X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libbcachefs%2Fjournal_io.c;h=c5bc58247146a2cdbc1eecb987db7ba667e9677f;hb=f0334bbc1032b3be485767ec804ec868a8b1e7f0;hp=bb9a1936c24cdd4014c8e39a134d4535213b1731;hpb=db931a4571817d7d61be6bce306f1d42f7cd3398;p=bcachefs-tools-debian diff --git a/libbcachefs/journal_io.c b/libbcachefs/journal_io.c index bb9a193..c5bc582 100644 --- a/libbcachefs/journal_io.c +++ b/libbcachefs/journal_io.c @@ -1,26 +1,55 @@ // SPDX-License-Identifier: GPL-2.0 #include "bcachefs.h" +#include "alloc_background.h" #include "alloc_foreground.h" #include "btree_io.h" #include "btree_update_interior.h" +#include "btree_write_buffer.h" #include "buckets.h" #include "checksum.h" +#include "disk_groups.h" #include "error.h" -#include "io.h" #include "journal.h" #include "journal_io.h" #include "journal_reclaim.h" #include "journal_seq_blacklist.h" #include "replicas.h" +#include "sb-clean.h" +#include "trace.h" -#include +static struct nonce journal_nonce(const struct jset *jset) +{ + return (struct nonce) {{ + [0] = 0, + [1] = ((__le32 *) &jset->seq)[0], + [2] = ((__le32 *) &jset->seq)[1], + [3] = BCH_NONCE_JOURNAL, + }}; +} + +static bool jset_csum_good(struct bch_fs *c, struct jset *j) +{ + return bch2_checksum_type_valid(c, JSET_CSUM_TYPE(j)) && + !bch2_crc_cmp(j->csum, + csum_vstruct(c, JSET_CSUM_TYPE(j), journal_nonce(j), j)); +} -static void __journal_replay_free(struct journal_replay *i) +static inline u32 journal_entry_radix_idx(struct bch_fs *c, u64 seq) { - list_del(&i->list); + return (seq - c->journal_entries_base_seq) & (~0U >> 1); +} + +static void __journal_replay_free(struct bch_fs *c, + struct journal_replay *i) +{ + struct journal_replay **p = + genradix_ptr(&c->journal_entries, + journal_entry_radix_idx(c, le64_to_cpu(i->j.seq))); + + BUG_ON(*p != i); + *p = NULL; kvpfree(i, offsetof(struct journal_replay, j) + vstruct_bytes(&i->j)); - } static void journal_replay_free(struct bch_fs *c, struct journal_replay *i) @@ -28,13 +57,13 @@ static void journal_replay_free(struct bch_fs *c, struct journal_replay *i) i->ignore = true; if (!c->opts.read_entire_journal) - __journal_replay_free(i); + __journal_replay_free(c, i); } struct journal_list { struct closure cl; + u64 last_seq; struct mutex lock; - struct list_head *head; int ret; }; @@ -46,104 +75,127 @@ struct journal_list { * be replayed: */ static int journal_entry_add(struct bch_fs *c, struct bch_dev *ca, - struct journal_list *jlist, struct jset *j, - bool bad) + struct journal_ptr entry_ptr, + struct journal_list *jlist, struct jset *j) { - struct journal_replay *i, *pos; - struct bch_devs_list devs = { .nr = 0 }; - struct list_head *where; + struct genradix_iter iter; + struct journal_replay **_i, *i, *dup; + struct journal_ptr *ptr; size_t bytes = vstruct_bytes(j); - u64 last_seq = 0; - int ret; - - list_for_each_entry_reverse(i, jlist->head, list) { - if (!JSET_NO_FLUSH(&i->j)) { - last_seq = le64_to_cpu(i->j.last_seq); - break; - } - } + u64 last_seq = !JSET_NO_FLUSH(j) ? le64_to_cpu(j->last_seq) : 0; + int ret = JOURNAL_ENTRY_ADD_OK; /* Is this entry older than the range we need? */ if (!c->opts.read_entire_journal && - le64_to_cpu(j->seq) < last_seq) { - ret = JOURNAL_ENTRY_ADD_OUT_OF_RANGE; - goto out; - } + le64_to_cpu(j->seq) < jlist->last_seq) + return JOURNAL_ENTRY_ADD_OUT_OF_RANGE; + + /* + * genradixes are indexed by a ulong, not a u64, so we can't index them + * by sequence number directly: Assume instead that they will all fall + * within the range of +-2billion of the filrst one we find. + */ + if (!c->journal_entries_base_seq) + c->journal_entries_base_seq = max_t(s64, 1, le64_to_cpu(j->seq) - S32_MAX); /* Drop entries we don't need anymore */ - if (!JSET_NO_FLUSH(j)) { - list_for_each_entry_safe(i, pos, jlist->head, list) { - if (le64_to_cpu(i->j.seq) >= le64_to_cpu(j->last_seq)) + if (last_seq > jlist->last_seq && !c->opts.read_entire_journal) { + genradix_for_each_from(&c->journal_entries, iter, _i, + journal_entry_radix_idx(c, jlist->last_seq)) { + i = *_i; + + if (!i || i->ignore) + continue; + + if (le64_to_cpu(i->j.seq) >= last_seq) break; journal_replay_free(c, i); } } - list_for_each_entry_reverse(i, jlist->head, list) { - if (le64_to_cpu(j->seq) > le64_to_cpu(i->j.seq)) { - where = &i->list; - goto add; - } - } + jlist->last_seq = max(jlist->last_seq, last_seq); - where = jlist->head; -add: - i = where->next != jlist->head - ? container_of(where->next, struct journal_replay, list) - : NULL; + _i = genradix_ptr_alloc(&c->journal_entries, + journal_entry_radix_idx(c, le64_to_cpu(j->seq)), + GFP_KERNEL); + if (!_i) + return -BCH_ERR_ENOMEM_journal_entry_add; /* * Duplicate journal entries? If so we want the one that didn't have a * checksum error: */ - if (i && le64_to_cpu(j->seq) == le64_to_cpu(i->j.seq)) { - if (i->bad) { - devs = i->devs; - __journal_replay_free(i); - } else if (bad) { + dup = *_i; + if (dup) { + if (bytes == vstruct_bytes(&dup->j) && + !memcmp(j, &dup->j, bytes)) { + i = dup; goto found; - } else { - fsck_err_on(bytes != vstruct_bytes(&i->j) || - memcmp(j, &i->j, bytes), c, - "found duplicate but non identical journal entries (seq %llu)", - le64_to_cpu(j->seq)); + } + + if (!entry_ptr.csum_good) { + i = dup; goto found; } - } + if (!dup->csum_good) + goto replace; + fsck_err(c, journal_entry_replicas_data_mismatch, + "found duplicate but non identical journal entries (seq %llu)", + le64_to_cpu(j->seq)); + i = dup; + goto found; + } +replace: i = kvpmalloc(offsetof(struct journal_replay, j) + bytes, GFP_KERNEL); - if (!i) { - ret = -ENOMEM; - goto out; + if (!i) + return -BCH_ERR_ENOMEM_journal_entry_add; + + i->nr_ptrs = 0; + i->csum_good = entry_ptr.csum_good; + i->ignore = false; + unsafe_memcpy(&i->j, j, bytes, "embedded variable length struct"); + i->ptrs[i->nr_ptrs++] = entry_ptr; + + if (dup) { + if (dup->nr_ptrs >= ARRAY_SIZE(dup->ptrs)) { + bch_err(c, "found too many copies of journal entry %llu", + le64_to_cpu(i->j.seq)); + dup->nr_ptrs = ARRAY_SIZE(dup->ptrs) - 1; + } + + /* The first ptr should represent the jset we kept: */ + memcpy(i->ptrs + i->nr_ptrs, + dup->ptrs, + sizeof(dup->ptrs[0]) * dup->nr_ptrs); + i->nr_ptrs += dup->nr_ptrs; + __journal_replay_free(c, dup); } - list_add(&i->list, where); - i->devs = devs; - i->bad = bad; - i->ignore = false; - memcpy(&i->j, j, bytes); + *_i = i; + return 0; found: - if (!bch2_dev_list_has_dev(i->devs, ca->dev_idx)) - bch2_dev_list_add_dev(&i->devs, ca->dev_idx); - else - fsck_err_on(1, c, "duplicate journal entries on same device"); - ret = JOURNAL_ENTRY_ADD_OK; + for (ptr = i->ptrs; ptr < i->ptrs + i->nr_ptrs; ptr++) { + if (ptr->dev == ca->dev_idx) { + bch_err(c, "duplicate journal entry %llu on same device", + le64_to_cpu(i->j.seq)); + goto out; + } + } + + if (i->nr_ptrs >= ARRAY_SIZE(i->ptrs)) { + bch_err(c, "found too many copies of journal entry %llu", + le64_to_cpu(i->j.seq)); + goto out; + } + + i->ptrs[i->nr_ptrs++] = entry_ptr; out: fsck_err: return ret; } -static struct nonce journal_nonce(const struct jset *jset) -{ - return (struct nonce) {{ - [0] = 0, - [1] = ((__le32 *) &jset->seq)[0], - [2] = ((__le32 *) &jset->seq)[1], - [3] = BCH_NONCE_JOURNAL, - }}; -} - /* this fills in a range with empty jset_entries: */ static void journal_entry_null_range(void *start, void *end) { @@ -157,73 +209,100 @@ static void journal_entry_null_range(void *start, void *end) #define JOURNAL_ENTRY_NONE 6 #define JOURNAL_ENTRY_BAD 7 -#define journal_entry_err(c, msg, ...) \ +static void journal_entry_err_msg(struct printbuf *out, + u32 version, + struct jset *jset, + struct jset_entry *entry) +{ + prt_str(out, "invalid journal entry, version="); + bch2_version_to_text(out, version); + + if (entry) { + prt_str(out, " type="); + prt_str(out, bch2_jset_entry_types[entry->type]); + } + + if (!jset) { + prt_printf(out, " in superblock"); + } else { + + prt_printf(out, " seq=%llu", le64_to_cpu(jset->seq)); + + if (entry) + prt_printf(out, " offset=%zi/%u", + (u64 *) entry - jset->_data, + le32_to_cpu(jset->u64s)); + } + + prt_str(out, ": "); +} + +#define journal_entry_err(c, version, jset, entry, _err, msg, ...) \ ({ \ - switch (write) { \ + struct printbuf _buf = PRINTBUF; \ + \ + journal_entry_err_msg(&_buf, version, jset, entry); \ + prt_printf(&_buf, msg, ##__VA_ARGS__); \ + \ + switch (flags & BKEY_INVALID_WRITE) { \ case READ: \ - mustfix_fsck_err(c, msg, ##__VA_ARGS__); \ + mustfix_fsck_err(c, _err, "%s", _buf.buf); \ break; \ case WRITE: \ - bch_err(c, "corrupt metadata before write:\n" \ - msg, ##__VA_ARGS__); \ + bch2_sb_error_count(c, BCH_FSCK_ERR_##_err); \ + bch_err(c, "corrupt metadata before write: %s\n", _buf.buf);\ if (bch2_fs_inconsistent(c)) { \ - ret = BCH_FSCK_ERRORS_NOT_FIXED; \ + ret = -BCH_ERR_fsck_errors_not_fixed; \ goto fsck_err; \ } \ break; \ } \ + \ + printbuf_exit(&_buf); \ true; \ }) -#define journal_entry_err_on(cond, c, msg, ...) \ - ((cond) ? journal_entry_err(c, msg, ##__VA_ARGS__) : false) +#define journal_entry_err_on(cond, ...) \ + ((cond) ? journal_entry_err(__VA_ARGS__) : false) #define FSCK_DELETED_KEY 5 -static int journal_validate_key(struct bch_fs *c, struct jset *jset, +static int journal_validate_key(struct bch_fs *c, + struct jset *jset, struct jset_entry *entry, unsigned level, enum btree_id btree_id, struct bkey_i *k, - const char *type, int write) + unsigned version, int big_endian, + enum bkey_invalid_flags flags) { + int write = flags & BKEY_INVALID_WRITE; void *next = vstruct_next(entry); - const char *invalid; - unsigned version = le32_to_cpu(jset->version); + struct printbuf buf = PRINTBUF; int ret = 0; - if (journal_entry_err_on(!k->k.u64s, c, - "invalid %s in jset %llu offset %zi/%u entry offset %zi/%u: k->u64s 0", - type, le64_to_cpu(jset->seq), - (u64 *) entry - jset->_data, - le32_to_cpu(jset->u64s), - (u64 *) k - entry->_data, - le16_to_cpu(entry->u64s))) { + if (journal_entry_err_on(!k->k.u64s, + c, version, jset, entry, + journal_entry_bkey_u64s_0, + "k->u64s 0")) { entry->u64s = cpu_to_le16((u64 *) k - entry->_data); journal_entry_null_range(vstruct_next(entry), next); return FSCK_DELETED_KEY; } if (journal_entry_err_on((void *) bkey_next(k) > - (void *) vstruct_next(entry), c, - "invalid %s in jset %llu offset %zi/%u entry offset %zi/%u: extends past end of journal entry", - type, le64_to_cpu(jset->seq), - (u64 *) entry - jset->_data, - le32_to_cpu(jset->u64s), - (u64 *) k - entry->_data, - le16_to_cpu(entry->u64s))) { + (void *) vstruct_next(entry), + c, version, jset, entry, + journal_entry_bkey_past_end, + "extends past end of journal entry")) { entry->u64s = cpu_to_le16((u64 *) k - entry->_data); journal_entry_null_range(vstruct_next(entry), next); return FSCK_DELETED_KEY; } - if (journal_entry_err_on(k->k.format != KEY_FORMAT_CURRENT, c, - "invalid %s in jset %llu offset %zi/%u entry offset %zi/%u: bad format %u", - type, le64_to_cpu(jset->seq), - (u64 *) entry - jset->_data, - le32_to_cpu(jset->u64s), - (u64 *) k - entry->_data, - le16_to_cpu(entry->u64s), - k->k.format)) { + if (journal_entry_err_on(k->k.format != KEY_FORMAT_CURRENT, + c, version, jset, entry, + journal_entry_bkey_bad_format, + "bad format %u", k->k.format)) { le16_add_cpu(&entry->u64s, -((u16) k->k.u64s)); memmove(k, bkey_next(k), next - (void *) bkey_next(k)); journal_entry_null_range(vstruct_next(entry), next); @@ -231,42 +310,45 @@ static int journal_validate_key(struct bch_fs *c, struct jset *jset, } if (!write) - bch2_bkey_compat(level, btree_id, version, - JSET_BIG_ENDIAN(jset), write, - NULL, bkey_to_packed(k)); - - invalid = bch2_bkey_invalid(c, bkey_i_to_s_c(k), - __btree_node_type(level, btree_id)); - if (invalid) { - char buf[160]; - - bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(k)); - mustfix_fsck_err(c, "invalid %s in jset %llu offset %zi/%u entry offset %zi/%u: %s\n%s", - type, le64_to_cpu(jset->seq), - (u64 *) entry - jset->_data, - le32_to_cpu(jset->u64s), - (u64 *) k - entry->_data, - le16_to_cpu(entry->u64s), - invalid, buf); + bch2_bkey_compat(level, btree_id, version, big_endian, + write, NULL, bkey_to_packed(k)); + + if (bch2_bkey_invalid(c, bkey_i_to_s_c(k), + __btree_node_type(level, btree_id), write, &buf)) { + printbuf_reset(&buf); + journal_entry_err_msg(&buf, version, jset, entry); + prt_newline(&buf); + printbuf_indent_add(&buf, 2); + + bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(k)); + prt_newline(&buf); + bch2_bkey_invalid(c, bkey_i_to_s_c(k), + __btree_node_type(level, btree_id), write, &buf); + + mustfix_fsck_err(c, journal_entry_bkey_invalid, + "%s", buf.buf); le16_add_cpu(&entry->u64s, -((u16) k->k.u64s)); memmove(k, bkey_next(k), next - (void *) bkey_next(k)); journal_entry_null_range(vstruct_next(entry), next); + + printbuf_exit(&buf); return FSCK_DELETED_KEY; } if (write) - bch2_bkey_compat(level, btree_id, version, - JSET_BIG_ENDIAN(jset), write, - NULL, bkey_to_packed(k)); + bch2_bkey_compat(level, btree_id, version, big_endian, + write, NULL, bkey_to_packed(k)); fsck_err: + printbuf_exit(&buf); return ret; } -static int journal_entry_validate_btree_keys(struct bch_fs *c, - struct jset *jset, - struct jset_entry *entry, - int write) +static int journal_entry_btree_keys_validate(struct bch_fs *c, + struct jset *jset, + struct jset_entry *entry, + unsigned version, int big_endian, + enum bkey_invalid_flags flags) { struct bkey_i *k = entry->start; @@ -274,7 +356,8 @@ static int journal_entry_validate_btree_keys(struct bch_fs *c, int ret = journal_validate_key(c, jset, entry, entry->level, entry->btree_id, - k, "key", write); + k, version, big_endian, + flags|BKEY_INVALID_JOURNAL); if (ret == FSCK_DELETED_KEY) continue; @@ -284,16 +367,36 @@ static int journal_entry_validate_btree_keys(struct bch_fs *c, return 0; } -static int journal_entry_validate_btree_root(struct bch_fs *c, - struct jset *jset, - struct jset_entry *entry, - int write) +static void journal_entry_btree_keys_to_text(struct printbuf *out, struct bch_fs *c, + struct jset_entry *entry) +{ + struct bkey_i *k; + bool first = true; + + jset_entry_for_each_key(entry, k) { + if (!first) { + prt_newline(out); + prt_printf(out, "%s: ", bch2_jset_entry_types[entry->type]); + } + prt_printf(out, "btree=%s l=%u ", bch2_btree_id_str(entry->btree_id), entry->level); + bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(k)); + first = false; + } +} + +static int journal_entry_btree_root_validate(struct bch_fs *c, + struct jset *jset, + struct jset_entry *entry, + unsigned version, int big_endian, + enum bkey_invalid_flags flags) { struct bkey_i *k = entry->start; int ret = 0; if (journal_entry_err_on(!entry->u64s || - le16_to_cpu(entry->u64s) != k->k.u64s, c, + le16_to_cpu(entry->u64s) != k->k.u64s, + c, version, jset, entry, + journal_entry_btree_root_bad_size, "invalid btree root journal entry: wrong number of keys")) { void *next = vstruct_next(entry); /* @@ -306,29 +409,46 @@ static int journal_entry_validate_btree_root(struct bch_fs *c, return 0; } - return journal_validate_key(c, jset, entry, 1, entry->btree_id, k, - "btree root", write); + ret = journal_validate_key(c, jset, entry, 1, entry->btree_id, k, + version, big_endian, flags); + if (ret == FSCK_DELETED_KEY) + ret = 0; fsck_err: return ret; } -static int journal_entry_validate_prio_ptrs(struct bch_fs *c, - struct jset *jset, - struct jset_entry *entry, - int write) +static void journal_entry_btree_root_to_text(struct printbuf *out, struct bch_fs *c, + struct jset_entry *entry) +{ + journal_entry_btree_keys_to_text(out, c, entry); +} + +static int journal_entry_prio_ptrs_validate(struct bch_fs *c, + struct jset *jset, + struct jset_entry *entry, + unsigned version, int big_endian, + enum bkey_invalid_flags flags) { /* obsolete, don't care: */ return 0; } -static int journal_entry_validate_blacklist(struct bch_fs *c, - struct jset *jset, - struct jset_entry *entry, - int write) +static void journal_entry_prio_ptrs_to_text(struct printbuf *out, struct bch_fs *c, + struct jset_entry *entry) +{ +} + +static int journal_entry_blacklist_validate(struct bch_fs *c, + struct jset *jset, + struct jset_entry *entry, + unsigned version, int big_endian, + enum bkey_invalid_flags flags) { int ret = 0; - if (journal_entry_err_on(le16_to_cpu(entry->u64s) != 1, c, + if (journal_entry_err_on(le16_to_cpu(entry->u64s) != 1, + c, version, jset, entry, + journal_entry_blacklist_bad_size, "invalid journal seq blacklist entry: bad size")) { journal_entry_null_range(entry, vstruct_next(entry)); } @@ -336,15 +456,27 @@ fsck_err: return ret; } -static int journal_entry_validate_blacklist_v2(struct bch_fs *c, - struct jset *jset, - struct jset_entry *entry, - int write) +static void journal_entry_blacklist_to_text(struct printbuf *out, struct bch_fs *c, + struct jset_entry *entry) +{ + struct jset_entry_blacklist *bl = + container_of(entry, struct jset_entry_blacklist, entry); + + prt_printf(out, "seq=%llu", le64_to_cpu(bl->seq)); +} + +static int journal_entry_blacklist_v2_validate(struct bch_fs *c, + struct jset *jset, + struct jset_entry *entry, + unsigned version, int big_endian, + enum bkey_invalid_flags flags) { struct jset_entry_blacklist_v2 *bl_entry; int ret = 0; - if (journal_entry_err_on(le16_to_cpu(entry->u64s) != 2, c, + if (journal_entry_err_on(le16_to_cpu(entry->u64s) != 2, + c, version, jset, entry, + journal_entry_blacklist_v2_bad_size, "invalid journal seq blacklist entry: bad size")) { journal_entry_null_range(entry, vstruct_next(entry)); goto out; @@ -353,7 +485,9 @@ static int journal_entry_validate_blacklist_v2(struct bch_fs *c, bl_entry = container_of(entry, struct jset_entry_blacklist_v2, entry); if (journal_entry_err_on(le64_to_cpu(bl_entry->start) > - le64_to_cpu(bl_entry->end), c, + le64_to_cpu(bl_entry->end), + c, version, jset, entry, + journal_entry_blacklist_v2_start_past_end, "invalid journal seq blacklist entry: start > end")) { journal_entry_null_range(entry, vstruct_next(entry)); } @@ -362,10 +496,22 @@ fsck_err: return ret; } -static int journal_entry_validate_usage(struct bch_fs *c, - struct jset *jset, - struct jset_entry *entry, - int write) +static void journal_entry_blacklist_v2_to_text(struct printbuf *out, struct bch_fs *c, + struct jset_entry *entry) +{ + struct jset_entry_blacklist_v2 *bl = + container_of(entry, struct jset_entry_blacklist_v2, entry); + + prt_printf(out, "start=%llu end=%llu", + le64_to_cpu(bl->start), + le64_to_cpu(bl->end)); +} + +static int journal_entry_usage_validate(struct bch_fs *c, + struct jset *jset, + struct jset_entry *entry, + unsigned version, int big_endian, + enum bkey_invalid_flags flags) { struct jset_entry_usage *u = container_of(entry, struct jset_entry_usage, entry); @@ -373,7 +519,8 @@ static int journal_entry_validate_usage(struct bch_fs *c, int ret = 0; if (journal_entry_err_on(bytes < sizeof(*u), - c, + c, version, jset, entry, + journal_entry_usage_bad_size, "invalid journal entry usage: bad size")) { journal_entry_null_range(entry, vstruct_next(entry)); return ret; @@ -383,21 +530,138 @@ fsck_err: return ret; } -static int journal_entry_validate_data_usage(struct bch_fs *c, - struct jset *jset, - struct jset_entry *entry, - int write) +static void journal_entry_usage_to_text(struct printbuf *out, struct bch_fs *c, + struct jset_entry *entry) +{ + struct jset_entry_usage *u = + container_of(entry, struct jset_entry_usage, entry); + + prt_printf(out, "type=%s v=%llu", + bch2_fs_usage_types[u->entry.btree_id], + le64_to_cpu(u->v)); +} + +static int journal_entry_data_usage_validate(struct bch_fs *c, + struct jset *jset, + struct jset_entry *entry, + unsigned version, int big_endian, + enum bkey_invalid_flags flags) { struct jset_entry_data_usage *u = container_of(entry, struct jset_entry_data_usage, entry); unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64); + struct printbuf err = PRINTBUF; int ret = 0; if (journal_entry_err_on(bytes < sizeof(*u) || bytes < sizeof(*u) + u->r.nr_devs, - c, + c, version, jset, entry, + journal_entry_data_usage_bad_size, "invalid journal entry usage: bad size")) { journal_entry_null_range(entry, vstruct_next(entry)); + goto out; + } + + if (journal_entry_err_on(bch2_replicas_entry_validate(&u->r, c->disk_sb.sb, &err), + c, version, jset, entry, + journal_entry_data_usage_bad_size, + "invalid journal entry usage: %s", err.buf)) { + journal_entry_null_range(entry, vstruct_next(entry)); + goto out; + } +out: +fsck_err: + printbuf_exit(&err); + return ret; +} + +static void journal_entry_data_usage_to_text(struct printbuf *out, struct bch_fs *c, + struct jset_entry *entry) +{ + struct jset_entry_data_usage *u = + container_of(entry, struct jset_entry_data_usage, entry); + + bch2_replicas_entry_to_text(out, &u->r); + prt_printf(out, "=%llu", le64_to_cpu(u->v)); +} + +static int journal_entry_clock_validate(struct bch_fs *c, + struct jset *jset, + struct jset_entry *entry, + unsigned version, int big_endian, + enum bkey_invalid_flags flags) +{ + struct jset_entry_clock *clock = + container_of(entry, struct jset_entry_clock, entry); + unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64); + int ret = 0; + + if (journal_entry_err_on(bytes != sizeof(*clock), + c, version, jset, entry, + journal_entry_clock_bad_size, + "bad size")) { + journal_entry_null_range(entry, vstruct_next(entry)); + return ret; + } + + if (journal_entry_err_on(clock->rw > 1, + c, version, jset, entry, + journal_entry_clock_bad_rw, + "bad rw")) { + journal_entry_null_range(entry, vstruct_next(entry)); + return ret; + } + +fsck_err: + return ret; +} + +static void journal_entry_clock_to_text(struct printbuf *out, struct bch_fs *c, + struct jset_entry *entry) +{ + struct jset_entry_clock *clock = + container_of(entry, struct jset_entry_clock, entry); + + prt_printf(out, "%s=%llu", clock->rw ? "write" : "read", le64_to_cpu(clock->time)); +} + +static int journal_entry_dev_usage_validate(struct bch_fs *c, + struct jset *jset, + struct jset_entry *entry, + unsigned version, int big_endian, + enum bkey_invalid_flags flags) +{ + struct jset_entry_dev_usage *u = + container_of(entry, struct jset_entry_dev_usage, entry); + unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64); + unsigned expected = sizeof(*u); + unsigned dev; + int ret = 0; + + if (journal_entry_err_on(bytes < expected, + c, version, jset, entry, + journal_entry_dev_usage_bad_size, + "bad size (%u < %u)", + bytes, expected)) { + journal_entry_null_range(entry, vstruct_next(entry)); + return ret; + } + + dev = le32_to_cpu(u->dev); + + if (journal_entry_err_on(!bch2_dev_exists2(c, dev), + c, version, jset, entry, + journal_entry_dev_usage_bad_dev, + "bad dev")) { + journal_entry_null_range(entry, vstruct_next(entry)); + return ret; + } + + if (journal_entry_err_on(u->pad, + c, version, jset, entry, + journal_entry_dev_usage_bad_pad, + "bad pad")) { + journal_entry_null_range(entry, vstruct_next(entry)); return ret; } @@ -405,44 +669,134 @@ fsck_err: return ret; } +static void journal_entry_dev_usage_to_text(struct printbuf *out, struct bch_fs *c, + struct jset_entry *entry) +{ + struct jset_entry_dev_usage *u = + container_of(entry, struct jset_entry_dev_usage, entry); + unsigned i, nr_types = jset_entry_dev_usage_nr_types(u); + + prt_printf(out, "dev=%u", le32_to_cpu(u->dev)); + + for (i = 0; i < nr_types; i++) { + if (i < BCH_DATA_NR) + prt_printf(out, " %s", bch2_data_types[i]); + else + prt_printf(out, " (unknown data type %u)", i); + prt_printf(out, ": buckets=%llu sectors=%llu fragmented=%llu", + le64_to_cpu(u->d[i].buckets), + le64_to_cpu(u->d[i].sectors), + le64_to_cpu(u->d[i].fragmented)); + } +} + +static int journal_entry_log_validate(struct bch_fs *c, + struct jset *jset, + struct jset_entry *entry, + unsigned version, int big_endian, + enum bkey_invalid_flags flags) +{ + return 0; +} + +static void journal_entry_log_to_text(struct printbuf *out, struct bch_fs *c, + struct jset_entry *entry) +{ + struct jset_entry_log *l = container_of(entry, struct jset_entry_log, entry); + unsigned bytes = vstruct_bytes(entry) - offsetof(struct jset_entry_log, d); + + prt_printf(out, "%.*s", bytes, l->d); +} + +static int journal_entry_overwrite_validate(struct bch_fs *c, + struct jset *jset, + struct jset_entry *entry, + unsigned version, int big_endian, + enum bkey_invalid_flags flags) +{ + return journal_entry_btree_keys_validate(c, jset, entry, + version, big_endian, READ); +} + +static void journal_entry_overwrite_to_text(struct printbuf *out, struct bch_fs *c, + struct jset_entry *entry) +{ + journal_entry_btree_keys_to_text(out, c, entry); +} + +static int journal_entry_write_buffer_keys_validate(struct bch_fs *c, + struct jset *jset, + struct jset_entry *entry, + unsigned version, int big_endian, + enum bkey_invalid_flags flags) +{ + return journal_entry_btree_keys_validate(c, jset, entry, + version, big_endian, READ); +} + +static void journal_entry_write_buffer_keys_to_text(struct printbuf *out, struct bch_fs *c, + struct jset_entry *entry) +{ + journal_entry_btree_keys_to_text(out, c, entry); +} + struct jset_entry_ops { int (*validate)(struct bch_fs *, struct jset *, - struct jset_entry *, int); + struct jset_entry *, unsigned, int, + enum bkey_invalid_flags); + void (*to_text)(struct printbuf *, struct bch_fs *, struct jset_entry *); }; static const struct jset_entry_ops bch2_jset_entry_ops[] = { #define x(f, nr) \ [BCH_JSET_ENTRY_##f] = (struct jset_entry_ops) { \ - .validate = journal_entry_validate_##f, \ + .validate = journal_entry_##f##_validate, \ + .to_text = journal_entry_##f##_to_text, \ }, BCH_JSET_ENTRY_TYPES() #undef x }; -static int journal_entry_validate(struct bch_fs *c, struct jset *jset, - struct jset_entry *entry, int write) +int bch2_journal_entry_validate(struct bch_fs *c, + struct jset *jset, + struct jset_entry *entry, + unsigned version, int big_endian, + enum bkey_invalid_flags flags) { return entry->type < BCH_JSET_ENTRY_NR - ? bch2_jset_entry_ops[entry->type].validate(c, jset, - entry, write) + ? bch2_jset_entry_ops[entry->type].validate(c, jset, entry, + version, big_endian, flags) : 0; } +void bch2_journal_entry_to_text(struct printbuf *out, struct bch_fs *c, + struct jset_entry *entry) +{ + if (entry->type < BCH_JSET_ENTRY_NR) { + prt_printf(out, "%s: ", bch2_jset_entry_types[entry->type]); + bch2_jset_entry_ops[entry->type].to_text(out, c, entry); + } else { + prt_printf(out, "(unknown type %u)", entry->type); + } +} + static int jset_validate_entries(struct bch_fs *c, struct jset *jset, - int write) + enum bkey_invalid_flags flags) { - struct jset_entry *entry; + unsigned version = le32_to_cpu(jset->version); int ret = 0; vstruct_for_each(jset, entry) { - if (journal_entry_err_on(vstruct_next(entry) > - vstruct_last(jset), c, + if (journal_entry_err_on(vstruct_next(entry) > vstruct_last(jset), + c, version, jset, entry, + journal_entry_past_jset_end, "journal entry extends past end of jset")) { jset->u64s = cpu_to_le32((u64 *) entry - jset->_data); break; } - ret = journal_entry_validate(c, jset, entry, write); + ret = bch2_journal_entry_validate(c, jset, entry, + version, JSET_BIG_ENDIAN(jset), flags); if (ret) break; } @@ -451,66 +805,94 @@ fsck_err: } static int jset_validate(struct bch_fs *c, + struct bch_dev *ca, + struct jset *jset, u64 sector, + enum bkey_invalid_flags flags) +{ + unsigned version; + int ret = 0; + + if (le64_to_cpu(jset->magic) != jset_magic(c)) + return JOURNAL_ENTRY_NONE; + + version = le32_to_cpu(jset->version); + if (journal_entry_err_on(!bch2_version_compatible(version), + c, version, jset, NULL, + jset_unsupported_version, + "%s sector %llu seq %llu: incompatible journal entry version %u.%u", + ca ? ca->name : c->name, + sector, le64_to_cpu(jset->seq), + BCH_VERSION_MAJOR(version), + BCH_VERSION_MINOR(version))) { + /* don't try to continue: */ + return -EINVAL; + } + + if (journal_entry_err_on(!bch2_checksum_type_valid(c, JSET_CSUM_TYPE(jset)), + c, version, jset, NULL, + jset_unknown_csum, + "%s sector %llu seq %llu: journal entry with unknown csum type %llu", + ca ? ca->name : c->name, + sector, le64_to_cpu(jset->seq), + JSET_CSUM_TYPE(jset))) + ret = JOURNAL_ENTRY_BAD; + + /* last_seq is ignored when JSET_NO_FLUSH is true */ + if (journal_entry_err_on(!JSET_NO_FLUSH(jset) && + le64_to_cpu(jset->last_seq) > le64_to_cpu(jset->seq), + c, version, jset, NULL, + jset_last_seq_newer_than_seq, + "invalid journal entry: last_seq > seq (%llu > %llu)", + le64_to_cpu(jset->last_seq), + le64_to_cpu(jset->seq))) { + jset->last_seq = jset->seq; + return JOURNAL_ENTRY_BAD; + } + + ret = jset_validate_entries(c, jset, flags); +fsck_err: + return ret; +} + +static int jset_validate_early(struct bch_fs *c, struct bch_dev *ca, struct jset *jset, u64 sector, unsigned bucket_sectors_left, - unsigned sectors_read, - int write) + unsigned sectors_read) { size_t bytes = vstruct_bytes(jset); - struct bch_csum csum; unsigned version; + enum bkey_invalid_flags flags = BKEY_INVALID_JOURNAL; int ret = 0; if (le64_to_cpu(jset->magic) != jset_magic(c)) return JOURNAL_ENTRY_NONE; version = le32_to_cpu(jset->version); - if (journal_entry_err_on((version != BCH_JSET_VERSION_OLD && - version < bcachefs_metadata_version_min) || - version >= bcachefs_metadata_version_max, c, - "%s sector %llu seq %llu: unknown journal entry version %u", - ca->name, sector, le64_to_cpu(jset->seq), - version)) { + if (journal_entry_err_on(!bch2_version_compatible(version), + c, version, jset, NULL, + jset_unsupported_version, + "%s sector %llu seq %llu: unknown journal entry version %u.%u", + ca ? ca->name : c->name, + sector, le64_to_cpu(jset->seq), + BCH_VERSION_MAJOR(version), + BCH_VERSION_MINOR(version))) { /* don't try to continue: */ - return EINVAL; + return -EINVAL; } if (bytes > (sectors_read << 9) && sectors_read < bucket_sectors_left) return JOURNAL_ENTRY_REREAD; - if (journal_entry_err_on(bytes > bucket_sectors_left << 9, c, + if (journal_entry_err_on(bytes > bucket_sectors_left << 9, + c, version, jset, NULL, + jset_past_bucket_end, "%s sector %llu seq %llu: journal entry too big (%zu bytes)", - ca->name, sector, le64_to_cpu(jset->seq), bytes)) { - ret = JOURNAL_ENTRY_BAD; + ca ? ca->name : c->name, + sector, le64_to_cpu(jset->seq), bytes)) le32_add_cpu(&jset->u64s, -((bytes - (bucket_sectors_left << 9)) / 8)); - } - - if (fsck_err_on(!bch2_checksum_type_valid(c, JSET_CSUM_TYPE(jset)), c, - "%s sector %llu seq %llu: journal entry with unknown csum type %llu", - ca->name, sector, le64_to_cpu(jset->seq), - JSET_CSUM_TYPE(jset))) { - ret = JOURNAL_ENTRY_BAD; - goto bad_csum_type; - } - - csum = csum_vstruct(c, JSET_CSUM_TYPE(jset), journal_nonce(jset), jset); - if (journal_entry_err_on(bch2_crc_cmp(csum, jset->csum), c, - "%s sector %llu seq %llu: journal checksum bad", - ca->name, sector, le64_to_cpu(jset->seq))) - ret = JOURNAL_ENTRY_BAD; - - bch2_encrypt(c, JSET_CSUM_TYPE(jset), journal_nonce(jset), - jset->encrypted_start, - vstruct_end(jset) - (void *) jset->encrypted_start); -bad_csum_type: - if (journal_entry_err_on(le64_to_cpu(jset->last_seq) > le64_to_cpu(jset->seq), c, - "invalid journal entry: last_seq > seq")) { - jset->last_seq = jset->seq; - return JOURNAL_ENTRY_BAD; - } fsck_err: return ret; } @@ -527,12 +909,12 @@ static int journal_read_buf_realloc(struct journal_read_buf *b, /* the bios are sized for this many pages, max: */ if (new_size > JOURNAL_ENTRY_SIZE_MAX) - return -ENOMEM; + return -BCH_ERR_ENOMEM_journal_read_buf_realloc; new_size = roundup_pow_of_two(new_size); n = kvpmalloc(new_size, GFP_KERNEL); if (!n) - return -ENOMEM; + return -BCH_ERR_ENOMEM_journal_read_buf_realloc; kvpfree(b->data, b->size); b->data = n; @@ -551,7 +933,7 @@ static int journal_read_bucket(struct bch_dev *ca, unsigned sectors, sectors_read = 0; u64 offset = bucket_to_sector(ca, ja->buckets[bucket]), end = offset + ca->mi.bucket_size; - bool saw_bad = false; + bool saw_bad = false, csum_good; int ret = 0; pr_debug("reading %u", bucket); @@ -559,35 +941,41 @@ static int journal_read_bucket(struct bch_dev *ca, while (offset < end) { if (!sectors_read) { struct bio *bio; + unsigned nr_bvecs; reread: sectors_read = min_t(unsigned, end - offset, buf->size >> 9); + nr_bvecs = buf_pages(buf->data, sectors_read << 9); + + bio = bio_kmalloc(nr_bvecs, GFP_KERNEL); + bio_init(bio, ca->disk_sb.bdev, bio->bi_inline_vecs, nr_bvecs, REQ_OP_READ); - bio = bio_kmalloc(GFP_KERNEL, - buf_pages(buf->data, - sectors_read << 9)); - bio_set_dev(bio, ca->disk_sb.bdev); - bio->bi_iter.bi_sector = offset; - bio_set_op_attrs(bio, REQ_OP_READ, 0); + bio->bi_iter.bi_sector = offset; bch2_bio_map(bio, buf->data, sectors_read << 9); ret = submit_bio_wait(bio); - bio_put(bio); + kfree(bio); - if (bch2_dev_io_err_on(ret, ca, + if (bch2_dev_io_err_on(ret, ca, BCH_MEMBER_ERROR_read, "journal read error: sector %llu", offset) || - bch2_meta_read_fault("journal")) - return -EIO; + bch2_meta_read_fault("journal")) { + /* + * We don't error out of the recovery process + * here, since the relevant journal entry may be + * found on a different device, and missing or + * no journal entries will be handled later + */ + return 0; + } j = buf->data; } - ret = jset_validate(c, ca, j, offset, - end - offset, sectors_read, - READ); + ret = jset_validate_early(c, ca, j, offset, + end - offset, sectors_read); switch (ret) { - case BCH_FSCK_OK: + case 0: sectors = vstruct_sectors(j, c->block_bits); break; case JOURNAL_ENTRY_REREAD: @@ -601,17 +989,13 @@ reread: case JOURNAL_ENTRY_NONE: if (!saw_bad) return 0; - sectors = c->opts.block_size; - goto next_block; - case JOURNAL_ENTRY_BAD: - saw_bad = true; /* * On checksum error we don't really trust the size * field of the journal entry we read, so try reading * again at next block boundary: */ - sectors = c->opts.block_size; - break; + sectors = block_sectors(c); + goto next_block; default: return ret; } @@ -627,8 +1011,26 @@ reread: ja->bucket_seq[bucket] = le64_to_cpu(j->seq); + csum_good = jset_csum_good(c, j); + if (bch2_dev_io_err_on(!csum_good, ca, BCH_MEMBER_ERROR_checksum, + "journal checksum error")) + saw_bad = true; + + ret = bch2_encrypt(c, JSET_CSUM_TYPE(j), journal_nonce(j), + j->encrypted_start, + vstruct_end(j) - (void *) j->encrypted_start); + bch2_fs_fatal_err_on(ret, c, + "error decrypting journal entry: %i", ret); + mutex_lock(&jlist->lock); - ret = journal_entry_add(c, ca, jlist, j, ret != 0); + ret = journal_entry_add(c, ca, (struct journal_ptr) { + .csum_good = csum_good, + .dev = ca->dev_idx, + .bucket = bucket, + .bucket_offset = offset - + bucket_to_sector(ca, ja->buckets[bucket]), + .sector = offset, + }, jlist, j); mutex_unlock(&jlist->lock); switch (ret) { @@ -649,17 +1051,18 @@ next_block: return 0; } -static void bch2_journal_read_device(struct closure *cl) +static CLOSURE_CALLBACK(bch2_journal_read_device) { - struct journal_device *ja = - container_of(cl, struct journal_device, read); + closure_type(ja, struct journal_device, read); struct bch_dev *ca = container_of(ja, struct bch_dev, journal); + struct bch_fs *c = ca->fs; struct journal_list *jlist = container_of(cl->parent, struct journal_list, cl); + struct journal_replay *r, **_r; + struct genradix_iter iter; struct journal_read_buf buf = { NULL, 0 }; - u64 min_seq = U64_MAX; unsigned i; - int ret; + int ret = 0; if (!ja->nr) goto out; @@ -676,35 +1079,57 @@ static void bch2_journal_read_device(struct closure *cl) goto err; } - /* Find the journal bucket with the highest sequence number: */ - for (i = 0; i < ja->nr; i++) { - if (ja->bucket_seq[i] > ja->bucket_seq[ja->cur_idx]) - ja->cur_idx = i; + ja->sectors_free = ca->mi.bucket_size; - min_seq = min(ja->bucket_seq[i], min_seq); - } + mutex_lock(&jlist->lock); + genradix_for_each_reverse(&c->journal_entries, iter, _r) { + r = *_r; - /* - * If there's duplicate journal entries in multiple buckets (which - * definitely isn't supposed to happen, but...) - make sure to start - * cur_idx at the last of those buckets, so we don't deadlock trying to - * allocate - */ - while (ja->bucket_seq[ja->cur_idx] > min_seq && - ja->bucket_seq[ja->cur_idx] > - ja->bucket_seq[(ja->cur_idx + 1) % ja->nr]) - ja->cur_idx = (ja->cur_idx + 1) % ja->nr; + if (!r) + continue; - ja->sectors_free = 0; + for (i = 0; i < r->nr_ptrs; i++) { + if (r->ptrs[i].dev == ca->dev_idx) { + unsigned wrote = bucket_remainder(ca, r->ptrs[i].sector) + + vstruct_sectors(&r->j, c->block_bits); - /* - * Set dirty_idx to indicate the entire journal is full and needs to be - * reclaimed - journal reclaim will immediately reclaim whatever isn't - * pinned when it first runs: - */ + ja->cur_idx = r->ptrs[i].bucket; + ja->sectors_free = ca->mi.bucket_size - wrote; + goto found; + } + } + } +found: + mutex_unlock(&jlist->lock); + + if (ja->bucket_seq[ja->cur_idx] && + ja->sectors_free == ca->mi.bucket_size) { +#if 0 + /* + * Debug code for ZNS support, where we (probably) want to be + * correlated where we stopped in the journal to the zone write + * points: + */ + bch_err(c, "ja->sectors_free == ca->mi.bucket_size"); + bch_err(c, "cur_idx %u/%u", ja->cur_idx, ja->nr); + for (i = 0; i < 3; i++) { + unsigned idx = (ja->cur_idx + ja->nr - 1 + i) % ja->nr; + + bch_err(c, "bucket_seq[%u] = %llu", idx, ja->bucket_seq[idx]); + } +#endif + ja->sectors_free = 0; + } + + /* + * Set dirty_idx to indicate the entire journal is full and needs to be + * reclaimed - journal reclaim will immediately reclaim whatever isn't + * pinned when it first runs: + */ ja->discard_idx = ja->dirty_idx_ondisk = ja->dirty_idx = (ja->cur_idx + 1) % ja->nr; out: + bch_verbose(c, "journal read done on device %s, ret %i", ca->name, ret); kvpfree(buf.data, buf.size); percpu_ref_put(&ca->io_ref); closure_return(cl); @@ -716,30 +1141,52 @@ err: goto out; } -int bch2_journal_read(struct bch_fs *c, struct list_head *list, - u64 *blacklist_seq, u64 *start_seq) +void bch2_journal_ptrs_to_text(struct printbuf *out, struct bch_fs *c, + struct journal_replay *j) +{ + unsigned i; + + for (i = 0; i < j->nr_ptrs; i++) { + struct bch_dev *ca = bch_dev_bkey_exists(c, j->ptrs[i].dev); + u64 offset; + + div64_u64_rem(j->ptrs[i].sector, ca->mi.bucket_size, &offset); + + if (i) + prt_printf(out, " "); + prt_printf(out, "%u:%u:%u (sector %llu)", + j->ptrs[i].dev, + j->ptrs[i].bucket, + j->ptrs[i].bucket_offset, + j->ptrs[i].sector); + } +} + +int bch2_journal_read(struct bch_fs *c, + u64 *last_seq, + u64 *blacklist_seq, + u64 *start_seq) { struct journal_list jlist; - struct journal_replay *i, *t; - struct bch_dev *ca; - unsigned iter; - size_t keys = 0, entries = 0; - bool degraded = false; - u64 seq, last_seq = 0; + struct journal_replay *i, **_i, *prev = NULL; + struct genradix_iter radix_iter; + struct printbuf buf = PRINTBUF; + bool degraded = false, last_write_torn = false; + u64 seq; int ret = 0; closure_init_stack(&jlist.cl); mutex_init(&jlist.lock); - jlist.head = list; + jlist.last_seq = 0; jlist.ret = 0; - for_each_member_device(ca, c, iter) { - if (!test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) && + for_each_member_device(c, ca) { + if (!c->opts.fsck && !(bch2_dev_has_data(c, ca) & (1 << BCH_DATA_journal))) continue; - if ((ca->mi.state == BCH_MEMBER_STATE_RW || - ca->mi.state == BCH_MEMBER_STATE_RO) && + if ((ca->mi.state == BCH_MEMBER_STATE_rw || + ca->mi.state == BCH_MEMBER_STATE_ro) && percpu_ref_tryget(&ca->io_ref)) closure_call(&ca->journal.read, bch2_journal_read_device, @@ -754,65 +1201,101 @@ int bch2_journal_read(struct bch_fs *c, struct list_head *list, if (jlist.ret) return jlist.ret; - if (list_empty(list)) { - bch_info(c, "journal read done, but no entries found"); - return 0; - } - - i = list_last_entry(list, struct journal_replay, list); - *start_seq = le64_to_cpu(i->j.seq) + 1; + *last_seq = 0; + *start_seq = 0; + *blacklist_seq = 0; /* * Find most recent flush entry, and ignore newer non flush entries - * those entries will be blacklisted: */ - list_for_each_entry_safe_reverse(i, t, list, list) { - if (i->ignore) + genradix_for_each_reverse(&c->journal_entries, radix_iter, _i) { + enum bkey_invalid_flags flags = BKEY_INVALID_JOURNAL; + + i = *_i; + + if (!i || i->ignore) continue; - if (!JSET_NO_FLUSH(&i->j)) { - last_seq = le64_to_cpu(i->j.last_seq); - *blacklist_seq = le64_to_cpu(i->j.seq) + 1; - break; + if (!*start_seq) + *blacklist_seq = *start_seq = le64_to_cpu(i->j.seq) + 1; + + if (JSET_NO_FLUSH(&i->j)) { + i->ignore = true; + continue; } - journal_replay_free(c, i); + if (!last_write_torn && !i->csum_good) { + last_write_torn = true; + i->ignore = true; + continue; + } + + if (journal_entry_err_on(le64_to_cpu(i->j.last_seq) > le64_to_cpu(i->j.seq), + c, le32_to_cpu(i->j.version), &i->j, NULL, + jset_last_seq_newer_than_seq, + "invalid journal entry: last_seq > seq (%llu > %llu)", + le64_to_cpu(i->j.last_seq), + le64_to_cpu(i->j.seq))) + i->j.last_seq = i->j.seq; + + *last_seq = le64_to_cpu(i->j.last_seq); + *blacklist_seq = le64_to_cpu(i->j.seq) + 1; + break; } - if (!last_seq) { - fsck_err(c, "journal read done, but no entries found after dropping non-flushes"); - return -1; + if (!*start_seq) { + bch_info(c, "journal read done, but no entries found"); + return 0; + } + + if (!*last_seq) { + fsck_err(c, dirty_but_no_journal_entries_post_drop_nonflushes, + "journal read done, but no entries found after dropping non-flushes"); + return 0; } + bch_info(c, "journal read done, replaying entries %llu-%llu", + *last_seq, *blacklist_seq - 1); + + if (*start_seq != *blacklist_seq) + bch_info(c, "dropped unflushed entries %llu-%llu", + *blacklist_seq, *start_seq - 1); + /* Drop blacklisted entries and entries older than last_seq: */ - list_for_each_entry_safe(i, t, list, list) { - if (i->ignore) + genradix_for_each(&c->journal_entries, radix_iter, _i) { + i = *_i; + + if (!i || i->ignore) continue; seq = le64_to_cpu(i->j.seq); - if (seq < last_seq) { + if (seq < *last_seq) { journal_replay_free(c, i); continue; } if (bch2_journal_seq_is_blacklisted(c, seq, true)) { fsck_err_on(!JSET_NO_FLUSH(&i->j), c, + jset_seq_blacklisted, "found blacklisted journal entry %llu", seq); - - journal_replay_free(c, i); + i->ignore = true; } } /* Check for missing entries: */ - seq = last_seq; - list_for_each_entry(i, list, list) { - if (i->ignore) + seq = *last_seq; + genradix_for_each(&c->journal_entries, radix_iter, _i) { + i = *_i; + + if (!i || i->ignore) continue; BUG_ON(seq > le64_to_cpu(i->j.seq)); while (seq < le64_to_cpu(i->j.seq)) { u64 missing_start, missing_end; + struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF; while (seq < le64_to_cpu(i->j.seq) && bch2_journal_seq_is_blacklisted(c, seq, false)) @@ -827,58 +1310,81 @@ int bch2_journal_read(struct bch_fs *c, struct list_head *list, !bch2_journal_seq_is_blacklisted(c, seq, false)) seq++; + if (prev) { + bch2_journal_ptrs_to_text(&buf1, c, prev); + prt_printf(&buf1, " size %zu", vstruct_sectors(&prev->j, c->block_bits)); + } else + prt_printf(&buf1, "(none)"); + bch2_journal_ptrs_to_text(&buf2, c, i); + missing_end = seq - 1; - fsck_err(c, "journal entries %llu-%llu missing! (replaying %llu-%llu)", + fsck_err(c, journal_entries_missing, + "journal entries %llu-%llu missing! (replaying %llu-%llu)\n" + " prev at %s\n" + " next at %s", missing_start, missing_end, - last_seq, *blacklist_seq - 1); + *last_seq, *blacklist_seq - 1, + buf1.buf, buf2.buf); + + printbuf_exit(&buf1); + printbuf_exit(&buf2); } + prev = i; seq++; } - list_for_each_entry(i, list, list) { - struct jset_entry *entry; - struct bkey_i *k, *_n; - struct bch_replicas_padded replicas; - char buf[80]; + genradix_for_each(&c->journal_entries, radix_iter, _i) { + struct bch_replicas_padded replicas = { + .e.data_type = BCH_DATA_journal, + .e.nr_required = 1, + }; + unsigned ptr; - if (i->ignore) + i = *_i; + if (!i || i->ignore) continue; - ret = jset_validate_entries(c, &i->j, READ); + for (ptr = 0; ptr < i->nr_ptrs; ptr++) { + struct bch_dev *ca = bch_dev_bkey_exists(c, i->ptrs[ptr].dev); + + if (!i->ptrs[ptr].csum_good) + bch_err_dev_offset(ca, i->ptrs[ptr].sector, + "invalid journal checksum, seq %llu%s", + le64_to_cpu(i->j.seq), + i->csum_good ? " (had good copy on another device)" : ""); + } + + ret = jset_validate(c, + bch_dev_bkey_exists(c, i->ptrs[0].dev), + &i->j, + i->ptrs[0].sector, + READ); if (ret) - goto fsck_err; + goto err; - /* - * If we're mounting in degraded mode - if we didn't read all - * the devices - this is wrong: - */ + for (ptr = 0; ptr < i->nr_ptrs; ptr++) + replicas.e.devs[replicas.e.nr_devs++] = i->ptrs[ptr].dev; + + bch2_replicas_entry_sort(&replicas.e); - bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal, i->devs); + printbuf_reset(&buf); + bch2_replicas_entry_to_text(&buf, &replicas.e); if (!degraded && - (test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) || - fsck_err_on(!bch2_replicas_marked(c, &replicas.e), c, - "superblock not marked as containing replicas %s", - (bch2_replicas_entry_to_text(&PBUF(buf), - &replicas.e), buf)))) { + !bch2_replicas_marked(c, &replicas.e) && + (le64_to_cpu(i->j.seq) == *last_seq || + fsck_err(c, journal_entry_replicas_not_marked, + "superblock not marked as containing replicas for journal entry %llu\n %s", + le64_to_cpu(i->j.seq), buf.buf))) { ret = bch2_mark_replicas(c, &replicas.e); if (ret) - return ret; + goto err; } - - for_each_jset_key(k, _n, entry, &i->j) - keys++; - entries++; } - - bch_info(c, "journal read done, %zu keys in %zu entries, seq %llu", - keys, entries, *start_seq); - - if (*start_seq != *blacklist_seq) - bch_info(c, "dropped unflushed entries %llu-%llu", - *blacklist_seq, *start_seq - 1); +err: fsck_err: + printbuf_exit(&buf); return ret; } @@ -911,10 +1417,9 @@ static void __journal_write_alloc(struct journal *j, * it: */ if (!ca->mi.durability || - ca->mi.state != BCH_MEMBER_STATE_RW || + ca->mi.state != BCH_MEMBER_STATE_rw || !ja->nr || - bch2_bkey_has_device(bkey_i_to_s_c(&w->key), - ca->dev_idx) || + bch2_bkey_has_device_c(bkey_i_to_s_c(&w->key), ca->dev_idx) || sectors > ja->sectors_free) continue; @@ -940,22 +1445,31 @@ static void __journal_write_alloc(struct journal *j, } /** - * journal_next_bucket - move on to the next journal bucket if possible + * journal_write_alloc - decide where to write next journal entry + * + * @j: journal object + * @w: journal buf (entry to be written) + * + * Returns: 0 on success, or -EROFS on failure */ -static int journal_write_alloc(struct journal *j, struct journal_buf *w, - unsigned sectors) +static int journal_write_alloc(struct journal *j, struct journal_buf *w) { struct bch_fs *c = container_of(j, struct bch_fs, journal); + struct bch_devs_mask devs; struct journal_device *ja; struct bch_dev *ca; struct dev_alloc_list devs_sorted; + unsigned sectors = vstruct_sectors(w->data, c->block_bits); + unsigned target = c->opts.metadata_target ?: + c->opts.foreground_target; unsigned i, replicas = 0, replicas_want = READ_ONCE(c->opts.metadata_replicas); rcu_read_lock(); +retry: + devs = target_rw_devs(c, BCH_DATA_journal, target); - devs_sorted = bch2_dev_alloc_list(c, &j->wp.stripe, - &c->rw_devs[BCH_DATA_journal]); + devs_sorted = bch2_dev_alloc_list(c, &j->wp.stripe, &devs); __journal_write_alloc(j, w, &devs_sorted, sectors, &replicas, replicas_want); @@ -987,57 +1501,24 @@ static int journal_write_alloc(struct journal *j, struct journal_buf *w, __journal_write_alloc(j, w, &devs_sorted, sectors, &replicas, replicas_want); + + if (replicas < replicas_want && target) { + /* Retry from all devices: */ + target = 0; + goto retry; + } done: rcu_read_unlock(); - return replicas >= c->opts.metadata_replicas_required ? 0 : -EROFS; -} - -static void journal_write_compact(struct jset *jset) -{ - struct jset_entry *i, *next, *prev = NULL; - - /* - * Simple compaction, dropping empty jset_entries (from journal - * reservations that weren't fully used) and merging jset_entries that - * can be. - * - * If we wanted to be really fancy here, we could sort all the keys in - * the jset and drop keys that were overwritten - probably not worth it: - */ - vstruct_for_each_safe(jset, i, next) { - unsigned u64s = le16_to_cpu(i->u64s); - - /* Empty entry: */ - if (!u64s) - continue; - - /* Can we merge with previous entry? */ - if (prev && - i->btree_id == prev->btree_id && - i->level == prev->level && - i->type == prev->type && - i->type == BCH_JSET_ENTRY_btree_keys && - le16_to_cpu(prev->u64s) + u64s <= U16_MAX) { - memmove_u64s_down(vstruct_next(prev), - i->_data, - u64s); - le16_add_cpu(&prev->u64s, u64s); - continue; - } - - /* Couldn't merge, move i into new position (after prev): */ - prev = prev ? vstruct_next(prev) : jset->start; - if (i != prev) - memmove_u64s_down(prev, i, jset_u64s(u64s)); - } + BUG_ON(bkey_val_u64s(&w->key.k) > BCH_REPLICAS_MAX); - prev = prev ? vstruct_next(prev) : jset->start; - jset->u64s = cpu_to_le32((u64 *) prev - jset->_data); + return replicas >= c->opts.metadata_replicas_required ? 0 : -EROFS; } static void journal_buf_realloc(struct journal *j, struct journal_buf *buf) { + struct bch_fs *c = container_of(j, struct bch_fs, journal); + /* we aren't holding j->lock: */ unsigned new_size = READ_ONCE(j->buf_size_want); void *new_buf; @@ -1045,42 +1526,50 @@ static void journal_buf_realloc(struct journal *j, struct journal_buf *buf) if (buf->buf_size >= new_size) return; - new_buf = kvpmalloc(new_size, GFP_NOIO|__GFP_NOWARN); + size_t btree_write_buffer_size = new_size / 64; + + if (bch2_btree_write_buffer_resize(c, btree_write_buffer_size)) + return; + + new_buf = kvpmalloc(new_size, GFP_NOFS|__GFP_NOWARN); if (!new_buf) return; memcpy(new_buf, buf->data, buf->buf_size); - kvpfree(buf->data, buf->buf_size); - buf->data = new_buf; - buf->buf_size = new_size; + + spin_lock(&j->lock); + swap(buf->data, new_buf); + swap(buf->buf_size, new_size); + spin_unlock(&j->lock); + + kvpfree(new_buf, new_size); } static inline struct journal_buf *journal_last_unwritten_buf(struct journal *j) { - return j->buf + j->reservations.unwritten_idx; + return j->buf + (journal_last_unwritten_seq(j) & JOURNAL_BUF_MASK); } -static void journal_write_done(struct closure *cl) +static CLOSURE_CALLBACK(journal_write_done) { - struct journal *j = container_of(cl, struct journal, io); + closure_type(j, struct journal, io); struct bch_fs *c = container_of(j, struct bch_fs, journal); struct journal_buf *w = journal_last_unwritten_buf(j); - struct bch_devs_list devs = - bch2_bkey_devs(bkey_i_to_s_c(&w->key)); struct bch_replicas_padded replicas; union journal_res_state old, new; - u64 seq = le64_to_cpu(w->data->seq); - u64 last_seq = le64_to_cpu(w->data->last_seq); - u64 v; + u64 v, seq; int err = 0; - bch2_time_stats_update(j->write_time, j->write_start_time); + bch2_time_stats_update(!JSET_NO_FLUSH(w->data) + ? j->flush_write_time + : j->noflush_write_time, j->write_start_time); - if (!devs.nr) { + if (!w->devs_written.nr) { bch_err(c, "unable to write journal to sufficient devices"); err = -EIO; } else { - bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal, devs); + bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal, + w->devs_written); if (bch2_mark_replicas(c, &replicas.e)) err = -EIO; } @@ -1089,18 +1578,25 @@ static void journal_write_done(struct closure *cl) bch2_fatal_error(c); spin_lock(&j->lock); + seq = le64_to_cpu(w->data->seq); + if (seq >= j->pin.front) - journal_seq_pin(j, seq)->devs = devs; + journal_seq_pin(j, seq)->devs = w->devs_written; - j->seq_ondisk = seq; - if (err && (!j->err_seq || seq < j->err_seq)) + if (!err) { + if (!JSET_NO_FLUSH(w->data)) { + j->flushed_seq_ondisk = seq; + j->last_seq_ondisk = w->last_seq; + + bch2_do_discards(c); + closure_wake_up(&c->freelist_wait); + + bch2_reset_alloc_cursors(c); + } + } else if (!j->err_seq || seq < j->err_seq) j->err_seq = seq; - if (!w->noflush) { - j->flushed_seq_ondisk = seq; - j->last_seq_ondisk = last_seq; - bch2_journal_space_available(j); - } + j->seq_ondisk = seq; /* * Updating last_seq_ondisk may let bch2_journal_reclaim_work() discard @@ -1109,7 +1605,8 @@ static void journal_write_done(struct closure *cl) * Must come before signaling write completion, for * bch2_fs_journal_stop(): */ - journal_reclaim_kick(&c->journal); + if (j->watermark != BCH_WATERMARK_stripe) + journal_reclaim_kick(&c->journal); /* also must come before signalling write completion: */ closure_debug_destroy(cl); @@ -1117,37 +1614,57 @@ static void journal_write_done(struct closure *cl) v = atomic64_read(&j->reservations.counter); do { old.v = new.v = v; - BUG_ON(new.idx == new.unwritten_idx); + BUG_ON(journal_state_count(new, new.unwritten_idx)); new.unwritten_idx++; } while ((v = atomic64_cmpxchg(&j->reservations.counter, old.v, new.v)) != old.v); + bch2_journal_reclaim_fast(j); + bch2_journal_space_available(j); + + track_event_change(&c->times[BCH_TIME_blocked_journal_max_in_flight], + &j->max_in_flight_start, false); + closure_wake_up(&w->wait); journal_wake(j); - if (test_bit(JOURNAL_NEED_WRITE, &j->flags)) - mod_delayed_work(system_freezable_wq, &j->write_work, 0); - spin_unlock(&j->lock); + if (!journal_state_count(new, new.unwritten_idx) && + journal_last_unwritten_seq(j) <= journal_cur_seq(j)) { + spin_unlock(&j->lock); + closure_call(&j->io, bch2_journal_write, c->io_complete_wq, NULL); + } else if (journal_last_unwritten_seq(j) == journal_cur_seq(j) && + new.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL) { + struct journal_buf *buf = journal_cur_buf(j); + long delta = buf->expires - jiffies; + + /* + * We don't close a journal entry to write it while there's + * previous entries still in flight - the current journal entry + * might want to be written now: + */ - if (new.unwritten_idx != new.idx && - !journal_state_count(new, new.unwritten_idx)) - closure_call(&j->io, bch2_journal_write, system_highpri_wq, NULL); + spin_unlock(&j->lock); + mod_delayed_work(c->io_complete_wq, &j->write_work, max(0L, delta)); + } else { + spin_unlock(&j->lock); + } } static void journal_write_endio(struct bio *bio) { struct bch_dev *ca = bio->bi_private; struct journal *j = &ca->fs->journal; + struct journal_buf *w = journal_last_unwritten_buf(j); + unsigned long flags; - if (bch2_dev_io_err_on(bio->bi_status, ca, "journal write error: %s", + if (bch2_dev_io_err_on(bio->bi_status, ca, BCH_MEMBER_ERROR_write, + "error writing journal entry %llu: %s", + le64_to_cpu(w->data->seq), bch2_blk_status_to_str(bio->bi_status)) || bch2_meta_write_fault("journal")) { - struct journal_buf *w = journal_last_unwritten_buf(j); - unsigned long flags; - spin_lock_irqsave(&j->err_lock, flags); - bch2_bkey_drop_device(bkey_i_to_s(&w->key), ca->dev_idx); + bch2_dev_list_drop_dev(&w->devs_written, ca->dev_idx); spin_unlock_irqrestore(&j->err_lock, flags); } @@ -1155,118 +1672,267 @@ static void journal_write_endio(struct bio *bio) percpu_ref_put(&ca->io_ref); } -void bch2_journal_write(struct closure *cl) +static CLOSURE_CALLBACK(do_journal_write) { - struct journal *j = container_of(cl, struct journal, io); + closure_type(j, struct journal, io); struct bch_fs *c = container_of(j, struct bch_fs, journal); struct bch_dev *ca; struct journal_buf *w = journal_last_unwritten_buf(j); - struct jset_entry *start, *end; - struct jset *jset; struct bio *bio; - struct bch_extent_ptr *ptr; - bool validate_before_checksum = false; - unsigned i, sectors, bytes, u64s; - int ret; + unsigned sectors = vstruct_sectors(w->data, c->block_bits); - BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb)); + extent_for_each_ptr(bkey_i_to_s_extent(&w->key), ptr) { + ca = bch_dev_bkey_exists(c, ptr->dev); + if (!percpu_ref_tryget(&ca->io_ref)) { + /* XXX: fix this */ + bch_err(c, "missing device for journal write\n"); + continue; + } - journal_buf_realloc(j, w); - jset = w->data; + this_cpu_add(ca->io_done->sectors[WRITE][BCH_DATA_journal], + sectors); - j->write_start_time = local_clock(); + bio = ca->journal.bio; + bio_reset(bio, ca->disk_sb.bdev, REQ_OP_WRITE|REQ_SYNC|REQ_META); + bio->bi_iter.bi_sector = ptr->offset; + bio->bi_end_io = journal_write_endio; + bio->bi_private = ca; - spin_lock(&j->lock); - if (c->sb.features & (1ULL << BCH_FEATURE_journal_no_flush) && - !w->must_flush && - (jiffies - j->last_flush_write) < msecs_to_jiffies(j->write_delay_ms) && - test_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags)) { - w->noflush = true; - SET_JSET_NO_FLUSH(jset, true); - jset->last_seq = cpu_to_le64(j->last_seq_ondisk); + BUG_ON(bio->bi_iter.bi_sector == ca->prev_journal_sector); + ca->prev_journal_sector = bio->bi_iter.bi_sector; - j->nr_noflush_writes++; - } else { - j->last_flush_write = jiffies; - j->nr_flush_writes++; + if (!JSET_NO_FLUSH(w->data)) + bio->bi_opf |= REQ_FUA; + if (!JSET_NO_FLUSH(w->data) && !w->separate_flush) + bio->bi_opf |= REQ_PREFLUSH; + + bch2_bio_map(bio, w->data, sectors << 9); + + trace_and_count(c, journal_write, bio); + closure_bio_submit(bio, cl); + + ca->journal.bucket_seq[ca->journal.cur_idx] = + le64_to_cpu(w->data->seq); } - spin_unlock(&j->lock); + + continue_at(cl, journal_write_done, c->io_complete_wq); +} + +static int bch2_journal_write_prep(struct journal *j, struct journal_buf *w) +{ + struct bch_fs *c = container_of(j, struct bch_fs, journal); + struct jset_entry *start, *end; + struct jset *jset = w->data; + struct journal_keys_to_wb wb = { NULL }; + unsigned sectors, bytes, u64s; + unsigned long btree_roots_have = 0; + bool validate_before_checksum = false; + u64 seq = le64_to_cpu(jset->seq); + int ret; /* - * New btree roots are set by journalling them; when the journal entry - * gets written we have to propagate them to c->btree_roots + * Simple compaction, dropping empty jset_entries (from journal + * reservations that weren't fully used) and merging jset_entries that + * can be. * - * But, every journal entry we write has to contain all the btree roots - * (at least for now); so after we copy btree roots to c->btree_roots we - * have to get any missing btree roots and add them to this journal - * entry: + * If we wanted to be really fancy here, we could sort all the keys in + * the jset and drop keys that were overwritten - probably not worth it: */ + vstruct_for_each(jset, i) { + unsigned u64s = le16_to_cpu(i->u64s); - bch2_journal_entries_to_btree_roots(c, jset); + /* Empty entry: */ + if (!u64s) + continue; + + /* + * New btree roots are set by journalling them; when the journal + * entry gets written we have to propagate them to + * c->btree_roots + * + * But, every journal entry we write has to contain all the + * btree roots (at least for now); so after we copy btree roots + * to c->btree_roots we have to get any missing btree roots and + * add them to this journal entry: + */ + switch (i->type) { + case BCH_JSET_ENTRY_btree_root: + bch2_journal_entry_to_btree_root(c, i); + __set_bit(i->btree_id, &btree_roots_have); + break; + case BCH_JSET_ENTRY_write_buffer_keys: + EBUG_ON(!w->need_flush_to_write_buffer); + + if (!wb.wb) + bch2_journal_keys_to_write_buffer_start(c, &wb, seq); + + struct bkey_i *k; + jset_entry_for_each_key(i, k) { + ret = bch2_journal_key_to_wb(c, &wb, i->btree_id, k); + if (ret) { + bch2_fs_fatal_error(c, "-ENOMEM flushing journal keys to btree write buffer"); + bch2_journal_keys_to_write_buffer_end(c, &wb); + return ret; + } + } + i->type = BCH_JSET_ENTRY_btree_keys; + break; + } + } + + if (wb.wb) + bch2_journal_keys_to_write_buffer_end(c, &wb); + w->need_flush_to_write_buffer = false; start = end = vstruct_last(jset); - end = bch2_btree_roots_to_journal_entries(c, jset->start, end); + end = bch2_btree_roots_to_journal_entries(c, end, btree_roots_have); - end = bch2_journal_super_entries_add_common(c, end, - le64_to_cpu(jset->seq)); + bch2_journal_super_entries_add_common(c, &end, seq); u64s = (u64 *) end - (u64 *) start; BUG_ON(u64s > j->entry_u64s_reserved); le32_add_cpu(&jset->u64s, u64s); - BUG_ON(vstruct_sectors(jset, c->block_bits) > w->sectors); - journal_write_compact(jset); + sectors = vstruct_sectors(jset, c->block_bits); + bytes = vstruct_bytes(jset); - jset->read_clock = cpu_to_le16(c->bucket_clock[READ].hand); - jset->write_clock = cpu_to_le16(c->bucket_clock[WRITE].hand); - jset->magic = cpu_to_le64(jset_magic(c)); + if (sectors > w->sectors) { + bch2_fs_fatal_error(c, "aieeee! journal write overran available space, %zu > %u (extra %u reserved %u/%u)", + vstruct_bytes(jset), w->sectors << 9, + u64s, w->u64s_reserved, j->entry_u64s_reserved); + return -EINVAL; + } - jset->version = c->sb.version < bcachefs_metadata_version_new_versioning - ? cpu_to_le32(BCH_JSET_VERSION_OLD) - : cpu_to_le32(c->sb.version); + jset->magic = cpu_to_le64(jset_magic(c)); + jset->version = cpu_to_le32(c->sb.version); SET_JSET_BIG_ENDIAN(jset, CPU_BIG_ENDIAN); SET_JSET_CSUM_TYPE(jset, bch2_meta_checksum_type(c)); - if (journal_entry_empty(jset)) - j->last_empty_seq = le64_to_cpu(jset->seq); + if (!JSET_NO_FLUSH(jset) && journal_entry_empty(jset)) + j->last_empty_seq = seq; if (bch2_csum_type_is_encryption(JSET_CSUM_TYPE(jset))) validate_before_checksum = true; - if (le32_to_cpu(jset->version) < bcachefs_metadata_version_max) + if (le32_to_cpu(jset->version) < bcachefs_metadata_version_current) validate_before_checksum = true; if (validate_before_checksum && - jset_validate_entries(c, jset, WRITE)) - goto err; + (ret = jset_validate(c, NULL, jset, 0, WRITE))) + return ret; - bch2_encrypt(c, JSET_CSUM_TYPE(jset), journal_nonce(jset), + ret = bch2_encrypt(c, JSET_CSUM_TYPE(jset), journal_nonce(jset), jset->encrypted_start, vstruct_end(jset) - (void *) jset->encrypted_start); + if (bch2_fs_fatal_err_on(ret, c, + "error decrypting journal entry: %i", ret)) + return ret; jset->csum = csum_vstruct(c, JSET_CSUM_TYPE(jset), journal_nonce(jset), jset); if (!validate_before_checksum && - jset_validate_entries(c, jset, WRITE)) - goto err; - - sectors = vstruct_sectors(jset, c->block_bits); - BUG_ON(sectors > w->sectors); + (ret = jset_validate(c, NULL, jset, 0, WRITE))) + return ret; - bytes = vstruct_bytes(jset); memset((void *) jset + bytes, 0, (sectors << 9) - bytes); + return 0; +} + +static int bch2_journal_write_pick_flush(struct journal *j, struct journal_buf *w) +{ + struct bch_fs *c = container_of(j, struct bch_fs, journal); + int error = bch2_journal_error(j); + + /* + * If the journal is in an error state - we did an emergency shutdown - + * we prefer to continue doing journal writes. We just mark them as + * noflush so they'll never be used, but they'll still be visible by the + * list_journal tool - this helps in debugging. + * + * There's a caveat: the first journal write after marking the + * superblock dirty must always be a flush write, because on startup + * from a clean shutdown we didn't necessarily read the journal and the + * new journal write might overwrite whatever was in the journal + * previously - we can't leave the journal without any flush writes in + * it. + * + * So if we're in an error state, and we're still starting up, we don't + * write anything at all. + */ + if (error && test_bit(JOURNAL_NEED_FLUSH_WRITE, &j->flags)) + return -EIO; + + if (error || + w->noflush || + (!w->must_flush && + (jiffies - j->last_flush_write) < msecs_to_jiffies(c->opts.journal_flush_delay) && + test_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags))) { + w->noflush = true; + SET_JSET_NO_FLUSH(w->data, true); + w->data->last_seq = 0; + w->last_seq = 0; + + j->nr_noflush_writes++; + } else { + j->last_flush_write = jiffies; + j->nr_flush_writes++; + clear_bit(JOURNAL_NEED_FLUSH_WRITE, &j->flags); + } + + return 0; +} + +CLOSURE_CALLBACK(bch2_journal_write) +{ + closure_type(j, struct journal, io); + struct bch_fs *c = container_of(j, struct bch_fs, journal); + struct journal_buf *w = journal_last_unwritten_buf(j); + struct bch_replicas_padded replicas; + struct bio *bio; + struct printbuf journal_debug_buf = PRINTBUF; + unsigned nr_rw_members = 0; + int ret; + + BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb)); + + j->write_start_time = local_clock(); -retry_alloc: spin_lock(&j->lock); - ret = journal_write_alloc(j, w, sectors); + ret = bch2_journal_write_pick_flush(j, w); + spin_unlock(&j->lock); + if (ret) + goto err; + + mutex_lock(&j->buf_lock); + journal_buf_realloc(j, w); + + ret = bch2_journal_write_prep(j, w); + mutex_unlock(&j->buf_lock); + if (ret) + goto err; + + j->entry_bytes_written += vstruct_bytes(w->data); + + while (1) { + spin_lock(&j->lock); + ret = journal_write_alloc(j, w); + if (!ret || !j->can_discard) + break; - if (ret && j->can_discard) { spin_unlock(&j->lock); bch2_journal_do_discards(j); - goto retry_alloc; + } + + if (ret) { + __bch2_journal_debug_to_text(&journal_debug_buf, j); + spin_unlock(&j->lock); + bch_err(c, "Unable to allocate journal write:\n%s", + journal_debug_buf.buf); + printbuf_exit(&journal_debug_buf); + goto err; } /* @@ -1282,69 +1948,45 @@ retry_alloc: bch2_journal_space_available(j); spin_unlock(&j->lock); - if (ret) { - bch_err(c, "Unable to allocate journal write"); - bch2_fatal_error(c); - continue_at(cl, journal_write_done, system_highpri_wq); - return; - } + w->devs_written = bch2_bkey_devs(bkey_i_to_s_c(&w->key)); - /* - * XXX: we really should just disable the entire journal in nochanges - * mode - */ if (c->opts.nochanges) goto no_io; - extent_for_each_ptr(bkey_i_to_s_extent(&w->key), ptr) { - ca = bch_dev_bkey_exists(c, ptr->dev); - if (!percpu_ref_tryget(&ca->io_ref)) { - /* XXX: fix this */ - bch_err(c, "missing device for journal write\n"); - continue; - } + for_each_rw_member(c, ca) + nr_rw_members++; - this_cpu_add(ca->io_done->sectors[WRITE][BCH_DATA_journal], - sectors); + if (nr_rw_members > 1) + w->separate_flush = true; - bio = ca->journal.bio; - bio_reset(bio); - bio_set_dev(bio, ca->disk_sb.bdev); - bio->bi_iter.bi_sector = ptr->offset; - bio->bi_end_io = journal_write_endio; - bio->bi_private = ca; - bio->bi_opf = REQ_OP_WRITE|REQ_SYNC|REQ_META; - if (!JSET_NO_FLUSH(jset)) - bio->bi_opf |= REQ_PREFLUSH|REQ_FUA; - bch2_bio_map(bio, jset, sectors << 9); + /* + * Mark journal replicas before we submit the write to guarantee + * recovery will find the journal entries after a crash. + */ + bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal, + w->devs_written); + ret = bch2_mark_replicas(c, &replicas.e); + if (ret) + goto err; - trace_journal_write(bio); - closure_bio_submit(bio, cl); + if (!JSET_NO_FLUSH(w->data) && w->separate_flush) { + for_each_rw_member(c, ca) { + percpu_ref_get(&ca->io_ref); - ca->journal.bucket_seq[ca->journal.cur_idx] = le64_to_cpu(jset->seq); + bio = ca->journal.bio; + bio_reset(bio, ca->disk_sb.bdev, REQ_OP_FLUSH); + bio->bi_end_io = journal_write_endio; + bio->bi_private = ca; + closure_bio_submit(bio, cl); + } } - if (!JSET_NO_FLUSH(jset)) { - for_each_rw_member(ca, c, i) - if (journal_flushes_device(ca) && - !bch2_bkey_has_device(bkey_i_to_s_c(&w->key), i)) { - percpu_ref_get(&ca->io_ref); - - bio = ca->journal.bio; - bio_reset(bio); - bio_set_dev(bio, ca->disk_sb.bdev); - bio->bi_opf = REQ_OP_FLUSH; - bio->bi_end_io = journal_write_endio; - bio->bi_private = ca; - closure_bio_submit(bio, cl); - } - } + continue_at(cl, do_journal_write, c->io_complete_wq); + return; no_io: - bch2_bucket_seq_cleanup(c); - - continue_at(cl, journal_write_done, system_highpri_wq); + continue_at(cl, journal_write_done, c->io_complete_wq); return; err: - bch2_inconsistent_error(c); - continue_at(cl, journal_write_done, system_highpri_wq); + bch2_fatal_error(c); + continue_at(cl, journal_write_done, c->io_complete_wq); }