1/*
  2 * Copyright (C) 2011-2017, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
  3 *
  4 * Parts came from evlist.c builtin-{top,stat,record}.c, see those files for further
  5 * copyright notes.
  6 *
  7 * Released under the GPL v2. (and only v2, not any later version)
  8 */
  9
 10#include <sys/mman.h>
 11#include <inttypes.h>
 12#include <asm/bug.h>
 13#include "debug.h"
 14#include "event.h"
 15#include "mmap.h"
 16#include "util.h" /* page_size */
 17
 18size_t perf_mmap__mmap_len(struct perf_mmap *map)
 19{
 20	return map->mask + 1 + page_size;
 21}
 22
 23/* When check_messup is true, 'end' must points to a good entry */
 24static union perf_event *perf_mmap__read(struct perf_mmap *map,
 25					 u64 *startp, u64 end)
 26{
 27	unsigned char *data = map->base + page_size;
 28	union perf_event *event = NULL;
 29	int diff = end - *startp;
 30
 31	if (diff >= (int)sizeof(event->header)) {
 32		size_t size;
 33
 34		event = (union perf_event *)&data[*startp & map->mask];
 35		size = event->header.size;
 36
 37		if (size < sizeof(event->header) || diff < (int)size)
 38			return NULL;
 39
 40		/*
 41		 * Event straddles the mmap boundary -- header should always
 42		 * be inside due to u64 alignment of output.
 43		 */
 44		if ((*startp & map->mask) + size != ((*startp + size) & map->mask)) {
 45			unsigned int offset = *startp;
 46			unsigned int len = min(sizeof(*event), size), cpy;
 47			void *dst = map->event_copy;
 48
 49			do {
 50				cpy = min(map->mask + 1 - (offset & map->mask), len);
 51				memcpy(dst, &data[offset & map->mask], cpy);
 52				offset += cpy;
 53				dst += cpy;
 54				len -= cpy;
 55			} while (len);
 56
 57			event = (union perf_event *)map->event_copy;
 58		}
 59
 60		*startp += size;
 61	}
 62
 63	return event;
 64}
 65
 66/*
 67 * Read event from ring buffer one by one.
 68 * Return one event for each call.
 69 *
 70 * Usage:
 71 * perf_mmap__read_init()
 72 * while(event = perf_mmap__read_event()) {
 73 *	//process the event
 74 *	perf_mmap__consume()
 75 * }
 76 * perf_mmap__read_done()
 77 */
 78union perf_event *perf_mmap__read_event(struct perf_mmap *map)
 79{
 80	union perf_event *event;
 81
 82	/*
 83	 * Check if event was unmapped due to a POLLHUP/POLLERR.
 84	 */
 85	if (!refcount_read(&map->refcnt))
 86		return NULL;
 87
 88	/* non-overwirte doesn't pause the ringbuffer */
 89	if (!map->overwrite)
 90		map->end = perf_mmap__read_head(map);
 91
 92	event = perf_mmap__read(map, &map->start, map->end);
 93
 94	if (!map->overwrite)
 95		map->prev = map->start;
 96
 97	return event;
 98}
 99
100static bool perf_mmap__empty(struct perf_mmap *map)
101{
102	return perf_mmap__read_head(map) == map->prev && !map->auxtrace_mmap.base;
103}
104
105void perf_mmap__get(struct perf_mmap *map)
106{
107	refcount_inc(&map->refcnt);
108}
109
110void perf_mmap__put(struct perf_mmap *map)
111{
112	BUG_ON(map->base && refcount_read(&map->refcnt) == 0);
113
114	if (refcount_dec_and_test(&map->refcnt))
115		perf_mmap__munmap(map);
116}
117
118void perf_mmap__consume(struct perf_mmap *map)
119{
120	if (!map->overwrite) {
121		u64 old = map->prev;
122
123		perf_mmap__write_tail(map, old);
124	}
125
126	if (refcount_read(&map->refcnt) == 1 && perf_mmap__empty(map))
127		perf_mmap__put(map);
128}
129
130int __weak auxtrace_mmap__mmap(struct auxtrace_mmap *mm __maybe_unused,
131			       struct auxtrace_mmap_params *mp __maybe_unused,
132			       void *userpg __maybe_unused,
133			       int fd __maybe_unused)
134{
135	return 0;
136}
137
138void __weak auxtrace_mmap__munmap(struct auxtrace_mmap *mm __maybe_unused)
139{
140}
141
142void __weak auxtrace_mmap_params__init(struct auxtrace_mmap_params *mp __maybe_unused,
143				       off_t auxtrace_offset __maybe_unused,
144				       unsigned int auxtrace_pages __maybe_unused,
145				       bool auxtrace_overwrite __maybe_unused)
146{
147}
148
149void __weak auxtrace_mmap_params__set_idx(struct auxtrace_mmap_params *mp __maybe_unused,
150					  struct perf_evlist *evlist __maybe_unused,
151					  int idx __maybe_unused,
152					  bool per_cpu __maybe_unused)
153{
154}
155
156void perf_mmap__munmap(struct perf_mmap *map)
157{
158	if (map->base != NULL) {
159		munmap(map->base, perf_mmap__mmap_len(map));
160		map->base = NULL;
161		map->fd = -1;
162		refcount_set(&map->refcnt, 0);
163	}
164	auxtrace_mmap__munmap(&map->auxtrace_mmap);
165}
166
167int perf_mmap__mmap(struct perf_mmap *map, struct mmap_params *mp, int fd)
168{
169	/*
170	 * The last one will be done at perf_mmap__consume(), so that we
171	 * make sure we don't prevent tools from consuming every last event in
172	 * the ring buffer.
173	 *
174	 * I.e. we can get the POLLHUP meaning that the fd doesn't exist
175	 * anymore, but the last events for it are still in the ring buffer,
176	 * waiting to be consumed.
177	 *
178	 * Tools can chose to ignore this at their own discretion, but the
179	 * evlist layer can't just drop it when filtering events in
180	 * perf_evlist__filter_pollfd().
181	 */
182	refcount_set(&map->refcnt, 2);
183	map->prev = 0;
184	map->mask = mp->mask;
185	map->base = mmap(NULL, perf_mmap__mmap_len(map), mp->prot,
186			 MAP_SHARED, fd, 0);
187	if (map->base == MAP_FAILED) {
188		pr_debug2("failed to mmap perf event ring buffer, error %d\n",
189			  errno);
190		map->base = NULL;
191		return -1;
192	}
193	map->fd = fd;
194
195	if (auxtrace_mmap__mmap(&map->auxtrace_mmap,
196				&mp->auxtrace_mp, map->base, fd))
197		return -1;
198
199	return 0;
200}
201
202static int overwrite_rb_find_range(void *buf, int mask, u64 *start, u64 *end)
203{
204	struct perf_event_header *pheader;
205	u64 evt_head = *start;
206	int size = mask + 1;
207
208	pr_debug2("%s: buf=%p, start=%"PRIx64"\n", __func__, buf, *start);
209	pheader = (struct perf_event_header *)(buf + (*start & mask));
210	while (true) {
211		if (evt_head - *start >= (unsigned int)size) {
212			pr_debug("Finished reading overwrite ring buffer: rewind\n");
213			if (evt_head - *start > (unsigned int)size)
214				evt_head -= pheader->size;
215			*end = evt_head;
216			return 0;
217		}
218
219		pheader = (struct perf_event_header *)(buf + (evt_head & mask));
220
221		if (pheader->size == 0) {
222			pr_debug("Finished reading overwrite ring buffer: get start\n");
223			*end = evt_head;
224			return 0;
225		}
226
227		evt_head += pheader->size;
228		pr_debug3("move evt_head: %"PRIx64"\n", evt_head);
229	}
230	WARN_ONCE(1, "Shouldn't get here\n");
231	return -1;
232}
233
234/*
235 * Report the start and end of the available data in ringbuffer
236 */
237static int __perf_mmap__read_init(struct perf_mmap *md)
238{
239	u64 head = perf_mmap__read_head(md);
240	u64 old = md->prev;
241	unsigned char *data = md->base + page_size;
242	unsigned long size;
243
244	md->start = md->overwrite ? head : old;
245	md->end = md->overwrite ? old : head;
246
247	if (md->start == md->end)
248		return -EAGAIN;
249
250	size = md->end - md->start;
251	if (size > (unsigned long)(md->mask) + 1) {
252		if (!md->overwrite) {
253			WARN_ONCE(1, "failed to keep up with mmap data. (warn only once)\n");
254
255			md->prev = head;
256			perf_mmap__consume(md);
257			return -EAGAIN;
258		}
259
260		/*
261		 * Backward ring buffer is full. We still have a chance to read
262		 * most of data from it.
263		 */
264		if (overwrite_rb_find_range(data, md->mask, &md->start, &md->end))
265			return -EINVAL;
266	}
267
268	return 0;
269}
270
271int perf_mmap__read_init(struct perf_mmap *map)
272{
273	/*
274	 * Check if event was unmapped due to a POLLHUP/POLLERR.
275	 */
276	if (!refcount_read(&map->refcnt))
277		return -ENOENT;
278
279	return __perf_mmap__read_init(map);
280}
281
282int perf_mmap__push(struct perf_mmap *md, void *to,
283		    int push(void *to, void *buf, size_t size))
284{
285	u64 head = perf_mmap__read_head(md);
286	unsigned char *data = md->base + page_size;
287	unsigned long size;
288	void *buf;
289	int rc = 0;
290
291	rc = perf_mmap__read_init(md);
292	if (rc < 0)
293		return (rc == -EAGAIN) ? 0 : -1;
294
295	size = md->end - md->start;
296
297	if ((md->start & md->mask) + size != (md->end & md->mask)) {
298		buf = &data[md->start & md->mask];
299		size = md->mask + 1 - (md->start & md->mask);
300		md->start += size;
301
302		if (push(to, buf, size) < 0) {
303			rc = -1;
304			goto out;
305		}
306	}
307
308	buf = &data[md->start & md->mask];
309	size = md->end - md->start;
310	md->start += size;
311
312	if (push(to, buf, size) < 0) {
313		rc = -1;
314		goto out;
315	}
316
317	md->prev = head;
318	perf_mmap__consume(md);
319out:
320	return rc;
321}
322
323/*
324 * Mandatory for overwrite mode
325 * The direction of overwrite mode is backward.
326 * The last perf_mmap__read() will set tail to map->prev.
327 * Need to correct the map->prev to head which is the end of next read.
328 */
329void perf_mmap__read_done(struct perf_mmap *map)
330{
331	/*
332	 * Check if event was unmapped due to a POLLHUP/POLLERR.
333	 */
334	if (!refcount_read(&map->refcnt))
335		return;
336
337	map->prev = perf_mmap__read_head(map);
338}