X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libbcachefs%2Ffs-io.c;h=f3b2e3eeb86b9599458e72974116768f0afb8fd2;hb=ba88873432ec9621b9913c00b31d7401357b5e55;hp=b2ea783fe38afc29855b47eca4803b84cf25e9f3;hpb=b485aae1bac95e3b5be235116e2bc43da85906c5;p=bcachefs-tools-debian diff --git a/libbcachefs/fs-io.c b/libbcachefs/fs-io.c index b2ea783..f3b2e3e 100644 --- a/libbcachefs/fs-io.c +++ b/libbcachefs/fs-io.c @@ -1,12 +1,15 @@ +// SPDX-License-Identifier: GPL-2.0 #ifndef NO_BCACHEFS_FS #include "bcachefs.h" #include "alloc_foreground.h" +#include "bkey_buf.h" #include "btree_update.h" #include "buckets.h" #include "clock.h" #include "error.h" #include "extents.h" +#include "extent_update.h" #include "fs.h" #include "fs-io.h" #include "fsck.h" @@ -15,6 +18,8 @@ #include "io.h" #include "keylist.h" #include "quota.h" +#include "reflink.h" +#include "trace.h" #include #include @@ -22,62 +27,207 @@ #include #include #include +#include #include #include #include #include -#include #include -struct quota_res { - u64 sectors; +/* + * Use u64 for the end pos and sector helpers because if the folio covers the + * max supported range of the mapping, the start offset of the next folio + * overflows loff_t. This breaks much of the range based processing in the + * buffered write path. + */ +static inline u64 folio_end_pos(struct folio *folio) +{ + return folio_pos(folio) + folio_size(folio); +} + +static inline size_t folio_sectors(struct folio *folio) +{ + return PAGE_SECTORS << folio_order(folio); +} + +static inline loff_t folio_sector(struct folio *folio) +{ + return folio_pos(folio) >> 9; +} + +static inline u64 folio_end_sector(struct folio *folio) +{ + return folio_end_pos(folio) >> 9; +} + +typedef DARRAY(struct folio *) folios; + +static int filemap_get_contig_folios_d(struct address_space *mapping, + loff_t start, u64 end, + int fgp_flags, gfp_t gfp, + folios *folios) +{ + struct folio *f; + u64 pos = start; + int ret = 0; + + while (pos < end) { + if ((u64) pos >= (u64) start + (1ULL << 20)) + fgp_flags &= ~FGP_CREAT; + + ret = darray_make_room_gfp(folios, 1, gfp & GFP_KERNEL); + if (ret) + break; + + f = __filemap_get_folio(mapping, pos >> PAGE_SHIFT, fgp_flags, gfp); + if (!f) + break; + + BUG_ON(folios->nr && folio_pos(f) != pos); + + pos = folio_end_pos(f); + darray_push(folios, f); + } + + if (!folios->nr && !ret && (fgp_flags & FGP_CREAT)) + ret = -ENOMEM; + + return folios->nr ? 0 : ret; +} + +struct nocow_flush { + struct closure *cl; + struct bch_dev *ca; + struct bio bio; }; -struct bchfs_write_op { - struct bch_inode_info *inode; - s64 sectors_added; - bool is_dio; - bool unalloc; - u64 new_i_size; +static void nocow_flush_endio(struct bio *_bio) +{ - /* must be last: */ - struct bch_write_op op; + struct nocow_flush *bio = container_of(_bio, struct nocow_flush, bio); + + closure_put(bio->cl); + percpu_ref_put(&bio->ca->io_ref); + bio_put(&bio->bio); +} + +static void bch2_inode_flush_nocow_writes_async(struct bch_fs *c, + struct bch_inode_info *inode, + struct closure *cl) +{ + struct nocow_flush *bio; + struct bch_dev *ca; + struct bch_devs_mask devs; + unsigned dev; + + dev = find_first_bit(inode->ei_devs_need_flush.d, BCH_SB_MEMBERS_MAX); + if (dev == BCH_SB_MEMBERS_MAX) + return; + + devs = inode->ei_devs_need_flush; + memset(&inode->ei_devs_need_flush, 0, sizeof(inode->ei_devs_need_flush)); + + for_each_set_bit(dev, devs.d, BCH_SB_MEMBERS_MAX) { + rcu_read_lock(); + ca = rcu_dereference(c->devs[dev]); + if (ca && !percpu_ref_tryget(&ca->io_ref)) + ca = NULL; + rcu_read_unlock(); + + if (!ca) + continue; + + bio = container_of(bio_alloc_bioset(ca->disk_sb.bdev, 0, + REQ_OP_FLUSH, + GFP_KERNEL, + &c->nocow_flush_bioset), + struct nocow_flush, bio); + bio->cl = cl; + bio->ca = ca; + bio->bio.bi_end_io = nocow_flush_endio; + closure_bio_submit(&bio->bio, cl); + } +} + +static int bch2_inode_flush_nocow_writes(struct bch_fs *c, + struct bch_inode_info *inode) +{ + struct closure cl; + + closure_init_stack(&cl); + bch2_inode_flush_nocow_writes_async(c, inode, &cl); + closure_sync(&cl); + + return 0; +} + +static inline bool bio_full(struct bio *bio, unsigned len) +{ + if (bio->bi_vcnt >= bio->bi_max_vecs) + return true; + if (bio->bi_iter.bi_size > UINT_MAX - len) + return true; + return false; +} + +static inline struct address_space *faults_disabled_mapping(void) +{ + return (void *) (((unsigned long) current->faults_disabled_mapping) & ~1UL); +} + +static inline void set_fdm_dropped_locks(void) +{ + current->faults_disabled_mapping = + (void *) (((unsigned long) current->faults_disabled_mapping)|1); +} + +static inline bool fdm_dropped_locks(void) +{ + return ((unsigned long) current->faults_disabled_mapping) & 1; +} + +struct quota_res { + u64 sectors; }; struct bch_writepage_io { - struct closure cl; - u64 new_sectors; + struct bch_inode_info *inode; /* must be last: */ - struct bchfs_write_op op; + struct bch_write_op op; }; struct dio_write { - struct closure cl; struct kiocb *req; + struct address_space *mapping; + struct bch_inode_info *inode; struct mm_struct *mm; unsigned loop:1, + extending:1, sync:1, + flush:1, free_iov:1; struct quota_res quota_res; + u64 written; struct iov_iter iter; struct iovec inline_vecs[2]; /* must be last: */ - struct bchfs_write_op iop; + struct bch_write_op op; }; struct dio_read { struct closure cl; struct kiocb *req; long ret; + bool should_dirty; struct bch_read_bio rbio; }; /* pagecache_block must be held */ -static int write_invalidate_inode_pages_range(struct address_space *mapping, +static noinline int write_invalidate_inode_pages_range(struct address_space *mapping, loff_t start, loff_t end) { int ret; @@ -87,8 +237,7 @@ static int write_invalidate_inode_pages_range(struct address_space *mapping, * is continually redirtying a specific page */ do { - if (!mapping->nrpages && - !mapping->nrexceptional) + if (!mapping->nrpages) return 0; ret = filemap_write_and_wait_range(mapping, start, end); @@ -110,32 +259,40 @@ static int write_invalidate_inode_pages_range(struct address_space *mapping, #ifdef CONFIG_BCACHEFS_QUOTA -static void bch2_quota_reservation_put(struct bch_fs *c, - struct bch_inode_info *inode, - struct quota_res *res) +static void __bch2_quota_reservation_put(struct bch_fs *c, + struct bch_inode_info *inode, + struct quota_res *res) { - if (!res->sectors) - return; - - mutex_lock(&inode->ei_quota_lock); BUG_ON(res->sectors > inode->ei_quota_reserved); bch2_quota_acct(c, inode->ei_qid, Q_SPC, -((s64) res->sectors), KEY_TYPE_QUOTA_PREALLOC); inode->ei_quota_reserved -= res->sectors; - mutex_unlock(&inode->ei_quota_lock); - res->sectors = 0; } +static void bch2_quota_reservation_put(struct bch_fs *c, + struct bch_inode_info *inode, + struct quota_res *res) +{ + if (res->sectors) { + mutex_lock(&inode->ei_quota_lock); + __bch2_quota_reservation_put(c, inode, res); + mutex_unlock(&inode->ei_quota_lock); + } +} + static int bch2_quota_reservation_add(struct bch_fs *c, struct bch_inode_info *inode, struct quota_res *res, - unsigned sectors, + u64 sectors, bool check_enospc) { int ret; + if (test_bit(EI_INODE_SNAPSHOT, &inode->ei_flags)) + return 0; + mutex_lock(&inode->ei_quota_lock); ret = bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors, check_enospc ? KEY_TYPE_QUOTA_PREALLOC : KEY_TYPE_QUOTA_NOCHECK); @@ -150,11 +307,13 @@ static int bch2_quota_reservation_add(struct bch_fs *c, #else +static void __bch2_quota_reservation_put(struct bch_fs *c, + struct bch_inode_info *inode, + struct quota_res *res) {} + static void bch2_quota_reservation_put(struct bch_fs *c, struct bch_inode_info *inode, - struct quota_res *res) -{ -} + struct quota_res *res) {} static int bch2_quota_reservation_add(struct bch_fs *c, struct bch_inode_info *inode, @@ -192,9 +351,9 @@ static int inode_set_size(struct bch_inode_info *inode, return 0; } -static int __must_check bch2_write_inode_size(struct bch_fs *c, - struct bch_inode_info *inode, - loff_t new_size, unsigned fields) +int __must_check bch2_write_inode_size(struct bch_fs *c, + struct bch_inode_info *inode, + loff_t new_size, unsigned fields) { struct inode_new_size s = { .new_size = new_size, @@ -205,15 +364,19 @@ static int __must_check bch2_write_inode_size(struct bch_fs *c, return bch2_write_inode(c, inode, inode_set_size, &s, fields); } -static void i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode, +static void __i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode, struct quota_res *quota_res, s64 sectors) { - if (!sectors) - return; + bch2_fs_inconsistent_on((s64) inode->v.i_blocks + sectors < 0, c, + "inode %lu i_blocks underflow: %llu + %lli < 0 (ondisk %lli)", + inode->v.i_ino, (u64) inode->v.i_blocks, sectors, + inode->ei_inode.bi_sectors); + inode->v.i_blocks += sectors; - mutex_lock(&inode->ei_quota_lock); #ifdef CONFIG_BCACHEFS_QUOTA - if (quota_res && sectors > 0) { + if (quota_res && + !test_bit(EI_INODE_SNAPSHOT, &inode->ei_flags) && + sectors > 0) { BUG_ON(sectors > quota_res->sectors); BUG_ON(sectors > inode->ei_quota_reserved); @@ -223,469 +386,619 @@ static void i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode, bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors, KEY_TYPE_QUOTA_WARN); } #endif - inode->v.i_blocks += sectors; - mutex_unlock(&inode->ei_quota_lock); } -/* normal i_size/i_sectors update machinery: */ - -static int sum_sector_overwrites(struct btree_trans *trans, - struct btree_iter *extent_iter, - struct bkey_i *new, bool *allocating, - s64 *delta) +static void i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode, + struct quota_res *quota_res, s64 sectors) { - struct btree_iter *iter; - struct bkey_s_c old; + if (sectors) { + mutex_lock(&inode->ei_quota_lock); + __i_sectors_acct(c, inode, quota_res, sectors); + mutex_unlock(&inode->ei_quota_lock); + } +} - *delta = 0; +/* page state: */ - iter = bch2_trans_copy_iter(trans, extent_iter); - if (IS_ERR(iter)) - return PTR_ERR(iter); +/* stored in page->private: */ - old = bch2_btree_iter_peek_slot(iter); +#define BCH_FOLIO_SECTOR_STATE() \ + x(unallocated) \ + x(reserved) \ + x(dirty) \ + x(dirty_reserved) \ + x(allocated) + +enum bch_folio_sector_state { +#define x(n) SECTOR_##n, + BCH_FOLIO_SECTOR_STATE() +#undef x +}; - while (1) { - /* - * should not be possible to get an error here, since we're - * carefully not advancing past @new and thus whatever leaf node - * @_iter currently points to: - */ - BUG_ON(bkey_err(old)); - - if (allocating && - !*allocating && - bch2_bkey_nr_ptrs_allocated(old) < - bch2_bkey_nr_dirty_ptrs(bkey_i_to_s_c(new))) - *allocating = true; - - *delta += (min(new->k.p.offset, - old.k->p.offset) - - max(bkey_start_offset(&new->k), - bkey_start_offset(old.k))) * - (bkey_extent_is_allocation(&new->k) - - bkey_extent_is_allocation(old.k)); - - if (bkey_cmp(old.k->p, new->k.p) >= 0) - break; +const char * const bch2_folio_sector_states[] = { +#define x(n) #n, + BCH_FOLIO_SECTOR_STATE() +#undef x + NULL +}; - old = bch2_btree_iter_next_slot(iter); +static inline enum bch_folio_sector_state +folio_sector_dirty(enum bch_folio_sector_state state) +{ + switch (state) { + case SECTOR_unallocated: + return SECTOR_dirty; + case SECTOR_reserved: + return SECTOR_dirty_reserved; + default: + return state; } - - bch2_trans_iter_free(trans, iter); - return 0; } -static int bch2_extent_update(struct btree_trans *trans, - struct bch_inode_info *inode, - struct disk_reservation *disk_res, - struct quota_res *quota_res, - struct btree_iter *extent_iter, - struct bkey_i *k, - u64 new_i_size, - bool may_allocate, - bool direct, - s64 *total_delta) +static inline enum bch_folio_sector_state +folio_sector_undirty(enum bch_folio_sector_state state) { - struct bch_fs *c = trans->c; - struct btree_iter *inode_iter = NULL; - struct bch_inode_unpacked inode_u; - struct bkey_inode_buf inode_p; - bool allocating = false; - bool extended = false; - bool inode_locked = false; - s64 i_sectors_delta; - int ret; - - bch2_trans_begin_updates(trans); + switch (state) { + case SECTOR_dirty: + return SECTOR_unallocated; + case SECTOR_dirty_reserved: + return SECTOR_reserved; + default: + return state; + } +} - ret = bch2_btree_iter_traverse(extent_iter); - if (ret) - return ret; +static inline enum bch_folio_sector_state +folio_sector_reserve(enum bch_folio_sector_state state) +{ + switch (state) { + case SECTOR_unallocated: + return SECTOR_reserved; + case SECTOR_dirty: + return SECTOR_dirty_reserved; + default: + return state; + } +} - bch2_extent_trim_atomic(k, extent_iter); +struct bch_folio_sector { + /* Uncompressed, fully allocated replicas (or on disk reservation): */ + unsigned nr_replicas:4; - ret = sum_sector_overwrites(trans, extent_iter, - k, &allocating, - &i_sectors_delta); - if (ret) - return ret; + /* Owns PAGE_SECTORS * replicas_reserved sized in memory reservation: */ + unsigned replicas_reserved:4; - if (!may_allocate && allocating) - return -ENOSPC; + /* i_sectors: */ + enum bch_folio_sector_state state:8; +}; - bch2_trans_update(trans, BTREE_INSERT_ENTRY(extent_iter, k)); +struct bch_folio { + spinlock_t lock; + atomic_t write_count; + /* + * Is the sector state up to date with the btree? + * (Not the data itself) + */ + bool uptodate; + struct bch_folio_sector s[]; +}; - new_i_size = min(k->k.p.offset << 9, new_i_size); +static inline void folio_sector_set(struct folio *folio, + struct bch_folio *s, + unsigned i, unsigned n) +{ + s->s[i].state = n; +} - /* XXX: inode->i_size locking */ - if (i_sectors_delta || - new_i_size > inode->ei_inode.bi_size) { - if (c->opts.new_inode_updates) { - bch2_btree_trans_unlock(trans); - mutex_lock(&inode->ei_update_lock); +/* file offset (to folio offset) to bch_folio_sector index */ +static inline int folio_pos_to_s(struct folio *folio, loff_t pos) +{ + u64 f_offset = pos - folio_pos(folio); + BUG_ON(pos < folio_pos(folio) || pos >= folio_end_pos(folio)); + return f_offset >> SECTOR_SHIFT; +} - if (!bch2_btree_trans_relock(trans)) { - mutex_unlock(&inode->ei_update_lock); - return -EINTR; - } +static inline struct bch_folio *__bch2_folio(struct folio *folio) +{ + return folio_has_private(folio) + ? (struct bch_folio *) folio_get_private(folio) + : NULL; +} - inode_locked = true; +static inline struct bch_folio *bch2_folio(struct folio *folio) +{ + EBUG_ON(!folio_test_locked(folio)); - if (!inode->ei_inode_update) - inode->ei_inode_update = - bch2_deferred_update_alloc(c, - BTREE_ID_INODES, 64); + return __bch2_folio(folio); +} - inode_u = inode->ei_inode; - inode_u.bi_sectors += i_sectors_delta; +/* for newly allocated folios: */ +static void __bch2_folio_release(struct folio *folio) +{ + kfree(folio_detach_private(folio)); +} - /* XXX: this is slightly suspect */ - if (!(inode_u.bi_flags & BCH_INODE_I_SIZE_DIRTY) && - new_i_size > inode_u.bi_size) { - inode_u.bi_size = new_i_size; - extended = true; - } +static void bch2_folio_release(struct folio *folio) +{ + EBUG_ON(!folio_test_locked(folio)); + __bch2_folio_release(folio); +} - bch2_inode_pack(&inode_p, &inode_u); - bch2_trans_update(trans, - BTREE_INSERT_DEFERRED(inode->ei_inode_update, - &inode_p.inode.k_i)); - } else { - inode_iter = bch2_trans_get_iter(trans, - BTREE_ID_INODES, - POS(k->k.p.inode, 0), - BTREE_ITER_SLOTS|BTREE_ITER_INTENT); - if (IS_ERR(inode_iter)) - return PTR_ERR(inode_iter); - - ret = bch2_btree_iter_traverse(inode_iter); - if (ret) - goto err; +/* for newly allocated folios: */ +static struct bch_folio *__bch2_folio_create(struct folio *folio, gfp_t gfp) +{ + struct bch_folio *s; - inode_u = inode->ei_inode; - inode_u.bi_sectors += i_sectors_delta; + s = kzalloc(sizeof(*s) + + sizeof(struct bch_folio_sector) * + folio_sectors(folio), gfp); + if (!s) + return NULL; - /* XXX: this is slightly suspect */ - if (!(inode_u.bi_flags & BCH_INODE_I_SIZE_DIRTY) && - new_i_size > inode_u.bi_size) { - inode_u.bi_size = new_i_size; - extended = true; - } + spin_lock_init(&s->lock); + folio_attach_private(folio, s); + return s; +} - bch2_inode_pack(&inode_p, &inode_u); - bch2_trans_update(trans, - BTREE_INSERT_ENTRY(inode_iter, &inode_p.inode.k_i)); - } - } +static struct bch_folio *bch2_folio_create(struct folio *folio, gfp_t gfp) +{ + return bch2_folio(folio) ?: __bch2_folio_create(folio, gfp); +} - ret = bch2_trans_commit(trans, disk_res, - &inode->ei_journal_seq, - BTREE_INSERT_NOFAIL| - BTREE_INSERT_ATOMIC| - BTREE_INSERT_NOUNLOCK| - BTREE_INSERT_USE_RESERVE); - if (ret) - goto err; +static unsigned bkey_to_sector_state(struct bkey_s_c k) +{ + if (bkey_extent_is_reservation(k)) + return SECTOR_reserved; + if (bkey_extent_is_allocation(k.k)) + return SECTOR_allocated; + return SECTOR_unallocated; +} - inode->ei_inode.bi_sectors += i_sectors_delta; +static void __bch2_folio_set(struct folio *folio, + unsigned pg_offset, unsigned pg_len, + unsigned nr_ptrs, unsigned state) +{ + struct bch_folio *s = bch2_folio(folio); + unsigned i, sectors = folio_sectors(folio); - EBUG_ON(i_sectors_delta && - inode->ei_inode.bi_sectors != inode_u.bi_sectors); + BUG_ON(pg_offset >= sectors); + BUG_ON(pg_offset + pg_len > sectors); - if (extended) { - inode->ei_inode.bi_size = new_i_size; + spin_lock(&s->lock); - if (direct) { - spin_lock(&inode->v.i_lock); - if (new_i_size > inode->v.i_size) - i_size_write(&inode->v, new_i_size); - spin_unlock(&inode->v.i_lock); - } + for (i = pg_offset; i < pg_offset + pg_len; i++) { + s->s[i].nr_replicas = nr_ptrs; + folio_sector_set(folio, s, i, state); } - if (direct) - i_sectors_acct(c, inode, quota_res, i_sectors_delta); - - if (total_delta) - *total_delta += i_sectors_delta; -err: - if (!IS_ERR_OR_NULL(inode_iter)) - bch2_trans_iter_put(trans, inode_iter); - if (inode_locked) - mutex_unlock(&inode->ei_update_lock); + if (i == sectors) + s->uptodate = true; - return ret; + spin_unlock(&s->lock); } -static int bchfs_write_index_update(struct bch_write_op *wop) +/* + * Initialize bch_folio state (allocated/unallocated, nr_replicas) from the + * extents btree: + */ +static int bch2_folio_set(struct bch_fs *c, subvol_inum inum, + struct folio **folios, unsigned nr_folios) { - struct bch_fs *c = wop->c; - struct bchfs_write_op *op = container_of(wop, - struct bchfs_write_op, op); - struct quota_res *quota_res = op->is_dio - ? &container_of(op, struct dio_write, iop)->quota_res - : NULL; - struct bch_inode_info *inode = op->inode; - struct keylist *keys = &op->op.insert_keys; - struct bkey_i *k = bch2_keylist_front(keys); struct btree_trans trans; - struct btree_iter *iter; + struct btree_iter iter; + struct bkey_s_c k; + struct bch_folio *s; + u64 offset = folio_sector(folios[0]); + unsigned folio_idx; + u32 snapshot; + bool need_set = false; int ret; - BUG_ON(k->k.p.inode != inode->v.i_ino); + for (folio_idx = 0; folio_idx < nr_folios; folio_idx++) { + s = bch2_folio_create(folios[folio_idx], GFP_KERNEL); + if (!s) + return -ENOMEM; + + need_set |= !s->uptodate; + } + + if (!need_set) + return 0; - bch2_trans_init(&trans, c); - bch2_trans_preload_iters(&trans); + folio_idx = 0; + bch2_trans_init(&trans, c, 0, 0); +retry: + bch2_trans_begin(&trans); - iter = bch2_trans_get_iter(&trans, - BTREE_ID_EXTENTS, - bkey_start_pos(&k->k), - BTREE_ITER_INTENT); + ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot); + if (ret) + goto err; - do { - BKEY_PADDED(k) tmp; - - bkey_copy(&tmp.k, bch2_keylist_front(keys)); - - ret = bch2_extent_update(&trans, inode, - &wop->res, quota_res, - iter, &tmp.k, - op->new_i_size, - !op->unalloc, - op->is_dio, - &op->sectors_added); - if (ret == -EINTR) - continue; - if (ret) - break; + for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents, + SPOS(inum.inum, offset, snapshot), + BTREE_ITER_SLOTS, k, ret) { + unsigned nr_ptrs = bch2_bkey_nr_ptrs_fully_allocated(k); + unsigned state = bkey_to_sector_state(k); + + while (folio_idx < nr_folios) { + struct folio *folio = folios[folio_idx]; + u64 folio_start = folio_sector(folio); + u64 folio_end = folio_end_sector(folio); + unsigned folio_offset = max(bkey_start_offset(k.k), folio_start) - folio_start; + unsigned folio_len = min(k.k->p.offset, folio_end) - folio_offset - folio_start; + + BUG_ON(k.k->p.offset < folio_start); + BUG_ON(bkey_start_offset(k.k) > folio_end); + + if (!bch2_folio(folio)->uptodate) + __bch2_folio_set(folio, folio_offset, folio_len, nr_ptrs, state); - if (bkey_cmp(iter->pos, bch2_keylist_front(keys)->k.p) < 0) - bch2_cut_front(iter->pos, bch2_keylist_front(keys)); - else - bch2_keylist_pop_front(keys); - } while (!bch2_keylist_empty(keys)); + if (k.k->p.offset < folio_end) + break; + folio_idx++; + } + + if (folio_idx == nr_folios) + break; + } + offset = iter.pos.offset; + bch2_trans_iter_exit(&trans, &iter); +err: + if (bch2_err_matches(ret, BCH_ERR_transaction_restart)) + goto retry; bch2_trans_exit(&trans); return ret; } -static inline void bch2_fswrite_op_init(struct bchfs_write_op *op, - struct bch_fs *c, - struct bch_inode_info *inode, - struct bch_io_opts opts, - bool is_dio) +static void bch2_bio_page_state_set(struct bio *bio, struct bkey_s_c k) { - op->inode = inode; - op->sectors_added = 0; - op->is_dio = is_dio; - op->unalloc = false; - op->new_i_size = U64_MAX; + struct bvec_iter iter; + struct folio_vec fv; + unsigned nr_ptrs = k.k->type == KEY_TYPE_reflink_v + ? 0 : bch2_bkey_nr_ptrs_fully_allocated(k); + unsigned state = bkey_to_sector_state(k); - bch2_write_op_init(&op->op, c, opts); - op->op.target = opts.foreground_target; - op->op.index_update_fn = bchfs_write_index_update; - op_journal_seq_set(&op->op, &inode->ei_journal_seq); + bio_for_each_folio(fv, bio, iter) + __bch2_folio_set(fv.fv_folio, + fv.fv_offset >> 9, + fv.fv_len >> 9, + nr_ptrs, state); } -static inline struct bch_io_opts io_opts(struct bch_fs *c, struct bch_inode_info *inode) +static void mark_pagecache_unallocated(struct bch_inode_info *inode, + u64 start, u64 end) { - struct bch_io_opts opts = bch2_opts_to_inode_opts(c->opts); + pgoff_t index = start >> PAGE_SECTORS_SHIFT; + pgoff_t end_index = (end - 1) >> PAGE_SECTORS_SHIFT; + struct folio_batch fbatch; + unsigned i, j; - bch2_io_opts_apply(&opts, bch2_inode_opts_get(&inode->ei_inode)); - return opts; -} + if (end <= start) + return; -/* page state: */ + folio_batch_init(&fbatch); -/* stored in page->private: */ + while (filemap_get_folios(inode->v.i_mapping, + &index, end_index, &fbatch)) { + for (i = 0; i < folio_batch_count(&fbatch); i++) { + struct folio *folio = fbatch.folios[i]; + u64 folio_start = folio_sector(folio); + u64 folio_end = folio_end_sector(folio); + unsigned folio_offset = max(start, folio_start) - folio_start; + unsigned folio_len = min(end, folio_end) - folio_offset - folio_start; + struct bch_folio *s; -/* - * bch_page_state has to (unfortunately) be manipulated with cmpxchg - we could - * almost protected it with the page lock, except that bch2_writepage_io_done has - * to update the sector counts (and from interrupt/bottom half context). - */ -struct bch_page_state { -union { struct { - /* existing data: */ - unsigned sectors:PAGE_SECTOR_SHIFT + 1; + BUG_ON(end <= folio_start); - /* Uncompressed, fully allocated replicas: */ - unsigned nr_replicas:4; + folio_lock(folio); + s = bch2_folio(folio); - /* Owns PAGE_SECTORS * replicas_reserved sized reservation: */ - unsigned replicas_reserved:4; + if (s) { + spin_lock(&s->lock); + for (j = folio_offset; j < folio_offset + folio_len; j++) + s->s[j].nr_replicas = 0; + spin_unlock(&s->lock); + } + + folio_unlock(folio); + } + folio_batch_release(&fbatch); + cond_resched(); + } +} - /* Owns PAGE_SECTORS sized quota reservation: */ - unsigned quota_reserved:1; +static void mark_pagecache_reserved(struct bch_inode_info *inode, + u64 start, u64 end) +{ + struct bch_fs *c = inode->v.i_sb->s_fs_info; + pgoff_t index = start >> PAGE_SECTORS_SHIFT; + pgoff_t end_index = (end - 1) >> PAGE_SECTORS_SHIFT; + struct folio_batch fbatch; + s64 i_sectors_delta = 0; + unsigned i, j; - /* - * Number of sectors on disk - for i_blocks - * Uncompressed size, not compressed size: - */ - unsigned dirty_sectors:PAGE_SECTOR_SHIFT + 1; -}; - /* for cmpxchg: */ - unsigned long v; -}; -}; + if (end <= start) + return; -#define page_state_cmpxchg(_ptr, _new, _expr) \ -({ \ - unsigned long _v = READ_ONCE((_ptr)->v); \ - struct bch_page_state _old; \ - \ - do { \ - _old.v = _new.v = _v; \ - _expr; \ - \ - EBUG_ON(_new.sectors + _new.dirty_sectors > PAGE_SECTORS);\ - } while (_old.v != _new.v && \ - (_v = cmpxchg(&(_ptr)->v, _old.v, _new.v)) != _old.v); \ - \ - _old; \ -}) + folio_batch_init(&fbatch); -static inline struct bch_page_state *page_state(struct page *page) -{ - struct bch_page_state *s = (void *) &page->private; + while (filemap_get_folios(inode->v.i_mapping, + &index, end_index, &fbatch)) { + for (i = 0; i < folio_batch_count(&fbatch); i++) { + struct folio *folio = fbatch.folios[i]; + u64 folio_start = folio_sector(folio); + u64 folio_end = folio_end_sector(folio); + unsigned folio_offset = max(start, folio_start) - folio_start; + unsigned folio_len = min(end, folio_end) - folio_offset - folio_start; + struct bch_folio *s; - BUILD_BUG_ON(sizeof(*s) > sizeof(page->private)); + BUG_ON(end <= folio_start); - if (!PagePrivate(page)) - SetPagePrivate(page); + folio_lock(folio); + s = bch2_folio(folio); - return s; -} + if (s) { + spin_lock(&s->lock); + for (j = folio_offset; j < folio_offset + folio_len; j++) { + i_sectors_delta -= s->s[j].state == SECTOR_dirty; + folio_sector_set(folio, s, j, folio_sector_reserve(s->s[j].state)); + } + spin_unlock(&s->lock); + } -static inline unsigned page_res_sectors(struct bch_page_state s) -{ + folio_unlock(folio); + } + folio_batch_release(&fbatch); + cond_resched(); + } - return s.replicas_reserved * PAGE_SECTORS; + i_sectors_acct(c, inode, NULL, i_sectors_delta); } -static void __bch2_put_page_reservation(struct bch_fs *c, struct bch_inode_info *inode, - struct bch_page_state s) +static inline unsigned inode_nr_replicas(struct bch_fs *c, struct bch_inode_info *inode) { - struct disk_reservation res = { .sectors = page_res_sectors(s) }; - struct quota_res quota_res = { .sectors = s.quota_reserved ? PAGE_SECTORS : 0 }; - - bch2_quota_reservation_put(c, inode, "a_res); - bch2_disk_reservation_put(c, &res); + /* XXX: this should not be open coded */ + return inode->ei_inode.bi_data_replicas + ? inode->ei_inode.bi_data_replicas - 1 + : c->opts.data_replicas; } -static void bch2_put_page_reservation(struct bch_fs *c, struct bch_inode_info *inode, - struct page *page) +static inline unsigned sectors_to_reserve(struct bch_folio_sector *s, + unsigned nr_replicas) { - struct bch_page_state s; - - EBUG_ON(!PageLocked(page)); - - s = page_state_cmpxchg(page_state(page), s, { - s.replicas_reserved = 0; - s.quota_reserved = 0; - }); - - __bch2_put_page_reservation(c, inode, s); + return max(0, (int) nr_replicas - + s->nr_replicas - + s->replicas_reserved); } -static int bch2_get_page_reservation(struct bch_fs *c, struct bch_inode_info *inode, - struct page *page, bool check_enospc) +static int bch2_get_folio_disk_reservation(struct bch_fs *c, + struct bch_inode_info *inode, + struct folio *folio, bool check_enospc) { - struct bch_page_state *s = page_state(page), new; - - /* XXX: this should not be open coded */ - unsigned nr_replicas = inode->ei_inode.bi_data_replicas - ? inode->ei_inode.bi_data_replicas - 1 - : c->opts.data_replicas; - struct disk_reservation disk_res; - struct quota_res quota_res = { 0 }; + struct bch_folio *s = bch2_folio_create(folio, 0); + unsigned nr_replicas = inode_nr_replicas(c, inode); + struct disk_reservation disk_res = { 0 }; + unsigned i, sectors = folio_sectors(folio), disk_res_sectors = 0; int ret; - EBUG_ON(!PageLocked(page)); + if (!s) + return -ENOMEM; - if (s->replicas_reserved < nr_replicas) { - ret = bch2_disk_reservation_get(c, &disk_res, PAGE_SECTORS, - nr_replicas - s->replicas_reserved, - !check_enospc ? BCH_DISK_RESERVATION_NOFAIL : 0); - if (unlikely(ret)) - return ret; + for (i = 0; i < sectors; i++) + disk_res_sectors += sectors_to_reserve(&s->s[i], nr_replicas); - page_state_cmpxchg(s, new, ({ - BUG_ON(new.replicas_reserved + - disk_res.nr_replicas != nr_replicas); - new.replicas_reserved += disk_res.nr_replicas; - })); - } + if (!disk_res_sectors) + return 0; - if (!s->quota_reserved && - s->sectors + s->dirty_sectors < PAGE_SECTORS) { - ret = bch2_quota_reservation_add(c, inode, "a_res, - PAGE_SECTORS, - check_enospc); - if (unlikely(ret)) - return ret; + ret = bch2_disk_reservation_get(c, &disk_res, + disk_res_sectors, 1, + !check_enospc + ? BCH_DISK_RESERVATION_NOFAIL + : 0); + if (unlikely(ret)) + return ret; - page_state_cmpxchg(s, new, ({ - BUG_ON(new.quota_reserved); - new.quota_reserved = 1; - })); - } + for (i = 0; i < sectors; i++) + s->s[i].replicas_reserved += + sectors_to_reserve(&s->s[i], nr_replicas); - return ret; + return 0; } -static void bch2_clear_page_bits(struct page *page) +struct bch2_folio_reservation { + struct disk_reservation disk; + struct quota_res quota; +}; + +static void bch2_folio_reservation_init(struct bch_fs *c, + struct bch_inode_info *inode, + struct bch2_folio_reservation *res) { - struct bch_inode_info *inode = to_bch_ei(page->mapping->host); - struct bch_fs *c = inode->v.i_sb->s_fs_info; - struct bch_page_state s; + memset(res, 0, sizeof(*res)); - EBUG_ON(!PageLocked(page)); + res->disk.nr_replicas = inode_nr_replicas(c, inode); +} - if (!PagePrivate(page)) - return; +static void bch2_folio_reservation_put(struct bch_fs *c, + struct bch_inode_info *inode, + struct bch2_folio_reservation *res) +{ + bch2_disk_reservation_put(c, &res->disk); + bch2_quota_reservation_put(c, inode, &res->quota); +} + +static int bch2_folio_reservation_get(struct bch_fs *c, + struct bch_inode_info *inode, + struct folio *folio, + struct bch2_folio_reservation *res, + unsigned offset, unsigned len) +{ + struct bch_folio *s = bch2_folio_create(folio, 0); + unsigned i, disk_sectors = 0, quota_sectors = 0; + int ret; + + if (!s) + return -ENOMEM; + + BUG_ON(!s->uptodate); + + for (i = round_down(offset, block_bytes(c)) >> 9; + i < round_up(offset + len, block_bytes(c)) >> 9; + i++) { + disk_sectors += sectors_to_reserve(&s->s[i], + res->disk.nr_replicas); + quota_sectors += s->s[i].state == SECTOR_unallocated; + } + + if (disk_sectors) { + ret = bch2_disk_reservation_add(c, &res->disk, disk_sectors, 0); + if (unlikely(ret)) + return ret; + } - s.v = xchg(&page_state(page)->v, 0); - ClearPagePrivate(page); + if (quota_sectors) { + ret = bch2_quota_reservation_add(c, inode, &res->quota, + quota_sectors, true); + if (unlikely(ret)) { + struct disk_reservation tmp = { + .sectors = disk_sectors + }; - if (s.dirty_sectors) - i_sectors_acct(c, inode, NULL, -s.dirty_sectors); + bch2_disk_reservation_put(c, &tmp); + res->disk.sectors -= disk_sectors; + return ret; + } + } - __bch2_put_page_reservation(c, inode, s); + return 0; } -int bch2_set_page_dirty(struct page *page) +static void bch2_clear_folio_bits(struct folio *folio) { - struct bch_inode_info *inode = to_bch_ei(page->mapping->host); + struct bch_inode_info *inode = to_bch_ei(folio->mapping->host); struct bch_fs *c = inode->v.i_sb->s_fs_info; - struct quota_res quota_res = { 0 }; - struct bch_page_state old, new; + struct bch_folio *s = bch2_folio(folio); + struct disk_reservation disk_res = { 0 }; + int i, sectors = folio_sectors(folio), dirty_sectors = 0; + + if (!s) + return; + + EBUG_ON(!folio_test_locked(folio)); + EBUG_ON(folio_test_writeback(folio)); + + for (i = 0; i < sectors; i++) { + disk_res.sectors += s->s[i].replicas_reserved; + s->s[i].replicas_reserved = 0; + + dirty_sectors -= s->s[i].state == SECTOR_dirty; + folio_sector_set(folio, s, i, folio_sector_undirty(s->s[i].state)); + } - old = page_state_cmpxchg(page_state(page), new, - new.dirty_sectors = PAGE_SECTORS - new.sectors; - new.quota_reserved = 0; - ); + bch2_disk_reservation_put(c, &disk_res); - quota_res.sectors += old.quota_reserved * PAGE_SECTORS; + i_sectors_acct(c, inode, NULL, dirty_sectors); - if (old.dirty_sectors != new.dirty_sectors) - i_sectors_acct(c, inode, "a_res, - new.dirty_sectors - old.dirty_sectors); - bch2_quota_reservation_put(c, inode, "a_res); + bch2_folio_release(folio); +} + +static void bch2_set_folio_dirty(struct bch_fs *c, + struct bch_inode_info *inode, + struct folio *folio, + struct bch2_folio_reservation *res, + unsigned offset, unsigned len) +{ + struct bch_folio *s = bch2_folio(folio); + unsigned i, dirty_sectors = 0; + + WARN_ON((u64) folio_pos(folio) + offset + len > + round_up((u64) i_size_read(&inode->v), block_bytes(c))); + + BUG_ON(!s->uptodate); + + spin_lock(&s->lock); + + for (i = round_down(offset, block_bytes(c)) >> 9; + i < round_up(offset + len, block_bytes(c)) >> 9; + i++) { + unsigned sectors = sectors_to_reserve(&s->s[i], + res->disk.nr_replicas); + + /* + * This can happen if we race with the error path in + * bch2_writepage_io_done(): + */ + sectors = min_t(unsigned, sectors, res->disk.sectors); + + s->s[i].replicas_reserved += sectors; + res->disk.sectors -= sectors; + + dirty_sectors += s->s[i].state == SECTOR_unallocated; + + folio_sector_set(folio, s, i, folio_sector_dirty(s->s[i].state)); + } + + spin_unlock(&s->lock); + + i_sectors_acct(c, inode, &res->quota, dirty_sectors); - return __set_page_dirty_nobuffers(page); + if (!folio_test_dirty(folio)) + filemap_dirty_folio(inode->v.i_mapping, folio); } -int bch2_page_mkwrite(struct vm_fault *vmf) +vm_fault_t bch2_page_fault(struct vm_fault *vmf) { - struct page *page = vmf->page; struct file *file = vmf->vma->vm_file; + struct address_space *mapping = file->f_mapping; + struct address_space *fdm = faults_disabled_mapping(); struct bch_inode_info *inode = file_bch_inode(file); - struct address_space *mapping = inode->v.i_mapping; + int ret; + + if (fdm == mapping) + return VM_FAULT_SIGBUS; + + /* Lock ordering: */ + if (fdm > mapping) { + struct bch_inode_info *fdm_host = to_bch_ei(fdm->host); + + if (bch2_pagecache_add_tryget(inode)) + goto got_lock; + + bch2_pagecache_block_put(fdm_host); + + bch2_pagecache_add_get(inode); + bch2_pagecache_add_put(inode); + + bch2_pagecache_block_get(fdm_host); + + /* Signal that lock has been dropped: */ + set_fdm_dropped_locks(); + return VM_FAULT_SIGBUS; + } + + bch2_pagecache_add_get(inode); +got_lock: + ret = filemap_fault(vmf); + bch2_pagecache_add_put(inode); + + return ret; +} + +vm_fault_t bch2_page_mkwrite(struct vm_fault *vmf) +{ + struct folio *folio = page_folio(vmf->page); + struct file *file = vmf->vma->vm_file; + struct bch_inode_info *inode = file_bch_inode(file); + struct address_space *mapping = file->f_mapping; struct bch_fs *c = inode->v.i_sb->s_fs_info; - int ret = VM_FAULT_LOCKED; + struct bch2_folio_reservation res; + unsigned len; + loff_t isize; + int ret; + + bch2_folio_reservation_init(c, inode, &res); sb_start_pagefault(inode->v.i_sb); file_update_time(file); @@ -696,456 +1009,382 @@ int bch2_page_mkwrite(struct vm_fault *vmf) * a write_invalidate_inode_pages_range() that works without dropping * page lock before invalidating page */ - if (current->pagecache_lock != &mapping->add_lock) - pagecache_add_get(&mapping->add_lock); + bch2_pagecache_add_get(inode); + + folio_lock(folio); + isize = i_size_read(&inode->v); - lock_page(page); - if (page->mapping != mapping || - page_offset(page) > i_size_read(&inode->v)) { - unlock_page(page); + if (folio->mapping != mapping || folio_pos(folio) >= isize) { + folio_unlock(folio); ret = VM_FAULT_NOPAGE; goto out; } - if (bch2_get_page_reservation(c, inode, page, true)) { - unlock_page(page); + len = min_t(loff_t, folio_size(folio), isize - folio_pos(folio)); + + if (bch2_folio_set(c, inode_inum(inode), &folio, 1) ?: + bch2_folio_reservation_get(c, inode, folio, &res, 0, len)) { + folio_unlock(folio); ret = VM_FAULT_SIGBUS; goto out; } - if (!PageDirty(page)) - set_page_dirty(page); - wait_for_stable_page(page); + bch2_set_folio_dirty(c, inode, folio, &res, 0, len); + bch2_folio_reservation_put(c, inode, &res); + + folio_wait_stable(folio); + ret = VM_FAULT_LOCKED; out: - if (current->pagecache_lock != &mapping->add_lock) - pagecache_add_put(&mapping->add_lock); + bch2_pagecache_add_put(inode); sb_end_pagefault(inode->v.i_sb); + return ret; } -void bch2_invalidatepage(struct page *page, unsigned int offset, - unsigned int length) +void bch2_invalidate_folio(struct folio *folio, size_t offset, size_t length) { - EBUG_ON(!PageLocked(page)); - EBUG_ON(PageWriteback(page)); - - if (offset || length < PAGE_SIZE) + if (offset || length < folio_size(folio)) return; - bch2_clear_page_bits(page); -} - -int bch2_releasepage(struct page *page, gfp_t gfp_mask) -{ - /* XXX: this can't take locks that are held while we allocate memory */ - EBUG_ON(!PageLocked(page)); - EBUG_ON(PageWriteback(page)); - - if (PageDirty(page)) - return 0; - - bch2_clear_page_bits(page); - return 1; -} - -#ifdef CONFIG_MIGRATION -int bch2_migrate_page(struct address_space *mapping, struct page *newpage, - struct page *page, enum migrate_mode mode) -{ - int ret; - - EBUG_ON(!PageLocked(page)); - EBUG_ON(!PageLocked(newpage)); - - ret = migrate_page_move_mapping(mapping, newpage, page, mode, 0); - if (ret != MIGRATEPAGE_SUCCESS) - return ret; - - if (PagePrivate(page)) { - *page_state(newpage) = *page_state(page); - ClearPagePrivate(page); - } - - migrate_page_copy(newpage, page); - return MIGRATEPAGE_SUCCESS; -} -#endif - -/* readpages/writepages: */ - -static bool bio_can_add_page_contig(struct bio *bio, struct page *page) -{ - sector_t offset = (sector_t) page->index << PAGE_SECTOR_SHIFT; - - return bio->bi_vcnt < bio->bi_max_vecs && - bio_end_sector(bio) == offset; + bch2_clear_folio_bits(folio); } -static int bio_add_page_contig(struct bio *bio, struct page *page) +bool bch2_release_folio(struct folio *folio, gfp_t gfp_mask) { - sector_t offset = (sector_t) page->index << PAGE_SECTOR_SHIFT; - - EBUG_ON(!bio->bi_max_vecs); - - if (!bio->bi_vcnt) - bio->bi_iter.bi_sector = offset; - else if (!bio_can_add_page_contig(bio, page)) - return -1; + if (folio_test_dirty(folio) || folio_test_writeback(folio)) + return false; - __bio_add_page(bio, page, PAGE_SIZE, 0); - return 0; + bch2_clear_folio_bits(folio); + return true; } /* readpage(s): */ static void bch2_readpages_end_io(struct bio *bio) { - struct bio_vec *bv; - int i; - - bio_for_each_segment_all(bv, bio, i) { - struct page *page = bv->bv_page; + struct bvec_iter_all iter; + struct folio_vec fv; + bio_for_each_folio_all(fv, bio, iter) { if (!bio->bi_status) { - SetPageUptodate(page); + folio_mark_uptodate(fv.fv_folio); } else { - ClearPageUptodate(page); - SetPageError(page); + folio_clear_uptodate(fv.fv_folio); + folio_set_error(fv.fv_folio); } - unlock_page(page); + folio_unlock(fv.fv_folio); } bio_put(bio); } -static inline void page_state_init_for_read(struct page *page) -{ - SetPagePrivate(page); - page->private = 0; -} - struct readpages_iter { struct address_space *mapping; - struct page **pages; - unsigned nr_pages; - unsigned nr_added; unsigned idx; - pgoff_t offset; + folios folios; }; static int readpages_iter_init(struct readpages_iter *iter, - struct address_space *mapping, - struct list_head *pages, unsigned nr_pages) + struct readahead_control *ractl) { - memset(iter, 0, sizeof(*iter)); + struct folio **fi; + int ret; - iter->mapping = mapping; - iter->offset = list_last_entry(pages, struct page, lru)->index; + memset(iter, 0, sizeof(*iter)); - iter->pages = kmalloc_array(nr_pages, sizeof(struct page *), GFP_NOFS); - if (!iter->pages) - return -ENOMEM; + iter->mapping = ractl->mapping; - while (!list_empty(pages)) { - struct page *page = list_last_entry(pages, struct page, lru); + ret = filemap_get_contig_folios_d(iter->mapping, + ractl->_index << PAGE_SHIFT, + (ractl->_index + ractl->_nr_pages) << PAGE_SHIFT, + 0, mapping_gfp_mask(iter->mapping), + &iter->folios); + if (ret) + return ret; - prefetchw(&page->flags); - iter->pages[iter->nr_pages++] = page; - list_del(&page->lru); + darray_for_each(iter->folios, fi) { + ractl->_nr_pages -= 1U << folio_order(*fi); + __bch2_folio_create(*fi, __GFP_NOFAIL|GFP_KERNEL); + folio_put(*fi); + folio_put(*fi); } return 0; } -static inline struct page *readpage_iter_next(struct readpages_iter *iter) +static inline struct folio *readpage_iter_peek(struct readpages_iter *iter) { - struct page *page; - unsigned i; - int ret; - - BUG_ON(iter->idx > iter->nr_added); - BUG_ON(iter->nr_added > iter->nr_pages); - - if (iter->idx < iter->nr_added) - goto out; - - while (1) { - if (iter->idx == iter->nr_pages) - return NULL; - - ret = add_to_page_cache_lru_vec(iter->mapping, - iter->pages + iter->nr_added, - iter->nr_pages - iter->nr_added, - iter->offset + iter->nr_added, - GFP_NOFS); - if (ret > 0) - break; - - page = iter->pages[iter->nr_added]; - iter->idx++; - iter->nr_added++; - - put_page(page); - } - - iter->nr_added += ret; - - for (i = iter->idx; i < iter->nr_added; i++) - put_page(iter->pages[i]); -out: - EBUG_ON(iter->pages[iter->idx]->index != iter->offset + iter->idx); - - page_state_init_for_read(iter->pages[iter->idx]); - return iter->pages[iter->idx]; + if (iter->idx >= iter->folios.nr) + return NULL; + return iter->folios.data[iter->idx]; } -static void bch2_add_page_sectors(struct bio *bio, struct bkey_s_c k) +static inline void readpage_iter_advance(struct readpages_iter *iter) { - struct bvec_iter iter; - struct bio_vec bv; - unsigned nr_ptrs = bch2_bkey_nr_ptrs_allocated(k); - - bio_for_each_segment(bv, bio, iter) { - /* brand new pages, don't need to be locked: */ - - struct bch_page_state *s = page_state(bv.bv_page); - - /* sectors in @k from the start of this page: */ - unsigned k_sectors = k.k->size - (iter.bi_sector - k.k->p.offset); - - unsigned page_sectors = min(bv.bv_len >> 9, k_sectors); + iter->idx++; +} - s->nr_replicas = page_sectors == PAGE_SECTORS - ? nr_ptrs : 0; +static bool extent_partial_reads_expensive(struct bkey_s_c k) +{ + struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k); + struct bch_extent_crc_unpacked crc; + const union bch_extent_entry *i; - BUG_ON(s->sectors + page_sectors > PAGE_SECTORS); - s->sectors += page_sectors; - } + bkey_for_each_crc(k.k, ptrs, crc, i) + if (crc.csum_type || crc.compression_type) + return true; + return false; } -static void readpage_bio_extend(struct readpages_iter *iter, - struct bio *bio, u64 offset, - bool get_more) +static int readpage_bio_extend(struct btree_trans *trans, + struct readpages_iter *iter, + struct bio *bio, + unsigned sectors_this_extent, + bool get_more) { - while (bio_end_sector(bio) < offset && + /* Don't hold btree locks while allocating memory: */ + bch2_trans_unlock(trans); + + while (bio_sectors(bio) < sectors_this_extent && bio->bi_vcnt < bio->bi_max_vecs) { - pgoff_t page_offset = bio_end_sector(bio) >> PAGE_SECTOR_SHIFT; - struct page *page = readpage_iter_next(iter); + struct folio *folio = readpage_iter_peek(iter); int ret; - if (page) { - if (iter->offset + iter->idx != page_offset) - break; - - iter->idx++; + if (folio) { + readpage_iter_advance(iter); } else { + pgoff_t folio_offset = bio_end_sector(bio) >> PAGE_SECTORS_SHIFT; + if (!get_more) break; - page = xa_load(&iter->mapping->i_pages, page_offset); - if (page && !xa_is_value(page)) + folio = xa_load(&iter->mapping->i_pages, folio_offset); + if (folio && !xa_is_value(folio)) break; - page = __page_cache_alloc(readahead_gfp_mask(iter->mapping)); - if (!page) + folio = filemap_alloc_folio(readahead_gfp_mask(iter->mapping), 0); + if (!folio) break; - page_state_init_for_read(page); + if (!__bch2_folio_create(folio, GFP_KERNEL)) { + folio_put(folio); + break; + } - ret = add_to_page_cache_lru(page, iter->mapping, - page_offset, GFP_NOFS); + ret = filemap_add_folio(iter->mapping, folio, folio_offset, GFP_KERNEL); if (ret) { - ClearPagePrivate(page); - put_page(page); + __bch2_folio_release(folio); + folio_put(folio); break; } - put_page(page); + folio_put(folio); } - __bio_add_page(bio, page, PAGE_SIZE, 0); + BUG_ON(folio_sector(folio) != bio_end_sector(bio)); + + BUG_ON(!bio_add_folio(bio, folio, folio_size(folio), 0)); } + + return bch2_trans_relock(trans); } -static void bchfs_read(struct btree_trans *trans, struct btree_iter *iter, - struct bch_read_bio *rbio, u64 inum, +static void bchfs_read(struct btree_trans *trans, + struct bch_read_bio *rbio, + subvol_inum inum, struct readpages_iter *readpages_iter) { struct bch_fs *c = trans->c; - struct bio *bio = &rbio->bio; + struct btree_iter iter; + struct bkey_buf sk; int flags = BCH_READ_RETRY_IF_STALE| BCH_READ_MAY_PROMOTE; + u32 snapshot; + int ret = 0; rbio->c = c; rbio->start_time = local_clock(); + rbio->subvol = inum.subvol; + bch2_bkey_buf_init(&sk); +retry: + bch2_trans_begin(trans); + iter = (struct btree_iter) { NULL }; + + ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot); + if (ret) + goto err; + + bch2_trans_iter_init(trans, &iter, BTREE_ID_extents, + SPOS(inum.inum, rbio->bio.bi_iter.bi_sector, snapshot), + BTREE_ITER_SLOTS); while (1) { - BKEY_PADDED(k) tmp; struct bkey_s_c k; - unsigned bytes; + unsigned bytes, sectors, offset_into_extent; + enum btree_id data_btree = BTREE_ID_extents; - bch2_btree_iter_set_pos(iter, POS(inum, bio->bi_iter.bi_sector)); + /* + * read_extent -> io_time_reset may cause a transaction restart + * without returning an error, we need to check for that here: + */ + ret = bch2_trans_relock(trans); + if (ret) + break; - k = bch2_btree_iter_peek_slot(iter); - BUG_ON(!k.k); + bch2_btree_iter_set_pos(&iter, + POS(inum.inum, rbio->bio.bi_iter.bi_sector)); - if (IS_ERR(k.k)) { - int ret = btree_iter_err(iter); - BUG_ON(!ret); - bcache_io_error(c, bio, "btree IO error %i", ret); - bio_endio(bio); - return; - } + k = bch2_btree_iter_peek_slot(&iter); + ret = bkey_err(k); + if (ret) + break; - bkey_reassemble(&tmp.k, k); - bch2_btree_trans_unlock(trans); - k = bkey_i_to_s_c(&tmp.k); + offset_into_extent = iter.pos.offset - + bkey_start_offset(k.k); + sectors = k.k->size - offset_into_extent; - if (readpages_iter) { - bool want_full_extent = false; + bch2_bkey_buf_reassemble(&sk, c, k); - if (bkey_extent_is_data(k.k)) { - struct bkey_s_c_extent e = bkey_s_c_to_extent(k); - const union bch_extent_entry *i; - struct extent_ptr_decoded p; + ret = bch2_read_indirect_extent(trans, &data_btree, + &offset_into_extent, &sk); + if (ret) + break; - extent_for_each_ptr_decode(e, p, i) - want_full_extent |= ((p.crc.csum_type != 0) | - (p.crc.compression_type != 0)); - } + k = bkey_i_to_s_c(sk.k); - readpage_bio_extend(readpages_iter, - bio, k.k->p.offset, - want_full_extent); + sectors = min(sectors, k.k->size - offset_into_extent); + + if (readpages_iter) { + ret = readpage_bio_extend(trans, readpages_iter, &rbio->bio, sectors, + extent_partial_reads_expensive(k)); + if (ret) + break; } - bytes = (min_t(u64, k.k->p.offset, bio_end_sector(bio)) - - bio->bi_iter.bi_sector) << 9; - swap(bio->bi_iter.bi_size, bytes); + bytes = min(sectors, bio_sectors(&rbio->bio)) << 9; + swap(rbio->bio.bi_iter.bi_size, bytes); - if (bytes == bio->bi_iter.bi_size) + if (rbio->bio.bi_iter.bi_size == bytes) flags |= BCH_READ_LAST_FRAGMENT; - if (bkey_extent_is_allocation(k.k)) - bch2_add_page_sectors(bio, k); + bch2_bio_page_state_set(&rbio->bio, k); - bch2_read_extent(c, rbio, k, flags); + bch2_read_extent(trans, rbio, iter.pos, + data_btree, k, offset_into_extent, flags); if (flags & BCH_READ_LAST_FRAGMENT) - return; + break; + + swap(rbio->bio.bi_iter.bi_size, bytes); + bio_advance(&rbio->bio, bytes); + + ret = btree_trans_too_many_iters(trans); + if (ret) + break; + } +err: + bch2_trans_iter_exit(trans, &iter); + + if (bch2_err_matches(ret, BCH_ERR_transaction_restart)) + goto retry; - swap(bio->bi_iter.bi_size, bytes); - bio_advance(bio, bytes); + if (ret) { + bch_err_inum_offset_ratelimited(c, + iter.pos.inode, + iter.pos.offset << 9, + "read error %i from btree lookup", ret); + rbio->bio.bi_status = BLK_STS_IOERR; + bio_endio(&rbio->bio); } + + bch2_bkey_buf_exit(&sk, c); } -int bch2_readpages(struct file *file, struct address_space *mapping, - struct list_head *pages, unsigned nr_pages) +void bch2_readahead(struct readahead_control *ractl) { - struct bch_inode_info *inode = to_bch_ei(mapping->host); + struct bch_inode_info *inode = to_bch_ei(ractl->mapping->host); struct bch_fs *c = inode->v.i_sb->s_fs_info; - struct bch_io_opts opts = io_opts(c, inode); + struct bch_io_opts opts; struct btree_trans trans; - struct btree_iter *iter; - struct page *page; + struct folio *folio; struct readpages_iter readpages_iter; int ret; - ret = readpages_iter_init(&readpages_iter, mapping, pages, nr_pages); - BUG_ON(ret); + bch2_inode_opts_get(&opts, c, &inode->ei_inode); - bch2_trans_init(&trans, c); + ret = readpages_iter_init(&readpages_iter, ractl); + BUG_ON(ret); - iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS, POS_MIN, - BTREE_ITER_SLOTS); + bch2_trans_init(&trans, c, 0, 0); - if (current->pagecache_lock != &mapping->add_lock) - pagecache_add_get(&mapping->add_lock); + bch2_pagecache_add_get(inode); - while ((page = readpage_iter_next(&readpages_iter))) { - pgoff_t index = readpages_iter.offset + readpages_iter.idx; + while ((folio = readpage_iter_peek(&readpages_iter))) { unsigned n = min_t(unsigned, - readpages_iter.nr_pages - + readpages_iter.folios.nr - readpages_iter.idx, - BIO_MAX_PAGES); + BIO_MAX_VECS); struct bch_read_bio *rbio = - rbio_init(bio_alloc_bioset(GFP_NOFS, n, &c->bio_read), + rbio_init(bio_alloc_bioset(NULL, n, REQ_OP_READ, + GFP_KERNEL, &c->bio_read), opts); - readpages_iter.idx++; + readpage_iter_advance(&readpages_iter); - bio_set_op_attrs(&rbio->bio, REQ_OP_READ, 0); - rbio->bio.bi_iter.bi_sector = (sector_t) index << PAGE_SECTOR_SHIFT; + rbio->bio.bi_iter.bi_sector = folio_sector(folio); rbio->bio.bi_end_io = bch2_readpages_end_io; - __bio_add_page(&rbio->bio, page, PAGE_SIZE, 0); + BUG_ON(!bio_add_folio(&rbio->bio, folio, folio_size(folio), 0)); - bchfs_read(&trans, iter, rbio, inode->v.i_ino, + bchfs_read(&trans, rbio, inode_inum(inode), &readpages_iter); + bch2_trans_unlock(&trans); } - if (current->pagecache_lock != &mapping->add_lock) - pagecache_add_put(&mapping->add_lock); + bch2_pagecache_add_put(inode); bch2_trans_exit(&trans); - kfree(readpages_iter.pages); - - return 0; + darray_exit(&readpages_iter.folios); } -static void __bchfs_readpage(struct bch_fs *c, struct bch_read_bio *rbio, - u64 inum, struct page *page) +static void __bchfs_readfolio(struct bch_fs *c, struct bch_read_bio *rbio, + subvol_inum inum, struct folio *folio) { struct btree_trans trans; - struct btree_iter *iter; - - page_state_init_for_read(page); - bio_set_op_attrs(&rbio->bio, REQ_OP_READ, REQ_SYNC); - bio_add_page_contig(&rbio->bio, page); + bch2_folio_create(folio, __GFP_NOFAIL); - bch2_trans_init(&trans, c); - iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS, POS_MIN, - BTREE_ITER_SLOTS); - - bchfs_read(&trans, iter, rbio, inum, NULL); + rbio->bio.bi_opf = REQ_OP_READ|REQ_SYNC; + rbio->bio.bi_iter.bi_sector = folio_sector(folio); + BUG_ON(!bio_add_folio(&rbio->bio, folio, folio_size(folio), 0)); + bch2_trans_init(&trans, c, 0, 0); + bchfs_read(&trans, rbio, inum, NULL); bch2_trans_exit(&trans); } -int bch2_readpage(struct file *file, struct page *page) -{ - struct bch_inode_info *inode = to_bch_ei(page->mapping->host); - struct bch_fs *c = inode->v.i_sb->s_fs_info; - struct bch_io_opts opts = io_opts(c, inode); - struct bch_read_bio *rbio; - - rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read), opts); - rbio->bio.bi_end_io = bch2_readpages_end_io; - - __bchfs_readpage(c, rbio, inode->v.i_ino, page); - return 0; -} - -static void bch2_read_single_page_end_io(struct bio *bio) +static void bch2_read_single_folio_end_io(struct bio *bio) { complete(bio->bi_private); } -static int bch2_read_single_page(struct page *page, - struct address_space *mapping) +static int bch2_read_single_folio(struct folio *folio, + struct address_space *mapping) { struct bch_inode_info *inode = to_bch_ei(mapping->host); struct bch_fs *c = inode->v.i_sb->s_fs_info; struct bch_read_bio *rbio; + struct bch_io_opts opts; int ret; DECLARE_COMPLETION_ONSTACK(done); - rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read), - io_opts(c, inode)); + bch2_inode_opts_get(&opts, c, &inode->ei_inode); + + rbio = rbio_init(bio_alloc_bioset(NULL, 1, REQ_OP_READ, GFP_KERNEL, &c->bio_read), + opts); rbio->bio.bi_private = &done; - rbio->bio.bi_end_io = bch2_read_single_page_end_io; + rbio->bio.bi_end_io = bch2_read_single_folio_end_io; - __bchfs_readpage(c, rbio, inode->v.i_ino, page); + __bchfs_readfolio(c, rbio, inode_inum(inode), folio); wait_for_completion(&done); ret = blk_status_to_errno(rbio->bio.bi_status); @@ -1154,71 +1393,103 @@ static int bch2_read_single_page(struct page *page, if (ret < 0) return ret; - SetPageUptodate(page); + folio_mark_uptodate(folio); return 0; } +int bch2_read_folio(struct file *file, struct folio *folio) +{ + int ret; + + ret = bch2_read_single_folio(folio, folio->mapping); + folio_unlock(folio); + return bch2_err_class(ret); +} + /* writepages: */ struct bch_writepage_state { struct bch_writepage_io *io; struct bch_io_opts opts; + struct bch_folio_sector *tmp; + unsigned tmp_sectors; }; static inline struct bch_writepage_state bch_writepage_state_init(struct bch_fs *c, struct bch_inode_info *inode) { - return (struct bch_writepage_state) { .opts = io_opts(c, inode) }; -} - -static void bch2_writepage_io_free(struct closure *cl) -{ - struct bch_writepage_io *io = container_of(cl, - struct bch_writepage_io, cl); + struct bch_writepage_state ret = { 0 }; - bio_put(&io->op.op.wbio.bio); + bch2_inode_opts_get(&ret.opts, c, &inode->ei_inode); + return ret; } -static void bch2_writepage_io_done(struct closure *cl) +static void bch2_writepage_io_done(struct bch_write_op *op) { - struct bch_writepage_io *io = container_of(cl, - struct bch_writepage_io, cl); - struct bch_fs *c = io->op.op.c; - struct bio *bio = &io->op.op.wbio.bio; - struct bio_vec *bvec; + struct bch_writepage_io *io = + container_of(op, struct bch_writepage_io, op); + struct bch_fs *c = io->op.c; + struct bio *bio = &io->op.wbio.bio; + struct bvec_iter_all iter; + struct folio_vec fv; unsigned i; - if (io->op.op.error) { - bio_for_each_segment_all(bvec, bio, i) - SetPageError(bvec->bv_page); - set_bit(AS_EIO, &io->op.inode->v.i_mapping->flags); + if (io->op.error) { + set_bit(EI_INODE_ERROR, &io->inode->ei_flags); + + bio_for_each_folio_all(fv, bio, iter) { + struct bch_folio *s; + + folio_set_error(fv.fv_folio); + mapping_set_error(fv.fv_folio->mapping, -EIO); + + s = __bch2_folio(fv.fv_folio); + spin_lock(&s->lock); + for (i = 0; i < folio_sectors(fv.fv_folio); i++) + s->s[i].nr_replicas = 0; + spin_unlock(&s->lock); + } + } + + if (io->op.flags & BCH_WRITE_WROTE_DATA_INLINE) { + bio_for_each_folio_all(fv, bio, iter) { + struct bch_folio *s; + + s = __bch2_folio(fv.fv_folio); + spin_lock(&s->lock); + for (i = 0; i < folio_sectors(fv.fv_folio); i++) + s->s[i].nr_replicas = 0; + spin_unlock(&s->lock); + } } /* * racing with fallocate can cause us to add fewer sectors than * expected - but we shouldn't add more sectors than expected: */ - BUG_ON(io->op.sectors_added > (s64) io->new_sectors); + WARN_ON_ONCE(io->op.i_sectors_delta > 0); /* * (error (due to going RO) halfway through a page can screw that up * slightly) * XXX wtf? - BUG_ON(io->op.sectors_added - io->new_sectors >= (s64) PAGE_SECTORS); + BUG_ON(io->op.op.i_sectors_delta >= PAGE_SECTORS); */ /* * PageWriteback is effectively our ref on the inode - fixup i_blocks * before calling end_page_writeback: */ - if (io->op.sectors_added != io->new_sectors) - i_sectors_acct(c, io->op.inode, NULL, - io->op.sectors_added - (s64) io->new_sectors); + i_sectors_acct(c, io->inode, NULL, io->op.i_sectors_delta); + + bio_for_each_folio_all(fv, bio, iter) { + struct bch_folio *s = __bch2_folio(fv.fv_folio); - bio_for_each_segment_all(bvec, bio, i) - end_page_writeback(bvec->bv_page); + if (atomic_dec_and_test(&s->write_count)) + folio_end_writeback(fv.fv_folio); + } - closure_return_with_destructor(&io->cl, bch2_writepage_io_free); + bio_put(&io->op.wbio.bio); } static void bch2_writepage_do_io(struct bch_writepage_state *w) @@ -1226,8 +1497,7 @@ static void bch2_writepage_do_io(struct bch_writepage_state *w) struct bch_writepage_io *io = w->io; w->io = NULL; - closure_call(&io->op.op.cl, bch2_write, NULL, &io->cl); - continue_at(&io->cl, bch2_writepage_io_done, NULL); + closure_call(&io->op.cl, bch2_write, NULL, NULL); } /* @@ -1235,115 +1505,178 @@ static void bch2_writepage_do_io(struct bch_writepage_state *w) * possible, else allocating a new one: */ static void bch2_writepage_io_alloc(struct bch_fs *c, + struct writeback_control *wbc, struct bch_writepage_state *w, struct bch_inode_info *inode, - struct page *page, + u64 sector, unsigned nr_replicas) { struct bch_write_op *op; - u64 offset = (u64) page->index << PAGE_SECTOR_SHIFT; - w->io = container_of(bio_alloc_bioset(GFP_NOFS, - BIO_MAX_PAGES, + w->io = container_of(bio_alloc_bioset(NULL, BIO_MAX_VECS, + REQ_OP_WRITE, + GFP_KERNEL, &c->writepage_bioset), - struct bch_writepage_io, op.op.wbio.bio); + struct bch_writepage_io, op.wbio.bio); - closure_init(&w->io->cl, NULL); - w->io->new_sectors = 0; - bch2_fswrite_op_init(&w->io->op, c, inode, w->opts, false); - op = &w->io->op.op; + w->io->inode = inode; + op = &w->io->op; + bch2_write_op_init(op, c, w->opts); + op->target = w->opts.foreground_target; op->nr_replicas = nr_replicas; op->res.nr_replicas = nr_replicas; op->write_point = writepoint_hashed(inode->ei_last_dirtied); - op->pos = POS(inode->v.i_ino, offset); - op->wbio.bio.bi_iter.bi_sector = offset; + op->subvol = inode->ei_subvol; + op->pos = POS(inode->v.i_ino, sector); + op->end_io = bch2_writepage_io_done; + op->devs_need_flush = &inode->ei_devs_need_flush; + op->wbio.bio.bi_iter.bi_sector = sector; + op->wbio.bio.bi_opf = wbc_to_write_flags(wbc); } -static int __bch2_writepage(struct page *page, +static int __bch2_writepage(struct folio *folio, struct writeback_control *wbc, void *data) { - struct bch_inode_info *inode = to_bch_ei(page->mapping->host); + struct bch_inode_info *inode = to_bch_ei(folio->mapping->host); struct bch_fs *c = inode->v.i_sb->s_fs_info; struct bch_writepage_state *w = data; - struct bch_page_state new, old; - unsigned offset, nr_replicas_this_write; + struct bch_folio *s; + unsigned i, offset, f_sectors, nr_replicas_this_write = U32_MAX; loff_t i_size = i_size_read(&inode->v); - pgoff_t end_index = i_size >> PAGE_SHIFT; + int ret; - EBUG_ON(!PageUptodate(page)); + EBUG_ON(!folio_test_uptodate(folio)); - /* Is the page fully inside i_size? */ - if (page->index < end_index) + /* Is the folio fully inside i_size? */ + if (folio_end_pos(folio) <= i_size) goto do_io; - /* Is the page fully outside i_size? (truncate in progress) */ - offset = i_size & (PAGE_SIZE - 1); - if (page->index > end_index || !offset) { - unlock_page(page); + /* Is the folio fully outside i_size? (truncate in progress) */ + if (folio_pos(folio) >= i_size) { + folio_unlock(folio); return 0; } /* - * The page straddles i_size. It must be zeroed out on each and every + * The folio straddles i_size. It must be zeroed out on each and every * writepage invocation because it may be mmapped. "A file is mapped - * in multiples of the page size. For a file that is not a multiple of - * the page size, the remaining memory is zeroed when mapped, and + * in multiples of the folio size. For a file that is not a multiple of + * the folio size, the remaining memory is zeroed when mapped, and * writes to that region are not written out to the file." */ - zero_user_segment(page, offset, PAGE_SIZE); + folio_zero_segment(folio, + i_size - folio_pos(folio), + folio_size(folio)); do_io: - EBUG_ON(!PageLocked(page)); + f_sectors = folio_sectors(folio); + s = bch2_folio(folio); + + if (f_sectors > w->tmp_sectors) { + kfree(w->tmp); + w->tmp = kzalloc(sizeof(struct bch_folio_sector) * + f_sectors, __GFP_NOFAIL); + w->tmp_sectors = f_sectors; + } + + /* + * Things get really hairy with errors during writeback: + */ + ret = bch2_get_folio_disk_reservation(c, inode, folio, false); + BUG_ON(ret); + + /* Before unlocking the page, get copy of reservations: */ + spin_lock(&s->lock); + memcpy(w->tmp, s->s, sizeof(struct bch_folio_sector) * f_sectors); + + for (i = 0; i < f_sectors; i++) { + if (s->s[i].state < SECTOR_dirty) + continue; - /* Before unlocking the page, transfer reservation to w->io: */ - old = page_state_cmpxchg(page_state(page), new, { - /* - * If we didn't get a reservation, we can only write out the - * number of (fully allocated) replicas that currently exist, - * and only if the entire page has been written: - */ nr_replicas_this_write = - max_t(unsigned, - new.replicas_reserved, - (new.sectors == PAGE_SECTORS - ? new.nr_replicas : 0)); + min_t(unsigned, nr_replicas_this_write, + s->s[i].nr_replicas + + s->s[i].replicas_reserved); + } - BUG_ON(!nr_replicas_this_write); + for (i = 0; i < f_sectors; i++) { + if (s->s[i].state < SECTOR_dirty) + continue; - new.nr_replicas = w->opts.compression - ? 0 - : nr_replicas_this_write; + s->s[i].nr_replicas = w->opts.compression + ? 0 : nr_replicas_this_write; - new.replicas_reserved = 0; + s->s[i].replicas_reserved = 0; + folio_sector_set(folio, s, i, SECTOR_allocated); + } + spin_unlock(&s->lock); - new.sectors += new.dirty_sectors; - BUG_ON(new.sectors != PAGE_SECTORS); - new.dirty_sectors = 0; - }); + BUG_ON(atomic_read(&s->write_count)); + atomic_set(&s->write_count, 1); - BUG_ON(PageWriteback(page)); - set_page_writeback(page); - unlock_page(page); + BUG_ON(folio_test_writeback(folio)); + folio_start_writeback(folio); - if (w->io && - (w->io->op.op.res.nr_replicas != nr_replicas_this_write || - !bio_can_add_page_contig(&w->io->op.op.wbio.bio, page))) - bch2_writepage_do_io(w); + folio_unlock(folio); - if (!w->io) - bch2_writepage_io_alloc(c, w, inode, page, - nr_replicas_this_write); + offset = 0; + while (1) { + unsigned sectors = 0, dirty_sectors = 0, reserved_sectors = 0; + u64 sector; - w->io->new_sectors += new.sectors - old.sectors; + while (offset < f_sectors && + w->tmp[offset].state < SECTOR_dirty) + offset++; - BUG_ON(inode != w->io->op.inode); - BUG_ON(bio_add_page_contig(&w->io->op.op.wbio.bio, page)); + if (offset == f_sectors) + break; - w->io->op.op.res.sectors += old.replicas_reserved * PAGE_SECTORS; - w->io->op.new_i_size = i_size; + while (offset + sectors < f_sectors && + w->tmp[offset + sectors].state >= SECTOR_dirty) { + reserved_sectors += w->tmp[offset + sectors].replicas_reserved; + dirty_sectors += w->tmp[offset + sectors].state == SECTOR_dirty; + sectors++; + } + BUG_ON(!sectors); + + sector = folio_sector(folio) + offset; + + if (w->io && + (w->io->op.res.nr_replicas != nr_replicas_this_write || + bio_full(&w->io->op.wbio.bio, sectors << 9) || + w->io->op.wbio.bio.bi_iter.bi_size + (sectors << 9) >= + (BIO_MAX_VECS * PAGE_SIZE) || + bio_end_sector(&w->io->op.wbio.bio) != sector)) + bch2_writepage_do_io(w); + + if (!w->io) + bch2_writepage_io_alloc(c, wbc, w, inode, sector, + nr_replicas_this_write); + + atomic_inc(&s->write_count); + + BUG_ON(inode != w->io->inode); + BUG_ON(!bio_add_folio(&w->io->op.wbio.bio, folio, + sectors << 9, offset << 9)); + + /* Check for writing past i_size: */ + WARN_ONCE((bio_end_sector(&w->io->op.wbio.bio) << 9) > + round_up(i_size, block_bytes(c)) && + !test_bit(BCH_FS_EMERGENCY_RO, &c->flags), + "writing past i_size: %llu > %llu (unrounded %llu)\n", + bio_end_sector(&w->io->op.wbio.bio) << 9, + round_up(i_size, block_bytes(c)), + i_size); + + w->io->op.res.sectors += reserved_sectors; + w->io->op.i_sectors_delta -= dirty_sectors; + w->io->op.new_i_size = i_size; + + offset += sectors; + } - if (wbc->sync_mode == WB_SYNC_ALL) - w->io->op.op.wbio.bio.bi_opf |= REQ_SYNC; + if (atomic_dec_and_test(&s->write_count)) + folio_end_writeback(folio); return 0; } @@ -1361,75 +1694,75 @@ int bch2_writepages(struct address_space *mapping, struct writeback_control *wbc if (w.io) bch2_writepage_do_io(&w); blk_finish_plug(&plug); - return ret; -} - -int bch2_writepage(struct page *page, struct writeback_control *wbc) -{ - struct bch_fs *c = page->mapping->host->i_sb->s_fs_info; - struct bch_writepage_state w = - bch_writepage_state_init(c, to_bch_ei(page->mapping->host)); - int ret; - - ret = __bch2_writepage(page, wbc, &w); - if (w.io) - bch2_writepage_do_io(&w); - - return ret; + kfree(w.tmp); + return bch2_err_class(ret); } /* buffered writes: */ int bch2_write_begin(struct file *file, struct address_space *mapping, - loff_t pos, unsigned len, unsigned flags, + loff_t pos, unsigned len, struct page **pagep, void **fsdata) { struct bch_inode_info *inode = to_bch_ei(mapping->host); struct bch_fs *c = inode->v.i_sb->s_fs_info; - pgoff_t index = pos >> PAGE_SHIFT; - unsigned offset = pos & (PAGE_SIZE - 1); - struct page *page; + struct bch2_folio_reservation *res; + struct folio *folio; + unsigned offset; int ret = -ENOMEM; - BUG_ON(inode_unhashed(&inode->v)); + res = kmalloc(sizeof(*res), GFP_KERNEL); + if (!res) + return -ENOMEM; + + bch2_folio_reservation_init(c, inode, res); + *fsdata = res; - /* Not strictly necessary - same reason as mkwrite(): */ - pagecache_add_get(&mapping->add_lock); + bch2_pagecache_add_get(inode); - page = grab_cache_page_write_begin(mapping, index, flags); - if (!page) + folio = __filemap_get_folio(mapping, pos >> PAGE_SHIFT, + FGP_LOCK|FGP_WRITE|FGP_CREAT|FGP_STABLE, + mapping_gfp_mask(mapping)); + if (!folio) goto err_unlock; - if (PageUptodate(page)) + if (folio_test_uptodate(folio)) goto out; - /* If we're writing entire page, don't need to read it in first: */ - if (len == PAGE_SIZE) + offset = pos - folio_pos(folio); + len = min_t(size_t, len, folio_end_pos(folio) - pos); + + /* If we're writing entire folio, don't need to read it in first: */ + if (!offset && len == folio_size(folio)) goto out; if (!offset && pos + len >= inode->v.i_size) { - zero_user_segment(page, len, PAGE_SIZE); - flush_dcache_page(page); + folio_zero_segment(folio, len, folio_size(folio)); + flush_dcache_folio(folio); goto out; } - if (index > inode->v.i_size >> PAGE_SHIFT) { - zero_user_segments(page, 0, offset, offset + len, PAGE_SIZE); - flush_dcache_page(page); + if (folio_pos(folio) >= inode->v.i_size) { + folio_zero_segments(folio, 0, offset, offset + len, folio_size(folio)); + flush_dcache_folio(folio); goto out; } readpage: - ret = bch2_read_single_page(page, mapping); + ret = bch2_read_single_folio(folio, mapping); if (ret) goto err; out: - ret = bch2_get_page_reservation(c, inode, page, true); + ret = bch2_folio_set(c, inode_inum(inode), &folio, 1); + if (ret) + goto err; + + ret = bch2_folio_reservation_get(c, inode, folio, res, offset, len); if (ret) { - if (!PageUptodate(page)) { + if (!folio_test_uptodate(folio)) { /* - * If the page hasn't been read in, we won't know if we + * If the folio hasn't been read in, we won't know if we * actually need a reservation - we don't actually need - * to read here, we just need to check if the page is + * to read here, we just need to check if the folio is * fully backed by uncompressed data: */ goto readpage; @@ -1438,15 +1771,17 @@ out: goto err; } - *pagep = page; + *pagep = &folio->page; return 0; err: - unlock_page(page); - put_page(page); + folio_unlock(folio); + folio_put(folio); *pagep = NULL; err_unlock: - pagecache_add_put(&mapping->add_lock); - return ret; + bch2_pagecache_add_put(inode); + kfree(res); + *fsdata = NULL; + return bch2_err_class(ret); } int bch2_write_end(struct file *file, struct address_space *mapping, @@ -1455,17 +1790,21 @@ int bch2_write_end(struct file *file, struct address_space *mapping, { struct bch_inode_info *inode = to_bch_ei(mapping->host); struct bch_fs *c = inode->v.i_sb->s_fs_info; + struct bch2_folio_reservation *res = fsdata; + struct folio *folio = page_folio(page); + unsigned offset = pos - folio_pos(folio); lockdep_assert_held(&inode->v.i_rwsem); + BUG_ON(offset + copied > folio_size(folio)); - if (unlikely(copied < len && !PageUptodate(page))) { + if (unlikely(copied < len && !folio_test_uptodate(folio))) { /* - * The page needs to be read in, but that would destroy + * The folio needs to be read in, but that would destroy * our partial write - simplest thing is to just force * userspace to redo the write: */ - zero_user(page, 0, PAGE_SIZE); - flush_dcache_page(page); + folio_zero_range(folio, 0, folio_size(folio)); + flush_dcache_folio(folio); copied = 0; } @@ -1475,24 +1814,33 @@ int bch2_write_end(struct file *file, struct address_space *mapping, spin_unlock(&inode->v.i_lock); if (copied) { - if (!PageUptodate(page)) - SetPageUptodate(page); - if (!PageDirty(page)) - set_page_dirty(page); + if (!folio_test_uptodate(folio)) + folio_mark_uptodate(folio); + + bch2_set_folio_dirty(c, inode, folio, res, offset, copied); inode->ei_last_dirtied = (unsigned long) current; - } else { - bch2_put_page_reservation(c, inode, page); } - unlock_page(page); - put_page(page); - pagecache_add_put(&mapping->add_lock); + folio_unlock(folio); + folio_put(folio); + bch2_pagecache_add_put(inode); + + bch2_folio_reservation_put(c, inode, res); + kfree(res); return copied; } -#define WRITE_BATCH_PAGES 32 +static noinline void folios_trunc(folios *folios, struct folio **fi) +{ + while (folios->data + folios->nr > fi) { + struct folio *f = darray_pop(folios); + + folio_unlock(f); + folio_put(f); + } +} static int __bch2_buffered_write(struct bch_inode_info *inode, struct address_space *mapping, @@ -1500,113 +1848,158 @@ static int __bch2_buffered_write(struct bch_inode_info *inode, loff_t pos, unsigned len) { struct bch_fs *c = inode->v.i_sb->s_fs_info; - struct page *pages[WRITE_BATCH_PAGES]; - unsigned long index = pos >> PAGE_SHIFT; - unsigned offset = pos & (PAGE_SIZE - 1); - unsigned nr_pages = DIV_ROUND_UP(offset + len, PAGE_SIZE); - unsigned i, copied = 0, nr_pages_copied = 0; + struct bch2_folio_reservation res; + folios folios; + struct folio **fi, *f; + unsigned copied = 0, f_offset; + u64 end = pos + len, f_pos; + loff_t last_folio_pos = inode->v.i_size; int ret = 0; BUG_ON(!len); - BUG_ON(nr_pages > ARRAY_SIZE(pages)); - for (i = 0; i < nr_pages; i++) { - pages[i] = grab_cache_page_write_begin(mapping, index + i, 0); - if (!pages[i]) { - nr_pages = i; - ret = -ENOMEM; - goto out; - } - } + bch2_folio_reservation_init(c, inode, &res); + darray_init(&folios); - if (offset && !PageUptodate(pages[0])) { - ret = bch2_read_single_page(pages[0], mapping); + ret = filemap_get_contig_folios_d(mapping, pos, end, + FGP_LOCK|FGP_WRITE|FGP_STABLE|FGP_CREAT, + mapping_gfp_mask(mapping), + &folios); + if (ret) + goto out; + + BUG_ON(!folios.nr); + + f = darray_first(folios); + if (pos != folio_pos(f) && !folio_test_uptodate(f)) { + ret = bch2_read_single_folio(f, mapping); if (ret) goto out; } - if ((pos + len) & (PAGE_SIZE - 1) && - !PageUptodate(pages[nr_pages - 1])) { - if ((index + nr_pages - 1) << PAGE_SHIFT >= inode->v.i_size) { - zero_user(pages[nr_pages - 1], 0, PAGE_SIZE); + f = darray_last(folios); + end = min(end, folio_end_pos(f)); + last_folio_pos = folio_pos(f); + if (end != folio_end_pos(f) && !folio_test_uptodate(f)) { + if (end >= inode->v.i_size) { + folio_zero_range(f, 0, folio_size(f)); } else { - ret = bch2_read_single_page(pages[nr_pages - 1], mapping); + ret = bch2_read_single_folio(f, mapping); if (ret) goto out; } } - for (i = 0; i < nr_pages; i++) { - ret = bch2_get_page_reservation(c, inode, pages[i], true); + ret = bch2_folio_set(c, inode_inum(inode), folios.data, folios.nr); + if (ret) + goto out; - if (ret && !PageUptodate(pages[i])) { - ret = bch2_read_single_page(pages[i], mapping); - if (ret) + f_pos = pos; + f_offset = pos - folio_pos(darray_first(folios)); + darray_for_each(folios, fi) { + struct folio *f = *fi; + u64 f_len = min(end, folio_end_pos(f)) - f_pos; + + /* + * XXX: per POSIX and fstests generic/275, on -ENOSPC we're + * supposed to write as much as we have disk space for. + * + * On failure here we should still write out a partial page if + * we aren't completely out of disk space - we don't do that + * yet: + */ + ret = bch2_folio_reservation_get(c, inode, f, &res, f_offset, f_len); + if (unlikely(ret)) { + folios_trunc(&folios, fi); + if (!folios.nr) goto out; - ret = bch2_get_page_reservation(c, inode, pages[i], true); + end = min(end, folio_end_pos(darray_last(folios))); + break; } - if (ret) - goto out; + f_pos = folio_end_pos(f); + f_offset = 0; } if (mapping_writably_mapped(mapping)) - for (i = 0; i < nr_pages; i++) - flush_dcache_page(pages[i]); - - while (copied < len) { - struct page *page = pages[(offset + copied) >> PAGE_SHIFT]; - unsigned pg_offset = (offset + copied) & (PAGE_SIZE - 1); - unsigned pg_bytes = min_t(unsigned, len - copied, - PAGE_SIZE - pg_offset); - unsigned pg_copied = iov_iter_copy_from_user_atomic(page, - iter, pg_offset, pg_bytes); - - if (!pg_copied) + darray_for_each(folios, fi) + flush_dcache_folio(*fi); + + f_pos = pos; + f_offset = pos - folio_pos(darray_first(folios)); + darray_for_each(folios, fi) { + struct folio *f = *fi; + u64 f_len = min(end, folio_end_pos(f)) - f_pos; + unsigned f_copied = copy_folio_from_iter_atomic(f, f_offset, f_len, iter); + + if (!f_copied) { + folios_trunc(&folios, fi); + break; + } + + if (!folio_test_uptodate(f) && + f_copied != folio_size(f) && + pos + copied + f_copied < inode->v.i_size) { + folio_zero_range(f, 0, folio_size(f)); + folios_trunc(&folios, fi); + break; + } + + flush_dcache_folio(f); + copied += f_copied; + + if (f_copied != f_len) { + folios_trunc(&folios, fi + 1); break; + } - flush_dcache_page(page); - iov_iter_advance(iter, pg_copied); - copied += pg_copied; + f_pos = folio_end_pos(f); + f_offset = 0; } if (!copied) goto out; - nr_pages_copied = DIV_ROUND_UP(offset + copied, PAGE_SIZE); - inode->ei_last_dirtied = (unsigned long) current; + end = pos + copied; spin_lock(&inode->v.i_lock); - if (pos + copied > inode->v.i_size) - i_size_write(&inode->v, pos + copied); + if (end > inode->v.i_size) + i_size_write(&inode->v, end); spin_unlock(&inode->v.i_lock); - if (copied < len && - ((offset + copied) & (PAGE_SIZE - 1))) { - struct page *page = pages[(offset + copied) >> PAGE_SHIFT]; + f_pos = pos; + f_offset = pos - folio_pos(darray_first(folios)); + darray_for_each(folios, fi) { + struct folio *f = *fi; + u64 f_len = min(end, folio_end_pos(f)) - f_pos; - if (!PageUptodate(page)) { - zero_user(page, 0, PAGE_SIZE); - copied -= (offset + copied) & (PAGE_SIZE - 1); - } + if (!folio_test_uptodate(f)) + folio_mark_uptodate(f); + + bch2_set_folio_dirty(c, inode, f, &res, f_offset, f_len); + + f_pos = folio_end_pos(f); + f_offset = 0; } + + inode->ei_last_dirtied = (unsigned long) current; out: - for (i = 0; i < nr_pages_copied; i++) { - if (!PageUptodate(pages[i])) - SetPageUptodate(pages[i]); - if (!PageDirty(pages[i])) - set_page_dirty(pages[i]); - unlock_page(pages[i]); - put_page(pages[i]); + darray_for_each(folios, fi) { + folio_unlock(*fi); + folio_put(*fi); } - for (i = nr_pages_copied; i < nr_pages; i++) { - if (!PageDirty(pages[i])) - bch2_put_page_reservation(c, inode, pages[i]); - unlock_page(pages[i]); - put_page(pages[i]); - } + /* + * If the last folio added to the mapping starts beyond current EOF, we + * performed a short write but left around at least one post-EOF folio. + * Clean up the mapping before we return. + */ + if (last_folio_pos >= inode->v.i_size) + truncate_pagecache(&inode->v, inode->v.i_size); + + darray_exit(&folios); + bch2_folio_reservation_put(c, inode, &res); return copied ?: ret; } @@ -1620,12 +2013,11 @@ static ssize_t bch2_buffered_write(struct kiocb *iocb, struct iov_iter *iter) ssize_t written = 0; int ret = 0; - pagecache_add_get(&mapping->add_lock); + bch2_pagecache_add_get(inode); do { unsigned offset = pos & (PAGE_SIZE - 1); - unsigned bytes = min_t(unsigned long, iov_iter_count(iter), - PAGE_SIZE * WRITE_BATCH_PAGES - offset); + unsigned bytes = iov_iter_count(iter); again: /* * Bring in the user page that we will copy from _first_. @@ -1637,11 +2029,11 @@ again: * to check that the address is actually valid, when atomic * usercopies are used, below. */ - if (unlikely(iov_iter_fault_in_readable(iter, bytes))) { + if (unlikely(fault_in_iov_iter_readable(iter, bytes))) { bytes = min_t(unsigned long, iov_iter_count(iter), PAGE_SIZE - offset); - if (unlikely(iov_iter_fault_in_readable(iter, bytes))) { + if (unlikely(fault_in_iov_iter_readable(iter, bytes))) { ret = -EFAULT; break; } @@ -1673,23 +2065,34 @@ again: } pos += ret; written += ret; + ret = 0; balance_dirty_pages_ratelimited(mapping); } while (iov_iter_count(iter)); - pagecache_add_put(&mapping->add_lock); + bch2_pagecache_add_put(inode); return written ? written : ret; } /* O_DIRECT reads */ +static void bio_check_or_release(struct bio *bio, bool check_dirty) +{ + if (check_dirty) { + bio_check_pages_dirty(bio); + } else { + bio_release_pages(bio, false); + bio_put(bio); + } +} + static void bch2_dio_read_complete(struct closure *cl) { struct dio_read *dio = container_of(cl, struct dio_read, cl); - dio->req->ki_complete(dio->req, dio->ret, 0); - bio_check_pages_dirty(&dio->rbio.bio); /* transfers ownership */ + dio->req->ki_complete(dio->req, dio->ret); + bio_check_or_release(&dio->rbio.bio, dio->should_dirty); } static void bch2_direct_IO_read_endio(struct bio *bio) @@ -1704,8 +2107,11 @@ static void bch2_direct_IO_read_endio(struct bio *bio) static void bch2_direct_IO_read_split_endio(struct bio *bio) { + struct dio_read *dio = bio->bi_private; + bool should_dirty = dio->should_dirty; + bch2_direct_IO_read_endio(bio); - bio_check_pages_dirty(bio); /* transfers ownership */ + bio_check_or_release(bio, should_dirty); } static int bch2_direct_IO_read(struct kiocb *req, struct iov_iter *iter) @@ -1713,7 +2119,7 @@ static int bch2_direct_IO_read(struct kiocb *req, struct iov_iter *iter) struct file *file = req->ki_filp; struct bch_inode_info *inode = file_bch_inode(file); struct bch_fs *c = inode->v.i_sb->s_fs_info; - struct bch_io_opts opts = io_opts(c, inode); + struct bch_io_opts opts; struct dio_read *dio; struct bio *bio; loff_t offset = req->ki_pos; @@ -1721,6 +2127,8 @@ static int bch2_direct_IO_read(struct kiocb *req, struct iov_iter *iter) size_t shorten; ssize_t ret; + bch2_inode_opts_get(&opts, c, &inode->ei_inode); + if ((offset|iter->count) & (block_bytes(c) - 1)) return -EINVAL; @@ -1733,8 +2141,10 @@ static int bch2_direct_IO_read(struct kiocb *req, struct iov_iter *iter) shorten = iov_iter_count(iter) - round_up(ret, block_bytes(c)); iter->count -= shorten; - bio = bio_alloc_bioset(GFP_KERNEL, - iov_iter_npages(iter, BIO_MAX_PAGES), + bio = bio_alloc_bioset(NULL, + bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS), + REQ_OP_READ, + GFP_KERNEL, &c->dio_read_bioset); bio->bi_end_io = bch2_direct_IO_read_endio; @@ -1759,15 +2169,23 @@ static int bch2_direct_IO_read(struct kiocb *req, struct iov_iter *iter) dio->req = req; dio->ret = ret; + /* + * This is one of the sketchier things I've encountered: we have to skip + * the dirtying of requests that are internal from the kernel (i.e. from + * loopback), because we'll deadlock on page_lock. + */ + dio->should_dirty = iter_is_iovec(iter); goto start; while (iter->count) { - bio = bio_alloc_bioset(GFP_KERNEL, - iov_iter_npages(iter, BIO_MAX_PAGES), + bio = bio_alloc_bioset(NULL, + bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS), + REQ_OP_READ, + GFP_KERNEL, &c->bio_read); bio->bi_end_io = bch2_direct_IO_read_split_endio; start: - bio_set_op_attrs(bio, REQ_OP_READ, REQ_SYNC); + bio->bi_opf = REQ_OP_READ|REQ_SYNC; bio->bi_iter.bi_sector = offset >> 9; bio->bi_private = dio; @@ -1780,12 +2198,14 @@ start: } offset += bio->bi_iter.bi_size; - bio_set_pages_dirty(bio); + + if (dio->should_dirty) + bio_set_pages_dirty(bio); if (iter->count) closure_get(&dio->cl); - bch2_read(c, rbio_init(bio, opts), inode->v.i_ino); + bch2_read(c, rbio_init(bio, opts), inode_inum(inode)); } iter->count += shorten; @@ -1794,481 +2214,745 @@ start: closure_sync(&dio->cl); closure_debug_destroy(&dio->cl); ret = dio->ret; - bio_check_pages_dirty(&dio->rbio.bio); /* transfers ownership */ + bio_check_or_release(&dio->rbio.bio, dio->should_dirty); return ret; } else { return -EIOCBQUEUED; } } +ssize_t bch2_read_iter(struct kiocb *iocb, struct iov_iter *iter) +{ + struct file *file = iocb->ki_filp; + struct bch_inode_info *inode = file_bch_inode(file); + struct address_space *mapping = file->f_mapping; + size_t count = iov_iter_count(iter); + ssize_t ret; + + if (!count) + return 0; /* skip atime */ + + if (iocb->ki_flags & IOCB_DIRECT) { + struct blk_plug plug; + + if (unlikely(mapping->nrpages)) { + ret = filemap_write_and_wait_range(mapping, + iocb->ki_pos, + iocb->ki_pos + count - 1); + if (ret < 0) + goto out; + } + + file_accessed(file); + + blk_start_plug(&plug); + ret = bch2_direct_IO_read(iocb, iter); + blk_finish_plug(&plug); + + if (ret >= 0) + iocb->ki_pos += ret; + } else { + bch2_pagecache_add_get(inode); + ret = generic_file_read_iter(iocb, iter); + bch2_pagecache_add_put(inode); + } +out: + return bch2_err_class(ret); +} + /* O_DIRECT writes */ -static void bch2_dio_write_loop_async(struct closure *); +static bool bch2_check_range_allocated(struct bch_fs *c, subvol_inum inum, + u64 offset, u64 size, + unsigned nr_replicas, bool compressed) +{ + struct btree_trans trans; + struct btree_iter iter; + struct bkey_s_c k; + u64 end = offset + size; + u32 snapshot; + bool ret = true; + int err; + + bch2_trans_init(&trans, c, 0, 0); +retry: + bch2_trans_begin(&trans); + + err = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot); + if (err) + goto err; + + for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents, + SPOS(inum.inum, offset, snapshot), + BTREE_ITER_SLOTS, k, err) { + if (bkey_ge(bkey_start_pos(k.k), POS(inum.inum, end))) + break; + + if (k.k->p.snapshot != snapshot || + nr_replicas > bch2_bkey_replicas(c, k) || + (!compressed && bch2_bkey_sectors_compressed(k))) { + ret = false; + break; + } + } + + offset = iter.pos.offset; + bch2_trans_iter_exit(&trans, &iter); +err: + if (bch2_err_matches(err, BCH_ERR_transaction_restart)) + goto retry; + bch2_trans_exit(&trans); + + return err ? false : ret; +} + +static noinline bool bch2_dio_write_check_allocated(struct dio_write *dio) +{ + struct bch_fs *c = dio->op.c; + struct bch_inode_info *inode = dio->inode; + struct bio *bio = &dio->op.wbio.bio; + + return bch2_check_range_allocated(c, inode_inum(inode), + dio->op.pos.offset, bio_sectors(bio), + dio->op.opts.data_replicas, + dio->op.opts.compression != 0); +} + +static void bch2_dio_write_loop_async(struct bch_write_op *); +static __always_inline long bch2_dio_write_done(struct dio_write *dio); + +static noinline int bch2_dio_write_copy_iov(struct dio_write *dio) +{ + struct iovec *iov = dio->inline_vecs; + + if (dio->iter.nr_segs > ARRAY_SIZE(dio->inline_vecs)) { + iov = kmalloc_array(dio->iter.nr_segs, sizeof(*iov), + GFP_KERNEL); + if (unlikely(!iov)) + return -ENOMEM; + + dio->free_iov = true; + } + + memcpy(iov, dio->iter.iov, dio->iter.nr_segs * sizeof(*iov)); + dio->iter.iov = iov; + return 0; +} + +static void bch2_dio_write_flush_done(struct closure *cl) +{ + struct dio_write *dio = container_of(cl, struct dio_write, op.cl); + struct bch_fs *c = dio->op.c; -static long bch2_dio_write_loop(struct dio_write *dio) + closure_debug_destroy(cl); + + dio->op.error = bch2_journal_error(&c->journal); + + bch2_dio_write_done(dio); +} + +static noinline void bch2_dio_write_flush(struct dio_write *dio) +{ + struct bch_fs *c = dio->op.c; + struct bch_inode_unpacked inode; + int ret; + + dio->flush = 0; + + closure_init(&dio->op.cl, NULL); + + if (!dio->op.error) { + ret = bch2_inode_find_by_inum(c, inode_inum(dio->inode), &inode); + if (ret) { + dio->op.error = ret; + } else { + bch2_journal_flush_seq_async(&c->journal, inode.bi_journal_seq, &dio->op.cl); + bch2_inode_flush_nocow_writes_async(c, dio->inode, &dio->op.cl); + } + } + + if (dio->sync) { + closure_sync(&dio->op.cl); + closure_debug_destroy(&dio->op.cl); + } else { + continue_at(&dio->op.cl, bch2_dio_write_flush_done, NULL); + } +} + +static __always_inline long bch2_dio_write_done(struct dio_write *dio) { - bool kthread = (current->flags & PF_KTHREAD) != 0; struct kiocb *req = dio->req; - struct address_space *mapping = req->ki_filp->f_mapping; - struct bch_inode_info *inode = dio->iop.inode; - struct bio *bio = &dio->iop.op.wbio.bio; - struct bio_vec *bv; - loff_t offset; - bool sync; + struct bch_inode_info *inode = dio->inode; + bool sync = dio->sync; long ret; - int i; - if (dio->loop) - goto loop; + if (unlikely(dio->flush)) { + bch2_dio_write_flush(dio); + if (!sync) + return -EIOCBQUEUED; + } - inode_dio_begin(&inode->v); - __pagecache_block_get(&mapping->add_lock); + bch2_pagecache_block_put(inode); - /* Write and invalidate pagecache range that we're writing to: */ - offset = req->ki_pos + (dio->iop.op.written << 9); - ret = write_invalidate_inode_pages_range(mapping, - offset, - offset + iov_iter_count(&dio->iter) - 1); - if (unlikely(ret)) - goto err; + if (dio->free_iov) + kfree(dio->iter.iov); + + ret = dio->op.error ?: ((long) dio->written << 9); + bio_put(&dio->op.wbio.bio); + + /* inode->i_dio_count is our ref on inode and thus bch_fs */ + inode_dio_end(&inode->v); + + if (ret < 0) + ret = bch2_err_class(ret); + + if (!sync) { + req->ki_complete(req, ret); + ret = -EIOCBQUEUED; + } + return ret; +} + +static __always_inline void bch2_dio_write_end(struct dio_write *dio) +{ + struct bch_fs *c = dio->op.c; + struct kiocb *req = dio->req; + struct bch_inode_info *inode = dio->inode; + struct bio *bio = &dio->op.wbio.bio; + + req->ki_pos += (u64) dio->op.written << 9; + dio->written += dio->op.written; + + if (dio->extending) { + spin_lock(&inode->v.i_lock); + if (req->ki_pos > inode->v.i_size) + i_size_write(&inode->v, req->ki_pos); + spin_unlock(&inode->v.i_lock); + } + + if (dio->op.i_sectors_delta || dio->quota_res.sectors) { + mutex_lock(&inode->ei_quota_lock); + __i_sectors_acct(c, inode, &dio->quota_res, dio->op.i_sectors_delta); + __bch2_quota_reservation_put(c, inode, &dio->quota_res); + mutex_unlock(&inode->ei_quota_lock); + } + + if (likely(!bio_flagged(bio, BIO_NO_PAGE_REF))) { + struct bvec_iter_all iter; + struct folio_vec fv; + + bio_for_each_folio_all(fv, bio, iter) + folio_put(fv.fv_folio); + } + + if (unlikely(dio->op.error)) + set_bit(EI_INODE_ERROR, &inode->ei_flags); +} + +static __always_inline long bch2_dio_write_loop(struct dio_write *dio) +{ + struct bch_fs *c = dio->op.c; + struct kiocb *req = dio->req; + struct address_space *mapping = dio->mapping; + struct bch_inode_info *inode = dio->inode; + struct bch_io_opts opts; + struct bio *bio = &dio->op.wbio.bio; + unsigned unaligned, iter_count; + bool sync = dio->sync, dropped_locks; + long ret; + + bch2_inode_opts_get(&opts, c, &inode->ei_inode); while (1) { - offset = req->ki_pos + (dio->iop.op.written << 9); + iter_count = dio->iter.count; - BUG_ON(current->pagecache_lock); - current->pagecache_lock = &mapping->add_lock; - if (kthread) - use_mm(dio->mm); + EBUG_ON(current->faults_disabled_mapping); + current->faults_disabled_mapping = mapping; ret = bio_iov_iter_get_pages(bio, &dio->iter); - if (kthread) - unuse_mm(dio->mm); - current->pagecache_lock = NULL; + dropped_locks = fdm_dropped_locks(); + + current->faults_disabled_mapping = NULL; + + /* + * If the fault handler returned an error but also signalled + * that it dropped & retook ei_pagecache_lock, we just need to + * re-shoot down the page cache and retry: + */ + if (dropped_locks && ret) + ret = 0; if (unlikely(ret < 0)) goto err; - /* gup might have faulted pages back in: */ - ret = write_invalidate_inode_pages_range(mapping, - offset, - offset + bio->bi_iter.bi_size - 1); - if (unlikely(ret)) - goto err; + if (unlikely(dropped_locks)) { + ret = write_invalidate_inode_pages_range(mapping, + req->ki_pos, + req->ki_pos + iter_count - 1); + if (unlikely(ret)) + goto err; - dio->iop.op.pos = POS(inode->v.i_ino, offset >> 9); + if (!bio->bi_iter.bi_size) + continue; + } - task_io_account_write(bio->bi_iter.bi_size); + unaligned = bio->bi_iter.bi_size & (block_bytes(c) - 1); + bio->bi_iter.bi_size -= unaligned; + iov_iter_revert(&dio->iter, unaligned); + + if (!bio->bi_iter.bi_size) { + /* + * bio_iov_iter_get_pages was only able to get < + * blocksize worth of pages: + */ + ret = -EFAULT; + goto err; + } - closure_call(&dio->iop.op.cl, bch2_write, NULL, &dio->cl); + bch2_write_op_init(&dio->op, c, opts); + dio->op.end_io = sync + ? NULL + : bch2_dio_write_loop_async; + dio->op.target = dio->op.opts.foreground_target; + dio->op.write_point = writepoint_hashed((unsigned long) current); + dio->op.nr_replicas = dio->op.opts.data_replicas; + dio->op.subvol = inode->ei_subvol; + dio->op.pos = POS(inode->v.i_ino, (u64) req->ki_pos >> 9); + dio->op.devs_need_flush = &inode->ei_devs_need_flush; + + if (sync) + dio->op.flags |= BCH_WRITE_SYNC; + dio->op.flags |= BCH_WRITE_CHECK_ENOSPC; + + ret = bch2_quota_reservation_add(c, inode, &dio->quota_res, + bio_sectors(bio), true); + if (unlikely(ret)) + goto err; - if (!dio->sync && !dio->loop && dio->iter.count) { - struct iovec *iov = dio->inline_vecs; + ret = bch2_disk_reservation_get(c, &dio->op.res, bio_sectors(bio), + dio->op.opts.data_replicas, 0); + if (unlikely(ret) && + !bch2_dio_write_check_allocated(dio)) + goto err; - if (dio->iter.nr_segs > ARRAY_SIZE(dio->inline_vecs)) { - iov = kmalloc(dio->iter.nr_segs * sizeof(*iov), - GFP_KERNEL); - if (unlikely(!iov)) { - dio->iop.op.error = -ENOMEM; - goto err_wait_io; - } + task_io_account_write(bio->bi_iter.bi_size); - dio->free_iov = true; - } + if (unlikely(dio->iter.count) && + !dio->sync && + !dio->loop && + bch2_dio_write_copy_iov(dio)) + dio->sync = sync = true; - memcpy(iov, dio->iter.iov, dio->iter.nr_segs * sizeof(*iov)); - dio->iter.iov = iov; - } -err_wait_io: dio->loop = true; + closure_call(&dio->op.cl, bch2_write, NULL, NULL); - if (!dio->sync) { - continue_at(&dio->cl, bch2_dio_write_loop_async, NULL); + if (!sync) return -EIOCBQUEUED; - } - closure_sync(&dio->cl); -loop: - bio_for_each_segment_all(bv, bio, i) - put_page(bv->bv_page); - if (!dio->iter.count || dio->iop.op.error) + bch2_dio_write_end(dio); + + if (likely(!dio->iter.count) || dio->op.error) break; - bio_reset(bio); - } - ret = dio->iop.op.error ?: ((long) dio->iop.op.written << 9); + bio_reset(bio, NULL, REQ_OP_WRITE); + } +out: + return bch2_dio_write_done(dio); err: - __pagecache_block_put(&mapping->add_lock); - bch2_disk_reservation_put(dio->iop.op.c, &dio->iop.op.res); - bch2_quota_reservation_put(dio->iop.op.c, inode, &dio->quota_res); + dio->op.error = ret; - if (dio->free_iov) - kfree(dio->iter.iov); + if (!bio_flagged(bio, BIO_NO_PAGE_REF)) { + struct bvec_iter_all iter; + struct folio_vec fv; - closure_debug_destroy(&dio->cl); + bio_for_each_folio_all(fv, bio, iter) + folio_put(fv.fv_folio); + } - sync = dio->sync; - bio_put(bio); + bch2_quota_reservation_put(c, inode, &dio->quota_res); + goto out; +} - /* inode->i_dio_count is our ref on inode and thus bch_fs */ - inode_dio_end(&inode->v); +static noinline __cold void bch2_dio_write_continue(struct dio_write *dio) +{ + struct mm_struct *mm = dio->mm; - if (!sync) { - req->ki_complete(req, ret, 0); - ret = -EIOCBQUEUED; - } - return ret; + bio_reset(&dio->op.wbio.bio, NULL, REQ_OP_WRITE); + + if (mm) + kthread_use_mm(mm); + bch2_dio_write_loop(dio); + if (mm) + kthread_unuse_mm(mm); } -static void bch2_dio_write_loop_async(struct closure *cl) +static void bch2_dio_write_loop_async(struct bch_write_op *op) { - struct dio_write *dio = container_of(cl, struct dio_write, cl); + struct dio_write *dio = container_of(op, struct dio_write, op); - bch2_dio_write_loop(dio); + bch2_dio_write_end(dio); + + if (likely(!dio->iter.count) || dio->op.error) + bch2_dio_write_done(dio); + else + bch2_dio_write_continue(dio); } -static int bch2_direct_IO_write(struct kiocb *req, - struct iov_iter *iter, - bool swap) +static noinline +ssize_t bch2_direct_write(struct kiocb *req, struct iov_iter *iter) { struct file *file = req->ki_filp; + struct address_space *mapping = file->f_mapping; struct bch_inode_info *inode = file_bch_inode(file); struct bch_fs *c = inode->v.i_sb->s_fs_info; struct dio_write *dio; struct bio *bio; + bool locked = true, extending; ssize_t ret; - lockdep_assert_held(&inode->v.i_rwsem); + prefetch(&c->opts); + prefetch((void *) &c->opts + 64); + prefetch(&inode->ei_inode); + prefetch((void *) &inode->ei_inode + 64); - if (unlikely(!iter->count)) - return 0; + inode_lock(&inode->v); + + ret = generic_write_checks(req, iter); + if (unlikely(ret <= 0)) + goto err; + + ret = file_remove_privs(file); + if (unlikely(ret)) + goto err; + + ret = file_update_time(file); + if (unlikely(ret)) + goto err; if (unlikely((req->ki_pos|iter->count) & (block_bytes(c) - 1))) - return -EINVAL; + goto err; + + inode_dio_begin(&inode->v); + bch2_pagecache_block_get(inode); + + extending = req->ki_pos + iter->count > inode->v.i_size; + if (!extending) { + inode_unlock(&inode->v); + locked = false; + } - bio = bio_alloc_bioset(GFP_KERNEL, - iov_iter_npages(iter, BIO_MAX_PAGES), + bio = bio_alloc_bioset(NULL, + bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS), + REQ_OP_WRITE, + GFP_KERNEL, &c->dio_write_bioset); - dio = container_of(bio, struct dio_write, iop.op.wbio.bio); - closure_init(&dio->cl, NULL); + dio = container_of(bio, struct dio_write, op.wbio.bio); dio->req = req; + dio->mapping = mapping; + dio->inode = inode; dio->mm = current->mm; dio->loop = false; - dio->sync = is_sync_kiocb(req) || - req->ki_pos + iter->count > inode->v.i_size; + dio->extending = extending; + dio->sync = is_sync_kiocb(req) || extending; + dio->flush = iocb_is_dsync(req) && !c->opts.journal_flush_disabled; dio->free_iov = false; dio->quota_res.sectors = 0; + dio->written = 0; dio->iter = *iter; - bch2_fswrite_op_init(&dio->iop, c, inode, io_opts(c, inode), true); - dio->iop.op.write_point = writepoint_hashed((unsigned long) current); - dio->iop.op.flags |= BCH_WRITE_NOPUT_RESERVATION; - - if ((req->ki_flags & IOCB_DSYNC) && - !c->opts.journal_flush_disabled) - dio->iop.op.flags |= BCH_WRITE_FLUSH; - - ret = bch2_quota_reservation_add(c, inode, &dio->quota_res, - iter->count >> 9, true); - if (unlikely(ret)) - goto err; - - dio->iop.op.nr_replicas = dio->iop.op.opts.data_replicas; - - ret = bch2_disk_reservation_get(c, &dio->iop.op.res, iter->count >> 9, - dio->iop.op.opts.data_replicas, 0); - if (unlikely(ret)) { - if (!bch2_check_range_allocated(c, POS(inode->v.i_ino, - req->ki_pos >> 9), - iter->count >> 9, - dio->iop.op.opts.data_replicas)) - goto err; + dio->op.c = c; - dio->iop.unalloc = true; + if (unlikely(mapping->nrpages)) { + ret = write_invalidate_inode_pages_range(mapping, + req->ki_pos, + req->ki_pos + iter->count - 1); + if (unlikely(ret)) + goto err_put_bio; } - return bch2_dio_write_loop(dio); + ret = bch2_dio_write_loop(dio); err: - bch2_disk_reservation_put(c, &dio->iop.op.res); - bch2_quota_reservation_put(c, inode, &dio->quota_res); - closure_debug_destroy(&dio->cl); - bio_put(bio); - return ret; -} - -ssize_t bch2_direct_IO(struct kiocb *req, struct iov_iter *iter) -{ - struct blk_plug plug; - ssize_t ret; - - blk_start_plug(&plug); - ret = iov_iter_rw(iter) == WRITE - ? bch2_direct_IO_write(req, iter, false) - : bch2_direct_IO_read(req, iter); - blk_finish_plug(&plug); - + if (locked) + inode_unlock(&inode->v); return ret; +err_put_bio: + bch2_pagecache_block_put(inode); + bio_put(bio); + inode_dio_end(&inode->v); + goto err; } -static ssize_t -bch2_direct_write(struct kiocb *iocb, struct iov_iter *iter) -{ - return bch2_direct_IO_write(iocb, iter, true); -} - -static ssize_t __bch2_write_iter(struct kiocb *iocb, struct iov_iter *from) +ssize_t bch2_write_iter(struct kiocb *iocb, struct iov_iter *from) { struct file *file = iocb->ki_filp; struct bch_inode_info *inode = file_bch_inode(file); - ssize_t ret; + ssize_t ret; + + if (iocb->ki_flags & IOCB_DIRECT) { + ret = bch2_direct_write(iocb, from); + goto out; + } /* We can write back this queue in page reclaim */ current->backing_dev_info = inode_to_bdi(&inode->v); + inode_lock(&inode->v); + + ret = generic_write_checks(iocb, from); + if (ret <= 0) + goto unlock; + ret = file_remove_privs(file); if (ret) - goto out; + goto unlock; ret = file_update_time(file); if (ret) - goto out; - - ret = iocb->ki_flags & IOCB_DIRECT - ? bch2_direct_write(iocb, from) - : bch2_buffered_write(iocb, from); + goto unlock; + ret = bch2_buffered_write(iocb, from); if (likely(ret > 0)) iocb->ki_pos += ret; -out: +unlock: + inode_unlock(&inode->v); current->backing_dev_info = NULL; - return ret; -} -ssize_t bch2_write_iter(struct kiocb *iocb, struct iov_iter *from) -{ - struct bch_inode_info *inode = file_bch_inode(iocb->ki_filp); - bool direct = iocb->ki_flags & IOCB_DIRECT; - ssize_t ret; - - inode_lock(&inode->v); - ret = generic_write_checks(iocb, from); if (ret > 0) - ret = __bch2_write_iter(iocb, from); - inode_unlock(&inode->v); - - if (ret > 0 && !direct) ret = generic_write_sync(iocb, ret); - - return ret; +out: + return bch2_err_class(ret); } /* fsync: */ -int bch2_fsync(struct file *file, loff_t start, loff_t end, int datasync) +/* + * inode->ei_inode.bi_journal_seq won't be up to date since it's set in an + * insert trigger: look up the btree inode instead + */ +static int bch2_flush_inode(struct bch_fs *c, + struct bch_inode_info *inode) { - struct bch_inode_info *inode = file_bch_inode(file); - struct bch_fs *c = inode->v.i_sb->s_fs_info; - int ret, ret2; - - ret = file_write_and_wait_range(file, start, end); - if (ret) - return ret; - - if (datasync && !(inode->v.i_state & I_DIRTY_DATASYNC)) - goto out; + struct bch_inode_unpacked u; + int ret; - ret = sync_inode_metadata(&inode->v, 1); - if (ret) - return ret; -out: if (c->opts.journal_flush_disabled) return 0; - ret = bch2_journal_flush_seq(&c->journal, inode->ei_journal_seq); - ret2 = file_check_and_advance_wb_err(file); + ret = bch2_inode_find_by_inum(c, inode_inum(inode), &u); + if (ret) + return ret; - return ret ?: ret2; + return bch2_journal_flush_seq(&c->journal, u.bi_journal_seq) ?: + bch2_inode_flush_nocow_writes(c, inode); } -/* truncate: */ - -static int __bch2_fpunch(struct bch_fs *c, struct bch_inode_info *inode, - u64 start_offset, u64 end_offset, u64 *journal_seq) +int bch2_fsync(struct file *file, loff_t start, loff_t end, int datasync) { - struct bpos start = POS(inode->v.i_ino, start_offset); - struct bpos end = POS(inode->v.i_ino, end_offset); - unsigned max_sectors = KEY_SIZE_MAX & (~0 << c->block_bits); - struct btree_trans trans; - struct btree_iter *iter; - struct bkey_s_c k; - int ret = 0; - - bch2_trans_init(&trans, c); - bch2_trans_preload_iters(&trans); - - iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS, start, - BTREE_ITER_INTENT); - - while ((k = bch2_btree_iter_peek(iter)).k && - !(ret = bkey_err(k)) && - bkey_cmp(iter->pos, end) < 0) { - struct disk_reservation disk_res = - bch2_disk_reservation_init(c, 0); - struct bkey_i delete; - - bkey_init(&delete.k); - delete.k.p = iter->pos; - - /* create the biggest key we can */ - bch2_key_resize(&delete.k, max_sectors); - bch2_cut_back(end, &delete.k); - - ret = bch2_extent_update(&trans, inode, - &disk_res, NULL, iter, &delete, - 0, true, true, NULL); - bch2_disk_reservation_put(c, &disk_res); - - if (ret == -EINTR) - ret = 0; - if (ret) - break; - - bch2_trans_cond_resched(&trans); - } + struct bch_inode_info *inode = file_bch_inode(file); + struct bch_fs *c = inode->v.i_sb->s_fs_info; + int ret, ret2, ret3; - bch2_trans_exit(&trans); + ret = file_write_and_wait_range(file, start, end); + ret2 = sync_inode_metadata(&inode->v, 1); + ret3 = bch2_flush_inode(c, inode); - return ret; + return bch2_err_class(ret ?: ret2 ?: ret3); } -static inline int range_has_data(struct bch_fs *c, - struct bpos start, - struct bpos end) +/* truncate: */ + +static inline int range_has_data(struct bch_fs *c, u32 subvol, + struct bpos start, + struct bpos end) { struct btree_trans trans; - struct btree_iter *iter; + struct btree_iter iter; struct bkey_s_c k; int ret = 0; - bch2_trans_init(&trans, c); + bch2_trans_init(&trans, c, 0, 0); +retry: + bch2_trans_begin(&trans); - for_each_btree_key(&trans, iter, BTREE_ID_EXTENTS, start, 0, k, ret) { - if (bkey_cmp(bkey_start_pos(k.k), end) >= 0) - break; + ret = bch2_subvolume_get_snapshot(&trans, subvol, &start.snapshot); + if (ret) + goto err; - if (bkey_extent_is_data(k.k)) { + for_each_btree_key_upto_norestart(&trans, iter, BTREE_ID_extents, start, end, 0, k, ret) + if (bkey_extent_is_data(k.k) && !bkey_extent_is_unwritten(k)) { ret = 1; break; } - } + start = iter.pos; + bch2_trans_iter_exit(&trans, &iter); +err: + if (bch2_err_matches(ret, BCH_ERR_transaction_restart)) + goto retry; - return bch2_trans_exit(&trans) ?: ret; + bch2_trans_exit(&trans); + return ret; } -static int __bch2_truncate_page(struct bch_inode_info *inode, - pgoff_t index, loff_t start, loff_t end) +static int __bch2_truncate_folio(struct bch_inode_info *inode, + pgoff_t index, loff_t start, loff_t end) { struct bch_fs *c = inode->v.i_sb->s_fs_info; struct address_space *mapping = inode->v.i_mapping; + struct bch_folio *s; unsigned start_offset = start & (PAGE_SIZE - 1); unsigned end_offset = ((end - 1) & (PAGE_SIZE - 1)) + 1; - struct page *page; + unsigned i; + struct folio *folio; + s64 i_sectors_delta = 0; int ret = 0; + u64 end_pos; - /* Page boundary? Nothing to do */ - if (!((index == start >> PAGE_SHIFT && start_offset) || - (index == end >> PAGE_SHIFT && end_offset != PAGE_SIZE))) - return 0; - - /* Above i_size? */ - if (index << PAGE_SHIFT >= inode->v.i_size) - return 0; - - page = find_lock_page(mapping, index); - if (!page) { + folio = filemap_lock_folio(mapping, index); + if (!folio) { /* * XXX: we're doing two index lookups when we end up reading the - * page + * folio */ - ret = range_has_data(c, - POS(inode->v.i_ino, index << PAGE_SECTOR_SHIFT), - POS(inode->v.i_ino, (index + 1) << PAGE_SECTOR_SHIFT)); + ret = range_has_data(c, inode->ei_subvol, + POS(inode->v.i_ino, (index << PAGE_SECTORS_SHIFT)), + POS(inode->v.i_ino, (index << PAGE_SECTORS_SHIFT) + PAGE_SECTORS)); if (ret <= 0) return ret; - page = find_or_create_page(mapping, index, GFP_KERNEL); - if (unlikely(!page)) { + folio = __filemap_get_folio(mapping, index, + FGP_LOCK|FGP_CREAT, GFP_KERNEL); + if (unlikely(!folio)) { ret = -ENOMEM; goto out; } } - if (!PageUptodate(page)) { - ret = bch2_read_single_page(page, mapping); + BUG_ON(start >= folio_end_pos(folio)); + BUG_ON(end <= folio_pos(folio)); + + start_offset = max(start, folio_pos(folio)) - folio_pos(folio); + end_offset = min_t(u64, end, folio_end_pos(folio)) - folio_pos(folio); + + /* Folio boundary? Nothing to do */ + if (start_offset == 0 && + end_offset == folio_size(folio)) { + ret = 0; + goto unlock; + } + + s = bch2_folio_create(folio, 0); + if (!s) { + ret = -ENOMEM; + goto unlock; + } + + if (!folio_test_uptodate(folio)) { + ret = bch2_read_single_folio(folio, mapping); if (ret) goto unlock; } + ret = bch2_folio_set(c, inode_inum(inode), &folio, 1); + if (ret) + goto unlock; + + for (i = round_up(start_offset, block_bytes(c)) >> 9; + i < round_down(end_offset, block_bytes(c)) >> 9; + i++) { + s->s[i].nr_replicas = 0; + + i_sectors_delta -= s->s[i].state == SECTOR_dirty; + folio_sector_set(folio, s, i, SECTOR_unallocated); + } + + i_sectors_acct(c, inode, NULL, i_sectors_delta); + + /* + * Caller needs to know whether this folio will be written out by + * writeback - doing an i_size update if necessary - or whether it will + * be responsible for the i_size update. + * + * Note that we shouldn't ever see a folio beyond EOF, but check and + * warn if so. This has been observed by failure to clean up folios + * after a short write and there's still a chance reclaim will fix + * things up. + */ + WARN_ON_ONCE(folio_pos(folio) >= inode->v.i_size); + end_pos = folio_end_pos(folio); + if (inode->v.i_size > folio_pos(folio)) + end_pos = min_t(u64, inode->v.i_size, end_pos); + ret = s->s[folio_pos_to_s(folio, end_pos - 1)].state >= SECTOR_dirty; + + folio_zero_segment(folio, start_offset, end_offset); + /* * Bit of a hack - we don't want truncate to fail due to -ENOSPC. * - * XXX: because we aren't currently tracking whether the page has actual + * XXX: because we aren't currently tracking whether the folio has actual * data in it (vs. just 0s, or only partially written) this wrong. ick. */ - ret = bch2_get_page_reservation(c, inode, page, false); - BUG_ON(ret); - - if (index == start >> PAGE_SHIFT && - index == end >> PAGE_SHIFT) - zero_user_segment(page, start_offset, end_offset); - else if (index == start >> PAGE_SHIFT) - zero_user_segment(page, start_offset, PAGE_SIZE); - else if (index == end >> PAGE_SHIFT) - zero_user_segment(page, 0, end_offset); + BUG_ON(bch2_get_folio_disk_reservation(c, inode, folio, false)); - if (!PageDirty(page)) - set_page_dirty(page); + /* + * This removes any writeable userspace mappings; we need to force + * .page_mkwrite to be called again before any mmapped writes, to + * redirty the full page: + */ + folio_mkclean(folio); + filemap_dirty_folio(mapping, folio); unlock: - unlock_page(page); - put_page(page); + folio_unlock(folio); + folio_put(folio); out: return ret; } -static int bch2_truncate_page(struct bch_inode_info *inode, loff_t from) +static int bch2_truncate_folio(struct bch_inode_info *inode, loff_t from) { - return __bch2_truncate_page(inode, from >> PAGE_SHIFT, - from, from + PAGE_SIZE); + return __bch2_truncate_folio(inode, from >> PAGE_SHIFT, + from, ANYSINT_MAX(loff_t)); } -static int bch2_extend(struct bch_inode_info *inode, struct iattr *iattr) +static int bch2_truncate_folios(struct bch_inode_info *inode, + loff_t start, loff_t end) +{ + int ret = __bch2_truncate_folio(inode, start >> PAGE_SHIFT, + start, end); + + if (ret >= 0 && + start >> PAGE_SHIFT != end >> PAGE_SHIFT) + ret = __bch2_truncate_folio(inode, + (end - 1) >> PAGE_SHIFT, + start, end); + return ret; +} + +static int bch2_extend(struct mnt_idmap *idmap, + struct bch_inode_info *inode, + struct bch_inode_unpacked *inode_u, + struct iattr *iattr) { - struct bch_fs *c = inode->v.i_sb->s_fs_info; struct address_space *mapping = inode->v.i_mapping; int ret; - ret = filemap_write_and_wait_range(mapping, - inode->ei_inode.bi_size, S64_MAX); + /* + * sync appends: + * + * this has to be done _before_ extending i_size: + */ + ret = filemap_write_and_wait_range(mapping, inode_u->bi_size, S64_MAX); if (ret) return ret; truncate_setsize(&inode->v, iattr->ia_size); - setattr_copy(&inode->v, iattr); - - mutex_lock(&inode->ei_update_lock); - ret = bch2_write_inode_size(c, inode, inode->v.i_size, - ATTR_MTIME|ATTR_CTIME); - mutex_unlock(&inode->ei_update_lock); - return ret; + return bch2_setattr_nonsize(idmap, inode, iattr); } static int bch2_truncate_finish_fn(struct bch_inode_info *inode, struct bch_inode_unpacked *bi, void *p) { - struct bch_fs *c = inode->v.i_sb->s_fs_info; - bi->bi_flags &= ~BCH_INODE_I_SIZE_DIRTY; - bi->bi_mtime = bi->bi_ctime = bch2_current_time(c); return 0; } @@ -2282,33 +2966,74 @@ static int bch2_truncate_start_fn(struct bch_inode_info *inode, return 0; } -int bch2_truncate(struct bch_inode_info *inode, struct iattr *iattr) +int bch2_truncate(struct mnt_idmap *idmap, + struct bch_inode_info *inode, struct iattr *iattr) { struct bch_fs *c = inode->v.i_sb->s_fs_info; struct address_space *mapping = inode->v.i_mapping; + struct bch_inode_unpacked inode_u; u64 new_i_size = iattr->ia_size; - bool shrink; + s64 i_sectors_delta = 0; int ret = 0; + /* + * If the truncate call with change the size of the file, the + * cmtimes should be updated. If the size will not change, we + * do not need to update the cmtimes. + */ + if (iattr->ia_size != inode->v.i_size) { + if (!(iattr->ia_valid & ATTR_MTIME)) + ktime_get_coarse_real_ts64(&iattr->ia_mtime); + if (!(iattr->ia_valid & ATTR_CTIME)) + ktime_get_coarse_real_ts64(&iattr->ia_ctime); + iattr->ia_valid |= ATTR_MTIME|ATTR_CTIME; + } + inode_dio_wait(&inode->v); - pagecache_block_get(&mapping->add_lock); + bch2_pagecache_block_get(inode); + + ret = bch2_inode_find_by_inum(c, inode_inum(inode), &inode_u); + if (ret) + goto err; - BUG_ON(inode->v.i_size < inode->ei_inode.bi_size); + /* + * check this before next assertion; on filesystem error our normal + * invariants are a bit broken (truncate has to truncate the page cache + * before the inode). + */ + ret = bch2_journal_error(&c->journal); + if (ret) + goto err; - shrink = iattr->ia_size <= inode->v.i_size; + WARN_ONCE(!test_bit(EI_INODE_ERROR, &inode->ei_flags) && + inode->v.i_size < inode_u.bi_size, + "truncate spotted in mem i_size < btree i_size: %llu < %llu\n", + (u64) inode->v.i_size, inode_u.bi_size); - if (!shrink) { - ret = bch2_extend(inode, iattr); + if (iattr->ia_size > inode->v.i_size) { + ret = bch2_extend(idmap, inode, &inode_u, iattr); goto err; } - ret = bch2_truncate_page(inode, iattr->ia_size); - if (unlikely(ret)) + iattr->ia_valid &= ~ATTR_SIZE; + + ret = bch2_truncate_folio(inode, iattr->ia_size); + if (unlikely(ret < 0)) goto err; - if (iattr->ia_size > inode->ei_inode.bi_size) + /* + * When extending, we're going to write the new i_size to disk + * immediately so we need to flush anything above the current on disk + * i_size first: + * + * Also, when extending we need to flush the page that i_size currently + * straddles - if it's mapped to userspace, we need to ensure that + * userspace has to redirty it and call .mkwrite -> set_page_dirty + * again to allocate the part of the page that was extended. + */ + if (iattr->ia_size > inode_u.bi_size) ret = filemap_write_and_wait_range(mapping, - inode->ei_inode.bi_size, + inode_u.bi_size, iattr->ia_size - 1); else if (iattr->ia_size & (PAGE_SIZE - 1)) ret = filemap_write_and_wait_range(mapping, @@ -2327,273 +3052,313 @@ int bch2_truncate(struct bch_inode_info *inode, struct iattr *iattr) truncate_setsize(&inode->v, iattr->ia_size); - /* - * XXX: need a comment explaining why PAGE_SIZE and not block_bytes() - * here: - */ - ret = __bch2_fpunch(c, inode, - round_up(iattr->ia_size, PAGE_SIZE) >> 9, - U64_MAX, &inode->ei_journal_seq); + ret = bch2_fpunch(c, inode_inum(inode), + round_up(iattr->ia_size, block_bytes(c)) >> 9, + U64_MAX, &i_sectors_delta); + i_sectors_acct(c, inode, NULL, i_sectors_delta); + + bch2_fs_inconsistent_on(!inode->v.i_size && inode->v.i_blocks && + !bch2_journal_error(&c->journal), c, + "inode %lu truncated to 0 but i_blocks %llu (ondisk %lli)", + inode->v.i_ino, (u64) inode->v.i_blocks, + inode->ei_inode.bi_sectors); if (unlikely(ret)) goto err; - setattr_copy(&inode->v, iattr); - mutex_lock(&inode->ei_update_lock); - ret = bch2_write_inode(c, inode, bch2_truncate_finish_fn, NULL, - ATTR_MTIME|ATTR_CTIME); + ret = bch2_write_inode(c, inode, bch2_truncate_finish_fn, NULL, 0); mutex_unlock(&inode->ei_update_lock); + + ret = bch2_setattr_nonsize(idmap, inode, iattr); err: - pagecache_block_put(&mapping->add_lock); - return ret; + bch2_pagecache_block_put(inode); + return bch2_err_class(ret); } /* fallocate: */ -static long bch2_fpunch(struct bch_inode_info *inode, loff_t offset, loff_t len) +static int inode_update_times_fn(struct bch_inode_info *inode, + struct bch_inode_unpacked *bi, void *p) { struct bch_fs *c = inode->v.i_sb->s_fs_info; - struct address_space *mapping = inode->v.i_mapping; - u64 discard_start = round_up(offset, PAGE_SIZE) >> 9; - u64 discard_end = round_down(offset + len, PAGE_SIZE) >> 9; - int ret = 0; - inode_lock(&inode->v); - inode_dio_wait(&inode->v); - pagecache_block_get(&mapping->add_lock); + bi->bi_mtime = bi->bi_ctime = bch2_current_time(c); + return 0; +} - ret = __bch2_truncate_page(inode, - offset >> PAGE_SHIFT, - offset, offset + len); - if (unlikely(ret)) +static long bchfs_fpunch(struct bch_inode_info *inode, loff_t offset, loff_t len) +{ + struct bch_fs *c = inode->v.i_sb->s_fs_info; + u64 end = offset + len; + u64 block_start = round_up(offset, block_bytes(c)); + u64 block_end = round_down(end, block_bytes(c)); + bool truncated_last_page; + int ret = 0; + + ret = bch2_truncate_folios(inode, offset, end); + if (unlikely(ret < 0)) goto err; - if (offset >> PAGE_SHIFT != - (offset + len) >> PAGE_SHIFT) { - ret = __bch2_truncate_page(inode, - (offset + len) >> PAGE_SHIFT, - offset, offset + len); - if (unlikely(ret)) - goto err; - } + truncated_last_page = ret; - truncate_pagecache_range(&inode->v, offset, offset + len - 1); + truncate_pagecache_range(&inode->v, offset, end - 1); - if (discard_start < discard_end) - ret = __bch2_fpunch(c, inode, discard_start, discard_end, - &inode->ei_journal_seq); -err: - pagecache_block_put(&mapping->add_lock); - inode_unlock(&inode->v); + if (block_start < block_end) { + s64 i_sectors_delta = 0; + + ret = bch2_fpunch(c, inode_inum(inode), + block_start >> 9, block_end >> 9, + &i_sectors_delta); + i_sectors_acct(c, inode, NULL, i_sectors_delta); + } + mutex_lock(&inode->ei_update_lock); + if (end >= inode->v.i_size && !truncated_last_page) { + ret = bch2_write_inode_size(c, inode, inode->v.i_size, + ATTR_MTIME|ATTR_CTIME); + } else { + ret = bch2_write_inode(c, inode, inode_update_times_fn, NULL, + ATTR_MTIME|ATTR_CTIME); + } + mutex_unlock(&inode->ei_update_lock); +err: return ret; } -static long bch2_fcollapse(struct bch_inode_info *inode, - loff_t offset, loff_t len) +static long bchfs_fcollapse_finsert(struct bch_inode_info *inode, + loff_t offset, loff_t len, + bool insert) { struct bch_fs *c = inode->v.i_sb->s_fs_info; struct address_space *mapping = inode->v.i_mapping; + struct bkey_buf copy; struct btree_trans trans; - struct btree_iter *src, *dst; - BKEY_PADDED(k) copy; - struct bkey_s_c k; - loff_t new_size; - int ret; + struct btree_iter src, dst, del; + loff_t shift, new_size; + u64 src_start; + int ret = 0; if ((offset | len) & (block_bytes(c) - 1)) return -EINVAL; - bch2_trans_init(&trans, c); - bch2_trans_preload_iters(&trans); + if (insert) { + if (inode->v.i_sb->s_maxbytes - inode->v.i_size < len) + return -EFBIG; - /* - * We need i_mutex to keep the page cache consistent with the extents - * btree, and the btree consistent with i_size - we don't need outside - * locking for the extents btree itself, because we're using linked - * iterators - */ - inode_lock(&inode->v); - inode_dio_wait(&inode->v); - pagecache_block_get(&mapping->add_lock); + if (offset >= inode->v.i_size) + return -EINVAL; - ret = -EINVAL; - if (offset + len >= inode->v.i_size) - goto err; + src_start = U64_MAX; + shift = len; + } else { + if (offset + len >= inode->v.i_size) + return -EINVAL; - if (inode->v.i_size < len) - goto err; + src_start = offset + len; + shift = -len; + } - new_size = inode->v.i_size - len; + new_size = inode->v.i_size + shift; ret = write_invalidate_inode_pages_range(mapping, offset, LLONG_MAX); if (ret) - goto err; + return ret; - dst = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS, - POS(inode->v.i_ino, offset >> 9), - BTREE_ITER_SLOTS|BTREE_ITER_INTENT); - BUG_ON(IS_ERR_OR_NULL(dst)); + if (insert) { + i_size_write(&inode->v, new_size); + mutex_lock(&inode->ei_update_lock); + ret = bch2_write_inode_size(c, inode, new_size, + ATTR_MTIME|ATTR_CTIME); + mutex_unlock(&inode->ei_update_lock); + } else { + s64 i_sectors_delta = 0; - src = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS, - POS_MIN, BTREE_ITER_SLOTS); - BUG_ON(IS_ERR_OR_NULL(src)); + ret = bch2_fpunch(c, inode_inum(inode), + offset >> 9, (offset + len) >> 9, + &i_sectors_delta); + i_sectors_acct(c, inode, NULL, i_sectors_delta); - while (bkey_cmp(dst->pos, - POS(inode->v.i_ino, - round_up(new_size, PAGE_SIZE) >> 9)) < 0) { - struct disk_reservation disk_res; + if (ret) + return ret; + } + + bch2_bkey_buf_init(©); + bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024); + bch2_trans_iter_init(&trans, &src, BTREE_ID_extents, + POS(inode->v.i_ino, src_start >> 9), + BTREE_ITER_INTENT); + bch2_trans_copy_iter(&dst, &src); + bch2_trans_copy_iter(&del, &src); + + while (ret == 0 || + bch2_err_matches(ret, BCH_ERR_transaction_restart)) { + struct disk_reservation disk_res = + bch2_disk_reservation_init(c, 0); + struct bkey_i delete; + struct bkey_s_c k; + struct bpos next_pos; + struct bpos move_pos = POS(inode->v.i_ino, offset >> 9); + struct bpos atomic_end; + unsigned trigger_flags = 0; + u32 snapshot; - ret = bch2_btree_iter_traverse(dst); + bch2_trans_begin(&trans); + + ret = bch2_subvolume_get_snapshot(&trans, + inode->ei_subvol, &snapshot); if (ret) - goto bkey_err; + continue; - bch2_btree_iter_set_pos(src, - POS(dst->pos.inode, dst->pos.offset + (len >> 9))); + bch2_btree_iter_set_snapshot(&src, snapshot); + bch2_btree_iter_set_snapshot(&dst, snapshot); + bch2_btree_iter_set_snapshot(&del, snapshot); - k = bch2_btree_iter_peek_slot(src); + bch2_trans_begin(&trans); + + k = insert + ? bch2_btree_iter_peek_prev(&src) + : bch2_btree_iter_peek_upto(&src, POS(inode->v.i_ino, U64_MAX)); if ((ret = bkey_err(k))) - goto bkey_err; + continue; - bkey_reassemble(©.k, k); + if (!k.k || k.k->p.inode != inode->v.i_ino) + break; - bch2_cut_front(src->pos, ©.k); - copy.k.k.p.offset -= len >> 9; + if (insert && + bkey_le(k.k->p, POS(inode->v.i_ino, offset >> 9))) + break; +reassemble: + bch2_bkey_buf_reassemble(©, c, k); - bch2_extent_trim_atomic(©.k, dst); + if (insert && + bkey_lt(bkey_start_pos(k.k), move_pos)) + bch2_cut_front(move_pos, copy.k); - BUG_ON(bkey_cmp(dst->pos, bkey_start_pos(©.k.k))); + copy.k->k.p.offset += shift >> 9; + bch2_btree_iter_set_pos(&dst, bkey_start_pos(©.k->k)); - ret = bch2_disk_reservation_get(c, &disk_res, copy.k.k.size, - bch2_bkey_nr_dirty_ptrs(bkey_i_to_s_c(©.k)), - BCH_DISK_RESERVATION_NOFAIL); - BUG_ON(ret); + ret = bch2_extent_atomic_end(&trans, &dst, copy.k, &atomic_end); + if (ret) + continue; - ret = bch2_extent_update(&trans, inode, - &disk_res, NULL, - dst, ©.k, - 0, true, true, NULL); + if (!bkey_eq(atomic_end, copy.k->k.p)) { + if (insert) { + move_pos = atomic_end; + move_pos.offset -= shift >> 9; + goto reassemble; + } else { + bch2_cut_back(atomic_end, copy.k); + } + } + + bkey_init(&delete.k); + delete.k.p = copy.k->k.p; + delete.k.size = copy.k->k.size; + delete.k.p.offset -= shift >> 9; + bch2_btree_iter_set_pos(&del, bkey_start_pos(&delete.k)); + + next_pos = insert ? bkey_start_pos(&delete.k) : delete.k.p; + + if (copy.k->k.size != k.k->size) { + /* We might end up splitting compressed extents: */ + unsigned nr_ptrs = + bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(copy.k)); + + ret = bch2_disk_reservation_get(c, &disk_res, + copy.k->k.size, nr_ptrs, + BCH_DISK_RESERVATION_NOFAIL); + BUG_ON(ret); + } + + ret = bch2_btree_iter_traverse(&del) ?: + bch2_trans_update(&trans, &del, &delete, trigger_flags) ?: + bch2_trans_update(&trans, &dst, copy.k, trigger_flags) ?: + bch2_trans_commit(&trans, &disk_res, NULL, + BTREE_INSERT_NOFAIL); bch2_disk_reservation_put(c, &disk_res); -bkey_err: - if (ret == -EINTR) - ret = 0; - if (ret) - goto err; - /* - * XXX: if we error here we've left data with multiple - * pointers... which isn't a _super_ serious problem... - */ - bch2_trans_cond_resched(&trans); + if (!ret) + bch2_btree_iter_set_pos(&src, next_pos); } - bch2_trans_unlock(&trans); + bch2_trans_iter_exit(&trans, &del); + bch2_trans_iter_exit(&trans, &dst); + bch2_trans_iter_exit(&trans, &src); + bch2_trans_exit(&trans); + bch2_bkey_buf_exit(©, c); - ret = __bch2_fpunch(c, inode, - round_up(new_size, block_bytes(c)) >> 9, - U64_MAX, &inode->ei_journal_seq); if (ret) - goto err; + return ret; - i_size_write(&inode->v, new_size); mutex_lock(&inode->ei_update_lock); - ret = bch2_write_inode_size(c, inode, new_size, - ATTR_MTIME|ATTR_CTIME); + if (!insert) { + i_size_write(&inode->v, new_size); + ret = bch2_write_inode_size(c, inode, new_size, + ATTR_MTIME|ATTR_CTIME); + } else { + /* We need an inode update to update bi_journal_seq for fsync: */ + ret = bch2_write_inode(c, inode, inode_update_times_fn, NULL, + ATTR_MTIME|ATTR_CTIME); + } mutex_unlock(&inode->ei_update_lock); -err: - bch2_trans_exit(&trans); - pagecache_block_put(&mapping->add_lock); - inode_unlock(&inode->v); return ret; } -static long bch2_fallocate(struct bch_inode_info *inode, int mode, - loff_t offset, loff_t len) +static int __bchfs_fallocate(struct bch_inode_info *inode, int mode, + u64 start_sector, u64 end_sector) { - struct address_space *mapping = inode->v.i_mapping; struct bch_fs *c = inode->v.i_sb->s_fs_info; struct btree_trans trans; - struct btree_iter *iter; - struct bpos end_pos; - loff_t block_start, block_end; - loff_t end = offset + len; - unsigned sectors; - unsigned replicas = io_opts(c, inode).data_replicas; - int ret; - - bch2_trans_init(&trans, c); - bch2_trans_preload_iters(&trans); - - inode_lock(&inode->v); - inode_dio_wait(&inode->v); - pagecache_block_get(&mapping->add_lock); - - if (!(mode & FALLOC_FL_KEEP_SIZE) && end > inode->v.i_size) { - ret = inode_newsize_ok(&inode->v, end); - if (ret) - goto err; - } - - if (mode & FALLOC_FL_ZERO_RANGE) { - ret = __bch2_truncate_page(inode, - offset >> PAGE_SHIFT, - offset, end); - - if (!ret && - offset >> PAGE_SHIFT != end >> PAGE_SHIFT) - ret = __bch2_truncate_page(inode, - end >> PAGE_SHIFT, - offset, end); - - if (unlikely(ret)) - goto err; - - truncate_pagecache_range(&inode->v, offset, end - 1); + struct btree_iter iter; + struct bpos end_pos = POS(inode->v.i_ino, end_sector); + struct bch_io_opts opts; + int ret = 0; - block_start = round_up(offset, PAGE_SIZE); - block_end = round_down(end, PAGE_SIZE); - } else { - block_start = round_down(offset, PAGE_SIZE); - block_end = round_up(end, PAGE_SIZE); - } + bch2_inode_opts_get(&opts, c, &inode->ei_inode); + bch2_trans_init(&trans, c, BTREE_ITER_MAX, 512); - iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS, - POS(inode->v.i_ino, block_start >> 9), + bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents, + POS(inode->v.i_ino, start_sector), BTREE_ITER_SLOTS|BTREE_ITER_INTENT); - end_pos = POS(inode->v.i_ino, block_end >> 9); - while (bkey_cmp(iter->pos, end_pos) < 0) { - struct disk_reservation disk_res = { 0 }; + while (!ret && bkey_lt(iter.pos, end_pos)) { + s64 i_sectors_delta = 0; struct quota_res quota_res = { 0 }; - struct bkey_i_reservation reservation; struct bkey_s_c k; + unsigned sectors; + u32 snapshot; + + bch2_trans_begin(&trans); + + ret = bch2_subvolume_get_snapshot(&trans, + inode->ei_subvol, &snapshot); + if (ret) + goto bkey_err; + + bch2_btree_iter_set_snapshot(&iter, snapshot); - k = bch2_btree_iter_peek_slot(iter); + k = bch2_btree_iter_peek_slot(&iter); if ((ret = bkey_err(k))) goto bkey_err; /* already reserved */ - if (k.k->type == KEY_TYPE_reservation && - bkey_s_c_to_reservation(k).v->nr_replicas >= replicas) { - bch2_btree_iter_next_slot(iter); + if (bkey_extent_is_reservation(k) && + bch2_bkey_nr_ptrs_fully_allocated(k) >= opts.data_replicas) { + bch2_btree_iter_advance(&iter); continue; } if (bkey_extent_is_data(k.k) && !(mode & FALLOC_FL_ZERO_RANGE)) { - bch2_btree_iter_next_slot(iter); + bch2_btree_iter_advance(&iter); continue; } - bkey_reservation_init(&reservation.k_i); - reservation.k.type = KEY_TYPE_reservation; - reservation.k.p = k.k->p; - reservation.k.size = k.k->size; - - bch2_cut_front(iter->pos, &reservation.k_i); - bch2_cut_back(end_pos, &reservation.k); + /* + * XXX: for nocow mode, we should promote shared extents to + * unshared here + */ - sectors = reservation.k.size; - reservation.v.nr_replicas = bch2_bkey_nr_dirty_ptrs(k); + sectors = bpos_min(k.k->p, end_pos).offset - iter.pos.offset; if (!bkey_extent_is_allocation(k.k)) { ret = bch2_quota_reservation_add(c, inode, @@ -2603,116 +3368,315 @@ static long bch2_fallocate(struct bch_inode_info *inode, int mode, goto bkey_err; } - if (reservation.v.nr_replicas < replicas || - bch2_extent_is_compressed(k)) { - ret = bch2_disk_reservation_get(c, &disk_res, sectors, - replicas, 0); - if (unlikely(ret)) - goto bkey_err; - - reservation.v.nr_replicas = disk_res.nr_replicas; - } + ret = bch2_extent_fallocate(&trans, inode_inum(inode), &iter, + sectors, opts, &i_sectors_delta, + writepoint_hashed((unsigned long) current)); + if (ret) + goto bkey_err; - ret = bch2_extent_update(&trans, inode, - &disk_res, "a_res, - iter, &reservation.k_i, - 0, true, true, NULL); + i_sectors_acct(c, inode, "a_res, i_sectors_delta); bkey_err: bch2_quota_reservation_put(c, inode, "a_res); - bch2_disk_reservation_put(c, &disk_res); - if (ret == -EINTR) + if (bch2_err_matches(ret, BCH_ERR_transaction_restart)) ret = 0; + } + + bch2_trans_unlock(&trans); /* lock ordering, before taking pagecache locks: */ + mark_pagecache_reserved(inode, start_sector, iter.pos.offset); + + if (bch2_err_matches(ret, ENOSPC) && (mode & FALLOC_FL_ZERO_RANGE)) { + struct quota_res quota_res = { 0 }; + s64 i_sectors_delta = 0; + + bch2_fpunch_at(&trans, &iter, inode_inum(inode), + end_sector, &i_sectors_delta); + i_sectors_acct(c, inode, "a_res, i_sectors_delta); + bch2_quota_reservation_put(c, inode, "a_res); + } + + bch2_trans_iter_exit(&trans, &iter); + bch2_trans_exit(&trans); + return ret; +} + +static long bchfs_fallocate(struct bch_inode_info *inode, int mode, + loff_t offset, loff_t len) +{ + struct bch_fs *c = inode->v.i_sb->s_fs_info; + u64 end = offset + len; + u64 block_start = round_down(offset, block_bytes(c)); + u64 block_end = round_up(end, block_bytes(c)); + bool truncated_last_page = false; + int ret, ret2 = 0; + + if (!(mode & FALLOC_FL_KEEP_SIZE) && end > inode->v.i_size) { + ret = inode_newsize_ok(&inode->v, end); if (ret) - goto err; + return ret; } - bch2_trans_unlock(&trans); - if (!(mode & FALLOC_FL_KEEP_SIZE) && - end > inode->v.i_size) { + if (mode & FALLOC_FL_ZERO_RANGE) { + ret = bch2_truncate_folios(inode, offset, end); + if (unlikely(ret < 0)) + return ret; + + truncated_last_page = ret; + + truncate_pagecache_range(&inode->v, offset, end - 1); + + block_start = round_up(offset, block_bytes(c)); + block_end = round_down(end, block_bytes(c)); + } + + ret = __bchfs_fallocate(inode, mode, block_start >> 9, block_end >> 9); + + /* + * On -ENOSPC in ZERO_RANGE mode, we still want to do the inode update, + * so that the VFS cache i_size is consistent with the btree i_size: + */ + if (ret && + !(bch2_err_matches(ret, ENOSPC) && (mode & FALLOC_FL_ZERO_RANGE))) + return ret; + + if (mode & FALLOC_FL_KEEP_SIZE && end > inode->v.i_size) + end = inode->v.i_size; + + if (end >= inode->v.i_size && + (((mode & FALLOC_FL_ZERO_RANGE) && !truncated_last_page) || + !(mode & FALLOC_FL_KEEP_SIZE))) { + spin_lock(&inode->v.i_lock); i_size_write(&inode->v, end); + spin_unlock(&inode->v.i_lock); mutex_lock(&inode->ei_update_lock); - ret = bch2_write_inode_size(c, inode, inode->v.i_size, 0); + ret2 = bch2_write_inode_size(c, inode, end, 0); mutex_unlock(&inode->ei_update_lock); } - /* blech */ - if ((mode & FALLOC_FL_KEEP_SIZE) && - (mode & FALLOC_FL_ZERO_RANGE) && - inode->ei_inode.bi_size != inode->v.i_size) { - /* sync appends.. */ - ret = filemap_write_and_wait_range(mapping, - inode->ei_inode.bi_size, S64_MAX); - if (ret) - goto err; + return ret ?: ret2; +} - if (inode->ei_inode.bi_size != inode->v.i_size) { - mutex_lock(&inode->ei_update_lock); - ret = bch2_write_inode_size(c, inode, - inode->v.i_size, 0); - mutex_unlock(&inode->ei_update_lock); +long bch2_fallocate_dispatch(struct file *file, int mode, + loff_t offset, loff_t len) +{ + struct bch_inode_info *inode = file_bch_inode(file); + struct bch_fs *c = inode->v.i_sb->s_fs_info; + long ret; + + if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_fallocate)) + return -EROFS; + + inode_lock(&inode->v); + inode_dio_wait(&inode->v); + bch2_pagecache_block_get(inode); + + ret = file_modified(file); + if (ret) + goto err; + + if (!(mode & ~(FALLOC_FL_KEEP_SIZE|FALLOC_FL_ZERO_RANGE))) + ret = bchfs_fallocate(inode, mode, offset, len); + else if (mode == (FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE)) + ret = bchfs_fpunch(inode, offset, len); + else if (mode == FALLOC_FL_INSERT_RANGE) + ret = bchfs_fcollapse_finsert(inode, offset, len, true); + else if (mode == FALLOC_FL_COLLAPSE_RANGE) + ret = bchfs_fcollapse_finsert(inode, offset, len, false); + else + ret = -EOPNOTSUPP; +err: + bch2_pagecache_block_put(inode); + inode_unlock(&inode->v); + bch2_write_ref_put(c, BCH_WRITE_REF_fallocate); + + return bch2_err_class(ret); +} + +/* + * Take a quota reservation for unallocated blocks in a given file range + * Does not check pagecache + */ +static int quota_reserve_range(struct bch_inode_info *inode, + struct quota_res *res, + u64 start, u64 end) +{ + struct bch_fs *c = inode->v.i_sb->s_fs_info; + struct btree_trans trans; + struct btree_iter iter; + struct bkey_s_c k; + u32 snapshot; + u64 sectors = end - start; + u64 pos = start; + int ret; + + bch2_trans_init(&trans, c, 0, 0); +retry: + bch2_trans_begin(&trans); + + ret = bch2_subvolume_get_snapshot(&trans, inode->ei_subvol, &snapshot); + if (ret) + goto err; + + bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents, + SPOS(inode->v.i_ino, pos, snapshot), 0); + + while (!(ret = btree_trans_too_many_iters(&trans)) && + (k = bch2_btree_iter_peek_upto(&iter, POS(inode->v.i_ino, end - 1))).k && + !(ret = bkey_err(k))) { + if (bkey_extent_is_allocation(k.k)) { + u64 s = min(end, k.k->p.offset) - + max(start, bkey_start_offset(k.k)); + BUG_ON(s > sectors); + sectors -= s; } + bch2_btree_iter_advance(&iter); } + pos = iter.pos.offset; + bch2_trans_iter_exit(&trans, &iter); err: + if (bch2_err_matches(ret, BCH_ERR_transaction_restart)) + goto retry; + bch2_trans_exit(&trans); - pagecache_block_put(&mapping->add_lock); - inode_unlock(&inode->v); - return ret; + + if (ret) + return ret; + + return bch2_quota_reservation_add(c, inode, res, sectors, true); } -long bch2_fallocate_dispatch(struct file *file, int mode, - loff_t offset, loff_t len) +loff_t bch2_remap_file_range(struct file *file_src, loff_t pos_src, + struct file *file_dst, loff_t pos_dst, + loff_t len, unsigned remap_flags) { - struct bch_inode_info *inode = file_bch_inode(file); + struct bch_inode_info *src = file_bch_inode(file_src); + struct bch_inode_info *dst = file_bch_inode(file_dst); + struct bch_fs *c = src->v.i_sb->s_fs_info; + struct quota_res quota_res = { 0 }; + s64 i_sectors_delta = 0; + u64 aligned_len; + loff_t ret = 0; - if (!(mode & ~(FALLOC_FL_KEEP_SIZE|FALLOC_FL_ZERO_RANGE))) - return bch2_fallocate(inode, mode, offset, len); + if (remap_flags & ~(REMAP_FILE_DEDUP|REMAP_FILE_ADVISORY)) + return -EINVAL; + + if (remap_flags & REMAP_FILE_DEDUP) + return -EOPNOTSUPP; + + if ((pos_src & (block_bytes(c) - 1)) || + (pos_dst & (block_bytes(c) - 1))) + return -EINVAL; + + if (src == dst && + abs(pos_src - pos_dst) < len) + return -EINVAL; + + bch2_lock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst); + + inode_dio_wait(&src->v); + inode_dio_wait(&dst->v); - if (mode == (FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE)) - return bch2_fpunch(inode, offset, len); + ret = generic_remap_file_range_prep(file_src, pos_src, + file_dst, pos_dst, + &len, remap_flags); + if (ret < 0 || len == 0) + goto err; - if (mode == FALLOC_FL_COLLAPSE_RANGE) - return bch2_fcollapse(inode, offset, len); + aligned_len = round_up((u64) len, block_bytes(c)); - return -EOPNOTSUPP; + ret = write_invalidate_inode_pages_range(dst->v.i_mapping, + pos_dst, pos_dst + len - 1); + if (ret) + goto err; + + ret = quota_reserve_range(dst, "a_res, pos_dst >> 9, + (pos_dst + aligned_len) >> 9); + if (ret) + goto err; + + file_update_time(file_dst); + + mark_pagecache_unallocated(src, pos_src >> 9, + (pos_src + aligned_len) >> 9); + + ret = bch2_remap_range(c, + inode_inum(dst), pos_dst >> 9, + inode_inum(src), pos_src >> 9, + aligned_len >> 9, + pos_dst + len, &i_sectors_delta); + if (ret < 0) + goto err; + + /* + * due to alignment, we might have remapped slightly more than requsted + */ + ret = min((u64) ret << 9, (u64) len); + + i_sectors_acct(c, dst, "a_res, i_sectors_delta); + + spin_lock(&dst->v.i_lock); + if (pos_dst + ret > dst->v.i_size) + i_size_write(&dst->v, pos_dst + ret); + spin_unlock(&dst->v.i_lock); + + if ((file_dst->f_flags & (__O_SYNC | O_DSYNC)) || + IS_SYNC(file_inode(file_dst))) + ret = bch2_flush_inode(c, dst); +err: + bch2_quota_reservation_put(c, dst, "a_res); + bch2_unlock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst); + + return bch2_err_class(ret); } /* fseek: */ -static bool page_is_data(struct page *page) +static int folio_data_offset(struct folio *folio, loff_t pos) { - EBUG_ON(!PageLocked(page)); + struct bch_folio *s = bch2_folio(folio); + unsigned i, sectors = folio_sectors(folio); - /* XXX: should only have to check PageDirty */ - return PagePrivate(page) && - (page_state(page)->sectors || - page_state(page)->dirty_sectors); + if (s) + for (i = folio_pos_to_s(folio, pos); i < sectors; i++) + if (s->s[i].state >= SECTOR_dirty) + return i << SECTOR_SHIFT; + + return -1; } -static loff_t bch2_next_pagecache_data(struct inode *vinode, +static loff_t bch2_seek_pagecache_data(struct inode *vinode, loff_t start_offset, loff_t end_offset) { - struct address_space *mapping = vinode->i_mapping; - struct page *page; - pgoff_t index; - - for (index = start_offset >> PAGE_SHIFT; - index < end_offset >> PAGE_SHIFT; - index++) { - if (find_get_pages(mapping, &index, 1, &page)) { - lock_page(page); - - if (page_is_data(page)) - end_offset = - min(end_offset, - max(start_offset, - ((loff_t) index) << PAGE_SHIFT)); - unlock_page(page); - put_page(page); - } else { - break; + struct folio_batch fbatch; + pgoff_t start_index = start_offset >> PAGE_SHIFT; + pgoff_t end_index = end_offset >> PAGE_SHIFT; + pgoff_t index = start_index; + unsigned i; + loff_t ret; + int offset; + + folio_batch_init(&fbatch); + + while (filemap_get_folios(vinode->i_mapping, + &index, end_index, &fbatch)) { + for (i = 0; i < folio_batch_count(&fbatch); i++) { + struct folio *folio = fbatch.folios[i]; + + folio_lock(folio); + offset = folio_data_offset(folio, + max(folio_pos(folio), start_offset)); + if (offset >= 0) { + ret = clamp(folio_pos(folio) + offset, + start_offset, end_offset); + folio_unlock(folio); + folio_batch_release(&fbatch); + return ret; + } + folio_unlock(folio); } + folio_batch_release(&fbatch); + cond_resched(); } return end_offset; @@ -2723,72 +3687,96 @@ static loff_t bch2_seek_data(struct file *file, u64 offset) struct bch_inode_info *inode = file_bch_inode(file); struct bch_fs *c = inode->v.i_sb->s_fs_info; struct btree_trans trans; - struct btree_iter *iter; + struct btree_iter iter; struct bkey_s_c k; + subvol_inum inum = inode_inum(inode); u64 isize, next_data = MAX_LFS_FILESIZE; + u32 snapshot; int ret; isize = i_size_read(&inode->v); if (offset >= isize) return -ENXIO; - bch2_trans_init(&trans, c); + bch2_trans_init(&trans, c, 0, 0); +retry: + bch2_trans_begin(&trans); - for_each_btree_key(&trans, iter, BTREE_ID_EXTENTS, - POS(inode->v.i_ino, offset >> 9), 0, k, ret) { - if (k.k->p.inode != inode->v.i_ino) { - break; - } else if (bkey_extent_is_data(k.k)) { + ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot); + if (ret) + goto err; + + for_each_btree_key_upto_norestart(&trans, iter, BTREE_ID_extents, + SPOS(inode->v.i_ino, offset >> 9, snapshot), + POS(inode->v.i_ino, U64_MAX), + 0, k, ret) { + if (bkey_extent_is_data(k.k)) { next_data = max(offset, bkey_start_offset(k.k) << 9); break; } else if (k.k->p.offset >> 9 > isize) break; } + bch2_trans_iter_exit(&trans, &iter); +err: + if (bch2_err_matches(ret, BCH_ERR_transaction_restart)) + goto retry; - ret = bch2_trans_exit(&trans) ?: ret; + bch2_trans_exit(&trans); if (ret) return ret; if (next_data > offset) - next_data = bch2_next_pagecache_data(&inode->v, + next_data = bch2_seek_pagecache_data(&inode->v, offset, next_data); - if (next_data > isize) + if (next_data >= isize) return -ENXIO; return vfs_setpos(file, next_data, MAX_LFS_FILESIZE); } -static bool page_slot_is_data(struct address_space *mapping, pgoff_t index) +static bool folio_hole_offset(struct address_space *mapping, loff_t *offset) { - struct page *page; - bool ret; + struct folio *folio; + struct bch_folio *s; + unsigned i, sectors; + bool ret = true; - page = find_lock_entry(mapping, index); - if (!page || xa_is_value(page)) - return false; + folio = filemap_lock_folio(mapping, *offset >> PAGE_SHIFT); + if (!folio) + return true; + + s = bch2_folio(folio); + if (!s) + goto unlock; - ret = page_is_data(page); - unlock_page(page); + sectors = folio_sectors(folio); + for (i = folio_pos_to_s(folio, *offset); i < sectors; i++) + if (s->s[i].state < SECTOR_dirty) { + *offset = max(*offset, + folio_pos(folio) + (i << SECTOR_SHIFT)); + goto unlock; + } + *offset = folio_end_pos(folio); + ret = false; +unlock: + folio_unlock(folio); return ret; } -static loff_t bch2_next_pagecache_hole(struct inode *vinode, +static loff_t bch2_seek_pagecache_hole(struct inode *vinode, loff_t start_offset, loff_t end_offset) { struct address_space *mapping = vinode->i_mapping; - pgoff_t index; + loff_t offset = start_offset; - for (index = start_offset >> PAGE_SHIFT; - index < end_offset >> PAGE_SHIFT; - index++) - if (!page_slot_is_data(mapping, index)) - end_offset = max(start_offset, - ((loff_t) index) << PAGE_SHIFT); + while (offset < end_offset && + !folio_hole_offset(mapping, &offset)) + ; - return end_offset; + return min(offset, end_offset); } static loff_t bch2_seek_hole(struct file *file, u64 offset) @@ -2796,26 +3784,34 @@ static loff_t bch2_seek_hole(struct file *file, u64 offset) struct bch_inode_info *inode = file_bch_inode(file); struct bch_fs *c = inode->v.i_sb->s_fs_info; struct btree_trans trans; - struct btree_iter *iter; + struct btree_iter iter; struct bkey_s_c k; + subvol_inum inum = inode_inum(inode); u64 isize, next_hole = MAX_LFS_FILESIZE; + u32 snapshot; int ret; isize = i_size_read(&inode->v); if (offset >= isize) return -ENXIO; - bch2_trans_init(&trans, c); + bch2_trans_init(&trans, c, 0, 0); +retry: + bch2_trans_begin(&trans); - for_each_btree_key(&trans, iter, BTREE_ID_EXTENTS, - POS(inode->v.i_ino, offset >> 9), + ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot); + if (ret) + goto err; + + for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents, + SPOS(inode->v.i_ino, offset >> 9, snapshot), BTREE_ITER_SLOTS, k, ret) { if (k.k->p.inode != inode->v.i_ino) { - next_hole = bch2_next_pagecache_hole(&inode->v, + next_hole = bch2_seek_pagecache_hole(&inode->v, offset, MAX_LFS_FILESIZE); break; } else if (!bkey_extent_is_data(k.k)) { - next_hole = bch2_next_pagecache_hole(&inode->v, + next_hole = bch2_seek_pagecache_hole(&inode->v, max(offset, bkey_start_offset(k.k) << 9), k.k->p.offset << 9); @@ -2825,8 +3821,12 @@ static loff_t bch2_seek_hole(struct file *file, u64 offset) offset = max(offset, bkey_start_offset(k.k) << 9); } } + bch2_trans_iter_exit(&trans, &iter); +err: + if (bch2_err_matches(ret, BCH_ERR_transaction_restart)) + goto retry; - ret = bch2_trans_exit(&trans) ?: ret; + bch2_trans_exit(&trans); if (ret) return ret; @@ -2838,22 +3838,31 @@ static loff_t bch2_seek_hole(struct file *file, u64 offset) loff_t bch2_llseek(struct file *file, loff_t offset, int whence) { + loff_t ret; + switch (whence) { case SEEK_SET: case SEEK_CUR: case SEEK_END: - return generic_file_llseek(file, offset, whence); + ret = generic_file_llseek(file, offset, whence); + break; case SEEK_DATA: - return bch2_seek_data(file, offset); + ret = bch2_seek_data(file, offset); + break; case SEEK_HOLE: - return bch2_seek_hole(file, offset); + ret = bch2_seek_hole(file, offset); + break; + default: + ret = -EINVAL; + break; } - return -EINVAL; + return bch2_err_class(ret); } void bch2_fs_fsio_exit(struct bch_fs *c) { + bioset_exit(&c->nocow_flush_bioset); bioset_exit(&c->dio_write_bioset); bioset_exit(&c->dio_read_bioset); bioset_exit(&c->writepage_bioset); @@ -2866,15 +3875,23 @@ int bch2_fs_fsio_init(struct bch_fs *c) pr_verbose_init(c->opts, ""); if (bioset_init(&c->writepage_bioset, - 4, offsetof(struct bch_writepage_io, op.op.wbio.bio), - BIOSET_NEED_BVECS) || - bioset_init(&c->dio_read_bioset, + 4, offsetof(struct bch_writepage_io, op.wbio.bio), + BIOSET_NEED_BVECS)) + return -BCH_ERR_ENOMEM_writepage_bioset_init; + + if (bioset_init(&c->dio_read_bioset, 4, offsetof(struct dio_read, rbio.bio), - BIOSET_NEED_BVECS) || - bioset_init(&c->dio_write_bioset, - 4, offsetof(struct dio_write, iop.op.wbio.bio), BIOSET_NEED_BVECS)) - ret = -ENOMEM; + return -BCH_ERR_ENOMEM_dio_read_bioset_init; + + if (bioset_init(&c->dio_write_bioset, + 4, offsetof(struct dio_write, op.wbio.bio), + BIOSET_NEED_BVECS)) + return -BCH_ERR_ENOMEM_dio_write_bioset_init; + + if (bioset_init(&c->nocow_flush_bioset, + 1, offsetof(struct nocow_flush, bio), 0)) + return -BCH_ERR_ENOMEM_nocow_flush_bioset_init; pr_verbose_init(c->opts, "ret %i", ret); return ret;