xref: /freebsd/sys/contrib/openzfs/cmd/zstream/zstream_redup.c (revision 5def4c47d4bd90b209b9b4a4ba9faec15846d8fd)
1 /*
2  * CDDL HEADER START
3  *
4  * This file and its contents are supplied under the terms of the
5  * Common Development and Distribution License ("CDDL"), version 1.0.
6  * You may only use this file in accordance with the terms of version
7  * 1.0 of the CDDL.
8  *
9  * A full copy of the text of the CDDL should have accompanied this
10  * source.  A copy of the CDDL is also available via the Internet at
11  * http://www.illumos.org/license/CDDL.
12  *
13  * CDDL HEADER END
14  */
15 
16 /*
17  * Copyright (c) 2020 by Delphix. All rights reserved.
18  */
19 
20 #include <assert.h>
21 #include <cityhash.h>
22 #include <ctype.h>
23 #include <errno.h>
24 #include <fcntl.h>
25 #include <libzfs.h>
26 #include <libzutil.h>
27 #include <stddef.h>
28 #include <stdio.h>
29 #include <stdlib.h>
30 #include <strings.h>
31 #include <umem.h>
32 #include <unistd.h>
33 #include <sys/debug.h>
34 #include <sys/stat.h>
35 #include <sys/zfs_ioctl.h>
36 #include <sys/zio_checksum.h>
37 #include "zfs_fletcher.h"
38 #include "zstream.h"
39 
40 
41 #define	MAX_RDT_PHYSMEM_PERCENT		20
42 #define	SMALLEST_POSSIBLE_MAX_RDT_MB		128
43 
44 typedef struct redup_entry {
45 	struct redup_entry	*rde_next;
46 	uint64_t rde_guid;
47 	uint64_t rde_object;
48 	uint64_t rde_offset;
49 	uint64_t rde_stream_offset;
50 } redup_entry_t;
51 
52 typedef struct redup_table {
53 	redup_entry_t	**redup_hash_array;
54 	umem_cache_t	*ddecache;
55 	uint64_t	ddt_count;
56 	int		numhashbits;
57 } redup_table_t;
58 
59 int
60 highbit64(uint64_t i)
61 {
62 	if (i == 0)
63 		return (0);
64 
65 	return (NBBY * sizeof (uint64_t) - __builtin_clzll(i));
66 }
67 
68 static void *
69 safe_calloc(size_t n)
70 {
71 	void *rv = calloc(1, n);
72 	if (rv == NULL) {
73 		fprintf(stderr,
74 		    "Error: could not allocate %u bytes of memory\n",
75 		    (int)n);
76 		exit(1);
77 	}
78 	return (rv);
79 }
80 
81 /*
82  * Safe version of fread(), exits on error.
83  */
84 static int
85 sfread(void *buf, size_t size, FILE *fp)
86 {
87 	int rv = fread(buf, size, 1, fp);
88 	if (rv == 0 && ferror(fp)) {
89 		(void) fprintf(stderr, "Error while reading file: %s\n",
90 		    strerror(errno));
91 		exit(1);
92 	}
93 	return (rv);
94 }
95 
96 /*
97  * Safe version of pread(), exits on error.
98  */
99 static void
100 spread(int fd, void *buf, size_t count, off_t offset)
101 {
102 	ssize_t err = pread(fd, buf, count, offset);
103 	if (err == -1) {
104 		(void) fprintf(stderr,
105 		    "Error while reading file: %s\n",
106 		    strerror(errno));
107 		exit(1);
108 	} else if (err != count) {
109 		(void) fprintf(stderr,
110 		    "Error while reading file: short read\n");
111 		exit(1);
112 	}
113 }
114 
115 static int
116 dump_record(dmu_replay_record_t *drr, void *payload, int payload_len,
117     zio_cksum_t *zc, int outfd)
118 {
119 	assert(offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum)
120 	    == sizeof (dmu_replay_record_t) - sizeof (zio_cksum_t));
121 	fletcher_4_incremental_native(drr,
122 	    offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum), zc);
123 	if (drr->drr_type != DRR_BEGIN) {
124 		assert(ZIO_CHECKSUM_IS_ZERO(&drr->drr_u.
125 		    drr_checksum.drr_checksum));
126 		drr->drr_u.drr_checksum.drr_checksum = *zc;
127 	}
128 	fletcher_4_incremental_native(&drr->drr_u.drr_checksum.drr_checksum,
129 	    sizeof (zio_cksum_t), zc);
130 	if (write(outfd, drr, sizeof (*drr)) == -1)
131 		return (errno);
132 	if (payload_len != 0) {
133 		fletcher_4_incremental_native(payload, payload_len, zc);
134 		if (write(outfd, payload, payload_len) == -1)
135 			return (errno);
136 	}
137 	return (0);
138 }
139 
140 static void
141 rdt_insert(redup_table_t *rdt,
142     uint64_t guid, uint64_t object, uint64_t offset, uint64_t stream_offset)
143 {
144 	uint64_t ch = cityhash4(guid, object, offset, 0);
145 	uint64_t hashcode = BF64_GET(ch, 0, rdt->numhashbits);
146 	redup_entry_t **rdepp;
147 
148 	rdepp = &(rdt->redup_hash_array[hashcode]);
149 	redup_entry_t *rde = umem_cache_alloc(rdt->ddecache, UMEM_NOFAIL);
150 	rde->rde_next = *rdepp;
151 	rde->rde_guid = guid;
152 	rde->rde_object = object;
153 	rde->rde_offset = offset;
154 	rde->rde_stream_offset = stream_offset;
155 	*rdepp = rde;
156 	rdt->ddt_count++;
157 }
158 
159 static void
160 rdt_lookup(redup_table_t *rdt,
161     uint64_t guid, uint64_t object, uint64_t offset,
162     uint64_t *stream_offsetp)
163 {
164 	uint64_t ch = cityhash4(guid, object, offset, 0);
165 	uint64_t hashcode = BF64_GET(ch, 0, rdt->numhashbits);
166 
167 	for (redup_entry_t *rde = rdt->redup_hash_array[hashcode];
168 	    rde != NULL; rde = rde->rde_next) {
169 		if (rde->rde_guid == guid &&
170 		    rde->rde_object == object &&
171 		    rde->rde_offset == offset) {
172 			*stream_offsetp = rde->rde_stream_offset;
173 			return;
174 		}
175 	}
176 	assert(!"could not find expected redup table entry");
177 }
178 
179 /*
180  * Convert a dedup stream (generated by "zfs send -D") to a
181  * non-deduplicated stream.  The entire infd will be converted, including
182  * any substreams in a stream package (generated by "zfs send -RD"). The
183  * infd must be seekable.
184  */
185 static void
186 zfs_redup_stream(int infd, int outfd, boolean_t verbose)
187 {
188 	int bufsz = SPA_MAXBLOCKSIZE;
189 	dmu_replay_record_t thedrr = { 0 };
190 	dmu_replay_record_t *drr = &thedrr;
191 	redup_table_t rdt;
192 	zio_cksum_t stream_cksum;
193 	uint64_t numbuckets;
194 	uint64_t num_records = 0;
195 	uint64_t num_write_byref_records = 0;
196 
197 #ifdef _ILP32
198 	uint64_t max_rde_size = SMALLEST_POSSIBLE_MAX_RDT_MB << 20;
199 #else
200 	uint64_t physmem = sysconf(_SC_PHYS_PAGES) * sysconf(_SC_PAGESIZE);
201 	uint64_t max_rde_size =
202 	    MAX((physmem * MAX_RDT_PHYSMEM_PERCENT) / 100,
203 	    SMALLEST_POSSIBLE_MAX_RDT_MB << 20);
204 #endif
205 
206 	numbuckets = max_rde_size / (sizeof (redup_entry_t));
207 
208 	/*
209 	 * numbuckets must be a power of 2.  Increase number to
210 	 * a power of 2 if necessary.
211 	 */
212 	if (!ISP2(numbuckets))
213 		numbuckets = 1ULL << highbit64(numbuckets);
214 
215 	rdt.redup_hash_array =
216 	    safe_calloc(numbuckets * sizeof (redup_entry_t *));
217 	rdt.ddecache = umem_cache_create("rde", sizeof (redup_entry_t), 0,
218 	    NULL, NULL, NULL, NULL, NULL, 0);
219 	rdt.numhashbits = highbit64(numbuckets) - 1;
220 	rdt.ddt_count = 0;
221 
222 	char *buf = safe_calloc(bufsz);
223 	FILE *ofp = fdopen(infd, "r");
224 	long offset = ftell(ofp);
225 	while (sfread(drr, sizeof (*drr), ofp) != 0) {
226 		num_records++;
227 
228 		/*
229 		 * We need to regenerate the checksum.
230 		 */
231 		if (drr->drr_type != DRR_BEGIN) {
232 			bzero(&drr->drr_u.drr_checksum.drr_checksum,
233 			    sizeof (drr->drr_u.drr_checksum.drr_checksum));
234 		}
235 
236 		uint64_t payload_size = 0;
237 		switch (drr->drr_type) {
238 		case DRR_BEGIN:
239 		{
240 			struct drr_begin *drrb = &drr->drr_u.drr_begin;
241 			int fflags;
242 			ZIO_SET_CHECKSUM(&stream_cksum, 0, 0, 0, 0);
243 
244 			assert(drrb->drr_magic == DMU_BACKUP_MAGIC);
245 
246 			/* clear the DEDUP feature flag for this stream */
247 			fflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo);
248 			fflags &= ~(DMU_BACKUP_FEATURE_DEDUP |
249 			    DMU_BACKUP_FEATURE_DEDUPPROPS);
250 			/* cppcheck-suppress syntaxError */
251 			DMU_SET_FEATUREFLAGS(drrb->drr_versioninfo, fflags);
252 
253 			int sz = drr->drr_payloadlen;
254 			if (sz != 0) {
255 				if (sz > bufsz) {
256 					free(buf);
257 					buf = safe_calloc(sz);
258 					bufsz = sz;
259 				}
260 				(void) sfread(buf, sz, ofp);
261 			}
262 			payload_size = sz;
263 			break;
264 		}
265 
266 		case DRR_END:
267 		{
268 			struct drr_end *drre = &drr->drr_u.drr_end;
269 			/*
270 			 * Use the recalculated checksum, unless this is
271 			 * the END record of a stream package, which has
272 			 * no checksum.
273 			 */
274 			if (!ZIO_CHECKSUM_IS_ZERO(&drre->drr_checksum))
275 				drre->drr_checksum = stream_cksum;
276 			break;
277 		}
278 
279 		case DRR_OBJECT:
280 		{
281 			struct drr_object *drro = &drr->drr_u.drr_object;
282 
283 			if (drro->drr_bonuslen > 0) {
284 				payload_size = DRR_OBJECT_PAYLOAD_SIZE(drro);
285 				(void) sfread(buf, payload_size, ofp);
286 			}
287 			break;
288 		}
289 
290 		case DRR_SPILL:
291 		{
292 			struct drr_spill *drrs = &drr->drr_u.drr_spill;
293 			payload_size = DRR_SPILL_PAYLOAD_SIZE(drrs);
294 			(void) sfread(buf, payload_size, ofp);
295 			break;
296 		}
297 
298 		case DRR_WRITE_BYREF:
299 		{
300 			struct drr_write_byref drrwb =
301 			    drr->drr_u.drr_write_byref;
302 
303 			num_write_byref_records++;
304 
305 			/*
306 			 * Look up in hash table by drrwb->drr_refguid,
307 			 * drr_refobject, drr_refoffset.  Replace this
308 			 * record with the found WRITE record, but with
309 			 * drr_object,drr_offset,drr_toguid replaced with ours.
310 			 */
311 			uint64_t stream_offset = 0;
312 			rdt_lookup(&rdt, drrwb.drr_refguid,
313 			    drrwb.drr_refobject, drrwb.drr_refoffset,
314 			    &stream_offset);
315 
316 			spread(infd, drr, sizeof (*drr), stream_offset);
317 
318 			assert(drr->drr_type == DRR_WRITE);
319 			struct drr_write *drrw = &drr->drr_u.drr_write;
320 			assert(drrw->drr_toguid == drrwb.drr_refguid);
321 			assert(drrw->drr_object == drrwb.drr_refobject);
322 			assert(drrw->drr_offset == drrwb.drr_refoffset);
323 
324 			payload_size = DRR_WRITE_PAYLOAD_SIZE(drrw);
325 			spread(infd, buf, payload_size,
326 			    stream_offset + sizeof (*drr));
327 
328 			drrw->drr_toguid = drrwb.drr_toguid;
329 			drrw->drr_object = drrwb.drr_object;
330 			drrw->drr_offset = drrwb.drr_offset;
331 			break;
332 		}
333 
334 		case DRR_WRITE:
335 		{
336 			struct drr_write *drrw = &drr->drr_u.drr_write;
337 			payload_size = DRR_WRITE_PAYLOAD_SIZE(drrw);
338 			(void) sfread(buf, payload_size, ofp);
339 
340 			rdt_insert(&rdt, drrw->drr_toguid,
341 			    drrw->drr_object, drrw->drr_offset, offset);
342 			break;
343 		}
344 
345 		case DRR_WRITE_EMBEDDED:
346 		{
347 			struct drr_write_embedded *drrwe =
348 			    &drr->drr_u.drr_write_embedded;
349 			payload_size =
350 			    P2ROUNDUP((uint64_t)drrwe->drr_psize, 8);
351 			(void) sfread(buf, payload_size, ofp);
352 			break;
353 		}
354 
355 		case DRR_FREEOBJECTS:
356 		case DRR_FREE:
357 		case DRR_OBJECT_RANGE:
358 			break;
359 
360 		default:
361 			(void) fprintf(stderr, "INVALID record type 0x%x\n",
362 			    drr->drr_type);
363 			/* should never happen, so assert */
364 			assert(B_FALSE);
365 		}
366 
367 		if (feof(ofp)) {
368 			fprintf(stderr, "Error: unexpected end-of-file\n");
369 			exit(1);
370 		}
371 		if (ferror(ofp)) {
372 			fprintf(stderr, "Error while reading file: %s\n",
373 			    strerror(errno));
374 			exit(1);
375 		}
376 
377 		/*
378 		 * We need to recalculate the checksum, and it needs to be
379 		 * initially zero to do that.  BEGIN records don't have
380 		 * a checksum.
381 		 */
382 		if (drr->drr_type != DRR_BEGIN) {
383 			bzero(&drr->drr_u.drr_checksum.drr_checksum,
384 			    sizeof (drr->drr_u.drr_checksum.drr_checksum));
385 		}
386 		if (dump_record(drr, buf, payload_size,
387 		    &stream_cksum, outfd) != 0)
388 			break;
389 		if (drr->drr_type == DRR_END) {
390 			/*
391 			 * Typically the END record is either the last
392 			 * thing in the stream, or it is followed
393 			 * by a BEGIN record (which also zeros the checksum).
394 			 * However, a stream package ends with two END
395 			 * records.  The last END record's checksum starts
396 			 * from zero.
397 			 */
398 			ZIO_SET_CHECKSUM(&stream_cksum, 0, 0, 0, 0);
399 		}
400 		offset = ftell(ofp);
401 	}
402 
403 	if (verbose) {
404 		char mem_str[16];
405 		zfs_nicenum(rdt.ddt_count * sizeof (redup_entry_t),
406 		    mem_str, sizeof (mem_str));
407 		fprintf(stderr, "converted stream with %llu total records, "
408 		    "including %llu dedup records, using %sB memory.\n",
409 		    (long long)num_records,
410 		    (long long)num_write_byref_records,
411 		    mem_str);
412 	}
413 
414 	umem_cache_destroy(rdt.ddecache);
415 	free(rdt.redup_hash_array);
416 	free(buf);
417 	(void) fclose(ofp);
418 }
419 
420 int
421 zstream_do_redup(int argc, char *argv[])
422 {
423 	boolean_t verbose = B_FALSE;
424 	int c;
425 
426 	while ((c = getopt(argc, argv, "v")) != -1) {
427 		switch (c) {
428 		case 'v':
429 			verbose = B_TRUE;
430 			break;
431 		case '?':
432 			(void) fprintf(stderr, "invalid option '%c'\n",
433 			    optopt);
434 			zstream_usage();
435 			break;
436 		}
437 	}
438 
439 	argc -= optind;
440 	argv += optind;
441 
442 	if (argc != 1)
443 		zstream_usage();
444 
445 	const char *filename = argv[0];
446 
447 	if (isatty(STDOUT_FILENO)) {
448 		(void) fprintf(stderr,
449 		    "Error: Stream can not be written to a terminal.\n"
450 		    "You must redirect standard output.\n");
451 		return (1);
452 	}
453 
454 	int fd = open(filename, O_RDONLY);
455 	if (fd == -1) {
456 		(void) fprintf(stderr,
457 		    "Error while opening file '%s': %s\n",
458 		    filename, strerror(errno));
459 		exit(1);
460 	}
461 
462 	fletcher_4_init();
463 	zfs_redup_stream(fd, STDOUT_FILENO, verbose);
464 	fletcher_4_fini();
465 
466 	close(fd);
467 
468 	return (0);
469 }
470