xref: /linux/arch/powerpc/perf/hv-24x7.c (revision 2ba9268dd603d23e17643437b2246acb6844953b)
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 static char *event_name(struct hv_24x7_event_data *ev, int *len)
146 {
147 	*len = be16_to_cpu(ev->event_name_len) - 2;
148 	return (char *)ev->remainder;
149 }
150 
151 static char *event_desc(struct hv_24x7_event_data *ev, int *len)
152 {
153 	unsigned nl = be16_to_cpu(ev->event_name_len);
154 	__be16 *desc_len = (__be16 *)(ev->remainder + nl - 2);
155 	*len = be16_to_cpu(*desc_len) - 2;
156 	return (char *)ev->remainder + nl;
157 }
158 
159 static char *event_long_desc(struct hv_24x7_event_data *ev, int *len)
160 {
161 	unsigned nl = be16_to_cpu(ev->event_name_len);
162 	__be16 *desc_len_ = (__be16 *)(ev->remainder + nl - 2);
163 	unsigned desc_len = be16_to_cpu(*desc_len_);
164 	__be16 *long_desc_len = (__be16 *)(ev->remainder + nl + desc_len - 2);
165 	*len = be16_to_cpu(*long_desc_len) - 2;
166 	return (char *)ev->remainder + nl + desc_len;
167 }
168 
169 static bool event_fixed_portion_is_within(struct hv_24x7_event_data *ev,
170 					  void *end)
171 {
172 	void *start = ev;
173 
174 	return (start + offsetof(struct hv_24x7_event_data, remainder)) < end;
175 }
176 
177 /*
178  * Things we don't check:
179  *  - padding for desc, name, and long/detailed desc is required to be '\0'
180  *    bytes.
181  *
182  *  Return NULL if we pass end,
183  *  Otherwise return the address of the byte just following the event.
184  */
185 static void *event_end(struct hv_24x7_event_data *ev, void *end)
186 {
187 	void *start = ev;
188 	__be16 *dl_, *ldl_;
189 	unsigned dl, ldl;
190 	unsigned nl = be16_to_cpu(ev->event_name_len);
191 
192 	if (nl < 2) {
193 		pr_debug("%s: name length too short: %d", __func__, nl);
194 		return NULL;
195 	}
196 
197 	if (start + nl > end) {
198 		pr_debug("%s: start=%p + nl=%u > end=%p",
199 				__func__, start, nl, end);
200 		return NULL;
201 	}
202 
203 	dl_ = (__be16 *)(ev->remainder + nl - 2);
204 	if (!IS_ALIGNED((uintptr_t)dl_, 2))
205 		pr_warn("desc len not aligned %p", dl_);
206 	dl = be16_to_cpu(*dl_);
207 	if (dl < 2) {
208 		pr_debug("%s: desc len too short: %d", __func__, dl);
209 		return NULL;
210 	}
211 
212 	if (start + nl + dl > end) {
213 		pr_debug("%s: (start=%p + nl=%u + dl=%u)=%p > end=%p",
214 				__func__, start, nl, dl, start + nl + dl, end);
215 		return NULL;
216 	}
217 
218 	ldl_ = (__be16 *)(ev->remainder + nl + dl - 2);
219 	if (!IS_ALIGNED((uintptr_t)ldl_, 2))
220 		pr_warn("long desc len not aligned %p", ldl_);
221 	ldl = be16_to_cpu(*ldl_);
222 	if (ldl < 2) {
223 		pr_debug("%s: long desc len too short (ldl=%u)",
224 				__func__, ldl);
225 		return NULL;
226 	}
227 
228 	if (start + nl + dl + ldl > end) {
229 		pr_debug("%s: start=%p + nl=%u + dl=%u + ldl=%u > end=%p",
230 				__func__, start, nl, dl, ldl, end);
231 		return NULL;
232 	}
233 
234 	return start + nl + dl + ldl;
235 }
236 
237 static unsigned long h_get_24x7_catalog_page_(unsigned long phys_4096,
238 					      unsigned long version,
239 					      unsigned long index)
240 {
241 	pr_devel("h_get_24x7_catalog_page(0x%lx, %lu, %lu)",
242 			phys_4096,
243 			version,
244 			index);
245 	WARN_ON(!IS_ALIGNED(phys_4096, 4096));
246 	return plpar_hcall_norets(H_GET_24X7_CATALOG_PAGE,
247 			phys_4096,
248 			version,
249 			index);
250 }
251 
252 static unsigned long h_get_24x7_catalog_page(char page[],
253 					     u64 version, u32 index)
254 {
255 	return h_get_24x7_catalog_page_(virt_to_phys(page),
256 					version, index);
257 }
258 
259 static unsigned core_domains[] = {
260 	HV_PERF_DOMAIN_PHYS_CORE,
261 	HV_PERF_DOMAIN_VCPU_HOME_CORE,
262 	HV_PERF_DOMAIN_VCPU_HOME_CHIP,
263 	HV_PERF_DOMAIN_VCPU_HOME_NODE,
264 	HV_PERF_DOMAIN_VCPU_REMOTE_NODE,
265 };
266 /* chip event data always yeilds a single event, core yeilds multiple */
267 #define MAX_EVENTS_PER_EVENT_DATA ARRAY_SIZE(core_domains)
268 
269 static char *event_fmt(struct hv_24x7_event_data *event, unsigned domain)
270 {
271 	const char *sindex;
272 	const char *lpar;
273 
274 	if (is_physical_domain(domain)) {
275 		lpar = "0x0";
276 		sindex = "core";
277 	} else {
278 		lpar = "?";
279 		sindex = "vcpu";
280 	}
281 
282 	return kasprintf(GFP_KERNEL,
283 			"domain=0x%x,offset=0x%x,%s=?,lpar=%s",
284 			domain,
285 			be16_to_cpu(event->event_counter_offs) +
286 				be16_to_cpu(event->event_group_record_offs),
287 			sindex,
288 			lpar);
289 }
290 
291 /* Avoid trusting fw to NUL terminate strings */
292 static char *memdup_to_str(char *maybe_str, int max_len, gfp_t gfp)
293 {
294 	return kasprintf(gfp, "%.*s", max_len, maybe_str);
295 }
296 
297 static ssize_t device_show_string(struct device *dev,
298 		struct device_attribute *attr, char *buf)
299 {
300 	struct dev_ext_attribute *d;
301 
302 	d = container_of(attr, struct dev_ext_attribute, attr);
303 	return sprintf(buf, "%s\n", (char *)d->var);
304 }
305 
306 static struct attribute *device_str_attr_create_(char *name, char *str)
307 {
308 	struct dev_ext_attribute *attr = kzalloc(sizeof(*attr), GFP_KERNEL);
309 
310 	if (!attr)
311 		return NULL;
312 
313 	attr->var = str;
314 	attr->attr.attr.name = name;
315 	attr->attr.attr.mode = 0444;
316 	attr->attr.show = device_show_string;
317 	return &attr->attr.attr;
318 }
319 
320 static struct attribute *device_str_attr_create(char *name, int name_max,
321 						int name_nonce,
322 						char *str, size_t str_max)
323 {
324 	char *n;
325 	char *s = memdup_to_str(str, str_max, GFP_KERNEL);
326 	struct attribute *a;
327 
328 	if (!s)
329 		return NULL;
330 
331 	if (!name_nonce)
332 		n = kasprintf(GFP_KERNEL, "%.*s", name_max, name);
333 	else
334 		n = kasprintf(GFP_KERNEL, "%.*s__%d", name_max, name,
335 					name_nonce);
336 	if (!n)
337 		goto out_s;
338 
339 	a = device_str_attr_create_(n, s);
340 	if (!a)
341 		goto out_n;
342 
343 	return a;
344 out_n:
345 	kfree(n);
346 out_s:
347 	kfree(s);
348 	return NULL;
349 }
350 
351 static void device_str_attr_destroy(struct attribute *attr)
352 {
353 	struct dev_ext_attribute *d;
354 
355 	d = container_of(attr, struct dev_ext_attribute, attr.attr);
356 	kfree(d->var);
357 	kfree(d->attr.attr.name);
358 	kfree(d);
359 }
360 
361 static struct attribute *event_to_attr(unsigned ix,
362 				       struct hv_24x7_event_data *event,
363 				       unsigned domain,
364 				       int nonce)
365 {
366 	int event_name_len;
367 	char *ev_name, *a_ev_name, *val;
368 	const char *ev_suffix;
369 	struct attribute *attr;
370 
371 	if (!domain_is_valid(domain)) {
372 		pr_warn("catalog event %u has invalid domain %u\n",
373 				ix, domain);
374 		return NULL;
375 	}
376 
377 	val = event_fmt(event, domain);
378 	if (!val)
379 		return NULL;
380 
381 	ev_suffix = event_domain_suffix(domain);
382 	ev_name = event_name(event, &event_name_len);
383 	if (!nonce)
384 		a_ev_name = kasprintf(GFP_KERNEL, "%.*s%s",
385 				(int)event_name_len, ev_name, ev_suffix);
386 	else
387 		a_ev_name = kasprintf(GFP_KERNEL, "%.*s%s__%d",
388 				(int)event_name_len, ev_name, ev_suffix, nonce);
389 
390 
391 	if (!a_ev_name)
392 		goto out_val;
393 
394 	attr = device_str_attr_create_(a_ev_name, val);
395 	if (!attr)
396 		goto out_name;
397 
398 	return attr;
399 out_name:
400 	kfree(a_ev_name);
401 out_val:
402 	kfree(val);
403 	return NULL;
404 }
405 
406 static struct attribute *event_to_desc_attr(struct hv_24x7_event_data *event,
407 				int nonce)
408 {
409 	int nl, dl;
410 	char *name = event_name(event, &nl);
411 	char *desc = event_desc(event, &dl);
412 
413 	/* If there isn't a description, don't create the sysfs file */
414 	if (!dl)
415 		return NULL;
416 
417 	return device_str_attr_create(name, nl, nonce, desc, dl);
418 }
419 
420 static struct attribute *
421 event_to_long_desc_attr(struct hv_24x7_event_data *event, int nonce)
422 {
423 	int nl, dl;
424 	char *name = event_name(event, &nl);
425 	char *desc = event_long_desc(event, &dl);
426 
427 	/* If there isn't a description, don't create the sysfs file */
428 	if (!dl)
429 		return NULL;
430 
431 	return device_str_attr_create(name, nl, nonce, desc, dl);
432 }
433 
434 static ssize_t event_data_to_attrs(unsigned ix, struct attribute **attrs,
435 		struct hv_24x7_event_data *event, int nonce)
436 {
437 	unsigned i;
438 
439 	switch (event->domain) {
440 	case HV_PERF_DOMAIN_PHYS_CHIP:
441 		*attrs = event_to_attr(ix, event, event->domain, nonce);
442 		return 1;
443 	case HV_PERF_DOMAIN_PHYS_CORE:
444 		for (i = 0; i < ARRAY_SIZE(core_domains); i++) {
445 			attrs[i] = event_to_attr(ix, event, core_domains[i],
446 						nonce);
447 			if (!attrs[i]) {
448 				pr_warn("catalog event %u: individual attr %u "
449 					"creation failure\n", ix, i);
450 				for (; i; i--)
451 					device_str_attr_destroy(attrs[i - 1]);
452 				return -1;
453 			}
454 		}
455 		return i;
456 	default:
457 		pr_warn("catalog event %u: domain %u is not allowed in the "
458 				"catalog\n", ix, event->domain);
459 		return -1;
460 	}
461 }
462 
463 static size_t event_to_attr_ct(struct hv_24x7_event_data *event)
464 {
465 	switch (event->domain) {
466 	case HV_PERF_DOMAIN_PHYS_CHIP:
467 		return 1;
468 	case HV_PERF_DOMAIN_PHYS_CORE:
469 		return ARRAY_SIZE(core_domains);
470 	default:
471 		return 0;
472 	}
473 }
474 
475 static unsigned long vmalloc_to_phys(void *v)
476 {
477 	struct page *p = vmalloc_to_page(v);
478 
479 	BUG_ON(!p);
480 	return page_to_phys(p) + offset_in_page(v);
481 }
482 
483 /* */
484 struct event_uniq {
485 	struct rb_node node;
486 	const char *name;
487 	int nl;
488 	unsigned ct;
489 	unsigned domain;
490 };
491 
492 static int memord(const void *d1, size_t s1, const void *d2, size_t s2)
493 {
494 	if (s1 < s2)
495 		return 1;
496 	if (s2 > s1)
497 		return -1;
498 
499 	return memcmp(d1, d2, s1);
500 }
501 
502 static int ev_uniq_ord(const void *v1, size_t s1, unsigned d1, const void *v2,
503 					size_t s2, unsigned d2)
504 {
505 	int r = memord(v1, s1, v2, s2);
506 
507 	if (r)
508 		return r;
509 	if (d1 > d2)
510 		return 1;
511 	if (d2 > d1)
512 		return -1;
513 	return 0;
514 }
515 
516 static int event_uniq_add(struct rb_root *root, const char *name, int nl,
517 				unsigned domain)
518 {
519 	struct rb_node **new = &(root->rb_node), *parent = NULL;
520 	struct event_uniq *data;
521 
522 	/* Figure out where to put new node */
523 	while (*new) {
524 		struct event_uniq *it;
525 		int result;
526 
527 		it = container_of(*new, struct event_uniq, node);
528 		result = ev_uniq_ord(name, nl, domain, it->name, it->nl,
529 					it->domain);
530 
531 		parent = *new;
532 		if (result < 0)
533 			new = &((*new)->rb_left);
534 		else if (result > 0)
535 			new = &((*new)->rb_right);
536 		else {
537 			it->ct++;
538 			pr_info("found a duplicate event %.*s, ct=%u\n", nl,
539 						name, it->ct);
540 			return it->ct;
541 		}
542 	}
543 
544 	data = kmalloc(sizeof(*data), GFP_KERNEL);
545 	if (!data)
546 		return -ENOMEM;
547 
548 	*data = (struct event_uniq) {
549 		.name = name,
550 		.nl = nl,
551 		.ct = 0,
552 		.domain = domain,
553 	};
554 
555 	/* Add new node and rebalance tree. */
556 	rb_link_node(&data->node, parent, new);
557 	rb_insert_color(&data->node, root);
558 
559 	/* data->ct */
560 	return 0;
561 }
562 
563 static void event_uniq_destroy(struct rb_root *root)
564 {
565 	/*
566 	 * the strings we point to are in the giant block of memory filled by
567 	 * the catalog, and are freed separately.
568 	 */
569 	struct event_uniq *pos, *n;
570 
571 	rbtree_postorder_for_each_entry_safe(pos, n, root, node)
572 		kfree(pos);
573 }
574 
575 
576 /*
577  * ensure the event structure's sizes are self consistent and don't cause us to
578  * read outside of the event
579  *
580  * On success, return the event length in bytes.
581  * Otherwise, return -1 (and print as appropriate).
582  */
583 static ssize_t catalog_event_len_validate(struct hv_24x7_event_data *event,
584 					  size_t event_idx,
585 					  size_t event_data_bytes,
586 					  size_t event_entry_count,
587 					  size_t offset, void *end)
588 {
589 	ssize_t ev_len;
590 	void *ev_end, *calc_ev_end;
591 
592 	if (offset >= event_data_bytes)
593 		return -1;
594 
595 	if (event_idx >= event_entry_count) {
596 		pr_devel("catalog event data has %zu bytes of padding after last event\n",
597 				event_data_bytes - offset);
598 		return -1;
599 	}
600 
601 	if (!event_fixed_portion_is_within(event, end)) {
602 		pr_warn("event %zu fixed portion is not within range\n",
603 				event_idx);
604 		return -1;
605 	}
606 
607 	ev_len = be16_to_cpu(event->length);
608 
609 	if (ev_len % 16)
610 		pr_info("event %zu has length %zu not divisible by 16: event=%pK\n",
611 				event_idx, ev_len, event);
612 
613 	ev_end = (__u8 *)event + ev_len;
614 	if (ev_end > end) {
615 		pr_warn("event %zu has .length=%zu, ends after buffer end: ev_end=%pK > end=%pK, offset=%zu\n",
616 				event_idx, ev_len, ev_end, end,
617 				offset);
618 		return -1;
619 	}
620 
621 	calc_ev_end = event_end(event, end);
622 	if (!calc_ev_end) {
623 		pr_warn("event %zu has a calculated length which exceeds buffer length %zu: event=%pK end=%pK, offset=%zu\n",
624 			event_idx, event_data_bytes, event, end,
625 			offset);
626 		return -1;
627 	}
628 
629 	if (calc_ev_end > ev_end) {
630 		pr_warn("event %zu exceeds it's own length: event=%pK, end=%pK, offset=%zu, calc_ev_end=%pK\n",
631 			event_idx, event, ev_end, offset, calc_ev_end);
632 		return -1;
633 	}
634 
635 	return ev_len;
636 }
637 
638 #define MAX_4K (SIZE_MAX / 4096)
639 
640 static void create_events_from_catalog(struct attribute ***events_,
641 		struct attribute ***event_descs_,
642 		struct attribute ***event_long_descs_)
643 {
644 	unsigned long hret;
645 	size_t catalog_len, catalog_page_len, event_entry_count,
646 	       event_data_len, event_data_offs,
647 	       event_data_bytes, junk_events, event_idx, event_attr_ct, i,
648 	       attr_max, event_idx_last, desc_ct, long_desc_ct;
649 	ssize_t ct, ev_len;
650 	uint32_t catalog_version_num;
651 	struct attribute **events, **event_descs, **event_long_descs;
652 	struct hv_24x7_catalog_page_0 *page_0 =
653 		kmem_cache_alloc(hv_page_cache, GFP_KERNEL);
654 	void *page = page_0;
655 	void *event_data, *end;
656 	struct hv_24x7_event_data *event;
657 	struct rb_root ev_uniq = RB_ROOT;
658 
659 	if (!page)
660 		goto e_out;
661 
662 	hret = h_get_24x7_catalog_page(page, 0, 0);
663 	if (hret)
664 		goto e_free;
665 
666 	catalog_version_num = be64_to_cpu(page_0->version);
667 	catalog_page_len = be32_to_cpu(page_0->length);
668 
669 	if (MAX_4K < catalog_page_len) {
670 		pr_err("invalid page count: %zu\n", catalog_page_len);
671 		goto e_free;
672 	}
673 
674 	catalog_len = catalog_page_len * 4096;
675 
676 	event_entry_count = be16_to_cpu(page_0->event_entry_count);
677 	event_data_offs   = be16_to_cpu(page_0->event_data_offs);
678 	event_data_len    = be16_to_cpu(page_0->event_data_len);
679 
680 	pr_devel("cv %zu cl %zu eec %zu edo %zu edl %zu\n",
681 			(size_t)catalog_version_num, catalog_len,
682 			event_entry_count, event_data_offs, event_data_len);
683 
684 	if ((MAX_4K < event_data_len)
685 			|| (MAX_4K < event_data_offs)
686 			|| (MAX_4K - event_data_offs < event_data_len)) {
687 		pr_err("invalid event data offs %zu and/or len %zu\n",
688 				event_data_offs, event_data_len);
689 		goto e_free;
690 	}
691 
692 	if ((event_data_offs + event_data_len) > catalog_page_len) {
693 		pr_err("event data %zu-%zu does not fit inside catalog 0-%zu\n",
694 				event_data_offs,
695 				event_data_offs + event_data_len,
696 				catalog_page_len);
697 		goto e_free;
698 	}
699 
700 	if (SIZE_MAX / MAX_EVENTS_PER_EVENT_DATA - 1 < event_entry_count) {
701 		pr_err("event_entry_count %zu is invalid\n",
702 				event_entry_count);
703 		goto e_free;
704 	}
705 
706 	event_data_bytes = event_data_len * 4096;
707 
708 	/*
709 	 * event data can span several pages, events can cross between these
710 	 * pages. Use vmalloc to make this easier.
711 	 */
712 	event_data = vmalloc(event_data_bytes);
713 	if (!event_data) {
714 		pr_err("could not allocate event data\n");
715 		goto e_free;
716 	}
717 
718 	end = event_data + event_data_bytes;
719 
720 	/*
721 	 * using vmalloc_to_phys() like this only works if PAGE_SIZE is
722 	 * divisible by 4096
723 	 */
724 	BUILD_BUG_ON(PAGE_SIZE % 4096);
725 
726 	for (i = 0; i < event_data_len; i++) {
727 		hret = h_get_24x7_catalog_page_(
728 				vmalloc_to_phys(event_data + i * 4096),
729 				catalog_version_num,
730 				i + event_data_offs);
731 		if (hret) {
732 			pr_err("failed to get event data in page %zu\n",
733 					i + event_data_offs);
734 			goto e_event_data;
735 		}
736 	}
737 
738 	/*
739 	 * scan the catalog to determine the number of attributes we need, and
740 	 * verify it at the same time.
741 	 */
742 	for (junk_events = 0, event = event_data, event_idx = 0, attr_max = 0;
743 	     ;
744 	     event_idx++, event = (void *)event + ev_len) {
745 		size_t offset = (void *)event - (void *)event_data;
746 		char *name;
747 		int nl;
748 
749 		ev_len = catalog_event_len_validate(event, event_idx,
750 						    event_data_bytes,
751 						    event_entry_count,
752 						    offset, end);
753 		if (ev_len < 0)
754 			break;
755 
756 		name = event_name(event, &nl);
757 
758 		if (event->event_group_record_len == 0) {
759 			pr_devel("invalid event %zu (%.*s): group_record_len == 0, skipping\n",
760 					event_idx, nl, name);
761 			junk_events++;
762 			continue;
763 		}
764 
765 		if (!catalog_entry_domain_is_valid(event->domain)) {
766 			pr_info("event %zu (%.*s) has invalid domain %d\n",
767 					event_idx, nl, name, event->domain);
768 			junk_events++;
769 			continue;
770 		}
771 
772 		attr_max += event_to_attr_ct(event);
773 	}
774 
775 	event_idx_last = event_idx;
776 	if (event_idx_last != event_entry_count)
777 		pr_warn("event buffer ended before listed # of events were parsed (got %zu, wanted %zu, junk %zu)\n",
778 				event_idx_last, event_entry_count, junk_events);
779 
780 	events = kmalloc_array(attr_max + 1, sizeof(*events), GFP_KERNEL);
781 	if (!events)
782 		goto e_event_data;
783 
784 	event_descs = kmalloc_array(event_idx + 1, sizeof(*event_descs),
785 				GFP_KERNEL);
786 	if (!event_descs)
787 		goto e_event_attrs;
788 
789 	event_long_descs = kmalloc_array(event_idx + 1,
790 			sizeof(*event_long_descs), GFP_KERNEL);
791 	if (!event_long_descs)
792 		goto e_event_descs;
793 
794 	/* Iterate over the catalog filling in the attribute vector */
795 	for (junk_events = 0, event_attr_ct = 0, desc_ct = 0, long_desc_ct = 0,
796 				event = event_data, event_idx = 0;
797 			event_idx < event_idx_last;
798 			event_idx++, ev_len = be16_to_cpu(event->length),
799 				event = (void *)event + ev_len) {
800 		char *name;
801 		int nl;
802 		int nonce;
803 		/*
804 		 * these are the only "bad" events that are intermixed and that
805 		 * we can ignore without issue. make sure to skip them here
806 		 */
807 		if (event->event_group_record_len == 0)
808 			continue;
809 		if (!catalog_entry_domain_is_valid(event->domain))
810 			continue;
811 
812 		name  = event_name(event, &nl);
813 		nonce = event_uniq_add(&ev_uniq, name, nl, event->domain);
814 		ct    = event_data_to_attrs(event_idx, events + event_attr_ct,
815 					    event, nonce);
816 		if (ct <= 0) {
817 			pr_warn("event %zu (%.*s) creation failure, skipping\n",
818 				event_idx, nl, name);
819 			junk_events++;
820 		} else {
821 			event_attr_ct += ct;
822 			event_descs[desc_ct] = event_to_desc_attr(event, nonce);
823 			if (event_descs[desc_ct])
824 				desc_ct++;
825 			event_long_descs[long_desc_ct] =
826 					event_to_long_desc_attr(event, nonce);
827 			if (event_long_descs[long_desc_ct])
828 				long_desc_ct++;
829 		}
830 	}
831 
832 	pr_info("read %zu catalog entries, created %zu event attrs (%zu failures), %zu descs\n",
833 			event_idx, event_attr_ct, junk_events, desc_ct);
834 
835 	events[event_attr_ct] = NULL;
836 	event_descs[desc_ct] = NULL;
837 	event_long_descs[long_desc_ct] = NULL;
838 
839 	event_uniq_destroy(&ev_uniq);
840 	vfree(event_data);
841 	kmem_cache_free(hv_page_cache, page);
842 
843 	*events_ = events;
844 	*event_descs_ = event_descs;
845 	*event_long_descs_ = event_long_descs;
846 	return;
847 
848 e_event_descs:
849 	kfree(event_descs);
850 e_event_attrs:
851 	kfree(events);
852 e_event_data:
853 	vfree(event_data);
854 e_free:
855 	kmem_cache_free(hv_page_cache, page);
856 e_out:
857 	*events_ = NULL;
858 	*event_descs_ = NULL;
859 	*event_long_descs_ = NULL;
860 }
861 
862 static ssize_t catalog_read(struct file *filp, struct kobject *kobj,
863 			    struct bin_attribute *bin_attr, char *buf,
864 			    loff_t offset, size_t count)
865 {
866 	unsigned long hret;
867 	ssize_t ret = 0;
868 	size_t catalog_len = 0, catalog_page_len = 0;
869 	loff_t page_offset = 0;
870 	loff_t offset_in_page;
871 	size_t copy_len;
872 	uint64_t catalog_version_num = 0;
873 	void *page = kmem_cache_alloc(hv_page_cache, GFP_USER);
874 	struct hv_24x7_catalog_page_0 *page_0 = page;
875 	if (!page)
876 		return -ENOMEM;
877 
878 	hret = h_get_24x7_catalog_page(page, 0, 0);
879 	if (hret) {
880 		ret = -EIO;
881 		goto e_free;
882 	}
883 
884 	catalog_version_num = be64_to_cpu(page_0->version);
885 	catalog_page_len = be32_to_cpu(page_0->length);
886 	catalog_len = catalog_page_len * 4096;
887 
888 	page_offset = offset / 4096;
889 	offset_in_page = offset % 4096;
890 
891 	if (page_offset >= catalog_page_len)
892 		goto e_free;
893 
894 	if (page_offset != 0) {
895 		hret = h_get_24x7_catalog_page(page, catalog_version_num,
896 					       page_offset);
897 		if (hret) {
898 			ret = -EIO;
899 			goto e_free;
900 		}
901 	}
902 
903 	copy_len = 4096 - offset_in_page;
904 	if (copy_len > count)
905 		copy_len = count;
906 
907 	memcpy(buf, page+offset_in_page, copy_len);
908 	ret = copy_len;
909 
910 e_free:
911 	if (hret)
912 		pr_err("h_get_24x7_catalog_page(ver=%lld, page=%lld) failed:"
913 		       " rc=%ld\n",
914 		       catalog_version_num, page_offset, hret);
915 	kmem_cache_free(hv_page_cache, page);
916 
917 	pr_devel("catalog_read: offset=%lld(%lld) count=%zu "
918 			"catalog_len=%zu(%zu) => %zd\n", offset, page_offset,
919 			count, catalog_len, catalog_page_len, ret);
920 
921 	return ret;
922 }
923 
924 #define PAGE_0_ATTR(_name, _fmt, _expr)				\
925 static ssize_t _name##_show(struct device *dev,			\
926 			    struct device_attribute *dev_attr,	\
927 			    char *buf)				\
928 {								\
929 	unsigned long hret;					\
930 	ssize_t ret = 0;					\
931 	void *page = kmem_cache_alloc(hv_page_cache, GFP_USER);	\
932 	struct hv_24x7_catalog_page_0 *page_0 = page;		\
933 	if (!page)						\
934 		return -ENOMEM;					\
935 	hret = h_get_24x7_catalog_page(page, 0, 0);		\
936 	if (hret) {						\
937 		ret = -EIO;					\
938 		goto e_free;					\
939 	}							\
940 	ret = sprintf(buf, _fmt, _expr);			\
941 e_free:								\
942 	kmem_cache_free(hv_page_cache, page);			\
943 	return ret;						\
944 }								\
945 static DEVICE_ATTR_RO(_name)
946 
947 PAGE_0_ATTR(catalog_version, "%lld\n",
948 		(unsigned long long)be64_to_cpu(page_0->version));
949 PAGE_0_ATTR(catalog_len, "%lld\n",
950 		(unsigned long long)be32_to_cpu(page_0->length) * 4096);
951 static BIN_ATTR_RO(catalog, 0/* real length varies */);
952 
953 static struct bin_attribute *if_bin_attrs[] = {
954 	&bin_attr_catalog,
955 	NULL,
956 };
957 
958 static struct attribute *if_attrs[] = {
959 	&dev_attr_catalog_len.attr,
960 	&dev_attr_catalog_version.attr,
961 	NULL,
962 };
963 
964 static struct attribute_group if_group = {
965 	.name = "interface",
966 	.bin_attrs = if_bin_attrs,
967 	.attrs = if_attrs,
968 };
969 
970 static const struct attribute_group *attr_groups[] = {
971 	&format_group,
972 	&event_group,
973 	&event_desc_group,
974 	&event_long_desc_group,
975 	&if_group,
976 	NULL,
977 };
978 
979 DEFINE_PER_CPU(char, hv_24x7_reqb[4096]) __aligned(4096);
980 DEFINE_PER_CPU(char, hv_24x7_resb[4096]) __aligned(4096);
981 
982 static unsigned long single_24x7_request(u8 domain, u32 offset, u16 ix,
983 					 u16 lpar, u64 *res,
984 					 bool success_expected)
985 {
986 	unsigned long ret;
987 
988 	/*
989 	 * request_buffer and result_buffer are not required to be 4k aligned,
990 	 * but are not allowed to cross any 4k boundary. Aligning them to 4k is
991 	 * the simplest way to ensure that.
992 	 */
993 	struct reqb {
994 		struct hv_24x7_request_buffer buf;
995 		struct hv_24x7_request req;
996 	} __packed *request_buffer;
997 
998 	struct {
999 		struct hv_24x7_data_result_buffer buf;
1000 		struct hv_24x7_result res;
1001 		struct hv_24x7_result_element elem;
1002 		__be64 result;
1003 	} __packed *result_buffer;
1004 
1005 	BUILD_BUG_ON(sizeof(*request_buffer) > 4096);
1006 	BUILD_BUG_ON(sizeof(*result_buffer) > 4096);
1007 
1008 	request_buffer = (void *)get_cpu_var(hv_24x7_reqb);
1009 	result_buffer = (void *)get_cpu_var(hv_24x7_resb);
1010 
1011 	memset(request_buffer, 0, 4096);
1012 	memset(result_buffer, 0, 4096);
1013 
1014 	*request_buffer = (struct reqb) {
1015 		.buf = {
1016 			.interface_version = HV_24X7_IF_VERSION_CURRENT,
1017 			.num_requests = 1,
1018 		},
1019 		.req = {
1020 			.performance_domain = domain,
1021 			.data_size = cpu_to_be16(8),
1022 			.data_offset = cpu_to_be32(offset),
1023 			.starting_lpar_ix = cpu_to_be16(lpar),
1024 			.max_num_lpars = cpu_to_be16(1),
1025 			.starting_ix = cpu_to_be16(ix),
1026 			.max_ix = cpu_to_be16(1),
1027 		}
1028 	};
1029 
1030 	ret = plpar_hcall_norets(H_GET_24X7_DATA,
1031 			virt_to_phys(request_buffer), sizeof(*request_buffer),
1032 			virt_to_phys(result_buffer),  sizeof(*result_buffer));
1033 
1034 	if (ret) {
1035 		if (success_expected)
1036 			pr_err_ratelimited("hcall failed: %d %#x %#x %d => "
1037 				"0x%lx (%ld) detail=0x%x failing ix=%x\n",
1038 				domain, offset, ix, lpar, ret, ret,
1039 				result_buffer->buf.detailed_rc,
1040 				result_buffer->buf.failing_request_ix);
1041 		goto out;
1042 	}
1043 
1044 	*res = be64_to_cpu(result_buffer->result);
1045 
1046 out:
1047 	return ret;
1048 }
1049 
1050 static unsigned long event_24x7_request(struct perf_event *event, u64 *res,
1051 		bool success_expected)
1052 {
1053 	u16 idx;
1054 	unsigned domain = event_get_domain(event);
1055 
1056 	if (is_physical_domain(domain))
1057 		idx = event_get_core(event);
1058 	else
1059 		idx = event_get_vcpu(event);
1060 
1061 	return single_24x7_request(event_get_domain(event),
1062 				event_get_offset(event),
1063 				idx,
1064 				event_get_lpar(event),
1065 				res,
1066 				success_expected);
1067 }
1068 
1069 static int h_24x7_event_init(struct perf_event *event)
1070 {
1071 	struct hv_perf_caps caps;
1072 	unsigned domain;
1073 	unsigned long hret;
1074 	u64 ct;
1075 
1076 	/* Not our event */
1077 	if (event->attr.type != event->pmu->type)
1078 		return -ENOENT;
1079 
1080 	/* Unused areas must be 0 */
1081 	if (event_get_reserved1(event) ||
1082 	    event_get_reserved2(event) ||
1083 	    event_get_reserved3(event)) {
1084 		pr_devel("reserved set when forbidden 0x%llx(0x%llx) 0x%llx(0x%llx) 0x%llx(0x%llx)\n",
1085 				event->attr.config,
1086 				event_get_reserved1(event),
1087 				event->attr.config1,
1088 				event_get_reserved2(event),
1089 				event->attr.config2,
1090 				event_get_reserved3(event));
1091 		return -EINVAL;
1092 	}
1093 
1094 	/* unsupported modes and filters */
1095 	if (event->attr.exclude_user   ||
1096 	    event->attr.exclude_kernel ||
1097 	    event->attr.exclude_hv     ||
1098 	    event->attr.exclude_idle   ||
1099 	    event->attr.exclude_host   ||
1100 	    event->attr.exclude_guest)
1101 		return -EINVAL;
1102 
1103 	/* no branch sampling */
1104 	if (has_branch_stack(event))
1105 		return -EOPNOTSUPP;
1106 
1107 	/* offset must be 8 byte aligned */
1108 	if (event_get_offset(event) % 8) {
1109 		pr_devel("bad alignment\n");
1110 		return -EINVAL;
1111 	}
1112 
1113 	/* Domains above 6 are invalid */
1114 	domain = event_get_domain(event);
1115 	if (domain > 6) {
1116 		pr_devel("invalid domain %d\n", domain);
1117 		return -EINVAL;
1118 	}
1119 
1120 	hret = hv_perf_caps_get(&caps);
1121 	if (hret) {
1122 		pr_devel("could not get capabilities: rc=%ld\n", hret);
1123 		return -EIO;
1124 	}
1125 
1126 	/* Physical domains & other lpars require extra capabilities */
1127 	if (!caps.collect_privileged && (is_physical_domain(domain) ||
1128 		(event_get_lpar(event) != event_get_lpar_max()))) {
1129 		pr_devel("hv permisions disallow: is_physical_domain:%d, lpar=0x%llx\n",
1130 				is_physical_domain(domain),
1131 				event_get_lpar(event));
1132 		return -EACCES;
1133 	}
1134 
1135 	/* see if the event complains */
1136 	if (event_24x7_request(event, &ct, false)) {
1137 		pr_devel("test hcall failed\n");
1138 		return -EIO;
1139 	}
1140 
1141 	return 0;
1142 }
1143 
1144 static u64 h_24x7_get_value(struct perf_event *event)
1145 {
1146 	unsigned long ret;
1147 	u64 ct;
1148 	ret = event_24x7_request(event, &ct, true);
1149 	if (ret)
1150 		/* We checked this in event init, shouldn't fail here... */
1151 		return 0;
1152 
1153 	return ct;
1154 }
1155 
1156 static void h_24x7_event_update(struct perf_event *event)
1157 {
1158 	s64 prev;
1159 	u64 now;
1160 	now = h_24x7_get_value(event);
1161 	prev = local64_xchg(&event->hw.prev_count, now);
1162 	local64_add(now - prev, &event->count);
1163 }
1164 
1165 static void h_24x7_event_start(struct perf_event *event, int flags)
1166 {
1167 	if (flags & PERF_EF_RELOAD)
1168 		local64_set(&event->hw.prev_count, h_24x7_get_value(event));
1169 }
1170 
1171 static void h_24x7_event_stop(struct perf_event *event, int flags)
1172 {
1173 	h_24x7_event_update(event);
1174 }
1175 
1176 static int h_24x7_event_add(struct perf_event *event, int flags)
1177 {
1178 	if (flags & PERF_EF_START)
1179 		h_24x7_event_start(event, flags);
1180 
1181 	return 0;
1182 }
1183 
1184 static struct pmu h_24x7_pmu = {
1185 	.task_ctx_nr = perf_invalid_context,
1186 
1187 	.name = "hv_24x7",
1188 	.attr_groups = attr_groups,
1189 	.event_init  = h_24x7_event_init,
1190 	.add         = h_24x7_event_add,
1191 	.del         = h_24x7_event_stop,
1192 	.start       = h_24x7_event_start,
1193 	.stop        = h_24x7_event_stop,
1194 	.read        = h_24x7_event_update,
1195 };
1196 
1197 static int hv_24x7_init(void)
1198 {
1199 	int r;
1200 	unsigned long hret;
1201 	struct hv_perf_caps caps;
1202 
1203 	if (!firmware_has_feature(FW_FEATURE_LPAR)) {
1204 		pr_debug("not a virtualized system, not enabling\n");
1205 		return -ENODEV;
1206 	}
1207 
1208 	hret = hv_perf_caps_get(&caps);
1209 	if (hret) {
1210 		pr_debug("could not obtain capabilities, not enabling, rc=%ld\n",
1211 				hret);
1212 		return -ENODEV;
1213 	}
1214 
1215 	hv_page_cache = kmem_cache_create("hv-page-4096", 4096, 4096, 0, NULL);
1216 	if (!hv_page_cache)
1217 		return -ENOMEM;
1218 
1219 	/* sampling not supported */
1220 	h_24x7_pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
1221 
1222 	create_events_from_catalog(&event_group.attrs,
1223 				   &event_desc_group.attrs,
1224 				   &event_long_desc_group.attrs);
1225 
1226 	r = perf_pmu_register(&h_24x7_pmu, h_24x7_pmu.name, -1);
1227 	if (r)
1228 		return r;
1229 
1230 	return 0;
1231 }
1232 
1233 device_initcall(hv_24x7_init);
1234