr_debug("%s: name length too short: %d", __func__, nl);
214 		return NULL;
215 	}
216 
217 	if (start + nl > end) {
218 		pr_debug("%s: start=%p + nl=%u > end=%p",
219 				__func__, start, nl, end);
220 		return NULL;
221 	}
222 
223 	dl_ = (__be16 *)(ev->remainder + nl - 2);
224 	if (!IS_ALIGNED((uintptr_t)dl_, 2))
225 		pr_warn("desc len not aligned %p", dl_);
226 	dl = be16_to_cpu(*dl_);
227 	if (dl < 2) {
228 		pr_debug("%s: desc len too short: %d", __func__, dl);
229 		return NULL;
230 	}
231 
232 	if (start + nl + dl > end) {
233 		pr_debug("%s: (start=%p + nl=%u + dl=%u)=%p > end=%p",
234 				__func__, start, nl, dl, start + nl + dl, end);
235 		return NULL;
236 	}
237 
238 	ldl_ = (__be16 *)(ev->remainder + nl + dl - 2);
239 	if (!IS_ALIGNED((uintptr_t)ldl_, 2))
240 		pr_warn("long desc len not aligned %p", ldl_);
241 	ldl = be16_to_cpu(*ldl_);
242 	if (ldl < 2) {
243 		pr_debug("%s: long desc len too short (ldl=%u)",
244 				__func__, ldl);
245 		return NULL;
246 	}
247 
248 	if (start + nl + dl + ldl > end) {
249 		pr_debug("%s: start=%p + nl=%u + dl=%u + ldl=%u > end=%p",
250 				__func__, start, nl, dl, ldl, end);
251 		return NULL;
252 	}
253 
254 	return start + nl + dl + ldl;
255 }
256 
257 static unsigned long h_get_24x7_catalog_page_(unsigned long phys_4096,
258 					      unsigned long version,
259 					      unsigned long index)
260 {
261 	pr_devel("h_get_24x7_catalog_page(0x%lx, %lu, %lu)",
262 			phys_4096, version, index);
263 
264 	WARN_ON(!IS_ALIGNED(phys_4096, 4096));
265 
266 	return plpar_hcall_norets(H_GET_24X7_CATALOG_PAGE,
267 			phys_4096, version, index);
268 }
269 
270 static unsigned long h_get_24x7_catalog_page(char page[],
271 					     u64 version, u32 index)
272 {
273 	return h_get_24x7_catalog_page_(virt_to_phys(page),
274 					version, index);
275 }
276 
277 static unsigned core_domains[] = {
278 	HV_PERF_DOMAIN_PHYS_CORE,
279 	HV_PERF_DOMAIN_VCPU_HOME_CORE,
280 	HV_PERF_DOMAIN_VCPU_HOME_CHIP,
281 	HV_PERF_DOMAIN_VCPU_HOME_NODE,
282 	HV_PERF_DOMAIN_VCPU_REMOTE_NODE,
283 };
284 /* chip event data always yeilds a single event, core yeilds multiple */
285 #define MAX_EVENTS_PER_EVENT_DATA ARRAY_SIZE(core_domains)
286 
287 static char *event_fmt(struct hv_24x7_event_data *event, unsigned domain)
288 {
289 	const char *sindex;
290 	const char *lpar;
291 
292 	if (is_physical_domain(domain)) {
293 		lpar = "0x0";
294 		sindex = "core";
295 	} else {
296 		lpar = "?";
297 		sindex = "vcpu";
298 	}
299 
300 	return kasprintf(GFP_KERNEL,
301 			"domain=0x%x,offset=0x%x,%s=?,lpar=%s",
302 			domain,
303 			be16_to_cpu(event->event_counter_offs) +
304 				be16_to_cpu(event->event_group_record_offs),
305 			sindex,
306 			lpar);
307 }
308 
309 /* Avoid trusting fw to NUL terminate strings */
310 static char *memdup_to_str(char *maybe_str, int max_len, gfp_t gfp)
311 {
312 	return kasprintf(gfp, "%.*s", max_len, maybe_str);
313 }
314 
315 static ssize_t device_show_string(struct device *dev,
316 		struct device_attribute *attr, char *buf)
317 {
318 	struct dev_ext_attribute *d;
319 
320 	d = container_of(attr, struct dev_ext_attribute, attr);
321 
322 	return sprintf(buf, "%s\n", (char *)d->var);
323 }
324 
325 static struct attribute *device_str_attr_create_(char *name, char *str)
326 {
327 	struct dev_ext_attribute *attr = kzalloc(sizeof(*attr), GFP_KERNEL);
328 
329 	if (!attr)
330 		return NULL;
331 
332 	sysfs_attr_init(&attr->attr.attr);
333 
334 	attr->var = str;
335 	attr->attr.attr.name = name;
336 	attr->attr.attr.mode = 0444;
337 	attr->attr.show = device_show_string;
338 
339 	return &attr->attr.attr;
340 }
341 
342 static struct attribute *device_str_attr_create(char *name, int name_max,
343 						int name_nonce,
344 						char *str, size_t str_max)
345 {
346 	char *n;
347 	char *s = memdup_to_str(str, str_max, GFP_KERNEL);
348 	struct attribute *a;
349 
350 	if (!s)
351 		return NULL;
352 
353 	if (!name_nonce)
354 		n = kasprintf(GFP_KERNEL, "%.*s", name_max, name);
355 	else
356 		n = kasprintf(GFP_KERNEL, "%.*s__%d", name_max, name,
357 					name_nonce);
358 	if (!n)
359 		goto out_s;
360 
361 	a = device_str_attr_create_(n, s);
362 	if (!a)
363 		goto out_n;
364 
365 	return a;
366 out_n:
367 	kfree(n);
368 out_s:
369 	kfree(s);
370 	return NULL;
371 }
372 
373 static void device_str_attr_destroy(struct attribute *attr)
374 {
375 	struct dev_ext_attribute *d;
376 
377 	d = container_of(attr, struct dev_ext_attribute, attr.attr);
378 	kfree(d->var);
379 	kfree(d->attr.attr.name);
380 	kfree(d);
381 }
382 
383 static struct attribute *event_to_attr(unsigned ix,
384 				       struct hv_24x7_event_data *event,
385 				       unsigned domain,
386 				       int nonce)
387 {
388 	int event_name_len;
389 	char *ev_name, *a_ev_name, *val;
390 	const char *ev_suffix;
391 	struct attribute *attr;
392 
393 	if (!domain_is_valid(domain)) {
394 		pr_warn("catalog event %u has invalid domain %u\n",
395 				ix, domain);
396 		return NULL;
397 	}
398 
399 	val = event_fmt(event, domain);
400 	if (!val)
401 		return NULL;
402 
403 	ev_suffix = event_domain_suffix(domain);
404 	ev_name = event_name(event, &event_name_len);
405 	if (!nonce)
406 		a_ev_name = kasprintf(GFP_KERNEL, "%.*s%s",
407 				(int)event_name_len, ev_name, ev_suffix);
408 	else
409 		a_ev_name = kasprintf(GFP_KERNEL, "%.*s%s__%d",
410 				(int)event_name_len, ev_name, ev_suffix, nonce);
411 
412 	if (!a_ev_name)
413 		goto out_val;
414 
415 	attr = device_str_attr_create_(a_ev_name, val);
416 	if (!attr)
417 		goto out_name;
418 
419 	return attr;
420 out_name:
421 	kfree(a_ev_name);
422 out_val:
423 	kfree(val);
424 	return NULL;
425 }
426 
427 static struct attribute *event_to_desc_attr(struct hv_24x7_event_data *event,
428 					    int nonce)
429 {
430 	int nl, dl;
431 	char *name = event_name(event, &nl);
432 	char *desc = event_desc(event, &dl);
433 
434 	/* If there isn't a description, don't create the sysfs file */
435 	if (!dl)
436 		return NULL;
437 
438 	return device_str_attr_create(name, nl, nonce, desc, dl);
439 }
440 
441 static struct attribute *
442 event_to_long_desc_attr(struct hv_24x7_event_data *event, int nonce)
443 {
444 	int nl, dl;
445 	char *name = event_name(event, &nl);
446 	char *desc = event_long_desc(event, &dl);
447 
448 	/* If there isn't a description, don't create the sysfs file */
449 	if (!dl)
450 		return NULL;
451 
452 	return device_str_attr_create(name, nl, nonce, desc, dl);
453 }
454 
455 static ssize_t event_data_to_attrs(unsigned ix, struct attribute **attrs,
456 				   struct hv_24x7_event_data *event, int nonce)
457 {
458 	unsigned i;
459 
460 	switch (event->domain) {
461 	case HV_PERF_DOMAIN_PHYS_CHIP:
462 		*attrs = event_to_attr(ix, event, event->domain, nonce);
463 		return 1;
464 	case HV_PERF_DOMAIN_PHYS_CORE:
465 		for (i = 0; i < ARRAY_SIZE(core_domains); i++) {
466 			attrs[i] = event_to_attr(ix, event, core_domains[i],
467 						nonce);
468 			if (!attrs[i]) {
469 				pr_warn("catalog event %u: individual attr %u "
470 					"creation failure\n", ix, i);
471 				for (; i; i--)
472 					device_str_attr_destroy(attrs[i - 1]);
473 				return -1;
474 			}
475 		}
476 		return i;
477 	default:
478 		pr_warn("catalog event %u: domain %u is not allowed in the "
479 				"catalog\n", ix, event->domain);
480 		return -1;
481 	}
482 }
483 
484 static size_t event_to_attr_ct(struct hv_24x7_event_data *event)
485 {
486 	switch (event->domain) {
487 	case HV_PERF_DOMAIN_PHYS_CHIP:
488 		return 1;
489 	case HV_PERF_DOMAIN_PHYS_CORE:
490 		return ARRAY_SIZE(core_domains);
491 	default:
492 		return 0;
493 	}
494 }
495 
496 /* */
497 struct event_uniq {
498 	struct rb_node node;
499 	const char *name;
500 	int nl;
501 	unsigned ct;
502 	unsigned domain;
503 };
504 
505 static int memord(const void *d1, size_t s1, const void *d2, size_t s2)
506 {
507 	if (s1 < s2)
508 		return 1;
509 	if (s2 > s1)
510 		return -1;
511 
512 	return memcmp(d1, d2, s1);
513 }
514 
515 static int ev_uniq_ord(const void *v1, size_t s1, unsigned d1, const void *v2,
516 		       size_t s2, unsigned d2)
517 {
518 	int r = memord(v1, s1, v2, s2);
519 
520 	if (r)
521 		return r;
522 	if (d1 > d2)
523 		return 1;
524 	if (d2 > d1)
525 		return -1;
526 	return 0;
527 }
528 
529 static int event_uniq_add(struct rb_root *root, const char *name, int nl,
530 			  unsigned domain)
531 {
532 	struct rb_node **new = &(root->rb_node), *parent = NULL;
533 	struct event_uniq *data;
534 
535 	/* Figure out where to put new node */
536 	while (*new) {
537 		struct event_uniq *it;
538 		int result;
539 
540 		it = container_of(*new, struct event_uniq, node);
541 		result = ev_uniq_ord(name, nl, domain, it->name, it->nl,
542 					it->domain);
543 
544 		parent = *new;
545 		if (result < 0)
546 			new = &((*new)->rb_left);
547 		else if (result > 0)
548 			new = &((*new)->rb_right);
549 		else {
550 			it->ct++;
551 			pr_info("found a duplicate event %.*s, ct=%u\n", nl,
552 						name, it->ct);
553 			return it->ct;
554 		}
555 	}
556 
557 	data = kmalloc(sizeof(*data), GFP_KERNEL);
558 	if (!data)
559 		return -ENOMEM;
560 
561 	*data = (struct event_uniq) {
562 		.name = name,
563 		.nl = nl,
564 		.ct = 0,
565 		.domain = domain,
566 	};
567 
568 	/* Add new node and rebalance tree. */
569 	rb_link_node(&data->node, parent, new);
570 	rb_insert_color(&data->node, root);
571 
572 	/* data->ct */
573 	return 0;
574 }
575 
576 static void event_uniq_destroy(struct rb_root *root)
577 {
578 	/*
579 	 * the strings we point to are in the giant block of memory filled by
580 	 * the catalog, and are freed separately.
581 	 */
582 	struct event_uniq *pos, *n;
583 
584 	rbtree_postorder_for_each_entry_safe(pos, n, root, node)
585 		kfree(pos);
586 }
587 
588 
589 /*
590  * ensure the event structure's sizes are self consistent and don't cause us to
591  * read outside of the event
592  *
593  * On success, return the event length in bytes.
594  * Otherwise, return -1 (and print as appropriate).
595  */
596 static ssize_t catalog_event_len_validate(struct hv_24x7_event_data *event,
597 					  size_t event_idx,
598 					  size_t event_data_bytes,
599 					  size_t event_entry_count,
600 					  size_t offset, void *end)
601 {
602 	ssize_t ev_len;
603 	void *ev_end, *calc_ev_end;
604 
605 	if (offset >= event_data_bytes)
606 		return -1;
607 
608 	if (event_idx >= event_entry_count) {
609 		pr_devel("catalog event data has %zu bytes of padding after last event\n",
610 				event_data_bytes - offset);
611 		return -1;
612 	}
613 
614 	if (!event_fixed_portion_is_within(event, end)) {
615 		pr_warn("event %zu fixed portion is not within range\n",
616 				event_idx);
617 		return -1;
618 	}
619 
620 	ev_len = be16_to_cpu(event->length);
621 
622 	if (ev_len % 16)
623 		pr_info("event %zu has length %zu not divisible by 16: event=%pK\n",
624 				event_idx, ev_len, event);
625 
626 	ev_end = (__u8 *)event + ev_len;
627 	if (ev_end > end) {
628 		pr_warn("event %zu has .length=%zu, ends after buffer end: ev_end=%pK > end=%pK, offset=%zu\n",
629 				event_idx, ev_len, ev_end, end,
630 				offset);
631 		return -1;
632 	}
633 
634 	calc_ev_end = event_end(event, end);
635 	if (!calc_ev_end) {
636 		pr_warn("event %zu has a calculated length which exceeds buffer length %zu: event=%pK end=%pK, offset=%zu\n",
637 			event_idx, event_data_bytes, event, end,
638 			offset);
639 		return -1;
640 	}
641 
642 	if (calc_ev_end > ev_end) {
643 		pr_warn("event %zu exceeds it's own length: event=%pK, end=%pK, offset=%zu, calc_ev_end=%pK\n",
644 			event_idx, event, ev_end, offset, calc_ev_end);
645 		return -1;
646 	}
647 
648 	return ev_len;
649 }
650 
651 #define MAX_4K (SIZE_MAX / 4096)
652 
653 static int create_events_from_catalog(struct attribute ***events_,
654 				      struct attribute ***event_descs_,
655 				      struct attribute ***event_long_descs_)
656 {
657 	unsigned long hret;
658 	size_t catalog_len, catalog_page_len, event_entry_count,
659 	       event_data_len, event_data_offs,
660 	       event_data_bytes, junk_events, event_idx, event_attr_ct, i,
661 	       attr_max, event_idx_last, desc_ct, long_desc_ct;
662 	ssize_t ct, ev_len;
663 	uint32_t catalog_version_num;
664 	struct attribute **events, **event_descs, **event_long_descs;
665 	struct hv_24x7_catalog_page_0 *page_0 =
666 		kmem_cache_alloc(hv_page_cache, GFP_KERNEL);
667 	void *page = page_0;
668 	void *event_data, *end;
669 	struct hv_24x7_event_data *event;
670 	struct rb_root ev_uniq = RB_ROOT;
671 	int ret = 0;
672 
673 	if (!page) {
674 		ret = -ENOMEM;
675 		goto e_out;
676 	}
677 
678 	hret = h_get_24x7_catalog_page(page, 0, 0);
679 	if (hret) {
680 		ret = -EIO;
681 		goto e_free;
682 	}
683 
684 	catalog_version_num = be64_to_cpu(page_0->version);
685 	catalog_page_len = be32_to_cpu(page_0->length);
686 
687 	if (MAX_4K < catalog_page_len) {
688 		pr_err("invalid page count: %zu\n", catalog_page_len);
689 		ret = -EIO;
690 		goto e_free;
691 	}
692 
693 	catalog_len = catalog_page_len * 4096;
694 
695 	event_entry_count = be16_to_cpu(page_0->event_entry_count);
696 	event_data_offs   = be16_to_cpu(page_0->event_data_offs);
697 	event_data_len    = be16_to_cpu(page_0->event_data_len);
698 
699 	pr_devel("cv %zu cl %zu eec %zu edo %zu edl %zu\n",
700 			(size_t)catalog_version_num, catalog_len,
701 			event_entry_count, event_data_offs, event_data_len);
702 
703 	if ((MAX_4K < event_data_len)
704 			|| (MAX_4K < event_data_offs)
705 			|| (MAX_4K - event_data_offs < event_data_len)) {
706 		pr_err("invalid event data offs %zu and/or len %zu\n",
707 				event_data_offs, event_data_len);
708 		ret = -EIO;
709 		goto e_free;
710 	}
711 
712 	if ((event_data_offs + event_data_len) > catalog_page_len) {
713 		pr_err("event data %zu-%zu does not fit inside catalog 0-%zu\n",
714 				event_data_offs,
715 				event_data_offs + event_data_len,
716 				catalog_page_len);
717 		ret = -EIO;
718 		goto e_free;
719 	}
720 
721 	if (SIZE_MAX / MAX_EVENTS_PER_EVENT_DATA - 1 < event_entry_count) {
722 		pr_err("event_entry_count %zu is invalid\n",
723 				event_entry_count);
724 		ret = -EIO;
725 		goto e_free;
726 	}
727 
728 	event_data_bytes = event_data_len * 4096;
729 
730 	/*
731 	 * event data can span several pages, events can cross between these
732 	 * pages. Use vmalloc to make this easier.
733 	 */
734 	event_data = vmalloc(event_data_bytes);
735 	if (!event_data) {
736 		pr_err("could not allocate event data\n");
737 		ret = -ENOMEM;
738 		goto e_free;
739 	}
740 
741 	end = event_data + event_data_bytes;
742 
743 	/*
744 	 * using vmalloc_to_phys() like this only works if PAGE_SIZE is
745 	 * divisible by 4096
746 	 */
747 	BUILD_BUG_ON(PAGE_SIZE % 4096);
748 
749 	for (i = 0; i < event_data_len; i++) {
750 		hret = h_get_24x7_catalog_page_(
751 				vmalloc_to_phys(event_data + i * 4096),
752 				catalog_version_num,
753 				i + event_data_offs);
754 		if (hret) {
755 			pr_err("failed to get event data in page %zu\n",
756 					i + event_data_offs);
757 			ret = -EIO;
758 			goto e_event_data;
759 		}
760 	}
761 
762 	/*
763 	 * scan the catalog to determine the number of attributes we need, and
764 	 * verify it at the same time.
765 	 */
766 	for (junk_events = 0, event = event_data, event_idx = 0, attr_max = 0;
767 	     ;
768 	     event_idx++, event = (void *)event + ev_len) {
769 		size_t offset = (void *)event - (void *)event_data;
770 		char *name;
771 		int nl;
772 
773 		ev_len = catalog_event_len_validate(event, event_idx,
774 						    event_data_bytes,
775 						    event_entry_count,
776 						    offset, end);
777 		if (ev_len < 0)
778 			break;
779 
780 		name = event_name(event, &nl);
781 
782 		if (event->event_group_record_len == 0) {
783 			pr_devel("invalid event %zu (%.*s): group_record_len == 0, skipping\n",
784 					event_idx, nl, name);
785 			junk_events++;
786 			continue;
787 		}
788 
789 		if (!catalog_entry_domain_is_valid(event->domain)) {
790 			pr_info("event %zu (%.*s) has invalid domain %d\n",
791 					event_idx, nl, name, event->domain);
792 			junk_events++;
793 			continue;
794 		}
795 
796 		attr_max += event_to_attr_ct(event);
797 	}
798 
799 	event_idx_last = event_idx;
800 	if (event_idx_last != event_entry_count)
801 		pr_warn("event buffer ended before listed # of events were parsed (got %zu, wanted %zu, junk %zu)\n",
802 				event_idx_last, event_entry_count, junk_events);
803 
804 	events = kmalloc_array(attr_max + 1, sizeof(*events), GFP_KERNEL);
805 	if (!events) {
806 		ret = -ENOMEM;
807 		goto e_event_data;
808 	}
809 
810 	event_descs = kmalloc_array(event_idx + 1, sizeof(*event_descs),
811 				GFP_KERNEL);
812 	if (!event_descs) {
813 		ret = -ENOMEM;
814 		goto e_event_attrs;
815 	}
816 
817 	event_long_descs = kmalloc_array(event_idx + 1,
818 			sizeof(*event_long_descs), GFP_KERNEL);
819 	if (!event_long_descs) {
820 		ret = -ENOMEM;
821 		goto e_event_descs;
822 	}
823 
824 	/* Iterate over the catalog filling in the attribute vector */
825 	for (junk_events = 0, event_attr_ct = 0, desc_ct = 0, long_desc_ct = 0,
826 				event = event_data, event_idx = 0;
827 			event_idx < event_idx_last;
828 			event_idx++, ev_len = be16_to_cpu(event->length),
829 				event = (void *)event + ev_len) {
830 		char *name;
831 		int nl;
832 		int nonce;
833 		/*
834 		 * these are the only "bad" events that are intermixed and that
835 		 * we can ignore without issue. make sure to skip them here
836 		 */
837 		if (event->event_group_record_len == 0)
838 			continue;
839 		if (!catalog_entry_domain_is_valid(event->domain))
840 			continue;
841 
842 		name  = event_name(event, &nl);
843 		nonce = event_uniq_add(&ev_uniq, name, nl, event->domain);
844 		ct    = event_data_to_attrs(event_idx, events + event_attr_ct,
845 					    event, nonce);
846 		if (ct <= 0) {
847 			pr_warn("event %zu (%.*s) creation failure, skipping\n",
848 				event_idx, nl, name);
849 			junk_events++;
850 		} else {
851 			event_attr_ct += ct;
852 			event_descs[desc_ct] = event_to_desc_attr(event, nonce);
853 			if (event_descs[desc_ct])
854 				desc_ct++;
855 			event_long_descs[long_desc_ct] =
856 					event_to_long_desc_attr(event, nonce);
857 			if (event_long_descs[long_desc_ct])
858 				long_desc_ct++;
859 		}
860 	}
861 
862 	pr_info("read %zu catalog entries, created %zu event attrs (%zu failures), %zu descs\n",
863 			event_idx, event_attr_ct, junk_events, desc_ct);
864 
865 	events[event_attr_ct] = NULL;
866 	event_descs[desc_ct] = NULL;
867 	event_long_descs[long_desc_ct] = NULL;
868 
869 	event_uniq_destroy(&ev_uniq);
870 	vfree(event_data);
871 	kmem_cache_free(hv_page_cache, page);
872 
873 	*events_ = events;
874 	*event_descs_ = event_descs;
875 	*event_long_descs_ = event_long_descs;
876 	return 0;
877 
878 e_event_descs:
879 	kfree(event_descs);
880 e_event_attrs:
881 	kfree(events);
882 e_event_data:
883 	vfree(event_data);
884 e_free:
885 	kmem_cache_free(hv_page_cache, page);
886 e_out:
887 	*events_ = NULL;
888 	*event_descs_ = NULL;
889 	*event_long_descs_ = NULL;
890 	return ret;
891 }
892 
893 static ssize_t catalog_read(struct file *filp, struct kobject *kobj,
894 			    struct bin_attribute *bin_attr, char *buf,
895 			    loff_t offset, size_t count)
896 {
897 	unsigned long hret;
898 	ssize_t ret = 0;
899 	size_t catalog_len = 0, catalog_page_len = 0;
900 	loff_t page_offset = 0;
901 	loff_t offset_in_page;
902 	size_t copy_len;
903 	uint64_t catalog_version_num = 0;
904 	void *page = kmem_cache_alloc(hv_page_cache, GFP_USER);
905 	struct hv_24x7_catalog_page_0 *page_0 = page;
906 
907 	if (!page)
908 		return -ENOMEM;
909 
910 	hret = h_get_24x7_catalog_page(page, 0, 0);
911 	if (hret) {
912 		ret = -EIO;
913 		goto e_free;
914 	}
915 
916 	catalog_version_num = be64_to_cpu(page_0->version);
917 	catalog_page_len = be32_to_cpu(page_0->length);
918 	catalog_len = catalog_page_len * 4096;
919 
920 	page_offset = offset / 4096;
921 	offset_in_page = offset % 4096;
922 
923 	if (page_offset >= catalog_page_len)
924 		goto e_free;
925 
926 	if (page_offset != 0) {
927 		hret = h_get_24x7_catalog_page(page, catalog_version_num,
928 					       page_offset);
929 		if (hret) {
930 			ret = -EIO;
931 			goto e_free;
932 		}
933 	}
934 
935 	copy_len = 4096 - offset_in_page;
936 	if (copy_len > count)
937 		copy_len = count;
938 
939 	memcpy(buf, page+offset_in_page, copy_len);
940 	ret = copy_len;
941 
942 e_free:
943 	if (hret)
944 		pr_err("h_get_24x7_catalog_page(ver=%lld, page=%lld) failed:"
945 		       " rc=%ld\n",
946 		       catalog_version_num, page_offset, hret);
947 	kmem_cache_free(hv_page_cache, page);
948 
949 	pr_devel("catalog_read: offset=%lld(%lld) count=%zu "
950 			"catalog_len=%zu(%zu) => %zd\n", offset, page_offset,
951 			count, catalog_len, catalog_page_len, ret);
952 
953 	return ret;
954 }
955 
956 #define PAGE_0_ATTR(_name, _fmt, _expr)				\
957 static ssize_t _name##_show(struct device *dev,			\
958 			    struct device_attribute *dev_attr,	\
959 			    char *buf)				\
960 {								\
961 	unsigned long hret;					\
962 	ssize_t ret = 0;					\
963 	void *page = kmem_cache_alloc(hv_page_cache, GFP_USER);	\
964 	struct hv_24x7_catalog_page_0 *page_0 = page;		\
965 	if (!page)						\
966 		return -ENOMEM;					\
967 	hret = h_get_24x7_catalog_page(page, 0, 0);		\
968 	if (hret) {						\
969 		ret = -EIO;					\
970 		goto e_free;					\
971 	}							\
972 	ret = sprintf(buf, _fmt, _expr);			\
973 e_free:								\
974 	kmem_cache_free(hv_page_cache, page);			\
975 	return ret;						\
976 }								\
977 static DEVICE_ATTR_RO(_name)
978 
979 PAGE_0_ATTR(catalog_version, "%lld\n",
980 		(unsigned long long)be64_to_cpu(page_0->version));
981 PAGE_0_ATTR(catalog_len, "%lld\n",
982 		(unsigned long long)be32_to_cpu(page_0->length) * 4096);
983 static BIN_ATTR_RO(catalog, 0/* real length varies */);
984 
985 static struct bin_attribute *if_bin_attrs[] = {
986 	&bin_attr_catalog,
987 	NULL,
988 };
989 
990 static struct attribute *if_attrs[] = {
991 	&dev_attr_catalog_len.attr,
992 	&dev_attr_catalog_version.attr,
993 	NULL,
994 };
995 
996 static struct attribute_group if_group = {
997 	.name = "interface",
998 	.bin_attrs = if_bin_attrs,
999 	.attrs = if_attrs,
1000 };
1001 
1002 static const struct attribute_group *attr_groups[] = {
1003 	&format_group,
1004 	&event_group,
1005 	&event_desc_group,
1006 	&event_long_desc_group,
1007 	&if_group,
1008 	NULL,
1009 };
1010 
1011 static void log_24x7_hcall(struct hv_24x7_request_buffer *request_buffer,
1012 			   struct hv_24x7_data_result_buffer *result_buffer,
1013 			   unsigned long ret)
1014 {
1015 	struct hv_24x7_request *req;
1016 
1017 	req = &request_buffer->requests[0];
1018 	pr_notice_ratelimited("hcall failed: [%d %#x %#x %d] => "
1019 			"ret 0x%lx (%ld) detail=0x%x failing ix=%x\n",
1020 			req->performance_domain, req->data_offset,
1021 			req->starting_ix, req->starting_lpar_ix, ret, ret,
1022 			result_buffer->detailed_rc,
1023 			result_buffer->failing_request_ix);
1024 }
1025 
1026 /*
1027  * Start the process for a new H_GET_24x7_DATA hcall.
1028  */
1029 static void init_24x7_request(struct hv_24x7_request_buffer *request_buffer,
1030 			      struct hv_24x7_data_result_buffer *result_buffer)
1031 {
1032 
1033 	memset(request_buffer, 0, 4096);
1034 	memset(result_buffer, 0, 4096);
1035 
1036 	request_buffer->interface_version = HV_24X7_IF_VERSION_CURRENT;
1037 	/* memset above set request_buffer->num_requests to 0 */
1038 }
1039 
1040 /*
1041  * Commit (i.e perform) the H_GET_24x7_DATA hcall using the data collected
1042  * by 'init_24x7_request()' and 'add_event_to_24x7_request()'.
1043  */
1044 static int make_24x7_request(struct hv_24x7_request_buffer *request_buffer,
1045 			     struct hv_24x7_data_result_buffer *result_buffer)
1046 {
1047 	unsigned long ret;
1048 
1049 	/*
1050 	 * NOTE: Due to variable number of array elements in request and
1051 	 *	 result buffer(s), sizeof() is not reliable. Use the actual
1052 	 *	 allocated buffer size, H24x7_DATA_BUFFER_SIZE.
1053 	 */
1054 	ret = plpar_hcall_norets(H_GET_24X7_DATA,
1055 			virt_to_phys(request_buffer), H24x7_DATA_BUFFER_SIZE,
1056 			virt_to_phys(result_buffer),  H24x7_DATA_BUFFER_SIZE);
1057 
1058 	if (ret)
1059 		log_24x7_hcall(request_buffer, result_buffer, ret);
1060 
1061 	return ret;
1062 }
1063 
1064 /*
1065  * Add the given @event to the next slot in the 24x7 request_buffer.
1066  *
1067  * Note that H_GET_24X7_DATA hcall allows reading several counters'
1068  * values in a single HCALL. We expect the caller to add events to the
1069  * request buffer one by one, make the HCALL and process the results.
1070  */
1071 static int add_event_to_24x7_request(struct perf_event *event,
1072 				struct hv_24x7_request_buffer *request_buffer)
1073 {
1074 	u16 idx;
1075 	int i;
1076 	struct hv_24x7_request *req;
1077 
1078 	if (request_buffer->num_requests > 254) {
1079 		pr_devel("Too many requests for 24x7 HCALL %d\n",
1080 				request_buffer->num_requests);
1081 		return -EINVAL;
1082 	}
1083 
1084 	if (is_physical_domain(event_get_domain(event)))
1085 		idx = event_get_core(event);
1086 	else
1087 		idx = event_get_vcpu(event);
1088 
1089 	i = request_buffer->num_requests++;
1090 	req = &request_buffer->requests[i];
1091 
1092 	req->performance_domain = event_get_domain(event);
1093 	req->data_size = cpu_to_be16(8);
1094 	req->data_offset = cpu_to_be32(event_get_offset(event));
1095 	req->starting_lpar_ix = cpu_to_be16(event_get_lpar(event)),
1096 	req->max_num_lpars = cpu_to_be16(1);
1097 	req->starting_ix = cpu_to_be16(idx);
1098 	req->max_ix = cpu_to_be16(1);
1099 
1100 	return 0;
1101 }
1102 
1103 static unsigned long single_24x7_request(struct perf_event *event, u64 *count)
1104 {
1105 	unsigned long ret;
1106 	struct hv_24x7_request_buffer *request_buffer;
1107 	struct hv_24x7_data_result_buffer *result_buffer;
1108 
1109 	BUILD_BUG_ON(sizeof(*request_buffer) > 4096);
1110 	BUILD_BUG_ON(sizeof(*result_buffer) > 4096);
1111 
1112 	request_buffer = (void *)get_cpu_var(hv_24x7_reqb);
1113 	result_buffer = (void *)get_cpu_var(hv_24x7_resb);
1114 
1115 	init_24x7_request(request_buffer, result_buffer);
1116 
1117 	ret = add_event_to_24x7_request(event, request_buffer);
1118 	if (ret)
1119 		goto out;
1120 
1121 	ret = make_24x7_request(request_buffer, result_buffer);
1122 	if (ret) {
1123 		log_24x7_hcall(request_buffer, result_buffer, ret);
1124 		goto out;
1125 	}
1126 
1127 	/* process result from hcall */
1128 	*count = be64_to_cpu(result_buffer->results[0].elements[0].element_data[0]);
1129 
1130 out:
1131 	put_cpu_var(hv_24x7_reqb);
1132 	put_cpu_var(hv_24x7_resb);
1133 	return ret;
1134 }
1135 
1136 
1137 static int h_24x7_event_init(struct perf_event *event)
1138 {
1139 	struct hv_perf_caps caps;
1140 	unsigned domain;
1141 	unsigned long hret;
1142 	u64 ct;
1143 
1144 	/* Not our event */
1145 	if (event->attr.type != event->pmu->type)
1146 		return -ENOENT;
1147 
1148 	/* Unused areas must be 0 */
1149 	if (event_get_reserved1(event) ||
1150 	    event_get_reserved2(event) ||
1151 	    event_get_reserved3(event)) {
1152 		pr_devel("reserved set when forbidden 0x%llx(0x%llx) 0x%llx(0x%llx) 0x%llx(0x%llx)\n",
1153 				event->attr.config,
1154 				event_get_reserved1(event),
1155 				event->attr.config1,
1156 				event_get_reserved2(event),
1157 				event->attr.config2,
1158 				event_get_reserved3(event));
1159 		return -EINVAL;
1160 	}
1161 
1162 	/* unsupported modes and filters */
1163 	if (event->attr.exclude_user   ||
1164 	    event->attr.exclude_kernel ||
1165 	    event->attr.exclude_hv     ||
1166 	    event->attr.exclude_idle   ||
1167 	    event->attr.exclude_host   ||
1168 	    event->attr.exclude_guest)
1169 		return -EINVAL;
1170 
1171 	/* no branch sampling */
1172 	if (has_branch_stack(event))
1173 		return -EOPNOTSUPP;
1174 
1175 	/* offset must be 8 byte aligned */
1176 	if (event_get_offset(event) % 8) {
1177 		pr_devel("bad alignment\n");
1178 		return -EINVAL;
1179 	}
1180 
1181 	/* Domains above 6 are invalid */
1182 	domain = event_get_domain(event);
1183 	if (domain > 6) {
1184 		pr_devel("invalid domain %d\n", domain);
1185 		return -EINVAL;
1186 	}
1187 
1188 	hret = hv_perf_caps_get(&caps);
1189 	if (hret) {
1190 		pr_devel("could not get capabilities: rc=%ld\n", hret);
1191 		return -EIO;
1192 	}
1193 
1194 	/* Physical domains & other lpars require extra capabilities */
1195 	if (!caps.collect_privileged && (is_physical_domain(domain) ||
1196 		(event_get_lpar(event) != event_get_lpar_max()))) {
1197 		pr_devel("hv permissions disallow: is_physical_domain:%d, lpar=0x%llx\n",
1198 				is_physical_domain(domain),
1199 				event_get_lpar(event));
1200 		return -EACCES;
1201 	}
1202 
1203 	/* see if the event complains */
1204 	if (single_24x7_request(event, &ct)) {
1205 		pr_devel("test hcall failed\n");
1206 		return -EIO;
1207 	}
1208 
1209 	return 0;
1210 }
1211 
1212 static u64 h_24x7_get_value(struct perf_event *event)
1213 {
1214 	unsigned long ret;
1215 	u64 ct;
1216 	ret = single_24x7_request(event, &ct);
1217 	if (ret)
1218 		/* We checked this in event init, shouldn't fail here... */
1219 		return 0;
1220 
1221 	return ct;
1222 }
1223 
1224 static void update_event_count(struct perf_event *event, u64 now)
1225 {
1226 	s64 prev;
1227 
1228 	prev = local64_xchg(&event->hw.prev_count, now);
1229 	local64_add(now - prev, &event->count);
1230 }
1231 
1232 static void h_24x7_event_read(struct perf_event *event)
1233 {
1234 	u64 now;
1235 	struct hv_24x7_request_buffer *request_buffer;
1236 	struct hv_24x7_hw *h24x7hw;
1237 	int txn_flags;
1238 
1239 	txn_flags = __this_cpu_read(hv_24x7_txn_flags);
1240 
1241 	/*
1242 	 * If in a READ transaction, add this counter to the list of
1243 	 * counters to read during the next HCALL (i.e commit_txn()).
1244 	 * If not in a READ transaction, go ahead and make the HCALL
1245 	 * to read this counter by itself.
1246 	 */
1247 
1248 	if (txn_flags & PERF_PMU_TXN_READ) {
1249 		int i;
1250 		int ret;
1251 
1252 		if (__this_cpu_read(hv_24x7_txn_err))
1253 			return;
1254 
1255 		request_buffer = (void *)get_cpu_var(hv_24x7_reqb);
1256 
1257 		ret = add_event_to_24x7_request(event, request_buffer);
1258 		if (ret) {
1259 			__this_cpu_write(hv_24x7_txn_err, ret);
1260 		} else {
1261 			/*
1262 			 * Assoicate the event with the HCALL request index,
1263 			 * so ->commit_txn() can quickly find/update count.
1264 			 */
1265 			i = request_buffer->num_requests - 1;
1266 
1267 			h24x7hw = &get_cpu_var(hv_24x7_hw);
1268 			h24x7hw->events[i] = event;
1269 			put_cpu_var(h24x7hw);
1270 		}
1271 
1272 		put_cpu_var(hv_24x7_reqb);
1273 	} else {
1274 		now = h_24x7_get_value(event);
1275 		update_event_count(event, now);
1276 	}
1277 }
1278 
1279 static void h_24x7_event_start(struct perf_event *event, int flags)
1280 {
1281 	if (flags & PERF_EF_RELOAD)
1282 		local64_set(&event->hw.prev_count, h_24x7_get_value(event));
1283 }
1284 
1285 static void h_24x7_event_stop(struct perf_event *event, int flags)
1286 {
1287 	h_24x7_event_read(event);
1288 }
1289 
1290 static int h_24x7_event_add(struct perf_event *event, int flags)
1291 {
1292 	if (flags & PERF_EF_START)
1293 		h_24x7_event_start(event, flags);
1294 
1295 	return 0;
1296 }
1297 
1298 /*
1299  * 24x7 counters only support READ transactions. They are
1300  * always counting and dont need/support ADD transactions.
1301  * Cache the flags, but otherwise ignore transactions that
1302  * are not PERF_PMU_TXN_READ.
1303  */
1304 static void h_24x7_event_start_txn(struct pmu *pmu, unsigned int flags)
1305 {
1306 	struct hv_24x7_request_buffer *request_buffer;
1307 	struct hv_24x7_data_result_buffer *result_buffer;
1308 
1309 	/* We should not be called if we are already in a txn */
1310 	WARN_ON_ONCE(__this_cpu_read(hv_24x7_txn_flags));
1311 
1312 	__this_cpu_write(hv_24x7_txn_flags, flags);
1313 	if (flags & ~PERF_PMU_TXN_READ)
1314 		return;
1315 
1316 	request_buffer = (void *)get_cpu_var(hv_24x7_reqb);
1317 	result_buffer = (void *)get_cpu_var(hv_24x7_resb);
1318 
1319 	init_24x7_request(request_buffer, result_buffer);
1320 
1321 	put_cpu_var(hv_24x7_resb);
1322 	put_cpu_var(hv_24x7_reqb);
1323 }
1324 
1325 /*
1326  * Clean up transaction state.
1327  *
1328  * NOTE: Ignore state of request and result buffers for now.
1329  *	 We will initialize them during the next read/txn.
1330  */
1331 static void reset_txn(void)
1332 {
1333 	__this_cpu_write(hv_24x7_txn_flags, 0);
1334 	__this_cpu_write(hv_24x7_txn_err, 0);
1335 }
1336 
1337 /*
1338  * 24x7 counters only support READ transactions. They are always counting
1339  * and dont need/support ADD transactions. Clear ->txn_flags but otherwise
1340  * ignore transactions that are not of type PERF_PMU_TXN_READ.
1341  *
1342  * For READ transactions, submit all pending 24x7 requests (i.e requests
1343  * that were queued by h_24x7_event_read()), to the hypervisor and update
1344  * the event counts.
1345  */
1346 static int h_24x7_event_commit_txn(struct pmu *pmu)
1347 {
1348 	struct hv_24x7_request_buffer *request_buffer;
1349 	struct hv_24x7_data_result_buffer *result_buffer;
1350 	struct hv_24x7_result *resb;
1351 	struct perf_event *event;
1352 	u64 count;
1353 	int i, ret, txn_flags;
1354 	struct hv_24x7_hw *h24x7hw;
1355 
1356 	txn_flags = __this_cpu_read(hv_24x7_txn_flags);
1357 	WARN_ON_ONCE(!txn_flags);
1358 
1359 	ret = 0;
1360 	if (txn_flags & ~PERF_PMU_TXN_READ)
1361 		goto out;
1362 
1363 	ret = __this_cpu_read(hv_24x7_txn_err);
1364 	if (ret)
1365 		goto out;
1366 
1367 	request_buffer = (void *)get_cpu_var(hv_24x7_reqb);
1368 	result_buffer = (void *)get_cpu_var(hv_24x7_resb);
1369 
1370 	ret = make_24x7_request(request_buffer, result_buffer);
1371 	if (ret) {
1372 		log_24x7_hcall(request_buffer, result_buffer, ret);
1373 		goto put_reqb;
1374 	}
1375 
1376 	h24x7hw = &get_cpu_var(hv_24x7_hw);
1377 
1378 	/* Update event counts from hcall */
1379 	for (i = 0; i < request_buffer->num_requests; i++) {
1380 		resb = &result_buffer->results[i];
1381 		count = be64_to_cpu(resb->elements[0].element_data[0]);
1382 		event = h24x7hw->events[i];
1383 		h24x7hw->events[i] = NULL;
1384 		update_event_count(event, count);
1385 	}
1386 
1387 	put_cpu_var(hv_24x7_hw);
1388 
1389 put_reqb:
1390 	put_cpu_var(hv_24x7_resb);
1391 	put_cpu_var(hv_24x7_reqb);
1392 out:
1393 	reset_txn();
1394 	return ret;
1395 }
1396 
1397 /*
1398  * 24x7 counters only support READ transactions. They are always counting
1399  * and dont need/support ADD transactions. However, regardless of type
1400  * of transaction, all we need to do is cleanup, so we don't have to check
1401  * the type of transaction.
1402  */
1403 static void h_24x7_event_cancel_txn(struct pmu *pmu)
1404 {
1405 	WARN_ON_ONCE(!__this_cpu_read(hv_24x7_txn_flags));
1406 	reset_txn();
1407 }
1408 
1409 static struct pmu h_24x7_pmu = {
1410 	.task_ctx_nr = perf_invalid_context,
1411 
1412 	.name = "hv_24x7",
1413 	.attr_groups = attr_groups,
1414 	.event_init  = h_24x7_event_init,
1415 	.add         = h_24x7_event_add,
1416 	.del         = h_24x7_event_stop,
1417 	.start       = h_24x7_event_start,
1418 	.stop        = h_24x7_event_stop,
1419 	.read        = h_24x7_event_read,
1420 	.start_txn   = h_24x7_event_start_txn,
1421 	.commit_txn  = h_24x7_event_commit_txn,
1422 	.cancel_txn  = h_24x7_event_cancel_txn,
1423 };
1424 
1425 static int hv_24x7_init(void)
1426 {
1427 	int r;
1428 	unsigned long hret;
1429 	struct hv_perf_caps caps;
1430 
1431 	if (!firmware_has_feature(FW_FEATURE_LPAR)) {
1432 		pr_debug("not a virtualized system, not enabling\n");
1433 		return -ENODEV;
1434 	}
1435 
1436 	hret = hv_perf_caps_get(&caps);
1437 	if (hret) {
1438 		pr_debug("could not obtain capabilities, not enabling, rc=%ld\n",
1439 				hret);
1440 		return -ENODEV;
1441 	}
1442 
1443 	hv_page_cache = kmem_cache_create("hv-page-4096", 4096, 4096, 0, NULL);
1444 	if (!hv_page_cache)
1445 		return -ENOMEM;
1446 
1447 	/* sampling not supported */
1448 	h_24x7_pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
1449 
1450 	r = create_events_from_catalog(&event_group.attrs,
1451 				   &event_desc_group.attrs,
1452 				   &event_long_desc_group.attrs);
1453 
1454 	if (r)
1455 		return r;
1456 
1457 	r = perf_pmu_register(&h_24x7_pmu, h_24x7_pmu.name, -1);
1458 	if (r)
1459 		return r;
1460 
1461 	return 0;
1462 }
1463 
1464 device_initcall(hv_24x7_init);
1465