X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libbcachefs%2Fbtree_io.c;h=d9abf90e1f8182d83117215e5bab1988755666b3;hb=5134a64cebf9cbc57d76f9e69a8627227803a061;hp=2f5097218f9c8c009f3f91b4d15f7ec7f585a9b5;hpb=cd9892e543845e045142ed29e4a5a9ce446a205e;p=bcachefs-tools-debian diff --git a/libbcachefs/btree_io.c b/libbcachefs/btree_io.c index 2f50972..d9abf90 100644 --- a/libbcachefs/btree_io.c +++ b/libbcachefs/btree_io.c @@ -14,18 +14,63 @@ #include "debug.h" #include "error.h" #include "extents.h" -#include "io.h" +#include "io_write.h" #include "journal_reclaim.h" #include "journal_seq_blacklist.h" +#include "recovery.h" #include "super-io.h" +#include "trace.h" #include -#include + +void bch2_btree_node_io_unlock(struct btree *b) +{ + EBUG_ON(!btree_node_write_in_flight(b)); + + clear_btree_node_write_in_flight_inner(b); + clear_btree_node_write_in_flight(b); + wake_up_bit(&b->flags, BTREE_NODE_write_in_flight); +} + +void bch2_btree_node_io_lock(struct btree *b) +{ + bch2_assert_btree_nodes_not_locked(); + + wait_on_bit_lock_io(&b->flags, BTREE_NODE_write_in_flight, + TASK_UNINTERRUPTIBLE); +} + +void __bch2_btree_node_wait_on_read(struct btree *b) +{ + wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight, + TASK_UNINTERRUPTIBLE); +} + +void __bch2_btree_node_wait_on_write(struct btree *b) +{ + wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight, + TASK_UNINTERRUPTIBLE); +} + +void bch2_btree_node_wait_on_read(struct btree *b) +{ + bch2_assert_btree_nodes_not_locked(); + + wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight, + TASK_UNINTERRUPTIBLE); +} + +void bch2_btree_node_wait_on_write(struct btree *b) +{ + bch2_assert_btree_nodes_not_locked(); + + wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight, + TASK_UNINTERRUPTIBLE); +} static void verify_no_dups(struct btree *b, struct bkey_packed *start, - struct bkey_packed *end, - bool extents) + struct bkey_packed *end) { #ifdef CONFIG_BCACHEFS_DEBUG struct bkey_packed *k, *p; @@ -33,16 +78,13 @@ static void verify_no_dups(struct btree *b, if (start == end) return; - for (p = start, k = bkey_next_skip_noops(start, end); + for (p = start, k = bkey_p_next(start); k != end; - p = k, k = bkey_next_skip_noops(k, end)) { + p = k, k = bkey_p_next(k)) { struct bkey l = bkey_unpack_key(b, p); struct bkey r = bkey_unpack_key(b, k); - BUG_ON(extents - ? bkey_cmp(l.p, bkey_start_pos(&r)) > 0 - : bkey_cmp(l.p, bkey_start_pos(&r)) >= 0); - //BUG_ON(bkey_cmp_packed(&b->format, p, k) >= 0); + BUG_ON(bpos_ge(l.p, bkey_start_pos(&r))); } #endif } @@ -51,9 +93,7 @@ static void set_needs_whiteout(struct bset *i, int v) { struct bkey_packed *k; - for (k = i->start; - k != vstruct_last(i); - k = bkey_next_skip_noops(k, vstruct_last(i))) + for (k = i->start; k != vstruct_last(i); k = bkey_p_next(k)) k->needs_whiteout = v; } @@ -78,7 +118,7 @@ static void *btree_bounce_alloc(struct bch_fs *c, size_t size, p = vpmalloc(size, __GFP_NOWARN|GFP_NOWAIT); if (!p) { *used_mempool = true; - p = mempool_alloc(&c->btree_bounce_pool, GFP_NOIO); + p = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS); } memalloc_nofs_restore(flags); return p; @@ -102,14 +142,14 @@ static void sort_bkey_ptrs(const struct btree *bt, break; for (b = a; c = 2 * b + 1, (d = c + 1) < n;) - b = bkey_cmp_packed(bt, + b = bch2_bkey_cmp_packed(bt, ptrs[c], ptrs[d]) >= 0 ? c : d; if (d == n) b = c; while (b != a && - bkey_cmp_packed(bt, + bch2_bkey_cmp_packed(bt, ptrs[a], ptrs[b]) >= 0) b = (b - 1) / 2; @@ -136,7 +176,7 @@ static void bch2_sort_whiteouts(struct bch_fs *c, struct btree *b) for (k = unwritten_whiteouts_start(c, b); k != unwritten_whiteouts_end(c, b); - k = bkey_next(k)) + k = bkey_p_next(k)) *--ptrs = k; sort_bkey_ptrs(b, ptrs, ptrs_end - ptrs); @@ -144,14 +184,13 @@ static void bch2_sort_whiteouts(struct bch_fs *c, struct btree *b) k = new_whiteouts; while (ptrs != ptrs_end) { - bkey_copy(k, *ptrs); - k = bkey_next(k); + bkey_p_copy(k, *ptrs); + k = bkey_p_next(k); ptrs++; } verify_no_dups(b, new_whiteouts, - (void *) ((u64 *) new_whiteouts + b->whiteout_u64s), - btree_node_old_extent_overwrite(b)); + (void *) ((u64 *) new_whiteouts + b->whiteout_u64s)); memcpy_u64s(unwritten_whiteouts_start(c, b), new_whiteouts, b->whiteout_u64s); @@ -176,144 +215,6 @@ static bool should_compact_bset(struct btree *b, struct bset_tree *t, } } -static bool bch2_compact_extent_whiteouts(struct bch_fs *c, - struct btree *b, - enum compact_mode mode) -{ - const struct bkey_format *f = &b->format; - struct bset_tree *t; - struct bkey_packed *whiteouts = NULL; - struct bkey_packed *u_start, *u_pos; - struct sort_iter sort_iter; - unsigned bytes, whiteout_u64s = 0, u64s; - bool used_mempool, compacting = false; - - BUG_ON(!btree_node_is_extents(b)); - - for_each_bset(b, t) - if (should_compact_bset(b, t, whiteout_u64s != 0, mode)) - whiteout_u64s += bset_dead_u64s(b, t); - - if (!whiteout_u64s) - return false; - - bch2_sort_whiteouts(c, b); - - sort_iter_init(&sort_iter, b); - - whiteout_u64s += b->whiteout_u64s; - bytes = whiteout_u64s * sizeof(u64); - - whiteouts = btree_bounce_alloc(c, bytes, &used_mempool); - u_start = u_pos = whiteouts; - - memcpy_u64s(u_pos, unwritten_whiteouts_start(c, b), - b->whiteout_u64s); - u_pos = (void *) u_pos + b->whiteout_u64s * sizeof(u64); - - sort_iter_add(&sort_iter, u_start, u_pos); - - for_each_bset(b, t) { - struct bset *i = bset(b, t); - struct bkey_packed *k, *n, *out, *start, *end; - struct btree_node_entry *src = NULL, *dst = NULL; - - if (t != b->set && !bset_written(b, i)) { - src = container_of(i, struct btree_node_entry, keys); - dst = max(write_block(b), - (void *) btree_bkey_last(b, t - 1)); - } - - if (src != dst) - compacting = true; - - if (!should_compact_bset(b, t, compacting, mode)) { - if (src != dst) { - memmove(dst, src, sizeof(*src) + - le16_to_cpu(src->keys.u64s) * - sizeof(u64)); - i = &dst->keys; - set_btree_bset(b, t, i); - } - continue; - } - - compacting = true; - u_start = u_pos; - start = i->start; - end = vstruct_last(i); - - if (src != dst) { - memmove(dst, src, sizeof(*src)); - i = &dst->keys; - set_btree_bset(b, t, i); - } - - out = i->start; - - for (k = start; k != end; k = n) { - n = bkey_next_skip_noops(k, end); - - if (bkey_deleted(k)) - continue; - - BUG_ON(bkey_whiteout(k) && - k->needs_whiteout && - bkey_written(b, k)); - - if (bkey_whiteout(k) && !k->needs_whiteout) - continue; - - if (bkey_whiteout(k)) { - memcpy_u64s(u_pos, k, bkeyp_key_u64s(f, k)); - set_bkeyp_val_u64s(f, u_pos, 0); - u_pos = bkey_next(u_pos); - } else { - bkey_copy(out, k); - out = bkey_next(out); - } - } - - sort_iter_add(&sort_iter, u_start, u_pos); - - i->u64s = cpu_to_le16((u64 *) out - i->_data); - set_btree_bset_end(b, t); - bch2_bset_set_no_aux_tree(b, t); - } - - b->whiteout_u64s = (u64 *) u_pos - (u64 *) whiteouts; - - BUG_ON((void *) unwritten_whiteouts_start(c, b) < - (void *) btree_bkey_last(b, bset_tree_last(b))); - - u64s = bch2_sort_extent_whiteouts(unwritten_whiteouts_start(c, b), - &sort_iter); - - BUG_ON(u64s > b->whiteout_u64s); - BUG_ON(u_pos != whiteouts && !u64s); - - if (u64s != b->whiteout_u64s) { - void *src = unwritten_whiteouts_start(c, b); - - b->whiteout_u64s = u64s; - memmove_u64s_up(unwritten_whiteouts_start(c, b), src, u64s); - } - - verify_no_dups(b, - unwritten_whiteouts_start(c, b), - unwritten_whiteouts_end(c, b), - true); - - btree_bounce_free(c, bytes, used_mempool, whiteouts); - - bch2_btree_build_aux_trees(b); - - bch_btree_keys_u64s_remaining(c, b); - bch2_verify_btree_nr_keys(b); - - return true; -} - static bool bch2_drop_whiteouts(struct btree *b, enum compact_mode mode) { struct bset_tree *t; @@ -356,11 +257,11 @@ static bool bch2_drop_whiteouts(struct btree *b, enum compact_mode mode) out = i->start; for (k = start; k != end; k = n) { - n = bkey_next_skip_noops(k, end); + n = bkey_p_next(k); - if (!bkey_whiteout(k)) { - bkey_copy(out, k); - out = bkey_next(out); + if (!bkey_deleted(k)) { + bkey_p_copy(out, k); + out = bkey_p_next(out); } else { BUG_ON(k->needs_whiteout); } @@ -382,19 +283,16 @@ static bool bch2_drop_whiteouts(struct btree *b, enum compact_mode mode) bool bch2_compact_whiteouts(struct bch_fs *c, struct btree *b, enum compact_mode mode) { - return !btree_node_old_extent_overwrite(b) - ? bch2_drop_whiteouts(b, mode) - : bch2_compact_extent_whiteouts(c, b, mode); + return bch2_drop_whiteouts(b, mode); } static void btree_node_sort(struct bch_fs *c, struct btree *b, - struct btree_iter *iter, unsigned start_idx, unsigned end_idx, bool filter_whiteouts) { struct btree_node *out; - struct sort_iter sort_iter; + struct sort_iter_stack sort_iter; struct bset_tree *t; struct bset *start_bset = bset(b, &b->set[start_idx]); bool used_mempool = false; @@ -403,13 +301,13 @@ static void btree_node_sort(struct bch_fs *c, struct btree *b, bool sorting_entire_node = start_idx == 0 && end_idx == b->nsets; - sort_iter_init(&sort_iter, b); + sort_iter_stack_init(&sort_iter, b); for (t = b->set + start_idx; t < b->set + end_idx; t++) { u64s += le16_to_cpu(bset(b, t)->u64s); - sort_iter_add(&sort_iter, + sort_iter_add(&sort_iter.iter, btree_bkey_first(b, t), btree_bkey_last(b, t)); } @@ -422,14 +320,7 @@ static void btree_node_sort(struct bch_fs *c, struct btree *b, start_time = local_clock(); - if (btree_node_old_extent_overwrite(b)) - filter_whiteouts = bset_written(b, start_bset); - - u64s = (btree_node_old_extent_overwrite(b) - ? bch2_sort_extents - : bch2_sort_keys)(out->keys.start, - &sort_iter, - filter_whiteouts); + u64s = bch2_sort_keys(out->keys.start, &sort_iter.iter, filter_whiteouts); out->keys.u64s = cpu_to_le16(u64s); @@ -445,7 +336,7 @@ static void btree_node_sort(struct bch_fs *c, struct btree *b, start_bset->journal_seq = cpu_to_le64(seq); if (sorting_entire_node) { - unsigned u64s = le16_to_cpu(out->keys.u64s); + u64s = le16_to_cpu(out->keys.u64s); BUG_ON(bytes != btree_bytes(c)); @@ -501,16 +392,10 @@ void bch2_btree_sort_into(struct bch_fs *c, bch2_btree_node_iter_init_from_start(&src_iter, src); - if (btree_node_is_extents(src)) - nr = bch2_sort_repack_merge(c, btree_bset_first(dst), - src, &src_iter, - &dst->format, - true); - else - nr = bch2_sort_repack(btree_bset_first(dst), - src, &src_iter, - &dst->format, - true); + nr = bch2_sort_repack(btree_bset_first(dst), + src, &src_iter, + &dst->format, + true); bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort], start_time); @@ -525,14 +410,11 @@ void bch2_btree_sort_into(struct bch_fs *c, bch2_verify_btree_nr_keys(dst); } -#define SORT_CRIT (4096 / sizeof(u64)) - /* * We're about to add another bset to the btree node, so if there's currently * too many bsets - sort some of them together: */ -static bool btree_node_compact(struct bch_fs *c, struct btree *b, - struct btree_iter *iter) +static bool btree_node_compact(struct bch_fs *c, struct btree *b) { unsigned unwritten_idx; bool ret = false; @@ -544,13 +426,13 @@ static bool btree_node_compact(struct bch_fs *c, struct btree *b, break; if (b->nsets - unwritten_idx > 1) { - btree_node_sort(c, b, iter, unwritten_idx, + btree_node_sort(c, b, unwritten_idx, b->nsets, false); ret = true; } if (unwritten_idx > 1) { - btree_node_sort(c, b, iter, 0, unwritten_idx, false); + btree_node_sort(c, b, 0, unwritten_idx, false); ret = true; } @@ -567,6 +449,24 @@ void bch2_btree_build_aux_trees(struct btree *b) t == bset_tree_last(b)); } +/* + * If we have MAX_BSETS (3) bsets, should we sort them all down to just one? + * + * The first bset is going to be of similar order to the size of the node, the + * last bset is bounded by btree_write_set_buffer(), which is set to keep the + * memmove on insert from being too expensive: the middle bset should, ideally, + * be the geometric mean of the first and the last. + * + * Returns true if the middle bset is greater than that geometric mean: + */ +static inline bool should_compact_all(struct bch_fs *c, struct btree *b) +{ + unsigned mid_u64s_bits = + (ilog2(btree_max_u64s(c)) + BTREE_WRITE_SET_U64s_BITS) / 2; + + return bset_u64s(&b->set[1]) > 1U << mid_u64s_bits; +} + /* * @bch_btree_init_next - initialize a new (unwritten) bset that can then be * inserted into @@ -576,16 +476,29 @@ void bch2_btree_build_aux_trees(struct btree *b) * * Returns true if we sorted (i.e. invalidated iterators */ -void bch2_btree_init_next(struct bch_fs *c, struct btree *b, - struct btree_iter *iter) +void bch2_btree_init_next(struct btree_trans *trans, struct btree *b) { + struct bch_fs *c = trans->c; struct btree_node_entry *bne; - bool did_sort; + bool reinit_iter = false; + + EBUG_ON(!six_lock_counts(&b->c.lock).n[SIX_LOCK_write]); + BUG_ON(bset_written(b, bset(b, &b->set[1]))); + BUG_ON(btree_node_just_written(b)); + + if (b->nsets == MAX_BSETS && + !btree_node_write_in_flight(b) && + should_compact_all(c, b)) { + bch2_btree_node_write(c, b, SIX_LOCK_write, + BTREE_WRITE_init_next_bset); + reinit_iter = true; + } - EBUG_ON(!(b->c.lock.state.seq & 1)); - EBUG_ON(iter && iter->l[b->c.level].b != b); + if (b->nsets == MAX_BSETS && + btree_node_compact(c, b)) + reinit_iter = true; - did_sort = btree_node_compact(c, b, iter); + BUG_ON(b->nsets >= MAX_BSETS); bne = want_new_bset(c, b); if (bne) @@ -593,116 +506,236 @@ void bch2_btree_init_next(struct bch_fs *c, struct btree *b, bch2_btree_build_aux_trees(b); - if (iter && did_sort) - bch2_btree_iter_reinit_node(iter, b); + if (reinit_iter) + bch2_trans_node_reinit_iter(trans, b); } static void btree_err_msg(struct printbuf *out, struct bch_fs *c, + struct bch_dev *ca, struct btree *b, struct bset *i, unsigned offset, int write) { - pr_buf(out, "error validating btree node %sat btree %u level %u/%u\n" - "pos ", - write ? "before write " : "", - b->c.btree_id, b->c.level, - c->btree_roots[b->c.btree_id].level); - bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(&b->key)); - - pr_buf(out, " node offset %u", b->written); + prt_printf(out, bch2_log_msg(c, "%s"), + write == READ + ? "error validating btree node " + : "corrupt btree node before write "); + if (ca) + prt_printf(out, "on %s ", ca->name); + prt_printf(out, "at btree "); + bch2_btree_pos_to_text(out, c, b); + + prt_printf(out, "\n node offset %u", b->written); if (i) - pr_buf(out, " bset u64s %u", le16_to_cpu(i->u64s)); + prt_printf(out, " bset u64s %u", le16_to_cpu(i->u64s)); + prt_str(out, ": "); } -enum btree_err_type { - BTREE_ERR_FIXABLE, - BTREE_ERR_WANT_RETRY, - BTREE_ERR_MUST_RETRY, - BTREE_ERR_FATAL, -}; +__printf(9, 10) +static int __btree_err(int ret, + struct bch_fs *c, + struct bch_dev *ca, + struct btree *b, + struct bset *i, + int write, + bool have_retry, + enum bch_sb_error_id err_type, + const char *fmt, ...) +{ + struct printbuf out = PRINTBUF; + va_list args; -enum btree_validate_ret { - BTREE_RETRY_READ = 64, -}; + btree_err_msg(&out, c, ca, b, i, b->written, write); -#define btree_err(type, c, b, i, msg, ...) \ + va_start(args, fmt); + prt_vprintf(&out, fmt, args); + va_end(args); + + if (write == WRITE) { + bch2_print_string_as_lines(KERN_ERR, out.buf); + ret = c->opts.errors == BCH_ON_ERROR_continue + ? 0 + : -BCH_ERR_fsck_errors_not_fixed; + goto out; + } + + if (!have_retry && ret == -BCH_ERR_btree_node_read_err_want_retry) + ret = -BCH_ERR_btree_node_read_err_fixable; + if (!have_retry && ret == -BCH_ERR_btree_node_read_err_must_retry) + ret = -BCH_ERR_btree_node_read_err_bad_node; + + if (ret != -BCH_ERR_btree_node_read_err_fixable) + bch2_sb_error_count(c, err_type); + + switch (ret) { + case -BCH_ERR_btree_node_read_err_fixable: + ret = bch2_fsck_err(c, FSCK_CAN_FIX, err_type, "%s", out.buf); + if (ret != -BCH_ERR_fsck_fix && + ret != -BCH_ERR_fsck_ignore) + goto fsck_err; + ret = -BCH_ERR_fsck_fix; + break; + case -BCH_ERR_btree_node_read_err_want_retry: + case -BCH_ERR_btree_node_read_err_must_retry: + bch2_print_string_as_lines(KERN_ERR, out.buf); + break; + case -BCH_ERR_btree_node_read_err_bad_node: + bch2_print_string_as_lines(KERN_ERR, out.buf); + bch2_topology_error(c); + ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology) ?: -EIO; + break; + case -BCH_ERR_btree_node_read_err_incompatible: + bch2_print_string_as_lines(KERN_ERR, out.buf); + ret = -BCH_ERR_fsck_errors_not_fixed; + break; + default: + BUG(); + } +out: +fsck_err: + printbuf_exit(&out); + return ret; +} + +#define btree_err(type, c, ca, b, i, _err_type, msg, ...) \ ({ \ - __label__ out; \ - char _buf[300]; \ - struct printbuf out = PBUF(_buf); \ + int _ret = __btree_err(type, c, ca, b, i, write, have_retry, \ + BCH_FSCK_ERR_##_err_type, \ + msg, ##__VA_ARGS__); \ \ - btree_err_msg(&out, c, b, i, b->written, write); \ - pr_buf(&out, ": " msg, ##__VA_ARGS__); \ - \ - if (type == BTREE_ERR_FIXABLE && \ - write == READ && \ - !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) { \ - mustfix_fsck_err(c, "%s", _buf); \ - goto out; \ + if (_ret != -BCH_ERR_fsck_fix) { \ + ret = _ret; \ + goto fsck_err; \ } \ \ - switch (write) { \ - case READ: \ - bch_err(c, "%s", _buf); \ - \ - switch (type) { \ - case BTREE_ERR_FIXABLE: \ - ret = BCH_FSCK_ERRORS_NOT_FIXED; \ - goto fsck_err; \ - case BTREE_ERR_WANT_RETRY: \ - if (have_retry) { \ - ret = BTREE_RETRY_READ; \ - goto fsck_err; \ - } \ - break; \ - case BTREE_ERR_MUST_RETRY: \ - ret = BTREE_RETRY_READ; \ - goto fsck_err; \ - case BTREE_ERR_FATAL: \ - ret = BCH_FSCK_ERRORS_NOT_FIXED; \ - goto fsck_err; \ - } \ - break; \ - case WRITE: \ - bch_err(c, "corrupt metadata before write: %s", _buf); \ - \ - if (bch2_fs_inconsistent(c)) { \ - ret = BCH_FSCK_ERRORS_NOT_FIXED; \ - goto fsck_err; \ - } \ - break; \ - } \ -out: \ - true; \ + *saw_error = true; \ }) #define btree_err_on(cond, ...) ((cond) ? btree_err(__VA_ARGS__) : false) -static int validate_bset(struct bch_fs *c, struct btree *b, - struct bset *i, unsigned sectors, - int write, bool have_retry) +/* + * When btree topology repair changes the start or end of a node, that might + * mean we have to drop keys that are no longer inside the node: + */ +__cold +void bch2_btree_node_drop_keys_outside_node(struct btree *b) +{ + struct bset_tree *t; + + for_each_bset(b, t) { + struct bset *i = bset(b, t); + struct bkey_packed *k; + + for (k = i->start; k != vstruct_last(i); k = bkey_p_next(k)) + if (bkey_cmp_left_packed(b, k, &b->data->min_key) >= 0) + break; + + if (k != i->start) { + unsigned shift = (u64 *) k - (u64 *) i->start; + + memmove_u64s_down(i->start, k, + (u64 *) vstruct_end(i) - (u64 *) k); + i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - shift); + set_btree_bset_end(b, t); + } + + for (k = i->start; k != vstruct_last(i); k = bkey_p_next(k)) + if (bkey_cmp_left_packed(b, k, &b->data->max_key) > 0) + break; + + if (k != vstruct_last(i)) { + i->u64s = cpu_to_le16((u64 *) k - (u64 *) i->start); + set_btree_bset_end(b, t); + } + } + + /* + * Always rebuild search trees: eytzinger search tree nodes directly + * depend on the values of min/max key: + */ + bch2_bset_set_no_aux_tree(b, b->set); + bch2_btree_build_aux_trees(b); + + struct bkey_s_c k; + struct bkey unpacked; + struct btree_node_iter iter; + for_each_btree_node_key_unpack(b, k, &iter, &unpacked) { + BUG_ON(bpos_lt(k.k->p, b->data->min_key)); + BUG_ON(bpos_gt(k.k->p, b->data->max_key)); + } +} + +static int validate_bset(struct bch_fs *c, struct bch_dev *ca, + struct btree *b, struct bset *i, + unsigned offset, unsigned sectors, + int write, bool have_retry, bool *saw_error) { unsigned version = le16_to_cpu(i->version); - const char *err; + struct printbuf buf1 = PRINTBUF; + struct printbuf buf2 = PRINTBUF; int ret = 0; - btree_err_on((version != BCH_BSET_VERSION_OLD && - version < bcachefs_metadata_version_min) || - version >= bcachefs_metadata_version_max, - BTREE_ERR_FATAL, c, b, i, - "unsupported bset version"); + btree_err_on(!bch2_version_compatible(version), + -BCH_ERR_btree_node_read_err_incompatible, + c, ca, b, i, + btree_node_unsupported_version, + "unsupported bset version %u.%u", + BCH_VERSION_MAJOR(version), + BCH_VERSION_MINOR(version)); + + if (btree_err_on(version < c->sb.version_min, + -BCH_ERR_btree_node_read_err_fixable, + c, NULL, b, i, + btree_node_bset_older_than_sb_min, + "bset version %u older than superblock version_min %u", + version, c->sb.version_min)) { + mutex_lock(&c->sb_lock); + c->disk_sb.sb->version_min = cpu_to_le16(version); + bch2_write_super(c); + mutex_unlock(&c->sb_lock); + } + + if (btree_err_on(BCH_VERSION_MAJOR(version) > + BCH_VERSION_MAJOR(c->sb.version), + -BCH_ERR_btree_node_read_err_fixable, + c, NULL, b, i, + btree_node_bset_newer_than_sb, + "bset version %u newer than superblock version %u", + version, c->sb.version)) { + mutex_lock(&c->sb_lock); + c->disk_sb.sb->version = cpu_to_le16(version); + bch2_write_super(c); + mutex_unlock(&c->sb_lock); + } - if (btree_err_on(b->written + sectors > c->opts.btree_node_size, - BTREE_ERR_FIXABLE, c, b, i, + btree_err_on(BSET_SEPARATE_WHITEOUTS(i), + -BCH_ERR_btree_node_read_err_incompatible, + c, ca, b, i, + btree_node_unsupported_version, + "BSET_SEPARATE_WHITEOUTS no longer supported"); + + if (btree_err_on(offset + sectors > btree_sectors(c), + -BCH_ERR_btree_node_read_err_fixable, + c, ca, b, i, + bset_past_end_of_btree_node, "bset past end of btree node")) { i->u64s = 0; - return 0; + ret = 0; + goto out; } - btree_err_on(b->written && !i->u64s, - BTREE_ERR_FIXABLE, c, b, i, + btree_err_on(offset && !i->u64s, + -BCH_ERR_btree_node_read_err_fixable, + c, ca, b, i, + bset_empty, "empty bset"); - if (!b->written) { + btree_err_on(BSET_OFFSET(i) && BSET_OFFSET(i) != offset, + -BCH_ERR_btree_node_read_err_want_retry, + c, ca, b, i, + bset_wrong_sector_offset, + "bset at wrong sector offset"); + + if (!offset) { struct btree_node *bn = container_of(i, struct btree_node, keys); /* These indicate that we read the wrong btree node: */ @@ -713,24 +746,24 @@ static int validate_bset(struct bch_fs *c, struct btree *b, /* XXX endianness */ btree_err_on(bp->seq != bn->keys.seq, - BTREE_ERR_MUST_RETRY, c, b, NULL, + -BCH_ERR_btree_node_read_err_must_retry, + c, ca, b, NULL, + bset_bad_seq, "incorrect sequence number (wrong btree node)"); } btree_err_on(BTREE_NODE_ID(bn) != b->c.btree_id, - BTREE_ERR_MUST_RETRY, c, b, i, + -BCH_ERR_btree_node_read_err_must_retry, + c, ca, b, i, + btree_node_bad_btree, "incorrect btree id"); btree_err_on(BTREE_NODE_LEVEL(bn) != b->c.level, - BTREE_ERR_MUST_RETRY, c, b, i, + -BCH_ERR_btree_node_read_err_must_retry, + c, ca, b, i, + btree_node_bad_level, "incorrect level"); - if (BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN) { - u64 *p = (u64 *) &bn->ptr; - - *p = swab64(*p); - } - if (!write) compat_btree_node(b->c.level, b->c.btree_id, version, BSET_BIG_ENDIAN(i), write, bn); @@ -739,82 +772,96 @@ static int validate_bset(struct bch_fs *c, struct btree *b, struct bch_btree_ptr_v2 *bp = &bkey_i_to_btree_ptr_v2(&b->key)->v; - btree_err_on(bkey_cmp(b->data->min_key, bp->min_key), - BTREE_ERR_MUST_RETRY, c, b, NULL, - "incorrect min_key: got %llu:%llu should be %llu:%llu", - b->data->min_key.inode, - b->data->min_key.offset, - bp->min_key.inode, - bp->min_key.offset); + if (BTREE_PTR_RANGE_UPDATED(bp)) { + b->data->min_key = bp->min_key; + b->data->max_key = b->key.k.p; + } + + btree_err_on(!bpos_eq(b->data->min_key, bp->min_key), + -BCH_ERR_btree_node_read_err_must_retry, + c, ca, b, NULL, + btree_node_bad_min_key, + "incorrect min_key: got %s should be %s", + (printbuf_reset(&buf1), + bch2_bpos_to_text(&buf1, bn->min_key), buf1.buf), + (printbuf_reset(&buf2), + bch2_bpos_to_text(&buf2, bp->min_key), buf2.buf)); } - btree_err_on(bkey_cmp(bn->max_key, b->key.k.p), - BTREE_ERR_MUST_RETRY, c, b, i, - "incorrect max key"); + btree_err_on(!bpos_eq(bn->max_key, b->key.k.p), + -BCH_ERR_btree_node_read_err_must_retry, + c, ca, b, i, + btree_node_bad_max_key, + "incorrect max key %s", + (printbuf_reset(&buf1), + bch2_bpos_to_text(&buf1, bn->max_key), buf1.buf)); if (write) compat_btree_node(b->c.level, b->c.btree_id, version, BSET_BIG_ENDIAN(i), write, bn); - /* XXX: ideally we would be validating min_key too */ -#if 0 - /* - * not correct anymore, due to btree node write error - * handling - * - * need to add bn->seq to btree keys and verify - * against that - */ - btree_err_on(!extent_contains_ptr(bkey_i_to_s_c_extent(&b->key), - bn->ptr), - BTREE_ERR_FATAL, c, b, i, - "incorrect backpointer"); -#endif - err = bch2_bkey_format_validate(&bn->format); - btree_err_on(err, - BTREE_ERR_FATAL, c, b, i, - "invalid bkey format: %s", err); + btree_err_on(bch2_bkey_format_invalid(c, &bn->format, write, &buf1), + -BCH_ERR_btree_node_read_err_bad_node, + c, ca, b, i, + btree_node_bad_format, + "invalid bkey format: %s\n %s", buf1.buf, + (printbuf_reset(&buf2), + bch2_bkey_format_to_text(&buf2, &bn->format), buf2.buf)); + printbuf_reset(&buf1); compat_bformat(b->c.level, b->c.btree_id, version, BSET_BIG_ENDIAN(i), write, &bn->format); } +out: fsck_err: + printbuf_exit(&buf2); + printbuf_exit(&buf1); return ret; } +static int bset_key_invalid(struct bch_fs *c, struct btree *b, + struct bkey_s_c k, + bool updated_range, int rw, + struct printbuf *err) +{ + return __bch2_bkey_invalid(c, k, btree_node_type(b), READ, err) ?: + (!updated_range ? bch2_bkey_in_btree_node(c, b, k, err) : 0) ?: + (rw == WRITE ? bch2_bkey_val_invalid(c, k, READ, err) : 0); +} + static int validate_bset_keys(struct bch_fs *c, struct btree *b, - struct bset *i, unsigned *whiteout_u64s, - int write, bool have_retry) + struct bset *i, int write, + bool have_retry, bool *saw_error) { unsigned version = le16_to_cpu(i->version); struct bkey_packed *k, *prev = NULL; - bool seen_non_whiteout = false; + struct printbuf buf = PRINTBUF; + bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 && + BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v); int ret = 0; - if (!BSET_SEPARATE_WHITEOUTS(i)) { - seen_non_whiteout = true; - *whiteout_u64s = 0; - } - for (k = i->start; k != vstruct_last(i);) { struct bkey_s u; struct bkey tmp; - const char *invalid; - if (btree_err_on(bkey_next(k) > vstruct_last(i), - BTREE_ERR_FIXABLE, c, b, i, + if (btree_err_on(bkey_p_next(k) > vstruct_last(i), + -BCH_ERR_btree_node_read_err_fixable, + c, NULL, b, i, + btree_node_bkey_past_bset_end, "key extends past end of bset")) { i->u64s = cpu_to_le16((u64 *) k - i->_data); break; } if (btree_err_on(k->format > KEY_FORMAT_CURRENT, - BTREE_ERR_FIXABLE, c, b, i, + -BCH_ERR_btree_node_read_err_fixable, + c, NULL, b, i, + btree_node_bkey_bad_format, "invalid bkey format %u", k->format)) { i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s); - memmove_u64s_down(k, bkey_next(k), + memmove_u64s_down(k, bkey_p_next(k), (u64 *) vstruct_end(i) - (u64 *) k); continue; } @@ -827,18 +874,20 @@ static int validate_bset_keys(struct bch_fs *c, struct btree *b, u = __bkey_disassemble(b, k, &tmp); - invalid = __bch2_bkey_invalid(c, u.s_c, btree_node_type(b)) ?: - bch2_bkey_in_btree_node(b, u.s_c) ?: - (write ? bch2_bkey_val_invalid(c, u.s_c) : NULL); - if (invalid) { - char buf[160]; + printbuf_reset(&buf); + if (bset_key_invalid(c, b, u.s_c, updated_range, write, &buf)) { + printbuf_reset(&buf); + bset_key_invalid(c, b, u.s_c, updated_range, write, &buf); + prt_printf(&buf, "\n "); + bch2_bkey_val_to_text(&buf, c, u.s_c); - bch2_bkey_val_to_text(&PBUF(buf), c, u.s_c); - btree_err(BTREE_ERR_FIXABLE, c, b, i, - "invalid bkey:\n%s\n%s", invalid, buf); + btree_err(-BCH_ERR_btree_node_read_err_fixable, + c, NULL, b, i, + btree_node_bad_bkey, + "invalid bkey: %s", buf.buf); i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s); - memmove_u64s_down(k, bkey_next(k), + memmove_u64s_down(k, bkey_p_next(k), (u64 *) vstruct_end(i) - (u64 *) k); continue; } @@ -848,104 +897,137 @@ static int validate_bset_keys(struct bch_fs *c, struct btree *b, BSET_BIG_ENDIAN(i), write, &b->format, k); - /* - * with the separate whiteouts thing (used for extents), the - * second set of keys actually can have whiteouts too, so we - * can't solely go off bkey_whiteout()... - */ - - if (!seen_non_whiteout && - (!bkey_whiteout(k) || - (prev && bkey_iter_cmp(b, prev, k) > 0))) { - *whiteout_u64s = k->_data - i->_data; - seen_non_whiteout = true; - } else if (prev && bkey_iter_cmp(b, prev, k) > 0) { - char buf1[80]; - char buf2[80]; + if (prev && bkey_iter_cmp(b, prev, k) > 0) { struct bkey up = bkey_unpack_key(b, prev); - bch2_bkey_to_text(&PBUF(buf1), &up); - bch2_bkey_to_text(&PBUF(buf2), u.k); + printbuf_reset(&buf); + prt_printf(&buf, "keys out of order: "); + bch2_bkey_to_text(&buf, &up); + prt_printf(&buf, " > "); + bch2_bkey_to_text(&buf, u.k); bch2_dump_bset(c, b, i, 0); - btree_err(BTREE_ERR_FATAL, c, b, i, - "keys out of order: %s > %s", - buf1, buf2); - /* XXX: repair this */ + + if (btree_err(-BCH_ERR_btree_node_read_err_fixable, + c, NULL, b, i, + btree_node_bkey_out_of_order, + "%s", buf.buf)) { + i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s); + memmove_u64s_down(k, bkey_p_next(k), + (u64 *) vstruct_end(i) - (u64 *) k); + continue; + } } prev = k; - k = bkey_next_skip_noops(k, vstruct_last(i)); + k = bkey_p_next(k); } fsck_err: + printbuf_exit(&buf); return ret; } -int bch2_btree_node_read_done(struct bch_fs *c, struct btree *b, bool have_retry) +int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca, + struct btree *b, bool have_retry, bool *saw_error) { struct btree_node_entry *bne; struct sort_iter *iter; struct btree_node *sorted; struct bkey_packed *k; - struct bch_extent_ptr *ptr; struct bset *i; bool used_mempool, blacklisted; + bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 && + BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v); unsigned u64s; - int ret, retry_read = 0, write = READ; + unsigned ptr_written = btree_ptr_sectors_written(&b->key); + struct printbuf buf = PRINTBUF; + int ret = 0, retry_read = 0, write = READ; + + b->version_ondisk = U16_MAX; + /* We might get called multiple times on read retry: */ + b->written = 0; - iter = mempool_alloc(&c->fill_iter, GFP_NOIO); - sort_iter_init(iter, b); - iter->size = (btree_blocks(c) + 1) * 2; + iter = mempool_alloc(&c->fill_iter, GFP_NOFS); + sort_iter_init(iter, b, (btree_blocks(c) + 1) * 2); if (bch2_meta_read_fault("btree")) - btree_err(BTREE_ERR_MUST_RETRY, c, b, NULL, + btree_err(-BCH_ERR_btree_node_read_err_must_retry, + c, ca, b, NULL, + btree_node_fault_injected, "dynamic fault"); btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c), - BTREE_ERR_MUST_RETRY, c, b, NULL, - "bad magic"); - - btree_err_on(!b->data->keys.seq, - BTREE_ERR_MUST_RETRY, c, b, NULL, - "bad btree header"); + -BCH_ERR_btree_node_read_err_must_retry, + c, ca, b, NULL, + btree_node_bad_magic, + "bad magic: want %llx, got %llx", + bset_magic(c), le64_to_cpu(b->data->magic)); if (b->key.k.type == KEY_TYPE_btree_ptr_v2) { struct bch_btree_ptr_v2 *bp = &bkey_i_to_btree_ptr_v2(&b->key)->v; + bch2_bpos_to_text(&buf, b->data->min_key); + prt_str(&buf, "-"); + bch2_bpos_to_text(&buf, b->data->max_key); + btree_err_on(b->data->keys.seq != bp->seq, - BTREE_ERR_MUST_RETRY, c, b, NULL, - "got wrong btree node (seq %llx want %llx)", - b->data->keys.seq, bp->seq); + -BCH_ERR_btree_node_read_err_must_retry, + c, ca, b, NULL, + btree_node_bad_seq, + "got wrong btree node (want %llx got %llx)\n" + "got btree %s level %llu pos %s", + bp->seq, b->data->keys.seq, + bch2_btree_id_str(BTREE_NODE_ID(b->data)), + BTREE_NODE_LEVEL(b->data), + buf.buf); + } else { + btree_err_on(!b->data->keys.seq, + -BCH_ERR_btree_node_read_err_must_retry, + c, ca, b, NULL, + btree_node_bad_seq, + "bad btree header: seq 0"); } - while (b->written < c->opts.btree_node_size) { - unsigned sectors, whiteout_u64s = 0; + while (b->written < (ptr_written ?: btree_sectors(c))) { + unsigned sectors; struct nonce nonce; - struct bch_csum csum; bool first = !b->written; + bool csum_bad; if (!b->written) { i = &b->data->keys; btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)), - BTREE_ERR_WANT_RETRY, c, b, i, - "unknown checksum type"); + -BCH_ERR_btree_node_read_err_want_retry, + c, ca, b, i, + bset_unknown_csum, + "unknown checksum type %llu", BSET_CSUM_TYPE(i)); nonce = btree_nonce(i, b->written << 9); - csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data); - btree_err_on(bch2_crc_cmp(csum, b->data->csum), - BTREE_ERR_WANT_RETRY, c, b, i, + csum_bad = bch2_crc_cmp(b->data->csum, + csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data)); + if (csum_bad) + bch2_io_error(ca, BCH_MEMBER_ERROR_checksum); + + btree_err_on(csum_bad, + -BCH_ERR_btree_node_read_err_want_retry, + c, ca, b, i, + bset_bad_csum, "invalid checksum"); - bset_encrypt(c, i, b->written << 9); + ret = bset_encrypt(c, i, b->written << 9); + if (bch2_fs_fatal_err_on(ret, c, + "error decrypting btree node: %i", ret)) + goto fsck_err; - if (btree_node_is_extents(b) && - !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data)) { - set_btree_node_old_extent_overwrite(b); - set_btree_node_need_rewrite(b); - } + btree_err_on(btree_node_type_is_extents(btree_node_type(b)) && + !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data), + -BCH_ERR_btree_node_read_err_incompatible, + c, NULL, b, NULL, + btree_node_unsupported_version, + "btree node does not have NEW_EXTENT_OVERWRITE set"); sectors = vstruct_sectors(b->data, c->block_bits); } else { @@ -956,71 +1038,104 @@ int bch2_btree_node_read_done(struct bch_fs *c, struct btree *b, bool have_retry break; btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)), - BTREE_ERR_WANT_RETRY, c, b, i, - "unknown checksum type"); + -BCH_ERR_btree_node_read_err_want_retry, + c, ca, b, i, + bset_unknown_csum, + "unknown checksum type %llu", BSET_CSUM_TYPE(i)); nonce = btree_nonce(i, b->written << 9); - csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne); - - btree_err_on(bch2_crc_cmp(csum, bne->csum), - BTREE_ERR_WANT_RETRY, c, b, i, + csum_bad = bch2_crc_cmp(bne->csum, + csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne)); + if (csum_bad) + bch2_io_error(ca, BCH_MEMBER_ERROR_checksum); + + btree_err_on(csum_bad, + -BCH_ERR_btree_node_read_err_want_retry, + c, ca, b, i, + bset_bad_csum, "invalid checksum"); - bset_encrypt(c, i, b->written << 9); + ret = bset_encrypt(c, i, b->written << 9); + if (bch2_fs_fatal_err_on(ret, c, + "error decrypting btree node: %i\n", ret)) + goto fsck_err; sectors = vstruct_sectors(bne, c->block_bits); } - ret = validate_bset(c, b, i, sectors, - READ, have_retry); + b->version_ondisk = min(b->version_ondisk, + le16_to_cpu(i->version)); + + ret = validate_bset(c, ca, b, i, b->written, sectors, + READ, have_retry, saw_error); if (ret) goto fsck_err; if (!b->written) btree_node_set_format(b, b->data->format); - ret = validate_bset_keys(c, b, i, &whiteout_u64s, - READ, have_retry); + ret = validate_bset_keys(c, b, i, READ, have_retry, saw_error); if (ret) goto fsck_err; SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN); - b->written += sectors; - blacklisted = bch2_journal_seq_is_blacklisted(c, le64_to_cpu(i->journal_seq), true); btree_err_on(blacklisted && first, - BTREE_ERR_FIXABLE, c, b, i, - "first btree node bset has blacklisted journal seq"); + -BCH_ERR_btree_node_read_err_fixable, + c, ca, b, i, + bset_blacklisted_journal_seq, + "first btree node bset has blacklisted journal seq (%llu)", + le64_to_cpu(i->journal_seq)); + + btree_err_on(blacklisted && ptr_written, + -BCH_ERR_btree_node_read_err_fixable, + c, ca, b, i, + first_bset_blacklisted_journal_seq, + "found blacklisted bset (journal seq %llu) in btree node at offset %u-%u/%u", + le64_to_cpu(i->journal_seq), + b->written, b->written + sectors, ptr_written); + + b->written += sectors; + if (blacklisted && !first) continue; - sort_iter_add(iter, i->start, - vstruct_idx(i, whiteout_u64s)); - sort_iter_add(iter, - vstruct_idx(i, whiteout_u64s), + vstruct_idx(i, 0), vstruct_last(i)); } - for (bne = write_block(b); - bset_byte_offset(b, bne) < btree_bytes(c); - bne = (void *) bne + block_bytes(c)) - btree_err_on(bne->keys.seq == b->data->keys.seq, - BTREE_ERR_WANT_RETRY, c, b, NULL, - "found bset signature after last bset"); + if (ptr_written) { + btree_err_on(b->written < ptr_written, + -BCH_ERR_btree_node_read_err_want_retry, + c, ca, b, NULL, + btree_node_data_missing, + "btree node data missing: expected %u sectors, found %u", + ptr_written, b->written); + } else { + for (bne = write_block(b); + bset_byte_offset(b, bne) < btree_bytes(c); + bne = (void *) bne + block_bytes(c)) + btree_err_on(bne->keys.seq == b->data->keys.seq && + !bch2_journal_seq_is_blacklisted(c, + le64_to_cpu(bne->keys.journal_seq), + true), + -BCH_ERR_btree_node_read_err_want_retry, + c, ca, b, NULL, + btree_node_bset_after_end, + "found bset signature after last bset"); + } sorted = btree_bounce_alloc(c, btree_bytes(c), &used_mempool); sorted->keys.u64s = 0; set_btree_bset(b, b->set, &b->data->keys); - b->nr = (btree_node_old_extent_overwrite(b) - ? bch2_extent_sort_fix_overlapping - : bch2_key_sort_fix_overlapping)(c, &sorted->keys, iter); + b->nr = bch2_key_sort_fix_overlapping(c, &sorted->keys, iter); u64s = le16_to_cpu(sorted->keys.u64s); *sorted = *b->data; @@ -1033,25 +1148,35 @@ int bch2_btree_node_read_done(struct bch_fs *c, struct btree *b, bool have_retry btree_bounce_free(c, btree_bytes(c), used_mempool, sorted); + if (updated_range) + bch2_btree_node_drop_keys_outside_node(b); + i = &b->data->keys; for (k = i->start; k != vstruct_last(i);) { struct bkey tmp; struct bkey_s u = __bkey_disassemble(b, k, &tmp); - const char *invalid = bch2_bkey_val_invalid(c, u.s_c); - if (invalid || - (inject_invalid_keys(c) && + printbuf_reset(&buf); + + if (bch2_bkey_val_invalid(c, u.s_c, READ, &buf) || + (bch2_inject_invalid_keys && !bversion_cmp(u.k->version, MAX_VERSION))) { - char buf[160]; + printbuf_reset(&buf); + + prt_printf(&buf, "invalid bkey: "); + bch2_bkey_val_invalid(c, u.s_c, READ, &buf); + prt_printf(&buf, "\n "); + bch2_bkey_val_to_text(&buf, c, u.s_c); - bch2_bkey_val_to_text(&PBUF(buf), c, u.s_c); - btree_err(BTREE_ERR_FIXABLE, c, b, i, - "invalid bkey %s: %s", buf, invalid); + btree_err(-BCH_ERR_btree_node_read_err_fixable, + c, NULL, b, i, + btree_node_bad_bkey, + "%s", buf.buf); btree_keys_account_key_drop(&b->nr, 0, k); i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s); - memmove_u64s_down(k, bkey_next(k), + memmove_u64s_down(k, bkey_p_next(k), (u64 *) vstruct_end(i) - (u64 *) k); set_btree_bset_end(b, b->set); continue; @@ -1063,7 +1188,7 @@ int bch2_btree_node_read_done(struct bch_fs *c, struct btree *b, bool have_retry bp.v->mem_ptr = 0; } - k = bkey_next_skip_noops(k, vstruct_last(i)); + k = bkey_p_next(k); } bch2_bset_build_aux_tree(b, b->set, false); @@ -1073,21 +1198,24 @@ int bch2_btree_node_read_done(struct bch_fs *c, struct btree *b, bool have_retry btree_node_reset_sib_u64s(b); bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&b->key)), ptr) { - struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev); + struct bch_dev *ca2 = bch_dev_bkey_exists(c, ptr->dev); - if (ca->mi.state != BCH_MEMBER_STATE_RW) + if (ca2->mi.state != BCH_MEMBER_STATE_rw) set_btree_node_need_rewrite(b); } + + if (!ptr_written) + set_btree_node_need_rewrite(b); out: mempool_free(iter, &c->fill_iter); + printbuf_exit(&buf); return retry_read; fsck_err: - if (ret == BTREE_RETRY_READ) { + if (ret == -BCH_ERR_btree_node_read_err_want_retry || + ret == -BCH_ERR_btree_node_read_err_must_retry) retry_read = 1; - } else { - bch2_inconsistent_error(c); + else set_btree_node_read_error(b); - } goto out; } @@ -1096,19 +1224,22 @@ static void btree_node_read_work(struct work_struct *work) struct btree_read_bio *rb = container_of(work, struct btree_read_bio, work); struct bch_fs *c = rb->c; + struct btree *b = rb->b; struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev); - struct btree *b = rb->bio.bi_private; struct bio *bio = &rb->bio; struct bch_io_failures failed = { .nr = 0 }; + struct printbuf buf = PRINTBUF; + bool saw_error = false; + bool retry = false; bool can_retry; goto start; while (1) { + retry = true; bch_info(c, "retrying read"); ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev); rb->have_ioref = bch2_dev_get_ioref(ca, READ); - bio_reset(bio); - bio->bi_opf = REQ_OP_READ|REQ_SYNC|REQ_META; + bio_reset(bio, NULL, REQ_OP_READ|REQ_SYNC|REQ_META); bio->bi_iter.bi_sector = rb->pick.ptr.offset; bio->bi_iter.bi_size = btree_bytes(c); @@ -1119,8 +1250,11 @@ static void btree_node_read_work(struct work_struct *work) bio->bi_status = BLK_STS_REMOVED; } start: - bch2_dev_io_err_on(bio->bi_status, ca, "btree read: %s", - bch2_blk_status_to_str(bio->bi_status)); + printbuf_reset(&buf); + bch2_btree_pos_to_text(&buf, c, b); + bch2_dev_io_err_on(bio->bi_status, ca, BCH_MEMBER_ERROR_read, + "btree read error %s for %s", + bch2_blk_status_to_str(bio->bi_status), buf.buf); if (rb->have_ioref) percpu_ref_put(&ca->io_ref); rb->have_ioref = false; @@ -1132,8 +1266,13 @@ start: &failed, &rb->pick) > 0; if (!bio->bi_status && - !bch2_btree_node_read_done(c, b, can_retry)) + !bch2_btree_node_read_done(c, ca, b, can_retry, &saw_error)) { + if (retry) + bch_info(c, "retry success"); break; + } + + saw_error = true; if (!can_retry) { set_btree_node_read_error(b); @@ -1144,6 +1283,17 @@ start: bch2_time_stats_update(&c->times[BCH_TIME_btree_node_read], rb->start_time); bio_put(&rb->bio); + + if (saw_error && !btree_node_read_error(b)) { + printbuf_reset(&buf); + bch2_bpos_to_text(&buf, b->key.k.p); + bch_info(c, "%s: rewriting btree node at btree=%s level=%u %s due to error", + __func__, bch2_btree_id_str(b->c.btree_id), b->c.level, buf.buf); + + bch2_btree_node_rewrite_async(c, b); + } + + printbuf_exit(&buf); clear_btree_node_read_in_flight(b); wake_up_bit(&b->flags, BTREE_NODE_read_in_flight); } @@ -1156,50 +1306,338 @@ static void btree_node_read_endio(struct bio *bio) if (rb->have_ioref) { struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev); + + bch2_latency_acct(ca, rb->start_time, READ); + } + + queue_work(c->io_complete_wq, &rb->work); +} + +struct btree_node_read_all { + struct closure cl; + struct bch_fs *c; + struct btree *b; + unsigned nr; + void *buf[BCH_REPLICAS_MAX]; + struct bio *bio[BCH_REPLICAS_MAX]; + blk_status_t err[BCH_REPLICAS_MAX]; +}; + +static unsigned btree_node_sectors_written(struct bch_fs *c, void *data) +{ + struct btree_node *bn = data; + struct btree_node_entry *bne; + unsigned offset = 0; + + if (le64_to_cpu(bn->magic) != bset_magic(c)) + return 0; + + while (offset < btree_sectors(c)) { + if (!offset) { + offset += vstruct_sectors(bn, c->block_bits); + } else { + bne = data + (offset << 9); + if (bne->keys.seq != bn->keys.seq) + break; + offset += vstruct_sectors(bne, c->block_bits); + } + } + + return offset; +} + +static bool btree_node_has_extra_bsets(struct bch_fs *c, unsigned offset, void *data) +{ + struct btree_node *bn = data; + struct btree_node_entry *bne; + + if (!offset) + return false; + + while (offset < btree_sectors(c)) { + bne = data + (offset << 9); + if (bne->keys.seq == bn->keys.seq) + return true; + offset++; + } + + return false; + return offset; +} + +static CLOSURE_CALLBACK(btree_node_read_all_replicas_done) +{ + closure_type(ra, struct btree_node_read_all, cl); + struct bch_fs *c = ra->c; + struct btree *b = ra->b; + struct printbuf buf = PRINTBUF; + bool dump_bset_maps = false; + bool have_retry = false; + int ret = 0, best = -1, write = READ; + unsigned i, written = 0, written2 = 0; + __le64 seq = b->key.k.type == KEY_TYPE_btree_ptr_v2 + ? bkey_i_to_btree_ptr_v2(&b->key)->v.seq : 0; + bool _saw_error = false, *saw_error = &_saw_error; + + for (i = 0; i < ra->nr; i++) { + struct btree_node *bn = ra->buf[i]; + + if (ra->err[i]) + continue; + + if (le64_to_cpu(bn->magic) != bset_magic(c) || + (seq && seq != bn->keys.seq)) + continue; + + if (best < 0) { + best = i; + written = btree_node_sectors_written(c, bn); + continue; + } + + written2 = btree_node_sectors_written(c, ra->buf[i]); + if (btree_err_on(written2 != written, -BCH_ERR_btree_node_read_err_fixable, + c, NULL, b, NULL, + btree_node_replicas_sectors_written_mismatch, + "btree node sectors written mismatch: %u != %u", + written, written2) || + btree_err_on(btree_node_has_extra_bsets(c, written2, ra->buf[i]), + -BCH_ERR_btree_node_read_err_fixable, + c, NULL, b, NULL, + btree_node_bset_after_end, + "found bset signature after last bset") || + btree_err_on(memcmp(ra->buf[best], ra->buf[i], written << 9), + -BCH_ERR_btree_node_read_err_fixable, + c, NULL, b, NULL, + btree_node_replicas_data_mismatch, + "btree node replicas content mismatch")) + dump_bset_maps = true; + + if (written2 > written) { + written = written2; + best = i; + } + } +fsck_err: + if (dump_bset_maps) { + for (i = 0; i < ra->nr; i++) { + struct btree_node *bn = ra->buf[i]; + struct btree_node_entry *bne = NULL; + unsigned offset = 0, sectors; + bool gap = false; + + if (ra->err[i]) + continue; + + printbuf_reset(&buf); + + while (offset < btree_sectors(c)) { + if (!offset) { + sectors = vstruct_sectors(bn, c->block_bits); + } else { + bne = ra->buf[i] + (offset << 9); + if (bne->keys.seq != bn->keys.seq) + break; + sectors = vstruct_sectors(bne, c->block_bits); + } + + prt_printf(&buf, " %u-%u", offset, offset + sectors); + if (bne && bch2_journal_seq_is_blacklisted(c, + le64_to_cpu(bne->keys.journal_seq), false)) + prt_printf(&buf, "*"); + offset += sectors; + } + + while (offset < btree_sectors(c)) { + bne = ra->buf[i] + (offset << 9); + if (bne->keys.seq == bn->keys.seq) { + if (!gap) + prt_printf(&buf, " GAP"); + gap = true; + + sectors = vstruct_sectors(bne, c->block_bits); + prt_printf(&buf, " %u-%u", offset, offset + sectors); + if (bch2_journal_seq_is_blacklisted(c, + le64_to_cpu(bne->keys.journal_seq), false)) + prt_printf(&buf, "*"); + } + offset++; + } + + bch_err(c, "replica %u:%s", i, buf.buf); + } + } + + if (best >= 0) { + memcpy(b->data, ra->buf[best], btree_bytes(c)); + ret = bch2_btree_node_read_done(c, NULL, b, false, saw_error); + } else { + ret = -1; + } + + if (ret) + set_btree_node_read_error(b); + else if (*saw_error) + bch2_btree_node_rewrite_async(c, b); + + for (i = 0; i < ra->nr; i++) { + mempool_free(ra->buf[i], &c->btree_bounce_pool); + bio_put(ra->bio[i]); + } + + closure_debug_destroy(&ra->cl); + kfree(ra); + printbuf_exit(&buf); + + clear_btree_node_read_in_flight(b); + wake_up_bit(&b->flags, BTREE_NODE_read_in_flight); +} + +static void btree_node_read_all_replicas_endio(struct bio *bio) +{ + struct btree_read_bio *rb = + container_of(bio, struct btree_read_bio, bio); + struct bch_fs *c = rb->c; + struct btree_node_read_all *ra = rb->ra; + + if (rb->have_ioref) { + struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev); + bch2_latency_acct(ca, rb->start_time, READ); } - queue_work(system_unbound_wq, &rb->work); + ra->err[rb->idx] = bio->bi_status; + closure_put(&ra->cl); } -void bch2_btree_node_read(struct bch_fs *c, struct btree *b, +/* + * XXX This allocates multiple times from the same mempools, and can deadlock + * under sufficient memory pressure (but is only a debug path) + */ +static int btree_node_read_all_replicas(struct bch_fs *c, struct btree *b, bool sync) +{ + struct bkey_s_c k = bkey_i_to_s_c(&b->key); + struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k); + const union bch_extent_entry *entry; + struct extent_ptr_decoded pick; + struct btree_node_read_all *ra; + unsigned i; + + ra = kzalloc(sizeof(*ra), GFP_NOFS); + if (!ra) + return -BCH_ERR_ENOMEM_btree_node_read_all_replicas; + + closure_init(&ra->cl, NULL); + ra->c = c; + ra->b = b; + ra->nr = bch2_bkey_nr_ptrs(k); + + for (i = 0; i < ra->nr; i++) { + ra->buf[i] = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS); + ra->bio[i] = bio_alloc_bioset(NULL, + buf_pages(ra->buf[i], btree_bytes(c)), + REQ_OP_READ|REQ_SYNC|REQ_META, + GFP_NOFS, + &c->btree_bio); + } + + i = 0; + bkey_for_each_ptr_decode(k.k, ptrs, pick, entry) { + struct bch_dev *ca = bch_dev_bkey_exists(c, pick.ptr.dev); + struct btree_read_bio *rb = + container_of(ra->bio[i], struct btree_read_bio, bio); + rb->c = c; + rb->b = b; + rb->ra = ra; + rb->start_time = local_clock(); + rb->have_ioref = bch2_dev_get_ioref(ca, READ); + rb->idx = i; + rb->pick = pick; + rb->bio.bi_iter.bi_sector = pick.ptr.offset; + rb->bio.bi_end_io = btree_node_read_all_replicas_endio; + bch2_bio_map(&rb->bio, ra->buf[i], btree_bytes(c)); + + if (rb->have_ioref) { + this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree], + bio_sectors(&rb->bio)); + bio_set_dev(&rb->bio, ca->disk_sb.bdev); + + closure_get(&ra->cl); + submit_bio(&rb->bio); + } else { + ra->err[i] = BLK_STS_REMOVED; + } + + i++; + } + + if (sync) { + closure_sync(&ra->cl); + btree_node_read_all_replicas_done(&ra->cl.work); + } else { + continue_at(&ra->cl, btree_node_read_all_replicas_done, + c->io_complete_wq); + } + + return 0; +} + +void bch2_btree_node_read(struct btree_trans *trans, struct btree *b, bool sync) { + struct bch_fs *c = trans->c; struct extent_ptr_decoded pick; struct btree_read_bio *rb; struct bch_dev *ca; struct bio *bio; int ret; - trace_btree_read(c, b); + trace_and_count(c, btree_node_read, trans, b); + + if (bch2_verify_all_btree_replicas && + !btree_node_read_all_replicas(c, b, sync)) + return; ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key), NULL, &pick); - if (bch2_fs_fatal_err_on(ret <= 0, c, - "btree node read error: no device to read from")) { + + if (ret <= 0) { + struct printbuf buf = PRINTBUF; + + prt_str(&buf, "btree node read error: no device to read from\n at "); + bch2_btree_pos_to_text(&buf, c, b); + bch_err(c, "%s", buf.buf); + + if (c->recovery_passes_explicit & BIT_ULL(BCH_RECOVERY_PASS_check_topology) && + c->curr_recovery_pass > BCH_RECOVERY_PASS_check_topology) + bch2_fatal_error(c); + set_btree_node_read_error(b); + clear_btree_node_read_in_flight(b); + wake_up_bit(&b->flags, BTREE_NODE_read_in_flight); + printbuf_exit(&buf); return; } ca = bch_dev_bkey_exists(c, pick.ptr.dev); - bio = bio_alloc_bioset(GFP_NOIO, buf_pages(b->data, - btree_bytes(c)), + bio = bio_alloc_bioset(NULL, + buf_pages(b->data, btree_bytes(c)), + REQ_OP_READ|REQ_SYNC|REQ_META, + GFP_NOFS, &c->btree_bio); rb = container_of(bio, struct btree_read_bio, bio); rb->c = c; + rb->b = b; + rb->ra = NULL; rb->start_time = local_clock(); rb->have_ioref = bch2_dev_get_ioref(ca, READ); rb->pick = pick; INIT_WORK(&rb->work, btree_node_read_work); - bio->bi_opf = REQ_OP_READ|REQ_SYNC|REQ_META; bio->bi_iter.bi_sector = pick.ptr.offset; bio->bi_end_io = btree_node_read_endio; - bio->bi_private = b; bch2_bio_map(bio, b->data, btree_bytes(c)); - set_btree_node_read_in_flight(b); - if (rb->have_ioref) { this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree], bio_sectors(bio)); @@ -1208,7 +1646,6 @@ void bch2_btree_node_read(struct bch_fs *c, struct btree *b, if (sync) { submit_bio_wait(bio); - bio->bi_private = b; btree_node_read_work(&rb->work); } else { submit_bio(bio); @@ -1219,14 +1656,14 @@ void bch2_btree_node_read(struct bch_fs *c, struct btree *b, if (sync) btree_node_read_work(&rb->work); else - queue_work(system_unbound_wq, &rb->work); - + queue_work(c->io_complete_wq, &rb->work); } } -int bch2_btree_root_read(struct bch_fs *c, enum btree_id id, - const struct bkey_i *k, unsigned level) +static int __bch2_btree_root_read(struct btree_trans *trans, enum btree_id id, + const struct bkey_i *k, unsigned level) { + struct bch_fs *c = trans->c; struct closure cl; struct btree *b; int ret; @@ -1234,19 +1671,21 @@ int bch2_btree_root_read(struct bch_fs *c, enum btree_id id, closure_init_stack(&cl); do { - ret = bch2_btree_cache_cannibalize_lock(c, &cl); + ret = bch2_btree_cache_cannibalize_lock(trans, &cl); closure_sync(&cl); } while (ret); - b = bch2_btree_node_mem_alloc(c); - bch2_btree_cache_cannibalize_unlock(c); + b = bch2_btree_node_mem_alloc(trans, level != 0); + bch2_btree_cache_cannibalize_unlock(trans); BUG_ON(IS_ERR(b)); bkey_copy(&b->key, k); BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id)); - bch2_btree_node_read(c, b, true); + set_btree_node_read_in_flight(b); + + bch2_btree_node_read(trans, b, true); if (btree_node_read_error(b)) { bch2_btree_node_hash_remove(&c->btree_cache, b); @@ -1267,8 +1706,14 @@ err: return ret; } -void bch2_btree_complete_write(struct bch_fs *c, struct btree *b, - struct btree_write *w) +int bch2_btree_root_read(struct bch_fs *c, enum btree_id id, + const struct bkey_i *k, unsigned level) +{ + return bch2_trans_run(c, __bch2_btree_root_read(trans, id, k, level)); +} + +static void bch2_btree_complete_write(struct bch_fs *c, struct btree *b, + struct btree_write *w) { unsigned long old, new, v = READ_ONCE(b->will_make_reachable); @@ -1286,83 +1731,53 @@ void bch2_btree_complete_write(struct bch_fs *c, struct btree *b, bch2_journal_pin_drop(&c->journal, &w->journal); } -static void btree_node_write_done(struct bch_fs *c, struct btree *b) +static void __btree_node_write_done(struct bch_fs *c, struct btree *b) { struct btree_write *w = btree_prev_write(b); + unsigned long old, new, v; + unsigned type = 0; bch2_btree_complete_write(c, b, w); - btree_node_io_unlock(b); -} -static void bch2_btree_node_write_error(struct bch_fs *c, - struct btree_write_bio *wbio) -{ - struct btree *b = wbio->wbio.bio.bi_private; - __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp; - struct bch_extent_ptr *ptr; - struct btree_trans trans; - struct btree_iter *iter; - int ret; - - bch2_trans_init(&trans, c, 0, 0); - - iter = bch2_trans_get_node_iter(&trans, b->c.btree_id, b->key.k.p, - BTREE_MAX_DEPTH, b->c.level, 0); -retry: - ret = bch2_btree_iter_traverse(iter); - if (ret) - goto err; - - /* has node been freed? */ - if (iter->l[b->c.level].b != b) { - /* node has been freed: */ - BUG_ON(!btree_node_dying(b)); - goto out; - } - - BUG_ON(!btree_node_hashed(b)); - - bkey_copy(&tmp.k, &b->key); - - bch2_bkey_drop_ptrs(bkey_i_to_s(&tmp.k), ptr, - bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev)); + v = READ_ONCE(b->flags); + do { + old = new = v; - if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(&tmp.k))) - goto err; + if ((old & (1U << BTREE_NODE_dirty)) && + (old & (1U << BTREE_NODE_need_write)) && + !(old & (1U << BTREE_NODE_never_write)) && + !(old & (1U << BTREE_NODE_write_blocked)) && + !(old & (1U << BTREE_NODE_will_make_reachable))) { + new &= ~(1U << BTREE_NODE_dirty); + new &= ~(1U << BTREE_NODE_need_write); + new |= (1U << BTREE_NODE_write_in_flight); + new |= (1U << BTREE_NODE_write_in_flight_inner); + new |= (1U << BTREE_NODE_just_written); + new ^= (1U << BTREE_NODE_write_idx); + + type = new & BTREE_WRITE_TYPE_MASK; + new &= ~BTREE_WRITE_TYPE_MASK; + } else { + new &= ~(1U << BTREE_NODE_write_in_flight); + new &= ~(1U << BTREE_NODE_write_in_flight_inner); + } + } while ((v = cmpxchg(&b->flags, old, new)) != old); - ret = bch2_btree_node_update_key(c, iter, b, &tmp.k); - if (ret == -EINTR) - goto retry; - if (ret) - goto err; -out: - bch2_trans_exit(&trans); - bio_put(&wbio->wbio.bio); - btree_node_write_done(c, b); - return; -err: - set_btree_node_noevict(b); - bch2_fs_fatal_error(c, "fatal error writing btree node"); - goto out; + if (new & (1U << BTREE_NODE_write_in_flight)) + __bch2_btree_node_write(c, b, BTREE_WRITE_ALREADY_STARTED|type); + else + wake_up_bit(&b->flags, BTREE_NODE_write_in_flight); } -void bch2_btree_write_error_work(struct work_struct *work) +static void btree_node_write_done(struct bch_fs *c, struct btree *b) { - struct bch_fs *c = container_of(work, struct bch_fs, - btree_write_error_work); - struct bio *bio; + struct btree_trans *trans = bch2_trans_get(c); - while (1) { - spin_lock_irq(&c->btree_write_error_lock); - bio = bio_list_pop(&c->btree_write_error_list); - spin_unlock_irq(&c->btree_write_error_lock); - - if (!bio) - break; + btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read); + __btree_node_write_done(c, b); + six_unlock_read(&b->c.lock); - bch2_btree_node_write_error(c, - container_of(bio, struct btree_write_bio, wbio.bio)); - } + bch2_trans_put(trans); } static void btree_node_write_work(struct work_struct *work) @@ -1371,25 +1786,44 @@ static void btree_node_write_work(struct work_struct *work) container_of(work, struct btree_write_bio, work); struct bch_fs *c = wbio->wbio.c; struct btree *b = wbio->wbio.bio.bi_private; + struct bch_extent_ptr *ptr; + int ret = 0; btree_bounce_free(c, - wbio->bytes, + wbio->data_bytes, wbio->wbio.used_mempool, wbio->data); - if (wbio->wbio.failed.nr) { - unsigned long flags; + bch2_bkey_drop_ptrs(bkey_i_to_s(&wbio->key), ptr, + bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev)); - spin_lock_irqsave(&c->btree_write_error_lock, flags); - bio_list_add(&c->btree_write_error_list, &wbio->wbio.bio); - spin_unlock_irqrestore(&c->btree_write_error_lock, flags); + if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(&wbio->key))) + goto err; - queue_work(c->wq, &c->btree_write_error_work); - return; - } + if (wbio->wbio.first_btree_write) { + if (wbio->wbio.failed.nr) { + } + } else { + ret = bch2_trans_do(c, NULL, NULL, 0, + bch2_btree_node_update_key_get_iter(trans, b, &wbio->key, + BCH_WATERMARK_reclaim| + BCH_TRANS_COMMIT_journal_reclaim| + BCH_TRANS_COMMIT_no_enospc| + BCH_TRANS_COMMIT_no_check_rw, + !wbio->wbio.failed.nr)); + if (ret) + goto err; + } +out: bio_put(&wbio->wbio.bio); btree_node_write_done(c, b); + return; +err: + set_btree_node_noevict(b); + if (!bch2_err_matches(ret, EROFS)) + bch2_fs_fatal_error(c, "fatal error writing btree node: %s", bch2_err_str(ret)); + goto out; } static void btree_node_write_endio(struct bio *bio) @@ -1397,14 +1831,17 @@ static void btree_node_write_endio(struct bio *bio) struct bch_write_bio *wbio = to_wbio(bio); struct bch_write_bio *parent = wbio->split ? wbio->parent : NULL; struct bch_write_bio *orig = parent ?: wbio; + struct btree_write_bio *wb = container_of(orig, struct btree_write_bio, wbio); struct bch_fs *c = wbio->c; + struct btree *b = wbio->bio.bi_private; struct bch_dev *ca = bch_dev_bkey_exists(c, wbio->dev); unsigned long flags; if (wbio->have_ioref) bch2_latency_acct(ca, wbio->submit_time, WRITE); - if (bch2_dev_io_err_on(bio->bi_status, ca, "btree write: %s", + if (bch2_dev_io_err_on(bio->bi_status, ca, BCH_MEMBER_ERROR_write, + "btree write error: %s", bch2_blk_status_to_str(bio->bi_status)) || bch2_meta_write_fault("btree")) { spin_lock_irqsave(&c->btree_write_error_lock, flags); @@ -1418,53 +1855,75 @@ static void btree_node_write_endio(struct bio *bio) if (parent) { bio_put(bio); bio_endio(&parent->bio); - } else { - struct btree_write_bio *wb = - container_of(orig, struct btree_write_bio, wbio); - - INIT_WORK(&wb->work, btree_node_write_work); - queue_work(system_unbound_wq, &wb->work); + return; } + + clear_btree_node_write_in_flight_inner(b); + wake_up_bit(&b->flags, BTREE_NODE_write_in_flight_inner); + INIT_WORK(&wb->work, btree_node_write_work); + queue_work(c->btree_io_complete_wq, &wb->work); } static int validate_bset_for_write(struct bch_fs *c, struct btree *b, struct bset *i, unsigned sectors) { - unsigned whiteout_u64s = 0; + struct printbuf buf = PRINTBUF; + bool saw_error; int ret; - if (bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key), BKEY_TYPE_BTREE)) - return -1; + ret = bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key), + BKEY_TYPE_btree, WRITE, &buf); - ret = validate_bset(c, b, i, sectors, WRITE, false) ?: - validate_bset_keys(c, b, i, &whiteout_u64s, WRITE, false); if (ret) + bch2_fs_inconsistent(c, "invalid btree node key before write: %s", buf.buf); + printbuf_exit(&buf); + if (ret) + return ret; + + ret = validate_bset_keys(c, b, i, WRITE, false, &saw_error) ?: + validate_bset(c, NULL, b, i, b->written, sectors, WRITE, false, &saw_error); + if (ret) { bch2_inconsistent_error(c); + dump_stack(); + } return ret; } -void __bch2_btree_node_write(struct bch_fs *c, struct btree *b, - enum six_lock_type lock_type_held) +static void btree_write_submit(struct work_struct *work) +{ + struct btree_write_bio *wbio = container_of(work, struct btree_write_bio, work); + BKEY_PADDED_ONSTACK(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp; + + bkey_copy(&tmp.k, &wbio->key); + + bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&tmp.k)), ptr) + ptr->offset += wbio->sector_offset; + + bch2_submit_wbio_replicas(&wbio->wbio, wbio->wbio.c, BCH_DATA_btree, + &tmp.k, false); +} + +void __bch2_btree_node_write(struct bch_fs *c, struct btree *b, unsigned flags) { struct btree_write_bio *wbio; struct bset_tree *t; struct bset *i; struct btree_node *bn = NULL; struct btree_node_entry *bne = NULL; - BKEY_PADDED(key) k; - struct bch_extent_ptr *ptr; - struct sort_iter sort_iter; + struct sort_iter_stack sort_iter; struct nonce nonce; unsigned bytes_to_write, sectors_to_write, bytes, u64s; u64 seq = 0; bool used_mempool; unsigned long old, new; bool validate_before_checksum = false; + enum btree_write_type type = flags & BTREE_WRITE_TYPE_MASK; void *data; + int ret; - if (test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags)) - return; + if (flags & BTREE_WRITE_ALREADY_STARTED) + goto do_write; /* * We may only have a read lock on the btree node - the dirty bit is our @@ -1479,33 +1938,53 @@ void __bch2_btree_node_write(struct bch_fs *c, struct btree *b, if (!(old & (1 << BTREE_NODE_dirty))) return; - if (!btree_node_may_write(b)) + if ((flags & BTREE_WRITE_ONLY_IF_NEED) && + !(old & (1 << BTREE_NODE_need_write))) return; - if (old & (1 << BTREE_NODE_write_in_flight)) { - btree_node_wait_on_io(b); - continue; - } + if (old & + ((1 << BTREE_NODE_never_write)| + (1 << BTREE_NODE_write_blocked))) + return; + + if (b->written && + (old & (1 << BTREE_NODE_will_make_reachable))) + return; + + if (old & (1 << BTREE_NODE_write_in_flight)) + return; + + if (flags & BTREE_WRITE_ONLY_IF_NEED) + type = new & BTREE_WRITE_TYPE_MASK; + new &= ~BTREE_WRITE_TYPE_MASK; new &= ~(1 << BTREE_NODE_dirty); new &= ~(1 << BTREE_NODE_need_write); new |= (1 << BTREE_NODE_write_in_flight); + new |= (1 << BTREE_NODE_write_in_flight_inner); new |= (1 << BTREE_NODE_just_written); new ^= (1 << BTREE_NODE_write_idx); } while (cmpxchg_acquire(&b->flags, old, new) != old); + if (new & (1U << BTREE_NODE_need_write)) + return; +do_write: + BUG_ON((type == BTREE_WRITE_initial) != (b->written == 0)); + + atomic_dec(&c->btree_cache.dirty); + BUG_ON(btree_node_fake(b)); BUG_ON((b->will_make_reachable != 0) != !b->written); - BUG_ON(b->written >= c->opts.btree_node_size); - BUG_ON(b->written & (c->opts.block_size - 1)); + BUG_ON(b->written >= btree_sectors(c)); + BUG_ON(b->written & (block_sectors(c) - 1)); BUG_ON(bset_written(b, btree_bset_last(b))); BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c)); BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format))); bch2_sort_whiteouts(c, b); - sort_iter_init(&sort_iter, b); + sort_iter_stack_init(&sort_iter, b); bytes = !b->written ? sizeof(struct btree_node) @@ -1520,12 +1999,43 @@ void __bch2_btree_node_write(struct bch_fs *c, struct btree *b, continue; bytes += le16_to_cpu(i->u64s) * sizeof(u64); - sort_iter_add(&sort_iter, + sort_iter_add(&sort_iter.iter, btree_bkey_first(b, t), btree_bkey_last(b, t)); seq = max(seq, le64_to_cpu(i->journal_seq)); } + BUG_ON(b->written && !seq); + + /* bch2_varint_decode may read up to 7 bytes past the end of the buffer: */ + bytes += 8; + + /* buffer must be a multiple of the block size */ + bytes = round_up(bytes, block_bytes(c)); + + if (bytes > btree_bytes(c)) { + struct printbuf buf = PRINTBUF; + + prt_printf(&buf, "btree node write bounce buffer overrun: %u > %zu\n", + bytes, btree_bytes(c)); + + prt_printf(&buf, "header: %zu\n", b->written + ? sizeof(struct btree_node) + : sizeof(struct btree_node_entry)); + prt_printf(&buf, "unwritten: %zu\n", b->whiteout_u64s * sizeof(u64)); + + for_each_bset(b, t) { + i = bset(b, t); + + if (bset_written(b, i)) + continue; + prt_printf(&buf, "bset %zu: %zu\n", t - b->set, le16_to_cpu(i->u64s) * sizeof(u64)); + } + + panic("%s", buf.buf); + printbuf_exit(&buf); + } + data = btree_bounce_alloc(c, bytes, &used_mempool); if (!b->written) { @@ -1541,26 +2051,18 @@ void __bch2_btree_node_write(struct bch_fs *c, struct btree *b, i->journal_seq = cpu_to_le64(seq); i->u64s = 0; - if (!btree_node_old_extent_overwrite(b)) { - sort_iter_add(&sort_iter, - unwritten_whiteouts_start(c, b), - unwritten_whiteouts_end(c, b)); - SET_BSET_SEPARATE_WHITEOUTS(i, false); - } else { - memcpy_u64s(i->start, - unwritten_whiteouts_start(c, b), - b->whiteout_u64s); - i->u64s = cpu_to_le16(b->whiteout_u64s); - SET_BSET_SEPARATE_WHITEOUTS(i, true); - } + sort_iter_add(&sort_iter.iter, + unwritten_whiteouts_start(c, b), + unwritten_whiteouts_end(c, b)); + SET_BSET_SEPARATE_WHITEOUTS(i, false); b->whiteout_u64s = 0; - u64s = btree_node_old_extent_overwrite(b) - ? bch2_sort_extents(vstruct_last(i), &sort_iter, false) - : bch2_sort_keys(i->start, &sort_iter, false); + u64s = bch2_sort_keys(i->start, &sort_iter.iter, false); le16_add_cpu(&i->u64s, u64s); + BUG_ON(!b->written && i->u64s != b->data->keys.u64s); + set_needs_whiteout(i, false); /* do we have data to write? */ @@ -1570,23 +2072,26 @@ void __bch2_btree_node_write(struct bch_fs *c, struct btree *b, bytes_to_write = vstruct_end(i) - data; sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9; + if (!b->written && + b->key.k.type == KEY_TYPE_btree_ptr_v2) + BUG_ON(btree_ptr_sectors_written(&b->key) != sectors_to_write); + memset(data + bytes_to_write, 0, (sectors_to_write << 9) - bytes_to_write); - BUG_ON(b->written + sectors_to_write > c->opts.btree_node_size); + BUG_ON(b->written + sectors_to_write > btree_sectors(c)); BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN); BUG_ON(i->seq != b->data->keys.seq); - i->version = c->sb.version < bcachefs_metadata_version_new_versioning - ? cpu_to_le16(BCH_BSET_VERSION_OLD) - : cpu_to_le16(c->sb.version); + i->version = cpu_to_le16(c->sb.version); + SET_BSET_OFFSET(i, b->written); SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c)); if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i))) validate_before_checksum = true; /* validate_bset will be modifying: */ - if (le16_to_cpu(i->version) < bcachefs_metadata_version_max) + if (le16_to_cpu(i->version) < bcachefs_metadata_version_current) validate_before_checksum = true; /* if we're going to be encrypting, check metadata validity first: */ @@ -1594,7 +2099,10 @@ void __bch2_btree_node_write(struct bch_fs *c, struct btree *b, validate_bset_for_write(c, b, i, sectors_to_write)) goto err; - bset_encrypt(c, i, b->written << 9); + ret = bset_encrypt(c, i, b->written << 9); + if (bch2_fs_fatal_err_on(ret, c, + "error encrypting btree node: %i\n", ret)) + goto err; nonce = btree_nonce(i, b->written << 9); @@ -1630,53 +2138,46 @@ void __bch2_btree_node_write(struct bch_fs *c, struct btree *b, c->opts.nochanges) goto err; - trace_btree_write(b, bytes_to_write, sectors_to_write); + trace_and_count(c, btree_node_write, b, bytes_to_write, sectors_to_write); - wbio = container_of(bio_alloc_bioset(GFP_NOIO, + wbio = container_of(bio_alloc_bioset(NULL, buf_pages(data, sectors_to_write << 9), + REQ_OP_WRITE|REQ_META, + GFP_NOFS, &c->btree_bio), struct btree_write_bio, wbio.bio); wbio_init(&wbio->wbio.bio); wbio->data = data; - wbio->bytes = bytes; + wbio->data_bytes = bytes; + wbio->sector_offset = b->written; + wbio->wbio.c = c; wbio->wbio.used_mempool = used_mempool; - wbio->wbio.bio.bi_opf = REQ_OP_WRITE|REQ_META; + wbio->wbio.first_btree_write = !b->written; wbio->wbio.bio.bi_end_io = btree_node_write_endio; wbio->wbio.bio.bi_private = b; bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9); - /* - * If we're appending to a leaf node, we don't technically need FUA - - * this write just needs to be persisted before the next journal write, - * which will be marked FLUSH|FUA. - * - * Similarly if we're writing a new btree root - the pointer is going to - * be in the next journal entry. - * - * But if we're writing a new btree node (that isn't a root) or - * appending to a non leaf btree node, we need either FUA or a flush - * when we write the parent with the new pointer. FUA is cheaper than a - * flush, and writes appending to leaf nodes aren't blocking anything so - * just make all btree node writes FUA to keep things sane. - */ + bkey_copy(&wbio->key, &b->key); - bkey_copy(&k.key, &b->key); + b->written += sectors_to_write; - bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&k.key)), ptr) - ptr->offset += b->written; + if (wbio->key.k.type == KEY_TYPE_btree_ptr_v2) + bkey_i_to_btree_ptr_v2(&wbio->key)->v.sectors_written = + cpu_to_le16(b->written); - b->written += sectors_to_write; + atomic64_inc(&c->btree_write_stats[type].nr); + atomic64_add(bytes_to_write, &c->btree_write_stats[type].bytes); - /* XXX: submitting IO with btree locks held: */ - bch2_submit_wbio_replicas(&wbio->wbio, c, BCH_DATA_btree, &k.key); + INIT_WORK(&wbio->work, btree_write_submit); + queue_work(c->io_complete_wq, &wbio->work); return; err: set_btree_node_noevict(b); b->written += sectors_to_write; nowrite: btree_bounce_free(c, bytes, used_mempool, data); - btree_node_write_done(c, b); + __btree_node_write_done(c, b); } /* @@ -1709,7 +2210,7 @@ bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b) * single bset: */ if (b->nsets > 1) { - btree_node_sort(c, b, NULL, 0, b->nsets, true); + btree_node_sort(c, b, 0, b->nsets, true); invalidated_iter = true; } else { invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL); @@ -1739,13 +2240,13 @@ bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b) * Use this one if the node is intent locked: */ void bch2_btree_node_write(struct bch_fs *c, struct btree *b, - enum six_lock_type lock_type_held) + enum six_lock_type lock_type_held, + unsigned flags) { - BUG_ON(lock_type_held == SIX_LOCK_write); - if (lock_type_held == SIX_LOCK_intent || - six_lock_tryupgrade(&b->c.lock)) { - __bch2_btree_node_write(c, b, SIX_LOCK_intent); + (lock_type_held == SIX_LOCK_read && + six_lock_tryupgrade(&b->c.lock))) { + __bch2_btree_node_write(c, b, flags); /* don't cycle lock unnecessarily: */ if (btree_node_just_written(b) && @@ -1757,78 +2258,70 @@ void bch2_btree_node_write(struct bch_fs *c, struct btree *b, if (lock_type_held == SIX_LOCK_read) six_lock_downgrade(&b->c.lock); } else { - __bch2_btree_node_write(c, b, SIX_LOCK_read); + __bch2_btree_node_write(c, b, flags); + if (lock_type_held == SIX_LOCK_write && + btree_node_just_written(b)) + bch2_btree_post_write_cleanup(c, b); } } -static void __bch2_btree_flush_all(struct bch_fs *c, unsigned flag) +static bool __bch2_btree_flush_all(struct bch_fs *c, unsigned flag) { struct bucket_table *tbl; struct rhash_head *pos; struct btree *b; unsigned i; + bool ret = false; restart: rcu_read_lock(); for_each_cached_btree(b, c, tbl, i, pos) if (test_bit(flag, &b->flags)) { rcu_read_unlock(); wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE); + ret = true; goto restart; - } rcu_read_unlock(); -} -void bch2_btree_flush_all_reads(struct bch_fs *c) -{ - __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight); + return ret; } -void bch2_btree_flush_all_writes(struct bch_fs *c) +bool bch2_btree_flush_all_reads(struct bch_fs *c) { - __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight); + return __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight); } -void bch2_btree_verify_flushed(struct bch_fs *c) +bool bch2_btree_flush_all_writes(struct bch_fs *c) { - struct bucket_table *tbl; - struct rhash_head *pos; - struct btree *b; - unsigned i; - - rcu_read_lock(); - for_each_cached_btree(b, c, tbl, i, pos) { - unsigned long flags = READ_ONCE(b->flags); - - BUG_ON((flags & (1 << BTREE_NODE_dirty)) || - (flags & (1 << BTREE_NODE_write_in_flight))); - } - rcu_read_unlock(); + return __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight); } -void bch2_dirty_btree_nodes_to_text(struct printbuf *out, struct bch_fs *c) -{ - struct bucket_table *tbl; - struct rhash_head *pos; - struct btree *b; - unsigned i; - - rcu_read_lock(); - for_each_cached_btree(b, c, tbl, i, pos) { - unsigned long flags = READ_ONCE(b->flags); - - if (!(flags & (1 << BTREE_NODE_dirty))) - continue; +static const char * const bch2_btree_write_types[] = { +#define x(t, n) [n] = #t, + BCH_BTREE_WRITE_TYPES() + NULL +}; - pr_buf(out, "%p d %u n %u l %u w %u b %u r %u:%lu\n", - b, - (flags & (1 << BTREE_NODE_dirty)) != 0, - (flags & (1 << BTREE_NODE_need_write)) != 0, - b->c.level, - b->written, - !list_empty_careful(&b->write_blocked), - b->will_make_reachable != 0, - b->will_make_reachable & 1); +void bch2_btree_write_stats_to_text(struct printbuf *out, struct bch_fs *c) +{ + printbuf_tabstop_push(out, 20); + printbuf_tabstop_push(out, 10); + + prt_tab(out); + prt_str(out, "nr"); + prt_tab(out); + prt_str(out, "size"); + prt_newline(out); + + for (unsigned i = 0; i < BTREE_WRITE_TYPE_NR; i++) { + u64 nr = atomic64_read(&c->btree_write_stats[i].nr); + u64 bytes = atomic64_read(&c->btree_write_stats[i].bytes); + + prt_printf(out, "%s:", bch2_btree_write_types[i]); + prt_tab(out); + prt_u64(out, nr); + prt_tab(out); + prt_human_readable_u64(out, nr ? div64_u64(bytes, nr) : 0); + prt_newline(out); } - rcu_read_unlock(); }