dedupe.c revision d28d27453837d4666bc0637020e0ead81f30db3e
1/*
2 * Small tool to check for dedupable blocks in a file or device. Basically
3 * just scans the filename for extents of the given size, checksums them,
4 * and orders them up.
5 */
6#include <stdio.h>
7#include <stdio.h>
8#include <unistd.h>
9#include <inttypes.h>
10#include <assert.h>
11#include <sys/types.h>
12#include <sys/stat.h>
13#include <sys/ioctl.h>
14#include <linux/fs.h>
15#include <fcntl.h>
16#include <string.h>
17
18#include "../lib/rbtree.h"
19#include "../flist.h"
20#include "../log.h"
21#include "../mutex.h"
22#include "../smalloc.h"
23#include "../minmax.h"
24#include "../crc/md5.h"
25#include "../memalign.h"
26#include "../os/os.h"
27
28FILE *f_err;
29struct timeval *fio_tv = NULL;
30unsigned int fio_debug = 0;
31
32void __dprint(int type, const char *str, ...)
33{
34}
35
36struct worker_thread {
37	pthread_t thread;
38
39	int fd;
40	uint64_t cur_offset;
41	uint64_t size;
42
43	unsigned long items;
44	int err;
45};
46
47struct extent {
48	struct flist_head list;
49	uint64_t offset;
50};
51
52struct chunk {
53	struct rb_node rb_node;
54	struct flist_head extent_list;
55	uint64_t count;
56	uint32_t hash[MD5_HASH_WORDS];
57};
58
59struct item {
60	uint64_t offset;
61	uint32_t hash[MD5_HASH_WORDS];
62};
63
64static struct rb_root rb_root;
65static struct fio_mutex *rb_lock;
66
67static unsigned int blocksize = 4096;
68static unsigned int num_threads;
69static unsigned int chunk_size = 1048576;
70static unsigned int dump_output;
71static unsigned int odirect;
72static unsigned int collision_check;
73
74static uint64_t total_size;
75static uint64_t cur_offset;
76static struct fio_mutex *size_lock;
77
78static int dev_fd;
79
80static uint64_t get_size(int fd, struct stat *sb)
81{
82	uint64_t ret;
83
84	if (S_ISBLK(sb->st_mode)) {
85		if (ioctl(fd, BLKGETSIZE64, &ret) < 0) {
86			perror("ioctl");
87			return 0;
88		}
89	} else
90		ret = sb->st_size;
91
92	return (ret & ~((uint64_t)blocksize - 1));
93}
94
95static int get_work(uint64_t *offset, uint64_t *size)
96{
97	uint64_t this_chunk;
98	int ret = 1;
99
100	fio_mutex_down(size_lock);
101
102	if (cur_offset < total_size) {
103		*offset = cur_offset;
104		this_chunk = min((uint64_t)chunk_size, total_size - cur_offset);
105		*size = this_chunk;
106		cur_offset += this_chunk;
107		ret = 0;
108	}
109
110	fio_mutex_up(size_lock);
111	return ret;
112}
113
114static int read_block(int fd, void *buf, off_t offset)
115{
116	ssize_t ret;
117
118	ret = pread(fd, buf, blocksize, offset);
119	if (ret < 0) {
120		perror("pread");
121		return 1;
122	} else if (!ret)
123		return 1;
124	else if (ret != blocksize) {
125		log_err("dedupe: short read on block\n");
126		return 1;
127	}
128
129	return 0;
130}
131
132static void add_item(struct chunk *c, struct item *i)
133{
134	struct extent *e;
135
136	e = malloc(sizeof(*e));
137	e->offset = i->offset;
138	flist_add_tail(&e->list, &c->extent_list);
139	c->count++;
140}
141
142static int col_check(struct chunk *c, struct item *i)
143{
144	struct extent *e;
145	char *cbuf, *ibuf;
146	int ret = 1;
147
148	cbuf = fio_memalign(blocksize, blocksize);
149	ibuf = fio_memalign(blocksize, blocksize);
150
151	e = flist_entry(c->extent_list.next, struct extent, list);
152	if (read_block(dev_fd, cbuf, e->offset))
153		goto out;
154
155	if (read_block(dev_fd, ibuf, i->offset))
156		goto out;
157
158	ret = memcmp(ibuf, cbuf, blocksize);
159out:
160	fio_memfree(cbuf, blocksize);
161	fio_memfree(ibuf, blocksize);
162	return ret;
163}
164
165static void insert_chunk(struct item *i)
166{
167	struct rb_node **p, *parent;
168	struct chunk *c;
169	int diff;
170
171	p = &rb_root.rb_node;
172	parent = NULL;
173	while (*p) {
174		parent = *p;
175
176		c = rb_entry(parent, struct chunk, rb_node);
177		diff = memcmp(i->hash, c->hash, sizeof(i->hash));
178		if (diff < 0)
179			p = &(*p)->rb_left;
180		else if (diff > 0)
181			p = &(*p)->rb_right;
182		else {
183			int ret;
184
185			if (!collision_check)
186				goto add;
187
188			fio_mutex_up(rb_lock);
189			ret = col_check(c, i);
190			fio_mutex_down(rb_lock);
191
192			if (!ret)
193				goto add;
194
195			p = &(*p)->rb_right;
196		}
197	}
198
199	c = malloc(sizeof(*c));
200	RB_CLEAR_NODE(&c->rb_node);
201	INIT_FLIST_HEAD(&c->extent_list);
202	c->count = 0;
203	memcpy(c->hash, i->hash, sizeof(i->hash));
204	rb_link_node(&c->rb_node, parent, p);
205	rb_insert_color(&c->rb_node, &rb_root);
206add:
207	add_item(c, i);
208}
209
210static void insert_chunks(struct item *items, unsigned int nitems)
211{
212	int i;
213
214	fio_mutex_down(rb_lock);
215
216	for (i = 0; i < nitems; i++)
217		insert_chunk(&items[i]);
218
219	fio_mutex_up(rb_lock);
220}
221
222static void crc_buf(void *buf, uint32_t *hash)
223{
224	struct fio_md5_ctx ctx = { .hash = hash };
225
226	fio_md5_init(&ctx);
227	fio_md5_update(&ctx, buf, blocksize);
228	fio_md5_final(&ctx);
229}
230
231static int do_work(struct worker_thread *thread, void *buf)
232{
233	unsigned int nblocks, i;
234	off_t offset;
235	int err = 0, nitems = 0;
236	struct item *items;
237
238	nblocks = thread->size / blocksize;
239	offset = thread->cur_offset;
240	items = malloc(sizeof(*items) * nblocks);
241
242	for (i = 0; i < nblocks; i++) {
243		if (read_block(thread->fd, buf, offset))
244			break;
245		items[i].offset = offset;
246		crc_buf(buf, items[i].hash);
247		offset += blocksize;
248		nitems++;
249	}
250
251	insert_chunks(items, nitems);
252	thread->items += nitems;
253	free(items);
254	return err;
255}
256
257static void *thread_fn(void *data)
258{
259	struct worker_thread *thread = data;
260	void *buf;
261
262	buf = fio_memalign(blocksize, blocksize);
263
264	do {
265		if (get_work(&thread->cur_offset, &thread->size)) {
266			thread->err = 1;
267			break;
268		}
269		if (do_work(thread, buf)) {
270			thread->err = 1;
271			break;
272		}
273	} while (1);
274
275	fio_memfree(buf, blocksize);
276	return NULL;
277}
278
279static int __dedupe_check(int fd, uint64_t dev_size)
280{
281	struct worker_thread *threads;
282	unsigned long nitems;
283	int i, err = 0;
284
285	total_size = dev_size;
286	cur_offset = 0;
287	size_lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
288
289	threads = malloc(num_threads * sizeof(struct worker_thread));
290	for (i = 0; i < num_threads; i++) {
291		threads[i].fd = fd;
292		threads[i].items = 0;
293		threads[i].err = 0;
294
295		err = pthread_create(&threads[i].thread, NULL, thread_fn, &threads[i]);
296		if (err) {
297			log_err("fio: thread startup failed\n");
298			break;
299		}
300	}
301
302	nitems = 0;
303	for (i = 0; i < num_threads; i++) {
304		void *ret;
305		pthread_join(threads[i].thread, &ret);
306		nitems += threads[i].items;
307	}
308
309	printf("Threads(%u): %lu items processed\n", num_threads, nitems);
310
311	fio_mutex_remove(size_lock);
312	return err;
313}
314
315static int dedupe_check(const char *filename)
316{
317	uint64_t dev_size;
318	struct stat sb;
319	int flags;
320
321	flags = O_RDONLY;
322	if (odirect)
323		flags |= O_DIRECT;
324
325	dev_fd = open(filename, flags);
326	if (dev_fd == -1) {
327		perror("open");
328		return 1;
329	}
330
331	if (fstat(dev_fd, &sb) < 0) {
332		perror("fstat");
333		close(dev_fd);
334		return 1;
335	}
336
337	dev_size = get_size(dev_fd, &sb);
338	if (!dev_size) {
339		close(dev_fd);
340		return 1;
341	}
342
343	printf("Will check <%s>, size <%lu>\n", filename, dev_size);
344
345	return __dedupe_check(dev_fd, dev_size);
346}
347
348static void show_chunk(struct chunk *c)
349{
350	struct flist_head *n;
351	struct extent *e;
352
353	printf("c hash %8x %8x %8x %8x, count %lu\n", c->hash[0], c->hash[1], c->hash[2], c->hash[3], c->count);
354	flist_for_each(n, &c->extent_list) {
355		e = flist_entry(n, struct extent, list);
356		printf("\toffset %lu\n", e->offset);
357	}
358}
359
360static void iter_rb_tree(void)
361{
362	struct rb_node *n;
363	uint64_t nchunks;
364	uint64_t nextents;
365	double perc;
366
367	nchunks = nextents = 0;
368
369	n = rb_first(&rb_root);
370	if (!n)
371		return;
372
373	do {
374		struct chunk *c;
375
376		c = rb_entry(n, struct chunk, rb_node);
377		nchunks++;
378		nextents += c->count;
379
380		if (dump_output)
381			show_chunk(c);
382
383	} while ((n = rb_next(n)) != NULL);
384
385	printf("Extents=%lu, Unique extents=%lu\n", nextents, nchunks);
386	printf("De-dupe factor: %3.2f\n", (double) nextents / (double) nchunks);
387
388	perc = 1.00 - ((double) nchunks / (double) nextents);
389	perc *= 100.0;
390	printf("Fio setting: dedupe_percentage=%u\n", (int) (perc + 0.50));
391}
392
393static int usage(char *argv[])
394{
395	log_err("Check for dedupable blocks on a device/file\n\n");
396	log_err("%s: [options] <device or file>\n", argv[0]);
397	log_err("\t-b\tChunk size to use\n");
398	log_err("\t-t\tNumber of threads to use\n");
399	log_err("\t-d\tFull extent/chunk debug output\n");
400	log_err("\t-o\tUse O_DIRECT\n");
401	log_err("\t-c\tFull collision check\n");
402	return 1;
403}
404
405int main(int argc, char *argv[])
406{
407	int c, ret;
408
409	while ((c = getopt(argc, argv, "b:t:d:o:c:")) != -1) {
410		switch (c) {
411		case 'b':
412			blocksize = atoi(optarg);
413			break;
414		case 't':
415			num_threads = atoi(optarg);
416			break;
417		case 'd':
418			dump_output = atoi(optarg);
419			break;
420		case 'o':
421			odirect = atoi(optarg);
422			break;
423		case 'c':
424			collision_check = atoi(optarg);
425			break;
426		case '?':
427		default:
428			return usage(argv);
429		}
430	}
431
432	if (!num_threads)
433		num_threads = cpus_online();
434
435	if (argc == optind)
436		return usage(argv);
437
438	sinit();
439
440	rb_root = RB_ROOT;
441	rb_lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
442
443	ret = dedupe_check(argv[optind]);
444
445	iter_rb_tree();
446
447	scleanup();
448	return ret;
449}
450