2 * Copyright (C) 2015 Google, Inc.
4 * Author: Sami Tolvanen <samitolvanen@google.com>
6 * This program is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License as published by the Free
8 * Software Foundation; either version 2 of the License, or (at your option)
12 #include "dm-verity-fec.h"
13 #include <linux/math64.h>
14 #include <linux/sysfs.h>
16 #define DM_MSG_PREFIX "verity-fec"
19 * If error correction has been configured, returns true.
21 bool verity_fec_is_enabled(struct dm_verity *v)
23 return v->fec && v->fec->dev;
27 * Return a pointer to dm_verity_fec_io after dm_verity_io and its variable
30 static inline struct dm_verity_fec_io *fec_io(struct dm_verity_io *io)
32 return (struct dm_verity_fec_io *) verity_io_digest_end(io->v, io);
36 * Return an interleaved offset for a byte in RS block.
38 static inline u64 fec_interleave(struct dm_verity *v, u64 offset)
42 mod = do_div(offset, v->fec->rsn);
43 return offset + mod * (v->fec->rounds << v->data_dev_block_bits);
47 * Decode an RS block using Reed-Solomon.
49 static int fec_decode_rs8(struct dm_verity *v, struct dm_verity_fec_io *fio,
50 u8 *data, u8 *fec, int neras)
53 uint16_t par[DM_VERITY_FEC_RSM - DM_VERITY_FEC_MIN_RSN];
55 for (i = 0; i < v->fec->roots; i++)
58 return decode_rs8(fio->rs, data, par, v->fec->rsn, NULL, neras,
59 fio->erasures, 0, NULL);
63 * Read error-correcting codes for the requested RS block. Returns a pointer
64 * to the data block. Caller is responsible for releasing buf.
66 static u8 *fec_read_parity(struct dm_verity *v, u64 rsb, int index,
67 unsigned *offset, struct dm_buffer **buf)
72 position = (index + rsb) * v->fec->roots;
73 block = position >> v->data_dev_block_bits;
74 *offset = (unsigned)(position - (block << v->data_dev_block_bits));
76 res = dm_bufio_read(v->fec->bufio, v->fec->start + block, buf);
77 if (unlikely(IS_ERR(res))) {
78 DMERR("%s: FEC %llu: parity read failed (block %llu): %ld",
79 v->data_dev->name, (unsigned long long)rsb,
80 (unsigned long long)(v->fec->start + block),
88 /* Loop over each preallocated buffer slot. */
89 #define fec_for_each_prealloc_buffer(__i) \
90 for (__i = 0; __i < DM_VERITY_FEC_BUF_PREALLOC; __i++)
92 /* Loop over each extra buffer slot. */
93 #define fec_for_each_extra_buffer(io, __i) \
94 for (__i = DM_VERITY_FEC_BUF_PREALLOC; __i < DM_VERITY_FEC_BUF_MAX; __i++)
96 /* Loop over each allocated buffer. */
97 #define fec_for_each_buffer(io, __i) \
98 for (__i = 0; __i < (io)->nbufs; __i++)
100 /* Loop over each RS block in each allocated buffer. */
101 #define fec_for_each_buffer_rs_block(io, __i, __j) \
102 fec_for_each_buffer(io, __i) \
103 for (__j = 0; __j < 1 << DM_VERITY_FEC_BUF_RS_BITS; __j++)
106 * Return a pointer to the current RS block when called inside
107 * fec_for_each_buffer_rs_block.
109 static inline u8 *fec_buffer_rs_block(struct dm_verity *v,
110 struct dm_verity_fec_io *fio,
111 unsigned i, unsigned j)
113 return &fio->bufs[i][j * v->fec->rsn];
117 * Return an index to the current RS block when called inside
118 * fec_for_each_buffer_rs_block.
120 static inline unsigned fec_buffer_rs_index(unsigned i, unsigned j)
122 return (i << DM_VERITY_FEC_BUF_RS_BITS) + j;
126 * Decode all RS blocks from buffers and copy corrected bytes into fio->output
127 * starting from block_offset.
129 static int fec_decode_bufs(struct dm_verity *v, struct dm_verity_fec_io *fio,
130 u64 rsb, int byte_index, unsigned block_offset,
133 int r, corrected = 0, res;
134 struct dm_buffer *buf;
135 unsigned n, i, offset;
138 par = fec_read_parity(v, rsb, block_offset, &offset, &buf);
143 * Decode the RS blocks we have in bufs. Each RS block results in
144 * one corrected target byte and consumes fec->roots parity bytes.
146 fec_for_each_buffer_rs_block(fio, n, i) {
147 block = fec_buffer_rs_block(v, fio, n, i);
148 res = fec_decode_rs8(v, fio, block, &par[offset], neras);
150 dm_bufio_release(buf);
157 fio->output[block_offset] = block[byte_index];
160 if (block_offset >= 1 << v->data_dev_block_bits)
163 /* read the next block when we run out of parity bytes */
164 offset += v->fec->roots;
165 if (offset >= 1 << v->data_dev_block_bits) {
166 dm_bufio_release(buf);
168 par = fec_read_parity(v, rsb, block_offset, &offset, &buf);
169 if (unlikely(IS_ERR(par)))
177 DMERR_LIMIT("%s: FEC %llu: failed to correct: %d",
178 v->data_dev->name, (unsigned long long)rsb, r);
180 DMWARN_LIMIT("%s: FEC %llu: corrected %d errors",
181 v->data_dev->name, (unsigned long long)rsb, r);
182 atomic_add_unless(&v->fec->corrected, 1, INT_MAX);
189 * Locate data block erasures using verity hashes.
191 static int fec_is_erasure(struct dm_verity *v, struct dm_verity_io *io,
192 u8 *want_digest, u8 *data)
194 if (unlikely(verity_hash(v, verity_io_hash_desc(v, io),
195 data, 1 << v->data_dev_block_bits,
196 verity_io_real_digest(v, io))))
199 return memcmp(verity_io_real_digest(v, io), want_digest,
200 v->digest_size) != 0;
204 * Read data blocks that are part of the RS block and deinterleave as much as
205 * fits into buffers. Check for erasure locations if @neras is non-NULL.
207 static int fec_read_bufs(struct dm_verity *v, struct dm_verity_io *io,
208 u64 rsb, u64 target, unsigned block_offset,
212 int i, j, target_index = -1;
213 struct dm_buffer *buf;
214 struct dm_bufio_client *bufio;
215 struct dm_verity_fec_io *fio = fec_io(io);
218 u8 want_digest[v->digest_size];
225 * read each of the rsn data blocks that are part of the RS block, and
226 * interleave contents to available bufs
228 for (i = 0; i < v->fec->rsn; i++) {
229 ileaved = fec_interleave(v, rsb * v->fec->rsn + i);
232 * target is the data block we want to correct, target_index is
233 * the index of this block within the rsn RS blocks
235 if (ileaved == target)
238 block = ileaved >> v->data_dev_block_bits;
239 bufio = v->fec->data_bufio;
241 if (block >= v->data_blocks) {
242 block -= v->data_blocks;
245 * blocks outside the area were assumed to contain
246 * zeros when encoding data was generated
248 if (unlikely(block >= v->fec->hash_blocks))
251 block += v->hash_start;
255 bbuf = dm_bufio_read(bufio, block, &buf);
256 if (unlikely(IS_ERR(bbuf))) {
257 DMWARN_LIMIT("%s: FEC %llu: read failed (%llu): %ld",
259 (unsigned long long)rsb,
260 (unsigned long long)block, PTR_ERR(bbuf));
262 /* assume the block is corrupted */
263 if (neras && *neras <= v->fec->roots)
264 fio->erasures[(*neras)++] = i;
269 /* locate erasures if the block is on the data device */
270 if (bufio == v->fec->data_bufio &&
271 verity_hash_for_block(v, io, block, want_digest,
273 /* skip known zero blocks entirely */
278 * skip if we have already found the theoretical
279 * maximum number (i.e. fec->roots) of erasures
281 if (neras && *neras <= v->fec->roots &&
282 fec_is_erasure(v, io, want_digest, bbuf))
283 fio->erasures[(*neras)++] = i;
287 * deinterleave and copy the bytes that fit into bufs,
288 * starting from block_offset
290 fec_for_each_buffer_rs_block(fio, n, j) {
291 k = fec_buffer_rs_index(n, j) + block_offset;
293 if (k >= 1 << v->data_dev_block_bits)
296 rs_block = fec_buffer_rs_block(v, fio, n, j);
297 rs_block[i] = bbuf[k];
300 dm_bufio_release(buf);
307 * Allocate RS control structure and FEC buffers from preallocated mempools,
308 * and attempt to allocate as many extra buffers as available.
310 static int fec_alloc_bufs(struct dm_verity *v, struct dm_verity_fec_io *fio)
315 fio->rs = mempool_alloc(v->fec->rs_pool, 0);
316 if (unlikely(!fio->rs)) {
317 DMERR("failed to allocate RS");
322 fec_for_each_prealloc_buffer(n) {
326 fio->bufs[n] = mempool_alloc(v->fec->prealloc_pool, GFP_NOIO);
327 if (unlikely(!fio->bufs[n])) {
328 DMERR("failed to allocate FEC buffer");
333 /* try to allocate the maximum number of buffers */
334 fec_for_each_extra_buffer(fio, n) {
338 fio->bufs[n] = mempool_alloc(v->fec->extra_pool, GFP_NOIO);
339 /* we can manage with even one buffer if necessary */
340 if (unlikely(!fio->bufs[n]))
346 fio->output = mempool_alloc(v->fec->output_pool, GFP_NOIO);
349 DMERR("failed to allocate FEC page");
358 * Initialize buffers and clear erasures. fec_read_bufs() assumes buffers are
359 * zeroed before deinterleaving.
361 static void fec_init_bufs(struct dm_verity *v, struct dm_verity_fec_io *fio)
365 fec_for_each_buffer(fio, n)
366 memset(fio->bufs[n], 0, v->fec->rsn << DM_VERITY_FEC_BUF_RS_BITS);
368 memset(fio->erasures, 0, sizeof(fio->erasures));
372 * Decode all RS blocks in a single data block and return the target block
373 * (indicated by @offset) in fio->output. If @use_erasures is non-zero, uses
374 * hashes to locate erasures.
376 static int fec_decode_rsb(struct dm_verity *v, struct dm_verity_io *io,
377 struct dm_verity_fec_io *fio, u64 rsb, u64 offset,
383 r = fec_alloc_bufs(v, fio);
387 for (pos = 0; pos < 1 << v->data_dev_block_bits; ) {
388 fec_init_bufs(v, fio);
390 r = fec_read_bufs(v, io, rsb, offset, pos,
391 use_erasures ? &neras : NULL);
395 r = fec_decode_bufs(v, fio, rsb, r, pos, neras);
399 pos += fio->nbufs << DM_VERITY_FEC_BUF_RS_BITS;
402 /* Always re-validate the corrected block against the expected hash */
403 r = verity_hash(v, verity_io_hash_desc(v, io), fio->output,
404 1 << v->data_dev_block_bits,
405 verity_io_real_digest(v, io));
409 if (memcmp(verity_io_real_digest(v, io), verity_io_want_digest(v, io),
411 DMERR_LIMIT("%s: FEC %llu: failed to correct (%d erasures)",
412 v->data_dev->name, (unsigned long long)rsb, neras);
419 static int fec_bv_copy(struct dm_verity *v, struct dm_verity_io *io, u8 *data,
422 struct dm_verity_fec_io *fio = fec_io(io);
424 memcpy(data, &fio->output[fio->output_pos], len);
425 fio->output_pos += len;
431 * Correct errors in a block. Copies corrected block to dest if non-NULL,
432 * otherwise to a bio_vec starting from iter.
434 int verity_fec_decode(struct dm_verity *v, struct dm_verity_io *io,
435 enum verity_block_type type, sector_t block, u8 *dest,
436 struct bvec_iter *iter)
439 struct dm_verity_fec_io *fio = fec_io(io);
440 u64 offset, res, rsb;
442 if (!verity_fec_is_enabled(v))
445 if (fio->level >= DM_VERITY_FEC_MAX_RECURSION) {
446 DMWARN_LIMIT("%s: FEC: recursion too deep", v->data_dev->name);
452 if (type == DM_VERITY_BLOCK_TYPE_METADATA)
453 block += v->data_blocks;
456 * For RS(M, N), the continuous FEC data is divided into blocks of N
457 * bytes. Since block size may not be divisible by N, the last block
458 * is zero padded when decoding.
460 * Each byte of the block is covered by a different RS(M, N) code,
461 * and each code is interleaved over N blocks to make it less likely
462 * that bursty corruption will leave us in unrecoverable state.
465 offset = block << v->data_dev_block_bits;
466 res = div64_u64(offset, v->fec->rounds << v->data_dev_block_bits);
469 * The base RS block we can feed to the interleaver to find out all
470 * blocks required for decoding.
472 rsb = offset - res * (v->fec->rounds << v->data_dev_block_bits);
475 * Locating erasures is slow, so attempt to recover the block without
476 * them first. Do a second attempt with erasures if the corruption is
479 r = fec_decode_rsb(v, io, fio, rsb, offset, false);
481 r = fec_decode_rsb(v, io, fio, rsb, offset, true);
487 memcpy(dest, fio->output, 1 << v->data_dev_block_bits);
490 r = verity_for_bv_block(v, io, iter, fec_bv_copy);
499 * Clean up per-bio data.
501 void verity_fec_finish_io(struct dm_verity_io *io)
504 struct dm_verity_fec *f = io->v->fec;
505 struct dm_verity_fec_io *fio = fec_io(io);
507 if (!verity_fec_is_enabled(io->v))
510 mempool_free(fio->rs, f->rs_pool);
512 fec_for_each_prealloc_buffer(n)
513 mempool_free(fio->bufs[n], f->prealloc_pool);
515 fec_for_each_extra_buffer(fio, n)
516 mempool_free(fio->bufs[n], f->extra_pool);
518 mempool_free(fio->output, f->output_pool);
522 * Initialize per-bio data.
524 void verity_fec_init_io(struct dm_verity_io *io)
526 struct dm_verity_fec_io *fio = fec_io(io);
528 if (!verity_fec_is_enabled(io->v))
532 memset(fio->bufs, 0, sizeof(fio->bufs));
539 * Append feature arguments and values to the status table.
541 unsigned verity_fec_status_table(struct dm_verity *v, unsigned sz,
542 char *result, unsigned maxlen)
544 if (!verity_fec_is_enabled(v))
547 DMEMIT(" " DM_VERITY_OPT_FEC_DEV " %s "
548 DM_VERITY_OPT_FEC_BLOCKS " %llu "
549 DM_VERITY_OPT_FEC_START " %llu "
550 DM_VERITY_OPT_FEC_ROOTS " %d",
552 (unsigned long long)v->fec->blocks,
553 (unsigned long long)v->fec->start,
559 void verity_fec_dtr(struct dm_verity *v)
561 struct dm_verity_fec *f = v->fec;
562 struct kobject *kobj = &f->kobj_holder.kobj;
564 if (!verity_fec_is_enabled(v))
567 mempool_destroy(f->rs_pool);
568 mempool_destroy(f->prealloc_pool);
569 mempool_destroy(f->extra_pool);
570 kmem_cache_destroy(f->cache);
573 dm_bufio_client_destroy(f->data_bufio);
575 dm_bufio_client_destroy(f->bufio);
578 dm_put_device(v->ti, f->dev);
580 if (kobj->state_initialized) {
582 wait_for_completion(dm_get_completion_from_kobject(kobj));
590 static void *fec_rs_alloc(gfp_t gfp_mask, void *pool_data)
592 struct dm_verity *v = (struct dm_verity *)pool_data;
594 return init_rs(8, 0x11d, 0, 1, v->fec->roots);
597 static void fec_rs_free(void *element, void *pool_data)
599 struct rs_control *rs = (struct rs_control *)element;
605 bool verity_is_fec_opt_arg(const char *arg_name)
607 return (!strcasecmp(arg_name, DM_VERITY_OPT_FEC_DEV) ||
608 !strcasecmp(arg_name, DM_VERITY_OPT_FEC_BLOCKS) ||
609 !strcasecmp(arg_name, DM_VERITY_OPT_FEC_START) ||
610 !strcasecmp(arg_name, DM_VERITY_OPT_FEC_ROOTS));
613 int verity_fec_parse_opt_args(struct dm_arg_set *as, struct dm_verity *v,
614 unsigned *argc, const char *arg_name)
617 struct dm_target *ti = v->ti;
618 const char *arg_value;
619 unsigned long long num_ll;
624 ti->error = "FEC feature arguments require a value";
628 arg_value = dm_shift_arg(as);
631 if (!strcasecmp(arg_name, DM_VERITY_OPT_FEC_DEV)) {
632 r = dm_get_device(ti, arg_value, FMODE_READ, &v->fec->dev);
634 ti->error = "FEC device lookup failed";
638 } else if (!strcasecmp(arg_name, DM_VERITY_OPT_FEC_BLOCKS)) {
639 if (sscanf(arg_value, "%llu%c", &num_ll, &dummy) != 1 ||
640 ((sector_t)(num_ll << (v->data_dev_block_bits - SECTOR_SHIFT))
641 >> (v->data_dev_block_bits - SECTOR_SHIFT) != num_ll)) {
642 ti->error = "Invalid " DM_VERITY_OPT_FEC_BLOCKS;
645 v->fec->blocks = num_ll;
647 } else if (!strcasecmp(arg_name, DM_VERITY_OPT_FEC_START)) {
648 if (sscanf(arg_value, "%llu%c", &num_ll, &dummy) != 1 ||
649 ((sector_t)(num_ll << (v->data_dev_block_bits - SECTOR_SHIFT)) >>
650 (v->data_dev_block_bits - SECTOR_SHIFT) != num_ll)) {
651 ti->error = "Invalid " DM_VERITY_OPT_FEC_START;
654 v->fec->start = num_ll;
656 } else if (!strcasecmp(arg_name, DM_VERITY_OPT_FEC_ROOTS)) {
657 if (sscanf(arg_value, "%hhu%c", &num_c, &dummy) != 1 || !num_c ||
658 num_c < (DM_VERITY_FEC_RSM - DM_VERITY_FEC_MAX_RSN) ||
659 num_c > (DM_VERITY_FEC_RSM - DM_VERITY_FEC_MIN_RSN)) {
660 ti->error = "Invalid " DM_VERITY_OPT_FEC_ROOTS;
663 v->fec->roots = num_c;
666 ti->error = "Unrecognized verity FEC feature request";
673 static ssize_t corrected_show(struct kobject *kobj, struct kobj_attribute *attr,
676 struct dm_verity_fec *f = container_of(kobj, struct dm_verity_fec,
679 return sprintf(buf, "%d\n", atomic_read(&f->corrected));
682 static struct kobj_attribute attr_corrected = __ATTR_RO(corrected);
684 static struct attribute *fec_attrs[] = {
685 &attr_corrected.attr,
689 static struct kobj_type fec_ktype = {
690 .sysfs_ops = &kobj_sysfs_ops,
691 .default_attrs = fec_attrs,
692 .release = dm_kobject_release
696 * Allocate dm_verity_fec for v->fec. Must be called before verity_fec_ctr.
698 int verity_fec_ctr_alloc(struct dm_verity *v)
700 struct dm_verity_fec *f;
702 f = kzalloc(sizeof(struct dm_verity_fec), GFP_KERNEL);
704 v->ti->error = "Cannot allocate FEC structure";
713 * Validate arguments and preallocate memory. Must be called after arguments
714 * have been parsed using verity_fec_parse_opt_args.
716 int verity_fec_ctr(struct dm_verity *v)
719 struct dm_verity_fec *f = v->fec;
720 struct dm_target *ti = v->ti;
721 struct mapped_device *md = dm_table_get_md(ti->table);
724 if (!verity_fec_is_enabled(v)) {
729 /* Create a kobject and sysfs attributes */
730 init_completion(&f->kobj_holder.completion);
732 r = kobject_init_and_add(&f->kobj_holder.kobj, &fec_ktype,
733 &disk_to_dev(dm_disk(md))->kobj, "%s", "fec");
735 ti->error = "Cannot create kobject";
740 * FEC is computed over data blocks, possible metadata, and
741 * hash blocks. In other words, FEC covers total of fec_blocks
742 * blocks consisting of the following:
744 * data blocks | hash blocks | metadata (optional)
746 * We allow metadata after hash blocks to support a use case
747 * where all data is stored on the same device and FEC covers
750 * If metadata is included, we require it to be available on the
751 * hash device after the hash blocks.
754 hash_blocks = v->hash_blocks - v->hash_start;
757 * Require matching block sizes for data and hash devices for
760 if (v->data_dev_block_bits != v->hash_dev_block_bits) {
761 ti->error = "Block sizes must match to use FEC";
766 ti->error = "Missing " DM_VERITY_OPT_FEC_ROOTS;
769 f->rsn = DM_VERITY_FEC_RSM - f->roots;
772 ti->error = "Missing " DM_VERITY_OPT_FEC_BLOCKS;
776 f->rounds = f->blocks;
777 if (sector_div(f->rounds, f->rsn))
781 * Due to optional metadata, f->blocks can be larger than
782 * data_blocks and hash_blocks combined.
784 if (f->blocks < v->data_blocks + hash_blocks || !f->rounds) {
785 ti->error = "Invalid " DM_VERITY_OPT_FEC_BLOCKS;
790 * Metadata is accessed through the hash device, so we require
791 * it to be large enough.
793 f->hash_blocks = f->blocks - v->data_blocks;
794 if (dm_bufio_get_device_size(v->bufio) < f->hash_blocks) {
795 ti->error = "Hash device is too small for "
796 DM_VERITY_OPT_FEC_BLOCKS;
800 f->bufio = dm_bufio_client_create(f->dev->bdev,
801 1 << v->data_dev_block_bits,
803 if (IS_ERR(f->bufio)) {
804 ti->error = "Cannot initialize FEC bufio client";
805 return PTR_ERR(f->bufio);
808 if (dm_bufio_get_device_size(f->bufio) <
809 ((f->start + f->rounds * f->roots) >> v->data_dev_block_bits)) {
810 ti->error = "FEC device is too small";
814 f->data_bufio = dm_bufio_client_create(v->data_dev->bdev,
815 1 << v->data_dev_block_bits,
817 if (IS_ERR(f->data_bufio)) {
818 ti->error = "Cannot initialize FEC data bufio client";
819 return PTR_ERR(f->data_bufio);
822 if (dm_bufio_get_device_size(f->data_bufio) < v->data_blocks) {
823 ti->error = "Data device is too small";
827 /* Preallocate an rs_control structure for each worker thread */
828 f->rs_pool = mempool_create(num_online_cpus(), fec_rs_alloc,
829 fec_rs_free, (void *) v);
831 ti->error = "Cannot allocate RS pool";
835 f->cache = kmem_cache_create("dm_verity_fec_buffers",
836 f->rsn << DM_VERITY_FEC_BUF_RS_BITS,
839 ti->error = "Cannot create FEC buffer cache";
843 /* Preallocate DM_VERITY_FEC_BUF_PREALLOC buffers for each thread */
844 f->prealloc_pool = mempool_create_slab_pool(num_online_cpus() *
845 DM_VERITY_FEC_BUF_PREALLOC,
847 if (!f->prealloc_pool) {
848 ti->error = "Cannot allocate FEC buffer prealloc pool";
852 f->extra_pool = mempool_create_slab_pool(0, f->cache);
853 if (!f->extra_pool) {
854 ti->error = "Cannot allocate FEC buffer extra pool";
858 /* Preallocate an output buffer for each thread */
859 f->output_pool = mempool_create_kmalloc_pool(num_online_cpus(),
860 1 << v->data_dev_block_bits);
861 if (!f->output_pool) {
862 ti->error = "Cannot allocate FEC output pool";
866 /* Reserve space for our per-bio data */
867 ti->per_bio_data_size += sizeof(struct dm_verity_fec_io);